Insights Into The Analysis of FDDEs

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 150

Insights into the Analysis of Fractional

Delay Differential Equations

Item Type Thesis or dissertation

Authors Osborne, Neil

Citation Osborne, N. (2021). Insights into the analysis of fractional delay


differential equations [Unpublished doctoral thesis]. University of
Chester.

Publisher University of Chester

Rights Attribution-NonCommercial-NoDerivatives 4.0 International

Usage policy The full-text may be used and/or reproduced in any format
or medium, without prior permission or charge, for personal
research or study, educational, or not-for-profit purposes
provided that: - A full bibliographic reference is made to the
original source - A link is made to the metadata record in
ChesterRep - The full-text is not changed in any way - The full-
text must not be sold in any format or medium without the formal
permission of the copyright holders. - For more information
please email researchsupport.lis@chester.ac.uk

Download date 01/03/2024 03:51:43

Item License http://creativecommons.org/licenses/by-nc-nd/4.0/

Link to Item http://hdl.handle.net/10034/626737


Insights Into The Analysis Of
Fractional Delay Differential
Equations

Thesis submitted in accordance with the University of Chester for the degree of
Doctor of Philosophy by Neil Osborne
July 2021
Acknowledgements

I would like to thank the following people for their help and support during this
research:

ˆ Dr F.G.Rodrigues (Departamento de Matematicas, Universidad de La Ser-

ena, Chile): for his tuition and insights in to the subject of Fractional
Calculus and other matters not just of mathematical interest.

ˆ Professor N.J.Ford and Dr J.Roberts (University of Chester) for their knowl-

edge, guidance and patience.

1
Declaration

The material being presented for examination is my own work and has not been
submitted for an award of this or another HEI except in minor particulars which
are explicitly noted in the body of the thesis. Where research pertaining to the
thesis was undertaken collaboratively, the nature and extent of my individual
contribution has been made explicit.

Signed Date

2
Contents

Glossary 6

Acronyms 7

1 Abstract 8

2 Introduction 9

3 Ordinary and Delay Differential Equations 14


3.1 Ordinary Differential Equations . . . . . . . . . . . . . . . . . . . 14
3.2 Delay Differential Equations . . . . . . . . . . . . . . . . . . . . . 24

4 Introduction to Fractional Calculus 32


4.1 Riemann-Liouville Fractional Integral Operator . . . . . . . . . . 34
4.2 Caputo Fractional Differential Operator . . . . . . . . . . . . . . . 40
4.3 Fractional Operator Properties . . . . . . . . . . . . . . . . . . . . 46
4.4 Fractional Ordinary Differential Equations . . . . . . . . . . . . . 50
4.5 Fractional Delay Differential Equations . . . . . . . . . . . . . . . 55

5 Existence and Uniqueness of Solutions 58


5.1 Ordinary Differential Equation IVP . . . . . . . . . . . . . . . . . 58
5.2 Delay Differential Equation IIP . . . . . . . . . . . . . . . . . . . 59
5.3 Fractional Ordinary Differential Equation IVP . . . . . . . . . . . 61
5.4 Fractional Delay Differential Equation IIP . . . . . . . . . . . . . 61

6 Solution Of Differential Equations 65


6.1 Solution of a Ordinary Differential Equation IVP . . . . . . . . . 65

3
6.2 Solution of a Delay Differential Equation IIP . . . . . . . . . . . . 68
6.3 Solution of a Fractional Ordinary Differential Equation IVP . . . 73
6.4 Solution of a Fractional Delay Differential Equation IIP . . . . . . 76
6.4.1 Validation of the New Solution . . . . . . . . . . . . . . . 88

7 Analysis of Fractional Delay Differential Equation IIP Solution 98


7.1 Solution Comparison: FDDE IIP versus DDE IIP . . . . . . . . . 98
7.2 Solution Continuity and Smoothness for FDDE IIP . . . . . . . . 99
7.3 Solution Properties . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.4 Issues for Numerical Methods (of Solution) . . . . . . . . . . . . . 119
7.5 Examples of FDDE IIP’s . . . . . . . . . . . . . . . . . . . . . . . 126
7.6 Extension of Solutions to Two Term FDDE IVP . . . . . . . . . . 133

8 Further Work 138

A Appendix 140

Appendix 140
A.1 Definition ([9] Definition 2.1) . . . . . . . . . . . . . . . . . . . . . 140
A.2 Definition (based upon [19] Eqn 2.4.51) . . . . . . . . . . . . . . . 141
A.3 Theorem (based upon [9] Theorem 3.8) . . . . . . . . . . . . . . . 141
A.4 Theorem (based upon [9] Theorem 3.7) . . . . . . . . . . . . . . . 142
A.5 Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
A.6 Theorem (Reference[9] Theorem 2.10) . . . . . . . . . . . . . . . . 144
A.7 Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

References 146

4
Glossary

A1 (0, T ] is the set of functions with an absolutely continuous first derivative on


the domain (0, T ] 9

α is the order of the fractional differential equation, 0 < α < 1 9

β is the power index of the term (s − a)β the fractional operators act upon 37

C0 (0, T ] is the set of continuous functions on the domain (0, T ] 17

C1 (0, T ] is the set of functions with a continuous first derivative on the domain
(0, T ] 49

c α
a Dt f (s) is the Caputo fractional derivative (see Appendix A.2) 141

f (t, y(t)) is the right side of the general ODE IVP 16

γ is the power index of the delay differential equation history function, tγ , where
γ ∈ R 25

α
a Jt f (s) is the Riemann-Liouville fractional integral (see Appendix A.1) 140

ϕ is the history function for the delay differential equations 9

τ is the delay period of the delay differential equations, where τ > 0 9

y0 is an initial value where y0 ∈ R 11

y1 , y2 are values of y where y1 , y2 ∈ R 17

5
yun is the unknown (un) function 68

yf ix is the known (fix=fixed) function 80

Acronyms

DDE IIP (one term) delay differential equation initial interval problem 11

DDE IIP (two term) delay differential equation with added term y(t) initial
interval problem 138

FDDE IIP (one term) fractional delay differential equation initial interval prob-
lem 9

FDDE IIP (two term) fractional delay differential equation with added term
y(t) initial interval problem 139

FODE IVP fractional ordinary differential equation initial value problem 11

ODE IVP ordinary differential equation initial value problem 11

6
1 Abstract

This thesis is concerned with determining the analytic solution, using the method
of steps, of the following fractional delay differential equation initial interval prob-
lem (FDDE IIP),

c α
0 Dt y(s) = −y(t − τ ) for t > 0, τ > 0, 0 < α < 1, and y ∈ A1 (0, T ]

y(t) = ϕ(t) for t ∈ (−τ, 0]

The properties of the analytic solution obtained are a surprise but they do sit
comfortably when compared with those of the analytic solutions of an ordinary
differential equation initial value problem (ODE IVP), a delay differential equa-
tion initial interval problem (DDE IIP) and an fractional ordinary differential
equation initial value problem (FODE IVP). Further the analytic solution for-
mula obtained is closely related to that of the analytic solution formula of the
DDE IIP. However, these insights into the analytic solution of the FDDE IIP we
have not seen before, and differ from those published elsewhere.

Keywords: fractional delay differential equation, initial interval problem, ex-


istence and uniqueness, analytic solution, continuity, smoothness

7
2 Introduction

It is well known that even a simple DDE IIP has discontinuities in the derivatives
of its solution [28]. The situation is complicated in more complex DDE’s which
we present in chapter 7.4. These cause a problem in some numerical methods of
solution which use these values from the derivatives. Some numerical methods
will trace the occurrence of the discontinuities and take some sort of average when
needed. So the question arises when we use fractional calculus of whether the
same problems occur in FDDE IIP’s. A literature search in 2015 at the time of
starting, revealed little in the way of analysis although there are many papers
solving such FDDE IIP numerically. So the primary motivation for this thesis
is to solve analytically, if possible, the linear FDDE IIP (one term) (fractional
delay differential equation initial interval problem) given below,

c α
0 Dt y(s) = −y(t − τ ) for t > 0, τ > 0, 0 < α < 1, and y ∈ A1 (0, T ]

y(t) = ϕ(t) for t ∈ (−τ, 0]

and to determine properties of the solution. The results we obtained were a


surprise to us. Subsequently, a later literature search in 2020, revealed work on
the same area of interest that differed in some ways from our own (for exam-
ple ’On Initial Conditions for Fractional Delay Differential Equations’, E.Kaslik,
R.Garrappa[15]).
There are a number of symbols/variables in this work and to help with clarity
note the following.

ˆ Symbols/variables relevant to the whole work will be defined when first

8
used, but may not be subsequently, but you will find them in the glossary.

ˆ Symbols/variables relevant to a particular topic/piece of analysis will be

defined locally or accepted as clearly defined in the analysis. Such variables


will not be in the glossary and may be redefined in a different context in
the subsequent work.

Furthermore, within this thesis we set τ = 1 (the delay term) in order to present
any plots.
A number of definitions and theorems have been collected together and in-
cluded in the appendix. The four appendix subsections A1 to A4 you may need
to reference at various times during this thesis and the appendix makes this a
simple task. The other subsections would just clutter the thesis more than it
needs to be, and so we have sited them in the appendix. Further there are a
number of sections/results in this work that are crucial to this thesis in that they
are either new material or important insights. We signpost these sections/results
below to help when reading/reviewing this thesis.

ˆ Consideration of the integral and integrand (see pages 45 and 47).

ˆ Solution of a FDDE IIP (see pages 78-90).

ˆ Validation of the New Solution (see pages 90-100).

ˆ Solution Comparison with DDE IIP (see pages 101-103).

ˆ Solution Continuity and Smoothness (see pages 103-117).

ˆ Issues for Numerical Methods (see pages 131-136).

9
It was during this work that we realised, to have a broader understanding
of the FDDE IIP we should not be viewing this in isolation. Starting with the
ordinary differential equation initial value problem (ODE IVP-see below) the
step to a delay differential equation initial interval problem (DDE IIP-see below)
provides a complication that is different to that of the step from the ODE IVP to
a fractional ordinary differential equation initial value problem (FODE IVP-see
below). The FDDE IIP is the culmination of starting with the ODE IVP and
travelling clockwise or anticlockwise around figure 1.

ˆ ODE IVP (Ordinary differential equation initial value problem):

dy
= −y(t) for t > 0 and y ∈ A1 (0, T ]
dt
y(0) = y0

ˆ DDE IIP (one term) (Delay differential equation initial interval problem):

dy
= −y(t − τ ) for t > 0 and y ∈ A1 (0, T ]
dt
y(t) = ϕ(t) for t ∈ (−τ, 0]

ˆ FODE IVP (Fractional ordinary differential equation initial value problem):

c α
0 Dt y(s) = −y(t) for t > 0, 0 < α < 1, and y ∈ A1 (0, T ]

y(0) = y0

10
ˆ Fractional delay differential equation initial interval problem (FDDE IIP):

c α
0 Dt y(s) = −y(t − τ ) for t > 0, 0 < α < 1, and y ∈ A1 (0, T ]

y(t) = ϕ(t) for t ∈ (−τ, 0]

And the relationships of these four differential equations can be viewed as,

Figure 1: Relationships between Differential Equations

So this thesis is broadly organised as follows,

ˆ Ordinary and delay differential equations: A short brief on the origin and

development of ODE IVP’s and DDE IIP’s where our first two differential
equations are defined.

ˆ Introduction to fractional calculus: The Riemann-Liouville fractional inte-

gral operator and the Caputo fractional differential operator.

11
ˆ Fractional ordinary and fractional delay differential equations: A short brief

on FODE IVP’s and FDDE IIP’s where our second two differential equa-
tions are defined.

ˆ Existence and uniqueness of solutions: All four differential equations will

be analysed for existence of a single unique solution.

ˆ Solution of differential equations: All four differential equations will be

solved. The solution for the FDDE IIP will be analysed, the aim being to
determine whether it is ’correct’, and we will conduct a continuity analysis.
Four examples of FDDE IIP’s with different history functions and different
differential equation orders α, will be presented.

ˆ Further work: Proposals for further work will be presented.

We begin with a short brief on ODE IVP’s and DDE IIP’s.

12
3 Ordinary and Delay Differential Equations

3.1 Ordinary Differential Equations

When undertaking a mathematical study of a dynamic process a key type of


equation that arises is that of the differential equation [22]. They express the
rates of change of a variable, called a dependant variable, say y(t), with respect to
another variable, called an independent variable, say t. For example, in dynamics,
dy
if we were modelling the position of an object we could have dt
, and/or higher
derivatives. Differential equations occur from a huge number of situations and in
a huge number of different varieties. An interesting introduction to differential
equations is the modelling of the size of a single species population [30]. It is worth
retelling as it shows how simply the differential equation emerges and combined
with an initial condition (y(t = 0) = y0 ) we have what is called an IVP. We are
then in a position to predict future behaviour of the population for t > 0, by
solving the ODE IVP. Inevitably we will have made some assumptions during
the modelling process and we will want to determine, by analysing the solution,
whether we should make any modifications to the ODE IVP. Hence we can go
through a process of changes to the ODE IVP to improve the accuracy of the
solution.
So to begin with we define P (t) as the population at a time t, then we can
argue that the increase in the population will be proportional to the population,
so the first ODE can be written as,

dP
= kP (t) where k > 0
dt

13
However, populations can reduce as well as increase, so we can define,

k = b − c where b = birth rate and c = death rate

so k can be both negative and positive, and the ODE becomes,

dP
= (b − c)P (t)
dt

Before we continue, we should note that population data is measured in integer


numbers, but we are considering P (t) to be a continuous function of time. We are
assuming that the population data is of the order of thousands, and so a small
change of a single individual is very small in respect to the overall numbers. If
we wanted to use discrete data we could use difference equations.
For the differential equation there are a number of descriptive terms used.

ˆ ’linear’: there is a formal definition of linearity, but for our purposes y(t)
dy
and dt
are raised to the power one,

ˆ ’ordinary’: there is only one independent variable t,

ˆ ’order’: the highest derivative is dP


dt
so is order one, and

ˆ ’constant coefficient’: the b − c is a constant, and not a function of t.

So we have an ordinary, linear, first order, constant coefficient differential equa-


tion. It is a fact that most differential equations cannot be solved in terms of
simple functions, that is to get an analytical solution. Before we attempt this we
determine indirectly (without solving the differential equation) whether,

ˆ a solution exists (existence), and,

14
ˆ can there be more than one solution (uniqueness)?

The existence and uniqueness theorem we will use is based upon the Picard Lin-
delöf result. This theorem considers properties of the ODE and the initial condi-
tion to prove existence and uniqueness. When looking at a differential equation
and initial condition it is not always easy to spot whether a solution is unique.
Consider the following two differential equations and initial conditions,

dy 2
= y3 where y(0) = 0
dt
dy 4
= y3 where y(0) = 0
dt

Both differential equations have the trivial solution y(t) = 0 which can be seen
by inspection, so solutions exist. The first however can be solved by separation
t3
of variables to have a further solution y(t) = 27
so the solutions are not unique.
The second when we attempt to solve by separation of variables fails to provide
a further solution and so the trivial solution is the only solution and is hence
unique. We now state the Picard Lindelöf result, a corollary of which is the
existence and uniqueness theorem we will use in this thesis.

Theorem 1 (based upon [9] Theorem 6.8)

c α
0 Dt y(s) = f (t, y(t))

y(0) = y0

Let 0 < α ≤ 1, y0 ∈ R and h∗ > 0. Define G = [0, h∗ ] × R and let the function
f : G → R be continuous and fulfil a Lipschitz condition with respect to the second

15
variable with a Lipschitz constant L > 0 that is independent of t, y1 , y2 . That is,

|f (t, y1 ) − f (t, y2 )|≤ L|y1 − y2 |

then there exists a uniquely defined function y ∈ C0 (0, T ] solving the initial value
problem.

Since there is no restriction on the distance h∗ besides f being continuous on


[0, h∗], we can confirm the following corollary, where we extend the distance h∗ ,

Corollary 1.1 (based upon [9] Corollary 6.9)


Let 0 < α ≤ 1, y0 ∈ R and h∗ > 0. Define G = [0, ∞) × R and let the function
f : G → R be continuous and fulfil a Lipschitz condition with respect to the
second variable with a Lipschitz constant L > 0 that is independent of t, y1 and
y2 . Then there exists a uniquely defined function y ∈ C0 [0, ∞) solving the initial
value problem.

For the population ODE, we need an initial condition to give us a starting


value, so we set P (0) = P0 , that is,

dP
= (b − c)P (t) where P (0) = P0 (1)
dt

We will now use the Corollary of the Picard Lindelöf result to show a solution
exists and there is at most one solution so it is unique. We consider the conditions
for the Corollary to be satisfied,

ˆ differential equation order α = 1 ⇒ 0 < α ≤ 1,

ˆ initial condition P (0) = P0 ∈ R,

16
ˆ f (t, y) = (b − c)P (t) is continuous on R

and fulfilling the Lipschitz condition,

|f (t, y1 ) − f (t, y2 )| = |(b − c)P1 (t) − (b − c)P2 (t)|≤ L|P1 − P2 |

for L = |b − c|> 0.

so there exists a uniquely defined function P ∈ C0 (0, ∞] solving the initial value
problem.
We now move on to solve the ODE IVP, and to be consistent with later work
where we solve a FODE IVP, we will use Picard iteration. The first step is to
integrate the ODE.

dP
= (b − c)P (t) (2)
dt

Z t Z t
dP
ds = (b − c) P (s)ds
0 ds 0
Z t
P (t) − P (0) = (b − c) P (s)ds
0
Z t
P (t) = P0 + (b − c) P (s)ds
0

We now have a Volterra integral equation of the second kind [5]. Picard iteration is
a process of successive iterations that converge to the solution, proved by Banach’s
fixed point theory. Now since we have shown that there is only one solution, the
iterations must converge to it. Consider the general Volterra integral equation of

17
the second kind,

Z t
f (t) = g(t) + k(t, s)f (s)ds
0

where k(t, s) is continuous on 0 ≤ s ≤ t ≤ T and g(t) is continuous on 0 ≤ t ≤ T .


