Wavelets on the Interval and Fast Wavelet Transforms

Albert Cohen, Ingrid Daubechies, Pierre Vial

To cite this version:

Albert Cohen, Ingrid Daubechies, Pierre Vial. Wavelets on the Interval and Fast Wavelet Trans-
forms. Applied and Computational Harmonic Analysis, Elsevier, 1993, �10.1006/acha.1993.1005�.
�hal-01311753�

HAL Id: hal-01311753

https://hal.archives-ouvertes.fr/hal-01311753
Submitted on 4 May 2016

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Wavelets on the Interval and Fast Wavelet Transforms
ALBERT COHEN

CEREMADE, Universite Paris-Dauphine, 75775 Paris Cedex 16, France

INGRID DAUBECHIES

Mathematics Department, Rutgers University, New Brunswick, New Jersey 08903, and AT&T Bell Laboratories, Murray Hill, New Jersey 07974

AND

PIERRE VIAL

Faculte des Sciences, St. Jerome, Universite de Marseille, 13397 Marseille Cedex 13, France

We discuss several constructions of orthonormal wavelet bases 4>(0 = (27r)-1/2 n

j=l
mo(2 -jo (1.4)
on the interval, and we introduce a new construction that avoids
some of the disadvantages of earlier constructions.
( 1.5)

where 1\denotes the Fourier transform, normalized by /(0

l. INTRODUCTION = (27r)- 112 J
dxe-i{xf(x). The functions 1/11,t{x) = 2-112
X 1/1(2 -j x - k), j, k E 7!.., then constitute an orthonormal
The construction of orthonormal wavelet bases or of pairs basis for L 2(1R). For fixedj E 7!.., the l{Jj.J<(x) = Ti124J(2-jx -k)
of dual, biorthogonal wavelet bases for L 2 ( IR) is now well are an orthonormal basis for a subspace ltj C L 2 (1R); the
understood. For the construction of orthonormal bases of spaces ltj constitute a multiresolution analysis, meaning in
compactly supported wavelets for L 2 ( IR), in particular, one particular that
starts with a trigonometric polynomial

( 1.1)
with

satisfying
jElL jElL

mo(O) = 1 ( 1.2)
and

and Projv,_J = Projv,f + 2: <f, 1/Jj.k)I/IJ,k·

kElL

(1.3)
(See Mallat [19], Meyer [20], or Daubechies [9, 10] for
more details.)
as well as some mild technical conditions. (Necessary and
sufficient conditions are given by Cohen [7] and Lawton [17, A consequence of ( 1.4) is that ;f>( ~) = m 0 (~/2);f>(~/2),
18]. A sufficient but not necessary condition, always satisfied or
in practice, is Im0 ( 0 I =1= 0 for all I~ I :s;; 1r I 2; see Mallat
[19]. One then defines the corresponding scaling function l{J l{J(x) = .fi 2: hnl/J( 2x - n),
and wavelet 1/J by

1
where the hn = (4>, 4>- 1 ,~~) are proportional to the Fourier spaces consisting of more regular functions. In particular
coefficients of m0 , hn = .ficn, where the Cn are as in (1.1). (Meyer [20]), if 1/1 E C(IR), then the 4>o.b k E 7L, and 1/1-J,Io
One can prove (see, e.g., Daubechies [10]) that smooth- j E N, k E 7L, provide an unconditional basis for the function
ness for 1/J implies that mo has to have a zero at 1r of suffi- spaces cs(IR), for all s < r. (One needs one "layer" of
ciently high multiplicity. More precisely, scaling functions in this case, because constant functions,
e.g., are in cs(IR) and cannot be written as combinations of

1/1 E Ck(IR) ~ f dx x 1ljl(x) =0 I= 0, ... , k

.p1.Jc. This layer of scaling functions is chosen at the coarsest
resolution level under consideration; it need not be the level
with label 0.) The reason wavelet bases (unlike Fourier se-
~ ries) can provide unconditional bases for C'-spaces is essen-
tially that the wavelets 1/1 have vanishing moments, as guar-
~m 1 0 = 0 I= 0, ... , k. anteed by ( 1.6) or ( 1.7).
d~ ~~rr (1.6) There is another way of interpreting the condition ( 1.6).
Since the functions 4>( · - n), n E 7L, are independent, it is
This in turn implies that m0 has at least 2k non-zero coeffi- equivalent to requiring that any polynomial of degree less
cients. than or equal toN - 1 can be written as a linear combination
By far the oldest example of such an orthonormal basis of of the 4>( x - n) (see Fix and Strang [ 13) and Cavaretta et
compactly supported wavelets is the Haar basis, with a/. [ 6]). Since ljl is orthogonal to all the 4>( · - n), this then
J
ensures that the first N moments of ljl, dx xnljl( x) for n = 0,
... , N- I, all vanish.
Except for the Haar basis, and unlike many examples with
and infinite support, the basic wavelet in an orthonormal basis of
compactly supported wavelets cannot have a symmetry or
1 for 0 ~ x < 1 antisymmetry axis. Symmetry can be recovered, without giv-
4>(x) = { ing up the compact support, if the orthogonality requirement
0 otherwise, is relaxed. In that case one builds two different (but related)
1 for 0 ,:;: x < 1I 2 multiresolution hierarchies of spaces, · · · C V2 C V 1 C V0 C
ljl(x)= - 1 for1/2:s;:x<1
v_l c V_z c 0 0
and
0
c Vz c VI c Vo c
0 0 0
c V_z c v_!
{ · · · , corresponding to two scaling functions 4> and ~ and two
0 otherwise. wavelets 1/1 and ~- They are defined by means of two trigo-
nometric polynomials m0 and m0 , solutions to
In this case 4> and 1/J have support widths 1, but they are not
continuous,andthereforeareunsuitableforthestudyofcon- mo(Omo(O + mo(~ + 7r)riio(~ + 1r) = 1; (1.8)
tinuous function spaces.
Other examples, with more smoothness, were constructed instead of ( 1.4), ( 1.5) we then have
in Daubechies [ 9] . They correspond to m0 of the type
4><0 = (27r)-l/2 TI mo(2-iO,
( 1.7) j~l

where QN( ~)is a polynomial of order N - 1 in e -if., obtained

4><0 = (27r)-l/2 n riio(2-i0,
j~!

by "spectral factorization" from

,fr(O = e-;u 2mo(f.l2 + 7r)c/>(f,/2),

~(0 = e-iU 2m0 (f,/2 + 7r)J>(02).

(Note that m0 is used in the definition of 1/J, and m 0 for ~-)

The resulting 4> and 1/J have support width 2N - 1, which is One has again that if some extra technical conditions are
the shortest possible under the constraint that m 0 has a zero imposed on mo, m0 , then the .P1 .~c and the ~i.Jc constitute Riesz
of order N at ~ = 7r. The smoothness of 4> and 1/1 in this family bases for L 2 ( IR). They are dual bases; i.e.,
of examples increases linearly with N; in particular, there
exists 11- = 0.2 such that </>, 1/J E C IJN.
These smoother wavelets provide not only orthonormal
bases for L 2 (1R), but also unconditional bases for function This duality is also reflected by the fact that

2
 !/lj,k 1.. ~' lj!j,k 1.. \1}, priori clear how to adapt them in such a way that the result
is an orthonormal basis of L 2 ( [ 0, 1) ) ,
A more detailed exposition, with proofs and examples, can Several solutions have been proposed for this problem.
be found in Cohen et at. [8 J; see also Cohen and Daubechies They all correspond to different choices of how to adapt the
[ 4 J. Symmetry for cf>, !/! and ~. ljJ is now possible because multiresolution hierarchy to the interval [ 0, 1J.
there exist symmetric solutions m 0 , m0 to ( 1.8). Two possi- A first solution consists in not doing anything at all. A
bilities exist: if m 0 , m0 have an even number of coefficients, function f supported on [ 0, 1J can always be extended to the
then cf>(x) is symmetric around x = 1/2, and !/J is antisym- whole line by puttingf(x) = 0 for x rt_ [0, 1). This function
metric around the same point; if m 0 and m0 have an odd can then be analyzed by means of the wavelets on the whole
number of coefficients, then cf> and !/! are both symmetric, real line. There are two things wrong with this naive ap-
cf>(x) around x = 0, !/f(x) around x = 1/2. proach. First of all, this kind of extension typically introduces
Smoothness for these ''biorthogonal'' wavelet bases again a discontinuity in fat x = 0 or 1, which is reflected by
requires a factorization similar to ( 1. 7) for m 0 or m0 • More "large" wavelet coefficients for fine scales (i.e., wavelet
precisely, we now have coefficients which do not decay very fast) near the two edges,
even if f itself is very smooth on [ 0, 1J. The (one-sided)
regularity of fat 0 or 1 is therefore not characterized by the
!/! E Ck(IR) => f dx x'ljJ(x) = 0 l = 0, ... , k decay of the (f, !/f1k) for j-+ -oo. The second "bad" aspect
is that this approach uses "too many" wavelets. At scale - j,
~ one finds (/, !/! -J,k) =1= 0 for the typically 2 i + 2N - 1 wavelets
corresponding to - ( 2N - 2) ~ k ~ 2 1 - 1; intuitively one
d' 1 mo
dC I
l;~rr
= 0 l = 0, ... 'k, should have to use only 2 1 wavelets, at scale - j, when look-
<,
ing at problems on [0, 1). We shall come back to this desir-
ability of having exactly 2 1 wavelets.
which is again equivalent to a factorization of type ( 1. 7) for A second solution consists in periodizing. In this case, one
m 0 • Similarly, smoothness for ljJ requires zero moments for expands a function f on [0, 1] into "periodized" wavelets
!/J, or a factorization similar to ( 1. 7) for m0. defined by
All the above concerns bases for [2(!R1). (These one-di-
mensional constructions can easily be extended to higher di- !/JI'_e~.k(x) = 2 112 }: !/f(2 1x + 2 11- k),
IEZ
mensions, but we stick to one dimension here.) In many ap-
plications, however, one is interested in problems confined with j ~ j 0 ~ 0 (for j < 0, the !/Jf!!'}.k vanish identically),
to an interval. Examples are numerical analysis (with bound- 0 ~ k ~ 2 1 - 1. These wavelets have to be supplemented
ary conditions at the edges of the interval) or image analysis by lowest resolution scaling functions c/>f!!'fo,b defined analo-
(where the domain of interest is the Cartesian product of two gously; the result is an orthonormal basis of [2([0, 1]), as-
intervals). To fix notations, let us assume that the interval is sociated with a multiresolution analysis in which V ~' is
[ 0, 1J. It is very easy to restrict the Haar basis for L 2 ( IR1) to spanned by the c/>f!!'},k· In this case, one has indeed exactly 2 1
a basis for L 2 ([0, 1]); starting from the collection {c/>o,k; k wavelets at scale - j; the number of scaling functions c/>f!!'j,k
E 7l} U { !/11.1<; j ~ 0, k E 7l}, which is an orthonormal basis in every V f!:J is also 2 1. Since obviously
for L 2 ( IR1), it suffices to take the restrictions of these func-
tions to [ 0, 1] . Since every one of them is supported either
in [ 0, 1J or in IR1 \] 0, 1 [, the collection that remains after all
the functions with restriction 0 have been weeded out, i.e.,
{ cf>o.o} U { !/f1,k; j ~ 0, 0 ~ k ~ 2 111 - 1}, is an orthonormal expanding a function on [ 0, 1J into periodized wavelets is
Haar basis for [2([0, 1]). Things are not so trivial when one equivalent to extending the original function into a periodic
starts from smoother wavelet bases on the line. In the ex- function with period 1 and analyzing this extension with the
amples ( 1. 7), both cf> and !/J have support width 2N - 1. In standard whole-line wavelets. It follows that, unless f itself
order to avoid having to deal with the two edges of [ 0, 1J at was already periodic, this construction introduces again a
the same time, we can choose to start from the basis discontinuity at x = 0, x = 1, which shows up as slow decay
{cf>-Jo,k; k E 7l} U {!/11.k;j ~- j 0 , k E 7l} for L 2 (1R1), where in the fine scale wavelet coefficients pertaining to the edges.
j 0 is chosen large enough so that none of the functions has Again, it is impossible to characterize the one-sided regular-
support straddling both 0 and 1 (i.e., 2io--l ~ N). Even so ity of f at 0 or 1 by simply looking at the decay of the
there will be 2N - 2 functions, at every resolution level and I(f, !/JJ.':> I for j-+ -oo, unless f is periodic.
at every end of [ 0, 1], among these orthonormal basis func- A third solution, often adopted in image analysis, is to
tions, that straddle an endpoint, so that their support is neither reflect at the edges. In this case, one extends the function f
completely in [ 0, 1J nor completely in IR1 \ ] 0, 1 [. It is not a on [ 0, 1] by mirroring it at 0 and 1:

3
 f(x) = /(2- x) for 1 ~ x ~ 2 "edge" functions have to be constructed explicitly. The re-
sult of Meyer's construction is an orthonormal family of
f(x)=f(-x) for -1~x~O.
wavelets in [ 0, 1], with N vanishing moments, and the same
regularity as the original 1/1; together with an orthonormal
Beyond -1 and 2 we mirror once more, and so on. The full
family of scaling functions on [0, 1) at the coarsest scale
extension is then defined by
under consideration, these adapted wavelets constitute an or-
thonormal basis for L 2 ([0, 1]). In addition, their regularity
f(x) = f(2n- x) 2n- 1 ~ x ~ 2n
and vanishing moment properties ensure that they are un-
f(x-2n) 2n~x~2n+l. (1.9) conditional wavelet bases for the Holder spaces C' ( [ 0, 1])
for all s < r, where r is the regularity of the original wavelet
If the original function on [ 0, 1] is continuous at 0 and 1, basis, 1/1 E C'. We recapitulate in Section 3 the main steps
then this extension is continuous at all the integers. Typi- of Meyer's construction (very briefly, without proofs), and
cally, however, the derivative of the extension has discon- give tables for the corresponding adapted "filter" coeffi-
tinuities at the integers. Expanding the extension ( 1. 9) of cients near the edges. Meyer's construction has two weak-
a function f on [ 0, 1] in a whole-line basis of wavelets nesses. Because the number of scaling functions at resolution
is equivalent to expanding the original function on j is larger than the number of wavelets, the construction can-
[ 0, 1] with respect to the "folded" wavelets t/J ~~d defined not be generalized to wavelet packets on the interval: in a
on [0, 1] by wavelet packet construction, wavelet coefficients get split as
well as scaling coefficients, using the same filters, and for
this it is essential that the two families have the same number
IEZ IEZ of coefficients at every scale. The fact that the number of
scaling functions is not a power of 2 is also a nuisance for
Starting from an orthonormal wavelet basis, this folding practical applications such as image analysis, where arrays
typically does not lead to an orthonormal wavelet basis on are typically squares with 256 X 256 or 512 x 512 pixels.
[ 0, 1]. If 1/11*' ;j,1* are two biorthogonal wavelet bases, with The other objection to Meyer's construction is that the ex-
t/J and ;j, both symmetric or antisymmetric around 1/2, then plicit construction of the edge functions involves the diago-
their folded versions are still biorthogonal on [ 0, 1], how- nalization of a matrix that becomes ill conditioned for rea-
ever. We discuss this in some more detail in Section 2, with sonably large N. We shall discuss this in more detail in Sec-
examples. The resulting biorthogonal multiresolution anal- tion 3.
ysis hierarchies on [ 0, 1] have 2 1 + 1 (symmetric case) or This paper presents a fifth solution, also derived from com-
2 1 ( antisymmetric case) scaling functions and 2 1 wavelets pactly supported wavelet bases for IR. Like Meyer's solution,
at resolution level j. Because the extension ( 1.9) typically it uses "interior" and "edge" scaling functions at every res-
has a discontinuous derivative, we again cannot expect to olution. We introduce fewer edge functions, however, tailor-
characterize arbitrary regularity of f by means of the wav- ing them so that the total number is exactly 2 1 at resolution
elet coefficients; decay of the (f, t/J JJ:1d) can characterize up j; moreover, as in Meyer's case, all the polynomials on
to Lipschitz regularity (a gain over the two previous "so- [ 0, 1] of degree ~ N - 1 can be written as linear combina-
lutions"), but not more. (One can do a little better by using tions of the scaling functions at any fixed scale. It then fol-
two different pairs of biorthogonal bases; see Section 2.) lows that all the corresponding wavelets, at the edge as well
A fourth solution was proposed in Meyer [ 21] . The start- as in the interior, have N vanishing moments, and this is
ing point of this construction is any one of the compactly sufficient to ensure that we again have unconditional bases
supported bases in Daubechies [ 9] , with N vanishing mo- for the C' ( [ 0, 1]) -spaces, with s < r if 1/J E C. This new
ments, and support 1/1 = support cf> = [ 0, 2N - 1] . The basis construction is explained in detail in Section 4, with many
on [ 0, 1] constructed by Meyer is derived from a multire- examples and consequences. After completing this work, we
solution analysis that "lives" on [0, 1]. At sufficiently fine learned that a similar construction was made independently
scales, the approximation spaces V ~~· 1 1 consist of 2 1 - 2N - by Jouini and Lemarie-Rieusset [16], and by B. Jawerth. A
2 "interior" functions, 2N - 2 "left edge" functions, and first announcement of the results was made jointly by B.
2N - 2 "right edge" functions. The complement spaces Jawerth and the present authors in [ 5]; extensions and ap-
W ~~· 1 1 are generated by 21 - 2N- 2 "interior" wavelets, N plications (developed independently of this paper) can be
- 1 "left edge" wavelets, andN- 1 "right edge" wavelets. found in L. Andersson et al. [1].
The total number of wavelets at scale j is thus 2 1 , but the
total number of scaling functions is larger, 2 1 + 2N- 2. The 2. FOLDING BIORTHOGONAL WAVELETS
"interior" functions are simply .p1k or cf>1t. as they were de-
fined on the whole line, for which the indices j, k happen to Given any reasonably decaying function f on IR, we define
correspond to support 1/11* or support cf>1* C [ 0, 1]. The its "folded" version by

