NORTH-HOLLAND MATHEMATICS STUDIES Studies in Computational Mathematics (1) Editors:
C. Brezinski University of Lille Vi...
114 downloads
248 Views
5MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
NORTH-HOLLAND MATHEMATICS STUDIES Studies in Computational Mathematics (1) Editors:
C. Brezinski University of Lille Villeneuvedxscq, France
L. Wuytack University of Antwerp Wilrijk, Belgium
NORTH-HOLLAND -AMSTERDAM
NEW YORK
0
OXFORD *TOKYO
136
NONLINEAR METHODS IN NUMERICAL ANALYSIS
Annie CUYT Luc WUYTACK University ofAntwerp Belgium
1987
NORTH-HOLLAND -AMSTERDAM
NEW YORK
OXFORD *TOKYO
‘.c Elsevier Science Publishers B.V.,
1987
Allrights reserved. No part of this publication may be reproduced, stored in a retrievalsystem, or transmitted, in any form or by any means, electronic, niechanical, photocopying, recording or otherwise, without the priorpermission of the copyright owner.
ISBN: 0 444 70189 3
Publishers:
ELSEVIER SCIENCE PUBLISHERS B.V. P.O. BOX 1991 1000 BZ AMSTERDAM THE NETHERLANDS
Sole distributors for the U.S.A. and Canada:
ELSEVIER SCIENCE PUBLISHING COMPANY, INC. 52VANDERBILT AVENUE NEW YORK, N.Y. 10017 U.S.A.
Library of Congress CataioginginPublifation Data
Cuyt, Annie, 1956Nonlinear methods in nurrerical analysis. (North-Holland mathematics studies ; 1 3 6 ) (Studies I n ccnputational mathemtics ; 1) Bibliography: p Includes index. 1. m r i c a l analysis. I. Wytack, L. (Luc), 194311. Title. 111. Series. IV. Series: Studies in ccnputational mathematics : 1.
.
QA297.C89 1987 ISBN 0-444-70189-3
519.4
(U.S.)
PRINTED IN THE NETHERLANDS
86-32932
To Annelies Van Soom from her mother. Annie Cuyt
PREFACE. Most textbooks on Numerical Analysis discuss linear techniques for the solution of various numerical problems. 01ily a small number of books introduce and illustrate nonlinear methods. This book accumulates several nonlinear techniques mainly resulting from the use of Pad6 approximants and rational interpolants. First these types of rational approximants are introducrd and afterwards methods based on their use are developed for the solution of standard problems in numerical mathematics : convergence acceleration, initial value problems, boundary value problems, quadrature, nonlinear equations, partial differential equations and integral equations. The problems are allowed to be univariate or multivariate. The treatment of the univariate theory results from a course given by the second author a t the University of Leuvcm and completed by the first author with many new theorems and numerical rcsiilts. The discussion of the multivariate theory is based on research work by thc first author. The text as it stands is now used for a graduate course in Numerical Analysis at the University of Antwerp. The book brings together many results of research work carried out at the University of Antwerp during the past few yrars. We particularly mention results of Guido Claessens, Albert Wambecq, Paul Van der Cruyssen and Brigitte Verdonk. Let us now give a survey of the conterits of this book and a motivation for the problems treated. Since continued fractions play an important role, Chapter I is an introduction to this topic. We mention somc basic properties, evaluation algorithms and convergence theorems. From the section dealing with convergence we can already learn that in certain situations nonlinear approximations are more powerful than linear approximations. The rervnt notion of branched continued fraction is introduced in the multivariate section and will be used for the construction of multivariate rational interpolants.
In Chapter I1 Pad6 approximants arc’ treated. They are local rational approximants for a given function. The problems of existence, unicity and computation are treated in detail. Also the convergence of sequences of Pad6 approxirnants and the continuity of the Pad6 operator which associates with a function its Pad6 approximant of a certain order, are considered. Again a special section is devoted to the multivariate case. We do not discuss the relationship between Pad6 approximants and orthogonal polynomials or the moment problem.
Preface
In Chapter 111 rational interpolants are defined. Their function values fit those of a given function at some given points. Many results of the previous chapter remain valid for this more general case where the interpolation conditions are spread over several points. In between the rational interpolation case and the Pad6 approximation case lies the theory of rational Hermite interpolation where each interpolation point can be assigned more than one interpolation condition. Some results on the convergence of sequences of rational Hermite interpolants are mentioned and multivariate rational interpolants are introduced in two different ways. The previous types of rational approximants are used in Chapter IV to develop several numerical methods for the solution of classical problems such as convergence acceleration, nonlinear equations, ordinary differential equations, numerical quadrature, partial differential equations and integral equations. Many numerical examples illustrate the different techniques and we see that the nonlinear methods are very useful in situations where we are faced with singularities. However, one must be careful in applying the nonlinear methods due to the fact that denominators in the formula can get small. We tried t o make the text as self-contained as possible. Each chapter also contains a problem section and a section with remarks that indicate extensions of the discussed theory. References to the literature are given at the end of each chapter in alphabetical order. In the text we refer to them within square brackets. Formulas and equations are numbered as (a.b.), where a indicates the chapter number and b the number of the formula in that chapter.
In preparing the text the authors did benefit from discussions with many colleagues and friends. We mention in particular Claude Brezinski (Lille) , Marcel de Bruin (Amsterdam), William Gragg (Lexington), Peter Graves-Morris (Canterbury), Louis Rall (Madison), Nico Temme (Amsterdam) , Helmut Werner (Bonn). We also thank Drs. A. Sevenster from North Holland Publishing Co who encouraged us to write this book and Mrs. F. Schoeters and Mrs. R. Vanmechelen who typed the manuscript. Ant werp
Annie Cuyt Luc Wuytack
1
.
CHAPTER I: Continued Fractions
$1.Notations and definitions
. . . . . . . . . . . . . . . . . . . . .
$2. Fundamental properties . . . . . . . . . . . . . 2.1. Recurrence relations for P,, and Qn . . . . 2.2. Euler-Minding series . . . . . . . . . . 2.3. Equivalence transformations . . . . . . . 2.4. Contraction of a continued fraction . . . . 2.5. Even and odd part . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
3
. . . . . . . . . . . . . . . . . . . . . . . . .
4
$3.Methods t o construct continued fractions . . . . . . . . . . . . . 3.1. Successive substitution . . . . . . . . . . . . . . . . . 3.2. Equivalent continued fractions . . . . . . . . . . . . . . 3.3. The method of Viscovatov . . . . . . . . . . . . . . . . 3.4. Corresponding and associated continued fractions . . . . . . . . . . . . . . . . 3.5. Thiele interpolating continued fractions
12 12 15 16 17 19
. 4 .5 .6 .7 . . . . . . . . .9
$4. Convergence of continued fractions . . . . . . . . . 4.1. Convergence criteria . . . . . . . . . . . . . 4.2. Convergence of continued fraction expansions . . 4.3. Convergence of corresponding cf for Stieltjes series
. . . . . . 20 . . . . . . 20
. . . . . . 25
. . . . . . 29
$5. Algorithms t o evaluate continued fractions . . . . . . . . . . . . 31 5.1. The backward algorithm . . . . . . . . . . . . . . . . 31 5.2. Forward algorithms . . . . . . . . . . . . . . . . . . . 3 2 5.3. Modifying factors . . . . . . . . . . . . . . . . . . . . 34 56 . Branched continued fractions . . . . . . . . . . . . . . . . . . 41 6.1. Definition of branched continued fractions . . . . . . . . . 41 6.2. A generalization of the Euler-Minding series . . . . . . . . 42 6.3. Some recurrence relations . . . . . . . . . . . . . . . . . 47 6.4. A multivariate Viscovatov algorithm . . . . . . . . . . . . 49 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
. . . . . . . . . . . . . . . . . . . . . . . . . . .
56
. . . . . . . . . . . . . . . . . . . . . . . . . .
58
Remarks References
2
“J’ai e u 1 ’honneur de prksenter d 1 ’Acadkmie e n 1802 un mkmoire 6oua le titre : Essai d’une me‘thode ge‘nkrale pour rkduire toutea sortea de skriea e n fraction8 continues. Aprka ce temps ayant eu occaaion de penaer encore ci cette matikre, j’ai f a i t de nouvelles rkjlezions qui peuvent aervir d perfectionner et simplifier la mkthode dont il a’agit. Ge sont ces rkfleziona que j e prksente maintenant d la sociktk savante. ”
B . VISCOVATOV - “De la me‘thode gknkrale pour rkduire toutea sortea de quantit4.9 e n fraction8 continues” (1805).
3
I.1. Notatdom and defindtdona
81. Notations and deflnitions. A continued fraction is a n expression of the form
bo+---
61
--
bl+
-____ b2 ___--___
+ b3+
02
...
a3
' * * +
bi
a;
+ . ..
where the ai and bi are real (or complex) numbers or functions and are respectively called partial numerators and partial denominators. Instead of the expression above we will most of t h e times use the following compact notations:
or
b o + f $ j? = l
(1.1.)
The truncation n=0,1,2,.
is called the nth convergent of the continued fraction (1.1.). If lim
n-+m
G, = C
exists and is finite, then the continucd fraction is said t o be convergent and C is called the value of the continued fraction. Clearly C, is a rational expression (1.2.)
+
where P, and &, are polynomials of a certain degree in the 2n 1 partial numerators and denominators 60, a l l61 I . . ., a,, 6,. The polynomials P, and Q n are respectively called the nth numerator and nth denominator of the continued fraction (1.1.).
4
1.2.1. Recurrence relation8 for
Pn and Qn
§2. Fundamental properties.
2.1. Recurrence relation8 for P, and Qn.
The nth numerators and denominators satisfy the same three-term recurrence relation, but with differentstarting values. This relation is given in the following theorem.
Theorem 1.1.
(1.3.)
Proof The proof is performed by induction. 0bviously
and so the formulas (1.3.) are valid for n = 1. Let us now suppose the validity of (1.3.) for n 5 k. We will prove i t for n = k + 1. We have, using (1.2.),
Consequently, by using (1.3.) for
p1
= k,
5
I. 2.2. Euler-Minding seriea
I
2.2. Euler-Minding 8 eries. It is easy now t o give an expression for the difference of two consecutive convergents of a continued fraction.
Theorem 1.2. If
Q n Qn-1
# 0, then cn- cn-l= ( - I ) n + l
al
a 2 . . .an
Q n Qn-i
(1.4.)
Proof One can show, by induction arid using the recurrence relations (1.3.), th a t for n > l
P,
Qn-l -
Qn
= (-l)n+lal a 2 . . .a,
From this (1.4.) follows immediately. This theorem can now be used to give a more explicit formula for Cn.
I
I. 2.5. Equivalence tranaformationa
6
Theorem 1.3. If
Qi
# 0 for 1 5 i 5 n, then Cn = bo
+
n i= 1
a l . , .ai (-l)i+l -
(1.5.)
Qi-1 Q i
Proof We have, by means of formula (1.4.),
=(-l)n+l
a1. . .a, Q n Qn-1
+ (-1)n
(11.
. .an-l
~
Qn-1 Qn-2
+
,
..+
(11 Qi
Qo
+ bo
I
The expression (1.5.) is the nth partial sum of the series
c 00
bo f
i= 1
(-1)i+1
631..
.a;
Qi-1 Q i
which is called the Euler-Minding series. Thus we have associated a series with a continued fraction such that the nth partial sum of the series equals the nth convergent of the continued fraction. This interrelation between series and continued fractions can be used to apply well-known results for series t o the theory of continued fractions. 2.3. Equivalence tranaformationa. Let p i into
# 0 for i 2 0. The transformation th at alters the continued fraction (1.1.) (1.6.)
is called an equivalence transformation. Clearly (1.1.) and (1.6.) have the same convergents. By performing equivalence transformations a continued fraction can be rewritten in a prescribed form. For instance, if a; # 0 for i 3 1, then (1.1.) can be rewritten as
I.2.4. Contraction of a continued fraction
7
(1.7.) by choosing P1 =
1 -
a1
and 1
pi = ; ; for i 2 2. t
Pt-1
Hence one can limit th e study of continued fractions to continued fractions of the form (1.7.). Such a continued fraction is called a reduced continued fraction [lo p. 4801. 2.4. Contraction of a continued fraction.
Let us consider the following problem. Suppose we are given a sequence (Cn},E- of subsequently differcnt elements and we want t o construct a continued fraction of which C, is the nth convergent.
Theorem 1.4. If C ,
# G,-l
for n
2 1, then the continued fraction bo
+ r=l
with
bi =
c;- c;-z - c;-z
Ci-1
has the elements of the sequence {Cn},en as convergents.
8
1.2.4. Contraction of a continued fraction
Prod We write
with P, = C , and &, = 1. Since the partial numerators and denominators of the continued fraction
bo
+
M
I
i= 1
with convergents C, must satisfy the relations (1.3.1,we get the following system of equations in the unknowns bo, ai and b;:
for
k2 2 :
b; Ci-1 + ai Gi-2 = Ci bi + a; = 1
A solution of this system is given by
(1 3.) I
If also C , # Cn-2 for n >_ 2, then by means of an equivalence transformation the continued fraction with partial numerators and denominators given by (1.8.), can be written as
1.2.5. Even and odd part
9
The formulas (1.8.) and (1.9.) can be used t o compute a contraction of a continued fraction, i,e. a continued fraction constructed in such a way t h a t its convergents form a subsequence of the sequence of convergents of the given continued fraction. We shall now illustrate this. 2.5. Even and o d d part. Consider a continued fraction with convergents { C n } n E ~ . The even part of this continued fraction is a continued fraction with converwhile , the odd part is a continued fraction with convergents gents { C Z ~ } , ~ N { C 2 n + l } n E ~Theorem . 1.4. enablcs us t o construct those even and odd parts. We now derive a formula for the even part and give an analogous formula for the odd part without proof. Consider the continued fraction
with convergents
The partial numerators and dcnomiriators of the even part as expressed in the partial numerators and denominators a ; and b i , are computed as follows. Let
use the formulas (1.8.) and perform an equivalence transformation with
We then get the following continued fraction
+
2
i- 2
with
1.2.5. Even and odd part
10
and
which is equal to
Analogous to formula (1.4.) one can prove t h a t
and that
(see problems (1) and (2) a t the end of this chapter). Finally we get for the even part
bo +
~ a 1 b2 _
lbi bz
+z 00
+ a2
I_
4 2 i - 2
l r n i - 2
I
__ _a 2 a_3 b4_ ~ ~ -
/ ( b z b3
+ a 3 ) b4 + b2 a 4
a2i--1
hi--1
b2i--4
+ a2i--1)
b2i
-t b2i-2 a2!
(1.1Oa.)
11
I.2.5. Even and odd part
In the same way one can prove that the odd part is
We illustrate this procedure with the following example. To compute C = 6 with r real positive, we first write r = b2 positive. So C2- b2 = a or
Hence
i= 1
If we take
t
= 102, we get 415- = 10 +
c-
i= 1
The even part of this continued fraction is
and the odd part is
2!
+a
with a
12
I.S.1. Succeaas've subatitutdon
83. Methods to construct continued fractions. We are now going t o describe some methods t h a t can be used t o write a given number or function as a continued fraction. Other techniques can for example be found in [lo pp. 487-5001 and [12 pp. 78-1501. Some convergence problems of such continued fractions are treated in the next section. 3 . j . Successive subatitution.
Let f be a given number or function. Write To = j and compute TI, T2,.. .,Tn+l such that
where b o , a; and b; are chosen freely and can be functions of the argument of f. In this way we get
By continuing this method of successive substitution we get an expression of the form (1.1.). It is important t o check whether f is really the value of this continued fraction or for which arguments of f this is true. Such problems are treated further on. We shall illustrate this method by calculating a continued fraction expansion for e z . Since
we choose bo = 1.
1.8.1. Successive substitution
13
So we get
TI= e z - l = z
b X + ( X
-
2
1
22 ++ ... 12
which suggests us t o write
with
Some easy computations show that
T2= X
1
2 1-(;
+...)
2-(;
2-2
So we can write
with
+...)
+
( y + .) ..
14
I.3.1. Successive substitution
T3
22 3
+ . . .)
= --(1
If we continue in this way, we find th at for a; = (i - 1). and i+l
which again suggests the choice
k=l
ai+l = i s
(d-
and
b,+l
= (i
bi = i
- s,
+ 1) - x, such th a t
1
1)xI + ~ (n1)s - _ _ _ In - 2 + Tn+l
i - 2
or by continuing the process [12 p.1301
eZ=l+
Another examp
I
2
__
11-x
i- 2
(1.1 1.)
is given by the construction of a continued fract.an for
C=
fi)= q T T x g
with s and y real positive. Proceeding as in the previous section we can write
C2= I ( x , Y) = b2(x, Y) + 4 5 , Y)
with b(s, y) = x + y and a(x, y) = sy. Then
and consequently
1.3.2. Equavalent continued fructiona
15
3.2. Equivalent continued fractionrr.
A series m
i=O
and a continued fraction bO f i=l
are called equivalent if for every n
2 0 the nth partial sum n i=O
of the series equals the nthconvergent
Remember that the Euler-Minding series and the continued fraction (1.1.) are equivalent, 'hmsforrning a given series into an equivalent continued fraction can also be done by means of formula (1.9.), with n
Gn =
C
d;
i=O
We obtain (1.12.)
In the following example Consider
czod ; is a power series.
16
I.S.S. The method of Vimovatov
Then after an equivalence transformation (1.12.) is
1+
1$
+cIT-
- ~ - _ _ (i - 2)! i! +’Tx 1 (%- I)! t.
00
i=2
If we perform another equivalence transformation, we get
(1.13 .) Remark t h a t a n equivalent continued fraction will converge if and only if the given power series converges. For our example this means that (1.13.) converges on the whole complex plane. Hence, by substitution of x by -2, (2
- l)z/
i=2
converges for all x. Consequently the continued fraction in the righthand side of (1.11.) converges for arbitrary 2.
3.3. The method of Viscovotou. This method is used t o develop a continued fraction expansion for functions given as the quotient of two power series [27]. Let
Then
I.3.4. Correeponding and uaaociated continued fraction8
17
This procedure can be repeated and if we define dk,i
for k
= dk--l,O
dk-2,i+l
- dk-2,O d k - l , i + l
> 2 and d 2 0, we finally get. (1.14.)
In case do; = 0 for i 2 1 then f ( x ) is a power series itself and we shall prove in the next section t h a t the method of Viscovatov can be used t o compute a corresponding continued fraction. If f(x) is the quotient of two polynomials and thus a rational function, this method can be used t o write f in the form of a continued fraction (see also problem (8)).
3.4. Cotreaponding and associated continued fraction8
A continued fraction
for which the Taylor series development of the nth convergent C,(z) origin matches a given power series
around the
M
i- 0
up t o and including the term of degree n is called corresponding t o this power series. In other words, for a corresponding continued fraction, if
then for every n we have e i = ci for i = 0 , . . ., n. A lot of methods t o construct corresponding continued fractions are treated in chapter 11.
18
I.3.4. Corresponding and aaaocdated continued fraction8
A continued fraction
for which the Taylor series development of the nth convergent C,(z) matches a given power series
c co
c; x i
i= 0
up to and including the term o degree 2n is called associatec A corresponding continued fraction can be turned into an associated one by calculating the even part. Let us now again consider the algorithm of Viscovatov. For the continued fraction (1.14.) we define
fo = f given by (1.14.) fl
= dl0 - doofo
fk
= &,02 fk-2
- dk--l,O
Then by induction it is easy to see th at the form
k
fk-1 fk(X)
= 2 , 3 , 4 , .. .
can be developed into a series of
i=O
One can also prove by induction th at for the kth convergent
pk _ - dl01 &k Id00 ~
ZOzl +... + I-ddl0 + -d.3051 d20
+
~
the relation
holds. Hence if f(s) is given by the series expansion f ( 2 ) = co
+ c1z +
C2Z2
+ ...
dk,Oz /dk--l,O
I
I.8.5. Thiele interpolating contanued fraction8
19
the algorithm of Viscovatov when applied t o (f(z)- C O ) / Z with d l ; = c l + ; for i 2 0, generates a continued fraction of the form
of which the Taylor series development of the kth convergent matches the power series (f(z) - cO)/z up t o and including the term of degree k - 1. In this way we obtain for f(z) the corresponding continued fraction co
+
( p -+ id20 d30zl +...
c l z l 4- dzozi
because the Taylor series development of the kth convergent matches the power series f ( z ) up to and including the term of degree k.
3.5. Thiele interpolating continued fract io n8. This technique is dealt with completely in chapter 111. It uses interpolation data and reciprocal differences. Thicle typo continued fractions will be constructed both for univariate and multivariate functions. In the univariate case continued fractions of the form
b"
+
2
i=l
/ad. - GI! 6;
will be used while we shall need branched continued fractions similar t o those in (1.29.) for the multivariate case.
I.d.1. Convergence criteria
20
54. Convergence of continued fractions.
4.1. Convergence criteria. The following result is a classical convergence criterion for reduced continued fractions, and is due to Seidel [21]. It dates from 1846.
Theorem 1.5. If
bi
> 0 for i 2
1, then the continued fraction
bo
converges, if and only if the series
+
c
i= 1
czl
bi
diverges.
Proof The Euler-Minding series for
It is an alternating series because a; = 1 and bi > 0 imply t h a t &i > 0 for i 2 1. The nth denominator & n is bounded below by 0 = min(1, b l ) . This can be proved by induction from the recurrence relation for Qn. If we put r n = Q n Qn-l then the rn are monotonically increasing because
Consequently
1.4.1.Convergence criteria
21
czl
Thus if bi diverges, the sequence {rn}nEmtends t o infinity. This implies, by a theorem of Leibniz on alternating series, t h a t the EulerMinding series
c:,
converges. On the other hand, if b , converges then we can prove t h a t rn is bounded above. To do so, we first prove that Qn
< (1 + 6 1 )(I + 62). . .(I + bn)
This is obvious for n = 1 . Assume now that it is also true for n the recurrence relations,
Consequently
Q~ because ez If we put
< ebl
e b 2 . . .ebn
> 1 + z for z > 0. M
i= 1
and u = e+
then
5 k, then using
1.4.1. Convergence criteria
22
This implies that the terms of the Euler-Minding series do not converge t o sero. I As an example, consider the continued fraction
An equivalence transformation rewrites it as I
I
11 -I- -11 l o + -(10 120
I
I
+
11 110
+
11 120
+...
Clearly this continued fraction converges. Note that the convergents satisfy (yn - l o = -
2
20
+ (Cn-* - 10)
such that for the value G of the continued fraction
c = 10 +
2 10+c
or
G2 = 102
m.
Since all convergents are positive we get c = The next theorem is valid for continued fractions of the form (1.1.)with bo = 0, also if the partial numerators or denominators are complex numbers. It was proved by Pringsheim [18] in 1899.
Theorem 1.8. The continued fraction
converges if jb;l 2 1u;I IC,( < 1 if n 2 1.
+ 1 > 1 for d 2 1. For the nth convergent G, we have
1.4.1. C o n v e r g e n c e c r i t e r i a
23
Proof First we prove the upper bound on G,. Let =
An(.)
an
+z
-__
b,
Then
Thus
lc,~=
0 82 0
.,.
8n(0)1
For the nth denominator Q n we have
and hence
JQnI - En-1 I 2
(IbnI - 1)(lQn-ll-
IQn-21)
Consequently IQnI
- IQn-11 2
nkW1 - 1) 2 nkm1 n
(1bkl
n
lakl
for n
2
1
This implies t h a t the IQ, are rnonotonically increasing and that the general term in the Euler-Minding series satisfies
So we have
1.4.1.Convergence criteria
24
and hence the Euler-Minding series converges absolutely. This implies convergence of the given continued fraction. I The earliest known convergence criterion for continued fractions with complex elements is the following result of Worpitzky [29] which can easily be proved using the preceding theorem. We don't give the original proof of 1865.
Theorem 1.7. The continued fraction
converges if
for e'
2 2.
Proof Apply the previous theorem t o the continued fraction
which is equal t o
up t o a n equivalence transformation with p i = 2 for all i
2 1.
I
An extensive treatment of the convergence problem of continued fractions is given in [ll pp. 60-1461. We also refer to the work of Wall [28] and Perron [17].
1.4.2. Convergence of c o n t i n u e d f r a c t i o n ezpamions
25
4.2. C o n v e r g e n c e of continued f r a c t i o n expanaions.
Let z be a complex number. Consider the continued fraction expansion
This expansion is only defined for those values of z for which the continued fraction converges. For some choiccs of bo(z),a i ( z ) and b i ( 2 ) a lot is known about the convergence of such a continued fraction expansion. We will give results for expansions of the following type:
(1.15a.)
and (1.15b.)
A lot of continued fractions can also be written in one of the forms above by means of an equivalence transformation. Theorem 1.8.
If for the continued fraction ( 1 . 1 5 ~ ~all) the b ; are real and strictly positive with b i divergent, then (1.15a.) converges in every closed and bounded subset G of the complex plane for which the distance to the negative real axis is positive. The convergence is uniform t o a function f (2) which is holomorphic for all z not on the negative real axis.
XE,
26
I.4.Z. Convergence of continued fraction ezpanaiona
Domain of uniform convergence of the continued fraction ( l . l 5 a . ) ,
t
Figure 1.1. This theorem was originally proved by Stieltjes [23, 28 p. 1201 in 1894. We illustrate it with t he following example. A continued fraction expansion for the function [lo p. 6141
is
'I +
1/21 11 4- -
Iz l 1
+
3/21
Iz l 1 --
+...
or after an equivalence transformation
According t o theorem 1.8. this continued fraction expansion converges for all z not on the negative real axis. The next result is due t o Van Vleck [20 p. 3941 and dates from 1904.
1.4.2. Convergence of continued fraction ezpansionrr
27
Theorem 1.9. If for the continued fraction (1.15b.) lim;-m a; = 0 with a; # 0, then (1.15b.) converges to a meromorphic function f(z).The convergence is uniform in every closed and bounded subset G of the complex plane that contains no poles of
f (4.
Jf for the continued fraction (1.L5b.) limi-m verges in the cut complex plane C\{z
12
=-
ai = a
x- x 2 4a’
# 0,
then (1.15b.) con-
l}
t o a function f ( z ) meromorphic in t h a t cut complex plane. The convergence is uniform in cvrry closed and bounded subset G of the complex plane that contains no poles of f(z) and no points of the cut {z
x
12 = - - - - , A
4a
> 1) -
D omain of uniform convergence of the continued fraction (1.15b.)
Figure 1.Z.
Let us also illustrate the previous theorem with an example. A continued fraction expansion for
28
1.4.2.Convergence of continued fraction expanaiona
is [l p. 881
-9122-
(2n + 1)(2n - 1)
r
given by z
-
3
Zn + .. . + z52 + z73 + . . . + _2n-f 1 --
only converges in t h e unit disc. Hence continued fraction expansions can have a larger convergence region. A related problem is the following. If a function f ( z ) is given, is it then possible t o construct a continued fraction expansion for j ( z ) such that t h e convergence region of the continued fraction expansion is exactly the domain of f(z)? This qiiestion has partly been answered when we discussed equivalent continued fractions. We do not study the problem in detai1 here but refer the interested reader to [20 pp. 386-4151 and [lo].
29
I.4.8. Convergence of correspondang continued fractions
4.3. Convergence of corresponding continued fractions for Stieltjea eeriea. The series
C di zi 00
f(z) =
(1.16.)
i= 0
is called a Stieltjes series if d; =
1"
t'dg(t)
where g(t) is a real-valued, bounded, nondecreasing function taking on infinitely many different values. The values d; are called moments of the function g ( t ) . If g ( t ) is constant for t > r with 0 < r < 00, then
The Stieltjes transform of g ( t ) is defined as
(1.17.)
A proof of the existence of this transform is given in [lo p. 5781. The series (1.16.) has convergence ridius 1/r and can be regarded as a formal power series expansion of F ( z ) which is analytic in the cut complex plane @\[$,m) [I0 p. 5811.
Theorem 1.10. The corresponding continued fraction for f(z)given by (1.16.) converges t o F ( z ) given by (1.17.) for all z in C \ [ t , co). The convergence is uniform OIL every closed and bounded subset of the cut complex plane. The proof which was originally given by Markov can be found in [17 p. 2021. A simple example of a Stieltjes series is
Here g ( t ) = t for 0
5 t 5 1 and
g ( t ) = 1 for
t 2 1.
30
1.4.3. Convergence of Corresponding continued fractions
The Stieltjes transform is 1 F ( z ) = -- ln(1- z ) z
As a consequence of theorem (1.10.) we get that the corresponding continued fraction for f converges to F for all z in C\[l, cm).
1.5.1. The backward algorithm
31
§5. Algorithms to evaluate continued fractions. If we want t o know an approximation for the value of a continued fraction, we must compute one or more convergents C,. The recurrence relations (1.3.) for the nth numerator and denominator provide a means t o calculate the ntA convergent since
This algorithm is called forward because i t is possible t o compute Cn+l from the knowledge of Cn with little extra work. We shall now discuss some other algorithms used for the computation of convergents.
5.1. The backward algorithm. The nth convergent C , can easily be calculated as follows: Put
rn+l,n = 0 and compute
Then
A drawback of this method is that it must fully be repeated for each convergent we want t o compute. It is impossible t o calculate Cnfl starting from Cn. But the algorithm appears to be niinierically stable in a lot of cases [2].
I. 5.2. Forward algor%'thms
32
5.2. Forward algorithms. The following theorem can be found in [14].
Theorem 1.11. The nth convergent of the continued fraction
i=l
is the first unknown
XI,,
li
of the tridiagonal system of linear equations
...
bl
-1
0
a2
b2
- 1
0
a3
b3
...
0
*.
...
-1
0
0
a,
=I
0
a1
0
(1.18.)
0
b,
(see also problem (5)). Consequently, algorithms for the solution of a linear tridiagonal system, and especially for the computation of t h e first unknown, are also algorithms for the calculation of C,. If backward Gaussian elimination is used to solve (1.18.), then the coefficient matrix of (1.18.) is transformed into a lower triangular matrix and the computation of ~ 1 =,r l , ~n is precisely the backward algorithm. It is also easy to see t h a t the computation of C , via the recurrence relations (1.3.) is equivalent with solving (1.18.) by means of the so-called shooting method. If we choose XI,, = z?; = 0 then according to (1.18,) z f L = --a1 and (0) Xk+l,,
If we choose
~
1 =,
(1) Zk+l,n
(0) - akxk-i,,
+ b k x k(0) ,r,
zit!,~ = 1 then
$1
k
= 2 , . .. ,? -I1
= b l - a1 and
- a k Z k( -11) , n + b k x k( 1, n)
k
= 21 . . . 1
-
The last equation of the linear system (1.18.) is after substitution of t h e first (n- 1) equations in it, merely a linear function g(xl,n). The zi,, we are looking for is a root of g(zl,,), in other words the intersection point of the z-axis with the straight line y = g(z).
I.5.2. Forward algortthma
33
so
or
After comparison of the starting values for the recursive formulas (1.3-)1when applied to
28
i=l
with those for the recu sive cornpiitation of
since the recurrence relations are identical.
2 k( 0, n)
34
1.5.3. Modifying factore
Another forward algorithm for the computation of C , is obtained when (1.18.) is solved by forward Gaussian elimination and backsubstitution. The resulting formulas are
ri,n = b;
+ t i -ail . n
e' = 2 , . . .,n
Finally C , = bo + ~ 1 , ~ . More algorithms for the calculation of C, can be found in the literature. We refer among others to [25] and [6].
5.3. Modifying factor8. Even efficient ways t o calculate C, do not guarantee that, C, is a good approximation for the value G = lim C , of the continued fraction. Since the nth n-yw convergent results from truncating
we shall call the chopped off part
the nth tail of the continued fraction. Clearly n-1
I
(1.19.)
and
35
I. 5.5. Modifying factore
Let again for n 2 1 &(5)
an = -
b,
+z
Then
and
Hence, in order to estimate G , it may be better to replace the tail T,,by a value different from zero. In many cases the tails do not even converge to zero. Suppose r,, is such an approximation for T,,.We shall then calI
I’,
= bo
+ a1
o 92 o
. . . o s,(m)
the nth modified convergent with 7, the nth modifying factor. The next theorems illustrate in which cases modifying factors are really worthwile, in the sense that
First of all we study the behaviour of the tails [17 p. 931.
Theorem 1.12. If the continued fraction (1.1.) is such that lim a; = a and lim b; = b with
+
t+m
c+m
a,b E G and if the quadratic equation x2 bx - a = 0 has two roots 5 2 with 1x11 < 1221 then lim T,,= z1 ,403
x1
and
36
1.5.8. Modifying factors
This behaviour of the tails suggests to choose
In order to study the effect of this modifying factor we rewrite the expression IC - rnl/1C - CnI as follows, Using the three-term recurrence relation for (1.19.) we find
Analogously
This leads to
with
where
ho = 03
If the continued fraction (1.1.) satisfies the conditions of theorem 1.12. then
1.5.3.Modifying factors we shall denote
D
+
37
- 1~11 d , = inax lam - a \ =
Ib
511
m>n
en = inax Ib, - bl m>n
En =
dnD D2 - 2 d ,
+ 0, = b + b,,
a, = a
b,
Tn = 51 + En B, = h,
+ 51
Using these notations we can formulate the next theorem [24].
