detc2007 34228fosp overview
TRANSCRIPT
-
8/3/2019 DETC2007 34228fosp Overview
1/18
AN OVERVIEW OF FRACTIONAL ORDER SIGNAL PROCESSING (FOSP)TECHNIQUES
YangQuan Chen Center for Self-Organizing and Intelligent Systems (CSOIS)
Department of Electrical and Computer EngineeringUtah State University
4160 Old Main Hill, Logan, Utah 84322, USAEmail: [email protected]
Rongtao SunPhase Dynamics, Inc.
1251 Columbia Dr.Richardson, TX 75081 USA
Anhong ZhouDepartment of Biological and Irrigational Engineering,
Utah State University4105 Old Main Hill, Logan, Utah 84322-4105, USA
Email: [email protected]
ABSTRACTThis paper presents a brief overview of some existing frac-tional order signal processing (FOSP) techniques where the de-
velopments in the mathematical communities are introduced; re-
lationship between the fractional operator and long-range depen-
dence is demonstrated, and fundamental properties of each tech-
nique and some of its applications are summarized. Specifically,
we presented a tutorial on 1) fractional order linear systems; 2)
autoregressive fractional integrated moving average (ARFIMA);
3) 1/f noise; 4) Hurst parameter estimation; 5) fractional orderFourier transformation (FrFT); 6) fractional order linear trans-
forms (Hartley, Sine, Cosine); 7) fractal; 8) fractional order
splines; 9) fractional lower order moments (FLOM) and 10) frac-
tional delay filter. Whenever possible, we indicate the connec-
tions between these FOSP techniques.
Keywords: Fractional order calculus, fractional order signal pro-
cessing, long range dependence, ARFIMA, 1/f noise, Hurstparameter, fractional order Fourier transformation, fractional or-
Corresponding author. Center for Self-Organizing and Intelligent Systems(CSOIS), UMC 4160, College of Engineering, Utah State University, Logan,
Utah 84322-4160, USA. Tel. 1(435)797-0148; Fax: 1(435)797-3054. URL:
http://www.csois.usu.edu/
der linear transforms, fractals, fractional order splines, fractionallower order moments (FLOM), fractional lower order statistics
(FLOS), symmetrical -stable process (SS), fractional delay.
1 Introduction
In recent years, fractional order signal processing (FOSP) is
becoming an active research area, due to the demand on analysis
of long-range dependence/ self-similarity in time series, such as
financial data, communications networks data and biocorrosion
noise [1]. We will show that FOSP is essentially based on the
idea of fractional order calculus (FOC).
Fractional order calculus is a generalization of the differ-
ential and integral operators [2]. It is the root of the fractionasystems described by fractional order differential equations. The
simplest fractional order dynamic systems include the fractional
order integrators and fractional order differentiators. The autore
gressive fractional integrated moving average (ARFIMA) mode
is a typical fractional order system. It is a generalization of au
toregressive moving average (ARMA) model [3]. The traditiona
models can only capture short-range dependence; for example
Poisson processes, Markov processes, autoregressive (AR), mov
1 Copyright c 2007 by ASME
Proceedings of the ASME 2007 International Design Engineering Technical Conferences & Computers andInformation in Engineering Conference
IDETC/CIE 2007September 4-7, 2007, Las Vegas, Nevada, USA
DETC2007-34228
-
8/3/2019 DETC2007 34228fosp Overview
2/18
ing average (MA), autoregressive moving average (ARMA) and
autoregressive integrated moving average (ARIMA) processes
[4]. For time series which possesses long-range dependence,
ARFIMA models give a good fit. In LRD (long range dependent)
processes, there is a strong coupling between values at different
times [5]. This indicates that the decay of the autocorrelationfunction is hyperbolic and decays slower than exponential de-
cay, and that the area under the function curve is infinite. We
can also say that their autocorrelation functions are power-law
distributed. 1/f noise is a signal that possesses long-range de-pendence. The power or square of some variable associated with
the random process, measured in a narrow bandwidth, is roughly
proportional to reciprocal frequency. In this paper, we will illus-
trate that the 1/f noise could be the output of a fractional ordersystem with input of white noise. The degree of LRD in time se-
ries is analyzed by estimating their Hurst parameter. Fractional
Fourier transform based estimator has been proved by this re-
search to have a better performance than the other existing esti-
mators for many time series including fractional Gaussian noise,biocorrosion processes data and Great Salt Lake water-surface-
elevation data. According to the concept of fractional Fourier
transform, some other fractional Linear transforms have been de-
veloped in the literature such as fractional Hartley transform [6],
fractional Sine transform [7] and fractional Cosine transform [7].
A brief explanation of fractals and fractional splines is pre-
sented. Finally, a concise introduction to fractional lower order
moments (FLOM) or fractional lower order statistics (FLOS) and
fractional delay is presented to make the FOSP techniques more
inclusive. The purpose of this paper is to present a brief overview
of various fractional order signal processing techniques.
2 Fractional Derivative and Integral
The historical developments culminated in two definitions
which are based on the work of Riemann and Liouville (RL) on
one hand and on the work of Grunwald and Letnikov (GL) on
the other hand [8]. Before explaining these two definitions, it
is necessary to introduce three related mathematical functions:
Gamma function, Beta function, and Mittag-Leffler function [2].
2.1 Three Special Functions
The gamma function extends the factorial function to non-
integer numbers. It is used in the definition of fractional calculus.
(z) =
0
euuz1du, for all z R (1)
Note that z could also be complex. Then, if the real part of the
complex number z is positive, the integral in (1) converges abso-
lutely. Using integration by parts,
(z + 1) = z(z). (2
The beta function has a form similar to the fractional integral/derivative of many functions. It also called the Euler integra
of the first kind defined by
B(p, q) :=
10
(1u)p1uq1du, p, q R+. (3
Its solution can be defined in terms of gamma functions,
B(p, q) =(p)(q)
(p + q)= B(q,p), p, q R+. (4
The Mittag-Leffler function is an important function that finds
widespread use in the world of fractional calculus [2]. It arise
out of the solution of non-integer order differential equations
The definition of the Mittag-Leffler function is
E(z) =
k=0
zk
(k+ 1), > 0. (5
The exponential function corresponds to the special case ofE(z)with = 1.
2.2 Definitions and Derivations
The formulation of fractional integrals and derivatives was a
natural outgrowth of integer order integrals and derivatives.
Let us start with
d
dtf(t) f(t) = lim
h0f(t) f(th)
h. (6
For the second order derivative, we know that
d2
dt2f(t) f(t) = lim
h0f(t) f(th)
h
= limh0
1
h
f(t) f(th)
h f(th) f(t2h)
h
= lim
h0f(t)2f(th) + f(t2h)
h2. (7
Similarly, for the third order derivative, we have
d3
dt3f(t) f(t) = lim
h0f(t)3f(th) + 3f(t2h) f(t3h)
h3. (8
2 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
3/18
Therefore, in general, we get
dn
dtnf(t) fn(t) = lim
h01
hn
n
j=0
(1)j
n
j
f(t jh), (9)
in which
nj
= n
(n1)(n2)...(nj+1)j!
= n!j!(nj)! .
The Grunwald-Letnikov definition of a fractional derivative
of order is given as
Dt f(t) = limh0
1
h
j=0
(1)j
j
f(t jh) (10)
where
j
= !
j!(j)! =(+1)
(j+1)(j+1) , Dt is a FOC operator.
A generalization of fractional derivative and integral operator is
aDt .
aDt =
d/dt R() > 0,
1 R() = 0,ta (d)
R() < 0.(11)
Similarly, the common formulation for the fractional inte-
gral can be derived directly from a traditional expression of the
repeated integration of a function. This approach is commonly
referred to as the Riemann-Liouville approach. It is shown in
(12) that the formula, usually attributed to Cauchy, for evaluat-
ing the n-th integration of the function f(t).
...
ta
nfold
f()d =1
(n)
ta
f()
(t)1n d (12)
for n N, n > 0. So,
aIt f(t) aDt f(t) =
1
()t
a
f()
(t)+1 d (13)
for a, R, < 0; andaD
t f(t) =
1
(n)dn
dtn
ta
f()
(t)n+1 d, (n1 < < n), (14)
where a and t are the limits ofaDt f(t).
It should be remarked that there are numerous definitions
of fractional differentiointegral operators such as Caputo, Rietz,
and so on. Interested readers are referred to [911] for detailed
discussions. For recent development, refer to [1215].
