Professional Documents
Culture Documents
Wavelet Analysis Basic Concepts and Applications 1St Edition Sabrine Arfaoui Online Ebook Texxtbook Full Chapter PDF
Wavelet Analysis Basic Concepts and Applications 1St Edition Sabrine Arfaoui Online Ebook Texxtbook Full Chapter PDF
Wavelet Analysis Basic Concepts and Applications 1St Edition Sabrine Arfaoui Online Ebook Texxtbook Full Chapter PDF
https://ebookmeta.com/product/environmental-biotechnology-basic-
concepts-and-applications-1st-edition-viswanath-buddolla/
https://ebookmeta.com/product/introductory-immunology-basic-
concepts-for-interdisciplinary-applications-3rd-edition-actor/
https://ebookmeta.com/product/pharmacokinetic-and-
pharmacodynamic-data-analysis-concepts-and-applications-5th-
edition-johan-gabrielsson/
https://ebookmeta.com/product/big-data-analysis-for-green-
computing-concepts-and-applications-1st-edition-rohit-sharma/
Wavelet Numerical Method and Its Applications in
Nonlinear Problems You He Zhou
https://ebookmeta.com/product/wavelet-numerical-method-and-its-
applications-in-nonlinear-problems-you-he-zhou/
https://ebookmeta.com/product/primary-mathematics-3a-hoerst/
https://ebookmeta.com/product/heterophase-polymerization-basic-
concepts-and-principles-1st-edition-hugo-hernandez/
https://ebookmeta.com/product/hydrochemistry-basic-concepts-and-
exercises-2nd-edition-eckhard-worch/
https://ebookmeta.com/product/rowland-and-tozer-s-clinical-
pharmacokinetics-and-pharmacodynamics-concepts-and-applications-
concepts-and-applications-5th-edition-derendorf/
Wavelet Analysis
Wavelet Analysis
Basic Concepts and Applications
Sabrine Arfaoui
University of Monastir, Tunisia
University of Tabuk, Saudi Arabia
Carlo Cattani
University of Tuscia, Italy
First edition published 2021
by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742
The right of Sabrine Arfaoui, Anouar Ben Mabrouk, and Carlo Cattani to be identified as authors of this work has been
asserted by them in accordance with sections 77 and 78 of the Copyright, Designs and Patents Act 1988.
Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers
have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright
holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowl-
edged please write and let us know so we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or
utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including pho-
tocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission
from the publishers.
For permission to photocopy or use material electronically from this work, access www.copyright.com or contact the
Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. For works that are
not available on CCC please contact mpkbookspermissions@tandf.co.uk
Trademark notice: Product or corporate names may be trademarks or registered trademarks and are used only for
identification and explanation without intent to infringe.
List of Figures ix
Preface xi
Chapter 1 Introduction 1
2.1 INTRODUCTION 5
2.2 WAVELETS ON R 6
2.2.1 Continuous wavelet transform 7
2.2.2 Discrete wavelet transform 10
2.3 MULTI-RESOLUTION ANALYSIS 11
2.4 WAVELET ALGORITHMS 13
2.5 WAVELET BASIS 16
2.6 MULTIDIMENSIONAL REAL WAVELETS 21
2.7 EXAMPLES OF WAVELET FUNCTIONS AND MRA 22
2.7.1 Haar wavelet 22
2.7.2 Faber–Schauder wavelet 24
2.7.3 Daubechies wavelets 25
2.7.4 Symlet wavelets 27
2.7.5 Spline wavelets 27
2.7.6 Anisotropic wavelets 29
2.7.7 Cauchy wavelets 30
2.8 EXERCISES 31
v
vi Contents
4.1 INTRODUCTION 51
4.2 DIFFERENT CONSTRUCTIONS OF CLIFFORD ALGEBRAS 52
4.2.1 Clifford original construction 53
4.2.2 Quadratic form-based construction 53
4.2.3 A standard construction 54
4.3 GRADUATION IN CLIFFORD ALGEBRAS 56
4.4 SOME USEFUL OPERATIONS ON CLIFFORD ALGEBRAS 57
4.4.1 Products in Clifford algebras 57
4.4.2 Involutions on a Clifford algebra 58
4.5 CLIFFORD FUNCTIONAL ANALYSIS 60
4.6 EXISTENCE OF MONOGENIC EXTENSIONS 67
4.7 CLIFFORD-FOURIER TRANSFORM 70
4.8 CLIFFORD WAVELET ANALYSIS 76
4.8.1 Spin-group based Clifford wavelets 76
4.8.2 Monogenic polynomial-based Clifford wavelets 82
4.9 SOME EXPERIMENTATIONS 92
4.10 EXERCISES 96
5.1 INTRODUCTION 99
5.2 BESSEL FUNCTIONS 99
5.3 BESSEL WAVELETS 105
5.4 FRACTIONAL BESSEL WAVELETS 107
5.5 QUANTUM THEORY TOOLKIT 119
Contents vii
Bibliography 209
Index 237
List of Figures
6.1 The normal reduced and centered density N (0, 1). 158
6.2 The gaussian density and its wavelet estimator. 159
6.3 Claw density. 159
6.4 Claw density wavelet estimator at the level J = 4. 160
ix
Preface
Nowadays, wavelets are applied almost everywhere in science. Both pure fields, such
as mathematics and theoretical physics, and applied ones, such as signal/image pro-
cessing, finance and engineering, apply wavelets. Although the references and/or the
documentation about wavelets and their applications are wide, it seems that with the
advancement of technology and the appearance of many phenomena in nature and in
life there still exist some places for more efforts and developments to understand the
new problems, as the existing wavelet methods do not provide good understanding
of them. The new COVID-19 pandemic may be one of the challenges that should be
understood.
On the other hand, especially for young researchers, existing references such as
books in wavelet theory are somehow very restricted. The majority are written for
specific communities. This is, in fact, not surprising and may be due to the necessity
of developing such references to overcome the concerned problems in that time.
Next, with the inclusion of wavelet theory in academic studies such as in mas-
ter’s and PhD programs, the scientific and academic communities have had a great
need to develop references in other forms. Students and generally researchers need
sometimes self-containing references responding to their need, to avoid losing time
in redeveloping existing results, which is a necessary step for both the generalization
and the experiments.
