Fractional calculus:

basic theory and applications

(Part I)
Diego del-Castillo-Negrete
Oak Ridge National Laboratory
Fusion Energy Division
P.O. Box 2008, MS 6169
Oak Ridge, TN 37831-6169
phone: (865) 574-1127
FAX: (865) 576-7926
e-mail: delcastillod@ornl.gov

Lectures presented at the

Institute of Mathematics
UNAM. August 2005
Mexico, City. Mexico

Outline

1. Introduction
2. Fractional calculus
3. Fractional diffusion and random walks
4. Numerical methods
5. Applications:
a) Turbulent transport
b) Transport in fusion plasmas
c) Reaction-diffusion systems
6. Some references
 1. Introduction
Here we introduce the notion of fractional integral as a straightforward
generalization of the standard, integer-order integral, and define the
fractional derivative as the inverse operation. To motivate the concept we
discuss two examples: Abel’s equation and heat diffusion. The concepts
discussed here are further elaborated in the next section.

What is a fractional derivative?

dn f L’Hopital (1695):
dx n “What if n=1/2?”

Leibniz (1695):
“This is an apparent paradox from which, one day, useful
consequences will be drawn”

It is a usual practice to extend mathematical operations,

originally defined for a set of objects, to a wider set of objects
+
r r !R " s s !R
+
n! n !N " #(s) s ! R

! xn " n #N + $ ! x% " % #C
This mathematical “games” have, eventually, important
physical applications
 Fractional integrals
Integer order integration
x x1 x n"1

a D #="n
x $ dx $ dx L $ dx # ( x )
1 2 n n
a a a

Interchanging the order 1 x

n"1
of integration this can be aD # =
"n
x $
(n "1)! a
( x " y ) # ( y ) dy
rewritten
! as

Extending this expression for non-integer n=! we get

Riemann-Liouville
! fractional integral
x
1 # "1
a D $"#
x=
%(# )
& ( x " y) $ ( y ) dy
a

$(µ + 1)
Example: 0 Dx"# x µ = x µ +#
! $(µ + # + 1)

!
Fractional derivatives
Having defined the fractional integral, define the
fractional derivative as the inverse operation a Dxµ a Dx" µ # = #

dN
µ
a Dx " =
dx N
[ a Dx#$ " ] ! =N"µ N = smallest integer > µ
!
N
d dm
aD " =
m
x
dx N
[ a D #(N #m )
x "] = m "
dx
!
Riemann-Liouville fractional derivative
! x
1 # ( y)
"
aD # =
x &xm ' " +1%m dy m "1 < # $ m
$(m % " ) a ( x % y)

#( " + 1) !"# Dxµ e ikx = (ik ) µ e ikx

Examples: 0 Dxµ x " = x "$ µ
! #( " $ µ + 1)

!
!
 An example: Abel’s equation
y (x,h) Let T(h) be the time it takes a particle
to go down sliding on the curve x = " (y)
v What is the relation between the function
g T(h) and "(y) ?
1 2
x Energy v! = g ( h " y )
conservation 2

h T
ds 1+ " '2 dy 1+ " '2
v= = $ 2g(h!# y)
dy = $ dt = T
dt dt 0 0

h
1 f (y)
f (y) = 1+ " '2 T(h) = # h"y
dy
! 2g ! 0

!#1/ 2 f
1 1/ 2
! T = " 0Dh
f = 0 Dy T
"

!
!

T(x,t)
Heat diffusion
x Consider a bar that is heated on one
end by a time dependent source S(t).
S(t) What is the relation between S(t) and
the temperature T(x,t) of the bar?
Heat diffusion equation Boundary conditions Initial conditions
"t T = # " x2 T T(x = 0,t) = S(t) T(x,t = 0) = 0
#
Laplace transform Tˆ (x,s) = $e " st
T(x,t) dt s Tˆ = " #x2 Tˆ
0
! t x2 ! ! #
x $ S(% )
T= &e 4 # (t$% )
3 / 2 d%
"(t) = $ T(x,t) dx
2"# 0 (t $ % ) ! 0

!
"(t) = 2# 0 Dt$1/ 2 S(t)
! !

