Professional Documents
Culture Documents
Signal Processing
journal homepage: www.elsevier.com/locate/sigpro
Review
a r t i c l e i n f o abstract
Available online 12 August 2010 A look into fractional calculus and its applications from the signal processing point of
Keywords: view is done in this paper. A coherent approach to the fractional derivative is presented,
Fractional derivative leading to notions that are not only compatible with the classic but also constitute a
Grunwald-Letnikov derivative true generalization. This means that the classic are recovered when the fractional
Fractional linear systems domain is left. This happens in particular with the impulse response and transfer
Riemann-Liouville function. An interesting feature of the systems is the causality that the fractional
Caputo derivative imposes. The main properties of the derivatives and their representations are
presented. A brief and general study of the fractional linear systems is done, by showing
how to compute the impulse, step and frequency responses, how to test the stability
and how to insert the initial conditions. The practical realization problem is focussed
and it is shown how to perform the input–ouput computations. Some biomedical
applications are described.
& 2010 Elsevier B.V. All rights reserved.
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
2. Fractional derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
2.1. Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
2.2. Existence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
2.3. Main properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
2.3.1. Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
2.3.2. Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2.3.3. Scale change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2.3.4. Time reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2.3.5. Time shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2.3.6. Derivative of a product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2.4. Group structure of the fractional derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2.4.1. Additivity and commutativity of the orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2.4.2. Associativity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2.4.3. Neutral element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
n
Corresponding author.
E-mail addresses: rmagin@uic.edu (R. Magin), mdo@fct.unl.pt (M.D. Ortigueira), igor.podlubny@tuke.sk (I. Podlubny), jtrujill@ullmat.es (J. Trujillo).
#
Eurasip member.
0165-1684/$ - see front matter & 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.sigpro.2010.08.003
R. Magin et al. / Signal Processing 91 (2011) 350–371 351
1. Introduction modeling such a process [28,135]. Later they showed that the
same results could be obtained using as starting point the
Fractional calculus has been attracting the attention of Grünwald–Letnikov derivative [140]. This theory was up-
scientists and engineers from long time ago. During this dated recently [146]. With this approach a linear system
period the main applications involved the use of the so called theory can be formulated in a fashion very similar to the
fractional integral operators to obtain explicit solutions of classic, being effectively a generalization in the sense of
regular models. However, most of the mentioned develop- obtaining the classic results when the order becomes integer.
ment was done by mathematicians [10,18,22–24,26–28,30]. This theory will be revised here, taking into account recent
Since the nineties of the last century fractional calculus has developments. We will consider the associated problems:
been rediscovered and applied in an increasing number establishment of the initial conditions and the stability of the
of fields, namely in several areas of physics [19,20,25,33,50, systems. A self-contained theory suitable for dealing with
80-83,85,86,116–119,133,162,167,168], control engineering problems like filtering, modeling, and realization is intended
[25,29,36,37,53,107,114,115,132,148,151,171-176] and signal to be present here.
processing [38,39,82,83,121,148,160,169]. A complete theory As referred to above the number of applications has
of the linear systems of fractional differential equation with been increasing. One of the areas where this can be
constant or variable coefficients can be found in the literature verified is the biomedical [20,45,46,67,87]. Here we
[18,20,22,28,30]. On the other hand, we must remark that in describe some of the recent applications in this field.
80% of the papers that appear in the scientific literature, in The now classic fractional Brownian motion (fBm)
the framework of the fractional calculus and their applica- modeling is also considered, as an application of fractional
tions, the corresponding authors use different fractional calculus [21,104,120,144]. We define a fractional noise
differential operators but at the end they contrast their that is obtained through a fractional derivative of white
model using a numerical approach based in a finite number noise. The fBm is an integral of the fractional noise.
of terms of the series that define the known Grünwald– The paper is organized as follows. In Section 2, we
Letnikov derivative [30]. Then they obtain excellent results. present the Grünwald–Letnikov fractional derivative and
Therefore we can conclude that a generalization of the linear its main properties and relations with other fractional
systems of differential equation is very useful to be used in derivatives like the so called Riemann–Liouville derivative
352 R. Magin et al. / Signal Processing 91 (2011) 350–371
and the Caputo derivative. Some examples of derivative If y = p, we put h= 9h9 to obtain the backward
computations are shown. Practical implementations and Grünwald–Letnikov derivative:
simulation are also considered. The introduction of the
fractional linear systems is done in Section 3. We define
P1 k a
k ¼ 0 ð1Þ f ðz þ khÞ
transfer function and impulse response and show how to k
Dab f ðzÞ ¼ lim ejpa ð3Þ
compute it. Stability and establishment of correct initial h-0 þ ha
conditions are also studied. The continuous to discrete It is important to enhance an interesting fact—when a
conversion is considered in Section 3.5 and the general is a positive integer we obtain the classic expressions for
input/output computations will be studied in section 4. In the integer order derivatives.
Section 5 we describe some applications in biomedical
engineering. To finish, we study the fractional Brownian
motion and present some conclusions. 2.2. Existence
Remarks: 1—In this paper we deal with a multivalued
expression za. As is well known, to define a function we It is not a simple task to formulate the weakest
have to fix a branch cut line and choose a branch (Riemann conditions that ensure the existence of the fractional
surface). It is a common procedure to choose the negative derivatives (1)–(3), although we can give some necessary
real half-axis as branch cut line. Unless stated to the conditions for their existence. To study the existence
contrary, in what follows we will assume that we adopt conditions for the fractional derivatives we must care
the principal branch and assume that the obtained about the behavior of the function along the half straight
function is continuous above the branch cut line. line z 7nh with nAZ + . If the function is zero for Re(z)o
With this, we will write ( 1)a =ejap. 2—Unless otherwise aAR (correspondingly Re(z)4a) the forward (backward)
stated, we will assume to be in the context of the derivative exists at every finite point of f(z). In the general
generalized functions (distributions). We always assume case, we must have in mind the behavior of the binomial
that they are either of exponential order or tempered coefficients. They satisfy
distributions.
a A
k r ka þ 1
2. Fractional derivative
meaning that f(z)(A/ka + 1) must decrease, at least as
2.1. Definitions A/k9a9 + 1 when k goes to infinite. For example considering
the forward case, if a 40, it is enough that f(z) be bounded
Similarly to the classic case, the known Grünwald– in the left half plane, but if a o0, f(z) must decrease to
Letnikov definition of fractional differential equation is zero to obtain a convergent series. In particular, this
introduced. We here introduce the following modification suggests that Re(h) 40 and Re(h)o0 should be adopted
of the mentioned fractional derivative by the limit of the for right and left functions,2 respectively in agreement
fractional incremental ratio [146]: with Liouville reasoning [10]. In particular, they should be
used for the functions such that f(z)= 0 for Re(z) o0 and
P1 k a
k ¼ 0 ð1Þ f ðzkhÞ f(z)=0 for Re(z) 40, respectively.3 This is very interesting,
k
Day f ðzÞ ¼ ejya lim a ð1Þ since we conclude that the existence of the fractional
9h9-0 9h9 derivative depends only on what happens in one half
complex plane, left or right. Consider f(z)=zb, with bAR
a
where stands for the binomial coefficients and with a suitable branch cut line. If b 4 a, we conclude
k
jy
h= 9h9e is a complex number, with yA( p,p]. The above immediately that Da[zb] defined for every zAC does not
definition is valid for any order, real or complex [70]. In exist, unless a is a positive integer, because the summa-
order to understand and give an interpretation to the tion in (1) is divergent.
above formula, assume that z is time and that h is real,
y =0 or y = p. If y =0, only the present and past values are 2.3. Main properties
being used, while, if y = p, only the present and future
values are used. This means that if we look at (1) as a
We are going to present the main properties of the
linear system, the first case is causal, while the second is
derivative presented above.
anti-causal1 [140].
In general, if y = 0, we call (1) the forward Grünwald–
Letnikov derivative: 2.3.1. Linearity
The linearity property of the fractional derivative is
P1 k a
k¼0 ð1Þ f ðzkhÞ evident from the above formulae. In fact, we have
k
Daf f ðzÞ ¼ lim ð2Þ
h-0 þ ha Day f ðzÞ þ gðzÞ ¼ Day f ðzÞ þDay gðzÞ ð4Þ
2
We say that f(z) is a right [left] function if f( N) =0[f(+ N) = 0].
3
By breach of language we call them causal and anti-causal
1
We will return to this subject later. functions borrowing the system terminology.
R. Magin et al. / Signal Processing 91 (2011) 350–371 353
similar to the McLaurin series and can be useful in provided 9e sh9 o1, that is if Re(s) 40. This means that
computing the primitive of some functions. the branch cut line of F(s) must be in the left hand half of
the complex plane. Then
2.4. Group structure of the fractional derivative a
ð1esh Þa 1esh a
Daf f ðzÞ ¼ lim a ¼ lim esz ¼ 9s9 ejya esz
n-0 þ h h-0 þ h
2.4.1. Additivity and commutativity of the orders
ð13Þ
We are going to apply (1) twice for two orders. We
have [138] iff yA( p/2,p/2), which corresponds to working with the
principal branch of (.)a and assuming a branch cut line in
Day ½Dby f ðtÞ ¼ Dby ½Day f ðtÞ ¼ Day þ b f ðtÞ ð9Þ the left hand complex half plane.
354 R. Magin et al. / Signal Processing 91 (2011) 350–371
Now, consider the series in (3) with f(z)= esz. Proceed- valid for non-positive integer orders. In terms of linear
ing as above, we obtain another binomial series: system theory, (15) tells us that the fractional forward
X differintegrator (a current terminology) is a linear system
1
a ksh
ð1Þk e with impulse response equal to the right hand side in (19).
k
k¼1 We could use (3) and obtain the impulse response of the
that is convergent to the main branch of anti-causal differintegrator by starting with u( t). The
procedure is similar and the result is [135]
FðsÞ ¼ ð1esh Þa
t fa1g
provided Re(s) o0. This means that the branch cut line of Dab dðtÞ ¼ uðtÞ ð20Þ
Ga
F(s) must be in the right hand half complex plane. We
The impulse responses (19) and (20) of the causal and
obtain directly for f(z)= esz
anti-causal differintegrators have sa as transfer functions
a
Dab f ðzÞ ¼ 9s9 ejya esz with regions of convergence Re(s)40 and Re(s)o0,
respectively.
with yA(3p/2,p/2), and
a
Dab f ðzÞ ¼ 9s9 ejya esz 2.5.4. The power function
valid iff yA(p/2,3p/2). These results can be used to The general power function does not have fractional
generalize a well known property of the Laplace trans- derivative as is easy to observe from (1), because it
form. If we return back to Eq. (2) and apply the bilateral increases without bound as t goes to 7N. This does not
Laplace transform happen with the causal (or anti-causal power). The results
Z þ1 obtained in the above closed section allow us to obtain the
FðsÞ ¼ f ðtÞest dt ð14Þ derivative of tbu(t). In the sequence of computations in the
1 following we shall be assuming that the exponents in
to both sides, we conclude that the powers are not negative integers. Using (18) again, we
obtain
L½Daf f ðtÞ ¼ sa FðsÞ for ReðsÞ 40 ð15Þ
Gðb þ1Þ ba
a
where for s we assume the principal branch and a cut line Daf t b uðtÞ ¼ t uðtÞ ð21Þ
Gðba þ1Þ
in the left half plane. With Eq. (3) we obtain
which generalizes the usual formula for bAZ and beN .