The present volume is composed of eight chapters. In the first introductory chap-
ter, a literal introduction is developed discussing generally the topic. Chapter 2 is
concerned with the presentation of the original developments of wavelet theory on
the real Euclidean space. This is also a preliminary chapter that will be of great help
for young researchers. Chapter 3 is more specialized and constitutes a continuation of
the previous one, in which some extending cases of wavelet theory and applications
have been provided. Chapter 4 is a very specialized part that is developed for the
first time to our knowledge. It is concerned with the presentation of wavelet theory in
a general functional framework based on Clifford algebras. This is very important as
these algebras contain all the Euclidean structures and gather them in one structure
to facilitate calculus. Readers will notice clearly that Clifford wavelet theory induces
naturally the Euclidean ones such as real and complex numbers, circles and spheres.
Chapter 5 is a continuation of the development of the theory in specialized fields
such as quantum theory. Next, in Chapter 6, statistical application of wavelets has
been reviewed. Topics such as density estimation, thresholding concepts, variance and
covariance have been detailed. Chapter 7 is devoted to wavelets applied in solving
partial differential equations. Recall that this field needs many assumptions on the
functional bases applied, especially the explicit form of the basis elements and their
xi
xii Preface
regularities. The last chapter is devoted to the link and/or the use of wavelet theory
in characterizing fractal and multifractal functions and their application. Each chap-
ter contains a series of exercises and experimentations to help understand the theory
and also to show the utility of wavelets.
The present book stems, in fact, from lectures and papers on the topics developed,
which have been gathered, re-developed, improved and sometimes completed with
necessary missing developments. However, naturally it is not exhaustive and should
be always criticized, sometimes corrected and improved by readers. So, we accept and
wait for any comments and suggestions.
We also want to stress the fact that we have provided in some chapters, especially
those on preliminary concepts that may be useful to young researchers, some exercises
and applications that are simple to handle with the aim to help the readers understand
the theory. We apologize if there are simpler applications and details that may be
more helpful to the readers but that have been left out from inclusion in this book.
This, in fact, needs more time and may induce delays in the publication of the book.
We hope that with the present form the readers become acquainted with the topics
presented.
The aim of this book is to provide a basic and self-contained introduction to
the ideas underpinning wavelet theory and its diversified applications. Readers of
our proposed book would include master’s degree students, PhD students and se-
nior researchers. It may also serve scientists and research workers from industrial
settings, where modeling real-world phenomena and data needs wavelets such as fi-
nance, medicine, engineering, transport, images and signals. Henceforth, the book
will interest practitioners and theorists alike. For theorists, rigorous mathematical
developments will be presented with necessary prerequisites that make the book self-
containing. For the practitioner, often interested in model building and analysis, we
provide the cornerstone ideas.
As with any scientific production and reference, the present volume could not have
been realized without the help of many persons. We thus owe thanks to many persons
who have helped us in any direction such as encouragements, scientific discussions
and documentation. We thank the Taylor & Francis Publishing Group for giving us
the opportunity to write and publish the present work. We also would like to express
our gratitude to our professors, teachers, colleagues, and universities. Without their
help and efforts, no such work might be realized. We would also like to thank all the
members of the publishing house, especially the editorial staff for the present volume,
Callum Fraser and Mansi Kabra, for their hospitality, cooperation, collaboration and
for the time they have spent on our project.
CHAPTER 1
Introduction
1
2 Wavelet Analysis: Basic Concepts and Applications
the so-called isotropic and anisotropic wavelets (see [11], [12], [20], [205], [218], [219],
[220], [229], [230], [248], [279], [295], [305], [357], [358], [363]).
Wavelet theory provides for functional spaces and time series good bases, allow-
ing their decomposition into spaces associated with different horizons known as the
levels of decomposition. A wavelet basis is a family of functions obtained from one
function known as the mother wavelet, by translations and dilations. Due to the
power of their theory, wavelets have many applications in different domains such as
mathematics, physics, electrical engineering and seismic geology. This tool permits
the representation of L2 -functions in a basis well localized in time and in frequency.
Wavelets are also associated with many special functions such as orthogonal poly-
nomials and hypergeometric series. The most well known may be the Bessel functions
that have been developed in both classic theory of Bessel functional analysis and the
modified versions in fractional and quantum calculus. As its name indicates, Bessel
wavelets are related to Bessel special function. Historically, special functions differ
from elementary ones such as powers, roots, trigonometric and their inverses, mainly
with the limitations that these latter classes are known for. Many fundamental prob-
lems such as orbital motion, simultaneous oscillatory chains and spherical body grav-
itational potential were not best described using elementary functions. This makes
it necessary to extend elementary functions’ classes to more general ones that may
describe well unresolved problems.
Wavelets are also developed and applied in financial time series such as market
indices and exchange rates. In [42], for example, a study of the largest transaction
financial market was carried out. The exchange market gave some high-frequency
data. Compared to other markets, such data can be available at long periods and
with high frequency. The data were detected for very small periods, which means
that the market is also liquid. Until 1990, economists were interested in intra-daily
data because of which the detection of some behaviors did not appear in the daily
analysis of data such as homogeneity.
A well-known hypothesis in finance is the homogeneity of markets where all in-
vestigators have almost the same behavior. The idea of nonhomogeneous markets is
more recent, and it suggests that investigators have different perceptions and differ-
ent laws. For the exchange market, for example, investigators can differ in profiles,
geographic localizations and also in institutional constraints. Another natural sug-
gestion can be done about traders. Naturally, traders investigating at short time in-
tervals allow some high-frequency behaviors in the change market. Long-time traders
are interested in the general tendency and the volatility of the market along a mi-
croscopic greed. Short-time traders, however, are interested in fractional perceptions
and so in macroscopic greed. This leads to the wavelet analysis of financial time
series.
Recently, other models have been introduced in modeling financial time series
by means of fractals, which are in turn strongly related to wavelets. For example,
in Olsen & Associates, operating the largest financial database, has noticed that the
tick frequency has strongly increased in one decade, causing problems in studying the
time series extracted from such a database. Such problems can be due to transmission
delays, input errors and machine damages. So, some filtering procedure has to be done
Introduction 3
before using the data. The first point that one must take into account in filtering time
series is their scaling behavior. Scaling laws were empirically observed by Olsen et al
until 1990 (see [306]). A time series X(t) has a scaling law if its so-called partition
function has the form
N/t
t X
Sq (t) = |X(jt)|q ∼ tτq +1
N j=1
where N stands for the size of the series for some appropriate function τ . This estima-
tion is well understood when merging wavelet tools into fractal models. (See Chapter
8.) In such models, Fourier analysis could not produce good results in estimating such
behavior. Indeed, Fourier transform of time series is generally limited because a single
analysis window cannot detect features in the signals that are either much longer or
shorter than the window size. Moving-window Fourier transform (MWFT) slides a
fixed-size analysis window along the time axis and is able to detect non-stationarities.
