12.07.2022 Views

DISCRETE WAVELET TRANSFORMS IN WALSH ANALYSIS

A review of discrete wavelet transforms defined through Walsh functions and used for image processing, compression of fractal signals, analysis of financial time series, and analysis of geophysical data is presented. Relationships of the discrete transforms considered with wavelet bases recently constructed and frames on the Cantor and Vilenkin groups are noted.

A review of discrete wavelet transforms defined through Walsh functions and used for image processing, compression of fractal signals, analysis of financial time series, and analysis of geophysical data is presented. Relationships of the discrete transforms considered with wavelet bases recently constructed and frames on the Cantor and Vilenkin groups are noted.

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

DOI 10.1007/s10958-021-05476-2

Journal of Mathematical Sciences, Vol. 257, No. 1, August, 2021

DISCRETE WAVELET TRANSFORMS IN WALSH ANALYSIS

Yu. A. Farkov UDC 517.518, 621.391

Abstract. A review of discrete wavelet transforms defined through Walsh functions and used for image

processing, compression of fractal signals, analysis of financial time series, and analysis of geophysical

data is presented. Relationships of the discrete transforms considered with wavelet bases recently

constructed and frames on the Cantor and Vilenkin groups are noted.

Keywords and phrases: Walsh function, Haar system, Weierstrass function, wavelet, frame, zerodimensional

group, discrete transform, image processing, signal coding, analysis of geophysical data.

AMS Subject Classification: 42C40, 65T60

1. Introduction. It was noted in the introduction to the monograph [1] that the interpretation

of Walsh functions as characters of the Cantor dyadic group was proposed by I. M. Gelfand. A wide

class of locally compact Abelian groups (called Vilenkin groups in modern literature) containing the

Cantor group as a special case was defined in [43]. Independently, Walsh functions as characters of

the Cantor dyadic group were studied by Fain (see [23]). Basic information from the theory of Walsh

series and transforms is presented in [25, 37]. Orthogonal wavelets with compact supports on the

Cantor group and the corresponding wavelets on the positive half-line were defined in [30, 35]. As was

noted in [14], approximative properties of orthogonal wavelets with compact support on the Cantor

group are significantly different from properties of Daubeshies wavelets. In addition, Lang wavelets

(see [30]) on a fixed segment of the positive half-line can have arbitrarily high dyadic smoothness, and

for classical wavelets, the smoothness increases with increasing of the length of the support (see [7,

Sec. 7.1], [34, Sec. 7.3], and [35]). The main methods of constructing orthogonal, biorthogonal, periodic,

and nonstationary wavelets on Vilenkin groups were discussed in recent papers [14–16] (see also the

references therein). The additive group of the field of p-adic numbers Q p and the Vilenkin group G p

are locally compact zero-dimensional groups (see [1, 27]). Constructions of wavelets on the field Q p

and their applications both in the theory of functions and in mathematical physics were analyzed

in [28, 39]. Recently, it was proved (see [8]) that in L 2 (Q p ) any orthogonal wavelet basis consisting of

functions with bounded spectrum is a modification of the Haar system.

It is well known that the fast discrete wavelet transform splits signals into low-frequency (approximating)

and high-frequency (detailing) components followed by partial sampling; the inverse transform

performs the signal recovery (see, e.g., [33, Sec. 7.3]). In this paper, we present a review of discrete

wavelet transforms defined by using generalized Walsh polynomials and used for image processing

(see [19]), compressing fractal signals (see [17, 36]), analysis of financial time series (see [36]), and

analysis of geophysical data (see [32, 36, 38]).

Let p be an integer, p ≥ 2, and let R + =[0, +∞). As usual, by Z, Z + ,andN we denote the sets

of integers, nonnegative integers, and natural numbers, respectively. We denote by 〈s〉 p the remainder

when dividing an integer s by p, andby[x] the integer part of x. Foreachx ∈ R + we set

x j = 〈[p j x]〉 p , x −j = 〈[p 1−j x]〉 p , j ∈ N.

Translated from Itogi Nauki i Tekhniki, Seriya Sovremennaya Matematika i Ee Prilozheniya. Tematicheskie

Obzory, Vol. 160, Proceedings of the International Conference on Mathematical Modelling in Applied Sciences

ICMMAS’17, Saint Petersburg, July 24–28, 2017, 2019.

1072–3374/21/2571–0127 c○ 2021 Springer Science+Business Media, LLC 127


These numbers are digits of the expansion of the number x in base p:

x = ∑ x j p −j−1 + ∑ x j p −j

j<0

j>0

(for p-periodically rational x, the decomposition with a finite number of nonzero terms is obtained).

