Filter Banks Using Wavelets

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL.
40, NO 9, SEPTEMBER 1992 2207
Wavelets and Filter Banks: Theory and Design

Martin Vetterli, Senior Member, IEEE, and Cormac Herley
Abstract-Wavelets, filter banks, and multiresolntion signal as a bandpass filter. Fine temporal analysis is done with
analysis, which have been used independently in 1 he fields of contracted (high-frequency) versions of the wavelet, while
applied mathematics, signal processing, and computer vision,
respectively, have recently converged to form a single theory. fine frequency analysis uses dilated (low-frequency) ver-
In this paper, we first compare the wavelet transform with the sions. The bandpass filters have thus constant relative
more classical short-time Fourier transform approaich to signal bandwidth or “constant-Q. The importance of constant
”
analysis. Then we explore the relations between wavelets, filter relative bandwidth when perceptual processes like the au-
banks, and multiresolution signal processing. We briefly re- ditory system are involved has long been recognized; for
view perfect reconstruction filter banks, which can be used both
for computing the discrete wavelet transform, and for deriving example, the musical scale introduced by Bach is expo-
continuous wavelet bases, provided that the filters meet a con- nentially spaced, and subband coding of speech typically
straint known as regularity. Given a low-pass filter, we derive uses an octave-band splitting of signals. The multireso-
necessary and sufficient conditions for the existence of a com- lution view of signals has been used in computer vision
plementary high-pass filter that will permit perfect reconstruc- for tasks like segmentation and object recognition.
tion. Posing the perfect reconstruction condition as a Bezout
identity, we then show how it is possible to find all higher de- From the above discussion, it is clear that several fields
gree complementary filters based on an analogy with the theory have developed similar ideas independently, and it is only
of diophantineequations. An alternative approach based on the recently that the connections have been fully recognized.
theory of continued fractions is also given. We use lhese results The activity in wavelets was initiated by Morlet’s work
to design highly regular filter banks, which generate biorthog- in geophysical signal processing [2]. A strong mathemat-
onal continuous wavelet bases with symmetries.
ical framework was built around the basic wavelet idea by
the “French school” [3], [4] and is documented in the
I. INTRODUCTION recent book by Meyer [ 5 ] , which also shows the connec-
tions to earlier results in operator theory. Tutorials on
All this time the guard was looking at her, first Ihrough a
telescope, then through a microscope, and then through an
wavelets are available in [6], [l], [7].
opera glass. Multiresolution approaches have been popular for com-
Lewis Carroll, Through the Looking Glass puter vision problems from range detection to motion es-
timation [8]. An important application to image coding
called a pyramid [9] is closely related both to subband
T HE analysis of nonstationary signals often involves a
compromise between how well transitions or discon-
tinuities can be located, and how finely long-trmi behav-
coding and to wavelets. Mallat used this concept of mul-
tiresolution analysis to define wavelets [lo], [ l l], [6], and
Daubechies constructed compactly supported orthonor-
ior can be identified. A typical example is the choice of mal wavelets based on iterations of discrete filters [12].
window length in the short-time Fourier transform. In The relation of these filters to classical maximally flat de-
wavelet analysis one looks at the signal at different signs [ 131 was recently noted by Shensa [ 141. In the signal
“scales” or “resolutions” : a rough approximation of the processing literature, work on filter banks goes back to
signal might look stationary, while at a detailzd level subband coding of speech [15], [16]. Orthogonal filter
(when using a small window) discontinuities become ap- banks were first derived by Smith and Barnwell [I71 and
parent. This multiresolution, or multiscale view of signal Mintzer [18], and were systematically studied by Vai-
analysis is the essence of the wavelet transfc’rm, which dyanathan [ 191, [20]. The biorthogonal case, especially
has recently become quite popular [l]. The goal is “to the linear phase case, was also studied [21]-[23].
see the wood and the trees.” In what follows we start by presenting a synthetic view
The wavelet analysis is performed using a single proof these diverse techniques, both as an introduction to the
totype function called a wavelet, which can be thought of various tools, and as a motivation for further develop-
ments. Section I1 introduces the wavelet transform and
Manuscript received August 24, 1990; revised June 25, 1391. The work compares it with the short-time Fourier transform in the
of M. Vetterli was supported in part by the National Science Foundation continuous time case. The discrete time case and its con-
under Grants ECD-88-11111, MIP-88-08277, and MIP-90-14189. The
work of C. Herley was supported in part by the National Science Foun- nection to the continuous case is then explored. Section
dation under Grant ECD-88- 11 11 1. 111 gives a brief overview of filter bank results in the or-
The authors are with the Department of Electrical Enginexing and Cen- thogonal and biorthogonal cases; and constructs biorthog-
ter for Telecommunications Research, Columbia Universlty, Vew York,
NY 10027-6699. onal wavelets as the limit of iterated regular filter banks
IEEE Log Number 9201587. (regularity means that the iterated filter converges to a
18353-587X/92$03.00 0 1992 IEEE
I
1
2208 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, NO. 9, SEPTEMBER 1992
continuous function). In Section IV, results on the alge-

braic structure of the FIR perfect reconstruction problem
are presented, and these will be used to construct new
wavelets. Section V presents design results of compactly
supported biorthogonal symmetric wavelets based on it-
erated filter banks.
A certain amount of review material has been included.
First, since much of the wavelet material may be unfa-
miliar to readers of this TRANSACTIONS we felt the review
might be of interest in its own right; second, this wavelet
background greatly assists the presentation of the original
material that follows. The material of Sections 11-D, I11
(from 111-C onward), and IV and V is novel. SOn e of the
main results of this paper were presented in cmdensed
form in [24], [25].
The focus on wavelets with symmetries is motivated
both by the fact that the problem was open, and for the
reason that in applications like image processing linear
phase is often desired [26]. Previous linear phase designs
were not concerned with regularity [23], [22]; the empha-
sis was on equal length filters with a lattice factorization.
This work specifically addresses the regularity question,
and proposes new techniques to produce regular designs.
The importance of regularity for signal processing appli-
cations is still an open question. While it is an elegant
mathematical result on the relation between continuous
and discrete systems, most filters used in signal process-
ing applications are “almost regular,” when a fiiite num-
ber of iterations is involved. Settling this question will of
(b)
course require substantial experimental evidence, and is
Fig. 1 . Continuous wavelet transform plots. The signal is shown on top;
beyond the scope of this paper. the axes are time horizontally, and log scale vertically. (a) Sineburst ana-
To give an idea of the effect of phase, we show In Fig. lyzed using symmetric wavelet. (b) Sineburst analyzed using asymmetric
1 the continuous wavelet transform plots of a burst of a wavelet.
sinewave analyzed with a) a symmetric wavelet designed
in Section V-A, and b) an asymmetric wavelet designed by R ( R being the set of positive reals), the set of inte-
+
in [12]. The horizontal axis represents time, .while the gers is Z . The inner product over the space of square-
vertical represents the log of scale (scale being the inverse summable sequences 12(2) is
of frequency). Scale increases from top to bottom; and the m
range is eight octaves for each of the plots. In other words,
the scale of the analyzing wavelet is smallest at he top of ( a ( n ) ,b(n)) = c
n = -m
a*(n)b(n)
the plot; and the largest scale is 28 times the smallest.
Note that unlike the Morlet wavelet used in [27 1, both of where a ( n ) , b(n) E 12(Z),and the asterisk * denotes com-
these wavelets are real, so magnitude and phase infor- plex conjugation. We define Ila(n)II; = ( a ( n ) ,a ( n ) ) ,and
mation do not have to be separately computed. For signal Ila(n)II = max, a(n). Similarly, over the space of square-
analysis, unlike coding, it is not necessary to use perfect integrable functions L2(R)we have the inner product:
reconstruction systems, since the representation is gen-
erally highly oversampled, and thus the continucus wave-
let transform example shown is for illustration purposes
( f ( x > g, ( x ) ) = jm
-m
f*(x)g(x) dr
only. wheref(x), g(x) E L2(R).The norm is given by Ilf(x)ll;

During the completion of this work, we becaine aware >
= ( f (x), f (x) . The z transform of a sequence is defined
of results developed independently by other researchers. by H(z) = Er= - m k ( n ) z Ln. The time reversed version of
Cohen et al.’s work on biorthogonal bases of wavelets a sequence which is nonzero for n = 0, 1 , * ,L - 1
[28], while of a more mathematical nature, leac s to very is fi(n) = h(L - 1 - n). We shall use the notation {h(n)}
similar designs. A recent manuscript by Rioul [29]. while = [k(O),h ( l ) , * . , h(L - l ) ] when we want to indicate
mainly focused on the orthogonal case, contains also some the coefficients of an FIR filter. Note that we will consider
results on biorthogonal systems, as well as the connection only filters with real coefficients, unless otherwise speci-
with filter banks. fied.
Notation: The set of real numbers will be rexesented Matrix and vector quantities will be denoted by bold-
I
VETTERLI AND HERLEY: WAVELETS AND FILTER BANKS 2209
face symbols; the asterisk * will denote Hermitian trans-

position, which, since we will consider only matrices with
real coefficients, is equivalent to ordinary transposition.
11. WAVELETS, MULTIRESOLUTION SIGNAL PROCESSING,

AND FILTERBANKS
A. The Wavelet Transform
Analysis of signals using appropriate basi j. functions
goes back at least as far as Fourier who used complex
sinusoids. The Fourier transform of a continuous time
signal x ( t ) is X,(w) = ( e j w f x, ( t ) > . A difficulty that has
often been cited with this approach is that, because of the
infinite extent of the basis functions, any time-local in-
formation (e.g., an abrupt change in the signal) is spread frequency
out over the whole frequency axis. Gabor addressed this
problem by introducing windowed complex sinusoids as
basis functions [30]. This leads to the doubly indexed
l-"-
windowed Fourier transform:
x W F ( w 7)
, = e-JW'w(t - 7)x(t)tit (1)
-m
where w ( * )is an appropriate window like a Gaussian.

I
That is, XWF(w,7) is the Fourier transform of x ( t ) win- time
dowed with w ( . ) shifted by 7 . Equivalently, the basis
functions are modulated versions of the window function (b)
(see Fig. 2(a)) [31]. The major advantage of the win- Fig. 2. Basis functions and time frequency resolution of the short-time
Fourier transform (STFT). (a) Basis functions. (b) Coverage of time-fre-
dowed or short-time Fourier transform (STFT I is that if a quency plane.
signal has most of its energy in a given time ini.erval [ - T,
TI and frequency interval [-Cl, a ] , then its STFT will be
localized in the region [-T, TI X [ - a , a ] , and will be is a low frequency function, while for small a , the basis
close to zero in time and frequency intervals nhere the function becomes a contracted wavelet, that is a short high
signal has little energy. A limitation of the S'I'FT is that, frequency function (see Fig. 3(a)).
because a single window is used for all frequencies, the The wavelet transform (WT) is defined as
resolution of the analysis is the same at all locations in
the time-frequency plane (see Fig. 2(b)).
Of course, the uncertainty principle excludes the pos-
X,(a, b) = y=
1
.\la J - m
Cm
h*
\
(-)t -a b / dt. (3)
sibility of having arbitrarily high resolution in both time The time-frequency resolution of the WT involves a dif-
and frequency, since it lower bounds the time-bandwidth ferent tradeoff to the one used by the STFT: at high fre-
product of possible basis functions by A T * Afl L (1 /47r), quencies the WT is sharper in time, while at low frequen-
where (AT)* and (Ais2)2 are the variances of the absolute cies, the WT is sharper in frequency (see Fig. 3(b)). The
values of the function and its Fourier transfom, respec- middle functions depicted in Figs. 2(a) and 3(a) are iden-
tively 1321. However, by varying the window used, one tical, and hence the time-frequency resolutions of the two
can trade resolution in time for resolution in frequency. methods are the same at that frequency.
In order to isolate discontinuities in signals one' would like Obviously, both the STFT in (1) and the WT in (3) are
to have some basis functions which are very :hart, while highly redundant when the parameters ( U ,7) and ( a , b )
some long ones are required to obtain fine frequency anal- are continuous. Therefore the transforms are usually eval-
ysis. An intuitively appealing way to achieve this is to uated on a discrete grid on the time-frequency and time-
have short high frequency basis functions, arid long low scale plane, respectively, corresponding to a discrete set
frequency ones. This is exactly what is achieLed with the of continuous basis functions. The question arises as to
wavelet transform, where the basis functions s re obtained whether there is a grid such that the set of basis functions
from a single prototype wavelet by translaticn m d dila- constitutes an orthonormal basis; which of course implies
tion/contraction [ 121, [ 111, 131, [4] : that there is no redundancy. Unfortunately, for the STFT,
this can happen only if w ( . ) is badly localized in either
time or frequency 1331, which is the reason that the STFT
is usually "oversampled" (a redundant set of points is
where a E R +,b E R . For large a , the basis function be- used), so that better behaved window functions can be
comes a stretched version of the prototype wavzlet, that used. In the wavelet transform case, however, it is pos-
I
I
2210 I E E E T R A N S A C T I O N S ON S I G N A L P R O C E S S I N G , VOL. 40, N O . 9, S E P T E M B E R 1992
uous, but is of interest because of its simplicity), where:

1 O I t < 1/2
h(t) = -1 1/2 5 t <1
[o otherwise.
The orthonormality is easily verified since at a given
scale, translates are nonoverlapping, and because of the
scale change by 2, the basis functions are orthogonal
across scale. The Haar basis is shown in Fig. 4(a). How-
ever, the Haar function is discontinuous, and is not gen-
erally appropriate for signal processing. A continuous set
of basis functions is given in Fig. 4(b). These functions
frequency are obtained from a compactly supported wavelet con-
4 structed by Daubechies [12], using a length 4 FIR filter.
Longer filters lead to smoother functions. It is interesting
to note that the translates and dilates of the functions in
both Figs. 4(a) and (b) form orthonormal bases for L2(R)
functions. The purpose of the present paper is to design
other continuous wavelets, having additional properties
like linear phase.
+ B. Multiresolution Signal Processing
time
From a signal processing point of view, a wavelet is a
Fig. 3 . Basis functions and time-frequency resolution of the viavelet trans- bandpass filter. In the dyadic case given in ( 5 ) , it is ac-
form (WT). (a) Basis functions. (b) Coverage of time-frequmcy plane. tually an octave band filter. Therefore the wavelet trans-
form can be interpreted as a constant-Q filtering with a set
of octave-band filters, followed by sampling at the re-
sible to design “nice” functions h ( . ) such that the set of spective Nyquist frequencies (corresponding to the band-
translated and scaled versions of h( forms an orthonor-
e ) width of the particular octave band). It is thus clear that
mal basis. By “nice” we mean that the function should by adding higher octave bands, one adds details, or res-
be at least continuous, perhaps with continuous deriva- olution, to the signal. Mallat [lo], [ I l l and Meyer [5]
tives also. Let us discretize the translation and dilation/ introduced the concept of multiresolution analysis and
contraction parameters of the wavelet in (2): used it to construct orthonormal bases of wavelets. This
multiresolution view can be interpreted as a successive
h,,(t) = a i m / 2 h(aomt - nbo), approximation procedure.
We will give a simple but intuitive explanation of mul-
m, n € 2 , a0 > 1, bo # 0
tiresolution and successive approximation. Call V0 the
which corresponds to a = a t and b = natbO.Note that space of all band-limited functions with frequencies in the
the translation step depends on the dilation, since long interval (-a, a). Then the set of functions
wavelets are advanced by large steps, and shoit ones by sin (n(x - k ) )
small steps. On this discrete grid, the wavelet transform #(x - k ) = sinc (x - k) = kEZ
T(X - k )
is thus
(6)
forms an orthonormal basis for VO.Similarly, call VPI the
space of band-limited functions with frequencies in the
Of particular interest is the discretization on a dyadic grid, interval ( - 2 a , 2a). Clearly, the set & . sinc (2x - k ) ,
which occurs for a. = 2, bo = 1. It is possible to construct k E Z is an orthonormal basis for K 1 . Also,
functions h( so that the set
e )
V0 c V-1. (7)
In particular, if x ( t ) E VO,then x(2t) E V - I . Now, call WO
m, n E Z, a0 = 2, bo = 1 (5) the space of bandpass functions with frequencies in the
interval (-2n, -n) U (a,2n). Then
v-,= V, 0 w, (8)
that is, WOis the orthogonal complement in V - , of VO.In
A classic example is the Haar basis (which is not contin- other words, V - I is equivalent to V0 plus some added de-
l
I
VETTERLI AND HERLEY: WAVELETS AND FILTER BANKS 221 I
-1.5'
-0.5 0 0.5 I 1.5
, -. 2
(a)
Fig. 4. Orthogonal system of scaled and translatrd wavelets (two scales only are shown). (a) Haar wavelet. (b) Daubechies's
wabelet based on a length-4 regular filter.
tail corresponding to WO.For completeness C I X (nx) has That is, +(x) can be written as
to be included in V,, and sin (nx)in WO.From the above
it is clear, by scaling, that if Vi is the space of band-lim-
ited functions with frequencies in the interval (-2-'n,
2-'n), then we get relations similar to (7) and (8): since it is the interpolation, by 4 ( 2 x ) , of the perfect half-
band low-pass filter. Note that 4 ( x ) and c, are both sym-
v ; c K-1 iEZ (9) metric. +(x) is called a scalingfunction because it derives
Vj-1 = v;0 W; i E z (10) an approximation in Vo of signals in I/- I . In V - the or-
thogonal complement WOto V, is given by the halfband
where W iis the space of bandpass functions w th frequen- high-pass signals. In the sampled domain, this is given by
cies in the interval ( - 2 - ' + ' n , -2-'n) U ( 2 - ' ~ T
, i +a).
' the halfband low-pass (12) modulated by ( - l ) , , and
Moreover, we have: shifted by one (to include sin ( n x ) ) .Thus, $(x) is the in-
terpolation thereof, that is,
*
v, c v1 c v, c v-1 c v-2 * * *
and, by iterating (10) $(x) = C

n = -m
( - l ) , ~ - , +14(2x - n). (13)
v;= W.+l 0 W + 2 0 K+3 0 * - (1 1)
Note that since the c,'s are symmetric we can reverse the
and, finally, the direct sum of all Wj's,j = i + 1, * . , sign as above, which we do for later convenience. Now
9
03, is equivalent to the space of square integrable func-

tions band limited to (-2-'+'a, 0) U (0, - 2 - ' + ' n ) . 4(x - k ) 1 $@ - k ) (14)
Let us now construct the wavelet that will span WO. since they cover disjoint regions of the spectrum. Also
First the set {4(x - k ) , k E Z } given by (6) constitutes a
<$(x - 4 , $(x - 0)= &,
basis for V,. Thus { h4(2x - k ) , k E Z} constitutes a
basis for V P I . Now, in the sampled version of K1, 4 ( x ) because the translates of 4(2x - n) in (13) are even, and
is given by the perfect halfband low-pass filtlx with im- thus the sign change is cancelled. Then the inner product
pulse response: reduces to (+(x - k ) , 4(x - 1 ) ) = It can be shown
that the $(x)'s span WO,and therefore, $(x) and its integer
sin (nn/2)
Jzc, = = discrete halfband filter. (12) translates form an orthonormal basis for WO.Thus, the
an /2 wavelet for this bandpass example is given by $(x).
I
1
Pictorially the situation is as shown in Fig. 5 . Fig. 5(a)

shows the imbrication of VI C Vo C V - , , and [b) shows
the added bandpass spaces W,. While the above example
may seem artificial, and leads to a wavelet $(XI which is
of infinite extent and has slow decay, the situatiam is con-
ceptually the same for all orthonormal wavelet bases [ 12 ].
Fig. 6 shows the corresponding division of the spectrum 7d4 7d2 x 2rr
for a wavelet that has compact support. In particular, (a)
given an orthonormal basis for Vo made up of @x) and its
integer translates, then, we can find coefficients c, such
that
m
because Vo C V - , . Then 7tl4 rdz rr 2rr

W
(b)
$(x) = fl=
c
-m
(-l)"C-,+,cb(2X - n) (16) Fig. 5 . Ideal division of spectrum using sinc filters. (a) Division into V,
,
spaces. Note that V, C V, . (b) Division into W,spaces.
~
and its integer translates form an orthonormal basis for

w,. SO { ~ ~ , (=x )2'/2+(2'x - j >,i , j E Z } con.;'titutes an space spanned by the rows of Ho is given by
orthonormal basis for L2(R), following (1 1). The impli-
cation of (15) is given graphically in Fig. 7 where it is H;H, - X.
shown how a scaling function can be obtained from a lin-
Note that multiplication by H; corresponds to upsampling
ear combination of its scaled versions. The first example
by 2 followed by convolution with a filter having impulse
is for illustration purposes only, while tht: second is an
actual scaling function from a wavelet basis.
response &(n) = [ h o ( - ~ I), h o ( ~- 2 ) , * * , h o ( l ) ,
hO(0)],(that is the time reversed impulse response of
ho(n)).Note also that in order for the set {ho(n - 2k), k
C. The Discrete Time Case E Z } to form an orthonormal basis L has to be even. For
Assume now that we deal with discrete time sequences if L were odd, then (ho(n),ho(n - L 1)) = ho(0)ho(L +
x ( n ) , n E Z which are square summable; that is, the space - 1) # 0 unless either ho(0)or ho(L - 1) is zero.
12(Z).Now, let us derive a coarse half-resolution approx- Call VP1 the space 12(Z) and call Vo the subspace of
imation to the original sequence. This can be done by fil- V-I spanned by the rows of Ho. Then call WOthe orthog-
tering with a halfband low-pass filter, followed by sub- onal complement of Vo in V - I :
sampling by 2 (discarding every odd-indexed sample). In
matrix notation, and assuming for the sake of simplicity V-, = V, 0 WO.
that the filter is FIR, the convolution with a filler having
-
impulse response [ho(0),ho(l), . , ho(L - 1) I followed Now the filter with impulse response h,(n) = (- 1)"
ho(L - 1 - n ) and its even shifted versions form an or-
by subsampling by 2 , corresponds to matrix multiplica-
tion of the infinite signal vector X[ * x( - l ) , ,c(O), x( l ) , thonormal basis for WO.First note that orthogonality of
-*-]by ho(n)and h l ( n ) with respect to even shifts is easily veri-
Assume further that the impulse response and its shifted fied, because of the sign change in hl(n):
versions by even shifts (i.e., the rows of the above ma-
trix) form an orthonormal set, that is, (h,(n - 21), ho(n - 2 k ) ) = 0 k, l E Z. (19)
(h& - 21), ho(n - 2k)) = Akl k , 1 E' 2. (18) In matrix notation calling H I a matrix based on hl(n) in
the same way that Ho in (17) was based on ho(n)we have
In matrix notation the equivalent of (18) is
HoH,* = I. H ~ H := 0
The projection of the original sequence x ( n ) onto the sub- thus V, I WO. Then from the orthonormality of
I
i . 2 ~
-0.4l ' I
0 5 IO 15 20 25 30 35 D
A
.. ---. '7--.
C.
.
0 50 100 I50 2w 250 300 350 400 450 MO
Fig. 6. Division of spectrum with a real watelei. (a) Dyadic stretches of the scaling function. (b) Fourier transform of the
stretched scaling functions, indicating the nesting of the V, spaces. (c) Dyadic stretches of the wavelet. (d) Fourier transform of
the stretched wavelets, indicating the arrangement of the W, spaces.
I
'5.5 0 0.5 1 1.5 2 25 3
Fig 7 Scaling function d(x) as a linear combination of scaled and shifted versions 4(2x - n ) (a) Hat function example (this
is not an orthonormal example) (b) Example based on a regular orthogonal 4-tap filter from 1121
~ ~~ ~~
ho(n - 21) it follows

( h l ( n - 21), hI(n - 2 k ) ) = ~5kl k, 1 E Z
since the sign change in h l ( n )is cancelled. In other words,
based on an orthonormal basis for Vo, we were able to
construct One for Just as in the continuous Again Fig. 8. Decomposition of V - , into Vo and WOusing multirate filters, and
in matrix notation recombination to achieve perfect reconstruction. H,*Ho . x i s the projection
of the signal x onto Vo, and H T H , . x is the projection onto WO.
H,H: = z (20)
and the projection of the original sequence onto WO is compose Vo into VI and W , and so on. Hence we find
given by
7 c 7-1 j = 0,1, * * -
HTH, * x.
5-1 = 7 0 Wj j = 0,1, ’ * *
Note that since we have projections onto orthorrormal and from which we have
complete subspaces:
.* v, c v, c v, c v-1
H;H, + H : H ~= I . *
and
It is also clear that by interleaving Ho and H I , we obtain
a block Toeplitz orthonormal matrix T: V-,=WoowloW,s.**.
. hl(L -
T =
0
0
-
which satisfies That is, the direct sum of all the W,’s is the space of square
TT* = T*T = I summable functions 12(2). Note that in contrast to the
continuous case there is a “maximum” resolution in the
that is, the two filter impulse responses ho(n)and h l ( n ) , discrete case given by V - which is determined by the
together with their even translates, form an orthonormal original sampling rate.
basis for Z2(2).Because this is such a fundarr,ental con- The decomposition of V - I into WO,W1, W2, etc, is es-
cept we will illustrate it with an example. Insiead of the sentially a wavelet transform on discrete sequences, since
rather trivial case where ho(n) = 1 / & * [ 1, I I and h l ( n ) it splits the original space in two, and then splits one of
= 1 /& * [ 1 , - 11 (which corresponds to the Walsh-Had- the resulting half spaces in two, etc. This is shown in Fig.
amard transform on successive blocks of 2 samples), we 9 using filters and subsamplers. Actually, if the filter
will choose ho(n) = [ a o a l ,a o b l ,bobl, -boa,] where a, ho(n) is the ideal halfband low-pass filter given by (12),
= cos CY, and b, = sin a,.Because a; +
b’ = 1 it is clear then hl(n) is the ideal halfband high-pass filter. Thus if
+
that Ilho(n)l12 = 1 . Also (ho(n 2), h o ( n ) ) = 0 and thus I/- I is the space of functions band limited to ( - 2 a , 2a)
ho(n)and its even translates form an orthonormz 1 set. Now as considered in the discussion of multiresolution analysis
choosing h l ( n ) = [ b o a l ,bobl, -aobl, a o a l ]makes h l ( n in Section 11-B, then Voand WOare the spaces of functions
+ 2 1) orthogonal to ho(n)while keeping it of n o m 1, and band limited to (-a, a) and ( - 2 a , - T ) U (a,2n), re-
orthogonal to its even translates. Thus the set (ho(n + spectively. That is, by iteration, the discrete system in
+
21), h l ( n 2 k ) ; I, k E Z } is an orthonormal basis for Fig. 9 computes exactly the discrete wavelet transform
12(Z). into octave bands (although with noncausal infinite im-
In Fig. 8, we recapitulate the above relations in the pulse response filters).
usual digital signal processing notation, using filters and Note that the two concepts of scale and resolution in-
sampling rate changes to denote the operators used so far. troduced above, while related, are not the same. The no-
First the inner products Ho * x and HI x are calculated. tion of resolution of a signal is related to its bandwidth.
Then the projections onto Vo and WOare found. Finally, This holds also in the sampled domain, but it is best
the original signal is reconstructed by adding rhe projec- thought of as the bandwidth of the equivalent continuous
tions from the two orthogonal subspaces. time signal. This definition indicates that an oversampled
What worked once will work again; that is we can de- version will not have more resolution than a critically
I
VETTERLI AND HERLEY: WAVELETS AND FILTER BANkS 2215
x q E l - y ( n ) resolution: halved
scale: unchanged
2 - hl(n)
(a)
2 - hl(n) -@-
2 - -@--
2 X q e I 4 ) resolution: halved
2 - ho(n) -@- _._.._. scale: doubled
resolution: unchanged
scale: halved
(C)
sampled version of the same signal. The notion of scale Fig. 10. Resolution and scale changes in discrete time (by factors of 2).
Note that the scale of signals is defined as in geographical maps. (a) Half-
is related to the size of a signal. We will adhere to the band low-pass filtering reduces the resolution by 2 (scale is unchanged).
convention used in the wavelet literature, that large scale (b) Halfband low-pass filtering followed by subsampling by 2 doubles the
denotes contraction of the signal, while small scale stands scale (and halves the resolution as in (a)). (c) Upsampling by 2 followed
for a dilated signal. Fig. 10 shows various multirate sig- by halfband low-pass filtering halves the scale (resolution is unchanged).
nal processing operations, and their effect upon resolution

