Proceedings of the 2nd IFAC

Workshop on Fractional Differentiation and its Applications

Porto, Portugal, July 19-21, 2006

FRACTIONAL CENTRAL DIFFERENCES AND DERIVATIVES

Manuel Duarte Ortigueira

UNINOVA and Department of Electrical Engineering

Faculdade de Cincias e Tecnologia da Universidade Nova de Lisboa
Campus da FCT da UNL, Quinta da Torre, 2825 114 Monte da Caparica, Portugal
Phone. +351 21 2948520, Fax. +351 21 2957786, E-mails: [email protected] and [email protected]

Abstract: Fractional central differences and derivatives are proposed in this paper. The
positive even and odd integer orders differences and derivatives are generalised to real
orders, obtaining two new different types of differences and derivatives. For each type,
suitable integral formulation is obtained. Their computations of the integrals lead to
generalisations of the well known Riesz potentials. A study for coherence is done by
applying the definitions to functions with Fourier transform. Copyright 2006 IFAC

Keywords: fractional central difference; fractional central derivative; Grnwald-Letnikov;

generalized Cauchy derivative.

1. INTRODUCTION Here we present a similar procedure for central

fractional derivatives. We must refer that we will not
In previous papers, we proposed a new approach for address the existence problem. We are mainly
introducing the fractional derivatives based on interested in obtaining a generalisation of a well
several steps: known formulation. We will introduce the general
framework for the central differences, considering
a) Use as starting point the Grnwald- two cases that we will call type 1 and type 2. These
Letnikov forward and backward derivatives are generalisations of the usual central differences for
b) With an integral formulation for the even and odd positive orders respectively. These
fractional differences and using the differences lead to central derivatives similar to the
asymptotic properties of the Gamma usual Grnwald-Letnikov ones.
function obtain the generalised Cauchy
derivative. For those differences, integral representations will be
c) The computation of the integral defining the proposed. From these representations we obtain the
generalised Cauchy derivative is done with derivative integrals by using the properties of the
the Hankel path to obtain regularised Gamma function. The integration is performed over a
fractional derivatives. two straight lines path that closes at infinite. Two
d) The application of these regularised generalisations of the usual Cauchy derivative
derivatives to functions with Laplace definition are obtained that agree with it when is
transform, we obtain the Liouville fractional even or odd positive integer. The computation of
derivative. those integrals over the referred path leads to
generalisations of the Riesz potentials. The most
interesting feature lies in the summation formulae for valid for a+b+c+d > -1, it is not very hard to show
the Riesz potentials. that:
+
To test the coherence of the proposed definitions we f(t) = f(t) (8)
c1 c1 c1
apply them to the complex exponential. The results and
show that they are suitable for functions with Fourier +
transform. The formulation agrees also with the f(t) = - f(t) (9)
c2 c2 c1
Okikiolu (1966) studies.
while
The paper outline is as follows. In section 2 we +
f(t) = f(t) (10)
present the type 1 and type 2 central differences and c2 c1 c2
their integral representations. Central derivative provided that + > -1. In particular, + = 0, and
definitions similar to Grnwald-Letnikov ones are the relations (8) and (9) show that when || < 1 and
presented in section 3 and their integral || < 1 the inverse differences exist and can be
representations obtained generalising the Riesz obtained by using formulae (1) and (2). We must
potentials. In section 4 we apply the definitions to the remark that the zero order difference is the identity
complex exponential to test the coherence of the operator and is obtained from (1). The zero order
definitions. At last we present some conclusions. type 2 difference will be considered later.

