View metadata, citation and similar papers at core.ac.

uk brought to you by CORE

provided by Elsevier - Publisher Connector

Computers Math. Applic. Vol. 28, No. 4, pp. 25-39, 1994

Copyright@1994 Elsevier Science Ltd
Printed in Great Britain. All rights reserved
08981221/94-87.00 + 0.00

Families of Biorthogonal Wavelets

A. KAROUI AND R. VAILLANCOURT
Department of Mathematics, University of Ottawa
Ottawa, ON, Canada KIN 6N5
rxvsg@acadvml.uottawa.ca

(Received January 1994; accepted February 1994)

Abstract-Several families of biorthogonal wavelet bases are constructed with various properties.
In particular, for a given filter, F,,(t), of finite length 2n + 1, a parametric family of dual filters,
F;;(E), of length 2N + 1 is constructed. The parametric nature of the dual filters makes it possible
to design the optimum dual filter ;FN” (5) corresponding to a fixed filter 3,,(t).

1. INTRODUCTION
Recently, the growing interest in orthogonal wavelets is due, in great part, to their ability to
represent wide classes of functions and operators without redundancy, and to the fast wavelet
transform which makes possible the computer implementation of new, very efficient algorithms.
In pure mathematics, wavelets are used to characterize some functional spaces, such as P(P)
for 0 < p < 00, Holder and Hardy spaces, etc., [l]. A proof by means of wavelets of the famous
T(1) theorem of David and Journe [2] on the L2-continuity of a class of linear singular integral
operators is found in [3, pp. 267-2781.
In numerical analysis, wavelets are used as an efficient tool for the rapid numerical application
of certain types of linear operators to arbitrary vector-valued functions [4]. They are also used
in the numerical solution of partial differential equations by mean of finite element methods
[l, pp. 57-601. They are widely and efficiently applied in engineering, for example, in sound
analysis [5], image processing [6-81 to cite but a few.
An orthonormal wavelet basis for L2(R) is a family of functions

r&(z) = 2-j&j (2% - “) , z E w, j, k E z, (1.1)

obtained by dilations and translations of a single (mother) wavelet II, E L2(W). Thus, any
function f in L2(W) can be expressed in terms of the wavelets $j,k:

(1.2)

where the equality holds in the strong L2-topology and the wavelet coefficients are given by the
scalar products

(f, ‘$‘jk) = Jrn f(z) ‘d’jk(s) dx* (1.3)

-00
This work was supported in part by the Natural Sciences and Engineering Besearcb Council of Canada under
Grant A 7691 and the Centre de recherches mathematiques of the Universite de Montreal. The authors would like
to express their thanks to practicing electrical engineer B. Karoui for his indispensable and continued assistance
with programming and computer expertise during the preparation of this work.

ww m4-c 25
 26 A.Kamm ANDRVAILLANCOUFLT

The (nonsmooth) Haar basis [9] constitutes the first known wavelets. J. 0. Stromberg [lo]
constructed the first orthonormal basis of the form (1.1) with a function $ of class Cm for an
arbitray integer m. In 1985, Y. Meyer [ll] constructed an orthonormal wavelet basis which
is an unconditional basis for various functional spaces. In [12], P. G. Lemarie constructed a
wavelet basis for J2(Wn) with bounded regularity, but with exponential decay. In 1988, Ingrid
Daubechies [13] constructed orthonormal wavelet bases with compact support and arbitrarily
high regularity.
We remark that the construction of the majority of useful wavelet bases is a consequence of the
design of some 2n-periodic functions called wavelet filters by the signal processing community
(see [14] in an early stage, and [15,16]. Moreover, any wavelet filter with finite or infinite length
is a finite impulse response (FIR) or infinite impulse response (IIR) filter, respectively.
In many applications [17, p. 113; 181, it is necessary to use linear phase FIR filters. Unfor-
tunately, such filters are impossible to design. On the other hand, biorthogonal wavelet bases
provide us with compactly supported symmetric wavelets [19]. Biorthogonal wavelets are formed
by a pair of families of dual (see Definition 2 below) wavelets, &k(z) and &k(z), derived from
two mother wavelets, $J(z) and J(z), respectively, and such that any function f in L2(R) can be
written in either forms:

(1.4)

Hence, f is decomposed by one family and reconstructed by the other. Ph. Tchamitchian [20]
constructed the first family of biorthogonal wavelets. In [20], it is shown that it is possible to
construct symmetric biorthogonal wavelet bases with arbitrary high preassigned regularity.
In this work, biorthogonal wavelet bases are constructed by an approach which differs from the
one used in [19]. By this new approach, it is possible to construct a new class of biorthogonal
wavelet bases with the following remarkable properties:
0 symmetry,
0 compact support,
0 regularity,
l the dual filter, corresponding to a fixed wavelet filter &(<), is given in parametric form.
This paper is divided as follows. In Section 2, some necessary or sufficient conditions are stated
for the construction of regular wavelets. In Section 3, we provide the numerical techniques for
the construction of the wavelets and estimating their regularities and to extend a fixed family of
biorthogonal wavelet filters to an infinite family. Numerical results are quoted in Section 4.

