International Journal of Scientific & Engineering Research, Volume 6, Issue 1, January-2015

ISSN 2229-5518

1492

Properties of wavelet coefficients of self-similar

time series
Vyacheslav V. Lyashenko1, Zh. V. Deineko2, M. Ayaz Ahmad3
Abstract The analysis and processing of sequence of the data presented in the form of time series, is one of the prevalent
methodologies in studying of various processes and the phenomena which are concerned to different fields of activity and researches. As a
practice shows, the most of investigated time series have a property of self-similarity. At the same time, among existing variety of different
ways and methods for the data processing which are presented in the form of a time series, it is possible to allocate the ideology of a
multiresolution wavelet-analysis. The essence of such ideology consists in carrying out of wavelet-expansion of investigated data and the
subsequent analysis of corresponding coefficients of such expansion wavelets coefficients. Thus, necessity in answer to a question about
properties of such wavelet coefficients is objective. It makes a basis of the given research. Hence, the basic properties of wavelet
coefficients of self-similar time series are formulated and proved on theoretical level in this work. Also, there are presented the results of
experiments which have confirmed a theoretical substantiation of the basic properties of wavelet coefficients of self-similar time series. They
have shown that the considered properties are characteristic for time series with different correlation structure.
Index Terms correlation structure, expansion, self-similarity, time series, wavelet coefficients, wavelet-analysis.

1 INTRODUCTION
For the solution of many practical problems representation
of the initial given and received results in the form of some
sequence, in particular in the form of time series is absolutely
natural.
An example of such problems can be:
Research of change of dynamics of various indicators of
banks functioning throughout some period of time [1];
The analysis of the data which represent results of some
experiments that were carried out in strict time succession [2];
Studying of the Internet traffic bandwidth properties which
are described by different characteristics of volumes of the
transferred information by means of datacom speed [3];
Per line or columnwise analysis of the two-dimensional image which is presented in the form of a matrix of data about
investigated object [4];
Studying and revealing of regularity among the processes
connected with the different natural phenomena: floods, avalanching, spread of fires, change of volumes of water in the
rivers depending on quantity of precipitations [5];
Such generalization of different data, in the form of interconnected sequence and time series allows to not only present
and describe compactly of investigated data, but also to consider the objective nature of communication between separate
elements of the interconnected sequence, to open the data
change regularity in their analyzed sequence. Finally, it also
does proved and claimed the presentation of various analyzed
data in the form of their some sequence.

At the same time, data presentation in the form of some sequence assumes use of certain procedures of their processing
that also depends on nature of considered sequence and a context of a solved problem. In particular, some sequence of data
could have a property of self-similarity that is quite typical for
data series which describe both dynamics of various economic
and technical processes, and the phenomena of natural character [6], [7], [8]. At the same time results of processing of initial
sequences of the data presented in the form of time series can
also have some similar properties. Such similarity of properties is caused by both the unified procedures of processing of
initial data, and the similar nature of data, for example, in the
form of their self-similarity. Thus considered time series of
various processes and the phenomenon can have quite difficult structure, contain local features of the various form and
time extent. However it does not limit possibility of consideration of the general properties of results of processing of various time series, reasoning from a generality of such property
as self-similarity of considered series. It also has formed a basis for a choice of subjects, the purpose and the primary goals
of the given research.

IJSER

1. Laboratory Transfer of Information Technologies in the risk reduction

systems, Kharkov National University of RadioElectronics, 14 Lenina Str.,
Kharkov 61166, Ukraine. E-mail: lyashenko.vyacheslav@mail.ru
2. Department of Media Systems and Technology, Kharkov National University of RadioElectronics, 14 Lenina Str., Kharkov 61166, Ukraine. E-mail:
dgv_zhanna@ukr.net
3. Physics Department, Faculty of Science, P.O. Box 741,
University of Tabuk - 71491, Saudi Arabia. E-mail: mayaz.alig@gmail.com