L½Dab f ðtÞ ¼ sa FðsÞ for ReðsÞ o0 ð16Þ
Eq. (21) can be considered valid for b a = 1 provided
where now the branch cut line is in the right half plane. that we write
These results have a system interpretation: there are two t a1 uðtÞ
systems (differintegrators) with the same expression for Daf ¼ dðtÞ ð22Þ
GðaÞ
the transfer function H(s) =sa, but different regions of
convergence. One is causal and the other is anti-causal. To see that this is correct, we use (18) to obtain
Later we will compute the corresponding impulse re- t a1 uðtÞ
sponses. The s =jo case will be considered later also. Daf ¼ Daf D
f
a þ 1 uðtÞ ¼ D1 uðtÞ ¼ dðtÞ
f
GðaÞ
plane. It is the backward Liouville derivative obtained function with a jump. This leads to
from (16) and (20): Z t
ðbÞ 1 X
n1
Z f ðtÞðttÞa1 dt f ða1iÞ ðaÞd ðtÞ
ðiÞ
ð1Þa 1 fRL þ ðtÞ ¼
Dab f ðtÞ ¼ f ðt þ tÞta1 dt ð25Þ GðaÞ a i¼0
GðaÞ 0
ð32Þ
These definitions were introduced both exactly with
The appearance of the ‘‘initial conditions’’ f (a 1 i)(a+ )
this format by Liouville [10]. Unhappily in the common
provoked some confusions because they were used as
literature the factor ( 1) a in (25) has been removed
initial conditions of linear systems. This is not correct in
and is called the Weyl derivative [22,30]. Although the
general. They represent what we need to join to the
above results were obtained for functions with Laplace
Riemann–Liouville derivative to obtain the Liouville
transform their validity can be extended to other func-
derivative (28) [145]. Now let us do a similar analysis to
tions [18,30]. ðnÞ
the Caputo derivative. The expression f½f ðtÞ uðtaÞd þ ðtÞg
states the integer order derivative of the function
2.7. Riemann–Liouville and Caputo derivatives f(t)u(t a). The so called jump formula gives [35,137,147]
X
n1
ðiÞ
The Riemann–Liouville and Caputo derivatives are yðnÞ ðtÞuðtaÞ ¼ ½yðtÞuðtaÞðnÞ yðn1iÞ ðaÞd ðtÞ ð33Þ
multistep derivatives that use several integer order i¼0
derivatives and a fractional integration [18,20,22, which leads to
24–28,30]. To present them, we use (19) and (20) to
Z t X
n1
obtain the following distributions [139]: 1 ðinÞ
fCðbþÞ ðtÞ ¼ f ðtÞðttÞa1 dt f ðn1iÞ ðaÞd ðtÞ
GðaÞ a
nÞ t n1 i¼0
dð
7 ðtÞ ¼ 7 uð 7tÞ, 0 o n o1 ð26Þ ð34Þ
GðnÞ
In this case, we can extract conclusions similar to those
and
8 we did in the Riemann–Liouville case. Relation (34) explains
n1
<7t
>
uð 7 tÞ for n o0
why sometimes the first n terms of the Taylor series of f(t)
ðnÞ
d 7 ðtÞ G ðnÞ ð27Þ are subtracted from it before doing a fractional derivative
>
: dðnÞ ðtÞ computation. It is like a regularization.
for n Z0
where nAZ. With them we define two differintegrations 2.8. Fourier transform of the fractional derivative and the
usually classified as left and right sided, respectively: frequency response
ðnÞ
flðaÞ ðtÞ ¼ ½f ðtÞuðtaÞd þ ðtÞd þ ðtÞ
ðnÞ
ð28Þ
Now, we are going to see if the above results can be
ðnÞ extended to functions with a Fourier Transform. We note
frðaÞ ðtÞ ¼
ðnÞ
½f ðtÞuðbtÞd þ ðtÞd þ ðtÞ ð29Þ
that the multivalued expression F(s)=sa becomes an analytic
with aobAR. The orders are given by a = n n, n being the function, as soon as we fix a branch cut line, in all the
least integer greater than a and 0o n o1. In particular, if a complex plane excepting the branch cut line. Computation
is integer then n = 0.4 From different orders of commut- of the derivative of functions with Fourier Transform is
ability and associability in the double convolution we can dependent on the way used to define (jo)a. If we define it
obtain distinct formulations. For example, from (28) we doing the limit as s-jo from the right we have
obtain the left Riemann–Liouville and the Caputo deriva- (
jap=2
a a e if o 4 0
tives [139]: ðjoÞ ¼ 9o9 ð35Þ
n o ejap=2 if o o 0
ðbÞ ðnÞ ðnÞ
fRL þ ðtÞ ¼ d þ ðtÞ ½f ðtÞuðtaÞd þ ðtÞ ð30Þ This means that the forward derivatives of a cisoid is
n o given by
ðnÞ (
fCðbþÞ ðtÞ ¼ ½f ðtÞuðtaÞd þ ðtÞ d þ ðtÞ
ðnÞ
ð31Þ jap=2
if o 4 0
a e
Daf ejot ¼ ejot 9o9 ð36Þ
For the right the procedure is similar. We are going to ejap=2 if o o 0
study more carefully the characteristics of these deriva- For x(t) =cos(o0t) we obtain
ðnÞ
tives. Consider (28). Let jðnÞ ðtÞ ¼ f½f ðtÞ uðtaÞd þ ðtÞg. We
have Da cosðo
f 0 tÞ ¼ 9 o0 9a cosðo0 t þ ap=2Þ ð37Þ
8 Z t
> It can be show [146] that these results are not valid in
< 1 f ðtÞðttÞn1 dt if t 4 a,
ðnÞ
j ðtÞ ¼ GðnÞ a the backward case.
>
:0 if t o a
2.9. Modeling, identification and implementation
So, in general when doing the second convolution in
(30) we are computing the integer order derivative of a As in the usual systems, modeling, identification and
implementation are very interesting tasks. In the frac-
4
All the above formulae remain valid in the case of integer tional case, they are slightly more difficult due to the fact
integration, provided that we put d(0)(t) = d(t). of having, at least, one extra degree of freedom—the
356 R. Magin et al. / Signal Processing 91 (2011) 350–371
fractional order. However, this difficulty increments the constants. In order to convert this time-domain repre-
possibilities of obtaining more reliable and robust sys- sentation into a model for fractional operations we take
tems. This is challenging and people working in the area the Laplace transform of both sides of (39). As seen in
have been giving different interesting answers. We can Section 2.5
refer to the following approaches:
L½t a uðtÞ ¼ Gð1aÞsa1 ð40Þ
2.9.1. Fractional devices and assuming that we can interchange the order of
The famous Curie law stating that the current in an integration for v and t we obtain
insulator decreases proportionally to a negative power of Z 1 a1 Z
1 v sinðapÞ 1 va1
time leads to the known ‘‘supercapacitors’’, which have sa1 ¼ dv ¼ dv
GðaÞGð1aÞ 0 s þv p 0 s þv
impedance proportional to 1/sa, with 0 o a o1 [34,116].
ð41Þ
Electrochemists have used the Constant Phase Elements
(CPE) description for over 60 years. The fractors (fractional which is the Stieltjes transform of va 1/G(a)G(1 a).5
capacitors) [53,99,105,170] and coils [162] have been Finally, solving for sa and if we let v =1/t where t is the
presented. The new terminology is ‘‘fractance’’ to indicate relaxation time corresponding to a particular value of v
an impedance with fractional order response. As these we obtain
devices become available commercially, we will be Z 1 Z 1
sinðapÞ ts dt sinðapÞ ts
sa ¼ ta ¼ ta dðlog tÞ
rewriting many of the rules for design of filters and p 0 tsþ 1 t p 0 tsþ 1
controllers [29,36,37,53]. ð42Þ
Thus, in this interpretation, we see that the fractional
2.9.2. Trans-finite Circuits
derivative operator is represented as an integral or
Infinite transmission lines are circuits with fractional
summation of Laplace domain terms that correspond to
behavior [65], but there are other interesting circuits with
high-pass filters, and by a similar derivation the fractional
similar characteristics like the tree fractance ( a tree of RC
integral operator is expressed in terms of an integral of
circuits) and chain fractance (a series of parallel RC)
low-pass filters. This is a unifying hypothesis because it
circuits [66,101,156].
extends in a natural way the usual progression of
modeling linear systems as a series of exponentials, which
2.9.3. Band-limited approximations typically increases as the degree of the integer order
It is an engineering approach. There are several ways of transfer function grows. With the above formulation, we
doing the design and implementation. We can refer to (a) see that the poles are logarithmically distributed.
the CRONE, which uses the Bode diagrams [36–39,107,
149–151] and (b) the continued fraction approaches
3. Fractional linear systems
[171,172]. Both construct pole–zero systems with inter-
laced poles and zeros.
3.1. Transfer function and frequency response
Other similar alternative is approximation by a
weighted summation of exponentials, which as the
number of elements increases toward infinity describes The results of the previous section are very important
fractional behavior. This concept has more recently been in applications since they allow us to introduce the useful
used by Anastasio [45] to approximate fractional order concept of transfer function. In fact, we define a linear
operators in his analysis of the vestibulo-ocular system. system through a fractional differential equation with the
The basic idea developed by Thorson and Biederman- general format
Thorson [167] is to represent a power law relaxation X
N X
M
decay in time (e.g., t a, where 0o a o1) by a sum of an Dvn yðtÞ ¼ bm Dnm xðtÞ ð43Þ
n¼0 m¼0
exponentials weighted in an appropriate manner. Starting
with the integral definition of the gamma function where the differentiation orders, nn, are, in the general
Z 1 case, complex numbers. As usual, we apply the LT to
GðaÞ ¼ xa1 ex dx, a 40 ð38Þ Eq. (43) and use the results of Section 2.5, to obtain the
0
transfer function of the system:
if we let x= ta, where t40, we can solve for ta: PM
Z 1 bm snm
1 HðsÞ ¼ PmN ¼ 0 ð44Þ
t a ¼ va1 evt dv ð39Þ n ¼ 0 an s
nn
GðaÞ 0
This integral can be interpreted as the Laplace trans- with region of convergence defined by Re(s)40 (causal
form of the function va 1/G(a). Hence, we see that (39) case) or Re(s)o0 (anti-causal case).
provides a representation for the power-law decay as a We may put the question of what happens with the
weighted integral of exponentials. Thus, between the frequency response of a given fractional linear system.
values of v and v +dv there exists an exponential e vtu(v) From the conclusions we presented in 2.8, we can say
with a weight, va 1/G(k). Here v has the units of (s) 1, that, having a causal fractional linear system with transfer
and can be viewed as a rate constant. The overall power
law relaxation given by (39) is the summation of all these 5
To obtain the last expression, we used the reflection property of
contributions for the entire range of possible rate the gamma function [14].
R. Magin et al. / Signal Processing 91 (2011) 350–371 357
function equal to H(s), the frequency response must be For this reason, we can invert this series term by term, to
computed from obtain
HðjoÞ ¼ lim HðsÞ ð45Þ X
s-jo
1
an t nn
f ðtÞ ¼ t n1 uðtÞ ð51Þ
This is in agreement with other known results. For n¼0
Gðnn þ nÞ
example, if the input to the system is white noise, with
unit power, the output spectrum is given by which is a special case of the two parameter Mittag–
SðoÞ ¼ lim HðsÞHðsÞ ð46Þ Leffler function, which is a generalization of the expo-
s-jo
nential, to which it reduces when n =1. This function is
well studied {see [18,20,28,65]}.6 Eq. (51) suggests us to
3.2. From the transfer function to the impulse response work with the step response instead of the impulse
response to avoid derivatives or working with non-regular
The general case represented in (43) is not easy to functions near the origin.
solve, because it is difficult to find the poles. For this When n =1/q, q being a positive integer, we obtain a
reason, in the following, we shall be restricting our different formulation for the inverse of the partial fraction
attention to the cases in which (49). Using the well known result of the sum of the first q
terms of a geometric sequence we obtain7 [22]:
the nn are irrational numbers but multiples of a given Pq
1 j¼1
aj1 s1jv
n, FðsÞ ¼ v ¼ ð52Þ
S a saq
the nn are any rational numbers; in this case, write
them in the format pn/qn. We conclude that the inverse LT of a partial fraction as
F(s)=1/(s1/q a) is a linear combination of q fractional
q
Let n be the greater common divider of the nn. Then derivatives of E0 ðt,aq Þ ¼ ea t uðtÞ.
nn = nn. We will assume that n o2, for stability reasons. The k41 case in (45) does not present great difficulties
With this formulation, Eqs. (43) and (44) assume the except some additional work. It can be obtained from the
general formats k= 1 case by repeated convolution or by differentiation
[135].
