9611
9611
9611
com
https://ebookmeta.com/product/fractional-differential-
equations-an-approach-via-fractional-derivatives-1st-
edition-bangti-jin/
OR CLICK BUTTON
DOWLOAD NOW
https://ebookmeta.com/product/fractional-stochastic-differential-
equations-applications-to-covid-19-modeling-1st-edition-abdon-
atangana/
https://ebookmeta.com/product/numerical-methods-for-fractal-
fractional-differential-equations-and-engineering-1st-edition-
muhammad-altaf-khan/
https://ebookmeta.com/product/methods-of-mathematical-modelling-
fractional-differential-equations-mathematics-and-its-
applications-1st-edition-harendra-singh-editor/
https://ebookmeta.com/product/natures-patterns-and-the-
fractional-calculus-fractional-calculus-in-applied-sciences-and-
engineering-1st-edition-west/
Fractional Order Systems Ivo Petráš Editor
https://ebookmeta.com/product/fractional-order-systems-ivo-
petras-editor/
https://ebookmeta.com/product/differential-equations-a-linear-
algebra-approach-1st-edition-anindya-dey/
https://ebookmeta.com/product/stochastic-models-for-fractional-
calculus-mark-m-meerschaert/
https://ebookmeta.com/product/fractional-inequalities-in-banach-
algebras-1st-edition-george-a-anastassiou/
https://ebookmeta.com/product/an-elementary-course-on-partial-
differential-equations-aftab-alam/
Applied Mathematical Sciences
Bangti Jin
Fractional
Differential
Equations
An Approach via Fractional Derivatives
Applied Mathematical Sciences
Volume 206
Series Editors
Anthony Bloch, Department of Mathematics, University of Michigan, Ann Arbor,
MI, USA
C. L. Epstein, Department of Mathematics, University of Pennsylvania,
Philadelphia, PA, USA
Alain Goriely, Department of Mathematics, University of Oxford, Oxford, UK
Leslie Greengard, New York University, New York, NY, USA
Advisory Editors
J. Bell, Center for Computational Sciences and Engineering, Lawrence Berkeley
National Laboratory, Berkeley, CA, USA
P. Constantin, Department of Mathematics, Princeton University, Princeton, NJ,
USA
R. Durrett, Department of Mathematics, Duke University, Durham, CA, USA
R. Kohn, Courant Institute of Mathematical Sciences, New York University,
New York, NY, USA
R. Pego, Department of Mathematical Sciences, Carnegie Mellon University,
Pittsburgh, PA, USA
L. Ryzhik, Department of Mathematics, Stanford University, Stanford, CA, USA
A. Singer, Department of Mathematics, Princeton University, Princeton, NJ, USA
A. Stevens, Department of Applied Mathematics, University of Münster, Münster,
Germany
S. Wright, Computer Sciences Department, University of Wisconsin, Madison, WI,
USA
Founding Editors
F. John, New York University, New York, NY, USA
J. P. LaSalle, Brown University, Providence, RI, USA
L. Sirovich, Brown University, Providence, RI, USA
The mathematization of all sciences, the fading of traditional scientific boundaries,
the impact of computer technology, the growing importance of computer modeling
and the necessity of scientific planning all create the need both in education and
research for books that are introductory to and abreast of these developments. The
purpose of this series is to provide such books, suitable for the user of mathematics,
the mathematician interested in applications, and the student scientist. In particular,
this series will provide an outlet for topics of immediate interest because of the
novelty of its treatment of an application or of mathematics being applied or lying
close to applications. These books should be accessible to readers versed in
mathematics or science and engineering, and will feature a lively tutorial style, a
focus on topics of current interest, and present clear exposition of broad appeal.
A compliment to the Applied Mathematical Sciences series is the Texts in Applied
Mathematics series, which publishes textbooks suitable for advanced undergraduate
and beginning graduate courses.
Fractional Differential
Equations
An Approach via Fractional Derivatives
123
Bangti Jin
Department of Computer Science
University College London
London, UK
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
v
vi Preface
witnessed exciting developments in recent years. We will also brief ly touch upon
these topics to give a f lavor, illustrating distinct influences of the nonlocality of
fractional operators.
The term FDE is a big umbrella for any differential equations involving one or
more fractional derivatives. This textbook only considers the models that involve
only one fractional derivative in either spatial or temporal variable, mostly of the
Djrbashian-Caputo type, and only brief ly touches upon that of the Riemann-
Liouville type. This class of FDES has direct physical meaning and allows
initial/boundary values as for the classical integer-order differential equations, and
thus has been predominant in the mathematical modeling of many practical
applications.
The book consists of seven chapters and one appendix, which are organized as
follows. Chapter 1 briefly describes continuous time random walk, which shows
that the probability density function of certain stochastic process satisfies a FDE.
Chapters 2 and 3 describe, respectively, fractional calculus (various fractional
integral and derivatives), and two prominent special functions, i.e., Mittag-Leffler
function and Wright function. These represent the main mathematical tools for
FDES, and thus Chapters 2 and 3 build the foundation for the study of FDES in
Chapters 4–7. Chapters 4–5 are devoted to initial and boundary value problems for
fractional ODEs, respectively. Chapters 6–7 describe mathematical theory for
time-fractional diffusion (subdiffusion) in Hilbert spaces and Hölder spaces,
respectively. In the appendix, we recall basic facts of function spaces, integral
transforms and fixed point theorems that are extensively used in the book. Note that
Chapters 4–7 are organized in a manner such that they only loosely depend on each
other and the reader can directly proceed to the chapter of interest and study each
chapter independently, after going through Chapters 2–3 (and the respective parts
of the appendix).
The selected materials focus mainly on solution theory, especially the issues of
existence, uniqueness and regularity, which are also important in other related
areas, e.g., numerical analysis. Needless to say, these topics can represent only a
few small tips of current exciting developments within the FDE community, and
many important topics and results have to be left out due to page limitation. Also
we do not aim for a comprehensive treatment and describing the results in the most
general form, and instead we only state the results that are commonly used in order
to avoid unnecessary technicalities. (Actually each chapter can be expanded into a
book itself!) Whenever known, references for the stated results are provided, and
the pointers to directly relevant further extensions are also given. Nonetheless, the
reference list is nowhere close to complete and clearly biased by the author’s
knowledge.
There are many excellent books available, especially at the research level,
focusing on various aspects of fractional calculus or FDEs, notably [SKM93] on
Riemann-Liouville fractional integral/derivatives, [Pod99, KST06] on fractional
derivatives and ODEs, [Die10] on ODEs with Djrbashian-Caputo fractional derivative.
Our intention is to take a broader scope than the more focused research monographs
and to provide a relatively self-contained gentle introduction to current
Preface vii
Part I Preliminaries
1 Continuous Time Random Walk . . . . . . . ....... . . . . . . . . . . . . . . 3
1.1 Random Walk on a Lattice . . . . . . . . ....... . . . . . . . . . . . . . . 3
1.2 Continuous Time Random Walk . . . . ....... . . . . . . . . . . . . . . 6
1.3 Simulating Continuous Time Random Walk . . . . . . . . . . . . . . . . 16
2 Fractional Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1 Gamma Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Riemann-Liouville Fractional Integral . . . . . . . . . . . . . . . . . . . . . 21
2.3 Fractional Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.1 Riemann-Liouville fractional derivative . . . . . . . . . . . . . . . 28
2.3.2 Djrbashian-Caputo fractional derivative . . . . . . . . . . . . . . . 41
2.3.3 Grünwald-Letnikov fractional derivative . . . . . . . . . . . . . . 51
ix
x Contents
xiii
xiv Acronyms
Perhaps the simplest stochastic approach to derive equation (1.1) is to consider the
random walk framework on a lattice, which is also known as a Brownian walk. At
each time step (with a time step size Δt), the walker randomly jumps to one of its
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 3
B. Jin, Fractional Differential Equations, Applied Mathematical Sciences 206,
https://doi.org/10.1007/978-3-030-76043-4_1
4 1 Continuous Time Random Walk
four nearest neighboring sites on a square lattice (with a space step Δx); see Fig. 1.1
for a schematic illustration.