2. BIORTHOGONAL WAVELET BASES

In this section, we describe the fundamental steps of the construction of wavelet bases. The
first step consists in designing a wavelet filter and constructing a scaling function. In the second
step, the wavelets are constructed.

2.1. Preliminaries

An orthonormal wavelet basis {‘$j,k; is directly related to a multiresol,ution analysis

j, k E Z)

(MRA) [21,22]. Let V. E L2(W) be the subspace spanned by the orthonormal functions 4(z - k),
k E Z. Define the space Vj obtained by dilating Vi by 2j,

f E Vj * f (2j’) E VI. (2.1)

An orthonormal basis of vj is given by {&k; k E Z}, with &k(z) = 2-ji2@ (2-j, - k). Then,

. *. c v2 c Vl c vi c v-1 c v-2 c - ** , (2.2)

 Familiesof BiorthogonalWavelets 27

and
fi tj- = {O},
-Xl -00
Since VI c VO,there exists a sequence, (cx~)~Ez, of complex numbers [23] such that the ftinction
C$(~-IS) E VI, by (2.1), satisfies the two-scale difference equation

(2.3)

where ai, = O(lnlVm) for any integer m 2 1. The Fourier transform of the first expression in
(2.3) is

The wavelet $J is defined by

r&2.9 = Ccz,ei*C
[?VSZ 1 6((E). (2.4)

$(z) = 2 C(-l)%l-&$(22 - n). (2.5)

nEZ

Let W’ be the orthogonal complement of 4 in q+,. It is shown in [l, pp. 71-731, that @on(z) =
Q!J(Z- n), n E Z, is an orthonormal basis of WO. It then follows that, for fixed j E Z, the sequence
(+&~Ez d&d as in (1.1) is an orthonormal basis of Wj. Moreover, UjCz Wj is dense in L2(W).

DEFINITION1. Scaling functions, orthornormd wavelets and wavelet filters are defined as follows:
(a) the functions (4jk)j,kEZ are called the scaling functions generated from the (father) scaling
function 4(z),
(b) the functions ($jk)j,kEZ are the orthonormal wavelets generated from the mother wavelet
ti(x),
(c) Fo(5) = fi (C,,, on einc) is a wavelet filter where the CX,are given by (2.3).
We shall use the following definition [24, p. 1511.
DEFINITION2. Two scaling functions q5 and 8, generating possibly different multiresolution
analyses of J?(W), are said to be dual scaling functions if their scalar product satisfy the following
condition:

(4(*-j),$(*-k)) :=Srn 4(X-j)J(X-lC)dX =Sj,k, j, k E Z. (2.6)

--oo
In the biorthogonal case, we shall need two dual scaling functions 4(x), F(z) satisfying

4(x) = 2 c %~(2X - n), S(x) = 2 C&&(2x - n). (2.7)

6Z NZ

By using condition (2.7) and the techniques employed for proving Theorem 1 in [l, pp. 72-731,
the reader can easily verify that the pair of functions
$(x) = 2 C(-l)“z&(22 - n), G(z) = 2 C(-l)nal_nJ(2s - n), (2.8)
nEZ nEZ

are (in general, nonorthonormal) dual wavelets assocated with the scaling functions 4(x), J(z).
We remark that if the scaling functions, $jk and $bjk, and their corresponding wavelets, ‘$jk
and $jkl are defined as in (l-l), then any f E L2(W), can be written in the forms

f(X) = c(h d’jk)&jk(X) + T;: x(h ‘hk&k(X)

kEZ nEZ kcZ
n<j
(2.9)
= c (f~&jk) d'jk(x) +c c (.f>6nk) ?hak(x)-
kEZ nEZ kEZ
n<j
 28 A. KAFL~UIAND FL VAILLANCOURT

By taking the Fourier transform of 4(x/2) and &z/2), we see by (2.7) that there exist a pair of
2x-periodic functions,

7710(E)
= C a,einE, (2.10)
nEZ

such that
2 ^,

VW) = mo(Ek#G)7 WE) = mowto (2.11)

Then the following important question arises: under what conditions on ms and &-, can one have
a pair of dual scaling functions and consequently a biorthogonal wavelet basis. An answer, given
in [19], is briefly summarized in the following subsection.