It is clear that

∞∑

[x] = x −j p j−1

j=1

andthereexistsanumberk = k(x) such that x −j =0forallj > k. By definition, the equality

z = x ⊕ y means that

z = ∑ 〈x j + y j 〉 p p −j−1 + ∑ 〈x j + y j 〉 p p −j

j<0

j>0

and, respectively, z = x ⊖ y, ifz ⊕ y = x.

Assuming that ε p =exp(2πi/p), on the interval [0, 1) we define the function

{

1, x ∈ [0, 1/p),

w 1 (x) =

ε l p, x ∈ [ lp −1 , (l +1)p −1) ,l∈{1,...,p− 1},

and continue it to the half-line R + with period 1. The Walsh generalized system {w l : l ∈ Z + } is

determined by the formula

w 0 (x) ≡ 1, w l (x) =

k∏ (

w1 (p j−1 x) ) ν j

, l ∈ N, x ∈ R + ,

where ν j are the digits of the p-ary decomposition of the number l:

l =

j=1

k∑

ν j p j−1 , ν j ∈{0, 1 ...,p− 1}, ν k ≠0, k = k(l).

j=1

The following notation will be used below.

γ(l) :=

k∑

ν j .

It is well known that the system {w l : l ∈ Z + } is an orthonormal basis in L 2 [0, 1]. In the case where

p = 2, the operations ⊕ and ⊖ coincide and the functions w l (x) are classical Walsh functions. Discrete

wavelet transforms with more than two input channels are used in multidimensional multirate signal

processing systems (see [41]). In the next section, using the functions w l (x), we define an orthogonal

discrete wavelet transform whose number of input channels coincides with the number p. Witha

special choice of parameters, we obtained from it the discrete transforms introduced by Lang in [30].

2. Orthogonal discrete wavelet transform. For an arbitrary natural n, wesetN = p n and

N 1 = p n−1 . According to [12], for any complex vector b =(b 0 ,b 1 ,...,b N−1 ) satisfying the condition

j=1

|b l | 2 + |b l+N1 | 2 + ···+ |b l+(p−1)N1 | 2 =1, 0 ≤ l ≤ N 1 − 1, (1)

we construct an orthonormal wavelet basis in the space of N-periodic complex sequences with the

standard scalar product. The condition (1) also appears in algorithms for constructing orthogonal

wavelet bases on the Vilenkin group G p (see [14, 15, 18]).

128


We denote by G(p, n) thesetofallvectorsb ∈ C N satisfying the condition (1). Using any method

of unitary matrix expansion 1 (see, e.g., [34, Sec. 2.6] and [14, p. 24]), for an arbitrary fixed vector

b (0) =(b (0)

0 ,b(0) 1 ,...,b(0) N−1

)ofG(p, n), we find the numbers b(s)

k

,0≤ k ≤ N − 1, 1 ≤ s ≤ p − 1, such

that the matrices

b (0)

l

b (0)

l+N 1

... b (0)

l+(p−1)N 1

M l =

b (1)

l

b (1)

l+N 1

... b (1)

l+(p−1)N

1

⎝.............................

⎠ , 0 ≤ l ≤ N 1 − 1,

b (p−1)

l

b (p−1)

l+N 1

... b (p−1)

l+(p−1)N 1

are unitary. Next, using the fast Vilenkin–Chrestenson transform, we calculate the coefficients

andthensetb (s)

k

k

= 1 N−1

b (s)

j

w j (k/N), 0 ≤ k ≤ N − 1, 0 ≤ s ≤ p − 1, (2)

N 1

c (s)

j=0

= c (s)

k

=0fork ≥ N.

The orthogonal discrete wavelet transform O(p, n) associated with the vector b (0) ∈ G(p, n), is

defined by formulas

a j−1,k = ∑

l⊖pk a j,l, d (1)

j−1,k = ∑

l⊖pk a j,l, ..., d (p−1)

j−1,k = ∑

l⊖pk a j,l, (3)

l∈Z +

c (0)

l∈Z +

c (1)

l∈Z +

c (p−1)

where the coefficients c (0)

k

of the low-frequency filter and the coefficients c (1)

k

, ..., c(p−1)

k

of the highfrequency

filters are defined by the formula (2) (see [7, Sec. 5.6] and [33, Sec. 7.3.2]). The dimension of

the vector a j with components a j,l is assumed to be a multiple of the number of input channels p. The

transform (3) turns the vector a j into an approximating vector a j−1 and the detalizing vectors d (1)

j−1 ,

..., d (p−1)

j−1

whose dimensions are p times smaller than the dimension of the vector a j . The inversion

formula for this transform is as follows:

a j,l = ∑

l⊖pk a j−1,k + c (1)

l⊖pk d(1) j−1,k

+ ···+ c(p−1)

l⊖pk d(p−1) j−1,k . (4)

l∈Z +

c (0)

The discrete transform O(2, 1) associated with the vector b (0) =(1/ √ 2, 1/ √ 2) coincides with the

classical Haar transform. In this case, the formulas (3) have the form

a j−1,k = a j,2k + a j,2k+1

2

, d j−1,k = a j,2k − a j,2k+1

2

.

