This action might not be possible to undo. Are you sure you want to continue?

BooksAudiobooksComicsSheet Music### Categories

### Categories

### Categories

Editors' Picks Books

Hand-picked favorites from

our editors

our editors

Editors' Picks Audiobooks

Hand-picked favorites from

our editors

our editors

Editors' Picks Comics

Hand-picked favorites from

our editors

our editors

Editors' Picks Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

**COMPACTLY SUPPORTED WAVELETS
**

∗

G. BEYLKIN

†

SIAM J. NUMER. ANAL. c 1992 Society for Industrial and Applied Mathematics

Vol. 6, No. 6, pp. 1716-1740, December 1992 011

Abstract. This paper describes exact and explicit representations of the diﬀerential operators,

d

n

/dx

n

, n = 1, 2, · · ·, in orthonormal bases of compactly supported wavelets as well as the rep-

resentations of the Hilbert transform and fractional derivatives. The method of computing these

representations is directly applicable to multidimensional convolution operators.

Also, sparse representations of shift operators in orthonormal bases of compactly supported

wavelets are discussed and a fast algorithm requiring O(N log N) operations for computing the

wavelet coeﬃcients of all N circulant shifts of a vector of the length N = 2

n

is constructed. As

an example of an application of this algorithm, it is shown that the storage requirements of the fast

algorithm for applying the standard form of a pseudodiﬀerential operator to a vector (see [G. Beylkin,

R. R. Coifman, and V. Rokhlin, Comm. Pure. Appl. Math., 44 (1991), pp. 141–183]) may be reduced

from O(N) to O(log

2

N) signiﬁcant entries.

Key words. wavelets, diﬀerential operators, Hilbert transform, fractional derivatives, pseudo-

diﬀerential operators, shift operators, numerical algorithms

AMS(MOS) subject classiﬁcations. 65D99, 35S99, 65R10, 44A15

1. Introduction. In [1] Daubechies introduced compactly supported wavelets

which proved to be very useful in numerical analysis [2]. In this paper we ﬁnd exact

and explicit representations of several basic operators (derivatives, Hilbert transform,

shifts, etc.) in orthonormal bases of compactly supported wavelets. We also present

an O(N log N) algorithm for computing the wavelet coeﬃcients of all N circulant

shifts of a vector of the length N = 2

n

.

Throughout this paper we only compute the nonstandard forms of operators since

it is a simple matter to obtain a standard form from the nonstandard form [2]. Meyer

[3], following [2], considered several examples of nonstandard forms of basic operators

from a general point of view. It is possible, however, to compute the nonstandard

forms of many important operators explicitly.

First, we explicitly compute the nonstandard form of the operator d/dx. The set

of coeﬃcients that deﬁnes all nonzero entries of the nonstandard form appears as the

solution to a system of linear algebraic equations. This system, in turn, arises as a

consequence of the recursive deﬁnition of the wavelet bases. The operator d

n

/dx

n

is

treated similarly to d/dx.

The computation of the nonstandard forms of many other operators reduces to

solving a simple system of linear algebraic equations. Among such operators are

fractional derivatives, Hilbert and Riesz transforms, and other operators for which

analytic expressions are available. For convolution operators, there are signiﬁcant

simpliﬁcations in computing the nonstandard form since the vanishing moments of

the autocorrelation function of the scaling function simplify the quadrature formu-

las. Moreover, by solving a system of linear algebraic equations combined with the

asymptotics of wavelet coeﬃcients, we arrive at an eﬀective method for computing the

nonstandard form of convolution operators. As examples, we compute the nonstan-

dard forms of the Hilbert transform and fractional derivatives. The generalization of

this method for multidimensional convolution operators is straightforward.

∗

Received by the editors August 6, 1990; accepted for publication (in revised form) October 21,

1991.

†

Schlumberger-Doll Research, Old Quarry Road, Ridgeﬁeld, Connecticut 06877. Present address,

Program in Applied Mathematics, University of Colorado at Boulder, Boulder, Colorado 80309-0526.

1716

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1717

Second, we compute the nonstandard form of the shift operator. This operator

is important in practical applications of wavelets because the wavelet coeﬃcients are

not shift invariant. Since the nonstandard and standard forms of this operator are

sparse and easy to compute, knowing these representations “compensates” for the

lack of shift invariance. The wavelet expansion of shifts of vectors or of matrices may

be obtained by applying the shift operator directly to the coeﬃcients of the original

expansion. The coeﬃcients for the shift operators may be stored in advance and used

as needed.

It is clear, however, that the particular manner in which sparseness of the shift

operator may be exploited depends on the application and may be less straightforward

than is indicated above. We present an example of such an application in numerical

analysis. Observing that there are only N log

2

N distinct wavelet coeﬃcients in the

decomposition of all N circulant shifts of a vector of the length N = 2

n

, we construct

an O(N log N) algorithm for computing all of these coeﬃcients. Using this algorithm,

we show that the storage requirements of the fast algorithm for applying the standard

form of a pseudodiﬀerential operator to a vector [2] may be reduced from O(N log N)

to O(log

2

N) signiﬁcant entries.

2. Compactly supported wavelets. In this section, we brieﬂy review the or-

thonormal bases of compactly supported wavelets and set our notation. For the details

we refer to [1].

The orthonormal basis of compactly supported wavelets of L

2

(R) is formed by

the dilation and translation of a single function ψ(x),

ψ

j,k

(x) = 2

−j/2

ψ(2

−j

x −k), (2.1)

where j, k ∈ Z. The function ψ(x) has a companion, the scaling function ϕ(x), and

these functions satisfy the following relations:

ϕ(x) =

√

2

L−1

k=0

h

k

ϕ(2x −k), (2.2)

ψ(x) =

√

2

L−1

k=0

g

k

ϕ(2x −k), (2.3)

where

g

k

= (−1)

k

h

L−k−1

, k = 0, · · · , L −1, (2.4)

and

_

+∞

−∞

ϕ(x)dx = 1. (2.5)

In addition, the function ψ has M vanishing moments

_

+∞

−∞

ψ(x)x

m

dx = 0, m = 0, · · · , M −1. (2.6)

1718 G. BEYLKIN

The number L of coeﬃcients in (2.2) and (2.3) is related to the number of van-

ishing moments M, and for the wavelets in [1], L = 2M. If additional conditions are

imposed (see [2] for an example), then the relation might be diﬀerent, but L is always

even.

The wavelet basis induces a multiresolution analysis on L

2

(R) [4], [5], i.e., the

decomposition of the Hilbert space L

2

(R) into a chain of closed subspaces

· · · ⊂ V

2

⊂ V

1

⊂ V

0

⊂ V

−1

⊂ V

−2

⊂ · · · (2.7)

such that

j∈Z

V

j

= {0},

_

j∈Z

V

j

= L

2

(R). (2.8)

By deﬁning W

j

as an orthogonal complement of V

j

in V

j−1

,

V

j−1

= V

j

⊕W

j

, (2.9)

the space L

2

(R) is represented as a direct sum

L

2

(R) =

j∈Z

W

j

. (2.10)

On each ﬁxed scale j, the wavelets {ψ

j,k

(x)}

k∈Z

form an orthonormal basis of

W

j

and the functions {ϕ

j,k

(x) = 2

−j/2

ϕ(2

−j

x − k)}

k∈Z

form an orthonormal basis

of V

j

.

The coeﬃcients H = {h

k

}

k=L−1

k=0

and G = {g

k

}

k=L−1

k=0

in (2.2) and (2.3) are

quadrature mirror ﬁlters. Once the ﬁlter H has been chosen, it completely determines

the functions ψ and ϕ. Let us deﬁne the 2π-periodic function

m

0

(ξ) = 2

−1/2

k=L−1

k=0

h

k

e

ikξ

(2.11)

where {h

k

}

k=L−1

k=0

are the coeﬃcients of the ﬁlter H. The function m

0

(ξ) satisﬁes the

equation

|m

0

(ξ)|

2

+|m

0

(ξ +π)|

2

= 1. (2.12)

The following lemma characterizes trigonometric polynomial solutions of (2.12)

which correspond to the orthonormal bases of compactly supported wavelets with

vanishing moments.

Lemma 1 (Daubechies [1]). Any trigonometric polynomial solution m

0

(ξ) of

(2.12) is of the form

m

0

(ξ) =

_

1

2

(1 + e

iξ

)

¸

M

Q(e

iξ

), (2.13)

where M ≥ 1 is the number of vanishing moments, and where Q is a polynomial such

that

|Q(e

iξ

)|

2

= P(sin

2

1

2

ξ) + sin

2M

(

1

2

ξ) R(

1

2

cos ξ), (2.14)

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1719

where

P(y) =

k=M−1

k=0

_

M −1 +k

k

_

y

k

, (2.15)

and R is an odd polynomial such that

0 ≤ P(y) +y

M

R(

1

2

−y) for 0 ≤ y ≤ 1, (2.16)

and

sup

0≤y≤1

_

P(y) +y

M

R(

1

2

−y)

¸

< 2

2(M−1)

. (2.17)

3. The operator d/dx in wavelet bases. In this section we construct the

nonstandard form of the operator d/dx. The nonstandard form [2] is a representation

of an operator T as a chain of triplets

T = {A

j

, B

j

, Γ

j

}

j∈Z

(3.1)

acting on the subspaces V

j

and W

j

,

A

j

: W

j

→W

j

, (3.2)

B

j

: V

j

→W

j

, (3.3)

Γ

j

: W

j

→V

j

. (3.4)

The operators {A

j

, B

j

, Γ

j

}

j∈Z

are deﬁned as A

j

= Q

j

TQ

j

, B

j

= Q

j

TP

j

, and Γ

j

=

P

j

TQ

j

, where P

j

is the projection operator on the subspace V

j

and Q

j

= P

j−1

−P

j

is the projection operator on the subspace W

j

.

The matrix elements α

j

il

, β

j

il

, γ

j

il

of A

j

, B

j

, Γ

j

, and r

j

il

of T

j

= P

j

TP

j

, i, l, j ∈ Z,

for the operator d/dx are easily computed as

α

j

il

= 2

−j

_

∞

−∞

ψ(2

−j

x −i) ψ

(2

−j

x −l) 2

−j

dx = 2

−j

α

i−l

, (3.5)

β

j

il

= 2

−j

_

∞

−∞

ψ(2

−j

x −i) ϕ

(2

−j

x −l) 2

−j

dx = 2

−j

β

i−l

, (3.6)

γ

j

il

= 2

−j

_

∞

−∞

ϕ(2

−j

x −i) ψ

(2

−j

x −l) 2

−j

dx = 2

−j

γ

i−l

, (3.7)

and

r

j

il

= 2

−j

_

∞

−∞

ϕ(2

−j

x −i) ϕ

(2

−j

x −l) 2

−j

dx = 2

−j

r

i−l

, (3.8)

where

α

l

=

_

+∞

−∞

ψ(x −l)

d

dx

ψ(x) dx, (3.9)

1720 G. BEYLKIN

β

l

=

_

+∞

−∞

ψ(x −l)

d

dx

ϕ(x) dx, (3.10)

γ

l

=

_

+∞

−∞

ϕ(x −l)

d

dx

ψ(x) dx. (3.11)

and

r

l

=

_

+∞

−∞

ϕ(x −l)

d

dx

ϕ(x) dx. (3.12)

Moreover, using (2.2) and (2.3) we have

α

i

= 2

L−1

k=0

L−1

k

=0

g

k

g

k

r

2i+k−k

, (3.13)

β

i

= 2

L−1

k=0

L−1

k

=0

g

k

h

k

r

2i+k−k

, (3.14)

and

γ

i

= 2

L−1

k=0

L−1

k

=0

h

k

g

k

r

2i+k−k

, (3.15)

and, therefore, the representation of d/dx is completely determined by r

l

in (3.12) or,

in other words, by the representation of d/dx on the subspace V

0

.

Rewriting (3.12) in terms of ˆ ϕ(ξ), where

ˆ ϕ(ξ) =

1

√

2π

_

+∞

−∞

ϕ(x) e

ixξ

dx, (3.16)

we obtain

r

l

=

_

+∞

−∞

(−iξ) | ˆ ϕ(ξ)|

2

e

−ilξ

dξ. (3.17)

In order to compute the coeﬃcients r

l

we ﬁrst note that any trigonometric poly-

nomial m

0

(ξ) satisfying (2.12) is such that

|m

0

(ξ)|

2

=

1

2

+

1

2

L/2

k=1

a

2k−1

cos(2k −1)ξ, (3.18)

where a

n

are the autocorrelation coeﬃcients of H = {h

k

}

k=L−1

k=0

,

a

n

= 2

L−1−n

i=0

h

i

h

i+n

, n = 1, · · · , L −1. (3.19)

The autocorrelation coeﬃcients a

n

with even indices are zero,

a

2k

= 0, k = 1, · · · , L/2 −1. (3.20)

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1721

To prove this assertion we compute |m

0

(ξ)|

2

using (2.11) and obtain

|m

0

(ξ)|

2

=

1

2

+

1

2

L−1

n=1

a

n

cos nξ, (3.21)

where a

n

are given in (3.19). Computing |m

0

(ξ +π)|

2

, we have

|m

0

(ξ +π)|

2

=

1

2

−

1

2

L/2

k=1

a

2k−1

cos(2k −1)ξ +

1

2

L/2−1

k=1

a

2k

cos 2kξ. (3.22)

Combining (3.21) and (3.22) to satisfy (2.12), we obtain

L/2−1

k=1

a

2k

cos 2kξ = 0, (3.23)

and hence, (3.20) and (3.18). (See also Remark 6 about vanishing moments of a

2k−1

.)

We prove the following:

Proposition 1.

(1) If the integrals in (3.12) or (3.17) exist, then the coeﬃcients r

l

in (3.12) satisfy

the following system of linear algebraic equations:

r

l

= 2

_

_

r

2l

+

1

2

L/2

k=1

a

2k−1

(r

2l−2k+1

+r

2l+2k−1

)

_

_

, (3.24)

and

l

l r

l

= −1, (3.25)

where the coeﬃcients a

2k−1

are given in (3.19).

