Orthogonal p-wavelets on R+

Yu.A. Farkov

Abstract
In this paper we give a general construction of compactly supported
orthogonal p-wavelets in L2 (R+ ) arising from scaling filters with pn
many terms. For all integer p ≥ 2 these wavelets are identified with
certain lacunary Walsh series on R+ . The case where p = 2 was
studied by W.C. Lang mainly from the point of view of the wavelet
analysis on the Cantor dyadic group (the dyadic or 2-series local field).
Our approach is connected with the Walsh – Fourier transform and
the elements of M -band wavelet theory.
Keywords: compactly supported p-wavelets, M -band wavelet
transform, Walsh functions, Walsh – Fourier transform, p-series lo-
cal field, signal processing.
AMS Subject Classification: 42A38, 42A55, 42C15, 42C40,
43A70.

1 Introduction
It is well-known that the scaling function ϕ for a system of p-wavelets with
compact support on the real line R satisfies a refinement equation of the type
N
X
ϕ(x) = ck ϕ(px − k).
k=0

The case where p = 2 have been studied by many authors in great detail (see,
e.g., [1], [2], and references therein). The wavelet theory for integer values of
This work was supported by the Russian Foundation for Fundamental Researches,
grant 02-01-00386.

4
p > 2 is also developed and have some applications in the image processing
context (see [3] for the bibliography and [4] for more recent results). In
this paper we give a general construction of compactly supported orthogonal
p-wavelets on the positive half-line.
Let p be a fixed natural number greater than 1. As usual, let R+ =
[0, +∞) and Z+ = {0, 1, . . .}. Denote by [x] the integer part of x. For x ∈ R+
and any positive integer j we set

xj = [pj x](mod p), x−j = [p1−j x](mod p), (1.1)

where xj , x−j ∈ {0, 1, . . . , p − 1}. It is clear that for each x ∈ R+ there exists
k = k(x) in N such that x−j = 0 for all j > k.
Consider on R+ the addition defined as follows: if z = x ⊕ y, then
X X
z= ζj p−j−1 + ζj p−j
j<0 j>0

with
ζj = xj + yj (mod p) (j ∈ Z \ {0}),
where ζj ∈ {0, 1, . . . , p − 1} and xj , yj are calculated by (1.1). We note that
this binary operation appears in the study of the dyadic Hardy spaces on R+
(see, e.g., [5]). It also is implicit in the book [6] where ⊕ is used for the study
of Walsh series and there applications in the image and date compression.
As usual, we write z = x ª y, if z ⊕ y = x.
Let n be a positive integer. Consider a refinement equation of the type
pn −1
X
ϕ(x) = p aα ϕ(px ª α). (1.2)
α=0

The case where p = 2 was studied by W.C. Lang [7]–[9] mainly from the
point of view of the wavelet analysis on the Cantor dyadic group (the dyadic
or 2-series local field). As noted in [9], the Cantor dyadic group is rather
different in structure than other groups for which wavelet construction have
been carried out.
Denote by 1E the characteristic function of a subset E of R+ . It was
shown in [7] that the function ϕ defined by
∞
X
ϕ(x) = (1/2)1[0,1) (x/2)(1 + a bj w2j+1 −1 (x/2)), x ∈ R+ , (1.3)
j=0

5
where 0 < a ≤ 1, a2 + b2 = 1, and {wm (x)} is the classical Walsh system,
satisfies the equation (1.2) with p = n = 2 and

a0 = (1+a+b)/4, a1 = (1+a−b)/4, a2 = (1−a−b)/4, a3 = (1−a+b)/4.

Moreover, the corresponding wavelet is given by

ψ(x) = 2a0 ϕ(2x ª 1) − 2a1 ϕ(2x) + 2a2 ϕ(2x ª 3) − 2a3 ϕ(2x ª 2).

Our purpose here is to extend these results for all possible p and n. In partic-
ular, a method is given to construct a system of wavelets that can be used to
decompose functions in L2 (R+ ) which is based on decimation by an integer
p > 2. Among motivations can be point out applications of the p-series local
fields to digital signal processing (e.g., [6, Ch.12]) and the similar results in
the wavelet theory on the line R (see, e.g., [10]). Our main results, Theorems
1 and 2, may be strengthened to provide p−wavelet bases for spaces beyond
L2 (R+ ) ( cf. [7, § 4], [9, § 4]).
For integer n ≥ 2, we denote by N0 (p, n) the set of all natural numbers
m ≥ pn−1 for which in the p-ary expansion

k
X
m= µj pj , µj ∈ {0, 1, . . . , p − 1}, µk 6= 0, k = k(m) ∈ Z+ , (1.4)
j=0

there is no n-tuple (µj , µj+1 , . . . , µj+n−1 ) that coincides with some of the
n-tuples

(0, 0, . . . , 0, 1), (0, 0, . . . , 0, 2), . . . , (0, 0, . . . , 0, p − 1).

