Introductory Notes on Fractional Calculus

xuru.org
[email protected]

July 31, 2006

Abstract
Our understanding of Nature relies on calculus, which in turn relies
on the intuitive concept of the derivative. It’s descriptive power comes
from the fact that it analyses the behavior at scales small enough that
its properties change linearly, so avoiding complexities that arise at
larger ones. Fractional Calculus generalizes this concept from integer to
non-integer order. Despite it seems not to have significant applications
in fundamental physics, research on this core concept could be valuable
in understanding Nature. These notes comprise an introduction to the
field.

1 Introduction
Fractional Calculus is the branch of calculus that generalizes the derivative
of a function to non-integer order, allowing calculations such as deriving a
function to 1/2 order. Despite “generalized” would be a better option, the
name “fractional” is used for denoting this kind of derivative.

!Note: These notes are not a summary of standard methods or an

academic text, so some conventional developments are omitted while
some other are introduced. The same happens with notation.

The derivative of a function f is defined as

f (x + h) − f (x)
D 1 f (x) = lim (1.1)
h→0 h
Iterating this operation yields an expression for the n-st derivative of a
function. As can be easily seen -and proved by induction- for any natural
number n,
n  
X n
D n f (x) = lim h−n (−1)m f (x + (n − m)h) (1.2)
h→0 m
m=0
2 1 Introduction

where
 
n n!
= (1.3)
m m!(n − m)!

or equivalently,
n  
n −n
X
m n
D f (x) = lim h (−1) f (x − mh) (1.4)
h→0 m
m=0

The case of n = 0 can be included as well.

Such an expression could be valuable for instance in a simple program
for plotting the n-st derivative of a function.
Viewing this expression one asks immediately if it can be generalized to
any non-integer, real or complex number n. There are some reasons that
can make us think so,

1. The fact that for any natural number n the calculation of the n-st
derivative is given by an explicit formula (1.2) or (1.4).

2. That the generalization of the factorial by the gamma function allows

 
n n! Γ(n + 1)
= = (1.5)
m m!(n − m)! Γ(m + 1)Γ(n − m + 1)

which also is valid for non-integer values.

3. The likeness of (1.2) to the binomial formula

n  
n
X n
(a + b) = an−m bm (1.6)
m
m=0

which can be generalized to any complex number α by

∞
X Γ(α + 1)
(a + b)α = aα−n bn (1.7)
n! Γ(α − n + 1)
n=0

which is convergent if

|b| < a (1.8)

There are some desirable properties that could be required to the frac-
tional derivative,

1. Existance and continuity for m times derivable functions, for any n

which modulus is equal or less than m.
2 Exponentials 3

2. For n = 0 the result should be the function itself; for n > 0 integer
values it should be equal to the ordinary derivative and for n < 0
integer values it should be equal to ordinary integration -regardless
the integration constant.

3. Iterating should not give problems,

D α+β f (x) = D α D β f (x) (1.9)

4. Linearity,

D α (af (x) + bg(x)) = a D α f (x) + b D α g(x) (1.10)

5. Allowing Taylor’s expansion in some other way.

6. Its characteristic property should be preserved for the exponential

function,

D α ex = ex (1.11)

2 Exponentials
The case of the exponential function is specially simple and gives some clues
about the generalization of the derivatives. Following (1.2),
α  
α ax −α
X α a(x+(α−n)h)
n
D e = lim h (−1) e =
h→0 n
n=0
α  
n α
X
ax −α
= e lim h (−1) (eah )α−n =
h→0 n
n=0
ax −α ah
=e lim h (e − 1)α =
h→0
α ax
=a e (2.1)

the above limit exists for any complex number α. However, it should be
noted that in the substitution of the binomial formula a natural number
has been considered. We shall deal with this problem later to get our first
generalization of the derivative. Applying this to the imaginary unit,

D α cos(x) + i D α sin(x) = D α eix = iα eix =

απ i απ
=e eix = ei(x+ 2 ) =
2
   
απ απ
= cos x + + i sin x + (2.2)
2 2
4 2 Exponentials

and

D α cos(x) − i D α sin(x) = D α e−ix = (−i)α e−ix =

−απ i απ
=e 2e−ix = e−i(x+ 2 ) =
   
απ απ
= cos x + − i sin x + (2.3)
2 2

solving this system we have the next definition for the sine and cosine deriva-
tives,
 
