Fractional Taylor Series for Caputo Fractional

Derivatives. Construction of Numerical

Schemes

David Usero
Dpto. de Matemática Aplicada. Facultad de CC. Quı́micas. Universidad
Complutense de Madrid. Spain

Abstract

A fractional power series expansion is obtained for Caputo fractional derivative as

a generalization of Taylor power series. The series obtained are independent from
the point in which fractional derivative is defined. This is used to obtain Euler and
Taylor numerical schemes to solve ordinary fractional differential equations. Finally
the methods derived are applied to integrate numerically a first and a second order
ordinary fractional differential equation using schemes of order α, 2α and 4α for
both equations.

Key words: Fractional derivative, Taylor series, numerical Euler method,

numerical Taylor methods of general order.
1991 MSC: 26A33, 30B10, 65R99, 33E12

1 Introduction

Fractional Calculus is a tool of Mathematical Analysis applied to the study of

integrals and derivatives of arbitrary order, not only fractional but also real.
Commonly this fractional integrals and derivatives are not known for many
scientist and up to recent years has been used only in a pure mathematical
context. But during this last decades this integrals and derivatives have been
applied in many context of sciences.

Nowadays it is impossible to describe a viscoelastic process without using

a fractional derivative. Fractional derivatives have also been used in Anom-
alous Diffusion description, where this derivatives can explain sub-diffusive
Email address: [email protected] (David Usero).

Preprint submitted to Elsevier 23 January 2008

and super-diffusive phenomena observed in real systems. Other applications
are Electromagnetic Theory, Circuit Theory, Biology, Atmospheric Physics,
etc. A wide description of such application together with an exhaustive defi-
nition of all these integrals and derivatives can be found in the books [8] [7],
[4].

The study of this fractional derivatives has two great difficulties, they do not
have a clear significance since they can not be associated to a tangent direction
as the usual first derivative is. This circumstance make impossible any intuitive
a priori analysis of the problem. The second, but not less, difficulty is their
complex integro-differential definition, which make a simple manipulation with
standard integer operators, a complex operation that should be made carefully.
A clear example of such situation is that summation rule is no fulfilled and so
the α derivative of the β derivative of a function is, in general, not equal to
the α + β derivative of such function.

It is not necessary at all to comment how important is the well known Taylor
power series of a given function in the history of mathematics, physics and
many other sciences. Power series have become a fundamental tool in the study
of elementary functions and also other not so elementary as can be checked in
any book of analysis.

In Physics, Chemistry and many other Sciences this power expansions has
allowed scientist to make an approximate study of many systems, neglecting
higher order terms around the equilibrium point. This is a fundamental tool
to linearize a problem which guaranties easy analysis.

In recent decades the power series expansion has been widely used in com-
putational science obtaining an easy approximate of a function [1], numerical
schemes to integrate a Cauchy problem [2], or gaining knowledge about the
singularities of a function by comparing two different Taylor series expansions
around different points [11] [12] [3].

In the context of the fractional derivatives, Taylor series has been developed
for different definitions [5], but in general they consist on series in powers of
xn+α , which in fact is not a purely fractional serie. Fractional Taylor serie has
been developed for Riemann-Liouville derivative [9], and in the present work
a similar study has been made for Caputo fractional derivative defined as
Z x
α 1 Df (t)
c Da f (x) = dt = I1−α
a Df (x) (1)
Γ(1 − α) a (x − t)α
where the fractional integral Iαa is defined as

1 Zx
Iαa f (x) = (x − t)α−1 f (t)dt. (2)
Γ(α) a

2
The use of this definition of the fractional derivative is justified since it has
”good physical properties” [7] as, for example that the derivative of a constant
is zero or that Cauchy problems requires initial conditions formulated in terms
of integer order derivatives interpreted as initial position, initial velocity, etc.

As it has previously mentioned, elementary manipulations with entire order

derivatives as the summation of index, Leibniz rule or the chain rule are not
valid for Caputo fractional derivative as also happen with Riemann-Liouville
definition. Taking this into account, and with no possibility of confusion it
will be used the following convection for the sequential derivative in order to
simplify notation:
mα
c Da = c Dαa c Dαa c Dαa ...c Dαa , (3)
| {z }
m times
which is not equivalent to the derivative of order mα. This is not the case for
the integral operator Iαa for which Iαa Iβa = Iα+β
a .

