(Mathematics Research Developments) Jacky Cresson - Fractional Calculus in Analysis, Dynamics and Optimal Control-Nova Science Pub Inc (2014)

MATHEMATICS RESEARCH DEVELOPMENTS
FRACTIONAL CALCULUS IN ANALYSIS,

DYNAMICS AND OPTIMAL CONTROL
No part of this digital document may be reproduced, stored in a retrieval system or transmitted in any form or
by any means. The publisher has taken reasonable care in the preparation of this digital document, but makes no
expressed or implied warranty of any kind and assumes no responsibility for any errors or omissions. No
liability is assumed for incidental or consequential damages in connection with or arising out of information
contained herein. This digital document is sold with the clear understanding that the publisher is not engaged in
rendering legal, medical or any other professional services.
Additional books in this series can be found on Nova’s website

under the Series tab.
Additional e-books in this series can be found on Nova’s website

under the e-book tab.
FRACTIONAL CALCULUS IN ANALYSIS,

DYNAMICS AND OPTIMAL CONTROL
JACKY CRESSON
EDITOR
New York
Copyright © 2014 by Nova Science Publishers, Inc.
All rights reserved. No part of this book may be reproduced, stored in a retrieval system or
transmitted in any form or by any means: electronic, electrostatic, magnetic, tape, mechanical
photocopying, recording or otherwise without the written permission of the Publisher.
For permission to use material from this book please contact us:
Telephone 631-231-7269; Fax 631-231-8175
Web Site: http://www.novapublishers.com
NOTICE TO THE READER

The Publisher has taken reasonable care in the preparation of this book, but makes no expressed or
implied warranty of any kind and assumes no responsibility for any errors or omissions. No
liability is assumed for incidental or consequential damages in connection with or arising out of
information contained in this book. The Publisher shall not be liable for any special,
consequential, or exemplary damages resulting, in whole or in part, from the readers’ use of, or
reliance upon, this material. Any parts of this book based on government reports are so indicated
and copyright is claimed for those parts to the extent applicable to compilations of such works.
Independent verification should be sought for any data, advice or recommendations contained in
this book. In addition, no responsibility is assumed by the publisher for any injury and/or damage
to persons or property arising from any methods, products, instructions, ideas or otherwise
contained in this publication.
This publication is designed to provide accurate and authoritative information with regard to the
subject matter covered herein. It is sold with the clear understanding that the Publisher is not
engaged in rendering legal or any other professional services. If legal or any other expert
assistance is required, the services of a competent person should be sought. FROM A
DECLARATION OF PARTICIPANTS JOINTLY ADOPTED BY A COMMITTEE OF THE
AMERICAN BAR ASSOCIATION AND A COMMITTEE OF PUBLISHERS.
Additional color graphics may be available in the e-book version of this book.
LIBRARY OF CONGRESS CATALOGING-IN-PUBLICATION DATA
Fractional calculus in analysis, dynamics, and optimal control / editor, Jacky Cresson (Laboratory
of Applied Mathematics, University of Pau, Pau Cedex, France).
pages cm
Includes index.
ISBN: (eBook)
1. Fractional calculus. I. Cresson, Jacky, editor of compilation.
QA314.C74 2014
515'.83--dc23
2013043481
Published by Nova Science Publishers, Inc. † New York

CONTENTS
Preface vii
Chapter 1 Local Fractional Derivatives 1
N. C. Dias and J. N. Prata
Chapter 2 Fractional Variational Embedding and Lagrangian Formulations 65
of Dissipative Partial Differential Equations
Jacky Cresson
Chapter 3 A Class of Fractional Optimal Control Problems and Fractional 127
Pontryagin's Systems. Variational Integrator and Existence
of Continuous/Discrete Noether's Theorems
Loïc Bourdin
Chapter 4 Fractal Traps and Fractional Dynamics 179
Pierre Inizan
Chapter 5 Numerical Approximations to Fractional Problems 201
of the Calculus of Variations and Optimal Control
Shakoor Pooseh, Ricardo Almeida and Delfim F. M. Torres
Index 241
P REFACE
Fractional calculus has received a big deal of attention in recent years due to its ability to
capture many particular features of Physics, Biology or Economy. Many Textbooks on the
subject are devoted to the theory of fractional calculus or deal with specific applications in
different fields. Mathematical monographs about the subject are most of the time dedicated
to fractional versions of classical mathematical fields like fractional differential equations,
fractional partial differential equations or fractional calculus of variations, etc. However,
the fractional version of many problems is not always supported by any mathematical con-
sideration apart from the generalization of known results. As a consequence, most of the
results in these monographs can not be used to solve classical problems of mathematics and
are of interest only for people working with fractional calculus.
A non exhaustive list of classical mathematical domains where fractional calculus can
be used to obtain new results is:
• Dynamical systems and in particular chaotic Hamiltonian systems
• Analysis of non differentiable functions
• Variational formulations of PDEs or ODEs
None of the previous fields is defined by or concerned with fractional calculus at the
beginning.
The first item was developed by Georges Zaslavsky and Rudolph Hilfer at the begin-
ning of the 1990. We can find a synthese of Zaslavsky’s work in his book Hamiltonian
Chaos and fractional dynamics (Oxford University Press, 2005). We refer to the article R.
Hilfer untitled Foundation of fractional dynamics (Fractals 3(3),549-556,1995) for a good
introduction to his point of view. Despite some progress, the tools and rigorous foundations
of fractional dynamics to study Hamiltonian chaos are far from being satisfying.
The two others items have appeared after the work of K. Kolvankar and A.D. Gan-
gal around 1997 about the local fractional calculus and its relation to analysis of irregular
curves (see K. Kolwankar and A. D. Gangal, Hlder exponents of irregular signals and local
fractional derivatives, Pramana J. phys. 48, No. 1 (1997), 4968) and in 1996 after the sem-
inal work of Fred Riewe about a fractional Lagrangian framework for dissipative systems
viii Jacky Cresson
(see F. Riewe, Nonconservative Lagrangian and Hailtonian mechanics, Phys. Rev. E (3),
53(2), 1890-1899, 1996).
The aim of this book is to provide examples where fractional calculus can be used to
solve classical problems of analysis or at least to provide a new point of view and to present
new results in the theory of fractional optimal control. All the contributions are self contain
and all the proofs are given in details.
The content of this book is as follows:
The first Chapter authored by Da Prata and Pinto deals with analysis and geometry of
non differentiable functions using the so-called local fractional derivative. The author re-
viewed the initial notion defined by K. Kolvankar and A.D. Gangall and the definition given
by F. Ben Adda and J. Cresson. They provide a full comparison of these two notions and
they relate precisely the order of differentiation to the fractal dimension of the underlying
curve as well as its Hölder regularity. They also extend the notion of local differentiation.
The second Chapter authored by Cresson reviews recent works on the use of a fractional
calculus of variations to find Lagrangian variational formulations for dissipative PDEs. This
Chapter also gives a complete introduction to the formalism of embedding which is used to
obtain these results.
The third Chapter authored by Bourdin studies the discrete/continuous fractional opti-
mal control by proving in particular a fractional Pontryagin’s theorem.
The fourth Chapter authored by Pierre Inizan deals with an abstract model associated
to chaotic Hamiltonian systems allowing to prove the emergence of fractional dynamics for
the asymptotic dynamical behaviour of the systems in the neighbourhood of fractal traps.
The construction uses a combination of G. Zaslavsky and R. Hilfer ideas.
The last Chapter authored by Poosed, Almeida and Torres deals more specifically with
numerical simulation of fractional optimal control problem.
All the Chapters contain many examples as well as numerical simulations.
We hope that the reader will find in this book many reasons to study fractional calculus
as well as new ideas and directions.
2 March 2013
Jacky Cresson
Laboratory of Applied Mathematics
University of Pau Avenue University
P. O. Box 1155
64013 Pau Cedex, France
jacky.cresson@univ-pau.fr
In: Fractional Calculus in Analysis, Dynamics ... ISBN: 978-1-62948-635-2
Editor: Jacky Cresson, pp. 1-63 c 2014 Nova Science Publishers, Inc.

Chapter 1
L OCAL F RACTIONAL D ERIVATIVES

N. C. Dias∗ and J. N. Prata†
Departamento de Matemática
Universidade Lusófona de Humanidades e Tecnologias
Lisboa, Portugal
and
Grupo de Fı́sica-Matemática
Univeridade de Lisboa
Lisboa, Portugal
PACS: 05.45-a, 52.35.Mw, 96.50.Fm

Keywords: Fractional calculus, fractals
AMS Subject Classification: 53D, 37C, 65P
1. Introduction
In the course of history, mathematical models of natural phenomena were based on as-
sumptions of regularity and determinism. It was believed that physical phenomena can,
by and large, be represented by analytic functions and the dynamics of physical systems
can be adequately described by differential equations. These beliefs were largely supported
by the successes of physical theories such as Newton’s mechanics, electromagnetic theory,
acoustics, heat transport, along with diffusion, quantum mechanics and relativity in the 20th
century. However, it became clear in more recent years that such phenomena as phase tran-
sitions, turbulence, and the rheology of polymeric materials could not be explained using
such an approach [27, 42]. In Mandelbrot’s own words [25] ”clouds are not spheres, moun-
tains are not cones, coastlines are not circles, and bark is not smooth, nor does lightning
travel in a straight line”.
In his renowned essay [25] Mandelbrot advocated a departure from this tradition of
smoothness, regularity and determinism, to one where roughness, irregularity and random-
∗
E-mail address: ncdias@meo.pt
†
E-mail address: joao.prata@mail.telepac.pt
2 N. C. Dias and J. N. Prata
ness are prevalent. It is not just the fact that natural phenomena are more accurately de-
scribed by the latter characteristics, but also from a mathematical standpoint, sets and struc-
tures which were reputed as exceptional or aberrant should in a sense be the rule. We will
recapitulate two notable theorems which prove that, from a topological point of view, the
sets of continuous functions and of infinitely differentiable functions are much more prof-
ligate at providing functions which are nowhere differentiable or nowhere analytic than the
opposite.
A paramount example of randomness and non-differentiability is Brownian motion.
The typical path of the Brownian particle is nowhere differentiable [27]. Unlike ordinary
Newtonian motion, one cannot write down differential equations for such processes. There
is an old answer to this problem, called fractional calculus. Fractional derivatives generalize
n α
the notion of derivative of order n ∈ N of a function ddxnf to non-integer orders ddxαf , α ∈
R+ . One of the truly remarkable properties of fractional derivatives is the fact that a non-
differentiable function f may nevertheless admit a fractional derivative of some order 0 <
α < 1. It is by now widely acknowledged that there are numerous situations were fractional
differential equations are very effective at describing the situation at hand [42, 44]. It is not
just the fact that non-differentiable (fractal) functions may emerge as solutions to fractional
differential equations, also the highly non-local nature of fractional derivatives is very useful
at addressing systems with long range interactions or prolonged memory effects.
On the other hand, this non-locality prevents the same kind of local description of the
geometry of the graph of a function that one has from traditional calculus of differentiable
functions. This prompted the search for local versions of fractional derivatives. The first
attempts were put forward by Kolwankar and Gangal [17, 18, 19] and several authors then
worked on the main mathematical properties of such derivatives [4, 7, 9] and applications
[21]. The local fractional derivatives are themselves (as are the non-differentiable functions
on which they act) fractal-like. There is a critical order of differentiability below which
the derivative vanishes, and above which it diverges. This is strongly reminiscent of the
Hausdorff-dimension as one computes the Hausdorff measure of a set below and above
its dimension. And indeed the two concepts are inextricably linked. The critical order of
differentiability is related to the fractal dimension of the graph of the function.
A structure theorem due to Chen, Yan and K. Zhang [7] proves that if the local fractional
derivative of some order 0 < α < 1 of a function exists in an interval, then it is zero, except
possibly in a set of vanishing Lebesgue measure. The interest in local fractional derivatives,
then becomes restricted to the following situations.
(i) The function is multi-fractal, and hence the critical order of differentiability varies in
the interval. This weakens the conditions of the Chen-Yan-Zhang Theorem and permits the
fractional derivative (of varying order) to be finite and non-zero. In this context, various
techniques in the framework of wavelet theory have been developed [15].
(ii) The set of zero Lebesgue measure on which the fractional derivative is non-vanishing
cannot be discarded as it has non-vanishing and finite Hausdorff measure of dimension
log 2
0 < s < 1 (e.g. the devil staircase function for the Cantor set for s = log 3 [40]). This
situation could be a physical model for a particle interacting at a set of instants which is
a fractal set (e.g. Cantor’s set). In such a case the particle’s trajectory is certainly not
a straight line, which is what one obtains if one neglects the set of vanishing Lebesgue
Local Fractional Derivatives 3
measure where the interactions occur. For this situation a certain type of fractal calculus
developed by Parvate and Gangal is more useful [32].
Certain fractal sets display a vanishing or divergent Hausdorff measure for all dimen-
sions s > 0. This is because the functions that one uses to ”resolve” these sets - the powers
xs - are not tailored for such sets. It is well known that other functions (called dimension
functions) may be more appropriate. By the same token, the fact that (in accordance with
the Chen-Yan-Zhang Theorem) the local fractional derivatives vanish almost everywhere
may in certain cases just be an indication that fractional derivatives of order α are not ap-
propriate and a fine-tuning using orders other than the powers may be more effective. As
we will see this is what happens for instance in the case of the Takagi function [39].
In this work we will follow this point of view. Our strategy to circumvent the Chen-
Yan-Zhang Theorem is then to develop a local fractional calculus of orders other than the
positive reals. We will develop the main properties of this type of calculus, which we
dubbed ψ-calculus. The fractal nature of the ψ-local fractional derivative manifests itself
again. Indeed, if it vanishes nowhere in an interval, then the set of points where it is positive
and where it is negative are both dense everywhere in that interval. Moreover, if one draws
the graph of the points where the ψ-derivative is positive (or negative) versus the value of
the ψ-derivative one obtains a dust-like struture.
Here is a brief outlook of the paper.
In the next section we dwell on the concept of negligible set from a topological point of
view. We revise the categories of Baire and Baire’s Theorem. We apply Baire’s Theorem in
the proofs of the Banach-Mazurkiewicz Theorem and of the Cater Theorem. The Banach-
Mazurkiewicz Theorem says that the set of functions that admit a finite (right) derivative at
a point in some interval is a negligible subset of the space of continuous functions on that
interval. We consider the Weierstrass and the Takagi functions as examples of functions
which are continuous but nowhere differentiable in an interval. The Cater Theorem, on the
other hand, states that the set of functions which are analytic somewhere in an interval are
a negligible subset of the space of infinitely differentiable functions on that interval. We
also give an example by Cater of a function which is infinitely differentiable but nowhere
analytic in an interval.
In section 3 we develop the concept of dimension function, a generalization of the topo-
logical and the fractal notions of dimension. As an application, we discuss various types of
Hölder continuity with respect to dimension functions. The Takagi function constitutes an
example of a function displaying a Hölder continuity with respect to one of such functions.
The Hausdorff measure and Hausdorff dimension are then presented in connection with the
dimension functions. The main result here is the relation between the Hölder behavior of a
function and the fractal dimension of its graph.
Section 4 presents our main results: a generalization of local fractional derivatives by
using generalized dimension functions as the order of differentiability instead of the more
widely used powers. We develop the complete calculus for such derivatives.
Finally in section 5, we consider some possible generalizations, such as Cresson and
Nottale’s quantum difference operators, the local fractional calculus in several dimensions,
and a generalization of Clarke’s derivative to fractional order.
Notation 1. We use the notation Vξ+ (ǫ) = [ ξ, ξ + ǫ ), Vξ− (ǫ) = ( ξ − ǫ, ξ ], and Vξ (ǫ) =
Vξ+ (ǫ) ∪ Vξ− (ǫ) = (ξ − ǫ, ξ + ǫ), for ǫ > 0.
The vector space of continuous real functions on an interval I - denoted C 0 (I) - is a
normed space with respect to the sup-norm
||f ||0 := sup |f (x)|. (1)

x∈I
Let U be any non-empty subset of a metric space Ω with metric d. The diameter of U
is defined as |U | := sup {d(x, y) : x, y ∈ U }. The ball of radius r > 0 and center x ∈ U
is B(x, r) := {y ∈ U : d(x, y) < r}.
We denote by int(M ) the set of interior points of some subset M of a topological space
X and by M its closure. The complement of M in X is M c = X\M .
2. Non-differentiable and Non-analytic Functions

2.1. Baire Categories. Baire’s Theorem
From the mathematical point of view there are various ways of assessing whether a certain
set or property can be deemed as negligible. For instance, one may say that a finite subset
of the integers is negligible in the set of integers, since the cardinality of Z is infinity.
From the point of view of cardinality, Z is not a negligible subset of Q since both are
countable. On the other hand, Q is a negligible subset of R both from the point of view of
cardinality as well as from the measure/probability point of view. Indeed, the measure
(see below) of Q is zero. One could therefore say that the probability of picking a rational
number randomly in some interval is zero. This may not come as a surprise, given that
Q is countable. Nevertheless, there are uncountable subsets of R which have equally a
vanishing (Lebesgue) measure. The most famous example is no doubt Cantor’s ternary set
C. If we consider instead a Hausdorff measure of suitable dimension (log 2/ log 3), then
Cantor’s set no longer has vanishing measure. In spite of this, after a simple inspection, one
realizes that Cantor’s set is rarefied in R. From the topological point of view it is nowhere
dense. In this case, measure and cardinality do not adequately describe this sparseness, as
C is not negligible from these points of view. If a metrizable topology is at our disposal in
some larger set X, then one can characterize this latter notion of a subset M ⊂ X being
negligible, by saying that M is nowhere dense in X. In the case of Cantor’s set, one could
express this idea by saying that there is no interval, where all the elements of this interval
could be approximated with arbitrary precision by elements of C.
These considerations lie at the heart of Baire’s classification of sets into categories, de-
scribing whether a subset M of a metric space is a countable union of nowhere dense sets, or
if the opposite holds. Within this framework we will reproduce here famous theorems which
show that differentiable functions are negligible in the larger space of continuous functions
and likewise that analytic functions are a negligible subset of the space of infinitely differen-
tiable functions. From the point of view of mathematical modeling of natural phenomena,
one could thus be tempted to say that over the years one has been resorting to a negligible
sector of the tool kit at our disposal.
Let us then establish the relevant definitions.
Definition 2. Let X be a metric space. A subset M ⊂ X is said to be rare (or nowhere

dense) if its closure M has no interior points. It is said to be meager or of the first category
if it is the union of countably many sets each of which is rare in X. If M is not meager it is
said to be nonmeager in X or of the second category. If the complement of M is meager,
then M is called comeager or residual. By abuse of language we will sometimes denote
by negligible set a meager set.
One way of assessing whether a set is of first category is through the following lemma.
Lemma 3. A subset M of a metric space X is meager if and only if M ⊂ ∪n∈N Mn , with

Mn closed and with empty interior for all n ∈ N.
Proof. Suppose that M is meager. Then M = ∪n∈N Fn , where int(Fn ) = ∅. Let Mn := Fn

for all n ∈ N. Then Fn is closed and int(Fn ) = ∅. Since Fn ⊂ Fn = Mn , we conclude
M ⊂ ∪n∈N Mn .
Conversely, suppose M ⊂ ∪n∈N Mn , with Mn = Mn and int(Mn ) = ∅. Define
Fn = Mn ∩ M for n ∈ N. We then have int(Fn ) ⊂ int(Mn ) = ∅ and M = ∪n∈N Fn .
Consequently, M is meager in X.
To prove Baire’s Theorem (Theorem 5. below), we need the following preliminary result
due to Cantor.
Lemma 4. (Cantor) Let X be a complete metric space and {Fn }n∈N a nested collection
of nonempty closed subsets of X such that X ⊃ F1 ⊃ F2 ⊃ · · · and |Fn | → 0 as n → ∞.
Then ∩n∈N Fn 6= ∅.
Proof. For each n ∈ N let us choose xn ∈ Fn . Since F1 ⊃ F2 ⊃ · · · , we have xn ∈ Fk ,

∀n ≥ k. By assumption, for each ǫ > 0, there exists k ∈ N such that |Fk | < ǫ. Since, for
n, m ≥ k, we have xn , xm ∈ Fk , we conclude that d(xn , xm ) ≤ |Fk | < ǫ. In other words,
(xn )n∈N is a Cauchy sequence in X. Given that X is complete there exists x ∈ X such that
xn → x as n → ∞. As all term of the sequence, after a certain order, belong to Fk and Fk
is closed, we conclude that x ∈ Fk , that is x ∈ ∩n∈N Fn .
There are various equivalent ways of stating Baire’s Theorem. A familiar formulation
says that a complete metric space is of second category in itself. Here we shall consider the
following convenient version of Baire’s Theorem.
Theorem 5. (Baire) Let X be a complete metric space and B1 , B2 , B3 , · · · a countable

collection of closed subsets of X with no interior points. Then B = ∪n∈N Bn has empty
interior as well.
Proof. To prove the theorem, we shall demonstrate an equivalent statement, namely that
any countable intersection of open, dense subsets of X is dense in X.
Let then A1 , A2 , A3 , · · · be a countable collection of open and dense subsets of X. We
want to prove that A = ∩n∈N An is dense in X, that is, for any nonempty open subset U of
X, U ∩A 6= ∅. Since A1 is open and dense in X, A1 ∩U 6= ∅. Moreover, A1 ∩U is open, as
it is the intersection of two open sets. We can then find another open set U1 ⊂ A1 ∩ U , such
that U1 ⊂ A1 ∩ U and |U1 | ≤ 1. To justify this, let a ∈ A1 ∩ U . There exists ǫ > 0 such that
B(a, ǫ) ⊂ A1 ∩ U . Obviously, we can then find another open ball centered at a of radius
δ ≤ 1 contained in B(a, ǫ). Let U1 = B(a, δ/2). We have: B(a, δ/2) ⊂ B(a, δ/2) ⊂
B(a, δ) ⊂ B(a, ǫ) ⊂ A1 ∩ U . Moreover, |U1 | = |B(a, δ/2)| = 2 · δ/2 = δ ≤ 1.
Similarly, since A2 is open and dense in X, A2 ∩ U1 6= ∅ and A2 ∩ U1 is an open
set. We can thus find another open set U2 ⊂ A2 ∩ U1 such that U2 ⊂ A2 ∩ U1 and
|U2 | ≤ 21 . By induction, we have a collection of closed sets Un 6= ∅, n ∈ N such that
U ⊃ U1 ⊃ U2 ⊃ U3 ⊃ · · · and Un → 0 as n → ∞. By Cantor’s Lemma, there
exists x ∈ ∩n∈N Un . In particular, x ∈ U1 and thus x ∈ U and x ∈ A1 . Moreover,
Un+1 ⊂ An+1 ∩ Un , ∀n ∈ N and thus x ∈ An , ∀n ∈ N. Consequently, x ∈ U ∩ A.
2.2. Continuous Nowhere Differentiable Functions

The following theorem due to Banach and Mazurkiewicz conveys the idea that the set of
functions which are differentiable at some point in an interval are, from a topological point
of view, a negligible subset of the space of continuous functions on that interval.
Theorem 6. (Banach, Mazurkiewicz) Let M be the set of functions with a finite right-
derivative at some point in [ a, b ) regarded as a subset of the metric space C 0 [a, b]. Then
M is meager in C 0 [a, b].
Proof. Let us denote by {Mn }n∈N the collection of subsets of C 0 [a, b]

n h i
Mn := f ∈ C 0 [a, b] : ∃t ∈ a, b − (b−a)
n such that
o (2)
f (t+h)−f (t) b−a
h ≤ n, for all h ∈ ( 0, n ] .
If g ∈ M , then g ∈ Mn for some (sufficiently large) n ∈ N, and thus M ⊂ ∪n∈N Mn . To

justify this claim, suppose limh→0+ g(t+h)−g(t)
h = g ′ (t+ ) ∈ R for some t ∈ [ a, b ). If we
fix K > |g ′ (t+ )|, there exists δ > 0 such that t + δ ≤ b and

g(t + h) − g(t)
≤ K, (3)
h
for all h ∈ ( 0, δ ]. If we choose n ∈ N such that n > K and b−an ≤ δ, then g ∈ Mn .

0
Since C [a, b] is complete, the result follows by Baire’s Theorem, if we prove that Fn
is closed with no interior points for all n ∈ N.
1) We prove that Fn is closed. Let f ∈ Fn . Then, there exists a sequence (fk )k∈N in
fk (t+h)−fk (t)
Fn such that fk → f uniformly h as k → i∞. Now let gk (t, h) = h , and let
f (t+h)−f (t) (b−a)
g(t, h) = h for t ∈ a, b − n and h ∈ ( 0, b−an . Since fk ∈ Fn , there exists
h i
tk ∈ a, b − (b−a) , such that |gk (tk , h)| ≤ n for all h ∈ ( 0, b−a

n n . Given the uniform con-
h i
vergence of (fk )k∈N , gk (t, h) also converges uniformly to g(t, h) in a, b − (b−a) n for fixed
h. On the other hand, the

h sequence (t
i k )k∈N , being bounded, has a convergent subsequence
(b−a)
tkm → c for some c ∈ a, b − n . We thus have
|gkm (tkm , h) − g(c, h)| ≤ |gkm (tkm , h) − gkm (c, h)| + |gkm (c, h) − g(c, h)| (4)
Since gk and g are continuous functions and gk (t, h) → g(t,h) uniformly, we conclude that
gkm (tkm ) → g(c, h) as km → ∞. But, for all h ∈ ( 0, b−a n , we have |gk (tk , h)| ≤ n, and
f (c+h)−f (c)
≤ n for all h ∈ ( 0, b−a

thus |g(c, h)| ≤ n, that is h n . In other words, f ∈ Fn
and Fn is closed.
2) Next we prove that int(Fn ) = ∅. We want to show that for all f ∈ Fn and any ǫ > 0,
B(f, ǫ) ∩ Fnc 6= ∅.
By Weierstrass’ approximation theorem [14], let p(x) be a polynomial on [a, b] such
that d(f, p) < ǫ/2. Let S denote the maximum slope of p(x) on [a, b], i.e.:

p(x) − p(y)
S := max : x, y ∈ [a, b] . (5)
x−y
We shall also consider a continuous function q(x) on [a, b] consisting of straight-line seg-
ments of slopes ±(S + n + 1) constrained so that |q(x)| < ǫ/2 for all x ∈ [a, b]. Define the
function g(x) = p(x) + q(x). It is straightforward to conclude that g ∈ B(f, ǫ). Indeed:
d(f, g) ≤ d(f, p) + d(p, g) = d(f, p) + ||q||h < 2ǫ + 2ǫ = iǫ.
On the other hand, we have for any t ∈ a, b − (b−a)
n :

g(t+h)−g(t) p(t+h)+q(t+h)−p(t)−q(t)
h =
h ≥
(6)
≥ q(t+h)−q(t)
p(t+h)−p(t)
h − h
One can then find a (sufficiently small) h ∈ ( 0, b−a

n ] for which the right-hand side is
≥ S + n + 1 − S = n + 1 > n and thus g 6= Fn .
Remark 7. Notice that if f admits a finite right-derivative at some ξ ∈ [ a, b ), then g(x) =
f (a + b − x) has a finite left-derivative at a + b − ξ and vice-versa. Consequently, the set
of functions with a finite left-derivative at some ζ ∈ ( a, b ] is equally of first category in
C 0 [a, b]. Finally, the set of functions which are differentiable at some point in the interval
is a subset of the former two sets, and hence must be of first category.
Before we end this section, we shall present two celebrated functions which are continu-
ous in an interval but nowhere differentiable in that interval: the Takagi and the Weierstrass
functions.
Takagi’s function was introduced in [39]. A recent survey of its properties and applica-
tions is [1].
Definition 8. Let φ(x) = dist(x, Z) denote the distance from x ∈ R to the nearest integer.
Then the Takagi function is defined on [0, 1] by:
∞
X 1
T (x) := n
φ(2n x). (7)
n=0
2
The continuity of T is easily established.
Lemma 9. Takagi’s function is continuous in [0, 1].
Proof. Notice that 0 ≤ φ(x) ≤ 12 for any x ∈ R. Consequently:

∞
∞
X 1
n
X 1
φ(2 x) ≤ <∞ (8)

2n 2n+1
n=0 n=0
for all x ∈ [0, 1], which establishes the uniform convergence of T in [0, 1]. Since φ is
continuous, the result follows.
Proposition 10. (Takagi) The Takagi function does not possess a finite derivative at any
point.
Proof. We will reproduce here the simple proof due to Billingsley [5]. Let Z+ denote the
set of non-negative integers. A dyadic rational of order m is a rational number of the form
x = 2km with k ∈ Z+ and m ∈ N. Put φk (x) = 2−k φ(2k x) for k = 1, 2, · · · . Fix a point x,
and, for each n ∈ N, let un and vn be dyadic rationals of order n with vn − un = 2−n and
un ≤ x < vn . Then:
n−1
T (vn ) − T (un ) X φk (vn ) − φk (un )
= , (9)
v n − un v n − un
k=0
since φk (vn ) = φk (un ) = 0 for all k ≥ n. But for k < n, φk is linear on [uk , vk ] with
slope φ′k (x+ ), the right-derivative of φk at x. Thus
n−1
T (vn ) − T (un ) X ′ +
= φk (x ). (10)
v n − un
k=0
Since φ′k (x+ ) = ±1 for each k, this last sum cannot converge to finite limit as n → ∞.
However, if T were differentiable at x, then (10) would have to be a convergent sequence,
and we have a contradiction. Hence T cannot have a finite derivative at x. Since x was
chosen arbitrarily, the result follows.
It will also be relevant for the sequel to consider the Hölder behavior of the Takagi
function. The following Proposition is due to Anderson and Pitt [3], and was proved in the
more general context of functions belonging to the so-called Zygmund space Λ∗d .
Proposition 11. (Anderson, Pitt) There exists a constant C > 0 such that, for all x ∈
(0, 1) and sufficiently small h 6= 0:
|T (x + h) − T (x)| ≤ C|h| log(1/|h|). (11)

T (x+h)−T (x)
It is also interesting to enquire about the existence of the limit limh→0 −h log2 (|h|) .
The following theorem is due to Krüppel [23].
Theorem 12. (Krüppel) Let x ∈ [0, 1] be a dyadic rational. Then:
T (x + h) − T (x)
lim = 1. (12)
h→0 1
|h| log2 |h|
Another characterization proved by Allaart and Kawamura [1] requires the following
definitions.
Definition 13. The binary expansion of a point x ∈ [ 0, 1 ) is denoted by

∞
X ǫn
x= = 0.ǫ1 ǫ2 · · · ǫn · · · , ǫn ∈ {0, 1} (13)
n=1
2n
where, obviously ǫn = ǫn (x). If the limit

n
1X
d1 (x) := lim ǫk (14)
n→∞ n
k=1
exists, we call d1 (x) the density (or long-time frequency) of the digit ”1” in the binary
expansion of x. In that case, the number
d0 (x) := 1 − d1 (x) (15)
is the density of the digit ”0”.
Definition 14. Let x ∈ [0, 1] be non-dyadic, and let (an )n∈N and (bP n )n∈N be the (unique)
∞ −an , 1 − x =
strictly
P∞ −bn increasing sequences of positive integers such that x = n=1 2
n=1 2 .
We say that x is density-regular if d1 (x) exists and one of the following holds:
(a) 0 < d1 (x) < 1; or

an+1
(b) d1 (x) = 0 and an → 1; or
bn+1
(c) d1 (x) = 1 and bn → 1.
Theorem 15. (Allaart, Kawamura) Let ∈ [0, 1] be non-dyadic. The limit in (12) exists if
and only if x is density-regular, in which case the limit is equal to d0 (x) − d1 (x).
The Weierstrass function was the first function to be proved to be continuous and
nowhere differentiable [41]. For a while it was regarded as a mathematical curiosity, but in
his 1926 investigation of turbulence, Richardson [34] observed that the velocity field of the
atmospheric wind is so erratic that it probably cannot be described by any analytic function
or indeed a differentiable function. He suggested Weierstrass’ function as a candidate to
represent the velocity field, since its properties of continuity and non-differentiability were
observed in the wind data. We have, since that time, come to realize that Richardson’s
intuition was superior to a half century of analysis regarding the nature of turbulence.
Definition 16. Fix λ > 1 and 1 < s < 2. Define the Weierstrass function Wλ,s : [0, 1] →
R by:
∞
X
Wλ,s (x) := λ(s−2)k sin(λk x). (16)
k=1
Lemma 17. The Weierstrass function is continuous in [0, 1].
Proof. Again the continuity is an immediate consequence of the uniform convergence of

the series in [0, 1]:
∞
∞

X X
λ(s−2)k sin(λk x) ≤ λ(s−2)k < ∞, ∀x ∈ [0, 1] . (17)

k=1 k=1
The following proposition expresses the Hölder and anti-Hölder behavior of the Weier-
strass function.
Proposition 18. Let Wλ,s denote the Weierstrass function for 1 < s < 2, λ > 1. Then:
(i) there exists a constant C1 > 0 such that
1
|Wλ,s (x + h) − Wλ,s (x)| ≤ C1 |h|2−s , for all h with |h| < ; (18)
λ
(ii) there exist positive constants C2 , δ0 with the following property: for each x ∈ [0, 1] and
0 < δ ≤ δ0 there exists y such that |x − y| ≤ δ and
|Wλ,s (x) − Wλ,s (y)| ≥ C2 δ 2−s , (19)
for large enough λ.
Proof. The proof presented here is given in Falconer’s book [12].
(i) Given 0 < |h| < λ−1 , let N ∈ N such that
λ−(N +1) ≤ |h| < λ−N . (20)
Then
PN (s−2)k sin λk (x + h) − sin(λk x) +

|Wλ,s (x + h) − Wλ,s (x)| ≤ k=1 λ
P∞ (s−2)k sin λk (x + h) − sin(λk x) ≤

+ k=N +1 λ (21)
PN (s−2)k λk |h|
P∞ (s−2)k
≤ k=1 λ + k=N +1 2λ
using that | sin(u) − sin(v)| ≤ |u − v| (a consequence of the mean-value theorem) on the

first N terms of the sum, and that | sin(u) − sin(v)| ≤ | sin(u)| + sin(v)| ≤ 2 on the
remaining terms. Summing these geometric series, we obtain:
|h|λ(s−1)N 2λ(s−2)(N +1)

|Wλ,s (x + h) − Wλ,s (x)| ≤ + ≤ c|h|2−s (22)
1 − λ1−s 1 − λs−2
where c is independent of h and we used (20).
(ii) In the same way, but splitting the sum into three parts - the first (N − 1) terms, the N th
term, and the rest - we get that:
Wλ,s (x + h) − Wλ,s (x) − λ(s−2)N sin λN (x + h) − sin(λk x) ≤

(23)
λ(s−2)N −s+1 2λ(s−2)(N +1)
≤ 1−λ1−s + 1−λs−2 ,
where we used (20) again.

(s−2)N
Suppose λ > 2 is large enough for the right-hand side of (23) to be less than λ 20
for all N . For δ < λ−1 , take N such that λ−N ≤ δ < λ−N +1 . For each
x ∈ [0, 1], we
may1
choose h with λ−N −1 ≤ |h| < λ−N < δ such that sin λN (x + h) − sin(λk x) > 10 .
So by (23)
|Wλ,s (x + h) − Wλ,s (x)| − λ(s−2)N sin λN (x + h) − sin(λk x) ≤

Wλ,s (x + h) − Wλ,s (x) − λ(s−2)N sin λN (x + h) − sin(λk x) ≤ λ(s−2)N

20
λ(s−2)N
⇒ |Wλ,s (x + h) − Wλ,s (x)| ≥ λ(s−2)N sin λN (x + h) − sin(λk x) −

20
λ(s−2)N λ(s−2)N λ(s−2)N λs−2 δ 2−s

≥ 10 − 20 = 20 ≥ 20 ,
(24)
which concludes the proof.
The non-differentiability of Wλ,s is then a trivial consequence of Proposition 18. (ii).
Corollary 19. Let λ > 1 and 1 < s < 2. Then the Weierstrass function Wλ,s is nowhere
differentiable in the interval [0, 1].
Proof. From Proposition 18. (ii), for each x ∈ [0, 1] and sufficiently large n ∈ N, we can
find yn ∈ [0, 1] \ {x} such that
1
|x − yn | ≤ , (25)
n
and 2−s
1
|Wλ,s (x) − Wλ,s (yn )| ≥ C2 . (26)
n
From (25) and (26), we then have

Wλ,s (x) − Wλ,s (yn )
≥ C2 ns−1 . (27)
x − yn
The right-hand side of the previous inequality diverges as n → ∞ and thus Wλ,s cannot be
differentiable at x.
2.3. Infinitely Differentiable, Nowhere Analytic Functions

By the same token, one can prove that the set of analytic functions is a negligible subset of
the set of infinitely differentiable functions. We will follow closely [6].
Let C ∞ [a, b] denote the set of functions which admit continuous derivatives of all or-
ders in [a, b]. The set C ∞ [a, b] is a subspace of C 0 [a, b] which is incomplete with respect
to the sup-norm of C 0 [a, b]. Indeed, the set of polynomials on [a, b] is a subset of C ∞ [a, b]
and, by Weierstrass’ approximation theorem [14], dense in C 0 [a, b]. Hence a sequence of
polynomials can converge to, say, a continuous non-differentiable function.
We thus need an alternative metric in C ∞ [a, b] to ensure completeness. For f1 , f2 ∈
C ∞ [a, b], we define
∞ (k) (k)
X 1 |f1 (x) − f2 (x)|
dF (f1 , f2 ) := sup (28)
2k a≤x≤b 1 + |f (k) (x) − f (k) (x)|
k=0 1 2
with the convention f (0) = f .
Proposition 20. The set F = (C ∞ [a, b] , dF ) is a complete, separable metric space.
Proof. The proof proceeds in several steps.

t
1) We start by showing that dF is a metric. First of all, since 0 ≤ 1+t ≤ 1 for all t ≥ 0, we
conclude that:
∞
X 1
0 ≤ dF (f1 , f2 ) ≤ <∞ (29)
2k
k=0
for all f1 , f2 ∈ C ∞ [a, b]. Clearly, dF (f1 , f2 ) = 0 if and only if f1 = f2 and dF (f1 , f2 ) =
t
dF (f2 , f1 ). It remains to prove the triangle inequality. Let g(t) = 1+t for t ≥ 0. Since
′ 1
g (t) = (1+t)2 > 0 for all t ≥ 0, we conclude that g is strictly increasing. Since for any
u, v ∈ R, |u + v| ≤ |u| + |v|, it follows that g(|u + v|) ≤ g(|u| + |v|) and thus:
|u + v| |u| + |v| |u| |v| |u| |v|

≤ = + ≤ + . (30)
1 + |u + v| 1 + |u| + |v| 1 + |u| + |v| 1 + |u| + |v| 1 + |u| 1 + |v|
Consequently:
(k) (k) (k) (k)
f1 (x)−f3 (x) + f3 (x)−f2 (x)

(k) (k)

(k) (k)
≤
1+ f1 (x)−f3 (x) + f3 (x)−f2 (x)

(31)
(k) (k) (k) (k)
f1 (x)−f3 (x) f3 (x)−f2 (x)

≤
(k) (k)
+
(k) (k)

1+f1 (x)−f3 (x) 1+f3 (x)−f2 (x)

and dF (f1 , f2 ) ≤ dF (f1 , f3 ) + dF (f3 , f2 ).
2) Let us now prove that F is complete. Let (fn )n∈N denote a Cauchy sequence in F , that
is limn,m→∞ dF (fn , fm ) = 0. Then
(k) (k)
|fn (x) − fm (x)|
sup (k) (k)
≤ 2k dF (fn , fm ) (32)
a≤x≤b 1 + |fn (x) − fm (x)|
(k)
Thus, for each fixed k ∈ N, the sequence of continuous functions (fn )n∈N converges
uniformly on [a, b], say to the continuous function gk . Then:
lim sup |fn(k) (x) − gk (x)| = 0, (33)

n→∞ a≤x≤b
and
(k)
|fn (x) − gk (x)|
lim sup (k)
= 0, (34)
n→∞ a≤x≤b 1 + |fn (x) − gk (x)|
for each k ≥ 0. By the uniform convergence, it is legitimate to interchange the limit with
the summation and obtain:
∞ (k)
X 1 |fn (x) − gk (x)|
lim sup = 0. (35)
n→∞ 2k a≤x≤b 1 + |fn(k) (x) − gk (x)|
k=0
(k)
It remains to prove that g0 ∈ F and g0 = gk for all k = 0, 1, 2, 3, · · · by induction on k.
(N −1)
For k = 0, there is nothing to prove. We assume g0 = gN −1 for some fixed N ≥ 1.
Clearly, the sequence (fn′ )n∈N is also a Cauchy sequence in F , because dF (fn′ , fm ′ ) ≤
′(k)
2dF (fn , fm ) by the definition of dF . Moreover, (fn )n∈N converges uniformly to gk+1 for
(N ) ′(N −1)
k ≥ 0. By the induction hypothesis g0 = g0 = gN . This completes the induction.
Thus g0 ∈ F and limn→∞ dF (fn , g0 ) = 0 follows from (33).
3) We finally prove the separability of F . Our aim will be to show that the set of polynomials
in x with rational coefficients is dense in F . For the sake of simplicity, we shall now assume
that the functions of F are defined on the interval [0, 1] (something which can be obtained
by a straightforward translation and dilation).
Take any ǫ > 0 and any f ∈ F . We prove by induction on n that for each n ≥ 0 there
is a polynomial p with rational coefficients such that
n
X
(i)
f (x) − p(i) (x) < ǫ, (36)

i=0
for 0 ≤ x ≤ 1.
For n = 0, we use Weierstrass’ approximation theorem [14] a find a polynomial q(x)
such that
ǫ
|f (x) − q(x)| < , (37)
4
for 0 ≤ x ≤ 1. We then approximate the coefficients of q with rational numbers to find the
appropriate polynomial p such that
ǫ
|p(x) − q(x)| < , (38)
4
for 0 ≤ x ≤ 1. And hence:
ǫ
|f (x) − p(x)| < , (39)
2
for 0 ≤ x ≤ 1.
Now assume that p can be found for some N ≥ 0. But, since f ′ ∈ F , there is a
polynomial p0 (x) with rational coefficients such that:
N ǫ
(i)
X (i+1)
f (x) − p0 (x) < , (40)

4
i=0
for 0 ≤ x ≤ 1.
1
− ξ < 4ǫ . Put

Let ξ ∈ Q such that f 2
Z x
p(x) = ξ + p0 (t)dt, (41)
1
2
for 0 ≤ x ≤ 1. Then
Rx Rx
1
f ′ (t)dt

|p(x) − f (x)| = ξ + 1 p0 (t)dt − f

2 − 1 ≤
2 2
(42)
1
R x
ǫ ǫ ǫ
+ 1 |p0 (t) − f ′ (t)| dt <

≤ ξ − f 2 4 + 4 = 2,
2
(i)
for 0 ≤ x ≤ 1. So p′ and p(i+1) = p0 and
PN +1 (i) (i)

i=0 f (x) − p (x) = |f (x) − p(x)| +

(43)
+ N f (i+1) (x) − p(i+1) (x) < ǫ ǫ ǫ
P
i=0 4 + 4 = 2,
for 0 ≤ x ≤ 1, which completes the induction proof on n.

Let n be chosen so that 2−n < ǫ. Then:
Pn 1 |f (k) (x)−p(k) (x)|
d(f, p) = k=0 2k sup0≤x≤1 1+|f (k) (x)−p(k) (x)|
+
P∞ 1 |f (k) (x)−p(k) (x)|

+ k=n+1 2k sup0≤x≤1 1+|f (k) (x)−p(k) (x)|
≤
(44)
Pn f (x) − p(k) (x) + P∞
(k) 1
≤ k=0 sup0≤x≤1 k=n+1 2k <
1
<ǫ+ 2n < ǫ + ǫ = 2ǫ
We next proceed to show that the set of functions which are analytic in some interval is
of first category in F .
For future reference, let us first define the following element of D, the Schwartz space
of functions in C ∞ with compact support in some interval [c, d]. Let then
 −(x−c)−2 (x−d)−2
 e , if c ≤ x ≤ d
h(x) = (45)
0

otherwise
This function will be dubbed a spike on (c, d). By a simple inspection, one realizes that
h does indeed admit an infinite number of continuous derivatives in R with h(k) (x) = 0
whenever x ≥ d or x ≤ c. Consequently, h cannot be analytic on any interval centered
around c or d.
Theorem 21. (Cater) The set F0 of nowhere analytic functions on [a, b] forms a residual
subset of F .
Proof. If a function is analytic in some interval, then it is analytic in every other subinterval.
It thus suffices to prove that the set of functions that are the limit of their Taylor series in
x − x0 on an interval I centered at x0 , with rational endpoints, is a first category set in F .
Fix an interval I ⊂ [a, b] with rational endpoints and center x0 .
For each j ∈ N, let:
( j )
X f (k) (x )
0
Uj := f ∈ F : (ξ − x0 )k − f (ξ) > 1 for some ξ ∈ I . (46)

k!
k=0
1) The set Uj is open. Indeed, let f ∈ Uj and define:

j
X f (k) (x )
0 k
ǫ = −1 + (ξ − x0 ) − f (ξ) (47)

k!
k=0
ea−b ǫ/2

where ξ is as in the definition of Uj . Now let f1 ∈ B f, 2−j 1+e a−b ǫ/2 . Then, by the
definition of dF , we have:

(k) (k)
(x0 )−f1 (x0 )

1 f
2k 1+f (k) (x0 )−f (k) (x0 )
≤
1

(48)
(m) (m)
(x)−f1 (x)

P∞ 1 f 1 ea−b ǫ/2
≤ m=0 sup a≤x≤b 2m (m)
< 2j 1+ea−b ǫ/2
1+f (m) (x)−f1 (x)

for all k ≥ 0. In particular, for 0 ≤ k ≤ j, we have:

(k) (k)
f (x0 ) − f1 (x0 )

ea−b ǫ/2
(k) (k)
ǫ
< f (x ) f (x ) < ea−b . (49)

a−b
⇔ 0 − 1 0
(k) 1 + e ǫ/2 2
(k)
1 + f (x0 ) − f1 (x0 )

Consequently

(k)
P j f (k) (x0 ) Pj f1 (x0 )
)k )k

k=0 k! (ξ − x0 − k=0 k! (ξ − x0 ≤
(50)

Pj (k) |ξ−x0 |k (b−a)k
< ea−b 2ǫ jk=0 < 2ǫ .
(k) P
≤ k=0 f (x 0 ) − f 1 (x )
0 k! k!
From (48), we also have:

ǫ ǫ
|f (x) − f1 (x)| < ea−b < , (51)
2 2
for all a ≤ x ≤ b. Consequently, from (49) and (50), we have:

Pj f (k) (x0 ) k

k=0 k! (ξ − x0 ) − f1 (ξ) =
1

(k)
f1 (x0 ) f (k) (x0 )
= jk=0
Pj
− x0 )k − − x0 )k +
P
k! (ξ k=0 k! (ξ

Pj f (k) (x0 )
+ (ξ − x0 )k − f1 (ξ) ≥

k=0 k!
P
f (k) (x0 )
≥ jk=0 (ξ − x0 )k − f (ξ) + f (ξ) − f1 (ξ) (52)

k!

(k)
f1 (x0 ) Pj f (k) (x0 )
− jk=0 )k )k
P
k! (ξ − x0 − k=0 k! (ξ − x0 >
P
f (k) (x0 )
> jk=0 (ξ − x0 )k − f (ξ) − |f (ξ) − f1 (ξ)| − ǫ
>

k! 2
P
f (k) (x0 )
> jk=0 (ξ − x0 )k − f (ξ) − ǫ = 1

k!
ea−b ǫ/2

and thus f1 ∈ Uj . Since f1 ∈ B f, 2−j 1+e a−b ǫ/2 was chosen arbitrarily, we conclude
a−b

e ǫ/2
that B f, 2−j 1+e a−b ǫ/2 ⊂ Uj , and Uj is open.
2) We now prove that for any fixed J ∈ N, the set ∪∞ j=J Uj is dense in F . Take any
f ∈ F and ǫ1 > 0. Let g be the spike on some interval (x0 , x0 + δ) ⊂ I. Let c be a
δ

number such that cg x0 + 2 > 2. Let p(x) be some polynomial in the countable and
dense subset of F mentioned in the proof of Proposition 20. such that dF (p, cg) < ǫ1 . Put
h(x) = f (x) + cg(x) − p(x) for a ≤ x ≤ b. Then, h ∈ B(f, ǫ1 ). Indeed:
dF (f, h) = dF (cg, p) < ǫ1 . (53)
Next, we notice that for large enough j ≥ J, p(j) = 0 and

j
X p(k) (x0 )
(x − x0 )k = p(x), (54)
k!
k=0
for all a ≤ x ≤ b. Moreover, since g is spike in (x0 , x0 + δ), g (j) (x0 ) = 0 for all j ∈ N
and
j
X g (k) (x0 )
(x − x0 )k = 0, (55)
k!
k=0
identically.
It follows that
j j
h(k) (x0 ) δ k f (k) (x0 ) δ k

X X δ
= − p x0 + . (56)
k! 2 k! 2 2
k=0 k=0
Thus for such j:

Pj h(k) (x0 ) δ k
δ

k=0 k! 2 − h x0 + 2 =
(57)
Pj f (k) (x0 ) δ k
δ
δ

= k=0 k! 2 − f x0 + 2 − cg x0 + 2 .
Given that, by assumption cg x0 + 2δ > 2, the previous equality can only hold provided

either j
X f (k) (x ) δ k
0 δ
− f x0 + >1 (58)

k! 2 2

k=0
or j
X h(k) (x ) δ k
0 δ
− h x0 + > 1. (59)

k! 2 2

k=0
In other words, either f ∈ ∪∞ j=J Uj or h ∈ ∪∞

j=J U
∞
j . Hence, ∪j=J Uj is an open and dense
subset of F . By Baire’s Theorem ∩∞ ∞

J=1 ∪j=J Uj is a residual set. Finally, the set of all

functions in F analytic on I is disjoint from ∩∞ J=1 ∪ ∞ U
j=J j and must be of first category.
Since there are only countably many intervals I ⊂ [a, b] with rational endpoints, it follows
that F \F0 is a first category subset of F .
We shall now construct a family of nowhere analytic functions in [a, b]. But first we
prove a lemma about subintervals of (a, b). Let {r1 , r2 , r3 , · · · , rn , · · · } be the set of ratio-
nal numbers in (a, b) enumerated.
Lemma 22. There exists a sequence of irrational positive numbers (vn )n∈N , such that vn <
b−a vi
n and [rn − vn , rn + vn ] ⊂ (a, b) for all n ∈ N, and none of the points ri , ri + 2 , ri + vi
(i = 1, 2, · · · , j − 1) lie in the closed interval [rj − vj , rj + vj ] for any j ≥ 2.
Proof. We proceed by induction. For n = 1, let v1 be any irrational number such that
[r1 − v1 , r1 + v1 ] ⊂ (a, b) and hence v1 < b − a. Now suppose that v1 , · · · , vN −1 have
been suitably selected for some N > 1. Then rN is not any of the rational numbers
v
r1 , · · · , rN −1 and is certainly not any of the irrational numbers r1 + v21 , · · · , rN −1 + N2−1 ,
r1 + v1 , · · · , rN −1 + vN −1 . It is hence possible to select a positive irrational number
vN such that vN < b−a N , [rN − vN , rN + vN ] ⊂ (a, b), and such that the closed interval
v
[rN − vN , rN + vN ] excludes all the points r1 , · · · , rN −1 , r1 + v21 , · · · , rN −1 + N2−1 and
r1 + v1 , · · · , rN −1 + vN −1 . This completes the induction.
Proposition 23. (Cater) For each n ∈ N, let cn be a positive number so small that
dF (0, cn hn ) < 21n , where hn is the spike on the interval (rn , rn + vn ) as in the previous
lemma. Moreover, let (an )n∈N be any bounded sequence of real P numbers and (wn )n∈N such
that wn = 1 if an 6= 0 and wn = 0 otherwise. Then the series ∞ n=1 an cn hn converges to
a nowhere analytic function in F if the sequence (wn rn )n∈N is dense in [a, b].
1
Proof. Let (wn rn )n∈N be dense in [a, b]. Since dF (0, cn hn ) < 2n and |an | ≤ M for some
fixed M and all n ∈ N, it follows that:
P∞ 1 |an cn hn (x)|
dF (0, an cn , hn ) = k=0 2k supa≤x≤b 1+|an cn hn (x)| ≤
P∞ 1 M |cn hn (x)|
≤ k=0 2k supa≤x≤b 1+M |cn hn (x)| ≤
(60)
P∞ 1 |cn hn (x)|
≤ k=0 2k supa≤x≤b (M + 1) 1+|c n hn (x)|
=
M +1
= (M + 1)dF (0, cn hn ) < 2n
Thus
N +j N +j N +j
!
X X X M +1 M +1
dF 0, ai ci hi ≤ dF (0, ai ci hi ) < i
< N −1 (61)
2 2
i=N i=N i=N
P∞
and n=1 an cn hn must converge in the complete metric space F , say to f .
Now fix any index N > 1 for which wN = 1 and aN 6= 0. By Lemma 22., for each i =
1, · · · , N − 1, the interval [rN − vN , rN + vN ] is a subinterval of (−∞, ri ) or (ri + vi , ∞)
or (riP , ri + vi ). For i = 1, · · · , N − 1, each spike hi is analytic in (rN − vN , rN + vN ) and
N −1
so is i=1 / (ri , ri +vi ) and rN + v2N ∈
ai ci hi . Note also that for i > N , rN ∈ / (ri , ri +vi ).
Thus:
(k) (k)
vN
hi (rN ) = hi rN + = 0, for i > N and all k ≥ 0. (62)
2
(k) (k)
Of course hN (rN ) = 0 for all k ≥ 0. Notice that limJ→∞ Ji=1 ai ci hi = f (k) uniformly
P
for all k ≥ 0, and
N
X vN vN
ai ci hi rn + = f rn + , (63)
2 2
i=1
so
Pj f (k) (rN ) vN k

limj→∞ k=0 k! 2 =
Pj 1 vN k P∞ (k)

= limj→∞ k=0 k! 2 i=1 ai ci hi (rN ) =
(64)
Pj 1 vN k PN −1 (k)

= limj→∞ k=0 k! 2 i=1 ai ci hi (rN ) =
PN −1 P∞ 1 vN k (k) PN −1 vN

= i=1 ai c i k=0 k! 2 hi (rN ) = i=1 ai ci hi rN + 2 .
But
N −1
vN X vN vN
f rN + = ai ci hi rN + + aN cN hN rN + . (65)
2 2 2
i=1
rN + v2N > 0, and cN 6= 0, aN 6= 0, we conclude that

Since hN
j
X f (k) (rN ) vN k vN
lim 6= f rN + (66)
j→∞ k! 2 2
k=0
and f is not analytic in the interval (rN − vN , rN + vN ).

Finally, if I is any subinterval of [a, b], then by hypothesis there is some index N > 1
for which aN 6= 0 and [rN − vN , rN + vN ] ⊂ I, so f is not analytic in I. Consequently, f
is nowhere analytic.
3. Measure and Dimension

From the point of view of Baire’s categories, we have come to the conclusion in the previous
section, that non-differentiable and non-analytic functions should play an important role in
modeling natural phenomena. In most complex phenomena where non-differentiability is
a prevalent feature, fractals of various natures are ubiquitous. Typically non-differentiable
functions are non-rectifiable and the length of an arc of its graph diverges [40]. This diver-
gence is intimately related with the fact that the dimension of the graph of a such a function
may be lower than 1. For each positive real number - the dimension of some fractal - we
can associate a fractal measure.
It is well known [12, 27] that the fractal dimensions of certain phenomena are more
suitably described by so-called dimension functions [12, 13] rather than real positive num-
bers (which are associated with the powers). For instance, with probability 1, Brownian
trails in Rn (n ≥ 2) have Hausdorff dimension 2. However, more delicate arguments reveal
that they have positive finite Hausdorff measure with respect to the functions [12]:
x2 log x1 log log log x1 , if n = 2

(67)
x2 log log x1 ,

if n ≥ 3
With these ideas in mind we recapitulate various ideas on generalized dimensions and
measure.
3.1. Dimension Functions

Definition 24. Consider real valued functions ψ defined on some interval of the form
(0, ǫψ ) with ǫψ (which may be different for different functions) a positive constant or ∞.
The set of dimension functions D is the set of such functions which in addition are contin-
uous, strictly increasing and such that limx↓0 ψ(x) = 0.
Definition 25. In D we define the following binary relations. We say that ψ, φ ∈ D are of
the same order of growth and write φ ∼ ψ if there exist real constants C1 , C2 such that
0 < C1 ≤ ψ(x)
φ(x) ≤ C2 for all 0 < x < min {ǫψ , ǫφ }. In particular, this is valid if the limit
ψ(x)
limx↓0 φ(x) exists and is a positive real number. Moreover, we say that the order of growth
ψ(x)
of ψ is greater than that of φ and write φ ≺ ψ, if limx↓0 φ(x) = 0. Finally, we have φ ψ
if φ ∼ ψ or φ ≺ ψ.
The following proposition is straightforward to prove. We leave the proof to the reader.
Proposition 26. The relations ∼ and in D are an equivalence relation and a partial
ordering, respectively.
Remark 27. We denote by D/ ∼ the set of equivalence classes in D and write [ψ] for
the equivalence class of ψ ∈ D. We should however mention that (D, ) is not a totally
ordered set as there are pairs of elements in D which are not comparable. Indeed, let
Z x
1 3
ψ(x) = τ 1 + sin dτ + x 2 (68)
0 τ
for x > 0. Clearly, ψ is continuous and limx↓0 ψ(x) = 0. Moreover

1 3√
ψ ′ (x) = x 1 + sin + x > 0, ∀x > 0 (69)
x 2
3
and so ψ is strictly increasing. Altogether, ψ ∈ D. Next, let φ(x) = x 2 ∈ D. For x > 0:
Z x
ψ(x) − 32 1
=x τ 1 + sin dτ + 1 (70)
φ(x) 0 τ
−1
Consider the sequence xn = π2 (2n + 1) . It follows that
3 R 2
ψ(xn ) π π(2n+1)
1

= 2 (2n + 1) τ 1 + sin dτ + 1 =
2
φ(xn ) 0 τ
(71)
π
3 n 2

1
o
= 2 (2n + 1) 2
π(2n+1) [1 + (−1)n ] + O (2n+1)2
+1
If n is odd, then ψ(x n)

φ(xn ) → 1, whereas
ψ(xn )
φ(xn ) → ∞, when n is even. We conclude that ψ and
φ are not comparable.
An important subset of D is the set of powers.
Example 28. Let us consider the powers xr , with r > 0. Clearly, these functions belong to
D. As a convention we shall use them as the representatives of the corresponding equiva-
lence classes. We denote the set of functions belonging to the equivalence classes of powers
by:
P := {φ ∈ D : ∃r > 0 such that φ ∼ xr } . (72)
By abuse of language we shall call this the set of powers.
The following lemmata present useful methods for obtaining new elements with differ-
ent orders of growth in D.
Lemma 29. Let φ1 , φ2 ∈ D be defined on (0, ǫφ1 ) and (0, ǫφ2 ), respectively. Then their
product φ1 · φ2 : (0, ǫ) −→ R, with ǫ = min {ǫφ1 , ǫφ2 }, belongs to D. Moreover, φ1 ≺
φ1 · φ2 and φ2 ≺ φ1 · φ2 .
Proof. The product of two non-negative, continuous and strictly increasing functions is
again continuous and strictly increasing. The product of two infinitesimals is again an
infinitesimal. On the other hand, we have limx↓0 φ1 (x)φ 2 (x)
φ1 (x) = limx↓0 φ2 (x) = 0 and
φ1 (x)φ2 (x)
limx↓0 φ2 (x) = limx↓0 φ1 (x) = 0, that is φ1 ≺ φ1 · φ2 and φ2 ≺ φ1 · φ2 .
Lemma 30. Let ψ1 : (0, ǫψ1 ) −→ R and ψ2 : (0, ǫψ2 ) −→ R belong to D and be such
that ψ2 ((0, ǫψ2 )) ⊂ (0, ǫψ1 ). Then the composition ψ1 ◦ ψ2 : (0, ǫψ2 ) −→ R belongs to D.
Furthermore (i) ψ2 ψ1 ◦ ψ2 if and only if x ψ1 ; (ii) ψ1 ψ1 ◦ ψ2 if x ψ1 ψ2 ;
and (iii) ψ1 ◦ ψ2 ψ1 if ψ2 ψ1 x.
Proof. The composition of two continuous and strictly increasing functions is a contin-
uous and strictly increasing function. Moreover, we have that limx↓0 (ψ1 ◦ ψ2 )(x) =
limτ ↓0 ψ1 (τ ) = 0, since limx↓0 ψ2 (x) = 0+ . Altogether, ψ1 ◦ ψ2 ∈ D. Let us now
prove the remaining statements: (i) limx↓0 (ψ1ψ◦ψ 2 )(x)
2 (x)
= limx↓0 ψ1ψ(ψ2 (x)
2 (x))
= limτ ↓0 ψ1τ(τ ) .
We conclude that ψ2 ≺ ψ1 ◦ψ2 if and only if x ≺ ψ1 . Now suppose that ψ1 ∼ x. Then there
exist C1 , C2 such that 0 < C1 ≤ ψ1x(x) ≤ C2 for all x ∈ (0, ǫψ1 ). Since, by assumption,
0 < ψ2 (x) ≤ ǫψ1 for all x ∈ (0, ǫψ2 ), we conclude that 0 < C1 ≤ (ψ1ψ◦ψ 2 )(x)
2 (x)
≤ C2 for all
x ∈ (0, ǫψ2 ). This means that ψ1 ◦ ψ2 ∼ ψ2 . Conversely, suppose that ψ1 ◦ ψ2 ∼ ψ2 , i.e.
there exist K1 , K2 such that 0 < K1 ≤ (ψ1ψ◦ψ 2 )(x)
2 (x)
≤ K2 for all x ∈ (0, ǫψ2 ). This means
that for all τ ∈ (0, ψ2 (ǫψ2 )) we have 0 < K1 ≤ ψ1τ(τ ) ≤ K2 . That is ψ1 ∼ x. (ii) Notice
that (ψ1ψ◦ψ 2 )(x)
1 (x)
= ψ1ψ(ψ2 (x)
2 (x))
·ψ 2 (x)
ψ1 (x) . Then result is then proved in the same way as (i). (iii)
ψ1 (x) ψ2 (x) ψ1 (x)
The proof is analogous since (ψ1 ◦ψ2 )(x) = ψ1 (ψ2 (x)) · ψ2 (x) .
Lemma 31. Let ψ ∈ D with limx↑ǫψ ψ(x) = ψ(ǫ− ψ ). Then its inverse function ψ
−1 :

0, ψ(ǫ− s
ψ ) −→ (0, ǫψ ) also belongs to D. Moreover, if for a given s > 0, ψ ≺ x , ψ ∼ x
s
1 1 1
or xs ≺ ψ, then x s ≺ ψ −1 , x s ∼ ψ −1 or ψ −1 ≺ x s , respectively.
Proof. Since ψ is strictly increasing, it is injective, and therefore admits an inverse. As

it is strictly increasing, continuous and limx↓0 ψ(x) = 0, limx↑ǫψ = ψ(ǫ− ψ ), we have that

− −
ψ ((0, ǫψ )) = 0, ψ(ǫψ ) . Consequently, ψ −1 is a function from 0, ψ(ǫψ ) onto (0, ǫψ ).
Moreover, as it is the inverse of a continuous and strictly increasing function, it is also
continuous and strictly increasing. Finally, since limx↓0 ψ(x) = 0+ , we conclude that
limx↓0 ψ −1 (x) = 0. Altogether, ψ −1 ∈ D. Finally, notice that:
s " #s " #s
ψ −1 (x) ψ −1 (ψ(x)) xs

x
lim 1 = lim 1 = lim 1 = lim (73)
x↓0 xs x↓0 (ψ(x)) s x↓0 (ψ(x)) s x↓0 ψ(x)
This relation yields the desired result.
Let us consider some more examples of functions in D.
Example 32. It is easy to show that the functions

 −1
1
σ1 (x) = log log 1 + x1



 , x ∈ 0, e−1
 σ (x) = log 1 + 1 −1 ,

x ∈ (0, +∞)

2 x
h 1 i−1 (74)


 σ3 (x) = e x − 1 , x ∈ (0, +∞)
 1
σ4 (x) = e− x2 ,

x ∈ (0, +∞)

all belong to D. Moreover, for any s > 0
xs σ3 (x) σ2 (x) σ4 (x)

lim = lim s
= lim = lim = 0. (75)
x↓0 σ2 (x) x↓0 x x↓0 σ1 (x) x↓0 σ3 (x)
This means that:

σ1 ≺ σ2 ≺ xs ≺ σ3 ≺ σ4 , ∀s > 0. (76)
Let us briefly mention that σ3 = σ2−1 , which illustrates Lemma 31..
Remark 33. At this stage it is important to remark that, while the product of two elements
in D/ ∼ is a well defined operation, the composition and inversion are not. Indeed, let
ψ1 , ψ2 , φ1 , φ2 ∈ D be such that ψ1 ∼ ψ2 and φ1 ∼ φ2 . That means that there exist
positive constants C1 , C2 , K1 and K2 such that 0 < C1 ≤ ψψ21 (x)
(x)
≤ C2 for all 0 < x <
φ1 (x)
ǫψ = min {ǫψ1 , ǫψ2 }, and 0 < K1 ≤ φ2 (x) ≤ K2 for all 0 < x < ǫφ = min {ǫφ1 , ǫφ2 }.
ψ1 (x)·φ1 (x)
Consequently, 0 < K1 C1 ≤ ψ2 (x)·φ2 (x) ≤ K2 C2 for all 0 < x < min {ǫψ , ǫφ }, i.e. ψ1 ·
φ1 ∼ ψ2 · φ2 . The product is thus a well defined operation (D/ ∼) × (D/ ∼) −→ (D/ ∼).
To prove that composition and inversion are not well defined in D/ ∼, we will give concrete
examples. Let ψ1 (x) = Lσ2 (x), ψ2 (x) = σ2 (x), φ1 (x) = M σ3 (x) = M σ2−1 (x) and
φ2 (x) = σ3 (x) = σ2−1 (x), where σ2 and σ3 are as in Example 32.. We shall further assume
that 0 < L < 1 and M > 0. And thus ψ1 ∼ ψ2 , φ1 ∼ φ2 . Moreover, ψ1−1 (x) = σ3 Lx
1
ψ1−1 (x)
and ψ2−1 (x) = σ3 (x). We thus have limx↓0 ψ2−1 (x)
= limx↓0 e x −1
L = +∞. And thus ψ1−1
e x −1
is not of the same order of growth as ψ2−1 . This means that inversion is not a well defined
M
operation (D/ ∼) −→ (D/ ∼). Likewise, we have (φ1 ◦ ψ1 ) (x) = and
(1+x−1 )1/L −1
h 1 i
(φ2 ◦ ψ2 ) (x) = x. But limx↓0 (φ 2 ◦ψ2 )(x) x
(φ1 ◦ψ1 )(x) = limx↓0 M 1 + x1 L − 1 = +∞. And thus
(φ1 ◦ ψ1 ) and (φ2 ◦ ψ2 ) are not of the same order of growth, which shows that composition
is not a well defined operation (D/ ∼) × (D/ ∼) −→ (D/ ∼).
An immediate consequence of Example 32. is the following proposition.
Proposition 34. The set of powers P is a bounded subset of D.
As for the set D, the following proposition holds.
Proposition 35. The set D has no maximal or minimal elements.
Proof. Suppose that D had a maximal element ψ. Then this would mean that there exists no
φ ∈ D such that ψ ≺ φ. However, from Lemma 29., for φ ∈ D, we have that φ · ψ ∈ D and
ψ ≺ φ · ψ, which is a contradiction. And so, D cannot have a maximal element. Likewise,
if D had a minimal element ρ, then there could be no φ ∈ D such that φ ≺ ρ. Take φ ◦ ρ for
φ ∈ D such that φ ≺ x. Then from Lemma 30., we conclude that φ ◦ ρ ∈ D and φ ◦ ρ ≺ ρ,
which is a contradiction.
Lemma 36. Let φ1 , φ2 ∈ P be such that φ1 ≺ φ2 . Then we can always find φ ∈ P and
ψ ∈ D\P such φ1 ≺ φ ≺ φ2 and φ1 ≺ ψ ≺ φ2 .
Proof. The proof is obviously independent of the representative of each equivalence class,
so we may safely assume φ1 (x) = xs and φ2 (x) = xr , with 0 < s < r. Let φ(x) = xn ,
with s < n < r. Then we have φ ∈ P and φ1 ≺ φ ≺ φ2 . This proves the first claim. Let
now ψ(x) = φ1 (x) · σ2 (x), where σ2 (x) is as in Example 32.. From Lemma 29. ψ ∈ D
and φ1 ≺ ψ. On the other hand, for u > 0:

ψ(x) s−u +∞ if u > s
lim u = lim x · σ2 (x) = (77)
x↓0 x x↓0 0 if u ≤ s
φ2 (x) xr−s
This proves that ψ ∈ D\P. Finally limx↓0 ψ(x) = limx↓0 σ2 (x) = 0. Altogether φ1 ≺ ψ ≺
φ2 .
Remark 37. Lemma 36. states that between any two elements φ1 , φ2 of P, we can always
find a third element of P. However, if either φ1 or φ2 (or both) do not belong to P, then
there may not exist any element of P between them. For instance, from Example 32., we
have that σ1 ≺ σ2 ≺ xs ≺ σ3 ≺ σ4 for any s > 0. This means that there exist no elements
of P between σ1 and σ2 and between σ3 and σ4 . Another example is as follows. For any
0 < n ≤ s < r, we have: xn ≺ xs σ1 (x) ≺ xs σ2 (x) ≺ xr , with xs σ1 (x), xs σ2 (x) ∈ D\P.
This means that there are no elements of P between xs σ1 (x) and xs σ2 (x).
In summary, between two elements ψ1 , ψ2 ∈ D\P there may or may not exist elements
of P. Similarly, the previous analysis also reveals that between an element φ ∈ P and an
element ψ ∈ D\P there may not exist an element of P. Indeed, since xs ≺ xs σ2 (x) ≺ xr
for any r > s > 0, then between xs ∈ P and xs σ2 ∈ D\P there exists no element of P.
On the contrary, between any two elements of D we can always find an element of D\P.
To prove this we need the following lemma.
Lemma 38. Let φ : (0, ǫφ ) −→ R, with ǫφ > 0 be continuous, strictly positive and such
φ(x)
that limx↓0 φ(x) = 0. Then there exists ψ ∈ D such that limx↓0 ψ(x) = 0.
Proof. Notice that we do not require φ to be strictly increasing, so that it may or may
not belong to D. Let us define f (x) = max0<u≤x φ(u), for x ∈ (0, ǫφ ). Obviously, f is
continuous, strictly positive, increasing (but not necessarily
p strictly) and limx↓0 f (x) = 0.
Take any ρ ∈ D and define ψ(x) = ρ(x)f (x) + f (x), for x ∈ (0, ǫψ ) and ǫψ =
φ(x)
min {ǫφ , ǫρ }. By construction ψ ∈ D. It remains to prove that limx↓0 ψ(x) = 0. Indeed, for
0 < x < ǫψ , p
φ(x) f (x) f (x)
0< ≤ p =p . (78)
ψ(x) ρ(x)f (x) + f (x) f (x)ρ(x) + 1
As x ↓ 0, the right-hand side vanishes.
Proposition 39. Let ψ1 , ψ2 ∈ D be such that ψ1 ≺ ψ2 . Then there exists η ∈ D\P such
that ψ1 ≺ η ≺ ψ2 .
Proof. Consider the function φ(x) = ψ 2 (x)

ψ1 (x) defined on (0, ǫ) with ǫ = min {ǫψ1 , ǫψ2 }. This
function is continuous, strictly positive and moreover limx↓0 φ(x) = 0. From Lemma 38.,
φ(x)
there exists ψ ∈ D such that limx↓0 ψ(x) = 0. Consider the function η1 (x) = ψ(x) · ψ1 (x).
From Lemma 29., η1 ∈ D and ψ1 ≺ η1 . On the other hand: limx↓0 ψη12(x) (x) φ(x)
= limx↓0 ψ(x) =
0, which means that η1 ≺ ψ2 . If η1 ∈ D\P, then the problem is settled. If η1 ∈ P, then
we repeat the previous procedure and find η3 ∈ D such that η1 ≺ η3 ≺ ψ2 . If η3 ∈ D\P,
then we terminate the proof. If η1 , η3 both belong to P, then from Lemma 36., we can find
η2 ∈ D\P such that ψ ≺ η1 ≺ η2 ≺ η3 ≺ ψ2 .
The fact that (D, ) is not totally ordered is unsatisfactory. This leads us to the follow-
ing definition.
Definition 40. A scale of dimension functions is a subset F of D such that: (i) P ⊂ F,

(ii) (F, ) is a totally ordered bounded complete set, and (iii) for any ψ, φ ∈ F such that
ψ ≺ φ, there exists ρ ∈ F such that ψ ≺ ρ ≺ φ. We refer to the latter property by saying
that F is dense. Two scales F1 , F2 are said to be compatible if F1 ∪ F2 is another scale.
A scale is maximal if there exists no other scale which contains it as a proper subset.
Remark 41. One may wonder about the existence of maximal scales. However, their exis-
tence is ensured by Hausdorff’s Maximality Principle [29]. This principle is equivalent to
Zorn’s Lemma and to the Axiom of Choice, and it states that in any partially ordered set,
every totally ordered subset is contained in a maximally totally ordered subset. A subset
is maximally totally ordered if, when it is enlarged in any way, fails to remain totally or-
dered. One should nevertheless stress that the maximally totally ordered set containing the
original totally ordered set may not be unique.
Lemma 42. Two scales F1 , F2 are compatible if and only if every pair of elements φ1 ∈
F1 \F2 and φ2 ∈ F2 \F1 are comparable.
Proof. This follows immediately from the definition and the fact that F1 ∪F2 = (F1 \F2 )∪
(F1 ∩ F2 ) ∪ (F2 \F1 ).
Example 43. The set of powers P is a scale, the trivial scale. In fact there is an order
preserving bijection from the set of positive reals onto P\ ∼ defined by r 7→ [xr ]. As the
set R+ is totally ordered, bounded complete and dense, so is the set of powers.
Example 44. Another interesting example which finds applications in the context of Brow-
nian motion (see (67)) is the scale F that consists of the equivalence classes of functions of
the form [40]:
1 β

α
φ(x) = x logn , α > 0, β ∈ R, n = 0, 1, 2, 3, · · · (79)
x
Here logn indicates

the iterated logarithm: log0 (x) = 1, log1 (x) = log(x), logn (x) =
log logn−1 (x) , for n ≥ 2.
3.2. Hölder-continuity
As an application, let us consider the so-called Hölder-continuity of a function in some set.
Definition 45. Let Ω1 , Ω2 be metric spaces with metrics d1 , d2 , respectively. A function

f : F ⊂ Ω1 → Ω2 is said to be locally ψ-Hölder continuous at u ∈ F for ψ ∈ D if there
exist constants K > 0 and 0 < ǫ < ǫψ such that
d2 (f (u), f (v)) ≤ Kψ (d1 (u, v)) (80)
for all v ∈ F with d1 (u, v) < ǫ. The function is said to be ψ-Hölder continuous in F if
there exists a constant K > 0 such that inequality (80) holds for all u, v ∈ F . Finally, if Ω1
has a partial ordering E, we say that f is locally right (resp. left) ψ-Hölder continuous
at u ∈ F if there exist constants K > 0 and 0 < ǫ < ǫψ such that (80) is valid whenever
v ∈ F with d1 (u, v) < ǫ and u E v (resp. v E u). By abuse of language if f is ψ-Hölder
continuous for ψ(x) = xr for some r > 0, then we shall simply say that f is r-Hölder
continuous.
Proposition 46. If f : F ⊂ Ω1 → Ω2 is ψ-Hölder continuous in F , locally ψ-Hölder

continuous at u ∈ F , locally right ψ-Hölder continuous at u ∈ F , or locally left ψ-Hölder
continuous at u ∈ F for ψ ∈ D, then it is, respectively, φ-Hölder continuous in F , locally
φ-Hölder continuous at u ∈ F , locally right φ-Hölder continuous at u ∈ F , or locally left
φ-Hölder continuous at u ∈ F for any φ ∈ D with φ ψ.
Proof. We prove the result when f is locally ψ-Hölder continuous at u ∈ F . The rest
is proved analogously. This means that there exist positive constants K, ǫ such that (80)
holds for any v ∈ F with d1 (u, v) < ǫ. Now take an arbitrary φ ∈ D such that φ ≺ ψ,
i.e. limx↓0 ψ(x)
φ(x) = 0. This means that there exists 0 < ρ < min {ǫψ , ǫφ , ǫ} such that
ψ(x) ≤ φ(x) for all 0 < x < ρ. For any v ∈ F such that d1 (u, v) < ρ, we have that:
d2 (f (u), f (v)) ≤ Kψ (d1 (u, v)) ≤ Kφ (d1 (u, v)) (81)
which means that f is locally φ-Hölder continuous at u ∈ F . Finally, suppose that φ ∈ D

is such that φ ∼ ψ. Then there exist constants C1 , C2 such that 0 < C1 ≤ ψ(x)
φ(x) ≤ C2 , for
all 0 < x < δ = min {ǫφ , ǫψ }. And so, for any v ∈ F such that d1 (u, v) < min {ρ, ǫ}, we
have:
d2 (f (u), f (v)) ≤ Kψ (d1 (u, v)) < C2 Kφ (d1 (u, v)) , (82)
which proves that f is locally φ-Hölder continuous at u ∈ F .
Now remember that a scale of dimension functions F is a totally ordered bounded
complete set. According to the previous proposition, if a function f is ψ-Hölder continuous
for some ψ ∈ F, then it is φ-Hölder continuous for all φ ∈ F with φ ≺ ψ. This motivates
the following definition.
Definition 47. Let F be a scale of dimension functions and let f : F ⊂ Ω1 → Ω2 be a

continuous function. We define the critical Hölder order in F , critical local Hölder order
at u ∈ F , critical local right Hölder order at u ∈ F , and critical local left Hölder order
at u ∈ F to be the suprema in F of each of the following sets, respectively.
{φ ∈ F : f is φ-Hölder continuous in F }
{φ ∈ F : f is locally φ-Hölder continuous at u ∈ F }

(83)
{φ ∈ F : f is locally right φ-Hölder continuous at u ∈ F }
{φ ∈ F : f is locally left φ-Hölder continuous at u ∈ F }
Obviously, the suprema may or may not exist, depending on whether these sets have an
upper bound or not, or on whether they are empty or not.
As an example, we consider the Weierstrass and the Takagi functions.
1

Lemma 48. TheTakagi function is locally ψ-Hölder continuous for ψ(x) = x log x ,
with x ∈ 0, e−1 .
Proof. First of all notice that ψ ∈ F, where F is the scale of Example 44. (with α =
β = n = 1 in (79)). Moreover, from (11), we conclude that T is indeed locally ψ-Hölder
continuous at every x ∈ [0, 1].
Likewise, for the Weierstrass function we have:
Lemma 49. The Weierstrass function Wλ,s has critical local Hölder order 2 − s.
Proof. From Proposition 18. (i) we know that Wλ,s is indeed locally (2 − s)-Hölder con-
tinuous in [0, 1].
On the other hand, suppose that Wλ,s is locally r-Hölder continuous at some x ∈ [0, 1]
for r > 2 − s. Then there exist ǫ, K > 0 such that
|Wλ,s (x) − Wλ,s (y)| ≤ K|x − y|r (84)
for all y ∈ [0, 1] with |x − y| < ǫ.