α απ
D sin(x) = sin x + (2.4)
2

and
 
απ
D α cos(x) = cos x + (2.5)
2

We could expect these relations for the sine and cosine derivatives to be
maintained in the generalization of the derivative.
Applying the above method we also can calculate the following,

D α cos(ax) + i D α sin(ax) = D α eiax = (ai)α eiax =

απ i απ
= aα e eiax = aα ei(ax+ 2 ) =
2
   
α απ α απ
= a cos ax + + ia sin ax +
2 2
(2.6)

and

D α cos(ax) − i D α sin(ax) = D α e−iax = (−ai)α e−iax =

απ i απ
= aα e− e−iax = aα e−i(ax+ 2 ) =
2
   
α απ α απ
= a cos ax + − ia sin ax +
2 2
(2.7)

Thus,
 
α α απ
D sin(ax) = a sin ax + (2.8)
2

and
 
απ
D α cos(ax) = aα cos ax + (2.9)
2
3 Powers 5

Indeed, the above result of the exponential can be applied to any function
that can be expanded in exponentials
∞
X
f (x) = an einx =⇒
n=−∞
∞
X ∞
X
=⇒ D α f (x) = an D α einx = an (ni)α einx =
n=−∞ n=−∞
∞ ∞
X απ i X απ
= an nα e 2 einx
= an nα ei(nx+ 2
)
(2.10)
n=−∞ n=−∞

Expanding the function in Fourier series,

∞ Zπ
X einx
f (x) = f (t)e−int dt =⇒
n=−∞
2π
−π
∞ απ Zπ
X nα ei(nx+ 2 )
α
=⇒ D f (x) = f (t)e−int dt (2.11)
n=−∞
2π
−π

This method can be useful for calculating fractional derivatives of trigono-

metric functions.

3 Powers
The case of powers of x also has some simplicity that allows its generaliza-
tion. The case of integer order derivatives
n−1
Y a!
D 1 xa = axa−1 =⇒ D n xa = xa−n (a − m) = xa−n (3.1)
(a − n)!
m=0

can be easily generalized to non-integer order derivatives

Γ(a + 1)
D α xa = xa−α (3.2)
Γ(a − α + 1)

which can be applied to any function that can be expanded in powers of x

∞
X
f (x) = an xn =⇒
n=−∞
∞ ∞
X X Γ(n + 1)
=⇒ D α f (x) = an D α xn = an xn−α (3.3)
n=−∞ n=−∞
Γ(n − α + 1)
6 4 Binomial Formula

Expanding the function in Taylor series,

∞
X D n f (a) n
f (x + a) = x =⇒
n!
n=0
∞
X D n f (a)
=⇒ D α f (x + a) = xn−α (3.4)
Γ(n − α + 1)
n=0

or expanding it in Laurent series,

∞ Z
1 X n f (t + a)
f (x + a) = x dt =⇒
2πi n=−∞ tn+1
C
∞ Z
1 X Γ(n + 1) f (t + a)
=⇒ D α f (x + a) = xn−α dt
2πi n=−∞ Γ(n − α + 1) tn+1
C
(3.5)

This can be an useful tool for calculating fractional derivatives. How-

ever, we should compare these results of powers with the previous results of
exponentials to see if they agree. With the result of exponentials (2.1),
∞ ∞
X (ax)n X an+α xn
D α eax = aα eax = aα = (3.6)
n! n!
n=0 n=0

but with the result of powers (3.4),

∞ ∞
X (ax)n X an xn−α
D α eax = D α = (3.7)
n! Γ(n − α + 1)
n=0 n=0

If we compare both results, we see that they only agree for integer values
of α. We shall see later where these discrepancies come from, and how they
can be avoided.

4 Binomial Formula
In the expression (1.2) the exponentian function allows the substitution of
the binomial formula as done in (2.1), but this is not possible for any given
function. For applying this substitution we require the following displace-
ment operator,

d h f (x) = f (x + h) (4.1)

whose iteration yields

d hα f (x) = f (x + αh) (4.2)

5 Functions of the Derivative 7

what allows the application of the binomial formula (1.6) for natural num-
bers and the generalized binomial formula (1.7) for complex numbers,
n  
n −n
X n m
D f (x) = lim h (−1) f (x + (n − m)h) =
h→0 m
m=0
n  
m n
X
−n
= lim h (−1) d n−m f (x) =
h→0 m h
m=0
dh − 1 n
 