!
 2. Fractional calculus
Here we present the precise definitions of the left and right Riemann-
Liouville (RL) fractional integrals and derivatives, and briefly discuss
some of the main properties of these operators. Of particular interest is
the singular behavior of the RL operators at the boundaries, and the
definition of the fractional derivatives in the Caputo sense. Further details
on the material discussed here can be found in [Samko-etal-1993],
[Podlubny-1999] and references therein.

Riemann-Liouville integrals

a x b
" >0
x
1 # "1
Left integral a Dx"# $ =
%(# )
& ( x " y) $ ( y ) dy x>a
a

b
!
1 # "1
Right integral x D $ =
"#
b
%(# )
& ( y " x) $!( y ) dy x <b
x
!

Given the reflection operator Q " (x) = "!(a + b # x)

!

Q [ a Dx"# $ ] = x Db"# [Q $ ] Q [ x Db"# $ ] = a Dx"# [Q $ ]

!

! !
 a +b"x
1 $ ( u)
Q[ a D $ ] =
"#
x & 1"# du z = "u + a + b
%(# ) a (a + b " x " u)

1 b
$ ( a + b % z) !
! =
"(# )
& 1%# dz= x Db%# [Q $ ]
x (z % x)

! Semi-group property
The fractional integrals satisfy the following important semigroup
property

a Dx"# a Dx" $ % = a Dx"# " $ % " >0

" >0
x Db"# x Db" $ % = x Db"# " $ %
!
!
!
!

x w
1 1 dw % (u)
a D "#
x a D"$
x %=
&(# ) &($ )
' (x " w)1"#
' (w " u) 1" $
du
a a

w
! u=w Fubini’s theorem

x w x x

" dw " duL = " du " dwL

u a a a u
x
x x
1 1 dw
=
"(# ) "($ )
% du & (u) %
(x ' w) (w ' u)1' $
1'#
a u
!

$(" ) $(# ) B(" , # )

!Beta function B(", # ) =
( x $ u)
1$" $ #
$(" + # )
x
1 % (u)
=
# + $)
"(!
' (x & u) 1&# & $
du = a Dx&(# + $ )%
a !
 Riemann-Liouville derivatives
n = [" ] + 1
a x b [#]=integral part of #
!

x
1 &n # ( u)
Left derivative a D # ="
x
$(n % " ) & x n
' " %n +1 du
a ( x % u)
n b
Right derivative ($1) &n # ( u)
"
xD # =
b
%(n $ " ) & x n
' " $n +1 du
! x (u $ x )

Reciprocity
a Dx" a Dx#" $ = $

x z
1 &n $ ( u)
a
"
D
x a
#"
D $=
x
%(" )%(n # " ) & x n
' dz ' du " #n +1 1#"
! a a ( x # z) (z # u)

Exchanging the order of the integrals and using the definition of the
! Beta function we get:
x
1 %n x ' B(# ,n $ # ) * 1 #n $ ( u)
=
"(# )"(n $ # ) % x n
- du & (u))( (x $ u)1$n ,+ =
"(n) # x n
& du (x % u) 1%n
a a

x
1 # ( u)
using
(n "1)!
$ du (x " u) 1"n
= a Dx"n
a
! !
"n
= n a Dx#n $ = $
"x
!

!
 However

k #" j

D "#
D $ = $ (x) " % [ a D
# #" j
$]
( x " a) j "1 # $ < j
a x a x x
j=1 x= a
&(# " j + 1)

!
! The previous formulae can be generalized as

a Dx" a Dx# $ % = a Dx" # $ % " #$ #0

k #" j

D "# $
D % = aD $ "#
% (x) " &![ a D $" j
%]
( x " a)
!a x a x x x
'(# " j + 1)
j=1 x= a

Composition of fractional derivatives

"n
" xn
[ a Dx# $ ] =a Dxn +# $ " #0

n+1 ( j ) j+" +n
% # n$ ( $ (a) (!
x + a)
aD '
"
x n*
= a Dx $ (x) + -
" +n

j= 0 ,(1+ j + " + n )
&# x! ) n "1 # $ < n

m % j%"
( x % a)
!
a D"
x a [ D $] = a D
#
x
" +#
x $ (x) % ' [ a D #% j
x $ ] x= a
&(1% "!% j) m "1 # $ < m
j=1

!
! In particular, fractional [ a x $ ]x= a = 0
D "# j j = 1,Ln
derivatives commute
if and only if
a D "
[
x a D $] = a D
#
x
#
x [ a D $]
"
x [ a Dx" # j $ ] x= a = 0
!
j = 1,Lm
!

