2464 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40. NO.

IO, OCTOBER 1992

The Discrete
. Wavelet Transform: Wedding the
A Trous and Mallat Algorithms
Mark J. Shensa

Abstract-In a general sense this paper represents an effort cessing. The other, the multiresolution approach of Mal-
to clarify the relationship of discrete and continuous wavelet lat and Meyer, originally used in image processing, em-
transforms. More narrowly, it focuses on bringing together two
separately motivated implementations of the wavelet trans- ploys orthonormal wavelets [6]-[ 101. The latter
form, the algorithme U trous and Mallat’s multiresolution de- algorithm, apart from its wavelet interpretation, was dis-
composition. It is observed that these algorithms are both spe- covered previously in the form of quadrature mirror filter
cial cases of a single filter bank structure, the discrete wavelet (QMF) filter banks with perfect reconstruction where it
transform, the behavior of which is governed by one’s choice finds application in speech transmission and split-band
of filters. In fact, the h trow algorithm, originally devised as a
computationally efficient implementation, is more properly coding [ 111-[ 131.
viewed as a nonorthonormal multiresolution algorithm for A glance at these two algorithms suffices to reveal
which the discrete wavelet transform is exact. Moreover, it is closely related structures. In fact, apart from the con-
shown that the commonly used Lagrange i~ trous filters are in straints on their filters, the decimated B trous [5] and Mal-
one-to-one correspondence with the convolutional squares of lat algorithms are identical. We are thus led to examine
the Daubechies filters for orthonormal wavelets of compact
support. the expanded family of algorithms encompassing both
A systematic framework for the discrete wavelet transform types of filters. In this vein, it is shown that the Lagrange
is provided, and conditions are derived under which it com- interpolation filters commonly employed by the B trous
putes the continuous wavelet transform exactly. Suitable filter algorithm are actually the squares (in a convolutional
constraints for finite energy and boundedness of the discrete sense) of the Daubechies filters for compact orthonormal
transform are also derived. Finally, relevant signal processing
parameters are examined, and it is remarked that orthonor- wavelets. We also derive conditions under which the dis-
mality is balanced by restrictions on resolution. crete implementation computes a continuous wavelet
transform exactly and find that they bear an intimate re-
lationship to the B trous constraints.
I. INTRODUCTION From a more general viewpoint, the situation is as fol-

W AVELETS are rapidly finding application as a tool

for the analysis of nonstationary signals [ 11-[5].
However, with the notable exception of orthonormal
lows: The algorithms to be discussed all are filter bank
structures (see Fig. 1). Their only distinguishing feature
is the choice of two finite length filters, a low-pass filter
wavelets [6]-[9], very little literature has been devoted to f and a bandpass filter g. The low-pass condition, ex-
linking discrete implementations to the continuous trans- pressed more precisely as C f k = J2, is necessary to the
form. As in the case of the discrete Fourier transform, construction of a corresponding continuous wavelet func-
these implementations (or filter banks) can, and should, tion. The bandpass requirement, while apparently not es-
stand on their own as abstract decompositions of discrete sential to all applications, ensures that finite energy sig-
time series: Their wide sweeping significance, however, nals lead to finite energy transforms (see Section VI).
lies in their interpretation as wavelet transforms. In a gen- Under these conditions, the filter bank output will be re-
eral sense, this paper represents an effort to clarify the ferred to as the discrete wavelet transform (DWT), a ter-
relationship of discrete and continuous wavelet trans- minology which will become clear in the course of the
forms. More narrowly, it focuses on bringing together two paper.
separately motivated implementations of the wavelet One class of DWT filter pairs are the Daubechies filters
transform. One of them, the algorithme U trous’ for non- [8] which yield orthogonal wavelet decompositions and
orthogonal wavelets [4],[5], was developed for music constitute, in more conventional terms, a QMF filter bank
synthesis [2] and is particularly suitable for signal pro- with perfect reconstruction. Another is that for which the
low-pass filter satisfies the i trous condition hk =
6 ( k ) / d 2 . Such filters, which simply serve to interpolate
Manuscript received May 21, 1990; revised August 5 , 1991. This work every other point, correspond to a nonorthogonal wavelet
was supported in part by the ONR and NOSC IR program.
The author is with the Naval Ocean Systems Center, San Diego, CA decomposition. As mentioned above, if they are further
92152-5000. restricted to be Lagrangian interpolators, they become the
IEEE Log Number 9202367. squares of the Daubechies filters, which is quite remark-
I .
Literally, “algorithm with holes,” this terminology, taken from [ 5 ] ,
refers to the fact that all the even coefficients of the relevant filter (with the able in consideration of the totally different derivations.
exception of the center) are zero. This also implies that a maximally flat filter with the same

U.S. government work not protected by U.S. copyright

I
SHENSA: DISCRETE WAVELET TRANSFORM 2465

of power/hertz.2 Certain weak "admissibility" condi-

tions are usually required on $ ( t ) for it to be a candidate
for an analyzing wavelet; namely, square integrability and

Fig. 1. A wavelet filter bank structure. The down-arrow indicates deci-

mation. The output of the transform is the family of signals w ' , forming
where $(a)is the Fourier transform of $ ( t ) . They ensure
the two parameter transform w i in the scale-time plane. Following termi- that the transformation is a bounded invertible operator in
nology to be introduced, w ' is the (decimated) discrete wavelet transform. the appropriate spaces [8], [16]. If $(U) is differentiable,
then it suffices that $ be zero mean, i.e., that S $(t) dt
= 0, for (1.2) to be satisfied.
number of vanishing derivatives at 0 and 7r is a Lagran- In the realm of signal processing, the significance of
gian interpolator; a fact which may aid in the design of (1.1) is probably (or, at least, traditionally) best grasped
maximally flat filters [ 141. by comparing it to the short-time Fourier transform:
A fundamental question is, when do these discrete im-
plementations yield exact (i.e., sampled) versions of a
continuous wavelet transform? Aside from regularity con-
ditions relating to smoothness [8], we find that i f f is a
' F(w, b) = S h(t - b)e'"'s(t) dt. (1.3)

trous, then the DWT coincides with a continuous wavelet Thus, to obtain F(w, b), one multiplies the signal by an
transform by a wavelet $ ( t ) whose samples $(n) form the appropriate window h (such as a Gaussian) centered at
filter g (i.e., g, = $ ( n ) ) . Even i f f is not B trous, the time b and then takes the Fourier transform. In mathe-
algorithm is exact provided the signal lies in an appropri- matical terms, (1.3) is an expansion of the signal in terms
ate subspace; however, in that insthnce, the sampled of a family of functions h(t - b) e'"', which are generated
wavelet values depend on f as well as g. This is the situ- from a single function h(t) through translations b in time
ation in the orthonormal case where, moreover, the filter and translations w in frequency. In contrast, the wavelet
g is almost completely determined from fthrough the con- transform (1. I) is an expansion in functions $((t - b ) / a )
straints of orthogonality. generated by translations b in time and dilations a in time.3
In the remainder of this introduction we define, and Thus, the continuous wavelet transform resembles a (con-
briefly motivate, wavelet transforms at various levels of tinuous) bank of short-time Fourier transforms with a dif-
discretization. Section I1 contains an abbreviated deriva- ferent window for each frequency. The significance of this
tion of the h trous algorithm followed by a description of is that, while the basis functions in (1.3) haveAthe same
the Mallat algorithm. (The uninitiated reader is particu- time and frequency resolution (that of h(t) and h (U)) at all
larly referred to [ 11, [6], [8] .) In Section I11 we define the points of the transform plane, those of (1.1) have time
undecimated DWT, relate it to the decimated transform, resolution (that of $ ( t / a ) ) which decreases with a and
and provide algorithms for its computation. Section IV frequency resolution (that of $ (aw)) which increases with
states and proves several theorems which delineate the re- a . This property can be a great advantage in signal pro-
lationship between the DWT and the continuous wavelet cessing since high frequency signal characteristics are
transform. It may be read independently of the algorithms generally highly localized in time whereas slowly varying
of Section I1 although the motivation for the constructions signals require good low frequency resolution.
may not be clear. Section V defines the Lagrange a trous As originally proposed by Morlet et al. [17], $ was a
filters and proves that they are the squares of the Daube- modulated Gaussian
chies filters. In Section VI, we formulate the inversion r ~ / ( ~ )= ejVote-t2/2 (1.4)
problem and provide filter constraints which ensure finite
energy and bounded operators. It concludes with a short and this function is still the prototypical analyzing wave-
examination of the tradeoffs involved in choosing the let for signal processing applications [I]. The window
bandpass filter, emphasizing the differences of the orthm- function $ ( t / a ) has Fourier transform $(aw) =
ormal and nonorthogonal cases. ae -(" -[vo/a112a 12,which has analysis frequency v o l a . We
emphasize that vo is simply a parameter which determines
A . Transform Dejinitions the analyzing wavelet; its role should not be confused with
that of a even though the scale axis is often expressed in
The continuous wavelet transform of a signal s ( t ) takes terms of frequency under the transformation a -, vo/u.
the form Observe that (1.4) only satisfies the admissibility condl-

W(a, b) = J a $ (G) s(t) dt (1.1)

tion (1.2) approximately (cf. [ 161, [ 181). As expected, its

*The energy density is given by ( W I Z[(da d b ) / a 2 ] ,an expression which

where $ is the analyzing wavelet, a represents a time di- is intimately linked to representations of the affine group (see [3], [Is],
lation, b a time translation, and the bar stands for complex [16]). Since a is proportional to bandwidth, 1 WIZis power/hertz. See Sec-
tion VI for a discussion of the discrete case.
conjugate. The normalization factor 1 / d a is perhaps most 3Alternatively, dilations in time may be considered contractions in fre-
effectively visualized as endowing IW(a, b)I2 with units quency since the Fourier transform of $ ( ? / a )is a$(aw).
2466 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, NO. IO. OCTOBER 1992

bandwidth is proportional to l / a , thus giving rise to a pictured in Fig. 1,

constant relative bandwidth; i.e., BW / ( v o l a ) = con-
stant. This feature is also reflected in the narrowing of the
time window at higher frequencies; i.e., at smaller a . In
general, one's choice of the function $ ( t ) is dictated by
[W'I, = C g,-][S'l' (1.7)
its time and frequency localization properties [3], [ 181.
We shall be exclusively concerned with discrete values where [w'],corresponds to w(2', 2'n) of (1.6) and so is
for a and b. In particular, we assume that a has the form the original signal s. The mysterious appearance of the
a = 2' where i is termed the octave of the transform. The filter f in (1.7) will be unraveled in the derivation of the
integral (1.1) then yields a wavelet series i trow algorithm in Section 11. Finally, we shall come full

