0% found this document useful (0 votes)
125 views131 pages

Mutib - Thesis - Numerical Methods For Solving Delay-Differential Equations

This chapter introduces delay-differential equations (DDEs) and discusses numerical methods for solving them. DDEs arise in areas like control systems, neural networks, and population dynamics. Existing numerical methods for DDEs have limitations. The chapter reviews these methods and recent work analyzing their stability properties to address those limitations. It lays the groundwork for adapting standard ODE solvers to DDEs and analyzing the properties of the adapted methods.

Uploaded by

785902070
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
125 views131 pages

Mutib - Thesis - Numerical Methods For Solving Delay-Differential Equations

This chapter introduces delay-differential equations (DDEs) and discusses numerical methods for solving them. DDEs arise in areas like control systems, neural networks, and population dynamics. Existing numerical methods for DDEs have limitations. The chapter reviews these methods and recent work analyzing their stability properties to address those limitations. It lays the groundwork for adapting standard ODE solvers to DDEs and analyzing the properties of the adapted methods.

Uploaded by

785902070
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 131

NUMERICAL METHODS FOR SOLVING

DELAY-DIFFERENTIAL EQUATIONS

BY

AHMAD NASSIR AL-MUTIB

THESIS SUBMITTED FOR THE DEGREE OF

DOCTOR OF PHILOSOPHY IN SCIENCE AT THE

VICTORIA UNIVERSITY OF MANCHESTER

MAY 1977
ProQuest N um ber: 13917012

All rights reserved

INFORMATION TO ALL USERS


The quality of this reproduction is d e p e n d e n t upon the quality of the copy subm itted.

In the unlikely e v e n t that the a u thor did not send a c o m p le te m anuscript


and there are missing pages, these will be noted. Also, if m aterial had to be rem oved,
a n o te will ind ica te the deletion.

uest
ProQuest 13917012

Published by ProQuest LLC(2019). C opyright of the Dissertation is held by the Author.

All rights reserved.


This work is protected against unauthorized copying under Title 17, United States C o d e
M icroform Edition © ProQuest LLC.

ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, Ml 4 8 1 0 6 - 1346
S" Q> & V o
TO THE MEMORY OF MY FATHER,

TO MY MOTHER; FOR HER SACRIFICE

AND ENCOURAGEMENTS,

TO MY SISTER MAYSOON AND BROTHERS

KHALID, HAMMAD, AND TAWFIEK.


ACKNOWLEDGEMENTS

I should lik e to express my deep thanks to Professor Joan Walsh *

fo r her useful comments and fo r a ll the help throughout the preparation

o f th is th e s is , and also to a ll the Numerical Analysis s t a f f a t the

U n ive rsity o f Manchester.

In a d d itio n , I would lik e to thank the U n iv e rs ity o f Riyad, Saudi

Arabia, fo r the scholarship which enables me to c a rry out th is research.

My thanks also to Miss Kendal Anderson fo r the e f f ic ie n t typin g

o f th is th e s is .

(iii)
STATEMENT

No p o rtio n o f the work re fe rre d to in th is the sis has been submitted

in support o f an a p p lic a tio n fo r another degree or q u a lific a tio n o f

th is or any other u n iv e rs ity or other in s titu te o f le a rn in g .

( iv )
PREFACE

This th e sis is concerned with, the adaptation o f numerical methods

fo r so lving i n i t i a l value problems in ordinary d iffe r e n tia l equations to

the problem o f d e la y -d iffe re n tia l equations.

In chapter I , the d e fin itio n o f delay d if fe r e n tia l equations and

the areas where they a rise are given5 together w ith a b r ie f discussion o f

numerical methods fo r solving i n i t i a l value problems in ODE, a review o f

e x is tin g numerical methods fo r solving DDE w ith th e ir advantages and

lim ita tio n s 3 and a discussion o f recent work on the s t a b ilit y prope rties

o f numerical methods.

In chapter I I , we s ta r t by givin g a review o f the work done on the

theory o f DDEs, then we sta te theorems sp e cifyin g con dition s fo r the

existence and uniqueness o f the s o lu tio n o f DDEs. Then we discuss one

o f the c h a ra c te ris tic s o f DDEs, th a t the d e riv a tiv e o f the so lu tio n may

have delay induced jump d is c o n tin u itie s . F in a lly we consider a n a ly tic a l

methods fo r solving a system o f lin e a r DDEs, and give re s u lts concerning

asymptotic prope rties o f the s o lu tio n .

In chapter I I I , we consider in general the d i f f i c u l t i e s a ris in g in

the adaptation o f numerical methods to solve DDE, and discuss ways to

overcome them. Then, we consider s u ita b le d e fin itio n s fo r s t a b ilit y o f

numerical methods fo r solving DDEs.

In chapter IV, we consider the analysis o f s p e c ific numerical methods

o f Runge-Kutta type fo r solving DDE. In p a r tic u la r , we consider ways o f

approximating the delay term and the e ffe c t o f th is approximation on the

pro p e rtie s o f the numerical methods considered.

F in a lly , in chapter V, a d e s c rip tio n o f algorithm s fo r solving i n i t i a l

fu n c tio n problem in DDE w ith a va ria b le stepsize is given. The p rope rties

o f these methods are v e r ifie d by considering a v a rie ty o f te s t problems.

(V)
CONTENTS

Page
CHAPTER I INTRODUCTION

1.1 In tro d u c tio n . i

1.2 D e la y -D iffe re n tia l Equations. 1

1.3 Numerical s o lu tio n o f i n i t i a l value ordinary

d iffe r e n tia l equations. 4

1.4 Review o f numerical methods fo r solving DDE. 8

1.5 S t a b ilit y analysis o f numerical methods

fo r so lvin g DDE. 16

1.6 Remarks. 18

CHAPTER I I THEORY OF DELAY-DIFFERENTIAL EQUATIONS

2.1 In tro d u c tio n . 20

2.2 Existence and Uniqueness o f the s o lu tio n

o f DDE. 20

2.3 D is c o n tin u ity o f the d e riv a tiv e o f the

s o lu tio n o f DDE. 23

2.4 Asymptotic prope rties o f the s o lu tio n o f a

lin e a r system o f DDEs. 25

CHAPTER I I I NUMERICAL METHODS FOR SOLVING DELAY-DIFFERENTIAL

EQUATIONS

3.1 In tro d u c tio n . 34

3.2 Numerical d if f i c u l t i e s in solving DDE. -34

3.2.1 D is c o n tin u ity o f the d e riv a tiv e . 35

3.2.2 The evaluation o f the delay term. 36

3.3 D e fin itio n s o f s t a b ilit y p rope rties o f numerical

methods fo r solving DDE. 41

3.4 Some numerical methods fo r solving ODE. 44

(V i)
CHAPTER IV THE ANALYSIS OF SPECIFIC NUMERICAL METHODS FOR DELAY-

DIFFERENTIAL EQUATIONS ■

4.1 In tro d u c tio n . 47

4.2 The adaptation o f numerical methods fo r

so lvin g DDE. 47

4.2.1 Kutta-Merson method fo r so lving DDE. 48

4.2 .2 The trapezium method fo r s o lv in g DDE. 57

4 .2 .3 The im p lic it Runge-Kutta method fo r

so lving DDE. 59

4.3 Local tru n c a tio n e rro r estimate o f methods

fo r so lvin g DDE. 60

4.3.1 LTE estim ate o f Kutta-Merson method

fo r solving DDE. 61

4.3 .2 LTE estim ate o f the trapezium method and

fo r the im p lic it Runge-Kutta method fo r

so lving DDE. 70

4.4 S ta b ility properties o f numerical methods

fo r so lving DDE. 71

4.4.1 Kutta-Merson method fo r so lving DDE. 71

4.4.2 The trapezium method fo r s o lv in g DDE. 73

4 .4 .3 The im p lic it Runge-Kutta method f o r so lving

DDE. 76

CHAPTER V NUMERICAL RESULTS AND ALGORITHMS

5.1 In tro d u c tio n . 82

5.2 D escription o f algorithm s fo r so lving DDE. 83 •

5.3 Test problems and numerical re s u lts . 87

5.4 Concluding remarks. 97

BIBLIOGRAPHY 115

( vi i )
CHAPTER I

INTRODUCTION

1.1 In tro d u c tio n

In th is chapter, we f i r s t s ta r t by d e fin in g d e la y -d iffe re n tia l

equations, and we describe some o f the areas where they a ris e ; we then

give a b r ie f discussion o f numerical methods fo r solving i n i t i a l value

problems in o r d in a ry -d iffe re n tia l equations. F in a lly we present a review

o f e x is tin g numerical methods fo r solving d e la y -d iffe re n tia l equations w ith

t h e ir advantages and lim ita tio n s ., and we discuss the recent work on

s t a b i l it y p rope rties o f these numerical methods.

1.2 D e la y -D iffe re n tia l Equations.

Many physical systems possess the feature o f having a delayed response

to in p u t c o n d itio n s , so th a t the ra te a t which processes occur depends not

on ly on the cu rre n t sta te o f the system but also on the past s ta te s .

Mathematical models o f such processes commonly re s u lt in d iff e r e n tia l

equations w ith a term in v o lv in g a time delay. Equations o f th is type are

c a lle d D e la y -D iffe re n tia l Equations (DDE).

A general form o f a f i r s t order DDE is


2;

u '( t ) » f ( t , u ( t ) , u ( t - d ( t , u ( t ) ) ) ) ( 1 -la )

where d ( t 3 u {t)) $ 0 is re fe rre d to as the "Delay11 and t - d ( t s. u ( t ) )

is re fe rre d to as the ‘•'Lag". In general the delay is a fu n c tio n both o f

t and o f the s o lu tio n u (t).

D e la y -D iffe re n tia l Equations arise in a wide v a rie ty o f fie ld s * from

the growth o f in fe c tio n to a model o f a h e a rt-1ung complex, re action to

X-ray treatm ent, and population growth. DDEs also a ris e in con trol theory*

machine to o lin g * and a n a ly tic number theory. A more d e ta ile d study o f the

a p p lic a tio n s o f DDEs can be found in Schmitt (1972), Halanay(1966), and

O guztoreli (1966).

The i n i t i a l con dition s needed to determine the s o lu tio n o f the DDE

( 1 . 1a) f o r t > t , g e nerally involve prescribed values o f u ( t) on an

in te r v a l. The precise nature o f these i n i t i a l conditions depends on the

nature o f the delay and the range o f values o f the argument t. I f the

s o lu tio n o f the problem is defined by an i n i t i a l value a t a s in g le p o in t,

we sh a ll re fe r to the problem, w ith Feldstein (1964), as an i n i t i a l value

d e la y -d iffe re n tia l equation (IVDDE), otherwise we s h a ll c a ll i t an i n i t i a l

fu n c tio n d e la y -d iffe re n tia l equation (IFDDE). In general, i f we seek the

s o lu tio n o f (1.1a) f o r t Q ^ t < T, we require an i n i t i a l fu n c tio n o f the

form

u ( t) = g ( t ) , m in (tQ, t * ) < t s t 0 ( 1 . 1b)

where t* = minffc - d ( t , u ( t ) ) ] , t e [ t Q, T ],

A more general DDE than (1.1a) is one w ith m u ltip le delays o f the

form

u '( t ) = f ( t , u( t ) , u ( t- d 1( t , u( t ) ) , . . . , u ( t-d q( t 9 u ( t ) ) ) ) ( 1 . 2)

where d ^ ( t, u ( t) ) 5 0 fo r i=l,2,...,q. A system o f n DDEs w ith

m u ltip le delays has the form

u * ( t) = f ^ t , . . . , U j(t-d ^ k ( t , ux( t ) , . . . , un( t ) ) ) , . . . ) (1 .3)

where i , j = 1 , . . . , n, k = 0, 1 ........ q, and dj k ^ 9 un^t ^ 5 0


3.

Before we give some examples o f DDEs, we mention th a t i f the d e riv a tiv e

a t the present time is a fu n c tio n o f the s o lu tio n a t the present time and

the s o lu tio n a t some fu tu re tim e, then we have an A dvanced-D ifferential

Equation o f the form

u * ( t) = f ( t , u ( t ) , u (t+ d (t, u ( t ) ) ) ) » d ( t , u ( t) ) 5 0 ( 1 *4)

Also i f the d e riv a tiv e depends on the s o lu tio n a t the present time and both

past and fu tu re times then we have a N e u tra l-D iffe re n tia l Equation o f the

form

u '( t ) = f ( t , u ( t ) , u f t - d j ( t , u ( t ) ) ) , u(t+d2( t 9 u ( t ) ) ) ) (1 .5 )

where d ^ t , u ( t ) ) , d2( t 9 u ( t ) ) 5 0 .

We sha ll not be concerned w ith these two types, but only w ith DDEs.

Examples o f (1 .1) are as fo llo w s :

Example (1 . 1 ) . A lin e a r IFDDE w ith a constant delay is

u '( t ) = b u (t - d ) , t 5 tQ ( 1 -6)

u ( t) = c tQ - d $ t $ tQ

This example illu s tr a te s many c h a ra c te ris tic s o f more general DDE

problems. Consider the above equation (1.6) in the form of the general

DDE (1 .1 ). The fu n c tio n s , f , d ( t , u ( t ) ) , and g ( t) are a ll a n a ly tic ,

nevertheless, i t is apparent th a t some d e riv a tiv e has d is c o n tin u itie s a t

every p o in t t Q + jd fo r j = 0, 1, 2 , . . . , and i t can be shown th a t a t

each p o in t t Q + jd , the f i r s t j d e riv a tiv e s o f u ( t) are continuous,

but the ( j+ l) s t d e riv a tiv e has a f i n i t e jump d is c o n tin u ity .

Example ( 1 .2 ) . An example o f a lin e a r IVDDE is the problemconsidered

by Fox e t al (1971),

u '( t ) = a u (t) + b u (A t), X is constant (1-7)

u (0) = c,

th is is a DDE only if o < A l < 1.

Example ( 1 . 3 ) . An example o f IFDDE with m u ltip le constant delays is given

by F e ld ste in and Goodman (1973)


4.

u '( t ) = u ( t - 1) - 3u (t - 72) + u ( t - tt) ( 1 .8)

u ( t) =1 fo r t ^ 0

Example (1 . 4 ) . An example o f a n o n-line ar IFDDE is given by Neves and

F e ldstein (1973)*

u '( t ) = i o ( t ) .u (& n (u (t))> t j 1 (1.9)

u ( t) =1 fo r t s< 1 .

W ithout knowing the s o lu tio n o f th is equation* i t is not possible to p re d ic t

th a t u<P>(t) has jumps a t t = e ( fo r p 5 2) and a t t = e2 ( fo r p 5 3)

The unique s o lu tio n is

t 1 ^ t £ e
u (t) «
e x p (t/e ) e ^ t $ e2

Further examples o f DDEs w ith d iffe r e n t kinds o f delays w ill be given

in chapter V.

1.3 Numerical s o lu tio n o f in . it ia l value ordinary d iffe r e n tia l equations

Since the methods to be considered here fo r forming approximations

to ( 1 . 1) are extensions o f those used fo r o rdina ry d iff e r e n t ia l equations*

i t is in order to discuss b r ie f ly the basic methods fo r i n i t i a l value

problems. A d e ta ile d discussion can be found in Lambert (1973).

Consider the i n i t i a l value problem

u 1( t ) = f ( t * u( t ) ) tQ$ t s T * (1.10a)

u ( t0) = c ( 1 . 10b)

where f is assumed to be L ip s c h itz continuous in u ( t) fo r a ll

t e [ t Q, T ], In a d d itio n * we w ill assume th a t a ll the p a r tia l d e riva tive s

of f up to an a r b itra ry but f i n i t e order e x is t* are continuous and

bounded.

Let y (t) be the numerical approximation to u ( t) * then a class o f

general k-ste p methods which generate the s o lu tio n y ( t n)* n = 0 * 1 , . . . , N*

t = t Q + nh* and Nh = T - t 0» is
5.

y (tr ) - S (fv} 0 <: r < k, (s ta rtin g values) (1.11a)

Examples o f these methods are

(1) The lin e a r m u ltiste p methods (LMM)

k
( 1 . 12)

(2) The one-step methods o f Runge-Kutta type, k = 1

m
(1.13)

m
kr = f ( t n + hV + hX br1lt1)
1*Hf

r = 1 , 2 , ..., m; 0 ,< a^ ^ 1

The q u a n tity en = y ^ p ) “ u(^ n) 1S c a lle d the global d is c re tiz a tio n

e rro r a t t n- I f th is e rro r tends to zero as' h -»- 0, we say th a t the

method is convergent, hence,

D e fin itio n 1 .1 . A method o f class (1.11) is convergent i f , when applied

to any problem o f class ( 1. 10) , y ( t R) u (t) as h -5- 0 , where n = ( t - t Q)/h

fo r any t e [ t Q, T l .

I f we s u b s titu te the tru e s o lu tio n u (t) in (1.11) then we define

the loca l tru n c a tio n e rro r (LTE) dn by

dr = u ( t r ) - s ( t r ) 0 $ r < k

where p(z)

dn represents the amount by which the tru e s o lu tio n u (t) f a i l s to s a tis fy

(1 .1 1 ). A minimal requirement fo r a convergent method o f class (1.11) is

consistency defined by.


6.

D e fin itio n 1 .2 . A method o f class (1.11) is said to be co n siste n t i f

max Id I 0 as h ->• 0 .
0$n$N
I f p is the la rg e s t in te g e r fo r which

d „ = 0(hp"KL) (1.15)

then the method ( 1 .11) is said to be o f order p .

To study any p e rtu rb a tio n in the s ta rtin g values o r in the d iffe re n c e

equation ( 1 . 11) as h •* Os nh = t , we have the fo llo w in g d e fin itio n .,

D efinition 1.3. The method (1.11) is said to be zero-sta ble i f no ro o t

o f the polynomial p(z) =


i a.z has modulus greate r than one, and i f
1=0
every ro o t w ith modulus one is simple.

We note th a t con sistent Runge-Kutta methods (1.13) are always zero-

sta b le since the on ly ro o t o f p(z) is 1 .

Now we present the fundamental re s u lt concerning the convergence o f

numerical methods o f class ( 1. 11) .

Theorem 1 .1 . A method o f class (1.11) is convergent i f and only i f i t is

both co n siste n t and ze ro -sta b le .

Theorem (1.1) was f i r s t proved fo r the case o f LMM (1 .1 2 ), by

D ahlquist (1956). A proof fo r the class (1.11) may be found in Isaacson

and K e lle r (1966).

As defined above, z e r o - s ta b ility ensures the stab le propagation o f

the e rro r in the l i m i t as h -> 0 , but in pra ctice we take h as a fix e d

p o s itiv e number, so we are concerned w ith the behaviour o f the global

d is c re tiz a tio n e rro r as n -> « w ith h fix e d . Id e a lly we would lik e a

r e s u lt o f the form: fo r a given fix e d steplength h and fo r a given

i n i t i a l value problem ( 1 . 10), the global d is c re tiz a tio n e r ro r en - y ( t ) - u ( t )

remains bounded as n «. The snag w ith th is requirement is the dependence

on the behaviour o f the s o lu tio n o f (1 .1 0 ). Accordingly a s p e c ific i n i t i a l

value problem o f the form

u '( t ) = a u (t), a is constant (1.16)

is considered as a te s t equation. We now make the assumption th a t when we


s u b s titu te (1.16) in ( 1. 11) , <k~ in ( 1. 11b) becomes lin e a r in y (t .),
t n+i
i = 0 , 1 , . . . , k. S p e c ific a lly , we assume th a t

W =aj 0V haM W e - 17>

where f o r the lin e a r m u ltiste p method ( 1. 12)

Y-j (ha) = e.., i = I,.,., k

and fo r the Runge-Kutta method (1.13)

Yx(ha) = 0; Y0(ha) = P ^ h a)/Qm(ha)

where Pm_^ is a polynomial o f degree a t most m-1 and Q is a polynomial

o f degree a t most m.

A fte r applying the method (1.11) to the IVP (1 .1 6 ), subtractin g.e qua tion

(1.14) from (1.11) and neglecting the local rounding e r r o r , we get

k
I - hayi (ha)>en+i = - / ° ( 0 e t y i + u (1.18)

0 S n ^ N -k

The s o lu tio n o f the d iffe re n c e equation (1.18) is o f the form: Complementary

Function + P a rtic u la r S o lu tio n . The p a rtic u la r s o lu tio n o f (1.18) can be

kept small i f the lo ca l tru n c a tio n e rro rs are s u f f ic ie n t ly sm all, and the

complementary fu n c tio n depends on the roots o f the s t a b ilit y polynomial

k
n ( r , h a ):= \ {a. - h a y ^ h a jlr 1 (1.19)
i =0 1 1

This leads us to the d e fin itio n o f absolute s t a b i l it y .

D e fin itio n 1 .4 . The method (1.11) is said to be a b s o lu te ly stable f o r a

given ha i f , fo r th a t ha, a ll the roots o f the s t a b ilit y polynomial

(1.19) l i e w ith in the u n it c ir c le . A region R(ha) o f the ha-complex

plane is said to be a region o f absolute s t a b ilit y o f ( 1. 11) , i f ( 1. 11) is

a b so lu te ly stable f o r a ll ha e R(ha).

The region o f absolute s t a b ilit y is determined on ly by the

c o e ffic ie n ts o f the method (1. 11) , and the corresponding la rg e s t

value o f h fo r which en w ill decrease is dependent on a, and


8.

hence on the p a rtic u la r d iffe r e n tia l equation whose s o lu tio n is sought*

For some class o f IVP in (1 .1 0) 3 we would lik e to fin d numericalmethods

whose stepsize h is not re s tric te d by the values o f a*, th is leads to

the fo llo w in g d e fin itio n .

D e fin itio n 1 .5 . A method is defined to be A -stable i f the absolute s t a b ilit y

region includes the h a lf plane Re(ha) < 0.

Here we give Lambert's d e fin itio n , Lambert (1973), o f the class o f

s t i f f d iffe r e n tia l equations, where A-stable methods are useful in s o lv in g .

Consider the lin e a r system o f d iffe r e n tia l equations

j j ' ( t ) = A u(t) + g ( t) (1.20)

where j j ( t ) is an m -dimensional v e c to r, A is an r n x m m a trix , and

£ (t) is a smooth fu n c tio n o f t. Let be the eigenvalues o f

the m a trix A, then

D e fin itio n 1*6 . The lin e a r system (1.20) is said to be s t i f f i f

( i ) Re(Ai ) < 0 , i = 1, 2, . . . , m

( ii) max |Re( A- ) | » min |Re( X- ) I -


Ui$m 1$i$m 1

1.4 Review o f numerical methods fo r solving DDE

I t appears in the lite r a t u r e th a t there is wide in te re s t in the

numerical s o lu tio n o f DDE and many approaches have been adopted fo r solving

p a rtic u la r equations. One o f the best surveys o f these methods is in

Cryer (1972), see also Baker (1976). Here, we give a review of'some o f

the general methods o f s o lu tio n .

Bellman (1961) , developed a technique f o r s o lv in g DDEs very

a ccu ra te ly a t the expense o f increasing the order o f the system. To use

his method we must have a delay o f the form d ( t ) , which must be monotone

decreasing, w ith d ( t) > 0. D efining the sequence { t n> by means o f

V ^ n ’ ^ n -r ' n?1’

the s o lu tio n on t $ t $ t ^ is given by the co n tin u a tio n process in

terms o f previous values on [t -j s t ]. We suppose the s o lu tio n has been


9.

found fo r tQ$ t $ ts and we define vQ( t ) = t , v x( t ) = t - d ( t ) 9

v n( t ) = v J v J t ) ......... ve( t) = v f v r (t)) - We then form ulate, on


I S

[ t s, t i ]» a system o f d iffe r e n tia l equations; l e t

ys(t) = y(t)

ys_T(* ) = y ( v i ( t ) )
0 - 21)
y s -2^ = y ( v2(t ))

y 0 (t) = y(vs(t))

fo r t s s t < t J+i . As t varie s in [ t s> t , 1» vr ( t ) varie s through

[ t s_r > t r+-| ] atld the o rig in a l equation ( 1. 1) now y ie ld s

dv~~~ {V vs - r ^ } = f (vs - r ^ ’ yr ^ » yr-i

w ith y r ( t s) = y ( t r ) , t s s t s t J+i ,fo r r =0» s. From these

equations we obtain, fo r t e [ts, ]

y ;(t) = f ( t , ys ( t ) , y s_1( t ) ) , y s( t s) = y ( t s)

y ; . - | ( t ) = f ( t , y s_ -|(t), y s_2( t ) ) . v ^ ( t ) , y s-1 ( t s) = y ( t s_1)


: ( 1 - 22 )

y'0W = f (t> y0(t)» y(vs+i ( t ) ) ) . v ^ ( t ) , y0( t s) = y ( t 0)

where y(v +- | ( t ) ) is a given i n i t i a l fu n c tio n defined on [ t * 9 t Q] .

In te g ra tio n o f the system (1.22) may be achieved w ith an automatic program

fo r a system o f the f i r s t order equations, By in te g ra tin g th is system

o f equations, the fu n ctio n s y (t) = ys( t) is determined over the in te rv a l

t s S t s t g+.j which enables us to continue fo r the next in te r v a l.

We observe here th a t an accurate s o lu tio n is obtained only a t the'

expense o f in te g ra tin g a large and growing system. However, the method

has the advantage th a t no storage o f previous fu n c tio n values is re q u ire d ,

only the i n i t i a l conditions need be saved. This technique has another draw

back, th a t the system formed by adding an equation fo r each in te r v a l, as


oaA Ko*u«v\
in (1 .2 2 ), may be unstable. Bellman^(1962) points out th a t t h is i n s t a b i l i t y
10.

was discovered fo r a c e rta in DDE which arises in number theory.

E l's g o l^ ts (1966) has proposed a method using the E uler formula

fo r solving DDEs w ith va ria b le delay o f the form d (t). He assumes th a t

the g rid can be constructed such th a t f o r each n 5 1 < n <: N+l ,

e ith e r t fi - d ( tn) s t Q (1.23)

or ‘ n " d( V = ^ (n ) q (n ) " n

E u le r's method then takes the form

hnf ( t V y (t n) ’ 9( t n- d ( t n) ) ) , i f V < i ( t n) St 0
(1.24)
y< W - y< V =
hnf ( t „ * y ( V ’ y (t q ( n p ) otherwise

y ( t 0) = g ( t 0) , and 0 S n j «.

This method does not work unless con dition (1.23) is s a tis fie d .

Condition (1.23) is s a tis fie d i f min d (t) > 0 o r i f t - d ( t ) is s t r i c t l y


t 0$t$T
monotone increasing on [ t 0, T ]. The case where the Euler method (1.24)

does not work is ca lle d by E l's g o lt's the "s in g u la r case" 9 see also

E l's g o lt's and Norkin (1973).

With these d i f f i c u l t i e s in mind., Feldstein (1964) proposed several

m o d ifica tio n s o f the Euler method to solve DDEs w ith a delay o f the form

d (t). His study is re s tric te d to the i n i t i a l value DDEs. We now describe

one o f his methods9 c a lle d EA-1.

Let the g rid be uniform and l e t t fi+-j = t n+h9 d e fin e :

V d< V where [z ] means the in te g e r p a rt o f z.


q(n) ¥

r(n ) = - — p q (n ), so 0 j r(n ) < 1

V d< V = ^q(n) + r (n>h ’

Then., F e ld s te in 's a lg orith m EA-1 is defined as fo llo w s

zo = y ( V

y0 = y ( V
11.

zn = y ,(n ) + h’ r <">f q(n) <■' ' 24^

fn = f (V V zn>

yn+l = y n + h f n

The order o f c a lc u la tio n is f , y n+^ 3 z ^ and values o f t *f and

yn are kept fo r la te r use9 w ith the p o s s ib ility o f large store requirements

F eldstein analysed (1.24*Ct)and showed th a t i t is convergent o f order one.

To get higher order one-step methods fo r solving DDEs9 Tavernini

(1971) has presented an in te re s tin g th e o re tic a l study o f numerical methods

generating the s o lu tio n in a continuous approximation.

For conveniences we w ill adopt the fo llo w in g n o ta tio n : the problem

considered is

u '( t ) = f ( t 9 u( t)s u ( t - d ( t 9 u ( t ) ) ) ) a F(u9 t ) (1.25)

u ( t) = g ( t ) 3 t e [ t * 9 t Ql

The uniform stepsize h - ( T - tQ)/m fo r some in te g e r m3 and the mesh

po ints are denoted by t^ - t 0+ ih s i = 0 S l s . . . 9 m. Define the sets

S = { ( t 9 s )|t0 $ t * T and 0 < s ,< T - t } ,

T.h = { t n,
os t ll 5 . . . , t m-1i }

Proceeding as in ordina ry d iffe r e n tia l equations., we define the tru e

increment fu n c tio n

( u ( t + S ) - u ( t ) )/<5 s 6 > 0
A (ts 6) = (1.26)
u '( t ) s 6=0

f o r every ( t 9 6) e S. We next approximate a by an increment fu n c tio n

and proceed to approximate the tru e so lu tio n u using the stepsize h by

a fu n c tio n y. To advance the s o lu tio n from tn to t ^*

n = Os I s . . . , m-1s

y ( t n+rh) = y ( t n) + rh ^ fy , t n , h9 r ) s r e [0 9 1] (1.27)

which is a continuous approximation o f the s o lu tio n on the in te rv a l

Lt s t I f the evaluation o f cj> requires the value o f y a t some


n n+1
p o in t other than a mesh p o in t9 say a t s 9 then we have e ith e r s $ t Q and

g(s) is usedj or s > t0 and the re la tio n

y (t+ rh ) = y ( t ) + rh<|)(y9 t , h 9 r ) 9 re [0 , 1 ]

is used w ith t = max(t. < s) and r = —r—.


i ' n
If <p is chosen to be F we obtain E u le r's method:

y (t+ rh ) = y ( t ) + rh F (y 9 t ) 9 r e [0, 139 t e Th (1.28)

The next approximation o f the tru e increment fu n c tio n is obtained byusing

a lin e a r combination o f interm ediate evaluations o f F:

A (t, rh ) = 4>(u9 t 9 h9 r ) = ^ a *(r)F (u , t+ b.h) (1*29)


i= l 1 1

where 0 £ $ 1 and ai ( r ) | e [0, 1] , i = 1, 2, . . . , q.

Suppose we have s p e c ifie d an increment fu n c tio n <|>which y ie ld s a

s o lu tio n o f order q -15 and we have progressed so fa r asto have fo r

t £ Th

y(s) [ t 0, t ]
y * ( S) =
S“ t '
y ( t ) + ( s - t ) f ( y 9 t s h9 —pp~)9 [ t 9 t+ h ]

We now use (1.29) as the increment fu n c tio n 9 to obtain

y (t+ rh ) = y ( t ) + rh ^ (y * 9 t 9 h9 r ) 9 re [0 , 1 ] (1.30)

Assuming th a t the tru e s o lu tio n u (t) is in >(q+l); [ t Q9 T ] and


C'M

u(q+1) is bounded by Y9 then

a ( 1 9. rh ) = ^ j 1 U [ Z) + (q+TyT
. _ (1.31)
rh

fo r t e [ t Q, T -h ] 9 and re [0, 1 ] 9 |x t | < 1. Also

q (b .h )J_1 M . (b .h )q
F(u. t+ b .h ) = ^ ( j- i) ; u (* ) + xi ,

fo r |x i t | $ 1 9 s© t - k a t
j -1
Therefore i f ^ (u 3 t 3 h3 r ) is to be a qth order approximation to

A ( t 9 rh) the a ^ ( r ) 's must s a tis fy the algebraic system

Bw( t ) = x ( r )
i -1 r ^
where ( B ) ^ = b. 9 w ^ r ) = a . ( r ) 9 v ^ —r—.

The m a trix B is non-singular i f and only i f the Vandermonde

determ inant formed from b 9. . . 9 b is non-zero. Hence fo r any choice o f


i Cj
d is tin c t b 's we can fin d polynomials al9 . . . # a on [0, 1 ] such th a t

,q
||*(u, t , h, r) - A(t, rh)|[ s Y [ l + ( q + 1 ) N ] (1.34)

i . e . the method (1.30) is o f order q 9 <xt t&a-vb ~

Tavernini shows th a t under c e rta in L ip s c h itz co n d itio n s on the

increment fu n ctio n s the above constructions y ie ld methods which are

convergent. A lso, he gives an example of a method o f order 29 then uses

the second order method to generate a th ir d order method and a th ir d order

to generate a fo u rth order method. His fo u rth order method is

y j ( t n+rh) =y ( t p) + rhF(y, t n)

y2( t n+rh) =y ( t p) + rh[(1 - £)F(y, t n)

+ £ F(y 1» t n+l>)]

y 3( t n+rh) =yftp) + rh t(l - §r + §r2)F(y, t„)

+ (2r - | r 2)F(y2, t n + ]h)

+ (- + | r 2)F(y2, t n+h)] •

y ( t n+rh) = y ( t n) + rh[(1 - | r + | r 2)F(y, t n)

+ (2r - | r 2)F(y3, t n + | h )

+ (- + | r 2) F(y3» V hU ( i. 3 5 )

where 0 s< r s< 19 y ( t ) is known fo r t $ t' .

Tavernini (1971) also mentions th a t i t is possible to obtain one-


14.

step methods by re placing the in te g ra l in the id e n tity

ft+ rh .

u (t + rh) - u (t) - F(u9 s)ds (1.36)


h

by an in te rp o la tio n polynom ial. H ill (1973) has developed a number o f

methods u s in g -d iffe re n t in te rp o la tio n polynomials fo r approximating the

integrand in (1.36) and then in te g ra tin g .

In another work, Tavernini (1969) has described lin e a r m u ltis te p

methods th a t apply to the DDE (1 .1 ) w ith a delay o f the form d (t). The

lin e a r m u lti step methods he studied have the form

k-1
“ ky ( t n+k-l + r h > + 7 0“ i< r >y ( W

= y< W ’

y ( t n+i - d ( t n+1. ) ) ) (1.37)

where 0 < r $ 1 and t^ +-j = t^+h. Conditions are imposed on the

c o e ffic ie n ts a ^ (r) and 3^ ( r ) which ensure th a t the fu n c tio n y (tT -v 'h ’)

is continuous in r.

M u ltiste p methods are also considered by Zverkina (19649 1965) f o r

DDEs which have a non-smooth s o lu tio n as in example 1. The basic idea o f

Zverkina is th a t the lo c a tio n o f jumps <5 in d e riv a tiv e s o f u ( t) can be

computed. Using th is inform ations the modified lin e a r m u lti step method

has the form

X “ l y ( t n + l) = y< W ’ y Ct n + i-d(t n+i)))+Tn<6> t 1 - 38)

where Tn(<5) depends on the jump 6 and is such th a t the order o f the

method is preserved. In p a rtic u la rs Zverkina considers the use o f Adams-

Bashforth methods o f orders 2 S 3S and 4. Numerical re s u lts fo r example (1)

are given fo r several m odified methods (1.38) and also f o r unmodified

methods, which show the s u p e rio rity o f the m odified methods. Hutchison

(1971) has extended the work o f Zverkina. He proved the convergence o f

the m odified m u ltis te p method (1 .3 8 ). Hutchison comments on the p ra c tic a l


15.

com plexity o f c a lc u la tin g the m o dificatio n T ( 6) to deal w ith jump

d is c o n tin u itie s , and to s im p lify i t he considers re w ritin g the method

(1.38) in terms o f backward differences,, and then m odifying the backward

d iffe re n ce s fo r jump d is c o n tin u itie s in the d e riv a tiv e a t p o in ts in the

in te rv a l o f in te g ra tio n .

Hutchison (1971) also studies a m odified two p o in t method o f the

form

+ y ( k ) ( t n) + «„]

where <$n denotes the jump in the kth d e riv a tiv e o f u ( t) at t . Method

(1.39) is im p lic it and Hutchison suggests various p re d ic to rs f o r use w ith

it. The methods (1.38) and (1.39) carr not be applied i f the p o ints o f

jump d is c o n tin u ity are not known, or i f the problem is s in g u la r as defined

by E l's g o lt 's in ,s e c tio n (1 .4 ).

F e ldstein and Goodman (1973) present a study o f the e ffe c t o f

d is c o n tin u itie s in the s o lu tio n or it s d e riv a tiv e s on the numerical

s o lu tio n , and they conclude th a t to preserve the order o f the numerical

method one should have the points o f d is c o n tin u ity o f the s o lu tio n and i t s

fir s t p+1 d e riv a tiv e s , where p is the order o f the method, as mesh

p o in ts . In another work, Goodman and Feldstein (1973) present a study o f

the propagation o f Round-off e rro r when solving the DDE (1 .1) by E u le r's

method (1 .2 4 ). They obtain by lin e a riz a tio n o f the DDE an estim ate and

bound fo r the accumulated ro u n d -o ff e rro r. Neves and F e ldstein (1973)

give a study o f the ch a ra c te riz a tio n o f the. points o f jump d is c o n tin u itie s

fo r DDE (1 .1 ).

Fox e t al (1971) consider the numerical s o lu tio n o f example (2) using

Chebyshev s e rie s , and give a very d e ta ile d discussion. They also consider

solving example ( 1 . 2) using the trapezium -rule coupled w ith an in te rp o la tio n

polynomial f o r evaluating the delay term. They use lin e a r and Hermite
16.

in te rp o la tio n and show the lack o f smoothness re s u ltin g from lin e a r

in te rp o la tio n .

F in a lly we mention some more recent Algorithms suggested f o r solving

DDEs. Opplstmp (1973) presented a Fortran program fo r solving DDE (1 .1)

w ith a delay o f the form d (t). His algorithm uses a fo u rth order e x p lic it

Runge-Kutta method w ith e ith e r Hermite in te rp o la tio n or Lagrange in te rp o la tio n

fo r evaluating the delay term y (t-d (t)) w ith the r e s tr ic tio n th a t the

stepsize is always less than the delay d (t) on the in te rv a l [t , t


n n+ i
For estim ating the LTE he uses the technique o f halvin g the step, and

comparing the so lu tio n s a t t +-j when using a stepsize o f h and two


1
steps o f 2h -

Neves (1975) presented a Fortran program which uses the Kutta-Merson

method fo r so lvin g DDE (1 .1 ). For evaluating the delay term y (t-d (t,y (t))),

his ro u tin e c a lls another ro u tin e which uses Hermite in te rp o la tio n i f

h < d ( t, u ( t ) ) , otherwise the in te rp o la tio n ro u tin e c a lls an e x tra p o la tio n

ro u tin e based on T a v e rn in i's fo u rth order method (1.35) to evaluate the

delay term. We note here th a t th is way o f evaluating each delay term may

need a la rg e number o f fu n c tio n evaluations, po ssibly as many as 15 per

step f o r evaluatin g each delay term. Also, Neves' program fo r solving

DDE cannot handle d ir e c tly a system o f DDE w ith m u ltip le delays. He

describes a way to overcome th is by augmenting the system o f DDE before

so lv in g .

1.5 S t a b ilit y an alysis o f numerical methods fo r solving DDE.

The question o f the s t a b ilit y o f numerical methods fo r solving DDEs

was f i r s t discussed by Brayton and Willoughby (1967). They considered

the system o f lin e a r neutral d iff e r e n tia l equations

j j 1( t ) + Au‘ ( t- d ) + B u(t) + Cu^(t-d) = 0 (1.40)

where A, B, and C are symmetric nxn m atrices, d is a constant,

ju ( t) is a vector o f dimension n. Equation (1.40) is a DDE i f A = 0.

Brayton and Willoughby proved th a t u jt) 0 as t -*■ «» i f I ± A and


17.

B ± C are p o s itiv e d e fin ite . The method they use is a f i n i t e d iffe re n c e


approximation o f the form

y (tn+1> - x (tn) = -A[y(tn. m+1). ♦

~ + ( W ) y ( t n)]

“ hC[v y ( tn_m+1) + ( i “ u ) y ( tn-m) ] ( i. 4 i)

where t ^ = ^ th , h " d/m, ^or some in te g e r m, and p is a constant

in [0 3 1 ]. Then., they show th a t y ( t n) -> 0 as t « provided th a t

(I ± A ), B + C, and (I ± A) - ( | - y)h(B ± C) are p o s itiv e d e fin ite .

Brayton and Willoughby also show by an example th a t i f the terms

in v o lv in g A and C are tre a te d as fo rc in g terms, so th a t the method

becomes

= • hB[>Ji ( ’t n+i') + ( 1 - u ) l( t n)]+ h F (tn) (1.42)

where F is a known fu n c tio n o f t n, th is can lead to an in c o rre c t r e s u lt.

W iederholt (1970, 1976) considers the s t a b ilit y o f several numerical

methods when applied to the lin e a r DDE,

u ( t) = a u (t) + b u ( t - l ) ( 1 .43)

where a and b are real constants. W iederholt defines the s t a b i l it y

region R(a, b) o f the DDE such th a t fo r a ll a, b e R(a, b) the s o lu tio n

u ( t) -> 0as t ^ oo. Ire ttin g h = 1/m, where m is an in te g e r,

W iederholt applies various numerical methods to the DDE (1.43) and gets

the s t a b ilit y polynom ials, then, fo r m = 1, 2, and 3, he determines

num erically the region R Jha, hb) in the (ha, hb)-plane in which a ll the

roots o f the s t a b i lit y polynomial are less than one. He then compares

Rm(ha, hb) w ith the s t a b ilit y region R(a, b) o f the DDE (1 .4 3 ). His

analysis is c a rrie d out fo r the Euler method, and fo r second and th ir d

order lin e a r m u lti step methods. His re s u lts show a change in the s t a b ilit y

region as one moves away from the ha-axis.

Cryer (1973, 1974) adopts a d e fin itio n s im ila r to the A - s ta b ilit y

d e fin itio n (1 .5 ) fo r ODE. He suggests as a te s t equation, the lin e a r DDE


18.

u1( t ) = b u (t-d ) t 5 0 (1.44)

u (t) = g ( t) -d <' t <: 0

where d is a p o s itiv e constant delay, g ( t) is a given continuous fu n c tio n ,

and b is a real constant. Since the asymptotic s t a b i l i t y properties o f

equation (1.44) are known, Bellman and Cook (1963), Cryer defines a lin e a r

m u ltistep method to be DAQ-sta b le i f the numerical s o lu tio n y ( t n) ->0 as

n -*• » f o r a l l values o f b fo r which the solu tio n o f (1.44) is

asym ptotica lly s ta b le . Cryer remarks th a t he has not considered equation

(1.44) w ith b complex, becausel i t t l e is known to date about the asymptotic

behaviour in t h is case. Cryer introduces several re la te d concepts and

analyses some s p e c ific numerical methods. He shows, via a long and elaborate

proof, th a t the trapezium method is DA0-s ta b le .

BarweYl (1974) fin d s the asymptotic s t a b i l i t y region R(b) o f equation

(1 .44), w ith b a complex constant. Then Barwell (1975) defines the lin e a r

m u ltistep method to be Q-stable i f the numerical, s o lu tio n y ( t ) ->- 0 as

n a. f o r a l l b e R(b). He shows th a t the backward Euler method is

Q-stable.

Barwell (1975) also considers the te s t equation

u ‘ ( t ) = a u (t) + b u (t-d ), t > 0, d > 0


(1.45)
u ( t) = g ( t ) , t E [-d , 0]

where a and b are complex. He proved, in Barwell (1974), th a t there

e x is ts a set o f values R(a, b) such th a t f o r any a, b e R(a, b) the

s o lu tio n o f (1.45) is asym ptotically stable. Using t h i s r e s u lt , he defines

the lin e a r multi step method to be P-stable i f the numerical so lu tio n

y ( t n) -*■ 0 as n « fo r a ll a, b e R(a, b). Then, he shows th a t the

f i r s t and second backward d i f f e r e n t i a t i o n methods are P -stable.

1 .6 Remarks

I t is clea r from the discussion in sections (1.4) and (1.5) th a t

there- is a need f o r d e tailed study o f the adaptation o f standard numerical

methods fo r solving ODEs to the problem o f DDEs. So we shall consider


19*
d if f e r e n t numerical methods and show how to approximate the delay term, and

the e f f e c t o f t h is approximation on the order o f the method, on the loca l

tru n ca tio n e r r o r , on the LTE estimate, and on the s t a b i l i t y properties of

the method. Then we give a description o f Algorithms f o r solving a system

o f DDE w ith m u ltip le delays and testiWmon a set o f examples.


CHAPTER I I

THEORY OF DELAY-DIFFERENTIAL EQUATIONS

2.1 In tro d u c tio n

In th is chapter we consider the problem o f solving DDEs from the

a n a ly tic a l p o in t o f view. We s t a r t by giving a short h is t o r ic a l review o f

the work done on the theory o f DDEs, then we state theorems spe cifying

conditions f o r the existence and uniqueness o f the so lu tio n o f DDEs.

Afterwards we discuss one of the c h a ra c te ris tic s o f DDEs, th a t the various

de rivative s o f the s o lu tio n may have lag-induced jump d is c o n tin u itie s .

F in a lly we consider a n a ly tic a l methods f o r solving a system o f lin e a r DDEs,

and give re s u lts concerning the asymptotic properties o f the s o lu tio n .

2.2 Existence and Uniqueness o f the so lu tio n o f DDE

The general form o f a DDE, as defined in chapter I , is

u '( t) = f { t , u (t), u (a (t, u ( t) ) ) ) , t 0 $ t $ T (2.1a)

u ( t) = g ( t ) , m in (t0, t * ) < t < t 0 (2.1b)

where t * = m in {a (t, u ( t ) ) > , f o r t e [ t QS T ], and a (t,u (t))= t-d (t,u (t))$ t

fo r t e [ t 0, T] (2.1c)

As defined e a r l i e r , d ( t , u ( t ) ) is called the delay and a ( t , u ( t )) is


21.

calle d the lag.

A fu n c tio n u ( t) is said to be a solu tio n to equation (2.1) provided

(i) u ( t) is a continuous extension o f g ( t) on [ t 0, T ],

( i i ) u '( t) s a tis fie s (2.1a) f o r t e [ t 0, T ], where the r i g h t hand

d e riv a tiv e is used a t t - t 0. -

Special cases o f the DDE (2.1) have received much non-numerical

a tte n tio n in the past few decades. See E l 's g o l 't s and Norkin (1973), Hale

(1971), Hajaviay and Yorkee (1971), Myshkis (195.1), Myshkis and E l ' s g o l ' t s

(1967), and Zverkin, e t al (1962), f o r extensive bib lio g ra p h ie s and

comprehensive surveys o f the l i t e r a t u r e . Also the general case o f (2.1)

received considerable a tte n tio n from D river (1961), Hale (1971), and

Halaijay and Yorkee (1971).

Before we state the existence and uniqueness re s u lts o f D river (1961)

we present the fo llo w in g d e f i n i t i o n and a remark.

Definition 2.1

(1) A fun ction W(t) is said to be L ip s c h itz continuous f o r

t e [a, bl i f there e x is ts a constant L 5 0 .su c h .th a t f o r any

t l9 t 2 e [a, b]

|W(tj) - W(t2) U L ltj - t 2 | (2.2)

(2) A fun c tio n V(x, y , z) is said to be uniform ly L ip s ch itz

continuous in i t s second and t h i r d arguments on [a 13 b x ] * [a2 , b2] x

[a3, b3] i f there e x is ts a constant L 5 0 such th a t i f y x , y 2 e [a2 , b2l

and i f - z 2 e [a 3, b3] then

IV(x, yx, zj) - V(x, y2, z2)l s H| yx - y2l +|zx - z2|}

f o r any x e [a i, b j l .

Remark 2.1

Suppose we have a ( t 0, g( t 0)) < t Q, a s t r i c t in e q u a lity . Then by

c o n tin u ity o f a ( t , u ( t ) ) , we can r e s t r i c t ourselves to a domain about

( t Q, g ( t Q)) in which a (t, u ( t ) ) < t Q. Then the fu n c tio n on the r i g h t

hand side o f the DDE (2.1a) becomes a known fun ction o f (t, u (t)). So,
22.

the theorems o f local existence and uniqueness o f the solu tio n o f ordinary

d i f f e r e n t i a l equations are applicable in th is case.

Theorem 2.1 Extended Existence, (D river (1961)).

Let f ( x , y , z) be defined and continuous over domain

D O [ t o9 T] x C [tQ, T] x C [tQ, T ], and l e t a ( t , u(t )) be defined and

continuous over a domain D*s> [ t Q, T] x C [t0, T ], l e t t * £ a ( t, u (t)) $ t

fo r ( t , u ( t ) ) e D*, and l e t g ( t) e C [t* , t Q] . Then there is a number

h > 0 such th a t a t le a s t one s o lu tio n u( t ) o f (2.1) e x is ts f o r

t* $ t < t Q + h. Moreover, the s o lu tio n can be extended to t * $ t < T,

where t Tx » and f o r any compact set FC D there is asequence of

numbers t Q < t x < t 2 < . . . -> T, such th a t

( t -, u ( t.j) , u (a (t. 9 u ( t . ) ) ) ) e D-F f o r i = 1, 2 ..........

Thus i f T is f i n i t e and cannot be increased, one may conclude th a t as

t T-0, one o f the fo llo w in g occurs:

(i) Tim sup{ | u ( t) | , [u ( a ( t , u { t ) ) ) | } = » ,

(ii) ( t, u (t), u (a (t, u (t)))) comes a r b i t r a r i l y close to the boundary

of D.

Remark 2.2

One cannot assert, in general, th a t u ( t) approaches the boundary o f

D as t -* T, as would be the case f o r ODE. See Myshkis (1951) f o r an

example. D river (1961) points out th a t ( i ) o f theorem (2.1) can be

eliminated i f f is bounded on i t s domain. Theorem (2.1) also holds fo r

t * - - « 9 i f we replace [ t * , t ] , by (t*, t ] where applicab le .

Theorem~2.2 Extended Uniqueness, (D riv e r (1961)).

Consider the DDE (2 .1 ). Let g ( t) be Lip sc h itz continuous on each

f i n i t e in te rv a l o f [ t * , t 0] , l e t f ( t , u ( t ) , u ( a ( t, u ( t ) ) ) ) be uniformly

L ip s ch itz continuous in i t s 2nd and 3rd arguments in i t s domain D. Let

a ( t , u (t)) be L ip s ch itz continuous i n ' i t s 2nd argument on

[ t 0, T] x {range o f u ( t ) [ t e [ t Q, T ] } . Then the s o lu tio n to the DDE (2.1)

is unique over any in te rv a l where a solu tio n e x is ts .


23.

Remark 2.3

The condition th a t g ( t) be L ip schitz continuous i s not necessary f o r

delays independent o f the s o lu tio n . E l 's g o l 't s and Norkin (1973).

F in a lly , we mention th a t Driver (1961) considers a more general DDE

than (2 .1 ). He considers a system o f DDEs with m u ltip le delays. Also, he

gives re s u lts concerning the dependence o f the s o lu tio n o f (2.1) onthe

i n i t i a l data and on the r i g h t hand side of (2.1a).

2.3 D isc o n tin u ity in the d e riv a tiv e o f the so lu tio n o f DDE

A n a tu r a lly occurring phenomenon in the s o lu tio n o f the DDE (2.1) is

a jump d is c o n tin u ity in the d e riv a tiv e . These d is c o n tin u itie s can occur

even i f f , a, and g are a n a ly tic in t h e i r arguments, as in example (1 .1 ).


J
In general, these jumps occur because the i n i t i a l fun c tio n does not

necessarily s a tis fy the d i f f e r e n t i a l equation. I f there is some le a s t

in te g e r j > 0 such th a t g ^ ( t Q - 0) $ u ^ ( t Q + 0) then g is calle d

j-in c o m p a tib le . T y p ic a lly , the i n i t i a l jumps a t t Q propagatecausing

ad d itio n a l d is c o n tin u itie s whose locations depend on the (unknown) s o lu tio n .

Consider, f o r example, a 1-incompatible i n i t i a l fu n c tio n , and suppose

the value o f the jump in the f i r s t d e riv a tiv e is

<5 = u 1( t Q + 0) - g ' ( t Q - 0).

Let a ( t , u ( t ) ) < t , then d if f e r e n t ia t in g (2.1a) y ie ld s

U<2>(t) = f 1 + f 2 . u ' ( t ) + f 3 .U1( a ( t , u ( t ) ) ) . a ^

where f^ = f . ( t , u ( t ) , u ( a ( t, u ( t ) ) ) ) the p a r tia l d e riv a tiv e w ith respect

to the i t h argument.

a (]) = ^ ( a ( t , U ( t» ) .

Then, a t some p o in t t 1 > t Q, where a ( t x, u ( t x)) = t 0 , we have .

+ 0) - - 0) = f g . S . J O ( 2.5)

a jump in the value o f the second d e riv a tiv e .

In general the i n i t i a l jumps in the d e riv a tiv e a t t0 can produce


24.

other jumps at subsequent points t^ in [ t Q, T l, defined by

t^ = m in{t > t ^ _ ^ | u ( t 3 u ( t ) ) = t^ f o r some j s a tis fy in g 0 ^ j $ i - 1 } (2.6)

To f in d the value o f higher d e riv a tiv e jumps a t t. one must obtain

approximations to values o f higher order d e riv a tiv e jumps a t e a r l i e r points

and develop in c re a sin g ly elaborate recursion formulae. Even i f the

evaluation o f jump d is c o n tin u itie s is possible, in general i t i s not possible

to obtain a p r io r i knowledge o f the exact loca tion o f the jump points

w ithout knowing the true s o lu tio n .

A recent study by Neves and Feldstein (1973) o f the c h a ra cte riza tio n

o f jump d is c o n tin u itie s leads to the fo llo w in g theorem,

Theorem 2.3

Consider the DDE (2 .1 ), which s a tis fie s the hypotheses o f theorem

(2 .2 ). Suppose a has i t s f i r s t p ^ 1 and f has i t s f i r s t p-1

p a r tia l de riv a tive s continuous over the appropriate domains. Let

g ( t) s C ^ ^ t * , t Q
j . Let t ^ e [ t Q, T) and k be an in te g e r in [1 , p ],

such th a t u ^ (t) is continuous at t = t fo r 0 ^ q ^ k-1 . Let


X#"" I
u ^ (t) be continuous on the in te rv a ls ) anc*

(t ^9 t -J + £] f o r some £ > 0. Assume there e x is ts a le a s t number

t e( t -j 9T) such th a t t is a zero o f integer m u l t i p l i c i t y m of

„ a ( t,u ( t) ) - t -= 0. Then u ^ (t) is continuous a t t - t fo r a ll q


1 &
where

(i) 0s q $ p if m is even,

( ii) 0 $ q $ min(p, m.k), i f m is odd.

Remark 2.4
/ '
Even i f k is maximal, u ( t) may have more continuous d e riv a tie s than

guaranteed by theorem (2 .3 ). If k is the lowest order d e riv a tiv e o f

u ( t) w ith a jump d is c o n tin u ity at t = t , , then theorem (2.3) implies


Xr“ *
th a t u ^ (t) is continuous at t = t , th a t i s , k+1 is the lowest order
X*

d e riv a tiv e o f u ( t) th a t can possibly have a jump d is c o n tin u ity at t = t .

If a ( t , u (t)) is s t r i c t l y increasing in t then t h is process is c a lle d

smoothing. If a is not s t r i c t l y increasing, smoothing need not occur.


25.

When a i s dependent on u ( t ) , the degree of smoothness and the loca tion o f

d e riv a tiv e jump points cannot be determined w ithout knowing the solution

u(t).

2.4 Asymptotic properties o f the solution o f a lin e a r system o f DDEs

Here we consider a system of lin e a r DDEs o f the form

JJ' ( t ) = Au(t) + B u (t-d ), t $ t (2.7)

t) = g ( t ) , -d s t $ t

where d is a p o s itiv e constant, A and B are constant nxn real

matrices, and _u is an n-dimensional vector.

One o f the basic a n a ly tic a l methods f o r fin d in g thesolu tion o f the

DDE (2.7) is "the continuation method" or "the method o f steps", see

Bellman and Cook (1963), E l 's g o l 't s and Norkin (1973). The idea o f the

continuation method is based on extending the s olu tion from in te rv a l to

in t e r v a l. For the in te rv a l [ t 0, t 0+d), the DDE (2.7) y ie ld s

u '( t) = Au(t) + B g (t-d ), t e [ t n,t n+d)


" ( 2.8)
U(t0) =l ( V

which is an i n i t i a l value ordinary d i f f e r e n t ia l equation; then by in te g ra tin g

(2.8) we can f in d the solu tio n in the form

u ( t) “ I j W te [ t 0, t 0+d) (2.9)

By re q u irin g the solu tio n o f the DDE to be continuous fo r t ^ t Q,

s u b s titu tin g (2.9) in (2.7) y ie ld s

u‘ ( t ) = Au(t) + B g .( t - d ) , t e [ t n+d, t n+2d)


“ o o
u (t„+d) =5x(t0+d).

and by solving (2.10) we can extend the solution to the next in te rv a l

[ t 0+d, t 0+2d) and so on, the solu tio n can thus be found fo r any required

f i n i t e value o f t. A de tailed in v e s tig a tio n o f the continuation method

f o r solving (2.7) is given in Tsoi (1975).

The continuation method is not helpful when one wants to study the
26.

behaviour o f the s o lu tio n f o r t i n d e f in it e ly larg e. A second fundamental

method f o r fin d in g the solu tio n o f (2 .7 ), which uses an approach s im ila r to

the one used in ODEs, is to b u ild up the s olu tion as a sum o f simple

exponential terms. For the system of lin e a r ODEs o f the form

y 1( t ) = A y(t) t 5 t (2.11a)

l ( t 0) (2.11b)

where K is a constant vector o f dimension n, we assume the s o lu tio n o f


Xt#
(2.11) to be o f the form y ( t ) = ce , where Ais a constant and _c an

n-dimensional constant vector. S u b stitu tin g y ( t ) in (2.11) y ie ld s

(A - A l ) y ( t ) 0 (2.12)

then y (t) is a n o n - tr iv ia l s o lu tio n of (2.11) i f and only i f A is a

ro o t o f the c h a ra c te ris tic equation

det(A - Al ) = 0 (2.13)

which has n roots x , A2 , . . . , x . So any s o lu tio n o f the form


A •t
y ( t ) = c^e , where is a constant vector, is a s o lu tio n o f (2.11).

Also, since the sum o f two solutions is a s olu tion o f (2 .1 1 ), assuming th a t

a l l the {A^} are d i f f e r e n t , the general solution o f (2.11) is

n A• t
lit) = .1 £ i e

By adopting the same approach fo r solving (2 .7 ) , we assume a solution

o f the form

u(t) = c es t , (2.15)

where s is a constant, can n-dimensional constant vector. S u b stitu tin g

jj(t) in (2.7) y ie ld s

(Is - A - Be'sd)u(t) = 0 (2.16)

Hence, jj(t) =£ e is a solution of (2.7) i f and only i f the number s

is a zero o f the transcendental function

H(s) = det(Is - A - Be'sd) (2.17)


Equation (2.17) is called the c h a ra c te ris tic equation o f ( 2 .7 ) 9 and i f sr

is a zero o f (2.17) 9 then s^ is called a c h a ra c te ris tic ro o t. To

summarize the r e s u l t s 9 we state the follo w in g theorems which is given in

Bellman and Cook (1963).

Theorem 2.4

The equation (2.7) is s a tis fie d by


c ■{*
l £ r (t> e r ( 2 . 18 )
r

where { s ^ } 9 r = 19 2 9. . . 9 is any sequence o f c h a r a c te r is tic roots o f

(2 .1 7 )9 £ r ( t ) is a vector polynomial

of degree
^ -JUtn'k
n less than the m u l t i p l i c i t y o f s^s and the sum (2.18) is

e ith e r f i n i t e or i n f i n i t e 9 with suita b le conditions to ensure convergence.

Although the r e s u lt o f th is theorem is s im ila r to th a t fo r the lin e a r

ODE (2 .1 1 )s there is one very important diffe re n ce . As we shall see l a t e r 9

there are in general i n f i n i t e l y many c h a ra c te ris tic roots of (2 .1 7 )9 and

therefore i n f i n i t e l y many exponential solutions o f DDE ( 2 . 7 ) 9 whereas

there are only a f i n i t e number o f roots fo r the ODE (2 .11).

From the re s u lts o f the la s t theorem i t is c le a r ly necessary, to have

a good idea o f the lo c a tio n o f the zeros o f H(s). F i r s t we discuss the

d is t r ib u t io n o f the zeros o f the c h a ra c te ris tic equation o f a single lin e a r

DDE o f the form

u1( t ) - au ( t ) + bu (t-d )
(2.19)
u ( t) = g ( t ) 9 -d ^ t ^ 0.

The c h a ra c te ris tic equation has the form

h(s) - s-a-be c*s ( 2 . 20 )

Since we can w rite

h(.s) - s{1 +e( s) } - be~sd ( 2 . 21 )


28.

where e(s) ->- 0 as |s| » 5 i t is reasonable to suppose the zeros of

h(s) and the zeros of

TT(s) = s - be"sd (2.22)

are close together f o r |s| large. A c tu a lly a change o f va ria ble converts

one function in to the other with a d if f e r e n t b. Assume s' - s-a, then

(2.20) y ie ld s

h(s) = s' - b 'e 's ' d (2.23)


_ a
where b* = be . Equation (2.23) is s im ila r to (2.22).

The zeros o f the fun ction TT(s) s a tis fy the re la tio n s

|sesd| = | b | , |b| f 0 (2.24)

or

Re(s) + ~ £ n |s | = ^ n | b | (2.25)

So, the zeros o f h(s) lie asym ptotically along the curve defined in (2.25)

and are o f the form given in fig u re (2 .1 ). Using the asymptotic form

(2 .2 5 )a i t is shown in Bellman and Cook (1963)’ th a t the roots sr of

(2.20) have the fo llo w in g prope rties:

(i) In general, equation (2.20) has an i n f i n i t e number o f roots in the

complex plane.

(ii) A ll the complex roots s^ occur in conjugate p a irs .

( iii) A ll the roots s^ l i e in the l e f t h a lf plane Re(s) < c , . f o r some

constant c.

(iv ) For Is | s u f f i c i e n t l y large, the zeros are spacedalong the curve

in fig u r e (2 .1 ), a t an asymptotic distance o f 2ir/d apart.

(v) If bd e^+ad j- 1 then a l l the roots are simple.

Remark 2.5

Property ( i i i ) o f the roots sr is c h a ra c te ris tic o f DDEs with a

constant delay, and is used in Bellman and Cook as an a lte rn a tiv e

d e f in it io n o f DDE. Also they show, i f the deviation o f the arguments of


29.

Im (s )

Fig. 2.1 The asymptotic d is t r ib u t io n of the .roots o f equation (2 .22).

the advanced and neutral d if f e r e n t i a l equations9 given in equation (1.4)

and (1.5) re s p e c tiv e ly , are constant, then fo r a lin e a r advanced d i f f e r e n t i a l

equation, a l l the roots s^ o f i t s c h a ra c te ris tic equation l i e in the

h a lf plane

Re(s) > c, c is a constant,

and f o r a lin e a r n e u tr a l- d if f e r e n t ia l equation, a l l the roots o f i t s

c h a ra c te ris tic equation l i e asym ptotically along a v e r tic a l lin e

Re(s) = c, c is a constant.

For the c h a ra c te ris tic equation, H(s) = 0, in (2 .1 7 ), o f a system of

lin e a r DDEs, i t is shown in Bellman and Cook th a t the zeros o f H(s) appear

in "chains", each chain being o f the type encountered f o r h (s ).

Before we consider the concept o f s t a b i l i t y o f a DDE,.we have the


fo llo w in g remark.
30.

Remark 2.6

Using the lin e a r transformation t - t ' d 3 equation (2.7) becomes

j j ' ( t ' . d ) - A u ( t'. d ) + B_u( d(t 1 - 1))

Let t 1) = u ( t 1.d ) 9 then

v ' ( t ' ) = d .A v ( t ') + d . Bv ( t 1 - 1) (2.26)

So5 the asymptotic properties o f the solu tion o f (2.26) as t ‘ -> ® are

the same as f o r (2 .7 ). I f the solu tio n o f (2.7) is required on the in te rv a l

[ t Q9 T ] s then the s o lu tio n o f (2.26) is required on the in te rv a l [ t Q/ d 9 T /d ],

D e fin itio n 2.2

The DDE (2.7) is ,d e fin e d to be stable i f f o r any s u f f i c i e n t l y small

i n i t i a l fun c tio n the s o lu tio n jj(t) approaches zero as t approaches

i n f i n i t y 3 i . e . f o r a small constant 6 > 09 lim ||_ u(t)|| - 0 , f o r


t-Ko

IIJdl(t) 11 < S9 -d ,< t ,< t 0>

This type o f s t a b i l i t y is commonly referred to as asymptotic

s ta b ility . To f in d the conditions whether the DDE (2.7) is stable o r n o t9

we have the fo llo w in g r e s u lt .

Theorem 2.5

A necessary and s u f f ic i e n t condition f o r a l l continuous solutions of

equation (2.7) to approach zero as t ~ is th a t a l l the c h a ra c te ris tic

roots have negative real parts.

In a way s im ila r to the one used f o r a system of lin e a r ODEs, using

a s i m i l a r i t y transformations the m atrix A can be w ritte n in the form

A = S_1Da5 (2.27)

where D^ = d ia g(x1#. . . 9 X^), the {X^} are eigenvalues o f the m atrix Aa

assumed to be d i f f e r e n t , and S is an nxn m a trix 3 i t s columns being the

eigenvectors corresponding to the eigenvalues fx ^ }- We have two cases:

(i) If AB = BA.

Then the matrices A and B have the same set o f eigenvectors, and

the matrix B can be w ritte n as


31.

‘ B = S^DgS (2.28)

where Dg = d ia g(ul 3 . . . *!.yn) 3 are d iffe r e n t eigenvalues of B. So

the c h a ra c te ris tic equation has the form

H(s) = d e t(s l - A - Be~Sc*) = 0

n -sd
= n (s - A, - y.e iU ) = 0 (2.29)
i=l 1 1

since d e tfS "1).d et(S ) = 1. Equation (2.29) leads us to

s - X - uesd = 0 (2.30)

where x and y are particular eigenvalues of A and B respectively,

(ii) If AB ^ BA,

The DDE (2.7) yields .

ju'(t) = S_:LD„Sju(t) + Bu(t - d)

Su'(t) = DASu(t) + SBS-1Su(t - d)

Let y ( t) = S_u(t) and SBS”1 = [pi -1, and assume the solution has the
'J
form y ( t ) = c:es t or y ..(t) = c ^ e s t 0 since c = [c l9 c2 , . . . , cn] T . Then,

se.es£ = A .c.es t + £ B ..c .e st_st*


1 1 "i aZi ij J

. e rl
or s = xi + b^e :>u (2.31)

. • l / i j Cj
fo r i = I , . . . , n, where b. = ,c. ^0.
i c-. i '

From the re s u lts o f equations (2.30) and (2.31) the best model f o r

studying s t a b i l i t y is the DDE

u1( t ) = a u (t) + bu(t - d) (2.32)

where a and b are, in general, complex numbers. The c h a ra c te ris tic

equation o f (2.32) is

h(s) = s - a - be"sd (2.33)

We now give re s u lts which specify conditions on a and b in ($.32) such

th a t the roots o f h(s) = 0 have negative real p a rts .


X? /

Fig. 2.2 The asymptotic s t a b i l i t y region o f the DDE (2.19).

(1) a and b are re a l.

This case is discussed in Bellman and Cook and t h e i r re s u lt is

Theorem 2.6

A ll roots o f equation (2.33) have negative real parts i f and only i f

(i) a < 1

(ii) a < -b < ''e2 +a2

where 0 is the ro o t o f 9 = a tan(ed) such th a t 0< ed < tt. If a ~ 0

we take 0 =
/
Figure (2.2) i l l u s t r a t e s the region o f the ( a 3 b) plane in which

the roots o f ($,.33) have negative real parts. The region in which -a > |b[

is the region o f s t a b i l i t y which does not depend on the delay d, while

the curve -b = J 02 + a2is a function o f the delay d. Wenotice here if

d 0 the s t a b i l i t y region tends to the h a lf plane a+b < 0.

(2) a — 0 and b complex.

This case has been considered by Barwell (1974) and his re s u lt is


33.
Theorem 2.7
T (j)
Let b = re % then a s u f f ic i e n t condition th a t a l l theroots o f

(2.33) have negative real parts is

( i ) Re(b) < 0 ( | < 4, < 4p ),

( ii) 0 < rd < min( f 4


(3) a and b are complex.

This i s also considered by Barwell (1974) and his r e s u lt is

Theorem 2.8

A s u f f ic i e n t condition th a t a l l the roots o f (2.33) have negative

re al parts is

Re(a) < “ | b [ .
CHAPTER I I I

NUMERICAL METHODS FOR SOLVING DELAY-DIFFERENTIAL

EQUATIONS

3.1 Intro d u ctio n

We mentioned in chapter I th a t our main concern is the adaptation of

numerical methods f o r solving ordinary d i f f e r e n t i a l equations to solve

d e la y - d if fe r e n tia l equations (DDE). In th is chapter we s t a r t by considering

in general the d i f f i c u l t i e s a r is in g in the adaptation o f numerical methods

to solve DDE5 and discuss ways to overcome them. Thens we consider

s u ita b le d e fin itio n s fo r the s t a b i l i t y o f a numerical method f o r solving

DDE. F i n a l l y 9 we describe some o f the popular methods f o r solving ODE

which w i l l be considered in d e ta il in the next chapter f o r solving DDE.

3.2 Numerical d i f f i c u l t i e s in solving DDE

The general form o f a DDE is

u ' ( t ) = f ( t , u ( t ) , u ( t-d ( ta u ( t ) ) ) ) a t Q < t $ T (3.1)

u ( t) = g ( t ) , m in(t* 9 t Q) $ t $ t Q

where t * - m in lt - d ( t , u ( t )) ], t e [ t Q, T] and g ( t) is the i n i t i a l

fu n c tio n .
35.

Here9 we assume th a t the DDE (3.1) s a tis fie s the conditions o f

theorems (2 . 1) and ( 2. 2) which guarantee the existence o f a unique solu tion

U(t) £ T].

Comparing the DDE (3.1) w ith an 0DE9 we have an a d dition al term

u(t - d ( t 9 u (t))) which needs to be evaluated f o r some values o f

t e [ t n9 t -j ] and u (t). We w i l l re fe r to the term u ( t - d ( t 9 u ( t ) ) )

a:s, the delay term. The main d i f f i c u l t y in adapting numerical methods f o r

solving ODE to DDE is the evaluation o f the delay term. The other d i f f i c u l t y ,

discussed in chapter I I 9 is the d e riv a tiv e d is c o n tin u itie s at some points

in the range [ t Q9 T] f o r DDE w ith certain i n i t i a l conditions.

3.2.1 D is c o n tin u ity o f the d e riv a tiv e

In the d e riv a tio n o f a numerical method o f order p 9 we assume th a t

the solu tion u ( t) has continuous derivatives up to p+1.

If u '( t) is L ip sc h itz continuous, then

u ( t n + h) = u ( t n) + h u ' ( t n) + 0(h*) (3.2)

So9 i f the d is c o n tin u ity is in the d e riv a tiv e u ^ ( t ) , where k <. p+19

and t e (t , t + h) 9 there is no guarantee th a t the order o f the method is

preserved. But since k $ 1, and the function f ( t 9 u (t) 9 u (t-d (t9 u ( t))))

in the DDE (3.1) is Lip sc h itz continuous in i t s second and t h i r d arguments,

we get a t le a s t a f i r s t order approximation o f the s o lu tio n .

Because o f th is in h e rite d property o f d is c o n tin u ity in the kth

d e riv a tiv e , we consider the adaptation of one-step methods ra th e r than

m u ltistep methods f o r solving DDE, since i f the d is c o n tin u ity is a t the

p o in t t or t n+h then the order o f the method is preserved on the

in te rv a l ( t 9 t n+h) •

Hence, in using a one-step method fo r solving DDE (3.1) the best

strategy fo r preserving the order is to fin d the points o f d is c o n tin u ity in

advance9 then to use them as mesh po in ts . But t h i s stra teg y is not possible

in general because the lo c a tio n o f a jump d is c o n tin u ity is not known ap r i o r i

f o r the case where the delay i s dependent on the s o lu tio n . So the best we

can hope f o r in a numerical method is th a t the local tru n c a tio n e rro r


36.

estimate o f the method w i l l detect the occurrence o f a d e riv a tiv e

d is c o n tin u ity , and w i l l then decrease.the stepsize so th a t the point o f

d is c o n tin u ity becomes a mesh p o in t.

3.2.2 The evaluation o f the delay term

Following the discussion o f the la s t section, we shall consider the

adaptation o f one-step methods o f Runge-Kutta type, described in chapter I ,

f o r solving DDE. We assume th a t the solu tion is s u f f i c i e n t l y d iffe r e n tia b le

in any required in t e r v a l, and th a t the points o f d e riv a tiv e d is c o n tin u ity

are always a t mesh po ints.

Suppose we have solved the DDE (3.1) up to the p o in t t , and l e t the

numerical s o lu tio n be y (t). We want to advance the solu tio n to the p o in t

using the Runge-Kutta method (1.13). Each time we evaluate the

fu n c tio n , the delay term must also be evaluated a t some p o in t. In the

fo llo w in g , we f i r s t discuss ways o f approximating the delay term, and then

we consider in general the e ff e c t of approximating the delay term on the

lo ca l tru n c a tio n e rro r and the order o f the Runge-Kutta method (1 .13).

3.2.2.1 Ways o f evaluating the delay term

Assume th a t the delay term needs to be evaluated a t the point s,


where

s - t - d ( t , u ( t ) ) , f o r some t e [t * t -j] .

The lo c a tio n o f the p o int s depends on the size o f the delay as fo llo w s :

(1) Small delay.

i.e . s g ( t n, t^^-j ] (3.3)

Here the method has to use extra p o la tio n f o r the evaluation o f the

delay term at s. One method o f e xtrapo lation is to make use o f the

s o lu tio n values at e a r l i e r mesh po ints, but th is approach has two d is ­

advantages, f i r s t th a t we lose the s e lf - s t a r t in g property o f the one-step

method, e sp e c ia lly f o r i n i t i a l value DDE where the delay at the s ta r tin g

p o in t is zero. The other disadvantage is th a t the s o lu tio n might have a

discontinuous d e riv a tiv e a t some o f the mesh p o in ts, and the extrapolation
37.

from one in te rv a l to another w i l l not be v a lid in general.

(2) Large delay.

i .e . (3.4)

If s £ t Q9 where t0 is the i n i t i a l p o in t, we use the i n i t i a l

fu n c tio n to evaluate the delay term. Otherwise t Q < s $ t , le t

s = t . + rH, 0 ,< j %
< n-1 (3.5)

where H= t - t ., and 0 < r « 1,

and we make use o f the s o lu tio n values at the mesh points t ^ , where

0 £ i £ j , to approximate the delay term. Hence the need arises to store

the previous s o lu tio n values and possibly the values o f the d e riv a tiv e ,

depending on the method o f approximation, f o r a s u f f ic i e n t back distance

depending on the nature o f the delay and the range o f in te g ra tio n , but at

most the back distance needed at the point


r tn is tn ■ to
n,

Because the so lu tio n may have discontinuous d e riv a tive s a t some o f

the mesh p o in ts, i t is recommended th a t the method o f approximating the

delay term at the point s should use only inform ation in the in te rv a l

[ty Another reason f o r not using an approximation method which

needs inform ation on more than one in te rv a l is t h a t, i f j = 0 in (3.5)

then f o r i n i t i a l value DDE, the only available information is on [ t Q, t x].

3.2 .2.2 The order o f approximating the delay term

Assume th a t the delay term is approximated by z(s) such th a t

u(s) - z(s) = e(s) (3.6)

where e(s) is the loca l tru ncatio n e rro r in approximating z(s) at

the p o in t s, and the order o f approximating the delay term is

e(s) = 0(Hq+1) (3.7)

where q i s an in te g e r, H/h = 0(1), and h = t n+-j - t . The Runge-Kutta

method (1,13) adapted fo r solving DDE (3.1) becomes

y ( t n+1) = y ( t n) + N>(tn , y , h) (3.8)


38.

where

m
* { t , y , h) = I'c .k .
0=1 J J

and

m
. > A + ' l“ j ) * y ( t n, + hi J y i ( 3- 9a)

kj = f ( t n + haj ’ V S ) + haJ>-

z ( t n + haj ' d(tn + haj ’ yj ( t n + haj U ) ) (3.9b)

fo r .j = 1, 2, . . . , m.
Ct v / f i l S £ lt\A -i pC ^C C L'-^, 6-jS C ^C ovv c £& & <J •s-mP'CT cK^jy^L-cJC - t i < ^U . ,^>-©£c<Tur

Consider the e f f e c t o f approximating the delay term on the local

tru n ca tio n e rro r and the order o f the method. By s u b s titu tin g the tru e

s o lu tio n in both sides o f equation (3 .8 ), the local tru ncatio n e r ro r o f the

method (3.8) is defined by

LTE = u ( t n+1) - u ( t n ) - h<J>(tn , u, h) (3.10)

where

m
* ( t n. U, h) = f Ic j k . j
o=i
m
yj ( t n +haj> = U( V + h I bj i ki (3.11a)

k. = f ( t + ha., y . ( t „ + h a .),
j v n j n j ' 9

z ( t n + ha^ - d ( t n + ha^, y ..( tn + h a p ) ) ) (3.11b)

fo r j = 1, 2 , . . . , m.

For the pth order Runge-Kutta method(3 .8 ), i f the delay term is

evaluated e x a c tly , the LTE is given by

^ = u( t n+-j) - u ( t n) - h f ( t n, u, h) (3.12)

= # ( t n. u ( t n))h p+1 + 0(hp+2)

where ij>(t, u ( t ) ) is a fu n ctio n o f the solu tio n and i t s d e riv a tiv e s , and
39.
_ m
* ( t , u, h) = I c X
j= l

m
yj ( t n + haj> = u<t n^ + \ l * j f i (3.13a)

kj = + hajS y p t n + h a p ,

u ( t n + haj ' d ( t n + haj ’ yj ( t n + haj n ) ) (3.13b)

Let £■ = ¥ i - ki , i = 1 ,..., m (3.14)


__ m
(3.15)
Ej = T [.bj 1 Ei ’ 3 = 1s- ” » m

and Sj = t n + haj - d ( t n + ha j . y p t H + t a p ) . (3.16)

j *"* 1 3 * * * grn
Subtracting (3.11) from (3.13) y ie ld s

y 3( t n + haj> = y j ( t n + haj> + S (3.17a)

and Ej = f ( l n + ha., y p i n + h a p ,

u( t n + h3j - d ( t n+ h a j, y j ( t n + h a j) ) ) )

- f(tn + h a j, y j ( t n + h a j) ,

z ( t n + haj - d ( t n+ haj, y j ( t n + h a j) ) ) ) (3.17b)

Then

d ( t n + h a j, y j ( t n + h a j)) = d ( t n + h a j, y j ( t n + h a j))

+ hK, X .
I 9J J

where l< = 3d(*> u ( t ) )


19j au(t)

fo r c [ y j ( t p + h a j) , y j ( t n + haj) + hEj]

and

« ( t n + haj - d ( t n + haj, y j ( t n + h a j)))


(3.18)
- U<S.i> + h K2.3-E3

where K .K, •
2 ,j dt n 1 jJ
2 ,j

fo r
n2,j e [Sj s Sj ' hl<l , j Ej 3

Using the assumption (3.6) fo r approximating the delay term, equation

(3.18) becomes

u(tn + haj - d(tn + haj, y j ( t n + haj)))

= z(Sj) + e(Sj) +h.K 2jjFj (3.19)

Using the re s u lts o f (3 .19), equation (3.17b) y ie ld s

Ej " hK5 ,j f j + K4,j e( sP j 1 , . . . , m. (3.20)

where

K5,3 _ K3 ,j + K4 ,3 ’K2,3

K = 3 f(t» u ,z )
3»J 3u ^ n ^? ± ^ s ) * frts j')
3 9j

for -<n3 .1 e t y j ( V ha3 ’ y . ( t +hap + hE ]

and if = d f ( t ^u>z )
4,j " 9Z
^ 4 ,j > & ~ +"4.

fo r e £u( sj) > u( sj ) + £ ( sj ) +

Using (3 .1 5 ), equation (3.20) can be w ritte n in the form

(I - hKB)E = e (3.21)

where K = diag(Kg -j, K5,m^


41.

Assuming th a t (I-hKB) /a non-singular, which is always true fo r

s u f f i c i e n t l y small h, then w ith the assumption (3 .7 ), equations (3.21)

y ie ld :-

Ej = K4 s j . £ ( sj ) + ° ( hq+2) (3.23)

fo r j - 1, 2, . . . , m.

Subtracting (3.12) from (3.10) y ie ld s

LTE - LTE = h [ ^ ( t n » u, h) - <f>( t n3 u, h)]


m
= h I c.E. (3.24)
j= l J J

Then, using the re s u lts o f (3 .2 3 ), the LTE o f the method (3.8). is

LTE = t ( t n, u( tn))hP+1 + 0(hP+Z)

m ~
+ h.I cj K4 , j e( sj ) + ° ( h ) ^3' 25)
J

Equation (3.25) shows th a t the order of the method is min(p, q-f-1).

So in order to preserve the order o f themethod we must have q 5 p -1 .

If q > p-1 then approximating the back values has no e f f e c t on the order

o f the method, and the p rin c ip a l pa rt of the LTE remains the same. The

case q > p-1 is useful f o r the estimation o f LTE, since we do not need to

evaluate the e r ro r o f approximating the back values in order to get an

estimate o f the p rin c ip a l LTE o f the method.

3.3 D e fin itio n s o f s t a b i l i t y properties of numerical methods f o r solving

DDE

In chapter I the d e fin itio n s o f absolute s t a b i l i t y and A - s t a b i l it y o f

numerical methods f o r solving ODE are given, and i t i s mentioned th a t they

are based on the -test equation (1.16), a lin e a r ODE. Here we consider the

extension o f these d e fin itio n s fo r a numerical method adapted f o r solving

DDE.

Assume th a t a numerical method is applied to the DDE (3.1) with a

fix e d stepsize h, then the global e rro r is defined by


42.

en = Y ( t n) - u ( t n) (3.26)

where t = t Q + nh, y ( t ) is the numerical solu tio n and u (t) is the

.exact s o lu tio n a t t.

In the s t a b i l i t y analysis o f a numerical method one is concerned not

w ith the source o f the e rro r but only w ith the behaviour o f the global

e r ro r as t » a f t e r some e rro r has been introduced. Since the

behaviour o f the global e rro r (3.26) depends on the behaviour o f the s o lu tio n

of the DDE (3 .1 ), we adopt a model DDE in order to study the s t a b i l i t y

properties o f a numerical method. In chapter I I , i t is shown th a t a

representative lin e a r DDE is

u1( t ) - a u (t) + b u ( t - l ) , t $ t (3.27)

u ( t) = g (t) -1 i t s t 0

where a and b are complex, and g ( t) is a continuous fu n c tio n .

Since the d e f in it io n o f absolute s t a b i l i t y is only concerned with

the case where the s o lu tio n u ( t) s a tis fie s

u ( t) -*■ 0 as t -*• «, ' (3.28)

we need to know the asymptotic s t a b i l i t y properties o f the solu tio n o f

(3 .27). In chapter I I , we give recent re s u lts about conditions on a and

b such th a t the s o lu tio n s a tis fie s (3.28).

For a numerical method f o r solving. (3.27) we expect the global e rro r

en -*■ 0 as n » i f the solution s a tis fie s (3 .2 8 ), which leads us to adopt

the fo llo w in g d e f i n i t i o n ,

D e fin itio n 3.1

A numerical method applied to the DDE (3.27) is said to be absolutely

stable f o r the stepsize h ? i f f o r any problem whose so lu tio n s a tis fie s

(3.28) the numerical solu tio n at step h s a tis fie s y ( t n) 0 as t

I f the absolute s t a b i l i t y o f a method is independent o f h, then we

get the fo llo w in g d e f in it io n s im ila r to the A - s t a b i l i t y d e f in it io n f o r ODE.


43.

D e fin itio n 3.2

A numerical method is said to be DA-stable i f f o r any s o lu tio n o f

(3.27) which s a tis fie s (3.28)* the numerical so lu tio n y (t ) + 0 as

t f o r anij h> 0 .

The d e f i n i t i o n o f D A -s ta b ility depends on knowing the necessary and

s u f f ic i e n t conditions on a and b such th a t the s o lu tio n s a tis fie s

(3 .2 8 ). Special cases o f D A -s ta b ility are discussed here. By assuming:

mh = .1* m£ I+ (3.29)

Cryer (19733 1974) considers a d e f in it io n o f D A -s ta b ility f o r lin e a r

m u ltis te p methods when using the model equation (3.27) w ith a = 0 and

b r e a l. Later* Barwell (1975) generalizes Cryer's d e f i n i t i o n byconsidering

the model equation (3.27) w ith a = 0 and b- complex. His d e f in it io n is

adopted here f o r one-step methods*

D e fin itio n 3.3

Let b = re 1^ and a = 0 in (3.27,). A numerical method is said to

be Q-stable i f under the conditions

( i ) Re ( b) < 0

( i 1) 0 < r < mi n(4^ “ <i>» <t>- J)

the numerical s o lu tio n y ( t ) *> 0 as t -* «» f o r a l l h s a tis fy in g (3.29).

Barwell (1975)* a f t e r g e ttin g a s u f f ic ie n t con dition on a and b in

(3.27) such th a t u ( t) 0 as t -> °°* considers the fo llo w in g d e f in it io n

D e fin itio n 3.4

A numerical method* applied to (3.27) is said to be P-stable i f under

the condition

Re(a) < - |b|

the numerical so lu tio n y ( t n) ->0 as t ->■ « fo r a ll h s a tis fy in g (3.29).

I t is c le a r from the d e fin itio n s (3.3) and (3.4) th a t i f the method

is P-stable then i t is A-stable* but i f i t is Q-stable then i t is not

necessarily A -stable.
44.

For a d e f i n i t i o n o f the absolute s t a b i l i t y region, we introduce

the fo llo w in g :

D e fin itio n 3.5

For the stepsize h, then

(1) If a and b are real in (3.27), theregion RP(a, b) in the

a,b-plane is calle d the P - s t a b i l it y region i f fo r any a, b e RP(a, b) the

numerical s o lu tio n o f (3.27) s a t is f ie s y (t ) + 0 as t 00.

(2) If a = 0 and b is complex in (3 .27), the region RQ(b) in

the b-plane is c a lle d the Q - s t a b ilit y region i f f o r any b e RQ(b) the

numerical s o lu tio n y ( t n) ^ 0 as t ^

3.4 Some numerical methods fo r solving ODE

Consider the i n i t i a l value problem in ODE o f the form

u ' ( t ) = f ( t , u(t ) ) , t 0s< t s< T (3.30)

u ( t 0) - c, c constant,

where u (t) is continuous and s u f f i c i e n t l y smooth. One o f the popular

methods f o r fin d in g the numerical solution o f (3.30) is the e x p l i c i t fo u rth

order Kutta-Merson method. Suppose we have the s o lu tio n o f (3.30) up to

the point t 9 and l e t y (t) be the numerical s o lu tio n a t t , then to

advance the s o lu tio n to the p o int t ^ , the Kutta-Merson method uses the

fo llo w in g :

Let h = t n+1 - t n.

y i =y 0 n) + -3h f ( v y ( V ) »

yz = y(tn) +l hf( V y ^ p U + ihf(£n + yi)>

y 3 = y(tn)+ lhf( V y t y ) + | hf(tn + '3h> yz )>

yk = y ( t p) + | h f ( t n, y ( t n)) - | h f ( t n + -gh, y2)

+ 2hf(tn + y 3)»

y5 = y(£n) + ]jhf( y y(£n)) + |hf(tn + y 3)

+ ^ h f(tn + h, y k ) ,
45.

the so lu tio n is taken to be

y (V |):= y s (3.3i)

This method uses f iv e function evalutionsper step. By s u b s titu tin g

the tru e s o lu tio n u( t n) the formulae, the LTE is

LTE = u ( t n+1) - y ( t n+1)

= iK y U(tn))h5 + 0(h6) (3.32)

The fun c tio n ^ ( t , u (t)) is in general a non-linear fu n c tio n in u { t) and

i t s d e riv a tiv e s . We notice here th a t the Kutta-Merson method y ie ld s two

values o f the s o lu tio n at the point t -j, y 5 and y^. Suppose th a t the

method is applied to the lin e a r ODE

u ' ( t ) = a u (t) + bt + c (3.33)

where a, b, and c are constants. I f we s u b s titu te the true s o lu tio n

u ( t ) , then

u( t n+-|) - y H = ^ <t n =. u ( t n) , h) (3.34)

= m hSu(5)( V + °<h6)
and

u ( t n+d ' y 5 = u( V ) h5 + ° ( h6) ( 3- 35)

= 720h5u(5)< V + ° ( h6)

Then, f o r the estim ation o f the p rin c ip a l part o f the LTE, we use

ITF a —3 h5 u (^)(t )
720 u ' n'

“ ^•(ys - y , ) • (3.36)

The estimate o f the LTE given in (3.36) holds exa ctly i f the equation is

o f the form (3 .3 3 ). For a small in te rv a l [ t ^ , t n+-j ] , one can assume th a t

f ( t , u( t )) in general behaves lin e a r ly .

One o f the disadvantages o f the Kutta-Merson method is the smallness

o f the absolute s t a b i l i t y region. Suppose we apply the method to the

d i f f e r e n t i a l equation
46.

u1( t ) = a u (t) a < 05 (3.37)

then the stepsize must s a tis fy

I ha| 2.78 (3.38)

Condition (3.38) is very r e s t r i c t i v e fo r c e rta in classes o f ODE.

To overcome the r e s t r i c t i o n on the stepsize, one can use A-stable

methods. Examples o f such methods are:

(1) The trapezium method.

The trapezium method is a second order i m p l i c i t one-step method o f

the form

y ( t n+1) = y(tn) + £ [ f ( t n. y ( t n)) + f ( t n+1, y ( t n+1) ) ] (3.39)

The LTE o f the method is

LIE = - + 0(li1>) (3.40)

(2) The fo u rth order i m p l i c i t Runge-Kutta method.

The method is

IS ■ f ( t n. y ( t n))

k2 = f ( £n + y< V + -27(5ki + Sk2 - k 3 »

k 3 = f ( t n + hs y O n) + f ( k i + 4k2 + k 3))

y< W + | ( k l + 4k2 + k3) (3.41)

The LTE has the form

LTE = «2( t n. u(tn))h 5 + 0(h6) (3.42)

where ijJ2( t , u( t )) is in general a non-linear fun c tio n o f u ( t) and i t s

d e riv a tiv e s . I f the method is applied to the lin e a r d i f f e r e n t i a l equation

(3 .37), then the LTE w i l l have the form

(3.43)

The main d i f f i c u l t y in using the methods (3.39) and (3.41) is t h e i r

im p lic itn e s s , so th a t we have to solve a system o f n o n-line ar equations at

each tim e-step. For the LTE estimate, a technique o f halving the stepsize

is often used.
CHAPTER IV

THE ANALYSIS OF SPECIFIC NUMERICAL METHODS

FOR DELAY-DIFFERENTIAL EQUATIONS

4.1 In tro d u c tio n

We'gave in the la s t chapter examples o f three one-step methods f o r

solving ODE, and we also mentioned the d i f f i c u l t i e s th a t numerical methods

should cope w ith when adopted to solve DDE. In th is chapter we study ways

o f approximating the delay term and the e ffe c t o f th is approximation on

the properties o f each o f the numerical methods considered. We continue

to w rite u ( t) f o r the exact so lu tio n and y ( t ) f o r the numerical s o lu tio n

a t the p o in t t.

4.2 The adaptation o f numerical methods fo r solving DDE

Before we consider any numerical method we assume th a t the s o lu tio n

is s u f f i c i e n t l y smooth on the range [ t Q, T ]. Weassume th a t the solu tio n

is approximated a t the mesh points t x, t 2, . . . , t and needs to be advanced

to the p o int t +-| w ith the stepsize h = t +-j - t • In general the

stepsize is not constant unless otherwise stated.

I t is mentioned in the la s t chapter th a t the delay term needs to be

evaluated each time there is a fun ction evaluation. The evaluation o f the
delay term depends on whether i t f a l l s in the in te rv a l [ t n, t +-j ] or not.

For each o f the fo llo w in g methods, we consider ways o f approximating the

delay term f o r the two cases o f small and large delay,

4.2.1 Kutta-Merson method f o r solving DDE

We stated in equation (3.31) the Kutta-Merson method f o r solving ODE.

We now consider using th is method fo r solving the DDE (3 .1 ) .

4.2.1.1 Small delay

The delay term has to be evaluated a t some p o int s, where

s = t - d (t, u ( t )) , t&-.(tn, t n+1] (4.1)

= t + rh , 0 < r < 1
n

Here we use an extension o f the idea suggested by Tavernini (1971) and

described in chapter I . Using t h is approximation we f in d a continuous

approximation to the intermediate values o f the s o lu tio n y J9 y 2 , y 3, y 4,

and y 5 in (3 .31).

In the in te rv a l (*tn , t ] we assume

ra
u(t + rh). “ u(t ) + rh I w . ( r ) u ' ( t + b.h) (4.2)
i= l

where 0 < r s< ] s 0 ^ b, <b„ < . . . < b $ 1 9 and w . ( r ) , i = 1, 2 , . . . , m ,


■*- m i
are functions o f r.

Assuming th a t the solu tio n is s u f f ic i e n t l y smooth, then

u (tn + rh) = u( j ) ( t n) + -([qh^ u(q+1)( t J + 0(hq+2) (4.3)


3=0
and

q-l (b.h)J (b.h)q , ^


|J ( 1 + 1 ) (bi h' q
u' ( t n + ^h) = I T ,—
—u^J
u<J+1) ((ttn)) + ■
—“l y - u (q+1)( t p) +0(hq+1)
— (4.4)
j =0 j: U ^

i = 1 , 2 , . . .,Tn

By s u b s titu tin g (4.4) in (4.2) and comparing i t w ith ( 4 .3 ) , we get


are d i s t i n c t .

By s u ita b le choice o f the we can adapt th is approach to the

Kutta-Merson method in the case o f condition (4.1) as fo llo w s :

( 1) q = 1 ■

bi = 0

Then W j( r ) = i , W j(l) = ) >

y i ^ n + rh) - u ( t n) + rh u1(vt ny

■ ) (4.6)

and

u(t rh) - y i ( t n + rh) = i l bi ^ u( 2 ) ( t n) + 0 ( h 3) (4.7)


n

( 2) q = 2 -
1
b2 = 0, b
2 3*

Then w j ( r ) = 1 - 3r Wj( 3) - 2>


2 Wl ^ = 4

w2 ( r ) = i21
r s
W2^ = 2’ «2(?) = I

y 2^bn + rh) = u ( t n ) + r h [ ( l - | r ) u ' ( t n) + f r u ' ( t n + ^ h) ] (4.8)

and

u(tn + rh) - y2( t n + rh) = -gh3r2(r - ^)u^3h t n) + 0(h1*) (4.9)

(3) q = 3

bl = b2 = "3’ b3 = 2

Then w ^r) = 1 - | r + 2r2 , Wl ( l ) = i

w2( r ) =| r * 6l"2 , w2(1) = - |

w3( r ) “ 4r2 - 2r , w3( l ),= 2-

ylf( t n + r h ) = u ( t n) + r h [ ( l - ~ r + 2r z ) u ' ( t )
(4.10)
6r 2 ’I u 1f t + - * ■ i n*# - 1
50.
(4) q = 3

bl = b2 = ~ Z ’ b3 = ^

Then wx( r ) = 1 - -|r + | r 2s wx( l ) = ^


w2( r ) = 2r - | r 2, w2( l ) = |

w3( r ) = - ^ r + | r 2, w3( l ) = 1

y 5( t n + r h ) = U( V + r h [ ( 1 " f r + | rZ ) u' ( t n)

+ (2 r - ^ r 2 ) u' ( t n + I h ) + (- | r + § r 2) u ' ( t n + h )] (4.12)

and

u ( t n + rh) - y 5( t n + rh) = ^ r 2( r - l ) 2u ^ ( t n) + 0(h5) (4.13)

Assuming th a t the s o lu tio n value is known f o r t $ t whenever i t

is required in the evaluation o f the d e riv a tiv e f , the modified. Kutta-

Merson method has the form:

ki =f < V y ( V ’ y (tn " d ( t n’

y x( t n +r h ) = y (t n) + r h A i

k2 = f ( t n + ] h , y x( t n + lh ),

y x( t n + ^h - d ( t n + l h , y x( t n + l h ) ) ) )

y 2(t n + r h ) = y ( V + rh [(l - | r ) k x + | r k 2]

k3 = f ( t n + 3 h’ y 2( t n+ 3h >’

y 2( kn + "311 " d( kn + ’3b ’ y 2^ n + '3,in ) )

y 3( t n + rh) = y ( t n) + r h [(1 - | r ) k x + | r k 3]

^ ” ^^bn + "2^ 3 y 3^kn

• y s (t n + l h “ d(kn + -Zht y 3 ^ n + I hn ) )

Y ^ n + rh) = y ( t n) + r h [(1 - | r + 2 r2)k x

+ ( ^ r - 6 r2)k + (4r2 - 2 r)kl[]

ks =^ n + h * y -t('fcn + h) ’
51

y4( t n + h " d(tn + h> y4( t n + h ) )) )

ys( t n + rh) = y ( t n) + r h l 0 - | r + | r 2)k1

+ (2r - 3 ^ ) ^ + (- + -|r2) kg]

and then

y ( t „ + h) = y 5( t n + h)

= y(tn) + h[^kj + fk, + lk5] (4.14)

We notice th a t the modified Kutta-Merson method (4.14) becomes the standard

Kutta-Merson method i f the delay is zero.

The f i r s t question a ris in g is "what is the order o f the modified

Kutta-Merson method (4.14)?". To fin d the order we need the fo llo w in g

r e s u lt . Let y -(t) be an approximation to u ( t) a t the p o in t t,

and l e t

e. j( t ) = u ( t) - y n.( t ) (4.15)

then

d ( t , y x- ( t ) ) = d ( t , u ( t ) ) - Dx( 5i ) -ei ( t )

where D ,(e .) ■ »& * . " ( * ) > .


l^ V 3U(t) (i|=g.

fo r e t u ( t ) , u ( t) + e ^ t ) ] ,

y .(t - d (t, y1(t))) = y ^ t - d (t, u (t))) - D2 ( n i , E p . e i(t)

d y .(t)
where D2(r,i , 5.) =
n -’W *
1

for ni e [t - d (t, u (t)), t - d (t, u (t)) - D1 ( 5 i ) . e i ( t ) ]

and

y n* ( t - d ( t 9 y i ( t ) ) ) = u ( t d ( t9 u (t))) £ i ( t - d ( t s u ( t ) ))

- D2(T1i 5 ^ i ) ,ei (t )

Then
52

f ( t , y ^ t ) , y .( t - d (t9 yi (t))))

- f ( t 9 u( t ) 9 u( t - d ( t , u (t ) ) ) ) -=*• Ei ( t )

= u* ( t ) - E.. ( t ) (4.16)

where E ^ t ) = K ^ e ^ t ) + K2 > i . e . ( t - d ( t , u ( t ) ))

’ K2j ¥ V si (4.17)

K = 3 f ( t » u »z >
< ? 1sir e (u (t). U (t) + e.(t)]
1 9i 3U

S’, , ;
K = 3 f ( t 9U-,Z)
2 ,i ~ 3z

I f e [u ( t - d ( t, u ( t ) ) ) j. u ( t - d ( t 9 u ( t ) )) + ( t - d ( t , u ( t ) ))

By making use o f the r e s u lt (4.16) together with the re s u lts we had

e a r l i e r f o r the intermediate values o f the solu tion y ( t ) 9 we now fin d the

order o f the method. Assume th a t we have the true s o lu tio n and i t s

d e riv a tiv e a t any p o int t $ t , le t t = t + r h 9 then:

£i(t) - u ( t ) - y x( t )

= + o(h3) (4.18)

= 0(h2)

£2 ( t ) - u ( t ) - y 2( t )

=^ - ^ (3)(tn) - j ! rE I(tn+>)

+ 0 ( h1*)

= 0(h3)

e 3( t ) = u ( t ) - y 3( t )

= . ^)u < 3> ( tn) + 0(hlt) (4.19)

= 0 ( h3)

ek {t ) = u ( t ) - y (t)
+ r h ( | r - 6 r2)E2( t n + l h ) + rh(4r2 - 2r)E3( t n + £) + 0(h=)

= 0(h**-) (4.20)

es( t ) = u ( t) - y5( t )

(4.21)

Thus the order o f the LTE o f the modified Kutta-Merson method (4.14)
is

es ( t n + h) - 0(h*») (4.22)

The r e s u lt o f (4.22) shows th a t, in general, the modified Kutta-Merson

method (4.14) is a t h ir d order method while the standard Kutta-Merson method

is a fo u rth order method. To overcome t h is problem of g e ttin g lower order

because o f the delay term, we go back to equation (4.5) and make use o f i t

to derive the intermediate value y 0( t + rh) as fo llo w s :


3' n '
For q = 2

le t b1 = 0, b2 = g f 0,

y 3( t n + r h ) = u ( t n) + r h [(1 - ^ ) u ' ( t n) + ^ u ' ( t n + fs)]

and

u ( t n + rh) - y 3( t n + rh) = rh(-^J-^- - w2( r ) - ^ P- - ]u ( 3) ( t n)

+ 0(hl>)

= I r ^ r " + (4.23)

If 3 = -3 we get the same r e s u lt as in (4.8) and (4 .9 ). By l e t t i n g


54.
= - |r , then

M r) = }

w2( r )

y 3( t n + rh) = u ( tn) + r h [ l u ' ( t n) + | u ' ( t n + | r h ) ] (4.24)


and

u ( tn + rh) - y 3( t n + rh) = 0(h1+) (4.25)

Using equation (4.24) the modified Kutta-Merson formulae (4.14) have the
form,

k i = f < V y ( V ’ y ( t n ‘ d ( t n’ y ^ n ^ ) )

y 1( t n + rh) = y ( t n) + rh kj

k2 = f ( t n+ ^ h » y i ^ n + ■3h>>

^ i^ n - d<t n + 3h’ y x ^ n + -3h>)))

y 2( t n + rh) = y ( t n) + r h [(1 - | r ) k j + | r k 2]

k3 = + 3 h, y 2( t n +.-ih),

y a ^ n + ! h ' d ( t n + l h - y a ^ n + 3 h ))))

k 3 <r) = f ( t n + | r h , y 2( t n + § rh ),

y a ^ n + l rh ' d<t n + f r h ’ y 2(t n + f ^ ) ) ) )

y 3 (tn + rh> = y^n) + rh[l ki + | k3 (r))

kK = f ( t n + j )1’ + | h>>

y 3( t n + | h - d( t n + ^h, y s( t n + ^ h ) ) ) )

y 4( t n + rh) = y ( t n) + r h [(1 - | r + 2r2)k1

+ ( | r - 6r2) k 3 + (4r2 - 2r)kH]

k5 = f ( t n + h, y ^ ( t n + h),

y^n + h ■ d^ n + h’ y i.( t n +

y s l ^ + rh) = y ( t n) + r h [(1 - | r + | r z)kj


55.

+ ( 2r - + {- l r + § r 2)k5l

Then

y ( t „ + h):= y 5( t n + h) (4.27)

Using the same approach as above f o r the order o f (4.14) we can

show th a t the modified Kutta-Merson method (4.27) is o f fo u rth order. The

method (4.27) uses* in general* s ix function evaluations per step while

the standard Kutta-Merson method uses f iv e function evaluations per step.

The method (4.27) reduces to the standard Kutta-Merson method i f the delay
1
is zero* and then kgOg) - k 3.

4 .2 .1 .2 Large delay

Assume th a t the delay term needs to be evaluated at the p o int s,

where

s = t - d ( t , u ( t ) ) , and t Q < s g t n, t E [ t n, t n+1] (4.28)

Let s = t . + rHi 9 0 $ j < n, 0 < r s 1* and H. - t . , - t . .


J u j J +i j
Following the re s u lts o f section (3 .3 ), we consider methods of

approximating the delay term which are of t h ir d order or higher, and which

use inform ation on a single in te rv a l [t^ ,

(1) Kutta-Merson method with Hermite in te rp o la tio n of the t h i r d degree.

Hermite in te rp o la tio n o f the t h i r d degree uses the s o lu tio n and i t s

d e riv a tiv e at t. and t

(4.29)

where P1 = ( l - r ) 2( l + 2 r ) , P2 = r 2(3-2r)

P3 - r ( l - r ) 2, P4 = - r 2( l - r )

I f we s u b s titu te the true solu tio n and i t s d e riv a tive in (4.29) the LTE

o f Hermite in te rp o la tio n is

e(s) = u(s) - y (s )

(4.30)

One o f the a lte rn a tiv e s to Hermite in te rp o la tio n (4.29) is to use


equation (4.12), which y ie ld s

y5(s) = y(tj) + - | r + |i"2)y'(tj)

+ (2r - | r 2) y ' ( t j + (- ^r+ | r 2)y ■( tj+ 1 ) ] (4.31)

which has the same LTE as in (4.30). I t can be shown th a t equation (4.31)

is equivalent to Hermite in te rp o la tio n in (4.29), since


H. ,
y ( t j +1) = y ( t j ) + ~ f f [ y '( t j ) + 4 y '( t j + £H..) + y ' ( t J+1) ] (4.32)
1
By e lim in a tin g y ' ( t . + pH.) between (4.31) and (4 .32), we get equation
J ^ J

(4.29).

(2) Kutta-Merson method with modified Hermite in te rp o la tio n o f the fo u rth

degree.

To get the modified Hermite in te rp o la tio n o f the fourth degree, we need

more inform ation on the in te rv a l [ t . , t .,,1 than the s o lu tio n and i t s


J
d e riv a tiv e values a t t^ and t - +^. Suppose the d e riv a tiv e value is given

at t . + b H., where 0 < b < 1, and l e t


J J

u ( t . + rH.) * w0( r ) u ( t . + H.) + w4( r ) u ( t . )


J J J j h J (4.33)
+ Hj [ w ^ r j u ' f t j ) + w2( r ) u ' ( t j + bH^) + w3( r ) u ' ( t j + H . ) ]

where w ^ (r), i ~ 0, 1 , . . . , 4, are functions o f r.

Assuming th a t the solu tio n is s u f f i c i e n t l y d if f e r e n t ia b le , then s u b s titu tin g

Taylor expansions a t the p o int t . f o r u ( t.+ r H . ) , u ( t- + H .) , u '( t.+ b H . ) 9


J J J J J J O
and u '( t.+ H .) in equation (4.33) and comparing the c o e ffic ie n ts o f H^,
J J

q = 0, 1 , . . 4 , y ie ld s

w0( r ) + “ l . M =1

w„(r) + w ^ r ) + w2( r) + w3(r) =r

TjW0(r) + bw2(r) + w3(r) = -^r2 (4.34)

^w0( r ) + b2w2(r ) + w3( r ) =| r 3


1 1
■^w0( r ) + b3w2(r ) + w3(r ) =^

I t can be shown th a t the system o f lin e a r equations (4.34) is


1
inco nsisten t i f b =^ so we can not use the d e riv a tiv e value at
57.

tj + t0 99t a fou rth de9ree Hermite in te rp o la tio n . I f we choose a


value o f b f then there is a unique solution o f the system (4 . 34).

Any choice o f b ^ 7^ w i l l cost us one more fun ction evaluation per step.
•j
Let b = ^ then the solu tion o f (4.34) is

w0( r ) = - 6r 2 + 16r3 - 9r4

wi ( r ) = r - 2r 2 + r 3

w2(r ) = ' T r2 " T r3 + T rh ‘ (4 -35)

w3(r) = | r 2 - |r 3 + - 1^

w4( r ) = 1 + 6r 2 - 16r3 + 9r4

And the local tru n ca tio n e rro r in the in te rp o la tio n formula (4.33) is

e(s) = - | r [ i "5 - w0( r ) - -^-w2( r ) - 5w3( r ) ] u % j i ?^j i [ v , t it-+J (4 .3 6 )

To use the modified Hermite in te rp o la tio n in (4.33) w ith (4 .3 5 ), we

need to store the so lu tio n and i t s d e rivative a t t-,n and the d e riv a tiv e
•j J
at t . +-^H.sa t each step.
J J

4.2.2 The trapezium method f o r solving DDE

Assume th a t the so lu tio n is advanced to the p o in t t , and the

fu n c tio n value is known at the point t , then the trapezium method has the

form,

y ( t n+1) = y ( t n) + J [ [ f ( t n, y ( t n), y(t„ - d(tn, y ( t n) ) ) ) .

+ f(V l- y(W ’ y< V l ' d ( t n+1 * y < W )))] ( 4 -37)

So, as well as solving a non-linear equation to f i n d the s o lu tio n

we neec* to evaluate the delay term at the point

s ■ t n+1 - d ( t n+1, y ( t n+1)) (4.38)

which depends on the value of the s olu tion at t ^. Whatever the method

we usef o r solving the non-linear equation, we need to evaluate the delay

term each time there is a function evaluation.

Since the trapezium mebhodis o f the second order, then from the r e s u lt

of (3.25) we can use lin e a r in te rp o la tio n or higher f o r approximating the


58

delay term.

4.2.2.1 Trapezium mdtiodwith lin e a r in te rp o la tio n

The approximation o f the delay term by lin e a r in te rp o la tio n depends

on the size o f the delay.

(1) Small delay.

Let s = t + rh 9 t n < s $ 11 , 0 < r ,< 1 (4.39)

Here we take advantage o f the im plicitn ess o f the m e th o d b y making

use o f the so lu tio n values a t t and t ,, to obtain


n n+1

z(s) = ( l - r ) y ( t n) + r y ( t n+1) (4.40)

The loca l e rro r in t h is approximation is

e(s) = ^ r ( r l ) h 2u<Z)(en). 5n e [tn» t n+1 ] (4.41)

(2) Large delay.

Let s = t . + rH ., 0 N < r < 1, 0 $ j < n, and H. = t . , n - t . .


J 0 0 J+l J
The lin e a r in te rp o la tio n y ie ld s

z(s) - ( l - r ) y ( t j ) + r y ( t ^ +1) (4.42)

w ith

e(s) = | r ( r - l ) H ? u(2) (S,)

4 .2 .2 .2 The trapezium method w ith Hermite in te rp o la tio n

We now consider using Hermite in te rp o la tio n based on the values o f

the solu tion and i t s d e riv a tiv e at two po ints.

(1) Small delay.

Let s = t n + rh , 0 < r ,< 1, t n < s t ^

Assume th a t the s o lu tio n and i t s d e riva tiv e are known a t the points

tn and t .j5 then Hermite in te rp o la tio n y ie ld s

z(s) = P,y(tn) + P2y ( t n+1) + P3h y ' ( t n) + P^hy1( t n+1) ( 4 .4 3 )

where Pl9 P2, Pgs and P^ have the same form as in (4 .2 9 ). Using the

trapezium method to get y ' ( t +-j)s since

y '^ n + P = " + § y (t n+f ’ y'^ V ^4 ’ 44^


59.

s u b s titu tin g (4.44) in (4.43) y ie ld s

z(s) = ( P l - ZPk M t n ) + (P2 + ■2Plt) y ( t n+1)

+ h(P3 - Ptf) y ’ ( t n) (4.45)

and

e(s) = X h3u(3)(t n) +

(2) Large delay.

Let s = t . + rH - 9 0 ^ r < l s 0 ^ j < n 3 and H- = t . - t..


J J J J T* » J

The form o f Hermite in te rp o la tio n is the same as described f o r Kutta

Merson method in equation (4.29).

4 .2 .3 The I m p l ic i t Runge-Kutta method fo r solving DDE

We described a fo u rth order i m p l i c i t Runge-Kutta method f o r solving

ODE in equation (3 .4 1 ). Here we consider i t s a p p lic a tio n to the DDE (3.1)

The i m p l i c i t method has the form

l<! = f ( V y ( V ’ y ( t n ' d(V y(V >)>

k2 = + y (t n) 8k2 ~ k3)>

y (tn + |h - d ( t n + } h , y ( t n) + ^ ( 5 k j + 8k2- k3) ) ) ) (4.46)

k 3 = f ( t n + h ’ yCtp) + ^ ( k l + 4 k 2 + k 3)>

y ( t n + h - d ( t n + h, y ( t n) + ^ ( k j + 4k2 +k3) ) ) ) (4.47)

y ( t n+P = y(V + ¥ (ki + 4k2 + k3) (4 -48)

So as well as solving a system o f non-linear equations, we need to

evaluate the delay term a t the p o in ts,

s i = kn + I h " d^ n + I*1’ y ( t n) + ^ - ( 5 ^ + 8k2 - k3) )

and

s2 =\ + h ' d ( t n + h’ y(V + F(ki + 4k2 + k3n

each time we evaluate the functions in (4.45) and (4.47) re s p e c tiv e ly .

The evaluation o f the delay term depends on the lo ca tio n o f s1 and s2 .


60.

(1) Small delay.

Assume th a t the delay term needs-to be evaluated a t the p o int s,

where

s = s 1 or s29 and s = t + rh 9 0 < r < 1

Using the s o lu tio n and i t s d e riv a tiv e values at the points tn and t .js

Hermite in te rp o la tio n y ie ld s

z(s) = P1y ( t n) + P2y ( t n+1) + hP3k 1 + hP^kj (4.49)

where P , P2, pg9 and P^ are defined in (4.29). Using the solu tio n value

from (4 .4 8 )9 (4.49) becomes

z(s) = (Pj + P2) y ( t n) + h(P3 + l p 2)kx

+ |hP2k2 + h(P,, + ^P2)k3 (4.50)

and

e(s) = ^ 2(1 - r ) W 4 ) ( t n) + 0(h5).

(2) Large delay.

s = s. or s99 such th a t s = t . + rH ., 0 ^ r < l , 0 ^ j < n ,


1 ^ J J
and - t j+1 - t j .

Since the i m p l i c i t Runge-Kutta method (4.4$) and the Kutta-Merson

method are o f the same order., the method given in section (4 .2 .1 .2 ) f o r

approximating the delay term w ith large delay can be used here. Hence3

we can use Hermite in te rp o la tio n o f the t h i r d degree (4.29) or the modified

Hermite in te rp o la tio n o f the fo u rth degree (4.33) to approximate the delay

term at s.

4.3 Local tru ncatio n e r ro r estimate o f methods f o r solving DDE

We have considered in the la s t section the adaptation o f s p e c ific

numerical methods f o r solving DDE. Here we consider the e ff e c t o f the

m o dificatio n o f each o f the methods on i t s LTE estimate. A ll the analysis

is considered when the method is applied to the lin e a r DDE

u ' ( t ) = a u (t) + bu (t-d ) (4.51)


61.

where a, b, and d are constant.

The use o f the lin e a r DDE (4.51) is reasonable, since a general DDE

behaves li n e a r l y on a small in t e r v a l.

4.3.1 LTE estimate o f Kutta-Merson method fo r solving DDE

Since the form o f the LTE o f Kutta-Merson method depends on the

lo c a tio n o f the delay, we consider the two cases o f small andlarge delay

separately.

4.3.1.1 Small delay

We assume th a t d < which implies th a t we need to evaluate the

delay term in the in te rv a l [ t n, t -j ] f o r a l l the fu n c tio n evaluations

required by the modified Kutta-Merson method (4 .27).

Since we are interested in the e ffe c t o f approximating the delay term

on the LTE and i t s estimate, f o r convenience we only consider the analysis

f o r the DDE

u ' ( t ) a bu (t - d) (4.52)

In order to f i n d the LTE} we assume th a t we have the exact solu tio n

and i t s d e riv a tiv e at the p o in t t , then we f in d the numerical s olu tion

at ■tp+i w^en the method is applied to (4.52), and compare i t with the

Taylor series expansion o f the exact solution u ( t n+-j).

We s t a r t from the Taylor series expansion o f u ( t n+^ ) 9

“( V P = U( V + hu‘ ( V + ! h2u(2>< V + ! h3u(3)(t n)


+ ^ fh -u (4) { t n) + + 0(h6) (4.53)

Using the DDE (4.52) and the Taylor expansion, we get:

u ( 5 > ( t n ) = b5u ( t n - 5d)

= b5 u ( t n ) + Dx

u ^ ( t n ) = b4 u ( t n - 4d)

= b^(l - 4bd)u(tn) + D2

u ^ 3 ) ( t n ) = b 3u ( t n - 3d)
62

= b3(l - 3bd + J | b 2d2) u ( t ) + D3

u^2) ( t n) = b2u( t n " 2d)

= b2(l - 2bd + 4b2d2 - - y b 3d3) u ( t ) + D

and

u ^ p t n) = b u (tn - d)

= b ( l - bd + | b 2d2 - | b 3d3 + ^ V d * * ) u ( t n) + Ds (4.54)

where

Di = K ^ n -J d 1. f o r some n . e [ t n> V , ] , i = 1, 2 , . . . , 5 (4.55)

and K^.(n^)s i - 1, 2S. . , 9 5, is a fu n ctio n of the s o lu tio n and i t s

d e riv a tiv e evaluated at n^. Then (4.53) becomes

“( V P = u( V + bhu( bn) + b2h24 ' « )“(*„)

+ b3h3(-g- - a + ^a2) u ( t )

+ b^h1* ^ - + 2a2 - ^ a 3 ) u ( t ^ )

+ b 5h 5 ( y j Q - + |a 2 - — a3 + -^ a '*)u (tn)

+ D + 0(h6) (4.56)

where <* = £ ( 4 .57)

and

V l = hD5 + + l h3D3 + A hl*D2 + W h5Dl <4 - 5 8 )

The modified Kutta-Merson method (4.27) applied to the DDE (4.52)

y ie ld s

ki = u‘ ( V

y x( t n + rh) = u ( t n) + r h u ' ( t n)

k2 = bu( V + (3 ' a) bhu' ( t n)

y2( t n + rh) = u ( t p) + r h [ (1 - | i " ) u ' ( t ) + | r k 2l

k3 = + (3 - « ) b)

= [b + b2h(-g- - a + 3a2) ] u (t ^ )
63.

+ [bh(-g- - Ja2) + b2h2( ^ j - -ja + ^a2 - ^aa)]u' (t n)

4 (1") = by2( t n + | r h - d)

y 3( t n + rh) = u ( t n) + r h [ l u ' ( t n) + | k ^ ( r ) ]

IS = b u (tn) + ( j - a ) b h [ | u ' ( t n) + - | k 3( l - a ) ]

= [b + b2h ( f - £ ) + b3h2( i - |» + « , 2

- f » 3) ) u ( t n) + [bh(-g - |a ) + b2h2( ^ - -ga

25 2 + -25
- -yga2 p .33W h3h3/( 1
) + b3h3 _ . 5 + ^172 o . ^1853 o

+ § a ‘t ) ] W ( t n) (4.59)
Then

y k ( t n + h) = u ( t n) + h [ l u ' ( t n) - | k 3 + 2k4 ]

=Cl + -^bh + b2h2( l - -^a2 ) + b3h3(-g- -

+ f a 2 - f a 3) ] u ( t n) + [lh + bh2 ( - l a + | a 2 )

+ b2h3( ^ r + b - i t *2 + • ^ a3) + b3hIt( 2^ - !<*

+ fa 2 - f p a 3 +f a 11) l u 1( t n) (4.60)

S u b s titu tin g u '( t ) from equation (4.54) y ie ld s

+ h) = [1 + bh +b2h2( l - a) + b3h3( l - a + -|a2 )

+, utfuij./
b h 1 - 2gd +23
-g^ 2p - 53
^ 3 3 + ^ 94h.\
)

+ b5h5( - -^a + ]Ia 2 - ^ , 3 + 4 3 ^ +^ 5 ) ] u ( t ^)

+ D6,2 (4.61)

where

D692 = ^ + bb2( ” I p + I " 2) + b2b3^2^ + I 01

_ -11,2
qu 4- l l y 3)) +13H
+ b3hl+T-l
(^4 -1 ++ -Q-az
ga llv 2- 3

+ ^ a ‘*)]D5 + b3h*»[|b2d2 - |b 3 d 3 + J J [ W m .

C^F ~ ^ ~ a^ + -I + b2h3 [- -|b3d3


+ ^ 1 r 1 i1 43 5 .34 ^ -i
+ ~2qTb “ J• 2° ~ TT T ® ■*
64.

+ - ^ b 5hd5(- 1 + | a ) (4.62)

Subtracting (4.61) from (4.56) y ie ld s

u^ n + f ' V f i + f = b‘,h‘, ('5“ ' T “ 2 " f z " 3

- ■gcl‘* ) L,( t n) + bShs (Tgo " F + W * 2 " TZ“ 3


7q -i
+ ^ - 2«5 ) u ( t n) + (D6 j1 - D6 ( 2 ) + 0(h 5) (4 .6 3 )

Nows we fin d y ( t + h) 9
5 n

y i,(tn + rh) = u ( t n) + r h [ ( l - | r + 2r 2)k 3

+ {-|r - 6r 2) k 3 + (4 r2 - 2 r)k H]

ks = b y ^ (tn + h - d)

= [b + b2h(Jj- + a - -^a2 + 2a3) + b3h2{-g-

- -|a - - ^ a 2 + 19a3 - -^ a 1* + 9a5) + b^h3^

U (tn) + [b h (-l - 2a + | a 2 - 2a3) + b2h2( ^ a 2 - ^ a 3

+ • ^ a '1 - 9a5 ) + b 3h 3(-2^- + -ja - 9a2 + -y ? ra 3

- ^ + ^ | 4 5 - 4£»6) + b - h M ^

_ 19 73 j 293 , , 1325 u 1687 *


W + 12° T T * + " 2 4 ^ ■ -STT*
, 140 6 25 7, - i ,. .
+ - y ^ b - ^ a 7) ] u ' ( t n) (4.64)

y 5( t n + b) = U(tp) + h l l k j + | k ^ + | k 5]

= [1 + |bb + b2h2(-j - ^a - "j^a2 + -^a3)

+ b3h 3(~ - | a + | a 2 + j| a 3 - -^ a 1* + | a 5)

+ h^h^f—1— -1 4- ^ 2 73,3 449 ij.


^48 3a 24a ' T 2 a + T f T 1

- f | a 5 + -j|a 6) ] u ( t n) + [^h + h2b ( i - -la

+ - ^ a 2 - ] a 3) + b2h3( ^ - - ] a 2 - ±a 3

+ ■ J a 1* - J a 5 ) + b3hi( ( ^ L ~ -got - -g*2 + 4j~a3

_ 515-a k ^ 89..s 43 R. , ukuI/ 1 19


4 8 “ + T “ 5 ' T?*6 ) + " T44
- ’j f a 7) l u' (tp ) (4.65)

S u b s titu tin g f rom (4.54) in to (4.65) y ie ld s

y 5 ( t n + h) = [1 + bh + b2h2( | - a ) + b3h3( | - a + | u2 )

+ b ^ h 1^ - ^ “ |c t + 2a2 - '^a 3 ) + b 5h 5 (—j


n
144 i f
, 301 ? 307 q , 157 u 79 q
+ “ 48^ " " 48^ + " TS r

+ " f * 7^ U( V + °6,3 (4.66)

where

ds #3 = + h2b( f " i p + j p 2 “ p 3)
1 3 + 85 ^
- | a 5)
“ 2° 2A0,

+ b3h4( - ^ - l a
37 3 _ 515 b + J§iL,5
- | a 2 + -g-a
43 ,
48 a + -g-a " TP
1 kl+UfW 1 __ 19 | 73 r,2 _ 293a 1325 , 1687
<144 144 75“ 72 a 3 +
144 a 144 a

+ - ^ a 6 - | | a 7)]D5 + bl,h5(-bd + | b 2d2 - | b 3d3

* 1 19 , 73 2293 3 .1325 „
24 ' '144 T44^ 7 ? IT 1 TTT*

- ~^ ^ a5 + ~ Y ja 7) + b3h^(-^b2d2 - ^ b 3d3

, 125.,,,,,,, 1 1. 3 2 , 37 3 515 4
24 ^ 4 8 - l a - ^a "T“ ‘ " ¥ “

+ ^ c *5 - | f a 6) + b2h3(- §b3d3 + J | [ W ) -

(■jf " t V _ ' j 0' 2 ' y 13 + f f 0*1* _ | “ 5) +

125u5u2rl4^ ^ r 7 2 1 3\ (4.67)
24 (g 201 "i "2a ” 3^ )

Thens the LTE o f the method when applied to the lin e a r DDE (4.52) is

LTE = u ( t n+1) - y 5( t n + h)

_ r 1 1 241 2 , 7 3 4 i*
720 If* "4801 I f 1 “ 301

+ | | a 5 - y | | a 6 + | a 7 ] b 5h 5 u ( t n )

+ (D6s1 - D6 j3 ) + 0(h6) (4.68)

And the LTE estimate o f the method y i e l d s 5


66.

LTE * | [ y 5( t n + h) - yit( t n + h)]

= b ^ f^ a - ^7,2 . 1 ^ 3 -

+ b5h5[72ff ' 45“ + fn “ 2 ■ I? 0 “ 3 + T lV 1*

' T l0 “ S + TZO" 6' 20“ 7] u (t n) + ^°6,3 ~ D6, 2^ + ° ( h6^ (4.69)

Hence, f o r small delay such th a t d ^ and assuming th a t the

s o lu tio n and i t s d e riv a tive s are bounded in the in te rv a l [ t , t n] , then


n n+i
equations (4 .5 5 ), (4 .58), (4 .62), and (4.67) y ie ld :

l D6, J s M! h6> l D6 >2l < M2h6> I b6 931 « M3h6 ( 4 - 70>


. . . . . . .

f o r some constants Mj9 M2, and M3. Since we are in te ste d in estimating

the p rin c ip a l p a rt o f the LTE, we can neglect the terms (D. D_ „) and
0 ,1 ° , O

(D6 g - D6 2) in (4.68) and (4.69) re s p e c tiv e ly . The s ig n if ic a n t p o int

in the LTE estimate (4.69) is the presence o f a term o f order 0(hh ) fo r

a ^ 0while the LTE o f the method is of order 0(h5) . So f o r small delay,

the LTE estimate tends to estimate a term o f order 0(h4) which is an

over estimate o f the p rin c ip a l p a rt of the LTE. If a = 0 then the LTE

estimate is equivalent to the p rin c ip a l part o f the LTE in (4 .68).

Consider the asymptotic behaviour o f the LTE and i t s estimate as

h + 0 in (4.68) and (4.69) re s p e c tive ly , since d is constant, then once

the stepsize is less than d, the formula which gives the LTE and i t s

estimate is d if f e r e n t . The case when h $ d is tre ate d in the fo llo w in g .

4 .3 .1 .2 Large delay

Assume th a t we have the solu tio n and i t s d e riv a tiv e a t the point t s

also assume th a t the delay term can be evaluated e x a c tly fo r any point s

such th a t s $ t , then we can tr e a t equation (4.51) as an ODE, and the

Kutta-Merson method y ie ld s

*1 = U< V + -3hu' ( t n)

y 2 = u ( t n) + ( jh u '( t n) + |h a y 1 + ^h b u (tn + | - d)

y 3 = u ( t n) + j j h u ' ( t n) + | h a y 2 + |h b u ( t p + -lh - d)
67.

yi,,u =(1 + ha + I *12®2 + I 11333 + 2^ h‘*a', ) u(t n)

+ l h b (1 + ^ i 2a2 + ^ L h 3a3) u ( t - d)

- -|hb(l - ih a - ^ 2a2) u ( t n + -jh - d)

+ 2hbu(tn + -jh - d) (4.71)

y 5,.U = (1 + ha + I h 2a2 + I h 3a3 + ^ h V +

+ ^ ( 1 + ha + ih 2a2 + ^h3a3 + ^ a 1* ) ^ - d)

+ -~h3ba2 ( l + -lh a )u (tn + -|h - d)

+ -^hb(l + ^h a)u (tn + -i-h - d)

+ ^ h b u (tn + h - d) (4.72)

From the re s u lts fo r the Kutta-Merson method given in (3.34) and


(3.35) 9

u ( t n + h) - y ^ u = ^ ( t n9 u ( t n) 9 h) (4 . 73)

and the LTE o f the method y ie ld s

LTE(u) = u ( t n + h) - y5jU

= ’K V Li(tn))hs + 0(h6) (4.74)

For the LTE estim ates we assume

’H V u(t n^ ) h5 = i M V u( V ’ h> + °(h 6) (4.75a)

which leads to

LTE(u) “ ^ (y5 - y j + 0(h6 ) (4.75b)

Now consider the e ff e c t of approximating the delay term on the LTE

and i t s estimate. Assume th a t the delay term in (4.51) is approximated by

z(s) at the point s such th a t

z(s) = u(s) + e(s) (4.76)

Thens the Kutta-Merson method applied to the DDE (4.5T) w ith the exact

s o lu tio n and i t s d e riv a tiv e at the point t , and using (4.76) f o r the

delay term, y ie ld s
68

y t+,z “ ^ + ha + l ^ 2 + ^ h3a3 + ) u ( t n)

+ -^hb(l + ygh2a2 + *y~h3a3) u (t^ - d)

- -?|hb.(l - --ha - -j^h2a2)z { t^ + i h - d)

+ 2hb z(tn + ^h - d)
(4.77)
and

y s,z = (1 + ha + ¥ 2a2 + ^ 3a3 + 2l hlfalf + m h5a5) u( t n5 '

**" 6 ^ 0 + ha + -^h2a2 + -gh3a3 + -^h^a^ )u (t^ - d)

+ l h 3ba2 (l + i h a ) z ( t n + ~h - d)

+ -|hb(l + ~ h a )z (tn + -~h - d)

+ <jhbz(tn + h - d)
(4.78)

By comparing (4.77) and (4.78) with u(t ^) we get

u^ n + l ^ y^ 9z ^u^ n + l^ y*+siP + ^y4 su ~ y^ 5z^

HW u( t n^ s h) • •

- ^h b (l - -|ha - ^

+ 2hbe(t^ + -^h - d) (4 . 79)

and

LTE(z) - u ( t n+1) - y ^

^ S i+ l^ y 5,u^ + *-y5,u ys 9z^

= * ( t n, u ( tn))h 5 + 0(h6)

+ gh3ba2(1 + ^ a^G(^p ^ "5^ ”*

+ -^-hb(1 + -^ha)e(tn + -^-h - d)

* +-g-hbs(tn + h - d) (4.80)

In the adaptation o f the Kutta-Merson method f o r solving DDE in

section ( 4 .2 .1 .2 ), we mentioned th a t the tru ncatio n e r ro r in approximating


the delay term s a tis fie s
69.

e(s) = 0(Hq+^ ) , where H/h = 0(1) (4-.81)

and q = 3 i f we use Hermite in te rp o la tio n (4.29) or q = 4 i f we use

the modified Hermite in te rp o la tio n (4.33).

If q = 4 then (4.79) and (4.80) reduce to the standard form o f

Kutta-Merson method, and so the LTE estimate is not affected by the

approximation o f the delay term.

If q = 3, then (4.79) and (4.80) reduce to

u ( t n+1) - y ,,j2 = ^ ( t , u ( t ) , h)

+ h b [ - . | e ( t n + l h - d) + 2 e (tn + - d )] + 0(h6) __ (4.82)

and

UTE(z) = u ( t n+1) - y S)Z

= * ( t n. u ( t n))h 5

+ h b [ | e ( t n + | h - d) + £ e ( t n + h - d) 1 + 0(h6 ) (4.83)

Equations (4.82) and (4.83) used w ith (4.75) y ie ld

LTE(z) = l ( y 5>z - y ^ ) + l h b i | c ( t n + ] h - d)

+ 3s O n + -jh " d) "■'5E:(tri + h - d)] + 0(h6 ) (4.84)

Equation (4.84) shows th a t to get an estimate o f the LTEof the method we

must estimate the e r ro r in approximating the delay term e( s) . Let us

consider an approach s im ila r to the one used f o r estim ating the LTE o f

Kutta-Merson method, l e t

s = t . + rH, 0 £ r < 1, and H ~t . . n - t -


3 3+1 3

then f o r the Kutta-Merson method (4.27) another estimate o f the delay term

is given by

y „ ( s ) = y ( t j ) + rH[(1 - | r + 2r 2) y ' ( t J.)

+ ( | r - 6r 2) y ' ( t j + -jH) + (4r2 - 2 r ) y '( t. +^ H) ] (4.85)

And the LTE f o r y 4(s) when the method (4.27) is applied to the lin e a r

ODE (3.37) is
70.

u ( t j + rH) - + rH) = -^H I+r 2( l - r )2u ^ ( t n) + 0 (H5) (4.86)

Equation (4.86) shows th a t the p rin c ip a l p a rt o f the LTE in y ^ (t + rH)

is equivale nt to the p rin c ip a l p a rt o f the e rro r in approximating the

delay term given in (4 .3 0 ), so th is approach can not be used here.

Since the purpose o f g e ttin g a LTE estim ate is to use i t fo r c o n tro llin g

the stepsize in va ria b le step alg orith m . So, i f e rro rs o f 0(h5) have been

c o n tro lle d a t e a r lie r stages, then they can be ignored in la te r in te rp o la tio n ,

unless the c o e ffic ie n t o f the e rro r term is very la rg e . Hence we can

assume th a t in general the e rro r introduced by the approximation o f the

delay term is small and the dominant pa rt in the LTE (4.83) is

* ( t n. u ( tn)) h 5

which depends on the present stepsize h. The stepsize h can be changed

to give the required accuracy.

4.3 .2 LTE estim ate fo r the trapezium method and fo r the im p lic it Runge-

Kutta method.

The technique o f halving the stepsize may be used fo r estim ating the

LTE fo r both the trapezium method and the im p lic it Runge-Kutta method, so

we give the treatm ent in general terms.

For a pth order method, to estim ate the LTE, we assume th a t i t has

the form

LTE = u ( t n))h p+1 + 0(hp+Z) (4.87)

Let the method o f approximating the delay term be o f order q. We have

shown in the la s t chapter th a t in order to preserve the order o f the method,

q must s a tis fy q 5 p,

For small delay, then the LTE always has the same form as in (4.87)

fo r q 5 p -1.

For large delay, i f q > p then the LTE o f the method has the same

form as (4 .8 7 ). If q = p-1, the LTE (3.25) o f a numerical method fo r

solving DDE is a combination o f two p a rts , the f i r s t o f the form (4.87)


71.
and the other depending on the e rro r in approximating the delay term. Hence,

to use step -ha lving in a va ria b le step alg orith m , when the c o e ffic ie n t of

the second p a rt o f the LTE (3.25) is not very la rg e , we assume th a t the

dominant p a rt o f the LTE is the one o f the form (4 .8 7 ), and neglect the

other p a rt introduced by the approximation o f the delay term.

In chapter V, we give re s u lts o f methods fo r so lving DDE w ith d iffe re n t

methods fo r approximating the delay term, and show th e ir e ffe c t on the LTE

estim ate o f the method.

4.4 S ta b ility pro p e rtie s o f numerical methods fo r so lving DDE

We discussed in the la s t chapter the d e fin itio n s o f s t a b ilit y of

methods fo r so lving DDE. The te s t equation used is the lin e a r DDE,

u '( t ) = a u (t) + b u ( t- l) t ^ t (4 . 88)

u ( t) = g (t)

where a and b are complex constants.

We now consider the s t a b ilit y properties o f the methods given in

section (4 .2 ). Assume th a t f o r each method, the numerical s o lu tio n o f

(4.88) is calcu la ted up to the p o in t t w ith a fix e d stepsize h such

th a t

t = t Q + nh, and mh = 1, m e I + (4.89)

Since the purpose o f in tro d u cin g P - s ta b ilit y and Q - s ta b ility , in

d e fin itio n s (3.3) and (3 .4 ), is to fin d methods which can be used in

p ra ctice w ith no r e s tr ic tio n on the stepsize because o f s t a b i l it y , and since


•j
h = —, then as m increases, h decreases. Hence, the im portant case is

to show th a t the method is Q-stable or P-stable fo r small values o f m.

For each o f the fo llo w in g methods fo r solving DDE, we give the re s u lts

fo r m = 1, 2, 3, and 4. Let z(s) be the approximation o f the delay

term a t s.

4.4.1 Kutta-Merson method fo r solving DDE

To advance the numerical s o lu tio n o f the DDE (4.88) to the p o in t


72.

S i+ l9 Kutta-Merson method y ie ld s ,

y x - y ( t n) + V ( t n) _

y 2 = y ( t n) + ^ h y '(tn) + l h a y x + |hbz(tn + ±h - 1)

ys = y ( V + + + f hbz( t n + | h - o

y^ = (1 + ha + gh2a2 + ~h3a3 + ^ h V * ) y ( t n)

+ ghb(1 + -ygh2a2 + ^ h 3a3) z ( t n - 1)

- |hb(1 - lha - i y , z a2) z ( t n + -lh - 1)

+ 2hbz(tn + gh - 1)

Then

y ^ H + l) - (1 + ha + “ h2a2 + -i-h3a3 + -^h ^a4 + -pp;h5a5) y ( t n)

+ gg(l + ha + ^ h 4 a4 )z (tn - 1)

+ gh3ba2(1 + I h a ) z ( t n + gh - 1)

+ ghb(1 + gha)z(tn + gh - 1)

+ ^hbz(tn + h - 1) (4.90)

Using co n d itio n (4 .8 9 ), and assuming th a t the values o f the s o lu tio n and

i t s d e riv a tiv e are stored at e a r lie r mesh p o in ts , then using Hermite

in te rp o la tio n fo r evaluating the delay term, we get

z ( t n + i h - 1) = z ( t n_n + | h )

= + ^ ha)y ('t'n-m^

+ 4j - ^ a ) y ( t n. m+1) + ^ h b y ( t n_zJ

" 27hby(t n-2m+l) ( 4 -91)

z ( t n + | h -1*1) = z ( t n_m + lh ) ' -

= ( g + ^ h a ) y ( t n_in) + ( g - g h a ) y ( t n_m+i )

+ X y ^t n-2m^ - ^k'y t ‘t n-2m+l^ (4 .9 2 )

Then (4.90) becomes


73.

y (tn+1) = (1 + ha + ^ + lh 3a3 + t T| ^ s a5)y(t n)

+ hb(l + yjha - ^ h 2a2 - yjgg^a3 - |$gh‘ta‘*)y(t n_m+l ^

+. hb(? + -J2*18 + T O h2a2 + T23ffh3aS + T Z g S ^ ^ ^ V m )

- h2b2( ^ + ^ha + - j ^ a 2 +. 6TOh3a3)y (tn_2ni+l>

+ h2b2(T2- + T^+ia + -^-h2a2 + ^ • h3a3)y(t n_Zm) (4.93)

A ll the so lu tio n s o f the d iffe re n ce equation (4.93) tends to zero as n ■* »

i f a ll the roots o f the c h a ra c te ris tic equation are in the u n it c ir c le .

The c h a ra c te ris tic equation is

s2m+1 - (1 + ha + -|jh2a2 + lh 3a3 + ^ a 1* + ^ S a S ^ m

- ^ (1 + ih a - p fh 2a2
o 648h a

- > + |h a +
W h2a2 + T O h3a3 +
^ h2b2
+ -gh2a2 + 55h3a3)S
+ 12 0 + I ha

- + -jha + |h 2a2 + ^ M a 3) = 0 (4.94)

When a and b are re a l, we give in fig u re (4 .1) the P - s ta b ilit y

regions fo r m = 19 2, 3, 4, and compare i t w ith the s t a b ilit y region o f

the DDE (4.88) in the a,b-plane. A ll the P - s ta b ility regions are closed

regions, they in te rs e c t w ith the a-axis a t the p o in t a = 3.54.m. The

in te rv a l on the a-axis gives the absolute s t a b ilit y in te rv a l fo r ODE

according to d e fin itio n (3 ,1 ).

If a - 0 and b is complex, we give in fig u re (4 .2 ) the Q - s ta b ility

region fo r m - 1, 2, 3, 4, and compare i t w ith the s t a b i l it y region o f the

DDE (4.88) in the b-plane.

4.4.2 The trapezium method fo r solving DDE

To advance the s o lu tio n from the p o in t t to the p o in t t , t , the


n n+ i
trapezium method y ie ld s

y ( t n+1) = y ( V + ^ t n + i))

+ - r ( z ( t n - 1) + z ( t n + h - 1)) ' (4.95)


74.

Vw

V n =* L

F ig . 4.1 P - s ta b ilit y regions o f the Kutta-Merson

method fo r solving DDE., w ith the s t a b ilit y

region o f the DDE (4 . 88)5 a and b re a l.


o . 1+

-l.k -a. 0. 6

--Q -z.

F ig . 4.2 Q - s ta b ility regions o f the Kutta-Merson

method fo r solving DDE, w ith the s t a b ilit y

region o f the DDE (4 .8 8 ), a = 0 and b

complex.
Using condition (4 .8 9 ), i t is cle a r th a t we get the same s o lu tio n a t t n+^

in (4,95) whether we use lin e a r or Hermite in te rp o la tio n fo r approximating

the delay term, then

y< W = y(V + ir< y (V + y< W >

+ T ( y ( t n-m) + <4-95)

Then the c h a ra c te ris tic polynomial is

(T - ^ ) 5 m+1 - (1 + ^ ) 5 m - ^ m- ^ = 0 (4.97) |

When a and b are re a l, we give in fig u re (4,3) the P - s ta b ilit y

region fo r tn = 1} 2 S 3, 4, and the s t a b ilit y region o f the DDE (4.88) in

the (a ,b )-p la n e .

When a = 0and b is complex, we give in fig u re (4 .4 ) the Q - s ta b ility

region o f the method fo r m =» 1, 2, 3, 4, and compare i t w ith the s t a b ilit y

region o f the DDE (4 .8 8 ),in the b-plane.

We mention here th a t Cryer (1973) proved th a t the method is Q-stable

fo r b re a l.

4 .4 .3 The im p lic it Runge-Kutta method fo r so lvin g DDE

By applying the im p lic it Runge-Kutta method to advance the s o lu tio n

o f the lin e a r DDE (4.88) from the p o in t tn to t -j» we get

1 + Iba +-rU zaz i£(1 + 3 .)


y(t ) = ------ 1 ------ E ------y ( t ) + ----- 6--------- f -------z ( t - 1)
1 - |h a t - ^ a 2 1 " ^ ha + T ^ 2®2

ih b hbf l ha. (4.98)


+ . lu . T (l T ‘
z ( t n + ■jh - 1) + r T T T T z ( t n + h " 1)
1 - ^ha + " 2 ' 1 - -£ha + ^ h 2a2

Using the assumption (4 .8 9 ), then i f we use Hermite in te rp o la tio n o f the

th ir d degree fo r approximating the delay term, z ( t + Jyh - 1) has the same

form as in (4 ,9 2 ), then (4.98) becomes


77.

Fig . 4.3 P - s ta b ilit y regions of the trapezium method,,

w ith the s t a b ilit y region o f the DDE (4 .8 8 ),

a and b are r e a l.
0-C

a -a .

- f -2.

F ig . 4.4 Q - s ta b ility region o f the trapezium

method, w ith the s t a b ilit y region o f

the DDE (4 .8 8 ), a = 0 and b complex.


79.

hb
6
+ (3 + h a )y (tn_m) + (3 - h a j y f t ^
ha h2az L

+ ^ V z m 5 " 2hby( V 2m+l>


(4.99)

Then the c h a ra c te ris tic polynomial is

(4.100)

For a and br e a l, we give in fig u re (4 .5) the P - s ta b ilit y regions

fo r m = 1, 2, 3, 4, and the s t a b ilit y region o f the DDE (4 .8 8 ).

If a ~ 0S b complex., the Q -s ta b ility c h a ra c te ris tic equation o f the

im p lic it Runge-Kutta method is the same as th a t o f the Kutta-Merson method

fo r so lvin g DDE (4 .9 4 ). Hence the Q -s ta b ility region fo r the im p lic it

Runge-Kutta is the same- as in fig u re (4.2) fo r m = 1, 2, 3, and 4.

Remark 4.1

I t is c le a r from the fig u re s (4.2) and (4.4) th a t the regions o f

Q - s ta b ility fo r m = 1, 2S 3, and 4 are a ll greater than or equal to the

s t a b ilit y region o f the DDE (4.88) in the b-plane fo r a ll the methods

considered. Also in fig u re s (4 .3 ) and (4.5) the P - s ta b ilit y regions fo r

the trapezium method and the im p lic it Runge-Kutta method are greater than

the s t a b ilit y region o f the DDE (4.88) in the a, b-plane, and so the

stepsize is not re s tric te d by the s t a b ilit y pro p e rtie s o f the method. For

the Kutta-Merson method, fig u re (4.1) shows the e ffe c t o f choosing a c e rta in

stepsize on the P - s ta b ilit y region. We w ill give numerical re s u lts in the

next chapter to show the e ffe c t o f the s t a b ilit y p ro p e rtie s o f the method

on the choice o f the stepsize.

Remark 4.2

The way we ca lcu la te the s t a b ilit y region, fo r example fig u re (4 .1 ),

is by ta kin g d iffe r e n t values o f (a, b ), the v e r tic a l and h o rizo n ta l axes,


80.

■m= l .

F ig . 4.5 P - s ta b ility regions of the im p lic it

Runge-Kutta method. With the s t a b ilit y

region o f the DDE (4 .8 8 )s a and b re a l.


81.

and fin d in g the roots o f the s t a b ilit y polynomial using the NAG lib r a r y

ro utine C02ADA. I f a ll the roots have magnitude less than one then we

accept the value o f (a 9 b) as p a rt o f the s t a b ilit y region. For the

P - s ta b ilit y regions we change, the values o f a and b by 0.25 each tim e,

and fo r the Q - s ta b ility region by 0.1 each tim e. I f i t appears in some o f

the fig u re s th a t the curves are id e n tic a l, th is is not e x a c tly so, but

they are the same f o r the accuracy we are using.

Remark 4.3

I t is not easy to prove P - s ta b ility o r Q - s ta b ility re s u lts fo r general

m. Our conjecture is th a t a ll the methods we have considered are Q -stable,

and the trapezium method and the im p lic it Runge-Kutta method fo r s o lvin g

DDE are also P -stable.


CHAPTER V

NUMERICAL RESULTS AND ALGORITHMS

5.1 In tro d u ctio n

The evaluation o f a numerical method fo r the s o lu tio n o f d iffe r e n tia l

equations depends on many fa c to rs includ ing

( i ) lo ca l tru n c a tio n e rro r,

( i i ) s t a b ilit y c h a ra c te ris tic s ,

( i i i ) e rro r propagation.

Mathematical analysis can u su ally give an evaluation o f these fa c to rs , as

in section (4.3) and (4 .4 ). However mathematical re s u lts fo r d iffe r e n t

methods may be incomplete or d i f f i c u l t to compare and they often re q u ire

s im p lify in g assumptions fo r any n o n -tr iv ia l problem. Thus numerical re s u lts

are fre q u e n tly used to te s t and compare d iffe r e n t methods.

For te s tin g and comparing numerical methods fo r so lv in g i n i t i a l

value ODE, Krogh (1970) presents a guide to the steps required when

producing a set o f re s u lts which should give a reasonably complete evaluation

o f a method. Hull e t al (1972) present a v a rie ty o f i n i t i a l value problems

fo r te s tin g a numerical method, also they suggest a basis fo r comparing

numerical methods based on how w ell they can handle r e la t iv e ly ro u tin e


83.

in te g ra tio n under a v a rie ty o f accuracy requirements.

Since any ODE is a special case o f DDE, every method fo r solving DDE

is also a method fo r solving ODE, so we w ill not be concerned w ith the

pro p e rtie s o f numerical methods fo r solving i n i t i a l value ODE. This is

in ve stig a te d very thoroughly in the lite r a tu r e , we mention here in p a rtic u la r

E nright e t al (1974a) fo r the evaluation o f Kutta-Merson as a general purpose

method fo r the numerical so lu tio n o f n o n - s tiff i n i t i a l value problems.

For the evaluation o f numerical methods fo r s t i f f i n i t i a l value problems,

where good s t a b ilit y p rope rties o f the method are u s e fu l, E nrig ht e t al (1974b)

discuss concepts fo r measuring the cost and r e l i a b i l i t y o f numerical

methods, and give p re lim in a ry re s u lts o f evaluating some im p lic it one-step

methods.

In th is chapter a fte r describing the p ra c tic a l considerations in

using a pth order one-step method fo r solving DDE, we present a s e le c tio n

o f te s t problems designed to show:

(1) the performance o f the method w ith a v a ria b le stepsize under

d iffe r e n t accuracy requirements,

(2) the e ffe ctive n e ss o f the local tru n c a tio n e rro r estim ate,

(3) the r e s tr ic tio n on the stepsize by the s t a b ilit y p ro p e rtie s o f

the method,

(4) the e ffe c t o f having d e riv a tiv e d is c o n tin u itie s a t points in

the range o f in te g ra tio n and the a b ilit y o f the method to cope

w ith them,

(5) the e ffe c t o f approximating the delay term.

5.2 D escription o f algorithm s fo r solving DDE

The implementation o f any numerical method fo r s o lv in g DDE in a

v a ria b le step alg orith m must co n sist o f the fo llo w in g :

( i) a formula fo r computing the next approximation to the s o lu tio n and

estim ating the local tru n c a tio n e rro r,

( i i ) a stra te g y fo r accepting the s o lu tio n and choosing the next stepsize.

Assume th a t the s o lu tio n is evaluated a t the p o in t t and i t is required


84.

to advance the s o lu tio n to the p o in t t ^ w ith the stepsize h = t .j - t .

One o f the basic requirements fo r any. numerical method fo r computing the

next approximation a t t +-j is the storage o f s u ffic ie n t back inform ation

so th a t the method can evaluate the delay term when i t is required a t

some p o in t t £ t n- The amount o f inform ation to be stored at each time

step depends on the method fo r approximating the delay term, but the in te rv a l

on which the info rm a tion is to be stored and the number of q u a n titie s to

be stored on th a t in te rv a l should be fle x ib le and adaptable fo r each problem,

depending on the nature o f the delay and the required accuracy re s p e c tiv e ly .

I f the delay term needs to be evaluated a t some p o in t t £ t Q, where tQ

is the i n i t i a l p o in t, then the i n i t i a l fun ction must be used.

For a stra te g y fo r accepting the so lu tio n and choosing the next

ste p size , assume th a t we are given a required e rro r tolerance e, and th a t

the lo ca l tru n c a tio n e rro r is estimated by E, then we make the fo llo w in g

te s ts :

(a) I f E > s, then re je c t the computed s o lu tio n . For choosing the next

ste p size , we have to take account o f the p o s s ib ility o f a p o in t o f jump

d is c o n tin u ity in the kth d e riv a tiv e o f the s o lu tio n , where k $ p+1.

Hence we re g is te r the p o in t t* - t + h as a possible p o in t o f d is c o n tin u ity

and then we choose the next stepsize as h /2 . Using the stepsize h/2

and the s o lu tio n and it s d e riv a tiv e values a t tn we c a lc u la te the next

approximation o f the s o lu tio n and the LTE estimate E, and repeat the

te s t in ( a ) .

(b) I f E$ e then accept the next approximation to the so lu tio n and le t

the next mesh p o in t ' be t Rjj= t n + h. For choosing the next stepsize

we make the fo llo w in g te s ts :

( i) I f E ^ e/2p+^ , keep the same stepsize and go to ( i i i ) .

( ii) I f E < e /2 ^ then: I f t * $ t n tthen double the stepsize

otherwise keep the same stepsize and go to ( i i i ) .

( iii) If t n+h > T, where T is a p o in t where the s o lu tio n is re quire d,

then we take the next stepsize to be h = T - t . Using the


85.

stepsize h and the s o lu tio n and i t s d e riv a tiv e values we

c a lc u la te the next approximation and repeat the te s t in (a ).

Remark '5 J

The stra te g y o f halving and doubling to get the next stepsize is

a rb itra ry ., and one can choose other fa c to rs fo r decreasing and increasing

the ste p size , but because o f the p o s s ib ility o f d e riv a tiv e jump d is c o n tin u ity

o f the s o lu tio n one should always have upper and lower lim its on the

choice o f the next ste p size , say h/2 and 2h, so th a t the stepsize w ill

not flu c tu a te w ild ly .

Following th is general discussion, we now describe the way we

implemented the method developed in section (4.2) f o r solving DDE.

5.2.1 Kutta-Merson method fo r solving DDE

We implemented the Kutta-Merson method (4.27) in an Algol program

f o r so lving a general system o f i n i t i a l fu n ctio n DDEs w ith m u ltip le delays

o f the form:

u '( t ) = f ( t , u ( t ) , . u ( t - d j ( t , u ( t ) ) , . . . , u ( t - d j t , u ( t ) ) ) ) ,

t0 S t S T (5.1)
and

U (t) = g ( t ) s t E [ t * , t 03 (5.2)

where

t * = min [ t - d j t , u ( t ) ) 9— , t- d ( t , u ( t ) ) ] ,
t 0*t$T ~ m "
d.|(t, jj ( t ) ) £ 0, i = 1 , o . . 3 tn5

and u^t) is an n dimensional ve c to r. The s o lu tio n is required on the

in te rv a l [ t Q, T ], The program requires three procedures depending on the

problem to be solved,

(1) a procedure to evaluate the s o lu tio n fo r t $ t , using the i n i t i a l

fu n ctio n (5 .2 ), o f the form:

INIFUN(TNp Y);

where Y is an n dimensional a rra y , which contains the so lu tio n value


a t the p o in t TN on e x it.

(2) a real procedure to evaluate the it h delay, o f the form:

D(TN, Y, I ) ;

where Y is an n dimensional array which contains the s o lu tio n value a t

the p o in t TN.

(3) a procedure fo r evaluating the d e riv a tiv e o f the form:

FUN(TN, Y, YD, F ) s

where TN is the value o f the dependent v a ria b le , Y is an n dimensional I

array which contains the value o f the s o lu tio n a t the p o in t TN, and YD

is an n x m array which contains the value o f the delay term which is a

ve cto r o f dimension n evaluated at the p o in t TN - D(TN, Y, I ) fo r

X 1, 2 , . * . , m. I

The other inform ation needed by the program is the size o f the system n,

the number o f delays m, the in te rv a l fo r the required inform ation to be j

sto re d , and the number o f q u a n titie s to be stored on th a t in te r v a l. The

main program consists o f three p a rts:

(1) a procedure fo r lo c a tin g and evaluating the delay term , j


(2) a procedure fo r advancing the s o lu tio n a t the p o in t t +-j,

(3) a procedure fo r s to rin g the required inform ation a t each in te r v a l.

5 .2 .2 . The trapezium method and the im p lic it Runge-Kutta method fo r

so lvin g DDE

Since we are in te re s te d in the good s t a b ilit y p ro p e rtie s o f the

trapezium method and the im p lic it Runge-Kutta method fo r solving DDE discussed

in section (4 .4 ), we implemented them only fo r a system o f lin e a r DDE o f

the form

u / ( t ) •== A jj(t) + B u (t-d ), t 5 tQ (5.3)

H (t ) = 9 ( t ) 9 t s [-d , t Q]

where jj( t) is an n dimensionalv e c to r, A and B are n * n matrices

and d is a constant.

The s tru c tu re o f the program fo r solving the system (5 .3 ) using the


trapezium method or the im p lic it Runge-Kutta method is b a s ic a lly the same

as the one described in section (5 .2 .1 ). The only d iffe re n c e here is th a t,

since these methods are im p lic it , we need to solve a system o f lin e a r

equations, fo r which we use two procedures. The f i r s t is to fa c to riz e a

m a trix in to a lower and upper tria n g u la r m atrices, interchanging the rows I

i f necessary, and te s t fo r s in g u la rity . The second is to solve a system

o f lin e a r equations using the re s u lts o f the f i r s t procedure. The f i r s t

procedure is used only when the stepsize is changed, and the second is

used a t each step.

5.3 Test problems and numerical re s u lts

The fo llo w in g te s t problems are chosen from a v a rie ty o f DDE, w ith

so lu tio n s e ith e r smooth or having d e riv a tiv e d is c o n tin u itie s a t some points

in the range o f in te g ra tio n . For each o f the problems solved on a given

range and fo r a required accuracy, we present the maximum global

d is c re tiz a tio n e rro r on the range o f in te g ra tio n as a measure o f the

r e li a b i l i t y o f the method and the number o f d e riv a tiv e evaluations on th is

in te rv a l as a measure o f the e ffic ie n c y of the method.

The f i r s t problem is chosen to show the meaning o f s tiffn e s s fo r

DDE, and the advantage o f having methods which have no r e s tr ic tio n on the.

stepsize because o f s t a b ilit y prope rties o f the methods. The remaining

examples are solved w ith the Kutta-Merson method fo r DDE. To show the

e ffic ie n c y o f the method in coping w ith points o f d is c o n tin u ity , we give

the re s u lts f i r s t by le t t in g the method choose the mesh points and then

comparing them w ith the re s u lts when the points o f d e riv a tiv e d is c o n tin u ity

are forced to be mesh p o in ts . To show the e ffe c t o f approximating the

delay term on the e ffic ie n c y and r e l i a b i l i t y o f the methods we solve a ll

the problems w ith the delay term evaluated e xa ctly and then compare i t

w ith the re s u lts when the delay term is approximated by one of the methods

given in section (4 .2 .1 .2 ).

In a ll the problems solved we use the s ta rtin g stepsize h = 0.1 and


88.

we use an absolute e rro r te s t unless otherwise s ta te d . A ll the c a lc u la tio n s

are performed on the CDC 7600 computer which has a 4 8 -b it mantissa

(equivalent to approximately fifte e n decimal d ig it s ) .

The fo llo w in g notation is used in the tables

e = the required e rro r tolerance,

ND = number o f d e riv a tiv e evaluations,

GE = the maximum global d is c re tiz a tio n e rro r on the in te rv a l

o f in te g ra tio n ,

* = the in te g ra tio n was terminated w ith stepsize very sm all.

Problem 5.1

u '( t ) = a u (t) + b u (t - d ), 0 < t £ T (5.4)


s •t
u (t) = E-e 1 , t e [-d , 0] (5.5)

where s* are some o f the real roots o f the c h a ra c te ris tic equation

h(s) = s - a - be"ds (5 .6 )

Equation (5 .4 ) is a lin e a r DDE w ith constant delay, and has a smooth

s o lu tio n o f the form

s.t
u (t) = e e 1 3 t $ 0

We have mentioned in section (2.4) th a t the c h a ra c te ris tic equation

(5.6) has an in f in it e number o f ro o ts . Also i t is c le a r from the form o f

equation (5.6) th a t i t has a t most two real ro o ts . Here we are going to

in v e s tig a te num erically the e ffe c t o f choosing d iffe r e n t values o f the

roots on the s t a b ilit y prope rties o f the method considered. The choice

o f the values o f a and b is such th a t a ll the roots have negative

real p a rts .

In a v a ria b le stepsize algorithm , fo r small values o f t the

stepsize is re s tric te d by re q u irin g small local tru n c a tio n e rro r, and fo r

large values o f t the stepsize is re s tric te d by the absolute s t a b ilit y


consideration s. Since the s o lu tio n is smooth, the method should take

successively la rg e r steps as t increases u n til the boundary o f the

s t a b ilit y region is reached, when i t should continue the in te g ra tio n

w itho ut in tro d u cin g s ig n ific a n t o s c illa tio n s in to the s o lu tio n . Also the

numerical re s u lts should show the effectiveness o f the loca l tru n ca tio n

e rro r estim ate in c o n tro llin g the stepsize and i t s re la tio n w ith the

absolute s t a b i l it y requirement. We give the numerical re s u lts fo r the

s o lu tio n on d iffe r e n t ranges; fo r the la rg e r ranges o f in te g ra tio n the

stepsize should get progress!vely la rg e r t i l l i t a rriv e s a t the boundary

o f the s t a b ilit y region o f the method, and i f any e rro r is introduced a t

e a rly stages o f the in te g ra tio n i t should not grow as the in te rv a l o f

in te g ra tio n gets la rg e r.

5.1-a

a = - 105

b - 1

d = &n(105 - 1)

_4-
The s o lu tio n : u ( t) = e , t 5 0.

The re s u lts are given in tables (5.1) and (5.2) when using the

trapezium method w ith lin e a r in te rp o la tio n and the trapezium method w ith

Hermite in te rp o la tio n re s p e c tiv e ly . In tab le (5 .5 ) the re s u lts are given

fo r the im p lic it Runge-Kutta fo r solving DDE w ith Hermite in te rp o la tio n .

These re s u lts show th a t the methods have achieved the required accuracy,

apart from e = 10" 10 in table (5.5) fo r the im p lic it Runge-Kutta method.

Also the re s u lts show th a t fo r the la rg e r in te rv a ls o f in te g ra tio n a

bigger stepsize is used. Table (5.1) and (5.2) show th a t there is no

s ig n ific a n t gain in using a higher in te rp o la tio n polynomial fo r approximating

the delay term w ith the trapezium method. In these ta b le s , changing the

required accuracy from 10“ 2 to 10"6 makes no d iffe re n c e to the re s u lts ;

th is is because the s ta rtin g stepsize is small and although the method


90.

increases i t successively i t a rriv e s a t the end o f the in te g ra tio n in te rv a l

before g e ttin g the optimal s te p s iz e .. In tables (5.3“)» (5 J£)5 and (5.6)

we give the re s u lts fo r th is problem using the same methods as before

but w ith the s ta rtin g stepsize equal to the in te rv a l o f in te g ra tio n ;

these re s u lts show th a t we need the same number o f d e riv a tiv e evaluations

on the d iffe r e n t in te g ra tio n in te rv a ls fo r the required accuracies 10~2

and lO "4 .

When we attempted to solve th is problem w ith the Kutta-Merson method

fo r so lvin g DDE., the stepsize was very s m a ll9 about 5 x 10~5 9 and the

global e rro r kept o s c illa tin g . This problem is an in te re s tin g one which

shows the kind o f s tiffn e s s encountered w ith DDE and the usefulness o f

having methods w ith good s t a b ilit y p ro p e rtie s .

5.1-b

a = -10

b = 1

d = an(9)

si

The s o lu tio n : u ( t) = e~^9 t $ 0.

In tables (5 .7 ) and (5.8) the re s u lts are given when the trapezium

method w ith lin e a r in te rp o la tio n and the trapezium method w ith Hermite

in te rp o la tio n are used re s p e c tiv e ly . In tab le (5.9) the re s u lts are

given when using the im p lic it Runge-Kutta method w ith Hermite in te rp o la tio n s

and in (5.10) the re s u lts fo r the Kutta-Merson method w ith Hermite

in te rp o la tio n . A ll the re s u lts showth a t we have achieved the required

accuracy., and ta b le (5 .8) shows th a t there is no s ig n ific a n t gain in

using Hermite in te rp o la tio n w ith the trapezium method. Table (5.10) shows

th a t fo r a low accuracy requirement ( e = 10“ 2) 9 the method needs almost

twice the number o f d e riv a tiv e evaluations when the in te rv a l o f in te g ra tio n

is doubled. Also the stepsize o s c illa te s between two values which shows
91.

the r e s tr ic tio n imposed by the s t a b ilit y requirement o f the method.

5.1-c

a - -0.54557126968

b = -0.45397452872

d = 10“ 3

s x = -1 , s2 = -10000

The s o lu tio n : u (t) = e_ t + t $ 0.

This is an in te re s tin g problem* where one would encounter s tiffn e s s


.Conipl<£

fo r an ODE w ith a^funheition s im ila r to the one above. In ta b le (5.11)

and .(5.12) the re s u lts are given fo r the trapezium method w ith lin e a r

in te rp o la tio n and the trapezium method w ith Hermite in te rp o la tio n

re s p e c tiv e ly . In ta b le (5.13) the re s u lts are given fo r the im p lic it

Runge-Kutta method. The poor re s u lts fo r high accuracy requirement are

m ainly due to the u n r e lia b ilit y o f the LTE estimate in c o n tro llin g the

stepsize a t an e a rly stage o f the in te g ra tio n * the stepsize gets progressively

la rg e r fo r large in te r v a l. Also in tables (5.11) and (5.12) there is no

s ig n ific a n t improvement when using a higher in te rp o la tio n formula w ith

the trapezium method.

In ta b le (5.14) we give the re s u lts when using the Kutta-Merson method

fo r so lvin g th is problem. The re s u lts show th a t the method performs very

w ell in achieving the required accuracy and the stepsize gets progressively

bigger fo r large in te r v a l.

5.1-d

a = 0

b =-1

d = 1CT3

Sjl = -1 .001001502672

s2 = -9118.006470403
92.
Sjt s2t
The s o lu tio n : u (t) = e + e , t ^ 0.

Since a ll the methods considered are Q -stable, there should be no

s t a b ilit y problem in solving th is problem. In tab les (5.15) and (5.16)

we give the re s u lts fo r the trapezium method w ith lin e a r in te rp o la tio n

and fo r the trapezium method w ith Hermite in te rp o la tio n re s p e c tiv e ly . In

ta b le (5.17) we give the re s u lts when using the im p lic it Runge-Kutta method

w ith Hermite in te rp o la tio n , and in ta b le (5.18) we give the re s u lts when

using Kutta-Merson method fo r solving DDE w ith Hermite in te rp o la tio n . For

high accuracy requirements the Kutta-Merson method and the im p lic it

Runge-Kutta achieve the required accuracy, but the trapezium method does

not because o f the .low order o f the method. Also in tables (5.15) and

(5.16) show th a t there is no s ig n ific a n t improvement in using higher order

in te rp o la tio n formula fo r approximating the delay term w ith the trapezium

method.

Remark 5.2

One w ill encounter s tiffn e s s s im ila r to th a t fo r a system o f lin e a r

ODE when the eigenvalues o f the matrices A or B in equation (5 .3)

d if f e r w idely re la tiv e to the time scale. From the case considered fo r

problem (5 .1) we suggest the fo llo w in g d e fin itio n o f s tiffn e s s f o r DDE:

"The lin e a r DDE (5 .4) is c a lle d s t i f f i f |a |d or |b |d is large r e la tiv e

to the time sca le ".

Problem 5.2

u 1( t ) = - u ( t - 1) , t 5 0

u(0) = 1

u ( t) = 0, -1 t < 0

Results to be given fo r t e [0, 6 ].

The delay: d = 1

Inform ation needs to be stored fo r a back distance = 1.

The s o lu tio n :
D is c o n tin u ity a t t = j 9 j = 0 9 1, 2, . . . 9 in the j t h d e riv a tiv e .

In ta b le (5.19) the re s u lts are given when the method is l e f t to

choose the mesh points and in ta b le (5.20) the re s u lts when the points o f

d e riv a tiv e d is c o n tin u ity are forced to be mesh p o in ts . In both cases we

considered three ways fo r approximating the delay term., by using the

tru e s o lu tio n to evaluate the delay-term , by using Hermite in te rp o la tio n

o f the th ir d degree3 and by using the modified Hermite in te rp o la tio n o f ;

the fo u rth degree (described in. section (4 .2 .1 .2 )) . These re s u lts show: j!


th a t there is no gain in using the modified Hermite in te rp o la tio n , though j
i t is more expensive in d e riv a tiv e evaluations and in the amount o f

info rm a tion needed a t each step, so we w ill not consider th is method fo r

oth er problems. Also the re s u lts o f ta b le (5.19) show th a t the method is

successful in choosing s u ita b le mesh points to achieve the required

accuracy, though i t uses a s lig h t ly la rg e r number o f d e riv a tiv e evaluatio ns.

Problem 5.3

u '( t ) = -u (e xp (l - 1/ t ) ) + 1, 0 < a $ t $ b

u ( t) = £ n ( t) , 0 < t $ a

Results to be given fo r t s [1 , 10].

The delay: d ( t) = t - exp(l - 1 / t ) .

Inform ation to be stored only fo r t e [1 , e ].

The s o lu tio n : u ( t) = £ n ( t) , t 5 a.

For th is problem the s o lu tio n is smooth and i t s d e riv a tiv e depends

only on inform ation in the in te rv a l [ 1, e ], and the delay is zero at the

s ta rtin g p o in t t = 1. In tab le (5.21) the re s u lts are given fo r th is

problem when using the Kutta-Merson method fo r solving DDE w ith f i r s t

the delay term evaluated e x a c tly and then w ith Hermite in te rp o la tio n to

approximate the delay term. The re s u lts show th a t there is not much

d iffe re n ce in the accuracy achieved and the number o f d e riv a tiv e evaluations
94.

is not a ffe cte d by approximating the delay term.

Problem 5.4

u '( t ) = - u ( t - 1 + e "t ) + s in (t - 1 + e“ t ) + c o s t,

t 5 0

u (t) = s in ( t ) , t ^ 0

Results to be given fo r t £ [0, 10 ].

The delay: d ( t) - 1 - e^

Inform ation needs to be stored fo r a back distance » 1.

The s o lu tio n : u ( t) = s in ( t ) .

This problem has a smooth p e rio d ic s o lu tio n * and the delay is zero

a t the s ta r tin g p o in t. In tab le (5.22) the re s u lts are given when the

problem is solved w ith the Kutta-Merson method fo r DDE* f i r s t w ith the

delay term evaluated e xa ctly and then w ith Hermite in te rp o la tio n fo r

approximating the delay term.

Problem 5.5

u 1( t ) = + [u (t/(l + 2 t ) 2 ) f 1+2t)\ t * 0

u(0) = 1

Results to be given fo r t e [0* 10].

The delay: d ( t) = t -------- — -


v 1 ( l+ 2t )2
1
Inform ation needs to be sotred only fo r t e [0* -g].

The s o lu tio n : u ( t) = e+ t .

Since the s o lu tio n o f th is is increasing very fa s t* we use a

r e la tiv e e rro r te s t. In ta b le (5.23) we give the maximum r e la tiv e global

e rro r and the number o f fu n c tio n evaluations when the problem is solved

by Kutta-Merson method fo r DDE, w ith the exact s o lu tio n fo r evaluating

the delay term and Hermite in te rp o la tio n fo r eva lu a tin g the delay term.

Problem 5.6

u '( t ) - c o s (t) . u (u (t) - 2), t £ 0

u ( t) = 1, t $ 0
95

Results to be given fo r t e [0, 10],

The delay: d ( t 9 u ( t) ) = t - u (t) + 2

The s o lu tio n : u (t) = s in ( t ) + 1, t 5 0.

By knowing the solu tion^ we notice th a t there is no need to store

any in fo rm a tio n 9 since d ( t 9 u ( t) ) ^ t 9 and so the method always uses the

i n i t i a l fu n c tio n fo r evaluating the delay term. The re s u lts are given in

ta b le (5.24) using the Kutta-Merson method f o r DDE.

A ll the fo llo w in g problems are solved by using the m odified Kutta-

Merson method fo r so lving DDE w ith Hermite in te rp o la tio n fo r approximating

the delay term and w ith the exact s o lu tio n fo r evaluating the delay term.

Problem 5.7

u1( t ) = ~r e x p (u (u (t) ~ Jtn(2).+ 1)) 9 t ^ l

u ( t) = 0, t * 1

Results to be given fo r t e [1 9 3 ].

The delay: d ( t 9 u ( t) ) = t - u ( t) + zn(2) - 1

The s o lu tio n :

Jin(t) 1 ^ t $ 2
u (t)
| + £n{2) - 19 2 ,< t ,< 3

The s o lu tio n has d e riv a tiv e d is c o n tin u itie s a t the po ints t = 1

and 2 in the f i r s t and second d e riv a tiv e re s p e c tiv e ly . Notice th a t the

back distance o f inform ation needed and the lo c a tio n o f points o f

d e riv a tiv e d is c o n tin u ity are not known a p r io r i . The re s u lts are given

in ta b le (5.25) and (5 .2 6 ).

Problem 5.8

u '( t ) = ^ u ( t ) . u ( £ n ( u ( t ) ) ) 9 t ^ 1

u ( t) = 19 t ^ 1

Results to be given fo r t e [1 9 e2] .

The delay: d ( t 9 u ( t) ) = t - n n (u ( t)).

The s o lu tio n :
96

1 s t ^ e
u(t) =
e x p (t/e ) e £ t £ e2

W ithout knowing the s o lu tio n o f th is problem one cannot determine

the po ints o f d e riv a tiv e d is c o n tin u itie s a t t - 1 and t = e in the

f i r s t and second d e riv a tiv e re s p e c tiv e ly . Also i t is not possible to

know the back distance fo r s to rin g the required in fo rm a tio n . The re s u lts

are given in tables (5.27) and (5 .28).

Problem 5.9

u ' ( t) = u2(t)

U ^(t) = -u2(exp(l-u2( t ) ) . [ u , ( t ) ] 2 .exp(t-u2( t ) ) ,

w ith

ux ( t ) = a n (t)

u2 ( t ) = 1 / t

Results to be given fo r t e [1, 10]

The delay; d ( t, jj( t )) = t - exp(l - u2( t ) )

The s o lu tio n :

ux (t ) = Jin(t)
t $ a > 0
u2 ( t ) = 1 / t

This problem is a system o f f i r s t order DDE w ith a smooth s o lu tio n .

The back distance fo r the inform ation to be stored is not known a p r io r i .

Results are given in tab le (5 .2 9 ).

Problem 5.10

For te [0, 1]

u j( t ) = u5( t - l ) + u3( t - l )

u * ( t) = U j f t - l ) + u2(t- 0 .5 )

Ug(t) = u3( t - l ) + u1(t-0 .5 )

u ^ (t) - u5 ( t - 1 ) .u ^ ( t- 1 )
97.

u j(t) = (t-1)

with f o r t e [-1 , Ol

U i( t) = \ ( t ) = us ( t ) = e x p (t+ l)

u2( t ) = exp(t+0.5)

u3( t ) - s in ( t + l)

The delays: dx « 0 .5 ; d2 = 1.

Inform ation to be stored fo r a back distance = 1.

The s o lu tio n :

ux ( t ) = je x p (t) - c o s (t) + e9 0 s t s 1

2exp(t) + exp(0.5) - 2, 0 $ t £ 0.5

u ,( t ) = e x p (t) + 2exp(t - 0.5) + t.e x p (0 .5 )

- 2t + 1.5exp(0.5) - 39 0.5 $ t £ 1

exp(t+0.5) - c o s (t) + 1 - exp(0,5)

+ s in ( l) 5 0 ^ t s< 0.5
u,(t) =
-c o s (t) + e x p (t-0 .5 ) - s in (t-0 .5 )

+ (t+ 0.5 )e - exp(+0.5) + s i n ( l ) , 0.5 < t £ 1

u4( t ) = 0.5exp(2t) - 0.5 + e a 0 ^ t £1

u5 ( t ) = e x p (t) + e - 1, 0 $ t $ 1

The re s u lts fo r th is problem are given in tables (5.30) and (5 .3 1 ).

5.4 Concluding Remarks

I t is c le a r from the numerical re s u lts o f the m odified Kutta-Merson

method fo r so lvin g DDE th a t i t performs very w ell and g e nerally achieves

the required accuracy. Also the LTE estimate o f the method succe ssfu lly

co n tro ls the stepsize to cope w ith the points o f d e riv a tiv e d is c o n tin u ity ;,

and there is no need to know these points in advance. The e ffe c t o f

approximating the delay term on the LTE estim ate o f the method is n e g lig ib l

and i t is s u ff ic ie n t to use the th ir d degree Hermite in te rp o la tio n w ith

the fo u rth order Kutta-Merson method fo r solving DDE.


98.

We mention here th a t the approach used in section (4 .2 .1 ) to modify

the Kutta-Merson' method to solve DDE can be used fo r any one-step method

o f Runge-Kutta type* the only consideration being the e ffe c t o f th is

m o d ific a tio n on the order o f the method. Also there is no advantage in

using a higher order in te rp o la tio n formula than necessary, and i t is

s u ff ic ie n t to use an in te rp o la tio n formula o f degree one less than the

order o f the method.

The s t a b il it y p ro p e rtie s o f numerical methods f o r solving DDE need

fu r th e r in v e s tig a tio n . I t would be in te re s tin g to know necessary and

s u ffic ie n t con dition s on a lin e a r DDE w ith constant delay and complex

c o e ffic ie n t such th a t the s o lu tio n is a s ym p to tica lly s ta b le , and then one

could use the more general d e fin itio n o f D A -s ta b ility suggested in section

(3 .3 ). I t would also be in te re s tin g to know the re la tio n between the

ro ots o f the c h a ra c te ris tic equation o f the lin e a r DDE and the roots o f

the s t a b i l it y polynomial o f the numerical method fo r solving DDE.

A& C &U TW C3 . 2 . 2 . 2 y ,
In te r v a l: [0 , 10] [0 S 20] [0 , 401'

£ GE ND GE ND GE ND

10“ 2 2.1201xl0“ 7 22 2.5467xl0"7 25 2.5467xl0~7 31


1 0 - 4 2.1201xl0-7 22 2.5467xl0"7 25 2.5467x]0"7 31

10"6 2.1201xl0"7 22 2.5467xl0“ 7 25 2.5467xl0“ 7 31

10"8 4.3247x10“ 9 121 5 .5272x]0~9 130 5.5272x10-9 136

TO-10 4.2721xlO "11 763 6.6280x10-11 808 6.6280xl0“ n

"•J
CO
Table 5.1 Problem (5 .1 -a ). Using the trapezium

method fo r DDE., w ith lin e a r in te rp o la tio n .

S ta rtin g stepsize h = 0 .1 .

I n te r v a l: [0 9 10] [0, 20] [0 S 40]

£ GE ND GE ND GE ND

10-2 2.1201x!0~7 22 2.2288xl0"7 25 2.2288xl0~7 31


10-4 2.1201x1 O’ 7 22 2.2288xl0"7 25 2.2288x10“ 7 31

10"6 2.1201xl0"7 22 2.2288xl0“ 7 25 2 .2288x]0~7 ' 31

1 0 '8 4.3247x10 "9 121 4.3443x10 "9 130 5.1801xl0~9 136

10"10 4.2721x]O^11 763 4.4023x!0-n 808 4.6652xl0_11 817

Table 5.2 Problem (5 .1 -a ). Using the trapezium

method fo r DDE., w ith Hermite in te rp o la tio n .

S ta rtin g stepsize h = 0.1 .


In t e r v a l: , [ 0 9 10] [ 0 , 2 0] [ 0 . 4 0]

e GE ND GE ND GE ND

10-2 6 .0 5 3 2 x 1 0 " 6 4 9 .5 1 1 3 x l 0 “ 6 4 9 .0 0 0 0 x1 0 -8 4

TO-4 6 .0 5 3 2 x l0 "6 4 9 . 5 1 1 3 x l0 - 6 4 9 . 0 0 0 0 x l0 " 8 4


1 cr>

1 .1 2 61 x1 0-6 17 1 .7 6 5 2 x ]0 “ 6 31 1 .7 6 5 2 x1 0 -8 57
o
00
l
r—
o

4 . 8 8 5 1 x l0 -9 . 205 1 . 1 4 9 2 x l0 - 8 399 1 .1 4 9 2 x 1 0 -8 476


1
O
O

4 .2 5 5 1 x l O " 11 733 7 .0 4 4 9 x l0 "u 780 7 .0 4 4 9 x 1 0” 11 791

Table 5.3 Problem (5 .1 -a ). Using the trapezium

method fo r solving DDE w ith lin e a r in te rp o la tio n .

S ta rtin g stepsize equal to the in te rv a l o f

in te g ra tio n .

In te r v a l: [0 , 10] [ 0 , 20] [ 0 , 40]

£ GE ND GE ND GE ' ND
N>

2.0 4 49 x1 0-8
oi

6.0532x1 O '6 4 4 5 .2 9 6 9 x l0 -5 4

TO-*1 6 .0 5 3 Z x l0 -6 4 2 .0449x10-8 4 5 .2 9 6 9 x l0 -5 4

1CT6 1 .1261x10-8 17 1 .1 2 6 1 x l0 -5 31 1 .1 2 6 1 x 1 0 -8 57
CD
I—

10-8
CO
CO
ln

205 4 .8 8 5 1 x 1 0 -9
1

4 .8 8 5 1 x l0 -9 399 482
X

1 0 - l °
4 .2 5 6 1 X 1 0 "11 733 4 .4 0 0 6 x l0 -n 780 6 .9 8 1 5 X 1 0 -11 788

Table 5.4 Problem (5 .1 -a ). Using the trapezium

method fo r solving DDE w ith Hermite

in te rp o la tio n . S ta rtin g stepsize equal

to the in te rv a l o f in te g ra tio n .
101.

In te r v a l: [0? 1 0 ] [ 0 , 20] [ 0 , 4 0]

£ GE ND GE ND GE ND

10“ 2 5 .7 6 8 2 x 1 0 “ 8 43 7 .4 5 9 3 x 1 0 " 8 49 5 .9 2 9 0 x ]0 “ 8 61

10” 4 5 .7 6 8 2 x l0 "8 43 7 .4 5 93 x1 0-8 49 5 .9 2 9 0 x l0 "8 61

1 0-6 5 . 7 6 8 2 x l0 “ 8 43 7 .4 5 9 3 x l0 -8 49 5 .9 2 9 0 x l0 -8 61

It) - 8 2 . 3 9 2 8 x l0 ” 8 61 2 .4 9 4 9 x 1 0” 8 73 2 .8 6 7 4 x l0 -8 73

00
t"-
10” 10 1 . 8 8 9 1 x l0 ” 9 169 1 ,9 1 7 8 x l0 “ 9 1 .9 8 7 2 x l0 -9 193

Table 5.5 Problem (5 .1 -a ). Using the im p lic it

Runge-Kutta fo r DDE9 w ith Hermite

In te rp o la tio n . S ta rtin g stepsize h = 0.1.

In te r v a l:
o

o
10, 20]

1
1
£ GE ND GE ND GE ND

10-2 3 .2 1 3 5 x ] 0 " 6 7 1 . 1 3 2 4 x ]0 “ 5 7 2 . 0 8 7 0 x l0 “ G 7

10 -1* 3 .2 1 3 5 x l0 ” 6 7 1 .1324x10-5 7 2 .0 8 7 0 x1 0 -6 7

TO-6 3 . 2 1 3 5 x l0 ” 5 7 1 . 1 3 2 4 x l0 ” 5 7 1 .1 4 7 4 x lQ - 5 17

io -8 8 .2 0 1 7 x l0 -8 49 8 . 3 6 8 5 x l0 “ 8 59 1 .3 9 9 5 x l0 "7 69
J—
1'
o

2 .0 0 9 x l0 “ 9 187 2 . 1 6 1 9 x l0 “ 9 203 2 .0 3 7 8 x l0 -9 213


o

Table 5.6 Problem (5 .1 -a ). Using the im p lic it

Runge-Kutta method fo r DDE* w ith Hermite

in te rp o la tio n . S ta rtin g stepsize equal to

the in te rv a l o f in te g ra tio n .
In te r v a l:

Io—
to, 10]

o 1
[0, 203

Hf
J

j
e GE ND GE ND GE ND

10"2 9.8359x10“ 4 22 9.8369x10"^ 25 9.8369x10"^ 31

1G"4 6.7366xl0“ 5 52 6.7366xl0” 5 58 6.7366x10-5 61

lO” 6 4.0378xl0"6 234 4.0378x10” 5 243 4.0378xl0"8 249


10-8. 1.5294xl0"7 1203 1.5294xl0“ 7 1239 1.5294xl0"7 1248

10"10 6.0019x10" 9 5695 6 „ 0019x10“ 9 5883 6.0019x10" 9 5907

Table 5.7 Problem (5 .1 -b ). Using the trapezium

method fo r solving DDE5 w ith lin e a r

in te rp o la tio n .

1------- 1
In te r v a l: [0, 10] [Q, 203

0
e GE ND GE ND GE ND

10"2 2.9208x10-^ 22 7.9526X10"11 25 9.2664x10"4 31

10"4 - 6.7366x10 "5 52 6.7366xl0"5 58 6.7366xl0"5 61

10"6 4.0378xl0"5 234 4.0378xl0"6 243 4.0378x10 "6 249

10"8 1 .5294xl0-7 1203 1.5294X10” 7 1239 1.5294xl0"7 1248

10“ 10 6 .0019x10 ” 9 5695 6 o0019xl0"9 5883 6.0019xl0"9 5907

Table 5.8 Problem (5 .1 -b ). Using the trapezium

method/for solving DDE3 w ith Hermite

in te rp o la tio n .
103.

In te r v a l: [0, 10] [0, 20] [0, 40]

£ GE ND GE ND GE ND

io -2 1.0070x10” ^ 43 1.0070x10 -4 49 1.0070x10-^ 61

IO-1* 2.4753xl0“ 5 55 2.4753x!0~5 67 2.4753x]0“ 5 67

io -6 6.7754x]0“ 7 121 6.7754x]0“ 7 133 6.7754x10 "7 139

1 0 -8 1.1398x10-8 307 1.1398xl0“ 8 337 1 .1398x10-8 349

IO” 10 3.5629xlO“ 10 795 3.5629xlO“ 10 891 3.5629x10” 10 909

Table 5.9 Problem (5 .1 -b ). Using the im p lic it

Runge-Kutta fo r solving DDE.


r—

In te r v a l:
o

t0, 20]

O
o
^
""I

:
i
£ GE ND GE ND GE ND

IO” 2 1.1365x10” 2 191 1.2277x10” 2 381 1 .2796x10"2 771

10“ ^ 1.4535x10” ^ 291 1 .5798xl0” 4 481 1 .7157X10” 4 866

IO "6 4.8387xl0"7 796 1.5523x10-8 976 1.5523x1O"6 1366

io -8 1.9454x10” 9 2181 1.1704x10-8 2441 1.5409xl0-8 2831


i—>
o

4.1740x]0“ n 7232 4.5500xl0-n 7871 1 .5266xlO” 10 8231


i
o

Table 5.10 Problem (5 .1 -b ). Using the Kutta-Merson

method fo r solving DDE.


In te r v a l: [0 , 10] [0, 20) [0, 40]

! ^ e GE ND GE ND GE ND
■ <M
2,3138xl0"2 84 2.3138x10 -2 87 2.3138x]0~2 90

; i o - 1* 1.3729xl0"3 177 1.3729xl0“ 3 183 1.3729x10“ 3 186

10 6 7.1458x10*5 589 7.1458x10 -5 604 7.1458xl0~5 610

io-® 3.4293x1 O'6 2597 3.4293xl0“ 6 2645 3.4293x10 "6 2654


10-10 *

Table 5.11 Problem (5 .1 -c ). Using the trapezium

method fo r solving DDE w ith lin e a r

in te rp o la tio n .

In te r v a l: [0 3 10] [0 S 20] [0 , 40]

e GE ND GE ND GE ND

10-2 2.3136x]0“ 2 84 2.3136x10 -2 87 2.3136x]0“ 2 90

10-^ 1.3729xl0-3 177 1.3729xl0-3 183 1.3729xl0-3 186

IO "6 7 .1455xl0-5 586 7.1455x10 -5 601 7.1455xl0-5 607

10-® 3 .4 2 9 )xl0 "5 2552 3.4291x)0-6 2600 3.4291xl0-6 2606


o1

*
t-l
o

Table 5.12 Problem (5 .1 -c ). Using the trapezium

method fo r solving DDE w ith Hermite

in te rp o la tio n .
105.

In te r v a l: [0 , 10] [0, 20] [0, 40]

e GE ND GE ND GE ND

10” 2 4.5458x!0-3 135 4.5458xl0“ 3 141 4.5458x10-3 147


10-* 7 .2317xl0-5 173 7.2317X10"5 179 7.2317xl0-5 185

10“ 6 4.8718xl0“ 6 261 4.8718xl0“ 5 279 4.8718xl0“ 6 291

IO "8 5.1782xl0"6 475 5.1782xl0"6 505 5.1782*10 "6 517

ilO “ 10 5.2048x10 "6 1065 5.2048x10” 6 1137 5.2048x10“ 6 1155

Table 5.13 Problem (5 .1 -c ). Using the im p lic it

Runge-Kutta method fo r s o lv in g DDE.


1—1

t~*
O
In te r v a l: [0, 20]
o

o
e GE ND GE ND GE ND

10-2 8.6411xlO“ It 181 8.6411xlO "4 193 6.3 79 1xl0"3 235

10“ ^ 3.1097xl0-5 256 3.1097xl0-5 274 7.8425xl0-5 316

10“ 6 1 .0375xl0“ 6 510 1 .0375xl0-6 540 1 .0375xl0-6 576

10“ 8 2.5686x10~8 1360 2.5686xl0“ 8 1426 2.5686xl0“ 8 1456

IO” 10 2.8114xlO-10 4100 2.8114xlO-10 4250 2.8114xlO-10 4280

Table 5.14 Problem (5 .1 -c ). Using the Kutta-Merson

method fo r solving DDE.


106.

In te r v a l: [0 , 10] [0 S 20] [0, 40]

£ GE ND GE ND GE ND

10” 2 2.0227x]0-2 84 2.0227x]0"2 87 2 .0227xl0~2 90

io - 1* 1.5528x10-3 172 1.5528xl0-3 178 1.5528xl0“ 3 181

IO'6 7.4131x10-5 589 7.4131xl0” 5 604 7.4131xl0-5 610

10-8 3 .4522xl0-6 2648 3.4522xl0“ 6 2696 3.4522xl0“ 6 2705

IO-10 *

Table 5.15 Problem (5 .1 -d ). Using the trapezium

method fo r solving DDE w ith lin e a r

in te rp o la tio n .
r“—

5_o 1
i

In te r v a l: [ 0 , 20] [ 0 9 4 0]
o

£ GE ND GE ND GE ND

10~2 2 .0 2 2 3 x l0 -2 84 2 . 0 2 2 3 x l0 " 2 87 2 .0 2 2 3 x]Q “ 2 90

10~4 1 .5 5 2 8 x 1 0 -3 172 1.5 5 28 x1 0-3 178 1 .5528x10-3 181

1Q"8 7 .4 1 2 2 x 1 0 " 5 582 7 .4 1 2 2 x l0 -5 598 7 .4 1 2 2 x1 0 -5 604

IO "8 3 . 4 5 1 7 x l0 ” 6 2555 3 .4 5 1 7 x l0 “ 6 2603 3 . 4 5 1 7 x l0 ” 6 2609

1 0 " 10 *

Table 5.16 Problem (5 .1 -d ). Using the trapezium

method fo r so lving DDE w ith Hermite

in te rp o la tio n .
107.

In te r v a l: [0, 10] [0, 20] [0, 40]

£ GE ND GE ND GE ND

TO'2 4.0057x]0"2 125 4.0057x10 "2 131 4.0057x10 "2 137

io - 1* 2.2729x10“ ^ 163 2.2729x10"^ 169 2.2729x10'* 175

IO '6 2.8902x]0“ 6 255 2.8902x10” 6 273 2.8902x1O"6 285

TO'8 1.0867xl0“ 7 475 1.0867xl0"7 505 1.0867xl0"7 517


O1
o

3.1965xl0“ 9 1059 3.1965x10"9 1131 3.1965x]0"9 1155

Table 5.17 Problem (5 .1 -d ). Using the im p lic it

Runge-Kutta fo r solving DDE.

In te r v a l: [ 0 , 10] [ 0 , 20] [ 0 S 40]

£ GE ND GE ND GE ND

10“ 2 3 . 3 6 3 2 x l0 “ 3 171 3 .3 6 3 2 x ]0 '3 183 6 .2 6 2 3 x l0 "3 225


•F
oi

2 .2 6 2 2 x ]0 “ 5 261 2 .2 6 2 2 x ] 0 “ 5 279 7 .6 8 8 3 x 1 0“ 5 321

10“ 6 1 . 0 3 1 4 x l0 “ 6 532 1.0 3 14 X 1 0"6 562 1 .0 3 1 4 x l0 "6 598

TO’ 8 1 .2 2 1 6 x 1 0-8 1431 1 .2 2 1 6 x ]0 ~ 8 1497 1 .2 2 1 6 x l0 "8 1533


f—i
o
O
1

1 .2 1 8 1 x lO “ 10 4459 1 .2 1 8 1 x ]0 “ 10 4621 1 .2 1 8 1 x lO “ 10 4657

Table 5.18 Problem (5 .1 -d ). Using the Kutta-Merson

method f o r solving DDE.


Delay term evaluated Modified Hermite
e xa ctly Hermite in te rp o la tio n
£ in te rp o la tio n

GE ND GE ND GE ND

10"3 4.1667xl0’ 3 151 4.1667xl0"3 151 4.1667xl0~3 173

IO"6 4.0690x10“ 6 411 4,0690x10"6 411 4.0690xl0“ 6 471


•jq-9 3.9737xl0"9 871 3.9737x10“ 9 871 3 .9737x]0-9 ion

Table 5.19 Problem (5 .2 ). Using the Kutta-Merson

method fo r solving DDE, w ith mesh

points chosen by the method.

Delay term evaluated M odified Hermite


e x a ctly Hermite in te rp o la tio n
£ in te rp o la tio n

GE . ND GE ND GE ND

10“ 3 3.1467xl0"3 156 3.1250xl0“ 3 136 3.1250x]0“ 3 158

10“ 6 3.0865x10 -6 361 3.0518xl0"6 376 3.0518xl0"6 434

10“ 9 3.0746xl0“ 9 976 2.9802x10~9 836 2.9802x10 "9 974

Table 5.20 Problem (5 .2 ). Using the Kutta-Merson method

fo r solving DDE, w ith d e riv a tiv e d is c o n tin u ity

points forced to be mesh p o in ts .


Del ay term evaluated
e xa ctly Hermite in te rp o la tio n
e
GE ND GE ND

10” 3 7.6724xl0~5 59 6.3382x10” * 59

IO” 6 2.4269xl0"7 209 2.5214 x l0 "7 209

IO"9 2.7417xlO“ 10 1163 8.6987xlO“ 10 1163

Table 5.21 Problem (5 .3 ). Using the K utta-

Merson method fo r solving DDE.

Delay term evaluated


e xa ctly Hermite in te rp o la tio n
£
GE ND GE ND

10“ 3 2.3406x10“ * 77 8.9346x10-* 77


IO"6 1.3794x10“ 7 492 3.6375xl0-7 492

10“ 9 3.3881xlO“ 10 1998 8.7088xl0"10 1998

Table 5.22 Problem (5 .4 ). Using the K utta -

Merson method fo r s o lv in g DDE.


no.

Delay term evaluated


e xa c tly Hermite in te rp o la tio n
E

GE+ ND GE+ ND

1 0 “ 3 8.8459xl0-6 175 4.6002xl0"5 175


1 0 - 6
3.4708xl0” 8 602 5.3247xl0“ 5 602
1 0 “ 9 1 .3762xlO“ 10 2421 1.3480x1 O’ 7 2421

Table 5.23 Problem (5 .5 ). Using the K utta-

Merson method fo r solving DDE.

t R elative global d is c re tiz a tio n e rro r.

e GE ND

10"3 2.3514x10” ^ 91

10” 6 1.3793x10” 7 584

10“ 9 3.3881x1 O'*10 2389

Table 5.24 Problem (5 .6 ). Using

the Kutta-Merson method

fo r solving DDE.
Delay term evaluated
Hermite in te rp o la tio n
e
e xa ctly

GE ND GE ND

IO "3 5 ,1 3 4 0 x 1 0 “ ^ 31 5 .2 5 4 3 x 1 0 “ 5 31

ID "6 2 .9 8 4 3 x 1 0 " 7 71 6 . 5 4 8 9 x l0 ~ 7 71

IO "9 1 . 3 3 7 0 x lO " 10 391 1.6096x1 CT10 391

Table 5.25 Problem (5 .7 ). Using the K utta -

Merson method f o r s o lvin g DDE. Mesh

points are chosen by the method.

Delay term evaluated


Hermite in te rp o la tio n
e xa ctly
£
GE ND GE ND

10~3 2.2431xlO“ 5 31 3.4644x10“ 5 31


10-6 2.9843x10“ 7 71 6.5489xl0“ 7 71
l 0-9 1.3370xl0“ 10 391 1.6096xl0~10 391

Table 5.26 Problem (5 .7 ). Using the K utta-

Merson method fo r solving DDE.

D eriva tive d is c o n tin u ity po ints are

forced to be mesh p o in ts .
Delay term evaluated
Hermite in te rp o la tio n
e e xa ctly

GE ND GE ND

10-3 2.5977xl0“ 2 71 2.6668xl0“ 2 71

IO "6 2 .6197xl0-5 316 2.6214xl0-5 316

IO "9 1.0635x10 "8 1076 1.0637xl0-8 1076

Table 5.57 Problem (5 ,8 ). Using the K utta-

Merson method fo r solving DDE.

Mesh points are chosen by the method.

Delay term evaluated


Hermite in te rp o la tio n
£ exa ctly

GE ND GE ND

IO "3 1 .1733x10" 3 67 1 . 1733xlG-3 56

IO "5 4.7077x10” 6 185. 4.7077xlQ“ 6 156


10-9 1 .0678xl0“ 9 1083 1.0678xT0“ 9 906

Table 5.28 Problem (5 f 8J . . ..Using the K utta -

Merson method fo r so lvin g DDE.

D e riv a tiv e d is c o n tin u ity p o in ts are

forced to be mesh p o in t? .
Delay term eveil uated
exa ctly Hermite in te rp o la tio n
e
. GE ND GE ND

IO’ 3 i.4 0 7 ix io - 1+ 59 1 .9658x10~^ 59


■JO-6 6.3735xl0“ 7 215 6.3735xl0"7 215

10~9 5.1470xlO™10 1148 3.7861xlO“ 10 1148

Table 5.29 Problem (5 .9 ). Using Kutta-Merson

method fo r solving DDE.


Delay term evaluated
e x a c tly Hermite in te rp o la tio n
£
GE ND GE ND

10“ 3 2.1790xl0“ 5 31 2.1790xl0” 5 31

10“ 6 1 .1089x10“ 7 106 1.1089x10” 7 106


10-9 1.4035x10“ 10 616 1.4035xl0"10 616

Table 5.30 Problem (5 .1 0 ). Using K utta-

Merson method fo r so lvin g DDE.

Mesh points are chosen by the

method„

Delay term evaluated


e xa ctly Hermite in te rp o la tio n
e
GE ND GE ND

IO” 3 2.1790xl0“ 5 ' 31 2.1790xl0“ 5 31

10“ 6 1.1089xl0"7 111 1.1089xl0“ 7 111

10“ 9 1 .4035xl0“ 10 621 1.4035x1O"10 621

Table 5.31 Problem (5 .1 0 ). Using K utta -

Merson method f o r so lvin g DDE.

D e riva tive d is c o n tin u ity po ints

are forced to be mesh p o in ts .


115.

BIBLIOGRAPHY

(1) ALLINGER, G., AND HYRON, H. (1976)

’’Approximate so lu tio n s o f d iffe r e n tia l equations w ith d e viating

arguments",

SIAM 0. Numer. A n a l., Vol. 3, No. 3, pp412-426.

(2) BAILEY, H.R., AND WILLIAMS, M.Z. (1966)

"Some re s u lts o f the d iffe r e n tia l d iffe re n c e equation


N
*<t) = I A . x ( t - T . ) " ,
1=0 1
d. Math. Anal. A ppl. 15, pp569-587.

(3) BAKER, C.T.H. (1976)

" I n i t i a l value problems fo r del a y - d iffe r e n tia l equations",

Modern numerical methods fo r ordinary d if f e r e n t ia l equations,

G. Hall and J.M. Watt (E d .), : r

Clarendon Press, Oxford.

(4) BARWELL, V.K. (1974)

"On the asym ptotic behaviour o f the s o lu tio n o f a d iffe r e n tia l

d iffe re n c e e q u a tio n ",

UTILITAS MATHEMATICA, V ol. 6, pp!89-194.

(5) BARWELL, V.K. (1975)

"Special s t a b ilit y problems fo r fu n c tio n a l eq uatio ns",

BIT 15, p p l30-135.

(6) BELLMAN, R.E. (1961)

"On the computational s o lu tio n o f d if fe r e n tia l d iffe re n c e

e q u a tio n s",

J . Math. Anal, and i t s Appl. 2, p p l08-110.

(7) BELLMAN, R .E., BUELL, O.D., AND KALABA, R.E. (1965)

"Numerical in te g ra tio n o f a d iffe r e n tia l equation w ith a decreasing

tim e -la g ",

Communication, ACM, Vol. 8, pp227-228.

(8) BELLMAN, R.E., BUELL, J .D ., AND KALABA, R.E. (1966)

"Mathematical experimentation in tim e -la g m o dula tion",


Communication, ACM, Vol. 9, pp752~754.

(9) BELLMAN, R.E., AND COOKE, K.L. (1963)

"D iffe re n tia l-D iffe re n c e Equations",

Academic Press, N.Y.

(10) BELLMAN, R.E., AND COOKE, K.L. (1965)

"On the computational s o lu tio n o f a class o f fu n c tio n a l

d if f e r e n t ia l equations",

J.-M ath. Anal, and i t s Appl. 12, pp495-500.

(11) BELLMAN, R .E ., AND KOTKIN, B. (1962)

"On the numerical s o lu tio n o f a d iffe r e n tia l- d iffe r e n c e equation

a ris in g in a n a ly tic number th e o ry".

Math. Comp. 16, pp473-475.

(12) BOFFI, V ., AND SCOZZAFAVA, R. (1967)

"A f i r s t order lin e a r D iffe re n tia l-D iffe re n c e Equations w ith

N Delays",

J . Math. Anal. Appl. 17, pp577-589.

(13) BRAYTON, R.K., AND WILLOUGHBY, R.A. (1967)

"On the numerical in te g ra tio n o f a symmetric system o f d iffe re n c e -

d iffe r e n tia l equations o f neutral ty p e ",

J. o f Math. Anal. Appl. 18, ppl82-189.

(14) BUTCHER, J.C . (1964)

" Im p lic it Runge-Kutta processes",

Math. Comp. 18, pp50-64.

(15) CHARTRES, B ., AND STEPLEMAN, R. ( 1 9 7 1 _

"Convergence o f d iffe re n c e methods fo r i n i t i a l and boundary value

problems w ith discontinuous data",

Maths, o f Comp., V ol. 25, No. 116, pp729-732.

(16) CHARTRES, B ., AND STEPLEMAN, R. (1975)

"Order o f convergence o f lin e a r m u lti step methods fo r fu n c tio n a l

d iffe r e n tia l equations",

SIAM J. Numer. A n a l., V ol. 12, No. 6, pp876-886.


117.
(17) COOKE, K.L, AND LIST, S.E. (1972)

"The numerical s o lu tio n o f in te g ro -d iffe re n tia l equations w ith

re ta rd a tio n ” 3

Tech. Rept. No. 72-4s Dept, o f E le c tric a l Engineering, U n iv e rs ity

o f Southern C a lifo rn ia , LOS ANGELES.

(18) CRYER, C.W. (1972)

"Numerical method fo r fu n c tio n a l d iff e r e n tia l equations".

Delay and Functional D iffe re n tia l Equations, S. Schmitt (E d i),

Academic Press, N.Y.

(19) CRYER, C.W. (1973)

"H ig h ly -s ta b le m u lti step methods fo r retarded d iffe r e n tia l

eq uatio ns",

Computer Science D ept., Tech. Rept. No. 190, U n iv e rs ity o f

Wisconson, Medison. Also SIAM J. Numer. A n a l., Vol. 11, No. 4,

(1974), pp788-797.

(20) CRYER, C.W., AND TAVERNINI, L. (1972)

"The numerical so lu tio n o f V o lte rra Functional D iffe r e n tia l

Equations by E u le r's method",

SIAM J. Numer. A n a l., V ol. 9, No. 1, ppl05-129.

(21) DAHLQUIST, G. (1956)

"Convergence and s t a b ilit y in the numerical in te g ra tio n o f

ordina ry d iffe r e n tia l equations",

Math. Scand. 4, pp33-53.

(22) DE NEVERS, K ., AND SCHMITT, K. (1971)

"An a p p lic a tio n o f the shooting method to boundary-value problems

f o r second order Delay equations",

J . o f Math. Anal, and A p p lic . 36, pp588-597.

(23) DRIVER, R.D. (1961)

"Existence theory fo r a d e la y -d iffe re n tia l system",

C ontribu tions to D iffe re n tia l Equations, Vol. 1, No. 3, pp317-335.

(24) DRIVER, R.D. (1962)


"Existence and s t a b ilit y o f s o lu tio n o f a d e la y -d iffe re n tia l

system",

Archive fo r Rational Mech. A nal., V ol. 10, No. 5, pp401-426.

(25) EL'SGOL'TS, L.E. (1966)

"In tro d u c tio n to the theory o f d iffe r e n tia l equations w ith

d e via tin g argument"

Holden-Day, N.Y. (Translated from Russian).

(26) EL'SGOL'TS, L .E ., AND NORKIN, S.B. (1973)

"In tro d u c tio n to the theory an d.a p p lica tio n o f d iffe r e n tia l

equations w ith d e via tin g argument",

Academic Press, N.Y. (Translated from Russian).

(27) ENRIGHT, W., BEDET, R., FARKAS, I . , AND HULL, T .E . (1974a)

"Test re s u lts on i n i t i a l value methods fo r n o n - s tiff ordinary

d iffe r e n tia l equations",

Tech. Rept. 68, Dept, o f Computer Science, U n iv e rs ity o f Toronto.

(28) ENRIGHT, W., HULL, T .E ., AND LINDBERG, B. (1974b)

"Comparing numerical methods fo r s t i f f systems o f ODEs",

Tech. Rept. 69, Dept, o f Computer Science, U n iv e rs ity o f Toronto.

(29) FELDSTEIN, A. (1964)

"D is c re tiz a tio n methods fo r retarded ord in a ry d iffe r e n tia l

eq u a tio n s",

Ph.D. th e s is . Dept, o f Maths., U n ive rsity o f C a lifo rn ia , Los

Angeles.

(30) FELDSTEIN, A ., AND GOODMAN, R. (1973)

"Numerical s o lu tio n o f ordinary and retarded d iffe r e n tia l

equations w ith discontinuous d e riv a tiv e s ",

Numer. Math. 21, ppl-13.

(31) FELDSTEIN, A ., AND GRAFTON,■ C.K. (1968)

"Experimental mathematics: an a p p lic a tio n to retarded ordinary

d iffe r e n tia l equations w ith in f in it e la g ".

Proceedings - 1968 ACM National Conference.


119.
(32) FOX, L ., MAYERS, D .F., OCKENDON, J .R ., AND TAYLER, A.B. (1971)

"On fu n ctio n a l d iffe r e n tia l equations",

J. In s t. Maths. A pplies. 8, pp271-307.

(33) GEAR, C.W. (1971)

"Numerical In itia l- V a lu e Problems in Ordinary D iffe r e n tia l

Equations",
\

P re n tic e -H a ll, Englewood C lif f s , N.J.

(34) GOODMAN, R ., AND FELDSTEIN, A. (1973)

"Round-off e rro r fo r retarded o rdina ry d iffe r e n tia l equations:

a p r io r i bounds and estim ates",

Numer. Math. 21, pp355-372.

(35) GRIMM, L. (1971)

"Existence and uniqueness o f no n -lin e a r neutral d if fe r e n tia l

e q u a tio n s",

B u lle tin , AMS (3) 77, pp374-376.

(36) HALANAY, A. (1966)

" D iffe r e n tia l Equations, S ta b ility , O s c illa tio n , Time la g s ".

Academic Press, N.Y.

(37) HALANAY, A ., AND YORKEE, J.A. (1971)

"Some new re s u lts and problems in the theory o f d iffe re n tia l-d e la y

e q u a tio n s",

SIAM Review, Vol. 13, No. 1, pp55-80.

(38) HALE, J. (1971)

"Functional D iffe r e n tia l Equations",

S pringer, N.Y.

(39) HALE, J . , AND CRUZ, M. (1970).

" S t a b ilit y o f fu n c tio n a l d iffe r e n tia l equations o f neutral typ e ",

J . o f D iffe r e n tia l Equations, V ol. 7, pp334-355.

(40) HALL, G., AND WATT, U.M. (Ed.) (1976)0 1

"Modern numerical methods fo r ordina ry d iffe r e n tia l equations",

Clarendon Press, Oxford.


120.

(41) HENRI Cl D P. (1962)

"D iscrete V ariable Methods in Ordinary D iffe r e n tia l E quations",

John VJiley, N.Y.

(42) HILL, D.R. (1973)

"An approach to the numerical so lu tio n o f d e la y -d iffe re n tia l

e q u a tio n s",

Ph.D. Thesis, P ittsburgh U n iv e rs ity .

(43) HILL, D.R. (1974)

"A new class o f one-step methods fo r the s o lu tio n o f V o lte rra

fu n c tio n a l d iffe r e n tia l equations",

BIT, 14, pp298-305.

(44) HULL, T .E ., ENRIGHT, W.H., FELLEN, B.M., AND SEDGWICK, B.M. (1972)

"Comparing numerical methods fo r ord in a ry d iff e r e n tia l equations",

SIAM J. Numer. Anal. 9, pp603-637.

(45) HUTCHISON, J. (1971)

"F in ite d iffe re n ce s o lu tio n to delay d if f e r e n t ia l equations",

Ph.D. Thesis, Rensselaer Polytechnic I n s t it u t e , Troy, New York.

(46) ISAACSON, E ., AND KELLER, H.B. (1966)

"A nalysis o f numerical methods",

W iley, N.Y.

(47) KEMPER, G.A. (1972)

"Lin ear m u ltiste p methods f o r a class o f fu n c tio n a l d if fe r e n tia l

e q u a tio n s",

Numer. Math. 19, pp361-372.

(48) KEMPER, G.A. (1975)

"S pline fu n c tio n approximation fo r s o lu tio n s o f fu n c tio n a l

d iffe r e n tia l equations",

SIAM J . Numer. A n a l., V ol. 12, No. 1.

(49) KROGH, F.T. (1970)

"On te s tin g a subroutine fo r the numerical in te g ra tio n o f

ord in a ry d iffe r e n tia l equations",


121

Je t Propulsion Lab., Tech. Memo. 217.

(50) LAMBERT, O.D. (1973)

"Computational Methods in Ordinary D iffe r e n tia l E quations",

Wi 1e y.

(51) LAPI DUS, L ., AND SEINFELD, d.H. (1971).

"Numerical S olution o f Ordinary D iffe re n tia l Equations",

Academic Press, N.Y.

(52) MYSHKIS, A.D. (1951)

"General theory o f d iffe r e n tia l equations w ith a retarded argument

American Mathematical Society T ra n s la tio n s , V ol. 55.

(53) MYSHKIS, A .D ., AND EL’ SGOL'TS, L. (1967)

"The sta te and problems o f the theory o f d iffe r e n tia l equations

w ith perturbed arguments",

RUSSIAN MATHS SURVEYS (2) 22, p p l9-57.

(54) NEVES, K.W. (1975)

"Automatic in te g ra tio n o f fu n ctio n a l d iffe r e n tia l equations: An

approach and A lg o rith m ",

ACM Transaction on Mathematical Software, V ol. 1, No. 4.

(55) NEVES, K.W,, AND FELDSTEIN, A. (1973)

"C ha racte risatio n o f jump d is c o n tin u itie s fo r state dependent

delay d iffe r e n tia l eq uatio ns",

Presented a t SIAM National Meeting, Hampton, Va.

(56) OGUZTORELI, M.N. (1966)

"Tim e-lag co n tro l systems",

Academic Press.

(57) OPPELSTRUP, J . (1973)

"D elay-1, A program fo r in te g ra tin g systems o f d e la y -d iffe re n tia l

e q u a tio n s",

Report TRITA-NA-7311, Royal In s titu te o f Technology, Stockholm,

Sweden.

(58) SCHMITT, K. (Ed.) (1973)


1220

"Delay and Functional D iffe re n tia l Equations and Their A p p lic a tio n s ".

Academic Press.

(59) TAVERNINI, L. (1969)

"Numerical methods fo r V o lte rra fu n ctio n a l d if fe r e n tia l equations",

Ph.D. Thesis, U n iv e rs ity o f Wisconsin a t Madison.

(60) TAVERNINI-, L. (1971)

"One-step methods fo r the numerical s o lu tio n o f V o lte rra fu n c tio n a l

d iffe r e n tia l equations",

SIAM 0. Numer. A n a l., Vol. 8, No. 4.

(61) THOMPSON, R .J. (1968)

"On some fu n ctio n a l d iffe r e n tia l equations: Existence o f solu tions

and d iffe re n c e ..approximations",

SIAM J. Numer. A n a l., Vol. 5, pp475-487.

(62) TSOI, A.C. (1975)

" E x p lic it s o lu tio n o f a class o f d e la y -d iffe re n tia l equations",

' i n t . * J. C o n tro l, V ol. 21, No. 1, pp39~48.

(63) VAN de LUNE, 0 ., AND WATTEL, E. (1969)

"On the numerical s o lu tio n o f a d iffe r e n tia l d iffe re n c e equation

a ris in g in a n a ly tic number th e o ry ",

Math. Comp. 23, pp417-421.

(64) WIEDERHOLT, L.F. (1970)

"Numerical in te g ra tio n o f d e la y -d iffe re n tia l equations",

Ph.D. th e s is , Univ. o f Wisconsin, Madison.

(65) WIEDERHOLT, L.F . (1976)

" S t a b ilit y o f m u ltis te p methods fo r d e la y -d iffe re n tia l equations".

Mathematics o f Computation, Vol. 30, No. 134, pp283-290.

(66) ZVERKIN, A.M., KEMENSKI, G.A., N0RKIN, S .B ., AND EL'SGOL'TS, L.E. (1962)

" D iffe r e n tia l equations w ith a perturbed argument",

Russian Math. Surveys, Vol. 17, pp61-146.

(67) ZVERKINA, T.S. (1964)

"A m o d ifica tio n o f f i n i t e d iffe re n ce methods fo r in te g ra tin g

You might also like