(2) If M ≥ 2, then equations (3.24) and (3.25) have a unique solution with a

ﬁnite number of nonzero r

l

, namely, r

l

= 0 for −L + 2 ≤ l ≤ L −2 and

r

l

= −r

−l

. (3.26)

Remark 1. If M = 1, then equations (3.24) and (3.25) have a unique solution but

the integrals in (3.12) or (3.17) may not be absolutely convergent. Let us consider

Example 3.2 of [1], where L = 4 and

h

0

= 2

−1/2

ν(ν −1)

ν

2

+ 1

, h

1

= 2

−1/2

1 −ν

ν

2

+ 1

, h

2

= 2

−1/2

ν + 1

ν

2

+ 1

, h

3

= 2

−1/2

ν(ν + 1)

ν

2

+ 1

,

where ν is an arbitrary real number. We have

a

1

=

1 + 3ν

2

(ν

2

+ 1)

2

, a

3

=

ν

2

(ν

2

−1)

(ν

2

+ 1)

2

,

and

r

1

= −

(1 +ν

2

)

2

2(3ν

4

+ 1)

, r

2

=

ν

2

(1 −ν

2

)

2(3ν

4

+ 1)

.

1722 G. BEYLKIN

The parameter ν can be chosen so that the Fourier transform ˆ ϕ(ξ) does not have the

suﬃcient decay to insure the absolute convergence of the integral (3.17).

For the Haar basis (h

1

= h

2

= 2

−1/2

), a

1

= 1 and r

1

= −

1

2

, and thus, we obtain

the simplest ﬁnite diﬀerence operator (

1

2

, 0, −

1

2

). In this case the function ϕ is not

continuous and

ˆ ϕ(ξ) =

1

√

2π

sin

1

2

ξ

1

2

ξ

e

i

1

2

ξ

,

so that the integral in (3.17) is not absolutely convergent.

Proof of Proposition 1. Using (2.2) for both ϕ(x − l) and

d

dx

ϕ(x) in (3.12) we

obtain

r

i

= 2

L−1

k=0

L−1

l=0

h

k

h

l

_

+∞

−∞

ϕ(2x −2i −k) ϕ

(2x −l) 2 dx (3.27)

and hence,

r

i

= 2

L−1

k=0

L−1

l=0

h

k

h

l

r

2i+k−l

. (3.28)

Substituting l = k −m, we rewrite (3.28) as

r

i

= 2

L−1

k=0

k−L+1

m=k

h

k

h

k−m

r

2i+m

. (3.29)

Changing the order of summation in (3.29) and using the fact that

L−1

k=0

h

2

k

= 1, we

arrive at

r

l

= 2r

2l

+

L−1

n=1

a

n

(r

2l−n

+r

2l+n

), l ∈ Z, (3.30)

where a

n

are given in (3.19). Using (3.20), we obtain (3.24) from (3.30).

In order to obtain (3.25) we use the following relation:

l=+∞

l=−∞

l

m

ϕ(x −l) = x

m

+

l=m

l=1

(−1)

l

_

m

l

_

M

ϕ

l

x

m−l

, (3.31)

where

M

ϕ

l

=

_

+∞

−∞

ϕ(x) x

l

dx, where l = 1, · · · , m, (3.32)

are the moments of the function ϕ(x). We note that (3.31) is well known if all moments

(3.32) are zero. The general statement follows simply on taking Fourier transforms

and using Leibniz’s rule. Using (3.12) and (3.31) with m = 1, we obtain (3.25).

If M ≥ 2, then

| ˆ ϕ(ξ)|

2

|ξ| ≤ C(1 +|ξ|)

−1−

, (3.33)

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1723

where > 0, and hence, the integral in (3.17) is absolutely convergent. This assertion

follows from Lemma 3.2 of [1], where it is shown that

| ˆ ϕ(ξ)| ≤ C(1 +|ξ|)

−M+log

2

B

, (3.34)

where

B = sup

ξ∈R

|Q(e

iξ

)|.

Due to condition (2.17), we have log

2

B = M −1 − with some > 0. The existence

of the solution of the system of equations (3.24) and (3.25) follows from the existence

of the integral in (3.17). Since the scaling function ϕ has a compact support there are

only a ﬁnite number of nonzero coeﬃcients r

l

. The speciﬁc interval −L+2 ≤ l ≤ L−2

is obtained by the direct examination of (3.24).

Let us show now that

l

r

l

= 0. (3.35)

Multiplying (3.24) by e

ilξ

and summing over l, we obtain

ˆ r(ξ) = 2

_

_

ˆ r

even

(ξ/2) +

1

2

ˆ r

odd

(ξ/2)

L/2

k=1

a

2k−1

(e

−i(2k−1)ξ/2

+ e

i(2k−1)ξ/2

)

_

_

, (3.36)

where

ˆ r(ξ) =

l

r

l

e

ilξ

, (3.37)

ˆ r

even

(ξ/2) =

l

r

2l

e

ilξ

, (3.38)

and

ˆ r

odd

(ξ/2) =

l

r

2l+1

e

i(2l+1) ξ/2

. (3.39)

Noticing that

2 ˆ r

even

(ξ/2) = ˆ r(ξ/2) + ˆ r(ξ/2 +π) (3.40)

and

2 ˆ r

odd

(ξ/2) = ˆ r(ξ/2) − ˆ r(ξ/2 +π), (3.41)

and using (3.18), we obtain from (3.36)

ˆ r(ξ) =

_

ˆ r(ξ/2) + ˆ r(ξ/2 +π) + (2|m

0

(ξ/2)|

2

−1)(ˆ r(ξ/2) − ˆ r(ξ/2 +π))

¸

. (3.42)

Finally, using (2.12) we arrive at

ˆ r(ξ) = 2( |m

0

(ξ/2)|

2

ˆ r(ξ/2) +|m

0

(ξ/2) +π)|

2

ˆ r(ξ/2) +π)). (3.43)

1724 G. BEYLKIN

Setting ξ = 0 in (3.43), we obtain ˆ r(0) = 2ˆ r(0) and thus, (3.35).

Uniqueness of the solution of (3.24) and (3.25) follows from the uniqueness of

the representation of d/dx. Given the solution r

l

of (3.24) and (3.25) we consider

the operator T

j

deﬁned by these coeﬃcients on the subspace V

j

and apply it to a

suﬃciently smooth function f. Since r

j

l

= 2

−j

r

l

(3.8), we have

(T

j

f)(x) =

k∈Z

_

2

−j

l

r

l

f

j,k−l

_

ϕ

j,k

(x), (3.44)

where

f

j,k−l

= 2

−j/2

_

+∞

−∞

f(x) ϕ(2

−j

x −k +l) dx. (3.45)

Rewriting (3.45)

f

j,k−l

= 2

−j/2

_

+∞

−∞

f(x −2

j

l) ϕ(2

−j

x −k) dx, (3.46)

and expanding f(x −2

j

l) in the Taylor series at the point x, we have

f

j,k−l

=

_

+∞

−∞

f(x) ϕ

j,k

(x) dx −2

j

l

_

+∞

−∞

f

(x) ϕ

j,k

(x) dx

+2

2j

l

2

2

_

+∞

−∞

f

(˜ x) ϕ

j,k

(x) dx,

(3.47)

where ˜ x = ˜ x(x, x − 2

j

l) and |˜ x − x| ≤ 2

j

l. Substituting (3.47) in (3.44) and using

(3.35) and (3.25), we obtain

(T

j

f)(x) =

k∈Z

__

+∞

−∞

f

(x) ϕ

j,k

(x) dx

_

ϕ

j,k

(x)

+2

j

k∈Z

_

1

2

l

r

l

l

2

_

+∞

−∞

f

(˜ x) ϕ

j,k

(x) dx

_

ϕ

j,k

(x). (3.48)

It is clear that as j →−∞, operators T

j

and d/dx coincide on smooth functions. Using

standard arguments it is easy to prove that T

−∞

= d/dx and hence, the solution to

(3.24) and (3.25) is unique. The relation (3.26) follows now from (3.17).

Remark 2. We note that expressions (3.13) and (3.14) for α

l

and β

l

(γ

l

= −β

−l

)

may be simpliﬁed by changing the order of summation in (3.13) and (3.14) and

by introducing the correlation coeﬃcients 2

L−1−n

i=0

g

i

h

i+n

, 2

L−1−n

i=0

h

i

g

i+n

and

2

L−1−n

i=0

g

i

g

i+n

. The expression for α

l

is especially simple: α

l

= 4r

2l

−r

l

.

Examples. For examples we will use Daubechies’ wavelets constructed in [1].

First, let us compute the coeﬃcients a

2k−1

, k = 1, · · · , M, where M is the number of

vanishing moments and L = 2M. Using relation (4.22) of [1],

|m

0

(ξ)|

2

= 1 −

(2M −1)!

[(M −1)!]

2

2

2M−1

_

ξ

0

sin

2M−1

ξ dξ, (3.49)

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1725

we ﬁnd, by computing

_

ξ

0

sin

2M−1

ξ dξ, that

|m

0

(ξ)|

2

=

1

2

+

1

2

C

M

M

m=1

(−1)

m−1

cos(2m−1)ξ

(M −m)! (M +m−1)! (2m−1)

, (3.50)

where

C

M

=

_

(2M −1)!

(M −1)! 4

M−1

_

2

. (3.51)

Thus, by comparing (3.50) and (3.18), we have

a

2m−1

=

(−1)

m−1

C

M

(M −m)! (M +m− 1)! (2m−1)

, where m = 1, · · · , M. (3.52)

The coeﬃcients r

e

are rational numbers since they are solutions of a linear system

with rational coeﬃcients (a

2m−1

in (3.52) are rational by construction). We note that

the coeﬃcients r

l

are the same for all Daubechies’ wavelets with a ﬁxed number

of vanishing moments M, notwithstanding the fact that there are several diﬀerent

wavelet bases for a given M (depending on the choice of the roots of polynomials in

the construction described in [1]).

Solving the equations of Proposition 1, we present the results for Daubechies’

wavelets with M = 2, 3, 4, 5, 6.

1. M = 2.

a

1

=

9

8

, a

3

= −

1

8

,

and

r

1

= −

2

3

, r

2

=

1

12

.

The coeﬃcients (−1/12, 2/3, 0, −2/3, 1/12) of this example coincide with one of the

standard choices of coeﬃcients for numerical diﬀerentiation.

2. M = 3.

a

1

=

75

64

, a

3

= −

25

128

, a

5

=

3

128

,

and

r

1

= −

272

365

, r

2

=

53

365

, r

3

= −

16

1095

, r

4

= −

1

2920

.

3. M = 4.

a

1

=

1225

1024

, a

3

= −

245

1024

, a

5

=

49

1024

, a

7

= −

5

1024

,

and

r

1

= −

39296

49553

, r

2

=

76113

396424

, r

3

= −

1664

49553

,

r

4

=

2645

1189272

, r

5

=

128

743295

, r

6

= −

1

1189272

.

1726 G. BEYLKIN

4. M = 5.

a

1

=

19845

16384

, a

3

= −

2205

8192

, a

5

=

567

8192

, a

7

= −

405

32768

, a

9

=

35

32768

,

and

r

1

= −

957310976

1159104017

, r

2

=

265226398

1159104017

, r

3

= −

735232

13780629

,

r

4

=

17297069

2318208034

, r

5

= −

1386496

5795520085

, r

6

= −

563818

10431936153

,

r

7

= −

2048

8113728119

, r

8

= −

5

18545664272

.

5. M = 6.

a

1

=

160083

131072

, a

3

= −

38115

131072

, a

5

=

22869

262144

,

a

7

= −

5445

262144

, a

9

=

847

262144

, a

11

= −

63

262144

,

and

r

1

=

3986930636128256

4689752620280145

, r

2

=

4850197389074509

18759010481120580

, r

3

=

1019185340268544

14069257860840435

,

r

4

=

136429697045009

9379505240560290

, r

5

=

7449960660992

4689752620280145

, r

6

=

483632604097

112554062886723480

,

r

7

=

78962327552

6565653668392203

, r

8

=

31567002859

75036041924482320

, r

9

=

2719744

937950524056029

,

r

10

=

1743

2501201397482744

.

Iterative algorithm for computing the coeﬃcients r

l

. We also use an itera-

tive algorithm as a way of solving equations (3.24) and (3.25). We start with r

−1

= 0.5

and r

1

= −0.5, and iterate using (3.24) to recompute r

l

. Using (3.43), it is easy to

verify that (3.25) and (3.26) are satisﬁed due to the choice of initialization. Table 1

was computed using this algorithm for Daubechies’ wavelets with M = 5, 6, 7, 8, 9. It

only displays the coeﬃcients {r

l

}

L−2

l=1

since r

−l

= −r

l

and r

0

= 0.

4. The operator d

n

/dx

n

in wavelet bases. Again, as in the case with the

operator d/dx, the nonstandard form of the operator d

n

/dx

n

is completely determined

by its representation on the subspace V

0

, i.e., by the coeﬃcients

r

(n)

l

=

_

+∞

−∞

ϕ(x −l)

d

n

dx

n

ϕ(x) dx, l ∈ Z, (4.1)

or, alternatively,

r

(n)

l

=

_

+∞

−∞

(−iξ)

n

| ˆ ϕ(ξ)|

2

e

−ilξ

dξ (4.2)

if the integrals in (4.1) or (4.2) exist (see also Remark 3).

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1727

Table 1

The coeﬃcients {r

l

}

l=L−2

l=1

for Daubechies’ wavelets, where L = 2M and M = 5, · · · , 9.