2.3 PropertiesThe properties of aD
t f(t) are summarized in the following
[16], [2] and [8].
1. If f(t) is an analytic function oft, the derivative aDt f(t) is
an analytic function oft and .2. The operation aDt gives the same result as the usual differ-
entiation of integer order n.
3. The operator of order = 0 is the identity operator.4. Fractional operators are linear:
aDt {a f(t) + bh(t)} = aaDt f(t) + baDt h(t). (15
5. For fractional integrations of arbitrary order > 0, >0, (Re() > 0,Re() > 0) the additive index law (semigroupproperty) holds:
aDt aDt f(z) = aD(+)t f(z) (16
6. Differintegration of the product of two functions
D[x(t)y(t)] =
k=0
k
x(k)(t)y(k)(t) =
k=0
k
y(k)(t)x(k)(t) (17
7. Laplace transform ofaDt :
From the Grunwald-Letnikov definition:
0
est0Dt f(t)dt= sF(s). (18
From the Riemann-Liouville definition:
0
est0Dt f(t)dt = sF(s)
n1
k=0
sk0Dk1t f(t)|t=0,(19
for n 1 < n, where F(s) = L{f(t)} is the normaLaplace transformation.
3 Fractional Linear Systems3.1 Transfer Function
Fractional linear system may be described by fractional dif
ferential equations [17]. The theory is supported by definitions ofractional derivative and integral. The integrators, differentiator
and constant multipliers are the simplest of these systems. The
lumped parameter linear systems are cascade, parallel or feed-
back of those simple systems [18, 19]. Assuming that the coeffi
cients of the equation are constant, the corresponding system is
called a fractional linear time-invariant (FLTI) system.
Define fractional linear time-invariant (FLTI) systems de
scribed by a differential equation with the following genera
3 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
4/18
form:
N
n=0
anDny(t) =
M
m=0
bmDmx(t) (20)
where D is the derivation operator and the is the order of differ-entiation. According to (18), applying Laplace transform to (20),
the following transfer function can be easily obtained, provided
that Re(s) > 0 or Re(s) < 0.
H(s) =Mm=0 bms
m
Nn=0 ansn
. (21)
3.2 Impulse Response
To obtain the impulse response from the transfer function,
we start by considering the simple case:
H(s) =1
s(22)
in which could be a fraction or a real number. According toRiemann-Liouville integral, the impulse response
h(t) = t1()
, > 0. (23)
Let us proceed with a further step by considering the following
transfer function:
H(s) =1
sa . (24)
For the inverse Laplace transform of (24), using the sum of the
first q terms of a geometric sequence with = 1/q we obtain:
H(s) =1
sa =
qj=1 a
j1s1j
saq . (25)
The inverse Laplace transform of (25) is
h(t) =1
a
q
j=1
ajE1j(t, aq) (26)
where
E1j(t, aq) = eaqt
0
1 j
k
t1+j+k
(j+ k)aq
k. (27
For the transfer function (21), we can expand H(s) in partiafractions like (24) when it is in commensurate order. Invert each
partial fraction and add the different partial impulse responses to
get the full impulse response.
For numerical simulation in Matlab of FLTI systems, a tool
similar to Matlab Control Systems Toolbox is provided and illus
trated in [20, 21].
4 Autoregressive Fractional Integrated Moving Average
A typical fractional system is autoregressive fractional in-tegrated moving average (ARFIMA) or fractional autoregressive
integrated moving average (FARIMA) [22]. It is a generalization
of autoregressive moving average (ARMA) model.
4.1 Moving Average Model (MA)
The notation MA(q) means a moving average model with qterms. An MA(q) model can be written as
yt = xt +1xt1 + ... +qxtq (28
for some coefficients 1,...,q. A moving average model is essentially a finite impulse response (FIR) filter.
4.2 Autoregressive Model (AR)
The notation AR(p) means an autoregressive model with pterms. An AR(p) model can be written as
yt = 1yt1 + ... +pytp +xt (29
for some coefficients 1,...,p. An autoregressive model is essentially an infinite impulse response (IIR) filter.
4.3 Autoregressive Moving Average Model (ARMA)
The notation ARMA(p, q) means a model with p autoregressive terms and q moving average terms. This model subsume
the AR and MA models,
yt = 1yt1 + ... +pytp +xt +1xt1 + ... +qxtq. (30
4 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
5/18
4.4 Autoregressive Integrated Moving Average(ARIMA)
An ARIMA model ARIMA(p, d, q) is a generalization ofARMA model. The p, d and q are integers greater than or equal
to zero and refer to the order of the autoregressive, integrated and
moving average parts of the model respectively.According to (30), an ARMA(p, q) model can be written as
(1p
i1
iLi)Xt = (1 +
q
i=1
iLi)t, (31)
where L is the lag operator such that LXt = Xt1, Xt is a giventime series and t are error terms. The error terms t are assumedto be normally distributed with zero mean such as Gaussian white
noise.
Then, the ARIMA model is generalized by adding a differ-
encing parameter d to (31).
(1pi1iLi)(1L)dXt = (1 +qi=1iLi)t, (32)(1L)d = k=0
dk
(L)k
where d is a positive integer.
A well-known special case, ARIMA(0, 1, 0) model, is givenby:
Xt = Xt1 + (33)
which is simply a random walk when is a white noise.
4.5 Autoregressive Fractional Integrated Moving Av-erage (ARFIMA)
The autoregressive fractional integrated moving average
model is generalized by permitting the degree of differencing to
take fractional values. The fractional differencing operator is de-
fined as an infinite binomial series expansion in powers of the lag
operator L. Fractionally differenced processes may exhibit long-
term memory (long-range dependence) or antipersistence (short
term memory) [22].
The usefulness of this model has been proved by numer-
ous studies. For instance, Diebold and ReRudebush applied
ARFIMA to real GNP (Gross National Product) data [23]; Baillieet al. found long memory in time series of inflation [24]; Carto
and Rothamn used an ARFIMA(0, d, 1) model for annual bondyields [25]; Bertacca et al. applied an ARFIMA based analysis
in sea SAR (Synthetic Aperture Radar) imagery [26]; Liu et al.
used ARFIMA for network traffic modeling [27].
An ARFIMA (p, d, q) process may be differenced a finiteintegral number until d lies in the interval [1
2, 1
2], and will then
be stationary and invertible [22]. This range is the most useful
set ofd.
1. When d= 12
, the ARFIMA (p, 12
, q) process is stationarybut not invertible.
2. When 12
< d < 0, the ARFIMA (p, d, q) process has a
short memory, and decay monotonically and hyperbolicallyto zero.
3. When d = 0, the ARFIMA (p, 0, q) process can be whitenoise.
4. When 0 < d< 12
, the ARFIMA (p, d, q) process is a stationary process with long memory, and is very useful in mod-
elling long-range dependence (LRD). The autocorrelation o
a LRD time series decays slowly as a power law function.
5. When d= 12
, the spectral density of the process is
s() =(ei)(ei)(ei)(ei)
{(1 ei)(1ei)}d
11 ...q11p
2
(2sin 12)2d C
(34
as 0. Thus the ARFIMA (p, 12
, q) process is a discretetime 1/f noise.
5 1/f NoiseWe may define a fractional stochastic process as the output
of a fractional dynamic system [28, 29]. The input could be sim-
ply white noise. Among various fractional random signals, the
1/f noise is of special importance. The 1/f noise is a randomprocess defined in terms of the shape of its power spectral den-
sity S(f). The power or square of some variable associated withthe random process, measured in a narrow bandwidth, is roughly
proportional to reciprocal frequency:
S(f) =constant
|f| . (35
The typical spectrum of 1/f noise is shown in Fig. 1.Models of 1/f noise were developed by Bernamont [30] in
1937 for vacuum tubes and by Mcwhorter [31] in 1955 for semi-
conductors. In [32], Keshner referred several examples of 1/
noise. From his considerations, we may conclude that those sig-
nals seem to belong to one of two types: those with spectrum ofthe form 1/f for every f and the others with that form only for fabove a given value. This means that the former can be consid-
ered as the output of a fractional integrator, while the second may
be the output of a low-pass fractional system with a pole near the
origin. For the first case, the system can be defined by the transfe
function H(s) = 1/s and its impulse response h(t) = t1/()with 0 < < 1. Let the input be a continuous-time stationarywhite noise w(t) with variance 2. The output can be written in
5 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
6/18
0 5000 100000
500
1000
1500
2000
2500
3000
3500Natural scale
Frequency
Spectrum
1/f noise
100
102
104
100
101
102
103
104
Loglog scale
Frequency
Spectrum(
db)
1/f noise
Figure 1. 1/f noise spectrum
convolution form as follows:
y(t) =1
()
t
w()(t)1d. (36)
The autocorrelation function of the output y(t) is then,
R() = 2||21
2(2) cos. (37)
There is an important property shown in (37). The autocor-relation of the 1/f noise decays slowly as a power law function,
which indicates the long-range dependence [33].
