QUADRATURE FORMULAE AND ASYMPTOTIC ERROR

EXPANSIONS FOR WAVELET APPROXIMATIONS

OF SMOOTH FUNCTIONS
WIM SWELDENSy z AND ROBERT PIESSENSy
Abstract. This paper deals with typical problems that arise when using wavelets in numerical
analysis applications. The rst part involves the construction of quadrature formulae for the cal-
culation of inner products of smooth functions and scaling functions. Several types of quadratures
are discussed and compared for dierent classes of wavelets. Since their construction using mono-
mials is ill-conditioned, also a modied, well-conditioned construction using Chebyshev polynomials
is presented. The second part of the paper deals with pointwise asymptotic error expansions of
wavelet approximations of smooth functions. They are used to to derive asymptotic interpolating
properties of the wavelet approximation and to construct a convergence acceleration algorithm. This
is illustrated with numerical examples.
Key words. wavelet, multiresolution analysis, quadrature formula, asymptotic error expansion,
convergence acceleration, numerical extrapolation
AMS subject classications. 65D32, 42C05, 41A30

1. Introduction.
1.1. Multiresolution analysis. We will brie y review wavelets and multireso-
lution analysis. For detailed treatments, one can consult [4, 8, 16, 18]. A multires-
olution analysis of L2 (R) is dened as a set of closed subspaces Vj with j 2 Z that
exhibit the following properties:
1. Vj  Vj+1 ,
2. v(x) 2 Vj , v(2x) 2 Vj+1 and v(x) 2 V0 , v(x + 1) 2 V0 ,
1
+[ 1
+\
3. Vj is dense in L2(R) and Vj = f0g,
j =?1 j =?1
4. A scaling function '(x) 2 V0 exists such that the set f'(x ? l) j l 2 Zg is a
Riesz basis of V0 .
Consequently, a sequence (hk) 2 `2 (Z) exists such that the scaling function satises
a renement equation
X
(1.1) '(x) = 2 hk '(2x ? k):
k
The set of functions f'j;l (x) j l 2 Zg with 'j;l (x) = 2j=2 '(2j x ? l), is a Riesz basis
of Vj . A complement space of Vj in Vj+1 is denoted by Wj , so Vj+1 = Vj  Wj , and
consequently
+1
M
Wj = L2 (R):
j =?1
 Manuscript recieved by the editors ??, accepted for publication ??
y Departement Computerwetenschappen, Katholieke Universiteit Leuven, Celestijnenlaan 200 A,
B 3001 Leuven, Belgium
z Department of Mathematics, University of South Carolina, Columbia SC 29208. This author is
Research Assistant of the National Fund of Scientic Research Belgium, and was partially supported
by ONR Grant N00014-90-J-1343.
1
2 WIM SWELDENS AND ROBERT PIESSENS

The complementary spaces are chosen such that

v(x) 2 Wj , v(2x) 2 Wj+1 and v(x) 2 W0 , v(x + 1) 2 W0 :
A function (x) is a mother wavelet if the set of functions f (x ? l) j l 2 Zg is a
Riesz basis of W0 . Since the mother wavelet is also an element of V1 , a sequence
(gk ) 2 `2 (Z) exists such that
X
(1.2) (x) = 2 gk '(2x ? k):
k
The set of wavelet functions f j;l (x) j l; j 2 Zg, with j;l (x) = 2j=2 (2j x ? l), is now
a Riesz basis of L2(R).
In the case of an orthogonal multiresolution analysis, the set of functions f'j;l (x) j
l 2 Zg is an orthonormal basis of Vj and the set f j;l (x) j j; l 2 Zg is an orthonormal
basis of L2(R). The orthogonal projection operators onto Vj and Wj , denoted as Pj
and Qj respectively, can then be written as
X X
Pj f (x) = h f; 'j;l i 'j;l (x) and Qj f (x) = h f; j;l i j;l (x):
l l
In the biorthogonal case, dual functions '~j;l (x) = 2j=2 '~(2j x ? l) and ~j;l (x) =
2j=2 ~(2j x ? l) exist such that
h 'j;l ; '~j;l i = l?l and h j;l ; ~j ;l i = j?j l?l for j; j 0 ; l; l0 2 Z:
0 0 0 0 0 0

The projection operators now can be written as

X X
Pj f (x) = h f; '~j;l i 'j;l (x) and Qj f (x) = h f; ~j;l i j;l (x):
l l
They are not orthogonal in general. The dual scaling function and wavelet satisfy
X
(1.3) '~(x) = 2 h~ k '~(2x ? k); ~(x) = 2 X g~k '~(2x ? k);
k k
and
X X
(1.4) '~(2x ? k) = hk?2l '~(x ? l) + gk?2l ~(x ? l):
l l
Dene now the Fourier transform of a function f (x) as
Z +1
f^(!) = f (x) e?i!x dx:
?1
Equations (1.1) and (1.2) can then be written as
(1.5) '^(!) = H (w=2) '^(!=2) and ^(!) = G(w=2) '^(!=2);
where H and G are 2-periodic functions given by
X X
H (!) = hk e?ik! and G(!) = gk e?ik! :
k k
 WAVELET QUADRATURES AND ERROR EXPANSIONS 3
Similar equations hold for the dual functions. A necessary condition for biorthogo-
nality is then

(1.6) 8! 2 R : H (!) H~ (!) + G(!) G~ (!) = 1
H (!) H~ (! + ) + G(!) G~ (! + ) = 0:
Given the coecients j;l = h f; '~j;l i of a function in Vj , one can nd its coef-
cients in the bases of the spaces Vj?1 and Wj?1 with decomposition formulae that
can be derived using (1.3),
p X p X
j?1;l = 2 h~ k?2l j;k ; and j?1;l = h f; ~j?1;l i = 2 g~k?2l j;k :
k k
The inverse step involves a reconstruction formula that can be derived from (1.4),
p X p X
j;k = 2 hk?2l j?1;l + 2 gk?2l j?1;l :
l l
When applied recursively, these formulae dene a transformation, the fast wavelet
transform [16, 17].
Examples:. Several families of orthogonal wavelets exist and are described in [2,
15, 18]. A well known example of orthogonal wavelets with compact support was
constructed by Ingrid Daubechies in [7]. Compactly supported biorthogonal wavelets
are described in [5]. The wavelets constructed in [4, 22] are semi-orthogonal wavelets,
which means that the wavelets that belong to one subspace Wj are not orthogonal but
the subspaces Wj are still mutually orthogonal. The scaling functions and wavelets
are compactly supported splines, but the dual functions do not have compact support.
1.2. Wavelets and Polynomials. The moments of the scaling function and
wavelet are dened as,
Z +1 Z +1
Mp = xp '(x) dx and Np = xp (x) dx with p  0:
?1 ?1
The scaling function is usually normalized with M0 = 1. Similar denitions hold for
the dual functions. The moments of a scaling function can be calculated from the
coecients hk with a p{term recursion relation,
p  
(1.7) Mp = 2p ? 11 Xp mM ;
i p?i
i=1 i
where the mi are the discrete moments of the sequence (hk ),
X
mi = hk ki:
k
The number of vanishing dual wavelet moments is denoted by N where N is at least
1,
N~p = 0 for 0  p < N; and N~N 6= 0:
This is equivalent with
^~(p) (0) = 0 for 0  p < N;
4 WIM SWELDENS AND ROBERT PIESSENS