The fixed-size window algorithm of MWFT limits the detection of cycles at wave-
lengths that are longer than the analysis windows, and non-stationarities in short
wavelengths (i.e., high frequencies) are smoothed. Use of the wavelet transform solves
this problem, because it uses narrow windows at high frequencies and wide windows
at low frequencies.
This book is devoted to developing the basic concepts of wavelet analysis neces-
sary for young researchers doing their Master’s level in science and researchers doing
doctoral studies in pure mathematical/physical sciences, as well as applied and inter-
acted ones by providing the basic tools required, with simple and rigorous methods.
It also aims to serve researchers at advanced levels by providing them the necessary
tools that will allow them to understand and adapt wavelet theory to their needs
such as supervision and development of research projects.
The book provides some highly flexible methods and ideas that can be manip-
ulated easily by undergraduate students, and thus may be of interest for Bachelors
in science by providing them a clear idea on what wavelets are, and thus permitting
them to decide in their scientific future.
Wavelets on Euclidean
Spaces
2.1 INTRODUCTION
Wavelets were discovered in the eighteenth century in petroleum exploration, and
since their discovery, they have proven to be powerful tools in many fields from
pure mathematics to physics to applied ones such as images, signals, medicine, fi-
nance and statistics. The study of their constructions and their properties, especially
in functional/signals decompositions on functional wavelet bases, has indeed grown
considerably.
In pure mathematics, wavelets constitute a refinement of Fourier analysis as they
compensate and/or resolve some anomalies in Fourier series. The first wavelet basis
has been, in fact, used before the pure mathematical discovery of wavelet bases since
the introduction of Haar system, which by the next has been proved indeed to be
a possible wavelet basis reminiscent of the regularity. Such a system dates back to
the beginning of the 20th century and was precisely discovered in the year 1909. It
was introduced in order to construct a functional basis permitting the representation
of all continuous functions by means of a uniformly convergent series. Recall that in
Fourier series, there are, as usual, many kinds of convergence that may be investigated
such as the point-wise convergence subject of the well-known Dirichlet theorem, the
uniform convergence which needs more assumptions on the series and the function,
the convergence in norm and Carleson’s or almost everywhere convergence. Each kind
of convergence requires special assumptions on the function. Although a Fourier series
converges in sense of Dirichlet, it does not imply that the graph of the partial Fourier
series converges to that of the function. This phenomenon is known as the Gibbs
phenomenon and is related to the presence of oscillations in the Fourier series near
the discontinuity points of the function. This means that the uniform convergence
is not sufficient. One of the challenging concepts in wavelet analysis is its ability to
describe well the behavior of the analyzed function near its singularities and join or
more precisely extend the notion of Littlewood-Palay decomposition.
To overcome some drawbacks of Fourier analysis, mathematicians have intro-
duced a bit of modification called the windowed Fourier transform by computing the
5
6 Wavelet Analysis: Basic Concepts and Applications
original Fourier transform of the analyzed signals on a special localized extra function
called the window. However, some situations remain non resolved especially with the
emergence of irregular signals or high-frequency variations. The major problem in the
windowed Fourier extension is due to the use of fixed window, which may not be well
adapted to other problems such as high fluctuations of non stationary signals. This
led researchers to think about a stronger tool taking into account nonlinear algo-
rithms, nonstationary signals, nonperiodical, volatile and/or fluctuated ones. It holds
that wavelets, since their discovery, have permitted to overcome these obstacles.
These powers are related simultaneously to many properties of wavelets. Indeed,
wavelet decomposition of functions joins Littlewood-Paley decomposition in many
cases. Wavelets provide simultaneous local analyses related to time–frequency. They
can be adapted to study-specific operators, especially differential and stochastic ones.
From the numerical and/or applied point of view, wavelets provide fast and accurate
algorithms, multi-resolution analyses as well as recursive schemes. These are very
important especially in big data analysis, image processing and also in numerical
resolution of partial differential equations.
In this chapter, we propose to review the basic concepts of wavelets as well as
their basic properties.
2.2 WAVELETS ON R
The first wavelet bases and thus analyses have been constructed on the real line R and
have been next extended to the general cases of the real/complex Euclidean spaces
Rm -Cm using different methods such as the natural tensor product.
Mathematically speaking, a wavelet or an analyzing wavelet on the Euclidean
space Rm may be defined in a large way as a function with specific properties that
may or may not be required necessarily as simultaneous assumptions. More precisely,
wavelet analysis is based primarily on the following points:
• An effective representation for standard functions,
• Robustness to the specification models,
• A reduction in the computation time,
• Simplicity of the analysis,
• An easy generalization and efficient, depending on the dimension,
• A location in time and frequency.
A wavelet is a function ψ ∈ L2 (R) that satisfies the following conditions:
• Admissibility,
dω
Z
|ψ̂(ω)|2 = Cψ < ∞. (2.1)
R+ |ω|
• Zero mean, Z +∞
ψ(0)
b = ψ(u)du = 0. (2.2)
−∞
Wavelets on Euclidean Spaces 7
By varying the parameters s and u, we may cover completely all the time–frequency
plane. This gives a full and redundant representation of the whole signal to be an-
alyzed (see [295]). This transform is called continuous because of the nature of the
parameters s and u that may operate at all levels and positions. The original signal
S can be reproduced knowing its CWT by the following relationship:
1 x−u dsdu
Z Z
S(t) = du,s (S)ψ . (2.7)
Cψ R s s2
It remains to notice that CWT is suitable for continuous time signals and those
representing varying singularities.
The function ψ ∈ L2 (R) satisfies some conditions such as the admissibility con-
dition and somehow describes Fourier-Plancherel identity and says that
dω
Z
|ψ̂(ω)|2 = Cψ < ∞. (2.8)
R+ |ω|
8 Wavelet Analysis: Basic Concepts and Applications
The function ψ has to also satisfy a number of vanishing moments, which is related
in wavelet theory to its regularity order. It states that
Z
p = 0, ..., m − 1, ψ(t)tp dt = 0. (2.9)
R
To analyze a signal by wavelets, one passes via the so-called wavelet transforms. A
wavelet transform is a representation of the signal by means of an integral form similar
to Fourier one in which the Fourier sine and/or cosine is replaced by the analyzing
wavelet ψ. In Fourier transform, the complex exponential source function yields the
copies eis. indexed by s ∈ R, which somehow represent frequencies. This transform
is continuous in the sense that it is indexed on the whole line of indices s ∈ R.
