SPRINGER BRIEFS IN MATHEMATICS

Adam Kubica
Katarzyna Ryszewska
Masahiro Yamamoto

Time-Fractional
Differential Equations
A Theoretical Introduction

123
SpringerBriefs in Mathematics

Series Editors
Nicola Bellomo, Torino, Italy
Michele Benzi, Pisa, Italy
Palle Jorgensen, Iowa, USA
Tatsien Li, Shanghai, China
Roderick Melnik, Waterloo, Canada
Otmar Scherzer, Linz, Austria
Benjamin Steinberg, New York, USA
Lothar Reichel, Kent, USA
Yuri Tschinkel, New York, USA
George Yin, Detroit, USA
Ping Zhang, Kalamazoo, USA
SpringerBriefs in Mathematics showcases expositions in all areas of mathematics
and applied mathematics. Manuscripts presenting new results or a single new result
in a classical field, new field, or an emerging topic, applications, or bridges between
new results and already published works, are encouraged. The series is intended for
mathematicians and applied mathematicians. All works are peer-reviewed to meet
the highest standards of scientific literature.
Titles from this series are indexed by Scopus, Web of Science, Mathematical
Reviews, and zbMATH.

More information about this series at http://www.springer.com/series/10030

Adam Kubica • Katarzyna Ryszewska •
Masahiro Yamamoto

Time-Fractional Differential
Equations
A Theoretical Introduction
Adam Kubica Katarzyna Ryszewska
Warsaw University of Technology Warsaw University of Technology
Warszawa, Poland Warszawa, Poland

Masahiro Yamamoto
Graduate School of Mathematical Sciences
The University of Tokyo
Tokyo, Japan

ISSN 2191-8198 ISSN 2191-8201 (electronic)

SpringerBriefs in Mathematics
ISBN 978-981-15-9065-8 ISBN 978-981-15-9066-5 (eBook)
https://doi.org/10.1007/978-981-15-9066-5

Mathematics Subject Classification: 35R11, 26A33

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020
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 Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface

Recently, fractional differential equations have attracted great attention and many
studies have been performed. However, there are not many works that cover the
theory of partial differential equations, so that unnecessarily lengthy arguments and
also arguments that lack rigor are sometimes presented, and it may be difficult to
gain unified views of the related fields.
This concise book provides rigorous treatments for time-fractional derivatives
in Sobolev spaces and solutions to initial boundary value problems for time-
fractional partial differential equations and establishes the foundation of the theory
for fractional differential equations. The results here should be fundamental also
for discussing other important topics such as nonlinear dynamical systems, optimal
control, and inverse problems for fractional differential equations. Although our
approach can work for more general fractional differential equations, focusing on a
differential equation with a single time-fractional derivative, we describe the theory.
The parts of the book have been presented as graduate courses at Sapienza
University of Rome and The University of Tokyo. The authors thank Professor
Bangti Jin (University College London) for valuable comments.
The first and second authors are supported partly by the National Science
Centre, Poland through 2017/26/M/ST1/00700 Grant, and the third author is partly
supported by Grant-in-Aid for Scientific Research (S) 15H05740 of Japan Society
for the Promotion of Science and by the National Natural Science Foundation of
China (nos. 11771270 and 91730303) and the “RUDN University Program 5–100.”

Warszawa, Poland Adam Kubica

Warszawa, Poland Katarzyna Ryszewska
Tokyo, Japan Masahiro Yamamoto
November 2020

v
Remarks for Readers

• In this book, we minimize the references. In Chap. 2, we use some basic

knowledge of functional analysis and operator theory. The readers can consult
suitable textbooks on functional analysis, and here we rely on Brezis [5], Kato
[13], Pazy [23], and Tanabe [27]. In particular, for the function space, we refer to
Adams [2]. Needless to say, there are other reputable source books.
• The readers who wish to use this book for quickly understanding a theoretical
essence of fractional differential equations are encouraged to skip the proofs of
Lemmata 2.2–2.4 and Theorem 2.1. They can skip also Sect. 2.6 of Chap. 2 and
Sects. 5.2 and 5.3 of Chap. 5.
• We use the following numbering: for example, Theorem 2.1 means the first
theorem in Chap. 2.

vii
Contents

1 Basics on Fractional Differentiation and Integration .. . . . . . . . . . . . . . . . . . . 1

2 Definition of Fractional Derivatives in Sobolev
Spaces and Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
2.1 Motivations .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
2.2 Preliminaries: Operational Structure of J α . . . . . . . .. . . . . . . . . . . . . . . . . . . . 11
2.3 Definition of Generalized ∂tα in Hα (0, T ) . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 21
2.4 Some Functions in Hα (0, T ) . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 26
2.5 Definition of ∂tα in H α (0, T ) . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 33
2.6 Adjoint of the Fractional Derivative in Hα (0, T ). .. . . . . . . . . . . . . . . . . . . . 38
2.7 Laplace Transform of ∂tα . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
3 Fractional Ordinary Differential Equations . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 47
3.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 47
3.2 Fundamental Inequalities: Coercivity .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 48
3.3 Well-Posedness for Single Linear Fractional Ordinary
Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 55
3.4 Alternative Formulation of Initial Value Problem .. . . . . . . . . . . . . . . . . . . . 59
3.5 Systems of Linear Fractional Ordinary Differential Equations . . . . . . . 62
3.6 Linear Fractional Ordinary Differential Equations
with Multi-Term Fractional Derivatives . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 69
4 Initial Boundary Value Problems for Time-Fractional Diffusion
Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
4.1 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
4.2 Some Results from Theorem 4.1 and Proposition 4.2 .. . . . . . . . . . . . . . . . 78
4.2.1 Interpolated Regularity of Solutions .. . . . . .. . . . . . . . . . . . . . . . . . . . 78
4.2.2 The Method by the Laplace Transform for Fractional
Partial Differential Equations . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 78
4.3 Proof of Theorem 4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 80
4.4 Proof of Theorem 4.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 88

ix
x Contents

4.5 Proofs of Propositions 4.1 and 4.2 . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 96

4.5.1 Proof of Proposition 4.1 . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 96
4.5.2 Proof of Proposition 4.2 . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98
4.6 Case Where the Coefficients of A(t) Are Independent of Time . . . . . . 101
5 Decay Rate as t → ∞.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
5.1 Main Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
5.2 Key Lemmata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
5.3 Completion of Proof of Theorem 5.1 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 115
6 Concluding Remarks on Future Works . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 121

A Proofs of Two Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 123

Notation . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 129
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 131
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 133
Chapter 1
Basics on Fractional Differentiation
and Integration

Derivatives of orders of natural numbers have been widely known and commonly
used since the origin of the calculus in the seventeenth century. On the other hand,
derivatives whose orders are not necessarily natural numbers seem not to be well-
known but have been considered since Leibniz (e.g., Ross [25]). Moreover, we can
refer to a historical paper [1] by N.H. Abel. He discussed the following problem:
Given a function f (t), find a function v(t) such that
 t
1
√ (t − s)−1/2 v(s)ds = f (t), 0 ≤ t ≤ T. (1.1)
π 0

This is an integral equation and appears to be related to the following mechanical

problem: we consider that a mass goes along a curve on the vertical surface only by
gravity starting from a point O to the end point P, and assume that the travel time
is given by a function f (t), where t is the distance in the vertical direction from O.
Then, determine a shape of such a curve realizing the traveling time f (t).
Abel estabilished that for given f ∈ C 1 [0, T ], the solution v to (1.1) exists
uniquely and is given by
 t
1 d
v(t) = √ (t − s)−1/2 f (s)ds, 0 ≤ t ≤ T, β > 0. (1.2)
π dt 0

Here and henceforth by (·) we denote the gamma function:

 ∞
(β) = e−t t β−1 dt, β > 0.
0

We note that (n + 1) = n! := n(n − 1) · · · 3 · 2 · 1 for n ∈ N.

We can regard (1.2) as the half-order derivative of f as follows. We note
t
0 (t − s) v(s)ds is a function which is obtained by (n + 1)-times
1 n
that (n+1)

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 1
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5_1
2 1 Basics on Fractional Differentiation and Integration

t
integrating v(t). Thus we can interpret √1 d
π dt 0 (t − s)−1/2 f (s)ds as a function
which is differentiated 12 = − 12 + 1-times. That is, setting n = − 12 formally in
t √
(n+1) 0 (t − s) v(s)ds and noting that π = (1/2), we see that the right-hand
1 n

side of (1.2) means the half-times derivative of f . Thus we can interpret that (1.2)
means that the half-order derivative of f is v when f is the half-times integration of
v, and establishes the inversion of the half-order integral in terms of the half-order
derivative.
As for other applications of fractional derivatives, we refer to Gorenflo and
Vessella [8].
Although it is later seen that the rigorous treatments are necessary, in this chapter
we intuitively discuss fractional calculus as an introduction.
Throughout the book we assume

0 < α < 1.

In this chapter, as functions we mainly consider u ∈ L1 (0, T ) satisfying du

dt ∈
L1 (0, T ), if not specified. We set
 t
1 d
Dtα u(t) := (t − s)−α u(s)ds (1.3)
(1 − α) dt 0

and
 t
1 du
dtα u(t) := (t − s)−α (s)ds. (1.4)
(1 − α) 0 ds

We call Dtα and dtα the Riemann–Liouville derivative and the Caputo derivative
respectively. The notations of the derivatives are different from e.g., Kilbas et
al. [15], Podlubny [24], and in Chap. 2, the fractional derivative ∂tα in fractional
Sobolev spaces is defined and plays an essential role rather than the classical
Riemann–Liouville derivative and the Caputo derivative.
Throughout this book, we interpret equalities in variables t, x, etc., as that they
hold for almost all t, x in domains under consideration, and we do not specify if
there is no fear of confusion.
In this book, we mainly consider the case 0 < α < 1, but for general α > 0, ∈ N,
choosing m ∈ N satisfying m − 1 < α < m, for u ∈ C m [0, T ], we can define dtα
and Dtα by
 t
1 d mu
dtα u(t) = (t − s)m−α−1 (s)ds
(m − α) 0 ds m

and
 t
1 dm
Dtα u(t) = (t − s)m−α−1 u(s)ds.
(m − α) dt m 0
1 Basics on Fractional Differentiation and Integration 3

Moreover, for α > 0, we define the Riemann–Liouville fractional integral

operator by
 t
1
J α f (t) = (t − s)α−1 f (s)ds, (1.5)
(α) 0

which we can regard as an operation of α-times integration.

β
Example 1.1 Let 0 < α < 1 and β > 0. Then t β , dtdt = βt β−1 ∈ L1 (0, T ) and we
have
(β + 1) β−α
dtα t β = Dtα t β = t . (1.6)
(1 − α + β)

In fact, this follows from the classical formula:

 b (p)(q)
(t − a)p−1 (b − t)q−1 dt = (b − a)p+q−1 , p, q > 0, a < b.
a (p + q)
(1.7)
We can prove also:
Lemma 1.1 Let u ∈ C 1 [0, T ]. Then for 0 < α < 1, we have

u(0) −α
dtα u(t) + t = Dtα u(t), 0 ≤ t ≤ T. (1.8)
(1 − α)

Proof Since u ∈ C 1 [0, T ], integration by parts yields

  s=0 
t
−α u(s)(t − s)1−α 1 t du
(t − s) u(s)ds = + (t − s)1−α (s)ds
0 1−α s=t 1−α 0 ds
 t
u(0) 1−α 1 du
= t + (t − s)1−α (s)ds.
1−α 1−α 0 ds

By differentiating both sides, the proof is complete by 1 − α > 0. 

The first-order derivative of a constant function vanishes identically. Also the
Caputo derivative with order α of a constant function is zero: ∂tα 1 = 0. However,
the Riemann–Liouville derivative is not zero. More precisely,

t −α
Dtα 1 = .
(1 − α)

The fundamental theorem of calculus reads

 t
du
(s)ds = u(t) − u(0), 0 < t < T,
0 ds
4 1 Basics on Fractional Differentiation and Integration

which indicates that the integration is a converse operation to the differentiation. We

can have a similar inversion for dtα and Dtα , whose special case is (1.2).
More precisely, as an important formula, we can prove

J α dtα u(t) = u(t), 0<t<T (1.9)

for u ∈ C[0, T ] ∩ C 1 (0, T ] satisfying

  
T  du 
 (t) dt < ∞ (1.10)
 dt 
0

and

u(0) = 0. (1.11)

Moreover,

J α Dtα u(t) = u(t), 0<t<T (1.12)

for u ∈ C[0, T ] ∩ C 1 (0, T ] satisfying (1.10).

Proof of (1.9) First we obtain
  s 
1 −α du
J α dtα u(t) =J α
(s − ξ ) (ξ )dξ
(1 − α) 0 dξ
 t  s 
1 du
= (t − s)α−1 (s − ξ )−α
(ξ )dξ ds
(α)(1 − α) 0 0 dξ
 t  t 
1 du
= (ξ ) (t − s)α−1 (s − ξ )−α ds dξ.
(α)(1 − α) 0 dξ ξ

t s
For the last equality, we used the exchange of orders of integrals: 0 0 · · · dξ ds =
t t  t du
0 ξ · · · ds dξ . By (1.7) we conclude that the right-hand side is 0 dξ (ξ )dξ =
u(t) − u(0), which completes the proof of (1.9). The proof of (1.12) is similar and
so is omitted. 
For the Caputo derivative, the usual formulae for derivatives of orders of natural
numbers do not hold. For example, for u, v ∈ C 1 [0, T ], we have

dtα (u + v)(t) = dtα u(t) + dtα v(t),

dtα (cu)(t) = cdtα u(t), where c is a constant,

1 Basics on Fractional Differentiation and Integration 5

but

dtα (uv)(t) = (dtα u(t))v(t) + u(t)(dtα v)(t) (1.13)

and
β α+β
(dtα (dt u))(t) = dt u(t). (1.14)

The violation of the Leibniz rules (see (1.13)) implies the lack of integration by
parts, which causes many difficulties in considering fractional differential equations.
As for (1.14), we consider u(t) = t α and β = α ∈ (0, 1). Then (1.7) yields
dt t α = (1+α)
2α −α , but d α (d α t α ) = d α ((α + 1)) = 0, which means that d 2α t α =
(1−α) t t t t t
α α α
dt (dt t ).
As for the successive derivatives, we can directly prove the following. Let 0 <
α, β < 1 and α + β < 1. Then
β α+β
(dtα (dt u))(t) = dt u(t), 0≤t ≤T
β
for u ∈ C 1 [0, T ] satisfying u(0) = (dt u)(0) = 0, and
β α+β
(Dtα (Dt u))(t) = (Dt u)(t)

for u ∈ C 1 [0, T ].
du
It is important that the derivative characterizes the behavior, that is, increase
dt
or decrease, of u(t). We can accept that such a property holds in a limited sense
for the fractional derivatives, because the fractional derivative at t involves the
integral over (0, t). On the other hand, Luchko [20] proves the following extremum
principle.
Lemma 1.2 Let u ∈ C[0, T ] ∩ C 1 (0, T ] satisfy
  
T  du 
 (t) dt < ∞. (1.15)
 dt 
0

We assume that u(t) attains the minimum at some t1 satisfying 0 < t1 ≤ T : u(t1 ) =
min0≤t ≤T u(t). Then

dtα u(t1 ) ≤ 0.

Proof We set y(s) = u(s) − u(t1 ), 0 ≤ s ≤ T . Then y(s) ≥ 0, 0 ≤ s ≤ T and

dtα u(s) = dtα y(s), 0 ≤ s ≤ T . For arbitrary ε > 0, we can choose small δ > 0 such
that δ < t1 and, since y(s) satisfies (1.15),
  δ   δ 
 1 α dy
 1  dy 
 (t1 − s) 
(s)ds  ≤  (s) ds < ε.
 (1 − α) ds (1 − α)(t1 − δ)−α 0  ds 
0
6 1 Basics on Fractional Differentiation and Integration

In defining dtα y(t1 ), we divide the integration interval into (0, δ) and (δ, t1 ), so that
for already chosen δ, we have
 δ  t1
1 dy 1 dy
dtα y(t1 ) = (t1 − s)−α (s)ds + (t1 − s)−α (s)ds =: I1 + I2 .
(1 − α) 0 ds (1 − α) δ ds
(1.16)

Then |I1 | < ε. Moreover, for δ > 0, since y ∈ C 1 (0, T ] and y(t1 ) = 0, we can
choose a constant Cδ > 0 such that |y(s)| ≤ Cδ |t1 − s|, δ ≤ s ≤ T . Therefore
integrating by parts, we have

(t1 − δ)−α y(δ) α t1
I2 = − − (t1 − s)−α−1 y(s)ds.
(1 − α) (1 − α) δ

By t1 − δ > 0 and y(s) ≥ 0, 0 ≤ s ≤ T , we see that I2 ≤ 0. Hence by (1.16), we

obtain ∂tα y(t1 ) ≤ ε. Here since ε > 0 is arbitrary, we let ε → 0 and the proof of
Lemma 1.2 is complete. 
Lemma 1.2 does not hold for α > 1. It follows from Lemma 1.2 that if dtα u(t) >
0, 0 ≤ t ≤ T , then u(t) ≥ u(0) for all t ∈ [0, T ].
In fact, assume that there exists t0 ∈ [0, T ] such that u(t0 ) = min0≤t ≤T u(t) <
u(0). Then t0 = 0. That is, 0 < t0 ≤ T . Lemma 1.2 yields that dtα u(t0 ) ≤ 0, which
contradicts the assumption dtα u(t) > 0. Thus the proof is completed.
This is much weaker than the usual property for du dt , and we do not know whether
u(t0 ) ≥ u(t1 ) if dtα u(t) ≤ 0, t0 < t < t1 with 0 < t0 < t1 < T .
Throughout this book, we set
 
W 1,1 (0, T ) := u ∈ L1 (0, T ); du dt ∈ L (0, T ) ,
1
1,1 (0, T ) := {u ∈ W 1,1 (0, T ); u(0) = 0} (1.17)
0W

with the norm

 
 du 
u W 1,1 (0,T ) = u L1 (0,T ) + 
 dt  1 .
L (0,T )

We note that u ∈ W 1,1 (0, T ) if and only if there exists an absolutely continuous
function 
u on [0, T ] such that u(t) = u(t) for almost all t ∈ (0, T ). Here and
henceforth we identify 
u with u for u ∈ W 1,1 (0, T ).
We close this chapter with the following lemma which generalizes Lemma 1.1 to
u ∈ W 1,1 (0, T ),
Lemma 1.3
(i) For u ∈ W 1,1 (0, T ), we have J 1−α u ∈ W 1,1 (0, T ) and

d 1−α du u(0) −α
J u(t) = J 1−α (t) + t , 0 < t < T.
dt dt (1 − α)
1 Basics on Fractional Differentiation and Integration 7

(ii) Let un , u ∈ W 1,1 (0, T ) and un −→ u in W 1,1 (0, T ). Then dtα un −→ dtα u in
L1 (0, T ).
(iii) dtα u = Dtα u for u ∈ W 1,1 (0, T ) satisfying u(0) = 0.
(iv) Let α, β > 0. Then

J α (J β u) = J α+β u, u ∈ L2 (0, T ).

By the Sobolev embedding (e.g., formula (1.20) on p. 193 in Kato [13],

Theorem 5.4 in Adams [2]), we see that W 1,1 (0, T ) ⊂ C[0, T ] and so u(0) makes
 t u ∈ W (0, T ). We can interpret Lemma 1.3 (i) so that the exchange of
sense for 1,1

dt and 0 · · · ds is possible in
d

 t
(t − s)−α u(s)ds.
0

Proof (i) and (iii) By u ∈ W 1,1 (0, T ), we have

 s du
u(s) = (ξ )dξ + u(0), 0 < s < T.
0 dξ

Hence
 t  s 
1 du
(J 1−α u)(t) = (t − s)−α (ξ )dξ + u(0) ds
(1 − α) 0 0 dξ
 t  s   t 
1 du 1
= (t − s)−α (ξ )dξ ds + (t − s)−α ds u(0)
(1 − α) 0 0 dξ (1 − α) 0
 t  t 
1 du 1 t 1−α
= (t − s)−α ds (ξ )dξ + u(0).
(1 − α) 0 ξ dξ (1 − α) 1 − α

Here we have exchanged the orders of integrals. Therefore

 t
1 du
(J 1−α u)(t) = (t − ξ )1−α (ξ )dξ
(1 − α)(1 − α) 0 dξ
1
+ t 1−α u(0), 0 < t < T.
(1 − α)(1 − α)

Taking the differentiation in t, noting that 1 − α > 0, we see that J 1−α u ∈

W 1,1 (0, T ) and

d 1−α du u(0) −α
J u(t) = J 1−α (t) + t . (1.18)
dt dt (1 − α)
8 1 Basics on Fractional Differentiation and Integration

Here we use Lemma A.1 in the Appendix, which is called the Young inequality for
the convolution:
 
1−α du 1 −α du
J (t) = t ∗ (t)
dt (1 − α) dt

and
     
 1−α du   1   du 
J  
≤ −α 
t    < ∞.
 
dt L1 (0,T ) (1 − α)  
L1 (0,T ) dt L1 (0,T )

Moreover, (1.18) directly implies conclusion (iii). 

Proof of (ii) We have
 t  
1 dun du
dtα un (t) − dtα u(t) = (t − s)−α (s) − (s) ds
(1 − α) 0 ds ds
 
1 d
= t −α ∗ (un − u) (t).
(1 − α) dt

Again the Young inequality Lemma A.1 for the convolution yields
 
1 d 
dtα un − dtα u L1 (0,T ) ≤ t −α  (un − u)
L1 (0,T )   1
(1 − α) dt L (0,T )

1 T 1−α
≤ un − u W 1,1 (0,T ) .
(1 − α) 1 − α

Thus the proof of Lemma 1.3 (ii) is complete. 

Proof of (iv) Exchanging the orders of the integrals and applying (1.7), we verify
 t  s 
1 1
J α J β u(t) = (t − s)α−1 (s − ξ )β−1 u(ξ )dξ ds
(α) (β) 0 0
 t  t 
1
= (t − s)α−1 (s − ξ )β−1 ds u(ξ )dξ
(α)(β) 0 ξ
 t
1 (α)(β)
= (t − s)α+β−1 u(ξ )dξ = J α+β u(t).
(α)(β) 0 (α + β)


Chapter 2
Definition of Fractional Derivatives
in Sobolev Spaces and Properties

2.1 Motivations

We consider a very simple equation

dtα u(t) = f (t), 0 < t < T, u(0) = a (2.1)

and find u(t) satisfying (2.1) with given f ∈ L2 (0, T ) and a ∈ R. Needless to say,
for α = 1, we have
 t
u(t) = f (s)ds + a, 0 < t < T.
0

If α = 1 and u satisfies (2.1) with given f ∈ L1 (0, T ), then u ∈ C[0, T ], so that

u(0) can be calculated by substituting t = 0 into u(t). In other words, for α = 1,
for any f ∈ L2 (0, T ) and a ∈ R, there exists a unique u ∈ H 1 (0, T ) satisfying
(2.1). The situation is different from the case of 0 < α < 1. In most of the existing
works, for f ∈ L2 (0, T ) or f ∈ L1 (0, T ), detailed regularity of the solution to
dtα u = f is not clarified and so the initial condition can only be justified if the
solutions are proved to possess a certain regularity at t = 0. The class f ∈ L2 (0, T )
is not sufficient for such regularity. For example, Theorem 3.3 (p. 126) in Podlubny
[24] assumes that f ∈ C[0, T ].
For fractional differential equations, the Laplace transform:
 ∞
(Lu)(p) := e−pt u(t)dt, p > p0 : some constant
0

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 9
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5_2
10 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

is convenient, but the justification of the initial value is indispensable in order to use
the Laplace transform. Because by some repetition of calculations, the formula

L(dtα u)(p) = pα (Lu)(p) − pα−1 u(0)

is known (e.g., p. 106 in [24]) and we have to justify the sense of u(0), which
requires a certain smoothness of u at t = 0. Such regularity at t = 0 is not well
established for f ∈ L2 (0, T ). Moreover, for such f, we have to justify the solution
formula for (2.1) (e.g., Kilbas et al. [15, p. 141]).
We illustrate with one simple example. Let 0 < α < 12 and
1
f (t) = t δ− 2 , 0 < t < T,

where δ > 0 is a constant. We consider

1
dtα u(t) = t δ− 2 , u(0) = a. (2.2)

Since δ − 12 > −1, we can formally apply the solution formula (e.g., [15, p. 141])
and obtain
 t
1 1 1
u(t) = a + (t − s)α−1 s δ− 2 ds = a + C0 t α+δ− 2 , (2.3)
(α) 0
where we set

 δ+ 1
2
C0 = .
 α+δ+ 1
2

Moreover, u(t) given by (2.3) cannot satisfy (2.2) if 0 < α < 12 and δ > 0 is
small such that α + δ − 12 < 0. Indeed, limt ↓0 u(t) = ∞, and so the initial condition
1
does not follow common sense. Furthermore we formally calculate dtα t α+δ− 2 :
 
1 t d α+δ− 1 α + δ − 12 t
(t − s)−α (t − s)−α s α+δ− 2 ds.
1 3
dtα t α+δ− 2 = (s 2 )ds =
(1 − α) 0 ds (1 − α) 0

However, since α +δ − 32 < −1, the integral does not exist. This means that formula
(2.3) does not hold for f ∈ L2 (0, T ) in general, although (2.3) is convergent for
e.g., f ∈ L∞ (0, T ) as a solution formula to (2.1). We can verify that (2.3) gives
dt ∈ L (0, T ), because we can calculate dt u pointwise. In
a solution to (2.1) if du 2 α

terms of an extended Caputo derivative, we will further discuss this later (see, e.g.,
(2.36)).
In some references (e.g., Podlubny [24, pp. 78–79]), it is emphasized that the
Caputo derivative dtα can admit the initial value problem (2.1), unlike the Riemann–
1
Liouville derivative Dtα . However, this simple example f (t) = t δ− 2 shows that the
2.2 Preliminaries: Operational Structure of J α 11

pointwise Caputo derivative dtα cannot make the initial value problem (2.1) well-
posed for general f ∈ L2 (0, T ).
Moreover, for guaranteeing the uniqueness of u satisfying dtα u = f ∈ L2 (0, T ),
we certainly need some extra conditions. In fact, since dtα 1 = 0 by the definition, if
u satisfies dtα u = f in (0, T ), then u + C1 satisfies the same equation with arbitrary
constant C1 . These discussions suggest the necessity for reconsidering the pointwise
Caputo derivative dtα and redefining the Caputo derivative within the framework of
L2 (0, T ).
The function space L2 (0, T ) is reasonable and convenient as data space. Hence
it is natural to formulate the initial value problem and define dtα for f ∈ L2 (0, T ) in
order to establish a more unified theory for fractional differential equations. Thus we
construct the theory where the fractional derivatives should be included in L2 (0, T ).
This is our main motivation in this book, and we construct a seemingly different
fractional derivative ∂tα , but we will prove that it is essentially same as the closure
operator of the Caputo derivative in 0 C 1 [0, T ] (see Sect. 2.3).
First in Sects. 2.2 and 2.3, we define the generalized fractional derivative ∂tα in
Sobolev spaces Hα (0, T ). Then in Chaps. 3 and 4, in terms of such ∂tα , we formulate
an initial value problem and prove the well-posedness.
Our formulation is similar to Zacher [30] (see also Kubica and Yamamoto [17]),
but more essentially relies on the property of the generalized fractional derivative
∂tα defined later in some Sobolev space. These properties are feasible for the
applications such as the clarification of the Sobolev regularity of solutions to initial
boundary value problems.

2.2 Preliminaries: Operational Structure of J α

By the direct observation of the definition (1.4) of the pointwise Caputo derivative,
we need the first derivative du α
ds (s) in order to define the derivative dt u of order
α < 1. In order to define such an adequate fractional derivative, which is denoted
by ∂tα , we should fulfill the following:
1. ∂tα should be well-defined in a subspace of the Sobolev space of order α;
2. the norm equivalence between ∂tα u L2 (0,T ) and some conventional norm of u
such as the norm in a Sobolev space.
For them,
• We will interpret J α as the fractional power of the operator defined by
 t
J u(t) = u(s)ds
0

with the domain D(J ) = L2 (0, T ).

• We define ∂tα as the inverse to J α .
These issues are done respectively in this section and Sect. 2.3.
12 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

The arguments in this section and a part of Sect. 2.3 are based on Gorenflo and
Yamamoto [9], Gorenflo et al. [11].
By L2 (0, T ) and H α (0, T ) we mean the usual L2 -space and the fractional
Sobolev space on the interval (0, T ) (see e.g., [2, Chapter VII]), respectively, and
we define the norm in H α (0, T ) by

    12
2
T T |u(t) − u(s)|2
u H α (0,T ) := u L2 (0,T )
+ dtds .
0 0 |t − s|1+2α

The L2 -norm and the scalar product in L2 are denoted by · L2 (0,T ) and (·, ·) =
(·, ·)L2 (0,T ) , respectively. By ∼ we denote a norm equivalence. Since J α is injective
in L2 (0, T ), by J −α we denote the algebraic inverse to J α .
We set
1
0H
α
(0, T ) = {u ∈ H α (0, T ); u(0) = 0} if < α ≤ 1.
2
We further define the Banach spaces
⎧ 1
⎪ α
⎨ 0H (0, T1 ), 2 < α < 1, 
 T |v(t )|2
Hα (0, T ) := v ∈ H (0, T ); 0
2 dt < ∞ , α = 12 ,
⎪
⎩ α
t
1
H (0, T ), 0 < α < 2

with the following norm:

⎧
⎪
⎨ v H α (0,T ) , 0 < α < 1, α = 12 ,
 1
v Hα (0,T ) = T |v(t )|2
2 (2.4)
⎪
⎩ v 2 + , α = 12 .
1 0 t dt
H 2 (0,T )

For later arguments, we need some convenient and not too narrow subspace of
Hα (0, T ). In particular, for H 1 (0, T ), such a subspace is more delicate.
2
We recall

0W
1,1
(0, T ) = {u ∈ W 1,1 (0, T ); u(0) = 0}

and introduce the following space:

Wα (0, T ) := u ∈ W 1,1 (0, T ); there exists a constant Cu > 0 such that

  
 du 
 (t) ≤ Cu t α−1 almost all t, u(0) = 0 . (2.5)
 dt 

Here Cu > 0 depends on a choice of u. We note that t β ∈ Wα (0, T ) for any β ≥ α.

2.2 Preliminaries: Operational Structure of J α 13

Henceforth by C > 0, C1 > 0, etc., we denote generic constants which are

independent of functions under consideration but dependent on α, T , while Cu
means that it depends on a function or a quantity u under consideration.
Then we can prove:
Lemma 2.1 Let 0 < α < 1. Then

Wα (0, T ) ⊂ Hα (0, T ).

Proof First we prove

 T  T
(s γ1 + t γ1 )|t − s|γ2 dsdt < ∞ if γ1 , γ2 > −1. (2.6)
0 0


Proof of (2.6) The conclusion (2.6) is trivial for γ1 ≥ or γ2 ≥ 0. Therefore we can
assume that −1 < γ1 , γ2 < 0. We note
 T  T  T  t 
(s γ1 + t γ1 )|t − s|γ2 dsdt = 2 (s γ1 + t γ1 )|t − s|γ2 ds dt. (2.7)
0 0 0 0

Indeed,
 T  T  T  t  T 
(s γ1 + t γ1 )|t − s|γ2 dsdt = + (s γ1 + t γ1 )|t − s|γ2 ds dt.
0 0 0 0 t

On the other hand,

 T  T   T  s 
(s γ1
+ t )|t − s| ds dt =
γ1 γ2
(s γ1
+ t )|t − s| dt ds
γ1 γ2
0 t 0 0
 T  t 
= (t γ1 + s γ1 )|s − t|γ2 ds dt.
0 0

In the last equality, we exchange s and t. Then

 T  T   T  t 
(s γ1
+ t )|t − s| ds dt = 2
γ1 γ2
(s γ1
+ t )|t − s| ds dt
γ1 γ2
0 0 0 0
 T  t   T  t 
=2 s γ1 (t − s)γ2 ds dt + 2 t γ1 (t − s)γ2 ds dt
0 0 0 0
 
(γ1 + 1)(γ2 + 1) T T t γ1 +γ2 +1
=2 t γ1 +γ2 +1 dt + 2 dt < ∞
(γ1 + γ2 + 2) 0 0 1 + γ2

by γ1 > −1 and γ2 > −1. Thus the proof of (2.6) is completed. 

