Contents INTRODUC'TION
vri
PREI.IMINARIES A N D NOTATION
xi
i ".
2 .;
'8-
:.. CHAPTER I . The What. Why. ant1 How of Wavelets
;t<-,.
i
-.-c.
--''jE- CHAPTEK 2
&
*a,?*
-
he Cont~nuouswavelet Transform
.*
CHAPTER %
D~bcrrteWdvcle! 1 ranhtorms: Frames
.r?,
.;!07
-
_ I L
&
'
CHAPTER 4 Tlrnc-Frequency Den\ity and O~thonormalBases
129
CHAPTER 5: Onh~~norrnal Bahes
167
CHAPTER 6. Orthonomlal Haw\ ol Compactly Supported Wavelets
of
Wavelets and Multtresolution Analysis
I
.---
2(5
CHAPTER 7: MOIL. A ~ O Uthe I Rceylartty of Compactly Supported Wavelet?,
25 1
CHAPTER X: Sy~nmstrylor Compactly Supported Wavelet Bases
1
CHAPTER 9 Ch;triiclcri~atlonof f:unclional Sjxkes by Means of Wavelets
-;' 289 .
"1
! $313 * t
C H ~ P T E K10 Grneraliza~~onr and T r e k s for d r t h o n u ~ n Wavelet l Baxr
\
-+:%ti
,,*
REFERENCES
<
35gz -+ SUBJECT INDEX 355
..kg+ 3 kfJTHOR INDEX
i
,
I
Introduction
Wavelets are a relatively recent alevelopment in applied mathematics. Their name itself was coined approximately a decade ago (Morlet, Arens, Fourgeau, and Giard (1982), Morlet (1983), Grossmann and Morlet (1984)); in the last ten years interest in them has grown at an explosive rate. There are several re+ sons for their present success. On the one hand, the concept of wavelets can be viewed as a synthesis of ideas which originated during the last twenty or thirty years in engineering (subband coding), physics (coherent states, renormalization group), and pure mathematics (study of Calderrjn-Zygmund operators). As a consequence of these interdisciplinary origins, wavelets appeal to scientists and engineers of many differentbackgrounds. On the other hand, wavelets are a fairly simple mathematical tool with a great, variety of possible applications. Already they have led to exciting appIications in signal analysis (sound, images) (some early references are Kronland-Martinet, &lorlet and Grossmann (1987), Mallat (1989b), (1989~);more recent references are given later) and numerical analysis (fast algorithms for integral transforas in Beylkin, Coifman, and Rokhlin (1991)); m k y other applications are being studied. This wide applicability also contributes to the interest they generate. \/ This book contains ten Iectures I delivered as the principal speaker at the CBMS conference on wavelets organized in June 1990 by the Mathematics Department at the University of Lowell, Massachusetts. According to the usual format o i the CBMS conferences, other speakers (G. Battle, G. Beylkin, C. Chui, A. .Cohen, R. Coifman, K. Grkhenig, J. Limdrat, S. MalIat, B. Torrbsani, and A. ljVillsky) provided lectures on their work related to wavelets. Moreover, three workshops were organized, on appiicatilins to physics and inverse problems (chaired by B. DeFacio), group theory arid harmonic analysis (H. Feichtiilger), and signal analysis (M. Vetterli). The audience consisted of researchers active in the field of wavelets as well as of mathematicians and other scientists and engjaeers who knew little about wavelets and hoped to learn more. This second ,group constituted the largest part of the audience. I saw it as my task to provide a tutorial on wavelets to this part of the audience, which would then be a solid grounding for more recent work exposed by the other lecturers and myself. Consquent!y, about two thirds of my lectures consisted of "basic wavelet theory," vii
'
.
viii
INTRODUCTION,
the other third being devotecl to more recent and unbublished work. This divislon is xeflected in the present write-up as well. As a result, I believe that this book be useful as an introduction to the subject, to be used either for individud reading, or for a seminar or graduate course. None of the other lectures or worl&hop papers presented at the CBMS conference haw been incorporated "8 As a result, this presentation is biased more toward my own work than the S conference was. In many instances I have included pointers to references her reading or a detailed exposition of particular applications, compleng the presknt text. Other books on wavelets published include Wuveiets me I;f.equency Methods (Combes, Grossmann, and Tcharnitchian (1987)), contains the proceedings of the International Wavelet Conference held in ille, France, in December 1987, Ondelettes, by Y. Meyer (1990) (in French; ish translation expected soon), which contains a mathematically more exed treatment than the present lectures, with fewer forays into other fields ver, Les Ondelettes en 1989, edited by P. G. LeniariC (1990), a collection of given at the Universitd Paris XI in the spring of 1989, and An Intwductzon avelets, by C. K. Chui (1992b), an introductior~from the approxirqation ory viewpoint. The pr~ceedingsof the International VIravelct Conference I n May 1989, held again in Marseille, are due to come out soon (Meyer (1992)) &oreover, many of the other contributors to the CBMS co~iference,a well as s(3pe wavelet researchers who could not attend, were invited f,o wnte an essay on their wavelet work, the result is the essay collection Wavelets and thew Applzcatzons (bskai et al. (1992)j) which can be considered a companion book to tl~isone. Ansther wavelet essay book is U'azleletu: A 2lrtoraal an Theory and Applzcutaons, edited by C. K. Chui (1992~);ir? addition, I know of several other wave!et essay books in preparztion (edited by J. Benedetto and M Frazicr, ano$her by M. Barlaud). as well as a monograph bj hl. Holschceider; thele was a*6pecial wavelet issue of IEEE %ns. Inform. Theorp in March of 1992; there @l be mother one, later in 1992,of &ms~ructrve Appmnmaizon Theory, and 7% in 1993, of IEEE %ns. Szgn. P m . In addition, several recent books inclgde claptels bn rawlets. Exarnp!es are MtiltzraCe Systems and Filter Banks by B. P. Vaidyanathaii (1992) and Quantum F/$yszcs, Relntzwty and Complec Sparztiqe: Towards a New Syntheszs by C. Kaiser (1990). Readers interested in the present lectures will find these books special issues useful for many details "qyFd other aspects not fully presented here. It is moreover clear that the subjgt i6 $till developing rapidly. T h S &ok more or less follows the path of my lectmes: each of the ten chap ters stands fs,~one of the ten lectures, presented in the order in which they were delivered,'lhe first chapter presents a quick overview of different aspects of the wavelet fransforrn. It sketches the outlines of a big fresco; subsequent chapters then fill in more detail. From there on, we proceed to the mntixluous wavelet transfokn (Chapter 2; with a short review of bandlimited functions and Shannon's theorem), to discrete but redundant wavelet trw~sforms(frames; Chapter 3) and to a general discwion of time-frequency density and the possible existence of orthonormal basis (Chapter 4). Many of the results in Chapters 2-4 can be formulated for the vindowed Fourier transform as well as the m e l e t
wq
INTRODUCTION
transform, and the two cases are presented in parallel, with analogies and differences pointed out as we go along. The remaining chapters all focus on orthonorma1 bases of wavelets. rnultiresolution analysis and a first general strategy for the construction of orthonormal wavelet bases (Chapter 5 ) ) orthonormal bases of compactly supported wavelets and their link to subband coding (Chapter 6), sharp regularity estimates for these wavelet bases (Chapter 7), symmetry for compactly supported wavelet bas& (Chapter 8) Chapter 9 shows that orthonorma1 bases are "good" bases for many functional spaces where Fourier methods are not well adapted. This chapter is the most mathematical of the whole book; most of its material is not connected t o the applications Biussed in other c h a p . ters, so that it can be skipped by readers-uninterested in this ppect of wavekt theory. I included it for several reasons: the kind of estimates usedin the proof are very important for'harmonic analysis, and similar (but more complicated) estimates in the proof of the "(1)"-theorem of David and JournC have turned ' out to be the groundwork for the applications to numerical analysis in the work of Beylkin, Coifrnan, and Rokhlin (1991) Moreover, the Calder6n-ZygmW theorem, explained in thls chapter, illustrates how techniques using d i f f e d t scales, one of the forerunners of wavelets, were used in harmonic analysis 1 before the advent of wavelets. Finalfy, Chapter 10 sketches several extensions of the constructio~sof orthononnd wavelet bases: to more than one dirnensh#, to dilation factors different from two (even nonirlteger), with the possibility &f better frequency localization, and to wavelet bases on a finite interval i n s t a d of the whole line. Every chapter concludes with a section of numbered "Notes," referred to in the text of the chapter by superscript nurhbers. These contain additional references, extra proofs excised to keep the text flowing, remarks, etc. -This book is a mathematics book: it states and proves many theorems. It also presupposes some mathematical background. In particular, I assume that the reader is familiar with the basic properties of the Fourier transform snd Fourier series. I also use some basic theorems of measure and integration theory (Fatou's lemma, dominated convergence theorem, Fubini's theorem; these c& be found in any good book on r e d analysis). In some chapters, familiarity with basic Hilbert space techniques is useful. A list of the basic notions and theorem used'in the book is given in the Preliminaries. The reader who finds that he or she does not know all of these prerequisites should not be dismayed, however; most of the book ean be followed wit5 just the ,basic notions of Fourier analysis. M6reover, I have tried t o keep a 6 %pedes~ trian pace in almost all tbe proofs, a t the risk of boring some'mthematically sophisticated read6rs. I hope therefore that these lecture not& twill interest people other than mathematicians. For this reason I have often shied away from the "Definition-Lemrna-Proposition-Theorem-Corollary" sequence, and I bave ttied t o be ilrtuitive in many places, even if this meant that the exposition became l e e succinct. I hope t o succeed in sharing with my readers some of the excitement that this interdisciplinary subject has brought into my scientific life. I want to take this opportunity t o express my gratitude to the many people who made the Lowell conference happen: the CBMS board, and thewathematics Department of the University of Lowell, in particular Professors G. Kaiser and
Qip
x
INTRODUCTION
M. B. Ruskai. The success of the conference, which unexpectedly turned out to have many more participants than customary for CBMS conferences, was due in large part to its very efficient organization. As experienced conference organizer I. M. James (1991) says, "every co&rence is mainly due to the efforts of a single individual who does almost all the work"; for the 1990 Wavelet CBMS conference, this individual was &ry &th Ruslcai. I am especially grateful to her for propotsing the conference in tbe first place, for organizing it in such a way that I had a minimal paperwork bad, while keeping me posted about all the developme;t?ts, and for generally being the organizational backbone. no small task. Prior to the conference I had the opportunity to teach much of this mathrid as a graduate course in the Mathehadies Department of thq University , Ann Arbor. My oneterm visit there was supported jointly by of ~ i G h i g m in a Visiting Profesorship for Women from the National Science Foundation, and by the University of Michigan. I would like to tW both institutions for their support. I would also like to thank all the facule a d students whp sat in on the course, and who provided feedback and useful suggestions. The manus~ript was typeset by Martina Sharp, who I thank Eor her patience and diligence, and for doing s wonderful job. I woulda't even have attempted to write this book without her. I am grateful to Jeff Lagarias for editorial comments. Several people heiped me spot typos in the galley proofs, and I lun grateful to dl of them; J would like to thank especially Pascal Auscher, Gerry Kaiser, Ming-3un Lai, and Martin Vetterli. All remaining mistakes are of course my responsiblity. I also would like to thank Jim Driscoll and Sharon Murrel for helping me prepare the author index. Finally, I want to thank my husband Etobert Calderbank for being extremely supportive md committed to our two-career-track with fainily, even though it occasionally means that he as wel! em I prove a few theorems less.
Ingrid Daubechies
ATtYT Bell Labomtones and Rutgers Universaty
I.
I
In this 8ec011dprinting several minor mistakes and many typographical errors have been corrected. T am grateful to everybody who helped me to spot them. I have also updated a few things: some of the previously unpublished references have appeared and some of the problems that were l i W as open have been wived. I have made no attempt t o include the many other interesting papers on wavelets that have appeared since the first printing; in any case, the list of references was not and is still not meant a I complete bibliography of the
subject. Ingrid Daubechies, Sept. 1992
Preliminaries and Notation
-
This prelimihary chapter fixes notation conventions and normalizatlqm. It also . states some basic theorems t b t will be used later in the boak. For those less familiar with Rilbert and Banach spaces, it contains a very brief (?%is primer should be used mainly as ru reference, to corfie back to in those instances when the reader comes across so& Hilbert or Banach space language that she or he is unfamiliar with. For most chapters, these concepts &re not used.) Let us start by some notation conventions. For x E W, we write lxj for the largest integer not exceeding x,
For example, 13/21 = 1, 1-3/21 = -2, 1-21 = -2, Similwly, {zl is the smallest integer which is larger than or equal to x. If a-0 (or oo), then we denote by O(a) any quantity that is bounded by a constant times a, by o(a) any quantity that tends to 0 (or oo) when a do= The end of a proof is always marked with a m; for clarity, many remarks or examples are ended with a o . In many pioofs, C denotes a "generic" constant, which need not ha* the same value throughout the proof. In chains of inequalities, I often use C1C',Clf1- - or C1, C2, C3,. . - to avoid confusion. We use the following convention for the Fourier transform (in one dimension):
With this normalization, one has
IIPIIL~= itrll~ lj(t)I 5 where
( 2 ~ ) - ' / ~11f B t l
9
xii
PRELIMINARIES AND NOTATION
Illversion of the Fourier transform is then given by
St,rictly &eaking, (0.0.1), (0.0.3) are well defined only if f , respectively 3f, are absolutely ixltegrabli.; for general L2-functions f , e.g., we should define Ff via a limiting p o c e k (see also below). We will implicitly assume t,llat the adequate limiting process is used in all cases, and write, with'a convenient abuse of notation, formulas similar to (0.0.1) and (0.0.3) even when a limiting process is understood. A standard property of the Fourier transform is:
$
with the notation I(') 7 f. If a function f is conlpactly supported, i-e.. f ( z ) = 0 if x < a or a: > 6, where -cx, < a < b < m. then its Fourier transform j(() is well defined also for cornplcx <,-and
If is mbreover infinitely differentiable, then the same argument can be applied to f (&I, leading to bounds on I<(' ~f(<)l. For a C" function f with support [a,b] thcre exist therefore constants so that the analytic extension of the Fourier transform of f satisfies
Conversely, any entire function which satisfies bounds of the type (0.0.4) for all N E N is tlie analytic extension of the Fourier transform of a Cm function with support in [a,b]. This is the Pdey-Wiener theorem. We will occasionally encounter (tempered) distributions. These are linear maps T from the set S(R) (consisting of all Cw functions that decay faster then any negative power (1 + to @, such that for all m, n E N,there exists C,,, for which I f I 5 $ 7 " SUP 10 + lxl)" f '"'(+)I x€
W
xiii
PRELIMINARIES AND NOTATION #
holds, for all f E S(W) The set of all such distributions is called S'(W). Any polynomially bounded function F can be interpreted as a distribution, with F (f ) 7 J dz F(z) f (x). Another example is the so-called "6-function" of Dirac, 6(f) =' f (0). A distribution T is said t o be supported in [a,b] if T ( f ) = 0 for all functions f: the support of whim has empty ~ntersectionwith [a,b]. One can define the Fourier transform 3 T or T of a disfribution T by ~ ( f=) ~ ( j(if) T is a function, then this coincides with our earlier definition). There exists a version of the Paley-Wiener theorem for distributions: an entire function T(() is the analytic extension of the Fourier transform of a distribution T in S'(R) supported in [a,b] if and only if, for some N E N,C N > 0,
Tlie only measure we will use is Lebesgue measure, on R and Rn. We will often denote the (Lebesgue) measure of S by IS(; in particular, ( [ a ,b]l = 6 - a (where b > a). , Well-known theorems from measure and integration theory which we wiil use include
Fatou's lemma. If f,, > 0, f,(x) -+ f(x) almost everywhere (n.e., the set of potnts where porntvnse convergence fazls has zero measure wath respect to Lebesgue measure), then
.
In parhcular, zf tkzs lim sup ts finste, then f zs zntegrable. (The limsup of a sequence is defined by limsup a, = nlim [sup {ak;k --r oo n-+m
2 n)j
;
has a every sequence, even if it does not have a limit (such as a, = limsup (which may be 00);for sequences that converge t o a limit, the limsup coincides with the limit.) Dominated convergence theorem. Suppose f , ( x ) + f (x) almost everywhere. If If,(x)J 4 g ( x ) for all n, and J dx g(x) < oo, then f is integmble,
Fubini's theorem. If $ dz[Jdy
If
(x,y)l] < oo, then
xiv
PRELIMINARlES AND NOTATION
i.e., the order of the rntegmtzons can be permuted.
In these three theorems the domain of integration can be any measurr subset of R (or It2 for F'ubini). When Hilbert spaces are used, they are usually denoted by IF1, u n l m t already have a name. We will foUow the mathematician's convention and scalar products which are linear in the first argument:
.
1
(All'*
+
X2u2,
v) = XI
( U I ,v ) 4- X z ( ~ 2 , v)
As usual. we have where B denotes the complex conjugate of a, and {u,u) 2 O for all , u E X. define the norm fluJJ of u by (0 11u1I2= (u,4 . In a Hilbert space, 111~11 = 0 implies u = 0, and all Cauchy sequences ( respect to (1 (1) have limits within the space. (Mote explicitly, if u, E 'H a llun - uml{becomes arbitrarily small if n, m are large enough-i.e., for all c there exists no, depending on c, so that llun - u,11 < c if n,rn 2 m-, then 1 exists u E H sothat the u, tend to u for n-roo, i.e., lin~,,, Ilu - u,l[ = A standard example of such a Hilbert space is L2(R), with
I
Here the integration runs from -oa to oo; we will often drop the integr bounds when the integral runs over the whole real line. Another example is P2(Z),the set of all square summable sequences of plex numbers indexed by integers, with
Again, we will often drop the limits on the summation index when wc over all integers. Both L2(1R)and t2(2)are infinite-dimensional Hilbert Even simpler are finite-dimensional Hilbert spaces,of which C ' is the sta example, with the scalar product
/
ck.
for u = (ul,..-,up), v = (ul,-.-,vk) E Hilbert spaces always have orthonormal bases, i.e., there exist fam~ vectors e,, in 31 (en, em) = &,m p
and
'
PRELIMINARIES AND NOTATION
XV
for all u E H. (We only consider separable Hilbert spaces, i.e., spaces in which orthonormal bases are countable.) Examples of orthonormal bases4me the Hermite functions in L2(R),the sequences e, defined by (e,), = 6,,3, ~ i t n, hj€Z in t2(Z) (i.e., all entries but the nth vanish), or the k vectors e l , . ,eh in ck defined by ( e r ) , = bt,rn, with 1 5 4, m. 5 k. (We use Kronecker's symbol b with the usual meaning: b*,, = 1 if 2 = j, 0 if i # j.) A standard inequality in a Hilbert space is the Cauchy-Schwarz inequality, 3'.
easily proved by writing (0.0.5) for appropriate linear combinations bf v and w. In particular, for f , g E L2(B),we have 7s'
*
,
and for c = (c+,fnEz, d = (dn)nEz E .t2@),
A consequence of (0 0.6) is
I I ~ I =I
SUP
v*
Il*ll5l
ICU,V)I=
SUP v,
IIvI1=l
1(21,2))1 a
(0.0.7)
/
"Operators" on 'H are linear maps from 'If t o another Hilbcrt space, often 3.1 ~tself.Explicitly, if A is an operator on 'HI then
An operator is continurn if Au - Av can be made arbitrarily small by making u - v small. Explicitly, for all c > 0 there should exist 6 (depending on c) so that llu - vl/ 5 6 implies IIAu - AvH < e . If we take v = 0, c = 1, then we find tfiat, for some b > 0, [Aull 5 1 if /lull < b. For any w E H we can define w' = IIAw'II 5 b-'IIwII. If llwfw; clearly ilw'll 5 b and therefore llAwll = IIAwll/llwll (w # 0)is bounded, then the operator A is called bounded. We have just seen that any continuous operator is bounded; the reverse is also true.. The norm IlAlf of A is defined by
'
It immediately follows that, for all u E H,
1 IAull 5 IIAII 1141 Operators-from H td C are called "linear functionals." For bounded linear functionals one has Riesz' representation theorem: for any t: H-tC, linear and
.
\ r
PREL,IMINARIES AND NOTATION
xvi boundrd, i.r', /t(u)l 5 C'llull - for all n H Ithere exists a unique vr E H so that e ( ~=) ( ~ ~ 7 ) ~ ) . An operator U from X I to Hz is an isometry if (Uv,U w ) = ( v , w ) for all v, w E 311, 11 is unitary if nloreover U'Ht = 7f2, i.e., every element vz E 7 - l ~can be written as v2 = Uv, for some vl E 'HI. If the en constitute an orthonormal basis ill f i l l and U is unitary, then the U e , constitute an orthonormal basis in 'Hz. The rpverse is atso true. any operator that rnaps an orthonormal basis to another orthonormal basis is un~tary. A sct D is c a l l a densc it1 31 if every u E H can be written as the limit of sonic sequence of u, in D. (One then says that the closure of D is all of 'H. The closure of a sgt S is obtained by adding to it all the v that can be obtained as 16ts of srqlicnccs in S.) If Ao is only defined for v E D, but we know that I
11 A ~ i l < f Cllvll
for a11 11 E
D,
(0.0 9)
then wo can extend A to 811 of 3-1 "by eontinuit,y." Explicitly: if u E H,find u, € D so that linl,,, un = 9 ~ .Then the un are rlecessarily a Cauchy sequence, and bccar~st.of (O.U.9), so are the Au,; the Au, have therefore a limit, which we call ATL(it does not deptmd on the particular sequence u, that was chosen). One can also deal with tlt~boundedoperators, i.e., A for which there exists no finite C sach that IfAuII Cllull holds for all u E. H. It is a fact of life that these: can usually only be defined on a dense set t)in 'H, and canndt be extended by the above trick (sinre they are not continuous). An example is in L2(R), where we can take D = C r (R) the set of all infinitely differentiable functions wlth compact support, for D. The dense set on which the operator is defined is called its tiornain. to a Hilbert 'I ne adjoint A* of a bounded operator A from a Hilbert space defined by space 'M2 (which may be 'HI itself) is the operator from 'H2 to
<
2
('1~1, A*u2) = (AW, ~
,
2 )
which should hold for all ul E 'HI, u2 E Hz. (The existence of A * is guaranteed by Riesz' representation theorem: for fixed 212, we can define a linear functional !on 7-f1 by t'(vl)= ( A u l ,uz). It is clearly bounded, and corresponds therefore v ) = d(u). It is easy t o check that the correspondence to a vector v so that (ul, a2+v is linear; this defines the operator A*.) One has i
(
I
I *I
=I
l l
11A* All = llA1I2 .
If A* = A (only possible if rl maps 7f to itself), then A is called self-adjoint. If a self-adjoint operator A satisfies (Au,u) 1 0 for all u E 31, then it isBcalleda positive operator; this is often denoted A > 0. We will write A > B if'A - B is a positive operator rer re-cl&s operators are special operators such that C, I (Aen,e,,)l is finite for all orthonormal bases in 31. For such a trace-class operator, C, (Ae,, en) is independent of the chosen orthonormal basis; we call this sum the trace of A,
PRELIMINARIES AND NOTATION
xvii
If A is positive, then it is sufficient to check whether En(Ae,,e,f is finite for only one orthonormal basis; if it is, then A is trace-class (This is not true for non-positive operators!) ' to itself consists of ali the The spectruma(A) of an operator A from H X E @ such that A - XId (Id stands for the Identity operator, I d u = u) does not havea bounded inverse. In a finite-dimensional Wilbert space, u ( A ) consists of the eigenvalues of A; in the infinite-dimensronal case, o ( A ) contairis aH the eigenvalues (constituting the point spectrum) but often contains other X as well, constituting the continuous spectrum. (For instance, in L2fW), multiplication of f (x) with sih T X has no point spectrum, but its continuous spect~umis (- 1,1] ) The spectrum of a self-adjoint operator consists of only real ~lurnbcrs;the spectrum of a positive operator contains on1y non- negative numbers. The spectral radius p(A) is defined by
It has the properties p(A)
< (IAl(
and p ( ~ =) Ilrn [ [ A ~ [ [ " ". n-m
Self-adjoint operators can be diagonalized. Thls is easiest to understand if their spectrum consists only of eigenvalues (as is the case in finite dimensions). One then h w u ( A ) = {A,; n E W } , with a corresponding orth~onormalfamily of e~genvectors, Ae, = Xnen .
It then follows that, for ail u E IH, Au =
(Au, e,)e,
=
(u, Aen)en =
h(u,
,
which is the "diagonalization" of A. (The spectral theorem permits us to generaIize this if part (or all) of the spectrum is continuous, but we will not need it in this book.) If two operators commute, i.e., ABu = B A u for all u E 'H, th6n they can be diagonalized simultaneously: there exists an orthonormal basis such that Ae, = crie, and Be, = One, . *
Many of these properties for bounded operators can also be formulated for unbouaded operators: adjoints, spectrum, diagonalization all exist for unbounded operators as well. One has t o be very careful with domains, however. For instance, generalizing the simultaneous diagonalization of commuting operators requires s careful definition of commuting operators: there exist pathological examples where A, 3 are both defined on a domain D, where AB and BA both make sense on D and are e q p l on D, but where A and B nevertheless are not
I
,
xviii
P R E L I M I N A ~ E SAND NOTATION
simultaneously diagondizable (because D was ehosen "too small"; see, and Simon (1971) for an ex'arnple). The proper definition of commuti~ bounded self-adjoint operators uses associated boupded operators: H commute if their associated unitary evolution operators commute. I adjoint operator H, the associated unitary evolution operators Ut are follows: for.any tl E D, the domain of H (beware: the domain of a sc operator is not just any dense set on which H is well defi~ed),UTv is th v ( t ) atetime t = T of the differential equation I
-
with initial condition v(0) = v . Banach spaces share many properties with but are more general th spaces, They are linear spaces equipped with a norm (which need E generally is not derived from a scalar product), complete with respr norm (i.e., all Cauchy sequences converge; see above). Some of thc we reviewed above for Hilbert spaces also exist in Ban& spaces; e.g operators, linear functionals, Spectrum and spectral radius. ' An ex1 Banach spwe that is not a Hilbert space is P(IP),the set of all functi~ such that 11fllLp (see(0.0.2)) is finite, with 1 5 p < m, p # 2. Anoth~ is Lm(R),the set d dl bounded functions on B,with I/ f = s supz The dual E* of a Banach space E is the set of all bounded linear on E; it is also a linear space, which comes with a natural norm (de (0.0.7)),with respect to which it is complete: E* is a Banach space it case of the p-spaces, 1 5 p c oo,it turns out that eiements of LQ,w q are related by p-' q-I = 1, d e f i ~ ebounded linear knctionals on 1 one has Hiilder's inequality, --
+
I/
de
f (4
5 (Ifllv
11i111~q
.
It turns out that all bounded linear functionals on LJ' are of this (Lp)* = Lq. In particular, L2 is its own dual; by N&Z' representatit (see above), every Hilbert space is its own dual. The adjoint A* of i A from El t o & is now an operator from Ei t o E l , defi~iedby
'
There exist different typea of bases in Banaeh spaces. (We will consider separable spaces, in which bases are countable.) The tute a Schauckr basis if, for all v E E, there exist unique jLn E N v = l i m ~ + , En=, /.inen (i.e., (Iv /.&,,&,)l-d as N-too). The requirement of the pn forces the e, t o be linearly independent, in tb po e,, can be in the closure of the linear span of all the others, i.e., t h y,, so that + = iirnN,, + ,,, Tme,,,.In r Schauder basis, 4 of the en may be important. A basis is called unconditional if in satisfies one of the following two equivalent properties:
-~ t = ,
x&,,
CHAPTER
1
1
The What, Why, and How of Wavelets
The wavelet transform is a tool tllat cuts up data IX functiolis or 'opibrators into different fruq~lelrcyco~nponenb,and thtt~rstuciies cnctl conll)oncr~twith a resolutiori nratclrcd to its wale. Forerurr~iersof this techl~iquewvrc ilivcntc(1inctepeader~tlyin pure a ria the ma tics (C:alderbn's r t w l u t i ~ xof~ tlic i(l(ntrtity in .hiumonic analysis see e g.: Caidcr6n (1964)), plrysics (coltcreut stirt.c.s for ttlc (ax b)group in quantum rnt~cjianics,first coristnrctctl by Aslakscr~allti Kliiucfcr (f9fi8), and linked to t l ~ chydrogen ato111Ifi~n~iltoiiian by Paul (1985)) irlrrl c~nginccring (QMF filters by FAstelallarid Gallarrd (197'7), and later QMF filtcra with cxact ' recorlstructior~property I)y -Smith aritf ~ ~ a r n w e(198G), ll Vt~.ttcrli(1!186) in i:lcctrical enginceri~~g; wavelcts were proposcd for tlrc* analysis of scisnrii. ci;rta by J . Morlet (T983)). The last.five years biiv~.scqu a synth(*sisl)c+twcv~~~ a11 t h c ~ diffc:rcnt approaches, which has been vcry fertile for dl tl~c*f i c b t d s r.orrc*t?rrlcvI. Let us .stay for a moment within the s i p h i~tlaly~is fr;~~ln>work. (Thc~tliscussiori can easily be translatad to other fields,) Tbe w;ivcllct tri~~rsforrrr of a signal evolving in time (e.g., the amplitude of the prcsrlrc on :rrl c*iirdr~~rn, for acoustical applications) depends on two variables: scak (or irrqucrlcy) ant1 ti~nc; wavelets provide a tool for tirne-frcque~~cy tocalizat&r~.Tlto first sclction tells us what time-frequency localization means and why it is of-ir~tcrt:st.Tltc rcniailiing sectiohs describe different types of wavelets.
+
1.1.
Time-frequency localization.
In m-my applications, gigen a signal f(t) (for the rnorni:~rt,wcb ;~sstlmethat t is a continuaus variable), one is interested in its frQuency rulrtcnt locally in ti&?. Thishissimilar to music notation, for example, which tells the player which
',
.
notes (= f r ~ u e n c .information) y t o play at any gIven smntent. The standard Fourier transform, *.
also gives a representation of: the frequency content of f , but information concerning time-localization of, c.g., high frequency bursts cannot be read off easily from 3f. Time-localization can be achieved by first windowing the signal f , so
2
CHAPTER 1
as tb cut off only a well-localized slice o f f , and then taking its Fourier transform:
This is the windowed Fourier transform, which is a standard technique for timefrequency local~zation.'It is even more familiar to signal analysts in its discrete version, where t and w arc assigned r ~ g u l k spaced l~ values: t = nto, w = m q , whcre m. n range over 2, and u",to > 0 are fixed. Then (1.1.1) becomes
This procedure is sdicmatically reprcwnfc~din Figure 1.1: for fixed n, the Tzl",(f correspond to the Fourier ccc~ficicntsof f ( )g(. - nto). If, for instance, g i s compactly strpportcci, then it is dcar that, with 'appropriately chosen wo, the Fourier coefficients ar,c sufficicnt to characterize and, if need be, to reconstruct f ( - ) g ( -- ato). Clranging n a~xrountsto shifting the "slices" by) steps of to and its multiples, allowing the rt3covery of all of f from the j). (We wili discuss this in Inore mathrn~aticaldetail in Chapter 3.) Many possible choices t~avebrcn proposed for the window funpiion g in signal analysis, most of which have compact slppurt a11d reasonable smoothness. Jn physks, (1.1.I) is related to coherent state representations; the gW*'(s) = etWag(s- $1 are the s to the Weyl -Heisenberg group (see,e.g., Klauder and coherent s t a t ~ associated Skagerstam (1985)). In this cor~text,a very popular choice is a Gaussian g. In all applications, g is suppawxi to be well concentrated in both time and frequency; if g and are both c o n c e n t r ~ daround zero, then ( T ~f )~( c j", t ) can be interpreted loosely a s the "content" off near time t and near frequency w. The windowed Fourier transform provides thus a description of f in the timefrequency plane.
yz(f)
cfE(
-. RG.1.1 The wandowed Founer trarufom:
the finctton f ( 2 ) is multipiitd with the unndow finctson g ( t ) , and the Founer c a e 5 e n t J or the @uct f (t)g(t) a= m W d ; the p m d ~ * ts then =peat& for tmnskited veraons of fhe d m , g(t t o ) , g(t Zto),
-
-
. ..
3
THF WHAT, WHY, AND HOW OF WAVELETS
The wavelet transform: Analogies and differences with the windowed Fourier transform.
1.2.
The wavelet transform provides a similar time-frequency description, with a few ~mportantdierences. The wavelet transform formulas analogous t o (1.1.1) and (1.1.2) are
%
and
8
-
.6
?E;
q:(f 1 = a0 p-
1
%'
/ dt f ( t )$(aimt - n b ) .
J
In b @ h h mwe assume thatJ! ,? satisfies
-
(for reasons explained in Chapters 2 and 3). Formula (1.2 2) is agmn obtained from (1.2.1) by restricting a, b to only d i s crete values: a = b = nboar in this case, with m , n ranging over Z, and Q > 1, > 0 fixed. One similarity between the wavelet and windowed Fourier transforms is clear: both (1.1.1) and (1.2.1) take the inner products of f with a family of functions indexed by two labels, e d ( s ) = ew*g(s - t) in (1.1.1), and @'sb(s) = 1a(-1f2jt(yb) in (1.2 1). The functions + O f b are called "aravelets"; the function $J is sometimes called "mother wavelet." (Note that and g are implicitly assumed to be real, even though this is by no means essential; if they are not, then complex conjugates have to be introduced in (1.1.1), (1.2. I).) A typical choice for 3 is $(t) = (1 - t2) exp(-t2/2), &hesecond derivative of the Gaussian, sometimes called the mexlcan hat function because it resembles a cross section of a Mexican hat. The mexican hat function is well lwalized in both time iand frequency, and satisfies (1.2.3). As a changes, the q F O ( s ) = lal-l/'$(s/a) I mver different frequency ranges (large values of the scaling parameter la\ corto smaH frequencies, or large scale .~cP*O; small values of la1 correspond .)Or'@ Changing the parameter b as well frequencies or very fine scale to move the time localization center: each $Plb(s) is lacalized around follows that (1.2.I), like (1.1.I), provides a time-frequency descripti~n difference between the wavelet and windowed Fourier transforms lies of the analyzing functions g Y s t and @asb,as shown in Figure 1.2. all consist of the same envelope function g, translated to the ion, and "filled in" with higher frequency oscillations. All the the value af w, have the same width. In contrast, the y F b have to their frequency: high frequency +a*b are very narrow, are much broader, As-aresult, the wavelet transform is better able tbt4he windowed Fourier transform to "zoom in" on very short lived high frequency phenomena, such as transients in signals (or singularities
ar,
+
.
$Jlt
CHAPTER
1
(b)
FIG I 2 Tkptcal shapes of (a) unndowu! E j u n e r tmnsform fvnctaans g W a t , and The P i ( x ) = ~-'~'g(x - t ) can be mewed as tmnrilated envelopes 3. (b) wavelets m" unth hzgker ff~puenctes the are ull copzes of the same ,%nctaons, translated
+".'
and compressed or s t r ~ t c h e d
in functions or integral kernels) This is illustrated by Figure 1 3, which shows windowed Four~ertransforms and the wavelet transform of the Same *a1 f defined by
I(t) = s i n ( 2 r y t )
.
+ sin(Z.rrv2t) + y i6(t - t l ) + li(t - t2)] .
In practice, thls signal is not give11by this continuous expression, but by samples, and adding a 6-function is then approximated by adding a constant to one sample only. In sampled vcrsion, we have then
.
For the example in Figure t.3a, ul = 500 Hz,vz = 1 kHz, 7 = 1/8,000 sec (i.e., we have 8,COO samples per second), a = 1.5, and n2 - nl = 32 (corresponding to 4 milliseconds betweeh the two pulses). The three spectrograms (graphs of
nq 6 ~ a w a - 6 ~ ov ja ~ y apap0 y u q o? am ppom I .~-5mny ??fiom pun S U L - 3 4 s a a . ~ a w~ uaarnlg uo?Jnlos;uR x m & u j q Jo uowdtuo3 (p) .(l,o 08 rpuodsauor, a?wrpco aw ' 3 3 ) ~ L Z DA3uanbaJ ~ o a u q u pun 'poylatu tanat RarG auvc ay$ ggn
ay)
'l(/),,O~I m o l d 0 b - p afi-q
(q) y w u o m ~ a~q m ayour ~ oj
-1lo w o @ w
? ~ P ~ U A I(3) qanq k 6 aavtpauuaaua 'apym = cuaz ?f3u19 = sanpn y&y) qaaat M6uy-n '(yd& av uo pabaptw lou n omyd ay$) panotd m I ( j ) , , d I fip :rulac6aymi~palpaso am a . r u mppm niopuam am
-auvld (a~uu!pro)m' ( v c m c q v ) ? ow ut
(I(~),,,J;I
W
02 p u o a ~ ~ y J o d a rpau6ar.m d ao
ruarag~pa;u~ rllm 5 l o ~ o J ~ wuu n~o , p?a m o p u ? ~(q) '(7)J p*c
0009 OOOE 0001 OOOC
r"-1
a y (~a ) -SI 31J
OOOP 000s
0
oooz 0031 0,
o " l
1-2 ;
i
OOOE
COSE
I
OOOP (3)
0
[,*! 1 .; ;,I: ;,,,
-
Om& OOOP
E
00s 000 1 00s 1 0002 OOSZ OOOE OOSE OOOP
SLBT3AVM 60 MOH CINV 'AHM ',LVHM 3HL
0 005 000 1 00s 1
oooz 00% OOOC OEC
OOOP
6 ,
CHAPTER 1
the modulus of the windowed Fourier transform) in Figure 1.3b use standard Hamming windows, with widths 12.8, 6.4,, and 3.2 milliseconds, respectively. (Time t varies horizontdly, frequency w vertically, on these plots; the grey levels indicate the value of ( F i n ( f ) l ,with black standing for the highest value.) As the window width increases, the resolution of the two pure tones gets better, but it becomes harder or even impossible to resolve the two pulses. Figure 1 . 3 ~ shows the modulus of the wavelet transform of f computed by means of the (complex) Morlet wavelet $ ( t ) = ~ e - ~ ~ ~ ~ ~ e-u2a2/4), ( e ' " ' with a = 4. (To make comparison with the spectrograms easier, a linear frequency axis has been used here; for wavelet transforms, a logbithmic frequency axis is more usual.) One already sees that the two impulses are resolved even better than with the 3.2 msec Hamming windoiv (right in Figure 1-3b), while t h e frequency resolution for the two pure tones is comparable with that obtained with the 6.4 msec Hamming window (middle in Figure 1.3b). This comparison of frequency reqlutions is illustrated mare clearly by Figure 1.3d: here sections of the spectrograms (i.e., plots bf I(Fi" f)(., t)l with fixed t) and of the wavelet transform modulus (1 (T""f )(-,b)l with fixed b) are compared. The dynamic range (ratio between the maxima and the "dipn between the two peaks) of the wavelet transform is comparable to that of the 6.4 msec spectrogram. '(Note that the flat horizontal "tail" for the wavelet transform in the graphs in Figure 1.3d is an artifact of the plotting package used, which set a rather high cut-off, a s compared with the spectrogram plots; anyway, this cut-off is already a t -24 dB.) In fact, our,ear uses a wavelet transform when analyzing sound, at least in the very first stage, The pressure amplitude oscillations are transmitted from the eardrum to the basilar membrane, which extends over tire whole length of the cochlea. The cochlea is rolled up as a spiral inside our inner ear; imagine it unrolled to a straight segment, so that the basilar membrane is also stretched out. We can then introduce a coordinate y along this segment. Experiment sad numerical simulation show that a pressnre wave which is a pure tone, f , ( f ) = eaWt, leads to a response excitation along the basilar membrane which has the s m e frequency in time, but with an envelope in y, Fu(t, y) = eWt &(y)- Iq a first approximation, which turns out to be pretty good for frequencies w above 500 Hz, the dependence on w of #,(y) corresponds to a shift by log w : there exists one function 4 so that 4, (y ) is very close to #( y -log w ). For a general excitation function f , f (t) = i;& j(w)eYt, it follows that the response function F(t,V ) is given by the corresponding superposition of "elementary response fuactions,"
If we now introduce a change of parameterization, by d e h h g
4 )= (
2
)
4 ,
G(a,t)= F(t,loga) ,
,f
T H E WHAT, WHY, A N D HOW OF WAVELETS P
then it &,$lows that t"'^,
G(a,t ) = 3:
$
J
dt' j(t') $ ( a ( t - t ' ) ) ,
which (up t o normalization) is exactly a wavelet transform. The dilation parameter comesin, of course, because of the logarithmic shifts in frequency in the &. The occurrence of the wavelet transform in the first stage of our own biological
:i
+
mustical analysis suggests that wavelet-based methods for acousticai analysis
.' haw a better chance than other rricthods to lead, e.g., to compression schemes 0 '
2..
"
yv$-?~detectable by our ear. ,
" ~ 2 ~ Different types of wavelet <
'
f -1,
e
*
3.f
transform.
*
%re exist many differ~~lt types of wavelet transfork, all starting from the ~ e " , b r r n u l a(1.2.1), s (1 2.2). In these rlotes we will distinguish between 11
.x.PP
witbib tile diwrt.tc wav~:lct't,raIlsfnnIIwe distinguish further between
BI. W u n d a n t discrcte systerns (franres) and
BZ. Ortl~onormal(and other) haws of wavelets. 1.31 The continuous wavelet transform. H ~ Rthe dilation and brans-. lation p&<metersa,h vary continuutivly over W (with the cowrai'nt a # a). The &&elt transforni is given by fonnula (1.2.1); a function can be reconstructed bhits wavelet transform by Incans of t h "resolution of &lentityn formula
(y
$ ), R I L ~( , ) denotes the L2-inner product. The v n b ( x )= CJ,depends only on $ and is giver1 by
trme CJ,< oo (otherwise (1.3.1) docs not make sense). If $Jis in L1(PI) the case in all examples of practical interest), then is continuous, so can be finite only if &(o) = 0, i s . , *(z) = 0. A proof for (1.3.1) ven in Chapter 2. (Note that we have implicitly assumed that $ is real; lex $, we should use instead of II,in (1.2.1). In some applications, lex t,b are useful.) la (1.3.1) can be viewed in two different ways: ( 1 ) as a way of re-cting f once its wavelet transform Fa" f is known, or (2) as a way to
I&
4
4
8
C
A
I 1 ~
I
write f as a srlporpositiorr of wnvckkts vsb; thc. c.ocalficicnts in this supcrpositior~ are exactly given I)y lho wztvcllc$t trarisfornl of j. Iloth pointq of view lead to intcrt?sting wpplicat.ions. The corresponclc~~~c-c! j(s) -+ ! ) ( ( I , L) rry)rcssc~~rt~s :i OII<:-variable filr~ction by u fut~ctionof two viiriiiI)l(!s, into wtri(.11 lots cd ~orr(!liitiolls arc \)uilt in (see Chaptcr 2). This r ~ d l ~ ~ l t l iofi ~t.l~c! l ~ yrc*i)rtmscbr~t,at i o ~ <:an t I)c. c~xploitc*rly;i brziutiful application is tllc cor~cl:ptof' t.11(. "skc~lc~t.oi~" of ;I sijiud, c*xtrit<.t.c!c! frorn ttlc continuous wavck11, t r i t ~ ~ ~ f o rwhit.11 ~ t i , (.an IN. 11sc.tl for ~)01liili(~ilr filleriilg (see, e.g., TorrCsarii (199 I), I)t.ljcritl cst irl. ( L!)Y'L)).
1.3.2. The d i s c r e t e but r c d u r ~ d a n twavelot transform-frames. In this case the dilatioii j)i~;ilnt.ti?r(1 i ~ l r Ic thi* trii~~sli~tiorl pi~riiltitbt~~t tmth ti~kc!O I I ~ Y discrete values. For rr wo r-floosc*rile iuttbgcbr (posit ivc* ;ul(l trcgtuivr!) powc!rs of one fixed dililtio~l~ ) i i ~ a l ~ t(LO f ~ --' t ~ 1, T i . v , (L ( I ; ; ' . AS cil~t'i~iyillustratetl ~. by Figun: 1.2, diffc~rc*tttvii111t~s of rrL c~orrc~si)t)~l(l to w;~vc~I(*t:, of tliffc*rcwt wi(1ths. It follows thitt t h t a tlisc.rt~tixirtion of t 11c. t.r;~~rslirtiotl ~xu.;utlc*tclr h sllo~tltidt:pt~ndon r n : r~arrow(lgigh frt'clllc~lw*y)w;tvolcts arc1 triiiwlrtCt~clI).y s~lr;tllsto~):, i l l ortlcr to covcr tlic wlrolc t i r ~ l c br;rllgil, w tlilc wiilcr (lower frc*clrto~tc.y) v.;ivclt~ts;it.('tr;ulslatcd by larger st.cp.s. Si~lc-c* f Ilc* wiclttl of d,(rc,, '"s)is ~)rol)ortiorl;~l to 1 ~ ; ; ' . w(*ch~ose tlicn!foro to cliscrct in, Ir 1)y b - titr,)(t:i1, w hi*rv / I , , 0 is liu.(l, i ~ t l ~11l <. Z. , Thv corresj,onding cfiscrc~tcblyiiil)c~ll(vlwi~vc:li.~s ;rrcl t.tic~rt.ft~ro - v
11
dl,,, ,,r
(.r
1
"A':
= 'Lo -- (to
m/2
(J(4t,,
1
,
.
{x
T L I , , , ~ )~); ; ~ IL~),))
(1.3.3)
Figlirc 1 . h shgws ~('ht!lllat,i('lillyt t ~ (li~&tii'o & of ti~llt~-fr(*(til~~rl(~y ~n~iili%irtioll c:c*tltcrs correspondi~lgtcr tlre cji,,,,,,. For a giver1 f ~ i t ~ c t i I, oi~ t.11~ i1111c~r pro(ll~'ts(f,@,;,,,,) then give exac-tly tht* clisc-rctc wi~vclt:t trruisft)rr~~ 'I;:: ( j) iL\ dc~fi~~o(l in ( 1:2.2) (we w u n l c agaiit thAt I/) is real). 111 t11(? dibcriatt?c&siB,tllc!rt? cfot?s not, cxist, i l t ~ I ' I I C M ~ . it b L r t ' ? i ~ ~of l ~ the ~ti~i~ identity" for~ni~la (zrlah)goi~sto ( 1.3.1) for thcb c.o~~t i rl~lons(.iLw.Rcu:onstruction of f from Tw"''(f j, if at i~llp c ~ st)lrb, i r~lustt hc-rt-fort.I)(. clo~rclI b v sornc othr:r rnems. Ttic followir~gqliestiolls nat.urally irrise:
(2) 1s it possible to rccoitstruet j in a o~~rncrici~lly st.;rl)lt*way fr+m TWib" (f)'? These questio~~s cor~cerrlthe rec.ovc1ry of f fro111 w
(1') Can any function Le written as a supcrpositio~~ of kt,'?
(2') Is there a ~lurnericallystable algorithn~to co~uputethe coefficients for such an expansion? Chapter 3 addresses these clncstions. A s in the c o ~ ~ t i r ~ l ~case, o u s these discrete wavelet transforms often provicic a very reduatiatlt description of the original
THE WHAT, WHY, AND HOW OF WAVELETS
$
is"&-
--
....
I . . . . . .
1.4. The lattrce~of tsme-jkquency Localwtwn for the wavelet t m f o r m and wan&unufonn. (a) The wavelet tnamfonnr &,,* aa lo& amund sn time. (thwu & w e , eg., for the weam@ that 141b h o w a n & g w q , at M list -t $ ( t ) = (1 - ~ ) e - ' ' / ~ ) ; 14,,(01 fhM peaks at *arbf W h d tzm the windowed FOUWk p ~ f ~ r m9m,n : ~8 @ m t e + s of &,n #r jhp%iq. (4) w m m u u l n t o antrmc,olp~ndsnjhqpmtq.
car*
10
CHAPTER 1
function. This redundancy can be exploited (it is, for instance, poesible to cmmwkle still obtaining raconstmtion of f with good precision), or eliminated to reduce the transform to jts bare essentials (such as in the i m e e compression work of W a t and Zhong (1992)).It . is in this discrete form that the wavelet transform is cloeest to the "~dransfarn"' of Fkazier and Jawerth (1988). The choice of the wavelet $ used in the continuoh waveiet transform or in ' frames of discretely labelled families of wavelets is esseatially only restricted by the requirement that C+,as defined hy (1.3.2), is finite. For practical reasom, one usually tho- 9 so that it is well concentrated in both the time and the frequency domain, but this still lebw a lot of freedom. In the oext section we will see how giving up most of this freedom allows us to build ort$cmormal baees of wavelets.
pute the wavelet transform only app-tdy,
1.3. Orthom~malw a d h :Multir-htiw analysis. For p m e very special choices of q5 agd q,b, the &,, cowtitute an orthonoma1 basis for L2(R). En partidat, if we chase cro = 2 , 4 = 1: then there e& $, with &od time-frequency localization properties, such that the
constitute an orthonormal basis for t2(~). (Forthe time being, and until Chap' ter 10, we restrict ourselves to 60 = 2.) The oldest m p l e of a h c t i o n for which the t,!~~,, defined by (1.3.4) constitute an orthonormai baais for L~(R) is the - H wfunction,
+
The Baar basis Em been lrnown sirwe Haar (19i0). Note that the Ha& function does not hsw goad tikfmquemy localization: its Fourier transform q((f decays like 1[1-"0r oo. Nevertheless we will use it hele for fllustktioa pwrpoees. What f o l l m b a p m f that the Haat family does indeed c o d U C e . an orthonormal b*aai. This proof b dierent from the one in most t w in fact, it will use multireaolution analysis as a tool. In order to prove that the q!~,,,,,(x) constitute an otthonormal lasirr,we need to eefablish that
<
-(
(1) the $Jm,- are ~rthononnal;~
(2) any La-function f. can be spproxhkhd, up to arbitratily a by a finite linear combiiation of the $J~,,.
d
precieion,
Ortbmomality is easy to establish. Since support (h+)* [ P n ,P ( n + 111, it blfgpps that two Hsat wsve1ete of the same scale (sthdvalue of m) never Overhpphg suppdrta paesibfe if the averlap, &Wtb8t ($m,n, tlm,rn#) = two wavelde have different sivmt, ea in Figure 1.5. It is eaey to check, hctr~evmr, e t if rn < mr, then support ($,,,,*) Iies wholly within a region where t,bm+,n*is
11
THE WHAT, WHY, AND HOW OF WAVELETS
mnstwit (as on the figure). It follows that tbe inner prodact of $m,n and $ J ~ I , , ~
ie thea proportional to the integral of JI itself, which is m.
L-A
-Ptc
f 5 Two Har wavekid, the support of the Ihonowcr" wavelet tn an tnLmral d e n the YMdrr* umuekt u cartant
, *?
u wmpktely can-
iq-fir
'
f rC
*-%%%$l_
Wmncentrate now on how well an arbitrary function f caa be appraorimated by lmrarrr combinations of Haar wavelets. Any f ia L ~ ( R can ) be arbitrarily well ,apprd.mated by a function with compact support which is p conrrtant on the @-J,(t+ 1)2-I[ (it suffices to take the support and j lage enough). We cqn tbarabnM r i c t ourselves to such piecewise constant functions only: mume f to be supported on [12Ja,2J~],and to be piemwise canstant on the [C2-Jo, ( L 1)2-*[, where JI and Jo can both be arbitrarily large (see Figure 1.6). Let p denote the most- d u e of p = f m [ W J n , (t 1)2-*I by ff . We now nt p as a sum of two pieces, jO = f1 4, where f l is an appraximation whicb is piecewise constant wer intervals Mce 8s large as o r i g i d l y , i,e., + I , ( ~ + ~ ) ~ - J O + I I = cmstunt = The values f: are given by the averhe two corresponding constant d u e s fw f', f: = $ (& (see The function 6 ' is pkewise constant with the same stepwidth aa & J iately bas
+
+
fir
fai
vL
$;Gptd
+
+
= $t+l - f: = i(pY+,'- &)= 6: is a linear combination of Bcaled and translated Haar functions: $+I
CHAPTER 1
/
BLOW UP
FIG 1 6 ( a ) A functron f unth mppmt 1-2'1, 2J1],preoewrse constunt on the [ k 2 - J ~ , (k + 1 ) 2 - ~ o [ ( b ) A blowup of a portron of j On every paw of mteruds, j rs repfoccd by t t s average ( 4f I); the dgemnce between f and f u 6l, a lanurr wndmdaon of H a m wavelets.
where f' is of the same type as fO, but with stepwidth &ice as lare. We can apply the same trick to f so that
',
with f a still supported on [-2J1, 2 J j,~ but piecewise constant on the even larger intervals [k2- J0+2, (k 1)2- J o + 2 [ . We can keep going like this; until we have
+
Here f
consists of two constant pieces (see Figure 1.7), with equal to the average of f ova [ 0 , 2 J 1 ( , aod f JO+JI J f J O + J 1 Il-2~l,ol S the awage of f over (-2", 0(. Even though'we have "filled out'' the whole support of f , we can still lreep going with our averaging trick: nothing stops us from widening our horizon from
/oJD+J1 f2+h
THE WHAT, WHY, A N D HOW OF.WAVELETS
i
FkG. 1.7. The wenrges off on (0, 2J1] and (-2'1, 01 a n be " s m d " out over the bigger htewcJs 10, 2 J l + 1 j , [ - 2 J 1 + 1 , 01, the dadrr o linear combanatton of vny stretched out
to z J l + ' , and writing f J l + J z = f J l + J 1 + 1 + d J 1 + J 2 + 1 , where f J1+Ja+l I[o,2Jl+l[
f f$+J2,
fJsiJa+l
- q1 f -Jll + J a
~ [ - ~ J I + ' , O= [
and *+ s I* q
.'
= ff $ + h - J 1 - l x - 1f
q5(2-h-'z + 1) (aee Figure 1.7). This can again be repeated, leading to gJl+h
-'
Ie
Jl+Ja
911ti
f
which
.
Wediately that
" ZZX
.
'
kl'&$e arbitrarily small by tathy tdliciently large K. As claimed,
f a~theref- i;b'ip~mdmatedto arbitrary precision by a finite linear combination of Haar wa-l The argume!nt we saw has implicitly used a "rnultiresolution"approach: we have written sucxessive coarser and coarser approximations to f (the f J ,
14
CHAPTER
1
averaging f over larger and larger intends), and at every'step we have written the difference between the approximation with resolution ZJ-', and the next In fact, we coarser level, with resolution 2', as a linear combination of the have introduced a ladder of spaces (V,),EZ representing the successive resolution levels: in this particular case, V, = (f E L2(It); f piecewise constant on the [23k,2J (k + I)[, k E Z). These spaces have the folIowing properties:
-
(1)
~~-cv~cv~c~cv-*cv-2c-~.;
(2)
n, v, = n, u,,,~, = L ~ ( R ) ;
(4)
f E Vo
+
f(.- n) E
for all n E 2.
Property 3 expresses that all the spaces are scaled -versions of one space (the L'multiresalution"aspect). In the Haar example we found then that there exists a function so that
+
The beauty of the rnultiresolution approach is that whenever a ladder of spaces t; satisfies the four properties a b m , together with ( 5 ) 34 E Vo so that the &,n(~)= $(x basis for V&,
- n) constitute an orthonormal
then there exists @ so that (1-3.5) holds. (In the Haar example above, we can take 4(x) = 1 if 0 5 x < 1, $(z) = 0 otherwise.) The $2,k constitute sutomatically an orthonormal basis. It turns out that $here are many examples of such "rnultiresolutio~d y s i s ladders," corresponding to many exampl& of orthonormal wavelet -bases. There exists an explicit recipe for the = $(22 - n) construction of $: since 9 E Vo C VdI,and the 4-1,~(z) constitute an orthonormal basis for V-1 (by (3) and (5) above), there exist a, @(2x - n). It then suffices to take (4,qL1,,) so that d(z) = a, = @(x) = C,(-l)na-,+l4421 -n). The function 4 is called a scalsng functaon of the mdtiresolution analysis. The comspondence multiresolubion analysis -+ orthonormal basis of wavelets will be explained in detail in Chapter 5, and further explored in subsequent chapters. This mdtiresolution appro& is also Linked with subband filtering, as explained in $5.6 (Chapter 5). Figure 1.8 shows some examples of pairs of functions 4, $ corresponding to different multiresdution analyses which we will encounter in later chapters. The Meyer wavelets (Chapters 4 and 5) bave compactly supported Fourier transform; 4 and 9 hhemselves are infinitely supported; they are shown in Figure t .8a. The Battle-Lemarih wavelets (Chapter 5) we spline functions (linear ifi Figure 1.8b, cubic in Figure I.&), with knots at Z for 4, at )Z for $. Both ) and have infinite support, and exponential decay; their numerical decay is faster than for the Meyer wavelets (for mmpwimn, the horizontal scale is the same in (a),
4
+
16
CHAPTER 1
(b), and (c) of Figure 1.8). The Haar wavelet, in Figure 1.86, has been known sirlce 1910. It can be viewed as the smallest degree Battle-Lemarid wavelet j$waar= $ B L , ~ )or also as the first of a family of compactly supported wavelets constructed in Chapter 6, haar = I@- Figure 1.843 plots the next member of the family of compactly supported wavelets N$J; 24 and 2$ both have support width 3, and are continuous. in this family of N+ (constructed in §6.4), the regularity increases linearly with the support width (Chapter 7). Finally, Flgure 1.8f shows another compactly supported wavelet, with support width 11; and less asymmetry (see Chapter 8). Notes.
1. *Thereexist other techniques for time-frequency localization than the windowed Fourier transform. A well-known example is the Wigner distribution. (See, e.g., Boashash (l!MJ) for a good review on the use of the Wigner distribution for signal analysis.) The advantage of the Wigner distribution is that, unlike the windowed Fourier transform or the wavelet transform, it does not introduce a reference function (such as the window function, or the wavelet) against which the signal has t o be integrated. The disadvantage is that the signal enters in the Wigner distribution in a quadratic rather than linear way, which is the cause of many interference phenomena. These may be useful in some applications, especially for, e-g., signals which have a very short time duration (an example is Janse and Kaiser (1983); Boashash (1990) contains references t o many more examples); for signals which last for a longer time, they make the Wigner distribution less attractive. Randrin (1989) shows how the absolute d u e s of both the windowed Fourier transform d the waveiet transform of a function can dso be obtained by "smoothing" its Wigner distribution in an appropriate way; the phase information is lost Lr this p r o c h however,.wd reconstruction is not possible any more.
2. The restriction bo = 1, mrreqxmding to (1.3-4), is not very serious: if (1.3.4) provides an o r t h o n o d basii, then so do the &,,,(x) = 2-=f2 , # 0 is arbi4(2-'z - nbo), with S(z) = Ibl-1/2$(bg1z), where trary. The choice a0 = 2 cannot be modifid by scaling, asd in fact cannot be chosen arbitrarily. The general construction of orthbnormal bases we will expose here can be made to work for all rational choices for at-, > 1, as shown in A d e r (1989), but the choice = 2 is the simplest. Different choices for a correspond of course to different q5. Although the constructive method for orthoMmnal wavelet bases, called multiresdutisn is rational, it is an open question whether analysis, can work only if there exist o r t h o n o d wavelet bases (necessarily not associsted with a ' multiresolution analysis), with good time-frequency Iodbtion, and with irrational a ~ .
CHAPTER
2
The Continuous Wavelet Transform
images of Lx-functions under the continuous wavelet transform constitute a ing kernel Hilbert space (r.k.H.s.). Such r.k.H.s.'s occur and are useful different contexts. One of the simplest examples IS the space of all -limited functkons, discussed in $321 and 2.2. In 52.3 we introduce the @Wept of band and time limiting; of c o u m no nonzero function can be strictly 2 -.: w e l i m i t e d (i.e., f (t) r 0 for t outside I-T,q) and band-limited (j(<) = 0 for ! a * &$[-S1, Q]), but one can still introduce time-and-bandJimiting operators We gmsent a short review of the beaut~fulwork of Landau, Pallak, and Slepian on this subject. We then switch to the continuous wavelet transform: the resolution &the identity in 52.4 (with a proof of (1.3.1)), the corresponding r.k.H.s. in $2 5 "".' 52.e we briefly show how the one-dimensio11al resuIts of the earlier sectlons -can be extended to higher dimensions. In 92.7 we draw a parallel with the continuous windowed Fcwier tramform. IB $2.8 we show how a different kind of time-aad-baa@imiting operator can be built from the continuous windowed Fourier transform or &om the wavelet transform. Finally, we comment in $2 9 on the "zoom-innproperty of the wavelet tiansf-.
7'
2..
Bandlimited functions and Shannon's theorem.
t
,$&'&tion f in .t2(R) is called bandlmdtcd if its Fourier transform F f has ;&Pact support, i.e., j(c) I 0 for #I > R Let up suppose, for simplicity, that 32 = a. Then f can be represented by its Fourier series (see Preliminaries),
18
f t foilaws that
where we have interchanged integral and summation in the th'rrd step, which 1s only a priori justifiable if Ic,I < a, (e.g., if o& finitely'mmy e, are nonzero). By a standard c o n t i r -argument, the final result holds for 811 bandlimited f (for every x, the series%.abealutely summable becaIf(n)f2= 27r 1% l2 < oo). Formula (2.1.1) hIls us that f is completely determined by its "sampled" values f (n). If we fiR the restriction SZ = 7r and 855ume support j c [-fl, 421. with 0 arbitrary, &en (2.1.1) becomes
C
C,
x,
=c
f (4
f(n;)
sin (flz- n x ) nz-nn.
*
n
the function is now determined by its samples f ( n f f ) , corresponding to a
w.
"sampling density" of R/n = (Weuse the notation IAl for the "size" of a , e t A C R, as rneamr&i by the hbesgue rnq!asure; ia this case Isupport jl = I[-fl,Q]l in) Thb -ling density is nsually d e d the Nyquist density. The expansion (2.1.2) g o 9 by the name of Shannon's theorem.
-
in (2.1.2) decsy very slowly (they ark The "elementary building W not even absolutely integrable). 'Y)versampw mdcei~it possible to write f as a superposition of h c t i o n s with fseter d w . Suppose that f is &i baudlimited ll in [-fl,fi] (i.e., support f c [-fl,Q]), but that f ia ssslpled at a rate (1 A) faster than the Nyquist rate,.with A > 0. Thm f can be recoverad born the f (nn/[fl(l+A)]) in the following way. Define gx by
+
THE CONTINUOUS WAVELET TRANSFORM
$25
14 s (1 + N Q ,
Figure 2.1). Because gx r i on support j, we have j(.$)= ft(<)tjA(<). We am now repeat the same construction as before. f )
-
=
C"
, e-~M~/l~(l+~)l
n
n
baw faster decay than
e;
G ~ ~ W ,
note that if A+Ov then as r decay by choosing gA smoother, but it to put too much effort into making very smooth: true, GA will decay for asymptotid1y large x, but the size of A impaaes some an the numen'ml decay of Gx. In other words, a COO choke of $A xkkaying faster than any inverse polynomial,
ax
IGx(z)l
< c;;V(A)(l+
xi)-'^+'),
CN(X)can be very large: it is related to the range of d u e . of Ntb,&&ptive of on [S?, R ( 1 + A)], so that it is roughly proportional to
*. A'J"
Whst
aA
8, k
the f (nxllnfl
if f is "unde~~ampled," i-e., if ~ u p j m f^ ~= 1-R,Q], bug d y kn-, ahere A > O? We have
+)I)
20
CHAPTER 2
where we have used that the ean"~/" have pcrkd 20, and where we have assumed A < (otherwise more terms would intervene in the sum in the last integrand). behave exactly as if they were the This means that the undersampled f (n,&-) Nyquist-spaced samples of a function of narrower bandwidth, the Fourier transform of which is obtained by Ufoldingover" f (see Figure 2.2). In the "folded" version of f , some of the high Frequency content of f is found back in lower frequency regions; only the 5 St(1 - 2A) are unaffected. This phenomenon is called aliasing; for undersampled acoustic signals, for instande, it is very clearly audible as a metallic clipping 'of the ~ p u n d .
3
2.2.
Bandlimited functions as a special case qf a reproducing kernel Hilbert space.
For any a,p,-m
< a < 0 5 m, the set of functions {f E L 2 ( ~ ) auppott ; j c [a, p])
constitutes a closed subspace of L2(R), i.e., it is a subspace, and all Cauchy sequences composed of elements of the subspace converge to an element of the subspace. By the unitarity of the Fourier transform on L2(R), it follows that the set of all bandlimited functions
is a closed subspace of L2(W). By the Paley-Wiener theorem (see Preliminaries), any function f in Bn has an analytic extension to an entire function on C,which we dso denote by f , and which is of exponential type. More precisely,
In fact; &r consists of exactly those Lz-functions fdr which there exists an analytic extension t o an entire function satisfying a bound of this type. We can
FIG.2.2. The Uuree tmnr f(€), f ( € - t ~-A)), (l and their nrm (thick lime).
and
f(€- m(1- Af) for ICI In(l -A),
THE CONTINUOUS WAVELET TRANSFORM
'
21
therefore consider &, to be a H~lbertspace of entire functions For f in Bn we have
sin R(x - y) .(x - 9 )
-
(The interchange of integrals in the last step is permissible if f E L1,i.e , if j is sufficiently smooth. Since, for all x, [n(x - .)]-'sin fl(x - ) IS in L2(R), the conclusion then extends t o all f in Bn by the standard trick explained in Prelirninaries.) Introducing the notation e , ( ~ )= slnQ *, r- we can rewrite (2 2.1)
f ( 4 = (f9ez) Note that e, E Bn, since 2,(<)
= ( 2 ~ ) - ' / ~e-'"< for
It1 > fl.
(2.2.2)
< R,
&(e) = 0 for
Formula (2.2.2) is typical for a reproducing kernel Hilbert space (r.k.H.s). In anr.k.H.s. 7foffunctions,themapassociatingtoafunctionfitsvalue f ( x ) at a point x is a continuous map (this does not hold in most Hilbert spaces of functions, in particular not in L2(R) itself), so that there necessarily exists e, E ? such i that f (z)= (f, e,) for d l f E 31 (by Riesz' representat~onlemma; see Preliminaries). One a h writes
'i.
.*I
'@ "here
& % ? .
* -$
-
K ( x ,y) = ez(y) is the reproducrng kerneb In the particular case of &, there even exist special x, = so that the eZmconstitute an orthonormal basis for B*,leadink to Shannon's formula (2.1.2). Such special z, need not exist in a general r.k.H.s. We will meet several examples of other r.k.H.s.'s in what follows.
%,I
3
2.3.
Band- and timelimiting.
I
h x t i o n s cannot be both band- and timelimited: if f is bandlimited (with arbitrary finite bandwidth), then f is the restriction t o R of an entire analytic function; if f were timelimited as well, support f C [-T,T] with T < oo, then f = 0 would follow (nontrivial d y t i c functions can only have isolated m). Nevertheless, many practical situations correspond to an eflecfive bandend timelimiting: imagine, for instance, that a signal gets transmitted (e.g. ' over a telephone line) in such a way that 6requencies above fi are lost (most redistic transmission means suffer from this kind of bandlimiting); imagine, also, that the signal (such ak a teleplione conversation) has a finite time durati6n. The transmitted signal ie then, for all practical purposes, effectively band- and timelimited. How can this be? And how well can a function be represented by
I
CHAPTER 2
22
i i
such a time- irnd bandlimited representation? Many researchers worked on these problems, until they were elegantly solved by the work of H. Landau, H. Pollack, and D. Slepian, in their series of papers Slepian and POW( M I ) and Landau and P o w (1961, 1962). An excellent review, with many more details than are given here, is Slepian (1976). The example mentioned above (signal with finite time duration transmitted over a bandlimiting channel) can be modeled as f o b let QT,Pn be the orthonormal projection operators in L2(R) defined by
and
(PnfY'(O = ] ( E ) fa14
Q,
- (PnfIA(()= 0
for
KI > R .
Then a signal which is timelimited to (-T, T ] satisfies f = QTf , and transmitting it over a channel with bandwidth R gives as end product Pnf = PnQTf (provided there is no other distortion). The operator PnQT represents the total time + band limiting process. How well the transmitted PnQTf approaches the original f is measured by IlpnQ~f f12/llf112 = (QT Pn QT f,f )/ftfIt2. The maximum value of this ratio is the largest eigenvalue of the symmetric operator QT Pn QT,even explicitly by
The eigenvalues and eigenfunctiions of this operator are now known explicitly because of a fortunate accident: QT PrrQT commutes with the second order . differential operator A, 1^
The eigenfunct~onsof this operator, which had been studied for different rea80ru bng before their conoection with band- and timelimiting warr discovered, are! called the prolate spheroidal wave functions, and many of their properties are known. Because A commutes with QT Po QT (and because the eigenvaluea of A are all simple), the prolate spheroidal wane frinctions are a h the e~nfunctiona pf QT Pn QT (with different eigenvaluea, of course). Mare specifically, if we denote the prolate spheroidal wave functions by $, n f N, ordered so thst the &orreaponding 'eige&mlues a,of A increase aa n incresses, then
A, decreeese as n increeses, and lim A, = 0 n-m
.
