SOLUTION TECHNIQUES AND ERROR ANALYSIS OF

GENERAL CLASSES OF PARTIAL DIFFERENTIAL

EQUATIONS

by
Wijayasinghe Arachchige Waruni Nisansala Wijayasinghe

A thesis
submitted in partial fulfillment
of the requirements for the degree of
Master of Science in Mathematics
Boise State University

May 2016
 © 2016
Wijayasinghe Arachchige Waruni
Nisansala Wijayasinghe
ALL RIGHTS RESERVED
 BOISE STATE UNIVERSITY GRADUATE COLLEGE

DEFENSE COMMITTEE AND FINAL READING APPROVALS

of the thesis submitted by

Wijayasinghe Arachchige Waruni Nisansala Wijayasinghe

Thesis Title: Solution Techniques and Error Analysis of General Classes of Partial
Differential Equations

Date of Final Oral Examination: 04 March 2016

The following individuals read and discussed the thesis submitted by student
Wijayasinghe Arachchige Waruni Nisansala Wijayasinghe, and they evaluated her
presentation and response to questions during the final oral examination. They found that
the student passed the final oral examination.

Barbara Zubik-Kowal, Ph.D. Chair, Supervisory Committee

M. Randall Holmes, Ph.D. Member, Supervisory Committee

Uwe Kaiser, Ph.D. Member, Supervisory Committee

The final reading approval of the thesis was granted by Barbara Zubik-Kowal, Ph.D.,
Chair of the Supervisory Committee. The thesis was approved for the Graduate College
by John R. Pelton, Ph.D., Dean of the Graduate College.
 DEDICATION

ෙමම කෘ% ෙ&

අඩං* +,ධා/තය/ ෙකෙර3

උන/6 ව/නා වූ,

ඇ:ම; ඇ% ව/නා වූ අය ෙ< න=

සහෘදයාෙණB

ඔබටF ෙම3 G6ම!

!
!
 ACKNOWLEDGMENTS

My heartiest gratitude goes to my advisor, Dr. Barbara Zubik-Kowal, for her

support, excellent guidance, and encouragement throughout my study. Without her
continuous, valuable assistance, this would not have been a fruitful work.

I like to express my appreciation to Dr. Uwe Kaiser and Dr. Randall Holmes for
their valuable suggestions and for having served on my committee.

My thanks also goes to Dr. Leming Qu, Dr. Jodi Mead, Dr. Grady Wright, Dr.
Jaechoul Lee, and Dr. Marion Scheepers for helping me with my coursework and
research.

Finally, I would like to thank my parents and my husband for their love and en-
couragement throughout this work.

v
 ABSTRACT

While constructive insight for a multitude of phenomena appearing in the physical

and biological sciences, medicine, engineering and economics can be gained through
the analysis of mathematical models posed in terms of systems of ordinary and
partial di↵erential equations, it has been observed that a better description of the
behavior of the investigated phenomena can be achieved through the use of functional
di↵erential equations (FDEs) or partial functional di↵erential equations (PFDEs).
PFDEs or functional equations with ordinary derivatives are subclasses of FDEs.
FDEs form a general class of di↵erential equations applied in a variety of disciplines
and are characterized by rates of change that depend on the state of the system.
As opposed to traditional partial di↵erential equations (PDEs), the formulation of
PFDEs, and hence, their methods of solution, are generally significantly complicated
by the functional dependence of the system. Consequently, mathematical analysis
has become essential to address important questions on PFDEs, their properties and
solutions. This thesis is devoted to a general class of parabolic PFDEs and works
out the details of the proof techniques of a related paper that help to address these
questions. In particular, we examine error bounds of approximate solutions with the
aim to address whether or not they converge to the exact solutions as a result of
refining the associated discretizations.

vi
 TABLE OF CONTENTS

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Numerical Solutions for a General Class of Systems of Ordinary

Functional Di↵erential Equations: Approximations and Definitions 6

3 Iterative Processes with General Splitting Functions for Semi-Discrete

Di↵erential Functional Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

4 Consistency Properties of Numerical Schemes for Partial Functional

Di↵erential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

5 Generalized Lipschitz Conditions for Numerical Schemes Applied to

Partial Functional Di↵erential Equations . . . . . . . . . . . . . . . . . . . . . 24

6 Theorems on Error Bounds for Numerical Solutions to General Par-

tial Functional Di↵erential Equations . . . . . . . . . . . . . . . . . . . . . . . . 30

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

vii
 1

CHAPTER 1

INTRODUCTION

Many real-life problems can be adequately modeled in terms of systems of di↵erential

equations, creating a general class into which all ordinary and partial di↵erential
equations fall. A variety of models written in terms of di↵erential equations feature
the independent time variable t, which plays a significant role in predicting the future
behavior of certain phenomena in question with the aid of a variety of di↵erent
available data, such as clinical, experimental, or field data. Although it is collected in
a limited period of time, this data can be used in tandem with systems of di↵erential
equations to provide information about the future.

However, in the context of mathematical models formulated in terms of di↵erential

equations, the derivative of the solution with respect to the time variable t depends
only on the solution at the present time. To obtain more realistic models that better
approximate the behavior of the investigated phenomena, models written in terms
of di↵erential equations are very often improved by incorporating information about
the past history of the solution, giving rise to functional di↵erential equations. As
opposed to traditional di↵erential equations, in models written in terms of functional
di↵erential equations, the derivative of the solution depends on its values at earlier
times in addition to its value at the present time. This flexible property provides
an important tool in mathematical modeling with functional di↵erential equations
 2

broadly applied in many scientific disciplines such as biology, medicine, physics, engi-
neering, economics, etc. Throughout its long history, functional di↵erential equations
have been investigated by many authors with respect to a multitude of aspects, about
which we refer the reader to [1], [2], [5]-[10], [12], [14]-[21], [23]-[28], [30], [31], [38]
for ordinary functional di↵erential equations and [3], [4], [11], [13], [22], [29]-[38]
for partial functional di↵erential equations. Aspects connected with modeling with
functional di↵erential equations are presented in [3], [4], [13], [14], [24]. One of the
main problems is that many of the equations have no analytical formulae for the exact
solutions and it has become essential to study their approximations in order to gain
insight about the solutions to the model equations and to conduct numerical simula-
tions. In order to get reliable approximate solutions, careful mathematical analysis of
their errors has to be conducted. In this thesis, we expand on the development of [37]
by filling in the details of the proofs to investigate errors of approximate solutions to
a class of parabolic partial functional di↵erential equations. For similar developments
and proof techniques for partial functional di↵erential equations as well as numerical
experiments for this class of equations, we refer the reader to [4, 22, 34, 35, 36, 38].

Originating from a multitude of areas of application to the real world around us,
partial functional di↵erential equations form a general class of problems that includes
partial di↵erential equations as one if its subclasses.

This thesis is devoted to partial functional di↵erential equations written in the

form
@u ⇣ @u @ 2u ⌘
(x, t) = f x, t, u(x,t) , (x, t), 2 (x, t) , (1.1)
@t @x @x

where u 2 C(B, R) is an unknown function (see below for B), x 2 [ L, L], t 2 [0, T ]
and f : [ L, L] ⇥ [0, T ] ⇥ C(D, R) ⇥ R ⇥ R ! R, represents any given continuous
 3

function. Another generalization in (1.1) is introduced by the argument u(x,t) ; for any
fixed x 2 [ L, L] and t 2 [0, T ], u(x,t) is a function. Such a functional argument allows
to generate di↵erential equations with e.g. a time delay and shift in space. Unlike for
classical partial di↵erential equations, the third argument u(x,t) in (1.1) is not a real
value but a real function defined on D (see Figure 1 for D) and called a functional
argument. The functional argument u(x,t) 2 C(D, R) for x 2 [ L, L], t 2 [0, T ], and
u 2 C(B, R), with L, T > 0, B = [ L̂, L̂] ⇥ [ ⌧0 , T ], D = [ ⌧ˆ, ⌧ˆ] ⇥ [ ⌧0 , 0], ⌧ˆ, ⌧0 0,
L̂ = L + ⌧ˆ, is defined as

u(x,t) (s, ⌧ ) = u(x + s, t + ⌧ ), (s, ⌧ ) 2 D.

Equation (1.1) is supplemented with the following initial conditon

u(x, t) = u0 (x, t), t 2 [ ⌧0 , 0], x 2 [ L̂, L̂], (1.2)

and boundary condition

u(x, t) = g(x, t), t 2 [0, T ], x 2 [ L̂, L] [ [L, L̂]. (1.3)

Here, f : [ L, L] ⇥ [0, T ] ⇥ C(D, R) ⇥ R ⇥ R ! R is a continuous function, and

u0 , g are given initial and boundary functions, respectively.

The partial di↵erential equation (1.1) describes a general class of problems. For
example, if f is defined by

f (x, t, , p, q) = ✏q + (0, 0)(1 (0, ⌧0 )), (1.4)

where ✏ is a positive constant, then (1.1) can be written in the following form
 4

@u @ 2u
(x, t) = ✏ 2 (x, t) + u(x, t)(1 u(x, t ⌧0 )).
@t @x

Here, the functional argument is given by t ⌧0 . Another class of examples can be

generated by defining f by

Z 0 Z ⌧ˆ
f (x, t, !, p, q) = a(t)q + a(t) !(s, ⌧ )dsd⌧, (1.5)
⌧0 ⌧ˆ

where a 2 C([0, T ], R+ ). Then, (1.1) can be written in the general form

Z 0 Z ⌧ˆ
@u @ 2u
(x, t) = a(t) 2 (x, t) + a(t) u(x + s, t + ⌧ )dsd⌧.
@t @x ⌧0 ⌧ˆ

An important subclass of the class of partial functional di↵erential equations

captured by (1.1) that may be most familiar to most readers is the entire class of
partial di↵erential equations written in the form

@u ⇣ @u @ 2u ⌘
(x, t) = f˜ x, t, u(x, t), (x, t), 2 (x, t) ,
@t @x @x

where f˜ : [ L, L] ⇥ [0, T ] ⇥ R3 ! R is any given function. This entire subclass is

another example that can be generated from the general class of equations (1.1) by
suitably defining f appearing on the right-hand side.
We have seen that not only do partial functional di↵erential equations o↵er a mod-
eling approach that more realistically portrays a wide class of real-world phenomena,
but also that their generality encompasses wide classes of subproblems, some of which
many readers have been acquainted with already in various real-world contexts.
The following figure illustrates the domain of the functional argument.
 5

Space(x)

L+𝜏̂

Boundary

𝜏̂
D (x,t)

𝜏̂ 𝜏̂
D T Time(t)
−𝜏0 (0,0)

−𝜏̂

D (x,t)

-L
Boundary

-L-𝜏̂

Figure 1.1: Graphical illustration of initial and boundary sets.