In wavelet theory, the situation is more unified. A CWT is also well known.
Firstly, a frequency, scale or a dilation or compression parameter s > 0 and a second
one related to time or position u ∈ R have to be fixed. The source function ψ
known as the analyzing wavelet is next transformed to yield some copies (replacing
the eis. ):
1 x−u
ψs,u (x) = √ ψ . (2.11)
s s
The CWT of a real-valued function f defined on the real line at the position u and
the scale s is defined by
Z ∞
ds,u (f ) = f (t)ψs,u (t)dt, ∀ u, s. (2.12)
−∞
By varying the parameters s and u, we cover completely all the time–frequency plane.
This gives a full and redundant representation of the whole signal to be analyzed (see
[295]). This transform is called continuous because of the nature of the parameters s
and u that may operate at all levels and positions.
So, wavelets operate according two parameters: the parameter u, which permits
translation of the graph of the source mother wavelet ψ, and the parameter s, which
permits compression or dilation of the graph of ψ. Computing or evaluating the
coefficients du,s means analyzing the function f with wavelets.
Theorem 2.1 The wavelet transform ds,u (f ) possesses some properties such as
and where
(τt f )(x) = f (x − t).
The proof of these properties is easy and the readers may refer to [163] for a review.
In wavelet theory, as in Fourier analysis theory, the original function f can be
reproduced via its CWT by an L2 -identity.
1 x − u dsdu
Z Z
f (x) = ds,u (f )ψ( ) 2 .
Cψ s s
dsdu
Z Z Z
ds,u (f )ds,u (g) = Cψ f (x)g(x)dx, ∀ f, g ∈ L2 (R).
s
Proof. We have
1 x−u 1
Z
{ fˆ(y)ψ̂(sy)e−iuy .
ds,u (f ) = f ∗ ψs (u) = √ f (x)ψ( )dx =
s s 2π
Consequently
1
Z Z
ds,u (f )ds,u (g)du = fb(y)ĝ(y)|ψ̂(sy)|2 dy.
u 2π y
Next, using Fubini’s rule, we observe that the last line is equal to
1 dadb
Z Z Z
sup F (x)G(x)dx − Ca,b (F )Ca,b (G)
||G||=1 Cψ 1/A≤a≤A |b|≤B a
1 dadb
Z
= sup Ca,b (F )Ca,b (G)
||G||=1 Cψ (a,b)∈[1/A,A]×[−B,B]
/ a
Now, Lemma 2.3 shows that the last quantity goes to 0 as A, B tends to +∞.
These are often called wavelet coefficients or detail coefficients of the signal S.
It holds that the set (ψj,k )j,k∈Z constitutes an orthonormal basis of L2 (R) and it
is called wavelet basis. A signal S with finite energy (in L2 (R)) is then decomposed
according to this basis into a series
∞ X
X
S(t) = dj,k ψj,k (t) (2.14)
j=0 k
called the wavelet series of S which replaces the reconstruction formula for the CWT.
Wavelets on Euclidean Spaces 11
For each j ∈ Z, Vj is called the approximation space at the scale or the level j. More
precisely, we have the following definition ([295]).
a) ∀j ∈ Z; V0 ⊂ V1 ⊂ .... ⊂ Vj ⊂ Vj+1 .
b) ∀j ∈ Z; f ∈ Vj ⇔ f (2.) ∈ Vj+1
[
e) Vj = L2 (R).
j∈R
f) ∀j ∈ Z; f ∈ Vj ⇐⇒ f (x − k) ∈ Vj
The property (a) reflects that the approximation of a signal at the resolution’s level
j + 1 contains the necessary information to yield the approximation at the level j.
Assertion (b) called dilatation’s property permits passing from a level of resolution
to another. The space Vj+1 contains signals that are coarser than Vj . Assertion (c)
means that a scaling function exists and permits the decomposition of the signal at the
starting level 0. The property (d) means that at very low resolution level (2−j −→ 0
as j −→ +∞) we lose all the details of the signal. At a minimal resolution, we
lose all the information about the signal. Assertion (e) implies that the signal may
be approximated with elements in Vj . At a maximal resolution, we reconstruct all
the whole signal. Finally, the property (f) of translation means that the space Vj is
invariant under integer translation.
Definition 2.5 The source function ϕ is called the scaling function of the MRA or
also the father wavelet.
It holds that this function generates all the subspaces Vj ’s of the MRA by acting
as dilation/translation parameters. Indeed, the property
(d) combined with
(f) in
j/2 j
Definition 2.4 implies that, for all j, ∈ Z, the set ϕj,k (x) = 2 ϕ(2 x − k) is an
k
orthogonal basis of Vj . From assertion (a), it holds that we may complete Vj in Vj+1
12 Wavelet Analysis: Basic Concepts and Applications
in the sense of the direct sum. Let Wj be the orthogonal supplementary of Vj in Vj+1 ,
that is
Vj+1 = Vj ⊕⊥ Wj . (2.16)
We will see that Wj plays a primordial role in representing the details of the analyzed
signal. This is why it is called the detail space at the level j. Iterating the relation
(2.16), we obtain an orthogonal decomposition for all J ∈ Z,
J,⊥
VJ = V0 ⊕⊥
M
Wj . (2.17)
j=0
The function ψ is called the wavelet function or the mother wavelet associated with
the scaling function ϕ of the MRA.