14 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

t
Now we complete the proof of Lemma 2.1. Indeed, u(t) = du
0 ds (s) for u ∈
Wα (0, T ). Hence
   t   
 t du    du  
|u(t) − u(s)| =  (ξ )dξ  ≤   (ξ ) dξ 
  
s dξ s dξ

≤Cu min |ξ | α−1

(t − s) ≤ Cu (t α−1 + s α−1 )|t − s|.
ξ ∈(s,t )

1
Let 2 < α < 1. Then

|u(t) − u(s)|2
≤ C(t 2α−2 + s 2α−2 )|t − s|1−2α .
|t − s|1+2α

Since 2α − 2 > −1 by 1
2 < α < 1 and 1 − 2α > −1 by 0 < α < 1, the inequality
(2.6) implies
 
T T |u(t) − u(s)|2
dsdt < ∞.
0 0 |t − s|1+2α

Next let 0 < α ≤ 12 . Similarly to (2.7), we can verify

    
T T |u(t) − u(s)|2 T t |u(t) − u(s)|2
dsdt = 2 ds dt.
0 0 |t − s|1+2α 0 0 |t − s|1+2α

Since
 
 du 
 (ξ ) ≤ Cu ξ α−1 ≤ Cu s α−1
 dξ 

for ξ ∈ (s, t), we choose δ > 0 such that 0 < δ < α

1−α ≤ 1. Then, by the mean
value theorem we can choose η ∈ (s, t) such that

|u(t) − u(s)|2 = |u(t) − u(s)|1−δ |u(t) − u(s)|1+δ

  1+δ
1−δ  du 

≤ 2 u L∞ (0,T )  dt (η) |t − s| ≤ C(ξ α−1 (t − s))1+δ

≤Cs (α−1)(1+δ)(t − s)1+δ

for 0 < s < t < T . Consequently, since α + αδ − δ > 0 and 1 + δ − 2α > 0 by

0 < α ≤ 12 and 0 < δ < 1−α
α
, the inequality (2.6) yields
    
T t |u(t) − u(s)|2 T t
dsdt ≤ C s (α+αδ−δ)−1 (t − s)(1+δ−2α)−1 ds dt < ∞.
0 0 |t − s|1+2α 0 0
2.2 Preliminaries: Operational Structure of J α 15

Thus we proved u ∈ H α (0, T ). Therefore u ∈ Hα (0, T ) = H α (0, T ) for 0 <

α < 12 . Next for 12 < α < 1, by u(0) = 0, we readily see that u ∈ Hα (0, T ) =
0 H (0, T ). Therefore Wα (0, T ) ⊂ Hα (0, T ) if 0 < α < 1 and α = 2 . Moreover,
α 1

for α = 12 , we have
   t
 t du 
|u(t)| =  (s)ds  ≤
1 1
Cu s − 2 ds = 2Cu t 2 ,
0 ds 0

T
and so 0 |u(tt )| dt < ∞, which implies that H 1 (0, T ). Thus the proof of
2

2
Lemma 2.1 is completed. 
In fact, the space Wα (0, T ) is a convenient subspace of Hα (0, T ).
Remark 2.1 For H 1 (0, T ), Lions and Magenes [19] use a different notation
2
1
2
0 H0(0, T ) (Remark 11.5 (p. 68) in [19] vol. I). However, we use H 1 (0, T ) as well
2
as Hα (0, T ), 0 < α < 1 throughout this book.
Henceforth we set
1
0C [0, T ] = {ϕ ∈ C 1 [0, T ]; ϕ(0) = 0}

X denotes the closure of Z by the norm of X.

and Z
Lemma 2.2
Hα (0,T )
1 [0, T ] = Hα (0, T ).
0C

Proof For 12 < α ≤ 1, we see that Hα (0, T ) = 0 H α (0, T ) := {u ∈

H α (0, T ); u(0) = 0}, so that the mollifier (e.g., Adams [2]) yields the conclusion.
Let 0 ≤ α ≤ 12 . By Lions and Magenes [19], we have

Hα (0, T ) = [0 H 1 (0, T ), L2 (0, T )]1−α

which is the interpolation space. Applying Proposition 6.1 (p. 28) in [19], for 12 <
γ < 1, we see that [0 H 1 (0, T ), L2 (0, T )]1−γ is dense in [0 H 1 (0, T ), L2 (0, T )]1−α
by 1 − γ ≤ 1 − α. Therefore

Hα (0,T )
Hγ (0, T ) = Hα (0, T )

1
for 2 < γ < 1. As is already proved, we see that

H (0,T )
1 [0, T ] γ = Hγ (0, T ).
0C
16 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

Both density properties yield

Hα (0,T )
1 [0, T ] = Hα (0, T )
0C

for 0 ≤ α ≤ 12 . Thus the proof of the lemma is completed. 

For 0 < α < 12 , the proof of the lemma is more direct by Theorem 11.1 (p. 55)
in [19].
Lemma 2.2 is useful, because, thanks to the lemma, in order to prove estimates
in Hα (0, T ), it usually suffices to prove them for 0 C 1 [0, T ].
Remark 2.2 By Theorem 11.1 (p. 55) in [19], we see that

H α (0,T ) 1
1 [0, T ] = H α (0, T ), 0≤α≤ (2.8)
0C .
2

We should distinguish Hα (0, T ) from H α (0, T ). For 1

2 < α ≤ 1, the mollifier yields
that 0 C 1 [0, T ] is dense in Hα (0, T ). To sum up,

1 [0, T ]
H α (0,T ) H α (0, T ), 0 < α ≤ 12 ,
0C = (2.9)
Hα (0, T ), 12 < α < 1.

By Theorem 2.1 in Gorenflo et al. [11], we know the following theorem.

Theorem 2.1 Let 0 < α < 1.
(i) J α : L2 (0, T ) −→ Hα (0, T ) is injective and surjective.
(ii) There exists a constant C > 0 such that

C −1 J α u Hα (0,T ) ≤ u L2 (0,T ) ≤ C J αu Hα (0,T ) (2.10)

for all u ∈ L2 (0, T ).

Moreover, we can prove

J α L2 (0, T ) ⊂ H1 (0, T ), α ≥ 1.

However, in this book we limit the range of α to 0 < α ≤ 1 and we omit further
characterization of J α L2 (0, T ).
In terms of Theorem 2.1, (J α )−1 exists and J −α = (J α )−1 . Then:
Theorem 2.2 There exists a constant C > 0 such that

C −1 J −α v L2 (0,T ) ≤ v Hα (0,T ) ≤ C J −α v L2 (0,T ) (2.11)

for all v ∈ Hα (0, T ).

2.2 Preliminaries: Operational Structure of J α 17

By Theorem 2.1, we see:

Corollary 2.1 Let 0 < α < 1. Then

J −α J α u = u, u ∈ L2 (0, T )

and

J α J −α u = u, u ∈ Hα (0, T ).

The first equality is directly seen by the definition, while the second equality
is verified as follows. For u ∈ Hα (0, T ), Theorem 2.1 (i) yields the existence of
w ∈ L2 (0, T ) satisfying u = J α w. Therefore J α J −α u = J α J −α (J α w) = J α w.
Hence J α J −α u = u for u ∈ Hα (0, T ).
Henceforth we write for example (2.10)

J −α v L2 (0,T ) ∼ v Hα (0,T ) ,

when there is no fear of confusion.

 t Now we prove Theorem 2.1. By 2J = J we 2denote the integral (Jy)(t) =
1

0 y(s)ds for 0 ≤ t ≤ T and by I : L (0, T ) → L (0, T ) the identity mapping. In

this section, we consider the space L2 (0, T ) over C with the scalar product (u, v) =
T
(u, v)L2 (0,T ) = 0 u(t)v(t)dt for complex-valued u, v, and Re η and Im η denote
the real and the imaginary parts of a complex number η, respectively. Let η̄ denote
the complex conjugate of η ∈ C. Henceforth if there is no fear of confusion, then
by (·, ·) we mean (·, ·)L2 (0,T ) , and R(A) and D(A) denote the range and the domain
of an operator A: AD(A) = R(A) respectively.
Lemma 2.3 For any u ∈ L2 (0, T ), the inequality Re (J u, u) ≥ 0 holds true and
R(I + J ) = L2 (0, T ).
Proof With the notations Re J u(t) = ϕ(t) and Im J u(t) = ψ(t), the following
chain of equalities and a final estimate can be easily obtained:
 T  t   T d
Re (J u, u) = Re u(s)ds u(t)dt = Re J u(t) J u(t)dt
0 0 0 dt
 T  
dϕ dψ 1 1
= ϕ(t) + ψ(t) dt = (ϕ(t)2 + ψ(t)2 )|tt =T
=0 = |J u(T )| ≥ 0.
2
0 dt dt 2 2

Therefore Re (J u, u) ≥ 0 for u ∈ L2 (0, T ). Next for λ = 0, we have the

representation
 t
(λI + J )−1 u(t) = λ−1 u(t) − λ−2 e−(t −s)/λu(s)ds, 0≤t ≤T (2.12)
0
18 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

for u ∈ L2 (0, T ). We will prove (2.12). We set v(t) = (λI + J )−1 u(t), that is,
 t
λv(t) + v(s)ds = u(t), 0 < t < T. (2.13)
0

Let u ∈ C 1 [0, T ]. Then

1
v(0) = u(0). (2.14)
λ
Differentiating (2.13) with respect to t, we have

dv 1 1 du
(t) + v(t) = (t).
dt λ λ dt
With (2.14), we obtain
 t 
1 −1t 1 du
v(t) = e λ eλs (s)ds + u(0) .
λ 0 dt

Using integration by parts, we have

 t
1 1
e−
t−s
v(t) = u(t) − 2 λ u(s)ds, 0 < t < T. (2.15)
λ λ 0

Next let u ∈ L2 (0, T ). Then the right-hand side of (2.15) defines a function v ∈
t s t t
L2 (0, T ). Exchanging the orders of the integrals: 0 0 dξ ds = 0 ξ ds dξ ,
we calculate
 t   t  s 
1 t 1 s−ξ
v(s)ds = u(s)ds − 2 e− λ u(ξ )dξ ds
0 λ 0 λ 0 0
 t  t  t 
1 1 ξ
e− λ ds e λ u(ξ )dξ
s
= u(s)ds − 2
λ 0 λ 0 ξ
 t  t 
1 1 s s=t ξ
= u(s)ds + e− λ e λ u(ξ )dξ
λ 0 λ 0 s=ξ
 t  t 
1 1 −t ξ 1 t
= u(s)ds + e λ e u(ξ )dξ −
λ u(ξ )dξ
λ 0 λ 0 λ 0

1 t t s
= e− λ e λ u(s)ds.
λ 0
t
This and (2.15) yield λv(t) + 0 v(s)ds = u(t), 0 < t < T for each u ∈ L2 (0, T ),
which is (2.12).
2.2 Preliminaries: Operational Structure of J α 19

Setting λ = 1, by (2.12) the operator (I + J )−1 u is defined for all u ∈

which implies that R(I + J ) = L2 (0, T ). The proof of Lemma 2.3 is
L2 (0, T ),
completed. 
In terms of Lemma 2.3, for 0 < α < 1, we can define the fractional power of J ,
which we denote by J (α) (e.g., Tanabe [27, Chapter 2]). More precisely, we have
the formula

sin πα ∞ α−1
J (α)u = λ (λI + J )−1 J u dλ, u ∈ D(J ) = L2 (0, T ) (2.16)
π 0

(see also Chapter 2, §3 in [27]).

Next we will prove that the fractional power J (α) of the integral operator J
coincides with the Riemann–Liouville fractional integral operator J α on L2 (0, T ):
Lemma 2.4

(J (α)u)(t) = (J α u)(t), 0 ≤ t ≤ T , u ∈ L2 (0, T ), 0 < α < 1.

Proof By (2.12), we have

 t
−1 −1
(λI + J ) J u(t) = λ e−(t −s)/λu(s)ds, u ∈ L2 (0, T )
0

t −s
and by change of variables η = λ , we obtain
 ∞
sin πα
λα−1 (λI + J )−1 J u(t) dλ
π 0
 ∞  t 
sin πα −(t −s)/λ
= λα−2
e u(s)ds dλ
π 0 0
 t  ∞ 
sin πα
= u(s) λα−2 e−(t −s)/λdλ ds
π 0 0
 t  ∞ 
sin πα
= u(s) η−α e−η dη (t − s)α−1 ds
π 0 0

(1 − α) sin πα t
= u(s)(t − s)α−1 ds.
π 0

Here we use also

 ∞
(γ + 1)
x γ e−ax dx = , a > 0, γ > −1. (2.17)
0 a γ +1

Now the known formula (1 − α)(α) = π

sin πα implies the statement of the
lemma.
20 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

Next we consider the differential operator

⎧
⎪
⎪ d 2 u(t)
⎨ (Su)(t) = − , 0 < t < T,
dt 2  (2.18)
⎪
⎪ D(S) = u ∈ H 2 (0, T ); u(0) =
du
⎩ (T ) = 0 .
dt

Here we note that the boundary conditions u(0) = du dt (T ) = 0 should be interpreted

as the traces of u in the Sobolev space H 2 (0, T ) (see e.g., [2, 19]). It is possible to
α
define the fractional power S 2 of the differential operator S for 0 ≤ α ≤ 1 in terms
of the eigenvalues and the eigenfunctions of the eigenvalue problem for the operator
S. More precisely, let 0 < λ1 < λ2 < · · · be the eigenvalues and ψk , k ∈ N the well
known corresponding normed eigenfunctions of S. It is easy to derive √
the explicit
(2k−1)2 π 2 √
formulas for λk , ψk , k ∈ N, namely, λk = 4T 2
and ψk (t) = √ sin λk t. In
2
T
particular, we note that ψk (0) = 0 and ψk ∈ H 2 (0, T ). It is known that {ψk }k∈N is
α
an orthonormal basis of L2 (0, T ). Then the fractional power S 2 , 0 ≤ α ≤ 1 of the
differential operator S is defined by the relations
⎧ α ∞ α2
⎪
⎪ S 2u = ψk )L2 (0,T ) ψk , u ∈ D(S 2 ),
α
⎨ k=1 λk (u,
D(S 2 ) = {u ∈ L2 ; ∞
α
k=1 λk |(u, ψk )L2 (0,T ) | < ∞},
α 2
(2.19)
⎪
⎪  1
⎩ u α =
∞
λα |(u, ψk ) 2 |2 2 .
D(S 2 ) k=1 k L (0,T )

α
According to [9], the domain D(S 2 ) can be described as follows:
α
D(S 2 ) = Hα (0, T ). (2.20)

The relation (2.20) holds not only algebraically but also topologically:
α α
S2v L2 (0,T ) ∼ v Hα (0,T ) , 0 ≤ α ≤ 1, v ∈ D(S 2 ). (2.21)
α
In particular, the inclusion D(S 2 ) ⊂ H α (0, T ) holds true.
Now we are ready to prove Theorem 2.1. 
Proof of Theorem 2.1 We first state the Heinz–Kato inequality (e.g., Theo-
rem 2.3.4 in Tanabe [27]): let X be a Hilbert space and linear operators A, B
in X satisfy D(A) = D(B), Re (Au, u) ≥ 0, Re (Bu, u) ≥ 0 for u ∈ D(A) and
R(I + A) = R(I + B) = X. We assume that there exists a constant C > 0 such that

Bu ≤ C Au , u ∈ D(A).

Then, for 0 ≤ α ≤ 1, we see that D(Aα ) = D(B α ) and

√
B α u ≤ eπ α(1−α)
C α Aα u , u ∈ D(Aα ).
2.3 Definition of Generalized ∂tα in Hα (0, T ) 21

The proof of the theorem is done by using the norm equivalence between
· α and Hα (0, T ), which is justified by the Heinz–Kato inequality.
D(S )
2
First of all, it can be directly verified that D(J −1 ) = J (L2 (0, T )) = 0 H 1 (0, T ),
dw(t)
(J −1 w)(t) = , and J −1 v L2 (0,T ) = v H 1 (0,T ) for v ∈ 0 H 1 (0, T ).
dt
Therefore by (2.21) we obtain the norm equivalence

J −1 v v ∈ 0 H 1 (0, T ) = D(J −1 ) = D(S 2 ).

1 1
L2 (0,T ) ∼ S2v L2 (0,T ) ,

Direct calculations show that

 T d(ϕ + iψ)
(ϕ − iψ)dt ≥ 0,
0 dt
1
and so both J −1 and S 2 satisfy the conditions for the Heinz–Kato inequality. Hence
the Heinz–Kato inequality yields

J −α v
α α
D(J −α ) = D(S 2 ).
α
L2 (0,T ) ∼ S2v L2 (0,T ) , v ∈ D(S 2 ), (2.22)
α
By (2.20) and (2.22), we have Hα (0, T ) = D(S 2 ) = D(J −α ) = R(J α ),
so that Theorem 2.1 (i) follows. By (2.21) and (2.22), the norm equivalence
J −α v L2 (0,T ) ∼ v Hα (0,T ) holds true for v ∈ D(J −α ) = R(J α ). Next, setting
v = J α u ∈ D(J −α ) with any u ∈ L2 (0, T ), by (2.21), we obtain the following
norm equivalence:
α
u L2 (0,T ) ∼ S 2 (J α u) L2 (0,T ) ∼ J αu H α (0,T ) , u ∈ L2 (0, T ).

Therefore Theorem 2.1 (ii) follows. 

2.3 Definition of Generalized ∂tα in Hα (0, T )

Next we show a generalization of Theorem 2.1 (ii).

Theorem 2.3 Let 0 ≤ α ≤ 1 and 0 ≤ β ≤ 1.
(i) Let 0 < α + β ≤ 1. Then J α : Hβ (0, T ) −→ Hα+β (0, T ) is surjective and

J αu Hα+β (0,T ) ∼ u Hβ (0,T ) , u ∈ Hβ (0, T ). (2.23)

(ii) Let 0 ≤ β − α ≤ 1. Then

J α−β u = J −β J α u, u ∈ Hβ−α (0, T ). (2.24)

22 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

Remark 2.3 Since J α is defined by the fractional power of the operator J and J −α
by J −α = (J α )−1 for 0 ≤ α ≤ 1, the general theory (e.g., Tanabe [27]) implies

J α (J β u) = J α+β u, u ∈ L2 (0, T ),

provided that α, β, α + β ∈ [−1, 1] and we understand

L2 (0, T ), γ ≤ 0,
D(J −γ ) =
Hγ (0, T ), γ ≥ 0.

Proof
(i) We note that

Hα (0, T ) = {J α ϕ; ϕ ∈ L2 (0, T )}, 0 ≤ α ≤ 1.

First let u ∈ J α (Hβ (0, T )), that is, u = J α ϕ with some ϕ ∈ Hβ (0, T ). By
Theorem 2.1 (i), there exists ψ ∈ L2 (0, T ) such that ϕ = J β ψ. Then, noting
that J α is defined as fractional power of J , we see that u = J α ϕ = J α (J β ψ) =
J α+β ψ, that is, u ∈ Hα+β (0, T ) by Theorem 2.1 (i). Hence J α (Hβ (0, T )) ⊂
Hα+β (0, T ). Next let u ∈ Hα+β (0, T ). Then Theorem 2.1 yields the existence
of ϕ ∈ L2 (0, T ) such that u = J α+β ϕ. Since J α+β = J α J β , we obtain
u = J α (J β ϕ) ∈ J α (Hβ (0, T )). Hence J α : Hβ (0, T ) −→ Hα+β (0, T ) is
surjective.
The norm equivalence (2.23) is verified as follows:

J αu H α+β (0,T ) ∼ J −α−β J α u L2 (0,T ) = J −β u L2 (0,T ) ∼ u Hβ (0,T ) ,

by u Hγ (0,T ) ∼ J −γ u L2 (0,T ) and Theorem 2.1 (ii).

(ii) Let u ∈ Hβ−α (0, T ). Replacing β by β − α in Theorem 2.3 (i) and applying it,
we see that J α u ∈ Hβ (0, T ) = D(J −β ), and so u ∈ D(J −β J α ). Conversely,
let u ∈ D(J −β J α ). Then J α u ∈ Hβ (0, T ). Again applying Theorem 2.3
(i), where we replace β by β − α, we have u ∈ Hβ−α (0, T ), which implies
Hβ−α (0, T ) = D(J −β J α ). Since D(J α−β ) = Hβ−α (0, T ) by the definition of
J α−β = (J β−α )−1 , we obtain D(J −β J α ) = D(J α−β ). For u ∈ Hβ−α (0, T ),
we set v := J α−β u ∈ L2 (0, T ). Then u = J β−α v ∈ Hβ−α (0, T ). Therefore

J −β J α u = J −β J α (J β−α v) = J −β J β v.

Here we used J α J β−α v = J β v by Theorem 2.3 (i). Hence J −β J α u = v =

J α−β u. Thus (2.24) is proved. Thus the proof of Theorem 2.3 is completed. 
Now we define the fractional derivative ∂tα in Hα (0, T ).
2.3 Definition of Generalized ∂tα in Hα (0, T ) 23

Definition 2.1 For 0 ≤ α ≤ 1, we set

∂tα u := J −α u, u ∈ Hα (0, T )

with the domain D(∂tα ) = Hα (0, T ).

We recall that, D(∂tα ) denotes the domain of the operator ∂tα .
Remark 2.4 In this book, we mainly consider ∂tα for the case of 0 < α < 1. For
α > 1, ∈ N, on the basis of ∂tα with 0 < α < 1, we can define as follows: Let
α = m + γ with m ∈ N and 0 < γ < 1. Then
 
γ d mu
∂tα u = ∂t
dt m

with

d m−1 u d mu
D(∂tα ) = u ∈ H (0, T ); u(0) = · · · = m−1 (0) = 0, m ∈ Hγ (0, T )
m
dt dt

(cf. (3.39)), and we can argue the isomorphism and fractional differential equations
in the same way as the later parts of this book, but we omit the details and postpone
them to forthcoming publications.
By Theorem 2.1, we note that Hα (0, T ) = J α L2 (0, T ). Therefore ∂tα in
Hα (0, T ) is well-defined and ∂tα u ∈ L2 (0, T ) for u ∈ Hα (0, T ). Moreover,
∂tα : Hα (0, T ) −→ L2 (0, T ) is surjective. Indeed, let v ∈ L2 (0, T ) be arbitrarily
given. By Theorem 2.1, we have ϕ := J α v ∈ Hα (0, T ) and so ∂tα ϕ = v by the
definition, which means that ∂tα : Hα (0, T ) −→ L2 (0, T ) is surjective.
On the other hand, we can prove

du
J −1 u = , u ∈ H1 (0, T ).
dt

Indeed, setting v = J −1 u, we have v ∈ L2 (0, T ) by Theorem 2.1 and u = J v,

t
that is, u(t) = 0 v(s)ds. By v ∈ L2 (0, T ), we see that u ∈ AC[0, T ], that is, u
dt (t) = v(t) for almost all t ∈ (0, T ), that
is absolutely continuous on [0, T ] and du
−1
is, dt (t) = (J u)(t) for almost all t ∈ (0, T ). Replacing α and β respectively by
du

1 − α and 1 in Theorem 2.3 (ii), we obtain

J −α = J (1−α)−1 = J −1 J 1−α ,

that is, J −α = d
dt (J
1−α ).
24 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

Noting that J −α = J −1 J 1−α and summing up, we can state:

Theorem 2.4 Let 0 < α < 1. Then ∂tα is an isomorphism between Hα (0, T ) and
L2 (0, T ). That is, ∂tα : Hα (0, T ) −→ L2 (0, T ) is injective and surjective, and

∂tα u L2 (0,T ) ∼ u Hα (0,T ) . (2.25)

Moreover,

d 1−α
∂tα u = J −α u = (J u) = Dtα u, u ∈ Hα (0, T ) (2.26)
dt
and

∂tα u = Dtα u = dtα u for u ∈ Hα (0, T ) ∩ 0 W 1,1 (0, T ). (2.27)

Indeed, (2.27) follows from (2.26) and Lemma 1.3. Thus we can calculate ∂tα u
by means of Dtα u for u ∈ Hα (0, T ).
We are tempted to assert (2.27) for u ∈ Hα (0, T ). However, dtα u does not directly
make sense for u ∈ Hα (0, T ) and with suitable extension of dtα , we can transfer
(2.27) to Hα (0, T ) (see Theorem 2.5).
Formula w = ∂tα u = dt d
(J 1−α u) in (2.26) can correspond to the classical
inversion for finding w solving J α w = u (e.g., Gorenflo and Vessella [8]) for u ∈
1,1 (0, T ), but our construction for ∂ α guarantees the formula for u ∈ H (0, T ),
0W t α
which is a wider space than the set of all absolutely continuous functions on [0, T ].
Moreover,
Proposition 2.1

J α Dtα u(t) = J α dtα u(t) = u(t)

for u ∈ 0 W 1,1 (0, T ).

This proposition means that J α is the left inverse to Dtα and dtα for u ∈
1,1 (0, T ).
0W

Proof By Lemma 1.3, we have

du
J 1−α (t) = dtα u(t) = Dtα u(t), u ∈ 0 W 1,1 (0, T ).
dt
Therefore
du
J α J 1−α (t) = J α dtα u(t) = J α Dtα u(t), u ∈ 0 W 1,1 (0, T ).
dt
2.3 Definition of Generalized ∂tα in Hα (0, T ) 25

By J α J 1−α = J we have
 t
α 1−α du du du
J J (t) = J (t) = (s)ds = u(t)
dt dt 0 ds

as u(0) = 0. Thus the proof of the proposition is completed. 

Henceforth we regard ∂tα as an operator with the domain D(∂tα ) = Hα (0, T )
if we do not specify. We can estimate ∂tα u L2 (0,T ) by the Sobolev norm but our
theorem establishes also the surjectivity for the convenience for applications to
fractional differential equations.
We conclude this section with the equivalence of ∂tα with the closed extension
of the Caputo derivative operator dtα . As for the closed extension and the closure of
an operator, see e.g., Kato [13] (Chapter III, §5). We consider the classical Caputo
derivative
 t
1 du
dtα u(t) = (t − s)−α (s)ds
(1 − α) 0 ds

with D(dtα ) = 0 C 1 [0, T ]. We consider dtα as an operator from D(dtα ) =

0 C [0, T ] ⊂ L (0, T ) to L (0, T ).
1 2 2

By dt we denote the closure in L2 (0, T ) of dtα with D(dtα ) = 0 C 1 [0, T ], which

α

is the smallest closed extension of dtα . Then we prove:

Theorem 2.5 We have D(dtα ) = Hα (0, T ), and

dtα = ∂tα = Dtα on Hα (0, T ).

This theorem means that our definition of ∂tα is consistent with the classical Caputo
derivative by considering the closure of the operator.
Proof of Theorem 2.5 We first prove that the operator dtα is closable.
Indeed, let un ∈ 0 C 1 [0, T ] and un −→ 0 in L2 (0, T ) and dtα un −→ v in L2 (0, T )
with some v ∈ L2 (0, T ). Then by Lemma 1.3 and Theorem 2.4, we have dtα un =
∂tα un , n = 1, 2, 3, . . . Therefore ∂tα un −→ v in L2 (0, T ). Theorem 2.4 yields
that un ∈ 0 C 1 [0, T ] ⊂ Hα (0, T ) is a Cauchy sequence in Hα (0, T ). Therefore
there exists some u ∈ Hα (0, T ) such that un −→ u in Hα (0, T ). Since un −→ 0
in L2 (0, T ), we see that u = 0, that is, un −→ 0 in Hα (0, T ). Hence again by
Theorem 2.4, we obtain ∂tα un = dtα un −→ ∂tα 0 = 0, that is, v = 0. Thus dtα is
closable.
Now we return to the proof of Theorem 2.5. We recall that u ∈ D(dtα ) if and only
if there exist un ∈ 0 C 1 [0, T ] such that un −→ u in L2 (0, T ) and dtα un is a Cauchy
sequence in L2 (0, T ). Since dtα un = ∂tα un for un ∈ 0 C 1 [0, T ] by Lemma 1.3 and
Theorem 2.4, we see that un is a Cauchy sequence in Hα (0, T ). Therefore there
exists u ∈ Hα (0, T ) such that un −→  u in Hα (0, T ) by Theorem 2.4. Since un −→
u in L2 (0, T ), we obtain  u = u and un −→ u in Hα (0, T ) and dtα un −→ ∂tα u in
26 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

L2 (0, T ). The definition of dtα yields

dtα u = lim dtα un = lim ∂tα un = ∂tα u.

n→∞ n→∞

Hence we proved that D(dtα ) ⊂ Hα (0, T ) and dtα u = ∂tα u for u ∈ D(dtα ).
H (0,T )
Conversely assume that u ∈ Hα (0, T ). Since 0 C 1 [0, T ] α = Hα (0, T ) by
Lemma 2.2, there exist un ∈ 0 C 1 [0, T ] such that un −→ u in Hα (0, T ). Then
∂tα un = dtα un −→ ∂tα u in L2 (0, T ) by Theorem 2.4. That is, un ∈ 0 C 1 [0, T ] is dtα -
convergent to u and so u ∈ D(dtα ) and dtα u = ∂tα u. Thus the proof of Theorem 2.5
is complete. 
Remark 2.5 In Sect. 2.5, we discuss the case where we start the operator dtα with
D(dtα ) = C 1 [0, T ], not 0 C 1 [0, T ].

2.4 Some Functions in Hα (0, T )

In this section, we show that the function t γ and some functions defined by the
Mittag-Leffler functions are in Hα (0, T ). Some of these inclusions are used in later
sections.
For α, β > 0, we define
 ∞ zk
Eα,β (z) = k=0 (αk+β) ,
∞ zk
(2.28)
Eα,1 (z) = k=0 (αk+1) , z ∈ C.

These functions are called the Mittag-Leffler functions and play important roles in
the fractional calculus (e.g., [15, 24]).
It is known that Eα,β (z) is an entire function in z with α, β > 0. The Mittag-
Leffler functions have been well studied and here we describe only a few of the
important properties used later.
Lemma 2.5
(i) Let 0 < α < 2 and β > 0. We assume that πα 2 < μ < min{π, πα}. Then there
exists a constant C = C(α, β, μ) > 0 such that

C
|Eα,β (z)| ≤ , μ ≤ | arg z| ≤ π.
1 + |z|

(ii) For λ ∈ R and α > 0, m ∈ N, we have

dm
Eα,1 (−λt α ) = −λt α−m Eα,α−m+1 (−λt α ), t > 0.
dt m
2.4 Some Functions in Hα (0, T ) 27

The proof of (i) can be found e.g., in [24]. The proof of (ii) is seen directly

because Eα,1 (z) is entire in z and we can differentiate Eα,1 (−λt α ) = ∞ (−λt α )k
k=0 (αk+1)
termwise.
Proposition 2.2 Let 0 < α < 1 and λ ∈ R. Then

Eα,1 (−λt α ) − 1 ∈ Wα (0, T ) ⊂ 0 W 1,1 (0, T ) ∩ Hα (0, T ).

Here we recall that Wα (0, T ) is defined by (2.5).

Proof The inclusion Wα (0, T ) ⊂ Hα (0, T ) is proved already in Lemma 2.1.
Moreover, Lemma 2.5 yields

d
Eα,1 (−λt α ) = −λt α−1 Eα,α (−λt α )
dt
and
 
d 
 Eα,1 (−λt α ) ≤ |λ|t α−1 |Eα,α (−λt α )| ≤ C|λ|t α−1 , t > 0.
 dt 

By Eα,1 (0) − 1 = 1
(1) − 1 = 0, the definition of Wα (0, T ) implies the
conclusion. 
Proposition 2.3 Let 0 < α < 1. Then
 t
(t − s)α−1 Eα,α (−λ(t − s)α )f (s)ds ∈ Hα (0, T )
0

and
 t 
 
 (t − s)α−1 Eα,α (−λ(t − s)α )f (s)ds  ≤C f
  L2 (0,T )
0 Hα (0,T )

for all f ∈ L2 (0, T ).

For f ∈ L2 (0, T ) we cannot expect more regularity than Hα (0, T ). For f ∈
L2 (0, T ),
we set
 t
v(t) := (t − s)α−1 Eα,α (−λ(t − s)α )f (s)ds.
0

Then ∂tα v is well-defined by v ∈ Hα (0, T ), and in Sect. 3.4 of Chap. 3, we prove

that v satisfies

∂tα v(t) = −λv(t) + f (t), 0<t<T

(see Proposition 3.2).

28 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

We can generalize the proposition: if f ∈ L2 (0, T ) and G is an entire function,

then
 t
(t − s)α−1 G((t − s)α )f (s)ds ∈ Hα (0, T ).
0

Proof We set
∞

(−λ)k s αk+α−1
gk (s) := , k = 0, 1, 2, 3, .., R(s) = gk (s).
(α(k + 1))
k=N

Then
∞
 
N −1
(−λ)k (t − s)αk+α−1
(t − s)α−1 Eα,α (−λ(t − s)α ) = = gk (t − s) + R(t − s).
(α(k + 1))
k=0 k=0
(2.29)

We choose N ∈ N such that

3
αN > − α. (2.30)
2
Then

αk + α − 1 ≥ αN + α − 1 > 0 if k ≥ N,

and so R ∈ C[0, ∞) and R(0) = 0. By the termwise differentiation, we obtain

∞
 ∞

d (−λ)k (αk + α − 1)s αk+α−2 (−λ)k s α(k−N)
R(s) = = s αN+α−2
ds (α(k + 1)) (α(k + 1) − 1)
k=N k=N
∞
 (−λ)N+j s αj
= s αN+α−2 = (−λ)N s αN+α−2 Eα,αN+α−1 (−λs α ).
(αN + α − 1 + αj )
j =0

ds ∈ L (0, T ). Therefore
Since αN + α − 2 > −1 by (2.30), we see that dR 1

 t  t
d dR
R(t − s)f (s)ds = (t − s)f (s)ds,
dt 0 0 dt

so that the Young inequality for the convolution (Lemma A.1 in the Appendix)
yields
  t 
d 
 R(t − s)f (s)ds  ≤C f
 dt  2 L2 (0,T ) ,
0 L (0,T )
2.4 Some Functions in Hα (0, T ) 29

t
which means that 0 R(t − s)f (s)ds ∈ W 1,1 (0, T ). Moreover,
  t   t
d 
 R(t − s)f (s)ds ≤C (t − s)αN+α−2 |f (s)|ds.
 dt 
0 0

Since αN > 32 − α by (2.30), we have 2αN + 2α − 4 > −1. Therefore the Cauchy–
Schwarz inequality and (2.30) yield
  t   t  12
d 
 
R(t − s)f (s)ds  ≤ C (t − s) 2αN+2α−4
ds f
 dt L2 (0,T )
0 0
C 3
= 1
t αN+α− 2 f L2 (0,T ) ≤ C1 f L2 (0,T ) .
(2αN + 2α − 3) 2

t
Again by (2.30), we see that αN + α − 1 > 0, so that 0 R(t − s)f (s)ds vanishes
at t = 0. Hence we proved
 t
R(t − s)f (s)ds ∈ 0 W 1,1 (0, T ) ∩ Wα (0, T ) ⊂ 0 W 1,1 (0, T ) ∩ Hα (0, T ).
0

Therefore by (2.27) we have

 t   t   s 
1 −α d
∂tα R(t − s)f (s)ds = (t − s) R(s − ξ )f (ξ )dξ ds.
0 (1 − α) 0 ds 0

Hence
  t   t   s 
 α  
−α  d

∂ 
R(t − s)f (s)ds  ≤ C (t − s)  R(s − ξ )f (ξ )dξ  ds
 t ds 0
0 0
 t
≤C (t − s)−α ds × C1 f L2 (0,T ) ≤ CT 1−α f L2 (0,T ) .
0

Therefore
  t 2 
 α  T
∂ R(t − s)f (s)ds  ≤ C 2 T 2−2α f 2
dt,
 t  L2 (0,T )
0 L2 (0,T ) 0

that is, (2.25) yields

 t 
 
 R(t − s)f (s)ds  ≤C f (2.31)
  L2 (0,T ) .
0 Hα (0,T )
30 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

Moreover, by the definition (1.5) of J α , the Eq. (2.29) implies

 t
(t − s)α−1 Eα,α (−λ(t − s)α )f (s)ds
0


N−1  t
= (−λ)k (Jαk+α f )(t) + R(t − s)f (s)ds.
k=0 0

Since 0 < αk + α ≤ αN for 0 ≤ k ≤ N − 1, Theorem 2.1 yields

Hαk+α (0, T ), αk + α < 1,

Jαk+α f ∈
H1 (0, T ), αk + α ≥ 1, k = 0, . . . , N − 1

and so


N−1 
N−1
(−λ)k Jαk+α f ∈ Hαk+α (0, T ) ∩ H1 (0, T ) = Hα (0, T ).
k=0 k=0

Thus with (2.31), the proof of Proposition 2.3 is complete. 