The present work is organized as follows. In section 2 a fractional McLaurin

serie expansion is obtained for a general function in terms of the Caputo
fractional derivative. The name of fractional MacLaurin serie is used since
this is centered in the point a used also to define the fractional integral and
derivative. In section 3 this power serie is generalized and centered in any
other point a1 > a within the radius of convergence of the power serie. In
section 4 the general fractional Euler and Taylor method are developed, and
those methods are applied to different Fractional Differential Equation (FDE)
of First order in section 5.1 and second order in section 5.2. Finally conclusions
and further lines are exposed.

2 Fractional MacLaurin power serie expansion

In the present section a fractional power serie for a function f (x) in terms
of its fractional derivatives c Dαa f (a) is obtained and this is called MacLaurin
power serie since the point in which the serie is constructed x = a coincides
with the point in which the Caputo derivative c Dαa is defined. This dependence
on the point in which the derivative is calculated is not useful at all since it
make impossible, for example, to obtain directly a numerical Euler scheme for
a Cauchy problem. It would be desirable using a serie in a point a1 different
to a in order to express the solution at point xn+1 as a function of the solution
in xn as this is done in numerical schemes for Cauchy problems. But buildings
can not be started by the roof, i. e. a MacLaurin serie is necessary to develop
more sophisticated expressions.

In order to obtain some interesting results first it is necessary to give some

definitions:

3
Definition 2.1 Let α ∈ R+ , Ω ⊂ R an interval such that a ∈ Ω, a ≤ x
∀x ∈ Ω. Then the following set of functions are defined:

a Iα = {f ∈ C(Ω) : Iαa f (x) exist and is finite in Ω}

a Dα = {f ∈ C(Ω) : c Dαa f (x) exist and is finite in Ω}

Also, some interesting properties of the Caputo’s definition of the fractional

derivative would be necessary in order to develop a power expansion similar
to the Taylor serie using fractional powers of the independent variable x.
Similar properties can be described for the Riemann-Liouville derivative [9].
First property correspond to the application of the integral operator Iαa to a
fractional derivative of a function.

Proposition 2.1 Let α ∈ (0, 1] and f (x) ∈ a Dα then

Iαa c Dαa f (x) = f (x) − f (a). (4)

This can be demonstrated applying equation (1) so Iαa c Dαa f (x) = I1a Df (x)
since the integral operators holds the summation of index. Then property 2.1
is obtained directly.

The next proposition is a zero order approximation of a function in terms of

its first α derivative.

Proposition 2.2 A function f (x) ∈ a Iα with the same conditions of propo-

sition 2.1 can be written as
(x − a)α
f (x) = f (a) + c Dαa f (ξ) (5)
Γ(α + 1)

for any ξ ∈ (a, x), being c Dαa f (x) continuous in [a, x].

In order to obtain expression (5) it is necessary to start from the same point
of proposition 2.1

1 Zx
Iαa c Dαa f (x) = (x − t)α−1 c Dαa f (t)dt (6)
Γ(α) a

and applying the Mean Value Theorem for the integral

α Z x α
c Da f (ξ) c Da f (ξ)
Iαa c Dαa f (x) = (x − t)α−1 dt = (x − a)α (7)
Γ(α) a Γ(α + 1)

for any ξ ∈ (a, x). Substituting expression (4) the desired proposition 2.2 is
obtained.

4
More sophisticated properties with cumulative integrals and differential op-
erators can be obtained and would be the starting point for a higher order
approximations.

Proposition 2.3 Let α ∈ (0, 1], m, n ∈ N and f (x) an analytic function in

Ω ⊂ R, f (x) ∈ a D(m+1)α with a, x ∈ Ω, a < x. Then

(x − a)mα
Imα mα (m+1)α
a c Da f (x) − Ia
(m+1)α
c Da f (x) = c Dmα
a f (a) (8)
Γ(mα + 1)

Since c Dkα
a f (a) is a constant value it can be written

(x − a)mα
Imα nα nα mα nα
a c Da f (a) = c Da f (a)Ia 1 = c Da f (a) . (9)
Γ(mα + 1)

On the other hand, applying proposition 2.1

I(m+1)α
a
(m+1)α
c Da f (x) = Imα α α mα mα mα mα mα
a Ia c Da c Da f (x) = Ia c Da f (x) − Ia c Da f (a),
(10)
and then equation (8) is obtained by direct substitution.

As it has been commented before, this proposition would be the initial point to
construct the power serie of a sufficiently well behaved function f (x). In order
to obtain a proper expression for this serie another proposition is needed.

Proposition 2.4 In the same conditions of proposition 2.3 and being m, k ∈

N, c Dmα mα
a f (x) continuous in [a, x], and c Da f (x) ∈ a Ikα ,

(x − a)kα
Ikα mα mα
a c Da f (x) = c Da f (ξ) (11)
Γ(kα + 1)

for any ξ ∈ (a, x).