= lim f (x) (4.3)
h→0 h

so that for any complex number α it can be generalized

dh − 1 α
 
α
D f (x) = lim f (x) (4.4)
h→0 h

This sheds more light on the derivative and its generalization. Using the
expression of the generalized binomial formula (1.7) for non-integer numbers,

dh − 1 α
 
α
D f (x) = lim f (x) =
h→0 h
∞
X Γ(α + 1)
= lim h−α (−1)n d hα−n f (x) =
h→0 n! Γ(α − n + 1)
n=0
∞
X Γ(α + 1)
= lim h−α (−1)n f (x + (α − n)h) (4.5)
h→0 n! Γ(α − n + 1)
n=0

In the case of integer values the summatory only extends α terms and it
is equal to the ordinary derivative.
Finally, it is obvious that as h goes to 0 the last equation is equivalent
to the following
∞
X Γ(α + 1)
D α f (x) = lim h−α (−1)n f (x − nh) (4.6)
h→0 n! Γ(α − n + 1)
n=0

5 Functions of the Derivative

It is worth to mention that being (4.4) the expression of the α-st derivative
of a function the derivative itself is
dh − 1
D = lim (5.1)
h→0 h
And being the α-st derivative or iterating the differentiation α times
powering it to α, applying other functions to the derivative could be also
8 5 Functions of the Derivative

considered. If the function applied to the derivative can be expanded in

powers of x,
∞
X ∞
X
n
g(x) = an x =⇒ g(D)f (x) = an D n f (x) (5.2)
n=−∞ n=−∞

The result is a weighted sum of different order derivatives. These func-

tions of the derivative are usually known as “formal differential operators”.
As an example, the exponential of the derivative applied to the exponential
would give the following result that could be valuable for calculating func-
tions of the derivative when both f and g can be expanded in exponentials
∞
X an xn
eax = =⇒
n!
n=0
∞ ∞
X an D n ebx X an bn ebx
=⇒ eaD ebx = = = eab ebx = eb(x+a) (5.3)
n! n!
n=0 n=0

If both functions f and g can be expanded in positive powers of x,

∞
X ∞
X
n
g(x) = an x , f (x) = bn xn =⇒
n=0 n=0
∞
X ∞
X ∞
X ∞
X
n m
=⇒ g(D)f (x) = an D bm x = an bm D n xm =
n=0 m=0 n=0 m=0
∞ ∞
X X Γ(m + 1)
= an bm xm−n =
Γ(m − n + 1)
n=0 m=0
∞ ∞
X X Γ(l + n + 1) l
= an bl+n x =
Γ(l + 1)
n=0 l=−n
∞ ∞
X X (l + n)! l
= an bl+n x =
l!
n=0 l=0
∞ ∞
X xl X
= (l + n)! an bl+n (5.4)
l!
l=0 n=0

Thus,
∞
X
g(D)f (x) = cn xn
n=0
∞
1 X
cn = (n + m)! am bn+m (5.5)
n!
m=0

At a first glance this seems not very interesting, but interesting properties
could be hidden under this apparent mess. We shall return later to this point
with the help of other tools.
6 Grunwald-Letnikov Derivative 9

6 Grunwald-Letnikov Derivative
Grunwald-Letnikov derivative or also named Grunwald-Letnikov differinte-
gral, is a generalization of the derivative analogous to our generalization by
the binomial formula (4.6), but it is based on the direct generalization of the
equation (1.4). The idea behind is that h should approach 0 as n approaches
infinity,
n  
m n
X
n −n
D f (x) = lim h (−1) f (x − mh) =
h→0 m
m=0
n
X n!
= lim h−n (−1)m f (x − mh) =
h→0 m! Γ(n − m + 1)
m=0
∞
X n!
= lim h−n (−1)m f (x − mh) =
h→0 m! Γ(n − m + 1)
m=0
x−a
h
X n!
= lim h−n (−1)m f (x − mh) (6.1)
h→0 m! Γ(n − m + 1)
m=0
being a negative infinity. However, the generalization is done so that any
wished a less than x can be chosen. The above equation can be generalized
now to get a formula equivalent to (4.6)
x−a
h
α −α
X Γ(α + 1)
D f (x) = lim h (−1)m f (x − mh) (6.2)
h→0 m! Γ(α − m + 1)
m=0
or equivalently,
 n
α X  !
α n m Γ(α + 1) x−a
D f (x) = lim (−1) f x−m
n→∞ x−a m! Γ(α − m + 1) n
m=0
(6.3)
These generalizations have the inconvenient that if α is a negative integer
they are not valid for the gamma function is infinity in negative integers and
zero. However, for negative α values the following can be used
−n
Y n+m−1 n+m−1
Y Y
l (−l) (−1)m l
 