W(2', n) 2! -
1
J2' s t - n
IC/ (7 )
s(t) dt. (1.5a)
circle in Section IV where we show, under quite general
conditions, that given filtersf and g there exists a function
$(t) with $(n) = gf k E-, such that the DWT acting on
the sampled signal is exactly the sampled output of the
We remark that finite energy for the wavelet transform is
continuous wavelet transform (i.e., of the wavelet se-
not at all equivalent to finite energy for the wavelet series. . ~ other words, the DWT with filter g defined by
r i e ~ ) In
It depends on the sampling grid as well as the function
gi & which was originally conceived as an ap-
$ ( t ) [3]. Thus, the admissibility condition (1.2) is not
proximation of the (continuous) WT for an arbitrary ana-
necessarily appropriate in the discrete case and shall be
lyzing wavelet $ A ( t ) , is exact for another wavelet func-
replaced with conditions on the relevant filters in Section
tion $ B ( t ) where $B(n) = gf for all n. Of course, if there
VI. In addition, we shall often take b to be a multiple of
is sufficient regularity, $ A ( t )and $B (t)will be close since
a4

s-(:
they coincide on the integers up to the length of g.
Before embarking on this voyage, we summarize, and
W(2', 2'n) 7 $ 7-n
42' ) s(t) dt. (1.5b) hopefully clarify, the plethora of transforms with a brief
analogy to the Fourier transform, Fourier series, the dis-
A logical step in applying the theory to discrete signals cretized z transform, and the discrete Fourier transform
is to discretize the integral in (1.5) (DFT). The Fourier transform of a continuous signal s(t)

S(W) k 1W

-w
e-j"'s(t) dt

The sample rate has been set equal to one. As indicated is a function of the continuous variable W . Restricting it
by 2% on the left-hand side, (1.6), as well as (1.5b), are to a discrete (one-dimensional) grid results in the coeffi-
decimated wavelet transforms. Octave i is only output cients of a Fourier series

s:w
every 2' samples. In this form the resulting algorithms will
not be translation invariant ([7]). This is easily seen by ~ ( 2 = ~ ~ e-j'*m') s(t) dt (1.9)
substituting s(k - r) for s(k) which produces w(2', 2'(n
- r / 2 ' ) ) , an integer translation of w(2', 2'n) only if r is
which in turn may be computed approximately by
a multiple of 2'. However, the invariance, which is lost
by decimation, is easily restored by separately filtering sZ(2.lrmA t ) =
k
ce-j21rmkArs(k At) At (1.10)
the even and odd sequences (see Section 111) or by using
an equivalent algorithm, also described in Section 111. Our the z transform of s, k s(n A t ) output at discrete points
major reason for starting with (1.5b) rather than (1.5a) is e -12lrmAI . If s(t) is band limited and sampled at an appro-
historical. It delineates the relationship of the DWT to the priate rate, A t = 1 / N , then the above may be computed
QMF filter banks and (orthonormal) wavelet structures al- exactly using the DFT
ready found in the literature [6]-[9]. It also simplifies our
derivation of the i trous algorithm and readily lends itself jm - c
l N exp (-N)
j2.lrmk sk.
(1.11)
to physical interpretation (see Section VI). Note that the N I
original Mallat algorithm [6] was decimated; i trow [4]
These correspond precisely to "(a, b ) , W(2', n ) , w(2', n ) ,
was not.
and undecimated w i . With wavelets, however, we have
Let g be the discrete filter obtained by truncating the
the additional difficulty of dealing with a whole class of
sampled wavelet function; i.e., g, = $(n). Then, pro-
functions $(I) rather than simply , j u t . Also complicating
ceeding from (1.6), we shall be able to amve at the DWT
things are its two-dimensional structure and the decimated
versions, which, due to their 2'n dependency on i, play a
4Physically, this reflects a need for less frequency sampling of the trans- distinguished role without analogy in the one-dimensional
form output at lower frequencies ( i . e . , larger scales U ) . Mathematically, b
= 2'n has its roots in the orthonormal wavelets where it suffices for in- case.
vertibility of the tranform [ 6 ] . The general case, however, is much more
complex [3]. Too sparse a sampling leads to incompleteness; oversampling 'The adjoint filter g l =-z_k is used to simplify future notation. It cor-
results in a redundant set of functions. responds to the integrand J . ( - t ) * s found in ( l . l ) , (1.5), and (1.6).
SHENSA: DISCRETE WAVELET TRANSFORM 2461

11. Two ALGORITHMS Some comment concerning filter definitions is also ap-
A. Notation propriate. Usage in the literature is uniform only up to the
adjoint. Also, the z transform is sometimes defined with
Decimation, which appears as a down arrow in Fig. 1,
a positive exponential which leads to similar differences
plays a pivotal role in all DWT algorithms. However, it
in the frequency domain. In keeping with signal process-
leads to operators which are not time invariant and present
ing applications we have chosen (2.5) as above, consist-
a potential source of confusion. It is thus worthwhile to
ent with the Fourier transform, and we shall define our
first establish some formal notation.
filters so that adjoints do not appear in convolutions. This
Signals and filters in boldface type will be treated as
produces a minimum of adjoints and greatly simplifies the
vectors, in which case * indicates discrete convolution and
notation. Unfortunately, it also results in the definition
yields a vector. The symbol f will be used for the Her-
g: = $(n) and the introduction off as an interpolation
mitian adjoint filter [ f ‘ ] k = f - k . Note that this is the com-
filter whereas g andf would be more natural. Note, also,
plex conjugate reversal and does not imply a conversion
that our filters are the adjoints of the filters defined in [8],
of a column vector to a row vector. The above mentioned
although their z transforms coincide since [8] defines the
decimation operator is represented by a matrix
z transform with a plus sign.
Akm 6(2k - m)
B. The A Trow Algorithm
We take the discretized wavelet series (1.6) as our
where Bkm is the Kronecker delta and 6(k) 6 ~Also
. of starting point. The difficulty in implementing (1.6) is that,
even for + ( t ) of finite support, as i increases, $ ( t ) must
significance is the dilation operator
be sampled at progessively more points, creating a large
computational burden. The solution posed by [4]is to ap-
proximate the values at nonintegral points through inter-
polation via a finite filter f t . The resulting recursion is
highly efficient and may be implemented with the filter
which dilates a vector by inserting zeros. Observe that A
bank structure of Fig. 1.
and 0 are transposes of each other, and that although they The interpolation is perhaps best introduced with an ex-
are linear, they are not time invariant; i.e., they are not
ample. Let f t be the filter (0.5, 1.O, 0.5). Then,
functions of k - m .
Convolution followed by decimation becomes [A( f *
$Ilk = C m & m [ f * SI, = [ f *SI2k = C m f 2 k - m S m . How- f - 2k$(k)
ever, a particularly insidious pitfall remains; namely,

(Af) *s # A ( f * s). (2.3)

This “associativity” problem may be avoided by replac-
ing convolution by f with a matrix F defined by

which shall occasionally be used in our proofs. A trivial approximates a sampling of $(t/2). With the help of the
calculation yields AFs = A(f * s). The symbol f will dilation operator 0, this may formalized as a general pro-
also be used for the adjoint of matrices. This is consistent cedure for dyadic interpolation. The steps are illustrated
with the above notation where [F’],,,
t
F,, = f n - , e in Fig. 2 . Let g be a filter defined by g: +(n); i.e.,
fm-n.
We define the Fourier transform s”(w) of a function s ( t )
by (1.8) and the z transform (on the unit circle) of a dis-
crete signal s by First we spread g’ to provide space in which to put the
interpolated values. The resulting filter is Dg’.Then we
sz(w) k C sne-j*”. (2.5) apply a filter f which leaves the even points fixed and
n
interpolates to get the odd points. This condition, that f
In the subsequent interplay between continuous and dis- be the identity on even points, is sufficiently important to
crete functions one must be careful to distinguish the usage warrant a separate defintion, which follows.
of these two transforms, Ignoring their differences can Definition 2.1: The low-pass filter f is said to be an a
easily lead to erroneous conclusions. In particular, al- trous filter if it satisfies
though the Fourier transform of s(2t) is s^(w/2)/2,

The J 2 is simply a convenient means of including the

normalization factor of (1.6) in the filter. The result of the
2468 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, NO. 10, OCTOBER 1992

some sense. This, in turn depends principally on choice

Dgt :
J
+(A)0
I
+io)
\\
0+U) 0 +(a
of the filter f. A major step towards treating this question
lies in the results of Section IV, as was outlined at the end
of the introduction. Since the algorithm is exact for some
pfV2 &(t), the question reduces to a) the quality of the ap-
dW): +(-I) x +(O) x +U) x +(2)
proximation $ = $B and b) the effect of this approxima-
tion on the wavelet integral (1.1). Inasmuch as $(n) =
Fig. 2. Diagram illustrating the dilation and interpolation of a function $B(n) for the finite set of integers n used to obtain g from
$ ( t ) : $ ( n / 2 ) = d2 f * (Dg+).
+

$(t), it is sufficient that these functions be smooth enough

and decrease fast enough at infinity. Conditions on f which
entire interpolation operation, as pictured in Fig. 2, is thus achieve regularity of the constructed wavelet $B are found
in [SI. Quantifying these statements remains a subject for
if' * ( o g t ) l n = [ ~ ~ ~ g ' l n future study.
= .f; - 2k $(k) C. Multiresolution Algorithm
1 Mallat's algorithm, illustrated in Fig. 3, has essentially
= J2 $(n/2). (2.10) the same tree structure as (2.12) (cf. [6]-[9]). Namely,
s i + l - A(h * s ' )
(2.13a)
Noting that $((k/2) - n) = $((k - 2n)/2) and insert-
ing the approximation (2.10) into (1.6), we obtain d"' = A(g * s'). (2.13b)

