Efg 149

Wavelet Analysis and
Its Applications
The subject of wavelet analysis has recently drawn a great deal of attention
from mathematical scientists in various disciplines. It is creating a common
link between mathematicians, physicists, and electrical engineers. This
book series will consist of both monographs and edited volumes on the
theory and applications of this rapidly developing subject. Its objective is to
meet the needs of academic, industrial, and governmental researchers, as
well as to provide instructional material for teaching at both the undergrad-
uate and graduate levels.
This fourth volume of the series is a compilation of twelve papers de-
voted to wavelet analysis of geophysical processes. In addition to an intro-
ductory review article written by the editors, this volume covers such
important areas as atmospheric turbulence, seismic data analysis, detec-
tion of signals in noisy environments, multifractal analysis, and analysis of
long memory geophysical processes. The series editor is grateful to Profes-
sor Foufoula-Georgiou and Dr. Kumar for their effort in compiling and ed-
iting this excellent volume, and would like to thank the authors for their
very fine contributions.
This is a volume in
WAVELET ANALYSIS A N D I T S A P P L I C A T I O N S
CHARLES K. CHUI, SERIES EDITOR

Texas A&M University, College Station, Texas
A list of titles in this series appears at the end of this volume.

Wavelets in
Geophysics
Edited by
Efi Foufoula-Georgiou
Si. Anthony Falls Hydraulic
Laboratory
Department of Civil Engineering
University of Minnesota
Minneapolis, Minnesota
Praveen Kumar
Universities Space Research Association
Hydrological Sciences Branch
NASA-Goddard Space Flight Center
Greenbelt, Maryland
®
ACADEMIC PRESS
San Diego New York Boston
London Sydney Tokyo Toronto
This book is printed on acid-free paper, fe)
Copyright © 1994 by ACADEMIC PRESS, INC.

All Rights Reserved.
No part of this publication may be reproduced or transmitted in any form or by any
means, electronic or mechanical, including photocopy, recording, or any information
storage and retrieval system, without permission in writing from the publisher.
Academic Press, Inc.

A Division of Harcourt Brace & Company
525 B Street, Suite 1900, San Diego, California 92101-4495
United Kingdom Edition published by

Academic Press Limited
24-28 Oval Road, London NW1 7DX
Library of Congress Cataloging-in-Publication Data
Wavelets in geophysics / edited by Efi Foufoula-Georgiou, Praveen

Kumar.
p. cm. — (Wavelet analysis and its applications ; v. 4)
Includes bibliographical references and index.
ISBN 0-12-262850-0
I. Foufoula-Georgiou, Efi, II. Kumar, Praveen. III. Series.
QC806.W38 1994
550'.l'5152433-dc20 94-16953
CIP
PRINTED IN THE UNITED STATES OF AMERICA

94 95 96 97 98 99 MM 9 8 7 6 5 4 3 2 1
To
Katerina, Thomas
and
Ilina
Contributors
Numbers in parentheses indicate the pages on which the authors'contributions begin.
J O H N D. ALBERTSON (81), Hydrologie Science, University of California at

Davis, Davis, California 95616
KEVIN E. BREWER (213), Department of Geological Sciences/172, University of
Nevada, Reno, Reno, Nevada 89557
YVES BRUNET (129), Laboratoire de Bioclimatologie, INRA, BP 81, 33883
Villenave dOrnon, France
CHIA R. CHU (81), Department of Civil Engineering, National Central
University, Chungli, Taiwan
SERGE COLLINEAU (129), Laboratoire de Bioclimatologie, INRA, BP 81, 33883
Villenave dOrnon, France
ANTHONY DAVIS (249), Universities Space Research Association, NASA-
Goddard Space Flight Center, Greenbelt, Maryland 20771
E F I FOUFOULA-GEORGIOU (1), St. Anthony Falls Hydraulic Laboratory,
Department of Civil Engineering, University of Minnesota,
Minneapolis, Minnesota 55414
NlMAL K. K. GAMAGE (45), Program in Atmospheric and Oceanic Sciences,
Astrophysics, Planetary and Atmospheric Sciences Department,
University of Colorado, Boulder, Colorado 80309
PETER GUTTORP (325), Department of Statistics, University of Washington,
Seattle, Washington 98195
CARL R. HAGELBERG (45), National Center for Atmospheric Research, Boulder,
Colorado 80307
J. F. HOWELL (107), Oceanic and Atmospheric Sciences, Oregon State
University, Corvallis, Oregon 97331
GABRIEL G. KATUL (81), School of the Environment, Duke University, Durham,
North Carolina 27708
PRAVEEN KUMAR (1), Universities Space Research Association, Hydrological
Sciences Branch, NASA-Goddard Space Flight Center, Greenbelt,
Maryland 20771
SARAH A. LITTLE (167), Department of Geology and Geophysics, Woods Hole
Océanographie Institution, Woods Hole, Massachusetts 02543
ix
x Contributors
PAUL C. LIU (151), NOAA Great Lakes Environmental Research Laboratory,

Ann Arbor, Michigan 48109
L. MAHRT (107), Oceanic and Atmospheric Sciences, Oregon State University,
Corvallis, Oregon 97331
ALEXANDER MARSHAK (249), Science Systems and Applications, Inc., Lanham,
Maryland 20706
MARC B. PARLANGE (81), Hydrologie Science, and Department of Agricultural
and Biological Engineering, University of California at Davis, Davis,
California 95616
DONALD B. PERCIVAL (325), Applied Physics Laboratory, University of
Washington, Seattle, Washington 98195
CHRIS J. PIKE (183), Centre for Cold Ocean Resources Engineering (C-CORE),
Memorial University of Newfoundland, St. John's, Newfoundland,
Canada AIB 3X5
NAOKI SAITO (299), Schlumberger-Doll Research, Old Quarry Road,
Ridgefield, Connecticut 06877, and Department of Mathematics, Yale
University, New Haven, Connecticut 06520
STEPHEN W. WHEATCRAFT (213), Department of Geological Sciences/172,
University of Nevada, Reno, Reno, Nevada 89557
WARREN WISCOMBE (249), NASA-Goddard Space Flight Center, Climate and
Radiation Branch, Greenbelt, Maryland 20771
Preface
Over the past decade, wavelet transforms have been formalized into a rigorous
mathematical framework and have found numerous applications in diverse
areas such as harmonic analysis, numerical analysis, signal and image pro-
cessing, nonlinear dynamics, fractal and multifractal analysis, and others.
Although wavelet transforms originated in geophysics (for the analysis of
seismic signals), it is only very recently that they are being used again in
the geophysical sciences. Properties t h a t make wavelets attractive are t i m e -
frequency localization, orthogonality, multirate filtering, and scale-space
analysis, to name a few.
This volume is the first collection of papers using wavelet transforms for
the understanding, analysis, and description of geophysical processes. It in-
cludes applications of wavelets to atmospheric turbulence, ocean wind waves,
characterization of hydraulic conductivity, seafloor bathymetry, seismic data,
detection of signals from noisy data, multifractal analysis, and analysis of long
memory geophysical processes. Most of the papers included in this volume were
presented at the American Geophysical Union (AGU) Spring Meeting in Balti-
more, May 1993, in a special Union session organized by us entitled "Applica-
tions of Wavelet Transforms in Geophysics." We feel t h a t this volume will serve
geophysicists as an introduction to the versatile and powerful wavelet analysis
tools and will stimulate further applications of wavelets in geophysics as well
as mathematical developments dictated by unique demands of applications.
The first chapter in this volume is a review article by Kumar and Foufoula-
Georgiou. The purpose of this article is to provide the unfamiliar reader with a
basic introduction to wavelets and key references for further study. Wavelet
transforms are contrasted with the Fourier transforms and windowed Fourier
transforms that are well known to geophysicists; this contrast highlights the
important property of time—frequency localization in wavelet transforms,
which is essential for the analysis of nonstationary and transient signals.
Continuous and discrete, as well as orthogonal, nonorthogonal, and biortho-
gonal wavelet transforms are then reviewed, and the concept of multiresolution
analysis is presented. Several examples of one- and two-dimensional wavelets
and information on wavelet construction are given. Finally, some sources of
available wavelet analysis software packages are included which may help the
interested reader get started in exploring wavelets.
The next four chapters present results from the application of wavelet
analysis to atmospheric turbulence. In Chapter 2, Hagelberg and Gamage
develop a wavelet-based signal decomposition technique t h a t preserves inter-
mittent coherent structures. Coherent structures in velocity and temperature
in the atmospheric boundary layer account for a large portion of flux transport
of momentum, heat, trace chemicals, and particulates. The authors' technique
partitions signals into two components: one containing coherent structures
characterized by sharp transitions and intermittent occurrence, and the other
xi
Xll Preface
containing the remaining portion of the signal (essentially characterized by

smaller length scales and the absence of coherent events). They apply this
decomposition to vertical velocity, virtual potential temperature, and buoyancy
flux density fields. Chapter 3, by Katul, Albertson, Chu, and Parlange, applies
orthonormal wavelets to atmospheric surface layer velocity measurements to
describe space-time relations in the inertial subrange. The local nature of the
orthonormal wavelet transform in physical space aids the identification of
events contributing to inertial subrange intermittency buildup, which can then
be suppressed to eliminate intermittency effects on the statistical structure of
the inertial subrange.
In Chapter 4, Howell and Mahrt develop an adaptive method for decom-
posing a time series into orthogonal modes of variation. In contrast to conven-
tional partitioning, the cutoff scales are allowed to vary with record position
according to the local physics of the flow by utilizing the Haar wavelet decom-
position. For turbulence data, this decomposition is used to distinguish four
modes of variation. The two larger modes, determined by spatially constant
cutoff scales, are characterized as the mesoscale and large eddy modes. The
two smaller scale modes are separated by a scale that depends on the local
transport characteristics of the flow. This adaptive cutoff scale separates the
transporting eddy mode, responsible for most of the flux, from the nontrans-
porting nearly isotropic motions. Chapter 5, by Brunet and Collineau, applies
wavelets to the analysis of turbulent motions above a maize crop. Their results
indicate that organized turbulence exhibits the same structure above a forest
and a maize crop, apart from a scale factor, and supports the interesting pos-
tulate that transfer processes over plant canopies are dominated by popula-
tions of canopy-scale eddies, with universal characteristics. The authors also
propose a methodology for separating turbulence data into large- and small-
scale components using the filtering properties of the wavelet transform.
In Chapter 6, Liu applies wavelet spectrum analysis to ocean wind waves.
The results reveal significant new insights on wave grouping parameteriza-
tions, phase relations during wind wave growth, and detection of wave breaking
characteristics. Chapter 7, by Little, demonstrates the usefulness of wavelet
analysis in studying seafloor bathymetry and especially identifying the loca-
tion and scarp-facing direction of ridge-parallel faulting. In Chapter 8, Pike
proposes a wavelet-based methodology for the analysis of high resolution
acoustic signals for sub-seabed feature extraction and classification of scatter-
ers. The key idea is that of displaying the energy dissipation in the t i m e -
frequency plane, which allows a more distinct description of the signal (com-
pared to the Fourier transform spectra) and thus provides a better means of
extracting seabed properties by correlating them to the attenuation of high
resolution acoustic signals. In Chapter 9, Brewer and Wheatcraft investigate
the wavelet transform as a tool for reconstructing small-scale variability in
hydraulic conductivity fields by incorporating the scale and location informa-
tion of each sample when interpolating to a finer grid. The developed multiscale
reconstruction method is compared to traditional interpolation schemes and is
used to examine the issue of optimum sample size and density for stationary
and fractal random fields.
Preface Xlll
Chapter 10, by Davis, Marshak, and Wiscombe, explores the use of wave-
lets for multifractal analysis of geophysical phenomena. They show the appli-
cability of wavelet transforms to compute simple yet dynamically meaningful
statistical properties of one dimensional geophysical series. Turbulent velocity
and cloud liquid water content are used as examples to demonstrate the need
for stochastic models having both additive (nonstationary) and multiplicative
(intermittent) features. Merging wavelet and multifractal analysis seems
promising for both wavelet and multifractal communities and is especially
promising for geophysics, where many signals show structures at all observa-
ble scales and are often successfully described within a multifractal frame-
work.
The wavelet transform partitions the frequency axis in a particular way:
it iteratively partitions the low-frequency components, leaving the high-fre-
quency components intact at each iteration. For some processes or applications
this partition might not achieve the best decomposition, as partition of the
high-frequency bands might also be necessary. Wavelet packets provide such a
partition and are used in Chapter 11, by Saito, for simultaneous signal
compression and noise reduction of geophysical signals. A maximum entropy
criterion is used to obtain the best basis out of the many bases t h a t the redun-
dant wavelet packet representation provides. The method is applied to syn-
thetic signals and to some geophysical data, for example, a radioactivity profile
of subsurface formation and a migrated seismic section. Finally, in Chapter 12,
Percival and Guttorp examine a particular measure of variability for long
memory processes (the Allan variance) within the wavelet framework and show
that this variance can be interpreted as a Haar wavelet coefficient variance.
This suggests an approach to assessing the variability of general wavelet
classes which will be useful in the study of power-law processes extensively
used for the description of geophysical time series. A fairly extensive bibliog-
raphy of wavelet analysis in geophysics is included at the end of this volume.
Several individuals provided invaluable help in the completion of this vol-
ume. Special t h a n k s go to the reviewers of the book chapters who volunteered
their time and expertise and provided timely and thoughtful reviews. The first
author t h a n k s Mike Jasinski, Hydrologie Sciences Branch at NASA-Goddard
Space Flight Center and the Universities Space Research Association for their
support during the completion of this project. We are grateful to Charu Gupta
Kumar, who converted most of the chapters to I^T^X format and typeset and
edited the entire volume. Without her expertise and dedication the timely
completion of this volume would not have been possible. Finally, we also t h a n k
the Academic Press Editor, Peter Renz, for his efficient help during the final
stages of this project.
Efi Foufoula-Georgiou Praveen Kumar

Minneapolis, Minnesota Greenbelt, Maryland
March, 1994
W a v e l e t A n a l y s i s in G e o p h y s i c s : A n I n t r o d u c t i o n
Praveen K u m a r and Efi Foufoula-Georgiou
Abstract. Wavelet analysis is a rapidly developing area of mathematical and

application-oriented research in many disciplines of science and engineering. The
wavelet transform is a localized transform in both space (time) and frequency, and
this property can be advantageously used to extract information from a signal that
is not possible to unravel with a Fourier or even windowed Fourier transform.
Wavelet transforms originated in geophysics in early 1980's for the analysis of
seismic signal. After a decade of significant mathematical formalism they are
now also being exploited for the analysis of several other geophysical processes
such as atmospheric turbulence, space-time rainfall, ocean wind waves, seafloor
bathymetry, geologic layered structures, climate change, among others. Due to
their unique properties, well suited for the analysis of natural phenomena, it is
anticipated that there will be an explosion of wavelet applications in geophysics
in the next several years. This chapter provides a basic introduction to wavelet
transforms and their most important properties. The theory and applications of
wavelets is developing very rapidly and we see this chapter only as a limited basic
introduction to wavelets which we hope to be of help to the unfamiliar reader and
provide motivation and references for further study.
§1. P r o l o g u e
T h e concept of wavelet transforms was formalized in early 1980's in a
series of p a p e r s by Morlet et al. [42, 43], Grossmann and Morlet [24], a n d
Goupillaud, Grossmann and Morlet [23]. Since this formalism a n d some
significant work by Meyer ([38, 39] a n d references therein), Mallat [30, 31],
Daubechies [13, 15], a n d Chui [6] among others, t h e wavelets have become
pervasive in several diverse areas such as m a t h e m a t i c s , physics, digital sig-
nal processing, vision, numerical analysis, and geophysics, t o n a m e a few.
Wavelet transforms are integral transforms using integration kernels called
wavelets. These wavelets are essentially used in two ways when studying
Wavelets in Geophysics 1
Efi Foufoula-Georgiou and Praveen Kumar (eds.), pp. 1-43.
Copyright 1994 by Academic Press, Inc.
All rights of reproduction in any form reserved.
ISBN 0-12-262850-0
2 P. Kumar and E. Foufoula-Georgiou
processes or signals: (i) as an integration kernel for analysis to extract

information about the process, and (ii) as a basis for representation or
characterization of the process. Evidently, in any analysis or representa-
tion, the choice of the basis function (or kernel) determines t h e kind of
information t h a t can be extracted about the process. This leads us to the
following questions: (1) what kind of information can we extract using
wavelets? and (2) how can we obtain a representation or description of the
process using wavelets?
The answer to the first question lies on the important property of
wavelets called time-frequency localization. The advantage of analyzing
a signal with wavelets as the analyzing kernels is t h a t it enables one to
study features of the signal locally with a detail matched to their scale,
i.e., broad features on a large scale and fine features on small scales. This
property is especially useful for signals t h a t are either non-stationary, or
have short lived transient components, or have features at different scales,
or have singularities. The answer to the second question is based on seeing
wavelets as the elementary building blocks in a decomposition or series
expansion akin to the familiar Fourier series. Thus, a representation of
the process using wavelets is provided by an infinite series expansion of
dilated and translated versions of a mother wavelet, each multiplied by an
appropriate coefficient. For processes with finite energy this wavelet series
expansion is optimal, i.e., it offers an optimal approximation to the original
signal, in the least squares sense.
In what follows we give a brief introduction to the mathematics of
wavelet transforms and where possible an intuitive explanation of these
results. T h e intention of this introduction is two fold: (i) to provide the
unfamiliar reader with a basic introduction to wavelets, and (ii) to prepare
the reader to grasp and appreciate the results of the articles t h a t follow
as well as the potential of wavelet analysis for geophysical processes. At
times we have sacrificed mathematical rigor for clarity of presentation in an
a t t e m p t to not obscure the basic idea with too much detail. We also hasten
to add t h a t this review is far from complete both in terms of the b r e a d t h of
topics chosen for exposition and in terms of the treatment of these topics.
For example, the important topic of wavelet packets has not been discussed
(see for example [53], [8], and the article by Saito [50] in this volume). It
is recommended t h a t the interested reader who is not a mathematician
and is meeting wavelet analysis for the first time begins with the book by
Meyer [39] and continues with the books by Daubechies [15] and Benedetto
and Frazier [3]. There are also several nice introductory articles on several
aspects of wavelets, as for example, those by Mallat [30, 31], Rioul and
Vetterli [47], and the article by Farge [19] on turbulence, among others.
Also, considerable insight can be gained by the nice book reviews by Meyer
[40] and Daubechies [18].
Wavelet Analysis in Geophysics: An Introduction 3
This article is organized as follows. Section 2 discusses Fourier and

windowed Fourier transforms and introduces continuous wavelet trans-
forms and their time-frequency localization properties. In section 3, the
wavelet transform is presented as a time-scale transform, and the wavelet
variance and covariance (alternatively called wavelet spectrum and cospec-
t r u m ) are discussed. A link is also made between non-stationary processes
and wavelet transforms akin to the link between stationary processes and
Fourier transforms. Section 4 presents some examples of commonly used
one-dimensional wavelets (Haar wavelet, Mexican hat wavelet, and Mor-
let wavelet). In section 5, discrete wavelet transforms (orthogonal, non-
orthogonal, and biorthogonal) are introduced and the concept of multires-
olution analysis presented. In section 6, we present extensions of contin-
uous and discrete wavelets in a two-dimensional space. Finally, in section
7 we present some concluding remarks and give information on obtaining
available software packages for wavelet analysis.
§2. T i m e - F r e q u e n c y A n a l y s i s
T h e original motive for the development of wavelet transform was (see
[23]) " . . . of devising a m e t h o d of acquisition, transformation and recording
of a seismic trace (i.e., a function of one variable, the time) so as to satisfy
the requirements listed below:
1. T h e contributions of different frequency bands (i.e., of the differ-

ent intervals of the Fourier conjugate variable) are kept reasonably
separated.
2. This separation is achieved without excessive loss of resolution in the

time variable (subject, of course, to the limitation of the uncertainty
principle).
3. T h e reconstruction of the original function from its "representation"

or "transform" is obtained by a method which is (a) capable of
giving arbitrary high precision; (b) is robust, in t h e sense of being
stable under small perturbations. "
T h e first two conditions essentially characterize the property known as

time-frequency localization. Recall t h a t although the Fourier transform of
a function / ( £ ) , given as
oo
f(t)e-^dt, (1)
/
-oo
gives t h e information about the frequency content of a process or signal,
it gives no information about the location of these frequencies in t h e time
domain. For example figures l a , b show two signals - the first consisting of
50 100 50 100
angular frequency angular frequency
Figure 1. Spectral and wavelet analysis of two signals. The first sig-
nal (a) (upper left) consists of superposition of two frequencies (sinlOt
and sin20£), and the second consists of the same two frequencies each
applied separately over half of the signal duration (b) (upper right).
Figures (c) (middle left) and (d) (middle right) show the spectra
of signals, i.e., |/(ω)| 2 vs ω, in (a) and (b) respectively, and (e) (lower
left) and (f) (lower right) show the magnitude of their wavelet trans-
forms (using Morlet wavelet) respectively.
two frequencies (sin 10t and sin 20t) superimposed for the entire duration
of the signal and the second consisting of the same frequencies, b u t each
one applied separately for half of signal duration. Figures lc,d show t h e
spectrum, i.e., |/(o;)| 2 , of these two signals, respectively. As is clearly
evident, the spectrum is quite incapable of distinguishing between the two
signals.
Time varying frequencies are quite common in music, speech, seismic
signals, non-stationary geophysical processes, etc. To study such processes,
one seeks transforms which will enable one to obtain the frequency content
of a process locally in time. There are essentially two methods t h a t have
been developed to achieve this (within the limits of the uncertainty principle
which states t h a t one cannot obtain arbitrary good localization in b o t h

time and frequency): (a) windowed Fourier transform, and (b) wavelet
transform. These two methods are discussed in the following subsections.
Figures le,f display t h e magnitude of the wavelet transform of t h e signals
shown in Figures l a , b , and clearly show the ability of the wavelet transform
to distinguish between the two signals.
2.1. W i n d o w e d Fourier t r a n s f o r m
2.1.1. Definition
In t h e Fourier transform framework, time localization can be achieved
by windowing the d a t a at various times, say, using a windowing function
#(£), and then taking the Fourier transform. T h a t is, the windowed Fourier
transform (also called the short-time fourier transform), G/(CJ, i ) , is given
by
(2)
(3)
where t h e integration kernel is gtj,t(u) = g(u — t)e~iuju. This transform

measures locally, around the point £, the amplitude of the sinusoidal wave
component of frequency ω. The window function g(t) is usually chosen
as a real, even function with the maximum concentration of energy in t h e
low frequency components. Notice t h a t the analyzing kernel <7ω,*(ΐί) has
the same support 1 for all ω and £, but the number of cycles vary with the
frequency ω (see Figure 2).
T h e representation of the function f(t) on the time-frequency plane,
i.e., (CJ,£) plane, thus obtained is called the phase-space representation.
T h e windowed Fourier transform is an energy preserving transforma-
tion or isometry, i.e.,
(4)
provided / ^ |g(£)| 2 dx = 1 (which we assume from here on). It is invertible

with t h e reconstruction formula given as ([31], eq. 15)
(5)
T h e parameters t and ω can b e assigned discrete values, say t = nt0
1
support is defined as the closure of the set over which the signal/process is non-zero.
Figure 2. Real (solid lines) and imaginary parts (dot-dashed lines)

of the analyzing kernel g{t)e~tuJt of the windowed Fourier transform at
different frequencies: (a) (top) ω = 3, (b) (middle) ω = 6 and (c)
(bottom) ω = 9. The dotted line indicates a Gaussian window function
g(t).
and ω = mcjo, and we obtain the discrete windowed Fourier transform

oo
f{u)g{u-nt0)e-im"°udu. (6)
/ -OO
For the discrete windowed Fourier transform to be invert ible, the condition
CJOÔ < 2π must hold (see [15], sections 3.4 and 4.1).
2 Λ . 2 . T i m e - f r e q u e n c y localization
In order to study the time-frequency localization property of the win-
dowed Fourier transform, we need to study the properties of |<7ω,*|2 and
\çju>,t | 2 since they determine the features of f(t) t h a t are extracted. Indeed,
using Parseval's theorem, equation (3) can be written as
1 f°° A
(7)
Figure 3. Uncertainties in time (top) and frequency (bottom) lo-

calization in a windowed Fourier Transform for a generic function g{t).
where ^ ^ ( u / ) is the Fourier transform of gu,t(u) and overbar indicates

complex conjugate. Let us define the standard deviations of gu^ and gu^
as ag and ag respectively, i.e.,
(8)
and
(9)
These parameters measure the spread of the function | ^ ? < | and \gUjt |, about
t and CJ, respectively (see Figure 3). Owing to the uncertainty principle,
the products of σ2, and σ | satisfy (see [31])
(10)
i.e., arbitrary high precision in b o t h time and frequency cannot b e achieved.

T h e equality in the above equation is achieved only when g(t) is t h e Gaus-
sian, i.e.,
(H)
W h e n the Gaussian function is used as a window, the windowed Fourier
transform is called the Gabor transform [22].
σ
σβ
' ΡJ
1 ^Λ 1 ^Λ
1 8, | §
° g .1 σ
*
1
σ Λ|
g)
1
1
1 ^
U
σ σ
8 . 8
1 ^Λ 1 ^Λ
1 8 | §
to
F i g u r e 4. Figure showing the phase-space representation using the

windowed Fourier transform.
Once a window function g(t) is chosen b o t h ag and σ§ are fixed. There-

fore, for any given to and CJQ, the time-frequency resolution can b e repre-
sented by the fixed size resolution cell [t0 ±σ9 χ ωο±σ9] (see Figure 4), i.e.,
the windowed Fourier transform at any point (£o 5 ô) in the phase-space
provides information about f(t) t h a t is localized with an uncertainty of ag
in the time domain and σ9 in the Fourier domain, and this localization is
uniform in the entire phase-space. In other words the entire phase-space
is uniformly layered with resolution cells or "bricks" of fixed dimensions.
This poses two kinds of limitations. Firstly, if the process has a transient
component with a support smaller t h a n σ^, it is difficult to locate it with
precision b e t t e r t h a n ag. Secondly if the process has important features
of differing sizes then we can not find an optimal g{i) for analyzing the
process. Therefore, window Fourier transform is more suited for analyzing
processes where all the features appear approximately at the same scale.
T h e wavelet transform addresses the limitations inherent in the windowed
Fourier transform.
2.2. Wavelet transform

In the windowed Fourier transform, the analyzing functions g„t for all
ω and t consist of the the same envelope g(t) filled in with sinusoids of
frequency ω. Due to the fixed envelope g(t), the resolution cell size in t h e
phase space given by [σ9 χ σ$] is the same for all ω and t. Since higher
frequency (or short wavelength) features have smaller support, it would be

desirable to have an analyzing function, say φ(ί), such t h a t its s t a n d a r d
deviation σψ is small when φ{ϊ) characterizes high frequency components
and vice-versa. This was achieved by decomposing the function f(t) using
a two p a r a m e t e r family of functions called wavelets (see [42] and [43]). One
of t h e two parameters is the translation parameter as in the windowed
Fourier transform case, but the other parameter is a dilation parameter λ
instead of the frequency parameter ω.
2.2.1. Definition
T h e wavelet transform of a function f(t) with finite energy is defined
as t h e integral transform with a family of functions xjj\)t{u) = ~7χΦ(Μγ^)
and is given as
(12)
Here λ is a scale parameter, t a location parameter and the functions ip\jt(u)

are called wavelets. In case φ\^(μ) is complex, we use the complex con-
jugate φχ t(u) in the above integration. Changing the value of λ has the
effect of dilating (λ > 1) or contracting (λ < 1) the function φ(ί) (see Fig-
ure 5a), and changing t has the effect of analyzing the function f(t) around
the point t. T h e normalizing constant -4^ is chosen so t h a t
for all scales λ (notice the identity φ(ί) = V>i,o(0)· We also choose the
normalization j \φ(ί)\2 dt = 1. The wavelet transform Wf(X^t) is often
denoted as t h e inner product (/, φ\$)-
Notice t h a t in contrast to the windowed Fourier transform case, t h e
number of cycles in the wavelet φ\^(η) does not change with t h e dilation
(scale) p a r a m e t e r λ but the support length does. We will see shortly t h a t
when λ is small, which corresponds to small support length, t h e wavelet
transform picks u p higher frequency components and vice-versa.
T h e choice of the wavelet φ{ί) is neither unique nor arbitrary. T h e
function φ{ί) is a function with unit energy chosen so t h a t it has:
1. compact support, or sufficiently fast decay, to obtain localization in

space;
2. zero mean, i.e., J ^ t / ^ ^ d i = 0, although higher order moments

|ψ(ω)|
Figure 5. Schematic illustration of the effect of dilation on a "generic"

wavelet (top) and the corresponding change on its Fourier transform
|^(ω)| (bottom). When the wavelet dilates, its Fourier transform con-
tracts and vice-versa, (a) λ < 1, (b) λ = 1, and (c) λ > 1.
may also be zero, i.e.,
tktp(t) dt = 0 for k = 0 , . . . , N - 1. (13)

/
T h e requirement of zero mean is called the admissibility condition
of the wavelet. It is because of the above two properties t h a t the
function i/j(t) is called a wavelet. The second property ensures t h a t
ip(t) has a wiggle, i.e., is wave like, and the first ensures t h a t it is
not a sustaining wave.
T h e inverse wavelet transform is given by ([15], eq. 2.4.4)

1 /«oo /*oo
f(t) = 7T / X-2Wf(X,u)<pXiU(t)dXdu (14)
W J-ooJO
where
(15)
T h e wavelet transform is also an energy preserving transformation, i.e., an

isometry (up to a proportionality constant), t h a t is,
(16)
2 . 2 . 2 . T i m e - f r e q u e n c y localization
In order to understand the behavior of the wavelet transform in the fre-
quency domain as well, it is useful to recognize t h a t the wavelet transform
W/(A,£), using Parseval's theorem, can be equivalently written as
(17)
Therefore, as in the windowed Fourier transform, we need to study the

properties of \ip\}t(u)\2 and |Â,<(<^)| 2 to understand the time-frequency
localization properties of wavelet transforms. Specifically, we need to un-
derstand the behavior of the standard deviations of |^λ,*| 2 and |^λ,*| 2 , i-e->
<τψχ>ί and σ? . Note t h a t , due to property (13), ψ\^(ω = 0) = 0. Conse-
quently, the center of passing band, CJ°~ , for tp\yt(t) is located away from
the origin ω = 0 (as shown in Figure 5b). It can be obtained as the center
of mass (or first moment about the origin) of the right lobe as
(18)
We therefore define the standard deviation (i.e., square root of the second
central moment of the right lobe) σ τ as
(19)
Similarly in the time domain the standard deviation σφλ t can b e obtained
as
(20)
where to is given as
(21)
It is easy to verify t h a t the following relationships hold:

1. T h e standard deviation σφχ t satisfies
σ
Φχ,ι = λσ
Φι,ο- ( 22 )
2. T h e standard deviation σ;, satisfies
3. T h e center of passing band αΛ corresponding to the wavelet iß\yt (u)

satisfies the relationship
ω°; = - ^ . (24)
From the above relationships one can easily see that as λ increases, i.e., as
the function dilates, both ω°7 and σ,?, decrease indicating that the center
of passing band shifts towards low frequency components and the uncer-
tainty also decreases, and vice-versa (see also figure 5). In the phase-space,
the resolution cell for the wavelet transform around the point (ίο,ω^ )
is given by [to ± λσ^ 1 0 x Alf0 ± Alf0] (see Figure 6) which has variable
dimensions depending on the scale parameter λ. However, the area of the
resolution cell [σφχ t x σ ? ] remains independent of the scale or location
parameter. In other words, the phase space is layered with resolution cells
of varying dimensions which are functions of scale such that they have a
constant area. Therefore, due to the uncertainty principle, an increased
resolution in the time domain for the time localization of high frequency
components comes with a cost: an increased uncertainty in the frequency
localization as measured by σ ? . One may also interpret the wavelet
transform as a mathematical microscope where the magnification is given
by 1/λ.
§3. Wavelets and Time-Scale Analysis

3.1. Time-scale transform
Useful information can also be extracted by interpreting the wavelet
transform (12) as a time-scale transform. This was well illustrated by Rioul
and Vetterli (see [47]) and is sketched below. In the wavelet transform (12)
when the scale λ increases, the wavelet becomes more spread out and takes
only long time behavior into account, as seen above. However by change
of variables, equation (12) can also be written as
(25)
HL
λσ^
Figure 6. Figure showing the phase-space representation using the

wavelet transform.
Since the mapping f(t) —> f(Xt) has the effect of contracting f(t) when
λ > 1 and magnifying it when λ < 1, the above equation indicates t h a t as
the scale grows, a contracted version of the function is seen through a fixed
size filter and vice-versa. Thus, the scale factor λ has the interpretation of
the scale in maps.
3 . 2 . S c a l o g r a m , w a v e l e t variance a n d covariance
From the isometry of wavelet transform (16) we have
(26)
In general, for two functions f(t) and g(t) (see [15], equation 2.4.2)
(27)
By considering the RHS of (26) we see t h a t |W/(A,£)| 2 /CÂ 2 can b e con-

sidered as an energy density function on the phase-space or (£, λ) plane, i.e.,
|W/(A, ί)\2ΑίΑλ/Οψ\2 gives the energy on the scale interval Δ λ and time
interval At centered around scale λ and time t. Flandrin (see [21]) pro-
posed to call the function |W/(A,£)| 2 a scalogram. In analogy, the product
Wf(\,t)Wg(\,t) can be called a cross scalogram.
Equation (26) can also be written as
(28)
where
(29)
gives the energy content of a function f(t) at scale λ, i.e., it gives the
marginal density function of energy at different scales λ. The function
E(X) has been referred to as wavelet variance (see [4]) or wavelet spectrum
(see [25]). In analogy, the function
(30)
has been referred to as wavelet covariance (see [4]) or wavelet cross-spectrum

(see [25]).
Notice t h a t for a given wavelet ip(t) the center of passing band ω°«
at scale λ is related to t h a t at unit scale through the relation (see equation
(24))
(31)
Using this relationship, the scale information can be translated to frequency

information. Using άω — —ω^ X~2dX and substituting in equation (28)
we get
n
(32)
By defining
(33)
the above equation can be written as
(34)
One would therefore expect t h a t Ε'(ω), and thus JE7(A), is related to the
power spectrum S/(u>) of f(t). This indeed is the case. It can be shown
(see [25]) t h a t
(35)
where Ξψχ (ω) is the spectrum of the wavelet at scale λ. T h a t is, E(s) is
the weighted average of the power spectrum of f(t) where the weights are
given by the power spectrum of ip\(t). This relation is interesting although

in characterizing a process through E(X) or Ε'{ώ), all location information
is lost, it does provide certain useful insight ([36], and [27]).
3.3. Non-stationarity and the Wigner-Ville spectrum

One reason for the remarkable success of the Fourier transform in the
study of stationary stochastic processes is the relationship between t h e
autocorrelation function and the spectrum as illustrated by the following
diagram:
T
X(t) £± Χ(ω)
R(r) = 8[X(t)X{t - r)] < = i S(u>) = \Χ(ω) | 2
where R(r) and 5(CJ) are the auto-covariance function and the power spec-
t r u m of the stochastic process X(t), respectively. If an analogous rela-
tionship could be developed for non-stationary processes using the wavelet
transform, then the properties of the wavelet transform could be harnessed
in a more useful way. It turns out t h a t , indeed, such a relationship can b e
developed.
T h e wavelet spectrum E(X) or Ε'(ω) discussed in the previous sec-
tion although interesting in its own right, takes us away from t h e non-
stationarity of the process since it is obtained by integrating over t. We,
therefore, need something else. This is provided by the Wigner-Ville spec-
t r u m . Let us define a general (non-stationary) covariance function R(t, s)
as
R(t,s)=S[X(t)X(s)].
T h e n t h e Wigner-Ville spectrum (WVS) is defined as (see [11] for a dis-
cussion of WVS and other time-frequency distributions)
oo
■R(t+^,t--)e-iurdT. (36)
/
The WVSx(t, ω) is an energy density function as
oo
WVSx(t,w)duj (37)
/
-co
i.e., we get the instantaneous energy by integrating over all frequencies,
and t h e total energy can be obtained as
(38)
16 P. Kumar and E. Foufoula-Georgiern
T h e relationship of interest to us is given by the relation between the scalo-

gram and the WVS
(39)
i.e., the scalogram can be obtained by affine smoothing (i.e., smoothing at

different scales in t and ω directions) of the WVS of X with the WVS of
the wavelet. This relationship has been developed by Flandrin (see [21]).
As of this writing, we are unaware of any inverse relation to obtain t h e
WVSx from the scalogram. We can put the key result of this subsection
in the following diagrammatic form:
We, therefore, see t h a t there is an inherent link between the study of non-
stationary processes and wavelet transforms akin to the link between sta-
tionary processes and Fourier transforms.
§4. E x a m p l e s of O n e - D i m e n s i o n a l W a v e l e t s
Due to the flexibility in choosing a wavelet, several functions have been
used as wavelets and it would be difficult to provide an exhaustive list. We
present here some commonly used wavelets (Haar wavelet, Mexican h a t
wavelet, and Morlet wavelet) in one-dimensional applications.
4.1. Haar wavelet

T h e Haar wavelet is the simplest of all wavelets and is given as
(40)
In a one-dimensional discretely sampled signal this wavelet can b e seen

as performing a differencing operation, i.e., as giving differences of non-
overlapping averages of observations. In two dimensions an interpretation
of the discrete orthogonal Haar wavelet transform has been given in [28].
4 . 2 . M e x i c a n hat w a v e l e t
T h e Mexican hat wavelet is the second derivative of t h e Gaussian e~l ' 2
given as (see Figure 7)
(41)
Figure 7. Mexican hat wavelet.
T h e constant is chosen such t h a t || φ \\2= 1. This wavelet, being the

second derivative of a commonly used smoothing function (the Gaussian),
has found application in edge detection (see [34] and [35]).
4.3. Morlet wavelet

T h e Morlet wavelet is given by
φ(ή = ττ" 1 / 4 ^-^- 6 " Φ* )e

-t2/2 (42)
which is usually approximated as
φ(ή _ _-1/4ρ-ίωο*ρ-*2/2 CJO > 5. (43)

Since for ωο > 5, the second term in (42) is negligible, i.e., φ(ί) « 0,
satisfying the admissibility condition. By Morlet wavelet we now refer t o
(43). This wavelet is complex, enabling one to extract information a b o u t
the amplitude and phase of the process being analyzed. T h e constant is
chosen so t h a t || φ \\2= 1. The Fourier transform of (43) is given by
-(ω-ω0)2/2
φ(ω) _=_π
-l/4 (44)
This wavelet has been used quite often in analysis of geophysical pro-
cesses (for e.g. see [45]) so we shall study it in a little more detail. T h e
Fourier transform of the scaled wavelet φ\}ο(ί) is given as
^λ,ο(ω) = χπ-ι/*β-^-^)'/2 = λ π _ι/46-^- ω ) *_
This wavelet h a s t h e property t h a t its Fourier transform is supported

Real and imaginary parts of Morlet wavelet
0.5l·
Oh
\y : . .\y
3
(a)
Spectrum of Morlet wavelet
0.8
0.6
0.2
-2 0 2 10
frequency
(b)
F i g u r e 8. (a) (top) Real (solid line) and imaginary part (dot-dashed

line) of Morlet wavelet (ωο = 5), and (b) (bottom) its Fourier trans-
form.
almost 2 entirely on ω > 0, centered at u°j — ωο/λ with a spread of

σ = l / λ . T h e wavelet φχ^ itself is centered at t with a spread of
Φχ,
σ
Φχ, = λ.
Figure 8a shows the real and imaginary parts of the Morlet wavelet at
unit scale and Figure 8b shows its Fourier transform (with ω υ = 5 ) . One can
interpret the results of analysis of real-valued processes using this wavelet
by plotting the square of the modulus and the phase, i.e., |(/,ip\,t)\ 2 and
-1 Im(f,il>xit)
tan Re(f,tf>x,t)
on two different plots. Figures 9a,b show these plots for
the analysis of a chirp signal (a signal whose frequency changes with t as at2
where a is some constant). The wavelet transform was obtained using the
Fourier transforms of the signal and the wavelet through implementation
of equation (17). T h e scales of analysis are plotted as the ordinate, and
the abscissa denotes t. The range of scales of the plots has been decided
using the following criteria. If At is the sampling interval of f(t) then t h e
center of passing band o;o/A m i n should be less t h a n or equal to the Nyquist
2
T h e Fourier t r a n s f o r m of (42) is s u p p o r t e d e n t i r e l y ο η ω > 0 b u t t h a t of (43) h a s a
negligible m a s s o n w < 0 for t h e c o n d i t i o n UJQ > 5.
frequency, i.e., o;o/A m i n < 2 π / 2 Δ ί , implying

UlpAt
Amin > . (45)
T h e m a x i m u m scale of analysis is obtained by considering t h e spread of

ip\j. Recognizing t h a t |^λ,ί| decays to 99.9% of its value at 3σψχ t , we
impose the condition 3σ^,λ f < (£ m a x — £ m i n )/2, i.e., the wavelet support
should b e contained within the d a t a range, giving
"Oiax *Ίηιη
(46)
T h e discretization of λ and t for implementation on discrete d a t a is dis-

cussed in the following sections.
Large values of | ( / , φχ^)\2 in the phase-space help us identify t h e scale
of the feature and its location on the t axis easily. Figure 9a clearly depicts
the decreasing scales in the signal with increasing t. In this figure we notice
t h a t large values of the squared modulus appear at relatively large scales on
the right hand side of the figure where there are no large scale features. This
is due to the apparent periodicity of the Fourier transform of the limited
extent signal. T h e phase plot helps us identify the change of phase of t h e
signal from 0 to 2π or — π to π. It is possible to count t h e number of cycles
in a signal. However, this depends upon the scale. As scale decreases, we
can see more waves and this gives rise to the bifurcation effect evident in
figure 9b. This is helpful in locating singularities and identifying fractal and
multifractal n a t u r e of processes (for example see [1]). Figure 9b shows some
edge effect at small scales due to the periodicity of the Fourier transform
of the limited extent signal. This periodicity can be eliminated by taking
the discrete Fourier transform of the chirp signal with a sufficient number
of zeros appended at the ends. For other methods see [26].
§5. D i s c r e t e W a v e l e t Transforms
W h e n the parameters λ and t in the wavelet transform (/, φχ^) take on
continuous values, it is called continuous wavelet transform. For practical
applications the scale parameter λ and location parameter t need t o b e
discretized. One can choose λ = λ™ where m is an integer and λ 0 is a
fixed dilation step greater t h a n 1. Since σψχ = \σφχ , we can choose
t = ntoX™ where to > 0 and depends upon ip(t), and n is an integer.
T h e essential idea of this discretization can be understood by an analogy
with a microscope. We choose a magnification, i.e., A^"m, and study t h e
process at a particular location and then move to another location. If t h e
magnification is large, i.e., small scale, we move in small steps and vice-
versa. This can be easily accomplished by choosing t h e incremental step
inversely proportional to the magnification (i.e., proportional to t h e scale
Modulus
Phase
Figure 9. Analysis of a chirp signal using Morlet wavelet: (a) (top)

square of the modulus and (b) (bottom) phase of (/, ip\,t).
\™) which t h e above m e t h o d of discretization of t accomplishes. We then

define
1 ,,f-nt0Ag\
Φ™,η(*) = ~κ^Φ(—3^—)
νΛο Λ
ο
= Ä0-m/V(A0-mi-ni0). (47)
T h e wavelet transform
(/, t/>m,»> = \-m/2 J f(t)ip(\-mt - nt0) dt (48)
is called t h e discrete wavelet transform.

In the case of the continuous wavelet transform we saw t h a t (f, ip\ft)
for λ > 0 and t G (—οο,οο) completely characterizes the function f(t). In
fact, one could reconstruct f(t) using (14). Using the discrete wavelet ipmin
(with φ decreasing sufficiently fast) and appropriate choices of λο and £o>
we can also completely characterize f(t). In fact, we can write f(t) as a
series expansion, as we shall see in the following subsections. We first study
orthogonal wavelets and then the general case.
5.1. Orthogonal wavelet transforms and multiresolution analysis

5.1.1. Orthogonal wavelet transforms
Consider the discrete wavelet transform for λο = 2 and t0 = 1, i.e.,
i f ne}rri
ra
</V» ( 0 = 2 - » / V ( 2 - t - n) = — φ ( - ¥ — ) . (49)
For t h e purpose of this subsection, let ipm,n{t) denote the above discretiza-
tion rather t h a n the general discretization given by equation (47). We will
also use t h e identity ôo(^) = Ψ(ί)· It is possible to construct a certain
class of wavelets ip(t) such t h a t ipm,n(t) are orthonormal, i.e.,
Φπι,η(^)Φτη',η'(ή dt = J m m / £ n n / (50)
/ ·
where 6{j is t h e Kronecker delta function given as
%J
\0 otherwise. ^ '
T h e above condition implies t h a t these wavelets are orthogonal to their

dilates and translates. One can construct ^ m ) „ ( i ) t h a t are not only or-
thonormal, b u t such t h a t they form a complete orthonormal basis for all
functions t h a t have finite energy [30]. This implies t h a t all such functions
f(t) can be approximated, up to arbitrary high precision, by a linear com-
bination of the wavelets ipmjn(t), i.e.,

oo oo
D
/(*)= Σ Σ m^m,n(t) (52)
m= — oo n = —oo
where the first summation is over scales (from small to large) and at each
scale we sum over all translates. The coefficients are obtained as
Dm,n = (ί,Ψπι,η) = I f ^)Ψm,η{1) dt
and, therefore, we can write

oo oo
/(*)= Σ Σ (/^m,n)^m,„(i). (53)

m— — oo n= — oo
From (53) it is easy to see how wavelets provide a time-scale representation
of the process where time location and scale are given by indices n and m,
respectively. T h e equality in equation (53) is in the mean square sense.
The above series expansion is akin to a Fourier series with the following
differences:
1. T h e series is double indexed with the indices indicating scale and
location;
2. T h e basis functions have the time-scale (time-frequency) localization
property in the sense discussed in section 2.2.
By using an intermediate scale mo, equation (53) can be broken up as
two sums
oo oo mo oo
/(*) = Σ Σ (ί'Ψ™,η)Ψ,η,η(*)+ Σ Σ </'</>m,n>V>m,n(i)·

m = m o + l n= — oo m— — con— — oo
(54)
It turns out t h a t one can find functions 0 m ) n ( O defined analogous to
0m,n(i) = 2 - m / 2 0 ( 2 - m i - n ) (55)
and satisfying certain properties enumerated in appendix A, such t h a t t h e
first sum on the RHS of equation (54) can be written as a linear combination
of φηΐο,η (see [30]), i.e.,
oo oo oo
(')= Σ Σ </>^m,n)tfm,n(t)· (56)

η= — oo m = m o + l n = — oo
Consequently,
oo m0 oo
/(*)= Σ </,4>m„,n><£mo,»W+ Σ Σ </>^m,n)tfm,n(t) (57)

n= — oo m=—oon= — oo
The function 0 m , n (O is called a scaling function and satisfies J φ(ΐ) dt = 1

among its other properties. For example, the scaling function correspond-
ing to the Haar wavelet is the characteristic function of the interval [0,1)
given as
(58)
The scaling functions and wavelets play a profound role in the analysis of
processes using orthogonal wavelets. This analysis framework is known as
the wavelet multiresolution analysis framework and is discussed below. Ap-
pendix B describes a class of orthogonal wavelets developed by Daubechies
[13] and Appendix C briefly discusses the implementation algorithm by
Mallat [30].
5.1.2. Multiresolution representation

Equation (56) states that all the features of the process /(£), that are
larger than the scale 2 m °, can be approximated by a linear combination of
the translates (over n) of the scaling function <f)(t) at the fixed scale 2 m °.
Let us represent this approximation by PmQj', i.e.,
(59)
Let us now define
(60)
so that equation (57) becomes
(61)
Since mo is arbitrary we also have
(62)
from which we can obtain by subtraction

(63)
or in general
(64)
This equation characterizes the basic structure of the orthogonal wavelet
decomposition (53). As mentioned before Pmf(t) contains all the informa-
tion about features in f(t) that are larger than the scale 2 m . From equation
(64) it is evident t h a t when we go from the scale 2 m to the next smaller

scale 2 m _ 1 , we add some detail to Pmf(t) which is given by Qmf(t). We
can, therefore, say t h a t Q m / ( i ) , or equivalently the wavelet expansion of
a function at any scale 2 m , characterizes the difference between the pro-
cess at two different scales 2 m and 2 m _ 1 , or equivalently at two different
resolutions.
Representation of a function within the nested structure of equation
(64) is called the wavelet multiresolution representation. Formally it con-
sists of a sequence of closed subspaces { V r m } m € 2 °f L2(H) where L2(H)
denotes the Hubert space 3 of all square integrable functions, and R and Z
denote the set of real numbers and integers, respectively. These subspaces
characterize the behavior of a function at different scales or resolutions.
For example, Vm characterizes functions at scale 2 m or equivalently at res-
olution given as 2 _ m samples per unit length. The subspaces satisfy the
following properties:
M l Vm C Vm-i for all m G Z, i.e., a space corresponding to
some resolution contains all the information about the space
at lower resolution, or equivalently, a space corresponding to
some scale contains all the information about the space at
larger scale.
M2 UÛ.ooVm is dense in L 2 ( R ) and f l ^ . ^ V ^ = {0}, i.e.,

as the resolution increases the approximation of the func-
tion converges to the original function and as the resolution
decreases the approximated function contains less and less
information.
M3 f(t) G Vm if and only if f(2t) G V m _i for all m G Z, i.e.,

all spaces are scaled versions of one space.
M4 f(t) G Vm implies f(t - ^ r ) G VmVk G Z, i.e., the space

is invariant with respect to the "integer translations" of a
function.
Notice t h a t since Vm C Vm-i we can write
Vm-i=Vm®Om (65)
where Om is the orthogonal complement of Vm in V m _i (i.e., Om is the
set of all functions in Vm-i t h a t are orthogonal to Vm) and 0 denotes
3
A Hubert space H is a vector space (possibly infinite dimensional) with an inner
product (.,.) which is complete with respect to the norm || / ||= (/, f)1/2 induced by
this inner product. A normed space is complete if every Cauchy sequence in that space
converges to an element of that space, i.e., for every sequence {/ n } C H such that
|| fm — fn |—y 0 as m, n -> oo, we have / „ - ^ / G H a s n - > o o [44].
the direct sum. Given this structure, representation of a function in Vm is

given by Pmf(t) and representation in 0m is given by Qmf(t) (compare
equation (65) with equation (64)). T h e operators Pm and Qm are orthog-
onal projection operators onto the spaces Vm and 0 m , respectively. Let
P^f and Qmf denote the discrete set of inner products { ( / , </>m,n)} and
{(ί,ψπι,η)}·, respectively. T h e set P^f gives the discrete approximation of
f(t) at scale 2 m and Q^f gives the discrete detail approximation of f(t).
Then, in simple words equation (65) says t h a t we need to add t h e infor-
mation contained in Q m / to P^f to go from one resolution (scale) to t h e
next higher resolution (smaller scale).
T h e multiresolution analysis framework is not unique. Several mul-
tiresolution frameworks can be constructed depending upon the choice of
the pair (φ,ψ). Recall t h a t the choice of either φ(ί) or ip(t) determines
the other. T h e simplest of all multiresolution frameworks is t h e one where
Vm is composed of piecewise constant functions. In this case the scaling
function is given by equation (58) and the wavelet is the Haar wavelet given
by equation (40). For examples of other pairs of (φ,φ) t h a t give rise t o
the multiresolution framework, see Appendix B and [13, 15] and [31]. For
algorithms to construct the pairs (φ,ψ), see [52, 51].
5.2. N o n - o r t h o g o n a l w a v e l e t t r a n s f o r m s
5.2.1. Frames
We saw in section 5.1 t h a t it is possible to find λο, to and ipmin(t)
as defined in equation (47) such t h a t ψγη,η{ί) are orthogonal. This allows
a function f(t) to be written as a series expansion as given in equation
(53). However, even if ^ m , n ( 0 are not orthogonal, the function f(t) can b e
represented completely as a series expansion under certain broad conditions
on the wavelet r/>(i), t0 and λο· These discrete wavelets which provide
complete representation of the function f(t) are called wavelet frames and
will be the subject of the next sub-section. We will see t h a t orthogonal
wavelets are a special case of this general framework. Let us first define
frames.
m a
A sequence of functions {ψη}η£Ζ Hilbert space H (see footnote
on page 24 for definition of Hilbert Space) is called a frame if there exist
two constants A > 0, B < oo, called frame bounds, so t h a t for all functions
f(t) in the Hilbert space H the following holds:
A||/||2<£|</,v>„>l2<B|l/ll2· (66)
n
T h e constant B < oo guarantees t h a t the transformation / —> { ( / , φ η ) }
is continuous and the constant A > 0 guarantees t h a t this transformation
is invertible and has continuous a inverse. This enables one to: (1) com-
pletely characterize the function, and (2) reconstruct t h e function from its
decomposition.
In general, a frame is not an orthonormal basis. It provides a redundant
representation of the function f(t). This is analogous, for example, t o
representing a vector in the Euclidean plane using more t h a n two basis
vectors. T h e ratio A/B is called the redundancy ratio or redundancy factor.
Redundant representations are more robust t o noise and therefore useful
when noise reduction is an issue.
W h e n A = B, the frame is called a tight frame. In this case there is a
simple expansion formula given as
Ϊ(ΐ) = \Σ(Ϊ,Ψη)ψη{ί). (67)

n
Notice t h a t this formula is very similar t o the one obtained for a n or-
thonormal set {φη}· In this case, however, {φη} niay not even be linearly
independent, i.e., there is a large degree of redundancy in the represen-
tation. Orthonormal bases arise as a special case. For a tight frame, if
A = B = 1 and if || φη \\= 1, then {φη} form an orthonormal basis and we
get t h e usual expansion formula. When {^ m , n } constitute a tight frame
then A = B = C ^ / i o l o g ^ o where Οψ is defined in equation (15) (see [15],
equation 3.3.8). However, in practice it is difficult t o get A exactly equal
to £?, b u t easier t o get A close t o B, i.e., e = - j — 1 <£ 1. Daubechies (see
[14], pg. 971) calls such frames snug frames. The expansion formula in this
case is given as
/(ί) =
ΐΓβΣ(/^>«+^ (68)
n
where t h e error 7 is of the order of 2x7 || / ||· T h e general case of A 96 B
is more involved and beyond the scope of this introduction (see [15], for
details).
5 . 2 . 2 . W a v e l e t frames
Now let L denote the transformation L : f(t) -l· {(/,ψγη,η)}, where
tpm,n(t) is defined by equation (47). We can characterize the function
f(t) through the wavelet coefficients { ( / , ipm,n)} provided the transform L
satisfies the condition (66), i.e.,
Α||/ιι 2 <ΣΣκ/>'/ν»>ΐ 2 ^ΐ/ΐΐ 2 · (69)

m n
Given discrete wavelets, we can obtain simple expansions such as in (67)

and (68), provided ^ m > n constitute a frame, i.e.,
W) = I Σ Σ</' VVn>^m,nW· (70)

m n
if {ψπι,η} is a tight frame, and

/(ί) =
ζΤβΣΣ<·Λ^>">^."+-Ύ (7i)
m n
when {ψπι,η} is a snug frame. Such frames can be constructed for certain
choices of λο and £o, provided ip(t) satisfies the admissibility condition,
i.e., J φ(ί) dt — 0, and has compact support or sufficiently fast decay. T h e
conditions for the choice of λο and to are described in Daubechies (see [15],
chapter 3). Here it suffices to say t h a t these conditions are fairly broad
and admit a very flexible range. For example, for the Mexican h a t wavelet
(as given in equation (41)), for λο = 2 and to = 1, the frame bounds are
A = 3.223 and B = 3.596 giving B/A = 1.116.
One can obtain B/A closer to 1 by choosing λο < 2. Grossmann et al.
[24] suggested decomposing each octave into several voices (as in music) by
choosing λο — 2 1 / / M where M indicates the number of voices per octave.
W i t h such a decomposition we get
ψ™η(ή = 2 - m / 2 A V ( 2 " m / M * - nto). (72)
For the Mexican h a t wavelet, by choosing M = 4 and to = 1 we can obtain
A = 13.586 and B = 13.690 giving B/A = 1.007. Such a decomposition
using such a multivoice frame enables us to cover the range of scales in
smaller steps giving a more "continuous" picture. For example, with M = 4
we get discrete scales at {λ = . . . , 1 , 2 1 / 4 , 2 1 / 2 , 2 3 / 4 , 2 , 2 5 / 4 , 2 3 / 2 , 2 7 / 4 , 4 , . . . }
as against {λ = . . . , 1 , 2 , 4 , . . . } for usual M = 1. Figure 9 was created using
Morlet wavelet with M = 4 and t0 = 1. For this decomposition A = 6.918,
B = 6.923 giving B/A = 1.0008. It should be noted t h a t Morlet wavelet,
which is not orthogonal, gives a good reconstruction under t h e framework
of equation (71). Multivoice frames are discussed extensively in Daubechies
([15], chapter 3) where more details on the values of A and B for different
choices of M and to are given for the mexican hat and the Morlet wavelet.
R e d u n d a n t representations such as the one presented above, in addition
to their noise reduction capability, are useful when representations t h a t are
close to the continuous case are sought (see for example [3, 32, 35, 5, 33]
and [49]).
5.3. B i o r t h o g o n a l w a v e l e t s
Under the wavelet multiresolution framework, the decomposition and
reconstruction of a function is done using the same wavelet, i.e.,
ί(*) = ΣΣ(Μ™.»)ψ™.»Μ (73)

m n
where { ( / , ^ m , n ) } are the decomposition coefficients. This however, can
severely limit the choice of wavelet φ(ί). For example, it has been shown
28 P. Kumar and E. Foufoula-Georgiern
(see [15], theorem 8.1.4) t h a t the only real and compactly supported sym-
metric or antisymmetric wavelet under a multiresolution framework is t h e
Haar wavelet. In certain applications however, real symmetric wavelets
which are smoother and have better frequency localization t h a n t h e Haar
wavelet may b e needed. In such situations, biorthogonal wavelets come
to t h e rescue. It is possible t o construct two sets of wavelets {ipm^n} and
{Ψτη,η} such t h a t
/(') = Σ Σ ^ - . " ) < / ν η ( ί ) (74)

m n
= EE</^™.»>^.«w· (75)
m n
T h a t is, one can accomplish decomposition using one set of wavelets and
reconstruction using another. T h e wavelets ißmyn(t) = ^ 7 7 ^ ( 2 ^ — n ) a n ( ^
ψηι,η^) = 2^72^(2^ ~ n ) n e e d t 0 Satis
fy
EEK/></Vn>|2 < B\\f\\* (76)

m n
ΣΣκ/'Vw*)! 2 ^ è\\f\\2 (77)

m n
{Ψπι,η,Ψτη',η') = ^mm'^nn' (78)
where B a n d B are some constants and condition (78) is t h e condition of

biorthonormality. Given such a biorthonormal set, it is possible t o con-
struct corresponding scaling functions {φηι,η} a n d {<j>m,n} such t h a t
\ψπι,η·)ψπι,η'/ — ^ηη'· v'*V
Notice t h a t nothing is said about the orthogonality of {ψγη,η}, {Ψπι,η},

{Φπι,η} a n d {Φπι,n) themselves. In general they form a linearly indepen-
dent basis. Also, there is no condition of orthogonality between t h e wavelets
ψ(ί) and V>(£), a n d t h e corresponding scaling functions φ(ί) and 0(£), re-
spectively. Given these wavelets and scaling functions, one can construct
a multiresolution nest, as in t h e orthonormal case, i.e.,
• · · c v2 c Vi c v0 c y_i c K-2 c · · ·
• · · C V2 C Vi C Vo C VLi C V-2 C ■ ■ ■
with Vm — s p a n { 0 m j „ } and Vm = s p a n { 0 m > n } and t h e complementary
spaces Om = s p a n { ^ m > n } and O m = s p a n { ^ m > n } . T h e spaces Vm a n d
Om (Vm and O m , respectively) are not orthogonal complements in general.
Equation (78), however, implies t h a t
Vm JL Öm and Vm J- O m . (80)
Another advantage of biorthogonal wavelets is (see [15], section 8.3)

t h a t one can have ip(t) and zp(t) with different vanishing moments. For
example, if ip(t) has more vanishing moments t h a n φ{ϊ), one can obtain
higher d a t a compression using ( / , V>m,n) and a good reconstruction using
the sum being restricted to some finite values.
§6. T w o - D i m e n s i o n a l W a v e l e t s
6.1. Continuous wavelets
T h e continuous analogue of wavelet transform (12) is obtained by treat-
ing u — (u 1,1*2) and t = (£1,^2) as vectors. Therefore for the two dimen-
sional case
λ>0
(81)
An analogous inversion formula also holds, i.e.,
(82)
T h e condition of admissibility of a wavelet remains t h e same, i.e.,
1. compact support or sufficiently fast decay; and
2. fft{t)dt = 0.
Two examples of two-dimensional wavelets are discussed in the follow-
ing subsection.
6.1.1. Two-dimensional Morlet wavelet

Define the vector t = (£1,^2) on the two-dimensional plane with \t\ =
\Jt\ -+-1\. Then the two dimensional Morlet wavelet is defined as
(83)
with Fourier transform
(84)
where Ω = {ω\, ω<χ ) is an arbitrary point on the two-dimensional frequency

plane, and Ω° = (ω^,ωί,) is a constant. T h e superscript Θ indicates the
Figure 10. Frequency support of two dimensional Morlet wavelet.
direction of the wavelet, i.e,
1
^ t a n - ^ . (85)
T h e properties of this wavelet are best understood from its spectrum.

Figure 10 shows the spectrum of the two-dimensional Morlet wavelet for
0 = 0 and λ = 1. This wavelet is no longer progressive as in the one-
dimensional case, i.e., its spectrum is not entirely supported on the positive
quadrant. Manipulating Ω 0 by changing Θ allows us to change the direc-
tional selectivity of the wavelet. For example, by choosing Ω° = (ω?, ω^) =
u;°(cos0,sin0), ω° > 5, 0 < θ < 2π we get the wavelet transform
T h e last equation is obtained by using Parseval's theorem. At any arbitrary

scale λ, equation (86) can be written as
\ /»OO /«OO 0 0
V^ J — OO J — OO
T h e above equation indicates t h a t the wavelet transform {ί,φθχ *) extracts

the frequency contents of the function f(t) around the frequency coordi-
nates (ωι/Χ,ω^/Χ) = (ω° cos0/A,u; 0 sin0/A) with a radial uncertainty of
σ? = Ι / λ , at the location t. Therefore, by fixing λ and traversing along
0, directional information at a fixed scale λ can be extracted, and by fix-
ing Θ and traversing along λ, scale information in a fixed direction can b e
obtained.
6.1.2. Halo wavelet

Often the directional selectivity offered by Morlet wavelet is not desired
and one wishes to pick frequencies with no preferential direction. Dallard
and Spedding [12] defined a wavelet by modifying the Morlet wavelet and
called it the Halo wavelet because of its shape in the Fourier space. T h e
wavelet itself is defined through its Fourier transform
ψ(Ω) = κβ-(Ι Ω Η Ω °Ι> 2 /2 (88)
where n is a normalizing constant. As can be seen from the above expression

this wavelet has no directional specificity.
6.2. Orthogonal wavelets

For two-dimensional multiresolution representation, consider the func-
tion / ( i i , £ 2 ) £ I / 2 ( R 2 ) . A multiresolution approximation of L 2 ( R 2 ) is a
sequence of subspaces t h a t satisfy the two-dimensional extension of prop-
erties M l through M4 enumerated in the definition of the one-dimensional
multiresolution approximation. We denote such a sequence of subspaces of
L 2 ( R 2 ) by ( ^ m ) m G z · T k e a P P r o x i m a - t i o n of the function / ( i i , Î 2 ) at t h e
resolution m, i.e., 2 2 m samples per unit area, is the orthogonal projection
on t h e vector space Vm.
A two-dimensional multiresolution approximation is called separable
if each vector space Vm can be decomposed as a tensor product of two
identical subspaces V^ of L 2 ( R ) , i.e., the representation is computed by
filtering t h e signal with a low pass filter of the form Φ(ίι, t2) = φ(ίι)φ^2).
For a separable multiresolution approximation of L 2 ( R 2 ) ,
Vm = t£®V£ (89)
where 0 represents a tensor product. It, therefore, follows (by expanding
y m + i as in (89) and using property M l ) t h a t the orthogonal complement
0m of Vm in Vm+\ consists of the direct sum of three subspaces, i.e.,
Om = {VL®Oln)®{Ol®V^)®{0]n®0]n). (90)
The orthonormal basis for Vm is given by
(2m*(2mti-n,2mt2-fc))(n fc)eZ2 = (2m^m(2mi1 -n)^»m(2mi2-fc))(n k ) ^ .

91
' < >
Analogous to the one-dimensional case, the detail function at the resolution
m is equal to the orthogonal projection of the function on to the space Om
which is the orthogonal complement of Vm in V^+i. An orthonormal basis
for Om can be built based on Theorem 4 in Mallat (1989a, pg. 683) who
shows t h a t if 4>{t\) ls the one dimensional wavelet associated with the
scaling function φ(ίχ), then the three "wavelets" Φ 1 ^ , ^ ) — Φ(ίι)Ψ(ί2)·>
Φ 2 ( ί ι , ί 2 ) = φ{Η)φ{ΐ2) and Φ 3 ( ί ι , ί 2 ) = Φ{Η)φ{ί2) are such t h a t
i V ^ m n / î ) ^ Trink") ^ m n k ) ( n k)£%2 J
is an orthonormal basis for Om.

T h e discrete approximation of the function f(ti,t2) at a resolution m
is obtained through the inner products
Pif = {(f,*rnnk)(nk)€Z*} = {(/. 0π,η^»*)(Μ)6Ζ» } (92)
T h e discrete detail approximation of the function is obtained by the inner

product of / ( i i , £ 2 ) with each of the vectors of the orthonormal basis of
Om. This is, thus, given by
Ä 7 = {(/,*L,*)(n,fc)<Ez2}> (93)
2
Ä / = {(/,*L,*)(„,fc)€z*K (94)
and
Q £ / = {(/,*3m„*)(n,fc)€Z'}· (95)
T h e corresponding continuous approximation will be denoted by Qmf(t),
Qmf(t) and Qmf(t) respectively. For implementation to discrete d a t a see
[31]·
T h e decomposition of Om into the sum of three subspaces (see equa-
tion (90)) acts like spatially oriented frequency channels. Assume t h a t we
have a discrete process at some resolution m + 1 whose frequency domain
is shown in Figure 11 as the domain of Pm+if. W h e n t h e same process
Frequency support of Q mf
Λ
Λ
3 1 Λ
3
Ψ (ω)
Frequency support of Q ^2f
Λ Λ
2 2
Ψζ(ω)
Frequency support of P f
Λ Λ
3 3
Ψ3(ω) Ψ (ω)
Frequency support of P m f
.Frequency support of Q mf
ω = ( œt , cot )
1 2
Figure 1 1 . Frequency support of wavelets in two dimensional mul-

tiresolution decomposition.
is reduced to resolution ra, its frequency domain shrinks to t h a t of P^f.

T h e information lost can be divided into three components as shown in
Figure 11: vertical high frequencies (high horizontal correlation), horizon-
tal high frequencies (high vertical correlation) and high frequencies in b o t h
direction (high vertical and horizontal correlations, for example, features
like corners). They are captured as Q ^ / , Q^f a n d Qmf respectively.
This property has been used in [29] to characterize the directional behav-
ior of rainfall. Wavelets with more t h a n three frequency channels can b e
constructed (see, for example, [10]) but are not discussed herein.
§7. C o n c l u s i o n s
Wavelet theory involves representing general functions in terms of sim-
pler building blocks at different scales and positions. Wavelets offer a ver-
satile and sophisticated tool yet simple to implement, and have already
found several applications in a wide range of scientific fields. T h e list of
Table I
Anonymous ftp site information for wavelet related software.
1 FTP site Directory References/Notes f
cs.nyu.edu pub/wave [35, 33], C programs 1
playfair.stanford.edu pub/software/wavelets MATLAB Scripts
simplicity.stanford.edu /pub/taswell MATLAB scripts
gdr.bath.ac.uk /pub/masgpn S software
pascal.math.yale.edu /pub/software/xwpl Wavelet Packet
Laboratory
[email protected] e-mail request for KHOROS
cml.rice.edu /pub/dsp/software MATLAB scripts
1 wuarchive.wustl.edu edu/math/msdos/modelling C programs
wavelet applications increases at a fast pace and includes to date signal pro-
cessing, coding, fractals, statistics, image processing, astrophysics, physics,
turbulence, mathematics, numerical analysis, economics, medical research,
target detection, industrial applications, quantum mechanics, geophysics
etc. (see, for example, the edited volumes by Chui [7], Ruskai et al. [48],
Benedetto and Frazier [3], Farge et al. [20], Meyer and Roques [41], Combes
et al. [9], and Beylkin et al. [2], among others). A general literature survey
on wavelets can be found in [46]. In geophysics, significant progress has
already been m a d e in studying and unraveling structure of several geophys-
ical processes using wavelets (see the extended bibliography of wavelets in
geophysics at t h e end of this volume). We have no doubt t h a t the study of
geophysical phenomena, which are by nature complex, and take place and
interact at a range of scales of interest, will continue to benefit from t h e
use of the powerful and versatile tools t h a t wavelet analysis has to offer.
There are a number of useful sources of information about wavelet pub-
lications and computer software for the implementation of wavelet analysis.
An electronic information service called "wavelet digest" exists (at the time
of this writing) on the Internet with the address waveletumath.scarolina.edu.
Several anonymous ftp sites exist on the Internet from where software for
wavelet analysis can be obtained. In Table I we provide a brief list (known
to t h e authors at the time this was written) solely for the purpose of infor-
mation to readers, without any recommendations, or reference to suitability
or correctness of these codes.
A . P r o p e r t i e s of Scaling F u n c t i o n
T h e scaling function satisfies the following properties:
1. j 4>(t)dt = 1, i.e., the scaling function is an averaging function-,

compare this with the wavelet t h a t satisfies / ψ(ί) dt = 0.
2. || φ(ΐ) ||= 1, i.e., the scaling function is normalized to have unit

norm.
3. f φτη,η(ί)ψπι',η'(t) dt = 0, i.e., the scaling function is orthogonal to

all the wavelets.
4. j ' Φτη,η(ί)Φπι,η'(ί) dt = Snn>, i.e., the scaling function is orthogonal

to all its translates at any fixed scale. Note t h a t unlike t h e wavelets,
t h e scaling function is not orthogonal to its dilates. In fact,
5. (j)(t) = Ση ϊι(η)φ(2ί — η), i.e., the scaling function at some scale can
b e obtained as a linear combination of itself at the next scale (h(n)
are some coefficients called the scaling coefficients). This is a two-
scale difference equation (see [16] and [17] for a detailed t r e a t m e n t
of such equations).
6. T h e scaling function and wavelet are related to each other. In fact,

one can show t h a t
φ{ΐ) = Υ/9{η)φ{2ί-η) (A.l)

n
where g(n) are coefficients derived from h(n). T h a t is, the wavelets
can be obtained as a linear combination of dilates and translates of
t h e scaling function.
For the particular case of Haar wavelet (see equation (40)) and t h e corre-
sponding scaling function (equation (58)) h(0) = h(l) = 1 and h{n) = 0
for all other n, and g(l) = —g(0) = 1 and g(n) = 0 for all other n.
B. Daubechies' Wavelets
Daubechies [13] developed a class of compactly supported scaling func-
tions and wavelets denoted as ( ^ , ρ/ψ). They were obtained through t h e
solution of the following two-scale difference equations:
2N-1
φ(ή = y/2 Σ Κη)Φ{2ί - n) (Β·ΐ)
2ΛΓ-1
tl>(t) = j2 Σ 9{η)φ{2ΐ-η) (Β.2)

n=0
where
g(n) = (-l)nh(2N - n + 1) for n = 0 , 1 , . . . , 2N - 1. (B.3)
For techniques to solve the above equations see [52]. T h e scaling coefficients
h(n) are obtained from solutions of high order polynomials ([15], chapter
6, and [51]) and satisfy the following constraints:
1. A necessary and sufficient condition for the existence of a solution

to the above two-scale difference equations is
2ΛΓ-1
Σ M») = \/2· (Β.4)

n=0
2. Integer translations and dilations of </>(£) and φ(ί) form an orthog-

onal family if the scaling coefficients satisfy
2JV-1
h n
J2 ( - 2k)h(n - 21) = Skl for all k and I. (B.5)
n=0
3. T h e constraints
27V-1
Σ {-l)n~lnkh{n) = 0 for k = 0 , 1 , . . . , N - 1 (B.6)
yield the result t h a t φ(ϊ) has N vanishing moments, i.e.,
/■tktp(t) at = 0 for k = 0 , 1 , . . . , N - 1. (B.7)
Daubechies' wavelets of this class have the following properties:
1. They are compactly supported with support length 2N — 1. T h e

scaling function also has the same support length.
2. As iV increases, the regularity of χφ and ΝΨ also increases. In fact

ΝΦ,ΝΨ £ CaN (the- set of continuous functions t h a t are a^ order
differential) where (iV, α^γ) pairs for some iV are given as (see [13])
{(2,0.5 - e), (3,0.915), (4,1.275), (5,1.596), (6,1.888), (7,2.158)}.
3. Condition (B.7) of vanishing moments of φ(ΐ) implies t h a t up to

Nth order derivatives of the Fourier transform of φ(ϊ) at the origin
are zero, i.e.,
-^φ^ω = 0) = 0 for k = 0 , 1 , . . . , iV - 1. (B.8)
This property essentially implies a form of localization of the Fourier

transform φ(ω).
Figures B l and B2 show the scale function and the corresponding

wavelets for N = 2 and TV = 5 respectively. T h e figures also depict the
magnitude of their Fourier transforms. As can be seen the regularity of t h e
wavelets and the scale functions increase as N increases.
10 20 30 40
Frequency (Radians)
10 20 30 40
Frequency (Radians)
F i g u r e B l . The figure shows Daubechies' scaling function (upper

left) and the wavelet (lower left) of order N = 2. The magnitude of
their Fourier transforms is shown, respectively, on the right hand column
plots.
C . I m p l e m e n t a t i o n A l g o r i t h m for O r t h o g o n a l W a v e l e t s
T h e implementation algorithm for wavelet multiresolution transforms
is simple. From a d a t a sequence {c^} (say at resolution level m = 0)
corresponding to a function f(t) we construct
ρ„/(*) = Σ <£#'-") (C.I)
for a chosen <£(£), i.e., assume { c ° } n € £ = { ( / , 0 O n ) } n G z · T h e d a t a se-

quence at lower resolution can be obtained by
ci = 5>(»-2*)<$. (C.2)
10 20 30 40
Frequency (Radians)
10 20 30 40
Frequency (Radians)
Figure B2. The figure shows Daubechies' scaling function (upper

left) and the wavelet (lower left) of order N = 5. The magnitude of
their Fourier transforms is shown, respectively, on the right hand column
plots.
= 1S
T h e detail sequence { ^ } f c G z {(/'î^)}fceZ obtained as
d\ = Y,g{n-2k)cl (C.3)
Equivalently the above two equations can be written in the matrix notation
{c1} = H{c0} and { d " 1 } = G{c 0 }. (C.4)
T h e matrices H and G are such t h a t
HH* = / , GG* = I (C.5)
and H*H and G*G are mutually orthogonal projections with
H*H + G*G = I (C.6)

where H* and G* are adjoints of H and G respectively and i" is t h e identity

matrix. Also,
GH* = 0 a n d HG* = 0. (C.7)
The algorithm given in equation (C.2) and (C.3) can be recursively im-
plemented and the d a t a and details at lower and lower resolutions can be
obtained. This also highlights another important feature. Once we have
the coefficients h{n), we never need to construct the scale function 0(f) or
the wavelet φ(ί) for implementation on a discrete d a t a set. T h e assump-
tion involved, however, is t h a t c^ = f f(t)(f)(t — n)dt, i.e., the integration
kernel is φ(ί) corresponding to the chosen /i(n)'s. T h e reconstruction of
the original sequence can be achieved by using
c™-1=2Y^h(n-2k)cÎ + 2j2g(n-2k)dZ (C.8)

k k
and is exact.
References
1. Arneodo, A., G. Grasseau and M. Holschneider, Wavelet transform of
multifractals, Physical Rev. Let, 61(20), 2281-2284, 1988.
2. Beylkin, G., R. R. Coifman, I. Daubechies, S. G. Mallat, Y. Meyer,
L. A. Raphael, M. B. Ruskai (eds.), Wavelets and Their Applications,
Jones and Bartlett Publishers, Boston, 1991.
3. Benedetto J. J., and M. W. Frazier (eds.), Wavelets: Mathematics and
Applications, CRC Press, Boca Raton, Florida, 1993.
4. Brunet, Y. and S. Collineau, Wavelet analysis of diurnal and nocturnal
turbulence above a maize crop, This Volume, 1994.
5. Burt, P . J., Fast filter transforms for image processing, Computer
Graphics and Image Processing, Vol. 16, 20-51, 1981.
6. Chui, C. K., An Introduction to Wavelets, Wavelet Analysis and its
Applications, Vol. 1, Academic Press, New York, 1992.
7. Chui, C. K., Wavelets - a Tutorial in Theory and Applications, Wavelet
Analysis and its Applications, Vol. 2, Academic Press, New York, 1992.
8. Coifman, R. R., Y. Meyer and V. Wickerhauser, Size properties of
wavelet-packets, in Ruskai et al., 453-470, 1992.
9. Combes, J . M., A. Grossman, P h . Tchamitchian (eds.), Wavelets:
Time-Frequency Methods and Phase-Space, Proc. of the Int. Conf.,
Marseille, Frace, Decemberl4-18, 1987, Springer-Verlag, 1989.
10. Cohen, A., Non-separable bidimensional wavelet bases, Revista
Mathematica Iberoamericana, 9(1), 51-137, 1993.
11. Cohen, L., Time-frequency distributions-a review, Proc. of the IEEE,
77(7), 941-981, 1989.
12. Daliard, T. and G. R. Spedding, 2-D wavelet transforms, Preprint,

1990.
13. Daubechies, L, Orthonormal bases of compactly supported wavelets,
Commun, on Pure and Appl. Math., Vol. XLI, 901-996, 1988.
14. Daubechies, I., T h e wavelet transform, time-frequency localization and
signal analysis, IEEE Trans. Info. Theory, 36(5), 1990.
15. Daubechies, I., Ten Lectures on Wavelets, SIAM, 1992.
16. Daubechies, I. and J. C. Lagarias, Two-scale difference equations I. ex-
istence and global regularity of solutions, SIAM J. Math. Anal., 22(5),
1388-1410, 1991.
17. Daubechies, I. and J. C. Lagarias, Two-scale difference equations II.
local regularity, infinite products of matrices and fractals, SIAM J.
Math. Anal., 23(4), 1031-1079, 1992.
18. Daubechies, I., Review of the Books: Wavelets and Operators, Y.
Meyer, (Cambridge University Press, New York, 1993, Cambridge
Studies in Advanced Mathematics, 37, Translation from the french
edition, Paris, 1990, by D.H. Salinger) and Wavelets: Algorithms and
Applications, Y. Meyer, (SIAM, Philadelphia, 1993, Translation form
the french edition, by R. D. Ryan), Science, 262, 1589-1591, 1993.
19. Farge, M., Wavelet transforms and their applications to turbulence,
Annu. Rev. Fluid Mech., 24, 395-457, 1992.
20. Farge, M., J. R. Hunt, J. C. Vassilicos (eds.), Wavelets, Fractals and
Fourier Transforms: New Developments and New Applications, Oxford
University Press, Oxford, 1991.
21. Flandrin, P., Time-frequency and time-scale, IEEE Fourth Annual
ASSP Workshop on Spectrum Estimation and Modeling, Minneapolis,
Minnesota, 77-80, 1988.
22. Gabor, D., Theory of communications, J. Inst. Elec. Eng., Vol. 93,
429-457, 1946.
23. Goupillaud, P., A. Grossmann and J. Morlet, Cycle-octave and related
transforms in seismic signal analysis, Geoexploration, 23, 85-102, 1984.
24. Grossmann, A. and J. Morlet, Decomposition of Hardy functions into
square integrable wavelets of constant shape, SIAM J. math. Anal.
15(4), 723-736, 1984.
25. Hudgins L. H., M. E. Mayer and C. A. Friehe, Fourier and wavelet
analysis of atmospheric turbulence, in Progress in Wavelet Analysis
and Applications, Y. Meyer and S. Roques, eds., 491-498, Editions
Frontiers, 1993.
26. Jawerth, B, and W. Sweldens, An overview of wavelet based multires-
olution analyses, to appear, SIAM review, 1994.
27. Katul, G. G., J. D. Albertson, C. R. Chu, and M. B. Parlange, Inter-
mittency in atmospheric turbulence using orthonormal wavelet, This
Volume, 1994.
28. K u m a r , P. and E. Foufoula-Georgiou, A multicomponent decomposi-

tion of spatial rainfall fields: 1. segregation of large and small-scale
features using wavelet transforms, Water Resour. Res., 29(8), 2515-
2532, 1993.
29. Kumar, P. and E. Foufoula-Georgiou, A multicomponent decomposi-
tion of spatial rainfall fields: 2. self-similarity in fluctuations, Water
Resour. Res., 29(8), 2533-2544, 1993.
30. Mallat, S., A theory for multiresolution signal decomposition: the
wavelet representation, IEEE Tran, on Pattern Anal, and Mach. Intel.,
11(7), 674-693, 1989.
31. Mallat, S, Multifrequency channel decomposition of images and wavelet
models, IEEE Trans, on Acoustics, Speech and Signal Anal. 37(12),
2091-2110. Dec. 1989.
32. Mallat, S., Zero-crossings of a wavelet transform, IEEE Trans. Inform.
Theory, 37(4), 1019-1033, 1991.
33. Mallat, S. and L. Hwang, Singularity detection and processing with
wavelets, IEEE Trans. Inform. Theory, 38(2), 617-643, 1992.
34. Mallat, S. and S. Zhong, Wavelet transform maxima and multiscale
edges, in Wavelets and Their Applications, M. B. Ruskai, G. Beylkin,
R. Coifman, I. Daubechies, S. Mallat, Y. Meyer, L. Raphael, eds., Jones
and Bartlett, Boston, 1992.
35. Mallat, S. and S. Zhong, Complete signal representation from multi-
scale edges, IEEE Trans. Pattern Anal. Machine Intell, 14(7),710-732,
1992.
36. Meneveau, C., Analysis of turbulence in the orthonormal wavelet rep-
resentation, J. Fluid Mech., 332, 469-520, 1991.
37. Meyer, Y., Ondolettes et applications, J. Annu. Soc. Math., Soc. Fran-
çaise Math., Paris, France, pp. 1-15., 1990.
38. Meyer, Y., Wavelets and Operators, (Translation from the French Edi-
tion, Paris, 1990, by D. H. Salinger) Cambridge University Press, New
York, 1993, Cambridge Studies in Advanced Mathematics, 37.
39. Meyer, Y., Wavelets: Algorithms and Applications, SI AM, Philadel-
phia, 1993.
40. Meyer, Y., Review of Books An introduction to wavelets by C. K. Chui,
and Ten Lectures on Wavelets by I. Daubechies, Bulletin (New Series)
of the American Mathematical Society, 28(2), 350-360, 1993.
41. Meyer, Y. and S. Roques (eds.), Progress in Wavelet Analysis and
Applications, Editions Frontieres, France, 1993.
42. Morlet, J., G. Arens, E. Fourgeau, and D. Giard, Wave propogation and
sampling theory- p a r t 1: complex signal and scattering in multilayered
media, Geophysics, 47(2), 203-221, 1982.
43. Morlet, J., G^ Arens, E. Fourgeau, and D. Giard, Wave propogation
and sampling theory- part 2: sampling theory and complex waves,
Geophysics, 47(2), 222-236, 1982.