For arbitrary p and n, the discrete Haar transform corresponds to the values

b (0)

0

= b (0)

1

= ···= b (0)

p−1 = √ 1 , b (0)

p

k

=0,k ≥ p;

then

c (s)

k

= √ 1 exp 2πiks , 0 ≤ k, s ≤ p − 1.

p p

In particular, for p =3wehave

c (0)

0 = c (0)

1 = c (0)

2 = c (1)

0 = c (2)

0 =

√ √ √

3

3 , c(1) 1 = c (2) 3

2 =

3 ε 3, c (1)

2 = c (2) 3

1 =

3 ε2 3, (5)

1 For a complete description of the set of all unitary extensions of a matrix by its first row, see [18, Theorem 9]; for

methods of matrix expansion in constructing multi-dimensional frames, see [29].

129


where ε 3 =exp(2πi/3). Lang (see [30]) considered the discrete transform O(2, 2) associated with the

vector b (0) =(1,a,0,b), where 0 <a≤ 1anda 2 + b 2 = 1. For this transform, nonzero coefficients

in (3) and (4) are determined by the formulas

c (0)

0 = c (1)

1 = 1+a + b

2 √ , c (0)

1 = −c (1)

0 = 1+a − b

2

2 √ ,

2

c (0)

2 = c (1)

3 = 1 − a − b

2 √ , c (0)

3 = −c (1)

2 = 1 − a + b

2

2 √ .

2

When processing a particular signal, a selected discrete wavelet transform is applied iteratively:

after the first step, the vectors d (1)

j−1

, ..., d(p−1)

j−1

remain in the memory, and the vector a j−1 passes

into the approximating vector a j−2 and detailing vectors d (1)

j−2

, ..., d(p−1)

j−2

, etc. At each step, the

dimensions of the vectors decrease p times. As a result, after j 0 steps, we obtain the following vectors:

a j−j0 , d (1)

j−1

, ..., d(p−1)

j−1

; ...; d(1)

j−j 0

, ..., d (p−1)

j−j 0

.

The vector a j is reconstructed from these vectors by the inverse transform.

If a multiple-scale analysis (MSA) of {V j } in L 2 (R + ) is constructed by using the vector b (0) (see [9,

15]), then, similarly to [33, Theorem 7.7], the formulas (3) and (4) allow one to quickly pass from one

level of approximating an arbitrary function f ∈ L 2 (R + ) by subspaces V j to another 2 . The freedom of

choosing the vector b (0) allows one to adapt the transform O(p, n) to the signal processed according

to the root-mean-square, entropy, or other criteria (see examples in [17, 19] and [33, Sec. 11.4.1]).

3. Coding of the generalized Weierstrass function. It was proved in a recent paper [36] that

for encoding some fractal signals, the discrete wavelet transform O(p, n) has advantages over the

discrete Haar transform and the zone coding method from [25, Sec. 11.3]. We illustrate this result by

the following example of coding values of the generalized Weierstrass function

W α,β (x) =

∞∑

α k e βkπix , 0 <α≤ 1, β ≥ 1 α .

k=1

As four original signals, we choose the values of the function W α,β (x) at the 243 nodes of the uniform

partition of the interval [0, 1) for the pairs of indices α and β listed in Table 1.

Let T N ) be the matrix composed of the numbers w(N)

l,k

= w k (l/N), 0 ≤ l, k ≤ N − 1.

As usually, we denote by TN ∗ the complex conjugate matrix T N. Then the direct and inverse discrete

multiplicative transforms defined in the space C N have the form

=(w (N)

l,k

̂x = N −1 T ∗ N x, x = T N ̂x, x ∈ C N .

The zone coding method Z(p) used in [25, Sec. 11.3] consists of applying the direct discrete multiplicative

transform to the vector x and subsequent setting to zero the last p n p n−1 components of the

vector ̂x, i.e., ̂x(p n−1 ), ̂x(p n−1 +1), ..., ̂x(p n − 1). Then, after applying the inverse discrete multiplicative

transform, we obtain the vector ˜x. The recovery error is estimated as δ = ‖x − ˜x‖ 2 / √ N.When

performing computational experiments, orthogonal discrete wavelet transforms O(3, 1) and O(3, 2)

were compared in [36] not only with the method of zone coding Z(3), but also and with the discrete

Haar transform H(3) defined by the coefficients (5). Results of calculations are presented in Table 1.

2 Recall that closed subspaces V j ⊂ L 2 (R +), j ∈ Z, formanMSAifV j ⊂ V j+1, ⋂ V j = {0}, and ⋃ V j = L 2 (R +). A

jth-level approximation of a function f ∈ L 2 (R +) is the orthogonal projection of this function onto V j.