The form of the iterations are,

Z t
fn (t) = g(t) + k(t, s)fn−1 (s)ds, for n=1,2, . . .
0

where f0 (t) = g(t)

For the first iteration we apply P (t) = P0 to the right side of the integral equation
to give,

Z t
P1 (t) = P0 + (b − c) P0 ds
0
Z t
= P0 + (b − c)P0 ds
0

= P0 + P0 (b − c)t

and for the second iteration we apply P1 (t) to the right side of the integral equa-
tion to give,

Z t
P2 (t) = P0 + (b − c) P0 + P0 (b − c)sds
0
Z t Z t
2
= P0 + P0 (b − c) ds + P0 (b − c) sds
0 0
P0 (b − c)2 t2
= P0 + P0 (b − c)t +
2

18
which on repeating leads to,

P0 (b − c)2 t2 P0 (b − c)3 t3
P (t) =P0 + P0 (b − c)t + + ...
2! 3!

X (b − c)n tn
= P0
n=0
n!

= P0 e(b−c)t (Malthus 1798)

where the value b − c is known as an eigenvalue.


In order to be able to investigate the solution we define P0 = 2000 and plot
some sample solutions. In the figure below the red line is the solution with b < c
and the green line is solution with b > c. From our knowledge of the exponential
function, we have endless population reduction towards P (t) = 0 (red line) and
endless population growth towards P (t) = ∞ (green line). There is also a solution
for b = c, a constant solution P (t) = 2000 shown as a blue dashed line. This
solution is described as unstable since any small perturbation from this solution,
either above or below P (t) = 2000, and the population moves away towards
infinity, P (t) → ∞, or towards zero, P (t) → 0. It is also worth noting that these
solutions extend to negative time, although for this population model this makes
little practical sense.

19
Figure 2: Solutions of the Differential Equation

Endless population growth does not seem practical since it is likely that factors
such as restricted resources, for example food, space, would eventually reduce the
growth rate. On the other hand the endless population reduction to zero could be
realistic as species extinction is a reality. Maybe if the time scale over which this
model was used was restricted then it could be of some use. This model certainly
has some problems but maybe a change can be made to improve the long term
behaviour.
Before we proceed we write down the ODE IVP in general terms. We will use
this later in our work as part of our analysis.

dy
= −y(t) where y(t) ∈ A1 [0, T ] (3)
dt
y(0) = y0

It is true that the process of modelling should start from a simple model with

20
a differential equation, and then improving with more sophistication to make
progress. Including too much sophistication in to a model from the outset can
cloud our understanding. The long term behaviour was modified by Verhurst
(1836), who changed the constant coefficient, with a function of t,

 
P (t)
b−c=r 1−
k

where,

ˆ r = intrinsic growth rate

ˆ k = carrying capacity

giving what is called the Logistic equation.

 
dP P (t)
= rP (t) 1 − where t > 0, P (t) ∈ A1 [0, T ]
dt k
P (0) = P0

This differential equation can be solved by separation of variables, to give,

P0 k
P (t) =
P0 − (P0 − k)e−rt

In the figure below we have the solutions,

ˆ in red P0 = 500, k = 2000 and r = 2,

ˆ in green P0 = 3000, k = 2000 and r = 1,

ˆ in blue dashed P0 = 2000, k = 2000 and r is not defined,

21
ˆ the ’t’ axis is a solution P0 = 0.

As with the previous differential equation we can describe the constant solutions
as either stable or unstable. First consider the constant solution P (t) = 0, any
small perturbation results in the solution moving away in a manner like the red
solution. Hence the constant solution P (t) = 0 is described as unstable. Whereas
the constant solution P (t) = 2000, in the blue dashed line, any small perturbation
results in the solution returning back towards the blue dashed line, either in the
manner like the green solution or the red solution. Hence this constant solution
is described as stable.

Figure 3: Solutions of the Logistic Differential Equation

The predicted behaviour of the single species population now has changed.
The red line could predict the population growth of a species with a low popu-
lation and an abundance of resources. The early growth rate is reduced as the
population increases towards the k = 2000 value, called the carrying capacity.
The green line could predict the population growth of a species with a high pop-

22
ulation and a lack of resources. The population falls to the carrying capacity. So
we can confirm this is an improved predictor of a single species population for
both the short and long term. A criticism of this model could be that the solu-
tions are very orderly. The red solution, and all solutions with 0 < P0 < 2000,
rise and asymptotically approach the carrying capacity k but never cross k. Sim-
ilarly, the green solution, and all solutions with P0 > k, fall and asymptotically
approach k, but never cross k. In studies of single species populations, and our
experience in other situations, we would expect that the population could os-
cillate about the carrying capacity k. However oscillatory behaviour is never a
characteristic of these differential equations. We next move on to DDE IIP prob-
lems where we will show that both exponential solutions and oscillating solutions
are experienced.

3.2 Delay Differential Equations

The next change made to this differential equation (Hutchinson 1948 + others)
involved including a delay term P (t − τ ). For this model this represents the delay
that occurs between when an individual is born and when they mature enough
to produce offspring [8]. The differential equation had the form,

 
dP P (t − τ )
= rP (t) 1 −
dt k

So the term P (t − τ ) has a delay in its argument by an amount τ . This delay


differential equation cannot be solved into simple functions and so other methods
have to be employed. But it is found that oscillating solutions are involved.
However we want to try and understand the role a delay term has on a solution.

23
To make further progress we will simplify the differential equation to a linear delay
differential equation, given by,

dy
= −y(t − τ ) for t > 0, τ > 0 and y(t) ∈ A1 (0, T ]
dt
y(t) = ϕ(t) for t ∈ (−τ, 0] (4)

This is the second differential equation that we will use later in this work in
our analysis. In order for a solution to this delay differential equation to be
determined we need more than just an initial condition [14]. We actually need
a function ϕ(t) continuous on −τ < t ≤ 0, called the history function. If we
consider a history function in the form of a power function, that is ktγ (where k
is a constant), then we are restricted in the values of γ that we can use. If,

ˆ γ ∈ (−∞, 0) then the power function is not continuous at t = 0, so is not

acceptable,

ˆ γ ∈ [0, ∞) except γ = 0, 1, 2, 3, ... the power function is not real valued on

(−τ, 0) so is not acceptable,

ˆ γ = 0, 1, 2, 3, ... are continuous on (−τ, 0) so is acceptable.

So the acceptable history functions are of the form ϕ(t) = k, kt, kt2 , . . .. In our
later work we will use the first two acceptable power functions as history func-
tions in our examples, that is, ϕ(t) = 1 and ϕ(t) = t.
The distinction between an ODE and a DDE is described in what is called
their dimensionality. The ODE requires just one piece of information, the initial
condition y(0) = y0 , to determine the subsequent behaviour of the solution. It is
hence described as being one dimensional. The DDE requires a history function

24
continuous on −τ < t ≤ 0 which has an infinite number of pieces of information
and is hence described as being of infinite dimension.

The ODE IVP (2) we solved previously had an exponential solution y(t) =
P0 e(b−c)t . We now show that the DDE IIP can have an exponential solution.
Consider,

dy
= −ηy(t − τ ) for t > 0, and η > 0 (5)
dt
y(t) = Aeλt for − τ < t ≤ 0 and A, λ are constants

We substitute the exponential solution, y(t) = eλt , into the DDE,

dy
= −ηy(t − τ ) ⇒Aλeλt = −ηAeλ(t−τ )
dt
⇒ Aeλt (λ + ηe−λτ ) = 0

⇒ λ + ηe−λτ = 0

⇒ λ = −ηe−λτ

⇒ λeλτ = −η

Note that λ is known as an eigenvalue. The above expression provides a condi-


tion which has to be satisfied in order for the DDE IIP to have an exponential
solution. If we define f1 (λ) = λeλτ and f2 (λ) = −η, we can plot the functions for
τ = 1 and various values of η,

25
Figure 4: Exponential Solution Outcomes

where

ˆ Red line: f2 (λ) = 1

ˆ Green line: f2 (λ) = −0.25

ˆ Blue line: f2 (λ) = −0.75

ˆ Black line: f1 (λ) = λeλτ

So we have three scenarios, the red line intersects the black line once giving a
single value for λ, the green line intersects the black line twice giving two values
of λ and the blue line does not intersect the black line, where complex eigenvalues
exist. Complex eigenvalues lead to oscillatory solutions via Euler’s formula, which
relates the complex exponential function to the trigonometric functions.

ex+yi = ex (cos(y) + isin(y))

In comparson the ODE IVP has one exponential solution only, and hence one
real value of λ.

26
We now determine analytically the minimum value of f1 (λ).

df1
= λτ eλτ + eλτ = 0 ⇒ eλτ (λτ + 1) = 0

⇒ λτ = −1
1
⇒ λ1 = −
τ

Hence the minimum point is,

 
1 1
(λ1 , f1 (λ1 )) = − , − e−1
τ τ

This means we can summarise the critical values of η,

ˆ η < 0 ⇒ λ1 > 0

ˆ η = 0 ⇒ λ1 = 0

ˆ 0 < η < τ1 e−1 ⇒ λ1 , λ2 < 0

ˆ η = τ1 e−1 ⇒ λ1 , λ2 = − τ1

ˆ η > τ1 e−1 ⇒ λ ∈
/R

Note that a Hopf bifurcation exists as η goes through the critical point η = τ1 e−1 ,
for more information on Hopf bifurcations see [13].
This analysis can also be carried out using the Lambert W function, see [6], [31].

We will now demonstrate how to show existence and uniqueness of solutions


whilst solving another DDE (given below) using the method of steps. This shows

27
that, at least sometimes, there are oscillating solutions of DDE’s. Consider,

 
dy π
= −y t − for t > 0 (6)
dt 2
 
π
y(t) = sin(t) for t ∈ − , 0
2

Using the method of steps we start with the first interval.

π
0<t≤ 2

Because we have restricted the domain in this step, we know that the right side
of the differential equation is,

   
π π
−y t − = −sin t −
2 2

This means we now have an ODE IVP defined on this interval, given by,

 
dy π
= −sin t − where y(0) = sin(0) = 0
dt 2

Our first task is to prove that a solution exists and that it is unique. We show
that the conditions for the Corollary of the Picard Lindelöf result are satisfied
(see corollary 1.1),

ˆ differential equation order α = 1 ⇒ 0 < α ≤ 1,

ˆ initial condition y(0) = 0 ∈ R,

ˆ f (t, y) = −sin(t − π2 ) is continuous on R, and,

ˆ fulfilling the Lipschitz condition.

28
Hence there exists a uniquely defined function y ∈ C0 (0, ∞] solving the initial
value problem. Next we solve the ODE IVP, integrating both sides of the dif-
ferential equation and attending to the initial conditions we obtain the solution:
”y(t) = sin(t). So our solution is given by,

π

sin(t), for −
 <t≤0
2
y(t) =
π

sin(t), for 0 < t ≤

2

where we have included the history function which in this case is also a solution
of the ODE. Note that this is not always the case.
In order for us to proceed in the next interval, we will need the following terms,

   
π π
y = sin ,
2 2
   
π π
y t− = sin t −
2 2

We now continue to the second step.

π
2
< t ≤ π:
So the ODE IVP for the second interval is,

     
dy π π π
= −sin t − where y = sin
dt 2 2 2

We show that the conditions for the Corollary of the Picard Lindelöf result are
satisfied (see corollary 1.1),

ˆ differential equation order α = 1 ⇒ 0 < α ≤ 1,

ˆ initial condition y( π2 ) = sin( π2 ) ∈ R,

29
ˆ f (t, y) = −sin(t − π2 ) being continuous on R, and,

ˆ fulfilling the Lipschitz condition.

Hence there exists a uniquely defined function y ∈ C0 [ π2 , ∞) solving the initial


π
value problem on the domain 2
< t ≤ π. Next we solve the ODE IVP, integrating
both sides of the differential equation and attending to the initial conditions we
obtain the solution: ”y(t) = sin(t). So our solution is given by,

π




 sin(t), for − 2
<t≤0


y(t) = sin(t), for 0 < t ≤ π2




sin(t), for π < t ≤ π


2

alternatively written as,

π
y(t) = sin(t) for − <t≤π
2

Repeating it can be shown that the sine wave is the solution of the DDE IIP
proving that oscillations can occur.
In this chapter we have discussed how differential equations occur and how
we can make modifications by changing coefficients and delay terms with an aim
of improving the behaviour of the solution. We have considered existence and
uniqueness of solutions using the Picard Lindelöf result and have solved DDE’s
using the method of steps. We now turn our attention to another development
of differential equations, that of fractional differential equations. We begin with
a short brief on fractional calculus.

30
4 Introduction to Fractional Calculus

Interest in integral equations stems back to the 1820’s with Abel, and follows with
many eminent mathematicians including Cauchy, Fredholm, Hilbert and Volterra
[17] [20] [1]. Integral equations can be used to model many situations, indeed Abel
was using them in his work on tautochrone curves. In 1823 Abel generalised the
tautochrone problem and in his solution he went on to understand, what we call
today, a Riemann-Liouville fractional order integral and a Caputo fractional order
differential [25] [18]. For a more modern use of fractional calculus in rheology
there an interesting read on Gordon Scott Blair [26].
Fractional calculus has a number of approaches. The one we are interested
in is the Riemann-Liouville fractional integral operator. This is based upon the
notion of extending the ’one-fold’ integral operator,

Z t
1
a Jt f (s) = f (s)ds
a

to the ’n-fold’ integral operator,[24],

Z t
n 1
a Jt f (s) = (t − s)n−1 f (s)ds
(n − 1)! a

This ’n-fold’ integral operator is limited by the discrete nature of the factorial
function. We can extend this operator by use of the Gamma function to replace
the factorial function. This creates the fractional integral operator of order α,
where α ∈ R+ , and we get the Riemann-Liouville fractional integral operator (see

31
Appendix A.1),

Z t
α 1
a Jt f (s) = (t − s)α−1 f (s)ds
Γ(α) a

Note that the misnomer ’fractional integral operator’ is retained even though
α may be a non-rational number, indeed in other areas of study α ∈ C. We
shall highlight two notions of a fractional differential operator, the first called
the Riemann-Liouville derivative and the second the Caputo derivative. The
definition of the Riemann-Liouville fractional derivative is given by,

RL α
a Dt f (s) = D[a Jt1−α f (s)]
 Z t 
d 1 −α
= (t − s) f (s)ds
dt Γ(1 − α) a
for a ≤ t ≤ T, where f (s) ∈ A1 [a, T ]

When used in a FDE IVP, that is,

RL α
a Dt y(s) = f (t)
 Z t 
d 1 −α
(t − s) y(s)ds = f (t)
dt Γ(1 − α) a

we solve by applying the R-L fractional integral operator (see [9] for details), to
give,

(t − a)α−1 1−α
y(t) = Jt y(s)]|t=a +a Jtα f (s)
Γ(α) a
Z t t
(t − a)α−1
Z
1 1
= (t − s)−α y(s)ds + (t − s)α−1 f (s)ds
Γ(α) Γ(1 − α) a t=a Γ(α) a

32
where the initial condition is given by,

Z t
1
1−α
a Jt y(s)|t=a = (t − s)−α y(s)ds
Γ(1 − α) a t=a

We are required to provide the value of this initial condition which is the fractional
integral of the function y(t) at t = a. However from a physical point of view it is
not clear what this means [12]. It turns out that the Riemann-Liouville fractional
derivative is easier to do analysis with, but it is difficult to use, in particular we
have to determine the initial condition. Whereas, the Caputo fractional derivative
is harder to do analysis with, but easier to use as the initial conditions are more
straight forward. It is for this reason that the Caputo fractional differential
operator is used where the initial conditions are of the form y(0) = y0 .
We will now present in more detail the Riemann-Liouville fractional integral
operator and the Caputo fractional differential operator.

4.1 Riemann-Liouville Fractional Integral Operator

At this stage it would be useful to explore a few of the characteristics of the


Riemann-Liouville fractional integral operator. The definition is given in ap-
pendix A.1 and is valid for α > 0, although we restrict our work to 0 < α < 1.

Z t
α 1
a Jt f (s) = (t − s)α−1 f (s)ds exists almost everywhere
Γ(α) a

for a ≤ t ≤ T, where f (s) ∈ A1 [a, T ]

33
For α = 0 the integral is undefined, but we note that,

Z t
α 1
lim a Jt f (s) = lim (t − s)α−1 f (s)ds
α→0 α→0 Γ(α) a

which we integrate by parts, to give,

t Z t
(t − s)α (t − s)α df
 
α 1
lim a Jt f (s) = lim − f (s) + ds
α→0 α→0 Γ(α) α a α ds
s=a
  Z t 
1 α α df
= lim − 0 + f (a)(t − a) + (t − s) ds
α→0 Γ(α + 1) a ds
Z t
df
= f (a) + ds
a ds

= f (a) + f (t) − f (a)

= f (t)

so we set a Jt0 f (s) = f (t), known as the identity, as we would expect in the
classical definition.
For α = 1, the integral simplifies to,

Z t
1
a Jt f (s) = f (s)ds
a

that is the classic integral.


Consider now an example where,

ˆ the fractional integral order α = 1


3

1
ˆ the function to be integrated is (s − 1) 2

ˆ the domain of integration is from 1 to t (where t > 1)

34
which is written as,

Z t
1 1 1 2 1
1 Jt (s − 1) =
3 2 (t − s)− 3 (s − 1) 2 ds
Γ( 13 ) 1

In order to be able to draw diagrams, in this section and the next section on the
Caputo fractional derivative, we have chosen t = 2.3, but note that in reality t
can take the values 1 < t < ∞.

Figure 5: Functions in the integrand

In figure 5 you can see the terms from the integrand, in black is the function
2
1 (t−s)− 3
to be fractionally integrated (s − 1) 2 and in green is the term Γ( 13 )
. This green
line is a weighting factor where the parts of the function to be integrated close to
the point when s = t contribute more to the fractional integral than those further
away (say, at s=1.2).
This Riemann-Liouville fractional integral can be solved using the Beta func-

35
tion (also known as Euler’s Beta function) to give the following formula,

Z t
α β 1
a Jt (s − a) = (t − s)α−1 (s − a)β ds
Γ(α) a
Γ(β + 1)
= (t − a)α+β for α > 0 and β > −1
Γ(α + β + 1)

Note that if α + β < 0 then at t = a the fractional integral is undefined.


So for our example above,

Z t
1 1 1 2 1
1 Jt (s − 1) =
3 2 (t − s)− 3 (s − 1) 2 ds
Γ( 13 ) 1
1
Γ( 2 + 1) 1 1
= (t − 1) 3 + 2
Γ( 13 + 21 + 1)
Γ( 32 ) 5
= 11 (t − 1)
6
Γ( 6 )

If we change the order α of the Riemann-Liouville fractional integral, the weight-


ing factor changes. In the following, figure 6, we show the weighting factor for
various values of α.