and scale (for simplicity only changes by factors of two low pass
are considered). I
D. Orthogonal Pyramids and Critical Sampling
qriginal difference
In computer vision and image coding [8], [9], a suc- signal signal
cessive approximation or multiresolution technique called
an image pyramid is sometimes used. It consists of deriv- low pass
ing a low resolution version of the original, then predict-
ing the original based on the coarse version, and finally 7 verS1on
taking the difference between the original and the predic-

o,nginal
tion (see Fig. 1l(a)). At the reconstruction, the prediction signal
is added back to the difference, guaranteeing perfect re-
construction. A shortcoming of this scheme is the over- (b)
sampling, since we end up with a low resolution version Fig. 11. Pyramid scheme involving a coarse low-pass approximation, and
a difference between the coarse approximation and the original. (a) Over-
and a full resolution difference (at the initial rate). We sampled case. (b) Critically sampled case, which is equivalent to a subband
briefly show below that, if the system is linear and the coding scheme.
low-pass filter is orthogonal to its even translates, then
one can actually subsample the difference signal after fil-
tering it. In that case, the pyramid reduces exactly to a
critically subsampled orthogonal subband coding scheme
like the one discussed in the previous subsection (a)
First, the prediction of the original based on the coarse
version is simply the projection onto the space spanned
by {ho(-n - 21), I E Z}. That is, calling the prediction
-
4 H(z) H(z) w..
{Fm
(b)
...
X:
H(z) H(z2) .... H(z2'-')
x = H;H,. X.
The difference signal is thus (c)

Fig. 12. Derivation of the equivalent iterated filter. (a) Subsampling by 2
d = (I - H,*Ho) * X. before a filter H(z) can be written as filtering by H ( z z ) followed by subsam-
pling. (b) Cascade of i filters each followed by subsampling by 2. (c)
But, because it is a perfect reconstruction system Equivalent filter, followed by subsampling by 2'.
I - H;Ho = HFHI
that is, d is the projection onto the space spanned by The signal d can be reconstructed by upsampling by 2
{ h i ( - n - 2 1 ) , I E Z}. Therefore, we can filter and sub- and filtering with fil(n).Then we have
sample d by 2, since, following (20)
HF(HIHTHl) x = HTH, x = d
HIH;Hl = H I .
and this, added to X = H; Ho x is indeed equal to x. In-
In that case, the redundancy of d is removed (d is now the notation of the multiresolution scheme described in
critically sampled), and the pyramid is equivalent to an Section 11-C the prediction X is the projection onto the
orthogonal subband coding system (see Fig. 1 l(b)). space Vo, and d is the projection onto WO.We have thus
I
~1
2216 IEEE TRANSACTIONS ON SIGNAL PROCESSING. VOL. 40. NO. 9, SEPTEMBER 1992
shown that pyramidal schemes can be critically sampled tering with H ( z ) is equivalent to filtering with H ( z 2 ) fol-
as well. A similar analysis is given in [29].It s interest- lowed by subsampling by 2 (see Fig. 12(a)).
ing to note that although pyramid coding schcmes have Therefore the cascade of i blocks of filtering operations
been in use for a decade or so, it had not been xeviously followed by subsampling by 2 is equivalent to a filter
noted that they could be subsampled under the conditions H'"(z) with z transform:
above. The discussion of the more general, biorthogonal I - I
case (see also Section 111-C) is given in Appendix .4.1 and ~ ' " ( z )= II
/=0
~ ( z ~ ' i) = 1, 2 * * (27)
~41.
followed by subsampling by 2'. We define H'''(z) = 1.
E. FIR Filter Banks and Compactly Supported Wavelets Assuming that the filter H ( z ) has an impulse response of
We now briefly show the connection betweei wavelets length L, the length of the filter H'"(z) is L"' = (2' -
of finite length and filter banks, as originally irvestigated 1) ( L - 1) -t 1 as can be checked from (27). Of course as
by Daubechies [12]. Note that it is common to use the i -+ 03 we get L'" 03. Now instead of considering the
-+
term compact support instead of finite length. First, as- discrete time filter, we are going to consider the function
sume we have an orthonormal basis of such functions 4(x) f " ) ( x ) which is piecewise constant on intervals of length
and $(x) which obey two-scale difference equations as in 1 / 2 ' , and equal to
(15) and (16): f"'(x) = 2 f / 2 h'"(n) n/2' Ix < (n + 1>/2'.
(4(x + 11, 4(x + 4 ) = 4 1 (21) (28)
($(x + 0,$@ + k)) = h, (22) Clearly, f'"(x) is supported on the interval [0, L - 11.
Note that the normalization by 2'12 ensures that if
(4(x + 0 , $(x + 4 ) = 0. (23) C (h"'(n))2= 1 then j ( f " ' ( ~ )dx) ~= 1 as well. Also, it
We will show these relations lead to a perfect Ieconstruc- can be checked that Ilh(')l12= 1 when \lh(*-')l12 = 1. A
tion FIR filter bank. The finite support of 4(x) means that fundamental question is to find out whether and to what
it can be written as a finite linear combination of the terms the functionf(')(x) converges as i + cx,. Fig. 13 shows
4 ( 2 x - n ) ; that is, finitely many of the c, are different two examples of such iterations. In Fig. 13(a) the first six
from zero. From (21) we get iterates of the filter with impulse response [1, 3, 3, 11
show that it converges rapidly to a continuous function;
( 4 ( 2 x - I), 4(2x - k)) = ;
6Ll. (24) while in Fig. 13(b) the iterates of the filter [ - 1 , 3, 3, - 1J
tend to a discontinuous function. In other words, different
Now, using (15) and (24),(21) can be written as (with n'
filters exhibit very different behavior. Of course when
= n - 21, and rn' = m - 2k):
constructing wavelets of compact support one would like
(4(x + 0 , 4(x + k ) ) them to be continuous functions, perhaps possessing con-
= (T c,4(2x + 21 - n), c c,,,4(2x + 2k
m
- m)
tinuous derivatives also. This can be achieved if the filter
meets certain regularity constraints; so that the limit func-
tion f'"'(x) is continuous. In [ 121 Daubechies gives such
a condition, which we now review.
First, assume that the filter H ( z ) has a zero at the half
cn'+21cn'+2k= ak/ sampling frequency, or H(eJ")= 0. This is not unreason-
able since H(z) is to be a halfband low-pass filter; in fact
from which it follows that IIc,II = h. iff'"(x), i + 03 is to converge to a continuous function
In other words the discrete filter, with impulr,e response at least one such zero is necessary [34], [ 3 5 ] . This to-
ho(n) = c , / h is orthogonal to its even translates, and gether with the fact that the filter impulse response is or-
with hi(n) = ( - l)"ho(L - n - 1) we obtain an orthogonal thogonal to its even translates is equivalent to Ch(n) =
perfect reconstruction FIR filter bank with orthogonal im-
pulse responses. Thus, compactly supported wavelet bases
H(1) = A. Define mo(z) = 1 / h * H(z), that is mo(l)
= 1. Now factor mo(z) into its roots at 7r (there is at least
lead to perfect reconstruction FIR filter banks. While the one by assumption) and a remainder polynomial R(z), in
converse does not always hold, and is not as immediate the following way:
to analyze, we discuss it here because it is the basis for
mdz) = [(1 + z-')/21NR(z).
the construction of compactly supported wal elets [ 121.
Considering the discrete time wavelet transform in Fig. Note that R(1) = 1 from the definitions. Now call B the
9 , one notices that the lower branch is an infinite cascade supremum of 1R(z)l on the unit circle:
of filters ho(n)followed by subsampling by 2 . Note that B = sup IR(e'")l.
subsampling the signal with z-transform X ( z ) by 2 results W€[O.2"]
in a new signal with z-transform Y(z): Then the following result from [12]holds.
Proposition 2.1 1121: I f B < 2N- I, then the piecewise
Y(z) = 1 / 2 - [x(z'/2)
+ x(--2i/2)J (26) constant function f '"(x) dejined in (28) converges point-
It is easily verified that subsampling by 2 followed by fil- wise to a continuous function f ( " ) ( x ) .
I
VETTERLI A N D HERLEY: WAVELETS AND FILTER B A N h S 2211
1 , [ 1 2 ] . This property is potentially useful for the

compression of smooth functions by wavelet transforms.
F. Bases of Orthonormal Wavelets Constructed from

Filter Banks
Finally, we show that if the filter used in (27) exhibits
the orthogonality with respect to even shifts described in
Section 11-C, and is regular (for example, if it meets the
conditions of proposition 2 . 1 ) then the infinitely iterated
filter bank generates orthogonal sets of wavelets. That is,
if filters ho(n)and hl(n) and their even translates form an
orthonormal set in 1*(Z),then we generate functions r#~(x),
$(x), which together with their integer translates, form an
orthonormal set in L2(R). These relations are derived in
[ 1 2 ] ;we rederive them since we shall use similar reason-
ing later in the biorthogonal case (Section 111-D).
First, using (27) and (28)
L- I
f " ' ( x ) = 2'/2 . mc= O ho(m) h!-"(n - 2l-l m)
n/2' Ix < (n + 1 ) / 2 ' . (29)

To write h$ - I)(n - 2' - ' m ) in terms off (' - "(x) observe
f('-I)(2x - m) = '2('-1)/2. h$-I)(n)
n/2'-' I2x - m < (n + 1)/2'-' (30)

that is, we have an expression for f (' - "(2x - m ) when x
is in the interval [(n + 2 ' - l m ) / 2 ' , ( n + 2 ' - l m + 1 ) / 2 ' ) .
Making the change of variable: n ' = n 2' - Im this gives +
A
L- 1
f'"(X) = 21/* . c ho(m) - f''"'(2x
m=O
- m). (31)
Recall that when the filter is regular, f'"(x) tends to a

continuous limit function 4 ( x ) as i 00. By taking the
-+
2 L
limit in (31), +(x) itself satisfies a two-scale difference
(a) (b)
equation :
Fig. 13. Iteration (28) for two simple filters. (a) [ l , 3, 3, I ] which con-
verges to a continuous function. (b) [-1, 3 , 3 , - 11 which conberges to a L- I
discontinuous function. $(x) = 2'12 * ho(n) . 4 ( 2 x - n). (32)
n =O
We can define also the bandpass function:

This is a sufficient condition to ensure pointw ise con- L- 1
vergence to a continuous function, and can be used as a
simple test. Note that more accurate methods are avail-
$(x) = 2'12 . c hl(n)
n=O
4 ( 2 x - n). (33)
able for testing regularity [ 3 6 ] , [ 121, [ 3 5 ] .The reason for
Assume now that we have filters such that the impulse
our interest in this bound is that we can make it a design
responses are orthogonal with respect to even shifts:
criterion to place a maximum number of zeros at T . This
approach is explored in Section V-A. (h,(n - 21), h,(n - 2k)) = 6, k, I E Z. (34)
Various methods for estimating the regularity index 0
We show now that $(x) is orthogonal to integer translates
such that $ ( x ) , 4 ( x ) E C p are available [ 1 2 ] , [as].We
of itself Vi. From the definition (28)f "'(x) is just the in-
shall use the methods of [35] to check the limit functions
dicator function on the interval [0, 1 ) ; so we immediately
that we design. These are especially useful if a filter fails
get orthogonality at the 0th level, that is, ( f " ' ( x - I ) ,
the test of proposition 2 . 1 .
f '"(x - k ) ) = Now we assume orthogonality at the
Note that if mo(z) has a zero of order N at z = - 1 then
ith level:
the wavelet generated will have N consecutive vanishing
moments. That is xk$(x) dx = 0, k = 0 , 1 , . ,N - (f"'(x - I ) , f"'(x - k)) = 6k, (35)
I
1
and prove that this implies orthogonality at thc: ( i -t 1)st

level :
=2 c n m
ho(n)ho(m)
(f';'(2x - 21 -- n),
f ' ; ' ( 2 ~ - 2k - m ) ) Call the above 2 X 2 matrix Hm(z), where the subscript
m indicates that it contains modulated versions of the fil-
= c ho(n)ho(n+ 21
n
- 2k) ters Ho(z) and H , ( z ) . Note that from the above equation
in order to eliminate the contribution to the reconstructed
= 8k,. signal from X( -2) (which represents an aliased version of
the signal) it is necessary and sufficient that the synthesis
Hence, by induction (35) holds for all i. So i.1 the limit filters are related to the analysis filters as follows:
i + 00:
[GO(Z),GI(z)l = C(z)[Hi(-z), - Ho(-~>l. (39)
Note that
The other orthogonality relations between $(A) and 4(x) det [Hm(z)l = Ho(z)HI(-z) - HO(-Z)HI(Z) (40)
now follow easily from (34):
= P(z) - P( -z) (41)
where P(z) = Ho(z)Hl(-z). We introduce the following
polyphase notation for the filters:
Hi(Z)= Hi0(Z2) + Z-";1(z2)
To end the section we show that $ ( x ) is orthogonal across
scale. First, across one scale: that is, Hio(z)contains the even-indexed coefficients of the
filter H i ( z ) , while Hil(z) contains the odd ones. Thus
($4
- 0,$(2x - k))
= (F hl(n) * 4(2x - 21 + n), $(2x - k)
c h l ( n ) ( 4 ( 2 x - 21 + n), $(2x - k)) or
['1 - 1'1 ['o O]z

