1992 - Meyer - Wavelets Algorithms and Applications
1992 - Meyer - Wavelets Algorithms and Applications
Algorith111s
Applications
Yves Meyer
CEREMADE
and
Institut Universit.aire de France
Translated a nd Rel'ised b1
Robert D . Ryan
Phibcldphi:l !0<>..;
by Roben D. Ryan.
p. em.
"Translation based on leQures onginally presented by the author
for the Spanish Institute in Madrid. Spain. Februa~ 1991"-CIP
galley.
ISBN 0-89871-309-9
I. Wavelets. I. Title.
QA403.3.M4913 1993
515'.2433-dc20
93-15 100
All rif:ht~ reser"ed. Pn ntcd in the Umted State~ of Amenca. !\<> pan of thi~ book may be reproduced.
stored. or tr.snsnuucd tn any manner w uhout t he wnuen pemussion of the Publisher. For mform~
tiC'In. wri te thl' StlCICt~ wr lndu~tn;tl and Applied lllathcmallt'>. 360(1 Uni versu~ Cit~ Science Center.
Philadelphia. Penn'~ h am.J J<11l~:!Mi!i
SlaJTL.
Contents
Translator's Foreword
Prerace
CHAPTER I.
1. 1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
CHAPTER 2.
2. 1
2.2
2.3
2.4
2.5
2.6
2.7
2 .8
2.9
2. 10
2. 11
CHAPTER 3.
3. 1
ix
xi
Introduction 33
Subband Codin{:: The Case of Ideal Filters 34
Qu3dra1Urt Mirror Fi lters 3(>
.~ .4 . The Trend and Auctuation
38
:<5 T he Timt>Scalc Algorithm of Mallat and the Time- Frequenc~ Algorithm Clf
Galnnd ~I}
.'.h Trend~ and Auctu:uion~ with OnhOnonnal Wavelet Ba.~~ 4(l
:>.i C'on\'ergcncc II' Wa\'eleL~ 4~
.'.!' Th.:- Wa\'clcts of Daubechie~ 4::<
~.::
~-~
:<.II
Conclu.~ion~
+l
,.
COl\'TE:>.'TS
CHAPTER 4.
4. 1
4.2
4.3
4.4
4.5
4.6
4.7
CHAPTER 5.
Time-Frequen~
5.1
5.~
5.3
5.4
5.5
5.6
5.7
5.8
CHAPTER 6.
6. 1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
6.9
CHAPTER 7.
7.1
72
7.3
7.4
75
CHAPTERS.
8.1
8.:!
8.~
i\.J
8.5
li.6
CHAPTER 9.
9.1
9 .2
93
9.J
95
ID
73
CHAPTER 10.
10. 1
10.2
10.3
10.4
10 .5
CHAPTER I I.
I I. I
11.2
I 1.3
11.4
11.5
129
119
126
Translator's Foreword
The question most often asked by tho~e who heard 1 was translating this book was, "How did you
get the job?'" Well. I asked for iL I mentioned to Professor Meyer in March 1992 that I had hem!
that he had wriuen a new book. He said. yes. and that the book was based on noces from Jeaures he
had given at the Spanish Institute in Madrid. He added that the book was being translated illlo
Spanish and that SIAM was interested in publishing an English edition. for which they would need a
translator. 1 volunteered to do the job: what you have is the result.
In addition to translating the text. I have tried to "work through" much of the mathematics to
correct typos. I have also added a line of explanation here and there where it seemed appropriate.
and sections of text and references have been updated. These revisions are not highlighted in any
panicular way (i.e. there are no translator's noces exc:ep~ at one reicreuce). but ratheT incorporated
into the text. These changes were made with Professor Meyer's blessing. Of course, there is
always the possibility that in the process of updating the manuscript I have introduced other errors:
for these I take full rel;ponsibility.
The great fun of this project has been the chance to work with Yves Meyer and other members
of "team wavelet." Ptofes.wr Meyer impm1ed both my French and my mathematics. and h is
enthusiasm and appreciation for m)' efforts kept things moving. Direct help also came from John
Benedetto. Marie Farge. Patrick Flandrin. Stephane Jaffard. and Hamid Krim. Alex Grossmann
gave moral suppon by assuring me that it was an important project. My sincere thanks to all of
these people. The work '''as done while I was a Liaison Scientist for the Office of Na~al Research
European Office. where my primary job was to report on mathematics in Europe. It was through
thi s work that I first made contact with the French wavelet communitv. and I thank the Office of
Naval Research for that opportunity. This was esse.ntially a weekenc:,-and evening project. and
hence a family project. In this conte xt. I thank my son. Michael J. Ryan. who kept house and
produced great d inners while I kept the electrons moving.
Roben D. Rvan
November 199:::
London. UK
"
Preface
The ..theory of wavelet.sM sunds at the intersection of the frontiers of mathematics, scientific
computin~. and signal processing. h s goal is to provide a coherent 5et of concepc.s, methods. and
al~orithms that are adapted to a variety of nonStationary signals and that are also suitabl.e for
numerical signal processing.
This book results from a series of lectures that Mr. Miguel Anoia Gallego, Director of the
Spanish lnstiwte. invited me to ~ive on waveletS and their applications. I have tried to fulfi ll. in the
followin~ pa~es. the objective the Spanish Institute set for me: to present to a scientific audience:
coming from different disciplines. the prospects that wavelets offer for signal and ima~e processing.
A description of the different al~orithms used today under the name "wavc:lets" (Chapters 2-7)
will be followed by an analysis of several applications of these methods: to numerical ima~e
proeessin~ (Chapter 8J, to fractals (Chapter 9 ), to turllulencc: (Chapcer 10), and to astronomy
(Chapter I I ). This v.ill take me out of my scientifiC domain: as a resuh, the last two chapters are
merely resumes of the ori~inal articles on which they are based.
I wish to thank the Spanish Institute for its generous hospitality as well as its Director for his
warm welcome. Additionally. I note the excellent orpnization by Mr. Pedro Corpas.
My thanks ~o also to my Spanish friends and eollea~ues who took the time: to anend these
lectures.
!l.i
CHAPTER
The purpose of this first chapter is to give the reader a fairly clear idea about
the scientific content of the following chapters. All of the "themes" that will be
developed in this study, using the inevitable mathematical formalism, already
appear in this "overture." It is written with a concern for simplic!_ty and clarity,
while avoiding as much as po6Sible the use of formulas and symbols.
Signal and image processing always leads to a collection of techniques or
procedures. But like all other scientific disciplines, signal and image processing assumes certain preliminary scientific conventions. We have sought, in this
first chapter, to describe the intellectual architecture underlying the algorithmic
constructions that will be presented in the following chapters.
1.1.
What is a signal?
technological activity. Signal processing is used in telecommunications (telephone and television). in the transmission and analysis of satellite images, and
in medical imaging (echography, tomography, and nuclear magnetic resonance),
all of which involve the analysis and interpretation of complex time series. A
record of stock price fluctuations is a signal, as is a rec:ord of temperature readings
that. permit the analysis of climatic variations and the study of global warming.
This list is by no means exhaustive.
Does there exist a precise definition of a signal that is appropriau. for the
field of scientific acti\'ity called '"signal processing"? A needlessly broad definition could include the sequence of lett.ers. spaces. and punctuation marks
appearing in Montaignes Essays. but the toOls v.e present do not apply to such
a sitmal. However. thC' structuralist analysis done by Roland Barth~ on lit.crary
t,c_,"f.S shares some amusin~ si.milarities with the multiresolution analysi." that WC'
describe in Chapter 4.
The signals we study will always be series of numbers and not !'Cries of letters.
words. or phrases. Thesc- numbers come from measurements. whicl1 arC' typical!~
1111\dC' using some recordin~ method. The signal." ultimately appear a.c: functions
of time. This is true for on~dimensional signals. The case of t.v.~imC'nsional
si~nlils will be examined in a moment.
Th e objecti11e.~ of .~ignal proce.uing o.rr to analy:r o.CCtJro.tely. rodr rffincutly.
trn nm it rapidly. o.nd tltrfl tc reconstruct rorrfull.v o.t tllr rccrit1rr tlw ddic:o.tr
CHAPTEH 1
These matrices are enormous. and as soon as one deals with a sequence of
images, tht> vol ume of numerical data that must be processed becomes immense .
Is it possible to reduce this volumt> by considering the "hidden laws" or correlations that exist among the differe.nt pieces of numerical information representing
the image? This question leads us naturally to define the goals of the scientific
discipline called "signal processing...
1.2. The goals of signal and image processing.
Experts in signal processing are called on t o describe. for a given class of signals.
aJ&orithms that lead to the construction of microprocessors and that allow certain
q~erations and tasks to be done automatically. T heS< tasks may be: analysis
ad diD.grwstics. coding. quantization and rompression. transmiSsion or storage.
and S1Jnthesis and reconstruction.
We will use several examples to illustrate the nature of t hese operations and
the difficulties they present. It will become clear that no unhersal algorithm'"
is appropriatt> for the extreme diversity of the situ ations encountered. Thus. a
large pan of this work is devoted to constructing .coding or analysis algorithms
that can be adapted to the signals that one processes.
Our first _example is the study of climatic Yariations and global warming.
'Thi5 example was discussed by Professor Jacques-Louis Lion!' at t he Spanish
Insti tute in 1990. and the following thought ~ wer e inspired by his talk!' [4).
Iu this example. one has fairly precisc t emperat u rc measurements from differ
eril poi nt.s in thr northem hemisphere that were taken <lver thr last two cemurit'l'.
and onc trie...; to disc-over if industrial acthity ha.~ caused global warming. The
extremt difficulty of the problem ari~c:; from tht exist euct' of sip1ificant natural
temperature fluctuations. Morco,er . thf.'S< fluctuation~ and thc rorrespoudiu~
cuinat.ic change- h11vc always cxist-<'d. a.~ wr learn from palroclimatology 16).
"Tt> specify 11 diagnostic, it is CS-"<'ntia.l t-< analy=.c. an d then \.u erase. thcst
natural flu ctuation~ lwhicb play thc rolt of u oi...;t) iu ordcr h ' han- aCCCS): t.u till'
..artificial" heating of the planet resulting from human activity. The diagnostic often depends on extracting a small number of significant parameters from
a signal whose complexity and size are overwhelming. Thus the analysis and
the diagnostic rely naturally on data compres~ion. If this compression is done
inappropriately, it can falsify the diagnostic.
Data compres8lon also occurs in the problem of transmis$aon. Indeed . transmission channels have a limited capacity. and it is therefore important to reduc<:
as much as possible the abundance of raw information so that it fits within tht
channel's "bit allocation."
One thinks. for example, of the digital telephone (Chapter 3 ) and the 64
Kbit/sec standard. which limits, without appeal, the quantity of information
that can be transmitted in one second.
A more surprising example appears in neurophysiology. The optic nerve's capacity to transmit visual information is clearly less than the volume of information collected by all the retinal cells. Thus, there must be "low-level processing"
of information before it transits the optic nerve. David Marr bas developed a theory that allows us to understand the purpose and performance of this low-level
processing. We present this theory in Chapter 8.
We now consider problems posed by coding and quantization. Different coding algorithms will be presented and studied in this work: subband coding,
transform coding, and coding by zero-crossings. In each case, coding involves
methods to transform the recorded numerical signal into another representation
that is, depending on the nature of the signals studied, more convenient for
some task or further processing. Quantization is associated with coding. The
'exact" numerical values given by coding are replaced with nearby values that
are compatible with the bit allocation dictated by the transmission capacity.
Quantization is an unavoidable step in signal and image processing. Unfortunately, it introduces systematic errors, known as "quantization noise.'" The
coding algorithms that are used (taking into account the nature of t he signals)
ou~;ht to reduce the effects of quantization noise when decoding takes place. One
of the advantages of quadrature mirror filters is that they "trap" this quantization noise inside well-defined frequenc~ channels. These filters will be studied in
Chapter 3.
The problems encountered in archiving data (as well as problems of transmi.."sion and reconstruction) are illustrated by the FBI"s task of storing the American
population "s fingerprints. Different image-compression .algorithms wer<' tested.
and a 'o.riant of th<' algorithm described in Chapter 6 gave the best results. This
<'Stablished a standard for fingerprint compression and reconstruction.
T he Ja..;t p-oup of operations consists of der.odi1tg, syntlte.<~is. and rc..<~toration.
S~ntht>Si." and d('('()diug ar<' th<' inverse operation." of coding and quruttization.
Tht' task il' t.o r('('()ust.ruct ru1 imag<' or audibl<' signal at th<' re<"<'h<'r from thC'
S<'ri<'l< of os rutd ] ":; that haY(' traveled over the transmission channel. Onl' thinks
of de<-odiuJ: zu1 <'nrodt-d mC'S.<>~.L~c. a.c: in crypt.ownphy, and thif' ruuuogy is correct
bt'C'nuse our C'IUlllOI rt'<'onstruct an inta(!t' or signal without kuowiu~ tht C't>ding
ctl~;orithm.
CHAPTE!t 1
1.8.
These "atorru;'' or "wavelets" have no more physical existence than the number system used to multiply the mass of the earth by that of the moon. Each
number system ha.s an internal logical coherence, but no scientific law asserts
that multiplication must of necessity be done in base 10 rather than base 2. On
tht' other hand, the number system used by the Romans is certainly excluded
because it is not suitable for multiplication.
Having different algorithms that allow us to code a signal by decomposing it
into "time-frequency atoms" presents us with a similar situation. The decision to
use one or the other of these algorithms will be made by considering their performance. How well they perform must be judged in terms of one of the anticipated
goaL~ of signal processing. An algorithm that i.~ optimal for compression am be
'disastrous for analysis: A stancuml energetic criterion for the compression could
cause detaiLs that are important for the analysis to be systematically neglected.
These thoughts v.'ill be developed and clarified in 1.6 and 1.7. It is time to
defint' wavelets, which we do in the next two sections.
1.4.
(1.1)
(l
- QIC:>
Tht> "mother v.avelt>t." t"(f). generateS tht' other v.avelets. t \ ...ltl (t ). o > 0.
b E R. of th<' fan1ily by change fJf scalt (th<' scale of ~/l(t ) is convcutiouully 1.
and tbnt of 1i1... b1( t ) i:< a > O) and translat.ion iu time (tb<' fuurt.ion t'(t} i.e;
romcut.ionl\lly c~nt<'rt'd around 0. ru1d t1... bl(tl is th<'n r.cnt.crt>d nruuud l).
CHAPTER 1
Thus we have
(1.2)
VJ{o,b)(t):;::
~VJC:b)
a> 0,
bE R.
Alex Grossmann and Jean Morlet have shown that, if V! (t ) is real-valued, this
collection can be used as if it were an orthonormal basis. This means that any
signal of finite energy can be represented as a. linear combination of wavelets
VJ{o,6)(t) and that the coefficients of this combination are, up to a. normalizing
factor, the scalar products f~oo /(t)1Pia,6)(t) dt.
These scalar products measure, in a. certain sense, the fluctuations of the
signal /(t) around the point b, at the scale given by a> 0.
It required uncommon scientific intuition to assert, as Grossmann and Morlet
did, that this new method of tim~scale analysis was suitable for the analysis and
synthesis of transient signals. Signal processing experts were annoyed by the
intrusion of these two poachers on their preserve and made fun of their claims.
This polemic died out after only a few years. In fact , the argument should
never have arisen because the methods of time-scale or multiresolution analysis
had existed for five or six years under various d1Sgwses: in signal analysis (under
the name of quadrature mirror filters) and in image analysis (under the name of
pyramid algorithms).
The first to report on this was Stephane Mallat. He constructed a. guide
that allowed the same signal analysis method to be recognized under very differeut presentations: wavelets, pyramid algorithms, quadrature mirror filters,
Littlewood-Paley analysis, David Marr's zero crossings...
Mallat's brilliant synthesis has been the source of many new developments.
One of the most notable of these is Ingrid Daubechies's discovery of orthonormal wavelet bases ha''ing preselected regularity and compact support. The
only previously knov.'ll case was the Haar system (1909). which is not regular. Thus 80 years separated Alfred Haar's work and its natural extension by
Daubechies !1].
Difficulties appeared when it was necessary to decompose a signal using Gabor wavelets. As long as one does only continuous decompositions (using all
frequencies and all time), Gabor wavelets can be used as if they formed an or
thonormal basis. But the corresponding discrete algorithms do not exist, or they
require so much tinkering that they become too complicated.
It is only very recently, by abandoning Gabor's approach, that two sep<V
rate groups have discovered time-frequency wavelets having good algorithmic
qualities. These special time-frequency wavelets, called Malvar wavelets, are
particularly well suited for coding speech and music. Moreover, they are equally
useful to the FBI for storing fingerprints.
The decomposition of a 8ignal in an orthononool bam of Malvar wavelets
tmitates writing mwic wing a mwical score. But this composition is misleading
because a piece of music can be un-itten in only one way, wheretls there exists
a nondenumeroble infinity of orthononool bases of Malvar wavelets. Choosing
one among these is equivalent to segmenting the given signal and then doing
a t raditional Fourier analysis on thl' delimited pieces. What is the best way
to choose this segmentation? This question leads us naturally to the following
section .
1.6.
CHAPTER I
1.7.
David Marr was fascinated and intrigued by the complex relations that exist
between the choice of a representation of a signal and the nature of the operations
or transformations that such a representation permits. He wrote !5, pp. 2o-21):
A representation is a formal system for making explicit certain entities or
types of information, together v.;th a specification of bow the s~stem does this.
And I shall call the result of using a representation to describe a given entity a
description of the entity in that representation.
For example. the Arabic. Roman and binary numerical ~stems are all formal
systems for representing numbers. The Arabic representation consists of a string
of symbols drawn from the set {0. 1. 2. 3. 4, 5. 6. 7. 8. 9} and the rule for
c.onstructing thl description of a particular integer 11 is that one decompose.<:
n into a sum of multiplC$ of powers of 10... A musical score pro,ides a "ay
of representing 11 l'ymphon~: the- alphabet allows tlw construction of 11 ''Tifl <'l
representation of words . ..
A repr~ntation. therefon. i:< not a foreign idea at all-we all us<' repr<"'<'J
tation~ all th<' time. HowPYcr. th<> notion that one can capturt' som<> a.c:;pect o:
reality b_,. making 11 descriptiou of it using a ~,nbol and that to do su em he
useful S<'<'DI~ w Ill<' a fascinating and powerful idea. But e\cn the simpl<' cxmuples WC' lnw<' di~cus.c:;ed int roduc<' soml' rather general ru1d important issut':< Lhat
arise' whcuenr vnc chooses w ust one particular representation.
For cxan1plt-. if om cho<>SE'l' t ht Arabic numerical representation. it. is<'<\..<:~ w
discover wh<'thC'r 11 number is a po"cr of 10 but difficult t.o discover wht'thC'r it is
11 powl'r of 2. If on<' chooses tht biunry representation . tht' situntion is rcvcrst'll.
10
CHAPTER l
Thus, there is a. trade-off; any particular repre8entation makes certain information uplicU at the upense of infonnation that is pU8hed into the background and
m4V be quite hard to recover.I
This issue is important, because how information is presented can greatly
affect how easy it is to do different things with it. This is evident even from
our numbers example: it is easy to add, to subtract and even t o multiply if the
Arabic or binary representations are used, but it is not at all easy t o do these
things-especially multiplication-with Roman numerals. This is a key rea.wn
why t he Roman culture failed to develop mathematics in the way the earlier
Arabic cult ures had . ..
There is an essential difference between Marr's considerations and the alger
rithms that we develop in the first six chapters. The difference is that the choice
of the best representation. according to Marr, is tied to an objective goal. For
tbt' problem posed by vision. the goal is to extract the contours, recognize the
edges of objects. delimit them. and understand their three-dimensional organization. In the algorithms we develop. the only criterion is internal t o the algorithm
and cons~ in reducing the amount of data. We have not yet studied situations
where one must also take into consideration an external criterion. In spite of
this difference, Marr's point of viev. is very close to the one we will present.
1.8. Terminology.
The elementary coostituents used for signal analysis and synthesis will be called.
depending on the circumstances. wavelets. time-frequency atoms. or v.avelet
packets (Chapter 8).
The wavelets used will be either Grossmann-Morlet wavelets of the form
.1_1;. (~).
..;a
a
a> 0 ,
j , k E '1...
k E !"\.
IE Z.
In thE' first two cases. wr will speak of time-scale algorithms: in the last ra.c;r.
we will speak of time-frequettr~- al~orithms. Later we will mix t hC' two point::: of
\iE!" and subject the Gabor-1\lah-ar wavelets to dyadic dilation.c; to construc-1 thl'
Daubechies v.avclets. On<' thu~ cnroumers g<'nC'ralized tim~fr<>qucncy at.oms.
\\'r v.ill lll*' only two "'very larg<' libraries.r T hr first consists of orthonormal
b8Se!: whO$(' element~> arC' wavelt't packets. In the second. t h<> v.-avelet. packeu: ar<>
replaced by thC' generalized tim~fr<'<)uency atoms. that W(' have just. d escribed.
1Thr
--
1.9.
11
Reader's guide.
Bibliography
Ttr. Lectures on Watoc-kL. Societ.y for lndu.~mal and Applied Mathem at
ic... P hiladelphia. PA. 1992.
D. G AllOI\. Tl1cMy of e<>mmumcatson . l. IE. 93 ( !946). PI' 429-457.
C . G ASQIC1 AtW F' . \\"tTOMS I\1 .Ar~aiy dr F01Jncr cl a ppliC<>tum.<. FiltrofiC. CalctJ
numenquc Or~ddcUes. Masson . Pe.rt~. ) !l9(l.
J. l . L tONS. La plane!<' Un"t: r Ol. d r$ mathhnottque.< ct dCJi IU,..-r ordanoteur.t. C'OUn< a
l'lnst itul dEspap>c. 1990.
D . MARl\ . \ "a..ntm. W . H. Fn-emM and (\.. 1\t'w York. HIH2.
I( I.
12;
:3j
14]
IS)
l6l
DAt B ECHIE~.
R . V AI''l"AIIl' AtW M. GHIL. S inj1Ular ~>lnnn anolpa.. 111 u o nltnror ciynamaa, with OJ'
plaC<>h<>u . t pcirodamahr tam<' rno . T'h~"'icn 0. 3~ (19~!1). PI' 35P-12..J.
CHAPTER
2.1.
Introduction.
The application of wavelets to signal and image processing is only a few years
old. But in looking back over the history of mathematics, we v.ill uncover at least
seven different origins of wavelet analysis. Most of this work was done around
the 1930s, and, at that time, the separate efforts did not appear to be parts of a
coherent theory. In particular, neither the word "wavelet" nor the corresponding
concept appeared in this research done a half-century ago. Only today do we
know that this work prefigured the theory of wavelets.
It is important to describe these seven sources in some detail. Each of them
corresponds to a specific point of view and a particular technique, v.hich, only
now, are we able to viev. from a common scientific perspective. What's more,
these specific techniques were rediscovered several yearsago by physicists and
mathematicians working on wavelets. For example, the Littlewood-Paley analysis (2.4), which dates from 1930, underlies Mallat's work on image processing.
Matthias Holschneider used, without knowing it, Lusin's technique (1930) to
clarify the fractal structure of Riemann's function (2.6 and 9.4). G~mann
and Morlet rediscovered Alberto Calder6n's identity (1960) 20 years later. And,
to spare no one, the author of these lines was not the first to construct a regular, well-localized orthonormal wavelet basis having the algorithmic structure
of Ha.ar's system (1909): J . 0. Stromberg bad done the same thing five years
earlier.
Does this mean that evel")-thing bad alread) been written and that '"t.earuv.-avelet" researchers appropriated-while making a great show of it-results discovered by others? This judgment must be qualified for two reasons. In the first
place. tht> physicists from l.eam wavelet"" were not aware of Calderon's work: it
was in completely good faith that tht-y presented as a revolutionru: innovation
rt>Sults that wcr<' about 20 years old. "Tt-am wavelet" researchers can b<' t~ecu...OO
of i~oranCE' but not of plagiarism. But above all, by rediscovering these lmown
scientific facts, thC' investigat.or.: gaw them new life and authority. Our debt t.o
Grossmann and 1\lorlC't is not so mucl1 for having rediscovered the idcntit-) of
Calderon as it is for ha,;n~; related it to processing nonstationru: si~als. Thil>
bold 10ynthesis cC>rtninly C>nrountcred r<.>sL"fwlce. and Calderon himself found t.hil'
W:<' of his work illron~mtul'.