4
 ftold(x) = L [f(x- 2n) + f(2n- x)]. (2.1) fined analogously. As in the periodized case (see Meyer
nEZ [ 20]), these spaces are non-trivial only for sufficiently fine
resolution. For j ;;;:,: 2 we have that
This function has the property that, for all x E IR, k E 7L,

rold( -x) = ffold(x) n

ffold(x + 2k) = rold(x). (2.2) + 1/J( -2-J X+ 2-J+ln- k)J = 0

For later convenience, note that

because L 1 1/J(x - //2) = 0 (a consequence of ifl(x) = fi
X Ln 8n<f>(2x- n), with Ln 8n = 0, Lk <f>(x- k) = 1). For
(2.3)
j = 1, we use ifl(x) + i/1(1 - x) = 0 to conclude that

an easy consequence of ( 2.2).

Suppose now that r/J, (/1 are two compactly supported wave-
lets giving rise to biorthogonal wavelet bases, with associated
scaling functions </>, 4> and multiresolution analyses · · · c V1
CV0 CV_ 1 C ···, ···CV1 CV0 CV_ 1 C ···,asinCohen
et a/. [ 8] . We "fold" these wavelets and scaling functions
and study the result. This is interesting only if we can exploit It follows that w~old = {0} for j ;;;:,: 1; similarly w~old = {0}
symmetries of</>, r/J, 4>, and (/1, and we therefore assume that for j ;;;:,: 1. On the other hand L* <f>(x - k) = 1 implies, for j
these functions have symmetry axes. There are two cases to ;;;:,: 1, that
distinguish: either </>( x) = </>( 1 - x), 1/J( x) = - i/1( 1 - x)
(same for 4>, (/1), corresponding to an even number of non-
zero taps in the filters m0 and m0 , or </>( x) = </>( - x),
ifl( 1 - x) = ifl( x) (same for 4>, (/1), with an odd number of
non-zero taps in the filters m0 and m0 • We start by considering
the first case, where </>, 4> are symmetric and 1/J, (/1 are anti- whereas, for j = 0, <f>(x) = </>(1 - x) leads to
symmetric around 1/2. The corresponding m0 , m0 are then
necessarily of the type <P~':lcd(x) = ,L [<f>(x- 2n- k) + </>( -x + 2n- k)]
n

= ,L [ </>( x - 2n - k) + </>( x - 2n + k + 1)]

= ,L <f>(x + m) = 1.

with p, p two polynomials such that It follows that the V~old (and V~oid) spaces collapse to only
the constant functions if j ;;;:,: 0. We therefore consider the
x 1+1+1 p(1- x)p(1- x) + (1- x)'+ 1+1 p(x)p(x) = 1. folded spaces only for j ~ 0. For j = -1 ~ 0, the equations
( 2.2) imply that
Several examples are given in Cohen et a/. [ 8]; in addition
C. Brislawn has drawn our attention to examples in which (2.4)
m0 and m0 have the same number of "taps" (i.e., the same
number of non-zero coefficients when written out as poly- and
nomials in e-i{). The wavelets 1/J, (/1 can be made arbitrarily
regular by choosing l, [large enough, and picking appropri- (2.5)
atep,p.
Define the space Vjold to be the closed linear span of the so that we can restrict ourselves to k = 0, ... , 21 - 1. The
<P~1d, k E l., where the closure is taken with respect to same happens with ~. 1/J, /jJ (with a change of sign in ( 2.4)
L 2 ([0, 1]). (Because of (2.2), any linear combination of for 1/J, If,).
<P ~~ld is completely determined by its restriction to [ 0, 1], so It turns out that the <f>t:!}~, 4>t:!)~·, 0 ~ k, k' :5; 21 - 1,
that we can identify the linear span of the <P~~kd with a sub- inherit the biorthogonality of the unfolded versions, as can
space of£ 2((0, 1]).) Spaces v~oid, Wjo1d, and w~old are de- be seen by using ( 2.3):

5
 J: dx~~}~(x)<i>~}~·(x) (An only slightly less trivial argument can handle the case
where 1/J, ~are not compactly supported but have good decay.)
If support 1/J = [ -K + 1, K] (remember that 1/J has a sym-
= L ((~-1,1" <1>-1,k'+2 1 + 1n} + <~-1,k• <1>-1,2 1 + 1n-k'-l}] metry axis at x = 1/2), then support 1/J-J.k = [(k - K +
nEZ 1)2-i, (K + k)2-i], so that 1/J~~~*(x) = I/J-1.k(x) for 0..,;; x
= L [bk,k'+2 1 • 1n + bk,2 1 + 1n-k'-l] = bk,k'· ..,;; 1 if 2K - 1 ..,;; 2 1 and K - 1 ..,;; k ..,;; 2 1 - K. If 2 1 ~ 2K
nEZ - 1 but 0 ..,;; k < K - 1, then 1/J~~~*(x) = 1/1-J,k(x)
- 1/J-J,-k-l(x) for 0..,;; x..,;; 1; similarly 1/J~~~k(x) = 1/J-J.k(x)
This biorthogonality implies, among other things, that the - I/J-J, 21•'+l-k(x) for 0..,;; x..,;; 1 and 2 1 - K < k..,;; 2i- 1.
Now take f E L 2 ([0, 1]) and define its extension to all of 1R
~~}~. k = 0, ... , 21 - 1 are all independent, so that V~}d is
by J<•'(x) = f(x) for X E (0, 1),/ex'(x) = 0 for X$ [0, 1].
a 2 1 -dimensional space. Completely analogously, we can de-
Define j 0 to be the smallest j such that 2 1 ~ 2K - 1. Then
rive

K-2
L [ L l<r",
j~Jo k~O
1/1-j.k> - <r", 1/1-j.-k-1>1
2

21-K
+ L I<text, 1/1-j.kW
k~K-l

proving that W~}d .l V~}d, W~}d .l V~}d, and that the 1/1~}~. 21-1

k = 0, ... , 21 - 1 constitute a basis for the 21 -dimensional

space W~)d.
In summary, we have two ladders
._;;; 2 }: }:
J~Jo k~-K+l

(same for the v~old) of nested 2 1 dimensional spaces v~r'

= 2 }: L I<text, 1/1-j,k> I 2, (2.8)
j=Jo k=-oc
and 21-dimensional complement spaces W~Jd, j ~ 0, with
where we have used in the final step that support 1/J-J.k n [0,
1] = 0 for k ..,;; - K or k ~ 2 1 + K + 1. The frame bounds
for the ~-*on the whole line then imply that ( 2.8) is bounded
1
by 2B J:oo dx IJ<•t(x)/ 2 = 2B fo dx /f(x)/ 2 , for some B <
(where the direct sum is not orthogonal). In addition, we oo. Since there are only finitely many 1/f~)~ to be considered
have the biorthogonality relations proved above, linking the for j < j 0 , this proves that, for all f E L 2 ( [ 0, 1]),
two hierarchies. Moreover V~old = v~old consists of only the
COnstant functiOnS, and one easily derives uj:O v~r =
U7~o V~)d = U([O, 1]) from the analogous relation for the
unfolded spaces. Any f E L 2 (/R) can therefore be written as for some B' < oo. The same is true for the <J> fold, ~told, and
the two upper bounds together with the duality ( 2.6) lead to
1 21-l
lower frame bounds, by Cauchy-Schwarz.
f = (/' .,l.fold)A.fold
'PO,O 'PO,O
+ lim
1 ~""
~
LJ
~
LJ
(f,' ,i,~ld }·'·~ld
'I' J,k 'I' J,k For numerical implementation in the L 2 ( 1R) case, one uses
J~o k~o
the filter coefficients c", or rather hn = fie", from ( 1.1), and
2L-1 1 21-1
the hn corresponding to m0 , rather than~. 1/J, <J>, ~themselves.
= L <t. <i>~l~*>~~l~*
k~O
+ lim1~ 00
L L <t ~~J?*>"'~J?k·
j~L k ~o The fi1ter coefficients to be used for the folded functions
~fold, .. · in the L 2 ( [ 0, 1]) case are simply correspondingly
folded versions of the h" and g" = ( -1 )"h-n+l· From ~1 .k(x)
This is not sufficient to ensure that the { ~~l~. k = 0, ... ,
= }:m hm~-J-1,2k+m(x) we immediately have
2L- 1} U {1/l~~d.k; j ~ L, k = 0, ... , 2 1 - 1} constitute a
Riesz basis, however. (See the similar discussion in Cohen ~~~~k(x) = L hm~~J~l.2k+m(x)
m
eta/. [ 8] .) We still need to establish frame bounds. To show
how the folded wavelets inherit frame bounds from their un- L h,_2k~~J~u(x). (2.9)
folded parents we exploit that 1/J, ~are compactly supported.

6
 + 1, L], then hm = 0 for m ,;;; - L or m
If support 4> = [ - L
> L, so that formula ( 2. 9) can be adopted as is for ( L - 1 ) I
2 :,;;; k ,;;; 21 - (L - 1 )/2. Fork outside this range, the index
2k + m may fall outside { 0, 1, ... , 2i+L - 1}, and we need
to apply (2.4), (2.5). For 0:,;;; k < (L- 1)12, for instance, FIG. 1. Low pass filtering scheme for the "folded" case, going from
V!o\d to V'ol!', in the case where support cf> = [ -1, 2], i.e., only h_,, ho, h,,
L
h 2 are *0; in fact h0 = h 1 = a, h_, = h 2 = f3 because of the symmetry
</>~~~k = 2: hm</>~}~l.2k+m conditions imposed (see text). The straight lines represent unaffected filter
m=-2k coefficients. In this case only two filter coefficients are affected by the fold-
-2k-l ing (wavy lines-~me at each end); the folded coefficient here is h0 + h_,
+ 2: hm</>~}~l.-21<-m-l = hu + h2.
m=-L+l
2: (hi-2k + h-2k-l-l)4>~J~l,/· (2.10)
For j = -J,;;; 0, one has
/~0

Something similar happens for 2 1 - ( L - 1) I 2 < k ,;;; 2 1 -

1; if 2 J < (L - 1), then even more folding terms step in, n
because the two edges come into play simultaneously. Sim-
ilar formulas can be derived for the high-pass filters linking n
the 1/J ~~d.k with 4> ~)~ 1,1 • This finishes our discussion of the case
where ¢, ~ and 1/J, ;jJ both have their symmetry axes at x = Similar equations hold for 4>, ;jJ; all the derivations are anal-
1/2. We end up with a multiresolution analysis with exactly ogous to the previous case. It follows that we now restrict
21 functions in the jth approximation space as well as in the ourselves to the V ~Jd, W~Jd spaces (and their - equivalents)
jth wavelet space, which means that these can be used for for j ~ -1 (as opposed to j ~ 0 in the previous case). For
wavelet packet constructions as well. Moreover, it is easy to j ~ 0,
find the "folded" filter coefficients from the original ones;
Fig. 1 illustrates how the coefficients near the edge are af-
fected in a simple case.
Let us now consider the case where </>, ~ have a symmetry
axis at x = 0, <f>(x) = </>( -x), (/J(x) = (/J( -x), and 1/J, ;jJ are
symmetric around x = 1/2, 1/1(1 - x) = 1/J(x), ;jJ(l - x) = (same for 4>, ;jJ), implying that
;jJ(x). This corresponds to an odd number of non-zero coef-
ficients in mo, mo; more precisely, mo, mo are of the type dim V~)d =dim v~r = 21 + 1,
dim W~Jd = dim W~}d = 21, (2.11)

at least for j ~ 0. For j = -1, we have dim Viold = 1 = dim

v~oid, dim w~old = 1 = dim W~otd. The independent basis
.
f unctwns m _1 can be t ak en as '1'-J,t.
. vfold A,fold k-- 0 , 1, .•• , 21·, t'n
W fold t k ,/,fold
_1 we a e 'I'_ J,b k -
- 0 1
, , · · ·, 2 i - 1 ·
with As in the previous case, one easily checks that

x 1+1p(l - x)p(1 - x) + (1 - x) 1+ 1p(x)p(x)

There are again several examples in Cohen et a/. [ 8], in-

= 1.
f dxi/J~~~k(x);jJ~~~k'(x) = 6k,k'
cluding some examples where 4> and 4> are very ''close.''
There exist no examples with the same number of non-van-
ishing coefficients for both m 0 and m0 in this case.
Ld.xi/J~~~k(x)4>~~~k'(x) = 0 = Ldx;jJ~~~k(x)</>~}~dx).
1

Define the folded scaling functions, the wavelets, and the The biorthogonality of the ¢, 4> is slightly different:
spaces they generate as before. One now checks that

1/J~~kd= 0 if j~2

1/J ~~~d = 1/J ~~~d for all kElL

<P J~fd = 2 j/2 if j ~ 1.
= 2: [Ok,k'+2i+ln + Ok,2i+ n-k'] •
1

7
If 0 < k < 2 1 , then this still equals 6*.*', but fork = 0 or k
= 2', we find 26k,k'. In order to preserve biorthogonality, we
are therefore forced to readjust the normalizations of <P~~~o
and <P ~~~ 2 1 by defining