Equation (11) is a generalization of corresponding equation from proposi-

tion 2.2. Using the definition of the integral operators

1 Zx
Ikα mα
a c Da f (x) = (x − t)kα−1 c Dmα
a f (t)dt (12)
Γ(kα) a

and applying the Mean Value Theorem for c Dmα

a f (x) equation (11) is obtained
directly.

Once demonstrated the last result, it is possible to obtain the desired ex-
pression for the fractional power serie for a function in terms of its Caputo
fractional derivatives of order α ∈ (0, 1].

5
Theorem 2.1 Let α ∈ (0, 1], n ∈ N and f (x) a continuous function in [a, b]
satisfying the following conditions:

(1) ∀j = 1, ..., n, c Djα jα

a f ∈ C([a, b]) and c Da f ∈ a Iα ([a, b]).
(n+1)α
(2) c Da f (x) is continuous on [a, b].

Then ∀x ∈ [a, b],

n
X
jα (x − a)jα
f (x) = c Da f (a) + Rn (x, a), (13)
j=0 Γ(jα + 1)

being
(x − a)(n+1)α
Rn (x, a) = c Da(n+1)α f (ξ) , a≤ξ≤x (14)
Γ((n + 1)α + 1)

Proof: Using proposition 2.3 for different values of m it is possible to write

• Case m = 0 → f (x) − Iαa c Dαa f (x) = f (a). Then f (x) = f (a) + Iαa c Dαa f (x).
(x − a)
• Case m = 1 → Iαa c Dαa f (x) − I2α 2α α
a c Da f (x) = c Da f (a) and then
Γ(α + 1)
(x − a)α
Iαa c Dαa f (x) = c Dαa f (a) + I2α 2α
a c Da f (x)
Γ(α + 1)
(x − a)2α
• Case m = 2 → I2α D
a c a
2α
f (x) − I 3α
D
a c a
3α
f (x) = D
c a
2α
f (a) . Then
Γ(2α + 1)
(x − a)2α
I2α 2α 2α
a c Da f (x) = c Da f (a) + I3α 3α
a c Da f (x)
Γ(2α + 1)
(x − a)mα (m+1)α (m+1)α
• General case → Imα a c Dmα
a f (x) = c D mα
a f (a) +I c Da f (x).
Γ(mα + 1) a

Substituting repetitively the resulting terms up to order n the following serie

is obtained
n
X
jα (x − a)jα
f (x) = c Da f (a) + I(n+1)α
a
(n+1)α
c Da f (x), (15)
j=0 Γ(jα + 1)

where the last term correspond to the integral form of the reminder term, so
it is possible to write

Rn (x, a) = I(n+1)α
a
(n+1)α
c Da f (x). (16)

Finally equation (14) is obtained applying proposition 2.4 with m = k = n + 1

to equation (16).

6
3 Fractional Taylor power serie expansion

In the previous section theorem 2.1 allows to obtain a fractional power serie for
a function in terms of its Caputo fractional derivatives evaluated at a, which
is, in some sense, the initial point of the independent variable x. The ideal
situation would be to obtain a similar serie with the derivatives evaluated in
any other point a1 > a, so the expansion can be constructed independently
from the starting point a.

The next theorem makes real this possibility at least up to fourth order and
gives an idea about how to obtain higher order approximations.

Theorem 3.1 Let α ∈ (0, 1] and f (x) a continuous function in [a, b] satisfy-
ing the following conditions:

(1) ∀j = 1, ..., 8, c Djα jα

a f ∈ C([a, b]) and c Da f ∈ a Iα ([a, b]).
(2) c D9α
a f (x) is continuous on [a, b].

Let a1 ∈ (a, b]. Then ∀x ∈ [a, b],

∆1 ∆2
f (x) = f (a1 ) + c Dαa f (a1 ) + c D2α
a f (a1 ) +
Γ(α + 1) Γ(2α + 1)
(17)
∆3 ∆4
+c D3α
a f (a1 ) + c D4α
a f (a1 ) + R4 (x, a1 , a),
Γ(3α + 1) Γ(4α + 1)

being ∆1 , ∆2 , ∆3 and ∆4 the differences given by

∆1 = [H α − Lα ]
h i
Γ(2α+1) α
∆2 = H 2α − L2α − Γ2 (α+1)
L ∆1
h i
Γ(3α+1) Γ(3α+1)
∆3 = H 3α − L3α − Γ(α+1)Γ(2α+1)
Lα ∆2 − Γ(2α+1)Γ(α+1)
L2α ∆1
h i
Γ(4α+1) Γ(4α+1) 2α Γ(4α+1)
∆4 = H 4α − L4α − Γ(α+1)Γ(3α+1)
Lα ∆3 − Γ2 (2α+1)
L ∆2 − Γ(3α+1)Γ(α+1)
L3α ∆1
(18)
and H = (x − a), L = (a1 − a). R4 (x, a1 , a) is the reminder term.