36
Figure 6: Weighting Factors

(t−s)α−1
ˆ for α = 0, the weighting factor term Γ(α)
is undefined. But for demon-
stration purposes a small value of α = 0.01 is used so the form of the
weighting factor is made clear and plotted as the red line.

ˆ for α = 1
3
plotted as the green line,

ˆ for α = 2
3
plotted as the green dashed line,

ˆ for α = 1 plotted as the blue line

Notice that for α = 1 the weighting factor is the function ’1’, that is, the Riemann-
Liouville fractional integral takes each point in the domain of integration with
equal weighting. This is the same as the classic integral, as we would expect.
Continuing with the same colour sequence, figure 7 shows Riemann-Liouville
fractional integrals for various α values.

37
Figure 7: Riemann-Liouville Fractional Integrals

1
ˆ for α = 0: the Riemann-Liouville fractional integral remains (t−1) 2 , shown

in red,

Γ( 23 ) 5
ˆ for α = 13 : the Riemann-Liouville fractional integral is Γ( 11 )
(t − 1) 6 , shown
6

as a green line,

Γ( 23 ) 7
ˆ for α = 23 : the Riemann-Liouville fractional integral is Γ( 13 )
(t − 1) 6 , shown
6

as a dashed green line, and finally,

3
ˆ for α = 1: the Riemann-Liouville fractional integral is 23 (t − 1) 2 , the same

as the classic integral, shown in blue.

We have seen in this chapter that the approach of repeated integrals which
is extended to the integral of arbitrary order α, produces the Riemann-Liouville
fractional integral operator. We have the issue that this operator is undefined
when α = 0, so it is defined that a Jt0 f (s) = f (t). And when α = 1, this fractional

38
integral operator is the same as the classic integral. We now move on to the
Caputo fractional derivative operator.

4.2 Caputo Fractional Differential Operator

The definition of the Caputo derivative where n < α < n + 1 and n ∈ N is given
by,

c α
a Dt f (s) =a Jtn+1−α Df (s)
Z t
1 dn+1 f
= (t − s)n−α n+1 ds
Γ(n + 1 − α) a ds
for a ≤ t ≤ T, where f (s) ∈ A1 [a, T ]

where we have the classic differential in the integrand,

df
= Df (t)
dt

In this work we restrict the value of α such that 0 < α < 1, and so the definition
changes to (see Appendix A.2),

c α
a Dt f (s) =a Jt1−α Df (s)
Z t
1 df
= (t − s)−α ds
Γ(1 − α) a ds
for a ≤ t ≤ T, where f (s) ∈ A1 [a, T ]

39
For α = 0,

Z t
c 0 df
a Dt f (s) = ds = f (t) − f (a)
a ds

For α = 1, the Caputo fractional differential is undefined, but note that,

Z t
1 df
lim c α
a Dt f (s) = lim (t − s)−α ds
α→1 α→1 Γ(1 − α) a ds

which we will integrate by parts to give,

t Z t
(t − s)1−α df (t − s)1−α d2 f
 
c α 1
lim a Dt f (s) = lim − + ds
α→1 α→1 Γ(1 − α) (1 − α) ds s=a a (1 − α) ds2
 Z t 2

1 1−α df 1−α d f
= lim − 0 + (t − a) + (t − s) ds
α→1 Γ(2 − α) ds s=a a ds2
Z t 2 
df df
= + 2
ds
ds s=a s=a ds
df df df
= + −
ds s=a ds s=t ds s=a
df
=
dt

df
so we define ca Dt1 f (s) = dt
, that is, the classic differential. Finally, we state
without proof that the fractional differential operator exhibits linearity.
We now consider an example where,

ˆ the fractional differential order α = 1


3

1
ˆ the function to be integrated is (s − 1) 2

ˆ the domain of integration is from 1 to t (where t > 1)

40
which is written as,

Z t
1 1 1 1 d 1
c
1 Dt (s
3
− 1) = 2
2 (t − s)− 3 (s − 1) 2 ds
Γ( 3 ) 1 ds

Figure 8: Functions in the integrand

In figure 8 you can see the terms from the integrand, in black is the term
1
d 1 (t−s)− 3
ds
(s − 1) 2 and in green is the term Γ( 23 )
. Note that the black dashed line is the
1
function to be fractionally differentiated (s − 1) 2 . The green line is the weighting
factor where the parts of the function to be differentiated close to the point when
s = t contribute more to the Caputo fractional differential than those further
away (say, at s=1.2).
This Caputo fractional differential can be solved using the Beta function (also

41
known as Euler’s Beta function) to give the following formula,

Z t
1 d
c α
a Dt (s
β
− a) = (t − s)−α (s − a)β ds
(1 − α) a ds
Γ(β + 1)
= (t − a)β−α for α > 0 and β ≥ −1
Γ(β − α + 1)

Note that if β − α < 0 then at t = a the fractional derivative is undefined.


So for our example above,

Z t
1 1 1 1 d 1
c
1 Dt (s
3
− 1) = 2
2 (t − s)− 3 (s − 1) 2 ds
Γ( 3 ) 1 ds
1
Γ( 2 + 1) 1 1
= (t − 1) 2 − 3
Γ( 12 − 13 + 1)
Γ( 32 ) 1
= 7 (t − 1)
6
Γ( 6 )

If we change the order α of the Caputo fractional differential, the weighting


factor changes. In the following plot we show the weighting factor for various
values of α.

42
Figure 9: Weighting Factors

ˆ for α = 0 plotted as the red line

ˆ for α = 1
3
plotted as the green line,

ˆ for α = 2
3
plotted as the dashed green line,

(t−s)−α
ˆ for α = 1, the weighting factor term Γ(1−α)
is undefined. But for demon-
stration purposes a value of α = 0.99 is used so the form of the weighting
factor is made clear and plotted as the blue line.

Continuing with the colour sequence the next plot is of Caputo fractional
differentials for various α values.

43
Figure 10: Caputo Fractional Differentials

1
ˆ for α = 0 the Caputo fractional differential remains (t − 1) 2 shown in red,

Γ( 32 ) 1
ˆ for α = 1
3
the Caputo fractional differential is Γ( 76 )
(t − 1) 6 shown as a green
line,

Γ( 23 ) 1
ˆ for α = 2
3
the Caputo fractional differential is Γ( 65 )
(t − 1)− 6 shown as a
dashed green line, and finally,

1
ˆ for α = 1 the Caputo fractional integral is 12 (t − 1)− 2 the same as the classic

differential, shown in blue.

In this chapter we have seen the approach taken for the Caputo fractional differ-
ential operator being based on,

c α
a Dt f (s) =a Jt1−α Df (s)

44
where we make use of the Riemann-Liouville fractional integral and the classic
differential. When α = 0 this operator simplifies to ca Dt0 f (s) = f (t) − f (a).
We have the issue that this operator is undefined when α = 1, so we define
c 1 df
a Dt f (s) = dt
the classic differential. We now consider some relationships that
exist concerning both these fractional operators.

4.3 Fractional Operator Properties

If we summarise the Riemann-Liouville fractional integral, plus the definitions for


α = 0 and α = 1 where f ∈ A1 [a, b], we have,

0
a Jt f (s) = f (t) for α = 0
t Z
α 1
a Jt f (s) = (t − s)α−1 f (s)ds for 0 < α < 1
Γ(α) a
Z t
1
a Jt f (s) = f (s)ds for α = 1
a

and the Caputo fractional differential, plus the definitions for α = 0 and α = 1,
we have,

Z t
c 0 df
a Dt f (s) = ds = f (t) − f (a) for α = 0
a ds
Z t
1 df
c α
a Dt f (s) = (t − s)−α ds for 0 < α < 1
Γ(1 − α) a ds
df
D1 f (s) = for α = 1
dt

45
First we have used the notation D1 to represent a classic differential which by
definition acts at or close to a point. The definition is,

df f (t + δt) − f (t)
D1 f (s) = = lim
dt δt→0 δt

It follows that in classic calculus, the expression,

Z t
1 1 df
a Jt [D f (s1 )] = ds = f (t) − f (a)
a ds

is carried out over an interval R, from s = a to s = t. We now consider the


fractional calculus counterpart expression. The Caputo fractional differential
operator 0 < α < 1 includes an integral. This means that they have a non-local
property caused by the interval of integration, from s = a to s = t.

Z t
1 df
c α
a Dt f (s) = (t − s)−α ds
Γ(1 − α) a ds

The fractional counterpart expression is,

Z t  Z s 
α c α 1 α−1 1 −α df
a Jt [a Ds f (s1 )] = (t − s) (s − s1 ) ds1 ds
Γ(α) a Γ(1 − α) a ds1
= f (t) − f (a)

but now the inner integral is a function of s the integrating variable of the outer
integral. So now the operation of integration is carried out over a domain R2 , as
shown below.

46
Figure 11: Domain of Integration

This property has an influence when solving a FDDE IIP which we will see in
due course. Although this is a relatively straight forward point we have not seen
it in any literature.

Let’s consider some other classic calculus expressions and their fractional
counterparts. Consider the fundamental theorem of calculus which relates differ-
entiation and integration. Until this theorem was published in the 17th century
the two disciplines were considered as separate. It states that,

Z t
dF
f (s)ds = F (t) ⇒ = f (t)
a dt

This theorem implies the existence of a primitive F (t) for a continuous function

47
f (t) where f ∈ C1 (0, T ]. The fractional counterpart also is true for 0 < α < 1,

Z t
α 1
a Jt f (s)
= (t − s)α−1 f (s)ds = F (t)
Γ(α) a
Z t
1 dF
c α
⇒ a Dt F (s) = (t − s)−α ds = f (t)
1−α a ds

However what is sometimes called a corollary or the second fundamental theorem


of calculus cannot be extended to fractional calculus, which we now show. First
for classic calculus we have,

Z t Z b Z t
f (s)ds = f (s)ds + f (s)ds for a ≤ b ≤ t
a a b
Z t Z a
= f (s)ds − f (s)ds
b b

= F (t) − F (a)

which shows that the integral from a to t is the sum of its constituent parts and
shows us how to solve definite integrals. But for fractional calculus this is not
true since,

Z t Z b Z t
1 α−1 1 α−1 1
(t − s) f (s)ds 6= (b − s) f (s)ds + (t − s)α−1 f (s)ds
Γ(α) a Γ(α) a Γ(α) b

(7)

for the reason that the first integral on the right side has the term (b − s)α−1 and
not (t − s)α−1 in the integrand. We have treated this integral in a natural way, as
the two occurrences of b are treated equally. However, we can see no problem in
principle, with using the second suggestion (t − s)α−1 where now the equality is
satisfied. This will mean the ’definite integral’ from a to b remains a function of

48
t. This we consider to be a decision made beforehand, in the same manner you
may choose Riemann-Liouville derivatives or Caputo derivatives, however with
properties as yet unknown.
Also, since the Caputo fractional differential operator has an integral operator,
we confirm that,

Z t Z b
1 −α f (s) 1 f (s)
(t − s) ds 6= (b − s)−α ds
Γ(1 − α) a ds Γ(1 − α) a ds
Z t
1 f (s)
+ (t − s)−α ds
Γ(1 − α) b ds

When we solve the FDDE IIP we will meet an expression of this form, albeit with
differing functions f (t) defined on each interval.
The extension of the classic integral, to the Riemann-Liouville fractional in-
tegral operator and then the Caputo fractional differential operator, feels math-
ematically coherent. Even so, we have shown, that you do not have to stray too
far and the properties of the classic calculus fail to be upheld by the fractional
calculus. The next step for us, is to start some investigations into fractional
differential equations which now follow.

4.4 Fractional Ordinary Differential Equations

A FODE includes a fractional differential operator, in this case we are using


a Caputo fractional differential operator [12] [16]. So if we continue the single

49
species model by changing the ODE IVP (1), to a FODE IVP, we get,

c α
0 Dt P (s) = (b − c)P (t) where P (t) ∈ A1 [0, T ] (8)

P (0) = P0

We will now solve the FODE IVP from above (8), using Picard iteration as we
did for the ODE IVP (1). The first step is to fractionally integrate the FODE of
order α,

α c α
0 Jt [0 Ds P (s1 )] =(b − c)[0 Jtα P (s)]


Z t  Z s 
1 α−1 1 −α dP
(t − s) (s − s1 ) ds1 ds
Γ(α) 0 Γ(1 − α) 0 ds1
(b − c) t
Z
= (t − s)α−1 P (s)ds
Γ(α) 0

We will proceed now in integral notation,

t
(b − c)
Z
P (t) − P (0) = (t − s)α−1 P (s)ds where we use Appendix A.3
Γ(α) 0
(b − c) t
Z
P (t) = P (0) + (t − s)α−1 P (s)ds
Γ(α) 0

In the previous use of Picard iteration we stated that the function in the integrand
which was k(t, s) = 1, had to be continuous on 0 ≤ s ≤ t ≤ T . We now have the
function k(t, s) = (t − s)α−1 which is not continuous at t = s (for 0 < α < 1).
In the book by Linz, [20], page 47, a statement is made that helps us with this
situation. It states that Picard iteration can be used to solve such an integral

50
equation where the integrand has a term of the form k(t, s) = (t − s)−k where
0 < k < 1. So using Picard iteration, for the first iteration we use P (s) = P0 ,
hence we get,

(b − c) t
Z
P1 (t) = P0 + (t − s)α−1 P0 ds
Γ(α) 0
(b − c)tα
= P0 + P0
Γ(α + 1)

and the next iteration is given by,

(b − c) t P0 (b − c)sα
Z  
α−1
P2 (t) = P0 + (t − s) P0 + ds
Γ(α) 0 Γ(α + 1)
P0 (b − c) t
Z t
P0 (b − c)2 1
Z
α−1
= P0 + (t − s) ds + (t − s)α−1 sα ds
Γ(α) 0 Γ(α + 1) Γ(α) 0
(b − c)tα (b − c)2 t2α
= P0 + P0 + P0
Γ(α + 1) Γ(2α + 1)

which if we repeat leads to,

(b − c)tα (b − c)2 t2α (b − c)3 t3α


P (t) = P0 + P0 + P0 + P0 ...
Γ(α + 1) Γ(2α + 1) Γ(3α + 1)

X ((b − c)tα )n
= P0
n=0
Γ(αk + 1)

P (t) = P0 Eα ((b − c)tα ) where Eα is the Mittag-Leffler function

The Mittag-Leffler function is an extension of the exponential function and often


occurs in fractional differential equations. It is defined as,


X tj
Eα (t) = where α > 0 and t > 0.
j=0
Γ(jα + 1)

51
Note the connection with the exponential function as follows,

∞ ∞
X tj X tj
E1 (t) = = = et
j=0
Γ(j + 1) j=0
j!

From chapter four of the book by K.Diethelm [9], the following properties of the
Mittag-Leffler function are drawn. We have,

ˆ convergence for all t,

ˆ as t → ∞ then Eα (−λtα ) → 0 where λ > 0

ˆ as t → ∞ then |Eα (λtα )| → ∞ where λ > 0

In order to be able to investigate the solution of the FODE IVP we define


P0 = 2000 and plot some sample solutions. In the figure below the red line is the
solution with b < c and the green line is solution with b > c. Hence we can have
endless population reduction towards P (t) = 0 (red line), or endless population
growth towards P (t) = ∞ (green line) or finally, there is also a solution for b = c,
a constant solution P (t) = 2000 shown as a blue dashed line. This solution
is described as unstable since any small perturbation from this solution, either
above or below P (t) = 2000, and the population moves away towards infinity,
P (t) → ∞, or towards zero, P (t) → 0. This property is the same as that for the
ODE IVP.

52
Figure 12: Solutions of the FODE IVP

However the behaviour of the ODE IVP and the FODE IVP differ near t = 0.
We pointed out that for the ODE IVP the solutions extend to negative time a
property of the exponential function. Clearly, the gradient of the solutions remain
bounded, however the gradient of the FODE IVP solutions become infinite at
t = 0. To see why we differentiate the solution,

(b − c)tα (b − c)2 t2α (b − c)3 t3α


P (t) = P0 + P0 + P0 + P0 ...
Γ(α + 1) Γ(2α + 1) Γ(3α + 1)
dP (b − c) (b − c)2 (b − c)3
= P0 αtα−1 + P0 2αt2α−1 + P0 3αt3α−1 + ...
dt Γ(α + 1) Γ(2α + 1) Γ(3α + 1)

Taking the first term in the limit as t → 0 with 0 < α < 0, we get,

(b − c)
for b > c lim P0 αtα−1 → +∞
t→0 Γ(α + 1)
(b − c)
for b < c lim P0 αtα−1 → −∞
t→0 Γ(α + 1)

53
hence the first term causes an infinite gradient at t = 0, either negative or positive
depending on the relative values of b − c.
So does this FODE IVP reflect the behaviour of a single species population,
in particular we have this infinite gradient (either positive or negative) at t = 0.
I suggest that you would argue that the initial gradient should be a bounded
number. Hence the ODE IVP which has the exponential function as a solution
would be a more suitable model.
Given below we take this FODE IVP as our third differential equation for use
in this thesis.

c α
0 Dt y(s) = −y(t) for t > 0 and y(t) ∈ A1 (0, T ]

y(0) = y0

4.5 Fractional Delay Differential Equations

We have now identified three differential equations which are related and are all
well understood,

ˆ ODE IVP

ˆ DDE IIP

ˆ FODE IVP

We now define the fourth differential equation, a fractional delay differential equa-
tion initial interval problem (FDDE IIP), which is related to the above differential

54
equations.

c α
0 Dt y(s) = −y(t − τ ) for t > 0 and y(t) ∈ A1 (0, T ]

y(t) = ϕ(t) for t ∈ (−τ, 0] and τ > 0.

The following figure (figure 13) shows the differential equations and their rela-
tionships.

Figure 13: Relationships between Differential Equations

The ODE IVP is the simplest differential equation and is well understood.
The step to a DDE IIP has certain complications and is also well understood.
The complications include,

ˆ the delayed term y(t−τ ), which means that data has to be provided (history

function) during the period −τ < t ≤ 0, in order for the DDE IIP to be

55
solved.

ˆ in general the first derivative (of the solution) has a jump discontinuity

at t = 0, and a pattern follows, that the second derivative has a jump


discontinuity at t = τ , etc. This is described as solution smoothing.