Fig. 1.1 Random walk on a square lattice, starting from the circle.
where the index j denotes the position on the lattice (grid), and p j (t) the probability
that the walker is at grid j at time t. It relates the probability of being at position j at
time t + Δt to that of the two adjacent sites j ± 1 at time t. The factor 12 expresses the
directional isotropy of the jumps: the jumps to the left and right are equally likely.
A rearrangement gives
p j (t + Δt) − p j (t) 1 p j−1 (t) − 2p j (t) + p j+1 (t)
= .
(Δx)2 2 (Δx)2
The right-hand side is the central finite difference approximation of the second-order
2 p(x, t) at grid j. In the limit Δt, Δx → 0+ , if the pdf p(x, t) is smooth
derivative ∂xx
in x and t, discarding the higher order terms in Δt and Δx in the following Taylor
expansions
leads to
∂t p(x, t) = κ∂xx
2
p(x, t). (1.2)
The limit is taken such that the quotient
is a positive constant, and κ is called the diffusion coefficient, which connects the
spatial and time scales.
1.1 Random Walk on a Lattice 5
Equation (1.2) can also be regarded as a consequence of the central limit theorem.
Suppose that the jump length Δx has a pdf given by λ(x) so that for a < b,
∫b
P(a < Δx < b) = a λ(x)dx. Then Fourier transform gives
∫ ∞ ∫ ∞
=
λ(ξ) e−iξ x λ(x)dx = 1 − iξ x − 12 ξ 2 x 2 + . . . λ(x)dx
−∞ −∞
= 1 − iξ μ1 − 12 ξ 2 μ2 + . . . ,
∫∞
where μ j is the jth moment μ j = −∞ x j λ(x)dx, provided that these moments do
exist, which holds if λ(x) decays sufficiently fast to zero as x → ±∞. Further, assume
that the pdf λ(x) is normalized and even, i.e., μ1 = 0, μ2 = 1 and μ3 = 0. Then
= 1 − 1 ξ 2 + O(ξ 4 ).
λ(ξ) 2
Denote by Δxi the jump length taken at the ith step. In the random walk model,
the steps Δx1 , Δx2 , . . . are independent. The sum of the independent and identically
distributed (i.i.d.) random variables Δxn (according to the pdf with a rescaled λ(x))
gives the position of the walker after n steps. It is also a random variable. Now we
recall a standard result on the sum of two independent random variables, where ∗
denotes convolution of f and g.
Theorem 1.1 If X and Y are independent random variables with a pdf given by f
and g, respectively, then the sum Z = X + Y has a pdf f ∗ g.
n,
By Theorem 1.1, the random variable Xn has a Fourier transform pn (ξ) = (λ(ξ))
and the normalized sum n− 2 xn has the Fourier transform
1
1 n
pn n− 2 ξ = 1 − 2n ξ + O(n−2 ) .
1 2
ξ2
Taking the limit as n → ∞ gives p(ξ) = e− 2 and inverting the Fourier transform
x2
gives the standard Gaussian distribution p(x) = (2π)− 2 e− 2 , cf. Example A.4 in the
1
appendix. This is precisely the central limit theorem, asserting that the long-term
average behavior of i.i.d. random variables is Gaussian. One requirement for the
whole procedure to work is the finiteness of the second moment μ2 of λ(x).
Now we interpret xn as the particle position after n time steps, by correlating the
time step size Δt with the variance of Δx according to the ansatz (1.3). This can be
easily achieved by rescaling the variance of λ(x) to 2κt. Then by the scaling rule for
the Fourier transform, pn (ξ) is given by
and taking limit as n → ∞ gives p(ξ) = e−ξ κt . By inverse Fourier transform, the
2
pdf being at certain position x and time t is governed by the standard diffusion model
6 1 Continuous Time Random Walk
It is the fundamental solution, i.e., the solution p(x, t) of (1.2) with the initial condi-
tion p(x, 0) = δ(x), the Dirac function concentrated at x = 0. Note that at any fixed
time∫t > 0, p(x, t) is a Gaussian distribution in x, with mean zero and variance 2κt,
∞
i.e., −∞ x 2 p(x, t)dx = 2κt, which scales linearly with the time t. This represents one
distinct feature of normal diffusion processes.
There are several stochastic models for deriving differential equations involving a
fractional-order derivative. We describe the continuous time random walk framework
(ctrw) due to Montroll and Weiss [MW65]. ctrw generalizes the random walk
model in Section 1.1, in which the length of each jump and the time elapsed between
two successive jumps follow a given pdf. We assume that these two random variables
are independent, though in theory one can allow correlation in order to achieve more
flexible modeling. In one spatial dimension, the picture is as follows: a walker moves
along the x-axis, starting at a position x0 at time t0 = 0. At time t1 , the walker jumps
to x1 , then at time t2 jumps to x2 , and so on. We assume that the temporal and spatial
increments Δtn = tn −tn−1 and Δxn = xn − xn−1 are i.i.d. random variables, following
pdfs ψ(t) and λ(x), respectively, known as the waiting time distribution and jump
length distribution, respectively. Namely, the probability of Δtn lying in any interval
[a, b] ⊂ R+ and Δxn lying in any interval [c, d] ⊂ R are given by
∫ b ∫ d
P(a < Δtn < b) = ψ(t)dt and P(c < Δxn < d) = λ(x)dx,
a c
respectively. Now the goal is to determine the probability that the walker lies in a
given spatial interval at time t. For given pdfs ψ and λ, the position x of the walker
can be regarded as a step function of t.
Example 1.1 Suppose that the distribution ψ(t) is exponential with a parameter τ > 0,
t
i.e., ψ(t) = τ −1 e− τ for t ∈ R+ , and the jump length distribution λ(x) is Gaussian
x2
−
with mean zero and variance σ 2 , i.e., λ(x) = (2πσ 2 )− 2 e
1
2σ 2 for x ∈ R. Then the
waiting time Δtn and jump length Δxn satisfy
where E denotes taking expectation with respect to the underlying distribution. The
position x(t) of the walker is a step function (in time t). Two sample trajectories of
ctrw with τ = 1 and σ = 1 are given in Fig. 1.2.
1.2 Continuous Time Random Walk 7
20 20
0 0
x
x
-20 -20
0 20 40 60 80 100 0 20 40 60 80 100
t t
Fig. 1.2 Two realizations of 1D ctrw, with exponential waiting time distribution (with τ = 1) and
Gaussian jump length distribution (with σ = 1), starting from x0 = 0.