w(2, 2n) ;z x f l - 2 n - Z m g m- St k
In keeping with the literature, we have replaced the filter
k ,m f with the filter h , which also serves to indicate that this
class of filters is constrained, as detailed below. We re-
= gn-m'zm'-kSk
k,m' mark that none of these filters are a trous filters. The con-
straints on h and g which ensure an orthonormal multi-
= [g * (A(? * s>>l, (2. 1)
resolution analysis [6]-[9] are
which is simply wi (1.7) with i = 1. Continuing induc- -
tively by replacing s in (2.11) with s i - ' , we find w(2', Ci [ & 2 j - n h z j - m + g 2 j - n g 2 j - m ~ = 6nm (2.14a)
2%) = wi for all ii where wh is given by (1.7), which can
(2.14b)
be rewritten for real f
s i + ' = A(f * si) (2.12a) C g, =0 (2.14~)
n
wi =g * si. (2.12b) h, = J 2 . (2.14d)
n
Except for decimation of the output (the undecimated ver-
sion will be derived in Section 111), this is the a trous al-
Recalling that Hm, k hm- ,and that A ' = D, we may
gorithm described in [4]. Thus, we see that the a trous rewrite (2.14a) and (2.14b) as
algorithm is simply a DWT for which the filter f (an in- (H'D)(AH) + (G'D)(AG) = Z (2.15)
terpolator) satisfies condition (2.9) and the filter g is ob-
tained by sampling an a priori wavelet function $(t). (AH)(G'D) = 0. (2.16)
Remark 1: The definition (1.6) is not so transparent as Furthermore, (2.15) and (2.16) imply (e.g., multiply
it might seem. It is, of course, intended to reflect an ap-
proximation to (1.5). From this viewpoint one might well
(2.15) on the left by A H ) '
consider a change of variables t --* t/2' before discretiz- (AH)(H'D) = z
ing (1.5). Such a procedure certainly alleviates the com- ) z.
( A G ) ( G ~ D= (2.17)
putational problem since it dilates s ( t ) (that is, samples s
at 2', values which are known) rather than contracting Thus, H'D and G t D are injections and (2.13). is an or-
$(t). However, unless the original function s(t) was highly thogonal decomposition of the discrete signal s'. That is,
oversampled (which begs the computational question), the si-' = H'Ds' + G t D w i with the scalarproduct ( H ' D s ' )
* (G'Dw') = 0. In fact, (2.15) is a paradigm for invert-
approximation is poor. More precisely, to accurately ap-
proximate s(t), and therefore also (1.5), we must sample ing the transform. These concepts are illustrated in Fig.
at least at the nyquist rate rnypfor s. Then the integral for 4.
octave i requires $(t) to be sampled at a rate 2' rnyq. Furthermore, from (2.14) it follows that (2.13) repre-
Remark 2: The derivation above, as well as that in [4], a decomPosition [6I9 [8I, [91 as follows:
of the a trous algorithm make no statements regarding the There exists a function ($(t) whose Fourier trans-
accuracy of the approximation (2.11) or even of the dis- form is given by
m
cretization from (1.5) to (1.6). The former is iterated over
(2.18)
i and, hence, to succeed must be numerically stable in
SHENSA: DISCRETE WAVELET TRANSFORM 2469

s s
scaling function form a basis for L 2 ( R ) ,and

d; = J2' s(t)$ ($ - n) = s(t)$;(t) (2.25)

provided

Fig. 3. The Mallat multiresolution algorithm. The down-amow indicates

decimation.
s; =
S s(t)$(t - k).

Then, if s(t) is in L2(R),the expansion of s(t) is

(2.26)

s(r) = , c c d;$;(t).
I=-w n
(2.27)

This follows from, (2.25) and orthonormality since com-

pleteness of the q5L(t) implies completeness of the $ i ( t ) .
Until recently, the only known orthonormal wavelets
with compact support (i.e., zero outside a finite interval)
were the Haar functions, generated by $(t) = 1 for 0 I
Fig. 4. Illustration of a single stage and its inverse in the multiresolution t < 1/2; - 1 for - 1 / 2 It < 0; and 0 elsewhere. Dau-
algorithm for orthonormal wavelets. bechies [8], [19] has uncovered an entire family of finite
length filters satisfying (2.14), demonstrating that the cor-
responding wavelets defined by (2.18), (2.23), and (2.24)
Expression (2.18) forms the basis of a recursion are orthonormal as a consequence of (2.15), (2.16) and
have compact support since they were generated by finite
length filters (see Section IV, eq. (4.10)). The first two
of these filters are
which in the time domain takes the form
1
4(t) = h-kJ2 4(2t - k). (2.20) h' = - (1, 1) (2.28a)
J2
k

With some additional computation (see Section IV), one and

may demonstrate that the translates and dilates of 4,
h' = 4 1~ (12 + J 3 , 3 + J 3 , 3 - J 3 , 1 - J3)

(2.28b)
have the property where the first component of h is on the left. The wavelets
corresponding to (2.28a) are exactly the Haar function
4 c 1( t ) = [AHlnk+:(t). (2.22)
mentioned above.
Some additional remarks relating the two algorithms are
Note that the above definitions differ in the sign of i from in order. The conditions ( 2 . 1 4 ~ )and (2.14d) effectively
those of [8]. make g a bandpass filter and h a low-pass filter (e.g., an
Finally, define interpolation filter) with the sum on g k analogous to the
$(t) A d2 g-k4(2t - k). (2.23) condition $(t) dt = 0. Also, d'" corresponds to w l .
The additional decimation appearing in (2.13b) would not
Then, using (2.14) and the above properties of 4, one can appear in a translation invariant version of Mallat's al-
show that the family of wavelets, gorithm (cf. Section 111). On the other hand, although the
discrete filters g play algorithmically identical roles, the
$ i ( t ) of the ?
trous
i algorithm are not the wavelet vectors
(2.24) of a functional expansion. Rather the $ k ( r ) are the duals
of a set of vectors for which the coefficients of the signal
are orthonormal ( 5 $L(t) $jk(t) dt = 6, &), and that the d' expansion are w:. That is, they are the coefficients of an
are the coefficients of the expansion of s ( t ) in terms of the expansion of the form s(t) = C l , ( s,-$; ) $L(t) where ( )
*:.6 indicates the L 2 inner product, and $ k ( t ) is the dual basis
More precisely, the translates and dilates +l(t) of the or frame (see [3], [15]) of 11.;. In Mallat's algorithm, since
the $k(t) are orthonormal, the basis and its dual coincide.
61n contrast, the wavelet functions $:(r) (1 / J 2 ' ) $((2/2') - n) of (1.5) Thus, in many senses, the discrete filters g are more fun-
are not generally orthogonal. It is the filter constraints (2.14a) and (2.14b) damental than the wavelets themselves. It is usually the
that ensure orthonormality. Dropping these two constraints in Section IV,
we develop a structure identical to (2.21)-(2.24); however, the constructed coefficients which are of major interest; the actual wave-
$:(r) need not be orthogonal. lets $k(t), let alone their duals, are rarely computed.
2470 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40. NO. IO. OCTOBER 1992

Finally, in anticipation of Section V, let us form the

squares of the two previous examples
1 1
hf *h = - (1, 1) * - (1, 1)
J2 J2
= (1/2, 1, 1/2) '
(2.29a)
i$ $+l
and
ht*h=(-'o2 1 2 0 2) (2.29b) Fig. 5 . The (undecimated) discrete wavelet transform. The filters D ' f are
16' 7 16' 16' 16 obtained from f by inserting 2' - 1 zeros between each pair of filter coef-
ficients. The operation of filtering is understood to mean convolution.
where * indicates convolution, and h given by (2.28a) and
(2.28b), respectively. Note that the filtersf = h * h / J 2
are 2i trous filters and that the interpolation equation (2.10) Note that the zeroth element of a series is invariant un-
holds exactly if $ ( t ) is a polynomial in t of degree one or der decimation so that w f and v3f should coincide at n =
three, respectively. 0. Utilizing this fact, we obtain the nth output of the un-
decimated discrete wavelet transform by translating the
111. WAVELET TRANSFORMS WITHOUTDECIMATION signal back by n samples and taking the decimated trans-
As has often been pointed out [7], the recursions (2.12) form at time zero. More precisely,
and (2.13) are not, in general, translation invariant. In Dejnitiun 3. I: Define the undecimated discrete wavelet
contrast, the original undecimated 2i trous algorithm [4] , transform G in terms of the decimated transform w by
[5],which is pictured in Fig. 5 and consists of entirely of
IQ; [w'(s@)],P [w'(T-nso)]o. (3.7)
convolutions, is patently translation invariant. In this sub-
section we shall use that property to provide a formal def- We see that the desired invariance is achieved
inition for an undecimated discrete wavelet transform w ,
then demonstrate that w ; = K&, and also show that @ is [ w ' (T, = w ' (T-, Tms O)IO
computed by the algorithm of Fig. 5. = [w'( T m - n so)l0
Let Tmbe the operation of translation by m; i.e.,
= [w'(sO)],-,. (3.8)
(Tms)n 2 sn-m* (3.1)
It is also clear from (3.6) and (3.7) that sampling vV; every
In order to include the dependency of w' on so, we add it 2' points produces exactly wf, that is,
as an argument, writing w ' ( s o ) . Equations (2.12a) and
(2.12b) become w:,= W i t n . (3.9)
w i ( s 0 )= G(AF)'sO. (3.2) Next, we show that @ may be computed by the filter
sequence pictured in Fig. 5. The proof is obtained by tak-
(Recall that Gmn& g, - ,.) Finally, we shall need two im- ing z transforms. From (3.2) and (3.7)
portant identities, which are proved in Appendix A.
Namely, for any F of the form Fmn= fm - n , we have the c fiie-'"" = c c [G(AF)']Om~,+ne-J"W
n m n
following lemmas.
Lemma 3.I: = [G(AF)']omeJm"s,(w). (3.10)
m
[(AF)'Ink = [ ( A W I O , ~ - * ~ ~ (3.3)
Applying Lemma 3.2 to (3.10), we have
Lemma 3.2:
1-1 wi):(w) A C *;e-'nw
k
[(AF)'],e'kw = eJ2ln0 IT
r=O
fZ(2'w). (3.4)
n

1 - 1

As expected, w' is not translation invariant, = gZ(2'w) f,(2"w)sz(w) (3.11)

r=O

[ w 1 ( ~ m s 4 ] n= c[~(~~ln,k~;-rn where i = 0 is understood to mean there are no factors of

f,.As described in the next paragraph, this is exactly the
= [G(AF)'I,,~+~~; z transform of the algorithm pictured in Fig. 5.
It is eady to see that D ' f is f with 2' - 1 zeros inserted
# [G(AF)'],,-,,,p;. (3.5) between every pair of filter coefficients and that its z trans-
form is f,(2'w). That is,
For example, [A F],, = f2, - , # [A F],,,, - ,,. However,
if we replace m by 2'm in (3.5) and use Lemma 3.1, the
last step becomes an equality so that
[ ~ ~ ( ~ 2 ~ r n s O=) l [ nw 1 ( s 0 ) 1 n - m . (3.6) and
Thus, translating s o by 2'm translates octave i by m. ( D ' f ) , = fz(2'w). (3.13)
SHENSA: DISCRETE WAVELET TRANSFORM 247 I

Equation (3.11) is then equivalent to

I
si+l = ( D ' f ) * si (3.14a)
*; = (Dig) * si (3.14b)
where s o A s. This is essentially the original (undeci-
mated) 2 trous algorithm found in [4] and [ 5 ] . However,
we would like to emphasize that, since the development
in this subsection has not made use of any filter con-
straints, the general equivalence of the decimated output
of the algorithm pictured in Fig. 5 (equations (3.14)) and Odd

Q
that of Fig. 1 (equations (2.12)) follows for arbitrary fil-
ters.