Put N(p, n) = {1, 2, . . . , pn−1 − 1} ∪ N0 (p, n). For example:

N(2, 2) = {2j+1 − 1 | j ∈ Z+ } = {1, 3, 7, 15, 31, . . .},

N(2, 3) = {1, 2, 3, 5, 6, 7, 10, 11, 13, 14, 15, 21, . . .},

k
X
N(3, 2) = { mj 3j | mj ∈ {1, 2}, k ∈ Z+ } = {1, 2, 4, 5, 7, 8, 13, . . .}.
j=0

For every m ∈ N(p, n), 1 ≤ m ≤ pn − 1, we choose a (real or complex)

number bm in such a way that

bj 6= 0 and |bj |2 + |bpn−1 +j |2 + |b2pn−1 +j |2 + . . . + |b(p−1)pn−1 +j |2 = 1 (1.5)

6
for j = 1, 2, . . . , pn−1 − 1. In the case p = n = 2 we have only one equality:
|b1 |2 + |b3 |2 = 1. Also, it is easy to see that

|b1 |2 + |b5 |2 = |b2 |2 + |b6 |2 = |b3 |2 + |b7 |2 = 1, if p = 2, n = 3

and

|b1 |2 + |b4 |2 + |b7 |2 = |b2 |2 + |b5 |2 + |b8 |2 = 1, if p = 3, n = 2.

The condition (1.5) is necessary for the orthonormality of our system {ϕ(· ª
k) | k ∈ Z+ } (see assertion (b) in Theorem 2 below).
For m ∈ N(p, n), 1 ≤ m ≤ pn − 1, we set

c(i1 , i2 , . . . , in ) = bm , if m = i1 p0 +i2 p1 +. . .+in pn−1 , ij ∈ {0, 1, . . . , p−1}.

Then for m ∈ N(p, n) using p-ary expansion (1.4) we write:

A(m) = c(µ0 , 0, 0, . . . , 0, 0), if k(m) = 0;

A(m) = c(µ1 , 0, 0, . . . , 0, 0)c(µ0 , µ1 , 0, . . . , 0, 0), if k(m) = 1;

...

A(m) = c(µk , 0, 0, . . . , 0, 0)c(µk−1 , µk , 0, . . . , 0, 0) . . .

. . . c(µ0 , µ1 , µ2 , . . . , µn−2 , µn−1 ), if k = k(m) ≥ n − 1.

And for s ∈ {0, 1, . . . , pn − 1} we put


 1, if s = 0,
(n)
ds = bs , if s = j + lpn−1 (1 ≤ j ≤ pn−1 − 1, 0 ≤ l ≤ p − 1),
0, if s = pn − lpn−1 (1 ≤ l ≤ p − 1).


For x ∈ [0, 1), let r0 (x) be given by

½
1, if x ∈ [0, 1/p),
r0 (x) =
εlp , if x ∈ [lp−1 , (l + 1)p−1 ) (l = 1, . . . , p − 1),

7
where εp = exp(2πi/p). The extension of the function r0 to R+ is defined by
the equality r0 (x + 1) = r0 (x), x ∈ R+ . Then the generalized Walsh functions
{wm (x)}(m ∈ Z+ ) are defined by
k
Y
w0 (x) ≡ 1, wm (x) = (r0 (pj x))µj ,
j=0

where
k
X
m= µj pj , µj ∈ {0, 1, . . . , p − 1}, µk 6= 0
j=0

(the classical Walsh system corresponds to the case p = 2).

For x, ω ∈ R+ , let
∞
³ 2πi X ´
χ(x, ω) = exp (xj ω−j + x−j ωj ) , (1.6)
p j=1

where xj , ωj are given by (1.1). Note that χ(x, m/pn−1 ) = χ(x/pn−1 , m) =

wm (x/pn−1 ) for all x ∈ [0, pn−1 ), m ∈ Z+ .
Theorem 1. The function ϕ given by the formula
X
ϕ(x) = (1/pn−1 )1[0,1) (x/pn−1 )(1 + A(m)wm (x/pn−1 )), x ∈ R+ ,
m∈N(p,n)
(1.7)
is a solution of the refinement equation (1.2) provided {aα } satisfy the linear
equations
pn −1
X
aα χ(α, sp−n ) = ds(n) (0 ≤ s ≤ pn − 1). (1.8)
α=0

Moreover, the system {ϕ(· ª k) | k ∈ Z+ } is orthonormal in L2 (R+ ).

It is easily seen that (1.7) coincides with (1.3) when p = n = 2 and
b1 = a, b3 = b.
Example 1. Suppose that b1 = b2 = . . . = bpn−1 −1 = 1. Then, by (1.5),
bm = 0 for m ≥ pn−1 and hence

1, if m ∈ {1, . . . , pn−1 − 1},

½
A(m) =
0, if m ∈ N0 (p, n).