44. Naylor, A. W. and G. R. Sell, Linear Operator Theory in Engineering
and Science, Springer-Verlag, New York, 1982.
45. Pike, C. J., Analysis of high resolution marine seismic data using the
wavelet transform, This Volume, 1994.
46. Pittner, S., J. Schneid, and C. W. Ueberhuber, Wavelet Literature
Survey, Institute for Applied and Numerical Mathematics, Technical
University Vienna, Wien, Austria, 1993.
47. Rioul, 0 . and M. Vetterli, Wavelets and signal processing, IEEE Signal
Processing Magazine, 14-38, Oct. 1991.
48. Ruskai, M. B., G. Beylkin, R. Coifman, I. Daubechies, S. Mallat, Y.
Meyers, and L. Raphael (eds.), Wavelets and Their Applications, Jones
and Bartlett Publishers, 1992.
49. Saito, N. and G. Beylkin, Multiresolution representations using the
auto-correlation functions of compactly supported wavelets, IEEE
Trans, on Signal Processing, 41(12), 3584-3590, 1993.
50. Saito, N., Simultaneous noise suppression and signal compression using
a library of orthonormal bases and the minimum description length
criterion, This volume, 1994.
51. Strichartz, R. S., How to make wavelets, The Amer. Mathematical
Monthly, 100(6), 539-556, 1993.
52. Strang, G., Wavelets and dilation equations: a brief introduction,
SIAM Rev., 31(4), 614-627, 1989.
53. Wickerhauser, V., Lectures on wavelet packets algorithms, Preprint,
Dept. of Math., Washington Univ., 1991.
The first author would like to thank Mike Jasinski and the Hydrologie
Sciences Branch at NASA-Goddard Space Flight Center for the support
during the course of this work. The second author would like to acknowl-
edge the support of National Science Foundation grants BSC-8957469 and
EAR-9117866, and NASA grant NAG 5-2108. We would also like to ac-
knowledge the thoughtful comments of Don Per ci val, Naoki Saito and Paul
Liu on this chapter.
Praveen Kumar
Universities Space Research Association/
Hydrological Sciences Branch (Code 974)
NASA-Goddard Space Flight Center
Greenbelt, MD 20771
[email protected]
Efi Foufoula-Georgiou
St. Anthony Falls Hydraulic Laboratory
Department of Civil and Mineral Engineering
University of Minnesota
Minneapolis, Minnesota 55414
e-mail: efiUmykonos.safhl.umn.edu
Applications of Structure Preserving Wavelet Decompositions
to Intermittent Turbulence: A Case Study
Carl R. Hagelberg and Nimal K. K. Gamage
Abstract. Coherent structures in velocity and temperature in the atmospheric

boundary layer account for a large portion of flux transport of scalars such as
momentum, heat, trace chemicals and particulates. Generally, coherent structures
are bounded by zones of concentrated shear in velocity or temperature fields.
We develop techniques based on the wavelet transform to provide a signal
decomposition which preserves coherent structures. The decomposition is used
to separate the signal into two components, one of which contains the important
structures as defined by the characteristics of the transform. Embedded within
the technique are a coherent structure detection mechanism, an analysis of in-
termittency resulting in an intermittency index, and filtering techniques. We
illustrate the dependence of the coherent structure detection mechanism on the
choice of analyzing wavelet demonstrating that anti-symmetric wavelets are ap-
propriate for detecting zones of concentrated shear, while symmetric wavelets are
appropriate for detection of zones of concentrated curvature.
We apply these techniques to the vertical velocity component, virtual po-
tential temperature, and buoyancy flux density records from two portions of data
gathered during the First International Satellite Land Surface Climatology Project
(ISLSCP) Field Experiment 1987 (FIFE 87). We analyze the buoyancy flux cor-
responding to velocity and temperature structures in a convective, unstable at-
mospheric boundary layer, and the same boundary layer in a state of declining
convection in late afternoon.
Spectral estimates of each of the partitions for the velocity signals are com-
pared to the non-partitioned velocity signals. The characteristics of the partition
reveal that the structure containing components of the velocity records follow a
classical spectral description having a —5/3 power law behavior in the inertial
range, while the non-structure components follow a —1 power law. We suggest
that classification of turbulence as weak or strong can be based on the intermit-
tency index and the scale distribution of coherent structures.
ISBN 0-12-262850-0
46 C. Hagelberg and N. Gamage
§1. Introduction
Various physical processes in the atmospheric boundary layer such as
the vertical transport of momentum and heat are associated with intermit-
tent, coherent events, such as convective updrafts and plumes occurring
due to localized heating and the resulting convection [18, 20]. Character-
izing the physics of these events through observations and their behavior
under varying conditions is still a topic of considerable continued study
[1, 3]. There is a need to verify existing models (conceptual, mathemati-
cal or numerical) with observational data, and the need to provide models
with parameter estimates for realistic simulation and continued insight into
the physics to further develop conceptual models. Additionally, there is a
need to determine from observation a set of parameters which best de-
scribe the processes associated with intermittent events. Identification of
such parameters will suggest new approaches to modeling.
Conditional sampling techniques have been extensively reported in
the literature as a means of extracting coherent events from data records
[1, 3, 17, 20, 24]. Typically, conditional sampling techniques extract co-
herent events from a record to form an ensemble. The ensemble is then
analyzed to characterize the events. This process loses important inter-
mittency information since the global placement of the events is neglected.
Various approaches to the sampling method have been used including "di-
rect" thresholding on the values of scalars [18], thresholding on computed
variance over short record patches [24, 25], visual identification of time do-
main characteristics [3], and thresholding time domain gradients [17]. In
the latter case, an indicator record is constructed applying a threshold on,
for example, the absolute value of the time derivative of velocity. This al-
lows the definition of intermittency as a function of threshold which is 0 if
no events are selected and 1 if the whole record is selected. This technique
does not attempt to address the issue of coherence. More recently, con-
ditional sampling methods have been developed based on variants of the
wavelet transform [4, 9, 20]. The techniques proposed and demonstrated
here, based on the wavelet transform, provide a general and objective means
for achieving the desirable characteristics of conditional sampling methods
and also provide a measure of intermittency which is directly related to
coherent events.
The wavelet transform has evolved in various disciplines which require
analysis of signals exhibiting intermittency in various forms. We attempt
to illustrate some properties of the wavelet transform and propose certain
new ideas useful to the study of atmospheric turbulence and the planetary
boundary layer.
Sharp edges contain significant energy at high wave number. A signal
consisting of intermittent sharp edges can have varying spectral charac-
Intermittent Turbulence 47
teristics depending on the spacing of the events. This is significant since

the physical processes of transport of m o m e n t u m and heat often occur in
concentrated events (in space or time) in the atmospheric boundary layer.
Consequently, a spectral description of such processes is not adequate to
charcterize t h e intermittency of events.
In this paper we partition signals into two components. One contain-
ing coherent structures characterized by sharp transitions and intermittent
occurrence, the other component containing the remaining portion of t h e
signal essentially characterized by smaller length or time scales and t h e
absence of coherent events. We envision an intermittency index, a struc-
ture type, and perhaps a small-scale spectral characterization as typical
parameters useful for characterizing the atmospheric boundary layer. T h e
structure type can be based on the analyzing wavelet. T h e intermittency
index is defined in Section 3 from the local extremes of the wavelet trans-
form. T h e spectrum may be computed directly from the wavelet transform
phase plane and is discussed in Section 2. We note t h a t the partitioning
we develop is different from a Fourier based low-pass/high-pass partition-
ing since a typical structure defined by a large gradient will have frequency
components in b o t h bands.
It is possible to show t h a t any symmetric wavelet is the second deriva-
tive of a smoothing function, whereas an anti-symmetric wavelet is the first
derivative of a smoothing function [4, 22]. T h e anti-symmetric wavelets are
optimally suited for shear-zone or micro-front detection [10]. These a n t i -
symmetric events contribute to the inertial range, dissipation range, and
to t r a n s p o r t processes. In contrast, the symmetric wavelets are optimal for
finding maximum curvature. Hence, if a signal contains a sharp transition,
the anti-symmetric wavelet will be highly coherent at the center of t h e
transition while the symmetric wavelet will be coherent at the edges of t h e
transition effectively defining the boundaries of the transition zone. Equiv-
alently, a zero-crossing of the symmetric wavelet transform will coincide
with t h e center of the transition. However, distinguishing between large
gradients and small gradients (over a given length scale) is more difficult
using a symmetric wavelet [22]. It requires an estimation of t h e angle of
the zero crossing in the transform [4], while the magnitude of t h e local
extremes of the anti-symmetric transform is a direct (relative) measure
of t h e sharpness [10]. Some of these properties are demonstrated using a
simulated d a t a set in Section 3.
Combining t h e optimal properties of anti-symmetry in space, and con-
centrated support in frequency leads to the natural choice of using t h e spline
wavelets for analyses where strong shear zones characterize t h e i m p o r t a n t
structures.
In Section 4 we apply the techniques presented to d a t a sampled during
F I F E 87 (described in Section 4). T h e vertical velocity, virtual potential
temperature, and (derived) buoyancy flux density fields obtained from the
atmospheric boundary layer are studied for two cases corresponding to late
morning and late afternoon on the same day. We compute the intermit-
tency index for each of the records. We then construct partition of the
signals and compute total buoyancy flux for various combinations of the
components. Additionally, spectra for the velocities and their decomposi-
tions are studied. A summary is provided in Section 5.
§2. The Wavelet Transform

2.1. The continuous case
We follow [22], where the wavelet transform is considered to be a type
of multi-channel tuned filter. This allows the implementation of a fast
algorithm and facilitates the interpretation of the wavelet transform as a
coherent structure detection mechanism. We review the necessary informa-
tion for the decomposition and reconstruction algorithms, though further
discussion may be found in [22] and [23]. The books [5] and [7] give com-
prehensive treatment of wavelet transforms and Multi-Resolution Analysis.
We limit our discussion to the properties most pertinent to our results.
To establish the notation, we begin with two functions, / and #, in
L 2 (R) and define their inner-product by
(/,<?):= Jf(t)g(t)dt. (1)

R
2
Hence, the L (R) norm of / may be written
ll/l|2 = ( / , / ) 1 / 2 .
Similarly, the convolution of / and g is
f*g(b):= J f{t)g{b-t)dt. (2)

R
The Fourier Transform of f(t) is
/(*)Λ =/(ω):= Je-^f(t)dt, (3)

R
and the inverse Fourier transform is denoted f(t) = /(u>) v .
Some further definitions specific to developing the wavelet transform
are the scaled version of a function,
/·(*):= ; / φ 0>*0); (4)

the scaled and translated version of a function,
/.»(*) := -8fC-^-) (-#0); (5)

and t h e reflection through the origin of a function,
/-(«):=/(-*)· (6)
2
A basic w a v e l e t is an L (H) function, xp{t), which satisfies the fol-
lowing admissibility criteria [5, 7]:
(i) zero mean: J dtip(t) = 0, and
(ii) regularity: Οψ := f du \ψ(ω)\2/\ω\ < oo.

Additionally we normalize the wavelet to have norm unity, t h a t is, \\ψ\\2 =
1, and since we are interested in time-frequency (equivalently, space-fre-
quency) localization, we specify t h a t ψ be a windowing function [5]. T h a t
is, the first moment, t\j)(t), should be integrable. Requiring first and higher
moments to be zero imposes higher regularity on the wavelet.
D e f i n i t i o n 2 . 1 . T h e continuous wavelet transform of a function f(t) G

L 2 ( R ) at a location 6, relative to the basic wavelet φ at scale s, is defined
by
Wsf(b) := (/,ψ.„) = J f(t) \φ(*-^)dt. (7)
R
Equivalently,
W.f(b):=f*t/>;{b). (8)
2
T h e continuous wavelet transform of an L ( R ) function is a function
of scale, s φ 0, and location, 6, and as such forms a surface over t h e (5, b)-
plane called the phase plane of the wavelet transform.
Note t h a t we have chosen a particular scaling for our definition of t h e
wavelet transform in equation (7) through equation (4). Other choices are
possible [2, 10], however, it has been argued t h a t the choice provided by
equation (7) facilitates the interpretation of the wavelet transform as a
measure of coherence of structures with respect to scale [10]. We note t h a t
[10] a t t e m p t e d to distinguish between scaling choices of yfs and s in t h e
definition of the wavelet transform by calling scaling by yfs the wavelet
transform and scaling by s the covariance transform. It appears t h a t t h e
emerging convention in the literature is to disregard this distinction by
using t h e n a m e "wavelet transform" for any choice of scaling, and we adopt
this point of view, making distinctions where necessary. Using the scaling
yfs leads to an energy preserving transform, t h a t is, a type of Parseval's
relation for wavelets. While this is attractive for interpreting the wavelet
transform as a decomposition of variance, it obscures the interpretation for
detecting coherent events.
In the notation of equation (7) the wavelet transform may b e inter-
preted as a covariance between the wavelet at a given dilation, φ8 and t h e
signal, / , at a specified location in the signal [10]. Using the notation of
equation (8) the wavelet transform may also be interpreted as t h e signal,
/ , filtered by a particular (non-uniform, band pass) filter, φ8 [6, 21]. We
will discuss the application of the wavelet transform to the construction of
filters and the filtering properties of some wavelets in Section 3.
T h e continuous wavelet transform may be inverted to recover the signal
as follows.
(9)
where Οψ is the constant defined in the regularity condition (ii).
2 . 2 . D y a d i c s u b s a m p l i n g in scale
We restrict the scale parameter, s to powers of 2. This allows the
implementation of the fast wavelet transform algorithm given by [22].
Thus, let s = 2·7 for j any integer. For this restriction to yield a
usefully invertible transform, the regularity condition (ii) must be modified
as follows. T h e function φ G £ 2 ( R ) (which satisfies the zero mean condition
(i)) is called a dyadic wavelet if there exist positive constants 0 < A < B <
oo such t h a t
(10)
which is called the stability condition for the dyadic wavelet [5, 22]. It has
been shown t h a t a dyadic wavelet also satisfies the regularity condition (ii)
[5].
D e f i n i t i o n 2 . 2 . T h e (continuous) dyadic wavelet transform of the signal

/ at the scale 2 J at location b is defined by
(11)
where we denote φ2j by φ$. Equivalently,

(12)
By restricting the scale parameter to powers of 2, we have in effect sub-

sampled the continuous wavelet phase plane. This is sometimes referred t o
as t h e semi-discrete wavelet transform. It has been shown t h a t a complete
reconstruction of the signal may be accomplished using another wavelet

which is dual to the analyzing wavelet [5, 7, 22]. T h a t is, t h e forward
wavelet transform is performed using the basic analyzing wavelet, φ, while
t h e inverse wavelet transform must be accomplished using t h e dual wavelet,
X, defined as any function capable of performing the inverse transform in
t h e following way.
D e f i n i t i o n 2 . 3 . Any function, χ G L 2 ( R ) , is a dyadic dual of a dyadic

wavelet, ψ, provided every function, / G L 2 ( R ) , can b e written as
oo .
/(*)= Σ 2-3^2/^/(6)x(2^-6))d6. (13)

i=-oo R
In order to find the dual wavelet, conditions have been established

which relate the dual to the analyzing wavelet. T h e following result is
proven in [5].
Theorem 2.1. Given a dyadic wavelet, φ, and any L2(H) function, χ,

such that
oo
sup Σ |χ(2> ω )| 2 < oo , (14)
uçRj= — oo
then x is a dyadic dual of φ if and only if
oo
Σ φ·@ω)χ@ω) = 1 . (15)
j=—oo
In equation (15), ψ* denotes the complex conjugate of φ. There may

b e many functions, χ, which will satisfy equation 15 for a given φ. T h e
condition given in equation (14) is a smoothness condition similar to t h e
stability condition imposed on the basic wavelet in equation (10). Equa-
tion (15) expresses t h e restriction to an energy preserving forward-inverse
transform relation.
T h e set of functions {ipj(t — b)} for — oo < j < oo and b G R , used
to define t h e continuous dyadic wavelet transform, forms a non-orthogonal
basis for L 2 ( R ) . An orthogonal basis for I / 2 ( R ) may b e obtained from this
set by subsampling t h e translation parameter b using 6*. = 2Jfc to obtain
t h e set fyjk(t) = ψ]{ί — 2Jfc)} for j and k integers. By normalizing t h e
orthogonal set we have an orthonormal basis of L 2 ( R ) . In fact t h e proper
normalization is given by the following.
= - L ^ ( — ) for s = 2j,b = 2jk. (16)

y
y/S S '
Thus, by choosing the normalization in equation (4) t o be ips(t) = 1/\fsip{t/s)

and subsampling in t h e orthogonal fashion, we arrive at t h e orthonormal
basis tyjk}· We denote t h e orthonormal wavelet coefficients obtained from
the signal / by Wjkf t o distinguish t h e m from t h e dyadic wavelet coeffi-
cients given by Wjf in equation (12).
Orthonormal wavelets provide certain advantages. For example, they
are self-dual in t h e sense t h a t χ = ψ in equation (13), with t h e integral over
b replaced by a sum over k. T h e orthonormal wavelet coefficients satisfy
preservation of energy as a type of Parseval's relation. T h a t is,
oo
2
ll/ll = Σ I^WI 2 · (17)
j,k= — oo
Using t h e orthonormal wavelet transform it is possible t o define t h e

wavelet spectrum in t h e following way [14, 27, 28]. Rewriting equation (17)
as
oo
2 E
ii/ii = Σ i
j= — oo
oo oo
= Σ Σ i^wi2, (is)
j= — oo k= — oo
we can define t h e wavelet spectrum of an Li function, / , as

oo
EJ= Σ i ^ i 2 · (19)
k= — oo
It is possible t o assign a correspondence between t h e Fourier wavenumber

and t h e scale of t h e wavelet transform. This can b e done formally by
assigning t h e wavenumber at t h e center of mass of t h e frequency content
of t h e wavelet at a particular scale [27]. However, for comparing power law
trends of wavelet spectra with trends in Fourier spectra t h e assignment is
essentially arbitrary as long as it remains consistent through t h e dilation
of t h e wavelet.
If t h e energy spectrum, |/(u;)| 2 /u;, obeys a scaling law, say | / ( C J ) | 2 / C J ~
P
CJ~ , then t h e wavelet spectrum is expected t o possess a scaling law through
the correspondence Ej <-» / ( ω ) | 2 ~ constant · 2^p~l\ (In our notation,
j -Λ —oo implies ω —> oo. T h a t is, large positive j corresponds t o increasing
scale; large negative j corresponds t o decreasing scale). T h e similarity
symbol, ~ , is used t o indicate t h a t for small scales (ω large, j negative a n d
large) t h e spectrum has a trend corresponding t o exponential decay.
2 . 3 . F i n i t e r e s o l u t i o n in d i s c r e t e a p p l i c a t i o n s
Since discrete d a t a are limited to finite resolution (smallest scale) de-
termined by a sampling rate it is not possible to perform an analysis at a
resolution finer t h a n some fixed scale. Choosing t h e smallest scale t o b e
unity (s = 23\ j = 0) means we cannot represent d a t a on scales between
grid points. Actual d a t a are also limited to some finite record length, say
s = V for j = J > 1. Furthermore, any definitions of functions on t h e finite
grid may fail to follow the continuous properties in the conditions given by
equations (10), (14), and (15). This is discussed in detail in [22]. T h e
solution is to introduce a smoothing function, φ, whose frequency content
represents the fractional loss of energy when performing a wavelet trans-
form followed by the inverse transform (under the restrictions imposed by
finite resolution). T h a t is,
oo
φ(ω) = Σψ(2*ω)χ(νω).
T h e relation in equation (15) then can be written
\φ(ω)\2 - \φ(^ω)\2 = ΣΦ*&»)*@ω) , (20)

i=i
for ψ*(23ω)χ(23ω) a positive, real, even function.

Denote by Sj the smoothing operator at scale j given by
5i/(i):=/*^(*), (21)
3 3
where φ^(ί) = 2~ φ^2~ ). T h e information content of the signal smoothed
at some coarse scale, Sjf{t), and the wavelet coefficients for all scales up to
the coarse scale, {Wj/(£)}i<j<./5 is sufficient to recover the original discrete
signal. We accomplish this using the fast wavelet transform algorithm
described in [22].
As noted, the orthonormal wavelet coefficients may be obtained from
the continuous dyadic wavelet coefficients through appropriate subsampling
and scaling. For real applications the d a t a record length is finite and an
orthogonal subsampling depends on which d a t a point is assigned as t h e
end of record. Since the wavelet transform is not shift invariant, a shift
in the location at which the wavelet coefficients are computed results in a
change in t h e wavelet coefficients, and hence a change in the estimate of
t h e spectrum resulting from the coefficients. By averaging over all possible
orthogonal subsamplings of the translation parameter we approach a shift
invariant estimate of the spectrum. An equivalent approach would be to
average over several globally translated orthogonal wavelet analyses.
W h e n computing spectra of finite resolution d a t a using traditional ap-

plications of the fast Fourier transform, it is well known t h a t t h e resulting
spectral estimates are highly dependent on windowing the d a t a ([13, 16]).
T h e wavelet transform creates its own window (the wavelet) and the char-
acteristics of the spectral estimate are then dependent on the regularity
of the analyzing wavelet. For an application of these ideas to atmospheric
d a t a see [9].
§3. Filtering P r o p e r t i e s of W a v e l e t s
3 . 1 . P a r t i t i o n s as filters
We wish to partition a signal into two components, one containing
significant structures and the other containing the remaining portion of
the signal. Once a partition is obtained, further analysis of each compo-
nent can proceed. For example, an intermittency index can be assigned
to the structure-containing component, and spectral characteristics of t h e
components can be analyzed. In certain communication theory applica-
tions t h e significant structures may correspond to a signal prior to being
transmitted, while the remaining portion of the signal corresponds to t h e
noise introduced during transmission. In the case of atmospheric turbu-
lence, structures may be defined as regions of sharp transition of significant
amplitude [10, 20], while the remaining portion corresponds to a different
physical flow characteristic (not necessarily regarded as "noise").
To create a partition of a signal one must begin with assumptions
regarding t h e n a t u r e of the signal. For example, Fourier (band-pass) filter-
ing partitions a signal into components based on frequency content. Some
criterion is used to determine the components, such as the location of a fre-
quency band or the location of multiple bands determined by a frequency
amplitude cutoff. T h e partition then consists of the portion of the signal
with frequency content inside the bands and the portion with frequency
content outside the bands. A filter consists of retaining only one element
of the partition. If the significant structures are band limited and t h e
remaining portion of the signal is outside the structure containing band,
then the signal will exhibit spectral gaps. T h e b a n d - p a s s filters can then
be prescribed according to the location of the spectral gaps.
However, if the structure of interest in the signal is a sharp transition
it contains significant high frequency content. T h e high frequency content
can not b e distinguished from the frequency contribution of several smaller
amplitude sharp transitions, or from colored or white noise. In this case
the component of interest in the signal is not usefully band limited and t h e
spectrum will not exhibit a clear spectral gap. A Fourier filtering technique
is not well suited to creating a partition of the signal in such a case. T h e
wavelet transform can help in this situation since it effectively limits t h e
spectral content under consideration to specific time or space locations.

T h e wavelet phase plane contains information on the coherence be-
tween t h e signal and the wavelet in b o t h position and scale. This can b e
condensed into a global statistic providing information a b o u t t h e scale a t
which there are t h e largest number of coherent events of significant ampli-
tudes. T h e global statistic for a signal, / , is called the wavelet variance
[10, 20] and is defined in the continuous case as
oo
2
D {s)= j[Wsf{b)fdb. (22)
— oo
This is similar to the wavelet spectrum of equation (19), b u t is not a vari-

ance decomposition of the signal with the choice of scaling given in equa-
tion (4). (It is possible to interpret the wavelet variance as proportional t o
the energy (variance) of the signal using appropriate logarithmic axes [4]).
However, the quantity D2 is a variance since it is the squared error a b o u t
zero of t h e wavelet transform (which removes the mean from t h e signal).
Note t h a t there are two situations which will contribute t o producing a
large value of the wavelet variance a t a given scale - one large amplitude
coherent event or several coherent events of lesser amplitude. For analy-
sis of atmospheric turbulence signals, b o t h situations may b e considered
significant.
We note t h a t when information about dominant physical length scales
of i m p o r t a n t structures are deduced from the wavelet variance, the contin-
uous scale transform, rather t h a n a dyadic subsampling, should b e consid-
ered. T h e subsampling, while providing complete information for inversion,
does not allow for a complete physical analysis since important physical
scales may be skipped.
While t h e wavelet variance is a reasonable means of characterizing
signals, a characterization by the type of events they contain, and a measure
of their intermittency would provide further description. We address this
issue in Section 3c.
Locating local maxima in the wavelet variance provides a means of
determining t h e scales a t which the structures in the signal are coherent
with t h e wavelet shape, and of significant amplitude a n d / o r number. I t is
possible t o imagine situations where there is more t h a n one local m a x i m a
in t h e wavelet variance. For example, if there is significant white noise in
t h e signal, t h e wavelet variance will have a large peak at t h e smallest scale,
and a somewhat smaller peak a t the scale of the coherent events. Another
example is a signal consisting of short pulses. One peak in t h e wavelet
variance corresponds to the scale of the length of the pulse and another peak
corresponds t o t h e spacing of t h e pulses, assuming t h e spacing and width
are different scales [8]. Finally, a superposition of two sinusoidal waves a t
different frequencies will produce two local maxima in the wavelet variance
[4]. In such a situation, each local maximum corresponds to a wavelength
and some choice as to their significance must be based on knowledge of t h e
n a t u r e of the signal.
Once the dominant event scale or scales are found, one of the obvious
methods leading to a partition of the signal is to limit the reconstruction
from the phase plane to the scales above and below some scale near t h e
dominant scale, or between two dominant scales. One approach to deter-
mining a partition if only one dominant scale is present is to reconstruct
from t h e phase plane information using the large scales down to some small
scale which is determined from the dominant scale by taking an appropri-
ate fraction of t h e dominant scale. The partition then consists of the large
scale reconstruction and the small scale reconstruction (that is, the remain-
ing portion of t h e signal). If there is a clear "scale gap" between peaks in
the wavelet variance, the scale for the partition could be chosen to be in
the gap [4]. We refer to this as scale threshold partitioning. Scale thresh-
old partitioning has the problem of neglecting the fact t h a t structures of
interest are multiscaled and may have significant information content even
at small scales. T h a t is, portions of the information content of a single
structure may fall into b o t h elements of the partition. This is an inherent
problem in partitioning signals in which multiscale structures are impor-
t a n t b u t are found in the presence of broad band noise. Another inherent
problem is determining the appropriate fraction of the dominant scale for
the threshold. Examples are given later in this section and in Section 4.
Even with these difficulties, a scale threshold partition works very well for
separating out white noise at the smallest resolvable scale.
In [23] the rate of growth or decay of significant "ridges" in the phase
plane is used as a means of quantifying what is meant by noise. This is
a measure of the regularity of a multiscale feature. For highly singular
features the wavelet transform will undergo a rapid increase in value at
small scales as the scale is decreased. Assuming t h a t the "noise" is less
regular t h a n the structures of interest and limited to small scales, a ridge
reconstruction ([23, 22]) is performed using only the portions of ridges
which do not exhibit an increase at small scales, or by extending ridges to
smaller scales keeping the amplitude fixed.
In [23] the use of ridge threshold filtering described above is motivated
by the interest in signal compression. For this study we have no need for
signal compression and therefore have the information of the entire phase
plane at our disposal. We can therefore employ a simpler approach to
partitioning which still allows for the multiscale nature of coherent events.
Furthermore, for applications to atmospheric turbulence data, we are inter-
ested in partitioning the signal rather t h a n an actual filtering (that is, we
retain b o t h components of the partition). T h e method we examine is t h e
reconstruction of the portion of the signal whose wavelet coefficients are

larger or smaller t h a n a particular threshold. Reconstruction of the signal
using coefficients larger t h a n some fixed value, | W y / ( n ) | > u; c , preserves
the structural features. The remaining coefficients, | W j / ( n ) | < wci are
used to reconstruct the portion of the signal not associated with t h e struc-
tures. We call this type of partitioning of a signal phase plane threshold
partitioning.
Note t h a t phase plane threshold partitioning requires the scaling chosen
in equation (4) for the definition of the wavelet transform. If some other
scaling is chosen then an appropriately scaled cutoff must b e used. For
the orthonormal decomposition, in which the scaling is chosen as 1/y/s
(equation 16), the phase plane cutoff is scale dependent and given by
T h e partitioning of the signal into a component containing structures

and a component containing the remainder allows the comparison of t h e
spectra of the respective components. We will demonstrate t h a t t h e spec-
tral characteristics of each component of a velocity record are quite differ-
ent. T h e partitioning can be accomplished using the wavelet partitioning
m e t h o d s described above. The partition is dependent upon t h e choice of
analyzing wavelet and the m e t h o d of thresholding used. We utilize these
characteristics to analyze buoyancy flux in the atmospheric boundary layer
in Section 4.
3 . 2 . D e t e c t i o n characteristics of anti—symmetric a n d s y m m e t r i c
wavelets
In this section we demonstrate certain general characteristics of t h e
differences between using symmetric versus anti-symmetric wavelets for
signal analysis. Figure l a is the vertical velocity record from an aircraft
measurement during the F I F E 87 field experiment (described in Section 4).
T h e signal contains structures resembling square or r a m p e d pulses on t h e
order of 10 seconds wide (40 meters). There are approximately eight or nine
such events. In order to illustrate the response of the wavelet transform to
such structures, consider the simulated signal of Figure 2. This consists of
four "pulses" of similar shape, unevenly spaced, with a resolution of 256
points. T h e shape was chosen to have some characteristics typical of t h e
structures observed in the F I F E 87 vertical velocity record.
T h e subsequent analyses are performed using the anti-symmetric quad-
ratic spline wavelet and the symmetric cubic spline wavelet shown in Fig-
ure 3. T h e wavelet transform of the simulated signal based on a symmetric
wavelet and an anti-symmetric wavelet at a single scale are shown in Fig-
ure 4. Note t h e magnitude of the symmetric transform is large at t h e
boundaries of t h e transitions, while the magnitude of t h e anti-symmetric
Figure 1. Morning case: (a) (top) Vertical velocity record shown as

the fluctuation about the mean, w' (m/s) = w — w, where w = —0.0045
(m/s), (b) (middle) Virtual potential temperature fluctuation, θ'ν (K).
Tv - 301.46 (K). (c) (bottom) Buoyancy flux density, w'9'v (ms _ 1 K).
Figure 1. Afternoon case (d) (top) Vertical velocity fluctuation, w'

(m/s), (e) (middle) Virtual potential temperature fluctuation, θ'ν (K).
(f) (bottom) Buoyancy flux density, w'0'v (ms - 1 K).
transform is large at t h e center of t h e transitions. This characteristic is

scale dependent since at scales larger t h a n a particular event size t h e lo-
cation of m a x i m u m coherence between the wavelet and t h e signal may b e
Figure 2. Simulated signal having features qualitatively similar to

features in the vertical velocity record (Figure la).
quadratic spline wavelet

cubic spline wavelet
Figure 3. Anti-symmetric quadratic spline (solid line) and symmetric

cubic spline (dashed line) used for wavelet analysis.
influenced by characteristics of the signal outside of the event. This is

illustrated in Figure 5, where the anti-symmetric transform is shown at
three scales. Note t h a t the location of maximal gradient, corresponding to
extremes of t h e transform, migrate as larger scales are considered [22].
We now t u r n to the notion of creating a partition of the signal. Con-
Figure 4. Anti-symmetric (long dashed line) and symmetric (short

dashed line) wavelet transforms of the simulated signal (solid line) at
the scale 2 3 .
sider t h e signal shown in Figure 6 consisting of the simulated signal given

in Figure 2 plus small amplitude sine wave with random phase. This is
intended to represent a situation where the desired features are the larger
scale structures which contain sharp transitions but are found in the pres-
ence of significant smaller scale information.
T h e dominant scale of structures must first be found using t h e wavelet
variance defined in equation (22). T h e dyadic equivalent of equation (22)
was used to compute an estimate of the wavelet variance of the simu-
lated signals of Figure 2 and Figure 6 using b o t h the symmetric and a n t i -
symmetric wavelets. The result is shown in Figure 7. Four curves are
shown corresponding to computing the wavelet variance of t h e simulated
signal without noise and the simulated signal with noise using either a
symmetric or anti-symmetric wavelet. The maximum wavelet variance for
scales larger t h a n the smallest occurs at the scale 2 5 and is apparently in-
dependent of the choice of wavelet. This independence would suggest t h a t
identifying structures in a signal is independent of the symmetry of t h e
wavelet used. However, further test cases using a variety of length scales
in t h e structures show t h a t is not always true.
Figure 5. Anti-symmetric wavelet transform (dashed lines) of the

simulated signal (solid line) at three scales. Note the migration of the
location of the extremes. The location and interpretation of a coherent
event is necessarily scale dependent.
Figure 6. Simulated signal which includes small scale information con-

sisting of a mix of white noise and coherent small scale structure.
600
400
Figure 7. Wavelet variance computed using the discrete form of equa-

tion (22)
Once the dominant scale of structures is selected from a (local) maxi-

m u m of t h e wavelet variance, the local extremes (either maxima or minima)
of t h e wavelet transform at t h a t scale provides the location of t h e struc-
tures. Moreover, a phase plane partition can then be formed based on a
fraction of the largest (in magnitude) of the local extrema at t h e scale of
m a x i m u m wavelet variance. This process is illustrated in Figure 8 where
t h e anti-symmetric quadratic spline wavelet has been used.
Once a threshold value has been determined by the m e t h o d in t h e pre-
vious paragraph, a partition of t h e signal may b e formed. A reconstruction
based on the wavelet transform values above the threshold forms one com-
ponent and the remaining portion of the signal forms the other component.
This is shown in Figure 9 where a factor of 0.3 of the largest of the local ex-
t r e m a was used (determined empirically based on the influence of changes
in t h e factor on the variance of the n o n - s t r u c t u r e component).
T h e result in Figure 9 may be compared to a simple scale threshold
partitioning shown in Figure 10, where the scale threshold is 2 1 . T h a t is,
t h e smallest scale information forms one element of t h e partition and all
scales above t h a t form the other. This is the best way in which to separate
Figure 8. Anti-symmetric wavelet transform (dashed line) of simu-

lated signal (solid line) at the scale of maximum variance (from Figure
7). The arrows indicate the location of the local extremes. A factor
times the largest of the extremes is used to determine the phase plane
threshold value.
out white noise, since it has a decorrelation length scale of two points. Scale
partitioning has been suggested as a method of filtering in [4].
Finally, the phase plane threshold and the scale threshold may be com-
bined to create a partition such as t h a t shown in Figure 11. We have
combined a phase plane threshold of a factor of 0.2 (also emperically deter-
mined as above) of the largest of the local extremes with a scale threshold
of 2 1 .
3.3. Quantifying intermittency

Intermittency has many manifestations depending on physical context.
One classic example is the laboratory flow constructed by ejecting fluid of
a higher t h a n ambient temperature from a jet. The time history of temper-
ature at a point downstream in the turbulent flow will exhibit intermittent
spikes as small cores of fluid of high temperature from the jet pass by t h e
sensor. As t h e sensor is moved from the center of the jet flow to the edge of
the domain, t h e intermittency effect increases. T h a t is, the fraction of t h e
Figure 9. Phase plane threshold partition of the simulated signal

shown in Figure 6. A constant was added to the large scale portion
to offset the plots.
record which is high t e m p e r a t u r e decreases as the sensor moves away from

t h e center of the domain. An intermittency index defined as the fraction of
t h e record occupied by high temperature fluid provides information a b o u t
the degree of entrainment of ambient fluid into the high t e m p e r a t u r e core.
See [26] for a more detailed discussion. In Figure l c the regions of sus-
tained positive buoyancy flux density are the structures of interest. These
structures could be identified by applying a threshold to t h e data, b u t t h e
intermittency (fractional record length where d a t a is large) is sensitive to
the choice of the threshold. Furthermore, thresholding does not take into
account the coherency of the structures. For the vertical velocity record
shown in Figure l a , intermittency depends on the type of event one wishes
to define. For example, we are interested in identifying regions of sharp
gradient of sufficient magnitude (which can not be identified using a thresh-
old on t h e d a t a ) . T h e wavelet transform provides a means of locating such
events. Once their location is known some measure of t h e intermittency of
these occurrences with respect to the total signal length can be devised.
T h e main purpose of this section is to propose a new measure of in-
-i r
50 100 150 200 250

Time (s)
F i g u r e 10. Scale partition (at scale 2 1 ) of the simulated signal shown

in Figure 6. A constant was added to the large scale portion to offset
the plots.
termittency computed directly from the wavelet transform information.

This measure of intermittency combined with the fraction of buoyancy flux
contributed by the structure component of buoyancy record provides an
important insight into the n a t u r e of the boundary layer convective pro-
cess. In Section 3b we have discussed finding the dominant scale using the
wavelet variance, and locating the local extremes in the wavelet transform
at t h a t scale. Using this information the signal is partitioned into a struc-
t u r e component and a non-strucuture component. As a measure of t h e
fraction of the record containing structures, an intermittency index can b e
formed as
support of structures
intermittency = (23)
total record length
For example, if the structure component of the buoyancy flux accounts for
most of the total flux and the intermittency index is small (very intermit-
tent coherent structures) then there are a few strong convective plumes in
t h e flow. In contrast, if the intermittency index is large (most of t h e record
occupied by coherent structures) then there is a more uniform t r a n s p o r t
(such as in regions of subsidence or secondary flow).
Figure 1 1 . Combination of a scale (at scale 2 1 ) and phase plane

threshold (at 0.25 times the largest extreme) to create a combination
partition of the simulated signal of Figure 6. A constant was added to
the large scale portion to offset the plots.
This definition of intermittency is dependent on t h e thresholding used

to locate local extremes in the wavelet transform, b u t does not depend on
thresholding the data.
§4. A p p l i c a t i o n s t o A t m o s p h e r i c D a t a
4 . 1 . D e s c r i p t i o n of t h e d a t a
We now apply the techniques described in Sections 2 and 3 to two
portions of d a t a collected during the First International Satellite Land
Surface Climatology Project (ISLSCP) Field Experiment 87 (FIFE87). A
description of t h e experiment and analysis of d a t a using other techniques
including conditional sampling may be found in the F I F E special issue of
J G R [15]. T h e purpose of this section is to demonstrate t h e utility of t h e
techniques presented in the previous sections by providing new insight into
t h e interpretation of specific data. This is an extension of t h e results shown
in [12] where a different segment of the d a t a was analyzed.
An objective of FIFE87 was to gather in situ boundary layer measure-
ments in conjunction with satellite remote measurements to obtain suitable

parameterization schemes. In part, the objectives are to determine the ac-
curacy of, and to calibrate, satellite d a t a and to relate the remotely sensed
d a t a to flux quantities. T h e intermittent nature of momentum, heat, and
buoyancy fluxes makes this task difficult and may lead to deceptive esti-
mates from remotely sensed data. The in situ d a t a used in this study were
sampled using the National Center for Atmospheric Research (NC AR) King
Air research aircraft in fair weather conditions over a 50 kilometer square
of uniform tallgrass over the Konza National Prairie in northeast Kansas—
the designated F I F E site. A case study for June 6, 1987 from the Intensive
Field Campaign 1 ( I F C l ) is reported in [11]. June 6 t h was a clear day des-
ignated as t h e "golden day" for FIFE87 I F C l . We study two portions of
the d a t a from J u n e 6 t h corresponding to late morning and late afternoon.
These two times present differing physical situations; late morning being
a strongly convective, developing mixed layer, and late afternoon having
subsiding thermal forcing and the resulting weak or decaying convection
combined with a well mixed and sheared boundary layer.
T h e measurements were obtained on June 6 t h at 650 meters above
ground in the morning and 1300 meters above ground in the afternoon
with 4 meter resolution along an east to west transect at the northern
edge of the site. The flights remained at a fixed position with respect to
the atmospheric boundary layer top (about 90% of ABL) to ensure consis-
tency in the sampled data. The d a t a were collected over the time intervals
11:26:46 L (Local time (L) = -6 U T C ) to 11:30:11 L for the morning leg,
and 15:55:56 L to 15:59:21 L for the afternoon leg. Figure l a - c shows
the vertical velocity component, the virtual potential temperature, and the
buoyancy flux density derived from the virtual potential t e m p e r a t u r e and
the vertical velocity for the morning leg. Figure l d - f shows the corre-
sponding fields for the afternoon. Note the intermittently occurring zones
of sharp transition in the vertical velocity (Figure l a and d), and t h e zones
of positive buoyancy flux (Figure l c and f). Note also the the number and
size of r a m p and plume structures (in b o t h velocity and temperature) is
different between the morning and afternoon legs.
4 . 2 . A n a l y s i s of t h e d a t a
Table 1 collects some length scale statistics computed using an a n t i -
symmetric wavelet transform for vertical velocity components and the vir-
tual potential t e m p e r a t u r e for the morning and afternoon legs. Listed are
the dominant scale lengths (as derived from the largest local maximum in
the wavelet variance), the number of occurrences of events of this size, and
t h e intermittency index as defined in Section 3c. Due to large amplitude
low frequency components in the temperature records, the largest variance
for an anti-symmetric wavelet will occur at the largest scale. We therefore
chose to use t h e next smallest scale at which a local maximum occurred.

T h e determination of dominant scales based on the dyadic wavelet trans-
form subsamples the continuous scale information t h a t would b e possible
using a continuous scale. However, t h e dominant scales are in t h e smaller
scales where there is a better sampling density, and the scale distinctions
at higher resolution are not particularly relevant to the present analysis.
Table 2 collects the same statistics based on using a symmetric wavelet.
T h e dominant length scales tend to be the same or slightly longer, and
the number of events the same or slightly fewer. Correspondingly, t h e
intermittency index is the same or slightly larger (less intermittent).
Both Table 1 and Table 2 show t h a t coherent structures are of smaller
scale in the afternoon t h a n the morning and are much more (spatially)
frequent. This can be seen by comparing, for example, the vertical velocity
record from the morning (Figure l a ) with the vertical velocity record from
the afternoon (Figure I d ) .
Table 1: Length scale analysis using an anti-symmetric wavelet

Dominant Number Intermittency
Field Length Scale (km) of events Index
Morning case:
w 1 10 0.59
θυ 0.5 18 0.53
*»% 0.5 20 0.61
Afternoon case:
w 0.125 76 0.57
θν 0.25 36 0.54
χυ'θ'υ 0.06 142 0.54
Table 2. Length scale analysis using a symmetric wavelet

Dominant Number Intermittency
Field Length Scale (km) of events Index
Morning case:
w 1 10 0.61
θν 0.5 19 0.57
w'9'v 1 8 0.49
Afternoon case:
w 0.125 77 0.59
θν .25 36 0.55
w'9'v 0.125 83 0.62
A combination phase plane threshold and scale partition has been cre-
ated for the velocity components (Figure 12a and c) and the buoyancy flux
densities (Figure 12b and d) for the morning and afternoon cases. Combi-
nation partitions were also created for the temperature records but are not
shown. An anti-symmetric wavelet analysis was used in each case. The
sharp vertical edges have been maintained in the larger scale partition in
each case.
Total buoyancy flux over the signal was computed by summing over
the flux density for each of five possibilities - from the original data, νϋ'θ'υ,
which will be the reference flux; from the structure partition of the derived
buoyancy flux density record, (u/0(,) str ; from the structure component of
the velocity with the original temperature data, u)'stx9'v; from the structure
component of the temperature with the original velocity data, t i / ^ s t r ) ' ;
and from the structure component of velocity with the structure component
of temperature, (wstTY(6v stTy. This was done for the morning case and
the afternoon case resulting in the ten numbers shown in Table 3 where the
total fluxes are shown as a percentage of the reference flux. The morning
reference flux was about 40% of the afternoon reference flux.
The flux transport computed using the structure component of velocity,
(u;str )'#(,, accounts for 69% of the reference flux in the morning and 58% of
the reference flux in the afternoon. The flux transport computed using the
structure component of temperature, u)f(6VstJ.y, is relatively higher in both
the morning (82%) and the afternoon (75%). Additionally, the total flux
computed using the structure component of the velocity with the structure
component of the temperature, (wstTy(6V8tT)\ is 58% in the morning and
42% in the afternoon indicating that the temperature structures are not
highly correlated with the velocity structures.
Table 3. Total flux analysis using structure partitions |

Flux density Total Flux 1
Morning case:
w'e'v Buoyancy flux 100%
(tu'É^tr Structure part of flux 100%
<Λ Flux due to velocity structures 69%
«/(<W Flux due to temperature structures 82%
(">str)'(0t> str)' Flux of temp. str. due to vel. str. 58%
Afternoon case:
ιν'θ'ν Buoyancy flux 100%
(u>'0C).tr Structure part of flux 100%
Flux due to velocity structures 58%
<X
t0'(e».tr)' Flux due to temperature structures 75%
(w S tr)'(0« str)' Flux of temp. str. due to vel. str. 42%
2 h
> o h
1 I — i 1 — i 1 — I 1 r η — i 1 — i 1 1 1 — i — i 1 p
Figure 12. Morning case: Combination phase plane and scale parti-
tions for (a) (top) vertical velocity and (b) (bottom) buoyancy flux
density, Vertical shifts have been added to offset the components.
~\—i—|—i—i—i—i—i—i—i—i—i—i—i—i—i—i—r
_J I i i i i I i i i i L
50 100 150 200
Time (s)
Figure 12. Afternoon case: Combination phase plane and scale par-
titions for (c) (top) vertical velocity and (d) (bottom) buoyancy flux
density, Vertical shifts have been added to offset the components.
In t h e lower portion of the convective boundary layer (CBL) the buoy-

ancy generation is due to heating resulting in coherent t e m p e r a t u r e struc-
tures, (evstT)', accounting for a large fraction of t h e buoyancy flux. Near
the top of the CBL t h e buoyancy is driven by accumulated m o m e n t u m ,
(wstr)'> of the plume structures [18]. T h e d a t a were gathered in a mid-
dle portion of t h e CBL (650 meters in the morning; 1300 meters in t h e
afternoon), and the percentage of buoyancy flux due to virtual potential
t e m p e r a t u r e structures being larger t h a n buoyancy flux due to velocity
structures reflects this in b o t h the morning and afternoon cases. Compar-
ing t h e velocity partition of the morning case (Figure 12a) to t h e velocity
partition of the afternoon case (Figure 12c) it is apparent t h a t t h e morning
case contains relatively larger, well-defined structures. This is reflected in
a comparison of the buoyancy flux densities as well (Figures 12b and d).
This is quantified in the length scale analysis summarized in Table 2.
T h e late morning boundary layer is characterized by vigorous convec-
tive activity which results in large plume structures penetrating from the
b o t t o m to t h e top of the CBL. By afternoon, vertical soundings taken si-
multaneously with t h e aircraft d a t a indicate t h a t significant vertical shear
of horizontal winds has developed and t h a t t h e CBL is well mixed. This
leads to entrainment of upper air into the CBL and breaks u p t h e plume
structures into smaller sized structures. Combined with t h e weakening
heating at t h e ground, this accounts for t h e observed smaller and more
numerous structures in t h e afternoon d a t a compared to t h e morning data.
This conceptual model is supported by the statistics gathered in Table 1.
T h e larger vertical velocity structures occur in the morning d a t a , b u t are
fewer in number, while the intermittency remains approximately t h e same.
T h a t is, t h e total area covered by vertical velocity structures is relatively
the same, b u t the size decreases by the afternoon. T h e virtual potential
t e m p e r a t u r e structures follow a similar, but less pronounced, p a t t e r n .
T h e Fourier spectra of the morning and afternoon vertical velocity
records are shown in Figure 13a and b, respectively. It is apparent t h a t t h e
size of t h e predominant structures decreases in t h e afternoon, since most of
the energy migrates to short wavelengths. Figure 14 shows t h e spectral es-
timates for the structure and n o n - s t r u c t u r e elements of t h e morning record
(Figure 14a) and the afternoon record (Figure 14b). T h e spectra are plot-
ted as log(k~5/3 /(&)) vs. log{k) to visualize a fc~5/3 power law behavior
as a flat region. In t h e inertia! range (200m and smaller) b o t h morning
and afternoon spectra indicate a —5/3 power law behavior (seen as a flat
region in Figure 13). T h e morning spectra contains more scatter due to
the smaller number of large coherent structures contained in t h e record (see
Table 1,2). In the afternoon when t h e record contained a large number of
small sized coherent structures, the spectra has less scatter. Many studies
have computed energy density spectra of observed turbulence records and
either established the presence or absence of t h e —5/3 power law region.

In [19] it was observed using d a t a from several experiments t h a t t h e —5/3
region is seen in strong turbulence but not in weak turbulence. It is sug-
gested in [12] t h a t t h e universal —5/3 slope spectra is found in t h e spectra
of the structure partition of velocity, while the n o n - s t r u c t u r e partition is
flat or has a —1 power law behavior. In Figure 14 we show t h e spectra of
the velocity structure component (thick line) and the n o n - s t r u c t u r e com-
ponent (thin line). T h e structure component of the record contains a —5/3
power law region while the non-structure component has a —1 power law
behavior. These results are consistent with [12]. T h e similarity of t h e
spectra of the structure component to spectra of strong turbulence and
those of the n o n - s t r u c t u r e to weak turbulence suggest t h a t a b e t t e r char-
acterization of turbulence may be achieved by the use of the intermittency
index and the size distribution of the coherent structure component of t h e
velocity.
§5. S u m m a r y a n d C o n c l u s i o n s
In this paper we have utilized the wavelet transform as a time (or space)
series analysis tool for the analysis of d a t a containing intermittent coherent
events. We use a non-orthogonal formulation of the wavelet transform, and
choose scaling to emphasize the edge and singularity detection capabilities
of the transform. T h e fast algorithm of [22] is implemented for this purpose.
We use local maxima of the wavelet variance (defined in Section 2) as
a means for identifying important scales in the data. This approach has
been used in other studies as well [10, 4]. It is important to note t h a t
this is a reduction of the wavelet phase plane and is subject to the bias of
non-characteristic events. For example, a single event with a large ampli-
t u d e may determine a local maximum in t h e wavelet variance when several
events of a differing scale may contain the desired structural information.
Using the wavelet transform for detecting coherent events is consid-
ered in Section 3. T h e locations of local extrema of an anti-symmetric
wavelet transform correspond t o locations of centers of large gradients in
the signal. Similarly, the locations of local extrema of a symmetric wavelet
transform correspond to locations of large curvature in the signal. Alterna-
tively, t h e zero crossings of the symmetric wavelet transform correspond to
locations of centers of gradients in the signal, b u t t h e magnitude (sharpness
of the gradient) is then only determined by the slope of t h e zero crossing.
Each type of detection is necessarily scale dependent which must b e consid-
ered when a t t e m p t i n g to identify structures. A crucial choice for structure
detection algorithms is between anti-symmetric and symmetric analyzing
wavelets. T h e specific type of wavelet, and its regularity are less critical in
applications to real data. This is due in part to the limitations imposed by
i 1 i i i i
103 101
Time (s)
| I I I I—I 1 Γ I i i i i—i—i r
M i l I I I L
10j 10z
wavelength (m)
Figure 13. Spectral variance density derived from the vertical velocity
component for (a) (top) the morning case, and (b) (bottom) the
afternoon case. The spectra have been multiplied by k5'3 so that a
—5/3 region will appear horizontal in the figure.
lo- 1 L
"Π—I 1 1 Γ
I I IΊ I L· J I 1 L·
10J 10z
wavelength(m)
Figure 14. Spectral variance density of structure (thick line) and non-
structure (thin line) components of w-velocity shown in Figure 12a and
c. (a) (top) Morning case, and (b) (bottom) afternoon case.
finite resolution and finite record length.

Having determined a dominant length scale, the wavelet transform at
the dominant scale of the signal can be used to identify t h e location of
coherent events of t h a t size (by locating local extremes in t h e transform).
We then define an intermit tency index based on the ratio of t h e support of
wavelets at these locations to the total record length.
In order to study the properties of the structure containing component
of a signal separately from the remaining component we utilize three types
of signal partitioning (Section 3). A scale threshold partition, a wavelet
phase plane partition, and a combination of both. The partitions may b e
visualized as taking the wavelet phase plane information derived from t h e
signal and partitioning it into two phase planes based on scale, magnitude
of t h e wavelet transform, or both. Each of the two new phase planes is
then used to reconstruct a signal, whose sum is the original signal. T h e
localization in time and frequency of the wavelet transform allows a par-
tition which retains sharp edge characteristics in the structure component
which not possible with traditional Fourier band pass techniques.
Finally, these tools are applied to atmospheric d a t a from the F I F E 8 7
experiment. T h e analysis performed here complements a previous study
which used d a t a from the same field experiment [12]. T h e vertical veloc-
ity components and the buoyancy flux density are partitioned using t h e
anti-symmetric wavelet transform for fields obtained during late morning
and late afternoon. An intermittency index and dominant scale length are
computed for each signal. Additionally, total fluxes were computed and
analyzed in terms of the structure components of the signals. We char-
acterize t h e evolution of the convective boundary layer by comparing t h e
characteristic scale lengths of velocity and temperature structures for t h e
morning and afternoon cases. Additionally, we summarize t h e contribution
of velocity and t e m p e r a t u r e structures to buoyancy flux by comparing t h e
contributions of various partitions of the signals. The findings are consis-
tent with t h e conceptual model of a convective boundary layer subject t o
increasing shear and mixing during the afternoon. T h e more uniform mix-
ing results in t h e temperature structures being stronger and well defined
in t h e afternoon.
T h e spectral energy density of each of t h e elements of the velocity
partitions are compared to the spectral density of the original velocity
signals. We have found t h a t the structure component has t h e classical
characteristics of a strong turbulence signal, containing an energetic region
at long wavelengths and a transition region having a —5/3 behavior. T h e
n o n - s t r u c t u r e component exhibits —1 spectrum usually associated with
weak turbulence. Thus, the structure element of the partition behaves
(spectrally) like t h a t of strong turbulence while the n o n - s t r u c t u r e element
behaves more like t h a t of weak turbulence. We suggest t h a t t h e spectral
characteristics of turbulent velocities are determined by the intermittency

and scale lengths of structures, the combination yielding t h e —5/3 spectral
slope, and provide a better classification t h a n weak and strong turbulence
based on t h e spectral characteristics.
In this study we have applied a wavelet analysis to signals derived from
the atmospheric boundary layer. We chose certain aspects of t h e types of
wavelets used (symmetric and anti-symmetric), and the way in which they
are used (non-orthogonal translates) based on known signal processing
characteristics of wavelets [22]. The technique encompasses many desir-
able features of traditional conditional sampling techniques. An objective
means for defining and quantifying intermittency of coherent events, and
for creating partitions of signals which preserve the sharp gradients char-
acteristic of coherent events has been presented. T h e objective (i.e. com-
putable) n a t u r e of these characteristics will allow large d a t a bases to be
processed in a consistent fashion. This should allow further study of t h e
n a t u r e of t h e evolution of structures in differing planetary boundary layer
scenarios.
References
1. Antonia, R. A., A. J. Chambers, C. A. Friehe, and C. W. Van A t t a ,
1979: Temperature ramps in the atmospheric surface layer. Journal of
the Atmospheric Sciences, 36, N o . l , 99-108.
2. Arneodo, A., G. Grasseau, and M. Holschneider, 1989: Wavelet Trans-
form Analysis of Invariant Measures of Some Dynamical Systems. In
Wavelets (Combes, J. M., A. Grossman, and P h . Tchamitchian, Eds.),
Springer-Verlag, 315 pg.
3. Bergstrom, H. and U. Hogstrom, 1989: Turbulent exchange above a
pine forest II. Organized structures. Boundary-Layer Meteorology 4 9
231-263.
4. Collineau, S., and Y. Brunet, 1993: Detection of turbulent coherent
motions in a forest canopy; P a r t I: wavelet analysis. Boundary-Layer
Meteorology 6 5 357-379.
5. Chui, C. K., 1992: An introduction to wavelets. Academic Press, Inc.
266 pg.
6. Daubechies, I., 1988: Othonormal bases of compactly supported
wavelets. Communications in Pure and Applied Mathematics, 4 1 , 9 0 9 -
996.
7. Daubechies, I., 1992: Ten lectures on wavelets. C B M S - N S F Regional
Conference Series in Applied Mathematics, Soc. for Industrial and
Appl. M a t h . , 357 pg.
8. Gamage, N. K. K., 1989: Modeling and analysis of geophysical turbu-
lence: use of optimal transforms and basis sets. P h . D . Thesis, Oregon
S t a t e University, 135 pg. [Available from University Microfilm, 305 N.

Zeeb Rd., Ann Arbor, MI 48106.]
9. Gamage, N. and W. Blumen, 1993: Comparative analysis of low-level
cold fronts: wavelet, Fourier and empirical orthogonal function decom-
positions. Monthly Weather Review (submitted).
10. Gamage, N. and C. Hagelberg, 1993: Detection and Analysis of Mi-
crofronts and Associated Coherent Events using localized transforms.
Jour, of Atm. Sei. , 50, No. 5, 750-756.
11. Grossman, R. L. 1992: Convective boundary layer budgets of moisture
and sensible heat over an unstressed prairie. Journal of Geophysical
Research, 9 7 , no. D17, 18,425-18,438.
12. Hagelberg, C. R. and N. K. K. Gamage, 1993 (submitted): Struc-
t u r e preserving wavelet decompositions of intermittent turbulence.
Boundary-Lay er Meteorology.
13. Harris, F. J., 1978: On the use of windows for harmonic analysis with
the discrete Fourier transform. Proc. of the IEEE, 6 6 , 51-83.
14. Hudgins, L. C. A. Friehe, and M. E. Mayer, 1993: Wavelet transforms
and atmospheric turbulence. Physical Review Letters, 7 1 , No. 20, 3279-
3282.
15. J G R special issue, November, 1992: First ISLSCP Field Experiment.
Journal of Geophysical Research, 9 7 , November 1992. 18343-19110.
16. Kristensen, L., P. Kirkegaard, C. W. Fairall, J. C. Kaimal, and D. H.
Lenschow, 1992: Advantages of tapering finite d a t a records for spec-
tral analysis. NOAA Technical Memorandum E R L W P L - 2 6 6 , National
Technical Information Service, Springfield, VA.
17. Kuznetsov, V. R., A. A. Praskovksy, and V. A. Sablenikov, 1992: F i n e -
scale turbulence structure of intermittent shear flows. Journ. of Fluid
Mechanics, 2 4 3 , 595-622.
18. Lenschow, D. H. and P. L. Stephens, 1980: T h e role of thermals in t h e
convective boundary layer. Boundary Layer Meteorology, 1 9 , 509-532.
19. M a h r t , L. and N. Gamage, 1987: Observations of turbulence in strati-
fied flow. Jour. Atm. Sei., 44, 1106-1121.
20. M a h r t , L., 1991: Eddy asymmetry in the sheared heated b o u n d a r y
layer. Jour. Atm. Sei., 4 8 , 472-492.
21. Mallat, S., 1989: Multifrequency channel decompositions of images
and wavelet models. IEEE Trans, on Acoustics, Speech, and Signal
Processing, 3 7 , No. 12, 2091-2110.
22. Mallat, S. and S. Zhong, 1992: Characterization of Signals from Multi-
scale Edges. IEEE Trans, on Pattern Anal, and Mach. Intel., 14, no.7,
710-732.
23. Mallat, S. and W. L. Hwang, 1992: Singularity detection and process-
ing with wavelets. IEEE Trans, on Inform. Theory, 3 8 , no.2, 617-643.
24. Schols, J. L. J., 1984: T h e detection and measurement of turbulent
structures in the atmospheric boundary layer. Boundary-Layer Mete-

orology, 2 9 , 39-58.
25. Taubenheim, J. 1989: An easy procedure for detecting a discontinuity
in a digital time series. Zeitschrift Meteorol., 3 9 , 344-347.
26. Tennekes, H. and J. L. Lumley, 1972: A first course in turbulence, T h e
M I T Press, Cambridge, Mass., 300 pp.
27. Yamada, M. and K. Ohkitani, 1990: Othonormal wavelet expansion
and its application to turbulence. Progress in Theoretical Physics, 8 3 ,
No. 5, 819-823.
28. Yamada, M. and K. Ohkitani, 1991: Orthonormal wavelet analysis of
turbulence. Fluid Dynamics Research, 8, 101-115.
This research was supported by the Advanced Study P r o g r a m of the Na-

tional Center for Atmospheric Research sponsored by t h e National Science
Foundation (CH), the Airforce Office of Scientific Research, Directorate of
Chemical and Atmospheric Sciences grant F49620-92-J-0137 (NG), and
NASA grant NAG5-904 (NG).
Carl R. Hagelberg
National Center for Atmospheric Research
P.O. Box 3000
Boulder, Colorado 80307
e-mail: [email protected]
Nimal K. K. Gamage
P r o g r a m in Atmospheric and Oceanic Sciences
Astrophysics, Planetary and Atmospheric Sciences Department
University of Colorado
Campus Box 391
Boulder, Colorado 80309.
Intermittency in Atmospheric Surface Layer Turbulence: The
Orthonormal Wavelet Representation
Gabriel G. Katul, John D. Albertson, Chia R. Chu, and

Marc B. Parlange
Abstract. Orthonormal wavelet expansions are applied to atmospheric surface

layer velocity measurements that exhibited about three decades of inertial sub-
range energy spectrum. A direct relation between the n t h order structure function
and the wavelet coefficients is derived for intermittency investigations. This rela-
tion is used to analyze power-law deviations from the classical Kolmogrov theory
in the inertial subrange. The local nature of the orthonormal wavelet transform in
physical space aided the identification of events contributing to inertial subrange
intermittency buildup. By suppressing these events, intermittency effects on the
statistical stucture of the inertial subrange is eliminated. The suppression of in-
termittency on the n t h order structure function is carried out via a conditional
wavelet sampling scheme. The conditional sampling scheme relies on an indicator
function that identifies the contribution of large dissipation events in the wavelet
space-scale domain. The conditioned wavelet statistics reproduce the Kolmogrov
scaling in the inertial subrange and resulted in a zero intermittency factor. A
relation between Kolmogrov's theory and Gaussian statistics is also investigated.
Intermittency resulted in non-Gaussian statistics for the inertial subrange scales.
§1. Introduction
According to the Kolmogorov theory [25] (hereafter referred to as K41),
the ensemble average of the n t h order velocity difference (Δΐί;) between two
points separated by spatial distance (r), in the inertial subrange, for high
Reynolds number flow is given by
<|Δ«,·η = ΛΓ η «6»*Γ* (1)
En* Foufoula-Georgiern and Praveen Kumar (eds.), pp. 81-105.
ISBN 0-12-262850-0
82 G. Katul et al.
where e is the turbulent energy dissipation rate
and Ui are the velocity components (i — 1,2,3), n is the order of the struc-
ture function, v is the kinematic viscosity, Kn is a universal constant inde-
pendent of the flow but dependent on n, r is the separation distance that
is much smaller than the energy containing length scale or integral length
scale (L) but much larger than the Kolmogorov microscale η(= [^/3/(e)]1/4),
and (·) is the ensemble averaging operator. The scaling laws of Equation (1)
have been confirmed by many experiments for n = 2 (e.g. the existence of
the 2/3 law for the structure function or —5/3 law for the power spectrum)
as originally proposed by Kolmogorov [25] and discussed in Monin and Ya-
glom ([42], pp. 453-527). However, the scaling laws in Equation (1) are
not accurate for n > 2, as evidenced by many other experiments (see e.g.
[1]). Deviations from these scaling laws have classically been attributed
to the intermittency in e. This intermittency results in an e(x) behavior
that resembles an on-off process. That is, the dissipation of turbulent ki-
netic energy occurs only in a small fraction of the fluid volume. Hence, the
breakdown of Equation (1) is attributed to the inequality between (e n ) and
(e) n , as noted by Landau (see footnote in [29]; p. 126). As a result, many
phenomenological models for intermittency corrections to K41 have been
proposed. Example phenomenological models include the lognormal model
[26], the ß-model [17], and other multifractal models ([40], [39] and [2]).
Many atmospheric surface layer (ASL) flows exhibit an inertial sub-
range that extends over many decades so that intermittency effects on
inertial subrange scaling becomes important (see [22]; [42], Ch.8). Refined
intermittency studies in the natural environment encounter difficulties due
to: 1) the limited sampling period over which steady state mean mete-
orological conditions exist, 2) the need for instrumentation that is free
of atmospheric contamination and possible calibration drifts due to tem-
perature and humidity changes, 3) the need for instrumentation that is
field robust and capable of providing all three velocity components (since
changes in wind direction are inevitable), and 4) the need for turbulence
conditions that allow the application of Taylor's frozen hypothesis with
minimum wavenumber distortion.
The first difficulty severely limits the number of data points that can
be used to evaluate the ensemble average in Equation (1). Typically, the
ergodic hypothesis is used to evaluate the ensemble average in Equation (1)
from the measured time averages. The convergence of the time average and
the ensemble average requires a very large number of measurements that
may not be available in many field studies due to unsteadiness in the mean
meteorological conditions. The second and third difficulties limit the use
Atmospheric Surface Layer Turbulence 83
of many laboratory fast response sensors such as hot wire probes t h a t are
capable of resolving scales as small as t h e dissipation scales b u t difficult
to operate for long periods in the natural environment. We note here t h a t
some success in using such instruments (e.g. triaxial hot wire probes) for
extended periods were reported [43]. However, hot wire probes are not
suited for ASL measurements in arid and semi-arid environments since t h e
air t e m p e r a t u r e fluctuation can be very large (up to 6°C in seconds).
T h e development of analyzing tools t h a t allow the study of intermit-
tency effects in t h e ASL from limited number of field measurements is
necessary. T h e purpose of this paper is to investigate t h e usefulness of or-
thonormal wavelet transforms to quantify intermittency effects on inertial
subrange scaling from ASL velocity measurements. For this purpose, we
develop a conditional sampling scheme t h a t is capable of identifying dissi-
pation events in the space-scale wavelet domain. This conditional sampling
scheme can identify the location of large dissipation events t h a t contribute
to inertial subrange intermittency.
T h e wavelet transform is applied to 56 Hz triaxial sonic anemometer
velocity measurements in the ASL t h a t exhibit an inertial subrange for
three decades. Since intermittency investigations typically utilize Fourier
power spectra and structure functions, we first establish a relation between
the wavelet coefficients and these statistical measures. Then, we introduce
conditional wavelet statistics t h a t are developed to isolate events responsi-
ble for deviations from K41. However, before we discuss these approaches,
we offer a brief review of wavelet transforms with emphasis on applications
to turbulence measurements.
§2. A n a l y s i s of T u r b u l e n c e using W a v e l e t T r a n s f o r m s
Wavelet transforms are recent mathematical tools t h a t can unfold tur-
bulence signals into space and scale [15]. Continuous wavelet transforms
have been applied to many turbulence measurements and proved t o be
successful in identifying local scaling exponents ([4], [14] and [3]), inter-
mittency visualization [30], and coherent motion in ASL flows ([10], [19],
[18] and [33]). Orthonormal wavelets are a discrete form of t h e continu-
ous wavelets; however, they have the added feature of forming a complete
basis with t h e analyzing wavelet functions orthogonal to their translates
([41]; [12]; [9], Chapter 1). T h e application of orthonormal wavelets has
yielded i m p o r t a n t new techniques in the study of turbulence ([53], [54],
[55], [37], [38] and [23]). For completeness, a brief review of continuous and
orthonormal wavelet transforms is given.
Analogous to Fourier transforms, wavelet transforms can be classified
as either continuous or discrete. T h e continuous wavelet transform is first
introduced followed by a motivation for using discrete wavelet transform.
84 G. Katul et al.
2.1. Continuous wavelet transforms

As shown by Grossmann et al. [20], the continuous wavelet transform
W(b, a) of a real square integrable signal f(x) (i.e. the integral of [/(#)] 2
from — oo to +oo is finite) with respect to a real integrable analyzing wavelet
φ(χ) (i.e. the integral of ψ(χ) from —oo to +oo is finite) can be defined as
W(b,a) = C;*-r / φ( )f(x)dx (3)

Va J-oo a
where a is a scale dilation, b is a position translation, and Cg is defined by

+oo
\Κ\~ι\φ*(Κ)\2άΚ <oo (4)
/ -OO
where K is the wavenumber and φ* is the Fourier transform of ψ(χ) given

by
1-oo
tP(x)e'ikxdx. (5)
/ -OO
T h e condition in Equation (4) ensures the locality of Cg in the Fourier

domain. T h e continuous wavelet transform is commonly viewed as a nu-
merical microscope whose optics, magnification, and position are given by
ψ(χ), α, and 6, respectively ([30] and [10]). T h e function ψ(χ) has to satisfy
the following conditions:
1. T h e admissibility condition, which requires t h a t
f
'—OO
r+oo
rjj(y)dy = 0. (6)
Simply stated, Equation (6) requires t h a t the average of φ(χ) be

zero.
2. T h e invertability condition, which requires at least one reconstruc-
tion formula for recovering the signal exactly from its wavelet coef-
ficients (see [15]).
The function f(x) can be retrieved from the wavelet coefficients by
W N _ i f+°° /*+°° i ,,x-b.TXT, ,.dbda

/ ( * ) = Cg 2 / / α-?ψ( a
)W(a,b) — aj - . (7)
J0 J-oo
Further details regarding the continuous wavelet transform theory can be
found in many references (e.g. [13], [11], [9] and [15]).
2.2. Orthonormal wavelet expansions

For actual turbulence measurements, discrete wavelet transforms are
preferred since f(x) is known at only discrete points Xj (whose spacing de-
pends on t h e resolution of the sensor and the sampling frequency). There-

fore, for discrete measurements, it is necessary to discretize the scale (a)
and space (b) domains of Equation (3). If f(xj) is defined by TV points, one
may consider simply discretizing the space domain of (3) by TV nodes and
the scale domain of (3) by N nodes (i.e. discretized by a series of dirac-
delta functions). In this manner, the wavelet transform of f(xj) requires
iV2 wavelet coefficients. Hence, the discretization discussed above forms an
over-complete description of f(xj) in the wavelet domain. Because of the
over-complete description of f(xj), some redundant information is injected
into t h e wavelet transform of f(xj). This redundant information may or
may not be advantageous depending on the application. If statistical anal-
ysis is to be performed on the wavelet coefficients, then the redundant
information can produce artificial correlations t h a t are a property of the
wavelet transform and not of turbulence.
As shown by Yamada and Ohkitani ([53], [54] and [55]), for space-
scale statistical relations, it is recommended t h a t the discretization of the
space and scale domains form a complete basis so t h a t N measurements are
characterized by N wavelet coefficients. Orthonormal wavelet transforms
are suited for this purpose since the basis function are orthogonal and the
m u t u a l independence of the wavelet coefficients is guaranteed.
Daubechies ([11]; [12], pp. 10), Mallat ([34] and [35]) and Meyer [41]
demonstrated t h a t using a logarithmic uniform spacing for the scale dis-
cretization with increasingly coarser spatial resolution at larger scale, a
complete orthogonal wavelet basis can be constructed. We choose t h e Haar
wavelet basis for its differencing characteristics, since we are interested in
developing explicit relations between the n t h order structure function and
the wavelet coefficients. Other important features of the Haar wavelet basis
are discussed in [33].
T h e Haar basis φ(χ) = ( α _ 1 / 2 ) ^ ( ( χ - 6 ) / α ) , where a = 2 m and b = 2mi
for i, m G Z, is given by
(8)
For this basis function, the wavelet coefficients WT^m+l\k) and t h e coarse
grained signal 5 ^ m + 1 ^ (k) at scale ra + 1 can be determined from t h e signal
S^m) at scale m by using
(9)
(10)
for m = 0 t o M - 1, i = 0 to 2M-m~l - 1, and M = log 2 (iV), N is the

86 G. Katul et al.
number of samples (integer power of 2) (see [5] and [6]). For t h e Haar
wavelet, the coarse grained signal defined by Equation (10) is a low-pass
filtered function obtained by a simple block moving average, while t h e
wavelet coefficients are computed from the high-pass filter in Equation (9).
T h e wavelet coefficients and coarse grained signal may be calculated by t h e
following pyramidal algorithm:
1. Beginning with m = 0, use Equations (9) and (10) to calculate t h e

signal S^ and the coefficients WT^ at the first scale by looping
over i from 0 to 2 M _ 1 — 1. This will result in S and WT vectors of
length N/2. The turbulence measurements are stored in S^.
2. Repeat step 1 with ra = 1 to calculate the next coarser scale's pair

of vectors 5 ( 2 ) and WT™ (each of length TV/4).
3. Repeat for larger scale m up to M — 1 to produce a series of S

and WT vectors of progressively decreasing length. Note t h a t at
m = M — 1 the coarse grained signal converges to a point.
This algorithm yields N—1 wavelet coefficients t h a t define the orthonormal

Haar wavelet transform of the measured turbulence signal. T h e above pyra-
midal procedure, which is the basis for Fast Wavelet Transforms ( F W T ) ,
requires about N computations vis-a-vis the N log 2 N computations for
Fast Fourier Transforms ( F F T ) . The set N — 1 discrete Haar wavelet coef-
ficients satisfies the conservation of energy condition
M _ m
JV-l M 2 -l
2
Σ/ω = Σ Σ wT^\i). (H)
Equation (11) states t h a t the sum of the square of the wavelet coefficients
conserves the norm of the signal and is similar to Parseval's identity in
Fourier expansions ([9], pp. 12).
§3. E x p e r i m e n t
T h e d a t a presented here were collected during an experiment carried
out on J u n e 27, 1993 over a uniform dry lakebed (Owens lake) in Owens
valley, California. The lakebed is contained in a large basin bounded by
the Sierra Nevada range to the east and the White and Inyo Mountains
to the west. T h e instrumentation site is located on the northeast end
of the lakebed (elevation=l,100 m ) . The site's surface is a heaved sand
soil extending uniformly 11 km in the North-South direction and 4 km in
the East-West direction. T h e three velocity components were measured
at z = 2.5 m above the surface using a triaxial ultrasonic anemometer
(Gill Instruments/1012R2) to an accuracy of ± 1 % . Sonic anemometers
Table I
Summary of energy, meteorological, turbulence, and surface roughness
conditions during the experiment. T h e net radiation was measured by
a Q6 Fritshen type net radiometer, the soil heat flux was measured by
two R E B S soil heat flux plates, the sensible heat flux was measured by
a Campbell Scientific eddy correlation system (sampling at 10 Hz), and
t h e friction velocity was measured by a triaxial sonic anemometer. T h e
m o m e n t u m roughness length was determined from 6 near-neutral runs
of (U) and it* using the logarithmic velocity profile.
Energy Conditions
Net radiation (Rn) 173 W m " 2
Soil Heat Flux (G) 78 W m " 2
Sensible Heat Flux (H) 90 W m " 2
Meteorological Conditions
Mean Horizontal Wind Speed ((U)) 2.68 m s " 1
Mean Air Temperature (Ta) 31.6°C
Turbulence Conditions
Friction Velocity (u+) 0.165 m s - 1
Root-Mean Square Velocity (au) 0.516 m s " 1
A t m o s p h e r i c Stability Conditions
Height above ground surface (z) 2.5 m
Obukhov Length (L) -3.98 m
Surface R o u g h n e s s
Momentum Roughness Length (z0) 0.13 mm
achieve their frequency response by sensing t h e effect of wind on t h e tran-

sit times of sound pulses traveling in opposite directions across a known
p a t h length dsi(= 0.149 m in this study). The sonic anemometer is well
suited for these experiments since it is free of calibration nonlinearities,
atmospheric contamination drifts, temperature effects, and time lag. T h e
main disadvantage of sonic anemometers is the wavenumber distortion due
to averaging over dsi. This distortion is restricted to wavenumbers in excess
of 2n/dsi(= 42.2 m - 1 ) as discussed in [51] and [16].
T h e sampling frequency fs was 56 Hz and t h e sampling period Tp was
9.75 minutes. T h e short sampling period was necessary to achieve steady
state conditions in the mean meteorological conditions. Recall also t h a t a
long sampling period may not b e necessary in this case since our intent is
to resolve interial subrange scales t h a t are much smaller t h a n the integral
length scale. T h e sampling at fs = 56 Hz for Tp = 9.75 minutes resulted in
N = 32,768 points which are displayed in Figure 1. Notice in Figure 1 t h a t
the velocity fluctuated by as much as 3 m s " 1 in a few seconds. A s u m m a r y
of t h e mean meteorological and turbulence conditions is presented in Table
1.
From Table 1, the ratio of the root mean square velocity au(= ((U —
88 G. Katul et al.
0 60 120 180 240 300 360 420 480 540 600

Time (s)
F i g u r e 1. Time variation of the measured longitudinal velocity (U)

at z = 2.5 m above the ground surface. The sampling frequency and
sampling period are 56 Hz and 9.75 minutes, respectively.
( t / ) ) 2 ) 1 / 2 ) to the mean horizontal wind speed (U) is 0.52/2.68 = 0.194

which is smaller t h a n 0.5. Hence, Taylor's hypothesis [48] can be used t o
convert time increments to space increments without significant distortion,
at least for the inertial subrange scales (e.g. [31], [44], [50] and [45]). T h e
resolvable wavenumber Κ^γ(= 2π/((U)(f s /2)~ 1 )) by the sonic anemome-
ter corresponding to the Nyquist frequency^ ( / s / 2 ) is 65.2 m - 1 . We note
t h a t this wavenumber is larger t h a n 42.2 m _ 1 , and therefore, we restrict
our statistical analysis to wavenumbers smaller t h a n 42.2 m _ 1 b u t display
the spectra for all measured wavenumbers.
§4. W a v e l e t S t a t i s t i c s
In this section, we first develop relations between the Haar wavelet co-
efficients and the Fourier power spectrum and the structure function, then
we discuss hdw wavelet transforms can be used to investigate intermit tency
effects on the inertial subrange.
4 . 1 . R e l a t i o n b e t w e e n w a v e l e t coefficients a n d Fourier p o w e r
spectrum
In Fourier analysis, the fundamental tool used to characterize turbu-
lence is the power spectral density function E(K). T h e function E(K)
represents t h e energy density contained in each wavenumber b a n d dK, and
thus provides information regarding the importance of each scale of mo-
tion. However, important spatial information regarding location of events
become implicit in the phase angle of Fourier transform. In this section,
we relate t h e Haar wavelet coefficients t o t h e Fourier power spectrum and
show how spatial information can be expressed in an explicit manner.
T h e variance of the turbulence measurement, in terms of t h e wavelet
coefficients, can be deduced from the conservation of energy
M m
M 2 - -l
2 λ
a = Ν~ Σ Σ {WT^\i]f. (12)
m—l i=0
T h e total energy TE contained in scale Rm(= 2m dy) can be computed

from the sum of the squared wavelet coefficients at scale index (m) so t h a t
nM — m -i
TE = N~1 ]T (WT< m >[i]) 2 (13)

2= 0
where dy(= / S _1 (C^)) is the measurement spacing in physical space. In

order to compare the wavelet power spectrum to the Fourier spectrum, we
define a wavenumber A^m, corresponding to scale Rm, as
Km = £-. (14)
Hence, the power spectral density function E(Km) is computed by dividing
TE by t h e change in wavenumber AKm(= 2π2~ m dy~l ln[2]) so t h a t
E{Km) (15)
- 2^(2)
where (·) is averaging in space over all values of (i) for scale index (m)
(see [37] and [53]). The adequacy of Equations (14) and (15) are dis-
cussed in Katul ([24], pp. 149, 186-187). In Equation (15), the wavelet
power spectrum at wavenumber Km is directly proportional to t h e aver-
age of t h e square of the wavelet coefficients at t h a t scale. Because the
power at wavenumber Km is determined from averaging many squared
wavelet coefficients, we expect t h e wavelet power spectrum to b e smoother
t h a n its Fourier counterpart. This is apparent in Figure 2 which displays
good agreement between Fourier and wavelet power spectrum for all three
decades of inert ial subrange wavenumber s.
90 G. Katul et al.
Fourier
Wavelet
-5/3
10-1 100 101 103

Wavert umber (m-1)
F i g u r e 2. A comparison of the Fourier and Haar wavelet power spec-

tra. Taylor's hypothesis is used to convert time to wavenumber domain.
The —5/3 power law is also shown.
T h e Fourier power spectrum was computed by square windowing 8192

points, cosine tapering 5% on each window edge, and averaging t h e re-
sultant 4 power spectra (N = 32,768). The wavelet power spectrum was
computed by: 1) Using the pyramidal algorithm in Equations (9) and (10)
to obtain t h e wavelet coefficients over position index (i) and scale index
(ra), and 2) Using Equation (15) in conjunction with computed wavelet
coefficients from step 1 to obtain the wavelet power spectrum. Windowing
is unnecessary for the wavelet spectrum. We note t h a t similar compar-
isons were reported by Hudgins et al. [21] and Barcy et al. [4] using t h e
continuous wavelet transform.
Since t h e wavelet power spectrum is directly proportional to t h e av-
erage of the squared wavelet coefficients, we can also calculate the spatial
standard deviation using
SD K
^ ^ = d ^ 2 ) [ { W T ( m ) [ i ] 4 > - ((WT{m)w2))2}*- ( 16 )
In Meneveau ([37] and [38]), it was pointed out t h a t a plot of E(Km) and
E^m) + SDE gives a compact representation of the energy and its spatial
o
c
•22
mO
σ>
\_
4)
C
c
0)
0)
cSo
2 3 2 3 2 3 2 3
10-2 1 0- 1 1 00 101 102
Wovenumber (m- 1 )
F i g u r e 3. The evolution of the coefficient of energy variation CV(K)

as a function of wavenumber (K). The solid line is the best fit regression
line through the data points.
variability at each scale which is referred to as the "dual spectrum". How-

ever, as shown by Katul and Parlange [23], a better dimensionless indicator
for the spatial energy variance is given by the coefficient of variation CVE
defined as
<*««.> _i§gg=i. (1 „
An example of the variation of CVE is shown in Figure 3.
Notice in Figure 3 t h a t CVE increases as the wavenumber increases
indicating increased turbulent activity at smaller scales. T h e increase in
CVE can be related to the increased spottiness or intermittency in t h e
dissipation r a t e and is discussed next. In Tennekes and Lumley ([46] pp.
66), the dissipation rate for locally isotropic turbulence is given by
du
(e) = 15i/ (18)
dx
where u is the velocity fluctuation in the longitudinal direction, and x is

identical to x\. Recall t h a t deviations from K41 are due to the spottiness
or intermittency in e. Tennekes [47] suggested t h a t intermittency results
92 G. Katul et al.
in a large spatial variance in the dissipation rate ( c e ) 2 around its m e a n

value (e). Hence, we arrive at the desired physical meaning of CVE'- This
dimensionless ratio can be interpreted as (σ€/(β)) if the following arguments
are adopted:
1. For locally isotropic turbulence, the standard deviation σ€ around

(e) is given by
* j 0.5
2'
I / Γ<9ΐ/1
du \ / l"r)?/l
du \ I
(19)
dx dx
Notice here the similarity between Equations (16) and (19).
2. T h e derivatives in Equations (18) and (19) (e.g. du/dx) can be

approximated by differencing operations (e.g. Au/Ax).
3. Using step (2), the Ax in the ratio (cr e /(e)) cancels out from t h e
numerator and denominator and we are left with the differencing
operations.
4. T h e differencing (Au) at scale (m) can be directly related to the

wavelet coefficients using Equation (9). This point is the subject of
section 4.2.
Hence, the increase in CVE, as shown in Figure 3, is a direct indication

t h a t € is becoming more and more intermittent at smaller scales. W h a t
is also important to note here is t h a t this result cannot be inferred from
the usual Fourier power spectrum. From a turbulent energy point of view,
a key difference between wavelet and Fourier transforms is t h a t Fourier
transforms are nonlocal and therefore distribute the energy uniformly in
space (i.e. SDE — 0).
4 . 2 . R e l a t i o n b e t w e e n w a v e l e t coefficients a n d s t r u c t u r e f u n c t i o n
Using Equation (9) the wavelet coefficients can be related to the n t h
order structure function, for any flow variable 0, using
(\φ(χ + r) - φ(χ)\η) ~ ( l ^ f f n ) - (20)

In t h e above equation, we made use of the following: 1) the separation
distance r = 2 m dy, 2) the wavelet coefficients are proportional to φ(χ +
r) — φ(χ) at position x = (2mi) dy, 3) the amplitudes of the Haar wavelet
coefficients are proportional to ( 2 m ) x / 2 , and 4) (·) is the averaging operator
of the wavelet coefficients over all values of the position index (i) at scale
index (m). To study intermittency effects on Equation (1), we modify t h e
above relation and propose a conditional structure function to be discussed

next.
4 . 3 . C o n d i t i o n a l s a m p l i n g a n d i n t e r m i t t e n c y effects o n K 4 1
In general, intermittency of turbulent fluids is symbolized by an on-off
process in the dissipation r a t e so t h a t at a certain time, the turbulent energy
is only active in a small fraction of the fluid volume [36]. In a one dimen-
sional cut (namely along the x direction), the signature of intermittency is
large isolated dissipation events within an overall passive surrounding fluid
([36], pp. 104-109). For t h a t purpose, we classify the wavelet coefficients
as either "dissipative" or "passive". The dissipative wavelet coefficients are
the coefficients directly influenced by the large localized dissipation events
discussed above, while the passive coefficients are not. T h e distinction be-
tween dissipative and passive must be based on some minimum dissipation
threshold criterion. Such a criterion is difficult to construct without some
relation between the velocity and the dissipation. Before we establish t h e
conditioning criteria, let us first define the indicator function 1^ at scale
index (m) by
J(m) f 0 if [WT^(i)? > F([WT(™\i)]>)

\ 1 otherwise ^ '
where F is a conditioning criteria t h a t allows discrimination between the

dissipative and passive wavelet coefficients. Based on Equation (21), F
may b e interpreted as the ratio of the dissipation at scale index (m) and
position index (i) to the mean dissipation, if the following arguments are
adopted:
1. T h e dissipation at position index (i) and scale index (m) is directly

proportional to the square of the velocity gradient at t h a t position.
T h a t is e at scale index (m) is proportional to (du/dx)2 at scale
index (m). This statement may be inferred from Equation (18)
without the averaging operation. Possible sources of deviation from
this statement are presented in [47].
2. T h e mean dissipation at scale index (m) is directly proportional to

((du/dx)2) at scale index (m). This statement is a direct conse-
quence of Equation (18).
3. Arguments resulting in Equation (20) are all valid.
Arguments (1) and (2) may not be valid for turbulence t h a t is locally
anisotropic. However, based on the numerous studies of local isotropy in
ASL flows (e.g. [22]; [42], Ch.8), we adopt the working hypothesis t h a t tur-
bulence is locally isotropic in the inert ial subrange. These three arguments
94 G. Katul et al.
result in a sequence of equalities given by:
e'm> _ [Hr-) _ (u^(x + Ax)-u^m'>(x))2 _{WT^m\i)f