We will now explain the relationship between ϕ and ψ. It holds from (2.18) that any
S ∈ L2 (R) is decomposed into components according to the subspaces Wj ’s in the
sense that
+∞
X
S= S Wj ,
j=−∞
Observing now that the wavelet coefficients are already defined by (2.12) or (2.13) as
Z +∞
dj,k (S) = S(t)ψj,k (t)dt =< S, ψj,k >
−∞
Wavelets on Euclidean Spaces 13
we obtain the wavelet series of S defined already by (2.14). Returning next to (2.16)
or also (2.17), we may write that
−1 X
X +∞
XX
S= < S, ψj,k > ψj,k + < S, ψj,k > ψj,k . (2.19)
j=−∞ k j=0 k
It results from the nesting property of the Vj ’s and the fact that Wj ⊂ Vj+1
that the first part in (2.19) is an element of V0 which is generated by the basis
(ϕk (t) = ϕ(t − k))k . Consequently, we may also write that
−1 X
X X
< S, ψj,k > ψj,k = Ck ϕk .
j=−∞ k k
In fact, it may truncate at any level J and apply the ϕJ,k instead of ϕk to obtain
X ∞ X
X
S(x) = CJ,k ϕJ,k (x) + dj,k ψj,k (x) (2.21)
k j=J k
which is known as the wavelet series decomposition of S at the level J. In fact, the
coefficients that appears in this decomposition are evaluated via the relation
Z +∞
CJ,k = S(t)ϕJ,k (t)dt
−∞
and are called the scaling or approximation coefficients of S. The first component
in VJ reflects the global behavior or the tendency of the whole signal S and the
second component relative to the dj,k ’s represents the details of S and thus reflects
the dynamic behavior of the signal.
gk = (−1)k h1−k .
For more details, we refer to [177], [236], [295]. We will now prove that these relations
allow obtaining all the decomposition of a signal from each other through specific
algorithms. Indeed, consider a signal S and its approximation coefficients Cj,k and
details dj,k . For j ∈ Z, we have
Hence,
X
Cj,k = hl Cj+1,2k+l .
l
This means that the approximation at level j is obtained from the level j + 1 by the
intermediate of a filter. We have, in fact, the following definition.
Definition 2.9 The sequence H = (hk )k is called the discrete low-pass filter.
Hence, X
dj,k = hS, ψj,k i = gl Cj+1,l+2k .
l
This means that the approximation at level j is obtained from the level j + 1 by the
intermediate of a filter. We have here also the following definition.
Definition 2.10 The sequence G = (gk )k is called the discrete high-pass filter.
Figure 2.1 presents the decomposition algorithm due to Mallat [295]. It consists of a
cascade algorithm permitting to obtain all the levels of resolution.
We will now explain the inverse algorithm which permits to obtain the level j + 1
from the level j. Recall that the orthogonal projection of S on the approximation
space Vj is given by (Hardle et al (1997) [236])
X X
SVj = Cj,k ϕj,k + dj,k ψj,k
k k
Wavelets on Euclidean Spaces 15
Cj+1
H G
↓2 ↓2
Cj dj
H G
↓2 ↓2
Cj−1 dj−1
and that
or else X X
Cj+1,n = hn−2k Cj,m + gn−2k dj,m . (2.23)
k k
3. gn = (−1)n hl−n ; ∀ n.
X
4. hn gn+2j = 0, ∀ j ∈ N. (called mutual orthogonality).
n
16 Wavelet Analysis: Basic Concepts and Applications
Cj−1 dj−1
↑2 ↑2
H G
Cj dj
↑2 ↑2
H G
Cj+1
As a result,
1 c 2 1
Z 2π X
kFϕ k2L2 (R) = kFϕ kL2 (R) = | λk eikξ |2 Γϕ (ξ)dξ.
2π 2π 0 k
1 X
Now, observing that the map Φ : λ = (λk ) → Φ(λ) = √ λk eikξ is an iso-
2π
morphism from `2 (R) topL2 ([0, 2π]), to get the proposition, the map K defined on
L2 ([0, 2π]) by K(f ) = f Γϕ should also be an isomorphism. This is the case if and
1
only if Γϕ and are both bounded (in L∞ ([0, 2π])).
Γϕ
Sometimes we need more regularity on the wavelet basis especially in the case of par-
tial differential equations. Besides, we sometimes need compactly supported wavelets
and/or wavelets with fast decay which are more adaptable and suitable for boundary
and/or limit conditions due to partial differential equations.
It is sometimes necessary to assume that the scaling function ϕ and its derivatives
to an order r (usually the order of the MRA) are of fast decay. This permits to
avoid the assumption on the overlap function Γϕ of ϕ to be upper bounded, as this
characterization becomes already holding. Indeed, whenever ϕ is of fast decay, its
Fourier transform ϕb is therefore infinitely continuously differentiable and is also of
fast decay. It satisfies, in particular,
K
|ϕ(ξ)|
b ≤
1 + |ξ|
18 Wavelet Analysis: Basic Concepts and Applications
Proof 2.2 Assume that the trigonometric series m is analytic on the annulus
|Imω| < R. Observe next that m is also periodic with period 2π. As a consequence,
the function ω 7−→ m(ω + 2π) − m(ω) is vanishing on all the annulus. Now applying
the complex integral, we get
Z 2π Z −iR+2π
λk = m(ω)e−ikω dω = m(z)e−ikz dz
0 −iR
Consequently,
|λk | ≤ Ce−kR .
We will now show that in Definition 2.4 some properties and/or assumptions may
be deduced from others, which thus reduces the number of assumptions and gives
more flexibility in the introduction of multiresolution analyses. We have precisely the
following result.
Properties 2.14 Assume that assertions (a), (b) and (c) in Definition 2.4 hold and
denote Φ, the function defined by its Fourier transform
1
b + 2kπ)|2 )− 2 .
X
Φ(ω)
b = ϕ(ω)
b |ϕ(ω
k
It holds that Z +∞ X
ϕ(x)dx = 1 and ϕ(x − k) = 1.
−∞ k
where the coefficients hj,k are the analogues of the filter coefficients hk in the two-scale
relation. More precisely,
Z +∞ 1
hj,k = ϕ(x)2 2 ϕ(2j x − k)dx.
−∞
Z +∞
Denote next Iϕ = ϕ(x)dx. We may write
+∞
k
Z +∞ k j
− 2j
hj,k − Iϕ 2 ϕ( j ) = hj,k − ϕ( )2 2 ϕ(2j x − k)dx
2 +∞ 2j
Z +∞ k j
≤ C |x − |2 2 |ϕ(2j x − k)|dx
+∞ Z 2j
+∞
−3 2j
≤ C2 |x||ϕ(x)|dx
+∞
−3 2j
≤ C2 .
As a result,
j k j
hj,k = Iϕ 2− 2 ϕ( j
) + O(2−3 2 ).
2
Consequently,
k j
)ϕ(2j x − k) + O(2−3 2 ).