The step from the ODE IVP to the FODE IVP provides different complications
and is well understood. They include,

ˆ the Caputo fractional derivative operator which, unlike the classic derivative

operator, which acts at or close to a point, acts over an interval from 0 to


t. This is described as a having a ’memory’.

ˆ due to the previous point, the solution has a Riemann-Liouville fractional

integral operator. Hence it includes in the integrand the term (t − s)α−1 for
0 < α < 1, so this integral is weakly singular.

The challenge of this thesis is to solve the FDDE IIP which we consider to be not
well understood. Further we hope to show that the remaining two steps in figure
13, from the,

ˆ DDE IIP to the FDDE IIP, and,

ˆ FODE IVP to the FDDE IIP

fit in coherently within the framework of this figure. This would mean we could
travel both clockwise and anti clockwise around the figure, starting at the ODE
IVP and arriving at the same FDDE IIP solution.

56
5 Existence and Uniqueness of Solutions

We now prove the existence and uniqueness of solutions to the above differential
equations. The well-known differential equations, the ODE IVP, DDE IIP and
FODE IVP, we will include fewer details and point you to the references, [14],
[9], [7], [4]. The FDDE IIP will be fully detailed but based upon the DDE IIP.
Clearly we will continue in the fully detailed analysis of the FDDE IIP in the
following chapters.

5.1 Ordinary Differential Equation IVP

Consider the ODE IVP, given by,

dy
= − y(t) for t > 0, y ∈ A1 (0, T ]
dt
y(0) = y0

We are now in a position to determine the existence and uniqueness of the


solution y(t) ∈ A1 [0, ∞) for the ODE IVP. We show that the conditions for the
Corollary of the Picard Lindelöf result are satisfied (see Corollary 1.1),

ˆ differential equation order α = 1 ⇒ 0 < α ≤ 1,

ˆ initial condition y(0) = y0 ∈ R,

ˆ f (t, y) = −y is continuous on R, and,

ˆ fulfilling the Lipschitz condition.

All the conditions of the corollary are satisfied so there exists a uniquely defined
function y ∈ A1 [0, ∞) solving the initial value problem.

57
5.2 Delay Differential Equation IIP

We now move on to the DDE IIP, consider,

dy
= − y(t − τ ) for t > 0, y ∈ A1 (0, T ]
dt
y(t) =ϕ(t) for t ∈ (−τ, 0], ϕ ∈ A1 (−τ, 0]

In the earlier example of a DDE (equation 6) we showed the existence and unique-
ness of a solution in each step in the method of steps. We repeat this method
now.

0 < t ≤ τ:
The ODE IVP is defined as,

dy
= −ϕ(t − τ ) where y(0) = ϕ(0)
dt

We show that the conditions for the Corollary of the Picard Lindelöf result are
satisfied (see Corollary 1.1),

ˆ differential equation order α = 1 ⇒ 0 < α ≤ 1,

ˆ initial condition y(0) = ϕ(0) ∈ R,

ˆ f (t, y) = −ϕ(t − τ ) is continuous on R, and,

ˆ fulfilling the Lipschitz condition.

Hence there exists a uniquely defined function y ∈ A1 [0, ∞) solving the initial
value problem. A point which needs clarification is the condition that f (t, y) =
−ϕ(t − τ ) is continuous on R. We have discussed that for the history functions

58
ϕ(t) to be continuous and real-valued on −τ < t ≤ 0 then they must be of the
form of a constant or kt, kt2 , .... Hence the functions f (t, y) = −ϕ(t − τ ) are
continuous on R.
We now quote the solution to this DDE in this first interval, taken from the
next chapter where we solve this DDE (10), and two further terms y(τ ) and
y(t − τ ).

Z t
y(t) = ϕ(0) − ϕ(s − τ )ds for 0<t≤τ
Z0 τ
y(τ ) = ϕ(0) − ϕ(s − τ )ds
0
Z t−τ
y(t − τ ) = ϕ(0) − ϕ(s − τ )ds
0

We will need these terms in the second interval, τ < t ≤ 2τ , when proving
existence and uniqueness.

τ < t ≤ 2τ :
The ODE IVP for the second interval is,

Z t−τ Z τ
dy
= −ϕ(0) + ϕ(s − τ )ds where y(τ ) = ϕ(0) − ϕ(s − τ )ds
dt 0 0

We show that the conditions for the Corollary of the Picard Lindelöf result are
satisfied (see Corollary 1.1),

ˆ differential equation order α = 1 ⇒ 0 < α ≤ 1,


ˆ initial condition y(τ ) = ϕ(0) − 0
ϕ(s − τ )ds ∈ R,

R t−τ
ˆ f (t, y) = −ϕ(0) + 0
ϕ(s − τ )ds is continuous on R, and,

59
ˆ fulfilling the Lipschitz condition.

Hence there exists a uniquely defined function y ∈ A1 [τ, ∞) solving the initial
value problem on the domain τ < t ≤ 2τ . Further, the DDE has a unique
piecewise defined solution y ∈ A1 [0, ∞) on the domain 0 < t ≤ 2τ .
In this thesis we restrict the DDE and FDDE analysis to two intervals, enough
for the principles and outcomes to be clear, but not too extensive that we get
clouded by algebra.

5.3 Fractional Ordinary Differential Equation IVP

Next we prove existence and uniqueness for solutions of the FODE IVP,

c α
0 Dt y(s) = −y(t) for t > 0, 0 < α < 1, y ∈ A1 (0, T ]

y(0) = y0

The only difference between this FODE IVP and the ODE IVP is the order of the
derivative, and since this order α satisfies the condition 0 < α ≤ 1 (see Corollary
1.1) it follows that existence and uniqueness is satisfied.

5.4 Fractional Delay Differential Equation IIP

Consider the FDDE IIP, given by,

c α
0 Dt y(s) = −y(t − τ ) for t > 0, 0 < α < 1, y ∈ A1 (0, T ]

y(t) = ϕ(t) for t ∈ (−τ, 0], ϕ ∈ A1 (−τ, 0]

60
We follow the same steps as the DDE IIP to show the existence and uniqueness
of a solution. We consider the first interval,

0 < t ≤ τ:
The FODE IVP is defined as,

c α
0 D t y(s) = −ϕ(t − τ ) where y(0) = ϕ(0)

We show that the conditions for the Corollary of the Picard Lindelöf result are
satisfied (see Corollary 1.1),

ˆ differential equation order 0 < α < 1 ⇒ 0 < α ≤ 1,

ˆ initial condition y(0) = ϕ(0) ∈ R,

ˆ f (t, y) = −ϕ(t − τ ) being continuous on R, and,

ˆ fulfilling the Lipschitz condition.

Hence there exists a uniquely defined function y ∈ A1 (0, ∞] solving the initial
value problem.
We now quote the solution to this FDDE IIP in this first interval, taken from
the next chapter when we solve this FDDE IIP (17), and two further terms, y(τ )
and y(t − τ ).

y(t) = ϕ(0) − 0 Jtα ϕ(s − τ )


Z τ
1
y(τ ) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
α
y(t − τ ) = ϕ(0) − 0 Jt−τ ϕ(s − τ )

61
We will need these terms in the second interval, when proving existence and
uniqueness.

τ < t ≤ 2τ :
The FODE IVP for the second interval is not so straight forward. The details
will become clear when we solve this FDDE IIP in the next chapter. For now we
present the FODE IVP without clarification,

Z τ
1 d
(τ − s)−α (ϕ(0) − 0 Jsα ϕ(s1 − τ ))ds + cτ Dtα y(s)
Γ(1 − α) 0 ds
α
= −ϕ(0) + 0 Jt−τ ϕ(s − τ )
Z τ
1
where y(τ ) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0

we rearrange the terms,

c α α
τ Dt y(s)= −ϕ(0) + 0 Jt−τ ϕ(s − τ )
Z τ
1 d
− (τ − s)−α (ϕ(0) − 0 Jsα ϕ(s1 − τ ))ds
Γ(1 − α) 0 ds

with initial condition,

Z τ
1
y(τ ) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0

We show that the conditions for the Corollary of the Picard Lindelöf result are
satisfied (see Corollary 1.1),

ˆ differential equation order 0 < α < 1 ⇒ 0 < α ≤ 1,


ˆ initial condition y(τ ) = ϕ(0) − 1
Γ(α) 0
(τ − s)α−1 ϕ(s − τ )ds ∈ R,

62
ˆ

Z t−τ
1
f (t, y) = −ϕ(0) + (t − τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
Z τ  Z s 
1 −α d 1 α−1
− (τ − s) ϕ(0) − (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0

being continuous on R almost everywhere, and,

ˆ fulfilling the Lipschitz condition.

Hence there exists a uniquely defined function y ∈ A1 [τ, ∞) solving the initial
value problem.

In summary we have proved that each of the four differential equations has a
solution, and further at most one solution. So in the next chapter we will solve
these differential equations exactly.

63
6 Solution Of Differential Equations

We begin by solving the three differential equations (ODE IVP, DDE IIP and
FODE IVP) that are well understood. We will plot the solutions and derivatives
of the solutions to exhibit their properties. We are particularly interested in the
DDE IIP method of solution (method of steps) as this is the method we will use
to solve the less well understood FDDE IIP. We will then solve the FDDE IIP
and plot its solution and derivatives of the solution.

6.1 Solution of a Ordinary Differential Equation IVP

We start with the ODE IVP, given by,

dy
= − y(t) for t > 0 and y(t) ∈ A1 (0, T ]
dt
y(0) = y0

For consistency in approach we solve using Picard iteration. Our first step is to
integrate the ODE,

Z t Z t
dy
ds = − y(s)ds
0 ds 0
Z t
y(t) − y(0) = − y(s)ds
0
Z t
y(t) = y0 − y(s)ds
0

The first approximate solution, given as y1 (t), is gained by substituting y(t) =


y0 in to the integrand of the integral on the right side of the Volterra integral

64
equation, and simplifying, that is,

Z t
y1 (t) = y0 − y0 ds = y0 − y0 t
0

Now we have an approximate solution y1 (t), we repeat this procedure to get an


improved approximate solution y2 (t),

Z t Z t
y2 (t) = y0 − y1 (s)ds = y0 − (y0 − y0 s)ds
0 0
Z t
t2
y2 (t) = y0 − y0 (1 − s)ds = y0 − y0 t + y0
0 2

continuing in this manner we get,

t t
s2
Z Z
y3 (t) = y0 − y2 (s)ds = y0 − (y0 − y0 s + y0 )ds
0 0 2
Z t 2
s t2 t3
y3 (t) = y0 − y0 (1 − s + )ds = y0 − y0 t + y0 − y0
0 2 2 6
2 3
 
t t
= y0 1 − t + −
2! 3!

which when repeated becomes,

t2 t3
 
y(t) =y0 1 − t + − + ...
2! 3!
=y0 e−t

65
where we note that the Taylor series is that of the exponential function e−t .
In order to visualise a solution we will now show an example.

dy
= − y(t) for t > 0 and y(t) ∈ A1 (0, T ]
dt
y(0) = 1

which has the solution,

y(t) = e−t
t2 t3
=1−t+ − + ...
2! 3!

Figure 14: Solution of the ODE

The figure above shows the solution and the first derivative. The second
derivative is the same as the original function shown in red. We continue now by

66
solving the DDE IVP.

6.2 Solution of a Delay Differential Equation IIP

Consider the DDE IIP, given by,

dy
= −y(t − τ ) for t > 0 and y(t) ∈ A1 (0, T ]
dt
y(t) = ϕ(t) for − τ < t ≤ 0 and ϕ(t) ∈ A1 (−τ, 0]

To solve this DDE IIP we use the method of steps, as we did in Section 3.2. We
will include plenty of detail in this subsection because we will be following this
method to solve the FDDE IIP.
Starting with the first step.

0 < t ≤ τ:
On this restricted domain the right-side of the DDE, y(t − τ ), is uniquely defined
by the history function, ϕ(t − τ ), so we may define an ODE IVP,

dyun
= −ϕ(t − τ ) for 0 < t ≤ τ
dt
yun (0) = ϕ(0)

where yun is an unknown function that satisfies the ODE IVP and which we seek
to determine.
To solve this ODE we integrate from 0 to t, that is,

Z t Z t
dyun
ds = − ϕ(s − τ )ds (9)
0 ds 0

67
and which we can simplify the left hand side to give,

Z t
yun (t) − ϕ(0) = − ϕ(s − τ )ds
0

and the solution is,

Z t
yun (t) = ϕ(0) − ϕ(s − τ )ds for 0 < t ≤ τ (10)
0

It is evident that

ˆ since ϕ(t − τ ) ∈ A1 (0, τ ] the solution yun (t) is continuous on 0 < t ≤ τ , and

furthermore,

ˆ since limt→0+ yun (t) = ϕ(0) the solution yun (t) is continuous with the history

function at the node t = 0.

We now move on to the next step, and follow the same idea.

τ < t ≤ 2τ :
We start with the DDE,

dy
= −y(t − τ )
dt

To ensure no details are overlooked we integrate over the whole domain from
s = 0 to s = t. So for the left side of the DDE we will get the terms,

Z τ  Z s 
d
ϕ(0) − ϕ(s1 − τ )ds1 ds
0 ds 0

68
with the solution from the first step as part of the integrand, and,

Z t
dyun
ds
τ ds

with the unknown solution. For the right side of the ODE we get the terms,

Z τ
ϕ(s − τ )ds
0

with the history function, and finally,

Z t Z s−τ 
ϕ(0) − ϕ(s1 − τ )ds1 ds
τ 0

with the solution from the first step. Putting this together we get,

Z τ  Z s  Z t
d dyun
ϕ(0) − ϕ(s1 − τ )ds1 ds + ds
0 ds 0 τ ds
Z τ Z t Z s−τ 
=− ϕ(s − τ )ds − ϕ(0) − ϕ(s1 − τ )ds1 ds
0 τ 0

We now show that there is a cancellation between the first and the third term.
If we take equation 9, that is,

Z t Z t
dy
ds = − ϕ(s − τ )ds
0 ds 0

Rs
where y(t) = ϕ(0) − 0
ϕ(s1 − τ )ds1 , giving,

Z t  Z s  Z t
d
ϕ(0) − ϕ(s1 − τ )ds1 ds = − ϕ(s − τ )ds
0 ds 0 0

69
and at t = τ we have,

Z τ  Z s  Z τ
d
ϕ(0) − ϕ(s1 − τ )ds1 ds = − ϕ(s − τ )ds
0 ds 0 0

so the first and third term cancel to give,

Z t Z t Z s−τ 
dyun
ds = − ϕ(0) − ϕ(s1 − τ )ds1 ds
τ ds τ 0

We now simplify and rearrange to give,

Z t Z s−τ 
yun (t) − y(τ ) = − ϕ(0) − ϕ(s1 − τ )ds1 ds
τ 0
Z t Z s−τ 
yun (t) = y(τ ) − ϕ(0) − ϕ(s1 − τ )ds1 ds
τ 0


and substitute that y(τ ) = ϕ(0) − 0
ϕ(s − τ )ds from the solution for 0 < t ≤ τ
to give the solution,

Z τ Z t Z s−τ 
yun (t) = ϕ(0) − ϕ(s − τ )ds − ϕ(0) − ϕ(s1 − τ )ds1 ds (11)
0 τ 0

for τ < t ≤ 2τ

We can repeat the argument earlier to determine that the solution around the
node t = τ is continuous.
In summary we have a piecewise defined solution,

Z t
0<t≤τ : y(t) = ϕ(0) − ϕ(s − τ )ds
0
Z τ Z t Z s−τ 
τ < t ≤ 2τ : y(t) = ϕ(0) − ϕ(s − τ )ds − ϕ(0) − ϕ(s1 − τ )ds1 ds
0 τ 0

70
In order to visualise a solution we will now show an example [28],

dy
= − y(t − τ ) for t > 0, y ∈ A1 (0, T ]
dt
y(t) = 1 for t ∈ (−τ, 0]

which has the piecewise solution, including an extra solution defined on 2τ < t ≤
3τ ,

0<t≤τ : y(t) = 1 − t (12)


(t − τ )2
τ < t ≤ 2τ : y(t) = 1 − t + (13)
2
(t − τ )2 (t − 2τ )3
2τ < t ≤ 3τ : y(t) = 1 − t + − (14)
2 6

Figure 15: Solution of the DDE

71
We continue now by solving the FODE IVP.

6.3 Solution of a Fractional Ordinary Differential Equa-

tion IVP

Consider the FODE IVP, given below in both fractional and integral notation,

c α
0 Dt y(s) = −y(t) for t > 0, 0 < α < 1, y(t) ∈ A1 (0, T ]

y(0) = y0

We start by fractionally integrating the FODE,

Z t  Z s  Z t
1 α−1 1 −α dy 1
(t − s) (s − s1 ) ds1 ds = − (t − s)α−1 y(s)ds
Γ(α) 0 Γ(1 − α) 0 ds1 Γ(α) 0
Z t
1
y(t) − y(0) = − (t − s)α−1 y(s)ds using appendix A.3
Γ(α) 0
Z t
1
y(t) = y0 − (t − s)α−1 y(s)ds
Γ(α) 0

where we get a Volterra integral equation of the second kind and now we use
Picard iteration to solve.

Our first approximate solution, given as y1 (t), is gained by substituting y(t) = y0


in to the integrand, that is,

Z t
1
y1 (t) = y0 − (t − s)α−1 y0 ds
Γ(α) 0
Z t
1
= y0 − y0 (t − s)α−1 ds
Γ(α) 0

= y0 − y0
Γ(α + 1)

72
Next we repeat the procedure to get y2 (t),

Z t

 
1 α−1
y2 (t) = y0 − (t − s) y0 − y0 ds
Γ(α) 0 Γ(α + 1)
Z t Z t
1 α−1 y0 1
= y0 − y0 (t − s) ds + (t − s)α−1 sα ds
Γ(α) 0 Γ(α + 1) Γ(α) 0
tα y0 Γ(α + 1)t2α
= y0 − y0 +
Γ(α + 1) Γ(α + 1) Γ(2α + 1)
tα t2α
= y0 − y0 + y0
Γ(α + 1) Γ(2α + 1)

repeating we get,

tα t2α t3α
 
y(t) = y0 1 − + − ...
Γ(α + 1) Γ(2α + 1) Γ(3α + 1)

which can be written as,

y(t) = y0 Eα (−tα )

We will now show an example of a FODE IVP, with α = 31 ,

Z t
1 1 dy
(t − s)− 3 ds = −y(t) for t > 0
Γ( 32 ) 0 ds

y(0) = 1

73
which has solution,

1
y(t) = E 1 (−t 3 )
3
∞ 1
X (−t 3 )j
=
j=0
Γ( 3j + 1)
1 2
t3 t3 t
=1− 4 + 5 − ...
Γ( 3 ) Γ( 3 ) Γ(2)

Figure 16: Solution of the FODE IVP

We continue now by solving the FDDE.

