(International Series of Numerical Mathematics

ISNM81:
International Series of Numerical Mathematics

Internationale Schriftenreihe zur Numerischen Mathematik
Serie internationale d' Analyse numerique
Vol. 81
Edited by
Ch. Blanc, Lausanne; R. Glowinski, Paris;
H. O. Kreiss, Pasadena; J. Todd, Pasadena
Springer Basel AG
Numerical Methods of
Approximation Theory,
Vol. 8
Workshop on Numerical Methods of Approximation Theory
Oberwolfach, September 28-0ctober 4, 1986
Numerische Methoden der

Approximationstheorie,
Band 8
Tagung iiber Numerische Methoden der Approximationstheorie
Oberwolfach, 28. September - 4. Oktober 1986
Edited by
Herausgegeben von
L. CoUatz, Hamburg
G. Meinardus, Mannheim
G. Niirnberger, Erlangen-Niimberg
1987 Springer Basel AG

Library of Congress Cataloging in Publication Data
(Revised for volume 8)
Numerische Methoden der Approximationstheorie.
(International series of numerical mathematics,
v. 16,26,30,42,52,59,67,81)
Selections from papers presented at the Tagung Uber
Numerische Methoden der Approximationstheorie held every
two years since 1971 at the Mathematisches Forschungs-
institut Oberwolfach, Schwarzwald.
Edited by L. Collatz and G. Meinardus.
Vol. 6-7 ed. by L. Collatz, G. Meinardus, H. Werner;
v.8- edited by L. Collatz, G. Meinardus, and
G. NUrnberger.
Vol. 5-8 also have English title: Numerical methods
of approximation theory.
Includes bibliographies.
1. Approximation theory. I. Collatz, L. (Lothar),
1910- II. Meinardus, GUnther, 1926-
III. Werner, Helmut, 1931- . IV. Tagung Uber
Numerische Methoden der Approximationstheorie.
V. Title: Numerical methods of approximation theory.
VI. Series: International series of numerical mathematics,
v. 16, etc.
QA221.N85 519'.4 72-363170
CIP-Kurztitelaufnahme der Deutscben Bibliothek

Numerical methods of approximation theory =
Numerische Methoden der Approximationstheorie /
Workshop on Numer. Methods of Approximation Theory.
- Basel ; Boston ; Stuttgart : Birkhauser
Auf d. Haupttitels. auch: Tagung Uber Numer.
Methoden d. Approximationstheorie
Bis Bd. 4 u.d.T.: Numerische Methoden der
Approximationstheorie
NE: Tagung Uber Numerische Methoden der
Approximationstheorie; PT
Vol. 8. Oberwolfach, 28. September - 4. Oktober
1986. - 1987.
(International series of numerical mathematics
Vol. 81)
NE:GT
All rights reserved.
No part of this publication may be reproduced, stored in a retrieval
system, or transmitted in any form or by any means, electronic,
mechanical, photocopying, recording or otherwise, without the prior
permission of the copyright owner.
© Springer Basel AG 1987

Urspriinglich erschienen bei Birkhiiuser Verlag Basel 1987
Softcover reprint of the hardcover 1st edition 1987
ISBN 978-3-0348-6657-6 ISBN 978-3-0348-6656-9 (eBook)

DOI 10.1007/978-3-0348-6656-9
Widmung
Die Herausgeber dieses symposiumbandes gedenken ihres Kollegen

und Freundes, Herrn
Professor Dr.rer.nat. Helmut Werner,
der so fruh von uns gegangen ist. Er war bis zur letzten Tagung
Uber Approximationstheorie im Mathematischen Forschungsinstitut
Oberwolfach Mitveranstalter und hat durch sein grosses Wissen
und sein Uberragendes Engagement diesen Tagungen zu einem be-
sonderen Ansehen verholfen. Viele Anregungen fUr die wissen-
schaftliche Arbeit jUngerer Mathematiker gehen auf seine Dis-
kussionsbeitrage im Oberwolfacher Institut zurUck. Hervorge-
hob en werden solI weiter sein stetes Hinwirken auf den Einsatz
mathematischer Methoden bei konkreten Anwendungen. Auch auf
diesem Gebiet stellt Professor Werner ein Vorbild dar.
Wir widmen daher den vorliegenden Band seinem Andenken.
Lothar Collatz GUnther Meinardus GUnther NUrnberger

Preface
This international conference was convened under the joint

chairmanship of L. Collatz (Hamburg), G. Meinardus (Mannheim),
and G. Nurnberger (Erlangen/Nurnberg). Responding to the strong
demand for an applications-oriented conference on approximation
theory, a group of 50 scientists from west Germany and abroad
took part. The total of 36 lectures offered a well-rounded, up-
to-date survey of the intricate developments in approximation
theory today. Multifarious aspects of approximation by splines,
rational functions, exponential functions, and polynomials with
special emphasis on numerical realizations were considered.
Other topics included the adaptation and application of ap-
proximation methods to differential equations, integral equa-
tions, quadrature formulas, and control problems. Further
applications concerned the development of mathematical models,
data fitting, computer-aided design, and problems in physics,
engineering sciences, and medicine.
The relaxed and pleasant working atmosphere facilitated inten-

sive scientific discussions. Special thanks are due to the mem-
bers of the Research Institute for their skilled assistance
throughout the conference, and to the institute's Director,
Professor Dr. Barner, for his kind cooperation.
Lothar Collatz Gunther Meinardus Gunther Nurnberger

Hamburg Mannheim Erlangen-Nurnberg
vorwort
Die internationale Tagung stand unter der Leitung von L.

Collatz (Hamburg), G. Meinardus (Mannheim) und G. Nurnberger
(Erlangen-Nurnberg). Aufgrund der grossen Nachfrage nach einer
anwendungsorientierten Tagung der Approximationstheorie nahmen
50 wissenschaftler des In- und Auslands teil. Die insgesamt 36
Vortrage boten einen reprasentativen Ueberblick uber die weit-
verzweigten aktuellen Entwicklungen in der Approximations-
theorie. Behandelt wurden verschiedenartigste Aspekte der
Approximation durch Splines, rationale Funktionen, Exponential-
funktionen und Polynome mit besonderer Betonung der numerischen
Realisierung. Ferner umfassten die Themen Adaption und Anwen-
dung von Approximationsmethoden auf Differentialgleichungen,
Integralgleichungen, Quadraturformeln und Kontrollprobleme.
Weitere Anwendungen betrafen die Entwicklung mathematischer
Modelle, Datenanpassung, Computer Aided Design und Probleme der
Physik, Ingenieurwissenschaften und Medizin.
Die freie und harmonisch Atmosphare fuhrte zu intensiven wis-

senschaftlichen Diskussionen. Hervorzuheben ist auch die ausge-
zeichnete Betreuung durch die Mitarbeiter des Forschungsinsti-
tuts und das verstandnisvolle Entgegenkommen des Instituts-
direktors, Herrn Professor Dr. Barner.
Lothar Collatz Gunther Meinardus Gunther Nurnberger

Hamburg Mannheim Erlangen-Nurnberg
Table of Contents
BAKER, T.H.C. - DERAKHSHAN, Mir S., Computed Approximations to

to Some Power Series 11
BERENS, H. - FINZEL, M., A Continuous Selection of the Metric Projection
in Matrix Spaces 21
CHUI, C.K. - HE, T.X., On Location of Sample Points for Interpolation
by Bivariate Cl Quadratic Splines 30
COLLATZ, L., Approximation von Eckensingularitaten bei Randwert-
aufgaben 44
DELVOS, F.-J., Periodic Area Matching Interpolation 54
DEUTSCH, F., An Exposition of Recent Results on Continuous Metric
Selections 67
GOLITSCHEK, M.v. - LIGHT, W.A., Some Properties of the Diliberto-
Straus Algorithms in C(SxT) 80
GOLITSCHEK, M.v. - SCHARDT, F. - WIEGAND, M., Mathematische
Auswertung ergospirometrischer Messungen 95
HANDSCOMB, D., Knot-Elimination; Reversal of the Oslo Algorithm 103
HAUSSMANN, W. - ZELLER, K., Fourier Methods in Practical
Approximation 112
KRABS, W., On Time-minimal Heating or Cooling of a Ball 121
MERZ, G., The Fundamental Splines of Periodic Hermite Interpolation
for Equidistant Lattices 132
NURNBERGER, G., Strong Unicity Constants in Chebyshev Approximation 144
RAHMAN, Q.I. - SCHMEISSER, G., On a Gaussian Ouadrature Formula
for Entire Functions of Exponential Type 169
RAHMAN, Q.I. - SCHMEISSER, G., 1-iarkoff Type Inequalities of Curved
l1ajorants 155
REIMER, M., Interpolation mit spharischen harmonischen
Funktionen 184
SCHABACK, R., Convergence Theorems for nonlinear Approximation
Algorithms 188
SCHEMPP, W., Orthogonal Polynomials in Opto-Electronics: Fiber
Optical Communication Systems 201
SCHMIDT, J.W., On the Convex Cubic C2 -Splines Interpolation 213
SCHWARZ, H.R., Rayleigh-Quotient-Minimierung mit Vorkonditionierung 229
WATSON, G.A., Data Fitting by Sums of Exponentials Using th~1 Norm 246
International Series of
Numerical Mathematics, Vol. 81
© 1987 Birkhauser Verlag Basel
COMPUTED APPROXIMATIONS TO SOME POWER SERIES
Christopher T H Baker and Mir S Derakhshan
Department of Mathematics, The University of Manchester.
1. The Problem and its Applications
We are interested in approximating certain formal power series (fps) of the type
(1.1) exp) = no + nIP + n 2p 2 +
To be more specific, we are concerned with techniques for computing the first N
coefficients of (1.1) where
(1.2) exp) = {«(3o+(31P+···+(3kPk) I (aO+alp+···+~pk))"
with p f (0,1] and where {aQ,(3Q} are specially chosen. We intend to concentrate, here, on
the case p = 2 which is of particular interest. Though we have particular applications in
mind and shall proceed in a fairly informal manner, we shall discuss (in terms which are
practically relevant) the efficiency of various algorithms for the problem under discussion.
The problem of computing (1.1) finds application in the following context. Suppose
that a k - step linear multistep method for the numerical solution of the initial- value
problem for a first-order ordinary differential equation is defined by its first and second
characteristic polynomials (cf. LAMBERT [3]):
(1.3) pep) = pko{p- 1) ; u(p) = pk(3(p-l),
wherein
(1.4)
(j( IL) = (30 +(31 p. +(321'2 + ... + (3kILk
We suppose the following: that p(l) = 0, p'(l) = u(l) =t 0, that (30 =t 0, and that,
with p*(p) = p(p)/(p-1), the polynomial
(1.5) p*( p) = a~pk- 1 +a~ pk- 2+ ... +ak- 1
is a Schur polynomial (Le. its zeros are located in the open unit disk centered on zero).
Observe that consequences of our assumptions are : 0(1) = 0, a'(l) = (3(1) =t 0, (3(0) t=-
O and 0{ p) = (1 - p)cx*( p) where cx*( p) is a polynomial with its zeros outside the closed
unit disk centered at zero, and a*(1) = p '(1).
With the given assumptions, which apply when {p,a} defines a strongly-stable
12
implicit linear multistep methods, it has been shown that the linear multistep formulae is
associated with formulae
(1.6) {1/r(v)}Poh(nh-W-'l'<t)dt .. h V Lj= oWnjl'<jh), v f (0,1]; n f 2:+
where
(1.7) Wnj = 11n-j
for j 0 .. j .. n (for some j 0 f l+). For the theoretical background in the case v = 1,
see WOLKENFELT [6] (the rules reduce to indefinite integration); for the case v f (0,1)
see LUBICH [4]. Formulae (1.6) find application in the discretization of Abel equations (v
= ~ corresponding to classical Abel equations) and Volterra equations (the case v = 1).
2. Recurrence Relations.
If we set
(2.1) {{3(p.)/o(p)} =: 14P.) = Wo +w, p. +w 2 '?- +
then
(2.2) O{p.) = {14P.)}V
and, clearly, determination of the coefficients of O{p') is facilitated by a knowledge of the
coefficients of 14P.). On rationalizing (2.1) and equating coefficients, we find (for n )0 0)
that
(2.3) O!oWn+ O!,Wn-, + ... + O!kWn-k = (3n
with the convention
(3Q = 0 for Q t {0,1, ... ,k}.
The values wQ for Q = 0,1, k are found by forward substitution, and the remaining
equations (2.3) give a recurrence :
(2.4) Wn = -{O!,Wn-,+ ... + O!kWn-k}/O!o (n > k)
determining wQ for Q > k «(3n = 0 if n > k). Notes. For some specific familiar choices
of {O!Q, (3 Q}, eqn (2.3) is considerably simplified. Thus, for the Adams - Moulton formulae
(AMFk) , O!o =1, O!, = -1, O!Q = 0 otherwise and (2.4) becomes Wn=(3n+Wn-1' so we
note that Wn= 1 for n > k. For the backward differentiation formulae (BDFk) with k .. 6,
(3Q = 0 unless Q = 0 and lim Wn = 1 as n ~ 00.
WOLKENFELT [6] discussing the properties and use of (2.3) to generate {wQ}
observes the intrusion of rounding error and proposes a variant method. Indeed, the
stability polynomial for (2.3) is p( p.) which is simple von Neumann, so that rounding error
cannot grow catastrophically but can accumulate linearly. However, employing the notation
in (1.5), summation of eqns (2.3) provides the result (for n > k):
(2.5) O!~Wn+O!~Wn-, +···+~-,Wn-k+' = rn
13
where (for all n > k)

(2.6) In = I; I := Q~"'k+Q~"'k-I+"'+ak-IWI .
A particular solution of the recurrence (2.5), since In = I is independent of n is ~ =
w; w:= I/tr(l) = IIp'(1). It follows that the general solution of (2.5) can be written
as ~ + ~ where ~ is the general solution of the homogeneous version of (2.5).
However, the solution we seek is determined by the values Wo ,WI ' "', wk which satisfy
the first k+l equations in (2.3). It follows that Wn = W + w~ where w~ = wn-w for n
= 0,1, "', k and
*" *.. *
(2.7) QOun+Qlun-1 +"'+~-IWn-k
It
°
We arrive at the result of WOLKENFELT [6], who proposes the solution of (2.7) to
determine {UJn} because (2.7) is a recurrence relation which is strictly stable (since p*( It) is
a Schur polynomial). Finding {wQ} determines the function (3(It)/a(It) as a fps.
3. Computation of {WQ}.
We have indicated an interest in the values {wQ}, and there are a number of
techniques which can be used to solve (2.3), of which we consider two.
(a) Direct method. For the Adams-Moulton rules, (2.4) allows direct evaluation of {wQ} as
we noted previously.
More generally, the use of (2.7) allows us to compute {w~} using
(3.1) uh = - {~uh- I + ... + ~uh-k}/<ib (n > k)
from whence un can be determined [6].
(b) FFT method. The expression in braces on the right-hand of (3.1) is, for each n, a
discrete convolution, and FFT techniques are known (cf. HENRIC! [2]) to play an effective
role in the determination of a family of convolutions. Even though the direct use of (3.1)
is preferable (k is small), we shall indicate how an FFT technique can be used for the
present purpose; it has some intrinsic interest and it serves to introduce the FFT (for
which we have further use later).
Let {a r }¥;- b be a sequence of (complex) numbers and let {aQ}¥ ;- 6 be the
discrete Fourier transform (DFT):
aQ = (11 jM)L~=~lr exp{-i21rrQIM}
(see §6); then we write
14
The 'natural' periodic extension of {ar}~;;- ~ of period M = 2N denotes the

sequence {ad with aN= aN + I = .... = a 2N - I defined to be zero and a 2jN +r = ar
for j = O. ±1. ±2. ...; r = 0.1 •...• 2N -1. {arl~;;- ~ will be called the restriction
of the sequence {arP~;;-~. When a known sequence {ar}~;; ~ defines its natural periodic
extension of period 2N it does so irrespective of any other values which might naturally
be assigned to ar for r t {0.1 •...• N-l}.
Theorem 1. Suppose {~}~oo. {sr}~oo to be periodic of period 2N (Le. q2jN+r = qr
and S2jN+r = Sr for j = 0, 1=1,±2, ... ). If
Pr = fj=ON- 1
~-J·sJ·' r = 0.1 •...• 2N -1;
and if [qr ~ Cit! p~-I and [sr ~ St!P~-' then [Pr ~ Pt! p~-I where Pt! =
J(2N)xqt!st!. and vice-versa.
The above theorem has a role [2] in forming the coefficients of the product of
two polynomials. If q( p) = q 0 + q ,p + ... + qN - , pN - I • s( p) = s 0 + s,p +
+ SN_,pN-I. and
(3.2)
then
(3.3) (r = O. 1. 2 •...• 2(N -1».
In consequence. we can apply the convolution theorem; suppose we have the natural
periodic extensions of {qrl~-'. {srl~-I of period 2N. and [qr ~ Cit!J2~-'. [sr ~
st!P~-'. If [Pr ~ Pt!P~-' where Pt! = J(2N)x(Cit!st!) for Q = 0.1 •...•2N-l.
we can find the coefficients {Pd using the relation
2N-l
)L
Pr = J(1/ 2N t!=!t! exp{+i1rrt!/N} (r = 0.1 •...• 2(N-l».
We shall use a non -standard variant of this result later.

Given the possibility of finding a product of two polynomials. the leading
coefficients in the formal power seriers representing a reciprocal can be determined using a
Newton iteration. If ')(p) = {yo + 1'IP + 1'2p2 + } we employ the notation
[')(P)]N to denote 1'0 + 1',p + 1'2p2 + ···+')'N_1pN-1.
Theorem 2 (Newton's iteration to find a reciprocal). If zlo](p) = llcl!o and
zIs+'](p) = [2 X zIs](p) - {zIs](p)p X o(P)hN (s = 0.1.2 •... ; N = 2s)
then )s+ '](p) = [lIo(p)hN.
It is instructive to compare the cost of the direct method with that based on the
use of the FFT. as indicated her. using FFT routines in the NAG library [5]. The times
to compute [c.(P)]N = [(3(p)/o(P)]N are shown in Table I. The direct method is clearly
most economical by a considerable margin. Note: Computing the weights {CUr} corresponding
to the BDFk rules. we observe from WOLKENFELT [6] that. when r > 59 for k = 4
15
and when r :. 227 for k = 6, Wr agrees with its limiting value limr~ Wr to
Direct Di rect FFT

N= k-4 k-6 k= 4,6
128 0.001 0.001 0.036
256 0.002 0.002 0.068
512 0.003 0.004 0.143
Table I (Timings in seconds on the CDC 7600)

15 significant figures. (There is therefore no need to compute excessively large numbers of
the coefficients Wr.)
4. Iterations for the Computation of {~}.
Implicit in our application of the FFT in §3 is the observation that the

coefficients of the product of two polynomials are expressible as a family of discrete
convolutions. The FFT provides a useful tool for large families of convolutions and allows
us to compute by an iterative technique the N leading coefficients of (1.1) with
(4.1) ~p.) = {wo + wlp. + w2p.2 + ... }P (p E (0,1]),
wherein Wo '4 0, and N = 2R , R E Z +. We shall concern ourselves with the cost in
computational time of iterative techniques in the special case P = i.
As in §3, particular cases of interest arise when we consider AMFk and BDFk
rules. For both these cases, we can assume the required coefficients wQ in
(4.2) ~p.) = {wo + WI P. + w2p.2 + ... }i
to have been found as in §3, but as HAIRER, LUBICH & SCHLICHTE [1] have shown,
there is no necessity to compute these coefficients in the case of BDFk rules, since, if we
normalize the coefficients with fJ 0 =I, the task in the case of BDF k rules is that of
finding
(4.3)
Since our problem can assume the form either (4.2) or (4.3), we remind the reader of
iterative techniques for computing the square root (4.2) or the reciprocal square root (4.3).
Theorem 3 (Heron's iteration for a square foot). If do](p.) = jW o and
(4.4) d r + l](p.) = U [dr](p.) + [w(P.)hN ><[l/dr](p.)hNhN
16
then d r + 1](p.) = [n(P.)hN where N = 2r , for r = 0,1,2,3, ....

To implement as an algorithm the iteration (4.4), a reciprocal [ltn[rkp.)hN is
required. The following shows that it can be computed by two steps of an iteration:
Theorem 4. If, with the notation of the preceding theorem, zlo](p.) = [1Idr-'](P.)]~N
and
(4.5) zls](p.) = [2 zls-'l(p.) - [{zls-'](p.)PhN dr](p.)hN (s = 1,2)
then z[ 2](p.) = [1Id r](p.)hN.
It is apparently simpler to exploit the formulation (4.3), when it is an alternative
to (4.2), since we can use an iteration which does not require a reciprocal of a fps:
Theorem 5 (Newton's iteration for the inverse square root). Suppose (30 = 1. If d o]( p.) =
11 ICl o and
(4.6) d r + '](p.) = H3xd r](p.) - [a.{p.)] >4: {dr](p.)PhNhN
then dr+,](p.) = [n(P.)hN = [{a.{P.)}-~hN , where N = 2r , r = 0,1.2, ....
The FFT is a useful tool to compute products of polynomials if the degrees are
large, but where one of the polynomials is of low degree direct multiplication is faster.
s. PRACTICAL ALGORITHMS.
The precise description of an algorithm requires the use of a programming

language and the specification of the computer on which it will be implemented. However,
by introducing a method of describing various tasks in terms of a pseudo - programming
language, we can display the expense of implementing methods suggested by the preceding
discussion. Thus, we suppose the calculations rely on the following procedures:
procedure extend (w,Q,m); comment Given wo'w" ... , wQ_, define WQ = WQ+, =
Wm-, =0 (m > Q).
procedure truncate (W,m); comment Given the sequence Wo ,w" ... , "hl-" return the
half-length sequence wo'w,. ... , w~m-1.
procedure lIt (w, w, m); comment Given {wo'w"
... , Wm - ,} by FFT techniques.
procedure invllt (w, W, m); comment Given {wo'w" ... , Wm-,}, compute the inverse
DFT {wo,W,. ... , Wm-,} by FFT techniques.
procedure prod (w, Cl, (3, m); comment Given {Clo'Cl" ... , O'm-,} and {(3o,(3" ... ,
(3m-,}. compute {wo.w, • ...• Wm-,} where "T = Clr x (3r·
procedure invlps (q, f, n, m); comment Given the ordered coefficients {fel, {~} of f(p.)
= fo +f,p. + ... +fn_,p.n-, and q(p.) = qo +q,p. + ... +qm_,p.m-, =
17
[1If(J.I)]m. where either n = m or n = ~m. evaluate the Newton improvement

q(J.I) := [2 x q(J.I) - [{q(J.I}P] x f(J.I)hm
and return the cofficients qo' q1' ·····.q2m-1 of [11 f(J.I)hm .;
begin comment performs the task equivalent to the following (see the remark below);
extend(q. m. 4m);
tfl(q. q. 4m);
extend(f. n. 4m);
tfl(i. f. 4m);
prod(f. q. f. 4m);
proti(q. q. f. 4m);
invtft (p. q. 4m);
Iruncate (p. 4m);
for i := m ~ 1 12 2m-1 do Qj := - Pi;

end code for invfps 0;
Remark: In practice. where one or both of q and f can be supplied as parameters. in lieu
of q and f. respectively. this will be more economical.
With the above. we are able to implement the iteration represented by (4.4):
procedure HERON (w. O. R); comment Given the coefficients {woo w1 • w2 • .... "N-1}
(N= 2R ). compute the first 2R coefficients Or of O(J.I) = {wo +w 1J.I+W 2J.12 + ... }~.;
begin comment code foUows;
0 0 := 11 IQo;
for r := 0 ~ 1 until R do
begin N := 2r;
extend(w. 2N. 4N); extend(O. 2N. 4N);
tfl(w. w. 4N); [[t(D.. O. 4N);
ifr=Othen
begin
So:=1I0 o;
inv[ps(S. O. 1.2)
end if;
~
begin inv[ps(S. O. N. N);
inv[ps(S. O. N. 2N)
end else;
comment S 0 .. ··.S2N -1 are correct fps coefficients of current [1/0( J.I)];
extend(S. 2N. 4N);
[[I(S. S. 4N);
18
prod{O. W. S. 4N);
invflt (v. n. 4N);
truncate (I', 4N);
for i:= N ~ 1 m 2N-l !!!L !7;. .- 0.5 x vi;
end r;
end Heron code 0;
The "building blocks" introduced above also permit a description of the code of
HAIRER, LUBICH. and SCHUCHTE [1], if we introduce a further procedure:
procedure multdirect(a. n. n. k. N. m); comment Compute directly the first m coefficients
{nrlW'-' of the product (a o + a,JL + +akJLk) x (no +n,JL + ... +
% _ , JLN - '), given their coefficients.
We shall describe a modified version of the code of [1], which has improved
efficiency when compared with the original.
procedure RECIPSQRT (a. n. R); comment Given the coefficients {a o' a" a2 , .... Ok}.
compute the first 2R +, coefficients nr of neJL) = {a o + a, JL + a 2 JL 2 + ... OkJLk) -1·

begin comment code follows;
no := 11 lao;
for r := 0 ~ 1 until R do
begin
N := 2r;
extend(n. N. 4N);
flt<n. n. 4N);
m := min(N + k.2N);
multdirect( a. n. k. 2N. m);
extend(a. m. 4N);
flt(o.. a. 4N);
prod{o.. n. o.. 4N);
comment finding a & recalling n is unnecessary here;
prod{o.. n. o.. 4N);
invflt (v. o.. 4N);
truncate (I'. 4N);
fQr i:= N ~ 1 m 2N-l !!Q. !7;. .- -0.5 x Vi;
end r;
end Recipsqrt 0;
We have to explain the comment (in the middle of the above code) that finding a
and recalling the stored n at this stage is unnecessary. The reader might expect to find.
19
in place of this comment. the following code (cf. [1]):

invfft (o,a., 4N);
truncate( o,4N);
extend( 0,2N ,4N) ;
fft(&.,0.4N) ;
The key to our algorithm. which shows that this is unnecessary. is the following result.
Proposition (Complement to Theorem 1). Suppose that a(p) = ao +a11l + ... +
a2N+k_lIl2N+k-l and that Z(1l) = Zo +Zlll + ... + ZN_lIlN - 1• Define a r = 0
for r = 2N+k. 2N+k+l •...• 4N-l and define Zr = 0 for r N. N+l •...• 4N-l.
and suppose that Car _ aQ]ri N - 1 and [Zr- Zdri N - 1. If ~Q 2jN x aQZQ and [1fr
_ ~Q]riN-l then
1fr = I~ arZr-j for r = 0.1 •...• 2N-1.
The preceding equality thus holds for a restricted class of suffices r. sufficient for
our application. The result is not the standard one since we do not employ the natural
periodic extension of the first 2N coefficients {ar } ~N - 1.
FFT routines in the NAG library [5] permit us to implement our various codes.
We show in Table IT some of the timings obtained in experiments using the NAG Mark 11
routines on the CDC 7600 of the Manchester University Regional Computer Centre. when
finding the coefficients nQ corresponding to BDF 4 rules.
N HAIRER et al. [1] Modified Version HERON [including

of [1] direct generation of [wQ) ]
128 0.034 0.018 0.047

256 0.064 0.035 0.090
512 0.136 0.074 0.182
1024 0.269 0.148 0.367
2048 0.534 0.294 0.726
Table IT (Comparative timings. in seconds. of codes implemented on the CDC7600)
6. CONCLUSIONS
The above descriptions of algorithms allow us to make a count of various

operations involved in the algorithms. For our purposes. Heron's algorithm is more
generally applicable than the other codes. The relative merits of different codes depend
upon the relative costs of the operations involved. Additionally. our description relies upon
20
codes for implementing fast Fourier transforms. Some routines provide for the storage of
sine and cosine values required when a first call to a procedure like ttt is made, so that
subsequent calls with the same sequence length M are cheaper. Other codes modify the
factors (l/IM) in the definitions of the FFT of a sequence of length M, to avoid
unnecessary multiplications.
To summarize, we wish to impress upon the reader the gains to be made by a
search for a good method, and analysis which leads to efficient encoding of an algorithm.
7. REFERENCES
[1] Hairer, E., Lubich, Ch., and Schlichte, M., Fast numerical solution of weakly singular
Volterra integral equations. Tech. Report, Dep. Math. University of Geneva, May 1986.
[2] Henrici, P., Fast Fourier methods in computational complex analysis. SIAM Rev. 21
(1979) pp 481 - 527 .
[3] Lambert, J.D., Computational methods in ordinary differential equations. Wiley, London
(1973).
[4] Lubich. Ch., Discretized fractional calculus. SIAM J. Math. Anal. 17 (1986) pp
704-719.
[5] Numerical Algorithms Group, The NAG Manual Mark 11. Numerical Algorithms Group,
Banbury Rd., Oxford.
[6] Wolkenfelt, P.H.M., The numerical analysis of reducible quadrature methods for
Volterra integral and integro-differential equations. Academisch Proefschrift, Amsterdam,
1981.
Christopher T. H. Baker
Department of Mathematics
The Victoria University of Manchester
Oxford Road
Manchester M13 9PL
England.
© 1987 Birkhiiuser Verlag Basel
A Continuous Selection of the Metric Projection
in Matrix Spaces
by
H. BERENS AND M. FINZEL

Mathematical Institute
University of Erlangen·Nuremberg
Erlangen, F. R. G.
1. Introduction
We denote by C mx n the vector space of complex m X n matrices over C, m, n e N, with elements
A,B, .... For A e C mx1 and Be c 1xn , l,m,n eN, AB denotes the matrix product of A and
B in c mxn , A· the adjoint of A in C 1xm • Instead of C nx1 we write cn, the vector space of
complex column vectors, we also write z, 'U, ••• to denote its elements. c mxn can be identified
with the vector space of linear transformations of c n into cm: for a given matrix A e C mx n, A
acts on z e c n via the matrix product Az.
For an A EC mx n, we denote its lingular value decompo.ition (SVD) by UEV·, where U
and V are unitary matrices in c mxm and c nxn , respectively, and E is a diagonal matrix in
c mx n with diagonal elements 0'1 ~ 0'2 ~ ••• ~ O'min(m,n) ~ 0, the so-called lingular value.
of A. We write E = diag(O'I,O''J'''',O'mln(m,n»). If A 1= 0 has rank r, i. e., 0'. > 0'.+1 = 0,
0< r :5 min(m, n), and if 1'1.117 '1.12, ... , 'UmJ and lVI, v'J,"" vnJ are the column representations of
U and V, respectively, then A = Ej=1 O'j 'Ujvj - for u e c m and v E c n 'Uv· denotes the tensor
product of'U and v in c mxn .
We are interested in the approximation of a given matrix in c mxn by matrices in c mxn of
restricted rank, say, ofrank less than or equal to k, 0:5 k < min(m,n), where c mxn is endowed
with the spectral norm II· 1100' If we equip c n and c m with the Euclidean norm 1·12 and if we
identify c mx n with the vector space of linear transformations on c n to C m, the spectral norm
on c mx n is equal to the associated operator norm. The spectral norm is invariant w. r. t. unitary
transformations, i. e., for A e c mxn and A = UEV· IIAlloo = IIElioo = 0'17 the largest singular
22
value of A. The Frobenius norm on e mxn is a second norm of interest to us. The matrix space
e mxn is vector space isomorphic to e mn • If we define the sesqui·linear form on e mxn by
1
(A,B) = trace A·B = trace BA· = -trace (A·B+BA·),
2
e mxn becomes an inner product space, isomorphic to e nm endowed with the Euclidean norm.
The associated norm 11·112 is the Frobenius norm, which is also invariant w. r. t. unitary trans·
formations, i. e., for A E e mxn and A = VI:V· IIAII~ = III:II~ = Ej~~(m,n) 11;.
Setting
Rk = {R E e mxn : rank(R) ~ k}, o ~ k < min(m, n),
distp (·; RIr)e mx n -+ R
: and pPR.. : e nx n -+ Rk denote the distance function and the metric
projection, respectively, e mx n being endowed with the Frobenius norm (p = 2) and the spectral
norm (p = 00), respectively.
Fix the rank k. For an A E e nxn \ RIr and A = UI:V·, we set
It is well· known that

mlll(m,n)
{dist2(A; RIr)P = E and
j=lr+l
that
and that 2 PR.. (A) is a singleton, whenever I1Ir > 1111'+ 1, i. e., Air is well· defined, independent of the
SVD of A. More precisely, the uniqueness set 2UR.. is equal to {A E e mxn : 11k> 1111'+d, a dense
open subset of e mxn • Moreover, the associated mapping e mxn 3 A H Air defines a continuous
selection of ooPR.. on 2UR.. which, however, has no continuous extension on e mxn •
We want to present a selection of ooPR.~ which is continuous on e mxn and which is in addition
'unny:
Theorem A. For A E e mx n \ Rk, 0~ k < min(m, n), and A = UI:V·, set
,,0
Llk = diag (0'1 - O'k+l, 0'2 - O'Ir+l,···, O'k - O'k+l, 0 , ••• , 0) and A~.. = U"~V·,
Ll A-
then A~ is well·deflned and e nxn 3 A H A~ is a continuous, sunny selection of ooPR. ••

The .alar property of the selection means, VA E e mxn and Vt E R+ A~ is the selected
element of ooPR..(A~+t(A-A~)). In particular, the theorem implies that RIr is a sun in c mxn •
The proof of Theorem A is elementary and given in the following section. In the remaining
sections we are going to study ooPR.. in greater detail. Among others, we are able to characterize
all solar points of ooPR..(A) for a given A in e nxn , e. g., we prove for So, the set of singular
matrices in c nxn
23
Theorem B. Let A E C nx n \ So. The matrix S in So is a solar point of A in ooPSo (A), exactly
when there exists a unitruy matrix W in c nxn such that
A = S + Un W and W· S + S·W is positive definite,
Un being the smallest singular value of A.