X
ϕ(x) = Iϕ ϕ( j
k
2
We now claim that Iϕ = 1. Indeed, let x0 be such that ϕ(x0 ) 6= 0. Let δ, η > 0 be
such that |ϕ(x) − ϕ(x0 )| ≤ whenever |x − x0 | ≤ η. Therefore,
X X
ϕ(x0 ) − Iϕ ϕ(x0 )ϕ(2j x − k) ≤ Iϕ |ϕ(2j x − k)|
k | kj −x0 |≤η
2
k
X
+ |ϕ( )| + |ϕ(x0 )| |ϕ(2j x − k)|
k
2j
| −x0 |>η
2j
j
+C + (2−3 2 ).
η k
Observe next whenever |x − x0 | ≤ then we necessarily get | j − x0 | > η, which
2 2
k η
implies that | j − x| ≥ . Hence,
2 2
C C
|ϕ(2j x − k)| ≤ ≤ .
(1 + |2j x − k|)4 (1 + |2j x − k|)2 (2j η2 )2
As a consequence,
X X 1 −3 2j
ϕ(x0 ) − Iϕ ϕ(x0 )ϕ(2j x − k) ≤ C η 2 + C + O(2 )
k k
(1 + |2j x 2 j
− k|) (2 2 )
j
≤ C2 −2j −2
η + C + O(2−3 2 ).
20 Wavelet Analysis: Basic Concepts and Applications
X 1
ϕ(y − k) − ≤ C.
k
Iϕ
X
Now observe that the function y 7−→ ϕ(y − k) is 1-periodic. Hence, for j large
k
enough, we get that such a function remains constant on the whole line R, which
means that
X 1
ϕ(y − k) = .
k
M
This yields
Z +∞ XZ k+1 Z 1 X 1
M= ϕ(y)dy = ϕ(y)dy = ( ϕ(y − k))dy = .
∞ k∈Z k 0 k∈Z
M
Hence, M 2 = 1 and \
thus may be chosen to be equal to 1.
We now prove that Vj = {0}. Let F be an element in the intersection. Hence, for
j
all j, we may write X
F = aj,k ϕj,k ,
k
where the aj,k are uniformly bounded. Moreover,
X
kF k2 = |aj,k |2 .
k
Consequently,
j
kχ[a,b] − Fj k22 ≤ C2− 2 .
Letting j → ∞ we observe Fj goes to χ[a,b] in L2 .
The adoption of these one-dimensional MRA and wavelets will be the starting point
to construct the multidimensional case. Indeed, let (Vj1 ) be a multiresolution analysis
of L2 (R) with a scaling function ϕ, a wavelet ψ and an orthogonal projections Pj1 .
Consider then orthogonal projection Pjd in L2 (Rd ) defined as the tensor product of d
copies of Pj1
Pjd = Pj1 ⊗ Pj1 ⊗ ... ⊗ Pj1 .
22 Wavelet Analysis: Basic Concepts and Applications
Denote Vjd = Pjd (L2 (Rd )). We have Vjd = Vj1 ⊗ Vj1 ⊗ ... ⊗ Vj1 . The closure in L2 (Rd )
of Vjd has an orthonormal basis
(
ϕdj,k = 2jd/2 ϕd (2j x − k) ; j ∈ R , k ∈ Rd
ϕd = ϕ ⊗ ϕ ⊗ ... ⊗ ϕ
Having in hand the explicit expression of Haar scaling function ϕ = χ[0,1[ , the latter
becomes √ X
χ[0,1[ (x) = 2 hk χ[k/2,(k+1)/2[ (x).
k
Next, because of the supports of the different functions appearing in the relation, we
obtain
hk = 0 , for all k 6= 0, 1
and thus, √ √
χ[0,1[ (x) = 2h0 χ[0,1/2[ (x) + 2h1 χ[1/2,1[ (x)
which yields immediately that
1
h0 = h1 = √ .
2
It results that the filter G is given by
1
g0 = −g1 = √ and gk = 0 , for all k 6= 0, 1.
2
The associated Haar mother wavelet is then expressed by
ψ = χ[0,1/2[ − χ[1/2,1[ .
We hereafter give the graphic illustrations of the Haar scaling function ϕ and the
Haar mother wavelet ψ (see Figure 2.3).
24 Wavelet Analysis: Basic Concepts and Applications
The subspace Vj is composed of square integrable functions that are of the form
j/2 j j j
aj,k 2 (1 − k + 2 x) , x ∈ [(k − 1)/2 , k/2 [,
j/2 j j j
f (x) = aj,k 2 (1 + k − 2 x) , x ∈ [k/2 , (k + 1)/2 [,
0 , others.
We now evaluate the filter coefficients. The two-scale relation (2.22) is written in the
present case as
√ X2
(1 − |x|)χ[−1,1[ (x) = 2 hk (1 − |2x − k|)χ[(k−1)/2,(k+1)/2[ (x).
k=−2
The coefficients hk are vanishing for |k| ≥ 3. Next by choosing for x the values −1,
−1 1
, 0, and 1, we obtain
2 2
1 1
h−2 = h2 = 0 , h−1 = h1 = √ and h0 = √ .
2 2 2
The G filter will be
1 −1
g0 = g2 = √ and g1 = √ .
2 2 2
The Schauder mother wavelet will be expressed by
1 1
ψ(x) = (1 − |2x|)χ[−1/2,1/2[ (x) − (1 − |2x − 1|)χ[0,1[ (x) + (1 − |2x − 2|)χ[1/2,3/2[ (x).
2 2
The following graph (Figure 2.4) represents Schauder ϕ and ψ.
∀ x, Φ(x) = Σ hk Φ(2x − k)
26 Wavelet Analysis: Basic Concepts and Applications
and √ X
∀ x, ψ(x) = 2 g(n)φ(2x − k)
where X
h(n − 2k)h(n − 2l) = δkl ,
n
X √
h(n) = 2,
n
and
g(n) = (−1)n h(1 − n)
If the filter H = (h(n))n∈Z is finite, the support of the scaling function is also finite.
Let
supp(φ) ⊂ [Nmin , Nmax ].
Hence,
For example for N = 2, we obtain the so-called Db4 wavelet, whose low-pass filter
coefficients hk are √ √
1+ 3 3+ 3
h0 = √ , h1 = √
4 2 4 2
√ √
3− 3 1− 3
h2 = √ , h3 = √
4 2 4 2
The next graph illustrates some wavelet function psi due to Daubechies. The repre-
sentation illustrates clearly the strong relation between the regularity, the number of
vanishing moments and the support length N .
Wavelets on Euclidean Spaces 27
We will start by the most simple example of wavelets and multiresolution analysis
due to the Haar case and show that, even though this is a nonregular case, it may serve
to construct higher order multiresolutions. These are the so-called spline wavelets.