14
CHAPTER 2
The role of Morlet and Grossmann in the wavelet saga can be compared to
that of Mandelbrot in the story of fractals. It is true that before Mandelbrot
there were Pierre Fatou and Gaston Julia, and certain of Mandelbrot 's discoveries
repeat their work. But Mandelbrot showed us the possibility of interpreting
the world around us with the help of a new concept, that of fractals. None
of the mathematicians working on Hausdorff dimension were ready, from their
training or experience, for such a leap into the unknown. Wavelets have gone
the same way, and one of our objectives is to construct the bridge that relates
signal processing to the different mathematical efforts that developed outside the
"theory of wavelets.
2.2.
Let's go back to the origins, that is to Joseph Fourier. As everyone knows, he asserted in 1807 that any 2r.-periodic function f(x) is the sum ao+ L~(ak coskx+
bt sinkx) of its "Fourier series." The coefficients ao. ak, and bk(k ~ 1) are calculated by
ao = 2171"
and by
112.
ak=7f
f(x ) coskxdx,
112.
bk =1r
f (x)sinkx dx.
When Fourier announced his surprising results, neither the notion of function
nor that of integral had yet received a precise definition. We can even say
that Fourier's statement. played an essential role in the evolution of the ideas
mathematicians have had about these concepts.
Before Fourier's work, entire series were used to represent and manipulate
functions, and the most general functions that could be constructed were endowed with very special properties indeed. Furthermore, these properties were
unconsciously associated v.oith the notion of function itself. By passing from a
representation of the form
(2.1)
to one of the form
(2.2)
ao -t- (a1 cosx + b1sinx) + (a 2 cos 2;r...,.. b:! sin 2;r) + ...
15
(a) They could modify the notion of function and find one that is adapted,
in a certain sense, to Fourier aeries;
(b) They could modify the definition of convergence of Fourier series; or
(c) They could find other orthogonal systems for which the phenomenon,
discovered by Du Bois-Reymond in the case of the trigonometric system, cannot
happen.
The functional concept best suited to Fourier series was created by Henri
Lebesgue. It involves the space L 2 (0, 21r] of (classes of) functions that are Squareintegrable on the interval [0, 21r]. The sequence
$. '
,jT. SID X
is an orthonormal basis for this space. Furthermore. the coefficients of the decomposition in this orthonormal basis form a square-summable series, and this
expresses the conservation of energy: The quadratic mean value of the developed function f (r.) is (up to a normalization factor) the sum of the squares of
the coefficients. Finally, the Fourier series of f converges to f in the sense of the
quadratic mean.
The second way that was followed to avoid the difficulty raised by Du BoisReymond was to modify the notion of convergence. The partial sums Sn(r.) are
replaced by the Cesaro sums C!n = ~(So++ Sn-d, and everything falls into
place.
The third route leads to wavelets. This was followed by Haar, who
asked bimseU this question: "Does there exist another orthonormal system
ho(r.). h 1(r. ), .... h.,(r.), ... offunctions defined on (0.1] such that for any tunc
tion f (r.). continuous on [0.1j. the series
(2.3)
...
16
CHAPTER 2
We can criticize the Haar construction on a couple of points. On one hand, the
"atoms" hn ( x) used to construct the continuous function f (x) are not themselves
continuous functions, and thus there is a lack of coherence.
But there is a more serious criticism. Suppose that, instead of being continuous on the interval [0,1], f(x) is a function of the class C 1 , which means f (x )
is continuous and has a continuous derivative. Then the approximation of f(x)
by step functions would be completely inappropriate. In this case. a suitable
approximation would be the one created from the graph of f(x ) by inscribing
polygonal lines.
The Ha.ar construction is suitable only for continuous functions. functions
square-integrable on [0,1j or, more generally, functions whose index of regularity
is near 0. We will see a little later what this means.
These two defects of the Haar system and the idea of approximating the
graph of /(x) with inscribed polygonal lines Jed Faber and Schauder to replace
the functions h..(x) of the Ha.ar system by their primitives. This research began
in 1910 and continued until 1920.
Define the "triangle function" A(x) by A(x) = 0 if x ~ [0.1], A(x ) = 2x if
0::; x::; 1/ 2, and A(x i = 2(1- x) if 1/2 ::; x ::; 1. Then Faber and Schauder
considered' the series An (x), n ~ 1, defined by
~f
(2.4) An (x)
= A{2jx- k)
for
n = 2j + k,
j ~ 0,
The support of An{x) is the dyadic interval In [k2- j, (k + 1)2- i]. and An (x )
is the primitive of hn(x) multiplied by 2 2i12 and zero outside In
For n = 0, we set Ao{x)
x, and we add the constant 1 to complete the
set of functions. Then the sequence 1, Ao(x) .... . An (x), ... is a Sclw.uder basis
for the Banach space E of continuous functions on [0.1]. This means that every
continuous function on [0.1] can be written as
ex
(2.5)
L o,. ~n(:r \
l
and that the series has the follov.ing properties: the convergence is uniform on
[0.1] and. as a consequence. the coefficients are unique.
We note that the Ha.ar system is not a Schauder basis of E because a Schauder
basis of a Banach space must be made up of vectors of that space. and the
functions h, are not continuous.
The calculation of thc coefficients in {2.5) is immediate. We have successively
flal =a and !{1) =a- b. whicl1 ghes a and b. This allows us to consider a
function ftx)- a - b:r. v.hicl1 is zero at 0 and 1. Oncc this reduction is made. wc
have .f(1 / 2) = o 1 which allow~ us to consider a function equal to zero at 0. 1.'2.
and 1. The calculation continues with f (1/4) = o ~ and ! {3/ 4) = o 3 and ~>Cl on.
If ~c do not wish to ~r / (:r) thi~ v.ay. tht> coefficients On can be comput.ed
directly by the fom1ula
~ ;6 ,
(- l ;
17
18
2.3.
CHAPTER
Brownian motion is a random signal. We will limit our d.iscuasion to the one-
dimensional caae. We thus write X(t,w) for the Brownian motion: t denotes
time, w belongs to a probability space n, and X (t, w) is regarded as a real-valued
function of time depending on the parameter w.
To obtain a realization of Brownian motion, we choose a particular orthonor, mal basis Z1(t), i E ! , for the usual Hilbert space L 2 (R). Then we know that
the derivative (in the sense of distributions) ftX(t,w) is written as
(2.7)
d
dtX(t,w) = L9;(..:)Z,(t),
EI
-7,;.1;
(2.8)
1;
DC
2 2:)-i12gn(w)l!.n(t).
I
..-here the 9n(;.;) are independent Gaussian random variables v.;th mean zero and
the same distribution.
lb verify that thl' function X (t . ..:) bclonp: to tht> Holdt>r Spacl'
for almost
all..: E 0 , it is sufficient to show that 2-J: 219n(...)l ~ C(:)2-Jr. If. for almost
all w n, onf.' had SUPn>O l9n(w)l < oc. then t.ht> trajectorie!' of th<' Brov."Jlian
motion would almOSt sur;ly belong to the spact> (' 112 But this is not tht> case.
and instead Wt' haVl' SUPn>209n(w)li ,llog n) < oc for almost all .... E n. Then
tbt> criterion for Holder regularity give:'
cr
(2.9)
V.'hert> C'(w)
all ..; E
19
We see, from this theorem of Paul Levy, the superiority of the Schauder basis
over the Fourier basis for studying local regularity properties.
The determination of the fractal exponents requires some extensions. These
extensions that allow the study of small, complicated details form part of "multiresolution analysis," which we define in the following chapters. These ideas are
already incorporated in t he definition of the Schauder basis itself, in particular,
within the mapping x .... 2' x - k. They are clearly absent in the trigonometric
system.
Patrick Flandrin 15] has extended this work to the case of fractional Brov.'Dian motion. as it was proposed by Benoit Mandelbrot and John W. van l'<ess
to model certain noise. Albert Benassi, Stephane Jaffard, and Daniel Rowe
13) have generalized these ideas to certain Gaussian- Markov fields. All of this
work demonstrates that multiresolution methods are adapted to the analysis and
synthesis of these processes.
2.4.
We have shown, with the example of Brownian motion, that the trigonometric
system does not provide direct and easy access to local regularity properties and
that these properties are clearly apparent when examined with other representations.
Similar difficulties are encountered when we try to localize the energy of a
function. To be more precise. the integral .].,. Jt' lf(x}fldx, which is the mean
value of the energy, is given directly by the sum of the squares of the Fourier
coefficients. However, it is often important to know if the energy is concentrated
around a few points or if it is distributed over the whole interval I0,2w]. This
determination can be made by calculating 21.,. J;" lf (x)! 4 dx, or more generally,
f02" Jf(x)l"dx for 2 < p < oo. When the energy is concentrated around a few
points. this integral will be much larger than the mean value of the energy. while
it will be the same order of magnitude when the energy is evenly distribute(!. We
write 11!11, = (i,. J;" lf(x)l"dx) 11P and, for obvious reasons of homogeneity. we
compare the norms llfll, to determine if the energy is concentrated or dispersed.
But if p is different from 2, we can neither calculate nor even estimate these
norms llf li, by examining the Fourier coefficients of f .
The information needed for this calculation is hidden in the Fourier series of
f: to reveal it. it is necessary t.o subject the series to manipulations that vere
discovered by Littlewood and Pal~ as long ago as 1930.
Lit.tl<'wood and Paley definE' the Mdyadic blocks !:l.1 f (x ) b~
i.,
(2. 10)
~if{:r)=
(akcosk:r+h.sinkx),
2J$ k<2J +l
where a 1,
+ I:~ lak coskx- b.- sink:r) denotes thr Fourier series of f. Then
oc
ft:r)
CHAPTER
20
and the fundamental result of Littlewood and Paley is that there exists, for
1 < p < oo, two constants C, ~ c, > 0 such that
11
(2.11)
c,ll/11, !S ( laol2 +
~ l~if(xW) ~
!S C, ll/11,.
'P
= =
If p
tb(~)
(2.12)
=1
if 1 +a !S 1~1 S 2- 2a.
and
(2.15)
,Z,(~)
=0
if I~ I S 1 -a or 1~1 ~ 2 + 2a,
L I~(Ti{):
=1
for all ~? 0.
-oc
WA\'.1-:LI:cTS
F.H.0~1
21
If j (:r:) bclong.c, t CJ U (R '), the sanae is trU(~ fur g(x), and ll/1!2
IIBll2 (the
c0nscrvation of energy).
If 1 < p < oc., there exist two constants C, ;::: c.p > 0 such that, for all
functions f belonging to V(IR"),
(2.16)
where
'
) 1/ p
This double inequality (2.16) doe!> not hold in the limiting case where
() = 0 [4J.
The Littlewood-Paley- Stein function g(x) provides a method for analyzing
f !x ) in which a major role is played by the ability to vary arbitrarily the scales
used in the analysis; by the same token, the notion of frequency plays a minor
role. The dilations of size 2i are present in the definition of the operators tl.;.
!'\'}vertheless. conditions (2.12} and (2.13) endow these operators with a frequential content. The sequence of operators~,, j E l. constitutes a bank of band-pass
filters. oriented on frequency intervals covering approximately an octave.
Thanks to the work of Marr and Mallat (which we describe in Chapter 8) the
Littlewood- Paley analysis provides an effective algorithm for numerical image
processing.
2.5.
1
I
(2.17)
fn l:r)d.x
t :rf,(:r)dx = 0
= Jo
for
n;:::: 1.
The Franklin system has advantages of both the Haar basis and the Schauder
basis. It. can be used to decompose any function fin L 2 (0,l ]. which the Schauder
basis does not allow. and it can be used to characterize the spaces C'. 0 <
r < 1. by li.f.f,. ' :S L TI- l / 2 - r. which the Haar basis does not allov.. Thus
tht Franklin sy:;tcm works as wdl in relative!) regular situations as it does in
rciatively irregulu.r sit.uations.
Thr wcaknr:;.~ of thr Franklin basi$ is that it no longer bas u simple> algorithmic s truct.un. Tht' functions of th( Franklin basis. unlike thOS<' of tht Haar basi.~
or t.hOS<' of tht' Sdmudcr basis. arr not derived from a fixed function t by integer
translation!' a.nd dy~k dilations. Thb defect caused the Fraukliu syst.cna tu br
abandomd ruad fo~otten for almost 40 years.
CHAPTER 2
22
Fortunately, the Franklin system has survived this disgrace. Zbigniew Ciesielski revived it in 1963 by showing the existence of an exponent -y > 0 and a
constant C > 0 such that
(2.18}
{2.19)
Thus, everything works as if /n (:r) = 2ii2 1/J(2J:r - k ) and 1/J(:r) were a Lipschitz
ftmction with exponential decay.
Today we have an asymptotic estimate for the functions /n(:r}. This estimate shows that, in a certain sense, the orthonormal wavelet basis discovered
by Strfunberg in 1980 was living hidden inside the Franklin system . We have, in
fact, for n = 2i + k, 0 $ k < 21,
(2.20)
where, for a certain constant C,
d(n) = inf(k,2;- k}.
(2.21}
(2.22)
l]
1+1
ll/2, 1), IO, 1/2]. l-1/2. 0].. .' [--2-.-2 .... ,
(2.23}
v'3) <
e Z.
and
t"
L2 (R).
decreases rapidl.' at
23
defined by z 2: +ill and 11 > 0. Then a function /(z +ill) belongs to H"(R) if
it is holomorpbic in the half-plane P and if
00
{2.25)
sup
11>0
(1
lf(z + iv)l"dx
1/ p
< oo.
- oo
When this condition is satisfied, the upper bound, taken over 11 > 0, is alao
the limjt as 11 tends to 0. Furthermore, I (2: + iy) converges to a function denoted
by l(z) when ll tends to 0, where convergence is in the &eDBe of the V -norm.
The space H~'(R) can thus be identified with a closed subepa.ce of V{R), which
explains the notation.
The Hardy spaces play a fundamental role in signal processing. One associates with a real signal l {t), -oo < t < oc., of finite energy, the analytic signal
F (t ) for which l(t), -oo < t < oo, is the real part. By hypothesis, the energy
of I is f~oo ll (t)lldt, and we require that F(t) have finite energy as well. This
implies that F belongs to the Hardy space H 2 {R). Then F(t)
f(t) + ig(t),
and the function g(t) is the Hilbert transform of /(t). For further infon:n&tion
about analytic signals, the reader may refer to [9, pp. 118-119]. One may alao
consult the rem&rkable exposition by Jean Ville, (10].
Read in the light of the theory of wavelets, Lusin 's work concerns the analysis
and synthesis of functions in H~'(R) using "atoms" or "basis elements," which are
the elementary functions of H"(R). These are, in fact, the functions (z- ()- 2 ,
where the parameter (belongs toP.
Thus one wishes to obtain effective and robust representations of the functions
f (z) in H~'(R) of the form
f (z) =
{2.26)
Ji
(z- ( )- 2 o (()dudv,
where ( = u + it: and where o(() plays the role of the coefficients. These
coefficients should be simple to calculate, and their order of magnitude should
provide an estimate of the norm of I in H~'(R).
The synthesis is obtained by the folloll.ing rule. We start with an arbitral'l'
measurable function o((), subject only to the following condition introduced by
Lusin: The quadratic functional A(:r) must be such that f~00 (A(:r))"d:r is finite.
This quadratic functional is defined by
(2.2i)
.4(.:rl =
where
f(.:r)
!r r
( lr(zl
ln
lo ltJ + it)l!?t.-!?dudv
= {(u.tr) e R2 .
t'
Notl' that this condition involves only tht' modulus of the coefficients o((}.
lf tbl' integral J~~(A(.:r))"d.:r is finite. then necessarily
f l:) ==
j{,,(;-()- o (()du dt
2
belongs to W '(R).
CHAPTER 2
24
11/ llp $
(2.28)
C(p)
(j
oe;.
- oc. (A (x ))Pdx
)Iff>
The left member of (2.28) is the norm off in HP(JR), as defined by (2.25). This
estimate, however, is sometimes very crude. If, for example, / (z) = (z + i )- 2 ,
one is Jed to choose the Dirac measure at the point i for o ( ( ) , and the second
member of (2.28) is infinite. This paradox arises because the representation
{2.29)
is not unique.
To obtain a unique decomposition. which we call the natural decomposition.
we restrict the choice too{( ) = ~ vf' (u + iv). When we do this. the two norms
11/ llp and IIAIIJ> become equivalent if 1 $ p < oo. Today this choice of natural
coefficients has an interesting explanation. This interpretation. which depends on
the contemporary formalism of wavelet theory, is given in the following section.
2.1.
Guido Weiss, in collaboration with Ronald R. Coifman, was the first to interpret.
as we have just done, Lusin's theory in terms of "atoms" and ..atomic decompositions." The "atoms" are the "simplest elements" of a function space. and the
objective of the theory is to find; for the usual function spaces. {1) the atoms
and {2) the ..assembly rules" that allov. the reconstruction of all the elements of
the function space using these atoms.
In the case of the holomorphic Hardy spaces of the last section, the atoms
were the functions (.::- ( )- 2 ( P . and the assembl~ rules werE" given b~ the
condition on Lusin's area function A\x).
For the spaces V[O, 21r], 1 < p < oc. the ..atoms cannot be the functions
cos b.: and sin kx, k ;:: 1, because this choice does not lead to assembly rules that
are sufficiently simple and explicit. to be useful in practice. Marcinkiev:icz showed
in 1938 t~t the simplest atomic decomposition for the spaces LP(0.1], 1 < p <
oc. is given by the Haar system. The Franklin basis would have served as well
and. from the scientific perspective given by v;avelet theory. the Franklin basis
and Littlewood-Paley analysis are naturally relat.ed.
One of the approaches to 'atomic decompositions.. i!' ghen by Calderon 1'
identity. To e>..-plain Calderon's identit~. we start with a function t:( x ) belonging:
to LZ(JR" ). Later in this history, Grossmann and r.1orlet callE-d this function an
analyzing wavelet. Its Fourier transform t7l~ ) is subject to thE" condition that
(~ .30)
i
I
oc
., dt
l1!"(t{ )l - = 1 for almost all ~ E IR" .
25
(2.31)
I)J(x )dx = 0.
}Rn
IP( - x ). w (x ) = t-nw(x f t ),
We write J,( x ) =
and J,,(x) = t - n.J;(x f t ). Let Q t
1
denote the operator defined as convolution with 1P1 ; its adjoint Q~ is the operator
defined as convolution with -1 .
Calderon's identity is a decomposition of the identity operator, written as
I =
(2.32)
oc
dt
Q,Q; -.
t
Grossmann and Morlet rediscovered this identity in 1980, 20 years after the
work of Calderon. However, with this rediscovery, they gave it a different interpretation by relating it to "the coherent states of quantum mechanics" [6].
T hey defined wavelets (generated from the analyzing wavelet '1/J) by
(2.33)
(X- b) ,
a> 0,
(2.34)
(2.35)
f oe {
Jo
}Rn
an+ 1
This is a linear combination of tht- original wavelets using the coefficients given
by the analysis.
We return to the specific case of thE' Hardy spaces H"{R) for 1 $ p < oc.
Tht- analyzing wavelet t (:.r) = ~ ( x + i)- 2 chosen by Lusin is the restriction to
thE' real a.xis of thE' funct ion ~(.: + i) - 2 : it is holomorphic in P and belongs t o
all of the Hardy spaces. The Fourier transform oft is J ({ ) = -2{e- { for~ ~ 0
and t(~l = 0 if~ $ 0.
Condition (2.30) is not satisfied but, on the other hand, we have
:>C
(:.?.36)
.. dt
l t lf ~) - =1
if { >0.
=0
if { $
o.
Condition (2.36) implies that thE' wavelets t"\ .,.1>1 generate H 2 (R) inst.cad of
L~ ( JR ) wht-n o > 0. h .;: R
26
CHAPTER 2
(2.37)
W (a,b) = (/,1/J(o.b)} = -1
1f
j"" f (x )(
- cc.
X -
aJa
b . )2 dx,
- Ul
ee.
dt
I,Z,(t{)l2 - = +X.
t
2.8.
if {
> 0.
Stromberg's wavelets.
The real version of the holomorphlc Hardy space H 1(R) is denoted by 'H 1 (R)
aud is composed of real-valued functions u(x ) such that u (x) + iv(x) belongs to
H1(R), where v(x) is also real-valued. In other words. u (x) belongs to 7t1 (R) if
and only if u and its Hilbert transform ii belong to L 1 (R).
Research on "atomic decompositions" for the functions in the Hardy space
7t1(R) takes two completely different approaches: one involves the atomic decomposition of Coifman and Weiss, and the other concerns the search for unconditional bases for the space 7t1 . Here is an outline of these theories.
Coifman and Weiss showed that the most general function f of 7t 1 can
be written as f (x ) = 2:~ >.kak(2 ), where the coefficients >.k are such t hat
IA.I:I < oo and where the ak(x ) are atoms of 'H 1 . This means that for
each ok{x), there exists an interval Ik such that ak (x ) = 0 outside of I .
lo(x)l :5 1/II~o l (I I I is the length of Ik ). and f 1 ak(x )dx = 0. These three
conditions imply that the norms of the o~o in 'H1 are bounded by a fixed constant
Co. The price to pay for this extraordinary decomposit ion is that it is not. given
by a linear algorithm.
Finding an unoonditional basis means constructing. once and for all. a sequence of functions b~o (:r) of 'H1 that are linearly independent. in a very strong
.sense. and that allov. all functions f of 7t 1 to be decomposed as
Eo""
cc
(::!.38)
J<x>=
L: v. bktxl.
0
(2.39\
Hen' thc g11 are specific functions in thc dual of 'H 1 . which is to say th<'y
spt'Citic BMO function:; .
RT<'
27
Thanks to the historical perspective that we have today, we can relate the
Littlewood- Paley decomposition (1930). the version of Franklin's basis given
by Stromberg (1927), and Calderon's identity (1960).
This first synthesis will be followed by a more inclusi':'e synthesis that encompasses the techniques of numerical signal and image processing. This second
synthesis will lead to Daubechies's orthonormal bases.
This first synthesis is based on the definition of the word "wavelet" and
on the concept of "wavelet transform." We will see that the success of this
synthesis depends on a certain lack of specificity in the original definition. At
this time. mathematicians had not created a general formalism covering all of
the examples we presented above. A physicist and an engineer. Grossmann and
Morlet. provided a definition and a way of thinking based on physical intuition
that was flexible enough to cover all these cases. Starting v.;th the GrossmannMorlet definition. we will present two other definitions and indicate hov. thry
are related.
The first definition of a wavelet. which is due to Grossmarm and Morlet. i.
quite brood. A watelet i.~ a function t ' in L 2 (R) whose Fourier transform 1it~)
satisfie.~ tile coudition
l~t~W' = 1 almost everywhere.
J:
28
CHAPTE(I
f x R" .
ln the Franklin- Stromberg theory, a is replaced by 2-; and b is replaced
by ka.-i, where j , k,E Z. ln other words, the analysis off in the FranklinStromberg basis is obtained by restricting the Littlewood- Paley analysis to the
"hyperbolic lattice" Sin (O.oo) x R consisting of the points (2-J, k2 - 1), j, k E Z .
The logical relations among these wavelets analyses are easy to verify.
We start with the Grossma.nn- Morlet analvsis. that is. the Calderon identit\.
This is written I
Jo"" Q,Q; ~. where QJJ) = f b 1 . This becomes I ::,
L~o.: tl1 .6.j in the Littlewood-fl"a.le~ theory, and if t
2-;. one bas Q1( /)
.6.; (!). Replacing t by 2-i and the integral J0"" u(t)~ by the sum L~o.: u (2-')
is completely classic.
To relate Littlewood-Paley analysis to the analysis that is obtained using the
orthogonal v:avelets of Franklin and Stromberg. wt> write t::;(z ) = 2i tf:(2i:r). and
tb;(x) 2iw(2i :r), where ~lxl t (-x}. We let tl1 ( / ) denote the convolution
product f 1P; and 6j : L 2 (R)- L2 (R} the adjoint ofthe operator .6.1 : L2 (R) L2 (R). Then .6-j (/) = f ;;1 . The coefficients o lf. k) of the decomposition off
in Stromberg's orthonormal basis are then given by
(2.41 )
o (j.k\ =
zn.!
f l:rk(2J: - in d.:
29
f (x ) =
lo JR..
J;-
j (x ) =
t 1.
-oc
Finally, for the one-dimensional case and for Stromberg's orthogonal wavelets,
t he last integral becomes a sum, and (2.43) becomes
oc.
(2.44)
f (x ) =
"'"
All of the preceding arguments remain at a fairly superficial level since the
hypot hesis on 1:' enables us to analyze only the space L2 of square-sum.mable
functions. This is the setting in which Grossmann and Morlet wrote their theoretical work. But this is evidently a sort of regression, for we have just shov."D
that, across a century of mathematical history, wavelet analysis was created
specifically to analyze function spaces other than L 2 Fourier analysis serves
admirably for L 2
If we want wavelets to be useful for the analysis of these other function spaces.
it is necessary to impose conditions on them in addition to those we have already
given. Up until now v.e have required only that the analysis preserve energy or,
which is the same thing. that the synthesis give an exact reconstruction.