THE CONTINUOUS WAVELET TRANSFORM
23
The eigenvalues A, depend on T and a, of course; an eaey scaling argument (substitute x = Tx',y = Ty' in the expression for (QT Pn QT f)(x)) showq that the A, depend only on the product TO. For fixed Tit,the behavior of A, as n increases is schematically represented in Figure 2.3. Typically, thaX, stay cloee to 1 for amdl n, plunge to zero near the t h m h d value 2TO/.rr, and stay c1-
0
10
20
30
FIG. 2.3. The u g e n v a l u o A,, for QTPnQT /or ZTQ/r = 25.
I
to zen, afterwards. More precisely, fqr any (arbitrarily small)e > 0, there exists a @ustant C,so that
which means that the "plunge regionn has width proportional to log(TR). S i tim,,, x-I log x = 0, the width of the plunge region b m e s aegligibly small, when compared to the t b t e s W value 2Tn/n, as T,R4m. In fact, (2.3-2) b a rigorous version of the fact that a time- and h W t e d region [-T, T j x Q] ca-ds to 2Tfl/.rr adegrees of freedom," i.a, khm exist (up to an error, mall,compared to TO) 2TQ/?r independent hmctions (and not more) that a& -tidy timefimitect to [ - T , q d badlimited to [-32, n]. Note that ,2Tfi/uis exactly the area of [-T, TJ x [-0,R], divided by 2n. This num-her is therefore equal to the number of sampling tjmes within [-T, T]specified by Shannon's theorem for a function with bandwidth R; this heuristic way of counting the "indepen&nt degrees of freedomn was part of the folklore of com&cation theory long before it was j d e d by Landau, Pollak, and Siepian. bdependentl~,it was also know to physicists that a region in phaate space ,.(= spacemomentum, or time-frequency such as here) with area S corresponds :&J S/2w "independent states" in the BBmihical limit (i-e., when S is much hrger tban li; the - d o n $ 1 2 ~corresponds to units such that fi = 1). We sin extend the definition of Nyquist deneity fmm ita o& sampling back-d, and use it fbr the critical time&equency density (2a)11 present in all
[-a,
-
-
onrmples. It is time to return to the wavelet tmdxm. In what follows we will develop tbsfxqtkms versions of both the wavelet t r d m a aod the windowed Fourier
w.
24
2.4.
CHAPTER 2
The continuous wavelet transform.
We restrict ourselves, for the time being, to one-dimensional wavelets. We always suppose that 1C, E L 2 ( R ) ; the analyzing wavelet should moreover satisfy the admissibility condition already mentioned in 5 1.3,
+
4
The role of this condition will soon b m e clear. If E L~(R), then is continuous and (2.4.1) can only be satisfied if 4(0)= 0, or J & $(x) = 0. On the other Band, if f tk @{z) = 0 and we impose a slightly stronger condition than integrability on +, namely I&(1 1x1)" I+(x)l < m for some a > 0, then 1$({)1 5 CK14, with B = rnin (a, I), and (2.4.1) is satisfied. It follows that, for all practical purposes, (2 4:l) is equivalent to the requirement that S & +(x) = 0. (In practice, we will impose far more stringent decay conditions on 11, than those n&ed in this argument.) We generate a doubly-indexed family of wavelets from $J by dilating and translating,
+
where a, b E R, a # 0 (we use negative as well as positive a a t this point). The normalization has been chosen so that ~I$J".~II = Il.rLI1 for all a, b. We will assume that = 1. The continuous wavelet transform with respect t o this wavelet family is then
.b)J 5 I(f 11. Note that J(T"""f)(cr, A function f can be recowred from its wavelet transform via the resolution of the identity, as foliows. PROPOSITION 2.4.1. Fm ail f ,g E L2(R),
25
THE CONTINUOUS WAVELET TRANSFORM
The expression between the first pair of brackets can be viewed as (2n)"? times the Fourier transform of Fa(<) = lal'12 &4); the second has a similar interpretation as ( 2 ~ ) ' /times ~ the complex conjugate of the Four~ertransform of Go(<)= 1a11'2 j ( < ) $(&I. By the unitarity of the Fourier transform it follows that
I(<)
-
(interchange is allowed by Fubini's theorem)
(make a change of variables = 4 in the second integral).
m
It is now clear why we imposed (2.4.1): if C* were infinite, then the resolution of the ideniity (2.4.2) would not hold. Formula (2.4.2) can be read as
with convergence of the integral "in the weak =a," i.e., taking the inner product of both sides of (2.4.4) with any g E L2(R): and commuting the inner product with the integral over a, b in the right-hand side, leads to a true formula. The convergence also holds in the following, slightly stronger sense: b
lim
A , -0 A1,B-w
f - C;'
* a2 (T-avf)(a,b)~pb =O.
(2.4.5)
A l l l a 5-42
lYiB
Here the integral stands for the unique element in L2(R) that has inner products with g E L2(R) given by
since the absolute value of this is bounded by
we can give a sense to the integral in (2.4.5) by Mew' lemma. (2.4.5) is then simple:
*
= sup 11911=~
(f -G1
da db
-2-
16159
<
.
sup 11911=1
/I
qX
The proof of
I
&& (TW"'f)(a,bj (Pav9)(a9b): 4 9
t4lA.l or I a l l A ~ or ( b ( > B
1
lf2
F$y Proposition 2.4.1, &he expression between the second pair of bgqdcets is )[g112 = 1, and. the expression between the firat p a of brackets conveqp t o zero as Al-+O,Az,B-+ao, because the infinite integral converges. This establishes (2.4.5). Formula (2.4.5), which shows that any f in L2(Et) can be arbitrarily well approximated by a su~erpoeitionof waveiets, may seem paradmdcak dter all, wavelets have integral zero, so how can any superposition of them (which necessarily still has integral zero) then be a good approximation to f k itself happens to have nonzero integral? The solution t o this paradix d m not lie (as solutions to paradoxes so often do) in the mathematical ~bpphe~s of the question. w e can easily make it at1 rigorous: if we take f E L1(R)j7 L~(R), and t if $ itself is in L1(R), then one easily checks that the
are iudeed all in L1(II) (with norm bounded by 2C;' 11f HL. fl$llLa l l $ t l L ~ B(A;'" - A;'/')), and that they have integral zero,whereas f itself, the function they are approaching as Al 4 0 , Az, B-rm, may well have nonzero integral.
27
THE CONTINUOUS WAVELET TRANSFORM
The explanation to this apparent paradox is that the limit (2.4.5) holds in L2sense, but not in L1-sense. As A14 0 , A2, B-oo,
, becomes a very flat, very stmtched-oiit function, which still has the same integral as f itself, but vanishingly small L2-norm. (This is similar to the observation that the functions g,(z) = (2n)" for ]z/< n, 0 otherwise, satisfy S g,, = 1 for all n , even though g,(x)-+O for all x, and 1%iILz = (2n)-1/2-+0 for n-oo; the & do not converge in L1(R).) Several variations on (2.4.4) are posrjible, in which we restrict ourselves to positive a only (as o p p d to the use of both positive and negative a in (2.4.4)). One possibility is to require that satisfy an admissibility condition slightly more stringent than (2.4.I), namely
.
+
-
Quality of these two integrals fol104s immediately if, e.g., 1// is 8 real function, because then = &<). The resolution of the identity then becomes, with
It(-<)
this new C*,
db T"" f (a,b) 9"lb, to be understood in the same weak or slightly stranger senses as (2.4.4). (The pmof of (2.4.7) is entirely analogous to that of (2.4.4).) Another variation occurs if f is a seal function, and if support 4 c [O,oo). In this case, one easily proves that
18
L
with
C* as defined by (2.4.1). (Toprove -(2.4.8), use that /(z)
= ('2~)-'/~
2bJ; 4 eszcf((), because j(-<) = &).) Formula (2.48) can of course be rewritten in terms of dl = fCe $ and & = Im $, two wavelets which are each , other's Hibert transform. Using a complex wavelet, even for the analpis of real functions, may have its advantages. In Kronland-Martinet, Morlet, and Grassmann (1987), for example, a complex wavelet q9 with support C 10, oo) is used, and the wavelet transform TYaVf is rep-@ by graphs of its modulus and its pbsee. If both f and tl, are sc+called "analytical signals," i.e., if support and h&Wt c [@ m), then Fvf (a,b) = 0 if a < 0, so that (2.4.4) immediately s i m ~ t o
4
4
=
CHAPTER 2
28
with C+ e 2 i n as defined by (24.1). Finally, we can adapt (2.4,9) to the casr where support d, C (O.cm), but support f @ [0,m). We write f = f+ f-, with support j+ c [0,m), support j- c (-00,01, $+ = 11, and we introduce $- (t)= $(-E); dearly sup&& &- C (-m,O]. Then ( f + , + ? b ) = 0 and (f-,~):~) = 0 for a > 0, so that, by straightforward application of (2.4.91,
+
$yb),
where (T;.Yf ) (a, b) = (f+ ,$ )': = (f, (Yay j) is defined analogously, and C3,isasin(2.4.1). . Another important variation consists of introducing a different function for the reconstruction than for the decomposition. More explicitly, if $2 satisfy that (2.4.11) I4 ICI-I 1 6 ~ ( t )116 2 ( ~ ) 1m <
.
T
then the same argument as in the proof of Proposition 2.4.1 shows that
with C+1.+2= 2~ (2.4.12) as
I4
41(C) 4 2 ( < ) .
If C+,,h# 0, then we can rewrite (2.4.13)
Note that and $2 may have very different properties! One may be inegub lar, the other smooth; both need not even be admissible: if (<) = 0(<) for €4, then &(o) # O is allowed. We will not use this extra freedom here. Iq HoWneider and Tchamitchian (1990), the freedom in the choices of 702 is exploited to prove some v e ~ yinteresting results (see also 52.9). One can, for instance, choose & to be compactly supported, support +2 c [-R, R], so that, for any 2, only the (f, @*) with - 21 5 lalR will contribute t o f (z)in the reconstruction formula (2.4.13); the set ((a, b); Ib - x( _< (aiR) is then called the "cone of influencen of & on 2. HoLschneider and Tchamitchian { l ~also ) prove that, with mild conditions on f , (2.4.13) is true pointwise as h 1 l as in the
L2-sense. PR~POSITION 2.4.2. Suppo~eh t & f L1(W),that $2 is di,fferentiable with 6 E L2(R),fiat x h E L1(R), and that $1 (0) = 0 = If f E L2(R) is bounded, then (2.4.13) holds pointwise in every point x where f is continuous,
A@).
i.e.,
(f,$:'b)
$ ; v ~ ( x ).
(2.4.14)
29
THE CONTINUOUS WAVELET TRANSFORM
f
Proof. 1. We can rewrite the right-hand member of (2.4.14) (before taking the limit) as
where all the changes of order of integration are permitted by fibini's theorem (the integral converges absolutely). Here M A ~ , is A defined ~ by
2. One easily computes that the Fourier transform of MA,,Az is
-
d" &!(a) &(a), w follow from a where ~ (= 0 (2r)'l2 c;;,, $ials,clA change of variables a+& in (2.4.16). Since ~ + ~ ( Ea )L2(R) and ( a ) is bounded, we have
4l
By (2.4.1 I),
M is also bounded, so that
implying that M , t i e inverse Fourier transform of M , is well defined, bounded, and continuous.
3. The decay of M is governed by the regularity of M. For checks that M ie differentiable in (, with
$0, one easily
Because srlrz E L1, 4 2 is differentiable, so that, for
= 0,
It follows that M is differentiable. Moreover, since & E L1,we have
which implies I4 then follows from
< CH[/&(<)I+ ISL(-C)~],so that &M
[/,
M
that M E
M( c ) I
(1
+
L1(R). Moreover, ~
.2,-1]
[/&+
'Iz
(1
( 0 =) ( 2 ~ ) ' ' c ~s:,h
E t2. It
I
/M(z)~
'I2
b?hb)61 = (0)
4. Using (2.4.17) we can rewrite (2.4.15) as
BecauseJM is continuous, integrable, and of integral 1, the first term tends to /(x) for A14 0 iff is bounded, irnd continuous in x. (Thisfollows kom a simple application of the dominated convergence theorem.) The second. term is bounded by
because M E
L2(R)by (2.4.18)- This term therefore tends to zero if
A2--*oo.
REMARK. In Holachneider and Tchamitchian (1990), this theorem ie proved under slightly more general conditions on f as well aa on h. o
31
THE CONTINUOUS WAVELET TRANSFORM i
+*$be reptoducisg kernel Hubert space underlying the continuous
- 4%-awavelet transform. &a
special urse of (2.4 2), we have, for f E L2(R),
In other words, Fa" maps L2(R)isomdtrkally into L'(R*; c;'ad2da db), the db space of all complex valued functions F on'R for which (I]PIJIZ = CJ' JJ do 7 l ~ ( ab)12 , converges; equipped with the m m 111 I/(,this is a Htlbert space The image TOVL2(It)constitutes only a s ~ , s y b p a c e ,not all of L'(R~; Cila-: da db); we will denote this subspace 71. The fgllowing.argument shows that H is a rAc.ff.6. Fqr any F E 'H, we can $ed f E L ~ ( R m ) that F = Fa" f, It follows then fqm-(i.4.2) tlwt
,I
.
Q
K (a, b; a', 6') = =
(a', bt)
L
(@a1,b', *avb)
4
Fbrmula (2.5.1) shows that 'H b-indeed an r.k.H.s. embedded as a subspace in ~'(3; C
-
2
<
-
F(a,b) = (f,+a*b)
=.
2a'/2/dm4
-
f(l) e-*(b+')(
= (2s)a3/' G(b+ ia), *% Gk is Baalytic on the upper hdf plane (Im r > 0). Moreover, one easily 3,
fZ
that
+
db alG(b *a)/' = /dz
If
(%)I2
,
32
CHAPTER 2
so that T"" can be interpreted as an isometry from H~ to the Bergman space of all analytic function on the upper half plane, square integral& with respect to the measure im z d(1m z ) d(Rez). On the other hand, one can Hove that any function in this Bergman space is asbciateti, via the wavelet transform with this particular $, to a function ixi H Z : the isometry is onto, and is therefore a unitary map. For qther choices of $, such as $ E H 2 with $(<) = Ng cfl e-t for 2 0, the image TwavN2 can be identified with other Bergrnan spaces of analytic functions on the upper half plane. Since T"""L2 or T""' HZ can be identified with a dproducihg kernel Hiibert s ace, it should be no surprise that there exist discrete families of points (a, b,) kuch thGt f is cornplet~lydetermined by, and cari be reconstructed from, (T""" f ) (a,, 6,). In particular, if Fa" f can be identifikd with a function in a Bergman space, then it b obviols that its v a l ~ l nat certain discrete families of points completely determine thc function, siricc it is, after ah, an analytic function. Reconstr~lctingit in a numerically stable way may be another matter: the situation is not as simple as i l l tlic I~andlirriitedcase, where there exists a special family of points x, sucli that t11~ex,, constitute an orthonormal basis for Bt2. There is no such corivcnier~torthonorrrrH1 basis e,;,b, in our T""L2 or P a Y H Z .We will see in thc next chapter how this pro1)lCrn can be tackled. Finally, before we leave this section, it should be rexnarked that (2.4.41, or the equivalent r.k.H.s. formulation, can be vicwctd as a consequence.of the theofy of square integrable group representatioxis. 1 do not wish to'dwell on this in detail here; readers who are interested in learning more about them should .consult the references in the notes. The are in fact tht! result of the action of the operators U ( u ,b), tiefincd by
+
.
'
$"tb
on the function $c The omrators O ( a ,b) are all unitary on L2(R), and cons&itute a representation of the a x bgroup:
+
U ( a ,b) U(a',b') = U(aat, b
+ ab') .
i
This group representation is irreducible (i.e., for any f # 0, there exists no nontrivial g orthogonal to dl the U(a,b) f , which is equivalent t o saying that the U ( a ,b) f s p m the entire space). The following result is true: if U is an irreducible unitary representation in ')I of a Lie-group G with left inmiant measure dp, and if for some f in 3.1,
.
then there exists a dense set D in 'H so that property (2.5.2) holds for any element f of D.Moreover, there exists a (possibly unbounded) operator A, well debed in D, so that, for all f E D and all hl,hz E 'H, -
33
TIIE CONTINUOUS WAVELET TRANSFORM
with C j = (A f , f). In tlie wavelet ease, the left invariant measure is a-2 dadb, anrl A is the opt:rator P ( A f )"(O = 14-I j ( < )Note that (2.5.3)is a grrleral resolution of the identity! In what follows, wc will not exploit this group structure underlying the
wavclct transforn~,miiinly because we will soon go to discretely labelled wavelet families, and tlrtsc (lo ~ ~corresporld o t to subgroups of the k bgroup. IJI quantum ~)hysic$,rcsolr~tionsof the ideiitity (2.5.3) have been studied and 11xd for mrwiy tlitfcrent groups G. The associateci families U ( g )f are there ca1k:cl cohere~itstatt.~,a name that was first used in connection with the Weyl Zieiseuberg group (siv also next x'ction), but later spilled over to all the other ,group; as'wbll ( i ~ l l eve11 ~ i to S O ~ ~ Irelated C constructions which w r e not generated by a group). An excellent review and a collection of important papers on this sutjjcct can Ije fo1111din Klauder and Skagerstani (1985). . he collerent states assuciatcbcl with thc: ax bgroup, which are r~owcalled wavelets, were first construj:tqJ by Aslisksc~rli11id Klauder (1968, 1969).
+
+
br"P
2.6. , ,
The contirluous wavelet transform in higher dimensions.
T k r c oxist st*vc>risIpossit>le cxt.onsions of (2.4.4) to L ~ ( w " )with 71 > 1. One possibility is to choose thc wavelct G L2(R") so that it is spherically sym~nctric. - Its Follricr transform is tllcn sph:rically symmetric as well,
2 ,
4K) = dl
.
!
and the ;utn~issii)ilityconciition twcornes
c
.)
" dt
- j9(t)l2 < m . t :
=2
'
Alang tlie wit: lincs as in the proof of Proposition 2.1 one can then prove that, for all f, g E L2(RTb),
"
a & ~ m d b ( T w a v f ) ( a , b ) ( ~ a v 9 ) ( a , b ~ = C y ( f l ~(2.6.1) )7 a J
here (Twavf ) ( a ,b) = ( f , $a*b), as before, and $ J ~ * ~ ( = X ) a-n/2 $(?I, f W+, a # 0, and b E Wn. Formula (2.6.1) can again be rewrithen as
=c~ll k" "
f
da
p
db (TBV f ) ( a ,b)
.
with
(2.6.2)
s
:
*moducerotations as well as dilations and translations. In two dimensions, for i-w, we then define
34
CHAPTER 2
where a > 0, b f It2,and where
Re is the matrix cos6 -sin9 sin 8 cos 9
>
.
The admissibility condition then becomes
and the corresponding resolution of the identity is 4
A similar construction can be made in dimensions larger than 2. Theae wavelets with rotqtior angles were studied by Murenzi (1989), and applied by Argoul et d. (1989) in a study of DbA (diffusion-limlted aggregates) and other two. dimensional fractals. 2.7.
Ptlrallels with the continuow windowed Fourier transform.
The windowed Fourier transform of a function f is given by
(2.7.1)
v""'"f)(w, t ) = (f, gW't),
where gW9'(x)= ewzg(z - t ) . Argumentg completely similar to those in the proof of Proposition 2.4.1 show that, for ail fl,f2 E L2(W),
which can be rewritten as
There is no admissibility condition in this case: any window function g in E2 will do. A convenient normalization for g is 11911 = 1. (The absence of a9 &himibility condition is due to the unimodularity of the Weyl-Heisenberg grou-. Groesmann, Morlet, end Paul (1985).) The continuous windowed Fourier transform can again be viewed aa a map frodZZ(R)to an r.k.H.s.; the functions F E T*'"L'(R) are all in L~(R') and moreover satisfy
where K(w, t ; w', t r ) = (@"*', pi). (We assume 11g11 = 1 here.) Ag& there exist very special choices for g which reduce this r.k.H.8. to a space of analytic functions: for g(x) = n-lI4 exp(-x2/2), one finds -
THE CONTINUOUS WAVELEZ' TRANSFORM
35
where q5 is an entire function. The set of all. entire functions # which can be obtained in this way constitutes the Bargmann Hilbert space (Bargmann (1961)): The obtained from g(a) = go(%) = exp(-z2/2) are often d i e d the canonical coherent states (see the primer in Klauder and Skiigerstam (1985)); the associate continuous windowed Fourier transform is the canonical coherent state representation. It has many beautiful and useful properties, of which we will explain one that will be used in the next section. Applying the differential d2 operator H = -&I "x 1 t o go(x) leads to g"lt
. '
'
+
,,- i.e., go is an eigenfunction of H witheigenvalue 0. In quantum mechanics lanb
'
-
-:, m
e , H is the harmonic oscillator Hamiltonian operator, and go is its ground state. (Strictly speaking, H is really twzce thestandard harmonic oscillator H d t o n i a n . ) The other eigenfunctions of H are given by higher or&r Hermite iimctions,
,
, fr
$
*Ki
A
satisfy
H4, = 2n 4, .
(The standard and easiest way to derive (2.7.4) is t o write H = A'A, $ where A = x + and A* is its adjoint A* = z - $, and to show that Ago = 0, A(A*)n = (Am)"A+ 2n(A*)n-a, so that Hq5, = an A*A(A*)n go = 2n(A*)"" go = 2n &; the n o ~ i z a t i o a i,~ can be computed easily as well.) It is well known that the (4,; n E N) form an o r t h o n o d basis for L2(R);they const~tutetherefore a "complete set of eigenfunctionsn for 11. Let us now consider the one-parameter firmilies = exp(-iHs)+. These wa the solutions to the equation
6,
*,
i&$s
= HJr,
,
'
(2.7.5)
t h initial condition qlb = @. In the very special case where &(z) = &"(x) = euz exp[-(2 -- t ) 2 / 2 ] ,we find Jt. = ew* g,Y.lt*, where w. = w ecs b n 28, t. = w sin 2s t cm 23, and a. = f (wt - w.t.) (as can easily be verified m c i t computation). That is, a canonical coherent state, when "evolved" &r (2.7.51, remains a c.anonica.1cohore~tetate (up to a phase factor which will important to us); the label (w,,t,) of the new coherent state is obtained the initial ( w , t) by a simple rotation in the time-frequency plane.
-
+
i b
-
-
+"3& The continuous transforms as tools to build useful operators.
-
, %%-ed
l l t i o s of the identity (2.4.4), (2.7.2) can be rewritten in yet another
way:
(2.8. la)
CHAPTER 2
where (., &)$ stands for the operator on L2(IR)that sends f to ( f , 4)4; this is a rank one projection operator (i.e., its square and its adjoint are both identical to the o ~ e r a t o ritself, and its range is one-dimensional). Formulas (2.8.1) state that a "superposition," with equal weights, of the rank one p~ojectionoperators corresponding to a family of wavelets (or a family of windowed Fourier functions) is exactly the identity operator. (As before, the integrals in (2.8.1) have to be taken in the weak sense.) What h a ~ p e n sif we take similar superpositions, but give different weights t o the different rank one projection operators? If the weight furiction is a t all reasonable, we end up with a well-defined operator, different from the identity operator. If the weight function is bounded, then the corresponding operator is as well, but in many examples it is advantageous t o consider even unbounded weight functions, which may give rise to ~nbounded operators. We will review a few interesting examples (bounded and unbounded,) in this section. U7estart with the windowed Fourier case Let us rewrite (2.8.lb) in the p, q (momentum, position) notation customary in quantum mechanics (rather than the w , t notation for the frequency-time plane), and insert a weight function
If w 6 L"(w~), then W may be unbounded and hence not everywhere defined; as a domain for 1%' we can then take ( f ; S l d p dq I ~ ( p , ~ ) I(f, l ' ~ P V Q <) ~ oo), which is dense for reasonable w and g.
Two useful examples in quantum rnrchanics are (1) w(p.9) = pZ, which leads to W = -$ Cg Id, where Cg = S@ if2lg(t)l2, and ( 2 ) w(p,g) = v(q), for which W is a multiplicative potential operator: (Wf )(x) = Vg(z)f (x). with V,(z) = J dq v(q)lg(z - q)12. Readers familiar with the basics of quantum mechanics will notice that in both cases the operator W corresponds pretty well to the "quantized version" of the phase space function w(p,q) (in units such that A = l ) , with a slight twist: t h e extra constant C; in the first case, the substitution of v * lg12 for the potential function v in the second case. In fact both formulas were used in Lie4 (1981) to prove that Thomas-Fermi theory, a semiclassicaI theory for atoms and molecules, is "asymptoticallyn correct (for Z400,i-e., for very heavy atoms); it gives the leading order term of the much more complicated quantum mechanical model. Lieb's proof used the two examples above in three dimensions (rather than one); the operators he really wanted to consider were, of course, -A = -a2= I -a2l a -d2 ,23 and V ( x ) = [x: 2; SO that he had to chat% an appropriate g and deal with the extra constant C, and the difference between V and V * lg12 by some other means. Note that choosing v(q) with an integrable singularity (such as the three-dimensional Coulomb potential) always leads to a nonsingular Vg: operators of type (2.8.2) cannot represent such singularities. Many othel applications of operators of type (2.8.2) exist. In pure m a t h e rnatics they are sometimes called Toeplitz operators, and whole books have been
+ +x Z ] - ~ / ~ ,
+
~
THE CONTINUOUS WAVELET TRANSFORM
3'1
written on them. In quantum optics they are also called "operators of type P," and again there exists an extensive literature on the subject (see Klauder and Skagerstam (1981)). But let us go back t o signal analysis, and see how (2.8.2) car, be used to build time-frequency localization operato'rs. - Let S be any measurable subset of Ht2. Let us return to tirne-frequency notation, and dkfine, via (2.8.2), the operator Ls corresponding to the indicator fu~ctionof S, a(w, t) = 1 if (w, t) E S, 0 if (w, t) 4 S,
.
It follows immediately from the resolution of the identity that
on the other hand, obviously ( L sf , f ) 2 0. In other words,
I f S is a bounded set, then the operator Ls is trace-class (see Preliminaries), since, for any orthonormal basis ( u ~ )in, L2(R), ~ ~ '
<
+
(order of integral and summation may be inverted by Lebesgue's dominated convergence theorem)
*4 where IS( is the measure of S. It follows that there exists then a complete set of eigenvectors for Ls, with eigenvalues decreasing t o zero,
%%i*
(4,; n E N) orthonormal basis for L ~ ( w.)
P
. Such an operator Ls has a very natural interpretation. If the window function g reasonably well localized and centered around zero in both time and frequency, 4
J'
CHAPTER 2
38
then (f,g"st) gWlt can be viewed as an "elementary component" of f , localized in the time-frequency plane around (w, t). Summing all these components gives f again; Ls f is the sum of only those components for which (w, t) E 5. Therefore, Ls f corresponds to thecextractionfrom f of only that information that pertains to the region S in the time-frequency plane, and the construction from that l e calized information of a function that "lives" on S only (or very nearby). This is the essence of a timefrequency -tion operator such as we saw in $2.3! We can now, moreover, study Ls for sets S much more general than rectangles [-0,521 x [-T, TI.(Note, however, that even for S = 1-R, Cl] x I-T, T j , our o g erators Ls tire different from the QT QT considered ia 52.3.) Unfortunately, for most choices of S and g, the eigenfunctions and eigenvalues of Ls are hard to characterize, and this construction is of limited usefulness. However, there is one choice of g and one particular family of sets S for which everything is transparent. Take g(x) = g,(z) = T-'/' ex&--z2/2), and SR = {(w,t ) ; u2 t2 R2). Let us denote the corresponding l o d b t i o n operator by LR,
+ <
These operators LR commute with the harmonic oscillator Hamiltonha H= + z2 - 1 from 52.7, as ca.n be seen by the following argument. Since
-&
e -*Ha
&.t
=
g,~.,t,
,
with a, = (wt - w,t,)/2 E R, we have
hence
I
If we substitute w' = w-,, t' = t-,, then one easily checks (use the explicit formules for a,, w., to at the end of $2.7) that % s t = g$'t' = exp [- (dt' - wt)] e-lH0 g t t t l . On the other hand, the domain of integration is M 8 n t under the transformation (w, t)+(wr, t') (because this transformation is simply a rotation in time-frequency space!). so that
4
and LR commutes with H, as announced. It fobws that there exists an orthonormal basis in which both LB anb H are diagonal (see PrdMnaries). But,
THE CONTINUOUS WAVELET TRANSFORM "
39
since the eigenvalues of \H are all n o n d e g e n d , there exists only one basis that d&gonalizea H,namely the Hermite fundions (see 52.7). It follows that the Hermitt: functions-& are necessarily the ~ c t i o n ofs LR. The eigenvalues of LR can be computed from
(There are many ways to compute this eqmmsion. One way, via the Bargmann Hilbert space, is explained in Note 3 at the enci d this chapter.) We then have
+FJ Y 1
3;
*
.LR & = h(RMnr
2
with
-- 2s
//
dwdt
1 n! Zn
I
exp
d + t 2 j R2
-r *hi&
is a so-called incomplete r-function. h this explicit formula for X,(R), &litis now poeaible to study its behavior as a function of n and R. I will summarize --$ 4"k?pdythe results here (details can be found in Dauhechies (1988)); Figure 2.4 also &shows a plot of &(R) for three different d u e s of R. For every R, the A.(R) %: * > ; , decrease monotonically as n increases; for smatl n they are close to 1, for large t% n c1ow to zero. The threshold value around wbicb they make this 'plunge," as ii5**( e e d , for example, by n t h , = max{n; & 1/21, is nth, R2/2. Note that %r%&s is again equd to sIZ2/27r, i.e.., the area of the timefrequency localization P w o n SR multiplied by the Nyquist density, just as in 52.3. The width of the &. w g e region is wider than in 52.3; however, -b
-
r
>
-c.
-9 h
p
$iZ t >
*
6
,%
&%; 73
*
%@ -pared i
$.*. $:,
-
*7 it&
-
# {n; l - ~ > X , 3 c ) s C , R , to the logarithmic width in (2.3.2)), but it is still negligible, for
R* when compared with nth. Another striking difference with 52.3 is that in this *case are i n m e n t of the size of the region SR eigenhnctions
the prolate spheroidal wave funetio11~):the R-dependence is completely ,$dncentmted in the An(R).
Ij: @dike
..-
CHAPTER 2
FIG 2 1
The etgenvalues A, ( R ) for R = 3, 5, and 7
Examples silnilar to all of the above exist for the continuous wavelet transform We can again insert a non-constant function w(a, b) in the integral in (2.8. la), and construct operators W different from the iderrtity operator. An example is w(a,b) -- a2 in three dimensions, with a spherically symmetric 1C, (where the resolution of the identity is given by (2.6.2)), i.e., ( W f ) ( x )=
c;'
B
"ah
u./,
d 6 ~ a 2 ( ~ " j ) ( a , b ) ~ " ~ ' ( z ) ,(2.8.3)
c+
4(()
where = 4(IEl) and (?+ = ( 2 ~ l )r d~s s#(s). Because the three-dimensional Fourier transform of g(x) = is g([) = %&/(,hi [(I) (in the sense of distributions), one easily checks that Wj can also be written a s
so that (Wf , g) represents the interaction Coulomb potential energy for two charge distributions f and g. This formula was used ,in, e-g., the relativistic stability of matter paper by Fefferman and de la Llave (1986). N0t-t (W f ,g) becomes "diagonal" in the representation (2.8.3) (which, incidentally, is why it turned out to be useful in Feff-an and de la Llave (1986)). Note also that this diagonal wavelet representEStion completely captures the s-ity of
41
THE CONTINUOUS WAVELET TRANSFORM
the kerrlel in (2.8.4) iio "clipping off" of the singularity as in the windowed Fourier case. This is due to the fact that wavelets can zoom in on singularities (an extreme version of very short-lived high frequency features!), whereas the ,, windowed Fourier functions cannot (see 51.2 or 32.9). We can also, as in the wlndowed Fourier case, choose to restrict the integral in (2.8. l a ) t o a subset S of (a, b)-space, thus defining time-frequency localization operators Ls. These are well defined for measurable S , and 0 Ls 5 1. For compact S not containing any points with a = 0, Ls is a trace-class operator. For general S, the eigenfunct~orisand eigenvalues may agair? be hard to characterize, ,Qut there exist again special choices of ?/I and S so that the eigenfunctions and eigenvalues of Ls are known explicitly. Their anaiysis is similar t o the windowed . -- Fourier case, but a bit more tricky. We will only sketch the results here; for full detaib the reader should consult Paul (1985) or Daubechles and Paul (1988). One such special is $(<) = 2< e-t for J 2 0, 0 for 5 0; the associated ; resolution of the identity from which we st@ is (see (2.4 9))
<
<
$J
where
++ = 4-(<) = $(-()
The operators LC = Ls, we consider are glven
$J,
+
with Sc = {(a,b) E W+ x W; a 2 f b2 1 _< 2aC), and C 2 1 In the representation of (a, b)-space as the upper half eompiex plane (z = h + za), the Sc correspond t o the disks Iz -zCI2 < C2- 1. The role of the harmonic oscillator Hamiltonian is now played by the operator H defined by
H(f)^(<)=
dL
d
For this H wehave then
3' '
where
b ( t ) + za(t) = ~ ( t=)
+
L
cost +sint
cost
- zsint '
with z = b ia. One easily checks that the flow z+z(t) preserves all the circles ]z - iCI2 = C2 - 1, as illustrated by Figure 2.5. It follows that H and the LC wmmute, so that they can be diagonalized simultsneously. The eigenvalues of H d l have degeneracy 2; for every eigenvalue En = 3 2n we can find two gigenfunctions,
+
(%43A(€) =
+
2& [ ( n 2 ) ( n
+ I)]-'/~
[ LZ(2t)e-c
for t 2 0
,
fort 5 0 ,
41
THE CONTINUOUS WAVELET TRANSFORM
the kerrlel in (2.8.4) iio "clipping off" of the singularity as in the windowed Fourier case. This is due to the fact that wavelets can zoom in on singularities (an extreme version of very short-lived high frequency features!), whereas the ,, windowed Fourier functions cannot (see 51.2 or 32.9). We can also, as in the wlndowed Fourier case, choose to restrict the integral in (2.8. l a ) t o a subset S of (a, b)-space, thus defining time-frequency localization operators Ls. These are well defined for measurable S , and 0 5 Ls 5 1. For compact S not containing any points with a = 0, Ls is a trace-class operator. For general S, the eigenfunct~orisand eigenvalues may agair? be hard to characterize, ,Qut there exist again special choices of ?/I and S so that the eigenfunctions and eigenvalues of Ls are known explicitly. Their anaiysis is similar t o the windowed . -- Fourier case, but a bit more tricky. We will only sketch the results here; for full detaib the reader should consult Paul (1985) or Daubechles and Paul (1988). One such special $ is $(<) = 2< e-t for J 2 0, 0 for 5 0; the associated ; resolution of the identity from which we st@ is (see (2.4 9))
<
where
++ = $, 4-(<) = $(-()
The operators LC = Ls, we consider are glven
+
with Sc = {(a,b) E W+ x W; a 2 f b2 1 _< 2aC), and C 2 1 In the representation of (a, b)-space as the upper half eompiex plane (z = h + za), the Sc correspond t o the disks Iz -zCI2 5 C2- 1. The role of the harmonic oscillator Hamiltonian is now played by the operator H defined by
H(f)^(<)=
dL
d
For this H wehave then
3' '
where
b ( t ) + za(t) = ~ ( t=)
+
L
cost +sint
cost
- zsint '
with z = b ia. One easily checks that the flow z+z(t) preserves all the circles ]z - iCI2 = C2 - 1, as illustrated by Figure 2.5. It follows that H and the LC wmmute, so that they can be diagonalized simultsneously. The eigenvalues of H d l have degeneracy 2; for every eigenvalue En = 3 2n we can find two gigenfunctions,
+
(%43A(€) =
+
2& [ ( n 2 ) ( n
+ I)]-'/~
[ LZ(2t)e-c
for t 2 0
,
fort 5 0 ,
(-c).
and;)t( )" (C) = (+$IA Here L : is the Laguerre polynomial (2 is a superscript, not a power) as given by the general formula
Since the operators LC comrnutk with the parity operation (IT>)"(() = j(-t), it follows that $ ,: are eigenfunctions for the LC as well (because of the degeneracy of H,not every eigehmction of H is a priori an eigenfunction of LC!).The corresponding eigenvalues of LC are
+;
that in this cam every (This means that LC has the same degeneracy as H , eigenfunction of H is in fact an eigedimction of LC as well.) We can therefore slso use t g,= .ji($: ',+ ) and $g= ($A- - $; ,-) as eigenfunciions;these
+;
-
4
43
THE CONTlNUOUS WAVELET TRANSFORM
3%
have the advantage that they are real. Figure 2.6 shows the plots of the first few 2;- $J,: (e for even, o for odd). A graph of X,(C), for various values of C, is given 2 in Figure 2.7. For reasonably large C , the X,(C) behave as we by now expect of @ * the eigenvalues of a time-frequency localization operator: they are close to 1 for 7 n small, with Ao(C) = 1and they plunge to 0 for a larger, C-dependent value of n. More precisely, for any -y E (0, I), the value of n for which A,(C) crosses y isequal t o n =qC+0(1) (Clarge), with ~ j ( 2 + 7 7 - ~ ) ( 1 - 2 C ' ~ )=~y~ 2 , s6 or 2q - ln(1 + 29) = - In y + Q(C-I). This implies that
+;
-+
-L
f?. Tg
# {eigepfunctions; A,(C) L 7 ) = 2#
&A
{n; L ( C )
2 71
= Z C F - ~( - I P T ) + O ( I ) ,
where F ( t ) = 2t - ln(1
+ 2 t ) . In particular,
9
# {eigenfunctions; X,(C) 2 1/21 = 2C J"'(In2) + O(1) .
(2.8.5)
In ~ r d e to r compare this with the Nyquist density, we first need to find the area in timefrequency space corresponding to LC.To do this, we go back t s the
+zb.
w e have
-
*,&J
@
Hence
k3' .A9*
Sc = {(a, b) E R+
x
W; a2 + d
+ 1 5 2aC) corresponds to the time-
frequency set sc=((u,t)E~2; *
'
9 C ^ +gw2 -+IS--
")
.
*
14
This co~espondsto a low frequency as well as a high frequency cut-off; see Figure 2.8 for a comparison between this timefrequency localization qet and the disks ' of the windowed Fourier case. lo The area of & is l$l= 6x(C - 1). Combining ' this with (2.8.4) gives
I ,C
# {eigenfunctions; X,(C) 2 1/21 = 1 ~"(ln2)+0(~-')
t&i
3~
which is different from the Nyquist density! This contradiction is only apparent, sad is due to the fact that the width d the "plunge region" of the &(C) is - proportional to C, and thus to I&/. We have indeed, for c > 0,
CHAPTER
2
FIG. 2.6. Plots of the ergenfunctrons $&,$: for n = 0 , 1, ond 2.
n FIG 2 7
The ezgenvalues X,(C)
for dzfferenl values of C
i
e
in contrast to the prolate spheroidal wave case, where the analog of the expression between curly brackets tends to zero as IS1-' log IS1 for ISI-oo, and with the windowed Fourier case, where it behaves like JSI-1/2 for ISI-cm. The fact that in the present case the width of the plunge region is of the same order as lScl itself results from the non-uniform time-frequency localization of the @ $*: it is an indication that we have to be careful with time-frequeocy-density-based intuition when dealing with wavelets. We will come back to this in Chapter 4. 2.9. I
.'
,, E
**i
a
-
The continuous wavelet transform as a mathematical zoom: The characterization of local regularity.
This section is entirely borrowed from Holschneider and Tchaniitchian (1990), who developed these techniql~esin part to study local regularity properties of a nondifferentiable functivn proposed by Riemar~n. THEOREM 2.9.1. Suppose that l d x ( l + 1x1) 1@(2)1 < oo, and $(0) = 0. If a bounded fvnrtzon f as Holder contznuotes with ercponent a, 0 < cr 5 1, i.e., .
then its wavelet tmlwfbnn sutzsfies
CHAPTER 2
FIG 2 8 (a) The set
& = {(t,w),
t2
+ 5 + 1 5 8)for d r f f e d values of C
(b) A
cornpanson between the tame-frequency l d a z a t w n sets fur the wtndwed Fourier tmnsform (the drrk SR = { ( t ,w ) , t 2 + w2 5 R ~ at) irfl) arid for the wavelet t r u w j m (at nght)
P m f S~nceJ & $(x) = 0 we have ($a*b,
(9) [fw
f) = J dz I ~ I - ~ / ~
- f(b11 ;
hence
The following is a converse theorem. THEOREM 2.9.2. Suppose that rs oompcrctlgj s u p p o w . Suppose also' that f E L2(R)u bounded and continuous. U, for some'u E 10, I[, the wavelet tmnsfm o f f satisfies ~(fi 5 ~ [ a l * + ' /,~ (2.9.1)
+
b
the^ f u Holder conitnuow wth e q p e n t a. 'L
47
THE CONTINUOUS WAVELET TRANSFORM
1. Choose +2 compactly supported and continuously differentiable, with dx h ( x ) = 0. Normalize4& s f h a t C#,* = 1. Then, by Pr~pasition 2.4.2.
da (f,Pb) *;*b(x) -w
-0
We will split the Litegral over a ink, two parts, Ial < 1 and la1 2 1, and call the two terms fss(x) (small axles) and fLs(x) (largescales). 2. F~rstof all, note that
r
fLS
5 C Next,
we look at I fLs(x
is bounded uniformly in s:
lI+lIIL1
/
do 14-3/2 = C, < 00
.
(2.9.2)
laiLi
+ h) - fLs(x)J for Ihl < 1:
Since J&(r+ t ) - ?,b2(z)l5 Cltl, and since support for some R < m, we can bound this by
(2.9.3)
+, support tCrz C [-R, R]
This holds for all /hi _< 1; together with the bound (2.9.2), we conclude that IfLs(x h) - fLS(x)( < Clhl for all h, uniformly in 2. Note that we did not evpn me (2.9.1)in this estimate: fLs is always regular.
+
3. The small d
e part fss b also uniformly bounded:
CHAPTER 2
4. We thcrefor~again only have to check I fss (x + h) - fss(x)1 fo; small h, such as lhl 5 1 . Ifsing again 1$2(2 t) - 1Ln(z)15 Cltl, we have
+
It follows that f is Hijlder continuous with exponent a. Together, Theorems 2.9.1 and 2.9.2 show that the Holder continuity of a function t a n be characterized by the decay in u of the absolute value of its wavelet transform. (Except for a = 1, where we do not have complete quivalence. ) Note that we did not assrime any regularity for .tD itself: apart from decay conditions on @, b e only exploited that $& $(x) = 0. (Although this condition is not stated expliritly in Theorem 2.9.2, $ nevertheless satisfies it: the bound (2.9.1) cannot hold otherwise.) Higher order difkrentiability of f and Holder continuity of its highest order well-defined derivative can be characterized similarly by means of the decay of the wavelet coefficients if 11 has more moments zero: in order to characterize f E Cn and Halder continuity with exponent a of f(") we will need a wavelet SO that J dx xm+(x) = 0 for m = 0 , 1 , .. . ,n. For such a wavelet we have, for a E 10, I[,
+
f
E
m = 0, . . - , n bounded 'and square integHijlder continuous with exponent a
C" , with all the f ("1,
rable, and
f(")
I (f,+a'b) 1 L ~la1"+'/*+~,uniformly in a . Again, we require no regularity for $. What is most striking about all these characterizationi is that they only involve the absolute value of the wavelet transform. Note that one can also derive regularity of f from the decay in w of the absolute value of its windowed Fourier transform T"'"(w, t), if the window g is chosen sufficiently smooth. In most cases however the value for the Holder expoaent computed from (w, t)1 will not be optimal. To obtain a true characterization, the phase of T"in(w, t) should also be taken into consideration, for instance via Littiewood-Paley type estimates (see, e.g., Frazier, Jawerth, and Weiss (1991)).
"'FI
-
49
THE CONrPINIJOUS WAVELET TRANSFORM
The wavciet transfornr cart also be uscd to charact.ctcrizelocc~lregularity, sorncthing that carlrlot be achivvt.d, cven when phase inforr~~at~on 1s taker1 irito a ~ count, by the windowed Fourier transfornl. Tile fidlowing two tlicorcms are again l~orrowedfrom Halschrtcider and Tchamitchiark (1990) THEOREM 2.9.3. Suppos~that d~ ( 1 -I-Irl) [+(x)] < DC) u ~ r dJ dx $(x) = 0. If a bounded furartzon f $ Hiildrr contmaorrs zn ro, 1~1th c q ~ u r r ~a n tE 10,1], r.e.,
theri
l- &
I(",$ U V ~ I tJt1) I <'c'f~1''~ ((cL~'' s
Proof. Dy transiatirlg everything we can
6%
< - C
J d.r
J.rlcx
Ibl")
~ s s t ~ r n rf rl ~ i r t :r,) --
0. Bcxause
(-;x - b ) 1
-
(lal" -Ilbl") .
- C' L
THEUREM 2.9.4. Suppose that 1/, is cornpartly .qul,yor ted. Suj~posealso that f t- L2(R)zs bounded and contznvous. If, for sonrr 7 IJ crntl ru C ]O,l[,
L
"*
I(f ,
@"qb)
1 5 CtuIVk1/2
unzformlg, u r b,
and 3.
$6 ' &,
X
1
then f t s Hiilder contznuous
'
271
.co urzth ercpcment a.
Proof.
e
C
i
1. The proof starts exactly like t h e proof of Theorcrn 2 9 2, ol which the first three points carry over without change, with 7 taking over the role of u irr point 3.
2. We therefore only have t o check Ijss(xo + h) - fss(xo)l for small h. By translating everything, we can assume xo = 0, and we obtain
'
h-b ~hla/~~~a~
50
CHAPTER 2
+
/
lhlllalll
$/-
-00
&(lala
+
-)
1+2
h-b
(TJ.-h
(-:)I
where we have assumed a > 7. (If a 5 7,things become simpler.) Let us denote the four terms in the right-hand side of (2.9.4) by TI, Tz, T3, and
T'
4. In the second term, we ase support & C
T-R,
R] to derive
5. Similarly, for sufficiently s d l Ihl,
5 Cfhp . f
6. Finally,
S
i theorerne f ~ higher r order local regularity can be proved. These theo-
rems justify the name "mathematical micmkope," which is sometimes bestowed on the wavelet transforrh. In Hdschneider and Tchamitchian (1990) these and other reeults were exploited to study the differentiability properties of the function defined by the Fourier series n-' sin(n2%z), first studied by Riernann.
xr=,
51
THE CONTINUOUS WAVELET TRANSFORM
Notes. -
I
1. All these B e r m spaces can be further transformed, via a (standard)
conformal map, to Hilbert space8 of analytic functions on the unit disk. 2. My main refkence for this paragraph am the articles by Grossrnann, Morlet, and Paul (1985, 1986). Their results can in fact be generalized to reducible representations as well, as bng as they have a cyclic vector (A. Grossrnann and T. Paul, private c o n m m i ~ t i a n ) .This is useful for the higher dimensional case, where the representations of the uz &group are reducible but cyclic.
+
3. The operator Pie obtained by "transporting" H to $he Bargmann Hilbert is particularly sirnpie: space, via the unitary map yh, h
= erp
1 (-;(d +
1')
- -ut 12(w + it)&(w 2
)
+ it)] ,
or ( H s @ ) ( z ) = 2z #'(z). It is obvious that the eigenfunctimk of HB are the monomials u,(z) = (2" ni)-'j2 zn. The following argument shows that these are indeed the fupctions in the Bargmantr space corresponding to the Hermite functions. One easily computes
$
1
1 = exp (-?(J t 2 ) - -ut 2 (-i)(w
+
c it) +C + it) ,
so that 4, .= (2" n!)-'/*(A*)"g,corresponds to (2" n!)-1/2(-i)nan = (-i)" &(r) in the Bargmann s&. (We use the normdization Ilt$llimgm. = / & J d p e-i(Z3fy') [WZ 4 i4l2,so that go itself corresponds to the constant function 1 in the Bargmann space.) In particular, this means that
&
['
+
-
(& ,g"*t) = exp - (a2 t 2 )- i w t ] (-ijn (zn n!)-lI2(w
+ a). .
4. Not every unbounded &tion leads to an unbounded operator; m e bounded operators can only be represented in this way if an unbounded weigtrt h e t i o n ie wed. In fact, KIauder 11966) proved that even some tr-eh operatars require a non-tempered distribution as weight f&tion! '
5.
Far 4 functiom w, one requires that W should be esaentidly self-adjoint on -this *domain.
,
CHAPTER 2
52
6. Yet another application in quantum nlechanics can be found in Daubechies and Klauder (1985), where it is shown how t o write the (mathematically not well defined) path int.egral for exp (-ith-) a s a limit of bona fide Wiener integrals (as the diffusion constant of the underlying diffusion process tends t o oo), provided H is of the form (2.8.2), with a weight function w(p,q) that does not increase too drastically for p, q-too. A similar theorem can be proved in the wavelet case (Daubechies, Klauder, and Paul (1987)).
7. Exactly the same arguments hold for all operators W of type (2.8.2) for which u(w, t) is rotationally symmetric, even if it is not an indicator function. An example is W(Y, t ) = exp[-cr(w2 + t2)J,fqr which it was first shown in Gori and Cuattari (1985) that the Hermite functions are the eigenfunctions (irrespective of a; the eigenvalues depend on a, of course!). 8. It is no coincidence that Fefferman and de la Llave would use a representation of type (2.8.3) for the operator (2.8.4): after all, Calder6n7sformula (to - which (2.4.4) is essentially equivalel;t) is part of a toolbox developed precisely for the study of singular integral operators (long before wavelets!) so that it is well adapted for treating the smjlar kernel in (2.8.4). In this particular instance, (2.8.4) makes sense even for nonadmissible (C+ cancels out); in Fefferman and de la Llave (1986), 2/, was taKen t o be the indicator function of the unit ball (which is nonadmjssible, since its integral does not vanish).
9. If we make an extra transformation, mapping the upper half plane {b +*in; a 2 0) to the unit disk (by means bf a conformal mapping), then everything becohes more transparent: the flow r-+z(t) then corresponds to a simple rotation around the center of the disk, and H as well as its eigenfunctions are given by simple expressions. See Paul (1985) or Seip (1991).
10. There exist many other choices of~!,t for which this analysis works. For each choice the set sc in time-frequency space corresponding t o Sc in (a,b)-space takes on adifferent shape. Explicit computations and a figure illustrating these different shapes can be found in Daubechies and Paul (1988).
CHAPTER
3
Discrete Wavelet Transforms: Frames
In this, the longest chapter in this book, we discuss various aspects of nonorthonormal, discrete wavelet expansions, tdgether with some parallels with the windowed Fourier transform. The, "frames" of the chapter title are sets of nonindependent vectors; they can nevertheless be used to write a straightforward and completely explicit expansion for every vector~inthe space. We will discuss wavelet frames as well as frames for the windowed Fourier transform; in the latter case, the'approach can he viewed as "oversampled" with respect to the Nyquist density in time-frequency space. A lot of the material in this chapter has been taken from Daubecliies (1990), updated here and there. A very nicely written review of frames (and of the continuous transforms as well), with some additional original theorems, is Heil and Walnut (1989).
*
'
d
u G ~
,t 2.i
,
,
2.
3.1.
Discretizing the wavelet transform.
In the continuous wavelet transform, we consider the family I
$"J'(x)
=
Ial-'/2
$
('0" --
, where b E W,,a E R+ with a # 0, and Q is admissible. For convenience, in ;". the discretization we restrict a to positive values only, so thal the admissibility '* 3 - condition becomes
q" I,
'
(see $2.4.) We would like to restrict a, b to discrete values only. The dlsaetiza-
* ' {ion of the dilation parameter seems natural: we choose o = a r l where m E 2, h d the dilation step a0 # 1 is fixed. For convenience we will assume ao > 1 @though it does not matter, since we take negative as well as positive powers m). For m = 0, it seems natural as well to discretize b by taking only . +he integer (positive and negative) multiples of dne fixed bo (we arbitrarily fix
$ > O), where 60 is appropriately chosen so that the $(z - nb) 'cover" the whole line (in a sense t o be made precise below). For different values of rn, the
-
CHAPTER 3
-
width of aim'2 + ( a L m ~is ) a F times the width of *(z) las measured, e.6, by width (f) = [ J d x x21f ( x ) [ ~ ] ' /where ~, we assume that $dz xi f = 0)) so that the choice b = nh aom will ensure that the discretized wavelets at level rn "cover" the line in the same way that the $(x - nb) do. Thus we choose a r aom,b = nboar, where m,n range over Z, and > 1, bo > 0 are fixed; the appropriate choices for ao, bo depend, of course, on the wavelet 1/, (see below). This corresponds to
(%)I2
We can now ask two questions: \
completely characterize f? Or, stronger, can we reconstruct f in a numerically stable way from the
(1) Do the discrete wavelet coefficients {f,Q,,,)
(f +m.n>? 1
(2) Can any function f be written as a superposition of "elenlentary building Can we write an easy algorithm to find the coefficients in blocks" I),,,?' such a superposition?
In fact, these questions are dual aspects of only one problem. We will see below that, for reasonable 11 and appropriate ao, bo, there exist $J, so that the answer CI
to the reconstruckion question is simply
It then follows that, for any g E
L2(R)
-
or g = En,,,(g,l/,m,n)I)m,n, at least in the weak sense;-this is effectively a prescription for the computat~onof the coefficients in s miperposition of ~ m . , leading to g. We will mostly focus on the first set of que~tionshere; for a more detailed discussion of the duality between (I) and (2), see Grijchenig (1991). In the case of the continuous wavelet transform, both quttstbns were answered immediately by the resolution of the identity, at least if was admissible. In the present discrete case there is no analog of the resolution of the identity,2 so we have to attack the problem some other way. We c8n aku, wonder whether there exists a "discrete admissibility condition," and wh& it is. Let us first give some mathematical content to the questions in (1). We will restrict ourselves mmtly to functions f E La(R), although dtrcrete families of wa&leta, l i i their continuously labelled cousins, can be used in many other h c t i o n spaces as well.
+
DISCRETE WAVELET TRANSFORMS. FRAMES
55
Functions can then be "characterized by means of their "wavelet coefficients" (f,$,,,) if it is true that *
(fit 1Clm,n) = (fi,
h,n)
for dlmtn E
implies fi s fi
,
or, equivakntly, if
,
(1,+m,n)=O
forallrn,n~Z=+ f =O.
A
But we want more than characterizabirity! we want b%@able t o reconstruct f in a numericdly stable way from the (f, Ilr,,,).In ord& for such an algorithm to exist, we must be sure that if the sequence ((fi, &,,n)),,nez is "close" to 4(f2, $m,ni)m,nEZcthe0 necessarily fi and f 2 were "close" as 4.In order to make thig,precise, we need topologies on the function spa& and qa the sequence space. P t h e function space L 2 ( R )we already have its Hilbert space topology; on tb-uence space we will choose a similar t2-topology, in which the distance sequences c1 = (c&,,),,,,,~~and 2 = (c$,,),,,~~ is measured by
1 4
"
- xe This implicitly assumm that the sequences ((f, $m,n))m,nEzare in C(Z2)thema
,"
mives, i.e., that Ern,,((f, &,,)I2 < oo for all f E L2(R). Ln practice, this is . bo problem. 'AS we will see below, any reasonable wavelet (which means that bas some decay in both time and frequency, and that J & * ( x ) = O ) , and any ' choice for a0 > 1, 4 > 0 leads to **
3
,<*bf
C I(II
< B IIfI12 .
@m.n)P
(3.1.2)
m,n
We will assume (without specifying any reetrictions yet on the $,,,; we will come back t a these later) that (3.1.2) holds. With the t2(z2) interpretation of ucloseness," the stability requirement means that if Em,,I(f , +,,,)I2 L small, n Uf 112 should be small. In particular, there should exist a < oo so that ,n [(f, $,n,n)I2 5 1 implies 11f ]I2 6 a. Now take arbitrary f E L~(W),and
< a. But thii means -1
$5 fbr dome A = a-I > 0. On the other hand, if (3.1.3) holds for all f, then the *+ distance Ilh -fzII emnot bear bit rail^ 1~ if Ern,,l(fi, - (k+ (Im,n)Ia
57
DlSCRETE WAVELET TRANSFORMS FRAMES <
I t follows that {el, e2, e s ) is a tight frame, but definitely not an orthonarr~)aJ basis: the three+vectorsr 1 , e2, e j are clearl~not linearly ~ndependent o
Fw
31
T h e ~ elhne vectors an C' m t c t u t e a laghi @me
4
Note tb&-in thls exn~rlplethe frame bound A = gives t h "redundancy ~ ratie" (three vcctorv In a two-dimensional space). If this redundancy patlo, as measured by A, IS equal to 1, then the tight frame is an orthonormal bmls PROPOS!T~ON 3 2 1 If (pJ)3)3E J t8 a hght h m e , wath frame bound A = 1, r f IIpJ11 = 1 for all 3 E J , then the pJ conahlute an orthonormal basas i Proof $rice ( j , cp,) = O for all g E J ~rnpliesf = 0, the vJspan all of 3.1 It -+ &ains to check that they are orthonormal We have, for any j E J , %
3'EJ
Since
Ih,n+12.
I(%, vJ~)I2 = ilnl14+
I I P ~ I I ~=
J'#J
a'€
llvJ11 = 1 , thls implies ( Q ) ,
fit)
= 0 for all 3'
J
# 3.
r
9
Formula (3 2.2) ~ v e as trivial way to recover f from the ( f , cp,), if the frame . Let us return to general frames, and see how things work there We troduce the frame opemtor. E F I N ~ O NIf ( g 7 1 ) I E J ~d a fMme SR 'F1, then tke b m e opero'w F as the JcJ < m) defined r opemtor jran X to t 2 ( J )= { c = ( ~ j ) ) ~[lcQ3 j ; =
b b
JEJ
.
( F f ) , = (f,9,) It follows from (3.2 1) that llFf 112 -< B 11f 112, i.e., F is bounded. The adjoint d F is easy to compute: *-
(F*c,f) = ( c . F f ) = C c , ( f , v , )
Ti
C *
f€J
= Cc,((p,,f), JEJ
90 that. 1 k
F*C
,-
=
C c,w
JEJ
,
6tl
DISCRETE WAVELET TRANSFORMS: FRAMES
g the operator (F0F)-I to the vectaq R y of vectors, which we denote by qj,
m W
q, = ( F * F ) - I
b
4"
$W %
-
i
8-'
I l f 112
l(fs$91'< A-I 4fHa -
5
(3.2.6)
3e J
i
%
vj .
The family turns out to be a frame aq well. PROPOS~TION 3.2.3. The ( @ , ) l E J c m t a t u l a ,fmme wth jhme m t a n t J B'landA-',
t. f
to an i n w h g W~
The assocsatsd frome operator : li --+ P(J), ( ~ f ) ,= dl, +,) satuficd P = F(F*F)-1, PP = ( F * F ) - l , PF = xd = F'P ~d FF FP' rs the mulogod pqstam opmtm, m P ( J ), onfo Ran ( F ) = Rr. (P) ,
'
.
17
1.
As an exercise, the reader can check that if a bounded operator S has a bounded inverse S-', and if S* = S, then ( $ - I ) *
= S"- It follows that
wce
v
51 G
- 2 f
l ( f l +,)I2 =
+
I((F*F)-'h
~ $ 1=~IIF(F*F)-'fli2
JG J
3E.J
= ( ( F ' F ) " ~ , F * F ( F F J - ~=~ () ( F * F ) - ' ~ j, )
.
13.2.7)
By (3.2.5),this implies (3.2.6); the 3, constitub a frame. Moreover, (3.2.7) implies dso that the frame operator @ atisfie P P = (F'Ff'l, i
2. (F(F*F)--'I),= ( ( F * F ) - i f ,9,)= V, @$) = (Pi),, P F= [ F ( F * F ) - * ] * F.;(F'F)-'F*F = M, F*F = F*F(FWF)-'= Id. also
QM
f
'2-
t-t +*\
-
+ ,&
+?tb
XI
bc
4
*
F
F F ' F ~= F ( F F ) - I
$ g.
.
(g
Ei
-
= F ( F * F ) - ~it, fnliows tha$ Ran c Ran (F). We have = P ( F * ~ ' hence ) ; Ran (F)C R8n (P). Consequently, F&n ( F ) Ran (F). Let P be the orthogonal projection operator onto Rgn (F).We wmt to prove that FF' = P, which is equivalent .to P F * ( F ~= ) Ff (i.e., BF* leaves elements of Ran (F) unchand) and PF'C = 0 for all c orthogonal t a Ran (F).Both assertJon% are easily cb&d
3. Since
:&cT
6
F * F= ~ ~f
*
ad cLRan(F) 3 (c, F f ) = O forall f E'H =+F'c=O #F*c=o.
60
CHAPTER 3
4
We will call J the dual frcrme of (cp,)J , J . It is easy to check that the dual frame of is the original frame ( c p 3 ) J c I hack again. We can rewrite some of the conclusions of Proposition 3.2 3 in a slightly less abstract form; P F = Id = F*F means that ,
This mems that we have a reconstruction formula for f from the (f, cp,)! At the same time we have also obtdned a recipe for writing f as a sttperposition of cp,, which demonstrates that the two sets of questions in $3.1 are indeed "dual." When given a frame ( P , ) ~ ~the~ only , thing we therefore need t o do, in order t o apply (3.2.8), is t o compute the cZr, = ( F * F ) - I rpj. We will come back to this soon. First we will address a question that often arises a t this point: I have stressed before that frames, even tight frames, are generally not (orthonormal) bases because the cp, are typically not linearly independent. This means that for a given f , there exist many different superpositions of the cp, which dl add up t o f . What then singles out the formula in the second half of (3.2.8) as especially interesting? We can get an inkling of the answer with a simple example. EXAMPLE. We revisit the simple example of Figure 3.1. We had there, for any v E c*,
3 Since C,=l el = 0 in this example, it follows that the following formulas are also t r u e
where a is arbitrary in C. (In this particular case, one can prove that (3.2.10) gives d l the possibie superposition formulas valid for arbitrary v . ) Somehuw, (3.2.9) seems more "economical" than (3 2.10) if a # 0. This intuitivestatement can be made mbre precise in the following way:
whereas
Likewise, the ( f , off into pj.
cii,) are the most
PROPOSITION 3.2.4. I f f af not alict eqwl ( f ,
@I),
"economical" coeEcients for a decompoeition
El,,
c, 9, f o r some c = (c,),~J E t ? ( ~ ) and , h n CjEJ Icj12 > CIEJ I(f, @j)I2=
, ",
61
DISCRETE WAVELET TRANSFOqMS FRAMES
1 Saying that f =
d
'
El,,c, cp, is equivalent to staying that f = F'c
2. Write c = a + b, where a E Ran (F)= Ran particular, a l. b; hence /(cl12= Ital12 Ifbf12.
+
(81,and b l. Ran ( F )
.,
In
+
.rl
3. Slnce a E Ran (F), there exists g E 31 so that a = pg,or c = Rg b. Hence f == F'c = F * ~ C JF'b + But b I Ran (F), so that F'b = 0, and F*F Id. It follows that f = g; hence c -i ~f b, and
+
,
4
-
CJEI( f, $,)12,
which is strictly larger than
unless 6 = 0 and c = Ff .
-.
This proposition can also be used t o see how the g3 play a special role m the fifit half of (3.2.8). We typically have nonuniqueness there as well: there may #st many other families ( u , ) , ~J SO that f = El, (f, cpl) u,. In our earl~er two-dimensional example, such other families are given by u, = +el a, where d is M arbitrary vector in c2.Since 3 e, = 0,we obv;ously have
+
r,=,
L,
3
2 )uJ = 3 ]=I
C(v,e])e] +
are "less economical" than the E,, in the sense that for 3
4)12
I(v,
+
%)I2
+3I(v,a)l2
,=I
2 Ilvl12 + 31(v,a)12 > 51tvl12 =
-
3
.,,I2 J=el
ilar inequality holds for every b e : if f =
CIE (f;
p,)ul, then J [(u,. g)I2 g)I2 f o r d l g E W , byPropasition3.2.4. , to the reconstruction issue. If we know $J = (F'F)-'q,, then (8.2 8) .'$@$, ~ie8hi.mhow to reconstruct f from the (f,v,). SOwe only need ti3 compute the +%@p+which involves the inversion of P F. If B and A are close to each other, i.e., B/A - 1 < 1, then (3.2.4) tells us that F*F is "close" to ~ db , that 1 W ~ F ) : is' UclOse"to Id, and @I "close" to ip3. More precisely,
tx,EI
*;,
%. .'3
#'t,; :".-
-a4 i
yr
-&
f=-
2
A+B
C (f,R)Vj + Rf , IEJ
(3.2.11)
CHAPTER 3
62
-m
Id.%This impliib where R = Id - 2 F'F; hence Id 5 R 5 B - A = &. If r is mall, we can drop the rest term Rf in (3.2.11), 1JRll I and we obtain a reconstruction formula for f which is accurate up .to an Laerror of 11f U. Even if r is not so small, we can mite an algorithm for the reconstruct~onof f with expoaeatial convergence. With the same definition of R, we have
+
A+B F'F = (Id - R) ;
-& -
2
<
hence (F*F)-' = (Id R)-'. Since \(RII & B < 1, t h e S e r b converges in norm, and its limit is (Id -R)-' . It followi that
Using only the zeroth order term in the reconstruction formula leads exactly to (3.2.1 1) with the rest term dropped, We obtain better appraximtions by truncating after N t e r n ,
with
as N increases, since &F which becomes exponentially the +? am be computed hy an iterative algorithfn,
< 1. In particular,
I
1
DISCRETE WAVELET' TRANSFORMS: FRAMES
with N
Q,!
2 =A + B 6cl +
N-1-
2 x
A+B
0;;'
( ~ m M , )
rnEJ
This may look daunting, but it is not so t e m t k in examples of practical interest, where maxiy (cp,, pc) are negligibly small.vThe same iterative technique can be applied directly to f :
f = ( F F ) " ' ( F ' F ) ~= N--.lim with
fN,
-
%OW that we have thoroughly explored abstract frame questions,-we return to
3.3. F'rames of wavelets. a numeridy stable reconstruction algom for f f r w the ( f , &,,), we require,iIlat $be $Jm,* constitute a frame. In we found an algorithm to reconstruct f from the (f, &,n) if the ,,$ , do ; for this-algorithm the r&io of the &frame bounds is important, k to ways of computing at least a bound on this ratio, later this section. First, however, we shm that the requirement that the ,,$ , saw in 83.1 that in order to have
nstitute a frame already imposes that qb is admissible
Admkeibility of the mother wavelet. --mi2 #(Grnx - nbo), m, n E Z, constrTHWREM 3.3.1. Jj the $,,,(z) =% fe a frome for L2(R)wtth frame bounds A, B, then y condition:
A9
1O"
# [-I
!$(F)\~< bolnao T B
(3.3.1)
(3.3.2) -w
64
CHAPTER 3
Proof. 1. We liave, for all f E LZ(R),
If we write (3 3 3) for f = uc, and add all the resulting inequalities, ~ 1 then ) ~ we weighted with coefficieiits cc 2 0 such that C, ~ ~ 1 1 <~ oo, obtain ..
In particular. then
if
C is any positive trace-class operator (see Preliminaries),
where the at are orthonormal, cc 2 0, and Cp,, cp = R C > 0. For any such operator, wc I ~ a v etherefore, by (3.3.4),
2 We now apply (3 3 5) to a very special operator C , constructed via the contznuous wavelet trallsforrn, with a dzflerent mother wave&. Take h to be any L~-functionsuch that support h c 10,oo), c& f-' fh(<)12< m, and define, as m Chapter 2, halb = a-'I2h (yb) for a, b E R, a > 0. If c(a, b) is a bounded, positive function, then
is a bounded, positive operator lsee 82.8). If, moreover, c(a,6) is integrable with respect to a-2 da db, then C is trace-class, and W d = f,"$ $-", @ c(a, b)llh112.6We will in particular choose c(a,b) = w(lbJ/a)if 1 <_ a 5 w, 0 otherwise, with w positive and integrable. We then have
db (-, hUvb) ha*' w and ds ~ ( J s 1hlj2 j ) = 2 In ao
):(
,
-
65
DISCRETE WAVELET TRANSFORMS FRAMES
3. The middle term In (3 3 5) becomes, far this C ,
z(c$rntn7 h , n ) = c m,n
m,n
/l
rnda
3 / O-mD d b w ( f )
(+mn,
h&)l2
But - 4 2
($~rn.n,h~'= ~)
=
I
a
-1/2
ak
@(U,~T
aom/2a - l I 2 Jdy +(y) h
(
- nbo) h y+nbo - bairn I
aim
Alter the change of var~ablesa' = airnu, Y = a i m b we thrhoSe obtain
2
2 2
Take now w(s) = X e-A " This function has only one local niax~rnurn and is monotone decreasing as Jslincreases. An elementary approxlni~tlol argument for integrals (the full details of which can be found in Daubechle (1990),Lemma 2 2) shows that for such.functions w and for any a,0 E R P > 0,
or, for our particular w ,
I
r;
with Ip(a, b)l
where
I w(0) = A. Consequently,
4
CHAPTER 3 which Ch as defined by (2.4.11. We can rewrite the first term in (3.3.7) a s
4. For the particular weight k t i o n w that we have chosen, we have dt w ( t ) = $,hence Tr C = llhIl2 1" ao. Substituting all our results in (3.3.9) we find
3
where IRI 5 X Ch l($l12. If we divide by llhllZ and let A tend to zero, then this proves (3 3 1). The negative frequency formula (5.3.2) is proved analogously. s
1 Formulas (33. I), (3.3.21 impme an a priori restriction on 11, namely that (-I < m and fwq1 ~ I - l < m. This is the aomc restrict~onas in the continuous case (see (2-4.6)).