If U:LV· is a SVD of A, then Q = UV· is such a unitruy matrix which, in addition, defines
a continuous selection of ooPS o'
Let E be a normed vector space over R with elements x, y, ... and norm II . II. A nonempty
subset K of E is said to be a ,un in E in the sense of VLASOV if it is an existence set, i. e.,
Vz E E PK(Z) =F 0, and Vz
E there exists a k E PK(Z) such that k E PK(k +t(z - k)), t E R+.
E
We call such an approximant a solar point of z in K. Proximinal convex sets in E are suns, and
if the norm is smooth, then suns in E are convex. In some sense suns in a normed vector space
are "convex" sets. Theorem A then states that Rk, 0 ~ k < min(m,n), is a sun in c mxn and
that the metric projection admits a continuous selection which preserves the solar property.
Solar points of the metric projection are characterized by the KOLMOGOROV condition. Let
(., ')s : E X E -+ R denote the semi·inner product on E defined by
Vz,y E E (y , z)s = 1->0+

lim
liz + tYl12 -ll zl12 = max{ (w y): w E 4l(z)}
2t '
where (.,.) : E· X E -+ R is the bilinear form on E· X E and 4l : E -+ E· the duality map, or

the subgradient of II . 112/2, see e. g. [2] for further details.
An element k E K is a solar point of Z exactly when
Vk' EK (k-k',z-k)s~O.
This is the so· called KOLMOGOROV condition. The characterization of solarity via the KOLMO·
GOROV condition was established by BROSOWSI.
Talking about convexity, solarity, and its characterization via the semi-inner product, we
consider notions which address a real vector space. In the final two sections we determine the
semi-inner product on c n ,< n, considered as a real vector space and equipped with the spectral
norm, and prove and discuss Theorem B.
24
~. The proof of Theorem A

To prove the theorem, we have to prove that the selection is well-defined. But this is an immediate
consequence of the following
Lemma. Let A E cmxn,A.,. 0, and let UEV· and XEY· be two SVDs of A, where
E = diag(O'I,""
~
n,
0'1,0'2, •• • ,0'2,
. . . ... ,0'.,
, ..
... ,0'.,0, ... ,0),
n,
nl +na + ... +n. =ranle(A), then there exist unitarymatricesS andT in c mxm and c nxn ,
respectively, such that X = US and Y = VT and
S = diag(W1 ,. • • ,W.,So) and T = diag(Wlo' .. ,W.,To),
where So,To, and WI, 1 ~ l ~ r, are unitary matrices in cmoxmo, mo + ranle(A) = m, in

C nox no, no + ranle(A) = n, and in c n , x n" respectively.
If m = n and if A is nonsingular then UV· = XY·.
The lemma needs no proof, it's just a matter of formulation. It follows that for a given
AE cmxn R./r and for the SVDs UEV· and XEY· of A UE~ V· = XE~Y·. Since the singular
\
values of a matrix A E c nx n are continuous functionals on C nx n, the selection is continuous.

To prove that it is sunny, it suffices to consider diagonal matrices. Ohviously, Vt E R+ E~ is the
selection of the metric projection of E~ + t(E - E~) in R./r. •
As a consequence of Theorem A we have
Corollary. The set R./r, 0 ~ Ie < min(m, n), is a sun in C mx n in the sense of VLASOV, c mx n
being endowed with the spectral norm.

25
S. The semi-inner product

We restrict ourselves to the matrix space c nxn , n e N, and define on c nxn X c nxn the sesqui·
linear form
VA,B e c nxn (A,B) = trace A· B = trace BA·.
Since the trace of a matrix is invariant w. r. t. similarity, for A = U1:V· and B = UOV·, U and
V being unitary inatrices, (A, B) = (1:, 0).
It is well·known that the dual of (c nxn , 11.11(0) is isometrically isomorphic to (c nxn , 1I·lId,
i. e., if 1 is a linear functional on c nx n, equipped with the spectral norm, then there exists a
unique matrix A in c nx n such that
n
VB e c nxn l(B) = (A,B) and 11111 = IIAlh = 111:lh = LUj,
j=1
where U1:V· is the SVD of A.

We are interested in a characterization ofthe faces of the unit ball of (C nx n, II· lid , i. e., for
a given A e C nx n, IIAlloo = 1, we are interested in a characterization of
{We c nxn :!R (W,A) = 1 and IIWIlI = I}.
Because of the invariance of the sesqui·linear form w. r. t. similarity, it suffices to characterize

the face associated with a diagonal matrix 1:.
Lemma. Let 1: = diag(u1!'''' Un), 1= Ul = ... = Up > Up+l ~ ... ~ Un ~ 0.

face(1:) = {Oe c nxn : O=diag(QTQ·,O), Qe C pxP , unitary,
p
T=diag(T1! ... ,Tp),Tl~ ... Tp~O, and LTj=I}.

j=1
The set of extreme points of the face is given by
extface(E) = {O e c nxn : 0 = qq., q e CP X {o}n-p and Iql2 = I}.
Let XTY· be a SVD ofthe matrix 0, where T = diagh, ... , Tn), Tl ~ .. , ~ Tn ~ 0, and
I:i~nlk(O) Tj = 1. It follows from
(ra~E) Cjj) : ; L
rank(E) rank(E) rank(O)
1 = !R (O,1:) = !R Uj Uj Cjj :; E ICjjl::;; L TI = 1
;=1 j=1 j=1 1=1
that Cjj ~ 0, 1 ::;; :j ::;; p, Cjj = 0, p < i ::;; n, and Ej=l Cjj = 1. In the final estimate of the series
of inequalities we used the SVD of 0: Cjk = Er~~k(O) TI aJjl'Ykl, 1 ::;; i, k ::;; n. Since, furthermore,
1- f,'H mE)" (tz;,.~) -~~O) (f,z;,.;,),

= "R
26
we have R(E~=1 ZjlYjd = 1, 1 ~ I ~ rank(C), giving Zjl = Yjl, 1 ~ I ~ rank(C), 1 ~ i ~ p, and

Zjl = Yjl = 0, 1 ~ I ~ rank (C) , p < i ~ n. Since the coefficients Cj" of C vanish for p < i ~ n or
p < k ~ n, the rank of C is less than or equal to p, which leads to C = diag( QTQ·, 0), where Q
is a unitary matrix in Cpx P , and T = diag(1'tt ... , 1'p), 1'1 ~ •• , ~ 1'p ~ °and Ef=l 1'1 = 1.
It remains to characterize the extreme points offace(~). Assume
p p
qq. = 0 L UI u,l'uj + PE 1'1 'II/,vj,

1=1 1=1
where q E CP, Iql~ = 1, U",'111 E CP, 1 ~ I ~ p, and u,ju,m = 'IIj'llm = 15'm, 1 ~ I,m ~ p,
UI ~ 0, 1 ~ I ~ p and Ef=1 UI = 1, 1'1 ~ 0, 1 ~ I ~ p and Ef=1 1'1 =1, 0, P> 0, 0 + P = 1.
From the representation
p p
q= 0 E UI (U,jq)U" + P E 1'1 ('IIjq)'111 = 0 U+ P 'II

1=1 1=1
and the estimates
p p
1u,1~ = E u?l ujql2 ~ E lujql~ = Iql2 = 1,
1=1 1=1
P p
I'III~ = E 1'?lfljql2 ~ E l'IIjql2 = Iql2 = 1,
1=1 1=1
and 1 = Iql~ ~ olu,l~ + PI'll 12 , we obtain that 1u,1~ = IflI2 = 1, Ul = 1, U2 = ... = up = 0, and
1'1 = 1, 1'2 = ... = 1'p = O. Hence, q = (Uiq)U,1 +{3 ('IIiq)'111' Since, furthermore, 1 = 1'U12 = l'Uiql
0
and 1 = I'III~ = l'IIiql, it finally follows that q = (Uiq)Ul = ('IIiq)'111 = U, = 'II, proving our claim• •
In [31 GRZASLEWICZ characterized the faces ofthe unit ball of the dual of .c(an ). Since the
lemma is central for our paper and since its proof is straightforward, we decided to sketch our
arguments.
Corollary. Let A E cnxn, A =F 0, and let UEV· and XEY· be two SVDs of A, where E =
diag(ul'" 0' Un), Ul = ... = Up > Up+l ~ ••• ~ Un ~ 0, then
~(A) = {W E c nxn : R (W,A) = IIAII~ = IIWIIB = U~{E)V· = X~{E)Y·,
where
Moreover,
VB E c nxn (B,A)s = max {R{Uq)· B{V q) : q E CP X {o}n- p and Iql~ = ut}

= max {R{Xq)"B{Yq): q E CP X {o}n- p and Iql~ = ut}.
27
It just remains to prove the formula for the semi·inner product. To do this, we point out
that
(B,A}s =max{lR (W,B): WE 4>(A)}
=max{lR (C,U-BV): CE ext 4>(E)}.
But for C = qq., q E e p X {o}n-p and Iql~ = 0'1> (C, U· BV) = trace U· BV qq" = q·U· BV q =
(U q)" B(V q). •
4. The proof of Theorem B

To prove the theorem, let A E e nx n be nonsingular, and let its norm be equal to one. The matrix
So E So is a solar point of A exactly when
VS E So (So - S,A - So)s ~ o.

We set A - So = XDY", where X and Y are unitary and D = diag(d1 , ... ,dn ), distoo(A; So) =
d = d 1 = ... = dr > dr+ 1 ~ '" ~ d n ~ o. Since 4>(A - So) = X4>(D)Y", the KOLMOGOROV
condition can be rewritten as
VT E So max{lR q. (To - T)q : q E e r x {O}n-r and Iql~ = I} ~ 0,
where So = XToY· and S = XTY·. We claim that r = n and that
Vq E en
Suppose that 1 ~ r < n. Then the matrix T = diag(rTo +1,,0), where rTO E e rxr denotes the
upper r Xr principal submatrix of To and 1r the identity matrix in e rx r, satisfies Vq E e r x {o}n-r,
Iql2 = 1, lR q" (To -T)q = -1, contradicting the KOLMOGOROV condition. Consequently D = d1,
and the KOLMOGOROV condition now reads
VT E So max lR q. (To - T)q ~ O.

!q!.=1
Suppose that Ro = (To + To)/2 is not positive semi·definite-To = Ro + iJ o, where Ro and
Jo are hermitian- then Ro = Z"AoZ, where Z E e nxn is unitary and Ao = diag(A1>""An),
Al ~ ••. ~ An, and An < O. Set T = Z· (Ao - AnI) Z, then T is hermitian and singular and
= An < 0, again a contradiction.
Vq E en, Iql2 = 1, lR q. (To - T)q
On the other hand, if Ro is positive definite, then the KOLMOGOROV condition is satisfied.
Indeed, for a given T E So select a vector q E en in the kernel of T, then lR q. (To - T)q =
li q"Toq ~ O.
Setting W = XY·, we verified that So E So is a solar point of A exactly when A = So + d W
and W" So + SoW is positive semi· definite. Let UEV· be a SVD of A. If qo E ker(So), Iqol~ = 1,
then Aqo = d W qo and, consequently, d = d IW qo 12 = IAqo 12 ~ 0' n' Furthermore, it follows from
28
that Un = d, Vn E ker(So} and WVn = Un, where (Vn,U n ) is the n·th singular pair, i. e., AV n =
UnU n and A·u n = O"nVn.
The fact that Q = UV· satisfies the asumptions of the theorem is easily checked. Since Q is
independent of a specific SVD of A, c nxn 3 A H A - O"nQ defines a continuous, sunny selection
•
Let E = diag(O'l,"" un}, 0"1 ~ ••• ~. Ur > Ur+l = ... = Un > 0, be a diagonal nonsingular
matrix in C nx n. If the unitary matrix W defines a solar point of E in So, i. e., E - unW = So E So
and W· So + Sd'W is positive semi· definite, it follows from the arguments given above that W =
diag(W"I n- r}, where Wr E c rxr is unitary and such that Vq E C r , Iql2 = 1, ~ q·W:Erq ~ Un,
Er = diag(Ul,"" O"r}.
while the set of solar points is given by
If Ul = U2 = u, then
With respect to the Frobenius norm we have for Ul > U2 and "1 = 0"2 = 0"
respectively.
We tried unsuccessfully to characterize the metric projection on C a x 3 onto the set of singular
matrices, just to get more information. We were, however, surprised when we found that the solar
points of the metric projection on c nx n onto So admit such a simple description, leading to a
second proof of Theorem A for m = nand k = n - 1.
In [1] BERENS AND SCHMID proved that the set St = {S E c nxn : Idet SI ~ t}, t E R+, is a
CHEBYSHEV set in c nx n, equipped with the spectral norm, and that c nx n \ St 3 A H A - dt Q
is the approximant of A in Sf> where Q = UV· , UEP being a SVD of A, and 0 < cit = min{ l' >
o : (Ul - 1') ... (u n - 1') = t} < Un' For t -> 0+ the net of approximants converges again towards
the selection given in Theorem A.
We want to conclude with the following extension of Theorem B

29
Theorem C. Let A E c nxn \ Rk, 0 S k < n. The matrix S in Rk is a solar point of A

in ooP~.(A), exactly when there exist unitary matrices X and Y in c nxn and an integer r,
Ic < r S n, such that A = S +XDY", D = diag(d1, ••• ,dn ), Ulr+l = d1 = ... = dr > d,+l ~ ...,
and X"SY +Y"S"X is not negative definite on any (r-k)·dimensional subspace ofC r x {o}n-',
i. e., V(r - k).dimensional subspace H ofC' x {o}n-r there exists a vectorh E H, Ihl2 = 1, such
that !R h" X" SY h ~ 0, Uk+! being the (k + 1)st singular value of A.
If UEV" is a SVD of A, then U and V are SUell unitary matrices, !' is the largest index,
k < !' S n, for which u, = Uk+l holds, and D = diag(u" .•. , u" U,+l!"" Un).
We omit the proof.
5. Bibliography
[I] Berens, H. and Schmid, H. J. An example of a Chebyshev set, in Proceedings of the Conference
on Constructive Function Theory-8B, Edmonton 1986, to appear.
[2] Berens, H. und Westphal, U. Kodissipative metrische Projektionen in normierten linearen
Raumen, in Linear Spaces and Approximation, Birkhauser Verlag 1978,119-130.
[3] Grz~slevicz, R.{1985} Faces of the unit ball of the dual of .c{Rn), Math. Ann. no, 535-540.
Mathematical Institute, University of Erlangen-Nuremberg, 8520 Erlangen, F. R. G.
ON LOCATION OF SAMPLE POINTS FOR INTERPOLATION

BY BIVARIATE C1 QUADRATIC SPLINES
C. K. Chui 1 and T. X. He 2
Texas A&M University
College Station, Texas
U.S.A.
1. Introduction
Let V be a vector space of continuous functions in a region D
in RS, s ~ 1, with dimension N, and ~ = {x 1 , ••• ,x N} be a set of N distinct
points in D. The problem of interpolation from V at the sample points in
AN is said to be poised if for any given set of data {Yl' ••• 'YN} c R ,
there exists one and only one f e: V such that f(x i ) = y., i = 1, ••• ,N. For
1
instance, if ~~ denotes the space of all polynomials of total degree d in

s variables, its dimension is given by the binomial coefficient
N= (d:S) .
Although the interpolation problem from 1 is always poised, a necessary and
~d
sufficient condition for the interpolation problem from ~~ to be poised is

that the sample points in AN do not lie on any algebraic (polynomial) curve
of degree d. Location of sample points in RS that guarantees poisedness

of the interpolation problem from ~~, s ~ 2, has been studied in [3] in
terms of generalized Vandermonde determinants. If we consider a univariate
spline space V, then the poisedness of the interpolation problem is com-
pletely characterized by the Schoenberg-Whitney condition, which essentially
says that each sample point xi lies in the interior of the support of a
lSuported by the U. S. Army Research Office under Contract No. DAAG

29-84-K-0154.
1,2Supported by the National Science Foundation under Grant No. DMS-86-2337.
31
B-spline which it may claim only to itself. In [2], it was pointed out that
the Schoenberg-Whitney condition does not directly generalize to the space
of bivariate linear splines on a triangulation, and in fact, the results in
[2] also seem to indicate that a poisedness condition which is both
necessary and sufficient is probably very difficult to obtain.
This paper is devoted to the study of interpolation by bivariate
C1 quadratic splines. There are immediately two obvious obstacles. First,
it is well known that the dimension is unstable with respect to the geometry
of the grid partition [6], and second, in general locally supported C1
quadratic splines do not exist. Hence, one must be modest in this study.
In this paper, we consider the criss-cross triangulation 6mn of the unit
square D = [0,1]2 defined by the grid lines

o = Xo < ••• < xm = 1, 0 = Yo < ••• < Yn 1
together with the two diagonals of each subrectangle
.. = [x.1- l'x.]
R1J 1
® [yo l'y,]
J- J
where i = 1, ••• ,m and j 1, ••• ,n. The vector space V will then be the
collection S~ = S~(6mn) = S~(6m,n;D) of all functions in C1(D) whose
restrictions to each triangular cell of the partition 6mn are in n~. The
dimension of this space is
N = dim S21 (6 mn ,D) = mn + 2m + 2n + 3,
and it has a basis consisting of minimal supported functions Bij [5]. In
Fig. 1, we give the B~zier net of each polynomial piece of B.. , using the
1J
notation:
x.-x. 1 y .-y. 1
A. = 1 1- B = J J-
1 xi +1-X i _1 j Yj+l-Yj-l'
A!1 = 1 - A.1 and B~J = 1 - BJ.• Note that the support of B1J
.. is properly
contained in [x.1- l'x·+
1 2
] ® [yoJ -1.,y.J+ 2].
32
o /'
~11
"I
o
../'i" ·' ............. , 0 ,"
I,'
1
o-
,
,A,'~.',-/-I"·'
.-
;'"
1'.., _ _ At.,"'., -
......... /
0
Fi g. 1
In [5J, it was also shown that

L B (x ,y) 1,
i ,j ij
L (-l)i+j(x.-x. l)(Y'-Y' l)B .. (x,y) = 0,

i ,j 1 1- J J- 1J
for all (x,y) in D, and that for each (i ,j ) with supp(B. .) n D * ~,

o 0 10,J O
the collection
(1) .. ) n D *~}
(i,j) * (i o ,j 0 ), supp(B lJ
1
forms a basis of S2.
The main objectives of this paper are to study the location of
sample points that guarantees the poisedness of the problem of interpolation
from S~ and to demonstrate a computational scheme of the corresponding
interpolants. These will be done in Sections 3 and 5, respectively.
Section 2 will be devoted to the discussion of certain smoothing and confor-
mality conditions that will be needed in our study of sample points
placement, and since we are also concerned with the order of approximation
as a result of interpolation at these pOints, we need a quasi-interpolation
formula which will be presented in Section 4.
33
2. Smoothing and conformality conditions
Let <A,B,C> be a non-degenerate triangle in R2 with vertices A,

B, C. We identify each point (x,y) E R2 with the barycentric coordinates
(m,v,w), so that A is represented by (1,0,0), B by (0,1,0), and C by
(0,0,1). In addition, we will express every polynomial p in 1T~ in the
B~zier formulation, namely:
p(x,y) = p(u,v,w) L a 2 ui vj wk ,
i+j+k=2 ijk i!j!k!
where {a ijk } is called the B~zier net of p relative to <A,B,C>. This net
uniquely determines the polynomial. In fact, if D, E, F denote the
mid-points of the edges AB, BC, and CA, respectively, then we have,
a 200 = p(A), a 020 = p(B), a002 = p(C)
1
all O = 2 (4p(D)-p(A)-p(B))
(2) 1
aOll = 2 (4p(E)-p(B)-p(C))
a lOI = 21 (4p(F)-p(C)-p(A)),
(cf. Fig. 2). For brevity, if no possible confusion could arise, we simply
label the above B~zier net as a l , ••• ,a 6 in the order of appearance in (2).
d dlj
a2 2
a_dS d
) c4
D a6 a c6
cs
b
b6 5 b c2
B E 2
Fig. 2 Fi g. 3
34
h3
6·
hS
a2 as
a6
Fig. 4.
There are two types of vertices in the triangulation 6 as shown
mn
in Fig. 3 and Fig. 4. To discuss the conformality conditions around these
vertices and the smoothing conditions across the edges that share these
vertices as common points of intersection, we label the B~zier nets of the
polynomial pieces on all the triangular cells in both figures. For
continuity, the B~zier coefficients of adjacent polynomials at the same
poi nts on the common edges must agree. So, assumi ng that thi s condition is
satisfied, we only discuss C1 conditions.
In the case described in Fig. 3, we have five independent C1
conditions, namely:
(3) a6 + b6 + c 6 + d6 4a 1
a6 + b6 2a4
b6 + c6 + 2b 5
( 4) c6 + d6 2c 4
d6 + a6 2d 5
We call (3) a C1 conformality condition and (4) C1 smoothing conditions.

The situation described in Fig. 4 is more complicated. There are nine
independent C1 conditions, with one conformality condition (either one of
(5) or (6)), and eight smoothing conditions given in (7) and (8).
35
(5) 2Ai Bj e6 + 2AiB'jd6 + 2AiBjh6 + 2AiBja6-Aid2 - Ai a2 = a1

(6) 2A;Bjf6 + 2AiBjg6 + 2AiBjb6 + 2AiBjc 6 - Bjf2 - Bj b2 = a1
2B j h6 + 2Bj a6 - a2 = a5
(7) 2B j e6 + 2Bjd 6 - d2 = d5
2Ai f 6 + 2Aig6 f2 f5
2Aib6 + 2Ai C6 b2 c5
2Ai f6 + 2Aig6 + 2B}6 + 2Bja 6
( 8) 2Ai f 6 + 2Aig6 + 2B j e6 + 2Bjd 6
3. Location of sample points

We now discuss the possible location of the sample points so that
the interpolation problem from 5~ is poised. Two very modest situations
will be discussed. First, after putting sample points at the (m+l)(n+1)
vertices
p .. =
lJ
(x.,y.),
1 J
°<
-
i < m,
-
0 < j < n,
--
where else are we allowed to place the remaining
N - (m+l)(n+1) = m + n + 2
1
sample points? Recall that N = dim 52. Next, suppose that we put
sample points at the centers
prj = (xi + (x i +1-xi)/2, Yj + (Yj+l-Yj)/2)
where i = 0, ••• ,m-1 and j = 0, ••• ,n-1. Where else could we place the
remaining
N - mn = 2m + 2n + 3
s amp 1e poi nts?
We first list two necessary conditions that arise from polynomial
interpolation.
(No!) At most six sample points are allowed in the closure of any
triangular cell, and if six points are placed there, then they
must not lie on any algebraic curve of degree two.
36
(N.2) At most three sample points are allowed to be collinear in the

closure of any triangular cell.
To formulate sufficient conditions, it is enough to show that the
zero function is the only one that interpolates the zero data. Let us first
consider the situation where sample points have been placed at the four
corners of the rectangle Rij and we wish to place four more points in it.
For this purpose, we may refer to Fig. 3 and use the conditions (3) and (4)
to determine the poisedness of the interpolation problem on R.. , where the
1J
dimension of Sl(~l ,R .. ) is eight, which is the same as the number of
2 ,1 1J
sample points. In the following, we list some admissible positions for the
remaining four points that guarantee poisedness.
(5·1) Place these points in the triangular cell <Po 1 .,P .. ,P~ 1.1>
1- ,J 1J 1- ,J-
with one of them on the edge y = y. and the other two noncol-
J
linear with the corner p.1- 1 ,J. or P1.J.. Then place another point
on any of the two edges x = xi _1 and x = xi.
(5. 2) Place one point on the edge x = x. and one on the edge y = y .•
1 J
Then place the other two anywhere in two different triangular
cells of R.. as long as condition (N.2) is satisfied and they
1J
do not both lie on the diagonal with positive slope.
(5·3) Place three points on three different edges of Rij , and place
the remaining one anywhere in R.. so that (N·2) is satisfied.
1J
Of course there are many other admissible positions but we do not
go into details here. We now return to the situation where sample points
have been placed at the (m+1)(n+1) vertices Pij • To state the admissible
locations of the other m + n + 2 sample points, we find it convenient to use
the notion of source and fl ow. Fix an arbitrary rectangle R. . and call it
1oJ 0
the source. From it we may construct two types of fl ows. A flow cannot go
through any vertex P..• The first type flows into the rectangl es R.
1J 10j
and Rij , 1 ~ j < jo' jo < j < nand 1 < i < i 0' i 0 < i < m. The second
o
type also flows away from Riojo but to two opposite corner rectangles,
37
say R11 and Rmn , so that the flow is monotone in the index (i,j). We
pl ace four sample poi nts in the source rectangl e R. . so that condit ions
1 oJ 0
(N.1) and (N·2) and one of the conditions (Sol), (S.2), (S03) are
satisfied. Next, using any flow of the first or second types, place one
point anywhere in each of the rectangles Rij that the flow passes thru,
with the exception of three edges of Rij , one where the flow enters and
the other two adjacent to it. The main tool to verify this result is to
start from the source and apply the conformality conditions (5) or (6) and
smoothing conditions (7) and (8) to go to the other rectangles, first
following the flow, then the rest of the rectangles Rij •
Next, suppose that we have placed sample points at all the centers
Pij of the rectangles Ri +1 j+1. There are many possible admissible loca-
tions for the remaining 2m + 2n + 3 sample points. For instance, we may
first fill the vertices Pio' Pin' = O, ••• ,m and Pi ., p. +1 .,
oJ 10 ,J
j = 1, ••• , n-1 for any i • There are three 1eft whi ch may be placed in
o
R. +1 . +1 for any j , so that conditions (No1) and (N·2) are satisfied
10 Jo 0
there. Another possibility is to fill all the center points Pij as well as
boundary vertex points P., P., P., and p., with the exception of P •
OJ mJ 10 1n mn
There are four remai ni ng ones whi ch may be shared by the four boundary edges
of D (again with the exception of Pmn ).
4. Approximation operators
Suppose that data values at all the grid points Pij and prj can be
measured or approximated. Then we may apply the following approximation
operators to estimate the function values anywhere in D. We set
(9) vmn (f) L f~.
lJ
B..
1J
i ,j
(10) W (f)
mn
where
38
f E C(D). Note that Vmn in (9) is a positive linear operator while Wmn in
(10) is not. We have the following
Proposition 1. V (f) = f for f(x,y) = I, x, y, xy, and W (g) = g for
- mn - - mn
2
!ll g E 11 2 •
Hence, as usual, these operators yield approximation orders of

O(h2) and O(h 3 ) respectively, where
h = ~a~ max{x i -x i _1 'Yj-Yj_1}.
1 ,J
More precisely, by using the standard notion of modulus of continuity under
the supremum norm IIII = II, the following result is an easy
D
consequence of Proposition 1.
Proposition 2.
(i) Ilf-Vmn(f) II ~ W(f ,4 h)
for all f E C(D).
(ii) Ilf-Vmn(f)11 ~ h[W(D(l,O)f,~)) + W(D(O,l)f,~)]
for all fECI (D).
(iii) Ilf-Vmn(f) II ~ {h 2 max IIDafl1

lal=2
for all f E C2 (0).
(iv) I If-Wmn (f) I I ~ i h2[W(D(2,O)f,~) + 2W(D{1,l)f,~) + W(D(O,2)f,~)]
for all f E C2 (D).
(v) Ilf-Wmn(f) II ~ -h h3 max IIDafl1

lal=3
3
for all f E C (D).
Here, the standard notation
O(i,j)f = ai +j f
axiayj
has been used.
39
5. Interpolation schemes
In [1], an interpolation scheme was introduced to take care of the
case when the sample points Pij are used. The approximation order by using
the scheme is 0(h 2 ). Note that the number of sample points used here is
smaller than the dimension of S~. In the following, we demonstrate an
interpolation scheme that takes care of the center points Pij as well as the
remaining 2m + 2n + 3 points. For simplicity, we only discuss the case
where these remaining sample points are placed at Pio' Pin' i = O, ••• ,m,
Pi oJ·' P.10+1 ,J. j = 1, ••• , n-1, and
(Xi +(x i +1- xi )/2, Yj ),

000 0
Pi j = (xi ,Yj +(Yj +1-Yj )/2),

o 0 000 0
and Pi +1 j . In Section 3, we have mentioned that the interpolation
o 0
problem with these N sample points is poised. That is, for any given data
{fr j , f k0 ' f kn ' fiR. ' f i +1 R.' fi j , f i j , f i +1 j }, i = 0, ••• , m-l, j 0, ••• ,

o 0 0 0 000 0
1
n-l, k = O, ••• ,m, and R. = 1, ••• ,n-l, there exists a unique s in S2'
satisfying
S( P~ .) f~ . 0, ••• ,m-l; j O, ••• ,n-l
1J 1J
S(P ko ) f ko k 0, ••• ,m
( 11)
J S(P kn ) f kn
S(P. R.) = f. R.
k
R.
0, ••• ,m
1, ••• ,n-l
10 10
S(P io+1R.) = fio+1R. R. = 1, ••• ,n-l
S(p; j) fi j , S(P; j) fi j , S(Pi +1 j) fi +1 j •

o 0 0 0 0 0 0 0 0 0 0 0
Write
40
m n
(12) S I I
a 00 Bo 0
i=-1 j=-1 lJ lJ
where x_I < xo ' y-l < Yo' xm+l > xm' Yn+l > Yn' and we fix one
coefficient
aO 1
0
2ff 0 - ~(fi 0 +f i +1 0 +f i 0 +1+ f i +10 +1)
1 oJ 0 oJ 0 oJ 0 0 J 0 oJ 0 0 J0
(This is necessary in view of the linear dependence of Bo 0') By using (2)

1J
and all the conformality and smoothing conditions in (2) - (8), we may solve
for
a i j' i 0 -1 2. i 2. i 0 + 1, j 0 - 1 2. j 2. j 0 + 1
in terms of the data there. Similarly, the same conformality and smoothing
conditions are used to compute the other coefficients, recursively, in terms
of the rest of the data values. We have the following recursive formulas.
1 1
+ ~ f; JO +1 + ~ f; +1 j +1
o 0 0 0
1 A
1 1
a; -1 JO - - [2f; j - 2 f; j - 2 f; j +1 - A;oa;ojoJ
o 0 A~ 0 0 o 0 0 0
10
1 1
ai -1 j +1 -~-- [-2B'j +1 f; j + 7 BJ +lfi j
o 0 Ai Bj +1 0 0 0 0 0 0
o 0
+ t(Bj +1+2)f; j +1 - A; Bj +1 ai j +1 J
000 o 0 0 0
41
1 1
a; -1 j -1 =- - [-2Bj f; j 2A; f; j + 2" (Bj +A; +2)f; j +
o 0 A!1 B!J 0 0 0 0000000
o 0
1 1
a; +1 j -1 [-2A{ +If; j -2Bj f; +1 jo + -ZA{ +If; j +
o 0 A; +1 B' ° 0 0 0 0 0 0 0 0
o Jo
and
for k = 2. 3 • •.• •n - JO o·
__
1_ [4f*
BJ'o -k+1 ; j -k+1
o 0 0
42
1 [f. -A' B a -A B'

A' B' '0 jo-k ;0 jo-k+l ;0-1 jo-k+l ;0 jo-k+l a;o jo-k
;0 jo-k+l
- A; Bj _k+l a; j -k+l]
o 0 0 0
for k
a; -k j
o
- A~ k IB a. k ]
'0- + 0 '0- 0
for j O,I, ••• ,n-l, k = 2,3""';0+1,
a; +k j = ___1___ [ 4f t +k-l j - Ai +k_ 1a; +k-2 j - Bj +1a; +k-l j+l

o A; 0+k 0 0 0 0
000
i
- Bja; +k-l j-l - (A; +k-l+ A +k+ Bj+Bj+l)a; +k-l j]
0
,
a.
0
+k n
1
[f; +k n+l
0
A~
0 0
,
+k Bn a. +k-l n
A~ ,
0 0
,
+kB~ a. +k-l n-l ,
A; +kBn
0
- A; +k B~ a; +k n-l]
0 0
43
1
ai +k 1 = [f i +k
I 0 Ai +kBo ai +k-l 0 Ai +kB~ai +k-l -1 -
o Ai +kBo 0 0 0 0 0
o
- Aio+kBo aio +k 0]
By using Proposition 2, we have the following:
Theorem. Let f £ C3(D) and the data values.i!:!.. (11) be defined Ex. the
functional values of f at the corresponding sample points. Then the spline
s in (12) that satisfies the interpolation conditions (11) ~ unique and
approximates f uniformly with error:
lis - fll ~ (: ~ h2 + b 3) lal=3

max IIDaf11
wh e re k = mi n{ mi n ( x. -x. 1 ,y . -y. 1) : i = 1, ••• , m, j =1 , ••• ,n }•

1 1- J J-
REFERENCES
1. Chui, C. K., H. Diamond, and L. A. Raphael, Interpolation by
bivariate quadratic splines on a nonuniform recangular grid, Trans.
Fourth U.S. Army Conference on Appl. Math. and Compo (1986), To appear.
2. Chui, C. K., T. X. He, and R. H. Wang, Interpolation by bivariate
linear splines, in Proc. Conf. in Memory of Haar, North-Holland, To
appear.
3. Chui, C. K. and M. J. Lai, Vandermonde determinants and Lagrange
interpolation in RS , in Nonlinear and Convex Analysis: Proc • .i!!. Honor
of !t Fan, Marcel Dekker, N. Y., 1987.
4. Chui, C. K. and R. H. Wang, On a bivariate B-spline basis,
Scientia Sinica, Vol. ~ (1984), 1129-1142.
5. Chui, C. K. and R. H. Wang, Concerning Cl B-splines on

triangulations of nonuniform rectanglar partitions, Approx. Theory
and Its Appl., Vol. 1 (1984), 11-18.
6. Morgan, J. and R. Scott, The dimension of the space of Cl piecewise

polynomials, ms.
7. Sha, Z., On interpolation by S21(~(2)),

m,n Approx. Theory and Its
Appl. Vol. l (1984), 71-82.
Charles K. Chui and T. X. He, Department of Mathematics, Texas A&M

University, College Station, Texas 77843, U.S.A.
© 1987 Birkhiiuser Verlag Basel
APPROXIHATION VON ECKENSINGULARITATEN BEI RANDvlERTAUFGABEN

L. Collatz, Hamburg
Abstract: There are different possibilities of the approximation

of singularities at corners occuring with elliptic boundary value
problems in the plane. Numerical experiences are mentioned.
Aufgabenstellung und Inhalt: Monotoniesatze sind bei Randwertauf-

gaben von linearen und nichtlinearen elliptischen Differential-
gleichungen zur EinschlieBung der Losungen, also zur Aufstellung
garantierbarer Fehlerschranken verwendet worden. Dabei sind sin-
gulare Stellen des Randes, Ecken, Unstetigkeiten in den vorgege-
benen Randwerten u.a. besonders zu berlicksichtigen. Die Wahl von
Systemen geeigneter Ansatzfunktionen ist dabei wichtig. Als Bei-
spiele werden zweidimensionale Aufgaben der Potentialtheorie flir
Bereiche mit Ecken bei verschiedenen Eckenwinkeln betrachtet und
es wird liber einige numerische Erfahrungen berichtet.
1. Die Approximation
Die hier betrachtete algorithmische Berechnung der Naherungslo-
sungen beruht auf Approximation und Optimierung (vgl. z.B. Mei-
nardus [67], Meinardus-Herz [82], Watson [80]u.a.). Man approxi-
miert die gesuchte Funktion u(x 1 , •.. ,x n ) oder kurz u(x) in einem
gegebenen Bereich B des Rn durch Funktionen ~(x,a1, ••• ,ap) einer
von Parametern a \) abhangenden Funktionenklasse ~(x,a), wobei die
.
Approximation als Optimierungsaufgabe fli~ die a\) geschrieben und

so der bequemen Behandlung auf einem Computer zuganglich wird.
vlenn Honotoniesatze gelten, ist es h2:ufig in nicht zu komplizier-

ten Fallen moglich und numerisch praktikabel,flir die gesuchte Lo-
sung u EinschlieBungen zu berechnen, d.h. Naherungsfunktionen v
und v zu ermitteln, so daB die Aussage
( 1) v < u < v
-
garantiert werden kann. Diese Methode ist an sehr verschiedenar-
tigen Problemklassen erprobt worden (vergl. z.B. Bohl [74],
Schroder [80], Collatz [68][81][85] u.a.).
45
Besondere Aufmerksamkeit beanspruchen etwaige Singularitaten,

Zwar sind einige Arten vie I untersucht worden, z.B. Unstetigkei-
ten in den vorgegebenen Randwerten, Ecken mit rechten Winkeln,
Spitzen, u.a., jedoch kommt man bei geringen Xnderungen des Typs
der Singularitat zu Fallen, bei denen die Theorie noch nicht ge-
nUgend ausgebaut ist oder bei denen man noch viel numerische Er-
fahrung sammeln mUBte, ehe man zu empfehlenswerten Vorgehenswei-
sen kommt.
2. Das Monotonieprinzip.
Dieses lautet in einem sehr einfachen FaIle: Im m-dimensionalen
punktraurn Rm der Punkte x=(x 1 ' ••• ,xm) sei ein zusammenhangender
Bereich B mit stUckweise glattem Rand dB gegeben. Mit
m 2
6u = ~ ~
j=1 dx. 2
J
und gegebenen Funktionen f(x) und g(x) werde fUr Funktionen u
einer Klasse K:
der vektorwertige Operator Tu gebildet :
(2) Tu=f- 6U ~ f(x) in B

lu - g(x) auf dB
Dann gilt fUr zwei beliebige Funktionen v,w aus K:
(3) Aus Tv > Tw folgt v > w in B U dB.
Dabei bedeute das Ungleichheitszeichen die klassische Ordnunq re-

eller Zahlen und fUr Funktionen gelte die Ungleichung punktweise
jeweils im ganzen betrachteten Bereich und bei T fUr jede Kompo-
nente. Nun werde angenommen, daB eine Losung u von
(4) Tu = 0 d.h. -6u = f(x) in B, u = g(x) auf dB
existierti dann folgt aus
(5) Tv < 0 < T~
die EinschlieBung (1).