8
Since
pn−1 −1
pn−1 , if 0 ≤ y < 1/pn−1 ,
X ½
χ(y, m) =
0, if 1/pn−1 ≤ y < 1
m=0

(see [6, § 1.5]), we have ϕ = 1[0,pn−1 ) . This function satisfies the equation
(1.2) when a0 = . . . = ap−1 = 1/p and aα = 0 for α ≥ p. Note that Theorem
1 is still true for n = 1 (the Haar case), if we assume N(p, 1) = ∅.
Example 2 (cf. [9, § 5.4]). Suppose ϕ is given by (1.7) with p = 2, n = 3,
and
b1 = a, b2 = b, b3 = c, b5 = α, b6 = β, b7 = γ,
where
|a|2 + |α|2 = |b|2 + |β|2 = |c|2 + |γ|2 = 1.
Then ϕ satisfies the equation
7
X
ϕ(x) = 2 aj ϕ(2x ª j)
j=0

with the coefficients

1
a0 = (1 + a + b + c + α + β + γ),
8
1
a1 = (1 + a + b + c − α − β − γ),
8
1
a2 = (1 + a − b − c + α − β − γ),
8
1
a3 = (1 + a − b − c − α + β + γ),
8
1
a4 = (1 − a + b − c − α + β − γ),
8
1
a5 = (1 − a + b − c + α − β + γ),
8
1
a6 = (1 − a − b + c − α − β + γ),
8
1
a7 = (1 − a − b + c + α + β − γ).
8

9
 2.5

1.5

0.5

0
1 2 3 4

−0.5

−1

2.5

1.5

0.5

0
1 2 3 4

−0.5

−1

Figure 1: The scaling functions of example 2 for a = 0.6, b = 0.4, c =

0.57, α = 0.8, β = 0.9165, γ = 0.8216 (top) and for a = 0.9, b = 0.1, c =
0.87, α = 0.4359, β = 0.9499, γ = 0.4931 (bottom).

10
 We give graphs of ϕ for certain values of a, b, c, α, β, γ (see Figs. 1– 2).
All plottings were generated using MatLab 6.5.
Example 3. Let ϕ be given by (1.7) with p = 3, n = 2, and
b1 = a, b2 = α, b4 = b, b5 = β, b7 = c, b8 = γ,
where
|a|2 + |b|2 + |c|2 = |α|2 + |β|2 + |γ|2 = 1.
Then ϕ satisfies the equation
8
X
ϕ(x) = 3 aj ϕ(3x ª j)
j=0

with the coefficients

1
a0 = (1 + a + b + c + α + β + γ),
9
1
a1 = (1 + a + α + (b + β)ε23 + (c + γ)ε3 ),
9
1
a2 = (1 + a + α + (b + β)ε3 + (c + γ)ε23 ),
9
1
a3 = (1 + (a + b + c)ε23 + (α + β + γ)ε3 ),
9
1
a4 = (1 + c + β + (a + γ)ε23 + (b + α)ε3 ),
9
1
a5 = (1 + b + γ + (a + β)ε23 + (c + α)ε3 ),
9
1
a6 = (1 + (a + b + c)ε3 + (α + β + γ)ε23 ),
9
1
a7 = (1 + b + γ + (a + β)ε3 + (c + α)ε23 ),
9
1
a8 = (1 + c + β + (a + γ)ε3 + (b + α)ε23 ),
9
where ε3 = exp(2πi/3).
We note, that for all p, n
n
p −1
1 X (n)
aα = n d χ(α, sp−n ) (0 ≤ α ≤ pn − 1), (1.9)
p s=0 s
which follows from (1.8). This relation is an analogue of the inverse discrete
Fourier transform (for the corresponding fast algorithm see, e.g., [11, p.459]).

11
 2.5

1.5

0.5

0
1 2 3 4

0. 5

1. 5

2. 5

0.2

0.1

−0.1 1 2 3 4

−0.2

Figure 2: The real part (top) and imaginary part (bottom) of scaling function
ϕ from example 3 with a = 0.3, b = 0.5, c = 0.8124, α = 0.4, β = 0.7, γ =
0.5916.

12
2 The Walsh – Fourier transform and mul-
tiresolution p-analysis
The Walsh – Fourier transform of a function f ∈ L1 (R+ ) is defined by
Z
˜
f (ω) = f (x)χ(x, ω)dx,
R+

where χ(x, ω) is given by (1.6). If f ∈ L2 (R+ ) and

Z a
Ja f (ω) = f (x)χ(x, ω)dx (a > 0),
0

then f˜ is defined as the limit of Ja f in L2 (R+ ) as a → ∞.

The properties of the Walsh – Fourier transform are quite similar to
those of the classical Fourier transform (see, e.g., [6, Ch.6] or [11, Ch.9] ). In
particular, if f ∈ L2 (R+ ), then f˜ ∈ L2 (R+ ) and

||f˜||L2 (R+ ) = ||f ||L2 (R+ ) .