JFe
d(<)12
14(<)/~
,+,
2 In defining the d~scretelylabelled we only took posrttue dilations a$' into consid6ratlon (the sign of rn affects whether af; is 2 3 or 1, but, a r > 0 for all m) Thii is tbe reason why formillas (3.3.1), (3.3.2) dlssoclate the positive and negative frequency domains. ff we had allowed negative d i i t e dilations as well+ then &hecondkion would b v i ? involved only 4 Kf-I M(t)12 (as bs easy to check by mimicking the above proof).
<
J-T
3. If the
,,+,
constitute a tight frame (A = B), then (3.3.1), (3.3.2) imply
constitute an orthonormal basis of L2(R)(such In particular, if the as the Haar basis, or other bases we will encounter), then
It is an easy exercise that the Baar basis does indied satisfy (3.3.8). Most of the orthon~rmalbases we will w i d e r are real, so that the h t equality. in (3.3.8) is trivially satisfied. J
4, A difierent proof of Propositioa 3.3.1 is given in Chui and Shi (1991). o
In all that fdm,sre wiU alwqys m u m e that fB is admiiible.
S
DISCRETE WAVELET TRANSFORMS- FRAMES
67
3.3.2. A sufficient condition and e s t b a t e e for the frame bounds, Not all choices for $, ao, 4 lead to frames of wavelets, even if $ is admisaible. In this subsection we derive some fairly general conditions on 9,m, bo under which we do ~ndeedobtain a fnune,$andwe estimate the corresponding frame bounds. To do this, we need to-estimate Em,, I ( j , tltm,n)12:
(by Plancherel's theorem for pertodic functions) 2.n = j ( t )f(< + 2 ~ k a ; ~ b o -i(go l) ~ ( 42nkb0'') t
I
I
1I ::, ir
c 114
60 m,&Z
+
Here Rest (f) is hounded by
I-.;".
(use Cauchy-Schwarz, and change vai-iables C = ( 2 n k b i 1 ~ "in the &nd facpx)
-
(~se Cauchy-Schwarz
on the swa over m)
.
CHAPTER 3
where B ( s ) = supc
CmEZ I&(GS)(t i ( a r ( + s)l. Putting (3.3.9) and (3.3.10)
together, we see that7
If the right-hand sides of (3.3 ll), (3.3 12) are strictly positive and bounded, then the 2hrn,- constitute a frame, and (3.3.11) givcs a lower bound for A, (3.3 12) an upper bound for 13. To make this work, we need that, for all 1 5 5 ao (other values o f t can be reduced to this range by multiplication with a suitable a?, except for 4 = 0;but this co~~stitutes a .set of measnre zero, and therefore does not matter), 0< < l r i ( ~ 021 B~< ; m€Z
xmEz l$(t$e))
moreover, ~$(~$e+s)lshould have sufficient decay at oo. "Sufficient" in this second condition means that [@( k ) P ( - $k)] ' I 2 converges, and that the sllm tends to 0 as tends to 0, ensuring that for small.enough 4 the first terms in (3.3.11), (3.3.12) dominate, so that the &,,, do indeed constitute a frame. In order t o ensure all this, it is sufficient t o require that the zeros of
x,;,
2
4 d o not "conspire," so that
for all ( # 0,
14((>(5 C(6la(1 +
with a > 0, 7 > a
1~(~)-7/~,
4
+ 1.'
(3.3.14)
These decay conditions on are very weak, and in practice k e will require much more! 1f is continuous, and decays at oo, then (3.3.13) is a necessary condition: if, for some b # 0, Xmfz\4(cb)12 5 c, then one can construct f E LZ(R), with I( f (1 = 1, so that (2n)-'h Ern,,((f, qm,n)I2 < Ze, implying A I 47r c/bo.'
4
69
DISCRETE WAVEZET TRANSFORMS FRAMES
If c can bechosen arbitrarily small, then there IS no finltc lower frame bound ( S e e also Chul and Shi (1991), where the stronger result A C, ~$(fl<)v< B is proved ) The following proposition summarizes our findings PILOPOSITION 3 3 2. If +, ag are 9uqh that
<
SUP
C
lSl
I$(G'<)I~ < ,
(3 3 15)
and zf A(.) = supr C, i$(ar()~ I$(@< + s)l d a y s at least as faqt as (1 + ls()-('+'), wtth c > 0, then there eztsb ( h l t h r > 0 such that the 1/,, constztute a jhme for all chosce,e < (bo),,,, For bo < (bOftM,the follounng ernresstons are fnrme bounds for the @,
( ~ = z{ '! A = %bo
-P
bo
2
~ < Iinf C I < ~ O m = - m ~ 4 ( a ~ k =~ oo~ ~2 (-
b ) ~ ( - ~ k ) ] ' " } ,
L f0
2
14~S0i2 +
I S I ~ I SmZ-00 ~W
2 b(3)
0( - ~ k ) ] ~ " }
k = -m
*#'J
+
The condztrons on 0 atad (3.3 15) are satsfied zf, e 9 , ,&<)I _< CI0,7>a+1 P m f U'e have dready carried out all the necessary estimates The decay of 0 ensures the existence of a (4)thc SO that Ek+o zk)]1/2 < inflll~lsao
x* I & ~ ~ c ) I ' lf
60 < (bo)thr.
[fl(3k)@(-
8
The moral of these technical estimates is simple. if y!~ is at dl "decent" (reasonable decay in time and frequency, J & + ( x ) = O), then there exists a whole range of %, so. that the corresponding $m,n constitute a frame Slnce our conditions on~!,t imply that -$is admissible in the sense of Chapter 2, this is not so surprising for values of ao, h~close to 1 , O respectively: we already know that the resolution of the identity (2.4.4) holds for such and it is reasonable to expect that a sufficiently fine discretization of the integration variables should not upset the reconstruction too much Surprisingly enough, for matly of practical interest, the range of "good" (m, 4) includes values which are quite 9v far from (1,O). We will see several examples below. But first we will look at ( t h e dual frame for a frame of wavelets, and discuss some variations on the basic scheme.
+,
J!+I
+
P
8.8.3.
~he'dualframe. As re saw ln 53.2, the dual lrame is defined by
CHAPTER 3
70
where F * F f = Ern,,(f,),$ ,, $m,n. We have an expucit formula for the inverse of F*F which converges exponentially fast, i.e., fke an,with a convergence ratio a proportional to ( f - 1). It is therefire usefol to have frame bounds A, B which are close to each other. Nevertheless, (3.2.8) necessitates, in principle, an infinite number of qb,,,,,to be computed. The situation is not quite as bad as one might expect: if we introduce the notation
EL,
-
then it is easy to check that, for all f E L2((R),
=
Jt follows that (F*F)-' and Dm cammute as well. In particular, since DmFg, ' = (F*F)-I D m P + = Dm (F*F)T' T n g ,
-
C
-
Unfortunately, F*Fand T" do not commute, so that we still have to compute, in principle, infinitely many &,,. In practice, one is interested only in functions "living" on a finite range of scales, on which F'F can be rearronably approxC (., +m,n) Ilm," (see the time-frequericy localization imated by CzL, section below, 53.5). l ~ ~ is an ~ integer, ~ tN =- a~'l-"'O,~ then one easily * checks that this truncated version of P F commutes with T ~so,that one only has to compute the N different q&, 0 < n N - 1 in this case. This number is still very large in many cases of practical interest, howewr. It is therefore especially advantageow to work with frames which are almost tight ("snug frames"), i.e., which have $ - 1 < 1: we can then stop after the zeroth order term of the reconstruction formula (3.2.1 l), avoid all the complications with the dud frame, and still have a high quality recomtruction of arbitrary f. On the uther hand, there exist very spacial choices of $, a.~,bo for wbich the &, are not c b t o a tight frame, but it so happens that all the . $ are generated by a single function, (3.3.17) +yn(2) = t,&m,n(~) = Q ~ ' $(O;"X ' - nf .
<
An example is provided by some of the biorthogonal bases that we will encounter in Chapter 8; andher example is given by the &transform of Frazier and Jawerth (1988) (see also Frazier, Jawerth, and Weiss (1991j).AC It is important t o realize that the &, and the $m,n may have very merent regularity properties, For instance, there exjst frameswhere $ itself is Cooand decays faster than any inverse polynomial, but where some of the are not in LP for d l p (implying that they have very slow decay). An example, due to P.,G. LemariB, is given in detail in Daubechiea (IM), pp, Q8&-!W.l0 Something similar may happen even if all the are generated by a six& b c t i o n 4: there exist examples where 3 E Ck (wit4 k arbitrarily large), but where is
Kn
$zm
4
DISCRETE WAVELET TRANSFORMS FRAMES
71
not continuous (The biorthogonal bases in Chapter 8 gtve examples where this happens; the first example was constructed by Tchamitchian (1989).) One can exclude such diiunilarities by imposing extra conditions on +,m,and bo (see Daubechies (19901, 5 II.D.2, pp. 991-992).
Some variations on the basic scheme. So far, we have not restricted the value of @, beyond a0 > 1. In practice, however, it is very conve nient to have a0 = 2. Going from one scale to the next then means doubling or halving the translation step, which is much more practical than if another ao is used. On the other hand, we have just seen th8t it is advantageous to use f r m with B / A - 1 < 1. Since our estimates (3,3.11), (3.3.12) for A, B give 3.3.4.
for all [ # 0, these two requirements together mply that CmEZ 14(2~<)1~ should be almost' constant for = 0, which is a very strong restriction on $, not generally satisfied. The Mexlcan hat funaion $(z) a (1 - x2) e-"I2, for instance, Lads to a frame with B / A close to 1, for ao 5 z ] / ~ but , certainly not for rn = 2 because the amplitude of the millations of 14~(2"())~ is too large. In order to remedy this situation, without having to gve up too much of our freedom in chooging a d its width in the frequency domam, we can adopt a method used by A Grossmann, R. Kronland-Martinet, and J. Morlet, and use different "voices" per octave. This amounts to wing several different wavelets, qbl, - - , and to look at the frame {@&,; m,n E 2, v = 1,. . . ,N). One can repeat the analysis of 53.3.2 (see, e.g., Daubechies (1990)), leading to the following estimates for the frame bounds of this rnultivoice frame:
XmEt
+
+*,
with
N
and
-
00
By choosing the $I, .,4n to have slightly staggered heqiiency localization centers, coupled with good decay at c6, one can achieve B / A - 1 << 1. (See the e$icamples in 93.3.5 below.) The time-fresmcy lrrt prresponding to d s m u i t i w i w 5 ~ a n es-Ettle ~ ~ werent b 4 + 1 . 4 q m exampla, with a
CHAPTER 3
FIG 3 2 The t ~ r n e : + ~ latttee for a scheme m t h four vorces I n thu w e It daflec-ent vorce waveleb ~ , l - - ~ , ore arsumed to be drlatrons of a mngle functwn JI, J s ( s ) 2 1 4 ( 2 ~ 1 sf ]g(t)l 4 ( w h d we assume to be even) peaks armrnd fwo, then M 143 1 unN be concenbmted anncnd f2 - ( ~ - ' ) / ' ~
,+'
four voices per octave, is ginren in Figure 3.2. Fof every dilation step, we find four different frequency levels (corresponding to the four differen4 frequency localizations of - - -,$4), all translated by the same translation step. Such a lattice can be viewed as the sujxqosition of four diierent lattices of the type in Figure 1.44 stretched by different &nts in the frequency direction. Each of these four sublattices has a Merent "density," which is reflected by the fact that typically the gDY have different Lz-norms. One choice favored by Grossm a n , Kronland-Martinet, and Morlet is to take "fractionallyn dilated versions of a single wavelet $:
(Note that these do indeed have different L2-norms!) In this cam N ~ simply Cz,=-, l i ( 2 m ' / N ~ ) ~and 2 ) this Cu=1 Cm=-ac 1 @ ( 2 ~ < ) 1bemrnes can easily be made to be almoet constant, by choosing N large enough.
DISCRETE WAVELET TRANSFORMS FRAMES
73
Fixing a0 = 2 allows also for a modification of the estimation techniques in $3.2,which may be useful in some instances. Let us go back to the estimate for Rest (f). We can rewrite k E Z, k # 0, as k = 2'(2k1 I), where P 2 0, k' E Z; the correspondence k-+(t', k') is one-teone. If ao = 2, then we can regroup different terms, and write
+
Rest (f) = -
c 2
jd~j~c~j~t+2+(2k~+i)~-!.~-m~)
ml.kttZ
~(zm'+fc)$/2t(2n1{ + 2
~ ( +2- l~) b 0 - l ) ] .
t-0
This leads to
where
These estimat'es are due to Ph. Tchamitchian. (Nl details of the derivation can be found in Daubechies (19901.) Note that &, unlike p, still takes the phases of into account; as a result, the estimates (3.3.21), (3.3.22) are often better than (3.3.11), (3.3.12) when is not a positive function. If is positive, then (3.3.11), (3.3.12) may be better. The estimates (3.3.21). (3.3.22) hold it we have one single voice per octave; they can of course also be extended to the multivoice
4
*vv
4
4
Case.
*
"4& ~
Y ."
A
3.3.5.
Examples.
d
A. Tit frames. The following construction (first proposed in Daubechies, G r m m n , and Meyer (1986)) leads to a family of tight wavelet frames. Let u bea Ck (or CQ))hnction from R to R that satisfies:
CHAPTER 3
74
(see Figure 3.3). An example of such a (C1)function v is
For arbitrary no > 1, bo > O we then define
4*(<)by e ~ t ~ t ? ~ ~ e ,
4+(6) = [In
- l l2
1
tIt
whcre 4 = 2n[h(ao2- I)]-', and $-(<) = G+(-(E) Figure 3.4 shows = 1 , and w as in (3 3.25). It is easy to check that
d+ for
erg = 2, 6"
where ~(o,,)
(c)
is the indicator function of
tfw, Q W I ~half
line (O,oo), i e.
x ( o , ~ ~ ) = 1 if 0 c < w, O otherwise. For any f E L2(R),one then has
hc. 3 3. The f i e v(a) defined by (3.3 25).
.
b+.
DISCRETE WAVELET TfUNSFORMS: FRAMES
If,
5,
= rn,n E Z, r = , ) J 2
i+*(.>l5 cnr a
%", I'
..$+ ?;? ;'ii c
J
4I ) .
It follows thet the + or -) is a tight frame for L'(w), with Frame bwnd One can uhe a variant to obtain a frame consisting of real wavelets: @ = Re $+ = I[@+ + and $2 = Cm @+ = - $-I gener&e 2 the tight 'frame {$&,, m, n E Z, A = 1 or 21. These frames h e not generated by translations and dilations of a single function; tflisis a naturel-consequence of the decoupling of positive and negative frequenctes in the construction. A more serious objection to their practical use is the fact that their Fourier transforms' are compactly supported, and that the sim of this support is relatrvely smdl (for reasonable ao;bo)- As a result, the decay of the wavelets is numerically rather alow: even though we may choose v to be Cm,so that the @* decay faster tban any inverse polynomial, , collection (q!$m,,;
Y
E P
.
$.
(1
+ t4)-
I
the value of CN turns out to be too large to be practicd. Note that we did nut introduce any restriction M w,bo in thii construction. ' \
B. The Mexican hat b c t i o n . The Mexican hat
function is the second derivative of the C a w i a n e-;'f2; if we normalize it so that its L'-norm hr 1, we obtain $(z)" =-1/4 (1 - z2)e-s2/2
-
f
43
.
Thb -hition (and dilated and translated versions d it) was plotted in Figure 1.a; if p u take one such plot, and imagine it .anMm n d its symmetry axb,
76
CHAPTER
3
then you obtain a shape similar to a Mexican hat. This function is popular in vision analysis (at least in theoretical expositions), where it was also christened. Table 3.1 gives the frame bounds for this function, as computed from (3.3.19), (3.3.20), with a0 = -2, for d i f f k n t values of bo and for a number df voices varying from 1 to 4. As soon a s we take 2 or more voices, the frame may be considered tight for all d~ 4 .75. Note that bo = .75 and ( w , ) , ~ ~ ~= , , 2'12 1.41 (intuitively corresponding to two voices per octave) are not small values fbr the Mexican hat function: the distance between the maximum of and its zeros is only 1, and the width of the positive frequency bump of (as measured by 1-23. For 4 (C -
4
-
[JF
+
-
C. A modulated Gaussian. This is the function most often used by R. Kronland-Martinet and J. Morlet. Its Fourier transform is a shifted Gaussian, adjusted siightly so that 4(0) = 0, -
-
Often &, is chosen so that the ratio of the highest and the second highdt maximum of is approximately i.e. 6 = ~ [ 2 In/ 2]$12 5.3364. ..; in practice ,one often takes Q = 5. For this value of to,the second term in (3.3.26) is so small that it can be neglected in practice. This Morlet-wavelet is complex, even though most applications in which it is used involve onIy real signals f . Often (see, e.g., Kronland-Martinet, MorIet, and Grossmann (1989)) the wavelet transform of a real signal withihis cornpiex wavelet is plotted in modul&phaseform, i.e., rather than (f, rC,m,n), Im (f, $rn,n), one plots l(f, $m,n)l and tan-' [Im (f, q!~~,,)/R.e(f, &,n)J; the phase plot is particularly suited to the detection of singularitis (Grossmann et al, (19879). For real f , one can exploit j(-6) = j(() to derive the foUowing frame bounds (this is analogous to what was done in $2.4 for real f):
i,
$J
A
IU,k , n ) 1 2 < B Jlf 11'
llf 11' 5
-
m,n€Z
for f red
I
I
DISCRETE WAVELET TRANSFORMS: FRAMES
77
Fhune boundr far wavelet f m m based on the Meziwn h t functwn +(z) = 2/fi n-'14(1 The dalatum p a m e t e r a0 = 2 in all we;N u the number of voicw.
x2)e-"'12-
8
%, These can, of c o r n , again be generaized to the multivoice case. Table 3.2gives ?-itbe tame bounds for Q = 2, several choices of 4, and number of voicos ranghg from 2 to 4. In practice, the number of voices ia often even higher.
r*
CHAPTER 3
F h m e boundr for wavelet * e s
-
b a d on the modulated Gawaaon, #(s) =n-'I4 (Cctoz T h e m i o n conatant oo = 2 in dl -; N i s the
e L ~ g / 2 fe - r a t a , wth €0 = *(2/ ln 2j1/l. nu& of wt-s.
-.
N =3 bo .5 1.
A 10,295 ' 5.147 1.5 3.366 2. 2.188 2.5 1.1% 3. .320
1 -
B
BIA
1.017 1.017 . 1.056 3 0 2 1.372 z.& 2.534 3.01 ' 8.024
113,467 5.234 3.W
D. A n example t h a t is easy to implement. So far we have not a d d r d the qucstion of how the wavelet coefficients (f,q,,,) are computed in practice. In real life, f is not given as a function, but in a sampled version. Computing the integrals Jdz f ( x ) $m,n(x) then requires some quadrature formulas. For the smallest scaleti (most negative m) of interest, this will not involve many samples of f, and one can do tbe computation quickly. For larger scales, however, one faces huge integrais, which might considerably slow down the ~01fiputatioaof the wavelet transform of any given function. Especially for o n - h e hpkmentations, orle should avoid having to compute these long integr&. A constructir7n achieving this is the so-called "algorithme a trous" (Hobcb&de;r et al. (1989)), which uses an interpolation technique to avoid lengthy cornputeftions (for details, I refer to their paper). Here I propose an analogous example (although it is not "Q trous"), by borrowing a leaf from multiresolution 84aapm and o r t h i o n d bases (to which we will come back), i.e., by i n t r o d ~ ~ d nang auxiliary Function 4. The basic idea is the fdlowing: suppose there exists a fundion q%80 that
DISCRETE WAVELET TRANSFORMS FRAMES
79
where in each case only finitely many coefficients are different from zero." (Such pairs of 4, abound; an example is given below. The "algorithme ti trown wnesponds to special 4 for which 1, all other even-indexed Q, = 0.) Here 4 does not have integral zero (but @ does!), and we will normalize 4 so that J& fix) = 1. Define, even though 4 is not a wavelet, &,+(x) = 2'*J2 &(2'*x - n);we take = 2 , 4 = 1. Then it is c h that
? r
$J
& problem (f, &,)
of findii the wavelet coefficients is reduced to mrnputing the (finite combinations of which will give the (f, $,)).* On the other
sa that the (f; b,,)can bc computed rec~rrsively,working from the smallest scale (where t& are easy to compute) to the largest wale Everything is done by simple, finite convolutions An example of a pair of functions satisfying (3.3.27), (3 3.28) is
,;* 1
L
r-
I #gi
& t ) = - @ - 21 r€
6
' $z ,
which corresponds to
$+*$b
@++
*R
6
&). = b
4
4
,* 5
e
ilfZ h
*
j
"fisrrs p
P
%
I
-21~5-1, + 2j3 , 4 - x2(1 + x / 2 ) , -1 f x 2 0 , $ - x2(1 - 1/21 , 0 < t < 1 ,
$
(2:
- 2)3
lo
9
l s ~ l 2 , otherwise .
a 110rmdbaWn constant chaten a, that JJ$J) = 1; one finds N = B&. 3.5s shows graph of 11; they are not unlike a Gaussian and its second
-$bimtive, piotted in Figure 32%for comparison. The function JI clearly satisfies &3.27) with do = N,d+l = -N/2, all other dk = 0, whereas **
CHAPTER 3
80 which implies
or Cg = 23 , C*l = 21, cf2 = i,allother ck = O For this$,ao = 2andbo = 1, the frame bounds are A = 73178, fl = 3.77107, corresponding to B / A = 2.42022; for a0 = 2, b~ = .5 we I~avcA = 2 33854, B = 2.66717, and B / A = 1.14053 (using bo = .5 means that the rt~urslonformula, linking the, ,@, with the 4,, and the cP,,, with the 4,- I ,,, , havc to be adapted, but this is easy). Here we have usecl only one voice. It is of course possible to choose severd different ,'Q corresppndirlg to different d i , that give rlse to a muftivoice scheme, closer *. to tight frames. This conciu(les our exarnplc se~ttoli,other examples are given in Daubechics (1990) (includ~ngune for which the ~ s t i n ~ a t e(3.3.21), s (3.3.22) outperform (3.3 1 1 , ( 312)) Many otlicr exanlples car) of course be consbructd. Tile wavclcts used in Mallat arid Zhong (19%)) are another examplc of the same tyyx as our last or1.e; in their case is choscn ta bo the first derivative of a fu~lction with non~zerointegral (so that S d x Q ( x ) = 0 but t @ ( z )# 0).
(b)
1
0
FIG 3.5.
JL] .-- -
.-
- --
-----
A n eurmpk
1
-
--
o& [l
-
---
.
-
L-
that w m y to cmplement: graphs of $,$I (ana), and a cornparwon
u n J a Cawstan and tts second denvatsue (an b).
3.4.
Frames for the windowed Fourier transform.
The windowed Fourier transform of Chapter 2 can also be discretized. The natural discretization for w , t in gWwt(x) = eWZ g(x - t ) is w = mq), t = nto, where wo, to > 0 are fixed, and m, n range over Z; the discretely labelled family is thus gm,n(x) = et-= g ( ~R~O)
.
.f
DISCRETE WAVELET TRANSFORMS: FRAMES
-
81
We can again seek answers to the sww questions as in the wavelet .case: for which choices of g, wo, to can a hetion be by the inner products (f, gmln);when is it possible to recopst~ctf in a numekally stable way from these inner products; can an e5cient &&ilBm be given to write f as a linear combination of the g,,,,,? The answers are ag&i pm*ided by the same abstract framework: stable numerical reconstruction of f from its windowed Fourier coefficients
~~
)r
i *
r
~7
h , n )
is only possible if the g,,, ,,sotilat
=
& f ( ~ eLamZ ) 9fz - e
lf
constitute a frame, i.e., if there erdst A
> 0, B < oo
*<
If the gm,, constitute a h e , then any function f E L 2 ( R ) can be written as
"
-"Whereg;nT. are the vectors in the dud fr-; (3.4.1) showsboth how to recover , f from the (f, g,,,,,) and how to write f as a superposition of the g,,,. The detailed analysis of frames of windowed Fourier functions brings out some features different from wavelet frames, due to the difkrences in their constmctions.
A necessary condition: Sufficiently h5gk t i m d k q u e n c - y denThe arguments in the proof of Theorem 3.3.1 can be wed for the windowed s@. burier case as well (with the obvious modifications),- 1 to the conclusion
* 3.4.1.
27r
A I- 11g1I2 L 3 woto
I
(3.4.2)
,,h any frame of windowed Fourier functions, with frame bounds A, B. This not impose any additional restrictions on g (we alwan wmme g G L2(R)). A ' consequence of (3.4.2) is that the kame bound for ory tight f r w equda $ 2 ~(u&)-' (if we choose g to have norm 1); in particular, if the g,,, constitute orthonormal basis, then woto = 2x. The t h e a c e of any owstmint on g in the inequalities (3.4.2) is similar to the
:-Fzwm ,s
I
b c e of an admissibility c o d t i o n for the continuous windowed Fourier tnubsk i n (see Chapter 2 ) , and quite amlike the constraint J 4 KJ-' 1 $ ( a 2 < oo on
&hemother wavelet, necessary for wavelet frames as well as for the continuous Wavelet transform. Another diEerence with the wavelet case is that the time and b x p n c y translation step and uij are constrained: there exist no windowed Pburier frames for pairs wo,to such that %to > 27r. Even more is true: if $d& > 2x, then for any choice of g E L 2 ( R ) there exists a corresponding T € L2(R) (f # 0 ) SO that f is orthogonal to dl the gm,, (z)= e i u Z g ( z -ntO). In this case the g,,, not only fail to constitute a frame, but the inner products
82
CHAPTER 3
Cf, ~ h n , ~are ) not even sufficient to determine f .
We are therefore restricted we even 'havetb chbose woto < 2?r. Note that no similar restriction w w,4 existed in the wavelet case! We wiM come back on these conditions in Cbpk,r 4, where we discuss in m e grester detail the role of time-hquency dm&y for wind d Fourier vemus wavelet frames; we postpone our prod of the neclessity of *to < 2r until then.
5 2n; in order to have good time- aud frequeacy-1&tbn,
to
A slrfaeient co&ti&
and estimates krs %beh m e bun*. do not necessarily tpmtituh a frame. An easy counterexampie is g(x) = 1 fm f) 5 x 5 1, o otherwe if I$ 7 1, then my func tion f supported in [I, to] will be orthogonal to all the h,,,bwever small uo is chosen. In this example essinf, C, (g(s - do)('= 0, and this is what stope the %+, frum being a frame. (Something M a r occurs fbr w a d ,see 53.3-1 Computations entirely similar to those in the wavelet csse shaw that
3.4.2
Even if woto
id
ft"fT
Ilf IId2
< %, the %,,
C I(f99mr)12
*
m.n
As in the wavelet ame, s&cientIy fast decay on g leads to d e w for P,so that ty choosing wo d en-, the s e e d terrrm in the rf&t-$and eides ofCf3.4.3), (3.4.4) caa be made aab&dlg smell- If C , Ig(x - nte)p is bomdd, and ltcmxkd b k o w by a strictly positive eamtmt (no "canspiring" of the -08 of g), then this implies that the h , constitu* a ~ E for W fdkbatly 4 UO, prieh &am&! bOtlllCts given by (3.4.3), (3.4.4). l%cpW*, ore b , t h e bllowiq prctpoaittan.
PROPOSITION 3.4.1. I f g ,
are suc-fr h t
DISCRETE WAVELET TRANSFORMS FRAMES
83
-
- nt01i b(z nto + $11 d a at hast ~ ad and rf P ( s ) = suposzg,, Cn fast as (1 + ~si)'('*'), mfh L > 0, then there ezwts (wO)thr > 0 SO that the gm,n(x) eCRUJb2g(z - nto) wnstttute a &me 11,heneuer < (@)thr For wo c ( w ~ ) the ~ ~nght-hand , ssdea of (3.4,3), (3.14) an? jhme bounds for the 9m,nThe conditions on and (3.4.5) are met if, e-g., Ig(z)l 5 C(1+ I X ~ ) - ~ with E
y
> 1. REMARK.The windowed Fourier case exhibits a symmetry under the Four~er
transform absent in the wavelet case. We have
$-
,
.
which impliea that (3.4 3), (3.4.4) still hold if we replace g , w, to by 8, to, wo, respectively, everywhere in the right-hand sides (including in the definit~onof p). Using this remark, we can therefore compute two est~mateseach for A and B, @ pick the highhest one for A, the lowest for B o 5.43.
5;
The dual frame. The dual frame 1s again defined by
F* F is now (F' F)f = C,,
(f, g,,,) g,,,,,. In this case one easily checks $hat F'F commutes mth translations by to as well as w~thmultiplications by i.e., if ( T f ) ( x )= f (s to), ( E f ) ( x )= e W xf ( z ) ,then *%ere
-
F*FT=TF'F,
F'FE=EFgF.
fdows that (F'F)" a h commutes with E and T, so that
-
gm,n = (F*F)" JiF = P T" (F'FL-I
g=(x)
T" g 9,
= elrnvozij(x - nto) = 4,,,(4
,
(F*Fj'lg. Unlike the generic wavelet case,the dud kame is dm* by a single function 4. This means that it is not aa important in the Fourier case that the frame be close to a tight frame: if B / A - I is ble, then one simply cumputts 3 to hi& precision, mce Bnd for all, with the two dual &ma.
dr~meawith compact support in time or hquency. The folmcthn, again &om Daubechiea, Grossmann, and Meyer (1986), arid to $3.3.6.A, leads to tight windowed Fourier W with arbitrarily
high regularity if wto< 2n. If support g C
[-$,
51,then
where we haw used that for any n, at most one value oft can contribute, because of the support property of g. Consequently,
and the frame is tight if and only if EnIg(x -nto)I2 = constant. For instance, if woto 2 n, then we can start again from a Ck or Cm-functionu satisfying (3.3.25) and define .
I
i
iI
I
sin
(
0
[
~
.
~
+ ( a ~ -. ~
~
o
- ) 5 ] 5, ~ 5 wo ~ - - t o ,
otherwise .
function (depending on the choice of u) with comp.et Then g is a Ck or support, Hgll = 1, and the g, constitute a tight frame with frame bound 27r(wohj)-' (as already followed from (3.4.2)). If woto < n, then this construction can easily be adapted. This construction gives a tight frame with compactly supported g. By taking its Fovxier transform, we obtain a frame for which the window function has compactly supported Fourier transf~rrn.'~
B. The Gaussian. In this case g ( t ) = n-li4 e - ~ ' / ~ Discrete . families of wiqdowed Fourier functions starting from a Gaussian window have been disc& ' extensively in the literature for many reasons. Gabor (1946) prop& their use for communication purposes (he proposed -to = 27r, however, which is inapprapriate: see below); because of the importance of the "canonical coherent states* in quantum mechanics (see Klauder and Skagemtarn (1985)) they are of interest to physicists; the link between Gaussian coherent states and the Bargrnann space of entire function makes it poesible to rewrite results concerning the gm,rn in
DISCRETE *AVELIW TRANSFOW.
7
.
85
FRAMES *
.
terms of sampling properties for the Bargmann space. Egploiting this lidc with entire functions, it was proved in Bargmann et al. (1971) and independently in Perelomov (1971) that the g, span aU of Gz(R) if and only if woto 5 2n; in Gacry, Grossmann, and Zak (1975) a different technique was used to show that if wotn = 2 7 ~ then .
even though the,,g , are "complete," in the sense that they span L~(IR).'~ (We will see ia Chapta 4 that this is a b t consequence of wo . to = 2n, and of the regularity of both g and g.) This is therefore an example of a family of g,,, where the inner products (f,gm,,) suffice t o characterize the function of f (if (fi, g,,,) = (fi, gm,n) for all n,n,then f i = f2), but where there is no n u m e r i d y stable reconstruction formula for f from the (f, g,,,). Bastiaans (1980,19%f)has constructed a d u d function 3 such that
f = C (f, 9m.n) b . n
=
-
1
(3.4 6 )
m.n
a#h 3m,n(~) e'mwo= g(z - nto), but comrergence of (3.4.6) holds only in a $'#wry weak (in the sense of distributions-see Jamsen (1981, 1984)), and wt even in the weak L2-sense; in fact, 3 itself is not in L*[w). to = 27r 1s thus completely undetstood, what happens if uoto < 27r? Table 3.3shows the values of the frame bounds A, B and of the ratio B/A, for " various values of wa.to, computed from (3.4-31, (3.4.4) and the analogous formulas uslng g. We find that the g,,, do constitute a frame, even for wo - to/(27r) = .95, &though B / A becomes very large so close to the 'Lcrit~cdl density. It turns out t0/(2n) = l / N , N E N, N > 1, the frame bounds can also be that when -computed via another technique, which leads to exact values (within the error tat ion) instead of lower, respectivelybupper,bounds for A, B.14For the to/(%) = f and Table 3.3 reveals these exact values as well; it surprising t o k e how close our bounds on A, B (which are, after d l , obtained hy-Schwarz inequality, and &&ht therefore be quite coarse) are t o ues Substituting these values for A, B into the aGroximation at the end of 83.2, we can compute 3 fixthese different choices of wo, to. e 3.6 shows plots of ij for the special case where wo = to = ( X ~ n ) l /with ~, X ues -25, .375, .5, -75, .95, and 1. Note that ~ a s t i k function ' 3, corresponds to X = 1 (lower right plot in Figure 3.6), has to be computed &ntly, since A = 0 for X = 1. FO; s m d A, the frame is very close t o tight, ij is close to g itxlf, as is illustrated by the near-Gaussian profile of ij for .25. As A increases, the frame becomes both less redundant (as reflected by ng maximum amplitude of 3) and less tight, causing 3 to defia$e more from a Gaussian. Because both 9 and j have (faster than) exponential , one can easily prove from the conve%ing series representation for 4 (see that 3 and 4 have exponential decay as well, if A > 0. It follows that good time-frequency locabation properties, for all the values of A < I in Figure 3.6, even though it is quite striking how ij tends to Bastiaans'
j.T
i,
C
I
86
CHAPTER 3 I
FIG 3 6 The dud /mme junciwn $'for G a w s ~ a ng and
= to =
(2r,4J1/2, &
X = 25,.375, 5 , 7 5 , 95, and 1. As A m m c l s e s , g h a t e s mom and mow fmm o Ga-n (rgectmng the rncseose of B / A ) , and its omplttude increases ar well ( k w e A + B deEnoses). For X = 1, g ta no longer square tntegmblc.
pathological as A increases. For A = 1 , dl-time-frequency lockization breaks down.15 The series of'plots in Figure 3-6suggests the conjecture, first formulated in Daubechies and Grossmann (1988), that, at teast for a u s s i a n g, the gmSn are a frame whenever woto .< 2 ~ .In Drtubechies (1990) it wes Shawn t h ~ thii t is indeed the case for u0to/(2?r) < .996. Using entire function methods, this conjecture has since been proved, by Lyubsrskii (1989) and independently by Seip and Wallsten (1990). There exist of course many other possible and popular choices for the window function g, but we will stop our list of examples here, and return to wavelets. 3.5.
I
I I
Time-frequency Iacalization.
One of our main motivations for studying wavelet transform (or windowed Fourier transforms) is that they provide a timefrequency picture, with, hopefully, good localization properties in both variables. We have awxted several times-that if q3 itself is well localbed in time and in frequency, then the frame Gnerated by will share that property. In this section we want to make this vague statement more precise. and 141 to be symmetric (true if, For the sake of convenience, we W u m e e.g., II,is real and symmetric-a good example is the Mexican hat f u n ~ t b n ) ' ~ , then 9 is centered around 0 in time 8nd near kto in frequency (with, e.g., b = E I $ K ) ~ ~ /4 [ Jl~j(c)v]). ~ If$ is well localized in time and frequency, then $, pjill similarly be well -1 wound comnbo in time and around fqmb , in frequency. Intuitively spadung, (f, b,,) then represents the "information contentn in f near time a?& d near the frequencies fq r n { o . If f itself is LLessentially localized" on two rectmgle9 in time-frequency space, meaning that, forsomeO<%
I
~
DISCRETE WAVELET TRANSMXUMS: FRAMES
87
TABLE 3.3 Values for the fmme boundP A , B and thew mtw B/A for the aue g(2) = r-'I4 exp(-z2/2), for drffereni d u e s of w, to For wolo = 1r/2 and x, the exoct d u e s can be computed ma the Zak tmruform (see Daubechtes and Gms8mann (1988)).
CHAPTER 3
where 6 is some small number, then this intuitive picture suggests that only those (f, +*,,)corresponding to m,n for which (aznbo, f iim w~thin or close to [-T, T] x (I-Q1,-&l l J [no,all)ate .needed t o reconstruct f to a very good approximation. The fobwing theorem itates that this is indeed trhe, thereby justifying our intuitive picture. THEOREM 3.5.1. Suppose t h t the $m,n(~)= a.-m/2 +(%-mx - &) wnstztute a jmme with frame bounds A, B, and suppose that
wzth a > 1, 0 > 0, 7 > 1. Then, for any c > 0 , there ezzsts a fintte set B, (Ro, Cll; T ) C z2 90 that, fix 011 f € L2(R),
1. If f satisfies (3.5.1) and (3.5.21, then the first two terms in the right-handA.
side of (3.5.3) are bounded by 26
11f 11;
choosing e = 6 then leads to
2. As €4, # Be(Oo, JZ1 ; T)doo (see proof, below):, infinite precision is only possible if infinitely many (f, &,n) are used. n Figure 3.7 gives a sketch ofthe set Bc(Oo, R1; T) far one particular value of e; the proof will show how we obtain this shape.
Pvf. 1. We define the set B, as
B,(Ro, R1 ; T) = {(m, n) E 2'; -
mo I m 5 m l ,
lnbol
5 aGmT+ t ) ,
where rno,rnl and t, to be dehed below,depend on & I , Q ~ TI , and e. The points (aomnbo, faornb)comqmnding to (m, n ) in such a set, do indeed fill a shape like Figure 3.7.
1
*
,
4
i
DISCRETE WAVELET TRANSFDaS: FRAMES
Ct
tr
t
L-----
FW. 3.7. Thc
set
-,---_I
Be(flo,R I , T ) of "wavelet io#ice p n b " needed for
raowtructron off sf f rs localrzed mostly an
2.
f-
C
(f, $m,n)
[-T,Tj
in time and an [-Cll,
appnmmate
an
-%I
U [Oo,fll] tn
+Zn
(m,n)EBt
=
SUP
=
s
SUP Uhll=l
7
1 C C
( f q
memo
,EZ
m>ml
[I(Pno.n,f.$m,n)I SUP fl'I1-l
h)
$m n ,$ ()*:,
(m,n)e~.
or
+
(f, +m,n)($xnv h)
(m,n)EB,
SUP #hH=l
C
(f h) -
Uhll=l
C
+
C
1((1*Pho,~1)f, +m,n)II ~($Ynv h)I
noSm<mr lntt,,l>.;-~+t
11( Q T +m,n) ~ ~ I + l((1- QTM,$m,n) I] I( $ x n j
'
h)I 9 (3.5.5)
where we have introduced (QT f )(z)= f(z)for 1x1 1 T, (QTf) (z)= 0
+=
<
otherwise, and (Pn,,n,f)"(O = if 00 It1 5 fh, (P~,,n,f )"(O = 0 otherwrse. since the constitute a kame with frame bounds B-I. A-'. y e have
=
[ /
1/2
e I~MII~]
(because Ilhtl = 1) .
I€I<%
Similarly,
It remains t o check that the other two terms in (3.5.5) can be bounded by
3. By the same Cauchy-Schwarz trick, we reduce the remaining two terms in
(3.5.5) t o
(3.5.6) It is therefore sufficient to show that for appropriate m, rnl, t each of the expressions between sqwm brackets is smaller than BE' llf 1!2/4.
4. We tackle the b t term in (3.5.6) by the same technique as in the-proof of
DISCRETE WAVELET TRANSFORMS FRAMES
Proposition 3.3.2:
where 0 < X < 1 will be fixed below. Since 11 + (u - s)~]-' (I + s2)[1 + (u + s ) ~ ] -is~ bounded uniformly in u and 8 , we have
Substituting this into (3.5.7) we find
The sum over t converge3 if yX > 1, ie., if X > 1/7; we can choose, e-g., A = i(1+ 7 - I ) . On the other hapd, for Ro L I 01,
CHAPTER 3
92
and
In all these estimates, the constants C, may depend on ao, bo, A, 8, and Ro, R1,no,and m l . Subdit~rtinginto (3.5.8), with the choice X = i ( 1 ? - l ) , leads to -y but are independent of
+
If rnl 2 (In m)-'((? - 1 ) - ' ln(4C7/Bt2) - In no]a d mo [r8-1(r - I)-' In (Bc2/4C7)- in Q l ) , then this leads to
'I (In Q)-'
,.
as desired.
5. The second term in (3.5.6) is easier to deal with. We have
C
l ( Q2I)~..,f+ .
JnhJ>a;"'T+t
+
The sum over n ~ p l i t sinto two parts, n 7 b;' (aCmT t ) , and n < - b ~ ' ( a & ~ Z ' - t ) . Let nl be the smallest integer larger then b;'(a;*T-tt).Then
+
,
-
(because JGmz6 1= n 4 - G m x > (n - ni)bo
t+Gm(T-x))
+
DISCRETE WAVELET TRANSFORMS F U M E S
The mlm over n < -6; that
93
'( a ~ ~ t)Tisf dealt with in the same way. It follows
wh~chcan be made srtlaller than B
c2((f 112/4 by
&wing
This concludes the proof
t The estimates for mo,m i , t that follow from this proof are very coarse; in practice, one can obtain much less coarse values ~f $ and hive faster decay than stated in the theorem (see, e.g., Daubcch~es(1990), p 996). For Iater reference, let us estimate # B,(Oo, R I ; T), as a function of !lo,i l l , T, and f We find
4
%
3
"
On the other hand, the area of the time-frequency region f-T, T] x (I-R1, Sto] U [h, a l l ) is 4T(F1 Qo) As Ro+O and T, ill -too, we find
-
ch is not independent of c. We will come back to this in Chapter 4. m 3.5.1 tells us that if tl, has reasonable decay in time and in ency, then frames generated by do indeed exhibit time-frequency 1 e ation features, at least with respect to time-frequency seb of the type T ] x ([-$I1, U [Q, $Il]). In practice, one is interested in l o c a l b on many other sets. A chirp signal,for instance, intuitively corresponds to nai region (possibly curved) in the timefrequency plane, and it should be to reconstruct it from only those +m,n for which ( a r n h , f%-mb) is in to this region. This turns out to be the case in practice (for chirp signals, others). It is harder to formulate this in a precise theorem, mainly one first has to agree on the meaning of "1ocdizationn on a prescribed uency set, when this set is not a union of rectangles as in Theorem 3.5. I.
+
CHAPTER 3
94
If we choose the interpretation in terms of the operators Ls defined i n $2.8 (i.e., f is mostly localized in S if II(1 - Ls) f 11 << I1 f It), then the theorem is almost trivial if the wavelets in the definition of Ls and in the frame both have good decay' properties. For any othQ timefrequency localization procedure (e.g., using Wigner distributions, or the sffine Wigner distributions of BeBertrand and Bertrand (1989)), we still expect results analogous to those in Theorem 3.5.1, but the proof will depend on the chosen localization procedure An entirely similar localization theorem holds for the windowed Fourier case. THEOREM 3.5.2. Suppose that the g,,,(x) = e'-~'g(r - nto) cowtotute a fmme mth fTnme bounds A, B, and suppose that
unth a > 1 . Then, for any f E L2(R), and all T, R > 0,
> 0,
c
there e m t t,,w,
.
> 0, such fhat, for dl
Proof. 1. By the ssme tricks as in points 2,3 of the proof of T b r e m 3.5.1, we have
IZ: \
+
mEZ IntoI>T+t,
liQ~f,*,.rl2
1'") . (3.5.12) = j(()for
,
where (QTj ) ( z ) = f (z) for Izl 5 T , 0 otherwise, and (Pnf)"(Z) 141 5 n, 0 otherwise. The theorem follows if we csn prove that the last tm tenns in (3.5.12) can be bounded by B1I2ci(f 11. Let us hst concentrah an the kust term.
DISCRETE WAVELET TRANSFORMS. FRAMES
%
d
> ;'+ :<
I '
$&
k ti ' p*a
*
+
95
One easily checks that the contribution for n > t ; ' ( ~ t,) is exactly equal to that for n < - t i l ( T t L ) ;we may restrict ourselves to negative n only, at the price of a factor 2. By reaefhhg y = E wo C if f! is positive, we see that ue may restrict owselves to negative C as well. Hence
+
sup
[l
-
+ (z+ nto)'~-~/*
I=l
CHAPTER 3
However, for any f i , v > 0, we have
I
It follows that
I
Let
7tl
be thc: smailcst integer larger than
00
<
I
I
[I
T + t,. Then
+ ( ( n - tc)toc t.)21-0+1/2
by the computation above. Putting it altogether, we have
where C3depends on UQ to, , a,and C,but not on T (or 0).
3. Similarly, one proves
Since cr > 1, it is clear that appropriate choices of t,, u, (independent of T or R!) make (3.5.15), (3.5.16) smaller than BE^ fl f 1i2/4. This concludes the proof. m
97
DISCRETE WAVELET TRANSFORMS FRAMES
<
Fgure 3 8 gives a sketch of the collect~or~ ( m , n )such that Iw( R + w e , Into( 5 T t,, as compared to the time-frequency rectangle [-T,T] x [-R,S2]. The "e-boxn has a d~fferentshape from that In F~gure3 7
+
FIG 3 8 The set of latirce pornts
Be needed
lor crppmnmote reconstructrota, vrcl Ule [-T, T]an hme and an 1-0, R]
wandowed Fmner tnirufonn, of a funcicon localued mosUv tn
compare agan the number of puiuts ;.the enlarged l k c Be = f.(m,n), lmwol<sZ+w,, InbJ
C/
-
1
3
f-l
Z T .I
6;s'+ % $;
+
*
..'
(n + u,)
+
2t;' (T t c ) --+ (woto)-' (3.5.17) 4TR In contrast t o the wavelet case, thts limit is independent of c We will come back bD the significance of this m Chapter 4. 5
*
-
# B. (T, n) - zw;
%
4TQ
3-8. Redundancy in frames: What does i t buy? yi-"h illustrated by the different tables of frame bounds, frames (wavelets or win' dowed Fourier functions) can be very redundant (as measured by, e.g.,
9
$ ifthe frame is cloae to tight, and ~f 811 the frame vectors are normalized). In ' applications (e.g., as in the work of the Marseille groups-see the papers of
-
@ tt*l i@
CHAPTER 3
98
i
Grossmann, Kronland-Martinet , Torresani) this redundancy is sought, because represent at ions cla3e to the continuous transform are wanted. It was noticed very early an by J Morlet (private communication, 1986) that this r e d y n b c y a1.w leads to robustness, in the sense that he could afford to store the wavdet coefficients (f, 11, ,) wlth low precision (only a couple of bits), and still reconstruct j with comparatively much higher precision. Intuitively, one can understand this phenomenon a.follows. Let (ql),EJ be a frame (not necyarily of wavelets or windowed ~ o d r i e rfunctions). If this frame is an orthonormal basis, then
p:
7f-.P(J1,
(Ff)]=(f,~.j) Is a unitary map, and the image of W under F is all of l?(J).If the frame is redundant, k c . , if the cp, are not independent, then the elements of F31 are sequencm with some correhtions bqilt into them, and F X = Ran (F) is a subspace of P2(.I), smaller than P2(J) itself. The more redundant the frame IS, the "smaller" Ran (F) will be. As shown in $3 2, the recoxlstnlction formula
involves a projection onto Ran ( F ) : i t can be rewritten as
and PC= 0 if c 1 Ran ( F ) . If the ( j , cp,) are adulterated by adding sump a, to each coefficiexlt (an example woutd be round-off error), the total e k t on the reconstructed h~nctionwould be Since p* includes a projection onto Ran (F),the component of the sequence a orthogonal t o Ran (F) does not contribute, and we expect 4 f - fappmx~~to be smaller than JIall. The effect should become more pronounced the "smaller" Ran ( F ) is, i.e , the more redundant the Frame is Let us make this more explicit with the twudimensiolial frame used as an example in 53.2, and by comparing it with an orthonormal basis. &fine 211 = 1 ( l , O ) , uz = (0, I), el = u*, ez = - p l , Q = $ul - iu,; (u,, 7.4,) constitutes an orthonormal basis for , 'c (el,ez,es) a tight frame with frame If we add q c to the coefficients (f, ul), where a, are independent bound random variables with mean zero and variance 1, then the expected error on the reconstruction will be
-921,
i.
If
add ale to the frame coefficients (f, e, ),, then we find
DISCRETE WAVELET TRANSPORMS. FRAMES
99
which represents a gain of over the orthonormal crrso! A similar argument can be applied to waveleior windowed Fourier frames. In order to confine ourselves to only Gnitefy #m,n or gmSn, we assume f to be LLessentiaEly localizedn on 1-T, T ] x ({-Q,, -%I U Po, Q,)) (wavelet case) x [S2J$ (windowed I, Fourief CBBB), so that there exists a finite set or 1-T, Be (see $3.5) fix wMch
+
.
i
-
(and similarly for the windowed Fourier case). Let us assume the frame is almost tight, #,,, 2: A-'tD,,,. Adding am,,6 t.u every (f, $m,n), under the assumptions E(a,,, a , t , , r ) 6,m~bnnt, E(a,,,) = 0, leads to
-
5 r2 ltf It2
+
@(# B,)A-'~
(3.6.1)
If we assume l(+rn,nI] = 1). If we "double the redundancyn by halving b,then tlte new frame would again be almost tight (see, e.g., (3.3.11), (3.3.12)), with a frame bound A' twice aa large; on the other hand, the new "c-box" B: would contain twice as many elements. It follows that
-21
$
i.e., doubling the redundancy leads to halving the dfed of adding errors to the wavelet coefficients. The same argument can be d e for the windowed Fourk - 'w. 31 As it stands, the argument above is rather heuristic. There are indications 6 %that it can be consideratly strengthened: the gain factor ohX?~ed by Morr kt was in fact larger than what would folbw from these arguments. hiore4 *$ ,%t ' Over, recently Munch (1992) showed'that for tight windowed Fourier frames ; with X = ( 2 ~ ) - ~ = wN ~"t ~, N E N, N > 1, tbe gaiin factor with respect to ' - the o r t h o n o d case i. in fact N-', and not N-', as would follow from our 0.-% I-. that A-I is integer in an essential way, but it is hard $ T e n t . His proof b believe tbat the same phenomenon would mt exist for nooninteger A*'; maybe holdsfor w&t frames! I put thia a$ a rbdlenge to the reader. . . . . it
' 6
*
s-"
'f"
.
loo 3.7.
CHAPTER 3
Some concluding remarks.
f n this chapter we have studied in some depth the reconstruction of
f from the
-m/2
sequence((f1 SDm,ti))m,nE~,whael(r,,n(~)=ao Il(qmz-nbo)(~dvariations thereof-see $3.3.4). We have seen that numerically stable reconstruction is only possible if the &,, constitute a frame, and we have derived a reconstruc, are a frame. One can, however, use other reconstruction tion formula if the,,$ formulas (provided the tlr,,, do constitute a frame: the necessity of that condition remains!) To conclude this chapter, let me sketch the appro& of S. Mallat, which addresses moreover the problem of shift-invariance. The discrete wavelet trausform, such as I have described it in this chapter, is highly non-invariant under translations. By thii I mean that two functions may be shifted verslons of each other, while their wavelet coefficients may be very different. This is already illustrated by the "hyperbolic lattice"" in Figure 1.4a, where the axis t = 0 plays a unique role. In practice one does not use an infinite number of scales, but cuts off very low and very high frequencies: only those m for which m l < rn 5 mo are used. The resulting truncated lattice is then invariant under translations by bo2"Q (choose a o = 2 for simplicity), which is, however, very large with respect to, e.g-, the sampling time step for f (in most applications, f is given in sampled form). If fl is a shifted version off obtained by a translat~on# nbo2w, then typically the hvelet c&cients of fl will be different from those of fo. Even if the shift is fibo2*, with ml m 5 m, then (fl, = (f, +m,n-zm-mii) if m 2 m, but no such formula can be written for m > m. For some applications (in particulat, all applications that involve "recognizing" f ) this can be a real problem. In a first approximation, the solution proposed by S. Mallat is the following:
<
Compute all the J & f ( r )+(2-"z - n2-"bo) = a,,,(/) (a special $Jsuch as in 93 3 5.D makes it-possible to compute these in C N log N operations, if j consists of N samples). This l i t of coefficients is invariant for shifts of f by nbo. At every level m, retain only those crm,,( f) that are local extrema (as a function of n). This effectively corresponds to a subsampling of the highly redundant am,,(f). In practice, the number of subsamples retained is proportional to 2'm times the original number; this is about the same number as one has in a not too redundant frame of the type described earlier, but the suhsampling is now adapted to f , and not imposed a priori by the hyperbolic lattice.
Together with this prescription for decomposition (here described in a simplified form) Maliat then p r o m a r d r u c t i o n algorithm, which works very well in practice (see Mallat (1991)).18 In M d a t and Zhong (1992), the procedure was extended to two dimensions, to treat images. One way to view Mailat's approach is to look upon the 2-m9(2-mx n2-mbo) as the underlying frame (note that the change in normalhtion caunters the larger number of frame vectors in every m-lwei) . Again, it is nemmuy that this family satisfy the frame condition
-
i
101
DISCRETE WAVELET TRANSFORMS: FRAMES
+
(3.1.4) for a stable reconstruction algorithm to exist, but once this condition is satisfied, several reconstruction algorithms can be proposed In this case, Mallat's extrema-algorithm is certainly more sophisticated than the standard frame inversion algorithm.
Notes. 1. If any f can be written as such a superposition, the
,
$m,n are ,also called "atoms," and the corresponding expansions "atomic decompositions." Atomic decompositions (for many spaces besides L2(R))have been studied and used in harmoruc analysis for many years: see, e.g., Coifmtp and Rochberg (1980) for atomic decompositions in entire function spaces.
2. This is true except for very special JI. If the qbm,, constitute an orthonormal h i s (see Chapters 4 and later), then the expansion with respect to this orthonormal basis provides a discrete "resolution of the identity."
3. The polarization identity recovers (f,g) from 11f f gll, 11f fig((:
That is, if (Jn)nENis any increasing sequence of fnite subsets of J, i.e., Jn c J,,, if n 5 m, tending to J as n tends to oo,i.e., U n E Jn ~ = J, then IIF'c - CIEJ., CI VJ11 0 for n-roo. The proof is in two steps:
-
If n2
2 nl
> no, then
I
whi& tends to 0 for w-oo. Hence the q,, = a Cauchy sequence, with limit 7 in L ~ ( R ) . For this 7 , and any f E L2(R),
This is proved
follows:
( ~ (+f91, f + 9 ) - (R(f - !I) f'- g ) (because R' = R) ;
cJpj constitute
,
CHAPTER 3
102 1-
(Rf,g) 5 2 B+*
1 1 1 1 1 I 2 + 11g1121 ; lllf + dl2+ (If - gl121= H
I(Rf,9)1 = ( R f , g ) ( R f , / I ( ~ f , g ) l
= (Rf *
i 31
(Rf ,9)9ll(RI,9)l)
B-A
jjg
1 6-A
55
Illf El2 + II(Rf,9)gll(Rf,g)l 1l21 Illf ll' + 11~1121;
IlRll = SUP#f,l=,,IgII=1 I(Rf,g)l 5
B-A rn .
6. Intuitively, % can be understood as a "supeiposition" of Jhe rank one trace-class operators (., ha*b)ha*b, with weights c(a, b). If c is integrable with respect to aT2da db, then the individual traces af {., halb)ha*b(which are all equal to I), weighted by the c ( a ,6 ) are "summable," so that the whole superposition has finite trace,
This handwaving argument can be made rigorous by approximation arguments. - 1
7. We use here the "essential infimum" (notation: ess inf) defined by
where IAJ stands for the Lebesgue measure of A c R. The difference between essinf, f ( x ) and inf, f(x) lies in the positive measure requirement: if f(0) = 0, f(x) = 1 for all x # 0, then inf, f(x) = 0, but ess inf, f(x) = 1, because f 2 1 except on a set of measure zero, which "does not count." In fact we could be pedantic, and replace inf or sup by essinf or-esssup in most of our conditions without invalidating them, but it is usually not worth it: in practice the expressions we are dealing with are continuous functions, for which inf and essinf coincide. In (3.3.11) the situation is different: even for very smooth 4 , the sum I$(U~F)~* is discontinuous at = 0, because 4 ( 0 ) = 0. For the Haar function, for instance, l$({)j = 4 ( 2 ~ ) - ' / ~ I ( ( - 'sin2 {/4, and CmeZ 1$(()12 = ( 2 ~ ) ~ ' ' if ( # 0, 0 if ( = 0. We therefore need to take the essential infimum; the infimum is zero.
8. This condition implies both the boundednas of decay of P(s):
zmEZ 14(ag6)12 and the
DISCRETE WAVELET TRANSFORh4S FRAMES
and
)
5 c2 sup
l
tg
aom (1
+ IaF2E + s12)-('-a)/2
=-a3
In the b t term we can use that, fm Is1 2 2, loF6 + ti > Is1 - 1 2 *, a(1 + + $I2)-' < 4{1 + /sI2)-I, for 1st 2, (I + aIgmt + s12)-I 5 1 < 5(1+ [sI2)-l. It &that the first term uro be bounded by C(1+ / s 1 ~ ) - ( 7 - ~as ) / soon ~ as a > 0, -y > a. For the second term that suPz,y,g (1 + y2)[1 + (z- y)2]-1[1 + (z3- v ) ~ ) - ' < 00 to bound the sum by C"(1 + Cz==a(l + loo*ft2)-c(~-Q)/2, whse b < E < 1 is wbitrary. Since 1 < 5 a,this can be bounded by if 7 > a. We have therefore, for 0 < p < y - a, Cm@-I-
lee
Em
If 6 is continuous and has decay at oo, then ili)(qf)l2 is mntinuous in c. errepr at 5 = 0. There exists therefore a a,that ~ $ ( @ c ) I5 ~ fr if K - QJ5 a. Define, for a' < a,a fundion f by f([) = (2a')-1/2 if 1.f - 5 a', i(c)= 0 atherwise. Then
k-bt-'
(me Cauchy-Wmm on the integral)
CHAPTER 3
3
SC
4-
2 ~ Sup ' f
19t&a$'{)l2. m~it
If /d(t)/< C(l + 1((2)-7/2 with y > 1, then this infinite s w is uniformly bounded in <, and we can choose cr' so that the whole right h d - s i d e of the inequaiity is 5 2c. 10. Beware of a mistake in the example on pp. 988-989 of Daubechies (1990). The fornlula for (b)" should read (b)" =C T, iii+,*, and leads to the conclusion that h $ L P ( R ) for .small p. I would like to thank Chui and Shi (1991) for pointihg this out to me.
11. This is slightly different from multiresolution analysis, where (3.3.27) would also contain a scaling factor 2:
12. One can ELLSO construct tight frames where neither g nor i have compact support. It is, for instance, possible to construct a tight frame in which both g and g have exponential decay. One way of doing this is t o start .from any windowed Fourier frame, with window function g, and to define the function G = (F*F)-"~ g, where P F = gm,,)gmVn.The functions Gm,,(x) = eamO" G ( r - n t o ) (same wo, to as in the h , n ) then constitute a tight frame. One has indeed
C,,,(-,
The explicit computation of G can be carried out by a series expamion for ( F * F ) - ' / ~analogous to the series for (F*F)-' in $3.2. If g and 4 have exponential becay (in particular if g is Gaussian), then the resulting G and its Fourier transform have exponential decay as well. -For more details, plots of examples, and an interesting application, see Daubechies, Jsffard, and J o u d (1991). t
I
1
I
C
105
DISCRETE WAVELET TRANSFORMS FRAMES 2
13. The proof in Bacry, Grossma~~n, and Zak (1975) uses the Zak transform, which we introduce and use in Chapter 4. Full details of their argument are also given in Daubechies (1990). It is ~nterestingthat their proof can be extended to show that the g,,, still span all of L2(R) if one (aay one) of the g,,, is deleted, but not lf two funct~onsare deleted.
F
14. These exact formulas use again the Zak transform. Their derivation is given in Daubechles and Grossmann (1988); it 1s also revlewed in Daubechies
( 1990). 15. In sonie applications, 3astiaans1 result is ~nterpreted(correctly) to TIISL: that one should "oversample" (I e., ci~ooscwuto < 27r) m order to restore stability. Nevertheless, even in buch an "oversampled" reginle, somet~mes Bastiaans' pathological dual function is st111used (sac, f a instance, Porat can be split into two families, and Zeevi (1988)). If wto= x, then the g,., . gm,zn and gm,zn+l, each of which can be corlsidcrcd to be a fam~lyof Gaussian windowed Fourier functions with w ~ t o= 2x, one generated by g i b i f , the other by g(x - t o ) . For both families, the badly cdnvergent (nonconvergent in LZ)expansion (3.4.6) can be wr~tten,axid a functio~ican Lc viewed as the average of the two expansions Thls is of course true in the sense of distrlbutio&, and in reasonable convergence (probably due to cancellations) seems to be achieved (us~nga truncated version of Bwtiaans' g- private communication by Zecv~(1989)), but far better timefrequency localization, and I suspect better convergence, in practice, would be achievable by using the optimal dual function ij (corr~spond~ng to X = .5 in Figure 3.6 in this case).
A
C
16. This symmetry is certa~nlynot necessary.
: 17.
9
It is in fact a true hyperbolic lattice with respect to the hyperbolic geometry on both the positive and negative frequency half plane.
$ 18, Note, however, that Y. Meyer has proved recently that the a,,,(f) which are local extrema in the construction above do not suffice to characterize XI tr* 5 f completely.
?
-.---
CHAPTER
4
4,
Time-Frequency Density and Orthonormal Bases
This chapter splits naturally into two parts. The first sectjon discusses the role of time-frequency density in wavelet transforms versus windowed Fourier transforms, In particular, for the windowed Fourier tramdorm, orthonormal bases we possible only at the Nyquist density but no such restriction exists for the wavelet case This leads naturally to the secand wtion, which discusses dire&posaibilitiea for orthonormal bases in the two cases. 5
4.1. i
The role of time-frequency density in wavelet and windowed Fourier frames.
We start w ~ t hthe windowed Fourier case. We mentioned in $3.4~1that a family 4
s
offunctioas(g,,,;m,n~Z),
,
g m , n ( ~= ) earnyz g(2 - d o )
-
.
(4.1 1)
cannot be a frame, whatever the choice of g, if w~ to > 2%. In fact, for any of g E L2(R), one C ~ Dfind f E L2(8t)80 that f # 0 but gmen)= O for m,n E Z. ff, for instance, wo = 27r, to = 2, then 5u~hs funct;iofif is easy to ruct: ( f , g,,,) = 0 for all m,n E 2 I& to
o = =
J dz
/d
1
2mmz
U,
f ( z ) s(x - en1
e2rlmz
C 1(..+1)g ( 2 + 1 - 2 n ) , CEZ
ZGE +
+
that it is sufficient to find f # 0 for which f (z I ) g(z I - 2n) * 0. -sE3efine ~AQW,for 0 5 s < I , ! E 2, f [ x + t ) = (-1)l g(z-4-1). Ckculy,
,,, +
-
+
-
f ( X e) g(3: t - 2n) = CtEZ(-l)' g(2 t - 1) g(=+ t 2n), which hms into its negative upon the substitution C = 2n - C' 1, and therefore 61p~abzero. The same constfiction can be used for any other pair wo,to Mth product 47r; a generalization of this ;construction ex$te if wo h > 2?r * " ~( 2d~ ) - ' w o t o is rational (aee Da~ibecbies(1990), p. 978). If wk(2x)- is far^ t4an 1 but irrational, tbea I know of no explicit construction for f # 0, 107
-
108
CHAPTER 4
f I g,,,. The existence of such an f was proved in Rieffel (1981), using arguments involving von Neumann algebras.' If only "reasonably nice" g are considered (i.e., g that have some decay in time as well as in frequency), and if we are only interested in proving that the gmVncannot constitute a frame (which is weaker than proving that there exists f I g,,,), then the following very elegant argument by H, Landau does the trick. If Ig(z)l 5 C(l z ~ ) - " / ~ , 1i(c)1 <- C(1 Ad the g,, mnstitute &.@me, then Theahem 3.5.2 tells us that functions f which are "essentially localizedn in [-T, T] x [-a,R] in the time-frequency pIane can be reconstructed, up to a small error, by using only the (f, g,) with 1-1 2 R, lntol 5 T . More precisely, iff is bandlimited t (IfII, then to [-a,01 and if & If
+
+
[hII>T
<
According to thi formula, all such hnctions can therefore be mitten, up to arbitrarily small error, as a superposition of the ,#,, with lml 5 w;' (0+6), In1 -< toT1(T+6), where b depends on the error allowed, but not on R or T. However, a corollary of the work of Landau, Pollak, and Slepian (see 82.3) is that the space of functions baridlimited to [-0,Q) and satisfying Xz.,r & 1f (x)I2 S 7 1f JI2 (0< 7 < 1. 7 fixed) contains at least - O(log (Q?')) dierdliit orthogonal functions (the appropriate prolate spheroidal wave functions). All these different orthonormal fiinctions can only be approximated by linear combinations of a fjnite number of gm,, if the number of gm,, exceeds that of the orthonormal ~ for any R, T. functions, i.e., if 27r-lQT -O(log ( R T ) )5 b i l t o - ' ( 0 + b ) ( +6), Taking the limit as 12,T--roo leads to (2x)-' < (woto)-' or woto < 27r. (This is really only a sketch of the proof. For full technical details, see Landau (1989).) For a31 pkctical purposes we e v ~ nhave to limit ourselves t o wo , to < 27r (strict inequality) if we *ant good timefkquency localization: frames for the limit case wo. to = 27r have necessarily bad localization properties in either time or frequency (or even in both). This is the content of the following theorem. THEOREM 4.'1.1. (Balian-Low) If the gm,n(s) = e2"'wg(x - n ) constitute a frome for LZ(W),then either $& x2fg(x)12= oo or e2(8(()l2 = 00. Before proceeding to the.proof of this theorem, let us review its history, and add some iemarks. Originally, the theorem was stated for orthonormal bases (instead of frames), independently by Balian (1981) and Low (1985). Their proofs we very similar, but contain a small technical gap which was filled by R.Coifman and S. Semmes; this proof c m then be extended to frames as well, as reported in Daubechies (1990), pp. 976-977. ~ubsequentl~, a different, very elegant proof for orthonormal bases was found by Battle (1988), which was generalized to frame in Daubechies and Janssen (1988). (This is the proof we give below.) Two well-known examples of functions g for which the family e2w'm2g(z- n) constitutes an orthonormal basii are
9
Je
1, OLxll, 0, otherwise
TIMEFREQUENCY%ENSITY AND ORTHONORMAL BASES
-sib n
109
and g(x) = 7 . In the first case, Sq C21p({)1* = 00,in the-second case J dx z219(x)12= 00- It is shown in Jenseq, Hoholdt, and Justesen (1988) that g with slightly better tirn&fiequency localization: they construct one can chg such that b&h g and jj are integrable (i.e., Sdz 1g(z)1 < a,j' 4 (#(()I < oo), but their dewy is still rather slow, as dictated by Theorem 4.1.1. Note that the choice wo = 21r, to = 1 in our formulation of Theorem 4.1.1 is not a serious restriction: the conclusion hofds whenever wo . to = 27r. To see &is, it su£6cesto apply the unitary operator (Uj ) ( x ) = ( 2 . r-1 ~) 1/2 ~ 9(2nw~'x); applying U to g,,,(z) = e'm"Ozg(z-nto) one finds ( U I Q ~ , ~ = ) (eZmimf X ) (ug)(xn)To prwe Theorem 4.1.1, we will use the so-called Zak transform. This transform is defined by (4.1.2) ( ~ f ) ( s1, ) = e2*auf(s- -
.
x
o
(€2 . .