46
3. Numerisches Verfahren
Es werden NaherungslBsungen v und v gesucht mit -
p q
{-
v ~a + L a v tj)v(x) + L b 1/!lJ(x)
(6) v=1 11=1 J!:
p q
v = va + L a v tj)v(X) + L blJ 1/!lJ (x)
v=1 11=1
Dabei sind v o,v o,tj)v,1/!lJ Funktionen der Klasse K, und die Parameter
a V ,a V ,b l1 ,b l1 sollen so bestimmt werden, daB (5) und damit (1)
gilt. Die Durchrechnung auf einem Computer erfolgt mit Hilfe der
Optimierung
(6) o < -
v-~ ..:: cS, IS = Min.
Hierbei wird man in der Regel, so fern die Differentialgleichung

inhomogen ist, eine spezielle LBsung Vo der inhomogenen Gleichung
durch va bzw. va approximieren und als tj)v regulare und als 1/!11
singulare LBsungen der Differentialgleichung verwenden.
4. Ecken-Singularitaten
Als solche sollen insbesondere in der Ebene Ecken mit dem Ecken-
winkel a betrachtet werden. Zahlt man Polarkoordinaten r,tj) von
der Ecke P aus, so bieten sich die Potential-Funktionen
(7) 1/! =rk1T/aSin(k~ .,) (k=1,2, ••• ,
a < a ..:: 21T)
an, welche auf den Tangenten an die Rand-

kurve im Punkte P verschwinden, Fig.1: man
hat "ausspringende" Ecken fUr a<1T und "ein-
springende" Ecken fUr a>7T. FUr a=1T ist die
Tangente stetig, fUr a=a tritt eine Spitze auf, und ein "Schlitz"
entspricht a=21T.
Die Funktion 1/! von (7) ist regular fUr

a = 1TIs fUr s = 1,2,3, •••
da sich in diesen Fallen 1/! auf ein Polynom reduziert.
47
Das besagt aber noch nicht, daB die Losung u an einer Ecke mit
Winkel a=n/s regul~r ist. Ein bekanntes Gegenbeispiel ist das
Torsionsproblem fUr einen Tr~ger von quadratischem Querschnitt
B={(X,y), ixi<1, iyi<1}, -~u=1 in B, u=O auf aB, wo im Punkt
P=(1,1) gilt uxx=O, Uyy=O also uxx+uyy=O; also ~u ist nicht ste-
tig bei an den Punkt P, Fig. 2. Ein ebenfalls sehr
,'1
Ann~herung
f:'ig.2
bekanntes Beispiel, allerdings bei der W~rmeleitungsgleichung,
ist LU=Ut-U =0 in B={(x,y), O<x<oo, O<t<oo} mit den Anfangs- und

xx
Randwerten:[U(X,O)=O (Anfangstemperatur im Stabe fUr O<X<OO sei Null
\U(O,t)=t (Stabende x=O wird proportional zur Zeit t
erw~rmt) .
Lu ist nicht stetig bei Annaherung an den Punkt (0,0) vom Be-
reichsinneren her, Fig. 3.
Man kann folgende Wege der numerischen Behandlung einschlagen:

1. Man berUcksichtigt bei jeder Ecke einen Term der Form (7), oft
wenigstens fUr k=1.
2. Man verwendetTerme der Form (7) nur fUr einspringende Ecken,
und l~Bt insbesondere bei geringen GenauigkeitsansprUchen
Terme der Form (7) bei ausspringenden Ecken fort.
3. Die Erfahrung hat gezeigt, daB es oft zweck~Big ist, auch
hohere singul~re Terme (k>1) zu berUcksichtigen, etwa fUr a=3~
auch k=1,5,7, ••• und fUr a=2n auch k=1,3,5 ••.
Das entspricht der Mitnahme weiterer Glieder in der Reihenent-
wicklunq nach Potenzen von zn/a mit z = rei~ = x + iy.
4. In der klassischen Arbeit von Kondrat'ev [67] hat Kondrat'ev

empfohlen, auch Glieder der Form
(8) zll(R,nz)V mit konstantem Il,\)
zu benutzen.
48
Generell gilt fUr diese verschiedenen Moglichkeiten, daB im

FaIle von Monotonie jede der genannten Moglichkeiten zu einer
garantierbaren Fehlerabsch~tzung fUhrt,und daB es nur eine
Frage des Ermessens und damit der numerischen Erfahrung ist,
mit welchem Grade von Genauigkeit man zufrieden sein will
und welchen Rechenaufwand man fUr vertretbar h~lt.
Andere Singularit~ten, wie z.B. Unstetigkeiten in den Randwerten,

verlangen andere Typen singul~rer Funktionen, z.B. logarithmi-
sche Terme bei Ringbereichen und dergl. (numerisches Beispiel
vgl. Nr. 5.F)).
Ungleich schwieriger und noch keineswegs genUgend theoretisch

erforscht sind r~umliche Singularit~ten, schon im R3 , wo auch
das Mittel der Funktionentheorie nicht mehr zur VerfUgung steht,
vergl. z.B. Arbeiten von Whiteman [85]. Nr. 5 E).
Hier solI Uber einige numerische Erfahrungen, insbesondere mit

Winkeln a, die nicht ganzzahlige Vielfache von n/2 sind, berich-
tet werden.
5. Numerische Erfahrungen
A) Rechte Winkel bei ein- und ausspringenden Ecken
Hier sind zahlreiche F~lle mit gutem Ergebnis durchgerechnet
worden
a) Das Torsionsproblem fUr einen Tr~ger vom
Querschnitt B verlangt in der x-y-Ebene eine
II I-6V-r I~
Funktion u zu ermitteln, die der Randwertauf-
gabe
(9) -l1u = 1 in B
Ft,.11- 18
(10) u=o auf dem Rande aB qenUgt.
Beispiele: Von Collatz wurde

als Bereich der Querschnitt
eines U-Tr~gers, Fig.4 behandelt
49
b) Ein anderes Beispiel ist die stromung urn eine Ecke, Fig. 5,
vgl. Collatz [gO].
B) Rechte und spitze Winkel

FUr das Torsionsproblem mit (9) (10) ist in Fig. 6 als Bereich B
ein Sechseck mit zwei einspringenden Winkeln der GroBe ~ und
vier ausspringenden Winkeln der GroBe igewahlt~ die 6pitzen
Winkel der GroBe 1brauchen nach dem oben gesagten nicht berUck-
sichtigt zu werden~ die Winkel bei P. mit a.=~ werden mit den
J J,(,
Funktionen ~j nach (7) erfaBt, wahrend die antisymmetrischen
Funktionen mit e2 =j, ••• auszulassen sind. Als Beispiel seien die
Fehlerschranken 0 beim Rechnen mit p regularen und q singularen
Funktionen angegeben. (Ich danke Herrn Dipl. Hath. Uwe Grothkopf
fUr die DurchfUhrung der numerischen Rechnung auf einem Computer).
(Man sieht die starke Verbesserung durch die singularen Terme) •
Werte von 0 (fUr !w-ul~Q in B)
mit Anzahl q der singularen Terme
p Polynomen q = 0 q"2 ~1T = ~ bzw .!.Q
vom Grad n ex 3 • 3
1 m= 0 2 0.3784
4 m 6 0.5259 0.0152
Bei sehr kleinem Winkel a wird man bei der Randbedingung (10)
erwarten, daB in der Umgebung der Ecke auch nur sehr kleine Wer-
te u auftreten werden, da die Tangentialebene dort u=O lau~
tet. bei geringen GenauigkeitsansprUchen wird man daher vielleicht
die Singularitat unberUcksichtigt lassen, bei hoheren Genauig-
keitsansprUchen aber sollte man die Singularitat erfassen.
C) Sehr kleine \'linkel a

Es werde hier als Beispiel eine Eigenwertaufgabe genannt (vgl.
z.B. r~ethoden bei Gorisch-Albrecht [84]). Erfahrungen mit kleinen
Winkeln wurden von Gorisch-Zimmermann [86] bei Berechnung der
Eigenschwingungen einer am Rande eingespannten rhombischen Mem-
bran mit Grundbereich B gemacht
mit B = {(x,y),lyl«1-lxl)tan (a)}, a=1T/24 (oder 7.5°) Fig.7
Hier lautet das Problem
50
(11) -~u(x,y) = AU(X,y) in B, u=O auf aBo

Die Autoren naherten mit Polynomen an,
beginnend mit r12cos(12~), also einem
polynom 12. Grades; dabei sind die Polar-
koordinaten von der Ecke P 1 =(-1,0) ge-
zahlt. Bei der Ecke P 2 =(0,tan(n/24) tritt der Eckenwinkel
8=11n/12 auf; n/8 ist nicht ganzzahlig; obwohl die entsprechenden
Singularitaten-Funktionen Wvon (2) nicht berUcksichtigt wurden,
sind die numerischen Resultate sehr gut, weil Polynome sehr hohen
Grades verwendet wurden; z.B. wurden die Schrank en fUr den ersten
Eigenwert A1
195.804 ~ A1 ~ 195.890
angegeben und analog sehr gute EinschlieBungen fUr A2 und A3 •
D) Erfahrungen mit Winkeln a

n
mit '2 < a < n und
wurden mit einem Bereich B nach Fig. 8 gesammelt. Man kann das
Beispiel als Torsionsproblem interpretieren; entstanden ist es
aus dem Temperaturproblem fUr einen dreidimensionalen Raum nach
Fig. 8, der im Innern gleichmaBig erwarmt wird mit -~u=1 und am
ganzen Rande die Temperatur u=O hat.
Mit den in Fig. 8 angegebenen Abmessungen

wurde zur Berechnung einer Naherung w
(auier Polynomtermen wie bei B) zunachst
nur die einspringende Ecke bei P mit Po-
larkoordinaten r und ~ berUcksichtig und
die ausspringendeEcke bei Q noch nicht
weiter beachtet. FUr die Funktionen
in (7) wurde
kn 4 8 12 16 24
-a = 5'5'~'~'~'···
verwendet. FUr die DurchfUhrung der Rechnung auf einem Computer
danke ich Herrn Eberhard Greiner.
51
Fur die Fehlerschranke 6 wurden die Werte erhalten:

Werte von 6 mit Iw-ul<6 in B
An:zahl der An:zahl der singularen Terme
Pol nomterme o 5
1 2,25 1.694
3 0.5625 0.1736
11 0.05188 0.02363
Die Experimente :zeigten aber, daB die Singularitat beiQ der GroBen-
ordnung nach einen ungefahr ebenso groBen EinfluB wie die Singu-
laritat bei P hat. Gerechnet wurde mit einer Naherungsfunktion
WI':m mit
n m t
( 12) w L cvPv + L bvv v + L aAw A
v=1 1l=1 ~=1
P1 1; P2r = Re(x+iy)r; P2r+1 = 1m (x+iy)r (r=1,2, ••• ) I
Dabei sind entsprechend Fig. 8 Polarkoordinaten r1'~1' von Q und

r2'~2' von P aus gezahlt. Der Ansatz ist also durch das Zahlen-
tripel (n,m,i) charakterisiert und es wird nun beschrieben, wie
die Verbesserung in der n-Richtung, in der m-Richtung und in der
i-Rich tung erfolgt. Die Ubergange werden symbolisch durch die
pfeile gekennzeichnet:
(n-2,m,i) ~
(n,m-1 , i) _ (n,m, i)
(n,m, i-1)----
6(n,m,i) bezeichnet die Fehlerschranke beim Rechnen mit dem Tri-
pel (n,m,i) fur das Beispiel (3,1,5) sind die entsprechenden
Fehlerschranken angegeben.
6(1,1,5) 0.7773 ...............
6(3,0,5) = 0,1736_ 6(3,1,5) 0.08456
15(3,1,4) 0,2018./"
52
Man sieht, daB eine Hinzunahme sowohl bei n, als aueh bei m oder
1 einen merkliehen EinfluB ausUbt, am st~rksten bei n. Aber der
EinfluB der Singularit~t bei Q ist ann~hernd so groB wie der bei
P.
E) Wie eingangs erw~hnt, mUBte man noch sehr viel mehr Erfahrun-
gen sammeln, sowohl bezUglieh der Theorie als aueh der Numerik,
insbesondere bei dreidimensionalen Singularit~ten und bei nicht-
linearen Problemen. (Versehiedene Arbeiten von Whiteman, z.B.
[85], Tolksdorf [83], Collatz [8Sa] u.a.).
F) Ein Beispiel eines ringfBrmigen Bereiches Fig. 9, bei welehem

verschiedene Arten von Singularit~ten auftre-
ten, wird bei Collatz-Grothkopf-Hayman [87]
durchgefUhrt.
G) Die von Kondrat'ev [67] empfohlene Ver-

wendung von Real- und lmagi~rteilen der
holomorphen Funktion (8) mit passenden ~
und v mUBte noeh ausgiebig erprobt werden.
Auch hier erBffnet sieh noeh ein weites Gebiet wissensehaftlicher
Forschung.
53
Literatur
Bohl, E. [74] Monotonie, Losbarkeit und Numerik bei Operatorglei-
chungen, Springer, 1974, 255 S.
Collatz, L. [68] Funktional Analysis und Numerische ~~thematik,
Springer, 1968, 371 S.
Collatz [80] Numerical treatment of some boundary-value problems
with hidden singularities, Proc. Congress on Approx. Th.III,
Acad.Press,1980 J 311-320.
Collatz, L. [81] Anwendung von Monotoniesatzen zur EinschlieBung
der L5sungen von Gleichungen; Jahrbuch Uberblicke der Mathe-
matik 1981, 189-225.
Collatz, L. [85] Einige Anwendungen der mehrdimensionalen Approxi-
mationstheorie zur LosungseinschlieBung bei RandwertaufqaPen,
Proc. Equadiff 6, Dep. of Math. Univ. Brno (1985), 367-372
Collatz, L. [85a] Inclusion of regular and singular solutionsof
certain types of integral equations, Intern.Ser.Num.Math.73
(1985) 93-102.
Collatz, L. - U. Grothkopf - W.K. Hayman [geplant 87] Modell ei-
ner Randwertaufgabe einer Pipeline mit verschiedenen Arten
von Singularitaten.
GQrisch, F. - J. Albrecht [84] Eine einheitliche Herleitung von
EinschlieBungssatzen fur Eigenwerte, Intern.Ser.Num.Math.69
(1984) 58-88.
Gorisch, F. - S. Zimmermann [86] On Trefftz's Method and its
application to Eigenvalue Problems, Z• Angew.Math.loo1ech. 66
(1986) T-304-~306.
Kondrat' ev [67] Berichte der Akademie der Wissenschaften, ~loskau,
(1967) •
Meinardus, G. [67] Approximation of functions, Theory and numeri-
cal methods, Springer, 1967, 198 p.
Meinardus G. - G. Merz [82] Praktische Mathematik I, B.I. Wissen-
schaftsverlag (1982) 491 s.
Schroder, J. [80] Operator inequalities,Acad.Press (1980), 367 p.
Tolksdorf, P. [83] Vortrag Tagung: Singularities and Constructive
Methods for their treatment. Proc. Oberwolfach 21.-26. Nov.
1983.
Watson, G.A. [80] Approximation Theory and Numerical Methods
John Wiley 1980
Whiteman, J.R. [85] Singularities in two- and threedimensional
elliptic problems and finite element methods for their
treatment, in Proc. Equadiff6, Purkyne Univ, Dep. of Math.
Erno 1985.
Prof. Dr. L. Collatz

Institut fur Angewandte ~~thematik
der Universitat Hamburg
BundesstraBe 55
D-2000 Hamburg 13
©1987 Birkhauser Verlag Basel
PERIODIC AREA MATCHING INTERPOLATION
Franz-Jurgen De1vos
o. Introduction
Anse10ne and Laurent [1 J , Schoenberg [1 2J , and

de Boor [2 J have given interesting applications of splines
to smoothing of histograms by matching the integral means of
splines between successive knots with the same means of a given
function. Subbotin [14J has also considered integral means
of cardinal splines ( see also [7 J ). These investigations
have been continued in a recent paper by Sheva1din [13J. It
is the purpose of our paper to study the corresponding problem
for classes of periodic functions which are constructed from
translates of a generating function with respect to a uniform
mesh. We apply our method to periodic splines of arbitrary de-
gree and trigonometric polynomials.
55
1. Spaces of translates
Let g e C2n have an absolutely convergent Fourier se-
ries, i. e.,
g(t)
with
1 2n 'kt
2n f g(t) e-~ dt ( k e 2Z
o
00
Moreover, we assume that g is real valued, i. e.,
ke2Z) •
Let n e :N be fixed and
j e 2Z )
define a uniform mesh. We relate the functions
n-1
L g(t-t,) eiktj (O<k<n)
j =0 J
and assume
( 0 < k < n )
In the papers [ 4 , 5] we have introduced the n-dimensional
space of periodic functions
= < 1 ,g( .-t 1 )-g, ••• ,g( .-t n _ 1 )-g >
which is translation invariant with respect to t1 [ 4 ]
Since V~(g) = <1,b 1 , ••• ,bn _ 1 > the function M defined by
n-1
(1 .1 ) M(t) (1 + L bk(t)/bk(O))/n
k=1
56
is an element of
1
V (g).
n
We have shown [ 4 J that M is a
periodic Lagrange function of V~(g), i. e.,
(1 .2 )
°0 ,J. O:;;j<n)
This property of M and the translation invariance of V~(g)

with respect to tl imply that we can define the interpolation
projector Sn on C2TI
n-l
(1 • 3 ) L f(t.) M(t-t.)
j =0 J J
Note that Sn(f) is the unique function in V~(g) with
(1•4) S (f)(t.) f (t. ) (O:>j<n)

n J J
For more information on interpolation by translation we refer to
thearticles [4,5,6,8J.
2. Periodic Histopolation
In this section we make the additional assumption that
the derivative of g is a periodic function of bounded varia-
tion. Then we have
h(t) := g' (t) ( t e ]R
In particular, h e L 2TI • We introduce the area matching func-
tionals
t.J+ 1
a. (f)
J
f f(t) dt (O:;;j<n)
tj
which are defined for any f e L2TI A function u e L2TI

57
is called an area matching interpolant of f e L 2n (also called
histopolant of f if u satisfies the the area matching
conditions
o :ii j < n
As a main result we will establish the following theorem on pe-
riodic area matching interpolation which is now called perio-
dic histopolation.
Theorem 1
For any f e L 2n there exists a unique histopolant
Hn(f) in the space V~(h) satisfying the area matching inter-
polation condition
(O:s;j<n)
We will carry out the proof by an explicit construction
of Hn(f) via the method of interpolation in V~(g) as de-

scribed in section ( see also [4, 5J ). For a given func-
tion f e L 2n we define the periodic integral of f by
t
J(f)(t) f f(s)ds - cO(f) t ( 0 :ii t < 2n )
o
Note that J(f) has a continuous periodic extension which
possesses the following properties
(2.1 ) J(f)(O) = 0 J(f)(2n) = 0
(2.2) DJ(f) (t) = f(t) - cO(f) a. e.

58
We claim that the periodic histopolant Hn(f) is given by
(2.3) u(t)
Lemma
The function u satisfies the area matching conditions
(2.4) a. (u)
]
= a. (f)
]
(O::ij<n)
Proof: Taking into account (1.4) and (2.2) we obtain
(tj + 1 -t j ) Co (f)
+ S (J(f))(t. 1) - S (J(f))(t.)
n ]+ n ]
tj +1 tj
+ f f(t) dt f f(t) dt
o o
t.]+ 1
f f(t) dt
tj
whence (2.4) follows.
Lemma 2
The histopolant u of f is an element of the function
space V~ (h) •
Proof: Since M e V1(g) we have the unique representation

n
n-1
(2.5) M(t) = YO + L y.(g(t-t].) -get))
j =1 J
59
Let N(t) = M'(t) Then we have

n-1
(2.6) N(t) L Yo (h (t-t 0 ) - h (t) )
j =1 J J
which implies
o :0 k < n )
Using formulas (2.3) and (1.3) we can conclude
n-1
(2.7) u(t) cOIf) + L J(f) (to) N(t-to)
j =1 J J
in view of J (f) (to) = 0 • This shows u e vn1 (h) which com-
pletes the proof of Lemma 2
Lemma 3
Let P e V~(h) be the function
n-1
1
L
0
(2.8 ) PIt) + (1 _l)N(t-to)

21T j =1 n J
Then we have
(2.9) °0 ,J 0
(O~j<n)
Proof: We use formula (2.7) to construct P . It follows
from (2.9) that
cO(P)
1
J(P)(to) = 1 - i ( 0 < j < n )
21T J n
Now Lemma 3 follows from (2.7) and Lemma 1 .

Lemma 4
The periodic histopolant H (f)

n
of f in V1 (h)
n V~(g' ) is
given by the formula

60
n-1
(2.10) L a. (f)P(t-t.)
j =0 J J
Proof: Note first that the translation invariance of V~(h)

with respect to t1 implies
O:;;j<n)
It follws from relation (2.9) that
(2.11 ) O.J , k O~j,k<n)
i. e. , P ( • -t .) , 0 :;; j < n , is a dual basis of the area mat-

J
ching functionals a k , 0 :;; k < n , in V1 (h) This implies
n
equality (2.10) and
Theorem 1 follows now from Lemmas 1 - 4 .
3. Trigonometric Histopolation
For simplicity we consider first the case n = 2m+1 . We
have shown [5J that the assumptions of section 1 are satis-
fied by the function

m
(3•1 ) g(t) L cos(kt)
k=1
The space V~(g) is just the n-dimensional space of tri-
gonometric polynomials of order m. Moreover, we have
( 3 .2) T
m
Thus, Theorem 1 is applicable .

61
Corollary 1
For any f e L2~ there is a unique trigonometric polyno-
mial Hn(f) e Tm ( n = 2m+1 ) satisfying the area matching
conditions
tj +1 tj +1
( 3.3 ) fH(f)(t)dt f f(t) dt (O:oj<n)
t. n t.
J J
Next we discuss the case n 2m. Then the function g
is given by
m-1
( 3.4 ) g(t) L cos(kt) + cos(mt)/2
k=1
The related function spaces V~(g) and v n1 (h) have the form
(3.5) V~(g)=<{COS(ko):k:om}> + <{sin(ko):k<m}> -. T~
(3.6) V~(h) = <{cos(ko):k < m}> + <{sin(ko):k:o m}> -. T~
Again, an application of Theorem 1 yields the
Corollary 2
For any f e L2~ there is a unique trigonometric polyno-

mial satisfying the area matching
conditions
(3•7 ) (O:oj<n)
62
4. Periodic Spline Histopolation
Again we consider first the case n = 2m+1 • As the gene-
rating function 9 we choose the shifted Bernoulli function
( 4.1 ) g(t) ( 0 ;:;; a < 1 q i: 2 )
The function space V~(g) is the space of periodic splines of
order q-1 wi th spline knots tj - a t1 ' j e :ll
(4.2) Sp(q-1,n,a)
( see [9,4J). Recall that
(4.3) B (t)
q
L (ik) -q exp (ikt)
k~O
(4.4) B' (t) Bq _ 1 (t)

q
This implies
(4.5) V 1 (h) Sp(q-2,n,a)

n
It follows from the interpolation theorem of ter Morsche [11J
( see also [3, 10J ) that the assumptions for 9 and h = g'
are valid. Thus an application of Theorem 1 yields the
Corollary 3
For any f e L 2n there is a unique periodic spline func-
tion Hn(f) e Sp(q-2,n,a) (n = 2m+1 ) satisfying the area
matching conditions
tj +1
(4.6) f H (f)(t) dt ( 0 ;:;; j < n )
t. n
J
63
We briefly list the modifications which have to be made
for the case n = 2m Taking into account the interpolation
theorem of ter Morsche [11J it follows that Corollary 3
remains valid for n = 2m in each of the following cases:
(4.7) q is even and a" 1 / 2
( 4.8) q is odd and a" 0
Finally we consider a more general class of function g
which are given by

00
(4.9) g(t) 2 ak cos(kt)/k

k=1
with
00
(4.10)
Moreover, it is assumed that

00
(4.11 ) h(t) - 2 a k sin(kt)

k=1
is a function of bounded variation Taking into account [5 J ,
Proposition 9, it follows that the assumptions for g and h
are valid. Thus, Theorem 1 holds for this class of functions
V1 (h) with no restriction on n.

n
Special cases are the following functions

00
h(t) 2 k- 2 r+1 sin(kt) ( r e :IN

k=1
t sin (t) / (cosh (b) -cos (t) )

00
-kb
h(t) - 2 e sin(kt) =-
k=1
( b > 0 )
64
Fig.1 The area matching fundamental function P e Sp(2,5,O)
Fig.2 The area matching fundamental function P e Sp(3,5,O)
Fig.3 The area matching fundamental function P e T2
References
P.M. Anselone and P.-J. Laurent A general method for

the construction of interpolating or smoothing splines,
Numer. Math. 1£(1968), 66-82 .
2 C. de Boor: Appendix to : I.J. Schoenberg, Splines and

histograms, in "Spline Functions and Approximation Theory"
( A. Meir and A. Sharma, eds. ), ISNM £l , 329-358 ,
Birkhauser Verlag, Basel, 1973 .
3 L. Collatz und W. Quade: Zur Interpolationstheorie der
reellen periodischen Funktionen, Sitzungsberichte der
PreuBischen Akademie der Wissenschaften 1Q(1938),383-429.
65
4 F.-J. Delvos: Periodic interpolation on uniform meshes,

Journal of Approximation Theory ( to appear )
5 F.-J. Delvos: Interpolation of odd periodic functions on

uniform meshes, in "Delay Equations, Approximation and
Application" ( G. Meinardus, G. Nurnberger, eds. ),
ISNM li, 105-121 , Birkhauser Verlag, Basel , 1985 .
6 F.-J. Delvos: Bernoulli functions and periodic B-splines,
Computing ( to appear ) •
7 F.-J. Delvos: Limits of abstract splines in "Inverse

and Ill-posed Problems" H. W. Engl and C. W. Groetsch,
eds. ), Academic Press, New York, 1987 ( to appear) .
8 F. Locher: Interpolation on uniform meshes by translates

of one function and related attentuation factors ,
Mathematics of Computation 11(1981), 403-416 .
9 G. Meinardus: Periodische Spline-Funktionen, in "Spline-

Functions, Karlsruhe 1975" ( K. Bohmer, G. Meinardus ,
W. Schempp, eds. ), Lecture Notes in Mathematics 501 (1976),
177-199
10 G. Merz Interpolation mit periodischen Spline-Funktionen

I, Journal of Approximation Theory lQ(1980), 11-19 .
11 H. ter Morsche On the existence and uniqueness of inter-
polating splines of arbitrary degree, in "Spline-Funktio=
nen" ( K. Bohmer, G. Meinardus, W. Schempp, eds. ),
Bibliographisches Institut, Mannheim, 1974 .
12 I. J. Schoenberg: Splines and histograms, in "Spline

Functions and Approximation Theory" ( A. Meir and A. Sharma,
eds. ), ISNM.£1.., Birkhauser Verlag, Basel, 1973 •
13 V. T. Shevaldin Some problems of extremal interpolation
in the mean for linear differential operators, Trudy
Mat. Inst. Steklov 164(1983), 233-273 ; English trans-
lation in Proc. Steklov Inst. Math. 164(1985)
66
14 Yu. Subbotin: Extremal problems of functional interpola-

tion and splines for interpolation in the mean, Trudy
Mat. lnst. Steklov 138(1975), 118-173, English translation
in Proc. Steklov lnst. Math. 138(1975) .
Franz-Jurgen Del v 0 s
Lehrstuhl fur Mathematik I
Universitat Siegen
Holderlin-Str. 3
D-5900 Siegen (w.-Germany)
AN EXPOSITION OF RECENT RESULTS ON CONTINUOUS METRIC SELECTIONS

Frank Deutsch
The Pennsylvania State University,
University Park, PA., U.S.A.
1. Introduction
The question of when a metric projection has a continuous selection has

received much attention in recent years. There has been a flurry of activity
in this area by mathematicians from allover the world. (At this writing, I
am aware of relevant work by researchers located in over 17 different
countries.) In spite of all this interest, or perhaps because of it, there
are still many questions which remain unanswered. The writer gave a survey of
continuous selections at a special meeting of the ~erican Mathematical
Society in January, 1982 (which appeared as [9]). At an Oberwolfach
conference a few years later, NURNBERGER and SOMMER [22] concentrated their
attention to the finite-dimensional subspaces in C[a,b] and gave a detailed
description of the results available in this setting. Still later, the writer
[10] announced some recent progress that had been made since the appearance of
[9] and [22].
It is the purpose of my talk today to announce the progress that
has been made mainly in the last two years, and to point out some of the open
problems which still remain. Many of these results have not yet appeared in
print. In a sense, this report may be regarded as an updating and supplement
to the papers [9], [10], and [22]. It is also hoped that this will provide
additional motivation for others to take up these problems.
2. Metric Selections
Let X be a normed linear space and M a finite-dimensional subspace. The

metric projection onto M is the (generally set-valued) mapping
PM : X ~ 2M defined by
68
(y E M I IIx-yll d(x.M)}
where d(x.M) = inf {IIx-yll lye M}. That is. PM(x) is the set of all
"best approximations" or nearest points to x from M. It is well-known and
easy to verify that PM(x) is a nonempty compact convex set for each x. M
is called a Chebyshev subspace if PM(x) is a singleton for each x. A
metric selection for M is any selection p for PM' That is.
p : X ~ M and p(x) E PM(x) for each x E X. (Metric selections are also
called "proximity maps" by some authors.)
Metric selections always exist by the Axiom of Choice. We will
only be interested in determining when metric selections exist which are
continuous or satisfy even stronger conditions. One reason for wanting to
know this is related to the construction of algorithms for determining best
approximations. For example. if it is known that continuous metric
selections do not exist for M. then every algorithm for determining best
approximations from M (is a metric selection for M and hence) is
unstable. On the other hand. if it is known that a continuous metric
selection for M exists. this provides us with a hunting license to seek a
stable algorithm for determining best approximations.
If X is strictly convex. then M is Chebyshev and PM is
continuous. Thus our main interest lies in non-strictly convex spaces. To
date. except in some special spaces. there seems to be no "useful" condition
on M or PM that characterizes when PM has a continuous selection.
(However. see Theorem 5.6 below.)
Throughout this report. unless otherwise specified. n will denote
a fixed positive integer and M will denote some arbitrary but fixed
n-dimensional subspace of X. The kernel of PM is the set
PM is called lower semi continuous (l.s.c) at Xo provided that for each

open set V with PM(X O ) nV ~ _. there exists a neighborhood U of Xo
such that PM(x) nV ~ _ for all x E U. PM is said to be l.s.c. if it is
l.s.c at each point of X. It is a well-known consequence of the Michael
selection theorem [21] that lower semicontinuity is sufficient for the
existence of a continuous metric selection. However. simple examples (see
e.g. [11]) show that it is not a necessary condition. Actually. KRUGER [15)
showed that lower semicontinuity is equivalent to the existence of a
69
continuous metric selection p which is "kernel preserving", i.e. p(x) 0

for every x E ker PM'
A continuity criterion for PM which is necessary for the
existence of a continuous selection and which, in many cases of interest, is
also sufficient, is called almost lower semicontinuity. Recall that PM is
almost lower semicontinuous (a.l.s.c.) at Xo if for each & > 0, there
exists a neighborhood U of Xo such that n{B&(PM(X» I x E U} ¢ _, where
PM is called a.l.s.c. if it is a.l.s.c. at each point of X. Almost lower

semicontinuity was introduced by the writer and KENDEROV [11] where it was
characterized as follows: PM is a.l.s.c. if and only if, for each E > 0,
there exists a continuous function PE : X ~ M which satisfies
PE(X) E B&(PM(X» for each x E X.
In some situations, the almost lower semicontinuity of PM is
equivalent to the existence of a continuous metric selection for M. These
cases include the following:
(1) When dim M = 1 [11];
(2) When (x E X I PM(x) is a singleton} is dense in X [11];
(3) When X is strictly convex (as a trivial consequence of (2»;
When X = C(T) (FISHER [13] and LI [17], independently).
(4)
In spite of these positive results, it was shown in just the last
few months by BROWN [6] and ZHIVKOV [30], independently, that there are
Banach spaces and finite-dimensional subspaces M such that PM is a.l.s.c.
but PM does not have a continuous selection. However, their examples are
not any of the classical Banach spaces with their "natural" norms.
By the results stated above, almost lower semicontinuity is
equivalent to the existence of a continuous metric selection in the classical
Banach spaces Lp(p), 1 < P < 00, and C(T). Whether this is also true in
L1 (p) is an interesting open problem.
2.1 Question. If M is an n-dimensional subspace (n > 1) of L1 (p) and

PM Is a.l.s.c., must PM have a continuous selection? In particular, what
is the answer in t1?
70
3. The Space Co(T)
In this section we will be working in the space Co(T) : all real continuous
functions x on a locally compact Hausdorff space T which "vanish at
infinity". i.e. {t E T I Ix(t)1 ~ EO} is compact for each E > O. Endowed
with the supremum norm IIxll ; sup {lx(t)1 It E T}. Co(T) is a Banach space.
If T is actually compact. then Co(T) ; C(T). the space of all real
continuous functions on T.
3.1 Theorem. Let M be a finite-dimensional subspace of C(T). T

compact. Then the following statements are equivalent.
(1) PM has a continuous selection;
(2) PM is a.l.s.c.;
(3) PM is 2-l.s.c.;
(4) For each x E ker PM with II xII 1. there exists Yo E PM(x) such
that for all y E PM(x).
int {t I x(t)[yo(t)-y(t)] ~ O} :J Z(PM(x» n x- 1 ({-1.1}). where

Z(y) ; {t E T I y(t) = O} and Z(PM(x»; n{Z(y) lyE PM(x)}.
Recall [11] that PM is said to be 2-1.s.c. at Xo if for each
& > O. there exists a neighborhood U of Xo such that
for all x 1 .x 2 E U.
The implications (1) ~ (2) ~ (3) are from [11]; LAZAR. MORRIS.
and WULBERT [16] verified (1) ~ (4); FISCHER [13] established (3) ~ (1);
and LI [18] verified (4) ~ (1).
STOVER [26] has shown that when T is a compact subset of ~.
then M has a continuous metric selection if and only if each x E C(T) has
a "strict" approximation. (See [26] for the somewhat involved definition of
strict approximation.) In fact. the strict approximation s(x) to each
x E C(T) defines a continuous selection for PM'
An intrinsic characterization of those subspaces M of C(T)
which have continuous metric selections has been made for certain compact
sets T.
71
3.2 Theorem. Let M be a finite-dimensional subspace of C(T). where T

is a compact locally connected Hausdorff space. Then the following
statements are equivalent.
(1) PM has a continuous selection;
(2) For each y € M.
(i) card {bd Z(y)} ~ dim {m € M I Z(m) ~ int Z(y)} :: r(y)
(ii) y has at most r(y) - 1 zeros with sign changes.
Here "card" means "cardinality of" and "bd" means "boundary of".
This result is due to LI [18]. In the important special case when
T : [a.b]. it had been established earlier by NURNBERGER and SOMMER (see
[22]). In the particular case when dim M : 1. this result is also valid
for ~ compact (resp. locally compact) Hausdorff space T. not necessarily
locally connected; see LAZAR. MORRIS. and WULBERT [16] (resp. DEUTSCH.
INDUMATHI. and SCHNATZ [12]).
The condition that everyone-dimensional subspace of Co(T) has a
continuous metric selection turns out to characterize T completely.
3.3 Theorem [7). Let T be a locally compact Hausdorff space. The

following statements are equivalent.
(1) Each one-dimensional subspace of Co(T) has a continuous metric
selection;
(2) T is discrete (1. e. each point in T is isolated) .
If T is the one-point compactification of the natural numbers.
then C(T) :
c. the space of all sequences x such that lim x(n) exists.
[t+oo
BLATTER [3] has verified the following result. Let M be a linear
sublattice of c which is "proximinal" (see definition in section 5). Then
PM has ~ continuous selection and ~ unique one i f and only i f either
M : {O}. M : c. or M: {ax I a €~} for some positive function x € c.
In this section. let (T.S.~) be a measure space and let Ll(~) denote the
space of all real ~-integrable functions x on T with the norm
IIxll: = J Ixld}-L.
T
72
We assume also that the dual space of LI (J-l) is L..,(J-l). (This will be the
case. e.g .• if J-l is a a-finite measure.) For any x E Ll(J-l). its support
is defined by
supp (x): = {t E T I x(t) ~ O}.
(Here. and in the sequel. sets in T are only defined up to a set of

J-l-measure zero.)
An atom in T is a set A E S such that J-l(A) > 0 and i f Be A.
either J-l(B) = 0 or J-l(B) = J-l(A). An element x E Ll (J-l) is said to satisfy
the Lazar condition [12) if whenever A and B are disjoint sets in S with
AU B = supp (x) and J I xl dJ-l = J IxldJ-l. then either A or B must be a

A B
union of finitely many atoms. If o ~ xl E LI (J-l). let [xl] denote the
one-dimensional subspace spanned by xl : [xl) = {OtX I I Ot E IR}.
There is an intrinsic characterization of those one-dimensional
subspaces of L1 (J-l) having continuous metric selections.
4.1 Theorem [12J. Then has a continuous
selection if and only if xl satisfies the Lazar condition.