X
N X
M
an Dnn yðtÞ ¼ bm Dmn xðtÞ ð47Þ
n¼0 m¼0 3.3. The stability problem
and
PM The study of stability of the fractional linear time
bm smn
HðsÞ ¼ PmN¼ 0 ð48Þ invariant (FLTI) systems we are going to do is based on the
nn
n ¼ 0 an s BIBO stability criterion, which implies stability when the
With a Transfer Function as in (48) we can perform the impulse response is absolutely integrable.
inversion quite easily, by the following steps: The simplest FLTI system is the system with transfer
function H(s) =sn with s belonging to the principal
(1) Transform H(s) into H(z), by substitution of sv for z. We Riemann surface. If n 4 0, the system is definitely
are assuming that H(z) is a proper fraction; otherwise, unstable, since the impulse response is not absolutely
we have to decompose it as a sum of a polynomials integrable, even in a finite interval. If 1o n o0, the
(inverted separately) and a proper fraction. impulse response remains a limited function when t
(2) The denominator polynomial in H(z) is the indicial increases indefinitely and it is absolutely integrable in
polynomial [59,60] or characteristic pseudo-polyno- every finite interval. Therefore, we will say that the
mial [22]. Perform the expansion of H(z) in partial system is in a wide sense stable. This case is interesting to
fractions. the study of fractional stochastic processes. If n = 1, the
(3) Substitute back sv for z, to obtain the partial fractions normal integrator, the system is in a wide sense stable.
in the form The case n o 1 corresponds to an unstable system, since
1 the impulse response is not a limited function when t goes
FðsÞ ¼ , k ¼ 1,2,. . . ð49Þ to +N.
ðsn aÞk
(5) Invert each partial fraction.
6
(6) Add the different partial impulse responses. An interesting implementation was done by Prof. Podlubny and
can be found at the site of MatLab. It is an implementation of the two
parameter generalized Mittag–Leffler function with precision control—
We are going to see how to invert F(s)= 1/(sn a) Using usage: mlf(alfa, beta, z, p) [153].
7
the properties of the geometric series, it is a simple task to With reason r =b/x, we obtain
obtain q1 q1 q
X 1rq X 1bq xq X
X
1 rj ) bj xj ¼ or xq bq ¼ ðxbÞ bj1 xqj
1r 1b=x
FðsÞ ¼ sn an snn ð50Þ j¼0 j¼0 j¼1
Consider the LTI systems with transfer function H(s) a 3.5. Discrete-time implementations
quotient of two polynomials in sn. The transformation
w= zq transforms the sector 0 r y r2p/q {y = arg(z)} to the It is not a simple task to obtain discrete-time
entire complex plane. So, the sector (p/2q)r y r(p/2q)+ implementation of a fractional differintegrator. There are
(p/q) is transformed to the left half plane. Consider the several algorithms that start from an s to z conversion and
first Riemann surface of z= sn defined by y =arg(s)A( p,p]. design an ARMA model [48,49,64,115,141,143,161,
This domain is transformed to j = arg(z)A( pa,pa]. 173,175]. However, they are mere approximations and
However the poles leading to instability must be inside there is no clear statement on the optimality of any
the sector ( pa/2,pa/2). We have two situations leading approch. It is an open subject needing additional research
to stability: efforts. The simplest way of doing such approximation
consists in starting from the forward GL derivative,
there are no poles inside the sector ( pa,pa] and remove the limit operation and truncate the series. This
there are poles but they are in the sectors ( pa, pa/ is not needed if we intend to compute the output of the
2) and (pa/2,pa). system being approximated for a causal imput. In fact, in
this case, the series becomes a finite sum:
The poles with argument equal to 7 pa/2 may lead to
Pbt=hc a
wide sense stable systems as in the usual systems. These k¼0
ð1Þk f ðtkhÞ
k
conclusions come from properties of the Mittag–Leffler Day f ðtÞ a ð54Þ
h
function [18]. To give a simple example, consider the
transfer function H(s)= 1/(sa +1), with 0 o a o2. It is easy In ths situation we must consider the so called
to see that there is no pole in the principal Riemann ‘‘short memory principle’’ [28], also known as ‘‘fixed
surface. Hence, it represents a stable system. memory principle’’, which is a useful tool for numerical
simulations in large time intervals. Taking into account
approximation (54) we can see, that if t b0 the final
3.4. The initial conditions sumation would be too large. From the calculation of
binomial coefficients above it follows that the past values
When looking for the output, y(t), to a given input, x(t), of the function f(t) near 0 have only small influence on the
we must consider the initial conditions. This is a problem new evaluated value of the function. Instead of using
that created much confusion and difficulties in the past ’’whole memory’’, only ’’recent past’’ of the function is
[22,96] motivated by the use of several different deriva- used, e.g. the interval (t T,t), where T is the ’’memory
tive definitions and of the one-sided Laplace transform. In length’’:
[137,147], we proposed a new way of looking at the
problem.
PbðtTÞ=hc k a
k¼0
ð1Þ f ðtkhÞ
As is well known, the solution of Eq. (47) has two k
Day;T f ðzÞ ð55Þ
terms: the forced (or evoked) and free (or spontaneous). ha
This second term depends only on the state of the system It is worth mentioning, that a similar approach was
at the reference. This state constitutes, or is related to, the introduced in Volterra’s work under the name ’’limited
initial conditions. These are the values at t =0 of variables after-effect’’ assumption [28].
in the system, which are associated with stored energy. It This continuous to discrete conversion is essentially
is the structure of the system that imposes the initial the following. Assume that h is a sampling interval. Then
conditions, not the eventual way of computing the deriva- we can also sample f(t) and Day f ðtÞ with the same interval.
tives. The solution is obtained with the fractional jump This is equivalent to performing the continuous to
formula [147] discrete (Euler) transformation:
X
n1
½ðnmÞg1
1z1
Dna ½yðtÞuðtÞ ¼ Dna ½yðtÞuðtÞ yðmgÞ d ðtÞ s¼
h
ð56Þ
0
and
that allows us to transform (47) to
PN a k
X
N 1
a k¼0 ð1Þk z
ðgiÞ 1z k
aj ½yðtÞuðtÞ sa ¼ ð57Þ
i¼0 h ha
X
M X
N X
i1
½ðnmÞg1
¼ bi ½xðtÞuðtÞðgiÞ þ ai yðgmÞ ð0Þd ðtÞ
i¼0 i¼1 0 where N is a ‘‘high enough’’ integer fixed according to the
XM
½ðnmÞg1
principle stated above. As can be seen we are doing a FIR
bi xðgmÞ ð0Þd ðtÞ ð53Þ approximation to the differintegrator—the impulse re-
i¼1
sponse is h(n)= ((( a)n)/n!)u(n) for n =0, 1, y, N. Using
We must note that (48) it is possible to obtain ARMA models for the
same operator {see [48,49,141,143]}. With these s to z
the initial conditions appear directly in the equation, transformations we arrive at the discrete-time signal
without using any transform and processing context and so obtain an easier and more
Eq. (53) is valid also in the time variant case. known framework.
R. Magin et al. / Signal Processing 91 (2011) 350–371 359
Other alternatives to the Euler transformation are the tions in the fractional calculus, and especially solving
bilinear transformation (Tustin) [62,143,171–176] fractional differential equations.
According to what we said above, the fractional
2 1z1
s¼ ð58Þ derivatives of order a can be approximated at all nodes
h 1 þz1
of the uniform grid t= nh, nAZ at once with the help of the
and the mixed operator (Al-Alaoui) [44,62] upper triangular strip matrix BðnaÞ :
h iT T
8 1z1 vnðaÞ ðaÞ
vn1 ... v1ðaÞ v0ðaÞ ¼ BnðaÞ vn vn1 ... v1 v0
s¼ ð59Þ
h 7 þz1
ð62Þ
As we want to obtain discrete equivalents to the
differintegrator, sa, the following considerations have to where
be mentioned [64]: nnðaÞ ¼ Da nðnhÞ ð63Þ
and
(1) sa, viewed as a causal operator, has a branch cut line ðaÞ
along the negative real axis for arguments of s in
h
0 hð1aÞ & & ðaÞ
hn1 hnðaÞ
ðaÞ
( p,p) but is free of poles and zeros. 0
hð0aÞ h1ðaÞ & & hn1
(2) A dense interlacing of simple poles and zeros along a
ðaÞ 1 0 0 h0ðaÞ hð1aÞ & &
line in the s plane is, in some way, equivalent to a Bn ¼ a ð64Þ
t ... ... ... & & &
branch cut (see the deduction of the Cauchy derivative).
(3) It is well known that, for interpolation or evaluation
0
... 0 0 h0ðaÞ h1ðaÞ
purposes, rational functions are sometimes superior 0 0 ... 0 0 h0ðaÞ
to polynomials, roughly speaking, because of their
ability to model functions with zeros and poles. In where hnðaÞ represents the impulse response according to
other words, for evaluation purposes, ARMA models the chozen method (Euler, Tustin or Al-Alaoui). From the
converge faster than the long MA (FIR). properties of the derivative those matrices BnðaÞ have also
(4) Trapezoidal (bilinear) rule maps adequately the the group structure presented in Section 2.4. This method
stability regions of the s plane on the z plane, and has been presented for the first time in [155] along with
maps the points s= 0 and s= N to the points z= 1 and several examples of numerical solution of ordinary
z = 1, respectively. fractional differential equations with the Riemann–
Liouvile and Caputo derivatives. From the viewpoint of
simplicity of usage, the matrix approach to numerical
The impulse response of the discrete-time linear
solution of fractional differential equations can be
system corresponding to the Tustin transformation is
compared to the Laplace transform method for solving
given by [143]
ordinary differential equations. Indeed, in both cases
a
2 ð1Þn ðaÞn X n
ðaÞk ðnÞk ð1Þk operators of differentiation are simply replaced with
hbil ðnÞ ¼ other symbols—in the case of the Laplace transform by
h n! k¼0
ðan þ 1Þk k!
a n
powers of the Laplace variable, and in the case of the
2 ð1Þ ðaÞn matrix approach by matrices of a known structure. For
¼ 2 F 1 ða,n,an þ 1,1ÞuðnÞ
h n! example, the famous Bagley–Torvik equation (see [28]
ð60Þ and references therein)
where 2F1(a,b,c, 1) is the Gauss hypergeometric function ay00 ðtÞ þ b0 Dt
ð3=2Þ
yðtÞ þ cyðtÞ ¼ f ðtÞ ð65Þ
that, for these arguments, does not have a closed form.
Similarly, the impulse response corresponding to the is discretized on a uniform grid with n nodes as
Al-Alaoui transformation can be computed following the ðaBð2Þ
ð3=2Þ
þ cBð0Þ
n þ bBn n ÞYn ¼ Fn ð66Þ
procedure used in [143] and is given by
where Yn is the vector of unknown values of y(t) at the
8 a ð7Þn ðaÞn Xn
ðaÞk ðnÞk ð7Þk
hbil ðnÞ ¼ uðnÞ discretization nodes, Fn the vector of values of the input
7h n! ð an þ 1Þk k!
a
k¼0 f(t) and BnðaÞ the triangular strip matrix representing the
n
8 ð1Þ ðaÞn 1 discrete analogue of the fractional derivative. Clearly, B0n is
¼ ð7Þn 2 F 1 ða,n,an þ 1,7 ÞuðnÞ
7h n! equal to the identity matrix En. The matrix approach has
ð61Þ been implemented in the form of a publicly available
It is interesting because it decreases quickly. With the Matlab toolbox.
above impulse responses, we can obtain ARMA models.