Proof: Since f (x) and c Djα

a f (x) for j = 1, 2, 3, 4 fulfill the conditions of theo-
rem 2.1, there must be fractional power series expansions for f (x), c Dαa f (x),

7
 2α 3α 4α
c Da f (x), c Da f (x), c Da f (x), and using them at x = a1 we have:

L α L 2α
f (a) = f (a1 ) − c Dαa f (a) Γ(α+1) − c D2α
a f (a) Γ(2α+1) −
L 3α L 4α
−c D3α 4α
a f (a) Γ(3α+1) − c Da f (a) Γ(4α+1) − R4,0

L α L 2α
α
c Da f (a) = c Dαa f (a1 ) − c D2α 3α
a f (a) Γ(α+1) − c Da f (a) Γ(2α+1) −
L 3α L 4α
−c D4α 5α
a f (a) Γ(3α+1) − c Da f (a) Γ(4α+1) − R4,1

L α L 2α
2α
c Da f (a) = c D2α 3α 4α
a f (a1 ) − c Da f (a) Γ(α+1) − c Da f (a) Γ(2α+1) −
3α 4α
(19)
L L
−c D5α 6α
a f (a) Γ(3α+1) − c Da f (a) Γ(4α+1) − R4,2

L α L 2α
3α
c Da f (a) = c D3α 4α 5α
a f (a1 ) − c Da f (a) Γ(α+1) − c Da f (a) Γ(2α+1) −
L 3α L 4α
−c D6α 7α
a f (a) Γ(3α+1) − c Da f (a) Γ(4α+1) − R4,3

L α L 2α
4α
c Da f (a) = c D4α 5α 6α
a f (a1 ) − c Da f (a) Γ(α+1) − c Da f (a) Γ(2α+1) −
L 3α L 4α
−c D7α 8α
a f (a) Γ(3α+1) − c Da f (a) Γ(4α+1) − R4,4 ,

being R4,i the corresponding reminder terms of the fourth order serie of every
iα
c Da f (x) at x = a1 .

Then substituting every serie from equation (19) in the fourth order serie of
f (x) and grouping corresponding terms, equation (17) with differences (18) is
obtained.

The reminder term R4 (x, a1 , a) results as a combination of the reminder terms

of expressions (19) and higher order terms with derivatives of order 5α, 6α,
7α and 8α evaluated at x = a. Explicit form is omitted because of its complex
form, but can be directly computed following indications given above.

The method outlined above can be extended up to any order obtaining series
similar to equation (17) for which the difference ∆k would be given by
k−1
X Ljα ∆k−j
∆k = H kα − Lkα − Γ(kα + 1) , (20)
j=1 Γ(jα + 1)Γ ((k − j)α + 1)

and the corresponding serie of order n would be

n
X
jα ∆j
f (x) = c Da f (a1 ) + Rn (x, a1 , a) (21)
j=0 Γ(jα + 1)

8
with Rn (x, a1 , a) being the reminder term of order nα.

4 Application to the construction of numerical schemes

One application of Taylor power serie is the construction of numerical schemes

to solve ODE Cauchy problems as is done in an elementary course of numerical
analysis [2]. In the present section a similar approach is going to be followed
in order to obtain numerical schemes to integrate the Cauchy problem for
Fractional Differential Equation (FDE).

Let start with a Cauchy problem for a general FDE given by:


 c Dα y(t) = f (y)
0

(22)
 y(0) = y0

with α ∈ (0, 1] and setting the origin at t = a = 0 for simplicity. It has been
set f (y) also for simplicity since in the problems studied below there is no
explicit dependence on t in the FDE but more general cases can be considered
setting f (t, y) instead.

To obtain different integration schemes for the problem (22) it is necessary to

transform the continuous variable t in a discrete analogue tj = jh with a fixed
step h.

Using the fractional Taylor expansion given at theorem 3.1 up to the corre-
sponding order it will make possible to obtain the schemes.