So, let ϕ = χ[0,1] be the Haar scaling function, which is equivalently the char-
acteristic function of the interval [0, 1]. Let r > 0 be in N. The spline wavelets are
defined inductively by
ϕ0 = ϕ and ϕr = ϕr−1 ∗ ϕ, ∀r ≥ 1.
ψ(X) = (2 − XAX)e−XAX/2
with m > 0. In one dimension, the positive half-line is a convex cone. Thus, a natural
generalization to two dimensions will be a wavelet whose support in spatial frequency
space is contained in a convex cone with apex at the origin. Let C ≡ C(α, β) be the
convex cone determined by the unit vectors eα , eβ , where α < β, β − α < π and for
all θ, eθ ≡ (cos θ, sin θ). The axis of the cone is ξαβ = e α+β . In other words,
2
C(α, β) = k ∈ R2 , α ≤ arg(k) ≤ β
= k ∈ R2 , k.ξαβ ≥ eα .ξαβ = eβ .> 0 .
C(α,
e β) = C(α
b , β).
b
where
π b π
b=β−
α , β =α+
2 2
and
eα .eαb = eβ .eβb = 0.
where η ∈ Ce and l, m ∈ N∗ . Note that such a wavelet is also supported by the cone
C. It satisfies the admissibility condition
d2 k
Z
C,η
AψC,η ≡ (2π)2 |ψblm (k)|2 < ∞.
lm |k|2
The following result of Antoine et al is proved in [13] and yields an explicit form for
the two-dimensional Cauchy wavelet.
Wavelets on Euclidean Spaces 31
2.8 EXERCISES
Exercise 1.
√
Denote φ = χ[0,1[ and φjk (t) = 2j φ(2j t − k), t ∈ R. Denote further E = L2 ([0, 1[)
the vector space of square-integrable functions on [0, 1[. Next for j fix V j , the vector
k k+1
space of constant functions, on { j ; j } for k ∈ {0; 2j − 1}.
2 2
a. Show that (φjk )k∈{0,2j −1} is an orthonormal basis of V j . < (φjk ); (φjh ) >= 0∀(k, h) ∈
{0, 2j − 1} × {0, 2j − 1}. and that the elements φjk are unitary.
Exercise 2.
Consider the same assumptions as in Exercise 1 above.
a. Find two functions (ψk1 )k∈{0,1} constituting an orthonormal basis of a vector space
W 1 such that V 2 = V 1 ⊕ W 1 .
b. Express the functions (ψk1 )k∈{0,1 } and (φ1k )k∈{0,1 } by means of (φ2k )k∈{0,3 }.
c. Express conversely the functions (φ2k )k∈{0,3 } in terms of (φ1k )k∈{0,1 } and
(ψk1 )k∈{0,1 }.
Exercise 3.
Consider the same assumptions as in Exercises 1 and 2 above.
a. Provide the 2j Haar wavelets (ψk1 )k∈{0,1 } which constitute an orthonormal basis
of W j = V j+1 V j .
d. Write the approximation coefficients ajk and the detail coefficients (djk ) by means
of the (aj+1
k ).
32 Wavelet Analysis: Basic Concepts and Applications
Exercise 4.
Consider the discrete signal S = [2 4 8 12 14 0 2 1] and the Haar multiresolution
analysis on R.
a. Decomposer S in V 0 ⊕2j=0 W j
b. Sketch the graphs of S relatively to V 2 , V 1 and V 0 .
b. Sketch the graphs of S relatively to the details W 2 , W l and W 0
Exercise 5.
Let ψ : R → C be wavelet in the Schwarz class such that ψ(0)
b = 0). Denote
Z
C= |xψ(x)|dx.
R
Exercise 6.
Let f, ψ : R → C be in the Schwarz class such that ψ(0) = 0) and denote
Z +∞ Z
dt
F (t) = ψs (t − u)Wf (u, s)du .
0 R s2
Show that F = Cψ f , where Cψ is a constant depending on ψ and not on f .
Exercise 7.
Let ψ be a wavelet such that ψ(0)
b = 0, where
Z
ψ(w)
b = ψ(t)e−izt dt, ∀w ∈ R.
R
Another random document with
no related content on Scribd:
had not loved the children with all the devotion he should have or he
would never have had the thoughts he had had—or so he had
reasoned afterward. Yet then as now he suffered because of the love
he should have given them, and had not—and now could not any
more, save in memory. He recalled how both boys looked in those
last sad days, their pinched little faces and small weak hands! Marie
was crushed, and yet dearer for the time being than ever before. But
the two children, once gone, had seemed the victims of his own dark
thoughts as though his own angry, resentful wishes had slain them.
And so, for the time, his mood changed. He wished, if he could, that
he might undo it all, go on as before with Marie, have other children
to replace these lost ones in her affection—but no. It was apparently
not to be, not ever any more.
For, once they were gone, the cords which had held him and Marie
together were weaker, not stronger—almost broken, really. For the
charm which Marie had originally had for him had mostly been
merged in the vivacity and vitality and interest of these two prattling
curly-headed boys. Despite the financial burden, the irritation and
drain they had been at times, they had also proved a binding chain,
a touch of sweetness in the relationship, a hope for the future, a
balance which had kept even this uneven scale. With them present
he had felt that however black the situation it must endure because
of them, their growing interests; with them gone, it was rather plain
that some modification of their old state was possible—just how, for
the moment, he scarcely dared think or wish. It might be that he
could go away and study for awhile now. There was no need of his
staying here. The neighborhood was too redolent now of the
miseries they had endured. Alone somewhere else, perhaps, he
could collect his thoughts, think out a new program. If he went away
he might eventually succeed in doing better by Marie. She could
return to her parents in Philadelphia for a little while and wait for him,
working there at something as she had before until he was ready to
send for her. The heavy load of debts could wait until he was better
able to pay them. In the meantime, also, he could work and whatever
he made over and above his absolute necessities might go to her—
or to clearing off these debts.
So he had reasoned.
But it had not worked out so of course. No. In the broken mood in
which Marie then was it was not so easy. Plainly, since he had run
across her that April day in Philadelphia when he was wiring for the
great dry goods store, her whole life had become identified with his,
although his had not become merged with hers. No. She was, and
would be, as he could so plainly see, then, nothing without him,
whereas he—he—Well, it had long since been plain that he would be
better off without her—materially, anyhow. But what would she do if
he stayed away a long time—or never came back? What become?