A. fold,n _ _!__ A. fold A, fold,n _ _!__ A, fold

'I'- j,O - ..fi 'I'- j,O, 'l'-j,21- ..fi'l'-}.21· FIG. 2. Low pass filtering scheme for the "folded" case, going from
V!o\d to V'~1d, in the case where support cf> = [ -2, 2], i.e., only h_ 2 , h_" h 0 ,
h~o h 2 are *
0; we have h_ 2 = h 2 , h_, = h 1 because of symmetry. This
(The extra "n" stands for "normalized.") The same has to diagram corresponds to the unnormalized formula ( 2.12); some of the af-
fected filter coefficients (wavy lines at edges) are equal to 12 times their
be done for ;;,~~~0 , ;;,~~~ 2 1. For j = -1 we also redefine
normal value in this case.

ffold,2(x) = ~ [f(x + 2n) + f(2n- x + 1)],

nEZ
All the other scaling functions remain untouched; i.e.,

and one finds dim VJold,z = 1 for j ;;:.: 0, so that only spaces
withj ~ 0 need be considered. One still has dim V~Jd· 2 = 2'
We can then write, for any /E L 2 ((0, 1]), + 1 for j ;;:.: 1, however.
To end this section, let us see for which values of s the
coefficients <J, 1/J~Jt> can be used to characterize f E C 5 ([0,
00 21-1 1]) in either of the two constructions. For 0 < s < 1, we
+ ~ ~ (f, ;p~~~k)l/l~~~k define CS ( [ 0, 1]) by the straightforward definition
1=0 k=O

2L :lO 21-t /E Cs([O, 1]) ~ 11/IJc. = sup{Jx- yJ-s lf(x)- f(y)J;

= ~ (f, ;;,~i~f)<f>~i\n + ~ ~ (/, ;jJ~~~k)I/J~~~*'
k=O j=L k=O 0 ~ X, y ~ 1} < 00. ( 2.13)

where we do indeed have dual frames. (The existence of Suppose that 1/1, <P E C, with r > 1. Then
frame bounds follows exactly as in the previous case.) For
numerical implementation, one needs the corresponding low- PROPOSITION 2.1. For 0 < S < 1, a function f on ( 0, 1)
and high-pass filters, which correspond again to folded ver- is in C([O, 1]) if and only if
sions of the original h", g", h", g". For instance, if support <P
= [ -L, L], then (2.9) can be used as is for L/2 ~ k ~ 2 1 sup 2i(s+l/2lJ<J, ;jJ~J.1)J < oo. (2.14)
j~O
- L/2. For 0 ~ k < L/2, we have O.s.::k~2'-1

A,fold _
'1'-j,k-
~
L,.
(h l-2k + h -2k-/ )A,fold ,
'1'-j-L,I• (2.12) Proof.
/;;.0

1. Assume first that f E CS ( [ 0, 1]). Because

the recurrence for the <P~~~f then follows by an easy adjust-

f )dx 1/J."J.k(x) = ~ fl dx[!/1-J,k(x + 2n) + I/J-J.k(2n- x)]

ment. Note that this adjustment is not even necessary: one
•tw - -
can keep working with the <P ~~~* as they are, provided that
0 n 0
at the reconstruction stage, the two extremal scaling coeffi-
cients get halved systematically at every step as one climbs
back up on the resolution ladder. Figure 2 i11ustrates how
( 2.12) affects the coefficients near the edge in a simple case.
= L: dx ;jJ_1,k(x) = 0,

This second case corresponds to what is sometimes done

we have
in image analysis when filters with an odd number of taps
are used; the trick of halving the extremal scaling coefficients
at the reconstruction stage is weJl known.
Remark. One can avoid some of the peculiarities of the
l(f, ;p~~~k)J =If dx[f(x)- /(2-'k)];jJ~~~k(x) I (2.15)
low resolution spaces in this second construction by folding
as -1/2, 1/2 instead of at 0, 1. The folding map is then
~C f dxJx- 2-ikJsJ;p~~~k(x)J.

8
We need to be concerned only about the behavior for large together with t/1( 1 - x) = ± t/1( x) this implies that
j. In that case, for most k, i.e., fork- 1 ~ k ~ 2 1 - k
(where support~ = [ -k + 1, k]), support ~-J,k C [0, 1], 21-[
~fold _ ~
so that 1/J-J,k(x)- t/1-J,k(x), and }: 11/J~~~k(x)- 1/J~~~k(y)l
k~O

(2.15) = C f_~, llilx- 2-1klsi~-J,k(x)l ~2 112 }:

1=-X)
lt/1(2 1x-l)-t/1(2 1 y-l)l. (2.16)

= C f_~, llilx- 2-1kls2 112 1~(2ix- k)l Because support 1/J C [- K + 1, K], at most 4K terms in this
sum can contribute, so that

~ c2-,(s+l/ 2 ) f_"', dylylsl~(y)l. (2.16) ~ 4K2 112 sup lt/1(2 1x- /)- t/J(2 1y- /)I
I

~ C 2 }12 min ( 1, 2 1 Ix - y I),

For 0 ~ k ~ k - 2 we have
where we have used that 1/J is bounded and Lipschitz. Con-
~ fold ~ ~
t/1 -J,k(x) = t/1-J,k(x) ± t/1-J,-k--l(x), sequently

so that lf(x)- f(y)l

Jo
~ C}: z-is2 1lx- Yl + C }: (2.17)
(2.15) ~ C L"' llilx- 2-ikl'2l 12 1~(2ix- k)l ,~o }~}o+l

+C f' llilx- 2-1kls2 112 1~(2 1 x + k + 1)1

with jo determined by 2 -Jo-t <
immediately implies that
Ix - yI ~ 2 -Jo. This now

~ 2C2-j(s+l/ 2 ) f_"'x dylylsl~(y)l Jf(x)- f(y)J ~ C'lx- yl2io(l-s)

+ C"2-}0S ~ C'"l X - yls,

+ cz-jsl2k + llszj!Z f_"', dyl~(2 1y)l and /E C 5 ([0, 1]). II

The arguments in this proof would not work for s ~ 1.
~ C'2-j(s+l!2)_ For s = 1, the only problem is the very last step: the naive
bound ( 2.17) would lead to an extra logarithmic factor in the
Something similar happens for 2 1 - k + 1 ~ k ~ 2 1 - 1. bound on 1/(x)- /(y)J, i.e., 1/(x)- f(y)l ~ Clx- yl
This proves that (2.14) holds. Ilog I x - y I 1. This can be avoided by replacing the Lip-
2. We now prove the converse. We have schitz space (obtained by takings= 1 in (2.13)) with the
larger space A* of functions of Zygmund class. On the whole
00 21-1 line A* ( IR) is defined by
1/( x)- f(y) I = I }: }: (/, ~~~~k) [ 1/J~~~k(x)
A.(~)= { f: ~-+ C; 11/IIA.
00 21-1
-1/J~~~k(y)]J ~C}:}: 2-J(s+I/ 2 )11/J~~~k(x)-t/J~~~k(y)J. f(x +h)+ f(x- h)- 2f(x)l }
=sup <co ·
}~o k~o
X~ I h '
(2.18)
But
as shown in Meyer [ 20],

f E A*(~) ¢* sup 2 3112 1(f, 1/J-J.k) I < co.

j,k

If we restrict to [ 0, 1], then the symmetric differences in

(2.18) make no sense at x = 0 or 1; to define A*([O, 1]) we

9
replace these symmetric differences by one-sided Lipschitz typically is only Lipschitz, even if f is very smooth on
bound estimates at the edges 0 and 1 (for x E]O, 1 [we keep [ 0, 1], we do not expect to be able to characterize CS ( [ 0, 1])
the symmetric difference): spaces, with s > 1, by means of folded wavelet bases. (For
1 < s < 2 we say that f E Cs([O, 1]) if and only if f is
A* ( [ 0, 1]) = { f : [ 0, 1] -+ C; differentiable on ] 0, 1 [, left differentiable at 1, right differ-
entiable at 0, and sup{ I x- y 1-s If( x) - f(y) - f'(x)(x-
f(x) + f(y)- 2f((x + y)/2)1 y) I; 0 :;;;; x, y ,;; 1} < oc.) Let us have a closer look at how
sup < oo,
I
Q,.;;x<:y"'l X - Y the proof of Proposition 2.1 fails when s > 1. First of all, in
the argument in point 1, we replacef(x) by f(x)- f(2- 1k)
sup If(x)- /(0)1 < oo, sup I /(1)- f(y)l < oo}. - (x - 2 -Jk )f'( 2 -Jk) in the evaluation of(/, ~~~~*); to make
J;;,x>O X O"'y<l 1- Y the rest of the argument work, we require that n
dx x ~~~~*(x) = 0 for 0,;; k,;; 2 1 - 1 and sufficiently large
It is then still true that j. This is easily seen to be equivalent to J; dx x~(x - m)
= 0, all m E 7L (look at sufficiently fine scales, and use
PROPOSITION 2.2. Let f be a bounded function on [ 0, 1].
Then J~, dx x ~(x) = 0), which in turn is equivalent to X f
dx H(x)~(x - m) = 0, all m E 7L, where H(x) is the hat
function,H(x) = 1- lxl if lxl,;; 1,H(x) = Ootherwise.
fE A*([O, 1]) ~ sup 2 3112 1(!, ~~~~*)I < oo. (2.19)
)0<0 One obvious construction of biorthogonal wavelet bases that
O~k:!!i21-l
satisfies this requirement is the case m 0 ( O = cos 2 (~I 2); we
have then H(x) = <P(x), which is necessarily orthogonal to
Proof all the ~(x - m). By playing around with the general recur-
1. We start by extending fto all of~ by defining sion relations for ~. 4> as well as for H, one can also show
that this is the only solution (see Appendix A). For these
f( x - 2n) if 2n ,;; x ,;; 2n + 1 special biorthogonal wavelet bases where <P = H
f""(x) ={ (2.20) (and only for those!), one does therefore get decay ~ 2 -J(s+ltZ)
f(2n- x) if 2n- 1 :;;;; x,;; 2n. for the(/,~~~~*) if /E CS([O, 1]) with 1 < s < 2. (Higher
values of s cannot be achieved.)
This extension is an "unfolded" version off, symmetric So much for the necessity part of condition ( 2.14). What
around every integer. It follows that f E A* ([ 0, 1]) ~ /"xt about the sufficiency? To prove that lf(x + t) - f(x) -
E A*(~). f'(x)tl ,;; Cltl', with 1 < s < 2, we would follow essen-
2. On the other hand, for j ;;, 0, 0 ,;; k :;;;; 2 1 - 1, tially the same procedure as in step 2 of the proof of Prop-
osition 2.1, using the bound

I-"_ _
dxf•''(x)•i•_1·,k(x) =
'I'
L 1

dxf(x)•;,c_ol1d.k(x),
'I'
(2.21)
11/J~~~*(x + t)- 1/J~~~*(x)- (1/J~~~*)'(x)tl

.;; C2 112 min(1, (2 1 ltl) 1+')

and, for all n E N,
if 1/1 E C'+ 1 • If l(f, ~~~~*)I .;; C2-J<s+ltzl, and s < r + 1,
(/ext, ~-j,k+21• 1 n) = (j•••, ~-j,k) then the conclusion follows. Unfortunately, in the special
biorthogonal bases where the necessary part of the theorem
(J•", ~-j,21• 1 n-k-l) = ±(j•", ~-j,k)·
works, the wavelet 1/1 is a finite linear combination of hat
functions, so that"' $. c+l for r > 0. There is therefore, as
Moreover 2 - 112 ( / " " , ~1 .*) is uniformly bounded for j ~ 0. It we expected, no single biorthogonal wavelet basis construc-
follows that
tion for which f E CS ( [ 0, 1]) can be characterized by decay
of the (!, ~~~~*), if s > 1. However, the above argument
sup 2 3112 1(!•", ~-1 .*)1 < 00 ~ shows that we can have a complete characterization of cs ([0,
j,k
1]) spaces with 1 < s < 2, if we are willing to use two
sup 2 3112 1(!, ~~~~*)I < oo. different pairs of biorthogonal bases, one pair 1/J, ~for which
j~O
0:!;ik=s;;21-I <P = H, and another pair 'lt, ~such that 'lt is (say) C 2 . Then

3. The equivalence (2.19) now follows immediately from /E Cs([O, 1]) ~ l(f, ~~~~*)I .;; C2-1 <s+lt 2 l
(2.21). •
and
This takes care of s = 1; s > 1 is a different matter. Be-
cause of (2.21 ), and because the extension defined by (2.20) 1(/, ~~~~*)I .;; C2-J<s+ltz) ~ /E CS([O, 1]).

10
 N-2
3. THE ORTHONORMAL WAVELET BASES ON [0, 1] +
A.. edge _ ~
LJ
Hedge A.. edge ~ h edge A.. half (3.2)
IN Y. MEYER'S CONSTRUCTION o/O,k - k,l '¥-1,1 LJ k,m '¥-l,m
1~-N+l m;;.N-1

We start by sketching the basic ideas in the Meyer [21]

construction. At sufficiently small scales, the two edges of with
[ 0, 1] ''decouple'' in the sense that wavelets or scaling func- N-2 -2k+N-2
tions near one edge are supported away from the other edge. edge_
H k,l -
One can therefore explain the idea by just looking at a half r~-N+l m~-N+l-2k

line, i.e., one edge only. Define

hedge _
k,m -
if X~ 0 r=-N+l

if X< 0
and the implicit assumption that hm = 0 for m < - N + 1 or
m > N. For the c/JS~!f, k ~ N- 1, the recursion is the same
as it was on the whole line.
The space V7alf can also be viewed as the space of all restric- c. Compute, for 0 ~ k ~ N - 2,
tions to [ 0, o:) of functions in \'}. Suppose that both cfJ and 1/J
N-2
have been shifted so that support cfJ = support 1/J = [- N + ,/,half ::: Proj' W holt 'I'
,[,half = ,[,half _ ~ (,/,half A.. edge) A.. edge
1, N]. (We work with the N-vanishing moment family 'I' O,k O,k
0 'I' O,k LJ '!'O,k •'¥0,1 '¥0,1 '
1~-N+l
( 1. 7).) Then the cfJ ~.~f with k :s;; - N vanish identically, and
the c/J~~lf with k ~ N - 1 are untouched. One can similarly
and orthonormalize the results to obtain
define i/J~11f; again, these are non-trivially affected by the
restriction only if - N < k < N - 1. In Meyer [ 21] it is N-2 N-1
established (simpler proofs were subsequently given in Le- , 1, edge _
'I' O,k -
~
LJ
B ,[,half
k,/'1' 0,/
+ A.. edge
Ck,lo/0,1 •
marie and Malgouyres [15]) that 1~0 1~-N+l

,-
1. The cfJ 7.~t N + 1 :s;; k :s;; N - 2, are all independent,
d. From 1/JS~Lf = ~m gmcfJ'-!_•(~2k+m one then computes
and orthogonal to the "interior" cfJ7~t with m ~ N - 1.
Introducing the natural definition W 7alt = V 7~{ n (V 7alt) -L, N-2
one finds that ,[,edge _
'f'O,k -
G edge A.. edge
k,/ '¥ -1,/
+ ~
LJ
edge A,. half
gk,m '¥-i.m• (3.3)
2. The interior 1/17.~£ (i.e., k ~ N - 1) are all in W7alf. m;?JN-1
3. However, the i/J7.~f with k = - N + 1, ... , -1 are in
V7au; i.e., Projw'•"!/J7.~f = 0 for -N + 1 :s;; k :s;; -1. with
4. On the other hand, the Projw'•"!/J7.~f with 0 :s;; k :s;; N-
2 are non-vanishing, independent, and orthogonal to the in- N-2 N-2k-2 N-2

terior 1/17.-.:.f with m ~ N- 1.