Moreover, we can prove:
Proposition 2.4 Let 0 < α < 1
2 and δ > 0 satisfy α + δ − 1
2 < 0. Then

1
t α+δ− 2 ∈ Hα (0, T ) = H α (0, T ).

Proof In view of Theorem 2.1, it suffices to verify

1
t α+δ− 2 = J α w0 (t),

where

Γ α+δ+ 1
2 1
w0 (t) := t δ− 2 ∈ L2 (0, T ),
Γ δ+ 1
2

but we prove directly. We set

1
θ = −α − δ + .
2

Then 0 < θ < 1

2 − α. It suffices to prove
 
T T |t −θ − s −θ |2
dsdt < ∞. (2.32)
0 0 |t − s|1+2α
2.4 Some Functions in Hα (0, T ) 31

Using
 T  T   T  s 
· · · ds dt = · · · dt ds,
0 t 0 0

similarly to (2.7), we have

   T  t −θ   T  T −θ 
T T |t −θ − s −θ |2 |t − s −θ |2 |t − s −θ |2
dsdt = ds dt + ds dt
0 0 |t − s|1+2α 0 0 |t − s|
1+2α
0 t |t − s|1+2α
  t −θ 
T |t − s −θ |2
=2 ds dt.
0 |t − s|
1+2α
0

We can prove
 T  t 
s (t − s) t ds dt < ∞
γ1 γ2 γ3
(2.33)
0 0

if γ1 , γ2 > −1 and γ1 + γ2 + γ3 > −2. 

Proof of (2.33) By γ1 , γ2 > −1, we have
 T  t  
(1 + γ1 )(1 + γ2 ) T γ1 +γ2 +1 γ3
s (t − s) ds t γ3 dt =
γ1 γ2
t t dt,
0 0 (2 + γ1 + γ2 ) 0

which is convergent by γ1 + γ2 + γ3 + 1 > −1. Thus the proof of (2.33) is

complete. 
Now we will prove
 T  t −θ  T t
|t − s −θ |2 −2θ −2θ |t − s |
θ θ 2
dsdt = t s dsdt < ∞. (2.34)
0 |t − s| |t − s|1+2α
1+2α
0 0 0

We choose 0 < μ < 1 such that


1 1−α
< μ < min 1, . (2.35)
2 − 2θ 1−θ
 
1
This is possible because 0 < 2−2θ < min 1, 1−α
1−θ by 0 < θ <
1
2 and 0 < α < 12 .

1−θ − 2−2θ = 2−2θ > 0 by 0 < α < 2 .

1
Indeed, 2−2θ < 1 is clearly seen, and 1−α 1 1−2α 1

By 0 < θ < 1, we can easily see that a + b ≥ (a + b) for a, b ≥ 0 and so

θ θ θ

|t θ − s θ | ≤ |t − s|θ . Hence

|t θ − s θ |2μ ≤ |t − s|2μθ

for 0 ≤ s ≤ t. On the other hand, the mean value theorem and θ < 1 yield

|t θ − s θ | ≤ max θ ηθ−1 (t − s) ≤ θ s θ−1 (t − s)

s≤η≤t
32 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

for 0 ≤ s ≤ t. Therefore

|t θ − s θ |2(1−μ) ≤ Cs 2(1−μ)(θ−1) |t − s|2(1−μ),

and so

|t θ − s θ |2 = |t θ − s θ |2μ |t θ − s θ |2−2μ ≤ Cs 2(1−μ)(θ−1)(t − s)2(1−μ)+2μθ .

Hence
   
T t |t θ − s θ |2 t −2θ s −2θ T t
dsdt ≤ C s 2(1−μ)(θ −1)−2θ t −2θ |t − s|1−2α−2μ+2μθ dsdt.
0 0 |t − s|1+2α 0 0

By 0 < α < 12 , 0 < θ < 1

2 − α and (2.35), we directly verify

2(1 − μ)(θ − 1) − 2θ > −1, 1 − 2α − 2μ + 2μθ > −1

and

{2(1 − μ)(θ − 1) − 2θ } + (−2θ ) + (1 − 2α − 2μ + 2μθ ) > −2,

so that (2.33) yields (2.34). Thus the proof of Proposition 2.4 is complete. 
Remark 2.6 It seems that we cannot directly prove the lemma without introducing
the parameter μ ∈ (0, 1).
1  δ+ 12
Since Proposition 2.4 proves C0 t α+δ− 2 ∈ Hα (0, T ), where C0 = , we
 α+δ+ 21
can apply Theorem 2.5 to calculate
1 1 1
∂tα (C0 t α+δ− 2 ) = dtα (C0 t α+δ− 2 ) = Dtα (C0 t α+δ− 2 )
 t
C0 d
(t − s)−α s α+δ− 2 ds
1
=
(1 − α) dt 0

C0 (1 − α) α + δ + 1
2 d δ+ 1 1
= t 2 = t δ− 2 .
(1 − α)  δ+ 3 dt
2

Here we used  δ + 32 = δ + 12  δ + 12 .
Therefore with the closure of dtα defined in Sect. 2.3 and ∂tα , we can justify
1 1
dtα (C0 t α+δ− 2 ) = t δ− 2 (2.36)

for 0 < α < 1

2 and small δ > 0 satisfying α + δ − 1
2 < 0.
2.5 Definition of ∂tα in H α (0, T ) 33

2.5 Definition of ∂tα in H α (0, T )

We have defined ∂tα u for u ∈ Hα (0, T ). Next, it is natural to define the fractional
derivative in H α (0, T ) which is wider than Hα (0, T ) for 0 < α < 1, although we
work mainly within Hα (0, T ) for discussing fractional differential equations.
Since Hα (0, T ) is a set of functions in H α (0, T ) which “vanish” in some sense
at t = 0, the following definition is naive but may be reasonable for 0 < α < 1 with
α = 12 : we define ∂tα v for v ∈ H α (0, T ) by

∂tα v, v ∈ H α (0, T ), 0 < α < 12 ,

∂tα v = (2.37)
∂tα (v − v(0)), v ∈ H α (0, T ), 12 < α < 1.

Then ∂tα u is well-defined because Hα (0, T ) = H α (0, T ) for 0 < α < 12 , and
H α (0, T ) ⊂ C[0, T ] for 12 < α < 1 which enables us to define v(0).
For 0 < α < 1 and α = 12 , we note that this definition gives the same result for
v ∈ Hα (0, T ), and so we can extend the domain of ∂tα from Hα (0, T ) to H α (0, T ).
1
On the other hand, the case of α = 1
2 is delicate, and we do not define ∂t2 u for
1 1
u ∈ H 2 (0, T )  H 1 (0, T ). For example, we note that 1 ∈ H 2 (0, T ) but 1 ∈
2
1 1
H 1 (0, T ). We do not define ∂t2 1 in L2 (0, T ), although we can calculate dt2 1 = 0
2
1
1 − 12
and Dt2 1 = t pointwise.
( 21 )
By the definition we note that

∂tα (u1 + u2 ) = ∂tα u1 + ∂tα u2 , u1 , u2 ∈ H α (0, T )

for 0 < α < 1 and α = 12 .

Throughout this book, we always interpret ∂tα u with 0 < α < 1 and α = 1
2 in
the sense of (2.37) for u ∈ H α (0, T ).
Now we show properties of ∂tα in conventional Sobolev space H α (0, T ).
Proposition 2.5
(i)

t −α
(1−α) , 0 < α < 12 ,
∂tα 1 = (2.38)
0, 12 < α < 1.

(ii)

Dtα u, 0 < α < 12 ,

∂tα u = (2.39)
dtα u, 12 < α < 1

for u ∈ C 1 [0, T ].
34 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

(iii) Let 12 < α < 1 and u ∈ H α (0, T ). Then there exists a sequence ϕk ∈ C 1 [0, T ],
k ∈ N such that ϕk −→ u in H α (0, T ) and ∂tα u = limk→∞ ∂tα ϕk in L2 (0, T ).
(iv) Let 0 < α < 12 . Then there exists a constant C > 0 such that

C −1 u − a H α (0,T ) ≤ ∂tα u L2 (0,T ) + |a| ≤ C( u − a H α (0,T ) + |a|)

for all a ∈ R and u ∈ H α (0, T ).

(v) Let 12 < α < 1. Then there exists a constant C > 0 such that

C −1 u H α (0,T ) ≤ ∂tα u L2 (0,T ) + |u(0)| ≤ C u H α (0,T )

for all u ∈ H α (0, T ).

Proof
(i) By (2.37), setting u(t) ≡ 1, we have ∂tα 1 = ∂tα (u − u(0)) = ∂tα (1 − 1) =
∂tα 0 = 0 for 12 < α < 1. For 0 < α < 12 , by (2.26) we can directly calculate
 t
d 1−α d 1
∂tα 1 =
(J 1) = (t − s)−α ds
dt dt (1 − α) 0
 
d 1 t −α
= t 1−α = .
dt (1 − α)(1 − α) (1 − α)

This completes the proof of (i).

(ii) Since u ∈ C 1 [0, T ] ⊂ H α (0, T ) = Hα (0, T ) for 0 < α < 12 and u is
absolutely continuous, equality (2.26) yields that ∂tα u = Dtα u for 0 < α < 12 .
Next we verify (2.39) for 12 < α < 1. By (2.37), we have ∂tα u = ∂tα (u − u(0)).
Applying Lemma 1.1 to u − u(0) and noting that u ∈ C 1 [0, T ], we obtain
 t
1 d
∂tα (u − u(0)) = dtα (u − u(0)) = (t − s)−α (u(s) − u(0))ds
(1 − α) 0 ds
 t
1 du
= (t − s)−α (s)ds = dtα u(t).
(1 − α) 0 ds

(iii) Since C 1 [0, T ] is dense in H α (0, T ), there exists a sequence ϕk ∈ C 1 [0, T ],

k ∈ N such that ϕk −→ u in H α (0, T ). By α > 12 , the Sobolev embedding
yields ϕk (0) −→ u(0). Therefore ϕk − ϕk (0) ∈ 0 C 1 [0, T ] and ϕk − ϕk (0) −→
u − u(0) in Hα (0, T ). By the definition (2.37), we see that

∂tα u = ∂tα (u − u(0)) = lim ∂tα (ϕk − ϕk (0)) = lim ∂tα ϕk

k→∞ k→∞

in L2 (0, T ).
2.5 Definition of ∂tα in H α (0, T ) 35

(iv) For 0 < α < 12 , we note that Hα (0, T ) = H α (0, T ), and so u ∈ H α (0, T )
implies u − a ∈ Hα (0, T ). Therefore, by Theorem 2.2, we have

u−a H α (0,T ) ∼ J −α (u − a) L2 (0,T ) = ∂tα u − ∂tα a L2 (0,T ) (2.40)

if u − a ∈ Hα (0, T ). Hence (2.38) and (2.40) imply

u−a H α (0,T ) ≥ u H α (0,T ) − a H α (0,T )

≥ C1 ∂tα u L2 (0,T ) − C2 ∂tα a L2 (0,T )

C2
= C1 ∂tα u L2 (0,T ) − t −α L2 (0,T ) |a| ≥ C1 ∂tα u L2 (0,T ) − C3 |a|,
(1 − α)

which is the second inequality in (iv). Next, (2.40) yields

u−a H α (0,T ) ≤ u H α (0,T ) + a H α (0,T )

≤ C4 ( ∂tα u L2 (0,T ) + ∂tα a L2 (0,T ) ) ≤ C5 ( ∂tα u L2 (0,T ) + |a|),

which is the first inequality in (iv).

(v) By (2.37) and Theorems 2.1 and 2.2, we have

∂tα u L2 (0,T ) = ∂tα (u−u(0)) L2 (0,T ) ∼ u−u(0) Hα (0,T ) ∼ u−u(0) H α (0,T ) .

Therefore

C6 ∂tα u L2 (0,T ) ≥ u H α (0,T ) − u(0) H α (0,T )

and

∂tα u L2 (0,T ) ≤ C6 ( u H α (0,T ) + u(0) H α (0,T ) ).

Directly, we see that

u(0) H α (0,T ) = |u(0)| 1 H α (0,T ) ≤ C7 |u(0)|.

Moreover, by the Sobolev embedding for 12 < α < 1, we see |u(0)| ≤

C8 u H α (0,T ) . Hence u(0) H α (0,T ) ≤ C7 |u(0)| ≤ C9 u(0) H α (0,T ) . Thus,
the proof of (v) and so the proof of Proposition 2.5 is completed. 
It seems that the definition of ∂tα in H α (0, T ) with 12 < α < 1 may be
artificial, but the definition is related to the closure of the classical Caputo derivative
operator. In Sect. 2.3, we introduce the closure in L2 (0, T ) of the Caputo derivative
36 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

operator dtα with the domain D(dtα ) = 0 C 1 [0, T ] and we establish Theorem 2.5.
Our discussion here is similar. We consider the classical Caputo derivative
 t
1 du
α
0 dt u(t) := (t − s)−α (s)ds, D(0 dtα ) = C 1 [0, T ].
(1 − α) 0 ds

To avoid confusion, by 0 dtα we denote the classical Caputo fractional derivative with
the domain C 1 [0, T ], and we consider 0 dtα as an operator from C 1 [0, T ] ⊂ L2 (0, T )
to L2 (0, T ).
First we prove:
Lemma 2.6 For 1
2 ≤ α < 1, the operator 0 dtα is closable.
Proof We have to prove that w = 0 if un ∈ C 1 [0, T ] = D(0 dtα ), n = 1, 2, 3, . . .,
un −→ 0 in L2 (0, T ) and 0 dtα un converge to w in L2 (0, T ). By un ∈ C 1 [0, T ],
n = 1, 2, 3, . . ., we see that un − un (0) ∈ 0 C 1 [0, T ] ⊂ Hα (0, T ). Therefore (2.27)
implies
α
0 d t un = 0 dtα (un − un (0)) = dtα (un − un (0)) = ∂tα (un − un (0)).

Hence ∂tα (un − un (0)) −→ w in L2 (0, T ), and so Theorem 2.4 implies un − un (0)
is a Cauchy sequence in Hα (0, T ), and there exists v ∈ Hα (0, T ) such that

un − un (0) −→ v in Hα (0, T ). (2.41)

In particular, un − un (0) −→ v in L2 (0, T ). Since un −→ 0 in L2 (0, T ), we have

−un (0) −→ v in L2 (0, T ) and so v(t) is a constant function as limit of the constant
functions: v(t) = v0 for 0 < t < T with some constant v0 . By v ∈ Hα (0, T ),
we see that v0 ∈ Hα (0, T ). For 12 < α < 1, since v ∈ Hα (0, T ) ⊂ C[0, T ]
implies v(0) = 0, we obtain v0 = 0. For α = 12 , since v0 ∈ H 1 (0, T ), we have
 T |v(t )|2  2
2 T 1 dt < ∞. Consequently, v = 0. The proof of Lemma 2.6 is
0 t dt = v0 0 t 0
complete. 
Now in this section, let 12 ≤ α < 1. By d α 2
t , we denote the closure in L (0, T )
of 0 dt . More precisely, u ∈ D(d
α
t ) if there exists a sequence un ∈ C [0, T ],
α 1

n = 1, 2, 3, . . . such that un −→ u in L2 (0, T ) and 0 dtα un , n = 1, 2, 3, . . ., are

convergent to some w ∈ L2 (0, T ). Then we define d t u = w.
α

Finally, we prove:
Theorem 2.6
(i)

1
D(d
t ) = H (0, T )
α α
if <α<1
2
2.5 Definition of ∂tα in H α (0, T ) 37

and

H 1 (0, T ) ⊂ D(d
1
t ) ⊂ H 2 (0, T ).
α
2

(ii) Let 0 < α < 12 , and let ∂tα be defined by (2.37) for all u ∈ H α (0, T ). Then

d
t = ∂t
α α
on H α (0, T ). (2.42)

Proof
First Step We prove

1
D(d
t ) ⊂ H (0, T ),
α α
≤ α < 1. (2.43)
2

Proof of (2.43) Let u ∈ D(d α α
t ). The definition of dt implies that there exists a
sequence un ∈ C [0, T ], n = 1, 2, 3, . . . such that
1

un −→ u, α
0 d t un −→ dα
t u in L2 (0, T ). (2.44)

Then un − un (0) ∈ 0 C 1 [0, T ] ⊂ Hα (0, T ). Lemma 1.3 and (2.27) yield

α
0 d t un = dtα (un − un (0)) = ∂tα (un − un (0)) −→ 
dtα u

in L2 (0, T ). By Theorem 2.4, it follows that un −un (0), n = 1, 2, 3, . . ., is a Cauchy

sequence in Hα (0, T ), and so we can find  u ∈ Hα (0, T ) such that

un − un (0) −→ 
u in Hα (0, T ) ⊂ H α (0, T ). (2.45)

In particular, un − un (0) −→ u in L2 (0, T ). Therefore, since un −→ u in L2 (0, T ),

it follows that −un (0) −→  u − u in L2 (0, T ), so that 
u − u is a constant function.
Hence −un (0) −→  u − u in H α (0, T ). In terms of (2.45), un −→ u in H α (0, T ).
This means that u ∈ H α (0, T ), and the proof of (2.43) is complete. 
Second Step We prove

Hα (0, T ) ⊂ D(dα
t ) (2.46)

and

d
t u = ∂t u,
α α
u ∈ Hα (0, T ). (2.47)
38 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

Proof of (2.46) and (2.47) By Lemma 2.2, for any u ∈ Hα (0, T ), we can choose
a sequence un ∈ 0 C 1 [0, T ] ⊂ Hα (0, T ), n = 1, 2, 3, . . ., such that un −→ u
in Hα (0, T ). Theorem 2.4 implies that ∂tα un −→ ∂tα u in L2 (0, T ). On the other
hand, Lemma 1.3 and (2.27) yield that 0 dtα un = ∂tα un , and so 0 dtα un −→ ∂tα u in
L2 (0, T ). Since un ∈ C 1 [0, T ], by the definition of 
dtα , we obtain u ∈ D(dα
t ) and
(2.47). 
Third Step The inclusions (2.43) and (2.46) imply

Hα (0, T ) ⊂ D(d
t ) ⊂ H (0, T ).
α α
(2.48)

1
To complete the proof of Theorem 2.6, for 2 < α < 1, we have to prove

H α (0, T ) ⊂ D(dα
t ),

dtα u = ∂tα (u − u(0)) for u ∈ H α (0, T ).

Let u ∈ H α (0, T ) be arbitrary. By α > 12 , by the Sobolev embedding we see that

u ∈ C[0, T ] and u − u(0) ∈ Hα (0, T ). By Lemma 2.2, we can choose a sequence
un ∈ 0 C 1 [0, T ], n = 1, 2, 3, . . ., such that

un −→ u − u(0) in Hα (0, T ), (2.49)

and in particular, un −→ u − u(0) in L2 (0, T ). Hence

un + u(0) −→ u in L2 (0, T ). (2.50)

On the other hand, in terms of (2.49) we see from Theorem 2.4 and (2.27) that
α
0 dt (un + u(0)) = 0 dtα un = ∂tα un −→ ∂tα (u − u(0))

in L2 (0, T ). With (2.50), we see that vn := un +u(0) ∈ C 1 [0, T ] such that vn −→ u

and 0 dtα vn −→ ∂tα (u − u(0)) in L2 (0, T ). Hence the definition of d t yields u ∈
α
α 
D(dt ) and dt u = ∂t (u − u(0)) for 2 < α < 1. Thus the proof of Theorem 2.6 is
α α 1

complete. 

2.6 Adjoint of the Fractional Derivative in Hα (0, T )

In Sects. 2.3 and 2.5, we discussed the two extensions of the classical Caputo
derivative by taking the closures. As another method for the extension, we here
argue the adjoint operator and show the adjoint equality.
The results in this section are not used in the later parts of the book.
2.6 Adjoint of the Fractional Derivative in Hα (0, T ) 39

We introduce:
Definition 2.2 The formal adjoint operator (dtα )∗ to dtα is defined by
 T 
1 dϕ
(dtα u, ϕ) = (u, (dtα )∗ ϕ) + t 1−α (t)dt u(0) (2.51)
(1 − α)(1 − α) 0 dt

for u ∈ C 1 [0, T ] and ϕ ∈ C 1 [0, T ] satisfying ϕ(T ) = 0.

Then:
Lemma 2.7 For each ϕ ∈ C 1 [0, T ] with ϕ(T ) = 0, we have

−1 T dϕ
(dtα )∗ ϕ(t) = (η − t)−α (η)dη.
(1 − α) t dη

Proof Let u ∈ C 1 [0, T ]. Then, exchanging the orders of the integrals, we have
  t
T 
1 −α du
=
(dtα u, ϕ) (t − s) (s)ds ϕ(t)dt
0 (1 − α) 0 ds
 T  T 
1 du
= (t − s)−α ϕ(t)dt (s)ds.
(1 − α) 0 s ds

On the other hand, by integrating by parts and ϕ(T ) = 0, we obtain

  t =T 
T
−α (t − s)1−α 1 T dϕ
(t − s) ϕ(t)dt = ϕ(t) − (t − s)1−α (t)dt
s 1−α t =s 1−α s dt
 T
1 dϕ
=− (t − s)1−α (t)dt.
1−α s dt

Therefore again integration by parts yields

 T  T 
1 dϕ du
(dtα u, ϕ) = − (t − s)1−α (t)dt (s)ds (2.52)
(1 − α)(1 − α) 0 s dt ds
 T  s=T
1 dϕ
=− (t − s)1−α (t)dt u(s)
(1 − α)(1 − α) s dt s=0
 T  T 
1 d dϕ
+ (t − s)1−α (t)dt u(s)ds
(1 − α)(1 − α) 0 ds s dt
 T 
1 dϕ
= t 1−α (t)dt u(0)
(1 − α)(1 − α) 0 dt
40 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

  T 
1 d 1−α dϕ
+ u, (t − s) (t)dt
(1 − α)(1 − α) ds s dt
  T   T 
−1 dϕ 1 dϕ
= u, (t − s)−α (t)dt + t 1−α (t)dt u(0)
(1 − α) s dt (1 − α)(1 − α) 0 dt

for u ∈ C 1 [0, T ] and ϕ ∈ C 1 [0, T ] satisfying ϕ(T ) = 0. The proof of Lemma 2.7
is completed. 
Next:
Definition 2.3 For u ∈ Hα (0, T ), we define dtα u = f if there exists f ∈ L2 (0, T )
such that

(u, (dtα )∗ ϕ) = (f, ϕ), ϕ ∈ C01 (0, T ).

The set of such u is defined as the domain of dtα .

We see that dtα u is well-defined, that is, dtα u is uniquely determined if u ∈
D(dtα ). Indeed, if dtα u = f and dtα u = g, then the definition implies (f − g, ϕ) = 0
for all ϕ ∈ C01 (0, T ). The density of C01 (0, T ) in L2 (0, T ) implies f = g.
Example 2.1 We consider dtα 1. First by Lemma 2.7, we have
  
T −1 T
−α dϕ
(1, (dtα )∗ ϕ) = (η − t) (η)dη dt.
0 (1 − α) t dη

Exchanging the orders of the integrals and integrating by parts, we obtain

 T  η 
−1 dϕ
(1, (dtα )∗ ϕ)
= (η) −α
(η − t) dt dη
(1 − α) 0 dη 0
 T  T
1 1 1−α dϕ 1
=− η (η)dη = η−α ϕ(η)dη.
(1 − α) 1 − α 0 dη (1 − α) 0

By the definition, we see that

1
(dtα 1)(t) = t −α ,
(1 − α)

which means (dtα 1)(t) = (Dtα 1)(t) = 1 −α , but dtα 1 = 0.

(1−α) t
2.6 Adjoint of the Fractional Derivative in Hα (0, T ) 41

We can readily prove

1
0C [0, T ] ⊂ D(dtα )

and the extended dtα u coincides with (1.4) for u ∈ 0 C 1 [0, T ].

Indeed, by Lemma 2.1 we note that 0 C 1 [0, T ] ⊂ Hα (0, T ). Setting
 t
1 du
f (t) = (t − s)−α (s)ds
(1 − α) 0 ds

and repeating the computations in (2.52), we see that (dtα u, ϕ) = (u, (dtα )∗ ϕ) =
(f, ϕ) for all ϕ ∈ C01 (0, T ). Thus the proof is complete. 
In fact, we can further prove:
Proposition 2.6

Hα (0, T ) ⊂ D(dtα )

and

dtα u = ∂tα u = Dtα u, u ∈ Hα (0, T ).

In the above example, we see that 1 ∈ D(dtα ), but 1 ∈ Hα (0, T ) for 12 ≤ α < 1.
Therefore Hα (0, T )  D(dtα ) for 12 ≤ α < 1.
Proposition 2.6 generalizes (2.27) with the extension dtα and we notice that dtα
cannot be calculated in general for u ∈ Hα (0, T ).
 δ+ 12
Example 2.2 Let 0 < α < 12 , δ > 0 and C0 = . Proposition 2.4 yields
 α+δ+ 12
1
t α+δ− 2 ∈ Hα (0, T ) and Proposition 2.6 implies
1 1 1 1
dtα (C0 t α+δ− 2 ) = ∂tα (C0 t α+δ− 2 ) = Dtα (C0 t α+δ− 2 ) = t δ− 2 .

Moreover, we can directly verify the above result by the definition of dtα . For this, it
suffices to prove
1 1
(C0 t α+δ− 2 , (dtα )∗ ϕ) = (t δ− 2 , ϕ), ϕ ∈ C01 (0, T ).
T T
Indeed, using Lemma 2.7, integration by parts and 0 t · · · dη dt
T η
= 0 0 · · · dt dη, we obtain
   
1
T 1 −1 T dϕ
(C0 t α+δ− 2 , (dtα )∗ ϕ) = C0 t α+δ− 2 (η − t)−α (η)dη dt
0 (1 − α) t dη
42 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

   
−C0 T dϕ η
α+δ− 12 −α 1 T 1 dϕ
= (η) t (η − t) dt dη = − ηδ+ 2 (η)dη
(1 − α) 0 dη 0 δ+ 1
2 0 dη
 T
1 1 η=T 1 1
=− [ϕ(η)ηδ+ 2 ]η=0 + ηδ− 2 ϕ(η)dη = (t δ− 2 , ϕ).
δ+ 1
2 0

Proof of Proposition 2.6 Let u ∈ Hα (0, T ) be arbitrary. By Lemma 2.2, there exist
un ∈ 0 C 1 [0, T ], n = 1, 2, 3, . . . , such that un −→ u in Hα (0, T ). By Theorem 2.4,
we note that ∂tα un = dtα un , n = 1, 2, 3, . . .. By Theorem 2.2, we have dtα un =
∂tα un −→ ∂tα u in L2 (0, T ). By using un (0) = 0, Lemma 2.7, un ∈ 0 C 1 [0, T ] and
Definition 2.3 imply

(dtα un , ϕ) = (un , (dtα )∗ ϕ), ϕ ∈ C01 (0, T ), n = 1, 2, 3, . . .

Letting n → ∞, we obtain

(u, (dtα )∗ ϕ) = (∂tα u, ϕ), ϕ ∈ C01 (0, T ).

Since ∂tα u ∈ L2 (0, T ) for u ∈ Hα (0, T ), Definition 2.3 yields that u ∈ D(dtα ) and
dtα u = ∂tα u. Combining (2.26), we complete the proof of Proposition 2.6. 
The adjoint equality (2.51) for u ∈ 0 C 1 [0, T ] is generalized as follows.
Proposition 2.7

(∂tα u, ϕ) = (u, (dtα )∗ ϕ), u ∈ Hα (0, T ), ϕ ∈ C01 (0, T ). (2.53)

Proof For u ∈ Hα (0, T ), by Lemma 2.2, we can choose un ∈ 0 C 1 [0, T ], n ∈ N

such that un −→ u in Hα (0, T ). Then, by Lemma 2.7 we have

(∂tα un , ϕ) = (un , (dtα )∗ ϕ), ϕ ∈ C01 (0, T ).

Moreover, (2.11) yields that ∂tα un −→ ∂tα u in L2 (0, T ). Therefore letting n → ∞,

we complete the proof of Proposition 2.7. 
2.7 Laplace Transform of ∂tα 43

2.7 Laplace Transform of ∂tα

For u belonging to a certain class, we can define the Laplace transform (Lu)(p) by
 ∞
(Lu)(p) := e−pt u(t)dt
0

for Re p > p0 : some constant.

The formulae of the Laplace transforms for fractional derivatives are well-known
(e.g., [15, 24]), provided that u satisfies a certain conditions. For example,

L(dtα u)(p) = pα (Lu)(p) − pα−1 u(0) (2.54)

for Re p > p0 : some constant. The formula (2.54) is convenient for solving
fractional differential equations. However, formula (2.54) requires some regularity
for u in order that u(0) is well-defined. Formula (2.54) does not make sense for all
u ∈ H α (0, T ) with 0 < α < 12 .
Moreover, such needed regularity should be consistent with the regularity which
we can prove for solutions to fractional differential equations. In particular, the
regularity for the formula concerning the Laplace transform should be not very
strong. Thus on the regularity assumption for the formula like (2.54), we have to
make adequate assumptions for u.
In this section, we state the formula of the Laplace transform for the fractional
derivative ∂tα in Hα (0, T ).
We set

Vα (0, ∞) := {u; u|(0,T ) ∈ Hα (0, T ) for any T > 0, (2.55)

there exists a constant C = Cu > 0 such that |u(t)| ≤ CeCt for t ≥ 0}.

Here u|(0,T ) denotes the restriction of u to (0, T ). Then we can state:

Theorem 2.7 The Laplace transform L(∂tα u)(p) can be defined for u ∈ Vα (0, ∞)
by
 T
L(∂tα u)(p) = lim e−pt ∂tα u(t)dt, p > Cu
T →∞ 0

and

L(∂tα u)(p) = pα Lu(p), p > Cu .

Here Cu > 0 in some constant depending on u.

44 2 Definition of Fractional Derivatives in Sobolev Spaces and Properties

Proof First for u ∈ Hα (0, T ), by Theorem 2.3 (i), we can see that

J 1−α u ∈ H1 (0, T ) ⊂ H 1 (0, T ), J 1−α u(0) = 0 in the trace sense (2.56)

and so
d 1−α
Dtα u = J u ∈ L2 (0, T ).
dt
d 1−α
Theorem 2.4 yields ∂tα u = J u for u ∈ Hα (0, T ). Let T > 0 be arbitrarily
dt
fixed. Then, in terms of (2.56), we integrate by parts to obtain
 T  T
d 1−α
e−pt ∂tα u(t)dt = (J u)(t)e−pt dt
0 dt 0
 t=T  T
= J 1−α u(t)e−pt +p e−pt J 1−α u(t)dt
t=0 0
   
e−pT T
−α p T
−pt
t
−α
= (T − s) u(s)ds + e (t − s) u(s)ds dt
(1 − α) 0 (1 − α) 0 0

=: I1 + I2 .

Since |u(t)| ≤ C0 eC0 t for t ≥ 0 with some constant C0 > 0, we estimate

 T  T
−pT −α C0 s −pT
|I1 | ≤ Ce (T − s) e ds = Ce s −α eC0 (T −s) ds
0 0
 T  ∞
−(p−C0 )T −α −C0 s −(p−C0 )T
= Ce s e ds ≤ Ce s −α e−C0 s ds
0 0
(1 − α)
= Ce−(p−C0 )T .
C01−α

Hence if p > C0 , then limT →∞ I1 = 0.