If x, y, ω ∈ R+ and x ⊕ y is p-adic irrational, than

χ(x ⊕ y, ω) = χ(x, ω)χ(y, ω), (2.1)

(see [6, § 1.5]). Thus, for fixed x and ω, the equality (2.1) holds for all y ∈ R +
except for countably many. It is known also, that the systems {χ(α, · )}∞ α=0
2
and {χ(· , α)}∞ α=0 are orthonormal bases in L [0, 1].
Accoding to [6, § 6.2] for any ϕ ∈ L2 (R+ ) we have
Z Z
ϕ(x)ϕ(x ª k)dx = |ϕ̃(ω)|2 χ(k, ω)dω, k ∈ Z+ . (2.2)
R+ R+

Let us denote by {ω} the fractional part of ω. For k ∈ Z+ , we have χ(k, ω) =

χ(k, {ω}). Thus from (2.2) it follows that
Z ∞ Z
X l+1
ϕ(x)ϕ(x ª k)dx = |ϕ̃(ω)|2 χ(k, {ω})dω
R+ l=0 l

Z 1 µX ¶
2
= |ϕ̃(ω + l)| χ(k, ω)dω.
0 l∈Z+

13
Therefore, a necessary and sufficient condition for a system {ϕ(·ªk) | k ∈ Z + }
to be orthonormal in L2 (R+ ) is
X
|ϕ̃(ω + l)|2 = 1 a.e. (2.3)
l∈Z+

Definition. A multiresolution p-analysis in L2 (R+ ) is a sequence of

closed subspaces Vj ⊂ L2 (R+ )(j ∈ Z) such that the following hold:
(i) Vj ⊂ Vj+1 forS all j ∈ Z.
(ii) The union Vj is dense in L2 (R+ ), and Vj = {0}.
T
(iii) f (·) ∈ Vj ⇐⇒ f (p ·) ∈ Vj+1 for all j ∈ Z.
(iv) f (·) ∈ V0 =⇒ f (· ⊕ k) ∈ V0 for all k ∈ Z+ .
(v) There is a function ϕ ∈ L2 (R+ ) such that {ϕ(· ª k) | k ∈ Z+ } is an
orthonormal basis of V0 .
The function ϕ is called a scaling function in L2 (R+ ).
By conditions (v) and (iii) the functions ϕ1,k (x) = p1/2 ϕ(px ª k) (k ∈
Z+ ) constitude an orthonormal basis in V1 . Since V0 ⊂ V1 , the scaling function
ϕ belongs to V1 and has the Fourier expansion
X Z
1/2
ϕ(x) = hk p ϕ(px ª k), hk = ϕ(x)ϕ1,k (x)dx.
k∈Z+ R+

This implies that

X X
ϕ(x) = p aα ϕ(px ª α), |aα |2 < +∞, (2.4)
α∈Z+ α∈Z+

where aα = p−1/2 hα . Under the Walsh – Fourier transform we have

ϕ̃(ω) = m0 (p−1 ω)ϕ̃(p−1 ω), (2.5)

where X
m0 (ω) = aα χ(α, ω). (2.6)
α∈Z+

When p = 2 a function ψ given by the formula

X
ψ(x) = 2 (−1)α aα ϕ(2x ª (α ⊕ 1)) (2.7)
α∈Z+

14
is a wavelet in L2 (R+ ), associated with the scaling function ϕ. Therefore,
the system of functions

ψj,k (x) = 2j/2 ψ(2j x ª k) (j ∈ Z, k ∈ Z+ )

is an orthonormal bases in L2 (R+ ) (cf. [1, § 5.1] and [9, § 3]).

Let p > 2. It follows from (2.3) and (2.5) that

|m0 (ω)|2 + |m0 (ω + 1/p)|2 + . . . + |m0 (ω + (p − 1)/p)|2 = 1 (2.8)

for a.e. ω ∈ [0, 1). Suppose that we have p − 1 functions

X X
ml (ω) = a(l)
α χ(α, ω), |a(l) 2
α | < +∞ (1 ≤ l ≤ p − 1),
α∈Z+ α∈Z+

such that
p−1
(ml (ω + k/p))l,k=0 (2.9)
is a unitary matrix for a.e. ω ∈ [0, 1) (for the problem of unitary extension
see, e.g., [10],[12],[13]). We set
X
ψl (x) = p a(l)
α ϕ(px ª α) (2.10)
α∈Z+

and
(l)
W0 = closL2 (R+ ) span {ψl (· ª k) | k ∈ Z+ } (1 ≤ l ≤ p − 1).