These new conditions are
(2.45)
(2.46)
(2.47)
W e can. for cxamplt>. impose the conditions that tt helonp: t.O the Schwartz
cia$ and that all of it!' moments vanish. \\'e can also requir<> t which " ill br thr
case' for Dauht'chit'l'.l' waYdets) that li'lil have m continuou.':' derivatives. t hat it
han .> compact support. and that its first r - 1 moments vanish.
T hr prop('rt.i<>:< of t hf Stromberg wswelC't art' imermediaw: It ha.-< cxpoucntiw
decay. a.o.; doe:< itl' first dC'ri,-ative. and
L:
t'(:r)d:r =
L:
:rt!(x)d:r
= o.
CHAPTER 2
30
2.10.
If history had stopped with this first synthesis, the Daubechies orthonormal
bases, which improve the rudimentary Haar basis, would never have been discovered.
A new start was made in 1985 by Stephane Mallat, a specialist in numerical
image processing. Mallat discovered some close relations among
(a) the quadrature mirror filters, which were invented by Croissier, Esteban,
and Galand for the digital telephone.
(b) the pyramid algorithms of Burt and Adelson, which are used in the context of numerical image processing, and
(c) the orthonormal wavelet bases discovered by Stromberg and his succesIIOrS.
These relations will be explained in the next two chapters. Using "Mallat's
program," Daubechies was able to complete Haar's work. For each integer r ,
Daubec:bies constructs an orthonormal basis for L 2 (R} of the form
(2.48)
j . k E Z,
(2.51)
1 is about 1/ 5. 'When r
= 0,
than that obtained with the Haar system. If the function being analyzed bas
m COIItinuous derivatives, where 0 S m S r + 1. then the coefficients a(j, k )
from iU decomposition in the Daubechies basis will be of the order of magnitude ~(m+l/2 )j. while it would be of the order 2- 3i/2 with the Haar system.
This means.tbat as soon as the analyzed function is regular, the coefficients one
keeps (those exceeding t he machine precision) will be much Jess numerous than
in the case of the Haar system. Thus one speaks of signal ~compression... Furthermore. this property has a purely local aspect because Daubechi1'.!' wavelets
havt> compact support.
Synthesis using Daubechies's wavelets also give!' better results than tht> Haar
~'Stem. In the latter case. a regular function is approximated by functions that
have strong discontinuities. This is very anno~-ing for image processing. as the
reader can veri~ by referring to the imagt> of Abraham Lincoln ou page i4 of
Marr's book [7). These remarkable qualities of Daubechies's base e>.."J>lain their
undisputed success.
2.11.
31
Conclusions.
(1) G.
(2)
(3]
!4)
(5)
!6:
!7:
!tt:
(9;
(10;
BATTLE.
No. 1. ( 19-.18).
CHAPTER
3.1.
Introduction.
It was with keen pleasure that I recently reread Claude Galand's thesis entitled
"Codage en sous-ba.ndes: theorie et applications a la compression numenque du
signal de parole." This thesis was defended on 25 March 1983 at the Signals and
Systems Laboratory of the University of Nice.
In his thesis, Galand carefully describes the quadrature mirror filters (which
he invented in collaboration with Esteban and Croissier) and their antic
ipated applications. He also posed some very important problems that would
lead to the discovery of "wavelet packets" (Chapter 7) and "Malvar's wavelets'"
(Chapter 6).
Galand 's work was motivated by the possibility of improving the digital tel~
phone, a technology that involves transmitting speech signals as se(zuences of O's
and 1's. However, as Galand remarked, these techniques extend far beyond the
digital telephone; facsimile, video, databases, etc., all travel over telephone lines.
At present, the bit allocation used for telephone transmission is the well
known 64 kilobits per second. Galand sought, by using coding methods tailored
to speech signals, to transmit speech well below this standard.
To 'llalidate the method he proposed, Galand compared it to two other tech
niques for coding sampled speech: predictive coding and transform coding.
Linear prediction coding amounts to looking for the correlations among sue
cessive values of the sampled signal. These correlations are likely to occur on
intervals of the order of 20 to 30 milliseconds. This leads one to cut the sampled
signal x (n ) into blocks defined by 1::; n::; N, N + 1::; n::; 2N, etc., and then
to seek. for each block, coefficients a~r . 1 ::; k ::; p. that minimizo:e the quadratic
mean -k I:~ !e(n )i2 of the prediction errors defined by
k= l
In general. p i.o; much smaller than/'.'. To transmit the block :r(n), 1::; n $ N . it
suffices to transmit th> first p values :z:(l) . . ,x(p). th> p coefficients a 1 .. a,..
and thc- prl'dict.ion erron; e(n ). Th> method i~ efficient if most of th> predk
tiou erron: art near 0. When thry fall below a certain threshold. th<>~ art' not
transmittt'<l. IUJd si~uifi<"Wlt comprC'SSiou can result.
3:4
34
CHAPTER3
Traruform coding consist.s in cutting the sampled signal into succe88ive blocks
of length N, as we have juat done, and then uaing a unitary transformation A to
traD8fonn each block (deooted by X) into another block (denoted by Y ). The
block Y is then quantized, with the hope that, for a suitable linear transformation, the Y blocks will have a simpler structure than the X blocks. Subband
coding will be presented in the next section.
For a stationary GaUBSian signal, the theoretical limits of the minim.al distortion that can be obtained by the three methods are the same. But, as Galand
showed, this assumes, in the case of subband coding, that the width of the frequency channels tends to 0 and that their number tends to oc. We will show that
these conditions cannot be satisfied (rigorously) for subband coding. To provide
an approximate solution, Galand proposed a treelike ~gement of quadrature
minor filters. This construction leads precisely to the wavelet packets, which w:
discuss in detail in Chapter 7.
In the cases of linear prediction coding and transform coding, the theoretical
limits of the minimal distortion are calculated as the lengths of the blocks tend
to infinity, while conserving the stationarity hypothesis.
If the three types of coding yield asymptotically the same quality of compression, why introduce subband coding? Galand saw two advantages: the simplicity
of the algorithm and the possibility that subband coding would reduce the unpleasant effects of quantization noise as perceived at the receiver. By quantizing
inside each subband, the signal would tend to mask the quantization noise, and
it would be less apparent.
The same argument bas been repeated by Adelson, H.ingorani, and Simoncelli
[1] for numerical image processing. The use of pyramid algorithms and wavelets
allows aspects of the human yisual system to be cleverly taken into account so
that the signal masks the noise. The perceptual quality of the reconstructed
image is improved even though the theoretical compression calculations do not
distinguish this method from the others.
3.2.
We follow Galand's example and begin with a deceptively simple case. For a
fixed m ~ 2, let I denote an interval of length 2r./ m \\"ithin (0,27r], and let l~
denote the Hilbert space of sequence (ck)kEZ verifying L:lckl2 < oo and such
that /(8) = L~oockeu.e is zero outside the interval I. This subspace will be
called a frequency channel.
If (ck)keZ is a sequence belonging to 1~. the subsequence (Cmk)kEZ prm;des
an optimal. compact representation of the original beca\l.SE'
( /(9} + f
(o + -2r.)
m
-r. -r,
((I . 2:7\1111))) = L
m
~
Ckm ComP
This means that the original sequence contains m times the numerical data
needed to reconstruct /(8) on I , knowing that /(8) is 0 out8ide of I .
The i<kal 1cheme for aubband coding consists in first jiltmng the incoming signal into m frequency channels associated with the intervals 12,.-1/m,
21r(l + 1)/ mJ, 0 S l $ m- 1, and then 1Ub1ampling the corresponding ou~
puts, retaining only one point in m . This operation, which consists in restricting
a sequence defined on Z to mZ, is called decimation and is denoted by m ! 1.
Thus the ideal subband coding scheme is illustrated as follows:
Yt(nm)-
Ym(nm)The scheme for recorutructing the original signal is the dual of the analysis
!jclleme. We began by extending the sequences y 1 (nm), ... , Jlm(nm) by inserting
O's at all' integers that are not multiples of m . Next we filter this "absurd
decision" by again using the filters F 11 ,Fm. The output returns (:z:(n)).
r--~.........,-(x(n ))-..--
- - Y m (nm)
One can. as Galand did. hope for the best, and tl) to replace the index
functions of the intervals 12,..1/m, 21r(l + 1)/mJ with more regular functions of
the form utm.r - 271"1). 0 S 1 S m - 1. If the w(:z:) = u:m(:r) were a finitr
trigonometric sun1. then the filters F1 Fm would have finit.t> length. which is
essential for applications.
But the Baliau- Lov. theorem (Chapter 6) tells us that such a function u:(x)
cannot be constructed if we demand that it be regular and well localized (uniformly in m ). Consequently, it Lc; not possible to realize the ideal subband codinl!
schemr just. described if we require that the filters F~o ... Fm have finite length
and. at thr samr timr . providr good frequency definition.
36
J.l.
CHAPTER3
Faced with the impos11'bility of realizing subband coding using m bands covering
the frequency 1paoe regularly and having finite-length filters (whoee length must
be Cm, u required by the Hei.lenberg unoertainty principle), Galand limited
himlelf to the cue m 2. He then had the idea to effect a finer frequency tiling
by IUitably iterating the ~band process. We will see in Chapter 7 that this
arborescent echeme leads directly to "wavelet packets," but we will also see that
theee "wavelet packets" do not have the desired spectral properties.
Subband coding using two frequellcy channels works perfectly. We are going
to describe it in detail.
The input signals :r(n) have already been sampled and are defined on the
integers n E Z; they are arbitrary sequences with finite energy:
.E
111011 =
(~ 1Xol2n ll')"'
and
lf>ili =(
The two filters Fo and F1 arc called quadrature mirTOr filter.~ if, for all signal.(
X of finite energy, one ha.~
(3.1)
= IIX I!:!.
Denote by T0 th<> operat.or DFCl: l:!(Z) - /:!(2Z) and. similarly, let T 1 denotE'
tht' operator DF1 : f:!(Z ) - /:!(2Z.l.
37
I= Tc)To + TjT1.
In (3.2), I : fi(Z) .... fl(Z) is the identity operator. What is much less evident is
that the vectors 1i)To(X) and TiTt (X) arc always orthogonal, as the following
theorem aseerts.
THEOREM 3.1. Let Fo(9) and Ft(9) denote the trarufer function.& of the
filt.e rs Fo and F} . Then the following:two properties are equivalent to each other
and to (3.1).
1 (
(3.3)
The matrix-
..J2
Fo(9)
(3.4)
Ft(9)
e e [0, 211").
Z2 (2Z)
is an isometric isomo:ryhism.
and F 1 .
. Here are some comments on these different properties.
Condition (3.2) is called the perfect reconstruction property: The input signal
X is the sum of two orthogonal signals 10To(X) and TjT1 (X), where the signals
To (X ) and Tt (X ) were given by the analysis. The operators T 0 = F0E and
Ti = FiE are applied to two sequences sampled on the even integers. These are
first extended in the crudest way, which is by replacing the missing values with
O's. Next, this seemingly nonsensical step is corrected by passing the sequences
through the filters F0 and F;, which are the adjoints of Fo and F 1 The correct
result is read at the output. The complete scheme, anal~'Sis and synth~, is
D OW classic:
--x
-...syntbesi$
Condition (3.4) means that quadrat ure mirror filters constitut.(' orthogonal
transformations of a particular type. while (3.3) allow:;: us to ronstruct. quadrature mirror filters t hat havt' finite impulse response .
To se<' this. W(' start with a trigonometric polynomial mo(9) = oo + Ott'ifl +
+oN ci!> P such that lmo(B)I:! + jm 0 (8+1r) j:! = 1 for all B. ('W<' will seE' further
CHAPTER3
38
on how to CODitniCt these polynomials.) Next, we write Fo(B) = v'2 mo(9) and
F 1(9) J2e"mo(9 + w). Then it follows directly that (3.3) i. .atiafied.
The followins jitJe uamplu illustrate the definition of quadrature mirror
6ltere.
Tbe firlt uample is essentially a counterexample becauee it is never UBed, for
a re&SOD that wiD become clear later. It consists in bypusing the operators F0
and F1 More preciaely, To(X) is the restriction of the sequence X= (z(n)),ez
to the even iute&ers, while T1(X) is the restriction of this t~equence to the
odd integers. Condition (3.1) is then trivially satisfied, and the Wlitary matrix (3.3) is
The aecond t!Z.111mple is more interesting. The filter operators are defined by
Fo((x(n)))
1
y'2(x(n) + z(n + 1}) and
1
The orthonormal
System.
The third emmple recaptures the ideal filters preseDted at the beginning of
the chapter. The 2w-periodic function mo(9) is 1 on [0,1r) and 0 on [1r, 211'), and
m1(8)
1- mo(B). Next, as above, define Fo(B) = v'2mo(9) and F1 (9) =
:Jjm1 (9).
In the fourth uample, mo(9) becomes the characteristic function of
[-w/2,,../2) when it is restricted to [-1r, w), and m1 (~) = 1- mo(9).
. The Wt emmple is a smooth modification of the preceding one. Assume
that 0 <a< wfl. We ask that mo(9) be 271'-periodic, equal to 1 on the interval
[-J +a, i-n], equal to 0 on [i +a, 3; - a], even, and infinitely differentiable.
In addition, we impose the condition
(3.5)
= v'2mo(9), F1 (8) =
v'2mt(9), and
='32
39
If theee conditions are aatisfied, the trend and the fiuctuation around this
trend of a signal X are defined, respectively, by Xo = 70To(X) aDd Xt =
7iTt(X).
The trend Xo is twice u regular u X and, u such, cao be auhlampled by
keeping ooly one point in two.
But this sul!sampllog is furnished here by CODStruction. Indeed, 70 : z2(2Z)-+
(l(Z) is a partial isometry, and To( X) constitutes the subsampling of the trend
Xo. In the same way, Tt(X) is t);le subsampllng of the fiuctuation X 1
3.5.
ro.
r"_
rN
ro
40
CHAPTER 3
of fluctootioru and bJI the ltut trend f N e l 2 (fo). The trrml/cmnaticn that mapa I
onto the aequence (r 1 , .. , r N, f N) u clearlJI ortlwgonal, and the invene u imme
diatel11 calculated, ba!ed on the perfect reconstruction properlJI of the quadrature
mirror jilten.
The significance of Mallat's algorithm stems &om the following observation:
For appropriate choice of the filters F0 and F1 , there are numerous cases where
the fluctuations r 1 , . , r N are, at different steps, extremely small. Coding the
signal thus comes down to coding the last trend f N as well as those coefficients
of the fluctuations that are above the threshold fixed by the quantization. Notice
that the last trend contains 2_,. times less data than the input signal and that
the gain is appreciable.
3.6.
(3.6}
Let
IP(X- k),
k E Z,
(3.i )
V;
J (x ) e l o
J (21 x)
l ~.
fit:r> =
z:=u..,,.k)~).k(:r l.
k
= \ :,
(3.8)
j. k E Z.
41
is an orthonormal bb.si~ for L2 (R) and, more precisely, for all j E Z, we have
(3.9)
dj(x>= I ) i"""i"<x>.
A:
The details, at a given scale, are thus linear combinations -of the "elementary
fluctuations," which are the wavelets related to that scale.
This analysis technique will be discussed again in the next chapter under the
name "multiresolution analysis."
Given two functions :p(:z:} and tP(XJ, called, respectively, "the father and
mother wavelet .~ it is possible to define two quadrature mirror filters Fo and F1
by way of the operators To = DF0 and T 1 DF1 This is done by relating the
approximation of the function space L 2 (R) t hat is given by the nested sequence
of subspa.ces l'J to the approximation of the real lineR that is given by the nested
sequence of grids r; = 2-i z .
To do this, we consider that the function :Pj.k(:z:} 2il 2 t,0(2ix-k) is centered
around the point k2-J , which would be the case if t,O(:r) was an even function .
We associate the point k2-i v.ith the function V'J,A: This gives a correspondence
between f ; and the orthonormal basis {lfJ,kk E Z} of V;. At the same time,
J2{r ;} is identified isometrically with \.j .
To define the operator To : /2 (r;+l ) - l 2 (f;), it is sufficient to define its
adjoint T0 : l 2 (f;) - 12 (fJ+1) . This adjoint 10 is a partial isometry. It is
constructed by starting v.i tb the isometric injection l'J C l'J+1 and by identifying
V; with t2 (f;} and V;+ 1 v.ith l 2 (fJ+ 1 ) , as explained above.
The orthonormal basis 1!1;.1< of W1 allows us to identify W; with 12 (f;) in the
same way. The isometric injection of H ~ C \'}~ 1 , interpreted v.ith this identification. becomes the partial isometr:y
Finally. the couple (;;. v) is represented by the couple (T 0 . TJ) or, which
amounts to the same thing, by the pair ( F0 F 1 } of the two quadrature mirror
filters. This crucial observation is due to Mallat.
Mallat also posro the converse problem: Given two quadrature mirror filters
F0 and F 1 is it possible to associate v.ith them two functions .,: and 1! ha,ing
properties (3.6 ). (3.i ). and (3.8)?
Although the comefS(' is incorrec-t in general. it is correc-t in numerous cases.
and this led to the construction of Daubechies's wavelets.
Our first and third e-xamples of quadrature mirror filters show that the converse is generally false. There art' no functions If and t/1 behind these numerical
algorithms.
The second example is related to tht' Haar system. ThC' function ,'(:r) is thC'
index function of 10.1 ). and \; is rompos!'<i of step functions that are constant
on each interval jk::?-J. (k + 1)2-J). k E Z.
ThC' fourth p.xampk IC'.ads to ~shannon's wavclC't.c;.~ Tlw fuuction -..~(.:r.) i!: thr
carqjnal sinC' ddimxi by i~;: .
42
CHAPTER3
Finally, the last example ia more interesting, bec&u8e both of the functions
~(z) aDd ,P(z) belong to the Schwartz class S(R), which COilliBts of the infinitely
(3.10)
1:
(3.11)
(3.12)
g(x)cU- = 1,
g(2k1r} = 0 if k E Z,
and
k # 0.
One can then restrict the filtered signal to the grid hZ.
We assume that these conditions are satisfied throughout the discussion. and
when /'\ tends to +oo, j and k being fixed. Refer to the "herringbone" schemt>
for the definitions of f,..,r,...r,..- 1 Here are the results 16].
THEOREM 3.2. Assume tha.t th( rmpulse responses of the quadroture mirror
fiJten Fo and F1 decrease rapidly at infinity and that the transfer function Fo(B)
of Fo 1atisjies F 0 (0} ~. F0 (8) ::!- 0 if-~ ~ 8 ~ ~ . Then Mallat's "herrinp
bont,. olporithm. applied to f g,... a. indicated above. converye. to the anoly$t.<
off in an orthonormal wavelet ba.<r.<.
43
00
lim
N-+oo
fN(k)=1_
lim TN(k)
N-+oo
/(z)ip(z - k)dz,
00
The functions V'(z) and ,P(z) are, respectively, "the father and mother" of
the orthonormal wavelet basis, as explained in the preceding sectiOilA very beautiful application of this result (to which we return in Chapter 4)
is the construction of the celebrated bases of Daubechies.
3.8.
PN(t)
= 1-
CN
1'(sinu) ~'~- du =
2
'Y1ceu.t
Jk J$21'1-1
=
h,.e-''"
with the constant CN > 0 chosen so that PN(Tr) 0. There exists (at least one)
finite trigonometric sum mo(t) = ~ ~N- 1
such that lmo(t)ll = PN(t)
and mo(O) = 1. The coefficients h" constitute the impulse respoDSl' of the filter
Fo and are real.
Under these conditions we knov: from Theorem 3.2 that the functions 9 and t/
exist and that they form a multiresolutional analysis. We now use these results
to construct cp and t " explicitly. The function cp. which we seek to construct,
must be a solution of the functional equation
21'\ - I
(3.13)
V'(z)
= v2. L
0
ht.V'(2z- k)
and
L:
<,?(x)dx = 1.
This functional equation follows from the inclusion V0 C V1 and the fi\Ct that
<,?(.T - .1.). k E Z. i!' an orthonormal basis for l (, and .J2<,?(2x- k). k E Z. il' M
orthonormal basi~ for \ "1
44
CHAPTER 3
(3.14)
and the principal difficulty in this coDBtruction is to show that one has cp({)
0(1{ 1-m) at infinity, where m = m(N) tends to infinity with N .
On the other hand, it is almost obvious that the support of cp(z} is in
I0,2N -1J.
The fact that cp(x- k), k E Z, is an orthonormal sequence is a direct consequence of Theorem 3.2 and the fact that mo(t) i= 0 on !- 1r f2,1T/ 2J.
To determine the Fourier transform 1}J({) of the wavelet t/J(z ), we write
m l (t) ei(l-2N)t mo(t + 1T). Then
(3.15)
3.9.
Conclusions.
The functions 1/JN used by Daubecbies to construct the orthonormal bases named
for her are new "special functions." These "special functions" never appeared in
previous work, and their only definition is provided by (3.14} and (3.15). This
means that the detour by way of quadrature mirror filters and the corresponding
transfer functions was indispen&aQ!e. In other words, it never would have been
possible to discover Duabechies's V.:avelets by trying to solve directly the existence
problem: Is there, for each integer r ~ 0. a function 1b{z ) of class cr such that
2il21/1(2i x- k), j, k E z, is an ortbonormaJ basis of L2 (R)?
The quadrature mirror filters will be used again in the algorithms for numerical image processing, which we describe in the next chapter.
Bibliography
R. HINCORANI. AND E. SIMO~CELLI. Ort}wgonol pyramid tromfurm.
for tm4ge codsng. SPIE, Visual Communications and Image Processi.nt; 11. Cambridge,
MA, October Zi- 29, 1987, Vol. 84 5. p p .
I. DAUBECHIES. Onhonorrn41!wu of compoctly 8Upportet/ woveleu. Comm. Pure Appl.
Math .. 41 (1988). pp. 909-996.
D . ESTEBA~ AND C. GALAND . Applicolson of qu;~droture msn-or filt er to split bond voice
codi"!) ~tems. ln\ernational Conferencc
Aooustia. Speech. a.nd Signal Processing,
Washi.n((I.On. DC. May 1977. pp. 191- 195.
C . GA LAND. Codogc en ow-bondu: theorsr et applsC4lU>ns ci lc com,....,...ton numenqut'
du ftgnal de perote. Thesis, Universh~ of !'\ice. Nice. France. March 1983.
S. MALLAT, A u.eo., farmultiruolul.son h9fl4l dt!a>mpOntion: The wavelet ,..,..es entatson.
IEEE Traru.. Pattern Aual. Macbuw ln&.ell.. 11 (1989). pp. Gi~-6!13.
Y . MEYER AND F'. PAIVA. Con~ dt' l'olgorithm~ de Mollet, pl't'prinL. CEREMAOE,
Uniwnity of Pari~ Dauphine. Pari$. F'nu.aC't' .
(1] E.
12;
!3j
(4)
(5]
(6;
H . A DELSO~ .
ro-ss.
on
CHAPTER
4.1.
Introduction.
CHAPTER4
46
4.2.
For the rest of the discussion, r; = 2-rl 2 will denote the sequence of nested
grids used for image processing. It often happens that the image is bounded by
the unit square. in which case we will speak of a 512 x 512 imag~ to indicate
that; 9; similarly, a 1024 x 1024 image will correspond to j
10.
At this point, we are working with images that are already digitized and
appear as numerical functions. The raw image that provided t hese digital images will be, for us, a function f (x. y). This function can b~ very irregular,
either because of noise or because of discontinuities in the image. For example.
discontinuities can be due to the edges of objects in the image.
The sampled images f; are defined on tht> corresponding grids f; = 2-iz 2
These sampled images are obtained fron1 tht> original physical image. that is,
/(%, 71), by the restriction operators Ri : L ~ (:R 2 ) - / 2 (f; ). The&' operators R;
47
will be defined in the folloolring pages. They are the same type aa thole u.eed in
DUJDerical analysis to dilc:reti.ze an irregular function or distribution.
The fundament4l dUcovt:rJI of Burt ond Adellon il the e:rilten of 1'Mriction
operator~ R; witA the prDf1ertJI that, for all initial im4gu f, the ompled imGgu
I;= R;(/) ore rel4ted to eoch other btl utremeiv mnple C4tuolitv relotion.f.
These causality relatioos, of the type "fine to eo&r&e," allow IJ- 1 to be calculated directly from /; without having to go beck to the original physical image
f(x , Jl).