A. An Alternative Implementation
A second possibility for the implementation of IV is to (zlj?a+l

use the algorithm in Fig., 1 and proceed directly from Fig. 6. Diagram of an implementation of the undecimated DWT.
(3.7). That is, to output iO.t',, one simply translates the sig-
nal by n and then computes wb. As an algorithm, this takes
the form a) compute wb; (b) translate the signal by one an exact implementation of (1.6) and for which $(t) and
sample; (c) go to step (a). Moreover, in view of (3.6), s ( t ) do (1.5) and (1.6) coincide? The general answer is
one need not reperform the entire recursion (,i.e., (2.12)) that we are able to construct such a rc/ provided the dis-
for every time point n in order to obtain w'. Rather, at cretized signal lies in the appropriate subspace of (cf.
each octave, the decimation is replaced by a split into even (2.26)). A somewhat surprising result is that it is neces-
and odd sequences, each of which is a starting point for sary and sufficient f o r f t o be a trous for condition (2.26)
the next octave (see Fig. 6). A couple of examples suffice to be dropped. Our approach shall be to mimic the con-
to convince one that if, at octave i , n mod 2' = 0, then struction of orthonormal wavelets outlined in Section II-C.
one takes the upper branch; if n mod 2' = 1, then one
takes the lower branch. A rigorous derivation follows from A. Construction of the Scaling Function 4
the formula (cf. [20])
We begin with the stipulation of the existence of a scal-
T m /2 seven m even ing function 4(t) with Fourier transform
ATms = . (3.15) 00
(JIa(,))
1 - w
T(rn- 1 ) / 2 ~ o d d m Odd
r= I
We remark that Figs. 5 and 6 are computationally
equivalent provided that the algorithm in Fig. 5 is imple- wherex(w) = ( f t ) , ( w ) is the z transform o f f t . To em-
mented efficiently. The code must be written so as to omit phasize the nonorthogonality of the corresponding wave-
multiplication by the zero elements of filters D'f.(They lets, we retain the symbol f rather than h. Note that this
are mostly zeros for i > 2.) Depending on the number of function 4 need not have (and in general does not have)
octaves, the computational burden still remains much all of the properties of the orthonormal 4 outlined in Sec-
greater than that of the decimated algorithm (i.e., Fig. 1); tion II-C.
however, there is considerable parallelism which may be For (4.1) to converge to a nonzero function, the factors
sufficiently exploited on suitable hardware to produce a must approach one. Thus, f , (0) = 1 , which implies
real-time implementation [5].
cfk = d2.
IV. THE DWT AS A N EXACTWAVELET TRANSFORM Even though 4 could be normalized differently by the in-
Regardless of the filters employed, one can, of course, clusion of a factor in (4.1), the filter f must necessarily
perform the recursions (2.12) or (2.13) on the sampled obey the low-pass condition (4.2); i.e., (2.14d). No,te also
signal s. Moreoever, provided that f (respectively, h) is that, under the chosen normalization, j + ( t ) dt = d(0) =
low pass and g bandpass, the procedure may be inter- 1. However, without spme additional conditions, the re-
preted physically as a bank of proportional bandwidth fil- lationship of 4 ( t ) to $ ( U ) remains somewhat tenuous.
ters (cf. [21]-[24] also Section VI). In the present section, Even under pointwise convergence, the limit may be a
we examine the mathematical significance of relaxing the highly discontinuous , fractal function [9]. Suitable regu-
filter constraints (2.9) and (2.14). Our goal will be to re- larity conditions for the inverse Fourier transform of a
late the more general filter bank to the continuous wavelet product of the form (4.1) to converge to a reasonably be-
transform, thus, in a sense, justifying the term DWT (cf. haved (e.g., L'(Z?), L2(R), and/or continuous) function
[25]). In this endeavor, the major questions which we shall may be found in [8] and [26]. The results are summarized
address are: for what functions $(t) is the recursion (2.12) in Appendix B.

I
2412 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, NO. IO. OCTOBER 1992

A time domain representation of 4(t) in terms off can We find, for example, that q j is zero for t < ti where tj
be derived from (4.1). Let x be the indicator function of - (tj- - N-)/2. With to = - 1/2, and with a similar
[-1/2,1/2) calculation for the right half interval, it follows that 4(t)
= limj-w qJ(t) is zero outside [-N-, N,].

B. Exactness
From (4.1) and Lemma 3.2, it is not difficult to see that To avoid confusion and stress their differences, let us
first recapitulate some definitons. Four different trans-
4(t) = lim Ck [(AF)i]okJ2i~(2it- k). (4.4) forms W(a, b), W(2', 2'n), w(2', 2h), and w; have been
1-03
mentioned ( ( l . l ) , (1.5), (1.6) and (2.12)). We retain a
In fact, the Fourier transform of (4.4) is7 terminology parallel to Fourier transforms, namely,
wavelet transform (WT), wavelet series,8 discretized
wavelet series, and discrete wavelet transform (DWT),
respectively. The first two transforms involve integrals of
a continuous signal; the latter two contain sums of sam-
pled signals. The first three utilize a continuous wavelet
function $(t), the last one employs the discrete filters g
and f. For consistency, we shall continue our develop-
which is just (4.1). Note that the existence of the function ment using decimated transforms; however, the results
(4.4) could have been taken as the starting point for our hold without change for undecimated transforms. This
analysis, since our proofs will not make use of (4.1). follows immediately, since they coincide for n = 0, and
We continue our parallel with Section 11-C. Define the undecimated transforms may be obtained at time n =
4k(t) by definition 4.1, as follows. no, by translating the signal by no samples and taking the
Definition 4.1: transform at n = 0. (See Section I11 where definition 3.1
remains valid for W(2', n) and w(2', n)).
Our starting point shall be a signal s(t) and discrete fil-
tersfandgwith w'definedby (2.12)ord1by(2.13), i.e.,
Then, substitution of 2't - n for t in (4.4) and the use of =
sl+l
A Fs'
Lemma 3.1 yield
w' = Gs'
J2'4(2't - n)
d'fl - Aw'. (4.11)
= lim Ck [(AF)']okJ2J+1x(2J+1t- 2'n - k)
,+m Recall that the matrices F, and G,, are given by f, -
= lim C [ ( A F ) ' - ' I ~ , ~ - ~,,J2'~(2't
, - k) and g , -,, respectively. Of course, we must also specify
j-m k an initialization of the recursion (4.11) for some i; for
example, for the zeroth octave so. The obvious choice is
= lim [(AF)'-'],kJ2JX(2Jt - k). (4.7)
J+W k s: A s(n) (4.12a)
On replacing i by -i, this becomes however, we shall also consider
4h(t) = lim C [(AF)'+'],d2'~(2~t - k). (4.8) s: G Ck - n)s(k) (4.12b)
j-w k

An immediate consequence of (4.8) is which relates to the discretized wavelet series w(2', 2h),
and
(4.9)
s: A j 4(t - n)s(t) dt (4.12~)
paralleling (2.22). Thus, we see that, despite their lack of which corresponds to the sampled WT (wavelet series).
orthogonality, the 4k(t) have retained most of their struc- For a given g, we shall construct a continuous function
ture. Furthermore, i f f i s a finite filter, then 4(t) has finite $(t) such that the DWT of (4.11) is an exact implemen-
support [8]. More, precisely, suppose the coefficients of tation of the discretized wavelet series under (4.12b) and
f are zero outside [-A'-, N+]. Let q J ( t ) of the wavelet transform under ( 4 . 1 2 ~ ) .
[(AFJ]okd2jx(2jt - k). Then, q J ( t ) converges to 4(t), Define $ ( t ) by
and, as in (4.9), we have
$(t) 2 4(t + k)gk = 4(c?- k ) g i (4.13)
(4.10)
'At times, we shall prefer the term sampled WT rather than wavelet se-
'This equality is immediate from Lemma 3.2 and i ( w ) = ries in order to emphasize its role as a restriction of the continuous trans-
(2 sin ( w / ~ ) ) / w . form.
SHENSA: DISCRETE WAVELET TRANSFORM 2413

... -
- lim [AF],J2 = 6,o. (4.20)
j+m

The import of Theorem 4.1 quickly follows. Equation

(4.12b) implies (4.12a) for arbitrary signals if and only if
The recursion (4.9) implies that q5,h(t) = Ck 4(n) = Ana. Thus,
[(AF)'Ink+;(t). Using this expression, along with (4.6), Corollary 4.1: The algorithm'(4.11) is an exact im-
the definitions (4.13) and (4.14) yield plementation of the DWT with so set equal to the sampled
signal if and only iff is an trous filter.
Furthermore, g is also related to $ ( t ) by sampling:
Corollary 4.2: Iff is an ? trous
i filter, then $(t) defined
by (4.13) satisfies $(n) = g:.
Wavelet coefficients of the type d' = Aw' - are easily '
obtained by replacing $(t) by J 2 $(2t). More precisely,
= zn-k$';(r) +
let $ ' ( t ) & d 2 $(2r) = J 2 Ckgk4(2t k ) , which coin-
cides with (2.23). Then, provided s o satisfies (4.12c),
= C [G(AF)'InkQ:(t). (4.15) (4.11) and (4.17) imply
d', = w i i '
If we take (4.12b) as the definition of so,the discretized
wavelet series (1.6) takes the form

w ( 2 , 2%) 2 C J.k(m>s(m>
m
= s $i;'(t)s(t) dt

=
C
ink
[G (A ~ 1 ' 1 n k Q i (m>s(m)

[G(A F)'s0],
which, by (4.11) is exactly wk. Furthermore, under
(4.16) = s $;'(t)s(r) dt. (4.21)

Hence, although the $;'(t) are not orthogonal, the Mallat

( 4 . 1 2 ~ )we have (again using (4.15))
algorithm still computes the wavelet transform. Of course,
W(2', 2'n) 1II

T ' , ( t ) s ( t ) dt
u n g r (4.12a) and h trous, or for (4.12b), we have d', =
Ck $ A' (k)s(k), the counterpart of (4.16). It is interesting

= s f; [G(AF)'In,4i(t)s(t) dt
that, in a sense, the decimated wavelet transforms (1.5b)
and (1.6) contain superfluous information. That is, they
are underdecimated by a factor of two, and, thus, w ' prop-
erty belongs to octave i 1. +
= [G(A~ ' l n k j 4; (t>s(r) dt
C. Summary
= [G(AF)'s0], (4.17) Let us summarize the results of this section. We are
again w',. given discrete filters f and g such that (4.4) is well de-
Finally, let us investigate the significance of the ii trow fined. Define $(t) by
condition; i.e., of the constraint (2.9), hk= hkO/J2.We $(t> 40 - k>skt (4.22)
prove the following theorem.
Theorem 4. I : f is an ii trow filter 4(n) = ana. with corresponding transforms sampled WT (wavelet
Proof: Letting i = -1, n = 0, and t = n in (4.9) series)
gives
4(n) = C [AF]okJ2 4(2n - k ) . (4.18) (4.23)
k
Then, and discretized wavelet series
4(n) = 6nO * J2hn = 6ao- (4.19) w(2', 2%) A c s;i,(k)s(k).
k
(4.24)
Conversely, suppose that f2,, = s n O / J 2 . Then, since
Let d,, * s stand for the scalar product Ck &(k)sk, and
x ( - k ) = Am, (4.8) with i = 0 and t = 0 implies #,(t) - s ( t ) for the L 2 scalar product j & ( t ) s ( t ) . Then
Q(-n) = lim [ ( A F ) J ] n o J 2 J
J+m f i s ii trous * $(n) = g:. (4.25)
= lim C [(AF)J-']nk[AF],J2 J2'-' For s discrete:
J+CD k
so = s andfis i trous 3 w(2', 2%) = wb (4.26a)
= lim [(A F)' - ']n0J2/- I
/-a
s: = & s 3 w(2', 2'n) = wi,. (4.26b)
2474 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40. NO. IO, OCTOBER 1992