5.1 Fractional Gaussian Noise
Fractional Gaussian noise is a kind of 1/f noise. It is long-range dependent with power law behavior in all frequencies [34].
Alternatively, the FGN process can be defined as a process Xi, i =1, 2... satisfying the condition for any timescales kand l [34]
(Z(k)i k) d= (
k
l)H(Z
(l)j l) (38)
where Zi is denoting the aggregated process of Xi on the
timescale k, Zj is denoting the aggregated process of Xi on the
timescale l, is the mean value of Xi, the symbold= stands for
equality in distribution and H is a positive constant (0 < H < 1)known as the Hurst parameter. Hurst parameter will be discussed
in detail in the next section.
It is shown in [34] that, the autocovariance function r() of
FGN Xi is given by
r() =1
2[(+ 1)2H + (1)2H]2H, > 0 (39
Apart from small , this function is very well approximated by
r = H(2H1)2H2 (40
which shows that autocorrelation is a power function of lag .
Considering a continuous time process Y(t) with autocorrelation Cov(Y(t),Y(t+ )) = c2H2, c = H(2H 1), we canalso obtain (39) by making Y(t) discrete using time intervals oany length and taking Xi as the average of Y(t) in the interval [i, (i + 1)]. This enables an approximate calculation of thepower spectrum of the process as
s(k)() 4
0c2H2 cos(2)d c12H (41
which is a power law of the frequency .
Since FGN is a very fundamental signal, many FGN generat
ing methods have been proposed. For example, FGN can be used
as a benchmark signal to validate the accuracy of the proposed
Hurst parameter estimators [35]. The relationship between frac-
tional order dynamic systems, long range dependence and power
law has been clearly illustrated in Fig. 2 [13].
.
-
: . . . ~ .
.
-
Fractional Calculus, LRD, Power Law,
ssH
1)( =
White Noise
)()(
1
=
t
th
cos)2(2)(
12
2
=
Ryy
u(t) y(t)
2/1 f
Power laws in
Signal/Systems
Probability distributionRandom processes (correlation functions)
noise (signal) generation via fractional dynamic system
y(t) is a Brownian motion when =1, i.e., process.2/1 f
Figure 2. Relationship between fractional order dynamic systems, long
range dependence and power law [13]
Some widely used FGN generation methods are briefly re
viewed below. Other related papers include [3639]
6 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
7/18
5.1.1 A Fast Fractional Gaussian Noise Genera-tor [40] This method is based on an algorithm to compute anapproximation to discrete fractional Gaussian noise. The algo-
rithm computes fast fractional Gaussian noise (FFGN) [40] as a
sum of a low frequency term and a high frequency term. The high
frequency term is a Markov-Gauss process and the low frequencyterm is a weighted sum of Markov-Gauss processes. Then, the
FFGN is given by
Xi = X(L)i +X
(H)i , (42)
where X(L)i and X
(H)i denote low and high frequency terms re-
spectively. Writing M(n)t for the nth Markov-Gauss process with
variance 1 and weight factor W12
n , we define
X(L)i =N
n=0
W1
2n M(n)t , (43)
Wn =H(2H1)(32H) (B
1HBH1)B2n(H1). (44)
The high frequency term X(H)i is used to compensate the defi-
ciency due to the approximation leading to (44). Finally, there
remains only the choice of B and N. Mandelbrot found 2,3 or 4
convenient for B [40], and Chi et al. (1973) recommend about
15-20 for N [41].
5.1.2 A Multiple Timescale Fluctuation Approach[34] It is shown in [34] that the autocorrelation function of theFGN process at the basic timescale could be approximated by
the weighted sum of exponential functions of the time lag. This
observation can lead to an algorithm to generate FGN. The best
(in terms of mean square error) approximation of (39) is given
by the following equation
r= 1.52(H0.5)1.32. (45)
It should be mentioned that this algorithm is based on the same
principle with the FFGN algorithm [40].
5.1.3 A Disaggregation Approach [34] [5] The dis-aggregation approach is a mathematical induction technique
which is made possible by the expressions of the statistics of
the aggregated FGN process [5]. The first step is sufficient to
describe the method application. Assume that the generation
is completed at the timescale k n and the time series will begenerated at the next timescale k/2. In the generation step, the
higher-level amount Z(k)i (1 < i < n/k) is disaggregated into two
lower-level amounts Z(k/2)2i1 and Z
(k/2)2i such that
Z(k/2)2i1 +Z
(k/2)2i = Z
(k)i . (46
Therefore, (46) can be used to generate Z(k/2)2i1 and then obtain
Z(k/2)2i . At this generation step, both the values of previous lower-
level time steps, i.e., Z(k/2)1 , ..., Z
(k/2)2i2 and the values of nex
higher-level time steps, i.e., Z(k)i+1, ..., Z
(k)n/k have become avail
able. To simplify the method, the correlations ofZ(k/2)2i1 with only
one higher-level time step behind and one ahead are considered
Thus, Z(k/2)2i1 can be generated from the linear relationship
Z(k/2)2i1 = a2Z
(k/2)2i3 + a1Z
(k/2)2i2 + b0Z
(k)i + b1Z
(k)i+1 +V (47
where a2, a1, b0 and b1 are parameters to be estimated and V isinnovation whose variance has to be estimated. The unknown
parameters can be estimated according to the correlations in the
form of Cor[Z(k/2)2iC1 ,Z
(k/2)2iC1+j] = rj , where rj is given by (39).
5.1.4 A Symmetric Moving Average Approach [34[42] The symmetric moving average (SMA) approach is intro-
duced by [42], which could be used to generate stochastic pro-cess with any kind of autocorrelation structure. For an inpu
white noise Wi, the SMA filter takes the weighted average of a
number ofWi to produce the output, for example
Xi =q
j=q
a|j|Wi+j = aqWiq + ... + a1Wi1 +
a0Wi + a1Vi+1 + ... + aqVi+q (48
where aj are the weights symmetric about a center (a0) that corresponds to the variable Wi and q theoretically is infinity but in
practice can be restricted to a finite number. The autocovariance
implied by (48) is
rj =qj
l=qa|j|a|j+l|, j = 0, 1, 2... . (49
It is also shown in [42] that the discrete Fourier transform
sa() of the aj sequence is related to the power spectrum of the
7 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
8/18
process sr() by
sa() =
2sr(). (50)
We can use (39) to approximate the inverse Fourier transform ofsa() to get aj:
aj
(22H)r032H [(j + 1)
H+0.5 + (j1)H+0.52jH+0.5], j > 0 (51)
In the end, the generation scheme (48) with coefficients aj can
lead to the SMA approach.
6 Hurst Parameter Estimation
The Hurst parameter H characterizes the degree of LRD. A
process is said to have long range dependence when 0.5
-
8/3/2019 DETC2007 34228fosp Overview
9/18
tional form expected, typically either FGN or FARIMA. Errors
may occur if the user fails to specify the model. Local Whittle is
a semi-parametric estimation method.
6.2.2 Periodogram Method [51] The periodogram isdefined by
I() =1
2N
Nj=1
Xjei j
2
, (56)
where is the frequency and i =1. For a series with finite
variance, I() is an estimate of the spectral density of the series.A log-log plot of I() should have a slope of 12H close to theorigin.
6.2.3 Wavelet Based Method Wavelet analysis hasbeen used with success to measure the Hurst parameter in re-
cent years [52], [53], [54], [55], [56], [50], [28], [57], [58], [59]
and [60]. To decide a suitable frequency-domain Hurst param-
eter estimation method for comparison, wavelet based estimator
seems to be the right choice. Wavelets can be thought of as akin
to Fourier series but using waveforms other than sine waves. The
Hurst exponent is calculated from the wavelet spectral density by
fitting a linear regression line through a set ofxj,yj points, wherexj is the octave and yj is the logarithm of the normalized power.
The slope of this regression line is proportional to the estimate
for the Hurst exponent. A 95% confidence interval can be given.