T t T T
As for I2 , by 0 0 · · · ds dt = 0 s · · · dt ds, we see

 T  T 
p
I2 = e−pt (t − s)−α dt u(s)ds
(1 − α) 0 s
 T  T −s 
p −pη −α
= e η dη e−ps u(s)ds.
(1 − α) 0 0
2.7 Laplace Transform of ∂tα 45

For p > C0 , since |u(s)| ≤ C0 eC0 s for s ≥ 0, we have

    
 T −s  −ps  ∞
 e −pη −α
η  
dη e u(s) ≤ C0 e −pη −α
η dη e−(p−C0 )s

0 0

for all s > 0 and T > 0, and the Lebesgue convergence theorem yields
 ∞  ∞ 
p −pη −α
lim I2 = e η dη e−ps u(s)ds
T →∞ (1 − α) 0 0

p (1 − α) ∞ −ps
= e u(s)ds = pα (Lu)(p)
(1 − α) p1−α 0

for p > C0 . Thus the proof of Theorem 2.7 is complete. 

Chapter 3
Fractional Ordinary Differential
Equations

3.1 Examples

First we consider simple fractional ordinary differential equations:

Dtα u(t) = −λu(t) + f (t), 0 < t < T, (3.1)

dtα u(t) = −λu(t) + f (t), 0 < t < T. (3.2)

We note that (3.1) and (3.2), etc. are considered pointwise.

It is known that as a well-posed problem, we have to attach some initial condition,
so that we usually discuss the following initial value problems:

Dtα u(t) = −λu(t) + f (t), 0 < t < T,

(3.3)
limt ↓0 J 1−α u(t) = a,

and

dtα u(t) = −λu(t) + f (t), 0 < t < T,

(3.4)
u(0) = a.

It is proved that there exists a unique solution to (3.3) and (3.4) respectively, and the
solutions are given by the following formulae (Kilbas et al. [15], Podlubny [24]).
For (3.3)

 t
u(t) = at α−1
Eα,α (−λt ) +
α
(t − s)α−1 Eα,α (−λ(t − s)α )f (s)ds, 0 < t < T.
0
(3.5)

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 47
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5_3
48 3 Fractional Ordinary Differential Equations

For (3.4)

 t
u(t) = aEα,1 (−λt ) + α
(t − s)α−1 Eα,α (−λ(t − s)α )f (s)ds, 0 < t < T.
0
(3.6)

Here we recall that Eα,β (z) with α, β > 0 are defined by (2.28).
Here we should expect that u(t) given by (3.5) and (3.6) are verified to
pointwisely satisfy (3.3) and (3.4) respectively. Moreover, in general, for f ∈
L2 (0, T ), we do not know whether the system

Dtα u(t) = −λu(t) + f (t), 0 < t < T,

u(0) = a

possesses a unique solution.

Even for a simple initial value problem (3.4) for a fractional ordinary differential
equation, there are very few existing works which systematically establish the
uniqueness and the existence of solutions in relevant Sobolev spaces with f ∈
L2 (0, T ), which is a natural choice of a function space for f . The purpose of this
chapter is to establish the unique existence of solutions to initial value problems for
fractional ordinary differential equations.
When we will treat the problem pointwise, for example, not in L2 (0, T ), we have
to formulate the initial condition in (3.4) and interpret dtα u for f ∈ L2 (0, T ). The
latter is not direct as is seen in Sect. 2.1 of Chap. 2. In this book, on the basis of
the fractional derivative ∂tα defined in Chap. 2, we will discuss exclusively in the
Sobolev spaces Hα (0, T ) with 0 < α < 1.
In Sect. 3.4, we will return to formulation (3.4) and formula (3.6), and discuss
some details.

3.2 Fundamental Inequalities: Coercivity

In the case of α = 1, it is trivial that

 t du 1 1
(s)u(s)ds = |u(t)|2 − |u(0)|2 , 0<t<T (3.7)
0 ds 2 2

for u ∈ C 1 [0, T ]. An inequality of the type (3.7) is useful for proving the uniqueness
of solutions. Thanks to (3.7), we can easily verify that if u ∈ H 1 (0, T ) satisfies

du
(t) = λu, 0 < t < T, u(0) = 0,
dt
3.2 Fundamental Inequalities: Coercivity 49

then u = 0 in (0, T ). Indeed, multiplying this equation by u and applying (3.7) and
the Gronwall inequality, we obtain u(t) = 0 for 0 < t < T .
In this section, we discuss inequalities corresponding to (3.7) within the Sobolev
space Hα (0, T ).
We prove:
Theorem 3.1
(i)
 T 1 1
u(t)dtα u(t)dt ≥ T −α u 2
− T 1−α |u(0)|2
0 2(1 − α) L2 (0,T ) 2(2 − α)
(3.8)
for all u ∈ W 1,1 (0, T ).
(ii)
 T 1
u(t)∂tα u(t)dt ≥ T −α u 2
(3.9)
0 2(1 − α) L2 (0,T )

for all u ∈ Hα (0, T ).

Theorem 3.2
(i)
 t
1 1
(t − s)α−1 u(s)dsα u(s)ds ≥ (|u(t)|2 − |u(0)|2), 0<t <T
(α) 0 2
(3.10)
for all u ∈ W 1,1 (0, T ).
(ii)
 t
1 1
(t − s)α−1 u(s)∂sα u(s)ds ≥ |u(t)|2 , almost all t ∈ (0, T )
(α) 0 2
(3.11)
for all u ∈ Hα (0, T ).
Similar inequalities have been known under a certain regularity condition on
u = u(t) and we refer for example, to Alsaedi et al. [3], Bajlekova [4], but we
need the inequalities in Hα (0, T ) for generalized fractional derivative ∂tα .
Henceforth C, Ck > 0 denote generic constants which are dependent of T and
α, but independent of choices of a, f , u(t), etc.
Proof of Theorem 3.1 (i) First we note

W 1,1 (0,T )
C 1 [0, T ] = W 1,1 (0, T ). (3.12)

Although this can be proved by a standard method based on the mollifier, we will
give the proof at the end of Theorem 3.2.
50 3 Fractional Ordinary Differential Equations

In view of (3.12), it suffices to prove (3.8) for u ∈ C 1 [0, T ]. Indeed, we assume

that (3.8) is already proved for u ∈ C 1 [0, T ]. For arbitrary u ∈ W 1,1 (0, T ), by
(3.12) we can choose a sequence un ∈ C 1 [0, T ] such that un −→ u in W 1,1 (0, T ).
Then
 T
un (t)dtα un (t)dt
0

T −α T 1−α
≥ un 2
− |un (0)|2 , n ∈ N. (3.13)
2(1 − α) L2 (0,T ) 2(2 − α)

By Lemma 1.3 (ii) we have dtα un −→ dtα u in L1 (0, T ). By the Sobolev embedding
W 1,1 (0, T ) ⊂ C[0, T ], the convergence un −→ u in W 1,1 (0, T ) implies un −→ u
in C[0, T ] and un (0) −→ u(0). Therefore, letting n → ∞ in (3.13), we obtain (3.8)
for each u ∈ W 1,1 (0, T ).
Now we prove (3.8) for u ∈ C 1 [0, T ]. Henceforth for simplicity, we set
1
g(t) = t −α , 0 < α < 1, t > 0.
(1 − α)
Then
 t du
dtα u(t) = g(t − s) (s)ds.
0 ds
We have
 t du
u(t)dtα u(t) = u(t) g(t − s) (s)ds (3.14)
0 ds
 t  t
du du
=− g(t − s) (s)(u(s) − u(t))ds + g(t − s)u(s) (s)ds
0 ds 0 ds
=: S1 + S2 .

dg
By integration by parts and ds (s) ≤ 0 for 0 < s < T , we see
 t  
d
S1 = − g(t − s) (u(s) − u(t)) (u(s) − u(t))ds (3.15)
0 ds
 t
1 d
=− |u(s) − u(t)|2 ds
g(t − s)
2
0 ds

1 s=t 1 t dg
=− g(t − s)|u(s) − u(t)| 2
− (t − s)|u(s) − u(t)|2 ds
2 s=0 2 0 dξ
1 s=t 1
≥− g(t − s)|u(s) − u(t)|2 = g(t)|u(t) − u(0)|2 ≥ 0.
2 s=0 2
3.2 Fundamental Inequalities: Coercivity 51

For the last inequality, we used

lim g(t − s)|u(s) − u(t)|2 ≤ C lim (t − s)−α |t − s|2 = 0

s→t s→t

by u ∈ C 1 [0, T ]. Moreover, we have

 t  t
du 1 d 1
S2 = g(s)u(t − s) (t − s)ds = g(s) ((|u(t − s)|2 )ds = ∂tα |u(t)|2 .
0 dξ 2 0 dt 2
(3.16)

Therefore Lemma 1.3 (i) yiellds

 t
1 d 1
S2 = g(s)|u(t − s)|2 ds − g(t)|u(0)|2 .
2 dt 0 2

T −α
Hence, since g(T − s) ≥ (1−α) for 0 < s < T , we obtain
 T  T  T  T
u(t)dtα u(t)dt = S1 dt + S2 dt ≥ S2 (t)dt
0 0 0 0
 T  T 
1 1
≥ g(s)|u(T − s)| ds − 2
g(t)dt |u(0)|2
2 0 2 0

1 T 1 T 1−α |u(0)|2
= g(T − s)|u(s)|2 ds −
2 0 2 (1 − α)(1 − α)
1 T −α 2 1 T 1−α |u(0)|2
≥ u −
2 (1 − α) L2 (0,T ) 2 (2 − α)

for u ∈ C 1 [0, T ]. Thus the proof of (3.8) for u ∈ C 1 [0, T ] is completed, and the
proof of (i) is completed. 
Hα (0,T )
Proof of Theorem 3.1 (ii) Since 0 C 1 [0, T ] = Hα (0, T ) by Lemma 2.2, it
suffices to prove (3.9) for u ∈ 0 C 1 [0, T ]. Indeed, for each u ∈ Hα (0, T ), there
exists a sequence un ∈ 0 C 1 [0, T ], n ∈ N such that un −→ u in Hα (0, T ). Then
∂tα un −→ ∂tα u in L2 (0, T ) by Theorem 2.2. Therefore by the passage to the limit,
we obtain (3.9) for each u ∈ Hα (0, T ).
The proof of (3.9) for u ∈ 0 C 1 [0, T ] is done as follows. For u ∈ 0 C 1 [0, T ],
equalities (2.27) and Lemma 1.3 (i) yield

∂tα u = dtα u.

Thus we can repeat the proof of (i) to prove (3.9) for u ∈ 0 C 1 [0, T ]. Thus the proof
of Theorem 3.1 is completed. 
52 3 Fractional Ordinary Differential Equations

Proof of Theorem 3.2 (i) First let u ∈ C 1 [0, T ]. Then (3.14) and (3.15) imply
 t
1 1 d
u(s)dsα u(s) := S1 + S2 ≥ S2 = (t − s)−α (|u(s)|2 )ds
2 (1 − α) 0 ds
1
= dsα (|u(s)|2 ), 0 < s < T.
2
Hence, since dsα c = 0 for a constant c, we see that

1 α 1
u(s)dsα u(s) ≥ d (|u(s)|2 ) = dsα (|u(s)|2 − |u(0)|2), 0<s<T
2 s 2

for u ∈ C 1 [0, T ]. By Theorem 2.4 and |u(s)|2 − |u(0)|2 ∈ 0 C 1 [0, T ], we have

d 1−α
dsα (|u(s)|2 − |u(0)|2 ) = J (|u(s)|2 − |u(0)|2 ) = J −α (|u(s)|2 − |u(0)|2 ).
dt
Therefore
1 −α
u(s)dsα u(s) ≥ J (|u(s)|2 − |u(0)|2 ), 0 < s < T , u ∈ C 1 [0, T ]. (3.17)
2
Since

J αu ≥ J αv in (0, T ) if u ≥ v in (0, T ),

which can be verified by (t − s)−α > 0 for 0 < s < t, by (3.17) we have

1 α −α
J α (udtα u)(t) ≥ J (J (|u|2 − |u(0)|2))(t)
2
1
= (|u(t)|2 − |u(0)|2), 0 < t < T,
2
and so
 t
1 1
(t − s)α−1 (dsα u(s))u(s)ds ≥ (|u(t)|2 − |u(0)|2 ), 0<t <T
(α) 0 2

for u ∈ C 1 [0, T ].
Next, let u ∈ W 1,1 (0, T ). Then by (3.12) we can choose uk ∈ C 1 [0, T ], k ∈ N
such that uk −→ u in W 1,1 (0, T ). Then dtα uk −→ dtα u in L1 (0, T ) by Lemma 1.3.
Moreover,
 t
1 1
(t − s)α−1 uk (s)dsα uk (s)ds ≥ (|uk (t)|2 − |uk (0)|2 ), k ∈ N. (3.18)
(α) 0 2
3.2 Fundamental Inequalities: Coercivity 53

The Young inequality for the convolution (Lemma A.1) implies

  t  t 
 1 1 
 (t − s)α−1
u (s)d α
u (s)ds − (t − s)α−1
u(s)d α
u(s)ds 
 (α) k s k
(α) 0 s  1
0 L (0,T )
 T
≤C (t − s)α−1 ds uk dtα uk − udtα u L1 (0,T )
0

≤ CT α uk dtα uk − udtα u L1 (0,T ) −→ 0

as k → ∞ by dtα uk −→ dtα u in L1 (0, T ) and uk −→ u in W 1,1 (0, T ) ⊂ C[0, T ].

Therefore we can choose a subsequence k  ∈ N tending to ∞ such that
 t  t
1 1
(t − s)α−1 uk  (s)dsα uk  (s)ds −→ (t − s)α−1 u(s)dsα u(s)ds
(α) 0 (α) 0

for almost all t ∈ (0, T ) as k  → ∞. Moreover, since uk  −→ u in L1 (0, T ) as

k  → ∞, with suitable subsequence k  ∈ N of k  , we have limk  →∞ uk  (t) = u(t)
for almost all t ∈ (0, T ). Letting k  → ∞ in (3.18) with k = k  , we complete the
proof of Theorem 3.2 (i). 
Proof of Theorem 3.2 (ii) Let u ∈ Hα (0, T ). By Lemma 2.2, we can choose a
sequence un ∈ 0 C 1 [0, T ] such that un −→ u in Hα (0, T ) and ∂tα un −→ ∂tα u
in L2 (0, T ). Moreover, ∂tα un = dtα un by (2.27). Theorem 3.2 (i) yields
 t
1 1
(t − s)α−1 un (s)∂tα un (s)ds ≥ |un (t)|2 , 0 < t < T.
(α) 0 2

Letting n → ∞, we complete the proof of (ii). 

Proof of (3.12) Let u ∈ W 1,1 (0, T ). We consider the mollifier defined as follows.
By the extension theorem, we can take  u ∈ W 1,1 (R) such that supp  u is compact
and u(t) = u(t) for 0 ≤ t ≤ T . Let χ ∈ C0∞ (R) satisfy χ ≥ 0 in R, supp χ ⊂
∞
{t; |t| ≤ 1} and −∞ χ(t)dt = 1. For ε > 0, we set
   ∞
1 t
χε (t) = χ , uε (t) = χε (t − s)u(s)ds (3.19)
ε ε −∞

for u ∈ L1loc (R). Then defining vε by (3.19) for v ∈ W 1,1 (R), we see (e.g., Adams
[2]) that vε ∈ C0∞ (R) and vε −→ v in W 1,1 (R). We set
 ∞

uε (t) = χε (t − s)
u(s)ds.
−∞
54 3 Fractional Ordinary Differential Equations

Since supp  uε ∈ C0∞ (R)

u is compact, we see that supp uε is compact. Therefore 
and  uε −→  u in W (R), that is, 
1,1 uε −→ u in W (0, T ). Since 
1,1 uε |[0,T ] ∈
C 1 [0, T ]. we have proved (3.12). 
Let  ⊂ Rn be a bounded domain wih boundary ∂ of C 2 -class, and let x =
(x1 , ..., xn ) ∈ Rn .
Henceforth we regard u = u(x, t) as a mapping from t ∈ (0, T ) to a function
u(·, t) in x and for example, we introduce a function space:
 T 
L2 (0, T ; L2 ()) := u = u(x, t); u(·, t) 2
L2 ()
dt <∞ .
0

We can similarly define C([0, T ]; L2 ()), C 2 ([0, T ]; H 1()), etc.

Now we formulate Theorems 3.1 and 3.2 in L2 (0, T ; L2 ()). Henceforth for
H0 () ⊂ L2 (), identifying the dual of L2 () of L2 () with itself, we define
1

H −1 () = (H10 ()) . Then H01 () ⊂ L2 () ⊂ H −1 () algebraically and
topologically (e.g., Brezis [5]). By H −1 () ·, ·H 1 () we denote the duality pairing.
0

Theorem 3.3
(i)
 T T −α T 1−α
(dtα u(·, t ), u(·, t ))L2() dt ≥ u 2
L2 (0,T ;L2 ())
− u(·, 0) 2
0 2(1 − α) 2(2 − α) L2 ()

for u ∈ W 1,1 (0, T ; L2 ()).

(ii)
 T T −α
H −1 () ∂tα u(·, t), u(·, t)H 1 () dt ≥ u 2
L2 (0,T ;L2 ())
0 0 2(1 − α)

for u ∈ L2 (0, T ; H01 ()) ∩ Hα (0, T ; H −1 ()).

Theorem 3.4
(i)
 t
1 1
(t − s)α−1 (dsα u(·, s), u(·, s))L2 () ds ≥ ( u(·, t) 2
L2 ()
− u(·, 0) 2
L2 ()
)
(α) 0 2

for each u ∈ W 1,1 (0, T ; L2 ()).

(ii)
 t
1 1
H −1 () ∂s u(·, s), u(·, s)H01 () (t − s)α−1 ds ≥
α 2
u(·, t) L2 ()
(α) 0 2

for each u ∈ L2 (0, T ; H01 ()) ∩ Hα (0, T ; H −1 ()).

3.3 Well-Posedness for Single Linear Fractional Ordinary Differential Equations 55

Proof In terms of the mollifier, we approximate u by smooth functions uε , ε > 0

∞
as follows. Let θ ∈ C0 (R ) satisfy θ ≥ 0, supp θ ⊂ {x ∈ R ; |x| ≤ 1}, and
n n

Rn θ (x)dx = 1. For ε > 0, we set


1 x
θε (x) := θ , uε (x, t) := u(y, t)θε (x − y)dy
εn ε Rn

(e.g., Adams [2]). Let u ∈ L2 (0, T ; H01()) ∩ Hα (0, T ; H −1 ()). Then

uε ∈ L2 (0, T ; C()) ∩ Hα (0, T ; C()), and ∂tα uε −→ ∂tα u, ∇uε −→ ∇u in
L2 (0, T ; L2 ()) as ε → ∞. Moreover, if u(·, 0) ∈ L2 () and 12 < α < 1, then we
see that uε (·, 0) −→ u(·, 0) in L2 () as ε → ∞. Therefore, applying Theorem 3.2
to uε , ε > 0 and letting ε → 0, we complete the proof of (i) and (ii) of Theorem 3.4.
Theorem 3.3 can be proved similarly by means of uε . 

3.3 Well-Posedness for Single Linear Fractional Ordinary

Differential Equations

In Chap. 2, we define ∂tα in Hα (0, T ). For example, the domain D(∂tα ) requires that
the element vanishes at t = 0 if 12 < α < 1. Thus in (3.4), we have to understand
that u ∈ D(∂tα ), and u(0) = 0 in the sense of the trace if 12 < α < 1. In order to
attach non-zero value to u(0) in some sense not only for 12 < α < 1 but also for
0 < α ≤ 12 , we interpret u(0) = a as u − a ∈ Hα (0, T ). Thus in this section, we
discuss an initial value problem for a linear fractional ordinary differential equation:

∂tα (u − a) = −λu + f (t), 0 < t < T,

(3.20)
u − a ∈ Hα (0, T ).

We note that if 12 < α < 1, then u − a ∈ Hα (0, T ) yields u(0) = a, which can
justify the initial condition in the pointwise sense. In view of Proposition 2.5 (i), we
can rewrite (3.20) as in the following lemma.
Lemma 3.1 In the sense of (2.37), the system (3.20) is equivalent to
 −α
∂tα u = −λu + (1−α)
t
a + f (t), 0 < t < T,
(3.21)
u ∈ H (0, T ), if 0 < α < 12 ,
α

and

∂tα u = −λu + f (t), 0 < t < T ,

(3.22)
u(0) = a. u ∈ H α (0, T ), if 12 < α < 1.

In the case of 12 < α < 1, the initial condition u(0) = a is understood as

limt →0 u(t) = a, because H α (0, T ) ⊂ C[0, T ].
56 3 Fractional Ordinary Differential Equations

Proof Let 0 < α < 12 . Then H α (0, T ) = Hα (0, T ) and u, a ∈ Hα (0, T ). Therefore

t −α
∂tα (u − a) = ∂tα u − ∂tα a = ∂tα u − a.
(1 − α)

Hence the equivalence between (3.20) and (3.21) is verified. Next, let 12 < α < 1.
Assume that u satisfies (3.20). By the Sobolev embedding, we see that u ∈ C[0, T ],
and u(0) = a. Consequently,

∂tα (u − a) = ∂tα (u − u(0)) = ∂tα u

by the definition (2.37) of ∂tα , which verifies that a solution to (3.20) satisfies (3.22).
On the other hand, assume that u satisfies (3.22). We see that u(0) = a implies that
u−a ∈ Hα (0, T ). Therefore the definition of ∂tα again means that ∂tα u = ∂tα (u−a).
Thus the proof of the lemma is completed. 
For α = 12 , we cannot have the corresponding problem like (3.21) and the
formulation (3.21) is not convenient. Thus we mainly consider (3.20) as the initial
value problem.
We prove:
Theorem 3.5 Let λ, a ∈ R be given. There exists a unique solution u to (3.20).
Moreover, for 0 < α < 1, we can choose a constant C > 0 such that

u−a Hα (0,T ) ≤ C(|a| + f L2 (0,T ) ) (3.23)

and

u H α (0,T ) ≤ C(|a| + f L2 (0,T ) ). (3.24)

In the pointwise sense, the unique existence of solutions to initial value problems
for fractional ordinary differential equations with Dtα and ∂tα has been well studied
(e.g., Kilbas et al. [15], Podlubny [24]), but such pointwise formulations meet
difficulty in several cases such as f ∈ L∞ (0, T ). Thus we need to formulate the
initial condition by (3.20).
Proof of Theorem 3.5 By Theorem 2.4, we can rewrite (3.20) as

J −α (u − a) = λu + f (t), u − a ∈ Hα (0, T ),

which is equivalent to

u(t) = a + λ(J α u)(t) + (J α f )(t)

 t  t
λ 1
=a+ (t − s) α−1
u(s)ds + (t − s)α−1 f (s)ds,
(α) 0 (α) 0

0 < t < T. (3.25)

3.3 Well-Posedness for Single Linear Fractional Ordinary Differential Equations 57

By Theorem 2.1, we see that λJ α : L2 (0, T ) −→ Hα (0, T ) is bounded and so is

a compact operator from L2 (0, T ) to itself (e.g., [2]). Now we assume that u(t) =
λ(J α u)(t) for 0 < t < T , that is,
 t
λ
u(t) = (t − s)α−1 u(s)ds, 0 < t < T.
(α) 0

Then
 t
|u(t)| ≤ C (t − s)α−1 |u(s)|ds, 0 < t < T.
0

Applying a generalized Gronwall inequality, we obtain u = 0 in (0, T ). For

completeness, we will prove it as Lemma A.2 in the Appendix. 
Consequently, the Fredholm alternative yields that there exists a unique solution
u ∈ L2 (0, T ) satisfying (3.25). Moreover, since u − a = J α (λu + f ) in L2 (0, T ).
Theorem 2.1 (i) implies that u − a ∈ Hα (0, T ). Thus the proof of the unique
existence of u is complete.
Moreover, (3.25) yields
 t  t
|u(t)| ≤ |a| + C (t − s)α−1 |f (s)|ds + C (t − s)α−1 |u(s)|ds, 0 < t < T.
0 0
(3.26)

We set
 t
R(t) = |a| + C (t − s)α−1 |f (s)|ds.
0

Applying the generalized Gronwall inequality Lemma A.2, we have

 t
|u(t )| ≤ CR(t ) + C (t − s)α−1 R(s)ds
0
  t   t   s 
≤ C |a| + (t − s)α−1 |f (s)|ds + C (t − s)α−1 |a| + (s − ξ )α−1 |f (ξ )|dξ ds
0 0 0
(3.27)

for 0 ≤ t ≤ T . We take the norms in L2 (0, T ). The Young inequality Lemma A.1
in the Appendix yields
 t 
  T 1−α
 (t − s)α−1 |f (s)|ds  ≤ t −α f ≤ f
  L1 (0,T ) L2 (0,T ) L2 (0,T ) .
0 L2 (0,T ) 1−α
58 3 Fractional Ordinary Differential Equations

Moreover,
 t  s   t  t 
(s − ξ )α−1 |f (ξ )|dξ ds = (s − ξ )α−1 ds |f (ξ )|dξ
0 0 0 ξ
 t (t − ξ )α
= |f (ξ )|dξ,
0 α

and so
 t  s    α
  t 
 (s − ξ ) α−1
|f (ξ )|dξ ds  ≤  f
  α 1 L2 (0,T )
0 0 L2 (0,T ) L (0,T )

again by the Young inequality. By exchanging the orders of the integrals, we can
similarly obtain
 t  s 
(t − s) α−1
(s − ξ ) α−1
|f (ξ )|dξ ds
0 0
   
t t (α)2 t
= |f (ξ )| (t − s) α−1
(s − ξ ) α−1
ds dξ = (t − ξ )2α−1 |f (ξ )|dξ,
0 ξ (2α) 0

and
 t  s  
 
 (t − s)α−1 (s − ξ ) |f (ξ )|dξ ds 
α−1
 
0 0 L2 (0,T )
 t 
(α)2 


≤  (t − ξ ) 2α−1
|f (ξ )|dξ 
 ≤C f L2 (0,T ) .
(2α) 0 L2 (0,T )

Consequently, (3.27) implies u L2 (0,T ) ≤ C(|a| + f L2 (0,T ) ). Hence the first

inequality in (3.20) yields

∂tα (u − a) L2 (0,T ) ≤ C(|λ| u L2 (0,T ) + f L2 (0,T ) ) ≤ C(|a| + f L2 (0,T ) ).

Thus the proof of (3.23) is completed. We can see (3.24) by u − a H α (0,T ) ≤

u − a Hα (0,T ) and
√
u−a H α (0,T ) ≥ u H α (0,T ) − a H α (0,T ) = u H α (0,T ) − T |a|,
√
which can be seen by a H α (0,T ) = a L2 (0,T ) = T |a| for constant a.
3.4 Alternative Formulation of Initial Value Problem 59

3.4 Alternative Formulation of Initial Value Problem

In this book, we exclusively formulate an initial value problem by (3.20). On the

other hand, in this section, we discuss

dtα u = p(t)u + f (t), 0 < t < T,

(3.28)
u(0) = a.

We recall (1.17):

du
W (0, T ) := u ∈ L (0, T );
1,1 1
∈ L1 (0, T )
dt
 
with u W 1,1 (0,T ) = u L1 (0,T ) +  du 
dt L1 (0,T ) .
Here and henceforth we assume p ∈ L∞ (0, T ) and f ∈ L2 (0, T ).
In (3.28), if u ∈ W 1,1 (0, T ), then the initial condition u(0) = a can be
immediately justified by W 1,1 (0, T ) ⊂ C[0, T ]. Moreover, by u ∈ W 1,1 (0, T ),
we can prove that dtα u ∈ L1 (0, T ) exists. In many monographs (e.g., Diethelm [6],
Kilbas et al. [15], Podlubny [24]), the initial value problem is formulated by (3.28).
Now we consider the formulation:
∂tα (u − a) = p(t)u + f (t), 0 < t < T,
(3.29)
u − a ∈ Hα (0, T ),

and compare with (3.28). Then we can prove:

Proposition 3.1
(i) Let p ∈ L∞ (0, T ) and f ∈ L2 (0, T ). If u ∈ W 1,1 (0, T ) satisfies (3.28), then u
satisfies (3.29).
(ii) Let p ∈ C 2 [0, T ] and f ∈ W 1,1 (0, T ). Then the unique solution u to (3.29) is
in W 1,1 (0, T ) and satisfies (3.28).
In (ii) we note that dtα u is defined by (1.4) because u ∈ W 1,1 (0, T ). The
proposition means that under assumptions f ∈ W 1,1 (0, T ) and p ∈ C 2 [0, T ], the
formulations (3.28) and (3.29) are equivalent. In (3.29), we note that in general we
cannot prove that u ∈ W 1,1 (0, T ) if p ∈ L∞ (0, T ) and f ∈ L2 (0, T ).
Proof of Proposition 3.1 (i) Let u ∈ W 1,1 (0, T ) satisfy (3.28). Then Lemma 1.3
(i) implies that dtα u = dtα (u − a) = Dtα (u − a) pointwise. Therefore (3.28) yields

Dtα (u − a) = p(t)u + f (t), 0 < t < T. (3.30)

Since u − a ∈ 0 W 1,1 (0, T ), applying J α to (3.30), by Proposition 2.1 and pu + f ∈

L2 (0, T ) we obtain

u − a = J α Dtα (u − a) = J α (pu + f ) ∈ R(J α ) = Hα (0, T ).

60 3 Fractional Ordinary Differential Equations

Hence u − a ∈ Hα (0, T ). Therefore (2.26) implies ∂tα (u − a) = Dtα (u − a).

Equation (3.30) yields (3.29). Thus we complete the proof of Proposition 3.1 (i). 
Proof of Proposition 3.1 (ii) For f ∈ W 1,1 (0, T ) and p ∈ C 2 [0, T ], we can prove
that u ∈ W 1,1 (0, T ) and u(0) = a. The proof is done in Theorem 3.6 (iii) later
for a more general case, and is postponed. Then Theorem 2.4 implies ∂tα (u − a) =
dtα (u − a) for u − a ∈ Hα (0, T ). By u − a ∈ W 1,1 (0, T ), the pointwise dtα defined
by (1.4) yields
 t
1 d
∂tα (u − a) = dtα (u − a)(t) = (t − s)−α (u − a)(s)ds
(1 − α) 0 ds
 t
1 du
= (t − s)−α (s)ds = dtα u(t).
(1 − α) 0 ds

Therefore (3.29) implies dtα u(t) = p(t)u + f (t) for 0 < t < T . Since u ∈
W 1,1 (0, T ) ⊂ C[0, T ], we have the initial condition u(0) = a in the sense of
limt →0 u(t) = a. Thus u satisfies (3.28) and the proof of (ii) is complete. 
Now we discuss a simple case (3.4) and clarify in which sense the solution
formula (3.6) should be understood.
In fact, we prove:
Proposition 3.2 Let f ∈ L2 (0, T ). Then u given by (3.6) satisfies

∂tα (u − a) = −λu + f (t), 0 < t < T,

(3.31)
u − a ∈ Hα (0, T ).

The existing references [15, 24] give a representation formula (3.6) for the
solution to (3.4), but it can be justified pointwise only if f has certain regularity
such as f ∈ W 1,1 (0, T ). In other words, for f ∈ L2 (0, T ) it is more consistent
for us to interpret (3.6) as solution formula for the initial value problem (3.31), not
(3.4).
Proof We set
 t
u1 (t) = aEα,1 (−λt ), α
u2 (t) = u2 (f )(t) = (t − s)α−1 Eα,α (−λ(t − s)α )f (s)ds.
0

Since u1 − a ∈ 0 W 1,1 (0, T ) ∩ Hα (0, T ) by Proposition 2.2, Eq. (2.27) yields

∂tα (u1 − a) = dtα (u1 − a) = dtα u1 = −λEα,1 (−λt α ) = −λu1 (t).