−n l=−n−m+1 l=n l=n
= = = =
m m! m! m!
(−1)m (n + m − 1)! (−1)m Γ(n + m)
= = (6.4)
m!(n − 1)! m! Γ(n)
so that for negative values equations (6.2) and (6.3) can be changed to
x−a
h
−α −α
X Γ(α + m)
D f (x) = lim h f (x − mh) (6.5)
h→0 m! Γ(α)
m=0
10 7 Riemann-Liouville Derivative

and
 n
α X  !
−α n Γ(α + m) x−a
D f (x) = lim f x−m (6.6)
n→∞ x−a m! Γ(α) n
m=0

The reason why x − a is considered in the above equations instead of a

single positive number can be seen considering the case of α = −1, which
must give the same result that integrating the function,

Zx n
−1
X Γ(m + 1)
f (t)dt = D f (x) = n→∞
lim h f (x − mh) =
h→0
m! Γ(1)
a m=0

n
X Znh
= n→∞
lim hf (x − mh) = n→∞
lim f (x − t)dt =
h→0 m=0 h→0
0
Zx
= n→∞
lim f (t)dt =⇒
h→0
x−nh
x−a x−a
=⇒ a = x − nh =⇒ n = , h= (6.7)
h n

Yes, in fractional derivatives limits must be considered as they must be in

integrals. Limits only vanish with integer order derivatives. This also means
that fractional derivatives are nonlocal, which may be the reason that makes
this kind of derivatives less useful in describing Nature. We shall see later
more about the integration -or better say differentiation- limits.
In the case of the derivative generalized by the binomial formula (4.6),
since n goes to infinity regardless of h, x − a must be infinity, so that the
derivative defined in (4.4) and (4.6) is equivalent to the Grunwald-Letnikov
derivative with a lower limit of negative infinity.

7 Riemann-Liouville Derivative
Riemann-Liouville derivative is the most used generalization of the deriva-
tive. It is based on Cauchy’s formula for calculating iterated integrals. If
the first integral of a function, which must equal to deriving it to −1, is as
follows

Zx
D −1 f (x) = f (t)dt (7.1)
0
7 Riemann-Liouville Derivative 11

the calculation of the second can be simplified by interchanging the integra-

tion order
ZxZt2 ZxZx Zx Zx
−2
D f (x) = f (t1 )dt1 dt2 = f (t1 )dt2 dt1 = f (t1 ) dt2 dt1 =
0 0 0 t1 0 t1
Zx
= f (t)(x − t)dt (7.2)
0

This method can be applied repeatedly, resulting in the following formula

for calculating iterated integrals,
Zx
−n 1
D f (x) = f (t)(x − t)n−1 dt (7.3)
(n − 1)!
0

Now this can be easily generalized to non-integer values, in what is the

Riemann-Liouville derivative,
Zx
α 1 f (t)
D f (x) = dt (7.4)
Γ(−α) (x − t)α+1
0

Note however that in the above formulas the election of 0 as the lower
limit of integration has been arbitrary, and any other number could be cho-
sen. Generally, the election of the integration limits in this and other gen-
eralizations of the derivative is indicated with subscripts. The Riemann-
Liouville derivative with the lower integration limit a would be
Zx
α 1 f (t)
a D x f (x) = dt (7.5)
Γ(−α) (x − t)α+1
a

The problem with this generalization is that if the real part of α is

positive or zero the integral diverges. So it only can be used to calculate
generalized integrals. However, this can be solved easily by deriving first by
ordinary derivative more than the amount necessary, thus making the re-
maining necessary differentiation negative and then applying the generalized
derivative for completing the rest in which will be a negative differentiation,
α
a D x f (x)= D n a Dxα−n f (x)
 
n = <(α) + 1 (7.6)