=
n
= 0. Hp(z2) = 2-' * H,(z)
Similar reasoning shows that across several s:ales $42'~ where Hp(z)is called the polyphase matrix. In particular,
- 1 ) is orthogonal to $(2'x - k) f o r j # i and k # 1. We
have shown the correspondence between the orthogonal- det [Hm(z)] = -22 -' det [Hp(z2)1 (42)
ity relations of the filter impulse responses, and those of
their infinite iterates. Besides orthogonality, complete- and
ness is required in order that {2-J/2$(2Jx - 1 I, j , I E Z >
be an orthonormal basis for L2(R),and this it8 proved in
[12]. Thus any functionf(x) E L2(R)can be written as = 1/2 * - P(-z"2)].
z'/2[P(z'/2) (43)
f(x) = c c (2-"2$(25x
J I
- 1),f(x))2-J/2$(2'x - I).
A . FIR Filter Banks
For convenience we repeat the following well-known
111. GENERAL FIR PERFECTRECONSTRUCTION FILTER fact [37], [38], [21].
BANKSAND BIORTHOGONAL WAVELETS Fact 3. I : For perfect reconstruction with FIR synthesis
Since we will be using perfect reconstruction filter
jilters after an FIR analysis, it is necessary and suficient
banks (PRFB's) to construct wavelets, we brizfly review that
the salient points here. For greater detail, a e refer for det [Hm(z)]= c z-~'-' where 1 E Z .
example to [19]-[22]. Assume that we have a filter bank
as in Fig. 8, with analysis filters Ho(z) and H,iz) but with Note that det [HJz)] is thus a pure delay as well. In order
general synthesis filters Go(z) and G,(z) inste2d of H0(z) to make the results that follow less arbitrary, we shall fix
and H,(z). Note that upsampling by 2 (inserting a zero c = 2. This involves no loss of generality, since filters
between every two samples) corresponds to simply re- that differ only by a multiplicative constant can be re-
placing z by z 2 in the z transform. Using (26) it is easily garded as equivalent.
seen that the output of the analysis/synthesi:; system is From this, and (40) it follows that P(z) can have only
I
a single nonzero odd-indexed coefficient: cients, that is

P(z) - P ( - z ) = 2p2/+1z-21-! (44) H&) H I(z - I ) = z -2n - Q(z 2 ) .
and we would normalize p 2 / +I = I . A polynomial P(z) So the zeros appear. in pairs at (a,-a). Proposition 4.3
satisfying this constraint is termed a valid polynomial. It below states that H&) cannot have such a pair of zeros if
is then clear that any factorization of a valid P(z) = PR is to be possible; thus for every zero at z = a in H o ( z ) ,
H,(z)H,( - z ) gives a possible FIR PRFB. Note that H,(z) there must be a corresponding zero at z = -a in H l ( z - I ) .
and H I (-2) are interchangeable. Given an FIR filter H&) This means that H l ( z ) has a zero at z = - 1 / a ;so H l ( z )
we use the term complementary filter to denote any H l ( z ) has the form given in (47). 0
such that P(z) = H,(z)Hl( -z) is a valid polynomial; that This necessary form of H l ( z ) means that
is H&) and H l ( z ) form a PR filter pair.
The qbove discussion indicates two possible design
H 1 ( z ) = -z -2kHo1(z- 2 ) + z -2k- H& -2).
methods for PRFB's [ 2 11. First the factorization method Choosing k = 0 leads to:
consists of finding a valid P(z) satisfying design criteria,
and then factoring it into H,(z) and H I (- z ) . Second, the
complementary filter method starts with a filter Ho(z) and
then solves a system of linear equations to find a comple-
Because it is a perfect reconstruction system
mentary filter leading to a valid P(z). (Note that this P(z)
can then be refactored as desired.) Of course once we have det [Hp(z)l = H O & ) H ~ ( Z - ' )+ H O ~ ( Z ) H O I ( Z=- ' 1
)
designed P(z) and factored it in terms of the analysis fil-
ters H,(z) and H I (-z) the synthesis filters follow directly (49)
from (39), with C(z) = c z-'. since det [H,(z)] has to be a delay, and is symmetric. This
is the polyphase equivalent of (45), because the even coef-
ficients of the autocorrelation H,(z) H,(z - I ) are given by
B. Orthogonal or Paraunitary Filter Banks (49). On the unit circle z = elw,this means that
In Section 111-C we have seen how to construct unitary IHoo(e'u)(2+ IHol(e'u)12= 1 (50)
operators based on filters which were orthogonal to their
that is they form a power complementary pair [38].
even translates. In filtering notation this means that the
The necessary form (47) also means that, substituting
even terms of the autocorrelation function are all zero,
into (40)
with the exception of the central one (which equals unity
for normalized filters). The autocorrelation of a filter H,(z) z -2k - '[H,(z)H,(z -I) + H,( -z)H,( -z -91 = 22 - 2 k - I.
is Hi( z )Hi(z - I ) ; thus the condition ( 1 8) becomes
Or, on the unit circle
H;(z)H~(z-'+
) 2 i (0, I}.
H,(-z)H;(-z~I) = E
+ = 2.
IHo(eJW)l2 (HO(ej(w+7))(2
(45) Note that the matrix Hp(z) in (48) has the following prop-
Further, the two filters H&) and H l ( z ) were orthogonal erty :
to each other at even translates (19), so the even terms of
the cross correlation are all zero:
Ho(z)HI(z-') + &(-~)Hl(-z-') = 0. (46)
If the two filters satisfy (45) and (46), their impulse re-
sponses and even translates form an orthogonal basis for
12(2).We are now ready to prove a fact on the form of
orthogonal FIR solutions:
Fact 3.2: Consider an FIR perfect reconstruction jilter
bank such that H,(z) satisfies (45). Then the length of ho(n) because of (49). In other words, it is unitary on the unit
has to be even. In order that ho(n - 2k) and h l ( n - 21) circle; such a matrix is called a paraunitary matrix [38];
form an orthogonal basis (that is H I(z) should satish (45) if it is also stable it is called lossless [39], [40]. This is
and (46j), it is necessary and suflcient that the extension of the allpass filter concept [41] to matrices
with polynomial entries.
H , ( z ) = z - 2 k - 'HO(-z -I). (47) The above discussion indicates two possible design ap-
Proofi That L must be even was already shown in proaches. The first is to find an autocorrelation function
Section 11-C. That a complementary filter given by (47) that has only a single even-indexed coefficient different
satisfies (45) and (46) and leads to perfect reconstruction from zero. This must be the central one, since an auto-
is easily verified by substitution. correlation function is symmetric. Such a function can be
Necessity is shown as follows: from (46) it is known factored into its "square roots" H&) and H o ( z - ' ) . In
that H o ( z ) H l ( z - ' ) is a polynomial with only odd coeffi- particular, zeros on the unit circle have to be double, since
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 40, NO. 9, SEPTEMBER 1992
the function must not change sign. This method was first FIR m e r bank rlavmg I I I I ~ ~pilase I iias umy L W W 1ioiiLeru
used by Smith and Barnwell [37] and Mintzer :18] to syn- taps, and is given by Ho(z) = z + -'z - 2 n I and H l ( z )
~
thesize orthogonal filter banks. The second niethod uses = ( z - / - Z - / - 2 n - 1 ) z -2n. (Note that paraunitary filters
lattice structures to synthesize paraunitary matrices, for can be factorized so as to approximate linear phase [17],
which complete factorizations have been given by Vai- ~421 ~ 9 1 . 1
dyanathan and Hoang [20]. To obtain longer real FIR filters, and still have exact
Note that designs in the filter bank literatlre did not linear phase, one has to give up orthogonality. The classes
address the question of regularity (see proposition 2.1). of linear phase solutions are indicated below.
Daubechies [ 121 specifically designed filters satisfying the Proposition 3.3: Linear phase perfect reconstruction
above orthogonality relations and having a maximum real FIRjlter banks usingjlters Ho(z)and H I (z) have one
number of zeros at K. of the following forms:
a) Both filters are symmetric and of odd lengths, dif-
C. Biorthogonal or General Perfect Reconstruction fering by an odd multiple of 2.
Filter Banks b) One jilter is symmetric and the other is antisym-
From fact 3.1, we know that perfect reconstruction re- metric; both lengths are even, and are equal or differ by
quires det [ H , (z)] to be an odd delay, and tl-e synthesis an even multiple of 2.
filters are given by (39) with C(z) = c z -'. c) Onejlter is of odd length, the other is even; both
Choose l = 1 for the purposes of this discussion. First have all zeros on the unit circle. Either both Jilters are
note that Go(z)H I( z ) and G l ( z )Ho(z) have on1 q odd coef- symmetric, or one is symmetric and the other is ant-
ficients, that is, isymmetric.
Proof: See Appendix A.2.
( go(n - 2k), h l ( n - 21)) = 0 (51) In class c) we find that P(z) = Ho(z)H I (-z) has zeros
only at the 2n - 1 roots of 5 I ; so the filters are of very
( g l ( n - 2k), ho(n - 2 1 ) ) = 0
(52)
little practical interest. We will not consider this trivial
(note the time reversal in the inner product). In matrix solution further.
notation It is also possible to get lattice structures that generate
HOG1 = 0 = HIGo linear phase FIR filter banks [23], [22]. Appendix A.3
(53)
indicates the lattice that will generate the low-pass Ho(z)
where H, and GI have been defined in a fashion similar to = (1 + z - I ) ~ (that
~ is the "most regular" low pass, fol-
(17). Now Go(z)Ho(z)= z-'Ho(z)Hl(-z) has a single lowing proposition 2.1) and its complementary filter
nonzero odd coefficient (because it is a PR sy:,tem). Sim- H,(z). A general discussion of complementary filters in
ilarly G l ( z ) H l ( zhas
) also only a single even-injexed coef- the biorthogonal case is given in Appendix A.3. The dis-
ficient different from zero. That is, cussion of the biorthogonal pyramid is presented in Ap-
pendix A. 1.
( g I ( n - 2 0 , h,(n)) = 61. (54)
Note that while in the FIR case it is necessary to give
In operator notation up orthogonality to get nontrivial linear phase solutions,
HOG0 = Z = HiGI. in the IIR case it is possible to have both; the filters are
(55)
noncausal, but good designs are possible [25].
Since we have a perfect reconstruction systerr we get
COHO + GIHI = I. (56) D. Biorthogonal Wavelets Based on Filter Banks
Of course the last equation above indicates that no non- To tie together the results of this section we now show
zero vector can lie in the column nullspaces of both Ho that the infinitely iterated biorthogonal perfect reconstruc-
and H I . Note that (55) implies that COHOand Cl H I are tion filter banks generate biorthogonal sets of functions.
each projections (since G,H,G,H, = G , H , ) .They project We denote by H$'(z) and G$'(z) the filters which are
onto subspaces which are not in general orthogonal. Note equivalent to the cascade of i blocks of filtering/subsam-
that (51) and (52) indicate, however, that tl-ere are or- pling in the analysis and synthesis sections, respectively.
thogonality relations between the filter impulse responses. We assume that both of the filters involved are regular.
Because of (51), (52), and (54) the analysis synthesis sys- From these define functions which are piecewise constant
tem is termed biorthogonal. In the special case where we on intervals of length 1/2':
have a paraunitary solution one finds: Go = H;: and G I
= H r , and (53) gives that we have projections onto sub-
f " ) ( x ) = 2'12 * h$'(n) n/2' Ix < (n + 1)/2'
spaces which are mutually orthogonal. Further consider- (57)
ations on biorthogonal systems can be found ir [28], [29].
One reason to use biorthogonal rather than orthogonal
bases is that the additional freedom allows us to have ar-
bitrary length linear phase filters. It is well known [12],
[23], [22] that the only orthogonal (or paraunitary) real Note that we assume regularity, that is in the limit as i -+
I
VETTERLI AND HERLEY: WAVELETS AND FILTER BANKS 222 1
00 both f " ' ( x ) and f " ' ( x ) converge to continuous func- orthogonal case [ 2 8 ] . Hence any function from L2(R)can
tions. Using the analysis in Section 11-F we.easily show be written
that f("( x ) andf(')(x) lead to limit functions which satisfy
the two-scale difference equations: f(x) = C C (2-j12$(2jx - l ) , f ( x ) ) 2-j'2$(2jx - 1 ) .
j 1
L-l
+(x) = 2Il2 c ho(n)