130


Table 1. Mean-square errors of compression of the values of the function W α,β

obtained for the zone coding method, the discrete Haar transform, and

orthogonal wavelet transform.

Function Z(3) H(3) O(3, 1) O(3, 2)

W 0.6;9 0.3774 0.2835 0.2068 0.1310

W 0.8;6 0.6660 0.5393 0.4826 0.4629

W 0.8;9 0.6572 0.4599 0.3448 0.2934

W 0.9;4 1.3786 0.9706 0.8763 0.8269

4. Biorthogonal discrete wavelet transform. To generalize the discrete transform (3) to the

biorthogonal case, we take two vectors b (0) =(b (0)

0 ,b(0) 1 ,...,b(0) N−1 )and˜b (0) =(˜b (0) (0) (0)

0 ,˜b 1 ,...,˜b

N−1 )such

that

b (0) ˜b(0)

l l + b (0)

l+N 1˜b(0)

l+N 1

+ ···+ b (0)

l+(p−1)N 1˜b(0)

l+(p−1)N 1

=1, 0 ≤ l ≤ N 1 − 1,

and the unitarity of the matrix M l is replaced by the condition M l ˜Ml = I, whereI is the identity

matrix, and ˜M l is defined in the same way as M l replacing the first row by the vector ˜b (0) .Afterthis,

similarly to (2), the coefficients c (s)

k

, ˜c(s) k

,0≤ k ≤ N − 1, 0 ≤ s ≤ p − 1, are defined by the biorthogonal

discrete wavelet transform

a j−1,k = ∑

c (0)

l⊖pk a j,l,

l∈Z +

d (1)

j−1,k = ∑

c (1)

l⊖pk a j,l,

l∈Z +

..., d (p−1)

j−1,k = ∑

c (p−1)

l⊖pk a j,l.

l∈Z +

(6)

The inversion formula is as follows:

a j,l = ∑

˜c (0)

l⊖pk a j−1,k + ˜c (1)

l⊖pk d(1) j−1,k

+ ···+ ˜c(p−1)

l⊖pk d(p−1) j−1,k .

l∈Z + (7)

Algorithms for constructing matrices for which the condition M l ˜Ml = I is satisfied are available

in [34, Sec. 2.6] and [22]. The corresponding wavelet biorthogonal systems on the Vilenkin group G p are

constructed in [20]. While in Fourier analysis, filters with a finite impulse response of a discrete wavelet

transforms are trigonometric polynomials, the transform filters (6) and (7) are Walsh polynomials

m s (ω) =

N−1

k=0

c (s)

N−1

k w ∑

k(ω), ˜m s (ω) =

k=0

˜c (s)

k w k(ω), s =0, 1,...,p− 1.

As in the orthogonal case, the transforms (6) and (7) allow one to quickly pass from one level of

approximation of a function f ∈ L 2 (G p ) to another (approximations are performed by using two MSA

{V j } and {Ṽj} defined by polynomials m 0 (ω) and ˜m 0 (ω), respectively; see [10]). In the case p =2,

the discrete transforms (6) and (7) were used in [19] for image processing. We also note that by

the orthogonal and biorthogonal discrete wavelet transforms defined above, similarly to well-known

constructions (see, for example, [33, 42]), one can construct wavelet packages and separable filter sets.

5. Nonstationary discrete wavelet transform. For each j ∈ Z + , we choose complex numbers

b (0)

j+1,k , k =0, 1,...,pj+1 − 1suchthat

∑p−1

∣b (0)

j+1,νp+l

l=0

∣ 2 =1, ν =0, 1,...,p j − 1, (8)

131


and introduce the sequence of filters

with the coefficients

p j+1 −1

j+1 (ω) := ∑

c (0)

j+1,k w k(ω), j ∈ Z + , ω ∈ R + , (9)

m (0)

k=0

p j+1

c (0)

j+1,k = ∑−1

p−j−1

l=0

b (0)

j+1,l w k((j +1)/l). (10)

Applying the inversion formula of the discrete Vilenkin—Krestenson transform (see [25, Sec. 11]), from

(9) and (10) we obtain

b (0)

j+1,k = m(0) j+1 ((j +1)/k), k =0, 1,...,pj+1 − 1, j ∈ Z + .

Next, we define the unitary matrices

b (0)

j+1,νp

b (0)

j+1,νp+1

... b (0) ⎞

j+1,νp+(p−1)

M ν,j =

b (1)

j+1,νp

b (1)

j+1,νp+1

... b (1)

j+1,νp+(p−1)

⎝.....................................