There are several methods, like the least-squares method 4. Input–output numerical computations in general
[48,49,141,143] and the continued fraction method [62–64] nonlinear systems
Another different way of doing the continuous to discrete
conversion is the so called matrix approach. The ‘‘matrix In various applications, e.g. in fluid mechanics, viscoe-
approach’’ to discretization of fractional integrals and lasticity, biology, physics and engineering [12,22,24,
derivatives has been developed by Podlubny [154,155]. It 26–28,50,61,73,97,98,168], considerable attention is gi-
is based on the use of triangular strip matrices. This method ven to ordinary and partial differential equations of
significantly simplifies many aspects of numerical computa- fractional order, due to their memory and hereditary
360 R. Magin et al. / Signal Processing 91 (2011) 350–371
properties [54–58,104,107,109]. However, most applica- m =0, 1, 2, y, N, which we will use in (71). This yields
tions have been directed toward modeling existing a Z 1
tm t a
situations where no outside interference has place. This ta1 yðtm tm tÞ dt ¼ f ½tm ,yðtm Þ þ yð0Þ m uðtm Þ
GðaÞ 0 Gð1aÞ
means that most studies consider the output of fractional
ð70Þ
systems under non-null initial conditions, but with null
input. This will be adopted here. According, for example, where we have also taken the input equal to zero. In this
to Momani [122–131] most fractional differential equa- relation we replace the integral by the quadrature formula
tions do not have exact analytic solutions, so approxima- Qm, and additionaly introduce the quadrature error Rm.
tion and numerical techniques [51,71–78,27] must be Thus, using the abbreviation gm(t)= y(tm tmt) yields
used. There are two main classes of methods for solving !
a
tm Xm a
yð0Þtm
fractional differential equations: the frequency-domain o g ðk=mÞ þ Rm ½gm ¼ f ðtm ,yðtm ÞÞ
methods and the time-domain methods. In this review we GðaÞ k ¼ 0 km m Gð1aÞ
deal mainly with time-domain methods. Finding accurate ð71Þ
and efficient methods for solving fractional-order differ-
where for okm we can write
ential equations (FODEs) is the goal of many research
works. Analytical and numerical methods for solving most o~ km Gð2aÞ
okm ¼ ð72Þ
of the FODEs must be used, as exact solutions cannot be að1aÞma
found easily. Some numerical methods for solving FODEs We finally solve the left-hand side for ym = y(tm) and get
were presented for instance in [52,71–78,106]. We will
X
m
ha tm
a
consider DE with the format ym ¼ ha f ðtm ,ym Þ o~ km yðtm khÞyð0Þ uðtm Þ
k¼1
Gð1aÞ
yðaÞ ðtÞ ¼ f ½t,yðtÞ þ axðtÞ ð67Þ
ð73Þ
where x(t) is the input, y(t) the output and tAR. Unless where the weights o
~ km are given by the substitution (74)
expressed we will assume that the derivative operation is and so
one step. This is important because we will need only one 8
initial condition, according to the results in Section 3.4. > 1 for k ¼ 0
>
>
>
> a for k ¼ m ¼ 1
We can then modify the above equation to make the >
<
o~ km 1a
initial condition appear explicitly in an equation that is ¼ 2 2 for k ¼ 1 and mZ2
Gð2aÞ >> 1a
valid for t Z0: >
> ðk1Þ
> þ ðk þ 1Þ1a 2k1a for 2 rk r m1
>
: ðk1Þ1a þ ða1Þka k1a for k ¼ mZ 1
ða1Þ
yðaÞ ðtÞ ¼ yð0Þd ðtÞ þ f ½t,yðtÞ þ axðtÞ, t 2 Rþ ð74Þ
or
t a
yðaÞ ðtÞ ¼ f ½t,yðtÞ þ axðtÞ þ yð0Þ uðtÞ, t 2 Rþ ð68Þ 4.2. Lubich’s difference methods
Gða þ 1Þ
It is a current practice to use here integral formulations. Lubich’s fractional difference methods form a subset to
We will use the Liouville integral seen in Section 2.6. fractional linear multistep methods, which were first
In the following we will present several approaches for presented by Lubich [111–113] and numerically imple-
solving Eq. (68), namely Diethelm’s method based on mented by Hairer, Lubich and Schlichte in [88] for a
quadrature and Lubich’s difference methods followed special type of Volterra integral equations. Consider again
with some information about Adams–Bashforth–Moulton that the input, x(t), is null. It can be shown that the FODE
method based on the Volterra integral equation, an can be rewritten as the Abel–Volterra integral equation
effective method for fractional order dynamical systems, Z t
Adomian’s decomposition method (ADM), variational yðtÞ ¼ ðttÞa1 f ½t,yðtÞ dt þyð0ÞuðtÞ ð75Þ
iteration method (VIM) and homotopy–perturbation 0
method (HPM). Lubich [112] showed that, if a 40 and given the method
order, pA{1,2,3,4,5,6}
dm y da y
a a by ¼ f ðtÞ, m1 o a r m ð94Þ 4.5. Homotopy–perturbation method (HPM)
dt m dt
subject to the initial conditions The HPM is a combination of the traditional perturba-
ðjÞ tion method and homotopy in topology. It solves the
y ð0Þ ¼ cj , j ¼ 1,0,. . .,m1 ð95Þ
FODEs by decomposing the complex problem to simple
where cj, j= 0,1, y, m 1 are arbitrary constants and y(t) is problems, and then the perturbation equation can be
a causal function of time. The system represented by (94) easily constructed by a homotopy in topology.
can be interpreted as composite fractional relaxation/ In the works of Momani and Odibat linear and
oscillation equation for the cases {0 o a r1, m=1} and nonlinear partial FODEs were solved [123–131,129,133]
{1 o a r2, m =2}, respectively. using this method. The problem and solution proposed in
Let J= D 1, the anti-derivative operator. If we apply the [40–42] can be written in the form
operator Jm to both sides of (94) and we use the initial Da1 y1 ðtÞ ¼ f1 ðt,y1 ,y2 ,. . .,yn Þ
conditions, we get Da2 y2 ðtÞ ¼ f2 ðt,y1 ,y2 ,. . .,yn Þ
uðtÞ ¼ f1 ðtÞ þaf2 ðtÞ þJ m f ðtÞ þ½aJ ma þ bJm uðtÞ ð96Þ ^
Dan yn ðtÞ ¼ f1 ðt,y1 ,y2 ,. . .,yn Þ ð102Þ
where
subject to the following initial conditions
X
m 1
ti X
m1
t ma þ i
f1 ðtÞ ¼ ci , f2 ðtÞ ¼ ci ð97Þ yk ð0Þ ¼ ck , k ¼ 1,2,. . .,n ð103Þ
i¼1
i! i¼1
Gðm a þ iþ 1Þ
where Dai
is the fractional derivative of yi of order ai,
According to Adomian [1,43] the solution y(t) can be where 0 o aI r1 and fi are arbitrary linear or nonlinear
decomposed to an infinite series of components: functions. The following homotopy can be constructed in
view of the HPM [91,93] as
X
1
yðtÞ ¼ yn ðtÞ ð98Þ Dai yi ¼ pfi ðt,y1 ,y2 ,. . .,yn Þ ð104Þ
n¼0
where i= 1, 2, y, n and p is an embedding parameter that
After substitution of the decomposition series (101)
changes from zero to unity [42]:
into both sides of (96) we obtain
X
1 X
1
If p =0, we will obtain the linear equation
yn ðtÞ ¼ f1 ðtÞ þ af2 ðtÞ þ Jm f ðtÞ þ ½aJ ma þbJ m yn ðtÞ
n¼0 n¼0 Dai yi ¼ 0
ð99Þ If p = 1, the homotopy (104) turns out to be the original
The iterates can be obtained from the previous system given in (102).
equation by the following recursive way:
The solution of system (104) can be expanded using
y0 ¼ f1 ðtÞ þ af2 ðtÞ þJ m f ðtÞ,
ma m ma m m
the parameter p
y1 ¼ ðaJ þ bJ Þy0 ¼ ðaJ þ bJ Þ½f1 ðtÞ þ af2 ðtÞ þ J f ðtÞ,
yi ðtÞ ¼ yi0 þpyi1 þp2 yi2 þ p3 yi3 þ ð105Þ
ma ma
y2 ¼ ðaJ m
þ bJ Þy1 ¼ ðaJ þ bJm Þ2 ½f1 ðtÞ þ af2 ðtÞ þJ m f ðtÞ,
Series of linear equations can be obtained after
^ substitution of (105) in (104) and collecting the terms
yk ¼ ðaJ ma þ bJ m Þyk1 ¼ ðaJ ma þ bJ m Þk ½f1 ðtÞ þ af2 ðtÞ þ J m f ðtÞ: with the same powers of p, in the form [42]
The components of y(t) are then defined as p0 : Dai yi0 ¼ 0
X
1 p1 : Dai yi1 ¼ fi1 ðt,y10 ,y20 ,. . .,yn0 Þ
yðtÞ ¼ ðaJ ma þ bJm Þk ½f1 ðtÞ þaf2 ðtÞ þ J m f ðtÞ ð100Þ
k¼0
p2 : Dai yi2 ¼ fi2 ðt,y10 ,y20 ,. . .,yn0 y11 ,y21 ,. . .,yn1 Þ
p3 : Dai yi3 ¼ fi3 ðt,y10 ,y20 ,. . .,yn0 ,y11 ,y21 ,. . .,yn1 ,y12 ,y22 ,. . .,yn2 Þ
To obtain the solution of (96) in a series form we
expand the operator in (100) using the binomial formula. ^ ð106Þ
R. Magin et al. / Signal Processing 91 (2011) 350–371 363
where the functions fi1, fi2,y, satisfy the following biophysics and physiology is directed toward linking
equation: molecular processes with whole organ (brain, heart, and
muscle) function by developing muliscale models that
fi ðt,y10 ,py11 þ p2 y12 þ , ,yn0 þ pyn1 þ p2 yn2 þ Þ
span the intermediate levels of structure (e.g., from the
¼ fi1 ðt,y10 ,y20 ,. . .,yn0 Þ þpfi2 ðt,y10 ,y20 ,. . .,yn0 ,y11 ,y21 ,. . .,yn1 Þ
centimeter dimensions of gross anatomy down to the
þ p2 fi3 ðt,y10 ,y20 ,. . .,yn0 ,y11 ,y21 ,. . .,yn1 ,y12 ,y22 ,. . .,yn2 Þ þ. . .:
submicron resolution of histology).