'
A priori, this is well defined only for f such that C If (s - t )1 converges for all s, in *icrrlar for / f ( z ) l5 C ( l + I X ~ ) - ( ' + ~ ) . It turns out, however, that this restrictive interpretation of Z can be extended t o a unitary map from L2(R)to L3([O,1j2). One way of seeing this is the following: e m , n ( ~=)e3nrmze(x- n), with e ( x ) = lofor0 5 x < 1, e ( x ) = 0 otherwise, constit utes an orthonormal basis for fi2(R).
i
-
i
-
, e2rrcms
e-2rrtn
(ZeI(5,t ) -
%a.
iw,
n;c
, '4.
1,
F I
;.
4'
fV
(Zerfs, t ) = 1 dmoet everywhere on fO,112
.
It foIlows that Z maps an orthonorma) basis of L3(R) to an orthonormal basis of ~ ~ ~ (112), 1 0 ,so that Z is ufiitary. We can extend the image of L2(R) under Z to a different space, isomorphic to L'([o, 1 j 2 ) . b m (4.1.2),we find, if we allow ( 5 , t) outside [o, 112,
k t us therefore define the space-2 by
2 = { F : it2+@;F(s, t + 1) = F(s,t ) , F ( s + 1 , t ) G e2"" ~ ( st ) ,
$8 s,;ithen Z is unitary between L2(R) and 2. The inverse map is easy as well: $$p snyF~Z, id**+
$"" @
&>
(z"F)(z) =
1
dt F ( z ,t ) ,
for
110
CHAPTER 4
if this integral 1s well defined (otherwise, we have to work with a limiting argument). The Zak transform has many beautiful and useful properties. As is often the case with beautiful and useful concepts, it was discovered several times, and it goes by many different names, according to the field in which one first learns of it. It is also known as the Weil-Brezin map, and it is claimed that even Gauss was aware of some of its pioperties. It srtfs also used by Gel'fand. J. Zak discovered it independently, and studied it systematically, first for applications to solid state physics, later in a wider setting. An interesting review article, geared mainly to applications in signal analysis, is Jamsen (1988). Only two of the many properties of the Zak transform will concern us here. x The first is that if g,,,(x) = e2"g(x - n),then
(as we already showed above, in the special case g = e). This implies
Equivalently, we have Z(F*F)Z-'= multiplication by f(Zg)(s,t)lzon 2, where F'F f = Em,,( j , g,,,) g., The second property we need concerns the relation between Z and the operators Q,P defined by (Qj)(x) = xf(x), ( Pf )(z)= -a f ' ( ~ )(or, more pmpxly, (Pf )" (6) = [ j ( ( ) )b. a e checks that Iz(Ql)l(s,t) = s ( Z f ) ( ~ , t)
1
@t(Zj)(s,t)
which means that dz x21f (zjP < QO, i-e.,Qf E L2(W), if and only if at(2f) E L2([0,112). Similarly, J @ t'fj(~)/~ < oo or P f E L2(R)if and only if %(Zj) E L2([0,112). We are now ready to attack the proof of Theorem 4.1.1.
Pmf of Theorem 4.1.1. 1, Assume that the g,,,
constitute a frame. Since
and since Z is unitary, this implies
TIMICTREQUENCY DENSITY A N D OEYTHONORMAL BASES
111
2 The dual frame vectors ij,,,,, are given by *
Pm,n
= (Fop)-'
gm,n
(see 53.2,$3 4 3). Since Z(FmF)Z-I = multiplication by JZg12,it follows that z3m.n = 1z91-2Z9m.n
which is in 2 hy (4.1.3) In particular, (4.1.4) implies that
with 2 4 = I/%. 3 S r t p p e now that J d x z2Ig(x)l2 < oo, Jcif <214(()12 < oo,i.e., that Qg, Pg E L ~ ( w ) .This will lead to a contrsdictbn, which will then prove the theorem Since Q g , Pg E L2(W), we hatie B),(Zg), &(Zg) E L2((0,1i2). Consequently,
are In L ~ ( [ D ,112);hence Qij, Pjj E
L2(R).
similarly,
5. Since Qg, Pjj f L2(&t),and since the dual frames,we have
(~rlr)mFEZ,
(3m,n)m,nE~ constitute
-
m,n
But (Qg, jj,,,,,)
= 1uk zg(x) e*2*inV = J d z g ( r ) c-
'(-
= (9-m,-n,
93).
3; - n) (z
-
- n ) ij(z - n ) (g,3m,n)
~mo~no)
CHAPTER 4
112
Similarly, (gm,n, PG) = (Pg, 9-,, )-,
. Consequently,
where the last term is again weU defined bmause Pg, Q$ € L2(R).
6. We have now reached aur cantradictian: (Qg, Pg) = (Pg, Qij) is impossible. For any two functions f i , j2 satisfyjng If'(x)l 5 C(l x2)-', l&(<)l 5 C(1 + t2)-*,-we have
+
0
1
2
= Jd~xf~(x)i/:o
On the other hand, since Pg. Qg E LZ(W), there exist gn satisfying Ign(x)l 5 C n ( l + x2)-', [&(<)I5 Cn(l+ t2)-' S U C ~that limn,, 9, = g, limn,, Pg, = Pg, lim,, Qgn = Q g . (Take for instance g, = ,C;, (9, H k ) H k , where Hk are the Herrnite functions.) A similar sequence Gn can be constructed for g: Then
Together with (4.1.6) this implies ( g , g ) = 0. However, from (4.1.5) we have (9, 3) = 1. This contradiction proves the theorem.* rn We can summarize o w findings so far as
wok = 2?r w o k < 27r
-
-
there exist frames, but they have bad timefrequency localization. frames (even tight frames) with excellent timefrequency localization are possible (see 53.4.4.A).
This is represented in Figure 4.1, &owing the three regions in the w, to-plane. As pointed out in 53.4.1, orthonormat bases are only possible in the ''border casen = 27r. In view of Theorem 4.1.1, this meana that all orthonofm81 bases of
TIMEFREQUENCY DENSITY AND ORTHONORMAL BASES
113
NO FRAMES FOR
w,t,
>2n
o0b = 2~
ONLY FRAMES WITH BADTIME FREWENCY
LOCALIZATION ARE POSSIBLE TIGHT FRAMES
POSSIBLE IF W b t O < 2x
+1
The mgons y l t o > 2u where m jhunu atr pussrble, and %to < 2n, &em t r m e - M e n c y laualtzakm errst, o n sepamted by the hyperbola woto = Z r , the only regton where orthononnal bas- 5re possrble
Fw.
hqht
frwnc?s trnlh ezcellent
the type (g,,,.,;
m,n E Z), with g,,,
as In (4.1.1), have bad time-frequency
localition. In fact, wo to is a measure for the Yime-frequency density" of the frame c6nstituted by the g,,,. We can for instance define this "density" as lim
A-00
#
4; tfw*nto)E AS) lxsl
9
(4 1 7)
where S is a "ressonable" set in R ' (with nonzero Lebesgue measure). This llmt is independent of S, and equal to (wo - to)-'. This "density" also emerged in the time-fresuencg localization discussion in $3.5; see (3.5.17). The restriction (wo - to)-' 2 (%)'I means that the time-freqmmcy $ensity of the Frame has to be at Ieast the Nyquist density (in its ugenaal&d" form; see $2.3). In fact, Theorem 4.1.1 tells us that we have to be s W 3 y a t m e the Nyqulst density if we want good timefrequency localization withr".r$ind& Fourier frames. Let us now turn to wavelets, where the situation is very different. It turns out that there is no "cleann definition of timefrequency density for wavelet exp i o n s . We already saw a first indication of this in the study of the localizatioxt operators in 52.8: for the windowed Fourier case, the number of eigenvalues in the transition region became negligible (as compared to the number of eigenvalues close to 1) when the area of the localizatiop region tended to infinity, whereas these two numbers were of the same order of magnitude in the wavelet case. This made it impossible t o make an accurate comparison with the Nyquist density. Something similar happene with discrete wavelet families. In the discussion of tbe-huency l d i t i o n with frames, in 33.5, we saw that a function that ia amentially concentrated in 1-T, x ((41, -no] x [Ro, QI]) in the time
\
I
i
114
CHAPTER 4
frequency plane can be approximated with good precision by a finite number of wavelets. Unlike the windowed Fourier case, the ratio of this number to the timefreqpency area 4T(R1-5)0) dependson the desired precision of the approximation (seeW(35.lI)), which makes it imjmsible (as in the continuous case) t o make a precise cowparison with the Nyquist density. On the other hand, if we try to define the andog of (4.1.7) for the hyperbolic lattices of Figure 1.4a, then we find that # ((mn);(WO, nto) f AS)
Rs(A) = -
lxsl
(where S is chosen so that the numerator is finite) does not tend to a limit as For S = [-T,T] x ([-2"', -2mo] u [2m0, 2m1j) and a0 = 2, for instance, Rs(X) oscillates between p ( l - 2m0-m1-1)/(1 - 2m-ml) and 2p(l - 2mo-ml-1 )/(I - 2mo-m1), where p depends on the chosen wavelet I). It might be this phenomenon, rather than the absence of an intrinsic time-frequency density for the frame, ,that causes the probkrn in counting the number of wavelets needed for time-frequency loealization. So let us probe a little deeper. As we mentioned before, there is no a priori restriction on the range of dllatlon and translation 'parameters in a wavelet frame: any choice of ao, 4 can be used t o define a tight frame with gpod localization in both time and fqquency for II, (see 53.3.5.A). In fact, froma (tight) frame with discretization parameters w, we can always constrkt a Merent (tight) frame with parameters m, boJ (same %) with b& arbitrary, by. simple d ~ a t b a It . ~is therefore not surprising , We can remove this dilational that we have no a priori restrictions on q-,4. freedom by fixing not only the ~ o r m a l i ~ t i oofn I), 11$11 = 1, but by also imFor r& ~ b we , could impose, for posing a fixed value for df El-' \$(<)t2. instance, E-' i & ~=i ~ I({-1 /&(<)12 = 1. A tight frame generated by a JI thus restricted would auhmatically have f'rame bound A = (see Theorem 3.3.1). A cornwith the formula A = for tight windowed Fourier frames suggesta thst, maybe (bo hao)-' w d d play the role of time-frequency density for the wavelet case. The following e p p l e destroys all hope in this direction. In the oe;rrt ssctian we will encounter @e Meyer wavelet 11; it has a compactly supportgd Fourier trqpsform 6 ' C (where k m a y be oo, as in 53.3.5.A; the two co-tiow w related) and the &,, = 2-"I2 e(2-"2 - n), m, n E Z 2morthonormal basis for L~(B).Let us, for this chapter only, define '.
&
&
4
are
&,n(z) =2-"I2 *(2-mx - nb)
,
where .(/, is the Meyer wavefet, and b > 0 is arbitrary. Consider the Mependent families F(b) = {q;, m,n E 2). As b changes, the "density" of the assbciated lattice changes as well. (Note that and 'r), are the s a n e for all the P(b)!) If any representation like Figure-4.1 held also for wavelets, then we woqld expect, since F(1) is an orthonormal baais f& Z2(Q that F(b) would not span dl of L2(R)if b > 1 ('hot enough" veekrrs), and that F(b) would not be linearly
TIMEFREQUENCY DENSITY AND 0ELTH90NORM#L BASES
115
independent ("too many" vectors) if b < 1. Yet one can prove (see Theorem 2.10 in Daubechies (1990),we also sketch this proof later irt this chapter) that for some c > 0, F(b) 1s a Riesz bass for L ~ ( R ) ,for any b E 11 - c, 1 €1. This example shows conclusively that lt is not always safe to apply "time-frequency space density intuition" to families of wave&+.
+
4.2.
Orthonormal bases.
4.2.1. Orthonormal wavelet bases. The conckusion of the last paragraph seems a rather negative point for wavelets no clean time-frequency density concept In thls section we emphasize a much more positive aspect: the existence of orthonormal wavelet bases with good t i m e - w n c y localization. Historically, the first orthonormal wavelet basis is the Haer basis, consfructed long before the term "wavelet" was coined. The basic wavelet is then, as we already saw in Chapter I,
3,
1, O < z < t -1, ij 5 x < 1,
0 otherwise
-
* W eshowed in $1 6 that the V,!J,,~(Z) = 2'"@~9(2-~2 n) constitute an or4honormal basls for L2(W).The Haar funcQon is not contmnuous, and its Fourier
KI-',
t;ransform decays only llke corresponding to bad frequencjr iodiz&ion. It may therefore seem that ttus basis IS no better than the winFoulier basis
$ 5
34 Z@ ,s,
3
, *a-
with 1, O ~ z < l ,
0, otherwise, which is alsb an orthonormal basis for L'[R). w, theHam W i already has advantages that this wlndowed Fourier k k s &oeswaC have. It twns out, for inStance, that the Ha& basis is an unconditional bwia for LP(R), f < p < m, whereas the windowed Fourier basis (4.2.2) is aat tf p J' We will come back to this in Chapter 9. For the anaiysis of s d ~ ( u n d i o n s the , dLmntinuous Hasi besis is ill suited An orthonormal wavelet basis with tim&fr-8queacyproperties complementary to the Haar basis is given by the L i t t l e d basis, ~ ~ %
C
.P &
7
%
ica = {
[2r)-l/2,
ol
r*
sg&52*,
( 8 k k k 2 -&xz)
I
I I
otpemise
a' @(x) = (TZ)-'
\
3s
I
.
It is easy to check that the &,n (z) = 2-*i3 $[2-"'~ - n) constitute indeed orthonormal basis for L2(R). We have H+,, ,1/ = 1for all rn,n E 2, sod
116
CHAPTER 4
-
By Proposition 3.2.1, this implies that {$;, m, n E 2) is an orthonormal basis for L2(R). The decay of $(x) is as bad ($(XI izl for x-+oo) as that of the orthonormd windowed Fourier h i s used in the Shannon expansion (2.1.1); both have excellent frequency localization, since their Fourier transforms are compactly supported. In the last ten years, Weral orthonormal wavelet bases for L2(R) have been constructed which share the best features of both the Haar basis and the Littlewood-Paley basis: these new constructions have excellent Iocalizatbn p r o p ertiedl in both time and frequency- The first constmction is due to Strornberg (1982); his wavelets have exponential decay and are in Ck ( k arbitrary but finite). Unfortunately, his construction was little noticed at the time. The next example is the Meyer basis mentioned above (Meyer (1985)), in which is compactly supported (hence E Cm)and Ck (k arbitrary, may be w). Unaware at that time of Stromberg's construction, Y. Meyer actually found this basis while trying to prove a wavelet equivalent of Theorem 4.1.1, which would have $imthe non-existence of these nice wavelet bases! Soon after, Tchaanitchian (1987) constructed the first example of what we shall call biorthogonal wavelet bases (see $8.3). The next year, Battle (1987) and LemariC (1988) used very &&rent methods to construct identical families of orthonormal wavelet beses with exponentially decaying $ E Ck (k arbitrary but finite). (Bsttle was inspired by t W q u e s in quantum field theory; LemariC reused some of Tchamitchian's computations.) Despite bving similar properties, the Battle-Jtemarj6 wavelets are diffemt from the Stmmberg wavelets. In the fall of 1986, S. Mallat and Y. Meyer developed the "muEtiresolution analysis" framework, which gave a satisfactory explanation for all these constructions, and provided a tool for the construction of yet other bases. But this is for later chapters. Before we get into multirdution analysis, let US review the construction of Meyer's wavelet baeis. The construction of is similar to the tight frame in 53.3.5.A.That f m e had redundancy 2 (twice Rtoo manyn vectors). To get rid of this redundancy, Meyer'e construction combines positive and negative frequencies (reducing a pair of functions to a single function). Ln order to achieve orthonormality, some clever
4
141
TIMEFREQUENCY DENSITY AND OfWHONORMAL BASES
tricks with phase factors are needed. More explicitly, we define
J! , I
4(c)=
{
(2~)-112edI2 sin [; v (&It] - I)]
,
$ 5 It)5
(2n)-'I2 ea'12 cos ,; v (&It,- I), ,
%f 5
0
otherwise,
5
117
by
7, ,
(4.2.3)
where v is a Ck or C* function satisfying (3.3.25), i.e.,
u(x) =
0 if x < O , 1 if x L 1 ,
(4.2.4)
with the additional property v(x) + v(1 - x) = 1 . The regularity of
4 is the same as that of v.
(4.2.5)
Figure 4.2 shows the shape of a
typical v and 141. In order to prove that the +m,,(x) = 2-m/2 $(2-mx - n) constitute an orthonormal basis, we only have t;o check that l/+(I = 1 and that the +rn,, make up a tight frame with frame constant 1 (see Proposition 3.2.1).
1
9
X
0
1
1
1
( 2 %)-Z
i
"I
i
0
i'i
$
h 1 Ls i
.
\
, -10
FIG. 4.2.
-5
0
5
10
I c ( , t ~ ~ t i o n r ~ g-gitmby(4.2.3)-(4.2.5). djr
i
L
CHAPTER 4
We have
But
(because v(s
+
11+(12
+ 1/2) = 1 - 4 1 1 2 - s) by (4.2.5))
= 1. To evaluate [(f,$-,,)I2, we use Tchamitchian's frame bound estimates (3.3.21), (3.3.22). We first prove that P1(2n(2k + 1)) = 0 for all k E 2,i.e., for all ( E R, hence
Em,,
I r O
Because of the support of 9, nonzero contributions to (4.2.6) are oniy poasible if IPS1 < !f m d I2'(( 2*(2k 1))l < implying 2'I2k+ 11 5 8/3. The only pairs (t,k) that do not violate this are (0,0), (0, -I), (1, O), and (1, -1). Let us check k = 0 (k = -1 is analogous). Then the left-hand side of (4.2.6) becorn*
+
+
9,
5
119
TIMEFREQUENCY DENSITY AND ORTHONORMAL BASES
-9 < < -%. For
One easily checks that both terms in (4.2.7) vani& 5 C within this interval, C = - + ?a with Q < a I1,We have
9
(4.2.7) = e-'(12 sin
-
- cm
:[
[;i
A
u ( l - a)] e1((+2z)A sin
u ( a ) ]sin
[i
Y(LI)]
+ Bin [X2 u(h)] ~ 1 [:6
~(a)]
On the other hand, one a i l y checks that for all # 0. It then follows from (3.3.21), (3.3.22) that Em1 3 ( 2 ~ [ )=1 ~ the &,, constitute a tight frahe with frame bound 1- (Similar computatio-zis can be used to prove that F(b) (see the end of 84-11 constitutes a R i a z bitsis for L2(R) if b is close to I . ~ ) This proof that the Meyer wavelets constitute an orthonormal basis relies on quasi-miraculous cancellations, using the interplay between the phase of and the special property (4.2.5) of v. Using multirec~oIutionanalysis, we will be able to explain away most of the miracle (see next chapter). Figure 4.3 shows a graph of @(z), with bhe C?L choice v ( z )= x4((35 - 8 - b+ 7 k 2 2 h 3 ) for 0 5 z I1. N4te tbt e w h if v E Cw,so that $ decays faster than any inverse polynomial, i.e., for'& N E N, tbere exists CN < do so that
This establishes (4.2.6).
4
-
J$(X)~
CN (1 + 1x12)- N ,
(4.2.8)
the numerical decay of .d, may be rather slow (i.e., inf (a; J+(x)J _< .001 JJ+IJL= for 1x1 > a) may be very large, reAecting a large CN iD (4.2.8)). The exponentially decaying wavelets of Stromberg or B a t t l e - M d W much faster numerical 'b decay, at the price of sacrificing regularity.
CHAPTER 4
In the matter of orthonormal bases then, wavelets seem to do quite a bit better than windowed Fourier functions: there are coastructions in which both gl, and have fast decay, in stark contrast with Theorem 4.1.1, which forbids simultaneous good decay for 9 and if g is a window function leading to an orthonormal basis. If I had written this chapter three years ago, this is probably where I would have stopped. But matters are not quite that simple: in the last few years, the windowed Fourier transform has led to a few surprises, which we will discuss briefly in the remainder of this chapter. %
4.2.2. The windowed Fourier transform revisited: 'Goodn orthonormat bases after all! One way in which one could try to generalize the windowed Fourier construction, so as to get round Theorem 4.1.'1, is to consider families 9,,,(2) that are not generated by a strict timefrequency lattice. This allows for a little leeway: Bourgain (1988) has constructed an orthonormal basis ( g j l j E J for t 2 ( R ) such that
6 [
uniformly in j E J, where. z, = dz ~19,(+)1~, = 4 (IJ,(E)(~.'(Note t-hat wavelet bases do not satisfy such a uniform bound. ) Giving up the lattice structure therefore permits better l d z a t i o n than allowed by the BalianLow theorem. However, Steger (private communication, 1986) proved that even slightly better localization then (4.2.9) is impossible: L ~ ( w does ) not admit an orthonormal basis ( g j ) j eJ satisfying
uniformly in j, if E > 0. This approach can therefore not lead to good timefrequency localization. There is another way in which we can try to break away fkm the lattice scheme (4.1.1). Note that in (4.2.9), (4.2.10), "time-frequency localization" stands for strong decay properties of the g,,, (gm,n)A away from the average values x,, This corresponds to a picture in which both gm,rn and (g,,,,)" have essentially one peak. Wilson (1987) proposes instead to construct orthonormal bases g,,, of the type
where
has two peaks, situated near
and
-9,
;
$
'3 a
4
4
?i
TIMEFREQUENCY DENSITY AND OFWHONORMAL BASES
121
with ,& ; 4, centered around 0. This ansata Fhangea the picture completely. W~lson(1987) proposes numerical evidence for the existence of such an orthonorma1 basis, with uniform exponential decay f& f, and dA, 4.; In his numerical construction he further "optimizes" the l d & i O n by requiring
Sullivan et al. (1'387) present arguments explaining both the existence of Wilson's basis and its exponential decay In both papers %ere are infinitely many functions-42; as rn tends to oo,the #$ tend to a Limit f u n c t b 42. The moral of Wllson's construction is that orthononnal b p r i t h good phase space localization seern possible after dl if &modal functions as in (4.2.13) are used. Note that many of our wavelet constructions, frames as well as the orthonormal bases we saw earlier, have these two peaks in frequency (one for > 0, one for < 0). In the case of frames, or for the continuous wavelet transform, the two frequency regions can be separ9ted (corresponding to one-frequency-peak functions; see 53.3.5.A or (2 4.9)), but this does not seem to be the case for orthonormal bases. We will see later that the two frequency peaks of $J need not be symmetric: there even exist examples with l]$b-2 JtSP 4 jJ;(<)12 arbitrarily small (but strictly positive!). However, there is no example, so far, of reasonably well-localized funct~ons$* wlth support c R* and such that the ($&,,; rn, R E Z, e = + or -) constitute an o r t h o r m a l basii lor L2(R),mrresponding to wavelet ba&s with only one "peak" ia f r m c y . (Equivalently, there is no example of a reasonably smooth h c t i o n 9 = $+ such that the functions 2m/2exp (2x2 2" ~) ~ ( 2 ~ m,n 0 , E 2, are an orthonormal basis of L2(~+).) It is believed, without proof so far, that qo sueh b&i exists.' But let us return to Wilson bases. If one giwa up the restriction (4.2.13) (if have exponential decay, then th& queditiea decay exponentially fast fm, in Im - m't, In - nfJ anyway), then Wilson's ansatz (4.2,11), (4.2.12) can be dramatically simplified. In Daubechies, Jaffard, and ~ o u r d(1991), * a construction is p r o p d that uses only one function 4 Explicitly, this conetruetion defines
<
<
? :
1
t k
,
(G)
A
gm,n(z) = fm(x - n),
m E N \ {0), n E
,
(4.2,14)
*
CHAPTER 4
etc -
.
with I E N, u = 0 or 1, and P = 0, o = O excluded. The result of all these phase factors and alternating signs is that
If we relabel Limp,,
in; (4.2.14) %y defining G,
, ~ r Et N,n f Z as
then
and for P > 0,
This comtruction {as'weU ae others mentimed below) shows therefore that the key to obtaining good time-fiequenq localization (4 can be chosen so that 4,
4
have exponential decay) and orthonormality in the windowed Fburier framework
is to use sines and cosines (alternated in an appropriate way) rather than complex exponentials. But let &I get back to (4.2.14), (4.2.15) and sliow how this construction can lead to an arthonormat basis. As 11wa1,we only need t o check lfgm,,,il = 1 and C:=l C n El(h, ~ qm,n)lZ = llh1I2. We immediately havesl1g~,nlf = IIfi II = IIbII, and for m > 1,
(weaseume q5 is red, for simplicity). Hence
/Jgrn,,lt= 1 for all m,n if
14 #(€I 4 ~ + 4 =~6 m0 . On the other hand,
TIMEFREQUENCY DENSITY AND OFCI'HONORMAL BASES
123
This equals ljhH2 if
A few simple manipulations lead to
f
If k is odd, k = 2k1+ 1, then this reduces to
which is zero, since the substitution 'P = - (e + 2k' + 1) transforms (4.2.21) in its negative. If k is even, k = 2k1,then (9.2.1 9) reduces to
The {g,,,; m E N \ {0), n E Z) therefore constitute an orthonormal basis if 4 is a real function satisfying (4.2.18) and (4.2.22). Nde that integratkmg(42.22) over t , between 0 and 2n, mttomatically leads to (4.2.18), so that we really have cmly the single condition (4.2.22) to satisfg. This %wmsout to be easy: we can take for i-ce support 4 C j-2a, 2x1, so th& (4.2.22) is automatically satisfied for k' # 0, and we only need t o check Q(f 2 ~ 4 = ) (21)-I~ This is true if, e.g.,
xez
+
otherwise ,* with u as in (4.2.4). if u is Cm,then the f,,, harve Becay faster thanany imretee polynomial, but, es for the Meyer basis, tkntmwical decay xnay be slow. Faster decay for the f, can be obtained with mmampctly supported #. To construct such a #, satisfying (4.2.221, we can again use the Zak transform, now nosPLalized so that ( Z h ) ( t l) ~ ,= (43r)'l2 rhiu h(4r(r - Q)
.
t€Z
'
CHAPTER 4
124
With this normalization, Z is again unitary from L2(R) to L2([0,1J2). It is not hard to check that (4.2.22) is equivalent to
(Full details are given in Daubechies, J d a r d , and J o d (1991).) This suggests the following technique for constructing $: %
Take any h such that
Define 4 by
If h. and both have exponential decay, then 4 turns out to have exponential decay as well. Figure 4.4 shows the graph of d, and 6 when h is a Gaussian. (Gaussians do indeed satisfy (4.2.24).) An interesting observatidn is that (4.2.23) is exactly equivalent to the requirement that the &,,(z) = e2*'4 (Z- $), or equivalently, the @,(<) = erint #J(( - m ) , with m, n E 2,comtitute s tight frame (with necessarily redundancy 2) for L2(R). The construction (4.2.25) can then be interpreted as the transition from a general frame, generated by h, to a tight frame, by application of ( F ' F ) - ' / ~ (see note 11 after Chapter 3, or Daubechies, Jaffard, and J o d (1991)). This W i h n basis can therefore be viewed as the result ef a clever "weeding" process on a (tight) frame with "twice too many" elements. Many variations on thi W i h n scheme are possible. Laeng (1990) has constructed an extension of the above scheme in which the frequency spacing need not be as regular as here. Auscher (1990) has reformulated the whole carnstruction: starting directly from (4.2.16), (4.2.17) as an ansatz, he derives all the results without use of the Fourier transform, and constructs different exampIes. In particular, he obtains examples where, in the notations of (4.2.17), the "windown 4 is compactly supported, which is very useful in applications. (The decay in frequency is less crucial, as long as it is 'C-reasonable.") These examples can also be viewed as the result of a "weeding" procedure on the tight frames with redundancy 2 obtained by taking wo& = K in 53.4.4.A. Other windowed Fourier bams using cosines and sines rather than complex exponentials, and leading to good timefrequency i d i t i o n , have been found by Malvar (1990) and Coifman and M e w (1990). Malvar's paper again uses alternating sines and cosines; he presents applications of his c o ~ c t i o nto rspeech coding. Coifman and Meyer's "localized sine basisn starts h m a partition of R in intervals, R = bj, a t t l l ,
U
JEL
;
; 7 7
i
h Y
j
TIMEFREQUENCY DENSITY AND OFWHONORMAL BASES
t
FIG.4.4. Tha functiow
t$
and
4 c o r n d i n g to (4.2.25)
af h(z) = 7r-l''
125
exp(-xa/2).
i
i
is
2~
< aj+l and limj,*, a, = foo. They then build window functions w, locatized around these I, = [a,, overlapping slightly with the neighboring with aj
'.
55 %b a:
w,(z) = 1 for
U, + t j
5 %5 @,+I -€,+I
,
>
8
where we assume that the e k satisfy a j +c, a,+l -€,+I for all j. Moreover, we require that w, and wj1 complement each other near aj: wj ( 2 ) = w,-~ (2aj -;z ) > %$ T:,.k' h and W ; ( X ) W;-~(Z) = 1 if 1x ajl 5 c,. (All this can be achieved with smooth ;. 4n, 3; one can take, for instance, w,(z) =sin(; v( 2-a&,+e i)] for Ix-a,l y, and l+C'ti)] for (X - ~ ~ + €j+l, ~ with l w satisfying (4.2.4) wl(z) = m[lv ( = - ~ ~ + ~ :Ti L~J., snd (4.2.5).) C o i and Meyer (1990) prwe that the family ( u , , ~ j, ; k E Z),
+
with
I
-
<
&
I
<
I
constitutes an orthonormal basis for L ~ ( R ) consisting , of compactly supported functions with fast decay in frequency. This basis has moreover a very interesting property: if for any j E Z we define Pj to be the orthogonal projection onto the space spanned by the {u,,~; k f 2) (Pjis "morally" the projection into [aj, aj+l]), then Pj P , + 1 is exactly the projection operator P,, associated ' to [aj, that we would have obtained if we had deleted the point a j + ~ from our "slicing" of R (i,e. if we had started with the sequence t i k , tik = ak if .k 5 j , iik = ak+l if k 1 j 1). This property makes it possible, to split and regroup intervals at will, adapted to the application one has in mind. A very nice discpion of this whole construction, with full details, is Auscher, W e b , and W i h a u s e r ( 1992). So there is, efter all, more to orthonormal windowed Fourier bases than was expected even only a few years ago. None of these bases, however, are unconditional bases for P ( R ) if p # 2. This is one point where wavelet bases have the advantage: they turn out to be unconditional bases for a much larger family of function spaces than even these "good" windowed Fourier bases. We will wme back to this in Chapter 9.
+
+
.
Notes. 1. Rieffel's proof does not produce an explicit f orthogonal to all the h,,. This is a challenge to the reader: find a (simple) corutruction of f I g,,, for all m, n, for arbitrary wo, to with %to > 2n.
2. For orthonormal bases the proof is much simpler. In this case we need not bother' with the Zak transform, which was only i n t r o d u d to prove that if Qg, Pg E L2, then Q j , Pfi E L3 as well. For orthonormal beses we can start directly with point 5, establishing (Qg, Pg) = (Pg,Qg), which is impoeeihk by point 6. This is the original elegant proof in Battle (1988).
3. If the $,,,(x) = ao-"I2 $ ( ~ - ~- nbo) x constitute a (tight) frBme, then 80.. do the $m,nn(x) = cro-m/2$#(--"x - &I), with @(x) = ( b ~ / b o ' ) ~ ~lr(boxlbo'). '~
4. To illustrate this, the following e;cample shows that the complex exponenexp (27rinz) do not constitute an unconditional basis for LP([O,lj) if p # 2. One can show (see Zygmmd (1959)) that
In both caees, x = 0 L the womt singularity, and the integrability of powers of these functions an (0, I] is determined by their behavior around
TIMEFREQUENCY DENSITY AND ORTHONORMAI, BASES
127
0. The first function is in LP for p < $, the second is not, even though the absolute values of their Fourier coeffic~entsare the same. This means that the functions exp (2rznx) do not constitute an unconditional basis for L ~ / ~ ( [I)), o, The Haar h i s adapted to the interval [0,11 cbnsists of { 4 ) ~ { ~ , 6m, ~ ,n~E; 2, m 5 0, 0 n 5 2Iml - 11, with r$(x) 3 1 on [O, 11. ThlS basis is orthonormal in L2((0,I]), and is an unconditional basis for LP([O, I]) if l
<
B
5. The following is a sketch of the proof that F(b) = {$kt,; m, n E Z], with $&,, as In (4 1 8), constitr~trsa Riesz basis (i.e., a IinearIy independent frame) for LZ(R) if b is close t o 1. First of all, we can still apply (3.3.21), (3.3.22) to find estimates for the frame bounds. For b # 1, &(27r(2k + l)/b) 0, but if b < 2, then only k = O* f l , lt2 1 4 to nonzero PI. In the cornputation of (4 2.6) (with (2k 1) replaced by (2k l)/b), only a finite number o f t contribute, so that this expression is continuous in b. I t follows that thc "rest terms" in (3 3.211, (3.3.22) are continuous in b as well; since (3.3.21)=(3.3 22)= 1 if b = 1, we have A > 0,B c oo lor b in "" a neighborhood of 1. It remains to pkvc* that the @&,, are independent. To do this, we construct the operator S(b),
+
-
i
+
+
!
s ( b ) j=
1( j *@k,n)&," . m,''
Clearly, S(h)$L,, = zl,&,,. To prove Independence of the ,$ ,: it is sufficient to prove that ((S(b) f((2 C({{ll, unifornlly in f E LZ(R), for some
i t
Using that for JHlkJ = J R k l Jwe , have
I+
*
*4
-
n
we obtain
128
CHAPTER 4
- Because of the support properties of 4, only m' = 0, f 1 contribute in this sum.If m' = 0 or 1, then any choice of n gives the same result; if m' = 1, then the sum may have one of two possible outcomes, dependingon whether - n is odd or even. On the other hand, wing the decay }$(x)t 5 CN(l ( ~ l ' ) - ~of ~,one easily checks that Cn,EZ I($&,, $:,,)I converges and . is continuous in b for m' = 0, f1. It follows that the coefficient of I} f)la in the right-hand side of (4.2.26) is coritinuous in b;-since it is 1 for b = 1, it is > 0 for b in a neighborhood of 1.
+
& They satisfy
instead.
7. After the first printing of this book, a proof was found by P. Auscher, to he published in the Colnptes Rendus dc llAcaddrnie Scientifique, Paris, under the title "I1 n'existe pas de basev d'ondelettes r6gui&res dans I ' c s p e de Hardy HZ(R)." Explicitly, he proves that it is inlpossible that 7 E C1and 1~[€)1+ }v1(5)15 C(1 + KI)-" with a > 1/2.
CHAPTER *
5
Orthonormal Bases of Wavelets and Multiresolution Analysis n
The first constructions of smooth ort honormal wavelet bases seemed a bit muaculous, a s illustrated by the proof in 54.2.A that the Meyer wavelets constitute " aa orthonormal basis. Thw situation changed with the advent of ~nultiresolution analysis, forlnulated in the fall of 1986 by Mallat and Meyer. Multiresolutlon d y s i s provides a natural framework for the understanding of wavelet bases, a& for the construction of new examples. The history of the formulation of m u ~ l u t i o analysis n is a beautiful exarnpte of applications stimulating theoretidl development. When he first learned about the Meyer'basis, Mallat was workingm&*age analysis, where the idea of studying images simultaneously a t different scales and comparing the results had been popular for many years (see, e.g., Witkin (1983) or Burt and Adelson (1983)) This stimulated him to view orthonormal wavelet bases as a tool to describe mathematically the "increment in informationn needed to go from a coarse epproximation to a bigher resolution approximation. This insight crystallized into multiresolution &pis (Mallat r) (1989), Meyer (1986)).
Ic t
I
.
5.1.
The basic idea.
A multiresolution analysis consists of a sequence of successive a-praximhtion spaces . More precisely, the closed subspaces VJ satisfy1
VJ
a .
- V2 C Vl
C
U4
Vo C V-l C = L2(a) ,
C
.
-'
(5.1.1)
.
r
(5.1.2)
JEZ
k
'
2
If we denote by P, the orthogonal projection operator onto 6 , then (5.1.2) ensures that lim,+-, P, f = f for all f E L2(R). There exist many ladders of spaces satisfying (5.1.1)-(5.1.3) that have nothing to do with Umultiresolutionn; the multiresolution aspect is a consequence of the additional requirement
CHAPTER 5
130
That is, all the spaces are scaled versions of the central space Vo. An example of spaces 5 satisfying (5.1.1)-(5.1.4) is
We will call this example the Haat murtiresolution cmalysis. (It ia i e a t e d with the Haar basis; see Chapter 1 or below.) Figure 5.1 shows what the projection of some f on the Haar spaces Vo,Vil might look like. This ex.ample d w exhibitf another feature that we require from a multiresolution analysis: invariance of fi tinder integer trabslations,
f
E
Vo
*
-
f (. n) E Vo for all n E Z .
Because of (5.1.4) this implies that if f €
1
(5.1.5)
I
5, then f (- - 2'n) E % for all n E 2.
Flnally, we requlre dso that there exiets t$ E Vo so that
i:
wher;, for dl j, n E Z. 4,,,(rJ= 2-3i2 B ( 2 - J t -n). ~ o ~ e t h e(5.1.6) i, and (5.1.4) imply that (d,,,; -nE Z) is orthonormal basis for 25 for all j E Z. This lest requirement (5.1.6) seems a bit more "cpatrived" than the other ones; we win see below that it can be relaxed considerably. In the example given a h , a possible choice for Q is the indicator filnctioa for [O, 11, $(a) =, 1 if Q 5 x 5 1, &(xi= 0 otherwise. We will often call 4 the "scaling function" of the multiresolution analysis The basic tenet of mdtiresolution analysis is that whenever rr collection of closed su€xigmessatisfies (5.1.I)-(5.1.6), then there exists an orthonormal wavelet basis {$J~,~; j, k E 2)of L~(R), $J,k(z) = Z - J / ~ + ( ~ - J X - k ) , such that, for all f in L 2 ( ~ ) ,
(Pfis the orthogonal projection onto %.) The wavelet J! ,T a, moreover, be constructed explicitly. Let us see how. For every 1 E Z, define W, to be the orthogonal complement of 6 in We have V,-I = V,.@ Wj (5.1.8) and
(If j > j', e.g., then W, c V j l l Wj,.)It follows that, for j < J ,
I
I
MUI,"lREsOLUTloN ANALYSIS
FIG 5 1
A functton j and ttr pro)eetrons onto V-1 and Vo
*
&
where dl thae aubepeces are orthogonal By virtue of (5 1.2) and (5.1 3), this implies
L ~ ( R =@ ) W, IEZ
.
(5 1-11)
a decompasition of ~ ~ into ( 1mutually ) orthogonal subspscee. Furtherwe, the '>W, spaad inherit the acdlng property (5.1.4) fmm the 6:
&rmula (5.1.7) is equivalent to saying that, for fixed 3, (GJSk; k E Z) constitutes 'an orthonormal basis for W, . Recsuse of (5.1.11) md (5.1 2), (5.1.3), this then atomatically implies that the whole collectim (+,,k; j, k E 2)is an orthouorma1 for L2(B).On the other hand, (5.1.12) ensures that if ($oScl k E Z) is an onormal basis for Wo, then k E 2) will likewme be an orthonormal for W,, for any j E 2. Our task thus reduces to finding 11, E Wosuch that (. k) constitute an orthonormad bash for Wo. To construct thii t,h, let us write out some interesting properties of 4 and Wo.
-
Since t$ E
c V-1,and the
are an orthonormal twb in V'l, we
h, 4-1,.
O= n
I
(5.1.13)
132
CHAPTER 5
with hn = (4,
(hn12= 1
and
$-I..)
.
(5.1.14)
nEZ
We can rewrite (5.1.13) as either
where convergence in either sum holds in La-sense. Formula (5.1.16) can be rewritten as = ~ ( [ / 2 &/2) ) (5.1.17) where 9
Equality in (5.1.17) holds pointwise almost everywhere. As (5.1.14) shows, mo is a 2lr-periodic function in
L2([0, 2x1).
2. The orthonormality of the $(, - k) leads to special properties for
no. We
have
implying
li(~ + 27r.f?)12= (2n)-l
a.e.
(5.1.19)
I
Substituting (5.1.17) leads to
-
(C = (12)
splitting the sum into even and odd t!, using the periodicity of q,and applying (5.1.19) once more gives
3. Let us now characterize Wo:f E Wois equivalent to f E Since f E V-1, we have
and f I Vo.
i
133
MULTIRESOLUTION ANALYSIS
I
q
with f,, = (f, &-I,,,).
This implies
where
mf is a 2n-periodic function in L2([0,2~]);convergence in (5.1.22) holds pointwise a.e. The constraint f L Vo implies f lhYk for all k, i.e.,
-
hence
where the series In (5 1.23) converges absolutely in L' (1-n,TI). Substituting (5.1.17) and (5 1 21), regrouping the sums for odd and even e (which we are allowed to do, because of the absolute convergence), and using (5.1.19) leads to
Since n o ( < ) and mo(C + n) cannot vanish together on a set of nonzero nieasure (because of (5 1.20)), this implies the existence of a 2n-periodic function X(C) so that
and This last equation c& be recast as A(<)
= e'C
2C)
,
.
(5.1.27)
where v is 2n-periodic. Substituting (5.1.27) and (5.1.25) into (3.1.21) gives f(() = rnO(t/2 + 4 40 4(2) (5.1.28) 1
where v is 2n-periodic.
-
CHAPTER 5
134
4. The general form (5.1.28) for the Fourier transform of f E Wo suggests that we take (5.1.29) = ee'* m ( S / 2 4- r)&/2)
&a
as a candidate for our wavelet. (5.1.28) can indeed be written as
Disregarding convergence questions,
so that the $(- - n) are a good candidate for a bas~sof Wo. We need to verify that the Q0.k are indeed an orthonormal basis for Wo. First of all, the properties of q a d ensure that (5.1.29) defines indeed an L2-function E V-I and _L & (by the analysis above), so that ~+5 E Wo. Orthon~nnalityof the #o,k is easy to check:
4
Now
Ce IS(< + 2 ~ 4 =P Ct lmo(t~+ =! + r)12 I & c /+~~ J ) I = = Imo(C/2 + t2 C P(t12 + 2rn)l2 + I ~ o ( ( / ~C ) I ~li(c/2 + r + 2 d i 2 n
n
= (24-' [Imo([/2)12 = (%)-I a.e.
+ Jrno((/2 -t?r)I2] (by'(5.1.20))
8.e.
(by (5.1.19))
.
Hence J & $ ( x ) 1C,(z- k) = dm. In order to check that the &,k are indeed a basis for all of Wo, it then suffices to check that any f E Wocan be written as
17n]2 < 00, or
with
-
n
P
MULTfRESOLUTION ANALYSIS
135
with 1 In-periodic and E L2((0,2~1).Let us go back to (5.1 28). We have = U(F)S(C),~ t: $h @ I Y ( E )= I ~21gI dc I X ( C ) IBY ~ . (5 1 22).
i(o
I
On the other hand, by (5 1.25),
Hence ff'@ Iv(i$)P= 2z (1f}I2 square integrable 23-periodrc 7
< 00, and f u of the form (5 1 30) w t h
-
We have thus proved the following theorem. THEOREM 5.1 1 If a ladder of c l o d subspaces (&)IEZ zn L2(lZ) satwfles $he andzttons (5.1.1) -(5.1.6), then them ezuts an assoctated orthonormal wavelet b u (@$,*, 1,k E Z) for L ~ ( w such ) h t
$
k
One possrhlity for the wnstructton of the umvelet #J rs 4
*#
4(0 = eg'2
:3
T!
ma(f/'l+
4kU2) ,
((ath p,-, 0s defined by (5.1.18), (5.1.Id)), or egustldentlly
I
h:
$(z)
k
&"
=
h x(-l)*-'
h-n- 4(2z - n ) .
n
t @ h P
( m t h convergenee of the last senes tn ~~-gense).
.?-
Note that II, is not determined uniquely& the mdtiresolution analysis ladder and requirement (5.1.31). if $Jsatisfies(5.1.31), then so will any y3# of the type
6% "
*
= P(t1$(€I ,
(5.13 3 )
with p 2~-periodicand lp(c)( = 1ae."In partiah, we can choose p(c) = pc etmP wUh m E 2, /pol = I, which correspasda to a phase change and a shift by m for 3, We will use this freedom to define, - b h d of (5.1.32),
CHAPTER 5
136
or occasionally 9n =
h - n + l + 2 ~1
with appropriately chosen N E 2. Of course we can take more general p in (5.1.33), but we will generally stick to (5.1.34) or (5.1.35).4 Even though every orthonormal wavelet basis of practical interest, known t o this date, is associated with a rnultiresolution analysis, it is possible to construct "pathological" II, such that the $,,*(z) = 2-jI2 $(2-3s - k) constitute an ort honormal basis for L' (W) but are not derivable from a multiresolution analysis. The following example (due to J. L. Journd) is borrowed from Mdlat (1989). Define (27r)-'I2
if
9 5 Kt 5 R
or 4r 5
1 ~ 51
otherwise .
,
(5.1.36)
x,
We immediately have 11+,,t (1 = (1+11 = 1. FUhermore, 27r (4(2j[)(' = 1 a.e. By Tchamitchian's criterium (3.3.21)-(3.3.22) the Ijl3,k therefore also constitute a tight frame with frame constant 1, provided
4+
One easily checks that support n[support (2k + 1)2n2'] has zero measure for all t 2 0. k E Z, so that (5.1.37) is indeed'satisfied. It. then follows from Proposition 3.2.1 that the Ilj$ constitute an orthonormal basis for L2(R). If ?I, were associated with a multiremlution analysis, then (5.1.29)and (5.1.17) would hold for the corresponding scaling function t$ (with possibly an extrti p({), Ip(<)J= 1 a.e. in the formula for +-see (5.1.33)). It then follows from (5..1.20) that (5.1.38) /#(6)12+ ~ & t ) d= ~ 1&([/2)l2 t
4
which implies, for [ # 0 ,
One easily checks from (5.1.36) that this implies
(
0
otherwise
.
If there existed a 2~-periodic% so that (5.1.17) held for this 4, then we would have Imo(<)l= 1 for 0 5 5 47117. By periodicity this would imply Imo(c)l = 1
4
137
MULTlRESOLUTION ANALYSIS
- as well for 27r 5 E 5 18n/7; hence I?q,(()l I&[)I = ( 2 ~ ) - I / 2for 27r 5 ( < 16n/7, even thodgh 1&2[)1 = 0 on this interval. This contradiction proves that this orthonormal wavelet basis is not derivable from a multiresolution analysis. Nute that @ has very bad decay. It is an open problem whether this kind of "pathologyn c$n persist if some smoothness is imp& upon (i.e., decay on T+!J).~ For later convenience, we note that ia terms of the h,, equation (5.1.20) can be rewritten a s
4
_
+ -
(This follows easily from writing out the explicit Fourier series for lrno(C)[' Im(C+ r)I2.)
-.*
6.2.
+
Examples.
Let us see what the recipe (5.1.34) gives for the Haar multiresolutio~ianalysis. In that case, 4(x) = 1 for 0 5 x < 1,0oth-, hence h, = f i
J
- n) =
& &(x) 4(2x
1/& 0
if n = 0,1,
otherwise.
1 if O < z < $ ,
-1
if
4
0 otherwise.
This is the Haar basis, which is no surp* we already saw in 51.6 that this wavelet basii is associated with the Haar multiresolution analysis. i The Meyer basis also fits neatly into this scheme. To see this, define 9 by r
$K) =
[
MI1 2~13,
(2~)-l/~, (2*)-1/2
[:Y
($1~1-
I)] , 2 ~ / 35
It1 5 4 ~ 1 3 ,
otherwise,
0
4
is plotted in a smooth function satidjing (4.2.4) and (4.2.5). 5.2. It ia an easy consequence of (425) that EkEz1)((+27rk)12 = (2~)-' , nrhich is equident to torthonormality of the #(- - k), k E Z (see 55.2). We then define Vo to be the closed subspace epanrled by this orthonormal set. Similarly, is the closed space spanned by the k E Z. The V, satisfy (5.1.1) if and if 4 E V-l,i.e., if and only if there exists a 2x-periodic function m, square -able on [O, 2x1, so that
where
iw
Y i8
+,,
CHAPTER 5
FIG. 5.2 The d s n g functm 4 for the Meyer basas, unUI v(x) = x4(35-- 842
-+ 70z2 -
2oz3).
In this particular c&, ma(()
=
and
%
can be easily constructed from 6 itself: This is 2~-periodicand in h2([0,2 4 ) ,
& CLEZ #(2(< + 2x4)).
~ ( t 1 2 i ) ( ~ 2 )=
6 C &t+ 4
4
C€Z
\
at) &/2)
=
+
(because [support 4(./2)] and [support J(- Ire)] do not m r l a p if e # 0)
= ict) (beeausu 6#(t/2) = 1 for 6 E support 8).
I leave it as an (easy)exerc-ke for the reader to check that the V, also satisfy properties (5.1.2), (5.1.3) ((5.1.4) and (5.1.5) are trivially satisfied already; see also 55.3.2). Let us now apply the recipe (5.1.29) to find $:
4(€)
eela %K/2
=
+
&t/2)
Ji; E l 2 C #({ + a(u+ 1)) & ~ / 2 ) ?EZ
=
6eel2 [& +
i-
- 2x11
(for all other C, the supports of the two facton, do not o v e t w.
It is w y to check (see also F i i 5.3) that tbit, is equivalent to (4.2.3). The phaae factor eel2 which w w needed for the Umiraculous cancellatione" in 54.2 occurs here naturally as a w~se~uence of the general analysis in $5.1. Before we discuse other 'emmples, we need Co relax condition (5.1..6). "C
139
MULTIRESOLUTION ANALYSIS
I
Relaxing some of the conditions.
3
.3.1. Riesz bases of scaling functions. The orthonormality of the 4 ( . - k ) (5.1.6) can be relaxed: we only need to require that the g5(- - k) const~tute a Riesz baas. The following argument shows how to construct an orthonormal h i s ##(- - k) for Vo starting from a Riesz basis {gl( - k); k E Z} of b. Tb &(. - k) are a &esz basis for Vo if apd only if they span 'I/o and if, for all (cki&z E
I
e2(z),
t
where J > 0, B < oo are independent of the c, (see Prelimideries). *
&t
and
so that (5.3.1) is equivalent to
We can therefore define q5# E L2(W) by
x,
+
1J6 (( 27re)12 = (27r)" ax., which means that the q5#( - k) are Clearly, orthonormal. On the other hand, the space ~ b #spanned by the q5#(. - k ) is given by
{ f; j = v J# with v 27r- perlodic, v E L ~ ( { o2x1)) , = {f; j = q with yl 27r- perlodic, vl E L ~ ( [ o2,~ 1 ) ) (use (5.3.2) and (5.3.3))
=
4
= Vo
(since the .($t
- n ) are a Riesz basis for VO)
Using the scaling function as a starting point. As &scribed in $5.1, a multiresolution analysis consists of a ladder of spaces (V,),EZ and a special function C#J E Vo such that (5.1.1)-(5.1.6) are satisfied (with (5.1.6) possibly relaxed as in 55.3.1). One can also try to start the mnstrudion from , an appropriate choice for the scaling function 4: after all, Vo can be constructed from the q5(. -k), and from there, all the other V, can be generated. This strategy is followed in many examples. More precisely, we choose # such that 5.3.2.
where
C, Ih12 < oo, and
-wethen define V.to be the c l o d subspace spanned by the 4j,k, k E 2,with &j,k(z)
= 2
(
2 - k). The conditions (5.3.4) and (5.3.5) are nk E 2)is a Ricm basis in each 5,srid
sary and sufiicient to ensure that {#j,k;
141
MULTIRESOLUTION ANALYSIS
that the V, satisfy the "ladder property" (5.1.1). It follows that the Vj satisfy (5.1.1), (5.1.4), (5.1.5), and (5.1.6); in order to make sure that we have a rnultiresolution analysis we need to check whether (5.1.2) and-(5.1.3) hold. This is the purpose of the following two propositions. PROPOSITION 5.3.1. Suppose 4 E L2(lR) satisfies (5.3.5), and define V, = Span {&k; k E 2). m e n n,,~ 5 = (0).
' +
Proof. \
1. By (5.3.5), the t$O,k constitute a Riesz basis for Vo. In particular, they constitute a frame for Vb, i.e., there exist A > 0,B < oo so that, for all L?
f
E Vo,
A
1111 51C ~
~ ( f a, s ) ! '
sB II~II~
(5.3.6)
k€Z
(see Preliminaries). Since 4 and the 4J,kare the images of Vo and the &,k under the unitary map (Dl f)(x) = 2 - ] I 2 f ( 2 - J x ) , it follows that, for all f ~ % , A ilfl12 l l(fl h.r)12 6 B 11f1I2, (5.3.7)
k~z with the same A,
B as in (5.3.6). .
2. Now take f E n,,~ 4. Pick e > 0 arbitrarily small. There exists a compactly supported and continuous j so that 11f - j l l L a 5 c. If we denote by P, the orthogonal projection on V, , then
i-
-+
llf
-PJ~II
= ll<(f
- j)ll 5 llf - ill 5 e ;
hence
*$
Ilf 11 I + H ~ ~ IforI all j
-
.
(R choeen so that [-R, R] contains the compact support of
j)
(5.3.8)
CHAPTER 5
142
with SR,J= UkEz[k-2-JR, k + 2 - $ R ] , where we assume that j is l&ge enough so that 2-3 R
< 3.
4. We can rewrite (5.3.9) as
cI
JR
t i , #J.t)~zs 2 ~ 1 1 i 1 1 2 - d~ x,(B)I,(Y)I~
k
(5.3.10)
where X, is the indicator function of SR,, i.e., xJ(y) = 1 if 9 E SR,, x, (y) = 0 if y @ SR,. For y @ 2, we obviously have xJ(3)-+ 0 for j -+ 00. It therefore follows from the dominated convergence theor$.m that (5.3.10) tends to 0 for j + oo. In particular, there exists a j such that (5.3.9) < r 2 ~ Putting . this together with (5.3.8), we find 11f 11 5 2 ~ Since . r was arbitrarily small to start with, f = 0. This pr9ves that (5.1.3) is satisfied. For (5.1.2) we introduce the additional hypotheses that is bounded and that J dx 4(x) # 0. PROPOSITION 5.3.2. Suppose that 4 E LZ(R)satisfies (5.3.5) and that, moreover, &(<) M bounded for a11 and mtinuous near = 0, with &(o) # 0. Define 5 as above. Then UJEz 5 = L 2 ( ~ ) .
4
<
Proof. 1. We will use again that (5.3.7) holds, with A, B independent of j.
€1~; Ez6)'.Fix c > 0 mbitrarily small. There exists a compactly supported "Co" function j so that ( 1 f < e. Consequently, for all
2. Take f
jllL1
J=-jEZ
On the other hand, by-(5.3.7),
3. By standard manipulations (see Chapter 3) we have
with
@a %
$!
143
MULTIRESOLUTION ANALYSIS
Since
is Cm, we can find C so that
It then follows that
c -
c ldlli- C (1 + r2PPJ)-'I2 J4 (1 + 11l2)-' LZO
(
use sup (1
+ y2)[l + (z-y)2]-1[l + ( z + y)2]-1 < ba twice
)
z.tt~IP
- C"2-J. <
(5.3-15)
4. Putting (5.3.12), (5.3.13), (5.3.14), and (5.3.15) together, we find
2~
*
1q
lj(2-
JolzI~ C ) 5I ZB C +~ cfl2-
J
.
(5.3.16)
Since d(t) is uniform% bounded as well as continuous in ( = 0, the lefb hand side of (5.3.16) converges to 27r[&(0)1~1( f I(:, (by the dominated convergence theorem) for J --+ 00.It therefore follows that
-
I I ~ I rI t~ &~ o ) ~ - ' c ~ ,
with C independenof c. Combining (5.3.17) with Ilf obtain Ilf l l ~ aI + I l f H ~ aI(1 + cl8(fi)l-')~. Since c was arbitrarily small, f = 0.
(5.3.17)
-fll~, 5
e, we
RE~~ARKS.
ri
::$
f
+
,
1. If slightly stronger conditions are imposed on 4, then Propositions 5.3.1 and 5.3.2 can be proved with easier estimates. In Micchelli (1990), for example, the aame conclusiom are derived if t$ is continuo~wand satisfies Jq5(x)J5 C(1+ 1~1)-'-~,CtEz$(x - Q)= const. # 0, which implies both 4 E L1 and J 4 ( 4 # 0.
*
.1
\Q
P
2. The extra condition that # be continuous in 0 in Proposition 5.3.3 is not necesssry. The following is an example of a multiresolution analysis in which the scaling function is not absolutely integrable. Let KM,p, be, respectively, the multiresolution spaces, the s d i g funGion, and the wavelet for the Meyer wavelet basis, with v E Coo(see 55.2). Let H be
eM
I
CHAPTER 5
144
<
<
the Hilbert transform, (Hj)"(t) = f ( ~if) > 0, -f (0if < 0. Define V, = H y M ,4 = ~ 4 Because ~ the . H'ilbert transform is unitary and commutes with scaling and with translations (in x), the V, still constitute a multiresolution aadysis, and the &,k are an orthonormal basis in Vo. But 6 is not continuous in 0. Because 0 @ support ( G M ) , 4 = ( H @ M ) A is a Coo-functionwith compact support, so that @ itself is Cw with fast decay. This is therefore an example of a very smooth wavelet with good decay, associated to a rnultiresolution analysis with bad decay for 4.6 Note also that 4M and 4 satisfy (5.1.17) with the same w ,illustrating that the c, in (5.3.4), or equivafently ma, do not determine 4 uniquely, and that decay of the c, as (n(-+oodoes not ensure decay for 4.7
3. If 4 is bounded, and continuous in 0, then the condition J(0) # 0 is necessary in Proposition 5.3.3. This can be seen asfollows. Take f E L ~ ( R ) , ' j # 0, with support j c I-R,R], R < m. If Uz, V, = L2(W),then f = limJ,, P- J j. But
6
as in (5.3.13). Since is continuous, the first term tends to A-' 27r]4(0)l2 J]f1I2 for f-too, by the dominated convergence theorem. The second term can be bounded exactly as in (5.3.15), so that this term tends to zero for J-roo. It follows that
Since 11f (1
#
C
w
0, this implies 6(0) # 0.
4. The argument in points 3 and 4 of the proof can also be wed to prove 14(0)12 5 B/2n. We have indeed
where (R1 can be bounded by C2-J for nice f . The other term tends to 2?rl$(o)(?llf 1' (see 4). 'Ibgether with remark 3 above, this implies A/27r 5 ( & o ) ( ~5 B/27r. In particular, if the &,k are orthonormal, then A = B and I$(o) ( = ( 2 ~-'I2. )
c
LO",4(0) # 0 (with 4continuous in 0) imply certain
3
'5. The conditions
E
restrictions on the c, aa well. l3quation (5.3.4) can be rewritten as
1
P
145
MULTIRESOLUTION ANALYSIS
with -(<) = C, c, cant . In particular, b(0) = %(0) &o), which implies ~ ( 0 = ) 1 (since &o) # 0) or
Moreover, (5.3.18) implies that m,-, is continuous, except pomibly near the zeros of In particular, mo is continuous in = 0. If, furthermore, lJ(<)l 5 C ( l I
6.
<
+
I&< +
4#
c
.
I&<
4
t$#(x)/c,@(<) d(r)
n
Together with c, = 2, this implies C, qn= 1 = C, czn+l. This is consistent with the admissibility condition for $.8 Note also that C, c2, = 1= czn+~is equivalent with M i d (1990)'s condition Ct d(x-e) = const. # O if I#(x)I C (1 1x1)-'-' and i f 4 i s ~ o n t i n u o u s . ~o
Cn
,
x,
8.
<
+
k,
C
All this suggests the following strategy for the construction of new orthonorma1 wavelet bases:
.
4
Choose # so that (1) 4 and have reasonably decay, (2) (5.3.4) and (5.3.5) ere satisfied, (3) J& 4 ( 4 2 0
If necessary, perform the "orthonormalization trick"
.e
Finally,
[Cr
G(f)
+2
+
= e"/2mo# ([/2 r)4#(2), with m,#(<) = ~ e ) l [Ct ~ ~l i~( 2/ ~ 2.1)12J-1/2, or equivalently
with m$(<) =
+
$;
e-i4.
~ ( 6 )
146
-
5.4.
More examplea: The BattleLemariB family.
CHAPTER 5
The Battle-Lernarid wavelets are associated with muItiresolutiori analysis ladders consisting of spline function spaces; in each case we take a B-spline witb knots at the integers as the original scaling function. If we choose C#J to be the piecewise constant spline, 1, 0 5 x 5 1 , 0 otherwise,
4
then we.end up with the H e e ~h i s . is the piecewise h e a r spline, The next
1 - 1 4 1 0 5 1x1 5 1, otherwise,
!.
plotted in Figure 5.4a. This 4 satisfies
,?
see Figtire 5.4h. Its Fourier trmsform is
and 2z
x,,ZI$(€+ 2nP)I2
1 -
(b1
J+
msc = + ( I
+ 2 m 2{/2).lO
;.-dl e(2xf
51 *(a++ 1
1 e( 2x-1 1
0 -1
n ~ 5.4. . &(a- 1).
0
1
The piwetube tincar 8-spiine 4; it racisjiee 4(2) = $4)(20
+- 1) f d(22) + 4
MULTIRESOLUTION ANALYSIS
147
Both (5.3.4) and (5 3.5) are satisfied, q5 E 1;' and l &$(x) = 1 # 0. The V, constitute a rnultiresolution analysis (consisting of piecewise linear functiom with knots at 232). Since q5 is not orthogonal to its translates, we need to apply the orthogonalization trick (5.3.3)
Unlike q5 itself, 4# is not compactly supported; its graph is plotted in Figurh 5.5a. To plot @#, the easiest procedure is to compute (numerically) the Fourier caefficlents of [I*+ 2 w2t/2]-1/2,
and to write 4#(x) =
C,
c,
d(x - n). The corresponding m$ is
FIG. 5.5. The a d i n g funetton p and Lhc wavelet T(, for the hnear splrne Battk-Lmrorit anrshuctton.
148
and
CHAPTER 5
4is given by
Again we can compute the Fourier coefficients d,, of [(I - sin2 4) (1 +
+
cos2 ~ / 2 ) - l ( l cos2 4)-1]1/2,and write
This function is plotted in Figure 5.54. In the next example q5 is a piecewise quadratic B-spline,
(
$ ( 2 +1)2 ,
( (I as plotted in F~gure5.6a. Now 4 satisfies =
-llx
otherwise ,
a 4 ( 2 +~ 1) + :4 ( 2 ~ )+ 2 #(25
-
1)
+
- 2)
(see Figure 5.6b); we have
g+g
a n d 2 ~G l i ( < + z ~ e ) l = ~ COSC+& C O S ~ [ = & + ? - ~ C O S E + $ C O S ~ C Again (5.3.4) and (5.3.5) are satisfied, and 4 E L1,with r d z )(XI # 0. TI. q5(. - k) are not orthonormal, and we need to apply the orthogonalization tric (5.3.3) to find q5# and rnf before we can construct 1C, Graphs of $# and ar given in Figure 5.7. In the general case, q5 is a B-spline of degree N, $J
where K
=.
0 if N is odd, K: = 1 if N is even. This q5 satisfies
#(x)
= 1 anc
if N = 2M is even
i f N = 2 M + 1 isodd.
MULTIRaSOLUTION ANALYSIS
5 6 The quadmtrc B-spltne 4, trarulated so that sotufies &(z) = +4(2z 1) :&(22) :4(22 - 1) + :4(22
Fw
+ +
+
t b knota am
at the tntqcra It
- 2)
tb(< +
Explicit formulas for El 27rt?)i2, for general N, can be found, e.g., in Chui (1992). In all cases, 4 satisfies (5.3.4), (5.3.5). For even N, 4 is symmetric for odd N, around x = 0. Except for N = 0, the r$(. - k) are around x = not orthonormal, and the orthogonalization trick (5.3.3) has to be applied. The result is that support 4# = W = support tl, for all the Battle-Lemarik wavelets. The "orthonormalized" $# has the same symmetry axis as $. The symmetry axis of .rjl always lies at x = (For N even, tl, is antisymmetric around this axis, for N odd, $ is symmetric.) Even though the supports of 4# and 9 "stretch out" over the whole line, 4# and $Jatill have very good (exponential) decay. To prove this, we need the following proposition. PROPOSITION 5.4.1. Assume that $ has ezponenttal decay, I$(z)l 5 C e-rlzl, and that, for some a 5 7 (a> O),
4,
i.
sup l(ep 4)"(<)I lalla
Assume also that 0 < a 5 El I$([ 27r1)12]-1/2. d(4)
+
< C (1 + l€o"-C
EL I ~ ( +C 27r&)12.
(5.4.1)
Def;ne 4# by J#(<) Then d# has ezponenttd decay as well.
=
I
CHAPTER 5
FIG.5 7. The scaL*ngfinctton 4 and the wavefet Q for tke quodmfic spline BotUc-Lcmohl conslructson.
1. The bound lt$(z)l 5 C eeal'l implies that 6([) Eas an analytic extamion i ( 2 ) E L2(R) for all It21 < a. The to the,strip IIm [I < a,and that^ same is true for =
6(€)
a-0.
2. For fixed b , define FtZ([,) =
and
+
+ it2)$(-(I
- B).Then
MULTIRESOLUTION ANALYSIS
(We have used that
xt If (x+2~4)15 (2n)-IJ d x If (x)l+ J dz I f ' ( ~ ) l . ' ~ ) +
Consequently, Ct Ft, ({I 2 4 converges absolutely if tIm t21 < 7. Similar bounds, combined with the dominated convergence theorem, show Q, (tl 2 4 is analytic in { = El 2f2 on the atrip lIrn [I < 7. that
+
+
+
'g %
F" ' i
$"
3. The function G(()= C, /J(E 2 ~ l )has ( ~ thus an-analytic extension t o JIm €1 < 7. Since G is periodic, with period 2n, and Gill > a > 0, this implies that there exists 6, p m i b l y smaller than 7 , sa that ReG(<)1 a/2 for IIm
4
4. On the other hand, (5.4.1) implies that
4
4#is analytic, and is
5 a. It follows that on tItn
i
Consequently,
5 C
,. (~OROLLARY 5.4.2.
; L
,+;
fi-
e-IZ2
,
for
< min ( 6 ,a) ,
rn
All the Battle-LemariC wcnrekb t,b and the corresponding dwnonncrl sding junctions $# have ezgonential decay.
153
MULTIHESOLUTION ANALYSIS
Regularity of orthonormal wavelet bases.
5.6.
For wavelet bases (orthonormel or not-see Chapter 8) there ki a link between the regularity of $J and the multiplicity of the aeto at ( = 0 of This is a cansequeace of the following theorem (stated sad proved in greater generality than needed here, for later convenience). THEOREM 5.5.1. Suppose f , j am two functions, not identicufly -tat, . ruch that (13,kl& ~ , v = ) 6jj&w
4.
= 2-3/2 f ( 2 - ? ~- k), Afk(x) = 2 - ~ / * f ( 2 -~ k). ~ S?~ppo8t!fkjt Ij ( ~ ) l5 C(I ~ x l ) - ~tkth , a > rn I, and SUppoJe ~ o f t E P , with j ( f )
with
+
+
bounded fur C I rn. Then
1. The idea of the proof is very simple. Choose 3, k, j', k' so that f,,&is rather spread out, and very much concentrated. (For this expwitory point ' . only,we assume that f has compact support.) On the tiny support of f,lbkt 4he slice of fic L'seenplby ~ , I can ~ I be replaced by its Taylor series, with as -many terms as are well defined. Since, however, $ & f,,k(z) f 3 ~ , k l( x ) = 0, ~ , I , ~ I
this impies that the integral of the product of j and a polynomial of order m is zero. We can then vary the beahions of BB given byk'. For each location the argument can be repeated, leading to a whole family of different polynomials of or& m which all give zero i n t w a l when multiplied with Thb leads ta $be dmired moment mudition. But let us be more precise as follows.
1.
2. We prwe (5.5.1) by induction on L. The following argument works for both the iriitid step and the inductive step. Assume J & znj(x) = 0 for n E N, n < e. (If e = 0, then this amounts to no assumption at all.) Since f(') is continuow (e < m), and since the dyadic rationals %-jk, ( j ,k f 33) are denee in It, there exist J,K eo that ~ ( ' ) ( z - ~ K#) 0. (Otherwise f (0 0 would follow, implying f I constant if t = 0 or 1, which we know not to be the case, or, if 2, f = polynomial of order t - 1 1 1, which would imply that f is not bounded and is therefore also excluded.) Moreover, for aryt c > 0 there exists d > 0 so thet
=
!>
I-
fa)
if
(E
C
-! n=O
f
(
-
2 ( x ~2
'
2-J KJ 5 6. Take now j > J , j > 0. Then
I<
9- 2 - J ~ 1 1
CHAPTER 5
Since $ dz x n j ( z ) = 0 for n
< P, the first term is equal to
)I \
(!!)-I
j(t)(z-J~)z-(r+l)~
(5.5.3)
4
1 1
1
Using the boiindedness of the fin), the second term can be bounded by
t
where we replaced the upper integration bound by oo in the first tern, And where we tiscd in the second tern1 that (1 2Jt)-l j (1 1 ) - I 2-1 ( 1 t ) - I for t 6. Note that C1,4 only depend on C, a,and e; they are independent of c, 6, and 3. Combining (5.5.21, (5.5.3), and (5.5.4)
+
&!& +
+
>
<
leads to
Here c can be made wbitrarily small, and for the corresponding b we can choose 3 sufficiently larp to make the second term arbitrarily small as well. It follows that J dx xCf (x) = 0.