⎠ ,

b (p−1)

j+1,νp

b (p−1)

j+1,νp+1

... b (p−1)

j+1,νp+(p−1)

where ν =0, 1,...,p j − 1. The first rows in M ν,j are formed by numbers from (8), and the remaining

row are defined as in [34, Sec. 2.6] or [14, p. 24]. Now, just like in (10), we set

p j+1

c (s)

j+1,k = ∑−1

p−j−1

l=0

b (s)

j+1,l w k((j +1)/l), s =1,...,p− 1,

and define the nonstationary discrete wavelet transform:

a j,k = ∑

j+1,l⊖pk a j+1,l, d (1)

j,k = ∑

j+1,l⊖pk a j+1,l, ..., d (p−1)

j,k

= ∑

l∈Z +

c (0)

l∈Z +

c (1)

l∈Z +

c (p−1)

j+1,l⊖pk a j+1,l. (11)

Its inversion formula is as follows:

a j+1,l = ∑

c (0)

j+1,l⊖pk a j,k + c (1)

j+1,l⊖pk d(1) j,k

+ ···+ c(p−1)

j+1,l⊖pk d(p−1) j,k

. (12)

l∈Z +

Numerical experiments performed in [21] demonstrate the possibility of using discrete transforms (11)

and (12) for compressing fractal signals. In these experiments, at each approximation level, the transform

coefficients (11) and (12) were chosen so that the projection of the signal onto the corresponding

approximating space was maximum (see [38]).

Some generalizations and modifications of the discrete wavelet transforms discussed above can be

obtained by extending the sets of admissible parameter values as a result of the transition from wavelets

to frames by analogy with the constructions in [2, 6]. In this connection, we note that according to [14,

18], a rigid frame in L 2 (G p ) can be constructed along any vector b =(b 0 ,b 1 ,...,b N−1 ) satisfying the

condition

∣ ∣

b 0 =1, ∣b l 2 ∣ ∣

+ b l+N1 2 ∣ ∣

+ ···+ b l+(p−1)N1 2 ≤ 1, 0 ≤ l ≤ N1 − 1. (13)

In addition, instead of above-mentioned constructions of wavelets and frames on the group G p ,

one can use their generalizations for the Vilenkin group G P associated with the sequence P =

132


{p 1 ,p 2 ,...,p j ,...}, p j ∈ N, p j ≥ 2. For example, for frame constructions similarly to (13) for each

j ∈ Z + , the complex numbers b (0)

j+1,k , k =0, 1,...,n j+1 − 1aretakensuchthat

where n 0 =1,n j =

with the coefficients

b (0)

j+1,0 =1,

p j+1 −1

l=0

∣ (0) b

j+1,νp j+1 +l

∣ 2 ≤ 1, ν =0, 1,...,n j − 1,

j∏

p l . Then the low-frequency filters are determined by the formula

l=1

n j+1 −1

j+1 (ω) = ∑

c (0)

j+1,k W k(ω)

m (0)

k=0

n j+1 −1

c (0)

j+1,k = ∑

n−1 j+1

b (0)

j+1,l W k((j +1)/l),

l=0

where W k (ω) are generalized Walsh functions associated with the sequence P. High-frequency filters

are defined in a similar way (see Sec. 2 for the case where p j = p for all j). Note that the algorithm

for constructing nonstationary orthogonal wavelets on the group G P is described in [16], which also

contains the corresponding generalization of the Haar system.

6. Periodic discrete wavelet transforms. As was noted in Sec. 2, the orthonormal wavelet basis

in the space of N-periodic complex sequences was constructed in [12] in an arbitrary complex vector

b =(b 0 ,b 1 ,...,b N−1 ) satisfying the condition (1). Another periodic discrete wavelet transform can be

obtained in the standard way by using the Lang wavelet periodization procedure (see [30]), similarly to

classical wavelets on the real line R (see, e.g., [33, Sec. 7.5.2]). In this section, we describe a construction

of a periodic discrete wavelet transform defined by using Dirichlet–Walsh-type kernels (see [10, 13,

17]).

For a given N = p n ,wesetx n,k = k/N, 0≤ k ≤ N − 1, and choose a vector b =(b 0 ,b 1 ,...,b N−1 )

whose components b k are nonzero. The functions Φ b N (x) andϕ n,k(x) are defined by the formulas

Consider the spaces

N−1

Φ b N (x) :=N ∑

−1

k=0

b k w k (x), ϕ n,k (x) :=Φ b N (x ⊖ x n,k), x ∈ R + .

V n := span { 1, w 1 (x), ..., w N−1 (x) } ,

W n (s) := span { w sN (x), w sN+1 (x), ..., w (s+1)N−1 (x) } , s =1,...,p− 1.

Recall that the scalar product in the space of 1-periodic functions on R + is determined by the formula

〈f,g〉 =

∫ 1

0

f(x)g(x) dx.

We see that the orthogonal direct sum of the spaces V n , W (1)

n

other words, if W n := W (1)

n

⊕···⊕W (p−1)

n

{ϕ n,k } N−1

k=0 is a basis of the space V n, and similar bases {ψ (s)

n,k }N−1 k=0

spaces W (s)

n , s =1,...,p− 1.