Coeﬃcients Coeﬃcients

l r

l

l r

l

M = 5 1 -0.82590601185015 M = 8 1 -0.88344604609097

2 0.22882018706694 2 0.30325935147672

3 -5.3352571932672E-02 3 -0.10636406828947

4 7.4613963657755E-03 4 3.1290147839488E-02

5 -2.3923582002393E-04 5 -6.9583791164537E-03

6 -5.4047301644748E-05 6 1.0315302133757E-03

7 -2.5241171135682E-07 7 -7.6677069083796E-05

8 -2.6960479423517E-10 8 -2.4519921109537E-07

9 -3.9938104563894E-08

10 7.2079482385949E-08

M = 6 1 -0.85013666155592 11 9.6971849256415E-10

2 0.25855294414146 12 7.2522069166503E-13

3 -7.2440589997659E-02 13 -1.2400785360984E-14

4 1.4545511041994E-02 14 1.5854647516841E-19

5 -1.5885615434757E-03

6 4.2968915709948E-06

7 1.2026575195723E-05 M = 9 1 -0.89531640583699

8 4.2069120451167E-07 2 0.32031206224855

9 -2.8996668057051E-09 3 -0.12095364936000

10 6.9686511520083E-13 4 3.9952721886694E-02

5 -1.0616930669821E-02

6 2.1034028106558E-03

M = 7 1 -0.86874391452377 7 -2.7812077649932E-04

2 0.28296509452594 8 1.9620437763642E-05

3 -9.0189066217795E-02 9 -4.8782468879634E-07

4 2.2687411014648E-02 10 1.0361220591478E-07

5 -3.8814546576295E-03 11 -1.5966864798639E-08

6 3.3734404776409E-04 12 -8.1374108294110E-10

7 4.2363946800701E-06 13 -5.4025197533630E-13

8 -1.6501679210868E-06 14 -4.7814005916812E-14

9 -2.1871130331900E-07 15 -1.6187880013009E-18

10 4.1830548203747E-10 16 -4.8507474310747E-24

11 -1.2035273999989E-11

12 -6.6283900594600E-16

1728 G. BEYLKIN

Proposition 2.

(1) If the integrals in (4.1) or (4.2) exist, then the coeﬃcients r

(n)

l

, l ∈ Z satisfy

the following system of linear algebraic equations:

r

(n)

l

= 2

n

_

_

r

2l

+

1

2

L/2

k=1

a

2k−1

(r

(n)

2l−2k+1

+r

(n)

2l+2k−1

)

_

_

, (4.3)

and

l

l

n

r

(n)

l

= (−1)

n

n!, (4.4)

where a

2k−1

are given in (3.19).

(2) Let M ≥ (n + 1)/2, where M is the number of vanishing moments in (2.6).

If the integrals in (4.1) or (4.2) exist, then the equations (4.3) and (4.4) have

a unique solution with a ﬁnite number of nonzero coeﬃcients r

(n)

l

, namely,

r

(n)

l

= 0 for −L + 2 ≤ l ≤ L −2, such that for even n

r

(n)

l

= r

(n)

−l

, (4.5)

l

l

2˜ n

r

(n)

l

= 0, ˜ n = 1, · · · , n/2 −1, (4.6)

and

l

r

(n)

l

= 0, (4.7)

and for odd n

r

(n)

l

= −r

(n)

−l

, (4.8)

l

l

2˜ n−1

r

(n)

l

= 0, ˜ n = 1, · · · , (n −1)/2. (4.9)

The proof of Proposition 2 is completely analogous to that of Proposition 1.

Remark 3. The linear system in Proposition 2 may have a unique solution

whereas integrals (4.1) and (4.2) are not absolutely convergent. A case in point

is the Daubechies’ wavelet with M = 2. The representation of the ﬁrst derivative in

this basis is described in the previous section. Equations (4.3) and (4.4) do not have

a solution for the second derivative n = 2. However, the system of equations (4.3)

and (4.4) has a solution for the third derivative n = 3. We have

a

1

=

9

8

, a

3

= −

1

8

,

and

r

−2

= −

1

2

, r

−1

= 1, r

0

= 0, r

1

= −1, r

2

=

1

2

.

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1729

The set of coeﬃcients (−1/2, 1, 0, −1, 1/2) is one of the standard choices of ﬁnite

diﬀerence coeﬃcients for the third derivative.

We note that among the wavelets with L = 4, the wavelets with two vanishing

moments M = 2 do not have the best H¨ older exponent (see [6]), but the representation

of the third derivative exists only if the number of vanishing moments M = 2.

Remark 4. Let us derive an equation generalizing to (3.43) for d

n

/dx

n

directly

from (4.2). We rewrite (4.2) as

r

(n)

l

=

_

2π

0

k∈Z

| ˆ ϕ(ξ + 2πk)|

2

(−i)

n

(ξ + 2πk)

n

e

−ilξ

dξ. (4.10)

Therefore,

ˆ r(ξ) =

k∈Z

| ˆ ϕ(ξ + 2πk)|

2

(−i)

n

(ξ + 2πk)

n

, (4.11)

where

ˆ r(ξ) =

l

r

(n)

l

e

ilξ

. (4.12)

Substituting the relation

ˆ ϕ(ξ) = m

0

(ξ/2) ˆ ϕ(ξ/2) (4.13)

into the right-hand side of (4.11), and summing separately over even and odd indices

in (4.11), we arrive at

ˆ r(ξ) = 2

n

( |m

0

(ξ/2)|

2

ˆ r(ξ/2) +|m

0

(ξ/2) +π)|

2

ˆ r(ξ/2) +π)). (4.14)

By considering the operator M

0

deﬁned on 2π-periodic functions,

(M

0

f)(ξ) = |m

0

(ξ/2)|

2

f(ξ/2) +|m

0

(ξ/2) +π)|

2

f(ξ/2) +π), (4.15)

we rewrite (4.14) as

M

0

ˆ r = 2

−n

ˆ r. (4.16)

Thus, ˆ r is an eigenvector of the operator M

0

corresponding to the eigenvalue 2

−n

and,

therefore, ﬁnding the representation of the derivatives in the wavelet basis is equivalent

to ﬁnding trigonometric polynomial solutions of (4.16) and vice versa. (The operator

M

0

is also introduced in [7] and [8], where the problem (4.16) with eigenvalue 1 is

considered.)

Remark 5. While theoretically it is well understood that the derivative operators

(or, more generally, operators with homogeneous symbols) have an explicit diagonal

preconditioner in wavelet bases, the numerical evidence illustrating this fact is of

interest, since it represents one of the advantages of computing in the wavelet bases.

If an operator has a null space (the actual null space or a null space for a given ac-

curacy), then by the condition number we understand the ratio of the largest singular

value to the smallest singular value above the threshold of accuracy. Thus, we include

the situation where the operator may be preconditioned only on a subspace. We note

that the preconditioning described here addresses the problem of ill conditioning due

1730 G. BEYLKIN

only to the unfavorable homogeneity of the symbol and does not aﬀect ill conditioning

due to other causes.

For periodized derivative operators the bound on the condition number depends

only on the particular choice of the wavelet basis. After applying such a precondi-

tioner, the condition number κ

p

of the operator is uniformly bounded with respect

to the size of the matrix. We recall that the condition number controls the rate of

convergence of a number of iterative algorithms; for example, the number of iterations

of the conjugate gradient method is O(

√

κ

p

). Thus, this remark implies a completely

new outlook on a number of numerical methods, a topic we will address elsewhere.

We present here two tables illustrating such preconditioning applied to the stan-

dard form of the second derivative (see [2] on how to compute the standard form from

the nonstandard form). In the following examples the standard form of the periodized

second derivative D

2

of size N ×N, where N = 2

n

, is preconditioned by the diagonal

matrix P,

D

p

2

= PD

2

P

where P

il

= δ

il

2

j

, 1 ≤ j ≤ n, and where j is chosen depending on i, l so that

N −N/2

j−1

+ 1 ≤ i, l ≤ N −N/2

j

, and P

NN

= 2

n

.

Tables 2 and 3 compare the original condition number κ of D

2

and κ

p

of D

p

2

.

Table 2

Condition numbers of the matrix of periodized second derivative (with and without preconditioning)

in the basis of Daubechies’ wavelets with three vanishing moments M = 3.

N κ κp

64 0.14545E+04 0.10792E+02

128 0.58181E+04 0.11511E+02

256 0.23272E+05 0.12091E+02

512 0.93089E+05 0.12604E+02

1024 0.37236E+06 0.13045E+02

5. Convolution operators. For convolution operators, the computation of the

nonstandard form is considerably simpler than in the general case [2]. We will demon-

strate that the quadrature formulas for representing kernels of convolution operators

on V

0

(see, e.g., Appendix B of [2]) are of the simplest form due to the fact that the

moments of the autocorrelation function of the scaling function ϕ vanish. Moreover,

by combining the asymptotics of wavelet coeﬃcients with the system of linear algebraic

equations (similar to those in previous sections), we arrive at an eﬀective method for

computing representations of convolution operators. This method is especially simple

if the symbol of the operator is homogeneous of some degree.

Let us assume that the matrix t

(j−1)

i−l

(i, l ∈ Z) represents the operator P

j−1

TP

j−1

on the subspace V

j−1

. To compute the matrix representation of P

j

TP

j

, we have the

following formula (3.26) of [2]:

t

(j)

l

=

L−1

k=0

L−1

m=0

h

k

h

m

t

(j−1)

2l+k−m

, (5.1)

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1731

Table 3

Condition numbers of the matrix of periodized second derivative (with and without preconditioning)

in the basis of Daubechies’ wavelets with six vanishing moments M = 6.

N κ κp

64 0.10472E+04 0.43542E+01

128 0.41886E+04 0.43595E+01

256 0.16754E+05 0.43620E+01

512 0.67018E+05 0.43633E+01

1024 0.26807E+06 0.43640E+01

which easily reduces to

t

(j)

l

= t

(j−1)

2l

+

1

2

L/2

k=0

a

2k−1

(t

(j−1)

2l−2k+1

+t

(j−1)

2l+2k−1

), (5.2)

where the coeﬃcients a

2k−1

are given in (3.19).

We also have

t

(j)

l

=

_

+∞

−∞

_

+∞

−∞

K(x −y) ϕ

j,0

(y) ϕ

j,l

(x) dxdy, (5.3)

and by changing the order of integration, we obtain

t

(j)

l

= 2

j

_

+∞

−∞

K(2

j

(l −y)) Φ(y) dy, (5.4)

where Φ is the autocorrelation function of the scaling function ϕ,

Φ(y) =

_

+∞

−∞

ϕ(x) ϕ(x −y) dx. (5.5)

Let us verify that

_

+∞

−∞

Φ(y)dy = 1 (5.6)

and

M

m

Φ

=

_

+∞

−∞

y

m

Φ(y)dy = 0 for 1 ≤ m ≤ 2M −1. (5.7)

Clearly, we have

M

m

Φ

=

__

1

i

∂

ξ

_

m

| ˆ ϕ(ξ)|

2

_

ξ=0

. (5.8)

1732 G. BEYLKIN

Using (5.8) and the identity ˆ ϕ(ξ) = ˆ ϕ(ξ/2)m

0

(ξ/2) (see [1]), it is clear that (5.7) holds

provided that

__

1

i

∂

ξ

_

m

|m

0

(ξ)|

2

_

ξ=0

= 0 for 1 ≤ m ≤ 2M −1, (5.9)

or (due to (2.12))

__

1

i

∂

ξ

_

m

|m

0

(ξ +π)|

2

_

ξ=0

= 0 for 0 ≤ m ≤ 2M −1. (5.10)

But formula (5.10) follows from the explicit representation in (2.13).

Remark 6. Equations (5.9) and (3.21) also imply that even moments of the

coeﬃcients a

2k−1

from (3.19) vanish, namely,

k=L/2

k=1

a

2k−1

(2k −1)

2m

= 0 for 1 ≤ m ≤ M −1. (5.11)

Since the moments of the function Φ vanish equation (5.4) leads to a one-point

quadrature formula for computing the representation of convolution operators on

the ﬁnest scale. This formula is obtained in exactly the same manner as for the

special choice of the wavelet basis described in [2, eqns. (3.8)–(3.12)], where the shifted

moments of the function ϕ vanish; we refer to this paper for the details.

Here we introduce a diﬀerent approach for computing representations of convo-

lution operators in the wavelet basis which consists of solving the system of linear

algebraic equations (5.2) subject to asymptotic conditions. This method is especially

simple if the symbol of the operator is homogeneous of some degree since in this

case the operator is completely deﬁned by its representation on V

0

. We consider

two examples of such operators, the Hilbert transform and the operator of fractional

diﬀerentiation (or antidiﬀerentiation).

The Hilbert transform. We apply our method to the computation of the

nonstandard form of the Hilbert transform

g(x) = (Hf)(y) =

1

π

p.v.

_

∞

−∞

f(s)

s −x

ds, (5.12)

where p.v. denotes a principal value at s = x.

The representation of H on V

0

is deﬁned by the coeﬃcients

r

l

=

_

∞

−∞

ϕ(x −l) (Hϕ)(x) dx, l ∈ Z, (5.13)

which, in turn, completely deﬁne all other coeﬃcients of the nonstandard form.

Namely, H = {A

j

, B

j

, Γ

j

}

j∈Z

, A

j

= A

0

, B

j

= B

0

, and Γ

j

= Γ

0

, where matrix

elements α

i−l

, β

i−l

, and γ

i−l

of A

0

, B

0

, and Γ

0

are computed from the coeﬃcients

r

l

,

α

i

=

L−1

k=0

L−1

k

=0

g

k

g

k

r

2i+k−k

, (5.14)

β

i

=

L−1

k=0

L−1

k

=0

g

k

h

k

r

2i+k−k

, (5.15)

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1733

and

γ

i

=

L−1

k=0

L−1

k

=0

h

k

g

k

r

2i+k−k

. (5.16)

The coeﬃcients r

l

, l ∈ Z in (5.13) satisfy the following system of linear algebraic

equations:

r

l

= r

2l

+

1

2

L/2

k=1

a

2k−1

(r

2l−2k+1

+r

2l+2k−1

), (5.17)

where the coeﬃcients a

2k−1

are given in (3.19). Using (5.4), (5.6), and (5.7), we

obtain the asymptotics of r

l

for large l,

r

l

= −

1

πl

+O

_

1

l

2M

_

. (5.18)

By rewriting (5.13) in terms of ˆ ϕ(ξ),

r

l

= −2

_

∞

0

| ˆ ϕ(ξ)|

2

sin(lξ) dξ. (5.19)

we obtain r

l

= −r

−l

and set r

0

= 0. We note that the coeﬃcient r

0

cannot be

determined from equations (5.17) and (5.18).