Now we derive the pdf of the total waiting time tn after n steps and the total jump
length xn − x0 after n steps. The main tools in the derivation are Laplace and Fourier
transforms in Appendices A.3.1 and A.3.2, respectively. We denote by ψn (t) the pdf
of tn = Δt1 + Δt2 + . . . + Δtn , and ψ1 = ψ. By Theorem 1.1, we have
∫ t
ψn (t) = (ψn−1 ∗ ψ)(t) = ψn−1 (s)ψ(t − s)ds.
0
Next we derive the pdf of the jump length xn − x0 after n steps. We denote by
λn (x) the pdf of the random variable xn − x0 = Δx1 + Δx2 + . . . + Δxn . Appealing
to Theorem 1.1 again yields
8 1 Continuous Time Random Walk
∫ ∞
λn (x) = (λn−1 ∗ λ)(x) = λn−1 (y)λ(x − y)dy,
−∞
n (ξ) is given by
and by the convolution rule for Fourier transform, λ
λ n,
n (ξ) = (λ(ξ)) n ≥ 0.
We denote by p(x, t) the pdf of the walker at the position x at time t. Since χn (t)
is the probability of taking n steps up to time t,
∞
p(x, t) = λn (x) χn (t).
n=0
and hence we obtain the following fundamental relation for the pdf p(x, t) in the
Laplace-Fourier domain
1 − ψ(z) 1
p(ξ, z) = . (1.6)
z ψ(z)
1 − λ(ξ)
t
Example 1.2 For the waiting time distribution ψ(t) = τ −1 e− τ and the jump length
2
− x2
distribution λ(x) = (2πσ 2 )− 2 e in Example 1.1, we have ψ(z) = (1 + τz)−1 and
1
2σ
σ2 ξ 2
= e− 2 . Thus, the Fourier-Laplace transform p(ξ, z) of the pdf p(x, t) of the
λ(ξ)
walker at position x (relative to x0 ) at time t is given by
σ2 ξ 2
p(ξ, z) = τ(1 + τz − e− 2 )−1 .
1.2 Continuous Time Random Walk 9
being finite or diverging. Below we shall discuss the following three scenarios: (i)
Finite and ς 2 , (ii) Diverging and finite ς 2 and (iii) Diverging ς 2 and finite .
(i) Finite characteristic waiting time and jump length variance. When both
characteristic waiting time and jump length variance ς are finite, ctrw does not
lead to anything new. Specifically, assume that the pdfs ψ(t) and λ(x) are normalized:
∫ ∞ ∫ ∞ ∫ ∞
tψ(t)dt = 1, xλ(x)dx = 0, x 2 λ(x)dx = 1. (1.7)
0 −∞ −∞
These conditions are satisfied after suitable rescaling if the waiting time pdf ψ(t) has
a finite mean, and the jump length pdf λ(x) has a finite variance. Recall the relation
(1.5): ψ(0) = 1 = λ(0). Since
∫ ∞
dk ψ
= e−zt (−t)k ψ(t)dt,
dz k 0
Similarly, we deduce
∫ ∞ ∫ ∞
λ (0) = −i xλ(x)dx = 0, (0) = −
λ x 2 λ(x)dx = −1.
−∞ −∞
Next, for τ > 0 and σ > 0, let the random variables Δtn and Δxn follow the
rescaled pdfs
Consequently,
1 − ψ(τz) 1
p(ξ, z; σ, τ) = .
z
1 − ψ(τz)λ(σξ)
The Taylor expansion of ψ(z) around z = 0 yields
10 1 Continuous Time Random Walk
The scaling relation here for κ is identical to that in the random walk framework, cf.
(1.3). Upon inverting the Laplace transform, we find
∫
1 ezt
dz = e−κ ξ t ,
2
p(ξ, t) =
2πi C z + κξ 2
and then inverting the Fourier transform gives the familiar Gaussian pdf p(x, t) =
x2
(4πκt)− 2 e− 4κ t . Last we verify that the pdf p(x, t) satisfies (1.2):
1
LF [∂t p − κ∂xx
2
p](ξ, z) = z p(ξ, z) − p(ξ, 0) + κξ 2 p(ξ, z)
= (z + κξ 2 ) p(ξ, z) − p(ξ, 0) = 1 − p(ξ, 0) = 0,
where the last step follows by p̃(ξ, 0) = δ̃(ξ) = 1. Hence, the pdf p(x, t) for x(t) − x0
at the position x (relative to x0 ) of the walker at time t indeed satisfies (1.2). The
ctrw framework recovers the classical diffusion model, as long as the waiting time
pdf ψ(t) has a finite mean and the jump length pdf λ(x) has finite first and second
moments. Further, the displacement variance is given by
∫ ∞
μ2 (t) = E p(t) [(x(t) − x0 )2 ] = x 2 p(x, t)dx = 2κt,
−∞
which grows linearly with the time t. One striking thing is that any pair of pdfs with
finite characteristic waiting time and jump length variance ς 2 lead to the same
result, to the lowest order.
(ii) Divergent mean waiting time. Now we consider the situation where the charac-
teristic waiting time diverges, but the jump length variance ς 2 is finite. This occurs
for example when the particle might be trapped in a potential well, and it takes a
long time for the particle to leave the well. To model such phenomena, we employ
1.2 Continuous Time Random Walk 11
a heavy-tailed (i.e., the tail is not exponentially bounded) waiting time pdf with the
asymptotic behavior
ψ(t) ∼ at −1−α, as t → ∞ (1.11)
for some α ∈ (0, 1), where a > 0 is a constant. One example of such pdf is
ψ(t) = α(1 + t)−1−α . The (asymptotic) power law decay in (1.11) is heavy tailed and
allows occasional very large waiting time between consecutive jumps. Again, the
specific form of ψ(t) is irrelevant, and only the power law decay at large time matters.
The parameter α determines the asymptotic decay of the pdf ψ(t): the closer is α to
zero, the slower the decay is and more likely the long waiting time is.
Now the question is whether such occasional long waiting time will change
completely the dynamics of the stochastic process∫in the large time. For the power
∞
law decay, the mean waiting time is divergent, i.e., 0 tψ(t)dt = +∞. So one cannot
assume the normalizing condition on the pdf ψ in (1.7), and the above analysis
breaks
∫ ∞down. Nonetheless,∫ in this part, the assumption on λ(x) remains unchanged,
∞
i.e., −∞ xλ(x)dx = 0 and −∞ x 2 λ(x)dx = 1.
15
10
5
x
-5
0 1000 2000 3000
t
(a) (b)
Fig. 1.3 ctrw with a power law waiting time pdf. Panel (a) gives one sample trajectory with
waiting time pdf ψ(t) = α(1 + t)−1−α , α = 34 , and standard Gaussian jump length pdf; Panel (b) is
a zoom-in of the first cluster of Panel (a).
In Fig. 1.3, we show one sample trajectory of ctrw with a power law waiting time
pdf ψ(t) = α(1 + t)−1−α , α = 34 , and the standard Gaussian jump length pdf, and an
enlarged version of the initial time interval. One observes clearly the occasional but
large waiting time appearing at different time scales, cf. Fig. 1.3(b). This behavior
is dramatically different from the case of finite mean waiting time in Fig. 1.2.
The following result bounds the Laplace transform ψ(z) for z → 0. The notation
Γ(z) denotes the Gamma function; see (2.1) in Section 2.1 for its definition.