,coincideswithV n+1 ;in

,thenV n ⊕W n = V n+1 . It was proved in [13] that the system

were constructed for each of the

,...,W (p−1)

n

133


We choose a number α ∈ (0, 1) and consider the case where the components of the vector b are

given by the equalities

{

α, if k =0ork = N − 1,

b k =

1, otherwise.

Then for Φ a N

(x), the following equality holds:

NΦ a N (x) =D N (x) − (1 − α)(1 + w N−1 (x)),

where

D N (x) :=

N−1

j=0

w j (x) =

{

N, x ∈ [0,N −1 ),

0, x ∈ [N −1 , 1).

The last equality expresses the well-known property of Dirichlet–Walsh-type kernels (see, e.g., [25,

Sec. 1.5]).

For any functions f n ∈ V n and g n ∈ W n , the following equalities are valid:

where

f n (x) =

N−1

k=0

g (s)

n

p−1

C n,k ϕ n,k (x), g n (x) =

N−1

(x) = ∑

In these formulas, the sequences of coefficients

C n = {C n,k },

D (s)

n

k=0

D (s)

n,k ψ(s) n,k (x).

s=0

g n

(s) (x),

= {D (s)

n,k

}, 1 ≤ s ≤ p − 1,

are uniquely determined by f n and g n , respectively. As was shown in [13], an arbitrary function

f n+1 ∈ V n+1 can be decomposed into the direct sum of the functions f n ∈ V n and g n

(s) ∈ W n

(s) ,and

also f n+1 can be reconstructed from this decomposition by using the following discrete transforms.

1. The direct periodic discrete wavelet transform

∑p−1

N−1

C n,ν = A (n)

pk+j,ν C n+1,pk+j, D (s)

where

j=0 k=0

p−1

n,ν = ∑

j=0

B (n)

pν+j,s C n+1,pν+j, 0 ≤ ν ≤ N − 1, (14)

1

⎧⎪ ⎨

A (n)

pk+j,ν = p + 1 − α

αpN , ν = k,

B (n)

⎪ ⎩ ε γ(ν)−γ(k) 1 − α

p

αpN , ν ≠ k, pk+j,s = p−1 ε −js

p .

Recall that we denote by γ(k) the sum of digits in p-ary decomposition of the number k.

2. The inverse periodic discrete wavelet transform

where

134

C n+1,pk+j =

N−1

ν=0

Q (n)

pk+j,ν C ∑

n,ν +

Q (n)

⎪⎨ 1 − ε γ(ν)

pk+j,ν = p

⎪ ⎩ − ε γ(ν)

p

p−1

ε js

p D (s)

n,k

s=1

, 0 ≤ k ≤ N − 1, 0 ≤ j ≤ p − 1, (15)

(1 − α) γ n+1,pk+j

, k = ν,

N

(1 − α) γ n+1,pk+j

, k ≠ ν.

N


Note that the decomposition and reconstruction algorithms based on the formulas (14) and (15)

are much simpler (both in structure and in the amount of arithmetic operations) compared with the

corresponding algorithms for the case of trigonometric wavelets constructed in [5].

7. Applications to the analysis of financial time series. Main results on applications of wavelet

transforms to the analysis of financial time series are presented in the books [24, 26]. In a recent paper

[31], shares of 16 large Russian companies from the beginning of 2010 to the end of 2016 were

analyzed. Increments of time series of stock values have a strongly marked chaotic nature and have a

large amplitude of individual interference, against which a weak common signal can be distinguished

only on the basis of its correlation in different scalar components of a multidimensional time series.

Classical methods of analysis based on correlations between adjacent samples are poorly effective in

processing financial time series, because from the point of view of the correlation theory of random processes,

stock price increments formally have all the signs of white noise (in particular, “flat spectrum”

and “delta-shaped” autocorrelation function). Due to this, in [31], the passage from original signals to

sequences of their nonlinear properties calculated in time fragments of small length was performed. As

these properties, the entropy of wavelet coefficients over a Daubechies basis, multifractality indicators,

and the autoregressive measure of signal nonstationarity were used. Measures of synchronous behavior

of the properties of time series in a sliding time window by using the method of principal components,

moduli values of all pairwise correlation coefficients, and a multiple spectral measure of coherence are

constructed, which is a generalization of the quadratic coherence spectrum between two signals. Using

the method proposed, two time intervals of synchronization of the Russian stock market were identified:

from December 2013 to March 2014 and from October 2014 to January 2016. Computational

experiments showed that the same result is obtained if the entropy of wavelet coefficients is calculated

using discrete transforms defined in Secs. 2 and 5.