2.2 Integral representations.

2. FRACTIONAL CENTRAL DIFFERENCES
Let us assume that f(z) is analytic in a region of the
2.1 Definitions complex plane that includes the real axis. To obtain
the integral representations for the previous
We introduce two types of fractional central differences we follow here the procedure used in
differences. Let > -1, hR+ and f(t) a complex Ortigueira (2005b,c). We only have to give
variable function. We define a type 1 difference by: interpretations to (1) and (2) in terms of the residue
theorem. For the first case, we must remark that the
+ (-1)k(+1) poles must lie at nh. This leads easily to
f(t) = f(t-kh) (1)
c1 (/2-k+1) (/2+k+1)
-
f(t) =
and a type 2 by1: c1
-w w
+ (-1)k(+1) f(t-kh+h/2) h +1 h
f(t) =
c2 [(+1)/2-k+1][(-1)/2+k+1]
(2) (+1)
- 2jh f(z+w) -w w
dw (11)
Let = 2N, NZ+, in the first. We obtain: + +1 + +1
Cc h 2 h 2
2N +N (-1)k(2N)!
c f(t) = (N-k)! (N+k)!f(t-kh) (3) The integrand function has infinite poles at every nh,
1
-N with nZ. The integration path must consist of
that can be written as infinite lines above and below the real axis closing at
+N 2N the infinite. The easiest situation is obtained by
c f(t) = (-1)k N-kf(t-kh)
2N
(4) considering two straight lines near the real axis, one
1
-N above and the other below (see figure 1).
A close look into (4) shows that aside a factor (-1)N it
is the current 2N order central difference. Similarly, Regarding to the second case, the poles are located
if is odd (=2N+1), the second is equal to current now at the half integer multiples of h, which leads to
central difference, aside the factor (-1)N+1. In fact, we
have: f(t) =
c2
2N+1 N+1 (-1)k(2N+1)! f(t-kh+h/2)
w 1 w 1
c f(t) = (5) - h +2 h +2
2 (N+1-k)! (N+k)!
-N (+1)
2jh f(z+w) -w w
dw (12)
and
+ +1 + +1
N+1 2N+1
Cc h 2 h 2
f(t) = (-1)k N+k f(t-kh+h/2)
2N+1
c (6) These integral formulations will be used in the
2
-N
following section to obtain the integral formulae for
In particular, with N=0, we obtain
1
the central derivatives.
c f(t) =f(t+h/2) - f(t-h/2).
2
With the following relation (Andrews et al, 1999)2:
+
3. FRACTIONAL CENTRAL DERIVATIVES
1
=
(a-k+1) (b-k+1) (c+k+1) (d+k+1) 3.1 Definitions
-
(a+b+c+d+1)
= (7) To obtain fractional central derivatives we proceed as
(a+c+1) (b+c+1) (a+d+1) (b+d+1)
usually (Ortigueira, 2005a,b; Samko et al, 1987):
divide the fractional differences by h (hR+) and let
1
Here we assume that is also non zero. h0. For the first case, and assuming again that >
2 -1, we obtain:
See page 123.

c1f(t) (+1) 1

Dc1f(t)
= lim = D =
c1 f(z+w) /2+1
2j /2 dw (21)
h0 h (w)
Cc
(-w) l r
(+1) + (-1)k and
lim f(t-kh) (13)
h0 h (/2-k+1) (/2+k+1)
- (+1) 1
2j
that we will call type 1 fractional central derivative. D = f(z+w) (+1)/2 dw
c2 (+1)/2
(w)l (-w)r
For the second case and assuming also that 0, we Cc
obtain the type 2 fractional central derivative given (22)
by for the type 2 derivative. Tthe subscripts l and r
mean respectively that the power functions have the
c2f(t)
f(t) = lim
Dc2 = left and right half real axis as branch cut lines. In
h0 h figure 3 we show the different segments used for the
(+1) + (-1)k f(t-kh+h/2) computation of the above integrals. For the type 1
lim
h0 h [(+1)/2-k+1] [(-1)/2+k+1] derivative, we have:
-

(14) (+1) 1
Formulae (13) and (14) generalise the positive
= -
2j f(z-x)
+1 -j/2 -j
e-j dx
integer order central derivatives to the fractional 1 0 x e e
case, although there should be an extra factor (-1)/2
(+1) 1
in the first case and (-1)(+1)/2 in the second case that 2j
j
= f(z-x) +1 j/2 j e dx
we removed. It is a simple task to obtain the x e e
4 0
derivative analogues to (8), (9), and (10):