Theorem 1.2 Let α ∈ (0, 1), and a > 0. If ψ(t) = at −1−α + O(t −2−α ) as t → ∞, then
ψ(z) = 1 − Bα z α + O(z) as z → 0
and we consider 0 < z < t0−1 so that zt0 < 1. Since ψ(0) = 1, we have
∫ ∞ ∫ t0
1 − ψ(z) = (1 − e−zt )ψ(t)dt = (1 − e−zt )(ψ(t) − at −1−α )dt
0 0
∫ ∞ ∫ ∞ 3
−zt −1−α
+ (1 − e )(ψ(t) − at )dt + (1 − e−zt )at −1−α dt =: Ii .
t0 0 i=1
Since 0 ≤ 1 − e−s ≤ s for all s ∈ [0, 1], the term I1 can be bounded by
∫ t0 ∫ t0
|I1 | ≤ ztψ(t)dt + ztat −1−α dt
0 0
∫ ∞ ∫ t0
≤ zt0 ψ(t)dt + az t −α dt ≤ cα,t0 z.
0 0
For the term I2 , using (1.12) and changing variables s = zt, we deduce
∫ ∞
|I2 | ≤ (1 − e−zt )ct −2−α dt
t0
∫ 1 ∫ ∞
≤ cz 1+α s−1−α ds + s−2−α ds
t0 z 1
−1 −α
≤ cz 1+α
(α ((t0 z) − 1) + (1 + α)−1 ) ≤ cα,t0 z.
Now we introduce the following rescaled pdfs for the incremental waiting time
Δtn and jump length Δxn : ψτ (t) = τ −1 ψ(τ −1 t) and λσ (x) = σ −1 λ(σ −1 x). By (1.6),
p(ξ, z; σ, τ) is given by
1 − ψ(τz) 1
p(ξ, z; σ, τ) = .
z
1 − ψ(τz)λ(σξ)
1.2 Continuous Time Random Walk 13
Under the power law waiting time pdf, from Theorem 1.2 we deduce ψ(z) = 1 −
= 1 − 1 ξ 2 + O(ξ 4 ) as ξ → 0. Hence,
Bα z α + O(z) as z → 0 and λ(ξ) 2
and then by the Fourier transform of the Wright function Wρ,μ (z) in Proposition 3.4,
we obtain an explicit expression of the pdf p(x, t)
With α = 1, this formula recovers the familiar Gaussian density. Then Lemma 2.9
implies that the pdf p(x, t) satisfies
C α
LF 0 Dt p(x, t) − κα ∂xx
2
p(x, t) = z α p(ξ, z) − z α−1 p(ξ, 0) + κα ξ 2 p(ξ, z)
= (z α + κα ξ 2 ) p(ξ, z) − z α−1 p(ξ, 0) ≡ 0,
since p(ξ, 0) =
δ(ξ) = 1. Thus the pdf p(x, t) satisfies the following time-fractional
diffusion equation with a Djrbashian-Caputo fractional derivative C α
0 Dt p in time t
(see Definition 2.3 in Chapter 2 for the precise definition)
C α
0 Dt p(x, t) − κα ∂xx
2
p(x, t) = 0, 0 < t < ∞, −∞ < x < ∞.
∫ ∞
d2
μ2 (z) = x 2 p(x, z)dx = − p(ξ, z)|ξ=0
−∞ dξ 2
d2
=− (z + κα z 1−α ξ 2 )−1 |ξ=0 = 2κα z −1−α,
dξ 2
which upon inverse Laplace transform yields μ2 (t) = 2Γ(1 + α)−1 κα t α . Thus the
mean square displacement grows only sublinearly with the time t, which, at large
time t, is slower than that in normal diffusion, and as α → 1− , it recovers the formula
for normal diffusion. Such a diffusion process is called subdiffusion or anomalously
slow diffusion in the literature. Subdiffusion has been observed in a large number
of physical applications, e.g., column experiments [HH98], thermal diffusion in
fractal domains [Nig86], non-Fickian transport in geological formations [BCDS06],
and protein transport within membranes [Kou08]. The mathematical theory for the
subdiffusion model will be discussed in detail in Chapters 6 and 7.
(iii) Diverging jump length variance. Last we turn to the case of diverging jump
length variance ς 2 , but finite characteristic waiting time . To be specific, we assume
an exponential waiting time ψ(t), and that the jump length follows a (possibly asym-
metric) Lévy distribution. See the monograph [Nol20] for an extensive treatment of
univariate stable distributions and [GM98] for the use of stable distributions within
fractional calculus.
The most convenient way to define Lévy random variables is via their character-
istic function (Fourier transform)
where μ ∈ (0, 2] and β ∈ [−1, 1]. The parameter μ determines the rate at which the
tails of the pdf taper off. When μ = 2, it recovers a Gaussian distribution, irrespective
of the value of β, whereas with μ = 1 and β = 0, the stable density is identical to
the Cauchy distribution, i.e., λ(x) = (π(1 + x 2 ))−1 . The parameter β determines the
degree of asymmetry of the distribution. When β is negative (respectively positive),
the distribution is skewed to the left (respectively right). When β = 0, the expression
μ) = e− |ξ | μ .
simplifies to λ(ξ;
Generally, the pdf of a Lévy stable distribution is not available in closed form.
However, it is possible to compute the density numerically [Nol97] or to simulate
random variables from such stable distributions, cf. Section 1.3. Nonetheless, fol-
lowing the proof of Theorem 1.2, when 1 < μ < 2, one can obtain an inverse power
law asymptotic [Nol20, Theorem 1.2, p. 13]
for some aμ,β > 0, from which it follows directly that the jump length variance
diverges, i.e., ∫ ∞
x 2 λ(x) dx = ∞. (1.16)
−∞
1.2 Continuous Time Random Walk 15
In general, the pth moment of a stable random variable is finite if and only if p < μ.
50
40
30
x
20
10
0
0 50 100 150 200 250
t
(a) (b)
Fig. 1.4 ctrw with a Lévy jump length pdf. Panel (a) is one sample trajectory with an exponential
waiting time pdf ψ(t) = e−t , and a Lévy jump length pdf with μ = 32 and β = 0. Panel (b) is a
zoom-in of Panel (a).
For ctrw, one especially relevant case is λ(x; μ, β) with μ ∈ (1, 2] for the jump
length distribution. In Fig. 1.4, we show one sample trajectory of ctrw with a
Lévy jump length distribution. Due to the asymptotic property (1.15) of the jump
length pdf, very long jumps may occur with a much higher probability than for an
exponentially decaying pdf like a Gaussian. The scaling nature of the Lévy jump
length pdf leads to clusters along the trajectory, i.e., local motion is occasionally
interrupted by long sojourns, at all length scales.
Below we derive the diffusion limit using the scaling technique, by appealing
to the scaled pdfs ψ∫τ (t) and λσ (x), for some τ, σ > 0. The pdf ψ is assumed to
∞
be normalized, i.e., 0 tψ(t)dt = 1. Using Taylor expansion, for small ξ, we have
(μ 1)
μπ
e−λ(ξ;μ,β) = 1 − |ξ | μ 1 − iβ sign(ξ) tan + O(|ξ | 2μ ).
2
Hence, for μ 1, we have
ψτ (z) = 1 − τz + O(τ 2 z 2 ) as τ → 0,
σ (ξ) = 1 − σ μ |ξ | μ 1 − iβ sign(ξ) tan μπ + O(σ 2μ |ξ | 2μ ) as σ → 0.