74
6.4 Solution of a Fractional Delay Differential Equation

IIP

We are now at the central part to this thesis, consider the FDDE IIP given below
which we will solve using the method of steps,

c α
0 Dt y(s) = −y(t − τ ) for t > 0, 0 < α < 1, y(t) ∈ A1 (0, T ]

y(t) = ϕ(t) for − τ < t ≤ 0, ϕ(t) ∈ A1 (−τ, 0]

We will proceed by starting with the interval 0 < t ≤ τ , and continue for τ <
t ≤ 2τ . The domain of integration will be followed closely because it deviates
from the simple intervals’ experienced when solving the DDE IIP by the method
of steps (see section 6.2). As we will show the domain of integration becomes an
area. This has no effect on the solution of the first interval but it results in an
extra term being present in the second interval.

0 < t ≤ τ:
On the restricted domain we define the FODE IVP in integral form as,

Z t
1 dyun (s)
(t − s)−α ds = −ϕ(t − τ ) (15)
Γ(1 − α) 0 ds
yun (0) = ϕ(0)

75
where yun (s) is an unknown function.
Applying Riemann integral to both sides of (15) we obtain,

Z t  Z s 
1 α−1 1 −α dyun (s)
(t − s) (s − s1 ) ds1 ds
Γ(α) 0 Γ(1 − α) 0 ds
Z t
1
=− (t − s)α−1 ϕ(s − τ )ds (16)
Γ(α) 0

We note on the LHS that we have two integral operators, and that the inner
integral is a function of s, which is the integrating variable of the outer integral.
We hence have an area domain over which we integrate, shown below.

Figure 17: First Domain of Integration

Continuing using appendix A.3, where

Z t  Z s 
1 α−1 1 −α df
(t − s) (s − s1 ) ds1 ds = f (t) − f (a)
Γ(α) a Γ(1 − α) a ds1

76
we can simplify the left hand side, to give,

Z t
1
yun (t) − ϕ(0) = − (t − s)α−1 ϕ(s − τ )ds
Γ(α) 0

hence, the solution is,

Z t
1
yun (t) = ϕ(0) − (t − s)α−1 ϕ(s − τ )ds for 0 < t ≤ τ. (17)
Γ(α) 0

Before we continue solving the FDDE we refer to two checks we have conducted
on our solution (see Section 6.4.1, part 1). Our first check was substituting our
solution into the FDDE, to see whether it is satisfied, and the second check was
where we compared our solution as limα→1 with that of the DDE IVP (see section
6.2). Both checks were satisfied.
We now continue solving the FDDE IIP, on the next interval.

τ < t < 2τ :
The FDDE IIP is,

Z t
1 dy(s)
(t − s)−α ds = −y(t − τ )
Γ(1 − α) 0 ds

The following analysis has extended lines of mathematics, so for clarity we look
at the LHS of the FDDE IIP in isolation and then return to the RHS of the
FDDE IIP. For the LHS we solve by fractionally integrating (of order α) and we
integrate s = 0 to s = t to ensure no details are overlooked. Hence we have,

Z t  Z s 
1 α−1 1 −α dy(s1 )
(t − s) (s − s1 ) ds1 ds (18)
Γ(α) 0 Γ(1 − α) 0 ds1

77
In the first step 0 < t ≤ τ , this had a domain of integration given by figure 17.
We now extend this domain for the second step to τ < t ≤ 2τ , that is,

Figure 18: Second Domain of Integration

Here we have three sub-domains over which the integration must take place,
which we label I1 , I2 and I3 .

ˆ I1 : Lower left triangle

ˆ I2 : Rectangle

ˆ I3 : Upper right triangle

Each In will have the form,

Z d  Z b 
1 α−1 1 −α dy(s1 )
In (a, b, c, d) = (d − s) (b − s1 ) ds1 ds
Γ(α) c Γ(1 − α) a ds1

where we have an outer integral in variable s acting upon an inner integral in


variable s1 . We must now complete the specific details for each sub-domain

78
choosing values for the variables/parameters a, b, c and d starting with I1 . Using
figure 18 for guidance, we note that,

ˆ s1 ranges from 0 to s (where 0 < s ≤ τ ),

and over this interval,


R s1
ˆ y(s1 ) = ϕ(0) − 1
Γ(α) s2 =0
(s1 − s2 )α−1 ϕ(s2 − τ )ds2 = yf ix (s1 )

which we label as yf ix (s1 ) to indicate that this is now ’fixed’. I1 is now partially
complete and given by,

Z d  Z s
1 1
I1 (c, d) = (d − s) α−1
(s − s1 )−α
Γ(α) c Γ(1 − α) 0
 Z s1  
d 1 α−1
ϕ(0) − (s1 − s2 ) ϕ(s2 − τ )ds2 ds1 ds
ds1 Γ(α) 0
Z d  Z s 
1 α−1 1 −α dyf ix
= (d − s) (s − s1 ) ds1 ds
Γ(α) c Γ(1 − α) 0 ds1

Next we complete the outer integral details, we note that,

ˆ s ranges from 0 to τ ,

which gives,

Z τ  Z s 
1 α−1 1 −α dyf ix
I1 = (τ − s) (s − s1 ) ds1 ds
Γ(α) 0 Γ(1 − α) s1 =0 ds1

In a similar manner we have,

Z t  Z τ 
1 α−1 1 −α dyf ix
I2 = (t − s) (τ − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) 0 ds1
Z t  Z s 
1 α−1 1 −α dyun
I3 = (t − s) (s − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) τ ds1

79
where yun (s1 ) is the unknown solution for the interval τ < t ≤ 2τ .
Putting these three terms together gives (18),

Z t  Z s 
1 α−1 1 −α
LHS = (t − s) (s − s1 ) Dy(s1 )ds1 ds
Γ(α) 0 Γ(1 − α) 0

= I1 + I2 + I3

Z τ  Z s 
1 1
α−1 −α dyf ix
LHS = (τ − s) (s − s1 ) ds1 ds
Γ(α) 0 Γ(1 − α) 0 ds1
Z t  Z τ 
1 α−1 1 −α dyf ix
+ (t − s) (τ − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) 0 ds1
Z t  Z s 
1 α−1 1 −α dyun
+ (t − s) (s − s1 ) ds1 ds (19)
Γ(α) τ Γ(1 − α) τ ds1

We now look back to the FDDE,

Z t
1 dy(s)
(t − s)−α ds = −y(t − τ )
Γ(1 − α) 0 ds

which we are solving using fractional integration (of order α). The RHS will now
be,

Z t
1
− (t − s)α−1 y(s − τ )ds
Γ(α) 0

which in the first step had the domain, simply the interval,

Figure 19

80
So we now extend this domain for the second step to,

Figure 20

For the interval 0 < s ≤ τ we replace y(s − τ ) in the integrand by ϕ(s − τ ),


and for the interval τ < t ≤ 2τ we replace y(s − τ ) by yf ix (s − τ ), where,

Z s−τ
1
yf ix (s − τ ) = ϕ(0) − (s − τ − s1 )α−1 ϕ(s1 − τ )ds1
Γ(α) 0

hence the RHS becomes,

Z t
1
− (t − s)α−1 y(s − τ )ds
Γ(α) 0
Z τ
1
=− (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
Z t  Z s−τ 
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0
Z τ Z t
1 α−1 1
=− (τ − s) ϕ(s − τ )ds − (t − s)α−1 yf ix (s − τ )ds
Γ(α) 0 Γ(α) τ
(20)

Up to this point we have fractionally integrated the LHS and RHS of the FDDE

81
IIP on the interval τ < t ≤ 2τ , putting these together ((19)=(20)), gives,

Z τ Z s
 
1 1
α−1 −α dyf ix
(τ − s) (s − s1 ) ds1 ds
Γ(α) 0 Γ(1 − α) 0 ds1
Z t  Z τ 
1 α−1 1 −α dyf ix
+ (t − s) (τ − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) 0 ds1
Z t  Z s 
1 α−1 1 −α dyun
+ (t − s) (s − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) τ ds1
Z τ Z t
1 α−1 1
=− (τ − s) ϕ(s − τ )ds − (t − s)α−1 yf ix (s − τ )ds
Γ(α) 0 Γ(α) τ

As a first step in simplification we show that there is a cancellation between the


first and fourth term. This cancellation also took place in the solution of the
DDE (see section 6.2). We take (16),

Z t  Z s 
1 α−1 1 −α dyf ix
(t − s) (s − s1 ) ds1 ds
Γ(α) 0 Γ(1 − α) 0 ds1
Z t
1
=− (t − s)α−1 ϕ(s − τ )ds
Γ(α) 0

where,

Z s1
1
yf ix (s1 ) = ϕ(0) − (s1 − s2 )α−1 ϕ(s2 − τ )ds2
Γ(α) 0

and let t = τ ,

Z τ  Z s 
1 α−1 1 −α dyf ix
(τ − s) (s − s1 ) ds1 ds
Γ(α) 0 Γ(1 − α) 0 ds1
Z τ
1
=− (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0

82
hence the first and fourth terms cancel to leave,

Z t  Z τ 
1 α−1 1 −α dyf ix
(t − s) (τ − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) 0 ds1
Z t  Z s 
1 α−1 1 −α dyun
+ (t − s) (s − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) τ ds1
Z t
1
=− (t − s)α−1 yf ix (s − τ )ds (21)
Γ(α) τ

It is at this stage that we can identify the FDDE that we are solving. It can be
taken from (21) by removing the outer fractional integrals from each term (see
(22)). We will need this equation later in the solution checks.

Z τ Z t
1 −α dyf ix 1 dyun
(τ − s) ds + (t − s)−α ds = −yf ix (s − τ )
Γ(1 − α) 0 ds Γ(1 − α) τ ds
(22)

Continuing we now simplify (21). In the first term we note that the inner integral
using variable s1 , is a definite integral of fixed limits and so maybe moved outside
of the outer integral, that is,

Z t  Z τ 
1 1
α−1 −α dyf ix
(t − s) (τ − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) 0 ds1
Z t Z τ
1 1 dyf ix
= α−1
(t − s) ds (τ − s)−α ds
Γ(α) τ Γ(1 − α) 0 ds
Z τ
(t − τ )α 1 dyf ix
= (τ − s)−α ds
Γ(α + 1) Γ(1 − α) 0 ds
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1
= (τ − s) ϕ(0) − (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1
=− (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0

83
d
where ds
ϕ(0) = 0. Hence this term can be substituted for the first term.
Next we simplify the second term by applying Appendix A.3, that is,

Z t  Z s 
1 α−1 1 −α f (s1 )
(t − s) (s − s1 ) ds1 ds = f (t) − f (a)
Γ(α) a Γ(1 − α) a ds1

so we have,

Z t  Z s 
1 α−1 1 −α dyun
(t − s) (s − s1 ) ds1 ds
Γ(α) τ Γ(1 − α) τ ds1
 Z τ 
1 α−1
= yun (t) − ϕ(0) − (τ − s) ϕ(s − τ )ds
Γ(α) 0

which we can use to substitute for the second term.


We apply both these simplifications to (21), to give,

τ Z s
(t − τ )α
Z  
1 −α d 1 α−1
− (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z τ
1
+ yun (t) − ϕ(0) + (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
Z t  Z s−τ 
1 α−1 1 α−1
=− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0

which we rearrange to give the solution,

Z τ
1
yun (t) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
α Z τ  Z s 
(t − τ ) 1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ 
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0
for τ < t ≤ 2τ

84
Before we continue we refer to two checks we have conducted on our solution
(see section 6.4.1, part 2). Our first check was substituting our solution into the
FDDE, to see whether it is satisfied, and the second check was where we compared
our solution as limα→1 with that of the DDE IVP (see section 6.2). Both checks
were satisfied.
In summary we have a piecewise defined solution,

0<t≤τ :
Z t
1
y(t) = ϕ(0) − (t − s)α−1 ϕ(s − τ )ds (23)
Γ(α) 0

τ < t ≤ 2τ :
Z τ
1
y(t) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
α Z τ  Z s 
(t − τ ) 1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ 
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0
(24)

We will now show an example of a FDDE IIP, with α = 31 ,

Z t
1 1 dy
(t − s)− 3 ds = −y(t − τ ) for t > 0 and y ∈ A1 (0, T ]
Γ( 23 ) 0 ds

y(t) = 1 for t ∈ (−τ, 0]

85
which has the piecewise solution,

1
t3
0<t≤τ : y(t) = 1 − 4
Γ( 3 )
1 2
τ3 (t − τ ) 3
τ < t ≤ 2τ : y(t) = 1 − 4 +
Γ( 3 ) Γ( 53 )

Figure 21: Solution of FDDE

We now bring together the figures of the four solutions for the ODE IVP,
DDE IIP, FDE IVP and FDDE IIP. We will make a comparison/discussion of
these solutions in section 7.1 and 7.3, but first we complete solution checks for
the FDDE IIP.

86
(a) ODE IVP (b) DDE IIP

(c) FODE IVP (d) FDDE IIP

Figure 22: Solution Comparison

6.4.1 Validation of the New Solution

In this section we will conduct two checks on the piecewise defined solutions we
determined for the FDDE IIP. This is an important section in this thesis as any
check that we can do will increase our confidence that our solution is correct.
With increased complexity the possibility of error becomes greater, so you will
find the checks for step two (where τ < t ≤ 2τ ) more involved than those for step
one (where 0 < t ≤ τ ). The first check we will carry out is checking that our
solution satisfies the FODE IVP’s we defined on each interval in the method of
steps. The second check is interesting as we determine whether the solution in
the limit as α → 1 approaches the solution of the DDE IIP, for each interval. We
begin with the first check on the first interval.

87
1) 0 < t ≤ τ :

1
Rt
Solution y(t) = ϕ(0) − Γ(α) 0
(t − s)α−1 ϕ(s − τ )ds

a) We substitute the solution into the LHS of the FDDE IIP (15), that is,

Z t
1 dy(s)
(t − s)−α ds = −ϕ(t − τ )
Γ(1 − α) 0 ds
Z t  Z s 
1 −α d 1 α−1
LHS = (t − s) ϕ(0) − (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s 
1 −α d 1 α−1
=− (t − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0

d
as ds
ϕ(0) = 0.

LHS = −ϕ(t − τ ) using Theorem 3 (see appendix A.4)

= RHS

so the solution satisfies the FDDE IIP.

b) Next in the limit as α → 1, we see whether the solution is the same as that
of the DDE IIP (10). That is,

Z t
1
y(t) = ϕ(0) − (t − s)α−1 ϕ(s − τ )ds
Γ(α) 0
 Z t 
1 α−1
lim y(t) = lim ϕ(0) − (t − s) ϕ(s − τ )ds
α→1 α→1 Γ(α) 0
Z t
= ϕ(0) − ϕ(s − τ )ds using appendix A.6.
0

which is the solution of the DDE IIP.

88
It follows that our solution for the first interval in the method of steps satisfies
both of our checks. We now continue to the next interval.

2) τ < t ≤ 2τ :

Solution

Z τ
1
y(t) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ 
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0

a) We substitute the solution into the left side (LHS) of the FDDE (22), that is,

LHS
Z τ  Z s 
1 −α d 1 α−1
= (τ − s) ϕ(0) − (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z t
1 dy(s)
+ (t − s)−α ds
Γ(1 − α) τ ds
Z τ  Z s 
1 −α d 1 α−1
= (τ − s) ϕ(0) − (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z t  Z τ
1 −α d 1
+ (t − s) ϕ(0) − (τ − s1 )α−1 ϕ(s1 − τ )ds1
Γ(1 − α) τ ds Γ(α) 0
Z τ Z s1
(s − τ )α
 
1 −α d 1 α−1
+ (τ − s1 ) (s1 − s2 ) ϕ(s2 − τ )ds2 ds1
Γ(α + 1) Γ(1 − α) 0 ds1 Γ(α) 0
Z s  Z s1 −τ  
1 α−1 1 α−1
− (s − s1 ) ϕ(0) − (s1 − τ − s2 ) ϕ(s2 − τ )ds2 ds1 ds
Γ(α) τ Γ(α) 0

89
We now separate the terms to help with simplification,

LHS
Z τ  Z s 
1 −α d 1 α−1
= (τ − s) ϕ(0) − (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z t  Z τ 
1 −α d 1 α−1
+ (t − s) ϕ(0) − (τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) τ ds Γ(α) 0
Z t
(s − τ )α

1 −α d
+ (t − s)
Γ(1 − α) τ ds Γ(α + 1)
Z τ  Z s1  
1 −α d 1 α−1
(τ − s1 ) (s1 − s2 ) ϕ(s2 − τ )ds2 ds1 ds
Γ(1 − α) 0 ds1 Γ(α) 0
Z t  Z s
1 d 1
− (t − s)−α (s − s1 )α−1
Γ(1 − α) τ ds Γ(α) τ
 Z s1 −τ  
1 α−1
ϕ(0) − (s1 − τ − s2 ) ϕ(s2 − τ )ds2 ds1 ds (25)
Γ(α) 0

To simplify we note that in the first and second terms we have,

d
(ϕ(0)) = 0
ds  Z τ 
d 1 α−1
ϕ(0) − (τ − s1 ) ϕ(s1 − τ )ds1 = 0
ds Γ(α) 0

To simplify the third term (T3) we note that the middle integral in variable
s1 has fixed limits and so maybe moved outside of both the derivative and the

90
outer integral, that is,

Z t
(s − τ )α

1 −α d
T3 = (t − s)
Γ(1 − α) τ ds Γ(α + 1)
Z τ  Z s1  
1 −α d 1 α−1
(τ − s1 ) (s1 − s2 ) ϕ(s2 − τ )ds2 ds1 ds
Γ(1 − α) 0 ds1 Γ(α) 0
Z τ  Z s1 
1 −α d 1 α−1
= (τ − s1 ) (s1 − s2 ) ϕ(s2 − τ )ds2 ds1
Γ(1 − α) 0 ds1 Γ(α) 0
Z t
(s − τ )α
 
1 −α d
(t − s) ds
Γ(1 − α) τ ds Γ(α + 1)

where we have left the first integral variable as s1 for clarity, but we will now
revert to integral variable s,