and, since '^~(0) = M~ 0 6= 0, also with

G~ (p)(0) = 0 for 0  p < N:
The sequence (~gk ) thus has N vanishing discrete moments. Using equation (1.6) we
see that this is also equivalent with
H (p) () = 0 for 0  p < N:
Consequently,
(1.8) ip'^(p) (2k) = k Mp for 0  p < N;
and, by the Poisson summation formula,
X
(1.9) (x ? l)p '(x ? l) = Mp for 0  p < N:
l
By rearranging the last expression we see that any polynomial with degree smaller
than N can be written as a linear combination of the functions '(x ? l) with l 2 Z.
The coecients in the linear combination themselves are polynomials in l. More
precisely, if p denotes the set of polynomials of degree p,
X X
(1.10) 8A 2 N ?1; 9B 2 N ?1 : A(x) = B (l) '(x ? l) = B (x ? l) '(l):
l l
This last equation holds because left and right hand side are polynomials that match
at every integer.
The number of vanishing dual moments determines the order of convergence of
the wavelet approximation of smooth functions. More precisely, if f (x) 2 C N , then
[11, 20, 21],
(1.11) kf (x) ? Pn f (x)k = O(hN ) with h = 2?n:
The conditions (1.8) are usually referred to as the Strang{Fix conditions.
1.3. Contents of the paper. This paper deals with typical problems that arise
when using wavelets in numerical analysis applications. Section 2 adresses construc-
tions of quadrature formulae to approximate the inner products h f (x); '~(2j x ? l) i .
We discuss and compare the accuracy of dierent quadrature formulae for several
types of wavelets.
In section 3 we develop pointwise asymptotic expansions for the error f (x) ?
Pn f (x) in powers of h, with h = 2?n. We show how the expansion can be used to
derive asymptotic interpolation properties and to accelerate the convergence. Also we
point out a simple trick to improve the accuracy of the rst term of the expansion.
2. Quadrature formulae.
2.1. General idea. The idea of a quadrature formula is to nd weights wk and
abscissae xk such that
Z +1 r
X
(2.1) f (x)'(x) dx  Q[f (x)] = wk f (xk):
?1 k=0
 WAVELET QUADRATURES AND ERROR EXPANSIONS 5
Definition 2.1. The degree of accuracy of a quadrature formula is q if it yields
the exact result for every polynomial of degree less than or equal to q.
The degree of accuracy determines the convergence order as follows: if f (x) be-
longs to C q+1 , then

(2.2) n;l ? 2?n=2 Q[f (2?nx)] = O(hq+1 ) with h = 2?n:

n;l
This can easily be seen using the Taylor expansion. It also follows from Peano's
theorem in [10]. Note that we do not impose any regularity conditions on '(x). The
number of abscissae r determines the eciency of a quadrature formula since the
number of function evaluations and algebraic operations is proportional to r. The
quadrature formula is usually constructed by demanding that
Q[xi] = Mi for 0  i  q;
which leads to an algebraic system. In case the abscissae xk are xed, this system is
linear in the unknowns wk . More ecient quadrature formula can be constructed by
also treating the abscissae as unknowns, cf. Gauss quadrature formulae.
In connection with a multiresolution analysis, the quadrature formula will be used
at a xed, nest level n to approximate the n;l . The coecients on the coarser levels
j;l and j;l with j < n, can then be calculated using the fast wavelet transform. Note
that this implies that the error will be O(2?n(q+1) ) for all coecients, independent of
their level j .
The abscissae should be chosen equidistant because in some applications the func-
tion f (x) is only known at equidistant abscissae and because then quadrature formulae
for neighboring coecients can share common points. Therefore we let the abscissae
be of the form k2?s +  , where  is an unknown.
Comparing equations (1.11) and (2.2), we see that the degree of accuracy should
be at least equal to N ? 1, otherwise the quadrature formula will ruin the convergence
order of the wavelet approximation.
2.2. Trapezoidal rule. A simple quadrature formula is the trapezoidal rule,
where
X
(2.3) Q[f (x)] = '(k) f (k):
k
In general the application of this rule is limited because it only has a degree of accuracy
equal to 1. Here however, the following preposition holds:
Preposition 2.2. If the scaling function satises the Strang-Fix condition (1.8),
the degree of accuracy of the trapezoidal rule (2.3) is equal to N ? 1.
This easily seen from equation (1.9) for x = 0. The trapezoidal rule can be used
in a multiresolution analysis. However in general it is not very ecient. In case '(x)
is not compactly supported, the sum in (2.3) has to be broken of which usually leads
to a large number of abscissae. But also when ' is compactly supported, this formula
is not really ecient: the Daubechies' orthogonal scaling functions have a support
length of 2N ? 1, such that r = 2N ? 2 = 2q while even with xed abscissae we can
achieve r = q + 1. Only in the case of cardinal B-splines, the trapezoidal rule is useful
because here r = N ? 1 = q.
6 WIM SWELDENS AND ROBERT PIESSENS

2.3. One point formulae. Since the integral of the scaling function is 1, we
can write a one point formula as Q[f (x)] = f (x1). Evidently, if x1 = M1 , the degree
of accuracy is equal to 1. In the case of orthogonal wavelets, the following theorem
holds:
Theorem 2.3. If '(x) is an orthogonal scaling function with N > 1, then
M2 = M21 .
Proof. Dene
m = h x; '(x) '(x ? m) i :
Because of the orthogonality holds that
?m = h x ? m; '(x ? m) '(x) i = m :
Consequently
X X
0= m m = h x; '(x) m '(x ? m) i ;
m m
and, since N > 1,
X
m '(x ? m) = x ? M1 :
m
Combining the last two equations yields M2 ? M21 = 0.
Note: This theorem was proven independently in case of Daubechies' scaling functions
in [14].
This means that for orthogonal scaling functions the degree of accuracy of the one
point quadrature is 2. Consequently it can be used in case N  3.
In [8, 9] Ingrid Daubechies constructed orthogonal scaling functions with compact
support that have N ? 1 vanishing moments,
(2.4) Mp = 0 for 1  p < N;
where again N is the number of vanishing wavelet moments. These wavelets were
called coi ets after Ronald Coifman who asked for their construction. We see from
(1.9) that they also satisfy
X
(2.5) kp'(k) = p for 0  p < N:
k
In this case the one point quadrature formula with x1 = 0 immediately has a degree
of accuracy of N ? 1. This formula was used in numerical analysis applications in [3].
The coi ets have the disadvantage that their support width is 3N ? 1. This results
in a 50% increase in computational cost for the fast wavelet transform in comparison
with the original Daubechies' wavelets with the same number of vanishing moments
whose support width is only 2N ? 1.
2.4. Practical aspects. In applications such as signal and image processing,
usually discrete samples al are given. Then there are several ways to start the mul-
tiresolution analysis. First one can construct a function a(x) 2 Vn [17],
p X
a(x) = h al 'n;l (x) with h = 2?n:
l
 WAVELET QUADRATURES AND ERROR EXPANSIONS 7
We can see that the continuous function a(x) will in a way \follow" the discrete
samples al . The quadrature formula can help us to nd a relationship between the
function a(x) and the discrete samples ai . Indeed, using the biorthogonal notation,
p p
h al = h a; '~n;l i and h a; '~n;l i = h [a(h (M1 + l)) + O(ht)];
so
al = a(h (M1 + l)) + O(ht):
This means that a(x) satises a quasi-interpolating property. Here t = 2 in general,
t = 3 for orthogonal wavelets and t = N for coi ets.
Secondly, one can consider the samples al as function evaluations, al = f (hl). This
corresponds to the one point quadrature formula with x1 = 0. Then the following
theorem is important:
Theorem 2.4. If f (x) 2 C N with f (i) (x) bounded for i  N , then (h = 2?n)
X X
f (hl) '(2nx ? l) = '(l) f (x ? hl) + O(hN ):
l l

Proof.
?1
NX iX
f (i) (x) (?ih! )
X
f (hl) '(2nx ? l) = (2n x ? l)i '(2nx ? l) + O(hN )
l i=0 l
NX?1 (?h) i X
= f (i) (x) i! li'(l) + O(hN )
i=0 l
X
= '(l) f (x ? hl) + O(hN )
l

This theorem states that taking function evaluations as coecients results in

approximating a dierent function f~n (x) = l '(l) f (x ? hl) with an error of O(hN ).
P