In order to define the restriction operators R;, we first consider the case d a
grid given by x = hk, J1 = hl, where h > 0 is the sampling step and (k, l) E ZxZ.
Very irregular functions should not be sampled directly and, therefore, the
image must be smoothed before it is discretized. This leads to the following
classic scheme
tz
(4.1}
c(k, l ) = ~
By v.'Titing fr?(Z, y)
(4.2)
jj
g ( k-
= (f,cph( -
kh, -lh)}
whert> \U. t) =If u(x.ylv(x , y )dx dy and where denotes the (dummy) variablt>
of integration.
Tht> extension operator enables us to extend a sequence c(k, l} defined on h'l':!
to a regular function on :R~ . In this sense, it is inverse to the filtering/ sampling
operation. We define the extension operator to be the adjoint of the restricti011
operator so it is given by
(4.3)
CHAPTER 4
48
The simplest examples are given by the spline functions. We consider the onedimensional case to simplify the notation. U we let T be the "triangle" function
T(x} sup(l-lxi,O), then (4.3} yields the usual piecewise linear interpolation
of a discrete sequence. A second choice is given by tp = T T, which is the basic
cubic spline.
Returning to the general case, we ask that the operator PhRh, composed of
the restriction operator followed by the extension operator, have the property,
that for all functions f E L 2 (R2 },
(4.4}
PhRh(/ ) --+
(4.5}
1<.0(0. 0)1 = 1,
49
Burt and Adelson aaked hO\\ to determine the functions r.p such that (4.6}
and (4. 7) are satisfied. Stated this way, the problem is very difficult, for most
of the usual choices of smoothing functions do not have these properties. To
resolve this difficulty, Burt and Adelson proceeded the other way around; that
iB to say, they sought to construct r.p from the operators T;.
For this it is necessary to derive some consequences of (4.7). The first is
that the operator To : P(Z2 ) -+ l 2 (2Z2 ) can be written as To ::: DFo, where
Fo: 12 (2 ) -.l2 (Z2 ) is a filter operator and where D : l2 (Z2 ) -+ l 2 (2Z2 ) restricts
a function defined OD Z2 to 2Z2 D is the decimation operator, which we have
already encountered in Chapter 3. The fact that To bas this form is a consequence
of the fact that To commutes v.-ith all even translations.
Thus, if X= (x(k)h:ez2, we can write
To(X )(2k) =
(4.8)
L w(2k -
l )x (l) ,
lEV
where w(k) is the impulse response of the filter F0 . For convenience, V! assume
(as Burt and Adelson did) that w(k) is real.
--,---,...,.
U we apply To to x (k) = J f (:r)r.p(x - k)d:r, then from (4.7) we get
J f (x)r.p (j - k)dx. Condition (4.8) is thus equivalent to
(4.9)
It
(4.10)
where
(4.11)
lr
(4.12)
l )fiCitH1'1l .
r.p({.f7)=mo
The second consequence that we derive from ( ~. 7) is that all of these conditions (for different. j) are in fact equivalent. This cau be seen by making thr
change of variables :r ,_ 2i :r and y .- 2i y in (4 .9 ). and intep-ating both sides
against f (:r.,y). We then bavt'
(4.13)
R1 _ 1 (f} (:.?-i+lk)
=L
IE Z'
w(2k -l)R;(f)(2-il).
50
CHAJ>TER4
T;(X)(Ti+ l k)
{4.14)
= E ~o~(2k -
l)x(2-1 l)
lEZ2
when X= (x(2-il)) belongs to l 2 {r;). Tbe point is that the eequence ~o~(k), k E
Z2 , is the same for all the operators T; . Working backwards, Burt and Adelson
began with a finite eequeoce of coefficients ~o~(k,l) such that E., E1 ~o~(k,l) = 1.
They defined mo({, 17) by (4.11) and then cl; by (4.12). Then the first question to
arise is whether the second member of (4.12) defines a square-integrable function.
If this is the case, we call this function <{;, we define the restriction operators R1
in terms of the Fourier transfonn of this function, and we define the transition
operators T; by (4.14). Then R;- 1 = T;Rj for all j E z.
rcr,)
analysis.
Burt and Adelson proposed a very ori~inal function for ..:. IUld this will lx
our third example. Takt> w(O) = 0.6: :(::::1 ) = 0.25: ..:t::::?) = - 0.05: and
51
w(k)
4.4.
Image compression is one of the uses of the pyramid algorithms. Burt and
Adelson's algorithm, which we describe in this section, will later be compared
with other algorithms (orthogonal pyramids and hi-orthogonal wavelets) that
perform better. All of the pyramid algorithms act on images that are already
sampled and never on the original physical image. In other words, the function
cp we have tried to construct using the sequence w(k) is never used. Then why
have we worked to learn its properties? The answer is given at the end of 4.6,
but we can indicate the idea here, We want, on reaching the summit of the
pyramid, to have a moderate, softened image and not one that is chaotic and
noisy; this is precisely where the regularity of the function cp comes into play.
The Burt and Adelson pyramid algorithms use only the transition operators
T; : 12 (r;) ..... 12 (r1 _ 1 ). All of these operators are the same, except for a change
of scale: therefore, we are going to assume that j
0.
The discussion of the algorithm begins with the definition of the trend and
the fluctuations around this trend for a sequence I belonging to /2 (f 0 ). This
trend cannot be To(/) because it "lives in a different universe" and cannot be
compared to I . To define the trend, it is neeessaz: to leave the coarse grid 2r0 ,
where To(/) is defined. and return to the fine grid r 0 where I is defined. This
is done by using the adjoint operator 10 : l 2 (2r0 ) ..... 12 (r0 ), and the trend off
is defined by TciTo(/).
Wt' clearly want th<' trend of a very regular function to coincid<' with that
function. Thi.c: leadg to t h<' r('Quircments that 70T0 ( 1) = 1 and, mor<' ~cnemlly.
52
CHAPTEH 4
that 7()To(P) = P for all polynomials P of degree less than or equal toN, for
some fixed integer N . This condition i8 equivalent to the following: The function
mo(~. '1) , defined by (4.11), must vaniab, along with all of its derivatives of order
less than or equal toN, at pointa (e11r, t21r), 1,2 E {0, 1}, with the exception
of the origin. At the origin, one must have
at all the points of this region. Thus the Burt and Adelson algorithm becomes
more effective as N increases.
To define the coding and compression algorithm of Burt and Adelson. we
begin with the "fine grid" r m = 2- '"Z2 and 8 numerical image fm sampled on
this line grid. This numerical image is. in fact, the restriction to r'" = 2- mz2 of
a physical image f e L 2 (R2 ) . This means that fm is the restriction in the usual
sense of the convolution product f Dm where Dm (:r:,y ) = 4mg(2'"x.2'"y). The
properties of the function g were indicated in 4.2: in addition. we shall assume
that the integral of g is equal to 1. However it is not necessary to '-return to the
physical image- f to use the algorithm.
Bun and Adelson replace fm by thr couple (trenc!. fluctuation). But thr
trend, which is given by T;,TmUna ). is completely defined by TmUm) Thb
amounts to saying that the trend is sufficiently regular that it can bt' coded by rt'taining one pixel in four. This coding is given by TmUm ). In summary. Burt &~ d
Adelson codr fm with the couple ITm(/m). fm - T;,Tmlfml;- Then thr fluct.ulltion. denot.ed by Tn is not processed further. They v:ritt' fno-1 = TmUm) and it crate t.hr procoourc. /no-1 is coded by lfno -':: Tn-1 ). v.herr f.,, _~= Tno-l Uno - 1)
and Tm - 1 = /no-1 - T,;,_.Tno-lUm-1).
53
(4.17)
Condition ( 4.17) is satisfied in all the regions v.here the image exhibits a certain
regularity.
If we suppose that the starting image fm is bounded on a square of side 1,
then the algorithm is stopped on reaching the summit of the pyramid, which is
the grid ro.
The image fm is coded by the sequence (/o, r 1 , ... , rm}, where / o, defined on
r(l, is a scalar and where the ri = (1- Tj1j)f1 , 1 ~ j ~ m , are the different
fl uctuations.
The diagram below gives a schematic description of the algorithm. The
columns correspond to the different grids r ;.
Tm
fm j
Tm
rm
Tm -1
r m-1
Tm-1
r---------------, i
1
Tm - 2
TJ
rm-2
T,
fo
r1
ro
Running the algorithm the other way, which is reconstructing fm from the
code. is very easy. Begin v.'ith / o and TJ. and compute .fl = Ti fo + r 1 . In the
same way. reconstruct h by h = Ti h + r2. and continue until fm is recovered.
As we will set> a little later, this algorithm is advantageous only if most
of the fl.uctuations are zero. Otherwise it is clearly disadvantageous because
information is wasted.
4.5.
Befor<' lea,'ing tlw first version of Burt and Adelson's algorithms. wt> ar<' going to
describe a continuous version of it. The interplay between the discrett> algorithms
and their continuous versions. which is implicit in the work of Burt and Adelson.
wa.. madc explicit by Mallat. and the author.
\\'( consider tl.Jc genrral cas<' becaUSt> dimension two plays no particular rok
in tht following definition.
A mult in::.H1luti011 analy.i.~ of L 2 (1R" ) i. an increasing sequence of clo.~ed subspace:< (\ ~ l.1 ez of L 2 (R" ) hmnng the f ollounng three properties:
CHAPTER4
n
00
(4.18)
V; = {o}.
-oo
""' V;
U
u den.te in L2 (Rn),
-oo
(4.19)
(4.20)
eZ
Recall that if H is a Hilbert spa.oe, a Riesz ba.$is (e; );EJ of H is. by definition,
an isomorphic image T : H- H of a Hilbert basis (/; ); EJ of H . (Note that T
is DOt necessarily an isometry.) Then each vector X e His decomposed uniquely
in a series
(4.21)
:z
= L a, e,
;EJ
where
even+oo.
We say that the multiresolution analysis (V; );ez is r-regular if it is p ossible
to choose the function 1,0(x ) in {4.20) so that, for all integers m ~ 0 and all
rER",
(4.22)
where a = (o 1 , .. on) is a multi-index satisfying o 1 + + o, :S r and where
ff' =(o/oxt} 0 1 (8/8xn) 0 " .
We return to the two-dimensional case. Here, a multiresolution analysis is. in
a oertain sense, a particular case of the pyramid algorithm. To see this. suppose
that the function o;. which is defined by (4 .1:?). has t he following additional property: There exist. two constants Cz ~ ('1 > 0 such that for all scalar sequences
(aJ:)J:eZl
55
The pyramid algorithma auociated with multireeolution analyaes are the only
ones that we will study in the following sections. They have some quite remarkable properties. For example, the restriction operator R; is then an isomorphism
between "'J and l 2 (r;). In this caae, the equation R;- 1 = T;R; can be eolved
directly; it is sufficient, in fact , to restrict the two sides of the space in V; in
order to invert R;.
Not all pyramid algorithms are related to a multiresolution analysis. A counterexample is given by one of the pyramid algorithms that we presented in 4.3.
In this example the function tp(:r) is 1/2 on [-1,1J and 0 elsewhere. Thus
(4.24)
= ~.'
4.6.
Shortly after the discovery of quadrature mirror filters by Esteban and GalaJid,
Woods and 0 'Nei1(13J had the idea to apply this technique to image processing.
They thus obtained the first example of an orthogonal pyramid. We are going
to set aside, for the moment, the specific construction carried out by Woods and
O'Neil using separable filters. Instead we will present the notion of an orthogonal
pyramid in complete generality. Then we will return to the particular case
where the quadrature mirror filters appear in the construction of an orthogonal
pyramid.
The Burt and Adelson algorithm is flawed because it replaces information
coded on JVJ pixels by new information whose description requires ~ !f2 pixels.
This criticism, which we will analyze in a moment, is not justified because in
many examples of real images, most of the values of the gray levels on the !f2
pixels are in fact zero, and the unfavorable pixel count, where !f2 Dec:omes
; N2, occurs very rarely.
Let us examine, however, why the information has been wasted or, more
precisely. where the inefficient coding occurs. At the start. the image f has been
coded on N2 pixels. Next, we replace this by the couple (To(/), (1- T0To)(f)J.
which is composed of the coding for the trend and the complete description of
the fluctuations around the trend. The description of To{f) requires
pixels.
whereas the description off- TtiTo(f) continues to require N 2 pixels. In all. we
use N2 +iN 2 pixels. At the ne>..'1. step . the pixel count becomes N2 +!N 2 +T6N 2
and so on ... At. the end. we will have used N2 + ! N:! + fr,J\1'2 + ~ 1. or
2
approx.imat.ely
pixels. The "wasted" pixels appear because the fluctuations
f - TjT; l /) have not been coded efficiently.
The orthogonal pyramids are a particular class of pyramid algorithm.~ that
code the fluctuations with ~ N 2 pixels. With this scheme there is no waste.
When the original image f is replaced by coding the trend and thr fluctuation.
the required pixels are ~ N 2 and ~ J\1'2, respective!~. and tht' volumt' of data
remains constant.
!!YJ
5N
5{j
CHAPTEH.4
We 1ay that a pyramid algorithm i.9 orthogonal if, in the 1eme of tile usual
1calar product of l 2 (ro), the trend TOTo(/) and the ftuctuatimu I - T0To(/)
around thu trend are orthogonal for each image f defined on the grid ro.
Write H
Ho TOTo(H), and H1 (1- TOTo)H. If the pyramid
Ho e H 1. Since the dimension of Ho is a
algorithm is orthogonal, then H
quarter that of H, the dimension of H 1 is 3/4dimH, as announced.
An equivalent definition of "orthogonal pyramids" requires the adjoint 1(j of
the operator To to be a partial isometry. (Recall that 7(i is defined on l 2 (r_I)
with values in l 2 (r 0 ) .) This takes us back to one of the characteristic properties.
in dimension one, of the "low-pass filter" To in a pair of quadrature mirror
filters, (T0 , TI). And this observation prompts us, in dimension two, to look for
the corresponding second filter, T1 We will see in a moment that three filters
are necessary in this case.
But before this, we show how to construct some pyramid algorithms. We
return to the "transfer function" mo(-'7) defined by (4.11). The P~Tamid algorithm is orthogonal if and only if
= rcro).
= 1.
This condition is completely analogous to the one on the transfer function mo({)
in the case of two quadrature mirror filters (Chapter 3, 3.3).
Continuing this comparison, we consider the function tp(x , y) in L2 (JR2 ) n
L 1(JR2 ) defined by (4.12) and normalized by II tp(x, y)dx dy
1. We might
expect that the sequence tp(x-k;y-1) , k.l E Z , is orthonormal. and this is true
in most cases. However, the proof involves a delicate limit process. passing from
the discrete to the continuous, and certain-orthogonal pyramids do not lead to
orthogonal sequences.
This difficulty already appeared in dimension one for tht> quadrature mirror
filters. The condition we assume here on mo({. f'J), which is sufficient to allow
passage from the "discrete to the continuous." is completely analogous to the
condition we used in dimension one.
It is sufficient to assume that mo({.f'J) -=1= 0 if-~ $ {$~and-~$ f'J $ ~
Then r,?(x - k , y - l ). k.l E Z, is an orthonormal basis of a closed subspace l0
of L 2 (1R 2 ) . At the same time. 2i r,?(2J:r - k . 2Jy -1). k.l E 'l}. is an orthonormal
basis for tht> subspact> l j: the extension operator P3 : i 2 (r;) ~ l j is an isometric
isomorphism: and tht> restriction operator R; : L2 (1R2 ) ~ P (f ;) is decomposed
into the orthogonal projection operat.or from L 2 (R2 ) ont.o \ j follmved by the
inverse isomorphism Pj 1 : V; - 12 (ri).
This allows us to define e>."J)licitl~ the transition operat.ors T; : /2 (r;) ~
2
l (r;-d. which. we recall. v.rert> defined implicitly by T; RJ :::: R;-J US<'
the operator P; to identify l2 (r;) with l j. and similar!~ US<' PJ- l to identi~
l 2 (r;-d with l j_ 1 . Ha,ing made these identifications. the t.ransition operator
T; : l 2 (f; ) -+ l 2 (r,_1 ) corresponds quitr simply t.o tht> ortho~onal projection of
l j on \ ~ _ 1 . which il< Pj _ 1T;P;-l in our notation.
fJ7
(4.26)
for all f belonging to l 2 (Z2 ). The left term is of coun;e computed in l 2 (z2),
whereas each term on the right is computed in l 2 (2Z2 ) .
One of the important results in the theory of orthogonal pyramids is the
existence and constructability of these operators R1o R2 , and R 3 . Furthermore,
if the impulse response w(k, l) of To decreases rapidly a t infinity, the operators
R 1 R 2 , and R3 can be constructed to have this same property.
Once R1o R 2 , and Ra are constructed, we can construct the corresponding
wavelets t/Ji, tP-J. and 1/Ja. Assuming that mo ({, 17) i= 0 if I{I $ ~ and 1'71$ ~ .
t hese three wavelets are defined by
(4.27)
1/l;(:c, y)
=4 L L w;(k.l)<p(2x + k , 2y + l) ,
It
= 1,2, or 3,
is deduced immediately from (4.26). Identity (4.28) provides a particularly elegant solution to the problem of coding the fluctuation f - 7(jT0 (/). This flur
t uation i:; exactly
CHAPTER4
fm
~=~=~=~=~= ~:
"'
""
~~ ~ ~
W2
w3
fl(fo)
Tbe wavelets appear in the asymptotic limit of this scheme. This limit is
&Ibn on the number of steps m , which must tend to infullty. We start with a
function J(x,JJ) belonging to L2 (R2 ). We "restrict" J to the fine grid r m
tbe classic acbeme. This means we have a fixed regular function g, which derapidly at infinity and whose Fourier transform g({, '1) satisfies g(O, 0) = 1
2k1r,2l1r) = 0 if (k,l ) =F (0,0). We write 9m(x,y) = 4mg(2mx,2my). Fi-
...,-put.a"
21
2;
J
J
'Albert Cohen established this result undt>r vel) general b~potheses [4]: of
themost .convenient is that mo(~ . f7) ~ 0 if 1{1~~and 1'11$ ~
beauty of this theol)' lead~ ont> to think that it provide$ the correct
lli.poa~e to the image-processing problem. indeed. the image i.< decomposed by
canalyN into information that is independent from one scale to another.
thil ~ with tM generol philosophy expressed in the introduction. These
~radent pocket$ of information arc represented by the trend m V
o and the
. .CQIGIDOIU fJ n;. WJhose orthogonal .um i.< equal to f . Thr charocteristic
ofVo-j is 2i. Furtltcmon:. caclt !; E H _, s. it.,r.lf dr.composed inte> e>rl.lwge>ual
~
59
or 1/Js.
The Haar system is the simplest example of orthogonal waveleta, and it has
been used for image proc:eeeing for a long time. However, it has the disadvantage that, following quantization, it introduces rather harsh "edge effects," thus
producing unpleasant images.
This prompts us to say a few words about the quantization problem. From
the point of view of the mathematician, aU orthonormal bases allow the signal to
be reconstructed exactly. This is not the point of view of the numerical analyst
or image specialist. In practice, the coefficients from the decomposition must
be quantized, whether we like it or not. These approximations arise from the
machine accuracy or are imposed by a desire to compress the d&ta. II it is true
that
f(x, Jl)
a;e;{z, Jl),
=L
jEJ
what happens to the first term when the a; are replaced by coefficients 0:; satisfying IO:;
I :S t:, where t: > 0 is related to the machine precision? Anything can
happen if we are not using a .robust orthonormal basis, and we know today that
many orthogonal wavelet bases are infinitely more robust than the trigonometric
basis.
. In spite of this, orthogonal wavelets (and the corresponding pyramid algorithms) have not completely satisfied the experts in image processing. The criticized defect is the lack of symmetry. The function cp(x) ought to be even, while
the function ,P{z) ought to be symmetric in the sense that ~(1 - x)
1/l{z).
These properties are satisfied by certain orthogonal waveJets, but they do not
bold for wavelets with compact support. The Haar system is the only exception.
This lack of symmetry is reflected in visible defects, again following quantization. These visible defects do not appear when one uses symmetric b~-ortbogonal
wavelets having compact support. We introduce these wavelets in the next
section.
-a,
4.7.
Hi-orthogonal wavelets.
(4.29)
/(:r }
=L
DO
L o;.~r1P;.~r(x),
-oo-oo
CHAPTEH 4
60
(4.30)
a;,lc
L:
/(z).j,j,lc(z )dx.
Here
and
Then the general theory of Cohen. Daubechies, and Feauveau tells us that,
for this choice of t/l{z ), 2il 2 tJJ(2i:r- k). j , k E Z. is a Riesz basis for L 2 (R) and
the dual basis has the same structure, which is given by 2il 2 tb( 2' :r - k), j , k E z.
However, in this particular case, the dual wavelet. is not a continuous function .
To repair this we need to take a more general approach.
We assume that 1/; belongs to a set of functions that are continuous, have
compact support. are linear on each interval [k/ 2, (k + 1)/2]. k E Z. and are
~'lllJiletric with respect to 1/ 2 (in the sense that tjl(l- z) = w(:r)). Then the
Cohen- Daubechies-Feauveau theory states that we can choose a tl: from this set
so that the dual wavelet tb is a function in the class c;r and has compact support.
Here is bow ,P and~ are constructed. Start with the triangle function '(z )
sup(1 -lxi.O), which v.'8S mentioned in 4.2. Write mo{{) (ca; U 2) 2 : then. by
construction, <P({) = mo{U2)<P{{/ 2). !\ext consider g_-.({) = c,. J," (sin t) 2 "'+1dt.
where CN > 0 is adjusted so that 9N(0} 1.
'
If moW is defined by mo({ )mo{{ ) = g,.({l. then
(4.31)
+ r. ) =
1.
(In the construction of the Daubechi~ v.-a,elets. one imposed till' condition
lm(l({)l2 9N({).)
Define It e L 2 {JR) by its Fourier transform
{-t32)
{i]
(4.33)
j ~(x)cp(x- k)dx = 0
if k -:1- 0,
and
=1
if k
= 0.
The function ~(:z:) is even , its support is the interval [-2N,2N), and ~(x) is in
the Holder space cr for all sufficiently large N.
It is clear that
v,+l
z.
and
m1 (U2),P(f./ 2).
CHAPTER 4
Bibliography
(1) M. AN'TOHINI , M. BAIU.AUD, J. DAUIIECHIES, AND P . MAT1UI':U, /mage coding wmg
we:& pontUotion in the uoveld trorujt1f"m dcm4in, IEEE lnternatia<W Conference
oo Acou.Uca, Speech, aDd SicnaJ Procaeing, Albuquerque, NM, April 1990, pp.
:t.m-2300.
(2) E. H. ADIWION, R . HINGORANI , AND E. SIMONcr:LLI, Orlhogon41 pymmid trorujormJ
far imAge coding, SPIE, Visual Communicatiooa and l.mage Procesaing II, Cam
bridae, MA, October 27- 29, Vol. 845, pp. 50-58.
(3} P . J . B URT AND E . H . ADELSOI" , The L4plocian wramid ,.. a compact image code,
IEEE Trans. Comm., COM-31 (1983), pp. 532- 540.
(4] A. COHI':N, Onddettu, anolv.U multiruoluttmu et troitement numo!rique du 8ig114l. Thteis, CEREMADE, Univet'tity or Paris-Dauphine, Parilo, France, Sept.ember 1990
(5) J. C. FEAUVEAU, Anolv-'e mulbrUolution par on.deleuu non orthogon4lu et b<uu d~
jiltru numbique, Theei.6, Univet'tity or Paris-South, Pari.&, France, January 1990.
(6] J. FROMEPIT, 7hlitement d 'im4gu et applicati<ml de 14 troruformee en ondelett.es, Thteis, CEREMADE, University of Paris-Dauphine , Paris, France, November 1990.
(7) J . FROMENT AND J . M . M OREL. Anolve multiechelle, Wicn 1ti.rio et ondelette$, in
Lu ondelett.es en 1989, P. G . Lemarie, eel., Lecture Notes in Mathematics 143S.
Sprinaer-Verlag, Berlin, N- York, 1990, pp. 51-80.
(8] S. MALLAT, A ~for multiruolution rign41 decompo.Stion: The wavelet repruenta
tion, IEEE Trans. Pattern Anal. Machine Intelligence. 11 (1989), pp. 674~93.
191 D. MARR, Vifton, A ~ invutiganon onto the human repruentation And
priiCUiing of WuCll in/orrno.hcn. W. H. Freeman and Co.. N- York, 1982.
(10] Y. MEYBR, Qrule1d;t.u, Hermann. Paris. 1990.