Various proximity measures based on wavelet coefficients for comparison time series were studied

in [3, 4]. In [36]; this technique was used for determining measures of proximity of financial time

series, which are the rates of various currencies against the US dollar. It was shown that not only

the discrete Daubechies transform is applicable for solving this problem, but also the discrete splash

transform O(2, 2). Measures of proximity between the time series corresponding to the Swiss franc

and 28 other time series were calculated. The exchange rate data were taken from the website Federal

Reserve Statistical Release 3 for the period from February 26, 1991, to December 31, 1998. The number

of samples for this period is 2048. With the help of the applied methodology, it is possible to identify

several currencies, the rates of which are closest to the Swiss franc, with respect to the following

proximity measures:

∣ ∣

∣ ∣

˜d

(x)

D 1 (x, y) =− ln

j,k

˜d (x)

j,k

(y) ˜d

j,k

∣ , D ∑

2(x, y) =1−

(y)

where

j,k

and ˜d

j,k

are normalized detailing coefficients of the series x and y, respectively. Namely,

in both cases, the first six places are occupied by the exchange rates of the following countries: The

Netherlands, Germany, Austria, Belgium, France, and Denmark.

REFERENCES

1. G. N. Agaev, N. Ya. Vilenkin, G. M. Jafarli, and A. I. Rubinstein, Multiplicative Systems of

Functions and Harmonic Analysis on Zero-Dimensional Groups [in Russian], Elm, Baku (1981).

j,k

˜d (x)

j,k

(y) ˜d

j,k

∣ ,

3 See www.federalreserve.gov/releases/h10/hist/default1999.htm

135


2. H. Bölcskei, F. Hlawatsch, and H. G. Feichtinger, “Frame-theoretic analysis of oversampled filter

banks,” IEEE Trans. Signal. Proc., 46, No. 12, 3256–3268 (1998).

3. E. V. Burnaev and N. N. Olenev, “Proximity measure for time series based on wavelet coefficients,”

in: Tr.XLVIIINauch.Konf.MFTI, Dolgoprudny (2005), pp. 108–110.

4. E. V. Burnaev and N. N. Olenev, “Proximity measures based on wavelet coefficients for comparing

statistical and calculated time series,” in: Collected Scientific and Methodical Papers [in Russian],

10, Izd. Vyatsk. Gos. Univ., Kirov (2006), pp. 41–51.

5. C. K. Chui and H. N. Mhaskar, “On trigonometric wavelets,” Constr. Approx., 9, 167–190 (1993).

6. Z. Cvetković and M. Vetterli, “Oversampled filter banks,” IEEE Trans. Signal. Proc., 46, No. 5,

1245–1255 (1998).

7. I. Daubechies, Ten Lectures on Wavelets, SIAM, Philadelphia (1992).

8. S. Evdokimov and M. Skopina, “On orthogonal p-adic wavelet bases,” J. Math. Anal. Appl., 422,

952–965 (2015).

9. Yu. A. Farkov, “On wavelets related to the Walsh series,” J. Approx. Theory., 161, No. 1, 259–279

(2009).

10. Yu. A. Farkov, “Biorthogonal wavelets on Vilenkin groups,” Tr.Mat.Inst.Steklova, 265, 110–124

(2009).

11. Yu. A. Farkov, “Periodic wavelets on the p-adic Vilenkin group,” p-Adic Numb. Ultr. Anal. Appl.,

3, No. 4, 281–287 (2011).

12. Yu. A. Farkov, “Discrete wavelets and the Vilenkin–Chrestenson transform,” Mat. Zametki, 89,

No. 6, 914–928 (2011).

13. Yu. A. Farkov, “Periodic wavelets in Walsh analysis,” Commun. Math. Appl., 3, No. 3, 223–242

(2012).

14. Yu. A. Farkov, “Constructions of MRA-based wavelets and frames in Walsh analysis,” Poincaré

J. Anal. Appl., 2, 13–36 (2015).

15. Yu. A. Farkov, “Orthogonal wavelets in Walsh analysis,” in: Generalized Integrals and Harmonic

Analysis (T. P. Lukashenko and A. P. Solodov, eds.), Izd. Mosk. Univ., Moscow (2016), pp. 62–75.

16. Yu. A. Farkov, “Nonstationary multiresolution analysis for Vilenkin groups,” in: Int. Conf. on

Sampling Theory and Applications, Tallinn, Estonia, 3-7 July 2017, Tallinn (2017), pp. 595–598.

17. Yu. A. Farkov and M. E. Borisov, “Periodic dyadic wavelets and coding of fractal functions,” Izv.

Vyssh. Ucheb. Zaved., 9, 54–65 (2012).

18. Yu. A. Farkov, E. A. Lebedeva, and M. A. Skopina, “Wavelet frames on Vilenkin groups and their

approximation properties,” Int. J. Wavelets Multires. Inform. Process., 13, No. 5, 1550036 (2015).