(c)(m) " /(au(^\2\ ~ ([u(m)(x + Ax)-u(™)(x)Y) ~ ((WT(™))2)
(22)
where e^ is t h e dissipation at position index (i) and scale ( m ) , (e)^m^ is
the mean dissipation at scale (ra), and [u^ (x + Ax) — u^m\x)]/Ax is t h e
finite difference approximation of t h e velocity gradient at scale (ra) and
position x. T h e first equality follows from arguments (1) and (2). T h e
second equality is due to t h e fact t h a t Ax required to convert gradients t o
differences (in a finite difference approximation of t h e derivative) cancels
out in t h e numerator and denominator, respectively. T h e third equality
is a direct consequence of Equation (20). For example, if F = 5, then all
squared wavelet coefficients t h a t are in excess of 5 times t h e average squared
wavelet coefficient at scale index (ra) are set to zero. This conditioning
criterion allows us to consider a conditional power spectrum Ec given by
E (Am) (23)
2^(2)
where (·) is now averaging in space over all non-zero values of [1^ WT^ (i)]2.
c
Hence, E represents t h e power spectrum of the passive fluid fraction. Also,
we can define t h e conditional n t h order structure function by
(\φ(χ + r) - φ(χ)\ψϊ ~ ((\l(m)WT(pr)) (24)
where ((·)) is averaging over all non-zero values of [1^ WT^ (i)]. These
conditional statistics can be computed by:
1) Using t h e pyramidal algorithm to calculate t h e Haar wavelet coef-
ficients at each scale index (m) and position index (i); 2) squaring these
coefficients to obtain t h e energy content at each scale index (m) and posi-
tion index (z); 3) averaging t h e squared wavelet coefficients for each scale
index (m); 4) dividing t h e squared wavelet coefficient (at space index i) by
the value computed in step (3); 5) If this ratio is larger t h a n some preset
value for F , then set this coefficient to zero, otherwise leave as is; 6) Use
Equation (23) or (24) t o determine t h e power spectrum or t h e n t h order
structure function with averaging performed over all non-zero values at
scale index (ra). Repeat t h e above steps for all values of (m) within t h e
inertial subrange. T h e adequacy of this conditional sampling criteria for
recovering K41 and characterizing intermittency is considered next.
10-1
F
h
^ 0 *
^10-2
>ô *
jyô*
l· ^yr y
l· ο ^ *'
>" '
- 9 ^ >'
3
Q s ^ * ' '
2 - ^s^ ** • Conditioned
o Unconditioned
10-3 s Slope = 0.68
: ^ô
"· Slope = 0.66
~S^ s
'"+ ! 1I , | | ! 1 1 1 11
2 3 4 2 3 4 2 3 4
10-1 1 00 10 1 102
Seporotion Distonce r ( m )
F i g u r e 4a. Comparison between the conditioned (F = 5) and uncon-

ditioned wavelet structure functions [D 2 (r)] for n = 2. The solid lines
are regression fits through the data points.
§5. R e s u l t s a n d D i s c u s s i o n
This section discusses the effects of intermittency on K41 using t h e
conditional wavelet analysis for three cases: 1) n = 2, 2) ra = 3, and 3)
n = 6. In each case, we check whether K41 is recovered when intermittency
is suppressed, and then we investigate the statistical structure of t h e events
responsible for deviations from K41 scaling. We do not present theoretical
details regarding intermittency models, b u t we focus more on the contrast
between the conditioned (intermittency suppressed) and unconditioned (in-
termittency active) statistics.
Case 1: n = 2
It is known t h a t intermittency effects are generally small and may
not be detectable for t h e structure function with n = 2 ([1] and [52]).
We test this hypothesis by comparing the unconditioned and conditioned
[F = 5) structure functions of Equation (24). T h e results are presented in
Figure 4aa.
Both conditioned and unconditioned second order structure functions
96 G. Katul et al.
exhibit scaling laws t h a t are in agreement with K41 (slope= 2/3) indicating
t h a t intermittency effects may not be significant for n = 2 (see [7] and [8]
for a possible physical explanation). We also present a summary of t h e
regression statistics for the regression model log[.D 2 (r)] = Alog[r] + B in
Table 2. Notice in Table 2 t h a t the coefficient of determination (R2) for t h e
regression model is in excess of 0.99; hence, the determination of scaling
laws from wavelet structure functions appears to be very reliable.
Table II
Summary of the regression statistics for the model
\og[(\AU\n}] = Alog[r] + B. The coefficient of
determination (R2) and the standard error of estimate
(SEE) are also shown. The conditioned statistics are
for the conditioning criterion F = 5.
n Slope Intercept #2 SEE Conditioned (C)/
(A) (B) Unconditioned (U)
2 0.680 -2.73 0.996 0.038 C
0.660 -2.55 0.996 0.036 u
3 1.010 -3.88 0.995 0.056 c
0.950 -3.53 0.995 0.055 u
6 2.008 -7.10 0.993 0.150 c
1.690 -5.76 0.987 0.170
u
One must note t h a t certain scale aliasing occurs due to the use of or-
thogonal and complete basis since the number of Haar modes characterizing
the frequency domain is relatively small (m = 15). This aliasing may in-
fluence the indicator function / ( m ) . Hence, intermittency characterization
by the indicator function can be overestimated or underestimated based
on the value of F. For t h a t purpose, we performed the same analysis for
F = 4 , 5 , 7 , and 10. The slope variation (for n = 2) did not differ by
more t h a n 0.004. This analysis clearly reveals the robustness of the pro-
posed conditional sampling scheme. Similar results were also obtained by
Yamada and Ohkitani [54].
Case 2: n = 3
As shown by Landau and Lifshitz ([29], pp. 123-128), a relation be-
tween t h e third order structure function and (r) is given by
|(AC/) 3 | = Î ( € ) r . (25)
T h e above relation was explicitly derived using the Navier-Stokes equations

for locally isotropic turbulence, and thus, is independent of any assumptions
implicit in K41 or any intermittency corrections to K41. If intermittency
does not alter the third order structure function behavior, our conditioned
iu-' E
:
2
10-2
^y^ x»
^y^ *
-D 2 4
y ^ \ '
S 10-3
2 *" • Conditioned
•
10-4 y^^S' o Unconditioned
Slope = 1.01
Slope = 0.95
I
^ A -X . _ ! , . . J. ι ι ι ι I
| , , —1 1—1 1 1 1 1 1 1 1 1 1 1 1 1
2 3 4 2 3 4 2 3 4
10-1 1 00 101 102
Seporotion Distance r (m)
Figure 4b. Same as Figure 4a but for n — 3.
and unconditioned statistics should b o t h reproduce the r 1 dependence of

Equation (25). Using Equations (20) and (24) with F = 5 and n = 3, we
compare the unconditioned and conditioned third order structure functions
in Figure 4b.
Both unconditioned and conditioned slopes (see Table 2) are in good
agreement with Landau and Lifshitz [29] predictions (which are also consis-
tent with K41). This analysis supports the insensitivity of the third order
structure function to intermittency corrections as noted by Anselmet et al.
[1]. Y a m a d a and Ohkitani [54] obtained similar results using an analogous
conditional analysis procedure, but a different wavelet.
Case 3: n = 6
T h e sixth order structure function can be related to the dissipation
correlation function from
((At/)6)
(e(x)e(x + r)) (26)
where t h e dissipation correlation function is given by
(e(x)e(x + r))~ (£) (27)

98 G. Katul et al.
Table III
Some values of the intermit tency parameter μ from various sources.
Source Measurements μ
Mahrt [32] Atmospheric mixed layer: 0.3-0.4
Convective and Nocturnal
Anselmet et ai [1] Turbulent and Duct flow 0.2 ± 0.05
Kuznetsov et al. [27] Wind-tunnel boundary layer 0.15-0.25
(Depends on external intermittency).
Monin and Yaglom [42] See Ch.8 for an extensive review 0.2-0.5
and μ is the intermittency parameter ([26], [17] and [1]). T h e value of μ

has been the subject of extensive research and its value appears to vary
between 0.15 and 0.5 (see Table 3). Thus, from Equations (26) and (27),
we see t h a t the sixth order structure function is related to μ by
((AU)6) ~ Γ2~μ. (28)
We now evaluate the performance of the conditional wavelet analysis

for reproducing K41 scaling and suppressing intermittency (i.e. μ = 0).
Using Equations (20) and (24) with F = 5 and n = 6, we compare t h e
unconditioned and conditioned sixth order wavelet structure function in
Figure 4c.
T h e slope of conditioned sixth order structure function is 2.0(= 2 — μ)
indicating t h a t intermittency is well suppressed (μ = 0) for the higher-order
statistics (and K41 is recovered). Recall t h a t the conditioning criteria is
based on second-order statistics (wavelet power spectrum), yet intermit-
tency was suppressed even in the sixth-order statistics.
T h e value of μ from the unconditioned sixth-order structure is 0.31 ( =
2 — 1.69), which agrees with many published values from laboratory and
atmospheric turbulence studies (see Table 3 for some examples). T h e con-
sistency in our value of μ with other laboratory experiments, in spite of
limited d a t a and sampling resolution, demonstrates the effectiveness and
robustness of orthonormal wavelet expansions for intermittency studies.
K41 and Non-Gaussian Statistics

Kraichnan [28] suggested t h a t K41 is consistent with the concept of
an inertial cascade if the velocity statistics within the inertial subrange do
not differ significantly from Gaussian. We therefore consider the velocity
statistics and their relation to K41 scaling. This relation is achieved by
noting t h a t conditional wavelet analysis recovers K41 and eliminates any
intermittency effects from the inertial subrange. Here Gaussian behavior
is tested by computing the conditioned and unconditioned wavelet flatness
factor for inertial subrange scales. T h e wavelet flatness factor (FF) at scale
10-1
10-2
10-3
10-4
**·
10-6
10-7 • Conditioned
o Unconditioned
- - Slope = 2.008
10-8
Slope = 1.69
10-9
2 3 4 3 4 3 4
io- 100 " ~ ' 10* 102
Seporotion Distance r (m)
F i g u r e 4c. Same as Figure 4a but for n = 6
index (m) is defined as
{(WT^\i])4)
FF(Rm) = (29)
((WTW\i})2)2
Using Equation (29) and the conditional wavelet analysis for F = 5, we

compute F F for all (m) corresponding to inertial subrange scales. T h e
results are summarized in Figure 4d.
Notice t h a t the conditioned FF(Rm) are nearly Gaussian (i.e. FF ~
3) for all inertial subrange scales while the unconditioned case is clearly
non-Gaussian with wavelet flatness factors up to 7. Also, note t h a t t h e
unconditioned flatness factor increases with decreasing separation distance.
This analysis clearly indicates t h a t K41 is associated with near-Gaussian
statistics and intermittency effects result in non-Gaussian statistics.
Finally, the wavelet flatness factor proposed in Equation (29) can also
be used t o infer t h e statistical properties of the horizontal velocity gradients
if the following arguments are adopted:
1. T h e differencing n a t u r e of the Haar wavelet transform, as can b e

noted from Equation (9), results in direct proportionality between
wavelet coefficients and horizontal velocity differences.
100 G. Katul et al.
• Conditioned
o Unconditioned
F i g u r e 4d. The wavelet flatness factor (FF) as a function of sepa-

ration distance (r) for inertial subrange scales. Both conditioned and
unconditioned cases are shown.
2. T h e dimensionless ratio in Equation (29) is the same for differences

and gradients since the division by the wavelet width or separation
distance t h a t is required to convert differences to gradients in nu-
merator and denominator cancels out. These two arguments are
due to M a h r t [33]. Using these two arguments, FF(Rm) in Equa-
tion (29) can be interpreted as a gradient flatness factor at scale Rm·
Notice in Figure 4d t h a t the unconditioned wavelet flatness factor
increases as the separation distance decreases. The flatness factor is
commonly used to measure the importance of the tails of the prob-
ability density function. Hence, the increase of FF in Figure 4d
indicates t h a t the tails of the horizontal velocity gradient probabil-
ity density function become more important as the scale decreases.
This analysis also indicates t h a t the increase in the gradient proba-
bility density tails, at smaller scales, is due to intermittency within
the inertial subrange (see Figure 4d). In contrast, no such increase
is noted for the passive wavelet coefficients.
Triaxial sonic anemometer velocity measurements at 2.5 m above a uni-
form dry lakebed (11 km by 4 km) were used to investigate intermittency in
t h e inertial subrange. T h e power spectrum of the horizontal velocity mea-
surements exhibited a —5/3 power law for three decades allowing detailed
investigation of scaling laws in the inertial subrange. In order to describe
space-scale relations in the inertial subrange, we utilized the orthonormal
wavelet representation. T h e orthonormal wavelet representation was well
suited for this investigation since the basis function are orthogonal and
m u t u a l independence of the expansion coefficients is guaranteed. In addi-
tion, it was shown t h a t the expansion coefficients can be related directly
to quantities commonly used in conventional turbulence analysis. Rela-
tions between t h e orthonormal wavelet coefficients and the Fourier power
spectrum, as well as relations with the nth order structure function were
established. A comparison between Fourier and wavelet power spectra was
also carried out. Good agreement between the two spectra was noted even-
though the Haar wavelet has poor localization in the frequency domain.
Since intermittency build up in the inertial subrange was due to local-
ized dissipative events, a conditional wavelet scheme was developed. T h e
conditional wavelet scheme relied on an indicator function t h a t identified
the wavelet coefficients directly influenced by large and localized dissipation
events. T h e conditional wavelet scheme efficiently suppressed intermittency
within the inertial subrange by removing these wavelet coefficients. K41
statistics, up to sixth order, were recovered when intermittency in t h e dis-
sipation was suppressed from the wavelet coefficients. It was also found
t h a t intermittency did not significantly influence second and third order
statistics in agreement with many other studies. T h e robustness of t h e
conditional wavelet scheme was also verified.
Finally, we demonstrated t h a t intermittency was directly responsible
for non-Gaussian statistics in the inertial subrange, while K41 was associ-
ated with near Gaussian statistics. T h e wavelet transform produced inter-
mittency factors comparable to values obtained from laboratory and other
field experiments.
References
1. Anseimet, F., Y. Gagne, E. J. Hopfinger, and R. A. Antonia, High-
order velocity structure functions in turbulent shear flows, J. Fluid
Mech., 140, 63-89, 1984.
2. Aurell, E., U. Frisch, J. Lutsko, and M. Vergassola, On t h e multifractal
properties of the energy dissipation derived from turbulence data, J.
Fluid Mech., 238, 467-486, 1992.
3. Argoul, F., A. Arneodo, G. Grasseau, Y. Gagne, E. J. Hopfinger, U.
102 G. Katul et al.
Frisch, Wavelet analysis of turbulence reveals the multifractal nature

of the Richardson cascade, Nature, 338, 51-53, 1989.
4. Barcy, E., A. Arneodo, U. Frisch, Y. Gagne, and E. Hopfinger, Wavelet
analysis of fully developed turbulence data and measurement of scaling
exponents, in Turbulence and Coherent Structures, ed. Metais, 0., and
M. Lesieur, Kluwer Academic Publishers, 450 pp., 1991.
5. Beylkin, G., R. Coifman, and V. Rokhlin, Fast wavelet transforms
and numerical algorithms I., Comm. Pure and Appl. Math., Vol.XLIV,
141-183, 1991.
6. Beylkin, G., R. Coifman, and V. Rokhlin, Wavelets in numerical anal-
ysis, Wavelets and their Applications, ed. Ruskai, M. B., Beylkin, G.,
Coifman, R., Daubechies, L, Mallat, S., Meyer, Y., Raphael, L., Jones
and Bartlett Publishers, Boston, 474 pp., 1992.
7. Chorin, A. J., Spectrum, dimension, and polymer analogies in fluid
turbulence, Phys. Rev. Lett., 60, 1947-1949, 1988.
8. Chorin, A. J., Constrained random walks and vortex filaments in tur-
bulence theory, Commun. Math. Phys., 132, 519-536, 1990.
9. Chui, C. K., An Introduction to Wavelets, Academic Press, Inc., 264
pp., 1992.
10. Collineau, S., and Y. Brunet, Detection of turbulent coherent motion
in a forested canopy, part I: Wavelet analysis, Bound. Layer Meteor.,
65, 357-379, 1993.
11. Daubechies, I., Orthonormal bases of Compactly supported wavelets,
Comm. Pure and Appl. Math., Vol.XLI, 909-996, 1988.
12. Daubechies, I., Ten Lectures on Wavelets, CBMS-NSF Regional con-
ference series in applied mathematics, S.I.A.M., 61, 357 pp., 1992.
13. David, Wavelets and Singular Integrals on Curves and Surfaces,
Springer-Verlag, 109 pp., 1992.
14. Everson, R., L. Sirovich, K. R. Sreenivasan, Wavelet analysis of the
turbulent jet, Physical Letters A, 145, 314-322, 1990.
16. Friehe, C. A., Fine-scale measurements of velocity, temperature, and
humidity in the atmospheric surface layer, Probing the Atmospheric
Boundary Layer, ed. Lenschow, D., American Meteorological Society,
269 pp., 1986.
17. Frisch, U., P. L Sulem, and M. Nelkin, A simple dynamical model of
intermittent fully developed turbulence, J. Fluid Mech., 87, 719-736,
1978.
18. Gamage, N., and C. Hagelberg, Detection and analysis of microfronts
and associated coherent events using localized transforms, J. Atmos.
Sei., 50, 750-756, 1993.
19. Gao, W., and B. L. Li, Wavelet analysis of coherent structures at the
atmosphere-forest interface, J. Appl. Meteor., 32, 1717-1725, 1993.

20. Grossmann, A., R. Kronland-Martinet, and J. Morlet,Reading and un-
derstanding continuous wavelet transforms, Wavelets: Time-Frequency
Methods and Phase Space, ed. Combes, J. M., Grossmann, A.,
Tchamitchian, Ph., Springer-Verlag, 315 pp., 1989.
21. Hudgins, L. H., M. E. Mayer, and C. A. Friehe, Fourier and wavelet
analysis of atmospheric turbulence, Paper presented at the Wavelets
and Applications Conference, Toulouse, France, June 8-13, 1992.
22. Kaimal, J. C., J. C. Wyngaard, Y. Izumi, O. R. Cote, Spectral char-
acteristics of surface layer turbulence, Quart. J .R . Met. Soc, 98,
563-589, 1972.
23. Katul, G. G., and M. B. Parlange, On the active role of temperature
in surface layer turbulence, J. Atmos. Sei., 1993,In press.
24. Katul, G. G., Coupled Processes Near the Land-Atmosphere Interface,
Ph.D Dissertation, University of California at Davis, 193 pp., 1993.
25. Kolmogorov, A. N., The local structure of turbulence in incompressible
viscous fluid for very large Reynolds numbers, Dokl. Akad. Nauk SSSR,
4, 299-303, 1941.
26. Kolmogorov, A. N., A refinement of previous hypotheses concerning
the local structure of turbulence in a viscous incompressible fluid at
high Reynolds number, J. Fluid Mech., 13, 82-85, 1962.
27. Kuznetsov, V. R., A. A. Praskovsky, and V. A. Sabelnikov, Fine-scale
turbulence structure of intermittent shear flows, J. Fluid Mech., 243,
595-622. 1992.
28. Kraichnan, R., Turbulent cascade and intermittency growth, Turbu-
lence and Stochastic Processes: Kolmogorov's Ideas 50 Years On, ed.
Hunt, J., Phillips, M., and Williams, D., Roy. Soc, 240 pp., 1991.
29. Landau, L. D., and E. M. Lifshitz, Fluid Mechanics, Pergamon Press,
536 pp., 1986.
30. Liandrat, J., and F. Moret-Baily, The Wavelet Transform: some ap-
plications to fluid dynamics and turbulence, European Journal of Me-
chanics, B/ Fluids, 9, 1-19, 1990.
31. Lumley, J., Interpretation of time spectra measured in high intensity
shear flows, Phys. Fluids, 6, 1056-1062, 1965.
32. Mahrt, L., Intermittency of atmospheric turbulence, J. Atmos. Sei.,
46, 79-95, 1989.
33. Mahrt, L., Eddy asymmetry in the shear-heated boundary layer, J.
Atmos. Sei., 48, 472-492, 1991.
34. Mallat, S, A theory for multiresolution signal decomposition: The
wavelet representation, IEEE Trans. Pattern Analysis and Machine
Intelligence, 11, 674-693, 1989a.
35. Mallat, S., Multiresolution approximations and wavelet orthonormal
bases of L2(R), Trans. Amer. Math. Soc, 315, 69-87, 1989b.
104 G. Katul et al.
36. McComb, W. D., The Physics of Fluid Turbulence, Oxford Science

Publications, 572 pp., 1992.
37. Meneveau, C , Analysis of turbulence in the orthonormal wavelet rep-
resentation, J. Fluid Mech., 232, 469-520, 1991a.
38. Meneveau, C , Dual spectra and mixed energy cascade of turbulence
in the wavelet representation, Physical Review Letters, 11, 1450-1453,
1991b.
39. Meneveau, C , and K. R. Sreenivasan, The multifractal nature of tur-
bulent energy dissipation, J. Fluid Mech., 224, 429-484, 1991.
40. Meneveau, C , and K. R. Sreenivasan, Simple multifractal cascade
model for fully developed turbulence, Phys. Rev. Lett., 59, 1427-1427,
1987.
41. Meyer, Y., Orthonormal Wavelets, Wavelets: Time-Frequency Methods
and Phase Space, ed. Combes, J.M., Grossmann, A., Tchamitchian,
Ph., Springer-Verlag, 315 pp., 1989.
42. Monin, A. S., and A. M. Yaglom, Statistical Fluid Mechanics Vol.11,
ed. Lumley, J., MIT Press, 874 pp., 1975.
43. Oncley, S. P., TKE dissipation measurements during the FLATS exper-
iment, Proc. 10th AMS Symp. on Turbulence and Diffusion, Portland,
Oregon, pp. 165-166, 1992.
44. Powell, D., and C. E. Elderkin, An investigation of the application
of Taylor's hypothesis to atmospheric boundary layer turbulence, J.
Atmos. Sei., 31, 990-1002, 1974.
45. Stull, R., An Introduction to Boundary Layer Meteorology, Kluwer Aca-
demic Press, 666 pp., 1988.
46. Tennekes, H., and J. L. Lumley, A First Course in Turbulence, MIT
Press, 300 pp., 1990.
47. Tennekes, H., Intermittency of the small-scale structure of atmospheric
turbulence, Bound. Layer Meteorol., 4, 241-250, 1973.
48. Taylor, G. I., The spectrum of turbulence, Proc. Roy. Soc, A, Vol.
CLXIV, 476-490, 1938.
49. Willis, G. E., and J. Deordorff, On the use of Taylor's translation
hypothesis for diffusion in the mixed layer, Quart. J. Roy. Meteor.
Soc, 102, 817-822, 1976.
50. Wyngaard, J. C , and S. F. Clifford, Taylor's hypothesis and high-
frequency turbulence spectra, J. Atmos. Sei., 34, 922-929, 1977.
51. Wyngaard, J. C , Cup, propeller, vane, and sonic anemometer in tur-
bulence research, Ann. Rev. Fluid Mech., 13, 399-423, 1981.
52. Yakhot, V., Z. S. She, and S. A Orzag, Deviations from the classical
Kolmogorov theory of the inertial range of homogeneous turbulence,
Phys. Fluids A, 289-293, 1989.
53. Yamada, M., and K. Ohkitani, Orthonormal expansion and its applica-
tion to turbulence, Prog. Theor. Phys.: Progress Letters, 86, 819-823,
1990.
54. Yamada, M., and K. Ohkitani, Orthonormal wavelet analysis of turbu-
lence, Fluid Dynamics Research, 8, 101-115, 1991a.
55. Yamada, M., and K. Ohkitani, An identification of energy cascade in
turbulence by orthonormal wavelet analysis, Prog. Theor. Phys., 86,
799-815, 1991b.
The authors would like to thank Scott Tyler for his assistance and support
at Owen's lake, and Teresa Ortenburger and Mike Mata for their help in the
data collection. We are grateful for the funding support from the National
Science Foundation (NSF) grant (EAR-93-04331), United States Geological
Survey (USGS), Water Resources Center (WRC) grant (W-812), Kearney
Foundation, and UCDAVIS superfund grant (5 P42ES04699-07).
Gabriel G. Katul
School of the Environment
Duke University
Durham, NC 27708-0328
USA
John D. Albertson
Hydrologie Science
University of California at Davis
Davis, CA 95616
USA
Chia R. Chu
Department of Civil Engineering
National Central University
Chungli, Taiwan
Marc B. Parlange
Hydrologie Science &
Department of Agricultural and Biological Engineering
University of California at Davis
Davis, CA 95616
USA
An Adaptive Decomposition:
Application to Turbulence
J. F. Howell and L. Mahrt
Abstract. Nine hours of 45 meter tower anemometer measurements are an-

alyzed to demonstrate an adaptive method for decomposing a time series into
orthogonal modes of variation. In conventional partitioning (or filtering) the cut-
off scales are specified a priori to be constant throughout the record. Applying
a constant cutoff scale is less effective if two different physical modes vary on
overlapping scales, since the statistical partitioning is then physically ambiguous.
For the turbulence data analyzed in this study, motions leading to a majority of
the momentum flux intermittently occur on small scales which otherwise lead to
little flux. To better separate the transporting motions from the more random
motions, the cutoff scale separating these two modes is allowed to vary with record
position.
The entire record of data is partitioned into four modes of variation using
a piece-wise constant (Haar) decomposition. The two larger scale modes are
characterized as the mesoscale and large eddy modes. Similar to conventional
partitioning, these two modes are determined by spatially constant cutoff scales.
The two smaller scale modes on the other hand, are separated by a scale which
depends on the local transport characteristics of the flow. Local extremes in
the spatial distribution of momentum flux determine the partitioning between
the two small scale modes. This leads to an adaptive cutoff scale, which better
isolates the transport mode responsible for a majority of the momentum flux.
The spatial variation of this cutoff scale allows the time series to be decomposed
into modes which are physically more pure as is verified in terms of traditional
statistics for each mode. For example, the gradient skewness in the longitudinal
wind component for the main transporting eddies is —0.23 using a constant cutoff
scale and —0.85 using the adaptive cutoff scale indicating that the shear driven
transporting eddies are more completely isolated using the variable cutoff scale.
The remaining small scale deviations define the fine scale mode, which consist of
non-transporting nearly isotropic motions.

ISBN 0-12-262850-0
108 J. Howell and L. Mahrt
§1. I n t r o d u c t i o n
A common issue in time series analysis is sorting out the different modes
of variation. If such modes overlap in Fourier space or are primarily local
or event-like, then traditional Fourier and eigenvector decompositions are
generally less effective in separating the different modes. As an alternative,
nine hours of turbulence d a t a are orthogonally decomposed into piece-wise
constants. Based on the decomposition statistics, four modes of variation
are defined. One of these modes includes a majority of the m o m e n t u m
flux and is described by a subrange of scales which depends on the record
position.
T h e Haar wavelet is the underlying basis in the decomposition applied
in this study. Decomposing the turbulence in terms of higher order wavelets
which are more compact in Fourier space yield similar results. Higher order
wavelets however, are slightly less efficient in capturing the sharp gradients
associated with the transport physics which is the primary goal of this
study. On the other hand, higher order wavelets do appear b e t t e r suited
for representing the larger scale, smoother variations. Additional reasons
for specifically using the Haar wavelet are provided at the beginning of
Section 2.1, and results of applying alternative orthogonal wavelets are
discussed at the end of Section 2.3.
In general, wavelets decompose global variance or energy in terms of
scale and position within the record. Wavelets are used extensively in
seismic analysis and are beginning to find their way into other geophys-
ical disciplines [3, 11, 12, 13, 18, 21]. Related to this study, Farge [9]
and Meneveau [23] discuss a variety of wavelet applications to turbulence.
Ways of constructing, describing, and implementing wavelet tools are many.
Daubechies [8] constructs wavelets which are compact in physical space,
and yet still provide a complete basis for decomposing the total sampled
variance. Viewing wavelet bases from a linear algebra perspective [24] can
be useful in solving numerical equations [1]. Spline wavelets are gener-
ally effective for interpolating and filtering d a t a [2, 7, 25]. Another com-
mon view is t h a t a wavelet basis set describes a multiresolution analysis
[10, 19, 22].
A multiresolution analysis is a convenient setting for locally describing
the different scales of variation in the data. Multiresolution techniques are
used in image analysis, for example, to store and transmit images compactly
[5,6]. T h e current development could be posed in terms of a multiresolution
analysis [16] or alternatively in terms of the wavelet analysis referred to
above. However, the methods in this study (Section 2) are kept simplified
such t h a t it is unnecessary to explicitly appeal to these closely related
topics.
T h e strategy in this study is to first decompose the turbulence time
An Adaptive Decomposition: Application to Turbulence 109
series into distinct modes. Because small scale variations in the turbulence
record correspond to two types of motions, an adaptive technique is used
to separate t h e m (Section 2.2). For a simple demonstration of t h e adaptive
technique, only t h e scale separating these two small scale modes is allowed
to spatially vary. Generally, the adaptive technique allows for t h e separa-
tion of physically distinct modes with overlapping scales. This separation
is not possible in conventional filtering with a specified response function or
distribution of weights. After decomposing the turbulence, the spatial dis-
tribution of m o m e n t u m flux will be reconstructed for the different modes.
Additional statistics are then computed for the different modes (Section
2.3). In Section 3 a brief physical interpretation of the results is provided.
§2. P a r t i t i o n i n g t h e T i m e Series
T h e d a t a used in this study consist of 9.1 hours of the three velocity
components measured 45 meters above flat terrain in near neutral condi-
tions [17]. T h e wind speed fluctuates about a mean value of ü = 12.8 m / s
throughout the 9.1 hours, and most of the energy is concentrated in t h e
1 minute or 1 km eddy motions (Section 2.3). For a turbulence d e p t h of
roughly 500 m the Reynolds number is more t h a n 10 8 . Additional statistics
describing this d a t a set can be found in [17] and [21].
T h e value of the longitudinal wind component at the i t h record position
is denoted as
«,·; i = l,2,...,2M (1)
where δ = ^ s is the width of a sampling interval and in t h e current analysis
M = 19, corresponding to 524,288 d a t a points. There are equivalent time
series of t h e cross stream component v and the vertical wind component
w.
2.1. Decomposition method

H a a r [14] first presented this decomposition m e t h o d over 80 years ago,
yet t h e utility of such a basic decomposition is only beginning to be realized,
primarily in the context of wavelets. Indeed, Daubechies [8] appealed to
the Haar basis set in introducing a fundamental set of equations describing
a unique sequence of orthogonal wavelet bases.
Two coefficients (—1 and + 1 ) essentially describe a Haar basis element.
T h e H a a r decomposition involves multiplying differences of arithmetic sums
of t h e d a t a by dyadic numbers (= 2P where p is an integer). Orthogonal
wavelet transforms usually involve more coefficients and irrational numbers.
T h e H a a r wavelet is (odd) symmetric while all other compactly supported
orthogonal wavelets are apparently asymmetric [8]. T h e Haar basis set
is sometimes discounted because a Haar transform value corresponds to
a Fourier spectral window which decays away from a central wavelength
at a slower r a t e t h a n higher order wavelets, an important consideration

in many applications. For the turbulence d a t a analyzed in this study, t h e
main transporting eddies often consist of sharp boundaries or microfronts.
Microfronts are efficiently decomposed with the Haar basis set because
they b o t h contain sharp gradients. Compactness in Fourier space seems
less important in this case. The Haar decomposition is now reviewed.
T h e longitudinal wind averaged over the entire record is defined as
i i
π=
^Σ>- (2)
Variations in the longitudinal wind are described by difference terms de-
fined as
i 2 m -i
Au(am; n) = — Σ ( w 2™(n-i)+j ~ u2m{n_lHJ). (3)
These difference terms correspond to Haar transform values and character-

ize the average change in the u signal on the scale am = 2mS across the
interval [am(n — 1), amn\. Specifically, (3) is half the difference between the
half interval averages, so for example, Au(ai; n) = ^ x (u2n — ^ 2 n - i ) · Inter-
vals of length a m are enumerated starting with n = 1 (the first translate)
and ending with n = 2 M _ m (the last translate). These non-overlapping
translated intervals completely cover the time series of length aw The
differences (3) are then computed for each translate and dyadic scale such
t h a t the sampled variations are completely decomposed.
T h e signal deviation from the record average at the ith record position
is reconstructed as
M
Ui - ü = Σ ( - l ) * A u ( a m ; n) (4)
i — 1 i— 1
n = 1 + int(——) ; £ = 1 + int(- r)
for i — 1 , 2 , . . . , 2 M . The second translation number ί represents the enu-

merated half intervals of length am/2. The translation numbers n and ί
correspond to the difference terms (3) which are necessary for reconstruct-
ing the d a t a value at the z th record position. Specifically, n corresponds to
the interval [am(n — l ) , a m n ] which includes the ith record position. T h e
integer ί is even (odd) depending on whether the d a t u m at the i t h record
position was added (subtracted) in the associated difference term. T h e
operator ini(*) yields the integer part of the number *.
T h e total sampled variance is obtained by squaring the left hand side
of (4) and averaging over the 2 M different record positions. This operation
corresponds to a sum over dyadic scales of mean square Haar transform
values. The total sampled variance is written as

M 2 M _ m
2 M 1
5ME(^'" ÏÏ 2
) = Σ ^Μ^ Σ [Ati(am;n)]2. (5)
i=l m=l n=l
The total covariance between the time series of the longitudinal and vertical
wind components is obtained by multiplying the u reconstruction (4) by an
equivalent expression for w, and averaging over all record positions. The
resulting covariance quantity is written as
2
l
-JM Ys(Ui * U)(Wi ~W) = (6)
i=l
M Λ 2 M-m
Σ ^Μ^ Σ û(am;n)Aw(am;n).
2
7Ti=l n=l
The identities (4)-(6) can be verified by substituting in the expression

for ü from (2), the expression for Au(am;n) from (3), and equivalent ex-
pressions for w and Aw(am;n). Verifying (4)-(6) is facilitated by the two
identities:
M
Σ 2m = 2 M + 1 - 2 (7)
m=\
171
i i
2.2. Modes of variation

Variations contributing to the local signal deviation from the record
average (4) are now partitioned into different modes. The sum of the dif-
ferent modes will be equal to (4), and the individual modes will locally
correspond to a continuous subrange of scales. Accordingly, the longitudi-
nal wind data associated with a given mode c (= 1,..., C) at the i t h record
position is described by a range of scales between a - and a m + so that
Ui-u= Σ (-l)eAu(am;n) (9)
. /l'■ — 1 x „ . . j' - 1N
n = 1 + int(——)
v ; t = 1 + int(- r)
2m 2m
where raf = M, ra^ = 1 , and ra+ = m~_1 — 1. A given mode c at the
i t h record position includes all the difference terms (3) associated with the
range of dilation scales a - through a + · The small scale cutoff for the
smallest scale mode is a\ = 2J, so m ^ = 1 is fixed. This leaves C — 1 scales

unspecified. In this study C = 4, so the values of mj~, ra^~ and m^ are to
be determined.
If t h e value of m~ is set to a constant, then t h e scale separating t h e
two corresponding modes will be independent of record position. For the
atmospheric wind data, the cutoff scales for the two largest scale modes,
namely the mesoscale (c = 1) and large eddy (c = 2) modes, are defined in
this manner. These two modes are defined to span the scales from 800 m
- 420 km. Here we have converted time to distance (Taylor's hypothesis)
using the mean wind speed ü = 12.8 m / s .
Because the horizontal variance drops off rapidly with scale at about
5 km (Figure 5, Section 2.3), 5 km is chosen as the small scale cutoff
for the mesoscale mode c = 1. The smallest dyadic scale greater t h a n 5
km is «13 = 6.5 km, so accordingly, mj~ = 13. At the 500 m scale the
u-w covariance (momentum flux) increases rapidly with decreasing scale.
Therefore, 500 m is chosen as the small scale cutoff for the large eddy mode,
c = 2, in order t h a t most of the flux remains to be captured. T h e smallest
dyadic scale greater t h a n 500 m is a 1 0 = 800 m, so accordingly m^ = 10.
T h e large eddy mode consist of motions which may generally span the
entire depth of the boundary layer whereas the smaller scale ( < a 9 = 400
m) motions are likely to be not as deep.
Since there is no obvious constant scale separating the two smaller
scale modes, the scale (or m ^ · ) is allowed to depend on the record posi-
tion according to the local physics of the flow. This allows random small
scale fluctuations to be effectively separated from small scale variations
associated with a transporting eddy. In other words, the t r a n s p o r t mode
is allowed to locally capture smaller scales if such scales account for sig-
nificant flux. For example, significant m o m e n t u m flux may be associated
with wind gusts or so called microfronts. Including small scales in the re-
construction of the transport mode at the position of a microfront more
effectively captures the flux. Between the gusts, the small scale motions
do not significantly transport momentum and therefore are not included in
the reconstruction of the transport mode.
Local transport occurs at locations where variations in the longitudinal
wind component u are in phase with variations in the vertical component
w. T h e analyzed eddies are centered according to the largest magnitudes
of the product A.u(am;n) x A n ; ( a m ; n ) within each of the 4,096 intervals
of width 07 « 100 m. Specifically, flux maxima are determined within the
non-overlapping intervals [a-j{n — l),ajn] for n = 1,2,3, . . . , 2 1 2 . These lo-
cal m a x i m a statistically represent local maxima in the spatial distribution
of m o m e n t u m flux, and thus determine the position and scale of a trans-
porting eddy. In this study we do not use a minimum threshold value for
conditioning the local flux maxima. The strongest m o m e n t u m flux event
within each non-overlapping 100 m interval is included in the t r a n s p o r t

mode, regardless of the sign or strength of the event.
T h e values of mî are determined by first requiring 1 < mî < 7 for
all record positions. The upper bound on mî ensures t h a t t h e t r a n s p o r t
mode (c = 3) includes, as a minimum, all scales between αγ = 100 m and
α 9 = 400 m. This choice is based on the observed global dependence of
covariance on scale (Section 2.3).
T h e next step in determining mî is to identify the largest value of
Au(am; n) x Aw(am; n) within in each a-j « 100 m interval. Say t h a t within
a particular interval of length αγ, this largest product corresponds to the
index values m* and n* associated with the transform interval [am* (n* —
1), am* n*]. The transport mode is then reconstructed down to the scale am*
for all of the record positions contained in the interval [a m * (n* — 1), a m * n*].
This reconstruction corresponds to including all of the difference terms (3)
associated with the telescoping intervals [am'(nf — l ) , a m / n ' ] where a m * <
a
m' < a7 a n d the translation number is n' = 1 + int( ™>-L· )·
T h e difference terms corresponding to the telescoping intervals are used
in reconstructing the transport mode c = 3. These terms determine m^~z·.
Considering the i t h record position contained within the given interval of
length αγ, this position will necessarily be contained in the interval [a m / (η'—
l ) , a m / n ' ] for one or more values of m!. For example, this position (by
definition) is always contained in the interval corresponding to m' = 7.
T h e value of mî at this given ith record position is simply the smallest of
the ml'. This value corresponds to the smallest of the telescoping intervals
in which this given ith record position is contained. A more complete
description (with figures) of a similar adaptive technique can be found in
[16]·
As an example, a 4 km segment of the mc distributions are plotted
in Figure 1. T h e straight horizontal lines indicate the spatially constant
scales separating modes 1 & 2 and 2 L· 3. The scale separating modes 3
& 4 (mî) varies from several meters to 100 m depending on the record
position. T h e local minima in mî occur at locations where the smaller
scale variations are identified as belonging to the transport mode. At these
locations, some of the flux occurs on small scales usually associated with
microfronts.
2.2.1. Definitions
T h e particular modal definitions, detailed below, are based on an inter-
pretation of t h e variance and covariance spectra (Section 2.3). T h e modal
reconstructions (Section 2.2.2) substantiate the partitioning. T h e largest
scale (mesoscale) mode leads to very little flux and consist of predominantly
horizontal motions. T h e large eddy mode leads to some flux, and for this
1
I ' ' ' ' I ' ' ' ' I ' ' ' ' I ' ' ' ' I ' ' ' ' I ' ' ' ' I ' ' ' '
U=19
m o d e of v a r i a t i o n
100km
15 = 1 mesoscale
l
c = 2 large eddy
n° 10 1 km
c = 3 transporting eddy
10 m
c = A fine scale
m= 1 I . . . . I . . . . I . . . . I . . . . I
198 199 200 201 202

km ->
Figure 1. The small scale cutoffs for the four different physical modes
of variation
mode the variance in the longitudinal wind component is significantly larger

than the variance in the cross stream and vertical wind components. The
next smallest scale, the transport mode contains a majority of the flux,
and the motions are more 3-dimensional than the large eddy mode. The
smallest resolved (fine) scale mode is responsible for very little net flux and
consist of nearly isotropic motions.
(i) The mesoscale mode, c— 1, at the i t h record position is defined as
«l,· = Σ (-l)*Au(am;n) (10)

m=13
i—1 i—1
= 1 + int(——) ; t = 1 + int(- r)
where Au(am; n) is defined in (3). The dilation scale am = 2mS = 2 m _ 4 s

« 2 m + 3 x 0.1 m, so this mode includes scales between a 13 « 6.5 km and
ai9 « 420 km.
(ii) The large eddy mode, c = 2, at the ith record position is defined as
12
«2,i= ^2 {-l)eAu(am]n) (U)
m=10
where n and £ are defined in (10). This mode includes scales between
αιο « 800 m and a\2 « 3.3 km.
(iii) T h e transport mode, c = 3, at the ith record position is defined as
9
u
î= Σ (-l)'Ati(a m ;n) (12)
m=mä i
where 1 < ra«^· < 7 is determined for each interval of length αγ « 100
m according to t h e above discussion. This mode includes scales between
a\ « 1.6 m and a 9 « 400 m.
(iv) T h e fine scale mode, c = 4, at the z th record position is defined as
m
i,i
u
*,i = ] £ (~l/Au(a m ; n) (13)
where m ^ = mî — 1. At record positions where ra^· = 1 the fine scale

mode is zero. This mode includes scales between ai « 1.6 m and a 6 « 50
m.
2.2.2. R e c o n s t r u c t i o n s
T h e sum of the different modes is equal to (4), so it follows t h a t
c
Ui -ü =^2uCii (14)
c=l
where C = 4 is the total number of modes in this case. Figure 2 shows

the different orthogonal modes of the longitudinal wind. T h e sum of the
different modes including the record average 12.8 m / s is equivalent to t h e
original d a t a (Figure 3). The large differences in the scales of variation
for t h e different modes is made apparent by plotting t h e modes on t h e
same horizontal scale. The transport mode c = 3 consist of variable width
blocks because the scale separating the two small scale modes depends
on t h e record position. The local minima in the separation scale occur
at locations where the block widths are more narrow as can be seen by
comparing Figures 1 and 2.
Segment lengths along the horizontal axes are now adjusted according
to t h e different scales of variations in order for the flux uc x wc to be plotted
for different c (Figure 4). For example, the mesoscale mode, c = 1, can be
viewed for the entire ( « 420 km) record while the fine scale mode, c = 4,
is suitably viewed on a segment only 400 m in length.
Figure 4 shows t h a t the transport mode captures most of t h e down-
ward m o m e n t u m flux while the fine scale mode leads to small flux with
nearly equal probability of upward or downward flux. In this sense, t h e
198 199 200 201 202

km->
Figure 2. The four orthogonal modes of the longitudinal wind com-

ponent u plotted on the same segment of the record shown in Figure
1.
partitioning is successful. In order to quantify t h e differences between t h e

four modes, globally averaged statistics are now computed.
§3. V a r i a n c e a n d Covariance S p e c t r a
A global estimate of the variance on the scale a m is the square of
the Haar transform (3) on the scale am averaged over all the translation
positions, written as
ηΙΥΊ — TO
1
var[ (15)
where A u ( a m ; n ) is defined in (3). Prom (5) the total variance summed

over all scales can then be expressed as
M
Var(u) = 2_\ var(u;am) (16)
m=l
where Var(u) corresponds to the left hand side of (5).

In order to obtain a more continuous estimate of the variance as a func-
tion of scale, the dilation a is allowed to take on values between orthogonal
O Q I I I I I | I I I I | I I I I | I I I I | I I I I | I I I I | I I I I | I I I I
5 I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I
198 199 200 201 202

km->
Figure 3. The sum of the four different modes which equals the orig-
inal data.
scales, and the difference (transform) intervals are allowed to overlap. This
leads to a difference t e r m defined as
Au(a; b) = — ^2(ui+j - ui+j-m) (17)

i=i
where t h e dilation scale is a = 2m J, and b = iS is the central position of t h e
transform interval. Provided there are N equally spaced d a t a points, (17)
is defined for i = ra,..., iV — ra so there are a total of N — 2m + 1 transform
values on the scale a = 2m8. This generalization leads to a non-orthogonal
estimate of t h e variance on the scale a defined as
1 N-m
var a
^) = jv_2m+i ΣM«;6)]2· (18)
i=m
Depending on the sampling, the non-orthogonal and orthogonal global es-
timates of a local variance may be nearly equal at a fixed dyadic scale, t h a t
is var*(u; a) « var(u; am) for a = am. T h e non-orthogonal estimate based
on t h e spatially overlapping transform intervals leads to smoother spectral
estimates with respect to scale.
An orthogonal estimate for the covariance between t h e longitudinal
(a)
4
2
s
, , | „ M | , , U | , U , | , , : H | , , , , | , , I , | H , , | , _
w I i i i iI i i I i Ii I i iI i i i i I i i i i ii i i i i i i i i_
fcl
0 -u. /•iJ" ^'· 1
0 M||v|W"' *|H^Yl||^lj
-2
50km X
-4 IIIIMIMNII ■ I M I I I I M M I I M I I M MI
200 400 175 200 225

km-»
l l l l | M J I | M I 1 | l l | l | l | l l | l l l ll | l l l l | l l l l""l""l""l'
(c)
i" idi :
IT 4
V)
*- F M \ 2
£ °
-4 500m
X
s -2 50^m
/I ■ I ■ ■ ■ ■ I ■ ■ ■ ■ I ■ ■ t ■ I ■
3 -4
198 200 202 199.8 200 200.2
F i g u r e 4. The product uc X wc of the four orthogonal modes. The

four modes are (a) the mesoscale mode c = 1, (b) the large eddy mode
c = 2, (c) the transport mode c = 3, and (d) the fine scale mode c = 4.
and vertical wind components on the scale am is written as

cyM — m
1
j(ti,w;am)
COtM = 9M_m ]T] Au(am',n)Aw(am;n), (19)
n=l
so from (6) the total covariance can be alternatively expressed as

M
Cov(u,w) = 2_] cov(u,w;am) (20)
where Cov(u,w) corresponds to the left hand side of (6).

A non-orthogonal estimate of the covariance on the scale a is defined
as
N—m
cov*(u,w;a) = N _ 2 m + l Σ Au(a;6)Aw(a;6) (21)
i=m
which provides b e t t e r scale resolution for estimating the scales of motion
dominating t h e vertical flux.
T h e orthogonal and non-orthogonal covariance spectra are plotted for
the time series of the longitudinal and vertical wind components (Figures
Figure 5. Variance in the longitudinal wind component u. The cir-

cles are orthogonal values while the curve represents non-orthogonal
estimates. The regions between the vertical dashed lines are denoted
MS (mesoscale), LE (large eddy), TE (transporting eddy), and FS (fine
scale) and correspond to four orthogonal modes of variation. The num-
bers in each region are the contributions of the corresponding mode
to the total variance. The fine scale mode corresponds to the shaded
region. The transport and fine scale modes both include variance on
dilation scales less than 100 m.
5-7). T h e numbers for each mode in the lower portion of the figures are
the relative contributions to the total covariances. These numerical values,
in addition to other statistics, are summarized in Table I.
Included in these statistics are gradients in the longitudinal wind com-
ponents. Gradients in a piece-wise constant signal are readily defined as
the difference between adjacent constant values divided by the distance
between the center positions of the associated segments. Consequently, for
t h e t r a n s p o r t mode the gradients are computed over variable distances,
since t h e piece-wise constants are of variable width (Figure 2, c = 3).
T h e skewness values of the gradients in the longitudinal wind compo-
nents for t h e different modes are listed in Table I. A positive (negative)
value of the skewness for the u gradients indicates the wind speed increases
(decreases) more rapidly in the downstream direction. Thus, t h e results
1 hour 1 minute 1 second

0.25 I π —
Σ · =Var(w)
2
j p 0.20 .= 1.4 m / s
J=, 0.15
CD
Ü
I °· 10
CO
>
£ 0.05
0.03
0.00
1000 100 10 1 0.1 0.01 0.001
dilation a (km)
Figure 6. As in Figure 5 except for the variance in the vertical wind

component w.
Table I
Haar wavelet decomposition statistics. Variances (var and cov) are in
m2/s2. T h e values in the parenthesis are the results if the adaptive step is
not taken to include additional small scale motions in the transport mode.
Mode Mesoscale Large Eddy Transport Fine Scale Total J
u var 1.043 1.474 1.456 0.411 4.3841
(1.263) (0.604)
v var 0.779 0.407 1.019 0.524 2.7294
(0.852) (0.691)
w var 0.027 0.170 0.731 0.457 1.3845
(0.526) (0.662)
u-w cov -0.049 -0.300 -0.438 -0.068 -0.8552
(-0.368) (-0.138)
u-w corr -0.29 -0.60 -0.42 -0.16 -0.35
(-0.45) (-0.22)
isotropy 0.03 0.18 0.59 0.98 0.39
(0.50) (1.02)
skewness of 0.048 -0.227 -0.846 -0.006 -0.220
u gradient (-0.230) (-0.019)
F i g u r e 7. As in Figure 5 except for the covariance between the time

series of the longitudinal wind component u and the vertical wind com-
ponent w.
show t h a t the transport mode is associated with rapid decreases in t h e

wind speed in the downstream direction, a characteristic of shear driven
t r a n s p o r t . T h e variable cutoff scale leads to a gradient skewness of —0.846
where as a constant cutoff scale results in a gradient skewness of only
—0.230. Thus, the spatially adaptive cutoff scale defines a more complete
physical partitioning.
Also listed in Table I are the correlation coefficients, defined in t h e
usual way as
C0V W)
/ ^ (22)
y/Var(u)Var(w)
and an isotropy coefficient which is calculated as
2 x Varjw)
}
Var(u) + Var(v) ^
If all three velocity components contain the same amount of variance, this
isotropy coefficient is unity.
T h e Haar is b u t one possible wavelet to use in decomposing t h e turbu-
lence. As mentioned at the beginning of Section 2.1, the Haar wavelet is an
element in a sequence of wavelets introduced by Daubechies [8]. Another
Table II
As in Table I except the numbers are D4 wavelet decomposition statistics.
Mode Mesoscale Large Eddy Transport Fine Scale Total
u var 1.052 1.556 1.364 0.372 4.3441
(1.207) (0.529)
v var 0.783 0.432 1.023 0.490 2.7282
(0.891) (0.622)
w var 0.029 0.176 0.737 0.442 1.3838
(0.563) (0.616)
u-w cov -0.069 -0.311 -0.421 -0.059 -0.8605
(-0.363) (-0.117)
u-w corr -0.40 -0.59 -0.42 -0.15 -0.35
(-0.44) (-0.20)
isotropy 0.03 0.18 0.62 1.02 0.39
(0.54) (1.07)
simple element of t h a t sequence is referred to as a D4 wavelet since it is

constructed from four coefficients whereas the Haar requires only two. A
distinction between different wavelets in the Daubechies sequence is local-
ity in physical space versus locality in Fourier space. For example, the D4
wavelet is more local in Fourier space t h a n the Haar whereas as the Haar
wavelet is more local in physical space. In order to study the sensitivity of
the present results to a specific wavelet, the D4 wavelet basis is applied to
the turbulence time series in the same manner as the Haar basis. T h e lack
of symmetry in the D4 wavelet was not a significant factor in this case.
T h e results of applying the Haar and D4 wavelet bases (Tables I and
II) are similar, though in the D4 wavelet decomposition (Table I) there is a
little more variance and covariance captured in the larger scale modes and
there is a little less variance and covariance in the two smaller scale modes.
Adapting the D4 wavelet decomposition to the spatial distribution of the
covariance has about the same effect as adapting the Haar decomposition.
T h e effect of adapting the decompositions is quantifiable in terms of the
differences between the values in parenthesis and the values immediately
above those numbers in Tables I and II. For example, when applying t h e
D4 wavelet (Table II), adapting the decomposition increases the u-w co-
variance (flux) in the transport mode by about 16%. Adapting the Haar
decomposition (Table I) leads to a 19% increase in the flux associated with
the t r a n s p o r t mode. Applying wavelets which are even more compact in
Fourier space lead to similar results. Though with higher order wavelets,
the ratio of the variance and covariance in the larger scales to t h a t of the
smaller scales tends to be greater.
Higher order wavelets appear to be better suited for representing t h e
larger scale modes, which normally have a smoother spatial structure. For
example, the variance and covariance spectra generated from decomposing

the turbulence d a t a with higher order wavelets contain peaks at the larger
( > 1 km) scales, which are b e t t e r defined t h a n in the Haar spectra. T h e
smaller ( < 1 km) scale turbulence and associated sharp gradients, however,
are efficiently captured by the Haar decomposition.
Another result is t h a t the total covariances captured by t h e D4 wavelet
(Table II) are slightly different t h a n t h e totals listed in Table I obtained
from t h e Haar decomposition. Total sampled covariances directly com-
p u t e d according to t h e left hand sides of (5) and (6) agree with t h e Haar
decomposition totals to more digits t h a n shown in Table I.
§4. P h y s i c a l I n t e r p r e t a t i o n
T h e turbulence d a t a analyzed in this study has been decomposed into
deviations of small scale averages from larger scale averages. T h e turbu-
lence measurements include 9.1 hours of wind tower d a t a which translates
to a b o u t 420 km using Taylor's hypothesis and a mean wind speed u = 12.8
m / s . T h e flow is partitioned primarily according to t h e scale dependence
of the u-w covariance (Figure 7). Deviations of 3.2 km averages from the
entire record average (variations on scales > 6.4 km) define the mesoscale
mode. Deviations from the 3.2 km averages represent t h e turbulence, which
in t u r n are partitioned into three different modes of variation (Section 2).
Based on the u-w covariance (see shoulder in Figure 7 around a = 2
km), t h e large eddy mode is defined by deviations of 400 m averages from
the 3.2 km averages, t h a t is variations on the scales between 800 m and
3.2 km. T h e exact scales defining t h e large eddy m o d e are somewhat
arbitrary, since the physics changes between the small scale and large scale
ends of this regime. At the small scale end of the large eddies, t h e vertical
motions become more significant. T h e large eddy mode corresponds to
motions which are more horizontal at the 45 m observation level compared
to the smaller scale motions as indicated by the small value of t h e isotropy
coefficient (Table I).
T h e large eddies may be the lower part of boundary layer scale motions
(on t h e order of 1 km deep) where t h e lower part of the eddies observed at
the tower level are forced by the ground to be more horizontal. In this case,
t h e t r a n s p o r t by the large eddies would increase with height. Roll vortices
are an example of such eddies [4]. W h e n observed from time series measured
from towers, such larger scale motions are sometimes referred to as inactive
eddies because they contribute significantly to the horizontal variance b u t
contribute little to t h e m o m e n t u m flux at levels closer to t h e ground. In this
case t h e motions are not considered to be traditional turbulence and must
be removed from t h e signal before traditional similarity arguments can b e
applied [15]. This concept may be most descriptive of t h e larger scale p a r t
124 J. Ho well and L. Mahrt
of the large eddies in the present partitioning. The weak vertical motions
t h a t occur on the large eddy scales, however, are well correlated with t h e
variations of the longitudinal wind component leading to significant (35%
of total) m o m e n t u m flux. While the mean shear seems to exert a greater
influence on the large eddy mode compared to the mesoscale mode (greater
gradient skewness, Table I), the shear effect on the large eddies is still small
compared to t h a t of the transporting eddies.
Deviations of the raw time series from the 400 m averages define t h e
two smallest scale turbulence modes. In order to distinguish the two types
of small scale motions, the scale separating the two modes varies depending
on the local behavior of the transport as discussed in Section 2. Specifically,
averages are computed over a sufficiently small scale ( < 400 m) in order
to resolve the local vertical transport of momentum. Deviations of these
smaller scale averages from the 400 m averages are included in the t r a n s p o r t
mode to capture a majority of the momentum flux, including local extremes
(Figure 4, panel c).
As a result, the fine scale structure is also determined by the variable
cutoff scale. This is in contrast to conventional high pass filtering where
the cutoff scale is constant throughout the record. A constant cutoff scale
is not used in this case because transporting eddies intermittently occur
on scales which are traditionally assigned to the fine scale structure. From
another point of view, the fine scale mode includes relatively larger scale
motions only at locations where small scale transport is absent. This means
t h a t motions occurring on a range of scales between 1.6 m and 50 m are of
the transporting eddy type at some positions while at other positions these
motions make u p the fine scale structure.
T h e t r a n s p o r t mode is characterized by strong gradient skewness (Ta-
ble I) reflecting t h e strong influence of the mean shear on the transporting
eddies. According to conventional expectations, the mean shear generates
eddy motions which transport higher m o m e n t u m toward the surface. In
terms of energetics, this momentum transport corresponds to conversion of
mean kinetic energy to turbulence kinetic energy. T h e d a t a d e c o m p o s i -
t i o n in t h i s s t u d y p r o v i d e s a definition of t h e t r a n s p o r t i n g e d d i e s
w h i c h allows q u a n t i t a t i v e verification of classical c o n c e p t s . Also
as expected, the value of the isotropy coefficient for the transporting eddies
is between t h a t of the large eddies and the fine scale structure. T h e main
transporting eddies are characterized by significant vertical motions which,
however, remain smaller t h a n the horizontal velocity components. This
is because the mean shear directly generates variance in the u-component
which subsequently induces vertical velocity fluctuations through pressure
fluctuations.
T h e remaining small scale deviations make up the fine scale structure.
Very little t r a n s p o r t is associated with the fine scale mode, and t h e energy
in t h e horizontal and vertical wind components are nearly equal. More

specifically, the isotropy coefficient is close to unity. T h e gradient skewness
is very small, since the fine scale structure does not receive energy directly
from t h e mean shear. Instead the fine scale structure receives energy from
the transporting eddies through the so called energy cascade to smaller
scales.
§5. C o n c l u s i o n
Geophysical time series generally consist of physically distinct modes of
variation, each occurring on a subrange of scales which depend on space and
time. Conventional decomposition or filtering techniques divide t h e time se-
ries according to scales which are constant in space and time. In this study,
distinct modes were isolated using a piece-wise constant (Haar wavelet) de-
composition which allows the scales defining a particular mode to vary with
record position. W i t h this approach sampled covariances are completely
and orthogonally decomposed. Partitioning the flow in this manner allows
assessment of t h e relative contributions of the different modes to traditional
statistics.
T h e turbulence d a t a analyzed in this study has been partitioned into
four modes of variation. Each mode is defined locally in terms of an upper
and lower cutoff scale. T h e cutoff scales for the two larger scale modes are
specified to be constant with respect to record position. T h e scale sepa-
rating t h e two smaller scale modes varies with position according to t h e
local m a x i m a in the spatial distribution of m o m e n t u m flux. A local mo-
m e n t u m flux is quantified in terms of the product of t h e difference t e r m s
Au(am;n) x A w ( a m ; n ) , which is equivalent to a product of wavelet co-
efficients at a fixed scale and position. Using a wavelet decomposition to
examine the spatial or temporal distribution of the scale dependent flux is a
promising approach for distinguishing distinct physical modes of variation.
Adapting the decomposition to the spatial distribution of m o m e n t u m
flux leads to an improvement in the small scale partitioning as interpreted
in terms of globally averaged statistics. T h e spatial dependence of t h e
cutoff scale allows the computed transport mode to capture more of the
m o m e n t u m flux; some of the flux occurring on scales traditionally included
as fine scale structure is now more correctly included in t h e t r a n s p o r t mode.
For t h e t r a n s p o r t mode, t h e gradients in the longitudinal wind component
are negatively skewed in the downstream direction, which verifies t h a t this
m o d e is primarily shear driven. If a constant cutoff scale were used, t h e
skewness of t h e gradients in t h e longitudinal wind for t h e main transporting
motions would b e only —0.230. After the additional small scale variations
are included by varying t h e cutoff scale, the gradient skewness is —0.846.
This change is a result of further resolving the microfronts associated with
m o m e n t u m transport. Moreover, the adaptive step leading to the spatially

varying cutoff scale reduces the u-w correlation for the computed fine scale
structure leading to a physically more pure decomposition.
Application of the adaptive technique to other geophysical time se-
ries requires t h a t the variable cutoff scale be posed in terms of the phys-
ical process of interest. The physics of the decomposition might also b e
posed in terms of sudden changes of a quantity associated with larger
scale variations. W i t h traditional filtering, for example, sudden climatic
changes, sharp frontal boundaries, or any near discontinuities in the time
series which lead to large scale changes will be partially partitioned into
the small scale part; the corresponding low pass filtered signal will include
only a smoothed version of the sharp changes. The adaptive decomposition
applied here can be constructed to include sharp changes in the larger scale
part of the signal, avoiding undue smoothing [16]. The basic goal of the
adaptive technique is to partition the flow according to the physics when
the subrange of scales describing a given physical mode varies spatially or
temporally. Since the present approach partitions the original time series
into separate time series for each mode, the coherent structures associated
with a particular physical mode can in t u r n be analyzed.
References
1. Alpert, B. K., Wavelets and other bases for fast numerical linear al-
gebra, Wavelets: A Tutorial in Theory and Applications (C. K. Chui,
ed.), Academic Press, 181-216, 1992.
2. Battle, G., Cardinal Spline Interpolation and the Block Spin Construc-
tion of Wavelets, Wavelets: A Tutorial in Theory and Applications (C.
K. Chui, ed.), Academic Press, 73-90, 1992.
3. Bradshaw, G. A., and T. A. Spies, Characterizing canopy gap structure
in forests using wavelet analysis, J. Ecology, 80, 105-215, 1992.
4. Brown, R. A., Longitudinal instabilities and secondary flows in the
planetary boundary layer: A review, Rev. Geoph. Space Phys., 1 8 ,
683-697, 1980.
5. Burt, P. J., The Pyramid as a Structure for Efficient Computation,
Multiresolution Image Processing and Analysis (A. Rosenfeld, ed.),
Springer Verlag, 6-35, 1984.
6. Burt, P. J., T h e Laplacian pyramid as a compact image code, IEEE
Trans. Commun., 3 1 , 532-540, 1983.
7. Chui, C. K., An Introduction to Wavelets, Academic Press, 1992.
Commun. Pure Appl. Math, 4 1 , 909-996, 1988.
Annu. Rev. Fluid Mech. 24, 395-457, 1992.
10. Feauveau, J.-C, Nonorthogonal multiresolution analysis using

wavelets, Wavelets: A Tutorial in Theory and Applications (C. K.
Chui, ed.), Academic Press, 153-178, 1992.
11. Gamage, N. and C. Hagelberg, Detection and analysis of microfronts
and associated coherent events using localized transforms, J. Atmos.
Sei., 50, 750-756, 1993.
12. Gamage, N. and W. Blumen, Comparative analysis of low level cold
fronts: wavelet, Fourier, and empirical orthogonal function decompo-
sitions, Mon. Wea. Rev. 121, 2867-2878, 1993.
13. Gambis, D., Wavelet transform analysis of the length of the day and
the El-Nino/Southern Oscillation variations at intraseasonal and inter-
annual time scales, Ann. Geophysicae, 10, 429-437, 1992.
14. Haar, A., Zur Theorie der orthogonalen Funktionensysteme, Mathema-
tische Annalen 69, 331-371, 1910.
15. Högström, U., Analysis of turbulence structure in the surface layer
with a modified similarity formulation for near neutral conditions, J.
Atmos. Sei., 47, 1949-1972, 1990.
16. Howell, J. F. and L. Mahrt, An adaptive multiresolution data filter:
application to turbulence and climatic time series, J. Atmos. Sei., 51,
1994.
17. Kristensen, L., D. H. Lenschow, P. Kirkegaard, and M. Courtney, The
spectral velocity tensor for homogeneous turbulence, Bound.-Lay. Me-
teor., 47, 149-193, 1989.
18. Kumar, P. and E. Foufoula-Georgiou, A new look at rainfall fluc-
tuations and scaling properties of spatial rainfall using orthogonal
wavelets, J. Appl. Meteor., 32, 209-222, 1993.
19. Madych, W. R., Some elementary properties of multiresolution anal-
yses of L2(Rn), Wavelets: A Tutorial in Theory and Applications (C.
K. Chui, ed.), Academic Press, 259-294, 1992.
20. Mahrt, L., Eddy Asymmetry in the Sheared Heated Boundary Layer,
J. Atmos. Sei. 48, 472-492, 1991.
21. Mahrt, L. and J. F. Howell, The influence of coherent structures and
microfronts on scaling laws using global and local transforms, J. Fluid
Mech., 260, 143-168, 1994.
22. Mallat, S., The theory of multiresolution signal decomposition: the
wavelet representation, IEEE Trans. Pattern Anal. Machine Intell., 7,
674-693, 1989.
resentation, J. Fluid Mech. 232, 469-520, 1991.
24. Strang, G., Wavelets and dilation equations: a brief introduction,
SIAM Review, 31, 614-627, 1989.
25. Unser, M. and A. Aldroubi, Polynomial splines and wavelets—a signal
processing perspective, Wavelets: A Tutorial in Theory and Applica-
tions (C. K. Chui, ed.), Academic Press, 91-122, 1992.
This material is based upon work supported by Grant DA A H04-93-G-0019

from the Army Research Office and Grant ATM-9310576 from the Physical
Meteorology Program of the National Science Foundation.
J. F. Howell
Oceanic and Atmospheric Sciences
Oregon State University
Corvallis, Oregon
USA
L. Mahrt
Oceanic and Atmospheric Sciences
Oregon State University
Corvallis, Oregon
USA
e-mail: mahrtUpbl.ats.orst.edu
Wavelet Analysis of Diurnal and Nocturnal Turbulence
Above a Maize Crop
Yves Brunet and Serge Collineau
Abstract. A set of daytime and nighttime turbulence data acquired above a

maize crop is analyzed, using various properties of the wavelet transform. Wavelet
variances and covariances provide characteristic duration scales for the turbulent
motions contributing most to the signal energy. Subsequent detection of their
signatures in the raw time series allows a characteristic gust frequency to be deter-
mined. Once normalized with friction velocity and canopy height, this frequency
appears to be similar to that observed above a pine forest, which supports the pos-
tulate that transfer processes over plant canopies are dominated by populations of
canopy-scale eddies, with universal characteristics. Conditional averaging of the
corresponding patterns provides a clear picture of the ejection-sweep processes.
Around the inversion time of the mean vertical temperature gradient, two scales
of motions are inferred from the observation of wavelet variances and covariances.
A technique based on the inverse wavelet transform is then developed for parti-
tioning the original signals into small- and large-scale components, using a cut-off
scale deduced from wavelet variances and covariances. Despite their wavelike as-
pect, the extracted large-scale components are not gravity waves. Their nature
is unclear but the proposed methodology has potential applications for studying
wave-turbulence interactions.
§1. Introduction
Daytime turbulence in the vicinity of vegetation canopies has been ex-
tensively investigated during the past ten years. It is now widely recognized
that, within and just above vegetation canopies, turbulent transport pro-
cesses are to a large extent dominated by intermittent, energetic coherent
structures, with length scales of the order of the canopy depth ([1], [18],
and [20]). A dynamic model for the formation and development of these
'canopy eddies' has been proposed recently ([11] and [19]).

ISBN 0-12-262850-0
130 Y. Brunei and S. Collineau
These structures may be seen in time series of wind velocity compo-

nents as periodic patterns revealing the existence of ejection-sweep pro-
cesses (a slow upward movement of air, followed by a strong downward
motion associated with an acceleration of horizontal velocity). In t h e pres-
ence of a mean vertical temperature gradient, periodic r a m p p a t t e r n s are
then apparent in time series of temperature (gradual rises followed by rel-
atively sharp drops). Collineau and Brunet [4] developed a methodology
based on the wavelet transform for studying such time series. From mea-
surements acquired above and within a pine forest in slightly unstable con-
ditions, Collineau and Brunet [5] determined a mean duration scale for
the active part of the coherent motions, using the wavelet variance, and
estimated the frequency distribution of time intervals between contiguous
structures, using the wavelet transform as a jump-detector. T h e latter ap-
proach enabled t h e m to perform conditional sampling on the time series.
This provided a clear picture of the ejection-sweep process, as well as an
evaluation of the relative contributions of coherent motions to m o m e n t u m
and heat fluxes.
Much less attention has been paid so far to nighttime turbulence. Paw
U et ai [18] presented some evidence t h a t similar processes occur, revealed
by inverse temperature ramps. Also, gravity waves have been observed
in a variety of plant canopies, just after sunset [17]. Both phenomena
are intermittent. Over plant canopies, turbulent coherent motions appear
quasi-periodically at a frequency depending, to a first approximation, on
friction velocity u* and canopy height h ([18] and [19]). Gravity waves
can be seen in time series of temperature and vertical velocity as trains
of oscillations characterized by the Brunt-Vaisala frequency. Discriminat-
ing between propagatory and turbulent events is a m a t t e r of importance.
Their respective behaviour towards pollutant diffusion, for example, has
consequences on diffusion modelling: turbulent processes induce t r a n s p o r t
of air, while linear waves do not. However, b o t h contribute to the variance
of vertical velocity, which is a critical input parameter in diffusion mod-
els. Not discriminating between waves and turbulence can thus lead to an
overestimation of the actual rates of diffusion [7].
Given these considerations, the aim of this paper is fourfold:
1. show t h a t the methodology proposed by Collineau and Brunet ([4]

and [5]) is also applicable over shorter plant canopies, where the
'signal' (i.e., organized turbulence) to 'noise' (i.e., random small-
scale motions) ratio is much smaller;
2. show t h a t organized turbulence in slightly unstable conditions ex-

hibits the same structure above a forest and a maize crop, a scale
factor apart;
Diurnal and Nocturnal Turbulence Above a Maize Crop 131
3. extend the analysis to stable conditions;
4. propose a methodology for performing scale separation in measured

time series, using the filtering capabilities of the wavelet transform.
For this, we use a micrometeorological d a t a set acquired above a maize

canopy. After a presentation of the site and the experimental procedures,
we briefly describe the characteristics of mean flow and turbulence. We
then recall the concept of wavelet variance (see [4] and [12]) and introduce
the wavelet covariance. Both are used to determine characteristic duration
scales of the energy-containing events in the time series. This enables an
objective detection scheme to be designed, using the good localization in
time of the wavelet transform. Conditional sampling is then performed on
the time series, leading to a clear picture of the ejection-sweep process.
Finally, we present a few preliminary results of wavelet decomposition of
turbulent d a t a into large and small scale components.
§2. E x p e r i m e n t a l P r o c e d u r e s a n d F l o w C h a r a c t e r i s t i c s
T h e experiment was carried out over a maize crop at Grignon, France
(48°51'N, 1°58Έ), from July to September 1990. T h e d a t a presented here
were obtained on August 13. At this date, the canopy height (evaluated
as the m e a n height of a set of 100 individual plants) was h = 1.55 m, with
a zero-plane displacement height d = 1.15 m (d is the apparent level of
m o m e n t u m absorption by the plants, used as the origin of heights in crop
studies). T h e leaf area index (LAI, or total leaf area per unit soil surface)
was 4.1. A meteorological mast was equipped with slow-response temper-
ature (shielded, aspirated thermocouples) and windspeed (cup anemome-
ters) sensors at 7 heights z ranging from z/h = 0.3 to z/h = 4.6. T h e
distance from the leading edge of the field was larger t h a n 200 m in the
direction of prevailing winds. Two three-dimensional sonic anemometers-
thermometers (model PAC-100, Dobbie Instruments), were also set u p on
the mast at heights z/h = 0.6 and 1.6. All three velocity components (u,
streamwise; t>, lateral; tu, vertical) and air t e m p e r a t u r e T were sampled at
20 Hz, a r a t e high enough for exploring a significant part of t h e inertial
subrange, and stored on a micro-computer.
Only t h e d a t a acquired above the canopy are used here. Linear trends
are first removed from all original time series, split ted into contiguous 30
min runs. In what follows, prime notations ΐί', w' and T' stand for t h e
fluctuations of ii, w and T around their mean values ïï, w and T, calculated
for each run.
Two sets of samples are used in this study, one ('diurnal') acquired from
13:30 to 15:30 U T and t h e other ('nocturnal') from 18:00 to 20:00 U T , just
before and after sunset (which occurred at 19:10 U T ) . An illustration of t h e
micrometeorological conditions encountered during these periods is given

in Table 1. For daytime samples mean horizontal windspeed is between 1.5
and 2 m s " 1 , and later drops to about 1 m s - 1 . In the first case the friction
velocity u* = (—u'w')1/2 is of the order of 0.4 m s - 1 , whereas in the second
case it drops from 0.35 m s - 1 at 18:00 UT to about 0 . 1 m s - 1 after 19:00. As
a result of radiative cooling of the underlying surfaces, the sensible heat flux
H = pCpW'T' (where p is the air density and cp the specific heat at constant
pressure) is directed towards the canopy throughout the second period,
with values between about —10 and —40 W m - 2 . For diurnal samples H
is positive, between 100 and 170 W m - 2 . The values taken by {z — d)/L
(where L is the Monin-Obukhov length) indicate that slightly unstable (or
near-neutral) conditions are prevailing in the afternoon, whereas light to
moderate stability occurs from 18:00 to 20:00 UT.
Table I
Summary of experimental conditions: statistical moments of u, w and T
(means, standard deviations, covariances) and stability parameter
\z - d)/L.
1 Time u T (JT u'w' w'T' — * - * ii
(m/s) (°C) (m/s) (m/s) (K) (m 2 /s 2 ) (K m/s) L
1 13:30 1.482 28.68 0.984 0.497 0.765 -0.111 0.134 -0.064 1

14:00 1.706 28.44 0.868 0.517 0.616 -0.139 0.113 -0.038
14:30 1.904 28.40 0.994 0.598 0.597 -0.193 0.116 -0.024
1 15:00 1.776 28.19 0.815 0.529 0.507 -0.170 0.080 -0.020 |
1 18:00 1.660 25.90 0.786 0.444 0.203 -0.113 -0.027 0.013 1
18:30 0.912 24.20 0.399 0.218 0.422 -0.030 -0.030 0.101
19:00 0.851 21.76 0.179 0.106 0.346 -0.006 -0.010 0.404
1 19:30 0.992 20.84 0.220 0.137 0.353 -0.010 -0.014 0.230 |
Figure 1 shows typical 200 s samples of instantaneous fluctuations,

extracted from the two data sets. All time series exhibit some degree of in-
termittency, with sporadic active periods particularly visible in the traces
of instantaneous cross-products. Closer inspection of Figure lb (night)
reveals typical recurrent features characterized by sharp increases in tem-
perature, associated with positive (respectively negative) fluctuations in
u (respectively w). These temperature patterns are commonly called 'in-
versed ramps' by analogy with their daytime counterpart (slow increases
followed by sudden drops), readily visible in Figure la. These tempera-
ture discontinuities can be unambiguously identified as the signatures of
turbulent coherent motions sweeping down into the canopy air space. In
stable conditions, this results in warmer air being imported from aloft by
energetic gusts, yielding a sudden increase in time series of temperature.
From this example, it therefore seems that in nightime and daytime
Figure 1. Typical examples of time series of velocity components, air

temperature and cross-products u'w' and w'T' above the maize crop
(z/h = 1.25) on August 13. (a): diurnal sample, (b): nocturnal sample.
For the clarity of these graphs the original signals (recorded at 20 Hz)
have been low-pass filtered at 2 Hz.
T(eC)
Figure 2. Mean air temperature profiles at various times, within and

above the canopy (the horizontal dashed line represents the canopy top).
conditions m o m e n t u m and sensible heat fluxes are transferred by the same

dynamic process, the only difference (sharp rises or drops in temperature)
being due to the inversion of the mean temperature profiles (Figure 2). No
regular wavelike oscillations are visually detectable in the nocturnal time
series. However, this may not always be the case and further analysis is
desirable, in order to investigate the possible existence of gravity waves.
For this, we need advanced d a t a processing tools, such as those provided
by the wavelet transform.
§3. W a v e l e t Variance a n d Covariance

In this paper, dilation over the non-dimensional scale a and translation
at b of the analyzing wavelet ψ are performed with the following normal-
ization factor (preserving the Li-norm):
a a (1)
We use only the continuous version of the wavelet transform of a function
/ , noted ( / , ψ ( α ^ ) and defined as:
+oo
/(i)V>(a'6) dt. (2)
/
-oo
Let / be a given experimental time series with energy Ef = ( / , / ) . We
define t h e wavelet variance WVf of / for t h e analyzing wavelet ψ as:

+oo
Κ/,^°·6)>|2Λ. (3)
/
-co
Because energy is conserved by wavelet transformation, t h e area under a
wavelet variance is proportional t o t h e signal energy, when appropriate log-
arithmic axes are used (see [4]). Due t o t h e 'built-in' smoothing process of
the wavelet variance (integration over t h e translation parameter), ensemble
averaging is not usually required t o obtain stable, readable variance graphs.
Collineau a n d Brunet [4] showed t h a t when a wavelet variance graph
exhibits a single peak a t a scale a, a characteristic duration scale D can b e
defined as:
D = άϋψ, (4)
where ϋψ is a duration constant which was shown t o b e an intrinsic prop-
erty of t h e wavelet ψ. Όψ can be calculated as t h e solution of a simple, first-
order differential equation, which was done analytically for a few common
wavelets (Haar, Mexican h a t , first derivative of a Gaussian function...). It
was demonstrated t h a t D can b e interpreted as t h e mean duration of t h e
elementary events contributing most t o t h e signal energy (D = r / 2 for a
sine function of period r ) .
T h e wavelet transform conserves not only t h e energy ( / , / ) , b u t also
the inner product (f,g) of two functions [6], since:
{Ϊ,9) = γτ / -{fâ'b)){g^(a'b))dadb, (5)
where
ΙΦΗΙ 2 ^ (6)
ω
-co
has a finite value, by definition of a wavelet (Φ being t h e Fourier transform

οΐψ). Just as a wavelet variance can b e considered as t h e wavelet equivalent
of a Fourier spectrum, one can define a wavelet covariance WCfg(a) from
Equations (2), (3) a n d (5), t h e wavelet equivalent of a cospectrum:
+oo
(f,4>^){g^)db. (7)
/
-co
To our knowledge, this property has not been used yet on experimental
d a t a sets 1 .
1
After this work was completed, it was brought to the authors' knowledge that
the concept of wavelet covariance has been independently described by L. H. Hudgins
(Wavelet Analysis of Atmospheric Turbulence, Thesis, University of California, Irvine,
March 1992).
0.2H
. O.H
""Too
D(s)
loo
D(s)
D(s)
Figure 3. Wavelet variances and covariances: WVU (triangles), WVW

(plus), WVT (crosses), WCUW (full squares) and WCWT (open squares).
(a): 14:00-15:00 UT, (b): 19:00-20:00 UT, (c): 18:00-18:30 UT.
We have calculated the wavelet variances of i/, w and T, and the wavelet
covariances of (u,w) and (u;,T), for all the available d a t a samples. T h e
Mexican h a t wavelet was chosen for this, because of its good localization
in frequency. Figure 3a shows the variances and covariances obtained from
14:00 to 15:00 U T , under slightly unstable conditions.
Single, well-defined peaks can be seen on the wavelet variance and
covariance curves. T h e wavelet variance for vertical wind velocity WVW
peaks at Dw « 1-2 s, whereas WVU and WCT peak at Du « DT ~ 3 -
4 s. From a qualitative point of view, these results are very similar to
those obtained by Collineau and Brunet [5] over a pine forest in similar
stability conditions. Only the magnitudes of the Z)-scales differ: in t h e
latter case typical values of 5-7 s and 10-15 s respectively were obtained
for the w-variance peaks, and the peaks for u and T. However, normalizing
the duration scales Du or DT by w* and h as suggested in the introduction
give t h e same mean value Du*/h « 0.85 for b o t h canopies. It therefore
seems t h a t over the maize crop the motions contributing to these peaks are
also typical canopy-scale eddies, of the same type as those depicted in [5],
This will be confirmed further.
T h e covariances WCUW and WCWT also exhibit unambiguous peaks, at
the same D scales as WVU and WVT (Figure 3a). Instead of using WCUW
and WCWT, Collineau and Brunet [5] calculated the wavelet variance of
series of instantaneous cross-products u'w' and w'T'. They were shown to
display a peak at the same scale as WVW. This is very similar to what hap-
pens with Fourier spectra and cospectra in the surface layer: the cospectra
peak at the same frequency as the u and T spectra (see [13] for a review),
whereas t h e spectra of the time series of cross-products such as u'w' and
w'T' peak at t h e same frequency as the w spectra.
Figure 3b shows the wavelet variances and covariances obtained from
19:00 to 20:00 U T , under conditions of moderate stability. T h e wT co-
variance is now negative, following the inversion in sensible heat flux, b u t
all curves are qualitatively similar to those in Figure 3a, with identical D
scales. T h e variances of u and w appear flatter t h a n in Figure 3a b u t this is
just a consequence of a change in the respective magnitude of t h e wavelet
variances and covariances. Full-scale graphs of WVU and WVW indeed show
well-marked peaks at the same D-scales as in Figure 3a. This confirms t h e
visual impression given by Figure l b , t h a t the same kind of structures as
those commonly seen in diurnal time traces are also present in this noctur-
nal sample. T h e normalized peak scales Du*/h are now about 3-4 times
larger t h a n in Figure 3a since ii* has dropped from 0.3-0.4 m s - 1 to a b o u t
0.1 m s - 1 . Using mean wind speed as a velocity scale would lead to a
smaller difference since ΰ has only dropped by a factor of 2.
Considering now t h e period from 18:00 to 18:30 U T (Figure 3c), two
peaks are apparent on all curves b u t WVW. Consequently, there seems t o
be two superimposed types of motions in this d a t a sample, which b o t h

contribute to the signal energy. The motions responsible for the left-hand
side peak are probably of the same nature as those visible in Figures 3a
and 3b, since here also D « 3.5 s. The second peak is located at D « 13 s
and its n a t u r e is yet unknown. We will focus on this in Section 5.
§4. J u m p D e t e c t i o n a n d C o n d i t i o n a l S a m p l i n g
Having determined a typical event duration, we are now interested in
how frequently these events occur. Several methods for detecting sharp
edges in digital time series, based on the use of localized transforms, have
been proposed recently ([12] and [14]). The combined localization prop-
erties of wavelets in b o t h time and frequency can be exploited together
usefully for designing efficient jump-detection algorithms. Collineau and
Brunet [5] showed t h a t using (Τ,φ^) (or ( w , ^ ( a , 6 ) ) ) as detection func-
tions of ramps (or large excursions from the mean) in time series of temper-
ature (or streamwise wind velocity), after determination of the scale a of
the wavelet variance peak, leads to an accurate detection of the structures
visible in t h e time series. T h e final decision step in the detection algorithm
requires an empirical threshold for first derivative like wavelets (e.g., the
Haar wavelet), but only a slope sign for the Mexican hat wavelet, which
yields zero-crossings in (T, φ(α^) (or (w, φ(α^)) whenever large, sharp rises
or drops occur in the series. W i t h a comparable reliability, the zero-crossing
algorithm is therefore simpler t h a n those based on thresholds, which is the
case of most standard jump-detection algorithms used in turbulence, such
as the Variable Interval Time-Averaging technique (see for example [21]).
However, it has been made clear in [15] t h a t detection through first or sec-
ond derivative like wavelets are equivalent algorithmic problems when the
objective is to classify the 'sharpness' of variations of the signal, which, in
the latter case, requires thresholding the slope of the wavelet transform.
Here, detection was performed on all time series of streamwise wind
velocity, using the Mexican hat wavelet. Wind velocity was chosen because
it has a higher signal-to-noise (as defined in the introduction) ratio t h a n
t e m p e r a t u r e in this particular d a t a set, and also because in this case the
slope sign criterion does not depend on whether the d a t a are diurnal or
nocturnal. Figure 4 shows histograms of the time intervals Δ between the
detected events, for two typical half-hour samples with similar values of
w*, in b o t h diurnal and nocturnal conditions. T h e two distributions are
strikingly similar and provide the same mean value Δ « 9 s. Normalizing
Δ as above, we obtain a mean value Au*/h « 2.1 over the whole d a t a set,
which is again in fairly good agreement with the value of 1.8 obtained over
the forest canopy already mentioned [5].
Finnigan ([9] and [10]) found t h a t the peak frequency of the Fourier
U.H--
ri
\l
14:30 1
0.3-
'55
iI
G 1 18:00
1^0.2-
L^
Probabi
1
o
, f~M Ι,Π,^, ,25

I
0 5 10 15 20
D(s)
Figure 4. Histograms of the time intervals between the events de-

tected in diurnal (14:30-15:00 UT) and nocturnal (18:00-18:30 UT) sam-
ples.
spectrum of u just above a wheat canopy was very close to t h e directly

measured arrival frequency of canopy-scale sweeps. So, if the time scale Δ
really is a measure of the mean time interval between successive canopy-
scale eddies, the inverse l / Δ should be directly comparable to t h e peak
frequency of the u spectrum. It should also be close to the peak frequency
of t h e T spectra and uw cospectra, which were shown in [13] to have all
the same normalized value fh/uh « 0.15 (±0.05), within and just above
a plant canopy (uh being the mean horizontal wind speed at canopy t o p ) .
Fourier spectra and cospectra for the sample 14:00-15:00 U T are shown in
Figures 5a and 5b, respectively. Except for the w spectrum, they all peak
at a frequency / « 0.1 Hz, which is in excellent agreement with 1/Δ = 0.11
Hz. As t h e mean wind speed at the canopy top is üh = 1.1 m s - 1 for this
sample, we also find fh/üh ~ 0.14, well in the range of expected values.
Having identified the relevant patterns, we can perform conditional
sampling and averaging in order to obtain the mean 'signature' of the de-
tected motions. Collineau and Brunet [5] showed t h a t conditional averaging
requires t h e use of a wavelet very well localized in time, which is not t h e
140 Y. Brunei and S. Colîineau
«o1
0.0Û1
001
I I I I I II
0.01
1 1—I I I I I II
0.1
1 1—I I I I HI
1
1 1—I I I I II I
10
f(Hz)
0.15
J . l j i 1 1—i Ii Ii 11
T—I I I ii
11 1 1—I I I I I II 1 1—I I I I III 1 1—I I I I II
10
Wooi o.oi o.i i
f(Hz)
Figure 5. u-, w- and Γ- Fourier spectra (a), and um- and wT-
cospectra (b) for the sample 14:00-15:00 UT. Spectral densities are mul-
tiplied by the frequency and divided by the signal variance.
case of the Mexican hat. Instead, we use the ' R a m p ' wavelet defined as:
0 -oo < x < -0.5