ð107Þ In building multiscale models one can either try to use
The following conclusions were made in the work of as much of the available anatomical and histological
Abdulaziz et al. [42]: knowledge as possible – building a highly complex
It is obvious that these linear equations can be easily structures with hundreds of components (organelles,
solved by applying the operator J ai , i.e., the inverse of the membranes, cells, extracellular matrix, etc.) – or try to deal
fractional operator Dai . Hence, the components yik, empirically with the complexity by developing whole
k= 0,1,2, y, of the HPM solution can be determined. That system descriptions (e.g., linear, non-linear, deterministic,
is, by setting p= 1 in (107) we can determine the entire or stochastic models) with embedded chaotic or fractal
HPM series solutions measures (fractal dimensions, Lyapunov exponents, non-
X
1 Gaussian probability distributions) that capture important
yi ðtÞ ¼ yik ðtÞ ð108Þ features of the observed behavior [2,31]. A diagram
k¼0 illustrating some of the relationships between these
The HPM series solution (106) can be approximated by approaches is shown in Fig. 1. In this figure the models
the following N-term truncated series: are characterized on the X-axis by their degree of linearity
X
N1 and on the Y-axis with respect to their deterministic nature.
fiN ðtÞ ¼ yik ðtÞ: ð109Þ Linear time-invariant causal (LTIC) system models cluster in
k¼0 the first quadrant, while stochastic, probabilistic models fall
According to the authors using this method, HPM in the fourth quadrant [4]. In this representation the
yields rapid convergence of the solution series in most methods of fractional calculus (linear, deterministic, but
cases, usually only a few iterations, leading to very non-integer order) bundle together in Fig. 1 within the relm
accurate solutions. of LTIC system models, where they interpolate between the
conventional integer order differential operators and
5. Biomedical applications extend the dynamics to fractional order [33].
Deterministic
Chaos LTIC
4
3
2
β
Hodgkin-Huxley
1
Chua-Hartley 0
0 1 2 3 4
α
Non-Linear Linear
Fractals
Turbulence
Brownian Motion
Random
Fig. 1. Illustration of the relationship between the principal types of models used to describe complex systems. For conventional linear time-invariant
causal (LTIC) models the governing differential equations take on only integer order.
comparison with ZD, then the tissue–electrode equivalent to construct similar fractional order dynamic models for
circuit reduces to a resistor in series with ZD, which can be the biomechanical properties of tissues [19]. For example,
approximated by two constant phase elements in series. Craiem and Armentano [67] have modeled the elastic
Thus, in the Laplace domain, the overall impedance can be properties of the aorta, in vivo in a Merino sheep, using a
written as fractional order generalization of the relationship be-
vðsÞ 1 1 tween stress s(t) and strain e(t). Their generalized Voigt
zðsÞ ¼ ¼ Rþ a þ b ð110Þ model consists of a spring in parallel with two ‘‘spring-
iðsÞ s Ca s Cb
pots’’ of fractional order a and b. The governing fractional
The corresponding impedance plane plot for (110) is order differential equation is
shown in Fig. 5 for the simple case of a =1/2 and b = 1. Such
plots match the data measured in experimental studies of db eðtÞ db eðtÞ
sðtÞ ¼ E0 eðtÞ þ Z1 a þ Z2 ð113Þ
Ovadia and Zavitz [152]. The transient voltage response of dt dt
this circuit to a step in applied current, such as the leading where E0 is the elastic constant for a spring, and Z1 and Z2
edge of a pacemaker pulse, is described in the time- represent the viscosities of two springpots in parallel with the
domain by spring. From this equation the complex modulus E*(o) can be
defined for sinusoidal signals as the ratio of stress to strain by
I0 t a I0 t b
VðtÞ ¼ I0 R þ þ ð111Þ
Ca Gða þ 1Þ Cb Gðb þ1Þ sðoÞ
E ðoÞ ¼ ¼ E0 þ Z1 ðjoÞa þ Z2 ðjoÞb ð114Þ
eðoÞ
which gives a power law response that corresponds
to that observed in heart stimulation experiments by The real part of E*(o) is defined as the storage modulus
Greatbatch and Chardack [87]. and the imaginary part of E*(o) is the loss or dissipation
Thus we observe that the basic cardiac tissue-electrode modulus. The storage modulus characterizes the elastic
impedance can be represented by a series combination of property of the arterial wall while the loss modulus
a resistor and two fractional lumped circuit elements. The describes the tissue’s ability to absorb energy. Both
overall transfer function for this model corresponds to the properties change with frequency and govern pulsatile
following fractional order differential equation: oscillations of the vessel walls in health and disease. This
model was found by Craiem and Armentano to give a
da VðtÞ da IðtÞ Ca dab IðtÞ better fit to in vivo data recorded from 2 to 30 Hz than a
Ca ¼ RCa þIðtÞ þ ð112Þ
dt a dta Cb dt ab
if we assume a 4 b.
We can use the correspondence between RC electric
circuits and viscoelastic networks of springs and dashpots
Fig. 3. A drawing of the tissue–electrode interface between cardiac Fig. 5. Impedance plane plot for two constant phase element impe-
muscle cells and an implanted electrode. (Redrawn from Ovadia and dances in series with a resistor. In this example, we set R= Ca = Cb = 1 and
Zavitz [152]). a = 1/2, b = 1.
Fig. 4. Tissue–electrode circuit model. RB is the bulk tissue resistance, Ra1 and Ra2 are electrode access resistances, y is the charge transfer resistance, C is
the dipole layer capacitance and ZD is the fractional Warburg impedance.
366 R. Magin et al. / Signal Processing 91 (2011) 350–371
Voigt model (single spring in parallel with a dashpot) or a such that the modulus and phase can be written as
fractional Voigt (single spring in parallel with single
9G ðoÞ9 ¼ rog =a20 ð1 þ w2 Þ, y ¼ tan1 ðG1 =Gd Þ ¼ py ð117Þ
springpot). A vector plot in the complex plane of the
complex modulus for this study is shown in Fig. 6. where g = 2 2y. The advantage of this model is that it
In particular, the model (114) captures the changes does not specify a particular Maxwell, Voigt or Kelvin
that arise in vessel wall elasticity when a vascular rheological model, but simply assumes an underlying
constriction is induced by the local administration of fractional order dynamics, oy, and then estimates the
phenylephrine [67]. The authors conclude that the a fractional power law parameters y and a0 from the MRE
springpot appears to describe the stretching of the elastic data. Sinkus first verifies this model for a tissue mimicking
fibers of the aorta (a is close to zero), while the b breast phantom at a fixed frequency of 65 Hz and then
springpot seems to represent a structural viscous beha- applies the model to human breast tissue by measuring
vior (b closer to 1). As expected the elastic contribution the dynamic modulus at 65, 75, 85 and 100 Hz. A complex
increases—a decreases from 0.20 to 0.11—following plane plot of Gd and Gl gives a straight line with a y value
administration of phenylephrine while the loss term is of approximately 0.13 for normal tissue. Analysis of 39
relatively unchanged (0.84–0.80). Thus, for a complex malignant and 29 benign tumors using this method gives
multi-scale tissue such as the arterial wall, the fractional a clear separation of the tumors from the normal (and
order model is able to characterize the important features fibrotic) breast tissue, and furthermore separates the
of its dynamic behavior. malignant from the benign tumors when individual cases
Fractional order models have also been used by Sinkus are plotted in a graph (Fig. 7) of y versus a0 (an increase in
et al. [164] to fit magnetic resonance elastography (MRE) specificity of about 20% at 100% sensitivity). In earlier
data from breast tumors. In this technique, MRI is used to studies this group was not able to classify breast tumors
image low frequency (50–1500 Hz) shear wave oscilla- on the basis of Gd and Gl alone, so this model provides a
tions in the breast. Wavelength and attenuation of the significant improvement in cancer detection.
vibrations directly reflect elastic shear modulus and In the three examples considered here, fractional order
viscosity of the tissue through the complex wave vector models were found to provide better fits to electrical
k(o)= b(o)+ ja(o). In MRE these tissue properties are and mechanical measurements made on living tissue.
mapped onto an elastogram image through an assumed Such studies need replication, but these findings provide
model of the tissue’s mechanical properties—usually a useful examples of cases where an extension of the
purely elastic spring with zero loss, or a Voigt spring/ ‘‘standard’’ integer order dynamic models of circuits and
dashpot model. In his study, Sinkus assumed a power law mechanical systems is warranted. Fractional order dy-
increase in attenuation with excitation frequency, namic models of complex, multiscale systems account for
a(o)= a0oy (where 0 oyo1), and invoked causality via anomalous dynamic behavior through a simple extension
the Hilbert transform to obtain the propagation constant of the order of the operations from integer to fractional. In
as b(o)= tan(py/2)a0oy. Thus, for the time-domain this extension is manifest through
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi incorporation to a variable degree of system memory
kðoÞ ¼ a0 oy ejp=2 1 þ ðtanðpy=2Þ2 Þ ð115Þ through convolution with a power law kernel exhibiting
fading memory of the past. Perhaps, in the future, the
k(o) is related to the complex shear modulus G*(o)
development of integrated space and time-domain frac-
through
tional order models will become a more standard
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
kðoÞ ¼ o r=G ðoÞ, G ðoÞ ¼ 9G ðoÞejy 9 ð116Þ component of linear systems analysis, at least when such
0.50
PHE
Dissipation modulus
0.40
0.30
CTL
0.20
0.10
0.00
0.95 1.05 1.15 1.25 1.35 1.45 1.55 1.65
Storage modulus
Fig. 6. Vector diagram (complex plane plot) of (114) for in vivo modulus
data from an aorta under control (CTL) conditions and following
application of a vasoconstrictive agent (PHE); redrawn from, Craiem Fig. 7. Plot of benign ( +) and malignant (’) breast tumor MRE data for
and Armentano [67]. 39 patients. These data are replotted from Sinkus et al. [164].
R. Magin et al. / Signal Processing 91 (2011) 350–371 367
models are applied to living systems. Clearly, when the the fractional Brownian motion. Let ra(t) be a fractional
structure in living systems is fractal, or when the noise. Define a process va(t), t Z0, by
measured signals exhibit anomalous properties, one Z t
should suspect that such dynamics might be best va ðtÞ ¼ ra ðtÞ dt ð123Þ
0
expressed by fractional order models. Much remains to
be done, and we look to the philosopher Henri Bergson to We will call this process fractional Brownian motion (or
provide inspiration, for, as Bergson [3] noted in his 1911 generalized Wiener–Lévy process). It is not difficult to
work Creative Evolution: ‘‘the present contains nothing show that it enjoys all the properties normally required
more than the past, and what is found in the effect was for fBm [21,120]:
already in the cause’’.
1. va(0) =0 and E{va(t)} =0 for every tZ0. If w(t) is
Gaussian, so are ra(t) and va(t). The proposed defini-
6. Fractional Brownian motion
tions do not need Gaussianity.
2. The covariance is [142]
Fractional Brownian motion was introduced first by
Kolmogorov [104]. Later, Mandelbrot and Van Ness s2
E½va ðtÞva ðsÞ ¼
[21,120] proposed it as a model for nonstationary signals, 2Gð2a þ 2Þcos ap
2a þ 1 2a þ 1 2a þ 1
with stationary increments, which is useful in under- ½9t9 þ 9s9 9ts9
standing phenomena with long range dependence and ð124Þ
with a frequency dependence of the form 1/fa, with a non-
integer [102,177–179]. In [144] an approach based on the valid for 9a9 o1/2. Putting H= a +1/2 wth HA(0,1),
fractional derivatives was proposed and will be described we obtain the usual formulation
next. VH h 2H 2H 2H
i
E½va ðtÞva ðsÞ ¼ 9t9 þ 9s9 9ts9 ð125Þ
Assume now that we are computing the fractional 2
derivative of the white noise, w(t), with power equal to s2. with
We define fractional noise by
s2
ra ðtÞ ¼ Da wðtÞ ð118Þ VH ¼
Gð2H þ1Þsin Hp
ð126Þ
If w(t) is Gaussian, we will call ra(t) fractional Gaussian The variance is readily obtained as
noise. As known, the autocorrelation function of the white
2H
noise is s2d(t). With some work, we obtain for the E½va ðtÞ2 ¼ VH 9t9 ð127Þ
derivative autocorrelation [136] 3. The process has stationary increments. Letting the
Gð2a þ 1Þ X
þ1
ð1Þk increments be defined by
Rar ðtÞ ¼ lim dðtkhÞ Z t
h-0 h2a Gðak þ1ÞGða þk þ 1Þ
1
Dva ðt,sÞ ¼ va ðtÞva ðsÞ ¼ ra ðtÞd t ð128Þ
ð119Þ s
For 9o9b p/T, the spectrum can be approximated by [2] Y. Bar-Yam, Dynamics of Complex Systems, Perseus Books, Read-
ing, MA, 1997.
s2 T 2
Sd ðoÞ 2H1
ð135Þ [3] H. Bergson, Creative Evolution, Dover, New York, 1998.