+ (+1) 1
D D f(t) = D
c1 c1 c1
f(t) (15)
=
2j
f(z+x)
+1 j/2
dx
2 0 x e
and
+
D D f(t) = - D f(t) (16) (+1)e-j/2 1
c2 c2 c1
=- 2j
f(z+x) +1 -j/2 dx
while 3 0 x e
+ where the integers refer the straight line segment
D D f(t) = D f(t) (17)
c2 c1 c2 used in the computation. Joining the four integrals,
again with + > -1. we obtain:

(+1)sin(/2) 1
3.2 Integral formulae D =-
c1
f(z-x) +1dx
0 x
Let us assume that the distance between the

horizontal straight lines in figures 1 and 2 is 2(h) (+1)sin(/2) 1
that decreases to zero with h. To obtain the integral -

f(z+x) +1dx =
formulae for the derivatives we are going to permute 0 x
the limit and integral operations in (11) and (12).
(+1) sin(/2) 1
With this permutation we must compute the limit of =-

f(z-x) |x|+1 dx (23)
two quotients of Gamma functions. As it is well
-
(s+a)
known, the quotient of two gamma functions As is not an odd integer we obtain
(s+b)
has an interesting expansion (Samko et al, 1987): 1 1
2(-) cos(/2)
Dc1 = f(z-x)|x|+1 dx (24)
(s+a) a-b N
= s 1 + cks-k + O(s-N-1) (18) -
(s+b)
1 that is the so called Riesz potential (Samko et al.,
as |s|, uniformly in every sector that excludes the 1987;Oldham and Spanier, 1974; Kilbas et al., 2006).
negative real half-axis. The coefficients in the series In the last step we used the reflection formula of the
can be expressed in terms of Bernoulli polynomials, gamma function.
but their knowledge is not important here. So, in the
conditions stated above, when h is very small For the other case, we compute again the integrals
(w/h+a) corresponding to the four segments. We have:
(w/h)a-b [ 1 + h.(w/h)] (19)
(w/h+b)
(+1) 1
where is regular near the origin. Accordingly to the
= -
2j
f(z-x) +1
-j(+1)/2
e-j dx
above statement, the branch cut line used to define a x e
1 0
function on the right hand side in (16) is the negative
real half axis. Similarly, we (+1) 1

=
2j f(z-x) +1 j(+1)/2 ej dx
(-w/h+a) x e
(-w/h)a-b [ 1 + h.(-w/h)] (20) 4 0
(-w/h+b)
but now, the branch cut line is the positive real axis.
With these results, obtain for the type 1 case
 (-/2)n
(+1) 1 h n= n! un (28)

= 2j f(z+x) +1 dx
x ej(+1)/2 where un is the discrete unit step Heaviside function.
2 0

(+1)e-j/2 1 As the discrete-time Fourier transform of hn is:

=- 2j
f(z+x) x+1 e-j(+1)/2 dx , H(ej) = FT[hn] = (1-e-jh)/2 (29)
3 0 the Fourier transform of Rb(n) is
Joining the four integrals, we obtain: S(ej) = limjh(1-z-1)/2 (1-z)/2 =
ze
(+1)sin[(+1)/2] 1 = (1-e-jh)/2 (1-ejh)/2 =
Dc2 =
f(z-x) +1 dx

0 x = |ejh/2 - e-jh/2| = |2 sin(h/2)| (30)
So,
(+1)sin[(+1)/2] 1 + (-1)n(+1)
-

f(z+x) +1 dx |2 sin(h/2)|= (/2-n+1) (/2+n+1) ejnh (31)
0 x -
The last integral can be rewritten as: We write, then:

c1e-jt = e-jt |2 sin(h/2)| (32)
0
1 1 So, there is a linear system with frequency response