(11) P . TCBANITCHIAN, BiorthogO'flGliU et Thi~ cLe. Opi.rateurs, Revista Matematica
lberoamerican~ 3 (1987) , pp. 163-189.
(12) M. VETTERU, Wavelet. and fil.ter bank,: &l4t~h.ips and new re.rolts, IEEE Conference on Ac;oustics, Speech and Signal Processing, Albuquerque. N M, April 1990, pp.
1723-1726.
(13) J. W. Wooos AND S. O'NEIL, Suhband coding of irn4ges. IEEE Trans. Ac:ous. Speech
Signal Process., 34 {1986), pp. 1276-1288.
CHAPTER
5.1.
Introduction.
Dennis Gabor (1946) and Jean Ville (1947) both addressed the problem ofdeveloping a mixed signal representation in terms of a double sequence of elementary
signals, each of which ocx:upies a certain domain in the time-~ plane. In
the following sections we will define what is meant by "time-frequency plane"
and "mixed representation," and we will suggest several choices for "elementary
signals."
. Roger Balian (1) tackled the same problem and expressed the motivation for
his work in these terms:
One is interested, in communication theory, in representing an oscillating
signal as a superposition of elementary wa!Jele~ each of which bas a rather
well defined frequency and position in time. Indeed, useful mformation is often
conveyed by both the emitted frequencies and the signal's temporal structure
(music is a typical example). The representation of a signal as a function of time
provides a poor indication of the spectrum of frequencies in play, while, on the
other hand, its Fourier analysis masks the point of emission and the duration of
each of the signal's elements. An appropnate representation ought to combine
the advantages of these two complementary descriptions; at the same time, it
should be discrete so that it is better adapted to communication theory.
Similar criticism of the usual Fourier analysis. as applied to acoustic signals,
is found in the celebrated work of Ville:
If wt> consider a passagt> [of music) containing several measures l v:hich is the
least that is nt>eded) and if a note. la for example, appears onct- in the passage. harmonic analysis vdll giw us tht> corresponding frequenc~ \'l'ith a certain
amplitude and a certai n phase. without localizing the la in time. But it is obvious that there arc moments during the passage when one does not bear the la.
The (Fourier] representation is nevertheless mathematically correct because the
phases of the notes near the la art- arr~ed so as to destroy this not' through
interferenct- when it is not heard and to reinforcr it, also through interference.
when it is heard: but if thert- is in this idea a cleverness that speak.c: well for
mathematical anru~-sis. on<' must not ignore thl' fact that it. i!' also 11 distortion
of reality: indet"<l whl'n th' la i~: not heard. thf' trur reason is that tht la i~ uot.
emittro.
f.:l
64
CHAPTEk 5
Thus it is desirable to look for a mixed definition of a signal of the sort advocated by Gabor: at each instance, a certain number of frequencies are present,
~ving volume and timbre to the sound as it is heard; each frequency is associated with a certain partition of time that defines the intervals during which
the corresponding note is emitted. One is thus led to define an instantaneous
spectrum as a function of time, which describes the structure of the signal at a
given instant; the spectrum of the signal (in the usual sense of the term), which
gives the frequency structure of the signal based on its total duratton, is then
obtained by putting together all of the instantaneous spectrums in a precise v.oay
by integrating them with respect to time. In a similar way, one is led to a distribution of frequencies with respect to time; by integrating these distributions,
one reconstructs the signal...
Ville thus proposed to unfold the signal in the time-frequency plane in such
a way that this development would lead to a Ilillced representation in timefrequency atoms. The choice of these time-frequency atoms would be guided by
an energy distribution of the signal in the time-frequency plane.
The time-frequency atoms proposed by Gabor are constructed from the function g(t ) = ?r- 114 exp( -t 2 / 2) and are defined by
(5.1)
w(t)
The parameters w and i{) are arbitrary real numbers, whereas h is positive. The
meaning of these three parameters is the following: w is the average frequency
of w (t ), h > 0 is the duration of w(t) , and i{) - h, i{) +hare the start and finish
of the "note" w(t). Naturally, all this depends on the convention used to define
the "pass band" of g(t).
The essent1al problem is to describe an algorithm that allows a given signal to
be decomposed, in an optimal way. as a linear combination of judiciously chosen
time-frequency atoms.
The set of all time-frequency atoms (v.'itb wand t 0 varying arbitrarily in the
time-frequency plane and h > 0 covering the whole scale axis) is a collection of
elementary signals that is much too large to pro\'ide a unique representation of
a signal as a linear combination of time-frequency atoms. Each signal adm1ts an
infinite number of representations. and this leads us to choose the best among
them according to some criterion. This criterion might be the one suggested
b~ \'ille: The decomposition of a signal in timt"-frequency atoms is relatt>d to
a synthesis. and this synthesis ought logically to be done in accordance with
an analysis. The analysis proposed by Ville v.'ill be described in the follov.-ing
sections. However, Ville did not explain how the results of the analysis would
lead to an effective synthesis.
A similar program (the definition of timt"-frequenry atoms. analysis. and
synthesis) was proposed by Jean-Sylvan Licnard [3j:
Wt consider the speech signal w be compo.~ed of elementary u>at>cform..~. tcf.
( winaowed sinusoids), rodz one defineJi by a small number of parameters. .4
watefoml model (ufm ) i..< a .nusoidal signal multiplsed by a windowing func tion. It i. not to be confused with tllr signal segment. uf . that it i. SUPJIO.~ed tCI
apJ>roximatc. It.~ total duration can b< dcromposeJi intc>attack (be/on: tJu max-
65
Lienard's '"time-frequency atoms" are thus different from those used by Gabor. They are, however, based on analogous principles. We have w(t) =
A(t) cos(wt + cp), where w represents the average frequency of the emitted "note"
and where the envelope A (t ) incorporates the attack and decay. T he principal
difference is that. in the '"atoms'" of Lienard. the duration of the attack and
that of the decay are independent. Thus Lienard's "atoms" depend on four indep endent parameters. and the optimal representation of a speech signal as a
linear combination of time-frequency atoms is more difficult to obtain. Some
empirical methods exist. and they lead to wonderful results for synthesizing the
singing voice. I had the chance to hear t he Queen of the Night"s grand aria
from Mozart's Magtc Flute interpreted b~ '"ttme-frequency atoms." This was
not a copy of the human voice; it involved the creation of a purely numerical
(superhuman) voice. This was commissioned by P ierre Boulez. the director of
the lnstitut. de Recherches Coordonnees Acoustique-Musiqu e.
01
---------~~.w
I
I
I
I
66
CHAPTERS
The time-frequency plane serves the acoustician the way music paper serves the
musician. To continue this metaphor, the signal anal)'lia that we seek to effect
ahould be compared to an exerciae called a "music:&l dictation," which consists
in writiz1 down the notes on hearing a passage of music.
We begin by presenting the point of view of Ville. We will then indicate how
to interpret the results in terms of the theory of pseudodifferential operators as
expressed in Hermann Weyl's formalism. This will bring us back to work done
by the physicist Eugene P. WigDer (1932).
Ville, eea.rc:hing for an "instantaneous spectrum," wanted tQ display the energy of a signal in the time-frequency plane and to obtain an energy density
W(t,{) having (at least) the following properties:
I:
(5.2)
1:
(5.3)
W(t.{)~ = lf(t)fl ,
wet,e)dt = 1icew,
where j({) denotes the Fourier transform of f . These two properties reflect the
program that we presented in the introduction: at each instant t , the function
W(t,{) gives an instantaneous Fourier analysis of the signal f (t ), and (5.2) is
the Plancherel formula. The same remark holds for (5.3):
comes from
li <eW
the contributions of all instants t . and one hopes that li({)l2 is more precise}~
analyzed by using these "individual contributions.'"
Properties (5.2) and (5.3) are clearly not sufficient to define W (t. ~) =
Wr (t. { l. We impose two other condit.ions. namely. "Mo~-ars formula'
(5.4)
which plays the role of ParseYaJ"s identit.y. and the requirement that if
(5.5}
then
(5.6)
n1(t. { ) = 2 c.\':p (
(t-
tof)
112
.,
.,
67
(5.7}
W(t,{)
L: /
Loo
(t+
i) 7 (t- i) e-f{~d-r.
We notice immediately that W(t, ~)is real and continuous in both variables. It
is easy to verify that W{t,{) has the properties indicated in (5.2) to (5.6).
If /(t) = f( - t), then W(O, 0) = 2
IJ(r)l2dr > W(t, {) for all other pairs
Loo
{t,{).
5.4.
We begin by treating the case of signals with finite energy. If W (t ,{) is the
Wigner- Ville transform of /{t), then
W (t, ~ - w) is the transform of e""' /{t),
W (t - to,{) is the transform of f(t- to),
and
W
Ga{)
is the transform of
~I(~). a> 0.
(t -to)
-1eiwr g - -
.Jh
is
2exp (
(t ~!o)2 -
h2(~- w)2).
Here are some other useful observations. The Wigner- Ville transformation
of a function characterizes the function up to multiplication by a constant of
modulus 1. W<' will prove this in the ne>.'t section when we establish the connection between the Wigner- Ville transform and the pseudodifferential calculus.
The Wigner- Ville transform of f (-t) is W (-t, -{) when the transform of f (t )
is W (t,{). Multiplying f (t) by a real or complex constant>. results in the transform W (t . {) being multiplied by 1>.12 Thus we need to consider only the case
where f~oc lf(t )l2dt = 1 when we are working with signals of finite energy.
Not. all function!< H'(t . { ) of thE' two variables t and {atE' the Wigner- Villc
transform of SOUIC' si~nal f (t) .
68
CtiAPTER 5
Q(t, {)
11
(5.8)
to)2)exp(~t).
pe
wherC' thC' terms W 1 (t. 0 and W2 (t. ~ ) art' tht> squart'" tcml." already calcull\tf'd
and wherC' H :l and n4 arc two ~cros.o:- L('rrru;. But thCSC' c-ross temts do not tend
to zero if W2 - w 1 or if t2 -
=4
t1
69
jj WJ. (t -
11.:
l 1 (s)f2(t- s)e-~d{
5.5.
j,-
(l\'('f
R"
r:(:r. y)
assudntt'<l with
CHAPTERS
70
the I)'Dlbol u(z, {) by
(27r)" K(x, ") =
(5.12)
j u ( z; ", {) e'C-IIHd{
= (27rtL(X;JI,x-~) .
This says that the symbol u(z,{) is the partial Fourier transform, in the
ftriable u, of the function L(z, u) and that the kernel K(x, y) that interests us is L(~,z- y). We can also write, in the inverse sense, L(z,JI)
K(x+ ! .z - J), and this allows us to recover the symbol u(z,{) by writing
a(z,{) =
(5.13)
K ( z + ~ . :z:-
~) e-fll(d11.
Thus we are led to hypotheses about the symbols that are the reflections. through
tbe partial Fourier transform, of hypotheses that we may v.-ish to make about the
Rrnels. If we admit all the kernel distributions K(:z:, y) belonging to S'(Rn x R.n),
then there will be no restrictions on a(x, { ) other than the condition that
symbol is
(p.J4)
u (:z:,{)
= 1 (x + ~) 7 (x-
e- ll(dJI.
kurning to dimension one, we have the following result: The Wigner- l'ille
lnru/orm of the function f is the Weyl ~~ of the orthogonal projection oper-
*f c:baracterizes
onto that function f .
/,
to
up
the real axis of a function F (.:) that is holomorphk in tht> upprr half-p lant'
71
1: [
{u(t
L:
W(t,{)d{) .
H, in particular, A(t )
1, the second member is zero, and W(t,{) cannot be
positive or zero unless it is concentrated on the curve { = .p'(t), which represents
the graph of the instantaneous frequency. Since 21rA 2 (t)
f~oo W(t,{)d{, the
"variance" of { is equal to -!(cf2 /dt2 )(log A(t)).
Here are two examples of the calculation of the instantaneous frequency.
Suppose first that the original signal / (t) is real, equal to 1 on the inter\'al
[-T, T], and 0 outsidt> this interval. The corresponding anal~<tic signal is then
F (t )
The phase c;(t ) of F { t) is continuous on the whole real line. odd. equal to
T (and thus -~ if t ~ - T ). and strictly increasing on [-T.T]. Tht>
instantaneous frequen~ is 0 outsid<> tht> interval [-T, T ]. suictly positive on
( -T, T ), equl\1 to "~ at 0. and increase; from ~ to +OC' &: t traverses tlw
interval [0. T ).
~ if t ;:::
CHAPTEk 5
72
AB one could have guessed, the in.ttantaneou.~ Fourier anol11. iJ propo$ed b11
Ville i3 not even a local propert11. Thu ~ Uult knowing the 6ign4l in an arbitro.r y klrge intenlaZ ~ at to u not .u[ficient to calculate the irutantaneo1U
frequency at t 0 . The operation responaible for this anomaly is the calculation of
the analytic BigDal F(t) associated with / (t); as everyone knows, the kernel of
the Hilbert transform decreases slowly at infinity.
This discussion shows that the signals to which the Ville theory applies are
necessarily academic signals (whose algorithmic structure does not change over
time) or asymptotic signals whose behavior on a short time interval is equivalent
to that. of a normal signal over a much longer duration.
The second example of a calculation of the instantaneous frequency is for
the signal/(t) = cost2 which is a chirp of infinite duration. The calculation of
the corresponding analytic signal is interesting because it exhibits two different
asymptotic behaviors depending on whether t tends to + :x. or - x . Indeed.
the Fourier transform of this analytic signal F (t) is 0 for { < 0 and is equal to
~ ( .;;i e-~' 1 4 + H1 ei!'/ 4 ) for { > 0. It follows that F (t ) is asymptotically
equal to e112 when t tends to +oc and to e- 112 when t tends to - oo. The
inStantaneous frequency of cost 2 is thus equal to 2t + E:(t) when t tends to + oc
and to -2t + E:{t) when t tends to - oc. ln both cases, E:(t) tends to 0 when lt i
tends to +oo.
5.7.
=('
j exp[i(ar
= 2r. (a:)
~ ) :: (3af~- ,..:))
1 1
113
A ((
73
wher!!
A(w)
= _!_
27r
f""
ei((" / 3J+w)ds
-oo
011
is the Airy function . Here again the Wigner-Ville transform of the function e'
is "essentially" concentrated around the curve ~
3ot2 , which is the graph of
the instantaneous frequencies. All of this must be put in quotation marks since
etats is not an analytic signal. But the signal is asymptotically analytic because
the Airy function decreases exponentially when w tends to +ex:.
5.8.
Bibliography
!I: R
12
l3i
!{
CHAPTER
6.1.
Introduction.
This chapter is the logical sequel of the preceding one. We will introduce algorithms that allow us to decompose a given signal .t(t) into a linear combination
of time-frequency atoms. The time-frequency atoms that we use are denoted by
/R(t) and are coded by the Heisenberg rectangles R (with sides parallel to the
axes and with area 1 or 27f depending on the normalization). If R = [a, b] x [o, 8],
we require that the function !R be essentially supported on the interval[a, b) and
that its Fourier transform
be essentially supported on [a,p]. We also ask that
the algorithmic structure of /R(t) be simple and explicit to facilitate numerical
processing in real time. The decomposition
iR
00
(6.1}
s(t)
= 2:o;/R1(t )
0
cannot be unique. and we take advantage of this flexibility by looking for optimal
decompositions, which here means they contain the fewest possible terms.
The point of view of Ville (and of numerous other signal-processing eicperts)
is that it is first necessary to understand the physics of the process and that '"the
algorithms will follow." Unfortunately, algorithms associated with the ViignerViUe transform have never followed, they have never existed, and the WignerVille transform is an analytic technique that does not lead to a synthesis or to
the transmission of information.
Our approach here is to favor synthesis and transmission over analysis. Thr
''time-frequency atoms that v.e use are completely explicit: They art' either
Malvar wavelets or wavelet packets. and we will immediately WTite dov.n tbt'
'atomic- decompositions~ of thE' type (6.1). This means that the synthesis will be
direct. whereas the analysis will consist in choosing- with the use of an entropy
criterion- the most effective synthesis which is the one that. leads to optimal
compression. Thus th> analysis proceeds according to algorithmic criteria and
not according to physics, and it is not at all clear that this approach lead~ t.o
a signal anal~'Sis that rt'veal!' physical properties having a real meaning. For
exaropk. MariC' Farg< had thC' idea to apply the algorithm to image!' of simulated t-W<~dimcusional turbulenc-e. Th> algorithm extractl"Cl. in order. coh<'r<'llt
structur(-:;. vorticit.y fihunrnl$. nnd w on down the seal<'. Thi..; i:> ama.ziu~: ht'-
CHAPTER 6
76
cause the algorithm is not based on an analysis that takes into consideration the
underlying physics of fully developed turbulence.
After these general remarks, it is time to specify the algorithms. There are
two options: Malvar wavelets and wavelet packets. With the first option, the
signal is segmented adaptively and optimally, and then the segments are analyzed
using classical Fourier analysis. The second option, "wavelet packets," reverses
the order of these operations and first filters the sigDal adaptively; the analysis
in the time variable is then imposed by the algorithm.
Ville (1947) proposed two types of analysis. He wrote: "We can either: first
cut the signal into slices (in time) with a switch; then pass these different slices
through a system of filters to anal~rzc them. Or we can: first filter different
frequency bands; then cut these bands intp slices (in time) to study their energy
variations." The first approach leads us to "Malvar wavelets" and the second to
"wavelet packets."
6.2.
The discovery of Malvar wavelets (Henrique Malvar, 1987) falls within the general framework of windowed Fourier analysis. The windov. is denoted b~ w(t),
and it allows the signal , (t) to be cut into "slices" that are regularly spaced in
time w(t-bl)s(t ), l = 0, 1, 2, ... The parameter b > 0 is the average length of
these "slices." Next, following Ville, one does a Fourier analysis on these slices.
which reduces to calculating the coefficients J e- i4klw(t - bl), (t)dt, where a > 0
must be related to band where k 0, 1, 2, ... This is thus the same as taking
the scalar products of the signal s(t) with the "wavelets"
Wi..t(t)
= eioklw(t- bl).
This analysis technique was proposed by Gabor (1946). in which case the w(t}
was the Gaussian. The "Gabor v.-avelets" lead to serious algorithmic difficulties
and. more generally, Low and Balian showed in the early 1980s that if u(t)
is sufficiently regular and well local..U:ed, then the functions w,.,1, k.l E Z. can
never be an orthonormal basis for L 2 (1R). More precisely. if the two integrals
J::000 (1 + 1tl) 2 1w(t)i2dt and f~oc ( l + 1{ 1)21w(~)j 2 d{ are both finite, the functions
Wk.l k.l E Z., cannot be an orthonormal basis of L 2 (R).
The crud(' v.indov. defined by u(t ) = 1 on the inter"al !O. 2r.) and u(t) = 0
elsev.here escapes this criterion. By choosing a = 1 and h = 21r, tht> windowE'd
analysis consists in restraining tht' signal t.o each interval !21;.. 2(/+l }7r) and using
Fourier series to analyze each of t.ht> corresponding functions. But tht' functi ons
obtained by this crude segmentation arC' not 21r-periodic. and the Fourier anal~sis
v.ill highlight this lack of periodicity and interpret it as a discontinuit,y or au
abrupt \'&fiation in the signal.
Ont> way to attenuatt' these numerical artifac~-ithout. howt'ver. elimi
nating them- is to use th(' Discret<' Cosine Transform (DC'T). We describC' tbC'
continuous version of this transform.
On each intervall21r., 2(1+ 1):-r). v.-e analyze the signal s(t) using the orthonor
mal basis composed ofthC' functions +. and .J..cos( ~ t) . k = 1.2.3.. .. If s(t)
\
.671'
"~
(6.2)
uu(t ) :;:::
(6.3)
e 2Z,
= 1,2, . ..
1 e 2Z,
k = O,
lE2Z+l,
and
(6.4)
~t) ,
= 1,2, ...
(6.5)
(6.6)
(6. 7)
~
2
=0
if t
~ -r.
w(t ) ~ 1 and
2
u (t) + u; ( -t)
=1
or t
3r.
if
1f
~ t ::; r..
Then the construction is the same, and the sequence""' defined by (6.2). (6.3).
and (6.4) is an orthonormal basis for L 2 (R). In Malvar's construction. thE> v.indow w(t) can be very regular (infinitely cti1Jerentiable, for example). but thE'
Fourier transform of u cannot havt- exponential decay. Condition (6.5) prevents
it. and this condition plays an essential role in the demonstrations.
Althou~h similar. the solutions found by Dnubechies. Jaffard. a11 d Jouruc. on
thP om hand. a.nd l>y l\1alw.r. on the other. arl' not linked b~ a logical connl'Ction.
6.3.
C'..oifman and tlu author tackled the problem of modifying tht' preceding constructions to obtain v.iudows with variable lengths that could bt> defined arbitrarily. The construction b~ Daubechies. Jaffard. and Journe d~ not extend to
this context. while t hat of Malvar gcueralizes. v.;thout tb<' sli~htest difficult.~. to
th( 01."<' of arbitrary windows.
78
CHAPTER6
We begin with an arbitrary partition of the real line into adjacent intervals
lcJo4j+J), where .. < - 1 < ao < 41 < 42 < .. . limJ-+oo a;
+oo, and
limJ--oo a; -oo. Write l; 4j+l - CJ and let OJ > 0 be positive numbers
that are ama1l enough 10 that l; ~ o1 + OJ+l for all j E
The windotva w;(t) that we use will eeeentially be the characteristic functions
of the intervals [a1,a;+l); the role played by the disjoint intervals (aJ-o;, CJ+o;)
is to allow the windows to overlap, which is necessary if we want the windows to
z.
be regular.
More precisely, we impose the following conditions:
(6.8,)
(6.9)
w;(t) = 1 if
(6.11)
tt!J (a;
(6.12)
+ a 1 S t S a;+l- OJ+t
ai
+ r ) + tt!J(a1 -
u:;-1(a; +
=1
-;-)
r ) = U:j(a1
if lr l S
r ) If lr l $
OJ
Or
Wj l t J
: \
I
:
I
aj
I
I
'
:
1:
_,'
'
2~j
'......: ----+-----'------;:o....,.;--.
:
---..--~I
-----T----
ai,. 1
1
~ ~----~'---~~
.. 1
...
2~/+1
I
~-
form:
(6.14)
r;:;
F
=
f .:.u1 lt ) ~ _::_ lf _,_
\' 1J
I
u"'-(t )
a1 )
if j E 2Z and k
= 1, 2, ...
up1 (t)
(6.15)
79
= {fw;(t) if j E 2Z and k = 0,
and
(6.16)
if j E 2Z + 1 and k
1, 2, . ..
The functions u;.-.(t), j E Z, k = 0, 1, 2, ... given by (6.13) are an orthonormal basis for L 2 (R), and the same is true for the functions defuied by (6.14),
(6.15), and (6.16).
Here are two graphs of Malvar wavelets:
4-------- - ---------
--1
.....,..
1
Not<' the ex"treme similarit) between the Malvar wavelets and thE' timt>frequenC'y atoms proposed by Lienard. In particular. the .Malvar wavelet:; 8T('
construc-ted v.itb an attack (whOSE' duration is 2o;). a stationary period (v.hich
lasts l; - o ; - o; +d and then a decay (which lasts 2a;+ll ThE' ability to
ch~. arbitrarily and independently. the duration of the attack, then that of
the stat ioutu') sC<'tion. and fumlly the duration of th<' relaxation is preciS<'ly what
CHAPTEH.6
80
6.4.
[-.-
w t 2x l
W !XI
, . . ---T"------.
. . _............
.:
t
8 '3
We can just as easily replace' all of the cosine;: in (6.13) by sines and thu.c:
obtain a second orthonormal basis for L 2 [0. ocj of the form
(6.18)
v;.k(x ) =
We nl').1 El).'tend w(x ) to the whole real linE' by making it an even function:
ul-x\ = UI{X). This give;: a natural even ext.ension for the functioul' uJ. k and
, .j.k
(!)
=- 1-
k, we have
j,kE Z,
1
- -21 ' 2 w(21 t - kr.),
...&.
j,k E Z ,
is an orthonormal basis for L 2 (R). To regain the usual form, 2i1 2 ~(2i t - k),
j , k E Z. it suffices to replace t by r.t.
It is clearl~ possible to require that w(x } be an infinitely differentiable func
tion, in v.hich case 1/J(t) will be a function in the Schwartz class S(R).
We recall the program of Ville. There were two possible approaches: either
segment the signal appropriately and follov. this by Fourier analysis, or pass the
signal through a bank of filters and then study the individual outputs of the
filter banks.
Here wr havr selected the second approach. The filter bank that wr US(' is
defined b~- the transfer functions u(:!-ix ). where w(x } is thE' even window used
above.
6.5.