Solving (5.17) with the asymptotic condition (5.18), we compute the coeﬃcients

r

l

, l = 0 with any prescribed accuracy. We note that the generalization for computing

the coeﬃcients of Riesz transforms in higher dimensions is straightforward.

Example. We compute (see Table 4) the coeﬃcients r

l

of the Hilbert transform for

Daubechies’ wavelets with six vanishing moments with accuracy 10

−7

. The coeﬃcients

for l > 16 are obtained using asymptotics (5.18). (We note that r

−l

= −r

l

and r

0

= 0.)

Table 4

The coeﬃcients r

l

, l = 1, · · · , 16 of the Hilbert transform for Daubechies’ wavelet with six vanishing

moments.

Coeﬃcients Coeﬃcients

l r

l

l r

l

M = 6 1 -0.588303698 9 -0.035367761

2 -0.077576414 10 -0.031830988

3 -0.128743695 11 -0.028937262

4 -0.075063628 12 -0.026525823

5 -0.064168018 13 -0.024485376

6 -0.053041366 14 -0.022736420

7 -0.045470650 15 -0.021220659

8 -0.039788641 16 -0.019894368

1734 G. BEYLKIN

Fractional derivatives. We use the following deﬁnition of fractional deriva-

tives:

(∂

α

x

f) (x) =

_

+∞

−∞

(x −y)

−α−1

+

Γ(−α)

f(y)dy, (5.20)

where we consider α = 1, 2 · · ·. If α < 0, then (5.20) deﬁnes fractional antiderivatives.

The representation of ∂

α

x

on V

0

is determined by the coeﬃcients

r

l

=

_

+∞

−∞

ϕ(x −l) (∂

α

x

ϕ) (x) dx, l ∈ Z, (5.21)

provided that this integral exists.

The nonstandard form ∂

α

x

= {A

j

, B

j

, Γ

j

}

j∈Z

is computed via A

j

= 2

−αj

A

0

,

B

j

= 2

−αj

B

0

, and Γ

j

= 2

−αj

Γ

0

, where matrix elements α

i−l

, β

i−l

, and γ

i−l

of A

0

,

B

0

, and Γ

0

are obtained from the coeﬃcients r

l

,

α

i

= 2

α

L−1

k=0

L−1

k

=0

g

k

g

k

r

2i+k−k

, (5.22)

β

i

= 2

α

L−1

k=0

L−1

k

=0

g

k

h

k

r

2i+k−k

, (5.23)

and

γ

i

= 2

α

L−1

k=0

L−1

k

=0

h

k

g

k

r

2i+k−k

. (5.24)

It is easy to verify that the coeﬃcients r

l

satisfy the following system of linear

algebraic equations:

r

l

= 2

α

_

_

r

2l

+

1

2

L/2

k=1

a

2k−1

(r

2l−2k+1

+r

2l+2k−1

)

_

_

, (5.25)

where the coeﬃcients a

2k−1

are given in (3.19). Using (5.4), (5.6), and (5.7), we

obtain the asymptotics of r

l

for large l,

r

l

=

1

Γ(−α)

1

l

1+α

+O

_

1

l

1+α+2M

_

for l > 0, (5.26)

r

l

= 0 for l < 0. (5.27)

Example. We compute (see Table 5) the coeﬃcients r

l

of the fractional derivative

with α = 0.5 for Daubechies’ wavelets with six vanishing moments with accuracy

10

−7

. The coeﬃcients for r

l

, l > 14, or l < −7 are obtained using asymptotics

r

l

= −

1

2

√

π

1

l

1+

1

2

+O

_

1

l

13+

1

2

_

for l > 0, (5.28)

r

l

= 0 for l < 0. (5.29)

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1735

Table 5

The coeﬃcients {r

l

}

l

, l = −7, · · · , 14 of the fractional derivative α = 0.5 for Daubechies’ wavelet

with six vanishing moments.

Coeﬃcients Coeﬃcients

l r

l

l r

l

M = 6 -7 -2.82831017E-06 4 -2.77955293E-02

-6 -1.68623867E-06 5 -2.61324170E-02

-5 4.45847796E-04 6 -1.91718816E-02

-4 -4.34633415E-03 7 -1.52272841E-02

-3 2.28821728E-02 8 -1.24667403E-02

-2 -8.49883759E-02 9 -1.04479500E-02

-1 0.27799963 10 -8.92061945E-03

0 0.84681966 11 -7.73225246E-03

1 -0.69847577 12 -6.78614593E-03

2 2.36400139E-02 13 -6.01838599E-03

3 -8.97463780E-02 14 -5.38521459E-03

6. Shift operator on V

0

and fast wavelet decomposition of all circulant

shifts of a vector. Let us consider a shift by one on the subspace V

0

represented

by the matrix

t

(0)

i−j

= δ

i−j,1

, (6.1)

where δ is the Kronecker symbol. Using (5.1) with the a

n

of (3.19) we have

t

(0)

l

= δ

l,1

, t

(1)

l

=

1

2

a

|2l−1|

, · · · . (6.2)

The only nonzero coeﬃcients t

(j)

l

on each scale j are those with indices −L+2 ≤ l ≤

L − 2. Also, t

(j)

l

→ δ

l,0

as j → ∞. As an example, the following Table 6 contains

the coeﬃcients t

(j)

l

, j = 1, 2, · · · , 8, for the shift operator in Daubechies’ wavelet basis

with three vanishing moments.

We note that the shift by an integer other than one is treated similarly. However,

if the absolute value of the shift is greater than L−2, then, on the ﬁrst several scales

j, there are nonzero coeﬃcients t

(j)

l

with l outside the interval |l| ≤ L − 2. As j

increases, all the nonzero coeﬃcients t

(j)

l

will have indices in the interval |l| ≤ L −1.

The importance of the shift operator stems from the fact that the coeﬃcients of

wavelet transforms are not shift invariant. However, as we have just demonstrated,

the nonstandard (and, therefore, the standard) forms of the shift operator are sparse

and easy to compute. By applying these sparse representations directly to the wavelet

coeﬃcients, in many applications we can eﬀectively compensate the absence of the

shift invariance of the wavelet transforms. For example, if the representation of a

vector in the wavelet basis is sparse, there is a corresponding reduction in the number

of operations required to shift such a vector. Speciﬁcally, in image processing the shift

1736 G. BEYLKIN

Table 6

The coeﬃcients {t

(j)

l

}

l=L−2

l=−L+2

for Daubechies’ wavelet with three vanishing moments, where L = 6

and j = 1, · · · , 8.

Coeﬃcients Coeﬃcients

l t

(j)

l

l t

(j)

l

j = 1 -4 0. j = 5 -4 -8.3516169979703E-06

-3 0. -3 -4.0407157939626E-04

-2 1.171875E-02 -2 4.1333660119562E-03

-1 -9.765625E-02 -1 -2.1698923046642E-02

0 0.5859375 0 0.99752855458064

1 0.5859375 1 2.4860978555807E-02

2 -9.765625E-02 2 -4.9328931709169E-03

3 1.171875E-02 3 5.0836550508393E-04

4 0. 4 1.2974760466022E-05

j = 2 -4 0. j = 6 -4 -4.7352138210499E-06

-3 -1.1444091796875E-03 -3 -2.1482413927743E-04

-2 1.6403198242188E-02 -2 2.1652627381741E-03

-1 -1.0258483886719E-01 -1 -1.1239479930566E-02

0 0.87089538574219 0 0.99937113652686

1 0.26206970214844 1 1.2046257104714E-02

2 -5.1498413085938E-02 2 -2.3712690179423E-03

3 5.7220458984375E-03 3 2.4169452359502E-04

4 1.3732910156250E-04 4 5.9574082627023E-06

j = 3 -4 -1.3411045074463E-05 j = 7 -4 -2.5174703821573E-06

-3 -1.0904073715210E-03 -3 -1.1073373558501E-04

-2 1.2418627738953E-02 -2 1.1081638044863E-03

-1 -6.9901347160339E-02 -1 -5.7198034904338E-03

0 0.96389651298523 0 0.99984123346637

1 0.11541545391083 1 5.9237906308573E-03

2 -2.3304820060730E-02 2 -1.1605296576369E-03

3 2.5123357772827E-03 3 1.1756409462604E-04

4 6.7055225372314E-05 4 2.8323576983791E-06

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1737

j = 4 -4 -1.2778211385012E-05 j = 8 -4 -1.2976609638869E-06

-3 -7.1267131716013E-04 -3 -5.6215105787797E-05

-2 7.5265066698194E-03 -2 5.6059346249153E-04

-1 -4.0419702418149E-02 -1 -2.8852840759448E-03

0 0.99042607471347 0 0.99996009015421

1 5.2607019431889E-02 1 2.9366035254748E-03

2 -1.0551069863141E-02 2 -5.7380655655486E-04

3 1.1071795597672E-03 3 5.7938552839535E-05

4 2.9441434890032E-05 4 1.3777042338989E-06

operator allows us to “move” pictures in the “compressed” form. The coeﬃcients

t

(j)

l

for the shift operators can be stored in advance and used as needed. It is clear,

however, that the method of using sparseness of the shift operator depends on the

speciﬁc application and may be less straightforward than is indicated above.

The following is an example of an application where, instead of computing shift

operators, we compute all possible shifts. We describe a fast algorithm for the wavelet

decomposition of all circulant shifts of a vector and then show how it may be used to

reduce storage requirements of one of the algorithms of [2].

We recall that the decomposition of a vector of length N = 2

n

into a wavelet

basis requires O(N) operations. Since the coeﬃcients are not shift invariant, the

computation of the wavelet expansion of all N circulant shifts of a vector appears to

require O(N

2

) operations.

We notice, however, that there are only N log

2

(N) distinct coeﬃcients and present

here a simple algorithm to compute the wavelet expansion of N circulant shifts of a

vector in N log(N) operations.

We recall the pyramid scheme

{s

0

k

} −→ {s

1

k

} −→ {s

2

k

} −→ {s

3

k

} · · ·

{d

1

k

} {d

2

k

} {d

3

k

} · · ·

(6.3)

where the coeﬃcients s

0

k

for k = 1, 2, · · · , N are given,

s

j

k

=

n=L−1

n=0

h

n

s

j−1

n+2k−1

, (6.4)

d

j

k

=

n=L−1

n=0

g

n

s

j−1

n+2k−1

, (6.5)

and s

j

k

and d

j

k

are periodic sequences with the period 2

n−j

, j = 0, 1, · · · , n.

In the pyramid scheme (6.3), on each scale j we compute one vector of diﬀerences

{d

j

k

}

k=2

n−j

k=1

and one vector of averages {s

j

k

}

k=2

n−j

k=1

. Instead, let us compute on each

scale j, (1 ≤ j ≤ n), 2

j

vectors of diﬀerences and 2

j

vectors of averages. We proceed

1738 G. BEYLKIN

as follows: let s

j−1

k

, k = 1, · · · , 2

n−j

be one of the vectors of averages on the previous

scale j −1 and compute

s

j

k

(0) =

n=L−1

n=0

h

n

s

j−1

n+2k−1

, (6.6)

s

j

k

(1) =

n=L−1

n=0

h

n

s

j−1

n+2k

, (6.7)

and

d

j

k

(0) =

n=L−1

n=0

g

n

s

j−1

n+2k−1

, (6.8)

d

j

k

(1) =

n=L−1

n=0

g

n

s

j−1

n+2k

. (6.9)

To compute the sum in (6.7) and (6.9), we shift by one the sequence s

j−1

k

in (6.6)

and (6.8).

Thus, stepping from scale to scale we double the number of vectors of averages

and of diﬀerences and, at the same time, halve the length of each of them. Therefore,

the total number of operations in this computation is O(N log N).

Let us organize the vectors of diﬀerences and averages as follows: on the ﬁrst

scale, j = 1, we set

v

1

= (d

1

k

(0), d

1

k

(1)) (6.10)

and

u

1

= (s

1

k

(0), s

1

k

(1)), (6.11)

where d

1

k

(0), d

1

k

(1), s

1

k

(0), and s

1

k

(1) are computed from s

0

k

according to (6.6)–(6.9).

On the second scale, j = 2, we set

v

2

= (d

2

k

(00), d

2

k

(01), d

2

k

(10), d

2

k

(11)) (6.12)

and

u

2

= (s

2

k

(00), s

2

k

(01), s

2

k

(10), s

2

k

(11)), (6.13)

where d

2

k

(00), d

2

k

(01), s

2

k

(00), s

2

k

(01) are computed from s

1

k

(0) according to (6.6)–

(6.9) and d

2

k

(10), d

2

k

(11), s

2

k

(10), and s

2

k

(11) from s

1

k

(1), etc. We claim that we have

computed all the coeﬃcients of the wavelet expansion of N circulant shifts of the

vector s

0

k

, k = 1, 2, · · · , N.

Indeed, the periodic sequence d

1

k

(0) contains all the coeﬃcients that appear if

s

0

k

is circulantly shifted by 2, 4, · · · , 2

n

(see (6.8)), and the periodic sequence d

1

k

(1)

contains all the coeﬃcients for odd shifts by 1, 3, · · · , 2

n

−1 (see (6.9)). By the same

token, the periodic sequences s

1

k

(0) and s

1

k

(1) contain all possible coeﬃcients for even

and odd circulant shifts of s

0

k

(see (6.6) and (6.7)). Repeating this procedure on the

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1739

next scale for both s

1

k

(0) and s

1

k

(1), we again obtain all possible coeﬃcients for odd

and even shifts which we collect in v

2

and u

2

, etc.