Theorem 1.13. Let the continued fraction
bo be such that Jim t-m
ai =
+
c-1 4 i= 1
6;
a and lim b; = b with a , b E C and let s-
00
51
be the
strictly smallest root of the quadratic equation x 2 + b x - a = 0 with
If also
(1.20c.) then
38
I. 5.3.Modifying factors
Proof Let us first show that in (1.20b.) we have Ib d, 5 D a / 3 and this implies
since
0 5 Hence or
16 +
+ 511 2
En. We know that
4- l ~ l 5l 1
Ib + 511
( D 2- 2 d n ) l b + 5 1 1 2 Dd, J b + 5 1 ) 2 En
In order to bound I(C - r n ) / ( C - Cn)\ we shall now calculate an upper bound for I(Tn - zl)/Tn\. Since lim Tn =
n-+m
and
51
Iiin b , = b
n-co
we have lim
n-m
En = 0
and and hence for fixed n We can then write for fixed k > max(m,ra)
1.5.9. Modifying factors
and
because
Using this upper bound for J f k - 1 -t Bk-11 we can also prove it to be an upper bound for \ & 2 + & 2 \ . Repeating this procedure as long as l a k l I d, and J B k - i l 5 e n , finally assures for k - I = n IFn
NOWsince dn _< la1/2 = I(b
+z
+ Pnl 5 E n
l ) I /~2 ~ and 2dn _< 2 D 2 / 3 we have
Next we shall compute an upper bound for Ihn/(h, + x1)1 which is the second factor in I(C - rn)/(C- Cn)l. Note already that
39
40
I. 5.3. Modifying
factora
By induction it follows that
bec,ause
This gives us
Using the estimatfesfor I(Tn - 21)/Tn\ and Jhn/(hn finish the proof. We have
+ XI)\
it is now easy t o
1.6.1. Definition of branched continued fraction8
41
56. Branched continued f'ractions.
6.1. Definition of branc hed continued fractions. If the denominators bi in the continued fraction
are themselves infinite expressions, then it is called a branched continued fraction. The b; are called the branches and we need a multi-index t o indicate a convergent. Consider for instance thc expression ai
1
The (n,mo, nl,...,nnfth convergent is then the subexpression
i= 1
We will use branched continued fractions to construct a multivariate Viscovatov algorithm for the computation of multivariate continued fraction expansions of the form
where k is the number of variables we are dealing with. Input of such an algorithm is a multivariate power series. Vice versa, given a branched coritiniied fraction, we can also construct an EulerMinding series of which the successive partial sums equal a given sequence of convergent s .
42
Z.6.2. A generalization of the Euler-Minding series
6.2. A generalization of the Euler-Minding aeries.
Let us consider continued fractions
( 1.21a.) for i = 0 , 1 , 2 , . . .. If C c ) denotes the nth convergent of (1.21a.) then according to (1.4.)
where we have written
We will now generalize (1.4.) for the branched continued fraction (1.21b.) Let us denote by Pn/Qnthe subexpression
(1.22.)
So P,/Qn is the (n,n,n - 1 , . . . , l , O ) t h convergent of (1.20.). Another subexpression we shall need is
1.6.2. A generalization of the Euler-Minding aerie8
43
( 1.23.) which is in fact the k t h convergent of Pn/Q,. These subconvergents can be ordered in a table
where we proceed in a certain row from one value to the next one by using (1.3.) for (1.22.) :
with RPj = 1 = Sin), R P ) = Cia) and Sl;' = 0. If we want t o develop a formula analogous t o (1.5.) for the branched continued fraction (1.21.) we must compute a n expression for t h e difference
(1.25.) Remark t h a t in comparison with P n - l / Q n - l the expression Pn/Qn contains an e x t r a term in each of the involved convergents of B;. Also Bn is not taken into account in P n - l / Q n - l . In order t o compute (1.25.) we must b e able to proceed from one row in the table of subconvergents t o the next row. The following t.heorem is a means to calculate the differences - RF-') and Sin) - sin-') Rk
44
1.6.2.A generalization of the Eder-Minding series
Theorem 1.14.
F o r n 2 2 a n d k = 1, ..., n - 1
and
Proof We shall perform the proof only for Rp' - Rf-') because it is completely analogous for sin) . Choose k and n and write down the recurrence relation (1.24.) for row n and row n - 1 in the table of subconvergents:
sin-')
R P ) = cn-kRk-l (k) ( n ) + a k R k(-n2 )
By subtracting we get
1.6.2. A generalization of the Euler-Minding series
45
The first three starting starting values are easy t o check and for R p ) -
I
again (1.4.) is used.
From t h e above theorem we see t h a t up t o an additional correction term the values R f ) - R P - l ) a nd Sin)- Sin-') also satisfy a three-term recurrence relation. By means of this result we can write for the numerator of ( 1 . 2 5 . ) :
because Rn-2 (n-1) / S L i l ) and Rn-l (n-1) /SLq1) are consecutive convergents of th e finite continued fraction n-1
t
1.6.2.A generalization of the Euler-Minding series
46
h this way
(1.26.)
We remark that (1.26.) reduces t o (1.4.) if the continued fraction (1.21.) is not branched because then R r ) = R p ) and Sikj = Sin) for all n 2 k. Consequently the classical Euler-Minding series will t u r n out to be a special case of the Euler-Minding series for branched continued fractions.
Theorem 1.15.
For n 1 2 the convergent C n , n , n - l , . . . , lof , ~ the branched continued fraction (1.21.) can be written as
I
&
Qi
i=2
Qi-1
Qi
47
1.6.8. Some recurrence relataono
Proof The result is obvious if we write
and insert (1.26.) for P i / Q i
I
- P,-1/&,-1.
As a result of the previous theorcm we can associate with the branched continued fraction (1.21.) the series
I
Qi
J
Qi
Qi- I
of which the successive partial sums equal the successive convergents Cn,n,n-i ,...,1,0 of (1-21.). 6.8. Some recurrence relations.
In order to formulate a multivariate Viscovatov algorithm we first show that a corresponding continued fraction can be obtained from a system of recurrence relations [ 161. Consider the problem of constructing a continued fraction expansion of the form
f(x) =
,f$ +
4 +
I 1
...
(1.27.)
for a given series expansion
Remark that (1.27.) coincides with (1.14.) after an equivalence transformat ion. Instead of using Viscovatov's algorithm, the coefficients a; can also
I . 6.3. Some recurrence relata'ons
48
be deduced from the following set of recurrence relations. Define fo = f
given by
(1.27.)
- fo
I1
=
fk
= akxfk--2 - f k - 1
a1
As mentioned in section 3.4.
fk(2)
k = 2,3,4,. . .
(1.28a.) (1.28b.)
can h e developed in a series of the form 00
i= 1
Equating coefficients in relation (1.28a.)
i= 1
i=l
we find
and for k 2 2 by means of (1.28b.)
we get
Using these formulas all the coefficients in the continued fraction (1.27.) can be computed. Hence we can also construct a continued fraction expansion of the form i= 1
1
I.S.4. A multivciriate Vaacovatov algorithm for the power series f(x) = c r ) reasoning t o (f(x) - co(0))/x.
49
+ c(l(l) z + c p ) z2+. . . by applying the previous
6.4. A multivariate Viacovatov algorithm
Let us now apply this reasoning t80the following problem [15]. We restrict ourselves t o the bivariate case only t o avoid notational difficulties. Given a double power series
try to End a branched continued fraction of the form
We define
fo = f fl
given by
= a l l - (1
fk =
(1.29.)
+ g 1 + h1)fo
a k k z y f k - 2 - ( 1 $. g k -k
A series expansion for
fk(z,
hk)fk-l
y) is then of the form
k
= 2 , 3 , 4 , .. .
(1.30a.) (1.30b.)
50
I.6.4. A multivariate Viscovatov algorithm
while gk(z) and h k ( y ) can be written as (1.31.f i= 1 00
(1.32 .)
Equating coefficients in formula { 1.30a.)
we obtain
i--1
and doing the same with (1.30b.)
we find for k 2 2
I.6.4.A multivariate Viscovatov algorithm
The coefficients a k + i , k and a k , k + i are computed by applying (1.27.) and (1.28.) to t,he series (1.31.) and (1.32.) As a consequence one can obtain a coutinued fraction expansion of the form
for a double power series
by applying the previous reasoning t o the power series
and compute the coefficients n;o and a o i in
51
52
I.6.4. A multivariate Viscovatov algorithm
by applyicg (1.27.) and (1.28.) to the series
i= 1
c
and
co
c$'yj
j= 1
To illustrate this technique we consider 'the following example. Take 1 1 + -12z 2 + z y + -g21 2 + 61 3 + -&+ --5y2+ 2 2 1 1 + 241 4 + 61 3y + -412 y 2 + -xy3 + -y4 + . . . 6 24 1 1 1 1 = 1 + z(l + + 6-z2 + . . .) + y ( l f -I/2 + -I/2 + . . .) 2 6 1 1 1 + q ( 1 + 12 + -9 + -61x 2 + -sy 1- -r/2 + . . .) 2 4 6
=l+z+y
--5
--5
1
-y3 6
--2
--2
--5
Since the problem is completely symmetric it is sufficient to calculate the coefficients aiO, d:") and el!,"' with i 5 j. Then
Using the above formulas we obtain
I.S.4. A multivariate Viscouatou algorithm and
53
1
a22 = 2
p) = 1 2 36
d y ) = 16.
*.-
Applying (1.27.) and (1.28.) to
gl(s) -- -
- -1+
-1x + o x 2 12 hl(?/) - - - - +1 i p +1 O y Y 2 X
2
+ ...
2
+...
and
we finally get for
(ez+Y
- e2
- eY
+ l ) / z y the branched continued fraction I
1 ~
11 +
( +
and for
xY14
I1+
.)1+
.. .)'
...
..
.) + (
+
..
.) + (
IF+ .. .) +
+ ..
er+g
1+
($+
I*+
XY
\I+
Iq+ ,*+
. .) + (,*+
I++
...) +
(,-+ + II+
(p+
4
lq+
4
+ (lqJ+
+ ...
I
I. Problems
54
Problems. (1)
Prove that
(3)
Prove formula (l.lOb,) for the odd part of a continued fraction.
(4)
The convergents of the continued fraction
with b;
> 0 for t 2 0 satisfy
a) c can},,^ is a monotonically increasing sequence. b) { C 2 n + l } n E is ~ a monotonically decreasing sequence. c) for n and rn arbitrary: Czrn+l > C2n
(5)
a) The nth numerator and denominator of the continued fraction
satisfy
Pn =
55
I . Problems
Qn=
...
0
63
...
0
*.
-1
0
an
1
0
bl
-
a2
b2
- 1
0
a3
0
...
bn
b) Also c) Prove theorem 1.11. a) Construct a continued fraction with convergents c n
+
= (1
%)(I + 71)..
+ 7n)
where rk(l ’yk) # 0 for k 2 0. b) Use it t o give a continued fraction expansion €or
-_ sin(nx) TX
- (1
(
- x)(l+ x ) 1 - ;)(I
+ ;)(I - ;)(1+
;). .
If are the convergents of a given continued fraction, construct a continued fraction with convergents
where a E R. This procedure is called the extension of a continued fraction.
(10)
How is the method of Viscovatov to be adapted if
d20
Give a continued fraction representation for gence.
and discuss its conver-
Prove formula (1.26.) using (1.24.) n - 1 times.
= O?
I. Remark5
50
Remarks. (1)
Facts about the history of continued fractions can be found in [5]. This history goes back t o Euclids algorithm t o compute the greatest common divisor of two integers (300 B.C.) but the first conscious use of continued fractions dates from the 16th century.
(2)
The notion of nth numerator and denominator satisfying a three-term recurrence relation can be generalized to compute solutions of a ( k 2) term recurrence relation with k + 1 initial data:
+
The (k + 1)-tuple of elements
is then called a generalieed continued Fraction [26].
(3)
More general forms of coutinued fractions where a i and bi are no longer real or complex numbers, are possible. We refer to the works of Fair is], Hayden [9], Roach [19], Wynn [30, 311 and Zemanian [32]. A lot of references on the theory of continued fractions can also be found in the bibliographies edited by Brezinski [4].
(4)
If a continued fraction
with nth convergent C,, converges to a finite limit C , then C - C , is called the nth truncation error. An extensive analysis of truncation errors is given in [I1 pp. 297-3281.
(5)
Another type of bivariate continued fraction expansions can for instance be found in [22]. They are of the form
I. Remarks
57
where the continued fractions B!J'(xy) are given by
i = 1,2, j = 1,2
and obtained by inverting power series. More types of branched continued fractions are given in [3] and [13].
58
I. References
References.
[
I ] Abramowite M. and Stegun I. Handbook of Mathematical functions. Dover publications, New York, 1968.
[
21 Blanch G . Numerical evaluation of continued fractions. SLAM Rev. 6, 1964, 383-421.
[
31 Bodnarc'uk P. and Skorobogatko W . (in Russian) Branched continued fractions and their applications. Naukowaja Dumka, Kiev, 1974.
[
41 Brezinski C.
[
History of continued fractions and Pad6 approximants. 51 Brezinski C. Springer, Heidelberg, 1986.
[
61 Cuyt A. and Van der Cruyssen P . Rounding error analysis for forward continued lraction algorithms. Comput. Math. Appl. 11, 1985, 541-564.
[
A bibliography on Pad6 approximation and related subjects. Publications ANO, Universiti: de Lille, France.
71 de Bruin M. and van Rossum €I.
Pad6 Approximation and its applications. Lecture Notes in Mathematics 888, Springer , Berlin, 1981.
[
81 Fair W. Noncommutative continued fractions. SIAM J. Math. Anal. 2, 1971, 226-232.
[
Continued fractions in Banach spaces. Rocky Mountain J. 91 Hayden 2’. Math. 4, 1974, 357-370.
[ 101 Henrici P.
Applied and computational complex analysis: vol. 2. John Wiley, New York, 1976.
[ 111 Jones W. and Thron W .
Continued fractions: analytic theory and applications. Encyclopedia of Mathematics and its applicalions: vol. 11, Addison-Wesley, Reading, 1980.
[ 121 Khovanskii A. generalizations t o Groningen, 1963.
[ 131 Kuchminskaya K.
The application of continued fractions and their problems in approximation theory. Noordhoff,
(in Russian) Corresponding and associated branched continued fractions for double power series. Dokl. Akad. Nauk Ukrain.SSR Ser. A 7, 1978, 614-617.
I. References
\
59
141 MikloSko J . Investigation of algorithms for numerical computation of continued fractions. USSR Cornputational Math. and Math. Phys. 16, 1976, 1-12.
[ 151 Murphy J . and O’Donohoc M.
A two-variable generalization of the Stieltjes-type continued fraction. J . Comput. Appl. Math. 4, 1978, 181190.
[ 161 Murphy J . and O’Donohoe M.
Some properties of continued fractions with applications in Markov processes. J . Inst. Math. Appl. 16, 1975, 5771.
[ 171 Perron 0.
Die Lehre von den Kettenbruchen 11. Teubner, Stuttgart,
1977.
[ 181 Pringsheim A.
Uber die Konvergenz unendlicher Kettenbriiche. S.-B.Bayer. Akad. Wiss. Math.-Nat. KI. 28, 1899, 295-324.
[ 191 Roach F .
Continued fractions over an inner product space. AMS Proceedings 24, 1970, 576-582.
[ 201 Sauer R. and Szabd F .
Mathematische Hilfsmittel des Ingenieurs 111.
Springer, Berlin, 1968.
[ 211 Seidel L.
Untersuchungen iiber die Konvergenz und Divergenz der Kettenbriiche. Habilschrift, Munchen, 1846.
[ 221 Siemaszko W.
Branched continued fractions for double power series. J . Comput. Appl. Math. 6, 1980, 121-125.
[ 231 Stieltjes T .
Recherches sur les fractions continues. Ann. Fac. Sci. Toulouse 8, 1894, 1-22 and 9, 1894, 1-47.
[ 241 Thron W. and Waadeland H .
Accelerating Convergence of Limit Periodic Continued Fractions K ( a n / l ) . Numer. Math. 34, 1980, 155-170.
[ 251 Van der Cruyssen P.
A continued fraction algorithm. Numer. Math.
37, 1981, 149-156.
[ 261 Van der Cruyssen P.
Linear Difference Equations and Generalized Continued Fractions. Computing 22, 1979, 269-278.
[ 271 Viscovatov B.
De la mCthode g6n6rale pour reduire toutes sortes de quantitCs en fractions continues. MCm. h a d . Impkriale Sci. St-Petersburg 1, 1803-1806, 226-247.
I. References
60
[ 281 Wall H.
Analytic theory of continued fractions. Chelsea, Bronx, 1973.
[ 291 Worpitzky J.
Untersuchungen uber die Entwickelung der monodronen und monogenen Funktionen durch Ket,tenbriiche. Friedrichs-Gymnasium und Realschule Jahresbericht, 1865, 3-39.
[ 30) Wynn P.
Continued Fractions whose coefficients obey a noncommutative law of multiplication. Arch. Rational Mech. Anal. 12, 1963, 273-312.
[ 311 Wynn P.
Vector continued fractions. Linear Algebra 1 , 1968, 357-395.
Continued fractions of operator-valued analytic func[ 321 Zernanian A. tions. J. Approx. Theory 11, 1974, 319-326.
61
.
CRAPTER II: Pad6 Approximants 51. Notations and definitions
. . . . . . . . . . . . . . . . . . . .
52 . Fundamental properties . . . . . . . . . 2.1. Properties of the Pad4 approximant 2.2. Block structure of the I’ad6 table . 2.3. Normality . . . . . . . . . . . .
. . . .
§3. Methods t o compute Pad6 approximants 3.1. Corresponding continued fractions 3.2. The qd-algorithm . . . . . . . 3.3. The algorithm of Gragg . . . . 3.4. Determinant formulas . . . . . 3.5. The method of Viscovatov . . . 3.6. Recursive algorithms . . . . . . 3.7. The <-algorithm . . . . . . . .
. . . . . . . . . . . . 76 . . . . . . . . . . . . 76
. . . . . . . .
. . . . . .
. . . .
. . . . . .
4.4.
. . . . . .
. . . .
. . . . . .
. . . .
. . . .
. . . .
. . . . 66 . . . . . 66 . . . . . 68 . . . . 72
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
79 83 85 89 90 92
. . . . . . . . . . . .
. . . .
. . . .
. . . .
. . . . .
. . . . .
. . . . .
. . . . .
96 96 97 98 99
54 . Convergence of Pad6 approximants . . . . . 4.1. Numerical examples . . . . . . . . . 4.2. Convergence of columns i n t h e Pad&table
4.3.
. . . .
63
Convergence of the diagonal elements Convergence of Pad4 approxirnants for Stieltjes series
55 . Continuity of the Pad6 operator . . . . . . . . . . . . . . . . . 100 $ 6. Multivariate Pad6 approximants . . . . . . . . . . . . . . . . . 104 6.1. Definition of multivariate Pad6 approximants . . . . . . . . 104 6.2. Block structure of the multivariate Pad6 table . . . . . . . . 109 6.3. The multivariate €-algorithm . . . . . . . . . . . . . . . 114
6.4.
The multivariate qd-algorithm
. . . . . . . . . . . . . . 115
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . .
118
Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . .
120
. . . . . . . . . . . . . . . . . . . . . . . . . .
122
References
62
‘‘Dam Ies applications d u calcul, il est ordinairement inutile, s i n o n imposaible, d’obtene’r le re‘sultat avec u n e cornp1;te exactitude. L e plus souvent, d1 s u f i t seulement d’en connaitre u n e valeur approche‘e telle que l’erreur commise e n adoptant cette valeur a u lieu d u re‘sultat exact soit infe‘rdeure A u n e limite donne‘e 6 priori. ”
H . PADh - “Sur l a re‘pre‘sentation approche‘e d ’une f o n c t i o n par des fractions rationelles ” (1892).
11.1. Notations and definition8
63
f 1. Notations and deflnitions. Consider a formal power series f(2)
= co
+ c1z + c 2 2 2 +
*. ,
(2.1.)
with cg # 0. In the sequel of the text we shall write d p for the exact degree of a polynomial p and w p for the order of a power series p (i.e. the degree of the first nonzero term). n) for f consists in finding The Pad6 approximation problem of order (m, polynomials
c m ...
p(2) =
UiZ’
i=O
and
i=O
such t ha t in the power series ( f q - p ) ( z ) the coefficients of z’ for + n disappear, i.e.
i = 0, . . .,m
(2.2.)
Condition (2.2.) is equivalent with the following two linear systems of equations
(2.3a .)
(2.3b.)
11.1.Nodotions and definitions
64
For n = 0 the system of equations (2.3b.) is empty. In this case ai = c; (a = 0 , . . ., n)and 6 0 = 1 satisfy (2.2.), in other words the partial sums of (2.1.) solve the Pad6 approximation problem of order (m,0). In general a solution for the coefficients a; is known after substitution of a solution for the b; in the left hand side of (2.3a.). So the crucial point is t o solve the homogeneous system of n equations (2.3b.) in the n + 1 unknowns b;. This
system has a t least one nontrivial solution because one of the unknowns can be chosen freely. The following relationship can be proved for different solutions of the same Pad6 approximation problem.
Theorem 2.1. If the polynomials p 1 , q1 and p2, 42 satisfy (2.2.), then p1 42 = p2 41.
Proof
The polynomial p1
42
- p q1 ~ can also be written as
Since
we have
pi 42 - p2 qi) 2 m
+n + 1
But (p1 42 - p241)(5) is a polynomial of degree at most rn + n. s o P1 (2'2 = P2 41.
A consequence of this theorem is t h a t the rational forms p1/q1 and
I
p 2 / q 2 are equivalent. Hence all nontrivial solutions of (2.2.) supply the same irreducible form. If p(z) and q ( z ) satisfy (2.2.) we shall denote by
the irreducible form of p/q normalized such t h a t qoC0) = 1.
111. Notation8 and definitions
65
This rational function rm,n(z)is called the Pad6 approximant of order (m, s) for f. By calculating the irreducible form, a polynomial may be cancelled in numerator and denominator of p / q . We shall therefore denote the exact degrees of PO and qo in rm,n respectively by m' and n’. As a conclusion we can formulalo t h c next theorem.
Theorem 2.2. For every nonnegative m and f exists.
92
a unique
Pad6 approximant of order ( m , n )for
66
11.2.1. Propertdea of the Pad6 approzimant
$2. Fundamental properties. 2.1.Properties of the Fade' approzirnant. Let rm,,, = po/qO be the Pad6 approximant of order (m,n) for 1. Although po and qo are computed from polynomials p and q that satisfy (2.2.), it is not necessary that p o and qo satisfy (2.2.) themselves. A simple example will illustrate this. Consider f ( z ) = 1 + z2 and take m = 1 = n. Condition (2.2.) is then equivalent with
{
bo = a0 b l = a1
{b, = 0 A solution is given by bo = 0 = a0 and b l = 1 = 01. So p ( z ) = z = q(z). Consequently P O = 1 = qo with w(f q O - P O ) = 2 < m + n + 1 and the corresponding equations (2.2.) do not hold. But it is easy to construct, from the knowledge of p o and qo, a solution of the system of equations (2.2.). Theorem 2.3.
If the Pad6 approximant of order (n,n) for f is given by
then there exists an integer 8 with 0 5 8 I min(m - m’,n - n'), such that p ( z ) = z" po(z) and q(z) = z " q o ( z ) satisfy (2.2.).
Proof Let p 1 , q1 be a nontrivial solution of (2.2.). Hence dPl 841 w(fq1
-PI)
5 IR. I 'Iz
L
m
+n+
1
Since the irreducible form of p1/q1 is po/qo we know that Pl(Z)
= t ( z )PO(Z)
Q l ( 5 )= t ( z )qo(2)
II.2.1.Properties of the Pad6 approzimant
67
with t ( z ) a polynomial of degree at most min(m - m', n - n'). If of the polynomial t ( z ) ,then
8
is the order
o 5 8 5 min(m - m', n - n') Since
4 P Q1 - PI 1 = w [ W qo - Po)]
= w [ z @ (Qo l -Po)] = 4 l ( z " 4 0 ) - (z" Po)] I
the proof is completed.
For another definition of Pad6 approximant we refer t o problem (2) at the end of this chapter. We have defined here the Pad6 approximation problem for a formal power series given around the origin. Of course it is possible to consider the same problem in any other point. Let
i=O
be a formal power series around thc point zo in the complex plane, with co
It is possible t o construct nontrivial polynomials
# 0.
and
i=O
such t h a t
(f 4 - P ) ( Z )
=
C
di(2-
~
0
)
~
i > m+n+l
This condition results in the same two systems of equations (2.3a.) and (2.3b.) and similar results hold.
68
11.2.2. Block structure of the Pad6 table
2.2. Block structure of the Pa& table. The Pad6 approximants rm,n for f can be ordered in a table for different values of rn and n:
ro,o
ro,?
r0,1
r0,3
...
This table is called the Pad6 table of f . As we already remarked, the first column consists of the partial sums of f. The first row contains the reciprocals of the partial sums of l/f (see also problem (4)). We now give part of the Pad6 table for exp(e) (table 2.1.) and for 1 + sin(z) (table 2.2.). In table 2.1. all the Pad6 approximants are different while in table 2.2. certain rm,ncoincide. In general the following result can be proved.
Theorem 2.4. Let the Pad6 approximsnt of order (m, n) for f be given by
Then (a)
w ( f qo - P O ) = m’
+ n + t + 1 with t 2 0
(b)
for k and k! satisfying m
(c)
m
5
m ’ + t and
fa
5
5k5
n‘+t
m’
+ t and n 5 l 5 n‘ + t :
11.2.2.Block structure of the Pad&table
09
Table 2.1. f ( z )= e z = 1
2
z3 z 4 + -+ + ... 2! 3! 4!
+ x+
1
-
.-
1-
I - X
-
1 1 1-2+-2
2
2
1
+1
-2
l + Z 2
l+z+
1+
$22
-3
2
1 + -z2 +
-2
1-
-2
3
1
I + -x
d. . .. 2 1 - -2 + -12 2 3
1 + -2 1
6
+ -12
-.
1 - - x1 + - 2 1 2
12
2
...
II.2.2. Block structure of the Pad& table
70
Table 2.2. f(z) = 1
1
x5 z7 + sin(x) = 1 + x - z3 ++ ... 5! 7! 3! -
1 1-x
-
1 1-2+x2
-__._
*
1 I - - S + S2 - - 5z 3 e
1 + -2 5
14-2
1-1-x
1 - -x+ 8
-2 6
1 + 2 + L 2
l + Z
1+x
.A. ..
1 + -x2 1
6
1
+2 - iz3
1 + 2 - ;z3
...
11.2.2. Block structure of the Pad6 table
Proof
To prove (a) we use theorem 2.3.:
This implies w(fq0
-Po)
2 m’ + n
+I
or
w(f qo - Po) = m
+n + 1 +t
with t 2 0 and t as large as possible. Let us now consider integers k and C such that
We put
Then clearly
and as a consequence of (a), U(f q - p ) = p n r + d + t + 8 + 1
Since
m‘+n’+t+a+l> k + L + l we have
71
11.2.8. Normality
72
because p , q is a solution of the Pad6 approximation problem of order (k,t ) . To prove (c) we again use theorem 2.3. It guarantees the existence of an integer 8 with 0 5 8 5 min(m - m‘, n - d) such t h a t m -192 + 1 Since
8
5 m - m’
5 w ( f q- p )
= m’
+d +t +8 + 1
we have VL+92+1 < m + n ’ + t + l
or n
< n’+t
An a1ogous Iy
rns m ’ + t
I
The previous property is called the block structure of the Pad6 table: the table consists of square blocks of size (t 1) containing equal Pad6 approximants.
+
2.3.Normality. As a result of theorem 2.4. we call a Pad6 approximant normal if it occurs only once in the Pad6 table. A criterion for the normality of an approximant is given in the next theorem.
Theorem 2.5. The Pad6 approximant
for f is normal if and only if
(a)
m = m‘ and n = n‘
(b)
~(Iqo-po)=m+n+~
11.2.5. Normality
73
Proof
If rm,* is riormal then (a) and (b) can be proved by contraposition. Suppose m’ < rn or n' < n. Then by theorem 2.4.(a):
~ ( qof - P O ) 2
m'
+ n’ + I
This implies
which is a contradiction with the normality of rm+(z). Suppose (b) is not valid. Using theorem 2.4. (a) we know that
4 Po - P O )
= m'
+ n’ + 1 + t
4 P qo - P O )
=m
+n + 1 +t
with t > 0. Since (a) must be valid:
with t Hence
> 0.
for all k and C satisfying m' 5 k 5 m! + t and n’ 5 C 5 n’ + t . This again contradicts the normality of rm,n(z). To prove t h a t (a) and (b) guarantee the normality of rm,n(z) we proceed as follows. Suppose rm,n(z)= rk,r(z) for certain k and C with k > m or C > n. For an integer B t h a t satisfies
w[(f
40
-Po)
57
2
k
+C+1
we find, by using (b), t h a t s>k-m This contradicts theorem 2.3.
or a > C - n
I
Normality of a Pad6 approximant can also be guaranteed by the nonvanishing of certain determinants.
74
II. 2.8. Normality
We introduce the notation Cm
... ...
Cm+n-1
...
Cm
Cm-1
Cm+l
Cm-n Cm+l-n
Cm
with det D,,o = 1. The following result c a n be proved [40 p. 2431. Theorem 2.6. The Pad4 approximant
For Stieltjes series this theorem and the following lemma [30 p. 605)lead to a remarkable result.
Lemma 2.1. Let p(t) be a real-valued, bounded, nondecreasing function defined on ( 0 , ~ ) and let the integrals
1
0)
ci =
t'dg(t)
exist for all e' 2 0. If g ( t ) has at least k points of increase then for all rn >_ 0 and for n = 0 , . . . , k we have
If g ( t ) has an infinite number of points of increase then for all m ,n 2 0
II. 2.9. N o r rnality
Clearly for Stieltjes series
with g ( t ) having infinitely many points of increase, the Iatter is true and hence we can conclude the following.
Theorem 2.7. Let f be a Stieltjes series and let g be a real-valued, bounded, nondecreasing function having infinitely many p0int.s of increase. Then for all m , n 2 0 the Pad6 approximant rrn,-, for
is normal.
75
II.3.1.Corresponding continued fractions
76
§3. Methods to compute Pad6 spproximents.
In the sequel of the text we suppose t h a t every Pad4 approximant in the Pad6 table itself satisfies the condition (2.2.). By theorem 2.3. this is the case if for instance min(m - m',n - n') = 0 for all m and n. A siirvey of algorithms for computing Pad6 approximants is given in [50] and P11.
3.1. Corresponding continued fractions.
The following theorem sball be used to compute the difference of neighbouring Pad4 approximants in the table.
Theorem 2.8.
If
and
P2
rm+k,n+C = -
42
with k, t
2 0 then a polynomial u(z) exists with
Proof
For the expression pi42 - p2q1 we can write
This completes the proof.
118.1. Corresponding continued fractions
77
Let us now consider the following sequence of elements on a descending staircase in the Pad4 table Tk
= {fk,ol
rk+l,ol rk+l,l) rk+2,i1..
.} for k 2 0
and the following continued fraction
Theorem 2.9. if every three consecutive elements in T k are different, then a continued fraction of the form (2.4.) exists with d k + i # 0 for i 2 1 and such t h a t the nth convergent equals the (n + l)thelement of Tk. Proof Put rk+;,j
=
Pi+ j
--
Q i+i
f o r i = j , j + l a n d j = 0 , 1 , 2 ,.... A continued fraction of which the nth convergent equals
is according t o theorem 1.4., given by
Here we have already used the fact t h a t QO = & I = 1. By theorem 2.8. we find that
78
II.3.1. Corresponding continued fractaone
i = 3 , 2 , ...
for certain nonzero numbers a; and b;. So the continued fraction (2.5a.) is
with
and
c
k+l P l =rk+l,O
=
CiXi
i=O
By performing an equivalence transformation we finally get a continued fraction of the form (2.4.). I In this way we are able to construct corresponding continued fractions for functions j analytic in the origin.
Theorem 2.10.
element of To ( n 2 O), then If the nth convergent of (2.4.) equals the ( n + (2.4.) is the corresponding continued fraction t o the power series (2.2.).
Proof Let Pn/Qn be the nth convergent of (2.4.). Then w ( j Qn - Pn) 2 n
because Pn/&, is also the ( n +
+ 1 and Qn(0) = 1
element of
To.