2 A MULTIRESOLUTION WAVELET-ANALYSIS AS A METHOD

OF A SEQUENCE ANALYSIS OF INVESTIGATED DATA

One of the methods of the sequence analysis of the data

presented in the form of time series is so-called multiresolution analysis method, on the basis of the theory of wavelettransformations [6], [9], [10], [11], [12].
A multiresolution wavelet-analysis transforms time series
to hierarchical structure by means of the wavelettransformations which results to the set of waveletcoefficients. On each new level of wavelet-expansion there is a
division of an approximating signal of the previous level of
detail (presented by some time series) on its high-frequency
component and on more smoothed approximating signal [10,
12]. The number of readout in investigated process, and,

IJSER 2015
http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 1, January-2015

ISSN 2229-5518

hence, the number of coefficients decreases every time in 2

times at increase of detail level on a unit.
Thus, a multiresolution analysis consists in splitting of an
investigated number into two components approximating
and detailing, with their subsequent crushing for the purpose
of change of level of expansion of a signal to the set level of
expansion. The number of practically used wavelets on j scale
coefficient sets level of expansion of a signal [10]. In practice, it
is quite spread the use of discrete wavelet transformation
(DWT) [4]. It is connected with that application DWT becomes
especially effective when the signal has high-frequency components of short duration and extensive low-frequency components. Such signals are also met more often in practice [10],
[12].
Discrete wavelets are used, as a rule, in steam with the discrete scaling-functions j j,k ( t ) [10], [12], [13], [14]. Scalingfunctions have the general area of the value assignment with
wavelets and define ratio between such values (form). According to discrete wavelet-transformation time series X(t ) , (
t = t 1, t 2 ,... ) consists of a set of coefficients detailing and
approximating [13], [14]:
N N

Na

X(t ) = apr(N, k )N , k (t ) + det( , k ) , k (t ) ,

(1)

X, j ,k Scalar product of investigated sequence of data

in the form of time series X(t ) and a mother wavelet on
corresponding level of expansion j .
Thus, the main tool for the research of studied processes is
processing of the wavelet- coefficients which have been received on different scales. Besides, allocated factors allow to
localize places of heterogeneity and differences of analyzed
signals, that is to carry out its spatial division on areas with
prominent features. As a result of DWT received series of coefficients has the define properties which allow to investigate
behaviour of the stochastic processes that have the properties
of self-similarity. These properties are the direct subject of the
given research. However before we consider them, lets make
some additional remarks which will be useful during the subsequent statement of a material.
The characteristic of a measure of self-similarity of stochastic process at the multiresolution wavelet-analysis are the values of Hursts indicator which is an indicator of complexity of
dynamics and correlation structure of time series [9], [15].
At the same time stochastic process X(t ) with a continuous
variable of time is called as self-similar in a narrow sense with
parameter H , 0 < H < 1 if for any material value a > 0 finite-dimensional distribution for X(at) are identical for finitedimensional distributions of aH X(at) [9], [16].
In other words, if for any k 1 , t 1, t 2 ,..., t k and any a > 0
[9, 16, 17]:

IJSER

k =1

= 1k = 1

where :

apr(N, k ) Approximating wavelet-coefficients of level N ;

det( j , k ) Detailing wavelet-coefficients of level j ;
N Chosen maximum level of expansion;
N j Quantity of detailing coefficients at j level of

expansion;
N a Quantity of approximating coefficients at level N ;
(t ) Mother wavelet-function;
(t ) Corresponding scaling-function.
At set the mother wavelet and (t ) corresponding scalingfunction approximating (t ) coefficients and aX (j , k ) detailing
coefficients of d X (j , k ) DWT for the process X(t ) can be defined
as follows [13], [14]:
+

aX ( , k ) = X(t ) .k (t )dt ,

(2)

d X (j , k ) =

(3)

X(t) j.k (t)dt ,

In particular, on each level of discrete wavelet transformation

(DWT) detailing coefficients represent features, detail of the
investigated signal, arising at transition from one scale to another and are equal [9], [15]:

d X (j , k ) = X, j ,k ,

(X(t 1), X(t 2 ),...X(t k ), )

(a H X(at 1), a H X(at 2 ),...,a H X(at k ))

