Computing with Daubechies' Wavelets

Adri B. Olde Daalhuis

Abstract. In the first two sections of this paper we describe a construction of

the compactly supported Daubechies' wavelets. In the third section we show
how a Discrete Wavelet Transform of a function can be obtained. We end this
paper with some interesting numerical illustrations.

§1 Wavelets with finitely many non-zero filter coefficients

In the theory of discrete wavelets the equation
00

<f>(x) = V2 L hn<f>(2x - n) (1.1)

n=-oo

plays a fundamental role. With the farther wavelet </>, satisfying this equation, we
have the accompanying mother wavelet 'ljJ defined by
00

'lj;(x) := ¥2- L (-l)nh1-n<f>(2x - n). (1.2)

n.=-oo

In Daubechies [1] for the first time a construction is given of wavelets resulting in
compactly supported orthonormal wavelet bases. (See also Heijmans [this volume,
§4].)
It is not difficult to put conditions on hn of (1.1) in order to obtain compactly
supported functions <f> and 'lj;. We only need to have just a finite number of {hn}
that are different from zero. In this section we give a proof of this property.

W&velets: An Elementary Treatment of Theory and Applications 93

Tom H. Koornwinder (ed.), pp. 93-105.
Copyright @ 1993 by World Scientific Publiarung Co., Inc.
All rights of reproduction in any form reaerved.
ISBN 981-02-1388-3
94 A. B. Olde Daalhuis

We use the fact that the solution <P of equation (1.1) can be constructed by the
following iteration scheme. Let

7Jo(x) = { ~ as-~<x<~
elsewhere.
(1.3)

That is, 7JO(x) is the characteristic function of[-~,~]. Next we define a sequence
of functions 7}1, l = 1, 2, ... , by writing
00

T/z(x) = v'2 L hn7Jl-l (2x - n). (1.4)

n=-oo

Then we have
<f>(x) = !-+co
lim r)l(x). (1.5)

The reader is referred to [1] for a proof of this constructive procedure, which may
also be used to draw pictures of solutions of (1.1).
Now, let just a finite number of filter coefficients hn of ( 1.1) be non-zero. That
is, assume that we have two integer numbers N_, N+ and that

n<N_, (1.6)
n>N+·

It is easily verified that the functions 7JL defined by (1.3) and (1.4) have compact
support. We have in fact
supp(711) = [Nz,-, Nz,+L
with
M - t
o,+ - ~'
Nz,+ = 2(N1-1,+ + N+).
Thus we have
Nz,---+ N_
as l ·-> oo,
Nz,+--+ N+
and it follows that
supp(<f>) c [N-,N+J, (1. 7)

and with ( 1.2)

supp(,,P) c [~(1 - N+ + N_), 1(1 + N+ - N_)]. (1.8)

In a recent paper ofLemarie-Rieusset & Malgouyres [3], it is proven that the support
of <:/> is indeed a connected interval.
Computing with Wavelets
95

§2 Daubechies' construction of compactly supported wavelets

Daubechies' construction of compactly supported wavelets starts with the function

We prescribe that all hn E lR, and just a finite number are non-zero. The function
H(€) is of fundamental importance in the theory. It generates the function ~(~):
00

~(€) = IT H(ri €),

j=l

where the function'(/;(€) denotes the Fourier transform of the function <P(x):

So, when H(€) is known, '(/;(€) and, hence, rjJ(x) can be constructed, and with (1.2)
the mother wavelet 1/J(x) can be constructed.
The construction uses the following important theorem:

Theorem 1. (Daubechies) Let hn be a sequence such that

(i) :En lhnllnl" < oo for some e > 0,
(ii) :En hn-2khn-2l = bk!,
(iii) :En hn = .J2.
i . e
Suppose also that H(€) = 2-2 :En hne-m can be written as

(2.1)
n

where
(iv) :E lfnllnl" < oo for some e > 0,
n I"
(v ) supeElR .r -inel < 2N-l .
£....,n Jne
Define

00

j=l

n
96 A. B. Olde Daalhuis

Then the </>ik(x) = 2-i/2<1>(2-ix - k) define a multiresolution analysis, and the

{,,Pj k} are the associated orthonormal wavelet basis.
This theorem and the following theorems are proven in [1]. The number Nin (2.1)
is crucial: when N is large the wavelet has interesting properties with respect to
approximations with the mother wavelets {iJ!jk}·
Remark. Starting with finitely many hn, we obtain finitely many fn, and (i) and
(iv) of Theorem 1 are obviously fulfilled.