Next we have to prove

∂tα u2 = −λu2 + f (t), 0 < t < T,

u2 ∈ Hα (0, T )
3.4 Alternative Formulation of Initial Value Problem 61

for each f ∈ L2 (0, T ). Proposition 2.3 immediately u2 ∈ Hα (0, T ). We will verify

the first equation in the above. First let f ∈ C0∞ (0, T ). Then
 t
u2 (f )(t) = (t − s)α−1 Eα,α (−λ(t − s)α )f (s)ds
0
 t
= s α−1 Eα,α (−λs α )f (t − s)ds.
0

Hence
 t C α
|u2 (f )(t)| ≤ C s α−1 ds f C[0,T ] = t f C[0,T ] .
0 α

In particular, u2 (f )(0) = 0. Moreover, using f ∈ C0∞ (0, T ), we can justify:

 t
du2 (f ) df
(t) = s α−1 Eα,α (−λs α ) (t − s)ds.
dt 0 dt

Therefore
   t
 du2 (f )  C α
 (t) ≤C s α−1 ds f = t f
 dt  C 1 [0,T ] C 1 [0,T ] ,
0 α

so that u2 (f ) ∈ 0 W 1,1 (0, T ). Consequently, u2 (f ) ∈ Hα (0, T ) ∩ 0 W 1,1 (0, T ) for

f ∈ C0∞ (0, T ). Therefore (2.27) yields
 t
1 d
∂tα u2 (f )(t) = Dtα u2 (f )(t) = (t − s)−α u2 (f )(s)ds, 0 < t < T.
(1 − α) dt 0

Exchanging the orders of the integrals, we calculate

 t
1 d
(t − s)−α u2 (f )(s)ds
(1 − α) dt 0
 t  s  
1 d
= (t − s)−α (s − ξ )α−1 Eα,α (−λ(s − ξ )α )f (ξ )dξ ds
(1 − α) dt 0 0
 t  t 
d 1 −α
= (t − s) (s − ξ ) Eα,α (−λ(s − ξ ) )ds f (ξ )dξ.
α−1 α
dt 0 (1 − α) ξ

On the other hand, using the power series of Eα,α (−ληα ) for η ≥ 0, we can directly
verify
 t
1
(t − η)−α ηα−1 Eα,α (−ληα )dη = Eα,1 (−λt α ), t > 0,
(1 − α) 0
62 3 Fractional Ordinary Differential Equations

which means
 t
1
(t − s)−α (s − ξ )α−1 Eα,α (−λ(s − ξ )α )ds
(1 − α) ξ
 t −ξ
1
= ((t − ξ ) − η)−α ηα−1 Eα,α (−ληα )dη = Eα,1 (−λ(t − ξ )α ).
(1 − α) 0

Consequently,
 t
d
∂tα u2 (f )(t) = Eα,1 (−λ(t − ξ )α )f (ξ )dξ.
dt 0

Since
d
Eα,1 (−λ(t − ξ )α ) = −λ(t − ξ )α−1 Eα,α (−λ(t − ξ )α )
dt

by Lemma 2.5 (ii) with m = 1, using f ∈ C0∞ (0, T ), we have

 t
∂tα u2 (f )(t) = −λ(t − ξ )α−1 Eα,α (−λ(t − ξ )α )f (ξ )dξ + Eα,1 (0)f (t),
0

that is,

∂tα u2 (f ) = −λu2 (f ) + f (t), 0<t<T

for f ∈ C0∞ (0, T ). Finally let f ∈ L2 (0, T ) be arbitrary. We choose fn ∈ C0∞ (0, T )
such that fn −→ f in L2 (0, T ) as n → ∞. Then we already proved that
∂tα u2 (fn ) = −λu2 (fn ) + fn (t) for 0 < t < T . Again Proposition 2.3 implies that
u2 (fn ) −→ u2 (f ) in Hα (0, T ) and ∂tα u2 (fn ) −→ ∂tα u2 (f ) in L2 (0, T ). Hence,
letting n → ∞, we obtain ∂tα u2 (f ) = −λu2 (f ) + f (t) also for f ∈ L2 (0, T ).
Thus the proof of Proposition 3.2 is completed. 

3.5 Systems of Linear Fractional Ordinary Differential

Equations
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
u1 (t) f1 (t) a1
⎜ .. ⎟ ⎜ .. ⎟ ⎜ .. ⎟
Let u(t) = ⎝ . ⎠ and F (t) = ⎝ . ⎠ ∈ (L (0, T )) , a = ⎝ . ⎠ ∈ RN
2 N

uN (t) fN (t) aN
and let P (t) = (pij (t))1≤i,j ≤N with pij ∈ L∞ (0, T ) for 1 ≤ i, j ≤ N be an N × N
matrix.
3.5 Systems of Linear Fractional Ordinary Differential Equations 63

In this section, we discuss a system of linear fractional ordinary differential

equations:

∂tα (u − a) = P (t)u(t) + F (t), 0 < t < T,

(3.32)
u − a ∈ (Hα (0, T ))N .

Similarly to Theorem 3.5, we can prove:

Theorem 3.6
(i) There exists a unique solution u to (3.32).
(ii) There exists a constant C > 0 such that

u − a (Hα (0,T ))N ≤ C(|a|RN + F (L2 (0,T ))N ),

(3.33)
u (H α (0,T ))N ≤ C(|a|RN + F (L2 (0,T ))N ).

(iii) We further assume that F ∈ (W 1,1 (0, T ))N and P ∈ (C 2 [0, T ])N×N . Then
u ∈ (W 1,1 (0, T ))N and u(0) = a.
(iv) Let F = 0 and P ∈ (C 2 [0, T ])N×N . Then u − a ∈ (Wα (0, T ))N .
(v) For F ∈ (C 1 [0, T ])N and P ∈ (C 2 [0, T ])N×N , we have u−a ∈ (Wα (0, T ))N .
The proof of (v) is the same as (iv) and we omit.
Theorem 3.6, in particular parts (iii) and (v) will play an important role in proving
the unique existence of solutions to initial boundary value problems in Chap. 4.
When we can assume stronger regularity than P ∈ (C 2 [0, T ])N×N , we can simplify
the proof but we omit details.
Moreover, we recall that

dv
Wα (0, T ) = v ∈ W 1,1 (0, T ); t 1−α ∈ L∞ (0, T ), v(0) = 0 .
dt

Proof of (i) By Theorem 2.4, we can rewrite (3.32) as

J −α (u − a) = P (t)u(t) + F (t),
(3.34)
u − a ∈ (Hα (0, T ))N ,

and so

u(t) = a + J α F (t) + J α (P u)(t)

 t
1
=a + J α F (t) + (t − s)α−1 P (s)u(s)ds, 0 < t < T. (3.35)
(α) 0

By Theorem 2.1 and P ∈ (L∞ (0, T ))N×N , using the compact embedding
(Hα (0, T ))N −→ (L2 (0, T ))N , we verify that u −→ J α (P u) is a compact
operator from (L2 (0, T ))N to (L2 (0, T ))N . Therefore in view of the Fredholm
64 3 Fractional Ordinary Differential Equations

alternative, for the unique existence of solution u to (3.32), it suffices to prove

u(t) = J α (P u)(t), 0 < t < T, (3.36)

implies u = 0 for 0 < t < T .

N 1
2 2 , we have
Setting |u| = j =1 uj

 t
|u(t)| ≤ C (t − s)−α P (L∞ (0,T ))N×N |u(s)|ds, 0 < t < T.
0

Hence, by the same argument as in Theorem 3.5, the generalized Gronwall

inequality (Lemma A.2) yields u = 0 in (0, T ). Thus the proof of (i) is
completed. 
Proof of (ii) By (3.35), we obtain
 t  t
1 1
u(t) = a + (t − s)α−1 F (s)ds + (t − s)α−1 P (s)u(s)ds
(α) 0 (α) 0

for 0 < t < T . Consequently, using P ∈ (L∞ (0, T ))N×N , we have

 t
|u(t)| ≤ CR(t) + C (t − s)α−1 |u(s)|ds, 0 < t < T, (3.37)
0

where we set
 t
R(t) = |a| + C (t − s)α−1 |F (s)|ds.
0

Thus, arguing in the same way as in Theorem 3.5, we can complete the proof of
Theorem 3.6 (ii). 
Proof of (iii) We set

Lα v := J α (P (t)v), v ∈ (L2 (0, T ))N , G(t) = J α (P (t)a) + J α F (t), 0 < t < T.

Moreover, for 0 < γ < 1 and m ∈ N, we define


d m−1 v d mv
Hm+γ (0, T ) = v ∈ H (0, T ); v(0)= · · · = m−1 (0) = 0, m ∈ Hγ (0, T ) .
m
dt dt

By the definition, we see

Hm+γ (0, T ) ⊂ H m+γ (0, T ). (3.38)

3.5 Systems of Linear Fractional Ordinary Differential Equations 65

Here H m+γ (0, T ) is a fractional Sobolev space (e.g., Adams [2]), and the norm in
H m+γ (0, T ) is defined by
  m 2
d v 
v H m+γ (0,T ) = v 2
L2 (0,T )
+ 
 dt m  2
L (0,T )

   m  1
T T d v d m v 2 1 2
+  (t) − (s) dtds . (3.39)
 dt m dt m  |t − s|1+2γ
0 0

Furthermore,
⎧
⎪
⎨ v H m+γ (0,T ) , 0 < γ < 1, γ = 2 ,
1
 1
v Hm+γ (0,T ) =  T 1  d m v 2 2 (3.40)
⎪
⎩ v 1 +m + 0 t  dt m (t) dt , γ = 12 .
2 H (0,T )

We show the two lemmata.

Lemma 3.2

J α (W 1,1 (0, T ))N ⊂ (W 1,1 (0, T ))N and Lα (W 1,1 (0, T ))N ⊂ (W 1,1 (0, T ))N

for 0 < α < 1.

Lemma 3.3 Lα (Hβ (0, T ))N ⊂ (Hα+β (0, T ))N for 0 < α < 1 and 0 < β ≤ 2.
At the end of the proof of Theorem 3.6, we give the proofs of the lemmata.
Now we proceed to the proof of (iii). Since u − a = J α (P (u − a) + P a) + J α F ,
in terms of Lα , Eq. (3.34) is rewritten as

u − a = Lα (u − a) + G in (0, T ). (3.41)

By Lemma 1.3 (i) and P a, F ∈ (W 1,1 (0, T ))N , we can easily verify

G ∈ (W 1,1 (0, T ))N . (3.42)

As is already proved in parts (i) and (ii), we see that u − a ∈ (Hα (0, T ))N .
Therefore, setting v1 = Lα (u − a) and w1 = G, Eq. (3.41) and Lemma 3.3 imply
u − a = v1 + w1 with v1 ∈ (H2α (0, T ))N and w1 ∈ (W 1,1 (0, T ))N . Substituting
this into (3.41), we obtain

u − a = Lα v1 + (Lα w1 + G).

By Lemmata 3.2 and 3.3, we see that Lα v1 ∈ (H3α (0, T ))N and Lα w1 ∈
(W 1,1 (0, T ))N . Setting v2 = Lα v1 and w2 = Lα w1 + G, we reach u − a = v2 + w2
with v2 ∈ (H3α (0, T ))N and w2 ∈ (W 1,1 (0, T ))N .
66 3 Fractional Ordinary Differential Equations

By 0 < α < 1, we can find k0 ∈ N such that 1 ≤ k0 α < 2. Indeed, we can

choose the minimum k0 ∈ N satisfying k0 α ≥ 1. Then we have (k0 − 1)α < 1.
Hence k0 α < 1 + α < 2 by 0 < α < 1.
The above arguments can be repeated up to k0 , That is,

u−a = 
v + v ∈ (Hk0 α (0, T ))N ⊂ (W 1,1 (0, T ))N and w
w with   ∈ (W 1,1 (0, T ))N .

Therefore we complete the proof that u − a ∈ (W 1,1 (0, T ))N .

Finally, in terms of (3.41), we have

|(u − a)(t)| ≤ |La (u − a)(t)| + |G(t)| (3.43)

and
  t 
 1 

|La (u − a)(t)| =  (t − s) (P (s)(u(s) − a))ds 
α−1
(α) 0

≤C P (L∞ (0,T ))N×N u−a (L∞ (0,T ))N t

α

in view of u − a ∈ (W 1,1 (0, T ))N ⊂ (L∞ (0, T ))N by the Sobolev embedding. We
can directly verify that |G(t)| ≤ Ct α . Therefore limt →0 |u(t) − a| = 0 by (3.43).
Thus the proof of (iii) is complete. 
Proof of (iv) First we prove: 
Lemma 3.4 Lα h ∈ (Wα (0, T ))N for h ∈ (Wα (0, T ))N .
Proof of Lemma 3.4 Since h ∈ (Wα (0, T ))N , we have h(0) = 0, h ∈
(L∞ (0, T ))N and Lα h(0) = 0. Therefore Lemma 1.3 (i) yields
 
d d d
(Lα h)(t) = J α (P (t)h(t)) = J α (P (t)h(t))
dt dt dt
   
dP (t) dh
=Jα h(t) + J α P (t) (t) .
dt dt

Moreover, we can directly verify that

J α L∞ (0, T ) ⊂ L∞ (0, T ). (3.44)

By P ∈ (C 1 [0, T ])N×N , we have dP

dt h ∈ (L∞ (0, T ))N , so that (3.44) implies
 
dP
α
J h ∈ (L∞ (0, T ))N .
dt
3.5 Systems of Linear Fractional Ordinary Differential Equations 67

By h ∈ (Wα (0, T ))N , we have

 
 dh 
 (t) ≤ Ch t α−1 ,
 dt 

and so
    t  
 α dh   dh 
J 
P (t) (t)  ≤ C (t − s) α−1 
 dt P (t) dt (s) ds
0
 t
≤ CCh (t − s)α−1 s α−1 ds ≤ Ch t 2α−1 ≤ C1 t α−1 .
0

Therefore
 
d 
 (Lα h(t)) ≤ C2 t α−1 ,
 dt 

which means that Lα h ∈ (Wα (0, T ))N . Thus the proof of Lemma 3.4 is
complete. 
By F = 0 we have G(t) = J α (P (t)a), 0 < t < T . In terms of P ∈
(C 1 [0, T ])N×N
and Lemma 1.3 (i), we see that G(0) = 0 and
   
d d dP (t) P (0)a α−1
G(t) = J α (P (t)a) = J α a + t .
dt dt dt (α)

By (3.44), we obtain J α dP
dt a ∈ (L∞ (0, T ))N and so dG 1−α
dt t ∈ (L∞ (0, T ))N .
Consequently,

G ∈ (Wα (0, T ))N . (3.45)

In view of Lemma 3.4, we see that

Ljα G ∈ (Wα (0, T ))N , j = 0, 1, 2, . . . . (3.46)

We choose k1 ∈ N such that

3 3
k1 α > , (k1 − 1)α ≤ . (3.47)
2 2
Similarly to the proof of (iii), in terms of (3.41), we will improve the regularity of
u − a. First by Lemma 3.3, u − a ∈ (Hα (0, T ))N yields Lα (u − a) ∈ (H2α (0, T ))N .
Consequently, setting v1 = Lα (u − a) and w1 = G, by (3.41) we see u − a =
v1 +w1 with v1 ∈ (H2α (0, T ))N and w1 ∈ (Wα (0, T ))N . Therefore, continuing this
argument, in view of Lemmata 3.3 and 3.4, we obtain u − a = vk + wk for k ∈ N,
68 3 Fractional Ordinary Differential Equations

where vk ∈ (H(k+1)α (0, T ))N and wk ∈ (Wα (0, T ))N provided that (k − 1)α ≤ 2.
Hence

+W
u−a =V ,  ∈ (Hk1 α (0, T ))N ,
V  ∈ (Wα (0, T ))N .
W

 ∈ (Hk1 α (0, T ))N implies

By the Sobolev embedding and (3.47), we see that V


dV
∈ (Hk1 α−1 (0, T ))N ⊂ (L∞ (0, T ))N ,
dt
 ∈ (Wα (0, T ))N . Thus u − a ∈ (Wα (0, T ))N , and the proof of
which means that V
Theorem 3.6 is complete. 
Now we prove Lemmata 3.2 and 3.3.
Proof of Lemma 3.2 The first inclusion was proved as Lemma 1.3 (i).
By P ∈ (C 1 [0, T ])N×N , we see that P (W 1,1 (0, T ))N ) ⊂ (W 1,1 (0, T ))N and so

Lα (W 1,1 (0, T ))N = J α (P (W 1,1 (0, T ))N ) ⊂ J α (W 1,1 (0, T ))N .

Thus Lα (W 1,1 (0, T ))N ⊂ (W 1,1 (0, T ))N follows and the proof of Lemma 3.2 is
complete. 
Proof of Lemma 3.3 In terms of (2.4), (3.39) and (3.40), using P ∈ (C 2 [0, T ])N×N ,
we obtain

P (Hβ (0, T ))N ⊂ (Hβ (0, T ))N , 0 < β ≤ 2. (3.48)

Indeed, for (3.48) we need the assumption P ∈ (C 2 [0, T ])N×N , not P ∈

(C 1 [0, T ])N×N .
In order to prove Lemma 3.3, thanks to (3.48), it suffices to prove

J α Hβ (0, T ) ⊂ Hα+β (0, T ) (3.49)

for 0 < β ≤ 2 and 0 < α < 1. 

Proof of (3.49) For 0 < α + β ≤ 1 and 0 < α < 1, estimate (3.49) is already
proved in Theorem 2.3 (i). We have to prove more general cases. Let α + β > 1. It
is sufficient to prove (3.49) for β = 1 + δ with 0 ≤ δ ≤ 1. Let v ∈ Hβ (0, T ) be
dt ∈ Hδ (0, T ). Hence v ∈ Hβ (0, T ) = H1+δ (0, T ) implies
arbitrarily given. Then dv
v ∈ W (0, T ) and v(0) = 0. Moreover,
1,1

d α dv
J v = Jα
dt dt
3.6 Linear Fractional Ordinary Differential Equations with Multi-Term. . . 69

by Lemma 1.3 (i) and so

d α
J v ∈ J α (Hδ (0, T )).
dt

If α + δ ≤ 1, then Theorem 2.3 (i) yields J α (Hδ (0, T )) ⊂ Hα+δ (0, T ). By the
definition of Hα+δ+1 (0, T ), we see

J α v ∈ Hα+β (0, T ). (3.50)

Next we assume 1 < α + δ < 2. We can represent α + δ = 1 + δ1 with some

δ1 ∈ (0, 1). By 0 ≤ δ ≤ 1 and Theorem 2.1 (i), we obtain Hδ (0, T ) = J δ L2 (0, T )
and it follows from α + δ = 1 + δ1 and Lemma 1.3 (iv) that

J α Hδ (0, T ) = J α (J δ L2 (0, T )) = J α+δ L2 (0, T ) = J δ1 (J 1 L2 (0, T )).

1
d
Since dt (J δ1 (J 1 w)) = J δ1 d(Jdt w) = J δ1 w for all w ∈ L2 (0, T ), the definition of
H1+δ1 (0, T ) yields J 1+δ1 w ∈ H1+δ1 (0, T ) for each w ∈ L2 (0, T ). Therefore

J α Hδ (0, T ) ⊂ H1+δ1 (0, T ) = Hα+δ (0, T ).

Thus (3.50) holds for all (α, β) satisfying 0 < β ≤ 2 and 0 < α < 1, and so the
proof of (3.49) is complete. Thus the proof of Lemma 3.3 is finished. 

3.6 Linear Fractional Ordinary Differential Equations

with Multi-Term Fractional Derivatives

We have discussed the unique existence of the solutions to initial value problems for
fractional ordinary differential equations on the basis of ∂tα in the Sobolev spaces
Hα (0, T ), and the method is widely applicable for example, to nonlinear equations.
Rather than comprehensive discussions, here we are restricted to linear fractional
ordinary differential equations with multi-term time fractional derivatives:


m
α
rj (t)∂t j (u − a)(t) = p(t)u(t) + f (t), 0<t <T (3.51)
j =1

with

u − a ∈ Hα1 (0, T ). (3.52)

70 3 Fractional Ordinary Differential Equations

Here

0 < αm < αm−1 < · · · < α1 < 1 (3.53)

and

r1 (t) > 0, 0 ≤ t ≤ T , r1 , . . . , rm ∈ L∞ (0, T ),

(3.54)
p ∈ L∞ (0, T ), f ∈ L2 (0, T ).

Then we can prove:

Theorem 3.7 We assume (3.53) and (3.54). There exists a unique solution u to
(3.51) and (3.52), and with some constant C > 0, we have

u L2 (0,T ) ≤ C(|a| + f L2 (0,T ) ), (3.55)

u−a Hα1 (0,T ) ≤ C(|a| + F L2 (0,T ) ) (3.56)

and
1
u H α1 (0,T ) ≤ C(|a| + f L2 (0,T ) ), if 0 < α < 1 and α = . (3.57)
2
Proof We can rewrite (3.51) as


m
J −α1 (u − a)(t) + (t)u(t) + f(t),
rj (t)J −αj (u − a)(t) = p 0 < t < T.
j =2
(3.58)
Here and henceforth we set
rj (t) p(t) f (t)
rj (t) = , j = 2, . . . , m, (t) =
p , f(t) = , 0 ≤ t ≤ T.
r1 (t) r1 (t) r1 (t)

Therefore (3.51) with (3.52) is equivalent to


m
w(t) = − rj (t)J α1 −αj w(t)+ p (t)a+f(t),
p J α1 w(t)+ 0<t <T (3.59)
j =2

and

w ∈ L2 (0, T ). (3.60)

The equivalence is directly seen by setting w = J −α1 (u − a), because u − a ∈

Hα1 (0, T ) if and only if u − a = J α1 w with w ∈ L2 (0, T ) by Theorem 2.1.
3.6 Linear Fractional Ordinary Differential Equations with Multi-Term. . . 71

Since α1 − αj > 0 for j = 2, . . . , m, we see that J α1 −αj : L2 (0, T ) −→

Hα1 −αj (0, T ) by Theorem 2.2. By the compactness of the embedding:

Hα1 −αj (0, T ) −→ L2 (0, T ),

we verify that J α1 −αj : L2 (0, T ) −→ L2 (0, T ) is compact. Similarly J α1 :

L2 (0, T ) −→  ∈ L∞ (0, T ), we see that the
mL (0, T ) αis−αcompact, αand by 2rj , p
2

operator − j =2 rj (t)J 1 j + p J 1 : L (0, T ) −→ L2 (0, T ) is compact.

Therefore if


m
v(t) = − rj (t)J α1 −αj v(t) + p
J α1 v(t), 0 < t < T, (3.61)
j =2

implies v = 0 in (0, T ), then the Fredholm alternative yields the unique existence
of w satisfying (3.59) for any a ∈ R and f ∈ L2 (0, T ). Let (3.61) hold. Then
m 
 t  t
|v(t)| ≤ C (t − s)α1 −αj −1 |v(s)|ds + C (t − s)α1 −1 |v(s)|ds
j =2 0 0

for 0 < t < T . Since α1 > α2 > · · · > αm > 0, we have

(t − s)α1 −αj −1 = (t − s)α1 −α2 +(α2 −αj )−1 ≤ T α2 −αj (t − s)(α1 −α2 )−1

and

(t − s)α1 −1 = (t − s)(α1 −α2 )−1 (t − s)α2 ≤ T α2 (t − s)(α1 −α2 )−1 ,

and so
 t
|v(t)| ≤ C (t − s)α1 −α2 −1 |v(s)|ds, 0 < t < T.
0

Noting that α1 −α2 > 0, we apply the generalized Gronwall inequality (Lemma A.2)
to verify that v = 0 in (0, T ). Thus the proof of the unique existence is complete.
The estimates (3.55)–(3.57) are proved similarly to Theorem 3.5. 
Chapter 4
Initial Boundary Value Problems for
Time-Fractional Diffusion Equations

4.1 Main Results

Let  ⊂ Rn be a bounded domain with boundary ∂ of C 2 -class and let ν =

ν(x) = (ν1 , . . . , νn ) be the unit outward normal vector to ∂ at x = (x1 , . . . , xn ).
Henceforth let

∂ ∂2 ∂
∂i = , ∂i2 = , i = 1, . . . , n ∇ = (∂1 , . . . , ∂n ), ∂s = .
∂xi ∂xi2 ∂s

We would like to discuss an initial boundary value problem for a time-fractional

partial differential equation, which we can write as follows at the expense of rigor:
⎧ α
⎨ dt u(x, t) + A(t)u(x, t) = F (x, t), x ∈ , 0 < t ≤ T ,
u(x, t) = 0, x ∈ ∂, 0 < t < T , (4.1)
⎩
u(x, 0) = a(x), x ∈ .

Let −A(t) be a uniform elliptic differential operator of the second order with
(x, t)-dependent coefficients:


n 
n
(−A(t)u)(x, t) = ∂i (aij (x, t)∂j u(x, t)) + bj (x, t)∂j u(x, t) + c(x, t)u(x, t),
i,j =1 j =1

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 73
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5_4
74 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

for x ∈  and t > 0, where aij = aj i , bj , c ∈ C 2 ([0, T ]; C 1 ()), and there exists
a constant μ0 > 0 such that


n 
n
aij (x, t)ξi ξj ≥ μ0 ξj2 , x ∈ , 0 ≤ t < ∞, (ξ1 , . . . , ξn ) ∈ Rn .
i,j =1 j =1
(4.2)

Remark 4.1 We can relax the regularity conditions of the coefficients similarly to
[17, 30], by approximating them by smooth functions. However, we omit the details
for simplicity.
The main purpose of this chapter is to formulate the initial boundary value
problem for t-dependent A(t) with initial value a and non-homogeneous term F in
L2 -spaces. We refer to Kubica and Yamamoto [17], Zacher [30] for such treatments,
and here we describe a unified approach within the framework by means of ∂tα in
the Sobolev space Hα (0, T ).
In particular, for more regular F , for example F ∈ L∞ (0, T ; L2 ()), there are
several works on the well-posedness for fractional partial differential equations and
we can refer to Gorenflo et al. [11], Li et al. [18], Luchko [21], Sakamoto and
Yamamoto [26]. Here we do not intend any complete list of the references.
We mainly assume that a and F are in some L2 -spaces. Then in general we
cannot prove that u ∈ C([0, T ]; L2 ()). Therefore, similarly to the case of
fractional ordinary differential equations, we must be careful in interpreting the
initial condition u(·, 0) = a in (4.1), which cannot imply that u(·, t) −→ a in
L2 () as t → 0.
Henceforth for H01 () ⊂ L2 (), identifying the dual of L2 () of L2 ()
with itself, we define H −1 () = (H10 ()) . Then H01 () ⊂ L2 () ⊂ H −1 ()
algebraically and topologically (e.g., Brezis [5]).
In terms of the definition ∂tα in Hα (0, T ) defined in Chap. 2, we formulate an
initial boundary value problem as follows:

∂tα (u − a)(x, t) + A(t)u(x, t) = F in H −1 (), 0 < t < T , (4.3)

u(·, t) ∈ H01 (), 0 < t < T, (4.4)

u − a ∈ Hα (0, T ; H −1 ()). (4.5)

For 12 < α < 1, we have Hα (0, T ) ⊂ H α (0, T ) ⊂ C[0, T ] by the Soloblev

embedding, and so (4.5) implies that u − a ∈ C([0, T ]; H −1()) and u(x, 0) =
a(x) is satisfied in the sense of limt →0 u(·, t) = a in H −1 ().
4.1 Main Results 75

Remark 4.2 Moreover, in view of (2.38), interpreting ∂tα u by (2.37), we note that
(4.3) is equivalent to
⎧
t −α
⎪
⎪ ∂t u + A(t)u = F +
⎪
α
(1−α) a, in H −1 (), 0 < t < T ,
⎨
if 0 < α < 12 ,
(4.6)
⎪
⎪ ∂tα u + A(t)u = F, −1
in H (), 0 < t < T ,
⎪
⎩
if 12 < α < 1.

We note that if u ∈ C 1 ((0, T ]; C())∩C([0, T ]; C())∩L2 (0, T ; C 2 ()) satisfies

(4.1), then u satisfies also (4.3)–(4.5).
Now we are ready to state our main results.
Theorem 4.1 We assume regularity aij , bj , c ∈ C 2 ([0, T ]; C 1 ()), 1 ≤ i, j ≤ n.
For F ∈ L2 (0, T ; H −1 ()) and a ∈ L2 (), there exists a unique solution u ∈
L2 (0, T ; H01 ()) satisfying u − a ∈ Hα (0, T ; H −1 ()) to (4.3)–(4.5). Moreover,
there exists a constant C > 0 such that

u−a Hα (0,T ;H −1 ()) + ∇u L2 (0,T ;L2 ()) ≤ C( a L2 () + F L2 (0,T ;H −1 ()) ).

(4.7)

In Zacher [30], and Kubica and Yamamoto [17] (Theorem 1.1), with the same
regularity conditions on a and F , the unique existence of the solution u to (4.3)–
(4.5) is proved with the regularity

J 1−α (u − a) ∈ 0 H 1 (0, T ; H −1 ()), u ∈ L2 (0, T ; H01 ()).

Noting that 0 H 1 (0, T ; H −1 ()) = H1 (0, T ; H −1()) by the definition, we can

verify that the class of solutions in [17, 30] is the same as in Theorem 4.1, that is,

J 1−α (u − a) ∈ 0 H 1 (0, T ; H −1 ()), u ∈ L2 (0, T ; H01 ())

if and only if

u − a ∈ Hα (0, T ; H −1 ()), u ∈ L2 (0, T ; H01 ()).

Indeed, setting α := 1 − α and β := α in Theorem 2.3 (i) of Chap. 2, we see

that J 1−α : Hα (0, T ) −→ H1 (0, T ) is surjective and isomorphism. Therefore
J 1−α (u − a) ∈ H1 (0, T ; H −1()) = 0 H 1 (0, T ; H −1 ()) if and only if u − a ∈
Hα (0, T ; H −1 ()), which means that the solution classes coincide.
In particular, for 12 < α < 1, we have

u−a C([0,T ];H −1 ()) ≤C F L2 (0,T ;H −1 ()) . (4.8)

76 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

Indeed, the estimate (4.8) is directly seen by (4.7), because Hα (0, T ) ⊂ H α (0, T ) ⊂
C[0, T ] if 12 < α < 1 by the Sobolev embedding.
Remark 4.3 By (2.40), we can rewrite (4.7) as

∂tα u L2 (0,T ;H −1 ()) + u L2 (0,T ;H01 ())

1
≤ C( a L2 () + F L2 (0,T ;H −1 ()) ), α = . (4.9)
2
Here ∂tα is defined by (2.37).
Proof of (4.9) For 0 < α < 12 , by (2.38) we have

t −α
∂tα (u − a) = ∂tα u − ∂tα a = ∂tα u − a.
(1 − α)

Therefore, since t −α L2 (0,T ) < ∞ for 0 < α < 12 , inequality (4.7) yields
 
 t −α 
∂tα u ≤ ∂tα (u − a) +
 a
L2 (0,T ;H −1 ()) L2 (0,T ;H −1 ())
(1 − α) L2 (0,T ;H −1 ())
≤ C( F L2 (0,T ;H −1 ()) + a L2 (0,T ;H −1 ()) ).

Next, let 1
2 < α < 1. Then (4.9) is seen by the definition (2.37). 
Remark 4.4 In Sakamoto and Yamamoto [26], in a special case where A(t) is
symmetric (i.e., bj = c = 0, 1 ≤ j ≤ n) and the coefficients of A(t) are
independent of t and F = 0, then for (4.1) it is proved that

lim u(·, t) − a L2 () =0

t →0

for each a ∈ L2 (). See also Sect. 4.5.

For more regular a and F , we can prove:
Proposition 4.1 We assume all the conditions in Theorem 4.1. Moreover, let p > α2 ,
and let F ∈ Lp (0, T ; H −1()) and a ∈ H01 (). Then there exists a constant C > 0
such that
α
u−a L∞ (0,T ;L2 ()) ≤ C(t 2 a H01 () + tκ F Lp (0,T ;H −1 ()) ). (4.10)

Here
pα − 2
κ= > 0.
2p
4.1 Main Results 77

In particular,

lim u(·, t) − a L2 () = 0. (4.11)

t →0

The proposition means that with more regular a and F , we can prove the
continuity of u(·, t) at t = 0 in L2 (), and so the initial condition (4.5) holds in
the usual sense.
Remark 4.5 For dtα u = div (p(x, t)∇u(x, t)) + F (x, t) in (4.1) with a = 0,
Jin, Li and Zhou (Theorem 2.1 in [12]) proves that u ∈ C([0, T ]; L2 ()) ∩
Lp (0, T ; H 2 ()) and dtα u ∈ Lp (0, T ; L2 ()) if F ∈ Lp (0, T ; L2 ()) with
p > α1 . In fact, we can interpret that p > α1 is a critical condition for the continuity
of u in t. Proposition 4.1 requires a stronger assumption p > α2 for p, but the weaker
spatial regularity is needed. The continuity of u at t = 0 should be exploited more
but here we omit details.
For more regular a and F , we can improve the regularity of u.
Theorem 4.2 We assume regularity aij ∈ C 2 ([0, T ]; C 1 ()), bj , c ∈
C 2 ([0, T ]; C 1 ()), 1 ≤ i, j ≤ n. For F ∈ L2 (0, T ; L2 ()) and a ∈ H01 (),
there exists a unique solution u ∈ L2 (0, T ; H 2() ∩ H01 ()) satisfying
u − a ∈ Hα (0, T ; L2 ()) to (4.3)–(4.5). Moreover, there exists a constant C > 0
such that

u−a Hα (0,T ;L2 ()) + u L2 (0,T ;H 2 ()) ≤ C( F L2 (0,T ;L2 ()) + a H01 () ).

Theorems 4.1 and 4.2 are corresponding results to the classical results for the
parabolic equation (i.e., α = 1) for which we refer to Evans [7], Lions and Magenes
[19], for example.
In the case of F = 0 in Theorem 4.2, we can further prove:
Proposition 4.2 We assume all the conditions in Theorem 4.2 and F = 0 and

max aij L∞ (×(0,∞)) , max bj L∞ (×(0,∞)) , c L∞ (×(0,∞)) < ∞.

1≤i,j ≤n 1≤j ≤n

Then there exists a constant C > 0 such that

u(·, t) L2 () ≤ CeCt u(·, 0) H01 () , t ≥0

for all the solutions u to (4.3)–(4.5).

78 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

4.2 Some Results from Theorem 4.1 and Proposition 4.2

4.2.1 Interpolated Regularity of Solutions

Our classes of solutions in Theorems 4.1 and 4.2 are flexible for interpolated
regularity properties. Although we can choose a general uniform elliptic operator,
we introduce the Laplacian with the homogeneous Dirichlet boundary condition:


n
−A0 u(x) = ∂k2 u(x), D(A0 ) = H 2 () ∩ H01 ().
k=1

1
−1
Then it is known that D(A02 ) = H01 () and H −1 () = (H01 ()) = D(A0 2 ).
Moreover, we can verify that for 0 ≤ θ ≤ 1, the interpolation spaces are given by:

[Hα (0, T ; L2 ()), L2 (0, T ; H 2() ∩ H01 ())]1−θ

= [Hα (0, T ; D(A00 )), L2 (0, T ; D(A0 ))]1−θ

= Hαθ (0, T ; D(A1−θ

0 )) ⊂ Hαθ (0, T ; H
2−2θ
())

(e.g., Yamamoto [29]). Here we consider only the case of a = 0 and F ∈

L2 (0, T ; L2 ()).
Therefore for each θ ∈[0, 1], in Theorem 4.2 we have u∈Hαθ (0, T ; H 2−2θ ()).
1
In particular, by D(A02 ) = H01 (), we see that u ∈ H 1 α (0, T ; H01()) by choosing
2
θ = 12 in Theorem 4.2, which is included in Theorem 1.4 in Kubica and Yamamoto
[17].