Let Wj be the orthogonal complement of Vj in Vj+1 . Then

p−1
(l)
M M
W0 = W0 , L2 (R+ ) = Wj ,
l=1 j∈Z
L
where denotes the orthogonal direct sum with the inner product of
2
L (R+ ). Moreover, the system of functions

ψj,k,l (x) = pj/2 ψl (pj x ª k) (j ∈ Z, k ∈ Z+ , 1 ≤ l ≤ p − 1) (2.11)

is an orthonormal bases in L2 (R+ ) (cf. [10], [14]).

15
3 Construction of p-wavelets
Let ϕ ∈ L2 (R+ ) satisfies the refinement equation (1.2). As before, we get
ϕ̃(ω) = m0 (p−1 ω)ϕ̃(p−1 ω), (3.1)
where
pn −1
X
m0 (ω) = aα χ(α, ω). (3.2)
α=0
Suppose that
pn −1
X
m0 (0) = 1 (i.e. aα = 1).
α=0
Put
∆s(n) := [sp−n , (s + 1)p−n ) for s ∈ Z+ .
(n) (n)
Then m0 (ω) is a constant on ∆s for each s and m0 (ω) = 1 on ∆0 . It
follows from (3.1) that
∞
Y
ϕ̃(ω) = m0 (p−j ω), ω ∈ R+ . (3.3)
j=1

(n)
We note that m0 (p−j ω) = 1 as p−j ω ∈ ∆0 (so product (3.3) is finite for
every ω ∈ R+ ).
We say that a function f : R+ 7→ C is W -continuous at a point x ∈ R+ ,
if for each ε > 0 there exists δ > 0 such that |f (x ⊕ y) − f (x)| < ε for
0 < y < δ. It is known that χ(α, · )(α ∈ Z+ ) are W -continuous functions.
By (3.2) and (3.3), the same is true for m0 and ϕ̃. Moreover, m0 and ϕ̃ are
uniformly W -continuous in [0,1) (cf. [6, § 2.3], [11, § 9.2]).
The collection {[0, p−j )| j ∈ Z} is a fundamental system of neighborhoods
of zero in the W -topology on R+ (cf. [6, § 1.2]).
Suppose that E is a W -compact set in R+ . The notation E ≡ [0, 1)
(mod Z+ ) means that for each x ∈ [0, 1) there exists k ∈ Z+ such that
x ⊕ k ∈ E. Denote by µ the Lebesgue measure on R+ .
We can now state the analogue of Cohen’s theorem (cf. [1, § 6.3] and [2,
§ 9.5]):
Theorem 2. Let
pn −1
X
m0 (ω) = aα χ(α, ω)
α=0

16
be a polynomial satisfying the following conditions:
(a) m (0) = 1.
P0p−1
(b) j=0 |m0 (sp−n ⊕ jp−1 )|2 = 1 for s = 0, 1, . . . , pn−1 − 1.
(c) There exists a W -compact set E such that int (E) 3 0, µ(E) = 1, E ≡
[0, 1)(mod Z+ ), and
inf inf |m0 (p−j ω)| > 0. (3.4)
j∈N ω∈E

If the Walsh – Fourier transform of ϕ ∈ L2 (R+ ) can be written as

∞
Y
ϕ̃(ω) = m0 (p−j ω), (3.5)
j=1

then ϕ is a scaling function in L2 (R+ ).

Remark 1. Assertion (b) of Theorem 2 is nothing but the statement
that for our polynomial m0 the equality (2.8) is true.
Remark 2. It is easy to check that m0 with the coefficients {aα } from
(1.8) satisfies all conditions of Theorem 2. For example, since

m0 (ω) 6= 0 for all ω ∈ [0, 1/p),

condition (c) holds for E = [0, 1). Therefore, the function ϕ given by (1.7) is
a scaling function in L2 (R+ ).
Theorems 1 and 2 tell us a general procedure to design p-wavelets in
2
L (R+ ) :
1. Choose a set of numbers {bm : m ∈ N(p, n), 1 ≤ m ≤ pn − 1} so that
(1.5) is true.
2. Compute {aα } by (1.9).
3. With m0 defined by (3.2) find
X
ml (ω) = a(l)
α χ(α, ω), (1 ≤ l ≤ p − 1),
α∈Z+

p−1
such that (ml (ω + k/p))l,k=0 is a unitary matrix.
4. Define ψ1 , . . . , ψp−1 by (2.10).
Remark 3. For wavelet construction the condition bj 6= 0 in (1.5) can
be replaced by assertion (c) of Theorem 2.
In connection with Remark 3 we give the following

17
 Example 4. Let p = 2, n = 3 and
½
1, if x ∈ [0, 1/4) ∪ [3/8, 1/2) ∪ [3/4, 7/8),
m0 (ω) =
0, if x ∈ [1/4, 3/8) ∪ [1/2, 3/4) ∪ [7/8, 1).