PARK (23) has extended this result as follows.
4.2 Theorem [23J. Suppose the n-dimensional subspace M has a basis

{x l ,x 2 ' ...• xn } which satisfies J-l{suPP (xi) n sUPP (x j )} o for all
i ~ j. Then PM has a continuous selection if and only if each Xj

satisfies the Lazar condition.
Because of the assumption on the supports of a basis. Park's result
is not a complete characterization of those subspaces having continuous
metric selections. This leads us to the next question.
4.3 Question. Is there an intrinsic characterization of the n-dimensional

(n > 1) subspaces of LI(J-l) which have a continuous metric selection?
The condition that everyone-dimensional subspace of Ll(J-l) have a
continuous metric selection forces the measure space to consist of a finite
number of atoms. More precisely, we have
4.4 Theorem (7). Everyone-dimensional subspace of L1 (J-l) has a continuous

metric selection if and only if Ll(J-l) is finite-dimensional.
73
Recently, PINKUS [24] has considered the following problem. Let T

denote a compact subset of Rm whose interior is dense in T, and let p
denote Lebesgue measure on T. Let Cl(T) denote the space of all real
continuous functions x on T with the Ll-norm:
II xII IIxlil = J Ixld/l·

T
4.5 Theorem (24). Let M be a finite-dimensional subspace of C1 (T). Then

PM has a continuous selection if and only if M is a Chebyshev subspace.
Moreover, if T is also connected and M is a finite-dimensional
non-Chebyshev subspace of C1 (T), then PINKUS [24] showed that PM has no
"L"" -continuous" selection. (An L"" -continuous selection for PM is a
selection s for PM such that IIxn-xll"" ---+ 0 implies
lis (x n ) - s (x) II"" ---+ o. Here 11°11"" denotes the supremum norm.)
Related to Theorem 4.5, LAZAR, MORRIS, and WULBERT [16] showed that
the metric projection onto any finite-dimensional subspace of Ll(T), the
completion of C1 (T), has no continuous selection. However, for more
general sets T and measures p other than Lebesgue measure, it is possible
that L1 (p) contain a finite-dimensional subspaec which has a continuous
(even linear) metric selection (see e.g. [8]).
5. Linear and Lipschitz Continuous Metric Selections
A selection p for the metric projection PM is called linear (resp.

Lipschitz continuous) if p(ax + py) = ap(x) + Pp(y) (resp. there exists
A> 0 such that IIp(x) - p(y)1I SAlix - yll) for all x, y S X and scalars
a,p.
Since for any selection p, IIp(x)1I S 211xll for all x, i t follows
that every linear selection is automatically Lipschitz continuous, and every
Lipschitz continuous selection is (uniformly) continuous.
The writer has characterized the subspaces having linear metric
selections as follows.
5.1 Theorem [8]. PM has a linear selection if and only if ker PM contains
a closed subspace N such that X = M ~ N. In this case, a linear selection
for PM can be defined by
74
p(m+n) ~ m (m+n E M Gl N).
That is. p is the "projection onto M along N".

The next two intrinsic characterizations were proved using Theorem
5.1.
5.2 Theorem. Let M be an n-dimensional subspace of Co(T). Then PM has

a linear selection if and only if M has a basis {x 1 .x 2 •.... x n } such
that
card (supp (xi)) ~ 2 (i=1.2 ..... n).
5.3 Theorem. Let M be an n-dimensional subspace of L1(~). Then PM has

a linear selection if and only if there exist n atoms A1 ..... An such that
for all y E M.
n
T\ U Ai
1
Theorems 5.2 and 5.3 were first proved in the case n = 1 by the
writer [8] and in the general case by an elegant argument of LIN [20].
Since Lp(~)' 1 < P < 00. is strictly convex. every finite-
dimensional subspace M is Chebyshev and PM is continuous. In fact. for
p = 2. L2(~) is a Hilbert space and PM is always linear. In case p ~ 2.
there is an intrinsic characterization of those finite-dimensional subspaces
of Lp(p) having linear metric projections.
§.4 Theorem. Let 1 < P < 00 • p ~ 2. and let M be an n-dimensional

subspace of Lp(~). Then PM is linear if and only if there exists a basis
{x 1 .x 2 ..... x n } of M such that for each i=1.2 •.... n. supp (xi) is purely
atomic and is the union of at most two atoms.
This result is due to PARK [23] who proved it using another
characterization of LIN [20].
The (set-valued) metric projection PM is called Lipschitz
continuous if there is a constant A> 0 such that
for all x.y E X.
where h denotes the Hausdorff metric:

75
h(A,B) : max {sup d(a,B), sup d(b,A)}.

aEA bEB
PARK [23] has shown that if M l§ ~ one-dimensional subspace and

PM is Lipschitz continuous, then PM has ~ Lipschitz continuous selection.
This leads to the next question.
5.5 Question. If M is n-dimensional, n > 1, and PM is Lipschitz

continuous, must PM have a Lipschitz continuous selection?
The converse of this question is false. The example given in [11]
shows that PM can have a Lipschitz continuous (even linear!) selection
without PM being Lipschitz continuous.
Recall that a subset K of the normed linear space X is called
proximinal i f PK(x): {y E K I IIx-yll = d(x,K)} is nonempty for every
x E X. If M is a linear subspace, then PM is always "homogeneous", i.e.
PM(o<x) = o<PM(x) for every scalar 0<, and "additive modulo Mit , i.e.
PM(x+y) = PM(x) + y for all x E X and y E M. However, selections for PM

will in general have neither of these properties. But the writer
(unpublished) has observed that if M is finite dimensional and PM admits ~
continuous selection, then PM admits ~ continuous selection which is also

homogeneous and additive modulo M.
Using this latter fact, PARK and the writer (unpublished) have
characterized those metric projections PM which admit continuous (resp.
Lipschitz continuous) selections as follows.
5.6 Theorem. Let M be a finite-dimensional subspace of X and let

Q : X - - ? X/M denote the quotient map: Q(x) = x + M. The following
statements are equivalent.
(1) PM has a continuous (resp. Lipschitz continuous) selection;
(2) There exists a closed subset N of ker PM such that (QI N)-l is
continuous (resp. Lipschitz continuous) and x - (QI N)-l(x+M) E M for all

x E X.
An analogous characterization of when PM has a linear selection

was given by the writer (8).
76
YOST [29] has shown that i f a subspace M of X has the "11 -ball
2
property", then M is proximinal, PM is Lipschitz continuous, and PM has
a continuous selection which is homogeneous and additive modulo M. (Recall
that M has the 1~ - ball property if, whenever x E X, r ~ 0, IIxll < r + 1,
and the closed ball B(x,r) meets M, then M n B(O,l) n B(x,r) ~ _.)
6. Some Miscellaneous Facts
BROWN, DEUTSCH, INDUMATHI, and KENDEROV [7] have shown that the following
property of a normed space is useful.
6.1 Definition. The normed linear space X has property (CS) provided: if
Xo and z ~ 0 are in X, then there exist E > 0 and cr E {1,-1) such
that for every x with "x-xo" < E either
(i) IIx-Q'zll s IIx-oozll for all c E R

or
(ii) IIx-crzll < IIx-oozll for all c > 1.
Examples of spaces with property (CS) include: (1) every strictly

convex space; (2) every 2-dimensional space; (3) every finite-dimensional
space whose unit ball has only a finite number of extreme points; (4) the
1
"double-cone" space X = 1R3 with the norm IIxll = Ix(1)1 + [x 2 (2) + x 2 (3)]2;
(5) every space with property (P) of BROWN [5].
6.2 Theorem [7]. X has property (CS) if and only if everyone-dimensional

subspace of X has a continuous metric selection.
This result was used to characterize the spaces of type ColT) and
L1 (p) which have property (CS): T must be discrete and dim L1 (p) < 00
(see Theorems 3.3 and 4.4 above). In these spaces, property (CS) coincides
with property (P) (see [4]), but not in general (e.g. the "double-cone"
space) .
6.3 Question. Is property (CS) equivalent to every finite-dimensional

subspace having a continuous metric selection? (The answer is yes in
ColT) and L1 (p) or, more generally, in any space in which property (CS)
is equivalent to property (P).)
If X and V are normed linear spaces. let L(X.V) denote the
77
set ot all bounded linear operators in L(X,Y) having rank n.
6.4 Theorem (KAMAL [14]). Let X be finite-dimensional and n a positive

integer. Then each n-dim~usional subspace of X* has a continuous metric
selection if and only if Kn(X,Co(T» is proximinal in L(X,Co(T» for each
locally compact Hausdorff space T.
It follows from Theorem 6.2 that, for n = I, the proximinality
condition of this theorem is equivalent to property (CS) for X*.
Let M be a proximinal linear subspace of the normed space
(X,U"U), let T be a compact Hausdorff space, and let C(T,X) denote the
space of all continuous functions f: T --+ X with the norm
UfU' = sup {Uf(t)U I t E T}. The set C(T,M) of all those f E C(T,X)
with range in M is a subspace of C(T,X). LIGHT and CHENEY [19] have shown
that if PM admits a continuous selection, then so also does the metric
projection PC(T,M)' A generalization of this is given in PROLLA and
CHIACCHIO [25].
In closing, although we have mainly considered the existence of
selections for the metric projection onto a linear subspace, it is perhaps
worth mentioning a few results in the non-subspace case. Here the theory is
still sparse.
BERENS [2] has considered the space X = ~nxn, the set of all
n x n-matrices equipped with the spectral norm. Among other things, he
showed that if K is either the (nonlinear) subset consisting of the rank k
(k < n) matrices or the subset of singular matrices, then it is possible to
construct a continuous selection for PK.
BARTKE [1] has shown that in the case of best approximation by
rational functions over a discrete set, the metric projection fails to have a
continuous selection.
Finally, in some problems of best approximation, the approximating
set depends on the element being approximated. That is, for each x E X a
proximinal set K(x) in X is given and one seeks best approximations to x
trom K(x), i.e. elements ot the set PK(x)(x). Then one asks whether the
set-valued mapping x r-+ PK(x)(x) has a continuous selection. (This
contains the metric projection PK as a special case by taking K(x) = K
for all x.) In the Banach space X = B[a,b] of all real bounded functions
on [a,b] with the supremum norm, UBHAYA ([27], [28]) has considered certain
collections of convex sets (K(x) I x E X} (e.g. K(x) is the set of all
convex functions y E X with x S y on [a,b]). For these examples, he
constructed Lipschitz continuous selections for the mapping x r-+ PK(x)(x).
78
7. References
1. K. Bartke, Diskete rationale Approximation, these proceedings.

2. H. Berens, On the best approximation in matrix spaces, these
proceedings.
3. J. Blatter, Metric projections of c onto closed vector sublattices,
preprint (Jan., 1986).
4. J. Blatter, P. Morris, and D. Wulbert, Continuity of the set-valued
metric projection, Math. Ann., 178 (1986), 12-24.
5. A. L. Brown, Best n-dimensional approximation to sets of functions,
Proc. London Math. Soc., 14 (1964), 577-594.
6. A. L. Brown, Set valued mappings, continuous selections and metric
projections, submitted for publication (Sept. 1986).
7. A. L. Brown, F. Deutsch, V. Indumathi, and P. Kenderov, work in
progress.
8. F. Deutsch, Linear selections for the metric projection, J. Functional
Anal., 99 (1982), 269-292.
9. F. Deutsch, A survey of metric selections, Fixed Points and Nonexpansive
Mappings (R. C. Sine, ed.), Contemporary Mathematics, 18, Providence,
1983, pp. 49-71.
10. F. Deutsch, Continuous selections for metric projections: some recent
progress, Approximation Theory V (C. K. Chui, L. L. Schumaker, and J. D.
Ward, eds.), Academic Press, New York, 1986, pp. 319-322.
11. F. Deutsch and P. Kenderov, Continuous selections and approximate
selections for set-valued mappings and applications to metric
projections, SIAM J. Math. Anal. Appl., 14 (1983), 185-194.
12. F. Deutsch, V. Indumathi, and K. Schnatz, Lower semicontinuity, and
continuous selections for set-valued mappings, J. Approx. Theory (to
appeal') .
13. T. Fischer, A continuity condition for the existence of a continuous
selection for a set-valued mapping, J. Approx. Theory (to appear).
14. A. Kamal, On proximinality and sets of operators. I. Best approximation
by finite rank operators, J. Approx. Theory, 47 (1986), 132-145.
15. H. Kruger, A remark on lower semi-continuity of the set-valued metric

projection, J. Approx. Theory, 28 (1980), 83-86.
16. A. J. Lazar, P. D. Morris, and D. E. Wulbert, Continuous selections for
metric projections, J. Funct. Anal., 3 (1969), 193-216.
17. W. L1, The problems of continuous selections in C(X), preprint (Feb.,
1986) .
18. W. L1. The characterization of continuous selections for metric
projections in C(K), preprint (May, 1986).
19. W. fl. Lj !~ht and E. W. Cheney, Approximation Theory in Tensor Product
Spaces, Lecture Notes in Mathematics, #1169, Springer-Verlag. New York.
1985.
79
20. P.-K. Lin, Remarks on linear selections for the metric projection, J.
Approximation Theory, 43 (1985), 64-74.
21. E. Michael, Continuous selections I, Annals of Math., 63 (1956),
361-382.
22. G. Nurnberger and M. Sommer, Continuous selections in Chebyshev
approximation, Parametric Optimization and Approximation (B. Brosowski
and F. Deutsch, eds.), ISNM, 72 (1984), Birkhauser Verlag, Basel
(pp. 248-263).
23. S.-H. Park, work in progress.
24. A. Pinkus, Continuous selections for the metric projection on Cl ,
preprint (Aug. 1986).
25. J. B. Prolla and A. O. Chiacchio, Proximinality of certain subspaces of
Cb(S;E), Rocky Mountain Math. J. (to appear).
26. V. Stover, The strict approximation and continuous selections for the
metric projection, doctoral dissertation, Univ. California, San Diego,
1981.
27. V. Ubhaya, Lipschitz condition in minimum norm problems on bounded
functions, J. Approx. Theory, 45 (1985), 201-218.
28. V. Ubhaya, Optimal Lipschitzian selection operator in quasi-convex
optimization, preprint (April, 1986).
29. D. Yost, Best approximation and intersections of balls in Banach spaces,
Bull. Austral. Math. Soc., 20 (1979), 285-300.
30. N. V. Zhivkov, A characterization of reflexive spaces by means of
continuous approximate selections for metric projections, J. Aprox.
Theory (to appear).
Prof. Frank Deutsch, Department of Mathematics, The Pennsylvania State

University, University Park, Pennsylvania 16802, U.S.A.
Numerical Mathematics. Vol. 81
SOME PROPERTIES OF THE DILIBERTO-STRAUS ALGORITHMS IN C(SxT)

by
M von Golitschck and W.A. Light
Introduction
The algorithm considered in this paper is due to Diliberto and Straus [2].
It produces, for a given element f E C(SxT). an element w E C(S) + C(T) which is
a best approximation to f from C(S) + C(T) using the usual supremum norm. We
begin with some notation. We will assume Sand T are compact Hausdorff spaces.
The linear subspace C(S) + C(T) will be denoted on occasion by W. Note that an
obvious interpretation of C(S). C(T) as subspaces of C(SxT) must be made. and that
with this interpretation W is closed. Operators M t : C(SxT) ... C(S) and Ms :
C(SxT) ... C(T) are defined as follows
(~f)(s) = ! sup f(s.t) + ! inf f(s.t) (f E C(SxT»

2 tET 2 tET
and
(~f)(t) = ! sup f(s.t) + ! inf f(s.t) (f E C(SxT».

2 SES 2 SES
The algorithm now starts by fixing f E C(SxT). We then define f 0 = f and
n = 1.2.3•......
The main historical thread is as follows. Firstly. this algorithm is a version of

the alternating algorithm which seems first to have been proposed in a Hilbert
space setting by von Neumann [8]. Secondly. Diliberto and Straus were only able
to establish that IIf nil ... dist(f.W). It remained to Aumann [1]. to show that the
81
algorithm actually converged. At the same time Golomb [4] generalised the results
of [2], and the central theme of these three papers can be found in [5].
Our purpose in this paper is twofold. Firstly, we wish to examine the
rate of convergence of the algorithm. Our limited conclusions will nevertheless
serve to show that the algorithm can converge slowly even on a finite grid.
Secondly, we shall show that in some sense the algorithm is "robust". This is done
by showing that the process of best approximation of functions f E C(SxT) by
elements of W is discontinuous in quite a strong sense, but that nevertheless the
algorithm can still cope with these discontinuities and produce a best approximation.
II. Convergence Results
We begin by pointing our that the work of Aumann did not in any way
involve estimates of the rate of convergence of the algorithm. Nor does it seem
possible to sharpen the critical tools used by Aumann so as to obtain such estimates.
This means that the best estimates come from [2] and [3].
2.1 LEMMA [3]. For every natural number k there is a function f E C(SxT) such
that the iterates in the algorithm satisfy
Thus the algorithm may produce very slow convergence initially. However, there
are results which to a certain extent ameliorate this initially gloomy picture. In
particular it follows from 2.3 below that IIf 211 = ... = IIf kll implies
IIf 211 .; k/(k-l) dist(f.W).
2.2 LEMMA [3]. For any f E C(SxT) the third and fourth iterates of the algorithm
satisfy
IIf 311 , 2 dist (f,W) and IIf 411'; dist(f,W).
For general n, such results become progressively harder to obtain.
2.3 LEMMA [3] and [7]. For any f E C(SxT), the iterates in the algorithm satisfy
82
n ~ 3
(ii) for any natural numbers N ~ 2 and k ~ I,
k+2
k+1
--L dist (f,W), n E IlL

logn
Proof As indicated in the statement (i) is from [3] while (ii) is from [7]. The proof
of (iii) is as follows. From (ii) we can obtain, by setting N = 4+k(m-I),
By applying this inequality recursively we obtain
For each n E IN, the appropriate choice of k and m, km , n leads to (iii). •

Our convergence result will refer to the case where Sand T have only
finitely many points, and we will develop the arguments first, interspersed with
results needed from other sources, and then state the result at the end. For
We also
assume that the iterates in the algorithm corresponding to the function f converge
to the function zf. The function f will be of type A if for the limit point zf we
have
(i) For each I , i , k there is a unique pair of points tao and ta.
1 1
such that z(s i,t a .) = max z(si,t) and z(si,tB.) = min z(si,t);
1 t I t
(ii) For each I , j , k there is a unique pair of points sa ., sa. such
J J
that Z{Sa/j) = m~x z(s,tj) and Z(SB j ,tj) = m~n z(s,t/
It is then immediate that for sufficiently large values of n, say n ~ N, the iterates
f n in the algorithm will also possess properties (i) and (ii). Then for a given f of
type A there exists an N such that for all n ~ N we have
(i)
83
2.4 DEFINITION A path is an ordered set of points {p I,P2' ... } in SxT such that the
line segments Pi - Pi+1 are alternately "horizontal" (parallel to the S-direction) and
"vertical" (parallel to the T -direction). If Sand T are infinite the path may consist
of a finite or infinite number of points. If the path consists of 2n points and PI- P2
is vertical then the path is called closed.
If {PI,P2, ... ,P2n} is a closed path R say, then a linear functional nR is
defined on C(SxT) by
2n
n (f) =..L .r (_l)if(p1·)
J 2n 1=1
The fundamental results of Diliberto and Straus can now be summarised.
2.5 THEOREM [2]. For the functional n. associated with a closed path R we have
(i) n.(w) = 0 for all w E W
(ii) If n .(f) = 0 for all closed paths J then fEW
(iii) IIn.1I = I and dist(f,W) = s~P n.(f) = I~mllfnll .
In addition if Sand T are finite sets then w E W is a best approximation to f E
C(SxT) if and only if there is a closed path • such that n .(f) = dist(f,W).
In a finite domain {sl' ... , sk} x {tl' ... , tk} the path
(sl,t l ), (s2,tl)' (s2,t2)' (s3,t2)' ... , (sk,tk)' (sl,tk)
will be called the basic closed path. It is illustrated in Figure l.
Fi gure
L)~l.)
~,>Jz )
.- -- --
·---·~·LI
84
If a function f is of type A on our finite domain, then by reordering Sand T if

necessary it is possible to decompose the domain so that the maximum and minimum
values of zf lie on "blocks" of basic closed paths. What this means is that we can
m m
write S = u S· T = U T· so that S· = (Sj : u(i) , j 'u(i+l)}, Ti = (tj : u(i)
i=l I' i=l 1 1
, j , u(i+l)} and 1 = u(l) < u(2) < u(3) < ... < u(m) = k. Furthermore on (Si x T)
U (S x T i) the maximum and minimum points of zf in each row and column lie on
the basic closed path in Si x T i. Figure 2 illustrates a possible configuration.
52. S3
~ ~
I
I I Figure 2
I o Maximum points of zf
I
I
I I Minimum points of zf
I I
,,r------- -~
L ____ ---.U
- \- - - -
'?>l
- - - - - - OO------li~":l1
With the same reordering we will ensure the same configuration for the iterates f n
in the algorithm once n j» N. For these values of n the algorithm becomes a
linear process. If we write
then
where A is a block matrix of the form

85
o
A
Since the block calculations are completely independent we consider the case where
A has only one block consisting of the basic closed path in all of SxT. An
elementary calculation shows that
fn+2 = ftl. - !. {fll . + fl1· I} - !.{f.1}. - f,1I. I + r. nl . - fl}. I . I} .
ij Ij 2 11 1,1+ 4 1j J,j+ 1- ,j j- ,j-
Here the subscripts need to be read modulo n. From this formula, one can easily
see that the value of fn+2 on the basic closed path is completely determined by the
value of rn on that same path. Also the rate of convergence of the algorithm in
SxT cannot be better than its rate of convergence on the basic closed path so we
now confine our attention to the basic closed path. If r~ represents the vector of
values of fn on the basic closed path so that
Then r:+2 = kB r~ where
-I o o o -I
B -I -I
-1 -1
o o -I -I
o o -1 -1
-1 -1
2.6 LEMMA The 2k x 2k matrix B has a complete set 0/ eigenvectors and eigenvalues
2(1 + cos(2rJl/k» r = 0,2,3, ..., k-l. In addition, i/ k is odd B has 0 as an
eigenvalue.
86
PROOF. If x = (xI,x2' ... , x2k)T and (B -ILl) x = 0 then this system has the form
1,2, ... ,2k,
where subscripts are again to be read modulo 2k. This is equivalent to
1,2, ... ,2k,
which in turn is equivalent to
1,2, ..., 2k.
At the point we see that IL=O is an eigenvalue with eigenvector given by x2i_1 = 0
and x2i arbitrary, I , i 'k. Putting Yi = x2i_1 we must now solve Yi + (2-IL)
Yi+ I + Yi+2 = 0 in such a way that Yk+ I = YI. This can only happen if the
difference equation has a periodic solution, and two linearly independent solutions
are given by Yi = cos i9 r , Yi = sin i9 r where (k+I)9 r = 9r + 2r1l and 9 '# O. If
9 = 0 only one independent solution is obtained by taking Yi = 1. The
corresponding eigenvalues are given by
2-u_ .. -2 cos 2r
~,
1l r 0,1.2 •... , k-I,
which gives
ILr = 2( I + cos(2r1l/k» , r = 0,1.2, ...• k-l.
Now for I , r , k-I the pairs of linearly independent eigenvectors gives us 2k-2
vectors which are of rank 2k-2. In addition the eigenvector (I,-I,I,-I,... ,I)T
corresponding to eigenvalue I and (0,1,0, ..., 0) T corresponding to value 0 make up
our complete set of eigenvectors. •
87
.
Let us now WrIte c·n+2m = emf·n were
h .
the elgenva Iues 0f e are 1 (
"2" I+cos
4.!! ), r = O,I, ... ,k-1 and 0. Since e has a complete set of eigenvectors, we may
write c~ as a linear combination of these vectors, say
2k 2k 2k
•
Cn +2m em I: oc· U· I: oc·emu· = i=1
I: oc·}. IJl u l· .
i= I I I i=1 I I I I
If we assume }.I = I, then since the algorithm converges and all other }.i have
I}.i l < I we can write
2k
oclul + ~oci}.Tui
Thus oci ul is the limit of the algorithm, which we have denoted previously by zf. •
Now, if we suppose }.2 = }.3 > }.i' 4 ' i , 2k, we can write
•• 2k
liz f - Cn+2mll = "2I: oci}.T ui" .
For sufficiently large values of m we can assume
and so
Z
•f - C n+2mll ~ e }.T
where e, e 1 are appropriate constants. This result may be summarised by saying
that the convergence depends on the length of the basic path (as indicated by the
parameter k). Our analysis enables us to state the following result.
2.7 THEOREM Let Sand T be finite sets and suppose f E CCSxT) is of type A.
Let zf denote the limit of the iterates of the Diliberto-Straus algorithm as applied to f
and let the longest basic closed path for zf have 2k points. Then there exists a
constant e and a number M such that
88
Ilf m - zfll ~ Cr m (l + cos(21l/k»m for all m ~ M.
This result points to a major weakness of the algorithm. For very long
basic closed paths the convergence rate is certainly no faster than a constant times
the approximate factor (l_(1l/k)2), which of course is very close to I for large k.
Worse still, this slow convergence can be guaranteed in a situation where the
eventual location of the maximum and minimum values in each row and column of
SxT is already known. Under these circumstances the problem can be regarded as
a linear problem as indicated in section 4 of [7], and such a viewpoint would lead
to a very rapid solution. Thus one may conclude that long extremal paths are
handled badly by the algorithm. For continua Sand T there are functions whose
extremal paths contain infinitely many points and it seems unlikely that one could
expect geometric convergence for such functions.
III Continuity Results
The results of the previous section highlight the poor performance of the
algorithm. We construct examples in this section which indicate that despite the
slow convergence, the algorithm is in some sense robust. Again we will begin with
Sand T being finite sets. The detail of the constructions to be carried out in
this section is rather complex and we choose to indicate the direction of
development of the arguments in preference to providing a careful analysis. For
convenience, we suppose our points lSi) are equally space in [0,1], so that si ih
o '" i '" 2k with the same assumption on the {tj}. We now define a function f on
SxT by f(si,tj) =1 - li-jlh. The function is sketched in Figure 3 below.
1
Figure 3
l-rh.
1..
1...
o
89
A second function g is defined by
I - li-j-Ilh ~ j-k+1
{
I - 1i-j+2klh , j-k
f(si,ti+k)=I-kh
tk I-kh
Figure 4
We now for convenience in later analysis define a function z on a grid contained

in [0,3]2 in accordance with the following programme.
(i) In [0,1]2 we take z = f
(ii) In [0,1] x [2,3] we rotate -f Tl/2 clockwise and translate to this
interval.
(iii) In [2,3] x [2,3] we rotate f Tl clockwise and translate
(iv) In [2,3] x [0,1] we rotate -f 3Tl/2 clockwise and translate
We take no further grid points in [0,3]2 so that z is completely defined. A sketch
is provided in Figure S.
90
1;{S,tJ-- 1
2
T Figure 5
2
S
3.1 LEMMA. We have dist(z,W) 1 and if IIz-wll dist(z,W) then w == o.
PROOF. The points labelled PI ,P2,P3' ... , PSk+4 describe a closed path R in the
domain. Hence by 2.5, dist(z,W) ~ n.(z) = 1. The function w = 0 gives
IIz-wll = 1 and so dist(z,W) = 1. By a result of von Golitschek and Cheney], if
wI and w2 are best approximations to z and. is a closed path such that nR(z)
dist(z,W) then w1=w2 on J. Hence w=o is the unique best approximation to z. •
We now modify the definition of the function g to a function g &
satisfying
g(si,9 j "I- i-I
g&(si,tj)
[ + & j = i-I,
- & j = i-I,
,
k +
, k
, i , 2k
We take & = ~h and define z& in the same way as z except using g& on [0,1]2.
We now claim that dist(z&,W) = 1. Following the argument in 3.1 we see easily
that dist(z&,W) ~ 1. We now construct a function w& E W such that IIz& - w&1I = 1.
It will be convenient to label the grid points in [0,1] on the S-axis as sO,sl, ... ,s2k as
I" • f I I I
before while those in [2,3] are labelled sO,sl' ...,s2k· The pomts to,t 1,t2' ... , t2k
are defined similarly. Now set w& = u& + v& where
91
i &
{ (2k-i) & k+1 , i , 2k
3.2 LEMMA We have dist(z&,W) I and the function w & e W is the unique best
approximation to z& from W.
PROOF. This is a long and tedious calculation. We will provide the details on
the square [0,1]2 and leave the rest for the tenacious reader to fill in. As it
happens this is the most difficult region. In Figure 6 we see a schematic
representation of z& and w& in this region.
(0,1) r----------,(I,I)
1-(2k+i-j)h
j=i-I (i+j-2k)& (j-i) &
I +(i- j-I)h j=k
j=i+k
l+(I-i+j)h Figure 6 (i-j) & (2m-i-j)&
(0,0) (1,0) i=k
We begin by supposing 'k, > k. Then
I - (2k+i-j)h-(i+ j-2k)h/2 i- j ~ -k
{
1 + (i-j-l)h - (i+j-2k)h/2 i-j < -k
- (2k+3i-j)h/2 i- j ~ -k
- (2-i+3j-2k)h/2 i- j < -k
92
Now, for such values of i,j

o .. 2k + 3i - j .. 2k + 3k - k 4k
and
I .. 2 - i + 3j - 2k .. 2 + 3k - 2k = 2-k.
Hence for such values of i,j 0 .. Z&(si,9 - w &(si,tj) '" 1. Now consider the line
i-j=1. If i '" k then z&(si,ti_I) = 1+& while w&(si,ti_I) = &, while if i > k then
z&(si,ti_I) = 1-& while w &(si,ti_I) = -&. Hence z&(si,ti_1 )-w &(si,ti_I) 1. In the
remaining cases it will be convenient to write i-j=r, and recall that r=1 has already
been considered.
Suppose now i,j '" k. Then -k '" r '" k and
Z&(si,tj) - w &(si,tj)
{ I + (r-I)h - rh/2
+ (l-r) -rh/2
-k '" r < I
I < r , k
{
+ ( !. - I)h -k'r<1
2
I + (l _ 3r)h I<r'k.
2
Now for -k , r '" I we have -I .. - (~+ I)h '" <i - l)h '" 0 while for I < r , k
-I'" (I - ~k)h '" (I - ~)h '" o.

Hence again 0 '" z&(si,t j) '" 1. Now consider i,j > k so that -k '" r '" k again.
Then
I + (r-I)h + rh/2 -k '" r < I

{
I + (l-r)h + rh/2
I + (~r - I)h -k '" r < I

{
= I + (l-r/2)h
Now for -k" r < I, -2 '" - ~k + I)h .. (~r - I)h '" 0
while for I < r '" k -I '" (I - ~)h '" (I - r/2)h '" 0 .
Thus again -I '" z&(si,tj) - W&(si,9 '" 1. Finally, we examine > k, '" k. In this
case
93
Z&(si,tj) - W&(si,tj) = I + (l-i+j)h - (2-i-j)h/2

I - (k + i/2 -3j/2 - I)h.
If we exclude the line i-j=1 then in this region i-j ~ 2. Then
°, k + i/2 - 3j/2 -1 , 2k - 1
and so
Thus on [0,1]2 we have Iz& - w &1 , I . A similar analysis on the other three
regions leads to IIz& - w &11 = I. The uniqueness then follows immediately from
similar considerations to 3.2. •
If we combine 3.1 and 3.2 then we have IIz-z&1I ,&. However, at the
point (so,tk) we have w& = k& so that IIw-w&1I = k& = ~h = 1 Thus we have
constructed two functions which can be very close on a finite grid while their best
approximations are far apart. This is the "ill-conditioning" of the best
approximation problem to which we wished to draw attention in this section.
This example can be extended to qO,3]2 in the following way. In
addition to the rectangular grid insert the leading diagonals and extend the
definitions of z and z& by defining z,z& to be the interpolants to the data on the
rectangular grid which are linear over the triangles. Define z,z& on the rest of
the grid in such a way that the linear interpolants to w,w& remain the best
approximations to z,z& on the appropriate subsets of [0,3]2. For example, linear
interpolation on [1,2] x [0,1] of the form
z(s,t) = ~(s-l)z(2,t) + (2-s)z(l,t)}
will suffice. Then the functions z,z& are 2-Lipschitz functions and w,w& remain
uniquely defined on the grids from which the construction began. Thus IIz-z,1I , &
while IIw-w &11 ~ 1
3.3 THEOREM Let z be in C(SxT) and let Az denote the set of all best
approximations to f from W '" C(S) + C(T). Then A : C(SxT) ... 2W is discontinuous in
the sense that there is no continuous selection from A, i.e. no continuous map
~ : C(SxT) ... W such that ~ E Az for each z E C(SxT).
94
PROOF. From the remarks preceding the theorem we can, for each n E IN,
construct functions f n,gn E C(SxT) such that
(i) IIf n -gn" , 1In, Ilf nll,lIgnll , 2
(ii) f nand gn are 2-Lipschitz
(iii) inf{lla-bll: a E Af n' b E Ag n } ~ i.
Now the set F = (f n):' U (gn):' is an equicontinuous bounded set in C(SxT) and so
is conditionally compact there. If there were to be a continuous selection 4> then
it would have to be uniformly continuous on F, but (i) and (ii) make this
impossible. •
In view of the poor conditioning of the problem it is hardly surprising
that the Diliberto-Straus algorithm converges slowly. Somehow the algorithm must
avoid the discontinuities in the best approximation operator as it converges and this
is probably responsible for the poor convergence.
References
1. Aumann, G. "Uber Approximative Nomographie, I, II and III", Bayer. Akad.
Wiss. Math-Nat. Kl. S.B. (1958), 137-155 ibid 1959, 103-109; ibid. (1960), 27-34.
2. Diliberto, S.P. and Straus, E.G. "On the approximation of a function of several
variables by the sum of functions of fewer variables," Pacific J. Math. 1
(1951), 195-210.
3. Von Golitschek, M and Cheney, E.W. "On the algorithm of Dilberto and Straus
for approximating bivariate functions by univariate ones", Numer. Funct.
Analysis and Optimisation 1 (1979), 341-363.
4. Golomb, M "Approximation by functions of fewer variables" in "On Numerical

Approximation," R. Langer, ed. University of Wisconsin Press, Madison,
Wisconsin, 1959, 275-327.
5. Light, W.A. and Cheney, E.W. "Approximation Theory in Tensor Product Spaces",
Springer Lecture notes in Mathematics vol. 1169 (1985), Heidelberg.
7. Light, W.A. and Cheney, E.W. "On the approximation of a bivariate function
by sums of univariate functions", J. Approx. Th. 29 (1980), 305-322.
8. Von Neumann, J. "Functional Operators, Vol. II" Annals of Mathematics Studies

22, Princeton University Press, 1950.
Acknowledgement Both authors were supported in this research by NATO grant

no. RG/0095.
Professor M von Golitschek, Institut fUr Angewandte Mathematik, Universitiit

WUrzburg, Am Hubland, Wurzburg.
Dr. W.A. Light, Department of Mathematics, University of Lancaster, Lancaster, LAI
4YL, U.K.
MATHEMATISCHE AUSWERTUNG ERGOSPIROMETRISCHER MESSUNGEN
M. v. Golitschek
Institut ftir Angewandte Mathematik und Statistik
der Universitat Wtirzburg
F. Schardt
Medizinische Poliklinik
M. Wiegand
Medizinische Poliklinik
1. Einleitung
Unter Ergospirometrie versteht man die Messung des Atemminuten-

volumens, der Sauerstoffaufnahme und der Kohlendioxydabgabe
unter korperlicher Belastung. Wahrend noch vor wenigen Jahren
die Berechnung der Primardaten nur aIle 30-60 Sekunden moglich
war, ergab sich mit der Entwicklung der Einzelatemzuganalyse
eine Ftille von Daten bei allen direkt und indirekt berechneten
parametern mit einem zum Teil sehr unruhigen Kurvenverlauf.
Die bisher angewandten mathematischen Methoden, wie z.B. gleiten-
der Mittelwert und Logarithmus konnten Knickpunkte, Krtimmungs-
punkte, Maxima und Minima nur unzulanglich bestimmen. Erst mit
der kubischen Splinefunktion eroffnete sich eine mathematische
Methode, die genannten Charakteristika der einzelnen Parameter
genau zu bestimmen. Gleichzeitig ergab sich damit die Moglich-
keit , Tangenten und Flachenintegrale der parameterkurven zu be-
96
rechnen. Es solI somit dargestellt werden, auf welche Art und

Weise die kubische Splinefunktion zur Bestimmung und Auswertung
von Charakteristika der einzelnen parameter herangezogen werden
kann.
2. Das mathematische Problem
Wir wert en die Belastungsphase und die Erholungsphase getrennt

aus. In jeder dieser beiden Phasen liegt folgendes mathemati-
sche Problem vor:
Zu den Zeiten t1 < t2 < ... < tm wurden die Werte

Y1'Y2""'Ym gemessen. Wesentlich ftir unser mathematisches
Verfahren ist, daB
(i) die Anzahl m der Messungen sehr groB ist,
(ii) die Zeitabstande t

- t. zwischen zwei
j+1 J
aufeinanderfolgenden Messungen nicht zu un-
regelmaBig sind,
(iii) tiber die Eigenschaften der zu erwartenden

Losungskurve nichts bekannt ist (wie etwa sttick-
weise Linearitat, exponentielles Wachstum, etc.).
Die Eigenschaften (i), (ii), (iii) erlauben, zur Auswertung

der MeBdaten kubische Splinefunktionen mit aquidistanten
Knoten zu verwenden:
Wir wahlen nE IN (hier: 10 ~ n ~ 40) und definieren den Knoten-

abstand h: = (t m - t 1) In und die Knoten
k = 0,1, ... ,n+2
insbesondere ist = tm
97
2
Mi t Sp, Sp E c [t l' tmJ , wollen wir die Menge aller kubischen
Splinefunktionen mit den aquidistanten Knoten {x 1 , .•. ,x n + 1 }
bezeichnen. Eine besonders einfache Basis bilden die sogenannten
B-Splines:
Definieren wir die Funktion
4+3 t 2 (I t I -2), -l~t~l
S(t) (2 - It 1)3 1 < It I ~ 2

0 It I > 2
so ist die B-Splinebasis {U , U 1 , ••• , U 2} von Sp gegeben

o n+
durch
und jede kubische Splinefunktion uE Sp hat eine eindeutige

Darstellung der Form
n+2
(2. 1 ) u ( t) = B c k Uk ( t )
k=o
Wir wahlen einen geeigneten Glattungsparameter ~ > 0 und wollen

als Losungskurve unserer Mel3reihe die Splinefunktion u* E Sp
finden, die den Ausdruck
m
2
(2.2) B (y r - u ( tr) ) + dt
r=l
bezuglich aller u E Sp minimiert.
Im Falle ~ =0 ware dies die Gaul3sche Methode der kleinsten

Quadrate, angewendet fur den linearen Unterraum Sp
tm
Den Glattungsterm ~
(u"(t))2 dt , J ~ > 0 haben wir in An-
tl
lehnung an eine Arbeit von Reinsch [2] und Schoenberg [3] hin-
98
zugenommen, da hierdurch die Anzahl n der Knoten wesentlich

groBer gewahlt werden kann als im Falle ~ = O. Wegen der groBen
Anzahl der Daten war die Anwendung des Verfanrens von Reinsch
nicht moglich. Denn in diesem Verfahren wird der Ausdruck
m
2
L) (Y
r - f ( tr» +
r=1
bezliglich aller Funktionen minimiert. Die Losungs-

kurve ist hier eine kubische Splinefunktion mit den m Knoten
t 1 , ••• ,t m • Da aber m bei unserem Problem sehr groB ist, werden
die Rechenzeit, insbesondere aber der Speicherplatzbedarf beim
Verfahren von Reinsch sehr groB.
J. Die Berechnung der Losungskurve
Wegen der Darstellung (2.1) der Splinefunktionen ist das Mini-

mierungsproblem (2.2) aquivalent zur Aufgabe, die Funktion
m n+2
F(c ,oo.,c 2) L) (y r - L)
o n+
r=1 k=o
bezliglich c , ••. , c 2 E JR zu minimieren.

o n+
Indem wir grad F = 0 setzen, erhal ten wir die n+J Normal-
gleichungen
(J.1 ) (B + ~ C)c = b
wobei c=(c , ... ,c2)T die unbekannten parameter sind, und

on+
die (n+J) X (n+J)-Matrizen B = (B ik ) C = (C ik ) und
T
b=(b , ... ,b 2) wie folgt zuberechnen sind:
o n+
99
m
L) Ui (t r ) Uk(t r )
r=1
t
m
Cik = J U'.'(t) Uk"(t) dt
t 1 ~
m
b.
~
= L) U.(t
~ r
) yr
r=1
i, k=O,1, •.• ,n+2. Bei der Berechnung der Bik und b.~ kann
man viel Rechenzeit sparen, falls man beachtet, daB die Basis-
funktionen die Eigenschaft Uk(t) = 0 fur
t ~ xk _2 ' haben. Die Matrix C hat die Gestalt
2 -3
°
-3 8 -6
°
° -6 14 -9
°
C = 6 h- 3
° -9 16 -9
16
°
-9
"- ... " ...
"-
...
"
-9 16 -9
°
° -9 14 -6
°
° -6 8 -3
° -3 2
Im Faile 13 > 0 ist die Matrix B + I3C stets posi tiv defini t.
Zur Auflosung von (3.1) haben wir das GauBsche Eliminations-
verfahren ohne Pivotsuche verwendet. Hierbei wird die Struktur
der Matrix B + I3C , Bik + I3Cik = 0 falls Ii - k I ~ 4 , nicht
zerstort. Naturlich sind auch andere Verfahren, wie z.B.
Cholesky-Zerlegung, erfolgreich.
100
Durch die Darstellung (2.1) der Losungskurve,
n+2
(3. 2 ) u*(t) = B c~ s( (t - xk)/h)
k=o
kann zu jeder gewlinschten Zeit t der Wert u*(t) schnell

berechnet werden. Auch hier wird die Rechenzeit erheblich klir-
zer, wenn die Eigenschaft s(x) =0 , falls Ix I ~ 2 , genu t z t
wird.
Beispiel: Atemminutenvolumen
Die MeBdaten sind durch punkte, die errechneten Krlimm- und

Knickpunkte sind durch senkrechte Linien markiert •
~~
. #.
I ',I
o 2.000
101
4. Auswertung der Losungskurve
Aus der Losungskurve u* sollen nun insbesondere
(i) absolute und relative Maxima und Minima.
(ii) Stellen groBter Krlimmung (Krlimmpunkte).
(iii) Knickpunkte
bestimmt werden.
Wahrend die Ermittlung der Extrema unproblematisch ist. treten

bei der Berechnung von Krlimm- und Knickpunkten erhebliche Schwie-
rigkeiten auf. Wir sind wie folgt vorgegangen:
Krlimmpunkte: Da u* eine kubische Splinefunktion mit den

Knoten x 1 ••••• x n + 1 ist. ist die 2. Ableitung u*" eine stetige
stlickweise lineare Funktion. deren Extrema (i.e. Krlimmpunkte)
nur in der Knotenmenge {x 1 ; ••• ;xn + 1 } liegen konnen. Hier
macht es sich besonders bezahlt. daB wir - allgemeiner als bei
der GauBschen Methode der kleinsten Quadrate - den Ausdruck
(2.2) minimiert haben. Denn auch bei groBem n (n= 30) liefert
unsere Methode bei genligend groBem Glattungsparameter ~ ausge-
zeichnete Losungskurven u* • die einerseits ausreichend glatt
und andererseits ausreichend flexibel sind: Die Dimension von
Sp ist n + 3 • der Abstand zwischen zwei aufeinander folgenden
Knoten ist h= (t m - tl )/n •
Knickpunkte: Wir definieren: Ein Knoten xk heiBt Knickpunkt.

wenn die Losungskurve u* entweder
(a) in xk _ 1 ~ t ~ xk nahezu linear und in
stark gekrlimmt ist oder

102
(b) in xk _ 1 ~ t ~ xk stark gekriimmt und in
Xk ~ t ~ xk+ 1 nahezu linear ist.
Mit anderen Worten, Knickpunkte sind Zeitpunkte, an denen u*

lineares Verhalten beendet oder beginnt. Bei der Berechnung der
Knickpunkte haben wir eine Konstante A gewi:i.hl t, A"" 0.2 ,
und festgesetzt:
Der Knoten xk ist Knickpunkt, falls
oder falls
lu*"(Xk-l ) I
Literatur
[1] de Boor, C.: 1978, A Practical Guide to Splines. Applied Math.

Sciences, 27. Springer-Verlag.
[2] Reinsch, C.H.: 1967, Smoothing by spline functions.

Numer. Math. 10, 177-183.
[3] Schoenberg, I.J.: 1964, Spline functions and the problem

of graduation, proc. Nat.Acad.Sci. 52, 947-950.
Prof. M. v. Golitschek, Institut fiir Angewandte Mathematik und

Statistik, Universiti:i.t Wiirzburg, 8700 Wiirzburg, Bundesrepublik
Deutschland.
KNOT-ELIMINATION; REVERSAL OF THE OSLO ALGORITHM
David Handscomb
Numerical Analysis Group, Oxford University, Oxford, England
Given a spline function of any order, we consider the possibility

of approximating it, to within an acceptable tolerance, by a
spline function of the same order but with fewer knots, and of
achieving this end through a finite algorithm.
1. The Oslo Algorithm
In computer-aided design, it is common to represent an arbitrary
curve by a sum of B-spline basis functions. It frequently
happens that one wants to modify the shape of the curve in a
certain subrange, but finds that the knots originally laid down
do not give sufficient flexibilty to permit the modification
desired. One therefore starts introducing additional knots.
When one does so, however, one needs to change the B-spline basis
to one generated by the new set of knots.

104
Fortunately every element of the old basis can be expressed as a
linear combination of elements of the new basis. so that as a
first step we can express the unmodified curve exactly in terms
of the new basis. The so-called Oslo algorithm (COHEN et al ..
1980) is an algorithm for deriving the coefficients of the new
expansion from those of the old. We give first a simple
derivation of the formulae on which the algorithm depends.
Let Sn denote the space of spl ine functions of order n wi th
... } S[j]
knots {zi' i = 6 •• , -1. O. 1. 2. and let
n
denote
that of spl ine functions of order n wi th knots {zi' i = ••• I
-1. O. 1. 2 . . . . . j-1. j+1 • . . . } ; then is a subspace of
S
n
The unnormalised B-spline
M . (x) x] € S ( 1. 1)
nl n
where
n-l
g [z; x] := (z - x)+ . (1. 2)
n
is independent of the ordering of the knots z.I-n ••• t Z.

1
[This is not true of the normalised B-spline Nni .] We wish to
express a B-spline of
105
M[~](x)
nl
.= g n [z.1-n , ... , z. l' z. l'
J- J+
... , Zl'; x] (1. 3)
(where j ~ i ~ j+n+1 ), in terms of Mn ,i_1(x) and Mni(x) .
Now
Mn +1 .(x)
,1
[(x-Z.1-n- 1) Mn,l-
. l(x) + (z.-x)
1 Mnl.(x)]/(z.-z.
1 1-n- 1)
(1. 4)
bu t, because Mn+ 1 ,1. is independent of the ordering of its
knots, we can also write
Mn +1 .(x)
,1
[(x-z.) M . l(x) + (z.-x)
J n,l- 1
M[~](x}]/(z.-z.}
nl 1 J
(1. 5)
and
Mn +1 .(x)
,1
[(x-z.
1-n-
1) M[~](x)
nl
+ (z.-x) M .(x)]/(z.-z.
J nl J l-n-
1)'
(1. 6)
Equating any two of the last three expressions, we get
[(z.-z.
J
. l(x) + (z.-z.)
l-n- 1) Mn,l- 1 J Mnl.(x}]/(z.-z.
1 l-n- 1)'
(1. 7)
To extend the notation to B-splines not having Zj in their

106
support, let represent Mni if i > j+n+l or Mn,l-

• 1 if
i < j. Then
(1. 8)
where
b .. (Zj - zmin(j,i-n-l))/(zmax(j,i) - zmin(j,i-n-l})'

n1J
(1. 9)
(zmax(j,i) - Zj)/(zmax(j,i) - zmin(j,i-n-l))·

(1. 10)
Note that b nij L0 , c nij L0 , b nij + c nij = 1 , and that
for i L j+n+l , c nij =0 for i ~ j .
If now we have any spline function in s[j] , expressed in the

n
form
s(x) = (1.11)
then we can rewrite it as a spline function in s , i n the form

n
s(x) (1.11)
by taking
107
(1.12)
for all i This is the basis of the Oslo algorithm for
inserting a knot.
2. Knot Removal
If we continue adding new knots. we are liable to end up with
unnecessarily many parameters to represent the curve. In order
to avoid this. we also need an algorithm for removing a knot.
This is not simply the inverse of the process of inserting a
knot. since when we remove the knot Zj we need to replace a
spline in S by an approximation in S[j]

n n
In the spirit of spline analysis. we let the approximation differ
from the original spline over a finite interval only; say over
(ZJ' zK) where J < j <K The di screpancy will then be a
spl ine function with support (ZJ' zK) ; let S denote the
nJK
subspace of all such spl ines (a trivial subspace i f K-J < n) .
and let S[j] denote the subspace of all such spl ines with no
nJK
discontinuity at Zj (trivial if K-J < n+1).
If K-J = n the problem is purely algebraic; the discrepancy can
only be a multiple ( T • say) of MnK • so that the coefficients
{a.[j]} of the approximation s[j] are related to the

1
coefficients {ail of the original spline s by the equalities

108
a. b a[j] + c a~j] i #- K, (2. 1)

1 n,i+l,j i+l nij 1
[in particular a. a[j] , i < j, a. a~j] i > j+n ]

1 i+l 1 1
[.] [j]
aK = b n ,K+l,j a Kil + c nKj a K + T , (2.2)
and calculating the coefficients is very straightforward.
We may be able to make the overall magnitude of the discrepancy
considerably smaller, however, by taking K-J > n . [Consider
the analogy of the "economisation" of polynomials, where to save
computational effort in evaluation we replace a polynomial of
degree n by one of degree n-l , and so on, which we do not by

n
deleting the x term but by subtracting the appropriate
multiple of the Chebyshev polynomial Tn']
So we are looking for an element of SnJK of minimum norm
subject to the condition that the discontinuity of its n_2 nd
derivative at is equal to that of the original spline s at
this knot; an equivalent problem is: given any element of SnJK
find a best approximation to it in S[j] For most practical

nJK
purposes the supremum norm is the most natural to choose, so that
the problem is one of minimax approximation. [In some design
contexts, approximation of one or more derivatives may also be
important; this leads to interesting allied problems.]

109
Now. although minimax approximation by spline functions is in
general rather tricky. since no Haar condition is satisfied. yet
it happens in this particular case that we can show that the
minimax approximation has the equioscillation property and is
unique.
3. Proof of Equioscillation Property of Minimax Approximation
The Schoenberg-Whitney theorem tells us that there is a unique
spline of order n > 1 . having knots at zl < z2 < ... < zk and
taking any given values at xl < x2 < < x k +n if and only if
(3. 1 )
We may deduce immediately (by introducing n fictitious data
points beyond each end knot at which sex) o ) that there is a
unique element of SnJK (K-J+1 ~ n > 1) taking any given values

at i f and only if
ZJ+i-1 < xi < zJ+i+n-1 . 1 ~ i ~ K-J-n+1 . (3.2)
From this we now deduce that there exists an element of SnJK
(K-J ~ n > 1) . changing sign at xl < x 2 < ... < x K- J - n - d

(d ~ K-J-n) . and vanishing at no other point of (zJ' zK) . if
110
and only if d ~ 0 and
Zj+l < xi < Zj+i+n+d-l ' 1 ~ i ~ K-J-n-d . (3.3)
In particular, there is no element of SnJK changing sign at
K-J-n+l points (as we can show directly using Rolle's theorem).
Now let sex) be an element of SnJK' let s[j](x) be an
approximation to it in S[j] let rex) = s[j](x) - sex) , and

nJK
let q(x) denote an arbitrary element of S[j] Then s[j](x)
nJK
is a minimax approximation to sex) if and only if there is no
q such that IIr + qll < IIrll .
Suppose that rex) has K-J-n+l equioscillating extrema. Then
there is no q(x) changing sign at K-J-n points, so that q(x)
cannot have the opposite sign to rex) at all its extrema, and
IIr + qll > IIrll for all q. Therefore s[j] is minimax (and, by
a similar argument, is unique).
Suppose conversely that rex) has no more than K-J-n-d (d ~ 0)
equioscillating extrema, these being located at el < e2 < ... <

e K- J - n - d These extrema are K-j-n-d of the possible K-J-n+l
zeros of r'(x) € Sn-l,JK Therefore from (3.3), provided that
n > 2 , we must have
ZJ+l < e i < zJ+i+n+d-l ' 1 ~ i ~ K-J-n-d (3.4)

111
[note that inequalities (3.4) are strict]. We can therefore find
points xl < x2 < ... < xK-J-n-d-1 such that both
e1 < x.
1
< e i +1 ' 1 ~ i ~ K-J-n-d-1 (3.5)
and
zJ+i+1 < x.
1
< zJ+i+n+d-1 1 ~ i ~ K-J-n-d (3.6)
Therefore there is an element q(x) of s[j] changing sign at

nJK
the points {xi} , and so having the same sign as rex) at all
its extrema; therefore s[j] is not minimax.
Note that the result is false for n 2.
4. References
Cohen, E. et al. (1980) Discrete B-splines and subdivision

techniques Computer Graphics and Image Processing li,
87-111.
Dr D.C. Handscomb, Numerical Analysis Group, Oxford University

Computing Laboratory, 8 - 11 Keble Road, Oxford OX1 3QD, ENGLAND
FOURIER METHODS IN PRACTICAL APPROXIMATION
Werner HauBmann and Karl Zeller
Abstract. On the basis of the Caratheodory-Fejer

method (in conjunction with some auxiliary tools) we
obtain simple but good substitutes for the Zolotarev
polynomials and certain generalizations. Thereby we
give rather sharp error bounds for the approximations
involved. Further we pOint out certain relations to
surnrnability theory.
1. Introduction
The Caratheodory-Fejer method for approximation is

connected with an eigenvalue problem and a recursion, see e. g.
Gutknecht-Trefethen [2,9]. This makes it difficult to obtain
error estimates which are easy to use and rather sharp. In certain
cases (which are important for practice) the situation is better.
The main case in question is given by Zolotarev polynomials and
the corresponding approximation (cf. Section 3). We also describe
a finer error estimate and an improved approximation (Sections 3
and 4). Then we treat certain extensions of these results, where
we consider generalized Zolotarev polynomials (where more than
two terms are prescribed). We describe a method based on products
and on coefficient estimates (Sections 6 and 7). Finally we out-
line a relation to surnrnability theory (methods of Norlund type,
see Section 8) and give some additional remarks (Section 9).
113
2. Notation
Given the numbers

2
m E IN u {OJ ; -1 <b < 1 a := -b/(1-b ) ,
we consider (for Izl 1} the function
m+1 m+2
z 1-bz az +
H (z)
1_b 2 1-b/z
and the modified partial sum

2m+1
V (z) .- az
m+2 +
L b n z m+ 1 - n - a b2m+2 z -m .
n=o
Thus we have
H(z} = V (z) + (b/z) 2m+2H (z) •
Passing to real parts and employing the Chebyshev polynomials

(with T_k := Tk ) we are lead to
00
n
h(x} aTm+ 2 (x} + L b Tm+1 -n (x)
n=o
(for -1 ~ x ~ 1) and
2m+1
v(x} n - ab 2m + 2 T _m(x}
= aT m+ 2 (x} + L b Tm+ 1 _ n (x}
n=o
satisfying
00
n
h(x} v (x) + b 2m + 2 (aT (x) + L b T- m- 1 - n (x) ) •
-m
n=o
We consider V as an approximation to H and v as an approxi-
mation to h in the sense of the Chebyshev norm
IIHII
sup 1 H (z) 1 resp. IIhll := sup Ih(x} I.
Izl=1 -1 ~ x ~ 1
The approximation error will be treated below in more detail. In

particular, h can be considered as a substitute for a Zolotarev
polynomial (with given initial terms aT m+ 2 + Tm+1 and the
remainder minimizing the Chebyshev norm). Thus h gives a good
approximation to a function whose Chebyshev expansion contains
these two terms while the higher order terms are small or
negligible.
114
3. Basic Estimates
The function H has constant modulus
1
1H (z) 1
1-b 2
on Izl = 1, and winding number m+1. From this one can deduce
results concerning good resp. best complex approximations.
In the present paper we are interested in real approxi-

mations. For this purpose we note that h has an extremal alter-
nation of length m+2 (including the pOints -1 and +1). The
latter property is relevant for the second part of the following
resul t (see [3]):
PROPOSITION 1. The polynomial v described in Section 2 satis-

fies
II vII
Furthermore, given any polynomial p of degree not exceeding m,

then the following estimate holds true:
IIv-pll
Thus in many cases II vII is only slightly larger than

the basic value 2
(1_b )-1 (appearing in complex approximation) ,
and the corresponding Zolotarev polynomial cannot have a sub-
stantially smaller norm. We mention that the estimates in Propo-
sition 1 improve results by Bernstein [1] and Reddy [6], see [3].
It will be our next goal to diminish the gap between

the two bounds in Proposition 1. This will be done in the next
section by employing a refined estimate. In the following section
we shall introduce a refined approximation (which is more compli-
cated yet useful for certain purposes). Such fine tools are at
present not available for the more general cases (where three or
more terms are prescribed) which we shall treat in the final
sections.
115
4. Compensation
The upper bound for HvH in Proposition 1 was received

by estimating the two terms in
V(z) H(z) - (b/z)2m+2 H (z)
separately. By utilizing the interplay between these components

we obtain a refinement. Therefore we consider the representation
m+1 -m-1
V(z) (1_b 2m+ 2 )H(Z) + b 2m+ 2• z -z 1-bz
1-b2 ·1-b/z
where we are interested in the real part. The first fraction is
purely imaginary:
m+1
z - z -m-1
with z = e it Hence the imaginary part of the second fraction is

relevant. This fraction has modulus one while its argument
oscillates in the interval 1:= [-2arc sin b,2arc sin b] (con-
sider the circles described by the numerator and the denominator).
For b = 1/V2 we have I = [- ~ , ~] . Now we put
{ I :in (2arc sin b) I for Ib I :5 11 V2 ,

else,
and see that the imaginary part is contained in the interval

[-S,S]. This leads to
PROPOSITION 2. The polynomial v introduced in Section 2 satis-

fies
with S as given in the preceding definition.
We remark that in many cases this bound is below the basic value
(1_b 2 )-1, namely for small Ibl. A special example of this kind
is given in Gutknecht-Trefethen [2, p. 360/361]. Thus there are
certain fine distinctions between real and complex Caratheodory-
Fejer approximation.
116
5. Extended Polynomial
Instead of v we consider a modified partial sum with

two more terms
2m+3
v(x) : = aT m+ 2 (x) + L b n Tm+ 1 _ n (x) - ab2m+4T_m_2 (x)
n=o
Collecting terms and observing the identity 1 - ab
we get
v(x)
Now we put
v*(x) v(x)
1+b 2m + 2
m
a'T m+ 2 (x) + L
n=o
where
1_b 2m + 2 -b 1_b 2m + 2
a' := a
1+b 2m + 2 1-b 2 · 1+b 2m+ 2
The relations connecting a with a' and b with b' (which

satisfies a' = -b'/(1-b,2) are not too difficult to describe
and to evaluate. For large m and small Ibl the relative
change is very small (e.g. below one percent). Therefore the
following result is useful both for theoretical and practical
purposes:
PROPOSITION 3. The polynomial v* introduced above satisfies

Hb 2m+ 4
IIv*1I ~
1+b 2m + 2 1-b 2
with the correspondence between a' and b as given by the pre-
ceding formulas.
We remark that we gain a factor b 2 , but b is somewhat larger

than b'. (Of course this can be made more precise by considering
the expression for a' in more detail and developing an
iteration scheme.) Also a compensation as in § 4 is available.
117
6. Products
In addition to b and a we introduce -1 <d < 1 and
c ._ -d/(1-d 2 ) and put
G(z) 1 1-bz 1 1-dz

:= 1_b2 ·1-b/z· 1 _ d 2·1-d/z
-1 2 -2 -1 2-2
(az+1+bz +bz + •.. )(cz+1+dz +dz + ... )
acz 2 + (a+c)z + (1+ad+bc)zo + ...
The terms (for n = -1,0,1,2, ..• ) are
Let Gn(z) denote the remainder of this power series, beginning

with the z-n term. We use
(if b ~ d) and obtain
PROPOSITION 4. For n = -1,0,1,2, .•• , b ~ d, and Izl = 1, the

remainder Gn(z) can be estimated as follows:
lad n + 1 1 Ib n + 1 1 (1-lbl)-1+ld n + 1 1 (1-ldl)-1 ICb n + 1 1
I Gn (z) I =:; 1 - Id I + Ib - d I + 1 - Ib I
The middle term of this estimate can be replaced by

Ibod n + .•. + bndol + Ib n + 1 1
1 - Idl (1-lbl)(1-ldll
which is useful in particular in the case b =d (or when b-d
is small). Of course, the role of band d can be interchange~
Without loss of generality, we may assume Ibl ~ Idl,

thus we can use estimates of the type
00
L (v+1) Ib v I =:;
v=n
for the middle term. The result can be used for approximation
purposes, especially combined with the following considerations.
118
7. Coefficients
For applications it is desirable to adjust the para-

meters ~, a and c such that one obtains a given polynomial:
~(acz2 + (a+c)z + (1+ad+bc) uz 2 + vz + w,
which we normalize by w = 1.
One method is to start with the preliminary value c = 0,

and to modify the coefficients iteratively, finally obtaining
the goal.
Another way is to look for certain sets of (~,a,c),
and to find out what set (u,v,1) is covered. To this end we

consider lal:5 t, Icl:5 t with ac:5 0. Because of the symmetry,
it is sufficient to treat the case 0:5 a:5 t, -t:5 c:5 0. Denote
S := {(a,c) °:5 a :5 t, -t:5 c :5 O}

and consider the map ~ S -+]R2 defined by
1
~(a,c) .- 1+ad+bc'(ac,a+c) -. (u, v)
Then the following result holds true:
PROPOSITION 5. Let T be the triangle (in the (u,v)-plane) with

2
vertices (O,-t), (O,t) and (-t ,0). Then
_1_ . T c ~(S)
1 +2t2
For the proof we mention that the map ~: S -+ T with

(p(a,c) = (ac,a+c) is bijective and bicontinuous. From ac:5 0,
we have ad ~ ° and bc ~ 0, hence 1:5 1+ad+bc:5 1+2t2 =: r. Thus the
deformation by the factor (1+ad+bc)-1 yields the result. 0
This can be refined a bit by using r* := 1+ad+bc :5 1-2ac

rather than r. Also other cases can be treated along the same
lines. The case ac> ° is somewhat more complicated, but the
idea remains the same (in particular for small a,c): First look
at (ac,a+c), then deform with the factor 1+ad+bc < 1.
119
8. Summability
Given r EJN, we consider the real parameters
Pk with Po 0, and Pk = 0 for k<O and k> r,

and the relation '"
zo+ ... +Pr z
r
-r Po
AZ
0 r Uo Z
0
+ u 1z
-1
+ u 2z
-2
+ ...
po/z + ... +Pr/ z
In the Caratheodory-Fejer approximation (see Gutknecht-Trefethen
[2,9]) the values uo, ... ,u r are given; then one determines A,
po, ... ,Pr (by an eigenvalue problem) and finally u r + 1 ' u r + 2 '
(by a recursion). Here we describe a connection to summability.
We employ the infinite triangular matrix P with ele-

ments
for n, k 0,1,2, ••••
The relation above can now be written as
AP* Pu
(multiply with the denominator and compare coefficients; the con-

volution on the right hand side corresponds to the product Pu).
Matrices of the type P are known in summability (little Norlund
methods or "allgemeine Zweierverfahren", see Zeller-Beekmann [10D.
In particular, one treats there the inverse matrix Q: It has
the elements
o r -1 00 n
:= where (poz + ... +Prz) = L q Z
n=o n
From this one deduces (e.g. by partial fractions) information
about Q and the summability domain of P. We mention
PROPOSITION 6. The vector u (uk) is given by AQp* u.
Since the vector p* is of finite type, only the first

columns of Q enter into the result, and these columns arealmost
equal (apart from a translation process). Thus in principle we
know the behaviour of the uk (for k > r) and can deduce infor-
mation about the error in Caratheodory-Fejer approximation.
120
9. Remarks
In Section 2 we have described modified partial sums of
the CF-series. In certain cases such expressions yield a sequence
of proxima (see Rivlin [7,8]). Further we mention pre-iteration
(a method to obtain a good initial alternation for a Remez algo-
rithm (cf. Meinardus [5], Hollenhorst [4]). Our approximations of
Zolotarev type can be used for this purpose since the extrema of
the main error term are accessible by analytic and numerical cal-
culations. Finally, the product considerations (Sections 6 and 7)
can be refined and extended to more than two factors.
References
1. Bernstein, S. N.: Collected Works, Vol. 1. Akad. Nauk SSSR,
Moscow 1952.
2. Gutknecht, M. H., Trefethen, L. N.: Real polynomial Chebyshev
approximation by the Caratheodory-Fejer method. SIAM J.Numer.
Anal. 19, 358-371 (1982).
3. Haussmann, W., Zeller, K.: Approximate Zolotarev polynomials.
Compo Math. Appl., to appear.
4. Hollenhorst, M.: Improved lower and upper bounds in polynomial
Chebyshev approximation based on a pre-iteration formula.
J. Approx. Theory 32, 170-188 (1981).
5. Meinardus, G.: Approximation von Funktionen und ihre numeri-
sche Behandlung. Springer: Berlin-Heidelberg-New York 1964.
6. Reddy, A. R.: A note on a result of Zolotarev and Bernstein.
Manuscripta math. 20, 95-97 (1977).
7. Rivlin, T. J.: Polynomials of best uniform approximation to
certain rational functions. Numer. Math. 4, 345-349 (1962).
8. Rivlin, T. J.: Some explicit polynomial approximations in the
complex domain. Bull. Amer. Math. Soc. 73, 467-469 (1967).
9. Trefethen, L. N., Gutknecht, M. H.: The Caratheodory-Fejer
method for real rational approximation. SIAM J. Numer. Anal.
20, 420-436 (1983).
10. Zeller, K., Beekmann, W.: Theorie der Limitierungsverfahren.
Springer: Berlin-Heidelberg-New York 1970 (2nd edition).
Werner HauBmann Karl Zeller

Department of Mathematics Department of Mathematics
University of Duisburg University of TUbingen
D-4l00 Duisburg D-7400 TUbingen
ON TIME-MINIMAL HEATING OR COOLING OF A BALL.
Werner Krabs
Fachbereich Mathematik der TH Darmstadt
1. The Problem
This paper is a supplement and partial correction of [2]. Considered is a

ball of homogeneous material of radius R> 0 which is heated or cooled in such
a way that the temperature at an arbitrary point in the ball only depends on
its distance r from the center of the ball and the time t?: O. The thermal
stress tangential to the surface of the ball at the time t (under suitable
physical assumptions) is then given by (see [1], p. 64)
R
a(R,t) = ~{-!J f S(r,t)r 2 dr - S(R,t)} ( 1.1)
I-I-! R 0
where S(r,t) denotes the temperature in the ball at the distance r from the
center and at the time to Further, E is the modulus of elasticity and I-! and
a are the coefficients of cross-extension and linear heat extension,
respectively. We assume
S(r,O) = SO(r) for all r E [O,R] (1. 2)
where SOEC[O,R] is some given temperature. The ball is assumed to be heated

or cooled on its surface in the form
S(R, t) = u(t) for all t;:: 0 (1.3 )
where u = u(t) is a control functi on whi ch, for every T > 0 is vari ab ly
chosen in C[O,T]o Under the above symmetry assumption the temperature S(r,t)
for t>O and O<r:,;R develops according to the heat equation
122
8t (r,t) = a~8(r,t) = ~~ (r 2 8 (r,t)) (1.4 )

r ()r r
where a> 0 denotes the coefficient of heat conductivity.