4.1 Fractional Euler Method

The Euler method uses the first order Taylor serie to obtain the integration
scheme. It is the easiest numerical integration method but the errors are usu-
ally very big so is not frequently used.

If we suppose that y(t) is a solution of problem (22), then according to the

fractional Taylor serie for y(t) at ti it can be written

∆1
y(ti+1 ) = y(ti ) + c Dα0 y(ti ) + R1 (t, ti , 0), (23)
Γ(α + 1)

9
with ∆1 = hα ((i + 1)α − iα ). Using the FDE of problem (22)

f (y(ti ))hα
y(ti+1 ) = y(ti ) + ((i + 1)α − iα ) + R1 (t, ti , 0) (24)
Γ(α + 1)

Fractional Euler method consist in approximate the function y(t) at ti , i =

0, ..., N for a set of values wi ≈ y(ti ) neglecting the reminder term. Then

f (wi )
wi+1 = wi + hα ((i + 1)α − iα )
Γ(α + 1) (25)
w0 = y0 ,

for every i = 0, ..., N .

Since the reminder term R1 (ti+1 , ti , 0) is of order h2α , the local truncation
error is of order hα and since α ∈ (0, 1] it would be hα > h. This method will
present, in general, higher errors than the classical Euler scheme.

Given equation (23) and using equation (14) it is possible to obtain

" #
2α Aδ1 iα Chα δ1 i2α (D − B)i2α
R1 (ti+1 , ti , 0) = h − + , (26)
Γ2 (α + 1) Γ(α + 1)Γ(2α + 1) Γ(2α + 1)

where ti = hi, δ1 = (i + 1)α − iα and A = c D2αα 2αα

0 y(0), B = c D0 y(ξ1 ) is
obtained applying the mean value theorem to the reminder term of y(ti ),
C = c D3αα
0 y(ξ2 ) is obtained applying the mean value theorem to the reminder
term of c Dαα 2αα
0 y(t) and finally D = c D0 y(ξ3 ) is obtained applying the mean
value theorem to the reminder term of y(ti+1 ). All of this quantities A, B, C
and D are unknown constants that can be bounded.

Using equation (26), local truncation error at step i + 1 can be defined as:

y(ti+1 ) − y est (ti+1 )

τi+1 (h) = , (27)
hα
where y est (ti+1 ) is the estimated value of y(t) at t = ti+1 using y(ti ). It is clear
from (24) that
R1
τi+1 (h) = α , (28)
h
and then τi+1 (h) neglects as hα .

4.2 Fractional Taylor Methods

Similar to Euler method, there exist higher order methods based on Taylor
series. They are commonly known as Taylor methods and they make possible

10
to obtain an approximation or any desired order. All of them has the same
difficulty: they require the evaluation of successive derivatives of the function
of orders lower than the required approximation. This is no easy in general
since many real problems make this an impossible task. In many other cases
as described in [11] it can be computed with symbolic mathematical software
obtaining series of hundred terms.

Depending on the order of the serie used for approximation it is possible to

use second, third, fourth order methods, etc. The most commonly used are
the second and fourth order methods and their fractional analogous would be
described below.

Using the same description given above for Euler Method, and approximating
the solution y(t) at every ti for a set of values wi ≈ y(ti ) the 2α Taylor method
for the FDE (22) is

f (wi )
wi+1 = wi + hα ((i + 1)α − iα ) +
Γ(α + 1)" #
α
D f (w ) Γ(2α + 1)
i ((i + 1) − i ) (29)
c 0 i 2α 2α α α α
+h2α (i + 1) − i −
Γ(2α + 1) Γ(α + 1)2
w0 = y0 ,

for every i = 0, ..., N .

With the expression given at Theorem 3.1 it is possible to obtain the 4α order
method given by:
α 2α 3α
f (wi ) c D f (wi ) c D f (wi ) c D f (wi )
wi+1 = wi + hα δ1 + h2α 0 δ2 + h3α 0 δ3 + h4α 0 δ4
Γ(α + 1) Γ(2α + 1) Γ(3α + 1) Γ(4α + 1)
w0 = y0 ,
(30)
for every i = 0, ..., N , and being δi the differences given by

δ1 = [(i + 1)α − iα ]
h i
Γ(2α+1) α
δ2 = (i + 1)2α − i2α − Γ(α+1)2
i δ1
h i
Γ(3α+1) Γ(3α+1)
δ3 = (i + 1)3α − i3α − Γ(2α+1)Γ(α+1)
iα δ2 − Γ(2α+1)Γ(α+1)
i2α δ1
h i
Γ(4α+1) Γ(4α+1) 2α Γ(4α+1)
δ4 = (i + 1)4α − i4α − Γ(3α+1)Γ(α+1)
iα δ3 − Γ(2α+1)2
i δ2 − Γ(3α+1)Γ(α+1)
i3α δ1
(31)

As it also happens with classical Taylor methods, the computation of the frac-
tional derivatives of f (y) will be, in general, a difficult task, increased this
time with the difficulties of the fractional derivative. This will make Taylor
method hard to be applied in many problems, since it will require huge calcula-

11
tions in order to estimate the remaining terms, being fractional Euler method
sufficiently accurate.