4.2.2 The Method by the Laplace Transform for Fractional

Partial Differential Equations

Let the elliptic operator A(t) in (4.1) be t-independent:


n 
n
(Av)(x) = − ∂i (aij (x)∂j v(x)) − bj (x)∂j v(x) − c(x)v(x), x ∈ ,
i,j =1 j =1

where aij , bj , c ∈ C 2 () and (4.2) is satisfied. A convenient way for constructing
a solution to (4.3)–(4.5) is by the Laplace transform (see Sect. 2.7 of Chap. 2),
which relies on formulae of Laplace transforms of time-fractional derivatives. For
the rigorous treatments, we have to specify the class of solutions u admitting such
formulae for the Laplace transforms, but it is not often clarified in view of the
consistency with the expected regularity of solutions to (4.3)–(4.5).
4.2 Some Results from Theorem 4.1 and Proposition 4.2 79

Within our framework, we can apply Theorem 2.7 of Chap. 2 concerning the
Laplace transform of ∂tα in Vα (0, ∞). Here we sketch such treatments. As related
arguments, see Kian and Yamamoto [14].
For arbitrary T > 0, let u ∈ L2 (0, T ; H 2() ∩ H01 ()) satisfy u − a ∈
Hα (0, T ; L2 ()) and (4.3)–(4.5) with a ∈ H01 () and F = 0. In particular,

∂tα (u − a) + Au = 0 in  × (0, T ).

Taking the scalar products in L2 () of both sides with arbitrarily fixed ϕ ∈ C0∞ ()
and integrating by parts, we have

(∂tα (u − a)(·, t), ϕ)L2 () + (Au(·, t), ϕ)L2 () = 0, t > 0.

Then

∂tα {((u − a)(·, t), ϕ)L2 () } + (u(·, t), A∗ ϕ)L2 () = 0, t > 0. (4.12)

Here


n 
n
∗
(A v)(x) = − ∂i (aij (x)∂j v) + ∂j (bj (x)v(x)) − c(x)v(x), x ∈ .
i,j =1 j =1

By u − a ∈ Hα (0, T ; L2 ()), we see that ((u − a)(·, t), ϕ)L2 () ∈ Hα (0, T ).
Proposition 4.2 yields u(·, t) L2 () ≤ CeCt a H 1 () for all t ≥ 0. Hence
0

|((u − a)(·, t), ϕ)L2 () |, |(u(·, t), A∗ ϕ)L2 () | ≤ CeC1 t , t ≥ 0.

Here C > 0 depends on a and ϕ.

Consequently, ((u(·, t) − a), ϕ)L2() ∈ Vα (0, ∞) and the Laplace transform of
(u(·, t), A∗ ϕ)L2 () :
 ∞
∗
L((u(·, t), A ϕ)L2 () )(p) := e−pt (u(·, t), A∗ ϕ)L2 () dt
0

exists for p > C1 . We take the Laplace transforms of both sides of (4.12) and apply
Theorem 2.7, so that

pα (L(u − a)(·, p), ϕ)L2 () + ((Lu)(·, p), A∗ ϕ)L2 () = 0, p > C1 ,

that is,

({pα L(u − a)(·, p) + A(Lu)(·, p)}, ϕ)L2 () = 0, p > C1

80 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

for all ϕ ∈ C0∞ (), which means

pα L(u − a)(·, p) + A(Lu)(·, p) = 0, p > C1

in L2 (). Thus we can justify the method by the Laplace transform for an initial
boundary value problem (4.3)–(4.5).

4.3 Proof of Theorem 4.1

The proof is based on what is called the Galerkin approximation (e.g., Evans [7],
Lions and Magenes [19]). The proof is composed of:
(i) Construction of approximating solutions in a family of finitely dimensional
subspaces: For this step, Theorem 3.6 (iii) plays an important role.
(ii) Uniform boundedness of the approximating solutions: The coercivity Theo-
rems 3.3 and 3.4 are essential.
(iii) Then we can prove the existence of solution as a weak convergent limit of a
subsequence of the sequence of the approximating solutions. The uniqueness
of the solution follows from the coercivity.
First we prove the existence of solutions.
Let 0 < λ1 ≤ λ2 ≤ · · · be the eigenvalues of − with the zero Dirichlet
boundary condition, which are numbered according to the multiplicities, that is, λm
appears -times if the multiplicity of λm is . By ρk ∈ H 2 (), k ∈ N, we denote an
eigenfunction of − for λk such that

1, k = ,
(ρk , ρ )L2 () = δk :=
0, k = .

Then

{ρk }k∈N is a basis in each L2 (), H01 (),

(ρk , ρ )L2 () = δk , H −1 () ρk , ρ H01 () = δk . (4.13)

Moreover,
⎧ ∞ ∞
⎪
⎪ a= k=1 (a, ρk )L2 () ρk in L2 (), a 2 = k=1 |(a, ρk )L2 () |
2,
⎪
⎪ L2 ()
⎪
⎪ there exists a constant C > 0 such that
⎪
⎪ 
⎪
⎨ a= ∞ −1 () and
k=1 H −1 () a, ρk H01 () ρk in H

⎪
⎪ C −1 ∞ −1
k=1 λk |H −1 () a, ρk H01 () | ≤ a H −1 ()
2 2
⎪
⎪ 
⎪
⎪
⎪ ≤C ∞ −1
k=1 λk |H −1 () a, ρk H01 () | ,
2 a ∈ H −1 (),
⎪
⎪  
⎩ C −1 ∞ k=1 λk |(a, ρk )| ≤ a
2 2
1 ≤C ∞ k=1 λk |(a, ρk )| ,
2 a ∈ H01 ().
H0 ()
(4.14)
4.3 Proof of Theorem 4.1 81

We note that

H −1 () ϕ, ψH01 () = (ϕ, ψ)L2 () for ϕ ∈ L2 () and ψ ∈ H01 ()

(e.g., p.136 in [5]).

By the density of C0∞ (0, T ; H −1()) in L2 (0, T ; H −1()), for any ε > 0, we
can choose F ε ∈ C0∞ (0, T ; H −1 ()) such that

Fε − F L2 (0,T ;H −1 ()) < ε. (4.15)

T
Indeed, we can construct such F ε by the mollifier 0 F (x, s)χε (t − s)ds, where
∞
χε ∈ C0∞ (R), ≥ 0 satisfies supp χε ⊂ (−ε, ε) and −∞ χε (t)dt = 1 for small
ε > 0.
Let N ∈ N be fixed arbitrarily. We look for


N 
N
uεN (x, t) = ε
pN,k (t)ρk (x), FNε (x, t) = H −1 () F
ε
(·, t), ρk H 1 () ρk (x)
0
k=1 k=1
(4.16)

satisfying

(∂tα (uεN − aN ), ρ )L2 () + (A(t)uεN , ρ )L2 () = H −1 () FNε , ρ H 1 () ,
0
uεN − aN ∈ Hα (0, T ; L2 ()), 1 ≤  ≤ N, 0 ≤ t ≤ T ,
(4.17)
where we set


N
aN = ck ρk , ck = (a, ρk )L2 () . (4.18)
k=1

ε and summing up over  = 1, . . . , N,

Multiplying the first equation in (4.17) by pN,
we have

(∂tα (uεN − aN ), uεN )L2 () + (A(t)uεN , uεN )L2 ()

= H −1 () FNε , uεN H 1 () , 0 ≤ t ≤ T. (4.19)

0

We rewrite (4.17) as
 
ε
∂tα (pN, − c ) = N k=1 k (t)pN,k (t) + f (t),
ε ε 0 < t < T,
(4.20)
ε
pN, − c ∈ Hα (0, T ), 1 ≤  ≤ N,
82 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

where
⎧
⎪
⎪ fε (t) = H −1 () F ε (·, t), ρ H 1 () ,
⎪
⎪  0
⎪
⎨ k (t) = ni,j =1 (∂i (aij (·, t)∂j ρk ), ρ (·))L2 ()

+ ni=1 (bi (·, t)∂i ρk , ρ )L2 () + (c(·, t)ρk , ρ )L2 ()
⎪
⎪  
⎪
⎪ = − ni,j =1  aij (x, t)(∂j ρk )(x)(∂i ρ )(x)dx
⎪
⎩ n
+ i=1 (bi (·, t)∂i ρk , ρ )L2 () + (c(·, t)ρk , ρ )L2 () .

By the regularity of F ε and aij , we know that k ∈ C 2 [0, T ] and fε ∈ W 1,1 (0, T ).
Consequently, we can apply Theorem 3.6 (iii) to see that there exists a unique
ε
solution pN, ∈ W 1,1 (0, T ) to (4.20) and pN,
ε
(0) = c for  = 1, . . . , N.
Next we estimate uN Since pN,k ∈ W (0, T ) and pN,k
ε ε 1,1 ε
(0) = ck , we see that
pN,k − ck ∈ Hα (0, T ) ∩ 0 W (0, T ) and so (2.27) yields
ε 1,1

∂tα (pN,k
ε
− ck ) = dtα (pN,k
ε
− ck ) = dtα pN,k
ε
, N ∈ N, 1 ≤ k ≤ N. (4.21)

Therefore,


N 
N
∂tα (uεN − aN ) = ∂tα (pN,k
ε
− ck )ρk = (dtα pN,k
ε
)ρk = dtα uεN . (4.22)
k=1 k=1

By (4.19) we have

(t − s)α−1 (dtα uεN (·, s), uεN (·, s))L2 ()


n
− (t − s)α−1 (∂i (aij (·, s)∂j uεN (·, s)), uεN (·, s))L2 ()
i,j =1


n
− (t − s)α−1 (bj (·, s)∂j uεN (·, s), uεN (·, s))L2 ()
j =1

− (t − s)α−1 (c(·, s)uεN (·, s), uεN (·, s))L2 ()

= H −1 () FNε (·, s), uεN (·, s)H 1 () (t − s)α−1 . (4.23)
0

By (4.2) we obtain


n
− (t − s)α−1 (∂i (aij (·, s)∂j uεN (·, s)), uεN (·, s))L2 ()
i,j =1


n 
= (t − s)α−1 aij (x, s)(∂j uεN (x, s))∂i uεN (x, s))dx
i,j =1 

≥ C0 (t − s)α−1 ∇uεN (·, s) 2

L2 ()
. (4.24)
4.3 Proof of Theorem 4.1 83

Moreover, fixing small δ > 0, we can choose a constant Cδ > 0 such that we have
 
    
 n  n
− (b (·, s)∂ u ε
(·, s), u ε
(·, s)) 
2 () ≤ C |∂j uεN (x, s)||uεN (x, s)|dx
 j j N N L 
 j =1   j =1

 
≤δ |∇uεN (x, s)|2 dx + Cδ |uεN (x, s)|2 dx. (4.25)
 

Similarly we see

|H −1 () FNε (·, s), uεN (·, s)H 1 () | ≤ δ ∇uεN (·, s) 2
L2 ()
+ Cδ FNε (·, s) 2
H −1 ()
.
0
(4.26)

On the other hand, by Theorem 3.4 (i) and uεN ∈ W 1,1 (0, T ; L2 ()), we obtain
 t (α)
(t − s)α−1 (dtα uεN (·, s), uεN (·, s))L2 () ds ≥ ( uεN (·, t) 2
L2 ()
− aN 2
(L2 ())N
).
0 2

Therefore application of (4.24)–(4.26) in (4.23) yields

 t
2
uεN (·, t ) L2 ()
+ (t − s)α−1 ∇uεN (·, s) 2
L2 ()
ds
0
 t
≤ aN 2
L2 ()
+ Cδ (t − s)α−1 ∇uεN (·, s) 2
L2 ()
ds
0
 t  t
+ Cδ (t − s)α−1 uεN (·, s) 2
L2 ()
ds + Cδ (t − s)α−1 FNε (·, s) H −1 () ds, 0 < t < T.
0 0

Fixing δ > 0 sufficiently small and absorbing the second term on the right-had side
into the left-hand side, we obtain
2
uεN (·, t) L2 ()
 t
≤ uεN (·, t) 2
L2 ()
+ (t − s)α−1 ∇uεN (·, s) 2
L2 ()
ds
0
 t
≤C aN 2L2 () +C (t − s)α−1 uεN (·, s) 2
L2 ()
ds
0
 t
+C (t − s)α−1 FNε (·, s) 2
H −1 ()
ds. (4.27)
0
84 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

The generalized Gronwall inequality Lemma A.2 in the Appendix, implies

2
uεN (·, t) L2 ()
 t
≤C aN 2L2 () +C (t − s)α−1 FNε (·, s) 2
H −1 ()
ds
0
 t  s 
+C (t − s)α−1 (s − ξ )α−1 FNε (·, ξ ) 2
H −1 ()
dξ ds, 0 < t < T.
0 0

Here
 t  s 
(t − s) α−1
(s − ξ ) α−1
FNε (·, ξ ) 2H −1 () dξ ds
0 0
 t  t 
= FNε (·, ξ ) 2
H −1 ()
(t − s)α−1 (s − ξ )α−1 ds dξ
0 ξ

(α)2 t
= (t − ξ )2α−1 FNε (·, ξ ) 2
H −1 ()
dξ.
(2α) 0

Since

(t − s)2α−1 ≤ C(t − s)α−1 , 0 ≤ s ≤ t ≤ T,

we have
 t
uεN (·, t) 2L2 () ≤C aN 2L2 () +C (t − s)α−1 FNε (·, s) 2
H −1 ()
ds.
0
T
Taking 0 · · · dt and applying the Young inequality on the convolution (Lemma A.1
in the Appendix), we have
2 2 2
uεN (·, t) L2 (0,T ;L2 ())
≤ C( aN L2 ()
+ FNε L2 (0,T ;H −1 ())
). (4.28)

By (4.22), applying Theorem 3.3 (i) we obtain

 T  T
(∂tα (uεN − aN )(·, t), uεN (·, t))L2 () dt = (dtα uεN (·, t), uεN (·, t))L2 () dt
0 0
 T 
≥ C0 uεN (·, t) 2L2 () dt − aN 2L2 () . (4.29)
0
4.3 Proof of Theorem 4.1 85

On the other hand, estimating the second term on the left-hand side and the right-
hand side of (4.19) by (4.29) and (4.25)–(4.26), we have
 T  T
uεN (·, t) 2L2 () dt + ∇uεN (·, t) 2
L2 ()
dt
0 0
 T  T
≤ Cδ ∇uεN (·, t) 2
L2 ()
dt + Cδ uεN (·, t) 2
L2 ()
dt
0 0
 T
+ Cδ FNε (·, t) 2
H −1 ()
dt + C aN 2
L2 ()
.
0

Therefore, fixing δ > 0 small and absorbing the first term on the right-hand side
into the left-hand side, we reach

∇uεN 2
L2 (0,T ;L2 ())
≤ C( aN 2
L2 ()
+ uεN 2
L2 (0,T ;L2 ())
+ FNε 2
L2 (0,T ;H −1 ())
).

Applying (4.28), we obtain

2 2 2
∇uεN L2 (0,T ;L2 ())
≤ C( aN L2 ()
+ FNε L2 (0,T ;H −1 ())
). (4.30)

Here C > 0 is independent of N ∈ N and ε > 0.

Now we estimate ∂tα (uεN − aN ) L2 (0,T ;H −1 ()). Since


N
∂tα (uεN − aN ) = ∂tα (pN,k
ε
− ck )ρk ,
k=1

we have


N
H −1 () ∂tα (uεN − aN ), ρ H 1 () = ∂tα (pN,k
ε
(t) − ck )H −1 () ρk , ρ H 1 () = 0
0 0
k=1

N
for  ≥ N + 1. For any ψ ∈ H01 (), we set ψN = =1 H −1 () ψ, ρ H01 () ρ .
Therefore (4.17) yields

H −1 () ∂t (uN − aN ), ψH 1 ()

α ε
0

= − H −1 () A(t)uεN , ψN H 1 () + H −1 () FNε , ψN H 1 () , ψ ∈ H01 ().

0 0
86 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

Hence

∂tα (uεN − aN )(·, t ) H −1 () = sup |H −1 () ∂tα (uεN − aN )(·, t ), ψH 1 () |
0
ψ∈H01 (), ψ H01 ()
=1

≤ sup |H −1 () A(t )uεN , ψN H 1 () |

0
ψ∈H01 (), ψ H01 ()
=1

+ sup |H −1 () FNε (·, t ), ψN H 1 () |

0
ψ∈H01 (), ψ H01 ()
=1

≤C sup |(∇uεN (·, t ), ∇ψN )L2 () |

ψ∈H01 (), ψ H01 ()
=1

+ sup FNε (·, t ) H −1 () ∇ψN L2 ()

ψ∈H01 (), ψ H01 ()
=1

≤ C( uεN (·, t ) L2 () + FNε (·, t ) H −1 () ).

Here we used also ψN H 1 () ≤ ψ H 1 () ≤ 1.

0 0
Since aN L2 () ≤ a L2 () and FNε L2 (0,T ;H −1 ()) ≤ F ε L2 (0,T ;H −1 ()) ,
combining (4.28) and (4.30), we obtain

∂tα (uεN − aN ) 2
L2 (0,T ;H −1 ())
+ uεN 2
L2 (0,T ;H01 ())

≤ C( a 2
L2 ()
+ Fε 2
L2 (0,T ;H −1 ())
). (4.31)

The sequences {uεN − aN }N∈N and {uεN }N∈N are bounded in H α (0, T ; H −1 ())
and in L2 (0, T ; H01()) respectively. Therefore we can extract a subsequence N 
of N ∈ N and uε ∈ L2 (0, T ; H01 ()) and v ε ∈ Hα (0, T ; H −1 ()) such that
uεN  −→ uε weakly in L2 (0, T ; H01 ()) and uεN  − aN  −→ v ε weakly in
Hα (0, T ; H −1 ()). Since aN  −→ a strongly in L2 (), we have uεN  −→ a + v ε
weakly in Hα (0, T ; H −1 ()). Therefore uεN  −→ uε and uεN  −→ a + v ε in the
sense of distribution, that is, in (C0∞ ( × (0, T ))) . Hence uε = a + v ε , that is,
v ε = uε − a. Therefore by (4.31) we have

uε − a 2
Hα (0,T ;H −1 ())
+ uε 2
L2 (0,T ;H01 ())
(4.32)

2 2
≤ lim

inf uεN  − aN  H α (0,T ;H −1 ())
+ uεN  L2 (0,T ;H01 ())
N →∞

≤ C( F ε 2
L2 (0,T ;H −1 ())
+ a 2
L2 ()
).
4.3 Proof of Theorem 4.1 87

Henceforth by XN we denote the closed subspace in H01 () spanned by ρ1 , . . . , ρN .

For any ψ ∈ XN , it follows from (4.17) that
 T  T
H −1 () ∂tα (uεN − aN ), ψH 1 () dt + H −1 () A(t)uN , ψH01 () dt
ε
0
0 0
 T
= H −1 () FN (·, t), ψH01 () dt,
ε
N ∈ N.
0

Consequently, letting N → ∞, we obtain

 T  T
H −1 () ∂t (u − a), ψH 1 () dt + H −1 () A(t)u, ψH01 () dt
α ε
0
0 0
 T
= H −1 () F
ε
(·, t), ψH 1 () dt. (4.33)
0
0

Since ψN ∈ XN and N ∈ N are chosen arbitrarily, it follows that (4.33) holds for
each ψ ∈ H01 (), so that

∂tα (uε − a) + A(t)uε = F ε in H −1 (), 0 < t < T .

By (4.32) we choose a sequence {εn }n∈N satisfying limn→∞ εn = 0 and u ∈

L2 (0, T ; H01 ()) such that ∂tα (uεn −a) −→ ∂tα (u−a) weakly in L2 (0, T ; H −1 ())
and uεn −→ u weakly in L2 (0, T ; H01()) as n → ∞. Since Fεn −→ F in
L2 (0, T ; H −1 ()) as n → ∞ by (4.15), we can verify (4.7) and

∂tα (u − a)(·, t) + A(t)u(·, t) = F (·, t) in H −1 (), 0 < t < T .

Finally, we have to prove the uniqueness of the solution to (4.3)–(4.5). To this

end, it is necessary to verify that if u ∈ Hα (0, T ; H −1 ()) ∩ L2 (0, T ; H01())
satisfies

n 
n
∂tα u − ∂i (aij (x, t)∂j u) = bj (x, t)∂j u + c(x, t)u in H −1 (), (4.34)
i,j =1 j =1

then u = 0 in  × (0, T ).
From (4.34) and integration by parts in view of u(·, t) ∈ H01 (), we have
⎛ ⎞

n
H −1 () ∂t u(·, s), u(·, s)H01 ()
α
+⎝ aij (·, s)∂j u(·, s), ∂i u(·, s)⎠
i,j =1 L2 ()


n
= H −1 ()  bj (·, s)∂j u(·, s), u(·, s)H 1 () + H −1 () c(·, s)u, uH 1 () , 0 < s < T.
0 0
j =1
88 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

Therefore
 t
H −1 () ∂t u(·, s), u(·, s)H01 () (t − s)α−1 ds
α
0
⎛ ⎞
 t 
n
+ (t − s)α−1 ⎝ aij (x, s)(∂j u(x, s))∂i u(x, s)dx ⎠ ds
0 i,j =1
⎛ ⎞
 t  
n
= (t − s)α−1 ⎝ bj (x, s)(∂j u(x, s))u(x, s)dx ⎠ ds
0  j =1
 t  
2
+ (t − s) α−1
c(x, s)u (x, s)dx ds.
0 

Applying (4.2) and (4.25), we derive

 t
H −1 () ∂t u(·, s), u(·, s)H01 () (t − s)α−1 ds
α
0
 t
+ (t − s)α−1 ∇u(·, s) 2
L2 ()
ds
0
 t  t
≤ Cδ (t − s)α−1 ∇u(·, s) 2
L2 ()
ds + Cδ (t − s)α−1 u(·, s) 2
L2 ()
ds.
0 0

By u ∈ Hα (0, T ; H −1 ()) ∩ L2 (0, T ; H01 ()), applying Theorem 3.4 (ii) to the
first term on the left-hand side and choosing δ > 0 sufficiently small, we have
 t
(α) (α)
u(·, t) 2L2 () ≤ u(·, t) 2L2 () + (1 − Cδ) (t − s)α−1 ∇u(·, s) 2L2 () ds
2 2 0
 t
≤ Cδ (t − s)α−1 u(·, s) 2L2 () ds, 0 < t < T .
0

The generalized Gronwall inequality Lemma A.2 in the Appendix implies u = 0 in

(0, T ). Thus the proof of Theorem 4.1 is complete.

4.4 Proof of Theorem 4.2

For F ∈ L2 (0, T ; L2 ()) and ε > 0, let F ε ∈ C0∞ (0, T ; L2 ()) satisfy F ε −
F L2 (0,T ;L2 ()) < ε. In terms of Theorem 3.6 (iii) and (v), we can argue similarly
ε
to the proof of Theorem 4.1 to construct an approximating sequence, that is, pN,k ∈
4.4 Proof of Theorem 4.2 89

ε (0) = c := (a, ρ )
(Wα (0, T ))N and pN,k k k L2 () for 1 ≤ k ≤ N. Here we have to
estimate


N
uεN (x, t) := ε
pN,k (t)ρk (x) in L2 (0, T ; H 2())
k=1


and ∂tα (uεN − aN ) in L2 (0, T : L2 ()), where aN = N k=1 ck ρk .
ε
Using (4.21), multiplying the first equation in (4.20) by dtα pN, , summing over
 = 1, . . . , N, we obtain


n
(dtα uεN (·, t), dtα uεN (·, t))L2 () − (∂i (aij (·, t)∂j uεN (·, t)), dtα uεN (·, t))L2 ()
i,j =1


n
− (bj (·, t)∂j uεN (·, t), dtα uεN (·, t))L2 () − (c(·, t)uεN (·, t), dtα uεN (·, t))L2 ()
j =1

= (FNε (·, t), dtα uεN (·, t))L2 () . (4.35)

The main part of the proof is the estimation of the second term on the left-hand
side. By uεN (·, t) ∈ H01 (), integration by parts yields
⎛ ⎞

n
−⎝ ∂i (aij (·, t)∂j uεN (·, t)), dtα uεN (·, t)⎠
i,j =1 L2 ()
⎛ ⎞

n
=⎝ aij (·, t)∂j uεN (·, t), dtα ∂i uεN (·, t)⎠ .
i,j =1 L2 ()
n
Now we consider ε α ε
i,j =1 aij (∂j uN )∂i dt uN (·, t). Since aij = aj i , 1 ≤ i, j ≤ n,
we have


n
aij (∂j uεN )∂i dtα uεN (·, t)
i,j =1

 
n
= aij ((∂j uεN )∂i dtα uεN + (∂i uεN )∂j dtα uεN ) + aii (∂i uεN )(∂i dtα uεN ) =: J1 + J2 .
i>j i=1
90 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

We set g(t) = 1 −α . Here we write uεN (s) = uεN (x, s), etc. Then
(1−α) t


J1 = aij ((∂j uεN )∂i dtα uεN + (∂i uεN )∂j dtα uεN )(x, t)
i>j
  t
= aij (x, t) g(t − s)((∂j uεN )(t)∂i ∂s uεN (x, s) + (∂i uεN )(t)∂j ∂s uεN (x, s))ds
i>j 0

  t
=− aij (x, t) g(t − s)((∂i ∂s uεN (s))(∂j uεN (s) − ∂j uεN (t))
i>j 0

+ (∂j ∂s uεN (s))(∂i uεN (s) − ∂i uεN (t))ds

  t
+ aij (x, t) g(t − s)((∂i ∂s uεN (s))∂j uεN (s) + (∂j ∂s uεN (s))∂i uεN (s))ds
i>j 0

=: J11 + J12 .

We have
  t
J11 = − aij (x, t) g(t − s){∂s (∂i uεN (s) − ∂i uεN (t))(∂j uεN (s) − ∂j uεN (t))
i>j 0

+ ∂s (∂j uεN (s) − ∂j uεN (t))(∂i uεN (s) − ∂i uεN (t))}ds

  t
=− aij (x, t) g(t − s)∂s ((∂i uεN (s) − ∂i uεN (t))(∂j uεN (s) − ∂j uεN (t)))ds
i>j 0

=− aij (x, t)[g(t − s)(∂i uεN (s) − ∂i uεN (t))(∂j uεN (s) − ∂j uεN (t))]s=t
s=0
i>j
  t dg
− aij (x, t) (t − s)(∂i uεN (s) − ∂i uεN (t))(∂j uεN (s) − ∂j uεN (t))ds
0 dξ
i>j

= aij (x, t)g(t)(∂i uεN (0) − ∂i uεN (t))(∂j uεN (0) − ∂j uεN (t))
i>j
 t 
dg
+ − (t − s) aij (x, t)(∂i uεN (s) − ∂i uεN (t))(∂j uεN (s) − ∂j uεN (t))ds.
0 dξ
i>j

Here we used

lim g(t − s)(∂i uεN (s) − ∂i uεN (t))(∂j uεN (s) − ∂j uεN (t)) = 0,
s↑t

which can be justified by ∂i uεN (x, ·), ∂j uεN (x, ·) ∈ Wα (0, T ) following from
Theorem 3.6 (v).
4.4 Proof of Theorem 4.2 91

On the other hand,

  t
J12 = aij (x, t) g(t − s)∂s (∂i uεN (x, s)∂j uεN (x, s))ds
i>j 0

= aij (x, t)dtα (∂i uεN ∂j uεN )(x, t).
i>j

Consequently,

1
J1 = aij g(t)(∂i uεN (0) − ∂i uεN (t))(∂j uεN (0) − ∂j uεN (t))
2
i=j
 t 
1 dg
+ − (t − s) aij (x, t)(∂i uεN (s) − ∂i uεN (t))(∂j uεN (s) − ∂j uεN (t))ds
2 0 dξ
i=j

1
+ aij dtα (∂i uεN ∂j uεN )(x, t).
2
i=j

Similarly, we can calculate

  t
J2 = aii g(t − s)∂i ∂s uεN (s)∂i uεN (t)ds
i 0

  t
=− aii g(t − s)∂i ∂s uεN (s)(∂i uεN (s) − ∂i uεN (t))ds
i 0

  t
+ aii g(t − s)∂i ∂s uεN (s)∂i uεN (s)ds
i 0
 t
1
=− aii g(t − s)∂s ((∂i uεN (s) − ∂i uεN (t))(∂i uεN (s) − ∂i uεN (t)))ds
2 0 i

1 t
+ aii g(t − s)∂s (|∂i uεN (s)|2 )ds
2 0
i

1 1 t dg 
= aii g(t)(∂i uεN (t) − ∂i uεN (0))2 + − (t − s) aii (∂i uεN (s) − ∂i uεN (t))2 ds
2 2 0 dξ
i i

1
+ aii dtα (|∂i uεN (t)|2 ).
2
i
92 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

Thus


n
aij ∂j uεN (x, t)∂i dtα uεN (x, t)
i,j =1

1 
n
= g(t)aij (x, t)(∂i uεN (0) − ∂i uεN (t))(∂j uεN (0) − ∂j uεN (t))
2
i,j =1
 t 
n
1 dg
+ − (t − s) aij (x, t)(∂i uεN (s) − ∂i uεN (t))(∂j uεN (s) − ∂j uεN (t))ds
2 0 dξ
i,j =1

1 
n
+ aij dtα ((∂i uεN )∂j uεN )(x, t).
2
i,j =1

By g(t) > 0, − dg
dξ (t − s) > 0 for 0 < s < t < T and (4.2), we obtain


n 
n
2 aij (x, t)∂j uεN (x, t)∂i dtα uεN (x, t) ≥ aij (x, t)dtα ((∂i uεN )∂j uεN )(x, t)
i,j =1 i,j =1
 t 
n
= g(t − s) ∂s (aij (x, s)∂i uεN (s)∂j uεN (s))ds
0 i,j =1
 t 
n
− g(t − s) (∂s aij (x, s))∂i uεN (s)∂j uεN (s)ds
0 i,j =1
 t 
n
+ g(t − s) (aij (x, t) − aij (x, s))∂s ((∂i uεN (s))∂j uεN (s))ds =: S1 + S2 + S3 .
0 i,j =1
(4.36)

Here we have
 t 
n
S1 = g(t − s) ∂s (aij (x, s)∂i uεN (s)∂j uεN (s))ds (4.37)
0 i,j =1
⎛ ⎞
 t 
n
= g(s)∂t ⎝ aij (x, t − s)∂i uεN (t − s)∂j uεN (t − s)⎠ ds
0 i,j =1
 t 
n
= ∂t g(s) aij (x, t − s)∂i uεN (t − s)∂j uεN (t − s)ds
0 i,j =1


n
− g(t) aij (x, 0)∂i uεN (x, 0)∂j uεN (x, 0).
i,j =1
4.4 Proof of Theorem 4.2 93

Furthermore
 t 
n
S2 = − g(t − s) (∂s aij (x, s))∂i uεN (s)∂j uεN (s)ds
0 i,j =1

and
 t 
n
S3 = rij (x, t, s)∂s (∂i uεN (s)∂j uεN (s))ds,
0 i,j =1

where

(t − s)−α
rij (x, t, s) = (aij (x, t) − aij (x, s)), 1 ≤ i, j ≤ n.
(1 − α)

We directly obtain
 t
|S2 | ≤ C (t − s)−α |∇uεN (·, s)|2 ds. (4.38)
0

To estimate |S3 |, we integrate by parts and apply

 
 α(t − s)−α−1 (t − s)−α 

|∂s rij (x, t, s)| =  (aij (x, t) − aij (x, s)) − ∂s aij (x, s)
(1 − α) (1 − α)
≤C|t − s|−α

by aij ∈ C 1 ([0, T ]; C()), so that

⎡ ⎤s=t

  n
|S3 | =  ⎣ rij (x, t, s)∂i uεN (s)∂j uεN (s)⎦
i,j =1 s=0
  
t n

− (∂s rij )(x, t, s)∂i uN (s)∂j uN (s)ds 
ε ε
0 i,j =1
 t
≤ C|∇uεN (x, 0)|2 + C (t − s)−α |∇uεN (x, s)|2 ds.
0
94 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

Therefore (4.36)–(4.38) imply


n
2 aij (x, t)∂j uεN (x, t)∂i dtα uεN (x, t)
i,j =1
 t 
n
≥ ∂t g(s) aij (x, t − s)∂i uεN (x, t − s)∂j uεN (x, t − s)ds
0 i,j =1


n
− g(t) aij (x, 0)∂i uεN (x, 0)∂j uεN (x, 0)
i,j =1
 t
− C|∇uεN (x, 0)|2 −C (t − s)−α |∇uεN (x, s)|2 ds, x ∈ , 0 < t < T .
0
(4.39)
T 
Taking 0  · · · dxdt and applying the Young inequality (Lemma A.1 in the
Appendix) to the fourth term on the right-hand side of (4.39), we have
 T 
n
2 (aij (·, t)∂j uεN (·, t), dtα ∂i uεN (·, t))L2 () dt
0 i,j =1
 T 
n
≥ g(s) (aij (T − s)∂i uεN (T − s), ∂j uεN (T − s))L2 () ds
0 i,j =1
 T
−C g(t)dt ∇a 2
L2 ()
− C( ∇a L2 () + ∇uεN 2
L2 (0,T ;L2 ())
).
0

Since
 T 
n
g(s) (aij (·, T − s)∂i uεN (·, T − s), ∂j uεN (·, T − s))L2 () ds
0 i,j =1
 T 
n
= g(T − s) (aij (·, s)∂i uεN (·, s), ∂j uεN (·, s))L2 () ds
0 i,j =1
−α
≥ CT ∇uεN 2L2 (0,T ;L2 ())
4.4 Proof of Theorem 4.2 95

by (4.2), we obtain
 T
[the second term on the left-hand side of (4.35)]dt (4.40)
0
⎛ ⎞
 T 
n
= ⎝ aij (·, t)∂j uεN (·, t), dtα ∂i uεN (·, t)⎠ dt
0 i,j =1 L2 ()

≥ − C( ∇uεN 2
L2 (0,T ;L2 ())
+ ∇a 2
L2 ()
).