When applied to orthononnal wavelet b, this theorem haa the following carollades: COROLLARY 5.5.2. If the + J , r ( ~ )= 2-5l3 ltr(2-f~ k ) constitute an orthononnal set in L~(R), with I+(x){ C (1 + IX~)-~-'-' ,$ € E ( R ) and boundedforesm, t h e n J d z d $ ( x ) = O f o r 4 ! = 0 , 1 , . - . , m .
<
-
Proof. Follows immediately from Theorem 5.5.1, with f = j= $.
m
155
MULTIRESOLUTION ANALYSIS
1. Other proofs can be found in Meyer (1990), Battle (1989). Both proofs *
6 &2j
'
work with the Fourier transform, unlike this one. Similar links between zero moments and regularity also constituted part of the "folk wisdom" among Calderbn-Zygmund theorists, prior to wavelets.
2. Note that we have not used multiresolution analysis to prove Corollary 5.5.2 or Theorem 5.5.1, nor even that the $,,t farm a basis: orthonormality is the only thing that matters. Battle's proof (which insplred this one) a h uses only orthonormelity; Meyer's proof wea the full framework of multiresolution analysis. o
"0 q=.
COROLLARY 5.5.3. Suppose the +,,k are orthononnd. Then st aa tmpossrble 9 has ezponential decay and that 3 E P , wafh all denvatrues bounded,
A?t
,t
that
i
d e s s @ E 0.
J ,
%3
%f. ?.a'
t 4 2
%
If $ E CaOwith bounded derivative, then by Thebrem 5.5.1, J dz zt+{x) = = 0 fm d l t E N. 0 for all t E N; hence $
$1
=o
2. If 11, has exponential decay, then 4 is analytic on some strip (Im
3b:
This is the trade-off announced at the end of the'last section: we have to for exponential (or faster) decay ip either time or frequency; we cannot ckave both. In practice, decay in z is often preferred over decay in [. A last consequence of Theorem 5.5.1 is the following factorization. ROLLARY 5.5.4. Assume that the @,,k m t t t u t e an orthononnal basts multsresolution analysas as described m $5.1. If and @ € C" w t h ~ ( ' 1 bounded for t 5 m, 1.14), factontm as c$ooge
m(€)= &&E
2
)
m+l
*L(O
9
(5.5.5)
L u 2n-penodic and E Om.
+1. ByCarollary5.5.2,
k
(
$41 c=o = O f o r e ~ m .
q .the other hand, G(()
+
4
= e - a c / 2 ~ ( < / 2 n ) 4((/2). Since both and are in P ,and $(0) # 0 (see Remark 3 at the end of §6.3.2), this means
4
I
1
CHWI'ER 5
166.
that mo,b m times Merentiable in = T , and
st'
=O
forl<m.
.
We will come back to the regularity of wavelet bases in Chapter 7. 6.6.
Camectian with subband fil&ring schemes.
Multireeolution atralysis leads nat"maUy to a hierarchical and fast.'scheme for tbe computation of the wavebt coefficients of a given function. Suppose that we have computed, or are given, the inner products of f with the at some given, fine d e . By nscaiing our ''unita" (or rescaling j ) we can assume that the b l of this fine d e is j = 0. It is then easy to compute the ( j , qJVk) for j 2 1. F i i of all, we have (see (5.1.34))
''
where gn = ($,
= (-1)" h-,+'.
Consequently,
Xt foUm that
(1,#l,k) = i.e., the
(f,#I,))
ete obtained
K=z (f. &,,'I
*
condvins the quence ((f,&,n>)rez with
(a-n)nez,and then retaining only the even eamples. Similarly, we have
which can be d to compute the (f, CpfC) by meana of the 8-tion* (convolution with & decimation by factor 2) from the (f, t$j'5-l,k), if these are
MUL-LUTION
ANALYSIS
known. But, by (5.1.15)
h-a 6j-~.n(z)
=
(5.6.3)
I
n 95
.*** whence
cf,
+
a*
..4,
h.k)
=
CK G
(f9
-
oj-1.n)
n
(5.6.4)
The procedure to follow is now cles: -ts from the (f, h,,),we compute tbe (f,$ l , k ) by (5.6.2),and the ( j , &,k) by (5.6.4). We can then apply (5.6.2), -" (5.6.4) again to compute the ( f , h C ) ? . : ( j , hL)from the ( j , 4l,n), etc.. . .: at of the c o r n > every step we compute not only the ~p~veW Coefficients ( j , *spending j-!evel, but also the ( j , C$j,k)for the same j-level, which are useful for computation of the next level wavelet coefficients. $a,:the The whole process can also be viewed as the computation of successively -.$*,i m m e r approximations of f , together with the difference in "information" be+ :.--., t-areen every two successive levels. In this vieolr we start out with a fine-scale : . t -5 -approximation to f , f" = Pof (recall that P, is the orthogonal projection onto , V ;.are will denote the orthogonal projection onto Wjby Q,), and we decompose P E V o = Vl@IVl into f O = f 1 + 6 1 , r b e r e f ' = = Plf is the w x t coarser approximation of f in the multiresolution analysis, and 6' = p - j' = QlfO = Ql f is what is "lost" in the transition f o -+ f l . In each of these V, , W,spaces we have the orthomrmal b d (41.k)kEZ, (+l,k)kEz, respectively, so that
., i
%
'
P = C c".,n, f1 ='C 41,*,P = C dt *I,,
% r.+w
n
n
F ' k r (56.2), (5.6.4)
-
n
give the effixt on-tha & e% okcthe ieni
ck =
h-lt
4,
9n-1L C: .
d: = n
a
ii = (dncZ and (Ab)k =
With the notation a = csn rewrite this as
cl=BcO,
basis
-+
.
(5.6.5)
Ena2*-n bn, we
d'=GcO.
The coarser approximation f1 E Vl = Vz GI W2dtk again be d e c o m p d Into f' = f 2 +a2, j2 E Vz, 62 E W2,with f2
=
C
:C h e n
n
@=x&
h,n
-
n
We again have c2=fjc1,
6=Gc1.
Schematically, all this can be represented as in F
i 5.8.
CHAPTER 3
FIG 5 8
Schem4kc rrpnsentokon of (5.6.5).
In practice, we will stop after a h i t e number of levels, which means we have rewritten the information in ((f, &,n))nE&= d ) as d l , dL,d3,. - - ,dJ and a find coarse approximation cJ, i.e., ((f,' @ j , k ) ) & ~ ,> = I , , J and ( ( j ', # ~ , l i ) ) b ~Since ~. all we have done is a succession of orthogonal basis transformations, the inverse operation is given by the adjoint matrices. Explicitly,
hence
(use (5.6,1), (5.6.3)) . In electrical engineering terms (5.6.5) and (5.6.6) are the analysis and synthesis steps of a subbandfiltenng scheme with exact reconstruction. In a twechgnnel subband filtering scheme, an incoming sequence is wnvolved with two different filters, one low-pass and one high-pass. The two resulting sequences are then subsampled, i.e., only the even (or only the odd) eatnee are retained. This is exactly what happens in (5-6.5). For readers unfamiliar with this "filtering" terminology, let me explain briefly what it means. Any square summable sequence ( c , , ) , ~can ~ be interpreted a& the sequence of sampled valw ~ ( nof) a bandlimited function 7 with suppod j. c [-?r, rr] (see Chapter 2),
(e)n,Z
sin r (x - ra) n
A filtering operation corresponds to the multiplication of j. with a 2~-periodic function, e.g.,
I
MULTIRESOLUTION ANALYSIS
The result is another bandlirnitd function, a * 7 ,
or
sin n ( x - n) n
is m d l y concentrated on (-n/2, n / 2 ] , hghThe filter is low-pass if pars if &ll-r,xl is mostly concentratad on {C; r / 2 5 I[) 5 n ) ; see Figure 5 9. The "idealn low-pass and high-pass filters a& lit(() = 1 if ](El < n/2, 0 if n / 2 < (
I
2
for n = 0 ,
for n = 2k, k # 0 ,
a,L =
- I k f ~ Rl = 2k 1)n
(2k
h ~5.9. . A low-rn
+
+1,
(sobd line) ad a high-pw filter (dashed kne)
%en the ideal law-pess filter Is apptied tr, 7 , the result is a bandlimited Eunction with support C [ - z / 2 , ?r/2j. Gnch a function is completely determined by its sampled values in 22, and we bavt (m'(2.1.2)) sin [ ~ (-x2n)/2] n(t-2n)/2 .
160
CHAPTER 5 \
Similarly, the result of applying the ideal h~gh-passfilter to 7 is a frequencyshifted version of a bandlimited fuaction with support c ( - r / 2 , r / 2 ] . Such a function is again completely determined by its sampled values on 22,
Since the even-indexed entries of the convolutions of the a:, a: with the cr suffice to characterize a~ * y and a~ * 7 completely, it makes sense to retain only them after the convolution. This is the rationale behind the decimation by a factor 2 in subband filtering, also called "downsampling." Reconstructing the orignal c, from the two filtered and decimated sequences,
is easy:
(since tiL + hH = 1)
/ I
Distinguishing between even and odd m, we find
This ctto also be rewritten as
This last operation can be seen ae the result of interleaving both the 4 and $.t with zeros (i.e., constructiog new sequences with m o odd ex@h, and with even entjes given by the consecutive
4, e);
convolving these interleaved (%peempledn)sequences with the mters aL, a*, respectively;
I
a
adding the two results.
MULTIRESOLUTION ANALYSIS
161
Schematically, (5.6.8) and (5.6.9) can be represented as in F i 5.10. The filter coefficients a:, a t for the ideal h - p a m and hii-pass filters aL, aH decay much too slowly to be useful. In practice, one prefers t% use the scheme in Figure 5.10 with filters ao,a', go,i ' with much faster decaying coefficients. This can only be achieved if the corresponding 2.rr-periodic functions aO,a', hO,iil are emoother than aL,an. Thii means aliasing can occur: jaul,lalllook like "rounded" versions of aL,dl (as in Figure 5.9), which means their support is larger than [-u/2, r / 2 J and {t; a12 I KI 5 n), respectively. Consequently, a" + 7, a' * 7 are not truly bandlimited with maxfmal frequency 7r/2, and sampling them as if they were leuds to aliasing, as explained in 52.1. This hrrs to be remedied in the rewmtrwtkm stage: iio, 5' need to match a0 and a' to get rid again of the aliasing present after the decomposition. And even this "matchingn is only possible if aa and ' a sre already matched in some way. To find the appropriate conditions on these filters, it is convenient to use the 'z-notation," in which s sequence ( G ) , , ~ is represented by the formal series a ( z ) = LEZ an zn. If z = e-'E is on the unit cirde, then this is nothmg but a Fourier series; sometimes it is convenient to consider general x E C rather than bI = 1. The decomposition stage of the subband filtering scheme in Figure 5.10 can then be written as
Here crO(z) 2) is the t-notation for the convolution of a' and c; is equal to t e formal sequente b, z2", i-e.,b(z) with all
"6
C,
removed. The mnstructitm stage is
Fs.
5.10. Schmotsc representdion of the h p d h a and nsamstructia &ages oerticol dashed lme) Jn a Nbbwd Ptaing scbtme. Every Ictkr (a0,a ] , .) m a boz npumL m v d a i o n rrnth the m n q m m h g quma; 2 1 stands for the duwmam-
(separated by UK
..
4the even 2 f for upmmphg by 2 (interknuang unth x m s ) . Inthe W " u u e l a O = a L , a* = a H , 6 0 = 2 a L , ~ ~ l = % H ; t h e / i n o l ~ w u l m t i d
plhg by 2 (ntoiRing tOthe~,S=c.
"?'
where d ( z 2 )is the z-notation for the upsampled version of ct (zero8 have bee4 interleaved: ,c9(r2)= Cn-cl, ban). The total effect is
In this expression, the second term contains the al'iasing effects: c(-z) comesponds to a shiftlng of the Fourier series C, & e-'"t by a, exactly what you would expect from aliesing due to sampling ;at half the N'jtq$st tat& In order to eliminate aliasing, we therefore need
The first subband coding schemes wjthout aliasing date back to Egteban and Galand (1977). Ln their work, as in most schemes that will be considered in these notes, the sequences are real; they choose
so that (5.6.11) is indeed satisfied, and (5.6.10) simplifies to
= 4, then al({) = Cn aie-"'f is the "mirror" of a0 with respect to the "half-band" value 6 = 7r/2, since &I(<) = a! e-'"€ = ao(z - <). Filters chosen as in (5.6.12) are therefore called "quadrature znirrw filters'' (QMF).In practice, one likes to work with FIR filters (FIR = finite impulse response; this means that only finitely many an are nonzero). Unhtunately, there exist no FIR a0 EO that ~ O ( Z-) ~ = 2, so that E cannot be identical to c in this scheme. It is nevertheless possible to find aO so that - ~ ~ ( - 2is )c~ h to 2, so that the output of the scheme is indeed dose to the input. There is by naw an immense literature on the design of various QMF; see the issues of IEEE IIFans. Acowt. Speech Sagnal Process. for the kt 15 years. There a h exist many generalizations to splitting into more than 2 bands (GQMF-eneralized QhlF). In Mintzer (1985), Smith and Bzmqpll(19@),and Vetterli (I-) p scheme different from (5.6.12) was propasf&
I f - ao is symmetric, a!,
En
a'
(E)
=w
a-1 aO(-z-'1
,
= aa(z-I) , iiE[r) = ttl(z-l) = zap(-z)
(6.6.M)
GO+)
.
It is easy t9 check that this W i d i t w again (5.6:11), end that (&&lQ)
becornea
MULTIRESOLUTION ANALYSIS
163
For t = eat and real a:, the expression between square brackets becomes &[la0(e-'c)12 t la0(-e-ac)12] = [lao(t)12-t laO(c =)I2]. There now exist FIR choices of a0 for which this is exactly 1, so that we have exact reconstruction in the subband filtering scheme. Smith and l3arnwell (1986) named filters chosen as in (5.6.13)13 "conjugate quadrature filters" (CQF), but the term has not become as generally popular as QMF. One last remark before we retilnr to wavelets. The whole purpose of subband filtering is of course not to just decompose and reconstruct: a simple wire, instead of the scheme in Figure 5.10, would be far cheaper and more efficient. T h e goal of the game is to do some compression or processing between the decomposition and reconstruction stages. For many applications (image analysis, for example), compression after subband filtering is more feasible than without filtering. Reconstruction after such' compression schemes (quantization) is then not perfect any more,14 but it is hoped that with specially designed filters, the distortion due to quantization can be kept small, although significant compression ratios are attained. We will come back to this (albeit briefly) in the next, chapter. Back to orthonormal wavelet bases. Formulas (5.6.5), (5.6.6) have exactly same structure as (5 6.8), (5 6.9), respectively. Going from one level in a &tiresolution analysis to the next coarser level and the corresponding level of~rrwlets,and then doing the reverse operation, can therefore be represented by B diagram similar to Figure 5.11. Here (h), = h-,, (g), = g_, (see above). If we assume that the h, are real, and if we also take into account that gn = (-l)"h-,+l, then we can identify Figure 5.11 with 5.10 with the choices
4
+
Up to a trivial sign change m a' and Zil, thie corresponds exactly with (5.6.13). This means-that every orthonormal wavelet bases associated with a multiresolution analysis gives rise to a pair of CQF filters, i.e., to a subband filtering &me with exact reconstruction. The reverse is not true: in an orthonormal besis construction we necessarily have aO(l) = C h, = 2'j2 (see Remark 5 at the end of §5.3.2), but there*exist CQF for which a (1) i~ close to, but not equal to, 2lI2. Moreover, all the examples of orthonormal bases we have see11 so far correspond to infinitely supported , and hence to non-bite aequencea h,; for sppiicationa, FIR filters are preferred. Js it possible to construct orthonormal wavelet baees corresponding to finite @tars? What does it mean for these filters to carrespond to, e.g., regular wavelets? Hoar can wavelets be uwfd in filtering contexts? All these are questions that will be addreseed in the n a t chapter,
a
Notes. 1. I chooee here the same neeting arder (the more negative the index, the the order kgtz the -):ee in the ladder of ~ e apacc& v Ttris is
.Q
that foltms naturally from'the notathi of non-orthogonal watteleta as iniSated by A. Grommann and J. MorIet. It is non-standard, however: Meyer (1990) uses the reverse ordering, more in accordance with establiihed practices in harmonic analysis. For applications in numerical analysis,Beytkin, Coifman, and Rokhlin (1989) find the ordering presented here the most practical.
I
2. We impose here no a priori regularity or decay on 4, udfke, e.g., Meyer (1990).
I
-
3. Equation (5.1.33) characterizes all the p d b l e Lemma 8.1.1 in Chapter 8.
$#.
This follow from
1. In case 4 has compact support, and we would like $ to have the seme compact support, (5.1.35) is the only poesible choice.
5. It is generally believed that time is no such 'pathologlcsl" example with continuous $. Another challenge for the reader! 6 (1991) has When this book wss in its last stages, I heard that M proved that if is compactly supported (continuous or not), then it is automatically ~ssociatedwith a multiresolution analysis. Thii aalves the open problem for one very important special case.
+
gk
bds
0. Note that there exins @ E l 4 ~a,tht the are an s r t h o w d for Vo and d# E L1(R),d i k e 4. It suflkes to take J#(f) = A(€) &), where X is 2n-periodic and A(() = sign(() . e g / a for KJ 5 x. Thh -4 : is again a Schmirtz function. The same Hilbert transfoh trick can tn applied to other multiresolution andyees, such aa the Battle-LmuuiC axbe or the constructions with compactly supported 9 in the next chap*.
.
7. If ue impoee that uniquely.
4 is continuow at ( = 0, then mo dbee determine 4
8. The continuity of 4,together with &(o)# 0, implim that ~ ( 0 ) 1 and that rn,, is costinuou m 6 = 0. It kUm t h t fe contincous in = 0.
<
MULTiRESOLUTION ANALYSIS
Since
+-
+ lrnt((+ *)I2 = 1, it f o l h that (+~L$#Iis contiWow Consequ~ntly,14(t)( = ~ r n f(~12) + r ) ]lt$((/2)l is continuwas
l4(~)l2
t
=m = 0 since J, h a to be admissible, this implies %#(lr) ~ ( x= )O This providm another derivation for (5 3.20).
in in
r'
,
ir
I
9. Proof We prove that C c z , = 1 = Cczn+l if 1q5(x)1 5 C(l IZ()-'-' and 4 is continuous
+
= 0; hence
Cr#(x - f ) = const. # 0
+ Define f (x) = Ct r$(x - 1). The conditions on 4 ensure that f is well defirretl and continuous We have
Hence f
IS continuous,
perlodic with period 1, and
It folfows that f a constant. e C, &(r 0 = c implies &%n) ' mo(tl2) t $ ( t / 2 ) ;hmce
-
= S , ~ ( ~ U ) - ' / ' c. But &(<) =
+
( n # ) 0,then J(m(2n 1)) = 0 would fdbm for all n E 2 , in contradiction with C t$(n .t 2m)12 > 0. Hence ~ ( m =) 0, or
If ~
C
~ 2 = n
1=C
*
~ ~ n + l .
10. An easy way to compute the Fourier awffieents of Et following:
-
i' $
4'
L
I
+2nZ)12 is the
For B-splme 4, these are easy to compute; see also Chui (19!42) for an explicit formula
* for 0 I y 5 2n,
CHAPTER 5
166
12. If f is given in "sampled" form, i.e., if we only know the f (n), then the inner products (f, can be computed.by a convolution (or filtering) operation, under the assumption that f e V,to start with (components of f orthogonal to I/o cannot be recovered). We have f = (f, h , k ) 40,k; hence f ( n ) = (f, 40,k) 4(n - k). Consequd1~,
xk
zk.
i.e., the (f,&,k)
(Emd(m) e-'mt)-l.
are the Fourier coefficienta of It follows that (f,
14
(Enf (n) e-'M)
C q k=) C,, ak-,, f (n), where
2*
a,,, =' ( 2 ~ ) - '
etM
0
13. For convenience, they chose a ' ( ~=) ~ ~ * - ' a ~ ( - z - Z0(z) ~ ) , = z 2 N aO(z-I), iil(2) = z ao(-z) rather than (5.6.13), with N E Z picked so that all the a', 6J are polynomials in z (no negative powers). One finds then E(z) = z~~ ~ ( a )which , corresponds t o a pure delay in the reconstruction.
14. This is an azgument used by fans of the Esteban-Galand-type QMF filters: these fail to give exact reconstruction from the start, but their deviation from exact reconstruction can be made small in comparison with digtottioas introduced by quantization.
i i 5
A1
-
Orthonormal Bases of Compactly
Supported Wavelets
Except for the Haar basis, all the examples of orthonormal wavelet bases in the previous chapter consisted of infinitely supported functions, as a result of the orthogonalization trick (5.3.3). To ccmstruct orthonormal examples in which # is compactly supported, it pays to start from q-(or, , equivalently, from the subband filtering schem55 6) rather thaa from 9 or the V,. Ia $6.1 we show how to construct so that (5.1.20J is saijsiled as well as (5.5.5) for some N > 0 (a necessary condition to have some regulGty for +). Not &verysuch is associated to an o r t h o n o d wavelet basis, hawever, an issue addressed in 556.2 and 6.3. The main results of these two sections are summarized in Theorem 6.3.6, at thi 4 of 56.3. Section 6.4 contains examples of crunpactly supported wavelets genemfhg orthonormal bases. The orthonormal wavelet bases thus obtained cannot, in general, be written in a closed d y t i c tom. Their graph can be computed with arbitrarily high precision, via ae algorithm that I cdl the * q a ~ cade algorithm," which is in fact a "rdinemt *en as ueed in c o r n aided design. All this is discussed in 56.5. A Iat of this material back to D a u W - (198Bb); for many of the results, better, simpler, or more general pro~&have hem found since, and I have given preference to these new ways of looking at thin&. These d i h n t a p prow$m are'borrowed mainly from MaUat (1989), Coben (1990),Lawton (1990, 1991), Meyer (1990), and Cohen, Daubechies, and F'eauveau (1992); for the link with rehemebt equations the referemxa are Cavtmtta, Dahmen, md Miccheili (1991) and Dyn and Levin (1990), as weli as earlier papers by these authara (see 56.5). .*
=
4
"
6.1. Construction of m.
t
In this chapter we are m & l y interested in constructing compactly supported wavelets $. The easiest way to ensure compact support for the wavelet $I is to Eamk the d i n g fundion $t with%* support (in its orthogona@d d o n ) . ft then f o U a &om tbe ddbRw6 ofthe k,
.e
T
168
CHAPTER 6
that only finitely many h, are nonzero, so that $ reduces to a finite linear combination oi compactly ~ilpportedfundions (we (5.1 34)), and therefore automaticallj~has wmpacf eugport itself. Choosing both $ and qb with- ampact support also has the advantage that the corresponding subband filtering scheme (see $5.6) uses only FIR filters. For compactly supported q5 the 2~-periodicfunction m,
becomes a trigonometric polynomial. As shown in Chapter 5 (see (5.1.20)), orthonormality of the b,,implies
where we have dropped the "almost everywherep became is necessarily cont tinuous, w that (6.1.1) has to hold for dl if it holds a.e. We are also interested in xm&g-+ and 4 reasonably regular. By Cerol1aty 5.5.4, this me& that rq-,should be of the form
with N 2 1, end C's trigonometric polynomial. Note %hateven without regularity constraint, we Beed (6.1.2) with M at least 1.' Putting (6.1.l), (6.1.2) tagether, it follows that we are W n g for
and
where L(c)= IL(()I2 is BfeO a polynomial in c a t . For our purpose it is connient to rewrite L(() aa a polynomial in ain22 = (1 - coe <)/2,
In terms of P, the constraint (6.1.4) become8
which should hold for all y we use b u t ' s theorem.'
E
[0,I], hence for all y
E
R To mlw (6.1.7) for P
169
COMPACTLY SUFPOEYTED WAVELETS ,$
THEOREM 6.1.1. If p l , p2 aw two polyromiab, of degree nl, nz, TWPCCttvelg, with no common zeros, then them ezijt mique polyromiab q g , qz, of d e p e nz - 1, nl - 1 , mpectiuely, so that
; k-
=1.
m(z) 4r(z) f pz(z) + **
(6.1.a)
.
t
Proof. 1. We first prbve existence; uniqueness follows later. We can assume that nr 2 n2 (by renumbering, if necessary). Since degree ( p z ) Idegree (pl), we can find pcdyoomialti a2{z),b(z),with degree ( a z ) = degree (pl) -
I
-
2. Similarly, we can find a3(x), h ( x ) , With degree(a3) = degree(p2) -degree (h), degree (4)< degree (b), that pl(4 = a3(4
bq(4
+ b3(5)
.
rc
We keep going with this procedure, with bn- 1 taking the role of pl in this last equation, and b, the role of b, b,-l(x) = &I+*(xS bn(x) + b n + l ( ~ .)
Since degree (b,) is strictly decreasing, this has to stop at some point, which Ss only possible if bN+l = 0 for some N,with bN # 0,
3. since
+
b ~ =-ON~b ~ - 1 b~
:
h
it follows that bN divides bNd2 as weil. By induction bN divides all the previous b,,, and pz, so that bN divides both pl snd pa. Since pa a d pz have no zeros in cxunmon, it follows tbat bN is e constant dr&arent ftom zero.
4.
?+*
8F
,
W e have now 9
bN
8 f;
= bN-2 - ON bN-1 = bN-2
- U N ( ~ NL UN-1 J bN-2)
= (1 -k t 3 ( ~z N - I ) ~ N-- ~ ON b ~ - s etc.. . .
!
I.
By induction 0
-
brv = ii#,, b ~ - k+ CN,). b~ -&-I ,
with i ~= --ON, , ~ 6 ~ =~1,16 ~ ~ =k &V,L + ~ - h , r ) a ~ - k ,&,N.~+I = iiN,k. It follom, again by induction, that degree(6~~,~) = degree(bNir-l)
-
.
170
CHAPTER 6
degree ( b ~ , ~ ) degree(h~,k) , = degree (bnr-c) - degree (bnr-;). k = N 1, we find
-
For
with degree ( ~ N , -1) N = degree ( p l ) - degree ( b ~1)- < degree ( P I ) , degree ( ~ N , N 1) - = degree (pz) - degree ( b N - 1) < degree ( p z ) . (We have used that degree ( 6 ~ - 1 )I 1; if degree (bN- 1 ) were 0, then bN would be zero.) It follows that ql = ~ N , N - ~ / B92N = , iiNs-l/bN. solve (6:1.8) and wtisfy the desired degree constraints.
5. It remains to establish uniqueness. Suppow q ~ g2 , and dl, & are two solution pairs to (6.1.8), both satisfying the degree restrictions. Then
Since pt , pl have no %era6in common, this imples that every zero of m is a zero of ql - 41, with &t least the same multiplicity. If ql # #*, then this means degree (qt - 4,) > degree ( B ) ,which is impossible since degree (ql ), degree (ql ) < degree ( p z ) Hence gl = Ql. It then follows immediately that
92 = 42-
1. For later convenience (Chtspter 8), we have stated Baout's theorem in gxwter generality than needed in the present dwpter. In fact, it holds under even mow general conditions: if pl and pa have zeros in common, then (6.1,8) can still be solved if its right-hand side is divisible by the greatest common denominator (g.c.d.) of p l , f i . The proof is still the same, but bN is now the g.c.d. of pl, pz instead of a constant. The argument in the proof is nothing but the m~istructionof the g.c.d. by Euclid's algorithm; it works in many other ftmeworlcs than the polynomials (in any graded ring, in algebraic terminology).
2. It 1s clear from the construction of pl, n,that if pr and haves~fyr a t i c ~ nal coefficients, then so will -and qz. This wilt be useful in Chapter 8. o Let us now apply this So the problem at hand, i-e., (6.1.7). By Theorem 6.1.1 there exist unique polynomials qf, qz, of degree J N - 1, so that
Substituting 1 - y for y in (6.1.9) U to
the dquenese of 91, a thus implies q2(y) = q l ( 1 - 9). It foUows that P ( y ) = ql(y) Is a sdution of (6.1.7). In thii case, we can find the expiicit form of ql
without ewn using EucM'e dgoritbm:
1
v
where we have written out explicitly the k t N terms of the Taylor expansion for (1 - Y ) - ~ . Since degree(g2)5 N 1, q~ is &ual to its Taylor expansion truncated after N terms, or-
-
'.1(.)="i:
; 1
1
k=O
N +( k-l)
yk.
-
This gives an explicit solution to (6.1.7). (Fortunately, it is positive for y f [O, I], candidate for JL(t)J2.)It is the unique lowest degree solution, which we will denote by There exist however many solutioni of higher degree. For any such higher degree solution,.webave so it is a good
P N b 3
Thts a&eady implies that P - PN is divisible by y N ,
Moreover, P(p)+f3(1-y)=0, i.e., P is antisymmetric with respect to follows.
1. We q summarize all our findings as
PROPOSITION 6.1.2. A trigonometric polyonid
satisfies (6.1.1)
and only
of We fonn
L(<)= IL([)j2 a n be written acr
L(<) = p(mn2 t/2) with
P(v) = PN(v) + lN
and R is an add poly"omial, chosen such
,
~ ( - 94 )
,
that P(y) > 0 for y E (0,I].
This proposition completely characterha (m,,(t)12. For out purposes we need however rno itself, not (mo/2.So how do we "extract the square rootn from L? Here a lemma by Riesz (see Polya a d S+ (1971)) c o w to our help. LEMMA 6.1.3.Let A be a positive trigonometric plyrwmial invariant under the substzktton - (; A is necessarily of the form
<+
M
~ ( € ) = ~ 4 n m with~ a,,,eR, m=O
76,
Then t h e e d l s a higonometric polyromial B of onler k4, a. e.,
--
Pnwrf.
1. W; dan write A(€) ="pA(cos\l), wherem is a polgaomiaiof d real coefficients.This polynomial can be factored,
mM e h
M
I
I
where the zeros cj of pA appear either in complex duplets c,, Ej7 or in real singlets. We ccin also write
where PA is a p02ynomtai 01 deep.ee 2M. For Izf = 1, we have
I
the ttwo polynomids in the rigfrt aPd left-hand sides of (6.1.13) therefore 8grw on ail of C.
I
-
JG.
2. Ifc, is red, then the zem oft -cj2+)z2 are ej f For l g l Z 1, tbae w t~ d (d-Ynt. if C, = i l ) of the fonn rjl ~7'.F'o~ {cjl < 1, the two zeroe are compb conjugate and of absolute value 1, i.e., they nre of the form ehj, e-&j. Sbwe lcj 1 < 1, such zero6 correepond to ''physicaln zeros of A (i-e., to values of for which A(<) = 0). In order not to cause any contrsdiction with A 2 0, th& z;eros must haw even multiplicity.
<
I
I
T.
173
COMPACTLY SUPPOBWBD W A V E L m
+
'
3. If cj ia not real, then we consider it together with c& =
nojnial
4f
5. The p l y -
(f- q z + 43)(4 - 9% + ir2) has four zeros, el f ,/-
ds.
and
One easily chedrs that the fouo~eermare all difkent, and form a quadruplet 3 , 2'; , f , , 27 l .
4. We therefore have
. '"
*
*
P.(Z) =
I
>
1
J
-ZJ)(z
- Zj)(% - %yl)(Z -
where we have regrouped the two different kinds of zerwr.
A
,,
1
5 aM
5.
L
For z = e-e on the unit circle, we have
Consequently,
where
fs clearly a trigonometric polynomial of order M with real coefficients.
m
1. This proof is constructive. It uses factorization of a polynomial of degree M ,however, which has to be done numerically and may lead to problems if M is large and some zeros are close together. Note that in this proof we
need to factor a polynomial of degree only M, unlike some other procedures, which factor directly PA,a polynomial of degree 2M. 2. This procedure of "extracting the square root" is also called spectral factorization in the engineering literature.
3. The polynomial B is not unique! For M odd, for instance, PA may have quadruplets of complex zeros, and 1 pair of real zeme. In each 2 quadrupiet we can choose to retain either r,, Z j to make up 8 ,or ql,z ;'; in each duplet -we can choose either rt or r;'. This d m already for different choices for B. Moreover, we can always muftiply B with e'M, n arbitrary in Z. o
-
.
Together, Propasition 6.1.2 and Lemma 6.1.3 tell us how to construct aU the possible trigonometric polynomials mo satisfying (6.1.1) and (6.1.2). It is not yet clear, however, whether any such ?no leads to orthonormal wavelet basis. In fact, some do not. This will be discussed in the next two sections. Readers who would like to skip most of the technicalities cap find the main results summed up in Theorem 6.3.6 at the end of $6.3. 0.2.
Correspondence with orthonormal wavelet bases.
We start by deriving a-formula for a candidate waling function 4. Once this is done, we will check when this candidate defines indeed a hona fide multiresoiution analysis. If a trigonometric polynomial n o is associated with a m u l t i r ~ l u t i o nanalysis as in 55.1, and if thp corresponding scaling function @ is in L1(R), then we know that for all <, (6-2.1) = ~ ( < / 2&/2) ) .
J(c)
(See (5.1.17). Continuity of and mo dm us to drop the "8.e.") Moreover, we know from Remark 3 fobwing Proposition 5.3.2 that necesearily J(0) # 0, hence ~ ( 0 =) 1. Becaw of (6.1.1) thi4 in turn implitm ~ ( x =)0. It follows that, for all k E 2, k # 0, &kn)
= d(2 Y(2m + 1 ) n )
(for some C 1 0, rn E Z)
176
COMPACTLY SUPPORTED WAVELETS
xt
+
f ~ i n c e I&( 2n!)12 = (2n)-' (see (5.1.29)),this fixes the normalization of 4: ?, &o)l = (2u)"12, or I $& d(x)l = 1. It is convenient to choose the phase of 4 I. so that Jdz flz) = 1, Taking all this into account, it follows from (6.2.1) that
This infinite product makes sense: since %(() = 2-'12 hn e-("€ satisfies
Cn
Cn lhnl fnl < oo, and ~
( 0 =) 1,
hence w
fl 1 ~ ( 2 - ' < ) I5 exp )=I
? The infinite product
'
in the right hand side of (6.2.2) therefore converges abeoW i y and uniformly on compact sets.4 A4i thin =lies generally whenever 4 E L1,m d the h, have sdeient decay. is a trigonometric polynomial (only finitely many of the in present c e , bi&e different from zero), and we are looking for q5 with compact support. Tog&her with the obvious constraint 4 E L2, compact support for 4 means I#I E L', so that the above discussion applies. It follows that (6.2.2) is the I only possible candidate (up to a constant phase factor) for the scaling function - corresponding to a trigonometric poiynomial w constructed tat in $6.1. We now need to check that 4 satisfies some basic requirements for a scaling function. First of all, q5 is square integrable: LEMMA 6.2.1. (Mailat (1989)) I f mo w a 2n-pen'odac function sutufyrng (6.1.1), and tf (2r)-'I2 %(2-'C) amvnyes pmntwe a.e., then its hmit
!
nzl
ts
i n i 2 ( ~ ) , ' a n Ipllr2 d 5 1.
Proof. I. Define
fk(C)
=
X ~ - ~ , ~ ]=( C 1 if)
(2%)-'I2
[n:=l m0(2-j()]
ICI 5 T , 0 otherwise.
Then fk
--+
~ ~ - r , * ~ ( 2 - ~where (),
6 pointwiae a.e.
2. Moreover,
= (2r)-'
n k
2h+lr
d(
1=1
lm(2-%)I2
(by the 2%-periodicityof mo)
CHAPTER 6
3. It follows that, for all k , llfkl12
=
I1fk-11Ia
1 ;
- . = Hfol12 '= 1 .
Consequently, by Fatou's lemma,
/4I ~ ( c )5I ~lim k
-+
sup 00
J
~fk(<)l'
5 1.
Second, since nb is a trigonometric polynomial, the following k3-a borrowed from Deslauriers arid Dubuc (1987) provos that 4 haa compact support. LEMMA6.2.2. rg) = 1. ~ ~ with~ 6 7, = I, U~CR , r [ 2 - 2 4 ) za an entrm functzon of exponential type. In pmticulor, it ia the F o u w tmmform of a dutnbutzon with suppord in fN1, Nrf. hf, By the Pdey-Wiener theorem for distributions, it is sufficient to prove that. r(2-a ie an entire function of exponentid type with bounds
,n:
c,":~,
c::~,
u
I
nzl
n n rt2-'0
r(2-'C)
' ( N I llm (1) 5 4 ( 1 + 1 ~ 1 ) ~ exp
for Im F 2 0 ,
5 c2(1 + I E I ) ~ =exp (NZ~b 40
for~m (50
.
00
,
I
1
ji.1
for some Cl, C2, M I ,M2. We will only prove the h t bound; the m n d is entirely analogous. Define
Then
00
JJ r(2-30
=
fi
r1(2-$€) ,
m we only need to prove a polynomial bound for For Irn C I0 we have
nGl ra(2-j€)hb
2 0.
COMPACTLY SUPPOFU'ED WAVELETS
I
<
Take arbitrary, with Im I
If
(€1
> 0. If )I)5 1, then
2 1, then there exists j o
<
> 0 so that 2J0 < KI < ~ J o * ' , and
(1
+ C)l0+' eC
5 eC ( 1 + C) exp Iln(l+ C) In I<)/In 21
Combining (6.2.3) for polynomial bound. '
I[( 5
1 and (6.2.4) for
> 1 establishes the desired 17
So far, so good. All this is not sufficient, however, to define a bona fide ecding function. A counterexample is
This satisfies (6.1.1), as well as m(0)= 1. Substituting it into (6.2.2) leads to5
1
3 , 05x53, 0 otherwise.
This is not a "goodnscaling function: the &,,,(x) = &(z-n) are riot orthonormal, even though satisfies (6.1.1). Another way of looking at this ia to eee that
178
I
CHAPTER 6
(5.1.19) is not satisfied:
xt
+
that this m e s that ~ IT!)? = 0 for ( = ?f,so that 4kdi (11.3.2) is 'hob satisfied: the h , n are not even a Rieez basis for the space they span.= &k order to amid this kind of mishap, we have to impom extra conditions on to make sure that 4 generates a true m u l t h l u t i o n analpis. These dnditione ensure that
tor dl (. Once (6.2.5) is sstiafied, everything else works: the -s V, = Span(4>,,; n E 2) constitute a muftiresolution ~nalysis(by 85.3.2); in Vj , the ($j,n)nEZ constitute an orthonormal basis. We define t(, by i
+(x) = \/i
z(-l)n Q(ZI-
n);
(6.2.6)
n
thie is automirtically compa~tlysupported because 4 is and becauae only finitely many h,, differ from zero. The ($,,k)J,kEZ constitute then an orthonormal basis of compactly supported wavelets for L2[R). Before we go into the conditions on that ensure (6.2.5), it b interesting to remark that even if (6.2.5) is not satisfied, the function $ defined by (6.2.6) still generates a tight Frame, as p r d in Lawton (1990). PROPOSITION 6.2.3. Let mo be o trigonometric polyrornial satisbrag (6.1.1) and m(0)A 1, and let 4, $ be llre compacffp supported L2-functions defined by (6.2.2), (6.2.6). Define, usual, q j , k ( x ) = 2-]I2 $(2-lx - k). Then,for all f E L2(W),
be., the ($,,&; j, k E 2)h t i t u t e a tight ftnme for L2(R).
1. First reanember that (6.1.1) eau abm be written as
(see (5.1.39)). 2. 'XBke f compactly supported sad CQD.Then ,all j :
I (f, +j,k)lz
caaobrgee for
I
179
COMPACTLY SUPPOIZTED WA-
S" %
-
Choose K
80
+
thst 2-3 suppot(j ) n[2-3 suppwt(f) kj is empty d k 2 K.
Then
cl -c
&&
wE2-J .uppoP(f)
z,
5
g
we-'
/d* Id,
Qy Id(v - k)12 dy I~(Y - nlr~- C)I=
~~~port(l1
-
(because, for every C, the mts (2-'suppart(f) C PIK),~ do~not overlap)
*s"-
+ +
..
q$"
6 Simlarly,
11411'
CkI( f ,v&,k) I2
converges for all j
It ~seasy to check that the right-hand side of (6.2.9) is absolutely summable (use that only finitely many h,, ate muzero), ~o that we may jnvert the order of the summations. 4. If n, m are even, n 3 2r, m =P 28, we have
hb-p = 4,. = 6,,,l.
(substitute k = 8 + t 4')
=
hhl
Similarly, for n = 2r + 1, m
28
+ 1 both odd,
(by (6.2.7))
.
180 5.
CHAPTER 6
If n = 2r is even and m = 2s + 1 is odd, then
-&
' k
(substitute k = s
6.
-.
+ t - C)
This establishes
for all rn, n. Consequently,
By "telescoping," we have J
7. The same estimates as in points 3 and 4 of the proof of Proposition 5.3.1 show that, for fixed continuous and compactly supported f, ((f , # J , ~ l2 ) It: if J is large enough, with t: arbitrarily small (J d+ pending on f and c). Similarly, the estimate in point 3 of the proof of Proposition 5.3.2 leads to
xk
6
with IR1 5 c if J is sufficiently large. Since is continuous tit = 0, and b(0) = (2?).)-1/2,the fimt term in thb right-hand side of (6.2.11) converges to If({)12 for J-oo (by dominated convergence: ]&{)I 5 ( 2 ~ ) - ' I 2 for all <,because 15 1 by (6.1.1)). C o m b i i all this with (6.2.10), we have !(it d'j,k)12 = !If112 j,k&
for all compactly supported CODfunctions f. Since these form a dense set in LZ,the result extends to all of L2(R)by the standard density argument. s Without any extra conditions on mo, we therefore already have a tight frame with fiame wmtent 1. By Propaition 3.2.1, this frame is an orthonormal h i s
COMPA&I'LY
181
SUPPORTED WAVELETS
if and only if I($i(= 1 (use that tftkj,kU = lI+11 for all j,k E Z), or equivalently, if @(x) Q(z- k) = i5k,o for all k E 2.' This in turn is equivalent to & id(€+ 2741' = (2x1-l. Using 14(t)1=1mo(t/2+n)l16(2)I (a consequence of (6.2.6)),this can be rewritten as
I&
'
S
+
with a(() = 2%C, I$(< 2rt)12. This is equivalent to
"E" rk
,
c::,
We have %(<) = h,, e-$*, with h ~#, 0 # h ~ =so, that lmo(<)12is a 2plynomia in urs C of degree N2- Nl.On the other hand, a(<)= a t e-"<, with at = ( 2 ~ ) - ' J & e"€ ($(€)/' = (2%)-'f & 4(z) &(z - I ) = 0 if L 1 N2 Nl, since support & c [NL,Nzj. Consequently, a*(<)- 1 is a poly1! nomial in cos C of degree N2 - Nl - 1. Howwer,by (6.2.13), a(()- 1 is zero '-. whenever ( n ~ ( ( ; )is( ~(Im(C) l2 and !%(C x)12 have no zera in common), so that this polpomial has s t least N2- Nl aenrs (wunting multiplicity). Since it L of degree N2- Nl - 1, it therefore has to vanish identically, i-e,, a(C)= 1, or -.-. +2zt)l2 = (ax)-'. This is another nsy to derive that (6.2.5) is necmmy ~~:m sufficient d for the $I~,~ to constitute an & w ) d basis. In the nonorthonormal example we ss(r above, with &(c) = i ( 1 e - 3 g ) , the recipe (6.2.6) for leads t o
zt
-
+
.zcl&c
*
+
~
h this case $ is indeed not nomdizzd, (I+(! = 3-'1'. If we define 4 = (J~tl-'$, then the $,,r are normaiized, and comtit\ae a tight frame with fiame mwtant 3: the "redundgncy factorn of the frame is 3. This is not so surprising once qne rrediees thbt the family (4jL)j,kEZ can be considered as the union of t h e e shifted copies of a ustretchedn Haar basis: I
3-
3
5-b
=
(4,,3*+1)(~) = 3 D I
+
2 =
03
w,
(& +,,k5) (x - 1/31 , (D3 $?) (2 - 2/31 3
#
with ,(D3f )(XI = 31/2f (32). But let us return to the condition (6.2.51,
$r it8
equivalent
(6.2.14)
182
CHAPTER 6
Several strategies have been developed, corresponding to conditions on m, to ensure that (6.2.5) or (6.2.14) hold. Most of these strategies involve proving that the truncated functions jk introduced in the proof of Lemma 6.2.1 (or some other truncated family) converge to not only pointwise, hut also in L'(R). Since it is not hard to show that, for every fixed k, the {fk(- n); n E Z)are orthonorma), this L2-convergence then automatically implies (6.2.14). Conditions on mo sufficient to easure this L2-convergenceare, e.g.,
4
inf
}m&)
1 >0
-
(Mailat (1989))
&1<*/3
. \
Neither of these conditions is necessuy' but both cover many interesting examples. Better bounds in ICl than (6.2-16) lesd ta regularity for @ and $; ws will come back to this in Chapter 7. Subtquently, newsmy and sufficientconditions on were found. We discuss these at length in the next section.
6.3. Necessary and sufiicient conditions for ortbonormality. Cohen (1990) identified a first necmary and sufictent condition on m)ensuring L2-con~ergence of the fk. Cohen's condition involves the structure of the zero-get of mo. Before starting his resilt, it is conwmbt to introduce a new &incept. DEFINITION. A compact set K is &led congncent to [-n,n] modulo 211 if
<
2. For at1 an [-a, r],them e ; . d r t s _ t ~ Z so t h a t ( + 2 t ? r ~K.
Typically, such a compact set K conwent to 1-r, u] can be viewed t& the result of some "cut-and-pastework" on 1-?r, u]- An example is given in Figure 6.1. We are now ready to state and pnwe Cohen's theorem. THEOREM 6.3.1. (Cohen (1990)) Assume that w a fritripondric pol& nomid satisfk,ng (6.1.1), with -(O) = t, and define 4 as in (6.2.2). Then the frdbwing am equivdent:
P,,Them ezish a compact set K congruent to [- r , r] madub On and wntarnmng a nesghborhood o/O so that inf inf
k>O eEK
{mo(2-kt)l > 0 .
I I
1
COMPACTLY SUPPOKFED WAVELETS
nc. e l x = f
set aonqrrcrnt to 1-T,
ii
b 4 ~ '
[ - ? f , - p w ] u [ - , - i ][ ~- f , P ] u [ 4 , ~ ] u [ ~ , U' Df -]W ~
s]modulo 21; at mn be v#cad as the rvuU of cuttang [ - r / 2 , - r / 4 ] and 15n/8,3n/4)a13 of I-*, sj and movnyf tht fint pia*r to ihe nght by 2m, the second b the Ieff
F
P k
r
REMARK. The condition (6.3.2) may seem a bit technfcal, and hard to verify in practice. Remember baweYer tbat K is compact, and is therefore bounded: K c [-R, R]. By the continuity of q and ~ ( 0 =) 1, it follows that lmo(2-k()l > un~formlyfor dl KJ 5 R, if k is larger than some ko. This means that (6.3 2) reduces to requiring that the ko functions ~ ( ( / 2 ) , ~ ( ( / 4 ) , - ., m0(2-*0[) have no, zero on K ,M equivalently, that m has no zero In K / 2 , K/4, - - , 2-&0K This is akeady much mote accessible! o
4,
Proof of lneorrem 6.3.1
I
proving (1) ~ = (2). t 1. We start Aaeurne that (6.3 1) holds, a equiudatly, &I&< Then, for ail E 1-n, n],there exists E N eo that
<
L
-.
7
E
k
c- . L
4' ?!
f? kP
+ 2nl)I2 = ( 2 r ) - ' .
z,r,
Since )is continuous, the finite sum id(*+ 2mL)12 is amtirmaus as well. Therefore there exists, for every in (-n, 4, a neighborhood (C; I< - €1 < Rt) so that, for all C in t b neighborhood,
2
a Gnite sub& of the coIlection of Since (-R, ?r] ie oompact, there intervals ((;I( -
C lkc + 2rl)i2 2
(&)-I
.
(6.3.3)
ItlSQ
t
2. I t W t h a t f o r ~ e r y ~ ~ [ - ~ , ~ , h e o t i s t s t ~ - & ~ d t ~ ~ that I&{+2 4 1 2 [8r(2& -t l)]-'P = C. Define now sets St, -4) < P 5 %
~
CHAPTER 6
for e # 0,
The St, - l o
5
e
t o form
I&O)1 '= ( 2 ~ ) - ' f 2> C , and since
4
a partition of I-r, KJ. Since is continuous, So contains a neigh-
borhood of 0. Now define
Then h' is dearly compact and congruent to I-n, a] modulo 213. By construction, I&(<)( > C on K ,and K contains a neighborhood of 0.
3. Next we show that K satisfies (6.3.2).As pointed out in the remark before the proof, we only need, to check that inftEK ( m o ( 2 - k ( ) l 0 for a finite number of k, 1 5 k 5 ko. For [ E K, we have that
is bounded below away from zero. Since 141 is also bounded, the first factor in the right-hand side of (6.3.4) has therefore no zeros on the compact set K . As a finite product of continuous functions it is itself continuous, so that
Since
5 1, we therefore have, for any k, 1 5 k 5 h,
This proves that (6.3.2) is satisfied, and biihes the proof (1) =+ (2). 4. We now prove the converse, (2)
Define pk(t) =
(2~)-I/2
* (1).
m0(2-j~)]' - X K ( ~ - ~ <whew ), XK is the
indicator function of K, XK(~)= 1 if ( E K, 0 otherwise. Since K contains a neighborhood of 0, pointwise for k-+m.
pk34
/
185
COMPACTLY SUPPORIBD WAVELETS
>
2 4
>
<
5. By assumption, I ~ ( 2 ' ~ , $ ) 1 C > O for k 1 and E K. On the other hand, we a h ha&, for any t , ]nso(C) mo(O)l C'IJI; hence Imo(E)I 2 1 - FIFI. Since K is bounded we can find4 so that ~-*c'IcI < ir( E K and k 2 b. Using 1 - x 2 e-2z Tor 0 5 x _< we find therefore, for
< E K,
-
< i,
4
We can rephrase this as
This implies
We can therefore apply the dominated convergence theorem and conclude
that j ~ ~ +in4 La.
6. The congruence of K with I-'17, A] aiSdu1o 27r means that for any 2 * - ~ e ~ d ifunction c f JtEX @ f (0= & f (C) = :J 4 f (F). 1"
particular,
=III
CHAPTER 6
. Since this implies J
Ipk(<)12e-'*
=
for all k. H e w
(because prc-+J pointwise, together with the dominance (6.3.5))
which is equivalent with (6.2.5) and therefore with (6.3.1). REMARK The "truncated" functions pk are not the same as the fk introduced in the proof of Lemma 6.2.1, but the following argument shows that L2-convergence of the pk implies L2-convergence of the fk. First of dl, K contains a neighborhood of 0, K > [-a,a] for some a, 0 <' a < n. Defille uk = (2n)-'I2 mo(2-'€1 X I - a , m l ( 2 - k ~ ) .Since XI-^,^^ .< X K , the same dominated convergence argument as for the pk applies, and vk+& in L2. Consequently, /(pk- ~ ~ 1 1 ~ 1fOr3 0~--+oo. Using the c o q p a x of K to [ - T , n] moddo 27r, one shows that (IpL- ~ k l l ~=a I(fk y l j L a - Hence I I -~illLa s llfk V ~ I I L Z 1luk - )NLZ-+O for LOO. Note that if Mdat's condition (6.2.15) ii satisfied,then we can simply take K = [ - T , n]; Cohen's condition is then trivially satisfied, and the are indeed orthonormal. The following C O gives ~another example of how to apply Cohen's condition. COROLLARY 6.3.2. (Cohen (1990)) Assme that r r ts ~ a trigonometric polynomial satzafpng (6.1.1), with ~ ( 0=)1, and dejine 4 as is (6.2.2). lf Q h no terns in [-n/3, 7r/3], then & A,, am orthonormal. P m f . We need only to construct a satisfactory compact eet X. Sine rca may have zeros in r / 3 < (€1 5 x/2, K = [-.rr, .rr] is no longer a g o d choice. But we start with thie choice as an anaatz, and "cut out" the raro8. More precisely, msume,that t h e r s a , o f m o h r / 3 < 5 n/2 are(: (They are nemsady finite in number, since ~ I O is a trigonometric polynomial.) Similarly, we have [t- 5 .. 5 ti for the zera d in - x / 2 I6 < -n/3. For every t &&we to be the bbmection with [-T, r] of a smaU open interval
n:=*
-
1
-
+
&,
<
<
--•
I
sc+.
I:
around$,dem~sotbat~~~~~wrb~aitherhotherorxitb I-r/3, */a] and so that JniojJ?< f . (If ',?= ir/2, ibm I; w i l l bc of the
COMPACTLY SUPPORTED WAVELETS
A very different approach to the derivation of conditions on
189 that ensure
(6.2.5) was initiated by Lawton (1990). Let us assume that n~ is of the form
i,e., h, = 0 for n < 0 or TL > N; m5 can always be brought into this forrn by multiplication by esNl<,corresponding to a shift of tb, by N1.Define ac = 1 N , and we J & b ( x ) #(x - Q). Since support($) C 10, N],at = 0 if < N into a (2N - 1)-dimensional vector can regroup the non-trivial at, ( ~ - N + I, - -,a&,' . . -, a ~ - l ) . Because $(x) = En hn 9(2r - n ) , the at satisfy
If follows that if we define the (2N - 1) x (2N - I ) matrix A by
where implicitly h, = 0 if m < 0 or m > N ,then
is
i.e., a is an eigenvector of A with eigenvalue 1. Note that 1 always an eigenvalue of A: if we defjne 0 by B = (0, - -.,0,1,0,. .- ,0) (1 in the central position), or Pc = 6c.0, then
by (6.2.71, i-e., AP = ,l? If the . eigenvalue 1 of A is nondegenerate, then a necessarily has to be a multiple of P, he., J & +(x) #(x - L) = ~ 4for,some ~ 7 E C. This implies I&(< 2nk)J2 = (27r)-'7; since 14(2nk)( = 0 for k # 0 (see the start of 56.2) a d j(0) = (2~)-')-l/~ by definition, it follows that 7 1, m that $ & +(x) +(x - Q)= 6[,0. We have thus a very simple sufficient criterion for orthonormality of the h,". THWREM 6.3.4. (Lawton (1990)) Assume that m~ is o trigonometric polynomial of the fonn f6.3.7), satismg (6.1.1) and m ( 0 ) = 1; define 4 as sn (6.2.2). If the eigenvalzre 1 of the (2N - 1) x (2N - I) m a t e A defined by (6.3.9) u nondagenemte, then the &,,_am orthonormal.
C,
+
-
I
CHAPTER 6 I
Orthonormality of the &,n can only fail if the characteristic equation for A has a multiple zero at 1. This indicates that among all the possible choicee for the h,n = 0,. . ,N (keep N fixed), the "badn choices (leading to nonorthonormal &,n) constitute a very "thin" set. (This statement is made more precise in Lawton (1990).) For N = 3, for instance, the only non-orthonormal choice (up to an overall phase factor) is ho = h3 = 1/2, hl = h2 = 0. Lawton's condition can be recast in tenns of trigolnometric polynomials. Define, as before, Ma(€) = Im(<)la, and d e h e the following operator Po,acting on 2~-periodicb c t i o n s f .
Clearly the constant polynomial 1 is invariant under Po by (6,l.l). everything out -in terms of Fourier coefficients, we have
(PO,), =
hn
k,n
h;~;
(
=
I ?
'
.
~h-y+m) h fm .
m
This is essentially the same expression as (6.3.8)! (We have not assumed that ,f = 0 for Iml > N, however, so it is not quite the same,) It bllowe that Lawton's condition is watisfied if we know that the only trigonometric polynornids invariant under Po are the constants. A priori it is not clear whether Lawton's condition is sufficient or not: it is conceivable that A has an eigenvector different from 0 with eigenvalue 1, but that a nemheless happens to be equal to B. However, in the spring of 1990 both Cohen end Lawton proved, indepsndently, that their two &ions are equivalent (a generabationagpemin Cobear, Daubechies, aad Pbauwu (1992) ae Theorem 4.3; see aleo Lawtan (1991)), implying the d c i e n c y of Lawton's condition. , THEOW6.3.5. Assunre that% is a trigonometniEpdynomicrl& that, (6.1.1) is satisfied and m&O) = 1. If there ezisb a cmnpad set K congruent to [-x, ~ rmoddo ] 2?r, cmtuining a neigbhood of 0, such that infklr ~ I I & ~ K lmo[2-"0l> 0,then &, only trigonmet& ~LyMmralrinw&&under A 8n the camstants.
I
I
1 4
[i !1
191
COMPACTLY SUPPOKIED WAVELETS
REMARK.This is sufficient to prow equivakmce. If we denote Lawton's original condition by (L), Cohen's condition by (C), Lawton's condition repbraad in $erms of Pi by (P)and the orthonormality of the&,, by (0),then we shady know (P) (A) (0) The implication (C) + (P)suffice8 to prim equivalence of a11 four conditions.
*
*
t
*
3
1. We will prove that the existence ofa noncomtant trigonometric polynomial f invariant for Po contradicts the existence of a compact set K with all the desired properties. Supf is such a oonconstant trigonometric polynomial, invariant for PO.Define f l ( 4 ) = f (<) - mint f (C), f2(t)= -f (6) + m a q f (0. Since f is not constant, at least one of fi, fi harc f3(0) # 0. Pick j so that f,(O) # 0, and define fo = f,. Then fo is nonnegative, fo(0)# 0, fa has a t least one zero, a d fo is invariant under Po.
I
i
i-
L
i li
2. Next we explore the zero set of fa, which turns out to have a very particular structure. If = 0 for 0 # ( E [O,2x[, then
fo(c)
Here Mo, fo are both nonnegative, and Mo(U2), Mo(2+7r) cannot vanish simultaneously, by (6.1.1). Therefore, either fo(Cj2) or fo(J/2 n) = 0. It follows that if wttpick one zero O # 6%E [0,2xf of fo, then we can aspciate to it a chain of zeros in [0,2r[, 6, - ..,&, - .. , with the property that t3+1equals either or $: n, or, equivalently, G = r<,+I,where T is the transformation 6 I-+ 2 ( mod (27r), which maps [O, 2 4 into itself. Being a trigonometric polynomial, fa ha8 only finitely many zeros, so that this chain cannot go on ad infiniturn. Note that the chain has at least two = would imply (1 = 0. Let r be the first index for dements, since which recurrence occurs, i.e., tr = & for some k < r. Then necessarily k = 1, because JE > 1 would lead to 41 = T ~ ' ~-
+
5:
k
Ii 6:
+
+
. c,-
this cycle. 3. If this cycle of zero8 does not exhaust the set of mos difkent from 0, then we can find 0 # C1 j = l , - . . , r - 1, for which fo((;) = 0. This can again be taken as a seed for a chain of zeros, & , G , . - - , G , . . .. Every element of this new chsin is neceesarily difierent from aU the 0, shyj~ = & would imply = ?'-I& = T'-~(,, i.e., would equal anw
+ c,,
'@ ' b
@
ct
&.By~he~unsupunaats.Ibova,C-Wthd~reafyckd-
i
CHAPTER 6
-192
for j , invariant under T , and disjoint from the first cycle. We can keep on constructing such cycles until exhaustion of the finite set of zeros of fo. The zero set of jo therefore oonsists of a union of finite invariant cycles for T . \
4. Now note that if fo(() = 0, then necessarily fof< -!- r ) # 0. indeed, since T< = T ( < n ) , both and C u would belong to the same cycle of zeros if fo(<) = 0 = fo(< x ) . If this cycle has length n, then it would follow that = rn<= = T"-'T(( n ) = ( ?r, which is impossible.
+
+
<
<
+ +
+
5. Finally, we remark that if fo(<) = 0 , then ' M ~ (+ C n) = 0. Indeed, for any [ so that jo(<)= 0, rf is also a zero for j o , and it follows that
+
Since fo(C) = 0 and f o ( t w ) # 0, this inlplies Mo(t + 17) = 0; hence mo(4 .r;) = 0. Therefore, the existence of jo implies the existence of a j = l , . . . , n - 1, cyclic set J l , . . . , < , , for T , with<] = = T < , , so that rno(<, A ) = 0 for all j. Since f o ( 0 ) # 0,we have # 0.
+
<,
+
6. We now show how these zeros t j + + for m0 are iacompatible with the = existence of K. Since T&+, = e l , T&, = t I , and, in particular, r n & , we have f, = 27rx1, where the x, c [O, I[ have the following binary representations: /
xi t2
= .dld2.-.tindl- - - d n d l . - . d n . - = .da-.-&dl*--dndl...d,.--
(d, =Oor 1)
G # 0, not all the dj are zero. Let us, for this point only, define -dSince = 1 - d for d = 0 or 1. Then <, + + = 27ry1 modulo 27r, with y, given by
We have q(27tyj) = 0, j = 1, -.- ,n. Suppose a compact set K existed with all the desired properties. Then there would be an integer t?, with a binary expansion with at moet a certain preassigned number L of digits (L depends only on the ~izeof K), so that 2 x 9 = 2 n ( 2 y l + l )has the property that r ~ t ~ ( 2 n 2 ' ~#y )0 for all k 10. We have
with ej = 1or 0 for j = 1 ,
..,L. We can also rewrite thia aa
193
COMPACTLY SUPPOKf'ED WAVELETS
t-
where el = 1 or 0 for j = I , . . . , L and e, = 0 if -J > L. The 2-ky are obtained by shifting the decimal point to the left. Since nq, is 2~-periodic, only the "tail," i.e., the part of the expansion of 2-'y to the right of the decimal point, decides whether ~ q , ( 2 r 2 - ' ~ )vanishes or not. If el = d l , then y/2 would have the same decimd part as 91, hence mo(27ry/2) = 0 would follow. Since mo(2-lry/2) # 0, we therefore have e 1 = d l . Similarly, we conclude ez = d,, ea = dn-1, etc It follows that e ~ + l ,. . ,e ~ + , are a h successively equal to dk, dk- 1, - . . ,dl, d,, . ,dk+ 1 for some k E {1,2,. .- ,n). Since the dl are not dl equal to 0, whereas e ~ =+ . .~- = e t + , = 0,this is a contradictiofi:'This finishes the proof.
t
:
With Theorem 6.3.5 we end orir discuosion of necessary and sufficient conditions on mo. The following theorem summarizes the main results of 586.2 and 6.3. THEOREM 6.3.6. Suppose mo rs a tngonomehc polynomial such that Imo(()12 Im([ + x)l2 = 1 and mo(0) = 1. Define 4 , @ by
+
i-
n
r
00
kt)
= (2~1-l'~
mo(2-'0 ,
Then 4, 11, are compactly supported Lz-functions, sattsfyng,
&
where h. u determined by ma via ~ ( t=) C, h, e-'"€. M o w e r , the T,~],~(x) = 2-]Iz $(2-f x - k), j , k E Z mnstitute a fight frame for L2(lR) with frame constant 1. This tight frame as an o r t b a d bcrsis if and only ifm-,. : satisfies one of the follovmzg equivalent conditions: e
There exists a compact set K, congruent to a neighborhood of 0, so that
[-T,a] modulo
ipf inf (rn.0(2-~.91>0
k>O (EK
k*
r
27r, containing
.
Them exists no nontrivial cycle {G, .t,,) in [0,2-/rl, invariant under [ w 2 ( modvlo27r, sueh t h a t r q ) ( ~ j + r ) = O f o r a l l j =1 , - - e n . The eigenvalue 1 of the [2(N2- Nl) - I] x mapix A &fined by
(where we assume h, = 0 for n < Nl, n
7:
[?(Nz - Nl)- 11-dimemional
> N2) j9 nondegenemte.
184
CHAPTER 6
From the point of view of a u b d filtering, this theorem tells us that, provided the high-pass filter has a null at DC (mo(lr)= 0, hence mo(O')= 1 with the appropriate phase choice), we &almostalways" have a corresponding orthonorma1 wavelet basis. The cormqmndence only fails "accidentally," as is illustrated by the last two equivalent necessary and sufficient conditions. In practice, one likes to work with filter pairs in which the low-pasa f3ter has no zeros in the band )(I 5 x / 2 , which is sufficient to ensure that the $&, are an orthonormal basis. But it is time to look at some examples! 6.4.
Examples of compactly supported wavelets generating an orthonormal basis.
All the examples we give in this section are obtained by apectrai factorization of (6.1.11), with different choices of N and R. Except for the Haar basis, we have no closed-form formula for t$(x), $(z);we- -will h-the next section how - - explain --the-plotg f o m . A f&t fmilly of examples, coastructed in Daubechies (1988b),corresponds to R -z 0 in (6.1 11). En the spectral factorization needed to extract C(<) from L(<) = pN (sin2 (/2), we retain systematically the zeros within the unit circle. For each N, the corresponding N- has 2N non-vanishing coefficients; we can Q that choose the phase of N ~ so
--/-----.
L
, N = 2 through 10. For fester implementation, Table 6.1 lists the ~ h for it makes sense to keep the factorization (6.1.10) explicit: the filter C is much shorter than m~(N t a p instead of 2N),and the filters are very easy to implement. Table 6.2 lists the coefficients of C(<), for N = 2 to 10. Figure 6.3 shows the plots of the correspondmg ~ 4 N+ , for N = 2,3,3,7, and 9. Both p,t#~ and Nt,$ have support width 2N - 1; their regulerity clearly increases with 6CpN, N. In fact, one can prove [see Chapter 7) that for w e N &, with C( e .2. Retaining systematicdIy the zeros within the unit circle in the spectrd facamounte to b i n g the 'minimum phase filter" nq,among torization all the possibilities once lmo12is fixed. Thiscorresponde to a very marked asymmetry in t$ and 9, as illustrated by Figure 6.3. Other choices may lead to i e s ~asymmetric 4,,)t although, as we will see in detail in Chapter 8, complete symmetry for 4, Q can not be achieved (except by the H w basis) within the hamework of compactly supported orthonormal wavelet bases. Table 6.3 lists the h,, for the "least asymmetricn 4, for N = 3 through 9, corresponding to $he same I;no12 as in Table 6.1, with a different "square root" m. We will come back in Chapter 8 on how this "lea& asymmetric square root" is determined. Figure 6.4 shows the corresponding Q and 9 functions. Figure 6.5 shows plots of I q t as functions of &, for the above examples, for N = 2, 6, and 10. These plota show that the subband filters for these
+,
r I
k
COMPACTLY SUPPORTED WAVELETS
2 %jikr ~ m&cunb ~ h (low-poaa n liUer) for the ~ p r ~ c ~ f l wavcletr unth e z b m d +sc and haghd number of tmnwhsng mommtr compuhbk! waih thar w p m wuith. The me & ~ c d 80 w ~ h= n
En
#
CHAPTER 6
TABLE 6.2
The coeficients
In
of fi C ( € )=
EnC,e-*4,
for N = 2 to 10. Normdirotwn:
EnCn = fi.
C O M P A ~ SUPPOFtTED Y WAVELETS
FIG.6.3. PYdr of the radag funduma ~4 tnul wavektu ~9 for the compcUt( wpported mmdttr tu& mccdmum number of wnhhing momenta for heir mpport width, and with the ezmmd phcrse choke, for N = 2,3,5,7, a d 9.
The low-pao filter cocficients for the " k t qpnmhicDmmpoetly nrppodd wovekb with N = 4 to 10. Luted hem a the CN,, = cN,n = 2. f i ~ . n ;a
maxtrnum number of vanishing momcntr, for
x,,
COMPACTLY SUPPOKTED WAVELETS
I
$ 8 -
g
FIO. 6.4. Plob of the rcdtng fincPian 4 and the wcrtrckt twmplly ruppwltd wa& and 10.