It is possible to make a similar analysis of the truncation error as the one

made for the Fractional Euler method. In this case it would be necessary to
obtain the expression for many reminder terms and calculations will become
very large to be outlined here. General details are given below.

In a general Taylor serie of order N the reminder term RN is of order h(N +1)α .
Then we will have to compute all the reminder terms of y(t) and all their N
fractional derivatives which also will be of the same order. Once substituted,
all the reminder terms obtained would be of such order or ever greater, so the
general expression of RN for y(ti+1 ) as a function of y(ti ) and its derivatives
will be of order h(N +1)α . Then the local truncation error defined in (27) will
be of order hN α . This result generalizes for the fractional Taylor method the
classical obtained for the truncation error of the Taylor method.

5 Numerical Study of Fractional Differential Equations

In the present section the three methods described above are going to be tested
in two different cases for which analytical solution is known.

In the first part the schemes are tested with a first order FDE and in the
last one with a second order FDE analogous to a fractional oscillator. In that
former case the three schemes are compared also with a symplectic scheme
obtained in [12] [10].

5.1 First Order FDE

The first problem studied is given by the first order Cauchy problem


 c Dα y(t) = λy
0

(32)
 y(0) = y
0

with α ∈ (0, 1], which solution is given by y(t) = y0 Eα (λtα ), being Eα (z)
the Mittag-Leffler function [4,8], that generalizes the elementary exponential
function ez .

This case is very simple and sequential fractional derivatives of the function
gives c Dkα k
0 y = λ y so Taylor Method can be made up to any order. Numerical
schemes for this simple case are:

12
Fractional Euler " #
E hα λ
wi+1 = wiE 1+ δ1 (33)
Γ(α + 1)
Fractional 2α Taylor
" #
T2 hα λ h2α λ2
wi+1 = wiT 2 1+ δ1 + δ2 (34)
Γ(α + 1) Γ(2α + 1)

Fractional 4α Taylor
" #
T4 hα λ h2α λ2 h3α λ3 h4α λ4
wi+1 = wiT 4 1+ δ1 + δ2 + δ3 + δ4
Γ(α + 1) Γ(2α + 1) Γ(3α + 1) Γ(4α + 1)
(35)

where the differences δj are defined in equation (31) and w0E = w0T 2 = w0T 4 =
y0 .

In figure 1 are shown the solutions for two different values of α and their
corresponding absolute values. Initial condition is x(0) = 1 in all cases and
parameter λ = 1. Calculations has been made with single precision and the
exact solution has been computed using a MATLAB routine made by I. Pod-
lubny [6]. As the order of approximation depends on a power of α it can be
observed that increasing the value of α decreases the errors as can be expected
since increasing α increases the order of the scheme and then it should be more
accurate.

5.2 Second Order FDE

The next problem studied correspond to a second order FDE given by



 2α 2


 c D0 y(t) = −ω y

 y(0) = y0 (36)




 c Dα y(0) = p0
0

and in that case two initial conditions for the function y(t) and its first frac-
tional derivative is needed.

This problem can be transformed into a system of two first order FDE


 c Dα y(t) = p(t)
0
(37)

 Dα p(t) = −ω 2 y(t)
c 0

13
with initial conditions y(0) = y0 and p(0) = p0 . This system is known as
Fractional Oscillator and has been studied in [12], [10,15] and in [12] [13,14]
adding an external periodic forcing. In all cases cited numerical integration
has been of maximum interest and a symplectic scheme has been used. This
scheme, developed partially by the author is explained at [12] and [10] and
correspond to

Symplectic Scheme

 i

 ω 2 hα X
 pS

 i+1 = p 0 − wk [(i + 1 − k)α − (i − k)α ]
Γ(α + 1) k=0
i (38)


 S hα X
α α
 w
 i+1
 = w 0 + pk+1 [(i + 1 − k) − (i − k) ]
Γ(α + 1) k=0

The main problem to compute algorithm (38) is the high number of operations
needed. In every step it calculates the a sum over all previous elements so
operations increase as N 2 and also does the computation time. This will not
happen with Euler and Taylor schemes because the number of operations
increase linearly with N .