Z τ  Z s 
1 −α d 1 α−1
T3 = (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z t
(s − τ )α
 
1 −α d
(t − s) ds (26)
Γ(1 − α) τ ds Γ(α + 1)

We continue with T3 by determining the last integral, hence,

Z τ  Z s 
1 −α d 1 α−1
T3 = (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z t
1 1 d
(t − s)−α (s − τ )α ds
Γ(α + 1) Γ(1 − α) τ ds
Z τ  Z s 
1 −α d 1 α−1
= (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
1 Γ(α + 1)
(t − τ )α−α
Γ(α + 1) Γ(α − α + 1)

where we have used appendix A.2, and simplifying we get,

Z τ  Z s 
1 −α d 1 α−1
T3 = (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0

91
So putting this together the LHS of the FDDE (25) is,

LHS
Z τ  Z s 
1 −α d 1 α−1
= (τ − s) ϕ(0) − (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z t  Z τ 
1 −α d 1 α−1
+ (t − s) ϕ(0) − (τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) τ ds Γ(α) 0
Z t Z τ
(s − τ )α

1 −α d 1
+ (t − s) (τ − s1 )−α
Γ(1 − α) τ ds Γ(α + 1) Γ(1 − α) 0
 Z s1  
d 1 α−1
(s1 − s2 ) ϕ(s2 − τ )ds2 ds1 ds
ds1 Γ(α) 0
Z t  Z s
1 −α d 1
− (t − s) (s − s1 )α−1
Γ(1 − α) τ ds Γ(α) τ
 Z s1 −τ  
1 α−1
ϕ(0) − (s1 − τ − s2 ) ϕ(s2 − τ )ds2 ds1 ds
Γ(α) 0

is simplified to,

Z τ  Z s 
1 −α d 1 α−1
LHS = − (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z τ  Z s 
1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s
1 −α d 1
− (t − s) (s − s1 )α−1
Γ(1 − α) τ ds Γ(α) τ
 Z s1 −τ  
1 α−1
ϕ(0) − (s1 − τ − s2 ) ϕ(s2 − τ )ds2 ds1 ds
Γ(α) 0

The first and second terms cancel giving,

Z t  Z s
1 −α d 1
LHS = − (t − s) (s − s1 )α−1
Γ(1 − α) τ ds Γ(α) τ
 Z s1 −τ  
1 α−1
ϕ(0) − (s1 − τ − s2 ) ϕ(s2 − τ )ds2 ds1 ds
Γ(α) 0

92
which we can simplify using Appendix A.4, that is,

Z t  Z s 
1 −α d 1 α−1
(t − s) (s − s1 ) f (s1 )ds1 ds = f (t)
Γ(1 − α) τ ds Γ(α) τ

to give,

 Z t−τ 
1 α−1
LHS = − ϕ(0) − (t − τ − s) ϕ(s − τ )ds
Γ(α) 0

which is equal to the RHS of (22), and so the solution satisfies the FDDE IIP.
To conclude the solution,

Z τ
1
y(t) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ 
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0

satisfies the FDDE IIP (22) in the interval τ < t ≤ 2τ .

b) Next in the limit as α → 1, we see whether the solution is the same as that of
the DDE IIP (11). That is,

Z τ
1
y(t) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ 
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0

93
lim y(t)
α→1
 Z τ
1
= lim ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
α→1 Γ(α) 0
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ  
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0

we consider,

ˆ the continuity of the Riemann-Liouville fractional integral with respect to the

order in the limit as α → 1. See Theorem 5 (Appendix A.6), hence,

 Z t  Z s−τ  
1 α−1 1 α−1
lim (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
α→1 Γ(α) τ Γ(α) 0
Z t Z s−τ 
= ϕ(0) − ϕ(s1 − τ )ds1 ds
τ 0

ˆ the continuity of the definite integral with respect to the order in the limit as

α → 1. See Theorem 6 (Appendix A.7), hence,

 Z τ  Z τ
1 α−1
lim (τ − s) ϕ(s − τ )ds = ϕ(s − τ )ds
α→1 Γ(α) 0 0

and,

τ Z s
(t − τ )α
Z  
1 −α d 1 α−1
lim (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
α→1 Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0

(τ −s)−α
now the term, Γ(1−α)
, has the form given below as α → 1,

94
Figure 23

so we conclude that,

τ Z s
(t − τ )α
Z  
1 −α d 1 α−1
lim (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds → 0
α→1 Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0

almost everywhere.
Putting these two limits into the solution we get,

lim y(t)
α→1
 Z τ
1
= lim ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
α→1 Γ(α) 0
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ  
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0
Z τ Z t Z s−τ 
= ϕ(0) − ϕ(s − τ )ds − ϕ(0) − ϕ(s1 − τ )ds1 ds
0 τ 0

which is the same as the DDE IIP solution (11).

95
It follows that our solution for the second interval in the method of steps sat-
isfies both checks.

96
7 Analysis of Fractional Delay Differential Equa-

tion IIP Solution

7.1 Solution Comparison: FDDE IIP versus DDE IIP

It is of interest to compare the solutions for the DDE IIP (see section 6.2) and the
FDDE IIP (see section 6.4). These two differential equations are closely related
and solved with the same method (method of steps). Further, we made a decision
on how we would interpret equation 7 describing it as ’in a natural way’.

0 < t ≤ τ:

DDE IIP: y(t) = ϕ(0) −0 Jt1 ϕ(s − τ )

FDDE IIP: y(t) = ϕ(0) −0 Jtα ϕ(s − τ )

There is a clear relationship between the DDE and the FDDE solutions. The
classic integral is replaced by the Riemann-Liouville fractional integral.

97
τ < t ≤ 2τ :

Z τ
DDE IIP: y(t) = ϕ(0) − ϕ(s − τ )ds
0

− τ Jt1 (ϕ(0) − 0 Js−τ


1
ϕ(s1 − τ ))
Z τ
1
FDDE IIP: y(t) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
Z τ
(t − τ )α 1
+ (τ − s)−α D[0 Jsα ϕ(s1 − τ )]ds
Γ(α + 1) Γ(1 − α) 0
− τ Jtα (ϕ(0) − 0 Js−τ
α
ϕ(s1 − τ ))

In this interval the DDE and FDDE solutions have the same clear relationship
where the classic integrals are replaced by the Riemann-Liouville fractional inte-
grals, and the definite integrals have a similar counterpart. There is also an extra
term in the FDDE solution, which originates from the square sub-domain in the
domain of integration. This only exists in the FDDE IIP problem and not in the
DDE IIP problem.
The DDE IIP is a well understood problem and solution. Our solution for
the FDDE IIP has a clear relationship with the DDE IIP solution, and the big
difference in the second interval with the extra term, we can clearly explain (see
page 81). This evidence supports our solution.
We now move on to an analysis of the continuity and smoothness of the FDDE
IIP solution.

7.2 Solution Continuity and Smoothness for FDDE IIP

We now consider the continuity and smoothness of the solution of the FDDE IIP
at the nodes t = 0 and t = τ . We will use two history functions ϕ(t) = 1 and

98
ϕ(t) = t, which we will write as ϕ(t) = tn for n = 0, 1. So for the FDDE IIP
given by,

c α
0 Dt y(s) = −y(t − τ ) for t > 0, τ > 0, 0 < α < 1, y ∈ A1 (0, T ] (27)

ϕ(t) = tn for − τ < t ≤ 0, ϕ ∈ A1 (−τ, 0], n = 0, 1

we proceed by using the solution before the node and determining its value at
the node. We will repeat this for the first derivative. We then use the solution
after the node and using a limit, to accommodate the interval falling short of the
node, determine its value as it approaches the node. We will repeat this for the
first derivative, and then make comparisons to determine continuity. We start
with the interval,

−τ < t ≤ 0:
First we have the history function ϕ(t) = tn , and so at the node t = 0,

t0 |t=0 = 1, for n = 0.


ϕ(t) |t=0 = tn |t=0 = (28)
t1 |t=0 = 0, for n = 1.

and the first derivative of the history function at t = 0, is given by,




0 |

t=0 = 0, for n = 0.
ϕ0 (t) |t=0 = (29)

1 |t=0 = 1, for n = 1.

0 < t ≤ τ:

99
We now consider the solution of the FDDE IIP and its derivative in the interval
0 < t ≤ τ . We consider the limit as t → 0+ , and determine continuity. The
solution of the FDDE IIP is given by,

y(t) = ϕ(0) −0 Jtα ϕ(s − τ )

where ϕ(t) = tn for n = 0, 1, hence,

Z t
n 1
y(t) = t |t=0 − (t − s)α−1 (s − τ )n ds
Γ(α) 0

We now determine the behaviour of our solution in the limit as t → 0+ .

For n = 0:

t

Z
1
y(t) = 1 − (t − s)α−1 ds = 1 − (30)
Γ(α) 0 Γ(α + 1)
lim y(t) = 1 compare with (28)
t→0+

For n = 1:

Z t
1
y(t) = 0 − (t − s)α−1 (s − τ )ds
Γ(α) 0
Z t Z t
1 α−1 τ
=− (t − s) sds + (t − s)α−1 ds
Γ(α) 0 Γ(α) 0
tα+1 τ tα
=− + (31)
Γ(α + 2) Γ(α + 1)
lim y(t) = 0 compare with (28)
t→0+

100
hence the solution y(t) is continuous at t = 0 for both history functions.

We now consider the first derivative of y(t).

For n = 0 we differentiate the solution (30), and consider the behaviour of y 0 (t)
in the limit as t → 0+ ,

tα αtα−1 tα−1
 
0 d
y (t) = 1− =− =− (32)
dt Γ(α + 1) Γ(α + 1) Γ(α)
lim y 0 (t) → −∞ compare with (29)
t→0+

For n = 1 we differentiate the solution (31), and consider the behaviour of y 0 (t)
in the limit as t → 0+ ,

tα+1 τ tα (α + 1)tα ατ tα−1


 
0 d
y (t) = − + =− +
dt Γ(α + 2) Γ(α + 1) Γ(α + 2) Γ(α + 1)
α
t τ tα−1
=− + (33)
Γ(α + 1) Γ(α)
lim y 0 (t) → ∞ compare with equation 29
t→0+

Neither y 0 (t) for n = 0 nor y 0 (t) for n = 1 is equal to the result (29), and so the
first derivative is not continuous at t = 0. Both expressions have a term tα−1
which in the limit as t → 0+ will form a vertical asymptote with the t = 0 axis,
either stretching in the positive or negative y direction. Hence the first derivative
is both discontinuous at t = 0 and unbounded. The behaviour is illustrated in
the graph below which applies to both history functions and where we have used
τ = 1 to plot the graph.

101
Figure 24: First Derivative of the Solution of the FDDE IIP

102
In summary at the node t = 0, the solution is continuous however the first
derivative of the solution is not, exhibiting a vertical asymptote with the y axis.

In order for us to undertake the same analysis at the node t = τ , we now state
the solution y(t) and y 0 (t) at the point t = τ .
From (30):

τα
y(t) |t=τ = 1 − for n = 0 (34)
Γ(α + 1)

and from (31):

τ α+1 ττα
y(t) |t=τ = − + for n = 1 (35)
Γ(α + 2) Γ(α + 1)

we leave the τ τ α unsimplified for clarity.


From (32):

τ α−1
y 0 (t) |t=τ = − for n = 0 (36)
Γ(α)

and from (33):

τα τ τ α−1
y 0 (t) |t=τ = − + for n = 1 (37)
Γ(α + 1) Γ(α)

103
τ < t ≤ 2τ :
The solution to the FDDE is given by,

Z τ
1
y(t) = ϕ(0) − (τ − s)α−1 ϕ(s − τ )ds
Γ(α) 0
α Z τ  Z s 
(t − τ ) 1 −α d 1 α−1
+ (τ − s) (s − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ 
1 α−1 1 α−1
− (t − s) ϕ(0) − (s − τ − s1 ) ϕ(s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0

where ϕ(t) = tn , for, n = 0, 1, hence,

Z τ
n 1
y(t) = t |t=0 − (τ − s)α−1 (s − τ )n ds
Γ(α) 0
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1 n
+ (τ − s) (s − s1 ) (s1 − τ ) ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z t  Z s−τ 
1 α−1 n 1 α−1 n
− (t − s) t |t=0 − (s − τ − s1 ) (s1 − τ ) ds1 ds
Γ(α) τ Γ(α) 0

For n = 0, we now simplify the solution, term by term, and show that continuity
exists at the node t = τ :

Z τ Z τ
1 α−1 n 1
(τ − s) (s − τ ) ds = (τ − s)α−1 ds
Γ(α) 0 n=0 Γ(α) 0
τα
=
Γ(α + 1)

For the third term after applying n = 0, we determine the R-L fractional integral

104
followed by the definite integral for s = 0 to τ .

Z τ Z s
(t − τ )α
 
1 −α d 1 α−1 n
(τ − s) (s − s1 ) (s1 − τ ) ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z τ Z s n=0
(t − τ )α

1 d 1
= (τ − s)−α (s − s1 )α−1 ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z τ
(t − τ )α sα
 
1 −α d
= (τ − s) ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α + 1)
(t − τ )α 1 Γ(α + 1)
= τ α−α
Γ(α + 1) Γ(α + 1) Γ(α − α + 1)
(t − τ )α
=
Γ(α + 1)

And for the final term after applying n = 0, we determine the R-L fractional
integral and then split the result into two R-L fractional integrals.

Z t  Z s−τ 
1 α−1 n 1 α−1 n
(t − s) t |t=0 − (s − τ − s1 ) (s1 − τ ) ds1 ds
Γ(α) τ Γ(α) 0 n=0
Z t  Z s−τ 
1 1
= (t − s)α−1 1 − (s − τ − s1 )α−1 ds1 ds
Γ(α) τ Γ(α) 0
Z t
(s − τ )α
 
1 α−1
= (t − s) 1− ds
Γ(α) τ Γ(α + 1)
Z t Z t
1 α−1 1 1
= (t − s) ds − (t − s)α−1 (s − τ )α ds
Γ(α) τ Γ(α + 1) Γ(α) τ
(t − τ )α 1 Γ(α + 1)
= − (t − τ )2α
Γ(α + 1) Γ(α + 1) Γ(2α + 1)
(t − τ )α (t − τ )2α
= −
Γ(α + 1) Γ(2α + 1)

105
putting this together we get,

τα (t − τ )α (t − τ )α (t − τ )2α
y(t) = 1 − + − +
Γ(α + 1) Γ(α + 1) Γ(α + 1) Γ(2α + 1)
τα (t − τ )2α
=1− + (38)
Γ(α + 1) Γ(2α + 1)

and,

τα
lim+ y(t) = 1 − compare with (34)
t→τ Γ(α + 1)

and so for n = 0, y(t) is continuous at t = τ .

We now repeat this analysis for n = 1, the terms in the solution are:

Z τ Z τ
1 α−1 n 1
(τ − s) (s − τ ) ds = (τ − s)α−1 (s − τ )ds
Γ(α) 0 Γ(α) 0
Z τn=1 Z τ
1 α−1 1
= (τ − s) sds − τ (τ − s)α−1 ds
Γ(α) 0 Γ(α) 0
α+1 α
τ ττ
= −
Γ(α + 2) Γ(α + 1)

For the third term after applying n = 1, we split the R-L fractional integral into
two (not shown) and solve them. We then have two definite integrals for s = 0
to τ which we solve.

106
Z τ Z s
(t − τ )α
 
1 −α d 1 α−1 n
(τ − s) (s − s1 ) (s1 − τ ) ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z τ Z s  n=1
(t − τ )α

1 d 1
= (τ − s)−α (s − s1 )α−1 (s1 − τ )ds1 ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α) 0
Z τ
(t − τ )α
 α+1
τ sα

1 −α d s
= (τ − s) − ds
Γ(α + 1) Γ(1 − α) 0 ds Γ(α + 2) Γ(α + 1)
Z τ
(t − τ )α dsα+1

1 1
= (τ − s)−α ds
Γ(α + 1) Γ(α + 2) Γ(1 − α) 0 ds
Z τ α

τ 1 −α ds
− (τ − s) ds
Γ(α + 1) Γ(1 − α) 0 ds
(t − τ )α
 
1 Γ(α + 2) α+1−α τ Γ(α + 1) α−α
= τ − τ
Γ(α + 1) Γ(α + 2) Γ(α + 1 − α + 1) Γ(α + 1) Γ(α − α + 1)
(t − τ )α
 
= τ −τ
Γ(α + 1)
=0

And for the final term after applying n = 1, we split the R-L fractional integral
into two and solve. We then split the resulting R-L fractional integral into two
and solve.

107
Z t  Z s−τ 
1 α−1 n 1 α−1 n
(t − s) t |t=0 − (s − τ − s1 ) (s1 − τ ) ds1 ds
Γ(α) τ Γ(α) 0 n=1
Z t  Z s−τ 
1 α−1 1 α−1
= (t − s) 0− (s − τ − s1 ) (s1 − τ )ds1 ds
Γ(α) τ Γ(α) 0
Z t  Z s−τ
1 α−1 1
= (t − s) − (s − τ − s1 )α−1 s1 ds1
Γ(α) τ Γ(α) 0
Z s−τ 
τ α−1
+ (s − τ − s1 ) ds1 ds
Γ(α) 0
Z t
(s − τ )α+1 τ (s − τ )α
 
1 α−1
= (t − s) − + ds
Γ(α) τ Γ(α + 2) Γ(α + 1)
Z t
1 1
=− (t − s)α−1 (s − τ )α+1 ds
Γ(α + 2) Γ(α) τ
Z t
τ 1
+ (t − s)α−1 (s − τ )α ds
Γ(α + 1) Γ(α) τ
1 Γ(α + 2) τ Γ(α + 1)
=− (t − τ )2α+1 + (t − τ )2α
Γ(α + 2) Γ(2α + 2) Γ(α + 1) Γ(2α + 1)
(t − τ )2α+1 τ (t − τ )2α
=− +
Γ(2α + 2) Γ(2α + 1)

putting this together we get,

τ α+1 ττα (t − τ )2α+1 τ (t − τ )2α


y(t) = − + + − (39)
Γ(α + 2) Γ(α + 1) Γ(2α + 2) Γ(2α + 1)

and,

τ α+1 ττα
lim+ y(t) = − + compare with (35)
t→τ Γ(α + 2) Γ(α + 1)

and so for n = 1, y(t) is continuous at t = τ .