82
CHAPTER6
(c;,a;+t] are combined into a single interval, namely, (a;- 11 a; +I] The other
intervals remain unchanged. This operation is called "merging." The inverse
operation consists in adding an extra point ex between the points a; and a;+ 1 ,
which results in replacing the interval !a;,a;+ 1) by the two intervals !a; , a ] and
!a,c;+tl This inverse operation is called "splitting."
A split-and-merge algorithm provides a criterion to decide when and where
to use one or the other of these elementary operations.
We are going to examine the effect of these operations on a Malvar basis.
We will show that an elementary operation induces an elementary modification
of the basis that is very easy to calculate. The following remark is the point of
departure for this discussion.
For each fixed j , let W; denote the closed subspace of L2 (R ) generated by
the functions u;,k(t), k = 0, 1,2, ... described by (6.13}. Then f (t ) belongs to
W; if and only if /(t) = w;(t)q(t), where q(t ) belongs to L 2 (a;- ex;, a;+ 1 + ex;+tl
and satisfies the following two conditions: q(a; + T) = q(a;- T) if ITI ~ex; and
q(c;+J +T) -q(a;+ 1 - T) if ITI ~ a;+l There are no conditions that need to
be satisfied on the interval (a;+ ex; ,a;+t - a;+t)
To use the suggestive language of Ronald Coifman and Guido Weiss, W; is
a "di~" with a positive polarity at 4j and a negative polarity at a;+J The
intervals (a; - a;, a; +a;) where the polarities interact are pairwise disjoint.
From here, the merging algorithm is trivial. Removing the point c; of the
partition amounts to replacing the two subspaces W;-t and W; by their direct
orthogonal sum W;-1 E9 W; without disturbing any of the other spaces Wp,
j' :/: j -1 and j' :/: j . But this, in turn, romes down to replacing the two windows
Wj-1 and Wj by the new window Wj defined by te; (t)
(wj_ 1 (t) + u"J(t)) 112 .
The two lengths l;-J and l; are replaced by l; = l;- 1 + l;, which changes the
fundamental frequency in (6.13).
To fix our ideas. we start with a segmentation with intervals of length 1,
c; = j , and choose w;(t) = w(t - j ) with a ; = 1/ 3. Next, we look at which
wind~'S can appear as a result of the merging algorithm, which consist s in
removing intermediate points in the partition. The resulting windows and the
corresponding wavelets will look like centipedes (see the figure of the second
Mah-ar wavelet). The localization of these centipedes in the time-frequency plane
is not optimal. This is because, in using the merging algorithm, we never change
the values of the numbers a ,. In our example. we always keep a , = 1.'3. T his
being the case. the algorithm allows us to replace only the partition consisting
of all the integer interYals of length 1 by a partition that has arbitrary integer
intervals [a;. a1 +1! The v."avelets that appear are given by (6.13). and thE'y are
approximate!~ localized in thE' time-frequenc~ plane around the rectangles
R(j, k ) = !a;. a; +l x
[kr.
1; (i:+l, lh].
Wt> must no"' provide thE' criterion that all09os us to decide when to US(' the
dynamic split-and-merge algorithm. This means that we need to establish a
numerical value to measure v.hat is gained or lost h~ adding or deleting n point
in the subdivision. This is the purp()S(' of thE' next section.
6.6.
83
Let H denote a Hilbert space and let (e;);EJ be an orthonormal basis for H . Let
x be a vector of H of length 1. We write x E ;EJ OJe;; the entropy of x relative
to the basis e; measures the number of significant terms in this decomposition.
This entropy is defined by exp(- EJeJ loJI2 log loJI 2 ).
If we have a collection (ej),eJ of orthonormal bases v.here w ranges over a
set n, we will choose for the analysis of % the particular basis (indexed by Wo)
that yields the minimum entropy.
This point of view poses three problems:
{6.24)
(6.25)
(6.26)
In image analysis, all of the algorithms are based on calculations that are
done from an energy function, which is defined as the quadratic mean value of
the gray levels; the algorithm used to search for an optimal basis for compression
does not escape this difficulty. The search for a Hilbert norm that is adapted
exactly to the structure of the image is still an unsolved problem.
A concrete example of where these difficulties arise is a program to transmit
radiographic images within large hospitals over the ordinary telephone lines. The
compression needed for transmission and the q~ty of the received image that
is required for the doctor to make a diagnosis are clearly antagonistic objectives.
The interested reader can refer to (3), where this problem is discussed.
6.7.
We v.ill examin<' in detail thE' particular case where the Hilbert space H is tht>
spact' of signals f (t ) with finite energy, which is defined by f~oe lf(tWdt. The
quality of the compression will be measured only by this criterion.
ThE' algorithm looks for '"the best basis. which is tht> one that optimizes
compression based on the reduction of transmitted data. The search is donE' by
comparing the scores of a whole family of orthonormal bases of L 2 (R). These
are Malvar bases. and thE'~ are obtained from segmentations of the real lin<' into
dyadic intervals. These intervals are systematically constructed in a scheme that
moves from "fine" to "coarse.~ This means that. we start. with a segmentation by
inten~s of length 2-q. wheT(' q ~ 0 is largt> enough to captur<' th<' fine."t. dt't.aib:
appcarin~ in thE' sip1al.
84
CHAPTER6
The process consists in removing, if necessary, certain points in the segmentation and in replacing, at the same time. two contiguous dyadic intervals I ' and
I" (appearing in the former segmentation) with the dyadic interval I = I' U I" .
For example, (2, 3) and (3, 4) can become (2, 4] with the disappearance of 3, but
[3,4) and (4,5] will never become [3,5). For the point 4 to disappear, it would
be necessary to wait for the eventual merging of the intervals (0, 4) and (4, 8).
By an obvious change of scale, we may assume that q = 0. The starting point
is thus the segmentation where the "fine grid" is Z. The intervals [aj, a j +ll of
6.5 are now [j, j + 1], and the first orthonormal basis to participatR in the
competition will be
(6.27)
u;.k(t )
= v'2cos [
1r (
+ ~) (t -
j)] w(t-
j ),
where j E Z, k = 1, 2, 3, ...
The other orthonormal bases that participate in the competition will all be
obtained from this first one by merging. The regrouping algorithm that merges
two orthonormal bases into one was described in detail in 6.5. We will limit its
application to situations where (aj_ 1 .aj ] and [aJoaj+ll are the left (I' ) and (/" )
right halves of a dyadic interval I .
Each partition of the real line into dyadic intervals of length greater than or
equal to 1 thus defines canonically one of the orthonormal bases that are allowed
to participate in the competition.
One reaches all of these partitions by iterating those elementary operations
that combine the left and right halves of a dyadic interval and by traversing this
tree structure, starting from the "fine grid."
We denote by I the collection of all the dyadic intervals I of length Il l ~ 1.
and if I = (aj , ai +I J is one of these dyadic intervals, we denote by WI the window
that was denoted by W j in 6.5. In the same way, we denote by W 1 the closed
subspace of L 2 (R) that was denoted by Wi and by w~k) the orthonormal sequence
defined by (6.13). which is now an orthonormal basis for li'J.
. If I' and I" are, respectively, the left and right halves of the dyadic interval
I , t hen we have WI= WI' e WI" and this direct sum is orthogonal.
The signal f (t ) that we wish to analyu optimally is normalized by
f~oc lf(t )l2dt = 1. To simplify the follov..;ng discussion. we assume in addition that f (t ) is zero outside the interval [1, T ] for some sufficiently large T .
Then f belongs to "H"L if L == [0. 21] and I is large enough.
It is eas~ to shov that if m tends to -oc. the entropy off in t he orthonormal
basis
l of li'1. I
10. 2m] , also tends to infinity. Thus there exists some \-alu'
of m after v.hich u~"l no longer enter!' into thr competition. In other words. the
dyadic partitions that come into play will. in fact , be the partitions of L 10. 2"' ]
(for sufficiently large m ) into dyadiC" intervals I of length I/ ~ ~ 1. The number
of partitions is thus finite, but it can be incredibl~ large (the order of magnitude
being 2<2"' l). It remains to find a fast algorithm to search for the "best basis.''
This is the algorithm that wr arr now going to describe. If I belongs to I .
then c1(k ) = / (1)u1(k ) (t.)dt.
uY
e(l) =-
(6.28)
L"" lc~kll
log lc~"')l 2 ,
and
(6.29)
where the lower bound is taken over all the partitions (Jp) of the interval/ iDto
dyadic intervals Jp belonging to I . If I = IJ,j + 1), then clearly e"(I) e(I).
The problem that we must solve is thus clearly reduced to finding the optimal
partition (Jp) when I L [0, 2m]. the largest of the dyadic~ involved ill
the competition. The calculation of e"(L) and the determination of the optimal
partition cannot be done directly because the number of cases to be considered
is too large. We will calculate e"(/} for Il l = 2" by induction on n. For n = 0.
we must calculate t:" (J) = e(I) for all intervals I= IJ,j + 1) in (O,r). Next we
proceed by induction on n, assuming that we have calculated e"(/) for Il l= 7!
and that we have determined the corresponding covering (J,.).
Suppose that I/ I = 2"+ 1 and let I ' and J" be the left and right halves of I.
There are t wo cases:
= =
(6.30)
(6.31}
If t:(/)
86
CHAPTER6
6.9.
We replace the real line R by the grid hZ, where h > 0 is the sampling step.
Thus the signal/ is given by a sampling denoted f(hk), k = 0, 1, 2, . . .. but
we will not discuss here the technique used to arrive at this sampling. We will
systematic:al.l forget h in all that follows.
A partition of Z will be defined by the intervals la;,a;+l], where a;-~ is an
')Dteger, in such a way that a; does not belong to Z. This point of view has also
~ been adapted in the context of the OCT.
We denote a number of points belonging to !a;,a;+l] n Z by l; =a;+ I -a; ,
~we let the numbers o; > 0 be small enough so that o; + oi+1 $ l;.
The windows w;(t) will be subject to exactly the same conditions as in the
cxmtinuous case. This means that
(6.32)
l
(6.33)
{6.34)
(6.35)
w;(t)
=0
w1 (t)
=1
on the interval
+ a;+1 ]
=1
if lr l $ a;.
0 S k Sl; -1,
J ez.
87
Bibliography
(1) R . COJPWAH ANDY . MEY!a, ~.,. l'cmoiJIH. Fowler 4/eMtrc. c. R. Aced.
Sci. Paria S6r. I (1991), pp. 29-201.
(2) 1. DA\IBECKJa, The ~ CroN/ann, time~ '-'UoCicm en4 ~ Oft4lsiN,
IEEE 'Iram. Inform. Theory, 36 (1990), pp. 961-1005.
(3)
(4)
(5]
(6)
(7)
(8)
S.
H.
S . MALVAR,
CHAPTER
7 .1.
Heuristic considerations.
A time-frequency analysis of a signal lets us express the signal as a linear rombination of time-frequency atoms. These time-frequency atoms are essentially
characterized by
(a) an arbitrary duration, t 2 - t 1 ; t 1 is the instant when the signal is first
heard (if it is a speech signal, for example}, and~ is the instant when it ceases
to be heard, and
(b) an arbitrary frequency, which is the average frequency; this average fre.
quency is the frequency of the emitted note in the case of a musical signal, while
the frequency spectrum given by Fourier analysis takes into consideration the
parasitic frequencies created by the note's attack and decay.
We also think of a time-frequency atom as occupying a symbolic region in
the time-frequency plane. This symbolic region is a rectangle with area 211" (the
Heisenberg uncertainty principle).
2F.
I
I
I
T I
wo-- -- 1
I
'I
_l_ __
I :
---
CHAPTER 7
90
(t~o~) and that the Fourier transform in({ ) of /n(t) is eaeentially supported by
the interval (wo- 1rjh, Wo + 1r/h). It is well known that there does not exist
a function with compact support whose Fourier transform also has compact
support. This leads ODe to consider the following, leas stringent conditions
(7.1)
(7.2)
L:
/ _ : ({ - Wo)
1iR(t)l 2 d{ $
21rCJ2h- 2 .
The time-frequency atom that optimizes this criterion (that is, for which
the constant C is the smallest possible) i.e; precisely the Gabor wavell't, and th,.
Gabor wavelet owes it success w this optimal localization in the time-frequency
plane.
We will see, on the other band, that the Gabor wavelets have a very disagreeable property, which makes them unsuitable for the time-frequency signal
analysis.
(7.3)
I:
(7.4)
' "" lo 1z -- 1~
-- . J i
(l
and
(7.5)
91
Thus remoteness of the rectangles in the time-frequency plane does not even
imply that the corresponding Gabor wavelets are almost orthogonal, and consequently the apparent heuristic simplicity of the time-frequency plane is completely misleading.
This catastrophic phenomenon results bom the arbitrariness of the h > 0
that are used in the definition of the time-frequency atoms. The rectangles
~. R 1 , . in 5ere's result have arbitrarily large eccentricity. When h = 1 all is
well, and the corresponding situation has been studied extensively. This is then
a form of windowed Fourier analysis where the sliding window is a Gaussian j4J.
Once we abandon the Gabor wavelets we have two options: Malvar wavelets
and wavelet packets. We will briefly indicate the advantages and disadvantages
of these two options.
If we use Malvar wavelets, then by their nature, the duration of the attack
or of the decay is not related to the duration of the stationary part. We can,
for example, have a Malvar wavelet for which the durations of the attack and
of the decay are of order 1, while the stationary part lasts T > > 1. If Wo is
the frequency corresponding to this stationary part, then the Fourier transform
of the Malvar wavelet v.ill be, at best, of the form jfcW:::::l~({- Wo), and it
cannot satisfy (7.2) since his of the order of magnitude ofT. On the other band,
the Malvar wavelets are constructed to be exactly orthogonal.
The implication of our observations is that the orthogonality of the Malvar
wavelets has been won at the prioe of tlieir localization in the time-frequency
plane, a localization that no longer guarantees the "minimal conditions" (7.1)
and (7.2).
One last remark about the Malvar wavelets is obvious but significant: Although they are given by a very simple formula, they are not obtained by trans-lation, change of scale, and modulation (or modulation by e*'A) of a fixed function g(t).
The option we propose in the following pages is that of wavelet .packets. The
advantages of this option are the following:
(a) Daubechies's orthogonal wavelets (Chapter 3) are a particular case of
wavelet packets.
(b) Wavelet packets are organized naturally into collections, and each collection is an orthonormal basis for L 2 (R).
(c) Thus one can compare the advantages and disadvantages of the various
possible deeompositions of a given signal in these orthonormal bases and select
th<' optimal collection of wavelet packets for representing the given signal.
(d) Wavelt't packets arE' described by a vel) simple algorithm 2i12wn(2i:r-k),
where j, k E z. n = 0. 1. 2.... , and v.here the supports of the Wn (:r) arE' in the
same fixed interval IO.L].
The integer n plays thE' role of a frequenc~. and it can be compared to the
integer k = 0. 1. 2 .... that occurs in thE' definition of the Malvar wavelets.
The pricE' to pa~ for these advantages is the same one that. comes up v.hen we
use Malvar v.-avelets. Indeed. if, to facilitate intuition, we associat<> the rectangle
R in thE' timt"-fr<'Quency plane defined by k2-' ~ t < (k + 1 )2-J aud n2i ~ { <
92
CHAPTER
(n+ 1)2i with the wavelet packet 2'12wn(2i t- k), then this choice does not meet
tbe conditions (7.1) and (7.2). Furthermore, we C&D110t do better by assigning a
frequency different from n to Wn(z), for although !1Wn!l2 1,
lim { inf
{7.6)
n-+oo
wER } _ 00
1.2.
We begin by defining a special sequence of functions tcn(z ), n = 0 , 1, 2, .... supported by the interval (0, 2N- 1}, where N !::, 1 is fixed at the outset. If A'= 1.
these functions Wn(z) constitute the Walsh system, which, we recall, is wellknown orthonormal basis for 2 [0, 1). If N !::, 2, the functions Wn(x ) are no
longer supported by [0,1]; however, the double sequence
n = 0, 1, 2, ... , k E Z
Wn (Z- k).
(7.7)
will be an orthonormal basis for L2 (R). This orthonormal basis will allow us
to do an orthogonal windowed Fourier analysis. Thus, for the moment, this
construction is similar to the Malvar wavelets. The difference occurs when the
dilations enter (the changes of variable of the form x- 2i x ).
We start with an integer N !::, 1 and consider two finite trigonometric sums
.,
mo()
l
rn
v2
(7.8)
2N-1
h,.e- "(
(7.9)
91<
(7.10)
= (- 1)"+1Ji2N -J-Ic
mo(O) = 1.
and. finally.
(i .ll)
or
m1 ({)
93
where
= i<e-iC -
1). ConditioD
v'2 h1
= ~(3 + v'3),
v'2h3 =
~(1- v'3),
= 0, 1, 2, . .. ,
2N - 1
(7.13)
W2n(x)
= J2
h~twn(2.x- k)
(7.14)
W2n+1(z )
= J2
21''- 1
9kWn(2x- k)
L
0
= 0 , is
2N - 1
(7.15)
wo(z) = J2
h~cwo(2.x- k).
(7.16)
Tf (x )
= J2 L
hkf(2x - k) ,
which becomes
(7.17)
CHAPTER 7
94
0/
h3
/ 0-----~----~o~
/
____r:::l_____
0
1/2
2_..
1/2
1 3/2
I
I
When f is normalized by
is given~
(i .lS)
~{~)
This diagram illustrates tbt> iterativt> schemt> for constructing ..; usin~ thr
characteristic function of [0.1] for thr initial ''alue /(). Here. fi+ 1 = T(/1 ) . and
~ have drawn the first fev; functious /o, It. and h The coefficients II{). h1 h2 .
and h3 are approximately ~
(l -1. J3).... , ~(
1- ,/3). as in thr second exh..
.., ample mentioned above. Then the st>queuet> f 1 converg~ uniformly to t hr fi.xt'<i
point ;,.."(r ).
95
(7.19)
Wn(X-
k),
= 0,1,2, ... ,
k E Z,
L 2 (.R) .
verify that
and that
W2n+l(.x}
= Wn(2x )- Wn(2x -
1).
This shows that, in the case of filters of length 1, the construction of wavelet
packets leads to the Walsh ~ystem.
The Walsh system Wn (x), n E N. is an orthonormal basis for L 2 (0, 1), and
it follows immediately that the double sequence Wn(x- k ). n E N. k E Z. is an
orthononnal basis for L~ (R).
In the general case of basic wavelet packets (filters longer than 1). the su~
ports of u:n(x- k) and U:n(x - k') are not necessarily disjoint ..-hen k :f; k1 so
that the orthogonality of the double sequence wn(x- k ).11 E N. k E Z , is mon>
subtle.
7.3.
CHAPTER 7
96
orthonormal baai8 for L 2 (R). This orthonormal basis is analogous to the Walsh
system but, for filters longer than l, it is more regular. That is, the frequency
localization of the functions Wn (:r) iB better than the frequency localization of
the functions in the Walsh system. Nevertheless, this frequency localization does
not yield an estimate of the type
(7.20)
uniformly in n (see [2]).
The general wavelet packets are the functions
neN,
(7.21)
j,k e
z.
These are much too numerous to form an orthonormal basis. In fact, we can
extract several different orthonormal bases from the collection (7.21). The choice
j
0, n e N, j , k e Z , leads to the orthonormal basiS described in the previous
section, while the choice n 1, j , k e Z. leads to the orthonormal wavelet basis
(Chapter 3).
We associate with each of the wavelet packets (7.21) the "frequency interval"
I,(j,n) defined by 2;n S < 2i(n+ 1). The following result describes certain
sets of wavelet packets that make up orthonormal bases for L 2 (R).
THEoREM 7.1. Let E be a eet ofpc.in (j,n), j e Z, n eN, 6uch that the
corTe8pOflding frequency intenJOO l(j, n) constitute a partition of [0, oo), up to a
countable eet. TMn the 6Ub&equence
(7.22)
(j.n) e E ,
Ire z.
L2 (R).
7 .4.
97
Splitting alcorlthma.
mo(8)
m1(1J) )
U(IJ) _ (
mo(8+1r) m1(8+1r)
15
.tary
UDJ
(7.23)
ht = h 'L:o21t- le1,
(e~o)~o ez.
and define
00
/21<+1
= v"2 EP2~o-1e1.
-co
-oo
Then the sequence (!,.), indexed by k E Z , is also an orthonormal basis for the
Hilbert space H.
Next, let Ho be the closed subspace of H generated by the vectors}21o, which
0
1
we denote by e~ >; similarly H1 will be generated by !21<+1 e~ >, k E Z.
0
Nothing prevents us from repeating on {H0 , e~ >) the operation we have done
on (H, e,.) and from iterating these decompositions, while keeping the same coefficients (o,) and (13~c) at each step.
An elementary example helps us to understand the nature of this splitting
algorithm. The initial Hilbert space is 2 [0, 21r) with tbe usual orthonormal basis
e, ~e"'', k E Z. The (21r-periodic) functions mo(IJ) and m1 (8) are (when
restricted to [0, 21r)) the characteristic functions of [0, 1r) and [1r, 211"). Then the
vectors f2,, k E Z, are -j; e12..,mo(IJ} and constitute a Fourier basis for the
interval [0, 1r), while the vectors ht+ 1, k E Z, constitute a Fourier basis for
the interval (1r, 21r). Finally, the subspace H 0 of His composed of the functions
supported on the interval (0, 1r}, while H1 is composed of the functions supported
on (1r,21r).
Iterating the splitting algorithm leads to subspaces that are naturally denoted
by H 1., ,... J) t: 1 0 or 1.... , ~i = 0 or 1, or even by Hr , where I denotes the
dyadic interval of length 2-; and origin t:J/2+ +~if2i . In fact. in the example
we have just studied. Hr is exactly the subspace of 2 (0, 211") consisting of the
functions that vanish outside the interval 211"1.
This example has guided the intuition of scientists working in signal processing. Assuming that the signal is sampled on Z, they have considered the
situation where (o,) and (B~~:) are two finite sequences and where mo(9) resembles the transfer function of a low-pass filter while m 1 (8) resembles that of a
high-pass filter. One requires. at least, that mo(O) 1 and that mo(8) does not
vanish in the interval 1- i .
By analogy with the preceding example, these scientists were led to believe
that the iterative scheme, v.hich we have called the splitting algorithm. v.-ould
providr a finer and finer frequency definition, as onr wanders through thr maz<'
of "channels" iu th<' follol':iu~ figure.
CHAPTER 7
98
H(O,O,O,O)
H(O,O)
Ho
H(O,l , O)
H(O,l)
HIO, 1.1 }
H!
H(l , O, O)
H(l. O)
H ( l,O, l)
H,
H (l,l, 0)
H( l.l )
LH(l.l.l)-i
H ( l , l,l , l )
The initial Hilbert space H is the direct sum of various combinations of these
subspaces. In particalar, H is the direct sum of all the subspa.ces at the same
"splitting level": at the first level there are 2 subspaces, at the next level there
are 4, then 8, 16, and so on.
To give a better understanding of the construction of wavelet packets and
the exact nature of the splitting algorithm, consider the case where the initial
Hilbert space is the space V; , j !:: 1, (in the language of multiresolution analysis)
with the orthonormal basis 2il2rp(2ix- k), k e Z. Next, suppose that the
splitting algorithm has operated j times. Then we arrive exactly at the sequence
of functions wn(:t - A:), k E Z, 0 :S n < 2i. More precisely, the integer n =
Eo+ 2E1 + + 2;- 1; - 1 is the index of the "frequency channel" H cc0 .c1 , ... . c,_ 1 )
The frequency localization of wavelet packets does not conform to the intuition of the scientists who introduced these algorithms, and the only case where
there is a precise relation between the integer n and a frequency in the sense of
Fourier analysis is the case where mo(B) and m 1 (8) are the transfer functions of
"ideal filters."
7.5.
Conclusions.
99
motion the algorithm for constructing the wavelet packets. From this processs
we obtain a huge collection of orthonormal buel for L2 (R).
It is then a question of determining, for a given signal, the optimal basis.
And again, the optimal ba8ia is the one (among all thoee in the wavelet paclcet)
that gives the most compact decompositioo of the signal
We determine this optimal basis by using a ~tcH:oanle" type strategy and
the method of merging. Thus we start from the finest frequency channels H1,
which are associated with the dyadic intervals I of length III 2-m. The integer
m is taken to be as large as possible, consistent with the ch06en precision. The
algorithm proceeds by making the following decision: It combines the left and
right halves, I' and I", of a dyadic interval I whenever the orthonormal basis of
Hr yields a more compact. representation than that obtained by using the two
orthonormal bases of Hr and Hr"
The discrete version of wavelet packets can also be used and is immediately
available. This is obtained by starting with the Hilbert space H
12 (Z) of
signal sampled on Z and the canonical orthonormal basis (c 11 ) 11ez : 1t is 1 at k
and 0 elsewhere. Here there is }>erlect resolution in position but no resolution
in the frequency variable. Next, we systematically apply the splitting algorithm
to improve the frequency definition until reaching the spaces H1 associated with
the dyadic intervals I of length Ill 2-m. Finally, we apply the algorithm to
ch006e the best basis (6:.7 ).