2.2. Necessary and Sufficient Conditions for the Existence of Biorthogonal Wavelet
Bases

From now on, we assume that the coefficients, on and &, of ms and 60, respectively, as
defined in (2.7), are real, satisfy the symmetry relations CY_, = on and i& = &, and are finite
in number. The last assumption is equivalent to the compact support property of the constructed
wavelets.
To construct a pair of dual scaling functions leading to a biorthogonal family of wavelets, we use
an efficient method (see [19]), which results from the biorthogonal version of the multiresolution
analysis. This method is essentially based on (2.4) which implies that

&) = fi m0 (2-j<), g(t) = fi GIe (2-Q). (2.12)

j=l j=l

To obtain scaling functions that lead to regular biorthogonal wavelets, the functions me(c) and
Co(E) have to satisfy certain conditions.
In [19], a set of conditions is provided on the dual 2n-periodic functions ms(c) and GO(<) and
consequently on the corresponding filters which are given, respectively, by

These conditions insure that our biorthogonal wavelet bases have preassigned regularities.
For completeness, we b_riefly summarize these conditions. It follows from the biorthogonality
conditon (2.6) on d, and 4, that mc and fig satisfy the identity:

m0(E)G(3 + m0(E + n)GotC+ r) = 1, vc E [O,nl. (2.13)

Hence, the coefficients (I~ and & satisfy the relation

c cYn(Yn = 1. (2.14)

Moreover, (2.12) implies that r$ and & are in L2(W) only if

mo(0) = ffio(O) = 1. (2.15)

On the other hand, if the scaling functions are to be continuous, it is necessary that me([) and
GO(~) vanish at < = TIT:
mc(7r) = %0(7r) = 0. (2.16)

Note that conditions (2.13) and (2.14) imply neither the biorthogonality of the scaling functions
nor that these are in L2(W). However, a positive answer is provided by the following proposition.
 Families of Biorthogonal Wavelets 29

PROPOSITION. Assume that both ma(c) and iiio(t) can be factored in the form:

l+e-iS L E
1 + e+ S(E),
(
mo(O= -j-
> f (C)9 Go(<)= -y-
( > (2.17)

where L, z > 1, and suppose, that for some k, k > 0,

Bk = s$f(~)f(2+ f (2k-‘<))1’k < 2L-1’2, (2.18)

(2.19)

Then, r$, & E L2(R) and

m 4(x)&(x - VI)dx = So,.
J -IX

PROOF. See [19]. I

It has been shown (see [19]) that if (2.18) and (2.19) are satisfied, then there exist two positive
numbers, E, ?> 0, and a positive constant c such that

L+hJ(Bk)/ b&)-E
< c (1 + El)- > (2.20)

L+log(Bb)/ leg(2)-z
< c (1 + ICI)- (2.21)

Now, by Theorem 3.8 in [19], if mc(5) and i;ic(<) satisfy (2.13) and if (2.20) and (2.21) are
satisfied, then the dual wavelets constructed from the scaling functions d(z) and F(z) generate
two biorthogonal wavelet bases, in the sense that any f E L2(W) can be written in both forms:

f(x) = c c(f, +jk)i&kb) = r;: x (f, &jk) @jk(+ (2.22)

jEZ kEZ jG3 kGZ

Thus, the problem of constructing a biorthogonal wavelet basis has been reduced to the much
easier problem of constructing 27r-periodic functions that satisfy conditions (2.13), (2.15), (2.18),
and (2.19).
REMARK 1. If there exists a constant c > 0 such that (2.20) and (2.21) are satisfied, then 4 and 4
belong to the Holder spaces F(R) and Cz(R), respectively, for all c < L - 1 - log( Bk)/ log 2 and
Z<L-1-log(B&log2.
In the previous sections, we have seen the basic theoretical steps for the construction of
biorthogonal wavelet bases; however, our actual construction relies on the special approxima-
tion techniques described in the next section.