For s(t) mtinuous: The minimum number of coefficients (unknowns)

equals the dimension of the space, i.e., satisfies 2N = M
s: = +,(t) s(t) * W(2', 2'n) = wf. (4.27) + 1 . Thus, the sums in (5.3) go from 1 to N. Moreover,
Ail the above results extend immediately to the undec- since for any n , the polynomials Q(x) L P(x - (n - 1)/2)
imated transforms, W(2', n ) , w(2', n ) , and @: by transla- also form a basis, (5.3) is equivalent to
tion invariance and definition 3.1. Also, the correspond- N -N
ing properties hold for d' with $(t) replaced by J 2 $(2f).
Note that the orthonormal wavelets are only exact under
P(i) = k= Ic akP(l - k) + k =C- I akP(-k) (5.4)

(4.26b) or (4.27) since their filters h are not a trous filters. which must hold for all P(x) of degree 1 2 N - 1.
Furthermore, they do not satisfy (4.25). We pick out the kth coefficient by letting P be the La-
grange polynomial
V . LAGRANGE INTERPOLATION FILTERS
A reasonable class of a trous interpolators to consider a
II (x-i)
LfN-'(x) =
i#j
i , j in [-N + 1, N I .
for (2.10) are those which are exact for polynomials P(t) II ( j - i)
of degree I M for some M , i.e., for which i#j
(5.5)
-P
$2 (;)
- = C f;-zkP(k).
k
(5.1)
L ~ N - ' , we get
~ , that replacing P in (5.4) with
Then, LfN-'(k) = 1 3 ~so

For reasons which will become clear very shortly, we shall aj = L ; ! J ' ( ~ ) f o r j = 1, - - ,N (5.6a)
call these filters Lagrange a trous filters. Since the a trous
filter f satisfies f2k = 13,,/J2, (5.1) is an identity for n a -J. = L ?J N - ' ( ~ > f o r j = 1, - - , N. (5.6b)
even. Let a contain the odd components o f f t
Inasmuch as the Lagrange polynomials L 2 N - ' form a ba-
J2fikp1 fork >0 sis for polynomials of degree less than or equal to 2N -
1 , these ak are in fact the unique solution to (5.4).
fork < 0. (5.2) It is also straiglitforward to see from (5.5) and (5.6) that
for k = 0 a is symmetric, i e., f o r j > 0
.
Then (5.1) is equivalent to II (1/2 - i ) iII (-1/2 + i )
i#l-j
- #j

P (f) = kFoakP (F - k)
aj =
II ( - j + 1 - i ) iII
i# I -j

II (1/2 - i )
#j
(-j + i )

+ k c< O akP (T - k) (5.3)

-
-
i#j

II
i#i
( j - i)
- a-j.
-
(5.7)
for n odd and for all polynomials P of degree IM . (Ac-
tually, a single value of n implies (5.3) for all n; see In summary, we have the following theorem.
(5.4).) We proceed to express the ak in terms of Lagrange Theorem 5.I : Let f be an a trous filter, i.e.,
polynomials, and to show that the above conditions are 1
essentially equivalent to f = h * h ' / 4 2 where h is an hk = J2 I30k. (5.8)
appropriate Daubechies filter.
Assume, furthermore, that f is real with symmetric sup-
A . Construction of a port described by k E [-2N + 1 , 2N - 11. Then, f is-a
Lagrange trous filter, that is, (5.1) holds for all poly-
First, we parameterize the family of filters a satisfying
(5*3)by + ''
the dimension Of the space Of
nomials P of degree 1 2 N - 1 , if and only if the odd
components off are determined by (5.6). Furthermore, f
mials for which it must hold. For such a relationship to
is necessarily symmetric.
exist, one must relate the length of the filter (the number
of unknowns) to M . To accomplish this, we shall assume
that a has exactly the minimum number of coefficients B. Relationship to Daubechies (QMF) Filters
needed to satisfy (5.3). We further assume that a has sym- In [8], Daubechies constructs essentially the entire class
metric support; i.e., there is an N such that ak = 0 for (kl of finite length filters h which satisfy (2.14) and fulfill
> N and ak # 0 for Jkl = N. This assumption is not suitable regularity conditions on (2.18). Explicitly, they
unreasonable, at least for symmetric wavelets $ ( t ) , since take the form
there is no reason a priori to distinguish between t and
- t , and one would even expect a to be symmetric. We (5.9)
shall see, in fact, that the weaker condition of symmetric
support joined with the previous constraints implies that where h ( z ) C h , P is the z transforms of h , and Q is
a actually is symmetric. an appropriately constrained polynomial. (In this section
 2475
SHENSA: DISCRETE WAVELET TRANSFORM

it is convenient express the z transform as a polynomial where i = 1, , 2N - 1. The signs work out since n
which we denoted 6 ( z ) where h,(w) = h ( e j w ) . )Her deri- - 1 + +
2k - i and n + 1 2k - i have the same parity
vation uses a specific g, which up to a phase factor is while n - i differs. Define Po@) = 1. Then, since the
given by polynomials PO(x)and Pi (x) for i = 1, -
, 2N - 1 form
g, = (-l)%(l - n) (5.10) a basis for polynomials of degree 1 2 N - 1, and since
(5.14b) implies (5.17) for i = 0, (5.17) must hold for
and, as a consequence, (2.14a) reduces to arbitrary polynomials P(x) of degree 5 2N - 1. Replac-
A h h t D = 1. (5.11) ing Pi by P in (5.17) and setting ak = bIkl yields (5.3).
Conversely, Theorem 5.1 implies that if a satisfies (5.3)
In other words, [hh t ] 2 n = A,, which is the B trous con- for all polynomials of degree 1 2 N - 1 and has symmet-
dition for hh t / d 2 . Finally, if Q is taken to be of minimal ric support, then it must be symmetric. Clearly (5.17)
degree (which turns out to be N - l ) , (el2
is unique [8]. must be also be satisfied. From (5.2) and reversing the
In other words, the squares of these filters are character- above algebra, thisis equivalent to (5.12a) and (5.14) with
ized completely by satisfying (2.14d), (5.1 l ) , and being J 2 f(z) replacing p (2) where f(z) is of degree of 2N - 1
of the form (5.9) with degree 2N - 1. We proceed to in z and also in z - ' . Letting n = 1 (or 0) in (5.14a), we
show that the h * h equal the Lagrange B trous filters. see thatf(z) has 2N roots at z = - 1. Sincefis symmetric,
Let h be the Daubechies filter of order 2N. Since h * f(z) must also have the form (5.12b). We conclude that
h is symmetric, the above conditions are equivalent to J 2 f = h * h t where h i s a Daubechies filter. Thus.
N N Theorem 5.2: There is a one-to-one correspondence be-
B(z) h * ht = 1 + c bk.Z2k-i+ c b k z - 2 k c '
1
tween the squares of Daubechies orthonormal wavelet fil-
ters h of length 2N and the Lagrange B trous filters f of
(5.12a)
length 4N - 1 given by f = h * h t / d 2 .
and Note that one can compute the h of length 2N by taking
b(z> = i ( i ( 1 + ~ ) ) ~ ( ; ( +1 z - ' ) ) N Q ( ~ ) Q ( ~ - ' )(5.12b)
. all possible square roots of the Lagrange a trous filters f.9
f is easily computed from (5.6) where its even compo-
That is, J 2 f = h * h if and only if f(z) is of the form nents given b y h k = 6,/.j2, its odd positive components
(5.12a), (5.12b) and is of degree 2N - 1 in z (respec- by f 2 k - = bk for k = 0 to N, and foraodd negative k by
tively, in z - ' ) . symmetry. Also, the spectrum o f f , f ( e I w ) = Jh(eJw)12,
Next, we show that the bk coincide with the ak of (5.3). presents a convenient method of computing the power
We multiply (5.12a) by z n for an arbitrary integer n, spectra of the h's. In another vein, since the h are maxi-
N N mally flat filters (i.e., have same number of vanishing de-
Z"b(Z) = Zn + c bkZn-1+2k+ c b k Z n + i - 2.k 45.13) rivatives at z = l and z = -l), Theorem 5.2 shows that
a maximally flat filter is a Lagrangian interpolator; a fact
The 2N zeros at z = - 1 in (5.12b) imply that which may aid in the design of such filters [ 141.

VI. WAVELET FILTERSIN SIGNAL PROCESSING

This section has its roots in a question which originally
and since &-1) = 0, we also have motivated the author to undertake this study: Inasmuch as
N the B trous and Mallat algorithms share the same recur-
2cbk=1. (5.14b) sions, why not choose the Daubechies filters since they
enjoy the additional advantage of orthonormality? The
Next, we note that strongest arguments in favor of orthonormality seem to be
mathematical elegance, ease of inversion, and, more sub-
tly, good numerical properties. The major drawback is a
lack of flexibility in filter design, in particular, an essen-
tially fixed relative bandwidth. On the other hand, we have
Define a set of polynomials Pi by seen that the DWT has a firm analytical basis independent
Pi@) 4 2x(2x - 1) * (2x - i + 1). (5.16) of the B trous approximation, even in the nonorthonormal
case. In the present section, we shall briefly examine the
Finally, setting the ith derivative of the right-hand side issues of inversion, boundedness, and adjustable relative
of (5.13) at z = - 1 to zero, utilizing (5.15) and definition bandwidth. (Another fundamental issue, regularity of the
(5.16), we obtain associated wavelet functions, appears in Appendix B.) In

(;> c
P; - =
k>O
bkPi ("-
2 + k) 'During the revision of this paper, it was brought to the author's attention
that an implicit relationship between the squared filters and Lagrange in-
terpolation had been independently noted in private conversations between
n + l
c
4- k > O bkP; ( 7 - k) (5.17) I . Daubechies and Ph. Tchamitchian. Similar observations are to appear in
t271.
2476 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40. NO. IO. OCTOBER 1992

particular, we wish to establish conditions under which

these properties obtain, expressed directly in terms of the
filters f and g rather than in terms, for example, of the
scale function +(t) or wavelet $(t). This view is in the
spirit of the DWT as an entity in its own right, and it is
certainly a necessary element in deciding which filters to
use in practice.
gl$