edge _
G k,l
~ ~ ~ C Hedge
- LJ LJ LJ k,s s,l
r=O m=-N+1-2k s=-N+l
These facts can be exploited to construct an orthonormal
wavelet basis adapted to the half line and the associated sub- and
band filtering scheme; the construction uses the following
steps: N-1 N-2
edge_
gk,m -
~
LJ
B k,rgm-2r + 2 Ck,shf,?!f,e'
a. Orthonormalize the c/J8~l, - N + 1 ~ k ~ N - 2; we ,~o s~-N+l

call the resulting functions cfJ ~?t<. Note that these are 2N - 2
different functions; even though cfJ 8~lt (x) = cfJ 8~~ _1 ( x - k + where we have again g. = 0 for n < 0 or n > 2N - 1.
N - 1) if k ~ N - 1 and x ~ 0, no such thing is true for the
orthonormal c/J8?1". We have The filter coefficients can then be used in a multiresolution
cascade, as usual. More precisely, the whole construction is
N-2 invariant for dilation of x by 2 j, and ( 3.2), ( 3.3) are valid if
c/Jo~r = 2 A*.,cfJSjlf, (3.1) 0, -1 are replaced by j, j - 1.
t~-N+1
The only steps in this program which are not completely
explicit are the two orthonormalization steps. These can be
where A is an invertible (2N- 2) X (2N- 2) matrix. done by a Gram-Schmidt procedure. If the Gram-schmidt
b. Since c/J8~ = ~m hmcfJ'-!_a/~2k+m, for all k ~ - N + 1, it procedure for the edge scaling functions starts with
follows that cf>8~~N+1, and moves on to larger values of k, so that Ak,I is an

11
upper triangular matrix, then the low pass filter coefficients "right" for a right edge (corresponding to (- oo, 0]). With
at the edge will still look "staggered"; i.e., out assumption that support .P = support t/J = [ - N + 1, N],
the "interior" functions on the half line ( -oo, 0] are the
H;dr = 0 if l > 2k + N, <Po.-t. t/Jo.-1 with l ~ - N, and the right edge functions are
therefore labelled with l ;;:;. - N + 1.
h~~e = 0 if m > 2k + N, At sufficiently coarse scales, the two edges start to interact,
in the sense that there are no "interior" scaling functions or
as illustrated in Fig. 3. This results in ''staggered'' 4J~~~·, i.e., wavelets, and all of them touch both 0 and 1. In that case
support 4J ~~:" = [ 0, N + k] . One can of course also choose one needs different definitions of the adapted scaling func-
to Gram-Schmidt orthonormalize the vectors in a different tions and wavelets (see Meyer [21]), leading to different
order, leading to other (non-staggered) filter coefficients at filter coefficients. We do not go into these here; we assume
the edge. Regardless of the procedure followed, we need the that we never reach coarser approximation levels than j 0 ,
overIap matnx, · I.e.,
· th e mner
· prod ucts (A. half , o.yo.t
'YO,k ,A.. half)
, - N + where j 0 is the smallest integer such that 2Jo-I ;;:;. N.
1 ~ k, l ~ N- 2, in order to compute the orthonormalization We have carried out these computations for N = 2, 3, ... ,
matrix A. These inner products can be computed explicitly, 10, starting from the wavelets and scaling functions with sup-
even though we do not have an analytic expression for 4J. port width [- N + 1, N], constructed in Daubechies [9], as
Because ¢(x) = fi ~~=-N+I hn¢(2x- n), we have well as for the less asymmetric variations from Daubechies
[11]. In Tables 1 and 2 we list the adapted edge coefficients
for the least asymmetric case, for N = 2 and 4, for both high
(4J~~lc, <P~~n = ioo dx ¢(x- k)¢(x- l)
and low pass filters. Tables for higher values of N can be
obtained electronically; see the note at the end of the paper.
= }: h,hn
m.n
I'"° dy¢(y - 2k - m) Figures 4 and 5 show what the edge scaling functions and
wavelets look like, in the cases N = 2 and 4, both at the left
and at the right edge. The functions plotted are at resolution
X ¢(y - 21 - n) . 0; i.e., they are in vs•lf, wgalf, where the half line is [0, oo)
for the left edge case, ( -oo, 0] for the right edge. Note that
It follows that the >-..*.1 = (¢8~ic, <PS~n satisfy not only the scaling functions, but also the wavelets are
"staggered"; theN- 1 wavelets at the left edge, e.g., have
N-2
supports [ 0, 2N - 2], [ 0, 2N - 3], ... , [ 0, N], respectively.
Ak,l = }: h,hnA2k+m,21+n• To achieve this, one needs an extra intermediate computation
m,n=-N+l
in step (c) above. First, observe that the t/J S~lf, <P S~1r with k
-N+1~k, l~N-2. (3.4) ~ N - 2 all are linear combinations of the <P 11_")~1 with l ~ 3N
- 4; it follows that the !j,~~lf are as well. On the other hand,
If one takes into account that X.,,, = 0 if r or s ~ - N, and the !frS~~* are all orthogonal to 4>8~4~- 3 , which is a linear com-
X.,.s = 6,,s if r or s ;;:;. N - 1, then ( 3.4) becomes a non- bination of the cpll_•j~* with 3N - 5 ~ k ~ 5N - 3. It follows
homogeneous linear system of ( N - 1) ( 2N - 3) equations that
for as many unknowns (X.u = ">1. 1.* because¢ is real). This
determines the overlap matrix of the ¢ S~lc, so that step (a)
can be carried out. In step (c) we need first the ( t/J S~Y, 0 = }: (!j,g~lf, cpll_•/~3N-5+m)(<I>-I,3N-5+m• <l>o.ZN-3)
m=O
cf>~~re) and then in addition the ( t/1 ~~1f, t/J ~~f) in order to carry
out the orthonormalization. Using t/IS~ir = ~m gmcpl>_•J~2k+m
and (3.1) together with (3.2), we can reduce those inner
_ ~half _
products to expressions involving only the filter coeffi- for aUk- 0, ... , N- 2. If we replace the tfro.* , k- 0, ... ,
~ half ( ~ half ( ~ half -I
cients, the matrix A, and the ( <P l>_•J~" <P ll_•j~s) = X.,,s which we N- 3, by t/lo.k - t/Jo.k, <P-uN-s) [ t/lo,N-2• 4J-1,3N-s)]
already know, so that step (c) too can be carried out with- X !j,g~~- 2 , leaving !frS~~- 2 untouched, then it follows that we
out problems. have obtained a new (equivalent) family, all but one of which
All this shows how Meyer's construction can be translated are linear combinations of the <P ll_•j~1 with l ~ 3N - 4, and
into an implementable scheme. On the interval, we have two therefore supported in [ 0, 2N - 3]. We can repeat this trick,
edges to take into account, so that we have to find adapted and end up with a family ~S~lf, k = 0, ... , N - 2, each
scaling functions and wavelets for the boundary at 1 as well; supported in [ 0, N + k] . These can then be orthonormalized,
they correspond to exactly the same computations, starting starting with the left-most one (i.e. k = 1), so that the stag-
from h., g" listed in the reverse order. We distinguish the two gered supports are preserved in the t/J ~~:".
families by replacing the superscript "edge" by "left" for Several techniques can be used to plot the 4> edge, t/1 edge func-
a left edge (corresponding to the half line [ 0, oo)) and tions. One possibility is to keep track of all the basis trans-

12
 ®®®®@@@@
I II I 1
rfJ?_J,k(x) = 2 112 r/1~~(2 1 x)
I I 1 I I 1
V I I I t/1~ 1 .zj-J(x) =2 112
r/lg.~)(2 1 (x- 1)).
I II 1 I I
I I I I
I 1 1 I
I I1I These statements are proved in Meyer [21] (borrowing ar-
11~1
guments from Meyer [20]). Basically, all that is needed is a
generalization of the proof of Proposition 2.1 to larger values
FIG. 3. Schematic representation of the non-zero H'!_,"" (solid lines), of s. The first step goes through if the wavelets are orthogonal
h~~:: (dashed lines) for k = -2, -1, 0, 1 in the case N = 3. The circled dots to polynomials of degree at least l s J, the second step if all
represent ''interior'' scaling functions; there are four edge scaling functions. the wavelets and scaling functions used (including those at
Note that for k = 1 there are more than six non-zero filter coefficients; for
the edges) are in C' with r > s. Both are easy to establish
general k, N there are 2(N + k) non-zero filter coefficients. We have as-
sumed here that the Gram-Schmidt procedure starts from the leftmost here, as follows. Let us restrict to the half line (the interval
4>~~~, and then works inward. is analogous). Since polynomials up to degree N - 1 can be
written as linear combinations of the cf>o.t on the whole line,
the same is true on the half line, using the cf> ~:f , k ~ - 2N +
2. By construction, the rfJ ~·}~*' j, k ~ N - 1 and the rfJ <-!7,~, j
~ 0, k = 0, ... , N - 2 are orthogonal to those restricted
formations, and to plot simply the appropriate linear com-
scaling functions, so that they are orthogonal to all polyno-
binations of the 4>~~~1 , which can be plotted by any of the
mials of degree N - 1 on [ 0, oo), and N - 1 ~ r ~ l s J. The
available techniques (using the refinement equation in a
second step is automatic: all the edge functions we have in-
"cascade" algorithm, as in Daubechies [ 9] or Daubechies
troduced are finite linear combinations of C'-functions.
[10], using ergodicity of an associated flow, as in Berger
Are there disadvantages of the Meyer construction? There
[ 3], or using subdivision starting from the best initial con-
are several, more or less serious. One (minor) disadvantage
ditions, as in Daubechies and Lagarias [ 12]). One can also
is that we introduced 2N - 2 scaling functions, but only N
use the cascade algorithm directly, with the adapted edge
- 1 wavelets at each edge. This results in dim W ~/1 = 2 1 ,
coefficients, to compute the 2 112 (4>~~:0. cf>-1 ,1), l ~ N- 1,
dim V !?/I = 2 J + 2N - 2; the inequality of these dimensions
which tend to cf>~f"(2 - 11) as J becomes large. This is how
the graphs in Figs. 4 and 5 were plotted.
After this explanation of the construction and the com-
putation of the filter or mask coefficients that would be used
in the implementation of these bases in applications, we tum TABLE 1
to a discussion of advantages and disadvantages. The Filter Coefficients H;:t,r, h;:tr and G~~<', g~~<' (as in (3.2),
The major advantage of the Meyer construction, indeed (3.3)) for the Case N = 2, for Left and Right Edges, for Meyer's
the main motivation for the construction, is that the resulting Construction
wavelet bases are unconditional bases for the Holder spaces
C. More precisely, if tjJ E C', then G•., or g.,,

Left side
{4>~~1 1
, r/l~.·lt; j ,;.;; 0, k ~ N - 1}
k = -1 -1 0.848528137424
U { cf>~f"; k = - N + 1, ... , N- 2} 0 -0.529150262213
k=O -1 0.132287565553 -0.512347538298
U { t/J ;;ge; j ,_;; 0, k = 0, ... , N - 2} 0 0.212132034356 -0.821583836258
1 0.838525491562 0.216506350947
is an unconditional basis for C ( [ 0, oo)) if s < r (with the 2 -0.484122918276 -0.125
understanding that for integer s Zygmund-class spaces have
to be used). On the interval, we have that Right side
k = -1 -3 0.484122918276 0.125
{cP-J ,k; N
0
- 1 ,;.;; k ,;.;; 2 Jo - N} -2 0.838525491562 0.216506350946
-1 0.212132034356 -0.821583836258
U { t/1-J.*; j ~ jo, N- 1 ,;.;; k,;.;; 2 1 - N} 0 -0.132287565553 0.512347538298
k=O -1 0.529150262213
U {tJ!?_J,k;j ~ jo, k = 0, ... , N- 2} 0 0.848528137424

U { t/1 ~ 1.k; j ~ j 0 , k = 2 1 - N + 1, . . . , 2 1 - 1} Note. In every case, the coefficients are listed from left to right ("outer-
most" coefficients first for the left side, last for the right side). The "inte-
is an unconditional basis for c•([O, 1]), where j 0 satisfies rior" h. are given by h_ 1 = (1 + ..J3 )/4..fi, h 0 = (3 + .[3 )14..fi, h 1 = (3
2 Jo-l ~ N and where we define - ..J3 )/4.J2, h2 = (1 - ..J3 )/4..f2.

13
 TABLE 2
The Filter Coefficients H:T, h:T and G;:!r g;:r
(as in (3.2), (3.3)) for the Case N = 4,
for Left and Right Edges, for Meyer's Construction

H,,, or h,; G,,~ or g,,~ H,,1 or ht.J G,,, or g,;

Left side Right side

k = -3 -3 0. 7983434920E + 00 k= -3 -9 -.7576592267E- 01 -.3222260624E- 01

-2 0.6022023488E + 00 -8 -.2963561006E- 01 -.1260377467E- 01
k = -2 -3 -0.3918024327E - 01 -7 .4975940883E + 00 .9927560389E - 01
-2 0.5194149822E - 01 -6 .8037308455E + 00 .2978757620E + 00
-1 -0.4817281609E + 00 -5 .2979273796E + 00 -.8039165735E + 00
0 0.8739021503E + 00 -4 -.9898430109E- 01 .4970737994E + 00
k = -1 -3 0.1774707150E- 01 -3 -.1322822645E - 01 .3110390157E - 01
-2 -0.2352740580E - 01 -2 .3237561509E - 01 -.7612569630E- 01
-1 -0.1232594861E + 00 -1 -.1031248132E- 02 .2424802864E - 02
0 -0.6575127688E - 01 0 -.3127686214E- 02 .7354216650E - 02
1 -0.9620570014E - 01 1 .1791458461E - 02 -.4212306812E- 02
2 0.9850684416E + 00 2 .1517778961E - 02 -.3568796441E- 02
k=O -3 -0.2636405192E- 01 0.5979027428E + 00 k = -2 -7 -.7658857852E - 01 -.3043501638E- 01
-2 0.3495099166E - 01 -0.7926434769E + 00 -6 -.2995738760E - 01 -.1190456375E- 01
-1 0.8114147375E + 00 -0.1659403671E- 01 -5 .4965300858E + 00 .9317065775E - 01
0 0.4440233637E + 00 0.6477069526E - 01 -4 .8099281192E + 00 .2811169922E + 00
1 0.3192581817E + 00 0.9713044594E - 01 -3 .2555397153E + 00 -.7578721046E + 00
2 0.1636579832E + 00 -0.1797030610E- 01 -2 -.6529437751E- 01 .4453190863E + 00
3 -0.4282797155E - 01 -0.3192898087E - 03 -1 .7039069384E - 01 -.1860863715E + 00
4 0.1094933054E + 00 0.8162911886E - 03 0 .7744666189E - 01 -.2220387757E + 00
k = 1 -3 -0.1670338745E - 01 0.4823971249E - 01 1 .7703595608E- 01 -.1783088893E + 00
-2 0.2214378721E - 01 -0.6395169431E - 01 2 .6526723504E - 01 -.1510687917E + 00
-1 -0.1643714751E - 01 0.3010034664E + 00 k = -1 -5 -.9324840456E- 01 -.1463677268E- 02
0 -0.1112580065E- 01 0.1718883936E + 00 -4 -.3647382930E - 01 -.5725128576E- 03
0.2995602574E + 00 -0.8873256413E + 00 -3 .2929411530E + 00 .6432723254E - 02
2 0.2728668922E - 01 -0.3991915695E - 01 -2 .8548427224E + 00 .1377492249E + 00
3 0.8472064764E + 00 0.2462565991E + 00 -1 .1612498760E + 00 .2189246565E + 00
4 -0.4270166998E + 00 -0.1524149055E + 00 0 .1056438759E + 00 .7323866487E + 00
5 -0.3309408518E - 01 -0.9080945357E - 02 1 .2816233337E + 00 - .4805345237E + 00
6 0.8460780753E - 01 0.2321619929E - 01 2 .2385999858E + 00 -.4071236849E + 00
k=2 -3 0.2727915769E - 02 -0.1162436086E - 02 k=O -3 -.5228106380E + 00
-2 -0.3616415322E- 02 0.1541048928E- 02 -2 -.2021221071E + 00
-1 -0.5206157868E - 01 0.2218479707E - 01 -1 .3973740339E + 00
0 -0.2836107693E - 01 0.1208539488E- 01 0 .4284467101E + 00
-0.4413123462E - 01 0.1880547055E - 01 .4477229118E + 00
2 -0.1285294872E- 01 0.5476976811E - 02 2 .3793246746E + 00
3 0.4543141690E + 00 -0.8114432093E - 01 k = I -1 -.8535656929E + 00
4 0.8282235028E + 00 -0.3089428579E + 00 0 .4620817304E + 00
5 0.3000539798E + 00 0.8028339176E + 00 .1836012602E + 00
6 -0.1037443976E + 00 -0.4957693566E + 00 2 .1555526555E + 00
7 -0.1262470890E - 01 -0.2962669767E - 01 k=2 -.6464210153E + 00
8 0.3227612835E - 01 0.7574314021E - 01 2 .7629809380E + 00