3. NUMERICAL TECHNIQUES FOR THE CONSTRUCTION

OF BIORTHOGONAL WAVELETS
The aim of this section is to provide and justify the numerical techniques used in the construc-
tion of biorthogonal wavelet bases. In particular, we shall prove that it is possible to construct
a pair of dual filters, such that the coefficients of one of them are given in parametric form. The
idea of our numerical method for the construction of dual trigonometric polynomials is based on
two major steps. The first step consists in the construction of a function, ms(,$), which satis-
fies the conditions of Section 2. In the second step, a dual function, GO(~), is determined in a
straightforward way. Both steps are analyzed in detail in the following subsections.
 30 A. KAROUIANDR. VAILLANCOURT

3.1.The Construction of mo(<)

By using a result from [13], one can easily see that if a 2n-periodic function, mo(<), satisfies
the first parts of (2.17), (2.15), and (2.16), respectively, that is

mo(<) = (q+F), mo(0) = 1, mo(?T) = 0, (3.1)

and (2.18) for some k > 0, then mo(<) is a candidate for generating a biorthogonal wavelet basis.
Hence, let no 2 1 be a positive integer and consider the function

(3.2)

We note that the symmetry of mo(<), that is mo(-<) = mo(<), implies the symmetry of the
associated wavelet. To have symmetric wavelets, we require that the coefficients on satisfy the
following relations:
cr, = CL,, 1 _< n I no. (3.3)

From condition (3.1) at t = 0 and X, we derive the pair linear equations:

cYo+2~cY,=1, cX(J+ 2 2(-l)“an = 0, (3.4)

n=l n=l

in the no + 1 unknowns c~j, j = 0, 1, . . . , no. By fixing no - 1 of these, one obtains a unique

solution to the system.
It remains to verify condition (2.18). This verification, in general, is not easy and cannot be
done explicitly. Hence, one resorts to numerical methods to find a good approximation to the
upper bound of Bk.
We remark that the problem simplifies considerably if, instead of estimating &, we estimate
the maximum of the absolute value of a piecewise polynomial which approximates the function
(f(M2E) . . . f (2k-‘J) 1. Th is method, which turns out to be very efficient in our case, is given
by the following theorem.

THEOREM 1. Consider a 2x-periodic function,

,M, R, -n<jIN. (3.5)

aj
Qj E

Suppose that mo(0) = 1 and ~-LO(<)can be factored in the form

l+eiE Lf(E),
mow= ( > 7j-- L 2 1.

If we write
FG) = f(lV(20 . * . f (2”~‘J) 7 (3.7)

then, for all e > 0 and k > 0, there exist a positive integer r and a finite partition of [0,27r], say
(Ii)iGI, such that
“;P IFk(t>?‘k - “YP 1% (Epl < E,
I

where, for each i, the function pF,(<) is equal to a polynomial of degree r if < E Ii, and 0
otherwise.
 Families of Biorthogonal Waveleta 31

PROOF. The proof consists in three parts.

(a) In the first part, if

Q&OS<) =ag+alcos~+...+a,cosn~, al4 E R O<_U<?J,

we obtain an estimate of the upper bound of supe jQn(cos<)l. For v = 1,2,. , . ,D, we
expand cos V< in a Taylor series around a point &, to be fixed later. Thus, we have

where

Since Q,(cos<) is 2n-periodic, it suffices to find the supremum over the interval [O,24:

Because
([274+W
[0,24 = u &a, I,= [?$s,y,

m=l

where [r] denotes the integer part of the real number r, it follows that

Now, for a fixed m, 1 2 m 5 ([Znn] f 1)/2, let cm = (2m - 1)/n. Then, for c E 1,, a
Taylor expansion of order r of Qa(cos <) around cm gives

*
Qn(cos5) = a~ + c a, c’
US1
Iu(t - tm)l’
j!
cos’j’(Em) + J&+1(S)
j=l

= PQ”,
(0 + Ja). (3.8)

(b) In the second part, if I’m,, is a function whose restriction on each Im is equal to the
polynomial P$=(<) and 0 outside, then it is clear that

Since, me(<) = (w>, f(t), then necessarily

mm =“c”a
+2 c PiPj COS((i
- .dC>.
--n -nsi, j$N-L
6%

Since f(c)f(E) is real and symmetric in e“c, then, with A4 = N - L + n, we have

f(S)fo = 5
js-A4
Tj eic”9 c-j = -Cj, 7-j =rj, 1 <j<M.
32 A. KAROUI AND R. VAILLANCOURT

By the first part of the proof, there exists a function Pf(z), associated with a finite
partition, (I,),, of [0,2n], such that

Since f(0) = 1, then sup< f(e) 2 1. Thus,

(c) In the third and last part, we prove the general case. Take any positive integer k, k 1 1,
and define the function Fk(c) by the product:

fi(<) = f(6) f(2<) *. . f(2k-‘t).