(5)

or

{X(t), t {t 1, t 2 ,..., t k }} {aH X(at), t {t 1, t 2 ,..., t k }},

(6)

where:
Means equality of finite-dimensional principles of distribution.
H Hurst's indicator which represents a measure of selfsimilarity of stochastic process.
In particular the formula (6) shows, that change of time
scale is equivalent to change of spatial scale of statuses. Therefore, typical realizations of self-similar process are visually
similar irrespective of a time scale on which they are considered. However it doesnt mean, that the process is repeated in
accuracy, it is more likely observed a similarity of statistical
properties, because of statistical characteristics do not vary at
scaling [9], [16].
Stochastic process X(t ) is statistically self-similar or strictly
self-similar in a broad sense, if the process aH X(at) has the
same statistical characteristics of the second order (expectation
value, dispersion and autocorrelation function), as X(t ) [9],
[16]:

(4)

where:

j,

1493

d X (j , k ) Detailing wavelet-coefficients k = 1, N j on level

IJSER 2015
http://www.ijser.org

M[X(t )] = aH M[X(at)] ,

(7)

D[X(t )] = a2H D[X(at)] ,

(8)

rx (t , s) = a2H rx (at , as) ,

(9)

International Journal of Scientific & Engineering Research, Volume 6, Issue 1, January-2015

ISSN 2229-5518

where:
M[X(t )] expectation value of the process X(t ) ,

D[X(t )] dispersion of the process X(t ) ,

rx (t , s) autocorrelation function (correlation function),
t , s time moments.
One of qualitative and quantitative characteristics of investigated sequence of data in the form of time series is the spectrum of its wavelet-energy [17].
The wavelet-energy characterizes energy of a signal on each
of expansion levels that corresponds to a certain range of frequencies. Value of wavelet-energy of corresponding level
shows the powerful of range of frequencies in a signal. At the
same time the wavelet-energy spectrum visually displays frequency structure of a signal.
The energy value Ej on the set level of wavelet-expansion j

with quantity of detailing wavelet-coefficients N j makes [16],

[17]:

1 Nj
Ej =
(dX (j , k))2 .
N j k =1

(10)

1494

Property 2.
The wavelet-coefficients received as a result of expansion of
process with stationary increments are stationary on each scale

2j .
This property follows from property of wavelet-functions

(t)dt = 0 [16], [17] and it guarantees stationarity of coefficients for SPSI.

For stationary detailing coefficients DWS for the process
with stationary p -increments, it is necessary to have zero
moments of wavelet-functions of p order.
At the same time the wavelet-function which has p -zero
moments allows to analyze more thin high-frequency structure of a signal by suppressing of slowly changing components of a signal (polynomial trend of p 1 order) [9], [17].
Considering, that in practice there are processes with stationary increments of p order, it is necessary to choose the mother
wavelet (t ) which have n > p zero moments then the detailing coefficients d X (j , k ) received at expansion, will be stationary on each level of expansion.
Property 3.
If there are moments of p order then for the waveletcoefficients which were received as a result of expansion of
process X(t ) , the following equality is carried out:

IJSER

The above listed explanations and designations allow to

formulate the basic properties of wavelet-coefficients of selfsimilar time series.

3 Properties of wavelet-coefficients of self-similar

time series
Property 1.
If casual process X(t ) is self-similar process with stationary
increments (SPSI), then detailing coefficients d X (j , k ) ,
k = 1, N j on each level of expansion j are self-similar. It
means equality of principles of distribution for series of wavelet-coefficients on each level of expansion with some scale:

(d X (j ,0), d X (j ,1),...,d X (j , N j 1))

1
j (H + )
2 (d
2

.
X (0,0), d X (0,1),...,d X (0, N j

(11)

1))

self-similar X(2j u) 2jH X(u) then [16, 17]:

dX (j , k ) = X(u)(2 u
j
2
2 X(2j u)(u

1
j (H + )
2
2

k )du

The conclusion of the formula (13) is based on the formula

(12) and properties of expectation value:
1

p
p j (H + 2)

M d X (j , k )
= M d X (0, k ) 2
=

j(H + 1)
p
2 =
.
= M d X (0, k ) M 2

As consequence, for the processes with a final dispersion

the expression (13) will be transformed into the following
form:

=
.