By the Poincaré inequality and uεN (·, t) ∈ H01 (), for small δ > 0, we can
estimate
 T
|[the third and the fourth terms on the left-hand side of (4.35)]dt
0

 T
≤ (Cδ ∇uεN (·, t) 2
L2 ()
+ δ dtα uεN (·, t) 2
L2 ()
)dt (4.41)
0

and
 
 T 
 (FNε (·, t), dtα uεN (·, t))L2 () dt 

0

≤ Cδ FNε 2
L2 (0,T ;L2 ())
+ Cδ dtα uεN (·, t) 2
L2 (0,T ;L2 ())
. (4.42)

Hence, applying (4.40)–(4.42) in (4.35), we derive

2 2 2
dtα uεN L2 (0,T ;L2 ())
≤ C ∇a L2 ()
+ Cδ ∇uεN L2 (0,T ;L2 ())
2 2
+ δ dtα uεN L2 (0,T ;L2 ())
+ Cδ FNε L2 (0,T ;L2 ())
.

Fixing δ > 0 sufficiently small and absorbing the third term on the right-hand
side into the left-hand side and applying (4.30), which was verified in the proof
of Theorem 4.1, we obtain

∂tα (uεN − a) 2
L2 (0,T ;L2 ())
= dtα uεN 2
L2 (0,T ;L2 ())

≤ C( ∇aN 2
L2 ()
+ FNε 2
L2 (0,T ;L2 ())
), (4.43)

where the constant C > 0 is independent of N ∈ N and ε > 0.

Based on the uniform bound (4.43) in N and ε, we can argue similarly to (4.31)–
(4.33). We omit the details, so that we can complete the proof of Theorem 4.2.
96 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

We note that the uniqueness of solutions is proved for less regular class, and the
uniqueness holds also within the stronger regularity of solutions in Theorem 4.2.

4.5 Proofs of Propositions 4.1 and 4.2

4.5.1 Proof of Proposition 4.1

As is already proved, u ∈ Hα (0, T ; H −1 ()) ∩ L2 (0, T ; H01 ()) satisfies

∂tα (u − a) + A(t)u = F in H −1 (), 0 < t < T ,

u(·, t) ∈ H01 (), u − a ∈ Hα (0, T ; H −1 ()).

Then we have

H −1 () ∂s (u − a)(·, s), u − aH 1 () + H −1 () A(s)(u − a)(·, s), u − aH 1 ()
α
0 0

+ H −1 () A(s)a, u − aH 1 () = H −1 () F (·, s), (u − a)(·, s)H 1 ()
0 0

for 0 < s < T . Here we note that for any ε > 0 there exists a constant Cε > 0 such
that
⎧
⎪
⎪ |H −1 () A(s)a, u(·, s) − aH 1 () | ≤ C a H 1 () (u − a)(·, s) H 1 ()
⎪
⎪ 0 0 0
⎪
⎪ ≤ ε (u − a)(·, s) 2H 1 () + Cε a 2H 1 () ,
⎪
⎪  
⎪
⎨ n
⎪ 0 0

j =1 H −1 () bj (·, s)∂j (u(·, s) − a), u(·, s) − aH01 () 
⎪
⎪ + |H −1 () c(·, s)(u(·, s) − a), u(·, s) − aH 1 () |
⎪
⎪
⎪
⎪
0
⎪ ≤ C u(·, s) − a L2 () u(·, s) − a H01 ()
⎪
⎪
⎪
⎩ ≤ ε u(·, s) − a 2 1 + Cε u(·, s)0 − a 2 2 .
H () 0
L ()
(4.44)

Here we used

bj (·, s)∂j (u(·, s) − a) H −1 () = sup |H −1 () bj (·, s)∂j (u(·, s) − a), ψH 1 () |
0
ψ H 1 ()
=1
0

= sup |(u(·, s) − a, ∂j (bj ψ))L2 () |

ψ H01 ()
=1

≤ sup u(·, s) − a L2 () ∂j (bj ψ) L2 () ≤ C u(·, s) − a L2 () .

ψ H01 ()
=1
4.5 Proofs of Propositions 4.1 and 4.2 97

Multiplying (t −s)
α−1
(α) and integrating in s over (0, t) and applying (4.2) and
Theorem 3.4 (ii), we obtain
 t
u(·, t) − a 2
L2 ()
+ (t − s)α−1 ∇(u(·, s) − a) 2
L2 ()
ds
0
 t
≤C (t − s)α−1 |H −1 () A(s)a, u(·, s) − aH 1 () |ds
0
0
 
 t  
 n

+  (t − s)α−1 −1
H () b j (·, s)∂j (u(·, s) − a), u(·, s) − a 1 ds
H0 () 

 0 j =1 
 t 
 
+  H −1 () c(·, s)(u(·, s) − a), u(·, s) − aH 1 () (t − s)α−1 ds 
0
0
 t
+C (t − s)α−1 F (·, s) H −1 () u(·, s) − a H 1 () ds.
0
0

Applying (4.44) and

F (·, s) H −1 () u(·, s) − a H01 () ≤ ε u(·, s) − a 2

H01 ()
+ Cε F (·, s) 2
H −1 ()
,

we obtain
 t
u(·, t) − a 2
L2 ()
+ (t − s)α−1 u(·, s) − a 2
H01 ()
ds
0
 t  t
≤ε (t − s) α−1
u(·, s) − a 2
H01 ()
ds + Cε (t − s)α−1 u(·, s) − a 2
L2 ()
ds
0 0
 t  t
+ Cε (t − s) α−1
a 2
H01 ()
ds + Cε (t − s)α−1 F (·, s) 2
H −1 ()
ds.
0 0

Choosing ε > 0 sufficiently small, we absorb the first term on the right-hand side
into the left-hand side, so that

u(·, t) − a 2
L2 ()
 t
2 2
≤ u(·, t) − a L2 ()
+C (t − s)α−1 u(·, s) − a H01 ()
ds
0
 t
2
≤C (t − s)α−1 u(·, s) − a L2 ()
ds
0
 t
+ Ct α a 2
H01 ()
+C (t − s)α−1 F (·, s) 2
H −1 ()
ds.
0
98 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

Since F ∈ Lp (0, T ; H −1 ()) with p > α2 , we can choose q > 1 satisfying 1

p + q1 =
2
p p
1 and q(α −1) > −1. Indeed, q = p−2 and p > 2
α imply q(α −1) = p−2 (α −1) >
−1. Therefore the Hölder inequality yields

 t  t  q1  t  p2
p
(t − s) α−1
F (·, s) 2
H −1 ()
ds ≤ (t − s)qα−q
ds ( F (·, s) 2
H −1 ()
) 2 ds
0 0 0

θ
≤ Ct q F 2
Lp (0,t;H −1 ())
,

where θ = q(α − 1) + 1 > 0. Thus

θ
2 2 2
u(·, t) − a L2 ()
≤ C(t α a H01 ()
+tq F Lp (0,T ;H −1 ())
)
 t
2
+C (t − s)α−1 u(·, s) − a L2 ()
ds, 0 < t < T.
0

Lemma A.2 in the Appendix implies

θ
α+ qθ
u(·, t) − a 2
L2 ()
≤ C((t α + t 2α ) a 2
H01 ()
+ (t q + t ) F 2
Lp (0,T ;H −1 ())
),

which completes the proof of Proposition 4.1.

4.5.2 Proof of Proposition 4.2

The proof is based on an estimate similar to Proposition 4.1.

By (4.3), we have

∂sα (u − a)(·, s) + A(s)u(x, s) = 0 in L2 (), 0 < s < t.

Multiplying both sides by (u − a)(x, s)(t − s)α−1 and integrating over  × (0, t),
we obtain
 t  
(∂s (u − a)(x, s))(u − a)(x, s)dx (t − s)α−1 ds
α
0 
⎛ ⎞
 t  
n
− ⎝ ∂i (aij (x, s)∂j u(x, s))u(x, s)dx ⎠ (t − s)α−1 ds
0  i,j =1
⎛ ⎞
 t  
n
+ ⎝ ∂i (aij (x, s)∂j u(x, s))a(x)dx ⎠ (t − s)α−1 ds
0  i,j =1
4.5 Proofs of Propositions 4.1 and 4.2 99

⎛ ⎞
 t  
n
− ⎝ bj (x, s)∂j u(x, s)(u(x, s) − a(x))dx ⎠ (t − s)α−1 ds
0  j =1
 t  
− c(x, s)u(x, s)(u(x, s) − a(x))dx (t − s)α−1 ds = 0.
0 

Since (u − a)(x, ·) ∈ Hα (0, T ) for x ∈  and u(·, s) ∈ H01 () for 0 < s < t, we
integrate by parts and use Theorem 3.4 (ii) where we note

H −1 () ∂s (u(·, s)−a), u(·, s)−aH 1 () =
α
(∂sα (u(x, s)−a(x))(u(x, s)−a(x))dx,
0


by ∂sα (u − a) ∈ L2 (0, T ; L2 ()), so that

⎛ ⎞
 t  
n
C0 u(·, t) − a 2
+ ⎝ aij (x, s)∂j u(x, s)∂i u(x, s)dx ⎠ (t − s)α−1 ds
L2 ()
0  i,j =1
⎛ ⎞
 t  
n
− ⎝ aij (x, s)∂j u(x, s)∂i a(x)dx ⎠ (t − s)α−1 ds
0  i,j =1
⎛ ⎞
 t  
n
− ⎝ bj (x, s)∂j u(x, s)u(x, s)dx ⎠ (t − s)α−1 ds
0  j =1
⎛ ⎞
 t  
n
+ ⎝ bj (x, s)∂j u(x, s)a(x)dx ⎠ (t − s)α−1 ds
0  j =1
 t  
− c(x, s)u(x, s)2 dx (t − s)α−1 ds
0 
 t  
+ c(x, s)u(x, s)a(x)dx (t − s)α−1 ds = 0.
0 

By (4.2) and using |u||a| ≤ 12 (|u|2 + |a|2), we obtain

 t  
C0 u(·, t) 2
L2 ()
− C1 a 2
L2 ()
+ C2 |∇u(x, t)|2 (t − s)α−1 dx ds
0 
 t  
α−1 1 α−1
− C3 ε|∇u(x, s)|(t − s) 2 |∇a(x)|(t − s) 2 dxds
0  ε
 t  
α−1 1 α−1
− C3 ε|∇u(x, s)|(t − s) 2 |u(x, s)|(t − s) 2 dxds
0  ε
100 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

 t  
1 α−1 α−1
− C3 ε|∇u(x, s)|(t − s) |a(x)|(t − s) 2 dxds 2
0  ε
 t  t
− C3 |u(x, s)| (t − s) dxds − C3
2 α−1
|a(x)|2(t − s)α−1 dxds ≤ 0.
0  0 

Hence, after applying the Cauchy–Schwarz

t  inequality, choosing ε > 0 suffi-
ciently small and absorbing 0  C32 ε2 |∇u(x, s)|2 (t − s)α−1 dxds into the term
t 
C2 0  |∇u(x, s)|2 (t − s)α−1 dxds, we obtain
 t
u(·, t) 2
L2 ()
+ C4 |∇u(x, t)|2 (t − s)α−1 dxds
0 
 t  
≤C (|∇a(x)| + |a(x)| )dx (t − s)α−1 dxds
2 2
0 
 t
+C |u(x, s)|2 (t − s)α−1 dxds, t > 0.
0 

Therefore
 t
u(·, t) 2
L2 ()
+ u(·, s) 2
H01 ()
(t − s)α−1 ds
0
 t
≤ Ct α a 2
H01 ()
+C u(·, s) 2
L2 ()
(t − s)α−1 ds, t > 0,
0

that is,
 t
u(·, t) 2
L2 ()
≤ Ct α a 2
H01 ()
+C u(·, s) 2
L2 ()
(t − s)α−1 ds, t > 0.
0

Consequently, noting that the constant C > 0 can be chosen uniformly in all t > 0
and applying Lemma A.2 in the Appendix, we can take a constant C0 > 0 such that
 t 
2 2 2
u(·, t) L2 ()
≤ C0 t α
a H01 ()
+ C0 e C0 t
(t − s) α−1 α
s ds a H01 ()
0
 
(α)(α + 1) 2α
≤ C0 t α + C0 eC0 t t a 2
(2α + 1) H01 ()

for all t ≥ 0. Thus the proof of Proposition 4.2 is complete.

4.6 Case Where the Coefficients of A(t) Are Independent of Time 101

4.6 Case Where the Coefficients of A(t) Are Independent

of Time

In the case where the coefficients of A(t) are independent of t, we can represent the
solution to the initial boundary value problem by the Mittag-Lefller functions (e.g.,
Sakamoto and Yamamoto [26]). In this section, we show that such a represented
solution coincides with the solution established by Theorem 4.1.
Let


n
−Av(x) = ∂i (aij (x)∂j v(x)) + c(x)v(x), x ∈ ,
i,j =1

where aij = aj i ∈ C 1 (), 1 ≤ i, j ≤ n, c ∈ L∞ (), c ≤ 0 in , and (4.2) is

assumed to hold. We consider the operator A with the domain D(A) = H 2 () ∩
H01 (). Then it is known that there exist eigenvalues of A which can be numbered
with their multiplicities

0 < λ1 ≤ λ2 ≤ · · · −→ ∞.

Henceforth if there is no fear of confusion, then we write the scalar product (f, g)
for f, g ∈ L2 (), in place of (f, g)L2 () .
Let ϕk , k ∈ N be an eigenfunction of A for λn such that

1, k = ,
(ϕk , ϕ ) := ϕk (x)ϕ(x)dx =
 0, k = .

Moreover, it is known that ϕk , k ∈ N, is an orthonormal basis in L2 ().

We can define a fractional power Aγ for γ ∈ R (e.g., Pazy [23]), and for a ∈
D(Aγ ), we have
∞
 γ
Aγ a = λk (a, ϕk )ϕk
k=1

where the series is convergent in L2 (), and

+ ∞
,1
 2γ
2
γ
A a L2 () = λk (a, ϕk )2 .
k=1
102 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

For a ∈ L2 () and F ∈ L2 (0, T ; H −1()), we consider

∞

v(x, t) = (a, ϕk )Eα,1 (−λk t α )ϕk (x)
k=1
∞ 
 t 
+ H −1 () F (·, s), ϕk H 1 () (t − s) α−1
Eα,α (−λk (t − s) )ds ϕk (x)
α
0
k=1 0

=: v1 (x, t) + v2 (x, t). (4.45)

We will verify:
Theorem 4.3 Let a ∈ L2 () and F ∈ L2 (0, T ; H −1 ()). Then v given by (4.45)
satisfies v − a ∈ Hα (0, T ; H −1 ()), v ∈ L2 (0, T ; H01 ()) and (4.3)–(4.5).
The theorem means that (4.45) coincides with the solution by Theorem 4.1. We
can similarly prove that v defined by (4.45) is the same as u in Theorem 4.2 if
a ∈ H01 () and F ∈ L2 (0, T ; L2 ()), but we omit the proof.
Proof It is sufficient to prove that v − a ∈ Hα (0, T ; H −1 ()) and v satisfies (4.3).

First Step First we note that D(A− 2 ) = H −1 () = (H01 ()) and there exist
1

constants C1 , C2 > 0 such that

1
C1 w H −1 () ≤ A− 2 w L2 () ≤ C2 w H −1 ()

for w ∈ H −1 (), and

∞

A − 12
w 2
= λ−1 2
k (w, ϕk ) .
k=1

∞
Since a = k=1 (a, ϕk )ϕk in L2 (), we have

∞

v1 (x, t) − a(x) = (a, ϕk )(Eα,1 (−λk t α ) − 1)ϕk (x). (4.46)
k=1

Second Step Theorem 2.4 implies

(v1 − a)(x, ·) Hα (0,T ) ∼ ∂tα (v1 − a)(x, ·) L2 (0,T )

for almost all x ∈ . Moreover, since Eα,1 (−λk t α ) − 1 ∈ Hα (0, T ) ∩ 0 W 1,1 (0, T )
by Proposition 2.2, we see that

dtα (Eα,1 (−λk t α ) − 1) = ∂tα (Eα,1 (−λk t α ) − 1),

4.6 Case Where the Coefficients of A(t) Are Independent of Time 103

so that
∞

∂tα (v1 − a)(x, t) = (a, ϕk )∂tα (Eα,1 (−λk t α ) − 1)ϕk (x)
k=1

∞

=− λk (a, ϕk )Eα,1 (−λk t α )ϕk (x). (4.47)
k=1

Therefore Lemma 2.5 (i) implies

∞

∂tα (v1 − a)(·, t) 2
H −1 ()
= (a, ϕk )2 |Eα,1 (−λk t α )|2 λk
k=1
⎛ 1
⎞2
∞
 α
2
λ t 2
≤Ct −α (a, ϕk )2 ⎝ k α ⎠ ≤ Ct −α a 2
,
1 + λk t L2 ()
k=1

that is,
2 2
∂tα (v1 − a) L2 (0,T ;H −1 ())
≤C a L2 ()
. (4.48)

Hence v1 − a ∈ Hα (0, T ; H −1()).

Third Step First we note an important property of Eα,1 , which is a consequence of
Lemma 2.5 (ii) and what is called the complete monotonicity:

d
Eα,1 (−ηα ) ≤ 0, Eα,α (−ηα ) ≥ 0, η ≥ 0, 0<α<1 (4.49)
dη

(e.g., Gorenflo et al. [10], Miller and Samko [22]). Therefore we obtain
 η  η
|s α−1 Eα,α (−λk s α )|ds = s α−1 Eα,α (−λk s α )ds
0 0
 η
1 d 1
=− Eα,1 (−λk s α )ds = (1 − Eα,1 (−λk ηα )). (4.50)
λk 0 ds λk

Here we recall: since Eα,1 (z) is an entire function in z and (αn + 1) = αn(αn),
we have
+∞ , ∞
d d  (−λk )n s nα  (−λk )n nαs nα−1
Eα,1 (−λk s ) =
α
=
ds ds (αn + 1) (αn + 1)
n=0 n=1
∞
 ∞

(−λk )n s nα−1 (−λk )m s mα
= = s α−1 (−λk ) = −λk s α−1 Eα,α (−λk s α ).
(αn) (αm + α)
n=1 m=0
104 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

By F ∈ L2 (0, T ; H −1 ()), we have

F (·, s), ϕk  :=H −1 () F (·, s), ϕk H 1 ()

0
1
− 12
F (·, s), A ϕk ) = λk2 (A− 2 F (·, s), ϕk ) ∈ L2 (0, T ).
1 1
= (A 2

Therefore Proposition 2.3 yields

 t
wk (t) := F (·, s), ϕk (t − s)α−1 Eα,α (−λk (t − s)α )ds ∈ Hα (0, T ),
0

and so (2.26) implies

d 1−α
∂tα wk (t) = (J wk )(t), t > 0. (4.51)
dt
Since Eα,α (η) is an entire function
 in η, using the Lebesgue theorem, we can
exchange 0 0 · · · dξ ds and ∞
t s
m=0 , so that
 t
1
J 1−α wk (t) = (t − s)−α wk (s)ds
(1 − α) 0
 t  s 
1
= (t − s)−α F (·, ξ ), ϕk (s − ξ )α−1 Eα,α (−λk (s − ξ )α )dξ ds
(1 − α) 0 0
 t + ∞
,
1 s  (−λk )m (s − ξ )αm
−α
= (t − s) F (·, ξ ), ϕk (s − ξ )α−1
dξ ds
(1 − α) 0 0 (αm + α)
m=0
∞
   
1 t s (−λk )m (s − ξ )αm+α−1
= (t − s)−α F (·, ξ ), ϕk  dξ ds.
(1 − α) 0 0 (αm + α)
m=0

Moreover, we exchange the orders of the integrals:

 t  s 
1 −α (−λk )m (s − ξ )αm+α−1
(t − s) F (·, ξ ), ϕk  dξ ds
(1 − α) 0 0 (αm + α)
 t  t 
1 (t − s)−α (s − ξ )αm+α−1
= F (·, ξ ), ϕk (−λk ) m
ds dξ
(1 − α) 0 ξ (αm + α)
 t
1 (1 − α)(αm + α) 1
= F (·, ξ ), ϕk (−λk )m (t − ξ )αm dξ.
(1 − α) 0 (αm + α) (αm + 1)
Therefore
∞ 
 t (−λk (t − ξ )α )m
J 1−α wk (t) = F (·, ξ ), ϕk dξ
(αm + 1)
m=0 0
 t
= Eα,1 (−λk (t − ξ )α )F (·, ξ ), ϕk dξ.
0
4.6 Case Where the Coefficients of A(t) Are Independent of Time 105

By (4.51), we have

d 1−α
∂tα wk (t) = J wk (t)
dt
 t d
= F (·, t), ϕk  + (Eα,1 (−λk (t − ξ )α )F (·, ξ ), ϕk dξ
0 dt
 t
= F (·, t ), ϕk −λk (t −s)α−1 Eα,α (−λk (t −s)α )F (·, s), ϕk ds = F (·, t ), ϕk −λk wk (t ).
0

Hence
∞

∂tα v2 (x, t) = F (·, t), ϕk ϕk (x)
k=1
- ∞   .
 t
+ − λk (t − s)α−1 Eα,α (−λk (t − s)α )F (·, s), ϕk ds ϕk =: S1 + S2 .
k=1 0
(4.52)

Therefore
∞
 1
∂tα v2 (·, t) 2
H −1 ()
= |(F (·, t), ϕk )|2 + S2 (·, t) 2
H −1 ()
λk
k=1

≤ F (·, t) 2
H −1 ()
+ S2 (·, t) 2
H −1 ()
.

Next
∞
 1
S2 (·, t) 2
H −1 ()
= (S2 (·, t), ϕk )2
λk
k=1
∞
  t 2
 
= λk  (t − s)α−1 Eα,α (−λk (t − s)α )F (·, s), ϕk ds 
k=1 0

and the Young inequality for the convolution (Lemma A.1) yields
2
S2 L2 (0,T ;H −1 ())
∞
   t 2
T  
= λk  (t − s)α−1 Eα,α (−λk (t − s)α )F (·, s), ϕk ds  dt
 
k=1 0 0

∞
  T 2  T
1
≤ λ2k |s α−1
Eα,α (−λk s )|ds α
|F (·, s), ϕk |2 ds.
0 λk 0
k=1
106 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

Here (4.49) and (4.50) imply

 T 2
λ2k |s α−1 Eα,α (−λk s α )|ds = (1 − Eα,1 (−λk T α ))2 ≤ 1. (4.53)
0

Consequently,
 ∞
T  1
S2 2
L2 (0,T ;H −1 ())
≤ |F (·, s), ϕk |2 ds
0 λk
k=1
 T
= F (·, s) 2
H −1 ()
ds = F 2
L2 (0,T ;H −1 ())
. (4.54)
0

Hence

∂tα v2 L2 (0,T ;H −1 ()) ≤C F L2 (0,T ;H −1 ()) .

With (4.46) and (4.48), since

∞

v(x, t) − a(x) = v(x, t) − (a, ϕk )ϕk = (v1 (x, t) − a(x)) + v2 (x, t),
k=1

we see that

∂tα (v − a) L2 (0,T ;H −1 ()) ≤ C( a L2 () + F L2 (0,T ;H −1 ()) ).

Fourth Step Here we verify that v ∈ L2 (0, T ; H01 ()) and estimate
1
A v L2 (0,T ;L2 ()), which is equivalent to ∇v
2
L2 (0,T ;L2 ()) .
First Lemma 2.5 (i) yields
∞ 2
 1  ∞

 
 A 2 (a, ϕk )Eα,1 (−λk t α )ϕk  = λk |(a, ϕk )|2 |Eα,1 (−λk t α )|2
 
k=1 L2 () k=1
⎛ ⎞2
∞
  2 ∞

1 α
2 2
1 λ t
≤C |(a, ϕk )|2 λk = Ct −α |(a, ϕk )|2 ⎝ k α ⎠
1 + λk t α 1 + λk t
k=1 k=1

≤ Ct −α a 2
L2 ()
.
4.6 Case Where the Coefficients of A(t) Are Independent of Time 107

Therefore, by 0 < α < 1, we obtain

∞ 2
 1 
 α 
 A 2 (a, ϕk )Eα,1 (−λk t )ϕk 
 
k=1 L2 (0,T ;L2 ())
 T 
≤C t −α dt a 2
L2 ()
=C a 2
L2 ()
.
0

In the same way as for the estimation of S2 2 , using Lemma A.1,

L2 (0,T ;H −1 ())
(4.53) and (4.54), we obtain
 + ∞   ,
 1  t 2
 2 
A F (·, s), ϕk (t − s)α−1
Eα,α (−λk (t − s) )ds ϕk 
α
 0 
k=1 L2 (0,T ;L2 ())
∞
   t 2
T  
= λk  F (·, s), ϕk (t − s)α−1 Eα,α (−λk (t − s)α )ds  dt
 
k=1 0 0

≤C F L2 (0,T ;H −1 ()) .

Thus v satisfies (4.4), (4.5) and the same estimate as (4.7).

Finally we prove that v satisfies (4.3). For each N ∈ N, we set


N 
N
aN (x) = (a, ϕk )ϕk (x), v1N (x, t) = (a, ϕk )Eα,1 (−λk t α )ϕk (x),
k=1 k=1

N 
 t 
v2N (x, t) = F (·, s), ϕk (t − s)α−1 Eα,α (−λk (t − s)α )ds ϕk (x)
k=1 0

and

v N (x, t) = (v1N + v2N )(x, t).

Similarly to (4.47) and (4.52), we have


N
∂tα (v1N − aN )(x, t) = − λk (a, ϕk )Eα,1 (−λk t α )ϕk (x)
k=1
+N ,

=−A (a, ϕk )Eα,1 (−λk t α )ϕk (x) = −Av1N
k=1
108 4 Initial Boundary Value Problems for Time-Fractional Diffusion Equations

and


N
∂tα v2N (x, t) = F (·, t), ϕk ϕk
k=1
+N  ,
 t
−A F (·, s), ϕk (t − s)α−1 Eα,α (−λk (t − s)α )dsϕk
k=1 0


N
= − Av2N (x, t) + F (·, t), ϕk ϕk .
k=1

Hence


N
∂tα (v1N − a)(x, t) + Av N = F (·, t), ϕk ϕk
k=1

for each N ∈ N. Since

 2 
 
N  T ∞

 
F − F (·, t), ϕk ϕk  = |F (·, t), ϕk |2 λ−1
k dt −→ 0
  0
k=1 L2 (0,T ;H −1 ()) k=N+1

as N −→ ∞ by F ∈ L2 (0, T ; H −1()), we see that

lim (∂tα (v N − a) + Av N ) = F in L2 (0, T ; H −1()).

N→∞

We have already proved that v satisfies (4.7), so that

lim (∂tα (v N − a) + Av N ) = ∂tα (v − a) + Av in L2 (0, T ; H −1 ()),

N→∞

which yields that v satisfies (4.3). 

Chapter 5
Decay Rate as t → ∞

5.1 Main Result

We consider
⎧ α
⎨ ∂t (u − a) + A(t)u(x, t) = 0 in L2 (), 0 < t < T ,
u(·, t) ∈ H01 (), 0 < t < T , (5.1)
⎩
u − a ∈ Hα (0, T ; L2 ()).

Here


n
− (A(t)v)(x, t) = ∂i (aij (x, t)∂j v(x)) + c(x, t)v, (5.2)
i,j =1

where

c≤0 on  × [0, ∞).

In the case of α = 1, we can prove that there exist constants C > 0 and θ > 0 such
that

u(·, t) L2 () ≤ Ce−θt a L2 () , t >0

for each a ∈ L2 () by the classical energy estimate. For 0 < α < 1, we know that

C
u(·, t) L2 () ≤ a L2 () , t >0 (5.3)
tα

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 109
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5_5
110 5 Decay Rate as t → ∞

in the case where the coefficients of A(t) are independent of t (e.g., [26]). On
the other hand, Vergara and Zacher [28] proved (5.3) for t-dependent A(t). For
t-dependent A(t), in a conventional energy estimate, we multiply the equation by
u − a and integrate over  and we may obtain only a weak estimate

C
u(·, t) L2 () ≤ α a H01 () , t > 0.
t2
This is weaker than we can expect in terms of (5.3) for t-independent A(t), because
the decay rate is α2 .
In this chapter, we modify the arguments in Vergara and Zacher [28] to prove
(5.3) within our framework of solutions in the case where the coefficients of A(t)
are dependent on time t. More precisely, we prove:
Theorem 5.1 In (5.2), we assume

aij = aj i ∈ C 2 ([0, ∞); C 1 ()), 1 ≤ i, j ≤ n, (5.4)

c ∈ L∞ ( × (0, ∞)), and there exists a constant μ0 > 0 such that


n 
n
aij (x, t)ξi ξj ≥ μ0 ξj2 , x ∈ , t ≥ 0, ξ1 , ..., ξn ∈ R (5.5)
i,j =1 j =1

and

c(x, t) ≤ 0, x ∈ , t ≥ 0. (5.6)

Then there exists a constant C > 0 such that

u(·, t) L2 () ≤ Ct −α a L2 ()

for each solution u to (5.1).

For the proof, we have to return to the construction of the solution to (5.1) by the
Galerkin approximation. By means of a conventional energy estimate which can be
obtained by multiplying the fractional equation by the solution u to integrate over 
by parts, we cannot reach the desired estimate u(·, t) L2 () ≤ Ct −α u(·, 0) L2 () .
Thus we need another kind of energy inequality, namely Lemma 5.1 by Vergara and
Zacher [28]. The lemma is technical and is shown in Sect. 5.2. Next, repeating the
argument by the Galerkin approximation based on the lemma, we can complete the
proof of the desired decay estimate.
5.2 Key Lemmata 111

5.2 Key Lemmata

The proof is based on another kind of energy inequality:

Lemma 5.1 We set
1 2 1
G(s) = s −s+ .
2 2

We assume that u ∈ C( × [0, T ]),

u(x, ·) ∈ W 1,1 (0, T ) for x ∈ , (5.7)

and

t 1−α ∂t u ∈ L∞ ( × (0, T )). (5.8)

Then for each t ∈ (0, T ] we have

⎧  
⎪ t u(·,s)
⎪
⎪ dtα u(·, t) 2 + 2α
(1−α) 0 (t − s)−α−1 u(·, t) 2 G L2 ()
ds
⎪
⎪ L2 ()
 
L2 () u(·,t ) L2 ()
⎪
⎨ 
2t −α u(·,0) L2 ()
+ (1−α) u(·, t) 2
L2 ()
G u(·,t ) ≤2 α
 u(x, t)dt u(x, t)dx,
⎪
⎪
L2 ()
⎪
⎪ = 0,
⎪
⎪ if u(·, t) L2 ()
⎩
dtα u(·, t) 2
L2 ()
≤ 0, if u(·, t) L2 () = 0.
(5.9)

Furthermore

u(·, t) L2 () dtα u(·, t) L2 () ≤ u(x, t)dtα u(x, t)dx. (5.10)


Proof Choosing M > 0 sufficiently large, we have uM := u+M > 0 on ×[0, T ].

We note that uM satisfies (5.7) and (5.8). We have

2 uM (x, t)dtα uM (x, t)dx − dtα uM (·, t) 2L2 ()

  t
2
= (t − s)−α (∂s uM )(x, s)(uM (x, t) − uM (x, s))dsdx
(1 − α)  0
  t    
2 −α uM (x, s) uM (x, s)
= |uM (x, t)| 2
(t − s) ∂s 1− ds dx.
(1 − α)  0 uM (x, t) uM (x, t)
112 5 Decay Rate as t → ∞

Hence, since G (s) = s − 1, we obtain


2 uM (x, t)dtα uM (x, t)dx − dtα uM (·, t) 2
L2 ()

  t    
2 2 −α ∂ uM (x, s)
=− |uM (x, t)| (t − s) G ds dx.
(1 − α)  0 ∂s uM (x, t)

We integrate by parts to have

 t     t−h   
uM (x, s)
∂ ∂ uM (x, s)
(t − s)−α
G ds = lim (t − s)−α G ds
0 ∂s
uM (x, t) h→0+ 0 ∂s uM (x, t)
 t−h    
uM (x, s) uM (x, s) s=t−h
= lim −α (t − s)−α−1 G ds + (t − s)−α G
h→0+ 0 uM (x, t) uM (x, t) s=0
 t−h      
uM (x, s) uM (x, t − h) uM (x, 0)
= lim −α (t − s)−α−1 G ds + h−α G − t −α G .
h→0+ 0 uM (x, t) uM (x, t) uM (x, t)

We shall show that

 
uM (x, t − h)
lim h−α |uM (x, t)|2 G = 0, t > 0. (5.11)
h→0+ uM (x, t)
Indeed, since
  
 −α  −α
h |uM (x, t)|2 G uM (x, t − h)  = h |uM (x, t) − uM (x, t − h)|2 ,
 uM (x, t)  2
assumption (5.8) yields
 t 
 ∂uM 
|uM (x, t − h) − uM (x, t)| =  (x, ξ )dξ 
t −h ∂ξ
 t    t
 1−α ∂uM  α−1
≤   dξ ≤ C(t − h) α−1
t 1−α ∂t uM
ξ ∂ξ
(x, ξ ) ξ L∞ (×(0,T )) dξ
t −h t −h

≤ h(t − h)α−1 t 1−α ∂t uM L∞ (×(0,T )) ,

so that (5.11) holds.