λ
2
These two identities and Fourier-Laplace convolution formulas yield
1
p(ξ, z) = lim p(ξ, z; σ, τ) = , (1.17)
z + κμ |ξ | μ 1 − iβ sign(ξ) tan μπ
2
where the limit is taken under the condition κμ = limσ→0+,τ→0+ τ −1 σ μ for some
κμ > 0, which represents the diffusion coefficient.
Next we invert the Fourier transform of p(ξ, z) and obtain
16 1 Continuous Time Random Walk
which is the characteristic function of a Lévy stable distribution, for which an explicit
expression is generally unavailable. This can be regarded as a generalized central
limit theorem for stable distributions.
One can verify by Fourier and Laplace inversion that the pdf p(x, t) satisfies
κ μ μ
∂t p(x, t) = 2μ (1 − β)−∞RDx p(x, t) + (1 + β)Rx D∞ p(x, t) ,
To simulate ctrw, one needs to generate random numbers for a given pdf, e.g.,
power laws and Lévy stable distributions. The task of generating random numbers
for an arbitrary pdf is often highly nontrivial, especially in high-dimensional spaces.
There are a number of possible methods, and we describe only the transformation
1.3 Simulating Continuous Time Random Walk 17
method, which is simple and easy to implement. For more advanced methods, we
refer to the monograph [Liu08].
The transformation method requires that we have access to random numbers
uniformly distributed in the unit interval 0 ≤ r ≤ 1, which can be generated by any
standard pseudo-random number generators. Now suppose that p(x) is a continuous
pdf from which we wish to draw samples. The pdfs p(x) and p(r) (the uniform
distribution on [0, 1]) are related by
dr dr
p(x) = p(r) = ,
dx dx
where the second equality follows because p(r) = 1 over the interval [0, 1]. Integrat-
ing both sides with respect to x, we obtain the complementary cumulative distribution
function F(x) in terms of r:
∫ ∞ ∫ 1
F(x) = p(x ) dx = dr = 1 − r,
x r
and the inverse function F −1 (x) is given by F −1 (x) = x − α − 1. This last formula
1
provides an easy way to generate random variables from the power law density ψ(t),
needed for simulating the subdiffusion model.
Generating samples from the Lévy stable distribution λ(x; μ, β) is more delicate.
The main challenge lies in the fact that there are no analytic expressions for the
inverse F −1 , and thus the transformation method is not easy to apply. Two well-
known exceptions are the Gaussian (μ = 2) and Cauchy (μ = 1, β = 0). One popular
algorithm is from [CMS76, WW95]. For μ ∈ (0, 2] and β ∈ [−1, 1], it generates a
random variable X ∼ λ(x; μ, β) as in Algorithm 1.
Exercises
where
arctan(β tan μπ μπ 2μ1
2 )
Bμ, β = and Cμ, β = 1 + β 2 tan2 .
μ 2
3: For μ = 1, compute
2 π W cos V
X= + βV tan V − β ln π .
π 2 2 + βV
Exercise 1.4 Prove that the pth moment of a stable random variable is finite if and
only if p < μ.
Exercise 1.6 Develop an algorithm for generating random variables from the Cauchy
distribution λ(x) = (π(1 + x 2 ))−1 and implement it.
Exercise 1.7 Implement the algorithm for generating random variables from Lévy
stable distribution, with μ = 1.5, and β = 0.
Exercise 1.8 Derive the diffusion limit for the ctrw with a power law waiting time
and Lévy jump length distribution, and the corresponding governing equation.
Chapter 2
Fractional Calculus
This function will appear in nearly every formula in this book! The usual way to
define the Gamma function Γ(z) is
∫ ∞
Γ(z) = t z−1 e−t dt, (z) > 0. (2.1)
0
Since Γ(1) = 1, (2.2) implies that for any positive integer n ∈ N, Γ(n + 1) = n!.
The following reflection formula [AS65, 6.1.17, p. 256]
π
Γ(z)Γ(1 − z) = , 0 < (z) < 1, (2.3)
sin(πz)
and Legendre’s duplication formula [AS65, 6.1.18, p. 256]
√
Γ(z)Γ(z + 12 ) = π21−2z Γ(2z) (2.4)
are very useful in practice. For example, one only needs to approximate √ Γ(z) for
(z) ≥ 12 and uses (2.3) for (z) < 12 . It also gives directly Γ( 12 ) = π.
It is possible to continue Γ(z) analytically into the left half complex plane C− :=
{z ∈ C : (z) < 0}. This is often done by representing the reciprocal Gamma
1
function Γ(z) as an infinite product (cf. Exercise 2.2)
1 n−z
= lim z(z + 1) . . . (z + n), ∀z ∈ C. (2.5)
Γ(z) n→∞ n!
1
Hence, Γ(z) is an entire function of z with zeros at z = 0, −1, −2, . . ., and accordingly,
Γ(z) has poles at z = 0, −1, . . .. The integral representation for Γ(z)
1
due to Hermann
Hankel [Han64] is very useful. It is often derived using Laplace transform (cf.
Appendix A.3). Substituting t = su in (2.1) yields
∫ ∞
Γ(z)
= uz−1 e−su du,
sz 0
which can be regarded as the Laplace transform of uz−1 for a fixed z ∈ C. Hence
uz−1 can be interpreted as an inverse Laplace transform
∫
z−1 −1 Γ(z) 1 Γ(z)
u =L = esu z ds,
sz 2πi C s
where C is any deformed Bromwich contour that winds around the negative real axis
R− = (−∞, 0) in the anticlockwise sense. Now substituting ζ = su yields
∫
1 1
= ζ −z e ζ dζ . (2.6)
Γ(z) 2πi C
The integrand ζ −z has a branch cut on R− but is analytic elsewhere so that (2.6) is
independent of the contour C. The formula (2.6) is very useful for deriving integral
representations of various special functions, e.g., Mittag-Leffler and Wright functions
(cf. (3.1) and (3.24) in Chapter 3). In practice, C can be deformed to facilitate the
1
numerical evaluation of the integral, which makes Γ(z) actually easier to compute
than Γ(z) itself [ST07].
For large argument, the following Stirling’s formula holds (for any > 0)
Further, for any two real numbers x and s, the following identity holds
often known as Wendel’s limit [Wen48], due to his elegant proof using only Hölder’s
inequality and the recursion identity (2.2); see Exercise 2.4 for the detail.
A closely related function, the Beta function B(α, β) for α, β > 0 is defined by
∫ 1
B(α, β) = (1 − s)α−1 s β−1 ds. (2.9)
0
2.2 Riemann-Liouville Fractional Integral 21
Proof The identities (2.11) and (2.12) are direct from the definition. Indeed,
Γ(α + 1) Γ(k + j + 1)
C(α, k + j)C(k + j, k) =
Γ(k + j + 1)Γ(α − k − j + 1) Γ(k + 1)Γ( j + 1)
Γ(α + 1) Γ(α − k + 1)
=
Γ(k + 1)Γ(α − k + 1) Γ( j + 1)Γ(α − k − j + 1)
= C(α, k)C(α − k, j),
showing (2.12). The formula (2.13) can be found at [PBM86, p. 616, 4.2.15.13].
and verify it by mathematical induction. Clearly, it holds for k = 1. Now assume that
the formula (2.14) holds for some k ≥ 1, i.e.,
∫ x
k 1
a I x f (x) = (x − s)k−1 f (s) ds.