11.8.2. The qd-dgorithm
Hence
(f-2) (i)
(O)=O
79
j = 0 ,... , n
because Q , ( x ) is nontrivial in a neighbourhood of the origin. So the Taylor series development of the nth convergent matches the given power series u p t o and including the term of degree n. In other words, (2.4.) is the corresponding continued fraction t o (2.1.). I
By continued fractions of the form (2.4.) one can only compute Pad4 approximants below the main diagonal in the Pad6 table, For the right upper half of the table one can use the reciprocal covariance property of Pad6 approximants, given in problem (4) a t the end of this chapter. We now t u r n t o the problem of the calculation of the coefficients d k + i in (2.4.) for k > 1 starting from the coefficicnts c o , c l , c 2 , . . ., and not from the knowledge of the Pad6 approximants. 3.2. The qd-algorithm. Consider the following continued fraction
If the coefficients qPt1) and er")
are computed as in theorem 2.9. then the convergents of gk equal the elements of 1 6 . If we calculate the even part of g k ( X ) we get
80
11.3.2. The qd-algorithm
If we calculate the odd part of
g k - ~ ( z ) we
get
The even part of gk(z) and the odd part of g k - l ( Z ) are two continued fractions which have the same convergents r k , o , r k + l , l , r k + 2 , 2 , . . . and which also have the same form. Hence the partial numerators and denominators must be equal, and we obtain (431 for k 2 1 and e 2 1
(2.6a.) (2.Ab.)
The numbers q y ) and e p ) are usually arranged in a table, where the superscript (k) indicates a diagonal and the subscript t! indicates a column. This table is called the qd-table. Table 2.9.
11.3.2. The gd-algorithm
81
The formulas (2.6.) can now also bc memorized as follows: e p ' is calculated such that in the following rhombus the sum of the two elements on the upper diagonal equals the sum of the two elements on the lower diagonal
and qcl is computed such t h a t in the next rhombus the product of the two elements on the upper diagonal equals the product of the two elements on the lower diagonal
e
1"’
*
Since the qd-algorithm computes the coefficients in (2.4.), it can be used t o compute the Pad6 approximants below the main diagonal in the Pad6 table. To calculate t h e Pad4 approximarits in the right upper half of the table, the qd-algorithm itself can be extended above the diagonal and the following results can be proved [32 pp. 615-6171.
82
II.3.2. The qd-algorithm
TGble 2.4.
...
... ... ...
1
-(z) = wo
f
and for k
2
+ w15 + 20222 + . . .
1
If the elements in the extended qd-table are all calculated by the use of (2.6.) using the above starting values, then the continued fraction
supplies the Pad6 approximants on the staircase
II.3.3. The algorithm of Gragg
83
To illustrate the qd-scheme we will again calculate some Pad6 approximants for the function exp(z). Table 2.3. looks like 0
1
2
0 1 3
_ -1 6
-
0
_ -1
12
1
-
0
4
1 5
_ -1 2ti
1
6
3 20
1 _-1.0
...
-.
0
...
From this we get
It is obvious t h a t difficulties can arise if the division in (2.6b.) cannot be performed by the fact t h a t e p ) = 0. This is the case if the Pad6 table is not normal for consecutive elements i n T k can then be equal. Reformulations of the qd-algorithm in this case are giver) in [I31 and [30]. 9.9. The algorithm of Gragg.
Let us now consider ascending staircases in the Pad6 table. Take
and consider the continued fractioii
118.8. The algorithm of Gragg
84
In an analogous way as for T k one can compute the coefficients f yl and *I_"’ such t h a t the athconvergent of this continued fraction equals the (n l)*h element of s k . Remark that fk(s) is not an infinite expression since s k is a finite sequence. If we compute the odd part of f k + l and the even part of fk, we again get continued fractions of the same form t h a t have the same convrrgents. This reasoning provides us with formulas for and "I [29]: for k 2 1 $4) = 0
+
where
1 -(z) = wo
f
and for k
+ w1z + w222 + . . .
2 1 and 1 5 !t 5 k - 1
These quantities are arranged in a table as follows. The superscript denotes now an upward sloping diagonal.
Table 2.5.
II. 9.4 Determinant for mu1QB I
For the computation of gt( k ) and et(k)..
8p)
and
+
85
fy' wc have similar rhombus rules as for the f(kj
t
and
3.4. Det e rminant formulas . One cau also solve the system of equations (2.3b.) and thus get explicit formulas for the Pad6 approximant. For
we write
II.3.4. Determinant formulas
86
Theorem 2.11.
If the Pad6 approximant of order ( m , n )for f is given by
and if D = det Dm,,, # 0, then
1
Po(%) = 5
and
Proof
L
Since D # 0 the homogeneous system (2.3b.) has a unique solution for the choice 60 = 1. Thus the following homogeneous system has a nontrivial solution:
c,+lbo
+ sbi + z2b2 + . . * +x"bn = 0 + c w b l + * . .+ cm+l-,,bn = 0
c,+,bo
+
(1 - qo(x))bo
Cm+n-Ibl
+ . . . + cmbn
=O
This implies that the determinant of the coefficient matrix of this system is zero:
=o
II.3.4. Determinant for mu1a8
87
and it proves the formula for qo(z). If we take a look at f ( z ) q o ( x ) we have
Because the polynomial po(z) contains all the terms of degree less than or equal t o rn of the series f(z)qo(z), we got the determinant expression for po(z) given above. I The determinant formula for qo(z) can also very easily be obtained by solving (2.3b.) using Cramer's rule after choosing bo = Dm+. The determinant expressions are of course only useful for srnall values of m and n because the calculation of a determinant involves a lot of additions and multiplications. They merely exhibit closed form formulas for the solution. From the proof of theorem 2.11. we can also deduce t h a t
(f QO
-Po)(.)
=
;1
cm+l
where 00
p k ( z )=
ci
zi = f(z) - F k ( z )
i=k+ 1
This gives a n explicit formula for the error ( f - r m,= ) ( z ) in terms of the coefficients c i in f(z). For Stieltjes series
with convergence radius
$, it is possible using the error formula, to indicate
88
II. 8.4. Determinant f~ rrnula8
within a finite set of Pad6 approximants which one is the most accurate on (-03, +[ . The most interesting result is the following one.
Theorem 2.12.
Let f be a Stieltjes series. For the Pad6 approximants in the set
and for those in
{rm,n I O
we have
i m + n I 2k + I}
This means that when k increases, the best PadC approximants for a Stieltjes series among the elements in successive triangles r0,o r1,o
ro,1
...
rO,k
are the elements on the descending staircase To,in other words they are the successive convergents of the corresponding continued fraction for the Stieltjes series. For other subsets of the PadC table similar results exist because, when f is a Stieltjes series, the errors ( f - r m , n ) ( z ) are linked by inequalities throughout the entire table. The interested reader is referred to [8].
11.3.5. The method of Viscovatov
89
3.5. The method of Viscovatov.
By the method of Viscovatov described in section 3.3. of chapter I, the recursive generation of staircase sequences Tk of Pad6 approximants is absolutely straightforward in the case of a normal Pad4 table. We have proved in secttion 3.4. of chapter I that the constructed continued fraction is corresponding. Hence it generates the elements on To. If the rnrthod of Viscovatov is applied to k i=O
for the construction of a corresponding continued fraction
then the elements on Tk are obt,airied from
However in order t o obtain the normaliz,cd
the normalization qo(0)= 1 has t o be built into the algorithm via a n equivalence transformation. We reformulate it as follows. For
i=O
we put
and for j
22
90
11.9.6. Recursive algorithms
Then
3.6. Recursive dgorithms.
It is also possible t o calculate Pad4 approximants on ascending staircases by means of a recursive computation scheme. To this end we formulate the following recurrence relations for which we again assume normality of the Pad6 table. First we introduce the notation m
i= 0
Theorem 2.13. If the Pad6 approximants
are normal, then
11.3.6.Recursive algorithm
91
Furthermore
because of the assumption of normality and because of the uniqueness of the Pad6 approximant. I This theorem enables us to calculate ~ 3 1 9 3when p1 /qI and p2 192 are given. We denote this by rm,n
--*
Theorem 2.14. If
then
Computationally this means rm-1,n
t Combining theorem 2.13. and 2.14. we can compute the elements on an ascending staircase in the Pad6 table, starting with the first column. Other algorithms exist for the computation of Pad6 approximants in a row, column or diagonal of the Pad6 table, instead of on staircases. We do not mention them here, but we refer t o [l], [36] and [41].
113.7. The r-algorithm
92
3.7. The t-algorithm. Consider again a continued fraction M
I
with convergents
Using theorem 1.2. of the previous chapter, we know t h a t
For the continued fraction gmwn(x) given by (2.5b.) we get for k = 2n
and for gm-,-l(z)
with k = 'Ln + 1 we have
1 -+ rm,n+l
- rm>n
__ _ _1 _ ~-_ r m , n - rm-1,n
QXn
P2ni-2 Q2n
- P2n Q2n3-2
Consequently the elements in a normal Pad4 table satisfy the relationship
II.3.7. The e-algorithm
93
where we have defined
The identity (2.8.) is a star identity which relates
and is often written as
( N - c)-*+ ( S - c)-I = ( E - c)-1+ ( W -
c)-I
If we introduce the following new notation for our Pad4 approximants
we obtain a table of €-values where again the subscript indicates a column and the superscript indicates a diagonal.
Table 2.6.
...
(3)
€0
...
The ci"') are the partial sums F,(Z) of the Taylor series f(z).
94
II.3.7. The c-algorithm
Remark the fact that only even column-indices occur. The table can be completed with odd-numbered columns in the following way. We define elements 1 m = 0 , 1 , .. . (m-n-1) - (m-n) €gn+l - €2,-1 - ~ _ _ (2.9.) n = 0 , 1 , .. . (m-n) - €::-n-i)
+
€2n
with
Table 2.7.
II.3.7. The c-algorithm
95
from which we can easily conclude by induction on n that
or (2.10.) The relations (2.9.) and (2.10.) are a means t o calculate all the etements in table 2.7. and hence also t o calculate all the Pad6 approximants in table 2.6. This algorithm is very handy when one needs the value of a Pad6 approximant for a given x and one does not want t o compute the coefficients of the Pad6 approximant explicitly. The c-algorithm was introduced in 1956 by Wynn [51]. To illustrate the procedure we calculate part of the completed €-table for f(z)= ez with z = 1. Compare the obtained values with e = 2.718281828.. .
Table 2.8.
I(z) = exp(4 x=1
1. 00 0 0 0 0 1.oooooo
3.000000 2.750000
2. 500000 6. 00 0 0 0 0
2.722222
2.666667 24. 0 0 0 0 0
- 2.000000 - 30.00000 - 264.0000
2.718750
2. 708333
2 .5 0 0 0 0 0
1.000000
2. 00 0 0 0 0
3 .0 0 0 0 0 0
2.000000
00
2. 000000
1.500000
1.000000
1. 00 0 0 0 0
0
0
0
0
2.727273
2 .6 6 6 6 6 7 1 0 .0 0 0 0 0
2 .7 1 8 7 5 0
2 .7 1 4 2 8 6 2 4 3 .0 0 0 0
2 .7 1 8 3 1 0
2 .7 1 7 0 4 0 30 1 2 .0 0 0 2 .7 1 8 2 5 4
-2280.000
120 . 000 0 2.718333
2. 716667 720.0000 2. 7 18056
Again computational difficulties can occur when the Pad6 table is not normal. Reformulations of the €-algorithm in this case can be found in [15] and [52].
I I . 4 . f . Numerical ezamples
96
§4. Convergence of Pad6 approximants.
Let us consider a sequence S = {ro, r l , r2, . . .} of elements from the Pad6 table for a given function f ( z ) . We want t o investigate the existence of a function F ( s ) with ri(z) = F'(z)
lim
i-cc
and t h e properties of t h a t function F ( z ) . In general the convergence of S will depend on the properties of f . Before stating some general convergence results we give the following numerical examples. One can already remark tha t t he poles of the elements in S will play an important role. A lot of information on the convergence of Pad6 approximants can also be found in [4].
4.1. Numerical ezamples. For f(x) = ez and r i ( z ) = rm,,.,(z)with m + n = i, we know [40 p. 2461 that lim
i-00
for all
ri(z) = ez
2
in C
We illustrate this with t he following numerical results. Table 2.9.
f(z) = e" z = l e = 2.718281828.. .
_-
2 3 4
1.000000 2.000000 2.500000 2.666667 2.708333
1
2
3
4
00
2.000000 2.660667 2,714286 2.727949 2.718254
3.000000 2.727273 2.7 18750 2.7 18310 2.718284
2.666867 2.716981 2.718232 2.718280 2.718282
3.000000 2.750000 2.722222 2.718750
Next we consider t he case t h a t f is a rational function. For 2+
f(z) = 21-
10
ZZ.4.2. Convergence of columna in the Pad6 table
97
the Taylor series expansion 10 + converges for 1x1
< 1. If
+ l o x 2 + z3 + 1oe4 + . . .
r;(z) = r i , l ( z )then i-1
ri(z) =
C
+
ckzk
k=O
CiZ' Ci+I
1--2
For i even the pole of r;(z) is 10 and for i odd the pole of r i ( z ) is &,. In these points the sequence r ; ( e )docs not converge t o f(z).
For
In( 1 + z) = 1 -- z + - - z3 + z4 ... Z 2 3 4 5 the Taylor series expansion converges for IzI < 1 while the diagonal Pad6 a g proximants r;,i(z) converge t o f for all z in C\(-oo, -I]. The following results illustrate this.
f(z) =
~
Table 2.10.
f(2)
=
In(] --
+ z) 2
f(1) = 0.69314718..
f ( 2 ) = 0.54930614..
i 0 1 2 3 4
i d 1) ____r~.
.
1.000000 0.700000 0.693333 0.693152 0.693147
r ;, i ( 2) 1.000000 0.571429 0.550725 0.549403 0.549313
4.2. Convergence of columna in the Pad6 table. First we take ti(.) = r i , o ( z ) ,the partial sums of the Taylor series expansion for f(z). The following result is obvious.
98
Theorem 2.15.
II.4.9. Convergence of the diagonal elements
I
with ! t > 0, then = {ri,l)}iEN If f is analytic in B(0,r) = {Z 121 < r) (I converges uniformly t o f on every closed and bounded subset of B(0, r). Next take rj(z) = ri,l(z)Jthe Pad4 approximants of order (i, 1) for f . It is possible to construct functions f t h a t are analytic in the whole complex plane but for which the poles of the r;,l are a dense subset of C [40 p. 1581. So in general S will not converge. But the following theorem can be proved [6].
Theorem 2.16. If f is analytic in B(0,r), Lhen a subsequence of {r,,l}aN exists which converges uniformly t o f on every closed and bounded subset of R ( 0 , r).
In [3] a similar result was proved for S = (r;,2(z)}iEm. For meromorphic functions f it is also possible t o prove the convergence of certain columns in the Pad4 table [22]. Theorem 2.17.
If f is analytic in B(0, r) except in the poles w1 . .
wk of f with total multiplicity converges uniformly to f on every closed and bounded subset of B ( O , r ) \ { ~ l , . . - , w k } .
n, then {r+};€N
4.9. Convergence of the diagonal elements.
In some cases a certain kind of convergence can be proved. It is called convergence in measure [39]. Theorem 2.18. Let f be meromorphic and G a closed and bounded subset of C. For every S in R$? there exists an integer k such t h a t for i > k we have lri,i(z) - f(x)I
for all
2
in G;
where Gi is a subset of G such t h a t the measure of G\G; is less than 6
L
and
II.d.4. Convergence of Pad4 approzimants for Stieltjes series
9!?
Generalizations of this result can tw found in [42].
4.4. Convergence of Pad6 approximants
for Stieltjes aeries.
Stieltjes series were introduced in chapter I. For such series
i=0
with convergence radius
$
and di =
t'dg(t)
where g ( t ) is a real-valued, bounded, nondecreasing function taking on infinitely many different values, one can prove t h a t the poles of the Pad6 approximants r;+k,i with k 2 -1 are simple and real positive. One obtains convergence of {r;,;}iEm t o an analytic function. T h e convergence is uniform on every closed and bounded subset of the cut corriplrx plane a:\[$, 00). Similar resuIts hold for { r i + k , i } i E N [8].
100
11.5. Continuity of the Pad6 operator
85. Continuity of the Pad6 operator. If rm,n is the Pad6 approximant of order ( m ,n) for j, then we call the operator Pm,nthat associates with f its ( m , n ) Pad6 approximant, the Pad6 operator. Here m and n are fixed. So we can adopt the notations
When we compute rm,n i n finite precision arithmetic the computed result is not exactly the (m,n) Pad6 approximant, but it differs slightly from it by rounding errors and data perturbations. Since we can consider the computed result as t h e exact (rn,n) Pad4 approximant of a slightly perturbed input power series, it is important t o study the effect of such small perturbations on the operator P. To measure the small perturbations we introduce a pseudo-norm for formal power series: ((cI(= max
O
lei1
with c = (co,. . ., c,+~)
and the supremum norm for continuous functions on an interval [ a , b ] :
The Pad6 approximants
were normalized such t h a t qo(0) = I . This implies the existence of an interval [a, b] around the origin where qo(z) is strictly positive. For given f(z) = c;zi we call a neighbourhood U, of j,the set of power oc, diz' such that IIc - dl( 5 6. The following lemma is needed series g(z) = CiG0 t o prove the continuity of the operator P.
czo
Lemma 2.2. If P j is normal, then a neighbourhood the approximant Pg is normal.
rJ6
of f exists such t h a t for all g in
Ua
11.5.Continuity of the P a d 4 operator
101
Proof The lemma is an immediate consequence of theorem 2.6. by virtue of t h e fact t h a t t h e determinants mentioned there are continuous functions of I = ( ' 0 , ' ' ' 1 ',+fZ)'
Theorem 2.18.
If rmTn= p o f q o is normal and 40 is strictly positive in [a, b ] , then there exist constants K and 6 (only depending on c = (co, . . ., c,+,) and [a, b ] ) such t h a t for every g(z) = d;z' with IIc - dJI 5 6 :
xzo
Proof The fact t h a t Pf is normal implies the existence of a neighbourhood U, of f such t h a t for all g in U, t h e approximant Pg is normal. Hence, by theorem 2.11., a determinant formula for Pg cau bc given.
For
g(2) =
c
i= 0
d;z'
we write
where
Since a; and b; are continously differentiable functions of c, a constant M exists such t h a t
11.5. Continuity of the Pad4 operator
102
and
for every power series g in U,. Write
We know that q o ( z ) is strictly positive in [a,b] and that 8o(Z) is a continuous function of d . Thus it is possible to construct a neighbourhood Ua of f with 6 5 c and a constant N such that
11-111 5
N
for all power series g in V6
80
Now it is possible t o bound
for all g in
I
u6.
So we have seen t hat in case of a normal approximant r,,,, the Pad6 operator is continuous. Let's take a look a t an example where the normality condition is not satisfied. Consider f(2) =
and take m = 1 = n. Then
and
1
+
OLZ
+ 52
11.5. Continuity of the Pad6 operator
r1,1(x)
=1
if
(Y
103
=0
If we write r,(z) =
a
+ (2- 1). Q-X
and r(x) = 1
then clearly for every x lirn r,(x) = r(z) U-+O
but
for every interval [a, b] around tlic origin. However, a weakening of the normality condition is possible, in order t o obtain a necessary and sufficient condition for the continuity of P. Theorem 2.20. The Pad4 operator P is continuous in f , if and only if min(m - m', n - n ' ) = 0 where m' and n’ are the exact degrees of numerator and denominator of Pf respectively.
The proof can be found in [48].
104
II. 6.1. Definition of multivariate Pa dC approzim ants
§S. Multivariate Pad6 approximants. 6.1, D ef ; nit i o n of mult iv ar i at e Pad k upp r o s i n a rat 8 .
We have seen in t he previous sections t h a t univariate Pad6 approximants can be obtained in several equivalent ways: one can solve t h e system of defining equations explicitly and thus obtain a determinant expression, one can set up a recursive scheme such as the t-algorithm or one can construct a continued fraction whose convergents lie on a descending staircase in the Pad6 table. In the past few years all these approaches have been generalized to the multivariate case [12, 33, 34, 35, 37, 38, 451 but mostly the equivalence between the different techniques was lost. However, for the following definition a lot of properties of the univariate Pad6 approximant remain valid, also the recursive computation and the continued fraction representation. We restrict ourselves t o t he case of two variables because the generalization for functions of more variables is straightforward. Consider the bivariate function f(s,y) with Taylor series dcvcllopment I(X,Y)
=
c
.
c ; j 2’
y'
.
i,j=O
around the origin. We know t ha t a solution of the univariate Pad6 approximation problem (2.2.) for
i=O
is given by
and
116.1.Definition of multivariate Pad& approzimantlr
1
X
...
2"
Cm+1
Cm
...
Cm+l-n
...
cm
Cm+n
Cm+n-1
105
+ +
Let us now multiply the jthrow in p(z) and q ( z ) by zmS-J-' ( j = 2 , . . .,n 1) and afterwards divide the j t h column in p ( z ) and q ( z ) by xi-' ( j = 2 , . . ., n. 1). This results in a multiplication of nunierator and denominator by zmn. IIaving done so, we get
c m
c
.
m-1
CiX'
i=O
i=O
c,
cm+15m+1
c,+,
xm+n
c,+,-1
1 c,+1
xm+l
Cm+n
if D = detD,,,
,
CiX'
m+n
c,
...
5"
...
xrn+n--l
.. .
1
...
Xrn
..*
cm+n-1 xm+n-l
# 0.
...
c
.
m-n i=O
Cm+l-n
C,
CiX' xm+l-n
Xm
1 Cm+l--n
xm+l-n
CmZm
This quotient of determinants can also immediately be written down for a bivariate function f(z, y). The s u m c i z i shall be replaced by the kth partial sum of the Taylor series dcvelopmenl of f(z,y) and the expression C k zk by an expression t h a t contains all the terms of degree k in f ( x , y ) . Here a bivariate term c,, zi y j is said to be of degree i + j.
cf=,
II. 6.1 Definition of multivariate Pad4 opprozimants
106
I
If we define m- 1
G
i+i=O
CijXiYJ
i+ j = O
c
c
CijXiYJ
i+j=m+n
eijxiyj
i+j=rn+n-
c
m--n
...
CijXiYj
i+j=O
c
...
CijXiYJ
i+i=m
1
(2.11.)
and
c
1
c
CijX'YJ
c
CijXiYj
i+j = m
i+j=m+l
i+ j = m + n
1
CijXiYJ
c
CijX'yJ
i+j=m+n-1
... ...
c
1 CijXiYJ
i+j=m+l-n
...
c
CijXiYJ
i+ j = m
(2.12.)
then it is easy to see that p ( x , y ) and q ( x , y ) are of the form
i+j=mn
(2.13.)
mn+n
Therefore we are interested in the following multivariate Pad6 approximation problem. Find bivariate polynomials p and q such that w p 2 mn
wq2mn
d p 5 mn + m aqLmn+n
w [ ( f s- p ) ( z ,Y ) ] 2 mn
+ m. + n + 1
(2.14.)
In comparison with the univariate problem (2.2.) the degrees of p and q and the order of p, q and f q - p have been shifted over mn.
11.6.1.Definition of multivariate Pad6 approzimcrnts
107
Theorem 2.21.
If p ( z , u) and q ( z ,y) are given by (2.1 I .) and (2.12.) then (fq - P N Z l Y> =
c
dij
yi
i-t j>mn+rn+n+l
Proof
If we put
then the conditions
(2.15a.)
where Ck = 0 if k
< 0. With this notation we have
c m
P(Z,
and
Y) =
k-0
Adz, I/)
108
11.6.1. Definition of multivariate Pad6 lapprozinaants
If we solve the homogeneous system for the & ( s , y), choosing Cm(W)
.
Gm+l-,(Z,V)
v ) .. .
CmIZ, v )
*.
B o ( z , y )= Cm+n-l(Z,
and using Cramer's rule, and if we substitute this solution in the system of equations defining the A ~ ( zy), , then we get precisely p(z, y) and q(z, V ) given by the determinant expressions (2.11 .) and (2.12.). m
So p(z,y) and q(z,y) given by (2.11) and (2.12.) constitute a solution of (2.14.).
Now the expressions (2.11.) and (2.12.) may be identically zero, but one can prove that problem (2.14.) always has at least one nontrivial solution [16]. For the definition of the ( m , n ) multivariate Pad6 approximant we first need the following property. Theorem 2.22.
If p1 , q1 and p2, 42 both satisfy the condition (2.14.) then
Proof
We proceed as in the univariate case. Write p1 92 - p2 q1 as
We know that
and consequently
11.8.2. Block structure of the multivariute Pad6 table
Now
3 ( ~ 1 q 2- P Z Q I )
5
2m.n
109
+m +n I
SO p i q 2 - p2q1 is identically zero.
The multivariate Pad6 approximant of order (m, n) for j ( z , y) is now defined as the irreducible form
of a rational function p ( z , y)/q(z, y) where p and q satisfy (2.14.). For these rm,n a lot of properties of univarjate Pad6 approximants remain valid.
6.2. Block structure of the multivariate P a d 6 table. From the previous section we easily conclude the following. Theorem 2.23.
For every rn and n a unique multivariate Pad6 approximant of order (m,n) for f(z, y) exists. Let us now first take a look at an example. Consider
f(z,y) = 1 +
=1
2
+ sin(zy)
0.1 - 1/ m
00
i= I
i- 0
(2i + I)!
+ 10Z -1- l 0 l z y + 1 0 0 o x y ~+ . . .
Take m = 1 = n. Then we have t o look for p(z,y) and q ( z , y ) of the form p(z,y) = a102 -tnoly
+ a20z2 + a l l z y + ao2v 2
110
II.6.2. Block atructure of the multiuariate Pad4 table
such that
According t o theorem 2.21. a solution is given by
p(sJ) =
I
+
1 10s l0lxy
= 102
A1
+ 1OOs2 - l0lsy
= 10s - l0lzy
1 1+1oz 1 0s l 0 l z y p(z,y) = 1ooosy2 l 0 l s y
1 l0lsy
1 102 q(s,y) = 1ooosy2 l 0 l s y
0 1 10s
1 1 10s
11.6.2. Block attucture of the multivariate Pad6 table
We obtain
111
+ 10z2 - 20.2211 + 10y2 = 2 - 1.0ly - I0.lzy + lOy2 4- 2.012y2 f;
- 1.01y
__ _.
r112(2,y)
In general the following results can be proved about the order and degree of numerator and denominator of rm,n(zJy). Since po/qo is the irreducible form of a rational function p / q with p and of the form (2.13.), we may write
and because of ( 2 . 1 5 ~ )also with
So we can define mI = a p o - wqo n' = aqo
-
wqo
0bviously
This definition of mr and n' is an extension of the univariate definition, because in the univariate case wqo = 0.
11.6.2. Block structure of the multivariate Pad6 table
11 2
Theorem 2.24. If the Pad6 approximant of order ( m , n )for f ( z , y ) is given by
then an integer 8 with 0 bivariate polynomial
5 5 5 min(m - m',n - d > and
a homogeneous
exist such that p(z,y) = S(z,y)po(z,y) and q(z,y) = S(z,y)qo(z,y) satisfy (2.14.). Proof Since
is computed from a solution of (2.14.), we may consider nontrivial polynomials pl(z, y) and q l ( z , y) and write
with
p1
and 41 satisfying (2.14.) and with
c
tij zi yJ
+ wqo
2 mn
aT
T(z,y) =
i+j=wT
a bivariate polynomial.
Clearly wql = w T
and thus
w T = mn - wqo with
5
2 0.
+8
11.6.2. Block structure of the multivariate Pad4 table Also
wT
5I~T
with
which implies
O
585
min(wc - m',n - n’)
Write i+ j=mra--wgc+s
Because
the proof is completed. Theorem 2.25.
If the Pad4 approximant of order
(m, 92) for
f m , n ( Z 1 Y)
j(z,yf is given by
Po
= --(x, Y) Qo
and mr = 8 p o - wqo and nf = ayo - wqol then: (a) w ( f q o -PO) = wqo (b) for k and fk,L(XI
(c) m
5 rn’
+ mr + nr+ t + 1 with t 2 0
!. satisfying m’ 5 k <. m’
t 4= rm,n(z,Yl
+ t and n 5 n' + t
it and
n'
5 e 5 n' + t :
113
114
116.9. The multivariate c-algorithm
The proof is based on the univariate version and can be found in [19, 201. From this theorem it must be clear that normality can be defined exactly in the same way as in the univariate case. Necessary and sufficient conditions for normality of rn,n(x, y) are given in problem (9) at the end of this chapter. As a conclusion we take a closer look a t the meaning of the numbers wqg, m’ and n . In the sollition p ( z , y ) and y(s,y) the degrees have been shifted over mn. By taking the irreducible form of p / q part of t h a t shift can disappear, but what remains in p 0 , q g and fq0 - po is a shift over wqg. Now m’ and a play the same role as in the univariate case: they measure the exact degree of a polynomial by disregarding the shi€t over wyo 6.3. The multivariate 6-algorithm. Because of the theorems 2.21. and 2.22. and using the notation i+j=k
c p =
5
. .
c i j 5’ y’ =
a’+ j = O
c rn
Ck(5,Y)
k=O
which are the partial sums of the bivariate Taylor series f ( z ,g), we have (m-n) t2n
- +m,n(z, Y)
Ii.6.4. The multivariate qd-algorithm
115
if the €-table is calculated using the formulas (2.9.) and (2.10.). Remark the analogy with the univariste theory: in both cases the cLm) are the mth partial sums and the same algorithm is used. The reformulation of the E-algorithm in rase the multivariate Pad4 table is not normal is also inspired by the univariatc results [20].
6.4. The multivariate qd-algorithm. First wc rewrite the univariate qd-algorithm in a form such t h a t it can immediately be generalized. Consider 1 tit. function f(z) =
c
c ; zi
i=O
and its Pad4 approximant of order ( m ,n).If rm+(z) is the (2n)'h convergent of the continued fraction
then we can also say t h a t rm,n(z) is the ( 2 r ~ ) *convergent ~ of the continued fraction
II.S.4. The multivariate qd-algorithm
116
with
We have simply included the factor z in Q f ' and E r ) . This last continued fraction can be generalized for a bivariate function in the same way as the formulas (2.11.) and (2.12.) were obtained: replace the expressioq c k z k by an expression t h a t contains all the terms of degree k in the bivariate series M
We define, for k
2
1 and t? 2 1
Y) =
c c
i+j=k+l
c;j
zi y j
____
c i j t iy i
i+j=k
of the continued In [18] one proves t h a t now rm,n(z,y) is the ( 2 7 ~ convergent )~~ fraction
C
m- n
i+ j=O
c i j z i y i+
c
Cij
i+j=
m--n+1
z'yjl
&Im-n+l)
- _ _ _ _(z, _y)l
1
We give a simple example. Consider
l
1
-1
Ei
m--n+l)(
i
z7Y -
...
(2.16.)
II.6.4. The multivariate qd-algorithm
=1
1 + z + y + -(2 + x y + y 2 ) + .. 2
Take m = 2 and n = 1. The Pad4 approximant r2,1(z,y) is given by
with
+ + UP) - i ( 2+ z2y + zy2 + y3)
= f(2 sy
Indeed
r 2 , ~ ( 2 g) ,
is the second convergent of the continued fraction (2.16.):
117
118
II. Problems
Problems. a) Show t h a t the normalization qo(0)= 1 is always possible for
b) Show t h a t p o ( 0 ) = co if qo(0) = 1. Let f(z) be analytic. The rational function
is the Pad4 approximant of order ( m ,n) for f if and only if
f(kj(0)= rt!n(0) for k = 0 , . . .,rn + n with j the largest possible integer ( j E
+j
Z).
Give an explicit formula for r l , l ( z ) using the Taylor coefficients of .f(z). Let be the Pad6 approximant of order (rn,n)for f. If f(0) # 0 then 740
7PO
~
(2)
is the Pad6 approximant of order ( m , n )for $ where 7 = i ( 0 ) . This property is called the reciprocal covariance of Pad6 approximants. a) Let rm,m(z)be the Pad6 approximant of order ( m , m )for f If c f(0)+ d # 0 then a +_ b rm,m(z) ~ . _ _ _ c + d rm,m(z) is the Pad4 approxirnant of order ( m , m ) for a
c
+ bf + df
This property is called the homograflc covariance of Pad6 approximants.
119
II. Problems b) Is this property also valid for off-diagonal approximants?
(6)
Prove theorem 2.6.
(7)
Prove theorem 2.14.
(8)
Let r = , = ( z ) be the Pad6 approximant of order (na,m) for f . Then r-..-(
<%)
is the Pad4 approximant of order (m, rn) for
(9)
Prove the following conditions for normality of a multivariate Pad4 approximant: rm,n(51 Y)
PO = -(XI 40
Y)
is normal if and only if
(10)
Formulate and prove t h e reciprocal and homografic covariance of multivariate Pad6 approximants.