! ! !
 Behavior near the lower terminal
$
# (k ) # (k ) (a) k&"
aD # =%
" "
a Dx ( x & a)
Let " (a) k
" (x) = $ ( x % a) x
k!
k= 0
k! k= 0

using
! k#" k$"
! D" ( x # a) k = $( k + 1) &
( x $ a)
a x ( x # a) D "
# = ' # (a) (k )
$( k + 1# " ) a x
k= 0 %( k + 1$ " )

m&1
$ (k ) (a) 1
! limx "a a D $ = limx "a '
#
x m "1 # $ < m
!
%( k + 1& # ) ( x & a)# &k
k= 0

!
! #
limx "a a D $ = % unless
x
" (k ) (a) = 0 k = 1,Lm "1

!
! !

Fourier-Laplace transforms
%
$ 1
Fourier "ˆ (k) = %e ikx
" (x) dx " (x) = &e $ikx
"ˆ (k) dx
2# $%
#$

Dx$ % ] = ("i k ) !%ˆ (k) $ ] = (i k ) $ˆ (k)

$ #
! F [ "# F [ x D+"
#

$ c +i$

Laplace "ˆ (s) = %e # st

" (t) dt! " (t) = %e st
"ˆ (s) ds
! c#i$
0

n$1
n "1 # $ < n
L[ 0 D # ] = s #ˆ (s) $ % sk
"
t
"
[ 0 Dt" $k$1 # ] t= 0
! k= 0
!

! transform of
These are natural generalizations of the Fourier-Laplace
regular derivatives
!
Note that we have taken a = "# b=#
 Some limitations of the RL definition
Although a well-defined mathematical object, the Riemman-Liouville
definition of the fractional derivative has some problems when it comes
to apply it in physical problems. In particular:

The derivative of a A 1 This might be an issue when

a Dx" A = using RL operators for
constant is not zero #(1$ " ) ( x $ a)"
writing evolution equations

The RL derivative limx "a a Dx# $ = % This might be an issue

is in general!singular when applying boundary
at the lower limit unless " (k ) (a) = 0 conditions

The Laplace
!
transform
of the RL
! n$1 This might be an
" ˆ
derivative L D"
0 t[# = s # ]
(s) $ % sk [ 0 Dt" $k$1 # ]t= 0 Issue when solving
k= 0
depends on Initial value problems
the fractional
derivative at zero
!

Caputo fractional derivative

These problems can be resolved by defining the fractional operators in the
Caputo sense. Consider the case 1 < " < 2
k +2$"
# (a) 1 # '(a) 1
&
# (k +2) (a) ( x $ a)
"
aD # $
x $ ='
%(1$ " ) (x $ " )" %(2 $ " ) (x $ " )" $1 k= 0 %( k + 3 $ " )
!

singular terms regular terms

k +2$"
! &
# (k +2) (a) ( x $ a)
a
"
D
x [# (x) $ # (a) $ # (1)
(a)] = '
%( k + 3 $ " )
k= 0

Define the Caputo C

Dx" # = a Dx" [# (x) $ # (a) $ # (1) (a)]
derivative by subtracting 0

the
! singular terms
x ' x k
# ""(u) # (k +2) (a) ( x $ a)
using & du = ( & x $ u % $1 du
! a ( x $ u)% $1 k= 0
k! a ( )
x
C 1 # &&(u)
a Dx" # =
$(2 % " )
' " %1 du
a ( x % u)
!
 x
C 1 & un #
In general
a
"
D #=
x
$(n % " )
' " %n +1 du
a ( x % u)
x
C ($1) n & un #
x
"
D #=
b
%(n $ " )
' " $n +1 du
! a ( x $ u )

C
and as expected: a Dx" A = 0
!
limx "a a Dx# $ = 0
n$1
! L[ 0 D # ] = s #ˆ (s) $ % sk
"
t
"
[ 0 # (" $k$1) ] t= 0
k= 0
!
Note that in an infinite domain

!C D$ % = Dx$ % C
D"# $ = x D"# $
"# x "# x

! !