This function can be seen as a \blurred" version ofPf (x) as '(l) is a low pass lter.
Now, f~n (x) will converge to f (x) for n ! 1 since k '(k) = 1. However, in general
this convergence is only O(h). In the case of the coi ets we see from (2.5) that
f~n (x) = f (x) + O(hN ).
Finally, one can consider the samples al as inner products,
al = 2n h f (x); '(0) (2nx ? l) i ;
where '(0) (x) is the box function [0;1]. One can then construct a recursive multires-
olution scheme as follows:
(0) = 2?n=2 a and  (n?j +1) = p2 X h
n;l (n?j )
l j ?1;l k?2l j;k :
k
Here is
(m) = 2j=2 h f (x); '(m) (2j x ? l) i ; with '(m) (x) = T m '(0) (x);
j;l
8 WIM SWELDENS AND ROBERT PIESSENS

and T an operator dened as

X
Tg(x) = 2 hk g(2x ? k):
k
(m)
!1 ' (x) = '(x). This scheme is applicable for a wide
It is proven in [7] that mlim
range of functions '(0) (x) with integral 1.
2.5. Multiple point formula.
2.5.1. Construction scheme. Since the degree of accuracy of a one point for-
mulae is limited, we construct multiple point integration formulae. In this section we
assume that '(x) has compact support [0; L], and satises a renement equation (1.1)
with L + 1 non-zero coecients hk . Although the construction is general, we focus on
scaling functions with compact support since in this case we have the extra limitation
that the abscissae should fall inside the integration interval. We construct an r point
quadrature formula with xk = dk ?  , dk = (k ? 1)2s and (r ? 1)2s ? L    0.
The range of the shift  is determined by the requirement that no abscissae should
fall outside the integration interval. In order to have a non-zero range for the shift  ,
the parameters r and s should be chosen such that (r ? 1)2s < L. This technique to
construct quadrature formulae is also used in [3] but there the shift  is given a xed
value.
Since there are r + 1 unknowns f; w1 ; : : : ; wr g, one can try to achieve a degree
of accuracy r. This results in the following system which is nonlinear in the unknown
,
r
X
(2.6) wk [dk ?  ]i = Mi 0  i  r:
k=1
The value of the shift  can be determined using the product polynomial (x). This
polynomial is dened as
r
Y r
Y r
X
(x) = (x ? xk ) = (x +  ? dk ) = pi( )xi;
k=1 k=1 i=0
where pi ( ) is a polynomial of degree r ? i. Since the degree of accuracy is r, the
quadrature formula gives the exact result for the product polynomial (x) so
r
X
0 = Qr [(x)] = pi ( ) Mi = ?( ):
i=0
The latter expression is a polynomial of degree r in  . For the quadrature formula to
exist, ?( ) must have a root in the interval [(r ? 1)2s ? L; 0]. However, the existence
of such a root is not theoretically guaranteed. If there is no root in this interval, an
arbitrary value for  must be chosen and one degree of accuracy is lost. Once  is
determined, the weights are the solution of the linear system formed by r equations
of (2.6). In order to construct ?( ) we write
r?j !
r?i
X r
X X
pi( ) = pi;j  j and ?( ) = Mi pi;j  j :
j =0 j =0 i=0
 WAVELET QUADRATURES AND ERROR EXPANSIONS 9
The coecients pi;j are symmetric (pi;j = pj;i) since the product polynomial is sym-
metric in  and x and can be found as p(i;jr) where
m
Y m m
X X?i
(m) (x) = (x +  ? dk ) = (m)  j xi :
pi;j
k=1 i=0 j =0
An algorithm to calculate the pi;j can be derived by writing
(m) (x) = (x +  ? dm ) (m?1)(x);
and identifying the coecients of the powers of x and  . A disadvantage of this
construction is that the system of equations (2.6) is ill-conditioned if r is large. In the
construction of Q13 for the Daubechies scaling function with N = 7, the condition
number of the linear system for the weights is 5:1016!
2.5.2. Modied construction. The ill-conditioning problem of the previous
section can be overcome if we use the basis of Chebyshev polynomials. This technique
is also used successfully in [12, 19]. The Chebyshev polynomial Tn (x) of degree n is
dened by T0 (x) = 1, T1 (x) = x and Tn (x) = 2 x Tn?1 (x) ? Tn?2 (x) for n > 1 [1].
Since the interesting properties of these polynomials only hold in the interval [?1; 1],
we transform the scaling function '(x) to this interval yielding a function ' (y). We
will use the notation y to indicate an independent variable that varies between ?1
and 1,
2' (y) = L'(x) with 2x = L(y + 1):
The renement equation (1.1) becomes
X
' (y) = 2 hk ' (2y ? 2k=L + 1):
k
We construct a quadrature formula,
Z 1 
L (y + 1)
 Z 1 r

' (y)f dy = ' (y)f (y)dy  X wk f (yk ) = Qr [f (x)];
?1 2 ?1 k=1
with yk = dk ?  , dk = 2dk=L ? 1, and  = 2=L. Let Mp denote the modied
   
moments,
Z 1
Mp = Tp (x) '(x) dx:
0
The new system can be written as
r
X
(2.7) wk Ti(dk ?   ) = Mi 0  i  r:
k=1
The solution procedure is similar to the one in the previous section. We construct a
polynomial ?( ), written as a linear combination of Chebyshev polynomials, and
try to nd a root in the appropriate interval. In order to construct ?( ) we write
r?i
r X
X
(y) = 2 ?(r ? 1) qi;j Tj ( ) Ti (y)
i=0 j =0
10 WIM SWELDENS AND ROBERT PIESSENS

and
r?j
r X
X
 
? ( ) = 2 ? (r ? 1) qi;j M Tj (  ):
i
j =0 i=0
Now let
m
Y m m
X X?i
(2.8) 2(m?1) (m) (y) = 2(m?1) (y +   ? dk ) = (m) T (  ) T (y )
qi;j j i
k=1 i=0 j =0
and
?1 mX
mX ?i?1
2(m?1) (m) (y) = 2 (m?1) (y +   ? d ) T (  ) T (y )
qi;j m j i
i=0 j =0
m m
X X?i mX ?2?i
?2 mX
= qi(?m1?;j1) Tj ( ) Ti (y) + m?1) T (  ) T (y )
qi(+1;j j jij
i=1 j =0 i=?1 j =0
mX ?i
?1 mX ?1 mX
mX ?2?i
+ (m?1) T (  ) T (y ) +
qi;j (m?1) T (  ) T (y )
qi;j
?1 j i +1 jj j i
i=0 j =1 i=0 j =?1
mX ?1?i
?1 mX
(2.9) ? 2 dm (m?1) T (  ) T (y ):
qi;j j i
i=0 j =0

An algorithm for the calculation of the qi;j = qi;j(r) can be found by identifying the
coecients of the Chebyshev polynomials of equal degree in x and in   in (2.8) and
(2.9). It is given in appendix A.
The condition number of the system for the construction of the same Q13 formula
as in the previous section is now 1:103 ! The roots of the polynomial ?( ) can
be found as the eigenvalues of its Chebyshev companion matrix. The eects of an
orthogonal basis on the condition of the roots of a polynomial is discussed in [13]. It
is stated there that the interval of orthogonality should contain the roots of interest.
This condition is satised in most cases here.
2.5.3. Calculation of the modied moments.. It is possible to calculate
the modied moment as a linear combination of the monomial moments using the
coecients of the Chebyshev polynomials. However, a considerable loss of signicant
digits will occur since these coecients tend to be large and dierent in sign. The
condition would essentially be as bad as in the construction of the previous section.
We need a formula to calculate the modied moments directly. We know that
Z 1  
Tp u ? 1 +2 2k=L ' (u) du:
X
Mp = hk
k ?1
In order to nd a recursion formula, we write this last, shifted and dilated Chebyshev
polynomial as a sum of Chebyshev polynomials of degree less than or equal to p,
 p
Tp y ? 1 + 2k=L  = 2?p X wi(p) (k) Ti(y);
2 i=0
 WAVELET QUADRATURES AND ERROR EXPANSIONS 11
Table 2.1
Errors of the integration rules.