2x + 1 -0.5 < x < 0
φ(χ) =
2z - 1 0 < x < 0.5
0 0.5 < x < +oo
which was shown by these authors to lead to slightly b e t t e r results t h a n

t h e H a a r wavelet, because its shape matches more closely t h e signature of
the turbulent structures (see Figure 6). Since this type of wavelet exhibits
a peak at j u m p times of the input signal, we had to calibrate t h e required
detection threshold. This was done so t h a t the number of detected events
matches t h e number of events seen by t h e Mexican h a t in t h e same time
series. For this exercise the detection was performed on t h e t e m p e r a t u r e
signal as in [5], b u t quite similar results would be obtained from t h e velocity
signals.
For each turbulent variable / , we first define N time-windows of length
Δ , centered at the N points detected in the temperature series. Then, for
each variable (if, w, T ) , the raw signal in each window is normalized by t h e
local m e a n fi and the local standard deviation σ / ί . Finally, the conditional
(normalized) average {/(£)} i s calculated over the N windows as:
i=l f*
with Ti — Δ / 2 < t < T{; + Δ / 2 , r,· being the ith detection time.
T h e averaged p a t t e r n s are presented in Figures 6a (diurnal samples)
and 6b (nocturnal samples). Both figures show a characteristic slow ({ti} <
0) upward ({w} > 0) movement of air, rapidly switching to a strong down-
ward ({w} < 0) motion associated with an acceleration of horizontal ve-
locity ({u} > 0). In the diurnal case a sharp t e m p e r a t u r e drop follows a
slow rise, whereas t h e opposite occurs in the nocturnal case. Such p a t t e r n s
have t h e characteristics of typical 'ejection-sweep' processes. It has t o b e
emphasized t h a t detection was performed on time series of t e m p e r a t u r e
only, which suggests a strong coupling between temperature, vertical and
streamwise velocities.
These results invite several comments. Firstly, they show t h a t above
our maize crop t h e motions responsible for t h e peaks in wavelet variances
and covariances, and objectively detected by t h e wavelet transform, do have
t h e 'signatures' of coherent eddies, such as those observed over a variety of
plant canopies or rough surfaces ([18], [19], and [20]). Secondly, t h e nor-
malized p a t t e r n s are very similar to those obtained by t h e same technique
above a forest canopy ([5]), with {T} peaking between ± 0 . 6 and ± 0 . 8 , and
{u} and {w} having equally smaller amplitudes, b o t h comprised between
0.8-
<Λ
0.6
a
VH {T}
> 0.4- {u}
cd
i—H
G 0.2- "{w}
O
■<-j
Ό 0-
e
o
X) -0.2- "**---.--- S
<
Nυ
'K! -0.4-
εo
t-<
-0.6-
(a)
-0.8
~A~ÎS Ï 3 F
t(s)
Figure 6. Normalized conditional averages {w}, {w} and {T}, as de-

fined in Equation (8), for the diurnal (a) and nocturnal (b) series.
about ±0.2 and ± 0 . 5 . Thirdly, apart from the difference in the tempera-
ture signature due to the opposite sign of the mean vertical t e m p e r a t u r e
gradient, turbulence seems to be dominated by the same processes in our
daytime and nighttime samples. However, this should not b e generalized
since a relatively small stability range has been explored ((z — d)/L between
- 0 . 0 6 4 and 0.404, as shown in Table 1).
We noticed in Figure 3c t h a t just after the inversion in t h e mean tem-
p e r a t u r e gradient, another peak was visible in the wavelet variances and
covariances, corresponding to longer duration scales. As gravity waves have
already been observed in similar conditions [17], one may reasonably won-
der whether those peaks could be due to such phenomena. This is t h e
objective of the next section.
§5. W a v e l e t D e c o m p o s i t i o n
T h e occurrence of linear gravity waves can be predicted theoretically
by a linear stability analysis of the dynamical flow equations, as t h e result
of Kelvin-Helmoltz instabilities (see [7] and [8]). This analysis provides a
characteristic wave frequency depending on the mean vertical t e m p e r a t u r e
gradient. This is the Brunt-Vaisala frequency N , defined as:
where g is the acceleration due to gravity (see for example [22]). As t h e

mean t e m p e r a t u r e gradient at the height of the sonic anemometer is of t h e
order of 0.3 K m - 1 in the period 18:00-18:30, the Brunt-Vaisala frequency
is N « 0.1 Hz (which corresponds to a characteristic time scale of a b o u t
10 s). On the other hand, the £)-scale corresponding to the second peak
in Figure 3c is 13 s; as the wavelet variance of a sine function peaks at a
frequency corresponding to half the period ([5]), the associated time scale
would be of 26 s, a value more t h a n twice as large as the former.
T h e linear stability analysis also predicts a phase shift of ± π / 2 between
w and T [22]. In fact, most experimental approaches for studying gravity
waves have relied on Fourier analysis, performed on time series of stream-
wise and vertical velocity, pressure and temperature. Cospectra, variance,
phase, q u a d r a t u r e and coherence spectra have been shown to exhibit spe-
cific features in t h e presence of gravity waves ([2], [7], and [16]). However
this is not always the case, for several reasons: (i) there is not always a
clear separation in the power spectra between the regions of waves and
turbulence, especially for w near the ground [16]; (ii) factors such as verti-
cal wind shear, advection or proximity to the surface may complicate t h e
situation and, for instance, p e r t u r b a t e the phase angles and decrease t h e
levels of coherence [3]; (iii) non-stationary waves cannot be easily detected
with Fourier-based methods if they only last for a few oscillations.
In our case, no specific feature could be visually detected in t h e vari-

ous spectra and cospectra. Nevertheless, it is still desirable to investigate
the n a t u r e of t h e second peak in the variances and covariances. For this,
the combined time- and frequency-localization properties of the wavelet
transform provide powerful tools.
From the definition of the wavelet transform, one can define an inverse
wavelet transform as:
-I Λ + 00 /« + 00 -I
m=7T / -(fîa'b)Wia'b)(t)dadb, (10)
a
W J-oo J0
by which the original signal can be retrieved from the whole set of wavelet
coefficients (/, φ^α^). This 'reconstruction' formula enables one t o fil-
ter t h e original function, by splitting the integration domain into specific
ranges of a scales. For instance, the wavelet transform can b e used as
a low-pass, high-pass or band-pass filter, provided t h a t adequate cut-off
scales are imposed in Equation (10).
In the present case, decomposition of the signal into two components
can b e simply performed by splitting the reconstruction integral over a
into large and small a scales. If the wavelet variance of the signal exhibits
two peaks separated by a gap (the wavelet equivalent of a spectral gap), an
obvious choice for the cut-off scale ac is the gap scale. A textbook example,
using t h e sum of two sine functions with different periods, was given in [4]:
the reconstructed small and large components of the signal allowed t h e
authors to retrieve accurately the respective original sine functions.
We applied this m e t h o d t o the series acquired from 18:00 t o 18:30
U T , after they were band-averaged a t 0.5 s in order to facilitate the com-
putation. T h e Mexican hat wavelet was chosen for its good frequency
localization. A cut-off duration scale Dc = 7 s was inferred from Fig-
ure 3c, corresponding, for this particular wavelet, to a cut-off dilation scale
ac = Dc/Όφ = Dcy/2 /π « 3.15. Figure 7 presents a small sample of the
extracted small-scale components w s , ws and Ts of the respective input
signals. Roughly speaking, what we see here is a representation of t h e
original series after low-pass filtering a t / / = 1/0.5 = 2 Hz and high-pass
filtering at fh = 1/DC = 0.14 Hz. All traces exhibit pseudo-periodic os-
cillations, sometimes combined into two characteristic types of events: one
set consists in slow, upward movements of cold air and the other in fast,
downward motions of warm air, each individual event lasting for about 3 s.
These p a t t e r n s correspond respectively to the ejection and sweep processes
described above. As they occur repeatedly all over the half-hour sample
with t h e appropriate time scales, it is likely t h a t most of t h e m are signa-
tures of t h e above-mentioned coherent structures, even if each individual
p a t t e r n is not a replicate of the averaged patterns showed in Figure 6b.
T h e large-scale components w/, wi and T/ shown in Figure 8 exhibit similar
0.08
Figure 7. A sample of small-scale components (u s , ws and Ts) ex-

tracted by the inverse wavelet transform (from the run 18:00-18:30 UT).
200
Figure 8. A sample of large-scale components (w/, wi and T/) ex-

tracted by the inverse wavelet transform (from the run 18:00-18:30 UT).
features, at larger time scales. It is now clear t h a t the underlying motions

are not gravity waves. In particular, wi and T/ are not out of phase, b u t
in phase opposition (just as ws and Ts). The scale difference between the
two types of components is illustrated in Figure 9a and 9b for vertical ve-
locity and temperature, respectively. The two large-scale components are
strongly linked. Both exhibit wavelike motions, with a periodicity of order
20-25 s, e.g. twice the duration scale associated with the second peak.
This scale separation is only based upon scale criteria. However, as
all large-scale components show repeatable features, they are not likely
to be due to unpredictable low-frequency trends, but rather result from a
deterministic, dynamic process, perhaps linked with the establishment of
the thermal inversion. At the present stage, the reasons for these features
are unclear. More d a t a collected during such transition regimes must be
analyzed, as well as d a t a acquired in much more stable conditions.
Nevertheless, this analysis emphasizes the filtering capabilities of the
wavelet transform, which appears to have great potential for extracting
specific components and investigating phase aspects.
T h e results of this experimental study confirm the importance of co-
herent structures in turbulent transfer processes between plant canopies
and the atmosphere. The features observed in time series of velocity com-
ponents and air temperature recorded above a maize crop are very similar
to those observed above other plant canopies, and quite consistent with
the picture of canopy turbulence acquired over the past few years. Wavelet
analysis has enabled us to detect characteristic signatures of coherent mo-
tions in all series, revealing the occurrence of ejection-sweep processes. T h e
scale of these motions is canopy-dependent and a comparison between the
pine forest and the maize crop suggests t h a t the occurrence frequency of
these motions scales with friction velocity and canopy height. In b o t h cases
a value of Au*/h « 2 was obtained for the mean time interval between
successive structures. Also, a unique value Du*/h « 0.85 was obtained for
the duration scale corresponding to the peak in the wavelet variances of
streamwise velocity and temperature. The wavelet covariances of uw and
wT were also observed to peak at the same scale. These results support
the postulate t h a t 'universal' mechanisms are responsible for the structure
of turbulence in the vicinity of plant canopies, in near-neutral conditions
([1], [11], and [19]).
Analysis of nighttime samples in light to moderate stability also proved
the existence of similar processes. In moderate stability ((z — d)/L œ 0.3)
the time scale Au*/h was found to be somewhat larger, but this needs
further confirmation since only one hour long sample was available for this
0.08
0.04
Figure 9. A sample of small- and large-scale components of vertical

velocity (a) and temperature (b) (from the run 18:00-18:30 UT).
study. In lighter stability, around the inversion time of the mean vertical
temperature gradient, another scale of motions was observed, for which no
satisfactory explanation can be given at the present time.
These results were obtained by performing a wavelet analysis of tur-
bulent time series, consisting in several steps. Firstly, computation of
wavelet variances and covariances provide characteristic duration scales
of the events contributing most to the signal energy. Secondly, the wavelet
transform is used for detecting these events, by 'looking' at the series at
these particular scales. As far as the detection itself is concerned, the use
of a wavelet such as the Mexican hat, which yields zero-crossings whenever
large discontinuities occur in the series, provides a simple detection scheme
which does not require adjustment of an empirical threshold. Conditional
averaging of the detected patterns can then be performed, enabling one to
extract from the time series a clean picture of the signatures of these mo-
tions. Compared to more traditional analysis techniques, this procedure
has the major advantage of being based upon a unique, self-consistent
line of mathematical treatments, which rely on the combined time- and
frequency-localization properties of the wavelet transform.
These properties have also been used for partitioning the original series
into large- and small-scale components, using a cut-off scale defined as the
gap scale between two peaks in the wavelet variances and covariances. This
has promising applications, for instance in the study of wave-turbulence
interactions.
References
1. Brunet, Y., J. J. Finnigan and M. R. Raupach, A wind tunnel study
of air flow in waving wheat: single-point velocity statistics, Boundary-
Layer MeteoroL, 1994, In press.
2. Caughey, S. J., Boundary-layer turbulence spectra in stable conditions,
Boundary-Layer MeteoroL, 11, 3-14, 1977.
3. Caughey, S. J. and C. J. Readings, An observation of waves and tur-
bulence in the earth's boundary layer, Boundary-Layer MeteoroL, 9,
279-296, 1975.
4. Collineau, S. and Y. Brunet, Detection of turbulent coherent motions
in a forest canopy. Part I: Wavelet analysis, Boundary-Layer MeteoroL,
65, 357-379, 1993.
5. Collineau, S. and Y. Brunet, Detection of turbulent coherent mo-
tions in a forest canopy. Part II: Time-scales and conditional averages,
Boundary-Layer MeteoroL, 66, 49-73, 1993.
6. Daubechies, I., The wavelet transform, time-frequency localization and
signal analysis, IEEE Trans. Information Theory, 36, 961-1005, 1990.
7. De Baas, A. F. and A. G. M. Driedonks, Internal gravity waves in a
stably stratified boundary layer, Boundary-Lay er Meteorol, 31, 303-

323, 1985.
8. Einaudi, F. and J. J. Finnigan, The interaction between an internal
gravity wave and the planetary boundary layer. Part I: The linear anal-
ysis, Quart. J. Roy. Meteorol. Soc, 107, 793-806, 1981.
9. Finnigan, J. J., Turbulence in waving wheat. Part I: Mean statistics
and honami, Boundary-Layer Meteorol., 16, 181-211, 1979.
10. Finnigan, J. J., Turbulence in waving wheat. Part II: Structure of mo-
mentum transfer, Boundary-Lay er Meteorol., 16, 181-211, 1979.
11. Finnigan, J. J. and Y. Brunet, Turbulent airflow in forests on flat and
hilly terrain, in Wind and wind-related damage to trees, M. P. Coutts, J.
Grace (eds.), Cambridge University Press, Cambridge, 1994, In press.
12. Gamage, N. K. K. and C. Hagelberg, Detection and analysis of mi-
crofronts and associated coherent events using localized transforms, J.
Atmos. Sei., 50, 750-756, 1993.
13. Kaimal, J. C. and J. J. Finnigan, Atmospheric boundary layer flows.
Their structure and measurement, Oxford University Press, New York,
1993.
14. Mahrt, L., Eigen structure of eddies; continuous weighting versus con-
ditional sampling, in Non-linear variability in geophysics: scaling and
fractals, D. Schertzer, S. Lovejoy (eds.), Kluwer Academic Publishers,
Dordrecht, 145-151, 1991.
15. Mallat, S. and S. Zhong, Characterization of signals from multiscale
edge, IEEE Trans, on Pattern Anal, and Mach. Intel, 14, 710-732,
1992.
16. Nai-Ping, Lu, W. D. Neff and J. C. Kaimal, Wave and turbulence struc-
ture in a disturbed nocturnal inversion, Boundary-Lay er Meteorol., 26,
141-155, 1983.
17. Paw U, K. T., R. H. Shaw and T. Maitani, Gravity waves, coher-
ent structures and plant canopies, in Reprints of the 9th Symposium
on turbulence and Diffusion, American Meteorological Society, Boston,
244-246, 1990.
18. Paw U, K. T., Y. Brunet, S. Collineau, R. H. Shaw, T. Maitani, Q.
Jie and L. Hipps, On coherent structures and turbulence above and
within agricultural plant canopies, Agric. For. Meteorol., 61, 55-68,
1992. Corrigendum, Agric. For. Meteorol, 63, 127, 1993.
19. Raupach, M. R., J. J. Finnigan and Y. Brunet, Coherent eddies in
vegetation canopies, in Proc. Fourth Australasian Conf. on Heat and
Mass Transfer, Christchurch, 75-90, 1989.
20. Raupach M. R., R. A. Antonia and S. Rajagopalan, Rough-wall tur-
bulent boundary layers, Appl. Mech. Rev., 44, 1-25, 1991.
21. Schols, J. L. J., The detection and measurement of turbulent structures
in the atmospheric surface layer, Boundary-Lay er Meteorol, 29, 39-58,
1984.
22. Stull, R. B., An introduction to boundary layer meteorology, Kluwer
Academic Publishers, Dordrecht, 1988.
Yves Brunet
Laboratoire de Bio climatologie
INRA
B P 81
33883 Villenave d'Ornon Cedex
France
email: [email protected]
Serge Collineau
Laboratoire de Bioclimatologie
INRA
BP81
33883 Villenave d'Ornon Cedex
France
Wavelet Spectrum Analysis
and Ocean Wind Waves
Paul C. Liu
Abstract. Wavelet spectrum analysis is applied to a set of measured ocean wind

waves data collected during the 1990 SWADE (Surface Wave Dynamics Experi-
ment) program. The results reveal significantly new and previously unexplored
insights on wave grouping parameterizations, phase relations during wind wave
growth, and detecting wave breaking characteristics. These insights are due to
the nature of the wavelet transform that would not be immediately evident using
a traditional Fourier transform approach.
§1. Introduction
Ever since Willard J. Pierson [18] adopted the works of John W. Tukey
[22] and introduced the power spectrum analysis to ocean wave studies,
Fourier spectrum analysis has been successfully and persistently used in
data analysis of wind-generated ocean waves. Over the past four decades,
with the increased availability of new instruments for measuring wind and
waves, spectrum analysis has continued to be the fundamental standard
procedure used for analyzing wind and wave data.
Fourier spectrum analysis generally provides frequency information
about the energy content of measured, and presumed stationary, time-
series data. Characteristic properties of waves such as total energy and
dominant or average frequency can be readily derived from the estimated
spectrum. This information, however, pertains only to the time span of
the measured data. Changes and variations within a time series cannot be
easily unraveled. As stationarity in the data simply represents a mathe-
matical idealization, its validity is usually regarded as an approximation of
the real wave field. The effectiveness of applying Fourier spectrum analysis
to a rapidly changing wave field, such as during wave growth or decay, is

ISBN 0-12-262850-0
152 P. Liu
uncertain. The emergence of wavelet transform analysis which can yield

localized time-frequency information without requiring that the time-series
be stationary has presented a rewarding and complementary approach to
the traditional Fourier spectrum analysis and has advanced significant new
perspectives for improved wave data analysis.
While wavelet analysis has been widely recognized as a revolutionary
approach applicable to many fields of studies, the application of wavelet
transform to wind wave data analysis is still in its infancy. This article
mainly presents the author's own attempt at understanding wind-generated
waves using the wavelet decomposition.
§2. W a v e l e t S p e c t r u m
Following a standard formulation [3], we briefly summarize the wavelet
transform. We start with a family of functions, the so-called analyzing
wavelets, ißab(t)i that are generated by dilations a and translations b from
a mother wavelet, ip(t), as
φα,(ή = ^ = φ ( ^ ) (1)
where a > 0, —oo < b < +oo, and f_™ ψ(ί)<ϋ = 0. The continuous wavelet
transform of a time-series, X(t), is then defined as the inner product of ipab
and X as
X(a, b) =< ψα1), X >= - = / X(t)tl>*( )dt (2)

a
y/\a\J-oo
or equivalently in terms of their corresponding Fourier transforms
+00
Χ(ω)ψ*(αω)βίοωαω (3)
/
-oo
where an asterisk superscript indicates the complex conjugate. In essence
the wavelet transform takes a one-dimensional function of time into a two-
dimensional function of time and scale (or equivalently, frequency).
In practical applications, the wavelets can be conveniently discretized
by setting a = 2s and b = T2S in octaves [4] to obtain
V.r(t) = 2 - / V ( 2 - t - r ) , (4)
where s and r are integers. Then the continuous wavelet transforms (2)
and (3) for time series data X(t) become
1 /*+°° t
x{s T) = x{t)r{ (5)
' 7Ploo v "T)dt
Wavelet Spectrum Analysis 153
and
(6)
In general, t h e studies of wavelet transforms and wavelet analysis are
centered on two basic questions [4]: (1) Do t h e wavelet coefficients com-
pletely characterize t h e time-series d a t a ? (2) C a n t h e original time series
be reconstructed from t h e wavelet coefficients? T h e answers t o b o t h of
these questions are clearly yes as evidenced by t h e voluminous literature
in recent years. In this paper we rely on t h e affirmative answer t o t h e first
question a n d concentrate on exploring t h e wavelet transform of measured
wind waves. It is an exciting a n d fruitful area for practical application of
the wavelet transform . As d a t a analysis on wind wave studies comprises
mainly of applications of statistics and Fourier transforms, t h e s u m m a r y
shown in Table 1 indicates t h a t wavelet transform analysis is a logical ex-
tension t o t h e currently available analyses.
In analogy with Fourier energy density spectrum, we can readily define
a wavelet spectrum for a data series X(t) as
WX(S,T)=X(S,T)X*(S,T) = \X(S,T)\2. (7)
There are other designations in t h e literature for Wx(s, r) [21]. Results

from t h e application of short-time Fourier transforms have been called spec-
trograms, whereas results from t h e application of wavelet transforms have
been called scalograms. Since in practice t h e scale, s, a n d translation,
r , can b e associated with a corresponding frequency, CJ, a n d time, £, (7)
can b e considered as a representation of t h e time-varying, localized energy
spectrum for a given time series.
We can similarly define a cross wavelet spectrum for t h e study of two
simultaneously measured d a t a sets X(t) and Y(t) as
WXY(S,T) = X(S,T)Y*(S,T) (8)
and accordingly,
T ( s τ ) = WXY(S,T)
VWXk(s,T)Wyk(s,T)
and
WXk(s,T)WYk(s,T)
as t h e complex-valued wavelet coherency and its square, t h e real valued
wavelet coherence, respectively, between t h e two d a t a sets. T h e functions
S W x y ( s , r ) a n d 3 W x y ( s , r ) in (10) are respectively t h e real a n d imagi-
nary parts of WXY (s, r ) , and hence t h e co- and quadrature- wavelet spectra
of X(t) and Y(t).
154 P. Liu
Table 1· Analogy of statistics, Fourier transform and wavelet transform

analysis.
Statistics Fourier Transform Wavelet Transform
Variance Frequency Spectrum Wavelet Spectrum
E(X2) SxH = xx* = |x|2 Wx(s,r)=XX* = \X\2
Covariance Cross Spectrum Cross Wavelet Spectrum
E(XY) SXY(u,) = ΧΫ* WXY(S,T) = XY*
Coefficient of Coherency Wavelet Coherency

Correlation
E XY SXYM p _ WXY(S,T)
r - ( ) 1
y/E{X)E{Y) y/SxWSyW y/Wx(s,T)WY(s,T)
Coefficient of Coherence Wavelet Coherence

Determination
2 _ \Ε[ΧΥ)γ ^2 _ PxyM2 Γ2 _ ΡΥχγ(8,τ)\2

' — E(X)E(Y) ~ WX(S,T)WY(S,T)
In implementing the applications, there are a number of well-defined

continuous wavelet forms available [6]. In this study we choose to use
the complex-valued, modulated Gaussian analyzing wavelet known as the
Morlet wavelet. This wavelet, originally proposed by Morlet et al. [16],
ushered in the present wavelet era and is given by
2
_ Jmt
ψ(ή = etmte -t /2 (H)
Its Fourier transform is
φ(ω) = y/2ê-("-m^2. (12)
Here we should point out that this wavelet is not an admissible wavelet
since a correction term is needed because φ(0) Φ 0. However, in prac-
tice choice of a large enough value for the parameter m, (e.g. m > 5)
generally renders the correction term negligible. In this study, we follow
Daubechies [5] and use m = π v/27^2. While there are admissible wavelets
available, the Morlet wavelet has been widely used in signal analysis and
sound pattern studies. Aside from its convenient formulation and histori-
cal significance, its localized frequency is independent of time, a feature of
particular advantage for wind wave studies.
Time Series of Wind Waves
1600 1650 1700 1750 1800
Wavelet Spectrum
0.3 '
N0.25 - -
fr 0.2 r*-~S\ Λ Γ\ '
; Ç^àVl â ^ Aß^k
Λ Λ Λ^
Φ
|0·15
u- 0.1
0.05 - 1 1
-
1650 1750 1800
time s
Figure 1. A sample plot of a time series of wind waves and its respec-
tive wavelet spectrum.
§3. Applications
In the following three subsections we present three wavelet transform
analyses of wind wave data leading to distinct results that would be dif-
ficult, if not impossible, to obtain from the usual Fourier transform. The
data used in the applications were measured during the recent SWADE
(Surface Wave Dynamics Experiment) program [25]. The wind and wave
data were recorded from a 3 m discus buoy during the severe storm of Oc-
tober 26, 1990. The buoy was located at latitude 38°22.1' N and longitude
73°38.9' W, with a water depth of 115 m near the edge of the continental
156 P. Liu
shelf offshore of Virginia in the Atlantic Ocean. Time series of wind and
waves were both recorded at 1 Hz from a combined design of a three-axis
accelerometer and magnetometer along with the Datawell Hippy system.
A total of 100 sets of data, each 1024 s in length, were used in the anal-
yses. The data, predominantly wind-generated waves, covered the entire
duration of the storm with wind speeds ranging from calm to 18 m/s and
significant wave heights approaching 7 m.
3.1. Wave grouping effects

Wind waves appear in groups; i.e., higher waves occur successively in
separated sequences. This phenomenon is well-known to seasoned sailors
and can sometimes be seen in wind wave recordings. Apart from being
a confirmed natural phenomenon, the existence of wave groups tends to
challenge the conventional notion that wave data can be considered as a
stationary process.
Wave data analysis, aimed at studying wave group characteristics, has
been confined to identifying individual groups by counting the number of
wave heights that exceed a prescribed height. A group is simply measured
by a group length which is the number of waves counted. While statistics
of the group lengths can be assessed, efforts have been generally directed
at correlating the mean group length with spectral properties of the data
[14].
In a wavelet transform analysis of wave data, an examination of a
contour plot of resulting wavelet spectrum of waves shows distinct energy
density parcels in the time-frequency domain. Figure 1 presents a simulta-
neous plot of sea surface elevations and the contours of their corresponding
wavelet spectrum. The contour patches shown in the figure clearly indi-
cate wave groupings that are visibly identifiable in the time series. The
boundary of a wave group can be readily specified by setting an appropri-
ate threshold energy level in the wavelet spectrum. Essentially there is a
localized time-frequency energy spectrum for each group of waves, which
is potentially more informative than previous approaches.
Based on the boundary specified for each wave group from the wavelet
spectrum, we have at least four relevant group parameters to characterize
a wave group:
(i) The group time length, tgy which is the difference between the max-
imum and minimum time scales the group boundary covered.
(ii) The total group energy, Eg, which is an integration of the local
wavelet spectrum over the time length tg.
(iii) The dominant group frequency, / p , which is the frequency of the
peak energy over the time length tg.
(iv) The dominant group wave height, hp, which can be obtained from
the time series as the maximum trough-to-crest wave height over
the time length tg.
The variability of these parameters indicates that wave groups are ap-
parently diverse, irregular, nonperiodic, and independent from each other.
The formidable task is to determine the significance and usefulness of these
parameters. Here we consider a simplified approach of forming two nor-
malized parameters:
• normalized group time length = tg * fp, and

• normalized total group energy = Eg/h'i.
W1 10° 101 102

Normalized Group Time Length
Figure 2. A scatter plot of normalized total group energy versus nor-

malized group time length, the solid line is a linear least square fit for
the data.
A scatter plot of these two normalized parameters, shown in Figure 2,

indicates a fairly well-defined linear relationship. As the normalized total
group energy is a measure of energy content, and the normalized group
time length is a measure of the number of waves with possibly the same
peak energy frequency, Figure 2 implies, that higher energy content in a
158 P. Liu
wave group tends to generate more waves in the group. This interesting
result, while intuitively understandable, is new.
A scatter plot of averages of dominant group wave heights versus sig-
nificant wave heights is shown in Figure 3. The significant wave height,
defined as the average of the highest one-third wave heights in the wave
record, is a familiar and widely-used parameter. For practical applications,
such as in engineering design, mean dominant group wave height would be
more pertinent than the significant wave height. Figure 3 shows that sig-
nificant wave heights are slightly less than the averages of dominant group
wave heights.
Figure 3. A scatter plot of mean group wave height versus significant-

wave height.
3.2. How do wind waves grow?

The wind and waves measurements in the S WADE program introduced
a new data collection practice, namely, that wind and waves were mea-
sured at the same resolution simultaneously. Previously, wind data were
merely collected as hourly averages. The availability of simultaneous high-
resolution wind and wave data has provided an unparalleled opportunity
to directly examine detailed wind action on waves, especially during wave
growth.
How do wind waves grow? It is a question that several generations of
scientists have addressed. In addition to the early work of Jeffreys [11] and
Ursell's [23] famous "nothing very satisfying" summary, modern conceptual
perceptions of wind waves primarily stem from the theoretical conjectures
of Phillips [17], Miles [15], and Hasselmann [8]. The current proliferation
of numerical wave models is basically developed from these early theories.
Numerous measurements of wave energy spectra with average wind speeds
have been conducted for the validation and possible enhancement of the
available models. Now with the latest SWADE measurements and the
advancement of wavelet transforms, we are able to examine wind wave
processes from new perspectives.
Time Series of Wind Speeds
1550 1600 1650 1700 1750 1800
Wavelet Spectrum
Figure 4. The time series of wind speeds corresponding to the wave

data of Figure 1 and its respective wavelet spectrum.
One way of analyzing simultaneously recorded wind and wave measure-

ments is through cross wavelet spectrum analysis. Figure 4 shows a part
of the wind speeds and their wavelet spectrum corresponding to the wave
data of Figure 1. There is no obvious relationship between the two time
series that we can deduce from the top parts of Figures 1 and 4. However, if
we consider the wavelet spectrum, a tract of high energy density contours
160 p. Liu
appears in b o t h spectra over the same frequency ranges and during t h e

time when highest wave heights occurred in the wave time series. Quali-
tatively we might infer t h a t wind and waves interact immediately during
wave growth.
frequency = 0.1131 : 0.1199- : 0.127... Hz
F i g u r e 5. Plots of three peak-energy frequency components versus

time. The five subgraphs from top down are, respectively, wavelet spec-
trum for wind speeds, wavelet spectrum for waves, the real part, the
imaginary, and the phase of coherence.
To see if we can verify this inference quantitatively, we calculate t h e

cross wavelet spectrum and their corresponding wavelet coherence for t h e
simultaneous wind and wave data. T h e results, expressed either in contour
or three-dimensional plots, are rather intricate and perplexing. It is not at
all clear what we can meanfully deduce. If, however, we plot t h e results for
individual frequencies, we can see some interesting results. Figure 5, corre-
sponding to the same d a t a of Figures 1 and 4, is an example of what these
plots can tell us. T h e five separate graphs in Figure 5 display, respectively
from top down, t h e wavelet spectrum for wind, t h e wavelet spectrum for
waves, t h e real p a r t , t h e imaginary p a r t , and t h e phase of wavelet coher-
ence. All of t h e plots contain the three frequency components of 0.1131,
0.1199, and 0.127 Hz for which the energy density is highest.
Note t h a t in Figure 1 there are five groups of waves t h a t can b e iden-
tified from t h e wavelet spectrum. In t h e second graph of Figure 5 in which
energy densities increase and decrease with respect t o time, only three
stronger groups (i.e., at time marks 1570, 1630, and 1695) are reflected
from the fluctations of these frequency components. T h e top graph of Fig-
ure 5 shows t h a t the wavelet spectrum components for wind speeds exhibit
similar, b u t more, energy fluctations with time. Some of the fluctuations
correspond closely to those of the waves. By examining the b o t t o m three
graphs of Figure 5, it shows quite clearly t h a t for t h e three wave groups
identified with appreciable energy contents, the real p a r t of their coherence
is close t o 1 1, their imaginary part close to 0, and their phase is also close
to 0. Therefore, during wave growth, t h e frequency components for peak
wave energy between wind and waves are inherently in phase. Wave groups
constitute t h e basic elements of wind wave processes, and the wave growth
are primarily taking place within the wave group.
As t h e growth of wind waves is an extremely complicated process,
the above results contribute still qualitatively toward an understanding of
t h e n a t u r e of how do waves grow. While we are accustomed to correlate
wave growth with "average" wind speeds, the results presented here clearly
show t h a t waves are in fact responding to wind speeds instantly. Further
detailed studies may challenge or counter more familiar notions of wind
waves. Using cross wavelet spectrum analysis not only introduces new
d a t a analysis techniques, it may also leads to new courses of exploration.
3.3. Detecting breaking waves

Wave breaking is a familiar phenomenon t h a t occurs intermittently
and ubiquitously on t h e ocean surface. It is visible from t h e appearance
of the whitecaps, yet it can not be readily measured with customary in-
struments. Wave breaking has been recognized as playing a crucial role
in accurate estimations of the exchange of gases between t h e ocean and
the atmosphere [24] and in the transfer of m o m e n t u m from wind t o t h e
ocean surface [1]. Most of t h e practical works on wave breaking [2], b o t h
1
Unlike the Fourier cross spectrum analysis where averaging can be used to avoid
coherence being identically one, we have to use the real and imaginary parts of the
coherence separately here, since their sum, the coherence, is indeed identically one in
this formulation. On the other hand, this approach successfully substantiates the use
of the co- and quadrature spectra [19] to signify their "in phase" and "out of phase"
properties, respectively.
162 P. Liu
in the laboratory and in the field, have been done with specialized methods
based on radar reflectivity, optical contrast, or acoustic output of the ocean
surface. Here we show that with the help of wavelet spectra [12], instead
of using specialized measurement devices, a basic wave-breaking criterion
can be easily implemented to wind wave time series to distinguish breaking
from non-breaking waves. This simple and fairly efficient approach can
be readily applied to indirectly estimate wave breaking statistics from any
available time series of wind-generated waves.
o: lambda = 1.35 x: lambda = 1.15

1
m
X
|x X
x
o x 1 ° * ° x II1 ï 1 *
_6I , ,i , , . ..,_. . ,1 1
1500 1550 1600 1650 1700 1750 1800
Time s
Figure 6. A sample plot of a time series of wind waves, same as Figure

1, with possible breaking waves marked by o's and x's indicating different
high frequency ranges \ωρ : ωη defined by the λ values given at the top
of the figure.
One of the most frequently used approaches for the study of wave break-
ing is the use of a limiting value of the wave steepness beyond which the
surface cannot be sustained [13]. Alternatively, assuming a linear dispersion
relationship, the wave surface will break when its downward acceleration
exceeds a limiting fraction, 7, of the gravitational acceleration, g, that is
ασ2 ^ 7#. The quantity ασ 2 can be calculated for a time series of wave
data since the local wave amplitude, a, is available from the measured time
series while the local wave frequency, σ, can be obtained from the wavelet
spectrum. In classical studies, it has generally been assumed that 7 = 0.5.

Recent laboratory studies [10] have shown that 7 is closer to 0.4. Some field
measurements [9] further indicate that the value of 7 should even be lower.
In this study we chose to follow the laboratory results and use 7 = 0.4.
o: lambda = 1.35 x: lambda = 1.15
18|-
16
ι
I12 o o
o°o
ce
2> 9p°
ω 10
O X
9> xä
X
'χΧ o x
**
I Ή H nil
4 6 8 10 12
Wind Speed m/s
Figure 7. Plot of the percentage of breaking waves with respect to

wind speeds. The o's and x's are the same as defined in Figure 6.
Since the wavelet transform provides an equivalent time-frequency spec-

trum, Wx(cj,f), for the wind wave time series, then there is a localized
frequency spectrum at each data point, X(U), given by
$iH = [Wx(u;,t)]t=ti.
It is not immediately clear which frequency should be used for σ in calcu-
lating ασ 2 . Because breaking events are generally associated with the high
frequency part of the spectrum, for each X{U) we chose to define a σ,· as
the average frequency [20] over the high frequency range, \ωρ : u>„, of the
localized spectrum at t = £,· as
l 1/2
Gi -
f:;yH")d"
S£9*M&» J
where ωρ is the localized frequency at the energy peak, ωη is the cut-off
164 P. Liu
frequency, and λ is a number greater than 1 that denotes the start of the
high frequency range beyond ωρ. The exact location of this high frequency
range has not been clearly defined. Considering this range as corresponding
to the familiar equilibrium range, one frequently used value of λ has been
1.35 [7].
To test this approach, Figure 6 presents an illustration of the analysis
where estimated breaking waves are marked on the same time series seg-
ment given in figure 1. The x's and o's represent the results with a high
frequency range between 1.15 and 1.35 times, respectively, of the local peak
energy frequency, CJP, and cut-off frequency, ωη. While the λ values of 1.15
or 1.35 has been chosen rather arbitrarily for comparisons, they are clearly
not always recognizing the same breaking waves. In general with the same
cut-off frequency, the lower end of the frequency range farther away from
the local peak frequency, i.e. large λ value, would yield higher local average
frequency σ and more breaking waves. Therefore an exploration of break-
ing waves could potentially serve to resolve the definition of the well-known
but still not yet well-defined equilibrium range. Figures 7 present plots of
overall percentages of breaking waves from all the data analyzed in this
study as a function of wind speed. While the data points are scattered
considerably, there is an approximate linear trend indicating an increase
in the percentage of breaking waves with an increase in wind speed. The
results shown in Figure 7 are in general accord with various available ob-
servations [9]. According to these results, breaking waves become prevalent
when wind speeds exceed 10 m/s.
At the present, the limiting fraction of downward wave acceleration
from the gravitational acceleration, 7, and the parameter locating the lo-
cal equilibrium range beyond local peak frequency, λ, are both tentative.
Therefore, the wavelet transform approach that leads to these results is
useful, convenient, and also exploratory. Perhaps a better simultaneous
measurement of wind-wave time series and wave breaking would suffice to
substantiate the approach. Unfortunately operational and sufficient instru-
ment for this simple purpose is still lacking.
§4. Concluding Remarks

We anticipate two groups of readers who might be benefit from this
paper, namely, those interested in wavelet applications and those interested
in wind wave studies. As this is a first attempt in applying the wavelet
transform to wind waves, the results are inevitably primitive. We hope we
have succeeded at least in demonstrating the rich potentials for wavelet
analysis. There are many important analysis issues that must yet be ad-
dressed, including rigorous basis for the cross wavelet spectrum analysis
which we have used here. For the wind wave studies, wavelets certainly
provide ample opportunities for d a t a analysis. From t h e encouraging re-

sults we reported here, we are justified in being optimistic.
References
1. Agrawal, Y. C., E. A. Terray, M. A. Donelan, P. A. Hwang, A. J.
Williams III, W. M. Drennan, K. K. K a h m a , and S. A. Kitaigorodiski,
Enhanced dissipation of kinetic energy b e n e a t h surface waves, Nature,
3 5 9 , 219-220, 1992.
2. Banner, M. L., and D. H. Peregrine, Wave breaking in deep water,
3. Combes, J. A., A. Grossmann, and P h . Tchamitchian (Eds.), Wavelets,
Time-Frequency Methods and Phase Space, 2nd ed. Springer-Verlag,
1989.
4. Daubechies, I., Ten Lectures on Wavelets, Society of Industrial and
Applied Mathematics, 1992.
5. Daubechies, I., T h e wavelet transform, time-frequency localization and
signal analysis, IEEE Trans. Inform. Theory, 3 6 , 961-1005, 1990.
6. Farge, M., Wavelet transforms and their applications t o turbulence,
7. Günther, H., W. Rosenthal, T. J. Weare, B. A. Worthington, K. Hassel-
m a n n , and J. A. Ewing, A hybrid parametrical wave prediction model,
J. Geophys. Res., 8 4 , 5727-5738, 1979.
8. Hasselmann, K., On the non-linear energy transfer in a gravity-wave
spectrum, P a r t 1, General theory, J. Fluid Mech., 12, 481-500, 1962.
9. Holthuijsen, L. H., and T. H. C. Herbers, Statistics of breaking waves
observed as whitecaps in the open sea, J. Phys. Oceanogr., 16, 290-297,
1986.
10. Hwang, P. A., D. Xu, and J. Wu, Breaking of wind-generated waves:
measurements and characteristics, J. Fluid Mech., 2 0 2 , 177-200, 1989.
11. Jeffreys, H., On the formation of water waves by wind, Proc. Roy. Soc,
bf A 107, 189-206, 1925.
12. Liu, P. C , Estimating breaking wave statistics from wind-wave time
series data, Annales Geophysicae, 4, 970-972, 1993.
13. Longuet-Higgins, M. S., On wave breaking and the equilibrium spec-
t r u m of wind -generating waves, Proc. Roy. Soc, A 3 1 0 , 151-159,
1969.
14. Masson, D. and P. Chandler, Wave groups, a closer look at spectral
methods, Coastal Engineering, 20, 249-275, 1993.
15. Miles, J. W., On t h e generation of surface waves by shear flows, J.
Fluid Mech., 3 , 185-204, 1957.
16. Morlet, J., G. Arens, I. Fourgeau, and D. Giard, Wave propagation
and sampling theory, Geophysics, 4 7 , 203-236, 1982.
166 P. Liu
17. Phillips, O. M., On the generation of waves by turbulent wind, J. Fluid

Mech., 2, 417-455, 1957.
18. Pierson, W. J., and W. Marks, The power spectrum analysis of ocean
wave record, Trans. Amer. Geophys. Un., 33, 834-844, 1952.
19. Priestley, M. B., Spectral Analysis and Time Series, Academic Press,
1981.
20. Rice, S. 0., Mathematical analysis of random noise, in Noise and
Stochastic Processes ( N. Wax, ed.), Dover, New York, 133-294, 1954.
21. Rioul, O., and F. Flandrin, Time-scale energy distribution: a general
class extending wavelet transforms, IEEE Trans. Signal Processing, 40,
1746-1757, 1992.
22. Tukey, J. W., The sampling theory of power spectrum estimates, Sym-
posium on Applications of Autocorrelation Analysis to Physical Prob-
lems, NAVEXOS-0-735, Office of Naval Research, 1949.
23. Ursell, F., Wave generation by wind, in Surveys in Mechanics, Cam-
bridge University Press, 216-249, 1956.
24. Wallace, D. W. R., and C. D. Wirick, Large air-sea gas fluxes associated
with breaking waves, Nature, 356, 694-696, 1992.
25. Weiler, R. A., M. A. Donelan, M. G., Briscoe, and N. E. Huang, Riding
the crest: A tale of two wave experiments, Bulletin Am. Meteorol. Soc,
72, 163-183, 1991.
This is GLERL Contribution no. 873. Partial financial support to initiate

this study was provided by the U.S. Office of Naval Research. This study
was also benefited from early collaborations with A. K. Liu and B. Chapron.
Paul C. Liu
NOAA Great Lakes Environmental Research Laboratory
Ann Arbor, Michigan
e-mail: liuUglerl.noaa.gov
W a v e l e t A n a l y s i s of Seafloor B a t h y m e t r y : A n E x a m p l e
Sarah A. Little
Abstract. 1-D wavelet analysis has been shown to be useful in studying bathy-
métrie profiles [7]. 2-D bathymétrie maps are less common than 1-D profiles,
but offer immensely more information about seafloor generation processes. 2-
D wavelet analysis is applied to swath-mapped bathymétrie data from the Mid-
Atlantic Ridge. Both image enhancement and feature identification are performed
with excellent results in the identification of the location and scarp facing direction
of ridge-parallel faulting. Wavelet image processing techniques enable computer
analysis of distribution and spatial patterns in faults to be performed without the
tedious job of transcribing hand picked and ruler-measured fault parameters from
printed images to a digital data base.
T h e wavelet decomposition of bathymétrie d a t a reveals structures and
p a t t e r n s which are easily overlooked in the raw data. In addition, it can
be used to isolate features of interest, such as fault scarps, for use in subse-
quent quantitative analysis. This paper describes the application of wavelet
analysis to seafloor topography. Much of seafloor bathymetry has been col-
lected by ships of opportunity traversing various sections of ocean. These
depth measurements are closely spaced in the along-track direction, b u t
t h e tracks are often many kilometers apart. T h e resultant d a t a sets are es-
sentially 1-dimensional spatial series. In a few areas of the seafloor, swath
bathymétrie surveys have been conducted which return detailed, 2-D, maps
of limited sections of the ocean. 1-D spatial series are more common, and
1-D wavelet analysis of bathymétrie profiles can be used to improve our un-
derstanding of the shape of the seafloor. Swath m a p p e d areas have much
more information t h a n single topographic profiles, of course, and I present
an example of image enhancement and fault scarp identification using 2-D
wavelet analysis.

ISBN 0-12-262850-0
168 S. Little
§2· 1-D T o p o g r a p h i e Profile

Little et al. [7] recently applied a wavelet decomposition to a 1-D,
1600 km long, bathymétrie profile from the northeast Pacific near Hawaii.
T h e profile runs parallel to the Murray Fracture Zone near 32°N, 152°W,
and crosses directly over a small Pacific bathymétrie high [9]. A 200 km
anomalous zone was discovered with the wavelet decomposition, which has
high-amplitude, long-wavelength topography and lacks the short-wavelength
topography which dominates both to the east and west. This "low-frequency"
zone is located between two fracture zones and the regional setting for this
area includes a bend in the Pioneer Fracture Zone to the north and a
break in the Murray Fracture Zone to the south. It is clear t h a t in this
topographically anomalous zone either the volcanism, tectonics or crustal
structure is significantly different from the surroundings. They interpreted
the anomalous crust to be the site of a short-lived, abandoned spreading
center which is associated with a regional, thermally induced, topographic
high.
T h e wavelet decomposition was particularly useful in finding the aban-
doned spreading center because these sites are marked by changes in the
roughness of the topography [10] which remain permanently frozen in the
crust after the spreading has ceased. The amplitude of the features of this
particular site were smaller t h a n other, known abandoned spreading cen-
ters, and these features are easily overlooked without the application of the
wavelet transform to decompose the signal into events at different scales.
T h e wavelet transform can be used to decompose a spatial series into
energy in a given frequency band and at a given location [4]. This can b e
contrasted to Fourier analysis which gives the energy in a given frequency
band for an entire signal. Although windowed Fourier transforms (or spec-
trograms) do give an estimate of frequency content as a function of location,
the spatial window is fixed, and the spatial resolution generally coarse. T h e
advantage of the wavelet transform is t h a t it matches its spatial localization
to the frequency of interest—fine spatial resolution for high frequencies and
broad spatial localization for low frequencies. In many cases, especially for
transient and variable signals such as faults and volcanic constructions, the
wavelet transform is superior to the Fourier transform for space-frequency
decomposition and event localization.
For spatial data, such as a bathymétrie profile, d a t a are decomposed
into energy at a given wavelet scale and at a given distance along track. In
the case of the Morlet wavelet used by Little et. al. [7], the wavelet scale
has a natural correspondence to spatial frequency.
T h e Morlet wavelet [5] has the properties of rapid decay in b o t h the
space and frequency domains, so as to maximally localize information in
b o t h domains, smooth so t h a t it is narrow-band and differentiable, and
Wavelet Analysis of Seafloor Bathymetry 169
complex so as to preserve phase information about the signal. T h e Morlet

wavelet is defined as a complex exponential, at some fixed frequency, times
a Gaussian with fixed variance:
w{t) = eîkth-£
where t is space, k is wavenumber, and σ is the Gaussian variance.
A single scale wavelet transform is the convolution of t h e wavelet at
t h a t scale with the data, b o t h of which are functions of the spatial domain.
For increased computational speed, this convolution may be computed in
the Fourier domain by multiplying the Fourier transform of the d a t a by
the Fourier transform of the wavelet, then taking the inverse transform of
the product. On a single scale, the wavelet transform can be thought of
as a linear filter. For a multiscale analysis, the wavelet shape is rescaled,
in powers of 2, to longer and shorter lengths, creating a bank of filters
of different sizes. T h e power in each size wavelet is normalized to one,
and then the wavelet is convolved with the data. In this way the series of
wavelet filters is used to scan for small- and large-scale events in the data.
W h e n a wavelet of a given size and shape is convolved with a spa-
tial series, the magnitude of the resultant series is a measure of the m a t c h
between the wavelet and d a t a series—small-scale events will m a t c h small
wavelets b u t not large wavelets. Therefore, when a bank of wavelets is
applied to a spatial series, one can b o t h identify transient events and dis-
tinguish events on different scales. In this way one can scan a d a t a series
for interesting features on a broad range of scales, and also identify re-
gions of t h e d a t a where events of a certain scale are entirely missing. Little
et. al. [7] used the Morlet wavelet transform, in this way, to discover t h e
anomalous low-frequency zone in a 1-D bathymétrie profile.
§3. 2 - D B a t h y m e t r y
T h e following section describes the application of wavelets generated
from spline functions to image enhancement and fault scarp identification
in 2-D swath-mapped bathymétrie data. T h e bathymétrie d a t a cover an
approximately 100 km x 70 km section of Mid-Atlantic Ridge near 29°N,
43°W, which is about 1/10 of the d a t a available in this region [13].
Topography near the Mid-Atlantic Ridge (MAR) is characterized by a
deep central median valley surrounded by faulted blocks leading out over
crestal mountains and down the outer flanks. These faulted blocks tend
to b e long linear ridges running parallel to t h e axis of spreading, ranging
in width from order of 10 m to 10 km. Swath m a p p e d areas of the M A R
reveal scales ranging from order of 100 m up to 100 km, and hence offer a
reasonable d a t a set with which to examine this major faulting process at
the MAR. Two major questions to ask about the M A R are 1) what does
170 S. Little
it look like? and 2) where are the faults located?

Two types of spline filters are used, a linear B-spline wavelet for visual
image enhancement to improve intuitive understanding, and the derivative
of a cubic spline for quantitative fault scarp edge detection.
3.1. Data
T h e d a t a used in this analysis was collected by a 16-beam Sea Beam
swath echo sounder. This instrument is mounted on the hull of a ship
and sends out sound in beams whose individual footprints are 2 | x 2 |
degrees, which translates to about 150 m on a side in a water depth of
3000 m. As the ship steams forward, it continuously collects an approxi-
mately 2 km wide swath of depth readings. These d a t a are tied to latitude
and longitude via satellite navigation and put together in a regional m a p
of bathymetry. There are places where the swaths do not overlap, and a
linear interpolation has been performed in these gaps so t h a t the d a t a can
be smoothly represented in a 2-D matrix of depth values. T h e d a t a are
shown in Figure 1 in a contour/gray scale image. There are 755 points
of longitude between 43.567°W and 42.552°W, and 495 points of latitude
between 28.581°N and 29.917°N. Depths range from 4000 m in the deep-
est areas (black) to 1800 m in the shallowest areas (white). There are two
white areas inside the image which represent no data, a large one centered
on longitude point number 200 and latitude point number 50, and a small
gap located at longitude point number 240, latitude point number 100. T h e
white areas to the left and right of the image also contain no bathymétrie
information. T h e central valley of the M A R runs up the middle of the
image at about 20° off vertical. On each side of the central valley are t h e
crestal mountains, and beyond them a small portion of the flanks. This
area of the Mid-Atlantic Ridge contains a ridge offset, visible in t h e upper
central part of the image, where the two major valleys are offset from one
another. T h e d a t a used in this paper have previously been analyzed for
fault locations using a curvature method [14], [15].
3.2. Filter design

In two-dimensions, the wavelet transform can be used to decompose
an image into a series of images, each of which contains information at a
specific location of features at a single scale (or spatial frequency). Further,
since images are 2-dimensional, each spatial frequency is coupled with an
orientation. These two parameters can be varied to produce a filter tuned
for features of nearly any width, length and orientation. T h e filter is a
matrix of coefficients which is convolved with a 2-dimensional image. T h e
resolution of location of a given feature in the wavelet transform is matched
to t h e scale of t h a t feature. Precision in location of fault scarps will b e
comparable to their width.
F i g u r e 1. A contour/gray scale image of interpolated Sea Beam

bathymetry between 42.552-43.567°W and 28.917-29.58ΓΝ, approxi-
mately an area of 100 km X 70 km. Depth ranges from approximately
4000 m (darkest areas) to 1200 m (lightest areas) and contours are shown
in 560 m intervals. The white areas represent no data, and there are
two no-data islands within the image: a large island centered at latitude
point 50 and longitude point 200; and a small one centered at latitude
point 100, longitude point 250. The nearly-vertical dark central valley
runs from south to north and is offset to the east at latitude point num-
ber 150. The crestal mountains are the light gray areas to the east and
west of the central valley before and after the offset zone.
T h e first step in the design of a suitable wavelet transform for image

analysis is the selection of a proper wavelet which, when convolved with
the d a t a , will enhance features of interest. A wavelet needs to b e a function
which integrates to zero and is square integrable [3]; (see also K u m a r and
Foufoula-Georgiou, this volume). There are a number of wavelet classes
in t h e literature which have been shown to meet these criteria and have
been used in d a t a analysis. These include the Morlet wavelet [11], t h e
Haar wavelet [4], [6], the B-spline wavelets [1], [2], [4], [8], [16], and t h e
Daubechies wavelets [3], [4].
172 S. Little
Bathymétrie profile at lat. point 300

-1000
-1500
g -2000
f, -2500
Û
-3000
-3500 Central valley
-4000
100 200 300 400 500
Sample point number, Longitude
F i g u r e 2. A bathymétrie profile running from west to east across the

image at latitude point 300 showing the profiles of ridges associated
with faulting. The plateaus at the far left and right are zones of no
data. From left to right, the profile comes up the flanks of the western
crest al mountains, crosses over seven or so major ridges associated with
faulting in the mountains, runs down to the central valley, up the eastern
crestal mountains and across five or so major ridges.
In the 2-D bathymetry d a t a under analysis, the features of interest

are relatively long, linear and narrow ridges with a preferred orientation
parallel to the axis of spreading. These fault blocks resemble straight lines
with cross sections t h a t look like ridges. An example of a profile across the
MAR, perpendicular to the axis of spreading, is shown in Figure 2. The
high plateaus at the extreme left and right are artifacts where no bathy-
métrie d a t a exist. Moving from left to right, one travels up the flanks, over
the western crestal mountains, across approximately 7 major fault ridges
(visible at this Sea Beam resolution), down into the central valley, up the
other side and over the eastern crestal mountains, crossing approximately
5 major fault ridges, and finally reaches the edge of available data. T h e
cross-section of these fault ridges suggests t h a t an edge detector would be
a good choice for a wavelet. For visual image enhancement it is desirable
to "bring out" the ridges, so t h a t they are easily seen by eye. For this, a
wavelet which would transform a given size class of ridges into highs and
valleys into lows would create an image where the ridges and valleys would
visibly stand out. One such detector for this purpose is a linear B-spline
wavelet [4], [2], [16]. For quantitative analysis, however, it is desirable to
have a wavelet which will transform the location of the fault scarp—that is
the transition from a low to a high (or vice versa) into a single high. T h e
filter formed from the derivative of the cubic B-spline achieves this [1] and
moreover, it can distinguish between a left facing edge and a right facing
edge (in profile), as described below and in Figure 3. This is very useful
for the analysis of fault distribution at the M A R because faults facing in

towards the central valley have been observed to be structurally different
from faults facing outward.
These two types of edge detecting filters will now be discussed.
3 . 3 . Linear B - s p l i n e w a v e l e t
Spline wavelets are constructed by a shifting, weighting and summing
of the B-spline functions, Ni(x), where x is the spatial variable, and i is
the order of the spline. Details of the construction of the linear and other
B-spline wavelets can be found in [2],
T h e linear B-spline function, N2(x), is defined as:
r o x <0
0<x < 1
1 <x < 2
ι o x>2.
T h e linear B-spline wavelet, W2, is formulated from N2 as follows:
W2 = ^[N2(2x)-W2(2x-l) + 10N2(2x-2)-W2(2x-3) + N2(2x-4)].
Figure 3a shows this wavelet (at 16 points) and the 1-D convolution
of this wavelet with three, noise free, 1-D ridges of different sizes. Notice
t h a t , in the transform, each step edge is converted to a low and a high.
T h e ridge whose width is most similar to the central portion of t h e wavelet
is transformed into a single high with two lower amplitude lows on either
side. This creates contrast which will visually "bring out" ridges.
T h e 2-dimensional filter is formulated to take advantage of the strongly
linear and oriented n a t u r e of the faults. Each linear ridge is qualitatively
like a long series of 1-D ridge-edges stacked together. Therefore, a 2-D filter
can be built by stacking up a number of 1-D ridge-edge detecting wavelets.
This is a departure from strict wavelet construction because the ridge-edge
parallel direction of this 2-D filter is not a wavelet. T h e 2-dimensional
version of t h e filter is specifically created in the following way: generate a
1-dimensional linear B-spline wavelet of a desired length; replicate this filter
to create a 2-dimensional matrix of identical filters; taper the edge-parallel
direction with a Hanning window [12]; and rotate the entire m a t r i x to t h e
desired orientation. A 3-D plot, contour/gray scale plot, and horizontal and
vertical profiles are presented in Figure 4 for this filter. Figure 4e shows
the filter rotated by —110 degrees from horizontal to match the dominant
orientation of the faults in the bathymetry. T h e length of the 1-dimensional
B-spline (BS) determines the width of the faults to be isolated. T h e number
of identical filters (NF) in the 2-dimensional matrix determines t h e fault
length over which to average (or smooth). If N F is large, then very long
174 S. Little
linear B-spline wavelet deriv. cubic B-spline

1 0.2
0.5 0
0
-0.5 -0.2
5 10 15 5 10 15
40 point ridge 40 point ridge

/ \
_r w
■V V -2 V
3 0 20 40 60 80 100 0 20 40 60 80 100
4js 10 point ridge 10 point ridge

If
0 _T~L _T~L_
-1
-2 F nAV
0 20 40 60 80 100 0 20 40 60 80 100
3 point ridge 3 point ridge

JL
4r -"V-
0 20 40 60 80 100 0 20 40 60 80 100
Sample Number
(a) (b)
F i g u r e 3. a) The top of the figure shows a 16 point linear B-spline

wavelet. Below is shown this wavelet convolved with ridge-edges of three
different sizes. The solid lines are the edges, and the solid-dot lines are
the resultant convolution. High amplitude is associated with narrow
ridges, and the edges of wide ridges. In addition, the high amplitude
is preceded by a low amplitude, which enhances visual contrast at the
edges, b) The top of the figure shows a 16 point derivative of a cubic B-
spline function. Below this is shown this filter convolved with the same
ridge-edges as shown in a). Notice that left-facing steps transform to
highs, while right-facing steps transform to lows. The location of these
edges is marked by the local maximum of the transform.
and straight faults will be identified; if N F is small, short and curved faults
will be identified as well as long and straight. A small N F , although it picks
up b o t h short and long faults, also picks up short curved features which are
volcanic rather t h a n faults, and hence there is a trade-off between getting
all the curved faults and getting too many short, non-fault features.
T h e 2-D linear B-spline filter, with BS=16 and N F = 7 is convolved, via
the 2-D Fourier transform, with the Sea Beam bathymetry and the results
shown in gray scale in Figure 5. Light areas correspond t o topographic
ridges, and dark areas correspond loosely to valleys. T h e fabric of t h e
3-D plot B-spline direction, BS=16
3 0.5
1 o (c)
< -0.5
5 10
Sample point
(a) NF direction, NF=7
(d)
BS direction 2 4 6
Sample point
Rotated-110
(b) (e)
Figure 4. a) A 3-D mesh plot of the 2-D linear B-spline filter with
BS=16 and NF=7. b) Contour/gray scale plot of this filter, dark is
high, c) Profile of this filter in the B-spline direction, d) Profile of
this filter in the NF direction, e) Contour/gray scale plot of this filter
rotated by —110° to match the dominant orientation of MAR-parallel
fault ridges.
topography is brought out visually in this image. Note t h a t although this

is a small scale wavelet, it is still possible to see large scale topographic
trends, e.g. the central valleys and crestal mountains. This is due to t h e
fact t h a t t h e 2-D filter used was not exactly zero mean. This filter makes
the image easier to study because most of the large scale topography has
been removed, but just enough allowed through to visually locate small
scale features in relation to the large trends such as the median valley and
crestal mountains. For instance, approximately 7 large bright linear ridges
are clearly on the left of the central valley, and fade out as they reach
the offset zone at latitude point 150. This filter transforms t h e d a t a to
an image which helps the viewer gain a qualitative understanding of t h e
topography.
176 S. Little
Figure 5. Application of filter shown in Figure 4e to bathymétrie data

shown in Figure 1. This is a filter for visually enhancing the image to
improve intuitive understanding of the topography. The lighter areas
correspond to bathymétrie ridges. The large scale topography is still
slightly visible, the dark central valley and lighter crestal mountains, due
to the non-zero mean of the filter. This enhances visual interpretation of
the image by enabling correlation between small scale ridges and larger
features while removing most of the effects of the large scale topography.
3.4, D e r i v a t i v e of cubic B-spline

Quantitative analysis of fault location, spacing and distribution re-
quires more t h a n the intuitive information gained from image enhancement.
T h e derivative of the cubic B-spline provides a filter which transforms a
step u p into a positive peak and a step down into a negative peak. This
provides a mechanism for locating fault scarps, t h a t is, the transitions from
valley to ridge (or vice versa), and classifying their facing direction. T h e
3-D plot B-spline direction, BS:
4 6
(a) Sample point
NF direction, NF=7
(d)
BS direction 10 0 NF direction 2 4 6
Sample point
Contour Rotated-110
(b) (e)
Figure 6. a) A 3-D mesh plot of the 2-D derivative of the cubic B-

spline function with BS=8 and NF=7. b) Contour/gray scale plot of
this filter, dark is high, c) Profile of this filter in the B-spline direction.
d) profile of this filter in the NF direction, e) Contour/gray scale plot
of this filter rotated by —110° to match the dominant orientation of
MAR-parallel fault scarps.
cubic B-spline function, N^x), is defined as:
0 £<0
0<x < 1
1 <x < 2
N4(x) = { § - ( x - 2 ) 2 + I(*-2)3 2< x < 3
|_I X + I ( X _3)2_I( X _3)3 3 < x <4
0 x >4.
A sixteen point long, 1-D filter created from the derivative of this
function is shown in Figure 3b. This filter is convolved with ridge edges of
three different sizes to show how it transforms a rising slope to a numerical
high, and a falling slope to a numerical low. In this way it locates fault
scarps and identifies t h e direction they face.
T h e 2-D filter is created from the derivative of the cubic B-spline t h e
same way as for the linear B-spline wavelet: Generate a 1-D filter of a de-
178 S. Little
Figure 7. A gray scale plot of the application of filter shown in Figure

6e to bathymétrie data shown in Figure 1. Dark corresponds to right-
facing fault scarps and light corresponds to left-facing fault scarps. This
image does not correspond to illuminated bathymetry and apparent 3-D
texture is misleading. The fault scarps predominantly face in toward the
central valley, dark on the left and light on the left. The faults also fade
out as they reach the offset zone.
sired length; replicate this filter to create a 2-dimensional matrix of identical

filters; taper the edge-parallel direction with a Hanning window; and rotate
the entire matrix to the desired orientation. The resulting filter is shown
in Figure 6 in a 3-D plot, a contour/gray scale plot, and horizontal and
vertical profiles. Figure 6e shows the filter rotated by —110 degrees from
horizontal to match the dominant fault scarp orientation.
A BS=8 and NF=7, 2-D derivative of a cubic B-spline filter is now
convolved with the bathymetry and the results, shown in Figure 7, show
excellent fault scarp identification. Dark gray corresponds to faults which
face easterly (to the right) and light gray corresponds to faults which face
westerly (to the left). The apparent surface texture in this figure is some-
what deceiving because it appears at a glance to be illuminated bathymetry.
It is not, as there are no true shadows in which features are hidden and
Right-Facing Scarps: BS=8, NF=7, cubic spline derivative
■' /
100 200 300 400 500 600 700

Sample point number
Figure 8. A plot of large-amplitude local minima (in horizontal rows)

of the transform shown in Figure 7. The local minima correspond to
the centers of right-facing fault scarps. There is a predominance of large
right-facing faults on the west side of the central valley, both below and
above the offset zone.
the 3-D effect is somewhat misleading. However, this d a t a can b e pro-

cessed quantitatively. A plot of large-amplitude local minima (calculated
in west-to-east lines) is shown in Figure 8. This figure shows the location
of large faults which face easterly. There are more large east facing faults
on t h e left-hand side of t h e central valley. Figure 9 shows large-amplitude
local maxima, and clearly identifies more large west facing faults on t h e
right-hand side. This implies t h a t there are larger faults facing in toward
the central valley t h a n facing away. This characteristic of the M A R has
been noted before from visual observations of bathymétrie maps. T h e 2-D
derivative filters offers a chance to quantify this qualitative observation.
Figure 10 shows b o t h left and right facing fault scarps, which gives a good
picture of t h e distribution of faulting, and also shows t h a t the faults fade
out in t h e offset zone which runs from the upper left, latitude point number
150, to the lower right, latitude point number 200.
180 S. Little
Left-Facing Scarps: BS=8, NF=7, cubic spline derivative
100 200 300 400 500 600 700

Sample point number
Figure 9. A plot of large-amplitude local maxima (in horizontal rows)

of the transform shown in Figure 7. The local maxima correspond to
the centers of left-facing fault scarps. There is a predominance of large
left-facing faults on the east side of the central valley, both below and
above the offset zone.
Bathymétrie d a t a from the seafloor contain the superposition of ridges,
valleys, and volcanos at many different scales. Wavelet analysis offers a use-
ful m e t h o d for decomposing the texture of the seafloor to help understand
processes which occur at many different scales. Much of the bathymétrie
d a t a available is in the form of 1-D spatial series. Scale decomposition
of these series has been shown to yield an understanding of geophysical
processes which originally were overlooked in the raw data. Smaller por-
tions of the seafloor have been swath-mapped, and these 2-D d a t a sets
offer an unusually good opportunity to decipher the complexity of seafloor
topography. This paper has presented two wavelet techniques for working
with swath-mapped bathymetry, one a qualitative image enhancement, and
the other a quantitative fault scarp identifier. A linear B-spline wavelet is
used to design a 2-D filter which produces an image in which the linear,
MAR-parallel fault ridges are easily seen in the context of the broader to-
All Scarps: BS=8, NF=7, cubic spline derivative
50
100
150
I
S 200
350
400
450
ι - , t1 .
100 200 300 400 500 600 700
Sample point number
F i g u r e 10. A plot of both local minima and local maxima, showing

the locations of fault scarps, regardless of facing direction. These data
can be used for quantitative analysis of fault spacing, orientation and
distribution.
pographic undulations. A second filter, designed from the derivative of a

cubic B-spline, is used to obtain quantitative information on the locations
and facing directions of fault scarps, t h a t is, the zones where t h e crustal
material was actually sheared. This quantitative information can b e fur-
ther processed to obtain fault spacing, preferred orientation, density and
distribution, parameters which can be used to understand the underlying
geophysical processes which occur at the MAR.
References
1. Canny, J., A computational approach to edge detection, Trans, on
Pattern Anal, and Machine Intelligence, 8(6), 679-698, 1986.
2. Chui, C. K., An Introduction to Wavelets, Academic Press, New York,
1992.
Comm. Pure Appl. Math. , 4 1 , 909-996, 1988.
4. Daubechies, I., Ten lectures on wavelets, CBMS Regional Conference
182 S. Little
Series in Applied Mathematics, SIAM, Philadelphia, PA, 1992.

5. Grossman, A. and J. Morlet, Decomposition of a Hardy functions into
square integrable wavelets of constant shape, SIAM J. Math. Anal.,
15(4), 723-736, 1984.
6. Haar, A., Zur Theorie der orthogonalen Funktionen-Systeme, Math.
Ann., 69, 331-371, 1910.
7. Little, S. A., P. H. Carter, and D. K. Smith, Wavelet analysis of
a bathymétrie profile reveals anomalous crust, Geophys. Res. Letts.,
20(18), 1915-1918, 1993.
8. Mallat, S., and Zhong, S., Characterization of signals from multiscale
edges, IEEE Trans, on Pattern Anal, and Machine Intelligence, 14(7),
710-732, 1992.
9. Mammerickx, J., Depth anomalies in the Pacific: Active, fossil, and
precursor, Earth Plant. Sei. Letts., 53, 147-157, 1981.
10. Mammerickx, J., and D. Sandwell, Rifting of old oceanic lithosphère,
J. Geophys. Res., 91(B2), 1975-1988, 1986.
11. Morlet, J., Sampling theory and wave propagation, in Issues in acoustic
signal/image processing and recognition, C.H. Chen (ed.), NATO A SI
Series, Vol. 1, Springer-Verlag, Berlin, 233-261, 1983.
12. Oppenheim, A. V., and R. W. Schäfer, Digital Signal Processing,
Prentice-Hall, 556, 1975.
13. Purdy, G. M., J.-C. Sempere, H. Schouten, D. DuBois, and R. Gold-
smith, Bathymetry of the Mid-Atlantic Ridge, 24°-31°N: A map series,
Mar. Geophys. Res., 12, 247-252, 1990.
14. Shaw, P. R., Ridge segmentation, faulting, and crustal thickness in the
Atlantic, Nature, 358, 490-494, 1992.
15. Shaw, P. R., and J. Lin, Causes and consequences of variations in
faulting style at the Mid- Atlantic Ridge, J. Geophys. Res., 98, 21,839-
21,851, 1993.
16. Vinskus, M., The discrete wavelet transform with applications to edge
detection, Master's thesis, Dept. Electrical and Computer Engineering,
Worcester Polytechnic Institute, 1993.
This work was supported by grant N00014-93-1-0179 from the Office
of Naval Research. WHOI contribution number 8617.
Sarah A. Little
Department of Geology and Geophysics
Woods Hole Océanographie Institution
Woods Hole
MA 02543, USA
A n a l y s i s of H i g h R e s o l u t i o n M a r i n e S e i s m i c D a t a U s i n g t h e
W a v e l e t Transform
Chris J. Pike
Abstract. A method is proposed for the analysis of high resolution acoustic

signals using the wavelet transform. For signals with intensities exhibiting fre-
quencies ranging from 100 Hz to 10 kHz, time-frequency displays of the signal
can lead to more robust means of estimating attenuation as well as quantification
of wavefield scatterers within shallow marine sediments. The time-frequency de-
compositions of the signals are accomplished using the wavelet transform and a
Morlet analyzing wavelet. Zero offset acoustic signals are analyzed and the mod-
ulus of the wavelet transform is displayed as a function of depth below seabed
versus wavelength. The results are discussed and related to the subseabed soil
conditions at an experimental field site.
T h e Centre for Cold Ocean Resources Engineering (C-CORE) is an
independently-funded research institute of Memorial University of New-
foundland in St. John's, Newfoundland, Canada. C-CORE is involved in
solving engineering problems related to resource development in the ocean
environment. T h e Centre has traditionally focused on a limited number of
research areas related to ocean resources and has built up expertise reflected
by t h e four research groups operating at C-CORE. They are t h e R e m o t e
Sensing group, t h e Geotechnical Engineering group, the Ice Engineering
group and t h e Seabed Geophysics group.
One of the areas of research engaged by the Seabed Geophysics group
has been t h e high resolution acoustic investigation of the sub-seabed. A
central goal of the acoustic program is the development of methods for
extracting marine soil properties by correlating acoustic properties derived
from acoustic signals with the geotechnical properties of the soil. During
t h e course of this research it was evident t h a t the high frequency acoustic

ISBN 0-12-262850-0
184 C. Pike
signals (500 Hz to 10,000 Hz) used were rapidly attenuated, especially the
higher frequencies, in those soils that contained many scatterers - boulders
and cobbles with dimensions comparable to the wavelengths of the acoustic
waves generated by the source. Spectra for the acoustic signals are usually
calculated by using the Fourier transform (FT) or the fast Fourier transform
(FFT). From the equation for the Fourier transform it is evident that this
equation is not well suited to the study of localized (in time) disturbances
in a signal nor is it appropriate to apply it to non-stationary time series.
The power spectrum of an acoustic signal is often analyzed in an attempt
to investigate attenuation of an acoustic signal. Techniques that compare
the spectra of the emitted and transmitted or reflected pulse are used to
assess attenuation as a function of frequency. The typical frequency range
for explortion seismic data is between 10 and 100 Hz. Over this range,
and for the length of duration of the seismic record, the basic assumptions
of stationarity are not grossly violated. High resolution seismic signals
can often exhibit severe loss of high frequencies as they travel down into
the earth and return again. A method that reflected the time-frequency
structure in high resolution signals would be more diagnostic and more
appropriate for the analysis of very high resolution signals.
Experiments designed to measure attenuation within the earth some-
times use down hole geophones to record the direct pulse from a source
near the surface or a core sample may be taken to the laboratory where
a high frequency acoustic wave is transmitted through the sample. If the
outgoing pulse and the transmitted pulse can be isolated then the spectra
can be compared. The isolation of the first arrival is usually accomplished
by using a short window Fourier transform (SWFT) and often the spectra
exhibit properties associated with the window function used as well as the
signal. The wavelet transform offers a means of avoiding this problem.
High resolving power of the acoustic wave is desirable for a range of
depths but this ability is compromised when multiple scatterers are present.
Morlet et al. ([21] and [22]) investigated the propagation of plane waves for
normal incidence, through periodic multi-layered media for wavelengths
ranging from much greater than the spatial period to periods on the or-
der of the spatial period. They found that for large wavelengths (16 times
the spatial period of the medium) the composite medium was transpar-
ent but phase-delaying; for short wavelengths (2 times the spatial period
of the medium) the signal was strongly attenuated and super-reflectivity
could occur; for intermediate wavelengths (8 times the spatial period of the
medium) velocity dispersion versus frequency appeared.
These observations are quite diagnostic of the medium, offering the geo-
physical investigators of shallow marine soils a new and potentially pow-
erful means of quantifying subseabed soil conditions through the use of
time-frequency methods. Morlet et al. ([21] and [22]) conclude that when
High Resolution Marine Seismic Data Analysis 185
frequency bandwidth spans four or five octaves, time-frequency m e t h o d s

will b e more sensitive to attenuation and velocity dispersion.
T h e wavelet transform ([21], [22], [6], [17], [27], [5] and [7]) is an appro-
priate tool for investigating the time-frequency n a t u r e of high resolution
acoustic signals. Displaying the energy distribution in the time-frequency
plane describes a signal more uniquely t h a n the F T spectra and could
provide a means for evaluating or quantifying the degree of backscatter
in different frequency bands and at different times. A complete m a p of
attenuation as a function of depth and frequency or wavelength can be
constructed and this type of information may be more readily correlated
to selected marine soil properties.
T h e main objective of this paper is to report on the application of
the wavelet transform to high resolution marine acoustic signals and its
use in t h e estimation of attenuation of acoustic wave energy for signals
reflected from subseabed reflectors. Usefulness of the wavelet transform
as a classifier for backscatter is discussed briefly with reference to wavelet
transform plots and depth-wavelength displays.
§2. A c o u s t i c - G e o t e c h n i c a l C o r r e l a t i o n s : P h y s i c a l a n d H i s t o r i c a l
Context
T h e motivation for research into correlations between acoustic/seismic
responses of a marine soil and the soil's geotechnical properties is to provide
the geotechnical engineer with cost-effective and reliable estimates of t h e
geotechnical properties of submerged soils. Reliable estimates of geotech-
nical properties are needed, for example, in the selection of a site and t h e
design of a structure t h a t is intended to be supported by the seabed. Such
structures can include gravity-based oil production platforms or support
columns for causeways or bridges. Porosity, density and grain size distri-
bution are some of the soil properties t h a t affect the acoustic response of
a signal and are also of interest to the geotechnical engineer [31]. Acoustic
wave properties related to these parameters through empirical relationships
include acoustic impedance, velocity and attenuation ([1], [9], [20] and [13]).
These properties can be determined from digital acoustic d a t a with greater
accuracy t h a n was possible in the past using analogue records.
T h e recent and rapid development of smaller, faster, and more power-
ful computers has had an impact upon many branches of science b u t none
so dramatic, perhaps, as in the area of high resolution sub-bottom marine
seismic reflection profiling. At about the same time as the first papers on
wavelet transforms were being written, analogue paper recordings of acous-
tic returns from t h e seabed and sub-seabed represented t h e current practice
in d a t a collection and presentation. Practitioners of land-based engineer-
ing seismic d a t a acquisition, and the oil and gas exploration industry, have
186 a Pike
been acquiring and processing digital data for almost as many years as there
have been computers. During the past ten years the marine high-resolution
(> 500 Hz) or site surveying industry has made the leap from analogue data
acquisition systems to digital data acquisition systems. This technological
leap can be attributed to the development of high-speed analogue-to-digital
computer boards and digital signal-processing chips, which together effec-
tively allow real-time data acquisition and processing of high-resolution
acoustic data.
The availability of digital acoustic data has led to the desire for a more
quantitative assessment of seismic signals. Empirical relationships between
acoustic properties and soil properties have been reported ([1], [32] and
[10]). Studies that relate acoustic properties to rock or sediment properties
involve the use of frequencies below 100 Hz ([26] and [12]) or above 10
kHz ([10] and [11]) but very few studies provide information in the 100
to 10,000 Hz range which is the typical range for many high-resolution
sub-bottom marine surveys. In soils characterized by large scatterers, such
as boulders in glacial tills, the 100-1,000 Hz range of frequencies is more
efficient in achieving greater depth penetration. Testing carried out in
laboratories, usually at very high (> 100 kHz) frequencies, on samples
collected in the field or on cone-penetrometer data acquired in situ, provide
data for correlations that are sometimes biased toward high frequencies.
But the means to correlate geotechnical properties of a soil over a more
complete range of frequencies is also needed. Some commercial acoustic
systems already offer classification schemes based upon how the transmitted
pulse is modified by the water-sediment interface upon reflection [18] or
internally upon transmission to deeper layers and subsequent return to the
receiver [29]. It is the seafloor sediment type that is most easily determined
by evaluating the seabed return as in Leblanc et al. [18] but characterization
derived from deeper returns becomes quite complex due to constructive and
destructive interference by scattered waves. Time-frequency analyses would
aid in assessing the degree of interference as well as capitalizing upon the
tuning or detuning of acoustic waves that can occur within the sub-bottom
layers.
There are several factors that contribute to the attenuation of an acous-
tic wave. Acoustic energy attenuates with increasing distance from the
source, a function described by the term geometrical spreading loss. Ge-
ometrical spreading losses are not related directly to the properties of the
medium in which the energy is propagating. Usually a means of correcting
or nullifying this effect is sought before attenuation due to the medium is
determined. Attenuation of acoustic energy can be divided into two basic
categories, intrinsic attenuation and apparent attenuation. Intrinsic atten-
uation is the process that degrades the amplitude of the signal such as
absorption due to frictional heating or viscous losses [14]. Apparent atten-
uation is due to processes like scattering, energy partitioning at acoustic

impedance contrasts, interbed multiples and mode conversions ([14] and
[30]). W h a t is more generally observed, however, is effective attenuation
which is a combination of intrinsic attenuation and apparent attenuation.
Estimation of intrinsic attenuation of propagating acoustic pulses can
provide information concerning grain size, moisture content, shear modulus
and Atterburg limits [13]. Estimates for effective attenuation are useful as
they can be related to the degree of scattering, and possibly to quantifi-
cation of t h e size of the scatterers within a soil. There are many m e t h o d s
for measuring or estimating attenuation of acoustic energy in sediments.
D a t a collected for the determination of attenuation generally fall into two
categories, in situ measurements and laboratory measurements, T h e fre-
quency content of the acoustic energy used to investigate the sediment can
be vastly different - from several hertz to millions of hertz. T h e connection
between t h e range of frequencies and measurements is not always clear.
Laboratory methods generally rely upon free vibration, forced vibra-
tion, wave propagation and observation of stress-strain curves [35]. Labo-
ratory studies involving wave propagation methods use frequencies in t h e
100 kHz to 1 MHz range [34]. In situ measurements t h a t rely upon wave
propagation can range from a few hertz to hundreds of kilohertz b u t as
noted in the preceding text there is a need for more information in t h e
range of frequencies between 100 Hz to 10 kHz range.
T h e estimation of attenuation using field or in situ measurements has
spawned many methods. In some situations it is possible to record t h e
outgoing acoustic pulse as a reference for comparison with the energy from
acoustic or seismic waves refracted through the sediments and returned t o
the surface to be recorded by geophones or hydrophones ([15], [14], [8] and
[33]). More reliable measurements can be made by drilling a hole down
through the sediment and recording the direct arrival of acoustic energy
from a source near the surface to a geophone placed at successively deeper
depths within t h e hole ([12], [24] and [26]). Similarly measurements can
be m a d e between two holes drilled through the rock or sediments [25] by
placing a source in one hole and a receiver in the other and varying the
depth of each.
Attention has been focused on extracting attenuation estimates from
seismic reflection data, particularly in the marine setting ([18], [29], [3] and
[36]). Schock et al. [29] estimate attenuation in the seabed by comparing
the correlated chirp signal from a reflector with a synthetically a t t e n u a t e d
chirp wavelet. From this measurement the frequency roll-off is determined
and plotted as a function of depth. The slope then gives the attenuation
coefficient in units of d B / m kHz. LeBlanc et al. [18], also using a chirp
system, a t t e m p t to estimate t h e surficial soil properties of the seabed by
correlating reflection losses at the water sediment interface with soil prop-
188 C. Pike
erties. Tyce [36] took advantage of naturally occurring sedimentary wedges

to estimate attenuation. This situation is useful because the decrease in
energy is easily observed and calculated as the wedge of sediment thickens.
A common technique for characterizing attenuation of acoustic signals
utilizes the spectral ratio between two pulses ([15], [14] and [12]). This
technique uses the ratio of the amplitude spectrum of the known unat-
tenuated source function with the amplitude spectrum of the a t t e n u a t e d
amplitude spectrum. A slope of this ratio against frequency yields the
attenuation, and terms t h a t are assumed to be independent of frequency
such as geometrical spreading or reflected energy can be ignored. A more
recent technique, introduced by Courtney and Mayer [2], employs a filter
correlation m e t h o d t h a t uses a series of bandpass filters on the a t t e n u a t e d
pulse and compares each pulse to the similarly bandpassed source pulse.
§3. T h e W a v e l e t Transform for A c o u s t i c Soil A n a l y s i s

3.1· T h e w a v e l e t t r a n s f o r m and M o r l e t w a v e l e t
T h e form of the wavelet transform used in this application is given by
Equation (6) of Goupillaud et al. [6]. Adopting the terminology common
to this volume the wavelet transform of / with respect to the analyzing
wavelet is:
(W f)(u, a) = |a|-i J Φ ( ^ ) / ( ί ) dt. (1)
Where f(t) is the time series or signal, Φ(£) is the analyzing wavelet and
^(^ϊτ) represents the dilated (a) and translated (u) wavelets. T h e | a | - 1 / 2
term serves to ensure t h a t the dilated wavelets have the same total energy
as the analyzing, or mother wavelet. In the analyses presented in this
paper, frequency or the natural logarithm of the frequency, is given by the
vertical axis and the time shifts are given by the horizontal axis for the
time-frequency displays. The frequency associated with each scale value,
a, was determined directly from the equation for the Morlet wavelet [6].
T h e Morlet wavelet used in these examples is described by the following
equation,
φ 6 ( ί ) = eibte~^ - V2e-êibte-t2. (2)
T h e Fourier transform of Equation (2) is
1
FT(9b(t)) = Φ 6 (ω) = e ^ ^ - e~* e ^ ^ . (3)
This wavelet is shown in Figure 1. The envelope of the wavelet and its
amplitude spectrum are Gaussian-modulated functions. The variable b in
Equations (2) and (3) has the same value of 5.336 and is the same as t h a t
given by Goupillaud et al. [6].
Figure 1. Morlet analyzing wavelet.
3.2. Estimating attenuation: a simple example