Fig. 7 . Illustration of a single stage and its inverse for the undecimated
Any software realization of the wavelet transform only algorithm found in Fig. 5 .
implements a finite number of octaves. Mathematically,
this reduces inversion to an algebraic question, one of version are much less stringent. In order to invert a stage
finding filters which satisfy certain (not overly restrictive) of the algorithm of Fig. 6, the filters p and q need only
equations. However, other considerations begin to come satisfy (cf. Fig. 7)
into play. Exact inversion requires finite filters, and, even
then, exceedingly long filters may not be useful. More- SI = p * SI+' + q * W'
over, the constrained problem is considerably more diffi-
cult to solve. An alternative approach, approximation by = ( p * f +q*g)*s'. (6.3)
truncated infinite filters, might be acceptable, but, once That is,
again, practical considerations dictate that the filters de-
p * f + q * g = 6 (6.4)
cay quickly. Similarly, the behavior of the DWT at infin-
ity (i.e., w ias i goes to infinity) becomes relevant. For where the Kronecker delta, 6 A 60,m, is the identity for
example, the condition C, g , = 0, the discrete counterpart convolution. This is a single equation, and consequently
of (1.2), is not necessary for inversion of a finite number less restrictive ht!n (6.2). If the polynomials formed by
of stages. However, it is necessary to finite energy and the z-transforms f ( z ) and g ( z ) are relatively prime, one
boundedness, which are desirable properties inasmuch as may apply the Euclidean algorithm for the greatest com-
they reflect directly on the numerical stability of the al- mon divisor (in this case, one) to find p and q. It has the
gorithm and/or its inverse (see, for example, [3]). advantage that finitefand g lead to finite p and q. Another
method is simply to solve the equation in frequency space,
A . Inversion
P z ( w ) f i ( w )+ qz(w)gz(w) = 1. (6.5)
To invert either the decimated or undecimated discrete
wavelet transform it suffices to invert a single stage (oc- There is almost too much flexibility in solving this equa-
tave); that is, to find S I , given S I + ' and w l + 'or W"'. tion, although it becomes much more restrictive if one
The equations for inverting the decimated algorithm are demands that the filters be finite or rapidly decreasing.
exactly analogous to those for the Mallat algorithm pic- Once again, a popular choice [30] is f , f , +
g,g, = 2,
which, for example, can be solved for g z by taking the
tured in Fig. 4. One seeks two filters p and q which invert
a single stage of the decimated DWT in Fig. 1; i.e., such square root of 2 - 1 fzI
as long as 1 f ,(U) l2 I2. (Or, vice
versa, it can be solved forf,.) The Daubechies (QMF)
that
filters h/d2 and g/d2 certainly satisfy this equation so
SI = PDs'+' + ells1+' that inversion for the undecimated version of the Mallat
= (PD)(AF)s' + (QD)(AG)s'. (6.1) algorithm is immediate. Another case, useful in signal
processing, is to choosefto be h trous, p = f t / 2 , and g
Equivalently, any filter with nonvanishing spectrum except possibly
where I f,(u)I equals d2 (see next subsection). Important
(PD)(AF) + ( Q D ) ( A G )= Z (6.2) questions of numerical stability, filter lengths, etc. cer-
where Z is the identity matrix. This type of equation, tainly remain to be answered, but are well beyond the
which in the frequency domain may be separated into two scope of the present paper.
equations comparable to (2.14a) and (2.14b), has been Finally, before departing from this subject, it should be
treated extensively (but not exhaustively) in the subband mentioned that inversion of the undecimated case in the
coding literature (cf. [12], [13], or even [8]). The QMF form of Fig. 5 also follows from (6.4). The inverting fil-
filters of the Mallat algorithm satisfy (6.2) with g , ( w ) = ters are just D'p and D'q. It is a simple matter to verify
f,(w + T ) , P = f t , and q = g' (i.e., (5.10) and (5.11)). that (6.4) implies that
A less restricted class is just p = f t and q = g'. The
general class of filters satisfying (6.2), so-called bior-
+
(D'p) * ( D ' f ) (D'q) * (D'g) = 6 . (6.6)
thogonal filters, are examined in [28] and [29]. It should (Inserting zeros in 6 just yields 6.)
be emphasized that for perfect reconstruction in applica-
tions all filters must be of finite length (FIR). This does B. Finite Energy and Boundedness
not imply that infinite filters (IIR) implemented by their The discrete wavelet transform is a m.apping of se-
truncations are not worthy of consideration [28]. quences s,, n = 1,. 2, * * * into the space of doubly in-
For the undecimated algorithm the requirements for in- dexed sequences wk, i , n = 1, 2, . * * . Finite energy for
 SHENSA: DISCRETE WAVELET TRANSFORM 2477

the signal is simply we see that the wavelet transform will have finite energy
c 1s,12 <
n
OQ. (6.7)
if and only if

One’s initial tendency is to look for the same condition

c 21 11wq2 + IIf”112 <
I L
00. (6.13)
on the wavelet transform, i.e., C i , , Iwk12 < OQ. However,
This is a necessary condition for the mapping w , from
the actual situation is not quite so transparent. Referring
Z2(Z)to Z2(Z2;2-’, 1) to be bounded. Of course, in prac-
back to the continuous case, we have [ 161
tice, implementations never compute an infinite number
of octaves. Nevertheless, the property (6.13) of finite en-
ergy can be quite important. Unbounded transformations
tend to have poor numerical behavior even when trun-
i.e., the weight (measure) da d b / a 2 results in units of cated. Similarly, the inverse will not be bounded unless
energy. For the DWT, a = 2’ and db is either 1 (the un- the series (6.13) is bounded below. A wavelet represen-
decimated case) or 2’ (the decimated case). Discretizing tation which has these properties,
and noting that d u / a = d(ln U ) , we see that
1
CI 2 11 + I(fm112IB \ ( S \ ( ~(6.14)
du db
-+
U2
(In 2) Ai A b
2’
= In 2 ($) (6.9)
AllsJ12 5
L

is called a frame (cf. [3], although here the ambient Hil-

since A i = 1 . Observing that octave i has a bandwidth, bert space is 1’ rather than L2(R).)We proceed to derive
up to a constant factor, of 1/2’, we assign (6.9) the fol- conditions on the filters f and g for (6.14) to hold.
lowing physical interpretation: The discrete wavelet Equations (3.14) imply
transform is in units of power/hertz. Multiplying by the
bandwidth 1 /2’ gives power, and additional multiplica-
tion by the time interval A b gives energy. Note that the
decimated version only outputs every 2’ points so that in
that case A b = 2’. In summary,
i) Iw:I2, I Wh12 are in units of power/hertz;
ii) octave i has bandwidth = 1 /2‘; Suppose that
iii) I W i 12/2’ and IwL are in units of energy.
max (If,(ct1)1~ + I gz(w)12)s 1. (6.16)
One should take care to note that for w ; , the energy weight U

du d b / u 2 is a constant independent of i so that IwfI2 is Then, in the time domain, (6.15) and (6.16) imply
discretized in a fashion so as to be both power/hertz and
energy /cell.
Let 1 1 ~ 1 1 ~ C, be the squared norm of s, and define L

(1f”(12& 2 lim sup 51 l(s‘((* (6.10) Adding 1) @‘-‘~~2/2’-’ to both sides and repeating for de-
i+m creasing octaves implies that
which corresponds to the DC energy (i.e., at w = 0). Fi-
nally, define the energy of the DWT by

E = c’ 721 I( w ’ + 1)s”(I2 (6.1la) Finally, letting J go to infinity, we get not only (6.13),
but also the right inequality of (6.14) with B = 1 .
E = l(~’11~.+IS"(^^. (6.1lb) However, the condition (6.16) is much too strong. That
I
is, the transformation g -+ Cg for a large enough constant
*
Energy conservation takes the form of the following Par- C would cause (6.16) to be violated even though C has
seval’s relationship for discrete wavelets: no effect other than to multiply the total energy by a con-
Dejnition 6.1: A particular choice of filters f and g is stant. In fact, the filters f and g produce finite energy
said to be energy conserving if, for some constant C , transforms if and only i f f and Cg yield finite energy.
Thus, to have finite energy, it is sufficient to find a C >
0 such that max, (1 fz(w)(2 + C1 gZ(w)l2 I2. Such a C
exists provided that 1 f,(U)’ I 2 and (1 /2) 1 gZ(w)l2/
One may also specify conservation for decimated trans- (1 - [ I /2] 1 f,(w)I2) is finite; i.e., is less than some finite
forms, in which case the 2’ is dropped. B = 1 /C. A similar argument holds for the lower bound.
Except for some clarifying remarks at the end, we re- If
strict the discussion in the remainder of this subsection to
the undecimated DWT. Following the above definitions, mwi n i <IfZ(U>l2+ I gz(412)2 1 (6.19)
2478 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, NO. IO, OCTOBER 1992

then (6.18) holds with the inequality reversed and the left b) Low-pass filter to obtain the lower half of the spec-
inequality of (6.14) holds with A = 1 . Once again, we trum (10, 7T/21).
apply the trick with the constant C and find that, for the c) Decimate to expand the lower half to [0, n].
inverse to be bounded, it is sufficient that there exist A = d) Go to a).
1/C > 0 such that (1/2)) gZ(w)l2/(1- [1/21 If,(w)12) In somewhat more detail: We first obtain the high fre-
2 A. In summary,
quency information by using g to filter the upper half of
Theorem 6.1: A sufficient condition for the undeci-
the spectrum of s'. The filter output is w ' . Then, in prep-
mated DFT IV and its inverse to satisfy (6.14) (that is, to
aration for the next octave, s' is low-pass filtered by f.
be bounded) is that, for all U , l f , ( ~ ) ( ~ 5 2 and
This retains the, as yet unexamined, low frequency con-
tents and also prevents the upper half of the spectrum from
aliasing (i.e., contaminating the low frequency contents)
in the dilation which follows. Finally, the operator A
To satisfy (6.20), one must have If,(w)l = 2 g,(w) = spreads" the remaining energy to fill the spectrum, pro-
0, and the multiplicities of the corresponding roots must ducing octave i +1 . The procedure then repeats itself,
+ I is bandpass filtered to get the spectral contents at
be identical. Note, also, that (6.20) can be used to give
an estimate of B / A , the so-called tightness of the frame. frequencies which are, in absolute units, one half the fre-
Whether these conditions are also necessary remains an quencies of the previous octave.
open question. One can, however, show from (3.11) and A potential problem is immediately apparent. If the
an examination of the power IVi(w) at w = 0 that a nec- bandwidth of gz(w) is less than ~ / 2 a, portion of the sig-
essary condition is g, (0) = 0 (equivalently, C, g, = 0). nal energy will be discarded; it never appears in w ' . One
This is the discrete analog of the admissibility condition possible remedy is to make g, sufficiently broad; how-
(1.2). The author conjectures that in the discrete case it ever, that would limit the resolution. Alternatively, we
is not a sufficient condition. (We remind the reader that may introduce so-called voices. That is, we can employ
even in those cases for which the DWT is exactly the sam- a bank of filters of the type g (see Fig. 9) in order to cover
pled WT, finite energy of the continuous wavelet trans- the entire upper half of the spectrum.
form does not imply that of the discrete transform.) We We formalize some of these concepts using the modu-
do have, however, the following theorem: lated Gaussian of (1.4) as an example. With the introduc-
Theorem 6.2: A necessary and sufficient condition for tion of an additional parameter P , +(t)becomes
energy conservation (6.12) is that for all w A ej v t e - - p 2 t 2 / 2 .
+(t) = (6.22)
1 1
- (IfZ(~>l2
2
+ 2I gz(w)I2)= 1 . (6.21) Its Fourier transform is given by