+
jar the "luwt -me&" with mazhwn number e j vanuhing momenh, for N = 4,6,8,
+(o)
COMPACTLY S U p p O m D WAVELETS
20 1
orthonormal bases are indeed very fiat at 0 and r , but very uround" in the transition region, near %/2. The filters can be made "steeper" in this transition region by a judicious choice of R in (6.1.1). Figure 6.6 shows the plot of lmol corresponding to N = 2 and R of degree 3 chosen such that I T T Q ( [ )has ~ ~ a zero * at t = 7n/9 (= 140°). This is much closer to a "realistic" subband coding filter. The corresponding "least asymmetricn function q5 is shown in Figure 6.7; it is less smooth than 4 4 (which has the same support width, but corresponds to N = 4 and R 3 O), but turna out to be smoother than 2r$ (for which has a zero of the same multiplicity, i.e., 2, at = T ) . In Chapter 7 we will come back 2 in greater detail to these regularity and flatness issues. The h, corresponding to .= Figure 6.7 are listed in Table 6.4., '
z ,z
$
f
3 S L
Rc. 6.7. The " h t aqmmetric" ading jiuaction 4 and wauekt w plotted in pioUn 6.6. ?g
+ cormqwd'ns to 1-1
All these examples correspond to real h,,, 4 and $, i-e., to I)( and 161symmet
,
k. ric arouad t = 0. It is also posaible t o construct (complex) examples with 161, ~ $ 1 @; p? t, fi
k
'
<
bncentrated much more on > 0 than on 4 < 0. Ttrlce for instance the m~ of the psevi~uaexaxn~~e, ahichsatisfea%(fq) = l,anddefioet~$([) = This obviously aatis6eS(6.l.l), since m~ d m , snd = 1. We can there fore construct tj# (4)= 4 ( 2 - j ( ) , @(() = c-*I2 & ( ~ 2 T ) @((/2);
4
nE1
+
m([-F).
202
CHAPTER 6
The coefictents for the low-pass filter r m s p o n d i n g to the sding.junctron in Figure 6.7.
these are compactly sap~~orterl L'-futlctions, and the @ f k , j, k E Z constitute a tight frarrie for iJ2(W), by Proposition 6.2.3. Moreover, since the only zeros of mu on 1-s,?rj are in = &%, * K , it follows that rn,#(<) = 0 only for = or - $. Coss 0 for 5 3 , and the ?+b:k rolistib~~tc an orttlonor~nalwavcllc~t.basis, by Corollary 6.3.2. Figure 6.8 plots [,no#(<)I, @(()I and l$#(()~;it is clear that J,-" 4 16#(<)l2is rnuch'larger than J!_ @ I$#( < ) I ' Note that the iicgativr frcyaency part of is much c l o s ~ rto thc origir~than thc ptmitive frcqi~encypart, as required by the liccessary condition @ /(I' I ) ' = 4 IS#(t)l2(n93.4). The existence of such ~ i in Cohen (1990); ill fact, for any c > 0 one "asymmetric." was first p o i n t ~ out can h i d an ortlionornial wavrlet basis such that 4 < f.
<
< &.
>
4#
s r
4
,:s
'
1G(()lZ
6.5.
The cascade algorithm: The link with subdivision or refinement schemes.
It can already be suspected frotn tlie figrlrcs in 36.4 that there is no closedforrrt analytic formula for the compactly supported (b(x),$(x) constructed here (except for the Haar case). Neverthclcss, we can, if 4 is continuous, compute $(x) with arbitrarily high precision for any given x; we also haye a fast algorithm to compute tlie plot of 4.' Let us see how this works. First of all, since 4 has compact support, and 4 E L1(R)with & #(x) = 1, we have PROPOSITION 6.5.1. I f f is a wntinu0u.g function on R, then, for all z E W,
5
I f f is uniformly continuow, then this pointwise convergence is uniform as well. If 'f is Holder continuous with exponent a ,
COMPACTLY SUPPORTED WAVELETS
T
-
'
I
'
'
I
FIG 6 8 i'lob of jmJ,(91, and 1$1 for an orthonormal wavelet bhvu whrw 1/, trnted m o w on posttrve than on negatloe frequencies
i
e
t h e n the convergence zs eqonentzally fast tn J
19
corlcrn
'
Proof. All the assertions follow from the fact that 2 3 4 4 2 5 ) is an "approx~mate bfunctlon" as 3 tends to m More precisely,
CHAPTER 6
(where we suppcwe support
9 C 1-8R]) .
If f is continuous, then this can be made arbitrarily small by choahg j 'sufficiently iarge. If f is uniformly continuous, bhen the choice of j can be made independently of x, and the convergence io uniform. If f is Holder continuous, then (6.5.2) foliaws immediately as well.
.
h u m e now that 4 itself is continuous, or even Hgder conthfbous with exponent a. (We will see many techniques to compute the Hiilder erlponent of 4 in the next chapter.) Take x to be any dyadic rational, z = 2-J K. Then Proposition 6.5.1 tells us that
1
Moreover, for j larger.+ than some jo,
l+($-rK) - 2312 (4,
#-j,2,-~~)(
C2-ja
I
(6.5.3)
where C,jo ard dependent on J or K. If 2 J - J is ~ integer, which is-automatically true if j 2 J, then the inner products (4, ) - j , v - ~ K ) are easy to compute. Under the assumption that the. &,, are orthonormal (which can be checked with any of the necessary and sufficient conditions on m,-, listed in Theorem 6.3.5), 4 is , the unique function f characterized by
(f (
h,n)
f
=
bo,n
j = 0
1
for j > O , k E 2 .
I
(6.5.4) (6.5.5)
f
We can we this ae input for the recaostruction algorithm of the subband 6ltering d a t e d with m~ (see 55.6). More specifically, we start with a low pess sequence cfl = &, and a highpass sequence dO, = 0, and we "crank the machine" to obtain
We then use d;;' = 0, to obtain, after another cranking, *
n I
etc. At every atage, the czf are equd to (4, 4-j,n). w e t h e r with (6.5.3), this means that we. have an algorithm with exponentially fast convergence to
I
compute the values of t$ at dyadic ratioads. W e can interpolate these k d ues and thus obtain a sequence of f p n c t i 5 approximating 4.9 We can, for instance, define $(z) to be the function, pi& constant on the intends 12-i(n- 1/21, 2-j(n 1/2)[, n E 2, such thbt #(2-jk) = 2jI2 (4, 4- j , k ) . Another possible choice is $ (z),p i e a r k linear ~n the [2-jn, 2-3 (n 1)]t a E it, so that 9;(2-Jk) = 2i/2{41 qLlSk). For both choices we have the following propodtion. PROPOSITION 6.5.2. If 4 & H oemtinwnu with ezponent a, then thew &hi:> 0 and jo E N so that?for j l y o ,
+
+
PmJ Take any z E R. For any j , choose n so that 2-in 5 z < 2-J(n I). By the definition of t$, $(z) is mmmarily a convex linear combination of 2112(4,9- ,,=) and 2iI2(+, 4- j,n+l), whether e = 0 or 1. On the other hand, if j is larger than some jo,
+
the same is true if we replace n by n + 1. It fdlm that e similar estimate holds for any convex combination, or Iq5(x) - $(%)I 5 C 2-la- Here C can be chosen independently of z,so that (6.5.8) fo110ws. This then is our fast aigorithm to compute approximate values of #(x) with arbitrariFy high precision:
PI Start with the sequence . - 0 - - .010- .- 0 - ., representing the r$(n), n G Z.
5F
F
b
2. Compute the qj (2-In),n E 2, by the machinen as in (6.5.7). At every step of this cascade,twice a s - n ~ values ~ ~ y are computed: values at "even pointsn 2-j(2k) are refined from the precious step,
-
and values at the "odd points" 2-j(2k
$(2-j(2k
+ 1))
+1) are domputed for the first*time,
h(k-wl
2
qJ-1
.
C
b
.
.
(6.5.10) : *
Both (6.5.9) and (6.5.10) can be viewed as convolutions.
3. Interpolate the G(2-jn) (piemwise constant if e = 0, piecewise linear if c = 1) to obtain r$(x) for non-dyadic z.
The whole algorithm was called the cascade algorithm in Daubechies and Lag&ias (1991), where 6 = 1 was chosen; in Daubechies (1988b), the choice c = 0 was made.1° All the plots of 4, $J in s6.4 and in later chapters are, in fact, plots of q j , with j = 7 or 8; at the resolution of these figures, the difference between r#~ and these T$ is imperceptible. A particularly attractive feature of the cascade algorithm is that it allows one to "zoom in" on particular features of 4. Suppose we already have c~mputedall the ~ g ( 2 - ~ n but ) , we would like to look at a blowup, with much better resolution, of 4 in the interval centered around 1. We could do this by computing all the ~ 7 ; ( 2 - ~for n ) very large J, and then plotting q;(x) only on the small interval of interest, corresponding to 25-4 . 15 5 n 5 2J-4 . 17 But we do not need to: by the "localn nature of (6.5.9), (6.5.10) much fewer computations suffice. Suppose h, = 0 for n < 0, n > 3. The computation of 7);(2TJn) only involves those rl;-, (~-~+'k)for which (n - 3)/2 5 k 5 n/2. Computation of these, jn tun, involves only the 77;-2(2-J+21) with (k - 3)/2 5 4' 5 k/2, or n/4 - 3/2 - 3/4 4' n/4. Working back to J = J - 4, we see that to compute 7$, on we only need the q ; ( Y 5 r n ) for 28 5 m < 34. We can therefore start the cascade from . . O . . .010 . . . 0 . . ., go five steps, select the seven values ~&(2-'rn), 28 5 m 5 34, use only these as the input for a new cascade, with four steps, and end up with a graph of q; on For larger blowups on even smaller intervals, we simply repeat the process; the blowup graphs in Chapter 7 have all been computed in this way. The arguments leading to the cascade algorithm have implicitly used the orthonormality of the @,,k, or equivalently (see $6.2, 6.3), of the &,,: we have characterized 4 as the unique function f satisfying (6.5.4), (6.5.5). The cascade algorithm can also be viewed differently, without emphasizing orthonormality at all, as a special case of a stationary subdivision or refinement scheme. Refinement schemes are used in computer graphics to design smooth curves or surfaces going through or passing near a discrete, often rather sparse, set of points. An excellent review is Cavaretba, Dahmen, and Micche!li (1991).'We will restrict ourselves, in this short discussion, to ohe-dimensional subdivision schemes.12 Suppose that we want a curve y = f (x)taking on the preassigned values f (n) = j,,. One possibility is simply to construct the piecewise linear graph through the points (n, f,); this graph has the peculiarity that, for all n,
12, 5 )
-
[s,gj< <
[E,31.
"
.
which gives a quick way to compute f a t half-integer points. The values of f at quarter-integer points can be computed similarly,
and so on for 2/4 + Z/8, etc. This provides a fast recursive algorithm for the computation of f at all dyadic rationals. If we choose to have a smoother spline interpolation than by piecewise linear splines (quadratic, cubic or even higher
,
207
COMPACTLY SUPPORTED WAVELETS
.,a
s#$
order splines), then the formulas analogous to (6.5.9), (6.5. lo), computing the f(2-Jn + 2-I-') from the f (2-Jk), would contain an infinite number of terms. It is possible t o opt for smoother than linear sprine approximation, with interpolation formulas of the type f(2-In
+ 2-I-')
= xakj(2-'(n
- k))
.
(6.5.13)
k
with only finitely many a k are nonzero; the resulting curves are no longer splines. An example is f ( 2 - J n +2-]-')
=
1 16 9 [f (2-In) 16
1)) + f(2-](n
+ 2))
-- [f(2-'(n
-
+-
+ f (2-3(n + 1))) .
(6.5.14)
This example was studied in detaiI in Dubuc (1986), Dyn, Gregory, and Levin (1987), and generalized in, for example, Deslauricrs and Dubuc (1989) and Dyn and L p i n (1989); it leads t o an almost C2-function f . {For details on methods to deterrrii~lethe regularity of j , .see Chapter 7.) Formula (6.5.14)describes an intcrpolat~onrefinement scheme, in which, a t every stage of the computation, the values computed earlier remain untouched, and only values a t intermediate points need to be computed. One can also consider schemes where a t every stage the values computed a t the previous stage are further "refined," corresponding t o a more general refinement scheme of t h e type fI+l(2-'-ln)
=
C
w n - 2 ~f1(2-'k)
.
(6.5.15)
k
Formula (6.5.15) corresponds in fact t o twoconvolution schemes (with two masks, in the terminology of the refinement literature), fJ+l(2-Jn) =
~ 2 ( n - k ) fJ(2-'k)
(6.5.16)
k
(the refinement of already computed values), and fJ+l(2-'n
+ 2-'-I)
~ 2 ( ~, I + I jj(2-Ik)
=1
(6.5.17)
k
(computation of values a t new intermediate points). In a sensible refinement scheme, the f, converge, as j tends t o oo, to a continuous (or smoother; see Chapter 7) function j,. Note that (6.5.15) defines the f, only on the discrete set 2-32. A precise statement of t h e "convergence" of j, to the continuous function f, is that -sm
m-+w
{
.-_ -- -
sup j20,k~Z
1ji12-m2-j*)
-
+
(6.5.18)
CHAPTER 6
208
where the superscript X indicates the initial data, f t ( n ) = A,. The refinement scheme is said to converge if (6.5.18) h& for all A E tm(Z); see Cavaretta, Dahmen, and Micchelli (1991). (It is also possible to rephrase (6.5.18) by first introducing continuous functions fj, interpolating the fj(2-jk); see below.) A general refinement scheme is an interpoiation scheme if mk = 6k,o, leading to fj+c(2-Jn) = fj(2-in). In both cases, general refinement scheme or more restrictive interpolation scheme, it is easy to see that the linearity of the procedure implies that the limit function f, (which we suppose c o n t i n u ~ u s is ~ ~given ) by
where F = F, is the "fundamental solution," obtained by the same refinement scheme from the initial data Fo(n) = bnp. This fundamental solution obeys a particular functional equation. To derive this equation, we first introduce functions fj(x) interpolating the discrete fj(2-ik): w
f3 (3)
=
(2-jk) u(2jx - k)
,
(6.5.20)
k
where w is a "reasonable" l4 function so that w(n) = bnS0. Two obvious choices w ( x ) = 1 for 5 x < 0 otherwise, or w(x) = 1 - 1x1 for 1x1 < 1, 0 otherwise. (These correspond to the two choicea in the exposition of the cascade algorithm above.) The convergence requirement (6.5.18) can then be rewritten as (1 - f&Il~-+0 for ,400. For the hndamental solution F,, we start from Fo(z)= w ( x ) . The next two appraxbating functions Fl,F2sat*
-;
are
1,
fix
Fl (x) =
FI( 4 2 ) w(2x - n)
x
(by (6.5.20)
n
=
urn u(2x-n)
(use (6.5.15) and Fo(n) = 6n,o)
n
This suggests that a aimilar formula should hold for all Fj, i.e.,
COMPACTLY SUPPORTED WAVELETS
Induction shows that this is indeed the case:
=
w,,-2k
tut
Fj-1(2-j+1~ - I ) w(2j+'z - n)
n,k,t
(by the induction hypothesis)
;>
Since F = F, = lim,,,
F, , (6.5.22) impties that the fundamental solution F
satisties
F(x) =
wk
F(2z - k)
.
(6.5.23)
k
It is now clear how our compactly supported settling functions 4 and the cascade algorithm fit into refinement schemes: on the one hand, 4 satisfies an equation of the type (6.5.23) (basically as a consequence of the multiresolution requirement Vo c V- and on the other hand the cascade algorithm corresponds exactly to (6.5.15), (6.5.20). Orthonormality in the underlying multiresolution framework made our life a little easier in the proof of Propxition 6.5.2, but similar results can be proved for refinement schemes, without orthonormality of the F(x - n). Some basic results for refinement schemes are: t
I
L
1b
If the refinement scheme (6.5.15) converges, then Enw2, = En Wzn+r = 1, and the associated functional equation (6.5.23) admits a unique continums solution of compact support (up to normalization). If (6.5.23) admits a continuous compactly supported solution F, and if the F ( z - n) are independent (i-e., the mapping P ( Z ) 3 A H &F(x - n) is one-to-one1'), then the subdivision algorithm converges.
C,
Micchelli (1991) and papers cited there. Note that the condition Enw2,, = EnW2n+l = 1 corresponds exactly to tbe requirements m(0)= 1,~ ( z=)0. 5 In a sense, constructions of compactly supported scaling functions and & wavelets can therefore be viewed an epecial cases of rehement schemes. I feel
CHAPTER 6
210
that there is a difference in emphasis, however. A general refinement scheme is associated with a scale of multiresolution spaces V, , generated by the F(2-3x - n), but typicdly no attention is paid a t all t o the complementary subspaces of the V, in c-1. Refining a sequence of data points in 1 steps corresponds t o finding a function in W-, of which the projection onto Vo, as gven by the adjoint of the refinement scheme (usually not an orthonormal projection), corresponds to the data sequence. There are many such functions in V-,, corresponding to the same data sequence, but the refinement scheme p i c b out the "minimal" ope. There is no interest in the study of the other non-minimal solutions in V - j , and how they differ from the unique refinement solution. This is natural: refine ment schemes are meant to build more "complicated" structures from simple ones (they go from Vo t o V-,). In contrast, wavelet analysis wants t o decompose arbztrary elements of V-, into building blocks in Vo and its complement. Here it is absolutely necessary to stress the importance of all the complement spaces Wt = Vt-l a Ve, and to have fast algorithms to compute the coefficients in those spaces a s well. This is where the wavelets enter, for which there is typically no analog in general refinement schemes. There is another, amusing link between orthononnal wavelet bases with compact support and refinement schemes: the mask associated with an orthonorma1 wavelet basis is always the "square root" of the mask of some interpolation scheme. More explicitly, define Ilfo(() = lmo(()12= En&, e-'*, i.e., hk+,. Then the urn are the mask coefficients for an interpolation urn = refinement scheme, since 2 ~ =2 ~ hk+2n= 6n,o(see (5.1.39)). In particular, as noticed by Shensa [1991), the interpolation refinement schemes obtained from the choice R r 0 in (6.1.11) are the so-called Lagrange interpolation schenes studied in detail by Deslauriers and Dubuc (1989),16of which (6.5.14) is an example. Note that it is impossible (except for the Raar case) for a finite orthonormal. wavelet filter mg t o be itself an interpolation filter as well: orthonormality implies jmo(t)I2 Imo(< r)I2= 1, while the interpolation requirement is equivalent t o hZn = -&6n,0, or mo(t) mo(C r ) = 1. If both requirements are met, then
z,
+
zk6
+
+
+
Assume that h, = 0 for n < Nl,n > Nz,and hN, # 0 # hN2. Then (6.5.24) already implies that either Nl = 0 or N2 = 0. Suppose Nl = 0 (N2= 0 is analogous); N2 is necessarily odd, N2= 2L 1. Take k = 2L in (6.5.24). Then
+
Since ho = 2-'I2, and h2L+l'# 0, this implies hl = 0. Similarly k = 2L - 2
COMPACTLY SUPPORTED WAVELETS
leads to
- which, together with hl = 0, hz, = 2-'/2~,,o implles hs = 0. It follows eventually that only ho and h2L+1are nonzero; they are both equal to 1/aso , that then forces the mask is a "staetched" Haar mask Orthonormality of the L = 0, or m,-,(O = ' i ( 1 + e - ' t ) , i.e., the Haar basis. If we lift the restriction that mo is a trigonometric polynomial, i.e., if 4, $ can be supported on the whole real line, then mo(<) + mo(< ?r) = 1 and Imo(<)12 I%(< ?r)I2 = 1 can be satisfied simultaneously by non-trivial mo; examples can be found in Evangelista (1992) or LemariCMalgouyres (1992).
+
+
+
Notes. 1. A compactly supported 4 E L2(R) is automatically in L1(w) It then follows from Remark 5 at the end of 35.3 that (0) = 1, mo(?r)= 0, i e , has a zero of multiplicity a t least I in ?r. that 2. In Daubechies (1988b), the solutions P of (6.1.7) were found via two corn-
binatorlal lemmas. The present more natural approach. using Bezout's theorem, was pointed out to me by Y. Meyer.
3. This formula for PN was already obtained in Hennann (1971), where maximally flat FIR filters were designed (without any perfect reconstruction schemes, however). 4. This convergence also holds if infinitely many h, are nonzero, but if they (hnl(l In\)' < oo for some c > 0. In that decay sufficiently fast so that case Isin nCI 5 ]n<~""('-") leads t o a similar bound.
+
5. We use here the classical formula sin x -= x
m
J-
m(2-Jz)
.
3=1
An easy proof uses sin 2a = 2 cos a sin cr to write 3
I-J cas(2-'x) 3=1
[
n J
=
j=l
sh(2-J+'s) 2 sin(2-I x)
-
sin x 2 ain(2- Jx) '
fer J+m. In Kac (1959) this formula is credited to which tends to Vieta, and used as a starting point for a delightful treatise on statistical independence.
1 h
kr
6.
This is true in general: if mo satisfies (6.1.1) and 4, as defined by (6.2.2),
generates a non-orthonormal family of translates h,,,then necessarily Cr 27rt)l2 = Q for m e 6. (See Cohen (1990b).)
I&( +
'
7.
The condition $ dx +(x) @(z - k ) = bi,o may seem stronger thsn 11311= 1, but eince the $jVr constitute a tight frame with frame constrcnt 1, the two are equimlent, by Propition 3.2.1.
8. Since +(XI ia a Bnite linear combination of translates of @x), fast alge rithrns td plot 4 lead to fast plots for 3. Throughout this section, we t restrict our attention to 9 only. 9. If 4 is not continuous, then the qj still converge to 4 iu Li (see $6.3). Moremet, they converge to & pointwise in every point wherei# is'contibuous.
101 The choice E = 1 was ueed in the proof of Propitdon 3.3 in Daubechiea (1988b), because the i)jl are abeolutely integrable, wherearr the qo are not. In Daubechia, (1988b) the convergence of the $ to 9 was actuahYproved first (using some extra technical conditions), and orthonormality of the h , n WEM then deduced from this convergence. 11. Note that there exiet many. other procedures for plotting graphs of wavelets. Instead of a mfinement cacpcade one can sleo stcrrt from appropriate #(n) and then compute the $(2-jk) directly from Nz) = h n 4 ( 2 ~- n). (In k t , when 9 is not continuoua, the M e 4gorithm may diverge, while this direct use of the 2 - d e equation with appropriate #(n)stilt converges. I would l i i to thank W. Sweldens for pointing this out to me.) This more direct aomputation can be done in a tree-like procedure; a different way of looking at this, avoiding the tree coe&ruction and leading to faster plots, uaes a dynamicel systems hamework, as developed by Bergei. and Wang (aee Berger (1992) for a review). The "zoom in* feature is loet, hawever.
fix,
12. Many experts on refinement or subdivieion schemes find the multidimenaional case much more interesting! 13. Thia is not a presentation with fullest generality! We merely suppose that the wk are such that there exists a continuoua limit. This already impliea Cwl,=Cwm+t=1.
14. For example, any compactly supported w with bounded variation wouId be "reasonablen here. 16. The following stretched Haar h c t i o n ahow8 how the F ( z - n) can fsil to be independent. W wo t w = 1,811other w, = 0. The solution to (6.5.23) is then (up to n o m d h f h ) F ( z ) = 1 for 0 5 z < 2 , 0 oth*. In this case the P-sequence A defined by A,, = (-1)" leads to EnAn F ( z -n) = 0 8.e. 'I
16. This hr no coinddenoe. If we Bx the length of the symmetric ater Mo = Iwf2,then the choice R r 0 means that Mo ia divisible by (1 coee) with the higheut poesible multiplicity compatible with its length arid the constraint Mo(e) Mo([ + r )= 1. On the other hand, Lagrange r e h e m a t
+
+
213
COMPACTLY SUPPORTED WAVELETS
schemes of order 2N - 1 are the interpolation schemea with the shortest length that reproduce all polynomials of order 2N - 1 (or less) exactly from'their integer samples. In terms of the filter W (c) = C, W, ei*, this mema
4
W ( 5 )+ W((-t- T ) = 1
(interpolation filter: w, = 6n,o)
and W(C) = 1 +
= 1
+ O((1 - ~
0 8 ~ ) ~ )
(see Cavaretta, Dahmen, and Mitxhelli
(1991), or Chapter 8).
+
The two requirements together mean that W(t n) has a zero of order 2N in E = 0, i.e., that W ( t + n) is divisible by (1 - coe E)*; hence W([) by (1 C W ~ It ) ~follows . that W = Mo.
+
-
CHAPTER
7
More About the Regularity of Compactly Supported Wavelets
The regular~tyof the Meyer or the Battle-LemariB wavelets is easy to assess: the Meyer wavelet has compact Fourier transform, so that it is Cw,and the BattleLemarik wavelets are spline functiorls, more precisely, piecewise polynomial of degree k, with (k - 1) continuous derivatives at the knots. The regularity of compactly supported orthonormal wavelets is harder to determine. Typically, they have a non-integer Holder exponent; moreover, they are more regular in some points than in others, as is already illustrated by Figure 6.3. This chapter presents a collection of tools that have been developed over the past few years to study the regularity of these wavelets. All of these techniques rely on the fact
.
C
(7.0.1) r(x) = cn o ~ -xn ) where only finitely many c,are nonzero; the wavelet $, BS a finite linear combination of translates of 4(2x), then inherits the same regularity properties. It follows that the techniques exposed in this chapter are not restricted t o wavelets alone; they apply as well to the Lasic tunctions in subdivision schemes (see $6.5). Some of the tools diussed here were in fact first dweloped for subdivision schemes, and not for wavelets. The different techniques fall into two groups: those that prove decay for the Fourier transform and those that work directly with 4 itself. We will illustrate each method by applying it to the family of examples ~ q constructed 5 in $6.4. It turns out that Fourier-based methods are better suited for asymptotic estimates (rate of regularity increase as N is increased in the examples, for instance); the second method gives more accurate local estimates, but is often harder to use. References for the results in this chapter are Daubechies (1988b) and Cohen (1990b) for 57.1.1; Cohen (1990b) and Cohen and Come (1992) for 57.1.2; Cohen and Daubechies (1991) for 87.1.3; hubechies and Lagarias (1991, 1992), Micchelli and Prautzsch (1989), Dyn and Levin (1989), and Rioul (1991) for $7.2; Daubechies (1990b) for 57.3.
6,
7.1. Fourier-based methods. he Fourier transform of equation (7.0.1) is
= nzo(tl2) 215
,
216
CHAPTER 7
3 x,
where m(<)= ~ e - ' * is a trigonometric polynomial. As we have seen many times before, (7.1.1) leads to I
where we have assumed ~ ( 0 = ) 1 and can be factorized as
J & +(x) = 1, as u ~ u d .Moreover,
1
where C is a trigonornet& polynomial as well; this leads to
I
A first method is based on a straightforward estimate of the growth of the infinite of the C(2-jt) as I(l-)oo.
I
I
7.11 Brute force methods. For 9 = n + P , n E N, 0 5 /3 < 1, we define Cato be the set of functions f which are n times continuously differentide and such that the nth derivative f is Holder continuous with exponent 0, i-e., ~f(") ( 2 )
-I(")( z i - t ) l 5 c ) t l P for a11 z,t .
It is well known and easy to check that if
I
then f € Ca.In partkuh, if 1/(015 C(1+ Itl)-l-Q-t ,then it follows that / E Ca.It follows that, if the growth for K]+oo of C(2-it) in (7.1.4) can
I
G,
be kept in check, then the factor ((1 - e-e)/y)Nensures 8mmthnss for LEMMA7.1.1. i / q = supt )t(c)(< 2N-a-11 then 4 E C".
Proof. 1. Since ~ ( 0=) 1, L(0)= f ;hence It(t)1
2. Now take any 6, with Hence m
L 1 + CKt. a
C$
~ u e n t l ~ ,
Kt 2 1. There exists J 3 1so that 2J-1 5 (€1 < 2J. J
w
.
MORE ABOUT COMPACTLY SUPPOKX'ED WAVELETS
Consequently, I ~ ( E )5) C"' (1 t
+ )[l)-n-l-t,and 4 E Ca.
217
Grouping together several L: leads to better ~stimates,as follows. LEMMA7.1.2. Define
i
r'
-
~
K
inf ECJ .
=
JGN
zjK < N - 1 - a , then&€ CP
Then K =lim,,,K,;
1. Take jz
> jl. Then j2 = njl + r 932
with 0
r
< r < jl, and
(93,)"
9;: . I
Consequently,
KJ, 5
nlogq,, j2
+ rlog91 5 K,, +
log 2
-
Cjlljz .
2. For any c > 0, there exists 30 so that K = inf, Kj > KJo - c. For j 2 jo we then have K1 5 K: c C j o / j K: c. Since c was arbitrary, it'
+ +
r
follows that
K
= lim,,,
J -W
+
KJ
3. if K: < N - 1 - a, then Kt < N - 1 - a for some t' E N. We can then repeat the argument in the proof of Lemmrr 7.1 . l , applying it to
n:;:
with &(() = 4 2 - I < ) , add with 2' plying the role of 2 in Lemma 7.1.1. This W to J&c)~5 ~ ( ItI)-N+"t;l l +5 C(1+ kl)-a-l-c, hence 4 E CP. a ;
The f 6 W n g lemma. &owe that in 'most cases, we will not be able to &sin
:
much better by the btute force method. LEMMA 7.1.3. There d t s a sequence
so
UIot
218
7
CHAPTER
1. By Theorem 6.3.1, the orthonormality of the &(- - n) implies the existence of a compact set K congruent to [ - A , T ] modulo 2a, such thht [&([)I 2 C > 0 for ( E K..Since K is congruent to I-n, lif ~ n LC d is periodic with period 2'+'n, we have
i.e., there exists Cc E 9 K so that (Lt(Ct)l = qr. Since K is compact, the 2-' Ct E K are uniformly bounded. We therefore have I
/
ICrl 5 2 (2
(7.1.7)
for 0 < C'.
2. Moreover, since
1 1
= Icos el21 5 1, we have for all ( r 2'K,
Putting it all together we find for tt = 2Ct
Since K = infr Kt, this is bounded below by a strictly positive constant. Let us now turn to the particular family of N@ Lonstructed in $6.4, and see how these estimates perform. We have
with
We start by establishing a few elementary properties of PN.
i
I
1
I
11
MORE ABOUT COMPACTLY SUPPOWTED WAVELETS i
LEMMA7.1.4. The polynomial PN ( x ) = il $ the followrng properties: 7F
05 x 5 y
A
05 x
$
c
<1
2 19
~~~~
=+ x - - ~ + 'P N ( x )2 p - N + l ~ N ( y ) , P N ( x )2 2 N - 1max ( 1 , 2 ~ ) ~ .- '
(7.1.8) .
(7.1.9)
F
Proof. 1 . I f 0 5 x 5 y, then
2. Recall (see 56.1) that
PN is the solution to
i, ~ ~ ( 3 )
On substituting z = it follows that PN(1/2) = 2 N - 1 . For x < we have P N ( z ) 5 = 2N-1 because PN is increasing. For x 2 3, applying (7.1.8)leads to PN( x ) 5 z N - 1 2 N - 1 P N ( i= ) 2 N - 1 ( 2 x ) N - 1 This . proves (7.1.9). m
i,
It is now easy to apply Lemmas 7.1.1 and 7.1.2. We have
Lemma 7.1.1 allows us to conclude that the Nt$ are continuous. In view of the graphs in Figure 6.3, which show the t$ to have increasing degree of regularity as N increases, this is clearly not optimal! Using 1C, for j > 1 immediately leads to sharper results. We have, for instance,
(because sin26 = 4 sin2€12 ( 1 - sin2€12)) .
4 I
4 + 4 (implying 4p(l - ,y) 5 f), then < p(N-l)by (7.1.9). In the remabiq window $ + 9 2
If eitha y 5 1/2 or y 2 [P~(u)P~(4y(1 y))] a,? $,wehave PN(v) PN(4y(l- Y))
<
'21(N-1)[ l d ( 1 - y)lN-'
C o ~ ~p, 5 l p(N-l) y 3-3(N-1)/2,and K2 5 (N 1)[2 - f 6 1 .It £ o U m that arymptoticdly, for huge N , N # E Q N with p = f - 1 u .1887. A sIightly better value can be obtained by estimating qr rather than p?; one finds then p N -1936. Note that y = is a fixed point for the map y w 4 y ( l - y), so that qk 1 [PN(3/4)Ikfor any k, leading to a lower bound on K: and an upper bound on the regularity of 4. In terms of <,y = sig $ = f co~~esponds to 4 = $;we already saw earlier that f play 8 epecial role because ( is an invariant cycle lor r n ~ l t ~ c a t i obyn 2 modulo 2n. In the next subsection, we will see how these invariant cycles can be used to derive decay estimates for +
9, T) 4.
9
7.1.2. Decay estimates from bvariaat cycles. The values of L: at an invariant cycle give riee to lower bounds for the decay of LEMMA 7.1.5. If f b , t l , - ., C [-n, .rr] is any non-trivial invariant cycle (z-e., b # 0) for the map T( = 2( (modulo 27r), uric tm= ~t,,,-~, m = l , - . - , N- 1, 7tMe1 = t o , then, for d l k N~,
-
where R =
kq~ IL(&,)l/(iW log2), and C
4.
> 0 is independent of k .
Proof. 1.
F i note that there existe Cl > 0 m that, for all k E N, Indeed, 2kMb = &I (mod %), ao that (7.1.10) f o U m if b # 0 or f T . We already know that to# 0; if b = fr , then = 0 (mod 2 ~ ahd ) hence & = 2 M - 1 =~ 0 (mod 2n), which ie impossible.
2. Now
L
MORE ABOUT COMPACTLY SUPPO-KI'ED WAWLETS
221
Since & ia a trigonometric polynomiai and L(0) = 1, there exists Cz so that I&(<)] 2 1 - C2 I,$/2 e-2 Ca for small enough. Hence, for r large enough, w
aD
Hence
We can apply this to the example at the en&of the last subsection: Lemma 7.1.5 ihpiies )&2"%)1 2 C(l + 1 2 " 9 1 ) - ~ +with ~ l? = log JL(F?j)L(-%)l/flogZ. If L has only real coeffic; 3"s (as is the case in most applications of practical interest), then It(-&)\= ": '?)I, and = log lC(F)(/ log 2. The next short 3, 3 ,-F, 2n 4~ 2n % invariant cyclea are {T, -7)) IT , ,- y6u) , etc.; each of them givea an upper bound for the decay exponent of 8. In some cases one of these upper bounds on a can be proved to be a lower bound as well. We firat prove the following lemma. LEMMA 7.1.6. Suppose that [-n,n] = Dl uD2 . . U D M ,and that thm ezicrts
,
q>Osothat
h
i? k
IL(0 t ( X ) . L(2""<)I Iq M [€DM. men 14(t)1 5 C(l+ ICI)-N+", with K = log p/ log 2.
,
:
Prw,f.
1. Let ua &hate. I
n I:
~ ( 2 - ~ < ) 1for , mme larjp but arbitrary j. Since C E Dmfor O T m me € {I, 2, M),
i;
We can now apply t,he same trick to 2-m<, and keep doing so until we cannot go on. At that point we have
with at most 111 - 1 different L-factors remaining (i.e.,r
< M - 1). Hence
with ql defined .as in (7.1.5). Consequently, with the definition (7.1.6),
and K = lim,,, Lemma 7.1.2.
K,
<
log q / log 2. The bound on
& now
follows horn
rn
In particular, one has the followirlg lemma. LEMMA 7.1.7. Suppose that r)
+
Then I&[)] 5 C(1 I
Of course, Lemma 7.1.7 is only applicable to very special t;in most cases (7.1.11) will not be satisfied: there even exist L for which C(%) = 0. Similar optimal bounds can be derived by using otherinvariant cycles as :reakpoints for a partition of [ - A , T T ] , and applying Lemma 7.1.6. Let us return to our "standardn example N ~ .In this case Cohen and Conze (1992) proved that try([) does indeed satisfy (7.1-1I), as follows. LEMMA7.1.8. For dl N E N, N 1, PN(y) =~~~~( N - '.+n)yn ~ntLjks
>
I
We start by proving yet another property of
PN.
223
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
LEMMA 7.1 9. P ~ ( x )=
N
(PN(x) - P ~ ( ~ ) X ~ .- ' ] 1-x
(7.1.14)
Proof.
1.
3. Combining (7.1 15) and (7 1.16) gives
(1 - x)P',(x) Since PN(l) =
=N
N-1 En=., (N - ln+ n )
=
(
*
I ) , (7.1.14) follows. m
~
~
i
We now tackle the proof of Lemma 7.1.8. Proof of Lemma 7.1.8.
1. Since PN(y) is increasing on [O,l], we only need to prove (7.1.13). 2. Define f (y) = PN(y) PN(4y(l
- 3)). Applying Lemma 7.1.9 leads to
CHAPTER 7
224
3. Since 4y(l - y) 5 y for y
1 3/4,
we can apply (7.1.8) to derive
Substituting this into (7.1.17) leads to
The quantity in square brackets equals *(I - g)PN(y) 1 0 for y 5 1, so that g(y) 5 o for 5 y I It follows that PN(y) PN(4g(l - y)) is decreasing on which prove6 (7.1.13) for y _< 9.
i.
(i, I],
4. For
5 y 5 1 we follow a different strategy. PN(?) by Lenrma 7.1.4, it suffices to prove
(e)N'l
Since PN(y) 5
But PN(4y(l- y)) 5 [I - 4y(l - ~ ) I - N= ( 2 -~ 1)-2N ( b m (1 PN(t)= 1 x N P N ( -~ x ) I), and
-
<
(7.1.19) where we have @ Lemma 7.1.4 again, as well ae
I
To prove (7.1.18) it is therefore sufficient to prove
S i both (2y - 1)-2 and y(2y - I ) - ~are decmsb on to verify that (7.1.20) hold for y = i.e., that
1,
Thie ia true for N 2 13.
I
[i,11, it d c e a
I
225
MORE ABOUT COMPACTLY SUPPORTED WAVELETS B
8
5. It remains to prwe (7.1.13) for 5 y 5 1end 1 hostepa: y 5 po = and y yo. For y hence, again by Lemma 7.1.3,
r
9,
3
Similarly P N ( ~5)
v,
- y)IN-'PN (4) = [16y(l - 9))N-l .
P N ( ~ P-( v)) ~ 5
33
>
< N 5 12. We do tbis in < 4y(l -y) _> 4,
( ~ ) ~ PN($), - l
-
PN(u) PN( 4 ~ ( 1 Y))
<
);(
so that N-1
(i)
PN
5 (:)"-'PN(:)
I l d ( l - v)]N - l
,
(7.1.21)
-
[i,
b e c a w g ( l - y) is decreasing on 1 . One checks by numerical computation that (7.1.21) is indeed smaller then [PN(!)I2 for 1 5 N 5 12. 6. F o r u p = PN(9) 5 ($"l
P N ( ~ P ( ~))PN(V)
5
Y;~+'PN(W)(~V - 1)-2
3
[py-
zN PN(M),
lN-I (7.1.22)
where the last inequality uses that (2y - I)-' armd y(2y - 1)-2 are both decreasing in [yo, 11. One checks by numerical computation that (7.1.22) issmaller than [PN($)]2for 5 5 N < 12.
< y 5 1. For the& 7. It remains to prove (7.1.13) for 1 5 N 5 4 and small values of N the polynomial PN(y) PN(4y(l- y)) - PN(:)' ha8 degree at most 9, and its roots can be computed easily (numerically). One checks tw there am none in j i , I], which fini&q the proof, because (7.1.13) is &bfiediny=l. rn '
It f o l h from Lemmas 7.1.8 and 7.1.7 that we know the exact r a y q t o t i c decay of N&):
IN & < ) ~ 5 c ( l + l t ~ ) - N + l -
IJ'JW'~/~)V/"""~ ,
(7.1.23)
225
I
CHAPTER 7
For the first few values of N this translates into estimates for a:
Nc$
E Ca"
with the following
We can also use Lemma 7.1.7 to estimate the smoothness of Nq5 as N - y o . Since
1 3 " ~ '5 P N ( ] )< aN-'
(uee Lemma 7.1.4 for the upper bound, (7.1.19) for the lower bound),
'
9
h
logIp~(3/4)1r -log3 21og2 21og2 N[1-
o(N-'log A!)] ,
&
12 implying that asymptotically, for large N , N) E CpN with p ,2075.' One does not, in fact, need the full force of Lemma 7.1.8 to prove this asymptotic result: it is eufficient to prove that
PN(Y)5 C 3N-1 for y 5
(7.1.24)
PN(v) P N ( ~ Y-($1) ~ 5 c 32(N-1) for 3 5 y 5 1 ,
I I
,
(7.1.25)
with C independent of N . The asymptotic result then followa immedi8tely from Lemma 7.1.6. The estimate (7.1.24) ia immediate from PN(y) 5 PN(5) 5: 3N-' for y < Q; the estimate (7.1.25) follows easily h m Lemma 7.1.4 as followe. u 5y5 then PN (y) P N ( ~ P-( y)) ~ 5 (4ylN-' (ley($ u)IN-l = [W2(1 Y)IN-l < - 32(N-1) b m w y2(l - p) is decreasing on [$,I]; if y 5 1, thm PN(P) P N ( ~ v-( ~u)) 5 ( ~ u ) ~ - ' P N ( ) )= ( 8 ~ ) N - 1 < 4 32(N-1). This much simpler argument for the exact asymptotic decay of is due to Volkmer (1991), who derived it independently of Cohen and Conze's work.
3
-
v,
-
a<
6
761.5. Littlewood-Paley type estimates. The estimatea in this eubeection we L1 or L2-estimates for (1 K I ) Qrather ~ than poinhbe decay estimatea for itself. The basic ides is the usual Littlewood-Paky technique: the Fourier transform of the function gets broken up into dyadic piecea (i.e., roughly 2jC 5 MI 5 2j+'C) and the integral of each piece' estim a M . If $g
+
,I
227
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
x;,
2JaU] < oo if o < - log A/ log 2, implying 4 E Ca.To obtain -ti; mates of this nature, we exploit the special structure of 4 ae the infinite product - ofmo(2-jc); the operator Po defined in 56.3 will be a basic tool in this derivation. we will first restrict ourselves to positive trigonometric po1ynomials Mo(<). (Later on,we will take 'Mo([)= Iw(t)12 to extend our results to non-poeitive m~.)As in $6.3 we define the operator Po acting on 27r-periodic functions by C[1+
,Po,,(<,
(f) + M o ( $ + r ) f
= ~ o ( f ) f
($+%)
.
This operator was studied by Come and Flaugi and several 6f, the results inthis subsection come from their work (Conze and Raugi (1990), Corn (1991)). Similar ideas were a h developed independently by E-irola (1991)and Villemoes (1992). A first useful lemma is the Wowing. LEMMA7.1.10.For all m > 0 and 411 2n-penodic junctiow f ,
Pmf. 1. By induction. For m = 1,
3nf2
*
.
= 2/_d2 MO,C,~,,, =
J2% -2u
4f(f)Mo
(f) .
CHAPTER 7
Since M' is a positive trigonometric polynomia3, it can be written as
+
'
One then finds that the (2J 1)-dimensional vector space of trigonometric polynomials defined by
is invariant for Po. The action of Po in VJ can be represented by a (25 (23 1) matrix which we will also denote by Po,
+
' (P0)kt = 2 a 2 k - t
with the convention a, = 0 if Irl
t
' - J s k , t?< J :
+ 1) x
/
> J . For Mo of the type
where L is a trigonometric polynomial such that L ( r ) # 0, the matrix Po has very special apectral properties. LEMMA7.1.11. The values 1, f ,- - .,2-2K+1 am eigenvduea for Po. The row vectors e, = (jk)i,-J, ..,J, k = 0, . .,2K- 1 genemte a subsplzce which is left invariant for Po. More precisely,
ekPo = 2-kek
+
linear combination of the en, n < k
.
I
I
MO-
I I
ABOUT CO&fPACnY SUPPOFWED WAVELETS
229
1. The factorization (7.1.28) is equivalent to "" I
a, jk(ll)l
- .
= 0 for k = O , - - - , 2 K 1
1.
This means that the Moreover, since Mo(0)= 1, C a 2 3 = %j+l = sum of each column in the matrix (7.1.27) is equal to 1; Q is thus a left eigenvedor of Po with eigenvalue 1.
it
: it.
2. For 0 < k 5 2K \
i
- 1, define gk = ekPo,i-e., (gk).. = 2 z j k oy-,
.
3
t For m even, m = 2!,
: 1
Hence
where
A,,, = Co?,(~j)"' I
F
= ~ 0 z J + 1 ( 2+j lIrn f
A amsequence of Lemma 7.1.11 is that the epacea Ek,
bi%b 1 < k 5 2K,are all right invariant hr Po. The main result of this subsection irr then the foflowing.
230
CHAPTER 7
THEOREM 7.1.12.Let X be the eigenvalue of POIEaK value. Define F,a by
with the largest absolute
-
If IXI < 1, then F € Ca" for dl c > 0.
2. The spectral radius p(Polh,)equals IX/. Since, for any 6 > 0, there exists C > 0 so that (IAnII I C(p(A)+ 6)n for ail n E N, it follows that
3. On the other hand, f(<) 2 1 for 5 < )(I< K. Together with the boundedneas of fl& ~ ~ ( 2 - j for t) 4 T ( d e d d as us& from IMo(t)I 1+ Cltt), t h ~mplies ~
<
(use Lemma 7.1.10)
L C'(jA)+ 6)"
.
By the arguhent at the start of this subsection, this implies P E Ca-'.
rn
In fact, a slightly stronger result can be p ? d . If we extend the definition of C" (n integer) to include 'all the functions for which the (n - 1)th derivative is in the Zypund class /
3 = {f; 0
i
If
(x + 9 )
+ f(x- 9 ) - 2f (41L Clvl
for all ~ 9 8 ,)
then it is also true that F E Ca if fiIhK is diagonal (i.e., we can drop the c in this case). Moreover, both this smoothness estimate and the estimate F E Ca-' for all E > 0 in Theorem 7.1.12 sre optimal if F has no eero on [-r,ri. For a proof, see Theorem 2.7 in Cohen and Daubechiea (1991). /
231
MORE ABOUT COMPACTLY SZFPPOR'IXD WAVELETS
REMARK. The same result can be derived via an equivalent technique, which uses an operator Pk dehed in the 88me way as Po, but with Mo(e) replaced by the factor L([)in (7.1.28). In this case we define X~ = ~I(P;),and we factorize F(() [2(sin[/2)flzK (2~)-'/' H z l L(2-JF) to obtain
-
so that F E Ca-' with a = 2K @. Thh J I U $ ~has ~ the advantage that we start directly with a smaller matrix P t , so that the computation of the spectral radius is simpler. The tp.o methods ate compkte1y equivalent, as shown by the following argument. If p is an eigmvalue'd Po, with eigenfunction f, €&*,
then j, can be w-ritten as K
f p ( 0 = (sin'
f)
sp(0
-
&placing Mo(() by its factorized form in
~ f J ' (= t , M.
(f)f, (d)
+Y
(f
+I)
fp
(g +
x)
9
+I)
9
we obtain after dividing by [sin2 f me2 f l N , 82zKg,(~) =
L (f)%
(f) + i(; +*) (f g,
so that the eigenvalues of Pk are exactly given by IcL = 22Kp.
o
will not be positive. (Indeed,in the framework of orthonormal In genera), wavelet bases, mo is never positive, except for the Haar basis; see Janssen (1992).) However, we can then define Mo = Irnot2; the same techniques lead to
'\ e
where
is the spectral radius for P;, with L(() = JC(<)12.Hence
I
232
CHAPTER 7
6. Consequently 4 C"" for a N + of resulting cr, for the first few
if a + < N + For the special
I
E
NC$
$6.4, the
.XL
- 4.'
I
.
1
values of N, are:
Theae are much better than the values obtained from the pointwise decay of 4 (see 47.1.2). The size of the matrix Pk increases with N (linearly), and I do not know of any way to determine the asyrnptotics of its spectral radius as N-+oo; for asymptotic estimates, pointwise decay of is the best method.
6
II I1 1
7.2.
A direct method.
e
The smoothness results obtained at the end of $7.1.3 for N4 with small N are still not optimal. Moreover, Fourier-based methods can only give information on the global Holder exponent, whereas it is clear from Figure 6.3 that 29, for instance, is smoother in some points than in others. In fact, we will see that there exists a whole hierarchy of (fractal) sets in which 2 4 has different Holder exponents, ranging from .55 to 1. Results such as these can be obtained by direct methods, not involving For the sake of simplicity, I will explain tbe setup in t4e general case, but expose the method in detail for the example 2i$ only, and (hen state the general theorems on global a+ local regularity without proof. Proofs can be found in Daubechies and Lagarias (1991, 1992). Similar results about the global regularity were also proved (independently, a ~ in d fact before Daubechies and Lagarias) in Micchelli and Prautzsch (1989), in the framework of subdivisiop schemes. The method is completely independent of wavelet theory. The starting point is the equation
6.
c:,
ck = 2, and we are interested in the compactly supported L1-solution with F, which: if it exists, is uniquely determined3 (up to its n o ~ s t i o n ) :since F E L1, F is continuous, and (7.2.1) implies
I
*
MORE ABOUT COMPAGTLY SUPPOKI'ED WAVELETS
233
4
with m(<) = ~ f = ~ = g qe-ik6, and Lemma 6.2.2 the. tells us that support F = (0, Kj. Equation (7.2.1) can be considered as a fixed point equation. Define, for t ' functions g supported on [O, K], Tg by K
(Tg)(x) =
C
Ck
9(2x - k) -
Then F solves (7.2.1) if TF = F. We will try to find this fixed point by the usual method: find a suitable Fa: define F, = T ~ F and ~ , prove that F, has a limit. To determine Fo, note first that (7.2.1) impoees a conatra.int on the values F(n), n E Z if F is continuous. Since support F = [0, K ] , we only need to determine the F(k), 1 5 k 5 K - 1; the other F(n) are zero. Substituting x = k, 1 _< k 5 K - 1 into (7.2.1) leads to K 1 linear equations for the K - 1 unknowns F(6); the system of equations, can also be read as the requirement that the -or (F(l), . .- ,F ( K 1)) be an eigmvector with eigenvalue 1 of a (K - 1) x (K - 1)-dimensional matrix derived from the c k . It turns out that, ,modulo some technical conditions (see below), this matrix does indeed have 1 as a nondegenerate eigenvalue, so that (F(l), - ..,F ( K - 1)) Can be fixed up to an overall multiplicative constant. Let us supthis is done. One can prove that F(k) # 0, k we can fix the mmnahation so that F(k) = 1. (All this will be illustrated by an exam& below.) Define now Fo(x) to be the piecewise linear function which takes exactly the values F(k) at the integers, i.e.,
-
-
xf=il
EL<'
Fo(z) = F(k)(k + 1 - x)
+ F(k + 1){x - k)
for k 5 x i k
+1 .
(7.2.2)
Successive applications of T define the Fj = TiPo, i.e.,
it eaeily follows that the Fj are piecewise linear with nodes at the 2-in E (0, K], n € M. To discuss whether the F, have a limit as j-+oo, and to study the regrilarity of this h i t , it is convenient to recad (7.2.3) in mother form. The key idea is to study the F,(x), Fj (z+l) Fj(z K - 1)simultaneously, for z E [O,lj. We define vj(z) E R~~ by
.
+
Fm 0 5 z 5 4, (7.2.3), together with support 4 c [O, K j , implies that F ~ + ~ ( x ) , Fj+l(z + 1),- ..,Fj+l(x + K - 1) are all linear combinations of F, (2x), Fj(2z + 1),.-.,4(2x + K - 1). More prec%ely,in i nof vj(x), q + l ( z ) = To vj(2x) for 0 5 z_<
,
(7.2.5)
where Tofe the K x K matrix defined by ( T ~ ) m n = %-n-1
15 m, n I K ,
(7.2.6)
'
234
CHAPTER 7
with the convention ck = 0 for k < 0 or k > K. Similarly, v,+i(s) = Ti ~ ~ ( 2 1) 2 - for
5 z5 1 ,
(7.2.7)
t
where v
(T1)mn =c2m-ny lSm,nIK. (7.2.8) Equations (7.2.5), (7.2.7) both apply for x = because of the special structure of TO,Ti -and vJ (in ~artkular,(TO)",= (TI), n + ~[v,(O)), = [v, (I)], for n = 2, - .. ,K) , the two equations are identical for x = $ . We can w m b i i (7.2.5) and (7.2.7) into a single vector equation as fallows. Every x e f0,1]can be represented by a binary sequence,
i:
with &(z) = 1 or 0 for all n. S t M y spea4ing, two possibilities exist for every dyadic rational x, i.e. every z of the type k 2-2 : we am replace the last digit 1 followed by all wos by a digit 0 followed by dl ones. This wili not cause a problem, but to be clear we h t b g u i s h these two sequences by a s u h i p t : q ( x ) for the sequence ending in zeros (the expansion "from above," ie., the expansion that will start by the same J - 1 digits as x 2- for J-*oo), ti; (z) for the sequepee ending in onea (expansion ''horn below"). For btance,
+
4
The two definition regions 0 5 z < and 3 < z 5 1 of (7.2.5) and (7.2.7) are completely-characterized by dl(z):dl(%) = 0 if z < 3, 1 if z > For every binmy sequence d = we alSO define its right shift r d by ( ~ d= ) ~&+%, n = 1,2,--..
4.
Itisthenclaarthatrd(x) = d(Zc)ifO_Cx< ;,rd(x)=d(&--1)if $ < z 5 1. (For z = we have two possib'Ilith 7d+ = d(O), rd= d(l).) Alth~ugh T is really defined on binary seq.llences, we will make a slight abuse of notation and write T X = y rather than 74s) = d(y). With this new notation, we can rewrite (7.2.5), (7.2.7) as the angle equation
4,
(4)
(i)
If the v, have a limit v, then this Pector-valued function u will therefore be a fixed point of the linear operator T'ddld by %
(Tw)(z) =
Tdl(t)
47x1 ;
T acte on all the wtor-valued functions w : [0,1] requirements = 0, [w(O)j, F(o)]i = 0, [ w ( ~ ) ] K
r [w(~)]L-I,
-
R~ t b t
k F= 2,.
-
satis& the t
- - ,N . -
17.2.10)
MORE: ABOUT COMPACIW SUPPOKMCD WAVELETS
235
(As a result of these conditions Tw is d&ed unambiguouely at tbe dyadic rationals: the two expansions leed to the aame reeutt.) What has all this recasting the equations into different forme done for us? Well, it follows from (7.2.9) that
which implies
-
In other words, information on the spectral properties of products of the Ts mairicee wiil help us to control the dierence v, - v,+t, so tpat we &n prove v, v , and derive smootbnesa for v. But let us turn to an example. For the function 3d (7.2.1) reads
Note thgt Q+C2
=
C I + C ~=
and 2c2 = Cl
+ 3c3 ,
1
~(e)
ck4
both of which are consequence8 of the divisibilityof = f cr e-&{ by (1 + e'e)2. The values 2@(1), 2#(2) are detenmined by the qmhm
Because of (7.1.13), the columns of M dl aum to 1, emuring that (1, 1) is s I& eigenvectnr of M with eigenvdue I. This eigenvAiue is nondegenemk; the right eigenvector for the same eigenvalue is therefore not o r t h o n o d to (1, I ) , which means it can -be normaliized so that the sum of it9 entrim is 1. Thie choke of (24(1), 2@)) lef&
The ~t~
To,TIare 3 r 8 matrim given by
236
CHAPTER 7
Becsuee of (7.2.13), To and Tl have a common left eigenvector el = (1, 1, 1) with eigewalue 1. Moreover, for all z E 10, I],
= (1
-
2 ) [2#(l)
+ 24(2)1 + 2 [24(1)+ od(2ll (use (7.2.2))
= 1.
I
It folious that, for all z E [O, 11, all j E N,
3
= el -vo(71z)
(because elTd= el. for d = 0, 1)
+
Coneequently, vo(y) - vc(y) E El = {w; el .w = wl w2 +US = 01, the space orthogonal to el. In view of (7.2.11), we therefore only need to study products of Td-matrices restricted to El in order to control the convergence of the uj. But more is true! Define es = (1, 2, 3). Then (7.2.14) implies
e2To = ie2 ezT1 = iez
+ aoel , + ale1 ,
eo+b-i
*tho. = = q . 0 1 = cl+2c3-+ = ei = ea - 2ao el, Cben (7.2.15) becomee
e! To =
3 e8
and e4 T~ =
3 e$ - 4 el,
(7.2.15)
v.
or ep T~ =
ifwedhe
4 e! - 3 tiel .
On the other hend, eq - ~ ( z = ) (I
- Z) C! - ~ ( 0 )+ z 4 - vo(l) =
-2
;
I; 1 ;
.T
1 'r
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
I
237
t
It follows that e! - [vo(x)- u~(x)]= 0. This means that we only need to study
,
products of Td-matrices r&icted to &, the space spanned by el and eg, in is oneorder to control v, - v,+t. But, because this is a simple example, dimensional, and TdlE,,is simply multiplication by some constant, namely the third eigenvalue of Td, which is for To, for TI. Consequently,
9
9
1 ~ 1
where we have used that the v1 M uniformly b o ~ n d L b .Since ~ < 1, (7.2.16) implies Ilvj I%) vj+t(x)ll 5 C 2-a, , with a = I log((1 + a)/4)l/log 2 = .550. It follows that the v,- have a limit function v , which is continuous since all the v, are and since the convergence is uniform. Moreover v automatically satisfies (7.2.10), since all the v, do, so that it can be "unfolded" into a continuous function F on [O,3].This function solves (7.2.1),so that 24 = F, and it is uniformly approachable by piecewise linear spline functions F, with nodes at the k2-3,
-
11 2 4 - F J l l t - o 5 C 2-03
(7.2.17)
It follows from standard spline theory (see, e.g., Schumaker (1981))~that 29is Holder continuous with exponent a = .550. Note that thls is better than the best result in 57.1 (we found cr = .5 - c at the end of 37.1.3). This Hijlder exponent is optimal: from (7.2.12) we have
hence
124(2-l) - 240)l = C(2-j)" . But this matrix method can do even better than determine the optimal Holder exponent. Since v ( x ) = Tdl(z)v(rx), we have, for t small enough so that z and x + t have the same j digits in their binary expansion,
This can be studied in exactly the same way as v,(x) - V ~ + ~ ( above; Z ) we fhd .
.
e:! [v(x)- V(Z
+t)] = t .
Fbr the remainder, only tbe T d l s matter, and we fmd
1
CHAPTER 7
238
where t itself is of order 2-j. With the notation r j ( x ) = f can be rewritten as
xi=,d , ( z ) , (22.18)
where P = 1 log l ( 1 - f i ) / ( l + &)jJ/log2. Suppose q ( z )tends to a Limit r ( x ) forj+m. If r ( z ) < = .2368,thcn the second term in (7.2.19) dominates the first, and v , hence 2 4 , is Holder continuous with exponent a @r(x). If r(x)> then the first term, of order 2-1, dominates, and 2 4 is Lipchits In fact, one can'even prove that 2 4 is differentiable in these points, which constitute a set of full measure. This establishes a whole hierarchy of fractal sets (the sets on which r ( x ) takes some preassigned value) on which 2 4 has different Holder exponents.. And what happens at dyadic rationals? Well,there you can define r * ( s ) , depending on whether you come- "from above" (associated with d + ( z ) ) or "From'belown (d- (x)); r+(x) = 0, r , (z)= I. As a consquence, 24 is left differentiable at dyadic rationals, x, but has Holder exponent -550 when z is approached from the right. This is illustrated by Figure 7.1, which shows blowups of 24,exhibiting the characteristic lopsided peaks at even very fine scales.
9
+
9,
In this example, we had twcs Usum rules'' (7.2.13), (7.2.14), decting that &(() = f C1 ck e-'N rss divleible by ( ( 1 e-q)/2)'. In general, .m is divisible by ( ( 1 + e ~ Y ) / 2 ) ~ipd , we b v e N sum rules. The mrbtqmce Enr
+
I
MORE ABOUT COMPACTLY SUPPOFtTED WAVELETS
239
will, h o m e r , be more than one-dimensional, which complicates wtimates. The general theorem about global regularity is as follows. THEOREM 7.2.1. Assume that the c k , k = 0, ..,K, satssfy K c k = 2
Ck=-,
6nd
K
C ( - I ) ~k'ck
= O forI=O, 1 , - - - , L .
(7.2.20)
k=O
For every m = 1,: ,. ,L + 1 , define Em to be the subspace of ltN orthogonal to U M = Span {el, - .,e,,,), where el = (11-', 23-', . .- ,NJ-') . Assume that there ezist 1/2 5 X < 1, 0 <_ e < L (e E N) and C > 0 such that, for all binary sequences ( d j ) j E Nand , dl rn E N,
1 . there ezists a non-trivral contsnuws L1-solution F for the two-scale equation (7.2.1) assocrated with the c,,, ,
2. this solution F is t times continuously differentiable, and 3. if X >
$,
then the llh derivative F ( ~ of ) f is Holder continuous, with expment at least tn A ( / In 2; if A = 112, then the eth denvatiue F(()of F is almost Lapschitz: it satrsfies
REMARK.The restriction X 2 means only that we pick the largest possible integer O 5 L for which*(7.2.21)holds with A < 1. If t = L, then n e c W y X 2' (see Daubechies and Lagarias (1992)); if L < L and X < then we could replace L by L+ 1 and X by 2X, and (7.2.21) would hdd for a larger integer t!. o A similar general theorem can be formulated for the local regularity fluctuations exhibiM by the example of 24. For-a precise statement, more details mil pro&, I refer to Daubechies and Lagarias'(1991, 1992). When applied to the Nt$, these methods lead to the following optin~alHolder exponents:
4
4,
These are clearly better than what was kbtained in §7.1.3;'m~reover,we see to our surprise that s4 is continuously differentiable, even though its graph seems to have a "peakn at z = 1. ~low&s show that thb is deceptive: the true maximum lies a little to the right of x = 1, md everything is indeed smooth (W Figure 7.2). The derivative of 3d is ~031tim10~6, but has a very small IIclder e q ~ ~ ~ eas n i!lurrtrated t, by Figure 7.3.
CHAPTER 7
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
241
Unfortunately, these matrix methods are too cumbersome to treat Iarge examples. Another, more recent "direct methodn has been developed in Dyn and Levin (1989)and Rioul (1991); when applied to the N4 with N = 2,3,4 it re produces .the a-values above; since it is computationally less heavy, it can sJBo taclde larger valuea of N with better results t h in $7.1.3 (see Rioul (1991)). , 1. Note the similarity of the matrices To, TI and Po in 57.1.3 (see (7.1.27))! Even the spectral analysis, with the nested invariant subspcrces, is the same. This shows that the result in Theorem 7.1.12 is indeed optimal: if iis the
spectral radius of 'polhX = . T I J then ~ ~ ,
so that X in (7.2.21) must be at least &, and the H6lder exponent is at most C Ilog XI/ bg 2 5 1 log i t / log 2. The difference between the two appro& is that the present method also givea optimal estimates if &([) is not positive, unlike $7.1.3.
+
2. The condition (7.2.21) suggests that infinitely many conditions on the To, TIhave to be checked before Theorem 7.2.1 can be applied. Fortunately, (7.2.21) can be reduced to equivalent conditions which can be checked in a finitetime computer search. For details, see Daubechies and Lagarias (1992). 3. In practice, it is not necessary to work with To,TI and restrict them to GK. One can also define directly the matrices Po, TI corresponding to the coefficients of m ( t ) / ( ( l + e-g) f 2)K; it turns out )that bounds on llTd, - - .TL]&, 11 are equivalent to bounds on ...f" 11 . 2-Lm (see Daubechiea and Lagarias (1992), 55). The matf'd are much smaller than Td((N K) x ( N - K) i d e a d of N x N). o
~fi,
-
Since this method works for any function satisfying an equation of the type (7.2.1), we can apply it to the basic functions in subdivision schemes. For the Lagpangian interpolation function cornspoxding to (6.5.14), a detailed @pis shma that F is "almoet" @: it ia C1, and P &isfie8
Thie had already been obtained previously by Dubuc (1986). But our matrix methods can do morel They can prove that F b ahalmost everywhere differentiable, and they can even compute F" where it ie well defined. For details, see again Daubechiee and Lagarias (1992). 7.3.
Compactly supported waveke with more regularity.
By Corollary 5.5.2, an o r t h o n o d bmb of *veleta can consist of CN-I wa~elete only if the basic wavelet $ has N Mniehing momenta. (We implicitly aesume
-243
CHAPTER 7
+
that $ stems from a multiresolution analysis and that 4, have sufficient decay; hoth.conditions are trivially satisfied for the compactly supported wavelet bases as constructed in Chapter 6.) This was our motivation to construct the Nq5, which lead to N~ with N vanishing moments. The asymptotie results in 57.1.2 show however that the ~ 4 ~ , q E5 CpN with p = .2. This means that 80% of the zero moments are "wasted," i-e., the same regularity could be achieved with only N / 5 vanishing moments. Something similar happens for small values of N. For instance, 29 is continuous but not C1, 34 is C1 but not 0, even though 3$ have, respectively, two and threb vanishing moments. We can therefore Usacrifice" in each of these t;wo cases one of the wishing moments end use the additional degree of W o r n to obtain 4 with a better Holder exponent than 2 9 or 3 4 4 have, with the same support width. This amounts to replacing Imo(() 1' = (cos2 $)Nfi(sin2 $ ) by lrnO(t)l2 = (m2$)N-1[~N-l(sin2f) a(sin2 f l Nc a d (see (6.1.11)), and to choose a so thkt the regderity of I$ is impmved. Examples for N = 21 3 are shown in Fivires 7.4 and 7.5; the corresponding h, are as follows:
+
These examples correspond to a choice of a such that mexl p (Tol~,),p (TI I )] is minimized; the eigenvalues of To,TI are then degenerate.8 One can prove that the Holder exponents of t b two functions are at leaat .5864,1.40198 respectively, and at moet .60017,1.4116; these last values are probably the true Hiilder exponents. For more detaile, see Daubechies (1990b). 7.4.
Regularity or d h i n g momeds?
The examples in the previoua section show that for fixed support width of #,I), or equivalently, for fixed length of the filters in the associated subband coding scheme, the choice of the h, that leads to maximum regularity is dieerent from the choice with maximum number N of mishing moments for @. The question
I
MORE ABOUT COMPACTLY SUPPOKI'ED WAVELETS
RG.7.4. The r d n g jbction & for the mo& ngJor
wavelet cmutruetum wat/i ouppmt
width 3.
FIG. 7.5. The sealing jhction 4 for
the m o s t regulw wauckt w ~ t r u c t a o nwih support
wdih 5.
then arises: what is more important, vanishing moments or regularity? The answer depends on the application, and is not always clear. Beylkin, Coifrnan, and RokhIin (1991) use compactly supported orthonormal wavelets to cmmpress large matrices, i.e., to reduce them to a sparse form. For the details of this a g plication, the reader should consult the original paper, or the chapter by Beylkin in R u d d et al. (1991); one of the things that make their method work is the number of vahihing moments. Suppase you want to decompose a function F(z) into wave1ets (strictly speaking, matrices should be modelled by a function of two mriabks, but the point is illustrated juet as well, and in a simpler way, with one variable). You compute all the wavelet M c i e n t s (F,qjCk), and to compress all that informtion, you throw away all the coefficieni;~smaller than some threshold r . Let us see what this meaw at some fine scale; j = -J , J € N and J "large." If F is CL-'and J, has L vanishing moments, then, for x near
244
CHAPTER 7
2- jk, we have
where R is bounded. If we rnultipiy this by $(2Jx - k ) and integrate,-then the @t L terms will not contribute because j & x C ~ ( x=) 0, 1 = 0, . . - ,L - 1. Consequently, (F,
=
j
I/&
(r- l-3k)L R ( x ) 2'12 ~ 1 ( 2 ~ kx)-
For J large, this Mil be negligibly smatl, u n k R is verp large near k2- J . After thresholding, we *ill therefore only reb*n fine-scale wawle3 coefficients near singularities of F or its derivatives. The e k t will be all the more pronounced if the number L of vanishing moments of q5 is largeg Note that the regularity of 1C, does not play a role at all in this argurnent3t seems that for Beylkin, Coifman, and Rokhlin-type applications the number of vanishing moments is far more important than the regularity of $. For other applications, regularity may be more relevant. Suppose you want to compress the information in an image. Again, you decompose into wavelets (twedimensional wavelets, e.g-, aseociated with a tensor product multiresolution analysis), and you throw away all the small coefficients. (This is a rather primitive procedure. In practice, one chooses to allocate more piecision to some coefficients than to others, by means of a quantization rule.) You end up with a representation of the type
where S is only a (small) subset of all the possible values, c h m n i11 function of I. The mistakes you have made will consist of multiples of the deleted $,,&. If these are very wild objecd, then the difference between I and f might well be much more perceptible than if 9 L%ploother. This is admittedly very much a band-waving argument, but it suggests that a t least some regulatity might be required. Some fire* experhe& rqmrkd in Antonini e t al. (1991) seem to confirm this, but more expdmWs am required for a convincing answer. The sum rules (7.2.20), egUidt%&to tbe divisibility of m(<) by (l+a"~)L+l, have another interesting m m m ~In. the example studied in detail, &, we saw that (7.2.13) and (7.2.14) impzied'tbt (we proved both for V j , so that they deo hold for v = lim,,,
of
Q rather than
v,
v,), or, in terms
245
MORE ABOUT COMPACTLY SUPPOFLTED WAVELETS
k
-
6
k dl z E p,11. h u s e support(&) = [O, 31, one essily checks that this implies, forallyER,
An polynomials of degree less than or equal to 1 can therefore be written as tinesr combinations of the #(z - n). Something similar happens in general: the conditions (7.2.20) ensure that all the polynomials of degree less than or egual to L can be generated by Linear combinations of the 4(x - n). (See Fix and Strang (1969), Cavaretta, Dahmen, ahd Micchelli (1991).) This can again be dt used to explain why the condition Fmlt=r = 0, C = 0, - .. ,L is useful in ~ u b b a n dfiltering schemes. Ideally, one wants the low frequency channel, after the filtering, to contain all the slowly changing features, and to find only true "high frequency" fehtures in the other channel. Polynomials of low degree are essentially slowly changing features, and the sum r u b (7.2.20) ensure that they (or their restrictions t o a large interval, to keep it all in LZ(R); we disregard border effects here) are in every V j , i.e , they are given completely by the low frequency channel. In the design of FIR filters for subband coding, it is not customaiy to pay much attention. to the number of vanishing moments of nao, which is reflected by the "flatness" of the filter at ( = T." What follows is yet another argument showing that in implementations where filters are cascaded, it is nevertheless important to have at least some zero moments. Suppax that we apply three successive low pass filtering decimation steps t o s signal. If we call the original signal f", with Fourier transform (6) = En2-e-'e, then the result of one !filtering decim~tionis the sequence fA-, where f '(4) = C, f: e-'* satisfies
+
t
+
The second term can be viewed as the result of aliasing, due to the lower sampling rate in the f Similarly, three such operations lead to
'.