T h e following example demonstrates the mechanics of the attenuation
analysis. A synthetic trace with the direct pulse from one of the normal inci-
dence traces collected at an offshore experimental field site in Terrenceville,
Newfoundland (described later), was convolved with a series of five spikes
placed at the 0.0003, 0.002, 0.004, 0.006 and 0.008 second time marks. A
geometrical spreading factor was then applied as well as an attenuation
factor, a, as follows:
A(t) = G(t)A0e-°vt. (4)
T h e geometrical spreading factor, G(£), is given by 1/vt where v is a con-
stant velocity of 1500 m / s in this example, but it can also be a function
of depth, and t is time in seconds. T h e definition used for the attenuation
coefficient, a, is t h a t given by Johnston and Toksöz [16] and it is assumed
to b e linear with frequency for the range of frequencies under investigation.
For t h e synthetic example a was assigned a value of 0.25 m _ 1 . In Figure 2a
the 0.010 second attenuated trace (without spreading losses) is displayed.
Figure 2b depicts the modulus of the wavelet transform, with the horizon-
tal axis representing time and the vertical axis the natural logarithm of t h e
frequency. T h e values range from —13.55 (black) to 10.35 (white). For all
the wavelet analyses presented the axes are labelled in this manner unless
otherwise noted.
Figure 2c depicts the spectral ratio for the wavelet transform of t h e
synthetic signal. T h e spectral ratio is a measure of t h e ratio between t h e
s p e c t r u m for the direct or emitted pulse and the spectrum for t h e reflected
of t r a n s m i t t e d pulse. Values range from —22.87 dB (white) to —2.47 dB
(black). This display, as well as the method used to calculate the spectral
ratios ([34] and [16]), bears further explanation by way of presenting t h e
190 C. Pike
0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
Time (sec)
F i g u r e 2 . Synthetic example; (a) 0.25 d B / m a t t e n u a t e d synthetic seis-

mic t r a c e (the direct water borne arrival from t h e sparker was used as
t h e wavelet for t h e synthetic); (b) modulus of t h e wavelet transform of
synthetic t r a c e in (a); (c) spectral r a t i o calculated from wavelet t r a n s -
form in (b).
relevant part of the derivation of the spectral ratio equation. After John-
ston and Toksöz [16], a plane wave propagating in a homogeneous medium
can be described by:
A(x,t) = A0e^kx-^ (5)
where attenuation may be represented by allowing either A:, the wavenum-
ber, or ω, the angular frequency, to be complex. Following [16] and in-
cluding a geometrical spreading term as in [34], the ratio of the reference
amplitude spectrum versus the spectrum at some time, ίχ, after taking the
natural logarithm of the ratio of the two amplitude spectra can be written
as:
A G
In —^ = (7 m - 7i)vtf + In -^ (6)
A\ Cxi
where v is velocity, t is one-way travel time, / is frequency, a(f) = 7 / is
the attenuation at / , and the left hand side of Equation (6), the spectral
ratio, is represented by the data displayed in Figure 2c. Equation (6) can
be interpreted in two ways; first the free variable can be the frequency, / ,
and then the slope, of a line fitted to the data when the spectral ratio is
plotted against frequency, would be:
slope = (7 m - 7i)ttf. (7)
The geometrical term, ln(G? i/G2), can be ignored if one assumes that the
geometrical spreading terms are independent of frequency. Alternatively,
travel time can be taken as the free variable but in this case the geometrical
factor does depend upon t and it will influence the value of the attenuation
derived from the slope of a line fitted to the curve of the spectral ratio
versus time calculation.
For this study the spectral ratio method was applied by taking the
frequency distribution for the time at which the maximum magnitude for
the direct pulse is greatest and using this spectrum as the reference spec-
trum. This occurs at approximately 0.0007 sec. in Figure 2b. Therefore,
any determination of the slopes of lines fitted to the data in the direction
parallel to the vertical axis satisfies Equation (7) and those that are fitted
to the data in the direction parallel to the horizontal axis are subject to
the time-dependency condition.
The procedure for estimating the attenuation from normally-incident
acoustic waves is demonstrated using the simple synthetic seismic example
described in the preceding text. The results of a comparison of the value
of the attenuation obtained using the synthetic trace corrected for geomet-
rical spreading losses and an uncorrected synthetic trace are displayed in
Figures 3a and 3b respectively. The synthetic trace was corrected for the
known geometrical spreading factor and the wavelet transform was then
calculated and the modulus displayed (Figure 2b). The mean power at
192 C. Pike
each time and for all frequencies in the transform was calculated and plot-
ted in Figure 3a. There are five broad peaks shown in Figure 3a associated
with each of the attenuated pulses in Figure 2a. The largest local m a x i m a
for a particular broad peak was chosen interactively and a least squares
fit was made to these selected values. The criterion for selecting the local
maxima is based upon providing a guideline for consistent picking, and in
selecting the large broad peaks it was assumed t h a t boundaries or acoustic
reflectors will reflect energy independent of frequency content which would
not be the case for random scatterers. The line and the points chosen
for the least squares fit are indicated (straight black line and black x's re-
spectively). T h e value of the slope was found to be 360 nepers/sec which,
when converted to meters by dividing out the velocity term, yields a value
of 0.24 nepers/meter. The conversion between nepers/unit length and dB
(decibels) per unit length is a (dB/unit length) = 8.686π (nepers/unit
length). The actual value used was 0.25 nepers/meter, a 4% difference be-
tween the actual and estimated value. If the geometrical spreading losses
are not removed then there should be some effect on the result for the
attenuation calculation, and this is observed in Figure 3b. Using the spec-
tral ratio d a t a from Figure 2c, the variance in the direction parallel to
the frequency axis was calculated for every time sample. The slope of a
least squares fit to the user-selected points from this plot yielded a value of
277 nepers/sec or 0.185 nepers/meter, a 27% difference. When all digital
sample values are considered, the least squares estimate for the corrected
d a t a is 0.29 nepers/meter, a 16% difference. As a final comparison t h e
primary event was windowed (150-point window) as well as t h e event at
0.008 sec. T h e amplitude spectrum was calculated for each event using t h e
fast Fourier transform and then the difference was divided by the separa-
tion between the two events yielding a value of 0.22 nepers/meter or a 13%
difference between the actual and estimated value.
T h e results given in Figures 3a and 3b are very encouraging, and in
spite of the large difference (16%) for the case where all values for a cor-
rected transform were used an order of magnitude for the attenuation is
nevertheless achievable. For real acoustic signals the zones or valleys in
Figure 3b, where the spectral ratio falls below 5 nepers with respect to
the primary signal, the interference due to scattering of acoustic energy
would tend to fill in these valleys possibly resulting in a lower value for the
attenuation estimate when including all values in the estimation.
Application of the wavelet transform to the problem of determining
attenuation offers the potential for automated order-of-magnitude deter-
minations of attenuation from digitally-recorded acoustic signals. As well,
the potential to surgically isolate energy in time and frequency can enhance
its application for sub-bottom attenuation measurements using reflected
seismic signals.
Time (sec)
Figure 3. (a) Attenuation curve for synthetic trace with correction

for geometrical spreading losses and (b) without correction for spreading
losses. (Note: see text for conversion between nepers and decibels)
194 C. Pike
§4. Experimental Field Program

Three-dimensional multichannel data were collected at a seabed test
site at Terrenceville near St. John's, Newfoundland. These data were
processed and interpreted conventionally [23]. Within the circular 78 m 2
test site the marine sediment was between 4 and 5 m thick overlying an
undulose and fractured bedrock. The depth of the water at the site was 10
m. Three distinct programs were carried out at the test site, an acoustic
program, a physical sampling program and a geotechnical program.
The acoustic program comprised four acoustic lines that intersected at
45 degrees at the centre of the test site (Figure 4a). Figure 4b shows a
simple schematic of the source/receiver geometry which was manipulated
into the survey grid shown in Figure 4a by scuba divers from the Ocean Sci-
ence Centre at Memorial University of Newfoundland. The 16 hydrophones
were arranged in four banks of four hydrophones per bank for recording
purposes and are labelled A through D. The acoustic source was mounted
on a tripod and was positioned approximately 40 cm directly above each
hydrophone of bank D (stations 5 through 8). These four stations were the
shot point locations for each profile of the survey. The hydrophones posi-
tioned beneath the source are referred to as normal incidence hydrophones.
The horizontal spacing between receiver stations was 1.42 m. The average
height above the seabed for hydrophones comprising bank A and C was 1.7
m and for bank D it was 1.3 m. The average source height was the same
as that for hydrophone banks A and C.
The sparker source used was developed specifically for the high res-
olution seismic research program at C-CORE [4] to deliver a repeatable,
broadband acoustic pulse. Up to 1080 joules ( J) of energy was delivered to
this source using a Huntec seismic energy power supply. Brüel and Kjœr
(B&K) 8105 omnidirectional hydrophones, with a flat frequency response
from 0.1 Hz-180 kHz, were used to measure the pressure field. Brüel and
Kjœr 2635 charge amplifiers, with built-in antialias filters, conditioned the
signals prior to digitizing. The data were sampled at 10 microseconds with
antialias filters of 0.2 Hz (low cut) and 30 kHz (high cut). The automated
acquisition system directed the analogue-to-digital recorder to trigger the
seismic energy source and then record the acoustic signals. The emitted
acoustic pulse, recorded on the normal incidence hydrophone, was then
correlated with a previously recorded reference pulse. If the correlation
coefficient exceeded a user-defined threshold, set at 0.95 for this survey,
the trace was accepted and all recorded channels were transferred to the
acquisition computers memory. The automated procedure continued, and
traces were summed with their respective counterparts in memory, un-
til each trace in memory consisted of 50 stacked or summed traces. These
summed traces were then written to the computer's hard drive. The acous-
Figure 4. (a) Site plan for experimental field program and (b) cross-
section of seabed recording geometry.
196 a Pike
tic source was then moved to stations 6, 7 and 8 with the above acquisition
procedure repeated for each source position. After the shot at station 8 the
entire b e a m was rotated about its central point in 45 degree increments and
the recording process described above was repeated. The complete source
receiver coverage is shown in Figure 4a. The four recorded profiles are la-
belled Beams 1 to 4 and station number 1 is identified for each b e a m with
a total of twelve stations per beam. The outer ring of sparker source point
locations, at a radius of 2.13 m, defines the extent of the subsurface normal
incidence coverage. T h e common depth point (CDP) coverage is defined
by a circle of radius 4.87 m, passing through stations 3 and 10 on beams 1
through 4.
T h e entire procedure described above was carried out for two separate
power settings, 480 J and 1080 J. Thus two complete surveys were acquired
for this site, and are referred to as the low and high power surveys.
T h e physical-sampling program provided marine soils and bedrock
cores for analysis and a geotechnical program provided some in situ mea-
surements using a limited series of cone penetrometer tests (CPT's) as well
as a series of seismic cone penetrometer tests (SCPT's) to acquire velocity
information. Figure 4a shows the locations of the C P T ' s and boreholes.
T h e cone penetrometer was capable of recording shear (S) and com-
pressional (P) wave data. This information was acquired when the cone
penetrometer was stopped at 1 meter intervals as the cone was pushed into
the seabed. T h e compressional wave arrival times were used to calculate
average and pseudo interval velocity values for the soils. T h e horizontal
offset between the three acoustic sources and the cone rods and the off-
set of the geophones from the cone tip were accounted for in all velocity
calculations.
Three continuously-sampled boreholes were drilled by rotary techniques,
using N-size casing. Standard penetration tests were conducted and repre-
sentative b u t disturbed soil samples were obtained in the overburden, using
a conventional 51 m m OD split-spoon (SS) sampler. Bedrock was cored in
NQ size. All borehole depths reported are with respect to the sea floor.
The exact hole locations were measured with respect to the underwater
acoustic survey by divers.
§5. D a t a A n a l y s i s
Acoustic d a t a analysis was divided into two phases with the first fo-
cused on determining subsurface stratigraphy by using the multichannel
acoustic data, the subsurface samples and results from the penetrometer
tests. T h e second phase involved analysis of the normal-incidence d a t a and
was concerned with measuring and quantifying changes in the waveform of
the reflected energy as well as finding some means to visually compare t h e
16 normal-incidence acoustic signals. Use of the wavelet transform was

investigated for this particular analysis.
5 . 1 . C o n v e n t i o n a l d a t a analysis a n d p r o c e s s i n g
Core samples extracted from the site confirmed the composition of
the marine sediments for this area. T h e sediments consisted of a fine to
medium-grained gravelly sand with some organic content overlying a sand-
gravel deposit of between 4.2 and 4.6 m thick. Sandwiched between t h e
fractured, layered bedrock and the sand-gravel deposit was a thin ( < 1.0
meter) till. T h e depth to bedrock ranged between 4.7 and 5.1 m below t h e
sea-bot torn.
T h e cone penetrometer d a t a were of limited value as this tool is de-
signed for use with fine-grained soils. However, some useful d a t a included
the cone tip resistance which provided information on soil stiffness. T h e
seismic d a t a were used to estimate independently the compressional (P)
wave velocity of the soil.
D a t a from the acoustic program were processed using conventional
common depth point procedures for multichannel seismic analysis [38]. An
example of a brute stack from the Beam 3 profile (Figure 4a) is given in
Figure 5. Two principal acoustic horizons are easily observed; the seabed
event, a positive (black) peak (trace excursions to the right), at about 2.0
ms and a second event, a trough at about 5.0 ms, b o t h indicated by t h e
arrow heads on either side of the plot. The seismic section is displayed with
two-way vertical travel times versus horizontal position in the profile. There
are many other weaker events evident but below about 6 ms the events
become more disjointed and incoherent. Figure 6 shows (a) a b r u t e stack for
the Beam 3 profile, (b) filtered and gained normal incidence trace (repeated
four times as a visual aid), (c) stratigraphie indicators from Borehole 1 and
(d) the geotechnical field report of the samples showing stratigraphy. T h e
reflection even at 5.0 ms in Figure 6a correlates with a boundary t h a t was
not inferred from the sampling program possibly due to low sample recovery
rates (see Figure 6d). A boundary was evident in the geotechnical analyses
and its existence corroborated from the force needed to push t h e cone
penetrometer past this boundary. It is indicated by the triangle symbol
(Figure 6c).
5 . 2 . W a v e l e t t r a n s f o r m analysis of a n a c o u s t i c signal
Using a Morlet analyzing wavelet, the wavelet transform was applied to
the normal incidence d a t a from the field program. These d a t a are recorded
at t h e locations in Figure 4 where t h e sparker source and hydrophones are
coincident. Figure 7a shows the normal incidence trace from profile Beam 3,
station 5 with a geometrical spreading correction applied to the d a t a . T h e
geometrical spreading correction was estimated using the average velocity
198 C. Pike
F i g u r e 5. Brute stack for profile Beam 1 (refer Figure 4a). Bandpass

filtered and gained for display.
function determined in phase I of the acoustic field program. T h e modulus

of the wavelet transform is presented in Figure 7b and the spectral ratio is
shown in Figure 7c.
T h e primary or direct arrival has a direct travel time of 0.0003 sec (Fig-
ure 7a). T h e next significant event is the reflection from the seabed, which
arrives at 0.0021 sec, followed by another clear event arriving at about t h e
0.005 sec mark at 8.4 ln(Hz). There are no distinct events between t h e
0.0021 sec event and the event at 0.005 sec but there is high frequency
energy in the signal which may be related to internal scattering of acous-
tic waves by cobbles and small boulders buried in the seabed. Following
F i g u r e 6. (a) Brute stack for profile Beam 3; (b) filtered normal in-
cidence trace for station 5, Beam 3 (trace is replicated four times for
visual aid); (c) borehole 1 showing identified stratigraphie boundaries;
(d) soil profile description based on physical samples and cores.
the 0.005 sec event are several more pronounced excursions from t h e signal
and these could be due to scattering from the fractured bedrock or internal
multiple reflections.
T h e modulus of t h e wavelet transform for the acoustic signal (Fig-
ure 7b) exhibits bright areas (shown as white) where the signal strength is
strong (note the direct arrival and the seabed event in Figure 7a). There
is more detail represented in this display t h a n would be found in a con-
ventional power spectra display. T h e modulus of the wavelet transform
associated with t h e direct pulse exhibits three lighter areas (see Figure 7b,
between 7.5 and 10 ln(Hz) and 0.00025 and 0.001 sec). Comparison be-
tween the waveform of the direct pulse (Figure 7a) and t h e modulus of t h e
200 C. Pike
wavelet transform for this pulse (Figure 7b) demonstrates t h e strength of

the transform for signal decomposition. The direct pulse (Figure 7a) has
an onset time of 0.0003 sec. Following this onset is the first positive peak,
followed immediately by the first trough, and then a second smaller positive
doublet. In Figure 7b the modulus of the wavelet transform of this pulse
indicates t h a t it consists of three primary frequency groupings. T h e large
main energy burst at about 8.5 ln(Hz) and 0.0004 sec associated with the
trough; t h e higher frequency elongated region between 8.7 and 9.2 ln(Hz)
associated with the first peak and the high frequency event at 9.8 ln(Hz)
associated with the doublet peak at 0.00075 sec. From this examination of
the direct water-borne pulse it can be seen t h a t the wavelet transform can
resolve very high frequencies in a signal, and t h a t interference p a t t e r n s due
to scattering are resolved well in time and space.
Recordings of this source in deeper water shows the possible cause of
the interference in this pulse. In deeper water the bubble pulse separates
completely from the initial pulse. In the water depths for these examples
the bubble pulse is on top of the initial pulse making the pulse appear as
a single burst. Detailed evaluation of the direct pulse allows for a b e t t e r
assessment of the subsequent reflected pulses and this assessment is possible
with the wavelet transform.
T h e first echo is a reflection from the seabed, an event at 0.0021 sec.
T h e source pulse has now undergone modification by the seabed as result
of the reflection and subsequent partitioning of the energy at the water-
seabed boundary. T h e two lower-frequency features are still present, b u t
t h e region t h a t was at the higher frequency appears brighter relative t o
the other two lower-frequency events (Figure 7b. Possible explanations
for these differences are: the lower frequency energy transmits more of
its energy into the seabed whereas more of the higher-frequency energy is
reflected, or more of the higher-frequency energy is scattered generating a
high degree of interference b o t h constructively and destructively.
At 9.1 ln(Hz) following the seabed event there is a line of bright, almost
merged regions extending for the length of the display. It is not very likely
t h a t 8100 Hz energy penetrated the seabed to any significant depth so this
energy probably represents energy backscattered from the sub-seabed. Fre-
quencies below 8100 Hz, for example the event at 0.005 sec, have values of
about 3600 Hz, which are of similar intensity in frequency content to the
direct pulse. There is not much energy evident between the seabed and
0.005 ms events for spectral values less t h a n about 3000 Hz possibly indi-
cating no major boundaries or acoustic impedances in this interval. This
interpretation is corroborated by the stratigraphie information presented
in Figure 6. Sediments at the test site included pebbles between 3 and
6 cm diameter which is approximately one quarter of the wavelength for
the dominant scattered energy above the 8100 Hz (9 ln(Hz)) line. Larger
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01
Time (sec)
F i g u r e 7. (a) Unfiltered, n o r m a l incident trace, from s t a t i o n 5, pro-

file B e a m 3 corrected for geometrical spreading losses, (b) Modulus of
t h e wavelet transform of t h e t r a c e in (a), (c) Spectral r a t i o based on
m o d u l u s of wavelet transform of ungained trace.
202 C. Pike
scatterers were assumed to be present as the cone penetrometer test en-

countered large rocks on several occasions requiring the boreholes t o b e
drilled out in order to proceed past these points. During the process of
drilling through the sediments to obtain a bedrock core a large boulder
was encountered resulting in the recovery of a core through a portion of
this boulder.
T h e spectral ratio, calculated in the same way as the synthetic trace
example (Figure 2c), is displayed in Figure 7c. Small spectral ratios are
shown as dark or black regions and values for large spectral differences with
respect to the direct or source pulse are displayed in white.
T h e average velocity function determined from the multichannel pro-
cessing of the common depth point acoustic d a t a in phase I [23] was used to
convert t h e scales for the axes of the wavelet transform modulus plots from
time into depth beneath the seabed and a constant velocity of 1800 m / s
was used to convert frequency to wavelength. The modulus was replotted
with t h e new axes and is shown in Figure 8 as a function of depth below
the seabed against wavelength values. The scattered energy observed above
8100 Hz in Figure 7 now corresponds to wavelength values in the 0.10 to
0.30 m range. T h e strong event at a depth of 3 m (Figure 8) corresponds
to reflected energy with a dominant wavelength of 0.4 m. Another event
at about 4.5 m has a wavelength of 0.7 m. Figure 9 shows the modulus
of the wavelet transform for four normal incidence examples from Beam 3
plotted as a function of depth below seabed versus wavelength. T h e strong
energy event observed at 3 m depth in Figure 8 is evident on all of the
panels for this profile indicating a continuous boundary reflector (note the
5.0 ms event in Figure 5 and 7).
T h e preceding discussion highlights some of the potential features of
the wavelet transform display t h a t offers a new diagnostic tool t o the high-
resolution geophysicist. The analysis should lend itself well to two- dimen-
sional image analysis procedures for feature extraction and correlation. By
utilizing an energy threshold condition coupled with a density measure
for the number of events per meter a classification system for subsurface
scatterers could be developed.
5 . 3 . E s t i m a t i o n of a t t e n u a t i o n
For the attenuation analysis a similar procedure was followed as de-
scribed in section 3.2. Figure 10 shows a comparison between the variance
of t h e spectral ratio (black dots) and the value for the variance of the cor-
rected d a t a using the estimated spherical spreading correction (solid line).
T h e points cluster fairly closely about the solid line indicating t h a t either
the spectral ratio or the corrected d a t a could be taken for analysis. Atten-
uation estimates were derived from the modulus of the wavelet transform
of the d a t a corrected for geometrical spreading losses. This approach was
F i g u r e 8. Modulus of the wavelet transform for the trace in Figure 7a

plotted against depth and wavelength (units are meters).
the most successful in the synthetic example.

Curves, similar to the one in Figure 3a, were produced for all four
stations (Figure 11) from profile Beam 3, b u t time axes were converted to
depth axes using the average velocity function. Peaks near the seabed and
at or near t h e 3.0 m depth were selected for each station. This procedure
was repeated for profile Beam 1 and the values for the attenuation in t h e
first 3 m are given in Table I.
T h e m e a n value is 4.6 n e p e r s / m (40 d B / m ) with a s t a n d a r d deviation
of 1.9 n e p e r s / m (17 d B / m ) . D a t a for coarse gravelly soils is sparse b u t
values for shallow (3 m) depth are similar [37]. Lee and Malloy [19] report
values for compressional wave attenuation t h a t range from a low of 5.3
d B / m for metalliferous sediments to a high of 80 d B / m for a silty clay.
204 a Pike
Figure 9. Modulus of the wavelet transform, for normal incidence

traces along profile Beam 3 at stations 5 and 6 (a and b respectively)
plotted as in Figure 8.
Figure 9. (contd.) Modulus of the wavelet transform, for normal

incidence traces along profile Beam 3 at stations 7 and 8 (c and d re-
spectively) plotted as in Figure 8.
206 C. Pike
Time (sec) x1(T
F i g u r e 10. Attenuation curves for station 5, Beam 3 (see text for

conversion between nepers and decibels).
Table I
Wavelet Transform Attenuation Values
Beam 1 (nepers/m) Beam 3 (nepers/m) |
1 Station 5 4.4 ÉÛ f
| Station 6 3.5 2.9
| Station 7 3.5 2.8
| Station 8 5.6 6^ [
§6. C o n c l u s i o n
Two analyses based upon the wavelet transform have been presented.
The qualitative analysis discusses the features observed in the modulus
of the wavelet transform in relation to the acoustic signal and the sub-
seabed. It represents a starting point from which feature extraction and a
quantitative analysis can begin with the goal of developing classifiers for
the scatterers. It should be noted t h a t only the Morlet analyzing wavelet
has been utilized in this study. Examination of alternate wavelets t h a t may
be more diagnostic for classification of scattering should be pursued. T h e
next stage of the study of wavelets for this application will employ other
0 0|
-1 -1
-2 -2l· ■ · : < - ■ · : y-A -2 -2
-3 -3l· ■·!· ■■•J·: :■· j -3 "31
1 - 4 ■■!■■<■■: H
Φ Q
Φ . p Φ Φ
Q :^» : : o
Q
-δ[· -·:·>Ν·- -| ;· j -5
-5 -5
-6l· Λ i- | -6
-6 -6
-7
-7 -7
7 8 9 7 8 9 7 8 9
(nepers) (nepers) (nepers) (nepers)
F i g u r e 1 1 . Attenuation curves for profile Beam 3, stations 5 to 8 (a-d

respectively), plotted against depth.
analyzing wavelets in the evaluation of a finite difference model of wavefield

scattering. Investigation into interference, tuning and detuning of acoustic
energy should provide a model for assessing complex scattering phenomena
which may answer questions concerning the size and n a t u r e of scatterers
and their effect upon acoustic wavelengths of the same order.
T h e attenuation analyses suffers from the short d a t a record length at
the TerrenceviUe test site, and limits the full potential of the analysis.
Digital sub-bottom profiler records with longer recording times of 20 ms
and greater were recently acquired using the same sparker source as used
at t h e field test site in TerrenceviUe. These d a t a were collected as p a r t of
a U.S. navy seabed imaging program off P a n a m a City, Florida. Extensive
tests were carried out at this site and all d a t a will be m a d e available to
the participants. T h e wavelet transform attenuation process will be refined
and applied in an analysis of these data.
Based upon the review of the current literature on acoustic-geotechnical
correlations for marine soils, there are several applications t h a t should ben-
efit from the wavelet transform analysis. Using C-CORE's 200-gravity
centrifuge facility, very high frequency acoustic d a t a will be collected in a
208 C. Pike
study of sediment loading and transport. The attenuation of high frequency

waves ( > 100 kHz) propagated through centrifuged sediment will b e ana-
lyzed using standard techniques t h a t rely on windowed Fourier transforms
or bandpassed signals which can then be compared to results obtained using
wavelet analysis.
Some authors have questioned the validity of estimating attenuation
using the m e t h o d of spectral ratios. Sams and Goldberg [28] point out t h a t
the effect of windowing compressional waves to isolate t h e m from slower
waves is included in the calculation of the spectral ratio. They conclude
t h a t for borehole studies the spectral ratio technique will not give good
results. This problem presents a good test for the application of the wavelet
transform and the results should prove superior since the window functions
can be avoided.
References
1. Buchan, S., D. M. McCann, and D. Taylor Smith, Relations between
the acoustic and geotechnical properties of marine sediments, Q. Jl.
Eng. GeoL, 5, 265-284, 1972.
2. Courtney, R. C. and L. A. Mayer, Calculation of acoustic parameters
by a filter- correlation method, J. Acoust. Soc. Am., 93, 1145-1154,
1993.
3. Dodds, D. J., Attenuation estimates from high resolution s u b b o t t o m
profiler echoes, Presented at the Saclant Asw Research Conference on
Ocean Acoustics Influenced by the Sea Floor, La Spezia, Italy, J u n e 9 -
12, NATO conference on b o t t o m interacting ocean acoustics, 173-191,
1980.
4. English, J. M., S. T. Inkpen and J. Y. Guigné, A new, high-frequency,
broadband seismic source. Presented at the 23rd Annual Offshore Tech-
nology Conference in Houston, Texas, O T C 6558, 1991.
5. Goupillaud, Pierre L., Three New Mathematical Developments on the
Geophysical Horizon, Geophysics: The Leading Edge of Exploration,
40-42, J u n e 1992.
6. Goupillaud, P., A. Grossmann and J. Morlet, Cycle-Octaves and Re-
lated Transforms in Seismic Signal Analysis, Geoexploration, 23, 8 5 -
102, 1984.
7. Goupillaud, P., A. Grossmann, J. Morlet, Cycle-Octave Representation
for Instantaneous Frequency Spectra, SEG Expanded Abstracts with
Biographies, 53rd Annual International SEG Meeting, 613-615, 1983.
8. Grant, F. S. and G. F. West, Interpretation Theory in Applied Geo-
physics, McGraw-Hill Book Company, New York, 1965.
9. Hamilton, E. L., Geoacoustic modelling of the sea floor, Journal of the
Acoustical Society of America, 68, 1313-1340, 1980.
10. Hamilton, E. L., Compressional-wave attenuation in marine sediments,

Geophysics, 37(4), 620-646, 1972.
11. Hamilton, E. L., Prediction of in-situ acoustic and elastic properties of
marine sediments, Geophysics, 36, 266-284, 1971.
12. Hauge, P. S., Measurements of attenuation from vertical seismic pro-
files, Geophysics, 46(11), 1548-1558, 1981.
13. Haynes, R., A. M. Davis, J. M. Reynolds and D. I. Taylor-Smith, T h e
extraction of geotechnical information from high-resolution seismic re-
flection data, In Offshore Site Investigation and Foundation Behaviour,
28, 215-228. Society of Underwater Technology, 1993.
14. Jacobson, R. S., G. G. Shor Jr., and L. M. Dorman, Linear inversion
of b o d y wave d a t a - P a r t II: Attenuation versus depth using spectral
ratios, Geophysics, 46(2), 152-162, 1981.
15. Jacobson, R. S., An investigation into the fundamental relationships
between attenuation, phase dispersion, and frequency using seismic
refraction profiles over structures, Geophysics, 52(1), 72-87, 1987.
16. Johnston, D. H. and M. N. Toksöz, Chapter 1, Definitions and ter-
minology, In Seismic Wave Attenuation, Eds. M. N. Toksöz and D.
H. Johnston, SEG Geophysical Reprint Series No. 2, Series Ed. F.K.
Levin, Society of Exploration Geophysics, 1981.
17. Kronland-Martinet, R., J. Morlet, A. Grossmann, Analysis of sound
p a t t e r n s through wavelet transforms, International Journal of Pattern
Recognition and Artificial Intelligence, 1, 273-301, 1987.
18. LeBlanc, L. R., L. Mayer, M. Rufino, S. G. Schock and J. King, Marine
sediment classification using the chirp sonar, J. Acoust. Soc. Am., 9 1 ,
107-115, 1992.
19. Lee, H o m a and R. J. Malloy, Acoustic retrieval of seafloor Geotech-
nics: Technical Note, Civil Engineering Laboratory, Naval Construc-
tion Battalion Centre, Port Hueneme, California 93042, P r o g r a m Nos.
Yf52.556.091.01.109, Sponsor Naval Facilities Engineering C o m m a n d
(1977).
20. McCann, C. and D. M. McCann, T h e acoustic properties of marine
sediments, Defence Oceanology International Brighton, 1990.
21. Morlet, J., G. Arens, E. Fourgeau and D. Giard, Wave propagation and
sampling theory - P a r t I: complex signal and scattering in multilayered
media, Geophysics, 47(2), 203-221, 1982.
22. Morlet, J., G. Arens, E. Fourgeau and D. Giard, Wave propagation
and sampling theory - P a r t II: sampling theory and complex waves,
Geophysics, 47(2), 222-236, 1982.
23. Pike, C. J., S. T. Inkpen, and J. Y. Guigné, Marine geophysical and
geotechnical site characterization; Terrenceville, Newfoundland, Con-
t r a c t report for Energy Mines and Resources, Bedford Institute of
Oceanography, March, C-CORE Contract No. 91-C5, 1991.
210 C. Pike
24. Portsmouth, I. R., M. H. Worthington and C. C. Kerner, A field study

of seismic attenuation in layered sedimentary rocks - 1 VSP data, Geo-
phys. J. Int., 113, 124-134, 1992a.
25. Portsmouth, I. R., M. H. Worthington and J. P. Neep, A field study of
attenuation in layered sedimentary rocks - II Crosshole data, Geophys.
J. Int., 113, 135-143, 1992b.
26. Ricker, Norman, The form and laws of seismic wavelets, Geophysics,
18(1), 10-40, 1953.
27. Rioul, Olivier and Marten Vetterli, Wavelets and signal processing,
IEEE Signal Processing Magazine, 14-38, 1992.
28. Sams, M. and D. Goldberg, The validity of Q estimates from borehole
data using spectral ratios, Geophysics, 55(1), 97-101, 1990.
29. Schock, S. G., L. R. LeBlanc and L. A. Mayer, Chirp subbottom profiler
for quantitative sediment analysis, Geophysics, 54, 445-450, 1989.
30. Schoenberger, M. and F. K. Levin., Apparent attenuation due to in-
trabed multiples, II, Geophysics, 43(4), 730-737, 1978.
31. Stoll, R. D., Sediment acoustics, Lecture notes in earth sciences, Vol-
ume 26, Spring-Verlag, Berlin, 155pp, 1989.
32. Taylor Smith, D., Geotechnical characteristics of the sea bed related
to seismo acoustics, In Ocean Seismo-Acoustics, Eds. T. Akal and J.
M. Berkson, Plenum Press, New York, 483-500, 1986.
33. Telford, N. M., L. P. Geldart, R. E. Sheriff and D. A. Keys, Applied
Geophysics 2nd Ed., Cambridge University Press, New York, 1990.
34. Toksöz, M. N., D. H. Johnston and A. Tinur, Attenuation of seismic
waves in dry and saturated rocks: 1 Laboratory Measurements, Geo-
physics, 44, 681-690, 1979.
35. Toksöz, M. N., and D. H. Johnston Chapter 2, Laboratory measure-
ments of attenuation, In Seismic Wave Attenuation, Eds. M. N. Toksöz
and D. H. Johnston, SEG Geophysical Reprint Series No. 2, Series Ed.
F. K. Levin, Society of Exploration Geophysics, 1981.
36. Tyce, R. C , Estimating attenuation from a quantitative seismic pro-
filer, Geophysics, 46, 1364-1378, 1981.
37. Tullos, F. N. and A. C. Reid, Seismic attenuation of gulf coast sedi-
ments, Geophysics, 34, 516-528, 1969.
38. Yilmaz, O., Seismic Data Processing, Doherty, S. M., Ed., Series: In-
vestigations in Geophysics 2, Neitzel, E. B. Series Ed., Society of Ex-
ploration Geophysicists, Tulsa, OK, 1987.
This research has been partially supported by the Natural Sciences

and Engineering Research Council of Canada (NSERC), ESSO Resources
Canada Ltd., Gulf Canada Resources Ltd., Mobil Oil Canada Ltd. and
Petro-Canada Resources Ltd.
Chris J. Pike
Centre for Cold Ocean Resources Engineering (C-CORE)
Memorial University of Newfoundland
St. John's, Newfoundland
Canada, A1B 3X5
Including Multi-Scale Information in the Characterization of
Hydraulic Conductivity Distributions
Kevin E. Brewer and Stephen W. Wheatcraft
Abstract. Transport in heterogeneous porous media is highly dependent on the

spatial variability of aquifer properties, particularly hydraulic conductivity. It is
impractical, however, when modeling a real situation, to obtain aquifer properties
at the scale of each grid block. Wavelet transforms are investigated as a tool to
assess the movement of information between scales, and to incorporate the scale
and location of hydraulic conductivity test data when interpolating to a fine grid.
Wavelet transforms are particularly useful as they include parameters that define
a spatial localization center that can be positioned in space, and changed in size.
As a result, one can incorporate multiple measured hydraulic conductivity values
in the wavelet transformation at a resolution matched to each measurement scale.
A multi-scale reconstruction method is developed here which uses forward
and inverse wavelet transforms in conjunction with a pseudo-fractal distribution
to fill in missing information around sparse data. This wavelet reconstruction
method is compared to several more traditional interpolation schemes with respect
to accuracy of solute transport prediction. The wavelet reconstruction algorithm
is then used to examine the issue of optimum sample size and density for stationary
and fractal random fields.
§1. Introduction
Various authors ([14] and [19]) have shown that the spatial variability
of hydraulic conductivity, K, produces most of the transport dispersion ob-
served at the field scale. If we could accurately reproduce the K-distribution
at a fine scale in a numerical model, there would be no need for stochas-
tic theories. However, it is impractical to collect fine scale K values for
each grid block in a high resolution model. Moreover, even if we had the
information, the model would likely contain enough grid blocks (and there-
fore enough unknowns) that it would take excessive computation time to

ISBN 0-12-262850-0
214 K. Brewer and S. Wheatcraß
be practical for routine use. Both of these factors (lack of information

and large number of unknowns) lead to development of stochastic theories,
upscaling methods, and concepts of effective parameters at coarser scales.
Our work is based on the premise t h a t it would be useful to have a
rigorous m e t h o d of creating a K-distribution at a particular scale, using K
information from larger and (perhaps) smaller scales. As a first step, we
have restricted ourselves to the problem of reproducing (or reconstructing)
a K-distribution at "the finest scale." By this, we mean t h a t given a limited
number of K test values from the smallest scale, and from several larger
scales, how do we reconstruct a complete K-distribution (or equivalently, a
complete K-grid) at the finest scale.
By test values, we mean the value of K obtained from an aquifer test,
permeameter test, etc. It is important to realize t h a t each K value has
an associated scale determined by the volume of the porous media used
to obtain the K value. For porous media properties, different K sampling
techniques inherently average different sample volumes, which can result
in different sample scales and values at the same physical location. We
have focused our research on how to best use this scale information from K
samples to enhance the characterization of porous media. Our results, with
appropriate modification, could also be applied to any other scale affected
property.
Summarizing, we know t h a t field (coarse) scale contaminant disper-
sion is mostly caused by the fine scale variability of the K distribution.
Ultimately, however, we cannot adequately sample at the fine scale to ac-
count for this variability, nor can we justify the excessive computation time
for solving practical fine grid problems. Therefore, we have two possible
options: (1) develop a method to adequately reconstruct the fine scale vari-
ability from coarser, multi-scale samples, a n d / o r (2), develop a m e t h o d to
properly coarsen t h e fine scale variability to more accessible scales. In b o t h
cases, the central issue is how K variability moves, or transforms, in b o t h
directions between scales.
In this paper we will develop a method t h a t uses the wavelet transform
to include multi-scale sample information in the characterization of K-
distributions. Figure 1 poses such a multi-scale reconstruction question
for a hypothetical situation of only two scales of measurement. Although
this figure shows the movement of coarse information to a fine scale, t h e
problem of coarsening fine scale information is similar. These problems are
somewhat analogous to, in electrical engineering terminology, a d a t a fusion
problem with incomplete information.
For our contaminant transport problems, the measure of reconstruction
accuracy of t h e K-grid at the finest scale is best determined by comparison
of t h e results of flow and transport models, since it is the variability of
the finest scale K values t h a t primarily controls transport. Although t h e
Characterization of Hydraulic Conductivity Distributions 215
F i g u r e 1. Schematic showing the problem of characterizing heteroge-

neous porous media on a fine grid utilizing information from different
scales.
accuracy of the individual reconstructed fine scale K-grid values at each

grid block is of some importance, it is of secondary interest. Better re-
construction techniques will result in higher accuracy of the reconstructed
breakthrough curves compared to the true (or known) breakthrough curves.
Reconstruction techniques should also honor the multi-scale sample values
exactly. Therefore, our primary concern in determining the success of a
reconstruction is not how individual finest scale element or node interpo-
lated values m a t c h known values, b u t how accurately simulated t r a n s p o r t
through each reconstructed media compares to the true transport results.
Our (or any) reconstruction m e t h o d can only be said to be success-
ful if it reconstructs actual field distributions. However, actual field d a t a is
rarely (if ever) available in enough detail to properly test our reconstruction
m e t h o d . Even in cases where high resolution field studies have been col-
lected [5], there are no field scale tracer tests associated with such d a t a sets.
Therefore, we have used various synthetic fine grid d a t a sets as "truth,"
with comparisons m a d e primarily with simulated transport results.
We will first show how wavelet transforms can be used for multi-scale
reconstruction by exploiting the inherent localization and frequency resolu-
tion of t h e wavelet transform. We have developed a wavelet reconstruction
m e t h o d in which sample location, scale, and values are all used t o recon-
216 K. Brewer and S. Wheatcraft
struct a fine scale grid. For this initial work, we have assumed t h a t there
is no noise in our sample data; t h a t is, our sample values perfectly rep-
resent the unknown information for the given sample location and scale.
Inclusion of measurement uncertainty is delegated to future research, and
might possibly follow along the lines of [2]. The wavelet reconstruction
technique is compared with other traditional interpolation techniques by
using a 256 x 256 node synthetically generated two-dimensional porous
media and solving the flow and transport equations.
Additionally, we have used a modification of the wavelet reconstruc-
tion technique, dubbed wavelet extrapolation, to investigate what, if any,
important scales of information exist in various synthetic representations
of porous media K distributions. This extrapolation m e t h o d creates es-
timates at each location for any scale finer t h a n the given, completely
sampled, scale. Here, we a t t e m p t to understand the relative importance of
each scale to identify the inherently significant coarse scale representations
of the fine scale variability.
§2. W a v e l e t B a s e d I n t e r p o l a t i o n
2.1. Discrete transform introduction
To understand our wavelet based reconstruction and extrapolation
techniques, we must first familiarize ourselves with some mathematical
details of wavelet transforms. The wavelet transformation is similar to
the Fourier transform in t h a t a d a t a set is transformed into the frequency
(or scale) domain. T h e Fourier transform (via spectral analysis) has been
used extensively to study the characterization of porous media [7], b u t a
limitation of the Fourier analysis technique is the assumption t h a t the in-
formation at each scale is homogeneous and can be essentially represented
by a single "power" value. W h a t makes the wavelet transform useful as
a tool for multi-scale reconstruction of porous media is t h a t this assump-
tion is not necessary, resulting in the preservation of positional information
(i.e. t h e heterogeneity) at each scale. Daubechies in [6] best expresses t h e
wavelet transform as "a tool t h a t cuts up d a t a or functions or operators
into different frequency components, and then studies each component with
a resolution matched to its scale." Thus, with the wavelet transform, one
can analyze d a t a sets with heterogeneous information at each scale, which
is commonly the case for porous media distributions.
T h e forward discrete wavelet transform consists of a convolution of a
basis function (sometimes confusingly called the wavelet function) with a
d a t a set. Unlike the Fourier transform, which uses only sines and cosines
as basis functions, the wavelet transform can use a variety of basis func-
tions. Basis functions are discussed in detail in [6], among others. Here we
discuss the important basis function characteristics t h a t will concern our
1 (a)
π
U
F i g u r e 2. Graphs of the (a) (left) Haar basis function, (b) (middle)
Daubechies4, and (c) (right) Daubechiesl2 basis functions.
reconstruction and extrapolation techniques.

Every basis function has localization (or positional) and frequency com-
ponents which control the scale-to-scale relationships, and the in-scale posi-
tional relationships. By scale-to-scale relationships we mean t h e amount of
coarsening (or conversely, the fining) from scale-to-scale. T h e basis function
components are commonly set is such a way t h a t the wavelet coefficients
on each scale are related to the next coarser or finer scale by a factor of 2,
and t h a t the transform convolution begins with every odd-indexed element
on each scale.
T h a t is, if for a one-dimensional d a t a set whose finest scale is 128
elements of 1 length unit, the next coarser scale will consist of 64 elements
representing 2 original-scale length units. Similarly, the next coarser scale
will consist of 32 elements representing 4 length units, and so on, until t h e
coarsest possible scale, which will consist of only one element representing
the entire 128 original-scale length. For our work, we have used t h e common
Haar and Daubechies family basis functions [6] which have a scale-to-scale
factor of 2. Each of these basis functions are also orthogonal, and so can
be conveniently used on two-dimensional d a t a sets by first performing each
transform on t h e row data, then the column data.
Figure 2 shows graphs of the basis functions themselves. T h e Haar
basis function is highly discontinuous, and as such, is often used to analyze
discontinuous d a t a sets. T h e Daubechies family basis functions are less
discontinuous, with the higher order functions becoming more smooth. T h e
basis function coefficients we used in the discrete transform equations are
listed in Table I.
T h e discrete wavelet transform is a transformation of information from
a fine scale t o a coarser scale by extracting information t h a t describes the
218 K. Brewer and S. Wheatcrafi
Table I
Basis function coefficients for the Haar and Daubechies4 functions.
Haar
ci 0.70710678
c2 0.70710678
Daubechies4
d 0.48296291
c2 0.83651630
c3 0.22414387
c4 -0.12940952
fine scale variability (the detail coefficients) and the coarser scale smooth-
ness (the smooth coefficients) according to:
{sm} = [H]{sm+1}; {dm} = [G]{sm+l} (1)
where s represent smooth coefficients, d represent detail coefficients, m is
the level, and H and G are the convolution matrices based on the wavelet
basis function. Higher values of m signify finer scales of information. The
complete wavelet transform is a process that recursively applies Equa-
tion (1) from the finest to the coarsest level (scale). This describes a scale
by scale extraction of the variability information at each scale. The smooth
coefficients generated at each scale are used for the extraction in the next
coarser scale.
The matrices H and G are created from the coefficients of the basis
functions, and represent the convolution of the basis function with the data.
The following is the general makeup for the H and G matrices for a four
coefficient basis function that has a scale-to-scale relationship of 2 (basis
functions that have more or less coefficients would be similar):
C\ C2 C3 C4 0 0 0 0 0 0
0 0 Ci C2 C3 C4 0 0 0 0
H = 0 0 0 0 ci c2 c3 c4 0 0 (2)
c3 c4 0 0 0 0 0 0 c\ c2
c4 -c3 c2 -ci 0 0 0 0
0 0 c4 —c3 c2 — c\ 0 0
G = 0 0 0 0 c4 —c3 c2 —ci (3)
C2 -d 0 0 0 0 0 0
c4 - c 3
As a forward wavelet transform example, Figure 3a shows a simple 4
element one-dimensional data set {6,4,5,3} and the first convolution appli-
cation using t h e simple Haar basis function resulting in two m = 2 level

detail coefficients and two m = 2 level smooth coefficients. Continuing in
Figure 3b, t h e transform convolution is applied to the two m = 2 level
smooth coefficients resulting in a single ra = 1 level smooth coefficient and
detail coefficient. Note t h a t the m = 2 level detail coefficients have not
been altered.
Alternatively, we can represent Figure 3a following the form of Equa-
tion (1) as:
r
6
Γ i 1 0 0 4
V2 x/2 <
0 0 1 1 5 (4)
V2 V2 .
3
for the smooth coefficient upscaling, and:
[ 6
2^- 1 1
V2 "V2 4
2 — 0 0 1 1
5 (5)
3
for the detail coefficient calculation. Continuing, we can represent the
final calculation in Figure 3b in matrix notation for the smooth coefficient
upscaling, and the detail coefficient calculation as, respectively:
H* ^} (6)
1-{V5 *}{"£ (7)
Traditionally, only the detail coefficients at every scale, and the coarsest
level smooth coefficients, are considered a complete set of wavelet coeffi-
cients. In our example of Figure 3, the complete set is {9, 1, 1.414, 1.414
}. In our work however, we are not so much interested with just t h e tra-
ditional wavelet coefficients, but in the entire suite of smooth and detail
coefficients at each scale (or level) since we will show in t h e next section
t h a t our multi-scale samples can be related to smooth coefficients at various
scales and locations.
T h e inverse discrete wavelet transform is similarly implemented via
a recursive recombination of the smooth and detail information from t h e
coarsest to finest level (scale):
{sm+1} = [H]'{sm} + [G]'{dm} (8)

where H' and G' indicate the transpose of the H and G matrices.
7.07 1.41 5.66 1.41

s, d, S2 d2
λ
■1/V5
m = 3{
.-{
.-3{
F i g u r e 3. Discrete wavelet transform using the Haar basis function on

an example data set {6,4,5,3}: (a) the transform of the original data, (b)
the completed transform resulting in a single m = 1 smooth coefficient,
and a single m = 1 detail coefficient, and two m = 2 detail coefficients.
Using the generated wavelet coefficients in our previous example of

Figure 3, the inverse transform using the Haar basis function initially pro-
ceeds to generate the m = 2 smooth coefficients from the m = 1 smooth
and detail coefficients, as shown in Figure 4a. Continuing, the finest scale
(third-level) coefficients are generated using the m = 2 level information,
as shown in Figure 4b.
In matrix notation, the Figure 4a transform is:
m-itH-%} (9)
9.00 1.00
s, d,
/ 1/V2 \ 1
n-2{
7.07 5.66
si S2
5.66 1.41
„ = 2{ s2 d2
1/V2 / / 1/V2 \ i
„.,{
Figure 4. Discrete inverse wavelet transform using the Haar basis

function on the wavelet coefficients from Figure 3: (a) the inverse trans-
form from the first-level, (b) the completed inverse transform resulting
in the original fine scale data values.
and t h e Figure 4b transform is:
41r
r 61
4 ±
4r 0
0
0
0
5 0 1 0 1 2 — • (io)
3 0 4r J 0 —fe
V2 s/2
Note t h a t t h e resulting finest level smooth information is exactly t h e

same as t h e original d a t a set. Overall, b o t h the forward and inverse wavelet
transformations, when applied over many scales (levels), can be pictorially
represented as in Figure 5.
Level
(m)
Figure 5. Exploded view of an one-dimensional wavelet transform.

Reprinted with permission from the SI AM Review, 31(4), pp. 614-627.
Copyright 1989 by the Society for Industrial and Applied Mathematics,
Philadelphia, Pennsylvania. All rights reserved.
2.2. Multi-scale reconstruction

T h e objective of multi-scale reconstruction is to completely determine
all finest scale values given a limited number of sample values from t h e
finest and several coarser scales, and possibly other necessary information.
This section introduces a multi-scale reconstruction algorithm based on the
wavelet transform.
As previously shown (see Figure 3 and Figure 4), the wavelet transform
can be represented by a linked pyramid (or lattice) of smooth and detail
coefficients. Because hydraulic conductivity (K) samples contain scale in-
formation and are an inherent average over some area or volume, they are
multi-scale samples and should be amenable to this reconstruction algo-
rithm. Therefore, our K samples at different scales and locations, for input
to our reconstruction algorithm, are equivalent to the smooth coefficients
on the lattice, b o t h representing an average of finer scale information.
To generate the desired finest scale values from coarser scale smooth
coefficients, the lattice relationships necessitate t h a t all other smooth and
detail coefficients, if unknown, also be generated. T h a t is, to generate
all the unknown finest scale values via application of the discrete inverse
wavelet transform, all the next coarsest scale smooth and detail coefficients
must be known or calculated. Since the discrete forward and inverse wavelet
transforms create relationships between adjacent levels (scales) of smooth

and detail coefficients, these coarser smooth coefficients can b e calculated
from even coarser scale (via interpolation using the inverse transform) and
from t h e finer scale (via upscaling using t h e forward transform) coefficients.
T h e same will be true for all unknown coefficients at any coarser scale,
except t h e coarsest possible scale, where only finer scale information can
be used. T h e problem is then of solving for all the unknown smooth and
detail coefficients on all levels with a limited number of known coefficients.
There are different potential solution techniques to solve this wavelet
transform based multi-scale reconstruction problem. Since our known sam-
ple d a t a is sparse (i.e., a large number of unknown smooth coefficients on
a number of different scales) and we are limiting our application to two-
dimensions, we have elected to use a simple cyclic iteration and averaging
technique which is in some ways similar to multigrid cycling. Other possible
techniques have been proposed [3], and may be utilized in future work.
Cyclic iteration essentially consists of starting at the coarsest scale,
proceeds to the finest scale, and then cycles to the coarsest and back down
to the finest until convergence is reached for each finest scale value. At
each scale, each unknown coefficient is calculated. For all, but t h e finest
and coarsest scales of interest, this means t h a t the value interpolated from
coarser information and the value upscaled from finer information is weight
averaged to t h e final estimate. In our work, we used equal weights for the
interpolated and upscaled values. T h e cyclic iteration (coarsest-to-finest-
to-coarsest, etc.) is necessary to "move" the limited information around
the lattice, as appropriate. Algorithm convergence is determined when
the mean, maximum, and variance of change in fine scale values between
iterations is below predetermined criteria, typically equal to two or three
orders of magnitude less t h a n the precision of the sample values.
T h e scales of interest, i.e. the scales inclusive of t h e finest and coarsest
in our cyclic iterations, are bounded on the fine-end and the coarsest-end
by t h e scales in the sample set. Our implementation requires complete
test value coverage at the coarsest scale to ensure each fine scale generated
value is supported by at least one known test value, and to bring closure
to the algorithm. It is important to note t h a t each smooth coefficient (i.e.,
sample value) contains all detail and coarse information from all higher
(coarser) scales.
In general, the requirement for complete test value coverage at the
coarsest scale is not necessary. For site characterization problems t h a t
have a statistically significant number of test values on any one particular
scale, a statistical relationship could be derived to assign stochastic sample
values to t h e remaining unknown sample locations at t h a t level. T h a t level
would become fully populated with real and stochastic values, and any
coarser scale test values would not be necessary.
Essentially all detail coefficient values at any scale cannot typically

be calculated, since enough spatial coverage of known smooth coefficients
to directly calculate detail coefficients will rarely occur. Therefore, based
on a pseudo-fractal relationship, we assign stochastic values for all detail
coefficients assuming an underlying infinite correlation at the finest scale.
The choice of a pseudo-fractal relationship for detail coefficients is arbitrary,
but reasonable, based on recent work such as [17] and [20]. It is possible to
use log-normally-distributed stationary random fields or a variety of other
statistical distributions. There is no inherent algorithmic limitation for use
of any other desired detail coefficient relationship.
Wornell and Oppenheim in [21] showed the fractal relationship between
expected values of the detail coefficients as:
E[(dmf] = σ22-^ (11)
2
where σ is a positive constant t h a t depends on the choice of the wavelet
basis function, 7 is related to the desired fractal dimension, and m is the
scale. Adopting Equation (11) for the assignment of specific detail coeffi-
cients, we have for our two-dimensional algorithm:
dm = / ( / 2 - i 6 ^ - 1 ) (12)
b = 2(3-D) + l (13)
where / is a fitting parameter, D is the desired fractal dimension (3.0 >
D > 2.0), and g is a Gaussian random number. We found / = 100 to be the
best fitting parameter value based on visual analysis of the reconstructed
d a t a sets. T h e random component in our work was generated via t h e
R A N I and GASDEV random number generators given in [13]. T h e choice
of fractal dimension is discussed below.
T h e reconstruction algorithm begins with the assignment of t h e sample
values to smooth coefficients at the appropriate location and level (based on
the location and scale of the sample). Since the wavelet transform smooth
coefficients are not normalized, the sample values are adjusted prior to their
assignment. For example, given a completely smooth d a t a set {1,1,1,1},
the m = 2 level smooth coefficients would not be 1 as expected from a
statistical mean, b u t would be 1.414 (for the Haar basis). In this case, a
m = 2 level sample value of 1 would be adjusted to 1.414 before assignment
to the smooth coefficient. This adjustment assumes t h a t the sample values
are representative of a statistical mean with the amount of adjustment
based on the basis function used. Fut her research is needed t o determine
the proper adjustment for actual hydraulic conductivity test values.
After the assignment of the sample values, the detail coefficients are
assigned during the initial downward (coarsest to finest scale) leg. After
calculating the finest scale estimates, the cyclic iterations begin, with con-
vergence calculations after each finest scale estimation. T h e algorithm ends
when convergence is reached. Pseudo-code for our wavelet reconstruction

algorithm is given in Table II.
Since t h e calculation for each coefficient is independent on each level,
the algorithm is inherently parallelizable. Using either distributed a n d / o r
parallel computing capabilities, a significant decrease in the execution time
for large problems would be experienced. Although not necessary for our
work, this capability may be highly desirable for large problems (especially
three-dimensional d a t a sets).
Wavelet reconstruction can also accommodate use of multiple detail
coefficient relationships a n d / o r wavelet basis functions. This allows incor-
poration of soft-information, such as known geologic relationships of values
which are different between scales. For example, large scale information
might be best modeled by a smoothing basis function, whereas the finer
scales could be b e t t e r modeled by a discontinuous (erratic) basis function.
This scenario could represent an interbedded sand and gravel river deposit
influenced by an overall hydraulic downstream particle sorting. In our ini-
tial work, however, we have not implemented these extensions and only use
one wavelet basis function and one stochastic detail coefficient distribution
for all scales.
2.3. Multi-scale extrapolation

We have also developed a pure stochastic extrapolation algorithm based
on the wavelet transform to study important scales of interest. This ex-
trapolation m e t h o d could be combined with the multi-scale reconstruction
algorithm for situations where the desired grid level is smaller t h a n the at-
tainable sample values, or when it is desirable to maintain the similarity of
the variable distribution upon closer inspection of a subset area of interest.
T h e algorithm requires complete sample coverage at only one scale,
and o u t p u t s all values at the desired finest scale. Detail coefficients are
stochastically generated, as discussed above, for all scales down to t h e
finest scale desired. Then, level-by-level, the smooth and detail coefficients
are used to interpolate the finer scale smooth coefficients via the inverse
wavelet transform. T h e algorithm stops when all the finest desired level
smooth values are generated. Pseudo-code for the extrapolation algorithm
is given in Table III .
§3. Traditional I n t e r p o l a t i o n M e t h o d s
T h e wavelet reconstruction technique was compared to three tradi-
tional interpolation methods: simple assignment, kriging, and conditional
simulation. These methods are briefly reviewed below. Kriging and con-
ditional simulation are classical geostatistical methods t h a t have been in
use in mining engineering and geological fields for over a decade. Journel
Table II
Pseudo-code of the wavelet multi-scale reconstruction algorithm.
C SETUP
READ D — FRACTAL DIMENSION
BETA = 2*(3-D)+l
C READ IN SAMPLE INFORMATION
FOR ALL SAMPLES
READ SAMPLE => LOCATION, VALUE, LEVEL
SMOOTH_COEF(LOCATION, LEVEL) = ADJUST_VALUE(SAMPLE VALUE)
KNOWN_FLAG(LOCATION, LEVEL) = 1
END FOR
C ASSIGN DETAIL COEFFICIENTS
FOR EVERY LEVEL
FOR EVERY LOCATION
DETAIL_COEF(LOCATION, LEVEL) = GAUSS*2**(-0.5*(LEVEL-1)*BETA)
END FOR
END FOR
C PERFORM INITIAL DOWN SWEEP
FOR LEVEL = MINLEVEL+1 TO MAXLEVEL-1
FOR EVERY LOCATION
IF KNOWN_FLAG(LOCATION, LEVEL) IS NOT 1 THEN
XWT = WAVELET_TRANSFORM(LEVEL, LOCATION, FORWARD_FLAG)
XIWT = WAVELET_TRANSFORM(LEVEL, LOCATION, INVERSE_FLAG)
SMOOTH_COEF(LOCATION, LEVEL) = AVERAGE(XWT, XIWT)
END IF
END FOR
END FOR
C CALCULATE THE FINEST SCALE (NO FORWARD TRANSFORM POSSIBLE)
LEVEL = MAXLEVEL
FOR EVERY LOCATION
SMOOTH_COEF(LOCATION, LEVEL) = XIWT
END IF
END FOR
C GO UP AND DOWN UNTIL CONVERGENCE IS ACHIEVED
CONVERGENCE_FLAG IS FALSE
DO UNTIL CONVERGENCE_FLAG IS TRUE
C GO UP
FOR LEVEL = MAXLEVEL-1 TO MINLEVEL+1 BY -1
FOR EVERY LOCATION
END IF
END FOR
END FOR
C GO DOWN
FOR LEVEL = MINLEVEL+2 TO MAXLEVEL-1
FOR EVERY LOCATION
END IF
END FOR
END FOR
C DO FINEST LEVEL
LEVEL = MAXLEVEL
FOR EVERY LOCATION
SMOOTH_COEF(LOCATION, LEVEL) = XIWT
UPDATE MEAN_CHANGE AND MAX_CHANGE
END IF
END FOR
IF MEAN_CHANGE < DESIRED_MEAN_CONVERGENCE _AND_
MAX_CHANGE < DESIRED_MAX_CONVERGENCE THEN
CONVERGENCE_FLAG IS TRUE
END IF
END DO
C WRITE THE FINEST LEVEL VALUES
FOR EVERY LOCATION
WRITE SMOOTH_COEF(LOCATION, MAXLEVEL)
END FOR
STOP END
Table III
Pseudo-code of the wavelet multi-scale extrapolation algorithm.
C SETUP
READ D — FRACTAL DIMENSION
BETA = 2*(3-D)+l
READ MAXLEVEL
READ OTHER VARIABLES
C READ IN SAMPLE INFORMATION ~ COMPLETE ON ONE SCALE
FOR ALL SAMPLES
READ SAMPLE => LOCATION, VALUE, LEVEL
SMOOTH_COEF(LOCATION, LEVEL) = SAMPLE VALUE
MINLEVEL = LEVEL
END FOR
C ASSIGN DETAIL COEFFICIENTS WHEN SWEEPING DOWN
FOR LEVEL = MINLEVEL+1 TO MAXLEVEL
FOR EVERY LOCATION
DETAIL_COEF(LOCATION, LEVEL) = GAUSS*2**(-0.5*(LEVEL-1)*BETA)
END FOR
END FOR
C WRITE THE FINEST LEVEL VALUES
FOR EVERY LOCATION
WRITE SMOOTH_COEF(LOCATION, MAXLEVEL)
END FOR
STOP END
and Huijbregts [9] and Hohn [8] detail the development and mathematical
justification behind these and other geostatistical methods.
3.1. Simple assignment

T h e first and simplest of the traditional methods is actually an assign-
ment method, rather t h a n an interpolation one, and consists of assigning a
sample value to each fine scale grid location based on the finest scale sample
encompassing t h a t fine scale grid location. For example, if an unknown fine
scale grid location was contained in the sampling area of a medium scale
sample and a coarse scale sample, the medium scale sample value would b e
assigned to this fine grid location. In other words, the medium scale sample
is the finest scale sample encompassing the unknown fine grid location. In
another example, if a different unknown fine scale grid location is only en-
compassed by a coarse scale sample, t h a t coarse scale sample value would
be assigned to the fine scale target location. Assignments are determined
in a like manner for all unknown fine grid locations. We have included t h e
simple assignment m e t h o d in our analysis primarily because it is expected
to represent the worst case in our reconstruction accuracy comparison.
228 K. Brewer and S. Wheatcrafl
3.2. Kriging
Kriging is a best linear unbiased estimator (BLUE) and uses the spatial
covariance of the samples to determine the best estimate at a particular
point. All samples are assumed to have point support and thus to be from
only one scale. As developed in [9], the kriging system consists of solving
for sample weights, λ:
[W}{\} = {B} (14)
where W is the sample-point to sample-point covariance matrix and B is
the sample-point to target-point covariance vector. The weights are then
used on the sample values to calculate an estimate at the target point. This
procedure is applied at each unknown grid point. The kriging algorithm in
[1] was used with a slight modification to incorporate the simple assignment
technique, described above, for those fine grid locations which were beyond
the algorithm range.
3.3. Conditional simulation

Conditional simulation is a method by which a correlated stochastic
random field on a fine scale, known as a simulation field, is conditioned to
limited sample information. The result is a correlated stochastic realiza-
tion that exactly matches sample values at the point these samples occur.
The unconditioned correlated random field is generated from the second
order statistics (correlation and mean) of the samples. The conditioning
is performed with a kriging-like algorithm. Like kriging, sample values are
expected to be from a single scale (the scale of the unconditioned field).
Proper generation of the unconditioned random field is very important
as an incorrect field will result in poor conditional simulation results. Turn-
ing bands methods are typically used to generate the correlated random
fields, and as pointed out by [16], using an inadequate number of bands
can result in striations or banding in the simulation field. For our work we
have incorporated the turning bands generator in [16] with a conditioning
algorithm based in part on the algorithm given in [12].
§4. Evaluation of Methods

4.1. Reconstruction comparison
The applicability and potential accuracy of the wavelet reconstruction
technique was investigated by comparing various reconstruction techniques
using a single generated sample set from a known synthetic two-dimensional
hydraulic conductivity distribution. A synthetic fine scale grid was used to
know, with certainty, the actual fine scale values and the resulting "true"
particle breakthrough curve. An equitable comparison could therefore be
made for each reconstruction technique.
F i g u r e 6. Sampling schematic.
T h e known fine scale grid was a two-dimensional set (128 x 128) of

K values ranging four orders of magnitude. T h e ln(üf) values were gener-
ated as a pseudo-fractal from a spectral fractional Brownian motion (fBm)
algorithm [18]. T h e input fractal dimension of 2.4 was representative of
geologic media scaling relationships ([10] and [17]).
T h e known grid was sampled at three different scales, as schematically
shown in Figure 6. There were four sample values (representing complete
area coverage) at the largest scale, nine sample values at a medium scale,
and 227 fine scale samples values grouped in three primary areas. All t h e
sample locations were chosen to mimic a typical site characterization sam-
pling plan using "best professional judgement," without regard to whether
it would favor one m e t h o d over another.
Obviously, "best professional judgement" is a limited m e t h o d for sam-
ple selection as it is based entirely on so called best guesses. A more
rigorous approach would have been to evaluate multiple sample sets in a
type of Monte Carlo process, in an effort to obtain the best sample suite
for each m e t h o d . These issues will be discussed in future work.
In general, the sample locating philosophy was: (1) to spread a limited
number of samples over the entire study area in an a t t e m p t to understand
the overall variability on each sampled scale, and (2), to group some of the
samples on each scale in areas of high variability. T h e nine medium scale
samples were therefore located throughout the domain, with a subset of
four samples grouped near the domain center. T h e three areas of dense
fine scale samples were located in areas of interesting medium and large
scale sample value variability. These fine scale sample groups were oriented
in line with the flow direction to maximize information t h a t directly affects
dispersion.
•
2-
T Sample •
• Known •
•
•
1.5- •
•
•
1-
• ▼ ▼
••
0.5- •
T
• ▼
▼ ▼ ▼ ▼ ' ▼
0- Ml 1 1
▼ ▼
1 1
0 —i5 1
10 15r 20 25 30 35 40
Lag
F i g u r e 7. Semivariograms for the entire, known pseudo-fractal syn-

thetic fine scale grid and the resulting sample set.
T h e medium and large scale sample values were calculated using a sim-
ple geometric mean of all known K values in each sample area [11]. This
sample set (sample value, location, and scale) was used as input for t h e
simple assignment and wavelet interpolation techniques. For the kriging
and conditional simulation interpolation techniques, the sample set was
modified by locating the medium and large scale samples at each sample
centroid, and discarding the sample scale information. T h e sample semi-
variogram, Figure 7, was used for determining the covariance structure for
the kriging and conditional simulation reconstruction techniques. A range
of 12, sill of 0.5, and a nugget of 0.0 was determined from the sample
semivariogram. The sample semivariogram is not similar to the known
semivariogram, and this difference is a likely cause for part of any resulting
error for the kriging and conditional simulation techniques.
Only one fine scale reconstruction realization from the sample set was
required for the deterministic simple assignment and kriging techniques.
The wavelet interpolation and the conditional simulation techniques, how-
ever, used a Monte Carlo ensemble average. The Daubechies4 wavelet basis
function [6] was used for the wavelet reconstruction technique.
For each of t h e fine scale reconstructions, as well as for t h e known grid,
the flow equation was solved with a multigrid approach [4] with no-flow side
r- 38000
I, A
\\ -37000
J| f\ -36000
-35000
N,\,v- - +- v ~ \^ - - - - - - -34000
~v^/*- -33000
j -32000
-31000
-30000
Figure 8. Cumulative Monte Carlo breakthrough curve statistics for

(a) the wavelet reconstruction, and (b) the conditional simulation tech-
niques.
boundaries and constant head up- and down-gradient boundaries. Parti-

cle tracking was used on each solved flow domain to simulate t r a n s p o r t .
Particles were released at the up-gradient boundary and tracked through
the flow domain. Since the sides were no-flow boundaries, all particles ex-
ited at t h e down-gradient boundary. T h e individual particle travel times
were used to create breakthrough curves. For t h e wavelet and conditional
simulation techniques, 50 Monte-Carlo realizations were found t o b e suffi-
cient to generate stable ensemble breakthrough curve statistics, as shown
in Figure 8.
T h e shaded contour maps shown in Figure 9 are t h e fine scale ln(AT)
values from t h e known grid, and the simple, kriged, conditional simula-
tion, and wavelet reconstructions. As shown, t h e fractal characteristic of
the known ln(AT) grid, the blocks of assigned values for t h e simple recon-
struction, as well as t h e expected smoothing of t h e point samples on t h e
kriged ln(K) reconstruction are readily observable. T h e conditional simu-

lation realization exhibits the correlated stochastic n a t u r e underlying t h e
technique. Although the wavelet In (if) reconstruction is not exact, it does
show the influence of the test scales (evidenced by the generally high grid
values in the southwest corner, consistent with large scale sample values),
indicating the coarser level information is indeed being transmitted to finer
scales. Resampling the wavelet reconstructions gives exact reproduction of
the original sample set values.
T h e simple and kriged particle breakthrough curves (Figure 10) exhibit
a "bench" which is not evident in the known breakthrough curve. This
bench is apparently due to the large difference in K values at the coarsest
sample scale directly influencing the reconstructions, since at t h a t scale, a
sample covers 50% of the domain perpendicular to the flow direction. T h e
wavelet breakthrough curve appears to exhibit some of the characteristics
of the known curve. The conditional simulation ensemble curve is very
smooth and generally indicates early breakthrough at most concentrations.
T h e mean error of each breakthrough curve can be calculated by:
where r; is the time for the i t h particle to travel through the reconstructed
domain, k{ is the time for the ith particle to travel through the known
domain, and for each of the n particles. These resulting mean error values
are shown in Table IV. As the percentages show, the wavelet and condi-
tional simulation techniques recreate breakthrough curves b e t t e r t h a n t h e
simple and kriging techniques. For this particular sample set, it appears
t h a t conditional simulation technique is slightly b e t t e r t h a n wavelet recon-
struction. This result may be an artifact of the sample suite or the basis
function used in the wavelet method.
4 . 2 . O p t i m u m scale
Addressing the question of whether there are optimum sample scales
for various synthetic two-dimensional hydraulic conductivity distributions,
the wavelet extrapolation technique was used with complete sample cover-
age at various scales. An optimum sample scale is defined as one where t h e
greatest reduction in breakthrough curve error occurs, which is indicative
of important scales representing the coarsening process of the fine scale
variability. Identification of optimum sample scales are also important in
developing an effective and efficient site characterization procedure. T h e
experiments consisted of using the wavelet extrapolation technique t o ex-
trapolate complete sample information at each scale down to t h e original
finest scale. T h e Daubechies4 wavelet basis function was again used for all
Figure 9. Fine scale ln(K) shaded contour maps for: (a) The known
grid.
Figure 9. (contd.) (b) The simple assignment reconstruction.

Figure 9. (contd.) (c) The kriging reconstruction.

Figure 9. (contd.) (d) One realization of the conditional simulation

reconstruction.
Figure 9. (contd.) (e) One realization of the wavelet reconstruction.

/
0.9-
0.8- /J
/ ,· r
/ / / ' /
0.7-
0.6-
S i' 1 / /
1
0.5- I ' /, /
! ' iM
0.4-
Known
0.3-
i ii 1
i 1/ ) ——— Simple
0.2- Kriged
■ · s- y
Wavelet
0.1- 1 ' lx
1 · l: ..... Cond. Sim.
n 1 / 1J: 1 1 1 1 1 1 1 1
0 100 200 300 400 500 600 700 800 900
Time
Figure 10. Particle breakthrough curves.
Table I V
Mean breakthrough curve error for reconstruction
techniques.
Reconstruction Technique Mean BTC error
Simple Assignment 50%
Kriging 48%
Wavelet Reconstruction Ensemble 31%
Conditional Simulation Ensemble 25%
extrapolations.
W i t h a 128 x 128 regular grid, 6 separate levels can be completely sam-
pled. Samples at each level were generated using the geometric mean of the
known fine scale values. Each complete level sample set is then used, sepa-
rately, to reconstruct the fine grid via the wavelet extrapolation technique.
T h e flow equations are solved and particle breakthrough curves are gener-
ated. Since the extrapolation technique is stochastic, a Monte Carlo suite
of thirty reconstruction realizations was necessary to ensure stability of t h e
first and second order moments of the ensemble breakthrough curve. These
techniques are similar to those described in Subsection 4.1. An example
of a breakthrough curve suite for all levels is given in Figure 11. This fig-
i 1 1 1 1 1 1 1 1
0 100 200 300 400 500 600 700 800
Time
F i g u r e 1 1 . Suite of breakthrough curves showing the increasing ac-

curacy of the reconstructions as finer levels of information are used.
ure visually shows the decrease in error (or equivalently, an increase in t h e

incorporation of the fine scale variability) as increasingly detailed sample
information is used.
T h e first group of numerical experiments used a synthetically generated
12 grid length (approximately 9% of the domain) correlated stochastic fine
scale 1η(/ίΓ) grid as the known values. The known field was generated via
the turning bands algorithm used previously. Figure 12 shows the resulting
mean and variance of the ensemble breakthrough curve mean error for each
level. T h e horizontal axis shows the number of samples in each level's
sample set. At the known fine grid level, 16,384 "samples," shows zero
error. This plot shows the expected continual decrease in mean error and
variance with finer sampling.
Using the mean values from the high slope part of Figure 12, and
differencing the error values between levels, we obtain Figure 13 which
shows the decrease in error from each previous level. This plot shows t h a t
the greatest decrease in error occurs at 64 samples, which is a sample size
of 16 x 16 grid units. This is very close to the correlation length of 12 and
appears t o indicate t h a t for correlated random fields, the optimum scale is
equivalent t o the correlation length.
To confirm t h a t this result was not due to the particular grid size
used, t h e experiment was re-run with a 512 x 512 correlated stochastic
50
45
p 40
LU
Φ 35-
30
5 -
O) 2 5 -
I i1
2 20- 11
CO
15-
Φ 1
ω 10-
c
co
Φ
-i 1 r-ü
10 100 1000 10000 100000
Number of Samples
Figure 12. Mean and variance of ensemble breakthrough curve errors.
10000 100000
Number of Samples
Figure 13. Mean error change for increasing finer samples levels, 128 X
128 correlated stochastic field with a correlation length of 12 units.
-■- (a) 128:12 I

9-
-#- (b) 512:12
8-
-*- (c) 512:48
CO 7
ir 7-
o
% 6-
5·-
CO
CO 4_
Ü
Ç 3-
0
• ^
^ 2-
1-
0- 1 1 1 — 1
10 100 1000 10000 100000
Number of Samples
F i g u r e 14. Mean error change for (a) 128 X 128:12, (b) 512 X 512:12,
and (c) 512 x 512:48 correlated stochastic fields.
two-dimensional field with a correlation length of 12 units, and also with

a correlation length of 48 units. The first field is effectively a reduction in
relative correlation length from 9% (12/128) of the overall domain length,
to 2% (12/512). T h e second field has the same percentage correlation
length as the first experiment (9%), except t h a t a finer overall grid is used.
T h e results for these two new experiments and the original experiments are
shown in Figure 14.
T h e results show t h a t the peak has shifted in the 512 x 512:12 case to
1024 samples. This is equivalent to a 16 x 16 unit sample size, which is again
nearly equivalent to the correlation length (12) of the field. T h e peak for
the 512 x 512:48 case is approximately equivalent to the first experiment,
and is again approximately equivalent to the correlation length of t h e field.
Although results from only one realization for each experiment is shown,
numerous other experimental replications of the stochastic fine grid field
confirm these results.
T h e second group of numerical experiments used a pseudo-fractal (i.e.,
infinite correlation) fine scale ln(K) grid for the known fine grid. T h e
pseudo-fractal field was generated with the spectral fBm algorithm de-
scribed before, with a representative fractal dimension (D=2.4). Figure 15
shows t h e resulting plot for a 128 x 128 grid, where 64 samples (16 x 16
grid units) appears optimum. Rerunning the experiment with a known fine
100 1000 10000 100000

Number of Samples
F i g u r e 1 5 . Mean error change for increasing finer samples, 128 X 128

pseudo-fractal field.
100 1000 10000 100000

Number of Samples
F i g u r e 1 6 . Mean error change for increasing finer samples, 512 X 512

pseudo-fractal field.
grid of 512 x 512 and the same fractal dimension, no change in the sample
o p t i m u m (64 samples) was evident, as shown in Figure 16. In this case,
however, the sample size is 64 x 64 grid units. T h e cause for the difference
in o p t i m u m sample size, yet consistency in sample quantity is not clear.
4 . 3 . W a v e l e t basis effects
As noted in Subsection 2.1, many different wavelet basis functions can
be used in the wavelet transform, and consequently, also in the wavelet
reconstruction and extrapolation techniques. Different wavelet basis func-
tions will preferentially move, between scales, different characteristics of
the target dataset. For example, use of the Haar basis function will em-
phasize discontinuities in the target dataset, and similarly, would enhance
reconstruction of a "discontinuous" fine grid field from a sample set. Other
wavelet basis functions, such as the Daubechies family used throughout
this research, emphasize the smoothness of the examined data.
T h e wavelet reconstruction experiments for fine scale ln(JiT) grids proved
to be highly dependent on the wavelet basis chosen. Figure 17 shows a
Haar basis wavelet reconstruction realization ln(Ä') shaded contour m a p .
T h e resulting discontinuous field is evident, especially compared with t h e
Daubechies4 reconstruction in Figure 9e. These discontinuities in the Haar
reconstruction, which are not inherent in the known grid, m a d e it unsuit-
able as a basis function for reconstructing our known fractal distributions.
An example of the effect on breakthrough curve error by utilizing dif-
ferent wavelet basis functions is shown in Figure 18. T h e Haar wavelet
basis function has greater breakthrough curve error t h a n the Daubechies4
wavelet function at all sample scales. The error decrease between scales
is also generally slower for the Haar basis. This indicates t h a t t h e H a a r
wavelet basis function does not interpolate the finer scale variability as
well as the Daubechies family. This effect is again primarily due t o the
difference in t h e smoothing properties between the Haar and Daubechies
wavelet basis functions. Since our known field (a pseudo-fractal) is rela-
tively smooth, the Daubechies4 wavelet function performs better. Similar
results should be expected for correlated stochastic grids.
§5. C o n c l u s i o n s a n d R e c o m m e n d a t i o n s
A primary purpose of this initial work was to investigate the usefulness
of t h e wavelet technique for reconstructing fine scale variability and mov-
ing multi-scale information. A secondary purpose was to use the wavelet
extrapolation technique to begin to understand what, if any, optimal scales
may exist. Based on the results discussed in Section 4, a number of con-
clusions can be stated:
Figure 17. Shaded contour map of fine scale ln(JsT) field reconstructed
using Haar basis function wavelet.
bu-.
^
o^ 45-
l_
2 40-
LU
CD 35-
£
3
ü 30-
25- b
13
2 20- H
sz
V
CO 15-
ω
m 10-
c
CO
φ h-
:> 0 1 1 1 r-ü 1
10 100 1000 10000 100000
Number of Samples
Figure 18. Effect of different wavelet functions on the mean break-

through curve error for the 128 X 128 pseudo-fractal field.
• First, we have demonstrated t h a t the wavelet reconstruction tech-

nique is able to incorporate the location and scale of each sample
in reconstructing a fine grid as evidenced visually and by the ex-
act resampling. T h e wavelet reconstruction provides a subjectively
b e t t e r representation of the known values t h a n the other compared
methods as shown by Figure 9. T h e resulting breakthrough curve
mean error for the wavelet technique was similar to t h e conditional
simulation technique error, and b e t t e r t h a n t h e others. Thus, we
are confident in stating t h a t the wavelet reconstruction technique is
a viable m e t h o d to incorporate scale information in fine scale recon-
structions. Further work is needed to understand the sensitivity of
this m e t h o d to the sample suite, and to determine optimum sample
suite characteristics.
• Second, our present results indicate t h a t choice of wavelet function

is important. Smoother wavelet functions, like t h e Daubechies fam-
ily, recreated the synthetic fractal and correlated stochastic distri-
butions better, as well as enabled a more rapid reduction in break-
through curve error. Future work is needed to determine what qual-
ities or properties of a wavelet basis function best aid reconstruction
of various porous media properties. It is hoped t h a t these wavelet
246 K. Brewer and S. Wheatcrafl
basis qualities might be directly related to geologic properties of

porous media.
• Third, an optimum sample scale appears to be related to the corre-

lation length of a correlated stochastic fine scale field. This implies
t h a t the maximum sampling benefit would be from samples at the
scale of the correlation length. Practically, this implies t h a t a site
characterization procedure should first be designed to determine the
correlation length of the hydraulic conductivity field, possibly via
sampling in one area at different scales. The procedure would be to
then cover the remaining site with samples at the correlation scale.
This type of site characterization procedure is quite different t h a n
what is currently considered "best professional judgement."
§6, C o n c l u d i n g R e m a r k s
Since fractals have correlation at all scales, the consistency of the opti-
m u m sample quantity (not scale) for the fractal field experiments is some-
what puzzling. One possible explanation might be t h a t there is a similarity
in the way wavelet techniques function compared to statistical relationships
of the number of required samples necessary for adequate representation
of a normal distribution. This may be an indication t h a t fractal distribu-
tions are best modeled by the wavelet method. Additional experiments are
needed.
Besides the areas of future research described above, part of a future
research plan should include increasing the robustness of the wavelet tech-
nique to handle irregular sample locations and overlapping sample volumes.
In addition, applying the reconstruction technique to a tightly controlled
field-generated multi-scale d a t a set is needed. Comparison of the recon-
struction of a fine scale grid, and various contaminant plumes should fur-
ther clarify t h e limitations and expectations of the wavelet reconstruction
application.
References
1. Carr, J. R., UVKRIG: A FORTRAN-77 program for universal Kriging,
Comp. and Geosci., 16, 211-236, 1990.
2. Chou, Kenneth C , A stochastic modeling approach to multiscale signal
processing, P h . D . Thesis, Massachusetts Institute of Technology, May
1991.
3. Chou, K. C , A. S. Willsky, A. Benveniste, and M. Basseville, Recur-
sive and iterative estimation algorithms for multi-resolution stochastic
processes, Proceedings of the 28th Conference of Decision and Control,
Tampa, Florida, IEEE, December 1989.
4. Cole, C. R. and H. P. Foote, Multigrid methods for solving multiscale

transport problems, in Dynamics of Fluids in Hierarchical Porous Me-
dia, Academic Press, 273-303, 1990.
5. Davis, J. M., R. C. Lohmann, F. M. Phillips, and J. L. Wilson, Ar-
chitecture of the Sierra Landrones Formation, central New Mexico;
depositional controls on the permeability correlation structure, GSA
Bulletin, 105, 998-1007, 1993.
6. Daubechies, I., Ten lectures on wavelets, p. 1, SI AM, Philadelphia,
1992.
7. Dykaar, B. B. and P. K. Kitanidis, Determination of the effective hy-
draulic conductivity for heterogeneous porous media using a numerical
spectral approach; 1, Method, Water Resour. Res., 28, 1155-1166,
1992.
8. Hohn, Michael E., Geostatistics and Petroleum Geology, Van Nostrand
Reinhold, New York, 1988.
9. Journel, A. G., and C. J. Huijbregts, Mining Geostatistics, Academic
Press, San Diego, 1978.
10. Mandelbrot, B. B., The fractal geometry of nature, W.H. Freeman and
Co., San Francisco, 1983.
11. Neuman, S. P., Statistical characterization of aquifer heterogeneities:
An overview, G.S.A., Special Paper 189, 81-102, 1982.
12. Pardo-Iguzquiza, E., M. Chica-Olmo, and J. Delgado-Garcia,
SICON1D, A FORTRAN-77 program for conditional simulation in one
dimension, Computers & Geosciences, 18, 665-688, 1992.
13. Press, W. H., B. P. Flanners, S. A. Teukolsky, and W. T. Vetterling,
Numerical Recipes: The art of scientific computing, 2nd ed., Cambridge
Univ. Press, Cambridge, 1992.
14. Smith, L., and F. W. Schwartz, Mass transport, 1, A stochastic analysis
of macroscopic dispersion, Water Resour. Res., 16, 303-313, 1980.
15. Strang, G., Wavelets and dilation equations,: a brief introduction,
SIAM Review, 31, 614-627, 1989.
16. Tompson, A. F. B., R. Ababou, and L. Gelhar, Application and use
of the three dimensional turning bands random field generator: single
realization problems, MIT Department of Civil Engineering, Report
Number 313, 1987.
17. Turcotte, D. L., Fractals and chaos in geology and geophysics, Cam-
bridge University Press, Cambridge, 1992.
18. Voss, R. F., Fractals in nature: From characterization to simulation, in
The Science of Fractal Images, H.-O. Peitgen, and D. Saupe, Springer-
Verlag, New York, 1988.
19. Warren, J. E., and F. F. Skiba, Macroscopic dispersion, Trans. S.P.E.,
231, 215-230, 1964.
20. Wheatcraft, S. W. and S. W. Tyler, An explanation of scale-dependent
dispersivity in heterogeneous aquifers using concepts of fractal geome-

try, Water Resour. Res., 24, 566-578, 1988.
21. Wornell, G. W., and A. V. Oppenheim, Estimation of fractal signals
from noisy measurements using wavelets, IEEE Trans. Sig. Proc, 40,
611-623, 1992.
This work was primarily funded by U.S. E.P.A. Grant CR817379, "Site
Characterization for Hazardous Waste Sites." This work has not been re-
viewed by EPA, and does not necessarily represent the position or opinions
of EPA. This work was also partially funded by the Desert Research Insti-
tute, Water Resources Center, Reno, Nevada.
Kevin E. Brewer
Department of Geological Sciences/ 172
University of Nevada, Reno
Reno, Nevada 89557
USA
e-mail: kevinbUhydro.unr.edu
Stephen W. Wheatcraft
Department of Geological Sciences/ 172
University of Nevada, Reno
Reno, Nevada 89557
USA
e-mail: steveuhydro.unr.edu
Wavelet-Based Multifractal Analysis of Non-Stationary and/or
Intermittent Geophysical Signals
Anthony Davis, Alexander Marshak, and Warren Wiscombe
A b s t r a c t . We show how some of the most popular wavelet transforms can be

used to compute simple yet dynamically meaningful statistical properties of a
one-dimensional dataset representative of a geophysical field or time-series. The
observed properties of turbulent velocity fields and of liquid water density in clouds
are used throughout as examples along with theoretical models drawn from the
literature. We introduce a diagnostic tool called the "mean multifractal plane,"
simpler than (but representative of) the two well-known multifractal hierarchies
of exponents H(p) and D(q), associated with structure functions and singular-
ity analysis respectively. It is used to demonstrate the pressing need for a new
class of stochastic models having both additive/non-stationary and multiplica-
tive/intermittent features.
§1. Introduction and Overview

Geophysical systems are notoriously complex, driven by external forc-
ing and ridden with internal instability due to their generically nonlinear
dynamics. Whether traceable to the break-up of larger ones or the am-
plification of smaller ones, structures are present on all observable scales.
The signals we depend on to probe these systems are naturally "rough"
and/or "jumpy" (i.e., non-differentiable and/or discontinuous) and little in-
sight is gained by statistical analysis unless the methods we adopt are fully
adapted to this situation. Traditional approaches include one-point his-
tograms (typically searching for Gaussianity), two-point auto-correlations
(typically searching for an exponential decay) and energy spectra (typically
searching for dominant frequencies). They are better suited for electrical
engineering applications than for geophysical research, although all of the
above listed statistics can be used in ways and circumstances we will de-
scribe in various parts of this paper. In the last couple of decades, we have

ISBN 0-12-262850-0
250 A. Davis et al.
seen Fourier analysis complemented by wavelet analysis, an increasing in-

terest in lognormal and power-law distributions and an explosion of ideas
centered on scale-invariance and fractals.
Scale-invariance is the natural framework for developing statistical
tools that account for all scales in presence at once. We focus on random
quantities that depend parameterically on scale - an idea idiosyncratic to
wavelet theory as well - and seek power-law behavior with respect to the
scale parameter for their statistical properties. If a single power-law expo-
nent is sufficient to describe all the statistics within a whole family, we talk
about monoscaling and a "monofractal" model is called for; otherwise (the
most general case), we are dealing with multiscaling and "multifractal"
models are in order. Multifractal statistics have proven useful in almost
every reach of nonlinear science, in the laboratory as well as in computa-
tional experiments (including the exploration of deterministic chaos). It
is hardly surprising that geophysical applications of multifractal and/or
chaos theory have been so successful, judging by the growing number of
researchers involved. However this says more about the geophysical com-
munity's data analysis needs and about the expectations raised by the
scaling paradigm than about the degree of maturity reached by multifrac-
tal statistical methodology. Many questions remain open on issues ranging
from the preprocessing required by certain procedures (cf. introductory re-
marks to section 4) to the validation of multiscaling studies (e.g., "Is the
observed multiplicity simply due to finite size effects?" ) but also on more
fundamental issues, for instance, the role of stationarity.
Because we depend on the spatial coordinate to perform all averaging
operations when analyzing data, our first task is to determine which aspects
of the signal are more likely to be stationary. In the upcoming section (and
the Appendix) we show that, within the framework of scale-invariance, this
can be done using spectral analysis. Assuming thereafter that the geophys-
ical process of interest is non-stationary but with stationary increments (a
frequent occurrence in nature), we turn to p-th order structure functions
in section 3. How can they be obtained from the wavelet transform of the
data? What are the general properties of the associated family of expo-
nents C(p)? How do they relate to geometrical, spectral and dynamical
concepts? Another option altogether is to derive a stationary "measure"
from the non-stationary signal, e.g., by taking absolute differences between
(appropriately separated) points; one can then apply some form of sin-
gularity analysis to it (we simply consider here the g-th order moments
of the coarse-grained measure). The same questions as above are asked
about this approach and associated family of exponents K(q) in section 4.
The most important outcome is the mapping of structure functions to non-
stationarity and of singularity analysis to intermittency, two challenging
aspects of nonlinear dynamics that happen to have well-defined statistical
Multifractal Analysis of Geophysical Signals 251
meanings. Often however, two large families of exponents is too much to

handle on practical grounds. In section 5 we briefly describe a way of re-
lating t h e m in an (almost) one-to-one fashion and go on to select a single
exponent from each approach. This enables us to devise a plot where one
axis, H\ = C(l) 5 measures the non-stationarity of the system and t h e other,
C\ = K'(l), its intermittency; we call this the "mean multifractal plane"
and discuss some possible applications, most importantly to the relatively
new field of multi-afline [57] stochastic modeling. Finally we summarize
and offer some concluding remarks in section 6.
We have surveyed the literature on a per topic basis: stationarity
and spectral analysis (section 2, Appendix), structure functions and multi-
affinity (section 3), singularity analysis and multiplicative cascade models
(section 4). The wavelet literature per se is already enormous b u t by-
and-large wavelet practitioners rely heavily on the existence of the inverse
wavelet transform [14]. Signal reconstruction is exact (no information is
lost) and, when m a d e approximate, it is generally in the spirit of d a t a
compression. Computational efficiency is often sought using t h e orthonor-
mality of many well-known wavelet bases [15]. These are not our concerns.
In statistical analysis, a large amount of information is thrown away in
the process of extracting a small number of sure quantities (i.e., averages)
from a large number of random ones (i.e., the data) and furthermore the
computational burden is slight in comparison to t h e overhead in d a t a pro-
duction. There is also a large body of multifractal literature, reviewed here
as needed. We will borrow heavily from the relatively small (but steadily
growing) number of papers t h a t use wavelets in connection with multi-
fractal models a n d / o r d a t a structures (generally pertaining to turbulence).
Most of these publications appeared in physics journals, emanating from
the group spearheaded by Alain Arnéodo of the Centre de Recherche Paul
Pascal (Pessac, France). His lectures and the papers he co-authored gave
us t h e impetus for this study.
§2. P r e l i m i n a r y C o n s i d e r a t i o n s o n S t a t i o n a r i t y , E r g o d i c i t y a n d
Scale-Invariance
2 . 1 . D a t a specification a n d s p e c t r a l analysis
We consider given a discrete set of N + 1 d a t a points:
fi = f(xi), Xi = U (i = 0 , l , . . . , J V ) , (1)
sampled from / ( # ) , x G [0, L], some scalar geophysical field (x is position) or
time-series (x is time). T h e overall length of the one-dimensional dataset is
L and / is t h e grid constant (or 1//, t h e sampling r a t e ) . We will furthermore
assume t h a t
i V = y » l . (2)
252 A. Davis et al.
In many cases, a time-series is really a picture of a geophysical field.

Consider for instance the time-series f(x) in Figure l a . It represents liq-
uid water content (LWC) captured from an airborne platform during F I R E
(First ISCCP Regional Experiment) while flying through an extended ma-
rine stratocumulus (StCu) deck off the southern coast of California on J u n e
30, 1987. (This type of cloud is of particular interest to climate modelers
due to its frequent occurrence over vast regions, hence a first order effect
on the planetary radiation budget.) Since the aircraft was flying at roughly
constant speed and altitude, we can view this as a snap-shot of the local
atmospheric LWC field, sampled along a linear transect (Taylor's frozen
turbulence hypothesis). The parameters of this dataset are / « 5 m and
N = 2 1 4 = 16384, hence L « 82 km. We note the "roughness" of this graph
which we will quantify in section 3. Figure l b presents a related dataset
ε(χ) based on absolute differences over a certain scale (to be discussed
shortly). We note the inhomogeneous "spikiness" of this field which, in the
tradition of turbulence studies, we view as a direct manifestation of inter-
mittency (sudden bursts of high frequency activity) in the original signal.
We will quantify intermittency in this sense in section 4.
T h e information in Equation (1) can be represented by a discrete
Fourier transform: fi-N/2 = îiî-N/2)^ h-N/2 = (* — N/2)/L (i =
0 , 1 , . . . ,iV). Being interested in finding the statistically robust features
of f(x), we start with the energy spectrum:
E/(k) = ( £ |/(±*)|2>, 1/L < fc < 1/2/, (3)

±
dropping subscripts for simplicity. This is probably the most popular statis-
tic in time-series analysis, beyond 1-point p.d.f.'s (i.e., simple histograms
of /-values as in Figure l a ) . T h e energy spectrum E£(k) of the ε-field in
Figure l b is computed in the same way and b o t h spectra for the d a t a in
Figure 1 are plotted in Figure 2.
In Equation (3) we use (·) to denote an "ensemble" average which in-
volves in principle every possible realization of the random process f(x).
In geophysical practice however, we are generally provided with the out-
come of a small number of experiments - possibly even a single one - and
we are forced to sacrifice spatial information to obtain spatial averages,
as estimates of their ensemble counterparts. Processes for which spatial
averages converge to the corresponding ensemble statistics in the limit of
large averaging sets are called "ergodic." Estimating an energy spectrum
with a single realization, dropping the triangular brackets in Equation (3),
amounts to making an ergodicity assumption. Throughout this paper, we
use the same notation for ensemble averages and their spatially obtained
estimates. The interested reader is referred to [32] and [29] for a detailed
analytical study of multiplicative cascade models where cases of ergodic-
ity violations can be found; this important class of models will be briefly
discussed in section 4 and in the Appendix.
410
time (sec)
F i g u r e 1. A horizontal probing of atmospheric liquid water content

(LWC), showing the internal structure of a marine stratocumulus deck
during FIRE, (a) The raw LWC data / ( x ) , sampled at 20 Hz (every / « 5
m, aircraft speed « 100 m / s ) for 820 s (L « 82 km). Also indicated is
the negatively skewed 1-point p.d.f. (the downwards spikes responsible
for this skewness are probably related to instabilities that entrain of dry
air from cloud top), (b) An absolute gradient field ε(χ) associated with
f(x) in panel (a). The differences are taken at the lower end (η) of
the non-stationary scaling regime defined by the structure functions (cf.
Figure 5a) or the energy spectrum (cf. Figure 2); in this case, we use a
distance of 4 pixels (η « 20 m ) .
254 A. Davis et al.
I i I i I i I i I i I i I i I i I i I i I i I i I i I . I i
lo
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 92(kL)
F i g u r e 2 . Scale-invariance of the spectral densities for the datasets in

Figure 1. Log (base 2) of the energy E(k) versus log 2 (fcL) for 2/L <
k < 1/2/. These bounds are set upwardly by the Nyquist frequency
and downwardly by the minimal noise reduction option using routine
S P C T R M (two non-overlapping partial spectra are averaged) in [50].
We also give a corresponding log-scale for r/l (using r « 1/k with L/l =
2 1 4 ) with several remarkable values highlighted. We notice the lack
of small scale (high k) information for the e-field which is formed by
taking differences over 4 grid points (corresponding to a distance η «
20 m). As advocated by Marshak et ai [40], the energy as well as
the (linearly discretized) wavenumbers are averaged inside octave-sized
bins. Absolute slopes (power-law exponents) are indicated in both cases,
showing this LWC dataset to be non-stationary (β > 1) over scales from
20 m to at least 5 km (50 s in Figure 1) while its associated ε-field
is stationary (ße < 1). Davis et ai [18] show that the transition to
stationary behavior (E(k) « constant) at r > R « 3 - 5 km (log 2 (fcL) <
2.5) is traceable to a single event (namely, the intense « tf-sized "dip
at « 700 s in Figure 1) and therefore is not robust with respect to the
addition of new data. (An improved estimate of this "integral" scale for
LWC in marine StCu is found to be « 20 km.)
2.2. Stationarity and the spectral exponent

Spectral analysis is often used to find dominant frequencies (hence
time-scales) in a time-series, viewed as a harmonic (and possibly determin-
istic) signal contaminated with r a n d o m noise, t h e latter having a charac-
teristically continuous spectrum. However, in natural systems it is usually
futile to try to separate a smooth signal from the noise. These systems are
turbulent and the signals they produce are noise. T h e prevailing type of
noise in N a t u r e is furthermore "scale-invariant," meaning t h a t it is free of
characteristic scales, at least over some large "scaling" range:
q<r<R (4)
which will depend on the type of geophysical signal. T h e corresponding
spectral densities reflect this by showing power-law distributions of energy
with respect to scale (r ~ 1/&):
Ef(k) oc k'ß, Ee{k) oc k~ß' (5)
over t h e associated range of wavenumbers. In the following, we will assume
for simplicity t h a t the dataset is scale-invariant in the full instrumentally
accessible range of scales (/ < r < L) in our general discussions. In our
data-specific discussions (Section 3.4 and Section 4.4), we relax this as-
sumption. It is noteworthy t h a t t h e exponent ß can b e retrieved from a
"wavelet spectrum" [21] under conditions spelled out by Fantodji et al. [20].
As shown in Figure 2, the LWC datasets scale over most of the accessi-
ble range of scales with ß « 1.4 for the LWC d a t a in Figure l a and ße « 0.7
for t h e associated absolute gradient field in Figure l b ; details on these and
several other related LWC time-series are given by Marshak et al. [40] and
Davis et al. [16]. Duroure and Fantodji [19] compare the scaling of Fourier-
and wavelet spectra for droplet density (rather t h a n LWC) in StCu as well
as for several stochastic models, similarly for Gollmer et al. [27] with liquid
water p a t h (the vertical integral of LWC retrieved from passive microwave
radiometry).
One-point p.d.f.'s make no use of the ordering of the data, leaving wide
open the possibility of correlations between /-values at different points
(more loosely, "structures" or "patterns" in the d a t a ) . T h e energy spec-
t r u m partially describes these correlations and has one advantage over
many other popular statistics, including auto-correlation functions: it re-
mains well-defined whether or not t h e process is "stationary," i.e., statisti-
cally invariant under translation. In the Appendix, we discuss stationarity
in some detail because, being a prerequisite for ergodicity, it is necessarily
invoked in every d a t a analysis task where spatial averaging is performed as
a surrogate for ensemble averaging. In the same Appendix, we argue t h a t
the spectral exponent can in fact be used to distinguish between stationary
(ß < 1) and non-stationary (ß > 1) scaling situations.
256 A. Davis et al.
According to this criterion, Figure l a illustrates non-stationarity, at

least for scales in t h e range 20 m « 4/ < r < L / 1 6 « 5 km (cf. Figure 2).
Although this may seem paradoxical to the casual observer, Figure l b il-
lustrates a stationary process. Indeed it is characterized visually by long
stretches with little activity suddenly interrupted by spikes coming in a
wide variety of heights and with different degrees of clustering. A priori
this sounds rather non-stationary. Conventional wisdom wants stationar-
ity to be the tendency t h a t "things are sort of the same at all times;" a
b e t t e r way of envisioning it would be a tendency to "quickly come back to
typical (i.e., most probable) values." T h e picture conveyed by t h e former
description is too restrictive, apparently excluding the strong deviations
from average values observed in Figure l b . We will substantiate this view
in the Appendix as well.
§3. Q u a n t i f y i n g a n d Qualifying N o n - S t a t i o n a r i t y w i t h S t r u c t u r e
Functions
In this section we interest ourselves primarily in non-stationary pro-
cesses with stationary increments (1 < ß < 3) which are ubiquitous in na-
ture. So, following t h e ideas surveyed in t h e Appendix, we will focus on t h e
field of increments of a random process f(x) over a distance r: Δ / ( Γ ; X) =
f(x + r / 2 ) - f(x - r / 2 ) , r / 2 < x < L - r / 2 (0 < r < L). T h e statistical
moments of | Δ / ( Γ ; Χ ) | are a two-point statistic known as "structure func-
tions of order p " and they are independent of x: ( | Δ / ( Γ ; X)\P) = ( | Δ / ( Γ ) | Ρ ) ;
this incremental stationarity is a necessary condition to feel justified in es-
timating ( | Δ / ( Γ ) | Ρ ) by averaging | Δ / ( Γ ; Χ ) | Ρ over x. In most cases small
increments are a frequent occurrence so one cannot take arbitrarily large
negative values for p.
Arnéodo, Bacry and Muzy have shown in a series of papers ([5], [7], [9]
and [45]) how wavelets can be explicitly used to perform structure function
analysis in its standard and more general forms. We summarize their ideas
in the upcoming subsection then review the implications of scale-invariance;
finally, we discuss turbulence studies and apply the technique to the LWC
d a t a presented above, underscoring the dynamical meaning of structure
function exponents in connection with non-stationarity.
3.1. W a v e l e t s for s t r u c t u r e function analysis

Define the "indicator" function of the open r-sized interval (—r/2, + r / 2 ) ,
r >0:
τ ( λ / 1 \x\ < r / 2 fa Λ
= (6a)
W ( 0 |*| > r/2 '
In the following, we will also need
Ψσ{χ) = -^-= exp(-x2/2a% (6b)
a Gaussian with vanishing mean and standard deviation σ.

T h e field Δ / ( Γ ; X) can be viewed as t h e convolution product of f(x)
with dlr(x) = δ(χ + r / 2 ) — S(x — r / 2 ) where δ(-) designates Dirac's gener-
alized function. So we have
Af(r;x) = [f*dlr](x)= If(xf)[S(x + r/2-x,)-S(x-r/2-x,)]dx\ (7)
where boundary complications are dealt with by making f(x) periodic with
period L. Since the kernel dlr of the integral operator in Equation (7) is
parameterized by a variable scale r, we are reminded of a wavelet transform.
Indeed, it is easy to show t h a t
Δ/(Γ; x) = lj f(x')[-dh(^-)] dx'. (8)

Muzy et al. [45] call -dli(x) = S(x - 1/2) - δ(χ+ 1/2) a "poor-man's
wavelet" and suggest a more general approach to increment estimation
based a priori on a restricted class of anti-symmetric wavelets Ψ(#), or-
thogonal in particular to constant functions but not to linear ones, t h a t
furthermore contain a single oscillation. Along with —dli(x), Figure 3
shows classical examples:
Φσ(α) = - V i / 2 0 * 0 = 4 ^ i / 2 ( z ) , (9a)
on the one hand, and
W
*«M = { r W>; <9b)
on t h e other. T h e latter and simpler (piece-wise constant) case is the

well-known Haar wavelet. In analogy with Equation (8), we consider t h e
continuous wavelet transform
Γ Φ [/](α,6) = - [ f(x)y(ZzlL)dxJ a > 0, b e ft, (10)

a J a
where the wavelet is real (so Φ is used instead of its conjugate Φ*). Notice
the non-standard normalization in Equation (10): 1/a appears, rather t h a n
1/y/a. This inhibits t h e strict orthonormality of the wavelet basis (not
required in the following) b u t ensures t h e constancy with respect to a of
the integral of t h e positive p a r t of the rescaled wavelet, a~l θ [ Φ ( χ / α ) ] where
θ ( · ) is t h e Heaviside step function. This facilitates scaling comparisons.
258 A. Davis et al.
dl,(x)
ψ Η (χ) >
ii
λ
\
Γ
ψ β (χ)
Figure 3. Acceptable wavelets for structure function analysis. The

first derivative VG(x) of (2/π)~ 1 / 2 exp(-2z 2 ); Haar' s piecewise-constant
wavelet Φ Η ( # ) and — dii(#), the "poor-man's wavelet" of Muzy et al.
[45], i.e., two J-functions of opposite sign. Notice that all three are anti-
symmetric but have ever smaller supports: all of 9ft, a compact subset,
two discrete points.
T h e premise of Arnéodo and his co-workers is t h a t , being increment-

like, the random wavelet coefficients in Equation (10) have statistical prop-
erties similar to those of Δ / ( Γ ; X) with scale r playing the role of t h e di-
lation parameter a and position x t h a t of the translation parameter b. In
particular, (|Τψ[/](α,6)| ρ ) will be independent of b and we are justified
in estimating this (absolute) moment by averaging over t h e location. In
summary, we expect t h a t
<|Δ/(Γ)|*> «-jr y |r*[/](r,6)|*db, 0 < r < L , p > - l , (11)
independently of the choice of Φ (say) in Equations (9a) and (9b). Using

a variety of numerical and analytical methods, Arnéodo et al. ([5], [7]),
Bacry et al. [9] and Muzy et al. [45] have shown this conjecture to hold
(in the sense of scaling described below) for a wide class of r a n d o m pro-
cesses. We refer to [45] (and references therein) for t h e restriction on p
in Equation (11). These authors also show how wavelets with more os-
cillations (e.g., successive derivatives of φσ(χ), orthogonal to higher order
polynomials) can be used to find the scale-invariant properties of a struc-
t u r e function-like statistic in Equation (11) for p < — 1, which is dominated
by regular points where t h e signal f(x) is differentiate.
3.2. Scale-invariant s t r u c t u r e f u n c t i o n s , g e n e r a l i t i e s
R a n d o m scale invariant fields/signals f(x) are characterized by scale-
conditioned statistics t h a t follow power-laws with respect to t h e scale pa-
rameter. Equation (11) describes a way of estimating such a statistic (r is
held constant during the spatial averaging), we thus anticipate
<|Δ/(Γ)|*>«Γ«*>, (12)
over some large range of scales [77, R] which may partially or entirely overlap
with the instrumentally accessible range [/,L].
Some general statements can be made about t h e family of exponents
ζ(ρ). Firstly, one exponent is known a priori, from definitions: ζ(0) = 0.
Secondly, ζ(ρ) will be a smooth differentiate function of p no m a t t e r how
rough the d a t a / ( # ) , being essentially a sum of many exponential functions
of p . Thirdly, it can be shown t h a t , if the signal f{x) has absolute bounds,
then ζ(ρ) is monotonically non-decreasing ([23] and [39]). Finally, it can
be shown on more general grounds t h a t ζ(ρ) is concave (ζ"(ρ) < 0).
To see this, choose units of length where L = 1 and / - u n i t s where
(\Af(l;x)\p) « 0 ( 1 ) ; we then have (\Af(r;x)\P) « r<&\ making explicit
the weak dependence of the prefactors on p. It follows t h a t — £ ( p ) l n r is
the second characteristic (or cumulant generating) function ln(exp(— ρξΓ))
of the random variables £ r = — In | Δ / ( Γ ; Χ)\ and is therefore convex [22].
Stationary processes of course have stationary increments b u t these
will scale trivially due to the invariance under translation: ζ(ρ) = 0 (scale-
independent increments). However, due to the effects of finite spatial res-
olution, even theoretical models lead to numerically small C(p)' s ( s e e [39]
for an example). At t h e other end of the non-stationarity scale, we have
continuous functions with bounded non-vanishing first derivatives, yielding
| Δ / ( Γ ; X)\ OC r (for almost every x) hence ζ(ρ) = p.
Any concave ζ(ρ) with ζ(0) = 0 allows the definition of a hierarchy of
exponents:
ζ
Hip) = -ψ- (13)
By "hierarchy" we mean a monotonie function, in this case non-increasing.
T h e reader is referred to Parisi and Frisch's [48] original paper on the mul-
tifractality of turbulent velocity signals/fields for t h e interpretation of the
ζ(ρ) and H (p) functions in terms of variable orders of singularity. ( "Sin-
gularity" is used here in the sense of Hölder-Lipshitz heuristics: how does
260 A. Davis et al.
h(x) in | Δ / ( Γ ; : Τ ) | OC rh^ vary, statistically speaking, with xl)

Obtaining ζ(ρ) and/or H(p) is the goal of this whole approach, provid-
ing us with a means to characterize non-stationarity, in both quantitative
and qualitative terms (examples to follow).
3.3· Geometrical, spectral and dynamical implications of
structure function statistics
Processes with a variable H(p) have recently been described as "multi-
affine" [57] or "non-stationary multifractals" [39], leaving processes with
a constant H(p) or linear ζ(ρ) to be "mono-affine" or "non-stationary
monofractals." This last class of random functions is indistinguishable
from "fractional" Brownian motions (used as examples in Section 3.4).
Such cases are often referred to as "self-affine" in the literature but this is
an unnecessarily restrictive view of statistical self-affinity.
The expressions multi- and mono-affinity refer in fact to a well-known
geometrical meaning of the p = 1 structure function. The quantity ( | Δ / ( Γ ) | )
can indeed be related to the fractal structure of the graph g(f) of / ( # ) ,
viewed as a statistically self-affine geometrical object in 2-dimensional Eu-
clidean space. We have [38]
Hx = H(l) = C(l) = 2 - DgU) > 0, (14a)
where Dg^ is the fractal dimension of g(f) or, in slightly different words,
Ηχ is the codimension of g(f) and is also known as the "roughness" ex-
ponent. This immediately tells us that the largest possible value for Hi
is 1, attained for (almost everywhere) differentiable functions which have
smooth graphs of non-fractal dimension Dg^ = 1, like any line or other-
wise rectifiable curve. At the opposite limit of Hi = 0 (stationarity), we
find graphs that fill space: £>#(/) = 2, a direct consequence of the necessary
discontinuity of stationary scaling processes discussed in the Appendix.
An infinite number of exponents is needed to describe multi-affine pro-
cesses completely. For instance, there is no general relation between Hi
and the spectral exponent which is however connected with the p = 2 case:
β = C(2) + 1 = 2H(2) + 1 > 1, (14b)
the Wiener-Khintchine relation for non-stationary scaling processes (cf.
Appendix). We will nevertheless retain Hi as the single most important
exponent in the whole H(p) hierarchy. The geometrical significance given
in Equation (14a) is not as important as the fact that Hi defines as simply
as possible the linear trend in C(p)? c f· Figure 5b. In summary, we view
Hi as a direct quantifier of the system's non-stationarity and the complete
ζ(ρ) or H(p) functions as a means to qualify this non-stationarity.
We can take this statement one step further than its purely statistical
sense. In the present circumstances, the expression "non-stationarity" is
dynamically correct. In the Appendix we recall that non-stationary scale-

invariant random processes, ubiquitous in nature, are characterized by large
wanderings from mean (or rather most probable) values with relatively long
return times. At very long times we expect a transition to stationary be-
havior to occur at least for geophysical fields which have absolute bounds.
In nonlinear dynamical systems (of ODEs), stationary (time-independent)
solutions generally occupy a very small portion of parameter space; an
even smaller portion are stable and even these are generally unstable with
respect to finite perturbations. In the chaotic regime time-dependent (non-
stationary) solutions starting close to each other wander away exponentially
fast at small times; at long times the outer size of the (usually strange)
attractor puts bounds on their relative deviation. Shifting from computa-
tional to natural systems we know, at best, the governing nonlinear system
of coupled PDEs (usually with forcing and dissipation terms); these are
also generically unstable and their solutions are attracted into a certain
state of "turbulence." At worst, we have no clue of the governing equa-
tions and rely primarily on observations for insight. The case of clouds
and many other geophysical systems is intermediate; we advocate a more
systematic use of structure functions, wavelet-based or not, to characterize
the non-stationary aspects of their dynamics.
3,4. Examples from theory, turbulence and cloud structure

In Figure 4 we see five random processes, all theoretical, with different
degrees of non-stationarity. At the top, we find a sample of "flicker" or 1 / /
noise (ß = 1) which is marginally stationary: ζ(ρ) = H(p) = 0. At the bot-
tom, we find a noiseless trend which represents everywhere differentiability:
£(p) = p and H{p) = 1. Between these two extremes, we find three samples
of "fractional" Brownian motion (fBm), illustrating less-and-less station-
a r y : C(p) = pHi or H(p) = Hi with Hi = 1/3,1/2,2/3. The relation
C(p) = pH\ n a s a simple probabilistic meaning. Using Equation (13), it
implies (\Af(r)\*>) « (|A/(r)|) p , i.e., that the p.d.f. of \Af(r;x)\ is narrow
enough to enable a simple dimensional argument to relate quantitatively all
moments. The Gaussian p.d.f.'s used in Figure 4 for fBm are good exam-
ples of weakly variable increments. These models however are monoscaling
(indeed mono-affine) whereas the next two examples argue that natural
processes tend to be multiscaling (hence multi-affine).
The most notable structure function analysis in the turbulence liter-
ature is due to Anselmet et al. [1] who determined ζ(ρ) for wind tunnel
flows at high Reynolds numbers. The authors find Hi « 1/3 and, more im-
portantly, confirm Kolmogorov's [34] prediction that £(3) = 1 (cf. section
5). They also extend their averaging procedures up to p = 18; this is not a
simple task because the larger the order p of the statistical moment in Equa-
tion (12), the more it is dominated by extreme events in Δ / ( Γ ; Χ ) which
262 A. Davis et al.
Ηι=0
Η,-1/2
Hi=2/3
F i g u r e 4 . Different amounts of the same kind of non-stationarity.

Samples of Mandelbrot's [38] fractional Brownian motion, B(x), 0 <
x < L = 1024, for # i = 0 , 1 / 3 , 1 / 2 , 2 / 3 and 1, from top to bottom.
The mid-point displacement algorithm was used [49]: take B(0) = 0
and B(L) = 1; generate a zero-mean Gaussian deviate with standard
deviation σ = y/21-^ - 1 (0 < # i < 1) and add it to [J5(0) + B(L)]/2
to obtain B(L/2)\ divide σ by 2Hl and repeat between x = 0, L/2 for
B(L/4), and x = L / 2 , L for B(3L/4); proceed similarly to smaller scales
(10 divisions in this case). The basic non-stationarity parameter H\ in-
creases from the marginally stationary case ( 1 / / noise at H\ = 0) to the
extreme case of a noiseless linear trend (B(x) = x/L at Hi = 1 ) . This
family of processes are all non-stationary (β = 2i?i + l > 1) with station-
ary increments (β < 3) and can be said to have the same "kind" of non-
stationarity, namely, the monoscaling (or mono-affine) kind: H(p) = H\.
Other (multi-scaling) types of scale-invariant non-stationary models are
listed in the text.
are poorly sampled (by definition) in d a t a analysis situations. T h e authors

found not only a nonlinear C(p)> thus establishing t h e multiscaling, b u t
used it for discriminating between the competing so-called "intermittency"
models. Anselmet and his co-workers used s t a n d a r d structure functions
based on two discrete values in their time-series. Applications of bone fide
wavelets to the analysis of turbulent signals abound and have been exten-
sively reviewed by Farge [21]. Of particular interest here are results on
velocity time-series generated in three-dimensional turbulent flows: Argoul
et al. [2] use the symmetric wavelets discussed in Section 4.1 below to visu-
alize the turbulent cascade process (rather t h a n perform scaling analyses)
while Muzy et al. [44] use the same symmetric wavelets as well as their
anti-symmetric counterparts discussed above in scaling analyses.
Figure 5a shows ( | Δ / ( Γ ) | Ρ ) for t h e LWC d a t a in Figure l a , using the
traditional discrete value calculation of structure functions. T h e scaling
range [77, R] where the exponents are defined is clearly marked. However,
Davis et al. [18] show t h a t the break t h a t defines the upper limit of R « 5
km can b e traced to a single event (a "dip" in LWC of approximately t h a t
width t h a t occurs c. 700 sec. into t h e flight on Figure l a ) . In other words,
this estimate of this transition point to large-scale stationary behavior,
( | A / ( r ) | p ) « constant, is not robust with respect to adding to t h e average
the four other LWC datasets obtained from the same instrument during
F I R E . A b e t t e r estimate (using all five flights) is found to be « 20 km
and, viewing R as a statistic, we are simply in presence of a violation of
ergodicity (another one is described in Figure 9). Assuming it is robust,
the special scale R is known as t h e "integral" scale of the process (see Ap-
pendix). In Figure 5b we have plotted ζ(ρ) and highlighted two remarkable
values: ζ(1) = H\ « 0.28, not unlike turbulent velocity, and ζ(2) which is
simply related to the spectral exponent β, as described in Equation (14b).
We notice the strong evidence of non-stationarity (ζ(ρ) is non-vanishing)
in marine StCu and their multifractality (in the sense of multi-affinity, ζ(ρ)
is nonlinear).
Multi-affine modeling is the synthetic counterpart of structure function
analysis, a relatively new field judging by the following literature survey
t h a t we believe to b e exhaustive at t h e time of publication. Schertzer and
Lovejoy [52] suggested the idea of low-pass power-law filtering in Fourier
space (or "fractionally" integrating) singular multiplicative cascade models;
the latter are described in some detail below. Cahalan et al. [12] describe
a class of "bounded" cascade models, shown by Marshak et al. [39] to have
ζ(ρ) = min[pii, 1] where t h e smoothing parameter H goes from 0 (singu-
lar cascades) to 00 (Heaviside step functions). Viscek and Barabâsi [57]
describe a variant of t h e mid-point displacement algorithm for generating
fBm t h a t yields processes with multiscaling structure functions and com-
pares their theoretical C(p)' s to those empirically determined for turbulent
264 A. Davis et al.
(a)
10 12 14
η=20 m R-5km |og (r/ |)
A Ψ 2
= -5 _O-O-Ö- & --° ° ° °
O-
-Θ -o
,Q--O D D Q
V -10 τ-·β-
.D-B"0-43
^ . ^ ^ O o O
cr"
^ ^ · _ ^ * - X"* X χ X
-2 -is ,-* -X-
>^*"
— Θ- p=i
-20 - -B- p=2
• - O- p=3
X-
p=4
(b)
Αζ(ρ)
0.6
0.5
β-1=ζ(2) / ; . ·
0.4
0.3 Η=ζ(1)
\ /
0.2
Ύ
Ύ
0.1
Ύ
Ρ
0 0.5 1 1.5 2Î 2.5 3 3.5 4
F i g u r e 5. Structure function analysis of cloud LWC. (a) T h e quanti-

ties l o g 2 ( | A / ( r ) | p ) are plotted for l o g 2 ( r / / ) = 0 , . . . , 13 and p = 1 , 2 , 3 , 4 .
We have indicated the scaling range [?/, R] and the slopes are the expo-
nents ζ(ρ). (b) The corresponding ζ(ρ) function, demonstrating the
multi-affinity of LWC and highlighting the remarkable values at p = 1,2
discussed in connection with Equations (14a,b).
velocity. Arnéodo et al. [5] and Muzy et al. [44] use recursively gener-
ated fields (akin to cascades but with negative weights as well as positive
ones) smoothed by fractional integration. Finally, Benzi et al. [11] have
recently proposed a wavelet-based technique for generating fields with any
prescribed ζ(ρ) function.
We now turn to an alternative (and currently more popular) multifrac-
tal approach to statistical data characterization, singularity analysis, and
another challenging issue in nonlinear processes, namely intermittency.
§4. Quantifying and Qualifying Intermittency with Singular

Measures
In this section we turn our interest to random measures ε(χ) that are
stationary (βε < 1); "measures" are non-negative processes defined in prin-
ciple only through integrals. Geophysical data generally does not come in
this (stationary, non-negative) format, a notable exception being rain rate
[29]. However, a stationary measure can always be derived from a signal
(say) by finite differencing and taking absolute values (cf. Figures la,b).
Other options are possible. For instance, working with one-dimensional tur-
bulent velocity signals, Meneveau and Sreenivasan [41] take squares rather
than absolute values while Schmitt et al. [54] use "fractional" differenti-
ation (a high-pass power-law filtering). Working on two-dimensional im-
agery, Tessier et al. [56] experiment with gradients and Laplacians and,
working with cascade models, Lavallée et al. [36] demonstrate numerically
this makes little difference in the ensuing singularity analysis anyway.
We will assume given some stationary random measure ε(χ) > 0, 0 <
x < L, either on a grid (as for data) or in the continuum limit (as in the case
of a theoretical model). Here again, wavelet-based techniques in singularity
analysis were first investigated by Arnéodo et al. ([3], [4]). Their basic ideas
are surveyed in the first subsection before reviewing the consequences of
scaling; finally, we illustrate the whole approach with turbulence results
from the literature and with our own LWC data, relating the new exponents
to the dynamical concept of intermittency.
4.1. Wavelets and "scaling functions" for singularity analysis

The basic idea in singularity analysis is to degrade the resolution with
which the measure e{x) is "observed" (mathematically speaking, via inte-
grals over intervals of variable lengths). This is very close to the concept
of wavelet analysis (viewed as a mathematical microscope with a variable
magnification or, more precisely, field of view). The simplest approach is
to generate a sequence of spatially degraded or coarse-grained measures
e(r;x), r / 2 < x < L — r / 2 (0 < r < L), obtained by averaging ε(χ) over
[x — r / 2 , x + r/2]. This is equivalent to a series of convolution products of
266 A. Davis et al.
ε{χ) with Ir(x), as defined in Equation (6a):
e(r;x) = - [ e{x')dx' = - [ ε(χ')ΙΓ(χ-χ')αχ' = ~[ε*/ Γ ](χ); (15)

r r r
Jx-r/2 J
again eventual boundary complications are eliminated by making ε(χ) L-
periodic. Making use of the symmetry of Ir(x), we can rewrite Equa-
tion (15) as a wavelet-type transform:
e(r;x) = l fs(X>)Ii(^)dx'. (16)

r J r
The kernel Ιχ (χ) is however more akin to a "scaling function" than a wavelet
since it is everywhere non-negative, hence a non-vanishing integral. In fact
h{x) is the scaling function of the Haar wavelet Φ//(χ) defined in Equa-
tion (9b). In the spirit of wavelet analysis however, there is no reason
not to consider other weighting functions than I\(x) in Equation (16). A
priori likely candidates need only to be symmetric, non-negative and to
have a unit integral as well as an 0(1) "width" (say, at half-height). For
instance, one could take a Gaussian-shaped function, ψ\/2 (x) in Equa-
tion (6b); a piece-wise linear function defined as the integral of — Φ//(χ) in
Equation (9b), namely ΨΗ(Χ) = max{0,min[l — x, 1 + #]}, would do just
as well. Figure 6a illustrates all three possibilities.
More in step with standard wavelet analysis, Arnéodo et al. ([3], [4])
relax the conditions of non-negativity and having unit integrals; such well-
known wavelets as Mexican and French "hats" can therefore be considered.
Figure 6b shows these famous examples:
ψ Μ ( χ ) = - x / 2 ^ y ' ( x ) = >/2π[1 - χ2]ψι{χ), (17a)
and
f i W< i
VF(x) = I - 1/2 1 < \x\ < 3 (17b)
[ 0 |a?|>3
The latter being a piece-wise constant approximation of the former, using
a compact support. Here again there is no need for an orthonormal basis of
functions. The common feature of the wavelets Φ ^ a n d ^ F · , currently used
in singular measure analysis, as well as the scaling-type functions Ji, φχ/2
and φπ is symmetry (hence orthogonality to linear functions in particular)
and a degree of simplicity (i.e., less than two oscillations). In analogy with
Equation (16), we consider the wavelet-type transform
1 /* x—b
Τφ[ε)(α,ο) = - ε(χ)φ( )dx, a > 0, b G 9ft, (18)
G J G
where φ(-) designates any of the above analyzing functions. Notice the
same non-standard normalization here as in Equation (10), ensuring that
F i g u r e 6. (a) Weighting functions for singularity analysis. T h e sim-

plest is the indicator function of the interval of interest (all points are
equally weighted), I\{x) in Equation 6a. A (non-normalized) Gaus-
sian centered on the interval and of the same width can be used,
φο(χ) = ψ\/2{χ) m Equation (6b), or else a piece-wise linear func-
tion, ΨΗ(Χ) — m a x { 0 , m i n [ l — #, 1 -f #]}. Notice the positive values, the
symmetry and the relative simplicity (i.e., one peak at most). These
kernels are not unlike the "scaling functions" of wavelet theory, (b)
Acceptable wavelets for singularity analysis. T h e second derivative of
exp(—x 2 /2) or "Mexican hat" Ψ Μ ( # ) is illustrated as well as the sim-
pler (piecewise-constant, compact supported) "French top hat" Φ^(^) in
Equation (17b). Notice again the symmetry and the relative simplicity
(i.e., less than two complete oscillations occur).
268 A. Davis et al.
the integral of the positive part of φ(χ/α), i.e., θ[φ(χ/α)], is oc a. This

facilitates the following scaling comparisons; if ε(χ) varies little, then so
will 7^[ε](α, b) with respect t o b o t h b and a.
Arnéodo et al. ([3] and [4]) show t h a t , within the general framework
of scale-invariant (fractal- and multifractal) measures, the random wavelet
coefficients in Equation (18) behave statistically like e ( r ; x ) with r « a
and x « b; see also Ghez and Vaienti ([25] and [26]). In particular, the
(|Τ^[ε](α,6)| 9 ) will be independent of b and comparable to the 1-point sta-
tistical moment of order q for s(r; x), namely (e(r; x)q) = (e(r)q). Although
a weaker statement t h a n ergodicity, the statistical stationarity of e(r; x) jus-
tifies estimating these moments by averaging Equation (18) with respect
to b. In summary, we have
(e(ry) « j-J fo[e](r,b)\* db, 0 < r < L, q G » , (19)
independently of the choice of φ. Some restrictions might have to be im-

posed on the domain of q. For instance, if s(r; x) can take null values then
we must take q > 0. If, over and above the spatial averaging in the r.h.s. of
Equation (19), bona fide ensemble-averages are taken then non-trivial sta-
tistical effects lead to the divergence of the higher positive order moments
for many interesting theoretical models of e(x) ([37], [33], [52] and [53]).
The divergence of moments is also discussed by Holley and Waymire [32]
and G u p t a and Waymire [29] who furthermore clarify the issue of ergodic-
ity within the theoretical framework of multiplicative cascades as models
for ε(χ).
One might object to using coefficients for continuously overlapping
wavelets in Equation (19) for the averaging, not being independent random
numbers. If required, this minor problem (redundancy is often viewed as
desirable) can be fixed by replacing t h e integral in Equation (19) by a sum
over wavelets with somewhat more disjoint supports:
^0 L
' J
BL(r)
where [·] means integer value and BL(T) = {b G [0, L], b = ir (i =

0 , . . . , [L/r] — 1)} is the relevant set of disjoint "boxes" of size r.
Before considering the scaling properties of the average measures ( s ( r ) 9 ) ,
we must point out t h a t Arnéodo's group does not systematically average
away t h e localization information in Τ^[ε](α, 6). On t h e contrary, its mem-
bers are actively researching "inverse" fractal problems. This consists in
retrieving from the structures they see in Τ^[ε](α, 6), viewed as a field in
(a, fc)-space, t h e dynamics of the system used to generate ε(χ) in t h e first
place. To this effect they have developed the "modulus m a x i m a wavelet
transform" m e t h o d and have successfully applied it t o a variety of systems

ranging from t h e deterministic Cantor measure (e.g., Arnéodo et al., [3]
and [4]) to r a n d o m diffusion limited aggregates (e.g., Arnéodo et al., [6]
and [8]).
4 . 2 . Scale-invariant singular m e a s u r e s , g e n e r a l i t i e s
For scale-invariant measures, we will seek the exponent K(q) in
(e(r)')ocr^). (20)
We are following t h e notations and conventions suggested by Schertzer and
Lovejoy [52]. An alternative b u t more widely used approach uses sums
rather averages over r-sized sets, i.e., the focus is on p(r;x) — re(r\x) «
Γ|Τ^[ε](Γ,χ)\ in one spatial dimension. One then defines the "partition
an(
function" Zq(r) = (Ep(rlx)q) ^ where, as suggested above, t h e sum
carries over disjoint r-sized boxes rather t h a n all possible x's as in Equa-
tion (19). There are precisely [L/r] such boxes at scale r in one dimension.
For scale-invariant measures, the scaling of Zq(r) is parameterized as rT^.
A little algebra leads to r(q) = (q — 1) — K(q).
Some general results follow from the above definitions. For t h e same
reasons as for structure functions, (s(r)q) and K(q) are smooth functions
of q and, in analogy with — ζ(ρ) in Equation (12), K{q) will b e convex
(Kf,(q) > 0), although we must still require t h a t the prefactors in Equa-
tion (20) depend only weakly on q. Proper normalization of the 1-point
p.d.f.'s of t h e measures e(r;x) requires K(0) = 0 in Equation (20), just
as ζ(0) = 0 in Equation (12). There is also an essential difference with
structure functions. Consider t h e case q = 1; the two averages in Equa-
tions (15)-(18) and (19) - inside one-dimensional "boxes" and over these
boxes when using I\ (·) - commute and we find (s(r)) = constant, hence
K(l) = 0. Since K(0) = K(1) = 0, convexity implies K(q) < 0 for
0 < q < 1 and K(q) > 0 elsewhere as well as K'(0) < 0 and K'{\) > 0.
It is important to know what to expect for weakly variable e(r; x) fields,
i.e., where (e(r)q) « (e(r))q. Equation (20) then yields K(q) = qK{l) = 0,
the ( s ( r ) 9 ) ' s are constant with respect to r; also t h e = ' s will apply in all
of t h e inequalities discussed in connection with Equation (20). Conversely,
it is easy to show t h a t finite K(q) 's (q φ 0,1) imply extremely singular
measures. To see this we first adopt units of length where L = 1 and
units for ε(χ) where (e(r)) = 1, 0 < r < 1. As soon as r < 1, we have
{e(r)q) < 1 for 0 < q < 1 (due to -K(q) > 0) and (e{r)q) » 1 for
q > 1 (—K(q) < 0). Taking t h e qr-th power of a non-negative r a n d o m
variable considerably reduces its range if 0 < q < 1 so, in order to obtain
(s(r) 9 ) <^C 1, we need a p.d.f. for e(r; x) t h a t is concentrated on vanishingly
small values. This is however incompatible with (s(r)) = 1 unless the
said p.d.f. is also very skewed in t h e positive direction. We can envision
270 A. Davis et al.
broad expanses of very small ε-values interrupted by many narrow "spikes,"

defined (say) by ε > 1, with a variety of heights, the larger the rarer. This
is what we mean by an "intermittent" process, i.e., one characterized by
sudden bursts of high-frequency activity, recalling that e(x) is computed
from gradients in the first place (a high-pass filtered version of the signal).
In the above discussion it is important to bear in mind that in practical
data-analysis situations a finite K(q) does not automatically imply a high
degree of intermittency. Indeed, at finite spatial resolution - necessarily the
case with data - a weakly variable measure will always yield small values
of K(q) (for q φ 0,1), a "residual" or "spurious" multiscaling. An example
of this behavior is discussed by Marshak et al. [39].
An extreme case of intermittency arises when e(x) is a single J-function
placed at random on [0,1]; our definitions lead to K(q) = q—l for q > 0 ([17]
and [39]). However, we will generally be dealing with measures which are
not as singular - nor intermittent - as this; instead of being concentrated
onto a single point, the measure is distributed over many smaller peaks.
Here too a hierarchy of exponents can be defined, this time a non-
decreasing one:
C(q) = ^ , (21a)
recalling that K(q) is convex and that K(l) = 0. It can be directly related
to the well-known non-increasing hierarchy of "generalized" dimensions
D(q)=l-C(q), (21b)
first introduced by Grassberger [28] and Hentschel and Procaccia [31] with
dynamical systems and strange attractors in mind. Following Parisi and
Frisch [48] who where focusing explicitly on structure functions, Halsey et
al. [30] established an interpretation for D(q) in terms of variable orders
of singularity. (In this case, "singularity" is used in the sense of measures:
how does a(x) in p(r;x) = rs(r;x) oc r _ a ( x ) vary with χΊ)
Obtaining K(q) and C(q) or D(q) is the goal of this approach, providing
us with a means to characterize intermittency, in both quantitative and
qualitative terms (examples to follow).
4.3. Geometrical, spectral and dynamical implications of
statistical singularity analysis
Measures with a variable C(q) are known as "multifractals" (one should
specify stationary). This leaves measures with non-vanishing constant C(q)
- hence K{q) oc (q—l) - to be "monofractals" (again stationarity being
implicit). In essence, a random (deterministic) fractal is a sparse subset of
space that can be described statistically (exactly) with a single exponent,
its fractal dimension. Whether viewed as a stationary random measure
(present section) or as a non-stationary random function (previous section),

a multifractal calls for many exponents; each of these is dominated by
a specific kind of event and can therefore be associated with the fractal
dimension of t h e set where such events occur.
Following as closely as possible our discussion of structure function
analysis, it is of interest to seek a simple b u t meaningful way of charac-
terizing the sparseness a n d / o r intermittency in the system, using a single
exponent rather t h a n a whole family. It is n a t u r a l to consider t h e mean
(i.e., t h e case q = 1). We already know t h a t (s(r)) is unit, independently
of r (K(l) = 0), when the measure is appropriately non-dimensionalized.
We would however like to know where t h e events t h a t contribute most to
{s(r)) occur. Recall t h a t an overwhelming majority of ε-values (i.e., the
most probable ones) are <C 1; these "typical" events fill space, leaving only
a sparse (fractal) set where e(r;x) > (e(r)) = 1. If "average" events (i.e.,
e(r;x) « 1) are already a rare occurrence, they cannot contribute enough
to t h e mean (e(r)) to offset t h e effect of the numerous small values. We
must go to much higher values hence somewhat rarer events. It can be
shown t h a t the codimension of the set where these "singularities" live is
d = C(l) = 1 - D(l) = K'(l) > 0, (22a)
where -D(l), the fractal dimension of this set, is known as the "information"
dimension ([28] and [31]). This tells us t h a t we can set the largest value
for C\ at 1 for D(l) = 0, attained in particular for randomly positioned
J-functions (all the measure is concentrated into a single point). Processes
with C\ > 1 are called "degenerate:" almost every realization is empty and,
every now-and-then, one occurs with a huge peak, perturbing t h e ensemble
average each time. At the opposite limit, C\ —> 0 (weak variability), we
find information everywhere: D{1) = 1.
An infinite number of exponents is needed to describe multifractals
completely. For instance, there is no general relation between C\ and the
spectral exponent which is however connected with the q = 2 moment:
βε = 1 - K(2) = 1 - C(2) = D(2) < D(0) = 1, (22b)
a special case of the Wiener-Khintchine theorem for stationary scaling pro-

cesses (cf. Appendix). We will nevertheless retain C\ as t h e single most
important exponent in the whole C(q) hierarchy. T h e geometrical signifi-
cance given in connection with Equation (22a) is not as i m p o r t a n t as the
fact t h a t C\ defines as simply as possible t h e linear trend in C(q) or D{q).
In other words, we have a first order estimate of the departure of the ε-field
from weak variability (non-intermittency) t h a t is readily obtained from a
K(q) graph (cf. Figure 8b). In summary, we view C\ as a direct quanti-
fier of intermittency in t h e system and the complete /*T(g)-, C(q)~ or D(q)
272 A. Davis et al.
function as a means to qualify this intermittency.

Unlike the concept of non-stationarity for structure functions, there is
no need to argue the dynamical relevance of intermittency. The concept
was imported from turbulence into statistics in the first place. We now
provide some examples both theoretical and empirical.
4.4. Examples from theory, turbulence and cloud structure
In Figure 7 we show four random processes, all theoretical cascade mod-
els ε(/; x) with / = L/1024 and log-normal multiplicative weights W, as first
proposed by Kolmogorov [35] and Obukhov [47] to emulate the variability
of the dissipation field in turbulence (ε oc du2 /dt). Clearly different degrees
of intermittency are present. At the top, we find a flat field (corresponding
to degenerate weights) which is trivially stationary and non-intermittent:
C(q) = 0. At the bottom, we find a very intermittent process: C(q) = qC\
for q < 1/Ci with Cx = 1/2 In 2 « 0.72 (σ2ηΐν = 1) and divergent mo-
ments for q > 2 In 2 « 1.34 (including q = 2). Between these two extremes,
we find two intermediate examples of log-normality, illustrating more-and-
more intermittency with C\ « 0.05 and 0.18. However, these models all
have the same "kind" of intermittency or multifractality. The same form
of C(q) applies throughout, there is no change in qualitative behavior (in
particular, there is always a threshold for divergence of higher moments).
Within the framework of the statistical theory of turbulence alone,
many other cascade-type intermittency models have been and are still be-
ing proposed (and discussed at length). Novikov and Stewart's [46] model
of "pulses within pulses," Mandelbrot's [37] model of "absolute curdling"
and Frisch et α/.'s [24] "/?-model" are all characterized by the frequent oc-
currence of null multiplicative weights; they are generically described by
K(q) = Ci(q— 1) or C(q) = C\ (i.e., monoscaling) with 0 < C\ < 1, for
q > 0 and K(q) = oo for q < 0. Schertzer and Lovejoy's [51] "a-model"
and Meneveau and Sreenivasan's [42] "p-model" have log-binomial weights;
they are generically described by K(q) = log2\pWq + p'W , g ], assuming for
simplicity that the cascade proceeds by divisions into two sub-intervals in
one spatial dimension, with q € U and parameters verifying p + pf = 1
(from K(0) = 0) and pW + p'W = 1 (from K(l) = 0), hence from Equa-
tion (22a) Ci = pW \og2 W + p W l o g 2 W. Benzi et al. [10] proposed a
"random /?-model" which also exhibits multiscaling. Schertzer and Lovejoy
[52] advocate a family of "universal" cascade models with log-Levy stable
weights that are generically described by K(q) = C\{qQ — q)/(a— 1) with
0 < a < 2 , f o r g > 0 and K(q) = oo for q < 0, the limiting cases a —> 0 +
and a —y 2~ leading back to monofractality and log-normality respectively.
Certain choices of parameters in the above can lead to the divergence of
higher moments, hence to an upper bound on the range of ςτ, namely q < qD
where qD is the solution of K{q) = q - 1 for q > 1 ([33], [37], [53] and [32]).
Muitifractal Analysis of Geophysical Signals 273
max/mean=1
cô.o
CÔ.05 max/mean=10
tA^wÄ
^=0.18 max/mean=52
AJL.L .Ad/taJ ' I A W U L U M I â .^MM^^^-^JKJ^LALML.
CÔ.72 max/mean=265
F i g u r e 7 . Different amounts of the same kind of intermittency. The

construction of an ε(1; x) field using a cascade model calls for recursively
dividing subsets of [0, L] into (say, twice) smaller ones and multiplying
the current value of ε(/; x) {l/L = 2 , 4 , 8 , . . . ) by independently chosen
non-negative random numbers W with unit mean. It can be shown
that, apart from the effects of divergence of higher order moments (see
text), K(q) = l o g 2 ( W 9 ) [43]. We used log-normal deviates which lead to
K(q) = C\q(q— 1), q < 1 / C i , and varied the fundamental intermittency
parameter C\ as indicated by changing the log-variance of the W s :
C\ = σι 2 η νν/21η2 with σ = 0.0,0.25,0.5,1.0. However, by enforcing
throughout the log-normality, we obtain the same "flavor" of intermit-
tency in different amounts. Other types of scale-invariant intermittency
are discussed in the text.
274 A. Davis et al.
η=20 m
(a)
12
10 \ o q-1
X - -Ξ-
- - O-
q=2
8 q=3
N< -X- q=4
ω
V
CsJ 6
ö) I
o.
O
4 I X:
ώ.
o- .o
2 I S
- El· .
I <\Jog 2 (r/l)
0 -Θ—Θ—Θ—©—Θ—Θ—Θ—Θ—e-
8 10 12 14
K(q) (b)
1
0.8
0.6
0.4
1 -β = K(2) .
0.2
^ - Κ Ό ) _
0>
0 0.5 1 1.5 2 2.5 3 3.5 4
F i g u r e 8. Singularity analysis of cloud LWC. (a) T h e q u a n t i t i e s

(e(r)q) are plotted for l o g 2 ( r / / ) = 2 , . . . , 4 and q = 1 , 2 , 3 , 4 . We have
indicated the lower end of the scaling range η and the absolute slopes
are the exponents K(q). (b) The corresponding K(q) function, demon-
strating the multifractality of absolute gradients of the LWC field such
as in Figure l b . We have highlighted the remarkable values at q = 1,2
discussed in connection with Equations (22a,b).
The most quoted singularity analysis of the dissipation field in turbu-

lence is due to Meneveau and Sreenivasan ([41], [42]) who determined (an
exponent family closely related to) K(q) for wind tunnel and atmospheric
boundary layer flows at high Reynolds numbers; see also [54]. Empirical
values for C\ fall in the range 0.2-0.3. Argoul et al. [2] show graphic evi-
dence of the cascade process using the wavelet in Equation (17a) directly on
the velocity field. In the same spirit of "inverse fractal theory" (i.e., finding
the internal dynamics from the data), Arnéodo et al. ([3], [4]) investigate
deterministic cascade models.
Figure 8a shows log 2 [(|s(r)|^)/(|ε(£)| 9 )] versus log 2 (r//) for q = 1,2,3,4
and r > η = 41 prepared for the time-series of LWC in Figure la. Specifi-
cally, we formed absolute differences over distances of 4 grid points (η — 20
m, the lower bound of the scaling regime as defined by the structure func-
tions in Figure 5a). Four such absolute gradient fields can be obtained and
one of these is illustrated in Figure lb. The (|e(r)| 9 )'s are then computed
in the standard (J r (x)-based) way by averaging first over space, as in Equa-
tions (15)-(19) and using disjoint r-sized boxes; then an average over all
four cases is performed. A priori, we wish to characterize intermittency in
the same range of scales [77, R] as for the non-stationarity previously but we
recall that, being caused by a single event, the break in scaling observed
at R « 5 km in Figure 5a is not robust [18] and we are justified in seeking
scaling behavior in (|e(r)| 9 ) up to r = L. The absolute slopes in Figure 8a
give the corresponding value of K(q). In Figure 8b we have plotted K(q)
and highlighted two remarkable values: K'(l) = C\ « 0.1 and K{2) which
is related to the spectral exponent βε of the ε-field, as described in the
Appendix. We notice the strong evidence of intermittency in the structure
of marine StCu (K(q) is non-vanishing) and that it is multifractal in nature
(in the sense of "multi-singular" measures, K(q) is nonlinear).
§5. Summary and Bi-Multifractal Statistics

5.1. Comparative multifractal analysis from the wavelet
theoretical standpoint
In the course of our survey, we have drawn many parallels between
the two multifractal statistical data analysis techniques. There are also
fundamental differences. To better appreciate these, we compare item-per-
item definitions, results and related concepts for structure functions and
singular measures.
276 A. Davis et al.
Statistical Structure Singular

Operation : Functions Measures
input :
nature :
/oo
function
Φ)
measure (ε(χ) > 0)
condition non-stationary stationary
(E(k)~k-P)-: (ßf > 1) (A < i)
Define £(r; x) > 0 : |Δ/(Γ;Χ)| = e(r;x) =
\f(x + r)-f(x)\ H:+rs{y)dy
A(z) in (ξ(ν;χΥ) C(*) -K(z)
ex rA^ :
solution (s) of z =0 z = 0,l
A(z) = 0 :
exponent hierarchy : H(z) = ζ(ζ)/ζ C(z) =
K(z)/(z-l)
spectral property : ßf = C(2) + 1 > 1 ββ = 1- K{2) < 1
analytical properties : continuity but discontinuity
non-differentiability and singularity
geometrical property : roughness sparseness
statistical/dynamical
property : non-stationarity intermit tency
Although wavelets yield only variants of existing techniques rather

t h a n radically new approaches in multifractal analysis, one does achieve
a higher degree of unification. In section 3 we h a d "differencing" (high-
pass) wavelets acting on t h e raw data. In section 4 we have "averaging"
(low-pass) wavelets acting on t h e absolute differences derived from t h e same
d a t a (or some other measure sensitive t o t h e high frequencies). One way
or another, we are focusing on t h e singularities of t h e gradient field and
this seems t o imply t h a t t h e two kinds of multifractality we observe (say,
in clouds) are really just two facets of a deeper one, independent of how
it is characterized. In this spirit, we now describe a merger of t h e two
approaches a n d its potential benefits.
5 . 2 . O n Ç(p)-to-/f(çr) c o n n e c t i o n s
Recalling t h a t e{x) is derived from f(x) for a given geophysical process
(e.g., governing cloud structure), b o t h / ( # ) ' s structure functions a n d s(#)'s
singularity analysis can produce evidence of multifractality. It is therefore
natural t o ask "Is it fundamentally t h e same multifractality?" and, if so,
"Is there a connection between ζ(ρ) a n d K(q)T\ T h e answer t o t h e former
and more philosophical question is probably "yes." Davis et al. [17] address
t h e latter question from an empirical perspective, suggesting t h a t t h e fields
| Δ / ( Γ ; Χ)\ a n d e(r; x) b e studied jointly. This can of course b e done within
the framework of wavelets by computing
(|Δ/(Γ; x)\"e(r; x)") *±J |T*[/](r, b)\f\^[e](r, b)\" db oc r*<™\ (23)
with the appropriate ranges for p and q defined in Equations (11) and (19);
concerning r, one should of course remain in the scaling regime defined in
Equation (4). The new exponent function X(p,q) reverts to known cases
for q = 0 and p = 0 and will be concave (dppX < 0, dgqX < 0) on its
domain of definition.
Davis et al. [17] show how the X(p,q) can be used to test the scaling
hypothesis that there exists two exponents t > 0 and s > 0 such that
ε(Γ;ι)=Μ. (24)
This generalizes Kolmogorov's [34] relation for fully developed turbulence
corresponding to s = 1 and t = 3; more precisely, Kolmogorov argued
Equation (24) not at every x but on average only, hence Ç(£ = 3) = s = l
since we know that (s(r)) is independent of r. If however Equation (24)
applies everywhere, then it follows (by taking p/t-th powers and averaging)
that
C(p) = (f)p-if(f), (25)
as often quoted in the case of turbulence in spite of the fact that Equa-
tion (24) has not been verified beyond Kolmogorov's average at p = t = 3
([55]; and references therein). So it appears that at least two new pa-
rameters are needed to connect £(·) and ΛΓ(·), the essential one being
s = ζ(ί) which allows Equation (24) to relate a stationary field (ε) to a
non-stationary one (/). Indeed, for small enough C\ the values of K(p/t)
in Equation (25) will be small as well (except possibly when p ^> i); in
turbulence theory K(p/3) is called an "intermittency correction" (usually
at p = 2). This makes (s/t)p reflect, to first order, the non-stationarity of
f(x) in Equation (25) since we then get ζ(ρ) « (s/t)p hence Hi « s/t (1/3
for turbulence).
5.3. The mean multifractal plane

The above ideas are important since they can in principle be used to
establish, from data alone, relations such as Equation (24) that can be
viewed as effective constitutive laws [17]. However in many practical situa-
tions we are not interested in dealing with the large number of exponents in
ζ(ρ) and/or K(q), let alone the larger number needed to represent X(p, q).
Can we restrict ourselves to a single exponent from each multifractal ap-
proach? The natural choices are Hi and Ci with no risk of redundancy,
even if Equation (25) applies. This minimal multifractal parameteriza-
tion enables us to define the "mean multifractal plane" with coordinates
278 A. Davis et al.
(Hi, Ci) G [0,1] ® [0,1], the upper limit on Ci being imposed only to ex-
clude degenerate gradient fields. Figure 9 represents an (ifi,Ci)-plot that
we have populated with empirical findings from the turbulence literature
(approximate positioning) and some of our LWC studies (precise position-
ing).
We notice on Figure 9 that the models discussed in this paper fall pri-
marily on either axis: additive processes (Section 3.3) have Ci = 0 and
Hi > 0 (they are neither intermittent nor stationary) while multiplicative
cascades (Section 4.3) have Hi = 0 and Ci > 0 (they are both stationary
and intermittent). We have added Mandelbrot's [38] "Devil's staircases"
which are simply integrals of cascade models, obeying Equations (24)-(25)
with s = t = 1 according to Equation (15). They are found at Hi = 1 and
finite Ci (they are almost everywhere differentiable but, having fully devel-
oped cascades as gradient fields, they are intermittent). A more familiar
case is found at Hι = Ci = 1: Heaviside steps, the integrals of Dirac £'s. It
seems that these standard scale-invariant models carefully avoid the locus
of the data, inside the square domain. This clearly demonstrates the need
for a new class of stochastic models that can access the whole multifrac-
tal domain and which are likely to have both additive and multiplicative
ingredients. A list of publications where such multi-affine models are de-
scribed is provided at the end of section 3; we believe this list (of five) to
be exhaustive at the time of publication but there is little doubt that many
more will follow. We also anticipate numerous geophysical applications for
(Hi, Ci )-plots. In particular, we foresee all sorts of data-to-data and data-
to-model intercomparisons for the purposes of validating numerical models
or retrievals from remote sensing for instance ([16] and [18]).
§6. Conclusions
Wavelets have been extensively used to show in graphic detail how
structures occur on all observable scales in many geophysical signals. A
natural environment for modeling such signals is provided by stochastic
processes which have no characteristic scale (sporting, e.g., power-law en-
ergy spectra) and these are best described by multiscaling or "multifractal"
statistics. Since scale-invariance prevails in nature as well, we are urged to
blend wavelet- and multifractal analysis techniques. The major benefit of
this merger for the wavelet community is the fact that multifractal statis-
tics have strong dynamical overtones; they can be used to characterize in
quantitative and qualitative terms both intermittency and non-stationarity.
The benefit for the multifractal community is a somewhat more unified view
of their two main tools, singularity analysis (targeting the intermittency)
and structure functions (targeting the non-stationarity). This connection
is timely because many important questions are still open about the inter-
C1 = 1
• .—. . . « « ^ • • •
' ^degenerate gradient fields y / / '
/ / / / / / / / ' </'
1.0 < /. /. <. £ £ / /....J. <, /X...M
Dirac δ-functions Heaviside step functions
-» more integration -»
\
« °- 3 1\ Turbulence
Rj <- dissipation field ...
CO & velocity field -»
•Ό
C
CO
PURELY ϊ
c MULTIPLICATIVE
Φ MODELS J
ts
1
| 30/06/87 FIRE LWC dataseT
£
c
ω 0.1 PURELY
o ADDITIVE Averages for FIRE-87 and
E MODELS; ASTEX-92 LWC data bases
2-D„
o.o — r 0.2 0.3 0.4 0.5 1.0
0.0 0.1
white noise fractional standard almost everywhere
to Brownian Brownian differentiate
Ί/Ρ noise motion motion functions
<0*fte1) (β=5/3) (ß=2) (ß*3)
less stationarity, more smoothness =§>
F i g u r e 9 . The "Mean Multifractal Plane" or ( # i , C i ) - p l o t . See text

for details on the various models and the point for LWC in marine
StCu (FIRE, 3 0 / 0 6 / 8 7 ) . Also indicated are points mapped to the av-
erages over whole LWC databases captured during FIRE in 1987 off
the southern Californien coast and the Atlantic Stratocumulus Transi-
tion Experiment (ASTEX) in 1992 near the Açores; see [18] and [16]
respectively for details. The proximity of these two points argues for
a degree of universality in the nonlinear dynamics that determines the
internal distribution of liquid water in marine StCu under rather dif-
ferent climatological conditions. The distance between these points and
the one obtained for the 3 0 / 0 6 flight during FIRE illustrates a dramatic
violation of ergodicity in this rather typical atmospheric process, hence
the need for vast databases to define the behavior of this and other
atmospheric fields in statistically robust quantitative terms.
280 A. Davis et al.
action of intermittent cascade-type fields and directly observable ones; for

instance, how does the intermittency of the dissipation field affect locally
the velocity field in turbulence?
The main ideas of multifractal analysis are presented from a wavelet
standpoint and amply illustrated with theoretical models and empirical
results drawn from the turbulence literature. New analyses of the liq-
uid water density field are also presented in sufficient detail to serve as
a tutorial. We describe how to systematically seek and how to interpret
connections between structure functions and singular measures for a given
type of data. Arguing that these connections are non-trivial to say the
least, we define the "mean multifractal plane" by judiciously selecting a
single exponent from each multiscaling approach, H\ for structure func-
tions and C\ for singular measures. Finally, we use the (i?i,Ci)-plot to
discuss several theoretical and empirical findings. This simple diagnostic
device will be helpful in many outstanding geophysical problems of find-
ing statistically robust and physically meaningful ways of intercomparing
data of different but related origins, numerical model output included. It
is used here to show the need for a newly introduced class of stochastic
models known in the physics literature as "multi-affine," to demonstrate
a violation of ergodicity in the distribution of liquid water in a particular
marine stratocumulus cloud deck, and to argue for a degree of universality
in the dynamics that determine these distributions.
A. Stationarity, Intermittency, Integral Scales and Stochastic

Continuity - A Tutorial in the Framework of
Scale-Invariance and Second Order Statistics
A . l . Stationarity versus ergodicity
A property of fundamental importance in the theory of stochastic pro-
cesses, especially from the standpoint of data analysis, is statistical "sta-
tionarity." In essence, a stationary quantity is statistically invariant under
translation. In principle one should talk about statistical "homogeneity" in
the spatial domain but we prefer the term "stationarity" borrowed from the
time domain; in this way we avoid any confusion with the idea of constant
fields.
When analyzing data, one generally uses the spatial coordinate to per-
form various averaging operations, as described in the main text. This
amounts to making an "ergodicity" assumption (i.e., that spatial averages
will converge towards their ensemble counterparts as the amount of data
increases). Stationarity is a far weaker assumption than ergodicity, but eas-
ier to work with on theoretical grounds and more reasonable in empirical
situations. For example, Holley and Waymire [32] and Gupta and Waymire
[29] show that multiplicative cascade models (which we will use to illustrate
stationarity further on) are generally non-ergodic. The highly correlated

fluctuations that characterize these models cause the (spatial) law of large
numbers to fail. Schertzer and Lovejoy [53] pursue these issues in more
quantitative terms, dependent on what portion of probability space (the
set of all possible realizations) is sampled.
Neither ergodicity nor stationarity can be established rigorously for
a dataset, no matter how long. We must rely on consistency checks and
we show here how scale-invariance eases this task. The central role of the
"integral scale" is also underscored. Throughout, we illustrate the con-
cepts with additive (non-stationary and non-intermittent) and multiplica-
tive (stationary and intermittent) scale-invariant models as well as a well-
known non-scaling example (Ornstein-Uhlenbeck processes). Finally, we
show that there exists a strong connection between stochastic continuity
and non-stationarity for scaling processes.
A.2. Stationarity and scale-invariance in physical space, the

integral scale
Let f(x) be a real random process defined either continuously (for a
theoretical model) or discretely (for data) on the interval [0,L]. A first
consequence of stationarity concerns 1-point statistics, e.g., for the mean
we find (f(x + r)) = (/(#)), 0 < r < L , 0 < # < L — r. A 2-point statistic
of obvious interest in time series analysis is {[f(x + r) — (f(x + r))][f(x) —
(f(x))]}', under stationary conditions, it reduces to the auto-correlation
function
{f(x + r)f(x)) = G(r), (A.l)
using a linear transformation that reduces (f(x)) to zero. The identity in
Equation (A.l) follows from stationarity and we notice that G(0) = (/(#) 2 ),
1-point variance. A popular application of G(r) is the estimation of the
"integral" correlation length:
(Α
R
= W)fG{r)dr· ·2)
To a first approximation, this length scale draws the line between correlated
(r < R) and uncorrelated (r > R) values of f(x) and f(x + r), a sort of
statistical period in the data.
For scale-invariant processes (i.e., with power-law 2-point statistics),
we can anticipate
G(r) oc Γ~μ, μ>0 (Α.3)
in stationary situations. The exponent μ must be positive since we gen-
erally expect less correlation as r increases. Equation (A.3) cannot be
substituted directly into Equation (A.2). Firstly, there is always an outer
limit to the range of scales where Equation (A.3) can apply; for a process
282 A. Davis et al.
defined on [0,L], we must require G(r) = 0 for r > L. There is also an in-
ner scale in most situations of practical interest; for 0 < r < /, we can take
f(x + r) « / ( # ) hence G(r) « G(0). Between these limits (with / <^ £ ) , we
assume Equation (A.3) to apply; more precisely, we take G(r) « G(0)[l/ry.
It follows t h a t
y « (l) m i n [M. (A.4)
So, for all practical purposes, the exponent μ in Equation (A.3) cannot ex-
ceed unity without leading to fields t h a t are uncorrelated from one pixel to
the next. If μ -> 0 (hence R& L), there is no trend towards decorrelation;
we view this as a symptom of non-stationarity and in Section A.5 will intro-
duce a 2-point statistic better adapted to this situation t h a n ( / ( # 4 - r ) / ( # ) ) .
A . 3 . A Fourier s p a c e criterion for s t a t i o n ä r i t y

T h e Wiener-Khintchine theorem states t h a t , under necessarily station-
ary conditions, G(r) in Equation (A.l) and E(k) in Equation (3) form a
Fourier transform pair. So, in particular, the integral scale R can be more
efficiently computed from
/»OO
R = E(0)/ E(k)dk, (A.5)

Jo
using fast Fourier transforms.
In their scaling versions Equations (A.l) and Equation (5), the Wiener-
Khintchine theorem translates to
β + μ = 1, (Α.6)
which implies, in particular (μ > 0), t h a t
β < 1. (A.7)
This criterion for stationarity is readily applicable to any kind of d a t a t h a t

comes on a grid and exhibits scaling behavior. Although not a proof of
stationarity for data, Equation (A.7) should be verified before computing
(f(x + r)f(x))-or any other statistic requiring stationarity - via spatial
averaging.
T h e simplest possible stochastic process is a sequence of independent
random numbers; viewed as a field / o ( # ) , x > 0, it is uncorrelated - or
rather "δ-correlated" - and stationary by construction: Go(r) oc S(r). T h e
subscript "0" stands for vanishing R in Equation (A.2) or (A.5), due to the
infinite denominator. Figure A l a provides a sample of such white (β = 0)
noise t h a t we will denote fo(x) for 0 < x < L = 4096 using zero-mean
unit-variance Gaussian deviates. We now describe a more interesting case
of scale-invariant stationarity.
K(x)
ff(x)
F i g u r e A l . Scale-Invariant Stationary and Non-Stationary Noises,

Along with an Intermediate Non-Scaling Case, (a) Unit-variance zero-
mean Gaussian white noise fo(x), 0 < x < L = 4096, which is scaling,
stationary (ß = 0 < 1) and discontinuous, (b) Brownian motion foo(x)
- essentially the running integral of (a) - is scaling, continuous and non-
stationary (ß = 2 > 1). (c) An Ornstein-Uhlenbeck process obeying
a simple stochastic ODE, Rf'n + /R — /o with an integral scale pa-
rameter R = L / 8 . Its spectrum is E(k) = £ ( 0 ) / [ l + (kR)2]. Being
non-scaling, this model can be at once stationary ( ( / R ( X + r)fn(x)) =
G(r) = G(0)exp[-r/R]) a n d continuous ({[fR(x + r) - / Ä ( X ) ] 2 ) =
2[G(0) — G(r)] -¥ 0 when r - > 0 ) . However, the scaling cases in panels
(a) and (b) are retrieved in the limits R —> 0 and R —> oo respectively,
hence the subscripts.
284 A. Davis et al.
step no. Ί (3 Ί =3 points)
< R^-
F i g u r e A 2 . Development of a Log-Normal Multiplicative Cascade.

The inset shows the growth of the "singularities" (spikes) through the
1-st, 2-nd and 4-th cascade steps. The 6-th step is illustrated in more
detail: ε(χ) is represented for 0 < x < L = 1 with a grid constant
/ = 3 ~ 6 = 1/729. Because it uses unbounded multiplicative weights,
this type of cascade process is patently non-ergodic ([32], [53]). How-
ever, the spatial average of this particular realization is approximately
equal to the ensemble average (namely, unity). Below the horizontal
axis, we have indicated the integral length scale, R « L / 5 in this case
(see text for details).
A.4. Multiplicative cascade models, stationarity and

intermittency
As a further illustration of scale-invariant stationarity, consider a mul-
tiplicative cascade process, traditionally denoted ε(#), 0 < x < L. These
are constructed by initially setting ε(χ) = EL then subdividing [0, L] into
(say) ra = 2 equal parts, [G, L/2] and [L/2, L], and multiplying e{x) in each
one by W\ and W[, both unit-mean non-negative random variables drawn
from the same distribution. This defines ε(/; x) for / = L/2. The procedure
is repeated ad infinitum:
n
Φ x) = eL fi W{, l/L = τη~η -+ 0. (A.8)
1
In Figure A2 we used m = 3, log-normal Ws with log-standard devi-

ation σιη \y = 0.5 and log-mean — σ2η w/2 in order to ensure proper nor-
malization (i.e., (W) = 1); cascade steps n = 1,2,4 (inset) and 6 are
illustrated. The statistical properties of
oo
ε(χ) = lim ε(Ζ; x) = [ J W{ (A.9)

""* î
have been investigated primarily in the turbulence literature ([35], [47] and
[37]), as a model for the intermittency of the dissipation field in very high
Reynolds number flows. These processes, known as "multifractals" [48],
are now widely applied in deterministic chaos [30].
Stationarity in the above sense [Equations (A.3)-(A.7)] follows from
[43]
(ε(χ + r)e(x)) ~ r~K^, K(2) = logm<W?> (A.10)
using notations consistent with those of section 4. Comparing with Equa-

tion (A.3), we have
με = K(2) > 0, (A.ll)
which is known as the "intermittency" parameter. The inequality in Equa-

tion (A.ll) follows directly from Schwartz's: (Wf) > (Wi)2 = 1. This
establishes stationarity (βε = 1 — με < 1) as soon as the variance of the
weights is finite, (W?) — (Wi)2 > 0. In this case "=" corresponds to degen-
erate multiplicative weights (Wi = 1) hence flat fields which are trivially
stationary. In Figure A2 we take m = 3 and the numerical values lead
to με « 0.23 and from Equation A.4 we find R/L « 0.22 for L/l = 729
(n = 6).
286 A. Davis et al.
Α.5· Non-stationary scale-invariant situations, up to the integral

scale
In non-stationary cases where
/?>1, (A.12)
the field f(x) can fluctuate wildly, given enough space, since there is rela-
tively more energy (variance) in the large scales. Being dependent on x as
well as r, the quantity in the l.h.s. of Equation (A.l) is no longer a simple
statistic, dependent only on the lag r. One can however work with incre-
ments of f(x): Δ / ( Γ ; X) = f(x + r) - f(x) with 0 < r < L, 0<x <L-r,
which will be less sensitive to the local value of /(·). The (2-nd order
2-point) statistical counterpart of Equation (A.l) is
<Δ/(Γ;202) = (Δ/(Γ)2) (Α.13)
where the identity assumes "stationary increments." The statistic in Equa-
tion (A. 13) is known in turbulence studies as a "structure function" [43]
and in geostatistics as a "variogram" [13].
In scale-invariant situations, we expect the 2-nd order structure func-
tion in Equation (A. 13) to obey a power-law:
( Δ / ( Γ ) 2 ) OC r«2\ 0 < C(2) < 2, (A.14)
where the exponent is denoted ζ(2) for consistency with section 3 in the
main text. We generally expect ( Δ / ( Γ ) 2 ) to be a non-decreasing function
of r, hence the sign of the exponent. The limit ζ(2) -> 0 (increments
independent of the scale involved) leads back to a stationary statistical
regime. The opposite limit ζ(2) —» 2 can be associated with the class of
(almost surely everywhere) differentiable functions since \f(x + r) — f(x)\ oc
r will not be a rare occurrence in this case.
It is easy to see why there is always an upper bound to the non-
stationary range for the scale parameter r in physical processes. If r —> oo
in Equation (A.14) then Δ / ( Γ ; Χ ) , hence f(x) itself, must also take arbi-
trarily large values whereas most geophysical fields have natural bounds.
Some are necessarily non-negative (e.g., albedo, rain rate, density of an ad-
mixture in turbulence); others are limited by in absolute value, often by the
input mechanism of some conserved quantity (e.g., albedo by the incident
radiant energy, velocity by the kinetic energy forced into the system). This
underscores the physical importance of the integral scale R; for chunks of
data of size r > R, we will find fluctuations of comparable magnitude and,
equivalently, a flatter spectrum for wavenumbers 0 < k < 1/R; from both
stand points, we are dealing with more stationary behavior.
The Wiener-Khintchine theorem can be generalized to non-stationary
situations where stationary increments prevail, leading to a Fourier con-
Muitifractal Analysis of Geophysical Signals 287
nection between ( Δ / ( Γ ) 2 ) and E(k) [43]. For their scaling versions, Equa-
tions (A. 14) and (5), this yields
β-ζ(2) = 1, (Α.15)
in analogy with Equations (A.6) and (A. 11). Notice that the spectral crite-
rion for non-stationarity in Equation (A. 12) is retrieved from (A14-A15).
We must furthermore require β < 3 in order to have stationary incre-
ments by requiring the gradient field (E\r/(k) = k2E/(k)) to be stationary
(/?-2<l).
A.6. Additive scale-invariant models, non-intermittent and
non-stationary
The classic example of a non-stationary stochastic process is Brownian
motion (often referred to as a Wiener-Lévy process):
foo(x) = /oo(0) + Γ /o(s') dx\ (A.16)

Jo
where we notice the additive nature of the resulting field, in contrast with
the multiplicative character of ε(/; χ) in Equation (A.8). Brownian motion
is not intermittent in the sense of Section A.4, even if the intermittency
is sought in the absolute gradient field |V/oo(#)| = |/o(#)|, as custom-
ary in turbulence studies (cf. section 4). If not completely flat, |/o(#)| is
very weakly variable in comparison with ε(1; χ) due to its close relation to
Gaussian white noise. Other examples of non-intermittent non-stationary
processes are discussed in the main text ("fractional" Brownian motions).
An important new class of non-stationary models with intermittent gradi-
ent fields ( "multi-affine" processes) is also evoked.
In Equation (A.16), we are essentially computing the position of a
particle performing a random walk as a function of time x > 0, the white
noise Jo(x) representing the independently distributed random steps. We
have (foo(x)) = {fo)x + /<x>(0) where both r.h.s. terms can be made to
vanish, leaving (/oo(#)) = 0? but this has no bearing on stationarity. To
wit, it can be shown that, still for (/ 0 ) = /oo(0) = 0, the 1-point p.d.f.
of /oo(#) is Gaussian with variance (foo(x)2) which is known to increase
linearly with x > 0. This consequence of the law of large numbers is already
proof of non-stationarity. In Fourier space, we come to the same conclusion
using Equation (A. 12) since the integral in Equation (A.16) corresponds
to a 1/k filtering which brings the vanishing spectral exponent of fo(x) to
β = 2 for /oo(#)· (This justifies the subscript "oo" standing for a formally
divergent R in Equation (A.5) due to the infinite numerator, symptom of
an incipient "infra-red catastrophe.") It is easy to see that Afoo(r;y) is
simply the integral of fo(x) over [y, y + r] and behaves statistically like
/oo(x) in Equation (A.16) with /oo(0) = 0 and x = r, independently of
288 A. Davis et al.
y. So (Δ/οο(^) 2 ) oc r, which proves t h a t ζ(2) = 1 in Equation (A. 14) for

this special case and, in turn, this confirms Equation (A. 15) for Brownian
motion.
A typical sample of Brownian motion is provided in Figure A l b . Like
e{l\x) in turbulent cascades Equation (A.8), there is no limit t o how big
Brownian motion can become, starting at /oo(0) = 0, b u t it takes a long
time t o get there and, when |/oo(#)| is large, it will of course take an equally
long time for it to come back t o zero, in all probability. As can be seen in
Figure A l b , t h e "zero-crossing" set {x > 0, foo(%) = 0} where foo(x) is at
its most probable value is in fact a very sparse set with a fractal dimension
of 1/2 [38]. This is a direct consequence of /oo(#)'s lack of stationarity. In
sharp contrast, the stationary cascade models ε(/; χ) will return to their
most probable value (which is very small when / < L ) very quickly after
a strong (necessarily positive) fluctuation. T h e typical delay will be « R
which is <C L for well-developed multiplicative cascades b u t « L (formally
oo, hence t h e choice of subscript) in Brownian motion.
A . 7 . S t o c h a s t i c c o n t i n u i t y w i t h s t a t i o n a r i t y or scale-invariance
but not both
If ( Δ / ( Γ ) 2 ) -+ 0 when r -» 0, then / ( · ) is said to be "stochastically"
continuous and this is t h e case for all non-stationary scaling processes obey-
ing Equation (A. 14). Of course processes t h a t are stationary per se have
stationary increments (if scaling prevails, then β < 1 implies β < 3). It is
easy to see t h a t
<A/(r)2) = 2 [ G ( 0 ) - G ( r ) ] . (A.17)
So stationary processes are stochastically continuous as long as G(r) is

continuous at r = 0; in particular, this implies G(0) < oo (finite 1-point
variance). This is not t h e case for exactly scaling processes since G(r) —> oo
if r —y 0 in Equation (3), one of t h e numerous reasons why there is always
a lower limit t o scaling regimes in physical systems. For smaller values of
the scale parameter, fields can be viewed as smoothly varying (an incipient
"ultra-violet catastrophe" is thus avoided).
T h e log-normal multiplicative cascade process in Figure A2 obeys Equa-
tion (A.2) and therefore illustrates stochastic discontinuity, as does t h e
white noise in Figure A l a (independent Gaussian deviates assigned to ev-
ery pixel) for which G(r) oc S(r). T h e Brownian motion in Figure A l b is
stochastically continuous since ( Δ / ( Γ ) 2 ) OC r. Finally, in Figure A l e we il-
lustrate an Ornstein-Uhlenbeck process denoted / # ( # ) , 0 < x < L = 4096,
which is at once continuous and stationary b u t non-scaling. Apart from t h e
variance σ 2 = GR(Q) of its 1-point Gaussian p.d.f., t h e only parameter of
an Ornstein-Uhlenbeck process is t h e integral scale R in its auto-correlation
function
GR(r) = GR(0) exp(-r/R). (A.18)
In Figure Ale, we took R = L/8 = 512. For the same process the structure
function is therefore
(Δ/Λ(Γ)2) = 2G ß (0)[l - exp(-r/i?)]. (A.19)
Notice the cross-over from non-stationary and continuous behavior (i.e.,
2
( Δ / Λ ( Γ ) 2 ) OC r) to stationary and discontinuous behavior (i.e., (A/#(r) ) «
constant) as r goes from <C R to ^> R. In essence, we are dealing with
Brownian motion (on scales r <& R) that is forced to come back to the
origin on a regular basis (at intervals of length « R). In the limit R —> 0
(vanishing correlation length) white noise is retrieved. For R —> oo (diverg-
ing correlation length) "unconstrained" Brownian motion is retrieved and
the fact that (//*(#)/#(#+ r)) —> constant > 0 in this case is characteristic
of non-stationary behavior, as anticipated in Section A.2.
A.8. Summary with analysis and modeling of geophysical data

in mind
Geophysical processes are driven by complex nonlinear dynamics that
typically unfold in both space and time. Can we view these as statistically
periodic above some scale and treat portions of equal or greater size as
statistically independent realizations? Or must we view the whole as one
single irreproducible experiment and hope that space/time averages con-
verge to a statistically meaningful value? In statistical jargon the former
hypothesis is called stationarity (statistical invariance under translation),
the latter as ergodicity (space/time and ensemble averages yield identical
answers). In all cases, we are required to characterize these processes from
data that always comes in finite quantities with some finite resolution and
neither stationarity nor ergodicity can be established rigorously. We ar-
gue that stationarity, the more general concept, is also more useful in data
analysis endeavors; it is also the more physically reasonable when it comes
to modeling the data.
In the framework of scale-invariance, even in the widespread physical
sense where a finite range of scales is involved, one can use a specific spec-
tral criterion for stationarity: the spectral exponent (β) is less than unity.
In numerous situations there exists a range of scales where non-stationary
behavior (β > 1) is observed; one can then study a stationary feature of
the data such as increments (as long as β < 3). Even if a strict ergodicity
assumption is physically unreasonable, we must still use spatial averag-
ing to some degree and it is important to focus exclusively on stationary
quantities. So an operational procedure for defining stationary conditions
and/or features is welcome.
290 A. Davis et al.
A given geophysical field can be stationary at large scales and non-

stationary at small ones. In fact this is to be expected on physical grounds
and the transition scale is known as the integral scale. Bearing this in
mind, there are two classes of stationary theoretical models that we find
particularly useful for tutorial purposes: Ornstein-Uhlenbeck processes and
multiplicative cascade processes.
• The former are patently non-scaling but have two distinct scaling
regimes: at large scales it behaves like (stationary) white noise and
at small ones like (non-stationary) Brownian motion, roughly mim-
icking natural processes. These processes are non-differentiable but
stochastically (i.e., almost everywhere) continuous and this is di-
rectly traceable to their non-stationary behavior in the small scale
limit.
• The latter are currently being used in turbulence studies (to model
the intermittency of the dissipation field), in hydrology (to model
the spatial distribution of rain), and in deterministic chaos (to model
the invariant measure supported by the strange attractor) amongst
many other applications. They are characterized by small scale
discontinuity (almost everywhere), itself traceable to very singular
behavior on sparse fractal sets.
The concept of intermittency - sudden and intense bursts of high frequency
activity - apparently conflicts with the conventional wisdom about station-
arity: "Things are almost the same at all times." We argue that this is an
unnecessarily restrictive outlook, excluding in particular strong deviations
from mean or most probable states. A better description of stationarity
that is compatible with extreme forms of variability (represented in par-
ticular by cascade models) is: "On short notice, things can quickly depart
from and return to typical values." By "short notice" and "quickly" we
understand one integral scale or so; to illustrate this point we show that
many integral scales fit into the overall length of the process in the case
of multiplicative cascades (although each one is still many pixels long, at
least for a large but finite number of cascade steps).
LIST OF SYMBOLS
Scale parameters:
• general
r temporal or spatial scale on which a statistic is to be
conditioned
k « 1/r, wavenumber (in absolute value)
• instrumental and/or computational
/ inner scale, sampling scale (data), grid constant (model)

L outer scale, overall physical size of dataset or model
• physical (to be viewed as statistical properties)
η lower bound of non-stationary scaling regime (where β > 1)
R upper bound of non-stationary scaling regime, integral scale
Stochastic processes or geophysical data-streams and related random
quantities:
• for general purpose
x time or space coordinate

B(x) fractional Brownian motion
f(x) a generic random function
fo(x) uncorrelated Gaussian fluctuations (white noise)
foo(x) Levy-Wiener process (Brownian motion), integral of fo(x)
ÎR(X) Ornstein-Uhlenbeck process with integral scale R
ε(χ) a generic random measure (ε(χ) > 0)
ε(/; x) cascade process on [0, L] (developed down to scale /)
• for structure functions
Δ / ( Γ ; X) increment of f(x) over scale r at x, e.g., f(x + r) — f(x)
h(x) local Holder exponent (in | Δ / ( Γ ; Χ ) | OC rh^)
• for singularity analysis
ε(η; χ) = | Δ / ( ^ ; χ)\, absolute small scale gradients used in
singularity analysis
p(r; x) integral of ε(χ) or ε(η; χ) over an r-sized box at x,
e.g., [x,x + r ] i n D = l
a(x) local order of singularity (in p(r; x) oc rQ^)
e(r; x) average of ε(χ) or ε(η; χ) over an r-sized box at x,
i.e., p(r;x)/rD
• for either
£(r;x) = \&f(r;x)\,-e(r;x) or p(r;x)
((£(r;x)z) scales like rA^)
Statistical properties:
E{k) energy (or power) spectrum, spectral density
G(r) = ( / ( r + x)f(x)), auto-correlation function (of a
stationary zero-mean process)
( | Δ / ( Γ ; Χ)\Ρ) p-th moment of stationary increments | Δ / ( Γ ; Χ)|,
structure function of order p
292 A. Davis et al.
(e(r; x)q) q-th moment of stationary measures e(r; x), used in

σ standard deviation of a Gaussian r a n d o m variable
Wavelet-related quantities:
• scaling-type functions
lr{x) = θ(χ + r / 2 ) — θ(χ — r / 2 ) , indicator function
of the interval ( - r / 2 , + r / 2 )
φσ(χ) = βχρ(—χ2/2σ2)/σ\/2π,
Gaussian function with variance σ2
ΨΗ{Χ) = max[0,min{l — x, 1 -f- x}], integral of Haar wavelet
• wavelet shapes
Φσ(*) = -Vi/20O
Φ#(#) Haar wavelet, see Equation (9b)
Φ Μ (Χ) Mexican hat, see Equation (17a)
Φ^(χ) French top hat, see Equation (17b)
-oil(a?) = -I[ (x) = S(x - r / 2 ) - i(a? + r / 2 ) ,
"poor man's" wavelet [45]
• wavelet transforms of random processes
ϊ φ [ / ] ( α ; b) for structure functions of / ( · ) : Φ = -dh, Φ#, ^ G ;
a = r and 6 = x
Τφ[ε] (α; 6) for singularity analysis of ε(·) :
0 = h, Ψ\, ΨΗ, ΨΜ, VF\ α = r and b = x
Exponents
D dimensionality of d a t a (e.g., number of points « (L/l)D)
• for second order statistics

β spectral exponent for D > l(E(k) ~ 1/k^)
μ = 1 — β > 0, scaling of autö-correlation function in
D = 1 (G(r) ~ 1/r")
• for structure functions (D = 1)
p order of statistical moment of | Δ / ( Γ ; # ) | in structure
function analysis
ζ(ρ) scaling of structure functions ( ( | Δ / ( Γ ; X)\P) ~ r ^ p ) )
H(p) = ζ(ρ)/ρ, a non-increasing hierarchy
Hi = JET(l) = C(l), index of non-stationarity (0 < Hi < 1)
Dg(f) = 2 — H\, fractal dimension of two-dimensional graph
of d a t a
• for singular measures (D > 1)
q order of statistical moment of s(r; x) in

K(q) scaling of singular measures ({e(r; x)q) ~ r~K^)
Γ 9
r(q) = (q — 1)D — i f (<?), scaling of ΣΡ( 1 x ) (using disjoint
boxes) ~ r r ^
C(g) = K(q)/(q — 1), a non-decreasing hierarchy
D(q) = r(q)/(q — 1) = D — C{q), a non-decreasing hierarchy
d = C ( l ) = ^ ' ( 1 ) , index of intermittency (0 < Cx < D)
• for bi-multifractal analysis (Section 5.2)
X ( p , q) scaling of joint moments ((\Af (r; x)\pe(r; x)q) ~ r x ^ )

s, t generalized Kolmogorov exponents in Equations (24)-(25)
Miscellaneous computational quantities:
Xi discrete time or space (a?,· = «7, i = 0 , . . . , N)

fi /-value associated with X{
m dividing ratio in cascade process
n number of cascade steps
N = L/l, number of g r i d / d a t a points in D = 1 (also m " in cascade
models
Mathematical functions:
θ(χ) = 0 for x < 0, = 1/2 for x = 0, = 1 for x > 0, Heaviside's

step function
S(x) = #'(·), Dirac's generalized function
References
1. Anseimet, F., Y. Gagne, E. J. Hopfinger, and R. A. Antonia, High-
order Velocity Structure Functions in Turbulent Shear Flows, Journal
of Fluid Mechanics, 140, 63-80, 1984.
2. Argoul, F., A. Arnéodo, G. Grasseau, Y. Gagne, E. J. Hopfinger, and
U. Frisch, Wavelet analysis of turbulence reveals the multifractal n a t u r e
of the Richardson cascade, Nature, 338, 51-53, 1989.
3. Arnéodo, A., G. Grasseau, and M. Holschneider, Wavelet transform of
multifractals, Physical Review Letters, 6 1 , 2281-2284, 1988.
4. Arnéodo, A., G. Grasseau, and M. Holschneider, Wavelet transform
analysis of some dynamical systems, in Wavelets, J. M. Combes, A.
Grossmann and P. Tchamitchian, (eds.), Springer, Berlin, 1989.
5. Arnéodo, A., E. Bacry, and J. F. Muzy, Wavelet analysis of fractal
signals, Direct determination of the singularity spectrum of fully de-
294 A. Davis et al.
veloped turbulence data, in Wavelets and Turbulence, Springer, Berlin,

1991.
6. Arnéodo, A., A. Argoul, E. Bacry, J. F. Muzy, and M. Tabard, Golden
mean arithmetic in the fractal branching of diffusion-limited aggre-
gates, Physical Review Letters, 68, 3456-3459, 1991.
7. Arnéodo, A., J. F. Muzy, and E. Bacry, Wavelet analysis of fractal
signals. Applications to fully developed turbulence data, in IUTAM
Symposium on Eddy Structure Identification in Free Turbulent Shear
Flows, Poitiers, October, 1992.
8. Arnéodo, A., A. Argoul, J. F. Muzy, and M. Tabard, Uncovering Fi-
bonacci sequences in the fractal morphology of diffusion-limited aggre-
gates, Physical Letters A, 171, 31-36, 1992.
9. Bacry, E., J. F. Muzy, and A. Arnéodo, Singularity spectrum of frac-
tal signals from wavelet analysis: Exact results, Journal of Statistical
Physics, 70, 635-674, 1993.
10. Benzi, R., G. Paladin, G. Parisi, and A. Vulpani, On the multifractal
nature of fully developed turbulence and chaotic systems, Journal of
Physics A, 17, 3521-3531, 1984.
11. Benzi, R., L. Biferale, A. Crisanti, G. Paladin, M. Vergassola, and A.
Vulpani, A random process for the construction of multiaffine fields,
Physica D, 65, 352-358, 1993.
12. Cahalan, R. F., M. Nestler, W. Ridgway, W. J. Wiscombe, and T.
Bell, Marine stratocumulus spatial structure, in J[th Int. Meeting on
Statistical Climatology, Rotrura, New Zealand, 1989.
13. Chistakos, G., Random Fields in Earth Sciences, Academic Press, San
Diego, 1992.
14. Chui, C. K., An Introduction to Wavelets, Academic Press, San Diego,
1991.
15. Daubechies, I., Orthonormal bases of compactly supported wavelets,
Communications in Pure and Applied Mathematics, 41, 909-996, 1988.
16. Davis, A., A. Marshak, W. Wiscombe. and R. Cahalan, Multifractal
characterizations of non-stationarity and intermittency in geophysical
fields, Observed, retrieved or simulated, Journal of Geophysical Re-
search, in print, 1994.
17. Davis, A., A. Marshak, and W. Wiscombe, Bi-Multifractal analysis
and multi-affine modeling of non-stationary geophysical processes, Ap-
plication to turbulence and clouds, Fractals, 3(1), 1994.
18. Davis, A., A. Marshak, W. Wiscombe. and R. Cahalan, The scale-
invariant structure of marine stratocumulus deduced from observed
liquid water distributions, II, Multifractal properties and model vali-
dation, Journal of the Atmospheric Sciences, submitted, 1994.
19. Duroure, C , and C. P. Fantodji, Wavelet analysis and Fourier anal-
ysis of stratocumulus microphysical data, report (Laboratoire de
Météorologie Physique de Clermond-Ferrand), 1993.