n=O
* 4(2x - n) (59) Note that $(x) and $(x) play interchangeable roles.
So, regular biorthogonal FIR filter banks lead to bior-
L- I thogonal bases of functions of finite length; it is easily
$(x) = 2'12 C
n=O
go(n) * $(2~ - n). (60) shown that the converse is also true. Assume that we have
functions $ ( x ) , 4 ( x ) , $(x), $(x) satisfying (59)-(62) and
Define also the associated bandpass functions: (63)-(66). Then it is easy to verify that they can be used
L-l to derive biorthogonal filter banks. For example, using
$(x) = 2Il2 * c hl(n) - 4(2x - n)
n=O
(61)
(63)
<i(x- 0 ,4(x - k))
L- 1
$(x) = 2'12 C
n=O
gl(n) $(2x - n). (62) = ( go(n) * $(2x - 21 - n ) ,
That &x) and 4 ( x ) are orthogonal with respect to integer

shifts is again shown inductively. We immediately get or-
m
h o ( m ) 4 ( 2 ~- 2k - WZ)
)
thogonality at the 0th level since f(O)(x) and f " ) ( x ) are = c c go@)
n m
* ho(rn) ( $(2x - 21 - n ) ,
each equal to the indicator function on the interval [0, 1):
( f ' ' ) ( x - l ) , f"' (x - k ) ) = hkl. Again assume orthogo- 4 ( 2-
~ 2k - m ) )
nality at the ith level: ( f ( ' ) ( x - l ) , f " ' ( x - k ) ) = hkl. = Cn &(n) ho(n + 21 - 2k) = 6 k l .
And then get
( f ( ' + y X- l ) , f ( i + l ) ( x- k ) ) The other filter biorthogonality relations follow from (64)-
(66).
= 2 (Cn go(n) . f " ' ( 2 x - 2k - n ) ,
C E. Filter Design
m
ho(rn)f(')(2x- 21 - rn)
) Propositions 2.1 gives a sufficient condition to ensure
= c go(n)
n
ho(n + 21 - 2k) pointwise convergence to a continuous function, which
hinges on the number of zeros at a. It is clear that a filter
= 6kl.
with a maximum number of zeros at a will not necessarily
be maximally regular. Ideally, to get a maximally regular
In the limit this gives filter (which is written H(z) = [(l +
z - ' ) / 2 I N F ( z )as in
Section 11-E) we would maximize N while simultaneously
<i(x- 0 ,4(x - k)) = 6kl. minimizing S U ~ ~ IF(ej")(. ~ [ ~ There, ~ is ~ a~ tradeoff in-
volved. However only N , the number of zeros at a, can
Once this is established
be easily controlled.
<$(x - 0 , $(x - 4 ) = 6kl The question of regularity is more involved for the
biorthogonal case than it is in the orthogonal case, since
follows immediately from ( 5 4 ) ; and the relations now we have to check the regularity of both the analysis
(Ho(z))and synthesis ( H I (-z)) low-pass filters (see Sec-
t$(x - I), rl/(x - k ) ) = 0 tion V-A). Fig. 14 illustrates the difficulty of making both
Ho(z) and H I ( - z ) regular for the length 4 linear phase
<$(x - 0,4 ( x - k ) ) = 0 case. Recall that the impulse responses of Ho(z) and
H , ( - z ) are given by [ l , a , a , 11 and [ l , -a, -a, 1 1 ,
come from ( 5 1 ) and ( 5 2 ) , respectively.
respectively [ 2 2 ] . The figure shows the scaling function
We have shown that the conditiork for perfect recon-
generated by [ l , a,a , 11 for a E [ - 3 , 31. At a = 3 we
struction on the filter coefficients lead to functions that
have that Ho(z) is very regular, while H I ( - z ) is very ir-
have the biorthogonality properties as shown above. Or-
regular (as shown also in Fig. 13). The difference be-
thogonality across scale is also verified, following the
analysis of Section 11-F: tween the regularity of the two filters clearly decreases as
a gets smaller, without ever leading to two regular linear
(2-J12$(2Jx- 1 ) , 2-'12$(2'x - k ) ) = 6, 6k,. phase perfect reconstruction filters of length 4 . Actually
at a = 0 there is neither pointwise nor L2 convergence
Thus the set { $ ( 2 ] x - I), $(2'x - k ) , i, j , k , 1 E Z } is [12] (the figure which shows the sixth iteration gives
biorthogonal. That it is complete can be verified as in the therefore an erroneous impression at a = 0).
4(z)
Fig. 14. Scaling function generated by using Iteration (28) for [ l , a , a , 11 and a E [-3, 31, (sixth iteration is shown)
I v . ALGEBRAIC S T R U C T U R E OF FIR S O L JTIONS Fact 4.1: Given a (x) and b (x)

A. Bezout 's Identity a(x)p(x) + b(x)q(x) = 4x1
From Section I11 it can be seen using (42), (40), and
(43) that the condition to achieve perfect reconstruction has a solution [ p(x), q(x)] gcd (a(x),b(x))divides c(x).
with FIR synthesis filters after an FIR analysis section can We now make use of these observations to examine the
be expressed in two equivalent forms: constraints on the filter banks.
Proposition 4.2: Assume that the jilters Ho(z) and H I(z)
~ o o ( z ) H , , ( z-
) H0,(z)H,o(z) = z - I (67) are FIR and causal. Then given one of the pairs [HOo(z),
H o I ~ z )[Hlofz),
~, Hllfz)19 [ H d z ) , HlOfz)l or [ & l ( z ) ,
H&)HI(-z) - H ~ ( - z ) H I ( z )= 2 ~ - ~ ' - ' . (68) H I (z)] in order to calculate the other pair necessary to
achieve perfect reconstruction it is necessary and SUB-
Each of these equations implies the other, and we use (67) cient that the given pair be coprime (except for possible
or (68) depending on which is more convenient for our zeros at z = 03).
purpose. These requirements of course considerably con- Proof: From fact 4.1 the gcd of each of these pairs
strain the possible solutions; both (67) and (C18) are spe- must divide the right-hand side of (67). Hence the only
cial forms of what is known as a Bezout identity [43], factors that they can have in common are zeros at z =
[44]. This identity arises in the Euclidean algorithm which M. U
calculates the greatest common divisor of two polyno- The above proposition is also proved using a different
mials ao(x)and a l ( x ) : argument in [45].
Proposition 4.3: A Jilter HO(z)has a complementary $1-
a&> = a gcd (a&>, a&)> ter if and only if it has no zeros in pairs at z = a and z
-
- -a.
where a,(x) is the last divisor of the algorithm and a is a
constant. It is well known that we can always write Proof: Ho(z) has a zero pair at ( a , - a ) if and only
if H&) has a factor A(z2). This happens if and only if
both Hoo(z) and Hol(z)have a common factor A @ ) , that
is, they are not coprime. Thus the absence of zero pairs
of the form ( a , - a ) and coprimeness are equivalent and
the proof is completed by using proposition 4.2. 0
If ao(x)and a , ( x ) are coprime then a,(x) has z m s at 0 or Proposition 4.4: There is always a complementary jil-
03 only; and the identification between (69), the Bezout ter to the binominal jilter:
identity, and (67) or (68) becomes clear. By examining
the implications of this, new results emerge; also, from H&) = (1 + z - ' ) ~= Hm(z2) + z-IH01(z2). (70)
this angle certain results that previously ha1 more in-
volved proofs (e.g., [45]) now become simpler. A review Proof: If HO,(z) and Ho,(z) had a common factor it
of the necessary properties of the Euclidean algorithm can would appear as a pair of zeros of Ho(z)at (a,-a);since
be found, for example, in [46]. The importilnce of the Ho(z) has zeros only at z = - 1 it cannot have such a
algorithm in the context of biorthogonal systens has also factor. 0
been noted in [28] and [47]. The following easily proved It should also be clear that for Ho(z) and H l ( z ) to form
fact will play an important role in the rest of this section a perfect reconstruction pair it is necessary that they be
1461. coprime. From fact 4.1 gcd [HO(z),Hl(z)] must divide the
...
right-hand side of ( 6 8 ) . This has a very clear signal pro- Proposition 4.5: Given a linear phase Ho(z) of odd
cessing interpretation: a zero common to Ho(z) and H l ( z ) length N , and its length N - 2 linear phase complement
would imply a transmission null in both channels of the H I(z), all higher degree odd length linear phase jilters
filter bank at some frequency, making reconstruction im- complementary to Ho(z) are of the form
possible.
H [ (z) = z - 2 m H I ( ~+) E(z)Ho(z)
B. An Analogy with Diophantine Equations where
m
The conditions under which a complementary filter to
some chosen Ho(z) exist were detailed in Section IV-A. It E(z) = ,Eai(z-2(r
r=I
- I) + -(4m-2i)
>.
has already been pointed out that such a filter could be Proof: Note that E(z) = E( -z), and that E(z)Ho(z)
found using Euclid's algorithm; another method 'of find- is symmetric about the point (N + 4m - 3 ) / 2 just as
ing a complementary filter based on solving a set of linear +
) Hence, z -2mH1(z) E(z)Ho(z)is easily seen
z - ~ " H , ( z is.
equations is given in [21] and Section V-A. The comple-
ment is not unique. If given a filter Ho(z),we can calculate
to be a linear phase solution of length N +
4m - 2 by
direct substitution.
one of its complements H , ( z ) , it is natural to wonder how +
We now show that all length N 4m - 2 solutions are
we may find others. The results of this section will show of this form. If H ; ( z ) is of length N + 4m - 2 then
that given any complementary filter, a simple mechanism
Ho(z)H;(-z) and Ho(z)E(z)Ho(-z) are both valid and of
exists for finding all others.
The basic idea again stems from our interpretation of
length 2N + 4m - 3 . So also is P " ( z ) = Ho(z)[Hi (z) -
E(z)Ho(- z ) ] . We can choose the coefficients of E(z) to
the condition for FIR perfect reconstruction as a Bezout set some of the end terms of P " ( z ) to zero. That is, for
identity. In number theory equations with integer coeffi- some choice of a l the coefficients of z o and z - ( 2 N + 4 m - 4 )
cients for which integer solutions are sought are known become zero, so that P " ( z ) is reduced in length by 4 (the
as diophantine equations [ 4 3 ] . For example, coefficients of z and z -(2N 4m - 5 , are already zero).
+
ax+by=c (71) Similarly, for each of the aiwe can reduce the length of
P " ( z ) by 4 . When a l , * * , am have been appropriately
where all quantities are integers and we seek the solution fixed P " ( z ) has length 2N - 3 , has powers of z - ' in the
( x , y ) is the most basic diophantine equation. Clearly, if range -2m to - 2 m + 2 N - 4 and is still valid. Since it con-
solutions of the equation
tains Ho(z) as a factor it must have the form
ax' + by' = 0 (72) P''(2) = z-2mHo(z)HI(-z)
are available then we can add them to ( x , y) and generate since the length N - 2 solution is unique by proposition
new solutions to ( 7 1 ) . If we replace integers by polyno- A.3. 0
mials then ( 7 1 ) is analogous to the equation for perfect From proposition 3.3 it follows that in the two nontriv-
reconstruction that we must solve. We exploit this fact in ial cases of linear phase solutions, the length of P(z) =
the rest of this section by noting that polynomial solutions H o ( z ) H l ( - z ) is 4n - 1 . In Appendix A.4 it is shown that
to an equation analogous to ( 7 2 ) are indeed easy to find; the solutions indicated in proposition 3.3 b) are special
this allows us to generate all possible complenientary fil- cases of those in a); that is, they can always be refactored
ters if we are first given one. In fact, if we work with the into the form a). It follows that all higher degree comple-
modulation version of the PR condition ( 6 8 ) , arid iden- mentary filters to a fixed Ho(z) are given by proposition
tify: a = -Ho(z) and b = Ho(-z) it is easy to see that 4 . 5 , unless they are trivial, in the sense of belonging to
[ x , y ] = [E(z)Ho( -z), E(z)HO(z)]sets the right-hand side class c) of proposition 3 . 3 .
to zero if E(z) = E( -z). Example: Consider Ho(z) = [ 1 , 4 , 6 , 4 , 11 and its unique
For the following proposition we make use of propo- length 3 complementary filter H , ( z ) = [ 1, 4 , 11 / 16. Let
sition A.3, which gives that the length N - 2 linear phase m = 2 . So we get
complementary filter to an odd length N linear phase filter which is linear phase, and complementary to Ho(z).
is unique if it exists. A further result allows us to use the diophantine equa-
I
I
tions to reach more general solutions. For this we use the domain; but the results there, like those of this section
result (from proposition A.2) that a length N filter has could be expressed in either modulation or polyphase no-
N - 2 length N - 2 complementary filters. tation. For the sake of definiteness consider the polyphase
Proposition 4.6: All filters of length N + 2m - 2 which version of the perfect reconstruction condition (67), and
are complementary to a length Nfilter Ho(z) have the form assume that Hoo(z) and Hol(z)are given. It can be seen
from proposition 4.6 that this lowest degree solution is in
Hi (z) = z -2kHl(z) + E(z)H&)
some sense fixed, and embedded at the core of any higher
where E ( . ) = E(-z) is apolynomial of degree 2 ( m - I ) , degree solution. The strong connections between Euclid’s
k E {O, I , * , m ) and H I(z) is a length N - 2 comple- algorithm, the Bezout identity and continued fraction ex-
mentary filter. pansions (CFE’s) is well known [43], [48]. In fact, we
Proof: That this is a solution is easily verified by now show that the canonic CFE of H l o / H I Iis the same
substitution. If as that of Hoo/Hol except for the last term; and further
m-1 that higher degree solutions are formed by adding terms
E(z) = c pIz-2’
r=O
to the CFE, the first terms remaining unchanged.
We define = H,, Do = H o l ,A - , = Hlo, and A.
then P“(z) = Ho(z)[HI (-2) - E(z)Ho(-z)] is reduced in = H I I . For the sake of simplicity we remove the phase
length by 2 for some choice of Po or 0, - Again the factor in (67). In this notation (67) becomes
argument is repeated; the length of P ” (z) can be reduced D-,(z)Ao(z) - A-,(z)D,(z) = 1. (73)
by 2 at each stage by fixing one of the 0,. In t i s way we
reduce the degree of P”(z)by 2m when all of the &’s are Now use Euclid’s algorithm starting with the pair D-,(z),
appropriately chosen. Since Ho(z) is still a factor, the re- Do(z). The first step gives
maining factor must be one of the length N - 2 solutions D - I ( z ) = b~(z)D,(z)+ Dl(z) deg DO > deg D1.
of which there are only N - 2. lI3
So far we have addressed only the problem of gener- Also do one division of the pair A - , ( z ) , Ao(z), denoting
ating higher degree solutions from lower ones; the next the remainder A I (z):
proposition shows that we can also go in the opposite di- A-l(z) = ao(z)Ao(z) + A , ( z ) deg A. > deg A l .
rection. We examine the linear phase case only, the ex-
tension to the general case is obvious. For the linear phase Together these equations give
case this says that given any complementary filter we can 1 = D-l(z)Ao(z) - A-,(z)D,(z)
generate all others using the results of proposition 4.5.
Proposition 4.7: If Hi2)(z) and H‘,” (z) are length N + = + Dl(z)Ao(z) - AI(z)Do(z).
(bok) - ao(z)>Ao(z)Do(z)
4m2 - 2 , and length N + 4ml - 2 linearphLse comple- However, since deg AoDo > deg D I A o and deg AoDo >
mentary jilters to the linear phase odd lenjlth N filter deg DoAl must have a. = bo, and hence
Ho(z), with m2 > m , then we can generate th? lower de-
gree solution from the higher as follows: DO(Z)Al(Z) - A,(z)D,(z) = - 1.
H(I)(~) = Z2(mz-m~)[~(2) Since this is of the same form, but of lower degree, than
(l,(Z) + E3(z)Ho’z)1
the equation that we started with (73), we can compare
where E3(z) = z - ~ - m( ‘ )~E~I (2) - E2(z). the second step of Euclid’s algorithm with a division of
Proof: Direct substitution. 0 Ao(z),Al(z) and this gives a l ( z ) = b,(z). The result is that
The subclass of all valid P(z)’s which can be factored we get a succession of Bezout identities:
as P(z) = Ho(z)Ho(z - I ) corresponds to paraur itary FB’s,
and generates orthonormal bases of wavelet:. Since all Dj- I(Z)Aj(Z) - Dj(Z)A,- l(Z) = (- 1)’ (74)
such P(z)’s are symmetric and positive they lorrn a sub-
class of those generated by the construction of proposition which are of decreasing degree. We find in turn that ao(z)
4.5 above. The case for which Ho(z) is chosen to be an = bo(z),a l ( z ) = b l ( z ) , * , a j ( z ) = bj(z), * , a N ( z )
even power of the binomial is treated in detail by Dau- = bN(z). Note that these outputs of Euclid’s algorithm are
bechies. In Appendix A.5 we establish the close relation the partial denominators of the canonical CFE of
between her results and proposition 4.5 above. The fact D- I ( z )/Do(z) [48l 1431:
that higher degree solutions contain lower degree ones is
also given in [47].
C . Continued Fraction Expansions