-
r3i2
[ p (g)
(8) (t) ma (5) +
I.
7 "folding" tams
(7.4.2) It follows that the product mo(() mo(X) m(q)plays an important role. l?igure 7.6 shows what this product loob like for the ideal low pass filter, mo([) = 1 5 a/2,Ofor u/2 I for I a. If the low pass filter is not ideal, then it will uleak" a little into the high-paes region 1r/2 I I(( 5 A. It is then important \
'
0 FIG. 7.6.
n/4
nt2,
3~14
n
Plbtr of mdt), mo(2*), mo(_QC)and of their pnnfvtfor the W bw-porr
ma.
to contain this leakage, especially when the filters are c d e d : it contribuks to the "folding" terms in (7.4.1) a d can lead to audible or visible aliasing once quantization is introduced and perfect reconstructipn ie no longer attained. In the ideal case'in Figure 7.6, the "bumpn of ~ ( q for)4 iE [3r/4, ?r) is killed In the product mo(<) m(2<) because %(() = 0 in this interval. The same happens for extra "bumps" of ~ ( qresulting ), in ~ ( 6 ~ ) ( q ) = 1 if ( E [0, x / 8 [ ,= 0 if ( E ]n/8,T ] . In the non-ideal case, a eimilar effect can be achieved by imposing that mo have a zero of reasonable multiplicity at ( = x , which "killsn the maximum of rnomo(X)in a concatenation. This phenomenon is illustrated in Figure 7.7, where a wavelet filter ie compared with a wn-wavelet perfect reconstruction filter. In Figure 7.7a we see plots of Im~(()l for two orthonormal perfect reconstruction firters (i.e., [mo(6)12 Imo(( %)I2= I), each with eight taps; the filter on the left componds to the example constructed in 56.4, with two vanishing momente (i.e., bas a zero with double multiplicity at ( = r),and an extra zero at ( = 77r/9. The filter on the right is not a wavelet filter, since m(n)# O and hence ~ ( 0# )I; it is conetructed mom according to standard wisdom, with an lequi-ripple" deeign: in this ceae the location of the nodm is choeen so that the amplitude of the two ripples ia the same es in the one ripple in the wavelet filter on the left, while keeping the trausition band as
~(e)
~(e)
+
+
MORE ABOUT COMPACTLY SUPPORTED WAVELETS
247
narmw as poasible within this constraint. The resulting fi1ter.i~a little steeper than the wavelet filter (its first zero is at ( = . 7 6 ~rather than .78n for the wavelet example), and seems therefore closer t o the ideal fiiter. (Of course, both are quite far from the ideal case, but remember that we have used only eigFit taps!) Figure 7.TowplotsI ~ ( t ) m,-,(q)j for these two examples, and Figure 7.70 blows up these plots in the region 7r/2 5 6 A. It is ciear that in the second (non-wavelet) case the leakage into this high frequency region is more important than in the wavelet case; this is true in L2-senseas well as in amplitude (the highest peak at right is about 3 dB higher thap at left). This effect can become even more pronounced when larger filters are considered."
~ ( x )
FIG.7.7. ComparLon of --h
:
<
thme cmcatenotimrs of two 8-tap low-paaa filters wtth Me perfect
(4p(oh of Imo(Oll (bl plob of b o ( 0 mo(2€) mo(W1, (4b l o w p a
ofbfmn/25€5+.
Notes.
,
1. Incidentally, thh prow that the statement in Remark 3 on p. 983 of D a u W e s (1988b)is m n g , I made a mistake in the extraction of a numerical value for p from Meyer's proof.
2. Thia ie ementially what was done in the Appendix of Daubechiee (1988b). Beware of the typos in that Appendix, however! 3. If the restriction that F is a compactly supported L1-frmction is removed, then many other solutions are possible. On the other hand* if we insist on compact support and F E L1, then necessarily ~ f =, PI1 ~ for some m € N, and F ia the mth derivative of the compactly supported L1-eblution to the equation obtained after replacing the by 2-m~k;n o - d y is . laet by the restriction ck = 2. For proofs, see Daubchiea and Lagarias (1991).
4. In all the examples we will consider, it is not really neceasery p choose the particular Fo conatructed later: the ,aIgorithm would work with any Fo with integral one.
5. We implicitly assume that the c k are real. Everything still carries through' for complex c k , but then v(x) E cK. 6. Equipping E l in this special case with the norm 11 1(a,-a - b, b)Hi2 = a2 bZ (equivalent with the standard Euclidian norm on E l ) , one finds SUPUEE,IIITo=III/IIbIII != . ~ ~ ~ , S \ ~ PIflTlulll/lllulll UEE, -859, 60 that Hfvj+l(x)- V ~ ( Z ) ~ ~5I Ajlllvl(x) - ~ o ( ~ ) lwith l J , A = .859, by (7.2.11). Et immediately follows that
+
J
Il%(z)ll
Ilvj(z)ll
+ CZi l v k ( ~-) ~ k - l ( ~ ) i l I k=l
5.
-
l l ~ o ~ ~ (1 ~ l l 4+ -
f
I
C l I l ~ l t-~ %iz)lll )
9
s
which b uqifarmly bounded in z and j. 7. The following argument a h givea a direct proof. Suppose that 2-b+') 5 p - x 5 2-3. 'Then there exists e E N eo that one of the two following alternatives holds: (e - 1)2-j 5 z y 5 l2-1 or (I - 1)2'j x 5 C2-J 5 y < (C 1)2-J, We will only diiuae the eecaad w;the first is similar. We then have . -
+
<
<
by (7.2.17). Because of the choice of C, there exists k E N so that zf -= k and f2-j = C2-J - k are both in [0, 1). We cab moreover 'choose binary expansions for zf and t 2 - j with coinciding first j digita (chcme
3:
-
249
MORE ABOUT COMPACTLY SUPPOWED W A V E L m
the expaneion ending in ones for t'2'j, and if x' is dyadic, the expansion ending in zercm for x'). It foUowe that
where we have used ITd, - - Tk,Ian < C 2-"3, the boundof vo, and %(u) ~ ( u ' E) El for all u,u'. We can eimilarly bound If, (y) ff(@-$)I; putting it all together lead^ to
-
-
- CNIz- *la , If( 4 - f (v)l< C 2-& < which proves the Holder ~~ptiauity with exponent a. 8. This corrects a mistake in the first priding, where a too large value was given for the Holder exponent for N = 2. I thank L. Villemoea and C. H d for pointing out this mistake to me. Incidentally, the N = 2 example is an instance where the best H b l e X in (7.2.21) is strictly larger thaa max(~(T0i~l)v P(TI~E,I )n] .t b cue P ( T O ~ E I=) ~ ( ~ 1 = ~ 1#, )a d Q I P ( ~ o ~ ~ a )czJ 1.09946.. 1 ~ l s . > 1. 0. Of course Beylkin, Coifman, and Bokhlin (1991) contains much more than thh! For a large class of matrices it turns oat t U after an o r t h o n o d b d i tranefm using wavelets, dense N x N matrices reduce,to sparse
atructurea with only O ( N ) entriae larger than the threshold c. The total L2-error made by throwing away all the entrieo below c turns out to be O(t), whicb b 8 much deeper rmult than the "comprdhn explained here, it is maentially the T(1) theorem of h v i d and JmmC, the proof of which u& "hardn analysis.
-
10. The argument below also holds for tbs biorthogonal ar (me Chapter 81, where the flatness of l r r ~ at l = 0 and at e = n need not be the same; it @ is the multiplicity of the zero at = 7c that comt4.
.d
4
11. In Cohen and Johnston (1992), 6ltere et% constructed which optimize rxitwb that are a mixture of "atsadu
4
I.
%
-
U 1%
CHAPTER
8
Symmetry for Compactly Supported Wavelet Bases
All the examples we have s e n so fat of compactly supported orthonormal wavelet bases ate compicuously non-symmetric, in contrast to the infinitely s u p p r t d wavelet bases we saw before, such as the Meyer and Battle-Lerniwik bases. In this chapter we discuss why this asyxnmetry occurs, what can be done about it, and whether anything should be done about it.
.
8.1.
Absence of symmetry for compactly supported orthonorma) wavetets.
In Chapter 5 we already saw that a m u l t ~ ~ l u t i oanalysis n does not determane 4, uniquely. Thii ib again borne out by the following lemma. LEMMA 8.1.1. If fn(x) = f ( -n) ~ and gn(x) = g(s -n), n E 2,mtadute d o n o n n a l b e $ of the same mbapoce E of L2(R),then there ezcstg a 21rd c f L n r h a(€), 14€)1= 1, so that g(4) = u(€)f(€), $J
Pmf-
-
C,
the f, are an o r t h o n o d basb for E 3 8, g = anI,,, with C, 1a,t3 = 11911~= 1. -IY, j(t) = a(€)f(t), with a(()=
1. Sin-
z,,
2. As shown in Chapter 5: orthonormality df the f (- - nj ie equivalent with 2m)12 = (%)-I a.e. Similarly I#(€2m)12= (2r)-I. It f o h m that (a(€)) = 1. r
C,
x, -
-
However, we idso have the following lemma LEMMA8.1.2. If (a,),Z u a finite dequence (dl but finitely muny a, eqwi O), and if [a(()l=i, then a, = a6,,, fw some no E Z,
.%. 1. Since la(€)12 =. 1,
x,,
~n
G = &,o.
(8.1.1)
2. Da&zenl,na so that a,, # O + a n , , ~ d a , ,=Oifnnz. 261
',
x,
3. By'(8.1.1), a m On+n,-nr = ,tnr-nl,o- But, by the definition of nl,na, the mua coarEdste of the single term a n , K , which is nonzero by definition. Hence nl = nl.
Together,'these.two lemmas imply that unnpactiy supported 4, d, are unique, for a g i v e n ~ m u l t ~ l u t i oanalysis, n up %oa shift. COROLLARY 8.1.3. If. f , g a n 60th compact&, supported, ' ad the fn = f (. - n ) , gn = 9(- - n), n E Z am both ortlrononncrl bases for Jh&same space E, then g { z ) = af (s- no) fm a m a E C, .la1. = 1 and no E 2hw,) By Lemma 8.1.1, a(<) = a ( ( ) j ( ( ) , with a, = j d z g(x)f(z - n). Bemuse f ,g have compact suppart, only finitely many a, # 0. Consequently,
fn pcuficuler, if 41, & are both compactly supported, and Who. normalizedn' ecaiing fundions for the p m e .multitePaolution analysis, then & is a shifted d o n of dl: the constant a is necessarily 1, because by convention dl: h(z)= 1 * j & q51(s) (see Chapter 5). This uniquenegs result can be. used to prove that, except for the,Haarb i r r ,dl red orthonormal wuvelet bases with compact support are asymmetric. THWREM 8.1.4. S u p s e that 4 and $, the eorrling function and wavctd associated with a multtmsolutwn analgsia, aw both real and sulrported If$ has either a s y u n e t r y or an ant-metry axis, then q5 is the Haar findion.
4 m that
-
= f, & @(a) #(x n) = 0 fin n < ,O, h, # 0. Since 4 is real, eo are the la,,. Let N be the-1 index for'which h, does not v d d x hn # 0, h, = 0 for n > N. Then N i a odd, becawe N even, N = 2m together with
1. We can always shift
.
2. S i n c d h = 0 ~ n < 0 , n > N , s u p g o l t ~ = [ 0 , ~ , b y ~ 6 . 2 2 ? T .b s I], where standard definition (5.1.34) then leads to suppad # = [-no, .no= The symmetry axis is therefore nacessarily at us heve either *: $ ( l - 2 ) = +(x) or $11 - 2 ) = -ql(x).
v.
4;
+
which meaw that the Wj-epacee ste invariant under the map s I+ Sic@ Vj = @ W*, 5 ie invarisht as well. k>j
-2.
S Y M M E n r Y FOR COmACTLY 8UPPOlWED WAVELET BAS=
4.
253
-
Dz?Bne now &(z)= #(N- z). Then the &- n) m a t e an orthmmmal basis of Yo(since Vo is h a r h t Ior z cr -XI, & &z) = $ & b(z) = 1, and eupport 4 = cruppoft 4. It Mows trOm Corollary 8.1.3 that 6 = 0, i:e., 4(N - z) = $(x). Consequently, h.
=
fi /&fl*)+(*-n)
=
fildz+(~-+)~~-iz+r.,'
5. On the other hand,
m
m
(use (8.1.2) on the eedond term)
By Lemma8.1.2, this implies hl, = &,,,-a for some W L 2,la1 = 2-'j2. Since we wumed ho # 0, this means t W h2, = dm,o a. By (8.t.2), hN = ho = a as well, and hi,+l = a 5 % . in general. The normalization Ch, = & (see Chapter 5 ) fix88 the value ofa,a = .
'
3
6. We have thus hzm =
-& 6m,o, hhn+l = $5 km,~
( t= h) ( l + e - i N t ) .
It f b k that d(<) = ( 2 ~ ) - ' I 2((I - e-we)/iN<), or 4(2) = N'l for 0 5 z 5 N,#(z)= 0 otherwise. If N = 1, then this gives exactly the Haat basis; if N > 1, then the q5(. - n) are not dhonormsl, which cbntradicts the aseumpti~lein the theorem. a
E
1 E*
r-t .
k
.
1. The nonexistence of sgmmetik or slbbpumtric ral compactly m p ~ d wavelets e h d d be no surprise to aOybody fam$h with subbaad coding: it haf & e d y bem noted by Smith and BMwell (1986) thw eymmetry i~ not mmjmtible with the the exact mumatmction property in subband atering, Tbe only extra result of Tbaran 8.1.4 is that symmbn 3 n d y implies symmetry for the h,,, but that is a rather intuitively true r d t anyway. 2. If the restriction that ) be real is W, them qmme!try i.possible, even if 4 is compsctly supported (Lawtan, private cammunication, 1990). o
254
CHAPTER 8
The asymmetry of all the ex+mpb plotted in $6.4 is therefore unavoidable. But why shauld we care? Symmetry is nice, but can't we do without? For some applications it does not r e d y matter at all. The nurnericd an* applications in Beylkin, ~oifman,'andRa$hlin (1991), for instance, work very well with very asymmetric wavelets. For other applications, the asymmetry can be a nuisance. In image coding, for example, quantization errors will often be most prominent around d g e s in the images; it is a property of our visual system that we are more tolerant of symmetric errors than asymmetric ones. In other words, less asymmetry would result in greater compressibility for the same perceptual errorV3 Moreover, symmetric filters make it easier to deal with the bomdapies of the image (see also Chapter lo), another reason why the subbsnd cdding engineering literature often sticks to symmetry. The following subeectiane discuss what we can do to make orthohormd wavelets less asymmetric, olr how we can recover symmetry if we give up ~rthonormalit~. 8.1.1. Closer to linear phase. Symmetric filters are ofiten called linear phase filters by engineers; if a filter is not symmetric, then its deviation from symmetry is judged by how much its phase deviates from a iinear function. More precisely, a filter with filter ooefficients a, is called linear plrcue if the phese of the function a(t) = Ena+-'* is a linear function of i.e., if, for some L E 2,
c,
This means that the a,, are symmetric around f, a,, = az-,. Note that according to this definition: the Haar filter no([) = (1 + e-4)/2 is not liniear phase, although the filter coefficientsare clearly~symm~tric. This is because the h:are symmetric around $ 2 in this '*
4
The phase has a discontinuity st r,vlke lmol = 0. If we extend the definition of linear phase to include also the slters for which the phase d a(() is piecewisq linear, with constant slope, and has discontinuities only where la({)l ism, then filters with the same symmetry as the Haar filter are also i n c l d To make a filter "close" to symmetric, the idea is then to juggle with its phase so that it is "almostn linear. Let ~ls .apply this to the "standard* construction of the Nd N@, 88 given ins6.4. In that errse we have
and the coefficients hn were determined by taking the "square rootn of PN via spedr factorization. Typicdly tbis means writing the polynomial L z ,defined by L e ) = pN (sin2t/2), 8s a product of (z Q ) ( Z- z ~ ) (-zZ; )(z - ~ y l or ( z - rr)( z - r;'), where y, rt sre the mplex, respectively, real roots of L, and selecting one jair (zt, 'ii)o\b aB each quadruple of complex mob, and oae value rt out of each pair of real roo@. Up to normalization, the resulting rno is
ca:
-
1'
)
255
SYMMETRY FOR COMPAGLZY SUPPoKIgD WAVELET BASES
then
~ m o ( E )=
(
1
+ e"C
)
N *
n
- d(e-* - %)
(e-*
t
n
(e-.
- .k)
-
k
*
The phase of ~ m cen o therefore be computed from the phase of each contribution. Since (e-4
- Rc e"ioc)(e-'C - I&
and (e"€
ebe) = e-g(e-4
- t t ) izs
the corre~pondingph contributi= arcts
and
k ( 4 ) ==ctg
(e-%/2
e,
cire
-3- 1
tlJ
USa!
+ R! tic)
-
- ,.[ 8 € / 2 ) ,
- 1) sin(
(R; ((i+ ii;)
(-rt
- 2Rr
)
- 2Rc mat
f) -
Let us choose the valuation of arctg ao thst & ia continuous in [0,2x], and #((0) = 0; as shrrwn by the example of the: Haar basis, this may not be the "true" phase: we have ironed out possible discontinuities Tome how llnear the phase is, this ironing out is exactly what we want to do, however. Moreover, we would like to extract only the nonlinear part of we therefore define
*(;
at(€)= at(<)
-*
*
--,6 * r ( 2 4 .
In $6.4 we sptmdically &oae all the zc, rc wi6abeolute d u e less than 1 when we COn8t~CtedNqk T h is a so-called ud phase" choice; it results in a total cut(€) = Cr *c(f) which is -vw'y nonlinear (see Figure 8.1). In order to obtain WIQ as cloee t o linear phase as possible, we have to choose the zeroe to retain from every quadruplet or din sucb a way that \Irwt(() is as close to zero as *ible. In practice, we have 21N/21 choices. This number can be reduced by -her factor of 2: for mry choii, the complementary choice f c h d n g aU the other zeros) lea& to tbe amp1ex conjugate ma (up to a phase shift), and therefore to the mirror image of 9. For N = 2 or 3, there is therefore &ectively only one pair @N, +N. For N 2 4, one can cornpate the 2LNI2J-' different graphs for .Jrht in order to find the doeest to linear phsse. The net r&st of a change of choice from y , 4 to 2.;' Z;' will be most significant if Rt focloar?to 1, and if at is cloee t o either 0 ot x. b F ' i 8.1 we show the graphs kr @$!for N = 4, 6, 8, 10, both for the original construction in $6.4, and for the case with flattest Qkt. Incidentally, in all cases the original constnlction corresponded to the lea& flat Qbt, i.e., ?athe m a t asyminetric 4. The "least +mymmetriP4 and associated with the flattest possible awt,were plotted in Figure 6.4 for N = 4, 6, 8, 10; the comespoxding Ntcr c d c i e n t s were given in %b1e 6.3, for all N from 4 to 10.
+,
FIG 8 1 The non-ltnmr ptwi of the phase of mo(t) for N = 4,6,8, and 10, (lorgut amp&&&) and for the *elor& Lo lighore* choue for the ezfrremal phwe eat
curve).
1. In this discussioq we have restricted ourselves to the case where and I.CIZ are given by (6.1.10)and (6.1.12), respectively. This means that the in Figure 6.4 are the least asymmetric possible, given that N moments of $ are zero, and that $ has support width 2N - 1. (Thiaie the minirnum width for N vanishing moments.) If 4 may have larger support width, then it can be made even more symmetric. These wider 801utiont3 correspond to a choice R f 0 in (6.1.11). The functions 4 in the next a u W i o n , for instance, are more symmetric than those in Figure 6.4, but have larger support width. 2. One can achieve even more symmetry by going a little beyond the "standard" multiresolution scheam exflained in Chapter 5. Suppose h,, are the coefficients associated to a Ystandard" multiresolutii d y d s and the corresponding orthonorxnal basis (compactly supported or 11~t).Define functio-
4'9 b2,Ilrl,
$j2by
(bl (z) =
1 C (-1)" Jz,
h-.+I
~~( nf a ,.
SYMMETRY FOR COMPACmY 8UPPORI"ED WAVEILm BA8E9
257
m e cdculations as in Cb#& 5 show that the functions q!~.!&(~) = ) - ~ + ~ ( 2 - ~ 1k), ~ $? 2jfl.k tZ) = 2-j-112$2(2A2j-1~ k) (j,k E 2)constitute an orthoao,rmal bas& for L2(R). S i the recursions above correspond to
Then the
-
nz,
+
-
4' esn be expect& to be el-
to linek phase than that of = & ( ~ - J O ~- o t *e * t w = dl(t), I&([) = d l ( € ) ; hence h(z) = C#J~(-Z),$t2(2) = *(-z). Figure 8.2 show 41, $1 computed from the h. for N = 2, i.r,ho = hl = h2 = h3 = (Unlike the previous comtmction, this "switchingwmakes a difference even for N = 2.) For the geest asymmetricn h, given in Table 6.3, this switching technique leads to slightly "bettern4, but seems o to have little effect on the phase of
A(€)
9,s,s.
9.
6.
1
-2
-1
0
FIG. 8.2. S* function 61 and w a d e l tothe4-taptmtodctfiUer~0f~.4.
1
&oinaf
2
bv epplying the '8tuitching h.ickW
258
CHAPTER 8
In 47.4 we saw one advantage uf having a high number of vanishing moments for $: jt led to high compr-lity because the fine d e wavelet coefticients 'of a function would be essentially zero where the function was smooth. Since J dz $(x) F 1, the same thing can never happen for the (f, #3,k). Still, if $ xt O(z) = 0 for t? 1, L,then we rad apply the same Taylor expansion . argument snd conclude that for J large, ( f , + - J , k ) 2D ' f (2-'k), with an error that is negligibly s d where f is smooth. This rnthat we have a particularly simple quadrature rule to go from the samples of / to its h e scale coefficients ( f , qLJ,,), b r . this reason, R. Coifman suggested in the spring of 19s9 that it might be worthwbile to construct orthonormal waveIet bases with vanishingamoments not only for q9, but also for 4.* In this section I give a brief account of how this can be done; more detaifs are gjven in Dmbechie%(1990b). Because they were Erst requested by Coifme (with a view to applying them for the algorithms in Beylkin, Cdfman, and Rakblin), I have n+ Qe resulting wavel@s "coi%ets. The goal is t o find 4 so that
- ..
+
-
--
+;
and / d z + ( z ) = ~ , Jdt=t+)=o,
L is then called tg &r of the &t. in terms of q; it is equivsld with
~=I,--.,A-I;
We &eady b o w how to exp-
What does (8.2.2) correspond to? It is equivalent to the condition Let us & whah $(o) = o meam for momo(€/2) 6f4/2), we h,ve
e = I , . . . , L - 1.
Consequently, J dZ x+(z)
that (8.2.2) is e q u i v d d
(8.2.2)
(8.2.1)
= 0,
Ft$
dI -lEi(l Because
d(e) =
-
witb &(o) = 0. Simib~ly,one MXS
o,&= l , . . - , ~l-, = n t h
1
259
SVMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
where 2 is a trigonometric polynomial. In addition to (8.2.3) and (8.2.4), mo will of course also have to satisfy 1~ i%(( r )la = 1. Let us specialize to L even (the easiest case, although odd L are not much harder), L = 2K. Then (8.2.3), (8.2.4) imply that we have to find twP trigonomtric polpornids PI, P2 uo that
2I)[(
+
+
K
(8.2.5)
But we already know what the generd fogm of such Pl , Pzye: (82.5) is n o t h i i other than the Bezout equation which we dready solved in p.1. fn particular, PIhas the form
where f is an arbitrary trigonometric polynomial. It then remains to taylor j in ~ ( 6 = ) ((1 e - * ) / ~ ) PI ~ ~(t) so tbat Imo(()l2 I%(€ +..lr)12 = 1 is satisfied. With the ansatz f(t) = 2K-1 fn e-'4, it is shown in Daukhies (1990) h w to reduce this "tayloring" to the solution of a system of K quadratic equati011~for K unknowns. A heuristic, perturbative argument suggests that this sygtem will have s solution for large K, and explicit numerid solutions are computed for K = 1,. . - ,5. Figure 8.3 sbowe the plots of the resulting 4, the -riding ooefficients are listed in M l e 8.1. It is clear from the figure that 4, $ are much more symmetric than the N#, N+ of 56.4, or ewen than the 6,$ in $8.2, but tbare is of course s price to pay: a-coiflet w i t h 2K v m i d b g momenta typically h& support width 6K 1, aa compared to 4K - 1 far ~ 4 .
+
+
+;
-
REMARIC.The m t z f (t) =
, ' I
2K-1 fn eli* xnTO
is not the only paeaible one, but it makes the computations easier. For smdl d u e s of K (4= 1,2,3), Merent w t z e s ste also tried out in Daubecbiee (1999b). It turns out that the smoothest coiflets (at least at these nmnll values for K) are not the moet symmetric ones; for K = 1, for instance, there exists a (very agymmetrie) coifiet with H6lder exponent 1.191814, whereas the coiflet of order 2 in Figure 8.3 ie ~ @C1; t both have support width 5. S h i k g a b d mjphity can be found for K = 2,3. For g n q b , d c i e n t s and mom details, aee Daubechiss (1990b). o 8.3.
Symmetric Morthogonal wavelet bum.
'
- eymmetry As mentioned above,'it is well known in the subband ~~ community that and exact recrmstruction are incompatible, if the same FIR filters are \liied h r e c o and~ dwmpositima ~ As soon as this laat.requitement is @ ~ ~ l l u p , s y m m e t r y i s ~ bThismrwrrthatwrepbcetheb~diagramof 1e.
/
/
CHAPTER 8
I
262
CHAPTER 8
mqm in terms of multiresolution analysis? What do d and dJ now stand for? (They were coefficients of orthogonal p j e c t i o n s in Chapter 5.) Is there an associated wavelet basis? How does it differ from the constructed earlier? The answer is that, provided the filters satisfy certain technical conditions, such a scheme corresponds to two dud wavelet bases, associated with two different multiresolution ladders. In this section we will see how to prove all this, and give several families of (syminetric!) examples. Except for an improved argument due to Cohen and Daubechies (1992), all these results are from Cohen;Daubechies, and Feauveau (1992). Many of the same examples are cJso derived independently in Vetterli end Herlqy (1990), who present a treatment from the =filter designn point of yieor.
FIG 8 4 Subband fiUcnng adreme wth e m d nem~ltructconBut ncbrutrucaon filters dcffsnntfmm the decompoation f i b s .
8.5.1.
Exact reconstruction. Since we have now four filters instead oft&, '
we have to rewrite (5.&.5), (5.6.6) as
and
In the z-notatron introduced in 55.6, tbis caa be reornitten as
'
+2
[-h(.)
sy-1)
+ j ( z ) W-z)]
#(-z)
.
Consequent)y,we require
ha)
-
+ !?(4S(z!
;.r
2
,
(8.3.2)
g(-z) = 0 , (8.3.2) where w sanune i;,d,iE,~to be p ~ t y y m b bsinoe the 61ters are all l?m (Fbr simplicity, rn wie the term "polynmhd" fa a dightly wider sense t b usual: & { x ) &(-a) i-#(z)
SYMMETRY FOR COMPACTLY SUPPOmED WAVELET BASES
zN:-Nx
we also allow negative powem. h other words. a,, zn is a p01yaoW in this terminology.) From (8.3.1) it fdlmw that & and have no common ma; consequently, (8.3.2) implies that
-
k(z) = -$(-~P(x)
= h(-z)dz),
for some polynomial p. Substitution into (8.3.1) l
(8.3.3)
d to
The ,only polynomials that divide amstants ara rnonomblq hence
for aome cr E C, k E 2,and (8.3.3) baco-&~~
Any choice for a and k will do; we choose a = 1, k = 1, which ma& equations (8.3.4) for g and g symmetric.Substitution into (8.3.1) gives
the
In terms of the filter coefficients, all this becomes
-
where we have implicitly assumed that dl the coe@ciits are real. These equations are obvious generalizations of (5.1.39), (5.134).
8.3.2. Scaling functions and wavelets. Because we have two pairs of tllfunction + wawbt: 4, $J and J,$. They
ters, we also have two pairs of &g are &fined by
where %(<) = $3 h.e-'4, ml(() = a n ~ o u 9 1 y .Note that (8.3.7) implie6
Cn
ml(<) = e'Y
h(< +r ) ,
$ICR e-'*; ~, 6 us defined
mt(€) = e*
&J([
+x) .
(8.3.10)
~ e m in w chapter 3 that in order togmtm@waveletF&az b m ~ ,
and$ have to dstEsfy $(o] = 0 =$@). A wweessry condition is thedam mg(@ 3 0 = 2
-
ex),
ml@); in terms of the polynomials h ( a ) this L equivalent t o h(- 1) = 0 = &(- 1). ~ t h t i t u t i o ninto ' (8.3.5) t h e ~ . l W b h(1) h(1) = 2,or
-
This impliee that w& can normdize both h and h so that 6, h, d! = C, h,,- Consequently, ~ ~ ( =01 = ) h ( O ) , and we can solve (8.3.8) by defining
The same arguments as in Chapter 6 show that these infhrite products converge uniformly on compact sets, md that Q and & have compact suppart, with s u p port width b;iven by the filter lengths. As fihite linear mmbinatione of 4 and 4, $ and also have compact support. This is by no means sufficient t o guarantee that the @,,k = 2 - ~ $42-ix / ~ - k ) and q I , k are dual RieQ& bases of wavelets, however. Indeed, even in the orthogonal case (reconstruction filtem = decotnp sition filters), it wss possibIe for $ to fail t~ generate aa orthonormal basi~(see 56.2, $6.3). In this nonorthogond case we have t o be even more careful. Let us summarize the different steps in the argument proving that we have dual wavelet bases (with certain restrictions]. First of all, if 4, 6 E L ~ ( R )(which wili have t o be proved tob! See below.) then we can define bounded operators Tjby
4
where
aa ustHrl.' A consequence of the definitions (8.3.8), (8.3.9) is
w h c r with the propetties of the 6lter eueflkients impoeed in $8.1, this implies (as can d y be checked by substitution)
. -
,
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASES
265
+
Thesame trick can be applied for other values of j; Uteleecoping" eU tb;e identities together lesds to \e
Exactly the same arguments as in Chapter 5, ueed in the estimatitm of (5.3.9), (5.3.13), re~p&i~ely,&OW t b t (Tjf,g)-4 0, ( T - j f , g ) d ( f , g ) , for 3 4 ~ . Coneeqmtly,
or, in a weak sense,
-
constitute dual Eesz bases. This is not sufficient to establish that the Jl,,c, or 3 , mry ~ f d ta-co~bi(yteframes; in thb ua, the For one thiDg, tiy convergam in (8.3.11) could depend cnrdarlly on the order of sum-. To 8 d
this, wet &db.iXlpOthat se
C
converge for all f E L2(@),or equivatently,
If t
,
b upper b o d hold, then it follow8 fnnn (8.3.11) that6
- so t h a we automatWy h.ve frames. But evm thehen the $j,k may merely be (red-) dud frank and not dud Rieez b, thie redundancy ie eliminated
by t h e m q w t
6j#,w)=
,
(8.3.13) which, t m d y IiJm in tbe o r t h o d cam (see g.2), can be ehown to be equivalent with (8.3.14) (&I. r , 6,v) = h r v . t
($~,k,
6j.9 6h.h.
If the conditions (8.3.12) arnd (8.3.14) are satisfied (we will come back to them shortly), then we do indeed have two multiresolution analpis ladders.
with
Ki =
Span (&,k;'k
E Z),
%
= Span {&o,~;k E Z).
The spice8
are again complemeata3 W, = Sp.o { Q,,r ; k E Z).W, = in V,,, , re8pecti~ely, - but &heywemk orthogonal complements: typically the mgle7 between V,, W,.or wjwill be smaller than 90'. This is the reason why we have to prove (8.3.12) idthis case, w k e a s i&was automatic in the orthonormal case. Another way of seeing this is the following. Because of the uon-orthogonality we have
,,
af 5,respectively,
c,
with a < 1, 0 > 1 (in the orthogonal case, equality holds, with a = @ = 1) Wike the orthonormal case, we catln~tt e b p e these inequalities to prove tbst the lb\3,& cordtute a Riesz basfrs: - t e l ~ i n wouId g fed to a blowup d the wmstmta: We therefore have to f d b a different strategy. Note that (8.3.13) i m p k that WJ I wj I $. The two rnultireaolution hierarchies and their sequences of con~plementspaces fit togetber like a giant zipper, and this is what allows us to control expressions like I( f,$ , , k ) ( 2 . But let us return to the conditions (8.3.12) and (8.3.14). We already saw how to tackle condition (8.3.14) in $6.3, in the simpler orthogonal case. Our strategy here is ementially the same. We again define an operator Po acting on
6,
x,:
21r-periodic functions,
POis defined analogwsly. In terms of the Fourier coefficients of f, the action of Po is given by
8 second operator
I
we wi¶lbe mostly iuterested in i~~variant trigonometric polpomiale for Po. This xlmsy that we can lestcict our attattion to the 2(N2- Nl)f l-dimensional subapace of f for which ff = 0 if C > N2 - N1 (we assuxlle h, a 0 if n < Nl or n > Na), on which Po ie repmxmted by s matrix, Theorems 6.3.1 and 6.3.4 have the foUanalog. THEOREM 8.3.1. The followng t h m statements an: equivalent:
SYMMETRY FOR COMPACTLY SUPPOEn'ED WAVELET BASES
267
1
2. il7rerrc &t strictly paittue trigonometric pdpnnials fo, JO invanant for Po, PO; there also exists a compact set K c o n p e n t to [-a,r]modulo 2n so that
3. Then? exist stnet18 poaittue tpgmometric pdynomiala fo, fo invariant f m Po, PO,and these am the 4 9 inwridnt polgwmials for Po, PO(up to nonrzalizatton).
The proof is very similar to the prooh in Chapter&, but a bit mGre compliIn $6.3, the functions fo, fo were simply constant;in this crtse, they are
cated.
+
A
lo([) d(( +
essentially fo(Z) = Cc 27tk')l2, = XI 2d)12. For details on how to adapt the proofs of $6.3to the preaent case,see Cohen, Dauhechies, and Feauveau (1992). Condition (8.3.14) therefore simply amounts to checking that two matrices have a nodegenerate eigenvalue 1 and that the entries of the corresponding eigenvectors define a strictly positive trigonometric polynomial. (Note that if the trigonometric polynomial takes negative values. then 4 $! L2(R). This happens for some exact reconstruction filter quadruplets.) Condition (8.3.12) is something we had not encmunte~edin the orthogonal case. It tums out that this condition is satisjkl if any of tbe three conditiw ia Theorem 8.3.2 holds. The proof of thi8 surprieing fact is in the following step#: First, one shows that the existem of an eigenvalue X of Po wi'th (A( 1 1, A # I would wntradict the square bitegmbility of 4. It follows therefore from Theorem 8.3.1 that dl the other &endues of Po have b lute value strictly smaller than 1 if the eigenvalue 1is nondegemerate md the associated eigavector corresponds to a strictly pmitivc trigonometric polynomial. The probf of this step uses I ~ m m a7.1.10. = 0 = &(n), we have obviously Mo(?r) = lnro(lr)12 = 0 = Ifho(?r) = M~(R).We saw in Chapter 7 that this means that the culurnns of the matrix representing PC all sum to 1, so that the row vector (of the
since - ( T )
I?:
appmpriate dimemion) with all ent,riee 1 is a left eigenvector for Po with eigenvalue 1. It follows from the first point that p, the spectral radius of POIE,, with El = { f; fn = 0), is stiictly rmdier than I. One then uees that f = 1 -COB ( is in El to pmve (*be estimatee are anslogoua to thoee in tne prmf of Thsonm 7.1.12)th.t &I~K 5 cI I '
x,
(c)
1p-1r5.,(1s2m,,
(?In.
Via Hiilk'. inequality this bpk 14 I&[Yc)I~(~-~~ < m for suEcientIy small 6, Thia can then be wed to pnwe a 'discretized" vereion, i.e., I nn)J2(1-6') S C < ap for aU E R, again for suWdently
+
<
small 6'. E3ecau88 ml is bwnded, qb ~#~tisfiee a h d s r bound,
On the other hand, one cad dr#r prwe that sup
C li(rt)lU'< m .
+SICI-<2u ,,z
since 6 is entire and &o) = 0, l&{)l 5 CKI for d c i e n t ~ y8 d I€(, SO that I&z~c)~"'is uniformly bounded for 5 Z r , and we only need to concentrate on j 2 0 in (8.3.16). But
c=-,
The second factor ia finite beesme $Jis compactly supported and in L2(R); the first factor is bounded by CAj, with (A/ < 1, se shown above. This establishes (8.3.M), which is also equivalent to
Finally, a combination of the Poieson summation formula and the ChuchySchwan. inequality leads to
It therefore follows from (8.3.15) and (8.3.16) that
more details concerning this arguntent, see Cohen and Daubechia (1992).
In order to ensure that we ham indeed two dual Rim baeee of
wavelets, rm
therefore only have to check that 1 is a nondegenerate eigenvalue of f i ,Po and that the cornpading trigonometric poIynomid is strictly positive.
sYMME%XYPOR C O M P A a Y SUPPOKi'ER WAYELEX BASES
#)g
8.3.3. EbguhdQ ud vanishing -8. If t b d ~ $*,h. W t u t e . dud Rim baees (of compactly supported wavelets, since we have asmmed the - "fittersto be FIR), then we can apply Theorenn 5.5.1 to link vanishipg momenta of one function with regularity for the other: if $ E C",then automati4 y J & z' d ( x ) = 0 C = 0, - - .,rnW9 This is eqwith for C = 0, .
,
,rn.
Becaase of (8.3.9) and &0) = 1, t& i m p k
k
led
for C = 0, -. ,m. By (8.3.10) this implies that Q is divieiWe by ((1 + e-g)/2Jm. In order to produce mg& 9,we thedore need to constrncb filtet pairs mo, r%o such that ma(() has a d t i j d e zero at = n. Note that nothing p m t s $ and 47, from having vwy d S k rqphrity properties, ss illllstnrted by some of the examplee below. If is mu& more regular then corresponding to many more vanishing moments for J, than for 4, then the two formulas
e
6
+,
both equally valid, have very different interpretatioarr (Tchamitcbian (1987)). In practice, (8.3.17)is much more useful thsn (8.3.18): on the one hand, the larp number of vanishing momenta of $ leads to much more "wmpreasion potential" in the regions where f is reasonably mooth (see 57.4); on the other hand, the uelementary building blocks" qjr are smoother. In Antonini d d. (1990) a. experiment was carried out with biorthogonal wavelets of this type: the eame filter pair wm used twice, the eecond time with roles of demmposition and reconstruction filters exchanged. The case ~0~eBpOnding ta (8.3.17) gave rise to much better d t s after quantiztrtion than (8.3.18). As we already mentioned in 57.4, it is not clear whether the high ,number of vanidhg moments of JI or the regularity of 6 ie the most important factor; it is possible that they are both important. 8.3.4. Symmetry. One advantage of biorthogonal mer orthonormel baeee is that q, rlro cap both be symmetric. if the filter correspondmg to haa an odd number of tape, and is symmetric, i.e. mo(-E) = esNmo((),. then m~ can be written as (8.3.19) m(<)= e-'Y Po(-€) , where f i is a pdynomial. it then f o l l m
thst & can be
of the eame
form,
tiio(€)
= e-'h(=€),
where fi ie any polynomid that aatkrfiw
(8.3.20)
.
we then have indeed
..
-
d o fh(0+ W J ( ( + ~ %(€+a) )
(8.3.22) = 1, which is tbe same as (8.3.5). F'olynomilrle & eolving (8.3.21) can only be found if po(z) and po(-z) have no common arm; once this is the case, there always
+
&st dutions by Bezout's theorem (see 86.1). Note that this also means that biorthogonal 'bases are much easier to construct than orthonormal bases: we need to ulve only linear equatione to find & sath&bg (83,21)once po is fixed, instead of.the spectral factorization needed in 56.1. ff the filter cwresponding to m,hae an even aumb4p of taps and is symmet& (such ss, eg.,the Haar W r ) , then mo -tidies no(-@= eLiYSjf w(€);
One can then again chooee +q, of the same type,
ty&tion (8.3.22) becomes
which meane that & wlvea tbe Beu>ut problem
m+ pgY{-~)
P~(z) with
___i_
h(-2)=
I
,
d(x)= 9 po(r).
- ,
EXAMPLES. All fbe extunplea we give here have both symmetry and some regularity. The trigonometric polynomials mf~and & 8t.e therefore of type (8.3.19), (8.3.20) or (8.3.23), (8.3.24), with po(m(),&(cart)" divisibb by . (1 e-%)' for m e C > 0. Since we are dealing with pal~cw,mi& in cost, P win 8~tomIdyb -; (1 + C Y ) ~ = r - 4 d i = 2e+[l+ws<). of type Consequently, m are looking for m,
+
if they ham an even number of tape (we have aseumed tbst k = 0, i.e., that the h,,, & are symmetric wound 01, or of type
of tym ie odd (agdn we have t h k = 0, axresponding to if the hl-, = h,,, ha-, = h,,). lo both awe, mWtution into (8.3.D)gives
SYMMETKY FOR COMPACTLY SUPPOKI'ED WAVELET BASES
,
r;zc)
271
+
= L + 2 in the first case, L = t + 1 in the second caee. If we define &(ma() = P (sin2 then (8.3.15) reduc- to
1).
an equation we already encountered in 56.1. The solutions to (8.3.26) are dl given by
where R ia an odd pofynomid (see Proposition 6.1.2). We now present three families of examples, based on dierent choicea for R and different factorieationa of P into qo and qB. Spline examples. Here we take R z 0, and &, = 1. it follows that = L - ~ ( / ~(UM R = d + I, SO that h(() = (CCM A = 2 i , or 8 ia a B-spline centered amundo, respectively ?. in the first caee, we then have, with N = 26,
i)",
in the second caee, with N = 26
+ 1,
In both cseee we can choose I fieely, subject to the mnstraint that the eigenvalue 1 of Po is nmdegeneiate and that the &dated eigenvector corresponds to a strictly poaitive trigonometric polynomial (see 58.4.2). The result is a family of biorthogonal bBees in which-4 is a spline function of compact mipport; for every preassigned order of this spline function (i.e., fixed I), there exists an infinity of choices for C, corresponding to different$,I (with increasing support widths) and different 4, with increasing number of vanishing moments. Note that is completely fixed by N dons, while nq,,hence 4 depe~rdson both N,[email protected] have plotted the functions fi,N$, and ly,,&, for the firat few values of N,N,in Figures 8.5-8.7 (fi 1 in F i 8.5, fi = 2 in Figure 8.6, fi = 3 in Figure 8.7); the &rresponding filters are given in Table 8.2. In all tbeee cases, the conditions derived in 98.4.2 are satidid. A striking feature in Figurea 8.58.7 is that from eome point on, increasing N (for fixed fi) does not alter the shape of 10 ,$; one txm the "wrbilclesn in the mmsponding d , N # and IV,N$ get ironed hut as N incre-.
6
-
I
; k i
/
,
CHAPTER 8
FIG. 8.5. Sp&ne e24mpb oith N = 1, N 8 and 6. Por N = 1 (not plot@ one H w borir. Ons her 'ruqlport 1 . ~ 4"'1-N 1,Nl#rtcppor2 1 . d = nrpggr( 1 . ~ ' 4
m
[-v, 91.
+
SYMMETRY FOR COMPACTLY SUPP0En'E:D WAVELET BMm
273
Fla. 8.8. &dim aamplu cuith fi = El N = 2,4,6, and 8. H a ruppwG I-N,N), npgmta,lv+=.~-a,~j= [-!#,$+I]. Aod~y1,tht~lotroj4.3aminlhttpbL.d obtuined by the om8cade algorithm (,app. #)ti2o6Jl uritA 6 or 8 i t c r ~ t i o n r . * O
CHAPTER 8
SYMMETRY FOR COMPACTLY SUPPOKI'ED WAVELET BASES
275
FIG.8.7. Splane ezamplea unth N = 3, N = 3,5,7, and 9. For N = 1 (not plottad), u not s p a n antegmble. Hen ~ p p w 3t . ~ = 9 [-N, N 11, mpport 3 , = w ~ m~ ~ 3t , ~ = 4 The f u n c t a o ~3 3 4 and 3,3+ are c20mpka of the fact that the w c o d e algonthm may daucge whJe the d r d algorithm dill convnpea (ace Note 11 d the end of Chaptar
[-v, v].
6).
+
278
CHAPTER 8
The functions I,s$ and 1 , 3 4 were first constructed in Tchamitchian (1987) as an example of two dual wavelet bases with very different regularity prop erties. Here they constitute the first non-orthonormal example of the family (N = 1 = N gives the Haar basis). As in the orthonorma1 case, arbitrarily high regularity can be attained with these examples, for both tj and As a spline function, ni,N& is piecearise polynomial of degree N - 1 and is at the knots; the regularity of f i V N $ can be assessed with any of the tech€ Cm niques in Chapter 7. Asymptotidly, for large A, one finds that if N > 4.165 N 5.165 (m 1). These spline examples have several remaxkable features. For one thing, all the filter coefficients are dyadic rationals; since division by 2 can be done very fast on a computer,this makes them very suitable for fast computations. Another attractive property is that the functions are known exactly and explicitly for all x, unlike the orthonormal compactly supported wavelets we saw before." One disadvantage they have is that mo and iiLoare very unequal in length, as is apparent from Table 8.2. This is reflected in very different support widths for 4 and & because they are determined by both mo and &, t,b and always have the same support width, given by the average of the filter lengths of mo, %, minus 1. The largedifference in filter lengths for rno, rkc,can be a nuisance in some applications, such as image analysis.
4.
+
+
-
cR-'
6
Examples with l&s disparate filter lengths. Even if we still take R r 0, it is possible to find mo and & with closer filter lengths by choosing an appropriate factorization of p(sin2 into qo(cm <) and &(cos [). For fixed L L there is a limited number of factorizations. One way to find them is to use spectral factorization again: we determine all the zeros (real and pairs of conjugated complex zeros) of P, so that we can write this polynomial as a product of real first and second order polynomials,
5)
+
Regrouping of these factors leads to all the possibilities for qo and h.Table 8.3 gives the coefficients for mo, r i for ~ ~three examples of this kind, for t+f = 4 and 5. (Note that t?+ = 4 is the smallest value for which a non-trivial factorization 2 = 4, the factorization of this type is possible, with qo, &I both real.) For is unique, for 4 + e^ = 5 there are two pwibilities. In both cases we have chosen of mo, 7fro as small as possible. The' l , so as to make the length d8-ce corresponding wavelets and waling functions are given in Figures 8.8 and 8.9. In all cases the conditions of 58-42 are satisfied.
!. +
,Biorthogonal bases close
to an orthonormal
basis. This first example of this f@y was suggested by M. Barlaud, whose mearch group in vision analysis tried out the filtersin 56A,6B for image coding (see Antonini et 4.(1990)). Because of the popularity of the Laplaciq pyramid scheme (Burt and Adelson (lW)),Barlaud wondered whether dual systems of wavelets could 8.3.5.
SYMMETRY FOR COMPACIZY SUPPOIWED WAVELET BASES
ma
The coc@wntr of m ~ , /or
on the spline m e " wath filters of
5 (me td). For each filter we haw olso gaven t +i = 4 (denoted N,fi). As in Tobk 8.2, multiplfing the entries below
srmilar length, a~nspondangto
the number of (cos2) factors with givcs the flter coeflcients h,,,
a
of %a+&ma
279
h,,.
mMMETRY FOR COMPACTLY SUPPO-D
WAVELm BASES
281
~.
These
be constructed, using the Laplacian pyramid fitter aa either mo or Alters are given explicitly by
- a e'w
+ .25e-'c + (.5 $ 2 4 + .&
-
ae2% .
'Ip0.3.27)
- For a
= -1/16, this reduces to the spline filter ,7&, aa described under the a = .05 is e9 "spline examples" above. For applications in Won, the pecially popular: even though the m-w # bas kss rgul*ity than 4 4 , it seems to lead to results that are better &om the point of view of visud perception. Fotlowing Barlaud's suggestion, we choee themfore a = .05 in (8.3.27), or
Candidates for r 3 ~ dual to this ~TQI have to satisfy
csn be chosen to be symmetric (since As s h m in $8.4.4, such ie symmetric); we 8180 opt for & divisible by (coe~ 1 2 (so ) ~that the corresponding 11, both have two zero moments). In other wwds, .r
4
where
By Theorem 6.1: 1, together.with the symmetry bf this equation for subetitution of x by 1 x, thia equation has a unique solution P of degree 2, which is easily found to be
-
One can check that both (8.3.28) and (8.3.29)satisfy all the conditions in 88.4.2. It follows that these m . ~and r%~ do indeed correspond to a pair of biorthogonal
282
CHAPTER 8
4,
wavelet bases. Figure 8.10 shows graphs of the corresponding 4, $Jand 4. All four functions are continuous but not differentiable. it is very striking how similar and ) are, or d and 4. This can be traced back to a similarity of &d-ria~,which is not immediately obvious from (8.3.27) and (8.3.30), but becomes apparent by comparison of the explicit numerical values of the filter coefficients, as in Table 8.4. In fact, both filters are very close to the (necesyarily nonsymmetric) filter corresponding to one of the orthonormal d e t s (see $8.3), which we list again, for comparison, in the third column in Table 8.4. This proximity of nzo to an orthonormal wavelet filter explains why the i h o d u d to WIO is so close to itself. A first application to w e analysis of thest: biorthogonal bases associated to the Laplacian pyramid is given in Antonini et al. (1990)-
4
FlG
8.10. Graphs of
4, $, 4,4 for the
b u n t h o g o d paar conrtrrrctcd from the Burt-
Adebon l o w - p a filter.
M. Barlaud's suggestion led to the aqidental discovery that the Burt filter is very elm to an orthonormal wavelet filter- (One wonders whether this closeness makes the filter so effective in applications?) This example suggested that maybe other biorthogod bases, with symmetric filters and rational filter coefficients, can be constructed by approximating and Usymmetrizing" e x i s t i i orthononnal wavelet fiIters, and computing the cmmsponding dual filter. The coiflet coefficients listed'in $8.3 were obtained via a construction method that naturally led to close to symmetric f i l h ; it ie natural, thergfore, to expect that symmetric biorthogonal filters cloee to im orthonormal ba& will in fact be cloee to these
SYMMETRY FOR COMPACTLY SUPPORTED WAVELET BASE8
b
6 f
283
F J t a u~c&aente for (mo)~,,~~, for the d u d fiuer (*)sryt computed an thy rectwn, and for correipondang to on o r t b m m d bosu of eot&tr ( r e the entncs for k = 1 m Table 8 1) a uecy close filter
t
coiflet bases. The analysis in $8.3 suggests, therefore,
(&: ')
m(<)= (cw ~ 1 2 ) ~ ~
+
(sin (/2)"
I' I
+ O((sin I / Z ) ' ~ )
k= l
:
In the examples below we have chosen in particular
(*-:+ ') (sin (/a* -+ .(sin ( / z ) ~ ~
~ (= 0 (me 6 1 2 ) ~ ~ k=O
and we have then followed the following procedure: 1 Find o such that
dE [I - Jw(()(~- (mo(t +
~)(~]l
ia minima (zero in the examples below). T ~ Ioptimizattan S crit~m can of caurae be replaced by other criteria (e.g., least sum of squares d dl the Fourier d c i e n t s of 1 - Imo(6)l2 - Irno([ r)12 instead of only the cod6cient of e* with t?= 0). For the cases K = 1, 2, 3, the emalleet root for a ir .861001748086,
+
2. R e p b this (irrational) "optimaln d& hr a by a close value expressible ss a simple f~action.'~For our ~aampIceu = .8 = 4/6 w8a chaeen for K--l,a=3.2=16/5forK=2&~*13forK=3. ForK=l,thia reduces then to the example above. 3. Since n ~is, now fixed, we can compute h.If we q u i r e that 6 be elso divisible by (cos [/2)2K, then where PK is a polynomial of degree 3K Daubechies (1990) a h m that
- 1.
The
anal*
~ m e
m in
284
CHAPTER 8
thereby determining already K of the 3K coefficients of PK. The othen, can be computed easily. For K = 2 and 3 we find
In Tabla 8.5 we liat the explicit numerical values of the filter d c i e n t s for w2 h and the cloeest coiflet, for K = 2 and 3. We have graphed , tj,a d tlr for both casts in F i p 8.11. It is wosthwhile to note that the computation of the biortbogonal filters m, A, aa explained by the above procedure, is nuch simpler than the computation in Daubechies (1990) of the orthonormal coiftet filters! This illustrates the greater flexibility of the construction of biorthogonal wavelet b a w versus orthonormal wavelet baees.
4,
NumeMwluuforthefiltmmo,~forbiorULcrponolborerclorebooi&tr,fo+thecarar1
K = 2 and 3 (roe rd). Thc third column lfaU the daficienb of the arihonormd dpCt fiUrr to which mtj and & am wry clone. I n order to wnpam thc di-t c o e m mmc w, w have ~ ~ p n a s cmeqtthing d in decimal notation; tn fact, the o o e f i e n t r of curd & om mtiocPclt.
SYh¶METRY FOR COMPACTLY SUPPOKI'ED WAVELET BASES
285
Notes. 1. In the sense that the
(. - n) are orthonormal, as are the &(.
- n).
2. Strictly speaking, Lemma 6.2.2 only proves support(+) c 10,N]. A recent paper by Lemarie and Malgo~lyres(1991) shows that support($) is necessarily an interval, which in this case then has to be [O, N].
3. Nevertheless, AWARE, Inc. uses the asymmetric filters from $6.4 with excellent results in image and video coding. Note also that 'perceptually" small or large errors are difficult to quantify matbematically; the norm most often used to measure "distance" is the t?-norm, but that is,rpore because this is the easiest norm to handle than for any othcr r F n . All experts agree that the 12-norm is not a good candidate for a "perceptualQ norm, but as far as I know, there is no agreement on a better candidate. 4. The N~ from $6.4 do not have this property. The graph of 1 N#(()Iis very ' d flat near ( = 0,showing that Si' j N$llt=o = 0 for l = 1 . - -N ,but the phase of
N4(()does not share this property.
5. The proof that Ti is a bounded operator is easy: if support 4 = [-Nl, Nz], then
hence
Similarly, one proves that all the Tj are boundad. 6.
We have
287
SYMMETRY FOR COMPACTLY SUPPOEtTED WAVELET BASES
7. The angle between two subspaces is defined as the minimum angle between elements, angle (E,F) =
inf
sEE, jEF
cos-'
I(e,f)l -------~leil llf 11
-
8. The proof in Cohen, Daubechies,-and Feauveau (1992) impoees a much etronger decay condition on namely I&[)\ 5 C(1+ I[j)-'/2-c (which is known not to be satisfied in even sdme orthonormal cases) in order to derive (8.3.15) and (8.3.16). The argument sketched here comes from Cohen and Daubechies (1992).
4,
9. The derivatives @('I, != 0 . .. m, are automatically bonded because $J has compactsupport. 10. In the case N = 2 = N, a curious phenomenon happens. The function 2,2#, although an element of L2((-2,2]) (hence also of L'([-2, 2]), has in fact a singularity at every dyadic rational. The true graph of 2,24 (or' p,2@)would therefore consist of a black rectangle (since our lines have some thickness), but the graph in Fig. 8.6 is nevertheless a close approximation in L2 or L' , although not in Loo.I would like to thank Win Sweldens for pointing this out to me. 11. Auscher (1989) and Chui and Wang (1991) contain another construction of non-orthonormal wavelet bases where one of the two wavelets, say $, is a compactly supported spline function, and is therefore aLso known exactly and explicitly everywhere. In this construction the W,-spaces are orthog* nal, unlike here, and W,= W, . As a result, the dual wavelet has infinite support (compact support for both can only be achieved by giving up the orthogonality of the W,), with erponentiai decay. The associated mdt i r d u t i o n analysis is the same asifor the BattleLemari6 wavelets; $ is c h m n so that it is orthogonal to the B-spline of the rkht order and all its integer translates, and $ is then given by;(<) = 2xk)A
+, 4
4
$(~)/[x, d((+
12. Choosing a rational leads to mo, f& with rational coefficients. Note that there ie nothing sacred about the original irrational values of a: changing criterion in point 1 will lead to slightly merent d u e s of a.
CHAPTER
9
*
Characterization of Functional Spaces by Means of Wavelets
The major message of this chapter is that tbeprthonormai basea we have discussed for the last four chapters also give good (i-e., unconditional) bases for many other spaces than L2,out-performing the Fourier basis functions in this respect. Almost all the material in this chapter ie borrowed from Meyer (1990), but it is here presented in (I believe) a more pedestrian way, accessible t o readers with a lower level of mathenlatical sophistication. (Meyer's book also contains much more on this subject than is explained in this chapter.) In $9.1 I start by reviewing a classic theorem of pure harmonic analysis, the Calder6n-Zygmund decomposition. It can be found atso in many textbooks (such as Stein (1970)); I include a detailed proof here as an illustrstiun of techniques using diierent (dyadic) scales, practiced in pure harmonic d y e i s long before waveleta came along. Together with some other ciasic theoreme, it leads to the proof that wavelets are an unconditional bases for LP, 1 < p < oo. Section 9.2 lists the characterizations, by means of wavelets, of other functional spaces, without proof. Also included is a short discussion on the detection of singularities with orthononnal wavelet bases. Section 9.3 tma& expansions of L'-functions by means of waveleta; since L1 has no unconditid bases, wavelet8 cannot do the impomible, but they still do a better job thsp Fburier expansions. Finally, 59.4 points out an amusing difference in emphnnin betftraen wavelet and Fourier expmiom.
9.1. Wavelets: Unconditional baaw hr W(R), 1 < p < oo. We start by proving the Calder6n-Zygmmpi position theorem. THEOREM 9.1.1. Suppose j iu a pwii& jtnctien in L1(R).Pi+ a > 0. Then R can be dewmp0ae.d as follow:
2. On the "good" set
G,f (z)Ia a.e.
3. Thc W"set B can be written an B=
U Qr, whm the Qk uc mm-anrhppingintmmb, keW
CHAPTER 9
290
and a 5 1
1. ch-
L-' of R.
Jq. &
~ ~ l - l
f(2)5
dl k E N
2a,
.
h
' L = 2' .SO thrrt 2-' da f ( ~ ) a. ~t i011ms thst dz f(x) 5 a for all k € Z This defines a first partition
f
+
2. Take a fixed interval Q = [kL,(k l)L[ in this h t partition. Split it into
.
+
i)~, '-I
two halves, [kL, (k $)L[ and [(k+ (k + l)L[. W e either of the halves, call it Q' and compute I*, = IQf J , & ,f(z). If 1, > a, then put Q' in the bag of intervals that will make up B. We have indeed
.
* . a < lo# < IQ'I-I iQdz f(z)= 21QI-' dz f(x) < 2a . *
]0
If < cr,- keep going (split inti halves, etc.), if necessary,ad infiniturn. Do the same for the other half of Q,and also for all the other intervals [kL, (k + l ) L [ . At the end we have a countable bag of "bad" intends which all satisfy (9.1.1); call their union 3 gnd the complement eet G. fql
3. By the construction of B, we find tbat for any z 6 B, there exists an infinite sequence of smaller and smaller inter& Q1, Q2, Q3, - . . SO that E Qn for e * v n, and IQnI*' Jq,, dg i ( g ) 5 a. In fact. 1Q,1 = )lQj-ll for every j, and QjC Qj,, . Because the Q, "shrink to" x,
Since the left eide is 5 a by c o ~ c t i o n it, follows that f(x) in G. '
< a a.e.
Note that the choice t = 2! implies that d l the intervals occurring in this proof are automatically dyadic intervals, i.e., of the form [k2-j, (k 1)2-I[ for some k, j E Z. Next we define Calder6n-Zygmunct operators and prove a ck'ical property. DEFINITION.^ A CeMer&n-Zygmund operator T on R w a n integral opemtor
+
for which the ant@
kernel satis*
and which define8 a bounded opetutor on L2(R).
I
-
291
CHARACTERIZATION OF FUNCTIONAL SPACES
THEOREM 9.1.2. A from L'
Caiderdn-Zygmund opemtor ia 1390 a bounded -tor
(R)to L L ( R ) .
The space L'(R)
in this theorem is ctefined as follows. DEFINITION. f E L L ( R ) tf there eztsb C > 0 so that, for all a > 0,
The infinurn of 1 1 C for which (9.1.4) holds (for all a 2 called tl f ll LL*
> 0) is sometimes
Indeed, if
1. If f E L1[R), then (9 1.4) b automatically satisfied. S,. = {x; lf(x)I 1 a),then
2. f(x) = IX" is in LLeak, since I{t; f(x) = IxI-B isnot in LLd IfP> 1.
121-'
2 a)i = 2 a
•
However,
The name Liid is justified by these examples: L L extends L1,and contains the functions f for which J 1f 1 ",justnmisses to be finite because of logarithmic singularities in the primitive of If 1. We are now ready for the proof of the theorem.
Proof of Theorem 9.1.2. 1. We want to estimate I{x; I Tf (x)l 2 all. We start by making a CalderbnZygmund decomposition of R for the function If 1, with threehold a. Define
'a w
9(4 = IQkI-I
b(z) =
[
/0,
d~ f (Y)
/ah
- 19k-l a j(y)
if s E interior of Qk
,
if r t interior of Ok.
Then f (z) = g ( x ) + b(x) a.e.; hence Tf = Tg + Tb. It follows that ITf (x)l 2 a is only poasible if either ITg(z)l > a/2 or ITb(x)f > a/2 (or
'--
CHA~JTER9
aQ2
*
both); consequently,
The theorem will therefore be proved if each of the terms in the right-hand eide of (9.1.5) is bounded by
5 1)
f)ILt.
2. We have
{ti
I%(z)1241
.I, dz 1Wz)12 =
IITgJtis 5 C I i g H ~, ~ (9.1.6)
became T is a bounded operator on t2.Moreover,
(use the definition of g, d 1f(x)l 5 a on G)
2a
j.az
ll(z)l =
llf l l ~. l .
Combining this with (9.1.6), we obtain
Qz
3. We now concentrate on b. Fot each k, we define new intenrsls by "stretching" the Qk:Q; hae the same center a k rn Qk, but twice its length. We define then B* = ukQ;,and C' = R\B4. Now
293
CHARAC'X'EZUZATIm OF ~ C T I O N A ISPACES ,
4. It remains to estimate ( ( x E Go;(Tb(x)( 2
$11.
We have
5. To estimate this iaet integral, we separate the different contributions to b. Define b k ( x ) by
(z) Then q x ) = Tb = Ck
if z$!Qk, I d y f (y) ,jf z E interior of Qk
.
C,
m,
bk(x) a.e., since the Qk do not overlap. Consequently, and
=
C /R\Q; dz i/a&dl,K(..u)
bk(Y)/
k
(yk is the center of Q h ; we can insert this extra term because dp bk(y) = 0)
The M e r e w e K(x, y) - K ( x , yk) can be eetimated by using the bound on the partial. derivative &K of K with mpect to its second variable:
5 C'
(after the substitution x = yk where (u(2 2, (v( <: 1) (independent of k) .
+ Rku, y = yk+ RLv,
Substitutin,: this into (9'1.10)yields
Together with (9.1.7), (9 1.8), and (9.1.9), this proves tbe theorem.
8
Once we know that T maps L2 to Lg and L1 to L L , we can extend T to other LP-spaces by interpolation theorem of Marcinkiewicz. THEOREM 9.1.3. If an apemtor T satssfies
el:
whenq14pl,qg~pr,thenfor= 1 PI If = 41 ~ + ~ , u n t h O < t < l , thew a u l a~ constant K , dcpen$tng on PI, ~ I , B , 41, and t , so that
IITf l l ~ q5 K Ilf IILP
.
Here Lteak stands for the space of all functions f for which
is finite. This theorem is remarkable in that it only needs weaker bounds et the two extrema, and nevertheless derive&bounds Gn Lo-norms (not L L ) for intermediate values q.' The proof of this theorern is outside the scope of this chapter; a proof of a more general version car be found in Stein and Weise (1971). The Marcinkienricz interpolation theorem implies that the L1 -, L i d boundedness proved in Theorem 9.1.2 is a d c i e n t to derive LP+ Lp boundedness for I < p < oo, a9 follows.
CHARACTERIZATION OF FUNCTIONAL SPACES
295
THEOREM 9.1.4. If T is an integnJ opemtor with integral kernel K satisfing (9.1.2) and (9.1.3), and if T rs bounded from LZ(R) t o L2(W), then T extend8 to a bounded operator from LP(R) to LP(R)for dl p with 1 < p < &. Proof. 1. Theorem 9.1.2 proves that T is bounded from L1 to Lkmk; by, Marcinkiewicz' theorem, T extends to a bounded operator from LJ'to LP for I < p 5 2. 2. For the range 2 5 p < oo, we use the adjoint
of T, defined by
It is associated to the integral kernel K ( X , y) = K(y, x), which also satisfies the conditions (9.1.2) and (9.13). On L ~ ( R )it, is exactly the adjoint in ad sense T*, so that it is bounded. It follows then from Theorem 9.1.2 that f is bounded from L1 to LLedc,and hence by Theorem 9.1.3, that it is bounded from LP to LP for 1 < p < 2. Since for + L = 1, : L P - r L P is the adjoint of T : Lq-rL'J, it follows that T-is bounJed for 2 < p < m. More explicitly, for readers unfamiliar with adjoints on Banach spaces,
:
-
s
I
sup
llf lit* Il~9lltPIC llf l l t q .
# € L' IIILP=~
i.
(Strictly speaking, changing the order of integrations in the third equality is not albwed for all f,g, but we can restrict to a den.% subspace where there is no such problem.)
I
6 -
We can now apply thie t o prove that if $ has some decay and some regularity and if the 7Clj,k(x) = 2-j/2 ,,!.,(2-Jz k) constitute an orthonormal basis for t2(R), then the $j,k atso provide unconditional bases for LJ'(R), 1 < p < oo. What we need to prove (see Preliminarieu) is that if
-
-
then ~ , k
\
for any choice of the W j , k = f1. We will assume that $Jis continuously differentiable, and that both $ and 11' decay faster than (1 -t lxl)-':
Then 4 E L P for 1 < p < OP, and f = Cj,kcj,k+,,k implies c,,k = &f ( 2 ) QJ,k (x), because of the orthonorrnality of the Jljlk. We therefore want to show that, for any choice of the w J , k = f1, T, defined by
is a bounded operator from LP to LP. We already know that Twis bounded from L2 to La, since
so the LP-boundedness will follow by Theorem 9.1.3 if we can prove that T, is an integral operator with kernel ~atisfying(9.1.2), (9.1.3). This is the content of the following lemma. LEMMA 9.1.5. Choose w j , k = f1, define K ( z , y ) XI,,wj,k+jlk ( ~ ) $ j , k ( ~ ) . Then exwta C < oo so that n and
/ K ( x v)I , 5
<
l $ ~ , k ( ~ )bh~,k(g)l l
.
C
2-j(l + (2-jz - kI)-l-'(l+
12-jg
- kl)-'-'
flk
(by (9.1.13))
-
Find jo E Z m that 2m 5 )z yl Ccvoparts: j < joand jzjo.
5 2fo+'. We split the sum
.
j into
CHARACTERTWTION OF FUNCTIONAL SPACES
+
2. Since Cr(l la - kl)-'-'(l+ values d a, 6,' we have
3. The part j
< jc
297
-
(b kI)'L-C is uniformly bounded for all
is a little Less easy.
Find k~ G z so t b t b 5 2 j Y 5 b+ 1, and define C,= E - b. Then
Sidariy,
2 Cor&wntly,
a0 that
+ p'y-kl>l+
with a
~JY,
I
CHAPTER 9
If the~eforefollows that ( K ( z ,y)l
< Cis - ~ 1 - l .