Following the proof of Corollary 19., for each suffciently large n ∈ N we can find
yn ∈ [0, 1] \ {x} such that:
1
|x − yn | < , (85)
n
and 2−s
|Wλ,s (x) − Wλ,s (yn )| C2 1
r
≥ r
≥ C2 nr−(s−2) (86)
|x − yn | |x − yn | n
As n → ∞ the right-hand side of (86) diverges and this contradicts (84). Consequently
2 − s is the local critical Hölder order of Wλ,s at every x ∈ [0, 1].
3.3. Hausdorff Measure

Definition 50. Let U be any non-empty subset of a metric space Ω with metric d. A count-
able (or finite) collection {Ui } of sets of diameter at most δ > 0 that cover F ⊂ Ω is called
a δ-cover of F .
Definition 51. Let F be a subset of Ω and ψ ∈ D. For any δ > 0 we define

(∞ )
∗
Hδψ (F ) := inf
X
ψ (|Ui |) : {Ui } is a δ-cover of F (87)
i=1
+
The set on the right-hand ∗ is obviously a non-empty subset of R and thus the infimum
side
exists. As δ decreases Hδψ (F ) increases, and we thus define:
∗ ∗
Hψ (F ) := lim Hδψ (F ). (88)
δ↓0
∗
We call Hψ (F ) the ψ-Hausdorff measure of F .
Theorem 52. The ψ-Hausdorff measure is an outer measure.
δ>
Proof. Given
∗
0 we may cover ∅ with a single set of diameter
∗ ǫ with 0 < ǫ ≤ δ. And
ψ
thus 0 ≤ Hδ (∅) ≤ ψ(ǫ) for any 0 < ǫ ≤ δ, giving Hδψ (∅) = 0. Consequently,
∗ ∗
Hψ (∅) = limδ↓0 Hδψ (∅) = 0.
If E ⊂ F ⊂ Ω, every δ-cover
of F is also a δ-cover of E, so taking the infimum over
∗ ∗
all possible δ-covers gives Hδψ (E) ≤ Hδψ (F ) for all δ > 0. Letting δ ↓ 0 gives
∗ ∗
Hψ (E) ≤ Hψ (F ).
Let {Fi } be a countable collection of subsets of Ω. Without loss of generality, we may
P∞ ψ ∗
assume that i=1 Hδ (Fi ) < ∞. For ǫ > 0, let {Ui,j , j = 1, 2, · · · } be a δ-cover
∗
of Fi such that ∞ Hδψ (Fi ) + 2ǫi . Then {Ui,j : i, j = 1, 2, · · · } is a
P
j=1 ψ (|U i,j |) ≤
δ-cover of ∪∞i=1 Fi and
∗ P∞ ψ ∗
Hδψ (∪∞
P∞ P∞ ǫ
F
i=1 i ) ≤ i=1 i=j ψ (|U i,j |) ≤ i=1 H δ (F i ) + 2i
=
∗ (89)
P∞ ψ ∗ (F ).
Hδψ
P∞
=ǫ+ i=1 (Fi ) ≤ ǫ + i=1 H i
Since this is true for any ǫ > 0, it follows that

∗ ∞
Hδψ
X
ψ
H (∪∞
i=1 Fi ) = lim (∪∞
i=1 Fi ) ≤ Hψ (Fi ), (90)
δ↓0
i=1

Definition 53. A set F ⊂ Ω is said to be Hψ -measurable for some ψ ∈ D if

∗ ∗ ∗
Hψ (A) = Hψ (A ∩ F ) + Hψ (A ∩ F c ) , (91)
for all A ⊂ Ω. Here F c denotes the complement Ω\F .
The following theorem due to C. Carathéodory [22] then establishes our measure space.
Theorem 54. Let Mψ denote the set of measurable sets in Ω with respect
∗ to the outer
∗
measure Hψ . Then Mψ is a σ-algebra and the restriction Hψ of Hψ to Mψ is a
measure.
Remark 55. The fact that Hψ is a measure means that besides satisfying all the properties
stated in the proof of Theorem 52., it also satisfies:
∞
X
Hψ (∪∞
i=1 Fi ) = Hψ (Fi ), (92)
i=1
for any countable family (Fi ) of mutually disjoint measurable subsets of Ω. Altogether
Ω, Mψ , Hψ is a measure space. Let us also mention that if Ω = Rn and ψ(x) = xs ,

s > 0, we recover the textbook Hausdorff measure of dimension s [12].
Remark 56. For most applications one is usually content with the collection of Borel sets
as a σ-algebra. This is the smallest σ-algebra containing all the open subsets of Ω. If,
as assumed, Ω is a metric space, then the open sets are just the open sets in the metric
topology. We denote the Borel set
∗ by B. It is important to remark that every Borel set is
measurable with respect to H ψ for any ψ ∈ D. In other words: B ⊂ Mψ for all ψ ∈ D.
Theorem 57. Let f : F ⊂ Ω1 −→ Ω2 be of ψ-Hölder type in F , with ψ ∈ D, i.e.

d2 (f (x), f (y)) ≤ ψ (d1 (x, y)) for any x, y ∈ F . If F is a Borel subset of Ω1 , then
−1
Hφ◦ψ (f (F )) ≤ Hφ (F ), for any φ ∈ D.
Proof. First of all notice that since F is a Borel set in Ω1 and f is a continuous function (as
it is Hölderian), then f (F ) is also a Borel set in Ω2 . Let {Ui } be an arbitrary δ-cover of F .
Then we have
|f (F ∩ Ui )| = sup {d2 (f (x), f (y)) , x, y ∈ F ∩ Ui } ≤
(93)
≤ sup {ψ (d1 (x, y)) , x, y ∈ F ∩ Ui } ≤ ψ (|Ui |)
where we used the fact that ψ is strictly increasing in the last step. We conclude that
{f (F ∩ Ui )} constitutes a ψ(δ)-cover of f (F ). Notice that as ψ, φ ∈ D, then also ψ −1 , φ ◦
ψ −1 ∈ D. Next we have:
X X X
φ ◦ ψ −1 (|f (F ∩ Ui )|) ≤ φ ◦ ψ −1 ◦ ψ(|Ui |) = φ(|Ui |). (94)
i i i
Taking the infimum over all δ-covers of F , we obtain:

( )
φ ◦ ψ (|f (F ∩ Ui )|) , {Ui } is a δ-cover of F ≤ Hδφ (F )
X
−1
inf (95)
i
Since there are more ψ(δ)-covers of f (F ) than those of the form {f (F ∩ Ui )}, we con-
φ◦ψ −1
clude that Hψ(δ) (f (F )) is lower or equal to the left-hand side of the previous equation.
φ◦ψ −1
Altogether: Hψ(δ) (f (F )) ≤ Hδφ (F ). Taking the limit δ ↓ 0, we obtain the desired result.
The following proposition describes the behavior of the measure under a change of
dimension function.
Proposition 58. Let ψ, φ, ρ ∈ D be such that φ ∼ ρ and φ ≺ ψ. Then there exist constants
C1 , C2 > 0 such that for any F ⊂ B:
C1 Hρ (F ) ≤ Hφ (F ) ≤ C2 Hρ (F ), Hψ (F ) ≤ Hφ (F ). (96)
Proof. Since φ ∼ ρ, there exist constants C1 , C2 > 0 and 0 < δ ≤ min {ǫφ , ǫρ } such that
C1 ≤ φ(x)
P∞ P∞
P∞ ρ(x) ≤ C2 for all x ∈ (0, δ). It then follows that C1 i=1 ρ(|Ui |) ≤ i=1 φ(|Ui |) ≤
C2 i=1 ρ(|Ui |) where {Ui } is an arbitrary δ-cover of F . Taking infima over all δ-covers
yields C1 Hδρ (F ) ≤ Hδφ (F ) ≤ C2 Hδρ (F ). Letting δ ↓ 0 we obtain C1 Hρ (F ) ≤ Hφ (F ) ≤
C2 Hρ (F ), which proves the first part. The second part is proved similarly. If φ ≺ ψ,
then limx↓0 ψ(x)
φ(x) = 0. That means that there exists δ with 0 < δ ≤ min {ǫ , ǫ } such
P∞ ψ φ
that ψ(x) ≤ φ(x), ∀x ∈ (0, δ). For any δ-cover {Ui }, we thus have: i=1 ψ (|Ui |) ≤
P∞ ψ φ
i=1 φ (|Ui |). Taking infima over all δ-covers yields Hδ (F ) ≤ Hδ (F ). Letting δ ↓ 0, we
obtain Hψ (F ) ≤ Hφ (F ).
Remark 59. If for φ, ρ ∈ D, φ ∼ ρ, then from the previous proposition, we may establish
an equivalence relation for measures. We write Hφ ∼ Hρ if and only if there exist positive
real constants C1 , C2 such that C1 Hρ (F ) ≤ Hφ (F ) ≤ C2 Hρ (F ) for every F ⊂ B. It is
easy to prove that this is indeed an equivalence relation. As before, for each equivalence
class in D/ ∼, we fix a representative φ ∈ [φ], which in turn, will fix a representative of the
equivalence class of Hφ . The relevant point here is that a set F is measurable with respect
to Hφ if and only if it is measurable with respect to any Hρ with ρ ∼ φ.
Proposition 60. Let φ, ψ ∈ D with φ ≺ ψ and let F ⊂ B be such that Hφ (F ) < ∞. Then
Hψ (F ) = 0.
Proof. Let δ be such that 0 < δ ≤ min {ǫφ , ǫψ } and {Ui } a δ-cover of F . Since
φ ≺ ψ, we have limx↓0 ψ(x) ψ(x)
φ(x) = 0, which means that φ(x) is bounded in (0, δ). Let
η(δ) = sup0<x≤δ ψ(x)
P∞
φ(x) . Obviously, limδ↓0 η(δ) = 0. We then have: i=1 ψ (|Ui |) =
P∞ ψ(|Ui |) P∞
i=1 φ (|Ui |) φ(|Ui |) ≤ η(δ) i=1 φ (|Ui |). Taking infima over all δ-covers of F , we get:
0 ≤ Hδψ (F ) ≤ η(δ)Hδφ (F ). Since Hφ (F ) < ∞, upon taking the limit δ ↓ 0, we obtain
Hψ (F ) = 0.
Propositions 58. and 60. reveal that, as we take different elements φ in some scale of
functions F and evaluate the corresponding measures Hφ (F ) of a set F ⊂ B, these will
jump from 0 to ∞. This motivates the following definition.
Definition 61. Let F ⊂ B and F be a scale of dimension functions. We define the Haus-
dorff dimension function of F to be the element of F:
ψH (F ) := supF ψ ∈ F/ ∼: Hψ (F ) = ∞ =

(97)
= infF ψ ∈ F/ ∼: Hψ (F ) = 0

Notice that HψH (F ) may be 0, ∞ or any real positive number.
Definition 62. The Hausdorff-Besicovitch dimension of a set F ⊂ B with Hausdorff

dimension function ψH (F ) := ψH is defined by:
log (ψH (x))

dimH (F ) := lim . (98)
x↓0 log(x)
Let us briefly verify that this definition makes sense. But before that let us define the
following quantities:
sH := sup {s > 0 : xs ≺ ψH } , rH := inf {r > 0 : ψH ≺ xr } . (99)
Notice that sH may be ∞, or it may not exist, if ψH ≺ xs for all s > 0. Likewise, rH may
be equal to zero, or it may not exist if xs ≺ ψH for all s > 0.
Proposition 63. The Hausdorff-Besicovitch dimension has the following properties. (i) For
ψH (x) = xs (s > 0), we have: dimH (F ) = s.
(ii) The Hausdorff-Besicovitch dimension is independent of the representative of the equiv-
alence class of ψH .
(iii) The Hausdorff-Besicovitch dimension is determined by the element of P ”closest” to
ψH . To be more specific: dimH (F ) = sH or dimH (F ) = rH .
log(xs )
Proof. (i) By definition dimH (F ) = limx↓0 log(x) = s, as expected. (ii) Next, suppose
φ(x)
that φ ∈ D is such that φ ∼ ψH . Then there exist C1 , C2 > 0 such that C1 ≤ ψH (x) ≤ C2 ,
for x ∈ (0, δ) and δ = min {ǫφ , ǫψH }. It then follows that:

φ(x)
log ψ (x)
ψH (x) H
limx↓0 log(φ(x))
log(x) = limx↓0 log(x) =

(100)
φ(x)
log(ψH (x)) log ψH (x)
= limx↓0 log(x) + limx↓0 log(x) = dimH (F ).
This means that the Hausdorff dimension does not depend on the representative of the
equivalence class to which ψH belongs. (iii) Finally, if ψH (x) = xu for some u > 0,
then 0 < dimH (F ) = sH = rH = u < ∞. Alternatively, suppose that ψH ∈ / P.
s
Since the scale F is a well-ordered set, then at least one of the sets {s > 0 : x ≺ ψH },
{r > 0 : ψH ≺ xr } is non-empty. Suppose that {s > 0 : xs ≺ ψH } is non-empty. This

means that 0 < sH ≤ ∞. We first assume that sH < ∞. Let us check that dimH (F ) = sH .
Since ψH ∈/ P, we have limx↓0 ψxHsH
(x)
= 0.

log x sH ψH (x) log ψH (x)
log (ψH (x)) x s H x s H
dimH (F ) = lim = lim = sH + lim . (101)
x↓0 log(x) x↓0 log(x) x↓0 log(x)

ψH (x)
log x sH ψH (x)
It remains to prove that limx↓0 log(x) = 0. Let us write φ(x) = x sH . By definition
of sH (and as ψH ∈/ P), we have limx↓0 φ(x)xδ
= ∞ for any δ > 0. That is, for any δ > 0,
we can find 0 < ǫ < 1 such that x < φ(x) < 1 ⇔ log xδ > |log (φ(x))|, ∀x ∈ (0, ǫ).
δ

And thus, we have:

< δ log(x) = δ.
log (φ(x))

log(x) log(x) (102)
In summary dimH (F ) = sH . If however, sH = ∞, then limx↓0 ψHxs(x) = 0 for all s > 0.

For sufficiently small |x| < 1, we have (log(x) < 0) and ψH (x) ≤ xs . Consequently,
log(ψH (x)) s)
log(x) ≥ log(x
log(x) . Taking the limit x ↓ 0, we obtain dimH (F ) ≥ s for any s > 0. And
thus dimH (F ) = sH = ∞.
If we assume instead that {r > 0 : ψH ≺ xr } is non-empty, then 0 ≤ rH < ∞. Following
the same steps, we can prove that 0 < dimH (F ) = rH < ∞ or 0 = dimH (F ) = rH .
Let us now explore some properties of the Hausdorff dimension and of Hausdorff di-
mension functions.
The following proposition is an immediate consequence of Theorem 57..
Proposition 64. Let f : F ⊂ Ω1 −→ Ω2 be ψ-Hölder continuous in F ∈ B with

ψ ∈ F and suppose that F has Hausdorff dimension function φH ∈ F. Then the Haus-
dorff dimension function of f (F ) is some ρH ∈ F such that ρH φH ◦ ψ −1 . Likewise:
log(φH ◦ψ −1 (x))
dimH (f (F )) ≤ limx↓0 log(x) .
Proof. Let φ ∈ F be such that φH ≺ φ. Then from Theorem 57. and Proposition 60.,
−1
we have that Hφ◦ψ (f (F )) ≤ Hφ (F ) = 0, which implies that the Hausdorff dimension
function ρH of f (F ) satisfies ρH φ ◦ ψ −1 for all φ ∈ F with φH ≺ φ. Consequently
ρH φH ◦ ψ −1 . Finally, we prove the last statement. By definition dimH (f (F )) =
limx↓0 log(ρ H (x)) −1
log(x) . If ρH ∼ φH ◦ ψ , then from Proposition 63. (ii), the result follows
immediately. Let us assume that ρH ≺ φH ◦ ψ −1 . For sufficiently small x, we may
assume that |x| < 1 (log(x) < 0) and ρH (x) ≥ φH ◦ ψ −1 (x). It then follows that
log(ρH (x)) log(φH ◦ψ −1 (x))
log(x) ≤ log(x) . Letting x ↓ 0, we obtain the desired result.
Theorem 65. Let Ω be a metric space. The Hausdorff dimension and the Hausdorff di-
mension functions with respect to a scale F of Borel subsets of Ω satisfy the following
properties.
(monotonicity) If E ⊂ F ⊂ B, then ψH (E) ψH (F ) and dimH (E) ≤ dimH (F ).

(countable stability) If F1 , F2 , · · · is a (countable) sequence of Borel subsets of Ω, then

ψH (∪∞ ∞
i=1 Fi ) = sup1≤i<∞ {ψH (Fi )} and dimH (∪i=1 Fi ) = sup1≤i<∞ {dimH (Fi )}.
Proof. Let E ⊂ F . As the ψ-Hausdorff measure is a measure, Hψ (E) ≤ Hψ (F ) for any

ψ ∈ F. It then follows that ψH (E) ψH (F ) and dimH (E) ≤ dimH (F ). This proves
monotonicity. Next consider a (countable) sequence F1 , F2 , · · · of Borel subsets of Ω .
From the monotonicity property, we immediately have that
sup1≤i<∞ {ψH (Fi )} ψH (∪∞

i=1 Fi ) (103)
and
sup1≤i<∞ {dimH (Fi )} ≤ dimH (∪∞
i=1 Fi ) (104)
for each i. Here sup1≤i<∞ {ψH (Fi )} is the supremum in the scale of dimension
functions F. On the other hand, given ψ ∈ F, if ψH (Fi ) ≺ ψ for all i, then
Hψ (Fi ) = 0 for all i and thus Hψ (∪∞ i=1 Fi ) = 0. This gives the opposite inequality
ψH (∪∞ F
i=1 i ) sup 1≤i<∞ {ψ H (Fi )}. The same reasoning leads to dimH (∪∞ i=1 Fi ) ≤
sup1≤i<∞ {dimH (Fi )}.
Regarding countable sets, we make the following conjecture.
Conjecture 66. Let F ⊂ B and F a maximal scale of dimension functions. Then Hψ (F ) =

0 for all ψ ∈ F if and only if F is countable.
Proof. We prove only sufficiency. Let F be a countable set. Then we may write it as a
sequence of points F = {x1 , x2 , · · · }. Since the ψ-Hausdorff
P∞ measure is a measure, we
have for any ψ ∈ F: Hψ (F ) = Hψ (∪∞ i=1 {x i }) ≤ i=1 H ψ ({x }). As the ψ-Hausdorff
i
measure of a set with a single point is zero for any ψ ∈ F, we obtain the result.
Proposition 67. Suppose that F can be covered by nk sets of diameter at most δk with
δk ↓ 0 as k → ∞. Moreover assume that nk ψ(δk ) stays bounded as k → ∞ for some
ψ ∈ D. Then Hψ (F ) < ∞.
Proof. If F is covered by nk sets of diameter at most δk , then by definition, 0 ≤ Hδψk (F ) ≤

nk ψ(δk ). If the right-hand side of this inequality stays bounded as k → ∞, then we
conclude that Hψ (F ) < ∞.
Remark 68. Under the conditions of the previous proposition, we conclude that the Haus-
dorff dimension function φH of F must be such that φH ψ.
Usually Hausdorff dimensions are difficult to determine. One sometimes resorts to

alternative definitions of fractal dimension. For subsets of Rn , the following definition is
useful.
Definition 69. Let F ∈ Rn and for each δ > 0 let Nδ denote the number of δ-mesh cubes
that intersect F . The lower and upper box-counting dimensions of F are given by:
log Nδ (F ) log Nδ (F )
dimB (F ) = limδ↓0 , dimB (F ) = limδ↓0 , (105)
− log δ − log δ
respectively. The box-counting dimension of F is:
log Nδ (F )
dimB (F ) = lim , (106)
δ↓0 − log δ
whenever the limit exists.
From the definition it follows immediately that:
dimH (F ) ≤ dimB (F ) ≤ dimB (F ). (107)
Let us now state some results concerning the dimension of graph of a function.
Proposition 70. Let f : I → R be a continuous function. Let 0 < δ ≤ 1, and m is the

least integer greater than or equal to 1δ . Then if Nδ is the number of squares of the δ-mesh
that intersect the graph of f ,
m−1
X m−1
X
δ −1 Rf [iδ, (i + 1)δ] ≤ Nf ≤ 2m + δ −1 Rf [iδ, (i + 1)δ] . (108)
i=0 i=0
Here Rf [x1 , x2 ] = supx1 ≤x,y≤x2 |f (x) − f (y)| is the range of the function f in the interval
[x1 , x2 ].
Proof. See [12].
Theorem 71. Let f : I → R be a continuous function, F a scale of functions and ψ ∈ F

x2
such that x ≺ ψ ≺ x2 and ψ(x) ∈ F. Moreover, suppose that there exist constants
0 < δ ≤ 1 and K > 0 such that:
|x − y|2
|f (x) − f (y)| ≤ K , (109)
ψ(|x − y|)
for all x, y ∈ I such that |x − y| ≤ δ. Then the graph Γ of f satisfies Hψ (Γ) < ∞ and has
a Hausdorff dimension function φH (Γ) ψ.
|x1 −x2 |2
Proof. If x1 , x2 ∈ I are such that |x1 −x2 | ≤ δ, then we have Rf [x1 , x2 ] ≤ K ψ(|x 1 −x2 |)
≤
2
δ x 2
K ψ(δ) , where we used the fact that ψ(x) is an increasing function as it belongs to F. If m
is the least integer greater than or equal to 1δ then m < 1 + 1δ . From Proposition 70. it then
follows that:
δ 2
Nδ ≤ 2m + mδ −1 K ψ(δ) ≤
h i h i (110)
δ
≤ (1 + δ −1 ) 2 + K ψ(δ) = 1
ψ(δ) 2ψ(δ) + δK + 2 ψ(δ)
δ + K .
Under the assumptions of the theorem, the terms inside the bracket stay bounded as δ →
0. We conclude that there exists some constant C > 0 such that Nδ ψ(δ) ≤ C. From
Proposition 67. the result of the theorem follows.
Theorem 72. Suppose that there exist real numbers C > 0, δ0 > 0, and ψ ∈ F with
2
x ψ ≺ x2 and xψ ∈ F. Moreover, let f : I → R be a continuous function for which
the following property holds: for each x ∈ I and 0 < δ ≤ δ0 there exists u such that
δ2
|x − u| ≤ δ and |f (x) − f (u)| ≥ C ψ(δ) . Then dimB (Γ) ≥ limδ↓0 log(ψ(δ))
log(δ) .
Proof. We shall consider again a δ-mesh and use the same notation as in Proposition 70. and
δ2
Theorem 71.. Under the assumptions of the theorem, we have Rf [x1 , x2 ] ≥ C ψ(δ) . Since
δ 2 C
m ≥ δ −1 and δ < δ0 , we conclude from Proposition 70. that Nδ ≥ mδ −1 C ψ(δ) ≥ ψ(δ) .
log(ψ(δ))
By definition we conclude that dimB ≥ limδ↓0 log(δ) .
Using the previous two theorems, we can easily obtain the box dimension of the graph
of the Weierstrass function.
Lemma 73. Let λ > 1 and 1 < s < 2. The graph Γλ,s of the Weierstrass function Wλ,s
has box dimension dimB (Γλ,s ) = s.
Proof. From Proposition 18. we realize that the conditions of Theorems 71. and 72. are
verified for ψ(x) = xs .
4. Fractional Calculus
In the previous sections we came across functions such as Takagi’s or Weierstrass’ functions
which are continuous everywhere in an interval, but nowhere differentiable in that interval.
If such functions are to represent some physical dynamical system or some other process,
then we would like to be able to write down equations for this sort of motion. We are never-
theless faced with a complicated problem. Since these functions are nowhere differentiable,
we cannot write down differential equations to which they would be solutions. There is an
old answer to this problem. It is called the fractional calculus [16, 28, 31, 33, 36].
Definition 74. Let f be a continuous function on I = [a, b]. The right and left Riemann-
Liouville fractional integrals of order α ≥ 0 of f at x ∈ I are defined by:
α (f )(x) := 1
Rx α−1 f (y)dy
Ia,− Γ(α) a (x − y)
(111)
α 1
Rb α−1
Ib,+ (f )(x) := Γ(α) x (y − x) f (y)dy
where Γ(·) denotes the Gamma function.
These integrals regarded as linear operators from C 0 (I) to C 0 (I) are bounded.
Lemma 75. Let f ∈ C 0 (I). Then
α (b − a)α α (b − a)α
||Ia,− (f )||0 ≤ ||f ||0 , ||Ib,+ (f )||0 ≤ ||f ||0 . (112)
|Γ(α + 1)| |Γ(α + 1)|
Proof. For any x ∈ [a, b], we have:

α 1
Rx
(x − y)α−1 |f (y)|dy ≤

Ia,− (f )(x) ≤
|Γ(α)| a
(113)
||f ||0 (x−a)α ||f ||0
≤ |Γ(α)| α ≤ |Γ(α+1)| (b − a)α ,
where we used αΓ(α) = Γ(α + 1). The first inequality in (112) follows immediately. The
second inequality is proved in a similar fashion.
Definition 76. The left and right Riemann-Liouville fractional derivatives of order α ≥
0 at x ∈ I are defined by:
α (f )(x) := dn n−α
Da,− dxn Ia,− (f )(x)
(114)
α (f )(x) dn n−α
Db,+ := (−1)n dx n Ib,+ (f )(x)
Here n − 1 denotes the integral part [α] of α.
Fractional derivatives partially solve our problem: it may be that a function is nowhere
differentiable in an interval, but that it admits nevertheless a fractional derivative of some
order 0 < α < 1. Indeed this what happens with the Weierstrass function. It can be
proved that it admits fractional derivatives (left and right) of order 0 < α ≤ 2 − r < 1
[26, 35]. Likewise Takagi’s function admits fractional derivatives (left and right) of any
order 0 < α < 1 [37].
However, fractional derivatives pose equally some problems. Let us consider a simple
example.
Example 77. Let f (x) = A(x − a)β and g(x) = B(b − x)β , with β > −1, x ∈ I and
A, B arbitrary real constants. For α ≥ 0 and n = [α] + 1, we obtain [33]:
AΓ(β+1)
α (f )(x) =
Da,− Γ(β+1−α) (x − a)β−α
(115)
α (g)(x) BΓ(β+1)
Db,+ = Γ(β+1−α) (b − x)β−α .
This example illustrates the kind of problem that one faces. For β = 0, we conclude that
the fractional derivatives of non-integer order of constant functions do not vanish. More-
over, from (115), we realize that the fractional derivatives depend crucially on the choice
of terminals a and b. This means that they have a non-local nature. Non-locality may play
to our advantage when trying to model certain memory and hereditary properties of various
materials and processes1 . However, it is very difficult to devise a connection between frac-
tional derivatives and the local geometry of the graph of a function as in standard differential
calculus. Moreover, there are some more practical difficulties: the fractional derivatives of
products obey a complicated rule (not the Leibniz rule) and those of the composition of
functions follow a generalization of the intricate Faà di Bruno formula.
These facts led Kolwankar and Gangal to introduce the so-called local fractional deriva-
tives [4, 7, 17, 18, 19].
1
See chapter 10 of [33] for a survey of applications of fractional calculus and [42, 44].
4.1. Local Fractional Derivatives

Definition 78. Let f ∈ C 0 (I) and α ∈ ( 0, 1] . The right (σ = +) or left (σ = −)
Kolwankar-Gangal local fractional derivative of order α of f at ξ ∈ I is given by:
dα f σ 1
(ξ ) := lim Dα [σ (f − f (ξ))] (x). (116)
dx α Γ(1 + α) x→ξσ ξ,−σ
If the limits exist, we say that f is right or left α-differentiable in the sense of Kolwankar
and Gangal (KG) at ξ.
Perhaps the most remarkable aspect of this definition is the fact that, under certain
circumstances, one may write a generalized Taylor formula. In the following theorem we
denote:
1
Fσ (ξ, x) := Dα [σ (f − f (ξ))] (x + ξ). (117)
Γ(1 + α) ξ,−σ
Theorem 79. Let f ∈ C 0 (I) be right (σ = +) or left (σ = −) α-differentiable at ξ ∈ I in

the KG sense. Then:
dα f σ
f (x) = f (ξ) + σ (ξ )|x − ξ|α + Rσ (x, ξ). (118)
dxα
The remainder satisfies
Rσ (x, ξ)
lim = 0, (119)
x→ξ σ |x − ξ|α
and is given by
x−ξ
dFσ (ξ, τ )
Z
Rσ (x, ξ) = σ |x − ξ − τ |α dτ, (120)
0 dτ
whenever this integral is well defined.
Corollary 80. Under the conditions of Theorem 79., we may write [9]:
dα f σ σ (f (x) − f (ξ))
(ξ ) = (Dxα f )(ξ σ ) := limσ . (121)
dxα x→ξ |x − ξ|α
Definition 81. If the limit in (121) exists or, equivalently, the Taylor formula (118,119)
holds, then we say that f is right (σ = +) or left (σ = −) α-differentiable at ξ in the
sense of Ben Adda and Cresson (BAC). We call (Dxα f )(ξ σ ) the right (σ = +) or left
(σ = −) α-derivative of f at ξ in the sense of BAC.
There is a crucial fact that may have been overlooked in [9, 17, 18, 19]. The condition
that the integral expression in (120) is well defined is in fact hard to fulfill. Thus, in general,
α-differentiability in the sense of KG and BAC are not equivalent. Concomitantly, the
Taylor formula (118,119) may not hold for α-differentiable functions in the sense of KG.
Indeed in (120), we need the quantity:
dFσ 1 d α
(ξ, τ ) = D [σ (f − f (ξ))] (τ + ξ). (122)
dτ Γ(1 + α) dτ ξ,−σ
α d
Since f (x) − f (ξ) vanishes for x = ξ, it is well known that the derivatives Dξ,−σ and dτ
commute [33]. We thus get:
dFσ 1 α d
(ξ, τ ) = Dξ,−σ [σ (f − f (ξ))] (τ + ξ).
dτ Γ(1 + α) dτ
However, the function f (x)−f (ξ) is, by assumption, non-differentiable. And so, in general,
the quantity (122) is not well defined.
Alternatively, we may require a (weaker) integrability condition for KG-differentiability
to imply BAC-differentiability [7].
Proposition 82. (Chen, Yan, Zhang) Let f ∈ C 0 [a, b] be such that the KG derivative
dα f σ α ∞
dxα (ξ ) exists at ξ ∈ (a, b) and Dξ,−σ [σ (f − f (ξ))] (x) belongs to L (ξ, ξ + δ) if σ =
∞
+ or to L (ξ − δ, ξ) if σ = − for some δ > 0. Then:
dα f σ
(ξ ) = (Dxα f ) (ξ σ ). (123)
dxα
Proof. We will prove the result for σ = +. The case σ = − is proved in a similar fashion.
α [(f − f (ξ))] (x) belongs to L∞ (ξ, ξ + δ), then it is integrable and the condition
Since Dξ,−
stated in (2.113), p.71 of [33] is satisfied and we have, for x ∈ (ξ, ξ + δ) that:

−α α [(f − f (ξ))] (x) =
Dξ,− Dξ,−
(124)
(x−y)α−1
h i
α−1
= f (x) − f (ξ) − Dξ,− [(f − f (ξ))] (x)

x=y Γ(α)
where x
1
Z
−α
Dξ,− (g(x)) = (x − y)α−1 g(y)dy (125)
Γ(α) ξ
and the evaluation here |x=y is understood as the limit x → ξ + . Moreover, since f is
continuous in [a, b], we have
Z x
α−1 1 f (y) − f (ξ)
Dξ,− [(f − f (ξ))] (x) = dy → 0 (126)
Γ(1 − α) ξ (x − y)α
as x → ξ + . Consequently:
−α α

f (x) − f (ξ) = Dξ,− Dξ,− [(f − f (ξ))] (x) , x ∈ (ξ, ξ + δ) . (127)
dα f +
Now the proof follows from the fact that dxα (ξ ) =
1 α [(f − f (ξ))] (x). We have
Γ(1+α) lim x→ξ + D ξ,−

f (x)−f (ξ) 1 −α α [(f − f (ξ))] (x) =
(x−ξ) α = (x−ξ) α D ξ,− D ξ,−
1
Rx
= (x−ξ)α Γ(α) ξ (x − y)α−1 Dξ,−
α [(f − f (ξ)) (y)] dy =
(128)
Rx α
= 1
(x−ξ)α Γ(α) ξ (x − y)α−1 α [(f
Dξ,− − f (ξ)) (y)] − Γ(1 + α) ddxαf (ξ + ) dy+
Γ(1+α) Rx α
+ (x−ξ) α Γ(α) ξ (x − y)α−1 ddxαf (ξ + )dy := I1 + I2 .
By a straightforward calculation, we have:

Γ(1+α) Rx α
I2 = (x−ξ)α Γ(α) ξ (x − y)α−1 ddxαf (ξ + )dy =
(129)
αΓ(α) dα f + (x−ξ)α dα f +
= (x−ξ) Γ(α) dxα (ξ )
α α = dxα (ξ ).
Thus we only need to show that I1 → 0 as x → ξ + . By definition we have that, for any
ǫ > 0, there exists 0 < η < ǫ such that
dα f +

1 α
Γ(1 + α) Dξ,− [(f − f (ξ)) (y)] − dxα (ξ ) < ǫ (130)

whenever y ∈ (ξ, ξ + η).

Therefore for any x ∈ (ξ, ξ + η):

Γ(1+α) R x α−1 1 α dα f +
|I1 | = (x−ξ) α Γ(α) ξ (x − y) Γ(1+α) Dξ,− [(f − f (ξ)) (y)] − dxα (ξ ) dy ≤

α
Rx 1 dα f +
≤ (x−ξ)α ξ (x − y)α−1 Γ(1+α) α [(f − f (ξ)) (y)] −
Dξ,− dxα (ξ ) dy ≤
α
Rx
≤ (x−ξ)α ξ (x − y)α−1 ǫdy = ǫ
(131)
which completes the proof.
There is an important structure theorem of Chen, Yan and Zhang that states that if
the local KG fractional derivatives of a function of order 0 < α < 1 exist for a Hölder
continuous function a.e. in an interval, then they vanish a.e. in that interval.
To prove that theorem we need some preliminary results and definitions [7].
dα f +
Lemma 83. (Chen, Yan, Zhang) Let f : [a, b] → R be continuous and such that dxα (ξ )
exists for some ξ ∈ (a, b). Then
1
f (ht + ξ) − f (ξ)
Z
lim (1 − t)−α dt (132)
h↓0 0 hα
exists and
1
dα f + 1 f (ht + ξ) − f (ξ)
Z
α
(ξ ) = lim (1 − t)−α dt (133)
dx Γ(1 + α)Γ(1 − α) h↓0 0 hα
Proof. By definition, we have:

x
dα f + 1 1 d f (t) − f (ξ)
Z
α
(ξ ) = lim dt. (134)
dx Γ(1 + α) x↓ξ Γ(1 − α) dx ξ (x − t)α
Let x
1 f (t) − f (ξ)
Z
Fξ (x) := dt, (135)
Γ(1 − α) ξ (x − t)α
which can also be written as

1
(x − ξ)1−α (ξ + (x − ξ)s) − f (ξ)
Z
Fξ (x) = ds, (136)
Γ(1 − α) 0 (1 − s)α
(t−ξ) α
if we change the variable to s = (x−ξ) . We observe from the definition of ddxαf (ξ + ) that
Fξ′ (x) must exist in a small interval ( ξ, ξ+δ ]. Clearly, Fξ (x) is also continuous in [ξ, ξ + δ]
if we define Fξ (ξ) = 0. Thus by the mean value theorem, we have, on one hand that for
each fixed h ∈ (0, δ), there is some ρh ∈ (0, h), such that
Fξ (ξ + h) − Fξ (ξ)
Fξ′ (ξ + ρh ) = . (137)
h
On the other hand, we have, by the equivalent definition of Fξ (x) above, that
Fξ (ξ+h)−Fξ (ξ) Fξ (ξ+h)
h = h =
h1−α
R1 f (ξ+(x−ξ)s)−f (ξ)
= Γ(1−α) 0 h(1−s)α ds = (138)
R1
= 1
Γ(1−α) 0 (1 − s)−α f (sh+ξ)−f
hα
(ξ)
ds
dα f + 1
Since by definition, the existence of dxα (ξ ) is equivalent to Γ(1+α) limx↓ξ Fξ′ (x), we have
1 dα f +
Γ(1+α) limh↓0 Fξ′ (ξ + ρh ) = dxα (ξ ). Therefore
1
f (ht + ξ) − f (ξ)
Z
lim (1 − t)−α dt (139)
h↓0 0 hα
exists and
1
dα f + 1 f (ht + ξ) − f (ξ)
Z
(ξ ) = lim (1 − t)−α dt (140)
dxα Γ(1 + α)Γ(1 − α) h↓0 0 hα
Stein and Zygmund [38] considered the α-fractional derivative in the sense of M. Riesz
for functions defined on R and its variations including the Weyl fractional derivative.
Definition 84. Let 0 < α < 1 and β = 1 − α. Given a measurable function f : R → R,

the β-th integral fβ for f is defined by
f (y)
Z
fβ (x) := 1−β
dy = (f ⋆ K1−β ) (x), (141)
R |x − y|
which is the convolution between f and Kγ (x) = |x|−γ . The α-fractional derivative of f
at x, denoted by f (α) (x), is defined by
d
f (α) (x) := fβ (x). (142)
dx
To characterize the existence of such α-fractional derivatives requires the following

definition [38].
Definition 85. The function f : R → R is said to satisfy Λα (0 < α < 1) at x if
Rx (t) := f (x + t) − f (x) = O (|t|α ) as t → ∞. (143)
On the other hand it satisfies the condition Nα2 at x if
[Rx (t)]2
Z δ
1+2α
dt < +∞ for some δ > 0. (144)
−δ |t|
Stein and Zygmund proved the following result [38].
Proposition 86. (Stein, Zygmund) Suppose f ∈ L1 (R) and satisfies the condition Λα for
each point x of a set E ⊂ R of positive measure. Then f (α) (x) exists almost everywhere in
E if and only if f satisfies Nα2 almost everywhere in E.
Remark 87. All the results of [38] remain valid if one replaces fβ by Weyl’s α-fractional
derivative: Z x
d f (y)
Iβ (x), where Iβ (x) := 1−β
dy. (145)
dx ∞ |x − y|
It is easy to see [20] that the KG local fractional derivative can be related to Weyl’s
fractional derivative as follows.
Remark 88. Let f be locally α-Hölder continuous in [a, b]. For any fixed ξ ∈ [a, b] we
define
+ f (x) − f (ξ), ξ < x < b,
fξ (x) := (146)
0, x ≤ ξ or x ≥ b,
and
f (x) − f (ξ), a < x < ξ,
fξ− (x) := (147)
0, x ≥ ξ or x ≤ a.
Let
x fξ+ (t) ∞ fξ− (t)
Z Z
Iβ+ fξ+ (x) = 1−β
dt, Iβ− fξ− (x) = dt, (148)
−∞ (x − t) x (t − x)1−β
and as usual β = 1 − α. Then the right (σ = +) and left (σ = −) KG local fractional

derivatives at ξ ∈ [a, b] are given by
dα f 1 d σ σ
f (ξ σ ) = limσ I f (x). (149)
dx α Γ(1 + α)Γ(1 − α) x→ξ dx β ξ
We then have the following remarkable theorem [7].
Theorem 89. (Chen, Yan, Zhang) Let f : [a, b] → R be locally α-Hölder continuous in
α
[a, b] for 0 < α < 1. Moreover, suppose that the KG local fractional derivatives ddxαf f (ξ ± )
α α
exist for almost all ξ ∈ (a, b). Then ddxαf f (ξ + ) = ddxαf f (ξ − ) = 0 for almost all ξ ∈ (a, b).
dα f
Proof. From Remark 88. we realize that if dxα f (ξ ± ) exist at some ξ ∈ (a, b), there is a
neighborhood Vξ (τ ) = (ξ − τ, ξ + τ ) ⊂ (a, b) with τ > 0, such that
d + + d − −
I f (x) exists for x ∈ Vξ+ (τ ), I f (x) exists for x ∈ Vξ− (τ ), (150)
dx β ξ dx β ξ
and β = 1 − α.
Given that, by assumption, f is locally α-Hölder continuous, we then conclude from
Proposition 86. that fξ+ and fξ− satisfy condition Nα2 a.e. in Vξ+ (τ ) and Vξ− (τ ), respectively.
Note that the function Rx (t) (143) when applied to fξ+ reads: Rx (t) = fξ+ (t+x)−fξ+ (x) =
f (t + x) − f (ξ) − f (x) + f (ξ) = f (t + x) − f (x) if x ∈ Vξ+ (τ ) and sufficiently small t.
A similar result holds for fξ− . This means that Rx (t) is independent of f (ξ) for x ∈ Vξ (τ ).
Thus Rx (t) = f (t + x) − f (x), and hence Nα2 holds for f at almost everywhere in Vξ (τ ).
For those x ∈ [a, b] for which Nα2 holds, there exists a sufficiently small δ > 0 such that
Rδ 2
(x − δ, x + δ) ⊂ (x − τ, x + τ ) and −δ [R x (t)]
|t|2α+1
dt < +∞. By Vitali’s equi-integrability
theorem for Lebesgue integrals [24], we have
h
[Rx (t)]2
Z
lim dt = 0. (151)
h↓0 0 |t|2α+1
dα f +
Since dxα (x ) exists, we have by Lemma 83. that
1
dα f + 1 f (ht + x) − f (x)
Z
(x ) = lim (1 − t)−α dt. (152)
dxα Γ(1 + α)Γ(1 − α) h↓0 0 hα
We will now prove that

1

−α f (ht + x) − f (x)
Z

lim
h↓0
(1 − t) hα dt = 0, (153)
0
which completes the proof.

Since f is locally α-Hölder continuous in [a, b], there is constant M > 0 such that
|f (x) − f (y)| ≤ M |x − y|α for y sufficiently close to x. For 0 < ǫ < 1, we have:
R1
−α f (ht+x)−f (x)
0 (1 − t) hα dt =
R1
−α f (ht+x)−f (x) dt + 1−ǫ (1 − t)−α f (ht+x)−f (x) dt (154)
R
= 1−ǫ (1 − t) h α 0 h α
:= I1 + I2 .
We then have for sufficiently small ǫ

R1 R1
I1 = 1−ǫ (1 − t)−α f (ht+x)−f (x)
dt ≤ M 1−ǫ (1 − t)−α tα dt ≤

h α
(155)
R1 M 1−α
≤M 1−ǫ (1 − t)−α dt = 1−α ǫ .
On the other hand, by applying the Cauchy-Schwarz inequality we obtain:

R 1−ǫ
I2 = 0 (1 − t)−α f (ht+x)−f (x)
dt ≤

h α
1
R 1−ǫ f (ht+x)−f (x) 1
R (1−ǫ)h f (s+x)−f (x) ds
≤ ǫα 0 hα dt = ǫα 0 hα h ≤
Rh
≤ 1
ǫα h1+α 0 sα+1/2 |f (s+x)−f
sα+1/2
(x)|
ds ≤ (156)
1 R 1
|f (s+x)−f (x)|2
R
1 h 2α+1 2 h 2
≤ ǫα h1+α 0 s ds 0 s2α+1
ds =
1
h |f (s+x)−f (x)|2
R
2
= √1 ds .
ǫα 2α+2 0 s2α+1
Since Nα2 holds for f at x, we have from (155) and (156)

Z 1
−α f (ht + x) − f (x) M 1−α

0 ≤ lim (1 − t) α
dt ≤ ǫ (157)
h↓0 0 h 1−α
for any sufficiently small ǫ > 0, which proves that (153) holds.
From Lemma 83. and Theorem 89., we realize that dividing the difference f (x + h) −
f (x) by the powers hα may not be sufficient to ”resolve” the details of certain types of
Hölder regularity. The fact that local fractional derivatives vanish almost everywhere, may
be an indication that a refinement may be required. Just as in the definition of Hausdorff di-
mension, dimension functions other than the powers may play an important role for certain
sets, it may be that difference-quotients of the form f (x+h)−f
ψ(h)
(x)
may be more adequate for
certain types of regularity.
Let F be a scale of dimension functions and define Q to be the subset of functions φ of
F/ ∼ such that:
φx (158)
Definition 90. A function f : I → R is right (σ = +) or left (σ = −) differentiable to

order ψ ∈ Q at ξ ∈ I, if there exists a real number λσ such that for x ∈ Vξσ (ǫψ ) ∩ I:
f (x) = f (ξ) + σλσ ψ (|x − ξ|) + Rσ (x, ξ), (159)
with
Rσ (x, ξ)
lim =0 (160)
x→ξ σ ψ (|x − ξ|)
ψ ψ
We denote by C+ (I)(resp. C− (I)) the set of right (resp. left) differentiable functions to order
ψ in [ a, b ) (resp. ( a, b ]). Finally, we say that f is ψ-differentiable if it is both right and
ψ ψ
left ψ-differentiable. In this case, we write C ψ (I) = C+ (I) ∩ C− (I).
Definition 91. From the Taylor formula (159,160) it follows immediately that
σ (f (x) − f (ξ)) σ (f (ξ + σh) − f (ξ))
λσ = limσ = lim , σ=± (161)
x→ξ ψ (|x − ξ|) h↓0 ψ(h)

We denote these limits by Dtψ f (ξ σ ) and call them the right (σ = +) and left (σ = −)
ψ-derivatives or ψ-velocities of f at ξ.
Notice that for ψ(x) = x we recover the ordinary right-sided and left-sided derivatives
and for ψ(x) = xr (0 < r < 1) we obtain the local fractional derivatives of Ben Adda and
Cresson.
Remark 92. A comment is now in order concerning the dependence of ψ-differentiability

on the representative of the equivalence class [ψ] ∈ Q. Take ψ1 , ψ2 ∈ [ψ]. Then there exist
ψ2 (x)
constants C1 , C2 such that 0 < C1 ≤ ψ 1 (x)
≤ C2 for 0 < x ≤ ǫ = min {ǫψ1 , ǫψ2 }. Then
we have for 0 < |x − ξ| < ǫ:
σ (f (x) − f (ξ)) σ (f (x) − f (ξ)) ψ2 (|x − ξ|)
= (162)
ψ1 (|x − ξ|) ψ2 (|x − ξ|) ψ1 (|x − ξ|)
We conclude that:

(i) if Dxψ2 f (ξ σ ) ≥ 0, then

0 ≤ C1 Dxψ2 f (ξ σ ) ≤ Dxψ1 f (ξ σ ) ≤ C2 Dxψ2 f (ξ σ );

(ii) if Dxψ2 f (ξ σ ) < 0,

then 0 ≤ C2 Dxψ2 f (ξ σ ) ≥ Dxψ1 f (ξ σ ) ≥ C1 Dxψ2 f (ξ σ ).
For all practical
purposes, we can
extract exactly the same information from the two deriva-
ψ1 ψ2
tives Dx f (ξ ) and Dx f (ξ σ ). Indeed, f is ψ1 -differentiable if and only if it is ψ2 -
σ
differentiable for all ψ2 ∼ ψ1 . Moreover, the sign of the derivatives is the same (which
means that we will be able to draw the same conclusions concerning the monotonicity of
the function f and the existence of extrema). We thus fix once and for all a representative
of each equivalence class in Q.
In this work we shall focus on ψ-diferentiability for ψ ≺ x. We may nevertheless

envisage the possibility of higher order differentiability:
Definition 93. Let f : I → R be of class C n (n ≥ 1) at ξ ∈ I, but not (n + 1)-times

differentiable at ξ. Moreover, let ψ ∈ F be such that xn ≺ ψ ≺ xn+1 . We say that f is
right (σ = +) or left (σ = −) ψ-differentiable at ξ if there exists a real number λσ such
that for any x ∈ Vξσ (ǫψ ) ∩ I:
n
X f (k) (ξ)
f (x) = (x − ξ)k + σλσ ψ (|x − ξ|) + Rσ (x, ξ), (163)
k!
k=0
with
Rσ (x, ξ)
lim = 0. (164)
ψ (|x − ξ|)
x→ξ σ
The numbers λσ are the ψ-derivatives of f at ξ:

(k)

σ f (x) − nk=0 f k!(ξ) (x − ξ)k
P

λσ = Dxψ f (ξ σ ) := limσ , σ = ±. (165)
x→ξ ψ (|x − ξ|)
4.2. ψ-differential Calculus

In this section we explore some of the properties of local fractional derivatives and the
main theorems of ψ-differential calculus. Some of them were derived in various papers
[4, 9, 10, 11] for differentiable functions to order α ∈ (0, 1). We include them here for
completeness and with the additional generalization to functions which are differentiable to
order ψ ∈/ P.
As in ordinary calculus, right (resp. left) ψ-differentiability entails right (resp. left)
continuity.
Proposition 94. Let f ∈ C 0 (I), ψ ∈ Q and ξ ∈ I.

(i) If f is right (resp. left) ψ-differentiable at ξ, then f is right (resp. left) continuous at ξ.
(ii) If f is ψ-differentiable at ξ, then f is continuous at ξ.

Proof. Let ρσ (x) := σ(fψ(|x−ξ|)
(x)−f (ξ))
, for x ∈ Vξσ (ǫψ )\ {ξ} ∩ I. If f is right (σ = +) or

left (σ = −) ψ-differentiable at ξ, then the limit limx→ξσ ρσ (x) = Dxψ f (ξ σ ) exists. We
thus have:

limσ σ (f (x) − f (ξ)) = limσ ρσ (x) · ψ (|x − ξ|) = Dxψ f (ξ σ ) · 0 = 0,
x→ξ x→ξ
which means that f is right (σ = +) or left (σ = −) continuous at ξ. This proves statement

(i). If f is ψ-differentiable at ξ, then from (i) f is right and left continuous at ξ, which
proves (ii).
The following result reveals that ψ-derivatives are true derivatives, in the sense that
they are linear operators which obey the Leibniz rule. Moreover, they satisfy the standard
quotient rule. We leave the simple proofs to the reader.
Proposition 95. Let f, g ∈ C 0 (I) be right (σ = +) (resp. left (σ = −)) ψ-differentiable

at ξ ∈ I, with ψ ∈ Q, and K ∈ R. Then, f + g, Kf and f · g are right (σ = +) (resp. left
(σ = −)) ψ-differentiable at ξ ∈ I and we have:

Dxψ (f + g) (ξ σ ) = Dxψ f (ξ σ ) + Dxψ g (ξ σ )

Dxψ (Kf ) (ξ σ ) = K Dxψ f (ξ σ ) (166)

Dxψ (f · g) (ξ σ ) = Dxψ f (ξ σ ) · g(ξ) + f (ξ) · Dxψ g (ξ σ )
Proposition 96. Let f, g ∈ C 0 (I) be right (σ = +) (resp. left (σ = −)) ψ-differentiable

at ξ ∈ I, with ψ ∈ Q and g(ξ) 6= 0. Then fg is right (resp. left) ψ-differentiable at ξ, and
we have:

ψ σ ) · g(ξ) − f (ξ) · D ψ f (ξ σ )

f D x f (ξ x
Dxψ (ξ σ ) = 2 . (167)
g [g(ξ)]
In the general case a chain rule is very difficult to derive. There is a particular case
where this is possible.
Letg and f be continuous functions

Proposition 97. on [a, b] and [g(a), g(b)], respectively.
φ σ r s σ
Suppose that Dx g (ξ ) and (Dx f ) g(ξ) exist for ξ ∈ (a, b), φ ∈ Q, 0 < r ≤ 1 and

sσ = sign Dxφ g (ξ σ ) . Then, f ◦ g is σ differentiable to order tr ◦ φ = φr at g(ξ), and
we have: r r
σ
Dxφ (f ◦ g) (ξ σ ) = sσ (Dxr f ) g(ξ)σs Dxφ g (ξ σ )

(168)

Proof. Notice that

σ(f (g(x))−f (g(ξ)))
φr (|x−ξ|) =
(169)
σ (g(x))−f (g(ξ))) g(x)−g(ξ) r

= sσ σs (f|g(x)−g(ξ)|r φ(|x−ξ|) .

Since Dxφ g (ξ σ ) exists, we have from from (159,160):

g(x) = g(ξ) + σsσ Dxφ g (ξ σ ) φ (|x − ξ|) + Rσ (x, ξ),

Rσ (x,ξ) σ
with limx→ξσ φ(|x−ξ|) = 0. We conclude that g(x) → (g(ξ))σs as x → ξ σ . Taking the
limit x → ξ σ in (169), we recover (168).
There are some particular cases of interest:
Corollary 98. Let g and f be continuous in [a, b] and [g(a), g(b)], respectively. Let ξ ∈
(a, b).
(i) If g is σ-differentiable to order φ ∈ Q at ξ and f is differentiable at g(ξ), then f ◦ g is

σ-differentiable to order φ at ξ and:

Dxφ (f ◦ g) (ξ σ ) = f ′ (g(ξ)) Dxφ g (ξ σ ). (170)
(ii) If g is differentiable at ξ with s = sign (g ′ (ξ)), and (Dxr f ) (g(ξ)σs ) exists, for 0 < r <
1, then f ◦ g is σ-differentiable to order xr at ξ and:
r
(Dxr (f ◦ g)) (ξ σ ) = s [(Dxr f ) (g(ξ)σs )] g ′ (ξ) .

(171)
(iii) It g is the scale transformation g(x) = λx (λ > 0) and (Dxr f ) (g(ξ)σ ) exists, then:
(Dxr (f ◦ g)) (ξ σ ) = λr (Dxr f ) (λξ σ ) . (172)
Proof. The proof follows immediately from Proposition 97.

Let us now analyze the structure of the ψ-derivatives of a function as ψ runs through Q.
Theorem 99. Let f ∈ C 0 (I) and ψ ∈ Q. If f is right (σ = +) or left (σ = −) ψ-

differentiable at ξ ∈ I, then

Dxφ f (ξ σ ) = 0, for all φ ∈ Q such that φ ≺ ψ. (173)
Proof. Let f be right or left ψ-differentiable at ξ and let φ ∈ Q be such that φ ≺ ψ. We

then have:

Dxφ f (ξ σ ) = limx→ξσ σ(fφ(|x−ξ|)
(x)−f (ξ))
= limx→ξσ σ(fψ(|x−ξ|)
(x)−f (ξ))
×

ψ(|x−ξ|) ψ(|x−ξ|)
× limx→ξσ φ(|x−ξ|) = Dxψ f (ξ σ ) × limx→ξσ φ(|x−ξ|) = 0.
n In viewof thistheorem, as
o Q ⊂ F with F a scale (which is bounded complete), the set
φ σ
φ ∈ Q : Dx f (ξ ) = 0 regarded as a subset of the chain Q has a supremum.
Definition 100. The right (σ = +) or left (σ = −) critical order of differentiability of

a function f ∈ C 0 (I) at a point ξ ∈ I is given by:
n o
σ
ψξ,c := supF φ ∈ Q : Dxφ f (ξ σ ) = 0 (174)
If the critical order is the same throughout the interval I, then we may simply write ψcσ .
A simple consequence of this definition and of Theorem 99. is:
Proposition 101. Let f ∈ C 0 (I) have critical order of differentiability ψξ,c

σ at ξ ∈ I with

σ ≺ x. Then D φ f (ξ σ ) does not exist for φ ∈ Q with ψ σ ≺ φ.
ψξ,c x ξ,c

Proof. Suppose that Dxφ f (ξ σ ) exists for ψξ,c
σ ≺ φ. Then, from Theorem 99., we would
have (Dxρ f ) (ξ σ ) = 0 for any ρ ∈ Q such that ψξ,cσ ≺ ρ ≺ φ. Note that such a ρ exists
because F is a scale (and hence dense). However, this is contradictory with the assumption
σ being the critical order of differentiability.
of ψξ,c
This behavior is easily illustrated with the following example.
Example 102. Consider the function f (x) = |x|r , 0 < r < 1. From (161), we easily
obtain for ψ ∈ Q:
 0, if ψ ≺ xr


ψ σ
Dx f (0 ) = σ, if ψ = xr , for σ = ±
∞, if xr ≺ ψ x

Also for the derivatives of order r, we have:

σ, if x = 0
(Dxr f ) (xσ ) = , for σ = ±
0, if x 6= 0
The last equation reveals that, contrary to ordinary derivatives, local fractional derivatives
do not obey Darboux’s theorem.
Proposition 103. Let f ∈ C 0 (I) be differentiable at ξ ∈ (a, b) to order ψ ∈ Q. Then

(i) r Dxψ f (ξ − ) < 0 ⇒ ∃ǫ > 0, such that ∀x ∈ (ξ − ǫ, ξ), we have r(f (x) − f (ξ)) > 0,
r = ±.

(ii) r Dxψ f (ξ + ) < 0 ⇒ ∃ǫ > 0, such that ∀x ∈ (ξ, ξ + ǫ), we have r(f (x) − f (ξ)) < 0,
r = ±.
Proof. From (159,160), we have, for sufficiently small ǫ > 0, and for x ∈ Vξσ (ǫ)\ {ξ}:

ψ σ
σ
f (x) − f (ξ) = σs Dx f (ξ ) ψ (|x − ξ|) + Rσ (x, ξ),

with sσ
= sign Dxψ f , and R
(ξ σ ) σ (x,ξ)
ψ(x−ξ) → 0 as x → ξ .
σ

If Dxψ f (ξ σ ) 6= 0, then the sign of σ(f (x) − f (ξ)) is sσ (for sufficiently small ǫ).
The converse result of this theorem is as follows:
Proposition 104. Let f ∈ C 0 (I) be differentiable at ξ ∈ (a, b) to order ψ ∈ Q. Then
(i) ∃ǫ > 0, such that ∀x ∈ (ξ − ǫ, ξ), we have r(f (x) − f (ξ)) > 0, r = ± ⇒
r Dxψ f (ξ − ) ≤ 0 .
∃ǫ >
(ii) 0, such that ∀x ∈ (ξ, ξ + ǫ), we have r(f (x) − f (ξ)) < 0, r = ± ⇒
ψ
r Dx f (ξ + ) ≤ 0 .
Proof. The proof follows the same rationale as the previous one. However, we cannot
ψ
discard the possibility that Dx f (ξ σ ) may vanish. This is certainly what happens if
ψ is below the critical order of differentiability. And even for critical orders a vanishing
derivative is not precluded.
An immediate consequence of this concerns the local extrema of functions.
Proposition 105. Let f ∈ C 0 (I) be differentiable at ξ ∈ (a, b) to order ψ ∈ Q.

(i) If ξ is a local maximum, then σ Dxψ f (ξ σ ) ≤ 0, σ = ±.

(ii) If ξ is a local minimum, then σ Dxψ f (ξ σ ) ≥ 0, σ = ±.
Proof. (i) Suppose that ξ is a local maximum. Then there exists ǫ > 0 such that f (x) −
f (ξ) ≤ 0 for all x ∈ Vξ (ǫ). For x ∈ (ξ − ǫ, ξ), we have (fψ(|x−ξ|)
(ξ)−f (x))
≥ 0. Taking the limit

x → ξ − , we obtain Dxψ f (ξ − ) ≥ 0. Likewise, for x ∈ (ξ, ξ + ǫ), we have (fψ(|x−ξ|)
(x)−f (ξ))
≤

0. If we take the limit x → ξ + , we conclude that Dxψ f (ξ + ) ≤ 0.
(ii) This is proved in a similar way.
We leave the proof of the converse result to the reader:
Proposition 106. Let f ∈ C 0 (I) be differentiable at ξ ∈ (a, b) to order ψ ∈ Q.

(i) If σ Dxψ f (ξ σ ) < 0, σ = ±, then ξ is a local maximum.

(ii) If σ Dxψ f (ξ σ ) > 0, σ = ±, then ξ is a local minimum.
We are now in a position to generalize some of the fundamental theorems of differential

calculus.
Theorem 107. (Rolle) Let f ∈ C 0 (I) be differentiable in (a, b) to order ψ ∈ Q. Moreover

suppose that f (a) = f (b). Then, there exists ξ ∈ (a, b) such that:

Dxψ f (ξ + ) · Dxψ f (ξ − ) ≤ 0. (175)
Proof. Since f is continuous in I, by Weierstrass’ Theorem, there exist, at least, two points
u, v ∈ I, such that f (u) ≤ f (x) ≤ f (v), ∀x ∈ I. That is, f (u), f (v) are a global
minimum and maximum of f in I, respectively. If u and v coincide with the a and b, then
ψ
from f (a) = f (b), we conclude that f is constant in I, and thus Dx f (ξ σ ) = 0, for any
ξ ∈ (a, b). If, u or v ∈ (a, b), then from Proposition 105., (175) follows immediately.
There is an alternative form of Rolle’s theorem, which will be useful in the sequel. Its
main advantage is the fact that we need to require only left or only right ψ-differentiability.
Theorem 108. Let f ∈ C 0 (I) with f (a)= f (b) and let ψ ∈ Q.

ψ
(i) If f is right ψ-differentiable at a and Dx f (a+ ) > 0 (resp.

Dxψ f (a+ ) < 0), then there exists a local maximum (resp. minimum)
of f in (a, b).
(ii) If f is left ψ-differentiable at b and Dxψ f (b− ) > 0 (resp.

Dxψ f (b− ) < 0), then there exists a local minimum (resp. maximum)
of f in (a, b).
Proof. Since f is continuous in I = [a, b], there exist u, v ∈ I, such that f (u) ≤ f (x) ≤
f (v), ∀x ∈ I.
ψ
(i) Suppose that Dx f (a+ ) > 0. If v = a or v = b, then from Proposition 105. (i),
we
would have a contradiction. And thus v ∈ (a, b). Conversely, let us assume that
ψ
Dt f (a+ ) < 0. If u = a or u = b, then from Proposition 105. (ii), we would again
have a contradiction. And thus u ∈ (a, b).
(ii) This is proved in a similar way.
An immediate consequence of this is the following.
ψ ψ
Theorem
109. Let ψ ∈ Q with ψ ≺ x,and let f ∈ C+ (I) (resp. f ∈ C− (I)) be such that
Dxψ f (x+ ) 6= 0, ∀x ∈ [ a, b ) (resp. Dxψ f (x− ) 6= 0, ∀x ∈ ( a, b ]).
n o
(+)
(i) The sets of points Fr = x ∈ [ a, b ) : sign Dxψ f (x+ ) = r , with r = ± are
both dense in I. n o
(−)
(ii) The sets of points Fr = x ∈ ( a, b ] : sign Dxψ f (x− ) = r , with r = ± are
both dense in I.
Proof. We start by proving that if the sign of the ψ-derivative remains constant throughout
a non-degenerate
interval, then the function is strictly monotone in that interval. Suppose
ψ +
that Dx f (x ) > 0, ∀x ∈ [ a, b ), but f is not a strictly increasing function in I. Then,
by continuity, it is always possible to find two points x1 , x2 , such that a ≤ x1 < x2 ≤ b
and f (x1 ) = f (x2 ). Consequently,
in the interval [x1 , x2 ] , f satisfies the conditions of
ψ +
Theorem 108. with Dx f (x1 ) > 0. It then follows that there exists a local maximum

at some v ∈ (x1 , x2 ). However, from Proposition 105. (i), we must have Dxψ f (v + ) ≤
0, which
is a contradiction. And so, f is strictly increasing. Conversely, suppose that
ψ
Dx f (x+ ) < 0, ∀x ∈ [ a, b ), but that f is not strictly decreasing in I. Again, there

exist a ≤ x3 < x4 ≤ b such that f (x3 ) = f (x4 ) and Dxψ f (x+ 3 ) < 0. By Theorem
108., there
exists a local minimum at some u ∈ (x3 , x4 ). From Proposition 105. (ii),
ψ +
Dx f (u ) ≥ 0, which is a contradiction. And thus, f is strictly decreasing. The proof

for Dxψ f (x− ) is analogous.
Now,
suppose that there exists some non-degenerate interval J ⊂ I, such that
ψ σ
sign Dx f (x ) is constant and non-zero in J. From the previous discussion, we con-
clude that f is strictly monotone in J. But by Lebesgue’s Theorem, f must differentiable
almost everywhere in J, which contradicts the hypothesis ψ ≺ x. This means that, given
ψ σ
x ∈ I and an arbitrary neighborhood of x, we can find points where Dx f (x ) is positive
and points where it is negative.
This leads to the fact that non-vanishing derivatives of order ψ ≺ x cannot be continu-
ous.
ψ ψ
Corollary 110. Let ψ ∈ Q with ψ ≺ x, and let f ∈ C+ (I) (resp.
f ∈ C− (I)). If the local
fractional derivative Dxψ f (xσ ) is continuous at ξ ∈ I, then Dxψ f (ξ σ ) = 0.

Proof. Suppose that Dxψ f (xσ ) is continuous and non-zero at ξ. Then, by continuity,

there exists ǫ > 0, such that sign Dxψ f (xσ ) remains constant and non-zero in Vξ (ǫ) ∩

I. However, this contradicts Theorem 109.. And thus Dxψ f (ξ σ ) = 0.
From these results, we can derive the Theorem 113. below concerning the fractal (Haus-
dorff) dimension of positive and negative parts of the graph of a local ψ-derivative function.
But first we need the following proposition and theorem. Here dimH (F ) will denote the
Hausdorff dimension of set F and Lθ is the line through the origin of R2 that makes an
angle θ with the horizontal axis. Finally, we denote the orthogonal projection onto Lθ by
projθ .
Proposition 111. A set F ⊂ Rn (n ≥ 1) with dimH (F ) < 1 is totally disconnected.