Exact solution of (37) is known and in terms of the complex variable Θ(t) =
Solutions for α=0.5 Solutions for α=0.9
400 250

200
300

150
200
100

100
50

0 0
0 1 2 3 4 5 0 1 2 3 4 5

Absolute errors for α=0.5 Absolute errors for α=0.9

300 100

250 80
200
60
150
40
100

50 20

0 0
0 1 2 3 4 5 0 1 2 3 4 5

Fig. 1. Plots for the solutions of the FDE (32) and absolute errors for h = 0.001 and
N = 5000. Continuous line correspond to exact solution, dotted line to computed
Euler solution, dashed line to computed 2α order Taylor solution and dashed-dotted
line to 4α order solution.

14
√ √
ωy(t) + ip(t)/ ω is Θ(t) = Θ(0)Eα (iωtα )

The particular form of problem (37) makes possible to compute the derivatives
of any order so the three methods described above can be obtained:

Fractional Euler Method



 E E pEi h
α
 wi+1 = wi +
 δ1
Γ(α + 1) (39)



wiE hα
 pE
i+1 = pE
i − δ1
Γ(α + 1)
Fractional 2α Taylor Method


 T2 T2 pTi 2 hα ω 2 wiT 2 h2α
 wi+1 = wi +
 δ1 − δ2
Γ(α + 1) Γ(2α + 1) (40)



ω 2 wiT 2 hα ω 2 pTi 2 h2α
 pTi+1
2
= pTi 2 − δ1 − δ2
Γ(α + 1) Γ(2α + 1)
Fractional 4α Taylor Method


 T4 T4 pTi 4 hα ω 2 wiT 4 h2α ω 2 pTi 4 h3α ω 4 wiT 4 h4α
 wi+1 = wi +
 δ1 − δ2 − δ3 + δ4
Γ(α + 1) Γ(2α + 1) Γ(3α + 1) Γ(4α + 1)



ω 2 wiT 4 hα ω 2 pTi 4 h2α ω 4 wiT 4 h3α ω 4 pTi 4 h4α
 pTi+1
4 T4
= pi − δ1 − δ2 + δ3 + δ4
Γ(α + 1) Γ(2α + 1) Γ(3α + 1) Γ(4α + 1)
(41)

where the differences δj are defined in equation (31) and w0E = w0T 2 = w0T 4 = y0
and pE T2 T4
0 = p0 = p0 = p0 .

In figure 2 are shown the solutions and their corresponding absolute errors
computed for the fractional harmonic oscillator (37) with ω = 1, initial con-
ditions x(0) = 1, p(0) = 0 and two different values of α. Exact solution has
been computed using the same MATLAB routine made by Prof. I. Podlubny
cited above with complex argument. An very small difference can be appreci-
ated between this exact solution and the one calculated with the symplectic
scheme, while a other Euler and Taylor methods has bigger errors as expected.
As also happened before, errors increase when decreases α.

6 Conclusions

Formally it is possible to obtain Fractional Taylor serie of general order N α

with α ∈ (0, 1] is obtained in a general form and its reminder term as it is
computed in the present work. This is done for a serie centered in the definition
point a called MacLaurin fractional serie and then generalized for any other

15
point a1 > a. This generalization is needed in order to define a numerical
Taylor scheme to integrate Fractional Ordinary Differential Equations (FDE).

This serie is used to construct numerical schemes of order 4α and also a general
method is outlined to obtain a Taylor Scheme of a higher general order N α.
Truncation errors of a Fractional Taylor Method of order N α are of order N α,
as it would be expected since this fractional methods are a generalization of
classical schemes.

This schemes have been applied to two different Fractional Differential equa-
tions, which general solutions are known, as a test for the method. In general,
solutions obtained by numeric computations require small computation time
compared with others, like the symplectic scheme (38). Most of the schemes
developed up to nowadays are computed inverting the differential equation
and evaluating the integral by any numerical method. In general this requiere
high computational times since the number of operations for the solution in
tn grows as n2 . This is not the case in Taylor methods since only the solu-
tion in tn is required in order to calculate the solution in tn+1 , and then the
computational time grows linearly with the number of steps.