19. Yu. A. Farkov, A. Yu. Maksimov, and S. A. Stroganov, “On biorthogonal wavelets related to the

Walsh functions,” Int. J. Wavelets Multires. Inform. Process., 9, 485–499 (2011).

20. Yu. A. Farkov and E. A. Rodionov, “Algorithms for wavelet construction on Vilenkin groups,”

p-Adic Numb. Ultr. Anal. Appl., 3, No. 1, 181–195 (2011).

21. Yu. A. Farkov and E. A. Rodionov, “Nonstationary wavelets related to the Walsh functions,” Am.

J. Comput. Math., 2, 82–87 (2012).

22. Yu. A. Farkov and E. A. Rodionov, “On biorthogonal discrete wavelet bases,” Int. J. Wavelets

Multires. Inf. Process., 13, No. 1, 1550002 (2015).

23. N. J. Fine, “On the Walsh functions,” Trans. Am. Math. Soc., 65, 372–414 (1949).

24. Wavelet Applications in Economics and Finance (M. Gallegati and W. Semmler, eds.), Springer,

Berlin (2014).

136


25. B. I. Golubov, A. V. Efimov, V. A. Skvortsov, Walsh Series and Transformations: Theory and

Applications [in Russian], Moscow (2008).

26. F. In and S. Kim, An Introduction to Wavelet Theory in Finance: A Wavelet Multiscale Approach,

World Scientific, Singapore (2012).

27. N. Kholshchevnikova and V. A. Skvortsov, “On U- andM-sets for series with respect to characters

of compact zero-dimensional groups,” J. Math. Anal. Appl., 446, No. 1, 383–394 (2017).

28. S. V. Kozyrev, A. Yu. Khrennikov, and V. M. Shelkovich, “p-Adic wavelets and their applications,”

Tr.Mat.Inst.Steklova, 285, 166–206 (2014).

29. A. Krivoshein, V. Protasov, and M. Skopina, Multivariate Wavelet Frames, Springer, Singapore

(2016).

30. W. C. Lang, “Fractal multiwavelets related to the Cantor dyadic group,” Int. J. Math. Math. Sci.,

21, 307–317 (1998).

31. A. A. Lyubushin and Yu. A. Farkov, “Synchronous components of financial time series,” Komp.

Issled. Model., 9, No. 4, 639–655 (2017).

32. A. A. Lyubushin, P. V. Yakovlev, and E. A. Rodionov, “Multivariate analysis of fluctuation parameters

of GPS signals before and after the mega-earthquake in Japan March 11, 2011,” Geofiz.

Issled., 16, No. 1, 14–23 (2015).

33. S. Mallat, A Wavelet Tour of Signal Processing, Academic Press, San Diego (1999).

34. I. Ya. Novikov, V. Yu. Protasov, and M. A. Skopina, Wavelet Theory, American Mathematical

Society, Providence, Rhode Island (2011).

35. V. Yu. Protasov and Yu. A. Farkov, “Dyadic wavelets and scaling functions on the half-line,” Mat.

Sb., 197, No. 10, 129–160 (2006).

36. E. A. Rodionov, “On applications of wavelets to the digital signal processing,” Izv. Saratov. Univ.

Nov. Ser. Mat. Mekh. Inform., 16, No. 2, 217–225 (2016).

37. F. Schipp, W. R. Wade, and P. Simon, Walsh Series: An Introduction to Dyadic Harmonic Analysis,

Adam Hilger, New York (1990).

38. Bl. Sendov, “Adapted multiresolution analysis,” in: Functions, Series, Operators, Memorial Conf.

in Honor of the 100th Anniversary of the Birth of Prof. G. Alexits (1899–1978), Budapest, Hungary,

August 9–13, 1999(L. Leindler et al., eds.), János Bolyai Math. Soc., Budapest (2002), pp. 23–38.

39. M. Skopina, “p-Adic wavelets,” Poincaré J. Anal. Appl., 2, 53–63 (2015).

40. S. A. Stroganov, “Estimates of the smoothness of low-frequency microseismic oscillations using

dyadic wavelets,” Geofiz. Issled., 13, No. 1, 17–22 (2012).

41. M. K. Tchobanou, Multidimensional Multi-Speed Signal Processing Systems, Tekhnosfera, Moscow

(2009).

42. M. Vetterli and J. Kovačević, Wavelets and Subband Coding, Prentice Hall, New Jersey (1995).

43. N. Ya. Vilenkin, “On a class of complete orthogonal systems,” Izv. Akad. Nauk SSSR. Ser. Mat.,

11, No. 4, 363–400 (1947).

Yu. A. Farkov

Russian Presidential Academy of National Economy and Public Administration, Moscow, Russia

E-mail: farkov@list.ru, farkov-ya@ranepa.ru

137

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!