While the vectors v

1

, v

2

, · · · , v

n

contain all the coeﬃcients, these coeﬃcients are

not organized sequentially. In order to access them, we generate two tables i

loc

(i

s

, j)

and i

b

(i

s

, j) in O(N log N) operations as follows. For each shift i

s

, 0 ≤ i

s

≤ N −1 of

the vector s

0

k

, k = 1, 2, · · · , N, let us write the binary expansion of i

s

,

i

s

=

l=n−1

l=0

l

2

l

, (6.14)

where

l

= 0, 1. For a ﬁxed scale j, 1 ≤ j ≤ n, we compute

i

loc

(i

s

, j) =

l=j−1

l=0

l

2

l

, (6.15)

and

i

b

(i

s

, j) =

l=j

l=n−1

l

2

l

, (6.16)

where i

b

(i

s

, j) = 0 if j = n. The number i

b

(i

s

, j) points to the begining of the subvec-

tor of diﬀerences in v

j

. Namely, the subvector of v

j

has indices between i

b

(i

s

, j) + 1

and i

b

(i

s

, j) +2

n−j

. Within this subvector (which is treated as a periodic vector with

the period 2

n−j

) the number i

loc

(i

s

, j) points to the ﬁrst element.

For all scales j, 1 ≤ j ≤ n, and shifts i

s

, 0 ≤ i

s

≤ N−1, we compute two tables in

(6.15) and (6.16). These tables give us the direct access to the coeﬃcients in vectors

v

1

, v

2

, · · · , v

n

for a constant cost per element.

We now brieﬂy describe one of the applications of the algorithm for the fast

wavelet decomposition of all circulant shifts of a vector in numerical analysis. The al-

gorithms of [2] are designed to evaluate the Calderon–Zygmund or a pseudodiﬀerential

operator T with kernel K(x, y),

g(x) =

_

+∞

−∞

K(x, y) f(y) dy (6.17)

by constructing (for any ﬁxed accuracy) its sparse nonstandard or standard form and

thereby, reducing the cost of applying it to a function.

Let us rewrite (6.17) as

g(x) =

_

+∞

−∞

K(x, x −z) f(x −z) dz. (6.18)

If the operator T is a convolution, then K(x, x − z) = K(z) is a function of z

only. The nonstandard form of a convolution requires at most O(log N) of stor-

age (see the previous section), while the standard form of [2] will contain O(N) or

O(N log N) signiﬁcant entries even for a convolution. Alternatively, the standard form

of K(x, x −z) = K(z) in variables x and z for the convolution operators contains no

more than O(log N) signiﬁcant entries for any ﬁxed accuracy, since the kernel depends

on one variable only.

1740 G. BEYLKIN

If we now construct the standard form of K(x, x−z) in variables x and z for pseu-

dodiﬀerential operators (not necessarily convolutions), we obtain “super”-compression

of the operator. Indeed, if these operators are represented in the form (6.18), then

the dependence of the kernel K(x, ·) on x is smooth and the number of signiﬁcant

entries in the standard form is of O(log

2

N).

The apparent diﬃculty in computing via (6.18) is that it is necessary to com-

pute the wavelet decomposition of f(x−z) for every x and thus, it appears to require

O(N

2

) operations. The algorithm of this section accomplishes this task in O(N log N)

operations. Therefore, the cost of evaluating (6.18) does not exceed O(N log N). The

advantage of such an algorithm is the reduced storage of O(log

2

N) signiﬁcant entries

for the standard form, which is an important consideration for the multidimensional

operators. We note that the extension of the algorithm for the fast wavelet decompo-

sition of all circulant shifts of a vector to the multidimensional case is straightforward.

REFERENCES

[1] I. Daubechies, Orthonormal bases of compactly supported wavelets, Comm. Pure Appl. Math.,

41 (1988), pp. 909–996.

[2] G. Beylkin, R. Coifman, and V. Rokhlin, Fast wavelet transform and numerical algorithms

I., Tech. Report YALEU/DCS/RR-696, Yale University, New Haven, August 1989; Comm.

Pure Appl. Math., 44 (1991) pp. 141–183.

[3] Y. Meyer, Le calcul scientiﬁque, les ondelettes et ﬁltres miroirs en quadrature, CEREMADE,

Universit´e Paris Dauphine, No. 9007.

[4] , Ondelettes et functions splines, Tech. Report, S´eminaire EDP, Ecole Polytechnique,

Paris, France, 1986.

[5] S. Mallat, Multiresolution approximation and wavelets, Tech. Report, GRASP Lab, Dept. of

Computer and Information Science, University of Pennsylvania, Philadelphia, PA.

[6] I. Daubechies and J. Lagarias, Two-scale diﬀerence equations, I. Global regularity of so-

lutions, preprint, Two-scale diﬀerence equations, II. Local regularity, inﬁnite products of

matrices and fractals, preprint.

[7] A. Cohen, I. Daubechies, and J. C. Feauveau, Biorthogonal bases of compactly supported

wavelets, preprint.

[8] W. M. Lawton, Necessary and suﬃcient conditions for constructing orthonormal wavelet

bases, J. Math. Phys., 32 (1991), pp. 57–61.

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS

1717

Second, we compute the nonstandard form of the shift operator. This operator is important in practical applications of wavelets because the wavelet coeﬃcients are not shift invariant. Since the nonstandard and standard forms of this operator are sparse and easy to compute, knowing these representations “compensates” for the lack of shift invariance. The wavelet expansion of shifts of vectors or of matrices may be obtained by applying the shift operator directly to the coeﬃcients of the original expansion. The coeﬃcients for the shift operators may be stored in advance and used as needed. It is clear, however, that the particular manner in which sparseness of the shift operator may be exploited depends on the application and may be less straightforward than is indicated above. We present an example of such an application in numerical analysis. Observing that there are only N log2 N distinct wavelet coeﬃcients in the decomposition of all N circulant shifts of a vector of the length N = 2n , we construct an O(N log N ) algorithm for computing all of these coeﬃcients. Using this algorithm, we show that the storage requirements of the fast algorithm for applying the standard form of a pseudodiﬀerential operator to a vector [2] may be reduced from O(N log N ) to O(log2 N ) signiﬁcant entries. 2. Compactly supported wavelets. In this section, we brieﬂy review the orthonormal bases of compactly supported wavelets and set our notation. For the details we refer to [1]. The orthonormal basis of compactly supported wavelets of L2 (R) is formed by the dilation and translation of a single function ψ(x), (2.1) ψj,k (x) = 2−j/2 ψ(2−j x − k),

where j, k ∈ Z. The function ψ(x) has a companion, the scaling function ϕ(x), and these functions satisfy the following relations: √ L−1 2 hk ϕ(2x − k),

k=0

(2.2)

ϕ(x) =

(2.3) where (2.4) and

√ L−1 ψ(x) = 2 gk ϕ(2x − k),

k=0

gk = (−1)k hL−k−1 ,

k = 0, · · · , L − 1,

+∞

(2.5)

−∞

ϕ(x)dx = 1.

**In addition, the function ψ has M vanishing moments
**

+∞

(2.6)

−∞

ψ(x)xm dx = 0,

m = 0, · · · , M − 1.

k (x)}k∈Z form an orthonormal basis of Wj and the functions {ϕj. i.k (x) = 2−j/2 ϕ(2−j x − k)}k∈Z form an orthonormal basis of Vj . the wavelets {ψj.3) is related to the number of vanishing moments M . BEYLKIN The number L of coeﬃcients in (2. If additional conditions are imposed (see [2] for an example). the decomposition of the Hilbert space L2 (R) into a chain of closed subspaces (2. (2.14) 1 |Q(eiξ )|2 = P (sin2 1 ξ) + sin2M ( 2 ξ) R( 1 cos ξ).. [5]. and where Q is a polynomial such that (2.1718 G. The coeﬃcients H = {hk }k=L−1 and G = {gk }k=L−1 in (2.12) is of the form (2. Any trigonometric polynomial solution m0 (ξ) of (2. On each ﬁxed scale j. it completely determines the functions ψ and ϕ.10) L2 (R) = j∈Z Wj . where M ≥ 1 is the number of vanishing moments. Lemma 1 (Daubechies [1]).7) such that (2. L = 2M . The following lemma characterizes trigonometric polynomial solutions of (2. then the relation might be diﬀerent. The function m0 (ξ) satisﬁes the k=0 equation (2.2) and (2. Vj = L2 (R). Once the ﬁlter H has been chosen. the space L2 (R) is represented as a direct sum (2.11) m0 (ξ) = 2−1/2 k=0 hk eikξ where {hk }k=L−1 are the coeﬃcients of the ﬁlter H. The wavelet basis induces a multiresolution analysis on L2 (R) [4]. 2 2 .12) |m0 (ξ)|2 + |m0 (ξ + π)|2 = 1. Let us deﬁne the 2π-periodic function k=L−1 (2.e.13) m0 (ξ) = 1 2 (1 + eiξ ) M Q(eiξ ).3) are k=0 k=0 quadrature mirror ﬁlters. and for the wavelets in [1]. but L is always even. j∈Z By deﬁning Wj as an orthogonal complement of Vj in Vj−1 .8) j∈Z · · · ⊂ V2 ⊂ V1 ⊂ V0 ⊂ V−1 ⊂ V−2 ⊂ · · · Vj = {0}.2) and (2.12) which correspond to the orthonormal bases of compactly supported wavelets with vanishing moments.9) Vj−1 = Vj ⊕ Wj .

Γj . ϕ(2−j x − i) ψ (2−j x − l) 2−j dx = 2−j γi−l .6) j βil = 2−j (3.15) P (y) = k=0 M −1+k k yk . The nonstandard form [2] is a representation of an operator T as a chain of triplets (3. ψ(2−j x − i) ϕ (2−j x − l) 2−j dx = 2−j βi−l . i. Bj . Bj .1) T = {Aj .4) The operators {Aj . dx . j j j The matrix elements αj .2) (3. Bj . and R is an odd polynomial such that (2. In this section we construct the nonstandard form of the operator d/dx. where Pj is the projection operator on the subspace Vj and Qj = Pj−1 − Pj is the projection operator on the subspace Wj .OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1719 where k=M −1 (2. and ril of Tj = Pj T Pj . Γj : W j → V j . 0 ≤ P (y) + y M R( 1 − y) 2 for 0 ≤ y ≤ 1. +∞ (3. (3.16) and (2. The operator d/dx in wavelet bases.8) where j γil = 2−j j ril = 2−j ∞ −∞ ϕ(2−j x − i) ϕ (2−j x − l) 2−j dx = 2−j ri−l .9) αl = −∞ ψ(x − l) d ψ(x) dx. Γj }j∈Z A j : Wj → W j . (3. βil . B j : Vj → W j . l. Bj = Qj T Pj . γil of Aj .5) αj = 2−j il ∞ −∞ ∞ −∞ ∞ −∞ ψ(2−j x − i) ψ (2−j x − l) 2−j dx = 2−j αi−l . and Γj = Pj T Qj . Γj }j∈Z are deﬁned as Aj = Qj T Qj . j ∈ Z.3) (3.7) and (3. il for the operator d/dx are easily computed as (3. acting on the subspaces Vj and Wj . 3.17) sup 0≤y≤1 1 P (y) + y M R( 2 − y) < 22(M −1) .

17) rl = −∞ (−iξ) |ϕ(ξ)|2 e−ilξ dξ.18) 1 1 |m0 (ξ)| = + 2 2 2 L/2 k=1 a2k−1 cos(2k − 1)ξ.1720 +∞ G. BEYLKIN (3. · · · .12) or. k=0 L−1−n (3. L − 1. therefore. ˆ In order to compute the coeﬃcients rl we ﬁrst note that any trigonometric polynomial m0 (ξ) satisfying (2. using (2. dx Moreover. dx (3. L−1 L−1 (3.20) a2k = 0.15) γi = 2 k=0 k =0 hk gk r2i+k−k . . and. (3.14) and βi = 2 k=0 k =0 gk hk r2i+k−k . dx d ψ(x) dx.2) and (2. the representation of d/dx is completely determined by rl in (3.12) is such that (3.3) we have L−1 L−1 (3. where an are the autocorrelation coeﬃcients of H = {hk }k=L−1 .16) we obtain +∞ 1 ϕ(ξ) = √ ˆ 2π +∞ −∞ ϕ(x) eixξ dx.12) rl = −∞ ϕ(x − l) d ϕ(x) dx.19) an = 2 i=0 hi hi+n . by the representation of d/dx on the subspace V0 . where ˆ (3. n = 1.13) αi = 2 k=0 k =0 gk gk r2i+k−k .12) in terms of ϕ(ξ). Rewriting (3. L/2 − 1.11) and γl = −∞ ϕ(x − l) +∞ (3. in other words. · · · . The autocorrelation coeﬃcients an with even indices are zero.10) βl = −∞ +∞ ψ(x − l) d ϕ(x) dx. L−1 L−1 (3. k = 1. (3.

ν2 + 1 h2 = 2−1/2 ν+1 . we obtain L/2−1 (3.26) rl = −r−l . 2(3ν 4 + 1) 1 + 3ν 2 .24) and (3.2 of [1].20) and (3.22) |m0 (ξ + π)|2 = 1 1 − 2 2 L/2 k=1 a2k−1 cos(2k − 1)ξ + 1 2 L/2−1 a2k cos 2kξ. If M = 1.25) 1 rl = 2 r2l + 2 k=1 a2k−1 (r2l−2k+1 + r2l+2k−1 ) .21) 1 1 |m0 (ξ)| = + 2 2 2 L−1 an cos nξ. (ν 2 + 1)2 a3 = ν 2 (ν 2 − 1) .19). (1) If the integrals in (3. (2) If M ≥ 2.22) to satisfy (2. and hence. rl = 0 for −L + 2 ≤ l ≤ L − 2 and (3. Computing |m0 (ξ + π)|2 .24) and (3.) We prove the following: Proposition 1.24) and (3. Let us consider Example 3.18).12). then equations (3.17) may not be absolutely convergent.25) have a unique solution but the integrals in (3. then the coeﬃcients r l in (3. Remark 1. k=1 Combining (3. l rl = −1. ν2 + 1 where ν is an arbitrary real number.23) k=1 a2k cos 2kξ = 0.OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1721 To prove this assertion we compute |m0 (ξ)|2 using (2.12) satisfy the following system of linear algebraic equations: L/2 (3. where L = 4 and h0 = 2−1/2 ν(ν − 1) .19).25) have a unique solution with a ﬁnite number of nonzero rl . ν2 + 1 h1 = 2−1/2 1−ν . n=1 where an are given in (3.17) exist. we have (3. (ν 2 + 1)2 . We have a1 = and r1 = − (1 + ν 2 )2 . ν2 + 1 h3 = 2−1/2 ν(ν + 1) . namely.11) and obtain (3. 2(3ν 4 + 1) r2 = ν 2 (1 − ν 2 ) . (See also Remark 6 about vanishing moments of a 2k−1 . then equations (3.12) or (3. l where the coeﬃcients a2k−1 are given in (3.21) and (3.12) or (3. (3.