Trapezoidal One point

n 5:2n rule formula Q5 Q5 Q10
0 5 7.08e-04 1.17e-02 6.13e-04 2.15e-03 -
1 10 4.17e-03 1.43e-03 9.78e-05 4.40e-05 1.03e-08
2 20 7.96e-04 1.76e-04 4.30e-06 6.51e-07 1.11e-12
3 40 1.15e-04 2.19e-05 1.52e-07 9.38e-09 4.21e-15
4 80 1.53e-05 2.74e-06 5.03e-09 1.38e-10 9.99e-16
5 160 1.98e-06 3.43e-07 1.61e-10 2.09e-12 -
6 320 2.50e-07 4.28e-08 5.10e-12 3.19e-14 -
7 640 3.15e-08 5.35e-09 1.60e-13 1.11e-16 -
8 1280 3.96e-09 6.69e-10 4.66e-15 - -
9 2560 4.96e-10 8.37e-11 2.22e-16 - -
10 5120 6.20e-11 1.04e-11 - - -

such that
?1 X !
M = 1 pX L
(p) 
p 2p ? 1 i=0 k=0 hk wi (k) Mi :
Numerical tests show this to be a stable recursion formula. The wi(p) (k) can be
calculated recursively. We will use the notation wi(p) = wi(p) (k) and  = 2k=L ? 1 for
simplicity here. Now
  pX+1
(2.10) y + 
Tp+1 2 = 2 ? (p+1) wi(p+1) Ti(y)
i=0
and
     
y
Tp+1 2 +  y + 
= (y + ) Tp 2 ? Tp?1 2 y + 
p+1
X p?1
X
= 2?(p+1) wi(?p)1 Ti(y) + wi(+1
p) T (y )
jij
i=1 i=?1
p p?1 !
X(p) X (p? 1)
(2.11) + 2 wi Ti (y) ? 4 wi Ti (y)
i=0 i=0
The algorithm can be found by identifying the coecients of the Chebyshev polyno-
mials of equal degree in (2.10) and (2.11). It is given in appendix A.
2.6. Numerical results. When calculating the coecients at a level n with
a quadrature formula, one usually wants to avoid evaluating (or \sampling") the
function f (x) at abscissae with spacing smaller then 2?n. This means that s  0.
For a certain r, one wants the maximal s so the abscissae spread out over the whole
integration interval. The maximal s, within the requirement that (r ? 1)2s < L,
however corresponds to the smallest admittance interval for  . As mentioned above,
there is no theoretical certainty that a quadrature formula with degree of accuracy r
exists. If for a formula with s > 0 no  can be found, one can always try to nd a
formula with spacing 2s?1.
12 WIM SWELDENS AND ROBERT PIESSENS

We constructed quadrature formulae for the Daubechies' orthogonal scaling func-

tions and veried that formulae Qr with 2  r  2N ? 1 exist for 2  N  10. For
r = 2, one of the weights is always zero and we return to the one point formula. In 9
of these 90 formulae, no  could be found for the maximal value of s and s was taken
one smaller. This was in most cases for r = N . The weights of these formulae can
vary in sign, but their absolute value does not grow too large when the number of
points increases.
We compare now dierent quadrature formulae in a practical example. We con-
struct several multiresolution trees, each with coarsest level 0 and nest level n, and
this for several n. We compare each time 0;0 . Notice however that the error is of the
same order in the whole multiresolution tree.
As an example we take for '(x) the Daubechies orthogonal scaling function with
N = 3, f (x) = sin(x) and
Z 5
(2.12) 0;0 = '(x) sin(x)dx  0:741104421925905
0
We compare the one point formula, Q5 , Q10 (with s = ?1 applied at level n ? 1), and
the trapezoidal rule. The total number of evaluations is then respectively 5:2n ? 4,
5:2n , 5:2n and 5:2n ? 1. We also use a formula Q5 were  was given a xed value
equal to ?1=2. The results are given in table 2.1. They show that for suciently
dierentiable functions f (x), it is useful to search for the optimal value of the shift  .
3. Asymptotic error expansions.
3.1. General idea. In the second part of the paper we study pointwise asymp-
totic error expansions. Dene therefore the error of the wavelet approximation as
En f (x) = f (x) ? Pn f (x);
and note that
1
X
(3.1) En f (x) = Qn+i f (x):
j =0
The general idea is to rst construct an expansion for Qn and then apply this formula
to nd the expansion for En . For simplicity we assume that the wavelets are orthogonal
and compactly supported.
3.2. Construction of error expansions. We know
X
(3.2) Qn f (x) = n;l (2nx ? l); with n;l = 2n h f (y); (2n y ? l) i
l
We suppose that f (x) 2 C N +1 with f (l) (x) bounded for l  N + 1. In the sequel we
will always have h = 2?n . We also suppose that the support of (x) is [A; B ] with
A; B 2 Z. In formula (3.2) both x and y belong to In;l = [h(A + l); h(B + l)] and
consequently jx ? yj  Lh with L = B ? A. We also suppose that AB  0 such that
hl 2 In;l . For n xed, at most L + 1 functions n;l (x) are non-zero at x, namely the
ones with 2n x ? B  l  2n x ? A. We can follow two strategies. First, using the
Taylor formula around hl, we can write
n;l = h f (hz + hl); (z) i
 WAVELET QUADRATURES AND ERROR EXPANSIONS 13
+N
MX p N +M +1
= h hp f (p) (hl) zp! + hN +M +1 f (N +M +1) ( ) (Nz+ M + 1)! ; (z) i
p=0
with  between hl and hl + hz
+M
NX
= f (p) (hl) Np!p + n;l hN +M +1 ;
p=N
with
jn;l j  max
2I
jf (N +M +1) ( )j N +M +1 :
n;l

Here j is dened as
Z B
1
j = j ! jzjj j (z)j dz:
A
Now
NX+M hp N X
(3.3) Qn f (x) = p f (p) (hl) (2nx ? l) + Kn hN +M +1 ;
p=N p! l
with
(3.4) jKn j  (L + 1) j?xmax
jL=2
jf (N +M +1) ( )j N +M +1 0 :
n

Consequently
NX+M hp N X f (p) (hl=2i)
Qn+if (x) = p!
p
2ip (2n+ix ? l) + 2i(NK+nM+i+1) hN +M +1 :
p=N l
Since the upper bound (3.4) of jKn+i j cannot grow as i increases, we still have a
O(hN +M +1) term if we sum the Qn+i f (x) and thus,
NX+M hp N X1 X f (p) (hl=2i)
(3.5) En f (x) = p (2n+ix ? l) + O(hN +M +1):
p=N p! i=0 l 2ip

The advantage of this formula is that the contributions of each subspace Wn+i can
be distinguished. The disadvantage is that, because of the double summation, it is
not very practical to work with. Therefore we now derive a second formula using the
Taylor formula around y = x in (3.2),
+M
NX )N +M +1 ; (2n y ? l) i
f (p) (x) (y ?p!x) + f (N +M +1) ( ) ((yN?+xM
p
n;l = 2n h + 1)!
p=N
with  between x and y
NX+M 2n f (p) (x)
= p ! h (y ? x)p; (2n y ? l) i + n;l hN +M +1 ;
p=N
14 WIM SWELDENS AND ROBERT PIESSENS

with
N +M +1
jn;l j  (NL+ M + 1)! max
2I
jf (N +M +1) ( )j 0 :
n;l

Now
2n h (y ? x)p; (2n y ? l) i = h (hz + hl ? x)p; (z) i
pX?N  
= h p p N (l ? 2n x)j :
p?j
j =0 j
Thus:
?N p
+M hp f (p) (x) pX
NX
n;l = p! j Np?j (l ? 2n x)j + n;l hN +M +1 ;
p=N j =0
and
+M
NX ?N
pX
Np?j (?)j  (2n x) + K hN +M +1 ;
Qn f (x) = hp f (p) (x) (p ? j )! j ! j n
p=N j =0
with j (x) 2 L2 ([0; 1]) dened as
X
j (x) = (x ? l)j (x ? l);
l
and
N +M +1
(3.6) jKn j  (L + 1) (NL+ M + 1)! j?xmax
jL=2
jf (N +M +1) ( )j 20 :
n