The reader can easily check that if K = (1 + 2”-‘) 2k-2, then

Fk(<)F(c) = F PjeidjE, d-j = -dj, /3-j = /3j, 1 < j 5 KM.

j=-KM

By choosing the partition

((2KMr]+1)/2

[vq = u 4,
l=l

and using the techniques of the first part of the proof, we construct a function PF, , whose
restriction on each of the intervals 1l is a polynomial of degree T-, and which is equal to
zero outside [0,24. Moreover, PF, (<) satisfies:

Again, since f(0) = 1, then supC ]Fk(<)] 1 1, which implies the following inequalities:

sip IpF,(<)I 2 s:P ]Fk(r)]2 +

IPjl
“c”
j=_KM lT + ‘>!

1
KM

sup ]F&)12 1+
1 IPjl
2
c suP(IFk(t)i2j=_KM
c CT+ I>! (3.9)
and

[s;p ll-%k(<)l] 1’2k - [s;P]Fk(<)]] II*1

5 [s!p]Fk(a]l’k{ [Itsup<,;k(<)]2j~M &]1’2k-1}

KM
1
5 const 1+ (3.10)
s’4’e!&&>I2j=_K,,,
c CT+ l)!

Finally, since mc(0) = 1 implies that sup< ]Fk(<)] 2 1, then the right-hand side of (3.10)
can be made arbitrarily small by choosing r arbitrarily big. I
 Families of Biorthogonal Wavelets 33

REMARK 2. AN APPROXIMATION TECHNIQUE. In practice, it is convenient to discretize the

function PF, (c) as follows. For a given finite partition,
(li)i, of [0,27r], the restriction of PF, (5)
on Ii is approximated by the constant value Fk(&)Fk(&), where & is the midpoint of the in-
terval Ii. Note that, from the proof of the previous theorem, it follows that the error made in
approximating sup( IFk (<) 1‘lk is bounded by

1/2k
(3.11)

where h is the size of the partition.

We shall use this technique whenever an approximation to the upper bound of a real function is
required. In particular, we shall use it to decide whether or not the trigonometric function m0([)
is a candidate for generating a biorthogonal wavelet basis.

3.2. The Construction of G&(t)

Once m0(<) is constructed, its dual, 60(c), is constructed in a straightforward way from the
identitv (2.13):
(3.12)

Since m0(0) = 1 and m0(~) = 0, then necessarily

iho = 1. (3.13)

If we require some regularity (at least continuity) on the wavelet q(z), then by an argument
given in [13], GO(<) also has to satisfy the condition:

f&)(n) = 0. (3.14)

It is trivial to see that if the two functions

(3.15)

and

No = 120+ (2k + l), (3.16)

n=-No

satisfy (3.12), then the number N0 + 7~0 has to be an odd integer.

If m0([) is of the form (3.15), then by choosing a dual symmetric function GO(<) of the
form (3.16) for some integer Ic, and applying conditions (3.12)-(3.14) one obtains the following
linear system of equations in &:

c
i+j=O
aipj = 1, C
i+j=2n
QiPj =o, l<n<nO+F-l, 40+22(-l)jflj
j=l
=O,

(3.17)
the solution of which are the coefficients of %0(E). The numerical method given in Remark 2
is then used to decide whether or not the dual trigonometric polynomial %0(t) generates a
biorthogonal wavelet basis.
Unlike other known methods for finding %0(E), this one has the interesting feature of providing
an infinite family of dual filters all of the same length. In fact, it is possible to extend the solution
 34 A. KAROUI AND R. VAILLANCOURT

of (3.17) in such a way as to generate an infinite set of dual filters all of the same length. This
is given by the following theorem.

THEOREM 2. Consider a 2n-periodic function,

mTz,(E) = 2 cyj eijc, Ctj EB, Cyj =cY-j, l<_jIns. (3.18)

j=-,o

Assume that, for some NO > no, there exists a real dual trigonometric function

?iiNO(5) = 2 @j eijc, Pj E & Pj = P-j, 1 <j 5 No, (3.19)

j=-No

such that m,,(E) and ?&N,,(t) satisfy condition (3.12) and fijiNo factors in the form