6.2.4 FrFT Based Estimator In [35] and [61], weproposed to use a fractional Fourier transform (FrFT) based es-
timator. It uses the spectrum calculated by FrFT for estimation.
The FrFT based local estimator has proven to have better per-
formances than other existing methods in extensive validation
experiments performed in [62].
7 Fractional Fourier Transform (FrFT)
Fractional Fourier transform may be considered as a frac-
tional power of the classic Fourier transform [63]. The first ideaof fractional power of the Fourier operator appears in 1929 [64].
Like the complex exponentials are the basis functions in Fourier
analysis, the chirps (signals sweeping all frequencies in a certain
frequency interval) are the basis in fractional Fourier analysis.
Since FrFT can provide a richer picture in time-frequency anal-
ysis [65] and it is nothing more than a variation of the standard
Fourier transform [66], it can improve the performance of some
signal processing applications.
Figure 3. Rotation concept of Fourier transform.
7.1 Formulation and DerivationLet us start with interpretation of a variable along a rotated
axis system. Like the time and frequency variables in the timefrequency plane, let x be the variable along the x-axis and isthe variable along the -axis (Fig. 3). Let xa and a represent therotated variables x and , respectively. We have
xaa
=
cosF sinFsinF cosF
x
, (57
where F = a/2 is the rotating angle. It is always assumed thata = xa+1. Therefore, xa and a are always orthogonal.
If f(x) is a time signal of the variable x, it lives on thehorizontal axis. Its Fourier transform (FT) F(f(x)) = F() is afunction of the frequency , and hence it lives on the vertical axisFourier transform changes signal f(x) in time domain x to F()in frequency domain , which corresponds to a counterclockwiserotation over an angle /2 in the (x,) plane. Since applying FTtwice to f(x) results in
(F2f)(x) = (F(F f))(x) =12
F()eixd = f(x), (58
F2 is called the parity operator. Thus, the time axis rotated over
an angle for F2.Similarly, it follows that
(F3f)() = (F(F2f))() =12
f(x)eixdx = F(), (59
which corresponds to a rotation of the representation axis over
3/2. If we apply one more F , we should get another rotationover /2, which brings us back to the original time axis. There-fore,
F4(f) = f or F4 = I. (60
In conclusion, the FT operator corresponds to a rotation of the
axis over an angle /2 in the time-frequency plane.
9 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
10/18
There are six definitions of FrFT [63]. The most intuitive
way to define FrFT is by generalizing the rotation concept of
classical FT. Like FT corresponds to a rotation in the time-
frequency plane over an angle F = /2, the FrFT will corre-spond to a rotation over an arbitrary angle F = a/2 with a
R.
For a more formal definition, FrFT can be defined through
eigenfunctions. We can also start with defining the classical FT.
According to (60), the eigenvalues are all in the set {1,i,1, i}and thus FT has only four different eigenspaces. A possible
choice for the eigenfunctions of the operator F that is generally
agreed upon is given by the set of normalized Hermite-Gauss
functions [67]:
n(x) =21/4
2nn!ex
2/2Hn(x), (61)
where Hn(x) = (1)nex2Dnex2 , D= d/dx is a Hermite poly-nomial of degree n. Therefore, there is an eigenvalue n suchthat
Fn = ein/2n. (62)
Thus, the eigenvalue for n is given by n = ein/2 = n with
= i = ei/2 representing a rotation over an angle /2.
Similarly, FrFT can be defined for a rotating angle F =a/2 by
Fan = eina/2n = ann =
nan. (63)
Since any function f in the eigenspaces can be expanded in
terms of these eigenfunctions f = n=0 ann with
an = 12nn!
2
Hn(x)ex2/2f(x)dx, (64)
we can get
fa := Faf = Fa[
n=0
ann] =
n=0
anFan =
n=0
aneina/2n. (65)
For computational purposes, (65) should be changed to an in-
tegral representation by replacing the an in the series by their
integral expression in (64):
fa() =
n=0
[
n(x)f(x)dx]eina/2n()
=
n=0
eina/2Hn()Hn(x)2nn!
e(x2+2)/2f(x)dx
=1
1e2iF
exp
2xeiF e2iF(2 +x2)
1 e2iF
exp
2 +x2
2
,
(66
where in the last step Mehlers formula [68] has been used.
n=0
eina/2Hn()Hn(x)2nn!
= exp 2xeiFe2iF(2+x2)1e2iF (1e2iF) (67
With the following expressions,
2xeiF1e2iF =ixcscF, (68
1
1e2iF
=e
i2 (
2 FF)
2|sinF| , (69
e2iF1e2iF +
1
2= i
2cotF, (70
where F = sgn(sinF), we can obtain a more tractable integrarepresentation ofF as (71).
fa() := (Faf)() =
ei2 (
2 FF)e
i2
2 cotF
2|sinF|
exp i x
sinF+
i
2x2 cotF
f(x)dx,
(71
where 0 < |F| < .
It is shown that defining via (71), the FrFT exists in the same
conditions as in which the FT exists [63]. Therefore, we have
10 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
11/18
proved the definition of FrFT in linear integral form as follows.
The ath order of fractional Fourier transform of f(x) is
fa() =
Ka(,x)f(x)dx, (72)
where Ka(,x) = AF exp[i(cotF22cscFx + cotFx2)],
AF =
1 i cotF, F = a/2.
7.2 Properties of the Kernel Function Ka(,x)
By Fb = FbaFa, we have
fa() := (Faf)() =
ei2 (
2 FF)e
i2
2 cotF2|sinF|
exp i xsinF + i2x2 cotFf(x)dx.
(73)
Using the expressions in (73) and the interpretation of the
FrFT as a rotation, it is directly verified that the kernel Ka has the
following properties [67].
If Ka(,x) is the kernel of the FrFT as in (72), then
1. Ka(,x) = Ka(x,);
2. Ka(,x) = Ka(,x);3. Ka(,x) = Ka(,x);4.
Ka(, t)Kb(t,x)dt = Ka+b(,x);
5. Ka(t,)Ka(t,x)dt = (x).
7.3 Convolution of FrFT
Since FrFT is a generalization of FT, it has similar prop-
erties as FT. Therefore we only mention the convolution prop-
erty here. FT transforms a convolution in time domain into a
product in frequency domain. This property remains true when
it concerns the convolution of two functions in the domain of
the FrFT Fa: if ga(xa) = fa(xa) ha(xa), then its FT becomesga+1(a) = fa+1(a)ha+1(a) [67]. Thus,
F{F[fa(xa)ha(xa)]} = F{fa+1(xa+1)ha+1(xa+1)}= fa+2(xa+2)ha+2(xa+2). (74)
7.4 FrFT of Some Common Functions
The fractional Fourier transform of some common functions
can be derived by the use of (72).
FrFT of Delta Function [63]
FrFTF{()}=
1 i cotFexp(i2 cotF) (75
FrFTF{( )}=
1 i cotFexp[i(2 cotF2cscF +2 cotF)] (76
Here, F is rotating angle of FrFT. As can be seen from (75) and(76), a delta function is transformed into a linear chirp function
by the FrFT.
FrFT of a Sinusoid [63]
FrFTF{exp(i2)} = 1 + i tanFexp[i(2 tanF2s+2 tanF)].
Notice that similar to the delta function, the FrFT of a sinusoid
is also a linear chirp function.
FrFT of a Constant [63]
FrFTF{C}= C
1 + i tanF exp(i2 tanF) (78
Since the DC term can be considered as a sinusoid with zero
frequency, the result is a special case of the previous FrFT pair
with = 0.
FrFT of a Chirp Function [63]
FrFTF{exp(i2)}=
1 + i tanF1 + tanF
exp
i2
tanF1 + tanF
(79
The FrFT of a linear chirp function is also a linear chirp function
with a special type of sweeping rate.
FrFT of a Gaussian Function [63]
FrFTF{exp(2)}= exp(2) (80
It is interesting to notice that the FrFT of a Gaussian function is
still a Gaussian function. In fact, the result is independent of the
transform angle F.
11 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
12/18
0
0.5
1
1.5
2
0
2
4
6
80
2
4
6
8
10
Fractional order
FrFT spectrum of a Sine function
Frequency
FrFTSpectrum
Figure 4. Fractional Fourier transform of a sine function.
40 30 20 10 0 10 20 30 400
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1symmetric delta
Figure 5. A sample of symmetric delta function.
