57
AN ITERATIVE METHOD FOR SOLVING NONLINEAR SYSTEMS OF EQUATIONS Item Type text; Dissertation-Reproduction (electronic) Authors Bryan, Charles Allen, 1936- Publisher The University of Arizona. Rights Copyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction or presentation (such as public display or performance) of protected items is prohibited except with permission of the author. Download date 28/08/2018 18:51:29 Link to Item http://hdl.handle.net/10150/284455

This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

Embed Size (px)

Citation preview

Page 1: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

AN ITERATIVE METHOD FOR SOLVINGNONLINEAR SYSTEMS OF EQUATIONS

Item Type text; Dissertation-Reproduction (electronic)

Authors Bryan, Charles Allen, 1936-

Publisher The University of Arizona.

Rights Copyright © is held by the author. Digital access to this materialis made possible by the University Libraries, University of Arizona.Further transmission, reproduction or presentation (such aspublic display or performance) of protected items is prohibitedexcept with permission of the author.

Download date 28/08/2018 18:51:29

Link to Item http://hdl.handle.net/10150/284455

Page 2: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

This dissertation has been 63—6284 microfilmed exactly as received

BRYAN, Charles Allen, 1936-AN ITERATIVE METHOD FOR SOLVING NON-LINEAR SYSTEMS OF EQUATIONS.

University of Arizona, Ph.D., 1963 Mathematics

University Microfilms, Inc., Ann Arbor, Michigan

Page 3: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

AN ITERATIVE METHOD FOR SOLVING

NON-LINEAR SYSTEMS OF EQUATIONS

by £

Charles,Bryan

A Dissertation Submitted to the Faculty of the

DEPARTMENT OF MATHEMATICS

In Partial Fulfillment of the Requirements For the Degree of

DOCTOR OF PHILOSOPHY

In the Graduate College

THE UNIVERSITY OF ARIZONA

19 6 3

Page 4: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

THE UNIVERSITY OF ARIZONA

GRADUATE COLLEGE

I hereby recommend that this dissertation prepared under my

direction by Charles Bryan

entitled An iterative method for solving non-linear systems

of equations

be accepted as fulfilling the dissertation requirement of the

degree of TTnnt.ny of Philnanphy

/ ) / A { e , / 3 tion D/rfrctor 7Date/ **7 / DissertationDfrector

/ /

After inspection of the dissertation, the following members

of the Final Examination Committee concur in its approval and

recommend its acceptance:*

^ Q. .SgAui.u x/—- 3 ) S Q / 6 3

a/j7y j ^ OR^F)

6.2

•This approval and acceptance is contingent on the candidate's adequate performance and defense of this dissertation at the final oral examination. The inclusion of this sheet bound Into the library copy of the dissertation is evidence of satisfactory performance at the final examination.

Page 5: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

STATEMENT BY AUTHOR

This thesis has been submitted in partial fulfill­ment of requirements for an advanced degree at The Univer­sity of Arizona and is deposited in The University Library to be made available to borrowers under rules of the Library.

Brief quotations from this thesis are allowable without special permission, provided that accurate acknowl­edgment of source is made. Requests for permission for extended quotation from or reproduction of this manuscript in whole or in part may be granted by the head of the major department or the Dean of the Graduate College when in their judgment the proposed use of the material is in the inter­ests of scholarship. In all other instances, however, permission must be obtained from the author.

SIGNED:.

Page 6: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

ACKNOWLEDGMENTS

The author would like to express his appreciation

to Dr. H. Melvin Lieberstein, whose kindly guidance has

helped to sharpen many of the results of this paper. This

paper was prepared under the sponsorship of the National

Science Foundation Grant NSF-G16249.

ii

Page 7: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

TABLE OF CONTENTS

Title: An Iterative Method for Solving Non-Linear

Systems of Equations.

Page

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

2. Some consequences for finite dimensional Banach spaces 4

3. Conditions for convergence in Tchebycheff norm . . 13

4. Choice of to 24

5. Conditions in a deletion space 29

6. Applications 43

Bibliography 50

iii

Page 8: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

AN ITERATIVE METHOD FOR SOLVING

NON-LINEAR SYSTEMS OP EQUATIONS

1. Introduction

The purpose of this paper is to establish suffi­

cient conditions to guarantee that the iteration

(11) (k+1) (k) i* 1 '""Xn > ji = 1,2,...,n7 u - } x i - x i w : r u n r n ^ u = 0 , 1 , 2 , . . . ] »

i, i 1 ' * * *' n '

•\ f u> a real number, and f_. ° i , will converge to a solu-1,3 TT7

J tion (unique in some neighborhood) of the system

(l»2) f (Xj ,Xg t • • • t) = 0 , i = 1,2,... ,n

where the functions f^ are real-valued functions of real

variables x., j = 1,2,...,n, and to establish computable 3

error bounds for the iterates. This iteration method was

first proposed by H. M. Lieberstein 193. The method is a

generalization of the Jacobi methodfl23 for solving a

system of linear equations. Along with this method Lieber­

stein also proposed the method

f ^(k+l) Jk+D Jk) (kK ,, (k+1)_ (k) *^*1 >xj »* * * »xn > (1.3) x. - x. -u>- ( k + 1 ) (k+1) (k)TkT7

i ,i 1 '•••'xi_l »xi »•••»xn )

which is a generalization of D. Young's overrelaxation for

linear systems [l3j.

1

Page 9: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

2

The advantage of methods (l.l) and (1.3) over

Newton's method is that the correction formulas are expli­

cit, whereas in Newton's method a linear system of equations

must be solved at each step of the iteration; this often

leads to approximating the solutions of the linear systems

by iterative techniques and hence applying iterations within

an iteration and the calculations necessarily become lengthy

and time consuming even when carried out on high speed com­

puters. Moreover in many cases, such as finite difference

analogs for differential equations, all calculations are

based on a single correction formula. Because of these

features, the method proves to be very efficient for use on

computers.

In section 2 the problem is reformulated in terms

of finite dimensional Banach spaces whereby the problem be­

comes that of finding conditions under which a certain

mapping has a fixed point obtainable by Picard iteration.

In section 3, using the comparison method developed by L. V.

Kantorovic [63 and analysis techniques similar to those • •

used by J. Schroder 111} , sufficient conditions are ob­

tained for the case when the space has a Tchebycheff norm.

In section 4 it is shown that under an appropriate defini­

tion of "best" the optimum value of co in (l.l) is one,

whereas it is well-known that this optimum is not neces­

sarily one in (1.3) when the f^ are linear. In section 5

Page 10: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

3

the conditions are generalized to apply to so-called dele­

tion spaces which include finite dimensional spaces with

Tchebycheff and ^ norms, thus yielding conditions other

than those of section 3. Uniqueness of the solution is

also shown here. Finally, in section 6 the application of

method (l.l) to finite difference equations is considered.

By using (1.3), which experimentally seems to converge

quite rapidly, to obtain a point at which the theory be­

hind (l.l) is applicable, we obtain by the theory of (l.l)

the existence of a solution of a system of finite difference

equations for a non-linear two-point boundary value problem.

A maximum error bound for the approximation is obtained of

the order of magnitude of 10~^ where the solution itself is

of the order of magnitude of unity.

Page 11: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

2. Some consequences for finite dimensional Banach spaces.

Let X be a real n-dimensional Banach space with a

basis e.^ , j = l,2,...,n. Corresponding to each element

e. of the basis consider the linear projection operator P. J J mapping X into X such that

(2.1) P4U) -Pj( - orj.j

and to each linear operator A mapping X into X there

corresponds the operator DA such that

(2.2) DA(x) = ZJ=1 [A(P.(X))J.