11. Remarks
120
Remarks. Condition (2.2.) is a linear condition and the polynomials p ( z ) and q(x) satisfying (2.2.) are not necessarily irreducible. Sometimes the Pad6 approximant of order (m, n) for f is defined in a non-linear way: let p f q be an irreducible rational function with numerator and denominator respectively of degree a t most m and n, then p/q is called t h e Pad6 approximant for f if [2] of order (m,n) P w(l--)>m+Pt+l Q
If this definition is used it is obvious t h a t the Pad4 approximation problem does not always have a solution. asNa basis for the polynomials. We have here used the functions { X ~ } ; ~ Instead one could also use a set of orthogonal polynomials. In this way one can e.g. study the Legendre-Pad6 approximation problem [26, 271 or thc Chebyshev-Pad6 approximation problem [14, 25, 281. More information about orthogonal polynomials and their link with the Pad4 approximation problem can be found in [lo]: almost all the recursive methods for computing sequences of Pad6 approximants can be derived from the theory of formal orthogonal polynomials. A general subroutine for the computation of Pad6 approximants is given in [lo] for the normal case and has been developed by Draux and van Ingelandt in the nonnormal case. Pad6 approximants have also been defined for matrix-valued functions [5, 461, for the operator exponential [24], for formal power series in a parameter with non-commuting elements of a certain algebra as coefficients [7, 231, for vector valued functions [21] and so on. Although Henri Pad6 was not the creator of the so-called Pad6 approximants, his name was given t o those approximants because of the extensive study of their properties in his thesis in 1892. Breeinski has edited a book containing the works of Pad6 and the story of his life [Q]. Pad6 approximants are in fact a special case of so-called Padb-type approximants [lo]. There a rational function p / q with numerator and denominator respectively of degree at most m and n is computed such that 4fq-p) 2 m+l
This supplies us with a linear system of m + 1 equations in rn
+ n+ 1
II. Remark8
+
121
unknowns. The remaining n 1 f r w parameters are used for a normalization and to insert some extra information about f if it is available, for instance some knowledge about singularities of f.
11. References
122
References.
[
11 Baker G.
Recursive calculation of Pad6 approximants. In [31], 83-92.
[
21 Haker G. 1975.
Essentials of Pad6 approxiruants. Academic Press, New York,
[
31 Baker G. and Graves-Morris P. Convergence of the Pad6 table. J. Math. Anal. Appl. 57, 1977, 323-339.
[
41 Baker G. and Graves-Morns 1’ Pad6 Approximants. Basic Theory. Encyclopedia of Mathematics and its applications: vol 13. AddisonWpsley, Reading, 1!481.
[
51 Hasu S. and Hose N . Two-dimensional matrix Pad4 approxirnants: existence, nonuaiqriencss and recursive computation. IEEE Trans. Automat. Control 25, 1980, 509-514.
[
T h e convergence of €’ad6 approximants. J. Math. Aual 61 Beardon A. hppl. 2 1 , 1968, 344-346.
[
Variational approach to t h e theory of operator 71 Bessis J . and Talrnan J . Pad6 approxirnants. Rocky Mountain J . Math. 4, 1974, 151-1 58.
[
Approximents cle Pad&ct applications. Cours de questions 81 Brezinski C. spkiales de Math I ([JCL), Uelgiunt, 1981.
[
91 Brezinski C. Henri Padd: (Euvrcs. Librairie Scient. e l ‘l’echn. Blanchard, Paris, 1984.
[ 101 Brezinski C.. Pad6 type approximation and general othogonal polynomials. ISNM SO, Rir khRuser Verlag, Basel, 1980 [ 11) HuJthed A .
Recursive algorithms for t h e Pad6 tat)le. two approaches. In [49], 211-230.
[ 121 Cbisholm .I. N-variable rational approximants I n [44], 23-42. [ 131 Clawsens G. and Wuytack L. On the computation of non-normal Pad6 approximants. J. Corn put. Appl. Math. 5, 1979, 283-289.
[ 141 GIenshaw C. arid Lord K.
liational approximations from Chebyshev
series. In [44], 95-113.
[ 151 Cordcllier F.
D6monstration alg6brique de i’extension de I’identit6 de Wynn aux tables de Pad6 non normales. In 1491, 36-60.
II. Refe re nc es
123
[ 161 Cuyt A.
Multivariate I’adh approximants. J. Math. Anal. Appl. 96, 1983, 283-293.
[ 171 Cuyt A.
The ~-algoril,tim and multivariate Pad6 approximants. Numer. Math. 40, 1982, 39-4fi.
[ 181 Cuyt A.
The qd-algorithm and multivariate Pad6 approximants. Numer. Math. 42, 1983, 250-269.
[ 191 Cuyt A .
Pad6 approxirnarits for operators: theory and applications. Lecture Notes in Mathematics 1065, Springer, Berlin, 1984.
[ 201 Cuyt A.
Singular rules for the calculation of non-normal multivariate Pad4 approximants. J. Con1 put Appl. Math. 14, 1986, 289-301
G e n e r a l i d (:-fractions and a multidimensional Pad6 [ 211 de Bruin M . table. Pti. D., Universily of hnistcrdam, 1974.
[ 221 de Montessus de Ballore 12
Sur les fractions continues algbbriqucs. Rend Circ. Mat. Palermo 19, 1905, 1-73.
[ 231 Draux A.
The Pad6 approxirnants in a non-commutative algebra and their applications. In [47], 117-131.
[ 24) Fair W. and L u k e Y .
t’atlh approximants t o the operator exponential. Numer. Math. 14, 1970, 379-382.
[ 251 Fike C.
Computer evaluation of Mathematical functions. Prentice Hall, New Jersey, 1968, 181-190.
1
Analytic coutinuatiou of scattering amplitudes and Pad6 261 FIeischer J . approximants. Nuclear Phys. 13 37, 1972, 59-76.
NonIincw I’ad@ approximants for Legendre series 271 Fleischer J. J . Math. Yhys. 14, 1973, 248-248.
[ 281 Gragg W.
Laurent-, Fourier- and Chebyshev-Pad6 tables, In [44], 61-
72.
[ 291 Gragg W.
The Pad4 table arid its relation t o certain algorithms of numerical analysis. SIAM Rev. 14, 1972, 1-62.
[ 301 Gragg W.
Matrix interpretations and applications of t h e continued fraction algorithm. Rocky Mountain J. Math. 4, 1974, 213-225.
II. Reference8
124
[ 311 Graves-Morris P.
Pad6 approxirnants and their applications. Academic
Press, London, 1973.
[ 321 IIenrici P.
Applied and computational complex analysis: vol. 1 &2. ,John Wiley, New York, 1976.
[ 331 Hughes Jones R.
General rational approximants in Approx. Theory 16, 1976, 201-233.
71
variabjes. J .
1
Rational approximation by a n interpolation 341 Karlsson 3. and Wallin H. procedure in several variables. In [44], 83-100.
I
General order Pad6 type rational approximants defined from 35) Leviu D. double power series. J . Inst. Math. Appl. 18, 1976, 1-8.
[ 361 Longman I.
Computation of the Pad6 table. Internat. J . Comput. Math. 3, 1971, 53-64.
Rational approximants t o holomorphic functions in n [ 371 Lutlerodt C. dimensions. J. Math. Anal. Appl. 53, 1976, 89-98.
A two-variable generalization of the Stieltjes-typc coutizlried fraction. -1. Comput. Appl. Math. 4, 1978, 181190.
[ 381 Murphy J . and O’Donnhoc M.
[ 391 Nuttall J.
The convergence of Pad6 approximants of mcromorphic fiiuctions. J . Math. Anal. Appl. 31, 1970, 147-153.
[ 401 Yerron 0.
Die Lehre von den Kettrribruchen 11. l’eubmr, Stiittgart,
1977.
[ 411 Pindor M.
A simpli6ed algorithm for calculating the Pad6 table derived from Baker and Longman schemes. J. Comp. Appl. Math. 2, 1976, 255-258.
[ 421 Pommerenke Ch.
Pad6 approxirnants and convergence in capacity. J . Math. Anal. Appl. 41, 1973, 775-780.
[ 431 Rutishauser N.
Der Quotienten-Differenzen Algorithmus. Mitteilungen Institut fur aogewandte Mathematik (ETH) 7, Birkhauser Verlag, Basel, 1957.
Pad6 and rational approximation : theory and applications. Academic Press, London, 1977.
[ 441 S a l f f . and Varga R.
11. Referenced
[ 451 Sciafe B.
125
Studies in Numerical Analysis. Academic Press, London,
1972.
[ 461 Starkand Y .
Explicit formulas for matrix-valued Pad6 approximants. J. Comput. Appl. Math. 5 , 1979, 63-66.
[ 471 Werner H. and Riinger H .
Pad6 approximation and its applications Lecture Notes in Mathematics 1071, Springer, Berlin, 1984.
[ 481 Werner II. and Wuytack L. O n the continuity of the Pad6 operator. SIAM J . Numer. Anal. 20, 1983, 1273-1280. [ 491 Wuytack L.
Pad6 approximation and its applications. Lecture Notes in Mathematics 765, Springer, Berlin, 1979.
[ 501 Wuytack L.
Commented bibliography on techniques for computing Pad6 approximants. In [49], 375-392.
[ 511 Wynn P.
On a device for computing the em(Sn) transformation. MTAC 10, 1956, 91-96,
[ 521 Wynn P. 175-195.
Singular rules for certain nonlinear algorithms. BIT 3, 1963,
127
CHAPTER mr Ratlonal Interpolants f l. Notations and definitions
.
. . . . . . . . . . . . . . . . . . . .
129
52 . Fundamental properties . . . . . . . . . . . . . . . . . . . . 131 2.1. Properties of t h e rational interpolant . . . . . . . . . . . . 131 2.2. T h e table of rational interpolants . . . . . . . . . . . . . 132 2.3. Normality . . . . . . . . . . . . . . . . . . . . . . . 135
. . . $ 3.Methods t o compute rational interpolants 3.1. Interpolating continued fractions . . . . 3.2. Inverse differences . . . . . . . . . . . . 3.3. Reciprocal differences . . . . . . . . . 3.4. A generalization of the qd-algorithm . . . 3.5. A generalization of the algorithm of Gragg 3.6. The generalized &-algorithm . . . . . . . 3.7. Stoer’s recursive method . . . . . . . . 54 . Rational Hermite interpolation
4.1. 4.2. 4.3. 4.4. 4.5.
. . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . . . . . . . . . . . .
. 138 . 138 139
. 143 . 146
. . . . . . . . 149 . . . . . . . . 151 . . . . . . . . 153
. . . . . . . . . . . . . . . . . 156 . . . . . . . . . 156
Definition of rational llermite interpolants The table of rational Ilermite interpolants Determinant representation . . . . . . . Continued fraction representation . . . . Thiele’s continued fraction expansion . .
. . . . . . . . . 159
. . . . . . . . . 162
. . . . . . . . . 163
. . . . . . . . . 164
85 . Convergence of rational Hermite interpolants . . . . . . . . . . . 187 $ 6. Multivariate rational interpolants . . . . . . . . . . . . . . . . 169 6.2. Interpolating branched continued fractions . . . . . . . . . 169 6.2. General order Newton-Pad6 approximants . . . . . . . . . 175
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . .
188
Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . .
190
. . . . . . . . . . . . . . . . . . . . . . . . . .
192
References
128
Die Lagrangesche Interpolationsformel, welche dazu dient, eine Rekhe v o n n W e r t h e n durch eine ganze F u n k t i o n { n - l l t e n Grades dnrzustellen, i s t v o n Cauchy durch eine Formel verallgemeinert morden, welche eine Rekhe v o n n f m W e r t h e n durch eine gebrochne Funktkon dnrstellt, deren Zahler u n d N e n n e r respektkve von (n-l)ten u n d mten Grade & z d . Mun k a n n die Formel dadurch deduciren, dass m a n die lineuren Gleichungen, v o n uielchen die Aujgabe ubhungt, uuflost, u n d die D e t e r m i n a n t e n , welche m a n fGr d e n Zahler u n d N e n n e r jindet, entwickelt. ”
C . JACOBI - “Uber die Darstellung ekner Rekhe yeyebner W e r t h e durch eine gebrochne rationale Funktion” (1845).
III.1. Notc~tionoand definitaorzs
129
$1. Notations and deflnitiona. Consider a function f defined on a subset G of the complex plane. Let { Z i } ; , - N be a sequence of different points belonging t o G. We still denote the exact degree of a polynomial p by a p and its order by w p . The rational interpolation n) for f cousists in finding polynomials problem of order ( m , m
and
n
with p ( z ) / q ( z )irreducible and such t h a t
e' = 0, . . m + n .)
(3.1.)
Instead of solving problem (3.1.) we consider the linear system of equations f ( z i ) q ( z i -) p ( z i )
=x
0
i=O
,...,m + n
(3.2.)
Condition (3.2.) is a homogeneous system of n-t n -+ 1 linear equatims in the + 2 unknown coefficients a; and 6; of p and q. Hence the system (3.2.) always has a t least one nontrivial solution. For different solutions of (3.2.) the following equivalence can be proved. m -t n
Theorem 3.1.
If the polynomials p i ,
q1 and p 2 , q 2 both satisfy (3.2.) then pIq2 = p z q l .
Proof For the polynomial p l q 2 - p 2 q l we can write (PI42
- P 2 q l ) f Z i ) = [(fq:!- p 2 h - ( f m - p1)q2](z;) = 0
i = 0 , . . .)rn + n
130
Since a ( p 1 ~ p2q1) m n zeros.
+
111.1.Notations and definition8
5 m +n
it must vanish identically for it has more than
I
Not all solutions of (3.2.) also satisfy (3.1.): it is very well possible t h a t the polynomials p and q satisfying (3.2.) are such t h a t p / q is reducible. Nevertheless, because of theorem 3.1., all solutions of (3.2.) have the same irreducible form. For p and q satisfying (3.2.) we shall denote by
the irreducible form of p/q, where qo(z) is normalized such that qO(x0) = 1, and we shall call tm,n(z)the rational interpolant of order ( m , n )for 1. The following result is a consequence of theorem 3.1.
Theorem 3.2. For every nonnegative m and n a unique rational interpolant of order (m, n) for j exists. Although the terminology "interpolant" is used i t may be t h a t r,,n(z) does not satisfy the interpolation conditions (3.1.) anymore [14]. A simple exampie will illustrate this. Let $0 = 0, z1 = 1, z2 = 2 and f(zo) = 0, f ( q )= 3, f(z2) = 3. Take m = n = 1. Then t h e system of interpolation conditions is
=0 3(bo 6 1 1 - (UO ~ 1 = ) 0 3(60 t 261) - (UO i-2 a l ) == O a0
+
+
A solution is p(s) = 32 and q ( z ) = 2. Thus p o ( z ) = 3 and qo(z) = 1 . Clearly
Note the similarity with the Pad6 approximation problem: the Pad6 approximant of order (m, n) did not necessarily satisfy condition (2.2.) anymore. We shall see t h a t many properties valid for Pad6 approximants can be generalized for rational interpolants [20].
III.2.I. Propertie8
of the rational interpolant
1.31
52. Fundamental properties. 2.1. Properties of the rational interpolant.
Let rm,n= p o / q o be the rational interpolant of order ( m , n ) for f . If po and yo do not satisfy the system of conditions (3.2.) themselves, it is easy t o construct polynomials p and y from po i L n d qo t ha t are a solution of (3.2.). Denote the exact degree of PO by m' and thc cxacl. degree of qo by n’.
Theorem 3.3.
If the rational interpolant of ordcr (rn,n) for f is
then a n integer a exists with 0 exist belonging t o { x u , . . ., + z},
<: H 5 min(m - m', rz - n’) and 8 points yl, . . .,y, s u c h t ha t
and
satisfy (3.2.).
Proof Let pl(z) and q ] ( z ) be a solution of ( 3 . 2 . ) . Then
(f q1)(2;) = p1 (z;)
i = 0 , . . ., m
+n
Hence if yl(zi) = 0 also p~(zi)= 0. Let {yl,. . .,ye} be the set of zeros of q1(z) belonging t o {Q,. .., z ~ + ~ We } . construct t(Z)
If
=
n;-,
(2 -?A)
132
III.2.2. The table of rational interpolanta
then a polynomial u exists with
Consequently 8
= at
5 min(m - m’,n - n’)
Since
we have
If we put p(z) = p o ( z ) t ( z ) and q ( z ) = qo(z) t ( z ) then p and q also satisfy (3.2.). I
As a conclusion we can say th at the rational interpolation problem (3.1.) has a solution if and only if p o ( z ) and qo(z) satisfy the system of equations (3.2.). 2.2. The table of rational interpolanta.
The rational interpolants of order (m,n) for f can again be ordered in a table: r0,o
r0,1
r0,2
ri ,o
71,l
r1,2
r2,o
r2,1
...
r3,0
r3,l
.. .
...
In the f i s t column one finds the polynomial interpolants for f and in the first row the inverses of the polynomial interpolants for j. By theorem 3.3. a t least
111.2.2. The tuble of rational interpolanta
m'
+ n' + t + 1 points ( 2 0 , . . .,
with t
~ ~ t + ~ t + t }
that
2 0 exist in ( 2 0 , .. ., z,+n} such
i = 0 , . . .,m'
rm,n(z;)= f(z;)
133
+ n’ + t
On the basis of this conclusion a property comparable with the block structure of the Pad4 table can be formulatcd. Theorem 3.4. Let
and let (a)
(f qo - p o ) ( z ; ) = 0
i = 0, , . ., m' 1- n’
(b)
(f qo - p ~ ) ( z i#) 0
i = m'
(c)
for k and t satisfying m'
(d)
m <m'+t andn
then
5
+t
+ n’ + t + 1, . . .,m +
5 k 5 m’ + t and nr 5 !? 5
n'
+ t:
n'+t
Proof For k and t satisfying m' R = min(k - m'> !?- n’) and
5k5
mr
+t
and n’
5 t 5 n’
-t t we define
For p ( z ) = p o ( z ) t ( z ) and q ( z ) = qo(z) t ( z ) we know from (a) that
Since k
+ ! 5 m' + n' + t +
8
we have
(f q - p ) ( z ; )= 0 i = 0, . . ., k + 1
134
111.2.2. The table of rational interpolants
and thus
because
Since p o / q o is the irreducible form of p / q , where p and q satisfy (jq-p)(z;)=O
i = O ,...,m+ n
it is possible, according t o theorem 3.3., t o find an integer 0 5 R 5 mdn(rn- m , n - n ) such t h a t
The upper bound on 8 implies the upper bounds m respectively. This finishes the proof.
8
with
+ t and n + t on m and n I
Remark that condition (b) in theorem 3.4. is only necessary t o assure t h a t the integer t in (a) is as large as possible. It is important t o emphasize t h a t the table of rational interpolants has the block structure described above, only when the interpolation points (20,.. ., + )z, are ordered in such a fashion t h a t the points xi in which (I4 - P ) ( Z i ) # 0 are the last points in the interpolation set. For chosen m and n such a numbering can always be arranged and then we know that all the rational interpolants lying in a square emanating rrom rrnf,=!and with rmt+t,n~+tas its furthermost corner are equal to p o / q o . Let us take a look at the following example. Let 2; = i for i = 0 , . . . , 5 with j ( z 0 ) = I, f ( x 1 ) = 1, f(x2) = 5 / 3 , j ( z 3 ) = 5/2, f(z4) = 17/5, f(z.5) = 13/5. Then
III.2,S.Normality
Po(4
r3,2(x) = -~~
4015)
-
135
1 + x2 1+ z
and (f qo - pa)(x;) = 0 for d = 0 , . . ., 4, while (f qo - po)(xs) # 0. This implies that r2,1 = r2,2 = r3,l = r3,2. If t h e interpolation points are not numbered such t h a t the conditions of theorem 8.4. are satisfied then it is possible for 2 dements in t h e table of rational interpolants t o be equal without having a square of equal elements. A disturbance in the numbering of the interpolation points in theorem 3.4. implies a disturbance in the block-structure. Let us illlistrate this. For 2; = i when i = 0 , . . . , 3 arid f ( x 0 ) = 2, f ( x 1 ) = 312, f ( ~ 2 )= 4/5, f(z3) = 112 we find
q , o ( z ) = r2,1(x) = 2 - 2 X
but r?,o(z)
=2 -
2 x - --x 1 5 10
2
and
One can also check t h a t the rational interpolation problem of order (k,!) for f as formulated in 13.1.) has a solut,iou if the integers k and f are such that
and if the conditions of theorem 3.4. are satisfied. 2.3. Normality. Again we call an entry of the table normal if it occurs only once in t h a t table. A necessary condition for the normality of the rational interpolant rm,n(x) is formulated in the following theorem.
138
KII.2.9. No rmaltty
Theorem 3.5.
If rm,,, = @
Qo
is normal and if (f qo - p~)(z;)= 0 for z' = 0 , . . ., m'
+ n’, then
m = m' and n = n’ ( j q o -PO)(%;)
f 0 for i = m + n + 1, m + n + 2
(IQo - po)(z;)
= o for i = 0,.. .,m l +
we know that rm,n= r m r , n i . Hence, if rm,n is normal, we have m = m' and n = n’. Now suppose that (f qo - p ~ ) ( z , + ~ ~ +=~ 0. ) Then
This contradicts the normality of r,,,+. If (I90 - ~ 0 ) ( ~ m + n + 2= ) 0 then
satisfy
(1q - p)(z;) = 0
for i = 0 , . , ., m
+n+2
Thus rm," = rm+l,n+l which is again a contradiction with the normality of rm,n-
I
111.2.9. Normality
137
Conclusion (b) in theorem 3.5. does not imply tha t ( f q o - p o ) ( z i ) f 0 for i
2m
+n + 1
That the conditions (a) and (b) are not sufficient to guarantee normality of r m , n ( z ) is illustrated as follows. I d z; = i for i = 0 , 1 , 2 , . . . and f(z0) = 0, f(z1) = I, f(z2) = 3, f(z3) = 4, f(zi) = i for i = 4 , 5 , 6 , .. .. For rn = 1 and n = 0 we find r,,+(z) = z with (a) and (t,) in theorem 3.5. satisfied. But rnL,n is not normal because r m , n = r k , t for k 2 3 and e 2 2. However, it is possible t o formulate a sufficient condition for the normality of rm.=.
Theorem 3.6.
If rm,n = YO
with m = m', n = n' and (fq0 - po)(z;) = 0 for a t most m + n + 1 points from the sequence {s;}~,]N, then T ~ is normal. , ~ Proof Let us suppose that rm,nis not normal, in other words that rm,n = r h , t for k >_ m, t? 2 n and with k t- !. > m + n. According t o theorem 3.3. an integer 8 exists with 0 5 8 5 min(k - m, 1 - ti) and a points {yl, . . . , y e } exist such that, the polynomials
and
satisfy
Hence ( f q o - p o ) ( Z i ) = 0 for a t least k + t? + 1 - 8 points in (z;)~&. Since 6 is bounded above by k - m and t? - n, we conclude th a t k + L + 1 - 8 > m + n 1 which contradicts the fact th at ( j q o - p o ) ( z ; ) = 0 for at most rn + n + 1 points from {zi);Em I
+
III.3.1. Interpolating continued fractions
138
$3. Methods to compute rational interpolants.
In the sequel of this chapter we suppose t h a t every rational interpolant r,,=(z) itself satisfies the interpolation conditions (3.1.). This is for instance satisfied if min(m - m', n - n’) = 0. 9.1. Interpolating continued fractions.
Theorem 3.7.
If
and rm+k,n+C
with k,C
2. 0 then
a polynomial
du
U(Z)
P2
=42
exists with
5 max(k - 1, t - I)
and
where
Proof As we assumed, the rational functions ?-m,n and rrn+k,=+( both satisfy [3.1.): for d = 0 , . . .,rn + n
(fql
-p l ) ( s i ) =0
(fq2
- p 2 ) ( ~ ;= ) 0 for i
= 0 , . . .,m
+ k + n + t?
Consequently (P1q2
- P 2 Q l ) ( Z i ) = KfQ2 - P 2 ) q l ] ( Z i ) - [ ( f q l
-p1)q2](2;)
=0
111.3.2. InvcrRp difference8
139
+
i = O , . . . , m n and thus a polynomial ~ ( z ) exists such that ) t~(z)B,+,+l(z). It is easy t o see t h a t a u 5 max(k - 1 , 1 - 1) (PlQZ' - p z q l ) ( ~ = since a ( p 1 q z - pzqi) 5 max(m + n + k , m n L). I
for
+ +
If we consider the staircase of rational interpolants ~ k = { ~ k , 0 , r k + 1 , 0 ~ r k + l , i ~ r k + 2 , i I for ~ ~ ~ }k
2 0
it is possible t o compute coefficients d;(i 2 0) such that the convcrgents of the continued fraction do + d l ( 2 - 2 0 ) dk(2 - 20). . .(z - X k - 1 )
+ + I
.
.
are precisely the subsequent elements of
Tk.
Theorem 3.8.
If every three consecutive elements in Tk are different, then a continued fraction of the form (3.3.)exists with d k + , # 0 for i 2 1 and such t h a t the nthconvergent equals the (n + 1lth element of Tk.
The proof is left t o the reader because it is completely analogous to the one given for theorem 2.9. We shall now describe methods t h a t can be used to calculate those coefficients d;(i 2 0).
3.2.Inverse d#e re nces . Inverse differences for a function f givcn in G are defined as follows:
po[z] = f(z) for every
for every
2
in G
.
Z O , Z ~ , . .,zk
in G
140
111.3.2. Inverse daflerencea
w e call p k [ Z 0 7 . . .,zk] the kth inverse difference of f in the points 20,. . ., Z k . Usually inverse differences depend on the numbering of the points 2 0 , . . .,Z k although they are independent of the order of the last two points. If we want t o calculate an interpolating continued fraction of the form
(3.4.)
we have to compute the inverse differences in table 3.1.
Table 3.1. Po I201
Theorem 3.9.
If di = p;[so,. . ,,xi] in the continued fraction (3.4.), then the Cn of (3.4.) satisfies
ltth
Cn(z;) = f(zi) for i = 0 , . . .,n if C , ( z i ) is defined. Proof
From the definition of inverse differences we know t h a t for n
2 1:
convergent
III.8.2. Inverse d i f e r e n c e s
141
With d; = cpi[xo,.. . l z;] it follows that C,, satisfies the imposed interpolation 1 conditions. The continued fraction po[zo]
2-20 2-z1 + ( c P l [ ~ O , X l+l-JP2[2UJ21,221I+ ... --
is called a Thiele interpolating continued frsction. To illustrate this technique we give the following exarnple. Consider the data: z i = i for i = 0, . . .,3, f ( ~=)1, f ( z l ) = 3, /(Q) = 2 and f ( x 3 ) == 4 . We get
3
1/2
4
1
4
3/10
The rational function
1
I
2 2-11 I+ + ___ + 2-21
I1/2
213
13/10
= -5. x 2 - 5 x - 6 42 - 6
= r(z)
142
III.3.2. Inueree differences
indeed satisfies r(zi) = f(zi)for 4 = 0 , . . ., 3. In the previous example difficulties occured neither for the computation of the inverse differences nor for t h e evaluation of r(z;). We shall illustrate the existence of such computational difficulties by means of some examples. Consider again the data: 2 0 = 0, z1 = I, z2 = 2 with f ( s 0 ) = 0, f(z1) = 3 = f ( ~ 2 ) . Then the table of inverse differences looks like 0 3
113
3
213
Hence
3
I
I
is not defined for x = 20, and thus we cannot guarantee t h e satisfaction of the interpolation condition r(x0) = f ( 2 0 ) . If we cousider the data: zi = i for i = 0 , . . . , 4 with ~ ( Z O= ) 1, f(z1) = 0, f ( z 2 ) = 2, f(23) = -2 and f ( q ) = 5 then p2[z0,21,23] is not finite. This does not imply the nonexistence of t he ratioual interpolant in question. A simple permutation of t h e interpolation d a t a enables us t o continue the computations. For zo = 0, z1 = 2, 22 = 1, z3 = 3 and z 4 = 4 we get
I
2
2
0
-1
113
-2
-1
-113
-3
5
z
-2
-917
The rational function
7/12
III. 8.9. Reciprocal differencea
143
satisfies r(zi) = f(z;) for d = 0,.. .,4. In order t o avoid this dependence upon the numbering of the data we will introduce reciprocal differences. 9.9. Reciprocal diferences.
Reciprocal differences for a function f given in G are defined as follows :
po[z]= f ( z ) for every z in G
w e call P k ( Z 0 , . . ., z k ] the kth reciprocal difference of the function f in the points 20,.. ., Z k . There is a close relationship between inverse and reciprocal differences as stated in the next property.
Theorem 3.10.
For k
2 2 and for all
. ., Z k in (7:
20,.
Proof
The relations above are an immediate consequence of the definitions. This theorem is helpful for the proof of the following important property. Theorem 3.11. pk(201..
.,Z k ] does not depend upon the numbering
of the points
X O , . . ., Z k .
I
144
III.3.9. Reciprocal diffe T ence 8
Proof We consider the continued fraction (3.4.)and calculate the kth convergent by means of the recurrence relations (1.3.):
i = 1, ..., k
For even k = Zj,this convergent is of the form
+ a l z + . . . + 0 j Z’ bo + 6 1 2 + . . . + b j z3
a0
____I____
and for odd k = 2 j - 1, it is of the form
+ . . . + a j zi+ biz + . . . + b3-1. zi-’
ao -t- a12
_ _ I _ -
bo
In both cases we calculate the coefficients of the terms of highest degree in numerator and denominator, using the recurrence relations for the kfh convergent and the previous theorem : for k even we get
and for k odd
Since Pk[%o,. . ., zk] appears to be a quotient of coeflicients in the rational interpolant, i t is independent of the ordering of the 20, .. ., z k because the rational interpolant itself is independent of th at ordering. I
111.3.3. Reciprocal differences
145
The interpolating continued fraction of the form (3.4.) can now also be calculated as follows: compute a table o f rcciprocal differences and put do = p o [ z o ] , d l = p 1 [zo, zl]and for i 2 2: d, =; p t [ z o , . . ., z z ]- p r - 2 [ z ~ , .. ., zi-21. U p t o now we have only constructcd rational interpolants lying on the descending staircase To. To calculate a ration:tl interpolant on T k with k > 0 one proceeds as follows. Obviously it is possiblr to construct a continued fraction of the form
(3.5.) whose convergtwts are the e l e m r n ~ sof ?i, Clearly CO,. . ., c k + l are the divrdtd difrerences f[zoj,. . f[z0, . , z k + l ] since rk,O and r k + l , O , the first two coriwrgents, are the polynomial interpolants for j of degree k and k + 1 respectivc3ly If we want t o calculate for instance r k + i t we need the (2!Jth convergent of ( 3 5.) In order to compute the coefkienl,s d k + * for = 2, . . .,2e we write .j
To define
8
we proceed as follows. Thc conditions rk+l.l(z;) = j ( z ; )
imply t h a t
So
A(Z)
8
for i = 0,.. ., k
+ 2t?
must satisfy
is the ( Z t -
convergcrit of the continued fraction
III.3.4.A generalization
146
of the qd-algorithm
Hence 8(z) belongs t o t h e descending staircase To in the table of rational interpolaats for the function -.q
f-P
As soon as t h e coefficients C O , . . ., c k + l are known, the function q / ( f - p) can be constructed and inverse or reciprocal diiferences for it can be computed. The coeficients dk+i with i 2 2 are preciscly those inverse differences. So finally the computation of an clement i n T k for f is reduced to the computation of an c l e m mt in 7’0 for 4_ _ I - P 3.4. A generalization of the qd-algorzthm.
Consider continued fractions of the form gk(2)
= co
+
c k
i= 1
c;(.
- 2 " ) ( 2 - el). . .(z
- Zi-1)
(3.6.)
Theorem 3.12.
If every three consecutive elements in Tk are different, then a continued fraction of the form (3.6.) exists with c k + J # 0, q?+') # 0, e y + ' ) # 0, 1 + q y + ' ) ( z o - z k + 2 ; - , ) # 0, 1 + e[,k+')(zo- ~ k + ~ # ; ) 0 for a 2 1 and such element of Tk. t h a t the nth convergent equals the ( n
+
IIZ.3.4.A generalization
of
the qd-algorithm
147
Proof
For t h e elements in Tk we put rk+i,j
=
Pk+i,j --__
qk+i,j
for i = j , j + l a n d j = 0 , 1 , 2 , . .
and for the convergents of gk(Z) we put, prl C, = -- for n = 0 , 1 , 2 , . . . with
&O
= &I = 1
Q R
lJsing theorem I .4. a continued fraction with nth convergent equal to
I:,
(2
r,
( n 2 0)
is, alter a n equivalence transforIr\ixLron,
(3.7.)
For
we find by means of theorem 3.7. (,hat f o r
. .(.
C k + 1 ( 5 - Z ).
+
with
Ck+1
1------r
# 0 since rk,o # r k + l , o .
-
1,
Zk)l
2
1,
’2-2
ai(z - z k + i ) /
148
III.S.4.A generalization of the qd-algorithm
For (3.7.) it is even true t h a t
we find t h a t (3.7.) can be written as (3.6.).