Wavelet packets offer a technique that. is dual to the one given by the Malvar
wavelets. In the case of wavelet packets, we effect an adaptive filtering, whereas
the Malvar wavelets are associated with an adaptive segmentation.
Bibliography
(1) R . COlf'MAK, Adopted multiruol11tion onoU, computotion, rign4l ,..ooeutng tmd ~
ator fhemv, ICM 90, Kyoto, Japan, S~Ved-c.
(2) R . COIFMAK, Y. MEYER. AND V . WJCKEJUIAUSER, Sire propemu of -..eld ,_Jt.
et.f, in Wavelets and Their Applications, M. B. Ruab.i, G . Beylkin, R. Coifman, I.
Daubec:hies, S. Mallat, Y. Meyer, L. Raphael, eda., Jones &.: Bartlett, Boston, MA,
1992, pp. 453-470.
(3) R . COIFMANN, Y . MEYER. S . Q UAKE , AND V . WICKEJUIAUSER, Signol ~9 and
c:ompreuion with tDGlldet poc.l:et.t, preprint, Yale Univmsity. Ne-.. Haven, CT. 1990.
(4) l. DAUB.ECHIES, The Vl01Jelet troruform., timefretluencv loc::olUGtion ond ftgnG1 ~.
IEEE Tr&DI. Inform. Theory. 36 (1990), pp. 961- 1005.
(S) E . Stat. Thesis, University of Paris-Dauphine, CEREMADE, Paris, France, 1992.
CHAPTER
Marr'5 p rogram.
Vision, A Comp1dational l nvt!3tigation into the Human &pruenl4tit>n and Proceuing of Vi.rual Inform4tion appeared in 1982, and it is somewhat analogous to
Descartes's Di.scours de la methode. Exactly as Descartes did, Marr takes us into
his confidence and speaks to us as if we were one of his friends cOlleagues from
his laboratory. Marr confides to us his intellectual progress and tells us about
his doubts, his hopes, and his enthusiasms. He gives a lively description ~ the
theories be has struggled with and rejected, and he explains his own research
program to the reader with a sense of jubilation.
We recall that the goal of Marvin Minsky's group at t he Massachusetts Institute of Technology (MIT) artificial intelligence laboratory was to solve the
problem of artificial vision for robots. The challenge was to construct a robot
endowed with a perception of its environment that enabled it to perform specific tasks. It turned out that the first attempts to coost.ruct a robot capable of
understanding its surroundings were completely unsuccessful.
These surprising setbacks showed that the problem of artificial vision was
much more difficult that it seemed. The idea then occurred to imitate, within
the limits imp<l6ed by the technology of robotics, certain solutions found in
nature. Marr. who li.'&S an expert on the human visual system. was invited to join
the MIT group and leave Cambridge, England, for Cambridge. Massachusetts.
According to Marr, the disappointments of the robotics scientists were due to
having skipped a step. They had tried to go directly from the statement of the
problem to its solution without having at band the basic scientific- under.;tanding
that ill nCCC!l.~ to ronstmc-t effcctivt' al_gorithms.
llll
102
CHAPTER
Marr's first premille is that there exists a science of vision, that it must be
developed, and that once there has been sufficient progress, the problems posed
by vision in robots can be solved.
Marr's eecond premise is that the science o human vision is no different from
the science of robotic (or computer) vision.
Marr's third premise is that it is as vain to imitate nature in the case of vision
as it would have been to construct an airplane by imitating the form of birds
and the structure of their feathers. On the other hand, he notes that the laws of
aerodynamics explain the flight of birds and enable us to build airplanes.
Thus it is important, as much for human vision as for computer vision, tv
establish scientific foundations rather than blindly seeking solutions.
To develop this basic science, one must carefully define the scope of inquir)'.
In the case of human vision, one must clearly exclude everything that depends on
training, culture, taste, etc., for instance, the ability to distinguish the canvas of a
m~r from that of an imitator. One retains only the mechanical or involuntary
aspects of vision, that is, those aspects that enable us to move around, to drive
a car, and so on.
We thus limit the following discussion to low-level vision: This is the aspect
of vision than enables us to recreate the three-dimensional organization of the
physical world around us from the luminous excitations that stimulate the retina.
The notion that low-level vision functions according to universal scientific
algorithms seemed, to some, an implausible idea, and it encountered two kinds
of opposition. In the first place, neurophysiologists had discovered certain cells
having specific visual functions. But Marr objected to this reductionist approach
to the problems of vision, and he offered two criticisms on this subject:
(a) After several very stimulating discoveries, neurophysiologists had not
made sufficient progress to enable them to explain the action of the human
visual system based on a collection of ad hoc cells.
(b) It would be absurd to look for the cell that lets you immediate!~ recognize
your grandmother.
On another front. Marr objected to attempts by psychologists to relate th<>
performance of the human \'isual system to a learning process: We recognize thE>
familiar objects of our environment by dint of having seen and touched them
simultaneously. Bela Julesz had made a fundamental discovery that eliminated
this as a working hypothesis.
Julesz made a systematic study of the response of the human visual. S~'Stem
when it V."&S presented with complete!~ artificial images-synthetic images having no significanc~which were computer-generated. random-dot stereograms.
If these synthetic images presented a certain ~formal structure- that stimulated
s~vision reflexes. the eye deduced. in several millisecond!' and ll:ithout thr
slightest hesitation. a three-dimensional organization of the image. This organizatioo "in relier is clearly only a mirage in which the mechanism of sterovision
finds itself trapped. This mechanism acts with the same speed. thf' same quality. and the same precision as if it v.we a matter of recognizing familiar objt'Cts.
Thus familiarity with t hf' objects that onE> sees plays no rolr in th!' primary
mechanism~; of vision. 1\tarr !'<'t out to understand thr algorithmic- ardtit<'<'t ure
103
of these mechanisms.
This venture can be compared to that of the 17th century physiologists who
studied the human body by comparing it with a complex and subtle machinean 888e1Dbly of bones, joints, and nerves whoee functioning could be explained,
calculated, and predicted by the same laws that applied to winches &Dd pulleys.
A century and a half later, Claude Bernard made a similar connection between
the organic functioning of the human body and progress in the naacent field of
organk chemistry. The synthesis of urea (Wohler, 1828) again reduced the gap
betwee.n the chemistry of life.and organic chemistry.
In their scientific approach, these researchers relied on solid, well-founded
knowledge, which came either from mechanics or from chemistry. They then
tried to effect a technology transfer and to apply results acquired in the study
of matter to life science.
But what Marr set out to do was much more difficult because the relevant
knowlege base, namely, an understanding of robotics, was too tenuous to serve
as the nucleus for an explanation of the human visual system.
Marr asserts that the problems posed by human vision or by computer vision
are of the same kind and that they are part of a coherent and rigorous theory,
an articulate and logical doctrine.
It is advisable, at the outset, to set aside any consideration of whether the
results will ultimately be implemented with copper wires or nerve cells and to
limit the investigation to the four properties of human vision that we wish to
imitate or reproduce in robots. These are
(a) the recognition of contours of objects, the contours that delimit objects
and structure the environment into distinct objects,
(b) the sense of the third dimension from twc:Hiimensional retinal images and
the ability to arrive at a three-dimensional organization of the physical world,
(c) the extraction of relief from shadows, and
{d) the perception of movement in an animated scene..
The fundamental questions posed by Marr are the following:
How is it scientifically possible to define the contours of objects from the
variations of their light intensity?
How is it possible to sense depth?
How is movement sensed? How do we recognize that an object bas moved
by examining a succession of_i~ages?
M.arr opened a very active
of contemporary scientific research by giving
each of these problems a precise algorithmic formulation and by furnishing parts
of thE' solution in the form of algorithms.
Marrs working hypothesis is that human vision and computer 'ision face the
samE' problems. Thus thE' algorithmic solutions can and must be tested 111.ithin
t he framework of robotics technology and artificial vision.
In case of success. it is necessary to investigate whether these algorithms art>
physiologically realistic. For example, Marr did not believe that neuronal circuits
used iterative loops. which are an essential aspect. of thE' existing algorithms.
Thi!' discussion raise$ th<' basic- problem of knowing the nature of thr reprr
.~rntatr cm (th!' terminology i~< dut w Marr) on 111.hich th<' alf!;orithm~ act. Marr
area
CHAPTEk 8
104
uses a simple comparison to help us underStand the implications of a representation. If the problem at band was adding integers, then the representation of
the integers could be given in the Roman system, in the decimal system, or in
the binary system. These three systems provide three represent4tions of the integers. But the algorithms used for addition will be different in the three cases,
and they wiU vary greatly in difficulty. This shows that the choice of this or that
repre$ento.tion involves significant cooaequences.
Marr was truly fascinated by the esaential and subtle role played by a representation. He wrote: "A representation, therefore, is not a foreign idea at all- we
use representations all the time. However, the notion that one can capture somt"
aspect of reality by making a description of it using a symbol and that to do so
can be useful seems to me a fascinating and powerful idea. . . r
8 .2.
Marr felt that image processing in the human visual system has a complex hierarchical structure, involving several layers of processing. The "low-level processing" furnishes a repre$enlation that is used by later stages of visual information
processing. Based on a very precise analysis of the functioning of the ganglion
cells, Ml!JT was led to this hypothesis: The basic representation ("the raw primal
sketch" ) furnished by the retinal system is a successioo of sketches at different
scales and these scales are in geometric progression. These sketches are made
with lines, and these lines are the famous zero-crossings that Marr uses in the
following argument:
TM firat of the three stages described above concerns the detection of intensity
changes. The two idea.s undering their detection are ( 1) that intensity ch4nges
occur at different scales in an image, and 40 their optimal detection requires the
we of operators of different sizes, and {2) that a sudden intensity change will
give rise to a peak or trough in the first deriootive or, equivalently, to a zerocrossing in the second derivative. These idea.s suggest that in order to detect
intensity changes efficiently, one should seo.rch for a filter that has two salient
charocteristics. First and foremost. it should be a differential operator, taking
either a first or second derivative of the image. Second. it should be capable of
being tuned to act at any desired scale. so that large filters can br used to detect
blurry shtulow edges. and 8mall ones to detect sharply focused fin e details in the
image.
Marr and Hildreth (1980) argued that the most satisfactory operator fulfilling those conditions is the filter t::.G , where 1::. is the Laplacian operator
{)2 f ox2 + lf2 !Otf and G stands for the two-dimensional Gaussian distribution
2
G(x , y ) = e-<:r'+s? l/ 2'"' which has standard deviation of o/?.. ~G is circularly svmmetric Mexican-hat-shaped operator whose distributaOTl Ul two dimensions may be expressed in term.. of the radial distance r from thr origin by th(
fortnula
~
2 (
r ) _,., ,2.,,
1::.G( r ) = - CT2 l - 2CT 2 c
...
*).
COMI' l~TER
105
the gray levels f (x , y ), the zero-crossmg. of Marr's theory are the lines of the
equation (f tP.r )(x, y) 0. Since the function ,P(x,y} is even, the convolutiou
product (f 1/J.r)(X,lf) is (up to a proportionality factor) the wavelet coefficient
of f , analyzed with the wavelet ,P. Thus the zero-crossings are defined by the
vanishing of the wavelet coefficients.
It remains to specify the values of a. These values, in geometric progression,
were discovered by Campbell, Robson, Wilson, Giese, and Bergen. based on neurophysiological experiments. T hese experiments led to the values oi = (1.75)Joo.
Marr's conjecture is that the original image f (x, y ) is completely determiited
by the sequence of lines where the functions (f
)(x,y) are zero. Interest
in this representation of an image stems from its invariance under translations.
rotations, and dilations (by p owers of 1.75).
We quote Marr: "Zer<r<:rossings provide a natural way of moving from au
analogue or continuous representation like the two-dimensional image intensity
values I (x, y) to a discrete, symbolic representation. A fascinating thing about
this representation is that it probably inCUIS no loss of information. The arguments supporting this are not yet secure . . . "
In the following pages, we propose to study Marr's conjecture. We will show
first of all that it is incorrect for periodic images covering an unbounded idea.
This does not exclude the possibility that the conjecture is true for bounded
images, t hat is, images having finite extent.
We will then examine Mallat's conjecture, which is a version of Marr's conjecture. Mal.lat's conjecture provides an explicit algorithm for reconstructing
t he image. This algorithm works very well in spite of the fact that Mal.lat's
conjecture is false. The counterexample that we coDBttUct is, in a certain sense,
more realistic than the one we present in the case o! Marr's conjecture. This
counterexample raises an exciting problem: 'Why does Mallat's algorithm work
so well?
"''a,
8.3.
(8.1 )
j (:r)
= sin:r..;. 2: a;..sinkx,
when
oc
(8.2)
1.
W<' art ~in!' w show that all choices of tlw coefficients a~: lead tl. t.h<'
sum<' zero-crossin~. For cxumplt' sin :r and sin :r + f, sin 2:r havt> tltt srum' zrr~
CHAPTER 8
106
croesinp. We verify t.uis assertion by .yat.ematically applying the following simple obeervation: H u(:r) and v(:r) are two continuous functions of :r, and if, for
aome constallt r E [0,1), lv(:r)l ~ r lu(:r)l for all :r, then u(:r) + v(:r)
0 is
equivalent to u(:r) 0.
Returning to (8.1), we define 96(:t}
= 6Ji;e-a'f'U' . Then
oc
- tf22 (fg6)(x )
d.t
Observe that lsin kxl ~ kl sinxl. which implies that lv(x)l ~ rju(x}l, where
r = E~ k 3 1okl < 1. Thus the zero-crossings of all the functions f (x ) are
x=mr., mE Z.
H we wish to have 0 ~ f (x) ~ 1. it is sufficient to add a suitable constant to
f (x ) (defined by (8.1)) and then to renormalize the result by multiplication with
a suitable positive constant. These two operations do not change the positions
of the zero-crossings.
A nontrivial two-dimensional counterexample is
f(x, y)
= sinxsiny+ L Oksinkxsinky,
2
where
8.4.
Mallat's conjecture.
The existence of these counterexamples and several remarks Marr made in his
book led Stephane Mallat to a more precise version of Marr's conjecture. He
gave it a formulation that was compatible "''ith the progress made in the early
1980s on pyramid algorithms for numerical image processing.
Mallat observed that numerical image processing using certain kinds of p~rra
mid algorithms (quadrature mirror filters) and Marr's approach represented two
particular examples of wavelet anal~sis of an image.
In fact. one has 6(/ 96 ) = E- 2 f t:l' " bert>
2
tr(x , y )
,l) ( :r + J?)
,. = -;;:1 ( 1 - --2
exp
- - ---
107
tzu.
tz
f(x)
i
I
-----~---T---,-------- -Xt
X2
X3
:
1
~------
lt_(f8.t)
dx
o
I
I
I
I
I
I
I
I
I
~(f 8~J) I:
dY~
101:!
CHAPTEH
Coding f with the double aequence (:r9_,, liN) meets two objectives: It is
invariant under translation, and it corresponds to a precise form of Marr's conjecture. One reads on page 68 of his book: "On the other band, we do have
extra information, namely, the values of the slopes of the curves as they cross
zero, since this corresponds roughly to the contrast of the underlying edge in
the image. An analytic approach to the problem seems to be difficult, but in an
empirical investigation, Nishihara (1981) found encouraging evidence supporting the view that a tv.'<Hiimensional filtered image can be reconstructed from its
zero-crossings and their slopes."
We are going to show that this conjecture is incorrect. However, this assertion
must be tempered because our counterexample depends on a specific choice for
the function 8(x ). If 8(x ) is the cubic spline, then we have a counterexample.
II, on the other hand, the cubic spline 8 (:r) is replaced with the function that is
equal to l+cos:r if lzl ~ 1r and to 0 if lzl > 1r (which is the Tukey window), then,
for all signals f (x) with compact support, reconstruction is theoretically possible
but unstable. In this case, it comes down to determining a function with compact
support from the knowledge of its Fourier transform in the neighborhood of zero.
which is an unstable problem.
To describe the counterexample. we make a change of scale so that the values
of 6 are 2w2-i rather than 2-i, j E z.
Then we define /o(:r) 1 + cosx if - w ~ :r ~ 1r and fo(x) = 0 if l:rl ~ x. We
next consider the function f 1 (x ) defined by
b (:r) = fo(x ) +
OC>
if
- r. S r S r.
and
il (:r} =O if i:r~~ :r.
We cb()()S(> the coefficients ak , k
l o~~:lk"'-
(8.3)
0. so that
0 as k - +oc
for all m
~ 1.
(8.4)
oe
(8.5)
jEZ.
1 SIS /1.'.
Then if 0 ~ {> S 'ff. (tfl f d:.r 2 }(/J 8~ ) and (d2 /d:.r 2 )(j 0 .. 6~ ) h avC' thC' 5anlC' zeros.
namely. at :r = ~ or at l:rl ~ r. + 2f. ThC' fa<'t that tbC' dC'r ivnt iVl'l' of / 1 6t
and / o 81- ar(' th; same at thes<' zcro-crossin(:l' r~ults from (t:.;,).
109
8.5.
We start v.i th a two-dimensional image f (x , y). From this we create the increasingly blurred versions at scales 6 = 2- i , j e Z , by taking the various
convolution products f 86. where, in two dimensions. 86(x, y) = 86(x)86(Y) =
t>- 2 8(x f 6)8(y/ 6). The function 8(x) is the basic cubic spline used in one dimension.
Next we consider the local maxima of the modulus of the gradient of f 8~.
We keep in memory the positions of these local maxima as well as the gradients at these points. The conjecture is that this data. computed for 6 = 2- 1 ,
characterizes the image who.se gray levels are given by f (x,y).
We will show that this conjecture is incorrect in this general form. This does
not exclude the possibility of its being true if (1) more restrictive assumptions are
made about the function f(x , y) or (2) the definition of the smoothing operator
is changed.
The counterexample in two dimensions will be f (x , y ) = 11 (x ) + / 1 ly). v.here
f 1 (:r) is the counterexample in one dlluension. One should observe bert> that
0('
f 1 (:r)
= f o\:r) -
and
for
- r. S r S :-:-
! 1 \.T\ :::,
~ fo(r) since
1 ..- cus(2k
+ 1):r S (2k +
112 (1
+ cos.:r ) .
CHAPTERS
110
The support of the function f(z, Jl) is thus the CTOSS -1!' $ X $ 11' Or -11' $ Jl $
since
7
9
8(/o(z) + /o(JI)) :5/(:r,JI) $ 8(/o(z) + /o(JI)).
Finally,
f(z,JI) 96(x)96(JI)
11'
Its length is (1 1z (ll 96)12 + 1 ~(/t 96}fl)112 It has a maximum if and only
if l!(h 96)1and lt 11 (11 96)1are at a maximum. But the set of functions ft (x )
has been constructed so that the positions of the maxima of l! (h 96)1 are independent of the choice of b and the same is true for the values of~ (b 96) at
these points when 6 211'2-i, j e z. Thus we have a counterexample in dimension two. Finding a counterexample whose support is a square is an important
unsolved problem.
8.6.
Conclusions.
All of this shows that Marr 's conjecture is doubtful. However, the counterexample that we have constructed is deficient because it does not have finite support.
Regarding Mallat's conjecture, one must distinguish between the problem
of unique representation and that of stable reconstruction. In our opinion, the
reconstruction is never stable (unless the class of images to which the algorithm
is applied is seriously limited). But it is, in certain cases. a representation that
defines the image uniquely.
Bibliography
S. G. MALLAT, M..Uifrequency ch4nnd decomposil>On$ of lf7149U and wavelet modeLs.
IEEE Ttans. on Acoustics. Speech Signal Proce!S., 37 (1989). pp. 2091- 2110.
(2] S. G. MALLAT AND W .-L . HWANC. Singulorit11 detectwn and procu8ing with wavdct.s.
[1]
Robotics Report No. 245. Courant Institute of Mathematical Sciences, Jlie,.. York. t\' Y.
March 1990.
(3] S. G . 1\lALLAT AND S. ZHONC . Completf ssgnal repruentatwn with multu cale ed9e...
Robotics Report No. 219. Courant institute of Mathematical Sciences, New York.
1\'Y. December 1989.
CHAPTER
9.1.
Introduction.
This title, in fact, pertains to the last three chapters of the book. In the next
chapter, wavelets will be used to analyze the multifractal structure of fully d~
veloped turbulence. Similarly, in Chapter 11, the astronomer Albert Bijaoui will
propose wavelets as a tool capable of elucidating the multifractal structure of
the hierarchical organization of the galaxies.
In this chapter, we show that Fourier analysis does not easily reveal the
multifractal structure of a signal and that ~velet analysis provides immediate
access to information that is obscured by tim~frequency methods. The two
examples we present have played a significant role in the history of mathematics
and, for this reason, this chapter has been written with extra care, by presenting
all the details of the arguments and demonstrations. We invite the reader to
accept our apologies, if he or she is not a mathematician, and to proceed directly
to Chapter 10. We hope that readers who are mathematicians will enjoy reading
this chapter as much as we have enjoyed writing it.
9.2.
We intend to shmo. that the function u (t) = I:;' 2-i cos(2' t) is nowhere differentiable and that the same is true for the function Ego' 2-i sin(2it). These proofs
will use wavelet analysis, which, in this example, appears in general outline as a
form of Littlewood- Paley analysis. The method we follow is due to Geza Freud.
Let 1/l(t) denote a function belonging to the holomorphic Hardr space H 2 (R).
Assume, in addition. that 11/l(t)l $ C/lt!3 if ltl ~ 1 and that thE' Fourier transform tb({) of 1!(t ) satisfies the following conditions:
(9.1)
(9.2)
tb({) = 0 if
(9.3)
tb(l )
{ ;::: 2,
= 1.
=
b~ [:).i j
lll
CHAPTER 9
112
filters that are arrayed on octave intervals. Thus the analysis of a real function
using the eequence 11;, j
0,1, 2, ... resembles a Littlewood- Paley analysis
that would be carried out on the analytic signal whose real part is f.
Freud's method is based on the following lemma.
LEMMA 9.1. Ld f(t) be a~ continuow functicn of the real varWble
t. Auurne that f is differentiable at to. Then
11;/(to)
= 2-it;
where t;-+ 0
aJ
j-+ +oo.
E:
e'
(9.4)
were
113
9.3.
(9.5}
-to I)],
where here and elsewhere. the constant. C may va.I)' from expression to expression.
Note that only the values h $ 1 are important in {9.6) since f (t ) is assumed
to be bounded; if (9.6) holds for these v-alues. then (9.7) is satisfied for all j ~ 0.
T H EOREM 9.1. Let f be a function of the real variable t that ism the class
C'0 and let to be a fixed real number. Assume that the wavelet coefficient.~ o (j. k )
of I satisfy, for j ~ 0 and all k E Z, the condition
(9.8)
'
(9.9)
1
0
ca
TJ(X)- <
:r
00.
Tllen f i..~ differentiable at to. and /'(to) can be computed by differenhating tht
Ulavelet expansion off trrn by tcrnl.
CHAPTER 9
114
Theorem 9.1 ia oot exactly the convene of Lemma 9.3 because we use two
additional conditions: a global regularity hypothesis on f and the Dini condition
{9.9).
The proof of Theorem 9.1 is so nice that we cannot resist the temptation to
present it.
In our notation, both the constant C and the function '7 may vary from place
to place. To be specific, when functions of the form TJ(Kx) = TJK (x ), K 2! 0,
appear in estimates, we replace these (which are finite in number) by the largest
of the '1x and rename it '7 Note that all the '1K satisfy the same conditions
as '7
L
j ~O
A:
where these three sums are taken over three intervals of j values.
Define the integer io by 2-io ~ ~ < 2 2-io and then define J1 2! i o by
2-J ~ lh l 2/ o < 2 2-J. Write
(9.10)
s~ (h) =
..I:
OSiSio
S2(h) is the analogous sum taken over j 0 < j ~ J1o and S3 (h) involves the values
j > il
We use two facts to estimate IS1 (h)l. On one hand. the functions 1i:;.A:(t) are
sufficiently regular so tha.t
(9.11)
and. on the other band. the support of tP;.k is the interval [k2-i. (k + L )2-i]. Once
to and hare fixed, the number of values of k for which ~;.~< (to) and 1Y;.~c ( to +h)
a.re not zero is no greater than 3L. For these values of k. (9.8) reduces to
lo(j.k)l ~ C2 - 3ii 2 TJ(2-i) since lk2 -i- tol ~ C2- i. Finally.