49o9 [4] E.N. Bruce, Biomedical Signal Processing and Signal Modeling, John
Wiley, New York, 2001.
[7] M.A. Chaudhry, S.M. Zubair, On a Class of Incomplete Gamma
We conclude that the proposed definition agrees with Functions with Applications, Chapman & Hall/CRC, 2002.
the Mandelbrot and van Ness results. [8] K. Diethelm, A.D. Freed, On the Solution of Nonlinear Fractional
The result expressed in (135) is interesting [21,120]: Differential Equations used in the Modeling of Viscoplasticity,
Springer, Heidelberg, 1999.
[9] K. Diethelm, Fractional Differential Equations, Theory and Numer-
if 0 oHo1/2, the spectrum is parabolic and corre- ical Treatment, TU Braunschweig, Braunschweig, 2003.
[10] S. Dugowson, Les différentielles métaphysiques, Ph.D. Thesis,
sponds to an antipersistent fBm; because the incre-
Université Paris Nord, 1994.
ments tend to have opposite signs, this case [11] S. Grimnes, O.G. Martinsen, Bioimpedance and Bioelectricity
corresponds to the integration of a stationary frac- Basics, Academic Press, San Diego, 2000.
[12] E. Hairer, S.P. Nørsett, G. Wanner, 2nd ed, Solving Ordinary
tional noise;
Differential Equations I: Nonstiff Problems, vol. 8, Springer-
if 1/2oHo1, the spectrum has a hyperbolic character Verlag, Berlin, 1993.
and corresponds to a persistent fBm; because the [14] P. Henrici, Applied and Computational Complex Analysis, vol. 2,
increments tend to have the same sign, this case John Wiley & Sons, Inc., 1991, pp. 389–391.
[15] R. Hilfer (Ed.), Applications of Fractional Calculus in Physics, World
corresponds to the integration of a nonstationary Scientific, Singapore, 2000.
fractional noise. [17] J. Keener, J. Sneyd, Mathematical Physiology, Springer, New York, 2004.
[18] A.A. Kilbas, H.M. Srivastava, J.J. Trujillo, Theory and Applications of
Fractional Differential Equations, Elsevier, Amsterdam, 2006.
[19] R.S. Lakes, Viscoelastic Solids, CRC Press, Boca Raton, FL, 1999.
[20] R.L. Magin, Fractional Calculus in Bioengineering, Begell House,
7. Conclusions Connecticut, 2006.
[21] B.B. Mandelbrot, The Fractal Geometry of Nature, W. H. Freeman
Fractional calculus models provide a relatively simple and Company, New York, 1983.
[22] K.S. Miller, B. Ross, An Introduction to the Fractional Calculus and
way to describe the physical and electrical properties of Fractional Differential Equations, Wiley, 1993.
complex, heterogeneous and composite biomaterials. [23] K. Nishimoto, Fractional Calculus, Descartes Press Co., Koriyama, 1989.
There is a multi-scale generalization inherent in the [24] K.B. Oldham, J. Spanier, The Fractional Calculus: Theory and
Application of Differentiation and Integration to Arbitrary Order,
definition of the fractional derivative that accurately Academic Press, 1974.
represents interactions occurring over a wide range of [25] A. Oustaloup, La Dérivation Non Entire, Editions Herme s, 1995.
space or time. Thus, we can avoid excessive segmentation [26] I. Podlubny, Fractional-Order Systems and Fractional-Order Con-
trollers, The Academy of Sciences, Institute of Experimental
or compartmentalization of tissues into subsystems or
Physics, Kosice, Slovak Republic, 1994.
subunits—a system reduction that often creates more [27] I. Podlubny, Numerical Solution of Ordinary Fractional Differential
computational and compositional complexity than can be Equations by the Fractional Difference Method, Gordon and
experimentally evaluated. Finally, fractional calculus Breach, Amsterdam, 1997.
[28] I. Podlubny, Fractional Differential Equations: An Introduction to
models suggest new experiments and measurements that Fractional Derivatives, Fractional Differential Equations, to Meth-
can shed light on the meaning of biological system ods of their Solution and Some of their Applications, Academic
structure and dynamics. Thus, by applying fractional Press, San Diego, 1999.
[29] J.M. Sabatier, O.P. Agrawal, J.A.T. Machado, Advances in Fractional
calculus to model the behavior of cells and tissues, we Calculus: Theoretical Developments and Applications in Physics
can begin to unravel the inherent complexity of individual and Engineering, Springer, 2007.
molecules and membranes in a way that leads to an [30] S.G. Samko, A.A. Kilbas, O.I. Marichev, Fractional Integrals and
Derivatives—Theory and Applications, Gordon and Breach Science
improved understanding of the overall biological function
Publishers, 1987.
and behavior of living systems. [31] M. Shelhamer, Nonlinear Dynamics in Physiology: A State Space
Approach, World Scientific, Singapore, 2007.
[32] M. Weilbeer, Efficient numerical methods for fractional differen-
tial equations and their analytical background, Ph.D. Thesis, Von
Acknowledgements der Carl-Friedrich-Gauß-Fakultät für Mathematik und Informatik
der Technischen Universität Braunschweig, 2005.
[33] B.J. West, M. Bologna, P. Grigolini, Physics of Fractal Operators,
The work of of Richard Magin was supported, in part,
Springer, New York, 2003.
by the NIH, NIBIB Grant R 01 EB 007537. [34] S. Westerlund, Dead Matter has Memory, Causal Consulting,
The work of Manuel Ortigueira was supported by Kalmar, Sweden, 2002.
[35] A.H. Zemanian, Distribution Theory and Transform Analysis, Dover
‘‘Fundac- a~ o para a Ciência e Tecnologia’’ (CTS multiannual
Publications, New York, 1987.
funding) through the PIDDAC Program funds. Igor [36] O.P. Agrawal, J.A.T. Machado, J. Sabatier (Eds.), Special issue on
Podlubny is partly supported by Grants APVV-0040-07 fractional derivatives and their applications, Nonlinear Dynamics
and VEGA 1/0390/10, 1/0404/08. The work of Juan Trujillo 38(1–4) (2004) December.
[37] J.A.T. Machado (Ed.), Special issue on fractional order calculus and
was supported in part by MICINN of Spain (Grants its applications, Nonlinear Dynamics 29 (1–4) (2002) July.
MTM2007-60246 and MTM2010-16499). [38] M.D. Ortigueira, J.A.T. Machado (Eds.), Special issue on fractional
signal processing and applications, Signal Processing 83 (11)
(2003) November.
References [39] M.D. Ortigueira, J.A.T. Machado (Eds.), Special section: fractional
calculus applications in signals and systems, Signal Processing 86
(10) (2006) October.
[1] G. Adomian, Solving Frontier Problems of Physics: The Decom- [40] O. Abdulaziz, I. Hashim, M.S.H. Chowdhury, A.K. Zulkifle, Assess-
position Method, Kluwer Academic Publishers, Boston, 1994. ment of decomposition method for linear and nonlinear fractional
R. Magin et al. / Signal Processing 91 (2011) 350–371 369
differential equations, Far East Journal of Applied Mathematics 28 [63] Y.Q. Chen, B.M. Vinagre, A new IIR-type digital fractional order
(1) (2007) 112–950. differentiator, Signal Processing 83 (11) (2003) 2359–2365.
[41] O. Abdulaziz, I. Hashim, E.S. Ismail, Approximate analytical [64] Y.Q. Chen, B.M. Vinagre, I. Podlubny, Continued fraction expansion
solution to fractional modified kdv equations, Mathematical and approaches to discretizing fractional order derivatives—an ex-
Computer Modelling, 49 (1–2) (2009) 136–145. pository review, Nonlinear Dynamics 38 (2004) 155–170.
[42] O. Abdulaziz, I. Hashim, S. Momani, Solving systems of fractional [65] T. Clarke, B.N.N. Achar, Hanneken, Mittag–Leffler functions and
differential equations by homotopy–perturbation method, Physics transmission lines, Journal of Molecular Liquids 114 (1–3) (2004)
Letters A 372 (2008) 451–459. 159–163.
[43] G. Adomian, A review of the decomposition method in applied [66] J.P. Clerc, A.-M.S. Tremblay, G. Albinet, C.D. Mitescu, a. c.
mathematics, Journal of Mathematical Analysis and Applications response of fractal networks, Le Journal de Physique-Lettres
135 (1988) 501–544. 45 (19) (1984).
[44] M.A. Al-Alaoui, Novel digital integrator and differentiator, Electro- [67] D. Craiem, R.L. Armentano, A fractional derivative model to
nics Letters 29(4) (1993) 376–378. describe arterial viscoelasticity, Biorheology 44 (2007) 251–263.
[45] T.J. Anastasio, The fractional-order dynamics of brainstem vesti- [68] V. Daftardar-Gejji, H. Jafari, Adomian decomposition: a tool for
bulo-ocular dynamics of brainstem vestibular–oculomotor neu- solving a system of fractional differential equations, Journal of
rons, Biological Cybernetics 72 (1994) 69–79. Mathematical Analysis and Applications 301 (2) (2005) 508–518.
[46] T.J. Anastasio, Nonuniformity in the linear network model of the [70] J.B. Diaz, T.J. Osler, Differences of fractional order, Mathematics of
oculomotor integrator produces approximately fractional-order Computation 28 (125) (1974).
dynamics and more realistic neuron behavior, Biological Cyber- [71] K. Diethelm, An algorithm for the numerical solution of differ-
netics 79 (1998) 377–391. ential equations of fractional order, Electronic Transactions on
[47] H.L. Arora, F.I. Abdelwahid, Solution of non-integer order differ- Numerical Analysis 5 (1997) 1–6.
ential equations via the adomian decomposition method, Applied [72] K. Diethelm, Generalized compound quadrature formulae for
Mathematics Letters 6 (1) (1993) 21–23. finite-part integrals, IMA Journal of Numerical Analysis 17 (3)
[48] R.S. Barbosa, J.A.T. Machado, I.M. Ferreira, Least-squares design of (1997) 479–493.
digital fractional-order operators, in: Proceedings of FDA’2004 [73] K. Diethelm, N. Ford, Analysis of fractional differential equations,
First IFAC Workshop on Fractional Differentiation and its Applica- Journal of Mathematical Analysis and Applications 265 (2002)
tions, 19–21 July 2004, Bordeaux, France. 229–248.
[49] R.S. Barbosa, J.A.T. Machado, I.M. Ferreira, Pole-zero approxima- [74] K. Diethelm, N.J. Ford, Numerical solution of the Bagley–Torvik
tions of digital fractional-order integrators and differentiators equation, BIT 42 (3) (2002) 490–507.
using signal modeling techniques, in: Proceedings of the 16th IFAC [75] K. Diethelm, N.J. Ford, A.D. Freed, A predictor-corrector approach
World Congress, July 2005, Prague, Czech Republic. for the numerical solution of fractional differential equations,
[50] H. Beyer, S. Kempfle, Definition of physically consistent damping Nonlinear Dynamics 29 (2002) 3–22.
laws with fractional derivatives, Zeitschrift fur Angewandte [76] K. Diethelm, A.D. Freed, The FracPECE subroutine for the numerical
Mathematik und Mechanik 75 (1995) 623–635. solution of differential equations of fractional order, Gesellschaft
[51] L. Blank, Numerical treatment of differential equations of für wissenschaftliche Datenverarbeitung, Göttingen, 1999.
fractional order, Numerical Analysis Report 287, 1996. [77] K. Diethelm, Y. Luchko, Numerical solution of linear multi-term
[52] N. Bildik, A. Konuralp, The use of variational iteration method, differential equations of fractional order Journal of Computational
differential transform method and Adomian decomposition Analysis and Applications, 2003.
method for solving different types of nonlinear partial differential [78] K. Diethelm, G. Walz, Numerical solution for fractional differential
equations, International Journal of Nonlinear Science and Numer- equations by extrapolation, Numerical Algorithms 16 (1997)
ical Simulations 7 (1) (2006) 65–70. 231–253.