Proof. The proof of this proposition can be found in [12].
Theorem 112. (Projection Theorem) Let F ⊂ R2 be a Borel set.
(i) If dimH (F ) ≤ 1 then dimH (projθ F ) = dimH (F ) for almost all θ ∈ [ 0, π ).
(ii) If dimH (F ) > 1 then projθ F has positive length (as a subset of Lθ ) and so has dimen-
sion 1 for almost all θ ∈ [ 0, π ).
Proof. The proof of this theorem can be found in [12].
ψ ψ
Theorem
Let ψ ∈ Q with ψ ≺ x, andlet f ∈ C+ (I) (resp. f ∈ C− (I)) be such
113.
that Dxψ f (x+ ) 6= 0, ∀x ∈ [ a, b ) (resp. Dxψ f (x− ) 6= 0, ∀x ∈ ( a, b ]). Then the
n o
(σ)
positive Γσ+ := x, Dxψ f (xσ ) : x ∈ F+ part of the graph of Dxψ f (xσ ) and the
negative partn o
(σ)
Γσ− := x, Dxψ f (xσ ) : x ∈ F− both have Hausdorff dimension greater or
equal to 1.

Proof. Let Dxψ f (xσ ) 6= 0, for x ∈ [ a, b ) if σ = + or x ∈ ( a, b ] if σ = −. Since,
(σ)
from Theorem 109., Fr is dense in I, it cannot be totally disconnected. From Proposition
(σ)
111., we conclude that dimH Fr = 1.
Let us now define
n o
Γσr := x, Dxψ f (xσ ) : x ∈ Fr(σ) , r, σ = ± (176)

Obviously, the graph of Dxψ f (xσ ) is Γσ = Γσ+ ∪ Γσ− . Consequently, from the Projection
Theorem 112., we conclude that:
dimH Γσr ≥ dimH Frσ = 1.
Likewise, we easily derive a generalized version of Cauchy’s mean value theorem.

Theorem 114. (Cauchy) Let f, g ∈ C 0 (I) be differentiable in (a, b) to order ψ ∈ Q.
Then, there exists ξ ∈ (a, b), such that:

σ [f (b) − f (a)] Dxψ g (ξ σ ) ≥ σ [g(b) − g(a)] Dxψ f (ξ σ ), (177)
or
σ [f (b) − f (a)] Dxψ g (ξ σ ) ≤ σ [g(b) − g(a)] Dxψ f (ξ σ ), (178)
Proof. Let us consider the function h(x) = f (x)(g(b) − g(a)) − g(x)(f (b) − f (a)). By
linearity, this function is continuous in I and ψ-differentiable in (a, b). Moreover, h(a) =
h(b). From the generalized Rolle Theorem 107., there exists ξ ∈ (a, b), such that

σ Dxψ h (ξ σ ) ≤ 0, or σ Dxψ h (ξ σ ) ≥ 0, for σ = ±.
The first inequality is equivalent to (177) and the second one to (178).
The following result generalizes L’Hopital’s rule.
Proposition 115. (L’Hôpital) Let f, g ∈ C 0 (I) be differentiable functions at ξ ∈ (a, b)

to order ψ ∈ Q. Suppose that there exists ǫ > 0 such that g(x) 6= 0, ∀x ∈
((ξ − ǫ, ξ + ǫ) \ {ξ})∩I. Moreover, suppose that f (ξ) = g(ξ) = 0 and that Dxψ g (ξ σ ) 6=
f (x)
0. Then, the limit limx→ξσ g(x) exists and we have:

ψ
f (x) D x f (ξ σ )
lim = . (179)
x→ξ σ g(x)
Dxψ g (ξ σ )
Proof. For x ∈ ((ξ − ǫ, ξ + ǫ) \ {ξ}) ∩ I, we have:

f (x) f (x) − f (ξ) σ (f (x) − f (ξ)) ψ (|x − ξ|)
= = .
g(x) g(x) − g(ξ) ψ (|x − ξ|) σ (g(x) − g(ξ))
If we take the limit x → ξ σ , we obtain (179).
Example 116. Let us consider the limit
T (x)
lim ,
x↓0 ψ(x)
where T (·) denotes Takagi’s function and
1

x log2 x if x > 0
ψ(x) = .
0 if x = 0
Clearly T (0) = ψ(0) = 0 and, from Krüppel’s Theorem 12., both are right ψ-differentiable
at x = 0. Since Dxψ T (0+ ) = Dxψ ψ(0+ ) = 1, from Proposition 115., we have:
T (x)
limx↓0 = 1.
ψ(x)
Let us now relate the critical order of differentiability with the critical order of Hölder
continuity. We start by proving the following lemma.
Lemma 117. If f ∈ C 0 (I) is locally right (resp. left) ψ-Hölder continuous at ξ ∈ I for
ψ ∈ Q, then it is right left) differentiable at ξ to all orders φ ∈ Q such that φ ≺ ψ.
(resp.
φ
Moreover, we have Dx f (ξ σ ) = 0, for σ = + (resp. σ = −).
Proof. If f is locally right (σ = +) or left (σ = −) ψ-Hölder continuous at ξ ∈ I with

ψ ∈ Q, then there exist positive constants K and ǫ such that:
|f (x) − f (ξ)| ≤ Kψ (|x − ξ|) , ∀x ∈ Vξσ (ǫ). (180)
For x ∈ Vξσ (ǫ)\ {ξ} we thus have:
|f (x) − f (ξ)| ψ (|x − ξ|)
0≤ ≤K . (181)
φ (|x − ξ|) φ (|x − ξ|)

Since φ ≺ ψ, upon taking the limit x → ξ σ , we obtain Dxφ f (ξ σ ) = 0.

Conversely, local fractional derivatives describe the local Hölder continuity of the func-
tion.

Proposition 118. Let f ∈ C 0 (I) be such that Dxψ f (ξ σ ) exists for ξ ∈ I, σ = ± and
ψ ∈ Q. Then f is right (σ = +) or left (σ = −) locally ψ-Hölder-continuous at ξ.

Proof. Since Dxψ f (ξ σ ) exists, we have from (159,160) for x ∈ I:

f (x) − f (ξ) = σ Dxψ f (ξ σ )ψ (|x − ξ|) + Rσ (x, ξ), (182)
Rσ (x,ξ) Rσ (x,ξ)
where ψ(|x−ξ|) → 0, as x → ξ σ . The latter means that there exists ǫ > 0, such that ψ(|x−ξ|)
is bounded in Vξσ (ǫ). In other words, there exist positive constants ǫ, C, such that:
|Rσ (x, ξ)| ≤ Cψ (|x − ξ|) , ∀x ∈ Vξσ (ǫ). (183)
We then have from (182,183) for x ∈ Vξσ (ǫ):
h i
|f (x) − f (ξ)| ≤ Dxψ f (ξ σ ) + C ψ (|x − ξ|) . (184)

This means that f is right (σ = +) or left (σ = −) locally ψ-Hölder continuous at ξ.

We may now establish the relation between critical order of differentiability and critical
Hölder order.
Theorem 119. A function f ∈ C 0 (I) has right (resp. left) critical order of differentiability
ψ ∈ Q at ξ ∈ I if and only if it has right (resp. left) critical Hölder order ψ at ξ.
Proof. Let f have right (σ = +) or left (σ = −) critical Hölder order ψ ∈ Q at ξ ∈ I.
Then f isright (resp. left) φ-Hölder continuous at ξ to all orders φ ∈ Q with φ ≺ ψ, and
we have Dxφ f (ξ σ ) = 0. We conclude that the critical order of differentiability ψξ,c σ is

such that ψ ψξ,c σ . Suppose that ψ ≺ ψ σ . Then D φ f (ξ σ ) exists for all φ ∈ Q with
ξ,c x
ψ ≺ φ ≺ ψξ,c σ . Notice that such φ always exist given that scales of functions are dense.
From Proposition 46., f is locally right (resp. left) φ-Hölder continuous at ξ. But this
contradicts the assumption that ψ is the right (resp. left) critical Hölder order. We conclude
σ .
that ψ ∼ ψξ,c
Conversely, suppose that f has right (σ = +) or left (σ = −) critical critical order of
σ at ξ. Then f is locally right (resp. left) φ-Hölder continuous at ξ for all
differentiability ψξ,c
φ ∈ Q such that φ ≺ ψξ,c σ . We conclude that the critical local right (resp. left) Hölder order
σ
ψ at ξ, satisfies ψξ,c ψ. Suppose that ψξ,c σ ≺ ψ. Then for all φ ∈ Q with ψ σ ≺ φ ≺ ψ,
ξ,c
f is locally right (resp. left) φ-Hölder continuous at ξ. This means that f is right (resp. left)
differentiable at ξ to all orders ρ ∈ Q such that ψξ,cσ ≺ ρ ≺ φ. But then ψ σ cannot be the
ξ,c
critical order of differentibility.
Proposition 120. Let f ∈ C 0 (I) be such that:

σ (f (ξ + σhn ) − f (ξ))
0 < lim

< +∞, σ=± (185)
n→∞ ψ(hn )
for some ψ ∈ D with ψ ≺ x and some sequence (hn )n∈N of positive numbers such that
hn ↓ 0. Moreover suppose that f is right (σ = +) or left (σ = −) ψ-Hölder continuous.
Then the right (σ = +) or left (σ = −) critical order of differentiability is ψ for all scales
F containing ψ. In particular, if ψ ∈ P this is true for all scales.
Proof. Let F be a scale containing ψ. Suppose that φ ∈ F is the critical order of dif-
ferentiability. Since f is ψ-Hölder continuous at ξ, from Lemma 48., the critical order of
differentiability must satisfy ψ φ. If ψ ≺ φ, then by Theorem 99., we should have
(Dxψ f )(ξ σ ) = 0, which contradicts (185).
Finally, suppose that ψ ∈ P. Since, by definition, every scale contains P (and hence ψ)
the result follows.
As an example, we consider the Weierstrass function.
Lemma 121. Let Wλ,s : [0, 1] → R denote the Weierstrass function for λ > 1 and 1 <
s < 2, and hn = λ−n . Then we have:
W (hn ) P+∞ k(s−2) sin λk =

limn→∞ λ,s 2−s
hn
= k=−∞ λ
(186)
= Wλ,s (1) + sin(1) + W 1 ,s (1)
λ
Moreover, s − 2 is the right critical order of differentiability at x = 0 with respect to all

scales.
Proof.
nP
Wλ,s (hn ) n−1
limn→∞ h2−s
= limn→∞ hs−2
n k=1 λk(s−2) sin(λk hn )+
n
P∞
+λn(s−2) sin(λn hn ) + k=n+1 λk(s−2) sin(λk hn ) =
nP o
n−1
λ(k−n)(s−2) sin(λk−n ) + sin(1) + λ(k−n)(s−2) sin(λk−n )
P∞
= limn→∞ k=1 k=n+1
(187)
In the first sum we perform the substitution l = n − k, whereas in the second sum we set
l = k − n. We thus obtain;:
W (hn ) 1 l(s−2)
= ∞ sin λ1l + sin(1) + ∞ l(s−2) sin(λl ) =

limn→∞ λ,s
P P
h2−s
n
l=1 λ l=1 λ
= W 1 ,s (1) + sin(1) + Wλ,s (1)

λ
(188)
which concludes the proof of (186).
Since Wλ,s is (2 − s)-Hölder continuous, we have from Proposition 120. that 2 − s is
the right critical order of differentiability of Wλ,s at 0 with respect to all scales.
5. Generalizations
There are various generalizations of the previous definitions worth investigating.
5.1. Quantum Difference Operators

A basic property of differentiable functions is that the quantities
σ (f (x + σǫ) − f (x))
∇ǫσ f (x) := (189)
ǫ
with ǫ > 0 and σ = ±1 converge and are equal when ǫ ↓ 0.

As a consequence the quantity

f (x + ǫ) + f (x − ǫ) − 2f (x)
aǫ f (x) = (190)
ǫ
converges to zero when ǫ ↓ 0.
Definition 122. Let h > 0 be a given real number. A function f ∈ C 0 is said to be ǫ − h-

differentiable at a point ξ if
aǫ f (ξ) < h. (191)
We can detect the non-differentiable character of a function by investigating its ǫ − h-

differentiability.
Definition 123. Let h > 0 be a given real number and f ∈ C 0 (I). The h-minimal resolu-
tion of f at a point x, denoted ǫ(f, h)(x) is defined as
ǫ(f, h)(x) := inf {aǫ f (x) < h} . (192)

ǫ
Similarly, the h-minimal resolution of f is
ǫ(f, h) := sup ǫ(f, h)(x). (193)

x∈I
Of course, if at some ξ ∈ I and for a given h, the h-minimal resolution is non-zero,

then f is non-differentiable at ξ.
In (193) we choose the sup, because if f is differentiable on a small set of points, the
h-minimal resolution is however non-zero.
Following Nottale [30] and Cresson [11], we then define
Definition 124. Let h > 0 and f ∈ C 0 (I). If ǫ(f, h) > 0, then for all 0 < ǫ < ǫ(f, h), we
define the right (σ = +) and left (σ = −) quantum difference operators of f at x ∈ I
as in (189). If ǫ(f, h) = 0, then ∇0+ f (x) = ∇0− f (x) = f ′ (x).
We then come to the following interesting result.
Proposition 125. Let Y : I → R be right (σ = +) or left (σ = −) locally ψ-Hölder

1
continuous in I with ψ ∈ Q such that xs ≺ ψ x n for some n ∈ N with n ≥ 2 and all
0 < s < n1 . Moreover, let f : R2 → R be of class C n+1 in R × I, where R denotes the
range of Y on I. Then we have for sufficiently small ǫ:
∂f
∇ǫσ f (Y (x), x) = ∂x (Y (x), x) +
(194)
Pn 1 ∂j f j
+σ j=1 j! ∂y j (Y (x), x) ǫ−1 (σǫ∇ǫσ Y (x)) + O (ψ(ǫ))
for σ = ± and where O (ψ(ǫ)) include all the terms of order ψ and higher.
Proof. We prove the result for σ = +. The proof for σ = − follows mutatis mutandis. We
have:
f (Y (x + ǫ), x + ǫ) = f Y (x) + ǫ∇ǫ+ Y (x), x + ǫ

(195)
Since Y is locally right ψ-Hölder continuous, we have for sufficiently small ǫ:
ǫ
ǫ∇+ Y (x) = |Y (x + ǫ) − Y (x)| = O (ψ(ǫ)) . (196)
Given that f is of class C n+1 , we conclude from (195,196) that we have the controlled
Taylor expansion up to order n:
f (Y (x + ǫ), x + ǫ) = f (Y (x), x) +
n+1 (197)
Pn 1 ǫ
i ∂k f
ǫj ǫ∇ǫ+ Y
P
+ k=1 k! i+j=k ǫ∇+ Y (x) ∂y i ∂xj
(Y (x), x) + O (x) .
As a consequence, it follows that

ǫ∇ǫ+ f (Y (x + ǫ), x + ǫ) =
n+1 (198)
Pn 1 ǫ
i ∂k f
ǫj ǫ∇ǫ+ Y
P
= k=1 k! i+j=k ǫ∇+ Y (x) ∂y i ∂xj
(Y (x), x) + O (x) .
By selecting terms of order less than or equal to one in ǫ on the right-hand side of the
previous equation and by taking into account (196), we obtain:
ǫ∇ǫ+ f (Y (x + ǫ), x + ǫ) =
h
∂f Pn 1 ∂if
i i
(Y (x), x) ǫ−1 ǫ∇ǫ+ Y (x) + O ǫ2 ∇ǫ+ Y (x)

=ǫ ∂x (Y (x), x) + i=1 i! ∂y i
(199)
Dividing by ǫ, the result follows.
Remark 126. It is interesting to remark that the case n = 2
∂f
∇ǫσ f (Y (x), x) = ∂x (Y (x), x) +
(200)
σǫ ∂ 2 f
+ ∂f ǫ
∂y (Y (x), x) (∇σ Y (x)) + 2! ∂y 2 (Y (x), x) (∇ǫσ Y (x))2 + O (ψ(ǫ))
is reminiscent of Itô’s formula.
We also note that Proposition 125. is the starting point for deriving the Schrödinger
equation in the context of scale relativity [11].
5.2. Local Fractional Directional Derivatives

Definition 127. Let X be a vector space, (Y, || · ||) a normed vector space, S a nonempty
subset of X, α ∈ ( 0, 1 ] and f : S → Y a given mapping. If for two elements x̄ ∈ S and
h ∈ X the limit
f (x̄ + λh) − f (x̄)
f (α) (x̄)(h) := lim (201)
λ↓0 λα
exists, then f (α) (x̄)(h) is called the directional derivative of order α of f at x̄ in the
direction h. If this limit exists for all h ∈ X, then f is called directionally differentiable
to order α at x̄.
Definition 128. A nonempty subset S of a vector space is said to be starshaped with re-
spect to some x̄ ∈ S, if for all x ∈ S and all λ ∈ [0, 1]:
λx + (1 − λ)x̄ ∈ S (202)
Definition 129. Let X be a vector space, S ⊂ X starshaped with respect to some x̄ ∈ S,

α ∈ ( 0, 1 ] and f : S → R a functional. The functional f is said to be α-convex at x̄ if
f (λx + (1 − λ)x̄) ≤ λα f (x) + (1 − λα )f (x̄) (203)
for all x ∈ S and all λ ∈ [0, 1].
Lemma 130. Let X be a vector space, S ⊂ X starshaped with respect to some x̄ ∈ S,

α ∈ ( 0, 1 ] and f : S → R a α-convex functional at x̄. Then for arbitrary x ∈ S, the
function ϕ : ( 0, 1 ] → R with
f (x̄ + λ(x − x̄)) − f (x̄)
ϕ(λ) = (204)
λα
is monotonically increasing.
Proof. Let 0 < s ≤ t ≤ 1. From the α-convexity at x̄, we have:

s t−s

f (x̄ + s(x − x̄)) − f (x̄) = f t (x̄ + t(x − x̄)) + t x̄ − f (x̄) ≤
s α s α

≤ t f (x̄ + t(x − x̄)) + 1 − t f (x̄) − f (x̄) = (205)
s α

= t (f (x̄ + t(x − x̄)) − f (x̄))
Consequently
ϕ(s) ≤ ϕ(t) (206)
and the result follows.
Theorem 131. Let X be a vector space, α ∈ ( 0, 1 ] and f : X → R a α-convex functional

at x̄ and at x̄ − h, for some h ∈ X. Then the derivative f (α) (x̄)(h) exists.
Proof. As before, we define

f (x̄ + λh) − f (x̄)
ϕ(λ) = , for all λ > 0. (207)
λα
Because of the α-convexity of f at x̄ − h, we have for all λ > 0

1 λ
f (x̄) = f 1+λ (x̄ + λh) + 1+λ (x̄ − h) ≤
α α (208)
1 1
≤ 1+λ f (x̄ + λh) + 1 − 1+λ f (x̄ − h)
Therefore, we have:
(1 + λ)α − 1

f (x̄ + λh) − f (x̄)
ϕ(λ) = ≥ (f (x̄) − f (x̄ − h)) (209)
λα λα
A simple investigation reveals that

(1 + λ)α − 1
0< < 1, for all λ > 0. (210)
λα
Consequently:
(1 + λ)α − 1

(f (x̄) − f (x̄ − h)) ≥ min {0, f (x̄) − f (x̄ − h)} (211)
λα
This means that the function ϕ is bounded from below. On the other hand, since f is
α-convex at x̄, we have from the previous Lemma that ϕ is monotonically increasing. Con-
sequently the limit
lim ϕ(λ) = f (α) (x̄)(h) (212)
λ↓0
exists.
Theorem 132. Let S be a nonempty subset of a vector space, α ∈ ( 0, 1 ] and f : S → R
a given functional.
(a) Let x̄ ∈ S be a minimal point of f on S. If the functional f has a directional derivative

of order α at X̄ in every direction x − x̄ with arbitrary x ∈ S, then
f (α) (x̄)(x − x̄) ≥ 0, for all x ∈ S. (213)
(b) Let the set S be starshaped with respect to x̄ ∈ S and f α-convex at x̄. If the functional
f has a directional derivative of order α at x̄ in every direction x − x̄ with arbitrary x ∈ S
and the inequality (213) is satisfied, then x̄ is a minimal point of f on S.
Proof. (a) Take any x ∈ S. Since f has a directional derivative of order α at x̄ in every
direction x − x̄, we have:
f (x̄ − λ(x − x̄)) − f (x̄)
f (α) (x̄)(x − x̄) = lim (214)
λ↓0 λα
Since x̄ is assumed to be a minimal point of f on S, we have for sufficiently small λ > 0
f (x̄ − λ(x − x̄)) ≥ f (x̄) (215)
Consequently, we obtain (213).
(b) Because of the α-convexity of f at x̄, we have for an arbitrary x ∈ S and all
λ ∈ ( 0, 1 ]
f (x̄ − λ(x − x̄)) = f (λx + (1 − λ)x̄) ≤ λα f (x) + (1 − λα )f (x̄) (216)
and hence
f (x̄ − λ(x − x̄)) − f (x̄)
f (x) ≥ f (x̄) + (217)
λα
Taking the limit λ ↓ 0, we obtain:
f (x) ≥ f (x̄) + f (α) (x̄)(x − x̄) (218)
With the inequality (213), we have for all x ∈ S:
f (x) ≥ f (x̄) (219)
Consequently x̄ is a minimal point of f on S.
5.3. Clarke Fractional Derivatives

Inspired by Clarke’s derivatives [8] we propose the following definition.
Definition 133. Let X be a vector space, S a nonempty subset of X, α ∈ ( 0, 1 ] and

f : S → R a given functional. If for x̄ ∈ S and h ∈ X the limit
(α) f (x + λh) − f (x)

fC (x̄)(h) := lim sup (220)
x→x̄,λ↓0 λα
(α)
exists, then fC (x̄)(h) is called the Clarke derivative of order α at x̄ in the direction h.
If this limit exists for all h ∈ X, then f is called Clarke differentiable to order α at x̄.
Example 134. Let f : R → R be given by
f (x) = |x|α (221)
for some α ∈ ( 0, 1 ]. Let us prove that

(α)
fC (0)(h) = |h|α (222)
for all h ∈ R. If h = 0, the result is trivial. Suppose that h 6= 0. With the aid of the triangle
and the Jensen inequalities, we have:
(α) |x + λh|α − |x|α |x|α + λα |h|α − |x|α

fC (0)(h) = lim sup α
≤ lim sup α
= |h|α (223)
x→0,λ↓0 λ x→0,λ↓0 λ
Now consider the sequences n12 n∈N and n1 n∈N as special cases. We have for sufficiently

large n:
12 + h α − 12 α = |h|α 1 + 1 α − 1 α =
1

n −α n n n nh nh
(224)
1 α 1 α α 1 α 1 2
|h|α |h|α

= 1+ nh − nh = 1+ nh − nh +O nh
As n → ∞ the previous expression converges to 1. Consequently, |h|α is both an accumu-

lation point of λ−α (|x + λh|α − |x|α ) for x → 0 and λ ↓ 0 and an upper bound of all
accumulation points. Hence, we obtain (222).
Theorem 135. Let S be a subset of a normed vector space (X, ||·||) with nonempty interior,
let x̄ ∈ int(S) be a given element, f : S → R a functional which is α-Hölder continuous
in a neighborhood of x̄ for α ∈ ( 0, 1 ], i.e. there exist K, ǫ > 0 such that
|f (x) − f (y)| ≤ K||x − y||α (225)
for all x, y ∈ B(x̄, ǫ). Then f is Clarke differentiable to order α at x̄ and

(α)
|fC (x̄)(h)| ≤ K||h||α for all h ∈ X. (226)
Proof. For arbitrary h ∈ X we have:

f (x + λh) − f (x)
≤ Kλ−α ||x + λh − x||α = K||h||α (227)
λα
for all x ∈ S and λ > 0 such that ||x− x̄|| < ǫ and ||x+λh− x̄|| < ǫ. Because of this bound
the limit superior appearing in the definition of the Clarke derivative exists. Furthermore,
we have:
(α)
|fC (x̄)(h)| = lim supx→x̄,λ↓0 f (x+λh)−f (x)

λα ≤
(228)
f (x+λh)−f (x)
≤ lim supx→x̄,λ↓0 λα ≤ K||h||α
which proves the result.
Definition 136. Let X be a vector space, α ∈ ( 0, 1 ] and p : X → R a functional. The

functional p is said to be α-sublinear, if:
(α − positive homogeneity) p(µx) = µα p(x), for all x ∈ X, µ ≥ 0

(229)
(subadditivity) p(x + y) ≤ p(x) + p(y), for all x, y ∈ X
Example 137. Let (X; || · ||) be a normed vector space, α ∈ ( 0, 1 ]. The functional
p(x) = ||x||α is α-sublinear. The functional q(x) = |l(x)|α , with l a linear functional
is α-sublinear.
x̄ ∈ int(S), α ∈ ( 0, 1 ] and f : S → R a functional which Clarke differentiable of order
(α)
α at x̄. Then the Clarke derivative fC (x̄) is an α-sublinear functional.
(α)
Proof. First of all notice that fC (x̄)(0X ) = 0. For arbitrary h ∈ X and µ > 0 we have:
(α) f (x+λµh)−f (x)
fC (x̄)(µh) = lim supx→x̄,λ↓0 λα =
(230)
= µα lim supx→x̄,λ↓0 f (x+(λµ)h)−f
(λµ)α
(x)
= µα f (α) (x̄)(h)
This proves α-positive homogeneity. Next we prove subadditivity. For arbitrary ha , h2 ∈

X, we have:
(α) f (x+λh1 +λh2 )−f (x)
fC (x̄)(h1 + h2 ) = lim supx→x̄,λ↓0 λα =
f (x+λh1 +λh2 )−f (x+λh2 )+f (x+λh2 )−f (x)

= lim supx→x̄,λ↓0 λα ≤
(231)
≤ lim supx→x̄,λ↓0 f (x+λh2 +λhλ1α)−f (x+λh2 ) + lim supx→x̄,λ↓0 f (x+λhλ2α)−f (x) =
(α) (α)
= fC (x̄)(h1 ) + fC (x̄)(h2 )
x̄ ∈ int(S), α ∈ ( 0, 1 ] and f : S → R a functional which is α-Hölder continuous and
α-convex in a neighborhood of x̄. If for a given h ∈ X, f is α-convex at x̄ − h, then the
Clarke and directional derivatives of order α at x̄ in the direction h coincide.
Proof. Since f is α-convex at x̄ and x̄−h, from Theorem 131., f (α) (x̄)(h) exists. Likewise,
(α)
as f is α-Hölder continuous in a neighborhood of x̄, by Theorem 135., fC (x̄)(h) also
exists. By the definition of these derivatives, we have automatically:
(α)
f (α) (x̄)(h) ≤ fC (x̄)(h). (232)
Let us now prove the converse inequality.

(α) f (x+λh)−f (x)
fC (x̄)(h) = lim supx→x̄,λ↓0 λα =
(233)
= limδ↓0,ǫ↓0 sup||x−x̄||<δ sup0<λ<ǫ f (x+λh)−f
λα
(x)
Since f is α-convex in a neighborhood of x̄, Lemma 130. leads to the equality:
(α) f (x + ǫh) − f (x)

fC (x̄)(h) = lim sup (234)
δ↓0,ǫ↓0 ||x−x̄||<δ ǫα
For an arbitrary µ > 0, we thus have:
(α) f (x + ǫh) − f (x)

fC (x̄)(h) = lim sup (235)
ǫ↓0 ||x−x̄||<µǫ ǫα
Since f is α-Hölder continuous in a neighborhood of x̄, we have for sufficiently small ǫ:

f (x+ǫh)−f (x) f (x̄+ǫh)−f (x̄)
ǫα − ǫα ≤ ǫ−α |f (x + ǫh) − f (x̄ + ǫh)| + ǫ−α |f (x) − f (x̄)| ≤
K K
≤ ǫα ||x − x̄||α + ǫα ||x − x̄||α ≤ 2Kµ
(236)
for some positive constant K. It follows that:
(α) f (x̄ + ǫh) − f (x̄)

fC (x̄)(h) ≤ lim + 2Kµ = f (α) (x̄)(h) + 2Kµ (237)
ǫ↓0 ǫα
Since µ > 0 was chosen arbitrarily, we conclude that

(α)
fC (x̄)(h) ≤ f (α) (x̄)(h). (238)
This completes the proof.

5.4. Comments
Some comments are in order.
1) The critical order of differentiability of Wλ,r is 2 − r with respect to all scales. It seems
however, that the local fractional derivative of order 2 − r exists nowhere. Can we prove
that? Nevertheless, since Wλ,r is (2 − r)-Hölder continuous its Clarke derivative exists (cf.
Definition 133. and Theorem 135.) everywhere.
2) If f : X → R is α-Hölder continuous in some neighborhood of x̄ ∈ X with α ∈ ( 0, 1 ],

then there exist K, ǫ > 0, such that:
f (x) − f (y)
−K ≤ ≤K (239)
||x − y||α
for all x, y ∈ B(x̄, ǫ) with x 6= y. It follows that:
f (x + λh) − f (x) f (x + λh) − f (x)

lim inf α
= ||h||α lim inf > −K||h||α (240)
x→x̄,λ↓0 λ x→x̄,λ↓0 (||h||λ)α
so that
f (x + λh) − f (x)
lim inf (241)
x→x̄,λ↓0 λα
also exists.
Acknowledgment
The authors would like to thank Rui Ferreira for drawing their attention to the
Banach-Mazurkiewicz Theorem 6. This work has been supported by the grant
PTDC/MAT/099880/2008 of the Portuguese Science Foundation (FCT).
References
[1] P.C. Allaart, K. Kawamura: The Takagi function: a survey. arxiv:math.CA/1110.1692.
[2] P.C. Allaart, K. Kawamura: The improper infinite derivatives of Takagi’s nowhere
differentiable function. J. Math. Anal. Appl. 372 (2010) 656-665.
[3] J.M. Anderson, L.D. Pitt: Probabilistic behaviour of functions in the Zygmund spaces
Λ∗ and λ∗ . Proc. London Math. Soc. 59 (1989) 558-592.
[4] A. Babakhani, V. Daftardar-Gejji: On calculus of local fractional derivatives. J. Math.

Anal. and Appl. 270 (2002) 66.
[5] P. Billingsley: Van der Waerden’s continuous nowhere differentiable function. Amer.
Math. Monthly 89 (1982) 691.
[6] F.S. Cater: Nowhere Analytic Functions. Amer. Math. Monthly 91 (1984) 618-624.
[7] Y. Chen, Y. Yan, K. Zhang: On the local fractional derivative. J. Math. Anal. and Appl.
362 (2010) 17-33.
[8] F.H. Clarke: Optimization and nonsmooth analysis. Wiley, New York (1983).
[9] F. Ben Adda, J. Cresson: About non-differentiable functions. J. Math. Anal. and Appl.
263 (2001) 721.
[10] F. Ben Adda, J. Cresson: Fractional differential equations and the Schrödinger equa-
tion. App. Math. Comp. 161 (2005) 323.
[11] J. Cresson: Scale calculus and the Schrödinger equation. J. Math. Phys. 44 (2003)
4907.
[12] K. Falconer: Fractal Geometry. Wiley (2003).
[13] H. Haase: Dimension functions. Math. Nachr. 141 (1989) 101.
[14] E. Hewitt, K. Stromberg: Real and Abstract Analysis. Springer-Verlag, New York
(1965).
[15] S. Jaffard, Y. Meyer: Wavelet Methods for Pointwise Regularity and Local Oscilla-
tions of Functions. Memoirs of the AMS (1996).
[16] A.A. Kilbas, H. M. Srivastava, J.J. Trujillo: Theory and Applications of Fractional
Differential Equations. North-Holland, Elsevier (2006).
[17] K. Kolwankar, A.D. Gangal: Fractional differentiability of nowhere differentiable

functions and dimensions. 6 (1996) 505.
[18] K. Kolwankar, A.D. Gangal: Hölder exponents of irregular signals and local fractional
derivatives. Pramana J. Phys. 48 (1997) 49.
[19] K. Kolwankar, A.D. Gangal: Local fractional derivatives and fractal functions of sev-
eral variables. Proc. Fractals in Eng. (1997).
[20] K. Kolwankar, J. Lévy Véhel: Measuring functions smoothness with local fractional
derivatives. Fract. Calc. Appl. Anal. 4 (2001) 49-68.
[21] K. Kolwankar, A.D. Gangal: Local fractional Fokker-Planck equation. Phys. Rev. Lett.
80 (1998) 214-217.
[22] K.B. Athreya, S.N. Lahiri: Measure Theory and Probability Theory. Springer (2006).
[23] M. Krüppel: On the extrema and the improper derivatives of Takagi’s continuous
nowhere differentiable function. Rostock. Math.+ Kolloq. 62 (2007) 41-59.
[24] J. Kurzweil, S. Schwabik: McShane equi-integrability and Vitali’s convergence theo-

rem. Math. Bohem. 129 (2004) 141-157.
[25] B.B. Mandelbrot: The Fractal Geometry of Nature. W.H. Freeman and Company, New
York (2006).
[26] A. le Méhauté: Fractal Geometries, Theory and Applications. CRC Press Inc. (2000).
[27] R. Metzler, J. Klafter: The random walk’s guide to anomalous diffusion: a fractional
dynamics approach. Phys. Rep. 339 (2000) 1.
[28] K.S. Miller, B. Ross: An Introduction to the Fractional Calculus and Fractional Dif-
ferential Equations. Wiley and Sons, New York (1993).
[29] G. Moore: Zermelo’s axiom of choide. Springer (1982).
[30] L. Nottale: Fractal Space-Time and Microphysics; Towars a Theory of Scale Relativ-
ity. World Scientific (1993).
[31] K.B. Oldham, J. Spanier: The Fractional Calculus; Theory and Applications of Differ-
entiation and Integration to Arbitrary Order. Dover (2002).
[32] A. Parvate, A.D. Gangal: Fractal differential equations and fractal time-dynamical
systems. Pramana J. Phys. 64 (2005) 389-409.
[33] I. Podlubny: Fractional Differential Equations. Academic Press (1999).
[34] L.F. Richardson: Atmospheric diffusion shown on a distance-neighbour graph. Proc.
Ry. Soc. London A 110 (1926) 709-737.
[35] A. Rocco, B.J. West: Fractional calculus and the evolution of fractal phenomena.
Physica A 265 (1999) 535.
[36] S.G. Samko, A.A. Kilbas, O.I. Marichev: Fractional Integrals and Derivatives: The-
ory and Applications. Gordon and Breach Science Publishers, Switzerland (1993).
[37] K.G. Spurrier: Continuous Nowhere Differentiable Functions. Dissertation, South
Carolina Honors College (2004).
[38] E.M. Stein, A. Zygmund: On the fractional differentiability of functions. Proc. Lon-
don Math. Soc. 14A (1965) 249-264.
[39] T. Takagi: A simple example of a continuous function without derivative. Phys.-Math.
Soc. Japan 1 (1903) 176-177. The Collected Papers of Teiji Takagi, S. Kuroda, Ed.,
Iwanami (1973) 5-6.
[40] C. Tricot: Curves and Fractal Dimension. Springer (1995).
[41] K. Weierstrass: On continuous functions of a real argument which possess a definite
derivative for no value of the argument. In: Königlisch Pressischen Akademie der Wis-
senschaften, Mathematische Werke von Karl Weierstrass, Berlin, Mayee and Mueller
(1895) vol.2, 71-74.
[42] B.J. West, M. Bologna, P. Grigolini: Physics of Fractal Operators. Springer (2003).
[43] S. Willard: General Topology. Dover (2004).
[44] G.M. Zaslavsky: Hamiltonian Chaos and Fractional Dynamics. Oxford (2005).

Chapter 2
F RACTIONAL VARIATIONAL E MBEDDING

AND L AGRANGIAN F ORMULATIONS
OF D ISSIPATIVE PARTIAL D IFFERENTIAL
E QUATIONS
Jacky Cresson∗
Laboratoire de Mathématiques et de leurs Applications de Pau,
Université de Pau et des Pays de l’Adour, Pau Cedex, France
PACS: 05.45-a, 52.35.Mw, 96.50.Fm

Keywords: Fractional calculus, fractional calculus of variations, variational formulation,
dissipative equations
1. Introduction
Many problems of physic cannot be formalized using the usual Lagrangian or Hamiltonian
formalisms as described for example in [1]. This is in particular the case for dissipative
systems. Many authors have tried to overcome this difficulty. The main reason behind
these generalizations is that the Lagrangian formalism provides efficient tools to study the
dynamics and properties of the underlying equations as well as an intrinsic structure (i.e.
not depending on coordinates systems) related to a first principle of physic by the least
action principle. As examples of previous attempts we can cite :
• Bateman ([4],[23])which constructs a mirror dynamics assuming that the description

of dissipative systems is physically incomplete.
• Tveter [41] formalism which preserves the form of the Hamiltonian equations.
• Riewe’s approach ([36],[37]) using fractional variational calculus.

∗
E-mail address: jacky.cresson@univ-pau.fr
66 Jacky Cresson
• Stochastic differential equations whose mean behaviour corresponds to a dissipative

systems.
This list is far from being exhaustive. All of these tentative are based on a particular as-
sumption. The formalism of Tveter is dedicated to the preservation of a Hamiltonian like
structure. This mathematical demand is not supported by any physic idea and the interpre-
tation of the new structure is not clear. The stochastic approach is not sufficiently general
due to mathematical difficulties in the effective computation of the mean dynamics. The
idea is to see dissipative phenomenon as the trace of an irreversible dynamics conducted by
a stochastic differential equation. We refer to ([31],§.2.2.3,p.59-60) for an example. Bate-
man’s work is also supported by a physic idea related to the irreversibility of the equation
but his construction appears as purely formal: the construction of the mirror dynamics is
not fixed by the theory. One must construct the equation for each case. Finally, Riewe’s
approach is supported by physic. The dissipative phenomenon induces forces which are of
fractional nature and which appear in the Lagrangian formalism via fractional derivatives.
His tentative fails for technical reasons which are discussed in details in Section 2.5 and can
not be used in its present form.
In this Chapter we follow Riewe’s idea and provide a fractional variational framework
for dissipative systems. The main advantage of this formulation is that dissipative terms are
then connected to the emergence of fractional terms in the Lagrangian formulation. As a
consequence, the interpretation of our formalism is clear.
Most of the material contains in this chapter can be found in many different articles
during the last years. As a consequence, it is difficult for the readers to have a global
view of the tools we have developed and the results we have obtained. This chapter can
be considered as an overview of our contributions in the direction of fractional calculus of
variations.
Our work in fractional calculus is part of a general program to study generalization of
ordinary or partial differential equations called embedding formalisms. This point of view
has evolved along the year and a first review can be found in [17]. Although this formalism
is continuously developed and modified we provide in Section 3 a general introduction to
this idea. We can formally resume this point of view as follows : try to find functorial or
categorical ways to generalize ODEs and PDEs and classify all the existing generalizations
using this formalism. Although this program is not completely fulfilled, we provide a first
tentative in this direction.
The main part of this chapter is then dedicated to the asymmetric fractional embedding
developed by Pierre Inizan and the author in [10] and its applications.
2. Variational Principles and Dissipative Systems

In this section we discuss the problem of constructing a variational principle for dissipative
systems. We review some past issues to solve this problem. Starting from the fact that irre-
versibility was the main obstruction, we propose to use a doubling phase space which takes
into account the evolution of the system toward past or future and different time derivatives
operators for each of these variables. We also discuss Riewe’s approach to dissipative sys-
tems using the fractional calculus and prove that it does not give a satisfying solution. This
Fractional Variational Embedding and Lagrangian Formulations ... 67
part can be avoided in a first lecture and must be considered as an heuristic support for the
mathematical framework of asymmetric fractional calculus of variation developed in the
next part.
2.1. Obstruction to Variational Principles: Helmholtz’s Conditions

The classical inverse problem of the calculus of variations is the following: Having a given
set of differential or partial differential equations is it possible to know if they derive from a
variational principle, i.e. as a critical point of an explicit Lagrangian functional. There exists
a huge literature on this subject. The Helmholtz’s conditions give necessary and sufficient
conditions for a given set of equations to be obtained as an Euler-Lagrange equation. We
refer to [34] for more details. Helmholtz’s conditions are algebro-analytic and are related
to the self adjointness of the differential operator attached to the equations.
2.1.1. Second Order Differential Operators

A second order differential equation on an interval [a, b] is defined by Oa,b (q) = 0 where
O is a second order differential operator, i.e.:
O : a < b 7−→ Oa,b : C 2 ([a, b], R) −→ C 0 ([a, b], R) (1)

q 7−→ Oa,b (q)
with:
Oa,b (q) : [a, b] −→ R (2)
t 7−→ Oa,b (q)(t) = O q(t), q̇(t), q̈(t), t

where q̇ (resp. q̈) is the first (resp. second) derivative of q and where:
O: R4 −→ R (3)
(x, v, w, t) 7−→ O(x, v, w, t).
Hence, a second order differential equation (independently of the interval [a, b]) is entirely
determined by the application O.
2.1.2. Lagrangian Systems and Euler-Lagrange Equations

A Lagrangian system derives from a variational principle. Precisely, let us consider two
reals a < b and the following Lagrangian functional:
La,b : C 2 ([a, b], R) −→ R (4)

Z b
q 7−→ L(q, q̇, t) dt,
a
where L is a Lagrangian, i.e. an application of the type:
L: R3 −→ R (5)
(x, v, t) 7−→ L(x, v, t).
Let C02 ([a, b], R) := {w ∈ C 2 ([a, b], R), w(a) = w(b) = 0} denote the set of varia-
tions. Then, q ∈ C 2 ([a, b], R) is said to be a critical point of La,b if for any variation w,
68 Jacky Cresson
DLa,b (q)(w) = 0. The calculus of variations allows to characterize the critical points of
La,b as the solutions on [a, b] of the following second order Euler-Lagrange equation:

∂L d ∂L
(q, q̇, t) − (q, q̇, t) = 0. (ELa,b )
∂x dt ∂v
A dynamical system governed by such an Euler-Lagrange equation is called a second or-
der Lagrangian system. We refer to [1, p.55-57] for more details concerning Lagrangian
systems.
2.1.3. Helmholtz’s Conditions

The classical Helmholtz’s result can be stated as follows:
Theorem 1 (Helmholtz’s condition). Let O be a second order differential operator. The
second order differential equation associated with O can be written as a second order
Euler-Lagrange equation if and only if all the Frechet derivatives of Oa,b are self-adjoint
for any real a < b.
We refer to [34, Theorem 5.92, p.364-365] for a detailed proof. As remarked by J-P.
Olver [34, p.365], the condition of self-adjointness of all the Frechet derivatives of Oa,b for
any real a < b are often referred as the Helmholtz’s condition.
Nevertheless, the Helmholtz’s condition can be more explicitly formulated:
Lemma 2. Let O be a second order differential operator. The operator O satisfies the
Helmholtz’s condition if and only if

2 d ∂O ∂O
∀a < b, ∀q ∈ C ([a, b], R), (q, q̇, q̈, t) = (q, q̇, q̈, t). (Hcont )
dt ∂w ∂v
Proof. Let a < b and q ∈ C 2 ([a, b], R). One can easily obtain that:
∂O ∂O ∂O
∀u ∈ C 2 ([a, b], R), DOa,b (q)(u) = (q, q̇, q̈, t)u + (q, q̇, q̈, t)u̇ + (q, q̇, q̈, t)ü.
∂x ∂v ∂w
(6)
Using integrations by part, it holds for any u ∈ C 2 ([a, b], R):
d2 ∂O

a,b ∗ ∂O d ∂O
DO (q) (u) = (q, q̇, q̈, t)u − (q, q̇, q̈, t)u + 2 (q, q̇, q̈, t)u . (7)
∂x dt ∂v dt ∂w
Finally, we obtain that DOa,b (q) = DOa,b (q)∗ if and only if:

d ∂O ∂O
(q, q̇, q̈, t) = (q, q̇, q̈, t). (8)
dt ∂w ∂v
This concludes the proof.
However, they do not give an idea of the physical origin of the obstruction to the ex-
istence of a variational principle. In the next Section, we discuss the case of dissipative
systems.
2.2. The Case of Dissipative Systems: Irreversibility as an Obstruction

to a Lagrangian Variational Structure
For dissipative systems a classical result of Bauer, [5], in 1931 states that a linear set of
differential equations with constant coefficients cannot derive from a variational principle.
In order to illustrate this idea, let us consider the classical linear friction problem given by:
q” = γ q̇ − U ′ (q), γ 6= 0. (9)
Using the notations of the previous section, the previous equation is associated to the appli-
cation
O(x, v, w, t) = U ′ (x) − γv + w. (10)
We have
∂O ∂O
= 1, = −γ, (11)
∂w ∂v
so that the Helmholtz’s condition is γ = 0. In other words, we have the following Lemma:
Lemma 3. Equation (9) admits a Lagrangian variational formulation if and only if γ = 0.
As a consequence, the presence of a dissipative (or damping depending on the sign of
γ) is an obstruction to the existence of a Lagrangian variational formulation. Another way
to interpret this result is to say that the underlying dynamics becomes irreversible in time.
2.3. Irreversibility
It must be note that the starting point of Riewe was the fact that most of classical processes
observed in Nature are nonconservative and then irreversible. Precisely, Riewe writes:
It is a strange paradox that the most advanced methods of classical mechan-
ics deal only with conservative systems, while almost all classical processes
observed in the physical world are nonconservative. Conservative systems are
time reversible by definition, while nonconservative systems exhibit the famil-
iar arrow of time due to irreversible dissipative effects such as friction.
As a consequence, most of the classical systems do not possess a Lagrangian varia-
tional structure due to the loss of reversibility. However, one can manage to overcome this
difficulty:
• Bateman [4], pointed out that this obstruction is only valid if one understand that the
variational principle does not produce additional equations. In particular, Bateman
constructs a complementary set of equations which enables him to find a variational
formulation. The main idea behind Bateman’s approach is that a dissipative system
must be seen as physically incomplete due to its underlying irreversibility.
• Contrary to dissipative systems, for a classical reversible system, the time evolution
in the two directions (past and future) of time is well described by a single differential
equation. In fact, we do not know what is the dynamical evolution if we reverse the
arrow of time for a dissipative system. This is for example the case of the diffusion
equation or the convection-diffusion equation. Actually, the diffusion phenomenon
is non reversible in time.
70 Jacky Cresson
2.4. Dealing with Irreversibility: Doubling Phase Space and Operators

Bauer’s theorem and Bateman construction give two distinct ways to overcome the obstruc-
tion to a variational formulation:
• To avoid Bauer’s obstruction to the existence of a variational principle, a natural idea
is to deal with a different kind of differential calculus. This is precisely what we
will do in the next part, following a previous attempt of Riewe ([36],[37]) where the
Riemann-Liouville and Caputo fractional calculus are used. However, as we will see,
the attempt of Riewe does not work and an appropriate modification of his formalism
has to be done.
• The lost of reversibility is interpret by Bateman as an incomplete description of a
system. We suggest to say that we have two different dynamical variables (x+ , x− )
which represent the evolution of the systems toward future or past. In experiments we
have only access to the time evolution x+ , i.e. that a physical process can be modelled
by a curve t 7→ x+ (t). The evolution of the same physical process if we reverse the
arrow of time is not known. But this doubling of phase space is only one part of the
dynamical modelization of a physical process. In experiment, we need an operator
acting on the phase variables which controls the dynamical behavior. In classical
mechanics, this is done using the ordinary differential calculus and in particular the
classical derivative with respect to time d/dt. In the new setting, one must look for
two time operators d+ and d− acting on x+ and x− which we call left and right time
operators in the sequel.
These simple remarks tell us that if we look for a variational formulation of a general
set of equations then the functional must at least be defined on the doubled phase space
X = (x+ , x− ) with a differential operator D given by DX = ( D+ α x , D α x ) where
+ − −
D+α and D α are right and left Riemann-Liouville or Caputo derivatives.
−
2.5. Riewe’s Approach to Dissipative Systems: Fractional Mechanics

In 1996-1997, Riewe ([36],[37]) defined a fractional Lagrangian framework to deal with
dissipative systems. Riewe’s theory follows from a simple observation: ”If the Lagrangian
n 2
d x
contains terms proportional to , then the Euler-Lagrange equation will have a
dtn
d2n x dx
term proportional to 2n . Hence a frictional force γ should derive from a Lagrangian
dt dt !2
d1/2 x
containing a term proportional to the fractional derivative ” where the notation
dt1/2
d1/2 d1/2 d1/2 d
1/2
represents formally an operator satisfying the composition rule 1/2
◦ 1/2
= .
dt dt dt dt
He then studied fractional Lagrangian functional using the left and right Riemann-Liouville
fractional derivatives which satisfy the composition rule formula. Let L a Lagrangian de-
fined by:
L : R × R × R × R −→ R,
(x, v+ , v− , w) 7−→ L(x, v+ , v− , w)
and x : [a, b] → R a trajectory. With this notation the functional studied by Riewe is given
by
Z b
1/2 1/2 dx
L(x) = L(x, D+ x, D− x, )dt.
a dt
He proved that critical points x of this functional correspond to the solutions of the gener-
alised fractional Euler-Lagrange equation

d ∂L 1/2 1/2 ∂L 1/2 1/2 ∂L 1/2 ∂L 1/2
(⋆ ) + D− (⋆ ) + D+ (⋆ ) = (⋆ ),
dt ∂w ∂v+ ∂v− ∂x
1/2 1/2
where ⋆1/2 := (x, D+ x, D− x, dx dt ).
Riewe derived such a generalised Euler-Lagrange equation for more general functionals
depending on left and right Riemann-Liouville derivatives D− α and D α with arbitrary
+
α > 0 (see [36], equation (45) p. 1894).
The main drawback of this formulation is that the dependence of L with respect to
1/2 1/2 1/2 1/2
D+ (resp. D− ) induces a derivation with respect to D− (resp. D+ ) in the equa-
1/2 1/2
tion. As a consequence, we will always obtain mixed terms of the form D− ◦ D+ x or
1/2 1/2
D+ ◦ D− x in the associated Euler-Lagrange equation. For example, if we consider the
Lagrangian
1 1 2
L(x, v+ , v− , w) = mw2 − U (x) + γv+ ,
2 2
we obtain as a generalised Euler-Lagrange equation
d2 x 1/2 1/2
m + γ D− ◦ D+ x + U ′ (x) = 0.
dt2
However, in general
1/2 dx 1/2
D− ◦ D+ x 6=
,
dt
so that this theory cannot be used in order to provide a variational principle for the linear
friction problem. This problem of the mixing between the left and right derivatives in the
fractional calculus of variations is well known (see for example Agrawal [2]). It is due to
the integration by parts formula which is given for f and g in C00 ([a, b]) by
Z b Z b
α α
f (t) D− g(t)dt = D+ f (t) g(t)dt.
a a
In ([36], p.1897) Riewe considered the limit a → b while keeping a < b. He then ap-
proximated D− α by D α . However, this approximation is not justified in general for a large
+
class of functions so that Riewe’s derivation of a variational principle for the linear friction
problem is not valid.
In [8], Cresson tried to overcome this problem by modifying the underlying set of vari-
ations in the fractional calculus. The set of variations is made of functions h satisfying
D−α h = D α h. A critical point of the fractional functional under this restriction is called a
+
weak critical point. In that case, we obtain that solutions of the linear friction problem cor-
responds to weak critical point of the functional but not an equivalence as in the usual case.
72 Jacky Cresson
The main problem is that the set of variations is too small to derive a Dubois-Raymond
result. We refer in particular to the work of Klimek, [27], for more details.
In the next section, we review the main result of [10] allowing us to obtain an equiv-
alence between solutions of the linear friction problem and critical points of a fractional
functional.
2.6. A Remark about Irreversibility, Causality and the Least Action Principle
The link between the least action principle and the causalilty principle has always been
ambiguous and it will be more present in the fractional calculus context. Poincaré [17]
sumed it up as follows :
L’énoncé du principe de moindre action a quelque chose de choquant pour

l’esprit. Pour se rendre d’un point à un autre, une molécule matérielle, sous-
traite à l’action donnée de toute force, mais assujettie à se mouvoir sur une
surface, prendra la ligne géodésique, c’est-à-dire le chemin le plus court. Cette
molécule semble connaı̂tre le point où on veut la mener, prévoir le temps
qu’elle mettra à l’atteindre en suivant tel ou tel chemin, et choisir ensuite le
chemin le plus convenable. L’énoncé nous la présente pour ainsi dire comme
un être animé et libre. Il est clair qu’il vaudrait mieux le remplacer par un
énoncé moins choquant, et où, comme diraient les philosophes, les causes fi-
nales ne sembleraient pas se substituer aux causes efficientes.
Henri Poincaré, La science et l’hypothèse, 1902.

In the fractional calculus, the derivatives are non-local, which makes the past and future
of the functions appear explicitly. Because of the simultaneous presence of left and right
derivatives, none of those equations respect the causality principle. This difficulty may have
been seen as a definite failure, and alternatives have notably been proposed in [7, 22] to get
around this problem. However, because we believe that the least action principle should
remain fundamental in any formalism, we choose to keep this approach - particularly the
one in [5], and we prove in this paper that causality may be respected. To this purpose,
we formulate the following remark: when one observe a phenomena which one wants to
describe using a differential equation, one only have access to the left derivatives of the
functions, i.e. to the differential operators dependent on the past values of the function.
Thus, this characteristic represents a trace of the time arrow, and the differential equation
becomes attached to the forward temporal direction. If we assume the existence of a similar
differential equation, but related to the backward temporal evolution, we obtain a formal
approach of irreversibility. Difficulties about causality inherent to the least action principle
can be solved using this formalism. More precisely, we show that using a new embedding,
termed asymmetric embedding, it is possible to obtain causal Euler-Lagrange equations. In
doing so, we observe that the information on the future lies in fact in the variations used
by this variational method. The virtual status of these could hence moderate the finalist
aspect of the least action principle. In addition, we prove that the asymmetric embedding is
coherent, i.e. that this procedure is globally compatible with the least action principle.
If a differential equation has a physical content, it should only involve left derivatives.
Indeed, the state of a system at time t should be fixed by its past states at times τ , τ < t. We
also note that if we study the reversibility of a system, equations describing the backward
evolution should only contain right derivatives [11]. This motivates the following definition.
Definition 4 (Causality). A fractional differential equation is said to be causal if it involves

fractional derivatives of a single type.
3. Embeddings Formalisms of ODEs and PDEs

3.1. Introduction
This part is an introduction to the formalism of embeddings for ordinary or partial differen-
tial equations developed in ([13],[14],[8],[6],[9],[10],[21],[19]) in different contexts. The
formalism of embedding takes its origin in the following, not exhaustive list of problems:
• Bath of a given biological or physical model : in many problems of Biology or

Physics, the model is first constructed by neglecting or forgetting the particular envi-
ronment of the experiment which is called a bath. For example, if one deals with the
Davidov’s model of energy transfer in proteins there exists a phonon bath. In general
the model without the bath is a given ordinary or partial differential equation. A nat-
ural question is then to look for the new equation of the dynamics when the bath is
taken into account.
• Turbulence : Fluid dynamics is modeled by partial differential equations. Solutions

of these equations must be sufficiently smooth. However, there exists turbulent be-
haviour which correspond to very irregular trajectories. If the underlying equation
has a physical meaning, then one must give a sense to this equation on irregular func-
tions. Why ? Because the partial differential equation is derived using first principles
of Physics. These principles are always valid despite the fact that the mathematical
formulation makes initially some regularity assumptions. This is the case for the tur-
bulent solutions of the Navier-Stokes equations. The previous remark has then leaded
many Physicist and Mathematicians to define ”generalized” solutions of PDEs. For
example, Jean Leray [29] has introduced what he has called a quasi-derivation and
the notion of weak solutions for the Navier-Stokes equations. This first work has a
long history and we refer to [3] for an overview’s of Jean Leray’s work going through
for example Laurent Schwartz’s distribution theory.
• Deformation quantization problems: The problem is to go from classical mechanics

to quantum mechanics trough a deformation involving the Planck constant. Roughly
speaking, we have a one parameter family of spaces and operators such that they re-
duce to usual spaces and operators in classical mechanics, i.e. Riemannian manifolds
and classical derivatives.
• Long term behaviour of Chaotic systems: Chaotic systems play an important role in
Physics. As an example, the dynamics of the Solar system, modeled by a n-body
problem, is known to be Chaotic over 100 Myr (see [28],[33]). The main characteris-
tic of these systems is that the long term behaviour can be considered as random, i.e.
74 Jacky Cresson
can be associated to a stochastic process. The characterization of such stochastic pro-

cesses is of course difficult. However, it leads naturally to the problem of describing
the dynamics satisfied by these stochastic processes knowing that at the beginning
we have a given differential equations.
• Discretisation of ODEs or PDEs: This is a classical problem in Numerical Analysis :
For a given ODE or PDE one want to construct a numerical scheme, based on discrete
data, which can be used to study the behaviour of solutions. The main problem is
to keep trace of some intrinsic properties of the equation at the discrete level, like
symmetries, maximum principle, variational structures, etc.
These problems although completely distinct have a common core : we need to define
what is the natural or canonical analogue of a given ordinary or partial differential equation
over a new vector space. This space can be formed for example of stochastic processes,
non-differentiable functions, discrete data, etc. Such a generalization of the initial equation
can be done using an embedding formalism.
An embedding formalism can be considered as the minimal amount of operators and
maps that one must defined in order to derive such a generalization. Indeed, an embedding
being fixed, we have defined three ways to extend a given equation : differential, integral
and variational. In a given embedding framework, each of these generalization are uniquely
determined.
The differential embedding use the differential operator point of view. As an exam-
ple of such a generalization, one can think to extension of PDEs in Schwartz’s distribution
sense. The integral embedding is based on the integral representation of the solutions of
a given equation. As an example of such generalization, one can consider Itô’s approach
to stochastic differential equations. The last generalization is based on the variational for-
mulation, when it exists, of the equation. An example of such a procedure is given by the
construction of variational integrators in numerical analysis.
3.2. Embedding Formalism

An embedding formalism is a collection of operators and mappings defined over a vectorial
set V . A given embedding formalism can be used to define various analogue of a differential
or partial differential equations, as well as other classical objects like functionals over V .
Many different generalizations of ODEs and PDEs can be included in this framework as we
will see in the next chapter. The interest to introduce such a general setting is to understand
on which structures and how many a generalization does depend. This is in general not
clear. Knowing where we have made a choice, we see what are the possibilities we have
in a given construction. As an embedding formalism is a finite number of data, we see that
all existing generalizations can be improved in a limited number of ways, the improvement
depending on the problem.
3.2.1. Definition
In this Section we define the notion of abstract embedding. We denote by C([a, b]; Rd )
the set of functions x : t ∈ [a, b] → Rd , d ∈ N∗ and by C i ([a, b], Rd ) the set of i-th
differentiable functions.
Definition 5 (Asbtract embedding). Let V be a given vector space. We assume that there
exists :
• A mapping ι : C([a, b], Rd ) → V .
• A linear operator D : V → V which takes place for a generalized derivative on V .
• A linear operator J : V → R which takes place for a generalized integral on V .
• An operator ⋆ : V × V → V such that
ι(f.g) = ι(f ) ⋆ ι(g). (12)
The operator J and D are related by the generalized fundamental theorem of differential
calculus
J ◦ D = D ◦ J = Id, (13)
where Id is the identity map Id : V → V , f 7→ f .
As we will see in the next Section, these elements are always sufficient in order to
provide an extension of an ordinary or partial differential equations over V . In order to
illustrate the previous definition, we give in the following Section an example of embed-
ding for which the operator D, J and the product ⋆ can be explicitly defined assuming the
existence of the map ι and two lift operators τ : V → C 1 and τD : V → C 0 satisfying
τD ◦ ι = Id. In that case, we derive all the properties of the operator D and J as for ex-
ample the generalized Leibniz formula or product formula. This example is important as
many important class of embedding formalisms can be put in this framework or in a close
analogue.
3.3. Example: Basic Embeddings

3.3.1. Construction of a Generalized Derivative
The operator d/dt acts naturally on the set of differentiable functions C 1 ([a, b], Rd ). Then,
a first idea is to define the operator generalized operator D on the subset ι(C 1 ) = V1 by
d
D◦ι=ι◦ , (14)
dt
meaning that we are looking to the following diagram :
d/dt d
f (f ) (15)
dt
ι ι
D
ι(f ) D(ι(f ))
d
The problem is that in general the quantity ι ◦ (f ) can not be written in term of ι(f ).
dt
Even if this is the case, then one must also extend the previously define operator to V .
76 Jacky Cresson
Another way to construct such an operator is to assume that there exists a lift operator
τ : V → C 1 ([a, b], Rd ). In this case, we can define D as follow :
Definition 6. A generalized derivative on V denoted by D is defined for all F ∈ V by

d
D(F) = ι τ (F) , (16)
dt
As a consequence, the following diagram commutes:
d
dt d
τ (F) (τ (F)) (17)
dt
τ ι
D
F D(F)
We have not in general that τ and ι are inverse one each others. So that D does not
reduce to d/dt over the image of ι(C 1 ).
3.3.2. Construction of a Generalized Integral

The mapping J must be constructed in such a way that we obtain an analogue of the fun-
damental theorem of differential calculus. As a consequence, we must take care of the
following diagram :
Z t
d
t
d d
Z
dt a
τ (F) (F) τD ( (F)) (18)
dt a dt
τ τD ι
D JD
F D(F) JD (D(F)).
It is easy to see that we need a new lifting operator τD is we want to satisfy the commu-
d
tativity. Indeed, as τ : V 1 → C 1 but : C 1 → C 0 , we must define for all F ∈ V 0 where
dt
V 0 = D(V 1 ) a natural element in C 0 such that
ι ◦ τD = ι ◦ τ = Id, (19)
and
d
τD ◦ D = ◦ τ. (20)
dt
As a consequence, we are lead to the following definition:
Definition 7. A generalized integral JD : V → V is defined for all F ∈ V by

Z t
JD (F) = ι τD (F) . (21)
a
Using the properties of ι, τ and τD , we verify that:

Lemma 8. For all F ∈ V 1 , JD (D(F)) = F + C, where C is a constant for D, i.e.
D(C) = 0.
Proof. We have by definition of JD and D that
Z t
JD (D(F)) = ι τD (∆(F)) ,
Za t (22)
d
= ι τD ι (τ (F)) .
a dt
As τD ◦ ι = Id over C 0 we deduce that
Z t
d
JD (D(F)) = ι (τ (F)) ,
a dt (23)
= ι (τ (F) + c) ,
where c ∈ R is a constant. As ι ◦ τ = Id over V , we deduce finally
JD (D(F)) = F + C, (24)
where C = ι(c) is a constant for D.
The generalized integral operator JD is then construct with respect to the generalized
derivative D and we then verify that we have the generalized fundamental result of differen-
tial calculus. However, as we are dealing with linear operators, one can define the integral
operator JD in a pure algebraic way as a solution of the equation JD ◦ D = Id. We must
also note that we can reverse the construction by defining first the generalized integral and
then the derivative.
3.3.3. Product Rule Formula

We keep the notation of the previous Section. We want to discuss what is preserve in the
new setting from the classical differential calculus. As an example, what can be said about
the product rule formula Z t Z t
df dg
·g =− f· ? (25)
a dt a dt
We define the ”product” on V compatible with the classical product of functions.
Definition 9. We denote by ⋆ the product on V defined by
V × V −→ V,
⋆: (26)
(F, G) 7−→ F ⋆ G = ι(τ (F) · τ (G)).
By definition of the map ι we verify that
·
(f, g) f ·g (27)
ι ι
⋆
(F, G) F ⋆ G.
78 Jacky Cresson
As a consequence, the formal data of the maps ι, τ and τD provide an example of

embedding.
Definition 10. Let ι : C(R, Rd ) → V , τ : V → C 1 and τD : V → C 0 such that τD ◦ι = Id

over C 0 and ι ◦ τ = ι ◦ τD = Id over V . We call basic embedding and
weZ denote by
t
d
E( ι, τ, τD )(V ) the embedding defined by D = ι (τ (F)) , JD (F) = ι τd (D)
dt a
and F ⋆ G = ι(τ (F) · τ (G)).
In the next Section, we study some properties of a basic embedding.
3.4. Properties of Basic Embeddings

In the following, we assume that we have fixed a basic embedding E( ι, τ, τD )(V ).
3.4.1. Leibniz Formula

Using the previous definition, we can prove that the operator D satisfied a Leibniz type
formula:
Lemma 11. Let F ∈ V 1 , G ∈ V 1 , then
D(F ⋆ G) = D(F) ⋆ G + F ⋆ D(G). (28)
Proof. By definition of D and ⋆ we have

d
D(F ⋆ G) = ι (τ (F) · τ (G)) . (29)
dt
By the classical Leibnitz formula, we obtain

d d
D(F ⋆ G) = ι (τ (F)) · τ (G) + τ (F) · (τ (G)) . (30)
dt dt
By linearity of D and the fact that ι(f · g) = ι(f ) · ι(g), we deduce

d d
D(F ⋆ G) = ι (τ (F)) · ι(τ (G)) + ι(τ (F)) · (τ (G)) . (31)
dt dt
As ι ◦ τ = Id, the definition of D and the definition of ⋆ this formula reduces to
D(F ⋆ G) = D(F) · G + F · D(G). (32)
In such an embedding the Leibniz relation is then preserved, meaning that the linear
operator D is a derivation on the set V .
3.4.2. Generalized Scalar Product

The classical scalar product on functions denoted by h·, ·i is defined by
Z
hf, gi = f · g. (33)
The natural analogue of the scalar product which is given by the previous embedding is:
Definition 12. We call generalized scalar product and we denote by (·, ·) : V × V → R
the quantity
(F, G) = JD (F ⋆ G). (34)
It must be noticed that the previous definition make sense in general. As a consequence,
one can look for the adjoint operator associated to D and denoted by D∗ , satisfying
(D(F), G) = (F, D∗ (G)). (35)
In the setting of the Eι,τ,τD (V ) embedding, we have:
Lemma 13. The adjoint operator associated to D is D∗ = −D.
Proof. By Lemma 11 we have for all (F, G) ∈ V 1 × V 1 that
D(F ⋆ G) = D(F) ⋆ G + F ⋆ D(G).
Using Lemma 8, we obtain
F ⋆ G = JD (D(F) ⋆ G) + JD (F ⋆ D(G)). (36)
3.4.3. Generalized Functionals

In this section, using the previous definitions, we give a natural analogue of an evaluation
functional and of a Lagrangian functional. We denote by F the set of functionals defined
over C(R, Rd ).
Evaluation Functionals: We introduce the following particular class of functional:

Definition 14. Let a : Rm → Rd be a function. The evaluation functional associated to a
is denoted by a capital letter A and defined by
C(Rm , Rd ) −→ F,
A: (37)
x 7→ A(x) := {t ∈ R, A(x)(t) = a(x(t))}.
A generalized evaluation functional over V is then defined by :
Definition 15. Let a be a given function and A its associated evaluation functional. Then
the embed evaluation functional denoted by A is defined for all F ∈ V by
A(F) = ι(A(τ (F))) = ι(a(τ (F)(t))). (38)
Most of the time we will write the same letter a for the evaluation functional A associ-
ated to a and its embedded version.
80 Jacky Cresson
Lagrangian Functionals: Let L : R × Rd × Rd → R be a function of class C 2 in all its

variables. The Lagrangian functional associated to L and defined over x ∈ C 1 (R, Rd ) is
given by
Z b
dx
L(x) = L t, x(t), dt. (39)
a dt
Using the previous construction we define the generalized Lagrangian functional as:
Definition 16. The generalized Lagrangian functional associated to L, and denoted by LE
is defined by for all X ∈ V 1 by
LE (X) = L(τ (X)). (40)
3.5. Differential, Integral and Variational Embedding of Differential

Equations
In this Section, we describe three different ways to generalize an ordinary differential equa-
tion. This reasoning will of course extend to partial differential equation.
An ordinary differential equation is usually given in differential form, i.e.
dx
= f (x, t), x ∈ Rn , t ∈ R. (41)
dt
However, one can also consider the integral form of the equation
Z t
x(t) = x(0) + f (x(s), s)ds. (42)
0
The differential form is related to dynamics via the time derivative. The integral form is
useful for proving existence and uniqueness of solutions or to study analytical properties of
solutions.
In order to give a meaning to a differential equation over a new set (stochastic processes,
non-differentiable functions, discrete sets) one can used the differential or the integral form.
In general, these two generalizations do not give the same object. In the differential case,
we need to extend first he time derivative. As an example of such a procedure, we can
look to Schwartz’s distributions [38] or backward/forward finite differences in the discrete
case. Using these new derivative one can then generalize differential operators and then
differential equations of arbitrary order. In the integral case, one need to give a meaning to
the integral over the new set. This strategy is for example used by K. Ito [30] in order to
define stochastic differential equations, defining first stochastic integrals.
In general, the integral form imposes less constraints on the underlying objects. This is
already the case in the classical case, where we need a differentiable function to write the
differential form but only continuity or weaker regularity to give a meaning to the integral
form.
Of particular importance for many applications in physics and mathematics is the case
Lagrangian systems governed by an Euler-Lagrange equation. The main property of these
equations is that there solutions corresponds to critical point of a Lagrangian functional,
i.e. is associated to a variational principle. Using a given embedding formalism we pro-
vide an analogous of the classical functional. By developing the corresponding calculus
of variations, we have obtained an embedded Euler-Lagrange equation. This extension of

the original Euler-Lagrange equation passing by the embedding of the functional is called
a variational embedding. Such constructions are well known in Numerical Analysis in
the study of discretisation of Lagrangian systems and give rise to variational integrators
([24],[32]).
We then have three ways to extend a given ordinary differential equation: differential,
integral or variational embedding. All these extensions are a priori different. The coherence
problem introduced in [14] in the context of the stochastic embedding consider the problem
of finding conditions under which these extensions coincide. We refer to [6],[21] and [14]
for such example.
3.6. Differential Embedding

An abstract embedding can be used to define an analogous of a given ordinary differential
equation on V ; Precisely, we consider a differential oeprator defined by
k
X di
P = ai (·). ,
dti
i=1
acting on C k (I, Rd ) functions.

Assume that the functional ai (.) keep sense over V . We denote by V k the subset of V
such that Di = D ◦ · · · ◦ D, i-th times, is well defined. Then the abstract embedding of P
is given by
Xk
Emb(P ) = ai (·) ⋆ Di , (43)
i=0
over V k .
If we look for the classical Euler-Lagrange equation associated to a Lagrangian function
and given by

d ∂L dx ∂L dx
(t, x, ) − (t, x, ) = 0, (44)
dt ∂v dt ∂x dt
where x : t ∈ R → x(t) ∈ R, the embedding is formally given by

∂L ∂L
D (T, X, DX)) − (T, X, DX) = 0, (45)
∂v ∂x
Of course, we have implicitely assumed that all the quantities here are well defined.
As an example, in the discretisation case, we obtain as a discrete embedded Euler-

Lagrange equation

∂L ∂L
∆+ (T, X, ∆+ X) − (T, X, ∆+ X) = 0, (46)
∂v ∂x
82 Jacky Cresson
3.7. Integral Embedding

Let us consider the following integral equation
Z t
x(t) = x(0) + f (s, x(s)) ds. (47)
a
Using an embedding formalism we define the following integral embedding of a differential

equation:
Definition 17. Let E be a fixed embedding formalism. The integral embedding of (47) is
defined by
X = ι(x(0)) + J(f (T, X)). (48)
The integral embedding is a priori different from the differential embedding although in
some cases the two procedures gives the same result. This is the case in the finite differences
embedding formalism.
The integral form of the Euler-Lagrange equations is given by
t
∂L ∂L
Z
(t, x(t), ẋ(t)) = (τ, x(τ ), ẋ(τ )) dτ + c, (49)
∂v a ∂x
for some constant c and all t ∈ [a, b]. Assume that a basic embedding Eι,τ,τD is fixed. The
integral embedding of the Euler-Lagrange equation is given by

∂L ∂L
(T, X, D(X)) = JD (T, X, D(X)) + C, (50)
∂v ∂x
where C = ι(c).
It is not a priori trivial that the integral and differential embedding gives the same result.
This problem is discussed in Section 3.9.
3.8. Variational Embedding

3.8.1. Lagrangian Systems
In this section, we recall classical definitions and theorems concerning Lagrangian systems.
We refer to [1] for a detailed study and for the proof of theorem 19.
Definition 18. A Lagrangian functional is an application defined by:
C 2 ([a, b], Rd ) → R,
Z b
L (51)
q 7→ L(q(t), q̇(t), t)dt.
a
where L is a Lagrangian i.e. a C 2 application defined by:
[a, b] × Rd × Rd → R,
L:
(t, x, v) 7→ L(t, x, v).
The functional (51) is also called action functional.