Once seen how integration limits must be specified in fractional deriva-

tives, we can return to the derivative of the exponential (2.1) and the deriva-
tive of powers (3.2) to see how their disagreement (3.7) was caused by having
12 8 Domain Transforms

different limits. With the exponential, supposing b a positive number,

Zx
−1 bx −1 bx ebx − eba eba
b e = a Dx e = ebx dx = =⇒ = 0 =⇒
b b
a
=⇒ a = −∞ (7.7)

while with the powers

Zx
xb+1 xb+1 − ab+1 ab+1
= a Dx−1 xb = xb dx = =⇒ = 0 =⇒
b+1 b+1 b+1
a
=⇒ a = 0 (7.8)
So both results were different in the case of the exponential function
because the limits of the derivatives were different. And the case of this dis-
crepancy is not casual. Differences in the results of derivatives with different
differentiation limits are indeed important.
Finally, it can be proved that the Grunwald-Letnikov derivative with
any given integration limits is equal to the Riemann-Liouville derivative
with the same limits for any complex number α with a negative real part.
This important result means that our derivative generalized by the binomial
formula in (4.4) and (4.6) is equivalent to the Riemann-Liouville derivative
with a lower limit of negative infinity provided that the real part of α is
negative,
Zx
dh − 1 α
 
α 1 f (t)
−∞ D x f (x) = lim f (x) = dt (7.9)
h→0 h Γ(−α) (x − t)α+1
−∞

This kind of Riemann-Liouville derivative with a lower limit of negative

infinity is known as the Weyl derivative.

8 Domain Transforms
The Laplace and Fourier transforms that serve to transform to the frequency
domain can be used to get generalizations of the derivative valid for functions
that allow such transformations. The Laplace transform is defined by
Z∞
e−tx f (x)dx

L f (x) = (8.1)
0

while its inverse transform is

a+i∞
Z
−1 1
etx f (x)dx

L f (x) = (8.2)
2πi
a−i∞
8 Domain Transforms 13

where a is chosen so that it is greather than the real part of any of the
singularities of f (x). An important property of the Laplace transform is
related to the transform of the n-st derivative of a function,

n−1
X
L D n f (x) = tn L f (x) − tm (D n−m−1 f )(0)
 
(8.3)
m=0

In the case that the terms in the summatory are zero the relation is
particularly simple, and for those kind of functions the derivative can be
generalized so that this property holds true for non-integer values of α

(D l 0 Dxα−m f )(0) = 0
 
l = 0...m − 1 m = <(α) + 1 =⇒
=⇒ L D α f (x) = tα L f (x)
 
(8.4)

for which the generalized derivative can be defined as

n o
D α f (x) = L−1 tα L f (x)

(8.5)

Keeping in mind the result of the generalized derivative of the exponen-

tial (2.1), the following development provides an easy understanding of the
reasons involved in the above generalization,
n  o
D α f (x) = D α L−1 L f (x) =
 
a+i∞
Z Z∞
 1
= Dα  etx e−tx f (x)dxdt =

2πi
a−i∞ 0
a+i∞
Z Z∞
1
= D α etx e−tx f (x)dxdt =
2πi
a−i∞ 0
a+i∞
Z Z∞
1
= α tx
t e e−tx f (x)dxdt =
2πi
a−i∞ 0
n o
= L−1 tα L f (x)

(8.6)

On the other hand, the Fourier transform is defined by

Z∞
e−itx f (x)dx

F f (x) = (8.7)
−∞

while its inverse transform is

Z∞
−1 1
eitx f (x)dx

F f (x) = (8.8)
2π
−∞
14 9 Convolution

This transform also has an analogous property related to the transform

of the n-st derivative of a function,

F D n f (x) = (it)n F f (x)

 
(8.9)

and the derivative can be generalized so that this property holds true for
non-integer values of α

F D α f (x) = (it)α F f (x)

 
(8.10)

yielding the following definition of the generalized derivative

n o
D α f (x) = F −1 (it)α F f (x)

(8.11)

The application of (2.1) can provide again an easy understanding of the

reasons involved in the generalization,
n  o
D α f (x) = D α F −1 F f (x) =
 
Z∞ Z∞
 1
= Dα  eitx e−itx f (x)dxdt =

2π
−∞ −∞
Z∞ Z∞
1
= α itx
D e e−itx f (x)dxdt =
2π
−∞ −∞
Z∞ Z∞
1
= (it)α eitx e−itx f (x)dxdt =
2π
−∞ −∞
n o
−1
(it)α F f (x)

= F (8.12)

In these two generalizations the implicit limits of differentiation should

be determined. In the case of Laplace transform, the generalized derivative
is a Riemann-Liouville derivative with a lower limit of 0, whereas in the
case of Fourier transform it is a Weyl derivative. Indeed, if we check for
instance the derivative generalized by the Fourier transform in the cases of
the sine and cosine functions -calculated in (2.4), (2.5), (2.8) and (2.9) with
the generalized derivative of the exponential that we have seen in (7.7) that
also is a Weyl derivative- we will find that they match perfectly.