[53] G.W. Bohannan, Analog fractional order controller in a tempera- [80] N. Engheta, Fractional curl operator in electromagnetics, Micro-
ture control application, in: Proceedings of the Second IFAC wave and Optical Technology Letters 17 (2) (1998).
Workshop on Fractional Differentiation and its Applications, July [81] N. Engheta, Fractional paradigm in electromagnetic theory, in:
2006, Porto, Portugal. D.H. Werner, R. Mittra (Eds.), Frontiers in Electromagnetics, IEEE
[54] B. Bonilla, M. Rivero, J.J. Trujillo, Linear differential equations of Press, 2000 (Chapter 12).
fractional order, in: Proceedings of the Second Symposium on [82] A. Fenander, Modal synthesis when modeling damping by use of
Fractional Derivatives and their Applications, FDTAs, SEP, 2005, fractional derivatives, American Institute of Aeronautics and
Long Beach, CA. Astronautics Journal 34 (5) (1996) 1051–1058.
[55] B. Bonilla, M. Rivero, J.J. Trujillo, On systems of linear fractional [83] L. Gaul, P. Klein, S. Kemple, Damping description involving
differential equations with constant coefficients, in: Proceedings fractional operators, Mechanical Systems and Signal Processing 5
of the International Symposium on Analytic Function Theory, (2) (1991) 81–88.
Fractional Calculus and their Applications, August 22–27, 2005 [84] A.J. George, A. Chakrabarti, The adomian method applied to some
University of Victoria, Victoria, Canada. extraordinary differential equations, Applied Mathematics Letters
[56] B. Bonilla, M. Rivero, J.J. Trujillo, On theory of systems of 8 (3) (1995) 91–97.
fractional linear differential equations, in: Proceedings of the [85] R. Gorenflo, F. Mainardi, Fractional diffusion processes: probability
Fifth International Conference on Multibody Systems, Nonlinear distributions and continuous time random walk, in: G. Rangarajan,
Dynamics, and Control, September 24–28, 2005, Long Beach CA, M. Ding (Eds.), Long Range Dependent Processes: Theory and
USA. Applications, Springer-Verlag, 2003.
[57] B. Bonilla, M. Rivero, J.J. Trujillo, On systems of linear fractional [86] R. Gorenflo, F. Mainardi, D. Moretti, P. Paradisi, Time-fractional
differential equations with constant coefficients, Applied Mathe- diffusion: a discrete random walk approach, Nonlinear Dynamics
matics and Computation 187 (9) (2007) 68–78. 29 (1–4) (2002) 129–143.
[58] B. Bonilla, M. Rivero, L. Rodrı́guez-Germá, J.J. Trujillo, Fractional [87] W. Greatbatch, W.M. Chardack, Myocardial and endocardial
differential equations as alternative models to the nonlinear electrodes for chronic implantation, Annals of the New York
fractional differential equations, Applied Mathematics and Com- Academy of Sciences 148 (1968) 235–251.
putation 187 (9) (2007) 79–88. [88] E. Hairer, C. Lubich, M. Schlichte, Fast numerical solution of
[59] L.M.C. Campos, Fractional calculus of analytic and branched nonlinear Volterra convolution equations, SIAM Journal on
functions, in: R.N. Kalia (Ed.), Recent Advances in Fractional Scientific and Statistical Computing 6 (3) (1985) 532–541.
Calculus Global Publishing Company, 1993. [91] J.H. He, Homotopy perturbation technique, Computer Methods in
[60] L.M.C. Campos, On a concept of derivative of complex order with Applied Mechanics and Engineering 178 (3–4) (1999) 257–262.
applications to special functions, IMA Journal of Applied Mathe- [93] J.H. He, A coupling method of a homotopy technique and a
matics 33 (1984) 109–133. perturbation technique for non-linear problems, International
[61] M. Caputo, Linear models of dissipation whose Q is almost Journal of Non-Linear Mechanics 35 (1) (2000) 37–43.
frequency independent, Part II, Geophysical Journal of the Royal [96] N. Heymans, I. Podlubny, Physical interpretation of initial condi-
Astronomical Society 13 (1967) 529–539. tions for fractional differential equations with Riemann–Liouville
[62] Y.Q. Chen, K.L. Moore, Discretization schemes for fractional order fractional derivatives, Rheol Acta 37 (2005) 1–7.
differentiators and integrators, IEEE Transactions on Circuits and [97] F. Huang, F. Liu, The fundamental solution of the space–time
Systems I: Fundamental Theory and Applications 49 (3) (2002) fractional advection–dispersion equation, Journal of Applied
363–367. Mathematics and Computing 18 (2) (2005) 339–350.
370 R. Magin et al. / Signal Processing 91 (2011) 350–371
[98] F. Huang, F. Liu, The time-fractional diffusion equation and [123] S. Momani, Numerical simulation of a dynamic system containing
fractional advection–dispersion equation, ANZIAM Journal 46 fractional derivatives, in: International Symposium on Nonlinear
(2005) 1–14. Dynamics, Shanghai, China, 2005.
[99] M. Itagaki, A. Taya, K. Watanabe, K. Noda, Deviations of capacitive [124] S. Momani, A numerical scheme for the solution of multi-order
and indutive loops in the electrochemical impedance of a dissolving fractional differential equations, Applied Mathematics and Com-
iron electrode, Analytical Sciences 18 (6) (2002) 641–644. putation 182 (1) (2006) 761–770.
[100] H. Jafari, V. Daftardar-Gejji, Solving a system of nonlinear [125] S. Momani, S. Abuasad, Application of He’s variational iteration
fractional differential equations using Adomian decomposition, method to Helmholtz equation, Chaos, Solitons & Fractals 27 (5)
Journal of Computational and Applied Mathematics 196 (2) (2006) 1119–1123.
(2006) 644–651. [126] S. Momani, K. Al-Khaled, Numerical solutions for systems of
[101] K. Kawaba, W. Nazri, H.K. Aun, M. Iwahashi, N. Kambayashi, A fractional differential equations by the decomposition
realization of fractional power-law circuit using OTAs, in: method, Applied Mathematics and Computation 162 (3) (2005)
Proceedings of the 1998 IEEE Asia-Pacific Conference on Circuits 1351–1365.
and Systems, APCCAS 1998, pp. 249–252, 24–27 November 1998, [127] S. Momani, Z. Odibat, Comparison between the homotopy
Chiangmai, Thailand. perturbation method and the variational iteration method for
[102] M.S. Keshner, 1/f noise, in: Proceedings of IEEE, vol. 70, no. 3, linear fractional partial differential equations, Computers and
March 1982, pp. 212–218. Mathematics with Applications 54 (7–8) (2007) 910–919.
[103] A.A. Kilbas, M. Rivero, L. Rodrı́guez-Germá, J.J. Trujillo, a-analytic [128] S. Momani, Z. Odibat, Homotopy perturbation method for non-
solutions of some linear fractional differential equations with linear partial differential equations of fractional order, Physics
variable coefficients, Applied Mathematics and Computation 187 Letters A 365 (5–6) (2007) 345–350.
(9) (2007) 239–249. [129] S. Momani, Z. Odibat, Numerical approach to differential equa-
[104] N.A. Kolmogorov, Wienersche Spiralen und Einige Andere tions of fractional order, Journal of Computational and Applied
Interessante Kurven im Hilbertschen Raum, Compte Rendus de Mathematics 207 (2007) 96–110.
l’Académie des Sciences de 1’Union des Républiques Soviétiques [130] S. Momani, Z. Odibat, Numerical comparison of methods for
Socialistes, vol. 26, no. 2, 1940, pp. 115–118. solving linear differential equations of fractional order, Chaos,
[105] M.S., Krishna, S., Das, K. Biswas, B. Goswami, Characterization of a Solitons & Fractals 31 (2007) 1248–1255.
fractional order element realized by dipping a capacitive type [131] S. Momani, R. Qaralleh, An efficient method for solving systems of
probe in polarizable medium, in: Proceedings of the Symposium fractional integro-differential equations, Computers and Mathe-
on Fractional Signals and Systems, Lisbon’09, Faculdade de matics with Applications 52 (3-4) (2006) 459–470.
Ciências e Tecnologia da Universidade Nova de Lisboa, Caparica, [132] C.A. Monje, B.M. Vinagre, A.J. Calderón, J.I. Suárez, I. Tejado, Some
Portugal, November 4–6. experiences in linear and nonlinear fractional order control,
[106] P. Kumar, O.P. Agrawal., An approximate method for numerical Presentation at the Symposium on Applied Fractional Calculus
solution of fractional differential equations, Signal Processing 86 (SAFC07), Industrial Engineering School (University of Extrema-
(2006) 2602–2610. dura), 15–17 October 2007, Badajoz, Spain.
[107] P. Lanusse,J. Sabatier, R. Malti, P. Melchior, X. Moreau, and A. [133] D. Moretti, A. Vivoli, Random walks and fractional diffusion,
Oustaloup, Past and some recent CRONE Group applications of Interdisciplinary Workshop ‘‘From Waves to Diffusion and
fractional differentiation, Presentation at the Symposium on Beyond,’’ 20 December 2002, Bologna, Italy.
Applied Fractional Calculus (SAFC07), Industrial Engineering [135] M.D. Ortigueira, Introduction to fractional signal processing, part
School (University of Extremadura), 15–17 October 2007, 1: continuous-time systems, IEE Proceedings on Vision, Image and
Badajoz, Spain. Signal Processing 1 (2000) 62–70.
[109] H.M. Liu, Variational approach to nonlinear electrochemical [136] M.D. Ortigueira, Introduction to fractional signal processing, part
system,, International Journal of Nonlinear Science and Numer- 2: discrete-time systems, IEE Proceedings on Vision, Image and
ical Simulations 5 (1) (2004) 95–96. Signal Processing 147 (1) (2000) 71–78.
[111] C. Lubich, On the stability of linear multistep methods for [137] M.D. Ortigueira, On the initial conditions in continuous-time fractional
Volterra convolution equations, IMA Journal of Numerical linear systems, Signal Processing 83 (11) (2003) 2301–2309.
Analysis 3 (4) (1983) 439–465. [138] M.D. Ortigueira, From differences to differintegrations, Fractional
[112] C. Lubich, Fractional linear multistep methods for Abel–Volterra Calculus and Applied Analysis 7 (4) (2004) 459–471.
integral equations of the second kind, Mathematics of Computa- [139] M.D. Ortigueira, J.A. Tenreiro-Machado, J. Sá da Costa, Which
tion 45 (172) (1985) 463–469. differintegration?, IEE Proceedings Vision, Image and Signal
[113] C. Lubich, Discretized fractional calculus, SIAM Journal on Processing 152 (6) (2005) 846–850.
Mathematical Analysis 17 (3) (1986) 704–719. [140] M.D. Ortigueira, A coherent approach to non integer order deriva-
[114] J.A.T. Machado, Analysis and design of fractional-order digital tives, Signal Processing, Special Section: ‘‘Fractional Calculus
control systems, Journal of Systems Analysis, Modelling and Applications in Signals and Systems 86 (10) (2006) 2505–2515.
Simulation 27 (1997) 107–122. [141] M.D. Ortigueira, A.J. Serralheiro, A new least-squares approach to
[115] J.A.T. Machado, Discrete-time fractional-order controllers, differintegration modelling, Signal Processing, Special Section:
Journal of Fractional Calculus and Applied Analysis 4 (1) (2001) ‘‘Fractional Calculus Applications in Signals and Systems 86 (10)
47–66. (2006) 2582–2591.