If we define
v(r, t) = r8 (r, t) for O:s; r :s; R, t 2: 0 (1. 5)
it follows that
vt(r,t) =avrr(r,t) for O<r<R, t>O (1. 6)
v(r,O) = r8 0 (r) for O:s; r:s; R, (1. 7)
v(O,t) = 0 and v(R,t) = Ru(t) ( 1.8)

for all t 2: 0,
if 8 = 8(r,t) is assumed to be bounded.

Conversely, a solution v = v(r,t) of (1.6), (1.7). (1.8) leads to a bounded
solution 8(r,t) = ~ v(r,t) of (1.2). (1.3), (1.4).
By [4], Kap. III, §2 there exists exactly one generalized solution v = v(r,t)
of (1.6), (107), (1.8) which is of the form
2r kn 2
v(r,t) = R r: [c k exp( - a(T) t)
00
k=l
+ (-1)k+1 kna ~ exp(-a(~)2(t-s))U(S)dS] sin ~r

o
where
R
ck = f p8 0 ()
P .
Sln kn p dp f or k En.
R Yol
o
We assume for the following that
80 (r) = 80 for all rE [O,R]

123
and some constant 80 , Then

2
ck = (_I)k+I ~n 80 for k EI'l
and we obtain as unique bounded solution of (1.2), (1.3), (1.4)

k+1
exp(-a(~)2t)80 1. sin kn r
00
8(r,t) = 2r r: (-1,
k=I n r R (1.9)
+ ~ ; (_I)k+lknj exp(-a(~)2(t-s))u(s)ds ..!. sin kn r.

" k= 1 0" r R
Insertion into (1.1) by use of (1.3) then yields
cr ( R) ra
,t =""l-U 6 r: 1 exp «-a 1{
kn) 2t )8 - u( t )
{:2" ~ 0
00
~ R k=I (kn) (1.10)

00 t
+ ~ r: f exp(-a(~)2(t-S))U(S)dS}.
R k=I 0
Problem of Time-Minimal Control: Given a target temperature 8E on the

surface of the ball. find the smallest possible time T>O and a control
function u E C[O.T] such that u(T) = 8E and, for some given value cr* > 0, the
thermal stress cr(R,t) tangential to the surface of the ball and given by
(1.10) satisfies the condition
sup Icr(R,t) I s cr* . (1.11)

tE[O,T]
On defining
(1.12 )
we obtain the equivalent problem: Given 8EE lR and d > 0, find the smallest
time T > 0 and some u E C[O,T] with u(T) = 8E such that
6 1 k 2
sup 1:2" r: --2 exp(-a(i) t)8 0 - u(t) +
00
(1.13 )
tE[O,T] R k=I (kn)
124
co t
+ ~ f exp(-a(~)2(t-S))U(S)dsl ~ d.
E (1.13)
R k=l 0
We shall see that this problem is closely related to the following

approximation problem: Given 6E ElR and T> 0, find some uT E C[O,T] such that
u(T) = 6E and the left-hand side of (1.13) becomes minimal for u = uT.
2. On Solving the Approximation Problem
For every T>O let UT = C[O,T] be equipped with the maximum norm
BUilT = max lu(t) I, u E Ur

tE[O,T]
If, for every u E UT, we define

t
KTu(t) =f k(t-s)u(s)ds (2.1)
o
where
k(t) = 6a ~ exp(_a(kn)2t ), t E (0 T]
if k~l T ' ,
(2.2)
then one can show that KT is a continuous (even compact) linear operator that
maps UT into UT and is positive, i.e.
u(t) ~ OvtE[O,T] ~ ~u(t) ~ OvtE[O,T].
Since
T
I~II ~ f k(t)dt ~ 1 - "T
o
with
6 1 kn 2
=~
co
"T E ~ exp(-a(T) T),
n k=l k
125
it follows that ~KT" <1 for every T>O. Therefore the linear operator
KT : UT - UT being defined by
KT u (t) = u (t) - KT u (t) , t E [0, T] , (2.3)
has a continuous inverse K~1 : UT - UT which is monotone, i.e.
If we define
(2.4)
then 80 EU T for every T>O and the approximation problem consists of finding
some u EC[O,T] with u(T) = SE that minimizes
We put
(2.5)
and
(2.6)
Obviously, we have
-1 ~
d(T,SE) = inf{llvllTI VEU T , KT (V+SO)EH(T,SE)}
= i nf { !Iv II T I v E UT '
-1
KT v (T ) = SE - KT-INSo (T) } .
If we defi ne, for every u E UT,
gT (u) = u (T) ,
126
then we can also write
where (KT1)* is the adjoint operator of KT1 being defined by
- 1)*U*(V)
(KT = u*( KT-1)
v *
VUE * an d v E Ur
UT
Further we have
with e ::: L
Therefore, by a well-known formula for d(T,8 E) in its last representation
(see, for instance, [3]), it follows that
-1 - -1 -
18E - KT 80 (T) 1 18E - KT 80 (T)1
d (T , 8E) _I * ~'---->-1--=--
i(KT ) gT' KT e(T)
If we now put
8E- KT-1 -80 (T)
(2.7)
vT = KTI e(T) e,
then
and
i . e. ,
If we further define
(2.8)
127
with vT being defined by (2.7), then
uT(T) = KT-1 VT(T)+KT

-1 -
60 (T) 6E, i.e. U T EH(T,6 E)
and
which implies that uT solves the approximation problem. This solution can be
represented in the form
-1
uT = 6E - KT 60 (T) K-1 e + K.:1 g
K:r l e(T) T ", 0
and can, therefore, be obtained by solving the two integral equations
(2.9)
and
(2.10)
for t E [O,T] with k = k(t) given by (2.2) and by defining

6C YT(T) _ _
uT(t) = _ uT(t) + YT(t), t E [O,T]. (2.11)
uT(T)
If 60 = 0, hence 80 = 0, then YT = 0 is the unique solution of (2.10) and a

solution uT E UT of the approximation problem is given by
6E _
uT(t) = -_- uT(t), t E [O,T], (2.12)
uT(T)
with uT E UT being the unique solution of (2.9) and

16 EI
(2.13)
128
3. On the Approximate Solution of the Approximation Problem for the Case of

Vanishing Initial Temperature
We assume 80 = 0, i.e., the initial temperature to be zero in which case a

solution of the approximation problem is given by (2.12) with UTEU T being
the unique solution of (2.9).
For every NElN and every T> a 1et
6a N kn 2
kN(t) = ~ E exp(-a(T) t), tE [O,T]. (3.1)
R k=l
In order to obtain an approximate solution of (2.9) we consider instead the

equation
t
uT,N(t) - 6 kN(t-S)UT,N(s)ds 1, t E [O,T], (3.2)
with kN = kN(t) being defined by (3.1). We first observe that the linear
operator KT,N : UT ~ UT given by
t
KT ,N u( t) = 6 kN ( t - s )u(s )ds , UE UT' (3 . 3 )
is compact. Furthermore, we have

T T
6
IIRT ,Nil ~ kN(t)dt ~ k(t)dt ~ 1- 13T 6
with 13T E (0,1). Therefore the operator
KT ,N u(t) = u(t) - KT, N u(t ) , u E UT'
has a continuouse inverse K~~N which is monotone. Consequently (3.2) has a

unique solution uT,N E UT which is taken as an approximation of the unique
solution UTEU T of (2.9). In turn
8E ~
uT N(t) = ~ uT N(t) (3.4)
, u (T) ,
T,N
is taken as an approximation of the solution uT E UT of the approximation
129
problem given by (2.12).

From (2,1) and (3.3) we infer, for every UEUT'
(KT - KT,N)u(t) 1:
k=N+1
3 0f exp(-a(~ )2(t-S))u(s)ds
R
t
which impl ies

6 1 6
max I (KT - KT N)u(t) I s 2" 1: -:-2 ~uUT s To IUl r
tE[O,T] , n k=N+1 k n N
Further, we have the error estimate
which in connection with
implies
which justifies the choice of uT,N as an approximation of uT.
4. On Solving the Problem of Time-Minimal Control for the Case of

Vanishing Initial Temperature
We again assume 80 = 0, First we observe that the unique solution uT E UT of

(2.9) has a representation of the form
t
uT(t) = 1 + f r(t-s)ds, tE [O,T], (4.1 )
o
where rEC(O,T] nL 1(0,T) is defined by
r ( t) = 1: kn ( t) , t E [0, T], (4.2)

n=l
with
130
t
k1 (t) = k(t) given by (2.2) and kn(t) = f kn_1 (t-s)k(s)ds (4.3)
o
for n = 2.3 •.... This implies that r is a positive function and hence the
function T ~ uT(T) is strictly increasing. By virtue of (2.13) and because of
uT(T) > 1 for all T> 0 we have
19 EI
d(T.9 E) = - _ - < 19E I· (4.4)
uT(T)
Thus. if d~ 19EI. for every T>O. there is some UEU T with u(T) = 9E and
IKTu DTsd (see (1.13)). namely u = uT defined by (2.12). Therefore the
problem of time-minimal control has no solution. In order to make this
problem meaningful we have to assume that d < 19EI. If we then show that
(4.5)
it follows from lim d(T.9 E) = 19 E I and the fact that T + uT(T) is strictly
T+O+
increasing. hence T + d(T.9 E) strictly decreasing (by (4.4)). that there is
exactly one T*>O such that d(T*.e E) = d. Obviously. T* is the minimal time
and uT* defined by (2.12) for T = T* is a time-minimal control.
Similar to UTEU T the unique solution UT.NEU T of (3.2) can also be

represented in the form
t
uT.N(t) = 1 + 6rN(t-s)ds. tE[O.T].
where rNEC[O.T] is defined by
with
t
kN• 1 (t) = kN(t) given by (3.1) and kN.n(t) = bkN.n_1(t-S)kN(s)dS
for n = 2.3 •.•.• From kN(t) s k(t) for all t ~ 0 it follows. by virtue of
(4.2). (4.3) that rN(t) sr(t) for all t~O and therefore uT.N(T) suT(T) for
131
all NEtl and all T > O. By using Laplace transformation for solving (3.2) one
can show that, for every NEI'l
This in connection with uT,N(T) ~ uT(T) for all NElN and all T > 0 impl ies
(4.5) and therefore ensures the solvability of the problem of time-minimal
control, if d < leEI with d given by (1.12).
References
[1] Parkus, H.: Instationare Warmespannungen. Wien 1959.

[2] Peter, G.: Optimale Anheiz- und AbkUhlvorgange mit der AuBenrand-
temperatur als SteuergroBe. Beitrage zur Analysis ~ (1981), 99-108.
[3] Psenicnyj, BoN.: Notwendige Optimalitatsbedingungen. Leipzig 1972
(Obersetzung aus dem Russischen).
[4] Tychonoff, A.N. und Samarksi, A.A.: Differentialgleichungen der
mathematischen Physik. Berlin 1959.
Author's address:
Fachbereich Mathematik der TH Darmstadt, SchloBgartenstr. 7, 6100 Darmstadt

THE FUNDAr~ENTAL SPLINES OF PERIODIC HER~lITE INTERPOLATION

FOR EQUIDISTANT LATTICES
Gerhard r'lerz, Gesamthochschule Kassel - Universitat, FRG
O.
We describe three methods for the construction of the fundamental splines of

periodic Hermite interpolation in the case of equidistant lattices. Apart
from the generalized Euler-Frobenius polynomials our main tools are the dis-
crete Fourier transform, a complex line integral representation and a certain
eigenvalue problem for the computation of null splines.
1.
Let be given the (r,N)-matrix ((y~p))), v = O(I)N-l, p = O(I)r-l; r ~ 1,

N > 1. We are interested in the construction of a N-periodic polynomial
spline function s of degree 2k+l and continuity class C(2k-r+l) (IR)
with knots in the integers such that
s(p)(v) = y(p) , v = O(I)N-l, p = O(I)r-l .

v
In order to have a well-posed problem the consistency condition r < k+l has
to be satisfied.
2.
If we denote the restriction of s to the interval v-I -< x -< v by pv

then it has been shown in [4] that after a suitable shifting of coordinates
the vector
q(t) o< t < 1

133
of N successive polynomial components of s results as the solution of a

system of N cyclic difference equations. Explicitly, we have
r-1
.9.(t) L W* Q (t)Wy(p) (1)
p=O p
where
W: = _1 (( ))N-1 21Ti
~ ~
11\1
11,\1 =0 ~ = exp N
denotes the matrix of discrete Fourier transform,

(p) 11
Qp(t):= diag h2k+1,r(t,~ ) 11 = 0(1)N-1, p 0(1)r-1, (2)
are certain diagonal matrices and
The functions h~kl1,r (t,z) , p = 0(1)r-1 , which occur in (2) are deter-
mined as follows: Define for Izl <1 the generalized Euler-Frobenius poly-
nomi a 1s by
n = 0,1,2, ...
put
r-1
h~kl1,r(t,Z):= ~O a~~~i~r (z) H2k +1_o (t,z) (3)
and fix the coefficients a(o,p)

2k+1,r
(z) in (3) by the interpolation conditions
h~kL,r (t,z) Iz=l = <5

po'
• p 0
,
0(1)r-1 (4)
In [2] it has been shown that the common determinant of the linear systems
(4) can be written in the form
~2k+1,r(z) = (-1)[r/2]0!1! ... (r-1)!(1-Z)(r-1)(4k-r+4)/2 H2k +1 ,r(1,Z) (5)
where H2k +1 ,r(1,z) is a polynomial of degree 2(k-r+1) in z (cf. [4],[5]).

134
Examples. (i) For r = 2 we obtain

2k (0)
(l-z) H2k+1,2(1,z)h2k+1,2(t,z) =
and
2k+1 (1)
(l-z) H2k +1,2(1,z) h2k +1,2(t,z) =
In the special case r = k = 2 we have
(Z2-6Z+1)h~~~(t,Z) = 4(z-l)(z2+4z+1)t 5 - 5(z-l)(3z 2+8z+1)t 4 +

+ 20z(z-l)(z+l)t 3 - 10z(z-l)2t 2 + z(z2_6z+1)
and (6)
(Z2-6Z+1)h~~~(t,Z) = (z3+11z2+11z+1)t5 - (4z 3+33z 2+22z+1)t 4 +
+ 2z(3z 2+14z+3)t 3 - 4z(z-l)(z+l)t 2 + z(z2_6z+1)t
(ii) Similarly, for r = 3,

(1-Z)4k-2H2k+1,3(1,Z)h~~ll,3(t,Z) =
= k(2k-1)[(2k-1)H 22k _2(1,z) - (2k-2)H2k_1(1,z)H2k_3(1,z)lH2k+1(t,z) +
2
+ (4k -l)[(k-1)H2k(1,z)H2k_3(1,z) - kH2k_1(1,z)H2k_2(1,z) 1 H2k (t,z)+
2 _ (1,z) - (2k-1)H2k(1,z)H2k_2(1,z)lH2k_1(t,z)
+ k(2k+1)[2kH 2k 1
4k-1 (1) _
(l-z) H2k+1,3(1,z)h2k+1,3(t,z)-
135
2
+ [k(2k+l)H 2k _1(1,z) - (k-l)(2k-l)H2k+l(1,z)H2k_3(1,z)lH2k(t,z) +
+ k[(2k-l)H2k+l(1,z)H2k_2(1,z) - (2k+l)H2k(1,z)H2k_l(1,z)lH2k_l(t,z) ,
2(l-z)4kH2k+l,3(l'Z)h~~ll,3 (t,z) =
2
[2k H2k_l(l,z)-(2k-l)H2k(1,z)H2k_2(l,z) 1 H2k +1(t,z) +
+ [(2k-l)H2k+l(l,z)H2k_2(l,z) - (2k+l)H2k(l,z)H2k_l(l,z)lH2k(t,z) +
+ [(2k+l)H~k(l,Z)-2k H2k+l(l,z)H2k_l(l,z)lH2k_l(t,z)
In the special case k = r = 3 this gives

(Z2+8Z+1)h~~§(t,Z) = -15(z-1)(z2_ 6z+1)t 7 + 35(z-1)(2z2_9z+1)t6 -
- 21(z-1)(6z 2-17z+1)t 5 + l05z(z-1)2t 4 - 35(z-1)z(z+1)t 3 + z(z2+8z+1)
(Z2+8Z+1)h~:§(t,Z) = -5(z3_7z2_7z+1)t7 + (24z 3-117z 2-128z+11)t 6 -
-3(15z 3-38z 2-49z+2)t 5 + 40z(z-1)(z+1)t 4 - 5Z(3z 2+8Z+3)t 3 + z(z2+8z+1)t ,
2(Z2+8Z+1)h~:§(t,Z) = -(z-1)(z2_22z+1)t 7 + (5z 3+216z 2+69z-2)t 6 -
- (lOz3_131z2+62z-1)t5 + lOz(z2_ 7z+1)t 4 - 5z(z-1)(z+1)t 3 + z(z2+8z+1)t 2
Remark. Although the method described so far works well in practice, the
following properties may be considered, in a certain sense, as more or less
significant limitations of its range of practicability:
(i) The construction of a real spline function requires a detour into the
complex domain.
136
(ii) Even in the presence of fast Fourier transform techniques the computa-
tional work involved in the construction of an interpolating spline function s
increases heavily with N.
The next paragraph is dedicated to an investigation of the problem of finding

out to what extent these difficulties may be reduced or even completely be
removed.
3.
The solution of our interpolation problem only requires the construction of

the respective fundamental splines
q (t)
...:.p = (ql ,p (t), q2 ,p (t), ... ,qN ,p (t)) T , P = O(I)r-l
defined by the special interpolation conditions
dO
- q. (t)1 opO 0'1 p,O O(I)r-l, j 1(I)N .
dt O J,P J
t=O
From (1) we obtain for p = O(I)r-l and j 1(I)N

1 N-l
qj,p(t) = N L ~-jflh(P) (t ~fl)
fl=O
2k+l,r '
By observing that
z N-l 1
res ~ =N
z =~ . . z
11 -1
the theorem of residues immediately leads us to
f (p) zN-l-j
h2k +1 ,r(t,z) ~ dz (7)
r~2 -c1
Here, C1 and C2 denote positively oriented circles about the origin with
radii PI = p and P2 = p-1 , respectively, where p < 1 is fixed in such
137
a way that C1 encloses the k-r+1 zeros zl < z2 < •.• < zk-r+1 of
H2k +1 ,r(1,z) which are located in the interior of the unit circle.
A crucial point in our investigations is now marked by the fact that according
to the symmetry relation
(p)
h2k +1 ,r(t,z -1) -_ ( -1) p z-1 h2k
(p)
+1,r ( 1-t,z )
the integral
f h~kl1 r(t,z) _z-N- dz

,
N-1-j
z -1
-1
via the substitution z + Z gets transformed into
_~-1
12 -
p
rr1 f (p)
h2k +1,r(1-t,z)
zj-2
I-;N dz
As a consequence, in the actual evaluation of (7) the N-th roots of unity

play no longer any role and we simply arrive at
(8 )
k-r+1
+ .E ( 1_z \!N)-l[(_1)Pzj-2
\!
res h(P) (l-t z)+zN-j-1 res
2k+1,r' \!
h(P) (t z)]
2k+1,r '
\!=1 Z=Z
\! Z=Z
\!
For further details of this method we refer to [5].
Remarks. (i) By the way we may mention that many qualitative properties of
the Hermite fundamental splines are more or less direct consequences of
formula (8).
(ii) In a recent paper, REIMER and SIEPMANN [7] have given a purely
algebraic derivation of a representation for the fundamental splines which
turns out to be analogous to (7) and (8) in the case of Lagrange data.
It seems possible to get on with a similar procedure in the present instance
of Herrnite interpolation.
138
4.
The following considerations are generalizations of arguments used by LEE [1]

in connection with Lagrange interpolation. They result in a new method for
the construction of the Hermite fundamental splines. At the same time we gain
insight into the intrinsic structure of expression (8) as a combination
of null splines and polynomial correction terms.
For reasons of simplicity we start with a look at the case k = r = 2. Here,
Obviously, the functions h~~~(t,O) may be characterized for p = 0,1 by

the Hermite interpolation conditions
h~~~ (1,0) = 1 h~~~ (1,0) =0

(10)
. (0)
h5 ,2 (1,0) = 0 h~~~ (1,0) = 1
and
h~~~ (0,0) = h~~~ (0,0) = ~~~~ (0,0) = ~ ~~~ (0,0) = 0 , (11)
respectively.
If we now define for p = 0,1
(12)
then with zl:= 3 - 2-v2 we arrive at
res (i-6z+1)-1 (13)

z=zl
i.e.
res h~~~(t,Z) = 5(10-71{2)t 2(t-l)2(2t+1{2-1) (14 )

z=zl
139
res h~:~(t,Z) = 4(7-51{;2)t 2(t-1)2(2t+1f:2-1) (14)

z=zl
As a consequence of (6) together with the relations
Hn (O,z) = z Hn (l,z) , n~ 1
and
a
at Hn (t,z) = n (l-z) Hn- 1 (t,z) n > 1
it follows that the polynomial f 5,2(t) of degree five on the right hand
side of (14) is up to a constant factor uniquely determined as solution
of the eigenvalue problem (cf. [3])
. .
f 5 ,2(0) = f 5 ,2(0) = f 5 ,2(1) = f 5 ,2(l) = 0
(15)
If one puts
then (15) leads to a homogeneous linear system with determinant
H5 ,2(1,z -1 ) = z-2-1
-6z +1
which for zl = 3 - 21{2 and with the normalization aD = 1 has the solu-
tion
(16)
In order to determine the factor Co in
we proceed as follows: The coefficient of t 5 in g~~~(t,Z) equals,

140
according to (6) and (12), 4(z-1)(l+4z+1). Consequently
Co = 4(ZI-I)(zi+4Z 1+1) = 80 (7-5-Vz) (18)
From (13) together with (16) - (18) we now conclude in accordance with
(14)
In the same way the factor c1 in
results from
In the general case the functions h~kll,r(t,O) and res h~kll r(t,z)
Z=Z v '
which occur on the right hand side of (8) in the components q.J ,p (t) of
the fundamental splines are given as follows:
1. The functions h~kll,r(t,O) are the unique solutions of the r Hermite

interpolation problems which arise from the conditions
, a = O(l)2k-r+l
for fixed p , p = O(l)r-l
2. Denote by Zv ' v = 1(I)k-r+l , the zeros of H2k +1,r(l,z) which lie

in the interior of the unit circle.
141
The functi ons res are constructed in two steps:

z=z v
(ex) For fixed v, v = 1(1)k-r+1 , determine the (normalized) polynomial
f~~ll,r(t) from the conditions (cf.(15))
~
dtO
f(v)
2k+1,r
(1) = 0 , ° = 0(1)r-1
~ f(v)
dtO 2k+1,r
(1) = z-l
v
dO f(v)
dtO 2k+1,r
(0), ° = r(1)2k-r+1
(p)
(8) Denote by ~2k+1,r ()
z the determinant which is obtained from the common
coefficient determinant ~2k+1,r(z) of the linear systems (4) by replac-
ing its first column by the vector
(0, ... , 0, 1,0, ... ,O)T

t
p-th component
and define for p = 0(1)r-1 (cf. (5))

(p) ( )
_--::____~;:,.:2k..:....+..;:1.!...,r'---z_v_______ (l-z ) 2k+ 1
[r] v
(-1) 7 0!1! ... (r-1)!(1-z )(r-1)(4k-r+4)/2
v
Then (cf. (12), (13»

(p)+ ,r ( t,z ) = c (v) f (v)+ ,r ()
res h2k -1
1 p 2k 1 t . res [H 2k +1 ,r(1,z)l
z=z v z=z v
Remark. The method just described is of utmost simplicity in the case k = r,

where only one null spline has to be determined. If we let k = r = 3 ,
then
142
and with the zero zl = 1{15 -4 of H7,3(I,z) = z2+8z+I we obtain
res h~~~(t,Z) = if-(7 m -27)t 3(t-I)3(6t-3+m )

z=zl
res h~~~(t,Z) = j(9m -35)t3(t-I)3(6t-3+m )

z=zl
res h~~~(t,Z) = ~(7m -27)t 3(t-I)3(6t-3+m)

z=zl
The components of the respective Hermite fundamental splines are now easily
available according to (8).
143
References
[1] LEE, D. (1986) A simple approach to cardinal Lagrange and periodic

Lagrange splines. J. Approximation Theory 47, 93 - 100
[2] LEE, S.L. and A. SHARMA (1976)Cardinal lacunary interpolation by

g-splines I. The characteristic polynomials. J. Approximation Theory ~,
85 - 96
[3] LIPOW, P.R. and I.J. SCHOENBERG (1973) Cardinal interpolation and spline
functions III. Cardinal Hermite interpolation. Linear Algebra and Its
Appl. ~, 273 - 304
[4] MEINARDUS, G. und G. MERZ (1980) Hermite-Interpolation mit periodischen

Spline-Funktionen. Numerical Methods of Approximation Theory, ISNM 52
(Birkhauser, Basel), 200 - 210
[5] MERZ, G. und W. SIPPEL (to appear) Zur Konstruktion periodischer Hermite-
Interpolationssplines bei aquidistanter Knotenverteilung. J. Approxima-
tion Theory
[6] REIMER, M. (1982) Extremal spline bases. J. Approximation Theory 36,

91 - 98
[7] REIMER, M. and D. SIEPMANN (1986) An elementary algebraic representation

of polynomial spline interpolants for equidistant lattices and its con-
dition. Numer. Math. 49, 55 - 65
Prof. Dr. Gerhard Merz, Gesamthochschule Kassel - Universitat, Fachbereich 17-

Mathematik, Nora-Platiel-Strasse 1, D-3500 Kassel, Federal Republic of Germany
STRONG UNICITY CONSTANTS IN

CHEBYSHEV APPROXIMATION
Giinther Niimberger
Mathematisches Institut
Universitat Erlangen-Niirnberg
Erlangen, W.-Germany
Abstract
We give formulas for computing or estimating strong unicity constants in Ohebyshev
approximation by finite-dimensional spaces, in particular by weak Ohebyshev spaces.
Results of this type were proved by several authors in the special case of Haar spaces
and spline spaces. Strong unicity constants are used to obtain error estimations for
approximations computed by Remez type algorithms.
Introduction
We investigate Ohebyshev approximation of functions in O(T) by functions from a
finite-dimensional subspace G of O(T) in the supremum norm 11/11 = SUPtET I/(t)l. A
function gl EGis called strongly unique best approximation of IE O(T), if there exists
a constant KI > 0 such that for all U E G, 11/- ull ~ 11/- gill + Killu - u/lI. The strong
unicity constant K(f) of I is defined to be the maximum of all such numbers K I .
These constants can be used to obtain error estimations in the computation of best
approximations: A Remez type algorithm for computing a best approximation glEG
of a given function IE OCT) yields a sequence of approximations (gm) in G converging
to g, and a sequence (..\m) of real numbers converging to 11/- g,1I which satisfy I..\ml S
11/- g,1I ~ 11/- gmll for all m (see e.g. HETTIOH & ZENOKE (S) and NURNBERGER
& SOMMER (8)). If gl is a strongly unique best approximation of I, then the error
IIgm - gill can be estimated as follows:
1 1
IIgm - u,1I S K(f) (11/- Umll-II/- g,lI) ~ K(f) (11/- gmll-I..\ml).
In this paper we show that for a given function I E O(T) the strong unicity constant
K(f) can be computed or estimated by the minimum of the norm of functions from
G which satisfy certain interpolation properties at the extreme points of the error 1-
U,. Results of this type are proved for arbitrary finite-dimensional spaces and for weak
145
Chebyshev spaces. They were established earlier by several authors for Haar spaces and
spline spaces ((I], [3], [4], [6], [9]). Most of our results were announced without proofs in
[7].
1. Finite-Dimensional Spaces
We use the following notation. Let T be a compact Hausdorff space. The space of all
continuous real-valued functions on T is denoted by OCT). Moreover, let {'h ... ,'n} be
a basis of an n-dimensional subspace a
of OCT). Given points tl, ••. , tn e T we denote
by
D ( 'I,···,'n
til ... ,tn
)
the determinant det ('i(tj))~,j=I. The set {th ... ,tn } is called poised w.r.t a,
if the
above determinant is different from zero. For a given function Ia e OCT) we define
E(Ia) = {t e T: Ih(t)1 = IIhll}

and
I h(t) > 0
sgn Ia(t) ={ 0 if h(t) = 0 .
-1 h(t) < 0
The number of points of a given set A is denoted by card A.
In the following we derive formulas for computing or estimating the strong unieity
constant in the case of arbitrary finite-dimensional spaces.
We need a characterization of strongly unique best appro:ximations due to WULBERT
(11].
Theorem 1.1 Leta 6e afinile-dimen,;onal ItullptJee ofO(T), I e O(T)\G tJ>nd" e G.