9 Convolution
The generalizations of the derivative as expressed in (4.6) and (7.4) suggest
that they can be formulated in terms of the convolution, which would be
important for the convolution is a very simple operation in the frequency
9 Convolution 15

spaces achieved by Laplace and Fourier transforms. The following devel-

opment shows how this is the case, and how after all the derivative of a
function is its convolution with certain function,

xα−1
Φα (x) = (9.1)
Γ(α)

which Laplace convolution with f (x) yields the Riemann-Liouville derivative

of order −α

Zx Zx
(x − t)α−1
Φα (x) ∗ f (x) = Φα (x − t)f (t)dt = f (t)dt =
Γ(−α)
0 0
−α
= D f (x) (9.2)

And since the transform of the function expressed in (9.1) gives the
following simple result

( )
Γ(α + 1) xα−1
α
= t−α

L x = =⇒ L {Φα } = L (9.3)
tα+1 Γ(α)

and the transform of the convolution is the multiplication of the transforms,

in the case that the function f fulfills all the requirements given previously
for the simplicity of the Laplace transform of the derivative, we again get
the equation (8.4), which was the key for the generalization by Laplace
transform

n o
L D −α f (x) = L Φα (x) ∗ f (x) = L Φα (x) L f (x) =
  

= t−α L f (x)

(9.4)

These last results showing that the generalized derivative is a convolution

with certain function, open the interesting question of what other kind of
operators would be defined if functions other than (9.1) had been chosen.
The answer is that the new operators defined would be functions of the
derivative. To see this, we can exploit the linearity of the convolution,
16 9 Convolution

supposing that the function g can be expanded in powers of x

∞
X
g(x) = an xn =⇒
n=−∞
∞
X ∞
X
n
=⇒ g(D)f (x) = an D f (x) = an Φ−n (x) ∗ f (x) =
n=−∞ n=−∞
 
∞ −n−1
X an x  ∗ f (x) =
=
n=−∞
Γ(−n)
 
∞ n
X a−n−1 x 
= ∗ f (x) =
n=−∞
Γ(n + 1)

= h(x) ∗ f (x) (9.5)

where
∞
X
h(x) = bn xn
n=−∞
a−n−1
bn = (9.6)
Γ(n + 1)

This proves the equivalence between functions of the derivative and con-
volutions. For any convolution there is an equivalent function of the deriva-
tive if the function implied can be expanded in powers of x and vice versa.
Now, the Laplace transform shows what was expectable,
   
L g(D)f (x) = L h(x) ∗ f (x) = L h(x) L f (x) =
 
∞ n
 X a−n−1 x  
= L L f (x) =
 Γ(n + 1) 
n=−∞
∞
Xa−n−1 L {xn } 
= L f (x) =
n=−∞
Γ(n + 1)
∞
X a−n−1 
= n+1
L f (x) =
n=−∞
t
∞
X
an tn L f (x) =

=
n=−∞

= g(t)L f (x) (9.7)

since the linearity of the Laplace transform combined with the result of (8.5)
10 Cauchy Integral Formula 17

shows clearly that

∞ ∞ n
X X o
an D n f (x) = an L−1 tn L f (x)

g(D)f (x) = =
n=−∞ n=−∞
 
∞
 X 
= L−1 an tn L f (x)

=
 
n=−∞
n o
= L−1 g(t)L f (x)

(9.8)

as well as (8.11) shows for the Fourier transform that

∞ ∞ n
X X o
an D n f (x) = an F −1 (it)n F f (x)

g(D)f (x) = =
n=−∞ n=−∞
 
∞
 X 
= F −1 an (it)n F f (x)

=
 
n=−∞
n o
= F −1 g(it)F f (x)

(9.9)

These are useful tools for the calculation of functions of the derivative.
As an example, the following case is considered with the help of Fourier
transforms
n o
cos(D) sin(x) = F −1 cos(it)F sin(x)

=
n
o
= F −1 cos(it)πi δ(t + 1) − δ(t − 1) =
= cosh(1) sin(x) (9.10)
and directly as in (5.2)
∞ ∞
X (−1)n 2n X (−1)n
cos(D) sin(x) = D sin(x) = (−1)n sin(x) =
(2n)! (2n)!
n=0 n=0
∞
X 1
= sin(x) = cosh(1) sin(x) (9.11)
(2n)!
n=0

yielding both methods the same result. The case of (5.3) also matches when
calculated with Fourier transforms.