 6

CHAPTER 2

NUMERICAL SOLUTIONS FOR A GENERAL CLASS OF

SYSTEMS OF ORDINARY FUNCTIONAL

DIFFERENTIAL EQUATIONS: APPROXIMATIONS AND

DEFINITIONS

The general class of equations given by (1.1) is written in terms of arbitrary functions
f and, in many cases, analytic solutions to these equations defined in the continuum
sense are unknown and approximated by numerical solutions computed on discrete
subsets. For any element of any of the discrete subsets (such elements are referred
to as grid-points), there exists an open neighborhood that is disjoint from the other
grid-points. The discrete subsets are finite and determined by parameters, a↵ecting
the coarseness of the corresponding discretizations. In the literature, the process of
semi-discretization has been also referred to as the Method of Lines. Letting the
values of these parameters approach zero causes the corresponding discretizations
to become finer. The goal of the thesis is to study error bounds of the approximate
solutions defined on the discrete subsets and to address the question of whether or not
they get closer to the exact solutions as the discretization becomes finer – a property
desired of discretization.

In this chapter, we construct approximate solutions to the general problem (1.1)–

(1.3). With this aim, we introduce spatial grid-points xj that we will use to replace
 7

the spatial derivatives in (1.1) by discrete operators. Let the spatial step-size h > 0
and M , M̂ be such that M h = L, M̂ h = ⌧ˆ and M, M̂ 2 N. Then, we define xj = jh,
for j = 0, ±1, ±2, .., ±M̃ , where M̃ = M + M̂ . Henceforth, we also use the notation
M0 = M 1 and n = 2M 1.

For each discretization parameter h, we define the vector function F = (F M 0 , . . . , FM 0 ) :

[0, T ] ⇥ Rn ⇥ C([ ⌧0 , 0], Rn ) ! Rn , the initial function ũ0 : [ ⌧0 , 0] ! Rn , and the
initial value problem
8
>
< ˙ (t) = F (t, (t), t ), t 2 [0, T ],
(2.1)
>
: (t) = ũ0 (t), t 2 [ ⌧0 , 0],

whose solution (t) 2 Rn depends on h and (as it will be shown in the next chapters)
converges to (u(x M 0 , t), ..., u(xM 0 , t)), as h ! 0. Note that n ! 1, as h ! 0, and
that the dimension of the system (2.1) increases as the discretization becomes finer.

The third input t 2 C([ ⌧0 , 0], Rn ) in (2.1) is defined by

t (⌧ ) = (t + ⌧ ),

for ⌧ 2 [ ⌧0 , 0], where t 2 [0, T ] and 2 C([ ⌧0 , T ], Rn ).

The vector function F in (2.1) can be defined as follows

2
Fi (t, z, !) = f (xi , t, Li,t !, i,t z, i,t z), (2.2)

where i = 0, ±1, ..., ±M 0 , t 2 [0, T ], z 2 Rn , and ! = (! M 0 , ..., !M 0 ) 2 C([ ⌧0 , 0], Rn ).

The operators Li,t : C([ ⌧0 , 0], Rn ) ! C(D, R) are defined in the following way
 8

xk+1 s t s xk t
[Li,t !](s, ⌧ ) = !k+i (⌧ ) + !k+1+i (⌧ ),
h h

where s 2 [ ⌧ˆ, ⌧ˆ], ⌧ 2 [ ⌧0 , 0], and k 2 N is such that xk 6 s 6 xk+1 and

8
>
< !j (⌧ ),
t
for j = 0, ±1, ..., ±M 0 ,
!j (⌧ ) =
>
: g(xj , t + ⌧ ), for j = ±M, . . . , ±M̃ ,

2
where g is defined in (1.3). The discrete operators i,t and i,t in (2.2) are defined for
t 2 [0, T ], i = 0, ±1, ..., ±M 0 , and z = (z M 0 , ..., zM 0 ) 2 Rn , by

t
zi+1 zit 1
i,t z = , (2.3)
2h
t
2 zi+1 2zit + zit 1
i,t z = , (2.4)
h2

where the vector z t = (z t M , ...zM

t
) 2 Rn+2 is defined by

8
>
< g(xi , t), for i = ±M,
zit =
>
: zi , for i = 0, ±1, . . . , ±(M 1).

As it will be shown in Chapter 4, the operators (2.3) and (2.4) approximate the first
and second order derivatives (respectively) at the point xi .

The initial function ũ0 : [ ⌧0 , 0] ! Rn in (2.1) is defined by

ũ0 (t) = (u0 (x M 0 , t), . . . , u0 (xM 0 , t)),

for t 2 [ ⌧0 , 0], where u0 : [ L̂, L̂] ⇥ [ ⌧0 , 0] ! R is the initial function given in the
original problem. The goal of this thesis is to work through the proof techniques of
[37] to help address the question of whether or not the components of the solution
 9

(t) 2 Rn to (2.1) converge, as h ! 0, to the values u(xi , t) of the exact solution to

the problem (1.1) with (1.2), (1.3).
 10

CHAPTER 3

ITERATIVE PROCESSES WITH GENERAL SPLITTING

FUNCTIONS FOR SEMI-DISCRETE DIFFERENTIAL

FUNCTIONAL SYSTEMS

In this chapter, we construct iterative procedures for solving the general problem
(2.1) and thus (1.1)–(1.3). The process is summarised in the form of the following
algorithm. Let (0)
: [ ⌧0 , T ] ! Rn be an arbitrary function. We define the sequence
of vector functions (k)
: [ ⌧0 , T ] ! Rn , where k = 0, 1, 2, . . . , recursively, by

(k)
˙ (k+1) (t) = G(t, (k+1)
(t), (k)
(t), t ), t 2 [0, T ],
(3.1)
(k+1)
(t) = ũ0 (t), t 2 [ ⌧0 , 0].

The functions G are chosen according to the given F and are referred to as splitting
(0)
functions. The function is referred to as a starting function, and the functions
(k)
are referred to as the successive iterates.

For example, if G is defined by

Gi (t, ⇣, z, w) = Fi (t, z1 , . . . , zi 1 , ⇣i , zi+1 , . . . , zn , w),

for i = 1, 2, . . . , n, where Fi is defined, for example, by (2.2), then (3.1) generates the
following iterative process of the Picard type in the functional sense
 11

⇣ ⌘
(k+1) (k) (k) (k+1) (k) (k) (k)
˙i (t) = Fi t, 1 (t), ..., i 1, i (t), i+1 (t), ..., n (t), t (3.2)

or if G is defined by

Gi (t, ⇣, z, w) = Fi (t, ⇣1 , . . . , ⇣i , zi+1 , . . . , zn , w),

then (3.1) generates another iterative process of the Picard type in the functional
sense

⇣ ⌘
(k+1) (k+1) (k+1) (k+1) (k) (k) (k)
˙i (t) = Fi t, 1 (t), ..., i 1 (t), i (t), i+1 (t), ... n (t), t . (3.3)

A common feature of these two processes is that the functional argument in F is

given by the previous iterate denoted by the superscript k, as is the case for standard
Picard iterations. The first of these processes has been referred to in the literature also
by terms such as Jacobi-Picard scheme, Jacobi-Picard waveform relaxation scheme,
Jacobi-Picard waveform method, Jacobi-Picard iteration scheme, or simply Jacobi
waveform relaxation. The second of these processes has been referred to in the
literature also by terms such as Gauss-Seidel-Picard scheme, Gauss-Seidel-Picard
waveform relaxation scheme, Gauss-Seidel-Picard waveform method, Gauss-Seidel-
Picard iteration scheme, or simply Gauss-Seidel waveform relaxation . Both of these
processes have also been collectively referred to by terms such as waveform relaxation,
dynamic iteration or simply, iterative processes. See e.g. [26], [35] for di↵erent naming
conventions. If the equaton in question is a classical equation without the functional
argument, then the name ‘Picard’ is not used in the above terms. The two iterative
processes of the Picard type in the functional sense di↵er in the dependence of the
 12

indices of the successive iterates, and one may be preferable to the other depending
on the problem.

In our error analysis, we will use the following definitions:

e(t) = U (t) (t),

e(k) (t) = (t) (k)
(t),
E (k) (t) = U (t) (k)
(t),

where t 2 [ ⌧0 , T ], k = 0, 1, . . . , U (t) = U1 (t), . . . , Un (t) , Ui (t) = u(xi , t), and

the functions u(x, t), v(t), v (k) (t) are solutions to the three problems (1.1)–(1.3),
(2.1), (3.1), respectively. Notice that e(t) is the error of semi-discretization (2.1)
and e(k) (t) is the error of iterative process (3.1), while E (k) (t) is the error of both
the semi-discretization (2.1) and iterative process (3.1). The prefix semi indicates
that the discretization corresponds to the spatial variable only. As mentioned earlier,
another name for the process of semi-discretization is the Method of Lines.

Henceforth, we will use the following assumptions. Suppose that, for

F : [0, T ] ⇥ Rn ⇥ C([0, T ], Rn ) ! Rn ,

there exist positive continuous functions n 2 C([0, T ], R+ ) such that

lim n (t) =0 (3.4)

n!1

and
U̇i (t) Fi (t, U (t), Ut )  n (t), (3.5)

for all t 2 [0, T ] and i = M 0 , . . . , M 0 . Moreover, suppose that u 2 C (4) (B, R) (class
 13

of 4-times continuously di↵erentiable funtions from B to R) and that for the given
function
f : [ L, L] ⇥ [0, T ] ⇥ C(D, R) ⇥ R ⇥ R ! R,

there exist positive continuous functions 1 , 2 , 3 2 C([0, T ], R+ ) such that

|f (x, t, !, p, q) ¯ , p̄, q̄)|  1 (t)|p

f (x, t, ! p̄| + 2 (t)|q q̄|
(3.6)
+ 3 (t) max |!(s, ⌧ ) !
¯ (s, ⌧ )|,
(s,⌧ )2D

¯ 2 C(D, R), p, q, p̄, q̄ 2 R.

for all x 2 [ L, L], t 2 [0, T ], !, !

For the iterative processes applied to (2.1), we assume that the functions

G : [0, T ] ⇥ Rn ⇥ Rn ⇥ C([ ⌧0 , 0], Rn ) ! Rn

satisfy
G(t, r(t), r(t), rt ) = F (t, r(t), rt ), (3.7)

for all t 2 [0, T ], r 2 C([ ⌧0 , T ], Rn ) (note that for any t 2 [0, T ], rt 2 C([ ⌧0 , 0], Rn ))
and there exist continuous functions µ1 2 C([0, T ], R), µ2 , µ3 2 C([0, T ], R+ ) such that

k& &¯ "[G(t, &, z, !) G(t, &¯, z, !)]kn (1 "µ1 (t))k& &¯kn , (3.8)

kG(t, &, z, !) G(t, &, z̄, !)kn  µ2 (t)kz z̄kn , (3.9)

kG(t, &, z, !) ¯ )kn  µ3 (t)k!

G(t, &, z, ! ¯ k0n ,
! (3.10)

for " 0, t 2 [0, T ], &, &¯, z, z̄ 2 Rn , !, !

¯ 2 C([ ⌧0 , 0], Rn ), where k · kn is an arbitrary
norm in Rn and
 14

k!k0n = max k!(⌧ )kn ,

⌧ 2[ ⌧0 ,0]

for ! 2 C([ ⌧0 , 0], Rn ).