Examples
1
(1) ho=h1 =- and
v'2

h _I+./3 h _3+./3 h _3-./3 h =1-J3

(2) 0 - 4\"'2 ' 1 - 4v'2 ' 2 - 4y'2 ' 3 4y'2 '
H(~) = [~(1 + e-ie)J 2 4[1 + v'3 + (1 - v'3)e-ie].
In the first example the function ef;(~) can be computed easily. We have
00 1 -i~
~m =II Hi+ exp(-iri~)) =
j=l
--~
i
.

Thus the father wavelet </J(x) and the associated mother wavelet iji(x) are the Haar
wavelets:
1 asO<x<!
</J(x) ={1 as 0 < x < 1 iji(x) = { -1 !
as < x < 1
0 elsewhere
0 elsewhere.
The first example is speciai with respect of symmetry because:

Theorem 2. (Daubechies) The Haar basis (associated to the above example (1)) is
the only orthonormal basis of compactly supported wavelets for which the associated
function <P has a symmetry axis.
The construction of the compactly supported wavelets from the function H ( ~) starts
by writing
(2.2)
The function Q(x) is a polynomial with real coefficients. It is easy to see that (ii)
of Theorem 1 can be rewritten as

(2.3)
Computing with Wavelets 97
With (2.2) we have
(2.4)
Because of the coefficients of Q(x) being real, we have Q(e-•€) = Q(ei€), and it
follows that IQ(e-•€)12 can be rewritten as a polynomial in cosE", or, equivalently,
as a polynomial in sin2 t~- So, with y = sin 2 H,
there is a polynomial P(y) such
that
(2.5)
and
P(y) ~ 0 Vy E [O, 1]. (2.6)
With this new polynomial, (2.3) becomes

YN P(l - y) + (1 - y)N P(y) = 1. (2.7)

Thus from a function H(~), fulfilling (2.2) and (2.3), we obtain a polynomial P(y),
fulfilling (2.6) and (2.7). With the next Lemma it follows that from a polynomial
P(y), fulfilling (2.6) and (2.7), we can obtain a function H(~), fulfilling (2.2) and
(2.3), and this H(O has real fin.

Lemma 1. (Riesz) Let A(~) = 'l:~=O a., cos~' with a., E lR, be a positive
trigonometric polynomial. Then there exists a trigonometric polynomial B(~)
'l:::=O b.. e-inf;, with bn E R, such that

The proof of this Lemma is given in [1], and it is constructive. We omit the proof.
A special solution of (2.6)-(2.7), which is given in [l], is

PN(Y) = ~ (N - ~ + j)yi. (2.8)

j=O J
Because of all coefficients of PN(Y) being positive, it fulfills (2.6), and the following
relations, which were found by Tom Koornwinder, show that PN(Y) fulfils (2.7).

PN(Y) = N(2NN-1) (1 - y)-N i l tN-1(1- t)N-ldt,

PN(l - y) = N(2N; l)y-N foy tN-1(1- t)N-ldt.

A general solution of (2.7) fulfils
P(y) = (1 - y)-N -yN (1 - y)-N P(I - y),
 A. B. Olde Daalhuis

thus PN(Y) fulfils this relation, and this gives

P(y) - PN(Y) = yN (I - y)-N [PN(l - y) - P(l - y)].

So, a general solution of (2.7) is of the form

P(y) = PN(Y) + YN R(y), (2.9)

with R(y) a polynomial that fulfils

R(l - y) + R(y) = 0. (2.10)

Thus P,v (y) is the polynomial-solution of (2. 7) of minimal degree, and the polyno-
mial R(y) has to be antisymmetric with respect to y = ~- Condition (2.6) gives
some extra restrictions on R(y).
To summarize, we have the following explicit characteriz ation of all solutions
H(f.) of (2.3), corresponding to only finitely many non-zero hn. In the following
theorem we have replaced R by a polynomial which is antisymme tric with respect
toy= 0.

Theorem 3. Any trigonometric polynomial solution of (2.3) is of the form

where NE JN, N? 1, and where Q is a polynomial such that

where R is an odd polynomial, with some extra restrictions.