Let L be a Lagrangian functional, we denote by DL(q)(w) the Fréchet derivative of L
in q along the direction w in C 2 ([a, b], Rd ), i.e.
L(q + εw) − L(q)
DL(q)(w) = lim .
ε→0 ε
An extremal (or critical point) of a Lagrangian functional L is a trajectory q such that
DL(q)(w) = 0 for any variations w (i.e. w ∈ C 2 ([a, b], Rd ), w(a) = w(b) = 0). Extremals
of a Lagrangian functional can be characterized as solution of a differential equation of
order 2 given by:
Theorem 19 (Variational principle). Let L be a Lagrangian functional associated to the
Lagrangian L and let q ∈ C 2 ([a, b], Rd ). Then, q is an extremal of L if and only if q is
solution of the Euler-Lagrange equation given by:

∂L d ∂L
∀t ∈ [a, b], (q(t), q̇(t), t) − (q(t), q̇(t), t) = 0. (EL)
∂x dt ∂v
3.8.2. Embedded Lagrangian Functionals

An abstract embedding E being fixed, we can define the analogous of a Lagrangian func-
tional over V .
Definition 20. Let L be an admissible Lagrangian system and L the associated functional.
The embedded Lagrangian functional of L under E, denoted by LE is given by
LE (X) = J(L(T, X, DX)), (52)
for all X ⊂ V 1 .
As V is a linear space, we can define the Frechet derivative of LE along a given direction
H ∈ V and denoted by
1
DLE (X)(H) = lim (LE (X + ǫH) − LE (X)) . (53)
ǫ→0 ǫ
The corresponding notion of critical points is again given by:

Definition 21. A vector X ∈ V is a critical point of the embedded Lagrangian functional
LE if DLE (X)(H) = 0 for all H ∈ V .
The characterization of the critical points of an embedded functional is based on a
calculus of variations that one must construct.
3.8.3. Formal Embedded Calculus of Variations

Lemma 22. Let H ∈ V01 and X ∈ V 1 , then the Frechet derivative of LE in X in direction
H is given by

∂L ∂L
DLE (X)(H) = J (T, X, DX) ⋆ H + D∗ (T, X, DX) ⋆ H . (54)
∂x ∂v
84 Jacky Cresson
Proof. Formally, we have for all ǫ > 0 sufficiently small and H ∈ V01 , V01 = V 1 ∩ V0 ,
LE (X + ǫH) = J(T,X, D(X))

∂L ∂L
+ǫJ (T, X, D(X)) ⋆ H + (T, X, DX) ⋆ D(H) ,
∂x ∂v
= LE (X)

∂L ∂L
+ǫJ (T, X, D(X)) ⋆ H + D∗ (T, X, D(X)) ⋆ H + . . .
∂x ∂v
(55)
up to terms of order ǫ2 .
In order to obtain the embedded Euler-Lagrange equation we need an analogue of the

Dubois-Raymond lemma.
Definition 23. An embedding E is said to satisfy the Dubois-Raymond property if
J(A ⋆ H) = 0 ∀ H ∈ V0 then A = 0. (56)
As a consequence, we can say that formally the expected Euler-Lagrange equation as-
sociated to the embedded functional is given by
Theorem 24. Let E be an embedding satisfying the Dubois-Raymond property. Let L be

an admissible Lagrangian and LE the associated embedded Lagrangian functional. An
element X ∈ V is a critical point of LE if and only if

∂L ∂L
(T, X, D(X)) + D∗ (T, X, D(X)) = 0 (57)
∂x ∂v
As an example, in the basic embedding case, the Euler-Lagrange equation obtained is

∂L ∂L
(T, X, DX) − D (T, X, DX) = 0, (58)
∂x ∂v
which coincides with the differential embedding of the Euler-Lagrange equation.
3.9. Coherence between Differential, Integral and Variational Embeddings

The previous formalism gives three different ways a priori to generalize a differential
equation. A natural question is then the problem of coherence between these generalized
versions of the same equation:
Coherence problem. Let E be a given embedding formalism. Do we have equivalence

between the differential, integral and variational embedding of a given equation ?
A natural answer can be given for what concerns the differential and variational embed-
ding:
Lemma 25. The differential and variational embedding coincide if and only if D∗ = −D.
This follows from the previous computations.

This is of course the case in the classical differential case, as well as in the Schwartz’s
distribution case. However, this is not satisfied in general. This multiplicity of associated
natural generalized equations can be considered as a serious difficulty. However, we can
also think that for some problems, we have some choices which can be justified depending
on the problem and what we are interested in.
Remark 1. In fact, it is also possible to obtain coherence by modifying the underlying

calculus of variations. Precisely, making some assumptions on the set of variations. For
example, assume that a variation H ∈ V satisfies
DH = D∗ H, (59)
we recover also a coherent embedding. However, this assumption has a main drawback :
The set of variations satisfying this symmetry property can be reduced to zero or too small
in order to ensure the validity of a Dubois-Raymond lemma. As a consequence, we have
no equivalence between critical point of the Lagrangian functional and the Euler-Lagrange
equation.
3.10. Example: Finite Differences Embedding

This Section deals with a simple example of embedding over discrete sets using finite dif-
ferences. We recover several results of the classical literature in Numerical Analysis in a
more compact and conceptual form.
3.10.1. Notations
Let I ⊂ R be a closed interval. We denote by C(I, R) the set of functions with value in
R defined over I. Let N ∈ N and a R, b > a. We pose h = b − a/N and we denote by
T = {ti } ∈ RN +1 the vector defined by tk = a + k.h, h = (b − a)/N , k = 0, . . . , N .
We denote by L([a, b]) the set of continuous functions which are piecewise linear and
by S+ ([a, b]) (resp. S− ([a,
Sb])) the set of step functions which are defined over subintervals
Ii = [ti , ti+1 ] such that i Ii = [a, b], constant over [ti , ti+1 [ (resp. ]ti , ti+1 ]) and right
(resp. left) continuous.
An element of RN +1 is always denoted by a capital boldface letter. For F ∈ RN +1 , we
denote by small letter the components F = {fi }.
3.10.2. Discretisation of Functions

We keep the notations of the previous chapter about abstract embedding . A basic way to
make a discretisation of a function is to associate a finite number of its values over a finite
interval. In the following, we choose a uniform distribution over the interval [a, b] but our
arguments and construction can be immediately extend to more general setting like a given
time scales in time scales calculus.
Definition 26. Let T ⊂ [a, b] be a finite discrete set with cardinal N + 1. The element of
T are denoted by ti , i = 0, . . . , N . Let f be a function f : [a, b] → R and Df ⊂ [a, b] its
86 Jacky Cresson
domain of definition. The discrete version of the function f over Tf = Df ∩ T, denoted by

F : Tf → R is given by F(ti ) = f (ti ).
In general, we identify F with the element of RN +1 given by
ι(f ) = F = {fk }k=0,...,N . (60)
The discrete analogue of the functional space can then be chosen as F(T, R) or RN +1 .
3.10.3. Backward and Forward Discrete Derivatives

The discrete analogue of the classical derivative can be derived in many ways. In this
section, we make the simplest construction leading to the classical backward and forward
derivatives.
The lifting map τ : RN +1 → L([a, b]) can be obtain by linear interpolation and is
defined by
fk − fk−1
τ (F) = fk + (t − tk ), tk ≤ t ≤ tk+1 , k = 0, . . . , N − 1. (61)
h
The function τ (F) is not derivable but we can always define a left and right derivative
in the points tk . As a consequence, we can define ∆+ and ∆− the two discrete operators
dσ
corresponding to the left and right derivative denoted by with σ = ±, by a commutative
dt
diagram :
Definition 27. The forward (resp. backward) discrete derivative denoted by ∆+ (resp. ∆− )
are defined by
dσ /dtdσ
τ (F) (τ (F)) (62)
dt
τ ι
σ
F ∆σ (F)
for σ = ±.
From the point of view of embedding two strategies can be considered. The first one
is to consider the set V = F(T, R). As d+ (L([a, b]) is only defined over [a, b[, we have
Td+ (L([a,b])) = T+ where
T+ = {t0 , . . . , tN −1 }. (63)
As a consequence, the operator ∆+ is not defined on T as F but on T+ .
The same can be done for ∆− and we define T− = {t1 , . . . , tN }.
Definition 28. The forward (resp. backward) discrete derivative ∆+ (resp. ∆− ) defined
over T+ (resp. T− ) is given for all F ∈ F(T, R) by
∆+ = ι d+ (τ (F)) resp. ∆− = ι d− (τ (F)) .

(64)
Another point of view is to extend the two operators over T by choosing an ad-hoc
extension of τ (F) outside of [a, b]. For example, we can extend the map τ by continuity
to f0 when t ≤ a and by fN when t ≥ b. In that case, simple computations lead to the
following expression for these linear operators over RN +1 :
Lemma 29. The discrete derivatives ∆σ : RN +1 → RN +1 are given by for all F ∈ RN +1

by
fk+1 − fk
∆+ (F) = , 0 ≤ k ≤ N − 1, 0 , (65)
h
and
fk − fk−1
∆− (F) = 0, , 1≤k≤N . (66)
h
It must be note that in general the backward and forward derivatives are not defined as
operators acting on RN +1 but directly by components. As a consequence, most of Text-
books in numerical analysis loose the differential character of this map.
An important property of these operators is the following characterisation of the kernel:
Lemma 30. Let σ = ±. We have ∆σ F = 0 if and only if F = f0 1, where 1 = (1, . . . , 1).
In the following a vector of the form λ1 will be called a constant.
3.10.4. Discrete Integration Operator

The discrete integration operator can be obtained by two different methods. The first one
follows the previous construction of a discrete derivative. The second one is algebraic and
only based on the fact that we want to construct a linear mapping respecting an analogue of
the fundamental result of differential calculus.
The Analytic Method: Assume that we have fixed as a discrete analogue of the derivative
the forward (resp. backward) discrete derivative ∆+ (resp. ∆− ). We want to define a linear
operator J+ : RN +1 × RN +1 (resp. J− : RN +1 → RN +1 ) such that
J+ ◦ ∆+ = Id (resp. J− ◦ ∆− = Id). (67)
As a consequence, we are looking the following diagram

Z t
Z t
d+ a
κ(F) d+ (F) d+ (F) (68)
a
κ ? ι
∆+ J+
F ∆+ (F) J+ (∆+ (F)).
Where we must defined correctly the map denoted by ? in the previous diagram. Indeed,
the operator d+ satisfies
d+ (L([a, b]) ⊂ S+ ([a, b]). (69)
88 Jacky Cresson
Or the image of RN +1 by the lift operator τ is L([a, b]) and not S+ ([a, b]). As a conse-
quence, we must defined a new lift operator κ+ : RN +1 → S+ ([a, b]) in order that the
previous diagram becomes commutative and satisfying
ι ◦ τ+ = ι ◦ τ. (70)
A natural mapping from RN +1 to S+ ([a, b]) is given by
RN +1 −→ M ([a, b]),
κ+ : (71)
F = {fk } 7−→ fk , tk ≤ t < tk+1 , k = 0, . . . , N − 1.
We then obtain:
Definition 31. The discrete integral J+ : RN +1 → RN +1 is defined for all F ∈ RN +1 by

Z t
+
J+ (F) = ι κ (F) . (72)
a
A computation leads to the following explicit formula:
Lemma 32. Let F ∈ RN +1 . We have
J+ (F) = {0, f0 .h, (f0 + f1 ).h, . . . , (f0 + · · · + fN −1 ).h} . (73)
We deduce the discrete version of the fundamental result of differential calculus:
Lemma 33. For all F ∈ RN +1 we have
J+ (∆+ (F)) = F − f0 1, (74)
where
1 = (1, . . . , 1) . (75)
The Algebraic Method: As ∆+ and J+ are linear map, we can search directly J+ as:
J+ (F) = M+ .F, M+ M(N + 1, N + 1), (76)
where M(N + 1, N + 1) is the set of (N + 1) × (N + 1) matrices.

The forward derivative ∆+ can be written as
1
∆+ (F) = A+ .F, (77)
h
where the matrix A+ ∈ M(N + 1, N + 1) is given by
 
−1 1 0 ... 0
 . . . .. ... 
 0 . 0 
. ..
 
A+ = 
 ..
 (78)
 .

 0 ... −1 1 
0 ... 0
The problem is then to define the matrix M+ such that

M+ .A+ .F = F + λ.1, λ ∈ R. (79)
The matrix A+ is non invertible. The problem has then not a unique solution.
We first rewrite the matrix A+ as
A+ = I+ + N+ , (80)
where N+ is a almost null matrices given by
 
0 ... 0
 .. 
N+ =  . . .
 . ...  (81)

 0 ... 0 
0 ... 0 1
and I+ is an invertible matrix defined by
 
−1 1 0 ... 0
 .. .. .. 
 0 . . .  0
. ..
 
I+ =  ..
 (82)


 .
 0 ... −1 1 
0 ... −1
As a consequence, we have
1
J+ (∆+ (F)) = (M+ .I+ + M+ .N+ ) .F. (83)
h
−1
Choosing M+ = h.I+ , we obtain
−1
J+ (∆+ (F)) = F + I+ .N + .F. (84)
−1
It is then sufficient to prove that I+ .N + .F is a constant for ∆+ . A simple computation
leads to
−1
I+ .N + .F = −fN .1. (85)
−1
The matrix I+ is given by
 
−1 . . . −1
 .. .. 
−1
 0 . . 
I+ =
 .. ..
.
 (86)
 . . 
0 ... 0 −1
alg
As a consequence, the algebraic characterization of J+ leads to the following operator J+ :
alg
Definition 34. The operator J+ : RN +1 → RN +1 is given by
alg
J+ (F) = (−(f0 + · · · + fN ), −(f1 + · · · + fN ), . . . , −fN ) . (87)
We have the following relation
alg
J+ (∆+ (F)) = F − fN .1. (88)
90 Jacky Cresson
3.10.5. The Product Formula

We define the ”product” on RN +1 compatible with the classical product of functions. We
denote by ⋆ the product on RN +1 defined by
RN +1 × RN +1 −→ RN +1 ,
⋆: (89)
(F, G) 7−→ F ⋆ G = {fk .gk }k=0,...,N ,
where F = {fk }k=0,...,N and G = {gk }k=0,...,N .
By definition of the map ι we verify that
·
(f, g) f ·g (90)
ι ι
⋆
(F, G) F ⋆ G.
3.10.6. Discrete Scalar Product

The classical scalar product on functions denoted by h·, ·i is defined by
Z b
hf, gi = f · g. (91)
a
The natural analogue of the scalar product which is given by the previous embedding is:
Definition 35. We call discrete scalar product and we denote by (·, ·) : RN +1 ×RN +1 → R
the quantity
(F, G) = J+,N (F ⋆ G). (92)
This scalar product is in fact degenerate as the equation (F, F) = 0 induces only
F = (0, . . . , 0, fN ), fN ∈ R. (93)
3.10.7. Discrete Integration by Part

The generalized integration by part formula can also be seen as the fact that the operator
−∆− and ∆+ are adjoint with respect to this scalar product structure.
The left and right derivatives are linked by the following integration by part formula
Z t Z t
+
d f · g dt = − f · d− g dt + (f (t)g(t) − f (a)g(a)), (94)
a a
over the set of functions f and g admitting right and left derivatives respectively and such
that f (a)g(a) = f (b)g(b) = 0.
This integration by part formula has a direct analogue over RN +1 . Indeed, we have for
all (F, G) ∈ RN +1 × RN +1
Z
Z
·
(κ(∆+ (F)), κ(G)) κ(∆+ (F)) · κ(G) (κ(∆+ (F)) · κ(G) (95)
κ κ ι
⋆ J+
(∆+ (F), G) ∆+ (F) ⋆ G J+ (∆+ (F) ⋆ G).
It is easy to prove that

κ ◦ ι = id over S+ ([a, b]). (96)
As d+ (L([a, b])) ∈ S+ ([a, b]) we deduce that
κ ◦ ι ◦ d+ ◦ τ (F) = d+ ◦ τ (F), ∀ F ∈ RN +1 . (97)
We deduce that
Z t Z t
(κ(∆+ (F)) · κ(G) = (κ ◦ ι ◦ d+ ◦ τ (F)) · κ(G),
a Za t
= (d+ ◦ τ (F)) · κ(G),
(98)
Za t
= (−d− ◦ τ (F)) · κ(G) + τ (F) · κ(G)
a
− τ (F)(a) · κ(G)(a).
The only problem that we have is that d− (L([a, b]) ⊂ S− ([a, b]) so that we can not use the
definition of J+ in order to obtain the result. Indeed we have in general
κ(−∆− (F)) 6= −d− ◦ τ (F). (99)
However, up to a set of zero measure, we have
κ(−σ(∆− (F))) = −d− ◦ τ (F), (100)
where
RN +1 −→ RN +1
σ: (101)
F = {Fk } 7−→ σ(F) = H = {Hk = Fk+1 }.
As a consequence, we obtain the natural object associated to −d− ◦ τ (F)) · κ(G) is
Z t Z t
(κ(∆+ (F)) · κ(G) = (κ(−σ(∆− (F))) · κ(G) + F ⋆ G − f0 g0 1. (102)
a a
Finally, we obtain the following generalized integration by part formula:
Lemma 36. For all (F, G) ∈ RN +1 × RN +1 , we have
J+ (∆+ (F) ⋆ G) = J+ (−F ⋆ σ(∆− (G))) + F ⋆ G − f0 g0 1. (103)
3.10.8. Discrete Dubois-Raymond Lemma

The discrete version of the Dubois-Raymond lemma is valid. We first introduce the set
V01 ⊂ RN +1 defined by
V01 = H ∈ RN +1 , h0 = hN = 0 .

(104)
Lemma 37. Let F ∈ RN +1 such that J+,N (F ⋆ G) = 0 for all G ∈ V01 then fk = 0 for
k = 1, . . . , N − 1.
92 Jacky Cresson
Proof. We take H = {0, f1 , . . . , fN −1 } ∈ V01 . Hence, we obtain
N
X −1
J+,N (F ⋆ G) = fk2 h = 0. (105)
k=1
A stronger result can be derived when F is replaced by ∆+ (F) :
Lemma 38. Let F ∈ RN +1 such that J+,N (∆+ (F)⋆G) = 0 for all G ∈ V then ∆+ (F) =
0.
Proof. The proof follows from a simple computation. As

N −1
X fk+1 − fk
J+,N (∆+ (F) ⋆ H) = hk h, (106)
h
k=0
we obtain, taking H = ∆+ (F) ∈ RN +1 that fk+1 − fk = 0 for all k = 0, . . . , N − 1, so

that F = f0 .1. As a consequence, F is a constant for ∆+ which concludes the proof.
3.10.9. Discrete Differential Embedding
Let x ∈ Rd , we consider the ordinary differential equations
dx
= f (x, t). (107)
dt
Using the finite differences embedding the discrete version of this equation is
∆+ X = f (X, T), X ∈ (RN +1 )d , T ∈ RN +1 . (108)
As ∆+ X is defined on T+ , we obtain for each i = 0, . . . , N − 1

xi+1 − xi
= f (xi , ti ), (109)
h
b−a
where ti = a + i .
h
3.10.10. Discrete Integral Embedding

The integral formulation of the previous ordinary differential equation is given by
Z t
x(t) = x(0) + f (s, x(s)) ds. (110)
0
The forward finite difference integral embedding of this equation is then given by
X = x0 1 + J+ (f (T, X)). (111)

By definition, this equation is equivalent to

i−1
X
xi = x0 + h f (tj , xj ), i = 1, . . . , N. (112)
j=0
As a consequence, we obtain
xi+1 − xi = hf (ti , xi ), i = 0, . . . , N − 1, (113)
which is the classical one step forward Euler scheme in Numerical Analysis.
The forward finite differences differential embedding of the equation is equivalent to
the differential case. As a consequence, we see that in this simple case, we have coherence
between the two discrete version of the equation.
3.10.11. Discrete Variational Embedding

The discrete calculus of variations is defined over discrete Lagrangian functionals which
are obtained using the discrete embedding that we have fixed.
Definition 39. Let L be an admissible Lagrangian function and L the associated functional.
Let σ = ±. The discrete forward (resp. backward) Lagrangian functional associated to L
is defined by
Lσ (X) = J+,N (L(T, X, ∆σ (X))). (114)
A discrete variation is a vector H ∈ RN +1 such that h0 = hN = 0.
Definition 40. A discrete critical point X ∈ RN +1 is a vector satisfying DLσ (X)(H) = 0

for all H ∈ V01 ⊂ RN +1 .
As usual the main difficulty is to characterize the critical point of a discrete Lagrangian
functional by explicit conditions on L and the components of X. The previous results lead
easily to :
Theorem 41 (Discrete Euler-Lagrange equation). Let L be an admissible Lagrangian func-

tion. A vector X ∈ RN +1 is a discrete critical point of the discrete Lagrangian functional
associated to L if and only if it satisfies the system

∂L ∂L
−σ∆− (T, X, ∆+ (X)) + (T, X, ∆+ (X)) = 0. (115)
∂v ∂x k=1...,N −1
Proof. Using a Taylor expansion of L, we prove easily that

∂L ∂L
DL+ (X)(H) = J+,N (T, X, ∆+ (X)).∆+ (H) + (T, X, ∆+ (X)).H . (116)
∂v ∂x
As H ∈ V01 , we have by Lemma 36 that

∂L ∂L
J+,N ( (T, X, ∆+ (X)).∆+ (H)) = J+,N (−σ∆− (T, X, ∆+ (X)) .H), (117)
∂v ∂v
94 Jacky Cresson
and

∂L ∂L
DL+ (X)(H) = J+,N −σ∆− (T, X, ∆+ (X)) .H + (T, X, ∆+ (X)).H .
∂v ∂x
(118)
Using Lemma 37 we conclude the proof.
We refer to [24] for a different formulation of the same result. As we will see in the
following, the discrete embedding formalism can be used as a guiding line to generalized
classical theorem. This point of view is illustrated in the discrete version of Noether’s
theorem [7].
4. Asymmetric Fractional Embedding

4.1. Notations and Assumptions
4.1.1. Domain
Let d ∈ N. We consider a smooth d-dimensional bounded convex domain Ω with boundary
∂Ω. Let (e1 , . . . , ed ) be the canonical basis for Rd . For any x ∈ Rd , we denote by xi the
i-th component of x in the canonical basis of Rd . Let 1 ≤ i ≤ d and x ∈ Ω. We denote by
δi,x the straight line of Rd defined by
δi,x = x + Span(ei )
and Ωi,x = Ω ∩ δi,x . As Ω is bounded and convex, Ωi,x is a segment. Then, it exists
ai,x ≤ bi,x such that
Ωi,x := {(x1 , . . . , xi−1 , t, xi+1 , . . . , xd ) | t ∈ [ai,x , bi,x ]}. (119)
4.1.2. Functional Sets

For two sets A and B, F(A, B) denotes the vector space of functions f : A → B. Let
a, b ∈ R, a < b. Let m, n ∈ N∗ and p, q ∈ N. Let U be an open subset of Rm or the finite
interval [a, b]. The vector space of functions U → Rn of class C p is denoted by C p (U ). For
f ∈ F([a, b] × Ω, R), we denote by:
∀t ∈ [a, b] ft : Ω −→ R ∀x ∈ Ω fx : [a, b] −→ R .
and
x 7→ f (t, x) t 7→ f (t, x)
Let C p,q ([a, b] × Ω) and C p ([a, b] × Ω) the functional spaces defined as follow:
C p,q ([a, b]×Ω) := {f ∈ F([a, b]×Ω, R)| ∀t ∈ [a, b], ft ∈ C p (Ω), ∀x ∈ Ω, fx ∈ C q ([a, b])},
and C p ([a, b] × Ω) := C p,p ([a, b] × Ω) when q = p. Let C0p (Ω) := {f ∈ C p (Ω) | f =

0 on ∂Ω}.
For p = 0, we introduce the following vector spaces:
0
C+ ([a, b]) := {f ∈ C 0 ([a, b]) | f (a) = 0},
0
C− ([a, b]) := {f ∈ C 0 ([a, b]) | f (b) = 0},
and for p ≥ 1, let
p
C+ ([a, b]) := {f ∈ C p ([a, b]) | f (k) (a) = 0, 0 ≤ k ≤ p − 1},
p
C− ([a, b]) := {f ∈ C p ([a, b]) | f (k) (b) = 0, 0 ≤ k ≤ p − 1}.
p
C0p ([a, b]) := C+ p
([a, b]) ∩ C− ([a, b]).
The set of absolutely continuous functions over [a, b] is denoted by AC([a, b]) and
AC p+1 ([a, b]) is the set defined by
AC p+1 ([a, b]) := {f ∈ C p ([a, b]), f (p) ∈ AC([a, b])}.
Then C p ([a, b]) ⊂ AC p ([a, b]). A natural functional space for the study of classical PDEs
is
F p,q ([a, b] × Ω) := {f ∈ F([a, b] × Ω, R)| ∀ t ∈ [a, b], ft ∈ C p (Ω), ∀ x ∈ Ω, fx ∈ AC q ([a, b])}.
Let F p ([a, b] × Ω) := F p,p ([a, b] × Ω) when q = p. For any integer m ≥ 0, H m (Ω)

denotes the Sobolev space of order m.
4.1.3. Fields
In this paper we are interested in fields u depending on time t ∈ [a, b] and space x ∈ Ω:
u : [a, b] × Ω −→ R .
(t, x) 7−→ u(t, x)
The notation ∇u(t, x) ∈ Rd is the gradient of x 7→ u(t, x) and ∂t u(t, x) ∈ R the partial
derivative of u according to t and ∂xi f the derivative of f in the i-th space-variable. The
Xd
divergence of a vector field F = (F1 , . . . , Fd ) is divF = ∂xi Fi . Let v : Ω → R,
i=1
1 ≤ i ≤ d and x ∈ Ω. We denote by vi,x the function defined by
vi,x : Ωi,x −→ R
(120)
y 7−→ v(x1 , . . . , xi−1 , y, xi+1 , . . . , xd ).
If v : Ω → R, we denote also by v its extension to Rd such that v(x) = 0 if x ∈ Rd \Ω.

For x, y ∈ Rd , x × y denotes the vector in Rd defined by
x × y := (x1 y1 , . . . , xd yd )t . (121)
4.2. Fractional Operators: The One-Dimensional Case

For a general overview of the fractional calculus and more details we refer to the classical
book of Samko, Kilbas and Marichev, [39].
96 Jacky Cresson
4.2.1. Fractional Integrals

Let β > 0, and f , g : [a, b] → Rn .
Definition 42. The left and right Riemann-Liouville fractional integrals of f are respec-
tively defined by Z t
β 1
( I+ f )(t) = (t − τ )β−1 f (τ )dτ,
Γ(β) Za
b
β 1
( I− f )(t) = (τ − t)β−1 f (τ )dτ,
Γ(β) t
for t ∈ [a, b], where Γ is the Gamma function.
4.2.2. Fractional Derivatives

Let α > 0. Let p ∈ N such that p − 1 ≤ α < p.
Definition 43. Let t ∈ [a, b], the function f is left (resp. right) Riemann-Liouville differen-
tiable of order α at t if the quantity
p p

α d p−α α p d p−α
D+ f (t) = ◦ I+ f (t) and D− f (t) = (−1) p ◦ I− f (t),
dtp dt
exist respectively. The value D+ α f (t) (resp. D α f (t)) is called the left (resp. right)
−
Riemann-Liouville fractional derivative of order α of f at t.
Exchanging the order of composition we obtain the left and right Caputo fractional
derivatives:
Definition 44. Let t ∈ [a, b], the function f is left (resp. right) Caputo differentiable of
order α at t if the quantity
dp p

c α p−α c α p−α p d
D+ f (t) = I+ ◦ p f (t) and D− f (t) = I− ◦ (−1) p f (t),
dt dt
exist respectively. The value c D+ α f (t) (resp. c D α f (t))
− is called the left (resp. right)
fractional derivative of order α of f at t.
In a general setting the subscript c refers to Caputo derivative, whereas no subscript is
used to specify the Riemann-Liouville fractional derivative.
4.2.3. Properties
Let us mention the following relations linking both Riemann-Liouville and Caputo frac-
tional derivatives:
Lemma 45. Let 0 ≤ α < 1 and f ∈ AC 1 ([a, b]) then D+
α and D α exist almost every-
−
where and
α (t − a)−α
D+ f (t) = c D+
α
f (t) + f (a)
Γ(1 − α)
α (b − t)−α
D− f (t) = c D−
α
f (t) + f (b) .
Γ(1 − α)
The proof can be found in ([39], thm.2.2, p.39). The following lemma concerns the
semi-group property and the integration by parts formula for the fractional derivatives of
both Riemann-Liouville and Caputo types:
Lemma 46. 1. Composition rule formula: Let f ∈ AC 2 ([a, b]), then
c 1/2 1/2 d 1/2 1/2 d

D+ ◦ c D+ = D+ ◦ c D+ = .
dt dt
2. Integration by parts formula: Let 0 ≤ α < 1. Let f ∈ AC 1 ([a, b]) and g ∈ C01 ([a, b])
then
Z b Z b
c α
( D+ f )(t)g(t)dt = f (t)(c D−
α
g)(t)dt
a a
Z b Z b
α
( D+ f )(t)g(t)dt = f (t)(c D−
α
g)(t)dt.
a a
We refer to [39], p.46 for a proof.
Lemma 47 (Regularity). Let α > 0 and p ∈ N.
1. If f ∈ C 1 ([a, b]), then c D+

α f ∈ C 0 ([a, b]).
+
1 ([a, b]), then D α f = c D α f , and D α f ∈ C 0 ([a, b]).
2. If f ∈ C+ + + + +
p+1 α f ∈ C p ([a, b]).
3. If f ∈ C+ ([a, b]), then c D+ +
The previous lemma is stated under more general but implicit assumptions (as f ∈
I+α (L1 )) in [39]. However, as we need explicit conditions on the functional spaces in order
to develop the fractional calculus of variations, we give a less general result but with explicit
functional spaces.
df α df
∈ C 0 ([a, b]), we have I+ 0

Proof. 1. As ∈ C+ ([a, b]) using (Thm. 3.1 of [39],
dt dt
p.53 with λ = 0).
2. It results from 1. and lemma 45.

p+1 df k
3. Let 1 ≤ k ≤ p. As f ∈ C+ ([a, b]), then ∈ C+ ([a, b])
dt
dk c α dk 1−α df 1−α (k+1)
k
D + f = k
I+ = I+ f .
dt dt dt
1−α (k+1)
As f (k+1) ∈ C 0 ([a, b]), we have I+ f ∈ C+0 ([a, b]) using (Thm. 3.1 of [39],
k
p.53 with λ = 0). Then c D+ α f ∈ C k ([a, b]) and d c α
D f (a) = 0 for 1 ≤ k ≤ p.
dtp k +
Moreover c D+α f (a) = 0 by 1., so that c D α f ∈ C ([a, b]).
+ +
A similar result holds for the right derivative.

98 Jacky Cresson
β β
Lemma 48. 1. If f ∈ AC([a, b]), then I+ f ∈ AC([a, b]) and I− f ∈ AC([a, b]).
β 0 ([a, b]) and I β f ∈ C 0 ([a, b]).
2. If f ∈ C 0 ([a, b]), then I+ f ∈ C+ − −
Proof. 1. We refer to Lemma 2.1 in [39, p.32].
2. We refer to Theorem 3.1 in [39, p.53], with λ = 0.
Lemma 48 and Lemma 45 directly provide the following results.
Lemma 49. 1. If f ∈ C p ([a, b]), then c D+

α f ∈ C 0 ([a, b]).
+
p α f = c D α f . In particular, D α f ∈ C 0 ([a, b]).
2. If f ∈ C+ ([a, b]), then D+ + + +
Similar results hold for the right derivatives.

For x > 0 the integers x and x are defined by x − 1 ≤ x < x and x − 1 < x ≤ x. A
preliminary lemma will be useful for the following ones.
p dp β β dp
Lemma 50. Let β > 0 and p ∈ N∗ . If f ∈ C+ ([a, b]), then ◦ I + f = I + ◦ f.
dtp dtp
1 ([a, b]),
Proof. We prove it by induction. For p = 1 and f ∈ C+
t
d β (t − a)β−1 1
Z
I+ f (t) = f (a) + (t − τ )β−1 f ′ (τ ) dτ
dt Γ(β) Γ(β) a
β ′
= I+ f (t),
since f (a) = 0.
p+1 p
Now let p ∈ N∗ and f ∈ C+ ([a, b]). Since f ′ ∈ C+ ([a, b]), we may apply the
induction hypothesis:
dp β ′ β dp ′
◦ I + f = I + ◦ f. (122)
dtp dtp
β ′ d β
From case p = 1, I+ f = I f . Hence (122) becomes
dt +
dp+1 β β dp+1
◦ I + f = I + ◦ f,
dtp+1 dtp+1
4.3. Fractional Operators: The Multidimensional Case

We generalise the previous definitions and properties on Ω by introducing the multidimen-
sional fractional operators.
For a function v : Ω → R, we denote by vi,x the function defined on Ωi,x , as in (119)

acting in the i-th component of v. The i-th partial fractional derivatives are given by:
∂iα v(x) := D+
α
vi,x (xi ), c α
∂i v(x) := c D+
α
vi,x (xi ),
the right fractional Riemann-Liouville and Caputo partial derivatives with respect to the i-th
component of v. In a same way the left fractional Riemann-Liouville and Caputo partial
derivatives are given by
α α
∂ i v(x) := D− vi,x (xi ), c ∂ α v(x) := c D−
α
vi,x (xi ).
i
The associated Riemann-Liouville and Caputo fractional gradient of u, denoted by ∇α u

and c ∇α u, are defined by
∂1α u(t, x)
c
∂1α u(t, x)
  
∇α u(t, x) :=  .. c ∇α u(t, x) :=  ..
, .
  
.  .
∂dα u(t, x) c ∂ α u(t, x)
d
The Riemann-Liouville and Caputo fractional divergence of v : Ω → Rd are defined by
d
X d
X
α
div v(x) = ∂iα vi (x) c α
div v(x) = c α
∂i vi (x).
i=1 i=1
The analogous definitions of the left fractional gradient and divergence hold adding a bar
on the previous symbols.
Lemma 51. Let γ = (γ1 , . . . , γd ) ∈ Rd , u ∈ AC 2 (Ω) and x ∈ Ω, then we have

div1/2 γ × c ∇1/2 u(x) = γ · ∇u(x).
1/2 
γ1 c ∂1 u

Proof. Since γ × c ∇1/2 u =  ..

 , then
 
.
1/2
γd c ∂d u
Xd
1/2 1/2
div1/2 γ × c ∇1/2 u(x) = γi ∂i ◦ c ∂i u(x).
i=1
Applying lemma 46 concludes the proof.
4.3.1. Fractional Green-Riemann Formula

The following lemma is a consequence of the unidimensional integration by parts formula
of lemma 45 and leads to the fractional version of the Green-Riemann theorem mixing
Caputo and Riemann-Liouville derivatives given in lemma 53.
Lemma 52. Let u ∈ AC 1 (Ω) and v ∈ C01 (Ω) then

Z Z
α
c
u(x) ∂ i v(x)dx = ∂iα u(x)v(x)dx.
Ω Ω
100 Jacky Cresson
α
Proof. We first extend u, v, c ∂ i v and ∂iα u over Rd by associating the value 0 if x ∈ Rd \Ω.
In this case, we have
Z Z
c α α
u(x) ∂ i v(x)dx = u(x) c ∂ i v(x)dx
Ω d
Z R Z ∞
c α
= u(x) ∂ i v(x)dxi dx1 . . . dxi−1 dxi+1 . . . dxd .
Rd−1 xi =−∞
As Ω is convex, there exists ai,x and bi,x such that ai,x ≤ bi,x and
Z ∞ Z bi,x
α
u(x) c ∂ i v(x)dxi = ui,x (xi )c D−
α
vi,x (xi ).
xi =−∞ ai,x
As ui,x ∈ AC 1 ([ai,x , bi,x ]) and vi,x ∈ C01 ([ai,x , bi,x ]), the integration by parts formula from
lemma 46 leads to:
Z bi,x Z bi,x
c α c α
ui,x (xi ) D− vi,x (xi )dxi = D+ ui,x (xi )vi,x (xi )dxi
ai,x ai,x
Z bi,x
= ∂iα u(x)vi,x (xi )dxi
ai,x
Z +∞
= ∂iα u(x)v(x)dxi .
xi =−∞
As a consequence, we have
Z Z Z ∞
c α α
u(x) ∂ i v(x)dxi = ∂i u(x)v(x)dxi dx1 . . . dxi−1 dxi+1 . . . dxd ,
Rd Rd−1 xi =−∞
Z
= ∂iα u(x) v(x)dx.
Ω
Lemma 53 (Fractional Green-Riemann theorem). Let Ω ∈ Rd be a smooth d-dimensional

bounded convex domain, u ∈ C01 (Ω, R) and v = (v1 , . . . , vd ) ∈ AC 1 (Ω)d , then we have
Z Z
α
v(x) · c ∇ u(x) dx = divα (v(x)) u(x) dx. (123)
Ω Ω
Proof. The proof results from lemma 52. On the canonical basis the scalar product is given
by
Z d Z
c α X α
v(x) · ∇ u(x) dx = vi (x) c ∂ i u(x) dx.
Ω i=1 Ω
Applying lemma 52, we have for any 1 ≤ i ≤ d,

Z Z
α
c
vi (x) ∂ i u(x) dx = ∂iα vi (x) u(x) dx.
Ω Ω
The definition of divα concludes the proof.

4.4. Differential Asymmetric Fractional Embedding

We adapt here the presentation done in [8].
Let M, N ∈ N∗ . If f ∈ F(RM +1 , RN ) and y ∈ F([a, b], RM ), we denote by f (y(•), •)
the function defined by
f (y(•), •) : [a, b] −→ RN
t 7−→ f (y(t), t).
Let p, k ∈ N. If f = {fi }0≤i≤p and g = {gj }1≤j≤p are two families of

F(Rn(k+1)+1 , Rm ) (g = ∅ if p = 0), with fj ∈ C 0 (Rn(k+1)+1 ) and gj ∈ C j (Rn(k+1)+1 )
for 1 ≤ j ≤ p, we introduce the operator Ofg defined by
Ofg : C k+p ([a, b]) −→ " F ([a, b], Rm )

p
#
di dk

X (124)
x 7−→ f0 + fi ⋆ i g i x(•), . . . , k x(•), • ,
dt dt
i=1
where, for two operators A = (A1 , . . . , Am ) and B = (B1 , . . . , Bm ), A ⋆ B is defined by
(A ⋆ B)(y) = (A1 (y)B1 (y), . . . , Am (y)Bm (y)) .
The fractional embedding presented in [8] consists in replacing d/dt by a fractional

derivative. Here we want to keep this idea, but additionnaly we want to split in two the
functional space of the trajectories, in order to make the asymmetry between left and right
fractional derivatives explicitly appear.
Let 0 < α < 1. For X = (x+ , x− ) ∈ C 1 ([a, b])2 (AC 2 ([a, b])2 would be sufficient),
we introduce the fractional derivative c Dα , defined by
c α
D X = c D+ α
x + , − c D−
α

x− .
The classical case is recovered for α → 1− (and not for α = 1).
Lemma 54. Let X ∈ C 1 ([a, b])2 . Then
c d
∀ t ∈ (a, b), lim Dα X(t) = X(t).
α→1− dt
Proof. Let X = (x+ , x− ) ∈ C 1 ([a, b])2 . Since x′+ ∈ C 0 ([a, b]), all points of (a, b) are
Lebesgue points of x′ . We may then apply Theorem 2.7 of [39, p.51]:
1−α ′
∀ t ∈ (a, b), lim I+ x (t) = x′ (t).
α→1−
We proceed likewise for x− .
Hence for k ∈ N∗ and a “suitable” function X,

( c D α ) k X = ( c D+
α k
) x+ , (− c D−
α k
) x− .
The following lemma provides an example of such “suitable” functions.

102 Jacky Cresson
Lemma 55. Let k ∈ N∗ . If f ∈ C+

k ([a, b]), then we have
c α
k
D+ f = c D+
αk
f.
A similar result holds for the right derivative.
Proof. We prove it by induction on k. For k = 1, the result is obvious. Now, let k ∈ N∗

k+1 k ([a, b]), we use the induction hypothesis: ( c D α )k f =
and f ∈ C+ ([a, b]). Since f ∈ C+ +
c D αk f =αk−αk (αk)
+ I+ f . We have αk ≤ k, so f (αk) ∈ AC([a, b]), and from Lemma 48,
( c D+
α )k f
∈ AC([a, b]). Moreover, from Lemma 49, ( c D+α )k f (a) = 0. We may then
apply Lemma 45:

α
D+ ◦ ( c D+
α k
) f = c D+
α
◦ ( c D+
α k
) f = ( c D+
α k+1
) f.
α ◦ ( c D α )k f = d 1−α αk−αk (αk)

On the other hand, D+ + ◦ I+ ◦ I+ f . We have f (αk) ∈
dt
C 0 ([a, b]), so we may use formula 2.21 of [39, p.34]:
1−α αk−αk (αk) β (αk)

I+ ◦ I+ f = I+ f ,
where β = 1 + αk − α(k + 1).

Since f (αk) ∈ C+
1 ([a, b]), from Lemma 50,
d β (αk) β (αk+1)
◦ I+ f = I+ f .
dt
We have αk + 1 ∈ {α(k + 1), α(k + 1) + 1}, so we consider two cases.
• If αk + 1 = α(k + 1), then
α(k+1)−α(k+1)
α
D+ ◦ ( c D+
α k
) f = I+ f (α(k+1)) ,
α(k+1)
= c D+ f.
• If αk = α(k + 1), then
α 1+α(k+1)−α(k+1)
D+ ◦ ( c D+
α k
) f = I+ f (α(k+1)+1) ,
α(k+1)−α(k+1) 1 (α(k+1)+1)
= I+ ◦ I+ f .
1 f (α(k+1)+1) (t) = f (α(k+1)) (t) − f (α(k+1)) (a). But

We have I+ α(k + 1) ≤ k, so
f (α(k+1)) (a) = 0.
Consequently,
α(k+1)−α(k+1)
α
D+ ◦ ( c D+
α k
) f = I+ f (α(k+1)) ,
α(k+1)
= c D+ f.
α(k+1)
In both cases, we have proved that ( c D+
α )k+1 f = c D
+ f , which concludes the
proof.
k ([a, b]) × C k ([a, b]), ( c D α )k X verifies
Consequently, if X = (x+ , x− ) ∈ C+ −
( c D α ) k X = ( c D+
αk
x+ , (−1)k c D−
αk
x− ),
and ( c Dα )k X ∈ C+
0 ([a, b]) × C 0 ([a, b]).
−
Let us now precise the splitting we are interested in.
Definition 56. Let k ∈ N and m, n ∈ N∗ . Let f ∈ F(Rn(k+1)+1 , Rm ). The asymmetric
representation of f , denoted by f˜, is defined by
f˜ : R2n(k+1)+1 −→ Rm
(x0 , y0 , . . . , xk , yk , t) 7−→ f (x0 + y0 , . . . , xk + yk , t).
Actually, the relevant functions will be in F([a, b], Rn ) × {0} or {0} × F([a, b], Rn ).
That is why we introduce the following “selection” matrix.
Let Mm,2m (R) be the set of real matrices with m rows and 2m columns. We note Im
the identity matrix of dimension m, and we introduce the operator σ defined by
σ : F([a, b], Rm )2 −→ Mm,2m (R)
X 7−→ (Im 0) if X ∈ F([a, b], Rn ) × {0} and X 6= 0,
(0 Im ) if X ∈ {0} × F([a, b], Rn ) and X 6= 0,
(0 0) otherwise.
Now we can define the asymmetric fractional embedding of a differential operator.
Definition 57. With the previous notations, the asymmetric fractional embedding of opera-
tor (124), denoted by Eα (Ofg ), is defined on a subset E α ⊂ F([a, b], Rn )2 , by
Eα (Ofg ) : E α −→ F ([a, b], Rm )

p
" #
f˜i ⋆ Dαi g̃i
f˜0 + σ(X)
X
+ c α k
X 7−→ X(•), . . . ,( D ) X(•), • .
f˜i ⋆(−1)i D− αi g̃
i
i=1
(125)
The definition set E α of Eα (Ofg ) depends on f and g. We also introduce the following
spaces:
α
E+ = E α ∩ (F([a, b], Rn ) × {0}) , α
E− = E α ∩ ({0} × F([a, b], Rn )) .
α , (125) becomes
In particular, for (x+ , 0) ∈ E+
p
" #

g
X
Eα (Of )(x+ , 0)(t) = f0 + fi ⋆ D+ gi x+ (t), . . . , ( c D+
αi α k
) x+ (t), t ,
i=1
and for (0, x− ) ∈ E−α , we have
p
" #

Eα (Ofg )(0, x− )(t) = f0 +
X
fi ⋆ (−1)i D−
αi
gi x− (t), . . . , (− c D−
α k
) x− (t), t .
i=1
104 Jacky Cresson
Remark 2. There exists of course several ways to define a fractional embedding because
of the different definitions of fractional derivatives. As it will be seen later, with our choice,
the action of a fractional Lagrangian is well defined, we may obtain a coherent and causal
embedding. This property is not shared by the others fractional embedding formalisms.
For the sake of clarity, we will often denote by x the integer which verifies x − 1 ≤ x <
x, where x ∈ R+ . We also denote by x the integer which verifies x − 1 < x ≤ x.
Precisions on E+ α and E α can be given thanks to the following lemma.
−
β+p β p
Lemma 58. Let β > 0 and p ∈ N. If f ∈ C+ ([a, b]), then c D+ f ∈ C+ ([a, b]). A
similar result holds for the right derivative.
β β p
Proof. If β ∈ N∗ , c D+ f (t) = f (β) (t) − f (β) (a), and c D+ f ∈ C+ ([a, b]). Else, let
(β) k
1 ≤ k ≤ p. Since f ∈ C+ ([a, b]), from Lemma 50,
dk c β dk β−β (β)
D f = I f
dtk +
dtk +
β−β
= I+ f (β+k)
β−β
Given that f (β+k) ∈ C 0 ([a, b]), I+ f (β+k) ∈ C+
0 ([a, b]), from Lemma 48. Hence
k
c D β f ∈ C k ([a, b]) and d c D β f (a) = 0. Moreover, c D β f (a) = 0 from Lemma 49.
+
dtk + +
c β p
Finally, D+ f ∈ C+ ([a, b]).
∂gi p+k p+k
Corollary 59. • If = 0 for all 1 ≤ i ≤ p, C+ ([a, b]) × C− ([a, b]) ⊂ E α , and
∂t
p+k p+k
for all X ∈ C+ ([a, b]) × C− ([a, b]), Eα (Ofg )(X) ∈ C 0 ([a, b]).
• If p = 0 (g = ∅) and k = 1, C 1 ([a, b])2 ⊂ E α and for all X ∈ C 1 ([a, b])2 ,

Eα (Of∅ )(X) ∈ C 0 ([a, b]).
p+k p+k
Proof. • Let X = (x+ , x− ) ∈ C+ ([a, b]) × C− ([a, b]). For all 1 ≤ j ≤ k,
c α j c αj j c αj c α j p p
( D ) X = ( D+ x+ , (−1) D− x− ) and ( D ) X ∈ C+ ([a, b]) × C− ([a, b]),
from Lemmas 55 and 58.
If x+ 6= 0 and x− 6= 0, Eα (Og )(X) = f˜0 X(•), . . . ,( c Dα )k X(•), • ∈ C 0 ([a, b]).

f
If x− = 0, let 1 ≤ i ≤ p. Since gi is of class C i , we have gi (x+ ) :
t 7→ gi (x+ (t), . . . , c D+
αk x (t), t) ∈ C i ([a, b]).
+ Moreover, gi (x+ )′ (a) =
k
X (j) ∂gi (x+ ) ∂gi (j)
∂j gi (x+ )(a) · x+ (a) + (a). Since = 0 and x+ (a) = 0 for all
∂t ∂t
j=1
1 ≤ j ≤ k, we obtain gi (x+ )′ (a) = 0. By induction, gi (x+ )(l) (a) = 0 for all 1 ≤
i ([a, b]), and from Lemma 49, D αi g (x ) ∈ C 0 ([a, b]).
l ≤ i. Hence gi (x+ ) ∈ C+ + i +
We proceed likewise if x+ = 0.
• Let X ∈ C 1 ([a, b])2 . We have f˜0 (X(•), c Dα X(•), •) ∈ C 0 ([a, b]) from Lemma
49, so Eα (Of∅ )(X) = f˜0 (X(•), c Dα X(•), •) is well defined and is a function of
C 0 ([a, b]).
In order to clarify those notations, we give here a short example.
Example. We set n = m = p = 1, k = 2, and we suppose that 0 < α < 1/2.
Let f0 , f1 , g1 : R3 × R −→ R be three functions defined by

f0 (a, b, c, t) = c + e−t cos b,
f1 (a, b, c, t) = 1,
g1 (a, b, c, t) = cos a.
The associated operator Ofg verifies
d2

d d
Ofg (x)(t) = 2 x(t) + e−t cos x(t) + cos(x(t)),
dt dt dt
for x ∈ C 2 ([a, b]) and t ∈ [a, b].
Moreover, for any (x+ , x− ) ∈ AC 2 ([a, b])2 , ( c Dα )2 (x+ , x− ) = ( c D+
2α x , c D 2α x )
+ − −
as it will be shown in Lemma 78. The asymmetric fractional embedding Eα (Ofg ) is hence
given by
Eα (Ofg )(x+ , x− )(t) = c D+
2α
x+ (t) + c D−
2α
x− (t) + e−t cos( c D+
α
x+ (t) − c D−
α
x− (t))
α
D+ cos(x+ (t) + x− (t))
+σ(x+ , x− ) α cos(x (t) + x (t)) .
− D− + −
For (x+ , 0) ∈ AC 2 ([a, b]) × {0}, the fractional embedding becomes

Eα (Ofg )(x+ , 0)(t) = c D+
2α
x+ (t) + e−t cos( c D+
α α
x+ (t)) + D+ cos(x+ (t)),
and for (0, x− ) ∈ {0} × AC 2 ([a, b]), we have
Eα (Ofg )(0, x− )(t) = c D−
2α
x− (t) + e−t cos(− c D−
α α
x− (t)) − D− cos(x− (t)).
The ordinary differential equations may be written by using operators Ofg . Following
[8], we consider the differential equations of the form
Ofg (x) = 0, x ∈ C p+k ([a, b]). (126)
Definition 60. With the previous notations, the asymmetric fractional embedding of differ-
ential equation (126) is defined by
Eα (Ofg )(X) = 0, X ∈ Eα. (127)
Consequently, if (x+ , 0) ∈ E+α , (127) becomes
p
" #
X
αi
f0 + fi ⋆ D + gi x+ (t), . . . , ( c D+
α k
) x+ (t), t = 0,
i=1
and for (0, x− ) ∈ α,
E− we obtain
p
" #
X
i αi
f0 + fi ⋆ (−1) D− gi x− (t), . . . , ( c D−
α k
) x− (t), t = 0.
i=1
We verify that for these two cases, the asymmetric fractional embedding respects
causality, in the sense of Definition 4.
106 Jacky Cresson
4.5. Variational Asymmetric Fractional Embedding-ODEs Case

4.5.1. Asymmetric Fractional Embedding of Lagrangian Systems
Let 0 < α < 1. Let L be a Lagrangian.
For X = (x1 , x2 ) ∈ R2n , Y = (y1 , y2 ) ∈ R2n , and t ∈ [a, b], the asymmetric repre-
sentation of L, denoted by L̃, verifies
L̃(X, Y, t) = L(x1 + x2 , y1 + y2 , t).
Given that
∂L ∂L
(x1 + x2 , y1 + y2 , t) = (x1 + x2 , y1 + y2 , t) = ∂1 L(x1 + x2 , y1 + y2 , t),
∂x1 ∂x2
we deduce ∂1 L̃(X, Y, t) = ∂1 L(x1 + x2 , y1 + y2 , t). Similarly, we note ∂2 L̃(X, Y, t) =

∂2 L(x1 + x2 , y1 + y2 , t).
Theorem 61. The asymmetric fractional embedding of (EL) is defined by

α ∂ L̃(X(t), c D α X(t), t)

c α D+ 2
Eα (EL) ∂1 L̃(X(t), D X(t), t) − σ(X) α ∂ L̃(X(t), c D α X(t), t) = 0.
− D− 2
(128)
α , (128) becomes
In particular, for (x+ , 0) ∈ E+
Eα (EL)+ ∂1 L(x+ (t), c D+

α α
x+ (t), t) − D+ ∂2 L(x+ (t), c D+
α
x+ (t), t) = 0, (129)
α,
and for (0, x− ) ∈ E−
Eα (EL)− ∂1 L(x− (t), − c D−

α α
x− (t), t) + D− ∂2 L(x− (t), − c D−
α
x− (t), t) = 0.
(130)
Proof. Equation (EL) may be written like (126) with k = 1, p = 1, f = {∂1 L, 1} and
g = {−∂2 L}. We conclude by using Definitions 57 and 60.
On the other hand, the asymmetric fractional embedding of the Lagrangian L, which
will be noted Lα , verifies
Lα (X)(t) = L(x+ (t) + x− (t), c D+

α
x+ (t) − c D−
α
x− (t), t),
for all (x+ , x− ) ∈ C 1 ([a, b])2 and t ∈ [a, b].

The associated action (51) now becomes
A(Lα ) : C 1 ([a, b])2 −→ R

Z b
X 7−→ L̃ (X(t), c Dα X(t), t) dt.
a
Remark 3. We see here the necessity to choose the Caputo derivative inside the functions.
If we had taken the Riemann-Liouville derivative, the action could be undefined even for
1 2
regular functions. For example, if L(x, v, t) = v − U (x) and x+ ∈ C 1 ([a, b]), with
2
x+ (a) 6= 0, we would have
2
α 1 x+ (a)
L(x+ (t), D+ x+ (t), t) ∼ (t − a)−2α ,
a 2 Γ(1 − α)
and A(Lα )(x+ , 0) would not be defined for α ≥ 1/2.
The obtention of the differential of the action first requires a formula for integration by
parts with fractional derivatives.
β
Lemma 62. Let β > 0. If f ∈ AC β ([a, b]) and g ∈ C0 ([a, b]), then we have the following
formula for fractional integration by parts:
Z b Z b
c β β
f (t) · D− g(t) dt = D+ f (t) · g(t) dt.
a a
Similarly, we have:
Z b Z b
c β β
f (t) · D+ g(t) dt = D− f (t) · g(t) dt.
a a
Proof. If β ∈ N∗ , this is the classical formula for integration by parts. Else, since g ∈
β
C0 ([a, b]), g β ∈ Lp ([a, b]), with p ≥ 1/β. Furthermore, f (β) ∈ L1 ([a, b]), so equation
2.20 of [39, p.34] is valid:
Z b Z b
c β β β−β
f (t) · D− g(t) dt = (−1) I+ f (t) · g (β) (t) dt.
a a
Moreover, for all 0 ≤ k ≤ β − 1, g (k) (a) = g (k) (b) = 0. Therefore, iterating the
classical integration by parts β times leads to:
b b
dβ β−β b
Z Z Z
β β−β (β) β
(−1) I+ f (t) · g (t) dt = I+ f (t) · g(t) dt = D+ f (t) · g(t) dt.
a a dtβ a
We proceed likewise for the other relation.
Lemma 63. Let X ∈ C 1 ([a, b])2 . We suppose that ∂2 L̃(X(•), c Dα X(•), •) ∈ AC([a, b]).
Then A(Lα ) is C01 ([a, b])2 -differentiable at X and for all H = (h+ , h− ) ∈ C01 ([a, b])2 ,
Z bh i
dA(Lα )(X, H) = ∂1 L̃ (X(t), c Dα X(t), t) + D−
α
∂2 L̃ (X(t), c Dα X(t), t) · h+ (t) dt
a
Z bh i
+ ∂1 L̃ (X(t), c Dα X(t), t) − D+α
∂2 L̃ (X(t), c Dα X(t), t) · h− (t) dt.
a
108 Jacky Cresson
Proof. Let H = (h+ , h− ) ∈ C01 ([a, b])2 and ε > 0. For all t ∈ [a, b], we have:
L̃(X(t) + εH(t), c Dα X(t) + ε c Dα H(t), t) = L̃(X(t), c Dα X(t), t)

+ ∂1 L̃(X(t), c Dα X(t), t) · (h+ (t) + h− (t))
+ ∂2 L̃(X(t), c Dα X(t), t) · ( c D+
α
h+ (t) − c D−
α
h− (t)) + o(ε).
Integrating this relation on [a, b] leads to:

Z b
A(Lα )(X + εH) = A(Lα )(X) + ε ∂1 L̃(X(t), c Dα X(t), t) · (h+ (t) + h− (t)) dt
a
Z b
+ε ∂2 L̃(X(t), c Dα X(t), t) · ( c D+
α
h+ (t) − c D−
α
h− (t)) dt + o(ε).
a
Since ∂2 L̃(X(•), DX(•), •) ∈ AC([a, b]), h+ ∈ C01 ([a, b]) and h− ∈ C01 ([a, b]), we
can use lemma 62:
Z b Z b
c α c α α
∂2 L̃(X(t), D X(t), t) · D+ h+ (t) dt = D− ∂2 L̃(X(t), c Dα X(t), t) · h+ (t) dt,
a a
and
Z b Z b
∂2 L̃(X(t), c Dα X(t), t) · c D−
α
h− (t) dt = α
D+ ∂2 L̃(X(t), c Dα X(t), t) · h− (t) dt.
a a
Z bh i
Finally, (h+ , h− ) 7→ ∂1 L̃ (X(t), c Dα X(t), t) + D− α
∂2 L (X(t), c Dα X(t), t) ·
Z bh a
i
h+ (t) dt + ∂1 L̃ (X(t), c Dα X(t), t) − D+ α
∂2 L (X(t), c Dα X(t), t) · h− (t) dt is lin-
a
ear, which concludes the proof.
Then we obtain a result similar to Theorem 19.
Theorem 64. Let X ∈ C 1 ([a, b])2 . We suppose that ∂2 L̃(X(•), c Dα X(•), •) ∈

AC([a, b]).
Then we have the following equivalence:
X is a C01 ([a, b])2 -extremal of the action A(Lα ) if and only if it verifies
∂1 L̃(X(t), c Dα X(t), t) + D−
α
∂2 L̃(X(t), c Dα X(t), t)

= 0,
(ELα ) ∀ t ∈ (a, b),
∂1 L̃(X(t), c Dα X(t), t) − D+
α
∂2 L̃(X(t), c Dα X(t), t) = 0.
(131)
α ∂ L̃(X(t), c D α X(t), t) and

Proof. Similar to Theorem 19. The only difference is that D− 2
α c α
D+ ∂2 L̃(X(t), D X(t), t) may not be continuous respectively in b and a.
Equation (131) is very restrictive since X must verify

α α
( D+ + D− )∂2 L̃(X(t), c Dα X(t), t) = 0.
This condition may not be related to the dynamics of the system and seems too strong.
For instance, for α ∈ (1, 2), functions which fulfill ( D+ α + D α )f = 0 are given in [27]
−
and are very specific. By restricting the set of variations, equations more relevant will now
be obtained.
Euler-Lagrange equations which have been obtained so far in [36, 2, 8, 22] involve both
left and right fractional derivatives. The following result provides a similar equation.
Corollary 65. Let x+ ∈ C 1 ([a, b]). We suppose that ∂2 L(x+ (•), c D+ α x (•), •) ∈
+
AC([a, b]). Then we have the following equivalence:
(x+ , 0) is a C01 ([a, b]) × {0}-extremal of the action A(Lα ) if and only if x+ verifies
∂1 L(x+ (t), c D+
α α
x+ (t), t) + D− ∂2 L(x+ (t), c D+
α
x+ (t), t) = 0, (132)
for all t ∈ [a, b).
Such an equation is not causal because of the simultaneous presence of c D+
α and D α .
−
Moreover, regarding (129), this procedure is not coherent. Those problems are solved with
the following results.
Corollary 66. Let x+ ∈ C 1 ([a, b]). We suppose that ∂2 L(x+ (•), c D+ α x (•), •) ∈
+
(x+ , 0) is a {0} × C01 ([a, b])-extremal of the action A(Lα ) if and only if x+ verifies
(ELα )+ ∀ t ∈ (a, b], ∂1 L(x+ (t), c D+
α α
x+ (t), t)− D+ ∂2 L(x+ (t), c D+
α
x+ (t), t) = 0.
(133)
Corollary 67. Let x− ∈ C 1 ([a, b]). We suppose that ∂2 L(x− (•), c D− α x (•), •) ∈
−
(0, x− ) is a C01 ([a, b]) × {0}-extremal of the action A(Lα ) if and only if x− verifies
(ELα )− ∀ t ∈ [a, b), ∂1 L(x− (t), c D−
α α
x− (t), t)− D− ∂2 L(x− (t), c D−
α
x− (t), t) = 0.
(134)
Equations (133) and (134) are causal. Moreover, they are respectively similar to (129)
and (130): (ELα )± ≡ Eα (EL)± . With such sets of variations, the asymmetric fractional
embedding is therefore coherent.
4.5.2. Derivatives of Higher Orders

Let α ∈ (0, 1) and k ≥ 2. We consider here generalized Lagrangian systems which involve
derivatives up to order k.
Definition 68. An extended Lagrangian is a function
L : Rn(k+1) × [a, b] −→ R
(y0 , y1 , . . . , yk , t) 7−→ L(y0 , y1 , . . . , yk , t),
which verifies the following properties:
110 Jacky Cresson
• L ∈ C 1 (Rn(k+1) × [a, b]),
• ∀ 1 ≤ i ≤ k, ∂i+1 L ∈ C i (Rn(k+1) × [a, b]).

The action is now defined by
A(L) : C k ([a, b]) −→ R

b
dk

d
Z
x 7−→ L x(t), x(t), . . . , k x(t), t dt.
a dt dt
We then obtain:
Lemma 69. Let L be an extended Lagrangian and x ∈ C k ([a, b]).
d dk
We suppose that for all 1 ≤ i ≤ k, ∂i+1 L x(•), dt x(•), . . . , dt k x(•), • ∈
AC i ([a, b]).
Then A(L) is C0k ([a, b])-differentiable at x and for all h ∈ C0k ([a, b]),
Z b" k
#
i
dk

i d
X
dA(L)(x, h) = ∂1 L + (−1) i ∂i+1 L x(t), . . . , k x(t), t dt.
a dt dt
i=1
Lemma 70. Let L be an extended Lagrangian and x ∈ C k ([a, b]).

d dk
We suppose that for all 1 ≤ i ≤ k, ∂i+1 L x(•), dt x(•), . . . , dt k x(•), • ∈
AC i ([a, b]).
x is a C0k ([a, b])-extremal for A(L) if and only if x verifies the Euler-Lagrange equation
k
" #
i
dk

X
i d
(ELk ) ∀t ∈ [a, b], ∂1 L + (−1) i ∂i+1 L x(t), . . . , k x(t), t = 0.
dt dt
i=1
(135)
Concerning the asymmetric fractional embedding, we start with the embedding of the
Euler-Lagrange equation.
The asymmetric fractional embedding of (135) is given by:
k
" #
X (−1)i D+αi ∂
i+1 L̃ c α k

Eα (ELk ) ∂1 L̃ + σ(X) αi ∂ X(t), . . . , ( D ) X(t), t = 0.
D− i+1 L̃
i=1
In particular, for (x+ , 0), we have

k
" #
X
Eα (ELk )+ ∂1 L + (−1)i D+
αi
∂i+1 L x+ (t), . . . , ( c D+
α k
) x+ (t), t = 0,
i=1
(136)
and for (0, x− ),
k
" #
X
αi
Eα (ELk )− ∂1 L + D− ∂i+1 L x− (t), . . . , (− c D−
α k
) x− (t), t = 0. (137)
i=1
Now we consider the embedding of the extended Lagrangian. First we need to set
a vector space for the trajectories, suitable for the calculus of variations. Let Fkα be the
functional space defined by
F α,k ([a, b]) = X ∈ C 0 ([a, b])2 | ∀ 1 ≤ i ≤ k, ( c Dα )i X ∈ C 0 ([a, b])2 .

We also introduce
F+α,k ([a, b]) = F α,k ([a, b]) ∩ (F([a, b], Rn ) × {0}) ,
F−α,k ([a, b]) = F α,k ([a, b]) ∩ ({0} × F([a, b], Rn )) .

The asymmetric fractional embedding of L, still denoted by Lα , is given by
Lα (X)(t) = L̃(X(t), . . . , ( c Dα )k X(t), t)

= L(x+ (t) + x− (t), . . . , ( c D+
α k
) x+ (t) + (− c D−
α k
) x− (t), t),
for all X = (x+ , x− ) ∈ F α,k ([a, b]) and t ∈ [a, b].

The associated action is now given by
A(Lα ) : F α,k ([a, b]) −→ R

Z b
X 7−→ L̃ X(t), c Dα X(t), . . . , ( c Dα )k X(t), t dt.
a
The variations should be of course in F α,k ([a, b]) and should be suitable for the inte-
gration by parts. The space C0k ([a, b]) is suitable (but may not be optimal). In particular,
C0k ([a, b]) ⊂ F α,k ([a, b]) from Lemmas 55 and 49.
The differential of the action is given by the following result.
Lemma 71. Let L be an extended Lagrangian and X ∈ F α,k ([a, b]).

We suppose that for all 1 ≤ i ≤ k, ∂i+1 L̃(X(•), . . . , ( c Dα )k X(•), •) ∈ AC αi ([a, b]).
Then A(Lα ) is C0k ([a, b])2 -differentiable at X and for all H = (h+ , h− ) ∈ C0k ([a, b])2 ,
" k
#
Z b X
αi
∂i+1 L̃ X(t), . . . , ( c Dα )k X(t), t · h+ (t) dt

dA(Lα )(X, H) = ∂1 L̃ + D−
a i=1
" k
#
Z b X
i αi
∂i+1 L̃ X(t), . . . , ( c Dα )k X(t), t · h− (t) dt.

+ ∂1 L̃ + (−1) D+
a i=1
Proof. Let H = (h+ , h− ) ∈ C0k ([a, b])2 and ε > 0. Similarly to Lemma 75, we have:
Z b
A(Lα )(X + εH) = A(Lα )(X) + ε ∂1 L̃(X(t), . . . , ( c Dα )k X(t), t) · (h+ (t) + h− (t)) dt
a
Z b k
X
+ε ∂i+1 L̃(X(t), . . . , ( c Dα )k X(t), t) · ( c D+
α i
) h+ (t) + (− c D−
α i
) h− (t) dt + o(ε).
a i=1
112 Jacky Cresson
Let 1 ≤ i ≤ k. Since h+ ∈ C0i ([a, b]), it verifies ( c D+

α )i h = c D αi , from Lemma 55.
+ +
By using Lemma 62, we have:
Z b Z b
∂i+1 L̃(. . .) · c D+
αi
h+ (t) dt = αi
D− ∂i+1 L̃(. . .) · h+ (t) dt.
a a
A similar relation holds for h− . Hence we obtain
A(Lα )(X + εH) = A(Lα )(X)

Z b" k
X
#
αi
∂i+1 L̃ X(t), . . . , ( c Dα )k X(t), t · h+ (t) dt

+ε ∂1 L̃ + D−
a i=1
" k
#
Z b X
i αi
∂i+1 L̃ X(t), . . . , ( c Dα )k X(t), t · h− (t) dt + o(ε).

+ε ∂1 L̃ + (−1) D+
a i=1
The terms in h+ and h− are linear, which concludes the proof.
We may still obtain coherent and causal embeddings, thanks to the following equiva-
lences.
Theorem 72. Let (x+ , 0) ∈ F+α,k ([a, b]).

We suppose that for all 1 ≤ i ≤ k, ∂i+1 L(x+ (•), . . . , ( c D+ α )k x (•), •) ∈
+
αi
AC ([a, b]).
(x+ , 0) is a {0} × C0k ([a, b])-extremal of the action A(Lα ) if and only if x+ verifies
k
" #
X
(ELk,α )+ ∀t ∈ (a, b], ∂1 L + i αi
(−1) D+ ∂i+1 L x+ (t), . . . , ( c D+
α k
) x+ (t), t = 0.
i=1
(138)
Let (0, x− ) ∈ F−α,k ([a, b]).
We suppose that for all 1 ≤ i ≤ k, ∂i+1 L(x− (•), . . . , (− c D− α )k x (•), •) ∈
−
AC αi ([a, b]).
(0, x− ) is a C0k ([a, b]) × {0}-extremal of the action A(Lα ) if and only if x− verifies
k
" #
X
αi
(ELk,α )− ∀t ∈ [a, b), ∂1 L + D− ∂i+1 L x− (t), . . . , (− c D−
α k
) x− (t), t = 0.
i=1
(139)
Proof. From Theorem 1.2.4 of [25], the fundamental lemma in the calculus of variations is
still valid for variations in C0∞ ([a, b]). Since C0∞ ([a, b]) ⊂ C0k ([a, b]), the result is proved.
Equations (138) and (139) are once again similar to (136) and (137): (ELk,α )± ≡
Eα (ELk )± . The following diagrams are thus valid:
+ −
Eα Eα
L Lα L Lα
C0k ([a,b]) {0}×C0k ([a,b]) C0k ([a,b]) C0k ([a,b])×{0}
+ −
Eα Eα
(ELk ) Eα (ELk )+ (ELk ) Eα (ELk )−
4.6. Variational Asymmetric Fractional Embedding-PDEs Case

In this part, we recall the classical definitions of the left and right Riemann-Liouville and
Caputo derivatives in the one dimensional case. We define the multidimensional fractional
analogous for partial derivatives, gradient and divergence. In particular, we prove a frac-
tional Green-Riemann theorem. We give a simplified version of the asymmetric fractional
calculus of variations introduced in [10] for which we refer for more details and weaker
assumptions on functional spaces. This formalism will be used in the last part to derive the
convection-diffusion equation from a variational principle.
4.6.1. Asymmetric Fractional Calculus of Variations

The fractional Euler-Lagrange equations obtained so far in [36, 2, 8, 22] involve both left
and right fractional derivatives. This is a main drawback, if ones want to recover PDEs with
order one derivative as composition of fractional derivatives of order 1/2. In this section, we
give a simplified version of the asymmetric calculus of variations introduced in [10] which
provides causal fractional Euler-Lagrange equations. We refer to [10] and [26] for more
details and in particular for weaker assumptions on the functional spaces.
Asymmetric Fractional Lagrangian Here the fractional derivatives are seen as partial
fractional derivatives according to t and x. With the notations ux and ut from section 1
4.1.2 we denote by
c α
D+ u(t, x) := c D+
α
ux (t), c
∇α u(t, x) := c ∇α ut (x) and ∇u(t, x) := ∇ut (x).
For a field U = (u+ , u− ) we define

c
Dα U (t, x) := c α
u+ (t, x), − c D−
α c
∇α U (t, x) := c
∇α u+ (t, x), −c ∇α u− (t, x) .