To prove this we may, without loss of generality, set C (6.23)

= 1 (i.e., redefine g by the constant factor C). Sufficiency
follows as above, with inequalities replaced by equalities. We define the bandwidth of $(U) as twice the interval
To prove necessity, we first note that energy preservation between points for which the modulus of (6.23) drops to
for arbitrary signals implies the energy must be conserved 1 / e of its peak value, i.e.,
for each stage. (For example, if the signal is s ', energy
must be preserved, and since it is preserved for so, the BW A 2d2 P. (6.24)
first stage must preserve energy.) From (6.15) this implies The filter g (which, here, equals g t ) is the sampled ver-
(6.21) with C equal to one. sion of (6.22), that is,
The decimated case seems to present problems. In or-
der for the above proofs to carry over, the even part of "- eJ~ne-82n2/2. (6.25)
the signal would need to compensate for the lack of the For convenience, we set the sample rate equal to one. The
factor 1/2' in (6.11b), but I ~ s ~ , ~ . , I #
~ ~ l / 2 l ~ s i 1 ~ *This
. following three restrictions on v and 0are necessary: First,
problem presents yet another area for additional research. in order that gz(w) lie in the upper half of the spectrum
(cf. Fig. 8), we require that
C. Resolution and Relative Bandwidth
7T
Considerable insight may be gained by viewing the al- - 5 v. (6.26a)
gorithm in the frequency domain. One stage of the deci- 2
mated DWT, illustrating (4.11) from this point of view, Next, in order that + ( t ) is admissible and analytic (see
is pictured in Fig. 8. Since we are dealing with the dis- [ 18]), we demand
crete wavelet transforms, $(z) = $(ej") is evaluated on the
unit circle. For convenience only the positive frequencies (6.26b)
are pictured. Briefly, the algorithm is
a) Bandpass filter the upper half of the spectrum to "A, which decimates, is a contraction. Thus, in the Fourier domain, it
yield w ' . is a dilation.
 2479
SHENSA: DISCRETE WAVELET TRANSFORM

0 ff

Fig. 8. Illustration of one stage of the algorithm, octaves i and i + 1 inclusive, viewed from the frequency domain.

g, considered as a filter on octave i is v/2', and its band-

width is 2 J 2 p/2'. For this reason it is simpler and less
3c E *e ambiguous to speak in terms of a relative bandwidth which
. . .. is independent of decimation. More precisely, we define
BW
RBW (6.28)
mean frequency '
(Any appropriate representative of the "center" of the
filter can replace the mean frequency in (6.28).) In the
case of g , we have
(2 J 2 p / 2 7 - -2 J 2 p
0 T/L R RBW = - (6.29)
(v/2') U
Fig 9 Plot of the power spectra of the bandpass filters for four voices
where $ ( U ) is given by (6 23) Also, since n / 2 5 v 5 n, we have 2 d 2 p/a 5 RBW
I4 d 2 p/?r, or approximately

Under (6.26b), g,(w) -

0 for w I0. (Reference [18] p 5 RBW I2p. (6.30)
recommends 0 Iv/5 as being sufficient.) Finally, in or-
In view of (6.30), we shall consider the parameter @ as,
der that the spectrum not be aliased, we set
essentially, the relative bandwidth of g .
v IT - J2p. (6.26~) The number of voices M which we would expect to need
to cover the upper half of the spectrum is
These may be summarized in
a/2 1
max (2nP, n/2) Iv 5 ?r - J2 6. (6.27) M---- (6.31)
2J2 p 26'
At this point a word of caution is advised. The band-
width ofthe discrete filter g (e.g., (6.25)) is 2 J2 p only The filter for voicej, which we denote is determined
when the sample rate is 1 . Since the ith octave is the re- by sampling the function $ ( t / a ' I 7 i.e.,
sult of i decimations by 2, sampling the original signal so [p],2 rl/(n/aJ> f o r j = 0, * ,M - 1
at a rate A r = 1 results in a Nyquist frequency of n/2'
for octave i (i.e., for si). Thus, the central frequency of and a 2 2'IM. (6.32)
2480 IEEE TRANSACTIONS ON SIGNAL PROCESSING. VOL. 40. NO. IO. OCTOBER 1992

The power spectra of the ,v are illustrated in Fig. 9. An VII. CONCLUSION

alternative would be to define the voices as frequency We have seen that the B trous algorithm bears an inti-
translations of the filter g. However, (6.32) seems to be mate relationship to Mallat's multiresolution algorithm.
a more natural definition since it maintains the affine Originally devised as a computationally efficient imple-
structure. (It is equivalent to taking a nondyadic value for mentation, it is more properly viewed as a nonorthogonal
the dilation parameter U in (1. l ) . ) Note that the band- multiresolution algorithm for which the discrete wavelet
widths of the voices decrease with j , and the spectral transform is exact. Moreover, the commonly used La-
spacing 2 d 2 p /CY'differs somewhat from that assumed in grange B trow filters are simply the convolutional squares
(6.31). of the Daubechies filters for compact orthonormal wave-
Control over the relative resolution is important be- lets.
cause, although the absolute resolution (,0/2') improves From a broader viewpoint, these two algorithms are in-
at higher octaves, at the same time, the analysis frequency stances of the discrete wavelet transform (DWT), which,
decreases. If one wishes to improve resolution at a given in more conventional terms, is simply a filter bank utiliz-
frequency, one has to better the relative resolution, i.e., ing decimation and two filters. There are two basic ver-
decrease 0. On the other hand, the standard tradeoffs ap- sions of the DWT one of which is simply the decimated
ply to choosing p. Small increases the relative resolu- output (octave i is decimated by 2') of the other. The
tion but also requires more voices and a longer filter g; decimated DWT is characterized by octaves a) obtained
thus, more computation. (The duration of $ ( t ) is on the by alternating a low-pass filter f with decimation, and b)
order of 2 J 2 / p so that to provide a reasonable approxi- tapped by a bandpass filter g to produce the output. The
mation to (6.22), the length of g must be proportional to undecimated DWT inserts i zeros between the elements
1/ p . ) Moreover, the increased length of g implies a wors- of the filters at octave i in lieu of decimation. (In the case
ening of the (relative) time resolution. The time-band- of voices, several g's are used.) Finally, we note that un-
width product is bounded below by the uncertainty prin- der very general conditions, there exists a function $(t)
ciple, and no amount of computation will simultaneously such that the filter bank outputs wi correspond to the sam-
produce arbitrarily small time and frequency resolution in pled wavelet transform
a single wavelet transform.
With respect to choosing the low-pass filterf, we note
that the spectrum of a longer filter f will generally have a
sharper cutoff. This cutoff is relevant because it prevents
thus, justifying the terminology discrete wavelet trans-
the energy in the upper half of the spectrum from leaking
form.
(aliasing) into the lower half under the decimation A. For
The personality of a given DWT is distinguished by the
most applications, a Lagrange B trous filter of length 7 (N
choice of filters. Iff satisfies the B trous condition f2n =
= 2) is sufficient [ 5 ] . One can, of course, also use the
hnO/J2,then g is the sampled version of $(t); i.e., g, =
-
asymmetric filters h. Their discrimination of temporal di-
$ ( - n). If finite length f and g obey the constraints of the
rection seems intriguing, but remains univestigated.
multiresolution algorithm, then the J2' $(2't - n) are the
Finally, how do the above considerations relate to or-
compact orthonormal wavelets. A number of fundamental
thonormal wavelets? The power spectra of filters h and g constraints have been discussed. In various combinations
obeying (2.14) satisfy [8], [20]
they have a bearing on the regularity of the wavelet func-
lh,(412 + I g,(4l2 = 2 (6.33a) tion, on the energy in the transform domain, and on the
boundedness and invertibility of the transform. In partic-
and ular, we have provided a set of conditions on the filters
Ihml = I gz(&7r)l= .j2. (6.33b) sufficient for the transform and its inverse to be bounded.
The signal processing properties of the discrete wavelet
It follows that, for positive frequencies (likewise, for neg-
transform depend particularly strongly on the choice of g.
ative frequencies), g and h must each maintain a band-
The general constraints mentioned above are not restric-
width on the order of 7r/2.11 Thus, in exchange for or-
tive on g; however, there is considerably less freedom in
thonormality one relinquishes control over bandwidth.
the orthonormal case. In particular, if orthonormality is a
The relative bandwidth is, essentially, fixed at
requirement, the half bandwidth of g (and, hence, the rel-
(n/2)/(37r/4) = 2/3. On the other hand, if one wishes
ative bandwidth of the wavelet) is no longer adjustable.
to give time and frequency localization equal weight (e.g.,
It remains fixed at approximately 7r/2.
= 1 SO that bandwidth = duration = 2d2), a relative
Many topics remain for investigation. Although con-
bandwidth in the neighborhood of 2 / 3 is in a sense opti-
siderable work has been done in finding filter pairs which
mal.
have a complementary set for the inverse transform ([81,
[28], [29]), it is far from exhaustive. The equivalence of
"For the orthonormal wavelets of compact support, the larger the value maximally flat filters (with equal order roots at 0 and T)
of N , the more rapid the asymptotic convergence of $(U) to zero; i . e . , as with Lagrange B trous filters as a design tool is perhaps
w + m [8]. This hints at a smaller bandwidth for $ and, hence, also for g,
for large N . However, the speed at which the $(w) fall off near w = 0 worthy of investigation. Many of our filter conditions on
appears to be fairly insensitive to N . energy are sufficient but possibly not necessary; a tight
SHENSA: DISCRETE WAVELET TRANSFORM 248 I

set of necessary and sufficient conditions for boundedness respectively. Also,

would be desirable. Finally, an investigation into the w:, = IT&. (B.3)
quality of the approximation of the DWT to the sampled
WT in the case where it is not exact could be fruitful. It In essentially all applications f is a low-pass filter and g
would perhaps lend more insight into the role of the reg- is high-pass. This rather vague qualification is quantified
ularity of the wavelet function in particular applications. below.
That is, in addition to the above definitions, it is ex-
APPENDIXA pedient to impose auxiliary conditions on the filters to en-
PROOFOF LEMMAS3.1 AND 3.2 sure a) that the DWT is related to some WT with a rea-
Lemma 3.1: sonably behaved scale function 4 ( t ) ; b) that the transform
have finite energy and be a bounded transformation; and,
[(A F)']nk = [(AF)'1O,k
- 21n. (A. 1) often, c) that it be invertible. The algebraic conditions for
Proof: For i = 1, we have invertibility are found in (6.2) and (6.4). At the time of
this paper no single set of necessary and sufficient con-
[AFIrrk =hn-k = [AF10.k-2n. (A. 2) ditions exist for the satisfaction of a) and b). Indeed, the
Then, by induction, definition of "reasonable behavior" of the scale function
ultimately depends on the application. In an attempt to
t(AF)'Iflk= f 2 n --m [(AF)'- ' l m k provide some degree of organization, we first list the can-
didate constraints, loosely labeled as low-pass, high-pass,
h n -m [ ( A F )-' ' 1 0 , k - 2 ' - Im or energy conditions. We then summarize their conse-
quences. If either of the filters is infinite it is assumed to
f-m [(A F)' - '10,k -21- lm -2ln satisfy the decay condition [8], [26]