0
0.5
1
1.5
2
40
20
0
20
400
0.2
0.4
0.6
0.8
1
Fractional order
FrFT spectrum of a symmetric delta function
Frequency
F
rFTSpectrum
Figure 6. Fractional Fourier transform of the symmetric delta function.
Figure 4 is an example of fractional Fourier transform of a
sine function. The orders of FrFT vary from 0.0 to 2.0.
For a symmetric delta function shown in Fig. 5, the FrFT
spectra with orders in (0,2) are plotted in Fig. 6.
8 Fractional Hartley and Sine/Cosine Transform
This section is re-written in concise form based on [6], [7
and [69].
8.1 Fractional Linear Transform
The derivation of fractional Fourier transform can be gener
alized to fractional linear transform. Let T be a linear transform
that maps function f(x) into the function g(). Then, the corresponding fractional transform ofT can be defined by
Ta{f(x)}= ga() (81
Let n(x) be an eigenfunction of linear transform T with eigenvalue n, then, we have
T{n(x)}= nn(). (82
Similar to (66), we can show that
Ta{f(x)} =
n=0
anann()
=
n=0
[
f(x)n(x)dx]ann()
=
Ka(x,)f(x)dx (83
where the transform kernel is defined by
Ka(x,) =
n=0
ann(x)n(). (84
8.2 Fractional Hartley TransformConsider the well-known Hartley transform [6] defined by
T{f(x)}= gH() = 12
f(x)cas(x)dx (85
where cas(x) = cos(x) + sin(x).
According to the linear fractional transform method, the
12 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
13/18
fractional Hartley transform can be given by
gaH() =
KaH(x,)f(x)dx (86)
where the Hartley kernel function is
KaH(x,) =
1 j cot
2ej
x2+2
2 cot[cos(xcsc) +
ej(2 ) sin(xcsc)] (87)
where = a/2. It is clear that the fractional Hartley transformgaH() of real function f(x) is complex valued except = k/2for integers k. Besides, the eigenvalues of Hartley transform
2
n = 1 for n = 0, 1,...,. (88)
Thus, the period of fractional Hartley transform is 2. Moreover,
the relationship between fractional Fourier transform and frac-
tional Hartley transform can be derived as
gaH() =1 + ej
a2
2fa() +
1ej a22
fa() (89)
8.3 Fractional Cosine and Sine Transforms
The fractional cosine and sine transforms [7] are now ac-
tively used in optics and signal processing.
The Cosine transform is defined as
CT{f(x)}= gC() = 12
cos(x)f(x)dx. (90)
The Sine transform is defined as
ST{f(x)}= gS() = 12
sin(x)f(x)dx. (91)
In [7], the fractional Cosine/Sine transforms have been de-
rived by taking the real/imaginary parts of the kernel of FrFT.
gaC() =
Re[KaF(,x)]f(x)dx, (92)
gaS() =
Im[KaF(,x)]f(x)dx. (93)
0 5 10 15 20 25 30 350
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
half delta
Figure 7. A sample of right half delta function
0
0.5
1
1.5
2
0
10
20
30
400
0.2
0.4
0.6
0.8
1
Fractional order
FrCT of a half delta function
Frequency
FrCTSpectrum
Figure 8. Fractional Cosine transform of the half delta function
As illustrations, Fig. 8 is the fractional Cosine transform of
the half delta function in Fig. 7. The orders of FrCT vary from
0.0 to 2.0. Figure 10 is the fractional Sine transform of the hal
delta function in Fig. 9. The orders of FrST vary from 0.0 to 2.0
9 Fractals9.1 What Are Fractals?
A huge amount of applications of FOC are motivated by
the research of fractals. In 1967, Mandelbrot investigated self
similarity in papers such as How Long Is the Coast of Britain?
Statistical Self-Similarity and Fractional Dimension [70]. After
that, Mandelbrot introduced the word fractal to denote an ob
ject whose Hausdorff-Besicovitch dimension is greater than it
13 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
14/18
35 30 25 20 15 10 5 01
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
half delta
Figure 9. A sample of left half delta function
0
0.51
1.5
2
40
30
20
10
00
0.5
1
1.5
Fractional order
FrCT of a half delta function
Frequency
FrCTSpectrum
Figure 10. Fractional Sine transform of the half delta function
topological dimension [71]. He illustrated fractals with images
based on recursion in [71].
A fractal is a geometric shape which is self-similar and has
fractional dimension. Self-similar has the same meaning as long-
range dependence. But for a geometric shape, a simple explana-
tion of LRD could be a small part of the shape has a similar
appearance to the full shape. Figure 11 is an example of fractals,
which is from [72].
9.2 Fractal Dimension (FD)
In fractal geometry, the fractal dimension is a statistical
quantity that gives an indication of how completely a fractal ap-
pears to fill space, as one zooms down to finer and finer scales
[73]. Fractals can have non-integer dimensions like 2.5. That is,
while we are in the 3 dimensional space, looking at a piece of
Figure 11. Sierpinsky Triangle
paper which is in 2 dimension, fractals are in between the integer
dimensions. However, fractals always have a smaller dimension
than what they are on. If you draw a Kochs Curve [74] on a
plane, the fractal cant have a dimension higher than the plane
which is 2.
It can be assumed that for any fractal object (of size P, made
up of smaller units of size p), the number of units (N) that fitsinto the larger object is equal to the size ratio (P/p) raised to thepower ofd, which is called the Hausdorff dimension [75].
D =logN
log(P/p). (94
There are some other ways to define fractional dimensions
such as box-counting dimension [76], information dimension
[77], and the correlation dimension [78].
As an example of calculating FD, consider Sierpinsky trian
gle shown in Fig. 11. It contains 3 identical triangles, each of
which requires 2X magnification to become identical to the en
tire figure. The fractal dimension is then log3/ log2, which isapproximately 1.58.
Long-range dependent time series can also be described by
a fractal dimension D which is related to the Hurst parameter
through D = 2
H [79]. Here, the fractal dimension D can be
interpreted as the number of dimensions the signal fills up. Besides, porous media model for the hydraulic system has fracta
dimension [80]. For example, the so called porous ball built by
the French group CRONE has been used in cars hydraulic cir
cuit. When the hydraulic gas passes through the porous ball
the pressure/speed shows a fractional order dynamics which is
apparently related to the FD of the ball [81]. Also, the prepa
ration of nanoparticles coated bio-electrodes is by polishing the
surface with fractal shapes. In addition, the diffusion behavio
14 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
15/18
of bioelectrochemical process will be fractional order dynamic,
which is related with FD [72].
10 Fractional Splines
10.1 SplinesThe first mathematical reference to splines was discovered
by Schoenberg in 1946 [82]. They are used for smoothing poly-
nomial approximations. A wide class of functions that are used
in applications requiring data smoothing can be referred to as
splines [83]. In [84], splines have been used to minimize the
measures of roughness.
A polynomial spline is a piecewise polynomial function
[85]. In its most general form a polynomial spline S : [a, b] Rconsists of polynomial pieces Pi : [xi,xi+1]R, where the knotssatisfy a = x0 < x1 < ... < xk2 < xk1 = b. The given k ponitsxi are called knots. The vector x = (x0,...,xk1) is called a knotvector for the spline. If the polynomial pieces on the subintervals
[xi,xi+1], i = 0,...,k2 all have a degree of at most n, the splineis said to be of degree n or of order n + 1 [86].
10.2 Fractional Splines
The fractional splines are a generalization of Schoenbergs
polynomial splines for all fractional orders [87, 88]. They in-
volve the fractional operator and many techniques derived from
wavelet and approximation theory. The piecewise power func-
tions of fractional degree a are the basic configuration of frac-
tional splines. Since the natural building blocks for the frac-
tional splines are Liouvilles one-sided power functions, the frac-
tional splines of degree a with the increasing sequence of knots
{xk}k Z as functions that can be written in the following form:
sa(x) = kZ
k(xxk)a+ (95)
where the xks are the knots of the spline. xa+ = max(0,x)
a is one
sided power function.
Before the fractional B-splines are discussed, a relevant def-
inition should be given as
a+f(x) = k
0
(1)k
a
k
f(x k). (96)
Similar to the classical B-splines, [87] defines the fractionalcausal B-splines by taking the (+ 1)th fractional difference ofthe one-sided power function
a+(x) =1
(a + 1)a+1+ x
a+ =
1
(a + 1) k0
(1)k
a + 1
k
(x k)a+. (97)
Figure 12 is an example of fractional B-splines with a > 0 [87].The fractional splines have the following properties [87]:
Figure 12. The fractional B-splines [87]
1) If a is an integer, fractional splines are equivalent to the
classical polynomial splines.