D+ u− (t, x)
For a generalised Lagrangian L(t, x, y, v, w, z), we denote ∂y L, ∂v L, ∂w and ∂z the partial

derivatives of L.
Definition 73. Let be a Lagrangian L defined as follows:
L : [a, b] × Ω × R × R × Rd × Rd −→ R
(t, x, y, v, w, z) 7−→ L(t, x, y, v, w, z)
The asymmetric representation of L is denoted by L̃ and given by:
L̃ : [a, b] × Ω × R2 × R2 × R2d × R2d −→ R

(t, x, U, V, W, Z) 7−→ L̃(t, x, U, V, W, Z)
114 Jacky Cresson
where

L̃ t, x, (u+ , u− ), (v+ , v− ), (w+ , w− ), (z+ , z− ) := L(t, x, u+ + u− , v+ + v− , w+ + w− , z+ + z− ).
The asymmetric fractional Lagrangian functional Lα is defined by:
Definition 74. Let L be a Lagrangian as defined in definition 73 of class C 1 ([a, b]×R3d+2 )

and L̃ its asymmetric representation. The associated asymmetric fractional Lagrangian
functional of order α is given by
Lα : C 1 ([a, b] × Ω)2 −→ R
Z bZ
U 7−→ L̃ (t, x, U (t, x), c Dα U (t), c ∇α U (t, x), ∇U (t, x)) dx dt.
a Ω
For convenience, we denote by Lα the functional defined for any U = (u+ , u− ) ∈

C 1 ([a, b] × Ω)2 by
Lα (U )(t, x) := L̃ (t, x, U (t, x), c Dα U (t), c ∇α U (t, x), ∇U (t, x))

= L(t, x, u+ (t, x) + u− (t, x), c D+ α u (t, x) − c D α u (t, x),
+ − − (140)
c ∇α u (t, x) − c α
+ ∇ u− (t, x), ∇u+ (t, x) + ∇u− (t, x)),
for any x ∈ Ω and t ∈ [a, b] so that

Z bZ
Lα (U ) = Lα (U )(t, x)dx dt.
a Ω
Let us note that as U ∈ C 1 ([a, b] × Ω) we have c DUα (t), c ∇αx U (t, x) and ∇x U (t, x) ∈
C 0 ([a, b] × Ω). Using the fact that L is C 1 ([a, b] × R3d+2 ) we obtain that Lα (C 1 ([a, b] ×
Ω) ⊂ C 0 ([a, b] × Ω) and conclude that Lα is well defined.
Asymmetric Calculus of Variations The next lemma explicits the differential of the
functional Lα defined on U ∈ (C 1 ([a, b] × Ω))2 in the direction (C01 ([a, b] × Ω))2
Lemma 75. Let U ∈ (C 1 ([a, b] × Ω))2 . Let ⋆α :=

c α c α
(t, x, U (t, x), D U (t, x), ∇ U (t, x), ∇U (t, x)) . We assume that
• ∀ x ∈ Ω, t 7→ ∂v L(⋆α ) ∈ AC 1 ([a, b]),
• ∀ t ∈ [a, b], x 7→ ∂w L(⋆α ) ∈ AC 1 (Ω),
• ∀ t ∈ [a, b], x 7→ ∂z L(⋆α ) ∈ C 1 (Ω).
Then Lα is (C 1 ([a, b] × Ω))2 -differentiable at U and in any direction H = (h+ , h− ) ∈

(C01 ([a, b] × Ω))2 , the differential of Lα is given by
Z b Z
α
∂y L(⋆α ) + D−
α
∂v L(⋆α ) + div (∂w L(⋆α )) − div(∂z L(⋆α )) · h+ (t) dx dt

DLα (U, H) =
aZ ΩZ
b
+ [∂y L(⋆α ) − D+
α
∂v L(⋆α ) − divα (∂w L(⋆α )) − div(∂z L(⋆α ))] · h− (t) dx dt.
a Ω
Proof. Using a Taylor expansion of the Lagrangian L, we obtain:

Z bZ
∂y L(⋆) · (h+ + h− ) + ∂v L(⋆) · ( c D+
α
h+ − c D−
α

DLα (U, H) = h− )
a Ω
+∂w L(⋆) · c ∇α h+ − c ∇α h− + ∂z L(⋆) · (∇h+ + ∇h+ ) dx dt.

As ∂v L(⋆) ∈ AC 1 ([a, b]) by assumption, using the integration by parts formula of lemma
46 with h+ (resp. h− ) in C01 ([a, b]) leads to
Z b Z Z b Z
c α c α α α

∂v L(⋆)·( D+ h+ − D− h− ) dx dt = D− ∂v L(⋆) · h+ − D+ ∂v L(⋆) · h− dx dt.
a Ω a Ω
As ∂w L(⋆) ∈ AC 1 (Ω) and h+ (resp. h− ) is in C01 ([a, b] × Ω), we can apply the fractional
Green-Riemann formula from lemma 53 to obtain
Z b Z Z b Z
c α
∇α h+ − c ∇α h− dx dt = div (∂w L(⋆)) · h+ − divα (∂w L(⋆)) · h− dx dt.

∂w L(⋆) ·
a Ω a Ω
The last part of the formula comes from the usual Green-Riemann theorem. This completes
the proof.
A consequence of the previous lemma is the following characterisation of the extremals

of Lα for asymmetric fractional functional as solutions of two fractional Euler-Lagrange
equations:
Theorem 76. Let U ∈ (C 1 ([a, b] × Ω))2 .

Let ⋆α := (t, x, U (t, x), c Dα U (t, x), c ∇α U (t, x), ∇U (t, x)). We assume that
• ∀ x ∈ Ω, t 7→ ∂v L(⋆α ) ∈ AC 1 ([a, b]),
• ∀ t ∈ [a, b], x 7→ ∂w L(⋆α ) ∈ AC 1 (Ω),
• ∀ t ∈ [a, b], x 7→ ∂z L(⋆α ) ∈ C 1 (Ω).
Then (C01 ([a, b] × Ω))2 extremals of Lα correspond to solutions of the following set of
fractional Euler-Lagrange equations:
∂u L(⋆α ) + D−α ∂ L(⋆α ) +divα ∂ L(⋆α ) −div ∂ L(⋆α )

v w z
α ∂ L(⋆α ) −divα ∂ L(⋆α ) −div ∂ L(⋆α )
= 0,
∂u L(⋆α ) − D+

v w z = 0.
4.6.2. Specialisation
This theorem is not applicable as it is for classical PDEs since as it provides a system of
PDEs. However, by restricting our attention to extremals over C 1 ([a, b] × Ω) × {0} over
variations in {0} × C01 ([a, b] × Ω), we obtain a more interesting version:
Theorem 77. Let u+ ∈ C 1 ([a, b] × Ω).

Let ⋆α+ := t, x, u+ (t, x), c D+
α u (t, x), c ∇u (t, x), ∇u (t, x) . We assume that

+ + +
• ∀ x ∈ Ω, t 7→ ∂v L(⋆α+ ) ∈ AC 1 ([a, b]),
• ∀ t ∈ [a, b], x 7→ ∂w L(⋆α+ ) ∈ AC 1 (Ω),

116 Jacky Cresson
• ∀ t ∈ [a, b], x 7→ ∂z L(⋆α+ ) ∈ C 1 (Ω).

Then (u+ , 0) is a {0} × C01 ([a, b] × Ω)-extremal of the action Lα if and only if u+ satisfies
∂y L(⋆α+ ) − D+ α
∂v L(⋆α+ ) −divα ∂w L(⋆α+ ) −div ∂z L(⋆α+ ) = 0,

for any x ∈ Ω, t ∈ [a, b].

The proof is a consequence of lemma 75.
Remark 4. In the previous part, we have heuristically justified the introduction of a doubled
phase space by saying that irreversibility induces a natural arrow of time. Our idea to focus
only on curves in C 1 ([a, b] × Ω) × {0} is precisely to say that we are interested in one
direction of time (here the future). However, the selection of a transverse set for variations
to the underlying phase space is not so clear. It means heuristically that the future depends
mostly on the virtual variations in the past. More work are needed in this direction.
5. Fractional Variational Formulation of Dissipative Ordinary

Differential Equations
5.1. Linear Friction
The differential equation of linear friction is
d2 d
m 2
x(t) + γ x(t) − ∇U (x(t)) = 0, (141)
dt dt
where t ∈ [a, b], m, γ > 0 and U ∈ C 1 (Rn ).
Even if U (x) is quadratic, it has been shown in [5] that this equation cannot be de-
rived from a variational principle with classical derivatives. But this can be done by using
d 1/2 1/2
fractional derivatives, since = c D+ ◦ c D+ , which is proved in the following lemma.
dt
Lemma 78. If f ∈ AC 2 ([a, b]), we have:
• if 0 < α < 1/2, c D+
α ◦ c D α f = c D 2α f ,
+ +
1/2 1/2
• if α = 1/2, c D+ ◦ c D+ f = f ′ ,
f ′ (a)
• if 1/2 < α < 1, for all t ∈ (a, b], c D+
α ◦ c D α f (t) = c D 2α f (t) +
+ + (t −
Γ(2 − 2α)
a)1−2α .
Proof. We prove it by induction on k. For k = 1, the result is obvious. Now, let k ∈ N∗
k+1 k ([a, b]), we use the induction hypothesis: ( c D α )k f =
and f ∈ C+ ([a, b]). Since f ∈ C+ +
c D αk f = αk−αk (αk)
+ I+ f . We have αk ≤ k, so f (αk) ∈ AC([a, b]), and from Lemma 48,
( c D+
α )k f
∈ AC([a, b]). Moreover, from Lemma 49, ( c D+α )k f (a) = 0. We may then
apply Theorem 45:

α
D+ ◦ ( c D+
α k
) f = c D+
α
◦ ( c D+
α k
) f = ( c D+
α k+1
) f.
α ◦ ( c D α )k f = d 1−α αk−αk (αk)

On the other hand, D+ + ◦ I+ ◦ I+ f . We have f (αk) ∈
dt
C 0 ([a, b]), so we may use formula 2.21 of [39, p.34]:
1−α αk−αk (αk) β (αk)
I+ ◦ I+ f = I+ f ,
where β = 1 + αk − α(k + 1).

Since f (αk) ∈ C+
1 ([a, b]), from Lemma 50,
d β (αk) β (αk+1)
◦ I+ f = I+ f .
dt
We have αk + 1 ∈ {α(k + 1), α(k + 1) + 1}, so we consider two cases.
• If αk + 1 = α(k + 1), then
α α(k+1)−α(k+1)
D+ ◦ ( c D+
α k
) f = I+ f (α(k+1)) ,
α(k+1)
= c D+ f.
• If αk = α(k + 1), then

1+α(k+1)−α(k+1)
α
D+ ◦ ( c D+
α k
) f = I+ f (α(k+1)+1) ,
α(k+1)−α(k+1) 1 (α(k+1)+1)
= I+ ◦ I+ f .
We have I+ 1 f (α(k+1)+1) (t) = f (α(k+1)) (t) − f (α(k+1)) (a). But α(k + 1) ≤ k, so
f (α(k+1)) (a) = 0.
Consequently,
α(k+1)−α(k+1)
α
D+ ◦ ( c D+
α k
) f = I+ f (α(k+1)) ,
α(k+1)
= c D+ f.
α(k+1)
In both cases, we have proved that ( c D+
α )k+1 f = c D
+ f , which concludes the
proof.
m γ
We consider the function L(x, v, w, t) = w2 − v 2 − U (x), which is an extended
2 2
Lagrangian.
The variations should be chosen in C02 ([a, b]), but the space
AC02 ([a, b]) = {f ∈ AC 2 ([a, b]) | f (a) = f (b) = 0}
is actually sufficient.
Theorem 79. Let x ∈ C 2 ([a, b]). Then x is solution of (141) if and only if (x, 0) is a
{0} × AC02 ([a, b])-extremal of the action A(L1/2 ).
118 Jacky Cresson
1/2 1/2 1/2 1/2 d
Proof. From Lemma 78 and its proof, c D+ ◦ c D+ x(t) = D+ ◦ c D+ x(t) = x(t),
dt
1/2
for all t ∈ [a, b]. Hence L(x(•), c D+ x(•), x′ (•), •) ∈ C 0 ([a, b]) and the action A(L1/2 )
is well defined.
Let t ∈ [a, b]. The partial derivatives of L verify:
1/2
• ∂1 L(x(t), c D+ x(t), x′ (t), t) = −∇U (x(t)),
1/2 1/2
• ∂2 L(x(t), c D+ x(t), x′ (t), t) = −γ c D+ x(t),
1/2
• ∂3 L(x(t), c D+ x(t), x′ (t), t) = m x′ (t).
1/2 1/2
Since x′ ∈ AC([a, b]), c D+ x = I+ x′ ∈ AC([a, b]), from Lemma 48. Conse-
1/2
quently, for all 1 ≤ i ≤ 2, ∂i+1 L(x(•), c D+ x(•), x′ (•), •) ∈ AC αi ([a, b]). Conditions
of application of Theorem 72 are hence fulfilled. Since C0∞ ([a, b]) ⊂ AC02 ([a, b]), the
choice of {0} × AC02 ([a, b]) for the variations is valid and Theorem 72 may be applied :
(x, 0) is a {0} × AC02 ([a, b])-extremal of the action A(L1/2 ) if and only if x verifies
h i
1/2 1 1/2
∂ 1 L − D + ∂ 2 L + D+ ∂3 L (x(t), c D+ x(t), x′ (t), t) = 0. (142)
1 x′ (t) = x′′ (t), (142) is exactly (141).

Given that D+
We see here the necessity of having a causal Euler-Lagrange equation. Indeed, an equa-
tion similar to (132) would have provide D− α ◦ c D α which is never equal to d .
+
dt
Furthermore, the choice of ( D ) instead of c Dαk in the asymmetric fractional em-
c α k
bedding is justified here. If we had taken c Dαk , the evaluation of the Lagrangian in
1/2
this example would have been L(x(t), c D+ x(t), x′ (t) − x′ (a), t), since c D+ 1 x(t) =
x′ (t) − x′ (a). Hence the initial condition x′ (a) = 0 should have been added to obtain
(141), which is too restrictive for the solutions of (141).
6. Variational Formulation of Dissipative Partial Differential

Equations
6.1. The Diffusion Equation
We are now interested in the diffusion equation
∂
u(x, t) = c ∆u(x, t), (143)
∂t
where t ∈ [a, b], x ∈ Ω, c > 0, and ∆ is the Laplace operator.
1 c
We consider the generalized Lagrangian L(u, v, w, x, t) = w2 − v 2 .
2 2
Theorem 80. Let u ∈ F(Ω × [a, b], R) such that
• ∀x ∈ Ω, t 7→ u(x, t) ∈ AC 2 ([a, b]),

• ∀t ∈ [a, b], x 7→ u(x, t) ∈ C 2 (Ω).
Then u is solution of (143) if and only if (u, 0) is a {0} × C02 (Ω × [a, b])-extremal of
the action A(L1/2 ).
Proof. Let x ∈ Ω and t ∈ [a, b]. The partial derivatives of L verify:

1/2 1/2
• ∂w L(u(x, t), ∇u(x, t), c D+ u(x, t), x, t) = c D+ u(x, t),
1/2
• ∀1 ≤ i ≤ n, ∂vi L(u(x, t), ∇u(x, t), c D+ u(x, t), x, t) = −c ∂xi u(x, t),
so conditions of Theorem 77 are fulfilled, and we have:

(u, 0) is a {0} × V α (Ω × [a, b])-extremal of the action A(L1/2 ) if and only if u verifies
n
" #
1/2 1/2
X
∂u L − ∂ x i ∂ vi L − D+ ∂ w L (u(x, t), ∇u(x, t), c D+ u(x, t), x, t) = 0. (144)
i=1
n
X 1/2 1/2 ∂
Given that ∂xi ∂xi u(x, t) = ∆u(x, t) and D+ ◦ c D+ u(x, t) = u(x, t), (144)
∂t
i=1
is exactly (143).
∂
Once again, causality is essential so as to obtain the term u(x, t).
∂t
6.2. The Fractional Wave Equation

In this section, we derive the fractional wave equation defined by Schneider and Wyss [40]
as the extremals of a fractional continuous Lagrangian systems.
The equation describing waves propagating on a stretched string of constant linear mass
density ρ under constant tension T is
∂ 2 u(t, x) ∂ 2 u(t, x)
ρ = T , (145)
∂t2 ∂x2
where u(t, x) denotes the amplitude of the wave at position x along the string at time t. The
wave equation corresponds to the extremals of the generalized functional associated to the
Lagrangian systems
ρ T
L(t, x, y, v, w) = v 2 − w2 . (146)
2 2
In [40], the authors define the fractional analogue of the wave equation by changing the
classical derivative by a fractional one. Using our notations, the definition of the fractional
wave equation is:
Definition 81. The fractional wave equation of order α is the fractional partial differential
equation
∂2u
−ρc D−2α
u = T 2. (147)
∂x
120 Jacky Cresson
A natural demand with respect to this generalization which is just a formal manipulation
on equations, is to keep a more structural property of the wave equation, namely the fact
that it derives from a least-action principle.
It is easy to prove that Theorem 80 is also valid replacing α = 1/2 by 0 < α < 1/2.
We then deduce :
Theorem 82. Let u ∈ F(Ω × [a, b], R) such that
• ∀x ∈ Ω, t 7→ u(x, t) ∈ AC 2 ([a, b]),
• ∀t ∈ [a, b], x 7→ u(x, t) ∈ C 2 (Ω).
Then u is solution of (147) if and only if (u, 0) is a {0} × C02 (Ω × [a, b])-extremal of
the action A(Lα ) with L defined by (146).
6.3. The Convection-Diffusion Equation

The convection-diffusion equation occurs in many physical problems such as porous media,
engineering, geophysics. It could model the dispersion of a pollutant in a river estuary, or
groundwater transport, atmospheric pollution, concentration of electron inducing an electric
current, heat transfer in a heated body. As many PDEs, the solution exists under conditions,
but is often not known explicitly. Even for linear convection-diffusion equation numerical
schemes are not always well understood. It is still a challenging problem to obtain efficient
and robust numerical schemes to solve the convection-diffusion equation due in particular
to the mixing between two different types of behavior, namely the convective and diffusive
regimes. Let us first recall the equation. Let Ω ⊂ Rd be an open subset with a Lipschitz-
continuous boundary ∂Ω. The time domain [a, b], 0 ≤ a ≤ b is arbitrary, but fixed. Let us
consider the general linear parabolic equation of second order:
ut + (γ · ∇)u − div(K · ∇u) + βu = f (t, x) in (a, b] × Ω,

u(t, x) = 0 in (a, b] × ∂Ω, (148)
u(a, x) = u0 (x) in Ω ,
where γ ∈ Rd , K ∈ Rd×d , β ∈ R.
As an example u is the concentration of a pollutant, transported by a flow of velocity
γ ∈ Rd . The tensor K represents the diffusivity of the pollutant specie. The creation or
destruction of the specie can be taken into account via βu, and f is the source term. The
unknown u is both depending on time and space.
We assume the coefficients are smooth, bounded and satisfying the following properties:
• K is symmetric, uniformly positive definite s.t.
K ∈ C 0 (a, b; L∞ (Ω)d×d ) and ∃λ1 , λ2 > 0 : λ1 |ξ|2 < ξKξ T < λ2 |ξ|2
• the convection γ is such that
γ ∈ C 0 (a, b; W 1,∞ (Ω)d ) and divγ = 0

• the reaction β is non-negative

β ∈ C 0 (a, b; L∞ (Ω)) and ∃β0 β ≥ β0 .
These assumptions guarantee that the problem (148) is well posed for f ∈
L2 (a, b; H −1 (Ω)), and every u0 ∈ L2 (Ω).
Let us notice that in the special case of absence of convection, when γ = 0, the station-
ary convection-diffusion equation is simply the Poisson equation. It is well known that the
Poisson equation derives from a variational principle also called least-action principle. This
means that the solution u of the Poisson equation is a minimizer of the following Lagrangian
functional
L : H01 (Ω) −→ R
1
Z Z
v 7→ L(v) = (K · ∇v) · ∇v dx − f v dx .
Ω 2 Ω
This is not the case of the convection-diffusion equation. For instance, the stationary
convection-diffusion equation admits the following weak formulation:
Z Z Z
(K · ∇v) · ∇φ dx + (γ · ∇)u φ + βuφ dx = f φ dx , for any φ ∈ H01 (Ω). (149)
Ω Ω Ω
Nevertheless the advective term is not symmetric in u and φ. As a consequence, the weak
formulation (149) does not derive from a potential, [42]. This can also be seen as the
convection-diffusion equation does not satisfy the so-called Helmholtz conditions, i.e. that
the Fréchet derivative of the Euler-Lagrange expression is not self-adjoint. We refer to
([34], Thm. 5.92, p.364) for more details.
Let us note that there were some attempt to construct variational formulation for the
convection-diffusion equation by Ortiz [35], where he resorts to a local transformation of
the solution by use of a “dual” problem. In this Section, we apply the previous fractional
formalism to obtain a fractional Lagrangian variational formulation.
Let us consider the reaction-convection-diffusion equation defined on [a, b] × Ω by
(148):
∂
u(t, x) + γ · ∇u(t, x) − div (K · ∇u(t, x)) + βu(t, x) = f (t, x). (150)
∂t
with constant coefficients γ ∈ Rd , K ∈ Rd×d and β ∈ R. As we already mentioned, this
equation does not derive from a variational principle in the classical sense. Nevertheless the
result of the previous section allows us to overcome this difficulty and obtain a variational
formulation of the convection-diffusion equation by mean of the asymmetric fractional La-
grangian. Let us defined the extended Lagrangian L given by
L: [a, b] × Ω × R × R × Rd × Rd −→ R
1 2 1 2 1 1
(t, x, y, v, w, z) 7−→ f (t, x)y − βy + v + (γ × w) · w − (K · z) · z.
2 2 2 2
The direct application of theorem 77 provides that the solutions of the convection-diffusion
equation are {0} × C01 ([a, b] × Ω)-extremals of the following asymmetric fractional func-
tional L1/2 defined for U = (u+ , u− ) by
Z bZ
1/2 1/2
L1/2 (U ) = L t, x, u+ (t, x) + u− (t, x), c D+ u+ (t, x) − c D− u− (t, x),
a Ω
c ∇1/2 u (t, x) − c ∇1/2 u (t, x), ∇u (t, x) + ∇u (t, x) dx dt.

+ − + −
122 Jacky Cresson
Namely, the following result holds:
Theorem 83. Let u ∈ F 2 (Ω × [a, b]). Then u is a solution of the convection-diffusion

equation (150) if and only if (u, 0) is a {0} × C01 ([a, b] × Ω) critical point of L1/2 .
Proof. Let x ∈ Ω and t ∈ [a, b]. Let t ∈ [a, b]. The partial derivatives of L verify:
1/2
• ∂y L(t, x, u(t, x), c D+ x(t), c ∇1/2 u(t, x), ∇u(t, x)) = f (t, x) − βu(t, x),
1/2 1/2
• ∂v L(t, x, u(t, x), c D+ x(t), c ∇1/2 u(t, x), ∇u(t, x)) = c D+ u(t, x),
1/2
• ∂w L(t, x, u(t, x), c D+ x(t), c ∇1/2 u(t, x), ∇u(t, x)) = γ × c ∇1/2 u(t, x),
1/2
• ∂z L(t, x, u(t, x), c D+ x(t), c ∇1/2 u(t, x), ∇u(t, x)) = −K · ∇u(t, x).
1/2
As u ∈ F 2 ([a, b] × Ω) we have that ux ∈ AC 2 ([a, b]) and as a consequence c D+ u ∈
AC 1 ([a, b]). We have also ut ∈ C 2 (Ω) so that using lemma 47, we deduce c ∇1/2 u ∈
C 1 (Ω). Moreover ∇u ∈ C 1 ([a, b]) so that the conditions of theorem 77 are fulfilled. From
1/2 1/2 d
lemma 46, as ux ∈ AC 2 ([a, b]) we have D+ ◦ c D+ u = u. Moreover, as ut ∈ C 2 (Ω),
dt
lemma 51 applies and we have

div1/2 γ × c ∇1/2 u(x) = γ · ∇u(x).
References
[1] Vladimir I. Arnold. Mathematical methods of classical mechanics, volume 60 of
Graduate Texts in Mathematics. Springer-Verlag, New York, second edition, 1989.
Translated from the Russian by K. Vogtmann and A. Weinstein.
[2] Om Prakash Agrawal. Formulation of Euler-Lagrange equations for fractional varia-

tional problems. J. Math. Anal. Appl., 272(1):368–379, 2002.
[3] Andler M., Jean Leray (1906-1998), Proceedings of the American Philosophical So-
ciety, Vol. 144, no. 4, 2000.
[4] Harry Bateman. On dissipative systems and related variational principles. Physical
Review, 38(1):815–819, 1931.
[5] P. S. Bauer. Dissipative dynamical systems. Proc. Nat. Acad. Sci., 17:311–314, 1931.
[6] L. Bourdin, J. Cresson, I. Greff, P. Inizan, Variational integrators for fractional La-
grangian systems in the framework of discrete embeddings, Preprint 2011.
[7] L. Bourdin, J. Cresson, I. Greff, A Continuous/discrete fractional Noether’s theorem,

Communication in Nonlinear Science and Numerical Simulation, 2012.
[8] J. Cresson. Fractional embedding of differential operators and Lagrangian systems. J.

Math. Phys., 48(3):033504, 34, 2007.
[9] J. Cresson, I. Greff, and C. Pierre. Coherent discrete embeddings for Lagrangian and
Hamiltonian systems. Arxiv-1107.0894, pages 1–24, 2011.
[10] J. Cresson and P. Inizan. Variational formulations of differential equations and asym-
metric fractional embedding. Journal of Mathematical Analysis and Applications,
Volume 385, Issue 2, 15 January 2012, Pages 975-997.
[11] J. Cresson and P. Inizan. Irreversibility, least action principle and causality. arXiv,
0812.3529v2, 2008. Unpublished.
[12] J. Cresson, Non-differentiable deformations of Rn , International Journal of Geomet-

ric Methods in Modern Physics, Vol. 3, no. 7 (2006) 1395-1415.
[13] J. Cresson and S. Darses (2006), Plongement stochastique des systèmes lagrangiens,
C.R. Acad. Sci. Paris Ser. I 342 (5), 333-336.
[14] J. Cresson and S. Darses, Stochastic embedding of dynamical systems. J. Math. Phys.,
48(7):072703, 54, 2007.
[15] J. Cresson, S. Darses, Théorème de Noether Stochastique, C.R. Acad. Sci. Paris, Ser.
I 344 (2007) 259-264
[16] J. Cresson, Introduction to embedding of Lagrangian systems, International Journal

of Biomathematics and Biostatistics, Vol. 1, no. 1, 23-31, 2010
[17] J. Cresson, Introduction to embedding formalisms and Lagrangian PDEs, Lecture

Notes TU München, 2011.
[18] J. Cresson, I. Greff, A non-differentiable Noether’s theorem, J. Math. Phys. 52,

023513 (2011); doi:10.1063/1.3552936 (10 pages).
[19] J. Cresson and I. Greff, Non-differentiable embedding of Lagrangian systems and

PDEs, 30.p, Journal of Mathematical Analysis and Applications Volume 384, Issue 2
(2011) Pages 626-646.
[20] J. Cresson and P. Inizan. About fractional Hamiltonian systems. Physica Scripta,
T136:014007, 2009.
[21] J. Cresson, D. F-M. Torres, A. Malinovska, Differential, integral and variational time
scales embeddings of Lagrangian systems, Computers Mathematics with applications,
2012.
[22] Frederico, Gastão S. F. and Torres, Delfim F. M., A formulation of Noether’s theorem
for fractional problems of the calculus of variations, J. Math. Anal. Appl., 334, 2007,2,
834–846.
124 Jacky Cresson
[23] Sk. Golam Ali, Benoy Talukdar, and Umapada Das. Inverse problem of variational
calculus for nonlinear evolution equations. Acta Phys. Polon. B, 38(6):1993–2002,
2007.
[24] Ernst Hairer, Christian Lubich, and Gerhard Wanner. Geometric numerical integra-
tion, volume 31 of Springer Series in Computational Mathematics. Springer-Verlag,
Berlin, second edition, 2006.
[25] L. Hörmander. The Analysis of Linear Partial differential Operators I. Springer-

Verlag, 1983.
[26] Pierre Inizan. Dynamique fractionnaire pour le chaos hamiltonien. PhD thesis. 2010.
[27] Malgorzata Klimek. On analogues of exponential functions for antisymmetric frac-

tional derivatives. Comput. Math. Appl., 59(5):1709–1717, 2010.
[28] Laskar J., Large-scale chaos in the solar system, Astron. Astrophys. 287, L9-L12
(1994).
[29] J. Leray, Sur le mouvement d’un liquide visqueux emplissant l’espace, Acta Math. 63
(1934), p. 193-248.
[30] K. Itô, On stochastic differential equations, Mem. Amer. Math. Soc. 4 (1951), 1-51.
[31] Làzaro Cami J.A., Stochastic geometric mechanics, Ph.D., University of Zaragoza,
2008.
[32] Jerrold E. Marsden and Matthew West. Discrete mechanics and variational integrators.
Acta Numer., 10:357–514, 2001.
[33] Marmi, S., Chaotic behaviour in the solar system following J. Laskar, Séminaire Bour-
baki, 51ème année, 1998-1999, Paper No. 854.
[34] Peter J. Olver. Applications of Lie groups to differential equations, volume 107 of
Graduate Texts in Mathematics. Springer-Verlag, New York, second edition, 1993.
[35] Michael Ortiz. A variational formulation for convection-diffusion problems. Internat.

J. Engrg. Sci., 23(7):717–731, 1985.
[36] Fred Riewe. Nonconservative Lagrangian and Hamiltonian mechanics. Phys. Rev. E
(3), 53(2):1890–1899, 1996.
[37] Fred Riewe. Mechanics with fractional derivatives. Phys. Rev. E (3), 55(3, part
B):3581–3592, 1997.
[38] L. Schwartz, Théorie des distributions, Publications de l’Institut de Mathématique de

l’Université de Strasbourg, No. IX-X. Nouvelle édition, entiérement corrigée, refon-
due et augmentée, Hermann, Paris, 1966.
[39] Stefan G. Samko, Anatoly A. Kilbas, and Oleg I. Marichev. Fractional integrals and
derivatives. Gordon and Breach Science Publishers, Yverdon, 1993. Theory and
applications, Edited and with a foreword by S. M. Nikol′ skiı̆, Translated from the
1987 Russian original, Revised by the authors.
[40] Schneider W.R., Wyss W., Fractional diffusion and wave equations, J. Math. Phys. 30
(1989) 134-144.
[41] F.T. Tveter, Deriving the Hamiltonian equations of motion for a nonconservative sys-
tem using a variational principle, Journal of Mathematical Physics, Vol. 39, no. 3
(1998), 1495-1500.
[42] M. M. Vainberg. Variational methods for the study of nonlinear operators. Holden-
Day Inc., San Francisco, Calif., 1964. With a chapter on Newton’s method by L. V.
Kantorovich and G. P. Akilov. Translated and supplemented by Amiel Feinstein.
[43] G.M. Zaslavsky. Hamiltonian Chaos & Fractional Dynamics. Oxford University
Press, Oxford, 2005.

Chapter 3
A C LASS OF F RACTIONAL O PTIMAL C ONTROL

P ROBLEMS AND F RACTIONAL P ONTRYAGIN ’ S
S YSTEMS . VARIATIONAL I NTEGRATOR
AND E XISTENCE OF C ONTINUOUS /D ISCRETE
N OETHER ’ S T HEOREMS
Loı̈c Bourdin∗
Laboratoire de Mathématiques et de leurs Applications de Pau,
Université de Pau et des Pays de l’Adour, Pau Cedex, France
Introduction
The control theory is the analysis of controlled dynamical systems. These systems are var-
ied: they can be differential, stochastic or discrete. The optimal control theory concerns the
determination of a control optimizing a certain cost. Consequently, this theory is strongly
connected to the 18th century classical mechanic (variational principles, Euler-Lagrange
equations, etc., see [6, 32, 47]). Since the second world war, this theory has a considerable
development and one can find applications in many domains: celestial mechanic [11], bi-
ology [15], hydroelectricity [16], economy [20, 22, 26], etc. The subject is widely treated
and one can find a lot of references on the subject, see for example [14, 24, 37, 57].
The fractional calculus, i.e. the mathematical field dealing with the generalization of the
derivative to any real order, plays an increasing role in many varied domains as economy
[17] or probability [44, 56]. Fractional derivatives also appear in many fields of Physics
(see [38]): wave mechanic [5], viscoelasticity [7], thermodynamics [39], fluid mechanic
in heterogeneous media [33, 59, 60], etc. A natural question then arises: can we develop
optimal control theories for fractional differential systems?
Recently, a subtopic of the fractional calculus gains importance: it concerns the varia-
tional principles on functionals involving fractional derivatives. This leads to the statement
∗
E-mail address: bourdin.l@univ-pau.fr
128 Loı̈c Bourdin
of fractional Euler-Lagrange equations, see [1, 9, 54]. A direct consequence is the emer-
gence of works concerning a particular class of fractional optimal control problems, see
[2, 3, 28, 29, 41] and references therein. These studies usually use a Lagrange multiplier
technique allowing to write these problems as problems of optimization without constraint
of augmented functionals. With a calculus of variations, authors then obtain a necessary
condition for the existence of an optimal control. This condition is commonly given as
the existence of a solution of a system of fractional differential equations called fractional
Pontryagin’s system.
In this chapter, we first give a new presentation of this result. Precisely, making an
additional assumption (see Condition (fx lip)), we rewrite directly these fractional optimal
control problems as simpler problems of optimization without constraint of functionals de-
pending only on the control. Although the method used is considerably inspired by the
Lagrange multiplier technique, it allows us to give a complete proof of this result using
only classical mathematical tools adapted to the fractional case: calculus of variations,
Gronwall’s Lemma, Cauchy-Lipschitz Theorem and stability of differential equations under
perturbations.
Nevertheless, the explicit computation of controls satisfying the above necessary con-
dition needs the resolution of a fractional Pontryagin’s system which is a main drawback.
Indeed, solving a fractional differential equation is in general very difficult. Consequently,
in this chapter, we suggest two deviously ways in order to get informations on the solutions
of a fractional Pontryagin’s system.
Firstly, we study the existence of classical conservation laws, i.e. functions which are
constant on each solution. Indeed, constants of motion, generally associated to physical
quantities, give strong informations on the solutions in the phase space for example. More-
over, they also can be used in order to reduce or integrate the equation by quadrature.
Previous results in this direction have been obtained by Torres and Frederico in [28, 29].
However, in each of these papers, the conservation law is not explicit but implicitly defined
by a functional relation. In this chapter, inspired by a recent result obtained in [12], we
prove a fractional Noether’s theorem providing an explicit conservation law for fractional
Pontryagin’s systems exhibiting a symmetry.
As it is done in [3], the second idea is to suggest a numerical approach. In this chap-
ter, we construct a numerical scheme preserving the variational structure of the fractional
Pontryagin’s systems. Indeed, this variational structure is intrinsic and induces strong con-
straints on the qualitative behaviour of the solutions. It seems then important to preserve
it at the discrete level. A variational integrator is a numerical scheme preserving the vari-
ational structure at the discrete level. We refer to Section 2 for more details concerning
the construction of a variational integrator and let us remind that the variational integrators
are well-developed in [36, 46] for classical Euler-Lagrange equations and in [13] for frac-
tional ones. In this chapter, we construct a variational integrator for fractional Pontryagin’s
systems and it is called shifted discrete fractional Pontryagin’s system.
Finally, adapting the strategy from the continuous level to the discrete one, we prove
a discrete fractional Noether’s theorem providing an explicit discrete conservation law for
shifted discrete fractional Pontryagin’s systems exhibiting a discrete symmetry. This result
is widely inspired from the discrete fractional Noether’s theorem proved in [12] for discrete
fractional Euler-Lagrange equations introduced in [13].
A Class of Fractional Optimal Control Problems ... 129
In this chapter, we also suggest:
• a link between classical optimal control problems and their fractional versions via the
Stanislavsky’s formalism, see [40, 56];
• a solved fractional example allowing to test numerical schemes in the strict fractional
case.
The chapter is organized as follows. In Section 1, we first remind classical definitions

and some results concerning fractional calculus (Section 1.1). Then, we present the class
of fractional optimal control problems studied and the usual strategy leading to the frac-
tional Pontryagin’s system (Section 1.2). Making an additional assumption on the problem
(see Condition (fx lip)), we develop a new proof leading to the same result with a differ-
ent presentation and using only classical mathematical tools adapted to the fractional case
(Section 1.3). In Section 1.5, we suggest a way in order to transit from a classical optimal
control problem to its fractional version via the Stanislavsky’s formalism. Subsequently,
we suggest some examples with, in particular, a solved fractional example allowing to test
numerical schemes in the strict fractional case (Section 1.6). Finally, in Section 1.7, we
state a fractional Noether’s theorem for fractional Pontryagin’s systems admitting a sym-
metry. Section 2 is focused on a variational integrator for fractional Pontryagin’s systems.
Indeed, Section 2.1 reminds usual definitions and proves some results on discrete fractional
derivatives. Then, in Section 2.2, we construct a variational integrator for fractional Pon-
tryagin’s systems and it is called shifted discrete fractional Pontryagin’s system. Finally,
after some numerical tests in Section 2.5, we state a discrete fractional Noether’s theorem
for shifted discrete fractional Pontryagin’s systems admitting a discrete symmetry (Section
2.5). Appendices A and B contain technical results and proofs non developed in Sections 1
and 2.
1. A Class of Fractional Optimal Control Problems

1.1. Definitions and Results about Fractional Calculus
Let us introduce the following notations available in the whole chapter. Let a < b be two
reals, let d, m ∈ N∗ denote two dimensions and let k · k be the euclidean norm of Rd and
Rm .
1.1.1. Fractional Operators of Riemann-Liouville and Caputo
The fractional calculus concerns the extension of the usual notion of derivative from non-
negative integer orders to any real order. Since 1695, numerous notions of fractional deriva-
tives emerge over the year, see [42, 52, 55]. In this chapter, we only use the notions of frac-
tional integrals and derivatives in the sense of Riemann-Liouville (1847) and Caputo (1967)
whose definitions are recalled in this section. We refer to [42, 52, 55] for more details.
Let g ∈ C 0 ([a, b], Rd ) and α > 0. The left (resp. right) fractional integral in the sense
of Riemann-Liouville with inferior limit a (resp. superior limit b) of order α of g is defined
130 Loı̈c Bourdin
by:
t
1
Z
α
∀t ∈]a, b], I− g(t) := (t − y)α−1 g(y) dy (1)
Γ(α) a
respectively:
b
1
Z
α
∀t ∈ [a, b[, I+ g(t) := (y − t)α−1 g(y) dy, (2)
Γ(α) t
where Γ denotes the Euler’s Gamma function. Let us remind that I− α g (resp. I α g) is con-
+
tinuous and can be continuously extended by 0 in t = a (resp. t = b). Let us note that I− 1g
(resp. −I+1 g) coincides with the anti-derivative of g vanishing in t = a (resp. t = b). For
0 g = I 0 g = g.
α = 0, let I− +
Now, let us consider 0 < α ≤ 1. The left (resp. right) fractional derivative in the sense
of Riemann-Liouville with inferior limit a (resp. superior limit b) of order α of g is then
given by:
α d 1−α
α d 1−α
∀t ∈]a, b], D− g(t) := I g (t) resp. ∀t ∈ [a, b[, D+ g(t) := − I g (t) ,
dt − dt +
(3)
provided that the right side terms are defined.
In the Riemann-Liouville sense, the strict fractional derivative of a constant is not zero.
Caputo then suggests the following definition. For 0 < α ≤ 1, the left (resp. right)
fractional derivative in the sense of Caputo with inferior limit a (resp. superior limit b) of
order α of g is given by:
α α

∀t ∈]a, b], c D− g(t) := D− g − g(a) (t) (4)

α α
resp. ∀t ∈ [a, b[, c D+ g(t) := D+ g − g(b) (t) ,
provided that the right side terms are defined. Let us note that if g(a) = 0 (resp. g(b) = 0),
then c D− α g = D α g (resp. D α g = D α g).
− c + +
In the classical case α = 1, the fractional derivatives of Riemann-Liouville and Caputo
both coincide with the classical derivative. Precisely, modulo a (−1) term in the right case,
we have D− 1 = D 1 = −D 1 = − D 1 = d/dt.
c − + c +
Finally, let us remind the following important result. If g ∈ C 1 ([a, b], Rd ) and 0 < α ≤
1, then we have:
α 1−α α 1−α
∀t ∈]a, b], c D− g(t) = I− ġ(t) and ∀t ∈ [a, b[, c D+ g(t) = −I+ ġ(t), (5)
α g (resp. D α g)
where ġ denotes the derivative of g. Consequently, in this case, we have c D− c +
is continuous and can be continuously extended in t = a (resp. t = b).
1.1.2. Some Properties and Results about the Fractional Operators

Let us precise that the properties and results developed in this section are well-known and
already proved in many references, see [42, 55] for example. However, these results are not
always exactly presented as we need. In this case, we give a detailed proof for the reader’s
convenience.
First, we remind two basic results concerning the fractional integrals. They are proved
in [42, 55] both using the Fubini’s theorem. The first one yields the semi-group property of
the fractional integral operators:
Property 1. Let g ∈ C 0 ([a, b], Rd ) and α1 , α2 ≥ 0. Then, we have:
α1 α2 α1 +α2 α1 α2 α1 +α2
∀t ∈ [a, b], I− ◦ I− g(t) = I− g(t) and I+ ◦ I+ g(t) = I+ g(t). (6)
The following second property is occasionally called fractional integration by parts. It
is very useful for calculus of variations involving fractional derivatives:
Property 2. For any g1 , g2 ∈ C 0 ([a, b], Rd ) and any α ≥ 0, we have:
Z b Z b
α α
I− g1 · g2 dt = g1 · I+ g2 dt. (7)
a a
Let us introduce the following convention: a function defined on ]a, b] (resp. [a, b[) is
said to be an element of C 0 ([a, b], Rd ) if and only if it is continuous on ]a, b] (resp. [a, b[)
and if it can be continuously extended in t = a (resp. t = b). From Section 1.1.1, we can
give the following examples:
• for g ∈ C 0 ([a, b], Rd ) and α ≥ 0, we have I−
α g, I α g ∈ C 0 ([a, b], Rd );
+
• for g ∈ C 1 ([a, b], Rd ) and 0 < α ≤ 1, we have c D−

α g, D α g ∈ C 0 ([a, b], Rd ).
c +
Now, we prove some results of composition between the left fractional operators. One can
easily derive the analogous versions for the right ones.
Property 3. Let g ∈ C 0 ([a, b], Rd ) and 0 < α ≤ 1. Then, c D−
α ◦ I α g is an element of
−
0 d
C ([a, b], R ) and for any t ∈ [a, b], we have:
α α
c D− ◦ I− g(t) = g(t). (8)
Let us assume additionally that c D− α g ∈ C 0 ([a, b], Rd ). Then, I α ◦ D α g is an element of
− c −
C 0 ([a, b], Rd ) and for any t ∈ [a, b], we have:
α α
I− ◦ c D− g(t) = g(t) − g(a). (9)
α g(a) = 0, we have for any t ∈]a, b]:
Proof. Let us prove the first result. Since I−
α α α α d 1−α α d 1
c D− ◦ I− g(t) = D− ◦ I− g(t) = ◦ I− ◦ I− g(t) = I g (t) = g(t). (10)
dt dt −
α ◦ I α g is continuous on ]a, b] and can be continuously extended by g(a) in
Hence, c D− −
t = a. Now, let us prove the second result. It is obvious for α = 1. Now, let us consider
α g ∈ C 0 ([a, b], Rd ), we have I 1−α g − g(a) ∈ C 1 ([a, b], Rd ).

0 < α < 1. Since c D− −
Combining the first result and Equality (5), we have for any t ∈]a, b]:
1−α 1−α
α d 1−α

g(t) − g(a) = c D− ◦ I− g − g(a) (t) = I− ◦ ◦ I− g − g(a) (t)
dt
α α α α

= I− ◦ D− g − g(a) (t) = I− ◦ c D− g(t). (11)
α ◦ D α g is continuous on ]a, b] and can be continuously extended by 0 in t = a.
Hence I− c −
The proof is completed.
132 Loı̈c Bourdin
Finally, we prove the following fractional Cauchy-Lipschitz type theorem. Let us pre-
cise that a version of this theorem is proved in [42, Part 3.5.1, Corollary 3.26, p.205]. It
gives the existence and the uniqueness of a global solution of a fractional differential equa-
tion.
Theorem 4 (Fractional Cauchy-Lipschitz theorem). Let 0 < α ≤ 1, F ∈ C 0 (Rd ×

[a, b], Rd ) and A ∈ Rd . Let us assume that F satisfies the following Lipschitz type con-
dition:
∃K ∈ R, ∀(x1 , x2 , t) ∈ (Rd )2 × [a, b], kF (x1 , t) − F (x2 , t)k ≤ Kkx1 − x2 k. (12)
Then, the following fractional Cauchy problem:

α

c D− g = F (g, t)
(13)
g(a) = A,
has an unique solution in C 0 ([a, b], Rd ). The solution g is an element of C [α] ([a, b], Rd )
(where [α] denotes the floor of α) and g satisfies:
α

∀t ∈ [a, b], g(t) = A + I− F (g, t) (t). (14)
Proof. Firstly, let g ∈ C 0 ([a, b], Rd ). From Property 3, we conclude that g is solution
of the fractional Cauchy problem (13) if and only if g satisfies Equality (14). Now, let us
define:
ϕ : C 0 ([a, b], Rd ) −→ C 0 ([a, b], Rd ) (15)

g 7−→ ϕ(g) : [a, b] −→ Rd
α F (g, t) (t).

t 7−→ A + I−
By induction, we prove that for any n ∈ N∗ and any g1 , g2 ∈ C 0 ([a, b], Rd ), we have:
K n (t − a)nα
∀t ∈ [a, b], kϕn (g1 )(t) − ϕn (g2 )(t)k ≤ kg1 − g2 k∞ . (16)
Γ(1 + nα)
Consequently, we have for any n ∈ N∗ and any g1 , g2 ∈ C 0 ([a, b], Rd ):
K n (b − a)nα
kϕn (g1 ) − ϕn (g2 )k∞ ≤ kg1 − g2 k∞ . (17)
Γ(1 + nα)
Since K n (b − a)nα /Γ(1 + nα) −→ 0, there exists n ∈ N∗ such that ϕn is a contraction.

Consequently, ϕ admits an unique fix point in C 0 ([a, b], Rd ). The proof is completed.
1.2. Presentation of the Problem and Usual Strategy

In this section, we first give a brief presentation of the class of fractional optimal control
problems interesting us in this chapter. Finally, we briefly remind the usual strategy (using
the Lagrange multiplier technique) leading to a necessary condition for the existence of an
optimal control, see [2, 3, 28, 29, 41] and references therein. From now and for all the rest
of the chapter, we consider 0 < α ≤ 1 and A ∈ Rd .
An optimal control problem concerns the optimization of a quantity depending on pa-

rameters given by a controlled system. Precisely, the aim is to find a control optimizing
a certain cost. Such a control is called optimal control. Many of these studies lead to the
use of the Lagrange multiplier technique. For example, we refer to [10, 25, 34, 35] for
controlled systems coded by ordinary differential equations.
In this chapter, we work in the following framework. We are interested in systems
controlled by the following fractional Cauchy problem:
α

c D− q = f (q, u, t)
(18)
q(a) = A,
where u denotes the control. The aim is to find a control u optimizing a quantity of the
form: Z b
L(q, u, t) dt, (19)
a
where q is solution of (18).
The common strategy is first to rewrite this problem as a problem of optimization under
constraint: Z b
arg min L(q, u, t) dt. (20)
(q,u) satisfying (18) a
Then, the Lagrange multiplier technique consists in the study of the critical points of the
following augmented functional:
Z b
α

(q, u, p) 7−→ L(q, u, t) − p · c D− q − f (q, u, t) dt, (21)
a
where p is commonly called Lagrange multiplier. Finally, with a calculus of variations, such
a strategy leads to the following result: a necessary condition for (q, u) to be a solution of
(20) is that there exists a function p such that the following fractional Pontryagin’s system
holds:
∂H

α
c D− q = (q, u, p, t)





 ∂w
 Dα p = ∂H (q, u, p, t)



+
∂x (22)
∂H


(q, u, p, t) = 0


∂v





q(a), p(b) = (A, 0),

where H(x, v, w, t) = L(x, v, t)+w·f (x, v, t). We refer to [2, 3, 28, 29, 41] and references
therein for more details.
In this chapter, we are going to assume that f satisfies a Lipschitz type condition, see
Condition (fx lip). This assumption will allow us to write directly the initial problem as a
simpler problem of optimization without constraint of the initial functional which is then
only dependent of the control u. Finally, a simple calculus of variations leads us to the
same result but with a new presentation. Let us remind that the method here developed is
widely inspired from the Lagrange multiplier technique. Nevertheless, it allows us to give
134 Loı̈c Bourdin
a complete proof only using classical mathematical tools adapted to the fractional case:
calculus of variations, Gronwall’s Lemma, Cauchy-Lipschitz Theorem and stability under
perturbations of differential equations.
1.3. New Presentation of the Result

In this section, we first give rigorously the definitions concerning the class of fractional
optimal control problems described in Section 1.2:
• The elements denoted u ∈ C 0 ([a, b], Rm ) are called controls;
• Let f be a C 2 function of the form:
f : Rd × Rm × [a, b] −→ Rd (23)
(x, v, t) 7−→ f (x, v, t).
It is commonly called the constraint function. We assume that f satisfies the follow-
ing Lipschitz type condition. There exists M ≥ 0 such that:
∀(x1 , x2 , v, t) ∈ (Rd )2 × Rm × [a, b], kf (x1 , v, t) − f (x2 , v, t)k ≤ M kx1 − x2 k;

(fx lip)
• For any control u, let q u,α ∈ C [α] ([a, b], Rd ) denote the unique global solution of the
following fractional Cauchy problem:
α

c D− q = f (q, u, t)
(CPαq )
q(a) = A.
q u,α is commonly called the state variable associated to u. Its existence and its
uniqueness are provided by Theorem 4 and Condition (fx lip);
• Finally, with this condition on f , the fractional optimal control problem described in
Section 1.2 can be rewritten as the simpler problem of optimization of the following
cost functional:
Lα : C 0 ([a, b], Rm ) −→ R (24)

Z b
u 7−→ L(q u,α , u, t) dt,
a
where L is a Lagrangian, i.e. a C 2 application of the form:
L : Rd × Rm × [a, b] −→ R (25)
(x, v, t) 7−→ L(x, v, t).
Hence, the existence and the uniqueness of q u,α for any control u allow us to rewrite di-
rectly the initial problem as a simpler problem of optimization without constraint of the
cost functional Lα : we do not need to introduce an augmented functional with a Lagrange
multiplier. Moreover, let us note that Lα is only dependent of the control.
A control optimizing Lα is called optimal control. A necessary condition for a control

u to be optimal is to be a critical point of Lα , i.e. to satisfy:
Lα (u + εū) − Lα (u)
∀ū ∈ C 0 ([a, b], Rm ), DLα (u)(ū) := lim = 0. (26)
ε→0 ε
In the following, we then focus on the characterization of the critical points of Lα . With an
usual calculus of variations, we obtain the following Lemma 5 giving explicitly the value
of the Gâteaux derivative of Lα :
Lemma 5. Let u, ū ∈ C 0 ([a, b], Rm ). Then, the following equality holds:
Z b
α ∂L u,α ∂L u,α
DL (u)(ū) = (q , u, t) · q̄ + (q , u, t) · ū dt, (27)
a ∂x ∂v
where q̄ ∈ C [α] ([a, b], Rd ) is the unique global solution of the following linearised Cauchy
problem:
 c Dα q̄ = ∂f (q u,α , u, t) × q̄ + ∂f (q u,α , u, t) × ū

−
∂x ∂v (LCPαq̄ )
q̄(a) = 0.

Proof. See Appendix A.4.
This last result does not lead to a characterization of the critical points of Lα yet. Then,
let us introduce the following elements stemming from the Lagrange multiplier technique:
• Let H be the following application
H : Rd × Rm × Rd × [a, b] −→ R (28)
(x, v, w, t) 7−→ L(x, v, t) + w · f (x, v, t).
H is commonly called the Hamiltonian associated to the Lagrangian L and the con-
straint function f ;
• For any control u, let pu,α ∈ C [α] ([a, b], Rd ) denote the unique global solution of the
 T
 Dα p = ∂H (q u,α , u, p, t) = ∂L (q u,α , u, t) + ∂f (q u,α , u, t)

×p
c +
∂x ∂x ∂x (CPα
p)

p(b) = 0.

pu,α is commonly called the adjoint variable associated to u. Its existence and its
uniqueness are provided by the analogous version of Theorem 4 for right fractional
derivative. Since pu,α (b) = 0, we can write c D+
α pu,α = D α pu,α .
+
Let us note that for any control u, the couple (q u,α , pu,α ) is solution of the following frac-
tional Hamiltonian system:

α q = ∂H
 c D− (q, u, p, t)


∂w
(HSα )
 αp = ∂H
 D+
 (q, u, p, t).
∂x
Finally, the introduction of these last elements allows us to prove the following theorem:
136 Loı̈c Bourdin
Theorem 6. Let u ∈ C 0 ([a, b], Rm ). Then, u is a critical point of Lα if and only if

(q u,α , u, pu,α ) is solution of the following fractional stationary equation:
∂H
(q, u, p, t) = 0. (SEα )
∂v
Proof. Let u, ū ∈ C 0 ([a, b], Rm ). From Lemma 5, we have:
Z b
∂L u,α ∂L u,α
DLα (u)(ū) = (q , u, t) · q̄ + (q , u, t) · ū dt. (29)
a ∂x ∂v
Then:
T !
b
∂L u,α ∂f u,α
Z
DLα (u)(ū) = (q , u, t) + (q , u, t) × pu,α · q̄
a ∂x ∂x

∂f u,α ∂L u,α
− (q , u, t) × q̄ · pu,α + (q , u, t) · ū dt. (30)
∂x ∂v
From Theorem 4, since q̄ is solution of (LCPq̄α ) and pu,α is solution of (CPαp ), we have:

α ∂f u,α ∂f u,α
q̄ = I− (q , u, t) × q̄ + (q , u, t) × ū (31)
∂x ∂v
and T !
u,α α ∂L u,α ∂f u,α u,α
p = I+ (q , u, t) + (q , u, t) ×p . (32)
∂x ∂x
Then, using the fractional integration by parts given in Property 2, we obtain:
b
∂f u,α ∂f u,α
Z
α u,α
DL (u)(ū) = p · (q , u, t) × q̄ + (q , u, t) × ū
a ∂x ∂v

∂f u,α ∂L u,α
− (q , u, t) × q̄ · pu,α + (q , u, t) · ū dt. (33)
∂x ∂v
Then:
T !
b
∂f u,α ∂L u,α
Z
α u,α
DL (u)(ū) = (q , u, t) ×p + (q , u, t) · ū dt
a ∂v ∂v
b
∂H u,α
Z
= (q , u, pu,α , t) · ū dt.
a ∂v
The proof is completed by the Dubois-Raymond’s lemma.
Let us note that Lemma 5 is proved in Appendix A.4 using Lemmas 28, 29 and 30.
These last three Lemmas are respectively a fractional Gronwall’s Lemma, a result of stabil-
ity of order 1 and a result of stability of order 2 for the fractional Cauchy problem (CPαq ).
Hence, the proof of Theorem 6 is only based on classical mathematical tools adapted to the
fractional case.
From Theorem 6, we retrieve the following result provided in [2, 3, 28, 29, 41] and
references therein:
Corollary 7. Lα has a critical point in C 0 ([a, b], Rm ) if and only if there exists (q, u, p) ∈
C [α] ([a, b], Rd ) × C 0 ([a, b], Rm ) × C [α] ([a, b], Rd ) solution of the following fractional Pon-
tryagin’s system:
∂H

α
c D− q = (q, u, p, t)





 ∂w
 Dα p = ∂H (q, u, p, t)



+
∂x (PSα )
∂H


(q, u, p, t) = 0


 ∂v




q(a), p(b) = (A, 0).

In the affirmative case, u is a critical point of Lα and we have (q, p) = (q u,α , pu,α ).
Let us note that the fractional Pontryagin’s system (PSα ) is made up of the fractional
Hamiltonian system (HSα ), the fractional stationary equation (SEα ) and initial and final
conditions. In practice, see Examples in Section 1.6, we use more Corollary 7 than Theo-
rem 6. Let us remind that Corollary 7 was already provided in [2, 3, 28, 29, 41] and ref-
erences therein without Condition (fx lip). However, this result is proved, in each of these
papers, using a Lagrange multiplier technique requiring the introduction of an augmented
functional.
As we have seen in this section, fractional Pontryagin’s systems emerge from the study
of a class of fractional optimal control problems. They have a variational structure in the
sense that they are obtained with a calculus of variations on functionals and there resolutions
give explicitly the critical points of these functionals. In Section 2, our aim will be to
provide them numerical schemes preserving this strong characteristic at the discrete level.
Moreover, let us make the following important remark: since a fractional Pontryagin’s
system emerges from a fractional optimal control problem, the main unknown is then the
control u. Consequently, the convergence of the numerical scheme constructed in Section
2 is going to be considered only with respect to u.
1.4. Remarks
1.4.1. The Classical Case
In the case α = 1, the fractional derivatives coincide with the classical one. Consequently,
the fractional optimal control problem studied coincides with the classical one. Then, in
this case, Corollary 7 is nothing else but the classical theorem obtained in [10, 25, 34, 35]:
Theorem 8. L1 has a critical point in C 0 ([a, b], Rm ) if and only if there exists (q, u, p) ∈
C 1 ([a, b], Rd ) × C 0 ([a, b], Rm ) × C 1 ([a, b], Rd ) solution of the following Pontryagin’s sys-
138 Loı̈c Bourdin
tem:
∂H


 q̇ = (q, u, p, t)



 ∂w
 ṗ = − ∂H (q, u, p, t)



∂x (PS1 )
∂H


(q, u, p, t) = 0


∂v





q(a), p(b) = (A, 0).

In the affirmative case, u is a critical point of L1 and we have (q, p) = (q u,1 , pu,1 ).
1.4.2. Link with the Fractional Euler-Lagrange Equation

Let us take the constraint function f (x, v, t) = v satisfying (fx lip). In this case, the frac-
tional optimal control problem studied is:
Z b α
c D− q = u
optimize L(q, u, t) dt under the constraint (34)
a q(a) = A.
Using Corollary 7, if this problem has a solution, then there exists a solution (q, u, p) ∈
C [α] ([a, b], Rd ) × C 0 ([a, b], Rm ) × C [α] ([a, b], Rd ) of the fractional Pontryagin’s system
(PSα ) here given by:  α
 c D− q = u



α p = ∂L


 D+ (q, u, t)


∂x (35)
 ∂L
(q, u, t) + p = 0






 ∂v

q(a), p(b) = (A, 0).

In the case of the existence of an optimal control u for problem (34), we obtain that the state
variable associated q u,α is solution of the following fractional Euler-Lagrange equation:

∂L α α ∂L α
(q, c D− q, t) + D+ (q, c D− q, t) = 0. (ELα )
∂x ∂v
According to the works of Agrawal in [1], we then obtain that q u,α is a critical point of the
following fractional Lagrangian functional:
Z b
α
q −→ L(q, c D− q, t) dt. (36)
a
We refer to [1] for more details concerning fractional Euler-Lagrange equations.
1.5. A Transition from the Classical to the Fractional Problem

As seen in Introduction, there is a large development in fractional calculus and consequently
concerning optimal control problems with controlled systems coded by fractional differen-
tial equations. In this section, we present a link between a classical optimal control problem
and its fractional version via the Stanislavsky’s formalism, see [40, 56]. Only for this sec-
tion, we assume that [a, b] = [0, τ ] with τ > 0 and 0 < α < 1.
1.5.1. The Stanislavsky’s Formalism

The governing equation for a fluid in a homogeneous porous medium is usually the classi-
cal Richards equation and it is derived from physical principles, see [53]. Nevertheless, the
fractional Richards equation plays an important role in the study of the behaviour of a fluid
in a heterogeneous medium, [33, 59, 60]. A natural question then arises: how physically
understand the emergence of the fractional order when the medium changes? Let us sketch
an example of answer. A heterogeneous porous medium presents a lot of stochastic het-
erogeneities implying complex geometric structure and then inducing stochastic retention
zones, [59]. Hence, we can assume that a change of medium induces a modification of the
time variable. Therefore, following the work of Stanislavsky consisting in the introduction
of a stochastic internal time (a ”slow” time), we remind that the emergence of a fractional
derivative can be the effect of a change of time. Let us give more details. However, we refer
to [40, 56] for a complete study.
Let us consider Tt a stochastic process of probability density function ρ(y, t). Tt repre-
sents the stochastic internal time (the ”slow” time). Stanislavsky assumes that the Laplace
transform Lap(y) of ρ(y, t) with respect to its first variable satisfies:
∀t ≥ 0, ∀s ∈ R, Lap(y) [ρ(y, t)](s) = Eα,1 (−stα ), (37)
where Eα,1 is the Mittag-Leffler function defined in Appendix A.1. A possible construction
of Tt is given in [56].
Stanislavsky studies the dynamical effects of this change of time. Precisely, for any
g ∈ C 0 ([0, τ ], Rd ) and any t ∈ [0, τ ], he studies the function Fα (g) defined by:

Fα (g)(t) = E g(Tt ) , (38)
where E designates the mean value. The main property is the following result proved in
[40, 56]:
Lemma 9. Let g ∈ C 1 ([0, τ ], Rd ). Then, we have:

α

∀t ∈ [0, τ ], E ġ(Tt ) = c D− Fα (g) (t). (39)
Hence, a classical derivative is transformed into a fractional one under the introduction
of the stochastic internal time of Stanislavsky.
1.5.2. Application on Fractional Optimal Control Problems

Now, let us talk about the consequences of the works of Stanislavsky on the class of frac-
tional optimal control problems studied in this chapter. We use the notations and definitions
of Section 1.5.1.
Let us assume that L and f are autonomous, i.e. L(x, v, t) = L(x, v) and f (x, v, t) =
f (x, v). Then, let us assume that f satisfies the following condition:
∀(q, u) ∈ C 1 ([0, τ ], Rd ) × C 0 ([0, τ ], Rm ), ∀t ∈ [0, τ ],

E f q(Tt ), u(Tt ) = f E q(Tt ) , E u(Tt ) . (40)
140 Loı̈c Bourdin
Such a condition is satisfied by linear autonomous constraint functions, as the examples

in Section 1.6. According to Lemma 9, one can prove that for any control u, Fα (q u,1 )
satisfies:
α u,1 ) = f F (q u,1 ), F (u)

c D− Fα (q α α
(41)
Fα (q u,1 )(a) = A.
Precisely, we have for any control u, Fα (q u,1 ) = q Fα (u),α .

Now, let us consider the optimal control problem studied in Section 1.2 in the classical
case:
Z τ
0 d
optimize the cost functional u ∈ C ([a, b], R ) 7−→ L(q u,1 , u, t) dt. (42)
0
Let us assume that the controlled system can be prone to changes inducing a modification
of the time variable as it is the case with porous media and the Richards equation. We are
then interested in the optimization of the following ”slow” cost functional:
Z τ
0 d
L Fα (q u,1 ), Fα (u) dt

u ∈ C ([a, b], R ) 7−→ (43)
0
which is relied to the following fractional cost functional:

Z τ
0 d
u ∈ C ([a, b], R ) 7−→ L(q u,α , u) dt (44)
0
as its restriction to the image of Fα . Hence, once an optimal control of the functional (44)
found, one can be interested in its projection on the image of Fα in order to approach an
optimal control of the functional (43).
Hence, the Stanislavsky’s formalism is an example relying a classical optimal control
problem to its fractional version via the introduction of a stochastic internal time.
1.6. Examples
In this section, we are going to study some examples of optimal control problems studied
in Section 1.3 both in classical and fractional cases.
1.6.1. An Applied Classical Linear-Quadratic Example

Classical linear-quadratic examples are often studied in the literature because they are used
for tracking problems. The aim of these problems is to determine a control allowing to
approach as much as possible reference trajectories, [57, Part 1.4, p.49]. In this section,
we study such an example, [24, Part 4.4.3, example 3, p.53]. More generally, a quadratic
Lagrangian is often natural (for example in order to minimize distances) and even if the
differential equations are frequently non linear, we are often leaded to study linearised ver-
sions.
In this section, we take α = d = m = 1. We consider an evolution problem of two
nondescript populations z and ξ interacting during a time interval [a, b]. We control the
injection or the discharge of the population ξ in the system: we then control the value of
the population ξ in real time. The interaction between z and ξ is governed by the following
linear differential equation:
ż = z + ξ. (45)
For any control ξ and any real z0 , we denote by zz0 ,ξ the unique solution of (45) associated
to ξ and satisfying the initial condition z(a) = z0 ∈ R. Let us assume that, for an initial
condition z(a) = X ∈ R, we know experimentally a satisfactory control ξX such that the
evolution of the population zX,ξX is healthy with respect to a nondescript constraint.
The problem is finally the following: assume that the initial condition is modified (i.e.
z(a) = Y 6= X), what is the control ξ minimizing the difference between ξ and ξX plus the
difference between zY,ξ and zX,ξX ? More precisely, we are looking for a minimizer of the
following functional:
1 b
Z
0
ξ ∈ C ([a, b], R) 7−→ (ξ − ξX )2 + (zY,ξ − zX,ξX )2 dt. (46)
2 a
With a change of variable A = Y − X, u = ξ − ξX , q = zY,ξ − zX,ξX and giving the
following quadratic Lagrangian and the following linear constraint function:
L : R2 × [a, b] −→ R and f : R2 × [a, b] −→ R (47)
(x, v, t) 7−→ (x2 + v 2 )/2 (x, v, t) 7−→ x + v,
the problem is reduced to find an optimal control u for the cost functional L1 associated to
L, f and A. See Section 1.3 for notations and definitions.
According to Theorem 8, we are interested in solving the Pontryagin’s system (PS1 )
here given by: 

 q̇ = q + u
ṗ = −q − p

(48)

 p+u=0
q(a), p(b) = (A, 0).

Hence, with p + u = 0, it is sufficient to solve the following linear Cauchy problem:


 q̇ = q + u
u̇ = q − u (49)
q(a), u(b) = (A, 0).

Using matrix exponentials, one can prove that it exists an unique solution (q, u) to Cauchy
problem (49) given by:
 h √ 1−R √ i

 q(t) = A cosh 2(t − a) + √ sinh 2(t − a)
2

∀t ∈ [a, b], h1 + R √ √ i (50)
 u(t) = A √ sinh 2(t − a) − R cosh 2(t − a)


2
where √
sinh 2(b − a)
R= √ √ √ . (51)
2 cosh 2(b − a) − sinh 2(b − a)
Finally, we conclude that u is the unique critical point of L1 . Hence, if there exists ξ
minimizing (46), then:
∀t ∈ [a, b], ξ(t) = ξX (t) + u(t). (52)
142 Loı̈c Bourdin
1.6.2. The Fractional Linear Quadratic Example

In this section, we consider the previous example of Section 1.6.1 in the strict fractional
case 0 < α < 1. Let us consider d = m = 1 and the couple (L, f ) given in Equation (47).
We want to express a critical point u of the cost functional Lα associated to L and f . See
Section 1.3 for notations and definitions.
Therefore, according to Corollary 7, we look for (q, u, p) solution of the fractional
Pontryagin’s system (PSα ) here given by:
α

cD q = q + u
 α−


D+ p = q + p
(53)

 p +u=0
q(a), p(b) = (A, 0).

Hence, from u + p = 0, it is sufficient to solve the following linear fractional Cauchy

problem:
αq = q + u

 c D−
α
D u = u − q (54)
 +
q(a), u(b) = (A, 0).
Although the fractional Cauchy problem (54) is linear, the presence of the left and the
right fractional derivatives is a main drawback to the explicit computation of a solution.
More generally, this characteristic implies very big difficulties in order to solve the most of
fractional Hamiltonian systems. Finally, the resolution of (PSα ) is the most of time still an
opened problem.
Nevertheless, in this example, we can hope getting informations on the solutions of
(PSα ) in dimension d = m = 2 with the help of the fractional Noether’s Theorem 14
proved in Section 1.7. Indeed, in this example, the fractional Pontryagin’s system (PSα )
admits a symmetry and a constant of motion can be obtained. We refer to Section 1.7 for
more details.
Let us also remind that we are going to make some numerical tests with the variational
integrator constructed in Section 2. This example will be treated and graphic representations
are going to be provided, see Section 2.4.2. Moreover, in dimension d = m = 2, we also
can apply the discrete fractional Noether’s Theorem 26 proved in Section 2.5. We refer to
Section 2.5 for more details.
1.6.3. A Solved Fractional Example

In this section, we give a solved fractional example in the sense that we give explicitly the
unique critical point of the cost functional Lα . Nevertheless, the fractional Pontryagin’s
system (PSα ) is still not completely resolved.
Let us take d = m = 1, [a, b] = [0, 1] and 0 < α ≤ 1. Let us consider the following
Lagrangian L and the following linear constraint function f :
L: R2 × [0, 1] −→ R and f: R2 × [0, 1] −→ R
(x, v, t) 7−→ (v 2 /2) + γ(1 − t)β x (x, v, t) 7−→ λx + µv,
(55)
where β, γ, µ, λ are parameters in R∗ . We are looking for a critical point u of Lα associated
to L and f . See Section 1.3 for notations and definitions.
According to Corollary 7, we have to solve the fractional Pontryagin’s system (PSα )

here given by:
α

c D q = λq + µu
 α−

D+ p = λp + γ(1 − t)β

(56)
 u + µp = 0

q(0), p(1) = (A, 0).

From Theorem 4, the Cauchy problem c D− α q = λq + µu with q(0) = A admits an unique

0
solution for any u ∈ C ([0, 1], R). Then, from u + µp = 0, it is sufficient to solve the
α
D+ p = λp + γ(1 − t)β
α
D− p0 = λp0 + γtβ
equivalent to (57)
p(1) = 0 p0 (0) = 0,
with the change of unknown p0 (t) = p(1 − t) for any t ∈ [0, 1]. The unique solution of this
last fractional Cauchy problem is given in [42, Chap.3, p.137] by:
Z t
(t − y)α−1 Eα,α λ(t − y)α γy β dy,

∀t ∈ [0, 1], p0 (t) = (58)
0
where Eα,α is the Mittag-Leffler function defined in Appendix A.1. In order to get a better
formulation, we make a change of variable which gives us:
Z t
∀t ∈ [0, 1], p0 (t) = γ (t − y)β y α−1 Eα,α (λy α )dy (59)
0
β+1 α−1
Eα,α (λy α ) (t)

= γΓ(β + 1)I− y (60)
= γΓ(β + 1)tα+β Eα,α+β+1 (λtα ). (61)
We refer to [42] for more details concerning the calculations. Finally, we obtain the unique
critical point u of Lα given by:
∀t ∈ [0, 1], u(t) = −µγΓ(β + 1)(1 − t)α+β Eα,α+β+1 λ(1 − t)α .