Note. In every case, the coefficients are listed from left to right ("outermost" coefficients first for the left side, last for the right side). In this case the
"interior" h11 are

h_, = -.07576571478950, h_, = -.02963552764600, h_, = .4976186676328,

h0 = .8037387518051, h, = .29785779560531, h, = -.0992195435766,
h, = -.01260396726203, h. = .03222310060405.

means that the adapted filters cannot be used for wavelet reasonably large values of N (N ~ 5 in practice), 4> is
packet constructions on [ 0, 1] . smooth and mostly concentrated on a fraction of its support
There is a second, more serious disadvantage, stemming [ -N + 1, N], so that \1</>~~Z- 2 \1 2 is very close to 1, and
from the nature of the construction itself, more precisely \14>~~~N+lll 2 is very small. One consequence is that the over-
from the definition of the 4>3~Lr as restrictions of tf>. For lap matrix (4>3~r, 4>3~)C) tends to be very ill conditioned

14
 2-.--.-
2
1 fl <!>left o .L J~-
0
-1
I v~'' 1

::l r <I> left

0,-1
-2 i
0 2 0 2
provided the denominator in the right hand side is positive.
4[
2~
\jlleft-1
0,0
A similar bound, with h-N+ 1, h-N+ 2 replacing hN, hN_ 1, re-
spectively, holds for r::::
d.x/tf>(x)/ 2, so that
;\
0 ~~---
,.

-2
._,I
Loo d.x\t/>(x-N+2)\2
0 2
= 1- J-N+2 d.x/tf>(x)/2
-N+1

I[ .
1 I
1
<I> right 1
\ 0.-1
\
•3%"' 1
0 \" _j
"~ I 0 _j 1
It follows that the condition number of the overlap matrix,
-2 -1 0 -2 -1 0 i.e.,

0 c::../\ ( N-2

"'~ ',"'
1
[ sup ~ x.k\;(<f,l~~.<f,l~jlf)]

-1

-2 -1
\i
0
.I L,[>.,f 2 =1 k,l=-N+1

X [ .
I Ill
L;(l';l'= 1
N-2
~
L,
k,I=-N+l
-(A.half
/-lk/-ll '¥OJ< ,
A,half)]-1
'¥OJ , (3.5)

FIG. 4. The orthonormal scaling functions and wavelets for N = 2 in

Meyer's construction, at a left edge (i.e., on [ 0, oo)) or a right edge (on is bounded below by
( -oo, 0]). Note that the scaling functions have "staggered" supports. The
2
[1- /h-N+zl 2 ][1- /hN/ 2 - /hN-1/ - /hN/ihN-11]
indexing convention is explained in the text.

x r1- 2 2
lh-N+1/ - lh-N+zl - lh-N+1iih-N+2ir
1
when N is large. The following trivial computation makes x [ihN/ 2 + lhN-1ilhNir 1. (3.6)
this more quantitative:
For N = 4, (3.6) is of the order of 10 3 , but its value increases
rapidly with N; for N = 10, it is -4 X 10 7 • Since (3.6) is
only a very coarse lower bound for (3.5), the true condi-
N
~ 2hmhm
f1 dxtf>(2x + 2N- 2- n) tioning number is in fact even larger. This ill conditioning
makes the computation of the adapted filter coefficients near
m,n=-N+1 0
the edges rather tricky; for N larger than 6, for instance, we
X 4>( 2x + 2N - 2 - m) already needed quadruple precision. The computation of the
overlap matrix itself, along the lines outlined before, is quite
:,;;; IhN / 2 f dy /tf>(y + N - 2) /
2 tricky itself as well, involving the inversion of a large badly
conditioned matrix.

+ /hN-1/ 2 f dy/tf>(y + N- 1)/

2
The disequilibrium between \14>~~- 2 \1 2 and !I<P~~~N+l\1 2 also
expresses itself in other ways. One application of wavelet
bases and multiresolution on the interval is the "natural"

+ 2/hN/ihN-11 f dy/tf>(y + N- 2)1\t/>(y + N- 1)\;

extension of functions living on the interval to functions on
the whole line. Since the edge-wavelets and scaling functions
can all be written as linear combinations of restrictions of
dy/tf>(y + N- 2)\ 2 :,;;; 114>11 2 = 1 (a very coarse
fo 1
using whole-line functions, one can extend them trivially by
bound), this leads to "gluing their tails on again," i.e., by replacing every <P~~j

15
 ,I
2

1\ .:~ I ~[ ·' <!>left

0,1
1 i\, <!>left
0,0

ofJ 0 2
lv--
4 6
0

-1
0
~ ~II
~ I
('',,,
IJI I
\

2
I
i:'

4 6
'If
0 I
-1 I
I \,

0 2 4 6
2 . 3

:~
, ~ nII~
.,. ,
I ',
2
1
~
II
II
<!>left
0.-2 !
<!>left
0,-3

0
~
I'
0 II ', -1 v
-1 ' -2
0 2 4 6 0 2 4 6 0 2 4 6

2 2'

on~v~j
11 'l' left 'l'left
1 4l
/': 0,2 0,0 j
o:
0
-1
-2
h ~I • -4

0 2 4 6 0 2 4 6 0 2 4 6
FIG. Sa. The orthonormal scaling functions and wavelets for N = 4 in Meyer's construction on the half line [ 0, oo). Note that the six scaling functions
<!>~"andthe three wavelets 1/J~" have supports (0, 4 + k].

by cP-,.k· If this is done for every edge term in the expansion 4. A DIFFERENT FAMILY OF ORTHONORMAL
of a function f on [ 0, cc)' the result is a smooth function r·' WAVELET BASES ON [0, l)
extending f to !R, with the appealing property that the high
frequency components in f spread out less to ( -oo, 0] than Our starting point is again the N vanishing moment family
the low frequency components. At any scale j, the extension ( 1.7); our goal is to adapt this family in such a way near the
is limited to [ -z- 1(2N- 2), oo). With the present construc- edges of the interval that we obtain unconditional bases for
tion this does not work so well in practice, however: because CS ( [ 0, 1]). As emphasized already in Sections 2 and 3, this
cPo~~"rHJ = cfJS~~N+I/IIcfJS~~N+III, the extension of cfJ&~~~+J to the can be achieved by retaining the interior scaling functions,
left half line has a huge amplitude, ranging up to and adding adapted edge scaling functions in such a way that
maxi<P(x)I!II<PS~~N+tll ~ llcfJ3~~N+III- • In fact, this is the
1
their union still generates all polynomials on [ 0, 1], up to a
reason Jawerth, in an application involving such exten- certain degree. Our construction in this section results from
sion operators for surface design in collaboration with taking this prescription literally. We start again by illustrating
Dahlberg, decided to develop a construction different from the construction on the half line [ 0, oo). The "interior" scal-
Meyer's. ing functions at scale 0 are the cPo,k with k ;, N - 1; they are
A third instance where one can feel the imbalance among supported on [ 0, cc). By themselves, the interior cPo.k do not
the cfJS~t is in the plots of the edge functions. Typically, even generate the constants on [ 0, oo), as is clear from cPo.k( 0)
cfJ~~~N+I has much faster high amplitude oscillations than cfJ = cfJ( -k) = 0 for all k ;, N - 1. Let us therefore add the
itself (the same oscillations are of course present in the tail constants "by hand." We define an edge function cfJ 0 by
of cfJ, but with exceedingly small amplitude); because of the
orthonormalization procedure, this oscillatory behavior can
c/J 0 (x) = 1- ~ cfJ(x- k).
typically spread to several edge scaling functions, as in Fig.
k~N-l
5. This is disturbing: we like to think of wavelets at a certain
scale as corresponding to a frequency band of about one oc-
The interior <Po,~c and this edge function cfJ 0 together generate
tave, and the edge wavelets in Figure 5, corresponding to just
all the constants on [ 0, oo). Moreover, because
one scale, clearly cover may octaves.
In the next section we present a different construction that
avoids all these problems, while still giving good bases for ~ c/J( X - k) = 1,
the C'-spaces. /c=:-Xl

16
 ·····~
$right

~:3 1
0,-2

I
/ /'\ \
0 ~ \/u
-1
-6 -4 -2 0 -6 -4 -2 0 -6 -4 -2 0
2 ~--------~
$right $right :'r,'_