In Section IV-B it was shown that any so1u:ion to (67)
or (68) could be written as the sum of lower degree so-
lutions and trivial higher degree solutions (trivial in the
sense that the right hand side of (44) becomes zero). In
Section IV-B we dealt exclusively with the modulation
I
VETTERLI AND HERLEY: WAVELETS AND FILTER HANKS 2225
bN]to denote a terminating CFE [49]. It follows that the Clearly, this structure is complete; all possible comple-
CFE’s of D - l(z) /Do(z) and A - l(z) /Ao(z)are identical for mentary filters can be generated by appropriate choice of
the first N + 1 terms. The terminal equation f o r j = N A N (2).
gives It warrants reiteration that the above analysis using the
polyphase notation, can be used to give equivalent results
in the modulation notation.
Remark that the Dj(z)’s are known, and DN is ;I scalar V . DESIGNRESULTS

(since it is the last divisor and D-,(z), Do(z) are assumed
to be coprime); hence we can choose A N ( z ) to be any
A . Linear Phase Filters with a Maximum Number of
polynomial and we are assured that Aj(z)’s are polynomial
Zeros at
-
for j = N - 1, * , 0, - 1. Therefore we get a valid We now design a linear phase P(z) = Ho(z)H I (-2) sat-
complementary filter [Ao(z),A - l(z)] for any appropriate isfying (68) and having the maximum possible number of
A N ( z ) . This can be expressed by writing u N + l(z) = zeros at z = - 1. Note that if P(z) has even a single zero
( - l ) N A N( z )DN. In summary, at z = -1 it cannot have any at z = 1 and vice versa.
This should be clear since we wish to have P(z) - P( - z )
equal to 22 - 2 1 - (see Section 111-A). We consider only
the case where P(z) is of odd length (see proposition 3.3);
hence it has an even number of zeros; since P(z) is linear
phase its zeros must occur in quads, and pairs on the unit
circle or real axis. It follows that P(z) must have an even
number of zeros at z = - 1 or none at all.
- ( - 1>”4N(z> DNI . (77)
Note that by proposition 4.4 we are assured that there
It can be shown that choosing A N ( z ) = 0 gives the same exists a complementary filter to the binomial of any de-
solution as that produced by Euclid’s algorithm. ‘That is, gree. We wish now to calculate R 2 k ( - z ) which is the
complementary filter to (1 + z In other words, we
wish to find the polynomial such that P(z) = (1 +
z - 1 ) 2 k R 2 k (is~ )valid, as defined by (44). That R 2 k ( z ) is
= [bok); ~ I ( z ) *, bN-1(41 linear phase is an immediate consequence of the fact that
If we divide (74) by DJ (z)DJ - 1 ( ~we
) get
*
P(z) and (1 + are. Because of this symmetry, on
equating the appropriate terms to zero it turns out that
only k of the equations are independent; so we get a set
(z) AJ- 1 (z) - (-1)’-
of k by k equations to be solved for the coefficients (ro,
(78)
(z) OJ- 1 (z) OJ(z)OJ- 1 (z) rl, . , rk-1) [ W .
which is the expression relating successive convergents to Example: If we choose Fo(z) = (1 + z we solve
a CFE [50]. This is precisely the continued fraction in the 3 by 3 system found by imposing the constraints on
(76), (77). Note that since the successive convergents sat- the coefficients of the odd powers of z - I of
isfy the decreasing degree Bezout identities (74), truncat-
ing (76) and (77) gives a lower degree FIR PRFB. ~ ( z =) (ro + r l z - l
+ r2zp2 + r l ~ - 3+ r , ~ - ~ )
There is more structure yet. Dividing (74) by
(1 + 62-I + 1 5 ~ +- ~~ O Z f- ~
DJ - ( z )AJ - (z) gives
+ 62-5 + P).
A,(z) DJ(z) - (- 1)’
AJ - 1(z) - 1 (z) -1 (z)OJ- l(z) . So we solve
Like (78) this relates successive convergents of a contin-

ued fraction; but the CFE here is that of a different group-
ing of the polyphase pairs. It turns out that we get
giving r6 = (3/2, -9, 19)/128.

-- - [-(-l)NAN(z)DN; bN(z),bN-,(z), * * e bl(z)l. In general therefore we solve the system
HOl(Z)
F2k * r2k = e2k (79)
Of course, these continued fraction relations give an-
other method for generating higher degree solutions to the where F2kis the k X k matrix, r2k = (ro, * -
, r(k- and
perfect reconstruction condition. By varying A,(z) in (77) e2kis the length k vector (0, 0, , 1).
for example we generate different complementary filters. Having found the coefficients of RZk(z) we factor it into
I
(b)
Fig. 15. Biorthogonal wavelets generated by he length 18 filters given in Tables I and 11. The filters have a maximum number
of zeros at z = - 1. (a) Time function and spc,ctnim of analysis wavelet. (b) Time function and spectrum o f synthesis wavelet.
TABLE I TABLE I1
IMPULSE RESPONSECOEFFJCJENTS O F IMPULSE RESPONSE COEFFICIENTS OF
BIORTHOGONAL Ho(Z);THE INFINITE BIORTHOGONAL Hl(Z); THE INFINITE
ITERATIONOF THESEFILTERS GIVE ITERATION OF THESEFILTERS GIVE
THE WAVELETS SHOWNI N FIG. 15 THE WAVELETS SHOWN I N F I G . 15
0.00122430 0.00122430
-0.00069860 0.00069979
-0.01183749 -0.01134887
0.01 168591 -0.01 141245
0.07 130977 0.02347331
-0.03099791 0.00174835
-0.22632564 - 0.0444 1 890
0.06927336 0.20436993
0.73184426 0,64790805
linear phase components; and then regroup these factors Fig. 15 shows the analysis and synthesis wavelets and
of RZk(z),and the 2k zeros at z = - 1 to form two filters: their spectra for one particular factorization of the k = 9
Ho(z) and H I (-z), both of which are to be regular. case, and Tables I and I1 list the coefficients of the filters
It turns out that for small k , ensuring that both Ho(z) Ho(z) and H l ( - z ) . Each of these filters has a factor (1 +
and HI(-z) meet the bound of proposition 2 . 1 can force z-I)~. We get Ho(z) = [(l + Z - ~ ) / ~ ] ~ Fwhere
~(Z)
one to choose filters of quite unequal length. E;or larger k , ~ [ ~= ,
S U ~ ~(Fo(ej")l ~ ~< ~28; so Ho(z)satisfies the
115.06
however, this problem eases, and it becomes possible to bound of proposition 2 . 1 . Hence the scaling function @(x)
get filters of the same, or approximately the same, length and wavelet $(x) generated by the infinite iterations, fol-
that generate regular symmetric wavelets. For Iexample, lowing (59) and ( 6 1 ) , converge to continuous functions.
I
VETTERLI A N D HERLEY: WAVELETS AND FILTER BANKS 2221
Fig. 16. Orthonormal wavelet generated by length 18 filter (from [12]). The filter has a maximum number of zeros at z = - 1.
(a) Time function. (b) Spectrum.
I
5 IO 15 20
(b)
Fig. 17. Biorthogonal wavelets generated by filters of length 20 and 24 given in Tables 111 and IV. The filters were designed
using the diophantine approach, and have better siopband performance than those in Fig. 15. (a) Time function and spectrum
of analysis wavelet. (b) Time function and spectrum of synthesis wavelet.
HI(
-2) can be similarly factored, and sup,,[o, 2nl B. More General Solutions: Diophantine Approach
IFl(eJw+")( = 211.3 < 28;hence $(x) and $(x) converge
to continuous functions also. While the design procedure of the previous section is
The method outlined in [35] yields the following esti- very simple, the spectra of the scaling function * ( U ) and
mates for the regularity index 0:for $(x) we find 2.46 < the wavelet +(U) are not as one might wish from low-pass
0,and for $(x) we get 3.55 < 0. and bandpass filters.
For comparison purposes we show in Fig. 16 the cor- Proposition 4.5 showed how to generate any valid lin-
responding plots for the orthonormal wavelet generated ear phase P(z) containing a given factor. For example, to
by a filter of the same length, as presented in [ 121. design a P(z) with 2k zeros at T ,we can calculate RZk( - z ) ,
I
TABLE 111 TABLE IV

IMPULSERESPONSE
COEFFICIENTS OF IMPROVED IMPULSERESPONSE
COEFFICIENTS OF IMPROVED
BIORTHOCONALHo(Z); THE INFINITE ITERATION BIORTHOCONALH1(Z);THE INFINITE ITERATION
OF THESE
FILTERS GIVETHE WAVELETS OF THESE
FILTERSGIVETHE WAVELETS
SHOWN IN FIG. 17 SHOWN I N F I G . 17
0.00133565 0.00465997
-0.0020 1229 0.00702071
-0.00577577 -0.01559987
0.00863853 -0.0232792 1
0.01279957 0.05635238
- 0.0236 1445 0.10021543
- 0.0 1900852 -0.06596 I5 1
0.04320273 -0.13387993
-0.0093 1630 0.38067810
-0,12180846 1.103981 18
0.05322 182
0.4 I5897 I4
as in the previous section. This solution, which has de-

gree 2k - 2, can then be used to generate all possible
solutions of degree 2k - 2 + 2m.
A second approach is to note that we need not place all
of the zeros at z = - 1 to start with. We could f x example
calculate the complementary filter to a factor (1 +
-
Z - ~ ) ~ ' [ U ~ ( Z ) U ~ ( ~ )u r ( z ) 1 2 where U ,( z ) represents a zero
pair on the unit circle. We are then assured csf having a
factor (1 + Z - ~ ) ~ [ U ~ ( Z ) U ~ ( Z* ) u l ( z ) ] to p ace in the
stopband of each of the filters.
In Fig. 17 we show the two symmetric wavelets $ ( t )
and $(t) and their spectra. These were generated by linear
phase filters of lengths 24 and 20, the coefficients of which
are listed in Tables I11 and IV. The filters weie designed
using a combination of the two approaches described
above. Each of the filters has only a single zero at z =
- 1, and neither of them meets the bound given in prop-
osition 2.1. However again using the estkmation methods
of [35] we find 0.79 < and 0.96 < 0.Note that the
stopband performance is much better than in Figs. 15 and
16.
Fig. 18 shows the wavelet and its spectrum for a better
orthogonal set of filters. The associated filter bank is
paraunitary; so H l ( z ) = H 0 ( z - I ) . The coeflicients are
listed in Table V. The regularity estimate here is 0.97 < 0
P. (b)
Fig. 18. Orthonormal wavelet generated by length 22 filter given in Table
V . The filter was designed using the diophantine approach, and has better
C. Root Loci of Higher Degree Solutions stopband performance than that in Fig. 16. (a) Time function. (b) Spec-
trum.
While we have shown that higher degree solutions can
give better results, it is still clear from Sections IV-B and
IV-C that these solutions are nonetheless very con- and plot the trajectory in the plane of the roots as (Y is
strained. To give a concrete example we briefly examine varied.
+
the case where H o ( z ) = (1 Z - ' ) ~ and the unique linear Fig. 19 shows the root locus for the region: (Y E [-2,
phase degree 4 complementary filter has impulse response 21. While the increased degree solution has more free-
coefficients h l ( n ) = [3, 18, 38, 18, 3]/256. V7e examine dom, it is clear that the roots move along very constrained
the m = 1 solution (from proposition 4.5) paths. In other words, it would be necessary to look at
solutions of considerably higher degree to get substantial
design freedom.
I
TABLE V PRFB's. If compact support is not desired, similar tech-