We can write this as

M
X
(3.7) Qn f (x) = hN +q f (N +q) (x) q(2nx) + Kn hN +M +1 ;
q=0
with
q
(3.8) q (x) =
X NN +q?j (?)j  (x):
(N + q ? j )! j ! j
j =0
Again we can sum the contributions all up such that,
M
X
(3.9) En f (x) = hN +q f (N +q) (x) q (2nx) + O(hN +M +1 );
q=0

with q (x) dened as the limit function of a uniformly convergent series,

1
X q(2ix)
(3.10) q (x) = i(N +q) :
i=0 2
 WAVELET QUADRATURES AND ERROR EXPANSIONS 15
This formula is more practical to work with but has the disadvantage that one cannot
distinguish the contributions of the dierent wavelet subspaces any more. The pro-
jection Qn f (x) belongs by denition to Wn . However, in formula (3.7) the rst term
doesn't belong completely to Wn . We can understand that for suciently smooth f (x)
and large n it has a big component in this space. As we will see, also the O(hN +1)
term of (3.7) has a component in Wn which can make the rst term a bad estimate
of Qn f (x) and consequently the rst term of (3.9) a bad rst estimate of En f (x). We
will come back to this later.
3.3. Interpolation. It is easy to see from their denition that p(x), p(x), and
 (x) are one periodic functions. So the general term consists of a power of h, the same
order of derivative of f (x) and a highly oscillating factor. In this section we take a
closer look at the rst term of the expansion. Therefore we rst derive some properties
of 0(x) and 0 (x). We use the notation (x) = 0 (x) and  (x) = 0 (x) N !=NN . Using
equation (1.2) we can obtain the following relations:
X X
(x) = (?1)l '(2x ? l) = 2 '(2x ? 2l) ? 1;
l l

(3.11) (x + 1=2) = ?(x); and  (x + 1=2) =  (x) ? 2 (x):

Using Poisson's summation formula the Fourier series of (x) can be written as
+1
X
(x) = ^(2k) exp(ik2x):
k=?1
In the Haar case, when '(x) = [0;1](x), (x) [0;1](x) is the Haar wavelet '(2x) ?
'(2x ? 1) and  (x) is a sawtooth function with  (x) = 2 ? 4x for 0  x < 1. The
(2j x) functions with j  0 are then called the Rademacker functions which are well
known in probability theory. Figures 3.1 and 3.2 show  (x) and (x) corresponding
to the Daubechies wavelets for dierent values of N . For N > 5, the plots of (x)
and  (x) almost coincide visually. For N = 10, they look like a shifted sine function.
This makes sense because the smoother (x), the faster the decay of ^( ) and the
more (x) will look like its fundamental frequency component.
Theorem 3.1. If '(x) is continuous, then (x) has at least two zeros in [0,1).
Proof. If '(x) is continuous, so is (x). The proof then follows directly from
formula (3.11). Moreover, if x0 is a zero in [0; 1) so is (x0 + 1=2) mod 1.
Theorem 3.2. If '(x) is continuous and N > 1, then  (x) has at least two zeros
in [0,1).
Proof. If '(x) is continuous, so is  (x) since the series (3.10) converges uniformly.
The proof follows from
N N ?2
 (0) =  (1) = 2N2 ? 1 (0) and  (1=2) = ? 22N ? 1 (0);
which means that  (x) has at least two changes of sign in [0; 1].
In the cases we considered, it turns out that (x) and  (x) have exactly two zeros
as can be seen in the gures. We will denote the zeros of  (x) in [0; 1) by x1 and x2
with x1 < x2. In the sequel we will suppose that the conditions of these theorems are
satised. Note also that
NN  (x) = xN ? P0 xN :
16 WIM SWELDENS AND ROBERT PIESSENS
2 2
.....
.... ..
.... .
.... . ..
..... ..
.....
....
N =2 N =2
..
... ..
... .. ..
....
... .. ... ...
.....
 (x)  (x)
.......................
......
.. .... .... ............. ..
..
. ..
... ........... ....................... ..
1 .
.
.. .......
....
..
....... ...........
.
..
..
..
..
1 ...
..
..
.... ..
..
...
. .
... ..
.
.... ..
... ....
.
..
...
. .. . ...
... ... ............. ..
.. .. . ..
.....
. ..
.. .
.
.. ...
.. . .....
..........
.
..
... ... ...
...
. ... .
..
. ... ..
.
...
...
. .. . ...
.. ..
0 0
. .. . ......
...
.
...
.. .
.
...... ....
. .. . ...
...
... .. ..
.
....
....
. ........... .
...
. .... ..
. ...
...
... ..
...
. .... .
.
....
...
...
.
.
...
... .... ...
. ... . ...
.. ... .. ........

?1 ?1
.
. . ....
..
...
... ...... ..
. ...
...
.
. ... ... ....... .
. ...
..
. .... .....
.... ..
. ...
...
. .
..
.
.....
...... ..
.
...
....
. ... .
..
. ... ..
.
....
...
..
.
....
..... ..
. ...
. . ...
..
.
....
......
.
.
. ..
...
. ..... .
. ...
.
. ..... .
. ...
. .
..
. ..... .
..
...

?2 ?2
. ...
. ...
..
. ...
.. ...
.
. ...
. ...
.
. ..
.
. ...
.. ...
.

0 1/2 1 0 1/2 1
2 2
.
......
.. ....
... ......
.. ...
..........
... ..
.... .........
N =3 N =3
... ..
...
. .... ...
... ...
.. .... ... ....
.. ... .... ....
...
 (x)  (x)
....
...
..
... .
.... ...
1 1
.. .. .
.. ...
.... .. .
.... ...
....
... .
.
.... ... .
.... ....
. .... . ....
..
.... ... .
...
....... ...
....
.... ... .
....
....
. ....
..
..
..
. ...
..
.. ... ..... ...
.....
.
. ..
.. ...
...... ...
.. ..
.
.....
. ..
... .
... ...
..
... ..
....
. ..
.. ..
.
...
...
. .
.. .. .. ..
0 .
..
.
.
...
. ...
..
.. 0 .
.
..
. ....
...
....
...
..
....
..... .... ....
.... . ...
...
. .....
... .... ...
...
.
.. ... ... ...
.. .... .... ...
... .... .. ...
...
. ..... ..
. ...
.
.... .... .. ...
.... ..

?1 ?1
.
. ... ..
... ... .. ...
... ... ... ..
..
...
. ..
... .
..
. ..
.. . ...
..
.. .. ...
.. ... .. ...
....
.
..
... .... ..
... ... ..
...
...
..
. ... .. ..
.
...
... .. ...
.. ... ..
. ..
.. . ..
.
.... ..
... ... ..
....
. ..
.... ...

?2 ?2
.
. ..
. ..
...
.
..
..
.
... ..

0 1/2 1 0 1/2 1
2 2
...................
....... ...... .........
.... ..... ....... .................
... .... ...... .....
... .... .... ...
... ... .... ...
... ... ... ...
... ... ...
1
.
.
..
.
.
.. ..
..
...
...
1 .
.
.
...
... ...
...
....
.
.. .. ...
... ....
..
.
. ...
.... .... ...
. ....
.
..
. ... ....
. ...
.. .... . ....
... ....
... ...
.
...
...
.
.... ....
.... .
... ...
...
.. .... . ....
..
. ... .
... ...
. .... . ...
... ... ...
0 0
. ... ..
. ..
.. ... ..
....
. ... ..
...... ..
..
... ..
.... ...
....
.
...
.. .
. ..
.. ..
..
.... ..
.. ..... ..
..
... .. ..
....
..
....
. .. .
. ..
..
.. ... .
...
....
. ... ..
.
..
..
.. .. ...
... ...

?1 ?1
.. .. .
... . ...
... ..

N =4 N =4
. .. . ..
.
.... ...
.. .
.
... ..
..
.
.. .... . ...
.
...

 (x)  (x)
... .. ... . .
...
... .
..
.... ... .
.....
. ... .
..
....
....... ..
.......
. ... ..
......
.
.................. ... .
...... ......
.....................

?2 ?2
0 1/2 1 0 1/2 1

Fig. 3.1.  (x) and  (x) for Daubechies' wavelets with N = 2; 3; 4.