C j - m t (A F - ' ~ mk ,- 2tn 3 E such that c Ifn[n e <

n
00. (B.4)
-
- [(AF)'lO,k- 2'n. (A.3) 1. Candidate Constraints
Lemma 3.2: i) Low pass

c [(hF)'],kejkW= eJ2lnW
i- 1
fZ(2'w). (A.4)
c f, = J 2
fl
r=O
(i.e., (1/2)fz(0) = 1).
Proof: For i = 1, we have ii) Energy
k [AF],eJk" = Ck f 2 f l - k e i k w ;lL(4125 1.
= eJ2""f,(w). (A.5) iii) Low pass
1
Then, by induction, = (1 + e'w>Nr(4
Jjfz'"' (B.7)
1-2
c [ ( A F ) ' l n k e J k=w c [ ( A F ) ] f l k e J k ( 2 ' - ' w )f,(2'W).
k k r=O
where ly(w)l 5 C < 1 / 2 ((B.7) implies thatf,(a) = 0).
iv) Energy: complementary low-padhigh-pass pair

APPENDIXB
Equation (B. 8) implies
SUMMARY OF FILTERCONSTRAINTS
The discrete wavelet transform 19' and the decimated a) high pass
discrete wavelet transform w' (or d') are defined for ar- Cg, = o (B.9)
bitrary filters f and g by n

sl+l = (D'f)* sf (B. la) (i.e. g z (0) = 0)

9

I9' = (D'g) * s f (B. 1b) b) low p a d h i g h pass

and ;If,(@l = 1 =$ sz(4 = 0 (B. loa)

si+l - A(f * si) (B.2a) gz(o) = 0 * ;[fi(o)l = 1. (B.lOb)
w' = g * SI (B.2b) v) Energy
di+t g Awi (B.2c) + Clgz(4l2) = 1.
;(If,(4l2 (B. 11)
2482 IEEE TRANSACTIONS ON SIGNAL PROCESSING. VOL. 40, NO. IO. OCTOBER 1992

2. Implications [ I I] D. Esteban and C . Galand, “Application of quadrature mirror filters

to split-band coding,” in Proc. ICASSP, May 1977, pp. 191-195.
Necessity for pointwise convergence of (4.1) to $(U): [I21 M. Vetterli, “A theory of multirate filter banks,” IEEE Trans.
(B.5). Acoust., Speech, Signal Processing, vol. 35, pp. 356-372, Mar. 1987.
(131 M. J . T . Smith and T . P. Barnwell, 111, “Exact reconstruction tech-
Sufficiency for (4.1) and (4.4) to :onverge in L ’ ( R ) , niques for tree-structured subband coders,” IEEE Trans. Acousf. ,
and L2(R) to continuous 4(t) and +(U), respectively: Speech, Signal Processing, vol. 34, pp. 434-441, June 1986.
(B.5)-(B.7). This is one of the central results of [8], which 1141 0. Hermann, “On the approximation problem in nonrecursive digital
filter design,” IEEE Trans. Circuir Theory. vol. 18, pp. 411-413,
also includes an examination of the decay of 4(t)and other May 1971.
regularity properties. Note that an important class of [I51 C . Heil and D. Walnut, “Continuous and discrete wavelet trans-
wavelets which does not fall under the domain of this forms,” SIAM Rev., vol. 31, pp. 628-666, Dec. 1989.
1161 A. Grossman and J . Morlet, “Decomposition of Hardy functions into
theorem is the Haar wavelets (2.28a). Pointwise conver- square integrable wavelets of constant shape,” SIAM J. Marh. Anal.,
gence still holds for the Haar wavelets, but they are not vol. 15, pp. 723-736, 1984.
continuous. [I71 J. Morlet, G . Arens, E. Fourgeau, and D. Giard, “Wave propagation
and sampling theory,” Geophysics. vol. 47, pp. 203-236, Feb. 1982.
Necessity for finite energy: (B .9). 1181 A. Grossmann, R. Kronland-Martinet, and J. Morlet, “Reading and
Sufficiency for finite energy and that the transformation understanding continuous wavelet transforms,” in Wavelets: Time-
be bounded: (B.6) and B < 03 in (B.8). Frequency Methods and Phase Space. Berlin: Springer, IPTI, 1989,
pp. 2-20.
Sufficiency for a bounded inverse: (B.6) and A > 0 in [ 19) I . Dautxhies, “Orthonormal bases of compactly supported wavelets
(B.8). 11: Variations on a theme,” preprint.
Necessity and sufficiency for energy conservation: [20] M. J . Shensa, “The discrete wavelet transform,” Naval Ocean Syst.
Center. Tech. Rep. 1426, June 1991.
(B. 11). [21] G . Gambardella, “A contribution to the theory of short-time spectral
analysis with nonuniform bandwidth filters,” IEEE Trans. Circuir
Theory, vol. 18, pp. 455-460, July 1971.
ACKNOWLEDGMENT [22] R. Altes, “Sonar for generalized target description and its similarity
The author would especially like to thank D. K . Le for to animal echolocation systems,” J . Acoust. Soc. A m e r . , vol. 59, pp.
97-206, Jan. 1976.
his meticulous reading of the manuscript and penetrating [23] M. Vetterli and C. Herley, “Wavelets and filter banks: Relationships
remarks. He is also grateful to D. Stein for several stim- and new results,” in Proc. ICASSP, Apr. 1990, pp. 1723-1726.
ulating discussions. Finally, he thanks an anonymous re- [24] A . Kumar and D. R. Fuhrmann, “The Frazier-Jawerth transform,”
viewer for greatly simplifying the proof of Theorem 5.2. in Proc. ICASSP, Apr. 1990, pp. 2483-2486.
1251 0. Rioul and P. Duhamel, “Fast algorithms for discrete and contin-
uous wavelet transforms,” IEEE Trans. Inform. Theory, Mar. 1992.
REFERENCES [26] S . Mallat, “Multiresolution approximations and wavelet orthonormal
bases o f L * ( R ) , ” Trans. Amer. Marh. Soc., vol. 315, ..
pp, 69-87, Sept.
[I] J. M. Combes, A. Grossman, and P. Tchamitchian, Eds.. Wavelets: 1989.
Time-Frequency Methods and Phase space, ~ ~ ~springer, l i ~IPTI,
: [27] R. Ansari, C. Guillemot. and J . F. Kaiser, “Wavelet construction
1989. using Lagrange halfband filters,” IEEE Trans. Circuits S y s f . , Sept.
[Z] A. Grossman, R. Kronland-Martinet, and J. Morlet, “Analysis of 1991.
sound patterns through wavelet transforms,” Inr. J. Parr. Recog. Ar- 1281 M. Vetterli and C. Herley, “Wavelets and filter banks: Theory and
tificial Infell., vol. I , pp. 97-126, Aug. 1987. design,” IEEE Trans. Signal Processing, vol. 40, no. 9 , Sept. 1992.
[3] I. Daubechies, “The wavelet transform, time-frequency localization, [29] A. Cohen, 1. Daubechies, and J.-C. Feauveau, ”Biorthogonal bases
and signal analysis,” IEEE Trans. Inform. Theory, vol. 36, pp. 961- of compactly supported wavelets,” in preparation.
1005, Sept. 1990. 1301 S . G . Mallat and S . Zhong, “Complete signal representation with
. . M. Holschneider, R. Kronland-Martinet, J. Morlet, and Ph. Tcham-
141 multiscale edges,” NYU Comput. Sci. Tech. Rep. 483, Dec. 1989.
itchian, “A real-time algorithm for signal analysis with the help of
the wavelet transform,” in Wavelets: Time-Frequency Merhods and
Phase Space. Berlin: Springer, IPTI, 1989, pp. 286-297.
151 P. Dutilleux, “An implementation of the algorithme a trous to com- Mark J. Shensa was born in Cleveland, OH, in
pute the wavelet transform,” in Wavelers: Time-Frequency Methods 1944. He received the B.A. degree in mathemat-
and Phase Space. Berlin: Springer, IPTI, 1989, pp. 298-304. ics from Princeton University in 1966, the M.A.
[6] S . G . Mallat, “A theory for multiresolution signal decomposition: degree in mathematics from the University of Cal-
The wavelet representation,” IEEE Trans. Putt. Anal. Machine In- ifornia, Berkeley. in 1969, and the Ph.D. degree
r e / / . , vol. 2 , pp. 674-693, July 1989. in electrical engineering from the University of
[7] S. G . Mallat, “Multifrequency channel decompositions of images and California, Berkeley, in 1971.
wavelet models, ” IEEE Trans. Acoust., Speech, Signal Processing, From 1971 to 1974, he was an Associate Pro-
vol. 37, pp. 2091-2110, Dec. 1989. fessor of Applied Mathematics at IME in Rio de
181 1. Daubechies, “Orthonormal bases of compactly supported wave- Janeiro, Brazil. In 1974 he joined Texas Instru-
lets,” Commun. Pure Appl. Marh., vol. 41, pp. 909-996, 1988. ments, Inc. Since 1980 he has been with the Naval
[91 1. Daubechies, “Orthonormal bases of wavelets with compact sup- Ocean Systems Center in San Diego, CA. He served as an Editor of the
port-connection with discrete filters, in Wavelers: Time-Frequency
” IEEE TRANSACTIONS O N ACOUSTICS, SPEECH, A N D SIGNAL
PROCESSING from
Methods and Phase Space. Berlin: Springer, IPTI, 1989, pp. 38- 1981 to 1984. He has made numerous contributions to technical journals
66. in the last twenty years. This work includes singular perturbation theory,
[IO] Y. Meyer, “Ondelettes, fonction splines, et analyses graduees,” adaptive signal processing, tracking and localization, and Boolean logic.
Univ. Torino, 1986. He is currently interested in wavelet transforms.