Thus, letting h → 0+, we obtain

2 uM (x, t)dtα uM (x, t)dx − dtα uM (·, t) 2
L2 ()

  t  
2α uM (x, s)
= (t − s)−α−1 |uM (x, t)|2 G dsdx
(1 − α)  0 uM (x, t)
  
2t −α uM (x, 0)
+ |uM (x, t)|2 G dx.
(1 − α)  uM (x, t)
5.2 Key Lemmata 113

Hence, by setting τ = 0 or = s, the Cauchy–Schwarz inequality yields

  
uM (x, τ )
|uM (x, t)|2 G dx
 uM (x, t)
  
1 2 1 2
= |uM (x, t)| + |uM (x, τ )| − uM (x, t)uM (x, τ ) dx
 2 2
1 1
≥ uM (·, t) + uM (·, τ ) 2L2 () − uM (·, t)
2
L2 () L2 () uM (·, τ ) L2 ()
2 2
+ ,
uM (·, τ ) L2 ()
= uM (·, t) L2 () G
2
.
uM (·, t) L2 ()

Thus (5.9) holds for uM . If u(·, t) L2 () =

 0, then letting M → 0, we obtain (5.9)
for u. Next, we assume u(·, t) L2 () = 0. Since G(η) ≥ 0 for η ≥ 0, inequality
(5.9) for uM yields

dtα u(·, t) + M 2
L2 ()
≤2 (u(x, t) + M)dtα (u(x, t) + M)dx


=2 (u(x, t) + M)dtα u(x, t)dx.


The right-hand side tends to 0 as M → 0 by u(·, t) L2 () = 0. Moreover,

 t   
1 d
dtα u(·, t) + M 2
L2 ()
= (t − s)−α (u(x, s) + M)2 dx ds
(1 − α) 0 ds 
 t   
1 −α
= (t − s) ∂s ( u(·, s) 2
L2 ()
+ 2M ∂s u(x, s)dx ds
(1 − α) 0 
 t
2M
= dtα u(·, t) 2
L2 ()
+ (t − s)−α ∂s u(x, s)dxds −→ dtα u(·, t) 2
L2 ()
(1 − α) 0 

as M → 0 by |∂s u(x, s)| ≤ Cs α−1 by (5.8). Hence dtα u(·, t) 2L2 () ≤ 0, and (5.9)
holds for u(·, t) L2 () = 0.
Finally, we have to prove (5.10). We can assume that u(·, t) L2 () = 0,
because (5.10) is trivial if u(·, t) L2 () = 0. By (5.8), we can directly verify
ds u(·, s) L2 () ∈ L (0, T ). First,
d 1

dtα u(·, t) 2
L2 ()
− 2 u(·, t) α
L2 () dt u(·, t) L2 ()
 t
2 d
= (t − s)−α ( u(·, s) L2 () − u(·, t) L2 () ) u(·, s) L2 () ds
(1 − α) 0 ds
 + , + ,
2 t u(·, s) d u(·, s) L2 ()
−α L2 ()
= (t − s) u(·, t) 2
−1 ds
(1 − α) 0
L2 () u(·, t) L2 () ds u(·, t) L2 ()
114 5 Decay Rate as t → ∞

 - + ,.
2 t d u(·, s)
−α L2 ()
= (t − s) u(·, t) 2
G ds
(1 − α) 0
L2 () ds u(·, t) L2 ()
 + ,
2α t u(·, s)
−α−1 L2 ()
=− (t − s) u(·, t) ds 2
L2 ()
G
(1 − α) 0 u(·, t) L2 ()
+ ,
2 u(·, s) L2 () s=t
+ (t − s)−α u(·, t) 2L2 () G 
(1 − α) u(·, t) L2 () s=0
 t + ,
−2α −α−1
u(·, s) L2 ()
≥ (t − s) 2
u(·, t) L2 () G ds
(1 − α) 0 u(·, t) L2 ()
+ ,
2t −α u(·, 0) L2 ()
− 2
u(·, t) L2 () G . Here we used also (5.8). Hence
(1 − α) u(·, t) L2 ()

 + ,
2α t u(·, s)
−α−1 L2 ()
dtα u(·, t) 2
+ (t − s) u(·, t) 2
G ds
L2 () (1 − α) 0
L2 () u(·, t) L2 ()
+ ,
2t −α u(·, 0) L2 ()
+ u(·, t) 2
L2 ()
G
(1 − α) u(·, t) L2 ()

≥ 2 u(·, t) α
L2 () dt u(·, t) L2 () .

Combining with (5.9), we reach (5.10). Thus the proof of Lemma 5.1 is complete.
Moreover, we show a generalized extremum principle under weaker assumptions
than Lemma 1.2.
Lemma 5.2 Let f ∈ W 1,1 (0, T ) attain its maximum over the interval [0, T ] at a
point t0 ∈ (0, T ]. If for each κ ∈ (0, T ), there exists β ∈ (0, 1] such that f ∈
1
W 1, 1−β (κ, T ), then (dtα f )(t0 ) ≥ 0 for every α ∈ (0, β).
Proof Firstly, we introduce the function g(t) := f (t0 ) − f (t) for t ∈ [0, T ].
We notice that g(t) ≥ 0 and (dtα g)(t) = −(dtα f )(t) for t ∈ [0, T ]. Then
g ∈ W 1,1 (0, T ) and g(t0 ) = 0. Hence, for κ ∈ (0, t0 ), the Hölder inequality yields
    
t0  dg   
|g(t)| ≤  (s) ds ≤  dg  1 |t − t0 |β for t ∈ [κ, t0 ]. (5.12)
 ds   dt  1−β
t L (κ,T )

Therefore for small h > 0 satisfying κ < t0 − h < t0 , we have

 
 dg 
|g(t0 − h)| ≤   1
 ds  1−β hβ . (5.13)
L (κ,T )
5.3 Completion of Proof of Theorem 5.1 115

We note that 0 < α < β. For fixed α ∈ (0, β), we obtain

 κ  t0
1 dg 1 dg
(dtα g)(t0 ) = (t0 − s)−α (s)ds + (t0 − s)−α (s)ds.
(1 − α) 0 ds (1 − α) κ ds

We fix ε > 0 arbitrarily. It follows from g ∈ W 1,1 (0, T ) that dg

ds ∈ L (0, T ). By the
1

Young inequality, we deduce that there exists sufficiently small κ > 0 such that the
first integral is smaller than ε. As for the second one, integration by parts, g ≥ 0 on
[0, T ] and g(t0 ) = 0 yield
 t0  t0 −h
dg dg
(t0 − s)−α (s)ds = lim (s)ds (t0 − s)−α
κ ds
κ ds
h→0+
 t0 −h
−α −α
= − (t0 − κ) g(κ) + lim h g(t0 − h) − lim α (t0 − s)−α−1 g(s)ds
h→0+ h→0+ κ
−α
≤ lim h g(t0 − h) = 0.
h→0+

For the last limit, we used (5.13) by noting 0 < α < β. Thus the proof of Lemma 5.2
is completed. 

5.3 Completion of Proof of Theorem 5.1

We apply the argument in Sect. 4.3 of Chap. 4 for proving the existence of u to
(5.1), and we use the same notations. Since the non-homogeneous term F (x, t) is
identically zero, we need not introduce the approximation parameter ε > 0 like
Sect. 4.3 of Chap. 4.
We recall that the function


N
uN (x, t) := pN, (t)ρ (x)
=1

satisfies
 N 

(dtα uN )(x, t)ρ (x)dx + aij (x, t)∂j uN (x, t)∂i ρ (x)dx
 i,j =1 

= c(x, t)uN (x, t)ρ (x)dx

116 5 Decay Rate as t → ∞

for  = 1, ..., N. In terms of (4.20), we apply Theorem 3.6 (iv) to see

pN, − c ∈ Wα (0, T ).

We multiply this identity by pN, (t) and sum it up from  = 1 to N:

 N 

uN (x, t)(dtα uN )(x, t)dx + aij (x, t)(∂j uN )(x, t)(∂i uN )(x, t)dx
 i,j =1 

= c(x, t)|uN (x, t)|2 dx.


Using (5.5) and (5.6), we obtain


uN (x, t)(dtα uN )(x, t)dx + μ0 ∇uN (·, t) 2
L2 ()
≤ 0.


We define

uN,η = uN + ηρN+1 ,

where η > 0 is a constant. Using the fact that dtα (ηρN+1 ) = 0 and the orthogonality
of {ρN }∞
N=1 , we obtain

uN,η (x, t)dtα uN,η (x, t)dx + μ0 ∇uN (t, ·) 2
L2 ()
≤ 0.


By Theorem 3.6 (iv) in Chap. 3, we see that the functions uN,η satisfy the
assumptions of Lemma 5.1, so that (5.10) and the Poincaré inequality yield
2
uN,η (·, t) α
L2 () dt uN,η (·, t) L2 () + c0 uN (·, t) L2 ()
≤ 0. (5.14)

Here the constant c0 > 0 is independent of N and η.

Since uN,η = uN + ηρN+1 and


N
(uN , ρN+1 ) = pN, (t)(ρ , ρN+1 ) = 0,
=1

we have

uN,η (·, t) 2
L2 ()
= uN (·, t) 2
L2 ()
+ η2 ,
5.3 Completion of Proof of Theorem 5.1 117

that is, c0 uN,η (·, t) 2

L2 ()
= c0 uN (·, t) 2
L2 ()
+ c0 η2 . Therefore

uN,η (·, t) α
L2 () dt uN,η (·, t) L2 () + c0 uN,η (·, t) 2
L2 ()

= uN,η (·, t) α
L2 () dt uN,η (·, t) L2 () + c0 uN (·, t) 2
L2 ()
+ c0 η 2 ≤ c0 η 2

by (5.14). Consequently,

uN,η (·, t) α
L2 () dt uN,η (·, t) L2 () + c0 uN,η (·, t) 2
L2 ()
≤ η 2 c0 .

We note that uN,η (·, t) 2

L2 ()
≥ η2 > 0, and we can divide by uN,η (·, t) L2 () to
obtain

dtα uN,η (·, t) L2 () + c0 uN,η (·, t) L2 () ≤ ηc0 . (5.15)

With the scalar product (·, ·) in L2 (0, T ), we have

(dtα uN,η (·, t) L2 () , ξ ) + (c0 uN,η (·, t) L2 () , ξ ) ≤ (ηc0 , ξ ) (5.16)

for all ξ ∈ C0∞ (0, T ) satisfying ξ ≥ 0. In view of (3.12), by Lemma 2.7 and
integration by parts, we know that (2.51) holds for u ∈ W 1,1 (0, T ) and ξ ∈
C0∞ (0, T ):

(dtα uN,η (·, t) L2 () , ξ)

 T 
α ∗ 1 −α
= ( uN,η (·, t) L2 () , (dt ) ξ ) − t ξ(t)dt uN,η (·, 0) L2 () .
(1 − α) 0

By the definition of uN,η , we can readily verify that uN,η (·, t) L2 () −→
uN (·, t) L2 () in L1 (0, T ) and uN,η (·, 0) L2 () −→ uN (·, 0) L2 () as η → 0.
Therefore

lim (dtα uN,η (·, t) L2 () , ξ) (5.17)

η→0

 T 
1
= ( uN (·, t) L2 () , (dtα )∗ ξ ) − t −α ξ(t)dt uN (·, 0) L2 () .
(1 − α) 0

Since uN (·, t) L2 () ∈ W 1,1 (0, T ) by Theorem 3.6 (iv), we apply (2.51) to the
right-hand side of (5.17) for uN (·, t) L2 () , and we obtain

lim (dtα uN,η (·, t) L2 () , ξ ) = (dtα uN (·, t) L2 () , ξ ).

η→0
118 5 Decay Rate as t → ∞

Therefore, letting η → 0 in (5.16), we reach

(dtα uN (·, t) L2 () + c0 uN (·, t) L2 () , ξ) ≤ 0 ξ ∈ C0∞ (0, T ), ξ ≥ 0.

(5.18)

We choose the mollifier χε with ε > 0 defined by (3.19). We note

θ (t) := dtα uN (·, t) L2 () + c0 uN (·, t) L2 () ∈ L1 (0, T ).

Again using the letter θ , we denote the zero extension of θ to R, and θ ∈ L1 (R).
Setting ξ(t) = χε (τ − t) with arbitrarily fixed τ ∈ (0, T ) and ε > 0, we see
that ξ ≥ 0and ξ ∈ C0∞ (0, T ) with sufficiently small ε > 0. By (5.18) we obtain
∞
θε (τ ) := −∞ θ (t)χε (τ − t)dt ≤ 0 for small ε > 0. We know that θε −→ θ
in L1 (R) as ε → 0 (e.g., [2]), so that we can choose a subsequence {εn }n∈N with
limn→∞ εn = 0 such that limn→∞ θεn (τ ) = θ (τ ) for almost all τ ∈ (0, T ). Hence,
letting n → ∞ in (5.18), we reach

dtα uN (·, t) L2 () + c0 uN (·, t) L2 () ≤0 for almost all t ∈ (0, T ). (5.19)

Using the power series of Eα,1 (−c0 t α ), one can directly verify that

v(t) = uN (·, 0) L2 () Eα,1 (−c0 t

α
)

satisfies

dtα v(t) + c0 v(t) = 0, v(0) = uN (·, 0) L2 () .

Now we will apply Lemma 5.2 to obtain that

uN (·, t) L2 () ≤ v(t) for every t ∈ [0, T ]. (5.20)

Proof of (5.20) We set f (t) := uN (·, t) L2 () − v(t). Then

dtα f (t) + c0 f (t) ≤ 0, t ∈ [0, T ], f (0) = 0. (5.21)

It suffices to verify f (t) ≤ 0 for 0 ≤ t ≤ T . Assume contrarily that there exists

t0 ∈ (0, T ] such that f (t0 ) = max f (t) > 0. From Theorem 3.6 (iv), we deduce
t ∈[0,T ]
that f ∈ W 1,∞ (κ, T ) for each κ > 0. Thus by Lemma 5.2, we have (dtα f )(t0 ) ≥ 0,
which means that dtα f (t0 ) + c0 f (t0 ) > 0. This is a contradiction with (5.21). Thus
we have verified (5.20).
Next, let ξ ∈ C0∞ (0, T ), ≥ 0 be arbitrary. Then (5.20) yields
 T  T
ξ(t) uN (·, t) 2
L2 ()
dt ≤ ξ(t)v 2 (t)dt.
0 0
5.3 Completion of Proof of Theorem 5.1 119

There exists a subsequence {uN } (still indexed by N) such that uN  u weakly

in L2 (0, T ; H01()), where u is the unique solution to (5.1) given by Theorem 4.1.
Thus by the weak lower semi-continuity of the L2 -norm, we have
 T  T
ξ(t) u(·, t) 2
L2 ()
dt ≤ ξ(t)v 2 (t)dt
0 0

for each ξ ∈ C0∞ (0, T ), ≥ 0. Hence

u(·, t) L2 () ≤ v(t) for almost all t ∈ [0, T ].

Since T > 0 is arbitrary and the solution u is uniquely defined in [0, ∞) by

Theorem 4.1, we reach u(·, t) L2 () ≤ v(t) for almost all t ≥ 0, which means

u(·, t) L2 () ≤ a L2 () Eα,1 (−c0 t

α
) for almost all t ≥ 0.

Finally, the asymptotic behavior of the Mittag-Leffler function implies

u(·, t) L2 () ≤ C1 t −α a L2 () , t > 0.

Thus the proof of Theorem 5.1 is complete. 

Chapter 6
Concluding Remarks on Future Works

The references on fractional differential equations are rapidly increasing and it is

impossible to refer to them to some substantial extent. Thus in this book with
limited pages, we concentrate on establishing one possible theory for the initial
boundary value problems for time-fractional partial differential equations. Our
theory redefines the time fractional derivative ∂tα keeping the consistency with the
classical Riemann–Liouville derivative and the Caputo derivative, and ∂tα should be
handled in a convenient way for the applications (e.g., the isomorphism between
Hα (0, T ) and L2 (0, T )).
We are obliged to give up comprehensive comparisons with other results. Also as
for possible future studies, we are restricted to mention a few topics as follows.
1. In this book, we discuss only the case of the order 0 < α < 1 of the time-
fractional derivative ∂tα . From the physical viewpoint, the case of α > 1, in
particular, 1 < α < 2 is important. The corresponding theory for the case α > 1
can be constructed similarly; see Remark 2.4 of Chap. 2. However, we postpone
the details to future works.
2 Here we do not discuss nonlinear equations

∂tα (u − a) + A(t)u = F (x, t, u, ∇u).

Needless to say, one purpose of the linear theory developed in this book is the
applications to the nonlinear theory.
3. Here we argue only the homogeneous Dirichlet boundary condition. For exam-
ple, the Neumann boundary condition should be considered and initial nonhomo-
geneous boundary value problems are demanded. The theory for common partial
differential equations such as α = 1, 2 has been established satisfactorily (e.g.,
Lions and Magenes [19]). Such a theory is not only mathematically interesting
but also is a basis for e.g., control problems, and widely applicable. As for
non-homogeneous boundary value problems for fractional partial differential
equations, we refer only to Yamamoto [29] and the references therein.

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 121
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5_6
122 6 Concluding Remarks on Future Works

4. There are various topics on inverse problems where we are required to determine
parameters such as coefficients and orders of the derivatives of fractional
differential equations by data of solutions which are over-determining in view
of the well-posedness of initial boundary value problems. As for surveys, see
the three chapters [16a–c] in the handbook edited by Kochubei, Luchko and
Machado. The current book has established the foundations for future studies of
inverse problems for fractional differential equations, and we can greatly expect
the comprehensive development of studies of inverse problems.
Appendix A
Proofs of Two Inequalities

For convenience, we prove two inequalities which have been used in this book.
We set
 t
(ρ ∗ v)(t) = ρ(t − s)v(s)ds, 0 < t < T .
0

Lemma A.1 (The Young Inequality for the Convolution) Let 1 ≤ p ≤ ∞, ρ ∈

L1 (0, T ) and v ∈ Lp (0, T ). Then

ρ∗v Lp (0,T ) ≤ ρ L1 (0,T ) v Lp (0,T ) .

Proof The proof for p = ∞ is straightforward. Let p = 1. Since

 T  t   T  T 
· · · ds dt = · · · dt ds (A.1)
0 0 0 s

we have
  t   T  t 
 T 
ρ ∗v =  ρ(t − s)v(s)ds  dt ≤ |ρ(t − s)||v(s)|ds dt
L1 (0,T )  
0 0 0 0
 T  T   T  T −s 
= |ρ(t − s)|dt |v(s)|ds = |ρ(η)|dη |v(s)|ds
0 s 0 0
 T  T 
≤ |ρ(η)|dη |v(s)|ds = ρ L1 (0,T ) v L1 (0,T ) .
0 0

For the second equality from the last, we changed the integral variables η = t − s.
Thus the proof in the case of p = 1 is complete.

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 123
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5
124 A Proofs of Two Inequalities

Finally, let 1 < p < ∞. Let q ∈ (1, ∞) satisfy 1

p + 1
q = 1, which is possible by
1 < p < ∞. We have to estimate
  t p
T  
 ρ(t − s)v(s) dt.
 
0 0

Then
1
1− q1
|ρ(t − s)v(s)| = |ρ(t − s)| q (|ρ(t − s)| |v(s)|),

and the Hölder inequality yields

 t  t 1
1− q1
|ρ(t − s)||v(s)|ds = |ρ(t − s)| q (|ρ(t − s)| |v(s)|)ds
0 0
 t  q1  t  p1
p q−1
≤ |ρ(t − s)|ds |ρ(t − s)| q |v(s)| dsp
0 0

1
 t  p1
≤ ρ q
L1 (0,T )
|ρ(t − s)||v(s)| ds p
.
0

Therefore, using (A.1), we obtain

 t
 p  T  t 
 T  p
 ρ(t − s)v(s)ds  dt ≤ ρ q 1 |ρ(t − s)||v(s)| ds dt
p
  L (0,T )
0 0 0 0
p  T  T 
= ρ L1 (0,T )
q
|ρ(t − s)|dt |v(s)|p ds
0 s
p  T  T −s 
= ρ q
L1 (0,T )
|ρ(η)|dη |v(s)|p ds,
0 0

where we changed the integral variables η = t − s. Hence

p p+q
p p p
ρ∗v Lp (0,T ) ≤ ρ q
L1 (0,T )
ρ L1 (0,T ) v Lp (0,T ) = ρ q
L1 (0,T )
v Lp (0,T ) ,

which implies the conclusion by 1

p + 1
q = 1. Thus the proof of Lemma A.1 is
complete. 
Lemma A.2 (A Generalized Gronwall Inequality) Let C0 > 0 be a constant and
0 < α < 1. Moreover, let r ∈ L1 (0, T ), ≥ 0 in (0, T ). We assume that u ∈ L1 (0, T )
satisfies
 t
0 ≤ u(t) ≤ r(t) + C0 (t − s)α−1 u(s)ds, 0 ≤ t ≤ T.
0
A Proofs of Two Inequalities 125

Then
 t
u(t) ≤ r(t) + C1 e C2 t
(t − s)α−1 r(s)ds, 0 ≤ t ≤ T. (A.2)
0

Here the constants C1 > 0 and C2 > 0 are dependent on α, C0 , but independent of
T > 0. We note that if C0 > 0 is independent of T > 0, then (A.2) holds for t > 0
with C1 , C2 > 0 which are independent of t > 0.
Proof We set
 s
(Lr)(s) = C0 (s − ξ )α−1 r(ξ )dξ, 0<s<T
0

for r ∈ L1 (0, T ). The assumption is written as

0 ≤ u(t) ≤ r(t) + (Lu)(t), 0 < t < T. (A.3)

Since Lv ≥ 0 for v ≥ 0 in (0, T ), we have

Lu ≤ L(r + Lu) = Lr + L2 u in (0, T ).

Then (A.3) yields

u ≤ r + Lu ≤ r + Lr + L2 u in (0, T ).

We successively estimate to obtain


N
u≤r+ Lk r + LN+1 u in (0, T ) (A.4)
k=1

for each N ∈ N. Exchanging the orders of the integrals, we have

 t  s 
(L2 r)(t) = C02 (t − s)α−1 (s − ξ )α−1 r(ξ )dξ ds
0 0
   
t t C 2 (α)2 t
= C02 r(ξ ) (t − s)α−1 (s − ξ )α−1 ds dξ = 0 (t − ξ )2α−1 r(ξ )dξ.
0 ξ (2α) 0

Here, by r ∈ L1 (0, T ) and the Young inequality, we note that all the functions
including r(ξ ) in the calculations here are in L1 (0, T ). Continuing the calculations,
we can reach

(C0 (α))k t
(Lk r)(t) = (t − s)kα−1 r(s)ds, k ∈ N. (A.5)
(kα) 0
126 A Proofs of Two Inequalities

1
We set θ = (C0 (α)) α . Substitution of (A.5) into (A.4) yields
 t +
N
,  t
θ αk (t − s)αk−1 θ α(N +1)
u(t ) ≤ r(t ) + r(s)ds + (t − s)(N +1)α−1 u(s)ds
0 (αk) ((N + 1)α) 0
k=1
 t +
N
,  t
θ αk s αk−1 θ α(N +1)
= r(t ) + r(t − s)ds + (t − s)(N +1)α−1 u(s)ds.
0 (αk) ((N + 1)α) 0
k=1
(A.6)

On the other hand, since Eα,1 (z) is an entire function in z ∈ C, we can verify
+ ∞
, ∞ ∞
d d  (θ s)αk  θ αk (αk)s αk−1  θ αk s αk−1
Eα,1 ((θ s)α ) = = =
ds ds (αk + 1) (αk + 1) (αk)
k=0 k=1 k=1

= s α−1 θ α Eα,α ((θ s)α ). (A.7)

Moreover, we have

1−α 1 C
|Eα,α (η)| ≤ C(1 + η) eη +
α
α
1+η

for all η ≥ 0 (e.g., Theorem 1.5 (p. 35) in Podlubny [24]). Therefore

1−α C
|Eα,α ((θ s)α )| ≤ C(1 + θ α s α ) α eθs + ≤ C1 eC2 s , s > 0,
1 + θ αsα

where the constants C, C1 > 0 are independent of s. Hence

max |Eα,α ((θ s)α )| ≤ C1 eC2 t

0≤s≤t

for all t ≥ 0. That is, (A.7) yields

∞ 
 θ αk (t − s)αk−1 
 
  ≤ C1 eC2 t (t − s)α−1 , 0 ≤ s ≤ t.
 (αk) 
k=1

Letting N → ∞ in (A.6), we obtain

 t
u(t) ≤ r(t) + C1 eC2 t (t − s)α−1 r(s)ds
0

θ α(N+1) t
+ lim (t − s)(N+1)α−1 u(s)ds, 0 ≤ t ≤ T.
N→∞ ((N + 1)α) 0
A Proofs of Two Inequalities 127

Let us choose N0 ∈ N such that (N0 + 1)α − 1 > 0. Then for each N ≥ N0 , we
have
 t   t
 
 (t − s)(N+1)α−1 u(s)ds  ≤ t (N+1)α−1 |u(s)|ds ≤ T (N+1)α−1 u L1 (0,T ) .
 
0 0

Consequently,
  t 
 θ N+1  α N
 (t − s) (N+1)α−1
u(s)ds  ≤ θ T α−1 (θ T ) u
 ((N + 1)α)  ((N + 1)α) L1 (0,T )
0

for each t ∈ [0, T ] and N ≥ N0 . The Stirling formula yields

(θ T α )N
lim = 0.
N→∞ ((N + 1)α)

Thus the proof of Lemma A.2 is complete. 

Notation

· L2 (0,T ) , (·, ·) L2 -norm, L2 -scalar product in L2 (0, T )

· L2 () , (·, ·)L2 () L2 -norm, L2 -scalar product in L2 ()

Dtα , 0 < α < 1 Riemann–Liouville derivative, see (1.3):

 t
1 d
Dtα u(t) = (t − s)−α u(s)ds
(1 − α) dt 0
dtα , 0 < α < 1 Caputo derivative, see (1.4):
 t
1 du
dtα u(t) = (t − s)−α (s)ds
(1 − α) 0 ds
J α, α > 0 Riemann–Liouville fractional integral operator, see (1.5):
 t
1
J α f (t) = (t − s)α−1 f (s)ds
(α) 0
∂tα , 0 < α < 1 Definition 2.1 in Sect. 2.3 of Chap. 2

W 1,1 (0, T ), 0 W 1,1 (0, T ) Definition (1.17)

Wα (0, T ) Definition (2.5)

1 [0, T ], H (0, T ), H α (0, T ), Section 2.2 of Chap. 2
0C α
H α (0, T ), H 1 (0, T )
0 0

Eα,β (z) Mittag-Leffler functions, see (2.28)

D(A) The domain of an operator A

R(A) The range of an operator A, i.e., R(A) = AD(A)

∂2
∂i = ∂ 2
∂xi , ∂i = , ∂s = ∂
Partial derivatives
∂xi2 ∂s

∇ = (∂1 , . . . , ∂n ) Gradient

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 129
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5
References

1. N.H. Abel, Résolution d’un problème de mécanique. J. Reine Angew. Math. 1, 97–101 (1826)
2. R.A. Adams, Sobolev Spaces (Academic, New York, 1975)
3. A. Alsaedi, B. Ahmad, M. Kirane, A survey of useful inequalities in fractional calculus. Fract.
Calc. Appl. Anal. 20, 574–594 (2017)
4. E.G. Bajlekova, Fractional evolution equations in Banach spaces, Doctoral thesis, Eindhoven
University of Technology, 2001
5. H. Brezis, Functional Analysis, Sobolev Spaces and Partial Differential Equations (Springer,
Berlin, 2011)
6. K. Diethelm, The Analysis of Fractional Differential Equations. Lecture Note in Mathematics,
vol. 2004 (Springer, Berlin, 2010)
7. L.C. Evans, Partial Differential Equations (American Mathematical Society, Providence, 1998)
8. R. Gorenflo, S. Vessella, Abel Integral Equations. Lecture Note in Mathematics, vol. 1461
(Springer, Berlin, 1991)
9. R. Gorenflo, M. Yamamoto, Operator theoretic treatment of linear Abel integral equations of
first kind. Jpn. J. Ind. Appl. Math. 16, 137–161 (1999)
10. R. Gorenflo, A.A. Kilbas, F. Mainardi, S.V. Rogosin, Mittag-Leffler Functions, Related Topics
and Applications (Springer, Berlin, 2014)
11. R. Gorenflo, Y. Luchko, M. Yamamoto, Time-fractional diffusion equation in the fractional
Sobolev spaces. Fract. Calc. Appl. Anal. 18, 799–820 (2015)
12. B. Jin, B. Li, Z. Zhou, Subdiffusion with a time-dependent coefficient: analysis and numerical
solution. Math. Comput. 88, 2157–2186 (2019)
13. T. Kato, Perturbation Theory for Linear Operators (Springer, Berlin, 1980)
14. Y. Kian, M. Yamamoto, On existence and uniqueness of solutions for semilinear fractional
wave equations. Fract. Calc. Appl. Anal. 20, 117–138 (2017)
15. A.A. Kilbas, H.M. Srivastava, J.J. Trujillo, Theory and Applications of Fractional Differential
Equations (Elsevier, Amsterdam, 2006)
16. A.N. Kochubei, Y. Luchko, J.A. Tenreiro Machado (eds.), Handbook of Fractional Calculus
with Applications, vol. 2 (De Gruyter, Berlin, 2019); (a) Y. Liu, Z. Li, M. Yamamoto, Inverse
problems of determining sources of the fractional partial differential equations, pp. 411–429;
(b) Z. Li, Y. Liu, M. Yamamoto, Inverse problems of determining parameters of the fractional
partial differential equations, pp. 431–442; (c) Z. Li, M. Yamamoto, Inverse problems of
determining coefficients of the fractional partial differential equations, pp. 443–464
17. A. Kubica, M. Yamamoto, Initial-boundary value problems for fractional diffusion equations
with time-dependent coefficients. Fract. Calc. Appl. Anal. 21, 276–311 (2018)

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 131
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5
132 References

18. Z. Li, X. Huang, M. Yamamoto, Initial-boundary value problems for multi-term time-fractional
diffusion equations with x-dependent coefficients. Evol. Equ. Control Theory 9, 153–179
(2020)
19. J.L. Lions, E. Magenes, Non-homogeneous Boundary Value Problems and Applications, vols.
I, II (Springer, Berlin, 1972)
20. Y. Luchko, Maximum principle for the generalized time-fractional diffusion equation. J. Math.
Anal. Appl. 351, 218–223 (2009)
21. Y. Luchko, Some uniqueness and existence results for the initial-boundary value problems for
the generalized time-fractional diffusion equation. Comput. Math. Appl. 59, 1766–1772 (2010)
22. K.S. Miller, S.G. Samko, Completely monotonic functions. Integral Transforms Spec. Funct.
12, 389–402 (2001)
23. A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential Equations
(Springer, Berlin, 1983)
24. I. Podlubny, Fractional Differential Equations (Academic, San Diego, 1999)
25. B. Ross, The development of fractional calculus 1695–1900. Hist. Math. 4, 75–89 (1977)
26. K. Sakamoto, M. Yamamoto, Initial value/boundary value problems for fractional diffusion-
wave equations and applications to some inverse problems. J. Math. Anal. Appl. 382, 426–447
(2011)
27. H. Tanabe, Equations of Evolution (Pitman, London, 1979)
28. V. Vergara, R. Zacher, Optimal decay estimates for time-fractional and other nonlocal
subdiffusion equation via energy methods. SIAM J. Math. Anal. 47, 210–239 (2015)
29. M. Yamamoto, Weak solutions to non-homogeneous boundary value problems for time-
fractional diffusion equations. J. Math. Anal. Appl. 460, 365–381 (2018)
30. R. Zacher, Weak solutions of abstract evolutionary integro-differential equations in Hilbert
spaces. Funkcial. Ekvac. 52, 1–18 (2009)
Index

A G

Adjoint operator 39 Galerkin approximation 80

Gamma function 1
Generalized extremum principle 114
C Generalized Gronwall inequality 124

Caputo derivative 2 H
Closable 25
Closure 25
Coercivity 48 Heinz–Kato inequality 20
Complete monotonicity 103 Homogeneous Dirichlet boundary condition
121

D
I

Decay rate 110

Initial boundary value problem 73
Initial condition 55
E Initial nonhomogeneous boundary value
problems 121
Initial value problem 47
Energy estimate 109 Initial value problem for a linear fractional
Extremum principle 5 ordinary differential equation 55
Interpolation space 15, 78
Inverse problems 122
F

L
Fractional ordinary differential equation 47
Fractional power 19
Fractional Sobolev space 12, 65 Laplace transform 9, 43
Fredholm alternative 57

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2020 133
A. Kubica et al., Time-Fractional Differential Equations, SpringerBriefs
in Mathematics, https://doi.org/10.1007/978-981-15-9066-5
134 Index

M R

Mittag-Leffler functions 26 Riemann–Liouville derivative 2

Mollifier 15, 53 Riemann–Liouville fractional integral operator
Multi-term time fractional derivatives 69 3

N S

Nonlinear equations 121 Sobolev embedding 7

P Y

Poincaré inequality 95 Young inequality 123