The followin, """eme"', an: equiflalent:
(i) The funefion " i, a "ron", unique buI approsimalion of I from G.
(ii) For ne,., nonlriflialfunetion g e G,
min (/(t) - gilt)) get) < o. (Ll)
fEEI,-,,1
Moreover, we need the following general formula for the strong unieity constant. (The
proof of Lemma 2.1 in NURNBERGER [6] for spline spaces can be applied to this general
case.)
Lemma 1.2 Let G 6e a finite-dimentional ,ub'paee of OCT) and" e G be a Ifron",

unique kIf approsimation of f e OCT) \ G. Then file laafle
K(f) = r:M fEW- I"n (f(t) - gilt)) get)

11,11=1 "
= min{I/II,II: g e G, "n (f(t) - g,(t» get) ~ 1 for all t e E(f - g,)}.
146
In the following we give formulas for the strong unicity constant which are simpler
than that in Lemma 1.2. In particular, we show that the strong unicity constant can be
computed or estimated by the norm of interpolating functions from G.
We first give a formula for K(f) in the ease that the dimension of Gis n and the error
has exactly n + 1 extreme points. Theorem 1.3 was proved by HENRY & ROULIER [41
for Hur spaces and by NURNBERGER [61 for spline spaces.
Theorem 1.3 Let G be a finite.dimentional ,ub'pace of O(T) and g, E G be a ,trongl"

unique but approsimation of I E O(T) \ G ,ueh that E(f - gil = {t l , ••• , tn+!}' For
each i E {I, ... ,n + I} let g; be the unique function in G ,uch that
j = 1, ... , n + I, j #= i. (1.2)
Then file hafle

K(I) = min{I/lIgdl : i = 1, ... , n + I}. (1.3)
Proof. We first show that the sets
i = 1, ... , n + I, (1.4)
are poised w.r.t. G. Suppose that there exists a set {tl"'" ti-I, ti+l!" .,tn+d which is
not poised. Then there exists a nontrivial function g E G such that
get;) = 0, i = 1, ... , n + I, i #= j.
By replacing g by -g, if necessary we may assume that
(f(t;) - g,(t;» get;) ~ 0, i= 1,,,.,n+1.
Since E(I - gil = {tl,,,.,t n+.} this contradicts Theorem 1.1. Since the sets (1.4) are
poised, the functions gl, ... , gn+! are uniquely determined.
By Lemma 1.2 there exists a function go E G such that sgn (I(t) - g,(t)) get) 5 1
for all t E E(I - g,) and K(f) = 1/ligoli. We set A =
{t E E(f - gil : go(t) =
sgn (I(t) - 9,(t»} and choose a point to E T such that Igo(to)1 = IIgoli. We may assume
that the function go is choosen such that card A is maximal. Then by the proof of Lemma
1 in SCHMIDT [91 there does not exist a function g E G such that
sgn (/(t) - g,(t» get) < 0, tEA, and sgn go (to) g(t o) > O. (1.5)
Then A #= 0, otherwise by selting g = go we get a contradiction to (1.5). We set A =

{WI"" ,.om}' Then we have m 5 n. Otherwise A =E(f - 9,) and by the definition of
A we get (f(t;) - g,(t;) go(t;) = (I(t;) - gl(t;)) sgn (f(t;) - 91(ti» = III - gIll> 0, i =
1, ... , n + 1. This contradicts Theorem 1.1. If m = n, then we have done. Suppose that
m < n. We choose arbitrary distinct points W m +!,. ", Wn-I fromE(I-g,)\{wl'" .,tOm }.
Then the set {WI!"., .on-I! to} is not poised. Otherwise, there exists a function 9 E G
147
such that ,(ei) = -sgn (/(ti) - ,,(ti», i = 1, ... , m, and ,(to) = sgn 'oCto), which
contradicts (1.5). Now, there exists a nontrivial function 'I e G such that
'I(ei) = 0, i = 1, ..• , n - 1, and '1 (to) = o. (1.6)

=
We set {en,en+l} E(I - ,,)\ {e., .•. , en-I}. Since {el! ... , en} is poised, itfollows
from (1.6) that 'I(en) ::F o. Then there exists a real number e such that
,o(en) + e'l(en) = sgn (/(en ) - g,(en». (1.1)

Then, since sgn (f(e n)':'" ,,(1I7n» ,o(1I7n) < 1, it follows that
c sgn (/(1I7n ) - g,(1I7n» gl(en) > O. (1.8)

Therefore, esgn (/(en+l) - ,,(en+l» 'I(en+l) < 0, otherwise together with (1.6) and
(1.8) this contradicts Theorem 1.1. This inplies that
sgn (/(en+l) - ,,(en+l» ('o(wn+l) + e'I(1I7n+l» < 1. (1.9)
It follows from (1.6), (1.1) and (1.9) that
sgn (/(t) - g,(t)) (,o(t) + e'l(t)) ::s; 1, t e E(I - g,).

Therefore, by (1.6) and Lemma 1.2 K(f) ::s; 1/l1'o+eglll ::s; 1/11go11 = K(f) which implies
that K(f) = Ill1go + egili. Since by (1.6) and (1.1)
card {t e E(f - gIl : ('o(t) + e'l(t)) = sgn (f(t) - ,,(t»} ~ m + 1,

this contradicts the maximality of card A. This shows that card A =n and proves
Theorem 1.3.
We now give a lower bound for K(/). Theorem 1.4 was proved by OLINE (3) for
Haar spaces and by NURNBERGER (6) for spline spaces.
Theorem 1.4 Ld G = 'lH'R {'ll ... , gn} H II ,uJ,ptUc 01 O(T), I e OCT) \ G _tl
" e G ,ucla thai 'lacre csi" n + 1 poi"" t l , ••• , tn+l e E(I - ,,) wi,1a
(1.10)
i=l, ... ,n+l, O'e{-l,l}. (1.11)

For C/lela i e {I, ... ,n + I} lei lai H flac unique I_dio,. ira G ,ucla flaGf
j = 1, ... ,,, + 1, j ::F i. (1.12)
K(f) ~ min{l/l1la;/1 : i = 1, ... ,,. + I}.

148
Proof. We set E = {tt, ••• , t"H} and show that

min max sgn (/(t) - g/(t» get) > O.
,EO fEEI/-"l
11,11=1
Suppose that there exists a function g e G such that IIgll = 1 and
sgn (/(t) - g/(t» get) ~ 0, teE. (1.13)
Since tT"+l ::F 0, there exists a basis {i., ... , i,,} of G such that for all j e {I, ... , n},
fj(ti) = 0, i = 1, •.. , n, i::F j, and fj(tj) = -sgn (f(tj) - g/(tj)). (1.14)
We may assume that {g., •.• ,g,,} has already this property. Then by (1.13) g 2:1=1 ajgj, =
where al ~ O, ••• ,a" ~ O. It follows from (1.11) that tT"H =
(n~=I(-l)iHtTtTi) and for
all j e {l, ... ,n},
tTj = (-l),,-j sgn giCt"H) (n

i#j
(_l)iH tTtTi) .
This implies that
(-l)"H tT tT"H sgn g(t"+1) = (_l)"H tTtT"H (Sgn t aj gAt"H») >

}=1
O.
On the other hand, it follows from (1.11) and (1.13) that
(_1),,+1 tT tT"H sgn g(t"H) = sgn (f(t"H) - g/(t"H)) sgn ,(t"H) ~ 0

which is a contradiction. This shows that a > O. Since a > 0, it follows from Theorem
1.1 that g/ is a strongly unique best approximation of /. Furthermore, by the proof of
Lemma 2.1 in NURNBERGER (6)
0= min{l/l1gl1 : g e G, sgn (f(t) - g/(t» get) ~ 1 for all teE}.
Now, analogously as in the proof of Theorem 1.3 we can show that there exists a function
go e G such that 1/11go11 = a ~ K(f), sgn (f(t) - gtCt» go(t) ~ 1 for all teE and
= =
card ({t e E: go(t) sgn (/(t) - g/(t»)}) n. This proves Theorem 1.4.
By using Theorem 1.3 we now give an upper bound for K(f) in the case that the
dimension of G is n and the error has exactly n + 1 extreme points. Corollary 1.5 was
proved by BLATT (I) for Haar spaces and by NURNBERGER (6) for spline spaces.
OoroUary 1.5 II G i, an n·dimeuional '.'paee oIO(T) and g/ e G i, a "rongl,

unique int approsimafion 01/ e OCT) \ G ,ue" that E(f - gIl contain, esadl, n + 1
point" then K(f) ~ ~.
149
Proof. We set {t., ... , t"+l} = E(/- 9/). Since 9/ is a strongly UDique best approxima-
tion of I, by the proof of Theorem 1.3 the sets
i = 1, ... ,n+ 1, (1.15)

are poised w.r.t. G. By CHENEY [2, p.73) there e:xist real numbers ~h ••• '~"+1 with
o ~ ~i ~ I and L::~II ~i = I such that for all 9 e G
n+l
L ~i sgn (/(ti) - 9/(ti» g(ti) = O. (1.16)
i=1
It follows from (1.15) that ~i > 0 for all i e {I, ... , n+ I }. Let the functions gh ... , 9n+l e
G be defined as in Theorem 1.3. Then by (1.16) for all i e {I, ... , n + I},
_1_< _1_ _ ~j
119jll - Ig;(tj)1 - (1 - ~j).
We set~, = minpj: i = 1, ... ,n+ I}. Then it follows that

n+1
1 = L ~i ~ (n + IP,
i=1
which implies that

_1_ < ~, <!.
11,,11 - (1 - ~,) - n
Finally, it follows from Theorem 1.3 that
K(f) = min {II:ill :i = 1, .•. , n + 1} ~ II~II ~ ~.

This proves Corollary 1.5.
2. Weak Chebyshev Spaces

In this section we give a formula for K (f) in the case of approximation by functions
from weak Chebyshev spaces.
An n-dimensional subspace G of O(T), where T is a compact subset of the real line,
is called weak Chebyshev, if there e:xists a basis {9I, ... ,gn} of G such that for all
points tl < ... < tn in T,
D( gl,···, 9n ) > o.
tl, .. . ,t" -
We use the following notation. Let G be an n-dimensional weak Chebyshev subspace
of O(T) and g/ e G be a strongly unique best approximation of Ie O(T) \ G. We define
y = {r = (t., ... ,t n ): tl < ... < t" in E(f - g/); {th ... ,t,,} is poised w.r.t. G;
150
sgn (f(ti) - UI(ti» sgn (f(ti+1) - UAti+1)) = -1,

i= 1, ... ,n-1,
or sgn (/(ti) - 91(ti» sgn (f(ti+l) - 91 (ti+1)) = -1,
i = 1, . .. ,n -1,
i ::f.i, for some ie {1, ... ,n-1}}.
For each 11 = (t 1 ,. •• , t,,) eY let U, be the unique function in G satisfying
i = 1, ... ,no
We now give a formula for K(f) in the case that G is weak Chebyshev and the error
has only finitely many extreme points which is the most important case in practice.
Theorem 2.1 was proved by SCHMIDT [91 for Haar spaces and by NURNBERGER [61
for spline spaces without the assumption that E(f - U/) is finite.
Theorem 2.1 Let G be an n·dimen,ional weak Ohd"hn ,ub'pace of O(T), where T i,

a compact ,ab,et of the real line and UI e G be a ,trongl, unique but approzimation of
f e O(T) \ G. If the ,et E(f - U/) i, finite, then
K(f) = min {1/119,1I : 11 e Y, 'In (f(t) - 91(t)) 9,(t) ~ 1 for all t e E(f - 9/)}'
Proof. We set e(t) = sgn (/(t) - Ut(t» for all t e T. By Lemma 1.2 there exists a
function UO e G such that e(t) Uo(t) ~ 1 for all t e E(f - U/) and K(f) = 1/11Uoll. We set
A = {t e E(f - Ut) : Uo(t) = e(t)} and choose a point to e T such that 190(to)1 = 119011·
Moreover, we define c(t) to be e(t), if teA and to be -sgn 90(t), if t = to. Then by the
proof of Lemma 1 in SCHMIDT [91 there does not exist a function ,eG such that
c(t) u(t) < 0, t e Au {to}.

We set G(t) = (91(t), ... , 9,,(t)) for all t e T, where {9l!'" ,U,,} is a basis of G. Then it
follows that there does not exist a z e Rn such that
< c(t) G(t), z > < 0, teA u {to}.

Since A u {to} is compact, it follows from CHENEY [2, p.191 that
o e cony {c(t) G(t) : teA u {to}}.

By Caratheodory's theorem (cp. CHENEY [2, p.17J) there exist points U1 < ... < Um+h
1 ~ m ~ n in A U {to} such that
o e CODV {C(Ui) G(Ui) : i = 1, ... , m + I}.

This implies that there exist real numbers Ah'''' Am+l > 0 such thatEi:~l Ai 1 and =
L:~11 Ai C(Ui) g(Ui) = 0 for all 9 e G. Then, since G is weak Chebyshev, by standard
151
arguments (cp. SINGER [10], p.178, Theorem 1.3, Proof of (2) ~ (4) ) there exists a
D'E {-1,1} such that
D' (-1); c(u;) = 1, i = 1, ... ,m + 1. (2.1)
Furthermore, we have to E {Ub"" Um +1}, otherwise E~il,x; C(Ui)gO(Ui) E:1 1 ,xi e(u;)' =
> O,which is a contradiction. Then there exists an integer p E {I, ... , m + I} such that
to = up. If there exists a function go and points 1£1, ••• ,um +! as above such that m = n
and {UI"'" Up-I' 1£1'+1,"" Un+l} is poised w.r.t. G, then the formula for K(f) is proved.
Suppose that a function and points with the above property do not exist. In the following
we will show that this leads to a contradiction. Let {gh"" 9n} be a basis of G. We first
show that there exist points UmH,"" Un such that the function 9 E G, defined by
()
gt=D' D ( gl, ..... • ,gn )
, tET,
Uh ••• , Up-I, 1£1'+1, ••• , Un, t
where D' E {-I, I}, is nontrivial and has the following property
(-1); get) ~ 0, t E [Ui,Ui+I] nT, i = 0, .. . ,p - 2;

(_1)P-I get) ~ 0, t E [Up-I,Up+l] n T; (2.2)
(_I)i+lg(t) ~ 0, t E [U;,Ui+I] nT, i = p+ 1, ... ,m + 1,
= =
where 1£0 -00 and UmH 00. To show (2.2) we first note that, since {gb"" gn} are
linearly independent, there exist points Wh'" ,Wn-l E T such that the function h E G,
defined by
h(t) = D ( gh ... , gn )
WI, ••• ,Wn-i, t
is nontrivial. We now consider all sets {UI, ••• , um+d as above such that there exist
points UmH < ... < Un in T for which the function g E G, defined by
9-()-D(91'
t - ...... ,gn) , tET,
Uh ••• ,1£,,-1, u,,+b"" Un, t
is nontrivial. If no such set {Uh"" um+d exists, then in the definition of g we set
UmH = WI, ••• , Un = Wn-l' If m =
n - 1, then, since G is weak Chebyshev, there exists
°
a sign D' E {-1,1} such that the function 9 = D'g satisfies (2.2). If m < n -1, then
we choose UmH as above to be maximal. If the maximum is obtained, then get) =
for all t E [UmH' 00) and therefore, since G is weak Chebyshev, the function 9 = D' g
satisfies (2.2). If the maximum is not obtained, then we choose instead of UmH a strictly
increasing sequence (UmH,,) converging to the maximum and define the corresponding
sequence of functions (g,) analogously as above. Then these functions satisfy (2.2) except
on (UmH",UmH) and in the subsequent arguments we may replace 9 by a function 9,
for a sufficiently large r. For the sake of shortness we only consider the case when the
maximum is attained. Therefore, we may assume that (2.2) holds. We assume that go is
152
chosen in such a way that m is maximal. Moreover, running from i = 1 to i = p - 1 we

choose the points Ui as above to be maximal. Then running from i p + 1 to i =
m+1 =
we choose the points 1Ii as above to be minimal. By replacing 9 by -g, if necessary, it
follows from (2.1) that
get) e(1I1) S 0, t E (-00,111) n T;

get) e(1Ii) ~ 0, t E [1Ii, UiH) n T i = 1, ... ,p - 2;
t E [U,-1I 1I,H) n T; (2.3)
get) e(1IiH) ~ 0, t E [1Ii, UiH) n T i = p + 1, ... , m;
get) e(UmH) SO, t E [Um+1, 00) nT,
We set
Al = {t E An (-00,111) : go(t) = -e(uI)},

A1 = l..f;:": {t E An [1Ii, Ui+l) : go(t) =e(Ui)},
A. = {t E An [U,-1I U,H) : 90(t) =e(u,_I)},
At = Ui=,H {t E An [Ui, UiH) : 90(t) =e(Ui+.)}'
A& = {t E An [UmH, 00) : go(t) = -e(umH)}.
Then it follows from the ma.ximality of m and the extremal choice of the points UI! ••• ,
get) = 0,t E uf=IA;. (2.4)

We now show that g(to) = o. Suppose that g(t o) ¢ o. It follows from (2.3) and (2.4)
that there exists a sufficiently small a > 0 such that e(t) (go(t) + ag(t» S 1 for all t E
E(f -g/). Moreover, by (2.1) sgn 9(to) = e(1I,_1 and then by (2.3) sgn 9(to) = sgn 90(to).
This impJies that
1190 + 01 911 ~ 190(to) + a9(to)1 > Igo(to)1 = 119011

and therefore
1 1
K(f) = 119011 > 1190 + 01911'
which contradicts Lemma 1.2. We now set
BI = {t E E(f - 9/) n (-OO,UI) : e(t) = -e(uI)},

B1 = Lf;::{t EE(f - 9/) n (1Ii,UHI): e(t) = e(1Ii)},
B. = {t E E(f - 9/) n (11,-1, 1I,H) : e(t) =e(1I,_I)},
B. = Ui=P+1 {t E E(f - 91) n (1Ii,1IHI) : e(t) =e( UHI)},
B, = {t E E(J - gf) n (1Im+l. 00) : e(t) = -e(1Im+l)}.
Then it follows that
there exists a point t E U~=IBi with 9(t) ¢ o. (2.5)

153
Otherwise, it follows from (2.3) that (/(t) - g,(t)) (-g(t)) ~ 0 for all t e E(f - gIl
which contradicts Theorem 1.1. As above it follows from (2.3) and (2.4) that there exists
a sufficiently small 01 > 0 such that for g()t = go + OIg we have e(t) g()t(t) ~ 1 for all
t e E(f - g,). We may assume tha.t 01 is chosen to be ma.ximal. It follows from (2.5)
that 01 is finite. We now set
A = {t e E(f - fIJ) : g()t(t) = ee,)},

0 1 = {t e An (-oo,al) : g()t(t) = -e(al)},
,
a, = l..!;:"1 {t e An (aj, aj+I) : g()t(t) = e(aj)},
~
O. = {t e An (a,-I. a,+1) : g()t(t) = e(a,_I)},

o. = U::',+I{t e An (aj,aHI): g()t(t) = e(ai+l)},
or. = {t e An (am+h 00) : g()t(t) = -e(am+l)}.
Then it follows that U~=1 OJ :/= I, otherwise, there exists a sufficienly small P > 0 such
that e(t) g()t+" ~ 1 for all t e E(f - g,), which contradicts the maximality of 01. This
finally leads to a contradiction. We first note that
1 1
K(f) = IIgoll ~ IIg()t1l = K(f),
since g(to) = o. If there exists a point a e O,UO.UO., then this contradicts the extremal
choice ofthe points ah ... , u,_11 u1+h ... , Um+h since obviously g(a) :/= o. If there exists
a point 111 E 0 1 U a" then we get a contradiction to the maximality of m, since obviously
g(1I1) :/= O. This proves Theorem 2.1.
Remark 2.2 We obtain the formula for K(f) in Theorem 2.1 also for the case when
E(f - gIl is not finite except that in the definition of Y the condition that the sets
{tho .. ,tn } are poised w.r.t G has to be omitted.
This can be shown by the following density argument. Let the assumptions of The-
orem 2.1 hold. There exists a minimal number of subsets El < ... < E, of E(f - g,)
such that E(f - 9,) = l!;=IEj and 1- 9, is constant on E j, i = 1, ••. ,po Then for each
i E {I, ... ,p} there exists an increasing sequence (Ej,m) of finite subsets of E j such that
since Ej is compact. We set Em = l!;=IEj,m for all m. Then by standard arguments

there exists a sequence (1m) in OCT) converging to I such that " is a strongly unique
best approximation of 1m and E(/m - 'I) = Em for all m. Since the sets E(fm - 'I) are
finite, we can now apply Theorem 2.1 to each function 1m and obtain the formula for
K(/m) as in Theorem 2.1. Then it is not difficult to verify that by going to subsequences
and taking limits we obtain the formula for K(f) as in Theorem 2.1 except that the
poisedness of the sets appearing in Y may be lost.
154
Some numerical examples concerning strong unicity constants for splines are given in
NURNBERGER [7].
Reference.
[I] Blatt, H.P. (1982) Strenge Eindeutigkeitskonstanten und Fehlerabschatzungen bei
linearer Tschebyscheff-Approxima.tion, in Numerical Methods of Approximation
Theory, Oollatz L., Meinardus G. and Werner H. eds. (Birkhauser, Basel), 9-25.
[2] Oheney, E.W. (1966) Introduction to approximation theory (McGraw-Hill, New
York).
[3] Oline, A.K. (1973) Lipschitz conditions on uniform approximation operators. J.
Approx. Theory 8, 160-172.
(4) Henry, M.S. and Roulier, J.A. (1978) Lipschitz and strong unicity constants for
changing dimension. J. Approx. Theory 22,85-94.
[5] Hettich, R. and Zencke, H. (1982) Numerische Methoden der Approximation und
semi-infiniten Optimierung (Teubner, Stuttgart).
[6] Niirnberger, G. (1982-83) Strong unicity constants for spline functions, Numer.
Funet. Anal. Optimiz. 5, 319-347.
[7] Nurnberger, G. (1983) Strong unicity constants for finite-dimensional subspaces,
in Approximation Theory IV, Ohui O.K., Schumaker L.L and Ward D. eds., (Aca-
demic Press, New York), 643-648.
[8] Niirnberger, G. and Sommer, M. (1983) A Remez type algorithm for spline func-
tions. Numer. Math. 41, 117-146.
(9) Schmidt, D. (1980) A characterization of strong unicity constants, in Approxima-
tion Theory ill, Cheney, E.W. ed. (Academic Press, New York), 805-810.
[10] Singer, I. (1970) Best approximation in normed linear spaces by elements of linear
subspaces (Springer, Berlin).
[11] Wulbert, D.E. (1971) Uniqueness and differential characterization of approxima-
tions from manifolds of functions. Amer. J. Math. 18,350-366.
G. Nurnberger, Mathematisches Institut, Universitat Erlangen-Nurnberg, BismarckstraBe

1 1/2, 8520 Erlangen, W.-Germany.
ON A GAUSSIAN QUADRATURE FORMULA FOR ENTIRE FUNCTIONS OF

EXPONENTIAL TYPE
Q.I. Rahman and G. Schmeisser
Dept. de Mathematiques et de Statistique, universite de Montreal

and Mathematisches Institut der Universitat Erlangen - Nlirnberg.
1. The quadrature formula of Gauss
Given a system of n distinct pOints x 1 ,x 2 , .•. ,x n let
n
w(x) := n(x-x ) and
v=1 v
(v= 1,2, ... ,n).
Then for all polynomials P of degree <n we have
n
P(x) = I P(x )Q, (x)
v=1 v v
so that
+1 +1
J P(x)dx where Wv := J 2v (x)dx
-1 -1
(v= 1,2, ••. ,n).
This is a quadrature formula for all polynomials of degree < n.

156
If we treat both Xv and wv as 2n unknowns, then, perhaps we

can arrange matters in such a way that the formula will be valid
for all polynomials of degree < 2n, (Le. linear combinations of
2 2n-1
the 2n powers 1,x,x , ... ,x ). This is indeed possible as was
found by Gauss who proved
Theorem A. If -1 < x 1 < x 2 < ••• < xn < 1 are the roots of the nth
Legendre polynomial then we can find positive constants w 1 ,
w2 , •.• ,wn such that
+1
(1) f P{x)dx
--1
for all polynomials of degree < 2n.
2. Turan's generalization of Gauss' formula
It was proved by Hermite [5] that for any given system of n

distinct pOints x 1 ,x 2 , ... ,x n and a positive integer k there
exist polynomials
(v 1,2, ... ,n)
each of degree :;,; kn-1 such that t{~) (x ) = 0 for all

vJ 11
v,ll,j,h (1:;';v:;';n,1:;';I1:;';n,o:;,;j:;';k-1,o:;';h:;';k-1) except for v=1l and
h=j; in the latter case
Hence the representation
n k-1
P{x) L L p{h){x)t h{x)
v=1 h=o v v
157
and the quadrature formula

+1
(2) J P(x)dx
-1
where +1
AVh := f ~Vh(x)dx (1 S v S n, aSh S k-1)
-1
are valid for all polynomials of degree ~ kn-1. Generalizing the

problem settled by Gauss, Turan [9] asked himself the following
question: "Does there exist a system of n real numbers x 1 ,x 2 '
.•. ,xn such that the quadrature formula (2) is true for all
polynomials of degree at most m where m;;: kn?"
For even k the answer is easily seen to be "no". In
fact, if for some real system x 1 ,x 2 , .•. ,xn the formula were
k
valid for all polynomials of degree S kn, then for P(x):={w(x)}
+1
we would obtain f P(x)dx since P (h) (x )
v
=a for
-1
0:;; h :;; k-1, 1 :;; v :;;; n. But P (x) being non-negative and not identi-
cally zero, its integral on [-1,1] cannot be zero. Hence for even
k, there does not exist any real system x 1 ,x 2 , ... ,xn such that
(2) is true for all polynomials of degree S kn. In addition, it
was proved by Turan [9] that in case of odd k there exists a
unique system of paints x1k,x 2k , ... ,x nk such that the quadratu-
re formula (2) is valid for all polynomials of degree S(k+1)n-1
if we take Xv xvk for v = 1,2, ... ,n. The paints x 1k '
x 2k ,.·.,xnk are characterized by the fact that the polynomial
n
TIn,k(x) := r-T(x-x vk )
v=1
minimizes the integral

+1
(3) Ik(P) := f Ip(x) I k + 1 dX
-1
158
in the class Pn of all monic polynomials of degree n. The case

k=1 of this result agrees with Theorem A since the properly
normalized Legendre polynomial of degree n does minimize the
Integral I1 (P) in Pn (see [7, Theorem 3.1.2]).
3. An analogue of Gauss' formula for entire functions of exponen-

tial type
A polynomial of degree <n is completely determined by its

values at n distinct points. A theorem of F. Carlson [3] says
something similar about entire functions of exponential type. If
f and g are two entire functions of exponential type less
than a >0 such that
f(vn/a) g(vn/a) for v 0,±1,±2, •..
then f(z) =g(z). In other words, an entire function of exponen-

tial type <a cannot vanish at all the points
vn/a (V = 0,±1,±2, ... )
without being identically zero. The example sin(az) shows that

the restriction on the exponential type cannot, in general, be
relaxed. More generally, according to a theorem of Valiron [10]
if f is an entire function such that
uniformly in a, and if f(z) vanishes at least once in each

interval
vn/a :;; x < (v+1) n/a (v 0,±1,±2, ... ),
then f(z) =0. For other results of this kind see [1, Chapter 9].
The theorems of Carlson, Valiron, etc. state that a
function of a specified class is completely determined by its
159
values on a certain set of points {A}oo • In many situations,

v v=-oo
the function can be represented by an interpolation series in
which the numbers f(A v ) appear as coefficients. Here is an
example:
Let {A}oo be an increasing sequence of real numbers
v v=-oo
such that A - A f: 0 > 0, t v 7T - A t~ L (v = 0, ± 1 , ± 2, ... ) with
v+1 v v °
O<L<~ and let
° N I
~(z) = lim(z-A )~ (1-Z/A )

N+oo 0 -N v
where the term corresponding to v=o is omitted from TT l
Further, let 0::: T < 0. Then for an arbitrary entire function f

of exponential type T bounded at the pOints {AV} , we have
[ 1, formula (10.5. 12) ]
00
(4 ) f(z) I
v=-oo
for each sufficiently large (depending on {A }) positive integer

v
O-T
k and all EE(o,~ ).
Let {A } = {A }+oo be an increasing sequence of
v v v=-oo
real numbers with AV tending to +00 as v + +00 and to - 00 as
v _00. We shall say that {A } is a uniqueness sequence of
+
v
~ °
if an entire function of exponetnial type <0 cannot
vanish at all the pOints AV without being identically zero. By
US(o) we shall mean the set of all uniqueness sequences of type 0.
The sequence {AV} will be called a sequence of class (p, T) with
0< P < 00, if there exists an entire function of order p type T
vanishing at all the points AV' It is clear that a sequence of
class (p,T) cannot be a uniqueness sequence of type ° if either
p < 1 or if p=1 and T<O. Let us denote by C(1,o) the set
of all sequences of class (1,0) and by US(1,o) the intersection
of US(o) and C(1,O). Each element{A)ofUS(O) has the
property that an arbitrary function of exponential type <0 is
completely determined by its values at the points AV but those
belonging to the subset US(1,o) are amongst the sparsest.
160
An entire function of exponential type which belongs to

L1 (-00,00) is necessarily bounded [1, Theorem 6.7.15] on the real
line. Hence if f is an entire function of exponential type
T <0 and belongs to L 1 (_00,00), then from (4) we obtain
00 00
(5) J f (x) dx
_00
00
where
_X)}k
:= J 1
lj!' (A )
lj! (x)
v
x=x--
v
{Sin
E:
E: (A
(A -x ~
v
dx, (V =0, ± 1 , ± 2, .•. ) .
In analogy with the problem considered by Gauss, we

may now ask if there exists an element {AV} of U8(1,0) with
the property that formula (5) remains valid for all entire func-
tions of exponential type <20 belonging to L1 (_00,00). The
answer is "yes" and we can, in fact, do a bit better. For this,
we need to recall Poisson's formula from the theory of Fourier
integrals, according to which, if
J
00
g(x) = - 1 e ixt G(t)dt

v2TI_ 00
and
J e -itxg(x)dx,
00
G(t) 1
J2;
_00
then (see [8, p. 60] and also [2] )
00 00
(6 ) L g(V) LG(2nv).
v=_oo v=-oo
If f
is an entire function of exponential type 20 belonging to
L1 (_00,00), and so also to L 2 (_00,00), then the function
n 2
g(z) := f(oz) is of exponential type 2n and belongs to L (_00,00).
Then, by the Paley-Wiener Theorem [1, Theorem 6.8.1] the Fourier
transform of g must vanish outside the open interval (-2n,2n).
Hence from (6) we obtain ([2], also see [4]):
161
Theorem B. Let 0 < 0 < 00. For every entire function of exponential
~ 20 belonging to L 1 (-00,00), we have
<X>
J f(x)dx
<X>
(7 ) ~ L
v=_oo
f (V1T/O) .
_00
Theorem B provides us with an example of a sequence {A }, viz.

v
{vn/o}, belonging to U8(1,0), which has the property (hereafter
called the "Gaussian property") that formllia (5) remains valid
for all entire functions of exponential type 20 belonging to
L1 (_00,00). We may ask if there are other elements of U8(1,0)
which have the same property. In order to answer this question
we start with the observation that if A ={A V} is an element of
U8 (1,0) I then so is 0v -o} for each 0 E lR. Therefore, we may and
hereafter we will assume that A =0. We then have to decide how
o
the other A 's should be distributed in order that {A} may
v v
have the Gaussian property. By assumption, there exist entire
functions fA of order 1 type 0 vanishing at all the pOints
AV(V = 0,±1,±2, ..• ). Let J A denote the set of all such func-
tions. It may be noted that if A has the Gaussian property, then
none of the functions fA in JA can belong to L2 on the real
line. The sequence A = 0v} will be said to have the "modified
L2 property" if there exist functions fA in J A such that
00
J -2 2
x fA (x)dx < 00. The sequence {V1T/O} which we know, has the Gauss-
-<X>
2
ian property also has the modified L property. In fact, we on-
ly need to think of the function sin(oz). Are there other ele-
ments {A } of U8(1,0) which have the Gaussian property as well
v
as the modified L2 property? We are going to prove that the
answer is "no" - see Theorem 2.
Let us note that if A = {AV} belonging to U8(1,0)
has the Gaussian property as well as the modified L2 property
then there is one and only one fA in JA such that
<X>
J x -2 fA2 (x) dx < 00 and 1. The existence of such a function

- 00
162
2
follows from the modified L property whereas its uniqueness is
a consequence of the Gaussian property. The main results will be
deduced from the following
Lemma. Let A = {AV} belonging to US(1,0) have the Gaussian