(62)
Although we obtain the unique critical point u of Lα , let us note that this example does
not provide a completely solved fractional Pontryagin’s system: the state variable q u,α
is still unknown. However, this example allows to test the quality of numerical schemes
giving approximations of critical points of cost functionals Lα . This will be done for the
variational integrator constructed in Section 2, see Section 2.4.
1.7. A Fractional Noether’s Theorem

As we have seen with concrete examples in Sections 1.6.2 and 1.6.3, fractional Pontryagin’s
systems (PSα ) are the most of time not resolvable explicitly. This is a strong obstruction in
order to express explicitly a critical point of a cost functional. In this section, we suggest a
deviously way in order to get informations on the solutions of (PSα ) and consequently on
the critical points of the cost functional Lα associated. Precisely, we are interested in the
existence of conservation laws for fractional Pontryagin’s systems admitting a symmetry
(see Definition 11).
144 Loı̈c Bourdin
In 1918, Noether proved the existence of an explicit conservation law for any classical
Euler-Lagrange equation admitting a symmetry. We refer to [6, 43, 50] for more details.
Adapting her strategy to the fractional case, Cresson, Torres and Frederico proved in 2007
a preliminary result giving a conservation law for any fractional Euler-Lagrange equation
admitting a symmetry, see [18, 27, 30]. Nevertheless, the conservation law obtained was
not explicit but only given implicitly via a functional relation. From this first result, we
have formulated in [12] a fractional Noether’s theorem providing an explicit formulation of
this conservation law via a transfer formula.
At the same time, in [28, 29], Torres and Frederico applied a similar strategy for frac-
tional Pontryagin’s system admitting a symmetry. Nevertheless, this conservation law is
also given implicitly via a functional relation. In this section, we are going to formulate a
fractional Noether’s theorem providing an explicit formulation of this conservation law via
an other transfer formula.
We first review the definition of a one parameter group of diffeomorphisms:
Definition 10. Let n ∈ N∗ . For any real s, let φ(s, ·) : Rn −→ Rn be a diffeomor-
phism. Then, Φ = {φ(s, ·)}s∈R is a one parameter group of diffeomorphisms of Rn if it
satisfies:
1. φ(0, ·) = IdRn ;
2. ∀s, s′ ∈ R, φ(s, ·) ◦ φ(s′ , ·) = φ(s + s′ , ·);
3. φ is of class C 2 .
Usual examples of one parameter groups of diffeomorphisms are given by translations
and rotations. The action of three one parameter groups of diffeomorphisms on an Hamil-
tonian allows to define the notion of a symmetry for a fractional Pontryagin’s system:
Definition 11. For i = 1, 2, 3, let Φi = {φi (s, ·)}s∈R be a one parameter group of dif-
feomorphisms of Rd , Rm and Rd respectively. Let L be a Lagrangian, f be a constraint
function and H be the associated Hamiltonian. H is said to be c D− α -invariant under the
α
action of (Φi )i=1,2,3 if it satisfies for any (q, u, p) solution of (PS ) and any s ∈ R:

α α

H φ1 (s, q), φ2 (s, u), φ3 (s, p), t − φ3 (s, p) · c D− φ1 (s, q) = H(q, u, p, t) − p · c D− q. (63)
From this notion, Torres and Frederico proved in [28, 29] the following result:
Lemma 12. Let L be a Lagrangian, f be a constraint function and H be the associated
Hamiltonian. Let us assume that H is c D− α -invariant under the action of three one pa-
rameter groups of diffeomorphisms (Φi )i=1,2,3 . Then, the following equality holds for any
solution (q, u, p) of (PSα ):

α ∂φ1 ∂φ1 α
D
c − (0, q) · p − (0, q) · D+ p = 0. (64)
∂s ∂s
Then, from this result, Torres and Frederico defined a notion of fractional-conserved
quantity. Nevertheless, this result did not provide exactly a constant of motion. In this
section, using a transfer formula, we are going to write the left term of Equation (64) as an
explicit classical derivative and then we obtain a real constant of motion. Let us provide
this transfer formula:
Lemma 13 (Transfer formula). Let g1 , g2 ∈ C ∞ ([a, b], Rd ) satisfying the following condi-
tion (C):
p−α (p)
the sequences of functions I− g1 − g1 (a) · g2 p∈N∗ and
(p) p−α
(g1 · I+ g2 )p∈N∗ converge uniformly to 0 on [a, b].
Then, the following equality holds:
"∞ #
α α d X r r+1−α
(r) (r) r+1−α
c D− g 1 · g2 − g1 · D+ g2 = (−1) I− g1 − g1 (a) · g2 + g1 · I+ g2 . (65)
dt r=0
Proof. In [12], under a similar condition than Condition (C), we have proved:
"∞ # "∞ #
α d X r r+1−α (r) α d X (r) r+1−α
D− g1 · g2 = (−1) I− g1 · g2 and − g1 · D+ g2 = g · I+ g2 . (66)
dt r=0 dt r=0 1
α g = D α g − g (a) , the proof is completed.

Consequently, writing c D− 1 − 1 1
A discussion is provided in [12] concerning the condition (C): one can prove that this
condition is satisfied for any couple of analytic functions for example. Then, combining
Lemmas 12 and 13, we prove:
Theorem 14 (A fractional Noether’s theorem). Let L be a Lagrangian, f be a constraint
function and H be the associated Hamiltonian. Let us assume that H is c D− α -invariant
under the action of three one parameter groups of diffeomorphisms (Φi )i=1,2,3 . Let (q, u, p)
be a solution of (PSα ) and let g denote ∂φ1 /∂s(0, q). If g and p satisfy Condition (C), then
the following equality holds:
"∞ #
d X r+1−α r+1−α
(−1)r I− g − g(a) · p(r) + g (r) · I+

p = 0. (67)
dt
r=0
This theorem then provides an explicit algorithmic way to compute a constant of motion
for any fractional Pontryagin’s systems admitting a symmetry. Nevertheless, the conserva-
tion law is only given as a series of functions: in most cases, it is not easily computable.
However, an arbitrary closed approximation of this quantity can be obtained with a trunca-
tion.
Let us note that the fractional linear-quadratic example developed in Section 1.6.2 gives
a concrete example of fractional Pontryagin’s system admitting a symmetry:
Example 15. Let us consider d = m = 2, the following quadratic Lagrangian and the
following linear constraint function
L: R2 × R2 × [0, 1] −→ R and f: R2 × R2 × [0, 1] −→ R2
(x, v, t) 7−→ (kxk2 + kvk2 )/2 (x, v, t) 7−→ x + v.
(68)
Then, let us consider the three one parameter groups of diffeomorphisms given by the
following rotations:
φi : R × R2 −→ R
2
(69)
cos(sθi ) − sin(sθi ) x1
(s, x1 , x2 ) 7−→ ,
sin(sθi ) cos(sθi ) x2
146 Loı̈c Bourdin
for i = 1, 2, 3 and where θ1 , θ2 ∈ R and θ3 = −θ1 . With these parameters, one can
prove that the Hamiltonian H associated to L and f is c D− α -invariant under the action of
(Φi )i=1,2,3 . Consequently, the fractional Pontryagin’s system given in (53) (in dimension 2)
is not resolvable but admits a symmetry and then admits an explicit conservation law given
by the fractional Noether’s Theorem 14. As said previously, this constant of motion is not
explicitly computable. However, truncating the series, one can provide an approximation
of this quantity.
2. Variational Integrator for Fractional Pontryagin’s Systems

In general, fractional differential equations are very difficult to solve. One can find some
solved examples in [42, 52, 55] using Mittag-Leffler functions, Fourier and Laplace trans-
forms. Additionally, fractional Pontryagin’s systems, as fractional Euler-Lagrange equa-
tions provided in [1], present an asymmetry in the sense that left and right fractional deriva-
tives are involved. It is an additional drawback in order to solve explicitly the most of
fractional Pontryagin’s systems. In this section, we then develop a numerical approach
treating them.
Let us remind that there exist many works concerning the statement of discrete operators
approaching the fractional derivatives (see [23, 31, 51]) and then concerning numerical
schemes for fractional differential equations (see [21, 45, 49, 48]). In particular, one can
find studies concerning the discretization of fractional Euler-Lagrange equations [4, 13] and
fractional Pontryagin’s systems [2, 3, 8, 19, 41].
Nevertheless, as we have seen in Section 1.3, a fractional Pontryagin’s system admits an
intrinsic variational structure: its solutions correspond to the critical points of a functional.
In this chapter, we want to construct a numerical scheme for fractional Pontryagin’s systems
preserving at the discrete level this strong property.
A variational integrator is a numerical scheme preserving the variational structure of
a system at the discrete level. Precisely, let us consider a differential system coming from
a variational principle (i.e. its solutions correspond to the critical points of a functional).
Then, a variational integrator is the numerical scheme constructed as follows:
• firstly, one have to define a discrete version of the functional;
• secondly one have to form a discrete variational principle on it.
Hence, a numerical scheme is obtained and it is called variational integrator. It preserves

the variational structure at the discrete level in the sense that its discrete solutions corre-
spond to the discrete critical points of the discrete functional. Let us remind that variational
integrators are well-developed for classical Euler-Lagrange equations in [36, 46] and let us
remind that we have developed a variational integrator for fractional Euler-Lagrange equa-
tions in [13]. In this section, we are going to construct a variational integrator for fractional
Pontryagin’s systems.
Let us introduce the following notations available in the whole chapter. Let N ∈ N∗ ,
h = (b − a)/N denote the step size of discretization and T = (tk )k=0,...,N = (a +
kh)k=0,...,N be the classical partition of the interval [a, b]. Let us assume that N is suffi-
ciently large in order to satisfy the following condition:
2hα M < 1, (cond h)
where M is the Lipschitz coefficient of the constraint function f , see Condition (fx lip).
2.1. Definitions and Results about Discrete Fractional Calculus

2.1.1. Reminder about Discrete Fractional Derivatives of Grünwald-Letnikov
For the sequel, we need the introduction of discrete operators approximating the fractional
derivatives of Riemann-Liouville and Caputo. As in [19, 23], let us define ∆α− and ∆α+ the
following discrete analogous of D−α and D α respectively:
+
∆α− : (Rd )N +1 −→ (Rd )N (70)

k
!
1 X
G 7−→ αr Gk−r
hα
r=0 k=1,...,N
and
∆α+ : (Rd )N +1 −→ (Rd )N (71)
N −k
!
1 X
G 7−→ αr Gk+r ,
hα
r=0 k=0,...,N −1
where the elements (αr )r∈N are defined by α0 := 1 and
(−α)(1 − α) . . . (r − 1 − α)
∀r ∈ N∗ , αr := . (72)
r!
These discrete fractional operators are approximations of the continuous ones. Indeed,
passing to the limit h → 0, these discrete operators correspond to the definition of the frac-
tional derivatives of Grünwald-Letnikov (1867) coinciding with the Riemann-Liouville’s
ones. We refer to [52] for more details.
Finally, according to Equation (4), we define c ∆α− and c ∆α+ the following discrete anal-
ogous of c D−α and D α respectively:
c +
α : (Rd )N +1 −→ (Rd N
c ∆− ) (73)

G 7−→ ∆α− (G − G0 ) k
k=1,...,N
and
α : (Rd )N +1 −→ (Rd N
c ∆+ ) (74)

G 7−→ ∆α+ (G − GN ) k .
k=0,...,N −1
Let us note that we preserve some continuous properties at the discrete level. In particu-
lar, G0 = 0 (resp. GN = 0) implies c ∆α− G = ∆α− G (resp. c ∆α+ G = ∆α+ G). Additionally,
148 Loı̈c Bourdin
in the classical case α = 1, these discrete fractional derivatives coincide with the usual
backward and forward Euler’s approximations of d/dt with a (−1) term in the right case:
Gk − Gk−1
∀k = 1, . . . , N, (∆1− G)k = (c ∆1− G)k = (75)
h
and
Gk − Gk+1
∀k = 0, . . . , N − 1, (∆1+ G)k = (c ∆1+ G)k = . (76)
h
2.1.2. Results Concerning the Discrete Fractional Derivatives

In this section, we prove two important properties preserved from the continuous level to the
discrete one. For the sequel, we first need the introduction of the following shift operators:
σ: (Rn )N +1 −→ (Rn )N and σ −1 : (Rn )N +1 −→ (Rn )N
G 7−→ Gk+1 k=0,...,N −1 G 7−→ Gk−1 k=1,...,N ,
(77)
where the integer n is d or m.
The first property is the following: considering the quadrature formula of Gauss as
approximation of the integral, we can prove the following discrete fractional integration by
parts:
Property 16 (Discrete fractional integration by parts). Let G1 , G2 ∈ (Rd )N +1 satisfying
G10 = G2N = 0, then we have:
N
X N
X −1
h (c ∆α− G1 )k · σ −1 (G2 )k = h σ(G1 )k · (c ∆α+ G2 )k . (DFIBP)
k=1 k=0
Proof. Since G10 = G2N = 0, we have c ∆α− G1 = ∆α− G1 and c ∆α+ G2 = ∆α+ G2 . Then, we
have:
N
X N
X −1 N
X k+1
−1 X
h (∆α− G1 )k ·σ −1 2
(G )k = h (∆α− G1 )k+1 · G2k =h 1−α
αr G1k+1−r · G2k .
k=1 k=0 k=0 r=0
(78)
Finally, since G10 = G2N = 0, the following equalities hold:
N
X N
X −1 X
k
h (∆α− G1 )k · σ −1 (G2 )k = h1−α αr G1k+1−r · G2k
k=1 k=0 r=0
N
X −1 N
X −1 N
X −1 NX
−r−1
= h1−α α 1
r Gk+1−r · G2k = h1−α αr G1k+1 · G2k+r
r=0 k=r r=0 k=0
N −1 NX
−k−1 N −1 N −k
! !
X X X
= h1−α G1k+1 · 2
αr Gk+r = h1−α 1
Gk+1 · 2
αr Gk+r ,
k=0 r=0 k=0 r=0
(79)
This last result is very useful for discrete calculus of variations involving discrete frac-
tional derivatives, see proof of Theorem 19. Secondly let us prove the following discrete
version of the fractional Cauchy-Lipschitz Theorem 14 proved in Section 1.1.2:
Theorem 17 (Discrete fractional Cauchy-Lipschitz theorem). Let F ∈ C 0 (Rd × [a, b], Rd )

satisfying the following Lipschitz type condition:
∃K ∈ R, ∀(x1 , x2 , t) ∈ (Rd )2 × [a, b], kF (x1 , t) − F (x2 , t)k ≤ Kkx1 − x2 k, (80)
with hα K < 1. Then, the following discrete fractional Cauchy problem:

α

c ∆− Q = F (Q, T )
(81)
Q0 = A
has an unique solution Q ∈ (Rd )N +1 .
Proof. We are going to construct by induction the solution Q of (81). Our method uses
the classical fix point theorem concerning the contraction mappings. Indeed, let us choose
Q0 = A. Then, for any k = 1, . . . , N , Qk has to satisfy:
k−1
X
Qk = hα F (Qk , tk ) + Q0 − αr (Qk−r − Q0 ). (82)
r=1
However, for any k = 1, . . . , N , the application hα F (·, tk ) + Q0 − k−1

P
r=1 αr (Qk−r − Q0 )
is a contraction and consequently admits an unique fix point. Hence, we first construct
Q1 , then Q2 , etc. By induction, we construct a solution Q of (81) and such a construction
assures its uniqueness.
2.2. Construction of the Variational Integrator

2.2.1. First Step of Construction
As said in introduction of this section, in order to complete the first step of construction of
a variational integrator, we have to provide a discrete version of Lα . In this way, let us give
the following definition:
• The elements U ∈ (Rm )N +1 are called the discrete controls;
• For any discrete control U , let QU ,α ∈ (Rd )N +1 denote the unique solution of the
following discrete Cauchy problem:
α

c ∆− Q = f (Q, U , T )
(CPαQ )
Q0 = A ∈ Rd .
QU ,α is called the discrete state variable associated to U . Its existence and its
uniqueness are provided by Theorem 17 and Conditions (fx lip) and (cond h);
• Finally, we define the following discrete cost functional:
Lαh : (Rm )N +1 −→ R (83)

N
,α
L(QU
X
U 7−→ h k , Uk , tk ).
k=1
150 Loı̈c Bourdin
Hence, we have provided a discrete version Lαh to the cost functional Lα . Now, the second
step of the construction of the variational integrator is to characterize the discrete critical
points of the discrete cost functional Lαh with the help of a discrete calculus of variations.
Let us make the following remark: such a characterization implies to be a necessary
condition for the existence of an optimizer of the discrete cost functional Lαh . In fact, in this
section, we have defined an actual discrete fractional optimal control problem.
2.2.2. Second Step of Construction

The second step of construction of a variational integrator consists in forming a discrete
variational principle on Lαh . Precisely, we focus on the characterization of its discrete criti-
cal points, i.e. the elements U ∈ (Rm )N +1 satisfying:
Lαh (U + εŪ ) − Lαh (U )

∀Ū ∈ (Rm )N +1 , DLαh (U )(Ū ) := lim = 0. (84)
ε→0 ε
With a discrete calculus of variations, we obtain the following discrete version of Lemma 5
giving explicitly the value of the Gâteaux derivative of Lαh .
Lemma 18. Let U , Ū ∈ (Rm )N +1 . Then, the following equality holds:
N
∂L ∂L U ,α
(QU ,α , Uk , tk )
X
DLαh (U )(Ū ) =h · Q̄k + (Q , Uk , tk ) · Ūk , (85)
∂x k ∂v k
k=1
where Q̄ ∈ (Rd )N +1 is the unique solution of the following linearised discrete fractional
Cauchy problem:
 c ∆α Q̄ = ∂f (QU ,α , U , T ) × Q̄ + ∂f (QU ,α , U , T ) × Ū

−
∂x ∂v (LCPαQ̄ )
Q̄0 = 0.

Proof. See Appendix B.3.
This last result does not lead to a characterization of the critical points of Lαh yet. As
in the continuous case, we then introduce the notion of discrete adjoint variable: for any
discrete control U , let P U ,α ∈ (Rd )N +1 denote the unique solution of the following shifted
discrete Cauchy problem:
∂H

α σ(QU ,α ), σ(U ), P , σ(T )

c ∆+ P =





 ∂x
T

∂L U ,α
∂f U ,α
= σ(Q ), σ(U ), σ(T ) + σ(Q ), σ(U ), σ(T ) ×P
∂x ∂x





PN = 0.

(σCPαP )
P U ,α is called the discrete adjoint variable associated to U . Its existence and its uniqueness
are provided by the analogous of Theorem 17 for right discrete fractional derivative and
by Conditions (fx lip) and (cond h). Let us note that, since PNU ,α = 0, we can write
α U ,α = ∆α P U ,α .
c ∆+ P +
The presence of shift operators in the definition of the discrete adjoint variable is the
consequence of the change of sums in the discrete fractional integration by parts (DFIBP)
(see Property 16). We refer to the proof of Theorem 19 for more details. We also refer to
Remark 21 for a discussion about the presence of the shift operators.
Finally, let us note that for any discrete control U , the couple (QU ,α , P U ,α ) is solution
of the following shifted discrete fractional Hamiltonian system:

∂H
 c ∆α− Q = Q, U , σ −1 (P ), T


∂w
(σHSαh )
∂H
 ∆α+ P =

 σ(Q), σ(U ), P , σ(T ) .
∂x
Finally, the introduction of this last discrete element allows us to prove the following theo-
rem:
Theorem 19. Let U ∈ (Rm )N +1 . Then, U is a discrete critical point of Lαh if and only
if (QU ,α , U , P U ,α ) is solution of the following shifted discrete fractional stationary equa-
tion:
∂H
Q, U , σ −1 (P ), T = 0. (σSEαh )

∂v
Proof. Let U , Ū ∈ (Rm )N +1 . From Lemma 18, we have:
N
" T #
X ∂L U ,α ∂f U ,α U ,α
h−1
DLα
h (U )(Ū )
= (Q , U k , tk ) + (Q , U k , tk ) −1
× σ (P )k · Q̄k
∂x k ∂x k
k=1
N T ! N
X ∂f U ,α −1 U ,α
X ∂L U ,α
− (Qk , Uk , tk ) × σ (P )k · Q̄k + (Q , Uk , tk ) · Ūk . (86)
∂x ∂v k
k=1 k=1
Then:
N −1 N
X X ∂f ,α
h−1 DLα
h (U )(Ū ) = (c ∆α
+P
U ,α
)k ·σ(Q̄)k − (QU
k , U ,
k kt ) × Q̄ k ·σ
−1
(P U ,α )k
∂x
k=0 k=1
N
X ∂L ,α
+ (QU
k , Uk , tk ) · Ūk . (87)
∂v
k=1
From the discrete fractional integration by parts (DFIBP) (see Property 16), we obtain:
N
X ∂f U ,α
h −1
DLαh (U )(Ū ) = (c ∆α− Q̄)k − (Q , Uk , tk ) × Q̄k · σ −1 (P U ,α )k
∂x k
k=1
N
∂L ,α
(QU
X
+ k , Uk , tk ) · Ūk . (88)
∂v
k=1
Since Q̄ is solution of (LCPαQ̄ ), we have:

N N
X ∂f ,α
X ∂L U ,α
h−1 DLα
h (U )(Ū ) = (QU
k , Uk , tk ) × Ūk · σ −1 (P U ,α )k +
(Q , Uk , tk ) · Ūk
∂v ∂v k
k=1 k=1
N T !
X ∂f U ,α −1 U ,α ∂L U ,α
= (Q , Uk , tk ) × σ (P )k + (Q , Uk , tk ) · Ūk .
∂v k ∂v k
k=1
152 Loı̈c Bourdin
Finally:
N
∂H ,α
QU
X
DLαh (U )(Ū ) = h −1 U ,α

∂v k , Uk , σ (P )k , tk · Ūk . (89)
k=1
Finally, from Theorem 19, we obtain the following result leading to the variational
integrator constructed:
Corollary 20. Lαh has a discrete critical point if and only if there exists (Q, U , P ) ∈
(Rd )N +1 × (Rm )N +1 × (Rd )N +1 solution of the following shifted discrete fractional Pon-
tryagin’s system:
∂H

α Q, U , σ −1 (P ), T

c ∆− Q =





 ∂w
 ∆α P = ∂H σ(Q), σ(U ), P , σ(T )



+
∂x (σPSαh )
 ∂H

Q, U , σ −1 (P ), T = 0


 ∂v




(Q0 , PN ) = (A, 0).

In this case, U is a discrete critical point of Lαh and we have (Q, P ) = (QU ,α , P U ,α ).
Let us note that (σPSαh ) is made up of the shifted discrete Hamiltonian system (σHSαh ),
the shifted stationary equation (σSEαh ) and initial and final conditions.
Hence, we have constructed the variational integrator (σPSαh ) for the fractional Pon-
tryagin’s system (PSα ). It is then a numerical scheme for (PSα ) preserving its variational
structure in the sense that the discrete solutions U obtained correspond to the discrete crit-
ical points of the discrete version Lαh of Lα .
Remark 21. Let us note that the variational integrator (σPSαh ) does not correspond with a
direct discretization of (PSα ) as it is done in [19]. There is an emergence of shift operators
caused by the conservation at the discrete level of the variational structure. However, it is
proved that the use of shifted numerical schemes allows to obtain more stability for some
fractional differential equations, see [48, 49].
Remark 22. Let us remind the following remark: since a fractional Pontryagin’s system
emerges from a fractional optimal control problem, the main unknown is then the control
u. Consequently, the convergence of the variational integrator (σPSαh ) is going to be con-
sidered only with respect to u. Let us note that the value of U0 does not take place in
the variational integrator (σPSαh ): it is a free value. Nevertheless, this is totally coherent
with the fact that this value does not take place neither in the definition of Lαh . Hence, in
numerical tests in Section 2.4, the error between an exact solution u of (PSα ) and a nu-
merical solution U obtained with (σPSαh ) is going to be evaluated on ku(tk ) − Uk k for
k ∈ {1, . . . , N } only.
2.3. Link with the Discrete Fractional Euler-Lagrange Equation

Let us take the constraint function f (x, v, t) = v satisfying (fx lip). In this case, applying
Corollary 20, we know that there exists a critical point of Lαh if and only if there exists a
solution (Q, U , P ) ∈ (Rd )N +1 × (Rm )N +1 × (Rd )N +1 of the shifted discrete fractional
Pontryagin’s system (σPSαh ) here given by:
 α

 c ∆− Q = U


 ∆α P = ∂L σ(Q), σ(U ), σ(T )



 +
∂x (90)
 ∂L −1
(Q, U , T ) + σ (P ) = 0






 ∂v
(Q0 , PN ) = (A, 0).

In the affirmative case, it implies that Q is then a discrete solution of the following discrete
fractional Euler-Lagrange equation:

∂L ∂L
(Q, c ∆α− Q, T ) + ∆α+ (Q, c ∆α− Q, T ) = 0. (ELαh )
∂x ∂v
Finally, according to our works in [13], we then obtain that Q is a critical point of the
following discrete fractional Lagrangian functional:
N
X
L Qk , (c ∆α− Q)k , tk .

Q −→ h (91)
k=1
We refer to [13] for more details concerning discrete fractional Euler-Lagrange equations.
2.4. Numerical Tests

In the following numerical tests, according to Remark 22, we are going to give graphic rep-
resentations only of discrete solutions U and the study of the convergence of the variational
integrator (σPSαh ) is only going to be evaluated on the convergence of the discrete control
to the continuous one.
2.4.1. The Linear-Quadratic Example

Linear-quadratic examples are often studied in the literature because they are used for track-
ing problems. The aim of these problems is to determine a control allowing to approach as
much as possible reference trajectories, [57, Part 1.4, p.49]. In this section, we study such
an example, [24, Part 4.4.3, example 3, p.53]. More generally, a quadratic Lagrangian is of-
ten natural (for example in order to minimize distances) and even if the constraint functions
are frequently non linear, we are often leaded to study linearised versions.
Let us choose d = m = A = 1 and [a, b] = [0, 1]. Then, let us take the following
quadratic Lagrangian and linear constraint function:
L : R2 × [0, 1] −→ R and f : R2 × [0, 1] −→ R (92)

(x, v, t) 7−→ (x2 + v 2 )/2 (x, v, t) 7−→ x + v.
154 Loı̈c Bourdin
Discrete solution U
0
−0.5
−1
−1.5
−2 alpha=1
alpha=3/4
alpha=1/2
alpha=1/4
−2.5
0 0.2 0.4 0.6 0.8 1
t
Figure 1.
Let us give the graphic representations of the numerical solutions U given by (σPSαh ) for
N = 500 and for α = 1, 3/4, 1/2, 1/4: We have seen in Sections 1.6.1 and 1.6.2 that the
fractional Pontryagin’s system (PSα ) is explicitly solved only in the classical case α = 1
and we obtained the following unique critical point of L1 :
√ √
cosh( 2) √ sinh( 2) √
∀t ∈ [0, 1], u(t) = sinh( 2t) − cosh( 2t), (93)
R1 R1
√ √ √
where R1 = 2 cosh( 2) − sinh( 2). Hence, we can only test the convergence of the
variational integrator (σPSαh ) for α = 1. We give the following
graphic representing the
logarithm of the error max |u(tk ) − Uk |, k = 1, . . . , N versus the logarithm of h and the
identity function for comparison: In this example with α = 1, the convergence seems then
obtained with order 1. Nevertheless, we do not know the exact solution of (PSα ) in the strict
fractional case 0 < α < 1. Consequently, we can not study the behaviour of the error in
this case.
2.4.2. A Solved Fractional Example

In this section, we are going to compute (σPSαh ) in the framework of the example solved in
the strict fractional case seen in Section 1.6.3. Consequently, for this example, we can test
the convergence of the variational integrator (σPSαh ) for any 0 < α ≤ 1.
Then, let us choose d = m = A = 1 and [a, b] = [0, 1]. Then, let us take the following
Lagrangian and linear constraint function:
L: R2 × [0, 1] −→ R and f: R2 × [0, 1] −→ R (94)

(x, v, t) 7−→ (1 − t)x + (v 2 /2) (x, v, t) 7−→ x + v.
alpha=1
−2.5
−3
−3.5
−4
−4.5
−5
−5.5
−6
−6.5 log(h)
log(error)
−7
−7 −6.5 −6 −5.5 −5 −4.5 −4 −3.5
log(h)
Figure 2.
Let us give the graphic representations of the numerical solutions U given by (σPSαh )
for N = 500 and for α = 1, 3/4, 1/2, 1/4: As we have seen in Section 1.6.3, the fractional
Pontryagin’s system (PSα ) is explicitly solved for any 0 < α ≤ 1 and we obtained the
following unique critical point of Lα :
∀t ∈ [0, 1], u(t) = −(1 − t)α+1 Eα,α+2 (1 − t)α ,

(95)
where Eα,α+2 is the Mittag-Leffler function with parameter (α, α + 2). Let us test the
convergence of the variational integrator (σPSαh ) for any 0 < α ≤ 1. We give the following

graphics representing the logarithm of the error max |u(tk ) − Uk |, k = 1, . . . , N versus
the logarithm of h and the identity function for comparison for α = 1, 3/4, 1/2, 1/4:
For this example, the convergence seems then obtained for any α = 1, 3/4, 1/2, 1/4 and
still with order 1. Hence, the graphics obtained in these Sections 2.4.1 and 2.4.2 make us
confident with respect to the quality of (σPSαh ) both in the classical and strict fractional
cases.
2.5. A Discrete Fractional Noether’s Theorem

In this section, we study the existence of discrete conservation laws for shifted discrete
fractional Pontryagin’s systems (σPSαh ). Precisely, following the same strategy as in the
continuous case, we introduce the notion of discrete symmetry for such systems and prove
a discrete fractional Noether’s theorem providing an explicit computable discrete constant
of motion. Let us note that this work is strongly inspired from our study in [12] where we
have provided a discrete fractional Noether’s theorem for discrete fractional Euler-Lagrange
equations admitting a discrete symmetry.
156 Loı̈c Bourdin
Discrete solution U
0
−0.5
−1
−1.5
−2
−2.5
−3
−3.5
alpha=1
alpha=3/4
−4 alpha=1/2
alpha=1/4
−4.5
0 0.2 0.4 0.6 0.8 1
t
Figure 3.
As in the continuous case, a discrete symmetry for a shifted discrete fractional Pon-
tryagin’s system (σPSαh ) is based on the action of three one parameter groups of diffeomor-
phisms on the Hamiltonian associated:
Definition 23. Let Φi = {φi (s, ·)}s∈R , for i = 1, 2, 3, be three one parameter groups of
diffeomorphisms of Rd , Rm and Rd respectively. Let L be a Lagrangian, f be a constraint
function and H be the associated Hamiltonian. H is said to be c ∆α− -invariant under the
action of (Φi )i=1,2,3 if it satisfies: for any (Q, U , P ) solution of (σPSαh ) and any s ∈ R

H φ1 s, Q , φ2 s, U , φ3 s, σ −1 (P ) , T − φ3 s, σ −1 (P ) · c ∆α− φ1 s, Q

= H Q, U , σ −1 (P ), T − σ −1 (P ) · c ∆α− Q. (96)

From this notion, we prove the following discrete version of Lemma 12:
Lemma 24. Let L be a Lagrangian, f be a constraint function and H be the associated

Hamiltonian. Let us assume that H is c ∆α− -invariant under the action of three one pa-
rameter groups of diffeomorphisms (Φi )i=1,2,3 . Then, the following equality holds for any
solution (Q, U , P ) solution of (σPSαh ):

∂φ1 ∂φ1
(0, Q) · σ −1 (∆α+ P ) − c ∆α− (0, Q) · σ −1 (P ) = 0. (97)
∂s ∂s
Proof. Let us differentiate (96) with respect to s and let us invert the operator c ∆α− and
alpha=1
−3.5
−4
−4.5
−5
−5.5
−6
−6.5
log(h)
log(error)
−7
−7 −6.5 −6 −5.5 −5 −4.5 −4 −3.5
log(h)
alpha=3/4
−3
−3.5
−4
−4.5
−5
−5.5
−6
−6.5
log(h)
log(error)
−7
−7 −6.5 −6 −5.5 −5 −4.5 −4 −3.5
log(h)
Figure 4.
∂/∂s. Taking s = 0, we finally obtain:
∂H ∂φ1 ∂H ∂φ2 ∂H ∂φ3

0, σ −1 (P )

(⋆) · (0, Q) + (⋆) · (0, U ) + (⋆) ·
∂x ∂s ∂v ∂s ∂w ∂s
∂φ3 −1
α −1 α ∂φ1
− 0, σ (P ) · c ∆− Q − σ (P ) · c ∆− (0, Q) = 0, (98)
∂s ∂s
where ⋆ = Q, U , σ −1 (P ), T . Since (Q, U , P ) is solution of (σPSαh ), we obtain (97).

Let us remind that our aim is to provide an explicit discrete constant of motion for
shifted discrete fractional Pontryagin’s systems (σPSαh ) exhibiting a discrete symmetry. Our
158 Loı̈c Bourdin
alpha=1/2
−2.5
−3
−3.5
−4
−4.5
−5
−5.5
−6
−6.5 log(h)
log(error)
−7
−7 −6.5 −6 −5.5 −5 −4.5 −4 −3.5
log(h)
alpha=1/4
−2
−2.5
−3
−3.5
−4
−4.5
−5
−5.5
−6
−6.5 log(h)
log(error)
−7
−7 −6.5 −6 −5.5 −5 −4.5 −4 −3.5
log(h)
Figure 5.
result is based on Lemma 24 and on the following implication:
∀G ∈ RN +1 , ∆1− G = 0 =⇒ ∃c ∈ R, ∀k = 0, . . . , N, Gk = c. (99)
Namely, if the discrete derivative of G vanishes, then G is constant. Consequently, our

aim is to write the left term of (97) as an explicit discrete derivative (i.e. as ∆1− of an
explicit quantity). In this way, we are going to use a discrete transfer formula as it is
done in [12] for discrete fractional Euler-Lagrange equations admitting a discrete symmetry.
Nevertheless, we have first to introduce some square matrices of length (N + 1). First,
B1 := IdN +1 and then, for any r ∈ {2, . . . , N }, the square matrices Br ∈ MN +1 defined
by:
∀i, j = 0, . . . , N, (Br )i,j := δ{1≤i≤N −1} δ{1≤j≤N −r} δ{0≤i−j≤r−1} − δ{j=0} δ{r≤i} ,
(100)
where δ is the Kronecker symbol. Secondly, we define the square matrices Cr ∈ MN +1
by:
∀r = 1, . . . , N, ∀i, j = 0, . . . , N, (Cr )i,j := δ{r≤i} δ{j=0} . (101)
Finally, we define the square matrices Ar ∈ MN +1 by:
∀r = 1, . . . , N, Ar := αr Br + βrα Cr , (102)
Pr
where βrα = k=0 αk . Examples of matrices Ar ∈ MN +1 for N = 5 are given in
Appendix B.4.
Lemma 25 (Discrete transfer formula). Let G1 , G2 ∈ (Rd )N +1 satisfying G2N = 0. Then,
the following equality holds:
N
hX
1 −1
i
G ·σ (∆α+ G2 ) − (c ∆α− G1 ) · σ −1 (G2 ) 1−α
=h ∆1− Ar × G1 · σ r−1 (G2 ) . (103)
r=1
Proof. See Appendix B.4.
Consequently, combining Lemmas 24 and 25, we prove:

Theorem 26 (Discrete fractional Noether’s theorem). Let L be a Lagrangian, f be a con-
straint function and H be the associated Hamiltonian. Let us assume that H is c ∆α− -
invariant under the action of three one parameter groups of diffeomorphisms (Φi )i=1,2,3 .
Then, the following equality holds for any solution (Q, U , P ) of (σPSαh ):
"N #
1
X ∂φ1 r−1
∆− Ar × (0, Q) · σ (P ) = 0. (104)
∂s
r=1
According to Equation (99), this theorem provides a discrete constant of motion for
any shifted discrete fractional Pontryagin’s systems (σPSαh ) exhibiting a discrete symmetry.
Moreover, this discrete conservation law is not only explicit but also computable in a finite
number of steps. Let us see a concrete example:
Example 27. Let us consider d = m = 2, the following quadratic Lagrangian and the
following linear constraint function:
L: R2 × R2 × [a, b] −→ R and f: R2 × R2 × [a, b] −→ R2
(x, v, t) 7−→ (kxk2 + kvk2 )/2 (x, v, t) 7−→ x + v.
(105)
Then, we consider the three one parameter groups of diffeomorphisms given by the follow-
ing rotations:
φi : R × R2 −→ R
2
(106)
cos(sθi ) − sin(sθi ) x1
(s, x1 , x2 ) 7−→ ,
sin(sθi ) cos(sθi ) x2
160 Loı̈c Bourdin
for i = 1, 2, 3 and where θ1 , θ2 ∈ R and θ3 = −θ1 . With these parameters, one can
prove that the Hamiltonian H associated to L and f is c ∆α− -invariant under the action
of (Φi )i=1,2,3 . Consequently, the fractional Pontryagin’s system (σPSαh ) admits a symme-
try and then admits an explicit discrete conservation law given by the discrete fractional
Noether’s Theorem 26.
We choose A = (1, 2), N = 100 and θ1 = θ2 = −θ3 = 1. Let us compute (σPSαh )
for α = 1, 3/4, 1/2, 1/4. Then, we denote Q = (Q1 , Q2 ) and P = (P 1 , P 2 ) the
discrete solutions obtained and we denote G = ∂φ1 /∂s(0, Q) = (−Q2 , Q1 ). We are then
interested in the value of:
XN
Ar × G · σ r−1 (P ) .

(107)
r=1
Let us see the graphics obtained by the computation of (σPSαh ) and by the computation of
the quantity given in Equation (107) for α = 1, 3/4, 1/2, 1/4:
As expected from Theorem 26, we obtain discrete constants of motion for this discrete
fractional Pontryagin’s system (σPSαh ) admitting a discrete symmetry and for any α = 1,
3/4, 1/2, 1/4. In this specific example, the constant obtained is zero.
A. Appendix
A.1. A Fractional Gronwall’s Lemma
Let us recall the definition extracted from [42, 52, 55] of the Mittag-Leffler function Eα1 ,α2
of parameters α1 , α2 ≥ 0:
∞
X tk
∀t ∈ R, Eα1 ,α2 (t) := . (108)
Γ(α1 k + α2 )
k=0
Now, let us give the following Lemma proved in [21, 58]. For the reader’s convenience, we
recall the proof:
Lemma 28 (A fractional Gronwall’s lemma). Let g ∈ C 0 ([a, b], R) and α > 0. Let us
assume that there exist K1 , K2 ≥ 0 such that:
α
∀t ∈ [a, b], 0 ≤ g(t) ≤ K1 I− g(t) + K2 . (109)
Then, g satisfies:
∀t ∈ [a, b], 0 ≤ g(t) ≤ K2 Eα,1 K1 (t − a)α .

(110)
Proof. Using Property 1, Assumption (109) implies by induction that for any n ∈ N∗ and
any t ∈ [a, b]:
n−1 n−1 k
X X K1 (t − a)α
0 ≤ g(t) ≤ K1n I−
nα
g(t) + K2 K1k I−
kα
(1)(t) = K1n I−
nα
g(t) + K2 . (111)
Γ(αk + 1)
k=0 k=0
Since g is continuous on [a, b] and non-negative, g is bounded by 0 and by a constant

K3 ≥ 0. Hence, using the Stirling asymptotic formula, we have for any t ∈ [a, b]:
n
K1 (t − a)α
Z t
n nα K1n nα−1
0 ≤ K1 I− g(t) = (t − y) g(y)dy ≤ K3 −→ 0. (112)
Γ(αn) a Γ(αn + 1) n→+∞
Discrete solutions (Q1,Q2) and (P1,P2), alpha=1

3.5
Q1
Q2
3 P1
P2
2.5
1.5
0.5
0
0 0.2 0.4 0.6 0.8 1
t
Discrete conservation law on the solutions (Q1,Q2) and (P1,P2), alpha=1

1
constant of motion
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
0 0.2 0.4 0.6 0.8 1
t
Figure 6.
Finally, making n tend to ∞ in Inequality (111), the proof is completed.

Let us note that E1,1 is the exponential function. Consequently, in the classical case
α = 1, the fractional Gronwall’s lemma is nothing else but the classical Gronwall’s lemma.
A.2. Result of Stability of Order 1

In this section, we use the notations and definitions given in Section 1.3. We prove the
following Lemma with the help of Lemma 28.
Lemma 29. Let u, ū ∈ C 0 ([a, b], Rm ). Then, there exists a constant C1 ≥ 0 such that:
∀|ε| < 1, ∀t ∈ [a, b], kq u+εū,α (t) − q u,α (t)k ≤ C1 |ε|. (113)
162 Loı̈c Bourdin
Discrete solutions (Q1,Q2) and (P1,P2), alpha=3/4

4
Q1
Q2
3.5
P1
P2
3
2.5
1.5
0.5
0
0 0.2 0.4 0.6 0.8 1
t
Discrete conservation law on the solutions (Q1,Q2) and (P1,P2), alpha=3/4

1
constant of motion
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
0 0.2 0.4 0.6 0.8 1
t
Figure 7.
Proof. From Theorem 4, we have for any |ε| < 1 and any t ∈ [a, b]:
q u,α (t) = A+I−

α
f (q u,α , u, t) (t) and q u+εū,α (t) = A+I−
α
f (q u+εū,α , u+εū, t) (t). (114)

Hence, for any |ε| < 1 and any t ∈ [a, b], we have:
kq u+εū,α (t) − q u,α (t)k = kI−α

f (q u+εū,α , u + εū, t) − f (q u,α , u, t) (t)k (115)

α
≤ I− kf (q u+εū,α , u + εū, t) − f (q u,α , u, t)k (t). (116)

4.5
Q1
4 Q2
P1
P2
3.5
2.5
1.5
0.5
0
0 0.2 0.4 0.6 0.8 1
t

1
constant of motion
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
0 0.2 0.4 0.6 0.8 1
t
Figure 8.
We have for any |ε| < 1 and any y ∈ [a, b]:
kf q u+εū,α (y), u(y) + εū(y), y − f q u,α (y), u(y), y k

≤ kf q u+εū,α (y), u(y) + εū(y), y − f q u,α (y), u(y) + εū(y), y k

+ kf q u,α (y), u(y) + εū(y), y − f q u,α (y), u(y), y k. (117)

From Condition (fx lip) and with a Taylor’s expansion of order 1 with explicit remainder,
164 Loı̈c Bourdin

4
Q1
Q2
3.5
P1
P2
3
2.5
1.5
0.5
0
0 0.2 0.4 0.6 0.8 1
t

1
constant of motion
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
0 0.2 0.4 0.6 0.8 1
t
Figure 9.
the following inequality holds for any |ε| < 1 and any y ∈ [a, b]:
kf q u+εū,α (y), u(y) + εū(y), y − f q u,α (y), u(y), y k

u+εū,α u,α
∂f u,α ε

≤ M kq (y) − q (y)k + |ε| q (y), ξ (y), y × ū(y)
. (118)
∂v
where ξ ε (y) ∈ [u(y), u(y) + εū(y)] ⊂ [−M1 , M1 ]m with M1 ≥ 0 independent of |ε| < 1
since u and ū are continuous on [a, b].
Thus, since ∂f /∂v, q u,α and ū are continuous, there exists M2 ≥ 0 such that for any
|ε| < 1, we have:

∂f u,α ε

∀y ∈ [a, b],
q (y), ξ (y), y × ū(y) ≤ M2
∂v

α ∂f u,α ε

(t) ≤ M3 , (119)
and then, ∀t ∈ [a, b], I− (q , ξ , t) × ū
∂v
where M3 := M2 (b − a)α /Γ(α + 1) is independent of |ε| < 1.

Finally, from Inequalities (116), (118) and (119), we have:

∀|ε| < 1, ∀t ∈ [a, b], kq u+εū,α (t) − q u,α (t)k ≤ M I−
α
kq u+εū,α − q u,α k (t) + M3 |ε|.
(120)
Using the fractional Gronwall’s Lemma 28, we conclude that:
∀|ε| < 1, ∀t ∈ [a, b], kq u+εū,α (t) − q u,α (t)k ≤ M3 |ε|Eα,1 M (t − a)α

(121)
≤ M3 |ε|Eα,1 M (b − a)α . (122)

Defining C1 := M3 Eα,1 M (b − a)α , the proof is completed.

A.3. Result of Stability of Order 2

following Lemma with the help of Lemmas 28 and 29.
Lemma 30. Let u, ū ∈ C 0 ([a, b], Rm ). There exists a constant C ≥ 0 such that:
∀|ε| < 1, ∀t ∈ [a, b], kq u+εū,α (t) − q u,α (t) − εq̄(t)k ≤ Cε2 . (123)
where q̄ is the unique global solution of the following linearised Cauchy’s problem:
 c Dα q̄ = ∂f (q u,α , u, t) × q̄ + ∂f (q u,α , u, t) × ū

−
∂x ∂v (LCPαq̄ )
q̄(a) = 0.

The existence and the uniqueness of q̄ are given by Theorem 4 and Condition (fx lip).
Proof. We proceed in the same manner than for Lemma 29. We have for any |ε| < 1 and
any t ∈ [a, b]:
kq u+εū,α (t) − q u,α (t) − εq̄(t)k

α
u+εū,α u,α ∂f u,α ∂f u,α
≤ I− f (q , u + εū, t) − f (q , u, t) − ε (q , u, t) × q̄ − ε (q , u, t) × ū
(t).
∂x ∂v
(124)
166 Loı̈c Bourdin
With a Taylor’s expansion’s of order 2 with explicit remainder, we have for any |ε| < 1
and any y ∈ [a, b]:
f q u+εū,α (y), u(y) + εū(y), y − f q u,α (y), u(y), y

∂f u,α ∂f u,α
−ε q (y), u(y), y × q̄(y) − ε q (y), u(y), y × ū(y)
∂x ∂v

∂f u,α u+εū,α
(y) − q u,α (y) − εq̄(y)

≤ ∂x q (y), u(y), y × q

1 2 ε ε
u+εū,α u,α
2
+ ∇ f ξ1 (y), ξ2 (y), y q
(y) − q (y), εū(y), 0 , (125)
2
where for any y ∈ [a, b]:
• ξ1ε (y) ∈ [q u,α (y), q u+εū,α (y)] ⊂ [kq u,α k∞ − C1 , kq u,α k∞ + C1 ]d according to
Lemma 29. Then, there exists M1 ≥ 0 such that ξ1ε (y) ∈ [−M1 , M1 ]d for any
|ε| < 1 and any y ∈ [a, b];
• ξ2ε (y) ∈ [u(y), u(y) + εū(y)] ⊂ [−M2 , M2 ]m with M2 ≥ 0 independent of |ε| < 1
and y ∈ [a, b] since u and ū are continuous on [a, b].
Since ∇2 f is continuous on the compact [−M1 , M1 ]d × [−M2 , M2 ]m × [a, b], there exists
M3 ≥ 0 such that for any |ε| < 1 and any y ∈ [a, b], we have:

1 2
∇ f ξ1ε (y), ξ2ε (y), y q u+εū,α (y) − q u,α (y), εū(y), 0)2

2
≤ M3 kq u+εū,α (y) − q u,α (y)k2 + 2kq u+εū,α (y) − q u,α (y)kkεū(y)k + kεū(y)k2 .

(126)
Then, from Lemma 29 and since ū is continuous, there exists M4 ≥ 0 such that:

1 2 ε ε 2
(y) − q (y), εū(y), 0) ≤ M4 ε2 .
u+εū,α u,α

∀|ε| < 1, ∀y ∈ [a, b], ∇ f ξ1 (y), ξ2 (y), y q

2
(127)
Consequently, from Inequality (125) and since f satisfies Condition (fx lip) (implying
k∂f /∂xk∞ ≤ M ), we have for any |ε| < 1 and any y ∈ [a, b]:
f q u+εū,α (y), u(y) + εū(y), y − f q u,α (y), u(y), y

∂f u,α ∂f u,α
−ε q (y), u(y), y × q̄(y) − ε q (y), u(y), y × ū(y)
∂x ∂v
≤ 2M kq u+εū,α (y) − q u,α (y) − εq̄(y)k + M4 ε2 . (128)
Finally, using the previous inequality and Inequality (124), we obtain for any |ε| < 1 and
any t ∈ [a, b]:
kq u+εū,α (t) − q u,α (t) − εq̄(t)k ≤ 2M I−

α
kq u+εū,α − q u,α − εq̄k (t) + M5 ε2 , (129)

where M5 := M4 (b − a)α /Γ(1 + α) is independent of |ε| < 1. Finally, from the fractional
Gronwall’s Lemma 28, we conclude that:
∀|ε| < 1, ∀t ∈ [a, b], kq u+εū,α (t) − q u,α (t) − εq̄(t)k ≤ M5 ε2 Eα,1 2M (b − a)α . (130)

Defining C := M5 Eα,1 2M (b − a)α , the proof is completed.

A.4. Proof of Lemma 5

In this section, we prove Lemma 5 and consequently, we use the notations and definitions
given in Section 1.3.
Let u, ū ∈ C 0 ([a, b], Rm ) and q̄ ∈ C [α] ([a, b], Rd ) the unique global solution of
(LCPαq̄ ). From Lemma 30, we have:
∀|ε| < 1, q u+εū,α = q u,α + εq̄ + hε , (131)
where khε k∞ ≤ Cε2 . In particular, since q̄ is continuous, there exists M1 ≥ 0 such that:
∀|ε| < 1, ∀y ∈ [a, b], [q u,α (y), q u+εū,α (y)] ⊂ [−M1 , M1 ]d . (132)
In the same way, since u and ū are continuous, there exists M2 ≥ 0 such that:
∀|ε| < 1, ∀y ∈ [a, b], [u(y), u(y) + εū(y)] ⊂ [−M2 , M2 ]m . (133)
We have:
Z b
α α
∀|ε| < 1, L (u + εū) − L (u) = L(q u+εū,α , u + εū, t) − L(q u,α , u, t) dt. (134)
a
With a Taylor’s expansion of order 2 with explicit remainder, we have for any |ε| < 1 and
any y ∈ [a, b]:

L q u+εū,α (y), u(y) + εū(y), y − L q u,α (y), u(y), y

∂L u,α ∂L u,α
−ε q (y), u(y), y · q̄(y) − ε q (y), u(y), y · ū(y)

∂x ∂v
∂L 1 2
q u,α (y), u(y), y · hε (y) + ∇2 L ξ1ε (y), ξ2ε (y), y εq̄(y) + hε (y), εū(y), 0 ,

≤
∂x 2
(135)
where for any |ε| < 1 and any y ∈ [a, b], ξ1ε (y) ∈ [q u,α (y), q u+εū,α (y)] ⊂ [−M1 , M1 ]d and
ξ2ε (y) ∈ [u(y), u(y) + εū(y)] ⊂ [−M2 , M2 ]m . Since L is of class C 2 , we obtain easily that
there exists M3 ≥ 0 such that for any |ε| < 1 and any y ∈ [a, b]:

L q u+εū,α (y), u(y) + εū(y), y − L q u,α (y), u(y), y

∂L u,α ∂L u,α
q (y), u(y), y · ū(y) ≤ M3 ε2 . (136)

−ε q (y), u(y), y · q̄(y) − ε

∂x ∂v
168 Loı̈c Bourdin
Consequently, we have for any 0 < |ε| < 1:

L(q u+εū,α , u + εū, t) − L(q u,α , u, t) ∂L u,α

∂L u,α
− (q , u, t) · q̄ − ≤ M3 ε.
(q , u, t) · ū
ε ∂x ∂v ∞
(137)
Hence:
b
Lα (u + εū) − Lα (u) ∂L u,α ∂L u,α
Z
lim = (q , u, t) · q̄ + (q , u, t) · ū dt. (138)
ε→0 ε a ∂x ∂v
B. Appendix B
B.1. Result of Stability of Order 1
In this section, we use the notations and definitions given in Section 2.2. Let us prove the
following discrete version of Lemma 29:
Lemma 31. Let U , Ū ∈ (Rm )N +1 . Then, there exists a constant C1 ≥ 0 such that:
+εŪ ,α ,α
∀|ε| < 1, ∀k = 0, . . . , N, kQU
k − QU
k k ≤ C1 |ε|. (139)
Proof. First, let us prove by induction that:

+εŪ ,α ,α
∀k = 0, . . . , N, ∃Rk ≥ 0, ∀|ε| < 1, kQU
k − QU
k k ≤ Rk |ε|. (140)
+εŪ ,α ,α
For k = 0, let us take R0 = 0 since QU 0 = QU 0 = A for any |ε| < 1. Let
k ∈ {1, . . . , N } and let us assume that the result (140) is satisfied for any n = 0, . . . , k − 1.
Let us prove that the result (140) is then satisfied for n = k. Since QU ,α (resp. QU +εŪ ,α )
is solution of (CPαQ ) associated to U (resp. to U + εŪ ), we have:
k
,α ,α ,α
QU = hα f (QU αr (QU
X
k k , U k , tk ) + A − k−r − A) (141)
r=1
and for any |ε| < 1:

k
+εŪ ,α +εŪ ,α +εŪ ,α
QU = hα f (QU αr (QU
X
k k , Uk + εŪk , tk ) + A − k−r − A). (142)
r=1
Consequently, for any |ε| < 1:
+εŪ ,α ,α U +εŪ ,α ,α
kQU
k − QU α
k k ≤ h kf (Qk , Uk + εŪk , tk ) − f (QU
k , Uk , tk )k
k
+εŪ ,α ,α
|αr |kQU − QU
X
+ k−r k−r k, (143)
r=1
and then, with the induction hypothesis, we have for any |ε| < 1:
+εŪ ,α ,α U +εŪ ,α ,α
kQU
k − QU α
k k ≤ h kf (Qk , Uk + εŪk , tk ) − f (QU
k , Uk + εŪk , tk )k
k
,α ,α
kf (QU f (QU
X
α
+h k , Uk + εŪk , tk ) − k , Uk , tk )k + |αr |Rk−r |ε|. (144)
r=1
Finally, using Condition (fx lip) and a Taylor’s expansion of order 1 with explicit remainder,
we prove:
+εŪ ,α ,α U +εŪ ,α ,α
∀ |ε| < 1, kQU
k − QU α
k k ≤ h M kQk − QU
k k
k
α
∂f U ,α ε X
+ h |ε| (Qk , ξ , tk ) × Ūk +
|αr |Rk−r |ε|, (145)
∂v
r=1
where ξ ε ∈ [Uk , Uk + εŪk ] ⊂ [−M1 , M1 ]m with M1 ≥ 0 independent of |ε| < 1. Hence,

since ∂f /∂v is continuous, we can conclude that there exists M2 ≥ 0 such that:

∂f U ,α ε
∀ |ε| < 1, (Qk , ξ , tk ) × Ūk

≤ M2 . (146)
∂v
Consequently, since hα M < 1 from Condition (cond h), we have:
k
!
+εŪ ,α ,α |ε|
kQU QU
X
∀ |ε| < 1, k − k k ≤ hα M2 + Rk−r |αr | . (147)
1 − hα M
r=1
We then define Rk := (hα M2 + kr=1 Rk−r |αr |)/(1 − hα M ) independent |ε| < 1 which
P
concludes the induction. To complete the proof, we have just to define C1 = max{Rk , k =
0, . . . , N }.
B.2. Result of Stability of Order 2

following discrete version of Lemma 30 with the help of Lemma 31:
Lemma 32. Let U , Ū ∈ (Rm )N +1 . Then, there exists a constant C ≥ 0 such that:
+εŪ ,α ,α
∀|ε| < 1, ∀k = 0, . . . , N, kQU
k − QU
k − εQ̄k k ≤ Cε2 , (148)
where Q̄ is the unique solution of the following linearised discrete fractional Cauchy prob-
lem:
 c ∆α Q̄ = ∂f (QU ,α , U , T ) × Q̄ + ∂f (QU ,α , U , T ) × Ū

−
∂x ∂v (LCPαQ̄ )
Q̄0 = 0.

Its existence and its uniqueness are provided by Theorem 17 and Conditions (fx lip) and
(cond h).
170 Loı̈c Bourdin
Proof. We proceed in the same manner that for Lemma 31. Let us prove by induction that:
+εŪ ,α ,α
∀k = 0, . . . , N, ∃Rk ≥ 0, ∀|ε| < 1, kQU
k − QU
k − εQ̄k k ≤ Rk ε2 . (149)
+εŪ ,α ,α
For k = 0, let us take R0 = 0 since QU 0 = QU 0 = A for any |ε| < 1 and
Q̄0 = 0. Let k ∈ {1, . . . , N } and let us assume that the result (149) is satisfied for any
n = 0, . . . , k − 1. Let us prove that the result (149) is then satisfied for n = k. Since QU ,α
(resp. QU +εŪ ,α ) is solution of (LCPαQ̄ ) associated to U (resp. to U + εŪ ) and since Q̄ is
solution of (LCPαQ̄ ), we have with a Taylor’s expansion of order 2 with explicit remainder:
+εŪ ,α ,α ∂f U ,α +εŪ ,α ,α
∀ |ε| < 1, QU
k − QU
k − εQ̄k = hα (Q , Uk , tk ) × (QU − QU − εQ̄k )
∂x k k k
X k
1 2 +εŪ ,α U ,α +εŪ ,α ,α
+ hα ∇ f (ξ1ε , ξ2ε , tk )(QU
k − Q k , ε Ū k , 0) 2
− αr (QU
k−r − QU
k−r − εQ̄k−r ),
2 r=1
(150)
where:
,α U +εŪ ,α ,α U ,α
• ξ1ε ∈ [QU
k , Qk ] ⊂ [kQU
k k − C1 , kQk k + C1 ] from Lemma 31. Then,
ξ1ε ∈ (−M1 , M1 ]d with M1 ≥ 0 independent of |ε| < 1;
• ξ2ε ∈ [Uk , Uk + εŪk ] ⊂ [−M2 , M2 ]m with M2 ≥ 0 independent of |ε| < 1.
Since ∇2 f (·, ·, tk ) is continuous, we conclude that there exists M3 ≥ 0 such that:

1 2 ε ε U +εŪ ,α U ,α 2

∀ |ε| < 1, ∇ f (ξ1 , ξ2 , tk )(Qk
− Qk , εŪk , 0)
2
+εŪ ,α ,α 2 U +εŪ ,α ,α
≤ M3 (kQU
k − QU
k k + 2kQk − QU 2
k kkεŪk k + kεŪk k ). (151)
Hence, from Lemma 31, there exists M4 ≥ 0 such that:

1 2 ε ε U +εŪ ,α U ,α 2
− Qk , εŪk , 0) ≤ M4 ε2 .

∀ |ε| < 1, ∇ f (ξ1 , ξ2 , tk )(Qk
(152)
2
From Equality (150) and Condition (fx lip), Inequality (152) and the induction hypothesis,
we obtain:
+εŪ ,α ,α +εŪ ,α ,α
∀ |ε| < 1, kQU
k − QU
k − εQ̄k k ≤ 2hα M kQU
k − QU
k − εQ̄k k
k
X
+ |αr |Rk−r ε2 + hα M4 ε2 . (153)
r=1
Finally, since 2hα M < 1 from Condition (cond h), we have:

k
!
+εŪ ,α U ,α ε2
∀ |ε| < 1, kQU
X
k − Q k − ε Q̄ k k ≤ hα M4 + Rk−r |αr | . (154)
1 − 2hα M
r=1
We then define Rk := (hα M4 + kr=1 Rk−r |αr |)/(1 − 2hα M ) which concludes the in-
P
duction. In order to complete the proof, we just have to define C := max{Rk , k =
0, . . . , N }.
B.3. Proof of Lemma 18

In this section, we prove Lemma 18 and consequently, we use notations and definitions
given in Section 2.2.
Let U , Ū ∈ (Rm )N +1 and Q̄ ∈ (Rd )N +1 the unique solution of (LCPαQ̄ ). From Lemma
32, we have:
+εŪ ,α ,α
∀k = 0, . . . , N, ∀|ε| < 1, QU
k = QU
k + εQ̄k + Hkε , (155)
where for any k = 0, . . . , N and for any |ε| < 1, kHkε k ≤ Cε2 . In particular, there exists
M1 ≥ 0 such that:
,α U +εŪ ,α
∀k = 0, . . . , N, ∀|ε| < 1, [QU
k , Qk ] ⊂ [−M1 , M1 ]d . (156)
In the same way, there exists M2 ≥ 0 such that:
∀k = 0, . . . , N, ∀|ε| < 1, [Uk , Uk + εŪk ] ⊂ [−M2 , M2 ]m . (157)
We have:
N h i
+εŪ ,α
X U ,α
∀|ε| < 1, Lα α
h (U +εŪ )−Lh (U ) = h L(QU
k , U k +ε Ū k , t k )−L(Q k , U k , t k ) . (158)
k=1
With a Taylor’s expansion of order 2 with explicit remainder, we have for any |ε| < 1 and
any k = 0, . . . , N :
∂L U ,α ∂L U ,α

U +εŪ ,α ,α
, Uk + εŪk , tk ) − L(QU , U , t ) − ε (Q , U , t ) · Q̄ − ε (Q , U , t ) · Ū

L(Qk k k k k k k k k k
∂x k ∂v k

∂L U ,α 1

2
≤ (Q , Uk , tk ) · Hkε + ∇2 L(ξ1ε , ξ2ε , tk )(εQ̄k + Hkε , εŪk , 0 , (159)
∂x k 2
,α U +εŪ ,α
where ξ1ε ∈ [QU k , Qk ] ⊂ [−M1 , M1 ]d and ξ2ε ∈ [Uk , Uk + εŪk ] ⊂ [−M2 , M2 ]m .
2
Since L is of class C , we obtain easily that there exists M3 ≥ 0 such that for any |ε| < 1
and any k = 0, . . . , N :

U +εŪ ,α ,α
L(Qk , Uk + εŪk , tk ) − L(QU k , U k , tk )

∂L ,α ∂L U ,α
− ε (QU , U , t ) Q̄ ε (Q , U , t ) Ū ≤ M3 ε2 . (160)

k k k · k − k k k · k
∂x ∂v
Consequently, we have for any 0 < |ε| < 1 and any k = 0, . . . , N :

L(QU +εŪ ,α , U + εŪ , t ) − L(QU ,α , U , t )
k k k k k k k
ε

∂L U ,α ∂L U ,α
− (Qk , Uk , tk ) · Q̄k − (Qk , Uk , tk ) · Ūk ≤ M3 ε. (161)
∂x ∂v
Hence:
N
Lα α

h (U + εŪ ) − Lh (U ) ∂L U ,α ∂L U ,α
X
lim =h (Qk , Uk , tk ) · Q̄k + (Qk , Uk , tk ) · Ūk .
ε→0 ε ∂x ∂v
k=1
(162)
172 Loı̈c Bourdin
B.4. Proof of Lemma 25

In this section, we use the notations and definitions given in Section 2.5. Let us prove
Lemma 25.
Let G1 , G2 ∈ (Rd )N +1 satisfying G2N = 0. First, let us denote for any k = 1, . . . , N :
h i
Xk := hα G1k · σ −1 (∆α+ G2 )k − (c ∆α− G1 )k · σ −1 (G2 )k . (163)
Our aim is to write X as an explicit discrete derivative (i.e. as ∆1− of an explicit quantity).
We have for any k = 1, . . . , N :
NX
+1−k
! k
!
X
Xk = G1k · αr G2k+r−1 − αr (G1k−r − G10 ) ·G2k−1 = α1 h(∆1− G1 ·G2 )k +Yk +Zk , (164)
r=0 r=0
where for any k = 1, . . . , N :

k
!
X
Yk := αr G10 · G2k−1 = βkα G10 · G2k−1 (165)
r=0
and
N −k k
" ! ! #
X X
Zk := G1k · αr G2k+r−1 − αr G1k−r · G2k−1 . (166)
r=2 r=2
Our aim is then to write Y and Z as explicit discrete derivatives. We then define for any
i = 0, . . . , N , Vi := h ir=1 Yr and Wi := h ij=1 Zj . Hence, we have ∆1− V = Y and
P P
∆1− W = Z and then, X = ∆1− (α1 hG1 · G2 + V + W ). Our aim is then to explicit V
and W . For any i = 0, . . . , N , we have:
i
X i
X N X
X N
Vi = h βrα G10 · G2r−1 = h βrα G10 · σ r−1 (G2 )0 = h βrα Cr (i, j)G1j · σ r−1 (G2 )j .
r=1 r=1 r=1 j=0
(167)
For any i = 0, . . . , N , we have:
i N −j j
" ! ! #
X X X
Wi = h G1j · αr G2j+r−1 − αr G1j−r · G2j−1 (168)
j=1 r=2 r=2
i N
X X −j j
i X
X
= h αr G1j · G2j+r−1 − h αr G1j−r · G2j−1 (169)
j=1 r=2 j=2 r=2
i N
X X −j i X
X i
= h αr G1j · σ r−1 (G2 )j − h αr G1j−r · G2j−1 (170)
j=1 r=2 r=2 j=r
i N
X X −j i X
X i−r
= h αr G1j · σ r−1 (G2 )j − h αr G1j · σ r−1 (G2 )j . (171)
j=1 r=2 r=2 j=0
The following equality holds for any r = 2, . . . , N and any i, j = 0, . . . , N :
δ{1≤j≤i} δ{2≤r≤N −j} − δ{2≤r≤i} δ{0≤j≤i−r}

= δ{1≤i≤N −1} δ{1≤j≤N −r} δ{0≤i−j≤r−1} − δ{j=0} δ{r≤i} . (172)
Consequently, we have for any i = 0, . . . , N :

N X
X N
Wi = h αr Br (i, j)G1j · σ r−1 (G2 )j . (173)
r=2 j=0
Finally, we have for any i = 0, . . . , N :

N X
X N
α1 hG1i · G2i + Vi + W i = h Ar (i, j)G1j · σ r−1 (G2 )j . (174)
r=1 j=0
Finally, the following equality holds:

N
hX i
X = h∆1− Ar × G1 · σ r−1 (G2 ) , (175)
r=1

Now, let us see some examples of matrices Ar ∈ MN +1 for N = 5:
α1 0 0 0 0 0 0 0 0 0 0 0
   
 β1α α1 0 0 0 0   0 α 2 0 0 0 0 
 α   α 
 β 1 0 α1 0 0 0   β 2 − α2 α2 α2 0 0 0 
A1 =  β α 0 0 α 1 0 0  , A2 =  β α − α 2 0 α 2 α 2
  ,
 1   2 0 0 

 β α 0 0 0 α1 0   β α − α2 0 0 α2 0 0 
1 2
β1α 0 0 0 0 α1 β2α − α2 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0
   

 0 α3 0 0 0 0 


 0 α4 0 0 0 0 

 0 α3 α3 0 0 0   0 α4 0 0 0 0 
A3 =  α
 , A4 =  

 β 3 − α3 α3 α3 0 0 0 


 0 α4 0 0 0 0 

 β3α − α3 0 α3 0 0 0   β4α − α4 α4 0 0 0 0 
β3α − α3 0 0 0 0 0 β4α − α4 0 0 0 0 0
and
0 0 0 0 0 0
 

 0 0 0 0 0 0 

 0 0 0 0 0 0 
.
A5 = 

 0 0 0 0 0 0 

 0 0 0 0 0 0 
α
β 5 − α5 0 0 0 0 0
References
[1] O.P. Agrawal. Formulation of Euler-Lagrange equations for fractional variational
problems. J. Math. Anal. Appl., 272(1):368–379, 2002.
[2] O.P. Agrawal. A general formulation and solution scheme for fractional optimal con-
trol problems. Nonlinear Dynam., 38(1-4):323–337, 2004.
174 Loı̈c Bourdin
[3] O.P. Agrawal. A formulation and numerical scheme for fractional optimal control
problems. J. Vib. Control, 14(9-10):1291–1299, 2008.
[4] O.P. Agrawal. A general finite element formulation for fractional variational problems.
J. Math. Anal. Appl., 337(1):1–12, 2008.
[5] R. Almeida, A.B. Malinowska, and D.F.M. Torres. A fractional calculus of vari-
ations for multiple integrals with application to vibrating string. J. Math. Phys.,
51(3):033503, 12, 2010.
[6] V. I. Arnold. Mathematical methods of classical mechanics, volume 60 of Graduate

Texts in Mathematics. Springer-Verlag, New York, 1979.
[7] R. L. Bagley and R. A. Calico. Fractional order state equations for the control of
viscoelastically damped structures. Journal of Guidance, Control, and Dynamics,
14:304–311, 1991.
[8] D. Baleanu, O. Defterli, and O.P. Agrawal. A central difference numerical scheme for
fractional optimal control problems. J. Vib. Control, 15(4):583–597, 2009.
[9] D. Baleanu and S.I. Muslih. Lagrangian formulation of classical fields within
Riemann-Liouville fractional derivatives. Phys. Scripta, 72(2-3):119–121, 2005.
[10] M. Bergmann. Optimisation aérodynamique par réduction de modèle POD et contrôle

optimal. Application au sillage laminaire d’un cylindre circulaire. PhD thesis, Institut
National Polytechnique de Lorraine, 2004.
[11] B. Bonnard, L. Faubourg, and E. Trélat. Mécanique céleste et contrôle des véhicules
spatiaux, volume 51 of Mathématiques & Applications. Springer-Verlag, Berlin, 2006.
[12] L. Bourdin, J. Cresson, and I. Greff. A continuous/discrete fractional Noether’s theo-

rem. Commun. Nonlinear Sci. Numer. Simul., 71:14–23, 2013.
[13] L. Bourdin, J. Cresson, I. Greff, and P. Inizan. Variational integrator for fractional
Euler-Lagrange equations. Appl. Numer. Math., 18(4):878–887, 2013.
[14] J.A.E. Bryson and Y.C. Ho. Applied optimal control. Hemisphere Publishing Corp.
Washington, D. C., 1975. Optimization, estimation, and control, Revised printing.
[15] R. Burlacu and A. Cavache. On a class of optimal control problems in mathemati-

cal biology. Sci. Ann. Univ. Agric. Sci. Vet. Med., 45:41–54 (2003), 2002. Annual
Symposium on Mathematics Applied in Biology & Biophysics.
[16] L. Cantaluppi. Optimal control of a dam under seasonal electricity prices. European
J. Oper. Res., 17(1):57–70, 1984.
[17] F. Comte. Opérateurs fractionnaires en économétrie et en finance. Prépublication

MAP5, 2001.
[18] J. Cresson. Fractional embedding of differential operators and Lagrangian systems. J.

Math. Phys., 48(3):033504, 34, 2007.
[19] O. Defterli. A numerical scheme for two-dimensional optimal control problems with
memory effect. Comput. Math. Appl., 59(5):1630–1636, 2010.
[20] G. Demange and J.-C. Rochet. Méthodes Mathématiques de la Finance. Economica,

1992.
[21] K. Diethelm. The analysis of fractional differential equations, volume 2004 of Lec-
ture Notes in Mathematics. Springer-Verlag, Berlin, 2010. An application-oriented
exposition using differential operators of Caputo type.
[22] N.M. Dmitruk. Optimization of the von Neumann model of an expanding economy
by optimal control methods. Vestsı̄ Nats. Akad. Navuk Belarusı̄ Ser. Fı̄z.-Mat. Navuk,
(4):48–53, 141, 2000.
[23] F. Dubois, A-C. Galucio, and N. Point. Introduction à la dérivation fractionnaire.