(14)

p
= M d X (0, k ) 2

(12)

X(u)(u k)du =

1
j (H + )
2 d

=2

j
2
k )2 du

(13)

1
j (H + )
2

This property of detailing coefficients follows from selfsimilitude of the process defined by property of scaling (compression/stretching) of mother wavelets. So if some SPSI is

1
jp(H + )
2 .

p
p
M d X (j , k ) = M d X (0, k ) 2

X (0, k )

IJSER 2015
http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 1, January-2015

ISSN 2229-5518
2
2
M d X (j , k ) = M d X (0, k ) 2j(2H + 1) .

1495

(15)

Property 4.
If X(t ) is SPSI, the correlation function of waveletcoefficients of j level decreases according with a ratio:

M[d X (j , k )d X (j , k + n)] n

2(H n )

, n , (16)

(a)

(b)

then the higher number of zero moments n of a mother

wavelet , the faster aspires to zero the correlation function

rx .
Property 5.
For the different levels of expansion j 1 j 2 and for all n
correlation of detailing coefficients of these levels d X (j 1, k )
and d X (j 2 , k + n) is equal to 0.
Property 6.
Detailing coefficients of DWS on each level of expansion j
have normal distribution with a zero average N(0, ) .

(c)
Fig. 1: A modelling time series signal-independent noise

IJSER

4 RESULTS OF EXPERIMENTAL RESEARCHES

The above listed properties of wavelet-expansion can be

used for the analysis of various real time series in practice.
In particular there was carried out the comparative analysis
of wavelet-expansion of series where stochastic components
have a different correlation structure: signal-independent
noise, processes of autoregress with short-term dependence,
fractal processes with long-term memory and non-stationary
fractal processes with a trend component. Researches were
spent for various modelling signals with small length of sample of an investigated series ( L = 128, 256, 512).
The process with short-term dependence was modeled by
the process of autoregress of the first order [9]:
(17)
X(t ) = 1X(t 1) + (t ) ,
where :
1 < 1 < 1 Some constant;
t Discrete time;
(t ) Independent from X(t ) values of a normal random
variable N(0, ) .
Gaussian fractal noise with Hurst's indicator H=0.8 has
been used as a model of fractal noise.
Non-stationary fractal process was modeled by a signal
which represents the sum of fractal noise and a polynom of a
different degree.
Signal-independent noise is presented by classical Brown
movement with Hurst's indicator H= 0.5.
The wavelet Dobeshi of 4th order db4 was used for the research of processes with a different correlation structure.
In figures 1-4 modelling time series (a), corresponding functions of correlation (b) and a wavelet-energy spectrum (c) are
presented.

IJSER 2015
http://www.ijser.org

(a)

(b)

(c)
Fig. 2: A modelling time series autoregress process

International Journal of Scientific & Engineering Research, Volume 6, Issue 1, January-2015

ISSN 2229-5518

(a)

1496

The confidential intervals of a significance value in 95 % are

shown with horizontal lines on the schedules of Figure 1 to Figure 4 of the correlation functions (b). For all modelling signals,
except a signal with a trend, autocorrelation coefficients lay in
corresponding limits. The number of readout 128 and 256 has
been chosen for modeling signals with Hurst's theoretically set
indicator, H =0.75.
Polynoms with different degree have been also generated
as a trend in the investigated series.
For the casual process of autoregress correlation function
exponentially quickly decreases to zero. Such dependence is
characteristic for many casual processes with short-term
memory.
For the fractal noise, with Hurst's indicator equal 0.75 the
correlation loss goes under the hyperbolic law that allows to
do a conclusion about presence of long memory in its dynamics.
As it specified above in Figure1 to Figure 4 the corresponding spectra of wavelet-energy which were received
with expansion of modelling signals by a mother wavelet db4
at 8 levels of expansion are also displayed. From presented
above it is possible to see, that wavelet-spectrum of fractal
and signal-independent noise have frequencies of all range
equally (Fig. 1 and Fig. 3). For the processes of autoregress
where the slow fluctuations which are defined with the large
coefficient of autoregress prevail, the spectrum shows the big
splash on low frequencies and absolutely insignificant on high
frequencies (Fig. 2).
If there is a trend in investigated signal (Fig. 4) then on low
frequencies the trend gives the big splash in coefficients. Thus,
influence of low-frequency components is appreciable.
During research properties of detailing and approximating
coefficients have been also analyzed at various levels of the
wavelet expansion and depicted in the figures such as (Figure
5 to Figure 8), where; (a) is for the function of autocorrelation
of approximating coefficients of expansion and (b) is for the
function of autocorrelation of detailing coefficients of expansion). Investigated modeling time series correspond to time
series which were presented on Figure 1 to Figure 4 found in
good agreement.