f(z+x) x+1 dx =
f(z-x)(-x)+1 dx given by:
0 - H1() = |2 sin(h/2)| (33)
we obtain that acts on a signal giving its central fractional
difference. Dividing by h (hR+) and computing the
1 sgn(x)
2(-)sin(/2)
D =- f(z-x) +1 dx (25) limit as h0, (33) gives:
c2
- |x| HD1() = || (34)
that is the modified Riesz potential (Samko et al, As is not an even integer:
1987). Both potentials (24) and (25) were studied
1 + (-1)n(+1)
also by Okikiolu (1966). These are essentially || = lim ejnh (35)
h0 h (/2-n+1) (/2+n+1)
convolutions of a given function with two acausal3 -
operators and are suitable for dealing with functions valid for > -1. The inverse Fourier Transform of
defined in R and that are not necessarily equal to || is given by (Okikiolu,1966):
zero at . In particular, they must be suitable for 1 --1
dealing with stationary stochastic processes. We do FT-1[||] = |t| (36)
2(-)cos(/2)
not go further in this way, by now.
and we obtain the impulse response:
1 --1
4. COEHRENCE OF THE RESULTS hD1(t) = |t| (37)
2(-)cos(/2)
leading to
4.1 Type 1 derivative
+
1
2(-)cos(/2)
--1
Dc1f(t) = f() |t-| d (38)
We want to test the coherence of the results by
considering functions with Fourier transform. To -
perform this study, we only have to study the that is coincides with (24).
behaviour of the defined derivatives for f(t) = e-jt,
t,R. In the following we will consider non integer 4.2 Type 2 derivative
orders greater than -1. We start by considering the
type 1 derivative. From (1) we obtain A similar procedure allows us to obtain

+ (-1)n(+1) c2e-jt =
c1ejt= e-jt ejnh (26)
(/2-n+1) (/2+n+1) + (-1)k(+1)
- e-jt e-jh/2 ejkh
where we recognize the discrete-time Fourier [(+1)/2-k+1] [(-1)/2+k+1]
-
transform of Rb(n)4, given by: (39)
(-1)n(+1) In order to maintain the coherence with the usual
Rb(n) = (27) definition of discrete-time Fourier transform, we
(/2-n+1) (/2+n+1)
This function is the discrete autocorrelation of change the summation variable, obtaining

(Ortigueira, 2000) c2e-jt =
+ (-1)k(+1)
e-jt e-jh/2 e-jkh
[(+1)/2+k+1] [(-1)/2-k+1]
-
3 (40)
We name acausal the operators that are neither
causal nor anti-causal. Now, the coefficients of the above Fourier series are
4 the cross-correlation, Rbc(k), between
In purely mathematical terms it is a Fourier series
with Rb(n) as coefficients.
 (-a)n This is formally similar to the Riesz-Feller potentials
hn= n! un (41)
(Samko et al, 1987).
and
(-b)n 4.3 The integer order cases
gn= n! )un (42)
with a = (+1)/2 and b = (-1)/2. It is interesting to use the central type 1 derivative
with = 2M +1 and the type 2 with = 2M.
Let Sbc(ej) be the discrete-time Fourier transform of
the cross-correlation, Rbc(k): For the first, /2 is not integer and we can use
Sbc(ej) = FT[Rbc(k)] (43) formulae (30) to (35). However, they are difficult to
manipulate. We found better to use (36), but we must
As Rbc(k) is a correlation, we conclude easily that
avoid the product (-).cos(/2), because the first
Sbc(ej) is given by:
factor is and the second is zero. To solve the
Sbc(ej) = limjh(1-z-1)(+1)/2 (1-z)(-1)/2 (44)
ze problem, we use
= (1-e-jh)(+1)/2 (1-ejh)(-1)/2 1 (+1) sin(/2)
= - (54)
-jh (+1)/2 2(-)cos(/2)
= (1-e ) (1-ejh)(+1)/2 (1-ejh)-1 (45)
(2M+1)! (-1)M
We write, then: (72) to get a factor equal to - . We