Since every linear operator A mapping X into X can be

represented as a matrix (a^.) it follows that

PiUWl - a^U)

and

C DA(X) 3 = tA(P.(x))] = Xi=l aiiPi(Pi(x))

= a..P.(x) IX 1 '

or equivalently, DA can be represented as a diagonal matrix

with the same diagonal as A. Finally if we consider any map

ping F(x) mapping X into X we may examine the mapping for

its differential properties. If F has a Frechet differen­

tial at x it is represented by

F'(x)fh] = lim ILx+fh)-F(x) PtO A

where x and h are elements of X and (5 is from the scalar

field C2» Definit ten, p. 490; 10, Thm. 1 , p. 185 J •

Page 12: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

With "these definitions and remarks in mind we re­

formulate the problem of convergence of (l.l). Let X be an

n-dimensional Banach space and F a Frechet differentiable

operator on X into X. Consider the mapping

(2.3) T(x) = x -WD^F'U^FU)]

where F'(x) is the Frechet derivative of F(x) aiid we write

for short

D-1F1 (x) = (DF'U))"1.

If elements of X are represented by column vectors and F(x)

by the column vector

^f (x11 t • /*!<*> \

f 2(x)

F(x) =

;n(x)/

'V\

f2 xl*x2'* n )

\ fn(xi»x2»•••»xn)/ then the Frechet derivative F'(x) is represented by the

Jaeobian matrix

F*(x) =

(fl,l(x)

f2,l*X

fl,2^

f2,2^

\fn,l*X* fn,2 X*

f2,nW

r UV n,nx 1

and hence

Page 13: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

DF'(x)=

f2,2^

If D F*(x) exists - that is, if f . . (x) =j= 0 for i = 1,2,..., 1»1

then.

V 0 • (x)/ n,nx '!

D_1F'(x)=

fl,lU* 0

1

2,2(x)

0

0

\

f—UV n,nx 'I

It is then seen that the iteration (l.l) is equivalent to

Picard iteration on T, where T is defined as in (2.3). We

shall use the following lemmas:

Lemma 1. Let A and B be any two linear operators on an

n—dimensional Banach space X into X. Then

DA -DB = D(A - B).

Proof• D(A-B)(x) » -J2=l I' i{(A-B)(P1(x))} =

- 1 U p i f A ( p i 2 i - i

= DA(x) - DB(x).

Lemma 2. Let X be an n-dimensional Banach space and F an

operator on X into X which is twice Frechet differentiable.

Page 14: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

Then

(2.4) D[F»(x)(h1,h2)] = (DP*(x)Lh2l)'[h^I ,

where (DF1 (X) [h2L ) 1 .is the Frechet derivative of DF' (x) [h^L

with respect to x, F"(x)(h^,h2) .is the Frechet differential

of F'(x )[hj] with respect to x, and

D[F» (x) (hx ,h2)] = ^ 1Pi F"(x)(h1,Pi(h2)].

Proof:

(DP'(x) h )' [h ] = ai™ -±- JU"(x+(>h )[h 1- DF'(x)[hJ

—v H^J

= Zi=lPi{F"(x)(ri(h2>'hl)]

and since F"(x) (P±(h2) ,1^) = F" (x) (h^ ,P^(h2))}

(DF1 (x) £h^J)'[hjl = 5 i=lPilF"(x) (hl'Pi(h2) ]l

= Dp-'UMh^)] .

Two properties of the spaces under consideration

will be assumed and spaces with these properties called

deletion spaces.

Definition: An n-dimensional Banach space X will be called

a deletion space if lor all linear operators A on X into X

the following inequalities are satisfied:

(2.5) IIDAII 5^ ||AII

(2.6) IjA-DAll $ IJAll.

Lemma 3. Finite dimensional Banach spaces with the Tcheby-

cheff and norms are deletion spaces.

Proof:

Case 1. Tchebycheff norm.

Page 15: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

8

When A is represented by the matrix (a j)

Thus

and

»•411 = 14iin 2j=l

= 4n - "All

lU-DAll = ?" ->n , .la. I £ IIAlt* liiin <£-j=1,j*i' ij '

Case 2. ^ norm.

Under the matrix representation for A

and in a manner similar to that of Case t,l conditions (2.5)

and (2.6) are satisfied.

It should be noted that Si spaces, l<p<-*°, are not

in general deletion spaces as can be seen by considering

the following example.

Let X be E with the 5^ norm. If A is a linear

operator on X then

II All = (p|ATA|)*

where piAt is the modulus of the eigenvalue of A with

largest modulus Cl2l. Let

1 -2 2 \

A = -2 1 2

\ 2 2 1

then

>T f> |ATA I = 9

/

Page 16: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

and

p| (A—DA)T(A-DA)| = 16.

Therefore IIA-DAll > IIAll and hence X is not a deletion space.

Ve will rely heavily on two well-known theorems of » •»

functional analysis which are stated here without proof,

i) Mean value Theorem L4, p.155^. Let X and Y be real

norraed linear spaces and y = T(x) be an operator with domain

DCX and range ECY where D is a convex open set. Let T be

Fxechet dif'f erentiable on D. Then

(2.8) l\ T(x+h)-T(x)II ± l/h»sup[llT'(x+th)ll ,Ottil].

(ii) Let X be a Banach space and A a linear operator

from X into X. If llAll^-1, then (I—A) exists and has domain

X where I is the identity mapping. Moreover C 8, p.1023

(2.9) \\ (I-A)_1ll i I-iXir .

The statement and proof of the following were dis­

covered by the author after examining a similar theorem of • •

Schroder Lll, p.l90l. However, the author feels that it is

quite probable that Kantorovic has established the same

result, but no reference has been found in the available

Russian literature.

Lemma 4. Let X be a Banach space, xQ an element of X, k a.

positive real number, and T a continuous operator defined on

S(xQ>k) = x|xeX, llx-xoll * k^

into X. The sequence {xn$ defined by

xn+l = T(*n>

Page 17: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

10 £

is well defined and converges in norm to an element x be­

longing to S(xQ,k) i£ there exist a positive real number

M, a set B, a^ mapping G from 0 into the reals, a number pQ

belonging to [0,M), a mapping 7 from Co,M) into the reals,

and a sequence of real numbers such that

^ B = i (p1, p2, p3, jo4, jo5)| p* real, Of i£5, Otfc M,

fL+ A M],

(2.10) ||T(x)-T(y)||* G^Jlx-y/l, llx-xjl ,lly-xQH ,UT(y)-xf/f/lT(y)-yli]

whenever both sides are defined,

(2.11) 0£G((o1,p2,p3,p4,p?) ± GU1,*2, 3, 4,^) for O.y 1

and (j^+ M '

(2.12) ||xo-T(xo)|U T(po)- po ,

(2.13) G(p-tr,p-£,tf-po,T(\r)-f,T(<r)j*) ± T(f»)-T(<r)

for whenever both sides are defined.

<2-14> Pn+l = T<Pn> '

(2-15) Pn+l '' fn '

<2-16> ill • r* A M

* and p - pQ k. Moreover

ii * ii / * Ux -xn|/ 4 f - fn.