It follows that for the history functions ϕ(t) = 1 and ϕ(t) = t the solution of (27)

108
is continuous at t = τ .

We now consider the first derivative of y(t).

For the history function ϕ(t) = 1:

From (38):

τα (t − τ )2α (t − τ )2α−1
 
0 d
y (t) = 1− + =
dt Γ(α + 1) Γ(2α + 1) Γ(2α)
2α−1
(t − τ )
lim+ y 0 (t) = lim+ compare with (36) (40)
t→τ t→τ Γ(2α)

For the history function ϕ(t) = t:

From (39):

τ α+1 ττα (t − τ )2α+1 τ (t − τ )2α


 
0 d
y (t) = − + + −
dt Γ(α + 2) Γ(α + 1) Γ(2α + 2) Γ(2α + 1)
2α 2α−1
(t − τ ) τ (t − τ )
= −
Γ(2α + 1) Γ(2α)
(t − τ )2α τ (t − τ )2α−1
 
0
lim y (t) = lim+ −
t→τ + t→τ Γ(2α + 1) Γ(2α)
2α−1
τ (t − τ )
lim+ y 0 (t) = lim+ − compare with (37) (41)
t→τ t→τ Γ(2α)

To determine whether we have continuity in y 0 (t) at t = τ we must consider the


behaviour of the term (t−τ )2α−1 . We have three differing cases which occur when
2α − 1 < 0, 2α − 1 = 0 and 2α − 1 > 0, for 0 < α < 1. For clarity we will graph
this behaviour (with τ = 1).

109
For case 1),

2α − 1 < 0 ⇒ 2α < 1
1
⇒α<
2

1
Figure 25: (t − τ )2α−1 , α = 3

It is clear from the graph that for n = 0 and n = 1, continuity is not possible for
1
0<α< 2
due to the unbounded behaviour.

For case 2),

2α − 1 = 0 ⇒ 2α = 1
1
⇒α=
2

1
Figure 26: (t − τ )2α−1 , α = 2

1
For α = 2
and the history function ϕ(t) = 1 we have, from (40):

(t − τ )2α−1
lim+ y 0 (t)|α= 1 = lim+ =1
t→τ 2 t→τ Γ(2α) α= 12

which is not equal to (36) and so is not continuous for any τ > 0.

110
On the other hand, for history function ϕ(t) = t, from (41) it follows that:

(t − τ )2α τ (t − τ )2α−1
 
0
lim y (t)|α= 1 = lim+ − = −τ
t→τ + 2 t→τ Γ(2α + 1) Γ(2α) α= 21

and from (37):

1 1 1
−τ 2 ττ−2 τ2
+ = −
Γ( 32 ) Γ( 12 ) Γ( 12 )

and so is not continuous at t = τ .

And finally case 3),

2α − 1 > 0 ⇒ 2α > 1
1
⇒α>
2

2
Figure 27: (t − τ )2α−1 , α = 3

1
For α > 2
and the history function ϕ(t) = 1 we have from (40):

(t − τ )2α−1
lim+ y 0 (t)|α> 1 = lim+ =0
t→τ 2 t→τ Γ(2α) α> 12

which is not equal to (36) and so is not continuous for any τ > 0. On the other
hand, for history function ϕ(t) = t from (41) it follows that:

(t − τ )2α τ (t − τ )2α−1
 
0
lim y (t)|α> 1 = lim+ − =0
t→τ + 2 t→τ Γ(2α + 1) Γ(2α) α> 12

111
and from (37):

τα τ τ α−1 τ α (α − 1)
− + =
Γ(α + 1) Γ(α) αΓ(α)

and so is not continuous at t = τ .


In conclusion, noting that 0 < α < 1;

ˆ The solution of the FDDE IIP, y(t), is continuous at t = 0 and t = τ .

ˆ The first derivative of the solution y 0 (t) is not continuous at t = 0 and t = τ .

ˆ At t = 0, y 0 (t) in the interval 0 < t ≤ τ is unbounded as t → 0+ . It displays

vertical asymptotic behaviour of the form ±tα−1 .

ˆ At t = τ , y 0 (t) in the interval τ < t ≤ 2τ displays differing behaviour as

t → τ + depending on the value of α. The details are,

– for 0 < α < 12 , we have a vertical asymptotic behaviour with t = τ + ,


1
for example (t − τ )− 3 .

– for α = 12 , we have a straight line across the whole interval.

1
– for 2
≤ α < 1 , we have a plot with an infinite gradient at t = τ + , for
1
example (t − τ ) 3 .

To exhibit these findings, we will solve some FDDE IIP’s and plot their solutions,
including first and second derivatives (see Section 7.5). We now discuss the
solution properties for all our differential equation solutions.

112
7.3 Solution Properties

In order to visualise the solutions of the four differential equations, the ODE
IVP, the DDE IIP, the FODE IVP and the FDDE IIP, we defined a simple initial
condition y(0) = 1 and history function y(t) = 1 for −τ < t ≤ 0. Also we have
1
used a value of α = 3
for the fractional derivatives. For the first three differential
equations we got the results given below (see figure 28).

(a) ODE IVP (b) DDE IIP

(c) FODE IVP

Figure 28: Solution Comparison

A question that occurred to us is, ’Can we using the properties of the above
solutions draw the solution of the FDDE IIP’ ? In order to attempt this consider
the diagram below (see figure 29).

113
Figure 29: Relationships between Differential Equations

The ODE IVP is mathematically the simplest of our differential equations. It


uses what is happening now to predict the future. Its solution is defined on the
whole ’t’ real axis. The initial condition has a well defined first derivative, that is,
the solution’s gradient at the initial condition is bounded. The DDE IIP presents
a certain complication, whereby a source of past information is used to determine
the future. There is a change in the form of the solution. We can describe this
change as ’transform (a)’ from an ODE IVP to a DDE IIP, which is seen as the
top horizontal arrow in the diagram. This transform is where the order of the
derivative remains the same, and the term y(t) changes to the delay term y(t−τ ).
The delay term is associated with the change of initial condition y(0) to a history
function φ(t) for −τ < t ≤ 0. Associated with this transform are the following
property changes to the solution. Firstly, the solution is broken into a piecewise

114
solution. Secondly, each piece of the solution has the same properties as the ODE
IVP, that is, they are defined on the whole ’t’ axis (even though we only use an
interval of it) and have a bounded first derivative at the initial condition. Finally
the smooth infinitely differentiable solution changes to a solution with a pattern
of derivatives that are not defined. That is, at t = 0 the first derivative is not
defined, at t = τ the second derivative is not defined, etc.
Next we consider ’transform (b)’ from the ODE IVP to the FODE IVP, shown
as the left vertical arrow on the diagram. This transform is where the term y(t)
remains with the argument t, and the order of the derivative changes from the
classic derivative to the fractional Caputo derivative of order α where 0 < α < 1.
The FODE IVP presents a different complication, whereby information from the
interval [0, t] is used to determine the future. There is a change in the form of the
solution from the ODE IVP, which is infinity differentiable and defined on the
whole t axis, to the FODE IVP, where the solution domain is restricted to the
initial condition to positive infinity, and the first derivative at the initial condition
is now not bounded.
The FDDE IIP and its solution exhibit a ’if and only if’ relationship. In other
words in the language of mathematical proofs we say they are equivalent (not
in an equivalent relationship which is another topic). An example of this would
1 3
be to say that 2
and 6
are equivalent. So all the aspects of the FDDE IIP are
portrayed in its solution. This also applies to all of the other differential equations
we are including in this work. So if all the aspects of each differential equation
are visible in the solution, can we, complete the following actions.

1. Apply ’transform (a)’ to the FODE IVP, shown as the bottom arrow on the

115
diagram, and draw the solution.

2. Apply ’transform (b)’ to the DDE IIP shown as the vertical arrow on the
right on the diagram, and draw the solution.

Now are these two graphs the same and are they equal to our solution of the
FDDE IIP?

(a) ODE IVP (b) DDE IIP

(c) FODE IVP (d) FDDE IIP

Figure 30: Solution Comparison

Our first impressions are that it does look as though we can travel both clock-
wise and anticlockwise around the diagram (see figure 30) and conclude with the
same form of solution. So we can apply to the ODE IVP the transformation
(a) and then transformation (b), or transformation (b) and then transformation
(a), and get the same result. But we only have two periods of the solution for

116
the FDDE IIP. Another period is really needed and hence we include this in the
proposals for further work.
Further, can we confirm in the solution that the FDDE IIP has both of the
complications described previously when determining the future behaviour? That
is, we need to consider that when the solution reaches t = τ both the delay compli-
cation y(t − τ ) and the fractional complication c0 Dtα y(s) are using the information
at y(0) and how will this develop?
So are we able to draw the solution of the FDDE IIP from the solutions of the
other differential equations? We are unsure and leave this as an open question
for discussion.

117
7.4 Issues for Numerical Methods (of Solution)

The paper by R.Garrappa and E.Kaslik titled ’On Initial Conditions for Frac-
tional Delay Differential Equations’ [15] employs a numerical method described
as a one step Adams-Bashforth method for the solution of FDDE’s. Kindly the
Matlab code has been put on the Mathworks website, and can be found with a
simple search. In the following figure 31 we have plotted four analytic solutions
of the FDDE IIP (with given α values and ϕ(t) history functions) against the
numerical solutions obtained from the Matlab code. It is not within the scope
of this thesis to make a detailed analysis of the performance of this numerical
method, and indeed the authors of the code do not claim it to be a high precision
numerical method. But it is of interest to us for comparison, so we do make
the following comments. Firstly, the numerical solution does confirm the sharp
changes in gradient seen in the analytic solution, however they are not at the
nodes t = τ and at t = 2τ . They are delayed past the nodes and furthermore this
delay increases. Indeed the numerical solution is becoming out of ’synchronisa-
tion’ with the analytic solution. Secondly, after the short distance of two steps
the difference between the analytic and numeric solutions are considerable.

118
(a) α = 13 , ϕ(t) = 1 (b) α = 23 , ϕ(t) = 1

(c) α = 13 , ϕ(t) = t (d) α = 32 , ϕ(t) = t

Figure 31: Solution/Numerical Method Comparison

Next, consider the analytical solution to the DDE IIP we have referred to
in this work. We have plotted the solution, the first derivative and the second
derivative on separate plots for clarity (see figure 32). It is known that numeri-
cal methods are best suited to finding solutions, both accurately and efficiently,
when the data they use is continuous. So we can foresee in the DDE IIP that
numerical methods that use the first derivative data could encounter a problem
at the jump discontinuity at t = 0. Further, the same problem exists with the
second derivative at t = τ , and a pattern of jump discontinuities continues with
the third derivative at t = 2τ . This pattern is referred to as ’propagation’ and
is a direct consequence of the DDE IIP. We note that after these discontinuities,

119
in the t axis, the plots become smooth and no such discontinuities are repeated.
The jump discontinuities will repeat at higher derivatives which are not likely
to be used in the numerical methods. Some numerical methods track the jump
discontinuities and then take some appropriate action to overcome the calculation
problems.

(a) Solution of DDE IIP(12) (b) First Derivative of Solution

(c) Second Derivative of Solution

Figure 32: DDE IIP Discontinuities in First and Second Derivatives

The occurrence of jump discontinuities does not end there. Consider this new

120
DDE IIP,

 
dy τ
= −y t − for t > 0, y(t) ∈ A1 (0, T ]
dt 2
y(t) = 1 for − τ < t ≤ 0

which has a similar solution as the original DDE IIP but where τ has been
replaced by τ2 . What we will now show is that the following DDE IIP with two
delay terms,

 
dy τ
= −y t − − y(t − τ ) for t > 0, y(t) ∈ A1 (0, T ]
dt 2
τ
y(t) = 1 for − < t ≤ 0
2

has more than just the discontinuities associated with the two delay terms. The
analytic solution and first/second derivatives are shown in figure 33. The solution
provides no surprises and the first derivative has a double jump height at t = 0.
The second derivative has the two propagated step discontinuities, the first at
τ
t = 2
is from the y(t − τ2 ) term and the second at t = τ is from the y(t −

τ ) term. But note there is a breakdown in smoothness at t = 2
which will
create a jump discontinuity in the third derivative. This has been created by
the combination of the two delays. So a numerical method that tracks the jump
discontinuities should be able to accommodate this extra discontinuity and its
propagated discontinuities.

121
(a) Solution of DDE IIP (two delay (b) First Derivative of Solution
terms)

(c) Second Derivative of Solution

Figure 33: DDE IIP (two delay terms) Discontinuities in First and Second Deriva-
tives

A final point on the complications due to discontinuities is known as clustering.


Consider a DDE IIP with two delay terms, one term being y(t − τ3 ) and the
τ
second term y(t − τ ). Due to the finite representation of 3
the propagation of the
discontinuity to t = τ will result in two delays very close together. The details of
these issues and others to do with numerical methods and DDE IIP can be seen
in [2] [27] [23].
We will now point to problems that higher order numerical methods could
have when solving the FDDE IIP. See figure 34 for plots of the analytic solution
and the first/second derivatives. Note that these plots include multiple lines, in

122
the interval τ < t ≤ 2τ , based on the order of the Caputo fractional derivative α,

ˆ full line plots for α < 21 ,

ˆ dashed line plot for α = 12 , and,

ˆ dotted line plot for α > 21 .

(a) Solution (b) First Derivative

(c) Second Derivative

Figure 34: FDDE IIP

If we start in the first interval, 0 < t ≤ τ , the first derivative has an infinite
discontinuity (at t = 0) which is repeated in the second derivative. In fact this
behaviour will now repeat for all subsequent derivatives. In the second interval,
τ < t ≤ 2τ , the behaviour of the solution is determined by the value of α. For
1
α < 2
there is an infinite discontinuity (solid line at t = τ ) which remains an

123
infinite discontinuity in the second and subsequent derivatives. Next, for α = 12 ,
there is a jump discontinuity (to a constant derivative, shown as a dashed line)
which becomes the zero function in the second derivative and remains so for the
1
subsequent derivatives. Finally for α > 2
there is a jump discontinuity (dotted
line) which becomes an infinite discontinuity in the second derivative and remains
so for the subsequent derivatives. So a numerical method should be able to ac-
commodate all these situations, all of which can be determined by the value of
α. In further work (see chapter 8) we have recommended that this research is
repeated for the interval 2τ < t ≤ 3τ . We would then be able to continue this
discussion. Further the impact of a FDDE IIP with two delay terms will need
to be assessed, in particular for the extra discontinuities and any problems with
clustering.
Finally, for the FDDE IIP (one term) we plot two history functions not previ-
ously considered, and then the resulting numerical solution (obtained using the
previously mentioned Matlab code). There are no simple analytic solutions avail-
able for these examples. The first plot for the history function cos(πt) exhibits
changes of gradient delayed to after the nodes at t = τ and t = 2τ whilst the
second does is smooth to the eye.

124
(a) α = 13 , ϕ(t) = cos(πt) (b) α = 31 , ϕ(t) = cos( πt
2 )

Figure 35: More Complicated History Functions

We will now present four examples of the solutions of our FDDE IIP to get a
further feeling for the nature of the solutions.

7.5 Examples of FDDE IIP’s

For the first two examples, we take the following FDDE with the history function
1
y(t) = 1, and apply two values for the fractional order, α = 3
and α = 23 ,

c α
0 Dt y(s) = −y(t − τ ) for t > 0

y(t) = 1 for − τ < t ≤ 0

For interest we include a comparison of the solutions of the DDE IIP with that
for FDDE IIP (30).

125
For 0 < t ≤ τ :

DDE IIP: y(t) = 1 − t



FDDE IIP: y(t) = 1 −
Γ(α + 1)

We see that there is a clear correlation between the solutions (39).

For τ < t ≤ 2τ : where we delay cancellation for clarity, the solutions are,

(t − τ )2
DDE IIP: y(t) = 1 − τ − (t − τ ) +
2
α
τ (t − τ )α (t − τ )α (t − τ )2α
FDDE IIP: y(t) = 1 − + − +
Γ(α + 1) Γ(α + 1) Γ(α + 1) Γ(2α + 1)

The FDDE IIP solution has an extra term, which cancels with the following term.
The DDE IIP solution also has the ability to cancel but in a different way, the
two τ terms. The post cancellation solutions are,

(t − τ )2
DDE IIP: y(t) = 1 − t +
2
α
τ (t − τ )2α
FDDE IIP: y(t) = 1 − +
Γ(α + 1) Γ(2α + 1)

Example 1: We set α = 13 ,

0 < t ≤ τ:

1
t3
y(t) = 1 − 4
Γ( 3 )

126
τ < t ≤ 2τ :

1 2
τ3 (t − τ ) 3
y(t) = 1 − 4 +
Γ( 3 ) Γ( 35 )

We plot the solutions on the appropriate intervals with τ = 1,

Figure 36: Solution and Derivatives of First Example

Note that all FDDE IIP’s will have a solution initial gradient in the first
interval of ±∞ depending on the sign of the initial condition. In this case we
have an initial gradient of −∞. In the second interval the initial solution gradient
is +∞ because the order α < 12 . In the next example α > 1/2 and so in the second
interval the initial solution gradient will be zero.

Example 2: We set α = 23 ,

127
0 < t ≤ τ:

2
t3
y(t) = 1 − 5
Γ( 3 )

τ < t ≤ 2τ :

2 4
τ3 (t − τ ) 3
y(t) = 1 − 5 +
Γ( 3 ) Γ( 37 )

Figure 37: Solution and Derivatives of Second Example

For the most part the figures represent the solutions well, except for the
solution in the second interval. Its gradient in red as limt→τ + is zero, which can
just be seen in the green curve.
For the last two examples, we take the following FDDE IIP with the history

128
1
function y(t) = t, and apply two values for the fractional order, α = 3
and α = 23 .

c α
0 Dt y(s) = −y(t − τ ) for t > 0

y(t) = t for − τ < t ≤ 0

For interest we include a comparison of the solutions of the DDE IIP with the
FDDE IIP.

For 0 < t ≤ τ :

t2
DDE IIP: y(t) = τ t −
2
τ tα tα+1
FDDE IIP: y(t) = −
Γ(α + 1) Γ(α + 2)

There is a clear correlation between the solutions.