For obtaining the Daubechies' wavelets, we choose R = 0. Thus for fixed N the
Daubechies' wavelets correspond to the trigonometr ic polynomial s H(f.) of minimal
degree. With this fixed N we obtain from the polynomial PN (y) the polynomial s
Q, H, and the numbers hn. Now we shall show that these hn, H, and Q fulfill the
conditions of Theorem 1.
The conditions (i) and (iv) are obviously fulfilled, and (ii) is equivalent to (2.3),
thus (ii) is fulfilled. The special form (2.2) of H(f.) gives that H(n-) = 0. Thus with
(2.3) we have 1 = IH(O)l 2 = Hl::>n) 2 , and (iii) is fulfilled.
n
Condition (v) is IQ(e-i~)I < 2N- 1 , and it is fulfilled because of
Computing with Wavelets 99

In the construction (via the proof of Lemma 1) of Q from IQl 2 , we obtain

all zeros of Q inside the unit circle, and this determines Q unambiguously, up to a
phase factor eiK{, K E 7l. We fix this phase factor so that Q contains only negative
frequencies, starting from zero, i. e.,
N-1
QN(e-i{) = L qne-ine with qo # O. (2.11)
n=O

Exrunples

(3)

( 4)
Q3(~) = H1+v'1o+Js+2Ko + 2(1- v'i.O)e-ie

+ ( 1 + vlo - Js + 2Ko)e- ie].

2

With this QN we have

N-1
HN(~) = [!(1 + e-ie)]N L qncin{
n=O
(2.12)
1 2N-! .
= 2-2 L hne-in(
n=O
Let us denote the corresponding</;, 'ljJ functions by </;N, 'I/JN· The theory of Section
1 gives
supp(r/;N) = [0, 2N - 1], supp('l/JN) = [-N + 1, N].
For N = 2 and N = 3 the associated hn can be calculated exactly, and for N ~ 10
Daubechies gives the numerical values of the hn in [l).
When N increases with 1, the number of non-zero hn increases with 2. And
when N increases the smoothness of </;N and 'I/JN increases:

Theorem 4. (Daubechies) There exists,\> 0 such that, for all NE JN, N :'.'.: 2,

</;N,1/JN E c>..N.

In Daubechies [2, p. 226] it is proven that for large N the optimal >. is
ln3
,\ = 1 - ln4 "'0.2075.
100 A. B. Olde Daalhuis

§3 Computing with the compactly supported Daubechies' wavelets

In this section we show how a Discrete Wavelet Transform of a function can be
obtained. We will use the Daubechies' wavelet <PN of the previous section, and,
again, the associated functions are </>jk(x) = 2-312<f>N(rix - k). A function will be
represented by a finite signal a= (ao, · · ·, aM-i)T. As an intermediate expression
we introduce A= L,f:,(j1 aj</>Kj, with M = 2K. Then the Discrete Wavelet Trans-
form of a appears to be constituted by 2K - 2 coefficients of the expansion of A in
the orthonormal 1/Jnm basis.
We choose N fixed, and with the filter-coefficients h.,. of <P N we define filters
HN, GN: i 2 - z2
00

(HNii)k = L
l=-oo
h1-2ka1, (3.1)

00

(GNii.)k = L 9L-2ka1, (3.2)

l=-oo

with again g,.. = (-l)"h1-n·

Now we let these filters work on the finite signal a= (a 0 , • • ·, aM-i)T, with M
even. In matrix-form HN is

(3.3)

ho h1
The matrix-form of GN is the same, with h,.. replaced by g,,.. Filtering with these
!M x M-matrices will cause edge effects for.N > 1. For eliminating these edge
effects, we make these matrices periodic in the following way

(3.4)
ho

ho h1
Computing with Wavelets 101

with 9N the same as 1-lN with h.,. replaced by g.,.. This is the same as leaving the
filters HN and GN unchanged and making the signal a periodic to a l 00 -vector.
The total filtering is now

(3.5)

The matrix [ ~:] M is orthonormal, i. e.,

(3.6)

This follows directly from the equations

00

L hn-2khn-2l = 8kz, (3.7)

n=-oo

00

L hn-2k9n-2l = 0, (3.8)
n=-oo

where (3.7) is condition {ii) of Theorem 1, and {3.8) follows from

00 00

L hn-2k9n-2l = L < </>1k, </>On >< </>0n, '!/J11 >=< </>1k, '!/Ju >= 0.
n=-oo n=-oo

We define

(3.9)

and we choose M = 2K, K E lN*. Let A E Vx such that

ak = ako =< A,</>Kk >. (3.10)

Then we define

akj =< A,</>K-j,k >, dkj =< A,'!/JK-j,k >. (3.11)

102 A. B. Olde Daalhuis

The Discrete Wavelet Transformation becomes (for K = 4)

ao,o ao,1 ao,2 ao,3

ai,o a1,1 a1,2 a1,3
a2,0 a2,1 a2,2 do,3
a3,o a3,1 ~ di,3
a4,0 a4,1 do,2 do,2
as,o as,1 dl,2 dl,2
as,o
a1,o [~:L
----+
as,1
a11 [~; J.
--+
d2,2
~
[~; J.
---->
d2,2
d3,2 (3.12)
as,o do,1 do,1 do,1
ag,o dl,l d1,1 d1,1
a10,o d2,l d2,1 d2,l
au,o d3,1 d3,1 d3,1
a12,o d4,l d4,l d4,l
a13,o ds,1 ds,1 ds,1
a14,o d6,1 d6,l d6,l
ais,o d1,1 d1,1 d1,1