Then from (3.3) we see that ϕ̃ = 1E where E = [0, 1/2) ∪ [3/4, 1) ∪ [3/2, 7/4).
Under the inverse Walsh – Fourier transform we obtain
1 1
ϕ(x) = 1[0,2) (x) + 1[0,4) (x)[w3 (x/4) + w6 (x/4)]
2 4
(see also Example 2 for a = c = 1, b = 0). By Theorem 2, this function
generate a multiresolution 2-analysis in L2 (R+ ).

4 Proofs
To prove Theorems 1 and 2, we need the following lemma (cf. [1, § 6.3] and
[2, § 9.5]):
Lemma 1. Under the conditions of Theorem 2 the system {ϕ(· ª k) | k ∈
Z+ } is orthonormal in L2 (R+ ).
Proof. For l ∈ N let
l
Y
µ[l] (ω) = m0 (ω/pj )1E (ω/pl ), ω ∈ R+ .
j=1

(n)
Since 0 ∈ int (E) and m0 (ω) = 1 on ∆0 , we obtain from (3.5)

lim µ[l] (ω) = ϕ̃(ω), ω ∈ R+ . (4.1)

l→∞

Also, by (a) and (c), there exists a number j0 such that

m0 (ω/pj ) = 1 for j > j0 , ω ∈ E.

Thus,
j0
Y
ϕ̃(ω) = m0 (p−j ω), ω ∈ E.
j=1

By (3.4), there is a constant c1 > 0 such that

|m0 (ω/pj )| ≥ c1 for j ∈ N, ω ∈ E,

18
and so
c−j
1 |ϕ̃(ω)| ≥ 1E (ω),
0
ω ∈ R+ .
Therefore
l
Y l
Y
|µ[l] (ω)| = |m0 (ω/pj )|1E (ω/pl ) ≤ c1−j0 |m0 (ω/pj )||ϕ̃(ω/pl )|
j=1 j=1

which by (3.5) yields

|µ[l] (ω)| ≤ c1−j0 |ϕ̃(ω)| for l ∈ N, ω ∈ R+ . (4.2)

Now, for l ∈ N we define

Z
Il (s) := |µ[l] (ω)|2 χ(s, ω)dω, s ∈ Z+ .
R+

Setting El := {ω ∈ R+ | p−l ω ∈ E} and ζ = p−l ω, we have

l
|m0 (ω/pj )|2 χ(s, ω)dω =
R Q
Il (s) = El
j=1
l−1 (4.3)
l
|m0 (ζ)|2 |m0 (pj ζ)|2 χ(s, pl ζ)dζ,
R Q
=p E
j=1

where the last integrand is 1-periodic.

Using the assumption E ≡ [0, 1)(mod Z+ ) and changing the variable we
get from (4.3)
Z p−1
1X l−1
Y
l−1 2
Il (s) = p |m0 (ξ/p + i/p)| |m0 (pj−1 ξ)|2 χ(s, pl−1 ξ)dξ.
0 0 j=1

Therefore, in view of Remark 1,

Z l−2
1Y
l−1
Il (s) = p |m0 (pj ξ)|2 χ(s, pl−1 ξ)dξ
0 j=0

which by (4.3) becomes

Il (s) = Il−1 (s).

19
 Since
Z 1 Z 1
2
I1 (s) = p |m0 (ξ)| χ(s, pξ)dξ = χ(s, ξ)dξ = δ0,s ,
0 0

where δ0,s is the Kronecker delta, we get

Il (s) = δ0,s (l ∈ N, s ∈ Z+ ). (4.4)

In particular, for all l ∈ N

Z
Il (0) = |µ[l] (ω)|2 dω = 1.
R+

By (4.1) and Fatou’s lemma we then obtain

Z
|ϕ̃(ω)|2 dω ≤ 1.
R+

Thus, from (4.1) and (4.2) by Lebesque’s dominated convergence theorem

it follows that Z
|ϕ̃(ω)|2 χ(s, ω)dω = lim Il (s).
R+ l→∞

Hence by (2.2) and (4.4),

Z
ϕ(x)ϕ(x ª s)dx = δ0,s , s ∈ Z+ .
R+

Proof of Theorem 1. Put X1−n = 1[0,1/pn−1 ) . For any x ∈ R+ we have

Z Z 1/pn−1
n−1 n−1
X1−n (ω ª m/p )χ(x, ω)dω = χ(x, m/p ) χ(x, ω)dω
R+ 0

= (1/pn−1 )1[0,1) (x/pn−1 )χ(x, m/pn−1 ) = (1/pn−1 )1[0,1) (x/pn−1 )wm (x/pn−1 ).
Consequently, taking the Walsh – Fourier transform of both sides of (1.7)
gives X
ϕ̃(ω) = X1−n (ω) + A(m)X1−n (ω ª m/pn−1 ). (4.5)
m∈N(p,n)

20
If ζ ∈ [0, 1/pn−1 ) and m ∈ N(p, n), then clearly
m m
ζ⊕ =ζ+
pn−1 pn−1

since [pn+j ζ](mod p) = 0 for all negative intergers j. Hence setting ζ = ω ª

m/pn−1 , we see from (4.5) that
 (n−1)
1,
 if ω ∈ ∆0 ,
(n−1)
ϕ̃(ω) = (4.6)
 A(m), if ω ∈ ∆m ,
0 otherwise,

where m ∈ N(p, n).