r
0,0 0,2
2 1 i
I
I,
0 ·--J
-1 '-------~--__J
-6 -4 -6 -4 -2 0

~--j 'i/'i
2

0 ,___ ___. 01
(',1 /\ 1
\ \ill
\j 'I
' 1
~~~
,1 right-~ ~·
'1'0,-1

right right \!
oI ,
i 1·

-1 'l'o.-3 -1 > 'l'o.-2 v

-1 .
-6 -4 -2 0 -6 -4 -2 0 -6 -4 -2 0
FIG. Sb. Same as Fig. Sa, but on the half line ( -oo, OJ.

we also have, for 0 ,.;; x < oo, Similar inclusions hold immediately if we scale by other in-
teger powers of 2, and we still have a hierarchy of nested
N-2 N-2
spaces.
t/J 0 (x) = ~ t/J(x- k) = ~ t/J(x- k),
k=-00 k~-N+I
This is essentially all there is to the construction we pro-
pose here. If we want the edge + interior scaling functions
showing that t/J 0 has compact support. It also shows, inci- to generate more polynomials than only the constants, then
dentally, that t/J 0 is orthogonal to all the interior tPo.J<· The we have to add in, by hand, more edge functions (for the
only thing that we have to check is that by adding functions polynomials up to degree L, we add in total L + 1 functions).
in this ad hoc way we do not leave the framework of a mul- If we work on the interval, then the same has to be done at
tiresolution hierarchy. We have, however, the other edge as well. On the other hand, as pointed out
earlier, for many applications it is desirable to have exactly
N+2k 2 i scaling functions of scale j when working on [ 0, 1]. Let
t/J(x - k) = .J2 ~ h 1_utfJ(2x - /) us count how much room this leaves us for adding extra
1~2k-N+I
functions at the edges. If we start from a minimal support N-
vanishing moment wavelet, then support t/J = [ - N + 1, N],
and and for j sufficiently large we have exactly 2 i - 2N + 2
interior scaling functions at scale j. This leaves room for
adding N - 1 ad hoc functions at each edge, so that the total
I~N-l k~N-l
family can generate polynomials of degree at most N - 2.
The unaltered whole-line scaling functions can generate all
3N-4 [ (/+N-1 )/2 J
polynomials up to degree N - 1 (see Section 1), so that we
= t/J 0 (2x) + I t/J(2x- /)[1 - fi ~ h 1_2k], seem to have "lost" one degree. In order to recover this one
extra degree (and so be able to characterize the cs ([0, 1])
I~N-l k~f (1-N)/2]

where we have used that hn = 0 for n < -N + 1 or n > N spaces for the same range of s as we could on all of ~), we
have to make room for one extra function at each edge of the
and 2." h2n = 11..J2 = 2." h2n+1· It follows therefore that
interval. For this reason we abandon the two outermost in-
Span{t/J 0 , t/Jo.k; k;;;;. N- 1} terior scaling functions, which corresponds to retaining only
C Span{t/J 0 (2·), tP-t,k; k;;;;. N- 1}. the tPo.J< with k ;;;;. N rather than k ;;;;. N - 1 on the half line.

17
 Let us now be more precise. For given N, we define the
• k ~
N edge functiOns <P , k = 0, ... , N- 1, on [0, oo) by

i/>k(x) = n~O
2N-2 (
:
)
c/J(x + n- N + 1) If I * 0 is odd, then m ~r l ( 1r) = 0 for r = 0, ... , N - 1
insures that i;><kl(27r/) = 0. This in turn, by applying the
equation again, implies that i/><kl(27r/) = 0 if I 0 is even. *
n~*
2N-2 ( )
= : c/J(x + n - N + 1) ( 4.1) Consequently the left hand side of ( 4.4) is constant; by in-
tegration from 0 to 1 one immediately establishes the rest of
( 4.4). Note that the C 1 can be computed from the hn in the
These are all compactly supported, and their supports are following way,
staggered, i.e., support i/> k = [ 0, 2N - 1 - k). The following
proposition summarizes some of their properties:
PROPOSITION 4.1. TheN functions i/>*, k = 0, ... , N - cl = fi .L hm
m
f dx x 1<P(2x- m)
1, are independent, and orthogonal to the cPo,m, m ~ N. To-
gether with the <Po.m, m ~ N, they generate all the polyno-
mials up to degree N - 1 on [0, oo). Finally, there exist
= 2-1-1/2 L hm
m
f dyc/J(y )(y + m )I
constants ak.IJ bk.m such that

k 3N-2-2k

i/>k(x) = ,L ak. 1i/> 1(2x) + ,L bk.mcP(2x- m). (4.2)

1~0 m~N

with Mr = 2- 112 Lm hmmr. In particular, M 0 = 1, so that

Proof
1. The independence of the 4>* follows immediately from C1 = (2 1 - ~
1) -1 r~
1
(I)r MrCt-r·
their staggered supports. Orthogonality with respect to the
cPo.m, m ~ N, is also immediate from ( 4.1) and the ortho-
normality of the c/J 0..,, n E ?L. 4. As a consequence of ( 4.4) we have, for k = 0, ... , N
2. For every k = 0, ... , N- 1, the function (1/k!)y(y - 1,
- 1 ) · · · (y - k + 1) is a polynomial in y of degree k. It
follows that these functions can be transformed by a simple ,L nkc/J(x - n - N + 1)
(triangular) transformation into the polynomials 1, x, ... , nEZ

xN - 1 • This is true in particular if y = n. Define now

= ,L [x- N + 1 - (x- n)]kc/J(x- n)
2N-2

4>*(x) = ,L nkc/J(x + n - N + 1),

n=O

k = 0, ... , N- 1, (4.3)

with again x E [0, oo). Then the i/>k and c/Jo.m, m ~ N generate we denote these polynomials pk(x). They are of degree k and
all the polynomials of degree up to N - 1 if and only if the their leading order term is exactly x*, since C 0 = 1. It follows
same holds for the 4>* and <Po.m, m ~ N. that theN polynomials Pk are independent; together they gen-
erate all polynomials of degree up toN - 1. From ( 4.3) we
3. We now establish some properties of the 4>*. We start
see that for x E [ 0, oo)
by noting that, fork = 0, ... , N- 1,

"'
.L <x- n)k<P<x- n) = ck := f"' dxx*<P<x). (4.4)
Pk(x) = (-1)k4>*(x) + ,L n*c/J(x- n- N + 1),
n=l
nEZ -oo

To see this, note that the left hand side is a periodic function so that the 4>*. together with the c/Jo.m, m ~ N, do indeed
with period 1; it is therefore completely characterized by its generate all the polynomials up to degree N - 1.
Fourier coefficients, which are, up to a constant, equal to the 5. It remains to establish the recurrence ( 4.2). Again, it is
derivatives i;><kl(27rl). Now sufficient to prove a similar recurrence for the 4>*. We have

18
 2N-2 N Substituting ( 4.6) into ( 4.5) immediately leads to the desired
J>*(x) = 2: n* 2: fihmcfJ(2x- 2N + 2 + 2n- m) recurrence. II
n=O m=-N+l
Remark. The only thing we have not spelled out in nu-
merical detail is the transition from the polynomials to en
the n k (and back again), but this triangular transformation is
t~-N+l n=O both well known and easy to derive (the entries of the tri-
3N-2 N-1 angular matrix satisfy easy recurrence relations). Otherwise
+ fi 2: cfJ(2x- r) 2: (n + N- 1 )kh2n+n everything in this construction is completely explicit. In par-
ticular, one easily checks that ak.k = 2 -k.
r=N n=-N+I
Define now
( 4.5)
V~~ =Span[ {«/l(2i·); k = 0, ... ,N- 1}
where we have used x ;?: 0, and we implicitly assume, as -------
always, that hm = 0 for m < - N + 1 or m > N. The second
u {cP-J.m; m ;?: N}].
term is already of the right form (the range on r is more (The superscript "left" stands here for a half line with an
limited in ( 4.2) because of the staggered supports for the endpoint and correspondingly adapted scaling functions at
¢k, a property which the J>k do not share); it remains to check the left end.) Proposition 4.1 establishes that the vt:7 consti-
that the first term can be recast into a linear combination of tute a multiresolution hierarchy,
the J> 1(2x), I = 0, ... , k. Note first that we can lift the
restriction on n in this sum: for n < 0 or n > 2N - 2 the · · · C V ~ft C V ~eft C V ~ft C V t:~ C V t:~ C · · · .
h2n-2N+2+t = 0 if- N + 1 ~ I ~ N- 1. Next note that
Since U1 Span{cfJ-J.m; m;?: N} already equals L 2((0, oo)),
we also immediately have U1 Vjert = L 2 ([0, oo)). One can
obtain an orthonormal basis for V ~tt by orthonormalizing the
¢ k, since they are already orthogonal to the orthonormal cPo.m;
scaling them leads to an orthonormal basis for every vj•ft. If
LJ h 2n-N+l-(N-I-!) (2n )*
-- 2-k+l/2""
one orthonormalizes by a Gram-Schmidt procedure, starting
with ¢N-t, and working down to lower values of k, then the
= 2-k+l!2 ±(k)(N- I-
r~o r
1)'
resulting orthonormal c/Jlef', k = 0, ... , N - 1, still have
staggered supports: support c/Jleft = [0, N + k]. (We have
chosen this indexing because the c/Jieft replace, in a way, the
X 2: h2n-N+I-(N-t-1)(2n- (N- [- 1))*-r c/J 0 ,~, I ~ N - 1. Consistent with the other notations, we de-
note by c/Jt:f).k(x) the functions 2 112c/Jiett(2 1x).) To carry out

= 2-k+l/ 2 ±(k)(N- /-
r~o r
1)'/.lw-t,k-n
(4.6)
the Gram-Schmidt orthonormalization explicitly, we need
again the overlap matrix ( ¢ *, ¢ 1). To compute this overlap
matrix, we use the recurrence ( 4.2). For k = 0, for instance,
we have

where we have used that m~s>(7r) = 0 for s = 0, ... , N-

a~.o ~ 11¢ 11
3

~ b~.m ~,
2
0 2 0 2
1, which implies Ln hn( -1 )"ns = 0, hence 11¢ 11 = +
m=N

from which we obtain l cb 0 11 2 • It then follows that

m

3N-4
1 1 ""
+ ao,oal,l 2(cP , cP ) + 2
~a ~~
L,. ba,mbl,m,
for r = 0, ... , N - 1, so that .'<!=!>

2: h2m-r(2m)' = f:.Lr.r = 2: h2m+I-r(2m + 1)' leading to an explicit formula for ( c], 0 , c], 1 ), since llcb 0 11 2 is
m m known. It is now clear how to proceed for higher values of
k. If all the (c],*, c], 1) for 0 ~ k, I~ K- 1 are known, then
we can compute the ( c],*, c],K) fork = 0, ... , Kin that order.
We obtain equations of the type

19
[1 - &auaK,K](~*, ~K) 1 is of course analogous. To simplify notation, we return to
the half line [ 0, oo). We define there W ~ett = V~~{ n (V ~a it ) J_;
= linear combination of the ( ~ ~K) with I
1
, <k the ifJ1.m, m ~ N all belong to W Jett, and we are looking for
N extra functions in W )eft, orthonormal to these 1/JJ.m. The
and of the (~ 1 , ~m) with I, m ~ K +constants. following proposition tells us where to look:
Since !a*.*aK,K = 2-k-K-I ~ L this immediately leads to a PROPOSITION 4.2. Define the functions ~*, k = 0, ... ,
numerically stable recursive scheme for determining the N- 1, by
(~*, ~ 1 ). (Equivalently, we have to invert an N(N + 1)/2
N-1
triangular system with a condition number bounded by 2, a ,],k = A, left _ ~ (A, left A, left) A, left
'+' 'P-1.1< LJ 'P-1/<• 'POm 'PO,m· ( 4.8)
distinct improvement over the situation in the previous sec- m=O
tion.)
The orthonormal <Pieft, constructed with staggered supports Then the~* are N independent functions in W~ft, orthogonal
along the lines indicated above, satisfy a recursion relation to the 1/Jo.m, m ~ N
similar to ( 4.2) and inherited by all the scales j. Explicitly, Proof
there exist constants Hi~~t and hie! (which can be computed 1. The ¢ 1:_fiJ" k = 0, ... , N - 1, are by construction or-
explicitly from the ak,t. bu in ( 4.2) and the orthonormali- thogonal to the <P- 1,m, m ~ N. Since the cPo.t. I/Jo.1 with I ~ N
zation procedure) such that are all linear combinations of the <P- 1.m, m ~ N + 1, it follows
that <P~~t.k .l ¢0 ,~, 1/10 •1 with I~ N. Consequently~* is nothing
N-1 N+U
but the orthonormal projection of¢~~~.* onto W~ett. Moreover,
A, left ~ Hiett A, left + LJ
~ h left A, ( 4.7) as a linear combination of functions orthogonal to the 1/Jom,
'P-j,k = LJ k,l 'P-j-1,1 k,m'P-j-lm·
/=0 m=N m ~ N, ~*is obviously orthogonal to them as well.
2. It remains to establish linear independence. First, note
that the 2N functions c/J~~.k and <P~it, k = 0, ... , N- 1, are
All this was on the half line. If we work on the interval all independent. This follows immediately from the stag-
[ 0, 1], and we start with a scale fine enough so that the two geredness of their supports: support c/J~~t·* = [0, N/2 + k/2],
edges do not interact, i.e., 2' ~ 2N, then there are 2' - 2N support¢~~~ = [0, N + k]. It follows that the~*, k = 0, ... ,
interior scaling functions <P-J.N• ... , <P-J.Z'-N-~> and we add N - 1, and the <P~~t, I = 0, ... , N - 1, also constitute a
N functions at each end. At 0, we have the functions defined family of 2N independent functions. This is only possible if
above, <fJ!l_,.k = ¢~~.*' k = 0, ... , N- 1. To obtain the extra the~*, k = 0, ... , N- 1, are independent. •
functions at 1, we first have to repeat the construction above Note that all the~* are supported in [ 0, 2N - 1). Because
for the half line ( - oo, 0] (or equivalent! y, repeat the con- of the recursion relation ( 4.7), the ~* can be written as a
struction on [0, oo) for the reflected coefficients h! = h-n+ 1 ). linear combination of cfJ ~~t.l and cfJ -l.m:
This leads to functions <P~~~~(x) = 2 112</Ji..ight(2'x), k = -1, N-1 JN-2
-2, ... , - N with support [ k - N + 1, 0 J (again the indexing
has been chosen so that the k-values complement the interior
~k = ck.tcP~~t.l +1: dk,mcP-l.m· 1:
( 4.9)
1=0 m=N
functions <P-J,m• m ~ - N- 1, where the outermost interior
function has not been retained, as before. If everything is The supports of the~* are not staggered. We can replace the
recast in terms of the #-construction, then <P i..ight ( x) = ~* by an equivalent family with staggered supports, by es-
<P !:'t~* (- x).) On the interval, the adapted scaling functions sentially the same trick as in Section 3.
at 1 are then given by <P~ 1 . 2 J_ 1 (x) = <P~~~'-- 1 (x- 1), I= 1,
PROPOSITION 4.3. There exists a family of N independent
... , N. Together, the ¢G_ 1·*' k = 0, ... , N- 1, <P-J,m• m =
functions~* in W~tt, k = 0, ... , N - 1, all linear combi-
N, . . . , 2 1 - N + 1, and <P ~ 1 , r = 2 1 - N, . . . , 2 1 - 1,
nations of the~*, so that support 1/J* C [0, N + k].
constitute an orthonormal basis for the 2'-dimensional space
v ~~·11. Proof
We now tum to the wavelets rather than the scaling func- 1. First note that the ~* are all orthogonal to <Po.ZN- 2
tions. As usual, we define W~/1 = V~~~1 n (V~~· 1 l)L. From L~=-N+l hmcP-lm+4N-4· It follows that, for all k,
dimension counting, it immediately follows that dim W ~/I
= 2'. On the other hand it is easy to check that the 21 - 2N
functions 1/J-J,m• m = N, ... , 2 1 - N- 1, are all in W~~· 1 1.
Since they are all orthonormal, we therefore need to add an 2. If dk,3N-z = 0 for all k, then choose ~N-l = rpN-I, and
extra 2N wavelets (Nat each edge) to provide an orthonor- proceed to the next step. If dk,JN-z *
0 for some k, reorder
mal basis for W~/1. We show here how to construct those *
the~* so that dN-I,JN-z 0, and define ~N-l = ~N-I, and,
at 0, the left end of the interval; the right end construction at fork < N- 1, if, <IJ,k = if,*- duN-z(dN-1.3N-z)- 1if,N- 1 • It

20
follows that the fi, <1>·* satisfy a recursion relation similar to same indexing conventions as for the scaling functions, we
( 4.9), with the upper limit on the sum over m replaced by introduce, for k = 0, ... , N - 1
3N- 4. Consequently support fi, (l).* C [0, 2N- 2] fork=
0, .. . ,N- 2.
3. We can now repeat these steps, ending up with the de-
sired support property after N - 1 steps. II The following theorem then holds.
The staggered support functions~* satisfy a relation of the THEOREM 4.4. Choose any J so that 2 1 ~ 2N. Then the
same type as ( 4.9), except that the upper bound on m is now collection
N + 2k. In a final step, these ~* can now be orthonormalized;
if we do this by a Gram-Schmidt procedure starting from k = U [ { !fJD_1.*; k = 0, ... , N - 1}
j~J
0 and working up to larger values of k, the staggered supports
are preserved, and we end up with an orthonormal family U { 1/1-},m; m = N, ... , 2 1 - N- 1}
1/Ji•fl, k = 0, ... , N - 1. For any j E 7/_ we define again
1/Jt:.~.k(x) = 2i 121/Ji•tt(2ix). Together with the 1/1-1, , m ~ N, the u {I/I~J,2f-N+k; k = 0, ... ' N- 1}]
1/Jt:.~.k• k = 0, ... , N - 1, provide an orthonormal basis for
U {cf>D_J,k;k= 0, ... ,N-1}
Wt:.f. Moreover, there exists constants Gl~~~ and gi•;!, such that
U {cP-J,m;m =N, . .. , 21 -N-1}
N-1 N+U