The general statement follows simply on taking Fourier transforms and using Leibniz’s rule.20). In order to obtain (3.31) l=−∞ lm ϕ(x − l) = xm + (−1)l l=1 m l Mlϕ xm−l . we obtain (3.30) rl = 2r2l + n=1 an (r2l−n + r2l+n ).1722 G. l ∈ Z.25) we use the following relation: l=+∞ l=m (3. then (3. where (3. In this case the function ϕ is not continuous and 1 1 sin 2 ξ i 1 ξ e2 .17) is not absolutely convergent. Using (2. We note that (3. L−1 2 k=0 hk Changing the order of summation in (3. a1 = 1 and r1 = − 2 .24) from (3.19).29) and using the fact that arrive at L−1 = 1. −∞ where l = 1.30). If M ≥ 2.12) we (3. ϕ(ξ) = √ ˆ 2π 1 ξ 2 so that the integral in (3.28) ri = 2 k=0 l=0 hk hl r2i+k−l . Proof of Proposition 1. BEYLKIN The parameter ν can be chosen so that the Fourier transform ϕ(ξ) does not have the ˆ suﬃcient decay to insure the absolute convergence of the integral (3. and thus.28) as L−1 k−L+1 (3.27) and hence.12) and (3. Using (3.33) |ϕ(ξ)|2 |ξ| ≤ C(1 + |ξ|)−1− .17).32) are zero. 1 For the Haar basis (h1 = h2 = 2−1/2 ). we rewrite (3.32) Mlϕ = +∞ ϕ(x) xl dx. ˆ . Substituting l = k − m. − 2 ). we (3.25). m.2) for both ϕ(x − l) and obtain L−1 L−1 +∞ d dx ϕ(x) in (3.31) with m = 1. · · · .29) ri = 2 k=0 m=k hk hk−m r2i+m . are the moments of the function ϕ(x). we obtain 1 1 the simplest ﬁnite diﬀerence operator ( 2 . where an are given in (3. ri = 2 k=0 l=0 hk hl −∞ ϕ(2x − 2i − k) ϕ (2x − l) 2 dx L−1 L−1 (3. Using (3.31) is well known if all moments (3. 0. we obtain (3.

43) r (ξ) = 2( |m0 (ξ/2)|2 r (ξ/2) + |m0 (ξ/2) + π)|2 r (ξ/2) + π)).25) follows from the existence of the integral in (3. ˆ ˆ ˆ and using (3.24) by eilξ and summing over l. The speciﬁc interval −L+2 ≤ l ≤ L−2 is obtained by the direct examination of (3. the integral in (3.12) we arrive at (3. ξ∈R |ϕ(ξ)| ≤ C(1 + |ξ|)−M +log2 B . Since the scaling function ϕ has a compact support there are only a ﬁnite number of nonzero coeﬃcients rl . Multiplying (3.40) and (3. and hence.37) r (ξ) = ˆ l rl eilξ . (3. rodd (ξ/2) = ˆ l r2l+1 ei(2l+1) ξ/2 .35) l rl = 0.24) and (3.OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1723 where > 0.2 of [1].18).24).38) and (3. Let us show now that (3. ˆ Due to condition (2.42) r(ξ) = r (ξ/2) + r (ξ/2 + π) + (2|m0 (ξ/2)|2 − 1)(ˆ(ξ/2) − r (ξ/2 + π)) .17). The existence of the solution of the system of equations (3. This assertion follows from Lemma 3. (3. we obtain from (3.17).34) where B = sup |Q(eiξ )|. where it is shown that (3. we have log2 B = M − 1 − with some > 0.39) Noticing that (3. 2 reven (ξ/2) = r (ξ/2) + r (ξ/2 + π) ˆ ˆ ˆ 2 rodd (ξ/2) = r(ξ/2) − r(ξ/2 + π).36) (3.41) reven (ξ/2) = ˆ l r2l eilξ . ˆ ˆ ˆ . using (2.36) r(ξ) = 2 reven (ξ/2) + rodd (ξ/2) ˆ ˆ 2 k=1 where (3. ˆ ˆ ˆ r ˆ Finally.17) is absolutely convergent. we obtain L/2 1 ˆ a2k−1 (e−i(2k−1)ξ/2 + ei(2k−1)ξ/2 ).

24) and (3. Using relation (4.45) Rewriting (3. fj.24) and (3. x − 2j l) and |˜ − x| ≤ 2j l.k (x) dx. Examples.k−l = (3.13) and (3.k (x) dx ϕj. 2 i=0 hi gi+n and i=0 L−1−n 2 i=0 gi gi+n . (3. we obtain +∞ (Tj f )(x) = k∈Z −∞ f (x) ϕj. M .46) fj.43).47) in (3. where M is the number of vanishing moments and L = 2M .35). .1724 G. Given the solution rl of (3.25) is unique.25) we consider the operator Tj deﬁned by these coeﬃcients on the subspace Vj and apply it to a j suﬃciently smooth function f . the solution to (3. we obtain r (0) = 2ˆ(0) and thus. · · · .k (x).48) 1 2 rl l 2 l +∞ f (˜) ϕj.k−l = 2−j/2 +∞ −∞ (Tj f )(x) = k∈Z 2−j l rl fj.8). (3.k (x) dx −∞ +22j f (˜) ϕj.22) of [1]. x −∞ It is clear that as j → −∞.14) for αl and βl (γl = −β−l ) may be simpliﬁed by changing the order of summation in (3.k (x) +2j k∈Z (3.k (x) dx − 2j l l2 2 +∞ +∞ f (x) ϕj.44) and using ˜ ˜ x (3.k (x).26) follows now from (3. and expanding f (x − 2j l) in the Taylor series at the point x. x −∞ where x = x(x.25).44) where (3.49) |m0 (ξ)|2 = 1 − (2M − 1)! [(M − 1)!]2 22M −1 ξ 0 sin2M −1 ξ dξ.13) and (3.k−l = 2−j/2 +∞ −∞ f (x) ϕ(2−j x − k + l) dx.k−l ϕj. Remark 2. let us compute the coeﬃcients a2k−1 . operators Tj and d/dx coincide on smooth functions. BEYLKIN Setting ξ = 0 in (3.14) and L−1−n by introducing the correlation coeﬃcients 2 L−1−n gi hi+n .45) (3. For examples we will use Daubechies’ wavelets constructed in [1]. k = 1.17). f (x − 2j l) ϕ(2−j x − k) dx.47) −∞ f (x) ϕj.k (x) dx ϕj. The relation (3. Substituting (3. ˆ r Uniqueness of the solution of (3. Using standard arguments it is easy to prove that T−∞ = d/dx and hence.24) and (3.35) and (3. The expression for αl is especially simple: αl = 4r2l − rl .25) follows from the uniqueness of the representation of d/dx. Since rl = 2−j rl (3. First. we have +∞ fj. we have (3. We note that expressions (3.

Solving the equations of Proposition 1. 2. 1095 r4 = − 1 . 128 a5 = 3 . M. we present the results for Daubechies’ wavelets with M = 2. Thus.OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1725 we ﬁnd. 1. 2920 75 .50) and (3. 3 r2 = 1 . 396424 128 . 49553 1 .50) where (3. 5. 2/3. (M − m)! (M + m − 1)! (2m − 1) where m = 1. 4. 743295 r3 = − 1664 . We note that the coeﬃcients rl are the same for all Daubechies’ wavelets with a ﬁxed number of vanishing moments M .18). 8 The coeﬃcients (−1/12. by computing (3.52) a2m−1 = (−1)m−1 CM . M = 2. 1024 a5 = 49 .51) |m0 (ξ)|2 = ξ 0 sin2M −1 ξ dξ. 1189272 1225 . 365 r3 = − 16 . · · · . 3. M = 3. that (−1)m−1 cos(2m − 1)ξ . −2/3.52) are rational by construction). 49553 r2 = r5 = 76113 . 128 2645 . a1 = and r1 = − 3. 8 1 a3 = − . 0. 1024 a3 = − 245 . The coeﬃcients re are rational numbers since they are solutions of a linear system with rational coeﬃcients (a2m−1 in (3. notwithstanding the fact that there are several diﬀerent wavelet bases for a given M (depending on the choice of the roots of polynomials in the construction described in [1]). (M − m)! (M + m − 1)! (2m − 1) m=1 2 M 1 1 + CM 2 2 CM = (2M − 1)! (M − 1)! 4M −1 . 1189272 r6 = − . by comparing (3. M = 4. 12 9 . 1024 a7 = − 5 . a1 = and 2 r1 = − . a1 = and r1 r4 = − = 39296 . 6. 365 r2 = 53 . 64 a3 = − 25 . 1024 272 . 1/12) of this example coincide with one of the standard choices of coeﬃcients for numerical diﬀerentiation. we have (3.

a1 = 16384 and r1 r4 r7 = − = = − 957310976 . 1159104017 1386496 .2) exist (see also Remark 3). The operator dn /dxn in wavelet bases.26) are satisﬁed due to the choice of initialization. 131072 847 . 8192 a7 = − 405 . 9379505240560290 4689752620280145 r6 = 483632604097 . 1159104017 17297069 . it is easy to verify that (3.43). Again. r9 = .25) and (3. M = 6. 32768 a9 = 35 . 262144 63 .1) or. and iterate using (3. 32768 265226398 . 8. (4. We also use an iterative algorithm as a way of solving equations (3. It only displays the coeﬃcients {rl }L−2 since r−l = −rl and r0 = 0.5 and r1 = −0.1) or (4. 2501201397482744 4. 8113728119 160083 . . the nonstandard form of the operator dn /dxn is completely determined by its representation on the subspace V0 . 18545664272 38115 . = −∞ (−iξ)n |ϕ(ξ)|2 e−ilξ dξ ˆ if the integrals in (4. 9. 10431936153 r5 = − r8 = − r6 = − 5. by the coeﬃcients (4. i. 5795520085 5 . alternatively. 6. 262144 r2 = G. 8192 a5 = 567 .24) and (3.. 112554062886723480 = a3 = − a9 = a5 = 22869 . 2318208034 2048 .2) rl (n) +∞ Iterative algorithm for computing the coeﬃcients rl .1726 4.24) to recompute rl . BEYLKIN a3 = − 2205 . Table 1 was computed using this algorithm for Daubechies’ wavelets with M = 5. M = 5. r8 = . as in the case with the operator d/dx. We start with r −1 = 0. r3 = . r5 = .5. 7. 262144 r3 = − 735232 . dxn l ∈ Z.25). 4689752620280145 18759010481120580 14069257860840435 136429697045009 7449960660992 .e. 19845 . 13780629 563818 . 131072 5445 . r2 = . 6565653668392203 75036041924482320 937950524056029 1743 . a1 a7 and r1 = r4 = 3986930636128256 4850197389074509 1019185340268544 . l=1 rl (n) +∞ = −∞ ϕ(x − l) dn ϕ(x) dx. Using (3. 262144 = − a11 = − r7 = r10 = 78962327552 31567002859 2719744 .

2363946800701E-06 -1.8814546576295E-03 3.1830548203747E-10 -1.2522069166503E-13 -1.7814005916812E-14 -1.10636406828947 3.6677069083796E-05 -2.30325935147672 -0.22882018706694 -5.5966864798639E-08 -8.2400785360984E-14 1.8996668057051E-09 6.6960479423517E-10 M =8 Coeﬃcients rl -0.6187880013009E-18 -4.1871130331900E-07 4.2079482385949E-08 9.2440589997659E-02 1.OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1727 Table 1 The coeﬃcients {rl }l=L−2 for Daubechies’ wavelets.12095364936000 3.88344604609097 0.3923582002393E-04 -5.8782468879634E-07 1.3352571932672E-02 7. where L = 2M and M = 5.4047301644748E-05 -2.4025197533630E-13 -4.32031206224855 -0.9686511520083E-13 M =9 11 12 13 14 1 2 3 4 5 6 -0.25855294414146 -7.28296509452594 -9.9938104563894E-08 7.3734404776409E-04 4.85013666155592 0.4613963657755E-03 -2.2026575195723E-05 4.6283900594600E-16 7 8 9 10 11 12 13 14 15 16 .86874391452377 0.6971849256415E-10 7.9583791164537E-03 1.9620437763642E-05 -4.0361220591478E-07 -1.89531640583699 0.0616930669821E-02 2.8507474310747E-24 M =7 1 2 3 4 5 6 7 8 9 10 11 12 -0.0189066217795E-02 2.2687411014648E-02 -3.1290147839488E-02 -6.4545511041994E-02 -1.4519921109537E-07 -3.2069120451167E-07 -2.0315302133757E-03 -7. l=1 Coeﬃcients rl -0.2035273999989E-11 -6.1034028106558E-03 -2.2968915709948E-06 1. 9.9952721886694E-02 -1.5885615434757E-03 4.82590601185015 0.1374108294110E-10 -5.6501679210868E-06 -2.7812077649932E-04 1.5241171135682E-07 -2.5854647516841E-19 l M =5 1 2 3 4 5 6 7 8 l 1 2 3 4 5 6 7 8 9 10 M =6 1 2 3 4 5 6 7 8 9 10 -0. · · · .

then the equations (4. The representation of the ﬁrst derivative in this basis is described in the previous section. If the integrals in (4. such that for even n (4. (n) rl = 0 for −L + 2 ≤ l ≤ L − 2.1) or (4. 8 .1) or (4. the system of equations (4. n = 1. (2) Let M ≥ (n + 1)/2.6) l n l2˜ rl (n) rl (n) = r−l . Equations (4. (4.8) (4.4) have (n) a unique solution with a ﬁnite number of nonzero coeﬃcients rl . then the coeﬃcients rl .2) are not absolutely convergent. where a2k−1 are given in (3. We have a1 = and 1 r−2 = − . (n − 1)/2.2) exist. The proof of Proposition 2 is completely analogous to that of Proposition 1. r2 = 1 . Remark 3. (n) (1) If the integrals in (4.7) l rl and for odd n (n) = 0.2) exist.19). However. n = 1. A case in point is the Daubechies’ wavelet with M = 2. 8 1 a3 = − . l n rl l (n) = (−1)n n!. and (4.4) do not have a solution for the second derivative n = 2. · · · .3) and (4. r1 = −1.5) (4.1728 G.3) and (4. 2 9 .1) and (4. ˜ (n) = 0. BEYLKIN Proposition 2. r0 = 0.3) and (4. · · · . namely. The linear system in Proposition 2 may have a unique solution whereas integrals (4.9) l n l2˜ −1 rl (n) rl (n) = −r−l . where M is the number of vanishing moments in (2.4) (n) rl 1 = 2n r2l + 2 + (n) r2l+2k−1 ) . ˜ (n) = 0. n/2 − 1.6).3) and (4. 2 r−1 = 1. l ∈ Z satisfy the following system of linear algebraic equations: L/2 (n) a2k−1 (r2l−2k+1 k=1 (4.4) has a solution for the third derivative n = 3.