20. Fantodji, C. P., C. Duroure, and H. R. Larsen, The use of wavelets for
the analysis of geophysical data, preprint, Physica, submitted, 1993.
Annual Review of Fluid Mechanics, 24, 395-457, 1992.
22. Feller, W., An Introduction to Probability Theory and its Applications,
vol. 2, Wiley, New York, 1971.
23. Frisch, U., From global scaling, à la Kolmogorov, to local multifractal in
fully developed turbulence, Proceedings of the Royal Society of London
A, 434, 89-99, 1991.
24. Frisch, U., P. L. Sulem, and M. Nelkin, A simple dynamical model for
intermittent fully developed turbulence, Journal of Fluid Mechanics,
87, 719-736, 1978.
25. Ghez, J. M., and S. Vaienti, On the wavelet analysis for multifractal
sets, Journal of Statistical Physics, 57, 415-420, 1989.
26. Ghez, J. M., and S. Vaienti, Integrated wavelets on fractal sets: I, The
correlation dimension, II, The generalized dimensions, Nonlinearity, 5,
777-804, 1992.
27. Gollmer, S. M., Harshvardan, R. F. Cahalan, and J. B. Snider, Wavelet
analysis of marine stratocumulus, in Proceedings of the 11th Interna-
tional Conference on Clouds and Precipitation, Montreal (Que.), Aug.
17-22, 1992.
28. Grassberger, P., Generalized dimensions of strange attractors, Physical
Review Letters A, 97, 227-330, 1983.
29. Gupta, V. K., and E. C. Waymire, A statistical analysis of mesoscale
rainfall as a random cascade, Journal of Applied Meteorology, 32, 251-
267, 1993.
30. Halsey, T. C., M. H. Jensen, L. P. Kadanoff, I. Procaccia, and B. I.
Shraiman, Fractal measures and their singularities: The characteriza-
tion of strange sets, Physical Review A, 33, 1141-1151, 1986.
31. Hentschel, H. G. E., and I. Procaccia, The infinite number of gen-
eralized dimensions of fractals and strange attractors, Physica D, 8,
435-444, 1983.
32. Holley, R., and E. C. Waymire, Multifractal dimensions and scaling ex-
ponents for strongly bounded random cascade, The Annals of Applied
Probability, 2, 819-845, 1992.
33. Kahane, J. P., and J. Peyrière, Sur certaines martingales de Benoit
Mandelbrot, Advances in Mathematics, 22, 131-145, 1976.
34. Kolmogorov, A. N., Local structure of turbulence in an incompressible
liquid for very large Reynolds numbers, Doklady Akademii Nauk SSSR,
30, 299-303, 1941.
35. Kolmogorov, A. N., A refinement of previous hypothesis concerning
the local structure of turbulence in viscous incompressible fluid at high
296 A. Davis et al.
Reynolds number, Journal of Fluid Mechanics, 13, 82-85, 1962.

36. Lavallée, D., S. Lovejoy, D. Schertzer, and P. Ladoy, Non-linear vari-
ability, multifractal analysis and simulation of landscape topography,
in Fractals in Geophysics, De Cola, L., and N. L a m (eds.), Klüver, 158-
192, 1993. of landscape topography: Multifractal analysis and simula-
tion, in De Cola, L., and Lam, N. (eds.), Fractal in Geography, Prentice
Hall, 1993.
37. Mandelbrot, B. B., Intermittent turbulence in self-similar cascades:
divergence of high moments and dimension of t h e carrier, Journal of
Fluid Mechanic, 62, 331-358, 1974.
38. Mandelbrot, B. B., Fractals: Form, Chance, and Dimension, Freeman,
W.H., and Co., San Francisco, 1977.
39. Marshak A., A. Davis, R. Cahalan, and W. Wiscombe, Bounded cas-
cade models as non-stationary multifractals, Physical Review E, 49,
55-69, 1994.
40. Marshak A., A. Davis, W. Wiscombe, and R. Cahalan, T h e scale-
invariant structure of marine stratocumulus deduced from observed
liquid water distributions, I, Spectral properties and stationarity issues,
Journal of the Atmospheric Sciences, submitted, 1994.
41. Meneveau, C , and K. R. Sreenivasan, The multifractal spectrum of
the dissipation field in turbulent flows, Nuclear Physics B, 2, 49-76,
1987.
42. Meneveau, C , and K. R. Sreenivasan, Simple multifractal cascade
model for fully developed turbulence, Physical Review Letters, 59,
1424-1427, 1987.
43. Monin, A. S., and A. M. Yaglom, Statistical Fluid Mechanics, vol. 2,
M I T Press, Boston, Mass, 1975.
44. Muzy, J. F., E. Bacry, and A. Arnéodo, Wavelets and multifractal
formalism for singular signals: Application to turbulence data, Physical
Review Letters, 67, 3515-3518, 1991.
45. Muzy, J. F., E. Bacry, and A. Arnéodo, Multifractal formalism for
fractal signals: The structure function approach versus t h e wavelet-
transform modulus-maxima method, Physical Review E, 47, 875-884,
1993.
46. Novikov, E. A., and R. Stewart, Intermittency of turbulence and spec-
t r u m of fluctuations in energy dissipation, Izvestija Akademii Nauk.
SSSR, Ser. Geofiz., 3, 401-412, 1964.
47. Obukhov, A., Some specific features of atmospheric turbulence, Journal
of Fluid Mechanic, 13, 77-81, 1962.
48. Parisi, G., and U. Frisch, A multifractal model of intermittency, in Ghil,
M., Benzi, R., and Parisi, G. (eds.), Turbulence and Predictability in
Geophysical Fluid Dynamics, North Holland, Amsterdam, 1985.
49. Peitgen, H.-O., and D. Saupe, (eds.), The Science of Fractal Images,
Springer-Verlag, New York, 1988.

50. Press, W. H., S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery,
Numerical Recipes in Fortran, second edition. Cambridge Un. Press,
1993.
51. Schertzer, D., and S. Lovejoy, On the dimension of atmospheric mo-
tions, in Tatsumi, T. (ed.) Turbulence and Chaotic Phenomena in Flu-
ids (IUTAM Symposium 1983), Elsevier North-Holland, New York,
1984.
52. Schertzer, D., and S. Lovejoy, Physical modeling and analysis of rain
clouds by anisotropic scaling multiplicative processes, Journal of Geo-
physical Research, 92, 9693-9714, 1987.
53. Schertzer D., and S. Lovejoy, Hard and soft multifractal processes,
Physica A, 185, 187-194, 1992.
54. Schmitt, F., D. Lavallée, D. Schertzer, and S. Lovejoy, Empirical de-
termination of universal multifractal exponents in turbulent velocity
fields, Physical Review Letters, 68, 305-308, 1992.
55. Stolovitzky G., P. Kailasnath, and K. R. Sreenivasan, Kolmogorovs
refined similarity hypothesis, Physical Review Letters, 69, 1178-1181,
1992.
56. Tessier, Y., S. Lovejoy, and D. Schertzer, Universal multifractals: The-
ory and observations for rain and clouds, Journal of Applied Meteorol-
ogy, 14, 160-164, 1993.
57. Viscek, T., and A.-L. Barabâsi, Multi-affine model for the velocity
distribution in fully turbulent flows, Journal of Physics A: Math. Gen.,
24, L845-L851, 1991.
This work was supported by the Department of Energy's Atmospheric

Radiation Measurement (ARM) project, grant DE-A105-90ER61069 to
NASA's Goddard Space Flight Center. We acknowledge Dr. Alain Arnéodo's
lectures on the elements of waveletry with many applications to scale-
invariant systems. We thank Phil Austin for providing the FIRE data.
We are also appreciative of the many fruitful discussions with T. Bell, R.
Cahalan, C. Duroure, M. Farge, S. Gollmer, D. Lavallée, S. Lovejoy, C.
Meneveau, D. Schertzer, Y. Tessier and T. Warn.
Anthony Davis
Universities Space Research Association
NASA-GSFC (Code 913)
Greenbelt
MD 20771
USA
e-mail: davisuclimate.gsfc.nasa.gov
298 A. Davis et al.
Alexander Marshak
Science Systems and Applications, Inc.
5900 Princess Garden Parkway
Lanham
MD 20706
USA
Warren Wiscombe
NASA Goddard Space Flight Center
Climate and Radiation Branch
Greenbelt
MD 20771
USA
Simultaneous Noise Suppression and Signal Compression
Using a Library of Orthonormal Bases
and the Minimum Description Length Criterion
Naoki Saito
Abstract. We describe an algorithm to estimate a discrete signal from its noisy

observation, using a library of orthonormal bases (consisting of various wavelets,
wavelet packets, and local trigonometric bases) and the information-theoretic cri-
terion called minimum description length (MDL). The key to effective random
noise suppression is that the signal component in the data may be represented
efficiently by one or more of the bases in the library, whereas the noise component
cannot be represented efficiently by any basis in the library. The MDL criterion
gives the best compromise between the fidelity of the estimation result to the data
(noise suppression) and the efficiency of the representation of the estimated signal
(signal compression): it selects the "best" basis and the "best" number of terms
to be retained out of various bases in the library in an objective manner. Because
of the use of the MDL criterion, our algorithm is free from any parameter setting
or subjective judgments.
This method has been applied usefully to various geophysical datasets con-
taining many transient features.
§1. Introduction
Wavelet transforms and their relatives such as wavelet packet trans-
forms and local trigonometric transforms are becoming increasingly popular
in many fields of applied sciences. So far their most successful application
area seems to be data compression; see e.g., [14], [6], [35], [30]. Meanwhile,
several researchers claimed that wavelets and these transforms are also use-
ful for reducing noise in (or denoising) signals/images [16], [7], [10], [21].
In this paper, we take advantage of both sides: we propose an algorithm
for simultaneously suppressing random noise in data and compressing the
signal, i.e., we try to "kill two birds with one stone."

ISBN 0-12-262850-0
300 N. Saito
Throughout this paper, we consider a simple degradation model: ob-

served data consists of a signal component and additive white Gaussian
noise. Our algorithm estimates the signal component from the data using
a library of orthonormal bases (including various wavelets, wavelet pack-
ets, and local trigonometric bases) and the information-theoretic criterion
called the Minimum Description Length (MDL) criterion for discriminating
signal from noise.
The key motivation here is that the signal component in the data can
often be efficiently represented by one or more of the bases in the library
whereas the noise component cannot be represented efficiently by any basis
in the library.
The use of the MDL criterion frees us from any subjective parameter
setting such as threshold selection. This is particularly important for real
field data where the noise level is difficult to obtain or estimate a priori.
The organization of this paper is as follows. In Section 2, we re-
view some of the important properties of wavelets, wavelet packets, lo-
cal trigonometric transforms which constitute the "library of orthonormal
bases" which will be used for efficiently representing nonstationary signals.
In Section 3, we formulate our problem. We view the problem of simultane-
ous noise suppression and signal compression as a model selection problem
out of models generated by the library of orthonormal bases. In Section 4,
we review the MDL principle which plays a critical role in this paper. We
also give some simple examples to help understand its concept. In Sec-
tion 5, we develop an actual algorithm of simultaneous noise suppression
and signal compression. We also give the computational complexity of our
algorithm. Then, we extend our algorithm for higher dimensional signals
(images) in Section 6. In Section 7, we apply our algorithm to several geo-
physical datasets, both synthetic and real, and compare the results with
other competing methods. We discuss the connection of our algorithm with
other approaches in Section 8, and finally, we conclude in Section 9.
§2. A Library of Orthonormal Bases

For our purpose we need to represent signals containing many transient
features and edges in an efficient manner. Wavelets and their relatives, i.e.,
wavelet packets and local trigonometric transforms, have been found very
useful for this purpose; see e.g., [14], [6], [35], [30]. As shown below, each
of these transforms (or basis functions) has different characteristics. In
other words, the best transform to compress a particular signal may not be
good for another signal. Therefore, instead of restricting our attention to a
particular basis, we consider a library of bases. The most suitable basis for
a particular signal is selected from this collection of bases. This approach
leads to a vastly more efficient representation for the signal, compared with
Noise Suppression and Signal Compression 301
confining ourselves to a single basis.

In this section, we briefly describe the most important properties of
these transforms. Throughout this paper, we only consider real-valued
discrete signals (or vectors) with finite length N (= 2 n ). Also we limit our
discussions to orthonormal transforms. Hence it suffices here to consider
discrete orthonormal transforms, i.e., the orthonormal bases of £2(N), the
iV-dimensional space of vectors of finite energy.
More detailed properties of these bases can be found in the literature,
most notably, in [2], [9], [13], [23], [22], [26], [33].
2.1. Wavelet bases
The wavelet transform (e.g., [13], [23]) can be considered as a smooth
partition of the frequency axis. The signal is first decomposed into low
and high frequency components by the convolution-subsamphng operations
with the pair consisting of a "lowpass" filter {hk} and a "highpass" filter
{gk } directly on the discrete time domain. Let H and G be the convolution-
subsamphng operators using these filters and H* and G* be their adjoint
(i.e., upsampling-anticonvolution) operations. It turns out that we can
choose finite-length (L) filters and satisfy the following orthogonality (or
perfect reconstruction) conditions:
HG* = GH* = 0, and H*H + G*G = /,
2
where I is the identity operator of i (N). Also we have the relation gk =
(—l)fc/i£_i_fc. The pair of filters {hk}klQ and {gk)k=o satisfying these
conditions are called quadrature mirror filters (QMFs).
This decomposition (or expansion, or analysis) process is iterated on
the low frequency components and each time the high frequency coefficients
are retained intact and at the last iteration, both low and high frequency
coefficients are kept. In other words, let / = {fk}%=o £ £2(N) be a vector
to be expanded. Then, the convolution-subsamphng operations transform
the vector / into two subsequences Hf and Gf of lengths N/2. Next, the
same operations are applied to the vector Hf to obtain H2f and GHf
of lengths N/4. If the process is iterated J (< n) times, we have the dis-
crete wavelet coefficients (Gf, GHf, GH2f,..., GHJ/, HJ+l / ) of length
N. As a result, the wavelet transform analyzes the data by partitioning
its frequency content dyadically finer and finer toward the low frequency
region (i.e., coarser and coarser in the original time or space domains).
If we were to partition the frequency axis sharply using the charac-
teristic functions (or box-car functions), then we would have ended up the
so-called Shannon (or Littlewood-Paley) wavelets, i.e., the difference of two
sine functions. Clearly, however, we cannot have a finite-length filter in the
time domain in this case. The other extreme is the Haar basis which par-
titions the frequency axis quite badly but gives the shortest filter length
302 N. Saito
(L = 2) in the time domain.

The reconstruction (or synthesis) process is also very simple: start-
ing from the lowest frequency components (or coarsest scale coefficients)
ff J+1 / and the second lowest frequency components GHJ/, the adjoint
operations are applied and added to obtain HJ f = H*HJ+1/ + G*GHJ f.
This process is iterated to reconstruct the original vector / . The compu-
tational complexity of the decomposition and reconstruction process is in
both cases O(N) as easily seen.
We can construct the basis vector w jtk at scale j and position k simply
by putting (GH3 / ) / = J/^, where δι^ denotes the Kronecker delta, and
synthesizing / = Wj^ by the reconstruction algorithm. Using these basis
vectors, we can express the wavelet transform in a vector-matrix form as
a = WTf,
where a G R ^ contains the wavelet coefficients and W G HNxN is an
orthogonal matrix consisting of column vectors Wj^· This basis vector has
the following important properties:
• vanishing moments: Σ/=ο ^mtüj,fc(0 = 0 for m = 0 , 1 , . . . , M — 1.
The higher the degrees of vanishing moments the basis has, the better
it compresses the smooth part of the signal. In the original construction
of Daubechies [12], it turns out that L = 2M. There are several other
possibilities. One of them is a family of the so-called "coiflets" with L = 3M
which are less asymmetric than the original wavelets of Daubechies [13].
• regularity: \wj)k(l + 1) - v*j,k(l)\ ^ c 2 ja

~ >
where c > 0 is a constant and a > 0 is called the regularity of the wavelets.
The larger the value of a is, the smoother the basis vector becomes. This
property may be important if one requires high compression rate since the
shapes of the basis vectors become "visible" in those cases and one might
want to avoid fractal-like shapes in the compressed signals/images [25].
• compact support: wj)k(l) = 0 for / £ [2'fc, Vk + (2j - 1)(L - 1)].

The compact support property is important for efficient and exact numer-
ical implementation.
2.2. Wavelet packet best-bases
For oscillating signals such as acoustic signals, the analysis by the
wavelet transform is sometimes inefficient because it only partitions the
frequency axis finely toward the low frequency. The wavelet packet trans-
form (e.g., [9], [22], [33]) decomposes even the high frequency bands which
are kept intact in the wavelet transform. The first level decomposition
is Hf and Gf just like in the wavelet transform. The second level is

H2f,GHf,HGf,G2f. If we repeat this process for J times, we end
up having JN expansion coefficients. Clearly, we have a redundant set
of expansion coefficients, in fact, there are more than 2 2 possible or-
thonormal bases. One way of selecting an efficient basis for representing
the signal or vector is to use the entropy criterion [9], [33]. We can think
of the wavelet packet bases as a set of different coordinate systems of R .
Then a signal of length N is a point in R ^ , and we try to select the most
efficient coordinate system out of the given set of coordinate systems to
represent this signal. The signal in an efficient coordinate system should
have large magnitudes along a few axes and small magnitudes along most
axes. In particular, the wavelet packet basis function becomes a unit vec-
tor along an axis of the coordinate systems. Then, it is very natural to
use the entropy as a measure of efficiency of the coordinate system. The
best-basis is the basis or coordinate system giving the minimum entropy for
its coordinate distribution. The computational complexity of computing
the best-basis is 0(N log2 N) as is the reconstruction of the original vector
from the best-basis coefficients.
R e m a r k . We would like to note that given a set of signals, the Karhunen-

Loève basis gives the global minimum entropy. However, it is very expensive
to compute; the cost is 0(N3) since it involves solving an eigenvalue prob-
lem. On the other hand, the wavelet packet best-basis can be computed
cheaply and is defined even for a single signal; see [34] for a comparison of
these two bases using images of human faces.
2.3. Local trigonometric best-bases

Local trigonometric transforms ([9], [22], [33], [2]) can be considered as
conjugates of wavelet packet transforms: they partition the time (or space)
axis smoothly. In fact, Coifman and Meyer [8] showed that it is possible to
partition the real-line into any disjoint intervals smoothly and construct or-
thonormal bases on each interval. In the actual numerical implementation,
the data is first partitioned into disjoint intervals by the smooth window
function, and then on each interval the data is transformed by the discrete
cosine or sine transforms (DCT/DST). Since it partitions the axis smoothly,
these transforms, i.e., local cosine or sine transforms (LCT/LST), have less
edge (or blocking) effects than the conventional DCT/DST. Wickerhauser
[33] proposed the method of dyadically partitioning the time axis and com-
puting the best-basis using the entropy criterion similarly to the wavelet
packet best-basis construction. The computational complexity in this case
is about 0(iV[log2 N]2). Local trigonometric transforms are clearly efficient
for the signals with localized oscillating features such as musical notes.
304 N. Saito
§3. P r o b l e m F o r m u l a t i o n
Let us consider a discrete degradation model
d = / + n,
where d, / , n G R and N = 2 n . T h e vector d represents the noisy
observed d a t a and / is the unknown true signal to b e estimated. T h e
vector n is white Gaussian noise (WGN), i.e., n ~ Λ/^Ο, σ2Ι). Let us
assume t h a t σ 2 is unknown.
We now consider an algorithm to estimate / from the noisy observation
d. First, we prepare the library of orthonormal bases mentioned in the
previous section. This library consists of the standard Euclidean basis
of R ^ , the Haar-Walsh bases, various wavelet bases and wavelet packet
best-bases generated by Daubechies' QMFs, their less asymmetric versions
(i.e., coiflets), and local trigonometric best-bases. This collection of bases
is highly adaptable and versatile for representing various transient signals
[7]. For example, if the signal consists of blocky functions such as acoustic
impedance profiles of subsurface structure, the Haar-Walsh bases capture
those discontinuous features b o t h accurately and efficiently. If the signal
consists of piecewise polynomial functions of order p , then the Daubechies
wavelets/wavelet packets with filter length L > 2(p + 1) or the coiflets
with filter length L > 3(p + 1) would be efficient because of the vanishing
moment property. If the signal has a sinusoidal shape or highly oscillating
characteristics, the local trigonometric bases would do the job. Moreover,
computational efficiency of this library is also attractive; the most expensive
expansion in this library, i.e., the local trigonometric expansion, costs about
0(iV[log 2 N]2) as explained in the previous section.
Let us denote this library by C = {#i,#2> · · · , ^ M } 5 where Bm repre-
sents one of t h e orthonormal bases in the library, and M (typically 5 to
20) is the number of bases in this library. If we want, we can add other
orthonormal bases in this library such as the Karhunen-Loève basis [1] or
the prolate spheroidal wave functions [13], [36]. However, normally, the
above-mentioned multiresolution bases are more t h a n enough, considering
their versatility and computational efficiency [7].
Since the bases in the library C compress signals/images very well, we
make a strong assumption here: we suppose the unknown signal / can be
completely represented by k ( < N) elements of a basis # m , i.e.,
/ = W ra a£>, (1)
where W m G R X is an orthogonal matrix whose column vectors are the
basis elements of ß m , and a m G R is the vector of expansion coefficients
of / with only k non-zero coefficients. At this point, we do not know the
actual value of k and the basis Bm. We would like to emphasize t h a t in
reality the signal / might not be strictly represented by (1). We regard (1)
as a model at hand rather than a rigid physical model exactly explaining f
and we will try our best under this assumption. (This is often the case if we
want to fit polynomials to some data.) Now the problem of simultaneous
noise suppression and signal compression can be stated as follows: find
the "best" k and m given the library C. In other words, we translate the
estimation problem into a model selection problem where models are the
bases Bm and the number of terms k under the additive WGN assumption.
For the purpose of data compression, we want to have k as small as
possible. At the same time, we want to minimize the distortion between
the estimate and the true signal by choosing the most suitable basis ß m ,
keeping in mind that the larger k normally gives smaller value of error.
How can we satisfy these seemingly conflicting demands?
§4. The Minimum Description Length Principle

To satisfy the above mentioned conflicting demands, we need a model
selection criterion. One of the most suitable criteria for our purpose is the
so-called Minimum Description Length (MDL) criterion proposed by Ris-
sanen [27], [28], [29]. The MDL principle suggests that the "best" model
among the given collection of models is the one giving the shortest descrip-
tion of the data and the model itself. For each model in the collection, the
length of description of the data is counted as the codelength of encoding
the data using that model in binary digits (bits). The length of description
of a model is the codelength of specifying that model, e.g., the number of
parameters and their values if it is a parametric model.
To help understand what "code" or "encoding" means, we give some
simple examples. We assume that we want to transmit data by first en-
coding (mapping) them into a bitstream by an encoder, then receive the
bitstream by a decoder, and finally try to reconstruct the data. Let L(x)
denote the codelength (in bits) of a vector x of deterministic or probabilis-
tic parameters which are either real-valued, integer-valued, or taking values
in a finite alphabet.
Example 4 . 1 . Codelength of symbols drawn from a finite alphabet.

Let x = (xi,X2?··· Ϊ#ΛΓ) be a string of symbols drawn from a finite al-
phabet X, which are independently and identically distributed (i.i.d.) with
probability mass function p(x), x £ X. In this case, clearly the frequently
occurring symbols should have shorter codelengths than rarely occurring
symbols for efficient communication. This leads to the so-called Shannon
code [11] whose codelength (if we ignore the integer requirement for the
codelength) can be written as
L(x) = — logp(x) for x £ X.
306 N. Saito
(Prom now on, we denote the logarithm of base 2 by "log", and the nat-
ural logarithm, i.e., base e by "In".) The Shannon code has the shortest
codelength on the average, and satisfies the so-called Kraft inequality [11]:
Σ 2" L(x) < 1, (2)
which is necessary and sufficient for the existence of an instantaneously

decodable code, i.e., a code such that there is no codeword which is the
prefix of any other codeword in the coding system. The shortest codelength
on the average for the whole sequence x becomes
TV N
L x ι χ{
( ) = Σ( ) = -Σ1ο&ρ(χί)'
i=l t=l
Example 4.2. Codelength of deterministic integers.

For a deterministic parameter j € Z^v = ( 0 , 1 , . . . , N — 1) (i.e., both the
encoder and decoder know TV), the codelength of describing j is written as
L(j) = logN since logN bits are required to index N integers. This can
also be interpreted as a codelength using Shannon code for a sample drawn
from the uniform distribution over ( 0 , 1 , . . . , N — 1).
Example 4.3. Codelength of an integer (universal prior for an integer).

Suppose we do not know how large a natural number j is. Rissanen [27]
proposed that the code of such j should be the binary representation of
j , preceded by the code describing its length logj, preceded by the code
describing the length of the code for log j , and so forth. This recursive
strategy leads to
L*(j) = log* j + log co = log j + log log j + · · · + logc 0 ,
where the sum involves only the non-negative terms and the constant
Co « 2.865064 which was computed so that equality holds in (2), i.e.,
]Cjii 2~L*(iï = i. This can be generalized for an integer j by defining
l
L*(j)= { if i = 0, (3 ,
^' ]^ log* \j\ +log4c 0 otherwise. ^'
L
(We can easily see that (3) satisfies ^2^L_QO 2~ ^ = 1.)
Example 4.4. Codelength of a truncated real-valued parameter.

For a deterministic real-valued parameter v € R, the exact code generally
requires infinite length of bits. Thus, in practice, some truncation must
be done for transmission. Let S be the precision and vs be the truncated
value, i.e., \v — v$\ < δ. Then, the number of bits required for v$ is the
sum of the codelength of its integer part [v] and the number of fractional
binary digits of the truncation precision J, i.e.,
L(t,,) = L * ( H ) + l o g ( l / i ) . (4)
Having gone through the above examples, now we can state the MDL
principle more clearly. Let M = {0 m : m = l , 2 , . . . } b e a class or collection
of models at hand. The integer m is simply an index of a model in the list.
Let x be a sequence of observed data. Assume that we do not know the
true model 0 generating the data x. As in [29], [24], given the index m, we
can write the codelength for the whole process as
L(x, 0 m , m) = L(m) + L(0 m \ m) + L(x | 0 m , m). (5)
This equation says that the codelength to rewrite the data is the sum of the
codelengths to describe: (i) the index m, (ii) the model 0 m given ra, and
(iii) the data x using the model 0 m . The MDL criterion suggests picking
the model 0 m * which gives the minimum of the total description length
(5).
The last term of the right-hand side (RHS) of (5) is the length of the
Shannon code of the data assuming the model 0 m is the true model, i.e.,
L(x | 0 m , m ) = - l o g p ( x | 0 m ,ra), (6)
and the maximum likelihood (ML) estimate 0 m minimizes (6) by the defi-
nition:
L(x | 0 m , m ) = - l o g p ( x | 0 m , m ) < - l o g p ( x | 0 m ,ra). (7)
However, we should consider a further truncation of 0 m as shown in Ex-
ample 4.4 above to check that additional savings in the description length
is possible. The finer truncation precision we use, the smaller the term
(7), but the larger the term L(0 m | m) becomes. Suppose that the model
0 m has km real-valued parameters, i.e., 0 m = (0 m > i,... ,0m,fcm)· Rissanen
showed in [27], [29] that the optimized truncation precision (δ*) is of order
1/v/ÏVand
min L(x, 0m,<$, ra, δ)
δ
= L(m) + L(em>s' | m) + L(x \ 0 m ,*., m) + 0(km)
« L(m) + ΣL*([0mj]) + ^logN + L(x | 0 m , m ) + 0(fc m ), (8)

i=i
where 0 m is the optimal non-truncated value given m, 0mj$* is its opti-
mally truncated version, and L*(-) is defined in (4). We note that the last
term 0(km) in the approximation in (8) includes the penalty codelength
308 N. Saito
necessary to describe the data x using the truncated ML estimate 0mj<$*

instead of the true ML estimate 0m. In practice, we rarely need to obtain
the optimally truncated value Bmj* and we should compute 0 m up to the
machine precision, say, 10~ 15 , and use that value as the true ML estimate
in (8). For sufficiently large JV, the last term may be omitted, and instead
of minimizing the ideal codelength (5), Rissanen proposed to minimize
km u
lo
MDL(x, 0 m , m) = L(m) + ^L*([0mj]) + -f S N + L(x | 0 m , m ) . (9)
i=i
The minimum of (9) gives the best compromise between the low complexity
in the model and high likelihood on the data.
The first term of the RHS of (9) can be written as
L(m) = - l o g p ( m ) , (10)
where p(m) is the probability of selecting m. If there is prior information
about m as to which m is more likely, we should reflect this in p(m).
Otherwise, we assume each m is equally likely, i.e., p(m) is a uniform
distribution.
R e m a r k . Even though the list of models M does not include the true
model, the MDL method achieves the best result among the available mod-
els. See Barron and Cover [4] for detailed information on the error between
the MDL estimate and the true model.
We also would like to note that the MDL principle does not attempt
to find the absolutely minimum description of the data. The MDL always
requires an available collection of models and simply suggests picking the
best model from that collection. In other words, the MDL can be considered
as an "oracle" for model selection [24]. This contrasts with the algorithmic
complexities such as the Kolmogorov complexity which gives the absolutely
minimum description of the data, however, in general, is impossible to
obtain [27].
Before deriving our simultaneous noise suppression and signal compres-

sion algorithm in the context of the MDL criterion, let us give a closely
related example:
Example 4.5. A curve fitting problem using polynomials.

Given N points of data (xj,y t ) G R 2 , consider the problem of fitting a
polynomial through these points. The model class we consider is a set of
polynomials of orders 0 , 1 , . . . , N — 1. In this case, 0 m = ( a o , a i , . . . , a m )
represents the m + 1 coefficients of a polynomial of order m. We also assume
that the data is contaminated by the additive WGN with known variance
σ 2 , i.e.,
Vi = /(«,·) +e<»
where /(·) is an unknown function to be estimated by the polynomial mod-
els, and e,· ~ JV(0, σ 2 ). To invoke the MDL formalism, we pose this ques-
tion in the information transmission setting. First we prepare an encoder
which computes the ML estimate of the coefficients of the polynomial,
(3o,... , S m ) , of the given degree m from the data. (In the additive WGN
assumption the ML estimate coincides with the least squares estimate.)
This encoder transmits these m coefficients as well as the estimation errors.
We also prepare a decoder which receives the coefficients of the polynomial
and residual errors and reconstruct the data. (We assume that the abscis-
sas {xi}iLi and the noise variance σ2 are known to both the encoder and
the decoder.) Then we ask how many bits of information should be trans-
mitted to reconstruct the data. If we used polynomials of degree N — 1,
we could find a polynomial passing through all N points. In this case, we
could describe the data extremely well. In fact, there is no error between
the observed data and those reconstructed by the decoder. However, we
do not gain anything in terms of data compression/transmission since we
also have to encode the model which requires N coefficients of the polyno-
mial. In some sense, we did not "learn" anything in this case. If we used
the polynomial of degree 0, i.e., a constant, then it would be an extremely
efficient model, but we would need many bits to describe the deviations
from that constant. (Of course, if the underlying data is really a constant,
then the deviation would be 0.)
Let us assume there is no prior preference on the order m. Then we
can easily see that the total codelength (9) in this case becomes
m -
MDL(y,em,m) = logN + V r ( [ 3 , · ] ) + ^ — logiV

l
i=o
t=l \ j=0 /
The MDL criterion suggests to pick the "best" polynomial of order m* by

minimizing this approximate codelength.
The MDL criterion has been successfully used in various fields such
as signal detection [32], image segmentation [19], and cluster analysis [31]
where the optimal number of signals, regions, and clusters, respectively,
should be determined. If one knows a priori the physical model to explain
the observed data, that model should definitely be used, e.g., the complex
310 N. Saito
sinusoids in [32]. However, in general, as a descriptor of real-life signals

which are full of transients or edges, the library of wavelets, wavelet packets,
and local trigonometric transforms is more flexible and efficient than the
set of polynomials or sinusoids.
§5. A Simultaneous Noise Suppression and Signal Compression

Algorithm
We carry on our development of the algorithm based on the information
transmission setting as the polynomial curve fitting problem described in
the previous section. We consider again an encoder and a decoder for our
problem. Given (fc,m) in (1), the encoder expands the data d in the basis
# m , then transmits the number of terms fc, the specification of the basis
m, and k expansion coefficients, the variance of the WGN model σ 2 , and
finally the estimation errors. The decoder receives this information in bits
and tries to reconstruct the data d.
In this case, the total codelength to be minimized may be expressed
as the sum of the codelengths of: (i) two natural numbers (fc, ra), (ii) (k +
1) real-valued parameters (am\a2) given (k, m), and (iii) the deviations
of the observed data d from the (estimated) signal / = Wmocm given
(A:, ra, dm , σ 2 ). The approximate total description length (9) now becomes
MDL(d,a^\a2,k,m)
= L(fc,ra) + L ( a ^ } , 5 2 \k,m) + L(d | a ^ , 5 2 , f c , m ) , (11)
where OL^ and σ 2 are the ML estimates of <Xm and σ 2 , respectively.

Let us now derive these ML estimates. Since we assumed the noise
component is additive WGN, the probability of observing the data given
all model parameters is
P(d I «LfcV,fc,m) = (2πσ2Γ"/2βχρ i - 1 1 * " ^ " " "Λ , (12)
where || · || is the standard Euclidean norm on R ^ . For the ML estimate

of σ 2 , first consider the log-likelihood of (12)
\np(d | a%\az,k,m) = -γ1η2πσζ - ^ — "-. (13)
Taking the derivative with respect to σ 2 and setting it to zero, we easily

obtain
52 = ^ | | d - W m a W | | 2 . (14)
Insert this equation back to (13) to get
lnp(d | oti\d\k,m) = - | l n ( ^ | | d - W m a £ > | | 2 ) _ E. (i5)
Let dm = W^d denote the vector of the expansion coefficients of d in the

basis Bm. Since this basis is orthonormal, i.e., Wm is orthogonal, and we
use the £2 norm, we have
||<f - Wma£\\* = \\Wm(Wld - «W)|| 2 = \\dm - *%ψ. (16)
From (15), (16), and the monotonicity of the In function, we find that
maximizing (15) is equivalent to minimizing
Κ,-α^ΙΙ2· (17)
Considering that the vector a™ only contains k nonzero elements, we can
easily conclude that the minimum of (17) is achieved by taking the largest
k coefficients in magnitudes of d m as the ML estimate of ctm\ i.e.,
S « = 0(l)dm = 0(t)(Wlrf), (18)
where Θ^ ^ is a thresholding operation which keeps the k largest elements in
absolute value intact and sets all other elements to zero. Finally, inserting
(18) into (14), we obtain
σ 2 = jjWWld - &^Wldf = 1 | | ( I - &^)Wld\\\ (19)
where / represents the N dimensional identity operator (matrix).

Let us further analyze (11) term by term. If we do not have any prior
information on (fc,m), then the cost L(k,m) is the same for all cases, i.e.,
we can drop the first term of (11) for minimization purpose. However, if
one has some prior preference about the choice of basis, knowing some prior
information about the signal / , L(fc, m) should reflect this information. For
instance, if we happen to know that the original function / consists of a
linear combination of dyadic blocks, then we clearly should use the Haar
basis. In this case, we may use the Dirac distribution, i.e., p(m) = J m , m o ,
where mo is the index for the Haar basis in the library £. By (10), this
leads to
L k if
L(h m)=f () ™ = ™o,
^ ' ' \ +oo otherwise.
On the other hand, if we either happen to know a priori or want to force
the number of terms retained (k) to satisfy k\ < k < ki, then we may want
to assume the uniform distribution for this range of k, i.e.,
L(k,m) = [ L ( m ) + log(fc2
- kl + X) if 1
ί ^ k ^ *2' (20)
v v
' ^ +oo otherwise. '
312 N. Saito
As for the second term of (11), which is critical for our algorithm, we
have to encode k expansion coefficients â^J and σ 2 , i.e., (fc+1) real-valued
parameters. However, in this case, by normalizing the whole sequence by
||d||, we can safely assume that the magnitude of each coefficient in a ' '
is strictly less than one; in other words, the integer part of each coefficient
is simply zero. Hence we do not need to encode the integer part as in (9)
if we transmit the real-valued parameter ||d||. Now the description length
of (œm ,σ 2 ) given (fc, m) becomes approximately ^ ^ l o g i V - h L*([S2]) +
L*([||d||]) bits since there are k + 2 real-valued parameters: k nonzero
coefficients, σ 2 , and ||d||. After normalizing by ||d||, we clearly have σ 2 < 1
(see (19)), so that L*([52]) = 1 (see (3)). For each expansion coefficient,
however, we still need to specify the index of the coefficient, i.e., where the
k non-zero elements are in the vector orm . This requires klogN bits. As
a result, we have
L{o£\a2 |fc,m) = |felogN + c, (21)

where c is a constant independent of (fc, m).
Since the probability of observing d given all model parameters is given
by (12), we have for the last term in (11)
L(d | a<£>,9*,k,m) = jlog||(J- e<*>)W£d|| 2 + c', (22)

where c' is a constant independent of (fc, m).
Finally we can state our simultaneous noise suppression and signal
compression algorithm. Let us assume that we do not have any prior infor-
mation on (fc, m) for now. Then, from (11), (21), and (22) with ignoring
the constant terms c and c', our algorithm can be stated as:
Pick the index (k*,m*) such that
AMDL(k\m*)= min ( ^HogiV + ^ log \\{I - Q^W^df ).

\<m<M
(23)
Then reconstruct the signal estimate
f = Wm.a{£). (24)
Let us call the objective function to be minimized in (23), the approx-
imate MDL (AMDL) since we ignored the constant terms. Let us now
show a typical behavior of the AMDL value as a function of the number
of terms retained (k) in Figure 1. (In fact, this curve is generated using
Example 7.1 below.) We see that the log(residual energy) always decreases
as k increases. By adding the penalty term of retaining the expansion coef-
ficients, i.e., (3/2)fclogiV (which is just a straight line), we have the AMDL
0 500 1000 7 1500 2000 2500

k
Figure 1. Graphs of AMDL versus k: AMDL [solid line] which is the
sum of the (3/2)fc log N term [dotted line] and the (ΑΓ/2) log(residual
energy) term [dashed line].
curve which typically decreases for the small fc, then starts increasing be-
cause of the penalty term, then finally decreases again at some large k near
from k = N because the residual error becomes very small. Now what we
really want is the value of k achieving the minimum at the beginning of
the fc-axis, and we want to avoid searching for k beyond the maximum oc-
curring for k near N. So, we can safely assume that k\ = 0 and &2 = N/2
in (20) to avoid searching more than necessary. (In fact, setting &2 > N/2
does not make much sense in terms of data compression either.)
We briefly examine below the computational complexity of our algo-
rithm. To obtain (fc*,m*), we proceed as follows:
Step 1: Expand the data d into bases B\,..., BM · Each expansion (includ-
ing the best-basis selection procedure) costs O(N) for wavelets, 0(iV log N)
for wavelet packet best-bases, and 0(N[\ogN]2) for local trigonometric
best-bases.
Step2: Let K(= &2 — &i + 1) denote the length of the search range for k.
For k\ < k < k*i, 1 < m < M, compute the expression in the parenthesis of
the RHS in (23). This costs approximately 0(N + 3MK) multiplications
and MK calls to the log function.
Step 3: Search the minimum entry in this table, which costs MK compar-
314 N. Saito
isons.
Step 4: Reconstruct the signal estimate (24), which costs O(N) for wavelets,
0(ΛΓ log N) for wavelet packet best-bases, and 0(7V[log ΛΓ]2) for local trigono-
metric best-bases.
§6· Extension t o Images

For images or multidimensional signals, we can easily extend our al-
gorithm by using the multidimensional version of the wavelets, wavelet
packets, and local trigonometric transforms. In this section, we briefly
summarize the two-dimensional (2D) versions of these transforms. For the
2D wavelets, there are several different approaches. The first one, which
we call the sequential method, is the tensor product of the one-dimensional
(ID) wavelets, i.e., applying the wavelet expansion algorithm separately
along two axes t\ and t2 corresponding to column (vertical) and row (hor-
izontal) directions respectively. Let / G κΛ 1 *^ 2 and Η,, G, be the ID
convolution-subsampling operations along axis tî = 1,2. Then this ver-
sion of the 2D wavelet transform first applies the convolution-subsampling
operations along the £χ axis to obtain fl = (Gif,GiHif,...,GiHllf),
then applies the convolution-subsampling operations along the t2 axis to get
the final 2D wavelet coefficients (G2fl,G2H2f1,... ,G2H22 fx) °f l e n gth
N\ x iV2, where J\ {< \ogN\) and J2 (< log7V"2) are maximum levels of
decomposition along t\ and t2 axes respectively. We note that one can
choose different ID wavelet bases for t\ and t2 axes independently. Given
M different QMF pairs, there exist M2 possible 2D wavelets using this
approach.
The second approach is the basis generated from the tensor product
of the multiresolution analysis. This decomposes an image / into four
different sets of coefficients, H\H2f, G\H2f, H\G2f, and G\G2f, cor-
responding to "low-low", "high-low", "low-high", "high-high" frequency
parts of the two variables, respectively. The decomposition is iterated on
the "low-low" frequency part and this ends up in a "pyramid" structure of
coefficients. Transforming the digital images by these wavelets to obtain
the 2D wavelet coefficients are described in e.g., [20], [13].
There are also 2D wavelet bases which do not have a tensor-product
structure, such as wavelets on the hexagonal grids and wavelets with matrix
dilations. See e.g., [18], [17] for details.
There has been some argument as to which version of the 2D wavelet
bases should be used for various applications [5], [13]. Our strategy toward
this problem is this: we can put as many versions of these bases in the
library as we can afford it in terms of computational time. Then minimiz-
ing the AMDL values automatically selects the most suitable one for our
purpose.
As for the 2D version of the wavelet packet best-basis, the sequen-
tial method may be generalized, but it is not easily interpreted; the ID
best-bases may be different from column to column so that the resultant
coefficients viewing along the row direction may not share the same fre-
quency bands and scales unlike the 2D wavelet bases. This also makes the
reconstruction algorithm complicated. Therefore, we should use the other
tensor-product 2D wavelet approach for the construction of the 2D wavelet
packet best-basis: we recursively decompose not only the "low-low" com-
ponents but also the other three components. This process produces the
"quad-tree" structure of wavelet packet coefficients instead of the "binary-
tree" structure for ID wavelet packets. Finally the 2D wavelet packet
best-basis coefficients are selected using the entropy criterion [33].
The 2D version of the local trigonometric transforms can be con-
structed using the quad-tree structure again: the original image is smoothly
folded and segmented into 4 subimages, 16 subimages, . . . , and in each
subimage the separable DCT/DST is applied, and then the quad-tree struc-
ture of the coefficients is constructed. Finally, the local trigonometric best-
basis is selected using the entropy criterion [33].
For an image of N = Ni x ΛΓ2 pixels, the computational costs are
approximately O(N), 0(N\og4 JV), 0(iV[log4 iV]2) for a 2D wavelet, a 2D
wavelet packet best-basis, a 2D local trigonometric best-basis, respectively.
§7. Examples
In this section, we give several examples to show the usefulness of our
algorithm.
Example 7.1. The Synthetic Piecewise Constant Function of Donoho-

Johnstone.
We compared the performance of our algorithm in terms of the visual qual-
ity of the estimation and the relative £2 error with Donoho-Johnstone's
method using the piecewise constant function used in their experiments
[16]. The results are shown in Figure 2. The true signal is the piecewise
constant function with N = 2048, and its noisy observation was created
by adding the WGN sequence with ||/||/||n|| = 7. The library C for this
example consisted of 18 different bases: the standard Euclidean basis of
Β Λ , the wavelet packet best-bases created with D02, D04, . . . , D20, C06,
C12, . . . , C30, and the local cosine and sine best-bases (On represents the
n-tap QMF of Daubechies and Cn represents the n-tap coiflet filter). In
the Donoho-Johnstone method, we used the C06, i.e., 6-tap coiflet with
2 vanishing moments. We also specified the scale parameter J = 7, and
supplied the exact value of σ2. Next, we forced the Haar basis (D02) to
316 N. Saito
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 2. Results for the synthetic piecewise constant function: (a)

Original piecewise constant function, (b) Noisy observation with (signal
energy)/(noise energy) = 7 2 . (c) Estimation by the Donoho-Johnstone
method using coiflets C06. (d) Estimation by the Donoho-Johnstone
method using Haar basis, (e) Estimation by the proposed method.
be used in their method. Finally, we applied our algorithm without spec-

ifying anything. In this case, the Haar-Walsh best-basis with k* = 63
was automatically selected. T h e relative £2 errors are 0.116, 0.089, 0.051,
respectively. Although the visual quality of our result is not too differ-
ent from Donoho and Johnstone's (if we choose the appropriate basis for
their m e t h o d ) , our method generated the estimate with the smallest rela-
tive £2 error and slightly sharper edges. (See Section 8 for more about the
Donoho-Johnstone method and its relation to our method.)
E x a m p l e 7 . 2 . A Pure White Gaussian Noise.

We generated a synthetic sequence of W G N with σ 2 = 1.0 and N = 4096.
T h e same library as in Example 7.1 (with the best-bases adapted to this
pure W G N sequence) was used. We also set the upper limit of search range
k2 = N/2 = 2048. Figure 3 shows the AMDL curves versus k for all bases
in t h e library. As we can see, there is no single minimum in the graphs,
and our algorithm satisfactorily decided k* = 0, i.e., there is nothing to
"learn" in this dataset.
x 10
-1
-2
0 500 1000 1500 2000 2500 3000 3500 4000 4500

k
Figure 3. The AMDL curves of the White Gaussian Noise data for all
bases. For each basis, k = 0 is the minimum value. The vertical dotted
line indicates the upper limit of the search range for k.
Example 7.3. A Natural Radioactivity Profile of Subsurface Formation.

We tested our algorithm on the actual field data which are measurement of
natural radioactivity of subsurface formation obtained at an oil-producing
well. The length of the data is N = 1024. Again, the same library was used
as in the previous examples. The results are shown in Figure 4. In this
case, our algorithm selected the D12 wavelet packet best-basis (Daubechies'
12-tap filter with 6 vanishing moments) with fc* = 77. The residual error is
shown in Figure 4 (c) which consists mostly of a WGN-like high frequency
component. The compression ratio is 1024/77 « 13.3. However, to be able
to reconstruct the signal from the surviving coefficients, we still need to
record the indices of those coefficients.
Suppose we can store each index by &,· bytes of memory and the preci-
sion of the original data is bf bytes per sample. Then the storage reduction
ratio Rs can be computed by
Rs (25)
~—W^b}—-7il +
b}>>
where r is a compression ratio. The original data precision was bf = 8
(bytes) in this case. Since it is enough to use b{ = 2 (bytes) for indices and
r = 13.3%, we have Rs « 9.40%, i.e., 90.60% of the original data can be
318 N. Saito
KJLi
. l*t ^ψΥ^/^
' ' \ * ^ , Λ^ * « Γ % Α ^ ' l/'vw*vv^^ww^j» \wj i^j
t i l î
IfV ;
llfS /ÂM/WV/X^J
Ιν ν / ^^-\^^/ντ
k/#^w^
0 200 400 600 800 1000
Figure 4. The estimate of the natural radioactivity profile of subsur-

face formation: (a) Original data which was measured in the borehole
of an oil-producing well, (b) Estimation by the proposed method, (c)
Residual error between (a) and (b).
discarded.
E x a m p l e 7 . 4 . A Migrated Seismic Section.

In this example, t h e d a t a is a migrated seismic section as shown in Fig-
ure 5 (a). T h e d a t a consist of 128 traces of 256 time samples. We selected
six 2D wavelet packet best-bases (D02, C06, C12, C18, C24, C30) as t h e
library. Figure 5 (b) shows t h e estimate by our algorithm. It automatically
selected t h e filter C30 and t h e number of terms retained as k* = 1611. If
we were t o choose a good threshold in this example, it would b e fairly dif-
ficult since we do not know t h e accurate estimate of σ 2 . T h e compression
rate, in this case, is (128 x 256)/1611 « 20.34. T h e original d a t a precision
was bf = 8 as in t h e previous example. In this case we have t o use bi — 3
(1 byte for row index, 1 byte for column index, and 1 byte for scale level).
If we p u t these and r = 20.34% into (25), we have Rs « 6.76%, i.e., 93.24%
of t h e original d a t a can b e discarded. Figure 5 (c) shows t h e residual er-
ror between t h e original and t h e estimate. We can clearly see t h e r a n d o m
noise and some strange high frequency patterns (which are considered t o
b e numerical artifacts from t h e migration algorithm applied).
(a) (b) (c)

F i g u r e 5. Results for the migrated seismic section: (a) Original seis-
mic section with 128 traces and 256 time samples, (b) Estimation by the
proposed method, (c) Residual error between (a) and (b). (Dynamic
range of display (c) is different from those of (a) and (b).)
§8. D i s c u s s i o n s
Our algorithm is intimately connected to the "denoising" algorithm
of Coifman and Majid [7], [10]. Their algorithm first picks t h e best-basis
from the collection of bases and sorts the best-basis coefficients in order of
decreasing magnitude. Then they use t h e "theoretical compression r a t e " of
the sorted best-basis coefficients { α ι · } ^ : 1 as a key criterion for separating
a signal component from noise. T h e theoretical compression r a t e of a unit
vector u is defined as c(u) = 2H^/N(u), where H(u) is the ^ 2 -entropy of
u w
TA, i.e., H(u) = — Σί=ι l l ° S i -> ^ d N(u) is the length of u. We note
t h a t 0 < c(u) < 1 for any real unit vector u , and c(u) = 0 implies u =
{Sifi0} for some z'o (the best possible compression), and c(u) = 1 implies
u = ( 1 , . . . , I)/ y/N(u) (the worst compression). T h e n t o decide how many
coefficients to keep as a signal component, they compare c{{ai]f_k+l)^ the
320 N. Saito
theoretical compression r a t e of the noise component (defined as the smallest

(N — k) coefficients), to the predetermined threshold r . They search k =
0 , 1 , . . . which gives an unacceptably bad compression rate: c({#i}£Lfc+i) >
τ . Their algorithm critically depends on the choice of the threshold r
whereas our algorithm needs no threshold selection. On t h e other hand,
their algorithm does not assume the W G N model we used in this paper;
rather, they defined the noise component as a vector reconstructed from
the best-basis coefficients of small magnitude.
Our algorithm can also be viewed as a simple yet flexible and effi-
cient realization of the "complexity regularization" m e t h o d for estimation
of functions proposed by Barron [3]. He considered a general regression
function estimation problem: given the d a t a (χί,νί)^, where {xi £ R p }
is a sequence of the (p-dimensional) sampling coordinates (or explanatory
variables) and {y,· E R } is the observed d a t a (or response variables), se-
lect a "best" regression function fa out of a list (library) Cjy of candidate
functions (models). He did not impose any assumption on the noise distri-
bution, but assumed t h a t the number of models in the list CN depends on
the number of observations N. Now the complexity regularization m e t h o d
of Barron is to find //y such t h a t
R(fN) = min i l Σ d(yi, /(*,·)) + jjL(f)J ,
where d(·, ·) is a measure of distortion (such as the squared error), λ > 0 is

a regularization constant, and L(f) is a complexity of a function / (such
as the L(m) + L(0m \ m) term in (5)). He showed t h a t various asymptotic
properties of the estimator fa as N —> oo, such as bounds on the estimation
error, the r a t e of convergence, etc. If we restrict our attention to the finite
dimensional vector space, use the library of orthonormal bases described in
Section 2, adopt the length of the Shannon code (6) as a distortion measure,
assume the W G N model, and finally set λ = 1, then Barron's complexity
regularization m e t h o d reduces to our algorithm. Our approach, although
restricted in the sense of Barron, provides a computationally efficient and
yet flexible realization of the complexity regularization method, especially
compared to the library consisting of polynomials, splines, trigonometric
series discussed in [3].
Our algorithm also has a close relationship with the denoising algorithm
via "wavelet shrinkage" developed by Donoho and Johnstone [16]. (A well-
written summary on the wavelet shrinkage and its applications can b e found
in [15].) Their algorithm first transforms the observed discrete d a t a into a
wavelet basis (specified by the user), then applies a "soft threshold" r =
σχ/ΙηΛΓ to the coefficients, i.e., shrinks magnitudes of all the coefficients
by the amount r toward zero. Finally the denoised d a t a is obtained by
the inverse wavelet transform. Donoho claimed informally in [15] that the
reason why their method works is the ability of wavelets to compress the
signal energy into a few coefficients. The main differences between our
algorithm and that of Donoho and Johnstone are:
• Our method automatically selects the most suitable basis from a

collection of bases whereas their method uses only a fixed basis
specified by the user.
• Our method includes adaptive expansion by means of wavelet pack-

ets and local trigonometric bases whereas their method only uses a
wavelet transform.
• Their method requires the user to set the coarsest scale parameter
J <n and a good estimate of σ 2 , and the resulting quality depends
on these parameters. On the other hand, our method does not
require any such parameter setting.
• Their approach is based on the minimax decision theory in statistics

and addresses the risk of the estimation whereas our approach uses
the information-theoretic idea and combines denoising and the data
compression capability of wavelets explicitly.
• Their method thresholds the coefficients softly whereas our method

can be said to threshold sharply. This might cause some Gibbs-like
effects in the reconstruction using our method.
Future extensions of this research are to: incorporate noise models

other than Gaussian noise, extend the algorithm for highly nonstationary
signals by segmenting them smoothly and adaptively, investigate the effect
of sharp thresholding, and study more about the relation with the com-
plexity regularization method of Barron as well as the wavelet shrinkage of
Donoho-Johnstone.
We have described an algorithm for simultaneously suppressing the ad-
ditive WGN component and compressing the signal component in a dataset.
One or more of the bases in the library, consisting of wavelets, wavelet pack-
ets, and local trigonometric bases, compress the signal component quite
well, whereas the WGN component cannot be compressed efficiently by any
basis in the library. Based on this observation, we have tried to estimate
the "best" basis and the "best" number of terms to retain for estimating the
signal component in the data using the MDL criterion. Both synthetic and
real field data examples have shown the wide applicability and usefulness
of this algorithm.
322 N. Saito
Acknowledgements
The author would like to thank Prof. R. Coifman and Prof. A. Barron
of Yale University for fruitful discussions.
References
1. Ahmed, N. and K. R. Rao, Orthogonal Transforms for Digital Signal
Processing, Springer-Verlag, New York, 1975.
2. Auscher, P., G. Weiss, and M. V. Wickerhauser, Local sine and cosine
bases of Coifman and Meyer and the construction of smooth wavelets,
in Wavelets: A Tutorial in Theory and Applications, C. K. Chui (ed.),
Academic Press, San Diego, 237-256, 1992.
3. Barron, A. R., Complexity regularization with application to artificial
neural networks, in Proceeding NATO A SI on Nonparametric Func-
tional Estimation, Kluwer, 1991.
4. Barron, A. R. and T. M. Cover, Minimum complexity density estima-
tion, IEEE Trans. Inform. Theory, 37 (4), 1034-1054, 1991.
5. Beylkin, G., R. Coifman, and V. Rokhlin, Fast wavelet transforms and
numerical algorithms I, Comm. Pure Appl. Math., 44, 141-183, 1991.
6. Bradley, J. N. and C. M. Brislawn, Image compression by vector quan-
tization of multiresolution decompositions, Physica D, 60, 245-258,
1992.
7. Coifman, R. R. and F. Majid, Adapted waveform analysis and de-
noising, in Progress in Wavelet Analysis and Applications, Y. Meyer
and S. Roques (eds.), Editions Frontieres, B.P.33, 91192 Gif-sur-Yvette
Cedex, France, 63-76, 1993.
8. Coifman, R. R. and Y. Meyer. Remarques sur l'analyse de fourier à
fenêtre, Comptes Rendus Acad. Sei. Paris, Série I, 312, 259-261, 1991.
9. Coifman, R. R. and M. V. Wickerhauser, Entropy-based algorithms
for best basis selection, IEEE Trans. Inform. Theory, 38 (2), 713-719,
1992.
10. Coifman, R. R. and M. V. Wickerhauser, Wavelets and adapted
waveform analysis, in Wavelets: Mathematics and Applications,
J. Benedetto and M. Frazier (eds.), chapter 10, CRC Press, Boca Ra-
ton, Florida, 1993.
11. Cover, T. M. and J. A. Thomas, Elements of Information Theory,
Wiley Interscience, New York, 1991.
Comm. Pure Appl. Math., 4 1 , 909-996, 1988.
13. Daubechies, L, Ten Lectures on Wavelets, volume 61 of CBMS-NSF
Regional Conference Series in Applied Mathematics, SIAM, Philadel-
phia, 1992.
14. DeVore, R. A., B. Jawerth, and B. J. Lucier, Image compression
through wavelet transform coding, IEEE Trans. Inform. Theory, 38

(2), 719-746, 1992.
15. Donoho, D. L., Wavelet shrinkage and W.V.D.: a 10-minute tour,
in Progress in Wavelet Analysis and Applications, Y. Meyer and
S. Roques (eds.), Editions Frontieres, B.P.33, 91192 Gif-sur-Yvette
Cedex, Prance, 109-128, 1993.
16. Donoho, D. L. and I. M. Johnstone, Ideal spatial adaptation by wavelet
shrinkage, preprint, Dept. of Statistics, Stanford University, Stanford,
CA, Jun. 1992, revised Apr. 1993.
17. Gröchenig, K. and W. R. Madych, Multiresolution analysis, Haar
bases, and self-similar tilings of Rn, IEEE Trans. Inform. Theory,
38 (2), 556-568, 1992.
18. Kovacevic, J. and M. Vetterli, Nonseparable multidimensional perfect
reconstruction filter banks and wavelet bases for Rn, IEEE Trans.
Inform. Theory, 38 (2), 533-555, 1992.
19. Leclerc, Y. G., Constructing simple stable descriptions for image par-
titioning, Intern. J. Computer Vision, 3, 73-102, 1989.
20. Mallat, S., A theory for multiresolution signal decomposition, IEEE
Trans. Pattern Anal. Machine Intell, 11 (7), 674-693, 1989.
21. Mallat, S. and Z. Zhang, Matching pursuit with time-frequency dic-
tionaries, IEEE Trans. Signal Processing, the special issue on wavelets
and signal processing, 41 (12), 1993, to appear.
22. Meyer, Y., Wavelets: Algorithms and Applications, translated and
revised by R. D. Ryan, SIAM, Philadelphia, PA, 1993.
23. Meyer, Y., Wavelets and Operators, volume 37 of Cambridge Studies
in Advanced Mathematics, translated by D. H. Salinger, Cambridge
University Press, New York, 1993.
24. Niblack, W., MDL Methods in Image Analysis and Computer Vision,
IEEE Conf. Comput. Vision, Pattern Recognition, Tutorial note, New
York, Jun. 1993.
25. Rioul, O., Regular wavelets: a discrete-time approach, IEEE Trans.
Signal Processing, the special issue on wavelets and signal processing,
41 (12) (1993), to appear.
26. Rioul, O. and M. Vetterli, Wavelets and signal processing, IEEE SP
Magazine, 8 (4), 14-38, 1991.
27. Rissanen, J., A universal prior for integers and estimation by minimum
description length, Ann. Statist., 11 (2), 416-431, 1983.
28. Rissanen, J., Universal coding, information, prediction, and estima-
tion, IEEE Trans. Inform. Theory, 30 (4), 629-636, 1984.
29. Rissanen, J., Stochastic Complexity in Statistical Inquiry, World Sci-
entific, Singapore, 1989.
30. Shapiro, J. M., Image coding using the embedded zerotree wavelet
algorithm, in Proc. SPIE Conf. on Mathematical Imaging: Wavelet
324 N. Saito
Applications in Signal and Image Processing, A. F . Laine (ed.), volume

2034, 180-193, 1993.
31. Wallace, R. S., Finding natural clusters through entropy minimization,
P h D thesis, School of Comput. Science, Carnegie Mellon University,
Pittsburgh, PA 15213, Jun. 1989.
32. Wax, M. and T. Kailath, Detection of signals by information theoretic
criteria, IEEE Trans. Acoust., Speech, Signal Processing, ASSP—33
(2), 387-392, 1985.
33. Wickerhauser, M. V., Lectures on wavelet packet algorithms, preprint,
Dept. of Mathematics, Washington University, St. Louis, Missouri,
Nov. 1991.
34. Wickerhauser, M. V., Fast approximate factor analysis, in Proc. SPIE
Conf. on Curves and Surfaces in Computer Vision and Graphics II,
volume 1610, 23-32, 1991.
35. Wickerhauser, M. V., High-resolution still picture compression, Digital
Signal Processing: A Review Journal, 2 (4), 204-226, 1992.
36. Wilson, R., Finite prolate spheroidal sequences and their applications
I: generations and properties, IEEE Trans. Pattern Anal. Machine
Intell., P A M I - 9 (6), 787-795, 1987.
Naoki Saito
Schlumberger-Doll Research
Old Quarry Road, Ridgefield, C T 06877
and
Department of Mathematics
Yale University
10 Hillhouse Avenue, New Haven, C T 06520
e-mail: saito @ridgefield, sdr.sib. com
L o n g - M e m o r y P r o c e s s e s , t h e A l l a n Variance a n d W a v e l e t s
Donald B. Percival and Peter G u t t o r p
Abstract. Long term memory has frequently been observed in physical time
series. Statistical theory for long-memory stochastic processes is radically different
from the standard time series analysis, which assumes short term memory. The
Allan variance is a particular measure of variability developed for long-memory
processes. This variance can be interpreted as a Haar wavelet coefficient variance,
suggesting an approach towards assessing the variability of general wavelet classes.
The theory is applied to a 'time' series of vertical ocean shear measurements for
which some drawbacks with the Haar wavelet are observed.
In a variety of applications, time series analysts have noticed t h a t
the estimated autocovariance sequence for their d a t a tends to decrease
rather slowly, indicating t h a t the series has 'long memory' in the sense t h a t
changes in t h e remote past continue to affect the present value of the se-
ries. Beran (1992) gives a good review of statistical and historical aspects
of long-memory processes. Time series t h a t are well modelled by long-
memory processes have been observed, for example, by Newcomb (1895) in
astronomy, by Gösset (Student, 1927) in chemistry, and by Smith (1938) in
agriculture. In geophysics a famous early example is Hurst's (1951) study of
the minimum annual height of the river Nile. This series has a sample auto-
covariance sequence (acvs) sT t h a t is approximately proportional to | τ | - 0 · 3 ;
i.e., the sequence decays hyperbolically, thus ruling out such s t a n d a r d time
series models as ARMA models t h a t have exponentially decaying autoco-
variance sequences (here sT is an estimate of sT = cov {Xt,Xt+r} when
{Xt} is a stationary process). T h e applicability of long-memory processes
to climate d a t a has been recently discussed by Raftery and Haslett (1989)
and Smith (1992).

ISBN 0-12-262850-0
326 D. Percival and P. Guttorp
T h e statistical properties of a long-memory process can b e quite dif-

ferent from those of a set of independent and identically distributed (iid)
observations. For example, t h e familiar variability properties of sample
averages of iid observations are far from valid for a long-memory process.
In fact, Smith (1938) observed t h a t sample averages in agricultural unifor-
mity trials h a d variances t h a t did not decrease at t h e rate of t h e number of
terms in t h e average, and deduced t h a t there must b e substantial spatial
correlation. To see t h e effect of long memory on t h e variability of averages,
consider a simple example of a long-memory process, namely, a fractional
Gaussian process {Xt} with self-similarity parameter 1/2 < H < 1 (Man-
delbrot and Wallis, 1969; such a process is t h e first difference of t h e frac-
tional Brownian motion defined by Mandelbrot and Van Ness, 1968). By
definition, this process has an acvs given by
,τ = !(|τ+1|2"-2|τ|2" + | τ - ΐ Γ ) , r = ±l,±2,...,
where so = var{X*} (note t h a t the case H = 1/2 corresponds t o t h e iid
case because then sT = 0 for τ φ 0, whereas {Xt} has long-memory if
1/2 < H < 1). Then var {X} = s0N2H~2, where X is t h e sample mean of
N observations (i.e., Σ * = ι Xt/N). Note t h a t , for values of H close t o 1,
the rate of decrease of variability in X is markedly different from t h e 1/N
rate of t h e iid case. Naive application of iid statistics t o t h e sample mean
of a long-memory process can thus be very misleading. For example, if we
fit a fractional Gaussian process to the Nile data, we obtain an estimate
of H = 0.85, implying t h a t the variance of t h e sample mean decreases
like 1/N03 instead of t h e usual 1/N rate. Numerically, obtaining 100
observations from this fractional Gaussian process is equivalent t o obtaining
only 4 observations from an iid process!
In spite of its slow rate of decay, t h e sample average of a long-memory
process is still a surprisingly efficient estimator of t h e mean level of t h e
process (Beran and Künsch, 1985; Percival, 1985). Unfortunately, t h e same
cannot b e said for other standard statistics. In particular, t h e sample
variance is a poor estimator of t h e process variance for a long-memory
process because it has b o t h severe bias and low efficiency (Beran, 1992).
Due t o these problems in estimation, Allan (1966) criticised t h e use of
the sample variance as a meaningful estimator of variability for stationary
processes with long memory and for nonstationary processes with infinite
variance. He proposed an alternative theoretical measure of variability
t h a t is now known as t h e Allan variance (see Section 2 below). In terms
of filtering theory, t h e Allan variance can b e interpreted as t h e variance of
a process after it has been subjected t o an approximate band-pass filter of
'constant Q ' (i.e., t h e ratio of t h e center frequency of t h e pass-band a n d t h e
width of t h e pass-band is a constant). T h e chief advantages of t h e Allan
Long-Memory Processes and Wavelets 327
variance for long-memory processes are two-fold: first, this variance can b e
estimated without bias and with good efficiency for such processes, and,
second, estimates of the Allan variance can in t u r n be used to estimate
t h e parameters of t h e long-memory process (as we note in Section 2, these
statements also hold for certain nonstationary power-law processes). T h e
Allan variance has been applied for over twenty-five years as t h e routine
time domain measure of frequency stability in high-performance oscillators.
In a recent article Flandrin (1992) briefly noted t h a t the Allan variance
can b e interpreted in terms of the coefficients of a Haar wavelet transform
of a time series. We explore this connection in detail in Section 2. As
we note in Section 3, the notion of the Allan variance can be generalized
to other wavelets to define a wavelet variance. T h e wavelet variance is
a useful way of summarizing the properties of the wavelet transform for
certain processes. In particular, the parameters of long-memory processes
can b e deduced from the wavelet variance. We also note in Section 3 t h a t
s t a n d a r d estimators for the Allan variance can be easily generalized to these
other wavelet variances.
In terms of computational complexity, the Allan variance is t h e sim-
plest of the wavelet variances. Unfortunately, it can be misleading for
certain processes of interest in geophysics. In the particular example we
consider in Section 4, we find t h a t the wavelet variance for Daubechies's
'least asymmetric' wavelet of order 8 (Daubechies, 1992) yields markedly
b e t t e r results (hereinafter we refer to this wavelet as the LA(8) wavelet).
This example demonstrates the usefulness for d a t a analysis of classes of
wavelets beyond the simple Haar wavelet. Our analysis also demonstrates
the importance of conducting a parallel spectral analysis to validate t h e
interpretation of a wavelet variance in terms of the parameters for a long-
memory process. Finally, in Section 5 we discuss some possible extensions,
including some general thoughts on assessing the variability of wavelet co-
efficients.
§2. T h e A l l a n Variance a n d t h e H a a r W a v e l e t
Suppose we have a time series of length N t h a t can b e regarded as one
portion of one realization from the stochastic process {YJ,i = 0 , ± 1 , . . . }
(for convenience, we assume t h a t the sampling interval between consecutive
observations is unity). Let
^ « ^ Σ ^ Η (1)
i=o
represent t h e sample average of k consecutive observations, the latest one of
which is Y%. T h e Allan variance at scale k is denoted by σ\ (k) and is defined
to be half t h e mean square difference between adjacent nonoverlapping
Yt(kys; i.e.,
al(k)=\E{[Yt{k)-Yt-k{k)f}.
In order for 0"y(fc) to be independent of the index £, we must impose a
stationarity condition on the process {!*}, namely, t h a t its first backward
difference Zt = Yt — Yt-\ is a stationary process. Note t h a t the Allan
variance at scale k is a measure of how much averages of length k change
from one time period of length k to the next.
To see why the Allan variance might be of interest for long-memory
processes, suppose momentarily t h a t {Yt} is a fractional Gaussian process
with self-similarity parameter 1/2 < H < 1. If we let 5 y ( · ) denote the
spectral density function (sdf) for {Yt} defined for frequencies / between
— 1/2 and 1/2, then we have
Μ/) = Μ/)|/Γ-2H
where Li(·) is a slowly varying function for | / | —>> 0 (Beran, 1992). Thus,
if we plot log(5y ( / ) ) versus log(/) for positive values of / close to zero, we
will observe (to a good approximation) a line with a slope of 1 — 2H. A
similar result holds for the Allan variance: we have
a2Y(k) = L2(k)k2H~2,
where L 2 O is a slowly varying function for k —l· oo (Percival, 1983, The-

orem 2.19). Thus, if we plot log(a y (fc)) versus log(fc) for large fc, we will
observe (to a good approximation) a line with a slope of 2H — 2. In the
analysis of frequency stability, it is common practice to produce plots of the
so-called 'σ-τ curve,' which is just a plot of an estimate of log^y(fc)) versus
log(fc). On such a plot, a slope of β with —1/2 < β < 0 would be indicative
of a stationary process with a self-similarity parameter H = ß -f 1.
As we mentioned above, the Allan variance is also well-defined when
{Yt} is not itself a stationary process but its associated first backward
difference {Zt} is. In such cases, we can define an sdf 5 y ( · ) for {Yt} in
terms of the sdf Sz{') for {Zt} via the relationship
Sy(f) Ξ 4sin
JgP-2
(7r/)
(this definition is motivated by the theory of linear filters). Suppose mo-
mentarily t h a t
Sy(f) = L3(f)\f\°,
where Ls(·) is a slowly varying function for | / | -> 0, and —3 < a < 0; i.e.,
SY(-) is a 'red' power-law process with exponent a greater t h a n —3. T h e n
we have
al(k) = L4(k)k~a-\
where L±(') is a slowly varying function for k —► oo. Thus, in principle,

the Allan variance can be used to deduce the parameters for certain 'red'
power-law processes.
Let us now consider the Haar wavelet transform of the time series Y\,
. . . , YJV , where we now assume that the sample size TV is a power of 2 so that
N = 2P for some positive integer p. By definition, this transform consists
of N — 1 'detail' coefficients and one 'smooth' coefficient s\ = Y2t=l Yt/N.
The detail coefficients djjk are defined for scales k = 1, 2, 4, . . . , N/2 and
- within the kth scale - for indices j = 1, 2, 3, . . . , N/2k as
1 p-i fc-i "I
dj k Y k l Y
> - ~frt Σ ^ ~ " Σ ^Jk-k-i . (2)
VZK
li=o i=o J
For example, we have
dj,! = [Y2j - i2i-i]/v/2, l<j<iV/2
d jl2 s [r4,· + n , - - i - Ytj-2 - YAJ-3]/2, l<J<N/4
djA = [Y8j + ---+Y8j-3-Ysj-4 Yaj-7)/2^2, l<j<N/8
djlN/2 = [YN + --- + YN/2+I-YN/2 Yi]/(V2)p.

We can now state the relationship between the Allan variance and the
Haar wavelet coefficients dj^. Using Equations 1 and 2, we have
<***=(£) [Y2jk(k)-Y2jk-k(k)].
If the first backward difference process {Zt} for {Yt} is stationary with zero
mean, we then have
var{d i)fc } = E{d\k) = Ê{[Y2jk(k)-Y2jk_k(k)]2} = ka2Y(k). (3)

Thus the Allan variance at scale k is directly related to the variance of the
Haar wavelet coefficient at that same scale. (Violation of the seemingly
innocuous assumption that E{Zt} = 0 can seriously impact our ability
to make sense of estimates of the Allan variance - see Percival, 1983, for
further discussion.)
In view of Equation 3, an obvious unbiased estimator for the Allan
variance is
N/2k N/2k
d
ϊγ^ = jf Σ h =^ Σ Fwik) - Y2jk-k(k)}*. (4)
j=l j=l
The above estimator is known in the frequency stability literature as the

'non-overlapped' estimator of the Allan variance because each value Yt in
the time series contributes to exactly one d2- k in the summation in Equa-
tion 4. T h e estimator <Ty(&) is not commonly used because of the superior
statistical properties of a closely related estimator known as the maximal-
overlap estimator (Greenhall, 1991). This estimator is defined as
**<*> Ξ 2(N-\k+l) Σ (F<(fc) - F - ( f c ) ) 2 · (*)

We can interpret the above estimator in terms of a filtering operation from
which we can obtain the Haar wavelet coefficients dj^ by subsampling
(also called 'downsampling'). At scale &, the filter t h a t yields the dj^s is
of length 2k and has coefficients
7 (l/y/2k, Z = 0,...,k-1;
Z,fc
~ \-l/>/2fc, Z = fe, . . . , 2 f e - l
(cf. Equation 2). Note t h a t Σζ=ό" "f k = 1. If we let

2k-l
Wt,k=^ïn,kYt-i, f = 2k,...,iV, (6)
1=0
then we have djtk — W2jk,kl i-e-> the drib's are obtained by appropriately
subsampling every 2/cth value of the Wife's. We refer to the Wife's as
the maximal-overlap Haar wavelet coefficients. We can now easily see the
difference between the non-overlapped and maximal-overlap estimators:
N / 2 k N
9 ~ 1 ~
? whereas fc w
v(fc) = Ü Σ ^ Μ > *y( ) = fe(^_2fc + 1) Σ *%; (7)
i.e., ôr y (k) makes use of just the N/2k subsampled Wt,k% whereas <5"y(fc)
makes use of all N — 2k -\- 1 of the Wtfk's. Thus the maximal-overlap
estimator does not use the usual decimated subseries of the discrete Haar
wavelet transform, but it can be said to use a uniformly sampled version
of the corresponding continuous transform. Note t h a t use of σ\ (k) r a t h e r
t h a n ^(k) imparts a certain degree of independence in the choice of the
origin (i.e., if we form a new time series by discarding Y\ and adding Y/v+i,
only one of the Wife's in the maximal-overlap estimator changes whereas
all of t h e m change in the nonoverlapped estimator). Note also t h a t the
definition of σ\(k) in Equation 5 holds for all sample sizes N (i.e., N need
not be a power of 2).
For computational purposes, it is more efficient to evaluate the right-
h a n d side of Equation 5 by computing the summations needed to obtain
t h e y t (fc)'s just once. To do so, let XQ = 0, and let Xt = Xt-i + Y% for
t = 1, . . . , N. Since Xt = Σ ω = 1 Υμ, we have
Yt{k) = -Y,Yt-j = {Xt-Xt-k)/k,

j=0
and hence we have
Yt(k) -Yt-k(k) = (Xt - 2Xt-k+Xt-2k)/k
and
Wtik = (Xt - 2Xt-k + Xt-2k)/V2k.
We can thus rewrite Equation 5 as
*'<*> = 2kHN-2k+l) t < * - 2 * - *+ * - » ) ' ·
§3. T h e W a v e l e t Variance
It is easy t o extend t h e ideas of t h e previous section t o wavelets other
t h a n t h e Haar wavelet. There are two approaches t h a t we can take t o gen-
erate t h e maximal-overlap wavelet coefficients Wt,k based upon a wavelet
of unit scale specified by t h e L\ filter coefficients /ΐο,ι, /ΐι,ι, . . . , /ΐχ,ι-ι,ι-
In t h e first approach, we start by generating t h e appropriate filter {h^k}
for each scale k based upon t h e wavelet for unit scale. This can b e done
easily by taking t h e inverse discrete wavelet transform of a properly placed
pulse (for details, see t h e subsection ' W h a t Do Wavelets Look Like?' in
Section 13.10 of Press et al., 1992). Let Lk b e t h e length of t h e wavelet
of scale k so t h a t , for example, L\ = 2 for t h e Haar wavelet while L\ = 8
for t h e LA(8) wavelet (the values of Lk for scales k = 2, 4, . . . , obey t h e
relationship Lk = (2k — l ) ( I q — 1) + 1 and can also b e obtained via t h e
recursive formula L2k — 2Lk + L\ — 2). For each scale k we then filter our
time series t o obtain
Lk-\
Wt,k= Σ Kkyt-h t= Lk,...,N
1=0
(cf. Equation 6).