I
To calculate the coefficients qLF+') and in (3.6.) one can use the following recurrence relations. Compute the even part of the continued fraction g k ( 2 ) and the odd part of the continued fraction gk-l(z). These contractions have the same convergents r k , O , r k + l , 1 , r k + 2 , 2 , . . . and they also have the same form. In this way one can check [2] that: for k 2 I
III.3.5. A generalization of the algorithm of Gragy
149
These coefficients are usually ordered as in the next table
Table 3.2.
Again the superscript denotes a diagonal i n the table and the subscript a column. Another qd-like algorithm exists for continued fractions of another form than the one given in (3.6.). Although it is computationally more efficient, it has less interesting properties and so we do not mention it here but refer t o [3].
3.5. A generalization of the algorithm of Gragg. The previous algorithm generalized the qd-algorithm and calculated elements on descending staircases. We can also generalize the algorithm of Gragg and calculate rational interpolants on ascending staircases [3]. To this end we assume normality of the table of rational interpolants. Consider for k 2 1 t h e staircase
and continued fractions of the form fk(2) = c g
+
c k
i= 1
C;(.
- 2 0 ) . . .(z - . & I )
-
Ck(2
- 20).
. . ( z - Zk-1)j 1 (3.8.)
150
111.8.5.A generalization of the algorithm of Gragg
Similar to theorem 3.12. one can prove that there exist coefficients f?) # 0 and 8:b) # 0 such t ha t the successive convergents of f k are the elements of s k , as soon as three consecutive elements of s k are different from each other. Making use of the relations existing between neighbouring staircases sk and S k + l we get the following recurrence relations: for k 2 1
f (1k ) - Ck-1 ck
(3.9.)
The coefficients
fr)
and
can be arranged in a two-dimensional table. Table 3.3.
f‘,“’
...
f I”
f I“’
...
Each upward sloping diagonal contains the coefficients which are necessary t o construct the continued fraction (3.8.). It is easy t o see th a t the formulas (3.9.) reduce to the corresponding algorithm of Gragg for the calculation of Pad6 approximants in case all the interpolation points coincide with the origin.
III.S.6. The generalized c-algorithm
151
3.6. The generalized c-algorithm. Let us again consider two neighbouring staircases Sm+nand Sm+n+l.Each of them can be represented by a continued fraction of the form (3.8.). The successive convergents of the continued fraction constructed from Sm+ncan be obtained by means of the forward recurrence relations (1.3.). If we write [4] rm,n=
Pm,n -~ Qm,n
then
(3.10.)
and (3.11.)
Consequently, using (3.10.),
Using (3.10.) and (3.11.) we get
152
111.3.6. The generalized
E-
algorithm
Combining these two relations, we obtain
Performing analogous operations on Sm+n+lwe obtain
Using this result it is possible to set up the following generalized E- algorithm [4], in the same way as the c-algorithm for Pad6 approximants was constructed from the star identity (2.8.) :
(")=o
6- 1
(-n-l)
€2 n
$)
=O
= rm,o(z)
m=O,l,
...
n=0,1,
...
m = 0 , 1 , . ..
111.3.7.Stoer ’5 recursive m e t h o d
153
9.7. Stoer’s recursive m e t h o d .
The use of recursive methods is especially interesting when one needs the function value of an interpolant and not the interpolant itself. Several recursive algorithms were constructed for the rational interpolation problem, one of which is the generalized ealgorithm. Other algorithms can be found in [lo, 16, 221. We shall restrict ourselves here t o t h o presentation of the algorithm described by Stoer. Let m p!&(z) =
c
ai
xi
i= 0
be defined by
i= 0
in other words, they solve the interpolation problem (3.2.) starting a t z, and let a& and b!$n indicate the coefficients of degree m and n in the polynomiais p!& and qk!n respectively. The following relations describe the successive calculation of the rational interpolants lying on the main descending staircase
with and
154
111.3.7.Stoer's recursive method
Proof We will perform the proof only for the first set of relations, because the second (i) and part is completely analogous. In case one wants to proceed from pn,n-l px:A)l to p?)- the degree of the numerator may not be raised. The coefiicient of the term of degree n + 1 in the right hand side of (3.12.) is indeed
To check the interpolation conditions in xi for i = j,. . ., j + 274 we divide the set of interpolation points into three subsets: (a) ( f q ( , i - pk),Jx,)
= -(xi - x
G+l)
3+2n) an,n-l
(b) (fq(,', - px',)(x;) = 0 for i = j + 1 , . . ., j
(d
(i)
(fqn,n--l - p n , n - - l ) ( Z i ) = 0
+ 2n - 1 since
Again these relations can easily be adapted for the calculation of rational interpolants on other descending staircases. To calculate the interpolants in
one starts with
1119.7.Stoer’8 recursive method
155
where the c; are divided differences of f. To calculate the interpolants in
one starts with
where the wi are divided differences of l / f . As for Pad4 approximants, one can also give explicit determinantal formulas for the numerator and denominator of rm,Jz). We will postpone this representation until the next section.
156
III.4.1. Definition of rational Hermite interpolants
54. Rational Hermite interpolation.
4.1. Definition of rational Hermite interpolante. Let the points ( Z , } ~ ~heNdistinct and let the numbers s;(i 2 0) belong to IN. Assume t ha t the derivatives f(')(z;) of the function f evaluated a t the point 2; are given for f = 0, . . ., 8 ; - 1. Consider fixed integers j , k, m and n with
i m+n+l=Ca; i- 0
+
k
The rational Hermite interpolation problem of order (m, n) for f consists in finding polynomials m
and
n
q(z) =
C
b; zi
with p / q irreducible and satisfying
In this interpolation problem 8i interpolation points coincide with z;, so 8; interpolation conditions must be fulfilled in 2;. Therefore this type of interpolation problem is also often referred t o a5 the osculatory rational interpolation problem 1211. In case 8; = 1 for all d 2 0 then the problem is identical t o the rational interpolation problem defined at the beginning of this chapter. In case all the interpolation conditions must be satisfied in one single point 20 then the osculatory rational interpolation problem is identical to the Pad6 approximation problem defined in the previous chapter.
III..&1. Definition of rational Hermite interpolants
157
Instead of considering problem (3.13.) we can look at the linear system of equations
(fQ- P ) ( " ( . i )
=0 (14 - P ) ( L ) ( z i + l= ) 0
for -! = 0,. . .,8 ; - 1 with i = 0 , . . . , j for C = O , ..., k - 1
(3.14.)
and this related problem always has a nontrivial solution for p ( z ) and q(z), since it is a homogeneous system of m + n 1 equations in rn + n + 2 unknowns. Again distinct solutions have the same irreducible form p o l 4 0 and we shall call
+
PO
rm,n = --
40
where 90 is normalized such that qO(z0) = 1, the rational Hermite interpolant of order (rn,n) for f. The rational Hermite interpolation problem can be reformulated as a NewtonPad6 approximation problem. We introduce the following notations: y~ = zo for -! = 0 , . . ., 80 - 1 y,qi)+l = z i for C = 0 , . . ., 8 ; c i i = 0 for i > j c i j = f[yi,. . .,t/i] for i 5 j
- 1 with d ( i ) = 8 0
+ 81 + . . . + 8i-1 (i 2. 1)
with possible coalescence of points in the divided difference f [ y i , . . .,g j ] . If we put
with
Bj(4 =
n
i
(2
- YL-1)
then formally
This series is called the Newton series for f. Problem (3.14.) is then equivalent with the computation of polynomials
158
111.4.1.Definit i o n of ratio no1 He rmit e interpol ant8
and
such that
Problem (3.15.) is called the Newton-Pad6 approximation problem of order (n, n) for f . To determine solutions p and q of (3.15.) the divided differences di
= (f q - p ) ( y o , .
..,yi]
for i = 0 , . . .,m+
must be calculated and put equal t o zero. The following lemma, which is a generalization of the Leibniz rule for differentiating a product of functions, is a useful tool.
Lemma 3.1.
For the proof we refer t o 119). Using lemma 3.1. it is now possible t o write down the linear systems of equations that must be satisfied by the coefficients a; and b; in p and q : COO COI
bo = QO
60
combo
+ c11 61 =
+
Clmbl
+ ...+
(3.16a.) cmmb,
= a,
(3.16b.)
III.d.2. The table of rational Hermite interpolant8
159
Since the problems (3.14.) and (3.15.) are equivalent, the rational function rm,= can as well be called the Newton-Pad6 approximant of order ( m ,n ) to I. In the same way as for the rational interpolation problem the following theorem can be proved. Theorem 3.14.
The rational Hermite interpolation problem (3.13.) has a solution if and only if the rational Hermite interpolant rm,== po/qo satisfies (3.14.). 4.2. The table of rational Hermite interpolants. Once again we will order the interpolants rm,nin a table with double entry: r0,o
r0,1
f0,2
...
r3.0
For a detailed study of the structure of the rational Hermite interpolation table we refer to [3].We will only summarize some results. They are based on the following property. Theorem 3.15.
If the rank of the linear system (3.18b.) is n - t then (up t o a normalization) a unique solution P(z) and q(z) of (3.16.) exists with
ap 5
m-t
aq<_n-t where at least one of the upper bounds is attained. Every other solution p ( z ) and q(z) of (3.18.) can be written in the form P(Z)
= B(z) 4 2 )
dz) = where 38 <_ t.
8(4
111.4.2. The table of rational Hermite interpolanta
160
Proof Since the rank of the linear system (3.16b.) is n - t , solutions pl, 41 and p 2 , 4 2 of (3.10.) can be constructed such that BPl dql
5m 5 n -t
and
by choosing the free parameters of the homogeneous system in an appropriate way. Then one can prove, just as in theorem 3.1., that P1Q2 = P2Ql
Since
we must have either ap1 5 m - t or 892 5 n - t. Hence there exists a solution 8ii L m - t and 8q 5 n - t and it is easy to see that it is unique (up t o a normalization). Now the polynomials P (i )(2)= Bi(z)P(z)
p, i j of (3.16.) with
with 0
5i5t
do also solve the Newton-Pad6 approximation problem of order
( m ,n) because already
(fij
-
=
c
iz m+ a+ 1
diBi(Z)
What’s more, they are a linearly independent set of solutions and hence span the
111.4.2. The table of rational Hermite interpolant8
161
solution space. Consequently, all other solutions p and q of 3.16. are a polynomial multiple of p and ij. Now if both ap < m - t and a9 < n - t we would have more than t + 1 linearly independent solutions and thus the dimension of the solution space would be greater than t + 1 which implies that the rank of the system (3.16b.) would be less t ha n n - t and this is a contradiction. I Before describing the shape of sets in the table of rational Hermite interpolants that contain equal elements, it is important t o emphasize that the structure of the table can only be studied if the ordering of the interpolation points {z;)~,=N remains fixed once it is chosen. Since the polynomials p and q constructed in the previous theorem have the property t ha t their degrees cannot be lowered simultaneously anymore unless some interpolation conditions are lost, we shall call them a minimal solution. This does not imply that p / g is irreducible. However we still have
Theorem 3.16. Let p(z) and g(z) be the minima) sohition of the Newton-Pad6 approximation problem of order ( m , n ) for f and let the rank of the linear system (3.16b.) be 92
- t.
a) If ap = m - t - tl then all the rational Hermite interpolants lying in the triangle with corner elements rm-t--tl,n-t, rm-t-tl,n+t+tl and rm+t,n--t are equal t o r,-t-tE,n-t. b) If aq = n - t - tz then all thc rational Hermite interpolants lying in the triangle with corner elements rm-t,n-tt--tl, rm-t,n+t and rm+t+tz,n-t--ta are equal t o r,-t,n-t-.ta. c)
If i > m+n-2~+ t a + l
with dm+n--2t+to+l # 0 then all the rational Hermite interpolants lying in the triangle with corner elements rm-t,n-t, r,+t+ts,n-t and rnr-t,n+t+tp are equal t o rm-t,n--t.
III.4.3. Determinant
162
representation
(m-t-t ,n-t) 1
1
(m, n ) (m+t,n-t)
(m+t,n-t) (m+t+t2,n-t-t,)
(m-t ,n-t)
(m-t,n+t) (rn-t,n+t+t3)
(rn+t+t,n-t) 3
Figure 3.1.
For the proof we refer to [5]. 4.3. Determinant repreaentation.
Theorem 3.17.
If the rank of the system of equations (3.16b.) is maximal,then (up to a normalization) rm,n = E! is given by
40
III.4.4. Continued fraction
representotion
163
where L= i
and
Fi,j(z) = 0 if i
>j
The proof is left as an exercise (see problem (5)). Again one can see that in case all the interpolation points coincide with one single point, these determinant formulas reduce to the ones given in the chapter on Pad6 approximants since the divided differences reduce to Taylor coefficients.
4.4. Continued fraction repreeentation. If one considers staircases
in the table of rational Hermite interpolants, one can again construct continued fractions of which the successive convergents equal the elements of Tk.It is easy to see that these continued fractions are of the form
+ - dk.+2(Z \
- Yk+l)I I
&+dZ
+/
- Yk+2)l
+...
The coefficients 6,. . .,dh+l are divided differences (with coalescence of points) and the other coefficients can still be obtained using the generalized qdalgorithm. The generalization of Gragg's algorithm and the generalized calgorithm also remain valid for the calculation of rational Hermite interpolants.
164
III.4.5. Thiele's continued fraction ezpansion
4.5. Thiele '8 continued fraction ezpansion. From theorem 3.9. we write formally
(3.17.)
We consider now the limiting case z ; -+ xo for i
p,(z) =
21 .
C P ~ ( ~ O ,* J. zjl
z; lim -+ 2 ...,j
a=o,
Then (3.17.) becomes I
I
I
This is a continued fraction expansion of around 5 0 . Formula (3.18.) is obtained from (3.17.) in the same way as we set u p a Taylor series development from Newton's interpolation formula. Since (3.18.)is formal, one has t o check for which values of z the righthand side really converges t o f(z). We can calculate the p j ( z ) using Thiele's method [17] :
and with
III.4.5. Thiele '8 continued fraction ezpunaion
165
we have
Now p j - l [ z o , . . . , z j - l ] does not depend upon the ordering of the points 20
,...,xj-1. s o
Consequently (3.19a.)
To calculate
one uses the reIationship
(3.19b.)
We apply the formulas (3.19.) to construct a continued fraction expansion of f ( z ) = ez around the origin: P o ( % ) = e"
tpl(z) = e-"
iPr(z0)
=1
pl(z) = e-"
pz(z) = -2e"
'pZ(Z0)
= -2
pZ(z) = -ez
p3(z) = - 3 ~ "
'p3(ZO)
= -3
p4(z) = 2e"
'p4(ZO)
=
= 5e-"
'P5(z0)
=
p3(z)
= -2e-=
P 4 ( 4 = e2
p5(2)
188
So we get
111.4.5. Thiele’e continued fraction ezpanaion
111.5. Convergence of rational Hermite interpolanta
167
55. Convergence of rational Hermite interpolants. The theorems of chapter I can be used to investigate the convergence of interpolating continued fractions. We shall now mention some results for the convergence of columns in the table of rational Hermite interpolants. r2,o, ] . . .}, in other words The first theorem deals with the first column { r o , o ,q , ~ it is a convergence theorem for interpolating polynomials. For given complex points (20,. , z j } we define the lemniscate
.
~ ( z o , ..,z,, t) = { z E C
1
~ (z zo)(z - 2 1 ) . . .(z - z,)l = r}
Broadly speaking, the convergence of an arbitrary series of interpolation does not depend on the entire sequence of interpolation points y; (as defined in the Newton-Pad6 approximation problem) but merely on its asymptotic character, as can be seen in the next theorem.
Theorem 3.18. Let the sequence of interpolation points (yo, y1 ,~ sequence
>it
Vk(j+l)+i
. .) be
2 , .
asymptotic t o the
= zi
for i = 0,.. . , j . If the function j(z) is analytic throughout the interior of the lemniscate B(z0 , .. ., z,, r) then the rm,0 converge to f on the interior of B(z0 , . . .,z j , r). The convergence is uniform on every closed and bounded subset interior to B(z0,. . . ] z j , r).
For the proof we refer to [20 p. 611 and 19 pp. 90-911. Let us now turn t o the case of a meromorphic function f with poles 2 0 1 , . . ., wn (counted with their multiplicity). For the rational Hermite interpolant of order (m,n) we write Pm,n
rm,n = __
4m,n
and for the minimal solption of the Newton-Pad4 approximation problem of order (m,n) we write fS,,,(z) and qm,n(z).
168
111.5. Convergence of rational Hermite inderpolanta
Let the table of minimal solutions for the Newton-Pad4 approximation problem be normal. According to [6] we then have &j,,+ = n. Let wim) ( (i = 1 , . . .,n) be the zeros of qm,n for rn = 0 , 1 , 2 , . . - and let p i = I(wi - Z O ) ( W ; - 2 1 ) .. .(wi - z j ) l with 0 < p1 5 p 2 5 . . . 5 pn 5 at < r for a positive constant a.
Theorem 3.19. If the sequence of interpolation points (yo, yl, 112, . . .} is asymptotic t o the sequence { Z O , Z ~ ,...,zj,zo,z1, . . .,zj,20, 21,. . .,z j , . . .}, if f is meromorphic in the interior of B(z0,. . ., zg’, r) with poles w 1 , . . , , w , counted with their multiplicity and if the table of minimal solutions for the Newton-PadB approximation problem is normal, then
+
= w , o(am>
i = 1, ..., n
and
uniformly in every closed and bounded subset of the interior of B(z0, . . .,z j , r) not containing the points ~ 1 , ...,W n . The proof is given in [4].
III. 6.1. Interpolating branched continued fraction8
169
86. Multivariate rational interpolants. We have seen t h at univariate rational interpolants can be obtained in various equivalent ways: one can calculate the explicit solution of the system of interpolatory conditions, start a recursive algorithm, calculate the convergent of a continued fraction or solve t he Newton-Pad6 approximation problem. We will generalize t h e last two techniques for niiiltivariate functions. These generalizations are written down for t he case of two variables, because the situation with more than two variables is only notationally more difficult. More details can be found in [7] and [8]. 6.1, Interpolating branched continued fractions.
Given two sequences { Z O , Z ~ , Z ? , . . .} and {yo,y1,y2 ,...} of distinct points we will interpolate the bivariate function f ( z , y ) at the points in (zo, X I , ~ 2 , . ..} X { y o , y l , y 2 , . . .}. 'TO this end we use branched continued fractions symmetric in t h e variables z a n d y and we define bivariate inverse differences as follows:
170
MI. 6.1. Interpolating branched continued fractions
Theorem 3.20. (3.20.)
with
Proof From theorem 3.9. we know that
Let us introduce the function gO(z, y) by
where
By calculating inverse differences
(0)
for go we obtain
III.6.l. Interpolating branched continued fractions
1
where ho(z,y) =
&I
yo, y]. So already
By computing inverse differences R (j 0, k) for ho we get
It is easy to see that
fl
- I/n
171
172
III.6.I . Interpolating branched continued fractione
From this we find by induction that
So we can write
where
If we introduce inverse differences which provides us with a function
C:l for hl
g 1 we can repeat the whole reasoning and inverse differences ?rj,k: (1 1
III. 6.1.Interpolating branched continued fraction8
173
In the same way as for ho we find
Finally we obtain the desired interpolatory continued fraction.
I
To obtain rational interpolants we are going to consider convergents of the branched continued fraction (3.20.). To indicate which convergent we compute we need a multi-index
The Fith convergent is then given by
with
For these rational functions the following interpolation property can be proved.
III. 6.1. Inte t p o lat ing branched c o ntinue d fr actt o )a8
174
Theorem 3.21. The convergent CK(Z,y) of (3.20.) satisfies C d Z t l , YC,) = I(ZC, t YC,)
for
( e l , &)
belonging to
Proof Let C = min (&, Cz). From theorem 3.20. we know that
where
Now f(%, yt,) = CE(ZC,, yt,) if and only if the following conditions are satisfied Cln
Q0
5 i 5 L : L1 I
mi,
and Lz 5 mi,
This is precisely guaranteed by saying ( t i , &) E I .
I
;...-
111.6.2. General order Newton-Pad6 approzimant8
175
For instance, if moz 2 m l , 2 ... 2 mnz and mOy 2 m l y 2 ... 2 mny the . as we can boundary of the set I is given by n = ( n , m o z J m o y.,.,mnr,mny), tell from the next picture which is drawn for n = 2.
m1y
mZy
-. I:.
0
-*-.
-. . 4
*
4
0
0
0
~
m2x
m1x mOx
Figure 3.8.
We illustrate this technique with a simple numerical example. Let the following d ata be given: z; = i for i = 0 , 1 , 2 , . . . and yg = j for j = O,1,2,. .. with f ( z i , Y i )= (i + j)'. Take A = ( 1 , 2 , 2 , 1 , 1 ) . Then we have to compute
The resulting convergent is
6.2. General order Newton-Pad6 approzimants.
Consider two sequences of real points { Z ; ) ~ ~ N and { Y , } ~ ~where N coalescent points get consecutive numbers. For a bivariate function f ( 2 , I/)we define the following divided differences
176
111.6.2. General order Newton-Padk approzimanta
(3.21a.) or equivalently
One can easily prove that (3.21a.) and (3.21b.) give the same result. When the interpolation points z;,. . ., zi+ri-l and yj, . . .lgj+sj-l coincide, then one must bear in mind that
We consider the following set of basis functions for the real-valued polynomials in two variables:
This basis function is a bivariate polynomial of degree i
+j .
III.6.2. General order Newton-Pad6 approzimants
177
With c k ; , J j = f [ z k , . . .,z i ] [ y c , .. ., yj] we can then write in a purely formal manner [l pp. 160-1641
C
f(z,1/> =
coi,oj
B i i ( z ,YI
(i,j)EINa
The following lemmas about products of basis functions B i j ( z ,y) and about bivariate divided differences of products of functions will play an important role in the sequel of the text.
Lemma 3.2. For k
+ L 2 i + j the product B i i ( ~ , y&(z,y) ) p=o
=
u-0
Proof We write B i j ( ~v ,) = Bio(2,v ) B o j ( z ,Y). Since Bio(z,y) is a polynomial in z of degree i we can write
and
with the convention th at an empty product is equal to 1. Consequently
u=o
p=o
which gives the desired formula if we put, ,A
= a,&.
178
111.6.2. General order Newton-Pedt apptoziments
A figure in IN2 will clarify the meaning of this lemma. If we multiply B;j(z,v) by &(Z, g) and k + l >_ a + J' then the only occuring Bpu(z,v ) in the product are those with ( p , v ) lying in the shaded rectangle.
I
k
i
k+i
Figure 3.3.
Lemma 3.3.
The proof is by induction and analogous t o the proof of th e univariate case. The definition of multivariate Newton-Padh approximants which we shall give is a very general one. It includes the univariate definition and the multivariate Pad6 approximants from the previous chapter as a special case as we shall see at the end of this section. With any finite subset D of IN2 we associate a polynomial
Given the double Newton series
111.6.2. General order Newton-Pad4 approzimanta
179
with c o i , ~ = , f[z0, . . ., z ; ] [ y o , .. ., y j ] ] we choose three subsets N , D and E of IN2 and construct an “ / D I E Newton-Pad4 approximant to f ( z , y ) as follows:
C
p(z,y) =
(i,j)EN
q(z,y)=
q j ~ ; (z,y) j b;j Rij (2, y )
( i , j W
(f Q - P)(z, y ) =
C
( N from “numeratorn)
(3.22a.)
(Dfrom “denominator”)
(3.22 b.)
d i j Bij (2,
Y)
( E from “equatiom”)
(3.22c.)
(i,j)EN2\E
We select N , D and E such that
D has n + 1 elements, numbered ( i o , j o ) , . . ., (in,in)
N C E E satisfies the rectangle rule, i.e. if (i,j) E E then (k,L) E E for k 5 iand L 5 j E\N has at least n elements.
Clearly the coeBcients d;, in
(f q - P N Z , Y ) =
C
dij
Bij(z,y)
(i,i)ENa
are dij
=(~~-P)[zo,...,z~~[Yo,...,Y~~
So the conditions ( 3 . 2 2 ~are ) equivalent with
(.f q - P)[zo,. . ., ~ i ] [ y o ,. ..,i/i] = 0 for (i,j) in E
(3.23.)
The system of equations (3.23.) can be divided into a nonhomogeneous and a homogeneous part:
( I q ) [ z o , . Z i ] [ y ~ ,..., Y j ] = P [ ~ o ., .,. Z i ] [ ~ o ,...,y j ] (f q)[zo, . . ., z ; ] [ y o ,. . ., y j ] = O for (i,j ) in E\N a,
Let’s take a look a t the conditions (3.23b.).
for ( i , j ) in N
(3.23a.) (3.23b.)
111.6.2.General order Newton-Pad6 approximants
180
Suppose that E is such t h a t exactly n of the homogeneous equations (3.23b.) are linearly independent. We number the respective n elements in E\N with ( h l ,k l ) , . . ., (hR?k,) and define the set = { ( h l i I),
.‘
. J (hnl
‘n)}
c E\N
( H from “homogeneous equations”)
By means of lemma 3.3. we have
Since the only nontrivial q [ z o , . . zP][yo,. . .,yu] are the ones with ( p , v) in D we can write .J
Remember that f [ z C c J . ,. z i ] [ y U ,..,yj] . = 0 if p > i or v > j. So the homogeneous system of n equations in n + 1 unknowns looks like
because
D
= {(iO, ~ O ) J * . * J (in?jn))
As we suppose the rank of the coefficient matrix t o be maximal, a solution q(z, y) is given by
111.6.2. General order N e w t o n - P a d 6 approximanta
181
By the conditions (3.23a.) and lemma 3.3. we find
Consequently a determinant representation for p(z y) is given by
(3.25b.)
If for all k, L 2 0 we have Q(ZkJYc)# 0 then
$(z,g) can be written as
with e i i = +[so,.. ., z i ] [ y o , .. yj]. Hence by the use of lemma 3.2. and since E satisfies the inclusion property .)
The following theorem describes which interpolation conditions are now satisfied by P / 9 .
182
111.6.2. General order Newton-Pad4 approzimanta
Theorem 3.22.
If
q ( Z k , yt)
# 0 for
( k , I!) in E then
where
If
r k = 1 = 6~
this reduces t o
Proof Given rk and 6~ for fixed (zk,yc), consider the following situation for the interpolation points, with respect to E
(3.26.)
I
I
, I
k Figure 3.4.
I
I 1
k+fk.1
I
*
111.6.2. Gener af order Nerut o n - f ad6 approzimant 8
and define
T
Using these d e h i t i o n s we rewrite I as
with
Because q ( z t , yt)
# 0 for ( k , l ) in R we have
183
IIl.S.2. General order Newton-Pad4 approximonte
184
To check the interpolation conditions we write
apf” R ‘J. . = p t - u ( B ~Boj) o
-
ax@ dy”
If we cover N2\E with three regions
because B;a(z,y) contains a factor (x - zk)pE+l, and
au
B~~
ay”
l(zkrYl)
= 0 for (i,j ) in B and
(I(,
v) in I
because Boj(x, y) contains a factor (y - Y $ ‘ E + ~ . Analogously
a’ B~~ axp
l(x*iYl)
= 0 for (i,j)in
C and
( p , v ) in I2
The most general situation for the interpolation points with respect to E is slightly more complicated but completely analogous to the one given in (3.26.). We illustrate this remark by means of the following figure:
111.6.2. General order Newton-Pad6 approzirnants
185
I
I
I
I
The proof in this case is performed in the same way as above.
I
We will now obtain the determinant representation given in theorem 3.17. for univariate Newton-Pad6 approximants, from the determinant representations (3.25a.) and (3.25b.). Consider the Newton interpolating series for f ( x , 0) and choose
+
If the points {(k,O) I m + 1 5 k 5 m n } supply linearly independent equations, then the determinant representations for p(z, 0) and q(z, 0) are
186
111.6.2.General order Newton-Pad6 approzimunts
We c a n also obtain the multivariate Pad6 approximants defined in the previous chapter as a special case. One only has t o choose
D = { ( i , j )I m n 5 i + j 5 mn+n} N={(i,j)Imn~%’+jImsa+m} b’= { ( i , j ) 1 mn 5 i t j _< mn + m t n }
because when all the interpolation points coincide with the origin, then
Bij
( 2 , ~= ) Z ’
yi
Let us now illustrate the multivariate setting by calculating a Newton-Pad4 approximant for 2 f ( z , ? / )= 1 + ___ sin(xy) 0.1 - y
+
with y j = ( j- 1)fi
i = 0 , 1 , 2 ' ... j = 0,1,2,. . .
The Newton interpolating series looks like 10 f(Z,YI = 1 + x+ zb + d3 O.l+fi O.l+&
Choose
111.6.2. General order Newton-Pad4 approzimants
Writing down the system of equations (3.23b.), it is easy t o check that
LI
= {(2? ')I
('1
2))
The determinant formulas for p(z, V ) and q(z, v ) yield 1 q ( z , f l )=
-
V-tJ;;
c02,01
c12,11
c01,02
L'11,02
c02,11
c01,12
100 1 0.01 - T (1 - 0.1 fi(Y
+
with
1
1
Finally we obtain
- 0.1+2-y 0.1 - y
+ d4)
187
IH. Prob lema
188
Problems. Let
o rm,= = P~40
be the rational interpolant of order (m,n) for f ( z ) with m' = apo and fir = r3g.o. Prove th at there exist at least m' + nr + 1 points {yl , .. .,ye} from the points ( 2 0 , . . ., z ,+,} such th a t rm,,(vi) = f(vi) for i = 1 , . . ., 5 . Formulate and prove the reciprocal and homografic covariance of rational inter pol ants. Prove theorem 3.8. Which interpolation conditions are satisfied by the nth convergent of the continued fraction (3.4) if a) d, = 0 b) d, = (x, Prove theorem 3.17 Prove the following result for the error function (f - rm,,)(z). Let I be ~ . an interval containing all the interpolation points 5 0 , . . ., z ~ + Then 'dz E I ,
3y, E I :
Compute rZ,z(z) satisfying rZ,z(si) = f(q) for i = 0 , . . ., 4 and r1,3(z) satisfying rl,s(zi) = f(zi) for i = 0 , . . ., 4 with z; = d (i = 0 , . . .,4) and f(z0) = 4, f ( z 1 ) = 2, f ( z 2 )= 1, f(z3) = -1, f ( Z 4 ) = -4If f ( z ) is a rational function
with at 5 m and 8 s 5 n then k with 0 5 k 5 2 max(m, n).
'pk[Zo,.
. .,X k - 1 ,
21
is constant for a certain
111. P r o ble m8
(9)
If for some k, p k [ Z O , . . tion.
(10)
Compute
r2,1(2)
(i = 0 , . . ., 3) and
of
.)
zk-1,2]
189
is constant then f(z) is a rational func-
r 2 , 1 ( z ; ) = f(z;) for i = 0 , . . . , 3 with z i = i = I , f(z1)= 3, f(z2) = 2, f(z3) = 4 J by means
satisfying f(z0)
a) the generalized qd-algorithm
b) the algorithm of Gragg (11)
Construct a continued fraction expansion using Thiele's method for f ( z ) = l n ( l + z) around 3: = 0.
(12)
Calculate the inverse differences for gl(z,y) and 7r!.li for h l ( ~y), and perform one more step in the proof of theorem 3.20. in order to obtain the contribution b---zl)(Y - Y l ) B 2 ( 2 , Y)
(!:i
with B 2 ( Z 1 Y ) = Pz[% 2 1 , Z2l[Yo,
in the continued fraction (3.20.).
Y l , Y21
III. Remarka
190
Remarks. (1)
Instead of polynomials
m
i=O
and
n i-0
one could also use linear combinations m
i= 0
and
of basis functions (g;}iEN, which we call generalized polynomials, and study the generalized rational interpolation problem
A unique solution of this interpolation problem exists provided ( g i ( Z ) ) ; ~ N satisfies the Haar condition, i.e. for every k 2 0 and for every set of distinct points (20,. . .,zk} the generalieed Vandermonde determinant
Examples of such interpolation problems can be found in [lS]. A recursive algorithm for the calculation of these generalized rational interpolants is given in [HI.
III. Re m ar ka
(2)
19 1
The Newton-Pad6 approximation problem (3.15.) is a linear problem in t ha t sense t hat rm,ncan he considered as the root of the linear equation
where p and q are determined by the following interpolation conditions (q f - p ) ( z j ) = 0
+
j = 0 , . . .,m n
Instead of such linear equations one can also consider algebraic equations
where the polynomials pi of degree m;are determined by
More generally we can consider for different functions fo(z), . . ., fk(Z) the interpolation conditions
An extensive study of this type of problems is made in [14] and [lo]. (3)
Rational interpolants have also been defined for vector valued functions [ l l , 251 using generalized vector inverses. For other definitions of multivariate rational interpolants we refer to [17] and [S]: Siemaszko uses nonsymmetric branched continued fractions while in [8] Stoer's recursive scheme for the calculation of univariate rational interpolants is generalized t o the multivariate case.