In the following chapters, we will show that the conditions (3.7)–(3.10) are satisfied
for the iterative processes (3.2) and (3.3) of the Picard type in the functional sense
and we will use them to derive error bounds for the general numerical schemes (2.1)
and (3.1).
 15

CHAPTER 4

CONSISTENCY PROPERTIES OF NUMERICAL

SCHEMES FOR PARTIAL FUNCTIONAL

DIFFERENTIAL EQUATIONS

In this chapter, we present results that we will apply to derive error bounds for numer-
ical solutions to partial functional di↵erential equations. In the theorem below, the
notation C (4) (B, R) refers to the class of 4-times continuously di↵erentiable functions
from B to R.

Theorem 4.1 ([37], Lemma 3.1). If u 2 C (4) (B, R) and f satisfies condition (3.6),
then F defined by (2.2) satisfies condition (3.5) with

c
n (t) = 1 (t) + 2 (t) + 3 (t) , (4.1)
n2

for t 2 [0, T ], where c is a positive constant that is independent on n.

Proof. Let i 2 {0, ±1, . . . , ±M 0 } and t 2 [0, T ] be arbitrary. From the definition of
U (t), equation (1.1), and definition (2.2), we have
 16

@u
U̇i (t) Fi (t, U (t), Ut ) = (xi , t) Fi (t, U (t), Ut )
@t
⇣ @u @ 2u ⌘
= f xi , t, u(xi ,t) , (xi , t), 2 (xi , t) Fi t, U (t), Ut
@x @x
⇣ @u @ 2u ⌘
= f xi , t, u(xi ,t) , (xi , t), 2 (xi , t)
@x @x
2
f xi , t, Li,t Ut , i,t U (t), i,t U (t) .

Hence

@u
U̇i (t) Fi (t, U (t), Ut )  1 (t) (xi , t) i,t U (t)
@x
@ 2u 2
+ 2 (t) (xi , t) i,t U (t) (4.2)
@x2

+ 3 (t) max u(xi ,t) (s, ⌧ ) Li,t Ut (s, ⌧ ) .

(s,⌧ )2D

By Taylor’s Theorem applied to u(xi+1 , t) and u(xi 1 , t), since u 2 C (4) (B, R), we
have,

h @u h2 @ 2 u h3 @ 3 u
u(xi+1 , t) = u(xi , t) + · (xi , t) + · 2 (xi , t) + · (✓i , t), (4.3)
1! @x 2! @x 3! @x3

with ✓i 2 (xi , xi+1 ), and

h @u h2 @ 2 u h3 @ 3 u
u(xi 1 , t) = u(xi , t) · (xi , t) + · (xi , t) · (⇠i , t), (4.4)
1! @x 2! @x2 3! @x3

with ⇠i 2 (xi 1 , xi ). From (4.3) and (4.4), we have

 17

@u @u Ui+1 (t) Ui 1 (t)

(xi , t) i,t U (t) = (xi , t)
@x @x 2h
@u u(xi+1 , t) u(xi 1 , t)
= (xi , t)
@x 2h

✓ ◆
@u 1 @u h3 @ 3 u h3 @ 3 u
= (xi , t) 2h · (xi , t) + · 3 (✓i , t) + · (⇠i , t)
@x 2h @x 6 @x 6 @x3
h2 @ 3 u @ 3u
= (✓i , t) + 3 (⇠i , t) .
12 @x3 @x

Since u 2 C (4) (B, R), there exists a constant C > 0 such that

@ 3u
(x, t)  C,
@x3

for all (x, t) 2 B, and

@u Ch2
(xi , t) i,t U (t)  .
@x 6

2L
From this and n + 1 = , we get
h

@u C 4L2 2CL2
(xi , t) i,t U (t)  · < . (4.5)
@x 6 (n + 1)2 3n2

For the second term of (4.2), we obtain

@ 2u 2 @ 2u Ui+1 (t) 2Ui (t) + Ui 1 (t)

(xi , t) i,t U (t) = (xi , t)
@x2 @x2 h2
(4.6)
@ 2u u(xi+1 , t) 2u(xi , t) + u(xi 1 , t)
= (xi , t)
@x2 h2

and apply
 18

h @u h2 @ 2 u h3 @ 3 u
u(xi+1 , t) = u(xi , t) + · (xi , t) + · 2 (xi , t) + · (xi , t)
1! @x 2! @x 3! @x3
h4 @ 4 u ˜
+ · (✓i , t),
4! @x4 (4.7)
h @u h2 @ 2 u h3 @ 3 u
u(xi 1 , t) = u(xi , t) · (xi , t) + · (xi , t) · (xi , t)
1! @x 2! @x2 3! @x3
h4 @ 4 u ˜
+ · (⇠i , t),
4! @x4

with ✓˜i 2 (xi , xi+1 ) and ⇠˜i 2 (xi 1 , xi ), respectively. From (4.6) and (4.7), we have

@ 2u 2
(xi , t) i,t U (t)
@x2

✓ ◆
@ 2u 1 2@ u 2
h4 @ 4 u ˜ h4 @ 4 u ˜
= (xi , t) h (x i , t) + ( ✓ i , t) + (⇠i , t)
@x2 h2 @x2 24 @x4 24 @x4
h2 @ 4 u ˜ @ 4u ˜
= ( ✓ i , t) + (⇠i , t) .
24 @x4 @x4

Since u 2 C (4) (B, R), there exists a constant C > 0 such that

@ 4u
(x, t)  C,
@x4

2L
for all (x, t) 2 B. From this and h = , we get
n+1

@ 2u 2 Ch2 4L2 C L2 C
(xi , t) i,t U (t)  = < . (4.8)
@x2 12 12(n + 1)2 3n2

Finally, for the third term in (4.2), we get

 19

(xk+1 s)
u(xi ,t) (s, ⌧ ) [Li,t Ut ](s, ⌧ ) = u(xi + s, t + ⌧ ) · (Ut )tk+i (⌧ )
h
(4.9)
(s xk )
· (Ut )tk+1+i (⌧ ) ,
h

for (s, ⌧ ) 2 D and k 2 N such that xk  s  xk+1 . Since (Ut )tj (⌧ ) = (Ut )j (⌧ ) =
Uj (t + ⌧ ) = u(xj , t + ⌧ ), for j = k + i and j = k + 1 + i, from (4.9), we get

u(xi ,t) (s, ⌧ ) [Li,t Ut ](s, ⌧ )

(xk+1 s) (s xk )
= u(xi + s, t + ⌧ ) u(xk + xi , t + ⌧ ) u(xk+1 + xi , t + ⌧ )
h h

and, by Taylor’s Theorem, we obtain

u(xi ,t) (s, ⌧ ) [Li,t Ut ](s, ⌧ )

"
(xk+1 s) (xi + xk xi s) @u
= u(xi + s, t + ⌧ ) u(xi + s, t + ⌧ ) + (xi + s, t + ⌧ )
h 1! @x

# "
2 2
(xi + xk xi s) @ u ˆ (s xk )
+ ( ✓i , t + ⌧ ) u(xi + s, t + ⌧ )
2! @x2 h
#
(xi + xk+1 xi s) @u (xi + xk+1 xi s)2 @ 2 u ˆ
+ (xi + s, t + ⌧ ) + (⇠i , t + ⌧ )
1! @x 2! @x2
" #
xk+1 s s xk
= u(xi + s, t + ⌧ ) 1
h h
" #
@u xk+1 s s xk
(xi + s, t + ⌧ ) (xk s) + (xk+1 s)
@x h h

@ 2u ˆ (xk s)2 (xk+1 s) @ 2u ˆ (s xk ) (xk+1 s)2

( ✓ i , t + ⌧ ) ( ⇠ i , t + ⌧ ) .
@x2 2 h @x2 h 2
 20

Upon rearranging this and using the inequality

@ 2u
(x, t)  C,
@x2

with C > 0, we get

" #
C
u(xi ,t) (s, ⌧ ) [Li,t Ut ](s, ⌧ )  (xk s)2 (xk+1 s) + (s xk )(xk+1 s)2
2h
" #
C 2
 h (xk+1 s) + h2 (s xk )
2h
" #
Ch Ch2
= xk+1 s + s xk =
2 2

2L
Therefore, since h = , we get
n+1

2CL2 2CL2
u(xi ,t) (s, ⌧ ) [Li,t Ut ](s, ⌧ )  < . (4.10)
(n + 1)2 n2

From (4.2), (4.5), (4.8), and (4.10), we get

2CL2 CL2 2CL2

|U̇i (t) Fi (t, U (t), Ut )|  1 (t) +  2 (t) +  3 (t) ,
3n2 3n2 n2

which shows that (3.5) holds with, for example, c = 2CL2 in (4.1), and finishes the
proof.

Corollary 4.1 ([37], Corollary 3.1). Suppose that there exists a constant d > 0 such
that
n o
max u(x, t) : x 2 [ L, L], t 2 [0, T ]  d
 21

where u is a solution of equation (1.1) with f defined by (1.4) for a class of functions
w 2 C(D, R) such that max{|w(s, ⌧ )| : (s, ⌧ ) 2 D}  d. Then the function f satisfies
condition (3.6). Moreover, if u is of class C 4 (B, R), then F defined by (2.2) satisfies
condition (3.5).

Proof. In order for Theorem 4.1 to be applied, it suffices to show (3.6). Let x 2
[ L, L], t 2 [0, T ], w, w̄ 2 C(D, R), p, q, p̄, q̄ 2 R be arbitrary. From the definition of
f , we get

f (x, t, w, p, q) f (x, t, w̄, p̄, q̄)

= "q + w(0, 0) 1 w(0, ⌧0 ) "q̄ w̄(0, 0) 1 w̄(0, ⌧0 )

= "(q q̄) + w(0, 0) w̄(0, 0) w(0, 0)w(0, ⌧0 ) + w(0, 0)w̄(0, ⌧0 )

w(0, 0)w̄(0, ⌧0 ) + w̄(0, 0)w̄(0, ⌧0 )

= "(q q̄) + w(0, 0) w̄(0, 0) w(0, 0) w(0, ⌧0 ) w̄(0, ⌧0 )

w̄(0, ⌧0 ) w(0, 0) w̄(0, 0) .

and

|f (x, t, w, p, q) f (x, t, w̄, p̄, q̄)|

 "|q q̄| + w(0, 0) w̄(0, 0) + w(0, 0) · w(0, ⌧0 ) w̄(0, ⌧0 )

+ w̄(0, ⌧0 ) · w(0, 0) w̄(0, 0)

⇣ ⌘
 "|q q̄| + 1 + w̄(0, ⌧0 ) + w(0, 0) max w(s, ⌧ ) w̄(s, ⌧ )
(s,⌧ )2D

 "|q q̄| + (1 + 2d) max w(s, ⌧ ) w̄(s, ⌧ ) .

(s,⌧ )2D

Therefore, f satisfies (3.6) with

 22

1 (t) = 0, 2 (t) = ", 3 (t) = 1 + 2d.

We now apply Theorem 4.1 and conclude that F defined by (2.2) satisfies (3.5) with

c
n (t) = (" + 1 + 2d),
n2

which finishes the proof.

Corollary 4.2 ([37], Corollary 3.2). Let d > 0 be as in Corollary 4.1 with f defined
by (1.5). Then f satisfies condition (3.6). Moreover, if u is of class C 4 (B, R), then
F defined by (2.2) and (1.5) satisfies condition (3.5).