IS1 (h)l ~ Ch 2
(9.12)
2i'1(2-i)
o(h).
OSiSio
(9.13)
I:
}o<1~J ..
2:
Jo<1:S.1
,,(:?--').
115
We have assumed that J~ 17(t)~ < oo and that f}(t) is an iDcreasing function. From this it follows that r:''7(2-.1) < oo. Furthermore, '7(2- ') decreases to 0, from which we conclude that lim,t-+ooi'1(2-.1)
0 and, hence,
that {log )'I( h) - 0. But j 1 - j 0 is of the order log with the result that
limh-o iS2(h) 0.
To bound IS3 (h)l, we use only the hypothesis on the global regularity off
as expressed in terms of the wavelet coefficients (9.3). We again take absolute
values inside the sum. By considering, as in the second case, the limitations
imposed by the compact support of', we deduce that IS3 (h)l $ Ch 2 , which
implies that limh-o *S3(h) = 0.
This completes the proof of Theorem 9.1; however, this is not the result that
will be used in the next section. We will need a "corollary of the proof" of
Theorem 9.1. The statement of the corollary itself follows.
COROLLARY. Let 8(x, y), x E R, 0 < t1 $ 1, be a mea.rurable function of
(x , y ) satufying, in addition, the following two conditions.
(9.14)
for all x E R
(9.15)
where the function '1 satUfies the hfiPOlhues of Thecrem 9.1, in particular (9.9).
Suppose that g(t ) is a function in the class C 2 with compact 8Upport. Define
11
1 00
f (t ) =
(9.16)
-oo
(t
dx.J!.
d .
tl
/ '(to)=
(9.17)
11 Lroooo
8(x , y )g'
(~ - X) dxd~.
tl
tl
Study
llli
CHAPTER
Not being able to prove that W(t} was nowhere differentiable, Weierstrass
considered the more lacunary series o(t} = I:;' b" cos(a"t), 0 < b < 1, and
be showed that if a is an odd integer 8lld if ab is sufficiently large, tbe.n o(t)
is nowhere differentiable. We have seen that the result is true for a
2 and
b = 1/2; in fact, ab ~ 1 is sufficient.
In 1916, G. H. Hardy proved that the Riemann function W(t ) is not differentiable at to in the following three cases:
(9.18)
(9.19)
(9.20)
to is irrational;
t0 = p/q with p =0 (mod 2) and q = 1 (mod 4);
to = p/ q with p = 1 (mod 2} 8lld q = 2 (mod 4).
00
F (t) =
(9.21)
n12 ein't.
= x + iy,
F (.: ) = ~ ~e'"""
(9.22)
~n2
,.
The most naturul wavelet analysis for such holomorphic functions is the analysis advocated by Lusin (2.5). The wavelet transform ofF(.:) is, up to a normalization, .F'(z) ir.l:~ e>q>(i1r11 2 .:)
!f(9(::) - 1). where 9(z) is the Jacobi
junction defined by
IX
(9.23)
9(.z ) =
2: e>.'J>(i1r112.:>.
:=X T-iy.
> 0.
-oc
It is easy to study the beba\'ior of the function 9(::) near the real axis by using
the functional equation satisfied by 9(: ). And everything v..-ould go well in the
best of all possible, worlds if the analyzing vavelet. w(t) = (t + i )- 2. satisfied
the minimal condition allowing the corollary of Theorem 9.1 to be used. namely,
that J~00 (1 + ltl}lt/J(t)ldt < oo. But this integral diverges.
Holschneider had the remarkabl<' idea to start vith the result of th<' analysi~
obtaine-d with the "bad wavelet." and then to do the s~nthcsi.s with th<' -~ood
wa\'t>leC g(t). Thus he cousidered 1\ function g(t ). in the class C2 and with
u;
where
g denotes the Fourier transform of g . We then have, for all .>. > 0,
(9.25)
and
oc.
(9.26)
II
1r
'
n2 .
1oo
[ 9 (i.-y x) 8(x + iy)dx dy = ~ t e-:2'
.
o
n
(9.27)
Jl
-oo
1r
Let to = ~":t! . To demonstrate that the second member of (9.27) is differentiable at to, we use the corollary of Theorem 9.1.
First of all, the contribution from y 2: 1 in (9.27) can be neglected. In fact,
19(z) - 11~ Ce- 11 when y 2: 1 and Z = X+ iy 80 that the integral in J1 from 1
to in.fi.n.ity yields a function in the class C2 or, more generally, having the same
regularity as g(t).
To establish the differentiability at t 0 , observe that
(9.28)
and
<y
~ 1
and that
(9.29)
if z = x + iy,
0 < Jl
1.
We can then apply the corollary of Theorem 9:1 and prove Gerver's theorem.
9.5.
Conclusions.
In V"iew of the example of the Riemann function . one would be tempted to conclude that wavelet analysis is better than Fourier analysis for studying fractal
structures. The Riemann function is given explicitly in terms of a Fourier series
and. yet, this exact spectral information yields no direct access to the pointv.rise regulari~ of the function . This conclusion. however. is incorrect for several
reasons.
First of all. the simplest and most precise method'for studying the Riemann
function has been found by ltatsu !9]. and it is not a timr-sct:Uc method. On the
contrary. ltatsus method is based on the judicious use of the Poisson summation
formullt. It is thus a "timf'-frequency"' method. and it ftives morr preci.~r results
than thOS(' obtained by wavelet analysis.
118
CHAPTER 9
In the eecond pla.ce, we know today that wavelet analysis is part of a much
larger diadpline, namely, 2-microlocal analysis (12), (10)). 2-microlocal analysis
saw the light of day at about the same time as wavelet analysis. The dictionary
allowing us to go from one type of analysis to the other has been worked out by
Jaffard (10].
Should one use wavelet analysis or 2-microlocal analysis? We refer the reader
to (ll), where this problem is carefully studied. My (subjective) point of view
is that 2-microlocal analysis is a more flexible and more precise instrument but
that its sophistication can discourage some scientists. Wavelet analysis offers an
apparent simplicity, 9:hich is one of th<> explanations of its success.
Bibliography
(1) A . ARNEODO, E . BACRY, AND J . f . M\:ZY. Wavelet. and mvltifrodolformolinn for
ingulor ngnoh: Applicotiom to turbukna doto, preprinl, Centre de Recherche
Paul Pascal, Pesaac, France, 1991.
[2] J. M . BONY, Second microlocolUotwn and J1"0P09ation of nngulorit~ for aemi-linMr
ht,tperbolic equotioru, Tanjguchi Symp. HERT. Katata, (1984), pp. 11-49.
(3] - -, Interaction du nngulorith pour l'iquation de Klein-Gordon non lineoin.
8eminaire Goulao~ 1983-1984, Ecole Polyteclullque, Centre de Matbematique.
Palaiaeau, France.
[4) P . L. BUTZER AND E. L. STARK, MR~nn' e=mple 8 of a oontin...,... nondifferenti4blt
function in tM light of t1110 kttera (1865) of ChrUtoffel to Prym, Bull. Soc. Math.
Belg., 38 (1986), pp. ~73.
[S) J . J . DutSTERMAAT, Selfmnilori.ty of 'Riemonn'a nondifferenJ>4ble ftmction, Nieuw
Arch. Wisk., 9 (1991), pp. 303-337.
(6) J. GERVER, The differentiability of the Riemann function at cm.un rational multipl~
ofr., Amer. J. Math., 92 (1970). pp. 33-SS.
(7] - -, More on the differentiobiltty of tM Riemann funchon. Amer. J. Math. , 93 (1970j.
pp. 33-41.
(8) M . HOJ..SCHNEIDER AND PH . TCHAMITCHIAI'> , Point!UUe onolyA< of Riemann'! '"non
differentiable" function.. lnventiones Mathematicae, lOS (1991 ). pp. 157- 176.
[9] S. ITATSt:. The diff~ti4hilit!l of 1M Raemann function. Proc. Japan Acad .. Ser. A .
Math. Sci. 57 (1981). pp. 492-495.
(10] S. JAFFARD, Pointwi.se m100thneaa. two macrolocoli.z'ation and wovelet c:oefficaents. Pui:r
licacions Mathematiques (Publicacions de Ia Universitat Autbnoma de Barcelona).
35 (1991), pp. 155-168.
(11] Y . Ill EYER. L 'anolyae par ondelette. d un objet multifractol: w function
;!-, sin n 2 t
de Raemann. Colloquium 1\l athematque de l'liniversite de Rennes. Rennes. Franc~ .
November 1991.
(12; H . QVEFFELEC. Derivobiliti de certoane. !ommes d e airte. de Feu.,~ locunotrc. Thesi~.
University of O rsay, Orsay. FrancE'. 1971.
L;7"
CHAPTER
10
10.1.
Introduction.
10.2.
CHAPTI::H. lU
120
and the statistical description is derived from the corresponding correlation functions. The mathematical tool adapted to this statistical approach is the Fourier
transform, and, by associating scale and freQuency in the usual way, Kolmogorov
and Obukhov arrived at e213 jkj- 1113 for the average spectral distribution of energy, where k is the vector-variable of the fourier transform. which was ~en
over the space variables Xt, x 2, and X3.
Various wind-tunnel experiments {Batchelor and Townsend (1940); and
Anselmet, Gagne, Hopfinger, and Antonio (1984)] have shown that the energy
associated with the small scales of a turbulent flow is not distributed uniformly
in space. This observation, that the support of the transfer of energy is spatially
intermitt.ent, has led several authors to hypothesize that this support is fract.al
{Mandelbrot (1975); and Frisch. Sulem, and 1\elkin (1978JJ or multifractal !Pa:isi
and Frisch (1985)].
Use of the Fourier transform does not elucidate the multifractal structurt'
of fully developed turbulence, and this observation leads us to an important
application of the wavelet transform.
10.3.
The conjecture of Parisi and Frisch 19] is based on experimental data supplied
by Anselmet, Gagne, Hopfinger, and Antonio. The computation that led to
their conjecture involved evaluating, as a function of the displacement l ~xl, the
average value of the pth power of the change in the vector velocity in a turbulent fluid, that is the average value of jv(x + ~. t) - v (x . t)IP. The surprising
result from these measurements was that they obtained a power law in terms
of I ~IC(p), where the exponent ({p} does not depend linearly on p. The interpretation given by Parisi and Frisch is that turbulent flow develops multifractal
singularities when the Reynolds number becomes very large.
The relation between the multi.fractal structure and the power law is given
by the following heuristic reasoning. To speak of a multifractal structure means
that, for each h > 0, there is a set of singular points v.'ith Hausdorff dimension
D (h) on which the increase in velocity acts like 16xlh. The contribution of these
"singularities of exponent h" to the average value of lt(x + Ax. t)- v(x. t)!" is
of the order of magnitude of the product l~ l"h i~1 3 -D(h l: the second factor
is the probability that a ball of radius l6xl intersects a fractal set of dimension
D (h).
\Vhen lAx I tends to 0, the dominant term is the one " ith tht' smallest possible
ex-ponent. which leads to
((p)
= hinf>(l{ph -r 3- D(h)}.
The C."\.-ponent ((p ) is thus given by the Legendre transform of function D(h).
which is. we recall. the Hausdorff dimension of the set of ex>ptional point :r.
where ltl:r + A:r. t)- v (x , t)l is the order of magnitude of IA:r 11' . The nonlinear
depcndena> of ((p) on p thus indi~ that th<' very abrupt \nrintions in velocity
correspond to a multifractal structure.
121
The wavelet transformation ~ the ideal tool for arwlyzing multifractal strt~C
tures, and Fh&ch has Mught to verif11 hi$ ronjecture by plunging into the VCJ
heart of the turbulent signal and traveling ocrcm the 6CGlu to CtJlculc.te the frru:t4
exponeni6.
The turbulent signal, which is the object of the analysis, is, in fact, a functiOII
of time obtained by hot-wire measurements. The turbulent flow comes from tbt
Modane wind tunnel, and the hot-wire technique provides the measurement a1
a given point of the velocity as a function of time. Taylor's hypothesis implies
that, for the particular condition." created in this tunnel. the ~segment in time
i~ equivalent to a ..segment in space'' (along the axis of the tunnel). F inally,
the turbulent signal is analy-t..ed with the wavelet transform and. thanks to a
two-dimensional color visualization, one displays a multifractal structure.
This phenomenological approach has been criticized by Everson, Sirovich, and
Sreenivasan [3]. These investigators have been able to show that wavelet analysis
of Brov;tuan motion produces very similar two-dimensional visualizations.
It then became a matter of urgency to move from the qualitative to the quan
titative and to extract the fractal exponents h and the corresponding Hausdorff
dimensions D (h) from the "Gagne signal." This research is being actively pursued by Alain Arneodo and his collaborators with the objective to confirm the
hypothesis of Frisch and Parisi.
10.4.
Farge's experiments.
CHAPTER
122
10
flow in two
Farge explains to us how and why she was led to use wavelet analysis in her
study of numerical simulations of two-dimensional fully developed turbulence l4j:
The use of the wavelet transform for the study of turbulence owes absoiutely
nothing to chance or fashion but comes from a necessity stemming from the
current development of our ideas about turbulence. If, under the influence of
the statistical approach, we had lost the need to study things in physical space.
the advent of supercomputers and the associated means of visualization have
revealed a zoology specific to turbulent flows, namely, the existence of coherent
structures and their elementary interactions, none of which are accounted for by
the statistical theory.. .
What Farge asks of wavelet analysis (or of any other form of "time-frequency"
analysis) is to decouple the dynamics of the coherent structures from the residual flow. The residual flow would play only a passive role in an action whose
10.5.
(3)
(4}
(5]
FRISCH , Wowlet on4l~u of tvrbulence re.-Lt the multifro,d4J n4tvre of~ Rich4rdon aucode, Nature 338 {1989), pp. 51- 53.
A . ARNtOOO. G . GllASSEAt;, AND M . HOLSCHNEIDER, WClvelet troruform ofmv!Ji]rGctoll, Pbys. Rev., 61 {1988), pp. 2281- 2284.
R . EVEI\SOI' , L. SJROVICH, AND K . R . SR.EENIVASAN, Wcweld onolV8U on ~ tvrbvlent
jet., Pbys. Lett. A , 145 {1990), pp. 314-324.
M . FAROE, 7'ronlfO'fTTik en on<Ulettu ccntinve et oppliaW.cn Q. Ia tvrbulence, Soci~
Mathematique !7an~. Pari.s, !7ance, May 5 , 1990.
A . M . KOLMOCOROV, A refinement of prevoqw hwothuu oonceming the loc4l tn.cture
of tu~ m wc:oouo incompruibk fluid Clt a high Reynold& number, J . Fluid
(9) G . PARISI AI'D U. FRISCH , 7\i.wlene.e and pnedid4bilitJI in geoph~ fluid d~m
ic.! Clnd climcote dynamics, M. Ghil, R. Benzi, and G . Parisi, eels., North-Holland .
Amsterdam. 1985, pp. 71-88.
flO) J . VOl< I"EilMANI' , Complete works. 1949. )Tta.nslator'soo&.e: The quotation i.s from On
the Pnncrples of Large Scole ComJ)Vtmg Mochinu ~ Herman H . Goldstine aod
J ohn von l"eumann. This paper ...-as never published. It coot&ios material given by
von Jl'eum&nn in a number of lectures, in particular ODe at a ITif'eting oo Ma~ 15, 1946.
of th<' l\lathematical Computing Advisory Board, OffiCf' of R~ and Inventions .
N"''Y Depanment. which in 1947 became lhe Office of Naval R~ .J
(11) N . ZAB\'SK\. ComJ)VtClt.i onolaJI'Iergt:tics. Physics Today, July {1984). pp. 2- 11.
CHAPTER
11
11.1.
Introduction.
Wavelets are being used by Albert Bijaoui and his collaborators to clarify the
hierarchical organization of distant galaxies and, possibly from this, to make
deductions about the formation of these galaxies.
The following lines are again "lecture notes," this time from original articles
by Bijaoui and others, and we encourage the interested reader to consult these
articles. Several are listed in the bibliography.
11.2.
The difficulties encountered today in the analysis of the distribution of the galaxies result, paradoxically, from technological progress in the construction of tel~
scopes.
On the one band. the quality of telescopes has been considerably improved
during the last 50 years. In 1950, astronomers could capture ten million galaxies.
Today, 100 million galaxies can be examined ... There is now such a quantity of
information that it is impossible to continue using traditional obsefvations based
on the astronomer's eye and judgment.
On the other band, the very nature of the i.mages coming from these telescopes
has undergone a revolution. Nov:, CCDs (Charge Coupled Devices) replace plates
of silver salts, and "chemical" photography is already a thing of the past. The
future will involvE> electronic reception of photons that come to us from the edge
of the Universe.
CCDs do not providE> images in the traditional sense but. rather. data that
arc suitable for various methods of processing. Tomorrow's telescope$
be
computers that can be accessed from a distance. These computers "ill. '\1\ith
proper soft."arc. automatically execute certain imagt>-processing algorithms. The
optimal utilization of the marvelous progress offered b~ these CCDs necessitates
implementing cffectiv(' algorithms for processing the received images. and this is
when wn\elets cnt.cr.
ill
126
11.3.
CHAPTER 11
This considerable technological progress has allowed us to study the threedimensional orpn1zatioD of the galaxies. The results that have been obtained
challenge the imqinUon. Far from being arbitrarily distributed like points
toeeed at random, the galaxies are organized according to extremely complex
geometric configurations, and these configuratioas contain information. In some
caaes, galaxies ue distributed in filaments; in other cases, they are spread over
huge surfaces that surround only void.
Galaxies aggregate into clusters, which are themselves organized into superclusters, although a galaxy can belong to a supercluster without being part of a
cluster.
One of the goals of contemporary astronomy is to analyze, in detail, the geometry of these hierarchical galactic organizatiODS. Thls analysis is particularly
important in the ease of the most distant galaxies.. Indeed, the images that we
receive from these galaxies date from the beginning of the creation of the Universe. It is possible to imagine that the fascinati..J:I geometry of these hierarchical
organizations of distant galaxies preserves traces of the process of "fragmentation" of the smooth, homogeneous primitive matter that made up the Universe.
11.4.
11.5.
We encounter the samt> situation here as tht> ont' that ar()S(' in tht> study of
turbulence. In that case. the passage from global fractal properties to local fractal properties required tht> use of .,-avelets. 1'ht> hope placed in thE' analysis of
astronomic ~ using wavelets is, .,.;tbout doubt, reasonable if one considers the arguments presented by Marr. Wavelet analysis is used to delimit the
boundaries of objects and, b~ so doing. to arri\~ at their thret>-dimensional organization. But it is precisely this tbret>-dimensional organization that is sought
b~ tht> astroph_vsici.sts.
On the other band, at. a more imaginative lt>,-el. it is tempting to com par<' thE'
flov. of the Unive'S<' over timE' to that of 11 ftuid and w continu<' this ml'taphor
l'lr
Bibliography
(1) PH. BENDJOYA, E . SLEZAK , AND CL. F ROESCHU , T7te IOIGWld tn~Nform, o new tool fm
utroid fomilv deUnninollon, Anru10m. a.nd A.tropbya., to appear.
(2) A . BUAOUl, Algorithmu de lo trran.foml4tion en onddettu, Application !'image u-
(3)
(4)
(5)
(6)
[7)
(8)
Index
G~Morlet theory, 28
rec:l.i8covered by Gro.mann aDd
Morlet, 13
Chqe Coupled Device~
in~y.125
chirp 8ipaJa,
Cieaielaki,22
~.
72
Balian
t~frequency reprweutation, 63
Balian-Low theorem, 35,76-77
Battle
~.31
Beousi
Gau.iao-Markov fields, 19
best basis. Su ol.o Malvar waveleu
approach to turbulence, 123
t~~cyt-.,77
99
Beylkin
waveletS aud numerical analYSis, 122
bi-orthosonal
Coifm&D-W.-, 26-27
com~on. See aLto lipal and imece
~
Daubechiea
bi-or1.hoc0Dal -veleu, with Cohen &Dd
~u. 5~
Calderon. 13
identity
drrompoaition of tlu> identity. 2t
129
130
INDEX
Daubechlee (continued)
con.t.ruction, 43-44
~11187.27
padceu, i l
O.U~Jalf&l'd-JCl!UM
wa..elet.a. 80
motion. 11insta.ntaneous , 66
2
L !0.I}. 29
Sl&tio~ ~als. !I
_...
s~em.
15-16
7~,
131
INDEX
numerical repreeentation, 2
image procee!ing. s~ GUO pyramid
aJaor itbms; signal and image
prooessing
m.tantaneous &equency, 70
wavelets, 22- 24
Mall&t
Mandelbrot
Jaffard
.....velets,
77, 80
Gaussian-Markov fields, 19
2-microloc:al e.nalysis, 118
solitons. 122
Lang. 116
Laplacian pyramid. 51
~besgue. 15
~marif..:llleYer wavelets. 80-Sl
~rav
.
stat~ICal theory of turbulence. 119
L~n~d
Lu.in
80-85
.
almllar to ...,.ve)et pacltet.&, 92, 96
auited to apeech aod muaic, 7
variable length wiDdowa, 77~
exampl5. 71
not a local property, 72
inst&ntaneoU& epectrum, 71
Ville's ideas, 64, 66
interpOlation operator, 47
ltattu
Riemann's function, 117
Daubechles-Jaffard-Journ~
~mace~.21
Brownian 1110tlon, 19
remark on "Ftane-Cult,u re," 7
1\ory of &ac:Ws, 14
turbuleDce, 120
Man:inldewicz
atomic decomposit ion for V'(0,1], 24
Ma:r
conjecture, 104-105, 110
counterexample. 105-106
Mallat'a formulation, 107- 108
ideas
astronomy, 1.26
optimal repreeentation, 9-10
repreaentation of visual information,
103- 104
adence of villon, 102- 103
VWion, 45, 101
wavelet. 10
merging. 81-62
wavelet packet. 96-99
Minsky. 101
Montaigne's EuoJI&. 1
Morlet. Sec Grossmann- Morlel
mot her wavelet , 5
first appearance in matbematics. 20
joined by fatheT ...avelet, 41. 43
Moyal's formula. 66
multifrac:tal. See ol.fo fractal
approach w structure of thC' Uni~. 11
hypothesi~' of Fnsch and Parisi a bout
turbulence:. 120
structure of Riemann ' f function. 17
INDEX
132
NyquWt. coDdltion, 47
O'Neil. See Woods and O'Neil
optimal algorithms, 7-9. See GUO energy
criterion; entropy criterion
orthogonal pyramids, 56. See GUo pyramid
algorithm
coding is efficient, 55, 57
exact reconstruction, 57
orthonormal wavelet baaes as limits. 5i
Parisi. See Frisch
perfect (exact) reconstruction property. 29. 3i
phy&ical image
filtering and sampling, 46-48
peeudodifferential calculw;, 66-07, 69
pyramid algorithm
bi-ortbogonal wavelets, 59-61
Burt and Adeleon's algorithms, 4&-50,
51- 53
cartographic example, 45
examples, ~51
Ade~~
sere
INUI::X
tmae-frequency ( contonued }
Lienard, 64--65
Lienatd 's compared to Malvar'e
wavelete, 79
Mou.rt'a Magi<; Flute, 6S
methods, 111
plane
ideu of Gabor, Balian, Ville, Lienard
63-65
wavelets, 6-7
time-41C&le analyais (wavelets). See a.Uo
multireeolution analysis
Grossmann-Morlet, H
Malvar's waveleu versus time-scaltwaveleu, 80
t ransfer function, 56
transient signals. 4. 6, 8
transition operators, 48-50
u&ed for c.ocfulg, 51-53
t rend and ftuctuation , 39
Mallat's "berringboneM algorithm, 40
pyramid algorithms, 51-52, 55-57
turbulence
Anleodo's wavelet. analysis, 121
hypothesis of Fn.c:b and Parisi, 120-121
Mar;e Farge's ideas, 75, 121- 122
atawtical theory, 119
Wickerhau.eer's analyais, 122
2-microlocal analysis, 118
unconditional basis, 26-27
Ville
analytic signals, 23
133
Riem.ann'a fuoction, 116-11 7
.Jgnals with fract.al ~ructure, 4
diacrete cue, 99
fmding optimal balls, 98-99
frequency localization, 98
general - velet peckeu, 95-96
roou in Galand'e ideu, 33-34, 36
we;emrue
fuoctioD,ll1
Zabuslcy
cobereDt structures in turbulence, 121- 122
zero-croesi..np. See Marrs conjecture
Zygmund. 20