[116] P.J. Mahon, G.L. Paul, S.M. Kesshishian, A.M. Vassallo, Measure- [142] M.D. Ortigueira, Riesz potentials and inverses via centred
ment and modelling of high-power performance of carbon-based derivatives, International Journal of Mathematics and Mathema-
supercapacitors, Journal of Power Sources 91 (1) (2000) 68–76. tical Sciences (2006) Article ID 48391.
[117] F. Mainardi, Linear viscoelasticity, Chapter 4, in: A Guran, A [143] M.D. Ortigueira, A.J. Serralheiro, Pseudo-fractional ARMA model-
Bostrom, O Leroy, G Maze (Eds.), Acoustic Interactions with ling using a double Levinson recursion, IET Control Theory and
Submerged Elastic Structures, Part IV: Nondestructive Testing, Applications 1 (1) (2007) 173–178.
Acoustic Wave Propagation and Scattering, World Scientific, [144] M.D. Ortigueira, A.G. Batista, On the relation between the
Singapore, 2002, pp. 97–126. fractional Brownian motion and the fractional derivatives, Physics
[118] F. Mainardi, The time fractional diffusion-wave equation, Radio- Letters A 372 (7) (2008) 958–968.
physics and Quantum Electronics 38 (1–2) (1995) 13–24. [145] M.D. Ortigueira, F.V. Coito, The initial conditions of Riemann–
[119] N. Makris, M.C. Constantinou, Fractional-derivative maxwell Liouville and Caputo derivatives, in: Proceedings of the Sixth
model for viscous dampers, Journal of Structural Engineering EUROMECH Conference ENOC 2008, June 30–July 4, 2008, Saint
117 (9) (1991) 2708–2724. Petersburg, Russia.
[120] B.B. Mandelbrot, J.W. Van Ness, The fractional brownian motions, [146] M.D. Ortigueira, J.J. Trujillo, Generalized GL fractional derivative
fractional noises and applications, SIAM Review 10 (4) (1968) and its Laplace and Fourier transform, in: Proceedings of the ASME
422–437. 2009 International Design Engineering Technical Conferences &
[121] V. Marinca, An approximate solution for one-dimensional weakly Computers and Information in Engineering Conference IDETC/CIE
nonlinear oscillations, International Journal of Nonlinear Science 2009, August 30–September 2, 2009, San Diego, CA, USA.
and Numerical Simulations 3 (2) (2002) 107–110. [147] M.D. Ortigueira, F.J. Coito, System initial conditions vs derivative
[122] S. Momani, An explicit and numerical solutions of the fractional initial conditions, Computers and Mathematics with Applications
KdV equation, Mathematics and Computers in Simulation 70 (2) special issue on Fractional Differentiation and Its Applications
(2005) 110–118. 59 (5) (2010) 1782–1789.
R. Magin et al. / Signal Processing 91 (2011) 350–371 371
[148] A. Oustaloup, Fractional order sinusoidal oscillators: optimization [172] B.M. Vinagre, I. Podlubny, A. Hernandez, V. Feliu, Some approx-
and their use in highly linear FM modulation, IEEE Transactions on imations of fractional order operators used in control theory and
Circuits and Systems CAS-28 (10) (1981). applications, Fractional Calculus and Applied Analysis 3 (3) (2000)
[149] A. Oustaloup, B. Mathieu, P. Lanusse, The CRONE control of 231–248.
resonant plants: application to a flexible transmission, European [173] B.M. Vinagre, I. Petras, P. Merchan, L. Dorcak, Two digital
Journal of Control 1 (2) (1995) 113–121. realisation of fractional controllers: application to temperature
[150] A. Oustaloup, F. Levron, F. Nanot, B. Mathieu, Frequency band control of a solid, In Proceedings of the European Control
complex non integer differentiator: characterization and synth- Conference (ECC2001), Porto, Portugal, 2001, pp. 1764–1767.
esis, IEEE Transactions on Circuits and Systems I: Fundamental [174] B. Vinagre, V. Feliu, Modeling and control of dynamic system using
Theory and Applications 47 (1) (2000) 25–40. fractional calculus: application to electrochemical, in: Proceedings
[151] A. Oustaloup, O. Cois, P. Lanusse, P. Melchior, X. Moreau, of the 41st IEEE Conference on Decision and Control, Las Vegas,
J. Sabatier, The CRONE approach: theoretical developments and NV, 2002, pp. 214–239.
major applications, in: Proceedings of the Second IFAC Workshop [175] B.M. Vinagre, Y.Q. Chen, I. Petras, Two direct tustin discretization
on Fractional Differentiation and its Applications, July 2006, Porto, methods for fractional-order differentiator/integrator, The Journal
Portugal. of Franklin Institute 340 (5) (2003) 349–362.
[152] M. Ovadia, D.H. Zavitz, The electrode-tissue interface in living [176] B.M. Vinagre, V. Feliu, Optimal fractional controllers for rational
heart: equivalent circuit as a function of surface area, Electro- order systems: a special case of the Wiener–Hopf spectral
analysis 10 (1998) 262–272. factorization method, IEEE Transactions on Automatic Control 52
[153] I. Podlubny, M. Kacenak, Mittag–Leffler function, /http://www. (12) (2007) 2385–2389.
mathworks.com/matlabcentral/fileexchange/8738S, 17 October [177] R.F. Voss, J. Clarke, 1/f noise in music: music from 1/f noise,
2005 (Updated 25 March 2009). Journal of Acoustics Society of America 63 (1) (1978) 258–263.
[154] I. Podlubny, T. Skovranek, B. Vinagre, Matrix approach to [178] R.F. Voss, J. Clarke, 1/f noise in speech and music, Nature 258
discretization of fractional derivatives and to solution of fractional (1975) 317–318.
differential equations and their systems, in IEEE Conference [179] W. Willinger, M.S. Taqqu, W.E. Leland, V. Wilson, Self-similarity
on Emerging Technologies Factory Automation ETFA 2009, in high-speed packed traffic: analysis and modelling of ethernet
September 22–25, Palma de Mallorca, Spain, doi: 10.1109/ traffic measurements, Statistical Science 10 (1) (1995) 67–85.
ETFA.2009.5347166, 2009.
[155] I. Podlubny, Matrix approach to discrete fractional calculus,
Fractional Calculus and Applied Analysis 3 (4) (2000) 359–386.
[156] V. Ramachandran, C.S. Cargour, M. Ahmadi, Cascade realization of
Further reading
the irrational immittance S1/2, in: IEE Proceedings, Part G, vol.
132, no. 2, April 1985, pp. 64–67.
[157] S.S. Ray, R.K. Bera, An approximate solution of a nonlinear fractional [5] H. Brunner, P.J. van der Houwen, The Numerical Solution of
differential equation by adomian decomposition method, Applied Volterra Equations, vol. 3, North-Holland Publishing Co., Amster-
Mathematics and Computation 167 (1) (2005) 561–571. dam, 1986.
[158] S.S. Ray, R.K. Bera, Solution of an extraordinary differential [6] A. Carpinteri, F. Mainardi (Eds.), Fractals and Fractional Calculus in
equation by adomian decomposition method, Journal of Applied Continuum Mechanics, CISM Courses and Lectures No. 378,
Mathematics 4 (2004) 331–338. Springer-Wien, New York, 1997.
[159] E.A. Rawashdeh, Numerical solution of semidifferential equations [13] J.H. He, Non-perturbative methods for strongly nonlinear
by collocation method, Applied Mathematics and Computation problems, Ph.D. Thesis, Berlin, 2006.
174 (2006) 869–876. [16] M. Inokuti, H. Sekine, T. Mura, General use of the Lagrange
[160] J. Sabatier, M. Aoun, A. Oustaloup, G. Gregoire, F. Ragot, P. Roy, Multiplier in Non-Linear Mathematical Physics, Pergamon Press,
Fractional system identification for lead acid battery state charge Oxford, 1978.
estimation, Signal Processing 86 (10) (2006) 2645–2657. [69] V. Daftardar-Gejji, H. Jafari, An iterative method for solving
[161] G.E. Santamarı́a, J.V. Valverde, R. Pérez-Aloe, B.M. Vinagre, nonlinear functional equations, Journal of Mathematical Analysis
Microelectronic implementations of fractional-order integrodif- and Applications 316 (2006) 753–763.
ferential operators, Journal of Computational and Nonlinear [79] G.E. Drăgănescu, Application of a variational iteration method to
Dynamics 3 (2) (2008) 021301. linear and nonlinear viscoelastic models with fractional deriva-
[162] I. Schäfer, K. Krüger, Modelling of coils using fractional derivatives, tives, Journal of Mathematical Physics 47 (8) (2006) Art. No.
Journal of Magnetism and Magnetic Materials 307 (1) (2006) 91–98. 082902.
[163] N. Shawagfeh, D. Kaya, Comparing numerical methods for the [89] T.H. Hao, Search for variational principles in electrodynamics by
solutions of systems of ordinary differential equations, Applied Lagrange method, International Journal of Nonlinear Science and
Mathematics Letters 17 (2004) 323–328. Numerical Simulations 6 (2) (2005) 209–210.
[164] R. Sinkus, K. Siegmann, T. Xydeas, M. Tanter, C. Claussen, M. Fink, [90] J.H. He, Variational iteration method for delay differential
MR Elastography of breast lesions: understanding the solid/liquid equations, Communications in Nonlinear Science and Numerical
duality can improve the specificity of contrast-enhanced MR Simulations 2 (4) (1997) 235–236.
mammography, Magnetic Resonance in Medicine 58 (2007) [92] J.H. He, Variational iteration method—a kind of non-linear
1135–1144. analytical technique: some examples, International Journal of
[165] H. Sun, A. Abdelwahed, B. Onaral, Linear approximation for Nonlinear Mechanics 34 (1999) 699–708.
transfer function with a pole of fractional order, IEEE Transactions [94] J.H. He, Variational iteration method for autonomous ordinary
on Automatic Control 29 (5) (1984) 441–444. differential systems, Applied Mathematics and Computation 114
[167] J. Thorson, M. Biederman-Thorson, Distributed relaxation pro- (2000) 115–123.
cesses in sensory adaptation, Science 183 (1974) 161–172. [95] J.H. He, Variational principle for some nonlinear partial differential
[168] P.J. Torvik, R.L. Bagley, On the appearance of the fractional equations with variable coefficients, Chaos, Solitons & Fractals 19
derivatives in the behavior of real materials, Journal of Applied (4) (2004) 847–851.
Mechanics (Transaction ASME) 51 (2) (1984) 294–298. [108] P. Linz, Analytical and numerical methods for Volterra equations,
[169] D. Valério, M.D. Ortigueira, J. Sá da Costa, Identifying a transfer SIAM Studies in Applied Mathematics 7 (1985).
function from a frequency response, ASME Journal of Computa- [110] H.M. Liu, Generalized variational principles for ion acoustic
tional and Nonlinear Dynamics, Special Issue Discontinuous and plasma waves by He’s semi-inverse method, Chaos, Solitons &
Fractional Dynamical Systems 3 (2) (2008) 021207. Fractals 23 (2) (2005) 573–576.
[170] P. Varshney, P. Gupta, G.S. Visweswaran, New switched capacitor [134] Z. Odibat, S. Momani, Application of variational iteration method
fractional order integrator, Journal of Active and Passive Electronic to nonlinear differential equations of fractional order, Interna-
Devices 2 (3) (2007) 187–197. tional Journal of Nonlinear Science and Numerical Simulation 1
[171] B.M. Vinagre, I. Podlubny, A. Hernandez, V. Feliu, On realization (7) (2006) 15–27.
of fractional-order controllers, in: Proceedings of the Conference [166] V.E. Tarasov, Fractional statistical mechanics, Chaos 16 (3)
Internationale Francophone d’Automatique, Lille, France, 2000. (2006).