Proof. We apply Theorem 4.1 with f defined by (1.5). Let x 2 [ L, L], t 2 [0, T ],
ˆ 2 C(D, R), p, q, p̂, q̂ 2 R be arbitrary. Since a is a positive function, from (1.5),
!, !
we get

Z 0 Z ⌧ˆ
|f (x, t, !, p, q) f (x, t, !
ˆ , p̂, q̂)| = a(t)q + a(t) !(s, ⌧ )dsd⌧
⌧0 ⌧ˆ
Z 0 Z ⌧ˆ
a(t)q̂ a(t) !
ˆ (s, ⌧ )dsd⌧
⌧0 ⌧ˆ
Z 0 Z ⌧ˆ
 a(t)|q q̂| + a(t) (!(s, ⌧ ) !
ˆ (s, ⌧ ))dsd⌧
⌧0 ⌧ˆ
Z 0 Z ⌧ˆ
 a(t)|q q̂| + a(t) |!(s, ⌧ ) !
ˆ (s, ⌧ )| dsd⌧
⌧0 ⌧ˆ
Z 0 Z ⌧ˆ
 a(t)|q q̂| + a(t) max |!(s, ⌧ ) !
ˆ (s, ⌧ )| dsd⌧
⌧0 ⌧ˆ (s,⌧ )2D
Z 0 Z ⌧ˆ
= a(t)|q q̂| + a(t) max |!(s, ⌧ ) !
ˆ (s, ⌧ )| dsd⌧
(s,⌧ )2D ⌧0 ⌧ˆ

= a(t)|q ⌧ ⌧0 a(t) max |!(s, ⌧ )

q̂| + 2ˆ !
ˆ (s, ⌧ )| ,
(s,⌧ )2D
 23

which shows that f satisfies (3.6) with

1 (t) = 0, 2 (t) = a(t), 3 (t) = 2ˆ

⌧ ⌧0 a(t).

By Theorem 4.1, the function F defined by (2.2) satisfies (3.5) with

c a(t)
n (t) = (1 + 2⌧0 ⌧ˆ)
n2

and the proof is finished.

We have proved some preliminary results that will provide useful stepping stones in
deriving error bounds for numerical solutions to general partial functional di↵erential
equations in the next chapters.
 24

CHAPTER 5

GENERALIZED LIPSCHITZ CONDITIONS FOR

NUMERICAL SCHEMES APPLIED TO PARTIAL

FUNCTIONAL DIFFERENTIAL EQUATIONS

In this chapter, we will show that the iterations (3.2) of the Picard type in the
functional sense satisfy conditions (3.7)–(3.10), which will be useful in deriving error
bounds for the general numerical schemes (2.1) and (3.1). The proof for iterations of
the type (3.3) is similar.

The main result can be summarised by the following theorem.

Theorem 5.1 ([37], Lemma 3.2). Suppose that f satisfies condition (3.6) and

@f
(x, t, !, p, q) 0, (5.1)
@q

where x 2 [ L, L], [0, T ], ! 2 C(D, R), p, q 2 R. Moreover, suppose that ⌫ 2

C([0, T ], R) is a function such that

@f
⌫(t)  (x, t, !, p, q) (5.2)
@q

for all inputs as in (5.1). Then, G defined by

 25

✓ t
◆
zi+1 zit 1
t
zi+1 2&i + zit 1
Gi (t, &, z, !) = f xi , t, Li,t !, ,
2h h2

satisfies condition (3.7). Moreover, if additionally k · kn is the infinity norm, then G

satisfies conditions (3.8)–(3.10).

Proof. First, we apply the definition of Gi and (2.2), obtaining

⇣ ri+1 (t) ri 1 (t) ri+1 (t) 2ri (t) + ri 1 (t) ⌘

Gi (t, r(t), r(t), rt ) = f xi , t, Li,t rt , ,
2h h2
⇣ ⌘
2
= f xi , t, Li,t rt , i,t r(t), i,t r(t) = Fi (t, r(t), rt ),

where t 2 [0, T ] and r 2 C([ ⌧0 , T ], R) are arbitrary. This shows that (3.7) holds.

In order to prove (3.8), we begin by using the definition of Gi and then we apply
the mean value theorem

⇣ zt zit t
zi+1 2&it + zit 1 ⌘
Gi (t, &, z, !) Gi (t, &¯, z, !) = f xi , t, Li,t rt , i+1 1
,
2h h2
⇣ t
zi+1 zit t
zi+1 &it + zit 1 ⌘
2¯
1
f xi , t, Li,t rt , ,
2h h2

@f ⇣ zt 2&it + zit 1 zi+1

t
&it + zit 1 ⌘
2¯
= (Q) i+1
@q h2 h2
2(&it &¯it ) @f 2(&i &¯i ) @f
= 2
(Q) = (Q),
h @q h2 @q
where t 2 [0, T ], &, &¯, z 2 Rn , ! 2 C([ ⌧0 , 0], Rn ) and Q 2 [ L, L]⇥[0, T ]⇥C(D, R)⇥
R2 . From (5.1), we get
@f
(Q) 0
@q

and conclude further that

 26

2" @f
&i &¯i " Gi (t, &, z, !) Gi (t, &¯, z, !) = 1+ (Q) &i &¯i
h2 @q
2" @f
= 1+ 2 (Q) &i &¯i ,
h @q

where " 0. Upon taking the infinity norm on both sides of the above equation, we
deduce that

2" @f
& &¯ " G(t, &, z, !) G(t, &¯, z, !) = 1+ (Q) & &¯ n .
n h2 @q

From this and (5.2), we obtain

2"⌫(t)
& &¯ " G(t, &, z, !) G(t, &¯, z, !) 1+ & &¯ n
n h2

= 1 "µ1 (t) & &¯ n ,

with
2⌫(t)
µ1 (t) = ,
h2

which shows that (3.8) holds.

In order to prove (3.9), we apply (3.6) and obtain

⇣ zt zit t
zi+1 2&i + zit 1 ⌘
Gi (t, &, z, !) Gi (t, &, z̄, !) = f xi , t, Li,t !, i+1 1
,
2h h2
⇣ z̄ t z̄it t
z̄i+1 2&i + z̄it 1 ⌘
f xi , t, Li,t !, i+1 1
,
2h h2
 27

t
zi+1 zit 1
t
z̄i+1 z̄it 1
t
zi+1 2&i + zit 1
t
z̄i+1 2&i + z̄it 1
 1 (t) + 2 (t)
2h 2h h2 h2
1 (t) ⇣ t t
⌘  (t) ⇣
2
⌘
 |zi+1 z̄i+1 | + |zit 1 z̄it 1 | + 2 |zi+1 t t
z̄i+1 | + |zit 1 z̄it 1 |
2h h
⇣  (t)  (t) ⌘⇣ ⌘
1 2 t t
= + 2 |zi+1 z̄i+1 | + |zit 1 z̄it 1 |
2h h
where t 2 [0, T ], &, z, z̄ 2 Rn , ! 2 C([ ⌧0 , 0], Rn ). Since

1 (t) 2 (t)
+ 2 0,
2h h

upon taking the infinity norm (with i = 1, . . . , n) on both sides of the above inequality,
we get
!
1 (t) 2 (t)
kG(t, &, z, !) G(t, &, z̄, !)kn  2 + 2 kz z̄kn ,
2h h

which shows that (3.9) holds with

!
1 (t) 2 (t)
µ2 (t) = 2 + 2 .
2h h

In order to prove (3.10), we apply (3.6) with 1 (t) = 2 (t) = 0 and obtain

⇣ zt zit t
zi+1 2&i + zit 1 ⌘
Gi (t, &, z, !) Gi (t, &, z, !
¯) = f xi , t, Li,t !, i+1 1
2h h2
⇣ t
zi+1 zit t
zi+1 2&i + zit 1 ⌘
1
f xi , t, Li,t !
¯,
2h h2

 3 (t) max Li,t !(s, ⌧ ) Li,t !

¯ (s, ⌧ ) ,
(s,⌧ )2D

for t 2 [0, T ], &, z 2 Rn , !, !

¯ 2 C([ ⌧0 , 0], Rn ). Then, from the definition of the
 28

operator Li,t we get

Li,t !(s, ⌧ ) Li,t !

¯ (s, ⌧ ) = Li,t (! !
¯ )(s, ⌧ )

xk+1 s⇣ t t
⌘ s xk ⇣ t t
⌘
= !k+i (⌧ ) !
¯ k+i (⌧ ) + !k+i+1 (⌧ ) !
¯ k+i+1 (⌧ ) ,
h h

where xk  s  xk+1 , and

xk+1 s t t
Li,t !(s, ⌧ ) Li,t !
¯ (s, ⌧ )  !k+i (⌧ ) !¯ k+i (⌧ )
h
s xk t t
+ !k+i+1 (⌧ ) ! ¯ k+i+1 (⌧ )
h
xk+1 s
 max k(! ! ¯ )(⌧ )k
h ⌧ 2[ ⌧0 ,0]
s xk
+ max k(! ! ¯ )(⌧ )k
h ⌧ 2[ ⌧0 ,0]
xk+1 s+s xk
= k! ¯ k0n = k!
! ¯ k0n
!
h

Upon taking the maximum over D on both sides of the above relations, we get

max Li,t !(s, ⌧ ) Li,t !

¯ (s, ⌧ )  k! ¯ k0n .
!
(s,⌧ )2D

Therefore,

G(t, &, z, !) G(t, &, z, !

¯) n
 3 (t)k! ¯ k0n ,
!

which shows that (3.10) holds with

µ3 (t) = 3 (t),

and finishes the proof of the theorem.

Having proven that G satisfies the conditions (3.7)–(3.10) under the assumptions
 29

of Theorem 5.1, we are now in a position to investigate the numerical schemes (2.1)
and (3.1) designed to find numerical solutions to general partial functional di↵erential
equations and to derive appropriate bounds on their errors, from which we can deduce
important convergence properties of the numerical algorithm.
 30

CHAPTER 6

THEOREMS ON ERROR BOUNDS FOR NUMERICAL

SOLUTIONS TO GENERAL PARTIAL FUNCTIONAL

DIFFERENTIAL EQUATIONS

In this chapter, we prove a sequence of results that can be used to deduce appropriate
bounds on the errors that we can expect to get numerically by applying the numerical
schemes (2.1) and (3.1) to solve a class of general partial functional di↵erential
equations.

The first of these results is a main theorem specifying a bound on the norm of
the error in terms of an integral that holds as long as the right-hand-side function f
satisfies an appropriate condition. In particular, we require a sharper condition than
(5.1).

Theorem 6.1 ([37], Theorem 4.1). Suppose the given function f satisfies conditions
(3.6) and (5.1), the step-size h > 0 is chosen in such a way that

@f h @f
(x, t, !, p, q) (x, t, !, p, q) 0, (6.1)
@q 2 @p

@f @f
for all x 2 [ L, L], t 2 [0, T ], ! 2 C(D, R), p, q 2 R, and(x, t, !, ·), (x, t, !, ·)
@p @q
are continuous functions for each x 2 [ L, L], t 2 [0, T ], and ! 2 C([0, T ], R).
Moreover, suppose that F satisfies (2.2) and u 2 C (4) (D, R). Then the errors e(t)
 31

satisfy !
Z t Z t
ke(t)kn  n (s)exp 3 (⌧ )d⌧ ds,
0 s

for t 2 [0, T ], where k · kn is the infinity norm in Rn .