IMPULSE RESPONSE OF IMPROVED
PARAUNITARY FACTORIZATION:
niques, using IIR filters generate orthogonal wavelet bases
THEINFINITE ITERATION GIVES with symmetries [25], [51].
THE WAVELET SHOWN IN
FIG. 18 A
APPENDIX
Ho(z) 1. Biorthogonal Pyramid

This Appendix extends the result of Section 11-D to the
ho(O) 0.055739
h d 1) 0.288322 biorthogonal case (see also Section 111-C). In this case,
hd2) 0.614682 the approximation X is equal to
h0(3) 0.608634
h0(4) 0.113646 3 = GoHo * X.
hd5) - 0.290892
ho(6) -0.13 1805 The difference signal becomes
h0(7) 0.162510
ho(8) 0.085330 d = (I - GoHo)x = GlHl . X.
ho(9) - 0.099666
ho(10) -0.042965 Because we have perfect reconstruction:
hdll) 0.060044
ho(12) 0.015233 GoHo + GI HI = I.
hO(W -0.032323
hd4) -0.001634 That is, we can filter the difference signal with hl(n) and
hO(W 0.014 199 subsample by 2, because
h d 16) - 0.002305
ho( 17) -0.004433 Hi GI HI = HI
ho(18) 0.00 1808
ho(19) 0.000646 from which d can be reconstructed by upsampling and in-
ho(20) -0.000577
ho(2 1) 0.0001 11 terpolation with G I , following the fact that
GI HI GI HI = GI HI
and the reconstructed signal, obtained by adding X to d ,
is equal to the original. That is, the biorthonormal pyra-
mid is equivalent to a critically sampled biorthogonal fil-
ter bank. Note that linear processing has been assumed
throughout, which would not be the case in coding appli-
cations.
2. Proof of Proposition 3.3