3. Fractional diffusion and

random walks
Here we discuss the close connection between fractional calculus (in
particular fractional diffusion equations) and the theory of continuous
time random walks (CTRW). We discuss the space-time scaling
properties of the solution of the fractional diffusion equation and the
relations with the Levy stable distributions. We present two examples of
Levy distributions one in the context of chaotic advection [del-Castillo-
Negrete-1998, 2000] and another in the context of financial mathematics
[Mantegna-Stanley-1994]. CTRWs were originally introduced in
[Montroll-Weiss-1965]. Further details can be found in the review papers
[Montroll-Shlesinger-1984], [Metzler-Klafter-2000], and references
therein. For a discussion in the context of turbulent transport in plasmas
see [del-Castillo-Negrete-etal-2004]. For a readable introduction to
stochastic processes and Levy distributions see [Paul-Baschnagel-1999],
[Bouchard- Georges-1990] and references therein. For a discussion on the
Green’s function of the fractional diffusion equation see [Mainardi et al.
Fractional Calculus and Applied Analysis, 4, 153-192 (2001)].
 The continuous time random walk
[Montroll-Weiss-1965]

!n ! n = waiting time ! (" ) = waiting time pdf

!n
! n = jump ! (" ) = jump size pdf

Master equation
% t
'% *
P(x,t ) = ! (x) & " (t' )dt' + & " (t # t' ) & $(x # x' ) P(x' ,t' )dx' dt'
)( ,+
t 0 #%

Contribution from particles Contribution from particles located at x’

that have not moved during (0,t) and jumping to x during (0,t)

The Montroll-Weiss master equation

ˆ 1! "˜ (s) 1
can be solved using P˜ (k,s) =
Laplace-Fourier transforms s 1 ! "˜ (s) #ˆ(k)

We want to take the long time, and large scale limit. To do this,
introduce the small parameters epsilon and delta and consider

1 $t' 1 $ x'
" * (t) = " & ) " * (x) = " & )
# %# ( # %# (

Then , the limit of interest is "#0 $ #0

In terms of Laplace transforms then
! !
L[" * (t)](s) = "ˆ (# s) F [ " * (x)](k) = "ˆ (# k )
!
1" #˜ ($ s) 1
rescaled MW equation Pˆ˜ * (k,s) =
! ! s 1" #˜ ($ s) %ˆ (& k)

!
 Gaussian Markovian case
Trapping pdf " (t) = µ e# µ $ Jumps pdf 1 2
"(x) = e%x / 2$
2# $
1
"˜ (# s) = $ 1% # s < t > +L 1/µ and $ transport scales
1+ # s < t >
!
2
k 2# 2 / 2 !
"ˆ (# k) = e
$ x
% 1$ # 2 x 2 k 2 /2 + L
! x2 #2
Continuum "#0 $ #0 "= finite
distinguished limit 2t $
!
ˆ ˆ
s P˜ "1 = " # k 2 P˜
!
!
Laplace
transform
[ ]
L ! t P˜ = s P˜ " # (x)
! !t P = " ! x2 P
Fourier
transform [ ]
F ! 2 x Pˆ = "k 2 Pˆ

Non-Gaussian, non-Markovian case

" (t) ~ t #( $ +1)
&
Trapping pdf "˜ (# s) $ 1%c1 (# s) + L
with “memory”

&
Jumps pdf "(x) ~ x
#(1+$ )
"ˆ (# k) $ 1%c 2 (# k ) + L
! !
x 2 #$
Continuum "#0 $ #0 "=
distinguished
! limit ! 2 t %&

ˆ ˆ
s ! P˜ " s ! "1 = " # k P˜
$

!
!
Laplace
transform [ "
]
L ! t P˜ = s P˜ # s $ (x)
" " #1

!t" P = # !|x$ | P
Fractional
Fourier
transform
[ ]
F !|x | Pˆ = # k
" "
Pˆ diffusion equation
 Solution of fractional diffusion equation
Consider the initial value problem of the asymmetric fractional diffusion
equation in an infinite domain

"t # = [ l $% D&x + r x D&% ] # " (x,t = 0) = " 0 (x)

(1" # ) (1+ # )
!l=" 2cos($% /2) r =" ! "1# $ # 1
2cos($% /2)

Using $ ! #
! F [ "# Dx$ %!
] (k) = ("i k ) % (k) F [ x D"# $ ] (k) = (i k ) $ (k)

! {[
"ˆ (k) = "ˆ 0 (k) exp!t l (#ik ) + r (ik )
$ $
]}

" "
% "$ k (
Then, using (mi k ) = k exp' #i *
& 2 k)

We can write the solution as "ˆ (k) = "ˆ 0 (k) Lˆ# $ (k)
!
2 , /6
where ˆL (k) = exp43 $ t k " .1+ i # k tan&( "% )+1 47
"# ! 4
5 - k ' 2 *0 48

And using the convolution theorem

!