This means that in case the scaling functions are spline functions,  (x) is a mono-
spline. Therefore, we might want to call  (x) in general a mono-wavelet.
The envelopes of the rst term are given by
(3.12) hN NN f (N ) (x)  and hN NN f (N ) (x)  ;
N! max N! min
with 1

max = xmax
2[0;1]
 (x) and min = xmin
2[0;1]
 (x):
 WAVELET QUADRATURES AND ERROR EXPANSIONS 17
2 2

...........
...... .........
.... ..... .........................................
..... ..... ........ ....
.... ..... .....
... ..... .....
....
... ..... ...
...
... ...
....
1 ..
.
...
...
.
.....
.....
.....
....
1 ...
...
.. ...
.
.
..
...
.
...
..
.. ... ... .
...
.
...
... ... ....
. ... .. ...
...
. ... ... ..
.
.. .. ..
... ...
.. ... .. ..
.. .. ..
. ... ... ..
. ... .. ..
... ... ..
.. ..
.. .. ..
.. .. ...
0
..
..
..
..
.
.
...
..
. .
0 ...
...
... ...
..
..
... . .. ...
.. ...
. .. ...
.. . ... ..
.. .. ...
... ... ...
... ..
... .. ..
... ... ..
... ...
... ... ... ...
... ... ..

?1 ?1
... .. ..
... ... ... ...
....
N =5 N =6
... ... ...
... .
......
. ...
... ..
.
... . ..
... ..... .... ....
.... ..... ... ...
.....
 (x)  (x)
... ..... ....
.... ...... ......
.....
..... ..... .........
.....................................
...... .....
..................

?2 ?2
0 1/2 1 0 1/2 1
2 2

N =8 N = 10
............................ .
..............................
...... ..... ......
..... .... ....
.....
.... .... ....
... ....
.. ....
...

 (x)  (x)
... ... ...
... ... .... ....
... ...
1 ...
1
... . ...
... ... ..
. ...
... ... ...
.
...
...
... ... .
.. ....
... .
.... .... ...
.
... ... ..
...
...
... .. . ...
... ... .
.. ...
... ... .. ...
... .. .. ...
... ... ..
. ...
... .
.... .... ...
... . .. ...
... ... ..
. ...
... ... ..
0 0
.
... .. ... ..
... ... .. ...
... ... ...
. ...
... ... . ...
... .... ...
. ...
... .. . ...
... ... ...
. ...
.. ... .
... .. ...
. ...
... .. ...
... ... ... ..
...
... .. ... ..
...

?1 ?1
...
... .
..
. ...
... ....
..
... ... ... ...
...
... ... ... ...
...
... ... .... .
.... ... ... .
.....
.
... ... ....
....
.... ... ....
..... ..
..... .....
....... ....
.....
...... .. ...........................
...........................

?2 ?2
0 1/2 1 0 1/2 1

Fig. 3.2.  (x) for Daubechies' wavelets with N = 5; 6; 8; 10.

Note that jmax j 6= jmin j so the oscillation is not necessarily balanced around the
axis. This dierence however becomes smaller for increasing N .
Theorem 3.2 now implies that the rst term of formula (3.9) will have at least
2n+1 zeros per unit length. For suciently small h, the approximation Pn f (x) will
thus interpolate the function f (x) also in roughly 2n+1 points per unit length. Note
that this is about twice the number of basis functions. The interpolation points zk
with Pn f (zk) = f (zk) satisfy asymptotically
z2k = (x1 + k) h + O(h2) and z2k+1 = (x2 + k) h + O(h2):
3.4. Numerical extrapolation. We can use the the error expansion to accel-
erate the accuracy of a wavelet series with extrapolation techniques. Indeed, the mul-
tiresolution scheme consists of a number of approximations at dierent levels which
can be used to estimate the components of the error expansion and eliminate them.
If m is the coarsest level of the multiresolution scheme and 2m x 2 Z, the asymptotic
error expansion at x, consists of powers of h whose coecients are independent of h
due to the periodicity of j (x). This means that classical extrapolation techniques
such as Richardson extrapolation become applicable.
Table 3.1 shows the results of a numerical experiment where f (x) = exp(?20 (x ?
0:5)2) and the wavelet used is the Daubechies
1 wavelet with N = 2 vanishing moments.
The rst column shows the the relative error of the approximation at x = 1=4 at levels
n from 2 to 9. The other columns are the relative errors of the values obtained with
18 WIM SWELDENS AND ROBERT PIESSENS
Table 3.1
The Richardson extrapolation table for x = 1/4.
E f (1=4)
n j n j
f (1=4)
2 1.2e-01 - - - - - - -
3 5.6e-02 3.3e-02 - - - - - -
4 1.7e-02 3.5e-03 6.8e-04 - - - - -
5 4.3e-03 1.9e-04 2.8e-04 2.5e-04 - - - -
6 1.1e-03 2.5e-06 3.0e-05 1.3e-05 5.5e-06 - - -
7 2.7e-04 2.3e-06 2.3e-06 4.3e-07 1.4e-08 7.3e-08 - -
8 6.6e-05 4.2e-07 1.5e-07 1.2e-08 1.3e-09 1.6e-09 9.9e-10 -
9 1.6e-05 6.2e-08 9.9e-09 3.4e-10 3.6e-11 1.6e-11 3.6e-12 3.1e-13

the following Richardson extrapolation scheme:

tn;1 = Pn f (1=4) for 2  n  9;
and
N +j ?2 tn;j ?1 ? tn?1;j ?1
tn;j = 2 2N +j?2 ? 1 for 3  n  9 and n  j  9:
The rst column shows a convergence of h2 , i.e. on each level the error is roughly
divided by 4. Every next column corresponds to eliminating one term of the expansion.
We see that in this case simple linear combinations result in an increase from 5 to
almost 13 accurate digits. This however only works for the points with 2m x 2 Z.
In a more general setting, it is also possible to consider f as a distribution. Note
that a distribution f is characterized by inner products h f;  i where  belongs to the
class of test functions. If f is a distribution we can approximate the inner products
h f;  i by h Pnf;  i . We can use the asymptotic error expansion of the wavelet
approximation to get an asymptotic error expansion for these inner products. For a
certain class of test functions this error expansion will consist of powers of h whose
coecients are independent of h. This again will make Richardson extrapolation
applicable.
3.5. A more accurate rst term. In this section we try to slightly modify the
rst term of the error expansion (3.9) so that it becomes a more accurate estimate of
En f (x). Therefore we rst consider the case of approximating f (x) = xN +1 at level
0. Equation (3:7) yields
Q0 xN +1 = (N + 1) NN x (x) + NN +1 (x) ? (N + 1) NN 1(x):
As mentioned above, this formula has the disadvantage that the components of each
subspace cannot be distinguished. Therefore we will try to rewrite this formula and
isolate its component in W0 . First we group the rst two terms,
Q0 xN +1 = (N + 1) NN (x + ) (x) ? (N + 1) NN 1(x); with  = (N N+N1)+1N :
N
Secondly, we isolate the component of 1(x) in W0 , by letting Q0 1(x) =  (x),  =
h 1(y); (y) i , and 1 (x) =  (x) ? 1 (x), such that (x) and 1(x) are orthogonal.
Then,
Q0 xN +1 = (N + 1) NN (x ? ) (x) + (N + 1) NN 1(x); with =  ? :
 WAVELET QUADRATURES AND ERROR EXPANSIONS 19
In appendix B an algorithm to calculate  is described. In this formula the second
term has no component in W0 . We see that in order to isolate the component in W0 ,
the modulating function (in this case x) has to be shifted over a distance . We now
need the following theorem.
Theorem 3.3. If  (x) is a periodic function with period one and r is the number
of vanishing moments of  (x) (x) and g(x) 2 C r , then Qn [g(x)  (2nx)] = O(hr)
where h = 2?n.
Proof. We know that
X
Qn [g(x)  (2nx)] = n;l (2nx ? l);
l
with
n;l = 2n h g(y);  (2n y) (2ny ? l) i = h g(h (y + l));  (y) (y) i :
The proof follows from the Taylor formula.
We assume that f (x) 2 C N +2 . Then from (3.7)
Qn f (x) = q0 (x) hN + q1 (x) hN +1 + O(hN +2);
where q0 (x) = f (N ) (x) 0(2nx), q1(x) = f (N +1) (x) 1(2nx), and, from (3.8),
0(2nx) = NNN! (2nx)
1(2nx) = NNN++11! (2nx) ? NNN! 1(2n x) = NNN! [? (2nx) + 1(2nx)] :
The problem here is that both q0(x) and q1 (x) have a component in Wn . Indeed,
since (x) (x) has a non vanishing integral, theorem 3.3 says that Qn q0(x) = O(h0)
and Qn q1 (x) = O(h0). We can solve this problem by grouping the components in
Wn ,
Qn f (x) = h NN! N f (N ) (x) ? h f (N +1) (x) (2nx) + hf (N +1) (x)1(2n x) + O(h2)
N h  i