1/Jt:.t}.k = L Gl~~tcf>t:_~t_l,/ + L gi•;!,ct>-j-l,m• (4.10) U {cP~J.zl-N+k; k = 0, ... , N- 1}
1=0 m=N

is an orthonormal basis for U([O, 1]). If r is the Holder

This completes our explicit construction, at least at a left end. index of cf>, 1/1 (i.e., cf>, 1/J E C), then this collection is also
The same of course has to be repeated at a right end. an unconditional basis for CS( [ 0, 1]) for s < r; a bounded
These right and left end functions can then be used to put function f is in CS ( [ 0, 1]) if and only if
together adapted wavelet bases on the interval; following the

l<f , -j,21-N+k >I .;;;

.1,1
'I'
c 2 -,<s+l/2),
TABLE 3
.
The Filter Coe ffic1ents
. Hedge hedge
k,t , k.J an
d Gedge
kJ g edge
k.J
( • (
as m )
4 .7 ,
(4.10)) for the Case N = 2, for Left and Right Edges where C is independent of j and m, k.
Proof The fact that we have an orthonormal basis fol-
H •. , or ht.t Gt.t or g.,, lows from the whole construction. The statements about
Left side CS([O, 1]) (where Zygmund spaces have to be used if sis
integer, as always) follow from observations made in Sec-
k=O 0 0.6033325119E + 00 -0.7965435169E + 00
1 0.6908955318E + 00 0.5463927140E + 00 tions 2, 3: the 1/J o_ 1·*' 1/1 _J.m , 1/1 ~ 1,1 are all orthogonal to all
2 -0.3983129977E + 00 -0.2587922483E + 00 polynomials of degree N- 1 ~ Lr J (since they are orthog-
k= 1 0 0.3751746045E- 01 0.1003722456E - 01 onal to the cf>D_,.k, cP-J,m and <P ~1.1 which together generate all
0.4573276599E + 00 0.1223510431E + 00 polynomials up to degree N - 1 on [ 0, 1]), and, as finite
2 0.8500881025E + 00 0.2274281117E + 00 linear combinations of C'-functions, they are in C' them-
3 0.2238203570E + 00 -0.8366029212E + 00
4 -0.1292227434E + 00 0.4830129218E + 00 selves. II
We have therefore achieved our goal: we have a basis with
Right side
the same good smoothness-characterization potential as
k = -2 -5 0.4431490496E + 00 0.2315575950E + 00 Meyer's interval construction, and with moreover the
-4 0.7675566693E + 00 0.4010695194E + 00 "right" number of scaling functions, and resulting from a
-3 0.3749553316E + 00 -0. 7175799994E + 00 numerically stable procedure. As on the whole line, we have
-2 0.1901514184E + 00 -0.3639069596E + 00
-1 -0.1942334074E + 00 0.3717189665E + 00 no explicit analytic expression for the wavelets and scaling
k = -1 -3 0.2303890438E + 00 -0.5398225007E + 00 functions on the interval. For practical applications, all that
-2 0.4348969980E + 00 0.8014229620E + 00 is really needed are the filter coefficients; in addition to the
-1 0.8705087534E + 00 -0.2575129195E + 00 hm, gm = ( -l)mh2N+l-m• we now also have the H~J, hl•;!,,
m~r' gl~~ (same at right) and their counter-parts. The goal
Note. In every case, the coefficients are listed from left to right ("outer-
most" coefficients first for the left side, last for the right side). The "inte- of all the explicit manipulations above was to obtain the con-
rior" h• are given by h_ 1 = (1 + fj )/4.fi, ho = (3 + fj )/4.fi, h, = (3 stants Hl".r, hi•!, G1~r. gi•! numerically, so that they could
- fj )14.fi, h 2 = (1 - fj )/4.fi (same as for Table 1). be used in numerical applications. We have carried out all

21
these computations for theN vanishing moment families with in Section 3; the overall behavior of their oscillation looks
support [ 0, 2N - 1] and closest to linear phase (as con- consistent with the behavior of the whole-line functions. It
structed in [10]), for N ::: 2, 3, ... , 10. We list the coeffi- is also quite striking that on [ 0, 1] the N functions cp iert, k :::
cients for N ::: 2 and 4, for both left and right edges, in Tables 0, ... , N - 1, are pure polynomials (of degree N - 1). This
3 and 4. (The full tables, for other values of N, can be ob- is natural, since all the scaling functions together on [ 0, oo)
tained electronically; see the note at the end of the paper.) generate the polynomials up to degree N - 1; since the in-
The corresponding scaling functions and wavelets for both terior scaling functions cp 0 ,., m ~ N, only start kicking in
edges are plotted in Figs. 6 and 7. These plots have again from x ~ 1 onward, the N adapted scaling functions at the
been obtained via the cascade algorithm. The adapted wave- left edge cannot be anything but polynomials themselves.
lets and scaling functions are now less oscillatory than those (The same is true of course for the right-adapted scaling

TABLE 4
10
The Filter Coefficients H':/", h':,t' and G~ g~K• (as in (3.2), (3.3)) for the Case N = 4, for Left and Right Edges

H,.1 or h,.1 G,.1 or g,,J H,, 1 or h,,1 G,,J or g,,1

Left side Right side

k=O 0 .90975392E + 00 -.75739704E - 01 k = -4 -11 .32210279E - 01 .75771168E- 01

1 .40416589E + 00 .32543918E + 00 -10 -.12598952E- 01 -.29637661E - 01
2 .89040317E- 01 -.68434906E + 00 -9 -.99108040E - 01 -.49764705E + 00
3 -.l1984192E- 01 .62004423E + 00 -8 .29771110E + 00 .80379367E + 00
4 -.30429084E - 01 -.18858513E + 00 -7 .80394959E + 00 -.29778999E + 00
k = 1 0 -.27285141E + 00 .16659597E + 00 -6 .49779209£ + 00 -.99201918E- 01
.50908154E + 00 -.48478431E + 00 -5 -.30235885£ - 01 .12853256E - 01
2 .62364244E + 00 .35646355E + 00 -4 -.67659162£- 01 .28761869E - 01
3 .46284008E + 00 .48398963E + 00 -3 -.17709184£- 01 .75281640£- 02
4 .24674764E + 00 -.60575438E + 00 -2 .19132441£- 01 -.81331898£- 02
5 -.17669532£- 01 .34518331£ - 01 -1 -.67756036£ - 02 .28803051£ - 02
6 -.45173645£- 01 .88249013E - 01 k = -3 -9 .32148741£ - 01 .75756811E - 01
k=2 0 .12611793£ + 00 .20825353E + 00 -8 -.12574882£ - 01 -.29632043E- 01
1 -.23085573£ + 00 -.40182281E + 00 -7 -.10276635£ + 00 -.49684271£ + 00
2 -.52799236£- 01 -.68721488E - 01 -6 .29864734E + 00 .80336362E + 00
3 .21926518E + 00 .33021352E + 00 -5 .81641197E + 00 -.30158150£ + 00
4 .46348071E + 00 .55802131E + 00 -4 .46061686E + 00 -.95048353£- 01
5 .70011973E + 00 -.59949744£ + 00 -3 .29213680E - 01 .13887372£ - 01
6 .41203257E + 00 -.69091991£- 01 -2 -.13907160E + 00 .30620323E - 01
7 -.26222762E- 01 .27853571E - 01 -1 .12900783E - 01 .45819595£- 02
8 -.67040697£- 01 .71209989£ - 01 k = -2 -7 .41268408£ - 01 -.69515191£- 01
k=3 0 -.29079804£- 01 .65485008E - 01 -6 -.16142013£- 01 .27190655E - 01
1 .59928071£ - 01 -.13495243£ + 00 -5 -.15813389£ + 00 .44993410£ + 00
2 .61764279£- 02 -.13908739£- 01 -4 .39377582£ + 00 -.67353457£ + 00
3 -.40211000E- 01 .90551421E - 01 -3 . 75400048E + 00 .68118565£ - 01
4 -.39525870£ - 01 .89008570E - 01 -2 .44880018E + 00 .50928676£ + 00
5 -.52599061£ - 01 .37334445£ + 00 -1 -.21916264E + 00 -.27262735E + 00
6 .32894945£ + 00 -.84046537£ + 00 k = -1 -5 .64379349E - 01 -.99241950£- 01
7 .79663789£ + 00 .31568494£ + 00 -4 -.25191808E - 01 .40503097E + 00
8 .49011302£ + 00 .12029765£ + 00 -3 .59477713E - 01 -.64952976E + 00
9 -.29432878£ - 01 -.13070202£ - 01 -2 .39191428E + 00 .60406780E + 00
10 -.75247623E- 01 -.33415072£ - 01 -1 .91547054E + 00 -.19827799E + 00

Note. In every case, the coefficients are listed from left to right ("outermost" coefficients first for the left side, last for the right side). The "interior"
h, are

h_, = .03222310060405, h_ 2 = -.01260396726203, h_, = -.0992195435766,

h0 = .29785779560531, h, = .8037387518051, h 2 = .4976186676328,
h, = -.02963552764600, h.= -.07576571478950.

Note that these are in the reverse order from the interior coefficients corresponding to Table 2. In order to get the same ordering as in Table 2, it suffices
to exchange "left" and "right" below, and to change the ordering everywhere.

22
 1.5

~·
2 of course-this amounts to adding a few extra sequences of
1 . 0,0 1.0 edge filter coefficients); if not, then we can live with just 2 1
q,left···l 0.5 - 2(L - K) scaling functions at each scale j.
0 ·····~
I The construction in Section 4, and the variation above,
-1 I assume that we want the scaling functions to generate
0 2
all possible polynomials up to a certain degree. If the
3 0 2 3
interval wavelets are used to solve a differential equation,
left then it may be useful to adapt the construction so that all
lj/0,0 the scaling functions and wavelets involved satisfy cer-
tain prescribed boundary conditions. Auscher [ 2] adapted
the original construction by Meyer in this way; his
scheme carries over entirely to the present construction
0 2 3 0 2 3 (with more numerical stability). The construction by Le-
marie-Rieusset, which is essentially the same as ours, ob-
tained independently, was carried out in view of this ap-

~ ~ ?~-
0.5
~~ri~~t . ::i .;,~U plication.
The same ideas apply of course to biorthogonal wavelet
bases. If one starts from a choice with (anti) symmetric
0 wavelets and scaling functions, with filters with an even
-0.5 L___-~-~----J
number of taps (i.e., q,, rjJ, ;J>, ijJ all have their symmetry
-3 -2 -1 0 -3 -2 -1 0 axis at 1/2, with q,, 4> symmetric and tjl, ijJ antisymmetric),
then the adapted scaling functions and wavelets at the right
·/~ ~ f ljl~i.~~t ; \.. ·~] edge can be chosen to be the mirrors of their left edge
equivalents. Since orthonormality is not an issue here, but

-~~-····\} ~
/ i
\ /v !
is replaced by biorthogonality, there is more freedom in
\ .JI right
1 lj/0,·2
•. i the choice of the edge functions. One can optimize the
adapted edge filters to have, e.g., a total sum of absolute
-2 -1 0 -3 -2 -1 0 values of their entries as small as possible. If the number
FIG. 6. The two edge scaling functions and wavelets in the case N = 2, of taps in the filters is odd, then it is impossible to have
for left and right sides (i.e., on [0, oc) or ( -oc, 0]). Note that we have two exactly 2 1 scaling functions at level j, and have adapted
adapted scaling functions at each edge, rather than one, because we have scaling functions at the edges that are mirror images of each
"sacrificed" the first interior function so as to be able to generate all linear
other. This construction seems therefore less appealing; the
polynomials (see text).
obstruction to mirroring is lifted if we allow 2 1 + 1
scaling functions at level j.
Note that in all these constructions we have restricted our-
selves to sufficiently fine scales so that the edges do not in-
functions on [ -1, 0] .) In the next section we discuss some teract. In theN vanishing moment case with minimal support,
more properties of and variations on our construction. this meant that our coarsest scale J was such that 21 ;;.. 2N.
What happens if we want to go further? Basically, there is a
5. DISCUSSION OF THE NEW CONSTRUCTION lot of freedom. If, for instance, N is a power of 2, so that 2 1
= 2N, then at the next stage we have exactly N scaling func-
Many variations are possible on the scheme of Section 4. tions which now touch both edges simultaneously. Since we
One can, for instance, start from completely different fami- want them to also generate theN-dimensional space of poly-
lies of whole-line wavelets. If q, has support [- L + 1, L], nomials up to degree N - 1 they all have to be polynomials
and m0 (0 has a zero of order K;;.. 1 at~= 11', with K < L, themselves. How we choose these polynomials determines
then there is no need to sacrifice the outermost interior scal- the filter coefficients at this level; note that this choice is
ing functions 4J-J.L-t and q,_1 ,2 J_L_ 2 : even if we retain them, completely unrestricted.
we have 21 - 2L + 2 interior functions, leaving enough room We conclude this paper by pointing out an important dif-
to add K extra functions at each end, so that all polynomials ference between wavelets on the line and wavelets on [ 0, 1],
up to degree K - 1 can be generated. If K < L - 1, we even which results in the necessity, in at least some applications,
have room to spare. If it is important that we have exactly 2 1 of preconditioning the data (e.g., an image) prior to their
scaling functions at scale j (as is the case in many applica- wavelet decomposition.
tions), then we can add an extra L - K functions almost Let us return to the example of the N-vanishing moment
arbitrarily (without leaving the multiresolution framework, family with minimal support. On the whole line, we have

23
 not have integral 1. The same problem exists in Meyer's
I_~x dx x 1jJ(x) =
1
0, I= 0, ... , N- 1, construction; the following proposition shows that it is in-
evitable there.
as well as PROPOSITION 5.1. Let 4Jg?f", k = -N + 1, ... , N- 2,
be the adapted scaling functions constructed as in Section 3
L n'g. = 0, I= 0, ... , N- 1, (for, say, a Ieftedge), withN;;;,: 2. Then J:N- 2
4J~?t<(x) 1 *
n for some k. Moreover, if the ¢~:/dge are the image of the
cfJ ~~r under any ( 2N - 2) X ( 2N - 2) unitary matrix, then
or equivalently f2N-2 dxA..# d ( )
Jo o/o:'t ge x* 1 fior some I.
Proof
L n (-1)"h. =
1
0. (5.1)
1. In Meyer's construction, ¢8~~~+ 1 has support [0, 1],
and L 2 -norm 1. By Cauchy-Schwarz,
This implies that if we apply our high and low pass filtering
to a sequence which is just a linear combination of polyno-
mial sequences, i.e.,

N-1
c.= La 1 n 1, (5.2) with equality only if ¢8~~~+ 1 ( x) = 1 on [ 0, 1]. Since this is
1=0 not the case (cfJ8~~~+J(x) = 0 for x > 1, and the function is
J
continuous), dxcfJ8~~~+ 1 (x) * 1.
the high pass filter yields exactly zero, 2. A different orthonormalization of the ¢8~/" in Section
3 would lead to cfJ~:'tdge related to the¢~~!" by
n N-2
~ U A-edge
L, /ko/0,1< '
There is a nice parallelism between the orthogonality of the k=-N+i

ljl(x - n) to polynomials and the orthogonality of the high

pass filter masks to polynomial sequences. This parallelism where U is a ( 2N - 2) X ( 2N - 2) -dimensional unitary
can also be expressed otherwise: just as the scaling functions matrix. It follows that support ¢~:/dge C (0, 2N- 2] for all
¢(x - n) generate all polynomials of degree up toN - 1, I, and
the low pass filtering leaves invariant the N-dimensional
space of sequences of type (5.2). This is because for any
such sequence one can find a polynomial p(x) such that I A,#,edge
dx "¥ 0,/
= ~ U
k
LJ lk
I dx A.
o/ O,k
edge.

c. = f dx p(x)¢(x - n). Consequently L. g.-2lcc• =

f dx p(x)2- 112 1jJ(x!2 - k) = 0, and L. h.-2Jcc. = If f dx ¢Wge(x) = 1 for all I, then this implies that
f dx p(x)2- 112 ¢(x/2 - k), leading to another polynomial L,IJ dxcfJZ:'tdge(x)l 2 = 2N - 2, hence L* lk N-
2 2

sequence. In particular, the sequence cn = 1 can be repre- 2

dxcfJ~~f(x)l 2 = 2N - 2 = if:N- dxl 2 • Since the 2N- 2
sented as c. = f dx ¢(x - n), so that Ln g. -2JcCn =
f dx2-1!21jJ(x/2- k) = 0. functions cfJ~~t" are orthonormal, this is only possible if
A. edge k } .
Things are not that simple on the interval [ 0, 1]. It is still x1o,ZN-2J E Span { '¥OJ< ; = - N + 1, ... , N - 2 . But th1s
true that the sequence with 2 j entries given by is again impossible because the cfJ~~t" are continuous in
2N- 2. •
In the construction of Section 4 there are only N functions
with support [ 0, 2N - 1], so that the same obstruction does
not hold. Nevertheless, imposing that at least for one ortho-
k = 0, ... , N- 1, m = N, ... , 2j normalization procedure all the scaling functions have inte-
gral 1 leads to very stringent requirements on the h., which
- N - 1, I = 0, ... , N - 1,
are not satisfied in the "canonical" case of the N vanishing
(5.3) moment family with minimal support. Once one has obtained
one orthonormal family of edge scaling functions, it is easy
gets mapped to the zero sequence by the high pass filters to test whether a possibly different orthonormalization could