(3.20)

where j(t) is a trigonometric function satisfying

sip If(C) f(25). . . f (2k-1J) (Ilk 5 23’2-E, (3.21)

for some positive integer k 2 1 and c > 0. Then, for all N = NO + 21, 1 a positive integer, there
exists a set S of trigonometric functions of length 2N + 1, dual to mno (E) and having parametric
coefficients.
PROOF. To prove, for a fixed positive intger 1, that there exists an infinite set of dual trigono-
metric functions of length NO + 21, it is enough to prove the result for N = NO + 2. Hence, if we
let
No+2
6&N(<) = C rj eijc,
j=-No-2

where
if -NeIj<Ne,
if No + 1 5 Ijl 5 NO + 2,
then,

where
No+2
fii& (E) = C Sj eijC,
j=-No-2

and Sj = 6-j. Since %N([) has to satisfy the identity

m,o(~)~N(J)$-m,o(Jf~)~N(~+n)=l, (3.22)

then the coefficients of G&(t) have to satisfy the following homogeneous system of (nc+Nc+3)/2
linear equations in Sj:

No+2
60 + 2 C (-1)jSj = 0,
j=l
(3.23)
no + NO + 1
&Sj = 0, O<nl
c 2 ’
i+j=2n
 Familiesof BiorthogonalWavelete 35

intheNs+3unknowns6j,j=O,l,... , NO + 2. Since No > no, this system has a parametric

solution of the form:
sj(bN0+2) = Tj 6N0+2, o_<jLNc+l.
Hence,
No+2
GN(<) = c,(t) + c Tj 6No+2 eije
j=-No-2

satisfies equation (3.12).

To prove that, under some conditions on the rj, we obtain dual filters that lead to the construc-
tion of an infinite family of biorthogonal wavelet bases, we consider the matrix, A, associated
with the linear system (3.17) in the unknowns ,0j. Since the elements of A are bounded, then
for all j, 0 5 j 5 NO + 2, we have (rj] 5 C for some constant C. Moreover, because %ho(<) is
symmetric and fiho(~) = 0, this function can be factored in the form

1 +eiE 2j’(F)
f&ov,(5)
= y---
( 1
This implies that

CN (5) = ~No (5) + f$,‘o (c) =

Since, [rjl 5 C, then there exist two real numbers, 1No+2 < LNo+p, such that for all 6No+2 E
11No+21~5No+2], we have

s’Ip If’(<)
j < 23’2 (2-“/2 - 2-1) .
If we write F(r) := f(c) + f’(s), then it is clear that

s;p jF(#(2<) *’ .2 (2k-15) (1’k < 23’2 x 2-d2.

Consequently, for all 6No+2 E (lNo+2,LNo+2], there exists a dual trigonometric function of length
No +2.
By repeating the above technique as many times ss required, one easily proves that there exist
two real numbers, IN < LN, such that for all 6,~ E [IN, LN], there exists a dual trigonometric
function 66N (5) of length 2N + 1, the coefficients of which depend linearly on the parameter 6~‘.
Furthermore, mno (E) and 6~6~(f) generate a biorthogonal wavelet basis. I

4. NUMERICAL RESULTS
The techniques of the previous section have been used to construct filters of length five, seven
and nine, respectively. The coefficients of the dual of each filter are given in parametric form. Let
mN(t) and GN(<) d enote the dual trigonometric functions that generate a set of biorthogonal
wavelet bases. Here, the integer N stands for the number of vanishing moments [4] of the
corresponding wavelets. Since, generally, the coefficients of ?%N(c) are given in parametric form,
we have used the numerical techniques of the previous section to obtain an approximation to the
range, [ZN,LN], of the parameter ~1,for which condition (2.19) is satisfied by ?%N(<).
To obtain the filters associated with mN([) and %A,(<), it suffices to multiply their coefficients
by u/z.
The decay associated with the Fourier transform of a scaling function c$(z) is defined as the
largest positive real number E such that, for some constant C, the following inequality holds:
00
_-ooI&)1(1 + Irl)‘& < c*
J
36 A. KAROUI AND R. VAILLANCOUFW

In this case the scaling function 4(z) and the corresponding wavelet Q(z) are at least of class C’-‘.
In Table 1, we list the coefficients Q, of mu and @, of iii~(t), for N = 2, 4, 6.
In Table 2, we give the range [1~,L,v] of the parameter 1-1appearing in ;ii~(<) of Table 1,
and list an estimate of the decays EN, TN associated with 4~ and &N, respectively, where the
parameter p is set, respectively, to 1.0, 1.25 and 2.5.