111.Reference8
192
References. Computing methods I. Addison Wesley,
[
11 Berezin J. and Zhidkov N . New York, 1965.
[
A generalization of the qd-algorithm. J. Comput. Appl. 21 Clsessens G. Math. 7, 1981, 237-247.
[
A new algorithm for osculatory rational interpolation. 31 Clsessens G . Numer. Math. 27, 1976, 77-83.
[
41 Claessens G.
[
A useful identity for the rational Hermite interpolation 51 Claessens G. table. Numer. Math. 29, 1978, 227-231.
[
61 Claessens G. On the Newton-Pad4 approximation problem. J. Approx. Theory 22, 1978, 150-260.
[
71 Cuyt A. and Verdonk B. General order Newton-Pad4 approximants for multivariate functions. Numer. Math. 43, 1984, 293-307.
[
81 Cuyt A. and Verdonk B. Computing 34, 1985, 41-61.
[
91 Davis Ph.
Some aspects of the rational Hermite interpolation table and its applications. Ph. D., University of Antwerp, 1976.
Multivariate rational interpolation.
Intcrpolation and approximation. Blaisdell, New York, 1965.
Contribution B l’approximation de fonctions de la variable complexe au sens Hermite-Pad6 et de Hardy. Ph. D., University of Grenoble, 1980.
[ 101 Della Dora J .
[ 111 Graves-Morris P. and Jenkins C . Generalised inverse vector valued rational interpolation. In [22], 144-156. ( 121 Larkin F.
Some techniques for rational interpolation. Comput. J. 10, 1967, 178-187.
An algorithm for generalized rational inter[ 131 Loi S. and Me Innes A. polation. BIT 23, 1983, 105-117.
[ 141 Liibbe W.
Ueber ein allgemeines Interpolationsprobiem und lineare Identitaten zwischen benachbarten Losungssystemen. Ph. D., University of Hannover, 1983.
III. Reference8
193
151 MuhIbach G . The general Neville-Aitken algorithm and some applications. Numer. Math. 31, 3978, 97-110.
[ 161 Saleer H .
Note on osculatory rational interpolation. Math. Comp. 16, 1962, 486-491.
[ 171 Siemaszko W.
Thielp-type branched continued fractions for twovariable functions. J. Corn put. Appl. Math. 9, 1983, 137-153.
[ 181 Stoer J.
Ueber zwei Algorithmen zur Interpolation mit rationalen Funktionen. Numer. Math 3, 1961, 285-304.
[ 191 Thiele T.
Interpolationsrechnung. Teubner, Leipzig, 1909.
[ 201 Walsh J.
Interpolation and approximation by rational functions in the complex domain. Amer. Math. SOC.,Providence Rhode Island, 1909.
[ 211 Warner D.
Hermite interpolation with rational functions. Ph. D., University of California, 1974.
[ 221 Werner H . and Biinger H .
Pad6 approximation and its applications. Lecture Notes in Mathematics 1071, Springer, Berlin, 1984.
[ 231 Wuytack L. On some aspects of the rational interpolation problem. SIAM. J. Numer. Anal. 1 1 , 1974, 52-80. [ 241 Wuytack L.
On the osculatory rational interpolation problem. Math. Comp. 29, 1975, 837 - 843.
[ 251 Wynn P.
Continued fractions whose coefficients obey a noncommutative law of multplication. Arch. Rational Mech. Anal. 12, 1963, 273-31 2. Ueber einen Interpolations-Algorithmusund gewisse andere Formeln, die in der Theorie der Interpolation durch rationale Funktionen bestehen. Numer. Math. 2, 1960, 151-182.
[ 261 Wynn P .
195
.
CBAPTER N:Applications
5 1. Convergence 1.1. 1.2. 1.3. 1.4. 1.5.
The The The The The
acceleration . . . . . . . . . . . . . . . . . . . . univariate t-algorithm . . . . . . . . . . . . . . . qd-algorithm . . . . . . . . . . . . . . . . . . . . algorithm of Bulirsch-Stoer . . . . . . . . . . . . . p-algorithm . . . . . . . . . . . . . . . . . . . . . multivariate €-algorithm . . . . . . . . . . . . . .
197
. 197 205
. 209 213
. 216
$2. Nonlinear equations . . . . . . . . . . . . . . . . . 2.1. Iterative methods based on Pad4 approximation . 2.2. Iterative methods based on rational interpolation 2.3. Iterative methods using continued fractions . . . 2.4. The qd-algorithm . . . . . . . . . . . . . . . 2.5. The generalized qd-algorithm . . . . . . . . .
. . . . . 220 . . . . . . 220 . . . .
. . . . . 227 . . . . . 233 . . . . 233 . . . . . 236
53. Initial value problems . . . . . . . . . . . . 3.1. The use of Pad4 approximants . . . . . 3.2. The use of rational interpolants . . . . 3.3. Predictor-corrector methods . . . . . 3.4. Numerical results . . . . . . . . . . . 3.5. Systems of first order ordinary differential
. . . . .
. 238 . 241 . 243
. . . . .
. . . . .
. . . . .
. . . . .
. . . . . . . . . . . . . . . . . . . . equations . . . .
238
244
. 247
.
54 Numerical integration . . . . . . . . . . . . . . . . . . . . . 250 4.1. Methods using Pad4 approximants . . . . . . . . . . . . . 251 4.2. Methods using rational interpolants . . . . . . . . . . . . 252 4.3. Methods without the evaluation of derivatives . . . . . . . . 253 4.4. Numerical results for singular integrands . . . . . . . . . . 254
55 . Partial differential equations
. . . . . . . . . . . . . . . . . . 257
§6.Integral equations . . . . . . . . . . . . . . . . . . . . . . . 260 0.1. Kernels of finite rank . . . . . . . . . . . . . . . . . . 260 . . . . . . . . . . . . . . 262 6.2. Completely continuous kernels
Problems
. . . . . . . . . . . . . . . . . . . . . . . . . . .
265
Remarks
. . . . . . . . . . . . . . . . . . . . . . . . . . .
267
. . . . . . . . . . . . . . . . . . . . . . . . . .
268
References
196
“It i s my hope that by demonstratdng the e m e with which the various transformations may be effected, their field of application might be widened, and deeper insight thereby obtained into the problems for whoae aoh6tion the tran8formatdons have been uaed.
P . WYNN (1956).
- “On a
device for computing the e m ( & )
tran~f~rmation~’
Z V . l . l . The univariate €-algorithm
197
The approximations introduced in the previous chapters will now be used to develop techniques for the solution of various mathematical problems: convergence acceleration, numerical integration, the solution of one or more simultaneous nonlinear equations, the solution of initial value problems, boundary value problems, partial differential equations, integral equations, etc. Since these techniques are based on nonlinear approximations they shall be nonlinear thernselves. We shall discuss advantages and disadvantages in each of the sections separately and illustrate their use by means of numerical examples.
§I. Convergence acceleration. 1.1. The univariate r-algorithm. Consider a sequence { a ; } ; , ~ of real or complex numbers with lim ai = A
i-
00
Since we are interested in the limiting value A of the sequence we shall try to construct a sequence (b;),.,N that converges faster to A , or
We shall describe here some nonlinear techniques that can be used for the construction of { b;},N. Consider the univariate power series
i-t
with Vai = a; - a;-1. Then clearly for the partial sums
we have
Fk(1)= ak
k = 0, 1,2, . . .
If we approximate f(z) by r ; , ; ( z ) , the Pad4 approximant of order (i,i) for f , then we can put b; = r;,;(l)
i = 0 , 1 , 2 , . ..
198
I V . l . l . The univariate €-algorithm
For the computation of b; the c-algorithm can be used:
Then
The convergence of the sequence {biIiEm depends very much on the given sequence { a , } ; , N . Of course the convergence properties of { b ; } ; , ~are th e same as those of the diagonal Pad6 approximants evaluated a t z = 1 and for this we refer to section 4 of chapter 11. In some special cases it is possible to prove acceleration of the convergence of ( a ; } i E N . A sequence {a;}iEN is called totally monotone if
Ak u ; > O
d,k=0,1,2
,...
where A k a , = Ak-' ai+i - Ak-' ai and Ao a; = a;. In other words, {ai}iEN is totally monotone if
a0 2 a1 2 a2 2 ... 2 0 Aao 5 Aal 5 h a 2 5 . .. 5 0 A2ao 2 A 2 a l 2 A2a2 2 . . . 2 0 and so on.
IV.1.1. The uniuariate 6-algorithm
199
Note t ha t every totally monotone sequence actually converges. The sequences (Xi}i,-~ for 0 <_ X 5 1 and {l/(i l ) } ; , ] ~are for instance totally monotone sequences. The close link with the theory of Stieltjes series becomes clear in the following theorem [4 p.811.
+
Theorem 4.1. The sequence { o ~ } ~ €isN totally monotone if and only if there exists a realvalued, bounded and nondecreasing function g on [0, 11 taking on infinitely many different values such that 1
a; = I 0
t' d g ( t )
i = 0,1,2,. ..
A sequence { a ; } ; , ~is called totally oscillating if the sequence
ai}iE]N
is totally monotone. One can see th at every convergent totally oscillating sequence converges to zero. For these sequences the following results can be proved [4 pp. 83-85]. Theorem 4.2.
If the 6-algorithm is applied to the sequence { U ; } ~ ~ ] N with 1imi-m a i = A and if there exist constants a and @ such that { a a ; + @ } i E ~is a totally monotone sequence, then lim
cfi
=A
lim
($2
k >_ 0 and fixed
=A
e 2 0 and k e d
f-00
k+m
If also lirn
i-oo
then
ai+l
-A # 1
-_-
ai-A
200
IV.1.1. The univariote t-algorithm
Theorem 4.3.
If the €-algorithm is applied t o the sequence { a ; } i E mwith limi-a; = A = 0 and if there exist constants cr and B such that (&a; + @ ) i E ~is a totally oscil-
lating sequence, then
We illustrate these theorems with some numerical results. Consider
This is a totally monotone sequence. In table 4.1. we have listed the values for n = 0 , . . ., 4 and m = 0, . . ., 10. So the values b i = can be found on the main diagonal. Clearly the convergence of the sequence { a i } i e ~is accelerated.
EE)
All these computations were performed using double precision arithmetic (56digit binary representation). To illustrate the influence of data perturbations and ( m - 4 obtained using single rounding errors we give in table 4.2. the same values E~~ precision input and single precision arithmetic (24-digit binary representation). In the tables 4.3. and 4.4. the respective results are given for the totally oscillating sequence
Here the convergence of the sequence {b;}+-N is much faster th a n th a t of the sequence {ai}i,-N. These examples illustrate that the influence of data perturbations and rounding errors can be important. When is nearly equal to one can loose a lot of significant digits. In [62] i t is shown how this misfortune may be overcome.
tP1)
Table 4.1.
56-digit binary reprea ent otio n %
m\n
S 0
1
2
3
4
0
0.10000000D+01
0.86666667D+00
0.52173913D+OO
0.43636364D-bOO
0.37874803D-I-00
1
0.50000000D+00
0.25000000D+00
0.17847058D+00
0.13801453DfOO
0.113Q8283D-tOO
2
0.33333333D-tOO
0.16668667D+00
O.llllllllD~00
0.85271318D-01
0.69561659D-01
3
0.25000000D+00
0.12500000D+00
0.83333333D-01
0.62500000D-01
0.50607287D-01
4
0.20000000D~00
0.10000000D+00
0.66666667D-01
0.50000000D-01
0.40000000D-01
5
0.16666667D+00
0.83333333D-01
0.55555556D-01
0.41666667D-01
0.33333333D-01
6
0.14285714D-kOO
0.71428571D-01
0.47610048D-01
0.35714286D-01
0.2857142QD-01
7
0.12500000D+00
0.62500000D-01
0.41666667D-01
0.31250000D-01
0.25000000D-01
8
O.lllllIllD~OO
0.55555556D--01
0.37037037D-01
0.27777778D-01
0.22222222D-01
9
O.lOOOOOOOD+OO
0.50000000D-01
0.33333333D-01
0.25000000D-01
0.20000000D-01
10
0.9090Q091D-01
0.45454545D-01
0.30303030D-01
0.22727273D-01
0.18181818D-01
k !Y
% F:
2. c P
3.
R
cs
hl
0
N
Table 4.2.
24 -digit binary repreaent ation
m\.
0
1
2
3
4
c..
0
0.10000000E+01
0.86666680E+00
0.52173018E+00
0.436363703+00
0.378748OOE+OO
k
1
0.500000003~00
0.250000083+00
0.17647058E+00
O.l3801455E+OO
0.113082853~00
Y
2
0.33333334E-kOO
0.16666664E+00
0.11111123E+00
0.852712323-01
0.69561742E-01
E
3
0.25000000E+00
0.12500004E+00
0.83333202E-01
0.62500395E-01
0.50605543E-01
4
0.20000000E+00
0.10000000E+00
0.66666812E-01
0.49907089E-01
0.40013745E-01
5
0.16666667E+00
0.83333351E-01
0.55554848E-01
0.416750723-01
0.332807823-01
6
0.14285715E+00
0.71428463E-01
0.47620941E:-01
0.356953373-01
0.28692538E-01
7
0.12500000E+00
0.62500104E-01
0.41664604E:-01
0.312740473-01
0.247051133-01 0.22477847E-01
8
O.llllllllE~00
0.55555556E-01
0.370379803-01
0.27756441E-01
0
0.10000000E+00
0.5000001OE-01
0.33332549E-01
0.2502411OE-01
O.l0564494E:-Ol
10
0.90909004E-01
0.45454517E-01
0.30304417E-01
0.226919743-01
0.18706985E-01
z 2. C
e
2.
f: m
m
it
rg 0 7
3
Table 4.3.
56-digit binary repreaentatran Ls,
m\.
5
0
0
0.10000000D+01
1
-0.50000000D~00
2
0.33333333D.fOO
3
4 5 6
7 8
-0.25000000D+00
1 0.40000000D~00 0.35714286D-01 -0.98039216D-02 0.40322581D-02
0.20000000D~00 -0.20408163D-02 -0.16666667D+00 0.14285714D-kOO -0.12500000D+00 0.111111llD+OO
0
-0.100OOOOOD+00
10
0.~0009001D-01
0.1 17370881)-02 -0.73637703D-03 0.40212508D-03 -0.34506556D-03 0.25125628D-03 -0.18860807D-03
2 0.25531015D-kOO -0.2 18078 10D-01 0.1 1454754D-02 -0.25960540D-03 0.83229297D--04 -0.330401 llD-04 0.15170805D-04 -0.77490546D-05 0.42861980D-05 -0.252506 123)-05 0.1 5651583D-05
3 0.18604651 D+OO 0.11833001D-01 -0.38107008D-03 0.34806706D-04 -0.73003358D--05 0.20314308D-05 -0.688Q106OD-06 0.27006940D-06 -0.11826320D-06 0.58510073D-07 -0.289770043)-07
4
0.14600870D+00 -0.96160448D-02 0.15341245D-03 -0.89086733D-05 0.10459978D-05 -0.20044207D-06 0.53279358D-07 -0.16203117D-07
; Y
z C
2. C P -.
l
a
cs m
a
cq
*
a 0
0.5656006OD-08 -0.220288523)-08 0.0378304OD-00
ts 0 W
01
0
Table 4.4.
b b
m+l Rd-digit binary representation
m n 0
0 0.10000000E~01
1
-0.500OOOOOE~00
2
0.333333348+00
3
4 5
-0.25000000E+00 0.20000000E+00 -0.16666667E+00
6
0.142857153+00
7
-0.12500000E~00
8
9 10
O.llllllllE~00 -0.10000000E+00 O.QO909094E-01
1
0.400000043+00 0.357142843-01 -0.880381823-02 0.40322733E-02 -0.20408179E-02 0.11736Q87E-02 -0.738376743-03 0.48213349E-03 -0.345061333-03 0.25125383E-03 -0.18860842E-03
2 0.25531918E-kOO -0.218878OOE-01 0.1145486OE-02 -0.25958874E-03 0.832260823-04 -0.3305387QE-04 0.15181173E-04 -0.77439190E-05 0.428871413-05 -0.25255961E-05 0,156517563-05
4
3 0.18604654E3+00 0.118331053-01 -0.38106500E-03 0.34898013E-04 -0.73038805E-05 0.20300561E-05 -0.68641026E-06 0.273580183-08 -0.11687172E-08 0.56633620E-07 -0.28493082E-07
0.146908823+00 -0.9816957OE-02 0.15340716E-03 -0.881250783-05 0.10442981E-05 -0.20828127E-06 0.56101026E-07 -0.138258973-07 0.860953913-08 -0.17431754E--08 0.183647753-08
* Y
k k
Y
z E
a.
a 3. C
h R
m
$-
-l
3
I V . 1.2. The qd- algorithm
205
Various generalizations t o accelerate the convergence of a sequence of vectors, matrices, and so on exist. We refer to [4] and [Sl]. The convergence of a multidimensional table of real or complex numbers is treated in section 1.5. 1.8. The qd-algorithm.
To accelerate the convergence of { ai}iEN we construct again
i- 1
We have seen in section 3 of chapter 11 th at it is possible t o construct a corresponding continued fraction I
I
I
where the coefficients d i can be calculated by means of the qd-algorithm. If we calculate the i t h convergent Ci(z) of this continued fraction, we can now put 6i = Gi(1)
i = 0 , 1 , 2 , .. .
The numbers 6; can be computed using a continued fraction algorithm from section 5 of chapter I. Remark that theoretically the numbers b 2 ; here are the same as the 6; computed by means of the €-algorithm. In practice there are however differences due to rounding errors. We will illustrate this numerically. The tables 4.5. and 4.6. illustrate the effect of perturbations when the qd-table is computed: the results in table 4.5. are obtained using double precision arithmetic while those in table 4.6. are obtaincd using single precision arithmetic. Input was
So f(z)= ez and limi-m a; = e = 2.718281828.. . From the tables 4.5. and 4.6. the values b; were computed using the three-term recurrence relations (1.3.) respectively in double and single precision. The results can be found in the respective tables 4.7. and 4.8., where they can also be compared with the values of the c-algorithm.
Table 4.5.
hl
0 Q,
k=l,2,
92
-0.16668667D-kOO 0.33333333D+00
-0.23800524D-01 0.142857 14D+00
-0.17857143D-01 0.12500000D+00
0.97222222D-01
-0.1388888OD-01 0.1 11111 11D+OO
-0.90900091D-02 0.90000001D-01
0.727272731)-01
0.18181818D-01
0.81818182D-01
-0.27272727D-01
-0.45454545D-01
0.54545455D-01
-0.36363636D-01
0.63636364D-01
(4
e5
0.55555556D-01
-0.44444444D-01
0.66666667D-01
-0.33333333D-01
0.77777778D-01
-0.22222222D-01
0.8888888OD-01
-0.11111111D-01 0.10000000D+00
-0.27777778D-01
(kf
95
-0.55555556D-01
0.60444444D-01
-0.41666667D-01
0.83333333D-01
(a)
e4
0.71428571D-01
-0.53571429D-01
0.80285714D-01
-0.35714286D-01
0.107 14286D-kOO
(k)
94
-0.7142857lD-01
0.05238005D-01
-0.47610048D-01
0.110047821)+00
(k) e3
0.10000000D+00
-0.66666667D-01
0.13333333D-kOO
-0.33333333D-01 0.16866867D+00
(k)
93
-0.10000000D+00
0.15000000D~00
-0.50000000D-01 0.20000000D+00
(k)
e2
O.I6666667D+OO
-0.83333333D--01 0.25000000D+00
56-digit binary reprerent ation
(k)
0.50000000D~00
...
Y t r
.4
n
Table 4.6. gd - table f o r gik’ =
Vak+1
v ak
_ I _
k
with ak =
C -j!1
k = 1,2, ...
j-0
4 - d i g i t binary representation
0.333333343+00
-0.833333433-01 0.25000000E~00
0.16666666E+00
0.14285713E-l-00
-0.17857134E-01 0.12500000E~00
-0.13888896E-01 0.111 llllOE+OO
0.888887723-01
-0.11111103E--01 0.100000003+00
-0.454895578-01
0.545119843-01
-0.36346123E-01
0.836478143-01
-0.272756813-01
0.72725780L?3-01
-0.1 81818443-01
0.818182303-01
-0.0090007OE-02 0.80809084E-01
-0.222224373-01
(k)
e5
0.555540553-01
-0.444474973-01
0.666824538-01
-0,333314843-01
0.77779025E-01
(4
95
-0.55559866B--01
0.694431893:-01
-0.416667613-01
0.83332881E-01
-0.277776273-01
0.972223 13E-01
(k)
e4
0.714307433-01
-0.535723123-01
0.89285374E-01
-0.357142693-01
0.107142813$00
(i)
94
-0.714278823-01
0.052385073-01
-0.47619178E-01
O.l1804755E+OO
-0.238085223-01
(4
e3
0.99Q89800E-01
-0.666665883-01
0.13333338E+00
-0.333333483-01
93
-0.10000007E+00
O.l4999998E+OO
-0.49999989E-01 0.20000000E+00
(k)
0.16666670E+00
IV.1.2. The qd-algorithm
208
Table 4.7.
lim a; = 2.71828182845904..
i-+co
j-0
56-digit binary repreaentation (0)
i
€2 i
1.000000000000000D+00 1.000000000000000D+OO 3.000000000000000D+OO 3.000000000000000Df00 2.7142857142857 14D+00 2.714285714285714D+OO 2.7 18309859154930D+00 2.718309859 154930D+00 2.718281718281718D+OO 2.718281718281718D+OO 2.7 18281828735696D+00 2.718281828735696D+OO 2.7182818284585631)+00 2.7 18281828458564D+00 2.718281828459046D+OO 2.718281828459046D+OO 2.7 18281828459045D+00 2.718281828459045D+OO Table 4.8. 1
a i = j -C 0,
lim ai = 2.718281828.. .
i-
00
24-digit binary represent ation 1
0 1
2 3 4 5
I 1.0000000E+00 3.0000000E+00 2.7142858E+00 2.7183099E+00 2.7 1828183+00 2.7182815Ei-00
1.0000000E+00 3.0000000E+00 2.7 I42858E+OO 2.71830993+00 2.71828183+00 2.71828223+00
.
IV.1.3. The dgorithm of Bulirach-Stoer
209
1.3. The algorithm of Bulirsch-Stoer. Let
(2,)
i
e be ~ a convergent sequence of points with
limi-+w z i = z When using extrapolation techniques to accelerate the convergence of we compute a sequence {bi};,N with
{U;},.~N
b; = limz-.z gi(z) where g i ( z ) is determined by the interpolation conditions gi(zi) = a j
j
0 , . . ., 9: with i
2
0
The point z is called the extrapolation point. Often polynomials are used for the interpolating functions 8 ; . We consider here the case of rational interpolation where we shall use Stoer’s recursive scheme given in section 3 of chapter 111. Using the notations of chapter I11 and writing
we first generalize the formulas (3.12.):
and
with pO,o(s) ( i ) = a, and qo,o(z) (i) = 1. We can easily see taking into account the interpolation properties of r(J+l)
m,n- 1
rk!n and
that
(4.1a .)
IV.1.3. The algorithm of Bulirech-Stoer
210
and analogously that
(4.1b .)
If aiT;ll
# 0, then we can write
or
Analogously
If we use the interpolating functions
IV.f.3.The algorithm of Bulirach-Stoer
and write 8 (; )' = r@,(z) for rn
with
= t-tb(z)= a j and
211
+ n = &,then
8-1 ( j1
= 0 for j = 0 , 1 , 2 , . . , and
Let ua compare t he algorithm of Bulirsch-Stoer with the €-algorithm, the qdalgorithm and the method of Neville-extrapolation based on polynomial interpolation (here bi = rri(z)). Input is the sequence {ai);,N = { ~ / r n ) ; ~ N
with lim;+a a; = 0. For the methods based on interpolation we use
z=
lim x i = O
i- oa
IV.i.3. The algorithm of Bulirsch-Stoer
212
The results are listed in table 4.9. Computations were performed in single precision arithmetic.
Table 4.9. a.=
1
-__
&Ti Bulirsch-
i ~
qd
Neville
Stoer
epsilon
Qi _____--_-- b i
s(0)
€ i(0)
C;(1)
0
1.00000000
1.00000000
1.00000000
1.00000000
1.00000000
2
0.5773502e
0.26964909
0.24118084
0.47414386
0.47414389
4
0.4 472 13 59
0.16024503
0.11 8 2 0 6 8 3
0.3095461 1
0.3095461 1
6
0.3779e450
0.11422185
0.07439047
0.22860453
0.22960511
8
0.33333334
3.08880618
0.06251375
0.18284784
0.18285738
10
0.30151 1 3 5
0.07302472
0.05196425
0.15728275
0,15730451
12
0.27735010
0.0617 1018
0.04684320
0.14 1 I5 857
0.14100026
14
0.25819880
0.11071037
0.04458576
0.12546575
0.12571160
0.04921301
0.10151753
0.10491328
0.34646210
16
0.24253564
I8
0.22941573
-2.86674523
0.04912448
0.10688034
0.10763902
20
0.2 1821788
-89.92565918
0.05485482
0.09624152
0.10446023
I V . l . 4 . The p-algorithm 1.4. The
2 13
p-algorithm.
Another technique is based on the construction of interpolating continued fractions of the form
If we take for g;(z) the ith convergent
and take z = 00 as extrapolation point, then we can put for i even:
6; = lim 30’2
g i ( z ) = do
+ d:! + . . . + di
Now the sequence b i can be calculated using reciprocal differences. Hence we can set up the following scheme:
k =0,1,2,. .. Zk-1. k = 1,2,3,... tk,l = a k - ak-1 Xk - Z k - j __ tk,j = tk-l,j-2 + tk,j-l - tk-1,j-1 tk,O
= ak
Zk
k=j,j+l
where in fact, in the notation of chapter 111, t k $ 3' = p [ Z k - j ,
. . .,zk]
Then clearly for a even
.,
bi = ~ [ z o , . .z;]
The values
tk,j
= ti,i
are ordered as in the following table
,...
and j = 2 , 3
,...
I V . l . 4 . The p-algorithm
214
to ,o t l ,o
tl,l
t2 ,o
t2,l
t2,2
t 3 ,O
t3,l
t3,2
with t k , o = a k and t k , - i = 0 for k = 0 , 1 , 2 , . . .. if we use { z ; } ~ ~as N interpolation points for the algorithm of Bulirsch and Stoer and {Z;};~IN with
as interpolation points for the p-algorithm, then for even 4:
(see problem (3)). Remark t ha t the computation of t i , i takes a smaller effort than th a t of 8y). For more properties of the p-algorithm we refer to [4]. We illustrate the influence with some numerical examples. In the tables 4.10. of the choice of the {z~},.~N and 4.11. several { b i } i E m were constructed, respectively for {ai);,N
={ l / m } i E N
and {a;)ieN = {I/('
+l ) } a ~
All computations were performed in single precision arithmetic. In each table the first sequence {b;}i,m is constructed with xi = &,the second with z i = i, the third with Z; = i2 and the fourth with xi = e i . The choice of the 2; greatly influences the convergence behaviour of the resulting sequence {b;};,N. We see that we do not necessarily obtain better results when the 2 ; converge faster to infinity. More information on this matter can be found in [8].
215
I V . l . 4 . The p-algorithm
Table 4.10.
___ i
-t ,. b; = t i , ; 6. b; 2 ti,; 1 - *,I 2; = i 2 z’- -2; = i a; I -_ _ ~ -
6
b-=t., x; = e' 1
1,t
0
1.00000000
1.00000000
1.00000000
1.oooooooo
1.00000000
2
0.57735026
7.07813832
0.24 11 8 0 8 8
0.50412308
0.48505624
4
0.44721358
0.02822608
0.11828871
0.30880488
0.30424682
6
0.37786450
-0.00126043
0.07431557
0.22015348
0.34178661
8
0.33333334
-0.00124686
0.06307473
0.17057084
0.30704348
10
0.30151135
-0.004163Q2
0.05177462
0.1 4054570
0.28143203
12
0.27735010
0.04758168
0.12608825
0.26141584
14
0.25818889
0.04435266
0.1 1 4 5 1 1 1 3
0.24517243
16
0.24253564
0.00017822
0.04088086
0.10530381
0.23163427
18
0.2284 1 5 7 3
0.00026065
0.05084635
0.00724047
0.22012024
0.21821788
0.00000440
0.03507820
0.08777263
0.21018887
20
0.00010506 -0.00047068
Table 4.11. a; =
i __
b*.
ai
0
1.00000000
2
0.33333334
4
0.20000000
6
0.14285716
8
0 . 1 1111111
10
0.08080008
=z
1 i+l
t.. 1,'
xi = -
b; = t %. *. bi = t i , ; .2 2; = i -~ --xi = 9,
1.00000000
b . = t *. , i
xi = e'
1.00000000
1.00000000
0.00000006
0.25000003
0.2401564 1
0.00000162
0.00000003
0.08080016
0.1543827 8
0.00000038
0.00000002
0.04545450
0.31652598
-0 .OOOOOO 15
0.02703313
0.08417003
0.00000003
0.01783462
0.07815881
-2.41421294
-0.00000042 0.00000018
1.00000000
IV.1.5. The multivariate e-algorithm
216
In [49] a study was made of some nonlinear convergence accelerating methods. The +algorithm was found to be the most effective one in many cases. Therefore we now generalize the t-algorithm for the convergence acceleration of multidimensional tables. 1.5. The multivariate t-algorithm. can he considered as a table with single entry. For its convergence acceleration we constructed the univariate function
A sequence {ai},.,N
c 00
f(z) =
VUiX'
i=O
with ai = 0 for i
< 0 and calculated
diagonal Pad6 approximants evaluated a t
x = 1. Let us now first consider a table (aj,k
}j,ken
with double entry and with A = limj,k.+OOaj,k = limj-w(limk+OO ajk) = limk-+w(limj+m ajk). To accelerate its convergence we introduce
We will again construct a sequence ( b i } i E ~ that will converge faster t o A in some cases [ll].To this end we now calculate bivariate Pad6 approximants for f ( z , y) and evaluate them at ( 5 , y) = ( 1 , l ) . If we denote by Ut =
j+k==t
Qjk
= at,O
+ at-1,l + . . + al,L-1 + a0,L
and start the c-algorithm with th e partial Bums of f( 1 , l )
IV.1.5. The multivariate 6-algorithm
approximants ri,i(z,y) evaluated a t
then the bivariate diagonal l’adi: (0) (e, y) = (1,l) are given by t2i . Hence we put bi
217
(0) = e2i
Let US generalize the idea for a table with multiple entry define
{ ~ j ~ . , , j ~ } ~ ~ We , , . . , ~ , ~ ~
W
f ( z l , .. .,zk) =
jl
, .. . ,j,=O
Vajl.,.j, Zil.
..zc
with
It is easy to prove that
Again multivariate Pad6 approxinlants for f ( z l , . . ., zk) can be calculated and evaluated a t (z,, . . .,zk)= ( 1 , .. ., 1) via the t-algorithm. Sirice
where now
the
6f)
for f(z1,. . .,Zk) are given by
IV.1.5. The multivariate €-algorithm
218
We illustrate this type of convergence acceleration with a numerical example. Suppose one wants to calculate the integral of a function G ( z ~ , ..,. Zk) on a bounded closed domain D of Rk. Let D = [0,1] x . . . x [0, I] for the sake of simplicity. The table { a i l . . . i k } j ~ , , , , , j can k ~ ~ be obtained for instance by subdividing the interval [O,1] in the lthdirection (t! = 1 , . . ., k) into 2jl intervals of equal length h~ = 2-Je(jt = 0, 1 , 2 , . . .). Using the midpoint-rule one can then substitute approximations
to calculate the ail...j, . We restrict oursclves again t o the case of two variables. With hl = 2-j and hp = 2-k we get
The column c0(4( l = 0, 1, . . .) in the 6-table given by
was also used by Gens [21] to start the €-algorithm for the approximate calculation of multidimensional integrals by means of extrapolation methods. He preferred this method to six other methods because of its simplicity and general use of fewer integrand evaluations. For
we have
[1
1
lim
j,k-cce
ajk =
1
dzdy = 2 In 2 = 1.386294361119891...
In table 4.12. one can find the a i , slowly converging to the exact value of the integral because of the singularity of the integrand in the origin. The 6; = r,,;(l, 1) converge much faster. For the calculation of the b i we need the c~ ( l = 0 , . . ., 2t).
IV.1.5. The multivariate €-algorithm
Hence b; should be compared with an information, i.e. j -tk = 2i.
ajk
which uses the same amount of
Table 4.12.
d
aii
1 2 3 4
1.166666606667 1.269017019048 1.325743700744 1.355532404415
219
bi = t2; (0)
1.330294906166 1.396395820203 1.386872037696 1.386308917778
220
IV.2.1. Iterative methods brrsed on P o d i approzimation
$2. Nonlinear equations.
Suppose we want to find a root xu of the nonlinear equation
f (4 = 0
Here the function f may be real- or complex-valued. If f is now replaced by a local approximation then a zero of th at local approximation could be considered as an approximation for x*. Methods based on this reasoning are called direct. One could also consider the inverse function g of f in a neighbourhood of the origin, if it exists, and replace g by a local approximation. Then a n evaluation of this local approximation in 0 could be considered as an approximation for x* since g(0)
.=z *
Methods using this technique are called inverse. 2.1. Iterative methods baaed on Pad6 approzimation.