adapted to the interval, but the sequence (5.3) is no longer have led to orthonormal functions with integral 1: this will
the sequence consisting of only 1 's: the edge functions do J
be the case if and only if L~:d I dx ¢~J(x) 1 2 = N. This last

24
 1.0 - 1.5[
1\. .
:I~~:;\1
I
,----,/-.~-. ~-1 1
~
05~1
0.5 \ 0,1 O <+.right

0 \ ____.---..
-0.5
-1 .0 L_.___ _ _ _ _----J
-o.~ _.I______
-6 -4 -2
.
0 -6 -4 -2 0
0 2 4 6 0 2 4 6
1.5 15 2

10l
~-~--~ ··~

~·~t:;·\v----'I
<1> right
0,-1
. /\ 0.2
<!>left l'
0.5 I \

0 \ __/-- J o
-o. 5 Lo~-2--4--6- -0.5 ..... ~ -6 -4 -2 0 -6 -4 -2 0
0 2 4 6

2 - . -...·
right A right
'l'o.-4 I 111
'1'0.-3

0 0

-1 -1 -1
0
\) 0
-1

.: ~;·~·-lA:
0 2 4 6 0 2 4 6 -6 -4 -2 0 -6 -4 -2 0

0
-1
..... . I
- ~ ~.1 h; J\~fl~
.....

'1'0,-1 1 jl

-2
0 2 4 6 -6 -4 -2 0 -6 -4 -2 0
FIG. 7a. The four edge scaling functions and wavelets for the case N = FIG. 7b. Same as in Fig. 7a for a right edge (i.e., on ( -oo, 0]).
4, for a left edge (i.e., on [ 0, oo)). Together with the interior scaling func-
tions, the edge scaling functions generate all polynomials of degree 3 or less.

:;;. N which coincides with p ( x) on [ 0, 2N ~ 1] and is sup-

ported on, say, [0, 4N]. Since p(x) is orthogonal to the
condition is not verified in the explicit examples computed
in Section 4.
cp ~~*, this linear combination reduces to a linear combination
of the c!Jo.m, m :;;. N, which implies that it vanishes identically
In practical examples (e.g., images) one would still like
on [ 0, 1] . Since p =I= 0 on [ 0, 1], this is a contradiction.
simple polynomial sequences such as 1 1 1 1 · · · or 1 2 3 4
There exists therefore a (unique) nonsingular N x N-ma-
· · · to lead to a zero high-pass component, however. This
trix A with nonsingular inverse such that
can still be achieved if we perform a prefiltering on the data.
The principle is simple. We compute two families of N-di-
N-1
mensional vectors,
(WI)k = L A~cm(VI)m.
(V1)* = J dx x 1cp(x - k) k, I= 0, ... , N ~ 1,
m=O

Given a sequence of data, we then apply this matrix to the

(W1)k = Jooo dx x 1 cp~•J(x) k, I= 0, ... , N ~ 1. ( 5 .4) first N entries,

N-1
The N vectors V1 are trivially independent: if their entries are ck = L Akmcm.
taken as successive columns of a matrix, the resulting deter- m=O

minant is a nonvanishing Vandermonde determinant. TheN

vectors W 1 are independent as well. If they were not, then This maps the polynomial sequences which we would like
there would be a polynomial p ( x) of degree N - 1 orthogonal to map to zero in the high pass filtering to sequences which
to all the cp~r, k = 0, ... , N ~ 1. On the other hand, there are polynomial after the first N entries but with specially
exists a finite linear combination of the cp ~J and the c!Jo.m, m tailored beginnings. The space of these modified polynomial

25
sequences is invariant under the low pass filtering operation, may well wonder whether this prefiltering is necessary at all.
and maps to zero under the high pass filters. Mter this pre- Instead of determining the edge wavelets so that they are
filtering operation, the data can be taken through as many orthogonal to polynomials, one can determine the high and
high and low pass filtering stages as desired, for a decom- low pass filters near the edges so that the high pass filters
position into wavelets (or wavelet packets, with best basis give zero when applied to polynomial sequences of low de-
search). From this decomposed form the data can be recon- gree. This is the point of view adopted by Herley and Vetterli
structed by the conjugate filtering operations, followed, in [14]; it avoids the need of preconditioning. The same phe-
the last instance, by a new filtering operation affecting only nomenon is, however, still present, in a disguised form: even
theN first entries, now using the matrix A - I . though 1 1 1 1 1 · · · and 1 2 3 4 5 · · · map to zero under the
To determine A explicitly, we would have to compute all high pass filters, the space of polynomial sequences is not
the quantities in ( 5.4). This can be done,. using the recursion preserved under the low pass filters. At the next stage, their
relations as we did above for similar integrals. One can avoid low pass versions are transformed into adulterated polyno-
all explicit computations by choosing convenient families of mial sequences, and these do not map to zero under the high
vr' W1 vectors, different from but equivalent to the Vt. WI. pass filtering operation. Our prefiltering introduces an extra
The V1 are defined by step, but after this (non unitary) step, everything is stable in
the sense that an originally polynomial sequence leads to zero
content in all the band pass channels afterwards. If one works
with biorthogonal instead of orthonormal wavelets, then
there is so much more freedom in the choice of the edge
the correspondence V1 ++ V7 is as follows. We know that there functions that it is possible to construct biorthogonal schemes
is one-to-one correspondence between polynomials of degree in which the low pass filters automatically preserve polyno-
N- 1 and the polynomial coefficient sequences of their ex- mial sequences, so that no pre filtering is necessary.
pansions in the <P(x- n). Define the polynomials q 1(x) by
APPENDIX A: PROOF OF AN IDENTITY USED
1
q 1(x) =- }: n(n- 1)· · ·(n- I+ 1)<P(x- n) IN SECTION 2
/! nEZ

=~(~)<P(x-N+1+n), l=O, ... ,N-1. PROPOSITION A.l. Let 1/J, ljJ be dual functions generating
biorthogonal wavelet bases as in Cohen et al. [ 8]. If
J dxH(x)lP(x- m) = Oforallm E 7L, whereH(x) = 1-
J
Then (V1)k = dx qt(x)<P(x - N + 1 + k). It follows that lxl for lxl ,;;; 1, H(x) = 0 otherwise, then <P(x) =
J
we should choose (W1)k = dx qt(x)<Pb~~-1-k(x). Note that H ( x - L) for some L E 7L.

( 4.1) implies that Proof We use the notations of Cohen et al. [ 8] through-
out this proof.
1. Using the recursion ljJ(x) = .fi Lk (-l)kh_k+ 1
¢(2x - k), we find that Lk ( -l)kh_k+Ia2m-k = 0 for all m
which means that the (W1)k are nothing but entries of the
E 7L, where am = J dx H(x)4>(2x + m). With the notation
a( 0 = Lm am e -im(, this can be rewritten as
orthonormalization matrix giving the transition ¢ 1 -+ <P~~,
which we have computed before. It is then easy to obtain A
and its inverse. Note that the triangular structure of the arrays mo(~ + 1r)a(O- mo(Oa(~ + 1r) = 0.
V 11 , W 11 , i.e., (V1)k = 0 = (W1)k if I> k, implies that A and
A -I have a similar structure: A~m = 0 = (A -I )~:m if k > m. Since mo(O, mo(~ + 1r) do not vanish together, this implies
All this concerned the left edge only. On the interval, we that
have to do two prefilterings: one on the leftmost N samples,
and another one (with a different matrix) on the rightmost a(O = mo(O v(~); (A.1)
samples. For the N vanishing moment family with minimal
support, we have computed these matrices explicitly. For N because only finitely many ak are non-zero ( 4> has compact
= 2 and 4 the results, for both left and right edge, are listed support), both a and v are trigonometric polynomials.
in Table 5; for other values of N they can be obtained elec- 2. Because mo(Omo(O + mo(~ + 1r)m 0 (~ + 1r) = 1, we
tronically (see the note at the end of the paper). conclude from ( A.1) that
If one is only interested in the discrete aspect of these
filtering operations, as in image subband filtering, then one v(20 = n1o(0a(0 + liio(~ + 1r)a(~ + 1r),

26
 TABLE 5
The Preconditioning Matrices Corresponding to the Edge Filters of Tables 3 and 4

N=2

Alef1 =
( 0.32540489E + 00
0.37158015E - 01
O.OOOOOOOOE + 00)
0.10014454E + 01 '

( 0.30779265E + 01 O.OOOOOOOOE + 00)

AI~:,=
-O.ll420457E + 00 0.99855668E + 00

Anght =
( O.l0898431E + 01
O.OOOOOOOOE + 00
-0.8008l323E + 00)
0.20962929E + 01 '

Ar~t= ( 0.91756331E + 00
O.OOOOOOOOE + 00
0.35052203E + 00)
0.47703258E + 00

N= 4

0.24899ll1E + 01 O.OOOOOOOOE + 00 O.OOOOOOOOE + 00 O.OOOOOOOOE + 00 )

-0.27529885E + Ol 0.16772l06E + 01 O.OOOOOOOOE + 00 O.OOOOOOOOE + 00
Alert=
( 0.16878414E + 01 -0.70753754E + 00 0.1l301451E + 01 O.OOOOOOOOE + 00
-0.40222212E + 00 0.17635443E + 00 -0.61621216E- 01 0.10068852E + 01
( 0.4016208E + 00 O.OOOOOOOOE + 00 O.OOOOOOOOE + 00
O.OOOOOOOOE + 00)
0.65922394E + 00 0.59622806E + 00 O.OOOOOOOOE + 00 O.OOOOOOOOE + 00
AI:~=
-0.18709675E + 00 0.37327395£ + 00 0.88484210E + 00 O.OOOOOOOOE + 00
0.33523746E - 01 -0.81584146£ - 01 0.54152199E- 01 0.99316192£ + 00
( 0.10003981E + 01 -0.22411543E - 02 ~0.18445Q47E ~ 01~0.73733049E- 02)
O.OOOOOOOOE + 00 0.10023l30E + 01 0.91704628E - 01-0.93100685E - 03
Anght=
O.OOOOOOOOE + 00 O.OOOOOOOOE + 00 0.78081762E + 00 0.37673864E + 00
O.OOOOOOOOE + 00 O.OOOOOOOOE + 00 O.OOOOOOOOE + 00 O.OOOOOOOOE + 00

( 0.99960208E + 00 0.22350928E - 02 0.23350830E - 0J-0.28464600E - 02 )

O.OOOOOOOOE + 00 0.99769238£ + 00 -O.ll7l7590E + 00 0.90053579E- 01
A~~ht = O.OOOOOOOOE + 00 O.OOOOOOOOE + 00 0.12807088E + 01-0.96398392£ + 00
O.OOOOOOOOE + 00 O.OOOOOOOOE + 00 O.OOOOOOOOE + 00 0.19979252E + 01

Note. For each of the cases N = 2, N = 4, we list A and its inverse A_,, first for the left side, then for the right side. They should be applied as follows:
arrange all the data points in one column vector d, with entries d, · · · d, (L "'= 2N). Apply A,•• to the N-vector with entries d 1 • • • dN, A rig!!, to the N-vector
with entries dL-N+l' · ·d, (in that order), and use the results to replaced,·· ·dN and dL-N+I · · ·d,; the dN+t' · ·d,_N are not touched. The resulting sequence
is the preconditioned sequence. After decomposition + reconstruction by means of interior and edge filters, the same procedure should be followed (with
A~~ and A~,:.,) to recover the original data.

This implies that a( O = (cos 2 ~ /2) v( 0; hence

vk = ..fi ~ h1a21c-l = ..fi ~ h1 J dx H(x);J>(2x + 2k- /)

4. Because f dx xlfr(x) == 0, m0 (0 is divisible by
I I

= Jdx H(x);J>(x + k).

cos 202, so that

v(O = v(~)Q(O,

3. On the other hand, H ( x) = ~ H ( 2x + 1) + H ( 2x) + where Q is a trigonometric polynomial. Consequently if z0

~ H(2x - 1 ), which leads to = e'~" corresponds to a zero of v (where lzol may be* 1),
then so does z0 /2. Since there can only be finitely many such
zeros, this implies that v( 0 == Ce-iL{ for some L E Z..
5. It follows that a(O = C cos 2 (U2)e-iL~; substituting
this into ( A.1) gives

27
 3. M. Berger, Random affine iterated function systems: Curve generation
and wavelets, SIAM Rev. 34 ( 1992), 361-385.
4. A. Cohen and I. Daubechies, A stability criterion for biorthogonal wave-
let bases and their related subband coding scheme, Duke Math. J. 68
or mo(O = C cos (02)e-LL~. leading to C
2
= 1 and cf>(x) = ( 1992), 313-335.
H(x- L). • 5. A. Cohen, I. Daubechies, B. Jawerth, and P. Vial, Multiresolution anal-
ysis, wavelets and fast algorithms on an interval, C. R. Acad. Sci. Paris
NOTE ON ELECfRONIC AVAILABILITY OF MORE Ser. I Math. 316 (1992), 417-421.
COMPLETE TABLES 6. A. S. Cavaretta, W. Dahmen, and C. Micchelli, Stationary subdivision,
Mem. Amer. Math. Soc. 93 ( 1991 ), 1-186.
Complete tables for N = 2 to 8, for Meyer's construction 7. A. Cohen, Ondolettes, analyses multiresolutions et filtres miroir en
quadrature, Ann. Inst. H. Poincare Anal. Non Lineaire 7 ( 1990), 439-
as well as the new construction, are available by ftp; node:
459.
research.att.com; user name: netlib, password: your e-mail
8. A. Cohen, I. Daubechies, and J. C. Feauveau, Biorthogonal bases of
address; directory: stat/ data. The file is in binary and is called compactly supported wavelets, Comm. Pure Appl. Math. 45 (1992),
wavelets.Z. You should uncompress it after acquiring it. You 485-560.
can also get the file by sending the message 9. I. Daubechies, Orthonormal bases of compactly supported wavelets,
Comm. Pure Appl. Math. 41 ( 1988), 909-996.
send /stat/ data/wavelets 10. I. Daubechies, "Ten Lectures on Wavelets," CBMS Lecture Notes, No.
61, SIAM, Philadelphia, 1992.
to [email protected]. The part of the tables concerning 11. I. Daubechies, Orthonormal bases of compactly supported wavelets. II.
only the new construction is also available from an ftp node Variations on a theme, SIAM J. Math. Anal. 24 ( 1993), 499-519.
in Europe; node: ftp.ensta.fr, user name: anonymous, pass- 12. I. Daubechies and J. Lagarias, Two-scale difference equations. I. Exis-
tence and global regularity of solutions, SIAM J. Math. Anal. 22 ( 1991),
word: your e-mail address, directory: pub, filename: tables;
1388-1410; Two-scale difference equations. II. Local regularity, infinite
this file is in ASCII. The tables in these files, like those in products of matrices and fractals, SIAM J. Math. Anal. 23 ( 1992), 1031-
this paper, are accurate to 10 -s. For most purposes this is 1079.
amply sufficient; more accurate tables (to 10- 14 ) have been 13. G. Fix and G. Strang, Fourier analysis of the finite element method in
computed by Mary Brewster and Greg Beylkin. Ritz-Galerkin theory, Stud. Appl. Math. 48 ( 1969), 265-273.
14. C. Herley and M. Vetterli, Time-varying orthogonal filterbanks and
ACKNOWLEDGMENTS wavelets, submitted for publication a shorter discussion is in C. Her ley,
J. Kovacevic, K. Ramchandran, and M. Vetterli, Time-varying ortho-
A. Cohen and P. Vial thank the Mathematics Center of AT&T Bell Lab- normal tilings of the time-frequency plane,/£££ Trans. Signal Process.,
oratories, where this work was carried out, for hospitality and support. We to appear.
also thank the referees for reading the manuscript very carefully and finding 15. P. G. Lemarie and G. Malgouyres, Support des fonctions de base dans
many small mistakes. We are grateful to Minh Do Kac for pointing out a une analyse multi resolution, C. R. A cad. Sci. Paris 313 ( 1991 ), 377-
wrong figure and to Mary Brewster and Greg Beylkin for drawing our at- 380.
tention to wrong entries in the tables. 16. A. Jouini and P. G. Lemarie-Rieusset, Analyses multiresolutions biorth-
Note Added in Proof It has come to our attention that there exists yet ogonales et applications, Ann. Inst. H. Poincare Anal. Non Linea ire, to
another approach to the construction of wavelet bases on the interval, with- appear.
out wrap-around effects, with 2' wavelets and 2' scaling functions at scale
j, in M. H. Freedman and W. H. Press, "Truncation of Wavelet Matrices:
17. W. Lawton, Tight frames of compactly supported wavelets, J. Math.
Edge Effects and the Reduction of Topological Control,'' University of Cal- Phys. 31 ( 1990), 1898-1901.
ifornia, San Diego, preprint (December 1992). 18. W. Lawton, Necessary and sufficient conditions for constructing ortho-
normal wavelet bases, J. Math. Phys. 32 ( 1991), 57--61.
REFERENCES 19. S. Mallat, Multiresolution approximation and wavelets, Trans. Amer.
Math. Soc. 315 ( 1989), 69-88.
1. L. Andersson, N. Hall, B. Jawerth, and G. Peters, Wavelets on closed 20. Y. Meyer, "Ondelettes et operateurs. I. Ondelettes. II. Operateurs de
subsets of the real line, in ''Topics in the Theory and Applications of Calder6n-Zygmund. III. Operateurs multilineaires," Hermann, Paris,
Wavelets" (Larry L. Schumaker and Glenn Webb, Eds.), Academic 1990, English translation, Cambridge Univ. Press, London/New York,
Press, Boston, to appear. 1993.
2. P. Auscher, Ondelettes a support compact et conditions aux limites, to 21. Y. Meyer, Ondelettes due l'intervalle, Rev. Mat. Iberoamericana 7
appear in J. Funct. Anal. Ill ( 1993), 29-43. (1992), 115-133.

28