So, for general K, the Discrete Wavelet Transformation is built up by simple

matrix-operations with orthonormal matrices. The special form of these matrices
make these matrix-operations, and the Discrete Wavelet Transform for general K,
easy to program. And the whole process is simple to invert. Notice that the
transformation is a process that transforms a signal of length 2K into a vector of
length 2K.
The djk are some of the coefficients of the expansion of A in the orthonor-
mal 1/Jjk basis. The remaining ao,K-1 and a 1,K-l are called the "mother-function
coefficients".

§4 What do Daubechies' wavelets look like,

and how do they work on signals?
The illustrations in this section are made on a Macintosh-II using a Pascal program,
which is based on the program given in Press [4]. In the illustrations, the length
of the signals M = 1024. First we show some figures of the rf> N and 1/J N for some
different values of N. Notice that a signal of 1/JN can be obtained by starting at the
right hand side of scheme (3.12) with a vector 8; = {8ikH';;t, where i E 2, · · ·, 1024
is fixed. This follows directly from (3.11). The figures are shown in Figures 1, 2,
and 3.
Notice that, according to Theorems 2 and 4, the smoothness of the wavelets
increases with N, and only for N = 1 the ef> N has a symmetry axis. The inverse
discrete wavelet transform of 810 + 8ss is shown in Figure 4.
Computing with Wavelets 103

Figure 1. Father (left) and mother wavelet for N = 1, the Haar functions.

Figure 2. Father (left) and mother wavelet for N = 2.

Figure 3. Father (left) and mother wavelet for N = 6.

Figure 4. The inverse discrete wavelet transform of 610 + 05g for N = 2.

104 A. B. Olde Daalhuis

Figure 5. The sinus-signal.

Since 10 lies early in the hierarchical range 9-16, that wavelet lies on the left
side of the picture. Since 58 lies in a later (smaller scale) hierarchy, it is a narrower
wavelet.
The discrete wavelet transforms of the sinus-signal (see Figure 5) are calculated
for N = 1 and N = 10. They are shown in Figures 6 and 7.

l.
Figure 6. Discrete wavelet transform of the sinus-signal for N = 1.

Figure 7. Discrete wavelet transform of the sinus-signal for N = 10.

The amplitudes of the transform for N = 10 are more concentrated at the left
part of the signal then the amplitudes of the transform for N = 1. This can be
explained by the smoothness of the underlying wavelets. The wavelets for N = 10
match better to the smooth signal. In other cases, when the original signal is more
singular, the wavelets for N = 1 match better to that signal.
Now we can truncate these wavelet transforms. For N = 1 there are 18 co-
efficients that have amplitudes larger then 0.05 times the maximum amplitude of
the transform. We set the remaining amplitudes to zero. With this kind of data-
reduction we have to record both the values and the positions of the non-zero
Computing with Wavelets 105

coefficients. Thus in the case of N = 1 we reduce to a vector of length 36. And in

the case of N = 10 we reduce to a vector of length 12. The following two pictures
show the result of truncations of the original signal from the two inverse discrete
wavelet transforms of the truncated vectors.

Figure 8. Original signal minus approximation signal for N = 1.

Figure 9. Original signal minus approximation signal for N = 10.

&i, when we start with a smooth signal, the data-reduction with N = 10 is much
better.
Remark. It is very important that vectors in wavelet space be truncated according
to the amplitude of the components, not their position in the vector. Keeping the
first 16 components of the vector would give an extremely poor approximation to
the original signal.

References
1. Daubechies, I., Orthonormal bases of compactly supported wavelets, Comm.
Pure and Appl. Math. 41 (1988), 909-996.
2. Daubechies, I., Ten Lectures on Wavelets, CBMS-NSF, Regional Conference
Series in Applied Mathematics # 61, SIAM, Philadelphia, Pennsylvania, 1992.
3. Lemarie..Rieusset, P.G. and G. Malgouyres, Support des fonctions de base dans
une analyse multi-resolution, C. R. Acad. Sci. Paris Ser. I Math. 313 (1991),
377-380.
4. Press, W.H., Wavelet transforms, Harvard-Smithsonian Center for Astro-
physics, No. 3184, 1991, preprint.

Adri B. Olde Daalhv.is

CW!
P.O. Box 4079
1009 AB Amsterdam
The Netherlands
IW<i@cwi.nl