For τ < t ≤ 2τ : where we delay simplification for clarity,

τ 2 τ (t − τ )2 (t − τ )3
DDE IIP: y(t) = τ 2 − − +
2 2 6
α α+1
ττ τ τ (t − τ )2α (t − τ )2α+1
FDDE IIP: y(t) = − − +
Γ(α + 1) Γ(α + 2) Γ(2α + 1) Γ(2α + 2)

The extra term seen in the FDDE IIP solution in the previous example is equal to
zero, so the solutions show a clear correlation. The post simplification solutions
are,

129
τ 2 τ (t − τ )2 (t − τ )3
DDE IIP: y(t) = − +
2 2 6
ατ α+1 τ (t − τ )2α (t − τ )2α+1
FDDE IIP: y(t) = − +
Γ(α + 2) Γ(2α + 1) Γ(2α + 2)

Example 3: We set α = 13 ,

0 < t ≤ τ:

1 4
τt3 t3
y(t) = 4 − 7
Γ( 3 ) Γ( 3 )

τ < t ≤ 2τ :

1 34 2 5
3
τ τ (t − τ ) 3 (t − τ ) 3
y(t) = 1 − +
Γ( 3 + 2) Γ( 32 + 1) Γ( 23 + 2)
4 2 5
τ3 τ (t − τ ) 3 (t − τ ) 3
= − +
4Γ( 34 ) Γ( 35 ) Γ( 83 )

130
Figure 38: Solution and Derivatives of Third Example

Example 4: We set α = 32 ,

0 < t ≤ τ:

2 5
τt3 t3
y(t) = 5 − 8
Γ( 3 ) Γ( 3 )

τ < t ≤ 2τ :

2 35 4 7
3
τ τ (t − τ ) 3 (t − τ ) 3
y(t) = 2 − 4 + 4
Γ( 3 + 2) Γ( 3 + 1) Γ( 3 + 2)
5 4 7
2τ 3 τ (t − τ ) 3 (t − τ ) 3
= 5 − 7 +
5Γ( 3 ) Γ( 3 ) Γ( 10
3
)

131
Figure 39: Solution and Derivatives of Fourth Example

7.6 Extension of Solutions to Two Term FDDE IVP

In an attempt to broaden the type of differential equations that our work applies
to, consider our DDE IIP example and solution.

dy
= − y(t − τ ) for t > 0, y ∈ A1 (0, T ]
dt
y(t) = 1 for t ∈ (−τ, 0]

132
which has the piecewise solution,

0<t≤τ : y(t) = 1 − t
(t − τ )2
τ < t ≤ 2τ : y(t) = 1 − t +
2
(t − τ )2 (t − 2τ )3
2τ < t ≤ 3τ : y(t) = 1 − t + −
2 6

We apply the transformation, y(t) = z(t)eat where a is a constant, first to the


DDE, so we have the terms,

dy dz
= az(t)eat + eat
dt dt
y(t − τ ) = z(t − τ )ea(t−τ )

and the DDE becomes,

dz at
az(t)eat + e = −z(t − τ )ea(t−τ )
dt
dz
+ az(t) = −e−aτ z(t − τ ) for t > 0, z ∈ A1 (0, T ]
dt

So we have transformed the DDE IIP in to a two term DDE IIP involving the
term z(t) and z(t − τ ), and importantly we can determine the solution. First we
now transform the history function,

y(t) = z(t)eat = 1 ⇒ z(t) = e−at for − τ < t ≤ 0 (42)

133
and finally we transform the solution,

0<t≤τ : z(t) = (1 − t)e−at


(t − τ )2 −at
 
τ < t ≤ 2τ : z(t) = 1 − t + e
2
(t − τ )2 (t − 2τ )3 −at
 
2τ < t ≤ 3τ : z(t) = 1 − t + − e
2 6

We plot the solution to the new DDE IIP for a = 1, where the red plot is the
solution z(t), the green plot is the first derivative and the blue plot is the second
derivative. What we can see is that similar continuity properties exist as for the
one term DDE IIP, whereby the first derivative of the solution is discontinuous
at the node t = 0 and the second derivative is discontinuous at t = τ . There is
the discontinuity in the second derivative at t = 0 which differs.

134
Figure 40: Solution of new DDE for a=1

So can we perform the same type of transform to a FDDE IIP to determine


the continuity properties of a two term FDDE IIP, that is one including the
term y(t)? We found that we cannot perform the corresponding transformation.
Consider the transform y(t) = z(t)Eα (atα ) applied to the history function (42),
that is,

y(t) = z(t)Eα (atα ) = 1 for − τ < t ≤ 0

The Mittag Leffler function is zero valued on the interval −τ < t ≤ 0, so the
history function z(t) is undefined. So no such transform exists, or anything
similar that would exhibit the same problem. Hence to investigate the solution
properties of the two term FDDE IIP, another approach will be required.

135
In the following subsection we will recommend some ideas for further work.

136
8 Further Work

It was a conscious decision to restrict the intervals upon which we solved the
FDDE IIP to 0 < t ≤ τ and τ < t ≤ 2τ . The results we obtained were unexpected
and so it was important that we spent more time exploring the solution and
conducting checks. Hence it is clear further work should include determining the
solution to the FDDE IIP in the next interval, from 2τ < t ≤ 3τ . The logic
in determining the domains of integration and cancellation of terms from the
τ < t ≤ 2τ interval can be transferred directly to the 2τ < t ≤ 3τ interval. It
is then likely that patterns will emerge that will make predictions of the solution
1
in further intervals possible. For example, the critical value of α = 2
determined
the type of behaviour of the solution in the interval τ < t ≤ 2τ . So will this
1
change to α = 3
for the interval 2τ < t ≤ 3τ ?
In our examples we illustrated the solution of the FDDE IIP using two history
functions, ϕ(t) = 1 and ϕ(t) = t. The first history function gave us an extra term
in the solution (albeit there was a cancellation), but the second history function
did not produce an extra term as it turned out to be equal to zero. The history
function ϕ(t) = t2 should be investigated to determine its influence on the solution
and to see if a pattern emerges for history functions t3 , t4 , . . ..
It would also be informative to solve the DDE IIP (two term), that is,

dy
+ ay(t) = −y(t − τ ) for t > 0 and y ∈ A1 (0, T ]
dt
y(t) = ϕ(t) for t ∈ (−τ, 0]

137
using the method of steps with the Laplace transform. This will provide a tem-
plate for the next item, the solution of a FDDE IIP (two term), given by,

c α
0 Dt y(s) + ay(t) = −y(t − τ ) for t > 0, 0 < α < 1, y(t) ∈ A1 (0, T ]

y(t) = ϕ(t) for − τ < t ≤ 0, ϕ(t) ∈ A1 (−τ, 0]

using the method of steps and the Laplace transform. Will there be a strong
correlation between the DDE IIP (two term) and FDDE IIP (two term) solutions.
Note that in the book by A.Kilbas et al (reference [19] theorem 5.15) there is a
theorem on the subject of solving a non homogeneous FODE IVP using the
method of Laplace transforms. Hence this could be used in the solution of the
FDDE IIP (two term).

138
A Appendix

A number of definitions and theorems have been collected together and included
in this appendix. It is hoped that this will help when reading this thesis. Of the
first four items, A1 and A2 are definitions and A3 and A4 are theorems that you
may be familiar with and occur numerous times in the thesis.

ˆ A1: Definition of the Riemann-Liouville fractional derivative

ˆ A2: Definition of the Caputo fractional integral

ˆ A3: Theorem a Jtα [ca Dsα f (s1 )]

ˆ A4: Theorem ca Dtα [a Jsα f (s1 )]

Rb
ˆ A5: Theorem 1
Γ(α) a
(b − s)α−1 ca Dsα f (s1 )ds

ˆ A6: Theorem limk→∞ supt∈[a+,b] | a Jtαk f (s) − a Jtα f (s)|

ˆ A7: Theorem
R a+τ R a+τ
limk→∞ sup 1
Γ(αk ) a
(a + τ − s)αk −1 f (s)ds − Γ(α)
1
a
(a + τ − s)α−1 f (s)ds

A.1 Definition ([9] Definition 2.1)

For α > 0, the Riemann-Liouville Fractional Integral operator of order α, is given


by,

Z t
α 1
a Jt f (s) = (t − s)α−1 f (s)ds
Γ(α) a

for a ≤ t ≤ b, where f (s) ∈ A1 [a, b]

139
0
For α = 0, a Jt f (s) = f (t) the identity operator.
Further,

α Γ(β + 1)
a Jt (s − a)β = (t − a)α+β for α > 0 and β > −1
Γ(α + β + 1)

A.2 Definition (based upon [19] Eqn 2.4.51)

For 0 < α < 1, the Caputo Fractional Differential Operator of order α, is given
by,

c α
a Dt f (s) = a Jt1−α Df (s)
Z t
1 df (s)
= (t − s)−α ds
Γ(1 − α) a ds
for a ≤ t ≤ b, where f (s) ∈ A1 [a, b]

c 0
For α = 0, a Dt f (s) = f (t) the identity operator.
c 1 df
For α = 1, a Dt f (s) = dt
the classic differential.
Further,

c α Γ(β + 1)
a Dt (s − a)β = (t − a)β−α for α < 1 and β > 0
Γ(β − α + 1)

A.3 Theorem (based upon [9] Theorem 3.8)

For f ∈ A1 [a, b] and α > 0

α c α
a Jt [a Ds f (s1 )] = f (t) − f (a)

140
and in integral form,

Z t  Z s 
1 α−1 1 −α df (s1 )
(t − s) (s − s1 ) ds1 ds = f (t) − f (a)
Γ(α) a Γ(1 − α) a ds1

A.4 Theorem (based upon [9] Theorem 3.7)

For f ∈ A1 [a, b] and α > 0,

c α α
a Dt [a Js f (s1 )] = f (t)

and in integral form,

Z t  Z s 
1 −α d 1 α−1
(t − s) (s − s1 ) f (s1 )ds1 ds = f (t)
(1 − α) a ds Γ(α) a

A.5 Theorem

For f ∈ A1 [a, b] and α > 0,

Z b
1
(b − s)α−1 ca Dsα f (s1 )ds = f (b) − f (a)
Γ(α) a

and in integral form,

Z b Z s
1 1 df (s1 )
(b − s) α−1
(s − s1 )−α ds1 ds = f (b) − f (a)
Γ(α) a Γ(1 − α) a ds1

141
Proof:

Z b Z s
1 1 df (s1 )
(b − s) (s − s1 )−α
α−1
ds1 ds
Γ(α) a a Γ(1 − α) ds1
Z b Z s
1 df (s1 )
= (b − s)α−1
(s − s1 )−α ds1 ds
Γ(α)Γ(1 − α) a a ds1

we change the order of integration,

Z b Z b
1 df (s1 )
= (b − s)α−1 (s − s1 )−α dsds1
Γ(α)Γ(1 − α) a ds1 s1

s−s1
and make the substitution, u = b−s1
, into the inner integral and simplify giving,

Z b Z 1
1 df (s1 )
= u−α (1 − u)α−1 duds1
Γ(α)Γ(1 − α) a ds1 0

now using the Beta function, where α, β ∈ R+ ,

Z 1
Γ(α)Γ(β)
tα−1 (1 − t)β−1 dt =
0 Γ(α + β)

we get,

b
df (s1 ) Γ(α)Γ(1 − α)
Z
1
= ds1
Γ(α)Γ(1 − α) a ds1 Γ(1)
Z b
df (s1 )
= ds
a ds1
= f (b) − f (a)

as required.

142
A.6 Theorem (Reference[9] Theorem 2.10)

Consider a sequence of positive numbers such that lim αk = α and a ≤ s ≤ t ≤


k→∞

a + τ . Then,

lim sup |a Jtαk f (s) − a Jtα f (s)| = 0


k→∞ t∈[a+,b]

and in integral form,

Z t Z t
1 αk −1 1
lim sup (t − s) f (s)ds − (t − s)α−1 f (s)ds = 0
k→∞ t∈[a+,b] Γ(αk ) a Γ(α) a

for  > 0, f ∈ C0 [a, a + τ ] and α ≥ 0.

A.7 Theorem

For a ≤ s ≤ a + τ and a sequence of positive numbers such that lim αk = α,


k→∞

Z a+τ Z a+τ
1 αk −1 1
lim sup (a + τ − s) f (s)ds − (a + τ − s)α−1 f (s)ds = 0
k→∞ Γ(αk ) a Γ(α) a

for  > 0, f ∈ C0 [a, a + τ ] and α > 0.


Proof:

Z a+τ Z a+τ
1 αk −1 1
(a + τ − s) f (s)ds − (a + τ − s)α−1 f (s)ds
Γ(αk ) a Γ(α) a
Z a+τ
(a + τ − s)αk −1 (a + τ − s)α−1
 
= f (s) − ds
a Γ(αk ) Γ(α)
Z τ αk −1
u uα−1
≤ |fmax | − du
0 Γ(αk ) Γ(α)

143
The integrand is zero when,

 α 1−α
uαk −1 uα−1

Γ(αk ) k
= =⇒ u∗ =
Γ(αk ) Γ(α) Γ(α)

hence

τ
uαk −1 uα−1
Z
|fmax | − du
0 Γ(αk ) Γ(α)
 Z u∗ αk −1 Z τ αk −1
uα−1 uα−1

u u
≤ |fmax | − du + − du
0 Γ(αk ) Γ(α) u∗ Γ(αk ) Γ(α)
u∗ τ 
uαk uα u αk uα
 
= |fmax | − + −
Γ(αk + 1) Γ(α + 1) 0 Γ(αk + 1) Γ(α + 1) u∗
τ αk τα uα∗ k uα∗
  
= |fmax | − +2 −
Γ(αk + 1) Γ(α + 1) Γ(αk + 1) Γ(α + 1)
→ 0 as αk → α

144
References

[1] B.Ross. A brief history and exposition of the fundamental theory of the
fractional calculus, Lecture Notes in Mathematics, vol. 457. Springer New
York, 1975, pp. 1–36.

[2] Zennaro M. Bellen A. Numerical Methods for Delay Differential Equations.


Clarendon Press, Oxford, 2003.

[3] Oana Brandibur, Roberto Garrappa, and Eva Kaslik. Stability of Systems
of Fractional-Order Differential Equations with Caputo Derivatives. Vol. 9.
8. 2021.

[4] A. Coddington E and N Levinson. Theory of Ordinary Differential Equa-


tions. Tata McGraw-Hill, 1987.

[5] C. Corduneanu. Integral Equations and Applications. Cambridge University


Press, 1991.

[6] Robert Corless, David Jeffrey, and Donald Knuth. A Sequence of Series for
the Lambert W Function. Jan. 1997, pp. 197–204.

[7] H.J. Dark. The Existence Theorem of Ordinary Differential Equations. 1940.

[8] O. Diekmann et al. Delay Equations: Functional-, Complex-, and Nonlinear


Analysis. Applied Mathematical Sciences. Springer New York, 1995.

[9] K. Diethelm. The Analysis of Fractional Differential Equations: An Application-


Oriented Exposition Using Differential Operators of Caputo Type. Lecture
Notes in Mathematics. Springer Berlin Heidelberg, 2010.

[10] K. Diethelm and N.J. Ford. Volterra Integral Equations and Fractional Cal-
culus: Do Neighbouring Solutions Intersect? 2012.

145
[11] Kai Diethelm. Smoothness Properties of Solutions of Caputo-Type Frac-
tional Differential Equations. eng. Vol. 10. 2. Institute of Mathematics and
Informatics Bulgarian Academy of Sciences, 2007, pp. 151–160.

[12] Kai Diethelm and Neville J. Ford. Analysis of Fractional Differential Equa-
tions. Vol. 265. 2. 2002, pp. 229–248.

[13] P. G. Drazin. Nonlinear Systems. Cambridge Texts in Applied Mathematics.


Cambridge University Press, 1992.

[14] R.D. Driver. Ordinary and Delay Differential Equations. Applied Mathe-
matical Sciences. Springer New York, 2012.

[15] Roberto Garrappa and Eva Kaslik. On initial conditions for fractional delay
differential equations. Vol. 90. 2020, p. 105359.

[16] R. Gorenflo and F. Mainardi. Fractional Calculus: Integral and Differential


Equations of Fractional Order. 1997.

[17] Rudolf Gorenflo and Francesco Mainardi. Fractional calculus: integral and
differential equations of fractional order, in: A. Carpinteri, F. Mainardi
(Eds.), Fractals and Fractional Calculus in Continuum Mechanics. Jan.
1997.

[18] K. Ishteva M. Properties and Applications of the Caputo Fractional Opera-


tor. Department of Mathematics Universtat Karlsruhe (TH), 2005.

[19] A. A. Kilbas, H. M. Srivastava, and J. J. Trujillo. Theory and Applications of


Fractional Differential Equations, Volume 204 (North-Holland Mathematics
Studies). USA: Elsevier Science Inc., 2006.

[20] P. Linz. Analytical and numerical methods for Volterra equations. 1985.

146
[21] M.L. Morgado, N.J. Ford, and P.M. Lima. Analysis and numerical methods
for fractional differential equations with delay. Vol. 252. Selected papers
on Computational and Mathematical Methods in Science and Engineering
(CMMSE). 2013, pp. 159–168.

[22] R. K. Nagle, E.B. Saff, and A. D. Snider. Fundamentals of Differential


Equations. Addison-Wesley, 2012.

[23] K.W. Neves and A. Feldstein. Characterisation of Jump Discontinuities for


State Dependent Delay Differential Equations. 1976.

[24] I. Podlubny. Fractional Differential Equations: An Introduction to Frac-


tional Derivatives, Fractional Differential Equations, to Methods of Their
Solution and Some of Their Applications. ISSN. Elsevier Science, 1998.

[25] Igor Podlubny, Richard L. Magin, and Iryna Trymorush. Niels Henrik Abel
and the birth of fractional calculus. Vol. 20. 5. Walter de Gruyter GmbH,
Jan. 2017.

[26] Sergei Rogosin and Francesco Mainardi. George William Scott Blair – the
pioneer of factional calculus in rheology. Vol. 6. Apr. 2014.

[27] Thompson S. Shampine L.F. Delay Differential Equations. Springer, 2009.

[28] Hal Smith. An Introduction to Delay Differential Equations with Applica-


tions to the Life Sciences. Texts in Applied Mathematics. Springer New
York, 2010.

[29] Hoang The Tuan, Ha Duc Thai, and Roberto Garrappa. An analysis of
solutions to fractional neutral differential equations with delay. Vol. 100.
2021, p. 105854.

147
[30] Open University. Mathematical methods and models MST209, unit 2,’First
order differential equations’. 2005.

[31] Sun Yi. Time-Delay Systems: Analysis and Control using the Lambert W
Function. 2009.

148

You might also like