Table1.The coefficients
CY,,
ofmN(<) and &, of&N(<).
-
N 12 Pn
-
2 0 0.550 0.5691056910+ 0.0609756098/~
*1 0.250 0.3656504065- 0.0670731707/~
f2 -0.025 -0.0833333333
*3 0.000 -0.1205284553+0.0701219512/~
f4 0.000 0.0487804877- 0.0304878804/~
f5 0.000 0.0048780487- 0.003048788Ofi
-
4 *0 0.5937500 0.5200237386+0.0113324174/.~
It1 0.3046875 0.2952157909- 0.0090430402fi
f2 -0.0468750 -0.0435768454- 0.0034340659~
13 -0.0546875 -0.0663172888+0.0140796703~
f4 0.0000000 0.0449397607- 0.0052655677~
*5 0.0000000 0.0227013993- 0.0053800366~
zt6 0.0000000 -0.0132413363+ 0.0034340659/~
f7 0.0000000 -0.0015999015+ 0.0034340659~
f8 0.0000000 0.0018665518-0.0004006410~
-
6 +0 0.480468750 0.6525501428+ 0.0020341714/~
fl 0.301562500 0.3116322510- 0.0012189717/.~
f2 0.026562500 -0.1470062921- 0.0008881367~
&3 -0.051562500 -0.0948617842+0.0021280625~
f4 -0.016796875 0.1028781814-0.0007285623~
f5 0.000000000 0.0376065002- 0.0010587927~
*6 0.000000000 -0.0437222610+0.0008615409~
f7 0.000000000 -0.0030134426+ 0.0001013459~
S3 0.000000000 0.0133070652- 0.0002885235j~
&Q 0.000000000 -0.0019276598+0.0000570196~
It10 0.000000000 -0.0017317653+0.0000265957/.~
*11 0.000000000 0.0005641356- 0.0000086637~
-
forN = 2,4,6,oftheparameterp appearing in &N(E)
Table2. The range[ZN,L.PJ],
of Table 1 and theoptimum lowerbounds,EN and FN, associated,
respectively,
with
&N(E) and zW(E).

Eight iterations of the constructive cascade algorithm given in [25, pp. 202-2051, produce a
good approximation to the graphs of the scaling functions and the corresponding wavelets.
In Figures 1, 2 and 3, we present three sets of graphs of $N(z), sN(z), @N(z), and &N(z),
corresponding to N = 2, 4, and 6, respectively. In these figures, the parameter p was set to 1.0,
1.25, 2.5, respectively.
Lastly, in Figure 4, the decays E(p), associated with the Fourier transforms zN(<), of the para-
metric scaling functions, are graphed against the parameter p, for N =2, 4, and 6, respectively.
 Families of Biorthogonal Wavelets 37

0.8

1.0
0.4

05
0.0 L
0.0

-0.4
. i
-6 -4 -2 0 2 4 6

0.8

1
-10 -5 0 5 10 -4 -2 0 2 4

F_igure 1. Dual scaling functions, @N(Z), g~( z ) , and corresponding wavelets, QN (z),
@N(X), for N = 2 and p = 1.0.

1.5
0.6
1.0 0.4

0.2
0.5
0.0
0.0 -0.2

-0.4
-0.5
-6 -4 -2 0 2 4 6 -4 -2 0 2 4 6
1.2
1
0.4 1

0.8

-10 -5 0 5 10 -7.5 -5.0 -2.5 0.0 2.5 5.0 7.5

Figure 2. Dual scaling functions, #N(I), FN (I), and corresponding wavelets, QN(z),
$N(z), for N = 4 and p = 1.25.
38 A. KAROUIANDR. VAILLANCOURT

0.4
1.0

0.2
0.5
0.0

0.0
-0.2

-0.5 i
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6

0.2

0.0

-0.2

-10 -5 0 5 10

Figure 3. Dual scaling functions, C#JN and corresponding wavelets, $N (z),

(z) , $N (L-C),
@N(Z), for N = 6 and p = 2.5.

1.6

1.4

1.2

1.0

0.8

0.6

-75 -5.0 -2.5 0.0 2.5 5.0 7.5 -40 -20 0 20 40

Figure 4. Graphs of decay function, sN(p), of $N(<) for N = 2, 4 and 6, respectively.