(b)

(c)

IJSER

Fig. 3: A modelling time series fractal noise

(a)

(b)

(a)

(c)

Fig. 4: A modelling time series the non-stationary process which is

presented by a polynomial trend with noise

(b)

Fig. 5: Functions of autocorrelation for approximating and detailing

coefficients of wavelet-expansion of modeling time series signalindependent noise

Further the correlation dependence of wavelet coefficients

IJSER 2015
http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 1, January-2015

ISSN 2229-5518

have been calculated mathematically and it shows that nonstationary properties are traced in a set of approximating coefficients. It is clearly mentioned in Figure 5 to Figure 8.

(a)

(b)

Fig. 6: Functions of autocorrelation for approximating and detailing

coefficients of wavelet-expansion of a modeling time series autoregress process

(a)

ynomial trend with a noise.

Autocorrelation function of fractal noise shows slow fading
under the hyperbolic law (Fig. 7), for the processes with shortterm memory autocorrelation function quickly decreases to 0
and shows a decrease according to exponential law (Figure 5
and Figure 6).
Both laws of distribution for detailing, and approximating
coefficients at different levels of expansion have been numerically investigated. It has been found out, that, detailing coefficients of DWS on each level of expansion j have normal distribution with a zero average.
At the same time researches have shown that only detailing
coefficients are normally distributed. Practically for all selective data the hypothesis has been accepted with a significance
value, = 0.05.
Finally, in we have drawn some histograms of distribution
of approximating and detailing coefficients for 8 level expansions for a mother wavelet db4. And we have depiceted in Figures 9 to12, in which (a) is the histograms of distribution approximating and (b) is the histograms of distribution of detailing
expansion) in all Figures 9-12.

IJSER
(b)

Fig. 7: Functions of autocorrelation for approximating and detailing

coefficients of wavelet-expansion of a modeling time series fractal
noise

(a)

1497

(a)

Fig. 9: Histograms of distribution of approximating and detailing

coefficients for investigated time series signal-independent noise

(a)

(b)

Fig. 8: Functions of autocorrelation for approximating and detailing

coefficients of wavelet-expansion of modeling time series the nonstationary process which is presented by a polynomial trend with a noise

The autocorrelation function of approximating coefficients

becomes a slowly decreasing curve (as it is in Fig. 8) for a pol-

(b)

(b)

Fig. 10: Histograms of distribution of approximating and detailing

coefficients for investigated time series autoregress process

Apparently from presented above, results of experiments

completely confirm theoretical calculations concerning properties of wavelet- coefficients of self-similar time series. At the same

IJSER 2015
http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 1, January-2015

ISSN 2229-5518

time it is possible to see that such properties are characteristic REFERENCES

for time series with a different correlation structure.
[1]

[2]

[3]

(a)

(b)
[4]

Fig. 11: Histograms of distribution of approximating and detailing

coefficients for investigated time series fractal noise
[5]

[6]

[8]

(a)