= r0;0 (x) hN + r0;1 (x) hN +1 + O(hN +2);

where
r0;0 (x) = NNN! f (N ) (x ? h ) (2nx) and r0;1 (x) = NNN! f (N +1) (x) 1(2n x):
In gure 3.3, 1(x) and 1 (x) are shown for the Daubechies wavelets. As a consequence
of the orthogonalization, 1 (x) is smaller then 1(x) and new rst term is more
accurate. From theorem 3.3 with (x) and 1(x) as  (x), respectively, we see that
Qn r0;0 (x) = O(h0) and Qn r0;1 (x) = O(h1). This means the component in Wn of the
O(hN +1) term will tend to zero if h ! 0. Now, for i > 0,
Qn+i f (x) = hN ! 2NiNN f (N ) (x) (2n+ix) +
N h

(N +1) ? 
h f 2i (x) ? (2n+ix) + 1(2n+ix) + O(h2)

= ri;0 (x) hN + ri;1 (x) hN +1 + O(hN +2);

20 WIM SWELDENS AND ROBERT PIESSENS
Table 3.2
The shift for the original Daubechies wavelets.
N N N N
1 0 6 -1.9942 11 -3.9906 16 -6.0034
2 -0.4301 7 -2.3909 12 -4.3924 17 -6.4068
3 -0.8187 8 -2.7892 13 -4.7947 18 -6.8104
4 -1.2077 9 -3.1888 14 -5.1973 19 -7.2142
5 -1.5996 10 -3.5893 15 -5.6003 20 -7.6179
Table 3.3
The shift for the \most symmetric" Daubechies wavelets.
N N
2 -0.4301 6 -0.0189
3 -0.8187 7 0.0683
4 0.0860 8 -0.0741
5 0.6132 9 0.3431

with
ri;0 (x) = NN! 2NiN f (N ) (x ? h) (2n+ix)
and
ri;1 (x) = N ! 2Ni(NN +1) f (N +1) (x) (2i ? 1) (2n+ix) + 1 (2n+ix) :
?

Lemma C.3 states that j (2ix) (x) has 2N vanishing moments if 0  j < N . Thus,
using theorem 3.3 with (2ix) and 1(2ix) as  (x), yields respectively Qn ri;0 (x) =
O(h2N ) and Qn ri;1 (x) = O(h2N ). So the only term with a component in Wn that is
independent of h is r0;0 (x). Consequently,
N
(3.13) En f (x) = h NN! N f (N ) (x ? h )  (2n x) + O(hN +1);
where shifting the modulating function, yields an O(hN +1) term with a component
in Wn that is O(h) and thus a more accurate rst term. This is illustrated with two
examples in the next section.
It is easy to see that the shift is zero for (anti-)symmetric wavelets. This shift
can be seen as a measure for the symmetry of the wavelet. Ingrid Daubechies showed
that, except for the Haar case, no symmetric compactly supported orthogonal wavelets
exist [7]. In [9] she constructed so called \most symmetric" wavelets, who have the
closest to linear phase of all wavelets with support length 2N ? 1 and N vanishing
moments.
In table 3.2 and 3.3 we give numerical values for in function of the number
of vanishing moments both for the original Daubechies wavelets and for the \most
symmetric" ones. For the original Daubechies the absolute value of the shift seems
to be increasing linearly with N . As could be expected, the shift is smaller for the
\most symmetric" ones.
3.6. Numerical Example. In this section we consider an example with
f (x) = exp(?20 (x ? 0:5)2);
 WAVELET QUADRATURES AND ERROR EXPANSIONS 21
2 1

.....
... ..
.... .
... ...
.
. N =2 N =2
1 ..
.....
.
.
.
.....
....
....
..
.
..
..
..
..
1 (x) 1/2 1 (x)
..
....
. ...
.. ....
..
...... .. .........
.
.
....... ............
. ... ...................... ............................ .
.. ...... ....
.
. ...... ........ ...... ......... ...
.
..... ....
.. ........ .......... ...
....
. ......
.... ... ...... .....
..
. ... .... ....
.... .....
0 0
. ............ ... .
. .... .... .
.
............. ...... ..... ....
.
. ...
.
...
... ... ....
..
....
....
.... ......
.. ......
.... .... .... ........ ..........
.. ...
... ..............
.... ...... ........
........ ..... ......
....
......... .
..
......
.. ..... ................................................
.. ..... .
.. .....
... ...

?1
.
. ...
... ...
....
-1/2
. ....
..
. ....
. ....
..
. ....
. ...
.
. ....
.. ....
.. ..

?2 -1
0 1/2 1 0 1/2 1
2 1

....
.......................
.... ...............
................... N =5
....
1 (x)
......

1 .
.
..
..
...
...
.
....
....
...
...
1/2 .................................
....... .....
.
.. ... ..... .....
.. ... ... ...
.. ... ... ...
.... .. .
..... ....
...
... .
.. .
..
. . ..
..... ....
....
... ... ..
.. .
. ..
..... ...
....
. ...
..
.. ...
. ...
...... ...
....
... . ...... ....
..
0 0
.. . ...
. ....
. ..
...
... .. .
...... .....
......
... .
..
.... .....
..
... ... ..
.
. ......
... .. ...
....
... ... .. ....
.. .
... .
..... ....
....
. ....
...
... ...
.. .
..... ....
.
.. ...
... .. .... ....
... ... .
. ....

?1
... ... ...
.... ......
... ... .
... ........

N =5
...
...
-1/2 ..............
... . ............
... ..
... ...
.... ......
..................
1 (x)
...
.... ...............
..........................

?2 -1
0 1/2 1 0 1/2 1
2 .....................
.....
1
... ....
...
.... ....
... ..
..... ....
.. ...

.
..
.
...
...
. ...
..
...
... N =8
1
... ...
...
...
...
...
1 (x) ...
..
...
.
..
1/2 ......
.....
....
.........................................
....... .....
.....
.....
.. ..
. .... ....
....
.. .
.. ....
... ..
...
.... .
..... .
..
.. ... ... .
...
.. .. ... .
.
.. ... ...
...
.
...
... .. ... ..
..
... ...
. ... .
....
.. ... .
...
0 0
.. .. ...
... .. ... ..
... .. ... ...
... .. ... ..
... .. ... ...
... .. ... ...
... .
. ... ...
.
. .
.. ...
...
... ....
..
... .. ... ...
.. .... ...
..
.. .... ...
..
.. .... ....
... ...

?1
...
.. ....
.
....
..... ..
......
.. ...... ..

N =8
......
-1/2
...
.. ... ......
........ .....
... .. ....................................
... ...
...
1 (x)
...
... ...
... ...
.
... ..
... ...
... ..
.... ...
... ....
.... ...
....

?2
......
.......................

-1
0 1/2 1 0 1/2 1

Fig. 3.3. 1 (x) and 1 (x) for Daubechies' wavelets with N = 2; 5; 8

where we look at the error in the interval [0; 1]. Figure 3.4 shows the error E6 f (x)
obtained with a computer program calculating P6 f (x) ? f (x). The wavelet used is the
Daubechies wavelet with 3 vanishing moments. One can check that the interpolating
properties are satised. On the same gure, the envelopes (3.12) of the rst term of
the expansion (3.9) are drawn dashed and the envelopes of the rst term of (3.13) are
drawn solid. Figure 3.5 shows the error E6 f (x) using the Daubechies wavelet with 9
vanishing moments and the same envelopes. We see that the rst term of the error
expansion already gives a reasonable 1approximation of the actual error and secondly
that shifting the modulating function indeed yields more accurate results. For both
these examples the inner products h f (x); 'n;l (x) i were calculated using a quadrature
22 WIM SWELDENS AND ROBERT PIESSENS

formula with an error of O(h2N ) so this in uence can be neglected.

x10 -4
1.5

0.5

-0.5

-1

-1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 3.4. The error and envelopes for N = 3 and n = 6.