2) The fractional splines are a-Holder continuous for a > 03) The fractional B-splines satisfy the convolution property
and a generalized fractional differentiation rule. Besides, they
decay at least like xa2.4) The fractional splines have a fractional order of approxi-
mation a + 1.5) Fractional spline wavelets essentially behave like frac
tional derivative operators.
11 Fractional Lower Order Moments (FLOM) andFractional Lower-Order Statistics (FLOS)
We all know that for random processes, the Gaussian modeis commonly used in statistic signal processing. This Gaussian
model assumption is usually acceptable and justifiable by the
Central Limit Theorem. A generalization of this assumption i
the so-called stable model which can be used to characterize
a broad class of non-Gaussian processes [8991] including un-
der water acoustic signals, low frequency atmospheric noise and
many man-made noises. It has been proven that using stable
model and fractional lower-order statistics (FLOS), additiona
benefits can be gained using this type of fractional order sig-
nal processing technique [90]. Note that, statistics of fractiona
lower order means < 2. When = 2, it is the Gaussian process which is the basis of second-order moment theory. When
> 2, it is called the high order statistics (HOS) [92,93]. When < 2, the density function has a heavier tail than Gaussian dis-tribution. Therefore, FLOS noises are more spiky. In addition
the variance or the second-order moment of a stable distribution
with < 2 does not exist. In fact, for a non-Gaussian stabledistribution with characteristic exponent , only the moments oorders less than are finite. Therefore, variance can no longer beused as a measure of dispersion and in turn, many standard sig-
nal processing techniques such as spectral analysis and all leas
15 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
16/18
squares (LS) based methods may give misleading results.
A simple test of infinite variance is to plot the running sam-
ple variance estimate Sn with respect to number of points n where
S2n = (nk=1(xk xn)2)/(n 1) and xn = nk=1xk/n. For finite
variance processes xk, Sn will converge to a constant value as n
increases. If Sn does not converge to a constant value, xk is anon-Gaussian infinite-variance process with fractional lower or-
der < 2.An excellent collection of papers on stable distributions,
processes and related topics can be found in [94] where it is
also clear that there is a natural link between LRD and heavy
tail or thick/fat/heavy processes characterized by FLM/FLOS.
A special case is the so-called SS (symmetrical -stable) pro-cess [89, 95] which finds wide applications in engineering and
non-engineering domains.
A very interesting fact is that, there is a strong link between
fractional calculus and stable processes [96, 97].
Due to the space limit, we shall not expand our discussion
here. However, the papers and web links cited in this sectionshould give a good start if the readers are interested in exploring
this FOSP technique.
12 Fractional DelayIn many digital signal processing tasks in communication
systems, it is often required to design a filter that implements a
fractional delay, that is, the block of zD with D a nonintegerpositive real number and z the Z-transform symbol. An excellent
review on fractional delay can be found in [98] with a tutorial
flavor.
Fractional delay filter design based on B-spline transform is
proposed in [99]. Closed-form design of all-pass fractional delayfilters is proposed in [100].
Most interestingly, in [101], improved design of digital
fractional-order differentiators approximating s using fractional
sample delay is proposed. This shows a connection between frac-
tional calculus and fractional delay filter. Once again, details
are omitted here in the interest of space limit. Interested readers
can consult the 207 papers at ieeeXplore.ieee.org from
1990-2007 on Fractional Delay.
13 ConclusionsIn this paper, we have given an overview of some exist-
ing fractional order signal processing (FOSP) techniques wherethe developments in the mathematical communities are intro-
duced; relationship between the fractional operator and long-
range dependence is demonstrated, and fundamental properties
of each technique and some of its applications are summarized.
Specifically, we presented a tutorial on 1) fractional order linear
systems; 2) autoregressive fractional integrated moving average
(ARFIMA); 3) 1/f noise; 4) Hurst parameter estimation; 5)fractional order Fourier transformation (FrFT); 6) fractional or-
der linear transforms (Hartley, Sine, Cosine); 7) fractal; 8) frac
tional order splines; 9) fractional lower order moments (FLOM)
and 10) fractional delay filter.
We believe that FOSP techniques will keep developing and
in many realistic applications, FOSP will find more and more
unique place that conventional signal processing methods cannoeven deliver the answer. It is expected the FOSP will become a
more dedicated research field with a strong emphasis on applica-
tion driven scenarios.
ACKNOWLEDGMENTWe wish to thank Review #1 for his/her two-page long review comments
Both reviewers helped improve this paper. This research was supported in par
by USU Skunk Works Research Initiative Grant ( Fractional Order Signal Pro
cessing for Bioelectrochemical Sensors, 2005-2006).
REFERENCES[1] Zaveri, N., Sun, R., Zufelt, N., Zhou, A., and Chen, Y., 2007. Evaluation
of microbially influenced corrosion with electrochemical noise analysis
and signal processing. Electrochimica Acta, 52 , p. (57955807).
[2] Loverro, A., 2004. Fractional calculus: history, definitions and applica
tions for the engineer. Department of Aerospace and Mechanical Engi
neering, Univeristy of Notre Dame.
[3] Ilow, J., and Leung, H., 2001. Self-similar texture modeling using
FARIMA processes with applications to satellite images. IEEE Trans
actions on Image Processing, 10 (5) , pp. 792797.
[4] Box, G., and Jenkins, G., 1976. Time series analysis: forecasting and
control, 2nd ed. Holden-Day, San Francisco.
[5] Koutsoyiannis, D., 2001. Coupling stochastic models of different tim
scales. Water Resources Research, 37 (2) , pp. 379392.
[6] Pei, S.-C., Tseng, C.-C., Yeh, M.-H., and Jian-Jiun, D., 1998. A new
definition of continuous fractional Hartley transform. 98. Proceedings o
the 1998 IEEE International Conference on Acoustics, Speech, and Signa
Processing, 1998. ICASSP, 3 May , pp. 14851488.
[7] Lohmann, A. W., Mendlovic, D., Zalevsky, Z., and Dorsch, R. G., 1996
Some important fractional transformations for signal processing. Optic
Communications, 125 (1) Apr. , pp. 1820.
[8] Samko, S., Kilbas, A., and Marichev, O., 1993. Fractional integrals and
derivatives: theory and applications. Gordon and Breach Science Pub
lishers. ISBN: 2-88124-864-0.
[9] Ortigueira, M. D., Machado, J. A. T., and da Costa, J. S., 2005. Which
differintegration. IEE Proceedings - Vision, Image, and Signal Process
ing, 152 (6) , pp. 846850.
[10] Magin, R. L., 2004. Fractional calculus in bioengineering. Critica
ReviewsTM in Biomedical Engineering, 32 (1-4) .
[11] Podlubny, I., 1999. Fractional Differential Equations. Academic Press.
[12] Vinagre, B. M., and Chen, Y., 2002. Lecture notes on fractiona
calculus applications in automatic control and robotics. In The 41s
IEEE CDC2002 Tutorial Workshop # 2, B. M. Vinagre and Y. ChenEds., pp. 1310. [Online] http://mechatronics.ece.usu.edu
/foc/cdc02 tw2 ln.pdf.
[13] YangQuan Chen, 2006. Ubiquitous Fractional Order Con
trols?. In the Proc. of The Second IFAC Symposium o
Fractional Derivatives and Applications (IFAC FDA06). July
2006, Porto, Portugal. pp. 1-12. Plenary lecture slides avail
able at http://mechatronics.ece.usu.edu/foc
/fda06/01ifac-fda06-plenary-talk#5-chen-utah.ppt
[14] J. A. Tenreiro Machado (Guest Editor), 2002. Special issue on fractiona
calculus and applications. Nonlinear Dynamics, 29 March , pp. 1385.
16 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
17/18
[15] Manuel Duarte Ortigueira and J. A. Tenreiro Machado (Guest Editors),
2003. Special issue on fractional signal processing and applications.
Signal Processing, 83 (11) Nov. , pp. 22852480.
[16] Chen, Y., 2003. Fractional-order calculus, fractional-order filter and
fractional-order control: an overview & some recent developments. In
USU Industrial Math Seminar. http://www.csois.usu.edu/people/yqchen/.
[17] Moshrefi-Torbati, M., 1995. Fractional calculus and its applications todynamic systems. PhD thesis, Institute of Sound and Vibration Research,
Southampton Univ., Oct.