Solutions for α=0.5 Solutions for α=0.9

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1
0 1 2 3 4 0 1 2 3 4

Absolute errors for α=0.5 Absolute errors for α=0.9

0.2 0.1

0.08
0.15

0.06
0.1
0.04

0.05
0.02

0 0
0 1 2 3 4 0 1 2 3 4

Fig. 2. Plots for the solutions and absolute errors for the FDE (37) with h = 0.001
and N = 5000. Black line correspond to exact solution, dotted line to computed
Euler solution, dashed line to computed 2α order Taylor solution, dashed-dotted
line to 4α order solution and long dashed line, that cannot be distinguished in this
scale, to the solution computed using the symplectic scheme.

16
Errors of the computed solution grows as could be expected and 4α order
method is the most accurate of the three schemes tested as shown in figures
1 and 2. This is the same for any initial condition. Higher order methods will
show best accuracy.

In general, increasing the number of terms will make possible, for a fixed time
step h, a most accurate solution that will be valid for a greater number of
iterations. But this will be true only within the radius of convergence of the
serie. This radius of convergence will be the distance of the initial solution to
the nearest singularity of the function.

This has been used with nonlinear ordinary differential equation of integer
order to localice this singularity by extending the problem to the complex
plane. Comparing two of this Taylor series of different orders it is possible to
estimate its radius of convergence and even localize them with high accuracy.
This has been done in [11] and applied to classical problems like the logis-
tic equation, the Pendulum, the Van der Pol Oscillator or the Henon-Heiles
system. There exist general routines that can be used to integrate classical
problems using Taylor methods as the one written in Fortran described in [3]
which also localize the singularity in a similar way. Nothing similar has been
done with fractional derivative problems and future work will be done in this
direction.

7 Acknowledgments

This work has been supported by the Ministerio de Educación y Ciencia of

Spain under the project “Modelos Fraccionarios de Evolución. Aplicaciones”
(MTM2005-05573) and “Convocatoria de Creación y Financiación de Grupos
de Investigación GR69/06”(910711).

References

[1] T. Apostol, “Calculus”, Blaisdell Publishing, Waltham, Massachusetts, 1990.

[2] R. Burden, J. Faires, “Numerical Analysis”, 6th. edition, Brooks-Cole

Publishing, Pacific Grove, 1997.

[3] Y. F. Chang, G. Corliss, “ATOMFT: solving ODE’s and DAE’s using Taylor
Series”, Computers Math. Applic., 28, 209-233, 1994.

[4] A. A. Kilbas, H. M. Srivastava, J. J. Trujillo, “ Theory and Applications of

Fractional Differential Equations”,Elsevier, 2006.

17
[5] T. J. Osler, “Taylor’s Series Generalized for Fractional Derivatives and
Applications”, SIAM Journal of Mathematical Analysis, 2, 37-48, 1971.

[6] I. Podlubny, “MLF.m” algorithm at file-exchange section of Mathworks

(http://www.mathworks.com).

[7] I. Podlubny, “Fractional Differential Equations”, Academic Press, New York

(1999).

[8] S. Samko, A. Kilbas, O. Marichev, “Fractional Integrals and Derivatives”,

Gordon and Breach Science Publishers, Amsterdam, 1993.

[9] J.J.Trujillo, M.Rivero, B.Bonilla, “On a Riemann-Liouville generalized Taylor’s

formula”, Journal of Mathematical Analysis and Applications, 231, 255-265,
1999.

[10] G. Turchetti, D. Usero and L. Vázquez, “Fractional Derivative and Hamiltonian

Systems”, Tamsui Oxford Journal of Mathematical Sciences, 18, 45-56, 2002.

[11] D. Usero, “Estimación numérica de los posibles polos de una ODE”, Dpto. de
Matemática Aplicada, Universidad Complutense de Madrid, 2001.

[12] D. Usero, “Propagación de Ondas No Lineales en Medios Heterogéneos”, Ph.D.

Thesis, Dpto. de Matemática Aplicada, Universidad Complutense de Madrid,
2004.

[13] D. Usero, L. Vázquez, “Fractional Derivative: A New Formulation for Damped

Systems”, in: L. Vázquez, R. S. MacKay, M. P. Zorzano (Eds.) “Localization
and Energy Transfer in Nonlinear Systems”, 2003 , pp.296-303.

[14] D. Usero, L. Vázquez, “Ecuaciones no Locales y Modelos Fraccionarios”,

Revista de la Real Academia de Ciencias Exactas, Fı́sicas y Naturales, 99n2,
203-223, 2005.

[15] L. Vázquez, “Singularity Analysis of a Nonlinear Fractional Differential

Equation” Revista de la Real Academia de Ciencias. Serie A: Matemáticas,
99, 211-217, 2005.

18