Proof. We apply Theorem 6.7 from the Appendix with ⇢(xi , t) = ei (t) = Ui (t) i (t),

t 2 [ ⌧0 , T ] and i = 0, ±1, ..., ±M̃ , where vi (t) = g(xi , t), for i = ±M, . . . , ±M̃ .
@⇢
Then, (xi , t) = ėi (t) = U̇i (t) ˙ i (t). We want to show that
@t

Z Z !
t t
max |ei (t)| = max |⇢(xi , t)|  n (s)exp 3 (⌧ )d⌧ ds.
i=0,±1,...,±M 0 i=0,±1,...,±M 0 0 s

Since
2
˙ i (t) = Fi (t, (t), t ) = f xi , t, Li,t t , i,t (t), i,t (t) ,

we get

@⇢
(xi , t) = U̇i (t) Fi (t, U (t), Ut )
@t
2 2
+ f xi , t, Li,t Ut , i,t U (t), i,t U (t) f xi , t, Li,t t , i,t U (t), i,t U (t)

2 2
+ f xi , t, Li,t t , i,t U (t), i,t U (t) f xi , t, Li,t t , i,t (t), i,t (t) .

From this and from the Mean Value Theorem, we get

Z 1 Z 1
@⇢ ˜⇢(xi , t) @f ˜(2) ⇢(xi , t) @f
(xi , t) (Qs )ds (Qs )ds
@t 0 @p 0 @q

= U̇i (t) Fi (t, U (t), Ut )

2 2
+f xi , t, Li,t Ut , i,t U (t), i,t U (t) f xi , t, Li,t t , i,t U (t), i,t U (t)

and
 32

Z 1 Z 1
@⇢ ˜⇢(xi , t) @f ˜(2) ⇢(xi , t) @f
(xi , t) (Qs )ds (Qs )ds
@t 0 @p 0 @q

 U̇i (t) Fi (t, U (t), Ut )

2 2
+ f xi , t, Li,t Ut , i,t U (t), i,t U (t) f xi , t, Li,t t , i,t U (t), i,t U (t) ,

where ˜⇢(xi , t), ˜(2) ⇢(xi , t) are defined by (6.12) in the Appendix and

Qs = (xi , t, Li,t vt , i,t v(t) + s ˜⇢(xi , t), 2

i,t v(t) + s ˜(2) ⇢(xi , t))

is a point from the domain of the function f . We now apply conditions (3.5) and
(3.6) to obtain

Z 1 Z 1
@⇢ ˜⇢(xi , t) @f ˜(2) @f
(xi , t) (Qs )ds ⇢(xi , t) (Qs )ds
@t 0 @p 0 @q

 n (t) + 3 (t) max Li,t Ut (s, ⌧ ) Li,t t (s, ⌧ ) .

(s,⌧ )2D

Since

max |Li,t Ut (s, ⌧ ) Li,t t (s, ⌧ )|

(s,⌧ )2D

xk+1 s⇣ ⌘ s xk ⇣ ⌘
= max (Ut )k+i (⌧ ) (vt )k+i (⌧ ) + (Ut )k+1+i (⌧ ) (vt )k+1+i (⌧ )
(s,⌧ )2D h h

xk+1 s s xk
= max (et )k+i (⌧ ) + (et )k+1+i (⌧ )
(s,⌧ )2D h h

xk+1 s s xk
= max ek+i (t + ⌧ ) + ek+1+i (t + ⌧ )
(s,⌧ )2D h h

xk+1 s s xk
= max ⇢(xk+i , t + ⌧ ) + ⇢(xk+1+i , t + ⌧ )
(s,⌧ )2D h h
 33

xk+1 s s xk
= max ⇢(xi + xk , t + ⌧ ) + ⇢(xi + xk+1 , t + ⌧ )
(s,⌧ )2D h h
!
xk+1 s s xk
 max + · max |⇢(xi + xj , t + ⌧ )|
(s,⌧ )2D h h j=k,k+1

= max · max |⇢(xi + xj , t + ⌧ )|

(s,⌧ )2D j=k,k+1

= max{|⇢(xi + xj , t + ⌧ )| : j = 0, ±1, . . . , ±M̂ , ⌧ 2 [ ⌧0 , 0]}

= k⇢(xi ,t) khn ,

where k is such that xk  s  xk+1 , we have

Z 1 Z 1
@⇢ ˜⇢(xi , t) @f ˜(2) ⇢(xi , t) @f
(xi , t) (Qs )ds (Qs )ds
@t 0 @p 0 @q
 n (t) + 3 (t)k⇢(xi ,t) khn .

We now apply Theorem 6.7 from the Appendix with

Z1 Z1
@f @f
G(xi , t, ⇢(xi ,t) ) = (Qs )ds, H(xi , t, ⇢(xi ,t) ) = (Qs )ds,
@p @q
0 0

and from the above inequality combined with (6.1), we get

ke(t)kn = max |ei (t)| = max |⇢(xi , t)|  ⌘(t), (6.2)

i=0,±1,...,±M 0 i=0,±1,...,±M 0

for t 2 [0, T ], where ⌘ is the solution to the initial value problem

8
>
< ⌘(t)
˙ = 3 (t)⌘(t) + n (t),
(6.3)
>
: ⌘(0) = 0.
 34

From (6.2) and (6.3), we have

Z Z !
t t
ke(t)kn  exp 3 (r)dr · n (s)ds
0 s

and the proof is finished.

The next theorem will be applied to derive an error bound for iterative processes
applied to problem (1.1)–(1.3).

Theorem 6.2 ([37], Lemma 5.1). Suppose F and G satisfy conditions (3.5) and
(3.7)–(3.10). Then,

Z Z !
t t ⇣ ⌘
kE (k+1) (t)kn  exp µ1 (s)ds µ2 (⌧ ) + µ3 (⌧ ) kE⌧(k) k0n + n (⌧ ) d⌧, (6.4)
0 ⌧

for all t 2 [0, T ] and k = 0, 1, 2, . . . , where k · kn is the infinity norm in Rn .

Proof. From the definition of the error E (k+1) (t), we get

Ė (k+1) (t) = U̇ (t) ˙ (k+1) (t)

(k+1) (k) (k)
= U̇ (t) G(t, (t), (t), t )

(k+1)
= U̇ (t) F (t, U (t), Ut ) + F (t, U (t), Ut ) G(t, (t), U (t), Ut )
(k+1) (k+1) (k)
+ G(t, (t), U (t), Ut ) G(t, (t), (t), Ut )
(k+1) (k) (k+1) (k) (k)
+ G(t, (t), (t), Ut ) G(t, (t), (t), t )

We now multiply both sides of the above relation by an arbitrary " < 0 and evaluate
the following infinity norm
 35

E (k+1) (t) + "Ė k+1 (t) n

⇣ ⌘
(k+1) (k+1)
= U (t) (t) ( ") G(t, U (t), U (t), Ut ) G(t, (t), U (t), Ut )
⇣
( ") U̇ (t) F (t, U (t), Ut ) + G(t, (k+1) (t), U (t), Ut ) G(t, (k+1)
(t), (k)
(t), Ut )
⌘
(k+1) (k) (k+1) (k) (k)
+G(t, (t), (t), Ut ) G(t, (t), (t), t )
n

(k+1)
U (t) (t) ( ")[G(t, U (t), U (t), Ut ) G(t, (k+1) (t), U (t), Ut )]
⇣ n

( ") U̇ (t) F (t, U (t), Ut ) + G(t, (k+1) (t), U (t), Ut ) G(t, (k+1) (t), (k)
(t), Ut )
⌘
(k)
+G(t, (k+1) (t), (k) (t), Ut ) G(t, (k+1) (t), (k) (t), t )
n

From this and (3.8), we get

E (k+1) (t) + "Ė (k+1) (t) n 1 ( ")µ1 (t) U (t) (k+1)

(t) n
⇣
( ") U̇ (t) F (t, U (t), Ut ) + G(t, (k+1) (t), U (t), Ut ) G(t, (k+1) (t), (k)
(t), Ut )
⌘
(k+1) (k) (k+1) (k) (k)
+G(t, (t), (t), Ut ) G(t, (t), (t), t ) .
n

Since " > 0, we use the scaling property of the norm and further conclude that the
following inequality holds:

E (k+1) (t) + "Ė (k+1) (t) n

E (k+1) (t) n
"µ1 (t) E (k+1) (t) n
(k+1) (k+1) (k)
+" U̇ (t) F (t, U (t), Ut ) + G(t, (t), U (t), Ut ) G(t, (t), (t), Ut )
(k+1) (k) (k+1) (k) (k)
+G(t, (t), (t), Ut ) G(t, (t), (t), t ) .
n

Upon dividing the above inequality by " < 0, we obtain

1 ⇣ (k+1) ⌘
E (t) + "Ė (k+1) (t) n E (k+1) (t) n  µ1 (t) E (k+1) (t) n
"
+ U̇ (t) F (t, U (t), Ut ) + G(t, (k+1) (t), U (t), Ut ) G(t, (k+1) (t), (k)
(t), Ut )
(k+1) (k) (k+1) (k) (k)
+G(t, (t), (t), Ut ) G(t, (t), (t), t ) .
n
 36

Notice that the right-hand side of the above inequality does not depend on " and by
taking " ! 0 on both sides, we deduce that

1 ⇣ (k+1) ⌘
D kE (k+1) (t)kn = lim E (t) + "Ė (k+1) (t) n
E (k+1) (t) n
"!0 "
 µ1 (t) E (k+1) (t) n
(k+1) (k+1) (k)
+ U̇ (t) F (t, U (t), Ut ) n
+ G(t, (t), U (t), Ut ) G(t, (t), (t), Ut ) n
(k+1) (k) (k+1) (k) (k)
+ G(t, (t), (t), Ut ) G(t, (t), (t), t ) n,

where D is the left-hand side derivative with respect to t. We now apply conditions
(3.5), (3.9), (3.10) to get that

D kE (k+1) (t)kn
(k) 0
 µ1 (t) E (k+1) (t) n
+ n (t) + µ2 (t) U (t) (k)
(t) n
+ µ3 (t) Ut t n
(k) 0
= µ1 (t) E (k+1) (t) n
+ n (t) + µ2 (t) E (k) (t) n
+ µ3 (t) Et n
(k)
 µ1 (t) E (k+1) (t) n
+ µ2 (t) + µ3 (t) kEt k0n + n (t).

Also, note that kE (k+1) (0) n

= 0 as the successive iterates satisfy the same initial
condition as U (0). Therefore, we conclude that

kE (k+1) (t) n
 (t),

where (t) solves the following problem

8
>
< 0
(t) = µ1 (t) (t) + ⇠(t),
>
: (0) = 0,

and
 37

(k)
⇠(t) = µ2 (t) + µ3 (t) kEt k0n + n (t).