[18] Ortigueira, M. D., 2000. Introduction to fractional linear systems. part
1. continuous-time case. Vision, Image and Signal Processing, IEE Pro-
ceedings, 147 (1) , pp. 6270.
[19] Ortigueira, M. D., 2000. Introduction to fractional linear systems. part 2.
discrete-time case. Vision, Image and Signal Processing, IEE Proceed-
ings, 147 (1) , pp. 7178.
[20] Xue , D., a nd Chen, Y., 2007. Solving Control Re-
lated Mathematic Problems in Matlab, 481 pages (in chi-
nese) ed. Tsinghua University Press, Beijing, China.
http://mechatronics.ece.usu.edu/controlmathbook/ .
[21] Xue, D., Chen, Y., and Atherton, D. P., 2007. Linear Feed-
back Control: Analysis and Design with MATLAB. SIAM Press.
http://mechatronics.ece.usu.edu/siambook/ .
[22] Hosking, J., 1981. Fractional differencing. Biometrika, 68 (1) , pp. 165
176.
[23] Diebold, F., and ReRudebush, G., 1989. Long memory and persistence
in aggregate output. Journal of Monetary Economics, 24 , pp. 189209.
[24] Baillie, R., Chung, C., and Tieslau, M., 1996. Analyzing inflation by the
fractionally integrated Arfima-Garch model. Journal of Applied Econo-
metrics, 11 , pp. 2340.
[25] Carto, N., and Rothamn, P., 1994. Fractional integration analysis of long-
run behavior for US macroeconomic time series. Economic Letters, 45 ,
pp. 287291.
[26] Bertacca, M., Berizzi, F., Mese, E. D., and Capria, A., 2004. A FARIMA
based analysis for wind falls and oil slicks discrimination in sea SAR im-
agery. Geoscience and Remote Sensing Symposium, 2004. IGARSS 04.
Proceedings. 2004 IEEE International, 7 , pp. 4703 4706.
[27] Liu, J., Shu, Y., Zhang, L., Xue, F., and Yang, O., 1999. Traffic mod-
eling based on FARIMA models. 1999 IEEE Canadian Conference on
Electrical and Computer Engineering, 1 , pp. 162167.
[28] Abry, P., and Veitch, D., 1998. Wavelet analysis of long range dependent
traffic. IEEE Transactions on Information Theory, 44 , pp. 215.
[29] Ortigueira, M. D., and Batista, A. G., 2004. A fractional linear sys-
tem view of the fractional Brownian motion. Nonlinear Dynamics, 38 ,
pp. 295303.
[30] Bernamont, J., 1937. Fluctuations de potentiel aux bornes dun conduc-
teur metallique de faible volume parcouru par un courant. Annales de
Physique, 7 , p. 71.
[31] McWhorter, A. L., 1955. 1/f noise and related surface effects in Germa-nium. PhD thesis, MIT, Cambridge, MA.
[32] Keshner, M. S., 1982. 1/f noise. Proceedings of the IEEE, 70 (3) ,pp. 212218.
[33] Mandelbrot, B. B., and Ness, J. W. V., 1968. Fractional Brownian mo-
tions, fractional noises and applications. SIAM Review, 10 , pp. 422437.
[34] Koutsoyiannis, D., 2002. Hurst phenomenon and fractional Gaussiannoise made easy. Hydrological Sciences Journal, 47 (4) , pp. 573596.
[35] Chen, Y., Sun, R., and Zhou, A., 2007. An improved Hurst parameter es-
timator based on fractional Fourier transform. In Proc. of the ASME De-
sign Engineering Technical Conferences, ASME, pp. DETC200734242.
3rd ASME Symposium on Fractional Derivatives and Their Applications
(FDTA07), part of the 6th ASME International Conference on Multibody
Systems, Nonlinear Dynamics, and Control (MSNDC).
[36] Reed, I. S., Lee, P. C., and Truong, T. K., 1995. Spectral representation
of fractional Brownian motion in n dimensions and its properties. IEEE
Trans. on Information Theory, 41 (5) , pp. 1439 1451.
[37] Li, M., and Chi, C.-H., 2003. A correlation-based computational mode
for synthesizing long-range dependent data. Journal of the Franklin In
stitute, 340 , p. 503514.
[38] Karagiannis, T., Faloutsos, M., and Riedi, R. H., 2002. Long-range de
pendence: now you see it, now, you dont!. In Proceedings of the Global
Internet.
[39] Chang, Y.-C., and Chang, S., 2002. A fast estimation algorithm onthe Hurst parameter of discrete-time fractional Brownian motion. IEEE
Trans. on Signal Processing, 50 (3) , pp. 554 559.
[40] Mandelbrot, B. B., 1971. A fast fractional Gaussian noise generator
Water Resources Research, 7 (3) , pp. 543553.
[41] Chi, M., Neal, E., and Young, G., 1973. Practical application of fractiona
Brownian motion and noise to synthetic. Water Resources Research, 9
pp. 15231533.
[42] Koutsoyiannis, D., 2000. A generalized mathematical framework fo
stochastic simulation and forecast of hydrologic time series. Water Re
sources Research, 36 (6) , pp. 15191534.
[43] Hurst, H. E., 1951. Long-term storage capacity of reservoirs. Proceed
ings American Society of Civil Engineering, 116 , pp. 770799.
[44] Mejia, J. M., Rodriguez-Iturbe, I., and Dawdy, D. R., 1972. Streamflow
simulation: 2. the broken line process as a potential model for hydrologic
simulation. Water Resources Research, 8 (4) , pp. 931941.
[45] Mandelbrot, B. B., and Wallis, J. R., 1969. Computer experiments with
fractional Gaussian noises. Water Resources Research, 5 (1) , pp. 228
267.
[46] Beran, J., 1994. Statistics for Long-Memory Processes. Chapman & Hall
New York.
[47] Taqqu, M., Teverovsky, V., and Willinger, W., 1995. Estimators for long
range dependence an empirical study. Fractals, 3 (4) , pp. 785788.
[48] Taqqu, M. S., and Teverovsky, V., 1998. Onestimating long-range depen
dence in finite and infinite variance series. A Practical Guide to Heavy
Tails: Statistical Techniques and Applications, R. J. Adler, R. E. Feldman
and M. S. Taqqu editors , pp. 177217.
[49] C.K. Peng, S.V. Buldyrev, M. S. H. S., and Goldberger, A., 1994. Mosaic
organization of DNA nucleotides. Physical Review E, 49 , pp. 1685
1689.
[50] Robinson, P. M., 1995. Gaussian semiparametric estimation of long
range dependence. Annals of Statistics, 23 , pp. 16301661.
[51] Geweke, J., and Porter-Hudak, S., 1983. The estimation and application
of long memory time series models. Journal of Time Series Analysis, 4
pp. 221238.
[52] Veitch, D., and Abry, P., 1999. A wavelet based joint estimator of th
parameters of long-range dependence. IEEE Trans. Info. Theory, Specia
Issue on Multiscale Statistical Signal Analysis and its Applications, 45 (3
, pp. 878897.
[53] Taqqu, M. S., Teverovsky, V., and Willinger, W., 1995. Estimators fo
long-range dependence: An empirical study. Fractals, 3 , pp. 785788.
[54] Stoev, S., Taqqu, M., Park, C., and Marron, J. S., 2005. On the wavele
spectrum diagnostic for hurst parameter estimation in the analysis of in-
ternet traffic. Computer Networks, 48 (3) , pp. 423445.
[55] Li, Y., Szu, H. H., Sheng, Y., and Caulfield, H. J., 1993. Wavelet pro
cessing and optics. Proceedings of the International Society for Optica
Engineering, 2051 , pp. 130141.
[56] Stoev, S., Taqqu, M., Park, C., Michailidis, G., and MarronJ. S., 2004. LASS: a tool for t he l ocal analysis of self
similarity in internet traffic. SAMSI Technical Reports (7)
www.eecs.umich.edu/ msim/Publications/lass-paper11.pdf.
[57] Stoev, S., Taqqu, M., Park, C., and Marron, J., 2004. Strenghts and limi
tations of the wavelet spectrum method in the analysis of internet traffic
In SAMSI Technical Reports.
[58] Beran, J., Sherman, R., Taqqu, M. S., and Willinger, W., 1995. Long
range dependence in variable-bit-rate video traffic. IEEE Transaction
on Communications, 43 (2/3/4) .