Now, let the polynomial
pn −1
X
m0 (ω) = aα χ(α, ω)
α=0

(n)
satisfy the condition (1.8), that is, m0 (sp−n ) = ds for 0 ≤ s ≤ pn − 1. Then,
by the definition of {A(m)}, from (4.6) we have
∞
Y
ϕ̃(ω) = m0 (p−j ω)
j=1

and so
ϕ̃(ω) = m0 (p−1 ω)ϕ̃(p−1 ω)
which gives (1.2). By Lemma 1 and Remark 2 the system {ϕ(· ª k) | k ∈ Z+ }
is orthonormal in L2 (R+ ).

For integer m let Em (R+ ) denotes the collection of all functions f on

R+ which are constant on [sp−m , (s + 1)p−m ) for each s ∈ Z+ . Further, we
set
Ẽm (R+ ) := {f : f is W -continuous and f˜ ∈ Em (R+ )}
and ∞ ∞
[ [
E(R+ ) := Em (R+ ), Ẽ(R+ ) := Ẽm (R+ ).
m=1 m=1

The following properties are true (see [6, § 6.2 and § 10.5]):

21
 1. E(R+ ) and Ẽ(R+ ) are dense in Lq (R+ ) for 1 ≤ q ≤ ∞.
2. If f ∈ L1 (R+ ) ∩ Em (R+ ), then supp f˜ ⊂ [0, pm ].
3. If f ∈ L1 (R+ ) ∩ Ẽm (R+ ), then supp f ⊂ [0, pm ].
For ϕ ∈ L2 (R+ ) we put

ϕj,k (x) = pj/2 ϕ(pj x ª k) (j ∈ Z, k ∈ Z+ )

and
Vj = closL2 (R+ ) span {ϕj,k | k ∈ Z+ } (j ∈ Z). (4.7)
Let Pj be the orthogonal projection of L2 (R+ ) to Vj . Also, we denote the
norm in L2 (R+ ) briefly by || · ||.
As an analogue of Proposition 5.3.1 in [1] (cf. Theorem 2.2 in [12]), we
have:
T
Lemma 2. If {ϕ0,k } is an orthogonal basis in V0 , then Vj = {0}.
Proof. Let f ∈ Vj . Given an ε > 0 we choose u ∈ L1 (R+ ) ∩ Ẽ(R+ )
T
such that ||f − u|| < ε. Then

||f − Pj u|| ≤ ||Pj (f − u)|| ≤ ||f − u|| < ε

and so
||f || ≤ ||Pj u|| + ε (4.8)
for every j ∈ Z.
Now, choose R > 0 so that supp u ⊂ [0, R). Then
Z R
j/2
(Pj u, ϕj,k ) = (u, ϕj,k ) = p u(x)ϕ(pj x ª k)dx.
0

Hence, by the Cauchy – Schwarz inequality,

X X Z R
2 2 2 j
||Pj u|| = |(Pj u, ϕj,k )| ≤ ||u|| p |ϕ(pj x ª k)|2 dx.
k∈Z+ k∈Z+ 0

Therefore, if j is chosen small enough so that Rpj < 1, then

Z Z
2 2 2 2
||Pj u|| ≤ ||u|| |ϕ(x)| dx = ||u|| 1SR,j (x)|ϕ(x)|2 dx, (4.9)
SR,j R+

22
 ª k | y ∈ [0, Rpj )}. It is easy to check that
S
where SR,j := k∈Z+ {y

lim 1SR,j (x) = 0 for all x ∈

/ Z+ .
j→−∞

Thus by the dominated convergence theorem from (4.9) we get

lim ||Pj u|| = 0.

j→−∞
T
In view of (4.8), this implies that ||f || ≤ ε, and thus Vj = {0}.
Proof of Theorem 2. Let a function ϕ be defined by Walsh-Fourier trans-
form (3.5) and the spaces Vj (j ∈ Z+ ) are given by (4.7). As before, since
ϕ̃(ω) = m0 (ω/p)ϕ̃(ω/p), we have
pn −1
X
ϕ(x) = p aα ϕ(px ª α),
α=0

which implies that Vj ⊂ Vj+1 . On account of Lemma 1, we see that conditions

(i) and
T (iii)–(v) of multiresolution p-analysis are satisfied. By Lemma 2 we
have Vj = {0}. Therefore, it remains to confirm that
[
Vj = L2 (R+ )

or, equvalently, [
( Vj )⊥ = {0}. (4.10)
Let f ∈ ( Vj )⊥ . Given an ε > 0 we choose u ∈ L1 (R+ ) ∩ E(R+ ) such
S
that ||f − u|| < ε. Then for any j ∈ Z+ we have

||Pj f ||2 = (Pj f, Pj f ) = (f, Pj f ) = 0

and so
||Pj u|| = ||Pj (f − u)|| ≤ ||f − u|| < ε. (4.11)
Choose a positive integer j so large that supp ũ ⊂ [0, pj ) and p−j ω ∈
[0, p−n+1 ) for all ω ∈ supp ũ. Then we put g(ω) = ũ(ω)ϕ̃(p−j ω). As the
2 j
system {p−j/2 χ(p−j k, ·)}∞
k=0 is an orthonormal bases in L [0, p ], we have