 Families of Biorthogonal Wavelets 39

REFERENCES
1. Y. Meyer, Wavelets and operators (Translated from the French by D.H. Salinger), In Cambridge Studies in
Advanced Muthematic, Vol. 37, Cambridge University Press, Cambridge, (1992).
2. G. David and J.L. Journe, A boundedness criterion for generalized Calderon-Zygmund operators, Annals of
Math. 120, 371-397 (1984).
3. Y. Meyer, Ondelettes et Opdrateurs II. OpCmteurs de Calderdn-Zygmund, Hermann, Paris, (1990).
4. G. Beylkin, R. Coifman, and V. Rokhlin, Fast wavelet transforms and numerical algorithms I, Comm. Pure
Appl. Math. 44 (Z), 141-183 (1991).
5. R. Kronland-Martinet, J. Morlet, and A. Grossmann, Analysis of sound patterns through wavelet transforms,
Internat. J. Pattern Recognition and Artificial Intelligence 1 (2), 273-301 (1987).
6. M. Antonini, M. Barlaud, P. Mathieu and I. Daubechies, Image coding using wavelet transform, IEEE
tins. Image Proc. 1 (2), 205-220 (1992).
7. J. Froment and S. Mallat, Second generation compact image coding with wavelets, In Wavelets: A Tutorial
in Theory and Applications, (Edited by C.K. Chui ), pp. 655-678, Academic Press, San Diego, CA, (1992).
8. M. Unser and A. Aldroubi, Polynomial splines and wavelets-A signal processing perspective, In Wavelets:
A Tutorial in Theory and Applications, (Edited by C. K. Chui), pp. 91-122, Academic Press, San Diego,
CA, (1992).
9. A. Haar, Zur Theorie der orthogonalen Funktionensysteme, Math. Ann. 69, 331-371 (1910).
10. J.O. Stromberg, A modified Franklin system and higher-order spline systems on ]w” as unconditional bases
for Hardy spaces, In Conference in Harmonic Analysis in Honor of Antoni Zygmund, Vol. II, (Edited by
W. Beckner et al.), pp. 475-493 Wadsworth Math. Series, Wadsworth, Belmont CA,, (1982).
11. Y. Meyer, Principe d’incertitude, bases hilbertiennes et algebres d’operateurs, In S~minaire Bourbaki
1985/86, Expose 662, Asterisque, pp. 209-223, Sot. Math. de France, Paris, (1987).
12. P.G. Lemarie, Ondelettes & localisation exponentielle, J. Math. Pures et Appl. 67 (3), 227-236 (1988).
13. I. Daubechies, Orthonormal bases of compactly supported wavelets, Comm. Pure Appl. Math. 41 (7),
909-996 (1988).
14. 0. Hermann, On the approximation problem in nonrecursive digital filter design, IEEE tins. Circuit
Theory CT-18 (l), 411-413 (1971).
15. P.P. Vaidyanathan, Theory and design of M-channel maximally decimated quadrature mirror filters with
arbitrary M, having the perfect reconstruction property, IEEE Trans. on Acoustics, Speech and Signal
Processing ASSP-35 (4), 476-492 (1987).
16. M. Vetterli, Filters banks allowing perfect reconstruction, Signal Processing 10 (3), 219244 (1986).
17. D.E. Dudgeon and R. M. Mersereau, Multidimensional Digital Signal Processing, Prentice-Hall, Englewood
Cliffs, NJ, (1984).
18. J.-C. Feauveau, Nonorthogonal multiresolution analysis using wavelets, In Wavelets: A Tutorial in Theory
and Applications, (Edited by C.K. Chui), pp. 153-178, Academic Press, San Diego, CA, (1992).
19. A. Cohen, I. Daubechies and J.-C. Feauveau, Biorthogonal bases of compactly supported wavelets, Comm.
Pure Appl. Math. 45 (5), 485-560 (1992).
20. Ph. Tchamitchian, Biorthogonalite et theorie des operateurs, Rev. Mat. Zberoamer. 3 (2), 163-189 (1987).
21. S. Mallat, Multiresolution approximations and wavelet orthonormal bases of L2(R), Tmns. Amer. Math.
Sot. 315, 69-88 (1989).
22. S. Mallat, A theory for multiresolution signal decomposition: The wavelet representation, IEEE nans. on
Patten Analysis and Machine Intelligence PAMI- (7), 674-693 (1989).
23. RR. Coifman, Y. Meyer, and V. Wickerhauser, Size properties of wavelet packets, In Wuvelets and Their
Applications, (Edited by M. B. Ruskai et al.), pp. 453-470, Jones and Bartlet, Boston, (1992).
24. C.K. Chui, An Introduction to Wavelets, Academic Press, San Diego, CA, (1992).
25. I. Daubechies, Ten Lectures on Wavelets, Society for Industrial and Applied Mathematics, Philadelphia,
PA, (1992).