Let
2;
be an approximation for the root xu of ri(x)
and let
Pi
= -(x) cli
be the Pad6 approximant of order ( m ,n) for f in z i . Then the next approximation x;+1 is calculated such that Pi(Zi+l)
=0
In case p ; ( z ) is linear ( m = 1) the value zi+l is uniquely determined. It is clear that this is t o be preferred for the sake of simplicity. A well-known method obtained in this way is Newton’s method (rn = 1, n = 0) which can be derived as follows. The Taylor series expansion for f(z) at zi is given by f(5) = f ( 5 ; ) f f ’ ( Z i ) ( Z
- 5 i ) + f‘ I ( x i ) ( 5 - X i ) 2 2
Hence the Pad6 approximant of order (1,O) for f a t
xi
+. . *
(4.2.)
equals
and we obtain (4.3.)
IV.2.1. Iterative methods boaed on Pad4 approzimation
221
Another famous method is Halley’s method based on the use of the Pad4 a p proximant of order (1,l) for f at X i : (4.4.)
Since the iterative procedures (4.3.) and (4.4.) only use information in the point z; to calculate the next iteration point z;+1 they are called one-point. The order of convergence of these iterative methods is given in the next theorem.
Theorem 4.4. If { z ~ } ~converges ~ N t o a simple root z* of f and if r i ( z ) is normal for every i then
For the proof we refer to [52]. Similar results were given in [IS], [43] and 1441. So the order is a t least rn n 1 and it depends only on the sum of m and ra. Consequently the order is not changed when Pad6 approximants lying on a n ascending diagonal in the Pad4 table are used. Iterative methods resulting from the use of (m,n) Pad6 approximants with n > 0 can be interesting because the asymptotic error constant C* may be smaller than when n = 0 [43]. The iterative procedures (4.3) and (4.4) can also be derived as inverse methods (see problem (5)). Let us apply Newton’s and Halley’s method t o the solution of
+ +
The root Z* = 0. We use xo = 0.09 as initial point. The next iteration steps can be found in table 4.13. Computations were performed in double precision accuracy (56-digit binary arithmetic).
222
IV.8.1. Iterative methods baaed on P a d 4 approzimation
Table 4.13.
i
Newton
Halley
Zi
2;
0
O.QOOOOOO0D-01
1
0.80997588D-01
2
0.6559965QD-01
3
0.43022008D-01
4
~.18500311D-01
5
0.34212128D-02
6
0.11703452D-03
7
0.13697026D -06
8
0.18760851D-12
9
0.35197080D-24
10
0.00000000D+00
0.90000000D -0 1 -0.40647645D-02 0.35821646D-06 -0.24514986D-18 0.00000000D+00
It is obvious that a method based on the use of (m,n) Pad4 approximants for f with n > 0 gives better results here: the function f has a singularity at z = 0.1. Observe that in the Newton-iteration zg is a good initial point in the sense that from there on quadratic convergence is guaranteed: 1zi+1
- 2.1 = 1zi+l I N c ' J z ~- z * I 2 for i 2
o
with C’ = 10. For the Halley iteration we clearly have cubic convergence from X I on. The formulas (4.3.) and (4.4.)can also be generalized for the solution of a system of nonlinear equations
which we shall write as
IV.d.1. Iterative methods based on Pad6 approzimation
223
Newton's method can then be expressed as [45]
where F’(xr),. . . ,xr)) is the Jacobiau matrix of first partial derivatives evaluated at (zy),. .. ,zt))with
Let us now introduce the abbreviations Fi = F ( z f ) ,. . . ,xk(i))
F , ! = F ( z (4 , ,... ,zk (4) To generalize Halley's method we first rewrite (4.4.) as
Then for the solution of a system of equations it becomes [I41
224
IV.2.1. Iterative method8 based o n P a d 6 appton'mdion
where the division and the square are performed componentwise and F " ( x 1 , ... ,x k ) is the hypermatrix of second partial derivatives given by
~~
a 2f k a2f k ... a2f k axkaxs ... axlaxk
dxlax2
.
.
I
az:
which we have t o multiply twice with the vector --F,!-'Fi. This multiplication is performed as follows. The hypermatrix F " ( x 1 , . . . , Z k ) is a row of k matrices, each le x k. If we use the usual matrix-vector nlultiplication for each element in the row we obtain a2fi
c
a2fk
c
k
asl ax;Yi ... i= 1 axlax;
Yi
...
k
i= 1
jin) vk
In [14] is proved that the iterative procedure (4.6.) actually results from the use of multivariate Pad6 approximants of order ( 1 , l ) for the inverse operator of ~ ( x l. ., . , x k ) at ( x v ) , . . . , x g ) ).
IV.8.1. Iterative method8 baaed on Pod6 approzimation
225
To illustrate the use of the formulas (4.5.) and (4.6.) we shall now solve the nonlinear system
{ (-
f 1 (‘J
Y) = -=+I/ = e-=-y
f&,y)
- 0.1
=0
- 0.1 = 0
which has a simple root at 12.0.1)) = (2.302585092994O46.. 0.
.
As initial point we take ( d 0 ) , g ( ’ ) ) = (4.3,2.0). In table 4.14.one finds the consecutive iteration steps of Newton’s and Halley’s method. Again Halley’s method behaves much better than the polynomial method of Newton. Here the inverse operator G of the system of equations F has a singularity near to the origin and this singularity causes trouble if we get close to it. For
we can write -0.5(ln(0.1 0.5(ln(0.1
+ u) + ln(O.l + v)) + u) - ln(O.l + v))
With (do), y(O)) = (4.3,2.0) the value do)= f2(z(O),y(O)) is close t o -0.1 which is close t o the singularity of G. For the computation of th e Pad6 approximants involved in all these methods the €-algorithm can be used. Another iterative procedure for the solution of a system of nonlinear equations based on the e-algorithm but without the evaluation of derivatives can be found in [5]. Since it does not result from approximating the multivariate nonlinear problem by a multivariate rational function, we do not discuss it here.
a
J9 0
Ne ton
0.43000000D+01
H Iev
,A4 0.20000000D+01
N N
0.43000000D+01
0.20000000D+01 0.5705728dD-kOO
1
-0.22427305D+02
-0.24720886D+02
0.287Q8400D+01
2
-0.21927303D+02
-0.24228888D-k.02
0.22495475D+Ol
-0.52625816D-01
3
-0.21427303D+02
-0.23720888D-kO2
0.23018229D+Ol
-0.44Q01947D-02
4
-0.20827303D+02
-0.23228888D+02
0.23025841D+01
-0.57154737D-05
5
-0.20427303D+02
-0.2272Q888D-kO2
0.23025851D-l-01
-0.Q7680305D--11
6
-0.1 QQ27303DS-02
-0.22228888D-kO2
0.23025851D+Ol
-0.17438130D--16
7
-0.18427303D+02
-0.21720888D-bO2
8
-0.18Q27303D+02
-0.2122Q888D-tO2
Q
0.18427303D+02
-0.20728888D)+02
10
-0.17827303D)+02
-0.20229888D-kO2
11
-0.17427303D+02
-0.19720888D-kO2
12
-0.16927303D+02
-0.10228888D+02
13
-0.16427303D+02
-0.18728888D+02
14
-0.15Q27303D+02
-0.18228888D-I-02
15
-0.15427303D+02
-0.17729888D+02
16
-0.14927303D+02
-0.17229888D+02
17
-0.14427303D+02
-0.16729888D+02
18
-0.13027303D+02
-0.1622Q888D-tO2
18
-0.13427303D+O2
-O.l572Q888D+02
20
-0,12!327303D+02
-0.15228888D-tO2
rn
y(4
3 e
m
J
?:
;r a
R e
Iv.8.B. Iterative methods based on rational interpolation
227
3.2.Iterative methods baa e d o n rut i o n a1 int erpol at i o n. Let
with pi and 9, respectively of degree m and n, be such that in an approximation z i for the root z * of f
fr)(zi-j)
with m that
= f(')(Zi-j)
I!
= 0 , . . .,8j - 1
(4.7.)
-+ n + 1 = C'tz08 ~ Then . the next iteration step z;+] is computed such Pi(zi4-1)
=0
For the calculation of zi+l we now use information in more than one previous point. Hence such methods are called multipoint. Their order of convergence can be calculated as follows. Theorem 4.5.
If
(Z;};~Nconverges to a simple root z* of f and f(n+n+l)(z) continuous in a neighbourhood of z* with
with n
> 0 is
where f ( k ) ( z *= ) 0 if k < 0, then the order of the iterative method based on the use of r,(z) satisfying (4.7.) is the unique positive root of the polynomial .i+1
- 8ozj- B1zj-l
The proof can be found in 1521.
- .. .- 8j = 0
228
IV.2.2.Iterative methods baaed on rational interpolation
If we restrict ourselves to the case 8[
=8
l = 0 , . , ., j
then it is interesting t o note that t h e unique positive root of
!=O
increases with j but is bounded above by 8 + 1 [53 pp. 46-52]. As a conclusion we may say t h a t the use of large j is not recommendable. We give some examples. Take m = 1, n = 1, 8 = I and j = 2. Then z;+1 is given by
The order of this method is 1.84, which is already very close to 8 + 1 = 2. Take m = 1, n = 1, 80 = 2, 81 = 1 and j = 1 . Then x;+1 is given by
The order of this procedure is 2.41. The ease m = I , n = 0,8 = 1 and j = 1 reduces to the secant method with order I .82. Let us again calculate the root of
with initial points close t o the singularity in x = 0.1. The successive iteration steps computed in double precision (56-digit binary arithmetic) are shown in table 4.15.
IV.2.2. Iterative method8 based on rational interpolation
229
Table 4.15.
(4.8.)
(4.0.)
secant method
Xi
zi
Xi
d
0.80000000D-01
0
1 2 3
0.90OOOO0OD-01
0.80000000D-01
0.80000000D-01
0.85000000D-01
O.QO0OOOOOD-01
O.QOOO00OOD-01
-0.15847802D-03 0.15131452D-06 -0.51538368D-13
4
0.20630843D-24
5
0.00000000D+00
- 0.17324153D-03
0.71095917D-01
0 . 4 6 6 2 1186D-10
0.64701421D-01
-0.6139231 1 D-25
0.46636684D-01
0.00000000D+00
0.30206020D-01 0.14080364D-01
6
0.425123451)-02
7
0.59843882D-03
8
0.28 4 30 08 1 D -0 4
0
0.15223576D-06
10
0.38727362D--10
11
0.58956894D-16
12
0.22832450D-25
By means of the multivariate Newton-Pad4 approximants introduced in section 8.2. of chapter 111 the previous formulas can be generalized for the solution of systems of nonlinear equations. We use the same notations as in chapter 111 and as in the previous section. For each of the multivariate functions f,(q,.. .,zk) with j = 1 , . ..,k we choose
D = N = ((0,. . . , O ) , ( l , O , . . .,O ) , ( 0 , 1 , 0 , .. . , O ) , . .., (0,...,o, 1)) H={(2,0,
..., 0 ) , ( 0 , 2 , 0,...10),..., (0,..., 0,2)) 2 INk
c Nk
IV. 2.2. Iterative methods baaed on rational interpolation
230
Here the interpolationset N U H expresses interpolation conditions in the points
zp)). . ., ($',
..
.)
Remark that this set of interpolation points is constructed from only three successive iteration points. The numerator of
with possible coalescence of points, is then given by
NO ,...,O ( 21, . . ,zk) c02,00,. ..(00
coo, ...,00,02
where
Nl,O
,...,0 (21> .
1
* 3
c12,00,...,00
0
zk
. . . NO ,...,0,1(21, . . ... 0 ...
*
coo, ...,00,12
>
zk)
IV.B.8. Iterative methods based o n rational interpolation
231
The values e,,tl,...,,Ltk are multivariate divided differences with possible coalescence of points. Remark that this formula is only valid if the set H provides a sys( i + l ), . . ., zk (i+l)) tern of linearly independent equations. The next iterationstep (21 is then constructed such that
p i k ( z y l ) ., .
.J
Z k( i + l ) ) = 0
For k = 1 and without coalescence of points this procedure coincides with the iterative method (4.8.). With k = 2 and without coalescence of points we obtain a bivariate generalixation of (4.8.). Let us use this technique t o solve the sgslem
{
e--l+Y e-Z-9
= 0.1 = 0.1
with initial points (3.2, -0.95), (3.4, -1.15) and (3.3, -1.00). The numerical results computed in double precision (56-digit binary arithmetic) are displayed in table 4.16. The simple root is (2.302585092994046.. . ,O.). In this way we can also derive a discretized Newton method in which the partial derivatives of the Jacobian matrix are approximated by difference quotients
N = H = ( ( 0 , ...,Q ) , ( l , O , ..., 0 ),,..J(Ol...lO,l)}
D
= ( ( 0 , . . .,n))
If we call this matrix of difference quotients AFi, then the next iterate is computed by meana of
we take the same system of equations and the same but fewer initial pointa as above. The consecutive iteration steps computed in double preciaion (56-digit binary arithmetic) can now be found in table 4.17.
As an example
232
IV.2.2. Iterative method8 baaed on rational interpolation Table 4.16.
i
~
y")
~-
0.320000OOD+Ol
-0.Q50000OOD-l-00
0.34000000D-k01
-0.11500000Df01
0
0.33000000D-k01
-0.10000000D~01
1
0.25249070D-kOl
-0.22072875D-kOO
2
0.22618832D-tOl
3
0.231278OQD)+Ol
-0.101644QOD-01 -0.51269373D-03
0.41 9 7 l Q 4 4 D - 0 1
4
0.23030978D-kOl
5
0.23025801D-kOl
0.40675854D-05
6
0.28025851Df01
-0.2568682QD-08
7
0.23025851 D+Ol
-0.1277881 6D-13
8
0.23025851Df01
-0.1 1350Q32D-16
Table 4.17.
i 0 1
~~
0.34000000D-k01
-0.11500000D-k01
0.33000000D-b0 1
-O.lO0OOOOOD+Ol
-0.29618530D-kOO
0.21743833Df01 0.20884Q33D-kOl
2
0.32743183D-l-01
3
0 . 2 2 1 1 4 2 1 1 Di-0 1
-0.84011352D+01
4
0.3651533RDt-0 1
-0.72149651D-kOl
5
-0.17Q00083D-kO4
0.208541 11 D i - 0 4
divergence
IV.2.3. Iteratave method3 wing continued fractions
233
The rational method is again giving better results. Now the initial points are
such t h at u = f l ( z , y ) is close t o -0.1 which is precisely a singularity of the inverse operator for the considered system of nonlinear equations. For a more stable variant of thc diacretized Newton method we refer to [24]. 2.3. Iterative method8 wing continued fractions.
If ri(z) is t h e rational interpolant of order (m, 1) for $(z), satisfying 1 ri(zw)) = --(&-L))
L = 0,.. . , m + 1
f
then ri(z) can be written in the form rj(z) = do
+ d l ( z - z(i-m-')
+ ldm(z
z(i-m-
-
) + . . . + dm-l(Z - z(i-m-1) 1)
). . .(. - z(i-2J)l
t
+I
2
-
1.. .(z- z ( i - 3 ) )
&-q
&+I
The coefficients d , ( j = 0 , 1 , . . ., m ) are divided differences while dm+l is an inverse difference. The root of ,!;(z) can be considered as an approximation for the root z* of f . SO z(i+l)
-2
(i-1)
- dm+l
This method can be compared with methods based on the use of rational interpolants of order (1, rn) for f(z).
2.4. The qd-algorithm. First we s t at e an important analytical property of the qd-scheme. To this end we introduce the following nokition. For the function f(s) given by its Taylor series development f(.) = co c 1 2 c 2 22
+
we define the H ankel-det erminant s
+
+
234
IV.Z.4. The gd-algorithm
We now call the series f(z)ultimately k-normal if for every R with 0 5 n 5 k there exists an integer M,, such t h a t for m 2 M,,the determinant Ifm,,,is nonzero.
Theorem 4.8. Let the meromorphic function f(z) be given by its Taylor series development in the disk B(0,r) = {z E C 121 < r ) and let the poles wi of f in B(0, r ) be numbered such that
1
0
< Iw] L
Iwzl
I ... < r
each pole occuring as many times as indicated by its order. If the series f(z) is ultimately k-normal for some integer k > 0, then the qd-scheme associated with this series has the following properties: a) for each n such that 0 < n 5 k and such t h a t Iwn-~l < (tun( < Iw,,+l(, where wg = 0 and, if f has exactly k poles, wk+l = 00, we have Iim m-+m
b) for each n such t h a t 0
1
qLm) = Wn
and such that Iw,(
Iim
m-cc
< Iw,,+lI,
we have
eim) = O
The proof can be found in [29 pp. 612-6131. As a consequence of this theorem the qd-algorithm is an ingenious tool t o determine the zeros of polynomials and entire functions. For if p(z) is a polynomial then the zeros of p ( z ) are the poles of the meromorphic function f(z) = b(z). AIIYq-column corresponding to a simple pole of isolated modulus would tend to the reciprocal value of that pole. It would be flanked by e-columns that tend t o zero. If moreover f ( z ) is rational the last e-column would be zero. Unfortunately the qd-scheme as generated in section 3 of chapter TI is numerically unstable. Rounding errors play an important role due to the divisions by small e-values. However, the rhombus rules (2.6.) and (2.7.) defining the scheme may be rearranged as follows:
IV.2.4. The qd-algorithm
235
In this way quotients are formed only with the quantities q p ) which do not tend t o zero and the qd-scheme is generated row by row. This is called the progressive form of the qd-algorithm. The problem of getting started is solved when we also calculate q y t l ) and elk+’) for negative values of k, i.e. if we calculate the extended qd-scheme as described in section 3 of chapter TI. In [XY pp.641-6541 it is also shown how to deal with the case in which several poles have the same modulus. To conclude this section we illustrate the preceeding theorem with the values qim) and el") for f(z) =
sin x 0.1 - x
which is an ultimately 1-normal function. Clearly with wo = 0,
w1
= 0.1 and
1 lirn qIm) = 10 = Wl
m-w
lim
m-oo
elm)
=0
Table 4.18. 0.10000000000000E+02
e
0 . Q Q S333 33 3 3333 3 E 4-0 1
ey)
=
0 . 1 ~ ~ ~ 6 6 6 ~ 6 6 ~ 6 ~ 7 ~ - n i
o.iooooooooooonnrc-t-o2
el")
=
0.83472454091016E-05
0.10000008347245E~+02
t:
=
-n.s3172454ooioie~-o5
0.100000000000001.;+02
$1
=
-0.108743778QQi25E-08
O.QOOOOQQQQS0l26E+O1
ey)
=
0.10000000000000E+02
e?
O . ~ O ~ O O O O O O O O O O ~ E + ~t:?) ~ 0.10n00000000000E-t02
-0.16686866666687E-01
-
=
0.1Q8743776QQ125E-08 0.27600144302181E-12 -0.276001 44382181 E-12
IV.R.5. The generalized qd-algorithm
238
2.5. The generalized qd-algorithm. The function f(s)can also be given by its Newton series 00
f(z)=
C f[zo, -
7
~ i ] ~ i ( z )
i=O
where the ~ [ z o. ., . , q] are divided differences with possible coalescence of points and where
ll j=l (z - zj-1) i
Bi(s) =
If we define the generalized Hankel-determinants
with Rm,o= 1 then we call the Newton series ultimately k-normal if for every n with 0 5 n 5 k there exists an integer Adn such that for m 1 Mn the determinant Hm,nis non-zero.
Theorem 4.7. Let the sequence of interpolation points ( 5 0 , 5 1 , 5 2 , .. .>be asymptotic to the sequence { ZO, zl,.. . , z,, 20, z1, . . . ,z i , 2 0 , z1, . . . ,z j , . . .} in the sense that
lim
k-oo
sk(j+l)+i =
d = 0 , . . . ,j
z;
Let the function f(s) be meromorphic in
..
~ ( 2 0 , ., Z j , r )
= {z E C
I
- zo)(z - 21).
. .( 5 - zj)l 5 r)
and analytic in the sequence of points { s ~ } ~and ~ Nlet the poles w i of B(z0,. . . , z j , r ) be numbered such that for wi
we have
=
I(Wi
- 20)
0 < w1 Iwp
1 . .
(Wi
- 2j)l
s . .. < r
f in
IV.2.5. The generalized qd-algorithm
237
where the poles are counted with their multiplicities. If the Newton series is ultimately k-normal for some integer k > 0, then the generalized qd-scheme associated with this series has the following properties : a) for each n such that 0 < n 5 k and such that w,-~ < w, < w,+1 where wo = 0 and, if f has exactly k poles, wk+l = r , we have
b) for each ra such that 0 < n 5 k and such that w , < w,+1, we have
For the proof we refer to [9].
238
IV.3.1. The we of Pad4 approzirnants
53. Initial value problems. Consider the following first order ordinary differential equation: dv
- =j(z,y)
dz
for
2
E [a,bj
(4.10.)
with y(a) = yo. When we solve (4.10.) numerically, we do not look for an explicit formula giving y ( z ) as a function of z but we content ourselves with the knowledge of y(zi) a t several points zi in [a, 61. If we subdivide the interval [a,b ] , k
U Izi-1 ,z i ]
[a, 61 =
i= 1
where zi =a
+ ik
i = 0,. . ., k
with
b-a h=--
k
for k
>0
then we can calculate approximations yi+l for y ( z i + ~ )by constructing local approximations for the solution y ( z ) of (4.10.) a t the point 2 ; . We restrict ourselves now t o methods based on the use of nonlinear approximations.
3.1. The use of Pad6 approximantcr. Let us try the following technique. If 8 i ( z ) is the Pad6 approximant of a certain order for y(z) a t z i then we can put Yi+l
=ai(Zi+l)
which is an approximation for y ( z i + l ) . For the calculation of ai(z) we would need the Taylor series expansion of y(z) at z;, in other words
Since the exact value of y(z;) is not known itself, but only approximately by y i , this Taylor series development is not known and hence this technique cannot be applied. However, we can proceed as follows. Consider the power series
IV.3.1. The u8e of Pad6 approzirnants
239
Let r i ( z ) be the Pad6 approximant of order ( m , n )for this power series. If we put z = zi+l, in other words z - x i = h , we obtain 8' = 0,..., k - 1
yi+r = r i ( z i + I )
Hence we can write ui+l
= Y i + hg(zir Yip h)
i = 0,. . .,k - 1
(4.11.)
where g is determined by r i . Such a technique uses only the value of zi and y i t o determine yi+1. Consequently such methods are called one-step methods. Moreover (4.11.) is an explicit method for t he calculation of v i + l . It is called a method of order p if the Taylor series expansion for g(z,y,h) satisfies Y ( z i + l ) - u ( z i ) - hg(ziJ Y ( z i ) ,h) = 0(hp+')
Clearly (4.11.) is a method of order (m + n) if r i ( z ) is a normal Pad6 approximant. The convergence of (4.11.) follows if g(z, y , h) satisfies the conditions of the following classical theorem [30 p. 711.
Theorem 4.8. Let the function g(zyy , h) be continuous and let there exist a constant L such that
g('J YJ
= f(’,
Y)
is a necessary and sufficient condition for the convergence of the method (4.11.), meaning that for fixed z E [ a ,b ] .
2 40
IV.3.1. The we of Podi approzarnants
From the fact that t i is a Pad4 approximant it follows th a t the relation g(z,y, 0) = f(z,y) is always satisfied (see problem (8)). The case n = 0 results in the classical Taylor series method for the problem (4.10.), If we take rn = n = 1 we get (4.12.)
If y(z) is a rational function itself, then using (4.11.) we get the exact solution Yi+l = Y(z;+I)
at least theoretically, if the degrees of numerator and denominator of r i ( z ) are chosen in a n appropriate way. Techniques based on the use of Pad4 approximants can be interesting if we consider stiff differential equations, i.e. if has a large negative real part [ZO]. An example of such a problem is the equation dY = xy dz
(4.13.)
-
with Re(1) large and negative. Since the exact solution of (4.13.) is
we have lim y ( z ) = lim
z-00
2-00
2% =o
and we want our approximations y; to behave in the same way. Dahlquist [15] defined a method to be A-stable if it yields a numerical solution of (4.13.) with Re(1) < 0 which tends to zero as i -+ 00 for any fixed positive h. He also proved t hat there are no A-stable explicit linear one-step methods. Take for instance the method of Euler (rn = 1, n = 0): Y i + l = ~i
+ hf ( z i j y i )
Y i + l = (1
+ hX)Yi
We get
= (1
+ hX)'+1 yo
IV.3.2. The
t~t: of
rational anterpolanta
241
Clearly lim yi+l = yo Iim (I i-
i-+W
00
+ /A)'+' = o
only if
)I
+ hX( < 1
So for large negative X the steplength h has to be intolerably small before acceptable accuracy is obtained. lo practice h is so small t h a t round-off errors and computation time become critical. The problem is t o develop methods t h a t do not restrict the stepsize for slability reasons. If (4.11.) results from the use of t h e Pad6 approximant of order (rn,m), (m, rn 1) or (m,m + 2) then one gets an A-stable method [16]. This can be seen as follows. If f ( z , y) = Xy then r,(z) is the Pad6 approximant for the power series
+
Hence Y i + l = rm,rt(hA) ~i
with h = ~ i + - lz;, where rm+(x) is the Pad6 approximant of order (rn,n)for e 2 . A-stability now follows from the following theorem.
Theorem 4.9. If m = n or m = n - 1 or m = n - 2 then the Pad6 approximant of order (m,n) for ez satisfies
For the proof we refer to [3, 181. 3.2. The uae of rational interpolants.
It is clear t h a t if the interpolation conditions are spread over several points, then the computation of yi+l will need several xi-^ and yi-t(t = 0, 1, . . .). Such methods are called multistep methods. Let r i ( z ) be the rational Hermite iuterpolant of order (rn,n) satisfying
242
IV.8.2. The me of rational interpolants
where (j
+ I)(R + 1) = rn + n + 1
Here
Then an approximation for y(zi+l) can be computed by putting
This is a nonlinear explicit multistep method. A two-step formula is obtained for instance by putting j = 1, m = 2, n = 1, and 8 = 1 and using theorem 3.17.:
where f, = j(x;,y;) and fi-1 = f(z;-1, yi-1). We can also derive implicit methods which require an approximation the caIculation of y;+1 itself, by demanding
v;+l
for
where (j
+ 2)(8 + 1) - 1 = m + n +
For m = 1 = n, j = 0 and
8
1
= 1 we get the formula
(4.15 .)
IV.3.3. Predictor-corrector method8
243
For more information concerning such techniques we refer t o [36]. Remark th a t multistep methods are never selfstarting. Both explicit and implicit ( j + 1)- step methods are of the form Yi+l =
j
C
acyi-c
+ h g ( z i + l , . . ., z i - j l
yi+lJ..
-J
~ i - j lh)
L=O
and they have order p if the Taylor series expansion of g satisfies i
y(zi+t) -
C
aty(zi-t)
- hg(zi+l
J
.
‘1
zi-j] y ( z i + l ) j , ..,y(zi-j), h ) = 0(hp+l)
L= 0
Hence (4.14.) is a third-order method if the starting values are third-order and (4.15.) is second-order. When applied t o a stiff differential equation one should keep in mind that linear multistep methods are not A-stable if their order is greater than two. The following result is helpful if af&’yl is real and negative. We know th a t we can write for problem (4.13.) y(z) = v(2i-j)
e(~-~i-j)X
with Re(X) large and negative. €leiice it is interesting to take a closer look a t rational Hermitc interpolants for exp(z) in some real and negative interpolation points and also in 0. Theorem 4.10.
5 m and 8qi 5 n is such th a t ri( t ) (0) = 1 = exp(*)(O)for 0 5 t 5 rn + n - t2 with e 5 m
If ri(z) = pi(z)/qi(z) with api a)
b) ri(tt) = exp(&) for & then lri(z)l
< 1 if
m
< 0, 1 5 k’ 5 j
5 n and
with &
# &, 1 5 !# k 5 j
z is real and negative.
For the proof we refer to [32]. 3.3. Predict0 r-correct or methode. For a solution of the initial value problem (4.10.) we have y’(z) = f ( z , y(z)) for every z in [a, b]
244
IV.S.4. Numeracal results
If we integrate this equation on the interval [zi-i, z i + t ] with j , 8 2 0 we get PZitl
Now f can be replaced by an interpolating function, through the points j(zi-1 ,yj-l)), . . ., which is easily integrated. ( z 8 ,f(z;, yi)), If t! = 1 we get p r e d i c t o r - m e t h o d s because they are explicit. If 8 = 0 we get corrector-methods because the value of y; is needed for the computation of y;. These implicit formulas can be used t o update an estimate of yfz;)iteratively. W h e n f is replaced by an interpolating polynomial we get the well-known methods of Adarns-Bashforth ( j = 0 and !. = 1) and Adams-Moulton ( 3 = 1 and 6 = 0). When f is replaced by an intc3rpolating rational function we can get nonliriear formulas of predictor or corrector type. Sincc the rational interpolant must be integrated, it is not recommendable to choose rational functions with a denominator of high degree.
3.4. Nu me rical r es ult 8 . Let u s compare two Taylor series methods ( m= 2 and m = 3) with the explicit method (4.12.) for t h e solution of the equation y‘ = I
+ y2
for z
2o
Y(0) = 1
The theoretical solution is g = tg(z -t 7r/4). We take the steplength h = 0.05. As can be seen in table 4.19. the second order rational method gives even better results t h a n t h e third order Taylor series method, a fact which can be explained by the singularity of the solution y(z) at z = 7 . To illustrate A-stability we will compare Euler’s method ( m= 1,n = 0) with the formulas (4.12.))(4.14.) and (4.15.) for the equation y’ = -25y
for z E [0,I]
Y(0) = 1
The solution is known to be y = exp(-25z). For the results in table 4.20. we chose the steplength h = 0.1. As expected Eiiler’s solution blows up while formulas based on the use of a “diagonal” entry of the Pad6 table or the rational Hermite interpolation table for t h e exponential decay quite rapidly. Similar results would have been obtained if “superdiagonal” entries of the Pad6 table were used. Remark that formula
Table 4.19. exact solution
+7)
Taylor series
Pad6 approximant
Taylor series
m=3
i
zi
1
0.05
0.1 10536D+01
0.1 105OOD+Ol
0.110526D+01
0.110533D+01
2
0.10
0.122305D+01
O.l22210D+Ol
0.122284D-kOl
0.122299D)+Ol
3
0.15
0.135600D-kOl
0.135449D+Ol
0.135573D+01
0.135598D)+01
4
0.20
0.150850D+01
0.1 5 0 5 8 2 D t 0 1
0.150795D+01
0.150831D+Ol
5
0.25
0.1 6858oD+oi
0.168150D-kOl
O.l68500D+Ol
0.188547D+01
6
0.30
0.188577D$-01
O.l88886D+Ol
0.188462D-kOl
0.189522D-t-01
7
0.35
0.214875D-kOl
0.213884D-kOl
0.214811D+Ol
0.214883D-kOl
8
0.40
0.246496D-kOl
0.244751D-kOl
0.24626lD+Ol
0.246335D-kOl
9
0.45
0.286888D+01
0.28398OD+Ol
0.286543D-I-01
0.286584D-I-01
10
0.50
0.340822D+Ol
0.335737D-t-01
0.340298D-t-01
0.340248D-kO 1
11
0.55
0.4 16936D4-01
0.407388D-kOl
0.416097D+Ol
0.415703D+01
12
0.60
0.533186D-t-01
0.513307D+Ol
0.531720D+Ol
0.5301 31D-t-01
13
0.65
0.7 3 4 0 4 4 D t 0 1
0.685144D-t-01
0.731087D+01
0.724568D-kOl
14
0.70
0.1 16814D+02
0.1 00697D-kO2
0.1160 1QDi-02
0.112431D-kO2
15
0.75
0.282383D-kO2
0 .I 77676D-kO2
0.277486D-i-02
0.232131D-kO2
tg(zi
m=2
m=l=n
*CD e
5 e
Table 4.20.
exact solution exp(-25si)
i 1
0.1
0.820850D-01
2
0.2
0.6737851)-02
3
0.3
0.553084D-03
4
0.4
0.453QQQD-04
5
0.5
0.372665D-05
6
0.6
0.305902D-06
7
0.7
0.251 100D-07
8
0.8
0.2061 15D-08
Q
0.Q
0.16QlQOD-OB
10
1.0
0.13887QD--10
m=l=n explicit one-step
Euler -0.150OOOD+01 0.225000D+01 -0.337500D+01 0.506250D-l-01 -0.758375D+01 0.113Q06D+02 -0.17085QD)+02 0.25628OD-l-02 -0.384434D)+02 0.576650D4-02
1
-0.111111D+OO 0.123457D-01 -0.137174D-02 0.152416D-03 -0.189351 D-04 0.188168D-05 -0.200075D-06 0.2323061)-07 -0.258117D-08 0.286797D-00
rn = 2,n = I
explicit multisteD
rn=l=n
implicit one-step
0.820850D-01
0.123047D+00
0.840728D-01
0.151 407D-01
-0.542641D-01
0.186302D-02
-0.4Q474OD-kOO
0.22023QD-03
-0.252173D+00
0.282073D-04
0.314055D-kOO
0.347083D-05
0.102175D+O1
0.427077D-06
0.16Ql37D-kOO
0.525507D-07
-0.lQl5Q6D-I-00
0.646622D-08
-0.6Q8115D-kOO
0.7Q5651D-OQ
IV.3.5. Systems of first order ordinary differential equatiow
247
(4.14.) based on a "subdiagonal" rational approximation is surely not producing an A-stable method. To obtain the results of table 4.20. by means of (4.14.) a second starting value yl was necessary. We took yl = exp(-2.5) = y(z1). For (4.15.) the expression f(zi,y;) = -25y; was substituted and y;+l was solved
from the quadratic equation. All these schemes can be coupled to mesh refinement and the use of extrapolation methods. If an asymptotic error expansion of y(zi) in powers of h exists, then the convergence of the sequence of approximations for y(z;), witb z i fixed, obtained by letting the stepsize decrease, ran be accelerated by the use of techniques described in section 1 [23]. 3.5. Systems of first order ordinary differential equations.