G. P. Kots and V. V. Lyashenko, Banking sectors of the economies of

European countries in the representation of statistical interrelation
between indices that characterize their development. European Applied Sciences. Vol. 1, pp. 461-465 (2012).
A. Khan, S. Joshi, M. Ahmad and V. Lyashenko, Some Effect of
Chemical Treatment by Ferric Nitrate Salts on the Structure and
Morphology of Coir Fibre Composites. Advances in Materials Physics
and Chemistry. Vol. 5, pp. 39-45 (2015).
A. Karpukhin, I. Kudryavtsev, A. Borisov, D. Gritsiv and H. Cho,
Computer Simulation of Chaotic Phenomena in High-Speed Communication Networks. Journal of Korean Institute of Information Technology. Vol. 11(3), pp. 113-122, (2013).
V. Lyashenko, O. Kobylin and M. A. Ahmad, General Methodology
for Implementation of Image Normalization Procedure Using its
Wavelet Transform. International Journal of Science and Research
(IJSR). Vol. 3(11), pp. 2870-2877 (2014).
M. Keblouti, L. Ouerdachi and H. Boutaghane, Spatial interpolation
of annual precipitation in Annaba-Algeria-comparison and evaluation of methods. Energy Procedia. Vol. 18, pp. 468-475 (2012).
L. Delbeke and P. Abry, Stochastic integral representation and properties of the wavelet coefficients of linear fractional stable motion. Stochastic Processes and their Application. Vol. 86 (2), pp. 177182 (2008).
W. E. Leland, M. S. Taqqu, W. Willinger and D. V. Wilson, On the
self-similar nature of Ethernet traffic (extended version). Networking,
IEEE/ACM Transactions on. Vol. 2(1), pp. 1-15 (1994).
M. W. Garrett and W. Willinger, Analysis, modeling and generation
of self-similar VBR video traffic. ACM SIGCOMM Computer Communication Review. Vol. 24 (4), pp. 269-280 (1994).
L. Kirichenko, T. Radivilova and Zh. Deineko, Comparative Analysis for Estimating of the Hurst Exponent for Stationary and Nonstationary Time Series. Information Technologies & Knowledge. Vol. 5 (1),
pp. 371388 (2011).
S. Mallat and W. L. Hwang, Singularity detection and processing
with wavelets. IEEE Transactions on Information Theory. Vol. 38 (2),
pp. 617643 (1992).
P. Abry, R. Baraniuk and P. Flandrin, The multiscale nature of network traffic: discovery analysis and modeling. IEEE Signal Processing Magazine. Vol 4 (2), pp. 518 (2002).
P. Flandrin, Wavelet analysis and synthesis of fractional Brownian
motion. IEEE Transactions on Information Theory. Vol. 38, pp. 910917
(1992).
N. Kingsbury, Image processing with complex wavelets. Philosophical Transactions of the Royal Society of London. Series A:
Mathematical, Physical and Engineering Sciences. Vol. 357 (1760), pp.
2543-2560 (1999).
C. E. Heil and D. F. Walnut, Continuous and discrete wavelet transforms. SIAM review. Vol. 31(4), pp. 628-666 (1989).
L. O. Kirichenko, Zh. V. Deineko and T. A. Radivilova, Application of
DFA Method in Fractal Analysis of Time Series of Different
Nature. Modern Stochastics: Theory and Applications. International
Conference, pp. 910 (2010).
P. Abry and F. Sellan, The wavelet-based synthesis for the fractional
Brownian motion proposed by F. Sellan and Y. Meyer: Remarks and fast
implementation. Applied and Computational Harmonic Analysis. Vol 3 (4),
pp. 377383 (1996).

IJSER
[7]

[9]

(b)

Fig. 12: Histograms of distribution of approximating and detailing

coefficients for investigated time series the non-stationary process
which is presented by a polynom with a noise

[10]

[11]

5 CONCLUSIONS
First of all, the basic moments, concerning use of a multiresolution wavelet-analysis as a method of the sequence analysis of the investigated data presented in the form of selfsimilar time series are considered in work. Such consideration
has caused necessity of carrying out of research of properties
of wavelet-coefficients of self-similar time series.
At theoretical level the basic properties of waveletcoefficients of self-similar time series are formulated and
proved.
These experiments have allowed to confirm a theoretical
substantiation of the basic properties of wavelet-coefficients of
self-similar time series. In particular, properties of approximating and detailing wavelet-coefficients, which had been
derived on different scales of discrete wavelet-expansion, have
been investigated. At that, it has been shown that theoretically
proved properties of wavelet-coefficients of self-similar time
series are characteristic for time series with a different correlation structure. It allows to use the considered properties of
wavelet-coefficients of self-similar time series in researches for
defferent applied problems.

[12]

[13]

[14]
[15]

[16]

1498

IJSER 2015
http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 1, January-2015

ISSN 2229-5518
[17] B. B. Mandelbrot and J. R. Wallis, Robustness of the Rescaled Range
R/S in the Measurement of Noncyclic Long-Run Statistical Dependence. Water Resources Research. Vol. 5, pp. 967988 (1969).

IJSER

IJSER 2015
http://www.ijser.org

1499