'
1
" ( x,t ) =
2#
(L $% (x & x') " 0 (x') dx'
&'

!
 Some examples
Gaussian Levy
" =2 " = 1.5
# =1 # =1
$ =0 $ =0
L L
! !

Levy Levy
" = 1.5 " = 1.5
# =1 # =1
$ = 0.35 $ = 0.5
L
L
! !

Probabilistic interpretation
'
Probability 1
" ( x,t ) = of being at x " ( x,t ) =
2#
(L $% (x & x') " 0 (x') dx'
&'
at time t
Transition probability
! !
In the case alpha=2, theta=0 the solution reduces to
1 $ x2 '
L2 0 (x) = exp% #
{
Lˆ"# (k) = exp $ t k 2
} 2 "t
(
& 4t )

Which as expected corresponds to a normalized Gaussian

2
! with variance " =2 !t

Consistent with the Brownian random walk for " = 2 the transition
probability is Gaussian.
!
What is probabilistic interpretation for " # 2 ?
!

!
 Levy distributions
Consider a set of independent, identically distributed random variables

{ln }n=1,2LN Prob( l < ln < l + dl) = p(l)

Given this, we would like to know what is the probability distribution of the
normalized sum
! !
1 N
SN = # ln " AN
BN n=1
Prob( x < SN < x + dx ) = P(x) = ??

In the context of a random walk the answer to this question gives the
probability of finding a particle at a given point at a given time provided we
!
know
! the probability distribution of the individual steps
An answer to this question is provided by the Central Limit Theorem according
to which:
#
P(x) " Gaussian provided l2 = $l 2
p(l) dl< #
N#$
"#

However, what happens if l 2 = " ?????

! !

A probability density is stable if there are constants a and b such that

$
p(a1 l+ b1 )" p(a1 l+ b1 ) = % dl p[a (z # l)+ b ] p(a l + b ) = p(az + b)
1 1 2 2
#$

The importance of the stable distributions is given by the following

Theorem (Levy-Khintchine)
!
A probability density L(x) can be a limiting distribution
N
1 Prob( x < SN < x + dx ) " L(x)
SN =
BN
#l n " AN
N #$
n=1

only if it is stable

! Gaussian is stable, !
The and according to the CLT we know that is the
attractor of of all processes by finite variance.

However, the family of stable distributions is much bigger.

 Levy-Khintchine canonical representation:
%
ikx
A probability density L" # (k) = e = & dx L a# (x) e ikx
$%
- ' k ' "& **0
Is stable if and only if "
L" # (k) = exp/i$ k + % c k )1+ i # tan) ,,2
/. ( k ( 2 ++21
0 <" #2 !"1 < # < 1

! But according
! to !
our previous calculation this is the Green’s fuunction
of the space fractional diffusion equation.

So we have gone full circle: from stochastic process to fractional diffusion

equations and back to stochastic processes. We first derived the fractional
diffusion equation as a continuum limit for non-Gaussian random walks. Then,
we solved the fractional equation, and showed that the solution is given in
terms of Levy stable distributions.

Space-time fractional diffusion equation

C
0 D"t # = [ l $% D&x + r x D&% ] # ! (x, t = 0) = " (x)

1
Symmetric case l= r = "
2cos(#$ /2)
!

Solution in terms of s " #1

! (k, s) = " % ! (k, t = 0)
Fourier-Laplace
! transforms s +$ k

s! "1 zn Mittag-Leffler
[ !
]
E! ("c t ) (s) = !
s +c
E! ( z) = #
n "(! n + 1)
function

1 $
! ( x, t ) =
2" %#$
(
e # ikx E& # ' k t & dk
(
)
 Self-similarity and non-diffusive scaling

Similarity Fundamental
!= t" # / $ x " ( x, t ) = t # $ / % K (& )
variable solution

1 #
K (! ) = $ cos(! z) E% ( & ' z( ) dz
" 0
!