c2e-jt = obtain finally
ejt (1-e-jh)(+1)/2 (1-ejh)(+1)/2 (1-ejh)-1 ejh/2 (2M+1)! (-1)M -2M-2
FT-1[||2M+1] = - |t| (55)
= ejt |2 sin(h/2)|+1 [2j sin(h/2)]-1
So, there is a linear system with frequency response and the corresponding impulse response:
given by: (2M+1)! (-1)M -2M-2
hD1(t) = - |t| (56)
H2() = |2 sin(h/2)|+1 [2j sin(h/2)]-1 (46)
that acts on a signal giving its fractional central Concerning the second case, = 2M, we use formula
difference. We can write also (50). As above, we have the product (-).sin(/2)
|2 sin(h/2)|+1 [2j sin(h/2)]-1 = that is again a .0 situation. Using
+ (-1)k(+1) 1 (+1) cos(/2)
= e-jkh (47) - = (57)
[(+1)/2+k+1] [(-1)/2-k+1] 2(-)sin(/2)
-
(2M)! (-1)M
we get a factor .We obtain then:

Dividing (46) by h (hR+) and computing the limit
as h0, it gives: sgn(t) (2M)!(-1)M -2M-1
FT-1[||2Msgn()] = |t| (58)

HD2() = -j ||sgn() (48)
As and
d||+1 sgn(t) (2M)!(-1)M -2M-1
= j(+1) ||sgn() hD2(t) = |t| (59)
j
d
we obtain from (36) using a well known property of As we can see, the formulae (56) and (59) allow us to
the Fourier transform: generalise the Riesz potentials for integer orders.
- sgn(t) --1 However, they do not have inverse.
hD2(t) = |t| (49)
(+1)2(--1)cos[(+1)/2]
or, using the properties of the gamma function
sgn(t) --1
5. CONCLUSIONS
hD2(t) = - |t| (50)
2(-)sin(/2)
We introduced a general framework for defining the
and as previously: fractional central differences and consider two cases
+
1 that are generalisations of the usual central
f() |t-| -1 sgn(t-)d
-
Dc1f(t) = -
2(-)sin(/2) differences. These new differences led to central
- derivatives similar to the usual Grnwald-Letnikov
(51) ones. For those differences, we proposed integral
It is interesting to remark that combining (34) with representations from where we obtained the
(48) as derivative integrals, similar to Cauchy) by using the
HD() = HD1()+jHD2() (52) properties of the Gamma function. The computation
We obtain a function that is null for < 0. This of those integrals led to generalisations of the Riesz
means that the operator defined by (50) is the Hilbert potentials. The most interesting feature lies in the
transform of that defined in (37) and the summation formulae for the Riesz potentials.
corresponding analytic derivative is given by the
convolution of the function at hand with the operator: To test the coherence of the proposed definitions we
--1 --1 applied them to the complex exponential. The results
|t| |t| sgn(t) show that they are suitable for functions with Fourier
HD(t) = -j (53) transform.
2(-)cos(/2) 2(-)sin(/2)
 REFERENCES

Andrews, G.E., Askey, R., and Roy R. (1999).

Special functions. Cambridge Univiversity
Press.
Henrici, P. (1974). Applied and Computational
Complex Analysis. vol. 1, John Wiley &
Sons, Inc., pp. 270-271.
Okikiolu, G. O. (1966). Fourier Transforms of
the operator H. Proceedings of Cambridge
Philosophy Society, 62, 73-78.
Oldham, K.B. and Spanier, J. (1974). The
Fractional Calculus. Academic Press.
Ortigueira, M. D. (2000). Introduction to
Fractional Signal Processing. Part 2:
Discrete-Time Systems. IEE Proc. On
Vision, Image and Signal Processing,
No.1, February, pp. 71-78.
Ortigueira, M.D. (2005). Fractional Differences
Integral Representation and its use to
define Fractional Derivatives. Proceedings
of the ENOC-2005, Fifth EUROMECH
Nonlinear Dynamics Conference.
Eindhoven University of Technology, The
Netherlands, August, 7-12.
Ortigueira, M. D. From Differences to
Differintegrations. To be published in
Fractional Calculus & Applied Analysis.
Kilbas, A.A., Srivastava, H.M., and J.J.Trujillo
(2006). Theory and Applications of
Fractional Differential Equations.
Elsevier.
Samko, S.G., Kilbas, A.A., and Marichev, O.I.
(1987). Fractional Integrals and
Derivatives - Theory and Applications.
Gordon and Breach Science Publishers.