T h e second approach t o obtaining t h e Wife's is t o modify t h e pyramid
algorithm for obtaining t h e usual wavelet coefficients (see Section 13.10 of
Press et a/., 1992, for a lucid description of this algorithm). T h e usual pyra-
mid algorithm makes use of two filters, namely, t h e wavelet filter {^z,i} a n d
the so-called scaling filter {<7/,i}. For t h e types of wavelets of interest here,
t h e scaling filter is defined in terms of t h e wavelet filter via t h e ' q u a d r a t u r e
mirror' relationship gî = ( — l ) / + 1 / i L i - / - i , i · Whereas t h e wavelet filter
resembles a high-pass filter, t h e scaling filter resembles a low-pass filter.
T h e usual pyramid algorithm uses the same scaling and wavelet filters at
each scale; i.e., at scale 2fc, these two filters are applied to the subsampled
o u t p u t from t h e scaling (low-pass) filter from scale k. For the maximal-
overlap estimator we must eliminate all subsampling, so formally we must
use different filters as we move from scale k to scale 2k. Forming these new
filters, however, is quite simple: the wavelet and scaling filters we need for
scale k in t h e maximal-overlap pyramid algorithm are obtained by insert-
ing k — 1 zeros between each of the coefficients in the wavelet and scaling
filters for unit scale (in effect, these zeros compensate for the elimination of
subsampling). T h e maximal-overlap pyramid algorithm avoids all multipli-
cations involving coefficients equal to zero by keeping track of the indices
of elements of the series t h a t need to be multiplied by the nonzero filter
coefficients.
Given the wavelet filter {7i/,i} and corresponding scaling filter {37,1}
for unit scale, the basic step of the maximal-overlap pyramid algorithm
takes as input
• a scale fc, which must be an integer power of 2, and
• a series X i , X25 · · · 5 ^Mk of length M* = N — (k — l)(Li — 1), where

N is t h e length of the time series (this need not be a power of 2);
and returns as o u t p u t two series of length M2A;, namely,
• a low-pass series X\ , X%\ · · ·, X)J and
• a high-pass series Λ # 1 Μ „ + 1 , X^l M 2 f c +2> ■ ■ · > Xp·

Given its two inputs k and {X*}, the basic step creates its two o u t p u t series
as follows.
1. Set m = fc(Li - 1) + 1.
2. For t = m, . . . , Μ&, do the outer loop of 2:
SetXt(Îîm+1=flb,iXt.
Set XN_Mh+t — ho,ixt·
Set u = t.
For / = 1 to L\ — 1, do the inner loop of 2:
Decrement u by k.
Increment Xt_m+i by gi,iXu>
Increment Xx_Mk+t by hlAXu.
E n d of the inner loop of 2.
E n d of t h e outer loop of 2.
W i t h t h e basic step so defined, t h e maximal-overlap pyramid algorithm

consists of t h e following steps.
1. Evoke t h e basic step with scale k = 1 and Xt = Yt for t = 1, . . . ,

M i = N t o obtain as o u t p u t t h e low-pass series X\ a n d high-pass
series x\h\ b o t h of length M 2 = N - Lx + 1. Set Vtyl = X{tl) a n d
Wt}\ = Xt i i-e-> t h e maximal-overlap wavelet coefficients of scale
1 are just t h e elements of t h e high-pass series.
2. Evoke t h e basic step again, b u t this time with scale k = 2 a n d

Xt = Vtti for t = 1, . . . , M-2 t o obtain as o u t p u t t h e low-pass series
X[l) and high-pass series x[h\ b o t h of length M 4 = TV - 3Li + 3.
Set Vti2 = Xt a n d Wtt2 = Xt i i-e-5 t h e maximal-overlap wavelet
coefficients of scale 2 are just t h e elements of t h e high-pass series.
3. Evoke t h e basic step for k = 4, 8, . . . , 2^ l o S2([^+î- 2 ]/î- 1 ])J- 1 5

where [#J refers t o t h e largest integer t h a t is less t h a n or equal t o
x. In particular, at t h e step in which t h e maximal-overlap wavelet
coefficients for scale k are calculated, t h e basic step is evoked with
Xt = Vttk/2 for t = 1, . . . , Mk = N - (k - l ) ( L i - 1) t o obtain as
o u t p u t t h e low-pass series X\ and high-pass series X\ , b o t h of
length M2k = N-(2k-l)(Ll-l). Set Vt,k = x\l) a n d Wt%k = Χ[Η);
i.e., t h e maximal-overlap wavelet coefficients of scale k are just t h e
elements of t h e high-pass series.
Note t h a t , in contrast t o t h e Haar wavelet, we can obtain only some of

t h e wavelet coefficients by subsampling t h e Wt,k'& - t h e missing coefficients
are those t h a t involve circularly wrapping t h e time series (see t h e discussion
in Press et a/., 1992). Under t h e same assumptions on {Yt} t h a t we used
to define t h e Allan variance, we can define t h e wavelet variance at scale k
as
vï(k) = E{WÏik}/k
(cf. Equation 3). T h e equivalent of t h e maximal-overlap estimator for this

wavelet variance would b e given by
N
1
ù W
r(k) = 771Γ7-Τ TT Y h
YK J
k(N - Lk + 1) £f *'*
(cf. Equation 7). A plot of t h e square root of this quantity versus scale
A: on a log/log scale yields a generalization of t h e σ-τ curve, from which
we can infer from regions of linearity t h a t a power law process might b e a
good model for our data.
350 450 550 650 750 850 950 1050

meters
Figure 1. Plot of vertical shear measurements (in inverse seconds) ver-
sus depth (in meters). This series was collected and supplied by Mike
Gregg, Applied Physics Laboratory, University of Washington. As of
1994, this series could be obtained via electronic mail by sending a mes-
sage with the single line 'send lmpavw from datas et s ' to the Internet
address s t a t l i b Q l i b . s t a t . c m u . e d u - this is the address for StatLib, a
statistical archive maintained by Carnegie Mellon University.
§4. A n A p p l i c a t i o n t o Vertical Shear M e a s u r e m e n t s

Here we illustrate the ideas of the previous section by examining a
'time' series of vertical ocean shear measurements. The d a t a were collected
by an instrument t h a t is dropped over the side of a ship and designed to
then descend vertically into the ocean. As it descends, the probe collects
measurements concerning the ocean as a function of depth. T h e ordering
variable of our 'time' series is thus depth. One of the measurements is
the x component of the velocity of water. This velocity is collected every
0.1 meters, first differenced over an interval of 10 meters, and then low-pass
filtered to obtain a series related to vertical shear in the ocean. Vertical
shear is thought to obey a power-law process over certain ranges of spatial
frequency, so this series is a useful candidate for examining how well t h e
Allan variance and other wavelet variances can deduce such a process.
Figure 1 shows the series of vertical shear measurements used in this
study. T h e series extends from a depth of 350.0 meters down to 1037.4 me-
ters in increments of 0.1 meters (there are 6875 d a t a values in all). There
are two thin vertical lines marked on the plot, between which there are 4096
values ranging from 489.5 meters to 899.0 meter. In what follows, we will
assume t h a t this subseries can be regarded as a portion of one realization of
a process whose first backward difference is a stationary process (we need
this assumption to apply meaningfully the methodology of Section 3).
T h e thick curves in the four rows of plots in Figure 2 show, respectively,
the squared modulus of the transfer function for the Haar wavelet (left-hand
1h
0.0 0.5 0.0 0.5

f f
Figure 2. Squared modulus of transfer functions (left-hand column of
plots) and phase functions (right-hand column) versus frequency for the
Haar wavelet (thick curves) and LA(8) wavelet (thin curves) for scales
1, 2, 4 and 8 (top to bottom rows).
column) and t h e associated phase function over the nominal pass-band of

the wavelet (right-hand column) for the four scales 1, 2, 4 and 8 (top
to b o t t o m rows). T h e thin curves in each plot show the corresponding
functions for the LA(8) wavelet, which has the following filter coefficients:
ho.i = 0.03222310060407815 h4A = 0.80373875180538600
hiji = 0.01260396726226383 h 5 |i = -0.49761866763256290
h2[i = -0.09921954357695636 h6[i = -0.02963552764596039
Η3Λ = -0.29785779560560505 h7[i = 0.07576571478935668
(the source of these coefficients is the 'N = 4' entry of Table 6.3, p. 198,
Daubechies, 1992, which gives the LA(8) scaling filter coefficients normal-
ized to sum to 2 - the above hi y s were obtained via the 'quadrature mirror'
relationship between scaling and wavelet filters with a renormalization so
t h a t Σ / = ο ^ ? ι = -0· Note t h a t the transfer functions for b o t h wavelets
roughly define a set of octave band filters; i.e., the transfer functions for
scale k are approximately concentrated between frequencies and l/4k and
l/2k (the spacing between the minor tick marks on the frequency axis is
1/16). T h e plots show t h a t the transfer functions for the LA(8) wavelet
are a b e t t e r approximation to a set of octave band filters t h a n those for t h e
Haar wavelet. T h e phase functions for the Haar wavelet are approximately
linear over the nominal pass-bands, whereas those for the LA(8) wavelet
are approximately constant and fairly close to zero. Thus the o u t p u t from
the Haar wavelet filters will be phase-shifted with respect to the input,
making it difficult to line up events at various scales with the original time
series. In contrast, because the filters for the LA(8) wavelet are approxi-
mately zero phase, we can more easily line up events at various scales with
the original time series. Note, however, t h a t the spans of Haar wavelet
filters for scales 1, 2, 4 and 8 are, respectively, 2, 4, 8 and 16, whereas
the corresponding spans for the LA(8) wavelet are 8, 22, 50 and 106. If
these wavelets are used as noncircular filters as discussed in Section 3, the
o u t p u t from the LA(8) wavelet will become increasingly shorter compared
to t h a t of the Haar wavelet as the scale increases. In fact, for the larger
scales the length of the LA(8) wavelet will exceed the length of the time
series, whereas the reverse will be true for the Haar wavelet.
Figure 3 shows (from b o t t o m to top) the renormalized o u t p u t s from
the Haar wavelet filters for physical scales 0.1, 0.2, 0.4, 0.8, 1.6 and 3.2
meters (because the distance between adjacent observations is 0.1 meters,
these physical scales correspond to, respectively, the unitless scales 1, 2,
4, 8, 16 and 32). In order to obtain physically meaningful units, it is
necessary to renormalize the Wife's by dividing by (2/c) 1 / 2 . Each of these
renormalized filtered series is drawn with the same vertical scale, so we
see t h a t the variability gets progressively larger as we move from shorter
to longer scales (the distance between minor tick marks on the vertical
scale is 1/s). T h e usual Haar wavelet transform for these scales can be
obtained by appropriately subsampling these filtered series. Note t h a t all
six of these filtered series appear to be approximately rescaled versions of
ΙΙ^ηΚΚ^
yvM^Jw^^ yVv#-
TO^^^^*>iH^ r*> » » « ^ » W ^ »»w"^Wa"WllVVv *WMh|'v|l'«'«%p>Ίν V
L M ^ ^ - N W ! mi^)/ii|Hp^nf * «Mt>»'WI>*Kllt>|i"'i i l l « ' * «P*«HI«I%II»H«I1.HIII>IH'«I.III '«M»*) i |i » yi
_L X I X J
350 450 550 650 750 850 950 1050
meters
Figure 3. Renormalized outputs of Haar wavelet filters for physical
scales 0.1, 0.2, 0.4, 0.8, 1.6 and 3.2 meters (bottom to top).
each other. We will see below t h a t this is caused by a form of 'leakage'

evidently due t o t h e fact t h a t t h e Haar approximation t o pass-band filters
is not good enough for this series and t h a t in effect longer scale fluctuations
are 'leaking' into these shorter scale fluctuations. Note t h a t there are two
bursts in t h e filtered series for t h e smaller scales, one near 450 meters a n d
the other near 1000 meters.
Figure 4 is a repetition of Figure 3, with t h e LA(8) wavelet used instead
of t h e H a a r wavelet (again, we have renormalized t h e Wife's by dividing by
(2&) 1 / 2 ). T h e vertical scales on Figures 3 a n d 4 are identical, so it is evident
\^f^^jt^Hêjjl^
^'■"»"«flNi»
Tf|r~
■ »#» >WH"* iH ♦»« H"»fr»»M i»«»1
" » » M Ni '■«H " Νι.ι»ιΜρφ..ι H"""-»"*' mp" Φ
J_ _]_ X X _L X J
350 450 550 650 750 850 950 1050
meters
Figure 4. Renormalized outputs of LA(8) wavelet filters for physical
scales 0.1, 0.2, 0.4, 0.8, 1.6 and 3.2 meters (bottom to top).
that there is much less variability at the smaller scales for the LA(8) wavelet
than for the Haar wavelet. While there is some correspondence between the
filtered series at different scales, the correspondence is much less marked
for the LA(8) wavelet than for the Haar wavelet. Each of the filtered series
in this figure is slightly shorter than the corresponding one in Figure 3 due
to the longer span of the filters for the LA(8) wavelet. Note that only part
of the usual LA(8) wavelet transform for these scales can be obtained by
subsampling these filtered series because we are not filtering the data in
Figure 1 as if it were circular (an assumption that would make no sense
10 1
10°
io- 1
■ ■
io- 2 J I I I ' i l l
2 1 1 2 3 2 1 1
10- 10- 10° 10 10 10 10" 10" 10° 10 10 2 1 0 3
scale (meters) scale (meters)
F i g u r e 5. 'σ-τ' curves for the Haar wavelet (left-hand plot) and the
LA(8) wavelet (right-hand).
at all for a series of ocean depth measurements!). Again we can see the
bursts in t h e filtered series for the smaller scales near 450 and 1000 meters.
An examination of the renormalized filtered series for physical scales longer
t h a n 3.2 meters shows no indication of these bursts. In the region marked
by t h e thin vertical lines between the two bursts, the filtered series for all
scales appear to be consistent with a stationarity assumption, so we have
chosen this subseries of 4096 values as a candidate for analysis using the
ideas discussed in the previous two sections.
T h e left-hand plot of Figure 5 shows the ' σ - τ ' curve, a popular analysis
tool in the frequency stability literature. This plot shows the square root
of t h e Allan variance (estimated using the maximal-overlap estimator) at
different physical scales plotted versus scale on a log-log plot. Theory
suggests t h a t regions of linearity correspond to a power-law process over a
particular region of frequencies, with the exponent of the power-law process
being related to the slope of the line (in log-log space). T h e first 7 values of
the σ-τ curve fall on such a line almost perfectly. The line drawn through
t h e m on t h e plot was calculated via least squares and has a slope of 0.83.
Since ay{k) (the square root of the Allan variance) varies approximately as
£-(c*+i)/2 for a p 0 w e r _ i a w process with exponent a (see Section 2), the σ-τ
curve strongly suggests the presence of a power-law process over scales of
0.1 to 6.4 meters with an exponent of a = —2.66 = —8/3. T h e right-hand
plot is t h e σ-τ curve corresponding to the LA(8) wavelet for the first 9
scales (recall t h a t filtered series for the longer scales cannot be obtained
due to t h e span of the filters for this wavelet), and it tells quite a different
story. T h e values are not aligned in an obvious straight line as in t h e case
of t h e Allan variance, and the slope t h a t we found using the first 7 scales
of t h e Allan variance certainly does not look reasonable for portions of t h e
340 1D. Pe rciva / and ' P . (jUttOJ
io 3 r
102[
101 I
10" h
10-4
10-4
<, ' "■
10-4 ■Λν.
• 1 .
i 1 1
" " . X'

10"4
·' ,1.1
10"4 ■ V
10-4
io-7 _ L _ L J
10": 10"2 10"1 10° 101
f (cycles/meter)
F i g u r e 6. Comparison of spectral estimates implied by σ-τ curves
with a WOSA spectral estimate.
LA(8) wavelet variance.

In Figure 6 we translate the σ-τ curves of the previous figure into es-
timates of t h e sdf over different octave bands (these have constant spacing
on a log frequency scale). For example, the wavelet variance for a scale of
0.1 meters maps onto t h e highest octave band; the one for a scale of 0.2 me-
ters maps onto t h e second highest octave band; and so forth. T h e thick
'staircase' on this plot corresponds to the Allan variance, while t h e thin
staircase corresponds to the LA(8) wavelet variance. In general, t h e LA(8)
wavelet variance yields an sdf estimate t h a t is lower t h a n the estimate

obtained from the Haar wavelet (i.e., the Allan variance). T h e difference
between the two sdf estimates is almost one order of magnitude in three
of the octave bands. The dots in this figure show a Welch's overlapped
segment averaging (WOSA) sdf estimate using a Hanning d a t a t a p e r and
a block size of 1024 d a t a points with segments overlapping by 50%, yield-
ing an sdf estimate with 13.4 equivalent degrees of freedom (a comparison
of the periodogram with direct sdf estimates using nontrivial d a t a tapers
indicated t h a t the periodogram suffers from leakage and t h a t tapering is
required here; for details on the WOSA sdf estimate, see, for example,
Section 6.17 of Percival and Waiden, 1993). Whereas the LA(8) wavelet
sdf estimate agrees fairly well with the WOSA spectral estimate, the Haar
wavelet sdf estimate is too high in several of the octave bands, indicating
some form of leakage evidently due to the fact t h a t the Haar wavelet forms
a fairly crude set of octave band filters.
§5. C o n c l u d i n g R e m a r k s
Here we make several remarks concerning the results of the previous
sections. First, a wavelet analysis of a time series t h a t can be modelled as
a power-law process can be used to deduce the properties of the underlying
process, b u t it is dangerous to do so without a careful look at a traditional
sdf estimate with good prevention against leakage. For example, if we had
computed just the left-hand σ-τ curve of Figure 5, we might have been
badly fooled by the degree to which the points line up as theory suggests
they should in the presence of a power-law process.
Second, as others have noted, the wavelet variance is a tool t h a t is
well-adapted for studying power-law processes; however, the wavelet vari-
ance can lead us to find 'power laws' (or 'fractal behavior') in d a t a t h a t
might be best modelled in other ways. Consider, for example, the octave
b a n d centered at 1 0 _ 1 cycles/meter in Figure 6. T h e wavelet-based esti-
mates of t h e power in this band are b o t h somewhat higher t h a n what the
WOSA sdf estimate suggests is reasonable (in fact, the deficiency of power
here can be a t t r i b u t e d to the preprocessing operation in which t h e veloc-
ity series was first differenced over an interval of 10 meters). This octave
band corresponds to a scale of 3.2 meters in the σ-τ curves of Figure 5.
For b o t h of these curves, the value at 3.2 meters can be obtained almost
exactly by linearly interpolating (on a log/log scale) between the points at
scales 1.6 and 6.4 meters. This suggests t h a t the wavelet-based estimates
might b e biased in the sense t h a t regions t h a t do not agree with a nominal
power-law behavior will tend to be 'filled in' in a manner consistent with
the power-law assumption. W h a t is vitally needed is a careful study of the
bias and variance of wavelet-based estimates of the sdf for processes t h a t
deviate from a power law in at least some octave bands.

Third, for purposes of estimating the wavelet variance, results obtained
by the frequency stability community using the Haar wavelet (i.e., the Allan
variance) strongly favor the maximal-overlap estimator. T h e superiority of
the maximal-overlap estimator is due to the fact t h a t it is formed by taking
the o u t p u t from a filter, summing the square of each term and then dividing
by the number of terms times the scale &, whereas the non-overlapped esti-
m a t o r takes exactly the same filter output, subsamples it, sums the square
of each subsampled term and then divides by the number of subsampled
terms times the scale k. It is thus intuitively reasonable t h a t the subsam-
pling operation yields an estimator of the Haar wavelet variance with larger
variance t h a n the nonoverlapped estimator (detailed analysis of t h e two es-
timators supports our intuition - see Greenhall, 1991). While it seems
obvious t h a t this result holds at least to some extent for wavelets other
t h a n the Haar wavelet, more work is needed to verify t h a t the maximal-
overlap estimator is indeed superior to the nonoverlapped estimator t h a t
would naturally fall out from the discrete wavelet transform (if we ignore
those parts of the transform t h a t correspond to circularly filtering a time
series).
Fourth, in spite of its poor properties at small scales, the Haar wavelet
seems to do quite well at long scales where in fact we cannot use the LA(8)
wavelet at all (in a noncircular fashion) because the lengths of the wavelet
and scaling filters are longer t h a n our time series.
Fifth, procedures have been worked out in the frequency stability com-
munity for placing confidence limits on the Allan variance under the fairly
stringent assumption t h a t the exponent of the power law is known a pri-
ori (see Greenhall, 1991, for details). In principle, we can obtain similar
confidence intervals for other wavelet variances. Also, by defining an esti-
m a t o r for the wavelet variance starting from a prewhitened direct spectral
estimate, we can in fact easily obtain statistically valid confidence intervals
based upon the sampling properties of direct spectral estimates. (Such an
estimator would also allow us to get a handle on the covariance between
estimates at different scales, a problem t h a t has yet to be addressed even
in the case of the Allan variance.)
Finally, Figure 4 points out t h a t the real potential value of wavelets is
in the area of time series with transient events. In the case of vertical shear
measurements, wavelet analysis shows t h a t these transients happen at just
a few small scales. Figure 4 shows t h a t the transient near 450 meters is
confined chiefly to scales of 0.8 meters and smaller and, moreover, t h a t t h e
interval of time over which the transient dissipates is shorter for a scale of
0.8 meters t h a n it is for smaller scales. These features in the d a t a would
be very hard to pick out using other analysis techniques.
References
1. Allan, D. W., Statistics of atomic frequency clocks, Proc. IEEE, 31,
221-30, 1966.
2. Beran, J., Statistical methods for data with long-range dependence,
Statist. Sei., 7, 404-16, 1992.
3. Beran, J. and H. R. Künsch, Location estimators for processes with
long-range dependence, Research Report 40, Seminar für Statistik,
ETH, Zürich, 1985.
4. Daubechies, I., Ten lectures on wavelets, SI AM, Philadelphia, 1992.
5. Flandrin, P., Wavelet analysis and synthesis of fractional brownian
motion, IEEE Trans. Info. Theo., 38, 910-7, 1992.
6. Greenhall, C. A., Recipes for degrees of freedom of frequency stability
estimators, IEEE Trans. Instr. Meas., 40, 994-9, 1991.
7. Hurst, H. E., Long-term storage capacity of reservoirs, Trans. Amer.
Soc. Civ. Eng., 116, 770-9, 1951.
8. Mandelbrot, B. B. and J. W. Van Ness, Fractional Brownian motions,
fractional noises and applications, SI AM Rev., 10, 422-37, 1968.
9. Mandelbrot, B. B. and J. R. Wallis, Computer experiments with frac-
tional Gaussian noises. Part 1, averages and variances, Water Resour.
Resear., 5, 228-41, 1969.
10. Newcomb, S., Astronomical constants (the elements of the four inner
planets and the fundamental constants of astronomy), Supplement to
the American Ephemeris and Nautical Almanac for 1897, U.S. Govern-
ment Printing Office, Washington, 1895.
11. Percival, D. B., The statistics of long memory processes, unpublished
Ph.D. dissertation, Department of Statistics, University of Washington,
1983.
12. Percival, D. B., On the sample mean and variance of a long memory
process, Technical Report 69, Department of Statistics, University of
Washington, 1985.
13. Percival, D. B. and A. T. Waiden, Spectral analysis for physical appli-
cations: multitaper and conventional univariate techniques, Cambridge
University Press, Cambridge, 1993.
14. Press, W. H., B. P. Flannery, S. A. Teukolsky and W. T. Vetterling, Nu-
merical recipes: the art of scientific computing (second edition), Cam-
bridge University Press, Cambridge, 1992.
15. Raftery, A. E. and J. Haslett, Space-time modelling with long-memory
dependence: assessing Ireland's wind power resource (with discussion),
J. Roy. Statist. Soc, Ser. C, 38, 1-21, 1989.
16. Smith, H. F., An empirical law describing heterogeneity in the yields
of agricultural crops, J. Agri. Sei., 28, 1-23, 1938.
17. Smith, R. L., Comment [on Beran, 1992], Statist. Sei., 7, 422-5, 1992.
18. Student, Errors in Routine Analysis, Biometrika, 19, 151-64, 1927.
Robert Spindel and John Harlett of the Applied Physics Laboratory, Uni-
versity of Washington, graciously provided discretionary funding to support
Percival's work on this manuscript. Guttorp's work was partially funded
by NSF grant DM S-9115756. The authors wish to thank Mike Gregg for
supplying the data used in Section 4 and also to thank Chuck Greenhall
and Emma McCoy for their helpful critiques.
Donald B. Percival
Applied Physics Laboratory
HN-10
University of Washington
Seattle, WA 98195
Peter Guttorp
Department of Statistics
GN-22
University of Washington
Seattle, WA 98195
e-mail: peterùstat. Washington, edu
Bibliography
1. Akujuobi, C. M. and A. Z. Baraniecki, Wavelets and fractals:

Overview of their similarities based on different application areas,
Proc. of IEEE-SP Int'l. Symposium on Time-Frequency and Time-
Scale Analysis, Victoria, 197-200, 1992.
2. Argoul, F., A. Arnéodo, J. Elezgaray, and G. Grasseau, Wavelet anal-
ysis of fractal growth processes, Proc. of 4th EPS Liquid State Conf,
Arcachon, France, 1988.
3. Argoul, F., A. Arnéodo, J. Elezgaray, G. Grasseau, and R. Murenzi,
Wavelet transform of fractal aggregates, Phys. Lett. A, 135, 327-336,
1989.
4. Argoul, F., A. Arnéodo, G. Grasseau, Y. Gagne, E. J.Hopfinger, and
U. Frisch, Wavelet analysis of turbulence reveals the multifractal na-
ture of the Richardson cascade, Nature, 338, 51-53, 1989.
5. Argoul, F., A. Arnéodo, J. Elezgaray, G. Grasseau, and R. Murenzi,
Wavelet analysis of the self-similarity of diffusion-limited aggregates
and electro deposition clusters, Physical Review A, 41(10), 5537-5560,
1990.
6. Arnéodo, A., F. Argoul, J. Elezgaray, and G. Grasseau, Wavelet trans-
form analysis of fractals: Application to nonequilibrium phase tran-
sitions, in Nonlinear Dynamics, G. Turchetti (ed.), World Scientific,
New Jersey, 130-180, 1988.
of multifractals, Phys. Rev. Lett, 61, 2281-2284, 1988.
analysis of some dynamical systems, in Wavelets, J. M. Combes, A.
Grossmann, and P. Tchamitchian (eds.), Springer-Verlag, Berlin,, 315
pp., 1989.
9. Arnéodo, A., E. Bacry, and J. F. Muzy, Wavelet analysis of fractal sig-
nals - direct determination of the singularity spectrum of fully devel-
oped turbulence data, in Wavelets and Turbulence, Springer-Verlag,
1991.
10. Arnéodo, A., E. Bacry, J. F. Muzy, Wavelets and multifractal formal-
ism for singular signals: Applications to turbulence data, preprint,
345
346 Bibliography
Centre de Recherche Paul Pascal, Pessac, 1991.

11. Arneodo, A., F. Argoul, E. Bacry, J. Elezgaray, E. Freysz, G.
Grasseau, J. F. Muzy, and B. Pouligny, Wavelet transform of fractals:
I. From the transition to chaos to fully developed turbulence. II. Op-
tical wavelet transform of fractal growth phenomena, in Wavelets and
Some of Their Applications, Y. Meyer et al. (ed.), Springer-Verlag,
1991.
12. Arnéodo, A., F. Argoul, E. Bacry, J. F. Muzy, and M. Tabard, Golden
mean arithmetic in the fractal branching of diffusion-limited aggre-
gates, Phys. Rev. Lett, 6 8 ( 2 3 ) , 3456-3459, 1992.
13. Arnéodo, A., J. F. Muzy, and E. Bacry, Wavelet analysis of fractal sig-
nals. Applications to fully developed turbulence data, IUTAM Symp.
on Eddy Structure Identification in Free Turbulent Shear Flows, Poi-
ters, October, 1992.
14. Arnéodo, A., F. Argoul, E. Bacry, J. Elezgaray, E. Freysz, G.
Grasseau, J. F. Muzy, and B. Pouligny, 1. Wavelet transform of frac-
tals: from the transition to chaos to fully developed turbulence. 2.
Optical wavelet transform of fractal growth phenomena, in Wavelets
and Applications, Y. Meyer (ed.), Masson, Paris and Springer, Berlin,
1992.
15. Bacry, E., A. Arneodo, U. Frisch, Y. Gagne, and E. Hopfinger,
Wavelet analysis of fully developed turbulence d a t a and measure-
ment of scaling exponents, in Turbulence and Coherent Structures, 0 .
Métais and M. Lesieur (eds.), Kluwer Academic Publishers, 203-215,
1991.
16. Bacry, E., J. F. Muzy, and A. Arneodo, Singularity spectrum of fractal
signals from wavelet analysis: Exact results, J. Stat. Phys., 70, 6 3 5 -
674, 1993.
17. Barbaresco, F., Time-frequency analysis of a complex radar signal by
a fast recursive processing of the wavelet transform for detection of
moving objects in clutters, in Proc. of IEEE-SP Int'L Symposium on
Time-Frequency and Time-Scale Analysis, Victoria, 505-508, 1992.
18. Barrodale, I., N. R. Chapman, and C. A. Zala, Estimation of bubble
pulse wavelets for deconvolution of marine seismograms, Geophys. J.
Roy. Astron. Soc, to appear, 1993.
19. Benzi, R. and M. Vergassola, Optimal wavelet analysis and its appli-
cation to two-dimensional turbulence, Fluid Dynamics Research, 8,
117-126, 1991.
20. Berger, M. A., Random affine iterated function systems - curve gen-
eration and wavelets, preprint, School of Mathematics, Georgia Inst.
of Technology, Atlanta, 38 p., 1991.
21. Berger, M. A., IFS Algorithms for wavelet transforms, curves and
surfaces, and image compression, preprint, School of Mathematics,
Bibliography 347
Georgia Inst, of Technology, Atlanta, 12 p.,

22. Berger, M.A. and Y. Wang, Multi-scale dilation equations and iter-
ated function systems, preprint, School of Mathematics, Georgia Inst.
of Technology, Atlanta, 30 p., 1991.
23. Bergstrom, H., and U. Hogstrom, Turbulent exchange above a pine
forest II. Organized structures, Boundary-Lay er Meteoroi, 4 9 , 2 3 1 -
263, 1989.
24. Berkhout, A. J., Least-squares inverse filtering and wavelet deconvo-
lution, Geophysics, 4 2 ( 7 ) , 1369-1383, 1977.
25. Berkooz, G., J. Elezgaray, and P. Holmes, Coherent structures in
r a n d o m media and wavelets, Physica D, 6 1 ( 1 ) , 47-58, 1992.
26. Bertrand, J., P. Bertrand, and J. P. Ovarlez, Dimensionalized wavelet
transform with application to radar imagery, Proc. of IEEE Int. Conf.
on Acoust., Speech and Signal Processing, Vol. 4, Toronto, 2909-2912,
1991.
27. Bijaoui, A. and A. Fresnel, Wavelets and astronomical image analysis,
in Wavelets, Fractals, and Fourier Transforms, M. Farge et al. (eds.),
Clarendon Press, Oxford, 1993.
28. Bijaoui, A., E. Slezak, and G. Mars, Universe heterogeneities from a
wavelet analysis, in Wavelets, Fractals, and Fourier Transforms, M.
Farge et al. (eds.), Clarendon Press, Oxford, 1993.
29. Bliven, L. F. and B. Chapron, Wavelet analysis and radar scattering
from water waves, Naval Research Reviews, 4 1 ( 2 ) , 11-16, 1989.
30. Bos, M. and E. Hoogendam, Wavelet transform for the evaluation of
peak intensities in flow-injection analysis, Anal. Chim. Ada, 2 6 7 ( 1 ) ,
73-80, 1992.
31. Bradshaw, G. A., Hierarchical analysis of spatial p a t t e r n and pro-
cesses of Douglas-fir forests using wavelet analysis, P h . D . thesis, Ore-
gon State Univ., 303 p., 1991.
32. Bradshaw, G. A. and T. A. Spies, Characterizing canopy gap structure
in forests using wavelet analysis, J. Ecology, 80, 105-215, 1992.
33. Broecker, W. S., D. Peteet, and R. Rind, Does the ocean-atmosphere
system have more t h a n one stable mode of operation? Nature, 3 1 5 ,
21-26, 1985.
34. Brotherton, T., T. Polland, R. Barton, and A. Krieger, Application of
time-frequency and time-scale analysis to underwater acoustic tran-
sients, in Proc. of IEEE-SP InVl. Symposium on Time-Frequency and
Time-Scale Analysis, Victoria, 513-516, 1992.
35. Bube, K., P. Lailly, P. Sacks, F. Santosa, and W. W. Symes, Si-
multaneous determination of source wavelet and velocity profile using
impulsive point-source reflections from a layered fluid, Geophysical
Journal, 9 5 ( 3 ) , 449-462, 1988.
36. Cahalan, R. F., M. Nestler, W. Ridgway, W. J. Wiscombe, and T.
348 Bibliography
Bell, Marine stratocumulus spatial structure, Pretorna, New Zealand,

paper presented at "4th International meeting on Statistical Clima-
tology," pp. 29-25, 1989.
37. Caldwell, J., Solution of Burgers' equation by Fourier transform meth-
ods, in Wavelets, Fractals, and Fourier Transforms, M. Farge et al.
(eds.), Clarendon Press, Oxford, 1993.
38. Chapman, N. R., I. Barrodale, and C. A. Zala, Measurement of sound-
speed gradients in deep-ocean sediments using /i-deconvolution tech-
niques, IEEE Journal of Oceanic Engineering, OE-9(l), 26-30, 1984.
39. Chapron, B., Application of higher order scale measurements, in Proc.
of IEEE-SP Int'l. Symposium on Time-Frequency and Time-Scale
Analysis, Victoria, 205-208, 1992.
40. Chapron, B. and F. L. Bliven, Scatterometer response and wavelet
transformation analysis of water wave surface, Traitement du Signal,
9(1), 27-31, 1992.
41. Chou, Kenneth C , A stochastic modeling approach to multiscale sig-
nal processing, Ph.D. Thesis, Massachusetts Institute of Technology,
May 1991.
42. Collineau, S. and Y. Brunet, Detection of turbulent coherent motions
in a forest canopy. Part I: Wavelet analysis, Boundary-Layer Meteo-
rol., 65, 357-379, 1993.
43. Combes, J. M., A. Brossmann, and Ph. Tchamitchian, Inverse prob-
lems and theoretical imaging, in Wavelets: Time-Frequency Methods
and Phase Space, J. Combes, A. Grossmann, and Ph. Tchamitchian
(eds.), Springer-Verlag, 1989.
44. Daneshvar, M. R. and C. S. Clay, Imaging of rough surfaces for impul-
sive and continuously radiating sources, J. Acoust. Soc. Am., 82(1),
360-369, 1987.
45. David, P. M. and B. Chapron, Underwater acoustic signal analysis
with wavelet process, J. Acoust. Soc. Am., 87(5), 2118-2121, 1990.
46. David, P. M. and B. Chapron, Underwater acoustic, wavelets and
oceanography, in Wavelets and Applications, Y. Meyer (ed.) Masson,
Paris and Springer, Berlin, 1992.
47. Davis, A., A. Marshak, and W. Wiscombe, Bi-multifractal analysis
and multi-affine modeling of non-stationary geophysical processes, ap-
plication to turbulence and clouds, Fractals, to appear, 1993.
48. Davis, A., A. Marshak, W. Wiscombe, and R. Cahalan, Multifractal
characterizations of non-stationary and intermittency in geophysical
fields, observed, retrieved or simulated, submitted to J. Geophys. Res.,
1993.
49. Donoho, D. L., I. M. Johnstone, G. Kerkyacharian, and D. Picard,
Density estimation by wavelet thresholding, Technical Report, Dept.
of Statistics, Stanford, Univ., 1993.
Bibliography 349
50. Duroure, C. and C. P. Fantodiji, Wavelet analysis and Fourier analysis

of stratocululus microphysical data, 1993, preprint.
51. Escalera, E. and A. Mazure, Wavelet analysis of subclustering: An
illustration, Abell 754, Astrophys. J., 388(1), 23-32, 1992.
52. Escalera, E., E. Slézak, and A. Mazure, New evidence of subclustering
in the coma cluster using the wavelet analysis, Astron. Astrophys.,
264(2), 379-384, 1992.
53. Everson, R. M. and L. Sirovich, A survey of wavelet analysis applied
to turbulence data, Report 89-182, Cent. Fluid Mech. Turbulence
Comput., Brown Univ., Providence, RI, 1989.
54. Everson, R., L. Sirovich, K. R. Sreenivasan, Wavelet analysis of the
turbulent jet, Physical Letters A, 146(6,7), 314-322, 1990.
55. Falconer, K. J., Wavelets, fractals and order-two densities, in
Wavelets, Fractals, and Fourier Transforms, M. Farge et al. (eds.),
56. Fantodji, C. P., C. Duroure, and H. R. Larsen, The use of wavelets
for the analysis of geophysical data, submitted to Physica, 1993.
57. Farge, M., Transformée en ondelettes continue et application a la
turbulence, Journée Annuelle de la Société Mathématique de France,
1990.
58. Farge, M., Continuous wavelet transform application to turbulence,
in Wavelets and Their Applications, G. Beylkin et al. (eds.), Jones &
Bartlett, Boston, 1991.
Annu. Rev. Fluid Mech, 24, 395-457, 1992.
60. Farge, M., The continuous wavelet transform of two-dimensional tur-
bulent flows, in Wavelets and Their Applications M. B. Ruskai et al.
(eds.), Jones & Bartlett, Boston, 275-302, 1992.
61. Farge, M., and G. Rabreau, Wavelet transform to analyze coherent
structures in two-dimensional turbulent flows, Proc. Scaling, Fractals
and Nonlinear Variability in Geophysics I, Paris, June 1988.
62. Farge, M., and G. Rabreau, Transformée en ondelettes pour détecter
et analyser les structures cohérentes dans les écoulements turbulents
bidimensionnels, C. R. Acad. Sei. Paris Ser. II, 307, 1479-1486,1988.
63. Farge, M. and M. Holschneider, Two-dimensional wavelet analysis of
two-dimensional turbulent flows, Proc. Sealing, Fractals and Nonlin-
ear Variability in Geophysics II, Barcelone, March 1989.
64. Farge, M. and G. Rabreau, Wavelet analysis of turbulent signals,
Proc. of Conf. on Chaos, Turbulence and Nonlinear Variability in
Geophysics, Barcelona, 1989.
65. Farge, M., M. Holschneider, and J. F. Colonna, Wavelet analysis of
coherent structures in two-dimensional turbulent flows, in Topological
Fluid Mechanics, H. K. Moffatt (ed.), Cambridge Univ. Press, 765-
350 Bibliography
776, 1990.
66. Farge, M., E. Goirand, and V. Wickerhauser, Wavelet packets analysis
compression and filtering of two-dimensional turbulent flows, LMD,
Ecole Normale Supérieure, Paris, 1991, Preprint.
67. Farge, M., E. Goirand, Y. Meyer, F. Pascal, and M. V. Wicker-
hauser, Improved predictability of two-dimensional turbulent flows
using wavelet packet compression, Fluid Dyn. Res., 1 0 ( 4 - 6 ) , 2 2 9 -
250, 1992.
68. Field, D. J., Scale-invariance and self-similar 'Wavelet' transforms: an
analysis of natural scenes and mammalian visual systems, in Wavelets,
Fractals, and Fourier Transforms, M. Farge et al. (eds.), Clarendon
Press, Oxford, 1993.
69. Flandrin, P., Time-frequency and time-scale, Proc. of J^th Acoust.,
Speech and Signal Processing Workshop on Spectrum Estimation Mod-
eling, 77-80, 1988.
70. Flandrin, P., Fractional Brownian motion and wavelets, in Wavelets,
71. Frick, P. and V. Zimin, Hierarchical models of turbulence, in Wavelets,
72. Gamage, N. K. K., Modeling and analysis of geophysical turbulence:
Use of optimal transforms and basis sets, Ph.D. Thesis, Oregon State
University, 135 pp., 1989.
73. Gamage, N. and W. Blumen, Comparative analysis of low level cold
fronts: wavelet, fourier, and empirical orthogonal function decompo-
sitions, Monthly Weather Rev., 1 2 1 , 2867-2878, 1993.
74. Gamage, N. K. K. and C. Hagelberg, Detection and analysis of mi-
crofronts and associated coherent events using localized transforms,
J. Atmos. Sei., 5 0 ( 5 ) , 750-756, 1993.
75. Gambis, D., Wavelet transform analysis of the length of the day and
the El-Nino/Southern Oscillation variations at intraseasonal and in-
terannual time scales, Ann. Geophysicae, 10, 331-371, 1992.
76. Genon-Catalot, V., C. Laredo, and D. Picard, Nonparametric esti-
mation of the diffusion coefficient by wavelet methods, Scandinavian
Jour, of Statistics, 1 9 ( 4 ) , 317-336, 1992.
77. Ghez, J. M. and S. Vaienti, On the wavelet analysis for multifractal
sets, J. Stat. Phys., 57, 415-420, 1989.
78. Giannakis, G. B., Wavelet parameter and phase estimation using
Cumulant slices, IEEE Trans, on Geoscience and Remote Sensing,
2 7 ( 4 ) , 452-455, 1989.
79. Glowinski, R., W. M. Lawton, M. Ravachol, E. Tenenbaum, Wavelet
solutions of linear and nonlinear elliptic, parabolic and hyperbolic
Bibliography 351
problems in one space dimension, in Computing Methods in Applied

Sciences and Engineering, R. Glowinski, A. Liehnewsky (eds.), SIAM
Press, Philadelphia, 55-120, 1990.
80. Gollmer, S. M., R. F. Harshvardan, and J. B. Snider, Wavelet analy-
sis of marine stratoculumus, Proceedings 11th International Conf. on
Clouds and Precipitation, Montreal, Quebec, August, 1992.
81. Goupillaud, P. L., Three new mathematical developments on t h e geo-
physical horizon, Geophysics: The Leading Edge of Exploration, June,
1992.
82. Goupillaud, P., A. Grossmann, and J. Morlet, Cycle-octaves and re-
lated transforms in Seismic signal analysis, Geoexploration, 2 3 , 8 5 -
102, 1984.
83. Goursat, M., Numerical results of stochastic gradient techniques for
deconvolution in seismology, Geoexploration, 2 3 , 103-119, 1984.
84. Grossmann A. and J. Morlet, Decomposition of Hardy functions into
square integrable wavelets of constant shape, SIAM J. Math. Anal.,
1 5 ( 4 ) , 723-736, 1984.
85. Grossmann, A., R. Kronland-Martinet, and J. Morlet, Reading and
understanding continuous wavelet transforms, in Wavelets: Time-
Frequency Methods and Phase Space, J. Combes, A. Grossmann, and
P h . Tchamitchian (eds.), Springer-Verlag, pp. 2-20, 1989.
86. Gurbatov, S. N. and A. I. Saichev, The self-similarity of D-
dimensional potential turbulence, in Wavelets, Fractals, and Fourier
Transforms, M. Farge et al. (eds.), Clarendon Press, Oxford, 1993.
87. Hagelberg, C. R. and N. K. K. Gamage, Structure preserving wavelet
decompositions of intermittent turbulence, Boundary-Layer Meteo-
rol., to appear, 1993.
88. Haynes, P. H. and W. A. Norton, Quantification of scale cascades in
t h e stratosphere using wavelet transforms, in Wavelets, Fractals, and
Fourier Transforms, M. Farge et al. (eds.), Clarendon Press, Oxford,
1993.
89. Heil, C. E. and D. F. Walnut, Continuous and discrete wavelet trans-
forms, SIAM Review, 3 1 ( 4 ) , 628-666, 1989.
90. Howell, J. F. and L. M a h r t , An adaptive multiresolution d a t a filter:
application to turbulence and climatic time series, J. Atmos. Sei., to
appear, 1994.
91. Hudgins, L. H., Wavelet analysis of atmospheric turbulence, P h . D .
thesis, University of California, Irvine, March, 300 p., 1992.
92. Hudgins, L. H., M. E. Mayer, and C. A. Friehe, Fourier and wavelet
analysis of atmospheric turbulence, in Progress in Wavelet Analysis
and Applications, Y. Meyer and S. Roques (eds.), Editions Frontieres,
491-498, 1993.
93. Hudgins, L. H., C. A. Friehe, and M. E. Mayer, Wavelet transform
352 Bibliography
and atmospheric turbulence, Physical Rev. Lett, Oct. 12, 1993.

94. Hunt, J.C.R., N.K.-R. Kevlahan, J. C. Vassilicos, and M. Farge,
Wavelets, fractals and Fourier transforms: Detection and analysis of
structure, in Wavelets, Fractals, and Fourier Transforms, M. Farge
et al. (eds.), Clarendon Press, Oxford, 1993.
95. Jones, J. G., P. G. Earwicker, and G. W. Foster, Multiple-scale cor-
relation detection, wavelet transforms and multifractal turbulence, in
Wavelets, Fractals, and Fourier Transforms, M. Farge et al. (eds.),
96. Kronland-Martinet, R., J. Morlet, and A. Grossmann, Analysis of
sound patterns through wavelet transforms, International J. of Pat-
tern Recognition and Artificial Intelligence, 1, 273-301, 1987.
97. Kumar, P. and E. Foufoula-Georgiou, A new look at rainfall fluc-
tuations and scaling properties of spatial rainfall using orthogonal
wavelets, J. Appl. Meteor., 32(2), 209-222, 1993.
tion of spatial rainfall fields: 1. segregation of large and small-scale
features using wavelet transforms, Water Resour. Res., 29(8), 2515-
2532, 1993.
tion of spatial rainfall fields: 2. self-similarity in fluctuations, Water
Resour. Res., 29(8), 2533-2544, 1993.
100. Larsonneur, J. L. and J. Morlet, Wavelets and seismic interpreta-
tion, in Wavelets, Time-Frequency Methods and Phase Space, J. M.
Combes, A. Grossman, P. Tchamitchian (eds.), Lecture notes on LPT1,
Springer-Verlag, 126-131, 1989.
101. Levy, S. and D. W. Oldenburg, The deconvolution of phase-shifted
wavelets, Geophysics, 47(9), 1285-1294, 1982.
102. Levy, S. and D. W. Oldenburg, Automatic phase correction of
common-midpoint stacked data, Geophysics, 52(1), 51-59, 1987.
103. Liandrat, J. and F. Moret-Bailly, The wavelet transform: Some ap-
plications to fluid dynamics and turbulence, European J. of Mech.
B/Fluids, 9(1), 1-19, 1990.
104. Lines, L. R. and T. J. Ulrych, The old and the new in seismic de-
convolution and wavelet estimation, Geophys. Prosp., 25, 512-540,
1977.
105. Little, S. A., P. H. Carter, and D. K. Smith, Wavelet analysis of
a bathymétrie profile reveals anomalous crust, Geophys. Res. Letts.,
20(18), 1915-1918, 1993.
106. Mahrt, L., Eddy asymmetry in the sheared heated boundary layer,
Journal of the Atmospheric Sciences, 48(3), 472-492, 1991.
107. Malakhov, A. and A. Yakimov, The physical models and mathemati-
cal description of 1/f noise, in Wavelets, Fractals, and Fourier Trans-
Bibliography 353
forms, M. Farge et al. (eds.), Clarendon Press, Oxford, 1993.

108. Mallat, S. and S. Zhong, Characterization of signals from multiscale
edges, IEEE Trans, on Patt. Anal, and Mach. Intel., 14, 710-732,
1992.
109. Mann, S. and S. Haykin, The adaptive chirplet: An adaptive general-
ized wavelet-like transform, Proc. of Society of Photo-Optical Instru-
mentation Engineers, 1565, 402-413, 1991.
resentation, Manuscript of the Center for Turbulence Research, Stan-
ford Univ., 1990.
resentation, J. Fluid Mech., 232, 469-520, 1991.
112. Meneveau, C , Dual spectra and mixed energy cascade of turbulence
in the wavelet representation, Physical Rev Lett., 11, 1450-1453,1991.
113. Meneveau, C. and K. R. Sreenivasan, The multifractal nature of tur-
bulent energy dissipation, J. Fluid Mech., 224, 429-484, 1991.
114. Meneveau, C , Wavelet analysis of turbulence: the mixed energy cas-
cade in Wavelets, Fractals, and Fourier Transforms, M. Farge et al.
115. Meyer, Y., Wavelets and Applications: Proceedings of the Interna-
tional Conference, Marseille, France, May 1989, Springer-Verlag, New
York, 1992.
116. Meyer, Y., Wavelets: Algorithms and Applications, SIAM Press,
Philadelphia, 1993.
117. Meyer, Y. and S. Roques, Progress in Wavelet Analysis and Applica-
tions, Editions Frontières, Gif sur Yvette, 1993.
118. Moffatt, H. K., Spiral structures in turbulent flow, in Wavelets, Frac-
tals, and Fourier Transforms, M. Farge et al. (eds.), Clarendon Press,
Oxford, 1993.
119. Morlet, J., G. Arens, I. Fourgeau, and D. Giard, Wave propagation
and sampling theory, Geophys., 47, 203-236, 1982.
120. Muller, J., Morphology of disordered materials studied by multifractal
analysis, in Wavelets, Fractals, and Fourier Transforms, M. Farge et
al. (eds.), Clarendon Press, Oxford, 1993.
121. Muzy, J. F., E. Bacry, and A. Arnéodo, Wavelets and multifractal
formalism for singular signals: Application to turbulence data, Phys.
Rev. Lett., 67(25), 3515-3518, 1992.
122. Muzy, J. F., E. Bacry, and A. Arneodo, Multifractal formalism for
fractal signals: The structure function aproach versus the wavelet-
transform modulus-maxima method, Phys. Rev. E, 47, 875-884,1993.
123. Meyers, S. D., B. G. Kelly, and J. J. O'Brien, An introduction to
wavelet analysis in oceanography and meteorology: with application
to the dispersion of Yanai waves, Monthly Weather Review, accepted
354 Bibliography
May 1993.
124. Otaguro, T., S. Takagi, and H. Satoy, P a t t e r n search in a turbulent
signal using wavelet analysis, Proc. 21st Japan Symp. on Turbulence,
Tokyo, J a p a n , 1989.
125. Pentland, A. P., Surface interpolation using wavelets, in Proc of 2nd
European Conference on Computer Vision, G. Sandini (ed.), Springer-
Verlag, 615-619, 1992.
126. Pentland, A. P. and B. Horowitz, A practical approach to fractal-
based image compression, in Data Compression Conference, J. A.
Storer and J. H. Reif (eds.), I E E E Comput. Soc. Press, Los Alamitos,
176-185, 1991.
127. P e r m a n n , D. and I. Hamilton, Wavelet analysis of time series for the
duffing oscillator: The detection of order within chaos, Phys. Rev.
Lett, 6 9 ( 1 8 ) , 2607-2610, 1992.
128. Persoglia, S., S. Sancin, and A. Vesnaver, Adaptive deconvolution by
lattice filters: Experience on synthetic and real data, Bollettino di
Geofisica Teorica ed Applicata, 2 7 ( 1 0 7 ) , 169-183, 1985.
129. Phuvan, S., T. K. Oh, N. P. Caviris, Y. Li, and H. H. Szu, Texture
analysis by space-filling curves and one-dimensional Haar wavelets,
Opt. Eng., 3 1 ( 9 ) , 1899-1906, 1992.
130. Pittner, S., J. Schneid, and C. W. Ueberhuber, Wavelet Literature
Survey, Institute for Applied and Numerical Mathematics, Technical
University Vienna, Wien, Austria, 1993.
131. Pouligny, B., G. Gabriel, J. F. Muzy, A Arnéodo, F. Argoul, and
E. Freysz, Optical wavelet transform and local scaling properties of
fractals, J. Appl. Cryst., 2 4 ( 5 ) , 526-530, 1991.
132. Powell, D. and C. E. Elderkin, An investigation of the application
of Taylor's hypothesis to atmospheric boundary layer turbulence, J.
Atmos. Sei., 3 1 , 990-1002, 1974.
133. R a m a n a t h a n , J. and O. Zeitouni, On the wavelet transform of frac-
tional Brownian motion, IEEE Trans, on Inform. Theory, 3 7 ( 4 ) ,
1156-1158, 1991.
134. Ranchin, T. and L. Wald, The wavelet transform for the analysis
of remotely sensed images, Int. J. Remote Sensing, 1 4 ( 3 ) , 615-619,
1993.
135. Rasmussen, H. 0 . , The wavelet Gibbs phenomenon, in Wavelets,
136. Redondo, J. M., Fractal models of density interfaces, in Wavelets,
137. Redondo, J. M., R. M. Gonzalez, and J. L. Cano, Fractal aggregates
in the atmosphere, in Wavelets, Fractals, and Fourier Transforms, M.
Bibliography 355
Farge et al. (eds.), Clarendon Press, Oxford, 1993.

138. Resnikoff, H. L. and C. S. Burrus, Relationships between t h e Fourier
transform and the wavelet transform, Proc. of the Society of Photo-
Optical Instrumentation Engineers, 1 3 4 8 , 291-300, 1990.
139. Ricker, N., T h e form and laws of propagation of seismic wavelets,
Geophysics, 18, 10-40, 1953.
140. Rioul, O. and M. Vetterli, Wavelets and signal processing, IEEE Sig-
nal Processing Magazine, 8 ( 4 ) , 14-38, 1991.
141. Rioul, O. and P. Duhamel, Fast algorithms for discrete and continuous
wavelet transforms, IEEE Trans, on Inform. Theory, 3 8 ( 2 ) , 569-586,
1992.
142. Rioul, O. and P. Flandrin, Time-scale energy distributions: A general
class extending wavelet transforms, IEEE Trans, on Signal Processing,
4 0 ( 7 ) , 1746-1757, 1992.
143. Saether, G., K. Bendiksen, J. Müller, and E. Froland, T h e fractal
dimension of oil-water interfaces in channel flows, in Wavelets, Frac-
tals, and Fourier Transforms, M. Farge et al. (eds.), Clarendon Press,
Oxford, 1993.
144. Saracco, G., Propagation acoustique en Régime harmonique & tran-
sitoire à travers un Milieu Inhomogène: Méthodes asymptotiques et
transformation en ondelettes, Thèse de l'Univ. d'Aix-Marseille II, Fac-
ulté des Sciences de Luminy, 232 p., 1989.
145. Saracco, G., A. Grossman, and P h . Tchamitchian, Use of wavelet
transforms in the study of propagation of transient acoustic signals
across a plane interface between two homogeneous media, in Wavelets,
Time-Frequency Methods and Phase Space, J. A. Combes, A. Gross-
m a n , P. Tchamitchian (eds.), Lecture Notes on I P T I , Springer-Verlag,
139-146, 1989.
146. Saracco, G. and P. Tchamitchian, A study of acoustic transmission
of transient signal in an inhomogeneous medium with t h e help of a
wavelet transform - application to an air-water plane interface, in Elec-
tromagnetic and Acoustic Scattering, Detection and Inverse Problems,
P.Chiapetta, B. Torréssani (eds.), World Scientific, 222-241, 1989.
147. Saracco, G., C. Gazanhes, J. Sageloli, and J.-P Sessarego, Analyse
Temps-echelle de la diffusion acoustique par des coques sphériques
élastiques en Régime Impulsionnel, J. Acoustique, 3 ( 4 ) , 381-392,
1990.
148. Saracco, G., P. Guillemain, and R. Kronland-Martinet, Characteriza-
tion of elastic shells by the use of the wavelet transform, in Proc. of
IEEE Ultrasonics Symposium, Vol. 2, Honolulu, 881-885, 1990.
149. Schiff, S. J., Resolving time-series structure with a controlled wavelet
transform, Opt Eng., 3 1 ( 1 1 ) , 2492-2495, 1992.
150. Schult, R. L. and H. W. Wyld, Using wavelets to solve t h e Burgers
356 Bibliography
equation - A comparative study, Physical Review A, 46(12), 7953-

7958, 1992.
151. Shann, W.-C, Finite element methods for Maxwell's equations with
stationary magnetic fields and Galerkin-wavelet s methods for two-
point boundary value problems, Ph.D. Thesis, Pennsylvania State
Univ., 132 p., 1991.
152. Sinha, B. and K. J. Richards, The wavelet transform applied to flow
around Antarctica, in Wavelets, Fractals, and Fourier Transforms,
M. Farge et al. (eds.), Clarendon Press, Oxford, 1993.
153. Slézak, E., A. Bijaoui, and G. Mars, Identification of structures from
Galaxy counts: Use of the wavelet transform, Astron. Astrophys.,
227(2), 301-316, 1990.
154. Stark, H.-G., Fractal graphs and wavelet series, Phys. Letters A,
143(9), 443-447, 1990.
155. Stark, H.-G., Continuous wavelet transform and continuous multiscale
analysis, J. Math. Anal. Appl., 169(1), 179-196, 1992.
156. Stark, J. and P. Bressloff, Iterated function systems and their appli-
cations, in Wavelets, Fractals, and Fourier Transforms, M. Farge et
al. (eds.), Clarendon Press, Oxford, 1993.
157. Stone, D. G., Wavelet estimation, Proc. of IEEE Int. Conf. on
Acoust., Speech and Signal Processing, 72(10), 1394-1402, 1984.
158. Strang, G., The optimal coefficients in Daubechies wavelets, Physica
D, 60(1-4), 239-244, 1992.
159. Strichartz, R. S., Wavelet expansions of fractal measures, J. Geom.
Anal., 1(3), 269-289, 1991.
160. Tewfik, A. H. and M. Kim, Correlation structure of the discrete
wavelet coefficients of fractional Brownian motion, IEEE Trans, on
Inform. Theory, 38(2), 904-909, 1992.
161. Turner, B. J. and M. Y. Leclerc, Conditional sampling of coherent
structures in atmospheric turbulence using the wavelet transform, J.
Atmos. and Ocean. Tech., to appear, 1994.
162. Vassilicos, J. C , Fractals in turbulence, in Wavelets, Fractals, and
Fourier Transforms, M. Farge et al. (eds.), Clarendon Press, Oxford,
1993.
163. Vergassola, M. and U. Frisch, Wavelet transforms of self-similar pro-
cesses, Physica D, 54(1,2), 58-64, 1991.
164. Vermeer, P. L., The multiscale and wavelet transform with applica-
tions in well log analysis, Ph.D. thesis, Techn. Univ. Delft, 200 p.,
1992.
165. Vermeer, P. L. and J.A.H. Alkemade, Multiscale segmentation of well
logs, in Wavelets, Fractals, and Fourier Transforms, M. Farge et al.
166. Wallace, R. S., Finding natural clusters through entropy minimiza-
Bibliography 357
tion, Ph.D. thesis, School of Comput. Science, Carnegie Mellon Univ.,

Pittsburg, PA 15213, Jun. 1989.
167. Wickerhauser, M. V., Lectures on wavelet packet algorithms, Techni-
cal Report. Dept. of Math., Washington Univ., 1992.
168. Wickerhauser, M. V., Acoustic signal compression with wavelet pack-
ets, in Wavelets - A Tutorial in Theory and Applications, C. K. Chui
(ed.), Academic Press, Boston, 679-700, 1992.
169. Willis, G. E. and J. Deordorff, On the use of Taylor's translation
hypothesis for diffusion in the mixed layer, Quart. J. Roy. Meteorol.
Soc, 102, 817-822, 1976.
170. Wilson, R., A. D. Calway, and E.R.S. Pearson, A generalized wavelet
transform for Fourier analysis: The multiresolution Fourier transform
and its application to image and audio signal analysis, IEEE Trans,
on Inform. Theory, 38(2), 674-690, 1992.
171. Wornell, G. W., A Karhunen-Loève-like expansion for l//-processes
via wavelets, IEEE Trans, on Inform. Theory, 36(4), 859-861, 1990.
172. Wornell, G. W. and A. V. Oppenheim, Estimation of fractal signals
from noisy measurements using wavelets, IEEE Trans. Sig. Proc,
40(3), 611-623, 1992.
173. Wornell, G. W. and A. V. Oppenheim, Wavelet-based representations
for a class of self-similar signals with application to fractal modulation,
IEEE Trans, on Inform. Theory, 38(2), 785-800, 1992.
174. Wyngaard, J. C. and S. F. Clifford, Taylor's hypothesis and high-
frequency turbulence spectra, J. Atmos. Sei., 34, 922-929, 1977.
175. Xu, J.-C. and W.-C. Shann, Galerkin-wavelet methods for two-point
boundary-value-problems, Numer. Math., 63(1), 123-144, 1992.
176. Yamada, M. and K. Ohkitani, Orthonormal wavelet expansion and its
application to turbulence, Prog. Theor. Phys., 83(5), 819-823, 1990.
177. Yamada, M. and K. Ohkitani, Orthonormal wavelet analysis of tur-
bulence, Fluid Dynamics Res., 8(1-4), 101-115, 1991.
178. Yamada, M. and K. Ohkitani, An identification of energy cascade
in turbulence by orthonormal wavelet analysis, Prog. Theor. Phys.,
86(4), 799-815, 1991.
179. Zhong, S., Edge representation from wavelet transform maxima,
Ph.D. Thesis, Univ. New York, 129 p., 1991.
180. Zubair, L., Studies in turbulence using wavelet transforms for data
compression and scale-separation, Thesis, Yale Univ., New Haven.
360 Subject Index
σ-τ curve, 328, 333, 339-341 97, 130-131, 138-139

Adaptive cutoff scale, 121 Conditional simulation, 225, 228,
Affine smoothing, 16 230-233, 245
Allan variance, 326 ff. Convolution, 169
Anemometer measurements, 107 Correlated random field, 228
Aquifer test, 214, 224 Correlations, 185-186
Atmospheric surface layer, 82-83, acoustic-geotechnical, 207
86 Covariance, 111
Attenuation, 184-187, 189, 1 9 1 - Cross scalogram , See scalogram
192, 202, 208 Cubic spline, 170
apparent, 186-187 Curvature, 45
coefficient, 187, 189 Cut-off scale, 129, 144, 148
compressional wave, 203 D a t a compression, 29
decibels, 192 Daubechies' wavelets, 35, 302, 315,
effective, 187 317
intrinsic, 186-187 Daubechies sequence, 122
nepers, 192 Denoising algorithm of Coifman
Autocovariance, 325 and Majid, 319
Backscatter, 185 Detail approximation, 25
Band-pass filter, 326 Detection, 45, 129, 131, 138, 141,
Bathymetry, 167 148
Best-basis, 303 Difference terms, 110
Biorthogonal wavelets, 27 Dilation parameter, See also scale
Boundary layer, 112 parameter, 9
Brownian motion, 287-289 Dispersion, 213-214, 229
fractional, 260-261 Dissipation, 82, 91-93, 97
Brunt-Vaisala frequency, 130, 143 Downsampling, see subsampling
Canopy, 129-132, 134, 137-139, Duration scale, 129-131,135,137,
141, 146 143-144, 146, 148
Cascade processes, 252, 285, 288 Dyadic numbers, 109
Chirp, 18, 187 Edge detection, 170
Cloud structure, internal, 261, 272 Energy density function, 13, 15
Coherent motion, 130, 132, 146 Exploration seismic data, 184
Coherent structure, 45, 47, 66, 69, Fast Wavelet Transforms, 86
73-74,126,129,144,146 Fault, 167
Coiflets, 302, 304, 315, 318 F I F E 87, 45, 47, 57
Complexity regularization, 320 Filter, 131, 133, 144, 146, 170,
Conditional averaging, 129, 139, 328, 330
148 band-pass, 326
Conditional sampling, 46, 50-53, high-pass, 331
55, 67, 78, 83, 9 3 - 9 4 , 9 6 - low-pass, 331, 334
Subject Index 361
octave band, 336 decomposition, 107, 110, 125

phase function, 335-336 transform, 109
q u a d r a t u r e mirror, 331, 336 Haar-Walsh basis, 301, 304, 315
scaling, 331-332 Haar wavelet, 16, 23, 85-86, 8 8 -
transfer function, 334-336 89, 92, 94, 96, 99, 108,
wavelet, 331-332 125, 135, 138, 141, 327
zero phase, 336 if.
Fine scale mode, 115 Halo wavelet, 31
Flux, 45, 48, 57-58, 65-66, 68, High-pass filter, 331
70-71, 73, 77 High resolution acoustic, 183, 185
Fourier Hilbert space, 24
cospectrum, 135, 137, 139- Hydraulic conductivity, 213-214,
140, 143-144 222, 224, 228, 232, 246
spectrum, 135, 137, 139-140, Image, 167
143-144 Inactive eddies, 123
Fourier analysis, 168 Inertial range, 45, 47, 73
Fourier spectral window, 109 Inertial subrange, 81-82, 87, 89,
Fourier transform, 3, 184, 188, 192, 93,98
216 Inner product, 9
short window ( S W F T ) , 184 Instantaneous energy, 15
Fractal, 224, 229-231, 241-243, Integral scale, 281-282
245-246, 341 in geophysical data-streams,
Fractional Brownian motion, 229 263
Fractional Gaussian process, 326 in multiplicative cascades, 272,
Frames 285-286
definition, 25 in Ornstein-Uhlenbeck processes,
frame bounds, 25 288
multivoice frames, 27 Integral transform, 9
redundancy ratio, 26 Integration kernel, 2, 5
snug frames, 26 Intermittency, 45, 48, 82-83, 88,
tight frames, 26 91, 93-94, 96, 98, 248,
wavelet frames, 25-26 252, 265, 270, 276, 285
Frequency stability, 327 exponents t h a t characterize,
Gabor transform, 7 270
Geometrical spreading, 186, 188- relation to sparseness, 271
189, 191-192, 197, 202 Intermittency index, 45
Geostatistics, 225, 227 Isometry, 5, 11
Globally averaged statistics, 116 Isotropie motions, 114
Gradient skewness, 107, 121 J u m p , 130, 138, 141
Gravity wave, 129-130, 134, 143, Kinetic energy, 124
146 Karhunen-Loève basis, 303-304
Haar, 109 Kolmogorov complexity, 308
basis set, 109 Kraft inequality, 306
362 Subject Index
Kriging, 225, 228, 230-233 white, 282

Large eddy mode, 112, 114 Noise reduction, 26
Least asymmetric wavelet, 327, 331, Non-overlapped estimator, 329-330,
335-342 342
Library of orthonormal bases, 300- Non-stationarity, 248, 255-256, 261,
303 276, 286
Linear B-spline, 170 exponents t h a t characterize,
Littlewood-Paley wavelet, 301 260
Local flux maxima, 112 relation to roughness, 260
Local isotropy, 91-93 with stationary increments, 256,
Local trigonometric transform, 303 286-287
Location parameter, 9, 12 Nonstationary process, 326-327
Long-memory process, 325 if. Ocean shear, 334 if.
Low-pass filter, 331, 334 Octave band filter, 336
Maximal-overlap Octaves, 185
estimator, 330-333, 339, 342
Orthogonal modes, 107, 115
pyramid algorithm, 332
Orthogonal projection operators,
Maximum likelihood estimate, 307,
25
310
Orthogonal wavelet decomposition,
Mesoscale mode, 114
23
Mexican h a t wavelet, 16, 135,137-
Orthonormal basis, 21
138, 141, 144, 148
Partition, 45, 47-48, 54, 56-57,
Microfronts, 112
60, 63-64, 66-67, 70, 7 3 -
Mid-Atlantic Ridge, 169
74, 77
Minimum description length prin-
Phase function, 335-336
ciple, 305-310
Modes of variation, 108 Phase-space, See time-frequency
M o m e n t u m flux, 108, 112, 125 analysis
Monte Carlo, 229-231, 238 Physically distinct modes, 109, 125
Morlet wavelet, 168 Piece-wise constant decomposition,
application, 4 107, 125
definition, 17 Porous media, 214, 216, 245-246
two-dimensional, 29 characterization, 214, 216, 223,
Multi-affine processes, 260, 263, 229, 232, 246
278 heterogeneous, 215-216
Multifractal analysis, 248, 275 Power-law process, 327-328, 334,
Multigrid, 223, 230 339, 341
Multiresolution Power spectra, 184, 199
analysis, 21 Pressure fluctuations, 124
nest, 24, 28 Prolate spheroidal wave function,
representation, 23-24 304
Noise, 255 Pyramid algorithm, 331-332
1 / / , 261 maximal-overlap, 332
Subject Index 363
Q u a d r a t u r e mirror filter, 301, 331, in the sense of Holder-Lipshitz

336 heuristics, 259
Quad-tree, 315 in the sense of singular mea-
R a m p , 130, 132, 138, 141 sures, 270, 276
R a n d o m processes, 252 Soil properties, 183, 185-187
auto-correlation function for, acoustic, 183
281 geotechnical, 183
as models for time-series, 251 Sparker, 194, 196-197, 207
one-point statistics, 255 Spectra, 45-48, 52, 54-55, 57, 7 3 -
two-point statistics, 256, 281 75, 77-78
Reconstruction algorithm, 144 Spectral density function, 328, 3 4 0 -
R e d u n d a n t representation, See frames 341
Reynolds number, 109 Spectral ratio, 188-189, 191-192,
Roll vortices, 123 198, 202, 208
Roughness, 168 Stationarity, 255, 280
Sample mean, 326-327 spectral criterion for, 255, 282
Sample variance, 326 vs. ergodicity, 280
Scale parameter, 9, 12 Stationary process, 325-326, 3 2 8 -
Scaling filter, 331-332 329, 334
Scaling function, 23 Statistical scale-invariance, 250, 259,
269, 280
properties, 34
exponents, 250, 255, 259, 269,
Scalogram, 13, 16
276
Sdf, see spectral density function
power-law behavior, 250, 255
Sea Beam, 170
scaling range, 255
Seafloor, 167
Stochastic continuity, 288
Seismic
Stochastic process
high resolution, 194 auto-covariance function, 15
multichannel, 197 non-stationary, 15
Self-similarity, 326, 328 power spectrum, 15
Semivariogram, 230 stationary, 15
Series expansion, 21 Stochastic theories, 213-214
Shannon code, 305 Storage reduction ratio, 317
Shannon wavelet, 301 Structure function, 82, 92-97, 256,
Shear driven, 107, 121, 125 276, 286
Short-time Fourier transform, 5 Subsampling, 330, 332-333, 336,
Sigma-tau curve, 328, 333, 3 3 9 - 338, 342
341 Taylor's hypothesis, 123
Similarity arguments, 123 Telescoping intervals, 113
Simultaneous noise suppression and Temperature, 129-134, 138, 141,
signal compression algo- 143, 146-148
rithm, 310-314 Tensor product, 31
Singularity analysis, 265, 276 Texture, 180
364 Subject Index
Theoretical compression rate, 319 Vertical ocean shear, 334 ff.

Time-frequency analysis, 3 Wavelet, 9
phase-space representation, 5, analyzing, 188
13 anti-symmetric, 28, 45, 47,
time-frequency localization, 3, 57, 61, 68, 70
6, 11 coherence, 153
time-frequency plane, 5 coherency, 153
time-frequency resolution cell, compact support, 9, 302
8 covariance, 13-14, 129, 131,
Time interval, 130, 138-139, 146 134-137, 141, 143-144,
Time-scale analysis, 12 146, 148
Time-scale representation, 22 cross-spectrum, 14, 153
Time series, 108, 129-134, 137- examples, 16, 31
138, 141, 143, 146, 148 expansion, 24
Topography, 167 filter, 331-332
Transfer function, 334-336 Haar, 327 if.
Transient, 8 least asymmetric, 327, 331,
Translation numbers, 110 335-342
Translation parameter, See also Morlet, 154, 188
location parameter, 9 mother wavelet, 2
Transport mode, 107, 112, 115 orthogonal, 108
Transport physics, 108 regularity, 36, 302
Turbulence, 128-131,138, 143,146, spectrum, 14, 153
148, 261, 272, 277, 285 symmetric, 28, 45, 47, 57, 69,
Turbulence data, 108, 125 74
Turning bands, 228, 239 transform, 131
Two-dimensional local trigonomet- two-dimensional, 29, 314
ric transform, 315 vanishing moments, 36, 302
Two-dimensional wavelet packet variance, 13, 129-131, 134-
transform, 315 138, 141, 143-144, 146,
Two-dimensional wavelet transform 148, 327, 331 ff
continuous, 29 Wavelet packet transform, 302
orthogonal, 31 Wavelet shrinkage, 320
separable, 31 Wavelet transform, 45-57, 61-69,
Uncertainty principle, 4, 7, 12 74, 77,129-130,134-135,
Variance 138, 141, 144, 146, 148,
Allan, 326 if. 301
sample, 326 admissibility condition, 10
of sample mean, 326 basis function, 216-218, 220,
wavelet, 327, 331 if 224-225, 230, 232, 243-
Variogram, 286 246
Velocity, 129-133, 137-138, 141, center of passing band, 1 1 -
143, 146-147 12, 14
Subject Index 365
continuous, 257, 266 phase relations, 161

definition, 9 significant wave height, 158
detail coefficients, 218-220, 2 2 2 - wave grouping, 156
225 Zero phase filter, 336
discrete, 19, 331, 342
extrapolation, 216-217, 225,
227, 232, 238, 243
inverse, 10,129-130,132,139,
144-145
multiresolut ion analysis, See
multiresolution analysis
non-orthogonal, See also frames,
25
orthogonal, 21
phase-space representation, 13
phase-space resolution cell, 12
reconstruction, 215-217,222-
226, 228, 230-233, 243,
245-246
smooth coefficients, 218-225
time-frequency localization, 11
two-dimensional, See two-dimensional
wavelet transform
White Gaussian noise, 304, 3 0 9 -
310, 315-316
Wiener-Khintchine relation for scal-
ing processes, 282
non-stationary situations, 260,
286
stationary situations, 271
Wigner-Ville spectrum, 15
Wind gusts, 112
Windowed Fourier transform, 208
definition, 5
discrete, 6
phase-space representation, 5
reconstruction formula, 5
Wind tower data, 123
Wind waves, 158
breaking waves, 161
equilibrium range, 164
localized frequency spectrum,
163
Detailed Table of Contents
Contributors ix
Preface xi
I. Wavelet Analysis in Geophysics: An Introduction . . . .1

1. Prologue 1
2. Time-Frequency Analysis 3
2.1 Windowed Fourier transform 5
2.2 Wavelet transform 8
3. Wavelets and Time-Scale Analysis . 12
3.1 Time-scale transform 12
3.2 Scalogram, wavelet variance and covariance 13
3.3 Non-stationarity and the Wigner-Ville spectrum . . . 15
4. Examples of One-Dimensional Wavelets 16
4.1 Haar wavelet 16
4.2 Mexican hat wavelet 16
4.3 Morlet wavelet 17
5. Discrete Wavelet Transforms 19
5.1 Orthogonal wavelet transforms and
multiresolution analysis 21
5.2 Non-orthogonal wavelet transforms 25
5.3 Biorthogonal wavelets 27
6. Two-Dimensional Wavelets 29
6.1 Continuous wavelets 29
6.2 Orthogonal wavelets 31
7. Conclusions 33
A. Properties of Scaling Function 34
B. Daubechies' Wavelets 35
C. Implementation Algorithm for Orthogonal Wavelets . . . . 37
References 39
II. Applications of Structure Preserving Wavelet

Decompositions to Intermittent Turbulence:
A Case Study 45
1. Introduction 46
367
368 Detailed Table of Contents
2. The Wavelet Transform 48

2.1 The continuous case . . 48
2.2 Dyadic subsampling in scale . . 50
2.3 Finite resolution in discrete applications . . 53
3. Filtering Properties of Wavelets . . 54
3.1 Partitions as filters . . 54
3.2 Detection characteristics of anti-symmetric and
symmetric wavelets . . 57
3.3 Quantifying intermittency . . 64
4. Applications to Atmospheric Data . . 67
4.1 Description of the data . . 67
4.2 Analysis of the data . . 68
5. Summary and Conclusions . . 74
References . . 78
Ill . Intermittency in Atmospheric Surface

Layer Turbulence: The Orthonormal
Wavelet Representation . . 81
1 Introduction . . 81
2. Analysis of Turbulence Using Wavelet Transforms . . . . . 83
2.1 Continuous wavelet transforms . . 84
2.2 Orthonormal wavelet expansions . . 84
3 Experiment . . 86
4 Wavelet Statistics . . 88
4.1 Relation between wavelet coefficients and Fourier
power spectrum . . 89
4.2 Relation between wavelet coefficients and
structure function . . 92
4.3 Conditional sampling and intermittency effects
onK41 . . 93
fi Results and Discussion . . 95
6. Summary and Conclusions . . 101
References 101
IV. An Adaptive Decomposition: Application

to Turbulence . . 107
1 Introduction . . 108
2. Partitioning the Time Series . . 109
2.1 Decomposition method . . 109
2.2 Modes of variation . . Ill
3. Variance and Covariance Spectra . . 116
4 Physical Interpretation . . 123
5. . . 125
References 126
Detailed Table of Contents 369
V. Wavelet Analysis of Diurnal and Nocturnal

Turbulence above a Maize Crop 129
1. Introduction 129
2. Experimental Procedures and Flow Characteristics . . . . 131
3. Wavelet Variance and Covariance 134
4. Jump Detection and Conditional Sampling 138
5. Wavelet Decomposition 143
6. Summary and Conclusions 146
References 148
VI. Wavelet Spectrum Analysis and Ocean

Wind Waves 151
1. Introduction 151
2. Wavelet Spectrum 152
3. Applications 155
3.1 Wave grouping effects 156
3.2 How do wind waves grow? 158
3.3 Detecting breaking waves 161
4. Concluding Remarks 164
References 165
VII. Wavelet Analysis of Seafloor Bathymetry:

An Example 167
1. Introduction 167
2. 1-D Topographic Profile 168
3. 2-D Bathymetry 169
3.1 Data 170
3.2 Filter design 170
3.3 Linear B-spline wavelet 173
3.4 Derivative of cubic B-spline 176
4. Conclusions 180
References 181
VIII. Analysis of High Resolution Marine Seismic

Data Using the Wavelet Transform 183
1. Introduction 183
2. Acoustic-Geotechnical Correlations: Physical and
Historical Context 185
3. The Wavelet Transform for Acoustic Soil Analysis 188
3.1 The wavelet transform and Morlet wavelet 188
3.2 Estimating attenuation: A simple example 189
4. Experimental Field Program 194
5. Data Analysis 196

5.1 Conventional data analysis and processing 197
5.2 Wavelet transform analysis of an acoustic signal . . . 197
5.3 Estimation of attenuation 205
6. Conclusion 206
References 208
IX. Including Multi-Scale Information

in the Characterization of Hydraulic
Conductivity Distributions 213
1. Introduction 213
2. Wavelet Based Interpolation 216
2.1 Discrete transform introduction 216
2.2 Multi-scale reconstruction 222
2.3 Multi-scale extrapolation 225
3. Traditional Interpolation Methods 225
3.1 Simple assignment 227
3.2 Kriging 228
3.3 Conditional simulation 228
4. Evaluation of Methods 228
4.1 Reconstruction comparison 228
4.2 Optimum scale 232
4.3 Wavelet basis effects 243
5. Conclusions and Recommendations 243
References 246
X. Wavelet-Based Multifractal Analysis of

Non-Stationary and/or Intermittent
Geophysical Signals 249
1. Introduction and Overview 249
2. Preliminary Considerations on Stationarity, Ergodicity
and Scale-Invariance 251
2.1 Data specification and spectral analysis 251
2.2 Stationarity and the spectral exponent 255
3. Quantifying and Qualifying Non-Stationarity with
Structure Functions 256
3.1 Wavelets for structure function analysis 256
3.2 Scale-invariant structure functions, generalities . . . 259
3.3 Geometrical, spectral and dynamical implications
of structure function statistics 260
3.4 Examples from theory, turbulence and
cloud structure 261
Detailed Table of Contents 371
4. Quantifying and Qualifying Intermittency with Singular

Measures 265
4.1 Wavelets and "scaling functions" for
singularity analysis 265
4.2 Scale-invariant singular measures, generalities . . . . 269
4.3 Geometrical, spectral and dynamical implications
of statistical singularity analysis 270
4.4 Examples from theory, turbulence and
cloud structure 272
5. Summary and Bi-Multifractal Statistics 275
5.1 Comparative multifractal analysis from the
wavelet theoretical standpoint 275
5.2 On ζ(ρΜο-Κ(ς) connections 276
5.3 The mean multifractal plane 277
6. Conclusions 278
A. Stationarity, Intermittency, Integral Scales and
Stochastic Continuity—A Tutorial in the Framework of
Scalé-Invariance and Second Order Statistics 280
A.1 Stationarity versus ergodicity 280
A.2 Stationarity and scale-invariance in physical space,
the integral scale 281
A. 3 A Fourier space criterion for stationarity 282
A.4 Multiplicative cascade models, stationarity
and intermittency 285
A.5 Non-stationary scale-invariant situations, up to
the integral scale 286
A. 6 Additive scale-invariant models, non-intermittent
and non-stationary 287
A. 7 Stochastic continuity with stationarity or scale-
invariance but not both 288
A.8 Summary with analysis and modeling of
geophysical data in mind 289
List of Symbols 290
References 293
XI. Simultaneous Noise Suppression and Signal

Compression Using a Library of Orthonormal
Bases and the Minimum Description
Length Criterion . . 299
1. Introduction 299
2. A Library of Orthonormal Bases 300
2.1 Wavelet bases 301
2.2 Wavelet packet best-bases 302
2.3 Local trigonometric best-bases 303
3. Problem Formulation 304

4. The Minimum Description Length Principle 305
5. A Simultaneous Noise Suppression and Signal
Compression Algorithm 310
6. Extension to Images 314
7. Examples 315
8. Discussions 319
9. Conclusions 321
References 322
XII. Long-Memory Processes, the Allan Variance

and Wavelets 325
1. Introduction 325
2. The Allan Variance and the Haar Wavelet 327
3. The Wavelet Variance 331
4. An Application to Vertical Shear Measurements 334
References 343
Bibliography 345
Subject Index 359

WAVELET ANALYSIS AND ITS APPLICATIONS
Charles K. Chui, Series Editor
1. Charles K. Chui (Editor), An Introduction to Wavelets: Wavelet

Analysis and Its Applications
2. Charles K. Chui (Editor), Wavelets: A Tutorial in Theory
and Applications
3. Larry L. Schumaker and Glenn Webb (Eds. ), Recent Advances in
Wavelet Analysis
4. Efi Foufoula-Georgiou and Praveen Kumar (Eds.), Wavelets
in Geophysics
373

Efg 149

Uploaded by

Copyright:

Available Formats

Efg 149

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Efg 149

Uploaded by

Copyright:

Available Formats

Wavelet Analysis and

CHARLES K. CHUI, SERIES EDITOR

A list of titles in this series appears at the end of this volume.

Copyright © 1994 by ACADEMIC PRESS, INC.

Academic Press, Inc.

United Kingdom Edition published by

Library of Congress Cataloging-in-Publication Data

Wavelets in geophysics / edited by Efi Foufoula-Georgiou, Praveen

PRINTED IN THE UNITED STATES OF AMERICA

J O H N D. ALBERTSON (81), Hydrologie Science, University of California at

PAUL C. LIU (151), NOAA Great Lakes Environmental Research Laboratory,

containing the remaining portion of the signal (essentially characterized by

Efi Foufoula-Georgiou Praveen Kumar

Praveen K u m a r and Efi Foufoula-Georgiou

Abstract. Wavelet analysis is a rapidly developing area of mathematical and

processes or signals: (i) as an integration kernel for analysis to extract

This article is organized as follows. Section 2 discusses Fourier and

1. T h e contributions of different frequency bands (i.e., of the differ-

2. This separation is achieved without excessive loss of resolution in the

3. T h e reconstruction of the original function from its "representation"

T h e first two conditions essentially characterize the property known as

which states t h a t one cannot obtain arbitrary good localization in b o t h

where t h e integration kernel is gtj,t(u) = g(u — t)e~iuju. This transform

provided / ^ |g(£)| 2 dx = 1 (which we assume from here on). It is invertible

T h e parameters t and ω can b e assigned discrete values, say t = nt0

Figure 2. Real (solid lines) and imaginary parts (dot-dashed lines)

and ω = mcjo, and we obtain the discrete windowed Fourier transform

Figure 3. Uncertainties in time (top) and frequency (bottom) lo-

where ^ ^ ( u / ) is the Fourier transform of gu,t(u) and overbar indicates

i.e., arbitrary high precision in b o t h time and frequency cannot b e achieved.

F i g u r e 4. Figure showing the phase-space representation using the

Once a window function g(t) is chosen b o t h ag and σ§ are fixed. There-

2.2. Wavelet transform

frequency (or short wavelength) features have smaller support, it would be

Here λ is a scale parameter, t a location parameter and the functions ip\jt(u)

1. compact support, or sufficiently fast decay, to obtain localization in

2. zero mean, i.e., J ^ t / ^ ^ d i = 0, although higher order moments

Figure 5. Schematic illustration of the effect of dilation on a "generic"

may also be zero, i.e.,

tktp(t) dt = 0 for k = 0 , . . . , N - 1. (13)

T h e inverse wavelet transform is given by ([15], eq. 2.4.4)

T h e wavelet transform is also an energy preserving transformation, i.e., an

Therefore, as in the windowed Fourier transform, we need to study the

It is easy to verify t h a t the following relationships hold:

1. T h e standard deviation σφχ t satisfies

2. T h e standard deviation σ;, satisfies

3. T h e center of passing band αΛ corresponding to the wavelet iß\yt (u)

§3. Wavelets and Time-Scale Analysis

Figure 6. Figure showing the phase-space representation using the

By considering the RHS of (26) we see t h a t |W/(A,£)| 2 /C^A 2 can b e con-

Equation (26) can also be written as

has been referred to as wavelet covariance (see [4]) or wavelet cross-spectrum

Using this relationship, the scale information can be translated to frequency

given by the power spectrum of ip\(t). This relation is interesting although

3.3. Non-stationarity and the Wigner-Ville spectrum

R(r) = 8[X(t)X{t - r)] < = i S(u>) = \Χ(ω) | 2

T h e relationship of interest to us is given by the relation between the scalo-

i.e., the scalogram can be obtained by affine smoothing (i.e., smoothing at

4.1. Haar wavelet

φ(ή _ _-1/4ρ-ίωορ-2/2 CJO > 5. (43)

^λ,ο(ω) = χπ-ι/β-^-^)'/2 = λ π _ι/46-^- ω ) _

/() = Σ Σ (ί'Ψ™,η)Ψ,η,η()+ Σ Σ </'</>m,n>V>m,n(i)·