Notice that
Z t ⇣Z t ⌘
(t) = ⇠(s) exp µ1 (⌧ )d⌧ ds,
0 s

for t 2 [0, T ]. Therefore,

Z t ⇣Z t ⌘
(k+1)
kE (t) n
 ⇠(s) exp µ1 (⌧ )d⌧ ds,
0 s

which implies (6.4) and finishes the proof.

In what follows, we will prove a sequence of preliminary results that will be useful
in proving Theorem 6.5.

Henceforth, we will use the following notation. Let t 2 [0, T ]. Then, the maximum
starting error will be denoted by

E(t) = max kE (0) (⌧ )kn .

⌧ 2[0,t]

We assume that the function µ1 has no roots in [0, T ], that is, either sign(µ1 ) = 1 or
sign(µ1 ) = 1, and we define

µ2 (⌧ ) + µ3 (⌧ )
r(t) = sign(µ1 ) max ,
⌧ 2[0,t] |µ1 (⌧ )|
n (⌧ )
n (t) = sign(µ1 ) max .
⌧ 2[0,t] |µ1 (⌧ )|

For the sake of simplicity of the subsequent proofs, we also introduce the following
definitions of four functions:
 38

Z t
A(t) = µ1 (⌧ )d⌧,
0
k 1 j
X A(t)
A(t)
↵k (t) = 1 e ,
j=0
j!
k 1
Sk (t) = r(t) ↵k (t),
k
X
k (t) = n (t) Si (t).
i=1

Theorem 6.3 ([37], Lemma 5.2). If µ1 has no roots in [0, T ], then all functions

k
sign(µ1 ) ↵k (t),

where k = 1, 2, . . . and t 2 [0, T ], are nondecreasing and nonnegative.

Proof. Throughout the proof, we use the notation

k
↵
˜ k (t) = ( sign µ1 ) ↵k (t)

˜ k0 (t)
and firstly show that ↵ 0, for all t 2 [0, T ]. From the definition of the function
A, we get

k 1 j 1 0 k 1 j
X ( 1)j A(t) A (t) X A(t)
↵k0 (t) = exp A(t) 0
exp A(t) A (t)
j=1
(j 1)! j=0
j!
k 2 j k 1 j
!
X ( 1)j+1 A(t) X ( 1)j A(t)
0
= exp(A(t))A (t) +
j=0
j! j=0
j!
k 1
0 ( 1)k A(t)
= exp A(t) A (t) .
(k 1)!

We now consider two opposite cases, sign(µ1 ) = 1 and sign(µ1 ) = 1. Suppose

firstly that sign(µ1 ) = 1. Then, from the definition, A(t) 0 and A0 (t) = µ1 (t) > 0.
 39

Since

k 1
k ( 1)k A(t)
˜ k0 (t)
↵ = sign(µ1 ) ↵k0 (t) k
= ( 1) exp A(t) A (t) 0
(k 1)!
k 1
A(t)
= exp A(t) µ1 (t) 0,
(k 1)!

↵
˜ k (t) is shown to be nondecreasing in the first case.

We now suppose that sign(µ1 ) = 1. Then, A(t)  0 and we get

k 1
k ( 1)k A(t)
˜ k0 (t)
↵ = sign(µ1 ) ↵k0 (t) = ↵k0 (t) = exp A(t) µ1 (t)
(k 1)!
k 1
A(t)
= exp A(t) µ1 (t) 0
(k 1)!

showing that ↵
˜ k (t) is nondecreasing also in the second case.

Since A(0) = 0, ↵
˜ k (0) = ↵k (0) = 0 and since ↵
˜ k (t) is nondecreasing on [0, T ], we
conclude that ↵
˜ k (t) 0, for t 2 [0, T ], which finishes the proof.

We now apply Theorem 6.3 to prove the next theorem on the nonnegativity and
monotonicity of r(t)Sk (t) and n (t)Sk (t) for k = 1, 2, . . . .

Theorem 6.4 ([37], Corollary 5.1). If µ1 has no roots in [0, T ], then the functions
r(t)Sk (t) and n (t)Sk (t), with k = 1, 2, . . . , are nondecreasing and nonnegative for all
t 2 [0, T ].

Proof. Let r̃(t) = r(t)Sk (t). Then, from the definitions of r(t) and Sk (t), we get
 40

⇣ ⌘
k 1
r̃(t) = r(t) · r(t) ↵k (t)

k
= r(t) ↵k (t)

!k
µ2 (⌧ ) + µ3 (⌧ )
= sign(µ1 ) max · ↵k (t)
⌧ 2[0,t] |µ1 (⌧ )|
!k
k µ2 (⌧ ) + µ3 (⌧ )
= sign(µ1 ) ↵k (t) · max .
⌧ 2[0,t] |µ1 (⌧ )|

Notice that the function !k

µ2 (⌧ ) + µ3 (⌧ )
max
⌧ 2[0,t] |µ1 (⌧ )|

is nondecreasing and nonnegative for t 2 [0, T ]. Moreover, by Theorem 6.3, the

function ( sign(µ1 ))k ↵k (t) is also nondecreasing and nonnegative for t 2 [0, T ].
Therefore, we conclude the same about r̃(t).

We now define ˜ (t) = n (t)Sk (t). From the definitions of the functions n (t) and
Sk (t) we get

!
⇣ ⌘
˜ (t) = n (⌧ ) k 1
sign(µ1 ) max · r(t) ↵k (t)
⌧ 2[0,t] |µ1 (⌧ )|
!
k 1 n (⌧ )
= sign(µ1 ) r(t) ↵k (t) max
⌧ 2[0,t] |µ1 (⌧ )|
!k 1 !
µ2 (⌧ ) + µ3 (⌧ ) n (⌧ )
= sign(µ1 ) sign(µ1 ) max ↵k (t) max
[0,t] |µ1 (⌧ )| ⌧ 2[0,t] |µ1 (⌧ )|

!k 1 !
k µ 2 (⌧ ) + µ 3 (⌧ ) n (⌧ )
= sign(µ1 ) ↵k (t) · max · max .
[0,t] |µ1 (⌧ )| ⌧ 2[0,t] |µ1 (⌧ )|

(6.5)

It can be observed that

 41

!k 1
µ2 (⌧ ) + µ3 (⌧ )
max
[0,t] |µ1 (⌧ )|

and
n (⌧ )
max
⌧ 2[0,t] |µ1 (⌧ )|

are nondecreasing and nonnegative as functions of t 2 [0, T ]. Moreover, by Theorem

k
6.3, we conclude that sign(µ1 ) ↵k (t) is nondecreasing and nonnegative on [0, T ].
Therefore, we further conclude from (6.5), that ˜ (t) has the same properties, which
finishes the proof.

The last result that is necessary in order to prove Theorem 6.5 can be summarised
by the following lemma.

Lemma 6.1 ([37], Lemma 5.3). The relation

Z t
exp A(⌧ ) ↵k (⌧ )µ1 (⌧ )d⌧ = exp A(t) ↵k+1 (t),
0

is satisfied for all t 2 [0, T ] and k = 1, 2, . . . .

Proof. Since A0 (t) = µ1 (t), we get

Z t Z t
exp A(⌧ ) µ1 (⌧ )↵k (⌧ )d⌧ = exp A(⌧ ) A0 (⌧ )↵k (⌧ )d⌧
0 0
Z t
d⇣ ⌘
= exp A(⌧ ) · ↵k (⌧ )d⌧
0 d⌧

From ↵k (0) = 0, we further conclude that

Z t
exp A(⌧ ) µ1 (⌧ )↵k (⌧ )d⌧
0
 42

" #⌧ =t Z t
= exp A(⌧ ) ↵k (⌧ ) + exp A(⌧ ) ↵k0 (⌧ )d⌧
0
⌧ =0
Z t
= exp A(t) ↵k (t) + exp A(0) ↵k (0) + exp A(⌧ ) ↵k0 (⌧ )d⌧
0
Z t
= exp A(t) ↵k (t) + exp A(⌧ ) ↵k0 (⌧ )d⌧.
0

Since
k 1
A(⌧ )
↵k0 (⌧ ) = exp A(⌧ ) A (⌧ )( 1) 0
, k
(k 1)!

we get
Z t
exp A(⌧ ) µ1 (⌧ )↵k (⌧ )d⌧
0

Z t k 1
0 k A(⌧ )
= exp A(t) ↵k (t) + A (⌧ )( 1) d⌧
0 (k 1)!
Z t
( 1)k k 1
= exp A(t) ↵k (t) + A0 (⌧ ) A(⌧ ) d⌧
(k 1)! 0
" #
k ⌧ =t
( 1)k A(⌧ )
= exp A(t) ↵k (t) +
(k 1)! k
⌧ =0

From A(0) = 0 and the definition of ↵k (t), we further conclude that

Z t k
A(t)
exp A(⌧ ) µ1 (⌧ )↵k (⌧ )d⌧ = exp( A(t))↵k (t) +
0 k!
k
!
exp A(t) A(t)
= exp A(t) ↵k (t)
k!

= exp A(t) ↵k+1 (t),

which finishes the proof.

 43

We now apply Lemma 6.1 and the previous two Theorems 6.3 and 6.4 to prove
the following theorem that supplies an explicit error bound for the successive iterates.

Theorem 6.5 ([37], Theorem 5.1). Suppose that the function F satisfies condition
(3.5) and G satisfies conditions (3.7)–(3.10). Moreover, suppose that the function µ1
has no roots in [0, T ]. Then,

E (k) (t) n
 r(t)E(t)Sk (t) + k (t), (6.6)

for t 2 [0, T ] and k = 1, 2, . . . .

Proof. We first show (6.6) for k = 1. By Theorem 6.2, we get

Z Z !
t t ⇣ ⌘
(1)
kE (t)kn  exp µ1 (s)ds µ2 (⌧ ) + µ3 (⌧ ) kE⌧(0) k0n + n (⌧ ) d⌧
0 ⌧
Z Z Z !
t t ⌧ ⇣ ⌘
 exp µ1 (s)ds µ1 (s)ds µ2 (⌧ ) + µ3 (⌧ ) E(⌧ ) + n (⌧ ) d⌧.
0 0 0

Then, using the definition of the function A(t), we further get

Z t ⇣ ⌘
(1)
kE (t)kn  exp A(t) A(⌧ ) µ2 (⌧ ) + µ3 (⌧ ) E(⌧ ) + n (⌧ ) d⌧
0
Z !
t
µ2 (⌧ ) + µ3 (⌧ ) n (⌧ )
= exp A(t) exp A(⌧ ) µ1 (⌧ ) E(⌧ ) + d⌧.
0 µ1 (⌧ ) µ1 (⌧ )

We now reduce the above integrand by considering its maximum and deduce that

Z !
t
(1) µ2 (s) + µ3 (s)
kE (t)kn  exp A(t) exp A(⌧ ) µ1 (⌧ ) max E(⌧ )d⌧
0 s2[0,t] µ1 (s)
Z !
t
n (s)
+ exp A(t) exp A(⌧ ) µ1 (⌧ ) max d⌧.
0 s2[0,t] µ1 (s)
 44

Therefore, since both maxima do not depend on ⌧ and the function E(t) is nonde-
creasing, we deduce that
! Z t
µ2 (s) + µ3 (s)
kE (1) (t)kn  exp A(t) max E(t) exp A(⌧ ) µ1 (⌧ ) d⌧
s2[0,t] µ1 (s) 0
!Z
t
n (s)
+ exp A(t) max exp A(⌧ ) µ1 (⌧ ) d⌧,
s2[0,t] µ1 (s) 0

which further implies that

! !!
µ2 (s) + µ3 (s) n (s)
kE (1) (t)kn  exp A(t) sign(µ1 ) max E(t) + max ·
s2[0,t] µ1 (s) s2[0,t] µ1 (s)
Z t
· exp A(⌧ ) µ1 (⌧ )d⌧.
0

We now use the definitions of the functions r(t), n to deduce that

⇣ ⌘Z t
(1)
kE (t)kn  exp A(t) r(t)E(t) + n (t) exp A(⌧ ) µ1 (⌧ )d⌧.
0

Therefore, since

Z Z " #⌧ =t
t t
exp A(⌧ ) µ1 (⌧ )d⌧ = exp A(⌧ ) A0 (⌧ )d⌧ = exp A(⌧ )
0 0
⌧ =0

=1 exp A(t)

we find that

⇣ ⌘⇣ ⌘
(1)
kE (t)kn  exp A(t) r(t)E(t) + n (t) 1 exp A(t)
(6.7)
⇣ ⌘⇣ ⌘
= r(t)E(t) + n (t) exp A(t) 1 .
 45

On the other hand, notice that, for k = 1, the right-hand side of inequality (6.6) is
written in the form

⇣ ⌘
r(t)E(t)S1 (t) + 1 (t) = r(t)E(t)↵1 (t) + n (t)S1 (t) = ↵1 (t) r(t)E(t) + n (t) .