Proof: Since Ho(z) and H I (-z) have linear phase so
does P(z). If P(z) has odd length, then both filters have
odd length or both have even length. Note that
2N
P(z) - P(-z) = ,zp ; z - ' [ l
r=O - (-1)7
N- I
Fig. 19. Locus of the movement of the roots of (80) in the z plane for a E
Since P(z) is of odd length, both P(z) and P( -z) are sym-
1-2, 21.
metric or antisymmetric about the point i = N . Hence,
P(z) - P( -z) is symmetric about this point; so the single
VI. CONCLUSION nonzero coefficient is p N , the central one. Hence P(z) is
The relationships between wavelets, multiresolution symmetric and not antisymmetric; which implies that
signal analysis, and filter banks have been developed, em- Ho(z) and H I (-z) are both symmetric or antisymmetric.
phasizing the equivalence between the fundamental ideas a) Lo and LI both odd. Now Lo + L1 - 1 is odd. The
of each of these fields, which is that functions or signals center of symmetry, which is (Lo + L l ) / 2 - 1 samples
be broken down into component signals at different reso- away from the end points, has to be odd. Thus (Lo +
lutions. In addition, strong similarities between the de- +
L l ) / 2 = Lo (Lo - L l ) / 2 has to be even. Thus (Lo -
tails of these techniques have been pointed out, and the L1)/2 is odd, and the length difference Lo - L , is an odd
interplay of ideas has enabled us to indicate new results multiple of 2. In particular, there are no same length so-
in each of the areas. lutions. Suppose that Ho(z) and H I (-z) can be both sym-
In particular, we derived biorthogonal compactly sup- metric or both antisymmetric. The latter possibility is
ported wavelets bases with symmetries using regular FIR ruled out, however, because the two polyphase compo-
I
I
2230 IEEE TRANSACTIONS ON SIGNAL PROCESSING. VOL. 40, NO. 9, SEPTEMBER 1992
nents are also antisymmetric when the length is odd, and by Gopinath [52].It is clear that the complementary filter
are therefore not coprime. Perfect reconstruction is then has only rational coefficients.
impossible following proposition 4.2. Now P(z) can be written as
b) Lo and LI both even. Again (Lo L , ) / 2 - 1 has to+
be odd so that the center term is not in the same set as the P(z) = Ho(z)Hl(-z) = (1 +z-')~~R~~(z).
+
end terms. Then Lo (Lo - L 1 ) / 2has to be even, and Since RZk(re") is linear phase it can be converted into
since Lo is even, this means that the length difference Lo a polynomial of degree k - 1 in cos (0 - j In r). Thus
- L1 is an even multiple of 2; and, for example, same the size of the factorization to be performed is halved, and
length solutions do exist. Also it was assumed that Ho(z) the accuracy of the solution is improved. If the exact ra-
and H l ( - z ) were both symmetric, but since Erl( -z) has tional coefficients of Rik(z)are used and this is factored
opposite symmetry to H l ( z ) when the length is even, the in cos z , the solution can be considerably more accurate
even length solution leads to a symmetric/anti symmetric than if a standard linear equation solver is used and direct
pair. factorization is performed. We will consider k = 6 to give
c) If P(z) has even length, then one filter has even an example of the importance of these considerations. De-
length, and one odd. Now we get note P&) which is exact, PA(z) which is found by multi-
2N- 1 plying out after using the lattice outputs given by propo-
P(z) - P(-z) = ,zp;Z-'[l
1=0
- (- 1)7 sition A . l and the efficient factorization; P t ( z ) is found
by multiplying out after using an ordinary linear equation
N- 1
solver, and direct factorization. We find 11 p6(n) -
p;(n)/lm= 0.00892 while 11 p6(n)- p;(n)ll, = 17.83739.
P(z) has only a single nonzero odd-indexed coefficient;
4. Restrictions on Complementary Filter Forms
but note that pi = fp2,- i = 0 , * * , N - 1 . So odd-
indexed coefficients are paired with even-indexed-ones. It Proposition A.2: A length N j l t e r has at most N - 2
follows that P(z) also has only one nonzero even-indexed complementaryjlters of length N - 2; and at most one
coefficient, and is of the form of length N - 1.
Proof: If H l ( t ) has length N - 2 then P(z) =
~ ( =~p j z1- J ( l k z 2 N - I - * j 1. (81) H o ( z ) H l ( - z ) has length 2N - 3 ; hence it has N - 2 odd-
Since P(z) has all its zeros on the unit circle (at the 2(N indexed coefficients. Thus we solve a square system of
- j ) - 1 roots of -t l ) , Ho(z) and H I (-z) also have all linear equations of size N - 2. The nonzero odd-indexed
zeros on unit circle. If Ho(z) and H l ( z ) are both to be an- coefficient can be placed in any of N - 2 positions.
tisymmetric, both must have zeros at z = 1 . This will If H l ( z )has length N - 2 then P(z) = Ho(z)Hl(-2) has
force H I (-z) to have a zero at z = 1 or z = - 1 depending length 2N - 2 ; one of the end terms is odd indexed while
on whether it is of odd or even length, respectively. This the other is even. There are N - 1 odd-indexed terms.
implies that P(z) has either a double zero at z = 1, or a Thus we solve a size N - 1 square system, but there is
pair at z = 1 and z = - 1 . Since P(z) contains only the only one possible position for the single nonzero odd-in-
2(N - j ) - 1 roots of k 1 both possibilities are ruled out. dexed term, which must appear at an end. 0
Hence the filters must have opposite symmetry. 0 Proposition A.3: For the linear phase case, if a $Iter
of even length N has a same length complementaryjlter
3. Lattice Structures Generating the Binomial Filter it is unique. For N odd if there is a length N - 2 comple-
and its Complementary Filter mentary j l t e r it is unique.
Proposition A. I : Let the polyphase components of Ho(z) Proof: N even gives N / 2 equations in N / 2 un-
and H I (z) be generated by the following lattice structure: knowns; the nonzero coefficient must be in the center.
Similarly, N odd gives (N - 1 ) / 2 equations in (N - 1 ) / 2
unknowns. 0
Note: An example of the equations to be solved in the
linear phase case is given in Section V-A. In propositions
then the choice A.2 and A.3 it is of course possible that the linear systems
to be solved are singular, in which case the complemen-
(a19 ff2, * f
f,)
9
tary filter does not exist (see Section IV-A for necessary
= ((2N + 1)/(2N - l ) , (2N - 1)/(2N - 3 ) , *
, and sufficient conditions). It is also possible that on solv-
ing the linear system one or more or the end terms of the
2N + 1) complementary filter found are zero, in which case it fails
to have the advertised length. For example, if we try to
leads to
find the linear phase length 5 complement to the length 7
Ho(z) = (1 + z-1)2N+1. filter Ho(z) = [ 1 , 0 , 0 , - 1, 0, 0, 11 we find H,(z) = [0,
0 , 1 , 0 , 01. For the purposes of propositions 4.5 and 4.6
An elegant proof of this proposition was pointed out to us this case, where the complementary filter has less than the
I
I
VETTERLI A N D HERLEY WAVELETS AND FILTER HANKS 223 1
expected length, has no effect provided that the zero end EeviC and N. T. Thao at Columbia University and the
terms are included. anonymous reviewers for suggesting significant improve-
Of the three classes of linear phase FIR solutions given ments in the manuscript.
in proposition 3.3, we note that solutions of class b) can
always be refactored to ones of class a). Suppose in a REFER EN cEs
class b) solution H l ( z ) is the antisymmetric filter; it then
necessarily has a zero 1 - z - l , which implies that H I (- z ) [ I ] J. M. Combes et al., Eds., Wavelets, Time-Frequency Methods, and
has a zero 1 + - I . This zero can be given to H o ( t ) . Both Phase Space (Lecture Notes on IPTI). New York: Springer-Verlag,
1989.
filters are now of odd length, and differ by an odd niultiple [2] P. Goupillaud, A. Grossmann, and J. Morlet, “Cycle-octave and re-
of 2. Hence, we have a class a) solution. The converse is lated transforms in seismic signal analysis,” Geoexploration, vol. 23,
pp. 85-102, 1984.
not in general true. [3] Y. Meyer, “Orthonormal wavelets,” in Wavelets, Time-Frequency
Methods and Phase Space (Lecture Notes on IPTI), J . M. Combes et
5. Relation to Work of Daubechies al., Eds. New York: Springer-Verlag, 1989.
[4] A. Grossmann and J . Morlet, “Decomposition of Hardy functions
We here establish the connection between proposition into square integrable wavelets of constant shape,” SIAM. J . Math.
4.5 and a result given in a different form in [12]. Anal., vol. 15, no. 4 , pp. 723-736, 1984.
Daubechies writes: P‘l’(z) = (1 + z-1)2kQ(’),’z);ex-
[ 5 ] Y. Meyer, Ondelettes, vol. 1 of Ondelettes et Opkrateurs. Paris:
Hermann, 1990.
cept for a phase factor this gives [6] S . Mallat, “Multifrequency channel decompositions of images and
wavelet models,” IEEE Trans. Acoust., Speech, Signal Processing,
vol. 37, pp. 2091-2110, Dec. 1989.
P(w) = 171 G. Strang, “Wavelets and dilation equations: A brief introduction,”
SIAMRev., vol. 31, pp. 614-627, Dec. 1989.
[8] E. A. Rosenfeld, Multiresolution Techniques in Computer Vision.
Since Q(w) is symmetric it can be written as a polynomial New York: Springer-Verlag, 1984.
in cos (w) = 1 /2( 1 - cos2 (w/2)) when restricted to the [9] P. J . Burt and E. H . Adelson, “The Laplacian pyramid as a compact
unit circle. So if we define y = cos2 (w/2) and we can image code,” IEEE Trans. Commun., vol. COM-31, pp. 532-540,
Apr. 1983.
rewrite the above as [IO] S . Mallat, “Multiresolution approximation and wavelets,” Tech.
Rep., Dep. Comput. Inform. Sci., Univ. Pennsylvania, Philadelphia,
P(w) = [COS’ ( ~ / 2 ) ] ~ Q j / 2 (-
1 COS‘ (~/2)) PA, Sept. 1987.
[ 1I ] S . Mallat, “A theory for multiresolution signal decomposition: The
wavelet representation,” IEEE Trans. Part. Anal. Machine Intell.,
vol. 11, no. 7, pp. 674-693, 1989.
Using the results of proposition 4.5 we know that higher [I21 I . Daubechies, “Orthonormal bases of compactly supported wave-
degree solutions are written lets,” Commun. Pure Appl. Math., vol. XLI, pp. 909-996, 1988.
[I31 0. Hermann, “On the approximation problem in nonrecursive digital
Q’(-z) = z - ~ ” Q , / ~ ( - z )+ E(z)(l - z--I)*’.. filter design,” IEEE Trans. Circuit Theory, vol. 18, pp. 411-413,
May 1971.
Since this is symmetric it can be written as a function of [14] M. J . Shensa, “The discrete wavelet transform: Wedding the i trous
and Mallat algorithms,” IEEE Trans. Signal Processing, to be pub-
cos (w): lished, Oct. 1992.
[I51 A. Croisier, D. Esteban, and C . Galand, “Perfect channel splitting
Q;/2(1 - cos2 (w/2)) by use of interpolation, decimation, tree decomposition techniques,”
= Q112(1 - cos2 (w/2)) + E 1 / 2 ((w))[sin2
~ ~ ~ (w/2)lk.
in Proc. Int. Conf. Inform. Sci.lSyst. (Patras, Greece), Aug. 1976,
pp. 443-446.
1161 R . E. Crochiere and L. R. Rabiner, Multirate Digital Signal Pro-
Clearly, this gives cessing. Englewood Cliffs, NJ: Prentice-Hall, 1983.
[I71 M. J . T. Smith and T . P. Barnwell, 111, “Exact reconstruction for
) Q I / ~ ( Y+) Y ~ E I / ~ ( Y )
Q ; / ~ ( Y= tree-structured subband coders,” IEEE Trans. Acoust., Speech, Sig-
nal Processing, vol. 34, pp. 434-441, June 1986.
where our constraint E(z) = E( - z ) translates to E l l 2 (y) [I81 F. Mintzer, “Filters for distortion-free two-band multirate filter
= E1/2(l - y), giving that E l l 2 (y) is symmetric about banks,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 33,
the point 1/2. The fact that E(z) is of degree 4m - 2 pp. 626-630, June 1985.
[19] P. P. Vaidyanathan, “Quadrature mirror filter banks, M-band exten-
gives that E l l 2 (y ) has degree 2m - 1. If we alter our sions and perfect-reconstruction technique,” IEEE ASSP Mag., vol.
notation to use zero phase polynomials throughout we 4 , pp. 4-20, July 1987.
would find E(z) = - E ( - z ) or E l l 2 (y) = -El 9 ( 1 - y), [20] P. P. Vaidyanathan and P.-Q. Hoang, “Lattice structures for optimal
design and robust implementation of two-band perfect reconstruction
which is precisely the requirement given in [ 121. QMF banks,” IEEE Trans. Acoust., Speech, Signal Processing, vol.
Further, Daubechies gives a closed form for y). 36, pp. 81-94, Jan. 1988.
In our notation this translates to the complementary filter (211 M. Vetterli, “Filter banks allowing perfect reconstruction,” Signal
Processing, vol. IO, no. 3, pp. 219-244, 1986.
to the binomial. [22] M. Vetterli and D. Le Gall, “Perfect reconstruction FIR filter banks:
Some properties and factorizations,” IEEE Trans. Acoust., Speech,
Signal Processing, vol. 37, pp. 1057-1071, July 1989.
ACKNOWLEDGMENT (231 T . Q. Nguyen and P. P. Vaidyanathan, “Two-channel perfect-recon-
struction FIR QMF structures which yield linear-phase analysis and
The authors gratefully acknowledge fruitful discussions synthesis filters, IEEE Trans. Acoust., Speech, Signal Processing,
”
with Dr. I. Daubechies, AT&T Bell Laboratories and 0. vol. 37, pp. 676-690, May 1989.
[24] M. Vetterli and C. Herley, “Wavelets and filter banks: Relationships
Rioul, CNET, Paris, who also provided the regularity es- and new results,’’ in Proc. IEEE Int. Conf. ASSP (Albuquerque, NM),
timates in Section V . They would like to thank J . Kova- Apr. 1990, pp. 1723-1726.
I
2232 I E E E T R A N S A C T I O N S O N S I G N A L P R O C E S S I N G , VOL. 40, N O . 9, S E P T E M B E R 1992
[25] C . Herley and M. Vetterli, “Linear phase wavelets: Tieory and de- [47] J.-C. Feauveau, “Analyse multirCsolution par ondelettes non orthog-
sign,” in Proc. IEEEInt. Con$ ASSP (Toronto, Canada), May 1991, onales et bancs de filtres numiriques,” Ph.D. dissertation, UniversitC
pp. 2017-2020. de Paris-Sud, 1990.
[26] M. Antonini, M. Barlaud, P. Mathieu, and I. Daubechies, “Image [48] D. E. Knuth, The Art of Computer Programming: Seminumerical Al-
coding using vector quantization in the wavelet transfc rm domain,” gorithms, vol. 11, seconded. Reading, MA: Addison-Wesley, 1981.
in Proc. IEEE Int. Con$ ASSP (Albuquerque, NM), Apr. 1990, pp. [49] A. Y. Khinchin, Continued Fractions. Chicago, IL: University of
2297-2300. Chicago Press, 1964.
(271 A. Grossman, R. Kronland-Martinet, and J . Morlet, “Reading and [50] E. W. Cheney, Introduction to Approximation Theory, second ed.
understanding continuous wavelet transforms,” in Wcvelets, Time- New York: Chelsea, 1981.
Frequency Methods and Phase Space (Lecture Notes on IPTI), J. M. [51] C. Herley and M. Vetterli, “Biorthogonal bases of symmetric com-
Combes er a l . , Eds. New York: Springer-Verlag, 191:9. pactly supported wavelets,” in Wavelets, Fractals and Fourier
[28] A. Cohen, I . Daubechies, and J.-C. Feauveau, “Biortiogonal bases Transforms, M. Farge, J. C . R. Hunt, and J. C. Vassilicos, Eds.
of compactly supported wavelets,” submitted for publication. Oxford University Press, to be published 1992.
[29] 0. Rioul, “A discrete-time multiresolution theory unifying octave- [52] R. A. Gopinath, personal communication, July 1990.
band filter banks, pyramid, and wavelet transforms,” IEEE Trans.
Signal Processing, to be published.
[30] D. Gabor, “Theory of communication,” J . Inst. Elec. .%g., vol. 93,
pp. 429-457, 1946.
Martin Vetterli (S’86-M’86-SM’90) was born in
[31] M. R. Portnoff, “Time-frequency representation of digi.al r.ignals and
Switzerland in 1957. He received the Dipl. El:
systems based on short-time Fourier analysis,” IEEE T,-an::.Acoust.,
lng. degree from the Eidgenossische Technische
Speech, Signal Processing, vol. 28, pp. 55-69, Feb. 1380.
Hochschule Zurich, Switzerland, in 1981, the
[32] R. Bracewell, The Fourier Transform and its Applicrifions, second
master of science degree from Stanford Univer-
ed. New York: McGraw-Hill, 1986.
sity, Stanford, CA, in 1982, and the Doctorat t s
[33] I. Daubechies, “The wavelet transform, time-frequency localization
Science degree from the Ecole Polytechnique
and signal analysis,” IEEE Trans. Inform. Theory, vol 36, pp. 961-
FCdkrale de Lausanne, Switzerland, in 1986.
1005, Sept. 1990. In 1982, he was a Research Assistant at Stan-
[34] N. Dyn and D. Levin, “Interpolating subdivision schemes for the
ford University, and from 1983 to 1986 he was a
generation of curves and surfaces,” School Math. S i . , Tel Aviv
Researcher at the Ecole Polytechnique. He has
Univ., 1990, preprint.
worked for Siemens and AT&T Bell Laboratories. In 1986, he joined Co-
[35] 0. Rioul, “Dyadic up-scaling schemes: simple criteria for regular-
lumbia University in New York where he is currently Associate Professor
ity,” SIAM J . Math. Anal., to be published, 1992.
of Electrical Engineering, a member of the Center for Telecommunications
[36] A. Cohen, “Ondelettes, analyses multirksolutions et traitement
Research, and Codirector of the Image and Advanced Television Labora-
numkrique du signal.” Ph.D. dissertation, UniversitC ’aris IX Dau-
tory. His research interests include multirate signal processing, wavelets,
phine, 1990.
computational complexity, signal processing for telecommunications, and
[37] M. J . T. Smith and T. P. Barnwell 111, “A procedure for designing
digital video processing.
exact reconstruction filter banks for tree structured subtiand coders,”
Dr. Vetterli is a member of SIAM and ACM, a member of the MDSP
in Proc. IEEE Int. Con$ ASSP (San Diego, CA), M a r . 1984, pp.
Committee of the IEEE Signal Processing Society and of the editorial boards
27.1.1-27.1.4.
of Signal Processing, Image Communication, and Annals of Telecommu-
[38] P. P. Vaidyanathan, “Multirate digital filters, filter banks, polyphase
nications. He received the Best Paper Award of EURASIP in 1984 for his
networks, and applications: A tutorial,” Proc. IEEF, vol. 78, pp.
paper on multidimensional subband coding, and the Research Prize of the
56-93, Jan. 1990.
Brown Bovery Corporation, Switzerland, in 1986 for his thesis, and the
[39] V. Belevitch, Classical Network Synthesis. San Francisco, CA:
IEEE Signal Processing Society’s 1991 Senior Award (DSP Technical Area)
Holden Day, 1968.
for a 1989 transactions paper with D. LeGall on FIR filter banks.
[40] P. P. Vaidyanathan and Z . Doganata, “The role of lo:,sle~~s systems
in modem digital signal processing,” IEEE Trans. E‘cluctrtion, vol.
32, pp. 181-197, Aug. 1989.
[41] A. V. Oppenheim and R. W. Schafer, Digital Signal Processing.
Englewood Cliffs, NJ: Prentice-Hall, 1975. Cormac Herley was born in Cork, Ireland, in
[42] I. Daubechies, “Orthonormal bases of compactly supported wavelets 1964. He received the B.E. (Elec.) degree from
11: Variations on a theme,” 1989, AT&T preprint. University College Cork in 1985, and the
[43] M. R. Schroeder, Number Theory in Science arid Cor,imunications, M S.E.E. degree from the Georgia Institute of
second ed. New York: Springer-Verlag, 1986. Technology, Atlanta, in 1987. Since 1990 he has
[44] T. Kailath, Linear Systems. Englewood Cliffs, NJ: ’rentice-Hall, been working towards the Ph.D degree in the
1980. Department of Electncal Engineenng, Columbia
[45] A. J. E. M. Janssen, “Note on a linear system occuning in perfect University, New York, NY.
reconstruction,” Signal Processing, vol. 18, no. 1, pp. 109-1 14, From 1987 to 1989 he was employed at Kay
1989. Elemetncs Corp., New Jersey. He is now a Re-
1461 R. Blahut, Fast Algorithms f o r Digital Signal Processit g . Reading, search Assistant at the Center for Telecommuni-
MA: Addison-Wesley, 1984. cations Research, Columbia University

Filter Banks Using Wavelets

Uploaded by

Copyright:

Available Formats

Filter Banks Using Wavelets

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Filter Banks Using Wavelets

Uploaded by

Copyright:

Available Formats

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL.

40, NO 9, SEPTEMBER 1992 2207

Wavelets and Filter Banks: Theory and Design

18353-587X/92$03.00 0 1992 IEEE

continuous function). In Section IV, results on the alge-

only. wheref(x), g(x) E L2(R).The norm is given by Ilf(x)ll;

face symbols; the asterisk * will denote Hermitian trans-

11. WAVELETS, MULTIRESOLUTION SIGNAL PROCESSING,

where w ( * )is an appropriate window like a Gaussian.

uous, but is of interest because of its simplicity), where:

and, by iterating (10) $(x) = C

03, is equivalent to the space of square integrable func-

Pictorially the situation is as shown in Fig. 5 . Fig. 5(a)

because Vo C V - , . Then 7tl4 rdz rr 2rr

and its integer translates form an orthonormal basis for

ho(n - 21) it follows

nal processing operations, and their effect upon resolution

taking the difference between the original and the predic-

The difference signal is thus (c)

1 , [ 1 2 ] . This property is potentially useful for the

F. Bases of Orthonormal Wavelets Constructed from

f " ' ( x ) = 2'/2 . mc= O ho(m) h!-"(n - 2l-l m)

n/2' Ix < (n + 1 ) / 2 ' . (29)

n/2'-' I2x - m < (n + 1)/2'-' (30)

Recall that when the filter is regular, f'"(x) tends to a

We can define also the bandpass function:

and prove that this implies orthogonality at thc: ( i -t 1)st

c h l ( n ) ( 4 ( 2 x - 21 + n), $(2x - k)) or

['1 - 1'1 ['o O]z

= 0. Hp(z2) = 2-' * H,(z)

a single nonzero odd-indexed coefficient: cients, that is

+(x) = 2Il2 c ho(n)

That &x) and 4 ( x ) are orthogonal with respect to integer

I v . ALGEBRAIC S T R U C T U R E OF FIR S O L JTIONS Fact 4.1: Given a (x) and b (x)

C . Continued Fraction Expansions

Remark that the Dj(z)’s are known, and DN is ;I scalar V . DESIGNRESULTS

Like (78) this relates successive convergents of a contin-

giving r6 = (3/2, -9, 19)/128.

TABLE 111 TABLE IV

as in the previous section. This solution, which has de-

TABLE V PRFB's. If compact support is not desired, similar tech-

Ho(z) 1. Biorthogonal Pyramid

2. Proof of Proposition 3.3

You might also like