(k − 1)! a
Then by definition and changing integration order, we deduce
∫ x ∫ x∫ s
k+1 k 1
a Ix f (x) = a Is f (s) ds = (s − t)k−1 f (t) dtds
a (k − 1)! a a
∫ x∫ x ∫ x
1 1
= (s − t)k−1 ds f (t) dt = (x − s)k f (s) ds.
(k − 1)! a t k! a
This shows the formula (2.14). Likewise, there holds
∫ b
k 1
x Ib f (x) = (s − x)k−1 f (s) ds. (2.15)
(k − 1)! x
One way to generalize these formulas to any α > 0 is to use the Gamma function
Γ(z): the factorial (k −1)! in the formula can be replaced by Γ(k). Then for any α > 0,
by replacing k with α, and (k − 1)! with Γ(α), we arrive at the following definition of
Riemann-Liouville fractional integrals. It is named after Bernhard Riemann’s work
in 1847 [Rie76] and Joseph Liouville’s work in 1832 [Lio32], the latter of whom
was the first to consider the possibility of fractional calculus [Lio32].
Definition 2.1 For any f ∈ L 1 (D), the left-sided Riemann-Liouville fractional inte-
gral of order α > 0 based at x = a, denoted by a Ixα f , is defined by
∫ x
1
(a Ixα f )(x) = (x − s)α−1 f (s) ds, (2.16)
Γ(α) a
and the right-sided Riemann-Liouville fractional integral of order α > 0 based at
x = b, denoted by x Ibα f , is defined by
∫ b
1
(x Ibα f )(x) = (s − x)α−1 f (s) ds. (2.17)
Γ(α) x
When α = k ∈ N, they recover the k-fold integrals a Ixk f and x Ibk f . Further, we
define for x → a+ by
(a Ixα f )(a+ ) := lim+ (a Ixα f )(x),
x→a
if the right-hand side exists, and likewise (x Ibα f )(b− ). In case α = 0, we adopt the
convention a Ix0 f (x) = f (x), in view of the following result [DN68, (1.3)].
Lemma 2.2 Let f ∈ L 1 (D). Then for almost all x ∈ D, there hold
α
lim a Ix f (x) = lim+ x Ibα f (x) = f (x).
α→0+ α→0
2.2 Riemann-Liouville Fractional Integral 23
∫h
Proof Let x ∈ D be a Lebesgue point of f , i.e., limh→0 h1 0 | f (x + s) − f (s)|ds = 0.
∫x
Then let Fx (t) = x−t f (s)ds, for any t ∈ [0, x − a]. Then it can be represented as
where ωx (t) → 0 as t → 0. Thus for each > 0, there exists δ = δ() > 0 such that
Then it follows from (2.18) and the fact 0 is a pole of Γ(z) that limα→0+ | a Ixα f (x) −
f (x)| ≤ . Since > 0 is arbitrary, it proves the first identity for x. This completes
the proof since the set of Lebesgue points of f ∈ L 1 (D) is dense in D.
The integral operators a Ixα and x Ibα inherit certain properties from integer-order
integrals. Consider power functions (x − a)γ with γ > −1 (so that (x − a)γ ∈ L 1 (D)).
It can be verified directly that for α > 0 and x > a,
α Γ(γ + 1)
a I x (x − a)γ = (x − a)γ+α, (2.19)
Γ(γ + α + 1)
and similarly for x < b
α Γ(γ + 1)
x Ib (b − x)γ = (b − x)γ+α .
Γ(γ + α + 1)
Indeed, (2.19) follows from changing variables s = a + t(x − a) and (2.10):
∫ x ∫
α γ 1 α−1 γ (x − a)γ+α 1
I
a x (x − a) = (x − s) (s − a) ds = (1 − t)α−1 tγ dt
Γ(α) a Γ(α) 0
B(α, γ + 1) γ+α Γ(γ + 1) γ+α
= (x − a) = (x − a) ,
Γ(α) Γ(γ + α + 1)
Clearly, for any integer α ∈ N, these formulas recover the integral counterparts.
Example 2.1 We compute the Riemann-Liouville fractional integral 0 Ixα f (x), α > 0,
of the exponential function f (x) = eλx , λ ∈ R. By (2.19), we have
24 2 Fractional Calculus
∞
(λx)k
∞ α k
0 Ix x
α
0 Ix f (x) = 0 Ixα = λk
k=0
Γ(k + 1) k=0 Γ(k + 1)
(2.20)
∞
x k+α ∞
(λx)k
k
= λ = xα = x α E1,α+1 (λx),
k=0
Γ(k + α + 1) k=0
Γ(k + α + 1)
zk
where the Mittag-Leffler function Eα,β (z) is defined by Eα,β (z) = ∞ k=0 Γ(kα+β) , for
z ∈ C, cf. (3.1) in Chapter 3. The interchange of summation and integral is legitimate
since Eα,β (z) is entire, cf. Proposition 3.1.
The following semigroup property holds for k-fold integrals: a Ixk a Ix f = a Ixk+ f ,
k, ∈ N0 , which holds also for a Ixα and x Ibα .
Theorem 2.1 For f ∈ L 1 (D), α, β ≥ 0, there holds
α β β α+β α β β α+β
a Ix a Ix f = a Ix a Ixα f = a Ix f, x Ib x Ib f = x Ib x Ibα f = x Ib f.
Proof The case α = 0 or β = 0 is trivial, since a Ix0 is the identity operator. It suffices
to consider α, β > 0. By changing integration order, we have
∫ x ∫ s
α β 1 α−1
(a Ix a Ix f )(x) = (x − s) (s − t)β−1 f (t) dtds
Γ(α)Γ(β) a a
∫ x ∫ x
1
= f (t) (x − s)α−1 (s − t)β−1 ds dt.
Γ(α)Γ(β) a t
By (2.10), we have
∫ x
β B(α, β) α+β
(a Ixα a Ix f )(x) = (x − s)α+β−1 f (s) ds = (a Ix f )(x).
Γ(α)Γ(β) a
The next result collects several mapping properties for a Ixα , and similar results
hold for x Ibα . (i) indicates that a Ixα and x Ibα are bounded in L p (D) spaces, 1 ≤ p ≤ ∞.
Thus for any f ∈ L 1 (D), a Ixα f and x Ibα f of order α > 0 exists almost everywhere
(a.e.). The second part of (ii) is commonly known as the Hardy-Littlewood inequality.
The spaces AC(D) and C k,γ (D) are defined in Section A.1.1 in the appendix.
Theorem 2.2 The following mapping properties hold for a Ixα , and α > 0.
(i) a Ixα is bounded on L p (D) for any 1 ≤ p ≤ ∞.
2.2 Riemann-Liouville Fractional Integral 25
(ii) For 1 ≤ p < α−1 , a Ixα is a bounded operator from L p (D) into L r (D) for
1 ≤ r < (1 − αp)−1 . If 1 < p < α−1 , then a Ixα is bounded from L p (D) to
p
L 1−α p (D).
(iii) For p > 1 and p−1 < α < 1 + p−1 or p = 1 and 1 ≤ α < 2, a Ixα is bounded from
1
L p (D) into C 0,α− p (D), and moreover, a Ixα f (0) = 0, for f ∈ L p (D).
(iv) a Ixα maps AC(D) into AC(D).
(v) If f ∈ L 1 (D) is nonnegative and nondecreasing then a Ixα f is nondecreasing.