(4. we arrive at (4. (n) into the right-hand side of (4. ˆ ˆ ˆ By considering the operator M0 deﬁned on 2π-periodic functions. we include the situation where the operator may be preconditioned only on a subspace.16) with eigenvalue 1 is considered.OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1729 The set of coeﬃcients (−1/2.2).13) ϕ(ξ) = m0 (ξ/2)ϕ(ξ/2) ˆ ˆ r (ξ) = ˆ l rl (n) 2π = 0 k∈Z |ϕ(ξ + 2πk)|2 (−i)n (ξ + 2πk)n e−ilξ dξ.16) and vice versa. then by the condition number we understand the ratio of the largest singular value to the smallest singular value above the threshold of accuracy. where the problem (4. Thus. While theoretically it is well understood that the derivative operators (or. Let us derive an equation generalizing to (3.14) r (ξ) = 2n ( |m0 (ξ/2)|2 r(ξ/2) + |m0 (ξ/2) + π)|2 r(ξ/2) + π)). operators with homogeneous symbols) have an explicit diagonal preconditioner in wavelet bases. We note that among the wavelets with L = 4.11). r is an eigenvector of the operator M0 corresponding to the eigenvalue 2−n and.11) where (4. Remark 4.14) as (4.) Remark 5. since it represents one of the advantages of computing in the wavelet bases. 1. and summing separately over even and odd indices in (4. ˆ therefore. ˆ rl eilξ .12) Substituting the relation (4. the wavelets with two vanishing moments M = 2 do not have the best H¨lder exponent (see [6]).15) (M0 f )(ξ) = |m0 (ξ/2)|2 f (ξ/2) + |m0 (ξ/2) + π)|2 f (ξ/2) + π). 0. more generally. we rewrite (4.2) as (4. We rewrite (4. (The operator M0 is also introduced in [7] and [8]. We note that the preconditioning described here addresses the problem of ill conditioning due . (4. ﬁnding the representation of the derivatives in the wavelet basis is equivalent to ﬁnding trigonometric polynomial solutions of (4.43) for dn /dxn directly from (4. If an operator has a null space (the actual null space or a null space for a given accuracy). −1. the numerical evidence illustrating this fact is of interest. but the representation o of the third derivative exists only if the number of vanishing moments M = 2.10) Therefore. ˆ ˆ Thus.11). 1/2) is one of the standard choices of ﬁnite diﬀerence coeﬃcients for the third derivative.16) M0 r = 2−n r. ˆ r (ξ) = ˆ k∈Z |ϕ(ξ + 2πk)|2 (−i)n (ξ + 2πk)n .

93089E+05 0. we have the following formula (3. the number of iterations √ of the conjugate gradient method is O( κp ).g. the computation of the nonstandard form is considerably simpler than in the general case [2].23272E+05 0. we arrive at an eﬀective method for computing representations of convolution operators. p D2 = P D 2 P where Pil = δil 2j . (j−1) Let us assume that the matrix ti−l (i. e. In the following examples the standard form of the periodized second derivative D2 of size N × N .37236E+06 κp 0. (j−1) .11511E+02 0. To compute the matrix representation of Pj T Pj . N 64 128 256 512 1024 κ 0.12604E+02 0. this remark implies a completely new outlook on a number of numerical methods.13045E+02 5.12091E+02 0. Table 2 Condition numbers of the matrix of periodized second derivative (with and without preconditioning) in the basis of Daubechies’ wavelets with three vanishing moments M = 3. Convolution operators. and where j is chosen depending on i. We recall that the condition number controls the rate of convergence of a number of iterative algorithms..10792E+02 0. Appendix B of [2]) are of the simplest form due to the fact that the moments of the autocorrelation function of the scaling function ϕ vanish. l ∈ Z) represents the operator Pj−1 T Pj−1 on the subspace Vj−1 . We will demonstrate that the quadrature formulas for representing kernels of convolution operators on V0 (see. by combining the asymptotics of wavelet coeﬃcients with the system of linear algebraic equations (similar to those in previous sections). l ≤ N − N/2j . the condition number κp of the operator is uniformly bounded with respect to the size of the matrix. l so that N − N/2j−1 + 1 ≤ i.26) of [2]: L−1 L−1 (5.14545E+04 0. Thus. where N = 2n . For convolution operators. We present here two tables illustrating such preconditioning applied to the standard form of the second derivative (see [2] on how to compute the standard form from the nonstandard form). For periodized derivative operators the bound on the condition number depends only on the particular choice of the wavelet basis. and PN N = 2n . a topic we will address elsewhere.1) tl (j) = k=0 m=0 hk hm t2l+k−m . This method is especially simple if the symbol of the operator is homogeneous of some degree. for example. p Tables 2 and 3 compare the original condition number κ of D2 and κp of D2 . Moreover. BEYLKIN only to the unfavorable homogeneity of the symbol and does not aﬀect ill conditioning due to other causes. is preconditioned by the diagonal matrix P .1730 G.58181E+04 0. After applying such a preconditioner. 1 ≤ j ≤ n.

where Φ is the autocorrelation function of the scaling function ϕ. and by changing the order of integration.26807E+06 κp 0.4) tl (j) +∞ = 2j −∞ K(2j (l − y)) Φ(y) dy.43633E+01 0.41886E+04 0. we obtain (5.0 (y) ϕj.43595E+01 0.16754E+05 0.OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1731 Table 3 Condition numbers of the matrix of periodized second derivative (with and without preconditioning) in the basis of Daubechies’ wavelets with six vanishing moments M = 6. N 64 128 256 512 1024 κ 0.5) Let us verify that Φ(y) = −∞ ϕ(x) ϕ(x − y) dx. We also have (5.7) Clearly.43542E+01 0.6) −∞ Φ(y)dy = 1 and (5.43640E+01 which easily reduces to (5.19).l (x) dxdy. we have (5.2) tl (j) = t2l (j−1) + 1 2 L/2 a2k−1 (t2l−2k+1 + t2l+2k−1 ).10472E+04 0. ξ=0 .43620E+01 0.3) tl (j) +∞ +∞ −∞ = −∞ K(x − y) ϕj.67018E+05 0. k=0 (j−1) (j−1) where the coeﬃcients a2k−1 are given in (3. +∞ (5.8) Mm = Φ 1 ∂ξ i m Mm = Φ +∞ −∞ y m Φ(y)dy = 0 for 1 ≤ m ≤ 2M − 1. +∞ (5. |ϕ(ξ)|2 ˆ .

The Hilbert transform. Remark 6. l ∈ Z.8)–(3.13) rl = −∞ ϕ(x − l) (Hϕ)(x) dx. But formula (5. We consider two examples of such operators. and Γ0 are computed from the coeﬃcients rl . Bj = B0 . .7) holds ˆ ˆ provided that (5. BEYLKIN Using (5. Aj = A0 .1732 G. which. This method is especially simple if the symbol of the operator is homogeneous of some degree since in this case the operator is completely deﬁned by its representation on V0 .10) follows from the explicit representation in (2.12)].13).12)) (5. denotes a principal value at s = x. |m0 (ξ + π)|2 ξ=0 = 0 for 0 ≤ m ≤ 2M − 1. where matrix elements αi−l . (3. Here we introduce a diﬀerent approach for computing representations of convolution operators in the wavelet basis which consists of solving the system of linear algebraic equations (5. eqns.v. namely. L−1 L−1 (5. The representation of H on V0 is deﬁned by the coeﬃcients ∞ (5.12) g(x) = (Hf )(y) = 1 p. where the shifted moments of the function ϕ vanish. it is clear that (5. We apply our method to the computation of the nonstandard form of the Hilbert transform (5. Equations (5. H = {Aj . s−x where p. the Hilbert transform and the operator of fractional diﬀerentiation (or antidiﬀerentiation). Γj }j∈Z .21) also imply that even moments of the coeﬃcients a2k−1 from (3.15) βi = k=0 k =0 gk hk r2i+k−k .8) and the identity ϕ(ξ) = ϕ(ξ/2)m0 (ξ/2) (see [1]).2) subject to asymptotic conditions. k=L/2 (5.14) αi = k=0 k =0 L−1 L−1 gk gk r2i+k−k . Since the moments of the function Φ vanish equation (5. B0 . π ∞ −∞ f (s) ds.10) 1 ∂ξ i m 1 ∂ξ i m |m0 (ξ)|2 ξ=0 = 0 for 1 ≤ m ≤ 2M − 1.11) k=1 a2k−1 (2k − 1)2m = 0 for 1 ≤ m ≤ M − 1. βi−l .9) and (3. we refer to this paper for the details.v.4) leads to a one-point quadrature formula for computing the representation of convolution operators on the ﬁnest scale.9) or (due to (2. Bj .19) vanish. Namely. (5. completely deﬁne all other coeﬃcients of the nonstandard form. and Γj = Γ0 . in turn. This formula is obtained in exactly the same manner as for the special choice of the wavelet basis described in [2. and γi−l of A0 .

588303698 -0. Coeﬃcients l M =6 1 2 3 4 5 6 7 8 rl -0.19).031830988 -0.026525823 -0. The coeﬃcients for l > 16 are obtained using asymptotics (5.128743695 -0. We note that the coeﬃcient r0 cannot be determined from equations (5.7).17) and (5. Solving (5.6). Using (5. By rewriting (5.19) rl = −2 0 |ϕ(ξ)|2 sin(lξ) dξ.019894368 .OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1733 and L−1 L−1 (5.17) rl = r2l + 1 2 L/2 a2k−1 (r2l−2k+1 + r2l+2k−1 ).077576414 -0. ˆ ∞ (5. We note that the generalization for computing the coeﬃcients of Riesz transforms in higher dimensions is straightforward.064168018 -0. (We note that r−l = −rl and r0 = 0.035367761 -0. (5. (5.045470650 -0.022736420 -0.024485376 -0. l = 1.18). We compute (see Table 4) the coeﬃcients rl of the Hilbert transform for Daubechies’ wavelets with six vanishing moments with accuracy 10−7 . 16 of the Hilbert transform for Daubechies’ wavelet with six vanishing moments.053041366 -0. and (5.18).16) γi = k=0 k =0 hk gk r2i+k−k . we compute the coeﬃcients rl . ˆ we obtain rl = −r−l and set r0 = 0. l ∈ Z in (5.13) in terms of ϕ(ξ).039788641 l 9 10 11 12 13 14 15 16 Coeﬃcients rl -0.18) rl = − 1 +O πl 1 l2M .) Table 4 The coeﬃcients rl .17) with the asymptotic condition (5. · · · .13) satisfy the following system of linear algebraic equations: (5. The coeﬃcients rl . k=1 where the coeﬃcients a2k−1 are given in (3.021220659 -0. l = 0 with any prescribed accuracy. we obtain the asymptotics of rl for large l.18).028937262 -0. Example.075063628 -0.4).

α The nonstandard form ∂x = {Aj . α The representation of ∂x on V0 is determined by the coeﬃcients +∞ (5. Using (5. and Γ0 are obtained from the coeﬃcients rl . The coeﬃcients for rl . (5. 1 l1+α+2M for l > 0.24) γi = 2 α k=0 k =0 hk gk r2i+k−k . and (5. 2 · · ·. Γ(−α) where we consider α = 1. (5. l > 14. Γj }j∈Z is computed via Aj = 2−αj A0 .7).25) 2 k=1 Example. Bj .26) (5. we obtain the asymptotics of rl for large l.6). We use the following deﬁnition of fractional derivatives: +∞ (5. It is easy to verify that the coeﬃcients rl satisfy the following system of linear algebraic equations: L/2 1 rl = 2α r2l + a2k−1 (r2l−2k+1 + r2l+2k−1 ) .23) and βi = 2 α k=0 k =0 gk hk r2i+k−k .21) rl = −∞ α ϕ(x − l) (∂x ϕ) (x) dx. (5.27) rl rl = 1 1 +O Γ(−α) l1+α = 0 for l < 0. or l < −7 are obtained using asymptotics (5. L−1 L−1 (5. then (5.22) αi = 2 α k=0 k =0 gk gk r2i+k−k . and γi−l of A0 . and Γj = 2 Γ0 . If α < 0. BEYLKIN Fractional derivatives. βi−l .29) rl rl 1 1 = − √ 1+ 1 + O 2 πl 2 = 0 for l < 0. We compute (see Table 5) the coeﬃcients rl of the fractional derivative with α = 0.1734 G. provided that this integral exists. l ∈ Z. L−1 L−1 (5.20) deﬁnes fractional antiderivatives. 1 1 l13+ 2 for l > 0.19).4).28) (5.5 for Daubechies’ wavelets with six vanishing moments with accuracy 10−7 . .20) α (∂x f ) (x) = −∞ (x − y)−α−1 + f (y)dy. where the coeﬃcients a2k−1 are given in (3. L−1 L−1 (5. −αj −αj Bj = 2 B0 . B0 . where matrix elements αi−l .

OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1735 Table 5 The coeﬃcients {rl }l . therefore. in image processing the shift .52272841E-02 -1.61324170E-02 -1.84681966 -0.69847577 2. Also. j = 1. if the absolute value of the shift is greater than L − 2. As j (j) increases.2) tl (0) = δl.49883759E-02 0. As an example. 2. in many applications we can eﬀectively compensate the absence of the shift invariance of the wavelet transforms.04479500E-02 -8.97463780E-02 l 4 5 6 7 8 9 10 11 12 13 14 Coeﬃcients rl -2. However.68623867E-06 4.78614593E-03 -6. if the representation of a vector in the wavelet basis is sparse.19) we have (6. (0) where δ is the Kronecker symbol.1 . We note that the shift by an integer other than one is treated similarly. Coeﬃcients l M =6 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 rl -2.01838599E-03 -5.45847796E-04 -4.24667403E-02 -1. The only nonzero coeﬃcients tl on each scale j are those with indices −L + 2 ≤ l ≤ (j) L − 2. there are nonzero coeﬃcients tl with l outside the interval |l| ≤ L − 2. l = −7.27799963 0.36400139E-02 -8.34633415E-03 2. as we have just demonstrated. · · · . (j) tl (1) 1 = 2 a|2l−1| .5 for Daubechies’ wavelet with six vanishing moments.91718816E-02 -1. all the nonzero coeﬃcients tl will have indices in the interval |l| ≤ L − 1. on the ﬁrst several scales (j) j. Shift operator on V0 and fast wavelet decomposition of all circulant shifts of a vector. By applying these sparse representations directly to the wavelet coeﬃcients. 8.82831017E-06 -1.1) with the an of (3. 14 of the fractional derivative α = 0. The importance of the shift operator stems from the fact that the coeﬃcients of wavelet transforms are not shift invariant. there is a corresponding reduction in the number of operations required to shift such a vector.73225246E-03 -6. for the shift operator in Daubechies’ wavelet basis with three vanishing moments. However.1) ti−j = δi−j. the following Table 6 contains (j) the coeﬃcients tl .77955293E-02 -2. tl → δl. ··· .1 . then. Speciﬁcally.38521459E-03 6.28821728E-02 -8. Let us consider a shift by one on the subspace V0 represented by the matrix (6. Using (5.92061945E-03 -7. For example. the standard) forms of the shift operator are sparse and easy to compute. · · · . the nonstandard (and.0 as j → ∞.

5859375 -9.1333660119562E-03 -2.3732910156250E-04 j=6 -4 -3 -2 -1 0 1 2 3 4 -4.99937113652686 1.1239479930566E-02 0. (j) tl Coeﬃcients l j=5 -4 -3 -2 -1 0 1 2 3 4 tl (j) -8.7055225372314E-05 j=7 -4 -3 -2 -1 0 1 2 3 4 -2.2974760466022E-05 j=2 -4 -3 -2 -1 0 1 2 3 4 0.7198034904338E-03 0. 0. · · · . 1.9237906308573E-03 -1.1482413927743E-04 2.1652627381741E-03 -1.0904073715210E-03 1.4860978555807E-02 -4.99752855458064 2. -1.1073373558501E-04 1.3304820060730E-02 2.8323576983791E-06 .5174703821573E-06 -1.11541545391083 -2.7220458984375E-03 1.9328931709169E-03 5.3411045074463E-05 -1.3712690179423E-03 2.5859375 0.1444091796875E-03 1.26206970214844 -5. Coeﬃcients l j=1 -4 -3 -2 -1 0 1 2 3 4 0.2418627738953E-02 -6.9574082627023E-06 j=3 -4 -3 -2 -1 0 1 2 3 4 -1.5123357772827E-03 6.1756409462604E-04 2.7352138210499E-06 -2. BEYLKIN Table 6 (j) The coeﬃcients {tl }l=L−2 for Daubechies’ wavelet with three vanishing moments. 8.1605296576369E-03 1.6403198242188E-02 -1.9901347160339E-02 0.2046257104714E-02 -2.1498413085938E-02 5.99984123346637 5.1736 G.1081638044863E-03 -5.171875E-02 0.171875E-02 -9.0836550508393E-04 1.87089538574219 0.96389651298523 0.1698923046642E-02 0.765625E-02 1. where L = 6 l=−L+2 and j = 1.0407157939626E-04 4.765625E-02 0.0258483886719E-01 0.4169452359502E-04 5.3516169979703E-06 -4.

that there are only N log2 (N ) distinct coeﬃcients and present here a simple algorithm to compute the wavelet expansion of N circulant shifts of a vector in N log(N ) operations. instead of computing shift operators. We proceed . We recall that the decomposition of a vector of length N = 2n into a wavelet basis requires O(N ) operations. 2. It is clear. let us compute on each k k=1 k k=1 scale j. j = 0.8852840759448E-03 0. we compute all possible shifts.1267131716013E-04 7. (1 ≤ j ≤ n).2778211385012E-05 -7. The coeﬃcients (j) tl for the shift operators can be stored in advance and used as needed. We recall the pyramid scheme {s0 } −→ {s1 } k k {d1 } k −→ {s2 } k {d2 } k −→ {s3 } k ··· (6.4) sj = k n=0 hn sj−1 n+2k−1 . that the method of using sparseness of the shift operator depends on the speciﬁc application and may be less straightforward than is indicated above. 1. · · · .7938552839535E-05 1. The following is an example of an application where. k k In the pyramid scheme (6.3). however.9366035254748E-03 -5. Instead.1071795597672E-03 2. the computation of the wavelet expansion of all N circulant shifts of a vector appears to require O(N 2 ) operations. and sj and dj are periodic sequences with the period 2n−j . We notice.0551069863141E-02 1.5) dj = k n=0 gn sj−1 n+2k−1 .99042607471347 5.9441434890032E-05 j=8 -4 -3 -2 -1 0 1 2 3 4 -1.3) {d3 } · · · k where the coeﬃcients s0 for k = 1.99996009015421 2. Since the coeﬃcients are not shift invariant. however.7380655655486E-04 5. We describe a fast algorithm for the wavelet decomposition of all circulant shifts of a vector and then show how it may be used to reduce storage requirements of one of the algorithms of [2].0419702418149E-02 0. k n=L−1 (6.6059346249153E-04 -2.5265066698194E-03 -4. 2j vectors of diﬀerences and 2j vectors of averages. n. n=L−1 (6.3777042338989E-06 1737 operator allows us to “move” pictures in the “compressed” form. on each scale j we compute one vector of diﬀerences n−j n−j {dj }k=2 and one vector of averages {sj }k=2 . · · · .OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS j=4 -4 -3 -2 -1 0 1 2 3 4 -1.6215105787797E-05 5. N are given.2976609638869E-06 -5.2607019431889E-02 -1.

Therefore. the periodic sequence d1 (0) contains all the coeﬃcients that appear if k s0 is circulantly shifted by 2. at the same time. 2n−j be one of the vectors of averages on the previous k scale j − 1 and compute n=L−1 (6. N . 3. 2n − 1 (see (6. stepping from scale to scale we double the number of vectors of averages and of diﬀerences and. n=L−1 (6. k = 1.6)– k k k k k (6. n+2k n=L−1 (6.8). · · · . 2. n=L−1 (6.8) dj (0) = k n=0 gn sj−1 n+2k−1 . d2 (11)) k k k k where d2 (00). · · · . d2 (01).9). j = 2.12) and (6.7) and sj (1) = k n=0 hn sj−1 . j = 1. d2 (10).6) and (6.13) u2 = (s2 (00).10) and (6.7)). d1 (1)) k k where d1 (0).11) u1 = (s1 (0).6) k and (6. we shift by one the sequence sj−1 in (6. s2 (01) are computed from s1 (0) according to (6. k k k k k On the second scale.9)). 4.9) and d2 (10). s2 (10). the total number of operations in this computation is O(N log N ).6)–(6.9). s1 (1)). We claim that we have k k k k k computed all the coeﬃcients of the wavelet expansion of N circulant shifts of the vector s0 . s2 (00). d2 (11). halve the length of each of them. n+2k To compute the sum in (6. and s1 (1) are computed from s0 according to (6. Thus. BEYLKIN as follows: let sj−1 . s2 (11)).9) dj (1) k = n=0 gn sj−1 . k Indeed. d1 (1). we set (6. etc.1738 G. · · · . Let us organize the vectors of diﬀerences and averages as follows: on the ﬁrst scale. and the periodic sequence d1 (1) k k contains all the coeﬃcients for odd shifts by 1.8)). 2n (see (6. By the same token. s2 (01). · · · .6) sj (0) = k n=0 hn sj−1 n+2k−1 . k k v1 = (d1 (0). d2 (01). and s2 (11) from s1 (1). k k k k v2 = (d2 (00). k = 1. the periodic sequences s1 (0) and s1 (1) contain all possible coeﬃcients for even k k and odd circulant shifts of s0 (see (6. s2 (10). s1 (0). Repeating this procedure on the k . we set (6.7) and (6.

j) = l=0 l2 l .17) g(x) = −∞ K(x. and shifts is .16) ib (is . the subvector of vj has indices between ib (is . j) + 1 and ib (is .15) and l is = l=0 l2 l . j) points to the ﬁrst element. 0 ≤ is ≤ N −1. y). j) and ib (is . While the vectors v1 .17) as +∞ (6. since the kernel depends on one variable only. · · · . we compute l=j−1 iloc (is . x − z) = K(z) in variables x and z for the convolution operators contains no more than O(log N ) signiﬁcant entries for any ﬁxed accuracy. 1 ≤ j ≤ n. y) f (y) dy by constructing (for any ﬁxed accuracy) its sparse nonstandard or standard form and thereby. v2 . 1. the standard form of K(x. In order to access them. j) = l=n−1 l2 l . These tables give us the direct access to the coeﬃcients in vectors v1 . these coeﬃcients are not organized sequentially. j) + 2n−j . If the operator T is a convolution. while the standard form of [2] will contain O(N ) or O(N log N ) signiﬁcant entries even for a convolution. we compute two tables in (6. The number ib (is . Within this subvector (which is treated as a periodic vector with the period 2n−j ) the number iloc (is . = 0. where ib (is . Let us rewrite (6.14) where (6. The algorithms of [2] are designed to evaluate the Calderon–Zygmund or a pseudodiﬀerential operator T with kernel K(x. +∞ (6. vn contain all the coeﬃcients. then K(x. Namely. N . 2. vn for a constant cost per element. we generate two tables i loc(is . j) in O(N log N ) operations as follows. let us write the binary expansion of is .16).15) and (6. l=j (6. k = 1. . · · · . k l=n−1 (6. We now brieﬂy describe one of the applications of the algorithm for the fast wavelet decomposition of all circulant shifts of a vector in numerical analysis. j) = 0 if j = n. v2 .18) g(x) = −∞ K(x. 1 ≤ j ≤ n. For all scales j. The nonstandard form of a convolution requires at most O(log N ) of storage (see the previous section). · · · . etc. j) points to the begining of the subvector of diﬀerences in vj . reducing the cost of applying it to a function. For a ﬁxed scale j.OPERATORS IN BASES OF COMPACTLY SUPPORTED WAVELETS 1739 next scale for both s1 (0) and s1 (1). 0 ≤ is ≤ N − 1 of the vector s0 . x − z) = K(z) is a function of z only. we again obtain all possible coeﬃcients for odd k k and even shifts which we collect in v2 and u2 . x − z) f (x − z) dz. Alternatively. For each shift is .

Daubechies and J. The apparent diﬃculty in computing via (6. ·) on x is smooth and the number of signiﬁcant entries in the standard form is of O(log2 N ). Yale University. [2] G. 41 (1988). Report YALEU/DCS/RR-696. Universit´ Paris Dauphine. 44 (1991) pp. Multiresolution approximation and wavelets. University of Pennsylvania. Global regularity of solutions.18). preprint. Philadelphia. J. I. Therefore. August 1989. Coifman. Ondelettes et functions splines. which is an important consideration for the multidimensional operators. Local regularity. Feauveau. Ecole Polytechnique. then the dependence of the kernel K(x. Pure Appl. Comm. . pp. 909–996. BEYLKIN If we now construct the standard form of K(x. Daubechies. Lawton. France. Math. R. Two-scale diﬀerence equations. the cost of evaluating (6. of Computer and Information Science. if these operators are represented in the form (6. Beylkin. 57–61. and J. M. GRASP Lab. Math. [6] I. we obtain “super”-compression of the operator. 9007. The advantage of such an algorithm is the reduced storage of O(log 2 N ) signiﬁcant entries for the standard form. e . 32 (1991). inﬁnite products of matrices and fractals. Mallat. I. 1986. Daubechies. [8] W. Dept.. x−z) in variables x and z for pseudodiﬀerential operators (not necessarily convolutions). CEREMADE. New Haven. it appears to require O(N 2 ) operations. Math. Report. [5] S. No. Cohen.18) is that it is necessary to compute the wavelet decomposition of f (x − z) for every x and thus. C. e [4] Paris. pp.. Phys. Lagarias. Meyer. Orthonormal bases of compactly supported wavelets. S´minaire EDP. Tech. REFERENCES [1] I. [7] A. Indeed.1740 G. [3] Y. preprint. Biorthogonal bases of compactly supported wavelets. Fast wavelet transform and numerical algorithms I. and V. les ondelettes et ﬁltres miroirs en quadrature. Necessary and suﬃcient conditions for constructing orthonormal wavelet bases.. Le calcul scientiﬁque. Comm. We note that the extension of the algorithm for the fast wavelet decomposition of all circulant shifts of a vector to the multidimensional case is straightforward. The algorithm of this section accomplishes this task in O(N log N ) operations. Rokhlin. Tech. Tech. Two-scale diﬀerence equations. Report. II. 141–183. preprint. PA.. Pure Appl.18) does not exceed O(N log N ).

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd