Nonlinear Analysis 69 (2008) 12991320

www.elsevier.com/locate/na

Limitations of frequency domain approximation for detecting chaos

in fractional order systems
Mohammad Saleh Tavazoei, Mohammad Haeri
Advanced Control Systems Lab., Electrical Engineering Department, Sharif University of Technology, Tehran, Iran

Received 7 April 2007; accepted 19 June 2007

Abstract

In this paper, we analytically study the influences of using frequency domain approximation in numerical simulations of
fractional order systems. The number and location of equilibria, and also the stability of these points, are compared between
the original system and its frequency based approximated counterpart. It is shown that the original system and its approximation
are not necessarily equivalent according to the number, location and stability of the fixed points. This problem can cause erroneous
results in special cases. For instance, to prove the existence of chaos in fractional order systems, numerical simulations have been
largely based on frequency domain approximations, but in this paper we show that this method is not always reliable for detecting
chaos. This approximation can numerically demonstrate chaos in the non-chaotic fractional order systems, or eliminate chaotic
behavior from a chaotic fractional order system.
c 2007 Elsevier Ltd. All rights reserved.

Keywords: Fractional order system; Chaos; Frequency domain approximation

1. Introduction

Fractional calculus as an extension of ordinary calculus has a 300-year-old history. It has been found that
the behavior of many physical systems can be properly described by using the fractional order system theory.
For example heat conduction [1], dielectric polarization [2], electrodeelectrolyte polarization [3], electromagnetic
waves [4], viscoelastic systems [5], quantum evolution of complex systems [6], quantitative finance [7] and diffusion
waves [8] are known dynamical systems governed by the fractional order equations. In fact, real world processes
generally or most likely are fractional order systems [9]. Furthermore, fractional order controllers such as the CRONE
controller [10], TID controller [11], fractional PID controller [12] and leadlag compensator [13] have already been
implemented to improve the performance and robustness of closed loop control systems.
In recent years, considerable attention has been paid to finding chaotic behaviors in fractional order models. For
example, it has been found that some fractional order differential systems such as the fractional order Chua circuit [14],
the fractional order Duffing system [15], the fractional order jerk model [16], the fractional order Chen system [17],
the fractional order Lu system [18], the fractional order Rossler system [19], the fractional order Arneodo system [20]

Corresponding author. Tel.: +98 21 66165964; fax: +98 21 66023261.

E-mail addresses: m tavazoei@ee.sharif.edu (M.S. Tavazoei), haeri@sina.sharif.edu (M. Haeri).

c 2007 Elsevier Ltd. All rights reserved.

0362-546X/$ - see front matter
doi:10.1016/j.na.2007.06.030
1300 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

and the fractional order NewtonLeipnik system [21] demonstrate chaotic behavior. In most of the literature in
the fractional chaos field, the existence of chaotic behavior in fractional order systems has been validated only by
numerical simulations. The methods used in numerical simulation of these chaotic fractional order systems are based
on two different viewpoints. The first viewpoint is that of simulating a fractional system by numerically solving
its fractional differential equations. Methods applied to find numerical solutions of differential equations directly
approximate the response of a fractional order system and are called time domain methods. One of the best methods
in this category is the improved version of the AdamsBashforthMoulton algorithm and it is proposed on the basis
of the predictorcorrector scheme. This method was introduced in [22,23] and has been used in some of the literature
such as [21,2426].
Another viewpoint as regards simulating a fractional order system is based on the frequency domain
approximations of the fractional operators. From the control theory viewpoint, the methods proposed for obtaining an
integer order continuous model, such as rational approximation of the irrational transfer function, can be divided into
two groups: methods using continued fraction expansions (CFE) and interpolation techniques, and the methods using
curve fitting or identification techniques [27]. For example, Carlsons method [28] and Matsudas method [29] belong
to the first group and Oustaloups method [30,31] and Charefs method [32,33] are classified in the second group. To
simulate a fractional order system by using the frequency domain approximations, the fractional order equation of
the system is first considered in the frequency domain and then the Laplace form of the fractional integral operator is
replaced by its integer order approximation (this approximation is often obtained by using methods from the second
group such as Charefs method). Then the approximated equations in the frequency domain are transformed into
the time domain. The resulting ordinary differential equations can be numerically solved by applying well known
numerical methods. The frequency domain methods have been used in many papers (for example [14,1620,34]) to
simulate the behavior of chaotic fractional order systems.
This paper is organized as follows. Section 2 briefly describes basic concepts in fractional calculus, fractional
systems, and the stability of these systems. We explain in Section 3 how the solution for a fractional order system is
approximated using the approximation methods mentioned. In Section 4, the number and location of equilibria, and
also the stability of these points, in the frequency based approximation model are compared with those for the original
system. In Section 5, we show that using frequency domain approximation in simulation of a fractional order system
may cause generation of fake chaotic behavior or may eliminate chaos from a fractional order chaotic system. Also,
on the basis of the stability theorem in a fractional order system, the required condition by which a fractional order
system demonstrates chaos and produces a one-scroll, two-scroll or multi-scroll chaotic attractor is discussed in this
section. Conclusions in Section 6 close the paper.

2. Introduction to fractional calculus

q
The differintegral operator, denoted by a D t , is a combined differentiation and integration operator commonly used
in fractional calculus. This operator represents taking both the fractional derivative and the fractional integral in a
single expression and is defined by
q
d

, q>0
dt q

q 1, q=0
a Dt = Z t (1)

(d )q , q < 0.


a

There are different definitions for fractional derivatives [35]. The most commonly used definitions are the
GrunwaldLetnikov, RiemannLiouville and Caputo definitions. These definitions are briefly discussed on the
following lines.
GrunwaldLetnikov definition:

dq f (t) t a q NX1
      
q j q t a
a Dt f (t) = = lim (1) f tj . (2)
d(t a)q N N j=0
j N
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1301

RiemannLiouville definition:
Z t
1
(t )q1 f ( )d q<0


(q) a

q
a D t f (t) = f (t) q=0
(3)

qn

f (t)] q>0
n
D [a D t
where n is the first integer which is not less than q, i.e., n 1 q < n and is the Gamma function:
Z
(z) = t z1 et dt. (4)
0

For a wide class of functions, the GrunwaldLetnikov and the RiemannLiouville definitions are equivalent [35]. The
Laplace transform of the RiemannLiouville fractional integral is given as follows:
n o
q
L 0 D t f (t) = s q F(s) q 0. (5)

Also, the Laplace transform of the RiemannLiouville fractional derivative is

n o n1
q qk1
X
L 0 Dt f (t) = s q F(s) s k 0 Dt f (0) n 1 < q n N. (6)
k=0

Unfortunately, the RiemannLiouville fractional derivative appears unsuitable for treatment by the Laplace transform
technique in that it requires the knowledge of the non-integer order derivatives of the function at t = 0. This problem
does not exist for the Caputo definition which is sometimes referred to as the smooth fractional derivative in the
literature.
Caputo definition:
t f (m) ( )
Z
1
d, m1<q <m
(m (t )q+1m

q
0 D t f (t) = dm
q) 0 (7)
f (t), q=m
dt m
where m is the first integer which is not less than q. It is found that the equations with RiemannLiouville operators
are equivalent to those with Caputo operators by the homogeneous initial conditions assumption [35]. The Laplace
transform of the Caputo fractional derivative is
n o n1
s q1k f (k) (0)
q
X
L 0 Dt f (t) = s q F(s) n 1 < q n N. (8)
k=0

Contrary to the case for the Laplace transform of the RiemannLiouville fractional derivative, only integer order
derivatives of function f appear in the Laplace transform of the Caputo fractional derivative. For zero initial
conditions, (8) reduces to
n o
q
L 0 D t f (t) = s q L{ f (t)}. (9)

A fractional order linear time invariant (FO-LTI) system can be represented in the following state-space form:
 q
D x = Ax + Bu
(10)
y = Cx

where x R n , u R r , y R p are states and input and output vectors of the system and A R nn , B R nr ,
C R pn , and q is the fractional commensurate order. Fractional order differential equations are at least as stable
as their integer order counterparts, because systems with memory are typically more stable than their memoryless
counterparts [36]. It has been shown that the autonomous system D q x = Ax, x(0) = x0 is asymptotically stable if
1302 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

Fig. 1. Stability region of the FO-LTI system with order 0 < q 1.

the following condition is satisfied [37]:

| arg(eig(A))| > q/2 (11)
where 0 < q < 1, and eig(A) represents the eigenvalues of matrix A. In this case, each component of the states decays
towards 0 like t q . Also, this system is stable iff | arg(eig(A))| q/2 and those critical eigenvalues which satisfy
| arg(spec(A))| = q/2 have geometric multiplicity 1. The stable and unstable regions for 0 < q 1 are shown in
Fig. 1.
Now, consider the following commensurate fractional order system:
D q x = f (x) (12)
where 0 < q < 1 and x R n . The equilibrium points of system (12) are calculated by solving the following equation:
f (x) = 0. (13)
These points are locally asymptotically stable if all the eigenvalues of the Jacobian matrix J = f / x evaluated at
the equilibrium points satisfy [36,37]
| arg(eig(J ))| > q/2. (14)

3. Numerical simulation of a fractional order system

Unlike the numerical solving of an ordinary differential equation, the numerical simulation of a fractional order
system is not unproblematic. In the literature of the fractional chaos field, two approximation methods have been
proposed for finding the response of a fractional order system numerically. The first method is an improved version of
the AdamsBashforthMoulton algorithm and is proposed on the basis of the predictorcorrector scheme [22,23]. To
explain the method we consider the following differential equation:
q
Dt y(t) = r (t, y(t)), 0t T
(k) (k)
(15)
y (0) = y0 , k = 0, 1, . . . , m 1,
q
where Dt is the Caputo fractional derivative. This differential equation is equivalent to the Volterra integral equation
dqe1 k Z t
X (k) t 1
y(t) = y0 + (t s)q1r (s, y(s))ds. (16)
k=0
k! (q) 0

By setting h = T /N and tn = nh (n = 0, 1, . . . , N ), (16) is discretized as follows:

dqe1 k n
X (k) t hq p hq X
yh (tn+1 ) = y0 + r (tn+1 , yh (tn+1 )) + a j,n+1r (t j , yh (t j )) (17)
k=0
k! (q + 2) (q + 2) j=0
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1303

where
dqe1 k n
p
X (k) t 1 X
yh (tn+1 ) = y0 + b j,n+1r (t j , yh (t j ))
k=0
k! (q) j=0

and
(n q)(n + 1)q ,
q+1
n j =0
a j,n+1 = (n j + 2)q+1 + (n j)q+1 2(n j + 1)q+1 , 1 j n

1, j =n+1
h q
b j,n+1 = ((n + 1 j)q (n j)q ).
q
The approximation error is determined as follows:
max |y(t j ) yh (t j )| = O(h p ) (18)
j=0,1,...,N

where p = min(2, 1 + q). The numerical solution of a fractional order system described by (12) can be obtained by
applying the method mentioned.
The second method, known as the frequency domain approximation, is based on the approximation of the fractional
order system behavior in the frequency domain. In [22], an algorithm has been proposed for calculating transfer
function approximations of 1/s q . In this method, the aim is to find zeros and poles of a transfer function that has
a similar amplitude diagram to 1/s q in a given frequency range. 1/s q has a Bode diagram characterized by a slope
of 20q dB/decade. Therefore, in this method the 20q dB/decade line is approximated by a number of zigzag
lines connected together with alternate slopes of 0 dB/decade and 20 dB/decade. According to this method, we
can obtain a linear approximation of the fractional order integrator with any desired accuracy over any frequency
band. The order of this linear approximation system depends on the desired bandwidth and accuracy. This method is
known as Charefs method. Oustaloups method [30] is very close to Charefs method. Carlsons method is another
frequency domain approximation that is introduced based on Newtons iterative method for root finding [28]. Also,
Matsudas method provides continuous approximations of a fractional order system obtained by CFE and identifying
a model from its amplitude [29]. To simulate a fractional order system by using frequency domain approximations,
first the fractional order equations of the system are considered in the frequency domain and the fractional integral
operator is replaced by its integer order approximation. Then the approximated equations in the frequency domain are
transformed into the time domain. The resulting ordinary differential equations can be numerically solved using well
known numerical methods like the RungeKutta algorithm and AdamsBashforthMoulton algorithm.
Due to the specificity of the error estimation bound in the first method, simulation results obtained by this method
are more reliable than the simulation results of the second method. In Section 5, we show that using the frequency
domain approximation in the numerical simulation may result in erroneous conclusions.

4. Analysis of the frequency based approximate model

In this section, the frequency based approximation model is analyzed and compared with the original system. In
the first part, the location of the fixed points of the approximated model is compared with that for the original system
and in the second part the stability of the fixed points in the approximated model is studied.

4.1. Fixed points

Suppose the original system is a commensurate fractional order system and described by the following equation:
D q x = f (x) (19)
where x = [x1 , x2 , . . . , xn ]T R n , f (x) = [ f 1 (x), f 2 (x), . . . , f n (x)]T and 0 < q < 1. The fixed points of system
(19) can be achieved by solving Eq. (13). To simulate the above system using the frequency domain approximations,
1304 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

the operator 1/s q is replaced by its integer order approximation. Assume that the approximating filter for operator
1/s q is given as
1 b p s p + + b1 s + b0
q
m . (20)
s s + + a1 s + a0
Then, the approximating model for the system in (19) is defined by the following equation
dm d dp d
x + + a 1 x + a 0 x = b p f (x) + + b1 f (x) + b0 f (x). (21)
dt m dt dt p dt
The fixed points of Eq. (21) are determined by solving the following equation:
f (x) = g x (22)
where g = a0 /b0 is the inverse of the DC gain of the approximating filter. If the DC gain of the approximating filter
is infinite, the fixed points of the approximated system (21) are the same as the equilibria of the original system (19);
otherwise the fixed point locations of the original system and the approximated one are different. The extent of this
inconsistency is related to the DC gain of the approximating filter; the higher the DC gain, the less the difference.
Unfortunately, a few frequency domain approximations have infinite DC gain. Example 1 shows that this problem not
only displaces the locations of the fixed points, but also changes the number of equilibria.

Example 1. Consider the following system:

Dq x = x 3 . (23)
This system has only one fixed point xe = 0, but if it is simulated using any frequency domain approximation with

positive and finite DC gain, the approximated system will have three equilibria, i.e., xe = g, 0, g, where g is the
inverse of the DC gain of the approximating filter. Now, suppose we want to simulate the following system using the
previous frequency domain approximation:

D q x = x 3 + cx. (24)
If 0 < c < g, the approximated
system has only one equilibrium xe = 0, but the original system has three fixed
points, i.e., xe = c, 0, c.
When the original system is defined using a more general form of
D qi xi = f i (x1 , x2 , . . . , xn ) i = 1, 2, . . . , n, (25)
its equilibria are obtained by solving the following set of equations:
f i (x1 , x2 , . . . , xn ) = 0 i = 1, 2, . . . , n. (26)
However, the fixed points of the approximated model, obtained by frequency domain approximations, are given by
solving the following set of equations:
f i (x1 , x2 , . . . , xn ) = gi xi i = 1, 2, . . . , n. (27)
where gi (i = 1, 2, . . . , n) is the inverse of the DC gain of the filter that approximates 1/s qi .

4.2. Stability of fixed points

In the previous part, we showed that the location and number of fixed points of the approximated system achieved
by the frequency domain approximations may not be the same as those for the original system. In this part, the stability
of fixed points in the approximated system is investigated and compared with the stability of the original equilibria.
Without loss of generality, we assume that the approximating filter (20) is strictly proper and p = m1. For simplicity,
we begin with a fractional order linear time invariant system described by
D q x = Ax (28)
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1305

where x R n , 0 < q < 1, and A R nn . The approximated model using the approximating filter (20) is determined
as
dm dm1 d
m
x + (am1 I bm1 A) m1 x + + (a1 I b1 A) x + (a0 I b0 A)x = 0. (29)
dt dt dt
The high order descriptor system (29) can be realized by a first-order state-space model as
z = M z, (30)
where

0 In
..
.

M = (31)


In
B0 B1 Bm1
and Bk = ak In bk A for k = 0, 1, . . . , m 1 [38]. The equilibrium of the approximated system is asymptotically
stable if and only if
Re(eig(M)) < 0, (32)
whereas the stability condition for the original system is determined by (11). Example 2 shows that the frequency
approximation may change the stability of a fixed point.

Example 2. According to (11), the fixed point x = [0 0]T in the following autonomous system:
1
 
0.9
D x= x (33)
1
is asymptotically stable if and only if || > 6.314. The behavior of this system can be simulated by frequency domain
approximations. In Table I of [14], the integer order approximation of operator 1/s 0.9 with the maximum discrepancy
of 2 dB from = 102 to 102 rad/s is given as
1 1.766s 2 + 38.27s + 4.914
.
s 0.9 s 3 + 36.15s 2 + 7.789s + 0.01000
When one uses the above approximation, the fixed point of the approximated system is asymptotically stable if and
only if || > 6.924. Another integer order approximation for operator 1/s 0.9 with the maximum discrepancy of 2 dB
from = 102 to 102 rad/s is given in Table 1 of [16]. This approximation, found by Charefs method, is given as
follows:
1 2.2675(s + 1.292)(s + 215.4)
.
s 0.9 (s + 0.01292)(s + 2.154)(s + 359.4)
The above approximation converts the stability condition || > 6.314 for the original system into || > 11.969 for
the approximated model.
When the original fractional order system is nonlinear time invariant, the stability investigation of its equilibria is
more complicated than that for the linear case. At first, we begin with a nonlinear system with only one state:
D q x = f (x), x R. (34)
Using the approximating filter (20) and Eq. (21), the approximated system is described by the following relation:

dm d dm1 d
m
x + + a 1 x + a 0 x = b m1 m1
f (x) + + b1 f (x) + b0 f (x). (35)
dt dt dt dt
To write Eq. (35) in first-order state-space form, the derivatives of the function f with respect to t are transformed
into derivatives of x with respect to t. This is done using Fa`a di Brunos formula.
1306 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

Theorem 1 (Fa`a di Brunos Formula [39,40]). When f , x, and t are scalars and f (x(t)) is a composite function, for
which all the necessary derivatives are defined, the r th total derivative of the function f with respect to t is defined as
!k 1 !k 2 !k r
dr r! x (1) x (2) x (r )
f (k) (x)
X
f (x(t)) = (36)
dt r k1 !k2 ! kr ! x 1! 2! r!

where the sum is over all non-negative integer solutions of the Diophantine equation k1 + 2k2 + + r kr = r , and
(k)
k := k1 + k2 + + kr . Also, f x (x) and x (i) respectively mean the kth derivative of the function f with respect to
x and the ith derivative of the function x with respect to t.

For instance, when r = 3 all solutions in non-negative integers of the equation k1 + 2k2 + 3k3 = 3 should be found.
These solutions are (k1 , k2 , k3 ) = (0, 0, 1), (k1 , k2 , k3 ) = (1, 1, 0), and (k1 , k2 , k3 ) = (3, 0, 0) with the corresponding
values of k being 1, 2, and 3, respectively. Therefore, Fa`a di Brunos formula says that
 ...   3
d3
  
3! (1) x 3! (2) x x 3! (3) x
f (x(t)) = f (x) + f (x) + f (x)
dt 3 0!0!1! x 3! 1!1!0! x 1! 2! 3!0!0! x 1!
... ...
= f x(1) (x) x + 3 f x(2) (x)x x + f x(3) (x)( x )3 .

By applying Fa`a di Brunos formula, Eq. (35) is converted into first-order state-space form with m states
. . . , x (m1) . Define z i = x (i) , i = 0, 1, . . . , m 1; then Eq. (35) is described by the following state-space
x, x,
model:
z = z
0 1
z = z
1 2
..



.





z m2 = z m1

m1
X (37)
z m1 = ai z i + b0 f (z 0 )




i=0



m1  z k1  z k2  z ki
i!
f z(k0 1 +k2 ++ki ) (z 0 )
1 2 i
X X
.

+ bi



i=1 k1 +2k2 ++iki =i
k1 !k2 ! ki ! 1! 2! i!

At a fixed point xe , the derivatives x, . . . , x (m1) are equal to zero. Hence, the Jacobian matrix of system (37)
x,
evaluated at an equilibrium point xe (i.e. z 0 = xe and z i = 0 for i = 1, 2, . . . , m 1) is

0 1 0
.. .. .. ..
. . . .

J = , (38)

0 0 1
b0 d a0 b1 d a1 bm1 d am1

where d = d f /dx|x=xe . An equilibrium point of the original system (34) is asymptotically stable if condition
(14) is satisfied, whereas the stability condition for the approximated model, described by (37), is related to the
approximating filter parameters. The equilibrium point xe of approximated system (37) is asymptotically stable if and
only if Re(eig(J )) < 0.
When the original system (19) has more than one state, to convert the approximating model (21) into first-order
state-space form, we use the generalized Fa`a di Bruno formula.

Theorem 2 (Generalization of Fa`a di Brunos Formula [41]). If f and t are scalars, x(t) =
[x1 (t), x2 (t), . . . , xn (t)]T is an n-vector and f (x(t)) is a composite function, for which all the necessary derivatives
are defined, then
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1307

dr XXX X r! k f
f (x(t)) = p p p
dt r r r Q
n
x1 1 x2 2 xn n
(i!)ki
r
Q Q
0 1 2 qi j !
i=1 i=1 j=1
r
(i) (i)
(x1 )qi1 (x2 )qi2 (xn(i) )qin ,
Y
(39)
i=1
where the respective sums are over all non-negative integer solutions of the Diophantine equations:
X
k1 + 2k2 + + r kr = r
X 0
q11 + q12 + + q1n = k1
X 1
q21 + q22 + + q2n = k2
2
..
X.
qr 1 + qr 2 + + qr n = kr
r
and p j is the order of the partial derivative with respect to x j and k is the order of the partial derivative:
p j = q1 j + q2 j + + qr j , j = 1, 2, . . . , n,
k = p1 + p2 + + pn = k 1 + k 2 + + k r .

For instance, the fourth derivative of a scalar composite function f (x1 , x2 ) with respect to t is calculated as
follows [41]:
d4 4 f 4 4 f 4 f 4 f 4 f 4
f (x 1 , x 2 ) = x
1 + 4 x
1 x
3
2 + 6 x
2 2
1 x
2 + 4 x
3
1 x
2 + x2
dt 4 x14 x1 x23 x12 x22 x13 x2 x24
3 f 3 f 3 f 3 f 2 3 f
+6 x 2 x + 6
3 1 1
x
2
1 x
2 + 6 x
1 x
2
2 + 6 x
2 x
2 + 12 x1 x1 x2
x1 x12 x2 x1 x22 x23 x12 x2
3 f 2 f ... 2 f ... 2 f ... 2 f ...
+ 12 x1 x2 x2 + 4 2 x1 x 1 + 4 x1 x 2 + 4 x 1 x2 + 4 2 x2 x 2
x1 x2
2 x1 x1 x2 x1 x2 x2
2 f 2 2 f 2 2 f f (4) f (4)
+3 x
+ 3 x
+ 6 x1 x2 + x1 + x .
x1
2 1
x2
2 2
x1 x2 x1 x2 2
By applying the generalized Fa`a di Bruno formula, (21) can be written in first-order state-space form with n m
(m1) (m1) (m1)
states x1 , x1 , . . . , x1 , x2 , x2 , . . . , x2 , . . . , xn , xn , . . . , xn . Suppose that xe R n is a fixed point of (21),
n
or in other words, xe R is one of the solutions of (22). By justification similar to that given for the case n = 1, the
Jacobian matrix of the approximated system, evaluated at the equilibrium point xe , is obtained as follows:

0 In 0
.. .. .. ..
. . . .

J = (40)


0 0 In
b0 D a0 In b1 D a1 In bm1 D am1 In
where D = f / x|x=xe . Therefore, the equilibrium point xe of the approximated system is asymptotically stable if
and only if
Re(eig(J )) < 0. (41)
This condition is not always equivalent to the one given in (14) for the original system. Sometimes, this problem
can yield erroneous consequences in our simulations that are performed on the basis of the frequency domain
approximations. Some of the erroneous consequences that are observed in simulation of chaotic fractional order
systems are shown in the next section.
1308 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

5. Effects of frequency domain approximation on chaotic behavior

In this section, we first discuss a necessary condition for the existence of chaos in a fractional order system. Then,
the influences of the frequency domain approximation on elimination or creation of chaos in numerical simulation of
the fractional order systems are considered.
In a 3-D nonlinear dynamical system, a saddle point is an equilibrium point on which the equivalent linearized
model has at least one eigenvalue in the stable region and one in the unstable region. In the same system, a saddle
point is called a saddle point of index 1 if one of the eigenvalues is unstable and the others are stable. Also, a saddle
point of index 2 is a saddle point with one stable eigenvalue and two unstable ones. For chaotic systems, it is proved
that scrolls are generated only around the saddle points of index 2. Moreover, saddle points of index 1 are responsible
only for connecting scrolls [4245].

5.1. Chaotic systems displaying one-scroll attractors

Assume that a 3-D chaotic system x = f (x) displays a one-scroll attractor. Hence, this system has a saddle point
of index 2 encircled by a one-scroll attractor. Suppose = j are unstable eigenvalues for this saddle point of
index 2. A necessary condition for fractional system D q x = f (x) to remain chaotic is that the eigenvalue is kept in
the unstable region, i.e.,
  || 2
 
1 ||
tan q > q > tan . (42)
2
Otherwise, the fixed point becomes asymptotically stable and then attracts the nearby trajectories.
The Rossler system [46] is a 3-D nonlinear system that can exhibit chaotic behavior. The attractor of the Rossler
system belongs to the one-scroll chaotic attractors family. This system is described by

x = (y + z)
y = x + ay (43)
z = b + z(x c).

For c2 > 4ab, the Rossler system has two fixed points, i.e.,
!
c c2 4ab c + c2 4ab c c2 4ab
O1 = , ,
2 2a 2a
!
c + c2 4ab c c2 4ab c + c2 4ab
O2 = , , .
2 2a 2a

The Jacobian matrix of system (43), evaluated at (x , y , z ), is

0 1 1
J = 1 a 0 . (44)
z 0 x c
It is shown that the fractional order Rossler system can display chaotic behavior as well [19]. The fractional order
Rossler system is defined by the following equations:
dq x
= (y + z)
dt q



q
d y
= x + ay (45)
dt q

q
d z = b + z(x c).


dt q
When (a, b, c) = (0.35, 1, 10), the Rossler system exhibits chaotic behavior (Fig. 2(a)). For the given parameters, the
fixed points and their corresponding eigenvalues are
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1309

Fig. 2. Phase plots in the xy plane projection and in the xyz space: (a) Rossler system where (a, b, c) = (0.35, 1, 10), (b) fractional order
Rossler system where (a, b, c) = (0.35, 1, 10) and q = 0.95 (time domain approximation) and (c) fractional order Rossler system where
(a, b, c) = (0.35, 1, 10) and q = 0.95 (frequency domain approximation).

O1 = (0.0351, 0.1003, 0.1003) : 1 = 9.9549 2,3 = 0.1700 j0.9842

O2 = (9.9649, 28.4711, 28.4711) : 1 = 0.3370 2,3 = 0.0111 j5.4283.
Hence, the fixed point O1 is a saddle point of index 2. According to (42), for q > 0.89, the fractional order Rossler
system with parameter set (a, b, c) = (0.35, 1, 10) has the necessary condition for remaining chaotic. Fig. 2(b) shows
the phase plots of the fractional order Rossler system for q = 0.95, when numerical simulation has been done on the
basis of the predictorcorrector scheme, described in Section 3. Simulation results confirm the existence of chaotic
behavior in this system.
The same system is also simulated using Charefs method in the frequency domain approximation. On the basis of
this method, 1/s 0.95 is replaced by the following transfer function [19] which has the maximum discrepancy of 1 dB:
1 1.2831s 2 + 18.6004s + 2.0833
.
s 0.95 s 3 + 18.4738s 2 + 2.6574s + 0.003
When numerical simulations are done on the basis of the above approximating filter, chaotic behavior disappears
(Fig. 2(c)). This problem is the direct consequence of undesired influences of the frequency domain approximation.

5.2. Chaotic systems displaying two-scroll attractors

When a 3-D chaotic system x = f (x) generates a double-scroll attractor, it has two saddle points of index 2,
surrounded by scrolls. Assume that i = i ji (i = 1, 2) are unstable eigenvalues of these saddle points of index
2. A necessary condition for fractional system D q x = f (x) to exhibit two-scroll chaotic attractor is the eigenvalues
1310 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

i (i = 1, 2) remaining in the unstable region, i.e.,

  | | 2
 
i 1 |i |
tan q > q > tan , i = 1, 2. (46)
2 i i

The Chen and Lu systems are samples of chaotic systems with double-scroll attractors. We continue by studying
the fractional order Chen and Lu systems and investigating the influences of frequency domain approximation in the
simulation of these systems.
The Chen system, introduced by Chen and Ueta in 1999 [47], is similar but not topologically equivalent to the
Lorenz system and displays more complicated dynamical behaviors. This system is described by

x = a (y x)

y = (c a)x x z + cy (47)
z = x y bz.

When (2c a)b > 0 this system has three fixed points, i.e.,

O = (0, 0, 0)
p p
C + = ( (2c a)b, (2c a)b, 2c a)
p p
C = ( (2c a)b, (2c a)b, 2c a).

The Jacobian matrix of the system evaluated at (x , y , z ) is

a a 0
J = c a z c x . (48)
y x b

The fractional order Chen system is described as follows:

dq x
= a (y x)
dt q




q
d y
= (c a)x x z + cy (49)

dt q
q
d z = x y bz.


dt q
In [17], chaotic behavior has been reported in a fractional order Chen system for parameter set (a, b, c) = (40, 3, 28)
and as low orders as q = 0.1, 0.2 and 0.3. For this parameter set, the fixed points and their corresponding eigenvalues
are

O = (0, 0, 0) : 1 = 32 2 = 20 3 = 3
C = (6.9282, 6.9282, 16) : 1 = 20.2304 2,3 = 2.6152 j13.5268.

Then, according to (46), the maximum fractional order q for which the fractional order Chen system remains chaotic
is q 0.88 and this is inconsistent with [17]. In other words, for q < 0.88, saddle points of index 2 become
asymptotically stable. It seems use of the frequency approximation for numerical simulation in [17] causes incorrect
outcomes. In this case, numerical simulation results based on the predictorcorrector scheme are given in Fig. 3 for
q = 0.88 (chaotic behavior) and in Fig. 4 for q = 0.87 (non-chaotic behavior).
In the following, we analytically investigate how use of the frequency domain approximations affects the simulation
results. Numerical simulations in [17] were done on the basis of the frequency domain approximations. The reference
of this paper for frequency domain approximations is Table I of [14]. This table has given approximations for 1/s q
with q = 0.1 0.9 in steps of 0.1. These approximations, obtained by trial and error, have a maximum discrepancy of
2 dB in the = 102 to 102 rad/s frequency band. In this table, the integer order approximations of operators 1/s 0.3 ,
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1311

Fig. 3. Simulation results based on the predictorcorrector scheme for a fractional order Chen system: (a, b, c) = (40, 3, 28) and q = 0.88.

Fig. 4. Simulation results based on the predictorcorrector scheme for a fractional order Chen system: (a, b, c) = (40, 3, 28) and q = 0.87.
1312 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

Table 1
Fixed points and corresponding eigenvalues of these points in the approximate model of a fractional order Chen system

Commensurate order of original system Fixed points Eigenvalues

1 = 420.8, 2 = 815.7, 3 = 9.4750, 4 = 8.5250
0
5 = 128.6, 6 = 0.5896, 7 = 0.5617, 8 = 6.6559
O = 0
9 = 0.4874, 10 = 0.0373, 11 = 0.0365, 12 = 0.0341
0
13 = 0.0024, 14 = 0.0024, 15 = 0.0024
q = 0.3

1 = 543.9, 2 = 8.358, 3 = 0.556, 4 = 0.0363

7.1388
5,6 = 0.0024 j0.0000, 7,8 = 8.901 j0.810
C = 7.1773
9,10 = 0.5751 j0.0239, 11,12 = 0.0369 j0.0006
15.9342
13 = 0.0024, 14,15 = 10.99 j327.3

1 = 1099.0, 2 = 2071.3, 3 = 12.335, 4 = 13.232

0
5 = 3057.3, 6 = 10.457, 7 = 0.7300, 8 = 0.7006
O = 0
9 = 0.0397, 10 = 0.0405, 11 = 0.6252, 12 = 0.0370
0
13 = 0.0022, 14 = 0.0022, 15 = 0.0022
q = 0.2

1 = 1387.4, 2 = 12.179, 3 = 0.6951, 4 = 0.0395

7.2823
5,6 = 12.708 j0.7544, 7,8 = 0.7145 j0.0248
C = 7.3495
9,10 = 0.0401 j0.0008, 11,12 = 0.0022 j0.0000
15.8859
13 = 0.0022, 14,15 = 55.50 j852.8

1 = 4071.4, 2 = 7390.1, 3 = 1000.2, 4 = 21.397

0
5 = 22.107, 6 = 0.9956, 7 = 1.0207, 8 = 19.751
O = 0
9 = 0.9316, 10 = 0.0463, 11 = 0.0471, 12 = 0.0441
0
13 = 0.0022, 14 = 0.0021, 15 = 0.0020
q = 0.1

1 = 4936.7, 2 = 21.271, 3 = 0.9910, 4 = 0.0462

7.3411
5,6 = 21.713 j0.5945, 7,8 = 1.0074 j0.0211
C = 7.4207
9,10 = 0.0467 j0.0006, 11,12 = 0.0022 j0.0000
15.8652
13 = 0.0021, 14,15 = 309.6 j3103.3

1/s 0.2 and 1/s 0.1 are given as

1 23.76s 4 + 224.9s 3 + 129.1s 2 + 4.733s + 0.01052

s 0.3 s 5 + 64.51s 4 + 252.2s 3 + 63.61s 2 + 1.104s + 0.002267
1 60.95s 4 + 816.9s 3 + 582.8s 2 + 23.24s + 0.04934

s 0.2 s 5 + 134.0s 4 + 956.5s 3 + 383.5s 2 + 8.953s + 0.01821
1 220.4s 4 + 5004s 3 + 5038s 2 + 234.5s + 0.4840
0.1
5 .
s s + 359.8s 4 + 5742s 3 + 4247s 2 + 147.7s + 0.2099
As mentioned in Section 4, use of the above approximations can lead to undesired effects in the location and stability
of the fixed points. When one uses these approximating filters, the approximated model becomes of order 15. Table 1
presents fixed points of the approximated model and their corresponding eigenvalues. In this case, the equilibria have
been determined using (22) and their corresponding eigenvalues have been obtained by calculating the eigenvalues of
Jacobian matrix J defined in (40).
It is observed that the locations of fixed points C are slightly different from the locations of the original equilibria.
Also, two of the corresponding eigenvalues of these fixed points (14 and 15 ) are unstable and then the fixed points
C are susceptible to surrounding by a chaotic double-scroll attractor, whereas the original equilibria C are stable
and, therefore, cannot be surrounded by a chaotic double-scroll attractor. Fig. 5 shows numerical simulation results
based on the frequency domain approximation when (a, b, c) = (40, 3, 28) and q = 0.3.
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1313

Fig. 5. Simulation results for a fractional order Chen system: (a, b, c) = (40, 3, 28) and q = 0.3 (frequency domain approximation).

The Lu system is another chaotic system with a double-scroll attractor, whose fractional order form has been
studied in some of the literature. The Lu system, found by Lu and Chen, is known as a bridge between the Lorenz and
the Chen systems and represents the transition from one to the other [48]. The Lu system is described by

x = a (y x)

y = x z + cy (50)
z = x y bz.

For bc > 0, the Lu system has three equilibria, i.e.,

O = (0, 0, 0)

C + = ( bc, bc, c) (51)

C = ( bc, bc, c).

The Jacobian matrix of system (50), evaluated at (x , y , z ), is

a a 0
J = z c x . (52)
y x b

For (a, b, c) = (35, 3, 28), this system demonstrates chaotic behavior. In this case, the fixed points and their
corresponding eigenvalues are as follows:

O = (0, 0, 0) : 1 = 35 2 = 3 3 = 28
C
= (9.1652, 9.1652, 28) : 1 = 19.7536 2,3 = 4.8768 j16.5495.
1314 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

Fig. 6. Simulation results based on the predictorcorrector scheme for a fractional order Lu system: (a, b, c) = (35, 3, 28) and q = 0.82.

The fractional order Lu system is described as follows:

dq x
q = a (y x)



dt
q
d y
= x z + cy (53)

dt q
q
d z = x y bz.


dt q
From (46), the minimum order that makes the fractional order Lu system behave chaotically is calculated as
q 0.81 for (a, b, c) = (35, 3, 28). Figs. 6 and 7 respectively show numerical simulation results based on the
predictorcorrector scheme, for q = 0.82 and q = 0.80. When q = 0.82, the fractional order Lu system is chaotic
and for q = 0.80 it is not. For q < 0.81, the fixed points C are stable. Unfortunately, this is not in agreement with
the conclusions of [18]. In [18], a chaotic attractor has been reported for the Lu system for (a, b, c) = (35, 3, 28) and
q = 0.1, 0.2, 0.3. The approximating filters that were used in [18] are the same as those in [17]. The inconsistency
mentioned arises due to usage of the frequency domain approximation in the numerical simulations. The calculated
fixed points of the approximated model and their corresponding eigenvalues are given in Table 2. It is observed that the
fixed points C of the approximated model are unstable even though the eigenvalues of the original fixed points C
settle in the stable region. Hence, the original fractional order system is not chaotic, whereas the approximated model
is. In Fig. 8, simulation results based on the frequency domain approximation are shown for (a, b, c) = (35, 3, 28)
and q = 0.2.

5.3. Chaotic systems displaying multi-scroll attractors

Recently, design of a chaotic system for generating multi-scroll chaotic attractors has been extensively studied [45,
4951]. Also, some papers introduced fractional order chaotic systems to generate multi-scroll chaotic attractors
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1315

Fig. 7. Simulation results based on the predictorcorrector scheme for a fractional order Lu system: (a, b, c) = (35, 3, 28) and q = 0.80.

Fig. 8. Simulation results for a fractional order Lu system: (a, b, c) = (35, 3, 28) and q = 0.2 (frequency domain approximation).
1316 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

Table 2
Fixed points and corresponding eigenvalues of these points in the approximate model of a fractional order Lu system

Commensurate order of original system Fixed points Eigenvalues

1 = 610.7, 2 = 886.9, 3 = 8.5517,
4 = 0.5626, 5 = 0.0365, 6 = 0.0024,

0
O = 0 7 = 128.6, 8 = 9.2845, 9 = 6.656,
0 10 = 0.5844, 11 = 0.0371, 12 = 0.4874
13 = 0.0341, 14 = 0.0024, 15 = 0.0024

q = 0.3 1 = 533.2, 2 = 8.348, 3 = 0.5560,

4 = 0.0363, 5,6 = 0.5765 j0.0188,
7,8 = 8.971 j0.6417,
9.4520
9,10 = 0.0023 j0.0000,
C = 9.5102
11,12 = 0.0369 j0.0005
27.9556
5,6 = 0.5765 j0.0188, 7,8 = 8.971 j0.6417
9,10 = 0.0023 j0.0000, 11,12 = 0.0369 j0.0005
13 = 0.0024, 14,15 = 65.08 j401.4

1 = 1586.4, 2 = 2254.1, 3 = 12.361, 4 = 0.7015

0
5 = 0.0397, 6 = 0.0022, 7 = 305.7, 8 = 10.457
O = 0
9 = 0.0370, 10 = 0.0022, 11 = 0.6252, 12 = 13.054
0
13 = 0.7244, 14 = 0.0404, 15 = 0.0022
q = 0.2

1 = 1360.8, 2 = 12.170, 3 = 0.6948, 4 = 0.0395

9.6483
5,6 = 12.766 j0.5957, 7,8 = 0.7159 j0.0195
C = 9.7501
9,10 = 0.0401 j0.0006, 11,12 = 0.0022 j0.0000
27.9223
13 = 0.0022, 14,15 = 194.6 j1046.1

1 = 5834.4, 2 = 8051.3, 3 = 0.9963, 4 = 0.0464

0
5 = 0.0021, 6 = 0.9956, 7 = 1000.3, 8 = 0.9316
O = 0
9 = 0.0441, 10 = 0.0020, 11 = 19.751, 12 = 21.969
0
13 = 0.0022, 14 = 0.0470, 15 = 1.0159
q = 0.1

1 = 4841.8, 2 = 21.264, 3 = 0.9907, 4 = 0.0462

9.7290
5,6 = 21.753 j0.4677, 7,8 = 1.0086 j0.0165
C = 9.8496
9,10 = 0.0468 j0.0005, 11,12 = 0.0022 j0.0000
27.9079
13 = 0.0021, 14,15 = 813.2 j3807.4

(for example [52]). The necessary condition for having multi-scroll chaotic attractors in these systems is similar to the
necessary condition discussed in Sections 5.1 and 5.2. It is shown here that using the frequency domain approximation
to simulate these system dynamics can lead to different and fallacious results as well.
Consider the following system:

x = y
y = z (54)
z = ax by cz + d f (x; k, h, p, q)

where f is called a saturated function series and described by

(2q + 1)k, x > qh + 1



k(x i h) + 2ik|x i h| 1, p i q

f (x; k, h, p, q) =
(2i + 1)k, i h + 1 < x < (i + 1)h 1, p i q 1
x < ph 1.

(2 p + 1)k,

The parameters k > 0 and h > 2 are respectively the slope and saturated delay time of the saturated function series.
Also, p and q are positive integers. For some appropriately chosen parameter set a, b, c, d, k, h, system (54) possesses
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1317

Fig. 9. Simulation results for a fractional order Rossler system: (a) maximum discrepancy of 3 dB, (b) maximum discrepancy of 2.8 dB, (c)
maximum discrepancy of 2.75 dB, and (d) maximum discrepancy of 2.7 dB.

( p + q + 2) saddle points of index 2 and ( p + q + 1) saddle points of index 1, and it has the potential to create a
maximum of ( p + q + 2)-scroll chaotic attractors [45]. In fact, the ( p + q + 2) saddle points of index 2 are responsible
for generating scrolls of the attractor and the ( p + q + 1) saddle points of index 1 are responsible for connecting these
( p + q + 2) scrolls to form a whole chaotic attractor [45]. As an example, where a = b = c = d = 0.7, k = 9, h = 18
and p = q = 2, system (54) generates a six-scroll chaotic attractor. Now, consider the fractional order counterpart of
the system (54):
q
d 1x
q =y

dtq 1



d 2y

=z (55)
dt q2

q3
d z = ax by cz + d f (x; k, h, p, q).



dt q3
Use of the frequency domain approximation in the simulation of the above system can produce undesired effects,
such as change in location and number of fixed points and also the stability of these points. For example when
a = b = c = d = 0.7, k = 9, h = 18 and p = q = 2, system (55) has 11 fixed points. However, when
q1 = q2 = q3 = 0.1, if one uses the frequency approximation of operator 1/s 0.1 given in Table I of [14] to simulate
the system, according to (22) the number of equilibrium points in the approximated system becomes 3. As another
observation of undesired effects of the frequency domain approximation on chaotic behavior, consider system (55) in
which q1 = 1, q2 = 1, q3 = 0.85. Figs. 9 and 10 show simulation results for this system based on frequency domain
approximations with different accuracy. In this simulation Charefs method has been used to approximate operator
1/s 0.85 in the frequency range = 102 to 102 rad/s. One can observe different behaviors such as a six-scroll
attractor, one-scroll attractor, period-3 orbit, period-1 orbit and convergence to a fixed point.
1318 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

Fig. 10. Simulation results for a fractional order Rossler system: (a) maximum discrepancy of 2.65 dB, (b) maximum discrepancy of 2 dB,
(c) maximum discrepancy of 1 dB, and (d) maximum discrepancy of 0.5 dB.

Due to the above observations, before simulating any fractional order system using frequency domain
approximation, the following questions must be answered: For reliability of the results, what accuracy is required?
and Which frequency range must be selected?. Regrettably, most of the existing papers that have used the frequency
domain approximation in their simulations did not answer these important questions.

6. Conclusions

It is well known that the frequency domain approximation of fractional order operators is a very good candidate
for realizing fractional order controllers [53]. Also, these approximation methods provide a simple procedure for
simulating the fractional order systems numerically. This technique has been greatly used to investigate chaos in the
fractional order systems as well. However, some limitations should be considered in the use of this technique. A
number of limitations arise from the approximating nature of these techniques. For example, none of the frequency
based approximations can demonstrate the role of history or long memory existing in the fractional order systems.
Furthermore, any approximating filter with finite DC gain changes the location of the fixed points. Also, in Section 4
it was shown that the stability of the fixed points in a frequency based approximated model is not necessarily the same
as the stability of the original fixed points.
In this paper, we also showed that two kinds of inaccuracies are expected, in detecting chaos via frequency domain
based numerical simulations:
The approximated model obtained by frequency domain methods exhibits chaos whereas the original system is not
actually chaotic.
The original system is chaotic but its frequency domain based approximation does not show chaotic behavior.
A mistake has occurred in the recent literature [17,18] that found the lowest order chaotic systems among fractional
order systems. In this work, we only considered the undesired effects of using frequency domain approximation,
which is the most popular method in the literature of the fractional chaos field. There are other methods for simulating
 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320 1319

fractional order systems, and some of them such as the improved version of the AdamsBashforthMoulton algorithm
(used in [21,2426]) or methods in the form of FIR filters (described in [54]) work more accurately. The problems of
using frequency domain approximations in simulating and detecting chaos in fractional order systems are very similar
to the limitations of using the function method described for limit cycle prediction [55,56].
Therefore, the outcome of this paper can be summarized in one sentence: Results of numerical simulations of
fractional order systems done on the basis of frequency approximation methods are not always reliable, especially in
detecting chaotic behavior in these systems.