α 1 (b − a)α
a Ix f L p (D) ≤ (x − a)α−1 L 1 (D) f L p (D) = f L p (D),
Γ(α) Γ(α + 1)
showing the assertion for a Ixα . The first part of (ii) follows similarly, and the proof
of the second part is technical and lengthy, and can be found in [HL28, Theorem 4].
To see (iii), consider first the case p > 1. Suppose first a + h ≤ x < x + h ≤ b. Then
By Hölder’s inequality, since (α−1)p > −(p−1) and ((p−1)−1 (α−1)p+1)p−1 (p−1) =
α − p−1 ,
∫ x+h 1 ∫ x+h p−1
p(α−1)
≤ c f L p (D) hα− p .
p p 1
|I1 | ≤ | f (s)| p ds (x + h − s) p−1 ds
x x
α− p1
This argument also yields |I2 | ≤ c f L p (D) h . Now for the term I3 , since
we have
∫ x−h
|I3 | ≤ h|1 − α| | f (s)|(x − s)α−2 ds
a
∫ x−h 1 ∫ x−h p−1
p p(α−2) p
p
≤ h|1 − α| | f (s)| ds (x − s) p−1 ds
a a
p−1
p−1
hα−
p 1
≤ |1 − α| p f L p (D) .
p + 1 − αp
26 2 Fractional Calculus
Thus the estimate also holds. This shows (iii). Next, if f ∈ AC(D), then f ∈ L 1 (D)
exists a.e. and f (x) − f (a) = (a Ix1 f )(x) for all x ∈ D. Then
α (x − a)α
a Ix f (x) = a Ixα a Ix1 f (x) + (a Ixα f (a))(x) = a Ix1 (a Ixα f )(x) + f (a) .
Γ(α + 1)
Both terms on the right-hand side belong to AC(D), showing (iv). Last,
∫
α (x − a)α 1
a I x f (x) = (1 − s)α−1 f ((x − a)s + a)ds,
Γ(α) 0
Proof For f , g ∈ L 2 (D), the proof is direct: by Theorem 2.2(i), for f ∈ L 2 (D),
∫b
α α
a I x f ∈ L (D), and hence a |g(x)||( a I x f )(x)| dx ≤ c f L 2 (D) g L 2 (D) . Now the
2
desired formula follows from Fubini’s Theorem. The general case follows from
Theorem 2.2(ii) instead.
The next result discusses the mapping to AC(D) [Web19a, Proposition 3.6]. It is
useful for characterizing fractional derivatives and solution concepts for fdes.
Lemma 2.4 Let f ∈ L 1 (D) and α ∈ (0, 1). Then a Ix1−α f ∈ AC(D) and a Ixα f (a) = 0
if and only if there exists g ∈ L 1 (D) such that f = a Ixα g.
2.3 Fractional Derivatives 27
Proof If there exists g ∈ L 1 (D) such that f = a Ixα g, then by Theorem 2.1,
1−α
a Ix f = a Ix1−α a Ixα g = a Ix1 g ∈ AC(D),
and since∫g ∈ L 1 (D), by the absolute continuity of Lebesgue integral, a Ix1−α f (a) =
x
limx→a+ a g(s)ds = 0. Conversely, suppose a Ix1−α f ∈ AC(D) and a Ix1−α f (a) = 0.
Let G(x) = a Ix1−α f so that G ∈ AC(D) and G(a) = 0. Then g := G exists for a.e.
x ∈ D with g ∈ L 1 (D), and G(x) = a Ix1 g. Furthermore, a Ixα G = a Ix1 f = a Ixα a Ix1 g =
1 α α 1 α
a I x a I x g. Since f , a I x g ∈ L (D), a I x f , a I x a I x g ∈ AC(D) and their derivatives
1 1
exist a.e. as L 1 (D) functions and are equal, that is, f = a Ixα g.
for some ξ ∈ (a, b). One fractional version is as follows [Die12, Theorem 2.1] (see
also [Die17]). The classical case is recovered by setting α = 1 and x = b.
Theorem 2.3 Let α > 0 and f ∈ C(D), and let g be Lebesgue integrable on D
and do not change its sign in D. Then for almost every x ∈ (a, b], there exists some
ξ ∈ (a, x) ⊂ D such that a Ixα ( f g)(x) = f (ξ)a Ixα g(x). If additionally, α ≥ 1 or
g ∈ C(D), then this result holds for every x ∈ (a, b].
∫x
Proof Note that a Ixα ( f g)(x) = Γ(α)
1
a
(x − s)α−1 f (s)g(s)ds. If α ≥ 1 and x ∈ (a, b],
α−1
then (x − s)α−1 is continuous in s. Hence, g̃(s) = (x−s) Γ(α) g(s) is integrable on [a, x]
and does not change sign on [a, x]. Thus by (2.21),
∫ x ∫ x
α
a I x ( f g)(x) = f (s)g̃(s)ds = f (ξ) g̃(s)ds = f (ξ)a Ixα g(x).
a a
If 0 < α < 1 and g is continuous, the same line of proof works. Finally, if 0 < α < 1
and g is only integrable, then one can still argue in a similar way, but the integrability
of g̃ holds only for almost all x ∈ D, cf. Theorem 2.2(i).
Now we discuss fractional derivatives, for which there are several different defini-
tions. We only discuss Riemann-Liouville and Djrbashian-Caputo fractional deriva-
tives, which represent two most popular choices in practice, and briefly mention
Grünwald-Letnikov fractional derivative. There are several popular textbooks with
extensive treatment on fractional derivatives [OS74, MR93, Pod99, KST06, Die10]
and also monographs [Dzh66, SKM93]. An encyclopedic treatment of Riemann-
Liouville fractional integral and derivatives is given in the monograph [SKM93],
28 2 Fractional Calculus
and the book [Die10] contains detailed discussions on the Djrbashian-Caputo frac-
tional derivative. Like before, for any fixed a, b ∈ R, a < b, which are assumed to be
finite unless otherwise stated, we denote by D = (a, b), and by D = [a, b]. Further,
for k ∈ N, f (k) denotes the kth order derivative of f .
if the limit on the right-hand side exists. The quantity RxDbα f (b− ) is defined similarly.
Due to the presence of a Ixn−α , RaDxα f is inherently nonlocal: the value of RaDxα f
at x > a depends on the values of f from a to x. The nonlocality dramatically
influences its analytical properties. We compute the derivatives RaDxα (x − a)γ , α > 0,
of the function (x − a)γ with γ > −1. The identities (2.19) and (2.2) yield
One can draw a number of interesting observations. First, for α N, RaDxα f of the
constant function f (x) ≡ 1 (i.e., γ = 0) is not identically zero, since for x > a:
R α (x − a)−α
aDx 1 = . (2.23)
Γ(1 − α)
2.3 Fractional Derivatives 29
In particular, for α ∈ (0, 1), (x − a)α−1 belongs to the kernel of the operator RaDxα and
plays the same role as a constant function for the first-order derivative. Generally, it
implies that if f , g ∈ L 1 (D) with a Ixn−α f , a Ixn−α g ∈ AC(D) and n − 1 < α ≤ n, then
n
R α
aDx f (x) = RaDxα g(x) ⇔ f (x) = g(x) + c j (x − a)α−j ,
j=1
in view of the identities (2.22) and (2.24), which gives directly the assertion.