Proof: By induction we show that

(2.20) (a) xn belongs to S(xQ,k)

(b) llxa-H-xn" 4 Pn+l-Pu

(c) »xn+l-xnll +K-xol14 fn+1" Po

where p^- pQ < M for all myO,

Page 18: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

XI

By hypothesis (2.20) holds for n = 0. Assume that

it holds for n* p-1. Then for n £ p-1

" xrn-l-xn" ~ Pn+l~fn

and hence

(2.21) l|xp-xol|f "xk+l-xk" - STiLl'Vl"^ ~

and therefore x^ belongs to S(xQ,k). Furthermore

II x -x ,1/ + II x ,-x IK M p p—1 p—i o

so that

llx .-.-x. Uil|T(x )-T(x ,)lli p+1 pM p' p-1

± G^llx-x .11 ,llx -x II , llx ,-x || ,|lx -x ll ,Mx -x ,|l} p p-l p o p-1 o ' p p p p-1 "*

or by applying (2.11), (2.13), (2.20b) and (2.21)

^ xp+l"~xp" " Pp"" fp-1' fp" Po' fp-l~ fio' fp' Pp~ /p-1^ 4

4 T(fp)-T(Pp_i) = pP+i-pP

and hence (2.20b) holds for n = p. Therefore

11 Vl-xpll+llVxo" -< fp+l-PpVp-Po = Pp+l"ft> and the induction is complete which proves that the sequence

{x^} is well defined. By the triangle inequality

(2"22> Wn" Since $ pnl is a convergent sequence, it is a Cauchy se­

quence; hence {x^ is a Cauchy sequence in X and therefore *

has a limit x in X. Since

IIX -X || £ P -Pi P - P n o ~ • n • o ' • o

Page 19: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

for all n * 0 it follows that

II x*-xQll & p*- fQ <• k

* / \ and hence x belongs to S(xo,k). Since T is continuous

x* = iiixn+l = ***» T(xn* = *u) = T<x*)»

Finally since

^ xn+p~xn'^ ~ P n+p~Pn ~ ^ ~ n

for all n and p greater than or equal to zero it follows

that

ll*#-xnIU p*-pn

for all n^O and the proof is complete.

Page 20: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

3. Conditions for convergence in Tchebycheff norm.

We examine first the case in which the norm is

given by

.. max | i " x , , = U U n U ± {

using the comparison method. This yields results which

allow us later (section 5) to obtain a somewhat more

general theorem using simple induction.

Since D^F^x) can be represented as a diagonal

matrix, the mapping T of (2.3) satisfies

P±(T(x)) = P.(x)-P. iD~1F'(x)CF(x)]} =

= P.(x)-D-1F'(x)^Pi(F(x))j .

Theorem 1. Let X be a finite dimensional Banach space with

Tchebycheff norm. Let F be a map ping of a set AtX into X.

The iteration

6.1) *n+l = V C F(xn>J

is well defined and the sequence {*x "j; will converge to x , *.

such that F(x ) = 0 if the following conditions are satis­

fied. There exist Xq belonging to A and positive real num­

bers r ,N,(3 .0 and H such that S(x .r )cA, F is twice —— o O O —— O ' 1 o o —

Frechet differentiable on S(xQ,ro), D"*^F,(xo) exists, and

(3.2) llF"(x)ll 4 N for x in S(x0,rQ),

(3.3) IID~1F' (Xq)|| 4 0o ,

(3.4) llD"1F«(xo) F(xo)ll 4 QQ ,

(3.5) l)D"1F,(xo)(F'(xO)-DF'(xo))l| i HQ,

(3.6) h(«) = H + )l- -i. ,

13

Page 21: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

14

( 3 . 7 ) ( 1 - Mh ( io ) ) 2 > 2 ( l+ t*r i f e iwO« u» l Q 0 N|3 ,

* (to) = max(l, 11- j-j)),

and

(3 8) 1—|M»|hi(to)—4(1—t«i»ili(u>) ) —2(1+ '«*»'fc(w))i«/iQoN(3q

(1+ iwilUoJJNp < ro *

Moreover

(3.9) llxn-x*IU I^LPWQ0

where

h(u/) + ll(u))( 1-^(1—Hdh(to))2—2(l+l?iu>)M )i«; Q N0 d(w)= 1 0 ° < l.

h(u>) + fc«*»)+ ,~jy( 1—Mht««>))2(1+ IftwWmO wQ N0

Proof! j

Let

(3.10) H(x) = D~1F'(XO)LP(X)] ;

then if x and y are elements of S(xo,rQ)

H'(x)-H'(y ) = D" 1 F' (x o ) tF ' (x ) -P ' (y ) ]

and by statement i) section 2, and (3.3)

(3.11) 11H ® (x) —H • (y) i I sup^llF" (tx+(l-t)y)ll ,0±t*.l^/|x-y lit

i0QNHx-y//.

Let fl N

5(f) ~\?2»

then E (p ) is a monotone increasing, concave (H"(p ) > 0 )

function on t0,«*>) into E0»°°)» and

(3 .12) l lH«(x) -H' (y ) l l i H' (nx-yn) .

The function

= l-Nft f T

Page 22: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

15

exists and is a bounded linear function on £0,00) for each

p such that 0 < jyjj- .

Since X is a deletion space, we obtain by Lemma 2

D-1F' (Xq)(DF1 (Xq)-DF' (x)) II £ 0ollD(F' (xo)-F« (x))||i

f ^IIF* (XQ)-F'(x)ll 4; @oN)|x-xoll

which is less than ofie if l|x-xoll <• k where k = min(rQ, jj j-) •

Hence by statement ii), section 2

(3.13) Rx = I-D'1F'(xo)tDF'(xo)-DF«(x)]

has a bounded inverse for each x which is an element of

S(xQ,k) such that

<3'14> K1"* l-MPJx-x0„ •

Now

DF'(x) = DF«(x )(l-D_1F«(x )D>F'(x )-DF'(X)}) =

- DF' (x )B o' x

so that

(3.15) D^F'U) = R~1D"1F,(xo)

for all x in S(xQ,k), or equivalently, T(x) is defined for

all x in S(xQ,k). Now by (2.3) and (3.15)

T(x)-T(y) = wD"1F'(x)^-F(x)- £ DF»(x)lT(y)-x]+

+ h DF'(y)[T(y)-y]+F(y)} =

= uR~1|-H(x)- i DH'(x)lT(y)-x] +

+H(y)+ i DH'(y)lT(y)-y]} = R fui

where

(3.17) f = -H(x)- jjjDH'(x)[T(y)-x]+H(y)+ £dH« (y)I T(y)-y] .

Page 23: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

16

We have

(3.18) P.(f) = -P.(H(x))+P.(H(y))-

- i P.^H'(x)[Pi(T(y)-x)J-H'(y)tPi(T(y)-y)]}.

Putting

(3.19) Pi(f(s)) = -P.(H(x))+P.(H(y))-

x ^ r H(x+s(Pi(T(y)-x))-H(x)

" w (, s

H(y+s(P.(T(y)-y))-H(y)

we have

(3.20) P.(f) = Pi(f(s)).

Denoting P^(x) by x^ we obtain

(3.21) fi(s) = -H.(x)+H.(y)- H.U+sU.-x.M.U)-

Hi(y+s(zi-yi))-Hi(y)J

where z = T(y). It now follows (fundamental theorem of

integral calculus) that rl

(3.22) f.(s) =- • (x) (x.-y. )dt-

- RTR ^j=iHi,j®(xryo)dt

where

(3.23) x = tx +*tl-t)y and i = j *

Further,

Page 24: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

17

(3.24)

- (1-A_j)uryj)dt" UJ

- 2:"=iHi,ju<.)(1-ii)(xryj) U9

so that

(3.25)

j#=i

since s can be restricted to be less than one.

Now since the space has a Tchebycheff norm, and remembering

(3.5), (3.6) and (3.9)

lf-(s)[^ 1 |h .(x+s(z.-x.) )-H . (x)| ||x-ylidt + 1 siu>i->0 J ,J '3

+ J X j=i Hi,j (*)~Hi^ j (x0)l fc<«»llx-yUdt+h(uj)|f x-yl/ o

and

* lilUUa [IW Hi,3("s(zi-"i) '""i.j®' Hx-y,ldt +

+l*i* n fej.ll Hij <=>-*!,, 'x0'lkl"lllx-yll<li+h(*i)llx-yll.

Page 25: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

Since the sums are finite there exist N and N1 such that

rtejHN, j < <2n-*N>>"hn,j(x) 1Hx-yldt + Jo

(3.26) £ 1 (x+s ( Zjj-Xjj) ) -H • (x)ll nx-ylldt + O

+ (llH' (x)-H' (x )llta»l|)c-yMdt+hMlx-yl| 0 °

and applying (3*12) to the integrands, we obtain ^

(3.27) (slIZjj-XjjlDllx-ylldt + £

1 +(H'(\IX-X II )lt(*>llx-ylldt+ h(U>) llx-yll. Jo 0

Now let p1 aiix-yll, p^=|lx-xol|, p^=»y-xQl|, p4=l/z-x/l,

and ^j^=llz-yH; then noting (3.23)

l|x-xollt tf1 +f?

U zN-xN«=|(zN-xN)t+(l-t)(zN-yN)|4 p4t+(l-t)p5,

and, since H'(p) is^increasing,

"f (s)IU^fl' (s^t^+d-t)^ )pXdt + i «

+ \ 1' (tp1+fJ)lfM^dt + hO^f1, Jo

or applying (3.10) t* 1 12

(3.28) ||f(3)||4Si,j(f4+f?) t- + mV.KPV+ •SV_)J +h >f1-

Therefore since the right hand member is independent of a

and the norm is continuous it follows from (3.20) that

(3.29) ,tfil^o[(PV) £_ +lw|lKv»)(f3p1+ (p*)2 )] +h(c*>)p1,

or applying (3.14), (3.16), and (3.29)jtherefore,

Page 26: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

+

,1 \ 2

19

(3.30) l|T(x)-T(y)|l 4 HuiHe 1# . |Jf||*

6 G li x-yll, l|x-xQ|/ ,l|y-xo|| ,l\T(y)-xll, llT(y)—ylfl

where

(3.31) olf1, P p5] = —(pW 1-NB p C

^ 2 + Rol kh») N(io(p3p1+ )+iMh(u»)p1 j.

Since M , N(3q, 1? (•«), and h(ui) are all positive it follows

that G satisfies (2.10) and (2.11) in Lemma 4 for 0* p*"/ jj75~« o

Now let p1 =p -0% p2 = f , p3 =<T , p =f-p, f5 =t-<rj

then noting (3.30), (3.31) becomes

(3.32) Gfcp-<r,p,0-j-p,f-<r] = I_Npi "IT*sT-p-q-)(f -<r) +

+ MV<hiNQo(T(p-a-)+ (p-cr)2)+Mh(w) (p-tf)J.#

Let

(3.33) T(p) = p + x.N^ jd + (l+M*»«>)-^ p2_ p -utol P hcu>)}

where C is any constant. Identifying "J as T(^) we obtain

in a manner similar to that used for (3.16)

i r N0 (3.34) f(f)-T(«r) = !_N p\-T- (2f-Mr)(f-<0 +

+ M l?l«o)N0 (^(p-t)+ (P-<*)2) •+• Mh(w)(p-ff )1 = 0 2

= cfa-'V , ,f(o-)-p,T(Tr)-a].

Since T(0) = C , if we choose C = l*»l Qq we obtain

(3.35) J|T(xo)-xo!| £ M Qq = T(0)-0.

Hence by applying (3.34) and (3.35) we see that (2.12) and

Page 27: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

tr

20

(2 .13 ) are satisfied for 0 = foST'f 4 H£" It only

remains to be shown that

<3-36> Pn * Pn+1 T T<Pn>

and

(3.37) P = P C min(r , 172--) n*«© «n 1 x o*

in order to satisfy the hypotheses of Lemma 4. .

For this purpose let us examine T(p). The fixed

points of T(p) are

1 — |u»\h(<«0~|( 1—i«w(hC««))2QqIu*\ (l+lu>l l?tw})N0o

1,2= (1+itoi I?(M)N0O "

by (3.5) and (3.6) these two roots are real und positive.

Let ^ 1— iwhM—A (1—Mh(W) )^—2Q lu>l (1+Ik«m»1|u»I )N{3 1

(3-38) P " II+»i»..))nao

and consider the image of the closed interval CO, p ] under

the mapping T(p). *

Since pi, p, the minimum of the two roots of the quadratric,

it follows that

(3 .39 ) q(p) = Q Q M + (l+I"l Uf*>) f — (1—Mh(<«) )p > 0.

Therefore r- _ (1-NB p)|M ll(w>N0 P+Mhtu>)J +Nfl q(p)

(3.40) T»(P) = 2 E J 2 1 ^ 2—L >0 (l-Nfl^r

* 1 for p^p l. and

T(p) J* °T(p*) = p*

Page 28: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

21

so that T maps \[.0,f* 3 into L P > P - l « , Furthermore with

Pn+1 = (Pn)' since 0 forp*P*, it follows that

T(f)7f or that (3.35) is satisfied. Now

„„ r u . ,lt2"V, x-21-hcil Nfi0?'D _

i gr" j+ u.n|JoF)3 >

Therefore, #

(3.41) Qtftp* T'(P) = T'(p*) = * lto>hCUj) = 01 (UJ) =

Ch (w) +h (<" )3 -k (uj ) 3

DI(U;) + *(AU)L+

where

(1- i<4h(to) )2-2(l+ MV (I« ))MQ0N60 .

Therefore

(3.42) 0-t d (UJ) < l .

Now by the mean value theorem and (3.41)

(3.43) ^T(p)-¥(^-)| £ ck («o) Ip-^l

and hence T(p) is a contraction map on Lo,p*3and the itera­

tion (3.35) will converge to the unique (by Banach's fixed

point theorem C10, p. 27l) fixed pointp . We have already

(3.38) shown that

* P * — r N o

/ \ * and by (3.7) f> < rQ ,

(3.44) p £ min( ,rQ)

r

Page 29: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

22

and the hypotheses of Lemma 4 are satisfied for T(x).

Thus there exists x such that T(x ) = x , and by (2.19)

(3.45) ||xn-x*|| < f* -fni •

£ It remains to be shown that F(x ) = 0. Solving (3.1) for

F(Xr) we obtain

(3.46) -F(xn) =>P'(xn)(xn+1-xn).

Now by applying statement (i), section 2 and the triangle

inequality

) I F ' ( X q)II - t l F ' ( x O )U £l]F'Un)-F'(xo)//£

i II xn-xoll sup [llF" (xQ+t (xn-xQ) )U, 0 it£ lji rQN

so '»

(3.47) J| F' (xn)jl i ||F• (xQ)|| +rqN.

W"e then apply Lemma 2 and obtain

(3.48) !|j)F'(xn)l! * llF»(xn))) i J/F'(xo)//+roN.

Using this, from (3.46) it follows that

II F (xn )il iilI xn+1-xnl| (1/ F1 (xQ )ll +roN).

Finally, since the norm is continuous

II F(x*)|| = 0

or F(x*) = 0

(which is immediate from (3.46) if one has first shown that

•DF'(x) is continuous).

Uniqueness is covered by the proof of the more

general Theorem 3, section 5.

Page 30: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

23

It is noted that the convergence of the sequence^pn}

to p* follows directly from (3.36) and (3.41) since p n

will converge to a fixed point of T(p) less than or equal * *

top . Since p is the only such fixed point the sequence *

must converge to p • However, no error bounds are obtained

by this argument.

Page 31: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

4. Choice of to .

In this section we will show that the optimum

to the convention that <o, is a "better" value of Ui than

U) 2 if o((^) ' (v) 30 "^at error bounds are cor­

respondingly related. It then follows that the best choice

of (O is that value which will minimize o( (w). To determine

this value we note the following

(I) Condition (3.5) can not be satisfied if

u) <, 0 and can be satisfied for 0 only as follows

value of <o is one. In order to do this we will agree

i f H 4. 1, o '

l'-"'2 lf H, < ~ •

(II) We have from (3.6) that

- 1 f°r 06UJ £ i

* ( « > =

1 for 1 i W ,

(III) By consideration of (I) and (II) we see

that (3.6) is satisfied if

(4 .1) 0 NO for 0 * t 2

o

24

Page 32: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

25

(4.2) g°N(3° L —f±£—* 1 (l-H )2 for £<*><1, (l-H,)2 2(1+I4

(2-«»(l+H ))2 , (4.3) Q NA < 2 < i4l-H )z for u> * l.

0 0 2(l+w) " 0

Furthermore for Wi { and (3.5) and (3.6) satisfied it

follows that

^ 4 , 4 ) 2N(3 Q * N0 O U-C«J

2N0 ^ since 0 4. rrpz—rr \ by (4.1). Therefore if (3.7) is

(l-Ho)2 2{2~uf)

satisfied for u) £ \ it is also satisfied for V = 1. Simi­

larly, if ——2(l+ui) ~ * *OT ^ "t0 ~ fol -ows

that

(4.5) l-H -o (1-H„)2—4B0Np0 ^

2N o ^ ( 1-HQ ) -/a? ( 1-HQ ) 2-2QON£OFC»( 1+TF)

^ ,

and hence if (3.5), (3.6), and (3.7) are satisfied for

jiw* 1, then they are also satisfied for 10= 1. Finally,

since

2- t o ( 1+HN) - J [ 2 -U>( 1+HQ )] 2-2QON0N6,( 1+ U+«)NB (4'6> rM = —^

is an increasing function for W?,l, if (3.5) and (3.6) are

Page 33: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

26

satisfied for some value of <o?l then

(4.7) (1-H ) - V ( l -H )2-4Q N0 ^ = r ( l ) < r ( » ) .

We have shown

Lemma 5 • If. co nditions (3.5), (3.6), and (3.7) are satis­

fied for any value of OJ between 0 and 2, they are satisfied

also for (o = 1.

With this lemma in mind we arrive at

Theorem 2. .If the conditions of Theorem 1 are satisfied

for any value of w between 0 and 2, then the best value of

(JJ jls one.

Proof:

We need only show that

<«•«> —a).

To do this we will consider three cases:

Case I. If (Hu>i£ , then

l-a»(l-H ) + (l-«/)-(1 (1-H ) 2-2kw(2-w) (4.9) clM = 2 °

(1-^) +l-uH l-Ho) + Vu? (l-Ho) 2-2ka/( 2-u>)

where k = and

(4.10) u(u+2(H -i-|)J^(l-HQ)g2k(l-..)](2^Ho-2)>2—«) CX(U>) = u

(2+ (Ho-2)+u)'

where

u = Ju»2 ( 1-HQ ) 2-2ku) (2-w) .

Page 34: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

Now

v/o>2(l-H )2-2kw(2-*) * w(l-H ) 2(l—H ) ; O v <P

so

2 (Hq-1 ) + Jw2( 1-Hq ) 2-2kui( 2- ») 0.

Since by (4.1)

k * v i &

(1-Hq)2 2(2-<*) 2 (l-«u)

it follows that

(4.11) u/(l-Ho)2-2k(l-u>) > 0.

Also 2-" 7 0

and since H ? 0, \ 0. o

(4.12) 2+ ui(Hq-2) » 0 ;

therefore,

(4.13) oM 4 0

and oC (u>) ^o((j) for Q*uj< £ .

Case II. If i, then

ojH +1- . o» 2 (L-H )2-2kw(l+ai) *(•») = —2 " ° .

H +1+Jw2 (1-H ) 2-2kw( l+«u)

and

(4.14) ot'(w) = -pa-ii0)2 -2ka»(l+fci) (/w2(l-Ho)22kw(l+«)+l

(u>2(l-H )2-(l+2u»)k) M +1) (w+1) ~ o o •

\ju>2( l-Ho)2-21uu(l+M)

Now by (4.2) and (4.9)

Page 35: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

28

4. ——— / ——— j therefore, (1-HQ) 2 2 ( l + t o ) l+2w

to(l-Ho)2 - k(l+2u/) 7 0,

and

(4.15) ot' (">) < 0.

Now c((S)><rf (£)? °< (l) for

Thus if (3.5), (3.6), and (3.7) are satisfied for some

(o^.l, then u> = 1 will minimize o( (<*).

Case III. If u> Tf. 1, then

UH +2)-l-wj(2-w(l+H ))2-2kw (l+o/) (4.16) =

(H +2)-l+ i(2-w( 1+H ))2-2kw(M (*} ' o

and [W(HQ+2 )-l](«+l fe—(1+Hq))(l+H>-k( 1+2^)3

/ x • u(H +3-u)+ ~ (4 .17) O((U*)= 2

(w(HQ+2)-1 + u)2

where

u = \[(2-m(l+Ho) )2-2kw(l+to)«»

Now,

H o +3_j(2-iu(l+H ))2-skw(l+w) 7 2-w(H +1) -

->K2-w(1+Ho)) 2-21uj(1+*J) >0,

and 2-w(1+Hq) > 0. By (3.5) it then follows that ©<• (<o) > 0,

and we have for 1 < 2

(4.18) o(( 1) 6 «< (cu).

Page 36: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

5. Conditions in a deletion space.

In section 3 conditions were obtained that would

guarantee the convergence of (l.l) when the norm associated

with the space was the maximum or Tchebycheff norm. In this

section it will be shown that the conditions are also suf­

ficient in any deletion space, in particular in a space,

and uniqueness will be obtained.

Theorem 3. Let X be a deletion space. Let F b£ a mapping

of set ACX into X. The iteration

x i = x -D-1F•(x )[F(x )] j n+1 n v n' v n'J _ , *

is well defined and the sequence will converge to x . #

such that F(x ) = Q if the following conditions are satis­

fied. There exist xQ belonging to A and positive real num­

bers r,N,(3 ,Q ,H such that S(x_,r)cA, F is twice Frechet —• O O O — — o 1 --im • • ——•

differentiable on S(x ,r), D~"'"F,(xo) exists, and

(5.1) IIF" (x)H 6 N for x in S(x ,r)t

(5.2) llD"1F'(xo)l|4 0Q,

(5.3) llD-1F'(xo)[F(xo)]|U Qq,

(5.4) (|D"1F»(xo)[F'(xo)-DF'(xo)]I| £ En < 1,

g Q N (5.5) b =

0 (1-H )2

and Qn ,

(5.6) r = ,° • (l-Jl^fb.) < r. 2b.(1-H.) u-"-,0o'

29

Page 37: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

"\

30

* / \ Moreover x belongs to S(xQ,ro) and

(5.7) )lx-x*\U _e£_ Q 1- d °

where

* « = Ho+1-(1-Ho) 4E7

H +l+(1-H )Jl-4b A * A ' * / * .

and x is unique in every closed sphere S(x^,k)C S(x^.r)

where

S(xQ,k) x|x« X,l|x-xQ\l £ kj

and k is. such that

k * 2b (I-H ) u+n^r0). o o

We will first show that the sequence is well

defined. To do this we show that there exist positive real

numbers r ,0 -Q and H such that n' n7*n n

(5.8) where f^ = DF'(xn),

(5.9) |lf-1[F(xn)]|j £ Qn,

(5.10) llf 1(F'(xn)-fn)|l i Hn < 1,

N0 Q (5.11) b = a a < i

n (1-H )2

and

(5.12) = rn, SUa,rn) C SU„,r0) . n n

These statements are true for n = 0 by hypothesis.

Let us assume that they are true for n = p; then

is defined, and by (5.10) and (5.11).

Page 38: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

(5 .X3) f i p = r p < r p 2 "p < r p .

1- l-4b 1- l-4b P P

Thus, there exists a positive number rp+ such that

s(Vi'Vi)csUP,rp)CA •

By (5.8)

|,fp1(fp+i~V" - WW

and hence by statement i), section 2

(5.14) ||fp1(fp+1-fp)jk(Jp||xp+1-xp||sup^®F' (xp+t(xp+1-xp) )J 'I), ,0*t*lj .

Applying Lemma 2 we obtain

(5.15) Hf^Up+i-y14

< &p#*p+l-*p# sup^((DP"(xp+txp+1)ll, 0 i ti 1 j ,

and, since X is a deletion space,

(5.16) sup |||P"(xp+txp+1)H,yo it! lj*

9p2pN = bp'1-Hp»2 < i •

Therefore the bounded linear operator

(5.17) hp =

has an inverse h~* such that P

(5.18) Hh^llS - "J • P 1-\(1-HP)

Moreover

Wx# - fP+i(x)'

and hence

(5.i9) #! = h;1*;1

Page 39: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

\32

exists and

(5.20) llfriJU — . P+1 l-b (l-H )2 VP+1

p p

Let

(5.21) Gp(x) = x-f'^FU)];

then

G (x ) = x ,, P P P+1

(5.22) G^(x) = fpX(fp- P'(x)),

and

(5.23) G"(x) = -f"V(x) . f Jr

Thus by statement i), section 2,

(5.24) H'p1(F(Xp+]^) ))l= Il0p(xp)"0p(xp+1)|l 1

i"VVi'sup{"°i(Vt(Vl"3Cp)"' ° 4 **

and by applying it again

(5.25) >l^(xp+1+t(xp-xp+1)l|4||a^Up)// +

+ |l-t|llxp-xp+1ll3up^||G»(xp+s(l-t)(xp+:L-xp)j(/,0ssal5

or for OStSl and by (5.25), (5.10), (5.9), (5.8), and (5.1)

(5.26) U^(xp+1+t(xp-xp+1))|li Hp+(l-Uep0pN£Hp+bp(l-Hp)2.

Therefore, by (5.24) and (5.26),

(5.27) llf;1 P(xp+1))l 4 Qp(Hp+bp(l-Hp)2)

and by (5.18), (5.19), and (5.27)

(5.28) llf;^ p(xp+1)|l =iih;1f;1(p(xp+1))i|t

n H +b (l-H )2 „ £ St, P P El =. i •

l-b_(l-H )

Page 40: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

33

Now by (5.19)

(5.29) =llhp1(fpl(P'(Vl)"Vl))"4

Let

(5.30) gp(x) = f"X(F'(x)-DF'U)) ;

then by the triangle inequality, (5.8), and Lemma 2

(5.31) llgp(xp+1)|U||gp(xp)|/ +//fp1{f(*p+1)-F'(xp)+DF'(xp)-

-M?'(xp+1)fl|i

i llgp(xp)U +Pp||F'(xp+1)-F'Up)-D(F'Up+1)-F'(xp))|l

and. since X is a deletion space by statement (i), section 2

we have

(5.32) Hgp(xp+1)IU llgp(xp)||+0pHF'(xp+1)-F'(xp)lli

il|gp(xp)ll+?p||xp+1-xp\| 3up^llF"(xp))l, 04Ul]

where -x = x +t(x ^,-x ). Therefore by (5.1), (5.9), and p p p+1 p

(5.12) we have

(5.33) llgp (xp<) |1 i|| gp (xp )|| + 9pQpN.

Therefore,

(5.34) <xp+1)-fp+1)ll i Hp+bp(l-Hp)2

and by applying (5.18) and (5.19)

(5.35) (vi'V111'- yv1-"/ • Hp+i • l-bp(l-Hp)2

Now

(5.36) 1-H j = (1-H ) (l-2hp(l-Hp)) > 0

l-bp(l-Hp)2

Page 41: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

34

• ince b * i and H * 1 or P P

(5-37) Vi41-

Also by (5.11), (5.20), (5.28), and (5.37)

(5.38) t V^U-"D)2

(i-Vi)2 ri-2bpd-Hp)]2

and

(5i39) l-4b , = 1~4bp >0

[l-2bp(l-Hp)]2

so that

(5.40) bp+1< i .

By (5.12) and the induction hypotheses

(5.41) SUp,rp) C SU0,r0) .

Applying (5.28), (5.36), and (5.39) to (5.12) we obtain

(5-42) VI • H*- M- -P+I p+I

= 2b (l-H ; P P *

then if *x-xp+1|| 4 rp+1)

(5.43) ||x-xpl|il|xp-xp+1|| +||xp+1-x|/±

£ Q +r -Q = r P P P P

or

(5.44) S(xp+1,rp+1) c S(xp,rp) c SU0,r0) .

Therefore the iterates x are well defined and belong to n

S(xo,rQ) . Since S(xQ,rQ) is closed we need only show

that (xn] is a Cauchy sequence to show that it has a limit

Page 42: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

35

in s^o'ro>-

For this purpose let

H +l-(l-H )Jl-4b (5.46) of = n

H +l+(l—H )J l-4b n v n'^ n

then applying the inductive definitions of H and b we n n

see that (5.47), o(n

for all n, or

(5.48) o(n =o<0 = «X

for all n.

Now by (5.28)

2n+l = 2nHn

and by a set of calculations

0 £ 2b (l-H )+Jl-4b < 2b +4l-4b < 1 " n* n' n n n

0 4 l-2bn(l-Hn)-n^bn " .

0 < (l+IIn)(l-Hn)[l-2bii(l-Hn)-h-4bn] =

= (l-Hn)[l-Hn-bn(l-Hn)2J -

" (1+Hn )t (1-Hn) fTUbn-bn (1-Hn) 2J

A (1+Hn) (l-bn( 1-Hn):2) - (1-Hn) n^4bn+bna-Hn)jF 1

Hn( l+Hn+( l_Hn) 4T„) fbn (1-Hn)2 (1+Hn) +bn( 1-Hj 3JI^b; <

< (1+Hn) (l-bn(1-Hn)2)-(1-Hn) JT5bn+bn( 1-Hn) 3 T4b^

(l+Hn+(l-Hn)n=4b;)(Hn+b0(l-Hn)2)

<• U-bn(l-Hn)2)(l+Hn-<l-Hn) g l-4ba)

Page 43: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

(5.49) Hn< «„ = .<„ = <*.

Therefore,

(5.50) Qn 6 ocnQ0»

and since «(< 1,

(5.51) tlxn+ -*nl| i Hk=l2n+k-l-^k=l °<n ±Qo = 2o A~C\

so that {x^ is a Cauchy sequence which converges to x

in SU0,rQ) .

Since by statement (i), section 2,

(5 .52) Hfn-f 0!l*«*n-xo\l sup{|lDF» (xo+t(xn-.xQ) )|| ,0iUl] i rQN ,

it follows that

(5.53) l|fn)|i llf0l|+r0N .

Now

(5.54) F(xn) = fn(xn+1-xn)

which implies that

(5.5$) UF(xn)J| t (tfl)+Nr0)Qn

or

(5.56) l|F(x*)l| = 0,

which is equivalent to

(5.58) F(x*) = 0.

The error bounds are obtained by letting p tend to

infinity in (3.51) yielding

(5.59) Ux*-x ft l erf11 Q . n ~ l-<* °

The uniqueness of x in every closed sphere S(xQ,k)

S(x.,r.) with iu (l+>Jl-4b~) is proved by contra-2b0(l-H0)

Page 44: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

37 *

diction. If u different from x were a solution in

S(x .k), then o '

(5.60) llu-x.li = ® n—— 0 2b (1-H )U^l-4b ), 0<6<1, 0 0

By induction it follows that

(5.61) ||u-xnl| £ m ^n (1+ 4l—4b^)» m +(l-rf)©< 1.

Obviously this is true for n = 0. Suppose it is true for

some n = 0, then Gp(u) = u where Gp is defined by (5.21).

Then by an argument similar to that used to obtain (5.28)

(5.62) llxp+1-u|| = ||Gp(xp)-Gp(u)|/i||xp-ul/(Hp+3pN//xp-u//).

Since by (5.28), (5.36), (5.38), and (5.39)

-P*1 i tt (l+Jl-4b ,)= 2VI(1"VD P+1

Qv (l+Jl-4b -2b (1-H )). 2b (1-ri ) P P P P P

2bpU-Hp)

"D 2 it follows that by adding and subtracting (m ^p(l+<ll-4b.u)

to (5.62) that

<*•*»> "vi"u//< r s - + p+i p+i

+ 2pHp C l-<ll-4b 1Tn>P^-(mp®)21 . U-H ) PJL 2bp

From the definitions and properties of Hp+ , Qp+ » and

b it follows that p+1

(5.64) 0 i(1-H )(1+H )(l-2b (l-4b )) = Jtr ir Jt Jt

=(1+Hp)(1-Hp)(l-2bp)-(l+Hp)(1-Hp)(l-4bp)

Page 45: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

38

or by a set of manipulations

Hp(l+H

pMl-Hp)Hp(l-4bp)*

i (1+H )(l-2b (1-H ))—(1—H )(l-4b ) if jf }r r

HpC1+Hp+(1—Hp)(l-4bp)+2jl-4bp]< (1+Hp)(l-2bp(1-Hp)-

-(1-H (l-4b )+2>ll-4b (H +b (l-H )2) P P P P P V

Hp(l+Jl-4bp) (l+Hp+(l-Hp)\Jl-4bp) *

£ (1+H -(1-H )i l-4b )(l+U-4b -2b (l-H )) P P P P P P

H (1+ l-4b )<, cx (l+Al-4b -2b (l-H )) = P P P P P P

= cx(l+>Jl-4b -2b (1-H )) P P P

(5 .65) , QA(1^) „ g |( 1 , ^ l r - )

2£pu-Hp) - 2b—rnr^T u+u p+i •

Thus since m% < 1 it follows that

mp© -(mp©)2 0,

and from (5.63) and (5.65)

IIu-x nlu p+1 ^ ((m p9 ) 2(l-o()+ o c m p G ) p "2Vi^-Vi'

or

l|u-x ju mPe2^i'1+J1-4Vi) («•+(!-.<)© ) =

P 2ViU"Vl'

= mP+le . 2VIU"VI'

Therefore mn0 Qn(l+fl-4bn) mnS (l-Hj mn© (1-H0)

Uu-V4 2bnH-5n) £ 5T) 4 fTI

Page 46: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

39

or

(5 . 6 6) i i lnx n -u | l= 0 .

Since f*nl converges to x this last statement implies *

that u = x which is a contradiction.

It must be noted that another error bound can be

obtained from (5.42), (5.44), and the remarks following

(5.44). Since n is arbitrary,

(5.67) llx -x II i ^n (l-^l-4b ) , for n ^ 0, n ~ 2b (1-H ) n

n n'

and this error bound is the one used for calculation in

Section 6.

In order to aid in possible computational applica­

tions we restate the above theorem in terms of the Tcheby-

cheff and norms, respectively.

Corollary 1. Let f (x^ ,... ,xq) be^ a real-valued function

H \ defined on ACE for i = 1,2,...,n, the sequence of Vectors

x(k) = (xj^ ,... >x^^) defined by (l.l) converges to a * * *

solution x = (x^,...,xn) of * *

f (Xx ,. • *,xn) = 0 3 i = 1 »2 ,... ,n ,

if there exist Xq = (x°, x^,... »x ) belonging to A and positive

real numbers r,N,0 ,Q , and H such that • ———— ' 0 0 o —— ———

S(xQ,r ) = £x|x«En, A,

f^ has continuous second order partial derivatives on A, and

(5.68) J",n jkW|« » for x in S(x0,r),

Page 47: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

40

<»•«> EM f-~ury| P„ .

(5-70) 11 ii n ] I~r^T I 4 So .

& Q N (5.72) b = -2-2— < i ,

(1-Hq)

and

(5-73) 2b (i-B ) = ro< r-o o

Moreover, x is in S(x ,r ) C. A and

max | (k) *W of Q

1* itn I l i i 1— oC o

where

1+H -(1-H )4l-4b o o o

1+H +(1-H )\|l-4b o o' * o * .

and x ijs unique in every closed sphere S(xQ,k) C S(xQ,r)

where

SUo'k) = \ x|x€En» if iin ixi-xil"kl

and

k " 2b0(l-H0) <l+^b0) •

Corollary 2. Let f^(x^,...,xq) Jbe a real-valued function

defined on A<- En for i = 1,2,..., n, the sequence of vectors

x(k) _ (x(k)ft#,x k ) defined by (l.l) converges to a solu-

"H* "M" tion x = (x^,..•,x^) of f^(x^,...,) = 0, i = l,2,»««,n,

if there exist xQ = (x°,...,x° ) belonging to A and posi­

tive real numbers r, N, , Q , and H such that — • ' o o ~ o • ———

Page 48: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

41

S(x o , r ) = ^x lx« f i n , 2 * r J c A »

has continuous second order partial derivatives on A, and

(5.74) *** 2 j=llfi,jk(x)M N f££ x in S(xo,r),

< 5 - 7 5 > To«» I r r fe l * s .

(5.76) z»

J . J O ' ' O

. 1 ° f. •(x ) JFJ O'

* Q » s *0

(5.77) ( 3Ej=i |f j — i , i xo )\- H0 x0)|» i-l»2,...,n,

M N (5.78) b = 0 0

° * i »

and

('•79> ro = 2b (LH ) U-fl=4b-0) < r. o o'

Moreover, x* is in S(xo,rQ) and n |x^-x*|i Q(

where

©< =

H +l-(l-H )U-4F o o 1 o

H +1+(1-H M l-4b_

and x i£ unique in every closed sphere S^xQ,k) C £>(xQ,r)

where

and

Stx0,kJ = £x|x«JSn, 2:°=1|xrx°| 4 k]

k 4 2b U-H ) •

Page 49: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

42

Thus through the corollaries we obtain two different

sets of conditions which will guarantee that the iteration

will converge (all norm topologies in finite dimensional

B-spaces being equivalent). The condition pertaining to

Hq in both cases is analogous to a condition of diagonal

domination which, as is well-known, is sufficient to guaran­

tee convergence when the f^ are all linear. It is this

condition which may somewhat limit the applications of the

method, even if a solution is known to exist.

Page 50: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

6. Applications*

The applications considered in this paper will be

to the finite difference analog to the two point boundary

value problem for a second order non-linear ordinary dif­

ferential equation. A few remarks will be made concerning

the finite difference analog to the Dirichlet problem for

a non-linear second order partial differential equation

with two independent variables.

Consider now the problem

(6.1) G(y"(x),y'(x),y(x),k) = 0 for a f x i b

(6.2) y(a) = yQ , y(b) = I

and along with it the finite difference analog with central

differences (y(x.) is replaced by y., y*(x.) is replaced by J J J

- 1 +l~ 1 -1 t and y"(x.) is replaced by +1~ .1 con-2h b2

sisting of the following set of equations

(6.3, t. (y Xb, . . ilSaZlzL,, . , J n 2h J J

for all j = l,2,...,n, where

" = fef • xo = a' x3 - yo = and yn+l = J-

We will now see what conditions on G will insure that con­

dition (5.71) (the "difficult" condition to satisfy) will be

satisfied. The following notation is adopted

43

Page 51: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

44

(j) a(y l'2y yj-l , yj*l'yj>-i,y ,x ) W1 " d IT 2h '

) y«

.0 . .0 >U)=

V h

(<i>. 3 Jy h

2h

.0 o

J J

2h J J

Then

k=J+1» j*n

L ( j )

f j 'k^l»* * * >yn^

G*J/(- h2 +G3 ( j ) ; k=j

0ii)(-T5)-82J)(-k)» j"1' j*1 n

^0 for all other k.

Now we see that in order for (5.71) to be satisfied in the

space with Tchebycheff norm

^j=ilfi,j(yi'*,,,yn)l 4 lfi,i<yJ»---»yJl>l

FOR X — X|2Y«**YU

or

(6.3) o(d)(_|)+GU)a))+jG(d)(^)+GU)(. i_)j4

jGp)(—§)+0 )) j=l,j$n

:MX,J n 1 n

(6.5) |e<n)(- )-G<n) |'|6in)(- •

Now if G is such that j=l,2,...,n, that is

(6.4)

Page 52: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

45

G^ and G^ have different signs and if h is chosen

such that

0 < h< 2 3 ,min

then

|G(J>+ | CF(J)|+ |G<J )_ | 0<J>|, JAO'J'-H^'}

for j = 1,2,...,n and hence (6.3), (6.4), and (6.5) are all

satisfied. If we restrict ourselves to i = l,2,...,n-l, we

see that these conditions will be satisfied if and only if

fied. Thus it is possible to examine the differential equa­

tion and predict whether or not (5.71) will be satisfied

when the space has a Tchebycheff norm. These conditions are

compatable with the conditions that Henrici uses to obtain

the existence arid uniqueness to the two point boundary

value problem of the class M. £5, p. 3473

Now in a similar manner we consider the Dirichlet

problem on the rectangle A:a x*b,c*y<d for the operator

of the form

where "&A is the boundary of A. We will use a uniform mesh

size on the rectangle such that nh = b-a and mh = d-c. This

will yield mn points where u must be determined by the equa­

tions^

G^^G^^* 0 when the mesh size is such that (6.6) is satis-

G(uxx>uyy>%uy,u,x,y) =0 in A

u = g(x,y) on "iA

Page 53: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

46

f j (U]_ »• • • »umn) =

u. , —2u.+u. , u. —2u.+u. u. , —u. , _ i _JL±S i IzS .1+1 izi,

,2 ,2 2h

u., -u. •l+m -l-m

» 2h .<yVy.i) = o,

j 'n, j + kPJ+1, jfcton, j* (m-l)n,

with appropriate changes to take care of the boundary con­

ditions. If we again consider the maximum norm we see that

the conditions G^^ 2^^ G^*^* 0, j = l,2,...,mn,

and

2 * mm (

Gu> 4 |G33>

opT 2

• |a(j) )

will imply that (5.71) is satisfied where the notation GM

is analogous to that used above.

We will now consider a specific problem, the two

point boundary value problem

y" + (yl)2__JL- + ^ =0, y 2"* ly ' 2"* <y+l)2 '

y(o) = 0, y(1) = 1. (j)

Ve first note that in the above notation G^ =1 for all j

0(S> 2. y<

, 3 " s i r " ( ^^J"1)2 (yj+i)J v 2h

GU> . (iZi, 2 2-x. h

J

Then if y?7-1 Gj G^ < 0 for all j,

Page 54: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

47

2 —T- :(1-y.i)(

y.i^-yi-l). ST3T h '•

2

Hence as a first approximation let y. = x. ; then

8 ( J ) J J

2 "TJT > 2 I = 2h > h for all j. G1

Therefore we will let h = .05, n = 19 and start with an

approximation y = x for the finite difference analog.

After calculations for determining the other constants we

see that the other conditions are not satisfied for any

r>0. Nevertheless by performing the iteration designated

by (1.3) with o» = 1.6 it was found on the IBM 7090 at the

U. S. Army Electronic Proving Ground, Ft. Huachuca, th tL± ict

y° = (y°,y°,...,y°) was obtained such that the requisite 1 2 n

conditions were satisfied and error bounds for y were

obtained from (5.67) by letting n = 0.

Page 55: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

The point y° and associated Tchebycheff error

bound are here listed:

yi = 0.05833235 r = .001

y°2 = 0.11414186 20 = .00000004

0 y3

= 0.16785497 0o = 0.49931097

o *4

= 0.21983032 H 0 = 0.99862184

o y5

= 0.27037553 N = 1.1

*6 = 0.31975935 b 0 = .001156299

y7 = 0.36822072 and

y8 = 0.41597560 lly*--y0ll

y9 = 0.46322248

y10 = 0.51014680

yll = 0.55692472

y12 = 0.60332645

y13 = 0.6507193

y14 = 0.69807087

y15 = 0.74595187

y16 = 0.79453991

y17 = 0.84402329

0 yl8

= 0.89460586

0 y19

= 0.94651308

This same problem was used by Lieberstein ["97 as

an example but he had no rigorous convergence analysis

and hence he did not exhibit either the existence of a

solution or error bounds.

Page 56: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

49

It is to be noted that (5*67) gives a better error

bound than (5.7), and that in either case difficulties with

accumulation of round-off error can be eliminated. Calcu­

lations for H , (3 , Q , N and b can be made by hand by n * n n n

assuming that xn is the initial point. This error bound

is actually obtained by the statement of the existence

theorem; that is, x must belong to S(xo,rQ) where

ro • 2bo(l-Ho)<1

and hence the error bound which we have quoted (5.67) arises.

Had we carried out another iteration on the trial problem

to obtain a point yW ve could use the error bound (5.7)

for the Tchebycheff norm but the bound is not as good as

that which we obtain in less time using (5.67). Also,

one should be warned about use of (5.7) unless the next to

the last iterate is taken to be y^°^ since otherwise a

serious unaccounted for error due to round-off may be

involved. In either case the computation for the error

can be monitored and, if it is needed, multiple precision

arithmetic can be used.

Page 57: This dissertation has been 63—6284 microfilmed …arizona.openrepository.com/arizona/bitstream/10150/284455/1/azu_td... · This dissertation has been 63—6284 microfilmed exactly

REFERENCES

1. Altman, M., A generalization of Newton's method, Bull. Acad. Polon. Sci. CI. Ill 3, (189-193) (1955).

2. Antosiewicz, H. A., Numerical analysis and functional analysis, A Survey of Numerical Analysis (485-517) McGraw-Hill, New York, 1962.

3. Bartle, R. G., Newton's method in Banach spaces, Proc. Amer. Math. Soc., 6 (827-831) (1955).

4. Dieudonne, J., Foundations of Modern Analysis, Academic Press (i960).

5. Henrici, Peter, Discrete Variable Methods in Ordinary Differential Equations, John Wiley and Sons, New York (1962).

6. Kantorovic, L. V., On Newton's method for functional equations, (Russian) Dokl. Akad. Nauk, SSSR n.s. 59, (1237-1240) (1948).

7. Kantorovic, L. V., Some further applications of the principle of majorants. (Russian) Dokl. Akad. Nauk. SSSR n.s. 80, (849-852) (1951).

8. Kolmogorov, A. N. and Fomin, S. V., Elements of the Theory of Functions and Functional Analysis Vol. 1, Graylock Press, Rochester, N. I., (1957).

9. Lieberstein, H. M., Overrelaxation for non-linear elliptic partial differential problems, MRC Tech. Summary Rept. 80, (1959).

10. Lusternik, L. V., and Sobolev, V. J., Elements of Functional Analysis, Frederick Ungar, New York,-0-961).

11. Schroder, J., Uber das Newtonsche Verfahren, Arch. Rat. Mech. Anal. 1 (2) (154-180) (1957).

12. Varga, Richard S.t Matrix Iterative Analysis, Prentice-Hall, (1962).

13. Young, D. M., Iterative methods for solving partial difference equations of elliptic type, Trans. Amer. Math. Soc. 79 (92-111) (1954).

50