Théorie et applications. Série des Techniques de l’ingénieur, 2009.
[24] L.C. Evans. An introduction to mathematical optimal control theory.
[25] J. Favier. Contrôle d’écoulements : approche expérimentale et modélisation de di-

mension réduite. PhD thesis, Institut National Polytechnique de Toulouse, 2007.
[26] A.V. Fedoseev. Optimal control in a three-sector model of the economy of an under-
developed country. Ž. Vyčisl. Mat. i Mat. Fiz., 12:925–942, 1085, 1972.
[27] G.S.F. Frederico and D.F.M. Torres. A formulation of Noether’s theorem for fractional
problems of the calculus of variations. J. Math. Anal. Appl., 334(2):834–846, 2007.
[28] G.S.F. Frederico and D.F.M. Torres. Fractional conservation laws in optimal control
theory. Nonlinear Dynam., 53(3):215–222, 2008.
[29] G.S.F. Frederico and D.F.M. Torres. Fractional optimal control in the sense of Caputo
and the fractional Noether’s theorem. Int. Math. Forum, 3(9-12):479–493, 2008.
[30] G.S.F. Frederico and D.F.M. Torres. Fractional Noether’s theorem in the Riesz-Caputo
sense. Appl. Math. Comput., 217(3):1023–1033, 2010.
[31] A. C. Galucio, J.-F. Deü, and F. Dubois. The Gα -scheme for approximation of frac-
tional derivatives: application to the dynamics of dissipative systems. J. Vib. Control,
14(9-10):1597–1605, 2008.
[32] R.V. Gamkrelidze. Discovery of the maximum principle. In Mathematical events of

the twentieth century, pages 85–99. Springer, Berlin, 2006.
[33] E. Gerolymatou, I. Vardoulakis, and R. Hilfer. Modelling infiltration by means of a

nonlinear fractional diffusion model. J. Phys. D: Appl. Phys., 39:4104, 2006.
[34] M.D. Gunzburger. Introduction into mathematical aspects of flow control and opti-
mization. Lecture series - Von KÃ¡rmÃ¡n Institute for Fluid Dynamics, 5:A1–A20,
1997.
176 Loı̈c Bourdin
[35] M.D. Gunzburger. Lagrange multiplier techniques. Lecture series - Von KÃ¡rmÃ¡n
Institute for Fluid Dynamics, 5:C1–C18, 1997.
[36] E. Hairer, C. Lubich, and G. Wanner. Geometric numerical integration, volume 31

of Springer Series in Computational Mathematics. Springer-Verlag, Berlin, second
edition, 2006. Structure-preserving algorithms for ordinary differential equations.
[37] M.R. Hestenes. Calculus of variations and optimal control theory. Robert E. Krieger
Publishing Co. Inc., Huntington, N.Y., 1980. Corrected reprint of the 1966 original.
[38] R. Hilfer. Applications of fractional calculus in physics. World Scientific, River Edge,
New Jersey, 2000.
[39] R. Hilfer. Fractional calculus and regular variation in thermodynamics. In Applica-

tions of fractional calculus in physics, pages 429–463. World Sci. Publ., River Edge,
NJ, 2000.
[40] P. Inizan. Dynamique fractionnaire pour le chaos hamiltonien. PhD thesis, Université
de Pau et des Pays de l’Adour, 2010.
[41] Z.D. Jelicic and N. Petrovacki. Optimality conditions and a solution scheme for frac-
tional optimal control problems. Struct. Multidiscip. Optim., 38(6):571–581, 2009.
[42] A.A. Kilbas, H.M. Srivastava, and J.J. Trujillo. Theory and applications of fractional
differential equations, volume 204 of North-Holland Mathematics Studies. Elsevier
Science B.V., Amsterdam, 2006.
[43] Y. Kosmann-Schwarzbach. Les théorèmes de Noether. Histoire des Mathématiques

(Palaiseau). Éditions de l’École Polytechnique, Palaiseau, second edition, 2006. In-
variance et lois de conservation au XXe siècle.
[44] P. Lévy. L’addition des variables aléatoires définies sur une circonférence. Bull. Soc.
Math. France, 67:1–41, 1939.
[45] Yanqin Liu and Baogui Xin. Numerical solutions of a fractional predator-prey system.
Adv. Difference Equ., pages Art. ID 190475, 11, 2011.
[46] J.E. Marsden and M. West. Discrete mechanics and variational integrators. Acta
Numer., 10:357–514, 2001.
[47] F. Martin-Robine. Histoire du principe de moindre action: trois siècles de principes

variationnels de Fermat à Feynman. Histoire des sciences. Vuibert, 2006.
[48] M.M. Meerschaert, H.-P. Scheffler, and C. Tadjeran. Finite difference methods for
two-dimensional fractional dispersion equation. J. Comput. Phys., 211(1):249–261,
2006.
[49] M.M. Meerschaert and C. Tadjeran. Finite difference approximations for fractional
advection-dispersion flow equations. J. Comput. Appl. Math., 172(1):65–77, 2004.
[50] E. Noether. Invariant variation problems. Transport Theory Statist. Phys., 1(3):186–
207, 1971. Translated from the German (Nachr. Akad. Wiss. Göttingen Math.-Phys.
Kl. II 1918, 235–257).
[51] K.B. Oldham and J. Spanier. The fractional calculus. Academic Press [A subsidiary
of Harcourt Brace Jovanovich, Publishers], New York-London, 1974. Theory and
applications of differentiation and integration to arbitrary order, With an annotated
chronological bibliography by Bertram Ross, Mathematics in Science and Engineer-
ing, Vol. 111.
[52] I. Podlubny. Fractional differential equations, volume 198 of Mathematics in Science

and Engineering. Academic Press Inc., San Diego, CA, 1999. An introduction to
fractional derivatives, fractional differential equations, to methods of their solution
and some of their applications.
[53] L.A. Richards. Capillary conduction of liquids through porous mediums. Phys., 1:318,
1931.
[54] F. Riewe. Mechanics with fractional derivatives. Phys. Rev. E (3), 55(3, part B):3581–
3592, 1997.
[55] S.G. Samko, A.A. Kilbas, and O.I. Marichev. Fractional integrals and derivatives.
Gordon and Breach Science Publishers, Yverdon, 1993. Theory and applications,
Translated from the 1987 Russian original.
[56] A. A. Stanislavsky. Hamiltonian formalism of fractional systems. Eur. Phys. J. B

Condens. Matter Phys., 49(1):93–101, 2006.
[57] E. Trélat. Contrôle optimal. Mathématiques Concrètes. Vuibert, Paris, 2005. Théorie
& applications.
[58] H. Ye, J. Gao, and Y. Ding. A generalized Gronwall inequality and its application to
a fractional differential equation. J. Math. Anal. Appl., 328(2):1075–1081, 2007.
[59] A. Zoia, M.-C. Néel, and A. Cortis. Continuous-time random-walk model of transport
in variably saturated heterogeneous porous media. Phys. Rev. E, 81(3):031104, Mar
2010.
[60] A. Zoia, M.-C. Néel, and M. Joelson. Mass transport subject to time-dependent flow
with nonuniform sorption in porous media. Phys. Rev. E, 80:056301, 2009.

Chapter 4
F RACTAL T RAPS AND F RACTIONAL DYNAMICS

Pierre Inizan∗
Institut de Mécanique Céleste et de Calcul des Éphémérides,
Observatoire de Paris, Paris, France
Abstract
Anomalous diffusion may arise in typical chaotic Hamiltonian systems. Accord-
ing to G.M. Zaslavsky’s analysis, this behavior is induced by sticky zones within the
phase space, which may trap trajectories for a long time. A description can be done by
means of fractional kinetics equations. However, the dynamical origin of those frac-
tional derivatives is still unclear. We provide in this article an attempt for a possible
explanation. Starting from R. Hilfer’s work, an expression for the average infinitesimal
evolution of trajectories sets is given by using Poincaré recurrence times. The fractal
structures of the traps, described by G.M. Zaslavsky, are then taken into account, and
it is shown that in this case, the derivative associated to this evolution may become
fractional, with order equal to the transport exponent of the diffusion process.
PACS: 05.45-a, 52.35.Mw, 96.50.Fm

Keywords: Fractional dynamics, Hamiltonian systems, chaotic dynamics
1. Introduction
Fractional calculus [26, 21, 22] is efficiently used in several fields of physics [12, 24]. For
example, it may be used to take into account memory effects and anomalous transport.
Several equations of physics have hence been generalized to non-integer orders so as to
provide new models. Among them, one find the Euler-Lagrange equation [23, 1, 2, 5, 7]
and the diffusion equation [20, 4, 9, 18, 27, 29, 6].
However, reasons for emergence of such operators are still unclear and the use of this
formalism often remains heuristic. R. Hilfer [10, 14, 11] and G.M. Zaslavsky [31, 34,
∗
E-mail address: pierre.inizan@imcce.fr
180 Pierre Inizan
25, 33] have tried through different ways to understand the physical origin of fractional
derivatives. Both of their models rely on the recurrence time notion.
The first of those authors studies the evolution operator of a subsystem and shows
that after a temporal renormalization, the associated infinitesimal generator is a fractional
derivative. However, the interpretation of this operator may seem difficult and the renor-
malization procedure is ambivalent.
Zaslavsky is interested in chaotic Hamiltonian systems. He makes fractional deriva-
tives appear in the diffusion equation related to the kinetic (i.e. probabilistic) description
of the system. Without a true justification for the introduction of those derivatives, he nev-
ertheless connects the transport exponent µ with the fractional orders of derivation and the
coefficients of the self-similar structures which appear in the phase space around resonance
areas.
In the present article, we study the dynamics of an Hamiltonian system, presented in
section 2. With ideas taken from Hilfer, we focus in section 3 on the evolution of a phase
space subset under the Hamiltonian flow. More precisely, the associated infinitesimal gener-
ator is considered. In several examples we show that it is proportional to the usual derivative
d/dt. Then we precise our model by taking into account the phase space structure described
by Zaslavsky and sumed up in section 4. In that case, we prove in section 5 that the infinites-
imal generator may turn into a fractional derivative of order µ. We discuss the relevance of
this exponent in section 6, before concluding in section 7.
2. Studied System
Let H be an Hamiltonian defined on a compact manifold M . The induced flow is denoted
φt . Let m be a measure defined on M . Let G be a measurable subset of M and m′ a
measure adapted to G (such that m′ (G) > 0). Two cases may happen: if m(G) > 0, then
we can choose m′ = m. Conversely if m(G) = 0 (important case in this article), m cannot
measure subsets of G, so m′ must differ from m.
Let us suppose that we only have access to G. Thus we are interested in the dynamics
restricted to G, and we consider the measurable space (G, T ′ , m′ ), where T ′ is a σ-algebra
of G over which m′ is defined. We introduce Gatt , the “attractive” subset of G:
Gatt = x ∈ G | ∃t0 > 0, ∀ t > t0 , φt x ∈ G .

If x ∈ Gatt , after some time it becomes possible to completely follow the trajectory starting
from x. The assumption that we only have access to G is thus invisible concerning the
dynamics on Gatt . Conversely, trajectories starting from G\Gatt leave G and then cannot
be tracked. Fortunately, from Poincaré recurrence theorem, if m(G) > 0, then almost all
trajectories come back into G. More precisely, we may define the Poincaré recurrence time
as
∀x ∈ G\Gatt , τG (x) = inf t > 0 | φt x ∈ G, ∃t0 ∈ (0, t), φt0 x ∈

/G .
We remark that if τG (x) < ∞, then φτG (x) x ∈ ∂G ∩ (G\Gatt ), where ∂G is the boundary
of G. Let Gext be the set of the starting points of trajectories which never come back into
G, i.e. points x such that τG (x) = ∞:
Gext = x ∈ G | ∃t0 > 0, ∀ t > t0 , φt x ∈

/G .
Fractal Traps and Fractional Dynamics 181
Theorem 1 (Poincaré recurrence theorem). The set Gext is negligible: m(Gext ) = 0.
In that case, G is said m-recurrent. Until the end, if m(G) = 0, we suppose that G
is m′ -recurrent. We note Grec the set of trajectories which alternatively wander inside and
outside G:
Grec = G\ (Gatt ∪ Gext ) .
We may remark that if x ∈ Grec , then φτG (x) x ∈ Grec . Although it is impossible to have
a continuous description of the dynamics within Grec , we may then track by intermittence
the trajectories stemming from this set, thanks to recurrence times. Following Hilfer [11],
we introduce the mapping
S : Grec −→ Grec
(1)
x 7−→ φ G (x) x.
τ
Iterations of S permit to follow the temporal evolution of a point of Grec . Let us remark
that for all k ≥ 1, S k x ∈ ∂G. Since m′ (G) = m′ (Gatt ) + m′ (Grec ), it is now possible to
track almost all trajectories starting from G, at least by intermittence.
The following characteristic times will also be useful:
n o
∀x ∈ Grec , τr (x) = inf t > 0 | φt x ∈ / G , τe (x) = inf t > 0 | φτr (x)+t x ∈ G .

The time τr (x) is the time that the trajectory starting from x stays in G before leaving, while
τe (x) is the time this trajectory then spends outside G (see figure 1). Those times verify
τG (x) = τr (x) + τe (x).
We may also note that by continuity of the Hamiltonian flow, if τr (x) = 0, then x ∈ ∂G.
Figure 1. Characteristic times.
In order to obtain global informations on the dynamics inside G, sets of trajectories -

i.e. evolution of subsets of G - should be studied, for instance through the evolution of their
measures.
182 Pierre Inizan
This problem is studied in details in [11, 10]. The operator S is redefined at precision
∆t on the set of measures on G and appears as a convolution product. For a measure ρ on
G and a subset A ⊂ G, Hilfer obtains
S(∆t)ρ(A)(t) = p∆t ∗ ρ(A)(t).
Then he focuses on the induced dynamics after a renormalization in time scale (con-
tinuous time-limit in [11] and ultra-long time limit in [10]) and obtains a new operator
associated to a new time step, S̃(∆t).
f In that case, he shows that the characteristic deriva-
tive of this operator,
more
precisely the infinitesimal generator [8, p.356] G associated to
the semi-group S̃(∆t)f and defined by
∆t≥0
f
S̃(∆t)ρ(A)(t)
f − ρ(A)(t)
Gρ(A)(t) = lim ,
∆t→0
f ∆t
f
may be equal (up to the sign) to the Marchaud fractional derivative of order α [26, p.109],
with α ∈ (0, 1). Actually, this approach is part of a deeper questioning on time structure
and irreversibility [14, 13].
However some points may still seem unclear, such as the signification of S(∆t) and the
renormalization procedures. Furthermore, the exponent α remains unspecified.
While keeping a similar approach, we propose here a simple dynamical model for which
we study the infinitesimal generator. In several examples, it is proportional to the ordinary
derivative. Then we use Zaslavsky analysis on Hamiltonian chaotic systems: it that case,
the generator may become a fractional derivative.
3. Construction of a Simple Model

Let us recall our objective: we would like to describe the dynamics of our system restricted
to G, in a global way, i.e. by considering sets of trajectories.
To do so, we introduce the mapping N given by
N : T ′ −→ F(R, R)
A 7−→ NA ,
where NA is a real-valued function defined by
NA : R −→ R+
′ t
t 7−→ m ((φ A) ∩ G).
Let A ∈ T ′ and t0 ∈ R. We want to know the infinitesimal evolution of NA (t0 ).

Several successive steps will lead us to a general formula.
3.1. Model with One Trap

Let ∆t > 0. We consider the following binary dynamics: all of the trajectories which leave
G are “trapped” within P ⊂ Γ, with P ∩ G = ∅. Then they come back after 2∆t, and stay
in G during a multiple of ∆t, until possibly leaving again.
We may then split G with the two following sets:
G0 (∆t) = {x ∈ G | τr (x) ≥ ∆t} , (2)

G1 (∆t) = {x ∈ G | τr (x) < ∆t, τe (x) = 2∆t} . (3)
We remark that G1 (∆t) may also be written as
G1 (∆t) = {x ∈ G | τr (x) = 0, τe (x) = 2∆t} ,

= {x ∈ G | τr (x) < ∆t} ,
= {x ∈ G | τr (x) = 0} .
This set is directly linked to trap P .

Hence we have G0 (∆t) ∩ G1 (∆t) = ∅ and G = G0 (∆t) ∪ G1 (∆t).
As in [10], we define the numbers
m′ (Gk (∆t))
pk (∆t) = , k ∈ {0, 1}.
m′ (G)
Those two quantities provide a probability density associated to recurrence times

(p0 (∆t) + p1 (∆t) = 1).
We suppose that these sets are “well mixed”:
∀B ∈ T ′ , m′ (B ∩ G0 (∆t)) = p0 (∆t)m′ (B), m′ (B ∩ G1 (∆t)) = p1 (∆t)m′ (B).
Starting from NA (t0 ) = m′ (A), we determine the following states. The shifts will
occur every ∆t, so we may just consider NA (t0 + n∆t), with n ∈ N.
1. At t+
0 , trajectories starting from A are splitting: some of them stay in G while the
others leave G during 2∆t. We note A0 the set of initial conditions of the first ones
and A1 the set of the second ones. Consequently, m′ (A0 ) = p0 m′ (A) and m′ (A1 ) =
p1 m′ (A) (we omit the dependance of p0 and p1 in ∆t).
2. At t0 + ∆t, only A0 is in G:
NA (t0 + ∆t) = m′ (A0 ) = p0 NA (t0 ).
Within the trap, A1 becomes A11 .

At t+
0 + ∆t, it is now A0 which splits similarly to A, and gives birth to A00 and A01 :
m (A00 ) = p0 m′ (A0 ) and m′ (A01 ) = p1 m′ (A0 ).
′
3. Trajectories which escaped from G at t+ +

0 come back at t0 + 2∆t. Consequently, at
t0 + 2∆t, only A00 is present in G:
NA (t0 + 2∆t) = m′ (A00 ) = p0 NA (t0 + ∆t).
At t+
0 + 2∆t, A00 splits into A000 and A001 , A11 comes back (it turns into A110 ), and
A01 stays outside G while becoming A011 .
184 Pierre Inizan
4. At t0 + 3∆t, G contains A000 and A110 . Hence we have

NA (t0 + 3∆t) = m′ (A000 ) + m′ (A110 ),
= p0 m′ (A00 ) + m′ (A1 ),
= p0 NA (t0 + 2∆t) + p1 NA (t0 ).
At t+
0 + 3∆t, A000 splits into A0000 and A0001 , A110 into A1100 and A1101 , A011
comes back and becomes A0110 . Concerning A001 , it stays outside G and turns into
A0011 .
5. At t0 + 4∆t, we find in G the sets A0000 , A1100 and A0110 :
NA (t0 + 4∆t) = m′ (A0000 ) + m′ (A1100 ) + m′ (A0110 ),
= p0 m′ (A000 ) + m′ (A110 ) + m′ (A01 ),

= p0 NA (t0 + 3∆t) + p1 NA (t0 + ∆t).
A sketch of the dynamics is given in figure 2.

An immediate generalization leads to
∀ n ∈ Z, NA (t0 + n∆t) = p0 NA (t0 + (n − 1)∆t) + p1 NA (t0 + (n − 3)∆t).
In particular,
NA (t0 + ∆t) = p0 NA (t0 ) + p1 NA (t0 − 2∆t). (4)
Keeping in mind definition (1), we note S(∆t) the operator of “infinitesimal” temporal
evolution, which leads to the next temporal step. In this example, time is discrete and takes
its values in t0 + ∆tZ + 2∆tZ = t0 + ∆tZ.
So we have
S(∆t)NA (t0 ) = NA (t0 + ∆t). (5)
Given that lim p0 (∆t)NA (t0 ) + p1 (∆t)NA (t0 − 2∆t) = NA (t0 ), S(∆t) verifies
∆t→0+
S(0) = id. (6)

Moreover, from (5), S(∆t) also verifies
∀∆t1 , ∆t2 > 0, S(∆t1 ) S(∆t2 ) = S(∆t1 + ∆t2 ). (7)
Let us remark that (4) may not be used to check property (7). By verifying (6) and (7),
(S(∆t))∆t≥0 defines a one-parameter semi-group. If NA possesses a left derivative at t0 ,
the associated infinitesimal generator G is given by
S(∆t)NA (t0 ) − NA (t0 )
GNA (t0 ) = lim , (8)
∆t→0+ ∆t
p1 (∆t) (NA (t0 − 2∆t) − NA (t0 ))
= lim , (9)
∆t→0+ ∆t
d
= −2p1 (0+ ) − NA (t0 ), (10)
dt
d
where NA (t0 ) is the left derivative of NA at t0 and p1 (0+ ) is the right limit of p1 at 0.
dt−
Figure 2. Model with one trap.
Remark 1. In this example, the function NA cannot be differentiable at t0 , unless

NA′ (t0 ) = 0. Indeed, S(∆t)NA (t0 ) = NA (t0 + ∆t) in that case, so we also have
d
GNA (t0 ) = + NA (t0 ).
dt
186 Pierre Inizan
3.2. Model with Two Traps

We generalize the previous example by supposing that there are now two sets P1 and P2
outside G, which “trap” trajectories during 2∆t and 3∆t respectively. During a cycle, a
trajectory wanders only through one trap. Trapped trajectories then stay in G during a
multiple of ∆t. Figure 3 illustrates this model.
Figure 3. Model with two traps.
As previously, we introduce the following sets:
G0 (∆t) = {x ∈ G | τr (x) ≥ ∆t} ,

G1 (∆t) = {x ∈ G | τr (x) < ∆t, τe (x) = 2∆t} ,
G2 (∆t) = {x ∈ G | τr (x) < ∆t, τe (x) = 3∆t} .
Once again, those sets form a partition of G. For k ∈ {0, 2}, we note pk (∆t) =
m′ (Gk (∆t))
. We still have p0 (∆t) + p1 (∆t) + p2 (∆t) = 1.
m′ (G)
By proceeding similaraly to the previous model, we find:
NA (t0 + ∆t) = p0 NA (t0 ) + p1 NA (t0 − 2∆t) + p2 NA (t0 − 3∆t).
Time evolves here in t0 +∆tZ+2∆tZ+3∆tZ = t0 +∆tZ. The infinitesimal evolution

operator S(∆t) once again verifies
S(∆t)NA (t0 ) = NA (t0 + ∆t),
thus semi-group properties (6) and (7) are still fulfilled.

Concerning the infinitesimal generator, we have
p1 (NA (t0 − 2∆t) − NA (t0 )) p2 (NA (t0 − 3∆t) − NA (t0 ))

GNA (t0 ) = lim + ,
∆t→0+ ∆t ∆t
d
= −(2p1 (0+ ) + 3p2 (0+ )) − NA (t0 ).
dt
3.3. Generalizations
Let {Pk }k∈N∗ be a set of traps with respective trapping times nk ∆t, nk ∈ N∗ . We still
assume that each time a trajectory leaves G, it is trapped by exactly one trap. Hence, if we
note
G0 (∆t) = {x ∈ G | τr (x) ≥ ∆t}
and for all k ∈ N∗ ,
Gk (∆t) = {x ∈ G | τr (x) < ∆t, τe (x) = nk ∆t} ,
we still obtain a partition of G.

For all k ∈ N, if we set
m′ (Gk (∆t))
pk (∆t) = , (11)
m′ (G)
the evolution of NA verifies
X
NA (t0 + ∆t) = pk (∆t)NA (t0 − nk ∆t). (12)
k≥0
Once again, the successive instants belong to t0 +∆tZ. Then S(∆t)NA (t0 ) = NA (t0 +
∆t), and X
S(∆t)NA (t0 ) = pk (∆t)NA (t0 − nk ∆t). (13)
k≥0
Now we consider any trapping times, denoted Tk (∆t) with k ∈ N∗ , and we suppose
they are well-ordered:
0 < T1 (∆t) < · · · < Tk (∆t) < · · · .

( )
X
We set T (∆t) = Tk (∆t)xk | I ⊂ N, Card(I) < ∞, ∀k ∈ I, xk ∈ Z .
k∈I
Tk1 (∆t)
In that case, as soon there exists (k1 , k2 ) ∈ (N∗ )2 such that / Q, the group
∈
Tk2 (∆t)
T (∆t) cannot anymore be written as τ0 Z, but is dense in R. In particular, it is no more
possible to define a minimal time step. But T (∆t) remains countable, so it is still possible
to move to the next step: the operator S(∆t) still makes sense, but is no longer equal to
NA (t0 + ∆t). Consequently, a generalization cannot be done with (12), but with (13):
X
S(∆t)NA (t0 ) = pk (∆t)NA (t0 − Tk (∆t)), (14)
k≥0
188 Pierre Inizan
where pk (∆t) is given by (11), with
G0 (∆t) = {x ∈ G | τr (x) ≥ T1 (∆t)} , (15)
and, for all k ∈ N∗ ,
Gk (∆t) = {x ∈ G | τr (x) < T1 (∆t), τe (x) = Tk (∆t)} . (16)
This formula is to be linked with expression (8) in [10].

We assume by now that ∆t 7→ p0 (∆t) has a right limit at 0, denoted p0 (0+ ). Now we
specify values of pk (∆t) and Tk (∆t) in the case of chaotic Hamiltonian systems.
4. Dynamical Traps and Anomalous Diffusion

Zaslavsky studies in [33] the general shape of chaotic Hamiltonian phase spaces. In chapter
12, he introduces the notion of dynamical trap so as to describe the behavior of trajectories
near KAM tori. This area possesses a self-similar structure: it is composed of imbricated
subsets Pk which verify
m(Pk+1 ) = λM m(Pk ), with λM < 1.
Moreover, the trapping times Tk associated also have a self-similar property:
Tk+1 = λT Tk , with λT > 1.
Trapping times are hence all the longer as traps are small. This analysis can also be
found in [32, 34].
This kind of structure has “macroscopic” consequences: when one studies diffusion of
particules through a probabilistic description of the system, the moment of order 2 is ruled
by the following law:
hx2 i ∝ tµ .
The classical case (normal diffusion) corresponds to µ = 1. The terms subdiffusion
and superdiffusion are respectively used for µ < 1 and µ > 1. See [33, part.3] and [28] for
more details.
Those anomalous diffusion phenomena can be described with the introduction of frac-
tional derivatives into some specific partial derivatives equations [20, 31, 25], [33, chap.
16].
One of the equations proposed by Zaslavsky [33, p.249], a simplified fractional Fokker-
Planck-Kolmogorov equation, is given by
∂β ∂α
P (x, t) = (A(x)P (x, t)) , 0 < β ≤ 1, 0 < α ≤ 2, (17)
∂tβ ∂xα
where P (x, t) is the probability to find the particule at position x at time t.
If we assume A constant, this equation leads to the following transport equation [33,
p.251]:
hxα i ∝ tβ .
The classical case corresponds to β = 1 and α = 2. The transport exponent [33, p.192]
is defined by
2β
µ= . (18)
α
According to Zaslavsky [33, p.251, p.263], the influence of the dynamical traps appears
through the following relation:
| ln(λM )|
µ= . (19)
ln(λT )
Equality between (18) and (19) provides a connection between the fractal structure of
the phase space and the fractional derivatives of (17). However, the justification for the
introduction of those derivatives in equations (16.3) and (16.4) of [33] is not clear. An
approach based on Continuous Time Random Walks (CTRW) [18, 20, 30] leads to such
derivatives, but those probabilistic models do not rely on the “microscopic” dynamics of
the trajectories.
We propose here to link the emergence of fractional operators with the self-similar
structure of the phase space described above. To do so, we simplify the original dynamics by
only considering the dynamical traps which appear in typical chaotic Hamiltonian systems
(see figure 4), with both of their fractal properties. In that case, we may use the model
presented in section 3.
Figure 4. Simplification of the original dynamics.
Coefficients λM and λT a priori depend on ∆t. Because of the dynamical definition of

traps Pk , the subsets Gk (∆t) also verify, for k ≥ 1,
m′ (Gk+1 (∆t)) = λM (∆t) m′ (Gk (∆t)).
190 Pierre Inizan
Concerning the characteristic times, we have, for all k ≥ 1,
Tk (∆t) = T1 (∆t) λT (∆t)k−1 .
We would like traps structure to become thiner when ∆t → 0, while remaining self-
similar. This leads us to assume
λM (∆t) = (λM )∆t and λT (∆t) = (λT )∆t , (20)
where by sake of lisibility, λM and λT are now two real numbers such that 0 < λM < 1
and λT > 1.
Remark 2. The transport exponent remains unchanged with definition (20):
| ln(λM (∆t))| | ln(λM )|

∀∆t > 0, µ = = .
ln(λT (∆t)) ln(λT )
Consequently, for all k ≥ 1,
m′ (Gk (∆t)) = (λM )(k−1)∆t m′ (G1 (∆t)),
and
(k−1)∆t
Tk (∆t) = T1 (∆t) λT . (21)
In order to obtain smaller characteristic times when ∆t → 0, we suppose that
lim T1 (∆t) = 0. (22)

∆t→0+
X
Using relation pk (∆t) = 1, we find that for all k ≥ 1,
k≥0
(k−1)∆t
pk (∆t) = (1 − p0 (∆t))(1 − λ∆t
M )λM . (23)
The infinitesimal evolution (14) of the system thus becomes

X
S(∆t)NA (t0 ) = p0 (∆t)NA (t0 )+(1 − p0 (∆t)) (1−λ∆t λk∆t k∆t

M) M NA t0 − T1 (∆t)λT . (24)
k≥0
The infinitesimal generators related to (24) can now be determined.
5. Fractional Infinitesimal Generator

Hölder conditions on NA appear in this part, so we need the following definitions. Let
Ω ⊂ R, f : Ω → R and α ∈ (0, 1].
Definition 1. Let x ∈ Ω. The function f satisfies the Hölder condition of order α at x if
∃c > 0, ∃η > 0, ∀ y ∈ Ω, |x − y| ≤ η ⇒ |f (x) − f (y)| ≤ c|x − y|α .

Definition 2. The function f locally satisfies the Hölder condition of order α if for all
x ∈ Ω, f satisfies the Hölder condition of order α at x.
Definition 3. The function f satisfies the Hölder condition of order α if
∃c > 0, ∀ x, y ∈ Ω, |f (x) − f (y)| ≤ c|x − y|α .
If α = 1, f is called Lipschitz continuous.
Now we go back to our problem and we begin to show that S(∆t) still fulfills (6).
Lemma 1. If NA satisfies the Hölder condition of order β, with β < µ, then
lim S(∆t)NA (t0 ) = NA (t0 ).

∆t→0+
Proof. The difference S(∆t)NA (t0 ) − NA (t0 ) verifies

X
S(∆t)NA (t0 ) − NA (t0 ) = (1 − p0 (∆t)) (1 − λ∆t λk∆t NA t0 − T1 (∆t)λk∆t

M) M T − NA (t0 ) .
k≥0
We remark that λM λβT < 1 if and only if β < µ. Given that NA satisfies the Hölder
condition of order β, we obtain
X β
|S(∆t)NA (t0 ) − NA (t0 )| ≤ (1 − p0 (∆t)) (1 − λ∆t
M) λk∆t
M T1 (∆t)λk∆t
T
k≥0
1 − λ∆t
≤ (1 − p0 (∆t)) T1 (∆t)β
M
∆t
1 − λM λβT
On the one hand,
1 − λ∆t
M + ln(λM )
lim (1 − p0 (∆t)) ∆t = (1 − p0 (0 )) β
,
∆t→0+ β ln(λ M λ )
1 − λM λT T
and on the other hand, lim T1 (∆t) = 0 from assumption (22).

∆t→0+
Consequently, lim (S(∆t)NA (t0 ) − NA (t0 )) = 0.
∆t→0+
As it has already be seen, definition (4) cannot be used to check property (7). So we
just assume that (7) is fulfilled.
We recall that the infinitesimal generator G associated to this semi-group verifies
S(∆t)NA (t0 ) − NA (t0 )

GNA (t0 ) = lim .
∆t→0+ ∆t
In [14, 10, 13, 11], Hilfer shows that fractional derivatives may appear as infinitesimal
generators of renormalized evolution operators. A similar result will now be obtained, but
without using any renormalization.
192 Pierre Inizan
For all ∆t > 0, we note

S(∆t)NA (t0 ) − NA (t0 )
G(∆t)NA (t0 ) = .
∆t
We also introduce the function f defined by
f : R+ × R+ −→ R
(∆t, y) 7−→ λM NA t0 − T1 (∆t)λyT − NA (t0 ) .
y
Consequently,
1 − λ∆t
M
X
G(∆t)NA (t0 ) = (1 − p0 (∆t)) f (∆t, k∆t).
∆t
k≥0
For all k ∈ N, we note

Z k+1
Ik (∆t) = f (∆t, k∆t) = f (∆t, k∆t) dx,
k
Z k+1
Jk (∆t) = f (∆t, x∆t) dx.
k
Hence G(∆t)NA (t0 ) can be written as
1 − λ∆t
M
X
G(∆t)NA (t0 ) = (1 − p0 (∆t)) Ik (∆t).
∆t
k≥0
5.1. Case µ > 1

In that case, λM λT < 1.
Theorem 2. If NA is differentiable and Lipschitz continuous on R, and if T1 is differentiable
at 0, then
d
GNA (t0 ) = −γ NA (t0 ),
dt
+ µ ′
where γ = (1 − p0 (0 )) T (0).
µ−1 1
X
Proof. First we prove that lim (Ik (∆t) − Jk (∆t)) = 0.
∆t→0+
k≥0
The function NA is differentiable, so is y 7→ f (∆t, y) for all ∆t ≥ 0, and
∂2 f (∆t, y) = ln(λM )f (∆t, y) − T1 (∆t) (λM λT )y NA′ (t0 − T1 (∆t)λyT ).
If we note c the Lipschitz constant, we have
|∂2 f (∆t, y)| ≤ ln(λM )cT1 (∆t) (λM λT )y + cT1 (∆t) (λM λT )y .
By setting c′ = c(1 + ln(λT )), we obtain
|∂2 f (∆t, y)| ≤ c′ T1 (∆t) (λM λT )y .

Let k ∈ N. Then
Z k+1
|Ik (∆t) − Jk (∆t)| ≤ |f (∆t, x∆t) − f (∆t, k∆t)| dx,
k
≤ ∆t sup |∂2 f (∆t, x∆t)|,
[k,k+1]
≤ c ∆tT1 (∆t) (λM λT )k∆t .

′
1
(λM λT )k∆t =
X
Given that lim ∆t and lim T1 (∆t) = 0, we infer
∆t→0+ | ln(λM λT )| ∆t→0+
k≥0
that X
lim (Ik (∆t) − Jk (∆t)) = 0.
∆t→0+
k≥0
Consequently,
X
− 1 − p0 (0+ ) ln(λM )

G(∆t)NA (t0 ) ∼ Jk (∆t). (25)
∆t→0+
k≥0
X
Now we can evaluate lim Jk (∆t). Firstly,
∆t→0+
k≥0
X Z ∞
x∆t
NA t0 − T1 (∆t)λTx∆t − NA (t0 ) dx.

Jk (∆t) = λM
k≥0 0
With substitution t = λTx∆t , we obtain

Z ∞
X 1
Jk (∆t) = t−(1+µ) [NA (t0 − tT1 (∆t)) − NA (t0 )] dt, (26)
∆t ln(λT ) 1
k≥0
1 T1 (∆t) ∞ −µ NA (t0 − tT1 (∆t)) − NA (t0 )
Z
= t dt.
ln(λT ) ∆t 1 tT1 (∆t)
NA (t0 − tT1 (∆t)) − NA (t0 )

Secondly, for all t ≥ 1, t−µ ≤ ct−µ , and t 7→ ct−µ is

tT1 (∆t)
integrable on [1, +∞). Moreover,
NA (t0 − tT1 (∆t)) − NA (t0 )

lim = −NA′ (t0 ).
∆t→0+ tT1 (∆t)
Consequently, by dominated convergence,

∞
1
X Z
lim Jk (∆t) = T ′ (0)(−NA′ (t0 )) t−µ dt,
∆t→0+ ln(λT ) 1 1
k≥0
1
= T ′ (0)NA′ (t0 ).
(1 − µ) ln(λT ) 1
194 Pierre Inizan
Finally, from (25),

ln(λM )
lim G(∆t)NA (t0 ) = − 1 − p0 (0+ ) T ′ (0)NA′ (t0 ),

∆t→0+ (1 − µ) ln(λT ) 1
µ
= − 1 − p0 (0+ ) T ′ (0)NA′ (t0 ).
µ−1 1
Remark 3. If NA verifies assumptions of theorem 2, then NA satisfies the Hölder condition

of order 1 and consequently fulfills conditions of lemma 1.
5.2. Case µ < 1

X
We can try to estimate lim Jk (∆t), assuming that NA is smooth enough and rapidly
∆t→0+
k≥0
decreasing in −∞, in order that all the following quantities are well-defined. We integrate
by parts (26):
Z ∞
X −T1 (∆t) NA (t0 − T1 (∆t)) − NA (t0 )
Jk (∆t) = t−µ NA′ (t0 − tT1 (∆t)) dt + .
µ∆t ln(λT ) 1 µ∆t ln(λT )
k≥0
Substitution u = T1 (∆t)t leads to

Z ∞
X −T1 (∆t)µ NA (t0 − T1 (∆t)) − NA (t0 )
Jk (∆t) = u−µ NA′ (t0 − u) du + .
µ∆t ln(λT ) T1 (∆t) µ∆t ln(λT )
k≥0
The integral is not problematic:

Z ∞ Z ∞
−µ ′
lim u NA (t0 − u) du = u−µ NA′ (t0 − u) du.
∆t→0+ T1 (∆t) 0
If T1 is differentiable at 0, then
NA (t0 − T1 (∆t)) − NA (t0 ) T ′ (0)
lim =− 1 N ′ (t0 ).
∆t→0+ µ∆t ln(λT ) µ ln(λT ) A
T1 (∆t)µ
Conversely, lim∆t→0+ = +∞. So we cannot find any infinitesimal generator.
∆t
The assumption on the differentiability of T1 at 0 should hence be replaced.
T1 (∆t)µ
In order that has a finite limit, we suppose that there exists b > 0 such that
∆t
T1 (∆t) ∼ b (∆t)1/µ .
∆t→0+
From a physical point of view, T1 (∆t) and ∆t are homogeneous to time, so we intro-
duce a constant of time τ such that b = τ 1−1/µ :
T1 (∆t) ∼ τ 1−1/µ (∆t)1/µ . (27)

∆t→0+
Under this assumption on T1 , a fractional derivative defined as follows will appear.

Definition 4. Let f : R → R and α ∈ (0, 1). The Marchaud fractional derivative of order
α is defined as
Z ∞
α α
D+ f (t) = u−(1+α) [f (t) − f (t − u)] du,
Γ(1 − α) 0
where Γ is the Gamma function.
This derivative is well-defined if f is bounded and locally satisfies the Hölder condition
of order δ, with δ > α. See [26, p.109] for more details.
Now we can enunciate the main result of the article.
Theorem 3. If NA satisfies the Hölder condition of order β and locally satisfies the Hölder
condition of order ν, with β < µ < ν, then
µ
GNA (t0 ) = −γ̃ τ µ−1 D+ NA (t0 ), (28)
where γ̃ = Γ(1 − µ)(1 − p0 (0+ )).

X
Proof. As previously, we firstly prove that lim (Ik (∆t) − Jk (∆t)) = 0.
∆t→0+
k≥0
Let k ∈ N. For all x ∈ [k, k + 1],
x∆t
NA (t0 − T1 (∆t)λTx∆t ) − NA (t0 )

f (∆t, x∆t) − f (∆t, k∆t) = λM
i
− NA (t0 − T1 (∆t)λk∆t T ) − NA (t0 )
h i
x∆t
+ λM − λk∆t
M NA (t0 − T1 (∆t)λk∆t
T ) − NA (t0 ) ,

= λxM NA (t0 − T1 (∆t)λTx∆t ) − NA (t0 − T1 (∆t)λk∆tT )
h i
x∆t
+ λM − λk∆t
M N (t
A 0 − T 1 (∆t)λ k∆t
T ) − N (t
A 0 ) .
Concerning the first right-hand member, we obtain the following inequality:

Z k+1
x∆t
NA (t0 − T1 (∆t)λTx∆t ) − NA (t0 − T1 (∆t)λk∆t

λM T ) dx
k
Z k+1
k∆t NA (t0 − T1 (∆t)λTx∆t ) − NA (t0 − T1 (∆t)λk∆t

≤ λM T ) dx,

k
Z k+1
k∆t T1 (∆t) λTx∆t − λk∆t β dx,

≤ λM T
k
β
(k+1)∆t
≤ λk∆t
M T1 (∆t)
β
λT − λk∆t
T ,
k∆t
≤ T1 (∆t)β λ∆t λM λβT

T −1 .
196 Pierre Inizan
For the second one, we have

k+1 h x∆t
Z i
λM − λk∆tM NA (t0 − T1 (∆t)λk∆t
T ) − NA (t0 ) dx

k
Z k+1
β kβ∆t
x∆t
λM − λk∆t

≤ T1 (∆t) λT M
dx,
k

(k+1)∆t
≤ T1 (∆t)β λTkβ∆t λM − λk∆t
M ,
k∆t
≤ T1 (∆t)β λ∆tM −1 λM λβT .
Consequently,
k∆t
β
|Ik (∆t) − Jk (∆t)| ≤ T1 (∆t)β λ∆t
T − 1 + λ ∆t
M − 1 λ λ
M T .
Since β < µ, λM λβT < 1. Then,

X λ∆t − 1 + λ∆t
M −1
|Ik (∆t) − Jk (∆t)| ≤ T1 (∆t)β T .
k≥0 1 − λM λβT
λ∆t ∆t
T − 1 + λM − 1 ln(λT ) + ln(λM )
Given that lim =− and lim T1 (∆t)β = 0,
λM λβT

∆t→0+ 1− β ∆t→0+
ln λM λT
we infer that X
lim (Ik (∆t) − Jk (∆t)) = 0.
∆t→0+
k≥0
Relation (25) is hence still valid here. Furthermore, (26) holds for µ < 1. Substitution
u = tT1 (∆t) leads to
Z ∞
X T1 (∆t)µ
Jk (∆t) = u−(1+µ) [NA (t0 − u) − NA (t0 )] du.
∆t ln(λT ) T1 (∆t)
k≥0
By definition, 0 ≤ NA (t) ≤ m′ (G) for all t ∈ R. Since we have also assumed that NA
locally satisfies the Hölder condition of order ν > µ, its Marchaud fractional derivative of
order µ is well-defined.
As a consequence,
Z ∞
Γ(1 − µ) µ
lim u−(1+µ) [NA (t0 − u) − NA (t0 )] du = − D+ NA (t0 ).
∆t→0 +
T1 (∆t) µ
With relation (27), we obtain
X τ µ−1 Γ(1 − µ) µ
lim Jk (∆t) = − D+ NA (t0 ).
∆t→0+ ln(λT ) µ
k≥0
Finally,
τ µ−1 ln(λM )Γ(1 − µ) µ
GNA (t0 ) = (1 − p0 (0+ )) D+ NA (t0 )
ln(λT ) µ
µ
= −(1 − p0 (0+ ))τ µ−1 Γ(1 − µ) D+ NA (t0 ).
We have deliberately let the constant τ appear in (28) for reasons of dimensional homo-
µ µ
geneity [16]: the relevant derivative is not D+ , but τ µ−1 D+ , in order to be homogeneous
to the inverse of a time.
Remark 4. So as to respect dimensional homogeneity, a constant of time τ ′ should have

been introduced for the traps constants:
′ ′
λT (∆t) = (λT )∆t/τ , λM (∆t) = (λM )∆t/τ .
However, from remark 2, the transport exponent should have eventually been un-
changed.
6. Discussion
6.1. Characterization of µ
From distribution (pk (∆t))k≥0 and characteristic times
(Tk (∆t))k≥0 (with T0 (∆t) = 0), we can evaluate moments hT α i∆t with α > 0, defined by
X
hT α i∆t = pk (∆t)Tk (∆t)α .
k≥0
In the case of dynamical traps described by (21) and (23), we obtain:


1 − λ∆t
M α
∆t T1 (∆t) if α < µ,
 (1 − p (∆t))

α 0
hT i∆t = 1 − λM λT α

 +∞ if α ≥ µ.
Consequently, if we note hT α i = lim hT α i∆t , parameter µ appears as a critical point:

∆t→0+

α 0 if α < µ,
hT i =
+∞ if α ≥ µ.
However, if m(G) > 0 and µ ≤ 1, hT i = ∞, which does not respect the Kac lemma
[17, 19]. Then we should assume m(G) = 0. This remark is closely akin to the approach
of Hilfer [11], where the fractional infinitesimal generator only appears for sets of measure
0.
Remark 5. If G is a section transverse to the Hamiltonian flow (a Poincaré section for

instance), then m(G) = 0, and since all the trajectories cross G, τr (x) = 0 for all x ∈ G.
Consequently, G0 (∆t) = ∅ and p0 (∆t) = 0, for all ∆t > 0. Hence p0 (0+ ) = 0, which
µ
leads to GNA (t0 ) = −Γ(1 − µ)τ µ−1 D+ NA (t0 ).
198 Pierre Inizan
6.2. Fractional Kinetic Equation

The model presented here do not explain the emergence of fractional derivatives in equa-
tions such as (17). However, the fractional exponent we have obtained is exactly the trans-
port coefficient (19). This result is compatible with relation (18) which involves the frac-
tional exponents of Zaslavsky.
Indeed, let us assume that S(∆t) could be applied to P (x, t) in order to describe “a
generalized shift of P (x, t) along t by ∆t” [33, p.246]. Then the temporal derivative asso-
ciated to the temporal evolution of P (x, t) is the infinitesimal generator G. In the case of
anomalous diffusion, exponents α and β in (17) become completely determined.
• If µ > 1 (superdiffusion), then the temporal derivative is classic: β = 1. Superdif-
fusion is exclusively taken into account by the spatial derivative of order α = µ2 .
Equation (17) thus becomes
∂P (x, t) ∂ 2/µ
= (A(x)P (x, t)) .
∂t ∂x2/µ
• If µ < 1 (subdiffusion), β = µ so α = 2: the temporal derivative is the only one to

be fractional. Consequently, (17) turns into
∂µ ∂2
P (x, t) = (A(x)P (x, t)) .
∂tµ ∂x2
In particular, if our model applies to P (x, t), fractional derivatives in space and time
cannot coexist.
7. Conclusion
The model which has been described in this article attempts to explain, from a dynami-
cal view point, the emergence of fractional derivatives in chaotic Hamiltonian systems. It
seems simplier than the formalism of Hilfer, in particular because no renormalization ap-
pears. Moreover, it strongly relies on fractal properties of the phase space. Our approach
is obviously perfectible on several aspects. It does not explain why T1 (∆t) should fulfill
(27), and condition m(G) = 0 imposed by the Kac lemma should be clarified. So as to test
the validity of the model, other systems should also be considered, in particular strongly
chaotic systems, where the distribution of recurrence times is similar to an exponential law
[3, 15]. Finally, we believe that there are still enough freedom degrees in our model for
allowing us to enhance it in forthcoming studies.
References
[1] O.P. Agrawal. Formulation of Euler-Lagrange equations for fractional variational
problems. J. Math. Anal. Appl., 272:368–379, 2002.
[2] D. Baleanu and S.I. Muslih. Lagrangian formulation of classical fields within
Riemann-Liouville fractional derivatives. Physica Scripta, 72:119–121, 2005.
[3] N. Buric, A. Rampioni, G. Turchetti, and S. Vaienti. Weak chaos and poincaré recur-
rences for area preserving maps. J. Phys. A, 36, 2003.
[4] A. Compte. Stochastic foundations of fractional dynamics. Physical Review E,

53(4):4191–4193, 1996.
[5] J. Cresson. Fractional embedding of differential operators and Lagrangian systems.

Journal of Mathematical Physics, 48(3):033504, 2007.
[6] J. Cresson. Inverse problem of fractional calculus of variations for partial differential
equations. Commun Nonlinear Sci Numer Simulat, 2009.
[7] J. Cresson and P. Inizan. About fractional Hamiltonian systems. Physica Scripta,
T136:014007, 2009.
[8] Feller. An Introduction to Probability Theory and Its Applications, Volume II. Wiley,
New York, 1971.
[9] R. Gorenflo, F. Mainardi, D. Moretti, and P. Paradisi. Time fractional diffusion: A

discrete randomw walk approach. Nonlinear Dynamics, 29:129–143, 2002.
[10] R. Hilfer. Foundations of fractional dynamics. Fractals, 3(3):549–556, 1995.
[11] R. Hilfer. Fractional dynamics, irreversibility and ergodicity breaking. Chaos, Solition
& Fractals, 5(8):1475–1484, 1995.
[12] R. Hilfer. Applications of Fractional Calculus in Physics. World Scientific Publ. Co.,
Singapore, 2000.
[13] R. Hilfer. Fractional evolution equations and irreversibility. In D. Helbing, H. Her-

rmann, M. Schreckenberg, and D. Wolf, editors, Traffic and Granular Flow ’99.
Springer, 2000.
[14] R. Hilfer. Remarks on fractional time. In L. Castell and O. Ischebeck, editors, Time,
Quantum and Information, page 235. Springer-Verlag, Berlin, 2003.
[15] M. Hirata, B. Saussol, and S. Vaienti. Statistics of return times: A general framework
and new applications. Commun. Math. Phys., 206:33–55, 1999.
[16] P. Inizan. Homogeneous fractional embeddings. Journal of Mathematical Physics,

49(8):082901, 2008.
[17] M. Kac. Probability and Related Topics in Physical Sciences. Interscience, New York,
1957.
[18] M.M. Meerschaert and H.-P. Scheffler. Limit theorems for continuous-time random
walks with infinite mean waiting times. J. Appl. Prob., 41:623–638, 2004.
[19] J.D. Meiss. Average exit time for volume-preserving maps. Chaos, 7(1):139–147,
1997.
200 Pierre Inizan
[20] R. Metzler and J. Klafter. The random walk’s guide to anomalous diffusion: a frac-
tional dynamics approach. Pysics Reports, 339:1–77, 2000.
[21] K.B. Oldham and J. Spanier. The Fractional Calculus. Academic Press, New York
and London, 1974.
[22] I. Podlubny. Fractional Differential Equations. Academic Press, 1999.
[23] F. Riewe. Nonconservative Lagrangian and Hamiltonian mechanics. Physical Review

E, 53(2):1890, 1996.
[24] J. Sabatier, O.P. Agrawal, and J.A. Tenreiro Machado. Advances in fractional calcu-
lus. Springer, 2007.
[25] A.I. Saichev and G.M. Zaslavsky. Fractional kinetic equations: solutions and applica-
tions. Chaos, 7(4):753–764, 1997.
[26] S.G. Samko, A.A. Kilbas, and O.I. Marichev. Fractional integrals ans derivatives:
theory and applications. Gordon and Breach, 1993.
[27] W.R. Schneider and W. Wyss. Fractional diffusion and wave equations. Journal of
Mathematical Physics, 30(1), 1989.
[28] M.F. Shlesinger, G.M. Zaslavsky, and J. Klafter. Strange kinetics. Nature, 363:31–37,
1993.
[29] A.A. Stanislavsky. Probability interpretation of the integral of fractional order. Theo-
ritical and Mathematical Physics, 138(3), 2004.
[30] A.A. Stanislavsky. Hamiltonian formalism of fractional systems. Eur. Phys. J. B,

49:93–101, 2006.
[31] G.M. Zaslavsky. Renormalization group theory of anomalous transport in systems

with hamiltonian chaos. Chaos, 4(1):25–33, 1994.
[32] G.M. Zaslavsky. Chaotic dynamics and the origin of statistical laws. Physics Today,
August 1999.
[33] G.M. Zaslavsky. Hamiltonian Chaos & Fractional Dynamics. Oxford University
Press, Oxford, 2005.
[34] G.M. Zaslavsky, M. Edelman, and B.A. Niyazov. Self-similarity, renormalization, and
phase space nonuniformity of hamiltonian chaotic dynamics. Chaos, 7(1):159–181,
1997.

Chapter 5
N UMERICAL A PPROXIMATIONS TO F RACTIONAL

P ROBLEMS OF THE C ALCULUS OF VARIATIONS
AND O PTIMAL C ONTROL
Shakoor Pooseh∗, Ricardo Almeida† and Delfim F. M. Torres‡

Center for Research and Development in Mathematics and Applications (CIDMA)
Department of Mathematics, University of Aveiro, 3810-193 Aveiro, Portugal
Keywords: fractional calculus of variations, fractional optimal control, numerical methods,

direct methods, indirect methods
AMS Subject Classification: 49K05, 49M25, 26A33
1. Introduction
A fractional problem of the calculus of variations and optimal control consists in the study
of an optimization problem in which the objective functional or constraints depend on
derivatives and integrals of arbitrary, real or complex, orders. This is a generalization of
the classical theory, where derivatives and integrals can only appear in integer orders.
1.1. Preliminaries
Integer order derivatives and integrals have a unified meaning in the literature. In con-
trast, there are several different approaches and definitions in fractional calculus for deriva-
tives and integrals of arbitrary order. The following definitions and notations will be used
throughout this chapter. See [19].
Definition 1.1 (Gamma function). The Euler integral of the second kind
Z ∞
Γ(z) = tz−1 e−t dt, Re(z) > 0,
0
is called the gamma function.
∗
E-mail address: spooseh@ua.pt
†
E-mail address: ricardo.almeida@ua.pt
‡
E-mail address: delfim@ua.pt
202 Shakoor Pooseh, Ricardo Almeida and Delfim F. M. Torres
The gamma function has an important property, Γ(z + 1) = zΓ(z), and hence Γ(z) =
(z − 1)! for z ∈ N, which allows to extend the notion of factorial to real numbers. Other
properties of this special function can be found in [5].
Definition 1.2 (Mittag–Leffler function). Let α > 0. The function Eα defined by
∞
X zj
Eα (z) = ,
Γ(αj + 1)
j=0
whenever the series converges, is called the one parameter Mittag–Leffler function. The
two-parameter Mittag–Leffler function with parameters α, β > 0 is defined by
∞
X zj
Eα,β (z) = . (1)
Γ(αj + β)
j=0
α

Definition 1.3 (Grünwald–Letnikov derivative). Let 0 < α < 1 and k be the general-
ization of binomial coefficients to real numbers.
• The left Grünwald–Letnikov fractional derivative is defined as
∞
GL α 1 X k α
a D t x(t) = lim (−1) x(t − kh). (2)
h→0+ hα k
k=0
• The right Grünwald–Letnikov derivative is

∞
GL α 1 X k α
t Db x(t) = lim α (−1) x(t + kh). (3)
h→0+ h k
k=0
In the above mentioned definitions, αk is the generalization of binomial coefficients to

real numbers, defined by

α Γ(α + 1)
= .
k Γ(k + 1)Γ(α − k + 1)
In this relation, k and α can be any integer, real or complex number, except that α ∈
/
{−1, −2, −3, . . .}.
Definition 1.4 (Riemann–Liouville fractional integral). Let x(·) be an integrable func-
tion in [a, b] and α > 0.
• The left Riemann–Liouville fractional integral of order α is given by
Z t
1
α
a It x(t) = (t − τ )α−1 x(τ )dτ, t ∈ [a, b].
Γ(α) a
• The right Riemann–Liouville fractional integral of order α is given by

Z b
1
α
t Ib x(t) = (τ − t)α−1 x(τ )dτ, t ∈ [a, b].
Γ(α) t
Numerical Approximations to Fractional Problems ... 203
Definition 1.5 (Riemann–Liouville fractional derivative). Let x(·) be an absolutely con-

tinuous function in [a, b], x(·) ∈ AC[a, b], and 0 ≤ α < 1.
• The left Riemann–Liouville fractional derivative of order α is given by
1 d t
Z
α
a Dt x(t) = (t − τ )−α x(τ )dτ, t ∈ [a, b].
Γ(1 − α) dt a
• The right Riemann–Liouville fractional derivative of order α is given by

Z b
α 1 d
t Db x(t) = − (τ − t)−α x(τ )dτ, t ∈ [a, b].
Γ(1 − α) dt t
Another type of fractional derivatives, introduced by Caputo, is closely related to the

Riemann–Liouville definitions.
Definition 1.6 (Caputo’s fractional derivative). For a function x(·) ∈ AC[a, b] with 0 ≤
α < 1:
• The left Caputo fractional derivative of order α is given by

Z t
1
C α
D
a t x(t) = (t − τ )−α ẋ(τ )dτ, t ∈ [a, b].
Γ(1 − α) a
• The right Caputo fractional derivative of order α is given by

Z b
−1
C α
t D b x(t) = (τ − t)−α ẋ(τ )dτ, t ∈ [a, b].
Γ(1 − α) t
Definition 1.7 (Hadamard fractional integral). Let x : [a, b] → R.
• The left Hadamard fractional integral of order α > 0 is defined by
t α−1 x(τ )
Z t
α 1
a It x(t) = ln dτ, t ∈]a, b[.
Γ(α) a τ τ
• The right Hadamard fractional integral of order α > 0 is defined by

Z b
α 1 τ α−1 x(τ )
I
t b x(t) = ln dτ, t ∈]a, b[.
Γ(α) t t τ
When α = m is an integer, these fractional integrals are m-fold integrals:
dτ1 τ1 dτ2
Z t Z τm−1
x(τm )
Z
m
a It x(t) = ... dτm ,
a τ1 a τ2 a τm
and
b b b
dτ1 dτ2 x(τm )
Z Z Z
m
t Ib x(t) = ... dτm .
t τ1 τ1 τ2 τm−1 τm
Definition 1.8 (Hadamard fractional derivative). For α > 0 and n = [α] + 1,

• The left Hadamard fractional derivative of order α is defined by
n
t n−α−1 x(τ )
Z t
α d 1
a Dt x(t) = t ln dτ, t ∈]a, b[.
dt Γ(n − α) a τ τ
• The right Hadamard fractional derivative of order α is defined by
d n
Z b
α 1 τ n−α−1 x(τ )
D
t b x(t) = −t ln dτ, t ∈]a, b[.
dt Γ(n − α) t t τ
When α = m is an integer, we have

m
d m

m d m
a Dt x(t) = t x(t) and t Db x(t) = −t x(t).
dt dt
1.2. Fractional Calculus of Variations and Optimal Control

Many generalizations to the classical calculus of variations and optimal control have been
made to extend the theory to cover fractional variational and fractional optimal control prob-
lems. A simple fractional variational problem, for example, consists in finding a function
x(·) that minimizes the functional
Z b
J[x(·)] = L(t, x(t), a Dtα x(t))dt, (4)
a
where a Dtα is the left Riemann–Liouville fractional derivative. Typically, some boundary
conditions are prescribed as x(a) = xa and/or x(b) = xb . Classical techniques have been
adopted to solve such problems. The Euler–Lagrange equation for a Lagrangian of the
form L(t, x(t), a Dtα x(t)) has been derived firstly in [30, 31]. Many variants of necessary
conditions of optimality have been studied. A generalization of the problem to include
fractional integrals, i.e., L = L(t, a It1−α x(t), a Dtα x(t)), the transversality conditions of
fractional variational problems and many other aspects can be found in the literature of
recent years. See [1, 4, 6] and references therein. Furthermore, it has been shown that a
variational problem with fractional derivatives can be reduced to a classical problem using
an approximation of the Riemann–Liouville fractional derivatives in terms of a finite sum,
where only derivatives of integer order are present [6].
On the other hand, fractional optimal control problems usually appear in the form of
Z b
J[x(·)] = L(t, x(t), u(t))dt → min
a
α

a Dt x(t) = f (t, x(t), u(t))
s.t.
x(a) = xa , x(b) = xb ,
where an optimal control u(·) together with an optimal trajectory x(·) are required to follow
a fractional dynamic and, at the same time, optimize an objective functional. Again, clas-
sical techniques are generalized to derive necessary optimality conditions. Euler–Lagrange
equations have been introduced, e.g., in [2]. A Hamiltonian formalism for fractional op-
timal control problems can be found in [9] that exactly follows the same procedure of the
regular optimal control theory, i.e., those with only integer-order derivatives.
Due to the growing number of applications of fractional calculus in science and engi-
neering (see, e.g., [11, 12, 33, 34]), numerical methods are being developed to provide tools
for solving such problems. Using the Grünwald–Letnikov approach, it is convenient to ap-
proximate the fractional differentiation operator, Dα , by generalized finite differences. In
[25] some problems have been solved by this approximation. In [13] a predictor-corrector
method is presented that converts an initial value problem into an equivalent Volterra inte-
gral equation, while [20] shows the use of numerical methods to solve such integral equa-
tions. A good survey on numerical methods for fractional differential equations can be
found in [16].
A numerical scheme to solve fractional differential equations has been introduced in
[7, 8], and [17], making an adaptation, uses this technique to solve fractional optimal control
problems. The scheme is based on an expansion formula to approximate the Riemann–
Liouville fractional derivative. The approximations transform fractional derivatives into
finite sums containing only derivatives of integer order.
In this chapter, we try to analyze problems for which an analytic solution is available.
This approach gives us the ability of measuring the accuracy of each method. To this end,
we need to measure how close we get to the exact solutions. We use the L2 -norm and define
the error function E[x(·), x̃(·)] by
Z b 21
E = kx(·) − x̃(·)k2 = [x(t) − x̃(t)]2 dt ,
a
where x(·) is defined on [a, b].
1.3. A General Formulation

The appearance of fractional terms of different types, derivatives and integrals, and the fact
that there are several definitions for such operators, makes it difficult to present a typical
problem to represent all possibilities. Nevertheless, one can consider the optimization of
functionals of the form
Z b
J[x(·)] = L(t, x(t), Dα x(t))dt (5)
a
that depends on a fractional derivative, Dα , in which x = (x1 , x2 , . . . , xn ), α =

(α1 , α2 , . . . , αn ) and αi , i = 1, 2, . . . , n, are arbitrary real positive numbers. The prob-
lem can be with or without boundary conditions. Many settings of fractional variational
and optimal control problems can be transformed to the optimization of (5). Constraints
that usually appear in the calculus of variations and are always present in optimal control
problems can be included in the functional using Lagrange multipliers. More precisely, in
presence of dynamic constraints as fractional differential equations, we assume that it is
possible to transform such equations to a vector fractional differential equation of the form
Dα x(t) = f (t, x(t)).

In this stage, we introduce a new variable λ = (λ1 , λ2 , . . . , λn ) and consider the optimiza-
tion of
Z b
J[x(·)] = [L(t, x(t), Dα x(t)) + λ(t)Dα x(t) − λ(t)f (t, x(t))] dt.
a
When the problem depends on fractional integrals, I α , a new variable can be defined as
z(t) = I α x(t). Recall that Dα I α x = x (see, e.g., [19]). The equation
Dα z(t) = Dα I α x(t) = x(t)
can be regarded as an extra constraint to be added to the original problem. However, prob-
lems containing fractional integrals can be treated directly to avoid the complexity of adding
an extra variable to the original problem. Interested readers are addressed to [4, 28].
Throughout this chapter, by a fractional variational problem, we mainly consider the
following one-variable problem with given boundary conditions:
Z b
J[x(·)] = L(t, x(t), Dα x(t))dt → min
a

x(a) = xa ,
s.t.
x(b) = xb .
In this setting Dα can be replaced by any fractional operator that is available in the liter-
ature, say, Riemann–Liouville, Caputo, Grünwald–Letnikov, Hadamard and so forth. The
inclusion of constraints is done by Lagrange multipliers. The transition from this problem
to the general one, equation (5), is straightforward and is not discussed here.
1.4. Solution Methods

There are two main approaches to solve variational, including optimal control, problems.
On the one hand, there are direct methods. In a branch of direct methods, the problem is
discretized over a mesh on the interested time interval. Discrete values of the unknown
function on mesh points, finite differences for derivatives, and, finally, a quadrature rule
for the integral, are used. This procedure reduces the variational problem, a continuous
dynamic optimization problem, to static multi-variable optimization. Better accuracies are
achieved by refining the underlying mesh size. Another class of direct methods uses func-
tion approximation through a linear combination of the elements of a certain basis, e.g.,
power series. The problem is then transformed into the determination of the unknown co-
efficients. To get better results in this sense, is the matter of using more adequate or higher
order function approximations.
On the other hand, there are indirect methods that reduce a variational problem to the
solution of a differential equation by applying some necessary optimality conditions. Euler–
Lagrange equations and Pontryagin’s maximum principle are used, in this context, to make
the transformation process. Once we solve the resulting differential equation, an extremal
for the original problem is reached. Therefore, to reach better results using indirect meth-
ods, one has to employ powerful integrators. It is worth, however, to mention here that
numerical methods are usually used to solve practical problems.
These two methods have been generalized to cover fractional problems, which is the
essential subject of this chapter.
2. Expansion Formulas to Approximate Fractional Derivatives

This section is devoted to present two approximations for the Riemann–Liouville, Caputo
and Hadamard derivatives that are referred as fractional operators afterwards. We introduce
the expansions of fractional operators in terms of infinite sums involving only integer order
derivatives. These expansions are then used to approximate fractional operators in prob-
lems of the fractional calculus of variations and fractional optimal control. In this way, one
can transform such problems into classical variational or optimal control problems. Here-
after, a suitable method, that can be found in the classical literature, is employed to find an
approximated solution for the original fractional problem. Here we focus mainly on the left
derivatives and the details of extracting corresponding expansions for right derivatives are
given whenever it is needed to apply new techniques.
2.1. Riemann–Liouville Derivative

2.1.1. Approximation by a Sum of Integer Order Derivatives
Recall the definition of the left Riemann–Liouville derivative for α ∈ (0, 1),
t
1 d
Z
α
a Dt x(t) = (t − τ )−α x(τ )dτ. (6)
Γ(1 − α) dt a
The following theorem holds for any function x(·) that is analytic in an interval (c, d) ⊃
[a, b]. See [6] for a more detailed discussion and [32] for a different proof.
Theorem 2.1. Let (c, d), −∞ < c < d < +∞, be an open interval in R, and [a, b] ⊂ (c, d)
be such that for each t ∈ [a, b] the closed ball Bb−a (t), with center at t and radius b − a,
lies in (c, d). If x(·) is analytic in (c, d), then
∞
α
X (−1)k−1 αx(k) (t)
a Dt x(t) = (t − a)k−α . (7)
k!(k − α)Γ(1 − α)
k=0
Proof. Since x(t) is analytic in (c, d), and Bb−a (t) ⊂ (c, d) for any τ ∈ (a, t) with t ∈
(a, b), the Taylor expansion of x(τ ) at t is a convergent power series, i.e.,
∞
X (−1)k x(k) (t)
x(τ ) = x(t − (t − τ )) = (t − τ )k ,
k!
k=0
and then, by (6),

∞
!
t
1 d (−1)k x(k) (t)
Z X
α
a Dt x(t) = (t − τ )−α (t − τ )k dτ. (8)
Γ(1 − α) dt a k!
k=0
Since (t − τ )k−α x(k) (t) is analytic, we can interchange integration with summation, so
∞
!
t
1 d (−1)k x(k) (t)
X Z
α k−α
a Dt x(t) = (t − τ ) dτ
Γ(1 − α) dt k! a
k=0
∞
(−1)k x(k) (t)

1 d X
= (t − a)k+1−α
Γ(1 − α) dt k!(k + 1 − α)
k=0
∞
(−1)k x(k+1) (t) (−1)k x(k) (t)

1 X
= (t − a)k+1−α + (t − a)k−α
Γ(1 − α) k!(k + 1 − α) k!
k=0
x(t)
= (t − a)−α
Γ(1 − α)
∞
(−1)k−1 (−1)k

1 X
+ + x(k) (t)(t − a)k−α .
Γ(1 − α) (k − α)(k − 1)! k!
k=1
Observe that
(−1)k−1 (−1)k k(−1)k−1 + k(−1)k − α(−1)k
+ =
(k − α)(k − 1)! k! (k − α)k!
k−1
(−1) α
= ,
(k − α)k!
since for any k = 0, 1, 2, . . . we have k(−1)k−1 + k(−1)k = 0. Therefore, the expansion

formula is reached as required.
For numerical purposes, a finite number of terms in (7) is used and one has
N
X (−1)k−1 αx(k) (t)
α
a Dt x(t) ≈ (t − a)k−α . (9)
k!(k − α)Γ(1 − α)
k=0
Remark 2.2. With the same assumptions of Theorem 2.1, we can expand x(τ ) at t,
∞
X x(k) (t)
x(τ ) = x(t + (τ − t)) = (τ − t)k ,
k!
k=0
where τ ∈ (t, b). Similar calculations result in the following approximation for the right
Riemann–Liouville derivative:
N
X −αx(k) (t)
α
t Db x(t) ≈ (b − t)k−α .
k!(k − α)Γ(1 − α)
k=0
A proof for this expansion is available at [32] that uses a similar relation for fractional
integrals. The proof discussed here, however, allows to extract an error term for this expan-
sion easily.
2.1.2. Approximation Using Moments of a Function

By moments of a function, we have no physical or distributive sense in mind. The naming
comes from the fact that, during expansion, the terms of the form
Z t
Vp (t) := Vp (x(t)) = (1 − p) (τ − a)p−2 x(τ )dτ, p ∈ N, τ ≥ a, (10)
a
resemble the formulas of central moments (cf. [8]). We assume that Vp (x(·)), p ∈ N,
denotes the (p − 2)th moment of a function x(·) ∈ AC 2 [a, b].
The following lemma, that is given here without a proof, is the key relation to extract
an expansion formula for Riemann–Liouville derivatives.
Lemma 2.3 (cf. Lemma 2.12 of [12]). Let x(·) ∈ AC[a, b] and 0 < α < 1. Then the
left Riemann–Liouville fractional derivative, a Dtα x(·), exists almost everywhere in [a, b].
Moreover, a Dtα x(·) ∈ Lp [a, b] for 1 ≤ p < α1 and
Z t
α 1 x(a) −α
a Dt x(t) = + (t − τ ) ẋ(τ )dτ , t ∈ (a, b). (11)
Γ(1 − α) (t − a)α a
The same argument is valid for the right Riemann–Liouville derivative and
Z b
α 1 x(b) −α
t Db x(t) = − (τ − t) ẋ(τ )dτ , t ∈ (a, b).
Γ(1 − α) (b − t)α t
Theorem 2.4 (cf. [7]). With the same assumptions of Lemma 2.3, the left Riemann–
Liouville derivative can be expanded as
α (t − a)−α
a Dt x(t) = x(t) + B(α)(t − a)1−α ẋ(t)
Γ(1 − α)
∞
X Γ(p − 1 + α)
− C(α, p)(t − a)1−p−α Vp (t) − (t − a)−α x(t) , (12)
p=2
Γ(α)Γ(1 − α)(p − 1)!
where Vp (t) is defined by (10) and

 
∞
1 X Γ(p − 1 + α)
B(α) = 1 + ,
Γ(2 − α) Γ(α − 1)p!
p=1
1 Γ(p − 1 + α)
C(α, p) = .
Γ(2 − α)Γ(α − 1) (p − 1)!
Proof. Integration by parts on the right-hand-side of (11) gives
α x(a) ẋ(a)
a Dt x(t) = (t − a)−α + (t − a)1−α
Γ(1 − α) Γ(2 − α)
Z t
1
+ (t − τ )1−α ẍ(τ )dτ. (13)
Γ(2 − α) a
Since (τ − a) ≤ (t − a),
τ − a 1−α

1−α 1−α
(t − τ ) = (t − a) 1− .
t−a
Using the binomial theorem, we have
∞
τ − a 1−α X Γ(p − 1 + α) τ − a p

1− = ,
t−a Γ(α − 1)p! t−a
p=0
in which the infinite series converges. Replacing for (t − τ )1−α in (13) gives
x(a) ẋ(a)
α
a Dt x(t) = (t − a)−α + (t − a)1−α
Γ(1 − α) Γ(2 − α)
 
1−α Z t X ∞ p
(t − a) Γ(p − 1 + α) τ − a 
+  ẍ(τ )dτ, t > a.
Γ(2 − α) a Γ(α − 1)p! t−a
p=0
Interchanging the summation and integration operations is possible, and yields

x(a) ẋ(a)
α
a Dt x(t) = (t − a)−α + (t − a)1−α
Γ(1 − α) Γ(2 − α)
∞ Z t
(t − a)1−α X Γ(p − 1 + α)
+ (τ − a)p ẍ(τ )dτ, t > a.
Γ(2 − α) Γ(α − 1)p!(t − a)p a
p=0
Decomposing the infinite sum, integrating, and doing another integration by parts, allow us
to write
x(a) ẋ(a) (t − a)1−α t
Z
α −α 1−α
a Dt x(t) = (t − a) + (t − a) + ẍ(τ )dτ
Γ(1 − α) Γ(2 − α) Γ(2 − α) a
∞ Z t
(t − a)1−α X γ(α, p)

p p−1
+ (t − a) ẋ(t) − p (τ − a) ẋ(τ )dτ
Γ(2 − α) p!(t − a)p a
p=1
∞
x(a) −α ẋ(t) 1−α (t − a)1−α X γ(α, p)
= (t − a) + (t − a) + ẋ(t)
Γ(1 − α) Γ(2 − α) Γ(2 − α) p!
p=1
∞ Z t
(t − a)1−α X γ(α, p)
+ (τ − a)p−1 ẋ(τ )dτ,
Γ(2 − α) (p − 1)!(t − a)p a
p=1
where
Γ(p − 1 + α)
γ(α, p) = .
Γ(α − 1)
Repeating this procedure again, and simplifying the results, ends the proof.
The moments Vp (t), p = 2, 3, . . ., can be regarded as the solutions to the following

system of differential equations:
V̇p (t) = (1 − p)(t − a)p−2 x(t)

(14)
Vp (a) = 0, p = 2, 3, . . .
As before, a numerical approximation is achieved by taking only a finite number of

terms in the series (12). We approximate the fractional derivative as
N
X
α −α 1−α
a Dt x(t) ≈ A(t − a) x(t) + B(t − a) ẋ(t) − C(α, p)(t − a)1−p−α Vp (t), (15)
p=2
where A = A(α, N ) and B = B(α, N ) are given by

 
N
1 X Γ(p − 1 + α) 
A(α, N ) = 1 + , (16)
Γ(1 − α) Γ(α)(p − 1)!
p=2
 
N
1 X Γ(p − 1 + α) 
B(α, N ) = 1 + . (17)
Γ(2 − α) Γ(α − 1)p!
p=1
Remark 2.5. This expansion has been proposed in [14] and a simplification has been made
Γ(p−1+α)
in [8], which uses the fact that the infinite series ∞
P
p=1 Γ(α−1)p! tends to −1, and concludes
that B(α) = 0, and thus
N
X
α −α
a Dt x(t) ≈ A(α, N )t x(t) − C(α, p)t1−p−α Vp (t). (18)
p=2
In practice, however, we only use a finite number of terms in series. Therefore,

N
X Γ(p − 1 + α)
1+ 6= 0,
Γ(α − 1)p!
p=1
and we keep here the approximation in the form of equation (15), [3]. To be more precise,
the values of B(α, N ), for different choices of N and α, are given in Table 1. It shows that
even for a large N , when α tends to one, B(α, N ) cannot be ignored.
Table 1. B(α, N ) for different values of α and N
N 4 7 15 30 70 120 170
B(0.1, N ) 0.0310 0.0188 0.0095 0.0051 0.0024 0.0015 0.0011
B(0.3, N ) 0.1357 0.0928 0.0549 0.0339 0.0188 0.0129 0.0101
B(0.5, N ) 0.3085 0.2364 0.1630 0.1157 0.0760 0.0581 0.0488
B(0.7, N ) 0.5519 0.4717 0.3783 0.3083 0.2396 0.2040 0.1838
B(0.9, N ) 0.8470 0.8046 0.7481 0.6990 0.6428 0.6092 0.5884
B(0.99, N ) 0.9849 0.9799 0.9728 0.9662 0.9582 0.9531 0.9498
Remark 2.6. Similar computations give rise to an expansion formula for t Dbα , the right
Riemann–Liouville fractional derivative:
N
X
α −α
t Db x(t) ≈ A(b − t) x(t) − B(b − t)1−α ẋ(t) − C(α, p)(b − t)1−p−α Wp (t),
p=2
where Z b
Wp (t) = (1 − p) (b − τ )p−2 x(τ )dτ.
t
The coefficients A = A(α, N ) and B = B(α, N ) are the same as (16) and (17) respec-
tively, and C(α, p) is as before.
Remark 2.7. As stated before, Caputo derivatives are closely related to those of Riemann–
Liouville. For any function, x(·), and for α ∈ (0, 1) for which these two kind of fractional
derivatives, left and right, exist, we have
C α x(a)
a Dt x(t) = a Dtα x(t) − ,
(t − a)α
and
C α x(b)
t Db x(t) = t Dbα x(t) − .
(b − t)α
Using these relations, we can easily construct approximation formulas for the left and right
Caputo fractional derivatives, e.g.,
C α
a Dt x(t) ≈ A(α, N )(t − a)−α x(t) + B(α, N )(t − a)1−α ẋ(t)
N
X x(a)
− C(α, p)(t − a)1−p−α Vp (t) − .
(t − a)α
p=2
2.1.3. Examples
To examine the approximations provided so far, we take some test functions, and apply (9)
and (15) to evaluate their fractional derivatives. We compute a Dtα x(t), with α = 12 , for
x(t) = t4 and x(t) = e2t . The exact formulas for the fractional derivatives of polynomials
are derived from
Γ(n + 1)
0.5 n
0 Dt (t ) = tn−0.5 ,
Γ(n + 1 − 0.5)
and for the exponential function one has
0.5 λt
0 Dt (e ) = t−0.5 E1,1−0.5 (λt),
where Eα,β is the two parameter Mittag–Leffler function (1).

Figure 1 shows the results using approximation (9). As we can see, the third approxi-
mations are reasonably accurate for both cases. Indeed, for x(t) = t4 , the approximation
with N = 4 coincides with the exact solution because the derivatives of order five and more
vanish.
Now we use approximation (15) to evaluate fractional derivatives of the same test func-
tions. In this case, for a given function x(·), we can compute Vp by definition, equation
(10). One can also integrate the system (14) analytically, if possible, or use any numerical
integrator. It is clearly seen in Figure 2 that one can get better results by using larger values
of N . Comparing Figures 1 and 2, we find out that the approximation (9) shows a faster
convergence. Observe that both functions are analytic and it is easy to compute higher-order
derivatives.
3 14
Analytic Analytic
N=1, E=0.26792 N=1, E=0.74738
N=2, E=0.13111 N=2, E=0.26928
2.5 12
N=3, E=0.028503 N=3, E=0.089841
2 10
Dα
Dα
0 t
0 t
1.5 8
1 6
0.5 4
0 2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t t
(a) 0 Dt0.5 (t4 ) (b) 0 Dt0.5 (e2t )
Figure 1. Analytic (solid line) versus numerical approximation (9).
3 14
Analytic Analytic
N=1, E=0.26792 N=1, E=0.74738
N=2, E=0.14821 N=2, E=0.40156
2.5 12
N=3, E=0.098334 N=3, E=0.26223
2 10
Dα
Dα
0 t
0 t
1.5 8
1 6
0.5 4
0 2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t t
(a) 0 Dt0.5 (t4 ) (b) 0 Dt0.5 (e2t )
Figure 2. Analytic (solid line) versus numerical approximation (15).
Remark 2.8. A closer look to (9) and (15) reveals that in both cases the approximations
are not computable at a and b for the left and right fractional derivatives, respectively. At
these points we assume that it is possible to extend them continuously to the closed interval
[a, b].
Following Remark 2.5, we show here that neglecting the first derivative in the expansion
(15) can cause a considerable loss of accuracy in computation. Once again, we compute the
fractional derivatives of x(t) = t4 and x(t) = e2t , but this time we use the approximation
given by (18). Figure 3 summarizes the results. Approximation (15) gives a more realistic
approximation using quite small N , 3 in this case.
2.5 12
Analytic Analytic
Approximate, B ≠ 0, N=3, E=0.098334 11 Approximate, B ≠ 0, N=3, E=0.26223
Approximate, B = 0, N=3, E=0.40046 Approximate, B = 0, N=3, E=2.0055
2 10
1.5 8
Dα
Dα
0 t
0 t
7
1 6
0.5 4
0 2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t t
(a) 0 Dt0.5 (t4 ) (b) 0 Dt0.5 (e2t )
Figure 3. Comparison of approximation (15) and approximation (18) of [8].
2.2. Hadamard Derivatives

For Hadamard derivatives, the expansions can be obtained in a quiet similar way [27].
2.2.1. Approximation by a Sum of Integer Order Derivatives

Assume that a function x(·) admits derivatives of any order, then expansion formulas for the
Hadamard fractional integrals and derivatives of x, in terms of its integer-order derivatives,
are given in [10, Theorem 17]:
∞
X
α
0 It x(t) = S(−α, k)tk x(k) (t)
k=0
and
∞
X
α
0 Dt x(t) = S(α, k)tk x(k) (t),
k=0
where
k
1 X k−j k
S(α, k) = (−1) jα
k! j
j=1
is the Stirling function.

As approximations, we truncate infinite sums at an appropriate order N and get the
following formulas:
N
X
α
I
0 t x(t) ≈ S(−α, k)tk x(k) (t),
k=0
and
N
X
α
0 Dt x(t) ≈ S(α, k)tk x(k) (t).
k=0
2.2.2. Approximation Using Moments of a Function

The same idea of expanding Riemann–Liouville derivatives, with slightly different tech-
niques, is used to derive expansion formulas for left and right Hadamard derivatives. The
following lemma is a basis for these new relations.
Lemma 2.9. Let α ∈ (0, 1) and x(·) be an absolutely continuous function on [a, b]. Then
the Hadamard fractional derivatives may be expressed by
t −α t −α
Z t
α x(a) 1
a Dt x(t) = ln + ln ẋ(τ )dτ (19)
Γ(1 − α) a Γ(1 − α) a τ
and
b −α
Z b
α x(b) 1 τ −α
t Db x(t) = ln − ln ẋ(τ )dτ.
Γ(1 − α) t Γ(1 − α) t t
A proof of this lemma, for an arbitrary α > 0, can be found in [18, Theorem 3.2].
Theorem 2.10. Let 0 < a < b and x : [a, b] → R be an absolutely continuous function.
Then
t −α t 1−α

α 1
a Dt x(t) = ln x(t) + B(α) ln tẋ(t)
Γ(1 − α) a a
∞
" #
t 1−α−p t −α

X Γ(p + α − 1)
− C(α, p) ln Vp (t) − ln x(t)
p=2
a Γ(α)Γ(1 − α)(p − 1)! a
with
 
∞
1 1 +
X Γ(p + α − 1) ,
B(α) =
Γ(2 − α) Γ(α − 1)p!
p=1
Γ(p + α − 1)
C(α, p) = ,
Γ(−α)Γ(1 + α)(p − 1)!
Z t
τ p−2 x(τ )
Vp (t) = (1 − p) ln dτ.
a a τ
Proof. We rewrite (19) as
t −α t −α
Z t
α x(a) 1 1
a Dt x(t) = ln + ln τ ẋ(τ )dτ
Γ(1 − α) a Γ(1 − α) a τ τ
and then integrating by parts gives
t −α t 1−α

α x(a) aẋ(a)
a Dt x(t) = ln + ln
Γ(1 − α) a Γ(2 − α) a
Z t 1−α
1 t
+ ln [ẋ(τ ) + τ ẍ(τ )]dτ.
Γ(2 − α) a τ
1−α
Now we use the following expansion for ln τt , using the binomial theorem,
t 1−α t 1−α ln τa 1−α

ln = ln 1−
τ a ln at
∞ p
t 1−α X Γ(p − 1 + α) ln τa

= ln · .
t p
a Γ(α − 1)p! ln a
p=0
This implies that
t −α t 1−α t 1−α

α x(a) aẋ(a) 1
a Dt x(t) = ln + ln + ln
Γ(1 − α) a Γ(2 − α) a Γ(2 − α) a
∞ −p t
X Γ(p − 1 + α) t τ p
Z
× ln ln [ẋ(τ ) + τ ẍ(τ )]dτ.
Γ(α − 1)p! a a a
p=0
Extracting the first term of the infinite sum, simplifications and another integration by parts
p p−1
using u = ln τa , du = (p) τ1 ln τa and dv = [ẋ(τ ) + τ ẍ(τ )]dτ , v = τ ẋ(τ ) yields
t −α t 1−α t 1−α

α x(a) 1
a Dt x(t) = ln + B(α) ln tẋ(t) − ln
Γ(1 − α) a a Γ(2 − α) a
∞ −p t
X Γ(p − 1 + α) t τ p−1
Z
× ln ln ẋ(τ )dτ.
Γ(α − 1)(p − 1)! a a a
p=1
A final step of extracting the first term in the sum and integration by parts finishes the
proof.
For practical purposes, finite sums up to order N are considered and the approximation
becomes
t −α t 1−α

α
a Dt x(t) ≈ A(α, N ) ln x(t) + B(α, N ) ln tẋ(t)
a a
N
t 1−α−p
X
+ C(α, p) ln Vp (t) (20)
a
p=2
with
 
N
1 X Γ(p + α − 1)
A(α, N ) = 1 + ,
Γ(1 − α) Γ(α)(p − 1)!
p=2
 
N
1 1 +
X Γ(p + α − 1) 
B(α, N ) = .
Γ(2 − α) Γ(α − 1)p!
p=1
Remark 2.11. The right Hadamard fractional derivative can be expanded in the same way.
This gives the following approximation:
−α 1−α
α b b
t Db x(t) ≈ A(α, N ) ln x(t) − B(α, N ) ln tẋ(t)
t t
N 1−α−p
X b
− C(α, p) ln Wp (t)
t
p=2
with
b p−2 x(τ )
Z b
Wp (t) = (1 − p) ln dτ.
t τ τ
2.2.3. Examples
In this section we apply (20) to compute fractional derivatives, of order α = 21 , for x(t) = t4
and x(t) = ln(t). The exact Hadamard fractional derivative is available for x(t) = t4 and
we have √
0.5 4 ln t
1 Dt (t ) = .
Γ(1.5)
For x(t) = ln(t), only an approximation of the Hadamard fractional derivative is found in
the literature:
0.5 1 0.5908179503 9 √
1 Dt ln(t) ≈ √ + 9t erf(3 ln t).
Γ(0.5) ln t Γ(0.5)
The results of applying (20) to evaluate fractional derivatives are depicted in Figure 4.
1.8 35
Analytic Analytic
1.6 N=3, E=7.1715e−016 N=3, E=0.72399
30 N=4, E=0.38
N=5, E=0.22964
1.4
25
1.2
1 20
0.8 15
0.6
10
0.4
5
0.2
0 0
1 2 3 4 5 6 7 8 9 10 1 1.2 1.4 1.6 1.8 2
t t
(a) 1 Dt0.5 (ln t) (b) 1 Dt0.5 (t4 )
Figure 4. Analytic versus numerical approximation (20).

2.2.4. Error Analysis

When we approximate an infinite series by a finite sum, the choice of the order of approx-
imation is a key question. Having an estimate knowledge of truncation errors, one can
choose properly up to which order the approximations should be made to suit the accuracy
requirements. In this section we study the errors of the approximations presented so far.
Separation of an error term in (8) concludes in
N
!
t
1 d (−1)k x(k) (t)
Z X
α −α k
a Dt x(t) = (t − τ ) (t − τ ) dτ
Γ(1 − α) dt a k!
k=0
∞
!
t
1 d (−1)k x(k) (t)
Z X
+ (t − τ )−α (t − τ )k dτ. (21)
Γ(1 − α) dt a k!
k=N +1
The first term in (21) gives (9) directly and the second term is the error caused by truncation.
The next step is to give a local upper bound for this error, Etr (t).
The series
∞
X (−1)k x(k) (t)
(t − τ )k , τ ∈ (a, t), t ∈ (a, b),
k!
k=N +1

is the remainder of the Taylor expansion of x(τ ) and thus bounded by (NM (t τ ) N +1

+1)! −
in which
M = max |x(N +1) (τ )|.
τ ∈[a,t]
Then,
d t

M M
Z
N +1−α
(t−a)N +1−α .

Etr (t) ≤
(t − τ ) dτ =
Γ(1 − α)(N + 1)! dt a Γ(1 − α)(N + 1)!
In order to estimate a truncation error for approximation (15), the expansion procedure
is carried out with separation of N terms in binomial expansion as
∞
τ − a 1−α Γ(p − 1 + α) τ − a p
X
1− =
t−a Γ(α − 1)p! t−a
p=0
N
Γ(p − 1 + α) τ − a p
X
= + RN (τ ), (22)
p=0
Γ(α − 1)p! t−a
where
∞
Γ(p − 1 + α) τ − a p
X
RN (τ ) = .
Γ(α − 1)p! t−a
p=N +1
Substituting (22) into (13), we get
α x(a) ẋ(a)
a Dt x(t) = (t − a)−α + (t − a)1−α
Γ(1 − α) Γ(2 − α)
 
N
(t − a)1−α t X Γ(p − 1 + α) τ − a p
Z
+ + RN (τ ) ẍ(τ )dτ
Γ(2 − α) a Γ(α − 1)p! t−a
p=0
x(a) ẋ(a)
= (t − a)−α + (t − a)1−α
Γ(1 − α) Γ(2 − α)
 
N
(t − a)1−α t X Γ(p − 1 + α) τ − a p 
Z
+ ẍ(τ )dτ
Γ(2 − α) a Γ(α − 1)p! t−a
p=0
1−α Z t
(t − a)
+ RN (τ )ẍ(τ )dτ.
Γ(2 − α) a
At this point, we apply the techniques of [8] to the first three terms with finite sums. Then,
we receive (15) with an extra term of truncation error:
(t − a)1−α t
Z
Etr (t) = RN (τ )ẍ(τ )dτ.
Γ(2 − α) a
τ −a
Since 0 ≤ t−a ≤ 1 for τ ∈ [a, t], one has
∞ ∞ ∞ 2
e(1−α) +1−α

X Γ(p − 1 + α) X 1−α X
|RN (τ )| ≤ = ≤
Γ(α − 1)p! p p2−α
p=N +1 p=N +1 p=N +1
∞ 2 2
e(1−α) +1−α e(1−α) +1−α
Z
≤ dp = .
p2−α
p=N (1 − α)N 1−α

Finally, assuming L2 = max x(2) (τ ), we conclude that

τ ∈[a,t]
2
e(1−α) +1−α
|Etr (t)| ≤ L2 (t − a)2−α .
Γ(2 − α)(1 − α)N 1−α
Remark 2.12. Following similar techniques, one can extract an error bound for the ap-
proximations of Hadamard derivatives. When we consider finite sums in (20), the error is
bounded by
2
e(1−α) +1−α t 1−α

|Etr (t)| ≤ L(t) ln (t − a),
Γ(2 − α)(1 − α)N 1−α a
where
L(t) = max |ẋ(τ ) + τ ẍ(τ )|.
τ ∈[a,t]
3. Direct Methods
There are two main classes of direct methods in the classical calculus of variations and
optimal control. On the one hand, we specify a discretization scheme by choosing a set of
mesh points on the horizon of interest, say a = t0 , t1 , . . . , tn = b for [a, b]. Then we use
some approximations for derivatives in terms of unknown function values at ti and, using
an appropriate quadrature, the problem is transformed to a finite dimensional optimization
problem. This method is known as Euler’s method in the literature [15]. Regarding Figure 5,
the solid line is the function that we are looking for, nevertheless, the method gives the
polygonal dashed line as an approximate solution.
x(tn )
xn−1
x2
x1 xi
x(t0 )
h
t
t0 t1 t2 ti tn−1 tn
Figure 5. Euler’s finite differences method.
On the other hand, there is the Ritz method, that has an extension to functionals of
several independent variables which is called Kantorovich’s method. We assume that the
admissible functions can be expanded in some kind of series, e.g., power or Fourier’s series,
of the form
X∞
x(t) = ak φk (t).
k=0
Using a finite number of terms in the sum as an approximation, and some sort of quadrature
again, the original problem can be transformed to an equivalent optimization problem for
ak , k = 0, 1, . . . , n.
In the presence of fractional operators, the same ideas are applied to discretize a prob-
lem. Many works can be found in the literature that use different types of basis functions to
establish Ritz-like methods for fractional calculus of variations and optimal control.
3.1. Euler-like Methods

The Euler method in the classical theory of the calculus of variations uses finite differences
approximations for derivatives and is referred also as the method of finite differences. The
basic idea of this method is that instead of considering the values of a functional
Z b
J[x(·)] = L(t, x(t), ẋ(t))dt
a
with boundary conditions x(a) = xa and x(b) = xb , on arbitrary admissible curves,

we only track the values at an n + 1 grid points, ti , i = 0, . . . , n, of the inter-
ested time interval [29]. The functional J[x(·)] is then transformed into a function
Ψ(x(t1 ), x(t2 ), . . . , x(tn−1 )) of the values of unknown function on mesh points. Assuming
h = ti − ti−1 , x(ti ) = xi and ẋi ≈ xi −x h
i−1
, one has
n
X xi − xi−1
J[x(·)] ≈ Ψ(x1 , x2 , . . . , xn−1 ) = h L ti , x i , ,
h
i=1
x0 = xa , xn = x b .
The desired values of xi , i = 1, . . . , n − 1, are the extremum of the multi-variable function

Ψ which is the solution to the system
∂Ψ
= 0, i = 1, . . . , n − 1.
∂xi
The fact that only two terms in the sum, (i − 1)th and ith, depend on xi , makes it rather
easy to find the extremum of Ψ solving a system of algebraic equations. For each n, we
obtain a polygonal line which is an approximate solution of the original problem. It has
been shown that passing to the limit as h → 0, the linear system corresponding to finding
the extremum of Ψ is equivalent to the Euler–Lagrange equation of the problem.
3.1.1. Finite Differences for Fractional Derivatives
In classical theory, given a derivative of a certain order, x(n) , there is a finite difference
approximation of the form
n
1 X k n
x(n) (t) = lim n (−1) x(t − kh),
h→0+ h k
k=0
n

where k is the binomial coefficient and

n n(n − 1)(n − 2) · · · (n − k + 1)
= , n, k ∈ N.
k k!
The Grünwald–Letnikov definition of fractional derivative is a generalization of this for-
mula to derivatives of arbitrary order.
The series in (2) and (3), the Grünwald–Letnikov definitions, converge absolutely and
uniformly if x(·) is bounded. The infinite sums, backward differences for the left and for-
ward differences for the right derivative in the Grünwald–Letnikov definitions for fractional
derivatives, reveals that the arbitrary order derivative of a function at a time t depends on
all values of that function in (−∞, t] and [t, ∞), for left and right derivatives respectively.
This is due to the non-local property of fractional derivatives.
Remark 3.1. Equations (2) and (3) need to be consistent in closed time intervals and we
need the values of x(t) outside the interval [a, b]. To overcome this difficulty, we can take

∗ x(t) t ∈ [a, b],
x (t) =
0 t∈/ [a, b].
Then we assume GL α GL α ∗ GL α GL α ∗
a Dt x(t) = a Dt x (t) and t Db x(t) = t Db x (t) for t ∈ [a, b].
This definition coincides with Riemann–Liouville and Caputo derivatives. The latter is
believed to be more applicable in practical fields such as engineering and physics.
Proposition 3.2 (See [25]). Let 0 < α < n, n ∈ N and x(·) ∈ C n−1 [a, b]. Suppose
also that x(n) (·) is integrable on [a, b]. Then, for every α, the Riemann–Liouville derivative
exists and coincides with the Grünwald–Letnikov derivative and the following holds:
n−1 t
x(i) (a)(t − a)i−α 1
X Z
α
a Dt x(t) = + (t − τ )n−1−α x(n) (τ )dτ
Γ(1 + i − α) Γ(n − α) a
i=0
GL α
= a Dt x(t).
Remark 3.3. For numerical purposes we need a finite series in (2). Given a grid on [a, b]
as a = t0 , t1 , . . . , tn = b, where ti = t0 + ih for some h > 0, we approximate the left
Riemann–Liouville derivative as
i
α 1 X α
a Dt x(ti ) ≈ α (ωk ) x(ti − kh), (23)
h
k=0
Γ(k−α)
where (ωkα ) = (−1)k αk = Γ(−α)Γ(k+1)

.
Similarly, one can approximate the right Riemann–Liouville derivative by
n−i
α 1 X α
t Db x(ti ) ≈ (ωk ) x(ti + kh). (24)
hα
k=0
Remark 3.4. The Grünwald–Letnikov approximation of Riemann–Liouville is a first order

approximation [25], i.e.,
i
α 1 X α
a Dt x(ti ) = (ωk ) x(ti − kh) + O(h).
hα
k=0
Remark 3.5. It has been shown that the implicit Euler method solution to a certain frac-
tional partial differential equation based on the Grünwald–Letnikov approximation to the
fractional derivative, is unstable [23]. Therefore, discretizing fractional derivatives, shifted
Grünwald–Letnikov derivatives are used and, despite the slight difference, they exhibit a
stable performance, at least for certain cases. The shifted Grünwald–Letnikov derivative is
defined by
i
sGL α 1 X α
a D t x(t i ) ≈ (ωk ) x(ti − (k − 1)h).
hα
k=0
Other finite difference approximations can be found in the literature. We refer here
to the Diethelm backward finite difference formula for Caputo’s fractional derivative, with
0 < α < 2 and α 6= 1, which is an approximation of order O(h2−α ) [16]:
 
−α i ⌊α⌋ k k
C α h X X (i − j) h (k) 
a Dt x(ti ) ≈ ai,j xi−j − x (a) ,
Γ(2 − α) k!
j=0 k=0
where 
 1, if i = 0,
ai,j = (j + 1) 1−α − 2j 1−α + (j − 1) 1−α , if 0 < j < i,
(1 − α)i−α − i1−α + (i − 1)1−α , if j = i.

3.1.2. Euler-like Direct Method for Fractional Variational Problems

As mentioned earlier, we consider a simple version of fractional variational problems where
the fractional term has a Riemann–Liouville form on a finite time interval [a, b]. The bound-
ary conditions are given and we approximate the problem using the Grünwald–Letnikov ap-
proximation given by (23). In this context, we discretize the functional in (4) using a simple
quadrature rule on the mesh points, a = t0 , t1 , , . . . , tn = b, with h = b−a n . The goal is to
find the values x1 , x2 , . . . , xn−1 of the unknown function x(·) at points ti , i = 1, . . . , n − 1.
The values of x0 and xn are given. Applying the quadrature rule gives
n Z
X ti n
X
J[x(·)] = L(ti , xi , a Dtα xi )dt ≈ hL(ti , xi , a Dtα xi )
i=1 ti−1 i=1
and by approximating the fractional derivatives at mesh points using (23) we have
n i
!
X 1 X α
J[x(·)] ≈ hL ti , xi , α (ωk ) xi−k . (25)
h
i=1 k=0
Hereafter the procedure is the same as in the classical case. The right-hand-side of (25) can
be regarded as a function Ψ of n − 1 unknowns x = (x1 , x2 , . . . , xn−1 ),
n i
!
X 1 X α
Ψ(x) = hL ti , xi , α (ωk ) xi−k . (26)
h
i=1 k=0
To find an extremum for Ψ, one has to solve the following system of algebraic equa-
tions:
∂Ψ
= 0, i = 1, . . . , n − 1. (27)
∂xi
Unlike the classical case, all terms, starting from the ith term in (26), depend on xi and we
have
n−i
∂Ψ ∂L X 1 ∂L
= h (ti , xi , a Dtα xi ) + h (ωkα ) (ti+k , xi+k , a Dtα xi+k ). (28)
∂xi ∂x h α ∂ a Dtα x
k=0
Equating the right-hand-side of (28) with zero, one has

n−i
∂L 1 X α ∂L
(ti , xi , a Dtα xi ) + α (ωk ) (ti+k , xi+k , a Dtα xi+k ) = 0.
∂x h ∂ a Dtα x
k=0
Passing to the limit, and considering the approximation formula for the right Riemann–
Liouville derivative, equation (24), it is straightforward to verify that:
Theorem 3.6. The Euler-like method for a fractional variational problem of the form (4) is
equivalent to the fractional Euler–Lagrange equation
∂L ∂L
+ t Dbα = 0,
∂x ∂ a Dtα x
as the mesh size, h, tends to zero.
Proof. Consider a minimizer (x1 , . . . , xn−1 ) of Ψ, a variation function η ∈ C[a, b] with
η(a) = η(b) = 0 and define ηi = η(ti ), for i = 0, . . . , n. We remark that η0 = ηn = 0
and that (x1 + ǫη1 , . . . , xn−1 + ǫηn−1 ) is a variation of (x1 , . . . , xn−1 ), with |ǫ| < r, for
some fixed r > 0. Therefore, since (x1 , . . . , xn−1 ) is a minimizer for Ψ, proceeding with
Taylor’s expansion, we deduce that
0 ≤ Ψ(x1 + ǫη1 , . . . , xn−1 + ǫηn−1 ) − Ψ(x1 , . . . , xn−1 )

n i
" #
X ∂L ∂L 1 X α
= ǫ h [i]ηi + [i] (ωk )ηi−k + O(ǫ),
∂x ∂ a Dtα hα
i=1 k=0
where
i
!
1 X α
[i] = ti , x i , α (ωk )xi−k .
h
k=0
Since ǫ takes any value, it follows that
n i
" #
X ∂L ∂L 1 X α
h [i]ηi + [i] (ωk )ηi−k = 0. (29)
∂x ∂ a Dtα hα
i=1 k=0
On the other hand, since η0 = 0, reordering the terms of the sum, it follows immediately
that
n i n n−i
X ∂L X X X ∂L
[i] α
(ωk )ηi−k = η i (ωkα ) [i + k].
∂ a Dtα ∂ a Dtα
i=1 k=0 i=1 k=0
Substituting this relation into equation (29), we obtain
n n−i
" #
X ∂L 1 X α ∂L
ηi h [i] + α (ωk ) [i + k] = 0.
∂x h ∂ a Dtα
i=1 k=0
Since ηi is arbitrary, for i = 1, . . . , n − 1, we deduce that

n−i
∂L 1 X α ∂L
[i] + α (ωk ) [i + k] = 0, for i = 1, . . . , n − 1.
∂x h ∂ a Dtα
k=0
Let us study the case when n goes to infinity. Let t ∈]a, b[ and i ∈ {1, . . . , n} such that
ti−1 < t ≤ ti . First observe that, in such case, we also have i → ∞ and n − i → ∞. In
fact, let i ∈ {1, . . . , n} be such that
a + (i − 1)h < t ≤ a + ih.
So, i < (t − a)/h + 1, which implies that
b−t
n−i>n − 1.
b−a
Then
lim ti = t.
n→∞,i→∞
Assume that there exists a function x ∈ C[a, b] satisfying
∀ǫ > 0 ∃N ∀n ≥ N : |xi − x(ti )| < ǫ, ∀i = 1, . . . , n − 1.
As x is uniformly continuous, we have
∀ǫ > 0 ∃N ∀n ≥ N : |xi − x(t)| < ǫ, ∀i = 1, . . . , n − 1.
By the continuity assumption of x, we deduce that

n−i
1 X α ∂L ∂L
lim (ωk ) α [i + k] = t Dbα (t, x(t), a Dtα x(t)).
n→∞,i→∞ h α ∂ a Dt ∂ a Dtα
k=0
For n sufficiently large (and therefore i also sufficiently large),

∂L ∂L
lim [i] = (t, x(t), a Dtα x(t)).
n→∞,i→∞ ∂x ∂x
In conclusion,
∂L ∂L
(t, x(t), a Dtα x(t)) + t Dbα (t, x(t), a Dtα x(t)) = 0. (30)
∂x ∂ a Dtα
Using the continuity condition, we prove that the fractional Euler–Lagrange equation (30)
holds for all values on the closed interval a ≤ t ≤ b.
3.1.3. Examples
Now we apply the Euler-like direct method to some test problems for which the exact
solutions are known. Although we propose problems for the interval [0, 1], moving to
arbitrary intervals is only a matter of more computations. To measure the errors related to
approximations, different norms can be used. Since a direct method seeks for the function
values at certain points, we use the maximum norm to determine how close we can get to
the exact value at that point. Assume that the exact value of the function x(·), at the point
ti , is x(ti ) and it is approximated by xi . The error is defined as
E = max{|x(ti ) − xi |, i = 1, · · · , n − 1}.
Example 3.7. Our goal here is to minimize a quadratic Lagrangian on [0, 1] with fixed
boundary conditions. Consider the following minimization problem:
( R1 2
2
J[x(·)] = 0 0 Dt0.5 x(t) − Γ(2.5) t1.5 dt → min
(31)
x(0) = 0, x(1) = 1.
Since the Lagrangian is always positive, problem (31) attains its minimum when
0.5 2
0 Dt x(t) − t1.5 = 0
Γ(2.5)
2 1.5
and has the obvious solution of the form x(t) = t2 because 0 Dt0.5 t2 = Γ(2.5) t .
To begin with, we approximate the fractional derivative by
i
0.5 1 X 0.5
0 Dt x(ti ) ≈ ωk x(ti − kh)
h0.5
k=0
for a fixed h > 0. The functional is now transformed into

Z 1 i
!2
1 X 0.5 2
J[x(·)] = ωk xi−k − t1.5 dt.
0 h0.5 Γ(2.5)
k=0
Finally, we approximate the integral by a rectangular rule and end with the discrete problem
n i
!2
X 1 X 0.5 2 1.5
Ψ(x) = h ωk xi−k − t .
h0.5 Γ(2.5) i
i=1 k=0
Since the Lagrangian in this example is quadratic, system (27) has a linear form and there-
fore is easy to solve. Other problems may end with a system of nonlinear equations. Simple
calculations lead to the system
Ax = b, (32)
in which
 Pn−1 2 Pn−1 Pn−1
Ai i=1 Ai Ai−1 Ai Ai−(n−2)

···
Pi=0
n−2 P n−2 2 Pi=n−2
n−2
A
 Pi=0 i i+1
 A A i · · · Ai Ai−(n−3) 
n−3 Pi=1 n−3 Pi=n−3
n−3

A= i=0 Ai Ai+2 i=1 Ai Ai+1 ··· i=n−4 Ai Ai−(n−4) ,
 
 . . . .
 .. .. . . ..


P1 P1 P1 2
i=0 Ai Ai+n−2 i=0 Ai Ai+n−3 · · · i=0 Ai
where Ai = (−1)i h1.5 0.5

i and b = (b1 , b2 , · · · , bn−1 ) with
n−i n−i
!
X 2h2 Ak 1.5 X
bi = t − An−i A0 − Ak Ak+i .
Γ(2.5) k+i
k=0 k=0
Since system (32) is linear, it is easily solved for different values of n. As indicated in
Figure 6, by increasing the value of n we get better solutions.
Let us now move to another example for which the solution is obtained by the fractional
Euler–Lagrange equation.
1
Analytic solution
0.9 Approximation: n = 5, Error= 0.03
Approximation: n = 10, Error= 0.02
0.7
0.6
0.35
x(t)
0.5
0.3
0.4
0.3 0.25
0.5 0.55 0.6
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
t
Figure 6. Analytic and approximate solutions of Example 3.7.
Example 3.8. Consider the following minimization problem:

R1
J[x(·)] = 0 0 Dt0.5 x(t) − ẋ2 (t) dt → min

(33)
x(0) = 0, x(1) = 1.
In this case the only way to get a solution is by use of Euler–Lagrange equations. The La-
grangian depends not only on the fractional derivative, but also on the first order derivative
of the function. The Euler–Lagrange equation for this setting becomes

∂L α ∂L d ∂L
+ t Db − = 0,
∂x ∂ a Dtα dt ∂ ẋ
and by direct computations a necessary condition for x(·) to be a minimizer of (33) is
α 1
t D1 1 + 2ẍ(t) = 0 or ẍ(t) = (1 − t)−α .
2Γ(1 − α)
Subject to the given boundary conditions, the above second order ordinary differential
equation has the solution

1 1 1
x(t) = − (1 − t)2−α + 1 − t+ . (34)
2Γ(3 − α) 2Γ(3 − α) 2Γ(3 − α)
Discretizing problem (33) with the same assumptions of Example 3.7 ends in a linear
1
Analytic solution
Approximation: n = 10, Error= 0.003
0.7
0.6
x(t)
0.5
0.4 0.58
0.3 0.57
0.2 0.56
0.1
0.51 0.52 0.53
0
0 0.2 0.4 0.6 0.8 1
t
system of the form

    
2 −1 0 0 ··· 0 0 x1 b1
 −1 2 −1 0 ··· 0 0  x2   b2 
    
 0 −1 2 −1 0 0 x3 b3
··· = , (35)
   
 
 .. .. .. .. .. .. ..  ..   .. 
 . . . . . . .  .   . 
0 0 0 0 ··· −1 2 xn−1 bn−1
where
n−i−1
h X k 0.5 0.5
bi = (−1) h , i = 1, 2, . . . , n − 2,
2 k
k=0
and
1
hX k 0.5 0.5
bn−1 = (−1) h + xn .
2 k
k=0
System (35) is linear and can be solved for any n to reach the desired accuracy. The analytic
solution together with some approximated solutions are shown in Figure 7.
Both examples above end with linear systems and their solvability is simply dependant
to the matrix of coefficients. Now we try this method on a more complicated problem, yet
analytically solvable, with an oscillating solution.
R1
Example 3.9. Consider the problem of minimizing 0 Ldt subject to the boundary condi-
tions x(0) = 0 and x(1) = 1, where the Lagrangian L is given by
4
16Γ(6) 4.5 20Γ(4) 2.5 5
L= 0.5
0 Dt x(t) − t + t − t0.5 .
Γ(5.5) Γ(3.5) Γ(1.5)
This example has an obvious solution too. Since L is positive, the minimizer is
x(t) = 16t5 − 20t3 + 5t.

Γ(ν+1) ν−α
Note that a Dtα (t − a)ν = Γ(ν+α) t .
The appearance of a fourth power in the Lagrangian, results in a nonlinear system as

we apply the Euler-like direct method to this problem. For j = 1, · · · , n − 1 we have
n i
!3
X
0.5
1 X 0.5
ωi−j ωk xi−k − φ(ti ) = 0, (36)
h0.5
i=j k=0
where
16Γ(6) 4.5 20Γ(4) 2.5 5
φ(t) = t + t − t0.5 .
Γ(5.5) Γ(3.5) Γ(1.5)
System (36) is solved for different values of n and the results are depicted in Figure 8.
0.5
−0.5
x(t)
−1
−1.5
−2 Analytic
Approximation: n = 5, E= 1.48e+000
−2.5 Approximation: n = 20, E= 3.01e−001
Approximation: n = 90, E= 6.18e−002
0 0.2 0.4 0.6 0.8 1

t

4. Indirect Methods
As in the classical case, indirect methods in fractional sense provide the necessary condi-
tions of optimality using the first variation. Fractional Euler–Lagrange equations are now
a well-known and well-studied subject in fractional calculus. For a simple problem of the
form (4), following [1], a necessary condition implies that the solution must satisfy a frac-
tional boundary value differential equation.
Theorem 4.1 (cf. [1]). Let x(·) have a continuous left Riemann–Liouville derivative of
order α and J be a functional of the form
Z b
J[x(·)] = L(t, x(t), a Dtα x(t))dt (37)
a
subject to the boundary conditions x(a) = xa and x(b) = xb . Then a necessary condition
for J to have an extremum for a function x(·) is that x(·) satisfies the following Euler-
Lagrange equation: (
∂L α ∂L
∂x + t Db ∂ a Dtα x = 0, (38)
x(a) = xa , x(b) = xb ,
which is called the fractional Euler–Lagrange equation.
Proof. Assume that x∗ (t) is the desired function and let x(t) = x∗ (t) + ǫη(t) be a family
of curves that satisfy boundary conditions, i.e., η(a) = η(b) = 0. Since a Dtα is a linear
operator, for any x(·), the functional becomes
Z b
J[x(·)] = L(t, x∗ (t) + ǫη(t), a Dtα x∗ (t) + ǫa Dtα η(t))dt,
a
dJ

which is a function of ǫ, J[ǫ]. Since J assumes its extremum at ǫ = 0, one has dǫ ǫ=0 = 0,
i.e., Z b
∂L ∂L α
η+ a D η dt = 0.
a ∂x ∂ a Dtα x t
Using the fractional integration by parts of the form
Z b Z b
g(t)a Dtα f (t)dt = f (t)t Dbα g(t)dt
a a
on the second term and applying the fundamental theorem of the calculus of variations
completes the proof.
Remark 4.2. Many variants of this theorem can be found in the literature. Different types
of fractional terms have been embedded in the Lagrangian and appropriate versions of
Euler–Lagrange equations have been derived using proper integration by parts formulas.
See [1, 3, 6, 22, 24] for details.
For fractional optimal control problems, a so-called Hamiltonian system is constructed

using Lagrange multipliers. For example, cf. [9], assume that we are required to minimize
a functional of the form
Z b
J[x(·), u(·)] = L(t, x(t), u(t))dt
a
such that x(a) = xa , x(b) = xb and a Dtα x(t) = f (t, x(t), u(t)). Similar to the classical
methods, one can introduce a Hamiltonian
H = L(t, x(t), u(t)) + λ(t)f (t, x(t), u(t)),
where λ(t) is considered as a Lagrange multiplier. In this case we define the augmented
functional as
Z b
J[x(·), u(·), λ(·)] = [H(t, x(t), u(t), λ(t)) − λ(t)a Dtα x(t)]dt.
a
Optimizing the latter functional results in the following necessary optimality conditions:
∂H
 a Dtα x(t) = ∂λ

α ∂H
D λ(t) = ∂x (39)
 t∂Hb
∂u = 0.
Together with the prescribed boundary conditions, this makes a two point fractional bound-
ary value problem.
These arguments reveal that, like the classical case, fractional variational problems end
with fractional boundary value problems. To reach an optimal solution, one needs to deal
with a fractional differential equation or a system of fractional differential equations.
The classical theory of differential equations is furnished with several solution meth-
ods, theoretical and numerical. Nevertheless, solving a fractional differential equation is a
rather tough task [12]. To benefit those methods, especially all solvers that are available to
solve an integer order differential equation numerically, we can either approximate a frac-
tional variational problem by an equivalent integer-order one or approximate the necessary
optimality conditions (38) and (39). The rest of this section discusses two types of approx-
imations that are used to transform a fractional problem to one in which only integer order
derivatives are present; i.e., we approximate the original problem by substituting a frac-
tional term by its corresponding expansion formulas. This is mainly done by case studies
on certain examples. The examples are chosen so that either they have a trivial solution or
it is possible to get an analytic solution using fractional Euler–Lagrange equations.
By substituting the approximations (9) or (15) for the fractional derivative in (37), the
problem is transformed to
N
Z b !
X (−1)k−1 αx(k) (t)
J[x(·)] = L t, x(t), (t − a)k−α dt
a k!(k − α)Γ(1 − α)
k=0
Z b
= L′ t, x(t), ẋ(t), . . . , x(N ) (t) dt
a
or
 
b N
Ax(t) B ẋ(t) C(α, p)Vp (t) 
Z X
J[x(·)] = L t, x(t), α
+ − dt
a (t − a) (t − a)α−1 (t − a)p+α−1
p=2
Z b
= L′ (t, x(t), ẋ(t), V2 (t), . . . , VN (t)) dt
a
V̇p (t) = (1 − p)(t − a)p−2 x(t)
Vp (a) = 0, p = 2, 3, . . .
The former problem is a classical variational problem containing higher order derivatives.
The latter is a multi-variable problem, subject to some ordinary differential equation con-
straint. Together with the boundary conditions, both above problems belong to classes of
well studied variational problems.
To accomplish a detailed study, as test problems, we consider here Example 3.8,
R1
J[x(·)] = 0 0 Dt0.5 x(t) − ẋ2 (t) dt → min

(40)
x(0) = 0, x(1) = 1,
and the following example.
Example 4.3. Given α ∈ (0, 1), consider the functional

Z 1
J[x(·)] = (a Dtα x(t) − 1)2 dt (41)
0
1
to be minimized subject to the boundary conditions x(0) = 0 and x(1) = Γ(α+1) . Since the
α
integrand in (41) is non-negative, the functional attains its minimum when a Dt x(t) = 1,
tα
i.e., for x(t) = Γ(α+1) .
We illustrate the use of the two different expansions separately.
4.1. Expansion to Integer Orders

Using approximation (9) for the fractional derivative in (40), we get the approximated prob-
lem
Z 1 "X N
#
˜
min J[x(·)] = C(n, α)tn−α x(n) (t) − ẋ2 (t) dt
0 n=0
(42)
x(0) = 0, x(1) = 1,
which is a classical higher-order problem of the calculus of variations that depends on
derivatives up to order N . The corresponding necessary optimality condition is a well-
known result.
Theorem 4.4 (cf., e.g., [21]). Suppose that x(·) ∈ C 2N [a, b] minimizes
Z b
L(t, x(t), x(1) (t), x(2) (t), . . . , x(N ) (t))dt
a
with given boundary conditions
x(a) = a0 , x(b) = b0 ,
(1)
x (a) = a1 , x(1) (b) = b1 ,
..
.
x(N −1) (a) = aN −1 , x(N −1) (b) = bN −1 .
Then x(·) satisfies the Euler–Lagrange equation
d2 N

∂L d ∂L ∂L N d ∂L
− + 2 − · · · + (−1) = 0. (43)
∂x dt ∂x(1) dt ∂x(2) dtN ∂x(N )
In general (43) is an ODE of order 2N , depending on the order N of the approximation
we choose, and the method leaves 2N − 2 parameters unknown. In our example, however,
the Lagrangian in (42) is linear with respect to all derivatives of order higher than two. The
resulting Euler–Lagrange equation is the second order ODE
N
X dn n−α d
(−1)n C(n, α) (t ) − [−2ẋ(t)] = 0
n=0
dtn dt
that has the solution

x(t) = M1 (α, N )t2−α + M2 (α, N )t,
where
"N #
1 X
n
M1 (α, N ) = − (−1) Γ(n + 1 − α)C(n, α) ,
2Γ(3 − α)
n=0
N
" #
1 X
M2 (α, N ) = 1 + (−1)n Γ(n + 1 − α)C(n, α) .
2Γ(3 − α)
n=0
Figure 9 shows the analytic solution together with several approximations. It reveals that
by increasing N , approximate solutions do not converge to the analytic one. The reason is
the fact that the solution (34) to Example 3.8 is not an analytic function. We conclude that
(9) may not be a good choice to approximate fractional variational problems. In contrast,
as we shall see, the approximation (15) leads to good results.
To solve Example 3.8 using (9) as an approximation for the fractional derivative, the
problem becomes
Z 1 XN
!2
˜
min J[x(·)] = C(n, α)tn−α x(n) (t) − 1 dt,
0 n=0
1
x(0) = 0, x(1) = .
Γ(α + 1)
The Euler–Lagrange equation (43) gives a 2N order ODE. For N ≥ 2 this approach is
1
inappropriate since the two given boundary conditions x(0) = 0 and x(1) = Γ(α+1) are not
enough to determine the 2N constants of integration.
1
Analytic
0.9 N=1
N=3
0.8 N=5
0.7
0.6
x(t)
0.5
0.4 0.7
0.3 0.68
0.2 0.66
0.1 0.62 0.64 0.66
0
0 0.2 0.4 0.6 0.8 1
t
Figure 9. Analytic versus approximate solutions to Example 3.8 using approximation (9)
with α = 0.5.
4.2. Expansion through the Moments of a Function

If we use (15) to approximate the optimization problem (40), with A = A(α, N ), B =
B(α, N ) and Cp = C(α, p), we have
 
Z 1 N
˜
X
J[x(·)] = At−α x(t) + Bt1−α ẋ(t) − Cp t1−p−α Vp (t) − ẋ2 (t) dt,
0 p=2
V̇p (t) = (1 − p)t p−2

x(t), p = 2, . . . , N, (44)
Vp (0) = 0, p = 2, . . . , N,
x(0) = 0, x(1) = 1.
Problem (44) is constrained with a set of ordinary differential equations and is natural to
look to it as an optimal control problem [26]. For that we introduce the control variable
u(t) = ẋ(t). Then, using the Lagrange multipliers λ1 , λ2 , . . . , λN , and the Hamiltonian
system, one can reduce (44) to the study of the two point boundary value problem
ẋ(t) = 12 Bt1−α − 21 λ1 (t),


 V̇ (t) = (1 − p)tp−2 x(t), p = 2, . . . , N,

p
−α −
PN (45)

 λ̇ 1 (t) = At p=2 (1 − p)tp−2 λp (t),
λ̇p (t) = −Cp t(1−p−α) , p = 2, . . . , N,

with boundary conditions

x(0) = 0, x(1) = 1,
Vp (0) = 0, p = 2, . . . , N, λp (1) = 0, p = 2, . . . , N,
where x(0) = 0 and x(1) = 1 are given. We have Vp (0) = 0, p = 2, . . . , N , due to (14)
and λp (1) = 0, p = 2, . . . , N , because Vp is free at final time for p = 2, . . . , N [26]. In
0.9 Analytic
N=2
0.8 N=5
N=10
0.7 N=16
0.6
x(t)
0.5
0.4 0.75
0.3
0.74
0.2
0.73
0.1 0.69 0.7 0.71
0
0 0.2 0.4 0.6 0.8 1
t
Figure 10. Analytic versus approximate solutions to Example 3.8 using approximation (15)
with α = 0.5.
general, the Hamiltonian system is a nonlinear, hard to solve, two point boundary value
problem that needs special numerical methods. In this case, however, (45) is a non-coupled
system of ordinary differential equations and is easily solved to give
 
N N
X C(α, p) X C(α, p) 
x(t) = M (α, N )t2−α − tp + 1 − M (α, N ) + t,
p=2
2p(2 − p − α) 2p(2 − p − α)
p=2
where
 
N
1 A(α, N ) X C(α, p)(1 − p)
M (α, N ) = B(α, N ) − − .
2(2 − α) 1−α (1 − α)(2 − p − α)
p=2
Figure 10 shows the graph of x(·) for different values of N .

Let us now approximate Example 4.3 using (15). The resulting minimization problem
has the following form:
 2
Z 1 N
˜
X
min J[x(·)] = At−α x(t) + Bt1−α ẋ(t) − Cp t1−p−α Vp (t) − 1 dt,
0 p=2
V̇p (t) = (1 − p)tp−2 x(t), p = 2, . . . , N, (46)

Vp (0) = 0, p = 2, . . . , N,
1
x(0) = 0, x(1) = .
Γ(α + 1)
Following the classical optimal control approach of Pontryagin [26], this time with
N
X
u(t) = At−α x(t) + Bt1−α ẋ(t) − Cp t1−p−α Vp (t),
p=2
1.4
Analytic: J=0
Approximation: N=2, J=0.0004116
1.2
0.8
x(t)
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
t
Figure 11. Analytic versus approximate solution to Example 4.3 using approximation (15)
with α = 0.5.
we conclude that the solution to (46) satisfies the system of differential equations
ẋ(t) = −AB −1 t−1 x(t) + N


−1 −p 1 −2 2α−2
λ1 (t) + B −1 tα−1 ,
P

 p=2 B Cp t Vp (t) + 2 B t
V̇p (t) = (1 − p)tp−2 x(t), p = 2, . . . , N,


λ̇1 (t) = AB −1 t−1 λ1 − N p−2 λ (t),
P

 p=2 (1 − p)t p
−1 −p

λ̇p (t) = −B Cp t λ1 , p = 2, . . . , N,

(47)
where A = A(α, N ), B = B(α, N ) and Cp = C(α, p) are defined according to Sec-
tion 2.1.2, subject to the boundary conditions
1
x(1) = Γ(α+1) ,

x(0) = 0,
(48)
Vp (0) = 0, p = 2, . . . , N, λp (1) = 0, p = 2, . . . , N.
The solution to system (47)–(48), with N = 2, is shown in Figure 11.
5. Conclusion
The realm of numerical methods in scientific fields is vastly growing due to the very fast
progresses in computational sciences and technologies. Nevertheless, the intrinsic complex-
ity of fractional calculus, caused partially by non-local properties of fractional derivatives
and integrals, makes it rather difficult to find efficient numerical methods in this field. It
seems enough to mention here that, up to the time of this manuscript, and to the best of
our knowledge, there is no routine available for solving a fractional differential equation as
Runge–Kutta for ordinary ones. Despite this fact, however, the literature exhibits a grow-
ing interest and improving achievements in numerical methods for fractional calculus in
general and fractional variational problems specifically.
This chapter was devoted to discuss some aspects of the very well-known methods for
solving variational problems. Namely, we studied the notions of direct and indirect methods
in the classical calculus of variations and we also mentioned some connections to optimal
control. Consequently, we introduced the generalizations of these notions to the field of
fractional calculus of variations and fractional optimal control.
The method of finite differences, as discussed here, seems to be a potential first candi-
date to solve fractional variational problems. Although a first order approximation was used
for all examples, the results are satisfactory and even though it is more complicated than in
the classical case, it still inherits some sort of simplicity and an ease of implementation.
The lack of efficient numerical methods for fractional variational problems is overcome,
partially, by the indirect methods of this chapter. Once we transformed the fractional vari-
ational problem to an approximated classical one, the majority of classical methods can
be applied to get an approximate solution. Nevertheless, the procedure is not completely
straightforward. The singularity of fractional operators is still present in the approximating
formulas and it makes the solution procedure more complicated.
Acknowledgments
Part of first author’s Ph.D., carried out at the University of Aveiro under the Doctoral
Program in Mathematics and Applications (PDMA) of Universities of Aveiro and Minho.
Work supported by FEDER funds through COMPETE — Operational Programme Factors
of Competitiveness (“Programa Operacional Factores de Competitividade”) and by Por-
tuguese funds through the Center for Research and Development in Mathematics and Appli-
cations (University of Aveiro) and the Portuguese Foundation for Science and Technology
(“FCT–Fundação para a Ciência e a Tecnologia”), within project PEst-C/MAT/UI4106/20-
11 with COMPETE number FCOMP-01-0124-FEDER-022690. Pooseh was also supported
by the FCT Ph.D. fellowship SFRH/BD/33761/2009; Torres by EU funding under the 7th
Framework Programme FP7-PEOPLE-2010-ITN, grant agreement no. 264735-SADCO.
References
[1] O. P. Agrawal, Formulation of Euler–Lagrange equations for fractional variational
problems, J. Math. Anal. Appl. 272 (2002), no. 1, 368–379.
[2] O. P. Agrawal, A general formulation and solution scheme for fractional optimal con-
trol problems, Nonlinear Dynam. 38 (2004), no. 1-4, 323–337.
[3] R. Almeida, S. Pooseh and D. F. M. Torres, Fractional variational problems depending

on indefinite integrals, Nonlinear Anal. 75 (2012), no. 3, 1009–1025.
[4] R. Almeida and D. F. M. Torres, Leitmann’s direct method for fractional optimization
problems, Appl. Math. Comput. 217 (2010), no. 3, 956–962.
[5] G. E. Andrews, R. Askey and R. Roy, Special functions, Encyclopedia of Mathematics

and its Applications, 71, Cambridge Univ. Press, Cambridge, 1999.
[6] T. M. Atanacković, S. Konjik and S. Pilipović, Variational problems with fractional

derivatives: Euler–Lagrange equations, J. Phys. A 41 (2008), no. 9, 095201, 12 pp.
[7] T. M. Atanacković and B. Stankovic, An expansion formula for fractional derivatives

and its application, Fract. Calc. Appl. Anal. 7 (2004), no. 3, 365-378.
[8] T. M. Atanacković and B. Stankovic, On a numerical scheme for solving differential

equations of fractional order, Mech. Res. Comm. 35 (2008), no. 7, 429–438.
[9] D. Baleanu, O. Defterli and O. P. Agrawal, A central difference numerical scheme for
fractional optimal control problems, J. Vib. Control 15 (2009), no. 4, 583–597.
[10] P. L. Butzer, A. A. Kilbas and J. J. Trujillo, Stirling functions of the second kind in the
setting of difference and fractional calculus, Numer. Funct. Anal. Optim. 24 (2003),
no. 7-8, 673–711.
[11] S. Das, Functional fractional calculus for system identification and controls, Springer,
Berlin, 2008.
[12] K. Diethelm, The analysis of fractional differential equations, Lecture Notes in Math-
ematics, 2004, Springer, Berlin, 2010.
[13] K. Diethelm, N. J. Ford and A. D. Freed, A predictor-corrector approach for the nu-
merical solution of fractional differential equations, Nonlinear Dynam. 29 (2002),
no. 1-4, 3–22.
[14] V. D. Djordjevic and T. M. Atanackovic, Similarity solutions to nonlinear heat con-

duction and Burgers/Korteweg-de Vries fractional equations, J. Comput. Appl. Math.
222 (2008), no. 2, 701–714.
[15] L. Elsgolts, Differential equations and the calculus of variations, translated from the
Russian by George Yankovsky, Mir Publishers, Moscow, 1973.
[16] N. J. Ford and J. A. Connolly, Comparison of numerical methods for fractional differ-
ential equations, Commun. Pure Appl. Anal. 5 (2006), no. 2, 289–306.
[17] Z. D. Jelicic and N. Petrovacki, Optimality conditions and a solution scheme for frac-
tional optimal control problems, Struct. Multidiscip. Optim. 38 (2009), no. 6, 571–581.
[18] A. A. Kilbas, Hadamard-type fractional calculus, J. Korean Math. Soc. 38 (2001),

no. 6, 1191–1204.
[19] A. A. Kilbas, H. M. Srivastava and J. J. Trujillo, Theory and applications of fractional

differential equations, North–Holland Mathematics Studies, 204, Elsevier, Amster-
dam, 2006.
[20] P. Kumar and O. P. Agrawal, An approximate method for numerical solution of frac-
tional differential equations, Signal Process. 86 (2006), 2602–2610.
[21] L. P. Lebedev and M. J. Cloud, The calculus of variations and functional analysis,
World Sci. Publishing, River Edge, NJ, 2003.
[22] A. B. Malinowska and D. F. M. Torres, Generalized natural boundary conditions for

fractional variational problems in terms of the Caputo derivative, Comput. Math. Appl.
59 (2010), no. 9, 3110–3116.
[23] M. M. Meerschaert and C. Tadjeran, Finite difference approximations for fractional

advection-dispersion flow equations, J. Comput. Appl. Math. 172 (2004), no. 1, 65–77.
[24] T. Odzijewicz, A. B. Malinowska and D. F. M. Torres, Fractional variational calculus

with classical and combined Caputo derivatives, Nonlinear Anal. 75 (2012), no. 3,
1507–1515.
[25] I. Podlubny, Fractional differential equations, Mathematics in Science and Engineer-

ing, 198, Academic Press, San Diego, CA, 1999.
[26] L. S. Pontryagin, V. G. Boltyanskii, R. V. Gamkrelidze and E. F. Mishchenko, The

mathematical theory of optimal processes, Translated from the Russian by K. N.
Trirogoff; edited by L. W. Neustadt Interscience Publishers John Wiley & Sons, Inc.
New York, 1962.
[27] S. Pooseh, R. Almeida and D. F. M. Torres, Expansion formulas in terms of integer-

order derivatives for the Hadamard fractional integral and derivative, Numer. Funct.
Anal. Optim. 33 (2012), no. 3, 301–319.
[28] S. Pooseh, R. Almeida and D. F. M. Torres, Approximation of fractional integrals by

means of derivatives, Comput. Math. Appl. 64 (2012), no. 10, 3090–3100.
[29] S. Pooseh, R. Almeida and D. F. M. Torres, Discrete direct methods in the fractional
calculus of variations, Comput. Math. Appl. 66 (2013), no. 5, 668–676.
[30] F. Riewe, Nonconservative Lagrangian and Hamiltonian mechanics, Phys. Rev. E (3)
53 (1996), no. 2, 1890–1899.
[31] F. Riewe, Mechanics with fractional derivatives, Phys. Rev. E (3) 55 (1997), no. 3, part
B, 3581–3592.
[32] S. G. Samko, A. A. Kilbas and O. I. Marichev, Fractional integrals and derivatives,

translated from the 1987 Russian original, Gordon and Breach, Yverdon, 1993.
[33] J. A. Tenreiro Machado, V. Kiryakova and F. Mainardi, Recent history of fractional

calculus, Commun. Nonlinear Sci. Numer. Simul. 16 (2011), no. 3, 1140–1153.
[34] J. A. Tenreiro Machado, M. F. Silva, R. S. Barbosa, I. S. Jesus, C. M. Reis, M. G.

Marcos and A. F. Galhano, Some applications of fractional calculus in engineering,
Math. Probl. Eng., Art. ID 639801 (2010) 34 pp.
INDEX
composition, 21, 22, 35, 70, 96, 113, 131

A computation, 66, 88, 89, 92, 128, 142, 160, 213
conduction, 177, 238
access, 70, 72, 180 conservation, 128, 143, 144, 145, 146, 152, 155, 159,
acoustics, 1 160, 161, 162, 163, 164, 175, 176
adaptation, 205 construction, viii, 23, 66, 70, 74, 77, 80, 85, 86, 87,
amplitude, 119 128, 139, 149, 150
anomalous diffusion, 63, 188, 198, 200 convention, 12, 20, 131
arrow of time, 69, 70, 116 convergence, 8, 10, 13, 62, 137, 152, 153, 154, 155,
asymmetry, 101, 146 193, 212
cost, 127, 133, 134, 140, 141, 142, 143, 149, 150
B
D
BAC, 36, 37
bedding, 118 damping, 69
Belarus, 175 deformation, 73
boundary value problem, 231, 234 degenerate, 49, 90
Brownian motion, 2, 24 derivatives, vii, 2, 3, 12, 15, 35, 37, 38, 40, 42, 43,
Brownian particle, 2 44, 45, 46, 49, 51, 58, 60, 61, 62, 66, 68, 70, 71,
72, 73, 86, 87, 90, 96, 97, 98, 99, 101, 104, 107,
109, 113, 116, 118, 119, 122, 124, 125, 127, 129,
C 130, 131, 137, 142, 146, 147, 148, 172, 174, 175,
177, 179, 180, 188, 189, 191, 198, 200, 201, 203,
calculus, vii, viii, 1, 2, 3, 34, 35, 44, 48, 61, 62, 63,
204, 205, 206, 207, 209, 212, 213, 214, 215, 217,
65, 66, 67, 68, 70, 71, 72, 75, 76, 77, 80, 83, 85,
219, 220, 221, 222, 223, 231, 232, 233, 236, 238,
87, 88, 93, 95, 97, 111, 112, 113, 123, 124, 127,
239
128, 129, 131, 133, 134, 135, 137, 138, 148, 150,
destruction, 120
174, 175, 176, 177, 179, 199, 200, 201, 204, 205,
determinism, 1
207, 220, 230, 232, 236, 237, 238, 239
differential equations, vii, 1, 2, 34, 62, 63, 66, 69, 74,
Cantor set, 2
80, 105, 123, 124, 128, 134, 138, 140, 146, 152,
case studies, 231
175, 176, 177, 205, 210, 231, 236, 238, 239
Cauchy problem, 132, 133, 134, 135, 136, 141, 142,
diffusion, 1, 63, 69, 113, 118, 120, 121, 122, 124,
143, 149, 150, 169
125, 175, 179, 180, 188, 199, 200
causality, 72, 105, 119, 123
diffusion process, 179
chaos, vii, 124, 176, 199, 200
diffusivity, 120
classes, 20, 24, 220, 232
dilation, 13
classical mathematical fields, vii
discrete data, 74
classical mechanics, 69, 73, 122, 174
discretization, 146, 152, 220
classical methods, 237
dispersion, 120, 176, 239
classification, 4
distribution, 73, 74, 85, 197, 198
closed ball, 207
divergence, 19, 95, 99, 113
closure, 4, 5
drawing, 61
coherence, 81, 84, 85, 93
dynamical systems, 122, 123, 127
complexity, 206, 236
242 Index
E H
electricity, 174 Hamiltonian, vii, viii, 63, 65, 66, 123, 124, 125, 135,
electromagnetic, 1 137, 142, 144, 145, 146, 151, 152, 156, 159, 160,
electron, 120 177, 179, 180, 181, 182, 188, 189, 197, 198, 199,
energy, 73 200, 205, 231, 234, 235, 239
energy transfer, 73 Hausdorff dimension, 3, 19, 30, 31, 32, 33, 42, 49,
engineering, 120, 205, 222, 239 50
environment, 73 Hausdorff-Besicovitch dimension, 30
equality, 17, 60, 135, 144, 145, 150, 156, 159, 172, heat transfer, 120
173 homogeneity, 59, 197
equivalence classes, 20, 24 hypothesis, 13, 19, 49, 98, 102, 116, 169, 170
Euler-Lagrange equations, 72, 82, 109, 113, 115,
122, 128, 138, 146, 153, 158, 173, 174, 198
evolution, 63, 66, 69, 70, 72, 73, 124, 140, 141, 179, I
180, 181, 182, 184, 186, 187, 190, 191, 198, 199
exponential functions, 124 identification, 238
identity, 75, 103, 154, 155
image, 76, 88, 140
F independent variable, 220
induction, 6, 13, 14, 17, 98, 102, 104, 116, 132, 149,
fluid, 127, 139 160, 168, 169, 170
force, 70, 72 inequality, 11, 12, 25, 32, 35, 42, 50, 57, 60, 164,
Ford, 238 166, 177, 195
formula, 35, 36, 42, 55, 70, 71, 75, 77, 78, 88, 90, integration, 71, 87, 90, 91, 97, 99, 100, 107, 111,
91, 97, 99, 100, 102, 107, 115, 117, 144, 145, 148, 115, 124, 131, 136, 148, 151, 176, 177, 208, 210,
158, 159, 160, 182, 188, 205, 208, 209, 211, 221, 216, 230, 233
223, 224, 238 internal time, 139, 140
foundations, vii, 199 inversion, 22
fractal dimension, viii, 2, 3, 19, 32
fractal properties, 189, 198
fractal structure, 189 J
fractional derivatives, vii, 2, 3, 35, 38, 40, 42, 43, 44,
46, 51, 61, 62, 66, 70, 73, 96, 97, 98, 101, 104, Japan, 63
107, 109, 113, 116, 124, 127, 129, 130, 131, 137, justification, 180, 189
142, 146, 147, 148, 174, 175, 177, 179, 180, 188,
189, 191, 198, 203, 204, 205, 212, 213, 215, 217,
221, 222, 223, 236, 238, 239 K
fractional differential equations, vii, 128, 138, 146,
152, 175, 177, 205, 231, 238 kinetic equations, 200
fractional differentiation, 205 kinetics, 179, 200
fractional order, 3, 139, 180, 200, 238
framework of wavelet, 2
L
France, 65, 127, 176, 179
friction, 69, 71, 72, 116
Lagrange multipliers, 205, 206, 231, 234
function values, 220
Lagrangian formalism, 65, 66
functional analysis, 238
Lagrangian formulation, 66, 174, 198
funds/funding, 237
laws, 128, 143, 155, 175, 200
lead, 76, 87, 93, 133, 135, 150, 182, 226
G Lebesgue measure, 2
Lie group, 124
geometry, viii, 2, 35 linear function, 59
graph, 2, 3, 19, 33, 34, 35, 49, 50, 63, 235 linear systems, 228
groundwater, 120 liquids, 177
growth, 19, 20, 22
Index 243
M Q
manifolds, 73 quantization, 73
manipulation, 120 quantum mechanics, 1, 73
mapping, 55, 75, 76, 87, 88, 181, 182
materials, 35
matrix, 88, 89, 103, 141, 228 R
matter, 206, 225
memory, 2, 35, 175, 179 radius, 4, 6, 207
metric spaces, 25 random walk, 63, 200
mixing, 71, 99, 120 real numbers, 17, 34, 190, 202
models, 1, 179, 180, 189 real time, 141
Moscow, 238 reasoning, 32, 80
multidimensional, 98, 113 recall, 82, 113, 120, 160, 182, 191
multiplier, 128, 132, 133, 134, 135, 137, 176, 231 recurrence, 179, 180, 181, 183, 198
relativity, 1, 55
relevance, 180
N renormalization, 180, 182, 191, 198, 200
requirements, 218
nature of turbulence, 9 resolution, 54, 128, 142
Niyazov, 200 rheology, 1
non-analytic functions, 19 rotations, 144, 145, 159
numerical analysis, 74, 87 roughness, 1
O S
obstruction, 66, 68, 69, 70, 143 shape, 188
one dimension, 113 signals, vii, 62
operations, 210 simulations, viii
optimization, 128, 133, 134, 140, 175, 201, 205, 206, Singapore, 199
220, 234, 237 smoothness, 1, 62
ordinary differential equations, 92, 105, 133, 176, solar system, 124
234, 235 solution, 66, 77, 83, 89, 117, 119, 120, 121, 122,
128, 132, 133, 134, 135, 136, 137, 138, 141, 142,
143, 144, 145, 149, 150, 151, 152, 153, 154, 156,
P 157, 159, 167, 168, 169, 170, 171, 173, 176, 177,
205, 206, 207, 212, 220, 221, 222, 226, 227, 228,
partial differential equations, vii, viii, 66, 67, 73, 74, 229, 230, 231, 233, 236, 237, 238
75, 95, 113, 115, 120, 123 sorption, 177
partition, 147, 186, 187 stability, 32, 128, 134, 136, 152
phase transitions, 1 state, 33, 72, 129, 134, 138, 143, 149, 174
physical phenomena, 1 states, 3, 23, 24, 38, 69, 72, 183
physical theories, 1 stochastic processes, 74, 80
physics, 80, 176, 179, 222 stress, 24
Planck constant, 73 structure, 2, 38, 45, 65, 66, 69, 90, 128, 137, 139,
Poisson equation, 121 146, 152, 180, 182, 188, 189, 190
pollution, 120 substitution, 53, 193
polymeric materials, 1 Switzerland, 63
porous media, 120, 140, 177 symmetry, 85, 128, 129, 142, 143, 144, 145, 146,
Portugal, 1, 201 155, 156, 157, 158, 159, 160
positive integers such that, 9
principles, 73, 122, 127, 139
probability, 4, 19, 127, 139, 183, 188 T
probability density function, 139
project, 237 techniques, 2, 176, 204, 207, 215, 219
proposition, 10, 19, 22, 25, 29, 31, 32, 49, 50 tension, 119
proteins, 73 terminals, 35
244 Index
thermodynamics, 127, 176 123, 128, 131, 133, 134, 135, 137, 148, 150, 174,
topology, 4, 28 175, 176, 199, 201, 204, 205, 207, 220, 230, 232,
trajectory, 2, 71, 83, 180, 181, 186, 187, 204 237, 238, 239
transformation, 45, 121, 206 vector, 4, 55, 56, 57, 58, 59, 60, 74, 75, 83, 85, 87,
translation, 13 93, 94, 95, 111, 205
transport, 1, 120, 177, 179, 180, 188, 189, 190, 197, velocity, 9, 120
198, 200
turbulence, 1, 9
W
V war, 127
Washington, 174
variables, 62, 66, 70, 80, 176 wavelet, 2
variations, vii, viii, 39, 65, 66, 67, 68, 71, 72, 81, 83,
85, 93, 97, 109, 111, 112, 113, 115, 116, 117, 118,

(Mathematics Research Developments) Jacky Cresson - Fractional Calculus in Analysis, Dynamics and Optimal Control-Nova Science Pub Inc (2014)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Mathematics Research Developments) Jacky Cresson - Fractional Calculus in Analysis, Dynamics and Optimal Control-Nova Science Pub Inc (2014)

Uploaded by

Copyright:

Available Formats

MATHEMATICS RESEARCH DEVELOPMENTS

FRACTIONAL CALCULUS IN ANALYSIS,

Additional books in this series can be found on Nova’s website

Additional e-books in this series can be found on Nova’s website

FRACTIONAL CALCULUS IN ANALYSIS,

NOTICE TO THE READER

LIBRARY OF CONGRESS CATALOGING-IN-PUBLICATION DATA

Published by Nova Science Publishers, Inc. † New York

• Analysis of non differentiable functions

• Variational formulations of PDEs or ODEs

L OCAL F RACTIONAL D ERIVATIVES

PACS: 05.45-a, 52.35.Mw, 96.50.Fm

||f ||0 := sup |f (x)|. (1)

2. Non-differentiable and Non-analytic Functions

Definition 2. Let X be a metric space. A subset M ⊂ X is said to be rare (or nowhere

Lemma 3. A subset M of a metric space X is meager if and only if M ⊂ ∪n∈N Mn , with

Proof. Suppose that M is meager. Then M = ∪n∈N Fn , where int(Fn ) = ∅. Let Mn := Fn

Proof. For each n ∈ N let us choose xn ∈ Fn . Since F1 ⊃ F2 ⊃ · · · , we have xn ∈ Fk ,

Theorem 5. (Baire) Let X be a complete metric space and B1 , B2 , B3 , · · · a countable

2.2. Continuous Nowhere Differentiable Functions

Proof. Let us denote by {Mn }n∈N the collection of subsets of C 0 [a, b]

If g ∈ M , then g ∈ Mn for some (sufficiently large) n ∈ N, and thus M ⊂ ∪n∈N Mn . To

for all h ∈ ( 0, δ ]. If we choose n ∈ N such that n > K and b−an ≤ δ, then g ∈ Mn .

h. On the other hand, the

One can then find a (sufficiently small) h ∈ ( 0, b−a

The continuity of T is easily established.

Lemma 9. Takagi’s function is continuous in [0, 1].

Proof. Notice that 0 ≤ φ(x) ≤ 12 for any x ∈ R. Consequently:

|T (x + h) − T (x)| ≤ C|h| log(1/|h|). (11)

Theorem 12. (Krüppel) Let x ∈ [0, 1] be a dyadic rational. Then:

Definition 13. The binary expansion of a point x ∈ [ 0, 1 ) is denoted by

where, obviously ǫn = ǫn (x). If the limit

d0 (x) := 1 − d1 (x) (15)

is the density of the digit ”0”.

(a) 0 < d1 (x) < 1; or

Lemma 17. The Weierstrass function is continuous in [0, 1].

Proof. Again the continuity is an immediate consequence of the uniform convergence of

|Wλ,s (x) − Wλ,s (y)| ≥ C2 δ 2−s , (19)

for large enough λ.

Proof. The proof presented here is given in Falconer’s book [12].

(i) Given 0 < |h| < λ−1 , let N ∈ N such that

λ−(N +1) ≤ |h| < λ−N . (20)

P∞ (s−2)k sin λk (x + h) − sin(λk x) ≤

using that | sin(u) − sin(v)| ≤ |u − v| (a consequence of the mean-value theorem) on the

|h|λ(s−1)N 2λ(s−2)(N +1)

where c is independent of h and we used (20).

where we used (20) again.

Wλ,s (x + h) − Wλ,s (x) − λ(s−2)N sin λN (x + h) − sin(λk x) ≤ λ(s−2)N

λ(s−2)N λ(s−2)N λ(s−2)N λs−2 δ 2−s

2.3. Infinitely Differentiable, Nowhere Analytic Functions

with the convention f (0) = f .

Proposition 20. The set F = (C ∞ [a, b] , dF ) is a complete, separable metric space.

Proof. The proof proceeds in several steps.

|u + v| |u| + |v| |u| |v| |u| |v|

and dF (f1 , f2 ) ≤ dF (f1 , f3 ) + dF (f3 , f2 ).

lim sup |fn(k) (x) − gk (x)| = 0, (33)

for 0 ≤ x ≤ 1, which completes the induction proof on n.

P∞ 1 |f (k) (x)−p(k) (x)|

1) The set Uj is open. Indeed, let f ∈ Uj and define:

for all k ≥ 0. In particular, for 0 ≤ k ≤ j, we have:

From (48), we also have:

for all a ≤ x ≤ b. Consequently, from (49) and (50), we have:

dF (f, h) = dF (cg, p) < ǫ1 . (53)

Next, we notice that for large enough j ≥ J, p(j) = 0 and

In other words, either f ∈ ∪∞ j=J Uj or h ∈ ∪∞

(monotonicity) If E ⊂ F ⊂ B, then ψH (E) ψH (F ) and dimH (E) ≤ dimH (F ).

sup1≤i<∞ {ψH (Fi )} ψH (∪∞