3.7. Remarks. Using a generalization of Bernoulli polynomials it is possible to

derive an asymptotic error expansion for the quadrature formula [6, 23]. As a result
it is possible to use extrapolation techniques similar to Romberg integration.
Future research includes the careful study of how the use of a quadrature formula
will in uence the error expansion. First experiment show that the use of a quadrature
formula with degree of accuracy q = N ? 1 can ruin the interpolating properties of
the wavelet expansion. Therefore one might want to use quadrature formulae with
q > N ? 1.
Acknowledgment. The authors are indebted to Pierre Verlinden for many fruit-
ful discussions. He also pointed out a shorter way for some of the proofs.
REFERENCES
[1] M. Abramowitz and I. A. Stegun, Handbook of mathematical functions, Dover Publications,
New York, 1965.
[2] G. Battle, A block spin construction of ondelettes, Comm. Math. Phys., 110 (1987), pp. 601{
615.
[3] G. Beylkin, R. Coifman, and V. Rokhlin, Fast wavelet transforms and numerical algo-
rithms I, Comm. Pure and Appl. Math., 44 (1991), pp. 141{183.
[4] C. K. Chui, An Introduction to Wavelets, Academic Press, San Diego, 1992.
[5] A. Cohen, I. Daubechies, and J. Feauveau, Bi-orthogonal bases of compactly supported
wavelets, Comm. Pure and Appl. Math., 45 (1992), pp. 485{560.
[6] W. Dahmen and C. A. Micchelli, Using the renement equation for evaluating integrals of
wavelets, SIAM J. Num. Anal., 30 (1993), pp. 507{537.
 WAVELET QUADRATURES AND ERROR EXPANSIONS 23
x10 -10
2.5

1.5

0.5

-0.5

-1

-1.5

-2

-2.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 3.5. The error and envelopes for N = 9 and n = 6.

[7] I. Daubechies, Orthonormal bases of compactly supported wavelets, Comm. Pure and Appl.
Math., 41 (1988), pp. 909{996.
[8] , Ten Lectures on Wavelets, no. 61 in CBMS-NSF Series in Applied Mathematics, SIAM,
Philadelphia, 1992.
[9] , Orthonormal bases of compactly supported wavelets II. Variations on a theme, SIAM
J. Math. Anal., 24 (1993), pp. 499{519.
[10] P. Davis and P. Rabinowitz, Methods of numerical integration, Academic Press, London,
1984.
[11] G. Fix and G. Strang, Fourier analysis of the nite element method in Ritz-Galerkin theory,
Stud. Appl. Math, 48 (1969), pp. 265{273.
[12] W. Gautschi, On the construction of Gaussian quadrature rules from modied moments,
Math. Comp., 24 (1970), pp. 245{260.
[13] , Numerical condition related to polynomials, in Recent Advances in Numerical Analy-
sis, C. de Boor and G. H. Golub, eds., Mathematics Research Center, The University of
Wisconsin, Academic Press, 1978, pp. 45{72.
[14] R. A. Gopinath and C. S. Burrus, On the moments of the scaling function, in Proceedings
of the ISCAS-92, San Diego, 1992.
[15] P.-G. Lemarie, Ondelettes a localisation exponentielle, J. de Math. Pures et Appl., 67 (1988),
pp. 227{236.
[16] S. G. Mallat, Multifrequency channel decompositions of images and wavelet models, IEEE
Trans. on Acoust. Signal Speech Process., 37 (1989), pp. 2091{2110.
[17] , Multiresolution approximations and wavelet orthonormal bases of L2 (R), Trans. Amer.
Math. Soc., 315 (1989), pp. 69{87.
[18] Y. Meyer, Ondelettes et Operateurs, I: Ondelettes, II: Operateurs de Calderon-Zygmund, III:
(with R. Coifman), Operateurs multilineaires, Hermann, Paris, 1990.
[19] R. Piessens and M. Branders, The evaluation and application of some modied moments,
BIT, 13 (1973), pp. 443{450.
[20] G. Strang, Wavelets and dilation equations: A brief introduction, SIAM Review, 31 (1989),
pp. 614{627.
[21] G. Strang and G. Fix, A Fourier analysis of the nite element variational method, in
24 WIM SWELDENS AND ROBERT PIESSENS

Constructive aspects of Functional Analysis, Rome, 1973, Edizione Cremonese.

[22] M. Unser, A. Aldroubi, and M. Eden, A family of polynomial spline wavelet transforms,
Signal Process., 30 (1993), pp. 141{162.
[23] P. Verlinden and A. Haegemans, An asymptotic expansion in wavelet analysis and its
application to accurate numerical wavelet decomposition, Numer. Algor., 2 (1992), pp. 287{
298.
A. Algorithms for the modied construction. Algorithm for the calculation of
the coecients qi;j in section 2.5.2:
q0(0);0 1=2
for m 1 (1) r
for i 0 (1) m
for j 0 (1) m ? i
(m)
qi;j qi(?m1?;j1) + qi;j
(m?1) +
?1
( m ? 1)
qi+1;j + qi;j (m?1) ? 2 d q (m?1)
+1 m i;j
if i = 1 then q1(m;j ) q1(m;j ) + q0(m;j ?1)
if j = 1 then qi;(m1 ) qi;(m1 ) + qi;(m0 ?1)
end for
end for
end for
Algorithm for the calculation of the coecients wi(p) in section 2.5.3:
w0(0) 1, w0(1)
, w1(1) 1
w0(2) 22 ? 3, w1(2) 4, w2(2) 1
for p 2 (1) : : :
w0(p+1) w1(p) + 2w0(p) ? 4w0(p?1)
w1(p+1) 2w0(p) + w2(p) + 2w1(p) ? 4w1(p?1)
for i 2 (1) p ? 1
p) + 2w(p) ? 4w(p?1)
wi(p+1) wi(?p)1 + wi(+1 i i
end for
wp(p+1) wp(p?)1 + 2wp(p)
wp(p+1+1) wp(p) = 1
end for
 WAVELET QUADRATURES AND ERROR EXPANSIONS 25
B. Calculation of  . In this section an algorithm to calculate  accurately is
described. We can write
 = h x; (x) (x) i
X
= 2 gk h x; (x) '(2x ? k) i
k
X X
= 1=2 gk h x; ((x + k)=2) '(x) i + 1=2 gk k h ((x + k)=2); '(x) i
k k
= 1=2 (p1 + p2 );
with
X X
p2 = gk k h ((x + k)=2); '(x) i = h1?k k h (x=2); '(x) i = 1 ? M1 ;
k k
and
X X
p1 = gk h x; ((x + k)=2) '(x) i = h x; (x=2) '(x) i = (?1)ll ;
k l
where
l = h x; '(x ? l) '(x) i :
The l can be found by solving the linear system
X
m = al?2m l + b2m ;
l
with
X X
ai = hk hk+i and bi = k hk hk+i:
k k
C. Proof of some Lemma's.
Lemma C.1. The function j (x) '(x) has N vanishing moments if 0  j < N
and no vanishing moments if N  j .
Proof.
X
h xp; j (x) '(x) i = h (x + l)p; xj (x) '(x + l) i = Mp Nj if 0  p < N
l
This is zero if 0  j < N and non-zero if p = 0 and j = N .
Lemma C.2. The following functions also have N vanishing moments if 0  j <
N and no vanishing moments if N  j : j (x) '(x ? l) with l 2 Z, j (2ix) '(x) with
i 2 N, j(x) '(x), and j (x) '(x).
Lemma C.3. The function j (2ix) (x) with i > 0 has 2N vanishing moments if
0  j < N and N vanishing moments if N  j .
Proof. For i = 1
X
h xp; (x) j (2x) i = 2 gl h xp; '(2x ? l) j (2x) i
l
X
= 2?p gl h xp; '(x ? l) j (x) i
l
X
= 2?p gl h (x + l)p; '(x) j (x) i
l
26 WIM SWELDENS AND ROBERT PIESSENS

If 0  p < N this is zero for all j because of lemma C.1. If N  p < 2N the inner
product in the summation is a polynomial of degree p ? N < N . To prove the lemma
for i > 1, we can follow the same reasoning as above and use lemma C.2.