[59] Stoev, S., Pipiras, V., and Taqqu, M., 2002. Estimation of the self
17 Copyright c 2007 by ASME
-
8/3/2019 DETC2007 34228fosp Overview
18/18
similarity parameter in linear fractional stable motion. Signal Processing,
82 , pp. 18731901.
[60] Karagiannis, T., Molle, M., and Faloutsos, M., 2004. Long-range depen-
dence: ten years of internet traffic modeling. Internet Computing, IEEE,
8 , pp. 5764.
[61] Sun, R., Chen, Y., Zaveri, N., and Zhou, A., 2006. Local analysis of long-
range dependence based on fractional Fourier transform. In Proceedingof the 2006IEEE MountainWorkshop on Adaptive and Learning Systems,
pp. 1318.
[62] Sun, R., 2007. Fractional order signal processing: Techniques and ap-
plications. Master of Science thesis, Dept. of Electrical and Computer
Engineering,Utah State University.
[63] Ozaktas, H. M., Kutay, M. A., and Zalevsky, Z., 2000. The Fractional
Fourier transform with Applications in Optics and Signal Processing.
John Wiley & Sons.
[64] Condon, E. U., 1937. Immersion of the Fourier transform in a continuous
group of functional transformations. Proceeding of National Academy
Sciences, 23 , pp. 158164.
[65] Almeida, L. B., 1994. The fractional Fourier transform and time-
frequency representations. IEEE Transactions on Signal Processing, 42
(11) , pp. 30843091.
[66] Zayed, A. I., 1996. On the relationship between the Fourier transform
and fractional Fourier transform. IEEE Signal Processing Letters, 3 (12)
, pp. 310311.
[67] Bultheel, A., and Martnez, H., 2002. A shattered survey of fractional
fourier transform. Manuscript.
[68] Namias, V., 1980. The fractional order Fourier transform and its applica-
tion in quantum mechanics. Journal of the Institute of Mathematics and
Its Applications, 21 , pp. 241265.
[69] Pei, S.-C., and Ding, J.-J., 2002. Fractional cosine, sine, and Hartley
transforms. IEEE Transactions on Signal Processing, 50 (7) , pp. 1661
1680.
[70] Mandelbrot, B. B., 1967. How long is the coast of Britain? statistical
self-similarity and fractional dimension. Science, 156 , pp. 636638.
[71] Mandelbrot, B. B., 1975. Stochastic models for the earths relief, the
shape and the fractal dimension of the coastlines, and the number-area
rule of islands. Proceedings of the National Academy of Sciences (USA),
72 , pp. 38253828.
[72] Chen, Y., 2005. Fractional order signal processing in biology/biomedical
signal analysis. In Fractional Order Calculus Day at Utah State Univer-
sity. http://mechatronics.ece.usu.edu/foc/event/FOC Day@USU/.
[73] Wikipedia, 2006. Fractal dimension wikipedia, the free encyclopedia.
[Online; accessed 24-July-2006].
[74] Vinoy, K., Abraham, J. K., and Varadan, V. K., 2003. On the relationship
between fractal dimension and the performance of multi-resonant dipole
antennas using Koch curves. IEEE Transactions on Antennas and Prop-
agation, 51 , pp. 22962303.
[75] Greenberg, J., and Kinsner, W., 1998. Characterizing behaviour of
birth/death processes using fractal measures. 1998. IEEE Canadian Con-
ference on Electrical and Computer Engineering, 1 May , pp. 373376.
[76] Schroeder, M. R., 1991. Fractals, Chaos, Power Laws: Minutes from an
Infinite Paradise. W H Freeman & Co (Sd).
[77] Parker, T. S., and Chua, L. O., 1987. Chaos: A tutorial for engineers.
Proceedings of the IEEE, 75 , pp. 9821008.[78] Grassberger, P., and Procaccia, I., 1983. Measuring the strangeness of
strange attractors. Physica, 9D , pp. 189208.
[79] Blok, H. J., 2000. On the nature of the stock market: simulations and
experiments. PhD thesis, University of British Columbia.
[80] Cushman, J. H., 1991. On diffusion in fractal porous media. Water
resources research, 27 (4) , pp. 643644.
[81] Altet, O., Moreau, X., Moze, M., Lanusse, P., and Oustaloup, A., 2004.
Principles and synthesis of hydractive CRONE suspension. Nonlinear
Dynamics, 38 , pp. 435459.
[82] Schoenberg, I. J., 1946. Contributions to the problem of approximation
of equidistant data by analytical functions, part a:on the problem of
smoothing or graduation. Quart. Appl. Math., 4 (1) , pp. 4499.
[83] Ahlberg, J. H., Nilson, E. N., and Walsh, J. L., 1967. TheTheory of Spline
and Their Applications. New York: Academic Press.
[84] Bartels, R. H., Beatty, J. C., and Barsky, B. A., 1987. An Introduction to
Splines for Use in Computer Graphics and Geometric Modeling . Morgan
Kaufmann.[85] Unser, M., 1999. Splines: a perfect fit for signal and image processing
Signal Processing Magazine, IEEE, 16 , pp. 2238.
[86] Wikipedia, 2006. Spline (mathematics) wikipedia, the free encyclope
dia. [Online; accessed 27-September-2006].
[87] Unser, M., and Blu, T., 2000. Fractional splines and wavelets. SIAM
Review, 42 (1) , pp. 4367.
[88] ORTIGUEIRA, M. D., 2003. A newsymmetric fractional b-spline. SIG
NAL PROCESSING, 83 (11) , pp. 23112318.
[89] Shao, M., and Nikias, C. L., 1993. Signal processing with fractiona
lower order moments: stable processes and their applications. Proceed
ings of the IEEE, 81 (7) , pp. 986 1010.
[90] Nikias, C. L., and Shao, M., 1995. Signal Processing with Alpha-Stable
Distributions and Applications . Wiley, New York.
[91] Nolan, J., 2007. Information on stable di st ri but ions
http://academic2.american.edu/
jpnolan/stable/
stable.html.
[92] Nikias, C. L., and Petropulu, A., 1993. Higher-order Spectra Analysis
A Nonlinear Signal Processing Framework. Prentice Hall, Englewoo
Cliffs, NJ.
[93] Swami, A., 2003. Higher Order Spectral Analysis t ool
box for M AT LA B. M AT LA B Central F ile E xchange
http://www.mathworks.com/matlabcentral
/fileexchange/loadFile.do?object
Id=3013&objectType=file.
[94] Nolan, J., 2007. Bibliography on sta
ble distributions, processes and related topics
http://academic2.american.edu/jpnolan/stable/StableBibliography.pdf.
[95] Gonzalez, J. G., Paredes, J. L., and Arce, G. R., 2006. Zero-order statis
tics: A mathematical framework for the processing and characterizatio
of very impulsive signals. IEEE Trans. on Signal Processing, 54 (10)
pp. 3839 3851.
[96] Gorenflo, R., and Mainardi, F., 1998. Fractional calculus and sta
ble probability distributions. Gorenflo, R. and Mainardi, F. : Frac
tional calculus and stable probability distributions, Arch. Mechanics
50, 377-388 (1998). Online: http://citeseer.ist.psu.edu/
gorenflo98fractional.html .
[97] Mainardi, F., Paradisi, P., and Gorenflo, R., 1999. Probabil
ity distributions generated by fractional diffusion equations. F
Mainardi, P. Paradisi and R. Gorenflo, Probability distributions gen
erated by fractional diffusion equations, in: J. Kertesz and I. Kon
dor, eds., Econophysics: an Emerging Science (Kluwer, Dordrecht
1999) in press, pp. 39. Online: http://citeseer.ist.psu.edu/
mainardi99probability.html .
[98] Laakso, T. I., Valimaki, V., Karjalainen, M., and Laine, U. K., 1996
Splitting the unit delay [FIR/all pass filters design]. IEEE Signal Pro
cessing Magazine, 13 (1) , pp. 30 60.[99] Olkkonen, J. T., and Olkkonen, H., 2007. Fractional delay filter based on
the b-spline transform. IEEE Signal Processing Letters, 14 (2) , pp. 97
100.
[100] Pei, S.-C., and Wang, P.-H., 2004. Closed-form design of all-pass frac
tional delay filters. IEEE Signal Processing Letters, 11 (10) , pp. 788
791.
[101] Tseng, C.-C., 2006. Improved design of digital fractional-order differen
tiators using fractional sample delay. IEEE Transactions on Circuits and
Systems I: Regular Papers, 53 (1) , pp. 193 203.
18 Copyright c 2007 by ASME