4 For the estimates on a, K, ByK, we write
and we follow the same technique; we obtain without difficulty
F'rom the discussion preceding the lemma it therefore follows that we have proved the following theorem THEOREM 9 1.6. -If t,h 2s C1 and j@(x)J,1t,hi(x)12 C(l i- Jxt)-'-', and af Wle $J,,~(z)= 2-3'' + ( 2 - J x - k) wnstit& an odwnonnal bash for L2(R), then the 3, k E Z) also wnstatute ah unmditwnaf basis for dl the tP-spaces, l
characterization of functiun spaces by means of wavelets.
Since the Ilr,,k constitute an unconditional basis for LP(R), there exists a characterization for functions f E LJ'(R) using ;dy the absolute values of the wavelet coefficients of f. In other words, even f , then we can decide whether f E P by looking only at the IV, 96,,k}1. Tbe explicit criterion is, again for 1 < p < oo,
I
For a proof that these are indeed equivalent characterizatiollliof LP(R),we Meyer
(1990). 1
Similarly, wavelets provide unconditiond bsses and characterizations far many other functional spaces. We list a few here, w i t b u t proofs.
I
The Sobolev epeces W8(R). Ttne Soblev syaces are defined by
.a@@,
CHARACTERIUTmN QF FUNCTIONAL SPACEB
*
Their characterization by meam of wavekt cadiciepta is
The H6lder spaces CD(R). For 0 < a < 1, we define
For s = n + s t , 0 c d < 1, wedefine
For integer values of s , the eppropriate in this ladder are not the traditional 0-spaces (consisting of the functians that am n times &ntinuowly M~fantiabie),nor even the Lipschtz-spaeee, but slightly larger e p a ~ e adebed
k
which takes the place of C1(R), and
For thie ladder of Holder spaces one bas the following characterization:
-
A W y integrable f is in CD(R)(a nonintepr) or A t (s = n integer) if and only if there exists C < oo so that, I(f,
-I ,
.
'
" *
h,r;)I 5c )
5C
ford ~ E Z , for all j
> 0,
kEZ.
(9.2.1)
We have imptieitly assumed here that 11 E C, with r > s. For proof8 and more examples, see Meyer (1990). Of the qunplea given here, the only epacea that can be completely characterized (with Uif and only if" conditions) by Fourier t r d o r m s are the +bolev apscee. The conditions (9.2.1) characterize gZoM regularity. LucaJ regularity can also be studied by means of coefiiciente with respect to an orthonwmd wavelet The moat general theorem is the foliowing, due to J&d (1989b). For simplicity, we assume that tC, hse o ~ m p b cbupport t a@ ie C1(the formulation of the theorem is slightly di&wt b r more general $).
.
300
CHAPTER 9
THEOREM 9.2.1. Iff is Ho'lder continuous with ezponent a,0 .< a < 1, at 20,
i.e.,
If (4 - f(zo)l 5 C b - sola ,
(9.2.2)
then m$(f, for j-mo. then
$-~,r)ldist (xo, a ~ p p o r t ( $ - ~ , ~ ) = )-0 ~ )(2-(4*)s
)
(9.2.3)
Conversely, if(9.2.3) holds and iff is known to be CY for some c > 0,
We do not have exact equivalence between (9.2.8) and (9.2.2) he&. The estimate (9.2.4) is in fact optimal, as is the condition f E CC:i f f is merely continuous, or if the logarithm in (9.2.4) is omitted, then counterexamples can be found (Jdard (1989b)). Non-equivalence of (9.2.2) and (9.2.3) can be caused by the existence of less regular points near s o , or by wild oscillations of f (s)nesr a (we, e.g., Mailat and Hwang (1992)). If we modify condition (9.2.3) dightly, then these problems are circumvented. More precieely (again with compactly supported p4 E C1),we have the following. THEOREM 9.2.2. Define, for c > 0,
.
I , for some E > 0 , and some a, O < a < 1,
then f i s Holder continuous with eqownt a an xu.
Pmf.
+
1. Choose any x in ]so->, zo c[. Since either -implies k E S(xo,j;e) , we have
' i t follows that
$j,k(~)
# 0 or p41,k(zo)# O
*
CHARACTERIZATION OF m C I ' I O N A C SPACES
I
301
2. Since $ has compact support, the number of k for which $j,k(x) # 0 or +j,k(xo) # 0 ia bounded, d o r m l y in j , by 2 Isupport ($)I. Consequently,
Since $ is bounded and C',
k
3. Now c
i
1. Similar theorems can, of cout&, be p
-
b jo so that 230 5 Jx
5 2jo+l. Then
.
d for V-spaces with a > 1.
2. If a = 1 (or more generdly, a E N), then the wry h t step of the proof does not work any more, because thesa~ond8erieswill not converge. That is why orre has to be more circumspect for a, and why the Zygmund
class eaters. 3. Theorems 9.2.1 and 9.2.2 are tho true if $ fLes infinite support, and 91, and haw good decay at oo (see Jaffard ( 1 ~ ) ) Compact . support for $J mdcecr the estimates easier. o P
of wavelet coefficients. Local regularity can therefore be etndiaa %y For practical purposes, one should b ~ , l it may ~ be :that ~ r large y value. of j ace needed to determine a in {9.85) reliably. This is illustrated by the following example. Take 3
+
f(=
-
=
if z l a - 1 , if a - l I z L o + l ,
2 e-f~-Ol C-lz-4 e-{2-al
>
[(z-a-$+q
u
ihbr function irs graphed in Figure 9.1 (with a = 0).
z,a+l;
This function has Hiilder 0, 1, 2 at z = a - 1, a, a 1, respectively, and is Cw elsewhere. *'kt then, for each of the three points zo = a - 1, a, or a + 1, compute
+
CHAPTER 9
{((f*$j.k)I; zo E support ($j,~)),and plot log Aj/ log 2. If a = 0, Aj = then the+ pIde fine up on etraight lines, with dope 1/2, 3/2 and 5/2, with pretty good d e u y , leading to good estimates for a. A decomposition in orthonormal wavelets is not translation invariant, however, and dyadic rationals, particularly 0, play a very special role with respect to the dyadic grid (2-fk; j, k E Z) of localization centers for our wavelet basis. Choosing different values for cr . illustrates this: for a = 1/128, we have very different (f, q3j,k), but still a reasonable line-up in the plots of logAj/ log2, with good estimates fox a;for irrational a, the line-up is much less impressive, and determining a becornea correspondingly leas preck. All this is illustrated in Figure 9.2, shawing the plots of logAj/log2 as a function of j, for zo = a - 1, a, a + 1 and for the three choices a = 0, 11128 and 8- 11/8 (we subtract 11/8 to obtain a cloae to zero, for programming convenience). ?b make the figure, ((f,$-f,k)I vm~computed for the relevant valuea of k end for j ranging from 3 to 10. (Note that this mean6 that f iteelf had to be dunpled with a resolution 2-17,in order t o have a reasonable accuracy for the j = 10 integrals.) For a = 0, the eight pointe line up beautifully and the estimate for a is accurate to l e . than 1.at all three locations. For a = 1/128, tbe points at the coarser resolution d e e dotnot dign as well, but if a + 3 is estimated from only the finat four resolution points, then the estimates are still within 2%. For the &ional choice a = fi 11/8 no a l ' i e n t cm be eeen at the discontinuity at a - 1 (one probably needs even d l e r scales), and the estimate for a -t3 at a,y h f is Lipechitz, is off by about 13% (interestixtgJy enough, the estim* would be much better if the d e 10 point were deieted); a t a 1, where 1' is Lipschitr, the. estimate is within 2.5%. This illustrates that to determine the lacs regularity of a function, it is more wefid to use very redundant wavelet families, where this translational m n - i n w c e in much l e a pronounced (discrete case) or absent (continuous caae).' (See H o b e i d e r and Tchamit&ian (1990), MBUat and Hwrrng (1992).) Another teason for using very redundant wavelet families for the chatacteriaation of load mgukwity is that then ody the number of wishing mornof JI limits
+4
-
I I
+
CHARACTERIZATION OF FUNCTIONAL SPACES
INDEX j OF THE SCALE
4 8 SCALE
-
8
1
0
-
Pro. 9.1. &dimotu of the H W awwnb of f(z a) (w Piam 9-11 41 a 1 (Cop), *,(mid&), o + 1 (am), E o r l l p d a l ~ k A j / h 2 , f o r d i d n m t ~ ~(Ihi.AOmr fa. wresptr(buwby1~.~ i s u J l c , ~ I v o J d l i k a t ~ ~ I b r ~ W ~ )
the xnaxhum regularity that can be characterized; the regularity of tj plays no role (see $2.9). If orthonormal bases are used, then we are necemady limited by the reguhrity of q6 itself, as is illustrated by choosing f = $. For thi choice we have indeed (f,q6-j,kJ = 0 for all j > 0, all k; it follows that with orthonormal wavelets we can hope to characterbe only regularity up to C 7 - c if $ E Cr.
9.3. Wavelets for L1([O,11). S i L1-spacesdo not have unconditional bases, wayeleta cannot provide one. 'Nevertheless, they still outperform Fourier analysis in somebeeme. We will illustrate this by a comparison of expansions in wave1eta versua Fourier aeries of L1([O, I])-functions. But first we must introduce " p e r i o d i i wavelets." Given a multireeolution analysis with scaling function 4 end wavelet tj, both with reasonable decay (say, I$(x)l, I@(x)l 5 C(l /zl)-'-'), we define
+
C
and
+ + EL +
Sha )(x 0. = 1: we have, for j 2 0, G ( z ) = 2-jIa Et#(2-fz- k 2 - j l ) = 2na,so that the for j 2 0, are d l identical one-dimensional spaces, amtaining only the constant functions. SimiIeily, because +(x C/2) = o , Wjw ~ = (0) for j _> 1. We therefore restrict our attention to the W.fm with j < 0. Obviously WPI c, :% a property inherited from the non-periodized spaces. Moreover, i~ st111 orthogonal to because
v,
y,
y,
v,
=
C
re2
(+jL+l~j~v,
h!~
=0
# j , ~ )
a
It Pdbn twt, ss in the non-ps.iodirmd ass,5y5= vjW @ WjW.Ths lpaca vim, WfPa are &lk+dimensiond: 8 h m 4iC+m2~jl = 41.k fbr m € 2, and the same is true for ), both and W J pare spanned by the 2111 function8 obtained from A = 0,1,.. ,2111 1. These 2b1 functione are moreaver orthonormal; in
a
%p -
I
I
1
I
CHARA~WTION OF FUNCTIONAL SPACES
e.g.,
W r we have, for 0 5 &,A1 5 2bI par
(J;,,
9
CL)
=
C
- 1,
(djC+2w~., @jp)= &,kt
.
t€Z
We ha* therefore a ladder of multireeolution w,
., (6) {c;
with successive orthogonal complements W r (of CfoPQ in Vy ), Wlw ,. and orthonormal basea { 4 , , k ; k = 0,. - 2111 - 1) in Viw, {#J,k; k = 0, ,2131 1) in WJPr. Since u , ~ - ~ % = ~L2([0,11) (this Follows again &om the correa-ng non-periodized version), the functiohs in U -j E N, k = 0, - . - ,2111 - 1) constitute an orthonormal be+& in L2[[0,11). We will relabel this basis as follows:
-
-
+J,(x) = r l z k ( x ) = gv (z k2-3)
for 0 S k 5 23 - 1
Then this basis has the following remarkable psoperty. THEOREM 9.3.1. If f i s o continuow penodrc jbnction with period 1, Uaen them d a, E C so that
1. Since the g, me orthonormal, we necesgarily have a, = ( f , h).D d m e
SN by N
La 6rst step we prc& that the SN are udifomly boundad, is., with C independent of f or N 2.
If N
= 2-f, then Su = Pmjvrr; hence -5
and this is uniformly bounded if f9.3.2) for N =. $.
J4(s)lIC(1+ JZ~)-"~. This ez3tablbht.s
hlstimatee exactly Wlar to those in point 2 shaw that the Lm-nnnn of the second sum b slso bounded by C 11f llL-, uniformly in j , which pravee (9.3.2) for all N.
I, I
Proof. 1. Suppose, for simplicity, that q9 is compactly supported, with support $ c [- L, L]. For sufEcientIy large j, this means that $J:*(Z) = $-,,*(x) if 12-Jk zo( 5 2-1. (Again, this is not crucial. For ~ m p a c t sup l ~ ported 9, one only has to be a little more careful in the estimabes below.')
-
2. Form = 23
+ k, a,
support $-,,r,
C
c
= Jdz j ( z ) ~ / J - ~ , ~ (Here Z).
-
+
[2"(k - L),2"(k L)] P - j ( z o - L - I), 2 - 3 ( ~ o + L + l ) ] (becam 12-lk- zol 5 2-1) ;
hence
(use f ( z ) - ~
-
( X O ) (5
- xo)fl(~o)=
and d u q e variables: y = 2, (z - xo))
- XO),
This has the following corollary. COROLLARY 9.4.2. If, for cJI m, Cl rn-=I2 I la,( 5 C2 m-3/2, wiU, is in Ca for all a *< 1, bui i s nowhere C1 > 0, C2 < oo, then CZzo am dzffernntiable. P m f . Immediate &om Theorem 9.2.2 and Lemma 9.4.1. m +
.
~13tUB now constmet a very psrticular function. ' W e a,= a p + k =
independently of k. Then
Pj,
CHARACTERIZATION OF PUNCTIONAL SPACES
where F(x) =
C,
+(z - m ) is a periodic function. We have
with
In the special case where $ = J"Msya (see Chapters 4 and 5 ) ) support 6 = {E; 'j 5 #I 5 ao that $(2m)# 0 only if n = f1. Moreover, $(-2a) = q(27r). Consequently, F ( x ) = A cas(2nx). and
9))
The "fulln wavelet series of the left-hand side has a lacunary Fourier expansion! If now the p, are chosen so that C12-J 5 231' 8, 5 Cz2-3, then we can apply Corollary 9.4.2," and conclude that the function is nowhere differentiable. For this special case, this is in fact a well-known result about lacunary Fourier series: 7, cm(A,z), with I-y,l < oo but %A, f . 0, defines a continuous, nowhere differentiablefunction. On the other hand, if we take a function with a localized singularity, but which is Gw elsewhere, such as, e,g., j (z) = (sin zx(-O, with 0 < a < 1, then its wavelet expansion will be more or less lacunary (all the coefficients decay very fast as -J-*oo, except the few for which 2'1~llc is close to the singularity), while the Fourier series is "full" : fn = -y,n-l+a O(C~+O), with yo # 0; the effects of the singularity are felt in all the Fourier coefficients.
xz,
+
Notes. 1. Tbere exist many different definitions uf Calder6n-Zygmmd operators. A discussion of these different definitiolls and their evolution is given a t the start of Meyer (1990, vol. 2). Note tkt$the bounds are infinite on the diagonal x = y; in general K will be singabran the diagonal. Strictly speaking, we should be more careful about what trappens on the diagonal. One way to make sure everything is well defined is to require that T is bounded h m V ta V' ('D is the set of aH compactly supported Cm functions, V' its dual, the spece of (non-tempered) distributions), and that if x $ support (f), then (Tf ) ( x ) = dy K(x, y) f (y). It then follows that K does not cornpletely determine T: the operator (TIf )(x) = (Tf )(x) m(x)f (x), with rn E Lw(R), has the same integral h 1 . See Meyer (1990, vol. 2) for a clear and extensive discuesion.
+
2. Note that Jf - (ILk,
mnstit%tes a (very convenient) abuse of notation. Aa rlpm,for a w p l e , by 11 1% - 11-' 1% + ll-'llrC, 2 Ilk - l)-lltt; #fqa$$I-* /Iry, the triangle inequality i.not satisfied, so that II.lILkd is not a 9 ~ e norm. "
+
+
310
I
CHAPTER 9
3. If the "weakn is dropped, then the theorem is known as the Riess-Thorin theorem; in this case K = and the restriction ql S pl, ~z 5 p2 is not neemarye
qc-t,
5. We can suppose without lorn of generality that a 2 0. Find k so that k a 5 k + 1. Then
<
6. In Note 9 of Chapter 5 ye eaw that
I-",
EL +
$ ( x P ) = conetank Since d;t QYZ) = 1, this ~~ is necessarily equal to 1,
8. By now, the reader has seen so many in8tances of this type of estimate that I leave the proof of Lemma 9.4.1 for non-compactly supported but welldecaying qh as an exercke.
I I
311
CHARACTEFUZATION OF FUNCTIONAL SPACES
9. Yes,Meyer'e waveIet does not have compact support, and the proof of kmmer 9.4.1 uses that $ is compactly supported. See, however, Note 8 above.
t1
\
i
'
CHAPTER
"
10
Generalizations a-nd Tricks for Orthonormal Wavelet Bases
This chapter mnaists of severat generalizations and extensions of k l i e r w t r u c %
tions. These are not treated with the same d e w as in the previous chapters. Some of the topics are still developing, and I expect that any write-up on them would look very different even two years from now. The sections cover multidimensional wavelets with dilation factor 2, via tensor product multire80lution analysis, or via nomeparable schemes; orthonormal wavelet bases with dilation factor different from 2, integer or non-integer; the "splitting trickn for better eequeency resolution (in fact, merely a specid case of the "wavelet packetsn of fhifmm and Meyer); wavelet basea on an intend.
10.1.
,
Multtdimensional wavelet basPrs with dilation factor 3.
For simplicity, we will consider only tbe two-dimensional case;higher dimemiom are analogous. One trivial way of constructing an orthonormal bssi for L ~ ( I I ~ ) , starting from an orthonormal wavelet basis g(r,,,(s) = 2-1i2 .tfi(2-fz - k) for L2(R),is aimply to take the tensor product' functions generated by two one~ o h:d
The resulting fundions are indeed wavelets, and ( 9 h,l;l; jl ,ja, kl,k3 E 2) is an orthonormal basis for L~(R~). In this basis the two variables zl and 23 are dilated v a t e l y . There exists another construction, more interesting for many eqplications, in which the dilations of the resulting orthonormal wavelet basis controt both variables simultaneously. In this construction, one considers the temor product oftwoone-dimensional m u l m t i o n gndyses rather than of tha mm%pading a- bases. More precisely, define tpama V,, j E Z; by J,
I
v* = VO 43 VO
= Sw(F(x, 9) = f (=lg(y);
F E V3M 3(21'2f.) ., E Vo
.
Then ;the Vj iorm a muttiresolution k k k in L = ( R ~ )
f,9 Vg) ,
\
'
CHAPTER
10
Generalizations a-nd Tricks for Orthonormal Wavelet Bases
This chapter conaists of several generalizations and exteadona of k i i e r coastructiom. T h e are not treated with the same deUd as in the previous chapters. Some of the topics are still developing, and I expect that any write-up on them would look very different even two years from now. The sections cover multidimensional wavelets with dilation factor 2, via tensor product multire80lution analysis, or via nomeparable schemes; orthonormal wavelet bases with dilation factor different from 2, integer or fion-integer; the "splitting trickn for better frequency resolution (in fact, mereIy a special case of the "wsvelet packets" of Chifinam and Meyer); wavelet bases on an interval.
10.1.
Multidimensional wavelet basPwr with dilation factor 3.
For simplicity, we will consider only the twodimensional caw; higher dimensions are analogous. One trivial way of comtructiag an orthonormal bas'i for L~(R'), starting from an orthonormal wavelet basis plr,,,(;c) = 2-JI2 +(2-Jz - k) for L2(R),is simply to take the tensor product' functions generated by two one dimemioasl beees:
The resulting fundions are indeed wavelets, and {* ,, ; jl,ja,kl,h € ZZ) ia an orthonormal basis for L~(R'). In this basis the two variabk z l d zl are dilated separately. There exists another construction, mare interesting for many applications, in which the dilations of the resulting orthonormal wavelet basis coatrot both variables simultaneously. In this construction, one considers the tenaof product oftwoo ~ e n s i o n m d u l m t i o n gndyses rather than of tha correspoading u a W baees. More precisely, d e b 8gmcea V,, j E Z;by
Then W Vjfimn cr multireso1ution bddet in L = ( R ~srtisfyiPg )
,
314
/
CHAPTER 10
I
~ i n d ' t h e+(. - n), n E Z, constitute an orthonormal bash for &, the produclr hmctiona i I
@~;nl,nl(~, Y) = #(z - nl) &(x- n 2 ) ,
n ~n2,
constitute an orthonormal basis for Vo, generated by the single function a. Similarly, the -
E
translations of
8
I
collstitute an orthonormal basis for Vj. As in the one-dimensionid case, we d e fine, for each j E: Z, the complement space Wjto be the orthogonal complement in Vj-1 of Vj. We have
Vj-1
=
&-,a V,-1
= (b$ W,) 0 (V, $ U/j)
= V,@&@[(Wj@Vj)@(~@Wj)@(~j@~j)]
=
v, 03 wj .
It follows that Wj -consists of three pit&, with orthonormal basta given by the +j.n,(z) @j,n,(~) (for Wj @ 41, 4j,nl(z) tlj,nS(y) (for 5 QP Wj), md +j,nl (x) +,,n2 (Y) (for Wj @ W j ) -Thie leads US to define t h wavelets, ~
( h , ~d ,stand for 'horimntal," “vertical," "diag&&" respectively; gee belw). Then { Y ; ~ , , ~nl,n2 ~ ; E Z, A = h, v or d ) is an orthonorpal basis for Wj, and
If, in tthie cmukruction, the original o n d i m e m i o n a l ~and JI have compact support, theq obviously m have @ and the aX.Moreaver, the interpretation in terme of p~bbandfiltering of B decomposition with reepect to such sp ort h o n o d bsels of compactly supported wavelets, as explained in $5.6, cMies aver to two dimensions. The filtering can be done on h" snd "columnsn in 'the twdimenaional army, cormponding to horiEontal and vertical directions in imegee, for example.' Figure 5.7 becomes, in two dimensions, the schematic representation in F i 10.1.
.
\
GENERALI3ATIONS AND TRICKS
FIG. 10.1. Scliaohc npnrento#on of rrpcoted low- an$ high-pcw fltmng, on rrnw and cdumns, for a b d i m m r i o n d waocllt &orition.
The dl*' correspond exactly to the wavelet co~fficiente(F,q;,,), with F = En $*o;n. In an image, horizontal edgea will show up in d'vh, vertical edges in dlvV,diagonal edges in dlld, as illustrated in the image example below. (This j u e t h the h, u, d superscripts.) Note that if the original image (8) consists of an N x N array, then (apart from border &ecta; see a h §10.6), every array d l s Xconsists of
$x
elements, and can therefore be r e p r ~ t e dby an image (magnitudes of the eoefkients corresponding to grey levels) of one quarter the size of the original. The whole scheme can therefore be represented as in Figure 10.2. Of course, one can d=mpaee 2 even further if &ore multimluti~n layere are wanted. Figure 10.3 shows this decomposition scheme on a red image, with three multireeolution layers.
*
dccdmpooition into two layem
Aa 10.2. Sdienoatic botudbhn of &we 10.3.
'
~entatkm of the
w
l
b
d
a
of t h Coro-d(-
raordst
thh.aaaosmad two-dimschemeo which have a temw product structure- 4 h aar dso~coneiderthe age in which one starts h m a dimensional multirwolution analysis (with the V, satisfying all the obvious gen-
316
CHAPTER 10 I !
.. Wa. 10.3. A d image, end itr mawkt &mptu&m into tlkat mul#ruohrlioR M. ~ n t & ~ l # ~ ~ ~ k t ~ ~ ~ ~ r a u t h o t t l b d ~ ~ , & * ~ , B A a r r p h a r t e , ~ ~ a h W p i d . t r r tk, c ? k h o m m 4~1 o d # o d ~ ( l b h i l # t ~ t l b d * ~ r n o IncsP~o d a l Y k e t o ~ M . B o r h d j a ~ t h i r j S l a r n .
.
GENERALIZATIONS AND TRICKS
317
eralizations of (5.1.1)-(5.1.6)) h whicb Vo is not a tensor product of two onedimensional Vo-spaces.' Some (but not all!) of the constructions done in one dimension can be repeated for this ease. More precisely, the multiresolution structure of the Vj implies that the cbrresponding d i g function 4P satisfies
for some sequence (hn)n,C. ric polynomial
O r t b o ~ t ofy the 99,n form the trigonom&
-
4
to satisfy
To construct an orthonomd basis of wavelets corresponding to this multirlution analysis, one has to find three wavelets rkl, 92, in Y - l , orthogonal to Voand such that the three spaces spsnned.by their respective integer translates are orthqpnal; moreqwr the qX(-- n) should also be orthonormal for each Sxed A. T!his i m p k that
where the ml ,mz ,m3 are such that the mritrix
is uni@ry. The analysis leading to this d i t i o n ia entirely similar to the onedimeandpis in $5.1; see, e.g., Meyer.fl990,§1I1.4).~ NOW%?& the number of wavelets to be constructed can be determined by an easy trhg."lEtrtwo dimensions, for example, Voie generated by the translates of m e firoctiimk $*,y), over the space VI1 is generated by the translates of @(ZX, 2 ~ ) f z?, .orequivalently, by the be2--$ates of the four functions 6(2z,2g), @(2x- I,&), a(%,2y I), a(%- 1,2y 1). V-l is therefore "foq times as big" as Vo. t+e other hand, earh of the w;-spaces is generated by the 2'-translates 0f a w e h p k t h *j(t, y), anrf is therefore "of the eame sizen ss Vo. It follows that bne n e (bres {= four minus one) ryscca W; (hen& three wavelets *j} to make up'tlkmmpbment d Vo in V-I. This rub may eound u b d - ~ v i n g , budo~a-*harsh, " tephm (and p&) it in more
z2;
-
-
CHAPTER 10
,
mathematical terms: the number of wavelets is equal to the number of different itself) of the subgroup z2in the group f z2. In, the general n-dimensionai caee, the same rule shows that there are 2" - 1 different functions mj to determine; they have to be su& that the 2" x 2"dimensional matrix coeets (diierent from
z2
h unitary, with r = 1,---,2",~ n 8d = (81,*..,8n)E (0,1)".~ In fact the unitarity requirement of (10.1.4) or (10.1.5) calls for a tricky balance: m ~m2, , m:, have to be found so that the first .row of (10.1.4) has unit norm, which seems h l e s e enough, but yte also simultaneomly need orthogc~ nality with and Bmong the other rows, which are all shifted versions (in or C) of the first row. These correlations between the rawe may be hard to juggle in practice. It is useful to untangle them first, which k be done via the so-cdled plyphase decompoeition. We write, e.g.,
<
me,,, j = 0, . - . ,3, are defined eimilarly from me,e = 1,- .,3. One easily checks that (10.1.3) is equivalent to
Similarly, all the other conditions ensuring unitarity of (10.1.4) can be recast in terms of the mrj;one finds that (10.1.4) is unitary if and only if the polyphase matrix
ie unitary. In n dimensions, one similarly defines
u
and the unitarity of U is eisnivalent to the unitarity of the polyphase matrix defined by &,8(~l~"',~n)= %-1,8(~1+".,~n) (10.1.7) The construction therefom boils down to the following question: given ma (&om (10.1.1), (10.1.2)), ean mr, ..,mp-1 be found such that (10.1.6) ie unitary? In the bmdimeneionel case, and if mo(& C) happens to be a d trigon* metric polynomial, then one can even dispense with the polyphase matrix: it is
-
,
319
GENERALIZATIONS AND TRICKS
+
check that the choice rnl(€, C) = e'emo(t n,C), rnn(<, () = e-'(E+O w(<, 5' -t R ) , m a ( €C) , = e-<mo(< x , C n ) makea (10.1.4) unitary. If m,-, is not real, then things are more complicated. At first sight one might even think the task is impossible in general in the n-dimensiond situation, where (10.1.7) is a 2" x 2"-matrix: after all, we need to find unit vectora, depending continuously on the ( n d y the second to last columns of (10.1.7)), orthogonal to a unit vector (the first column of (10.1.7)), i.e., tangent to the unit sphere. But it b well known that "it is impossible to p m b a sphere," i.e., there exist no nowhere-vaaisbing continuous vector Wds tangent to the anit rephere, except in real dimensions 2, 4, or 8. The first column in (10.1.7) does not describe the full sphere, however; in W , becauae it is a mntinuous function of n variabiea (the - . . ,<=) in a 2n-diensional space*and 2" > n, it only describes a compact set of meamre zero. Thia fact saves the day and'make6 it p d b l e to construct ml,.-.,rnzn-l, aa shown by Grdcheaig (1987); see also sII1.6 in Meyer (1990). Griichenig's proof ia not constructive; a different, ccHlstructive proof is given in Vial (1992). Unfortunately, these constructions can not force compact support for the 92:even if Q is a trigonometric polynomial (only finitely many hn # O), the m, are not necessarily. -to
+
+
<,
[,,
\
t
10.2.
Onedimensional orthonormat wavelet b &tion factor larger t h a n 2.
s
with integer*
dilatm fBGtOr 3. A multiresolution analthe m e. way tie for dilation 2, i.e., by (5.1.1)-(5.1.6),exceptthat(5.1.4)ierepWby
For illustration purposes, let us ch-
ysb for dilation 3 is defined in exactly
r F.
1
f'
We can use the same trick again se above: Va je genersted by the integer trenslatee of one function, i.e., by the c$(x - n), whUe V-1 is generated by the 4(3x - n), or,'equivalently, by the integer traa&& of three functions, c$(3z), 4(32 l), and 0(3x - 2). V- is "three timee a big" as &, and two spaces of the "same siren ss & are n d e d to comp-t q and constitute V-r: re rill need two spaca G ,W$, or two wavelets, snd @. We can again introduce mo, ml, m by
i
O r t h o n o d t y of the whole family defined by
h,n(~) '
r
(e,, which defined can be
are the mt,
{&,,,$~,,,a,+; n f Z), where
is JJOW
-
3-31' d3-j~n)
malogously), again tos
#j,n
aevural orthonormality conditiom on m by the ru?q-t that the matrix
-
,
CHAPTER 10
is unitary. Again, one can restate this in terms of a polyphase matrix, removing the correlations betwem' the ram. Explicit choim of ma, ml ,q fot which (10.2.1) i s indeed unitary haw been constructed in the AQQPliterature (see,e.g., VPridyanathan (1987)). The question is then again, as ia4hapter 6, whether these Biters correspond to bona fide La-funct-#, @I; and q2,whdher the $,, constitute an orthonomal basie, and what t-h regtrlarity is of all these RecessariIy have functions. We know, from Chapter 3, that 10' and $"mast integral zero, corresponding to ml(0)= 0 = q ( 0 ) . Since the first row of (10.2 1) must have norm 1 for aII (, it follows that ~ ( 0 =) 1 (which is necessary anyway for the convergence of the infinite product fl,"3, m~(3-It)which defines &<)). The first column of (10.2.1) must a h have norm 1 for all c, so that 1 c-Y*-?'( m(0)= 1 implies tq,($) = 0 = m($),i-e., ~ ( is t divisibIe ) by + 'If, moreover, m y smoothness for $J', j8 desired, then we need additional ' vanishing moments of qj1,@: which by exactly the & m e argument as before, lead t o divisibility of m(C)by ((1 + e-4 4- e-2*)/3)Lif +', @ E CL' j. One is thus led to looking for mo of-the type n ~ ( 6=) ((I e-9 e-a*)j'3)NC(() such that fmo(()12 Imo(e Ima(( %)I2 = 1. If m +is~ a trigonometric polynomial, this means that L = lLJ2is again the solution to a Bezout problem. The minimal degree solution leads to functions 9 with arbitrarily high regularity; however, the regularity index only grows logarithmically with N (L. Vilemoes, private communication),' Once m~is fixed, ml and rnz have to be determined. The design scheme explained in Vaidyanathan et 81. (1989) gives a way t o do this. In this scheme, the matrix (10.2.1) (or rather, its z-notation equivalent) is written as a product of similar matrices the entries of which are much l&r degree pdynomials, with only a few parameters determining each factor If one imposes that the first column of a product of such matrice3 is given by the m-, we have fixed, then the d u e s of these parameters are fixed @ewise, and mi, rnl cap be read off from the product matrix? If the compact support constraint is iiied, then other wnstructione are pweible. In Auscher (1989) one csn Bnd escarnples where t$ and @ are CaOfunctions with fast decay (and infinite support). One final remark bout dilation factor 3. We have aeen that ~ I O must neceed y be divisible by (1 e'q e-2q)/3. This factor doea not vanieh for ( = ?r (unlike the factor (1 + e'*)/2 for the dilation factor 2 case). However, if we want to interpret rno ae a law-paes fitter, then %(A) = 0 would be a good idea. Xb ensure this, we need C(m) = 0, which means going beyand the h e s t degree solution to the BQout equation for JCI2. $ J ~
+
+ y)?+
+
+
+
+
+
i
321
GENERALIZATIOfJS AND TmCKS
F
C z
Similar constructions can be' made for integer dilation factors. For non-prime dilation factors a, one can generate mxptable r n from ~ constructions for the factors of a, although not all possible sorutions for dilation a can be obtained in this way. For a = 4, e.g., one cao atart from a scheme with dilation 2 and filters r q , and m l , and one can define ttce filters jiLO,fil,m2,m~(still orthonormal; the'distinguirshes them from the dilation factor 2 filters) by
(It is Beft to the reader as an exercise @ prove that this leads indeed to an orthonormal basis. One easily checks Ebat the 4 x 4 analogue of ('10.2.1) is unitary.) Note that the function 4 is the m e for the factor 4 and the factor 2 construGtions! We will come ?m& t o t k in $10.5. i
10.3. Murtidimenstonal wavelet bases with matrix dilations. This is a generalization both 510.1 and $10.2: the mdtiresolution spaces are subspaces of L2(Rn),and the basic dilation is a matrix D with lnteger entries (so that DZn c Zn) such that all its eigendues have absolute d u e strickly larger than 1 (so that we are indeed dilating in dl directions). The number of wavelets is again determined by the number of ccJeets of DZn; one introduces again m ~ml,, , . . , and the orthonormality d i t i o n a can again be formulated as a unitarity requirement for a matrix constructed from the w ,r n,~. . . The analysis for these matrix dilation r~lsesis quite a bit harder than for the onedimensional case with dilation 2, and, depesdiog on the matrix &asen, there are a few surprises. One surprbe is that generalizing the Haar basis (i.e., chooeing so that sll its nonvanishing coacients are equat) leads in many cases to a function 4 which Is the indicator functbn of a selfsimilar set with fractal boundary, tiling the plane. For two dimensions, with D = ( f ), e.g, one finds that $t can be the indicator function bf the twin dragon set, as shown in Grkhenig and Madych (1992) and Lawton snd Etesnikoff (1991). Note that such hctal tiles may occur even for D = 2 Id if Q is chosen 'boa-canonicallyn (e.g., m(L6)= $(I + c - ~ c+ e-*((+C) + c-'(e+X)) in two dimensions--see Grijchenig and Medych (1992)). For more complicated mg (not all c d c i e n t s are quid), the problem is to control regularity. Zero moments for the do not lead to factorization of in these m u l t i d i m e h a l --(becatwe it is not t m f h h t to know wrm of a mufti-db1e polynomial to factorize it), and one has to resort to other trickrr to m&rol the decay of A particularly interesting case is given by the "quincunx lattice," te., the twc+dimenaional case where D Z ~=, {(m, n); m + n E 22). In this caae there ia only one other m e t , and therefore o* QIM waWet to coastruct, so that the choice for ml is as straightfom-ard as it wae br dilation 2 in one d i m d o n . The
-1
+,
6.
k 4
6
CHAPTER 10
322
conditions on r r ~ml , reduce to the requirement that the 2 X 2 matrix
be unitary. It is convenient to chocwe
Note that any 0 t t b 0 ~ 2 n a &lis for dilation factor 2 in m e dimension 8 u b m a t i d y gives rise to a pair of candidates for ma, ml for the quincupx scheme: it suffices to take mot€,C) = n#(<) (where is the one-dimensionnl Glter).' Di&erent choices for D can be made, however. Two poeaibilities studied in detaiI in Cohen and Daubechies (t991) and KowikviC and Yetterli (1992) are D l = ( { - ] ) a n d D 2 = ( i -{). T h e m & f o r m o W t o v e r y d & r e n t wavelet bases for these two matrices; ia psrtidar,if one d e w , vim the mech* nism explained above, the filter mo from the "standardn one-dimensional wavelet filters ~ r n oin w.4, then the resulting 4 are inaeasingly regular tf Dz is choeen (with regularity index proportioual to N), whereas chOOBing Q1 Ieads to 4 which are at most continuous, regardless of W , Other choicm for D may lead to yet other families, with different rmarity properties again. One can of c o w also c h m to construct two lh&hogonal bases rather than one orthonormcri W i, as in 58.3; for the ch~lcesDz,DJ several possibilities are expIored in Cohen and Daubechies (1991) and KuvdkviC. and Vetterli (1992). Ln 'this biorthogond case, one can again derive filters fmim dimensionad constmdions. If one starts from a symmetric biortbqpal filter pair in one dimemion, where aU the filters are polynomials in cos then it d c e e to replace cui ( by 3 (coa& coe () in every filter to obtain symmetric biorthogonal filter pairs for the quincunx case8 Because of the symmetry of ttiese exarnp1e3, the matrices Dl and DZlead to the m e functions q5,6 in this case. One finds again that symmetaic biorthogonal bases with arbitrarily high regd-arity are possible (see Cohen and Daubechies (1991)). The quincunx case is of intereat in image prvlceesing because it treats the different directione more hoxnogen80~~1y t h the m b b (tensor-product) twdhe11siond Bcheme: instead of havipg two favorite directions (horizontals and verticals), the quincunx schemes treat horizontals, verticals, and diagonals on the BBme footing, without introducing redundancy to achieve this. The first quincunx subband filtering schemes, with aliasing cancellation but without exact reconstruction ( w h e had not h e n disc;overed men for one dimension at the time) sre given in VetterIi (1984); Fkauveeu (1990) contains orthonormal and biorthogond schemes, and linlts them to wavelet t m e q Vetterli, Kovaikvi6, and Le Gdl (1990) dimmaxs the we of perfect reconstruction quinamx fiiterhg schemes for HDTV applications. In Antonini, Barlaud, and Mathieu (1991) biorthogonal quincunx decompositioae combined with vector quantization gip t3pectacular results for image compaction.
<,
+
Q-TXIONl
10.1.
orth011omm1 r
O-C)-~~I
323
Fl ANDCKS
m tmm ~ with ~
iwm-&
dilation factom. &tion factors 1 2.' In one dimension, we have so fer only dimmed Non-integer dilation factom are alao posgie, hmmret. Within the framework of a multiremlution d y s i s , the dilation Earnor must be r s t M 1 O(for a proof, see Auscher (1989)). It had already beem pointed out by G. David in 1985 that the const~uctionof the Meyer wavelet could be generabed to d i M m a = for k E N,k 2 I; Aucher (1989) contains constructions for arbitrary r&iond a (w also Auscber'e paper in Ruskai et d* (1992)). Let us illustrate for a .= how the factor 2 scheme has to be adaptrd& We start again ftom a m u l t ~ l u t i o n analysis, deeDcd u in (5.1.1)-(5.1.0), with f instadof 2 for the dilation h t . We have a&n t$ E c V-* r ~~an(+(q -n)), so that
v, 4
;
(The nason for the superecript 0 will soon become clear.) Coneequently,
and o r t h o n o ~ t o yf the
+(a
- 2)m i w
On the other hand, #(. - 1) is also in GIand can therefore also be written ae a (Meredt) linear combination of the 9(#x - n),
-
Orthonormality of the 4(x 2C - I), and orthogonality of the #(x respect to the 4(2 2L) then impha
-
All thb xWam+thatwe have in fact trsnr rrq)-functions,
- U - 1) with
324
CHAPTER 10
What about ml ? We define again, for j E 2, the space Wjto be the orthogonal complement in of Vj- Note that V--1 is generated by the (P($x - n), n E Z, or,equivalently, by the even integer translates of t h m functions, namely
mrmponding to n = 3, n = 3C -k 1, and n = 3t + 2. The space &, is generated by the 2 2 tnrnslates of two functions, d(x -24) and 4(x- 2t- I), t E 2. It fallow8 that the complement space Wo is generated by the ZZ ktudate of a single function, Wo = Span (#(. -2n); n E 2); (UWo is half the&eof%.") We expect therefore an orthonormel basis of the type J r j V r ( ~ )= ($)-$i2t,b ( [ f ) j x - 2k), j,k E 2. This function can also be wikten as a iinesr mtab'don of the #(! x 4,
-
,
+
and orthonormality of the +(x - 2n), plus orthogonality with respect to the q5(x - 2n), #(I - 2n - 1) implies
4 x,
With the definition ml (0= g.e-"'(, the conditions (10.4.2), (10.4.4)(10.4.7)are equivalent to the unitarity of the matrix
This matrix looks identical to (10.2.1),but t bBimilarity is deceptive: in (10.4.8) the first two columns are both given by low-paes filters, because they are .both related to the scaling function 4 (mg(O) = 1 = 4(0)), whereas the second column in (10.2.1)corresponds to a hi&-pm filter. Such mi,ml am indeed be constructed (8% Auscher (1989)for detsfls snd,graphs). Note that 4 and rnj are closely related. The Fourier transform of '(10.4.1),(10.4.3)are
implying
I
,
325
GENERALIZATIONS AND TRICKS
4
which should hold for almoet d l (. 1f is continuow, then the following argument shows that vanishes on m e intervals. Since J(0) = ( 2 ~ ) - ' / ~there , exists a ao that for ICI a, 2 ( 2 ~ ) - ' / ~ / 2 .Consequently, for I(I < a,
8
<
+
since 4,mi are also 2?r-periodic, this implies d ( c + 2r) = 0 = 4 ( C 27r) for I(( a. It follows that li(J 3r)l = 0 for ICI a. In particular, this means that 4 cannot be com&tly supported (compact support for 4 m-that is entire, and non-trivial entire functions can only have isolated zeros). Neverthelese, subband filtering schemes with rational'noointeger dilation factors, in patticular with dilation $, have been propoeed and constructed by KovdeviC and Vetterli (1993), with FIR filters. The basic idea is simple: starting from cO,one can first decompose into three subbands, by means of a scheme as in 510.2, and then regroup the two lowest frequency bands by means of a synthesis filter corresponding to dilation 2; the result of this operation is cl, whili the third, highest frequency band after the first decomposition is d l . The corresponding block diagram is Figure 10.4. If all the filters are FIR, then the whole scheme is FIR as well. But didn't we just prove that there does not exist a multiresolution analysis for dilation factor with FIR filters? The solution to this paradax is that the block diagram above does not correspond9to the construction described earlier. A detailed snaiysie of F i r e 10.4 shows that this scheme uses two different functions 4' and 4a, with Vo generated by the 6' ( x - 2n), #2(x - Zn),n E 2. The argument used to prwe that 6 cannot have compact aupport then no longer applies? and d l , t$2 can indeed have compact support. The analog of (10.4.9) ie naw an equation relating the twedimensional vectors ($((), @(E)) and (bl($ I ) ,P ( $c)), however, and it is hard to see how to formulate conditions on the filtere that result in regularity of #', #=.
<
(+
<
4
One may weiibnder what the rationale is for t h w fractional dilation factors. The answer is that they may provide a sharper frquency loudhation. If the
,
326
CHAPTER 10
dilation factor is 2, then t,&is easentiaUy localized between r and 27r, as illustrated by the Fourier tramform of a "typical" 4[, in Figure 10.5. For some applications, it may be useful to have wavelet bases that have a bandwidth narrower than one octave, and fractional dilation wavelet bases are one poesible answer. A different m w e r is given in Coheq and Daubechiea (1890), eummarized in the next section.
no.10.5.
Mod&
of ~ l o & ( f ) l , with
N+
OJ
defied in fS.4.
10.6.
,Better frequency remlution: The splJttlng trick
Supp-
that h,,g, are the filter coefkienta assodrtad to an orthcrnormal baaia with dilation factor 2, i.e.,
wavelet
satieties and gn = (-1)"
h-*+I
.
Then we have the following lemma LEMMA10.5.1. Take any function f (not necessarily connected 20 wavelets in any wag) 80 that Me f (. n), n E Z a n orthonomud. Dujlne
-
GENERALXZATIONS AND TRICKS
Proof. 1. Since J d3: f (z) f (2
2
PI(€)=
x
- n) = &,o,
h,e-"h<)
=
we have
fi mo(t)/(~)-
(10.5.3)
n
Consequently,
d
= 2 (2r1-l [Imo(t)12 + Im(t+ r fPI
-
,r-I
(by (10.5.1))
.
(use (10.5.2) and (10.5.3))
This implies
The orthonormality of the F2(+- 2k) is proved analogously, using &(Q =
fi e-*€ mo(€ + r)f(O.
and
-
-
which proves that the Fi(z 2k) and the F2(x 2C) are orthogonal. 4. Finally, the Fl(-- Zk), F2(.- 2k) span all of E
GENERALIWTIONS AND TRICKS
Since the W1-eprrces are all dilated copies of WQ, we cart construct corresponding Wj, aad th& u d b Is i@rt a basis for L2(R) = @ W,.W e define thus
ortho~)N bases fm ~€2-
I
*;,t(4 - 2-j/2 91(2-fz, s), &(z) = 2-]I1 $2(2-'z the
{$jV1.,,
'
-24 ;
LZ(R).Since $ I ,
j , t E 2) constitute an o r t h o n o d bssis of
GZ
4, $I and & both have better frequency localization than $ itself (at the price of having larger supports in 2-space!). The splittipg up of hquency space corresponding to the W,on one hand, and the W1, W f on the a h a b d , ia repmmtd kbematbUy in Rigam 18.7. Note that &ueucy is &ibsated ll logarithmicaUy,even wi& (bs #$& are obtained by "splitting"
qSk.
ths .pEi#ing of Vo (conuponding morr o bandroidtA 4 s )inkr (a) WE,WS, and t-5 or No (b) wI,w:, W J , and Vaf i ~10.7. . S -
npnrmiation of
v,
or
lua to.
W ( E$(€I, ) i2(c) = ml(e) $(t).~tf o ~ h that 14'([)1'+(4~(€)1~ = 216(f)1" asillwtratedbq Figure10.8, which a h shows that $', $a do indeed "aplitn into two piece, corresponding to its 1-t and highest Ufrequency halves." Computing the melsdents of a fundb with resplft to the ,#,; E.D again be done via a subband filtering seheme, in which one additional etep of
BY -t~ction,
$I(€)
JZ
=
4
etr
M
'
1Q.8. h p h r of 11J1((5)\, i&'(€)(i luherre Loru-w fitkr Thc cd linc u the p p h of 4 I$(€)).
rquol to I@-(€),
dzfbrsd in 36.4.
high- aud low-paas splitting b added after the bbstandardnhigh-pass filtering. Schematidy, have now Figure 10.9. Note that the dilation factor 4 scheme propoeeid at the end of 510.2 (derived from a factor 2 scheme) aIso contains these functione $I, (The wavelets in that scheme are eseentially $(XI,the original wavelet for factor 2, and fi +l(2x), fi @ ( 2 x ) , with $I, +2 defined as above.) If one wwb with a tensor product muitiresolution analysis in higher dimensions, then the splitting trick can be applied >ina elective way. Figure 10.10 shorn how one can,e.g., use the splitting trick in two himemione to achieve an orthonormal wavelet basis with better angular resolution in the frequency plane than a "standard" wavelet basis. Figure 10.10a visualthe construction of $10.1 in the frequency plane: the small central square corresponds to (say) Vo; adding to it the two vertical rectangles companding to Wg = Wo@ Vo,the two horizontal rectangles for W$ = @I Wo and the four corner squares for W$= Wo@ WOleada to the bigger square representing V-1. The structure then repeate in the next annulus,to conetitute V-2. The angular m l u t i o n in the Fourier plape of this scheme is not very good, as shown by the figure. Figure 10.10b a h m whet the same twdimmsional wnstructioa looks like whan one starts from a onedimensional multiresolution d y s i with ~ dilation factor 4, a~ given at the end of 510.2. In thia case the onedimensional scheme haa dready three wavelets, so that the twodhwmiod product scheme ends up with 2 x 3 + 32 = 15 wavelets. Fie 10.10b repr~~ents one step (with &tion 4)
GENERALIZATIONS AND TRICKS
FIG. 10.9. Schematic npmsentation of t k d $ d jifienng openation# when workmg with Ydplitwwavckb.
in the rnultiresolution scale, as compared to two step (with dilation factor 2), i.e., two successive annuli in Figure 10.l&. The d r a l part of the two pictures is identical; the only diffe&nce between the tkro pictures is that the outer annulus of Figure 10.10a is split into many pieces to gSve Figure 10.10b, while.the inner annulus is untouched. This corresponds to a "spiitti* of one level out of two" in terms of the splitting trick lemma, ae pointed o i t above. The result ie good angular resoluhion for some wavelets +omsponding to the outer isyer in Figure 10.10b), bad for others (corresponding to the most central redanglea in Figure 10.lob). Figure 10.10~shows the same picture again, with two steps in a multireolution analysis with dilation factor 3, but for a product structure starting from the two 3 octave bandwidth wavelets constructed in this section rather than the one octave bandwidth wavelet $. The ion is the same, but t b e n are now 2 x 2 + 22 = 8 wavelets (aa o p p d ta 3 f a Figure 10.10a, and 15 for Figure 10.10b). Figure 1MOc can be obtained from F i e 10.10a by splitting every w u l u s ' (inner as well as outer) into halvea by cuts in both horizontal and vertical directions. This improves the angular resolution on the squares in the corners (corresponding to the W! of F i i 10.10a), but does nothing for the angular resolution of the rectangles ( c o r r e s p o to ~ W; a W" in Figure 2a), which were split better in the outer annulus & F i e 1O.lOb. T ie beet sngular resolution can be obtained by giving up the product etrudtufe altogether and just carving up every one of the W; spaces of Figure 10.10s vertically and/or horizontally, by applying the "splitting tridP in z and/or gy,until the desired resolution is achieved. An example is given in'~igure 10.10d. This etill curresponds to an orthonormal basis, and to a fmt algorithm for decompoeing and recon8truCting functions,although the mganhtion iR somewhat m e complex. If even better w g d ~ bresolution Q reqb& then one can repeat the eplitting tri& where and as many times as tucmmy. 8
t
'
10.6.
Wavelet packet bases.
The better re801ution warnlets d the pxwioua &ion are in fact ardy special msee in a h u t i f u l mnstruction d CMfimm 4hfeyer, called wavelot pncbts.
:
CHAPTER 10
FIG. 10 10 Vwwls&~m of the locducrtton an the Founer plane tuwdsmensiond muttamduhon ~clrtnrcs,as uplasnd m Ulc k t .
d s e v e d by v o w
The present section is only a short description of their construction; more details can be found in, e.g., Coifman, Meyer, and Wickerhauser (1W2), and applications to acoustic stgnab and images are given in Wickerhauser (I-, 1992). Start with a "usual"multiresolution analysis for factor 2, andconsider only the b,W,spaces with j 5 0. The decompsaition
cormpoads to a frequency splitting as ech&maticallyrepresented in F i e 10.7a. Heuristically, W-l has "twice the sizen of Vo and Wo(which have the %me size"), W-2 has "four times this aize," etc. One can ima@ne reducing all of them, via the splitting trick, to subspaas of the same sizq W-lgets eplit once, W 4 twice, etc. This c o m n c i s to defining a hast of W o r n $4, ,...,e,, where 4 stands for the W-rspsce (and for the number of aplittinga this or rnl, at the jth q m x b undixg~e),end c, = 0 or 1 hdicatea t b choice,
.,
GENERALIZATIONS AND TRICKS
splitting. Explicitly,
-
Clearly, the $ c , ~ ,, +&(z)axe all linear combiiations of the $ ( p z - k), and the splitting trick lemma (applied e times) proves that {$c,rl,. , r c ( - - n); €1, - - - ,EC = 0 or 1, n E Z) is an orthonormal basis of Span {+(2c - -k ) ; k cz Z) = W-c. It follows that {h,,,, ,e,(- - n); I E N, n E Z,CI,-.. ,ct = 0 or 1) U {+(. - n); n E Z) is an orthonormal basis for L2(R).Note that this basis corresponds to the integer translates of functions which all have more or less equivalent frequency localization (in bands of width more or less A, starting with 5T for the 4(. - n), .rr 5 I 27r for $(- - n), .).l1 This is very similar in flavor to the windowed Fourier transform and the Wilson bases of $4.2.B, while still having the same ease of computation, via subband filtering schemes, as wavelet bW5. There are of course many intermediate solutions between wavelets on one hand and the wavelet packet basis described above: one may choose to split some W-c spaces less often, or to split some of its subspaces more than others. Every such choice corresponds to an orthonormal basis; moreover, there exist efficient algorithms (based on computations of the "entropy of a function" with respect to the different splittings) to determine which of this whole library of choices is the most efficient for given signal; see e.g., Coifman and Wickerhauser (1992).
10.7. Wavelet bases on an interval.
AD the one-dimensional wavelet constructions we have d i s c 4 so far lead to
'
bases for L2(R). In many applications one is interested in only part of the real line: numerid analysis computations generally work on an interval, images are concentrated on rectangles, many systems to analyze sound divide it in chunks. All tbp involve decompositions of functions f supported on an interval, say [O, 11. could, of course, decide tb use standard wavelet bases to d y z e f, mi%the function equal to zero outside (0,11, but this introduces an artificial "jump" at tBe edges, reflected in the wavelet coefficients.l2 It is, mareover, not efEcient c o m p u ~ It. is therefoe useful to develop wavelets adapted to "life on an interval." A first way of achkvhg this is to use the periodized wavelets described in 59.3. These are camp-dy qcient, but using them amou~ftsto analyzing the periodized function f , defined by f(z) = f (2 - 1x1) (where 1sj denotes the largest integer not exceeding z),by means of the usual (non-periodid) wavelets. Unless f is already periodic, we have again introduced a "jumpn at
334
CHAPTER 10
the boundaries 0 , l ,'which will be reflected by large fine scale wavelet coefficients near 0 and 1 . There exists another solution which does not present this inconvenience, proposed by Meyer (1992),based on orthonormal wavelet bases with compact s u p port. In this construction, the wavelets with a support contained in [O,l] that does not reach either 0 or 1 are kept as are, but this family is supplemented at the edges by specially adapted functions. Let us illustrate how the idea works on the half line rather than an interval. This simplification allows us to deal with only one boundary, and to disregard the adaptations needed at the coarsest scales, where both boundaries of [0,11 have to be dealt with simdtaneously. Define
vFf 3 =
span
{@y2r;k g Z) .
Vdf
The space can also be viewed as the space' of all restrictions to [O, oo) of functions in 5. If we suppose that the original scaling function t$ has support [O,2N - 11, then e,$'(x) = 0 if k < -2N + 1; we therefore only consider the r#~;,$' with k > -2N 1. All but 2N - 2 of these are untouched by the restricting procedure: 4;,$f(~) = t $ ~ ~ , ~if( kx j2 0; these functions are therefore . still orthonormal. The 2N - 2 functions & s f ,k = - 1 , - . - ,-(2N - 2) are ind& pendent of each other and of the 4,,k with k 0 . Define now wJef to be the orthogonal complement in of Vjha". If, for convenience, we shift 11 so that it is also supported on (0,2N - 11, then the (restrictions of $j,k to [O, oo)) are clearly in W;df if k 2 0 , since they are orthogonal to all the &!', and lie in What about the with k = - 1 , . .- ,-(2N - 2)? ( I f k is even smaller, k < -2N + 1 , then = 0.) It turns out (see Meyer (1992)) that the +$ff with k = - N , - ( N + I ) , - - ,. -(2N - 2) are in qhdf, i.e., they are orthogonal to w,ef. The other $;ff, k = -1,. - .,-(N - 1) contribute to W,w$in fact, we ha- that { k 2 -(2N - 2 ) ) U k 2 -N(-1))
+
V,yf
%fqf.
I I
I
1
jl
>
$kgf
~!JF tjkf
{$iff;
qY.l3In order to orthonormalizethis basis, one
is a ($n-orthogonal) basis for proceeds in the following steps:
&sf,
(1)O r t h o n o d m the k = -1, - . ,-(2N - 2). The resulting functions &, k = -1, ,.. ,-(2N-2) are automstically orthogonal to the &,k, k 2 0, and together they provide an orthonormd basis for If we define
v,-.
(
( x )= 2
>
( x ) , j s 2,k = -1,. .. ,-(2N - 2 ) ,
0) U {&,,k; k = - 1 , - - . , - ( 2 N then { 4 j , k ; k basis for 9- for any j E 2.
- 2 ) ) is an
orthonormal
(2) Project the+&',
k = -1,...,-(N - I) ants ~ ~ ~ & i & i x t g , d < )& k3*
IN-2
-7%
%-
(3) Orthonormalize the JI: The d t i w &, k = -1, ,-(N - I), ~ h with the +t,ff,k 2 0, provide an orthonord bash for WOhrtf.We can again define a a a
.
then (4 , k ; k = - 1,. - ,-(N - 1))Cl {Jttek; k 2 0) is an orthonomd h i s for w$. The union of all these b.as (jranging over Z) givea a b.ds for L2(l0,a)). The resulting beses me not only orthonormal basea for La([O,m)), but provide also unconditional bases for the Holder spaces restricted to the half-line (i.e , they even handle regularity properties at 0 "correctly"), etc.; see Meyer 71992) for pro&. To implement all thie in practice, one needs to compute extra filter coefficients at the boundaries, corresponding to the expansion of the k= -l,...,-(N - I), &,kt k = -I,.--,-(2N 2), in terms of the &-l,c, .t = -1,-.-,-(2N -2) and the 1 = 0 , . . - , 4 N - 5 . Thesecan becomputed &om the origrnal ht; tables are given in Cohen, Daubechies, and Vial (1992); this paper el%a contains an alternative to Meyer's canetruction, involving fewer additional functions at the edges (only N imtmd of 2N -2), while still handling the regularity properties correctly, even at the edge. One last remark about wavelet bases on an idetval. h image analpis, it is customary to treat border effects by extending the image, beyond t,he barder, by its reflection: this extension avoids the discontinuity that follows from periodization or extending by zero (although there still ie a discontinuity in the dsrimtive, however). It is well known that thie means that border e f k t s are .. mmlmlaed, and that no extra coefficients (to deal with the borders) haw to be introduced, provided that the filters wed are symmetric. The same trick can be used to provide biorthogonal wavelet bases on [O,l], with much less &ort than the orthonord wavelet W i on the interval of Meyer (1992) or Cohen, Daubechies, and Vial (1992). If f is a function on R, then we can define a function on [O,1] by "fold- ' in< its graph at 0 and 1. The first "fold," at 0, amounte to replacing f(s) by f ( x ) f (-x). Folding back the two tails (one from the original f , the other from the folded over negative part) eticking out beyond 1 legde to f (x)+ f (-z) f(2 - z) f(z 2). If we keep folding like this, theP we end up with
-
e
.
+
+
+
+
Note, h r later convenience, that"
a
330
CHAPTER 10
Talre now $J,4, $wo wavelets which generate biorthogod &let
baea of
la(^);
#,J, as codtructed in $8.3; aseume also tha$ t$,d are symmetric around 4, 4(1- z) = r#(z),i(l- 2 ) = a x ) , and that qb, & am antbymmetric around f, $(I - r ) = -$(z), $(l - r) = -d(z). (Exampleq
wi& .misted scaling b c t i o n s
were constructed in 88.3.) Apply the "folding" technique to the t,bjtb and $,,k, 6
$51'
is defined .odogou81yYWe will r~triiotour attention to j with J 5 0, for which (10.7.2) can be rewritten as
Fbr good measure we dm de0m #', &(I x ) , we find
-
$?Yk
4;T;
< 0, or j
since flz) = #(I
= -J
- x ) , &x) =
Obvlnuly, $ ~ ~ L + , ~ . = , for rn E Z, so that we only need to consider the durn k = 0, - ,zJ+' 1. Monwer, = &&,whih mema 1. A similar argument sharve we can rest& orrmelves to only k = 0, ... , that we only need to consider the $3tk for k = 0, . ,2J - 1. RaznmlgMy, the and @yk,, with 0 5 k, kt 5 zJ - 1, are still biorthogond on [O, 1). To prove this, use (10.7.1):
-
mk
-
q5t'&J+l-k-
-
@yk,
-
Thb biorthogonality implies, w n g other things,that the k = 0, ..,2' I Mdl independat, and proride a tmi~for v!' = {jbld ; f E V-J} . (The
GENERALIZATIONS AND TRICK8
-
@Jd
W Y , ~=
r m e is tm En the #:$".) We can .Lo define ~p.ca by I/M; f E W - J ) .ObviowiY,w?jd isspanned by t h e m k , k=&-*-,@M w e o m , computstiona similar to (10.7.4) sbow that b
337 4.
1.
v'?
pmving that W?Jd I and that the &"j',,, 0 5 k 5 2J - 1 are independent. It follows that the. "foldedn structures iaherKt all the properties (nesting of the spaces, biorthogonelity, basis properties,. ) from the unfolded originals. The filter coefficientscorresponding to these folded Morthog6nal basecr are Likewise obtained by folding at the edges corxe8pondii to s = 0,1; if $,$,4,4 are compactly supported, then only the filter coefficients new the borders will be a f h t d . Examples are given in Cohen, Daubdies, and Vial (1992). Bemuse analyzing f on [O, 11with these folded biorthogonal wavelets amounts to the same as extending f to all of R by reflections and analyzing this extension with the oiigiaal biotthogonal wavelets, we cannot hope, however, to characterize Holder spaces on [O, 11 beyond Holder exponent 1 with this technique. Thiil is progrem with respect to what periodized wavelets can do, but it 18 less petfbrmant than the o r t h o d wavelet baaee on p,1).Fa'more detaila: see %hen, Daubdies, and Vial (1992). - 4
"
:
-
Notes. .'
-
1. One example is the foilowing. Let r be the hexagonal lattice, = .(me1 n2e3; nl, ns E Z),where el = (l,O), es = (1/2, &/2); r d~+ a partition of R' into equilsteral triangles. Define % to be the space of continuous functions in L2(R) that are pieceftriee aEm on theee triangles The urkhonmal besia fm this muitireaWion d y s i a is m d x w t e d in J&d (1989). Biurthogond beeerr of compactly supported mveleta with this hexagonal symmetry ate construct& in Cohtm and Schlenkm (1091).
+
\*
2, lb me-dimensional conditions of 55.1 can also be crrclt in a matrix i~ that. use d(t) = ml(tfi) &/2), .nd tb. c~~~c&tion% are = 1, Im1(€)12+ Im1(€ + 412= 1, mo(0 m1(0+ ensuring the orthonormality of the (h,n; i ,E Z), of th*m,& rn5 E Z) and the orthogonality of these ttoa Waf vectors, xwpectissl@ But t h e e conditiom b e equivalent to the r&@remmt that
;mL
is
3. If one prefers to index the entrim df U with the n u m h 1,. .,2" rather than entrim of (0, I)", then one clrn mumber 8 E {O, lIn by defining @ = 1+ q22-l E {I,.* -,r).
x;-l
4. At present, I know of no explicit scheme that provides an infinite family uf w ,for dilation factor 3, with regularity .pawing proportionally to the filter support width.
5. The same can be done for dilation factor 2. where the factor matrices are even simpler. The bwic idea is that if im(i)12 +lmo((+r)l2 = 1, then, for any y E R, n E Z,&(() = (1 $)-'I2 [ ~ ( f +) ye-i(2n+1)(tq-,(
+
+
L
will also satisfy I ~ $ ( ( ) I ~ + I ~ ( ( + * = ) ~1.' then it is convenient to & m e n = N
xE'10'43e - ' * .
2N+l ane-'Y, Urn(() = C=o
+ 1,
Ieading to tr$(c)
.)
=
All this can be rewritten in matrix to&:
The whole operation raises the degree ,of @ 2.. One can, moreuver, prove (Vaidyanathan and Homg (1988)) that my trigonometric polync~ mial mq satisfying Imo(€1lz Iw(<+ r )l2 = 1 .can be obtained by letting products of such 7-matrices act on two-tap filters. We have not used this method in our const~ctiomin 93.4 becauae it does not preserve the divisibility of m~ by ( I + e-e): impoeing that the final is divisible by (1 e-*) leads to highly nonlinear constraints on the parameters yj. The method hae the advantage, however, that it leada to easy implementations of the filters (wing the 3 directly), and that roundoff emam on the yj do not mar the exact reconstruction property. Anymy, a &nih Rimilnt technique can be d Eor more than two channels, as ahown in lh&m&a, Vsidysnathan, and Nguyen (1988) or, with mop prsdid xnatri%factors, in Vddyanathan et d. (1989). Them matrix hctorhtion tediniquere go back to the circuit theory work of Bele-
+
+
vitch (1968).
6, TI& method for determining ml, m was indicated to me by W. Lawton and R Gopinath (personal c o d c a t h , 1990). t 7. Such a onedimembnd filter would not be uaefui in practice, hcmmr! 8. This haa been obeerved by maay authors. The oldest reference eeems to be McClellan (1973). One csn also, replace the onedimensional am( by (acast (1 a)coe C)/2 for a E R arbitrary, but there seema little point in not making the symmetric choice a = f .
+ -
9. One can argue that some of the higher-dbmio$ schemea dbcuseed in f 10.3 correspond to noninteger d i l a t i i . In two dimensions, fw artample, and Drr = -1) both satisfy @ .t 16 Id, the matricee Dl =
(i -1)
(l
-
GEN$XtALIZATIONS AND TRICKS
so that a single dilation can be viewed as a dilation by a rotation and/or a reflection).
dr
* 4
r
339
f i (combined with
10. If one slao considers orthonormal wavelet not resulting from multires olution analysis, then it is not known whether irrational dilation factors me allowed or not.
11. For large C, the concentration of 4t,,,,..,+(is not as good as this discussion suggests, however; see Coifman, Meyer, and W i r h a u s e r (1992). This is already noticeable on Figure 10.8, where &2 have "sidelobes."
tiR,
12. In image analysis one often extends f beyond the boundaries of the image by reflection; this extension is continuous, but its derivative still has a "jump." We come back to this at the end of $10.7. .(
13. Some of these assertions are highly non-trivial! A sizeable part of Meyer (1992)is devoted to their proof; simpler proofs have recently been found by Lemari6 and Malgouyres (1991). 14. We have
-
"i. 2 3
7;%
References
34
Published pspar and papem that are &ukd ba pubkat60n sn labelled by the yesr in which they appamd or will appear;other pmprhb u e hbeW by the yest in which they were written. Tbis may play havoc with tho chronological , but then, due to d i n c e s in publicstion oped ofMerent journde, BO may Otiicid publication dates. M. ANTOWINI, M. BARLAUD, AND P. MATHIEU (1991), Image d a n g wsng fanace vector qwntktron of wavelet m e w , Proc. IEEE Internat. Conf. Acouat. Signal Speech Procam., pp. 2273-IM76. M. ANTONINI, hi. BARLAUD, P. MATHIEU, AND t. D A ~ I E(lwl), S finage coding wsng umaelet tmmfowna, IEEE lkans. Acoust. S i g d Speech Process., to appear. F AROOUL, A. A-Do,
J ELBZOARAY, G. G
tro~~fawn of two-domc~lonol+to1
~ u AND. R. MURENZ~ (1989), Wavclct oggrrgoto, Phyk Lett. A, 135, pp. 327-336.
F. Amout, A. A R N ~ D OG. , GRASSEAU,Y. Glsarrr, E. S. H O P F I N GAND ~ , U. FRISCH (1@0B), Wavelet andym of turbrJenoe m v d o the mJtifnrctal nuturn of the Rachanbon cororrdc, Nature, 338, pp 51-53. A ARNBODO,F. A ~ O U L ,J. ELEZQARAY, -D G. WMU Wavdet tmmfarm . . phase 11988), o w of jmctdr: Appliaati4n to tmnritMnr, in Nmliiieur Dynamh, G. 'hmhetti, ed., World !%eat&, S i , p. 130.
E. W. ASLAKSEN AND J. Ft. KLAUDBR (1W18), U d m y ~
e ~of the taBne a Iprorup, J. ~ Math. *., 9, pp. 206-211; we as0 C o n t h w ~ t~prrermtotumthcmy arang the afine gmup, 3- Math. Php., 10 (1989). pp. BiW-227S.
P. A U S (1989), ~ Odeletto D a u p b , Paricr, Raoce.
ct
w, Ph.D, Thesie, Univer6it6 Paris,
-(1990),S~trl(propcrtiuf ~ r w ~ k r u d r u c a u a o m p k r w i t h c ~ m , p o e t m p p o d , preprint, UnivemW de brim, Rasros, in W&: Mathemati- and Applicatii, J. b a d s t t o and M.Ftaziar, edik, CRG Prar,to appear.
-(1992), Wavdct
boru for
h2(R),
mliond aEildian factor, in Rwkai et d. (1992).
pp. 439-452.
P. A V S C H ~G.RWIRSS, , AND M-V. W l m ~ g (m), g ~ L d .ine and C O bor ~u of Cd&m 4M e ~ c r a ~ * ~ H m n & h t w o d e bin, Chui(l9fQb).