Moreover, from the definition of the function ↵1 (t), we get

↵1 (t) = 1 exp A(t) .

Therefore, from (6.7), we have

kE (1) (t)kn  r(t)E(t)S1 (t) + 1 (t),

which shows that (6.6) is satisfied for k = 1.

We now suppose that (6.6) is satisfied for a certain k > 1. From the definition of
the maximum norm k · k0n , we have

kE⌧(k) k0n = max kE⌧(k) (s)kn = max kE (k) (⌧ + s)kn ,

s2[ ⌧0 ,0] s2[ ⌧0 ,0]

for ⌧ 2 [0, T ]. Therefore, from (6.6), we find that

⇣ ⌘
kE⌧(k) k0n  max r(⌧ + s)E(⌧ + s)Sk (⌧ + s) + k (⌧ + s) .
s2[ ⌧0 ,0]

Since, by Theorem 6.4, all functions n (t)Si (t), where i = 1, 2, . . . , are nondecreasing,
from the definition of the function k, we conclude that k is also nondecreasing.
Moreover, the function E(t) is nondecreasing and, by Theorem 6.4, the functions
r(t)Sk (t) have the same feature. Therefore, the function that is being maximized on
 46

the right-hand side of the above inequality is also nondecreasing, which implies that

kE⌧(k) k0n  r(⌧ )E(⌧ )Sk (⌧ ) + k (⌧ ).

Therefore, by Theorem 6.2, we find that

Z Z !
t t ⇣
kE (k+1) (t)kn  exp µ1 (s)ds µ2 (⌧ ) + µ3 (⌧ ) r(⌧ )E(⌧ )Sk (⌧ )
0 ⌧
!
⌘
+ k (⌧ ) + n (⌧ ) d⌧.

From this and the definition of the function A(t), we further deduce that

Z
(k+1) A(t)
t
A(⌧ ) µ2 (⌧ ) + µ3 (⌧ ) ⇣ ⌘
kE (t)kn  e e r(⌧ )E(⌧ )Sk (⌧ ) + k (⌧ )
0 µ1 (⌧ )
!
n (⌧ )
+ µ1 (⌧ ) d⌧.
µ1 (⌧ )

We now consider maxima of the two quotients in the above integrand and obtain that
!Z
µ (s) + µ (s) t ⇣ ⌘
(k+1) A(t) 2 3 A(⌧ )
kE (t)kn  e max e r(⌧ )E(⌧ )Sk (⌧ ) + k (⌧ )
s2[0,t] µ1 (s) 0
!Z
t
n (s)
µ1 (⌧ ) d⌧ + eA(t) max e A(⌧ ) µ1 (⌧ ) d⌧.
s2[0,t] µ1 (s) 0

From this and the definitions of r(t) and n (t), we deduce that

Z t ⇣ ⌘
(k+1) A(t) A(⌧ )
kE (t)kn  e r(t) e r(⌧ )E(⌧ )Sk (⌧ ) + k (⌧ ) µ1 (⌧ )d⌧
0
Z (6.8)
t
+ eA(t) n (t) e A(⌧ )
µ1 (⌧ )d⌧.
0
 47

We now consider the first term, T1 , on the right-hand side of (6.8) and, from the
definitions of Si (t) and k (t), we obtain the following expression for T1

Z !k
t
µ2 (s) + µ3 (s)
T1 = eA(t) (sign µ1 )r(t) e A(⌧ ) ( sign µ1 )k ↵k (⌧ ) max
0 s2[0,⌧ ] |µ1 (s)|

Z k
! !i 1
t X n (s) µ2 (s) + µ3 (s)
A(⌧ )
E(⌧ )|µ1 (⌧ )|d⌧ + e max max
0 i=1
s2[0,⌧ ] |µ1 (s)| s2[0,⌧ ] |µ1 (s)|
!
( sign µ1 )i ↵i (⌧ )|µ1 (⌧ )|d⌧ .

By extending the above maxima from [0, ⌧ ] to [0, t] and interchanging the order of
summation and integration, we deduce that

!k Z t
A(t) µ2 (s) + µ3 (s) A(⌧ )
T1  e (sign µ1 )r(t) max E(t) e ( sign µ1 )k ↵k (⌧ )
s2[0,t] |µ1 (s)| 0

! k
!i 1 Z
n (s)
X µ2 (s) + µ3 (s) t
A(⌧ )
|µ1 (⌧ )|d⌧ + max max e ( sign µ1 )i
s2[0,t] |µ1 (s)| i=1
s2[0,t] |µ1 (s)| 0

!
↵i (⌧ )|µ1 (⌧ )|d⌧ .

We now apply Lemma 6.1 and obtain

!k
µ 2 (s) + µ 3 (s)
T1  eA(t) r(t) max E(t)( sign µ1 )k ( 1)e A(t) ↵k+1 (t)
s2[0,t] |µ1 (s)|
! k !i 1 !
n (s) X µ 2 (s) + µ 3 (s)
+ max max ( sign µ1 )i ( 1)e A(t) ↵i+1 (t) .
s2[0,t] |µ1 (s)| i=1 s2[0,t] |µ1 (s)|

From the definitions of r(t), n (t), and Si (t), we obtain

 48

k
X
k+1 i
T1  r(t) E(t)↵k+1 (t) + n (t) ( 1) r(t) ↵i+1 (t)
i=1
k
(6.9)
X
= r(t)Sk+1 (t)E(t) + n (t) Si+1 (t).
i=1

We now consider the second term, T2 , on the right-hand side of (6.8) and obtain
the following expression for it:

Z t h i⌧ =t
A(t) A(⌧ ) 0 A(t) A(⌧ )
T2 = e n (t) e A (⌧ )d⌧ = e n (t) e = n (t) eA(t) 1
0 ⌧ =0

= n (t)↵1 (t) = n (t)S1 (t).

(6.10)

From (6.8), (6.9), and (6.10), we deduce that

k+1
X
(k+1)
kE (t)kn  T1 + T2  r(t)Sk+1 (t)E(t) + n (t) Si (t) + n (t)S1 (t)
i=2
k+1
X
= r(t)Sk+1 (t)E(t) + n (t) Si (t) = r(t)Sk+1 (t)E(t) + k+1 (t),
i=1

which shows that (6.6) holds and finishes the proof.

We now compare error bounds by means of the following theorem.

Theorem 6.6 ([37], Lemma 6.1). Suppose

1 (t) = 0, 3 (t) = %2 (t), n (t) = c0 h2 (1 + %)2 (t),

µ1 (t) = h 2 2 (t), µ2 (t) = h 2 2 (t), µ3 (t) = %2 (t)

where %, , c0 > 0. Then,

Z t ⇣Z t ⌘
k (t) < n (s) exp 3 (⌧ )d⌧ ds,
0 s
 49

for all t 2 [0, T ], k = 1, 2, . . . .

Proof. Let !
Z t Z t
⌘(t) = n (s) exp 3 (⌧ )d⌧ ds,
0 s

for t 2 [0, T ]. Then, from the definition of n (t) and 3 (t), we deduce that

Z Z Z !
t t s
⌘(t) = c0 h2 (1 + %) 2 (s) exp %2 (⌧ )d⌧ %2 (⌧ )d⌧ ds
0 0 0
Z !Z Z !
t t s
= c0 h2 (1 + %) exp %2 (⌧ )d⌧ 2 (s) exp %2 (⌧ )d⌧ ds.
0 0 0

Since

Z Z ! " Z !#s=t
t s s
1
2 (s) exp %2 (⌧ )d⌧ ds = exp %2 (⌧ )d⌧
0 0 % 0
Z !!
s=0
t
1
= 1 exp %2 (⌧ )d⌧ ,
% 0

we further obtain

Z ! Z !!
t t
1
⌘(t) = c0 h2 (1 + %) exp %2 (⌧ )d⌧ 1 exp %2 (⌧ )d⌧
0 % 0
Z t ! !
1+%
= c0 h 2 exp %2 (⌧ )d⌧ 1
% 0

and !
Z t
d⌘
(t) = c0 h2 (1 + %)2 (t) exp %2 (⌧ )d⌧ .
dt 0

Let
µ2 (⌧ ) + µ3 (⌧ ) n (⌧ )
µ(t) = max , (t) = max .
⌧ 2[0,t] |µ1 (⌧ )| ⌧ 2[0,t] |µ1 (⌧ )|

Then, from the definition of k (t), n (t), Si (t), and r(t), we find that
 50

k
X k
X i 1
k (t) = n (t) Si (t) = sign(µ1 ) (t) r(t) ↵i (t)
i=1 i=1
k
X i 1
= sign(µ1 ) (t) sign(µ1 )µ(t) ↵i (t)
i=1
k
X i 1 i
= (t) µ(t) sign(µ1 ) ↵i (t).
i=1

Since

n (⌧ ) µ2 (⌧ ) + µ3 (⌧ )
max = c0 h4 (1 + %) 1
, max = 1 + %h2 1
,
⌧ 2[0,t] |µ1 (⌧ )| ⌧ 2[0,t] |µ1 (⌧ )|

the function k (t) can be written in the form

k
X
1 i 1 i
k (t) = c0 h4 (1 + %) 1
1 + %h2 sign(µ1 ) ↵i (t).
i=1

Therefore, from the relation

Z
|µ1 (t)| A(t) ⇣ t
i d↵i
⌘i 1
sign(µ1 ) (t) = e |µ1 (⌧ )|d⌧
dt (i 1)! 0
Z Z t
h 2 2 (t) ⇣ 2
t ⌘i 1 ⇣
2
⌘
= h 2 (⌧ )d⌧ exp h 2 (⌧ )d⌧
(i 1)! 0 0

we obtain

⇣ Z t ⌘
d k 2 2
(t) = c0 h (1 + %)2 (t) exp h 2 (⌧ )d⌧
dt 0
Z t
1 ⇣ ⌘i 1
Xk
2
· ( h + %) 2 (⌧ )d⌧
i=1
(i 1)! 0

and
 51

⇣ Z t ⌘ ⇣ Z t ⌘
d k 2 2 2
(t) < c0 h (1 + %)2 (t) exp h 2 (⌧ )d⌧ exp ( h + %) 2 (⌧ )d⌧
dt
⇣ Z t 0
⌘ d⌘ 0
2
= c0 h (1 + %)2 (t) exp % 2 (⌧ )d⌧ = (t).
0 dt
(6.11)
Notice that k (0) = 0 and ⌘(0) = 0. Therefore, from (6.11), we conclude that

k (t) < ⌘(t), for all t 2 [0, T ] and k = 1, 2, . . . , which finishes the proof.