X Z pj
2
|ck (g)| = p −j
|g(ω)|2 dω, (4.12)
k∈Z+ 0

23
where Z pj
−j/2
ck (g) = p g(ω)χ(p−j k, ω)dω.
0
Observing that
Z
ϕ(pj x ª k)χ(x, ω)dx = p−j ϕ̃(p−j ω)χ(p−j k, ω),
R+

by the Plancherel relation we get

Z pj
−j/2 −j
p (u, ϕj,k ) = p g(ω)χ(p−j k, ω)dω.
0

Thus, in view of (4.12),

X Z pj
2 2
||Pj u|| = |(u, ϕj,k )| = |ũ(ω)ϕ̃(p−j ω)|2 dω. (4.13)
k∈Z+ 0

(n)
Since m0 (ω) = 1 on ∆0 and since p−j ω ∈ [0, p−n+1 ) for ω ∈ supp ũ, it
follows from (3.5) that ϕ̃(p−j ω) = 1 for all ω ∈ supp ũ. Furthermore, because
supp ũ ⊂ [0, pj ), we obtain from (4.11) and (4.13)

ε > ||Pj u|| = ||ũ|| = ||u||.

Consequently, we conclude

||f || < ε + ||u|| < 2ε,

which implies (4.10).

References
[1] I. Daubechies, Ten lectures on wavelets, SIAM, Philadelphia, 1992.

[2] Satoru Igari, Real analysis — with an introduction to wavelet theory,

AMS Translations of mathematical monographs, 177, Providence, Rhode
Island, 1998.

24
[3] Q. Sun and Z. Zhang, M -band scaling function with filter having vanishing
moments two and minimal length, J. Math. Anal. and Appl., 222 (1998),
225–243.

[4] Malay K. Kundu and Mausumi Acharyya, M -band wavelets: application

to texture segmentation for real life image analysis, Int. J. Wavelets, Mul-
tiresolution and Inf. Processing, 1 (2003), 115–149.

[5] B.I. Golubov, A modified strong dyadic integral and derivative, Matem.
Sbornik, 193 (2002), 37–60 (in Russian); English transl. in Sbornik:
Mathematics, 193 (2002).

[6] B.I. Golubov, A.V. Efimov, and V.A. Skvortsov, Walsh series and trans-
forms, Moscow: Nauka, 1987; English transl.: Kluwer, Dordrecht, 1991.

[7] W.C. Lang, Orthogonal wavelets on the Cantor dyadic group, SIAM J.
Math. Anal., 27 (1996), 305–312.

[8] W.C. Lang, Fractal multiwavelets related to the Cantor dyadic group, Int.
J. Math. and Math. Sci., 21 (1998), 307–317.

[9] W.C. Lang, Wavelet analysis on the Cantor dyadic group, Houston J.
Math., 24 (1998), 533–544.

[10] G.V. Welland and M. Lundberg, Construction of compact p-wavelets,

Constr. Approx., 9 (1993), 347–370.

[11] F. Schipp, W.R. Wade, and P. Simon, Walsh series: An introduction to

dyadic harmonic analysis, Adam Hilger, Bristol and New York, 1990.

[12] R.Q. Jia and Z. Shen, Multiresolution and wavelets, Proc. Edinburg
Math. Soc., 37 (1994), 271–300.

[13] Chen Di-Rong, On the existence and constructions of orthonormal

wavelets on L2 (Rs ), Proc. Amer. Math. Soc., 125 (1997), 2883–2889.

[14] Yu.A. Farkov, Orthogonal wavelets on locally compact abelian groups,

Funkts. Anal. i Prilozhen., 31 (1997), 86–88 (in Russian); English transl.
in Funct. Anal. Appl., 31 (1997).

25
[15] Yu.A. Farkov, Compactly supported orthogonal wavelets on locally com-
pact abelian groups, To appear in Izv. Russian Akad. Nauk. Ser. Mat.,
69(3) (2005), 193–220 (in Russian); English transl. in Izvestiya: Mathe-
matics, 69 (2005).

Higher Mathematics and Mathematical Modelling Department,

Moscow State Geological Prospecting University
23, Ulitsa Miklukho-Maklaya, Moscow 117873, Russia
farkov@msgpa.ru

26