Had he thought of that then? Yes, he had. He had even thought that
once away he might not feel like renewing this situation which had
proved so disastrous. And Marie had seemed to sense that, too. She
was so sad. True he had not thought of all these things in any bold
outright fashion then. Rather they were as sly, evasive shadows
skulking in the remote recesses of his brain, things which scarcely
dared show their faces to the light, although later, once safely away
—they had come forth boldly enough. Only at that time, and later—
even now, he could not help feeling that however much Marie might
have lacked originally, or then, the fault for their might was his,—that
if he himself had not been so dull in the first instance all these black
things would not have happened to him or to her. But could she go
on without him? Would she? he had asked himself then. And
answered that it would be better for him to leave and build himself up
in a different world, and then return and help her later. So he fretted
and reasoned.
But time had solved all that, too. In spite of the fact that he could
not help picturing her back there alone with her parents in
Philadelphia, their poor little cottage in Leigh Street in which she and
her parents had lived—not a cottage either, but a minute little brick
pigeon-hole in one of those long lines of red, treeless, smoky
barracks flanking the great mills of what was known as the
Reffington District, where her father worked—he had gone. He had
asked himself what would she be doing there? What thinking, all
alone without him—the babies dead? But he had gone.
He recalled so well the day he left her—she to go to Philadelphia,
he to Boston, presumably—the tears, the depression, the
unbelievable sadness in her soul and his. Did she suspect? Did she
foreknow? She was so gentle, even then, so trustful, so sad. “You
will come back to me, dearie, won’t you, soon?” she had said, and so
sadly. “We will be happy yet, won’t we?” she had asked between
sobs. And he had promised. Oh yes; he had done much promising in
his life, before and since. That was one of the darkest things in his
nature, his power of promising.
But had he kept that?
However much in after months and years he told himself that he
wanted to, that he must, that it was only fair, decent, right, still he
had not gone back. No. Other things had come up with the passing
of the days, weeks, months, years, other forces, other interests.
Some plan, person, desire had always intervened, interfered,
warned, counseled, delayed. Were there such counselors? There
had been times during the first year when he had written her and
sent her a little money—money he had needed badly enough
himself. Later there was that long period in which he felt that she
must be getting along well enough, being with her parents and at
work, and he had not written. A second woman had already
appeared on the scene by then as a friend. And then—
The months and years since then in which he had not done so!
After his college course—which he took up after he left Marie,
working his way—he had left Boston and gone to K—— to begin a
career as an assistant plant manager and a developer of ideas of his
own, selling the rights to such things as he invented to the great
company with which he was connected. And then it was that by
degrees the idea of a complete independence and a much greater
life had occurred to him. He found himself so strong, so interesting to
others. Why not be free, once and for all? Why not grow greater?
Why not go forward and work out all the things about which he had
dreamed? The thing from which he had extricated himself was too
confining, too narrow. It would not do to return. The old shell could
not now contain him. Despite her tenderness, Marie was not
significant enough. So—He had already seen so much that he could
do, be, new faces, a new world, women of a higher social level.
But even so, the pathetic little letters which still followed from time
to time—not addressed to him in his new world (she did not know
where he was), but to him in the old one—saying how dearly she
loved him, how she still awaited his return, that she knew he was
having a hard time, that she prayed always, and that all would come
out right yet, that they would be able to be together yet!—she was
working, saving, praying for him! True, he had the excuse that for the
first four years he had not really made anything much, but still he
might have done something for her,—might he not have?—gone
back, persuaded her to let him go, made her comfortable, brought
her somewhat nearer him even? Instead he had feared, feared,
reasoned, argued.
Yes, the then devil of his nature, his ambition, had held him
completely. He was seeing too clearly the wonder of what he might
be, and soon, what he was already becoming. Everything as he
argued then and saw now would have had to be pushed aside for
Marie, whereas what he really desired was that his great career, his
greater days, his fame, the thing he was sure to be now—should
push everything aside. And so—Perhaps he had become sharper,
colder, harder, than he had ever been, quite ready to sacrifice
everything and everybody, or nearly, until he should be the great
success he meant to be. But long before this he might have done so
much. And he had not—had not until very recently decided to revisit
this older, sweeter world.
But in the meantime, as he had long since learned, how the
tragedy of her life had been completed. All at once in those earlier
years all letters had ceased, and time slipping by—ten years really—
he had begun to grow curious. Writing back to a neighbor of hers in
Philadelphia in a disguised hand and on nameless paper, he had
learned that nearly two years before her father had died and that she
and her mother and brother had moved away, the writer could not
say where. Then, five years later, when he was becoming truly
prosperous, he had learned, through a detective agency, that she
and her mother and her ne’er-do-well brother had moved back into
this very neighborhood—this old neighborhood of his and hers!—or,
rather, a little farther out near the graveyard where their two boys
were buried. The simplicity of her! The untutored homing instinct!
But once here, according to what he had learned recently, she and
her mother had not prospered at all. They had occupied the most
minute of apartments farther out, and had finally been compelled to
work in a laundry in their efforts to get along—and he was already so
well-to-do, wealthy, really! Indeed three years before his detectives
had arrived, her mother had died, and two years after that, she
herself, of pneumonia, as had their children. Was it a message from
her that had made him worry at that time? Was that why, only six
months since, although married and rich and with two daughters by
this later marriage, he had not been able to rest until he had found
this out, returned here now to see? Did ghosts still stalk the world?
Yes, to-day he had come back here, but only to realize once and
for all now how futile this errand was, how cruel he had been, how
dreary her latter days must have been in this poor, out-of-the-way
corner where once, for a while at least, she had been happy—he
and she.
“Been happy!”
“By God,” he suddenly exclaimed, a passion of self-reproach and
memory overcoming him, “I can’t stand this! It was not right, not fair. I
should not have waited so long. I should have acted long, long since.
The cruelty—the evil! There is something cruel and evil in it all, in all
wealth, all ambition, in love of fame—too cruel. I must get out! I must
think no more—see no more.”
And hurrying to the door and down the squeaking stairs, he
walked swiftly back to the costly car that was waiting for him a few
blocks below the bridge—that car which was so representative of the
realm of so-called power and success of which he was now the
master—that realm which, for so long, had taken its meaningless
lustre from all that had here preceded it—the misery, the loneliness,
the shadow, the despair. And in it he was whirled swiftly and gloomily
away.
IX
PHANTOM GOLD