In this last theorem, we have demonstrated a relation between error bounds,

specifically, that the latter bound is sharper. Particularly, we conclude from the form
of the error bounds that the numerical schemes indeed converge towards the exact
solutions as the step-size h tends to zero – a necessity for any numerical scheme –
from which we deduce that the schemes produce robust results.
In this thesis, we investigated general parabolic partial functional di↵erential equa-
tions and their approximate solutions constructed by means of spatial discretization
and iterative processes of the Picard type in the functional sense. We used di↵erential
inequalities and proved a variety of theorems on the approximate solutions and their
errors. Particularly, on the basis of the results that we have proved, we have arrived
at important convergence properties of the considered numerical schemes.
 52

REFERENCES

[1] C.T.H. Baker, C.A.H. Paul, 1996, A global convergence theorem for a class
of parallel continuous explicit Runge-Kutta methods and vanishing lag delay
di↵erential equations, SIAM J. Numer. Anal., 33:1559–1576.
[2] C.T.H. Baker, 2014, Observations on evolutionary models with (or without) time
lag, and on problematical paradigms. Math. Comput. Simulation 96, 4–53.
[3] B. Basse, B.C. Baguley, E.S. Marshall, W.R. Joseph, B. van Brunt, G.C. Wake,
D.J.N Wall, 2003, A mathematical model for analysis of the cell cycle in cell lines
derived from human tumours, J. Math. Biol. 47:295–312.
[4] B. Basse, Z. Jackiewicz, B. Zubik-Kowal, 2009 Finite-di↵erence and pseudo-
spectral methods for the numerical simulations of in vitro human tumor cell
population kinetics. Math. Biosci. Eng. 6, no. 3, 561–572.
[5] A. Bellen, M. Zennaro, 2003, Numerical methods for delay di↵erential equations,
Numerical Mathematics and Scientific Computation, Oxford University Press.
[6] A. Bellen, S. Maset, M. Zennaro, N. Guglielmi, 2009, Recent trends in the
numerical solution of retarded functional di↵erential equations. Acta Numer.
18, 1–110.
[7] M. Bjorhus, 1994, On dynamic iteration for delay di↵erential equations, BIT 43,
325–336.
[8] M. Bjorhus, 1995, A note on the convergence of discretized dynamic iteration,
BIT 35, 291–296.
[9] H. Brunner, H. Xie, R. Zhang, 2011, Analysis of collocation solutions for a class
of functional equations with vanishing delays. IMA J. Numer. Anal. 31, no. 2,
698–718.
[10] H. Brunner, C. Ou, 2015, On the asymptotic stability of Volterra functional
equations with vanishing delays. Commun. Pure Appl. Anal. 14, no. 2, 397–406.
[11] S. Brzychczy, 1987, Existence of solutions for non-linear systems of di↵erential-
functional equations of parabolic type in an arbitrary domain, Ann. Polon. Math.
47, 309–317.
 53

[12] A. Feldstein, A. Iserles, D. Levin, 1995, Embedding of delay equations in an

infinite-dimensional ODE system, J. Di↵erential Equations 117, 127–150.

[13] S.A. Gourley, Y. Kuang, 2005, A delay reaction-di↵usion model of the spread of
bacteriophage infection, SIAM J. Appl. Math. 65:550–566.

[14] S.A. Gourley, S. Ruan, 2012, A delay equation model for oviposition habitat
selection by mosquitoes. J. Math. Biol. 65, no. 6-7, 1125–1148.

[15] N. Guglielmi, E. Hairer, 2007, Sti↵ delay equations. Scholarpedia, 2(11):2850.

[16] N. Guglielmi, E. Hairer, 2016, Ernst Path-regularization of linear neutral delay

di↵erential equations with several delays. J. Comput. Appl. Math. 292, 785–794.

[17] K.J. in ’t Hout, M.N. Spijker, 1991, Stability analysis of numerical methods for
delay di↵erential equations. Numer. Math. 59, no. 8, 807–814.

[18] K.J. in ’t Hout, 1997, Stability analysis of Runge-Kutta methods for systems of
delay di↵erential equations. IMA J. Numer. Anal. 17, no. 1, 17–27.

[19] K.J. in ’t Hout, B. Zubik-Kowal, 2004, The stability of Radau IIA collocation
processes for delay di↵erential equations. Math. Comput. Modelling 40 no. 11-12,
1297–1308.

[20] A. Iserles, Y, Liu, 1997, On neutral functional-di↵erential equations with pro-

portional delays. J. Math. Anal. Appl. 207, no. 1, 73–95.

[21] A. Iserles, J. Terjki, 1995, Stability and asymptotic stability of functional-

di↵erential equations. J. London Math. Soc. (2) 5, no. 3, 559572.

[22] Z. Jackiewicz, Y. Kuang, C. Thalhauser, B. Zubik-Kowal, 2009, Numerical

solution of a model for brain cancer progression after therapy. Math. Model.
Anal. 14 no. 1, 43–56.

[23] Z. Jackiewicz, M. Kwapisz, E. Lo, 1997, Waveform relaxation methods for

functional di↵erential systems of neutral type, J. Math. Anal. Appl. 207, 255–285.

[24] Y. Kuang, 1993, Delay Di↵erential Equations with Applications in Population

Dynamics, Academic Press, Boston.

[25] S. Maset, M. Zennaro, 2014, Good behavior with respect to the sti↵ness in
the numerical integration of retarded functional di↵erential equations. SIAM J.
Numer. Anal. 52, no. 4, 1843–1866.

[26] U. Miekkala, O. Nevanlinna, 1987, Convergence of dynamic iteration methods

for initial value problems, SIAM J. Sci. Stat. Comput. 8, 459–467.
 54

[27] L.F. Shampine, 2005, Solving ODEs and DDEs with residual control, Appl.
Numer. Math., 52:113–127.

[28] L. F. Shampine, I. Gladwell, S. Thompson, 2003, Solving ODEs with MATLAB,

CUP.

[29] J. Szarski, 1976, Uniqueness of the solution to a mixed problem for parabolic
functional-di↵erential equations in arbitrary domains, Bull. Acad. Polon. Sci.
Ser. Sci. Math. Astron. Phys. 24, 814–849.

[30] S. Thompson, L.F Shampine, 2006, A friendly fortran 90 DDE solver, Appl.
Numer. Math., 56:503–516.

[31] S. Thompson, 2007, Delay-di↵erential equations. Scholarpedia, 2(3):2367.

[32] J. Wu, 1996, Theory and Applications of Partial Functional Di↵erential Equa-
tions, Springer, New York.

[33] B. Zubik-Kowal, 2008, Delay partial di↵erential equations. Scholarpedia,

3(4):2851.

[34] B. Zubik-Kowal, 1997, The method of lines for parabolic di↵erential-functional

equations, IMA Journal of Numerical Analysis, 17, 103–123.

[35] B. Zubik-Kowal, S. Vandewalle, 1999, Waveform relaxation for functional-

di↵erential equations, SIAM J. Sci. Comput. 21:207–226.

[36] B. Zubik-Kowal, 2000, Chebyshev pseudospectral method and waveform re-

laxation for di↵erential and di↵erential-functional parabolic equations, Appl.
Numer. Math. 34:309–328.

[37] B. Zubik-Kowal, 2004, Error bounds for spatial discretization and waveform
relaxation applied to parabolic functional di↵erential equations, Journal of Math-
ematical Analysis and Applications, 293, 496–510.

[38] B. Zubik-Kowal, 2014, Numerical algorithm for the growth of human tumor cells
and their responses to therapy, Appl. Math. Comput. 230, 174–179.
 55

APPENDIX

In this thesis, we apply the following one-dimensional version of [34, Theorem 1].
For this application, we use the symbol Fc ({xi : i = 0, ±1, . . . , ±M̃ } ⇥ [ ⌧0 , 0], R)
to denote a class of functions continuously di↵erentiable with respect to the second
argument. We also use similar notation for functions on similar domains.

Theorem 6.7. We assume that the following conditions are satisfied.

(i) : [0, T ] ⇥ R+ ! R+ is continuous, nondecreasing with respect to the second

input and such that there exists the right-hand maximum solution ! on [0, T ] of
the initial-value problem
8
>
< !(t)
˙ = (t, !(t)),
>
: !(0) = 0,

(ii) G, H : [0, T ] ⇥ {xi : i = 0, ±1, . . . , ±M 0 } ⇥ Fc ({xi : i = 0, ±1, . . . , ±M̂ } ⇥

[ ⌧0 , 0], R) ! R satisfy the inequality

h
G(P )  H(P ),
2

where P is any point in the domain of G and H,

(iii) ⇢ 2 Fc ({xi : i = 0, ±1, . . . , ±M̃ } ⇥ [ ⌧0 , T ], R) is such that ⇢(xi , t) = 0 if

t 2 [ ⌧0 , 0] and i = 0, ±1, . . . , ±M̃ or if t 2 [ ⌧0 , T ] and i = ±M, . . . , ±M̃ , is
 56

di↵erentiable with respect to the second input and satisfies the inequality

@⇢ ˜⇢(xi , t)G(xi , t, ⇢(x ,t) ) ˜(2) ⇢(xi , t)H(xi , t, ⇢(x ,t) )

(xi , t) i i
@t
 (t, k⇢(xi ,t) khn ),

for i = 0, ±1, . . . , ±M 0 , where

⇣ ⌘
˜⇢(xi , t) = 1 ⇢(xi+1 , t) ⇢(xi 1 , t) ,
2h
⇣ ⌘
˜(2) ⇢(xi , t) = 1 ⇢(xi+1 , t) 2⇢(xi , t) + ⇢(xi 1 , t) ,
h2
k⇢(xi ,t) khn = max{|⇢(xi + xj , t + ⌧ )| : j = 0, ±1, . . . , ±M̂ , ⌧ 2 [ ⌧0 , 0]}.
(6.12)

Then,
|⇢(xi , t)|  !(t),

0
for all t 2 [0, T ], i = 0, ±1, . . . , ±M .

For clarity, we omit the proof, which is too technical for the purpose of this thesis.
We refer the reader to [34], where the proof is presented for the multi-dimensional
case.