10 Cauchy Integral Formula

Another way of generalizing the derivative to non-integer order is given by
the Cauchy integral formula that plays a key role in complex analysis,
I
n! f (t)
D n f (z) = dt (10.1)
2πi (t − z)n+1
18 11 Properties

Despite its generalization to any complex number α seems straightfor-

ward, it must be taken into account that while being n integer there is an
isolated singularity at t = z, being it non-integer there is a branch point,
what means that the integration contour has to be chosen carefully. Other-
wise, the generalization only involves changing the factorial to the gamma
function, so defining what is known as the Cauchy-type fractional derivative
Z
α Γ(α + 1) f (t)
z0 D z f (z) = dt (10.2)
2πi (t − z)α+1
C (z0 , z + )

where supposing that the branch line starts at t = z and passes through z0 ,
the contour C starts at t = z0 , encircles t = z once in the positive sense and
returns to t = z0 where now the integrand has a different value.
It can be proved that this generalization of the derivative is equivalent to
the Riemann-Liouville derivative with a lower limit of z0 for the appropriate
values of α in which both derivatives are defined.

11 Properties
Fractional derivatives satisfy quite well all the properties that one could
expect from them, despite some of them are only characteristic of integer
order differentiation and some other have restrictions. For instance, the
property of linearity (1.10) is fulfilled, while that of the iteration (1.9) has
some restrictions in the cases that positive differentiation orders are present.
Some properties that include summatories can be generalized changing
the summatories into integrals. One such property is the expansion in Taylor
series
Z∞ t
a Dzf (z0 )
f (z) = (z − z0 )t dt (11.1)
Γ(t + 1)
−∞

and other is the Leibniz rule

Z∞
α Γ(α + 1)
a D z f (z)g(z) = D α−t f (z) a Dzt g(z)dt (11.2)
Γ(t + 1) Γ(α − t + 1) a z
−∞

These and other generalized properties can be applied to the study of

special functions, which often can be expressed in terms of simple formu-
las involving fractional derivatives. For instance, Gauss’s hypergeometric
function can be expressed as

Γ(a)x1−c b−c xb−1

2 F 1 (a, b; c; x) = 0 Dx (11.3)
Γ(b) (1 − x)a
12 Local Operators 19

12 Local Operators
Despite of some interesting results, we have left unanswered important ques-
tions about the functions of the derivative. For instance, it would be very
interesting to know if integer order derivatives are or not the only kind of
functions of the derivative that are local, and if are or not other kind of local
operators that are not functions of the derivative.
Looking at (5.2) it seems that in fact the only local functions of the
derivative are integer order derivatives and their finite sums, being non-
locality the result of adding infinite terms with displacement functions
steadily increasing the distance to the point in which the function is evalu-
ated. However, this impediment can be overcome by defining the derivative
in the following way

D α f (x) = lim D α f (x) (12.1)

a→x a x

so that the limit assures locality. In order to avoid another feature of frac-
tional derivatives that sometimes is not wanted -the fact that the derivative
of a constant is not zero- it is usual to define the local fractional derivative
as

D α f (x) = lim x Dyα f (y) − f (x)

(12.2)
y→x

or similar forms. So there can be operators being local and taking an infinite
number of evaluations of the function, provided that the points in which
these evaluations are carried out -or the value of their displacement function-
remains being infinitesimal. The question now is if there are other local
operators besides the local fractional derivative and functions of it.
It could be argued that if a function admits its expansion in Taylor
series, all the information about the function is indeed in its derivatives, so
that any operator giving information about the function will be unnecessary
and reducible to a sum of integer order derivatives. However, on the one
hand there are functions that are not differentiable but are local fractional
differentiable, and on the other hand, even if local operators could be reduced
to ordinary derivatives for differentiable functions, that does not assure that
ordinary derivatives are the most appropriate local operators for all the
tasks.
Characterizing all local operators is important, for locality makes of them
tools that could be useful in understanding Nature.