1 & x )
" ( x, # t ) = $ /%
"( $ / % ,t + Self-similar scaling
# '# *

Moments xn = n
# x " (x,t) dx = t n$ / %
#& n
K(&) d&
!
>1 Super-diffusion
n n " /# 2"
! x ~t
# Sub-diffusion
<1
!
! !
!

# determines the asymptotic scaling in space

Space similarity
! ( x, t ) = t "# / $ K (% ) != t" # / $ x
variable

1 #
K (! ) = $
" 0
cos ( ! z ) E% ( & ' z
(
) dz

!"# K (! ) ~ ! "( 1+ # )
!
For fixed t and large x
! (1+" )
~x
! ( x, t0 ) ~ x "( 1+# )
 % determines the asymptotic scaling in time

! ( x, t ) = | x | "1 # "$ / % K (# "$ / % ) ! = t | x | "# / $ Time similarity

variable

1 #
K (! ) = $ cos(! z) E% ( & ' z( ) dz
" 0

!"# K (! ) ~ ! "( 1+ # )

!"0 K (! ) ~ 1+ ! "(1"# ) !
For fixed x:
~ t! ~ t !"
& t " for t ~ 0
(
! ( x 0, t ) ~ '
()t # " for t $ % t

An example from finance

The simplest models of the fluctuations of assets prices assume an underlying
geometric Brownian motion, and lead to Gaussian distribution of price variations.

In 1963 by Mandelbrot pointed out

the existence of “fat tails” due to large events.
Based on this he proposed to use Levy stable
distributions to model prices.
Gaussian
Levy
Mategna and Stanley showed that the
probability distribution of price variations
of the S&P500 strongly departs from a
Gaussian and is well-fitted near the center
by a Levy stable distribution of index #=1.4.

[Mategna-Stanley, Nature 1995]

 Passive scalar diffusion
r r r
!r (t) = r (t) " r (0)
r = ensemble average
r (t)
r
M(t ) = !r = mean
r r 2
" 2 (t) = [!r # !r ] = variance
r
P(!r,t) = probability distribution

homogenous, Brownian M(t ) = V t

isotropic random lim ! 2 (t) = D t
t! "
turbulence walk r
P("r,t) = Gaussian

V= transport velocity D=diffusion coefficient

Coherent structures can give rise to

anomalous diffusion
x
trapping region trapping
transport event
region

flight
event
exchange region t

Coherent ! 2 (t) ~ t "

structures correlations lim
t! " P(x,t ) = non - Gaussian
 An example from chaotic advection

Quasigeostrophic flow

2! " v

The rotating annulus experiment

Solomon, Weeks, Swinney,

Phys. Rev. Lett. 71, 3975 (1993).

Test particle transport

Experiment Model

Solomon, Weeks, Swinney, PRL, 71, 3975 (1993) [del-Castillo-Negrete-1998]

 Non-Gaussian distribution functions

Probability distribution functions of particle

displacements.

P P

Fat tails

x x
Gaussian
[del-Castillo-Negrete-1998]
Solomon, Weeks, Swinney, PRL, 71, 3975 (1993)

Hamiltonian chaotic advection model

dx "# dy "#
=! =
dt "y dt "x
2
! = " ln [cosh x ] + % j =1 # j $ j (x) cosk j (y " c j t)

tanh(x) zonal flow regular neutral modes

u0

x
[del-Castillo-Negrete-1998]
 integrable flow
! 1 = " ln [cosh x ]+ # 1 !1 (x) cos k1 y + c1 x

shear flow transport

trapping region

separatrix

chaotic transport ! = !1 + " 2 ! 2 (x) cos( k2 y # $ t )

nonintegrable perturbation

chaotic orbit

exchange region
stochastic layer

Self-similar probability distributions

t! / 2 P
P

!x
! = ("x # "x ) t #$ / 2
X = !x " !x

collapse of curves ! P* (X, t) = t "# / 2 F ( X t "# / 2 )

X n ~ t n! / 2 *
P (X,t ) = !
" /2 *
P (!
"/2
X, ! t)