Nonlinear techniques can also be used to solve a system of first order ordinary differential equations
dzj = f j ( X , 2 1 , 2 2 , . . . , Z t ) dX
where the values z,,! and the functions approaches are possible. If we introduce vectors
fj
j = 1, ..., k are given for j = I , . . .,k. Several
then one method is t o approximate the solution componentwise using similar techniques as in the preceding sections. So for instance (4.12.) becomes
Y ( z ; + ~N) Y;:+l where
and
= Y;
Y;) ] + h [*F(zi,2F2(Zi, Y;)- hF'(xil yi)
(4.16.)
248
IV.3.5. Systems of first order ordinary differential equation8
and the addition and multiplication of vectors is performed componentwise. In other words (4.16.) is equivalent with
For more information on such techniques we refer to [56,35, 381. Another approach is not based on componentwise approximation of the solution vector Y ( s )but is more vectorial in nature. Examples of such methods and a discussion of their properties is given in [57, 7 , 27, 131. The nonlinear techniques introduced here c a n also be used to solve higher order ordinary differential equations and boundary value problems because these can be rewritten as systems of first order ordinary differential equations. Again same of the nonlinear techniques prove to be especially useful if we are dealing with stiff problems. A system of differential equations
dY -= F ( z , Y ) dx Y ( a ) = Yo with
is called stiff if the matrix
IV.3.5. System8 of first order ordinary differential equations
249
has eigenvalues with small and large negative real part.
{
Consider for example
+
= 9 9 8 ( ~x )~ 199822 (z)
Zl(0j = 1
- - - -99921(2) - 19992z(x)
22(0)
=0
The solution is
so that again
lim z I ( x ) = 0 = lim
z 4 m
DO’%
z2(x)
where both z1 and a2 contain fast and slow decaying components. For a discussion of stiff problems we refer to [20 pp.209-2221.
IV.4. Numerical integration
250
§4. Numerical integration.
s,
b
Consider I = j ( z ) d z . Many methods to calculate approximate values for I are based on replacing j by a function which can easily be integrated. The classical Newton-Cotes formulas are obtained in this way: j is replaced by an interpolating polynomial and hence I is approximated by a linear combination of function values. In some cases the values of the derivatives of j ( x ) are also taken into consideration and then linear combinations of the values of f ( z ) and its derivatives a t certain points are formed t o approximate the value I of the integral. This is for instance the case if polynomial Hermite interpolation is used. In many cases the linear methods for approximating I give good results. There are however situations, for example if f has singularities, for which linear methods are unsatisfactory. So one could t r y t o replace f by a rational function r and consider
[
r(x)dz
as an approximation for I . But rational functions are not that easy to integrate unless the poles of r are known and the partial fraction decomposition can be formed. Hence we use another technique. Let us put
Then
1 = y(b) If f is Riemann integrable on [a,b ] , then y is continuous on [a, b ] . If f is continuous on [a, 61, then y is differentiable on [a,b] with y’(z) = j ( z ) and y(a) = 0
So I can be considered as the solution of a n initial value problem and hence the techniques from the previous section can be used. We group them in different categories.
I V . 4 . l . Methoda w i n g Pad6 approximantd
251
4.1. Methods using Pad6 approzirnants. Let us partition the interval [a, 61 with steplength h = ( b - a ) / k and write Z; = a
+ ih
a = 0, ...,k
and (4.17.)
s,”’
f(t)dt. where y ; approximates y ( z i ) = If r; is the Pad6 approximant of order ( m ,n ) for t l , ; ( h )then we can put
and consider y k as an approximation for I . In this way Yi+l
which means
= yi
+ hg(zi, h)
i = 0, ...,k - 1
(4.18.)
l;’’
f ( t ) d t N hg(z;,h)
If m = 1 = ra we can easily read from (4.12.) t h a t (4.18.) results in (4.19.) Formulas like (4.18.) use derivatives of f(z) and are nonlinear if n > 0. From the previous section we know that (4.18.) is exact, in other words that yk = I , if y ( z ) is a rational function with numerator of degree m and denominator of degree ra. For n = 0 formula (4.18.) is exact if y ( z ) is a polynomial of degree m, i.e. f ( z ) is a polynomial of degree m - 1. The obtained integration rule is then said t o be of order m - 1. The convergence of formula (4.18.) is described in the following theorem which is only a reformulation of theorem 4.8.
252
IV.4.2. Method8 wing rational interpolanta
Theorem 4.11. Let y; be defined by (4.18.). Then lim
h-0
yi(h) = y(z) for fixed z E [a, b]
====i(h)
if and only if g(z,O) = f(z). Instead of (4.17.) one can also write t l , i ( h )= T/i
+ lata,i(h)
with
and compute Pad6 approximants r , of a certain order for tz,;(h).Let us take
m = 1 = n . If we define
Yi+i
= yi
+ hr;(h)
then we get
4.2. Method8 using rational interpolanta.
If we again proceed as in the section on initial value problems we can construct nonlinear methods using information in more than one point. Since these methods are not self starting but need more than one starting value their use is somewhat limited and rather unpractical. An example of such a procedure is the following. Let ti(.) be the rational Hermite interpolant of order ( 2 , l ) satisfying
and let
IV.4.3. Methods without the evaluation of derivatives
253
Then we know from formula (4.14.) t h a t
with f; = f(zi) and f i - 1 = f ( z i - I ) . Often rational interpolants are preferred t o Pad6 approximants for the solution of numerical problems because the use of derivatives of f is avoided. As mentioned, a drawback here is the necessity of more starting values. Another way t o eliminate the use of derivatives, now without the need of more starting values, is the following.
4.3. Methods without the evaluation of derivatives. One can replace derivatives of j ( z ) in formula (4.18.) by linear combinations of function values of f ( z ) without disturbing the order of the integration rule. To illustrate this procedure we consider the case m = 1 = n. Then 1591
We will compute constants a,p and 7 such t h a t for
we have gfZ,
h) - t ( s ,h) = O(hrn+%) = O(h2)
For ~ ; + i= ~i
+ h t ( z ; ,h)
this would imply y(si+l) - Yi+l = ~(hm+m+l)= O ( P )
Condition (4.20.) is satisfied when a+B=2 B7 = -1
(4.20.)
254
IV.4.4. Numerical resulta for singular integranda
In other words, for 7 # 0, 27 a=---
p = -1
+1 7
-.
7
so
For 7 = 1 we get the integration rule
In this way we approximate (4.21.)
4.4. Numerical results for singular integranda We will now especially be interested in integrands regular in [a,b] but with a singularity close t o the interval of integration and on the other hand in integrands singular in Q or b . The problem of integrating a function with several singularities within [a, 61 can always be reduced t o the computation of a sum of integrals with endpoint singularities. If f ( z ) is singular in 6 then the value of the integral is defined by
and is assumed t o exist. We shall compare formula (4.19.) with Simpson’s rule ( m = 2, n = 0 ) and with a (2k)-point Gaussian quadrature rule that isolates the singularity in the weight function. If f(t) can be written as w ( t ) h ( t ) where w ( t ) contains the endpoint singularity of f(t) and h(t) is regular then the approximation
I
wlh(tl) + . . .
+ 402kh(t2k)
does not involve function evaluations in singular points. We use a (2t)-point formula because on [zi,2;+1] for i = 0 , . . ., k - 1 both (4.19.) and Simpson’s rule
IV.4.4. Numerical reaulta for aingular integranda
255
need two function evaluations. Since / is singular in b = zk, we take f ( z k ) = 0 in Simpson's rule which means that the singularity is ignored. In (4.19.) the singularity of f is no problem since f and f’ are only evaluated in 2 0 , . . ., xk--1 [58]. Our first numerical example is
d t = 3.04964677.. with a singularity in
t = In 3 = I .09861228. . . and the second example is
+
2(1_ _ ~ t)sint _ _ _ _ _cost d t = 2
We shall also compare the different integration rules for the calculation of e'dt = 1.71828182. . .
which has a smooth integrand. Because of the second example the weight function w ( t ) in the Gaussian quadrature rule was taken t o be W ( t )=
1
J1-t
.___
All the computations were performed in double precision arithmetic and the double precision values for the weights w i (i = 1 , . . .,2k) and the nodes ti (d = 1 , . . .,2k) were taken from [I pp. 916-9191. Remark t ha t the nonlinear techniques behave better than the linear techniques in case of singular integrands. However, for smooth integrands such as in table 4.23., the classical linear methods give better results th a n the nonlinear techniques. Also, if the singularity can be isolated in the weight function such as in table 4.22., Gaussian quadrature rules are very accurate. In general, little accuracy is gained by using nonlinear techniques if other methods are available for the type of integrand considered [51].
256
IV.4.4. Numerical reeults f o r singular integranda
As in the previous section all these schemes can be coupled to mesh refinement and extrapolation.
Table 4.21. 1
I=J
et
o (3 - et)2
k = 16
dt = 3.04964677.. .
Gaussian
Simpson's
formula
quadrature
rule
(4.19 .)
3.1734660 3.0773081 3.0564777
3.28067 65 3.0841993 3.0531052
3.1006213 3.0573798 3.0510553
Table 4.22.
I = J 2(1 - t ) sin t -tc o s t dt = 2 0 G - j
k=4 k=8 k = 16
Gaussian
Simpson's
formula
quadrature
rule
(4.19.)
2.0000000 2.0000000 2.0000000
1.7504074 1 A267581 1.8786406
1.9033557 1.8797832 1.go48831
Table 4.23. 1
I = 1etdt = 1.71828182.. . 0
k=4 k=8 k = 18
Gaussian
Simpson's
formula
quadrature
rule
(4.18.)
1.7205746 1.7204038 1.7188196
1.7182842 1.7182820 1.7182818
1.7284991 1.7206677 1.7188592
IV.5. Partial diflerential equatiom
257
§5. Partial differential equations. Nonlinear techniques are not frequently used for the solution of partial differential equations. We will describe here a method based on the use of Pad6 approximants t o solve the heat conduction equation which is a linear problem. For other illustrations we refer to [54, 48, 17, 22, 281. All these techniques first discretize the problem so that thr: original partial differential equation is replaced by a system of equations which is nonlinear if the partial differential equation is. Another type of techniques which we do not consider here are methods which do not discretize the original problem but solve it iteratively by means of a procedure in which subsequent iteration steps are differentiable functions [13]. Linear techniques of this type arc recommended for linear problems and nonlinear techniques can be used for nonlinear partial differential equations. Let us now concentrate on the heat conduction equation. Suppose we want t o find a solution u ( z , t )of the linear problem
a+, t ) - a%+, t ) ~
at
.~
82
~
a
t>O
(4.22 .)
with u(2,O) = u ( 2 )
u(0,t ) = a
4 4 t)= P The domain [a, 61 with
x
[0,oo) is replaced by a rectangular grid of points (z;, ti)
zi = a + ih ti = j A t
i = 0,.. .,k + 1
j = 0, 1 , 2 , . . .
where
and At is the discretization step for the time variable. We first deal with the discretization in the space variable z and introduce u;(t) = u ( z ; , t ) for
t20
Then using central differences, (4.22.) can for instance be reduced t o
258
I V.5. Part id digere nt ial
equ at ion8
for t > 0 and i = 1, . . . , k. In general we c a n write
(4.23.)
where A is a real symmetric positive definite k x k matrix and depends on the chosen approximation for the operator d 2 / d s 2 . If we introduce the notations
then the exact solution of (4.23.) is U ( t ) = eWtAV where e-tA is defined by
Using the discretization in the time variable t we can also write (4.24.)
I V.5. Partial d ifler e nt ial equ at i o ns
259
If rm,.,(t)= p ( t ) / q ( t ) is the Pad6 approximant of order (n,n) for e-t then (4.24.) can be approximated by
U ( t + A t ) = [q(At.A)]-' [P(At.A)] U ( t )
(4.25 .)
Varga proved t ha t for n 2 m this is an unconditionally stable method [ 5 5 ] , meaning t ha t initial rounding errors remain within reasonable bounds as the computation proceeds independent of the stepsize used. If rn = 1 and n = 0 then (4.25.) means
U(t + At) = ( I
- AtA)U(t)
or equivalently if A = (&.!)kXk
ui(ti+l) = u i ( t j ) - ~t
k
C aitut(tj)
.!=I
which is the well-known explicit method to solve (4.22.). The solution a t level t = ti+l is determined from the solution at t = t , , For nt = 1 = n we obtain from (4.25.)
or equivalently
which is the method of Crank-Nicholson [50]. The operator a2/ax2is replaced by the mean of a n approximation for the partial second derivative a t level t = and the same approximation at t = t i .
IV.6.1. Kernel8 of finite rank
260 56. Integral equations.
As in the previous section we shall discuss linear equations for which the use of nonlinear methods is recommendable. Those interested in nonlinear integral equations are referred to [lo, 121 where methods are indicated for their solution. If the integral equation is rewritten as a differential equation then techniques developed for the solution of initial value problems can also be used. We restrict ourselves t o the discussion of an inhomogeneous Fredholm integral equation of the second kind (the unknown function f appears once outside the integral sign and once behind it):
f ( 4-
J'." K ( z ,y)f(v)dy
= g(z)
xE
1% bl
(4.28.)
Here the kernel K ( z , y ) and the inhomogeneous right hand side g(z) are given real-valued continuous functions. Fredholm equations reduce to Volterra integral equations if the kernel K ( z , y ) vanishes for g > z which produces a variable integration limit.
6.1.K e r n e l 8 of finite rank. Formally the solution of (4.26.) can be written as a series. P u t
and
If we define
and
then (4.26.) reduces to
IV.6.1. Kernel8 of finite rank
261
The series (4.27.) which is a power series in X, is called the Neumann series of the equation (4.26.). Convergence of the Neurnann series for certain values of X depends on the properties of the kernel K ( z ,y). If K ( z ,y ) is bounded by IWz,y)l
<M
for (z,Y) E
b,b]
X [a, b ]
then clearly the series (4.27.) converges uniformly t o f(s)in [a, b] if 1
1x1 < ---___ M ( b - a) If Pad6 approximants in the variable X are constructed for (4.27.) then they may have a larger convergence region than the series itself. Especially interesting is the case t ha t the kernel is degenerate, in other words k
~ ( zY), =
C Xi(Z)Yi(y)
i= 1
with ( X i } and {Y;} each linearly independent sets of functions. Such a kernel is also said to be of finite rank k. Let us try to determine f(z) in this case. If we put
then (4.26.) can be written as (4.28.) Multiply (4.28.) by Yj and integrate to get
or equivalently
IV.6.2. Completely continuoua kernela
262
(4.29 .)
which is the determinant of the coefficient matrix of system (4.29.), then D(X) is a polynomial in X of degree at most k. In case D(X) # 0 the solution of (4.29.) is given by
with
Dij
the minor of the ( i , j ) t h element in D(X). SO (4.28.) can be written as
which is a rational function in X of degree a t most k both in numerator and denominator. Since the series development of (4.30.) coincides with the Neumann series (4.27.) we know that the solution f(z) is equal to the Pad6 approximant of order (k,k) for the Neumann series. So in this case the Pad6 approximant is the exact sum of the series because the sum is a rational function [2 p. 1761. 6.8. Completely continuoua kernels.
The equation (4.26.) can be rewritten
EUI
(I-XK)f=g where the linear operators I and K are defined by
(4.31.)
263
IV.6.2. Completely continuoua kernela
We shall consider square-integrable functions f with
Suppose now t ha t { f;}iEN is a bounded converging sequence of functions. Then the operator K is said t o be completely continuous if for all bounded { f ; ) , ~ the sequence { K fi}iENcontains s subsequence converging t o some function h ( s ) with
- h(2)l2ds -+ 0 as i -+
llKfi - hll = \i[lKfi(Z)
00
A basic property of completely continuous transformations is th a t they can be uniformly approximated by transformations of finite rank. Thus there is a n infinite sequence { K ; } ; Lof~ kernels of finite rank i such th a t
where lim
i-+w
t;
=0
When a completely continuous kernel K is replaced by Ki then the solution f; of
( I - Mi)!; = g is given by the Pad4 approximant ri,i of order (i,;), as explained in the previous section. In case K is completely contiuuous the exact solution f of (4.31.)is a meromorphic function of X [31 p. 311 arid
lim f, = f
i-
00
2 64
Iv.6.2. Completely continuous kernels
in any compact set of the X-plane except at the finite number of X-poles of f [8].Since
we consequently have
lim
ri,i
1-m
=f
in any compact set of the X-plane except at the finite number of X-poles of f.Thus the solution f is the limit of a sequence of diagonal Pad6 approximants. This result is not very useful since each approximant in the sequence is derived from a different Neumann series, with kernel K ; and not with kernel K . However a similar result exists for the sequence r i , i derived from the Neumann series (4.32.) i= 1
where the operator K' is defined by
Theorem 4.12. If ri,i is the Pad4 approximant of order (d, i) t o the Neumann series (4.32.) of the integral equation (4.28.) with completely continuous kernel K then the solution f is given by lim
i-oc
ri,i
=f
in any compact set of the X-plane except a t the finite number of poles of f and at limit points of poles of ri,i as i -+ 00. More information on this subject can be found in [2 pp. 178-1821.
I V . Problems
265
Problems. Show t ha t with
EF’= an for n 2 0,
which coincides with Aitken’s A2-process to accelerate the convergence of a sequence. Give a n algorithm similar to the one in section 1.4. for the calculation of t,,i, but now based on the use of inverse differences instead of reciprocal differences, Show that if the algorithm of Bulirsch and Stoer is used with the interpolation points zi (lim,+m s; = 0) and the p-algorithm is used with the interpolation points z: = 1,s; (1imi4- 2;. = oo),then t i ; = B~( 0 ) . Compare the amount of additions and multiplications performed in the c-algorithm and the qd-algorithm when used for convergence acceleration. Derive the formulas (4.3.) and (4.4.) using inverse interpolation instead of direct interpolation. Derive the formulas (4.8.) and (4.9.) based on the use of rational interpolants. in section 2.3. for successive values of Organize the computation of d,+l d such t ha t a mininum number of operations is involved. Prove that one-step explicit methods for the solution of initial value problems based on the use of Pad6 approximants are convergent if g(z, g, h) given by (4.11.) is continuous and satisfies
IV. Problem8
266
(9)
Check the formulas (4.12.) and (4.14.).
(10)
Write down formula (4.16.) for the solution of
- 99821 (2)+ 199822( Z) - - - -9992+) (11)
- 199922(.)
ZI(0) = 1 ZZ(0) = 0
Construct a nonlinear numerical integration rule based on the use of Pad4 approximants of order ( 2 , l ) . Afterwards eliminate the use of derivatives as explained in section 4.3.
I V . Remnrka
267
Remarks. Nonlinear methods can also be used for the solution of other numerical problems. We refer for instance to [40] where the solution of linear systems of equations is treated, t o [19] for analytic continuation, to [34] for numerical differentiation and to (421 for the inversion of Laplace transforms.
An important link with the theory of numerical linear algebra is through QR-factorization. Rutishauser [47] proved th a t the determination of the eigenvalues of a square malrix A can be reduced to the determination of the poles of a rational functiom .f built from th e given matrix. In this way decomposition techniques for A to compute its eigenvalues are related t o the qd-algorithm when used to compute poles of meromorphic functions. Univariate and multivariate continued fractions and rational functions are also often used t o approximate given functions. For univariate examples we refer t o (33, 421. The bivariate Beta function is a popular multivariate example because it has numerous singularities in a quite regular pattern. For numerical results we refer to [89, 26, 121.
As a conclusion we can say t h a t every linear method has its nonlinear analogue. In case linear methods are inaccurate or divergent, it is recommendable t o use a similar nonlinear technique. The price we have to pay for the ability of the nonlinear method to cope with the singularities is the programming difficulty to avoid division by small numbers.
2 68 References.
1
1) Abramowite M. and Steguo Handbook o Mathematical functions. Dover publications, New York, 1968.
[
21 Baker G. and Gamrnel J. The Pad6 approximant in theoretical physics. Academic Press, New York, 1970.
[
31 33aker G. and Graves-Morris P. Pad4 approximants: basic theory. Encyclopedia of Mathematics and its applications: vol 13. AddisonWesley, Reading, 1981.
\
41 Brezinski C. Algorithnies d’accklkration de la convergence. Editions Tecbnip, Paris, 1978.
[
51 Rrezinski C. Application de 1’6-algorithme B la r6solution des syst6mes non linhaires. C. R. Acad. Sci. Paris Sbr. A 271, 1970, 1174-1177.
[
61 Bulirsch R . and Stoer J. Fehlerabschatzungen und Extrapolation mit rationalen Funktionen bei Verfahren vom Richardson-Typus. Numer. Math. 6, 1964, 413-427.
[
On the stability of rational Runge 71 Calvo M. and Mar Quemada M. Kutta methods. J. Comput. Appl. Math. 8, 1982, 289-292.
[
81 Chisholm J. Solution of linear integral equations using Pad6 Approximants. J. Math. Phys. 4, 1963, 1506-1510.
[
91 Claessens G. Convergence of multipoint Pad6 approximants. Report 77-26, University of Antwerp, 1977.
[ lo] Clarysse 7 .
Rational predictor-corrector methods for nonlinear Volterra integral equations of the second kind. In [60], 278-294.
[ 111 Cuyt A.
Accelerating the convergence of a table with multiple entry. Numer. Math. 41, 1983, 281-286.
[ 121 Cuyt A.
Pad6 approximants for operators: theory and applications. Lecture Notes in Mathematics 1065, Springer Verlag, Berlin, 1984.
[ 131 Cuyt A.
Pad6 approximants in operator theory for the solution of nonlinear differential and integral equations. Comput. Math. Appl. 6, 1982, 445-466.
Iv. Reference8
269
Abstract Pad6 approximants for 141 Cuyt A. and Van der Cruyssen 1’. the solution of a system of nonlinear equations. Comput. Math. Appl. 9, 1983, 617-624.
[ 151 Dahiquist G.
A special stability problem for linear multistep methods. BIT 3, 1963, 27-43.
[ 161 E61e B.
A-stable methods and Pad4 approximations t o the exponential. SIAM J. Math. Anal. 4, 1973, 671-680.
[ 171 Fairweather G.
A note on the eficient implementation of certain Pad6 methods for linear parabolic problems. BIT 18, 1978, 100-108.
[ 181 Frame J .
The solution of equations by continued fractions. Amer. Math. Monthly 60, 1953, 293-305.
[ 191 Gammei J.
Continuation of functions beyond natural boundaries. Rocky Mountain J. Math. 4, 1974, 203-206.
[ 201 Gear C .
Numerical initial value problems in ordinary differential equations. Prentice-Hall Inc., New Yersey, 1971.
[ 211 Genz A .
The approximate calculation of multidimensional integrals using extrapolation methods. Ph. 1). in Appl. Math., University of Kent, 1975.
[ 221 Gerber P. and Miranker W .
Nonlinear difference schemes for linear partial differential equations. Computing 11, 1973, 197-212.
[ 231 Gragg W.
On extrapolation algorithms for ordinary initial value problems. SLAM J. Numer. Anal. 2, 1965, 384-403.
[ 241 Gragg W. and Stewart G .
A stable variant of the secant method for solving nonlinear equations. SIAM J. Numer. Anal. 13, 1976, 889-903.
[ 251 Graves-Morris P .
Pad6 approximants and their applications. Academic
Press, London, 1973.
[ 261 Graves-Morris P. , Hughes Jones R . and Makiuson G . The calculation of some rational approximants in two variables. J. Inst. Math. Appl. 13, 1974, 311-320.
[ 271 Hairer E.
Nonlinear stability of RAT, an explicit rational Runge-Kutta method. BIT 19, 1979, 540-542.
IV. References
270
1
Pad6 approximants, fractional step methods 281 Hall C. and Porsching T. and Navier-Stokes discretizations. SLAM J . Numer. Anal. 17, 1980, 840851.
[ 291 Henrici P.
Applied and computational complex analysis: vol. 1. John Wiley, New York, 1974.
[ 301 Henrici P.
Discrete variable methods in ordinary differential equations. .John Wiley, New York, 1962.
[ 311 Hoheisef G.
Integral equations. Ungar, New York, 1968.
[ 32) Iserles A.
On the generalized Pad6 approximations to the exponential function. SLAM J. Math. Anal. 16, 1979, 631-636.
[ 33J Kogbethnts E .
Generation of elementary functions. In [46], 7-35.
[ 341 Kopal Z.
Operational methods in numerical analysis based on rational approximation. In 1371, 25-43.
[ 351 Lam bert J.
Computational methods in ordinary differential equations. John Wiley, London, 1973.
A method for the numerical solution of [ 361 Larnbert J. and S6aw B. g' = f ( z , y) based on a self-adjusting non-polynomial interpolant. Math. Comp. 20, 1966, 11-20.
[ 371 Langer R.
On numerical approximation. University of Wisconsin Press, Madison, 1959.
[ 381 Lee D.and Preiser S.
A class of nonlinear multistep A-stable numerical methods for solving stiff differential equations. Internat. J. Comput. Math. 4, 1978, 43-52.
I 391 Levin D.
On accelerating the convergence of infinite double series and integrals. Math. Comp. 35, 1980, 1331-1345.
[ 401 Lindskog G.
The continued fraction methods for the solution of systems of linear equations. BIT 22, 1982, 519-527.
[ 411 Longman I.
Use of Pad4 table for approximate Laplace transform inversion. In [25], 131-134.
The special functions and their approximations. Academic 421 Luke Y . Press, New York, 1969.
IV. References
271
[ 431 Merz G .
PadCsche Naherungsbruche und Iterationsverfahren hoherer Ordnung. Computing 3, 1968, 165-183.
[ 441 Nourein M.
Root determination by use of Pad6 approximants. BIT 16, 1976, 291- 297.
[ 451 Ortega J. and RheinboIdt W.
Iterative solution of nonlinear equations in several variables. Academic Press, New York, 1970.
[ 461 Ralston A. and Wilf S.
Mathematical methods for digital computers. John Wiley, New York, 1960.
Der Quotienten-Differenzen AIgorithmus. [ 471 Rutishauser H. Math. Phys. 5, 1954, 233-251.
z. h g e w .
On time-discretiEations for linear time-dependent partial differential equations. University of Manchester, Numer. Anal. Report 5, 1974.
[ 481 Siemieniuch J . and Gladwell I .
Numerical comparison of nonlinear convergence accelerators. Math. Comp. 38, 1982, 481-499.
[ 491 Smith D. and Ford W. [ 501 Smith G.
Numerical solution of partial differential equations. Oxford University Press, London, 1975.
[ 511 Squire W.
In defense of linear quadrature rules. Comput. Math. Appl. 7, 1981, 147-149.
[ 521 Tornheim L.
Convergence of multipoint iterative metho&. J . Assoc. Comput. Mach. 11, 1964, 210-220.
[ 531 n a u b J .
Iterative methods for the solution of equations. Prentice-Hall Inc., New York, 1964.
[ 541 Varga R.
Matrix iterative analysis. Prentice-Hall Inc., Englewood-
Cliffs, 1962.
[ 551 Varga R .
On high order stable implicit methods for solving parabolic partial differential equations. J . Math. Phys. 40, 1961, 220-231.
[ 561 Wambecq A.
Nonlinear methods in solving ordinary differential equations. J. Comput. Appl. Math. l , 1976, 27-33.
Iv. References
272
[ 571 Wambecq A.
Rational Runge-Kutta methods for solving systems of ordinary differential equations. Computing 20, 1978, 333-342.
[ 581 Werner H . and Wuytack L.
Nonlinear quadrature rules in the presence of a singularity. Comput. Math. Appl. 4, 1978, 237-245.
[ 591 Wuytack L.
Numerical integration by using nonlinear techniques. J. Comput. Appl. Math. 1, 1975, 267-272.
[ 601 Wuytack L.
Pad6 approximation and its applications. Lecture Notes in Mathematics 765, Springer, Berlin, 1979.
1
611 Wynn P . Acceleration techniques for iterated vector and matrix problems. Math. Comp. 16, 1962, 301-322.
[ 621 Wynn P. 175-195.
Singular rules for certain non-linear algorithms.
BIT 3, 1963,
273
SUBJECT INDEX A Aitken A2-process, 265 A-stable method, 240 asymptotic error constant, 221
B backward algorithm, 3 1 block-structure, 68 multivariate -, 109 Bulirsch-Stoer algorithm, 209
c continued fraction, 3 associated -, 18 branched -, 41, 169 contraction of -, 7 convergence of -, 20 convergent of -, 3 corresponding -, 17,76 equivalent -, 15 evaluation of -, 31 even part of --, 9, 80, 84 - expansion, 25 extension of--, 56 generalized -, 56 interpolating -, 138, 169 odd part. of -, 9,80,84 reduced -, 7 value of -, 3 convergence - acceleration, 197 - of continued fraction, 20 - in measure, 98 - of Pad6 approximants, 96 - of rational interpolants, 167 corrector method. 244
274
Subject in dez
covariance homografic --, 118, 119 reciprocal -, 118,119
D denominator partial -, 3 nth --, 3 determinant formulas, 85, 162, 181 direct method, 220 divided difference, 145, 163, 233 multivariate -, 175
E c- algorithm
univariate -, 92, 197 multivariate -, 114 €-table, 94,198 equivalence transformation, 6 Euler-Minding series, 5,46 explicit method 239 extrapolation, 209 - point, 209
F
forward algorithm, 32 Fredholm integral equation, 260
G Gauss quadrature, 254 generalized - polynomial, 190 - rational interpolant, 190 - continued fraction, 56
Subject in dez
H Hankel determinant, 233 Halley’s method, 221, 223
I
implicit method, 242 inverse difference, 139,233 bivariate -, 169 inverse method, 220 iteration one-point -, 221 multipoint -, 227
K kernel degenerate -, 261 - of finite rank, 261 completely continuous -,
L
lemniscate, 167
M modified convergent, 35 modifying factor, 35 multipoint method, 227 multistep method, 241
N
Neumann series, 261 Newton’s method, 220, 223 discretized -, 231
263
275
2 76
Subject indez
Newton-Coates formulas, 250 Newton-Pad6 approximation problem, 168 Newton-Pad4 approximant, 159 Newton series, 157 multivariate -, 236 numerator partial -, 3 nth 3 -.-)
0 one-point method, 221 one-step method, 221, 239 order - of iteration method, 221, 227 - of one-step method, 239 - of multistep met,hod, 243
P Pad6 aproximant, 65 convergence of -, 96 homographic covariance of -, 118 multivariate -, 109 normal -, 72,114, 119 reciprocal covariance of -, 118 Pad6 operator, 100 continuity of --, 100 Padk-type approximation, 120 predictor-corrector method, 244
Q qd- algorithm , 79 extended 81 generalized -, 148, 236 multivariate -, 116 progressive form of -, 236 -$
Subject indez
qd-table, 80 extended
-, 82
R rational interpolant, 130 generalized -, 190 multivariate -, 169 normal -, 135 table of -, 132 rational interpolation problem, 129 rational Hermite interpolant, 157 table of -, 159 rational Hermite interpolation problem, 156 reciprocal difference, 143 recursive algorithm, 90, 153, 209 recurrence relations, 4 p-algorithm, 213 S
sequence totally monotone -, 198 totally oscillating -, 199 series Stieltjes -, 29, 99 ultimately k-normal -, 234, 236 Stieltjes series, 29, 99 Stieltjes transform, 29 stiff differential equation, 240 Stoer’s method, 153
T
Thiele continued fraction expansion, 164 Thiele interpolating continued fraction, 19 branched -, 141. 170
277
Subject indez
2 78
U unconditionally stable method, 259 ultimately k-normal series, 234, 236
V Viscovat ov method of -, 16 multivariate algorithm of
W weights, 255
-,
49