References

[1] V.G. Jenson, G.V. Jeffreys, Mathematical Methods in Chemical Engineering, 2nd ed., Academic, New York, 1977.
[2] H.H. Sun, A.A. Abdelwahad, B. Onaral, Linear approximation of transfer function with a pole of fractional order, IEEE Trans. Automat.
Control 29 (1984) 441444.
[3] M. Ichise, Y. Nagayanagi, T. Kojima, An analog simulation of non-integer order transfer functions for analysis of electrode process, J. Electro-
Anal. Chem. 33 (1971) 253265.
[4] O. Heaviside, Electromagnetic Theory, Chelsea, New York, 1971.
[5] R.L. Bagley, R.A. Calico, Fractional order state equations for the control of visco-elastically damped structures, J. Guid. Control Dyn. 14
(1991) 304311.
[6] D. Kusnezov, A. Bulgac, G.D. Dang, Quantum levy processes and fractional kinetics, Phys. Rev. Lett. 82 (1999) 11361139.
[7] N. Laskin, Fractional market dynamics, Physica A 278 (2000) 482492.
[8] A.M.A. El-Sayed, Fractional order diffusion wave equation, Internat. J. Theoret. Phys. 35 (2) (1996) 311322.
[9] P.J. Torvik, R.L. Bagley, On the appearance of the fractional derivative in the behavior of real materials, Trans. ASME 51 (1984) 294298.
[10] A. Oustaloup, X. Moreau, M. Nouillant, The CRONE suspension, Control Eng. Pract. 4 (8) (1996) 11011108.
[11] B.J. Lurie, Tunable TID controller, US patent 5, 371, 670, December 6, 1994.
[12] I. Podlubny, Fractional order systems and PI D controllers, IEEE Trans. Automat. Control 44 (1) (1999) 208214.
[13] H.F. Raynaud, Inoh A. Zerga, State space representation for fractional order controllers, Automatica 36 (2000) 10171021.
[14] T.T. Hartley, C.F. Lorenzo, H.K. Qammer, Chaos in a fractional order Chuas system, IEEE Trans. CAS-I 42 (1995) 485490.
[15] P. Arena, R. Caponetto, L. Fortuna, D. Porto, Chaos in a fractional order Duffing system, in: Proc. ECCTD, Budapest, 1997, pp. 12591262.
[16] W.M. Ahmad, J.C. Sprott, Chaos in fractional order autonomous nonlinear systems, Chaos Solitons Fractals 16 (2003) 339351.
[17] J.G. Lu, G. Chen, A note on the fractional order Chen system, Chaos Solitons Fractals 27 (3) (2006) 685688.
[18] J.G. Lu, Chaotic dynamics of the fractional order Lu system and its synchronization, Phys. Lett. A 354 (4) (2006) 305311.
[19] C. Li, G. Chen, Chaos and hyperchaos in the fractional order Rossler equations, Physica A 341 (2004) 5561.
[20] J.G. Lu, Chaotic dynamics and synchronization of fractional order Arneodos systems, Chaos Solitons Fractals 26 (4) (2005) 11251133.
[21] L.J. Sheu, H.K. Chen, J.H. Chen, L.M. Tam, W.C. Chen, K.T. Lin, Y. Kang, Chaos in the NewtonLeipnik system with fractional order, Chaos
Solitons Fractals (2006) (in press).
[22] K. Diethelm, N.J. Ford, A.D. Freed, A predictorcorrector approach for the numerical solution of fractional differential equations, Nonlinear
Dynam. 29 (2002) 322.
[23] K. Diethelm, N.J. Ford, A.D. Freed, Detailed error analysis for a fractional Adams method, Numer. Algorithms 36 (2004) 3152.
[24] J.H. Chen, W.C. Chen, Chaotic dynamics of the fractionally damped van der Pol equation, Chaos Solitons Fractals (2006) (in press).
[25] L.J. Sheu, H.K. Chen, J.H. Chen, L.M. Tam, Chaos in a new system with fractional order, Chaos Solitons Fractals 31 (2) (2007) 12031212.
[26] W.C. Chen, Nonlinear dynamics and chaos in a fractional order financial system, Chaos Solitons Fractals (2006) (in press).
[27] B.M. Vinagre, I. Podlubny, A. Hernandez, V. Feliu, Some approximations of fractional order operators used in control theory and applications,
Fract. Calc. Appl. Anal. 3 (3) (2000) 945950.
[28] G.E. Carlson, C.A. Halijak, Approximation of fractional capacitors (1/s)1/n by a regular Newton process, IRE Trans. Circuit Theory CT-11
(2) (1964) 210213.
[29] K. Matsuda, H. Fujii, H optimized wave-absorbing control: Analytical and experimental results, J. Guid. Control Dyn. 16 (6) (1993)
11461153.
[30] A. Oustaloup, La Commande CRONE, Editions Hermes, Paris, 1991.
[31] A. Oustaloup, F. Levron, B. Mathieu, F.M. Nanot, Frequency-band complex noninteger differentiator: Characterization and synthesis, IEEE
Trans. Circuits Syst. I 47 (1) (2000) 2539.
[32] A. Charef, H.H. Sun, Y.Y. Tsao, B. Onaral, Fractal system as represented by singularity function, IEEE Trans. Automat. Control 37 (9) (1992).
[33] A. Charef, Analogue realization of fractional order integrator, differentiator and fractional PI D controller, IEE Proc. Control Theory Appl.
153 (6) (2006) 714720.
[34] J.S.H. Tsai, T.H. Chien, S.M. Guo, Y.P. Chang, L.S. Shieh, State-space self-tuning control for stochastic chaotic fractional-order systems,
IEEE Trans. Circuits Syst. I 54 (3) (2007) 632642.
[35] I. Podlubny, Fractional Differential Equations, Academic Press, New York, 1999.
[36] E. Ahmed, A.M.A. El-Sayed, H.A.A. El-Saka, Equilibrium points, stability and numerical solutions of fractional order predatorprey and
rabies models, J. Math. Anal. Appl. 325 (1) (2007) 542553.
[37] D. Matignon, Stability results for fractional differential equations with applications to control processing, in: Computational Engineering in
Systems Applications, Lille, France, IMACS, IEEE-SMC, vol. 2, July 1996, pp. 963968.
1320 M.S. Tavazoei, M. Haeri / Nonlinear Analysis 69 (2008) 12991320

[38] G.R. Duan, Parametric control systems design: Theory and applications, SICE-ICASE International Join Conference, Korea, October 1821,
2006, pp. I-15: I-23.
[39] S. Roman, The Formula of Faa di Bruno, Amer. Math. Monthly 87 (1980) 805809.
[40] W.P. Johnson, The curious history of Fa`a di Brunos formula, Amer. Math. Monthly 109 (2002) 217234.
[41] R.L. Mishkov, Generalization of the formula of Faa di Bruno for a composite function with a vector argument, Internat. J. Math. Math. Sci.
24 (7) (2000) 481491.
[42] L.O. Chua, M. Komuro, T. Matsumoto, The double scroll family, IEEE Trans. Circuits Syst. 33 (1986) 10721118.
[43] C.P. Silva, Shilnikovs theoremA tutorial, IEEE Trans. Circuits Syst. I 40 (1993) 675682.
[44] D. Cafagna, G. Grassi, New 3-D-scroll attractors in hyperchaotic Chuas circuit forming a ring, Internat. J. Bifur. Chaos 13 (10) (2003)
28892903.
[45] J. Lu, G. Chen, X. Yu, H. Leung, Design and analysis of multi-scroll chaotic attractors from saturated function series, IEEE Trans. Circuits
Syst. I 51 (12) (2004) 24762490.
[46] O.E. Rossler, An equation for continuous chaos, Phys. Lett. A 57 (5) (1976) 397398.
[47] G. Chen, T. Ueta, Yet another attractor, Internat. J. Bifur. Chaos 9 (1999) 14651466.
[48] J. Lu, G. Chen, A new chaotic attractor coined, Internat. J. Bifur. Chaos 12 (3) (2002) 659661.
[49] S.M. Yu, J.H. Lu, W.K.S. Tang, G. Chen, A general multiscroll Lorenz system family and its DSP realization, Chaos 16 (2006) 110.
[50] J. Lu, G. Chen, Generating multiscroll chaotic attractors: Theories, methods and applications, Internat. J. Bifur. Chaos 16 (2006) 775858.
[51] J. Lu, S. Yu, H. Leung, G. Chen, Experimental verification of multi-directional multi-scroll chaotic attractors, IEEE Trans. Circuits Syst.-I 53
(1) (2006) 149165.
[52] W.M. Ahmad, Generation and control of multi-scroll chaotic attractors in fractional order systems, Chaos Solitons Fractals 25 (2005) 727735.
[53] I. Petras, I. Podlubny, P. OLeary, L. Dorcak, B. Vinagre, Analogue Realization of Fractional Order Controllers, FBERG, Technical University
of Kosice, Kosice, 2002.
[54] I. Petras, A note on the fractional-order Chuas system, Chaos Solitons Fractals (2006) (in press).
[55] S. Kou, K. Han, Limitations of the describing function method, IEEE Trans. Automat. Control 20 (2) (1975) 291292.
[56] S. Engelberg, Limitations of the describing function for limit cycle prediction, IEEE Trans. Automat. Control 47 (11) (2002) 18871890.