∞
(λx)k ∞ RD α x k
x
R α λx
0Dx e = R0Dxα = λk 0 (2.25)
k=0
Γ(k + 1) k=0
Γ(k + 1)
∞
λ k x k−α ∞
(λx)k
= = x −α = x −α E1,1−α (λx).
k=0
Γ(k + 1 − α) k=0
Γ(k + 1 − α)
Example 2.3 For α > 0 and λ ∈ C, let f (x) = x α−1 Eα,α (λx α ). Then the Riemann-
Liouville fractional derivative R0Dxα f is given by
∞
(λx α )k ∞ RD α x (k+1)α−1
x
R α
0Dx f (x) = R0Dxα x α−1 = λk 0
k=0
Γ(kα + α) k=0
Γ(kα + α)
∞
λ k x kα−1
= = λx α−1 Eα,α (λx α ).
k=0
Γ(kα)
The interchange of the fractional derivative and summation is legitimate since Eα,β (z)
is an entire function, cf. Proposition 3.1 in Chapter 3. That is, x α−1 Eα,α (λx α ) is
invariant under R0Dxα , and is a candidate for an eigenfunction of the operator R0Dxα
(when equipped with suitable boundary conditions).
Remark 2.2 Note that in Examples 2.2 and 2.3, the obtained results depend on the
starting value a. For example, if we take a = −∞, then direct computation shows
R α λx
−∞ Dx e = λα eλx, λ > 0.
where ci, i = 0, . . . , 4 are arbitrary constants. Indeed, for (i), one first applies a Ixα to
β
both sides, and then a Ix . Thus two extra terms have to be included: (i) and (ii) include
two constants, whereas (iii) includes only one. Note that if we restrict f ∈ C(D),
then the singular term (x − a)α−1 (or (x − a)β−1 ) must disappear, indicating that a
composition rule might still be possible on suitable subspaces. Indeed, it does hold
γ
on the space a Ix (L p (D)), γ > 0 and p ∈ [1, ∞], defined by
γ p γ
a I x (L (D)) = { f ∈ L p (D) : f = a Ix ϕ for some ϕ ∈ L p (D)},
γ γ
and similarly the space x Ib (L p (D)). A function in a Ix (L p (D)) has the property that
its function value and a sufficient number of derivatives vanishing at x = a.
Lemma 2.5 For any α, β ≥ 0, there holds
R αR β α+β α+β
aDx aDx f = RaDx f, ∀ f ∈ a Ix (L 1 (D)).
α+β α+β
Proof Since f ∈ a Ix (L 1 (D)), f = a Ix ϕ for some ϕ ∈ L 1 (D). Now the desired
assertion follows directly from Theorem 2.6(i) below and Theorem 2.1.
2.3 Fractional Derivatives 31
k
( f g)(k) = C(k, i) f (i) g (k−i) .
i=0
Theorem 2.5 Let f and g be analytic on D. Then for any α > 0 and β ∈ R,
∞
R α
aDx ( f g) = C(α, k)(RaDxα−k f )g (k),
k=0
∞
R α α−β−k β+k
aDx ( f g) = C(α, k + β)(RaDx f )(RaDx g).
k=−∞
Proof Since f , g are analytic, the product h = f g is also analytic. Hence, there
exists a power series representation of h based at x,
∞
h(k) (x)
h(s) = (−1)k (x − s)k ,
k=0
k!
and then applying the operator RaDxα termwise (which is justified by the uniform
convergence of the corresponding series), we obtain for n − 1 < α ≤ n, n ∈ N,
∫ x
∞
R α 1 dn (−1)k (x − s)k
aDx h(x) = (x − s)n−α−1 h(k) (x)ds
Γ(n − α) dx n a k=0
Γ(k + 1)
∞ ∫ x
dn (−1)k h(k) (x)
= n Γ(n − α)Γ(k + 1)
(x − s)k+n−α−1 ds
k=0
dx a
∞
dn (−1)k (x − a)k+n−α h(k) (x)
= .
k=0
dx n Γ(n − α)Γ(k + 1)(k + n − α)
Next we simplify the constant on the right-hand side. By the identity (2.2),
1 C(k + n − α − 1, k)
= .
Γ(n − α)Γ(k + 1)(k + n − α) Γ(k + n − α + 1)
Now by the combinatorial identity (2.11), we further deduce
32 2 Fractional Calculus
(−1)k C(α − n, k)
= .
Γ(n − α)Γ(k + 1)(k + n − α) Γ(k + n − α + 1)
Consequently,
∞
dn (x − a)k+n−α h(k) (x)
R α
aDx h(x) = C(α − n, k) .
k=0
dx n Γ(n + k − α + 1)
where we have used the identity (2.13). Applying the classical Leibniz rule, inter-
changing of summation order and the identity (2.12) give
∞
∞
(x − a)k+j−α
R α
aDx ( f g) = g (k) (x) C(α, k + j)C(k + j, k) f (j) (x)
k=0 j=0
Γ(k + j + 1 − α)
∞
∞
(x − a)k+j−α
= C(α, k)g (k) (x) C(α − k, j) f (j) (x) .
k=0 j=0
Γ(k + j + 1 − α)
Now using the identity (2.26) completes the proof of the first assertion. The second
assertion can be proved in a similar but more tedious manner, and thus we refer to
[SKM93, Section 15] or [Osl70] for a complete proof.
The next result gives the fundamental theorem of calculus for RaDxα : it is the left
inverse of a Ixα on L 1 (D), but generally not a right inverse.
Theorem 2.6 Let α > 0, n − 1 < α ≤ n, n ∈ N. Then
(i) For any f ∈ L 1 (D), RaDxα (a Ixα f ) = f .
(ii) If a Ixn−α f ∈ AC n (D), then
2.3 Fractional Derivatives 33
n−1
(x − a)α−j−1
αR α R α−j−1
a I x aDx f = f − aDx f (a+ ) .
j=0
Γ(α − j)
Proof (i) If f ∈ L 1 (D), by Theorem 2.2(i), a Ixα f ∈ L 1 (D), and by Theorem 2.1,
n−α α
a Ix a Ix f = a Ixn f ∈ AC n (D).
n−1
(x − a) j
n−α
a Ix f = cj + a Ixn ϕ f , (2.27)
j=0
Γ( j + 1)
j−n+α
for some ϕ f ∈ L 1 (D), with c j = RaDx f (a+ ). Consequently,
αR α
a I x aDx f = a Ixα ϕ f . (2.28)
Now applying a Ixα to both sides of (2.27) and using Theorem 2.1, we obtain
n−1
(x − a)α+j
n
a Ix f = cj + a Ixα+n ϕ f .
j=0
Γ(α + j + 1)
n−1
(x − a)α+j−n
f = cj + a Ixα ϕ f .
j=0
Γ(α + j − n + 1)
This together with (2.28) yields assertion (ii). Alternatively, by definition, we have
∫ x
αR α 1 dn
a I x aDx f (x) = (x − s)α−1 n (a Isn−α f )(s)ds
Γ(α) a ds
∫ x
d 1 dn
= (x − s)α n (a Isn−α f )(s)ds .
dx Γ(α + 1) a ds
Then applying integration by parts n times to the term in bracket gives
∫ x n
1 dn−j n−α (x − a)α−j
(x − s)α−n−1 a Isn−α f (s)ds − a I f (s)| s=a .
Γ(α − n) 0 j=1
ds n−j s Γ(α + 1 − j)
Note that the expression makes sense due to the conditions on f (x).