2
property as well as the modified
~~~~~~~~~~~~~~~=
L property.
-1 Further, let
be the function in J A such that ~A(z) := z fA(z) belongs
2
to L (-00,00) and ~A(o) = 1. Then ~A minimizes the integral
J ,f(x) ,2dx over all entire functions of exponential type ° who-

-00
se value at the origin is 1.
Proof. Let f ~ ~A be an arbitrary entire function of exponen-

tial type ° assuming real values on the real axis and belonging
to L 2 (_00,00) with f(o) = 1. Then the function
h(z) := f(z) - ~A (z) ~ 0
belongs to L 2 (_00,00) and h(o) = o. Further, fA(Z)h~Z) is an

entire function of exponential type 20 which vanishes at all
the points A. Since {A } is supposed to have the Gaussian
v v
property, it follows that
00 00 00 00
2
J f 2 (x)dx J ~~(X)dX +
J
h (x)dx + 2J f (x) h (x) dx
A x
_00 -00 -00 -00
00 00
J ~~(X)dX + J h 2 (x)dx
-00 -00
>J~~(X)dX.
_00
If f(z) := L avz has complex coefficients and J 'f (x) ,2dx < 00 1
V=O -00
then the above reasoning may be applied to the function L ClVZ v

V=O
163
where c. Re a v for v = 0,1,2, ... to obtain

v
00 00 00
00
(8)
f [f (x) [2x ~ f IV=o
I c.vx V[2 dx >
f [ (j) It (x) [2 dx .
-00 -00 -00
00 00
2
The conclusion that [(j) It (x) [ dx is, of course,
00 _00 -00
trivial if f [f(x) 12dX 00.
-00
Applying the lemma to the sequence {vTI/a} we con-

clude that if f is an entire function of exponential type a
such that flo) f 0, then
]!,
a
_00 -00
i.e.
f
00
[ f (0) I 2 ~ ~ [f (x) [ 2 dx
_00
which, clearly. remains true if flo) happens to be o. The same

conclusion obviously holds for the function f(z+x) for every
o
x E lR and so we have
o
Theorem 1. Let f be an entire function of exponential type a

2
belonging to L (-00,00). Then for every x E lR
o
00
[ f (x ) I 2
o
~ ~TI f [f (x) [ 2 dx
-00
sin(az)
where equality holds if and only if f(z) := c
z
In view of (8) we also have
Theorem 2. The sequence {vTI/a} is the only element of US(1,a)

which has the Gaussian property as well as the modified L2 pro-
perty.
164
Remark 1. Notice the similarity between the (normalized) nth

Legendre polynomial
n n(n-1) n-2
x - 2 (2n-1l x + •••
whose zeros solve the problem considered by Gauss and the function
sin(oz)
o
whose zeros solve the corresponding problem for entire functions

of exponential type. Whereas P minimizes the integral
n
1
J I P (x) 1
2 dx over all monic polynomials of degree :s n the function
-1 00
x -2 1f(x) 12 dx
so(z) minimizes the integral
J
_00
over all entire
func·tions f of exponential type 0 such that f (0) = 0 and

f'(o) =1.
4. An analogue of Turan's formula for entire functions of

exponential type
Again, let Ii. : = {A) ~=-oo be an increasing sequence of real num-

bers with AO and A -+ too as v -+ ± 00. Given k E :IN , the
=0
v
sequence Ii. will be called a k unig;ueness seg;uence of t:i!2e 0 i f
an entire function of exponential type <ko cannot vanish along
with all its derivatives of order :Sk-1 at the points AV without
being identically zero. The sequence {vn/o} is a k uniqueness
sequence of type 0 for each k E :IN. It is clear that a k unique-
ness sequence of type 0 must also be a uniqueness sequence of
type o. By k-US(o) we shall mean the set of all k uniqueness
sequences of type 0 and by k-US(1,o) the intersection of k-US(o)
and C(1,o). Each element tAvJ ofk-US(o) has the property that
an arbitrary entire function of exponential type <ko is com-
ple~e~y deterEUined by its values and those of its derivatives of
165
order :£k-1 at the points \) but the elements of the subset

k-US(1,a) are amonst the sparsest. In many situations, the
function can be represented by an interpolation series in which
(k-1 )
the numbers f(Av), ... ,f (AV) appear as coefficients, i.e.
00 k-1
f(z) = L L f(h) (A)Lvh(Z)
V=-OO h=o
where
= 0 for all v,].l,j,h
except for V=].l and h=j; in the latter case
In such a situation, the quadrature formula

00
00 k-1
(9 )
f
_00
f(x)dx L L
v=-oo h=o
w f(h)(A)
vh v
where 00
wvh : =
f_00Lvh(x)dx (-00<v<00,0~h~k-1)
is valid for all entire functions of exponential type <ka pro-

vided f E L 1 (_00,00) and term by term integration is justified.
Formula (9) is to be compared with (2).
In analogy with Turan's result mentioned in Section 2,
Olivier and Rahman [6] proved the following quadrature formula.
Theorem C. Let k be an odd positive integer and a > o. Further,

let ao,o =1 whereas for 0 ~ j ~ k-1 let a j ,k-1 =- j\ 1jJ (j) (0)
where 1jJ(z):=
(k-1)/2
r-r1
j =
(1 + ~2).
J
2
If f is an entire function of
exponential type (k+1)a belonging to L 1 (-00,00), then

166
00
k-1
J f(x)dx
1T
a j~O (2a)j a j ,k-1
-00
j even
The example (sin(aZ) )k(sin~Ez/2))2 shows that such a formula

cannot hold for all entire functions of order 1 type ka+E for
any E > 0 if k happens to be even.
At this stage it is natural to ask the following
question.
Question 1. Are there other elements {AV} of k-US(1,a) such

that formula (9) remains valid for all entire functions of expo-
nenti~ltype (k+1)a belonging to L1 (_oo,oo)? Here it is understood
that k is an odd positive integer.
In order to give an answer to Question 1 we need to introduce

two definitions.An element A = {AV} of k-US(1,a) will be said
to have the "Turan property" if forumula (9) remains valid for
all entire functions of exponential type (k+1)a belonging to
L 1 (_oo,00). We shall say that A has the "modified Lk+1 property"
if there exists a function fA in JA such that
00
-2 k+1
J x fA (x) dx < 00. We are now ready to state
_00
Theorem 3. For each odd positive integer k, the sequence {v1T/a}

is the only element of k-US(1,a) which has the Turan property
as well as the modified Lk+1 property.
This result which provides an answer to Question 1 will be proved

elsewhere.
Remark 2. From Theorem C it can be deduced that for each odd

positive integer k, the function (z) := 1 sin(az) (whoses
a a
zeros solve the "problem of Turan" for entire functions of expo-
nential type) minimizes the integral
167
00
J x- 2 If(x) I k + 1dx over all entire functions of exponential type a

-00
such that f(o) 0, f' (0) 1. For comparison we recall that the
n
polynomial TI k:= r-T(x-x Vk ) whose zeros solve the correspon-
n, v=1
1
ding problem for polynomials minimizes the integral JIP(X) Ik + 1dx
-1
over all monic polynomials of degree n.
References.
1. Boas, R.P., Jr. (1954) Entire functions (Academic Press, New
York) .
2. Boas, R.P., Jr. (1972) Summation formulas and bandlimited
signals. T8hoku Math. J. 24, 121-125.
3. Carlson, F. (1914) Sur une classe de series de Taylor
(Thesis, Uppsala).
4. Frappier, C. and Q.I. Rahman (1986) Une formule de quadrature
pour les fonctions entieres de type exponentiel. Les Annales
des Sciences Mathematiques du Quebec 12, 17-26.
5. Hermite, Ch. (1878) Sur la formule d'interpolation de Lagrange.
J. reine angew. Math. 84, 70-79.
6. Olivier, P. and Q.I. Rahman (1986) Sur une formule de quadra-
ture pour des fonctions entieres. Modelisation mathematique
et Analyse numerique 20, 517-537.
7. Szego, G. (1967) Orthogonal polynomials, 3rd edn (Amer. Math.
Soc., Providence, Rhode Island).
8. Titchmarsh, E.C. (1948) Introduction to the theory of Fourier
integrals, 2nd edn (Oxford University Press, London).
9. Tur~n, P. (1950) On the theory of the mechanical quadrature.
Acta Scient. Math. ~, 30-37.
10.Valiron, G. (1925) Sur la formule d'interpolation de Lagrange.
Bull. Sci. Math. (2) ~, 181-192, 203-224.
168
Qazi Ibadur Rahman, Departement de Mathematiques etdeStatistique,

Universite de Montreal, P.Q. H3C 3J7 CANADA.
Gerhard Schmeisser, Mathematisches Institut der Universitat

Erlangen - NUrnberg, Bismarckstr. 1 1/2, 0-8520 Erlangen,
FEDERAL REPUBLIC OF GERMANY.
MARKOFF TYPE INEQUALITIES FOR CURVED MAJORANTS
Q.I. Rahman and G. Schmeisser
Dept. de Mathematiques et de Statistique, Universite de Montreal

and Mathematisches Institut der Universitat Erlangen-Nlirnberg.
1. Introduction
Let ITn be the set of all polynomials of degree at most n with

complex coefficients. As usual denote by
T (x) := cos(m arc cos x)

m
and
u (x) := (1-x 2 )-1/2 sin «m+1)arc cos x)
m
the Tschebyscheff polynomials of degree m of the first and se-

cond kind, respectively. Furthermore, let us write II ·11 for the
maximum norm on the unit interval, i.e.
II p II : = max Ip (x) I •
-Hx:;;1
Extending the well-known Markoff inequality of his brother Andrei

to higher order derivatives W.A. Markoff [9] proved the following
Theorem A. If P E ITn such that
(1) II p II :; 1,
170
then
(2) IIp(j) II :;;T(j) (1) = n 2 (n 2 _12)(n 2_2 2 ) ••• (n 2 _(j_1)2)

n 1.2.3 ••••• (2j-1)
(j = 1,2, ••• ).
In this connection it is of interest to recall an in-

equality of S. Bernstein [14] which says that under hypothesis
(1)
(3) I p' (x) I :;; n(1-x 2 ) -1/2 for -1 < x < 1.
Hence at a fixed point inside the unit interval Ip(j) (x) I may
be much smaller than the right hand side of (2).
At the conference on Constructive Function Theory held
in Varna, Bulgaria in the year 1970, P. Turan asked a problem
which in a generalized form may be stated as follows.
Problem 1. Let ~ be a non-negative function on [-1,1]. Given a

polynomial p E II such that
n
(4) Ip(x)I:;;~(x) for -1:;;x:;;1,
find a bound for II p (j) II •
Problems of this type also occured in approximation

theory, notably in the work of Dzyadyk [5]. Note that for a
majorant ~ positive in the open unit interval, (4) may be written
as Ilpll :;; 1 with the weighted sup-norm
~
1
Ilpll~:= sup ~(x) Ip(x) I·
-1<x<1
The most interesting cases are those where ~ vanishes at ± 1 since

in view of (2) and (3) it can be expected that the behaviour of
pIx) near the endpoints of the unit interval has a strong influ-
171
ence on the size of the desired bound. In fact, Turan originally

asked his problem for majorants like
(5) q>(x) (circular majorant),
2
(6) q>(x) := 1-x (parabolic majorant),
etc.
Problem 1 has been extensively studied by Rahman and
Pierre & Rahman in a series of papers [13;11;12]. One of their
most general results may be stated as follows.
Theorem B. Let
q>(x) := (1_X)A/2(1+x)\1/2
where A and \1 are non-negative integers. Define
>../2 \1/2
S(1-X) (1+x) Tm-(A+\1)/2(x) if A,\1 are both even
,
l(1_X) (A+1)/2(1+x) (\1+1)/2 u (x) i f A,\1 are both
m-1-(A+\1)/2
odd.
Then for every polynomial pEn satisfying (4)
n
(7) IIp(j) 1/ ~max{IIA~j) 1/ ,1/ A~~~ 1/ }
provided that
(8) j ~ A?
In the case of the majorant q>(x) =1 in (1) we have

A = \1 = 0 and hence (8) imposes no restriction on (7); in fact,
one can show that Theorem B reduces to Theorem A. In the case of
the circular majorant (5) we have A = \1 = 1 and again (8) imposes
no restriction on (7). However, in the case of the parabolic majo-
rant (6) we have to choose A = \1 = 2 which requires j ~ 2. Hence
172
we do not get a bound for IIp'll. A result which fills this gap
will be given in Section 2.
In the case of the majorant (5), Turan in 1975 also
asked the following
Problem 2. Given a.polynomial of degree at most n with real co-

efficients such that
0;;; p(x) ;;; \p(x) for -1;;;x;;;1,
find a bound for " p'" .
In the case of the majorant \P (x) '" 1, Problem 2 is

easily reduced to Problem 1. In fact, if 0;;; p (x) ;;; 1 for
-1 ;;; x;;; 1, then q (x) : = 2p (x) -1 satisfies the hypothesis (1).
Hence "p''';;; n 2 /2 by Theorem A. Equality is obviously attained
for
T (x)+1
n
p(x)
2
For the circular majorant (5), however, Problem 2 seems to be very

much different from Problem 1. We shall discuss it in some detail
in Section 3.
Finally we recall the following inprovement of Theorem A
due to Duffin & Schaeffer [4].
Theorem C. Let p E TIn such that
Ip(cos V'TT/n) I ;;; for v o,1, ••• ,n.
Then (2) holds.
A new short and elegant proof was recently given in

[3]. Since Theorem C has applications in the theory of interpo-
lation and numerical differentiation [1i8] it is natural to ask,
if an analogous improvement is also possible in the case of cur-
ved majorants.We shall turn to this question in Section 4.
173
2. Problem 1 for parabolic majorant
We start with the following
Proposi tion 1. For n f: 2 let pEIl such that

n
2
(9) Ip(x)I:>1-x for -1:> x :> 1.
Then
Proof. Let
w(z) := e i(n-2)z s~n

. 2z f ( z ) : = p (cos z).
Then wand f are both entire functions of exponential type nand
( 11) I f (x) I :> I w(x) I for x E lR •
Moreover, w has real zeros only and its Phragmen-Lindelof indica-

tor hw satisfies
n <: - (n-4) = h (n/2).

w
In this situation it follows from Theorem 11.7.2 in [2] that (11)

extends to the first derivatives so that
I f' (x) I :> I w' (x) I for x E lR •
This yields
p'( cos x) I ~ li(n-2)sin x + 2cos xl
which is (10).
174
Remark 1. The polynomial p*(x) := (1-x 2 )T n _ 2 (X) satisfies the

hypothesis of Proposition 1 and for odd n ~ 3
Ip~(o) I = IT~_2(0) I = n-2.
Remark 2. It follows from [12, Theorem 1'] that under the hypo-
thesis of Proposition 1
Ip' (0) I ~ n-3 for even n::: 6.
Concerning Problem 1 for parabolic majorant we can draw

from Proposition 1 and the remarks the following conclusions:
(i) For all n ~ 4 we have II p' II ~ n-2.
(ii) This estimate is best possible for all odd n ~ 5. Equality
can only be attained i f II p' II = I p' (0) I.
(iii) For all even n ~ 6 we have II p' II < n-2.
Finally we find by elementary calculation:
(iv) For n E {2,3,4} the best possible inequality II p' II ~ 2
holds. For n E {2, 3} equality can only be attained if
IIp'll = Ip'(±1)1. For n=4 the example p(x):=(1-x 2 )(2x 2 -1l
shows that equality can be attained with II p'll =Ip' (x) I at
all xE{±1,±1/2}.
There is still something to be done. So far we do not
know the best possible bound in the situation under (iii). Investi-
gating the polynomial p*(x) introduced in Remark 1 near the
origin we find that
2
Ilp~ll~n-2-;n + O(n- 2 ) as n+ oo •
Is the right hand side also an upper bound for even n? Using
techniques developed by W.A. Markoff in [9] and extended by
Voronovskaja [16] and Gusev [7] we can show that the answer is
indeed "yes". The proof being rather long will be given elsewhere.
Here we content ourselves with stating the complete result.
175
Theorem 1. Let n <: 2 and p E IIn such that
for -1;:;; x ;:; 1.

Then
The inequality is sharp for n=2,4 and all odd n. For even n the
inequality
1T2 -2
IIp'll:£n-2- an +O(n)
is an asymptotically sharp estimate.
3. Problem 2 for circular majorant.
We start with two auxiliary results.
Lemma 1. Let
n-2 n
R.(x) := (x 2 -1)T n _ 1 (X) 2 n(x-x)
v=o v
and
R. v (x) : = R.(x)
x-xv
where -1, xn = 1 and
(2v-1 1T \
v = 1,2, ••• ,n-1.
Xv := cos\n_1 ·2)
Then R.' (x) ~o for t;, ;:;;x;:;; 1 and v = 0,1, •.• ,n where
v n
1T
t;,n := cos 3(n-1)
Proof. Denoting the largest zero of R.~ by nn we clearly have

176
R,' (x) ~ 0 for all x ~ n • Further, if x E ] n ,1 [, then R, (x) < 0

n n n n
since all the zeros of R,n except 1 lie to the left of nn.
Since
(x+1)R,n(x)
R, (x)
v x-xv
we conclude that
(x+1) R,~ (x) 1+xV
R,~(x) = -~-"2 R,n (x) ?: 0
(x-xv)
for all x E [n , 1] and v=o,1, ••. ,n-1 as well. It is now enough

n
to prove that I',;n ~ nn· This is obviously done by verifying that
R,' (I',; ) ~ o. Since
n n
R,~(cos 6) = cos(n-1)6 - (n-1) (tan ~)Sin(n-1)6
the proof is completed by a simple trigonometric calculation.
Lemma 2 [13, Theorem 2]. If P E II such that

n
for -1 < x < 1,

then
X 2 )1/2
ip'(x)i ;;;((n-1)2 + 2 for -1 < x < 1.
1-x
We are now prepared to prove the following
Theorem 2. If P E lIn such that
2 1/2
o;;;p(x);;;(1-x)
(12 ) for -1;;; x ;;; 1 ,
then n-1 . -2(2v-1 1T\
( 13) IIplll~n~1 L Sl.n \n-1 • '4)
v= 1
v odd
Proof. Using the notation of Lemma 1 we may represent p(x) by

the interpolation formula of Lagrange as
177
p(x)
and obtain by differentiation
p' (x) =
Since
v . (2V-1 1T\
£,1 (xv) = (-1) (n-1) s~n \ n-1 2' }
and
0:;; p(x ) :;; .
s~n (2V-1
-_.- 1T )
v n-1 2
we conclude with the help of Lemma that
n-1
I p' (x) I :; n~1 2 £,~ (x)
v=1
v odd
Clearly £,1 (x) increases with x on the interval in question,
v
i.e.
. -2 ( 2v-1 1T )
£'~(x) :;; £,~(1) = s~n \ n-1 .4'
and so
n-1
. -2 (2V-1 ~\ =: N
Ip' (x) I :;; n~1 2
v=1
s~n \n-1 4) n
v odd
for and by symmetry for -1:;; x :;; -~n as well.
For we have by Lemma 2
~2 1/2
I p' (x) I :; ( (n-1)2 + ~)
1-~n
1/2
=(n-1)2 + cot2(3(n~1)))
(1 + .-2..)
2
1 / 2 (n-1) <.!.§. (n-1) < N
2 n·
1T 1T
This proves the inequality in (13).

178
Remark 3. We note that
Nn = (1.74245 •.• )n + 0(1) as n-+ oo •
It is seen as follows. Using the trapezoidal rule on [a,b] where
1T 2+(-1)n
a := 4n , b
4n
we find b
1T I (4V+1
00
JCi>X
a
1TN
n n v=o 4n
.1T)-2 + 0(1)
as n -+ 00. It can be shown [6] that the infinite series on the

right hand side is equal to
n 2 (1 + ~ G)
1T2
where G 0.91596 ••. is Catalan's constant. Hence
(cot x -~) 1T 1TN - n1T(1 + ~G) + 0(1)

n 1T2
x=4n
or equivalently
Nn (1 + 8G)n + 0(1) = (1.74245 ••• )n + 0(1).

1T2
It is natural to ask for the precision of the estimate

(13). An answer is given in the following
Remark 4. Let Pm(X) be the Legendre polynomial of degree m

with the normalization P (1) = 1. Then [15, p. 165, (7.3.8)]
m
Now, if n is even and m:= (n-2)/2, then

179
71 n-2 2 2
p*(x) :="2 . -2- (l-x )Pm(x)
satisfies the hypothesis of Theorem 2 and
Hence the right hand side of (13) cannot be replaced by anything

smaller than (71/2 + o(l))n. We are inclined to believe that the
best possible bound in (13) is of the form
where as n -+ 00 •
4. A Duffin & Schaeffer type result
Let p E II such that

n
(14 ) Ip(x) I ~ (1-x 2 ) 1/2 for -1 ~ x ~ 1.
As a consequence of Theorem B it was shown in [12] that
( ') dj 2
(15 ) II p J II 5; - , (x -1) Un - 2 (x) I
- dx J x=l
-L
2j-l
in particular IIp'll ~ 2(n-l). Inequality (15) is obviously best

2
possible since p(x) = (x -1)U n _ 2 (x) satisfies the conditions
imposed on p(x). In view of Theorem C it is natural to ask if
(15) is already true under the weaker hypothesis that (14) holds
only at n+1 appropriately chosen pOints in [-1,1]. The answer
is very surprising. It is "no" for the first derivative and "yes"
for all higher order derivatives. In fact, the following theorems
hold whose proofs will be given elsewhere.
180
Theorem 3. Howsoever we choose an infinite triangular matrix
(1 )
x1
(n) •
x1 ••••
of nodes with
-1 < (n) < x(n) < ••• < x (n) <=

= xo 1 n
there always exists a sequence of polynomials p (x) and a se-

n
quence of points T E [-1,1] such that p (x) is of degree at
n n
most n
for v o,1, ..• ,n

and
I p' n (T n ) I :; ~7T (1 + 0 (1) ) n log n
as n tends to infinity.
On the other hand we have
Theorem 4. Let
2v-1 7T
cos(n:::-:!. 2)' v= 1,2, •.• ,n-1.
If P E TIn such that
for v o,1, ••. ,n,

181
then
( .) dj 2
lip J II ~ (x -1)U n _ 2 (x) Ix=1 for j 2,3, •••
and dx j
2
II pIli ~ (n-1) (1i'log (n-1) + 3)
2
= 1i' (1 + 0 ( 1) ) n log n
as n -+ 00
5. Acknowledgement
Theorem 2 was first proved in [10, Theoreme 7] but the proof

presented here offers certain simplifications. The example men-
tioned in Remark 4 is also contained in [10, p. 76].
182
References
1. Berman, D.L. (1952) Solution of an extremal problem of the

theory of interpolation. Dokl. Akad. Nauk SSSR 87, 167-170
(in Russian). -
2. Boas, R.P.,Jr. (1954) Entire functions (Academic Press, New

York) •
3. Duffin, R.J. and L.A. Karlovitz (1985) The Markoff-Duffin-

Schaeffer inequalities abstracted. Proc. Natl. Acad. Sci.
USA 82, 955-957.
4. Duffin, R.J. and A.C. Schaeffer (1941) A refinement of an

inequality of the brothers Markoff. Trans. Amer. Soc. 50,
517-528.
5. Dzyadyk, V.K. (1956) On a constructive characteristic of

functions satisfying the Lipschitz condition Cl (0 < Cl < 1)
on a finite segment of the real axis. Izv. Akad. Nauk SSSR
(seriya mat.) 20, 623-642.
6. Gradshteyn, I.S. and I.M. Ryzhik (1980) Tables of integrals,

series and products (Academic Press, New York) •
7. Gusev, V.A. (1961) Derivative functionals of algebraic

polynomials and W.A. Markov's theorem. Izv. Akad. Nauk SSSR
Ser. Mat. 25, 367-384 (Russian); English trans. Appendix:
Transl. Math. Mongraphs, vol. 28 (Amer. Math. Soc., Providen-
ce, R.1., 1970).
8. Haverkamp, R. (1980) Approximationsfehler der Ableitungen

von Interpolationspolynomen. J. Approximation Theory 30,
180-196.
9. Markoff, W.A. (1916) tiber Polynome, die in einem gegebenen

Intervalle moglichst wenig von Null abweichen. Math. Ann 22,
218-258.
10. Pierre, R. (1977) Sur les polynomes dont Ie graphe sur [-1,1]
se trouve dans une region convexe donnee (Th~sis, Universite
de Montreal).
11. Pierre, R. and. Q.I. Rahman (1976) On a problem of Turan

about polynomials. Proc. Amer. Math. Soc. ~, 231-238.
12. Pierre, R. and Q.I. Rahman (1981) On a problem of Turan

about polynomials II. Can. J. Math. ~, 701-733.
13. Rahman, Q.I. (1972) On a problem of Turan about polynomials

with curved majorants. Trans. Amer. Math. Soc. 163, 447-455.
183
14. Rahman, Q.I. and G. Schmeisser (1983) Les inegalites de

Markoff et de Bernstein (Les Presses de l'Universite de
Montreal, Montreal).
15. Szeg8, G. (1967) Orthogonal polynomials (Third edition, Amer.
Math. Soc., Providence, Rhode Island).
16. Voronovskaja, E.V. (1970) The functional method and its appli-
cations. Trans. Math. Monographs, vol. 28 (Amer. Math. Soc.,
Providence, R.I.).
Qazi Ibadur Rahman, Departement de Mathematiques et de Statistique,

Universite de Montreal, P.Q. H3C 3J7 CANADA.
Gerhard Schmeisser, Mathematisches Institut der Universitat
Erlangen - Nurnberg, Bismarckstr. 1 1/2, 0-8520 Erlangen,
FEDERAL REPUBLIC OF GERMANY.
INTERPOLATION MIT SPHARISCHEN HARMONISCHEN FUNKTIONEN
M. Reimer
FB Mathematik, Universitat Dortmund
In [3], [4] haben wir gezeigt, wie Interpolationsnormen bei Vor-

liegen einer orthogonalen Struktur mit Hilfe des kleinsten Eigen-
wertes einer positiv-definiten symmetrischen Systemmatrix abge-
schatzt werden k6nnen. Das bedeutet, daB die gerade in hochdi-
mensionalen Raumen problematische Berechnung des Maximums der
Lebesgue-Funktion durch die sehr viel einfachere Bestimmung eines
kleinsten Eigenwertes ersetzt werden kann. Das Verfahren wurde
von uns an einer Reihe numerischer Beispiele erprobt. Offen ge-
blieben ist die Frage, ob die Ergebnisse nicht zu pessimistisch
sind. Wir konnen sie dadurch beantworten, daB wir sie ftir einen
vollen InterpolationsprozeB positiv beantworten. Dieser ProzeB
wird in Raumen spharischer harmonischer Funktionen definiert,
auf deren Schltisselrolle bei der Behandlung von Polynomraumen
tiber der Sphare wir zunachst hinweisen wollen.
Wir betrachten bei festem r E IN ,{ 1} die Raume lP r (;{r), \1 E IN ,

\1 \1 0
aller Restriktionen von Polynomen (homogenen harmonischen Poly-
nomen) des Grades \1 in r Variablen auf die Sphare sr-1:=
{x E JRr II x I = 1 }. In ihnen wird durch das Oberflachenintegral ein

inneres Produkt induziert, beztiglich dessen die Raume ;{~ ortho-
gonal zueinander stehen. Da die harmonische Fortsetzung eines
Elements aus lP rwieder aus lP r ist, gilt sogar die orthogonale
\1 \1
185
Dekomposition
lI? r\1 = lH\1

*r
E&
lH*r '" *r
\1-1 '" ••. Ell lHo ' (1)
wobei wir anmerken, daB wir fUr die praktische DurchfUhrung die-
ser Dekomposition in [2] ein sehr einfaches Verfahren angegeben
haben. In Verallgemeinerung von (1) kann man auch die Unterraume
der Gestalt
\1 *r
lI? Ell v lH v
E (2)
v=o
betrachten, wobei EV = 1 oder 0 anzeigt, daB lli~ in der direkten
Summe vorkommt oder nicht vorkommt. Jeder Raum (2) ist rotations-
invariant. Umgekehrt laBt sich aber auch jeder rotationsinvari-
ante Unterraum lI? von lI?r in der Form (2) darstellen. Das ergibt
\1
sich daraus, daB sein reproduzierender Kern ein zonales Poly-
nom P(xy) mit PElI?1 ist. So ist z. B. der reproduzierende Kern
von ~~ von der Ge:talt P v (xy) mit P v = const • c~-2/2,
const +
0 . Stellt man aber P in der Form
dar, so ergibt sich auf Grund der Reproduktionseigenschaften der

Kerne zunachst die Beziehung
P(xy) = { P(xt)P(ty)dt
It =1
\l \l
= L L cvc\ { Pv(xt)P\(ty)dt.
v=o \=0 It =1
*r ergibt sich hieraus

Wegen der Orthogonalitat der lHv
\1 \1 2
\ C P = P = \ P
L. vv L C vv'
v=o V=O
woraus C v E{0,1} und schlieBlich auch (2) folgt.
186
Wegen dieser SchlUsselrolle der ijr fUr den Aufbau der rotations-
~
invarianten unterraume von F r solI die Interpolationsaufgaoe
zunachst in E~ untersucht w:rden. Die Dimension dieses Raumes
ist N = 2~+1. Wahlt man das Knotensystem
~
jeweils geeignet (z. B. "extremal"), so ist die Norm des entspre-

chenden Interpolationsoperators l : c (S 2 ) .... ij3 j edenfalls in der
~ ~
Form
1
c ~2 < II ll-lll, ~ 2~+1 , (3)
c > 0, beschrankt ([1],[4]), wobei die untere Schranke fUr be-

liebige projektionen gilt. Eine Verscharfung dieses Ergebnisses
gelang B. SUNDERMANN [6]. Er zeigte, daB aquidistant auf einem
Breitenkreis nahe dem Aquator verteilte Knoten zu einer Inter-
polati~nsnorm
fUhren. In einer gemeinsamen Arbeit [5] haben wir nun versucht,

dieses Ergebnis mit Hilfe der Eigenwertmethode zu erzielen. Dies
gelang, da die Systemmatrix (P~(tjtk» sich in diesem Fall als
zyklisch erweist, so daB ihre Eigenwerte im wesentlichen durch
die Koeffizienten der Tschebyscheff-Fourier-Entwicklung des Po-
lynoms P (a+(1-a)x), a = sin 2 £, £ geographische Breite, gegeben
~
sind. Diese Koeffizienten erfUllen bezUglich des Parameters a
eine bemerkenswerte Differentialgleichung, die eine gUnstige
Selektion des Parameters a = a gestattet, so daB der kleinste
~
Eigenwert A(a ) besonders groB wird. Dies fUhrt zu der Abschat-
~
zung
187
mit --4-
7T 3/4
<0.85.
Damit ist das Ergebnis von SUndermann mit unserer Eigenwertme-

thode wiedergewonnen, wobei sogar die O-Konstante konkretisiert
werden konnte.
Literatur
1. Daugavet, I. K.: Some applications of the Marcienkiewicz-

Berman identity. Vestnik Leningrad Univ., Math. 1, 321 - 327
(1974) .
2. Reimer, M.: Best approximation to polynomials in the mean

and norms of coefficient-functionals. In: Multivariate Appro-
ximation Theory (W. Schempp and K. Zeller, ed.) ISNM~,
Birkhauser Verlag, Basel 1979.
3. Interpolation on the sphere and bounds for the La-

grangian square sums. Erscheint in Resultate der Mathematik.
4. - , - und Stindermann, B.: A Remez-type algorithm for the cal-

culation of extremal fundamental systems for polynomial spa-
ces on the sphere. Computing 37, 43-58 (1986).
5. - , - und - , - : Gtinstige Knoten fUr die Interpolation mit

homogenen harmonischen Polynomen. Erscheint in Resultate der
Mathematik.
6. SUndermann, B.: Projektionen auf Polynomraume in mehreren

Veranderlichen. Diss. Dortmund 1983.
Prof. Dr. M. Reimer

Fachbereich Mathematik
Universitat Dortmund
Postfach 50 05 00
D-4600 Dortmund 50
Bundesrepublik Deutschland
CONVERGENCE THEOREMS FOR

NONLINEAR APPROXIMATION ALGORITHMS
R. Schaback
Institut fur Numerische und Angewandte Mathematik

der Universitiit Gottingen
Gottingen, FRG
The author gratefully dedicates this paper to the memory of his academic
teacher, Prof. Dr. Helmut Werner.
Abstract. For a large class of algorithms, including trust-region-, projection-, line-

search-, and Levenberg-Marquardt methods, a quantitative global convergence proof based
on a general convergence theory is given.
1 Introduction
We use the notations and conventions of M.J.D. POWELL's survey paper [8] and
summarize them here as follows:
Let
h : Rm ~ R, with h(x) ~ 00 for IIxll ~ 00
be a convex and continous substitute for 11.11 or 11.11 2 , bounded from below. We try to
minimize F(x) := h(J(x)) for a function
f:Rn~Rm, m~n,fECl,
but in practice we minimize the normalized linearization gain

'I/J(x) := F(x) - inf h(f(x} + yo f(x)s), x ERn (1)
11·II:;;l,.ER"
and are content with stationary points, i.e. zeros of 'I/J. The analysis can be restricted to
a domain H c Rm where
(2)
189
holds. Furthermore, the inequality
IF(x + s) - h(J(x) + V f(x)s)11 S Lllsllw(s) (3)
with the modulus w(.) of continuity of V f is frequently used (see section 2 of [8]).
The algorithms we consider here will simply be defined as mappings
leading to the iteration

(4)
A function cJ> : R n x Rn - t R will be needed to evaluate the progress of the algorithms.
Standard choices will be (15) or (10), corresponding to POWELL's two convergence the-
orems in [8], but in principle the convergence theory does not depend on cJ>.
2 General Convergence Theory

The following presentation of general convergence theorems combines POWELL's two
approaches (for line-search- and trust-region methods) into one, extends them (e.g. to
projection methods like DEUFLHARD's [3] and to linear convergence results) and includes
the quantitative global convergence theorem of SCHABACK [9]. We keep the treatment
fairly general, because we want to apply the results under different circumstances in a
forthcoming paper.
The first axiom (see [8] and [9], for instance) concerns the descent of the error:
Axiom 2.1 For all x ERn,
F(x + s(x)) S (1 - l)F(x) + lcJ>(X, s(x)), 1 E (0,1). (5)
The second axiom bounds F(x) - cJ>(x,s(x)) from below with respect to 1jJ(x) (see [8]):
Axiom 2.2 For all [, > 0 there exists some f > 0 such that 1jJ(x) :;,. [, implies
F(x) - cJ>(X,8(X)) :;,. Eo
Theorem 2.1 Fur the iteratiun (4) of a method satisfying Axioms 2.1 and 2.2,
(6)
Proof: \\'henever 1t'l(xd :;,. [" F(Xk) - F(Xk+d :;,. (, but hand F are bounded from below .
•
This theorem combines Powell's theorems 1 and 2 from :8], while the next one incorporates
a quantitative result of SCHABACK [9 1 based on the following
190
Axiom 2.3 For all x E Rn
F(x) - ~(x,s(x)) ~ I\:g(TjJ(x)) (7)

for some I\: > 0 and some nonnegative continous function g : R -> R vanishing only at
zero.
Theorem 2.2 If Axioms 2.1 and 2.9 hold, then
(8)
Proof: By summation of
We shall apply this theorem in the cases g(x) = x2 and X4.
For linear convergence we use

Axiom 2.4
F(x) - ~(x, s(x)) ~ I\:(F(x) - F(x")), I\: >0 (9)
for all x from a neighbourhood U of a critical point x' of F.
This means that the algorithm achieves at least a fixed fraction of the total possible local
error decrease.
Theorem 2.3 Assume that {Xk} has a limit point x·. Then axioms 2.1 and 2.4 imply a
linear convergence step
whenever Xk E U.
Proof: Obvious from (5) and (9). See also [9] .•
3 A Class of Algorithms
We consider
1
~(x,s) := h(f(x) + Vf(x)s) + -sTB(x)s (10)
2
with a symmetric n x n matrix B(x) and
(11)
with certain functions J.t(x), M(x): R -> R, M(x) 2: O. The algorithms are assumed
to minimize ~(x,s) with respect to s over some domain S(x) c Rn having the following
properties:
191
1. S(x) = P(x)S(x) for a linear projector P(x) satisfying
IIVf(x)(s - P(x)s)11 -::: u(x)llsll, s E Rn
for some function u(x) ~ O. This is done to cover methods like DEUFLHARD's (see
e.g [3]). If P(x) is not needed, u(x) = 0, but it may improve the numerical behaviour
of algorithms to project on "relevant" eigenspaces of V fT(X)V f(x) in case of rank
loss of the Jacobian.
2. S(x) contains a ball
Kr(:r:) := {s E R n Illsll -::: r(x)} , r(x) E (0,00]
3. S(x) is contained in KR(z),R(x) E [r(x),oo].
The solution s(x) of the minimization should exist (but must not necessarily be unique)
and lie in Kp(z):
iis(x)li ~ p(x), 0 ~ p(x) ~ R(x). (12)
We consider R(x) as a controllable a-priori bound on the solution in the sense of trust-
region methods, while p(x) is an a-posteriori bound on the actual solution. For regularized
line-search- and Levenberg-Marquardt methods, R(x) = 00, but there are bounds p(x)
on IIs(x)ll. For instance, Levenberg-Marquardt methods have
B(x) = M(x)· Id,S(x) = Rn,IL(X) = -M(x),r(x) = R(x) = 00,

and
IIs(x)1I2 -::: 2~~;?1 for M(x) > 0 (13)
Ils(x)112 -::: 2~i;jl for M(x) > 21F(x)l. (14)
To cover line-search methods, we finally take
s(x) := t(x)s(x)
with a stepsize t(x) E (0,1] to make the definition of our algorithm class complete.
This includes
• line search methods
• projection methods
• Levenberg-Marquardt methods
• trust region methods

192
together with a fairly free choice of second order terms B(x). It will allow a theoretical
comparison, because we shall give general conditions for convergence theorems like those
in the preceding section. Note that in contrast to POWELL's approach we consider only
"genuine" trust-region steps and regard adjustment of control parameters as an "inner"
problem for each iteration.
The parameters It(x), M(x), R(x), t(x), and u(x) are those the algorithm really uses at the
step x f-> x+s(x). However, any realistic implementation does not fix them beforehand, but
rather updates them according to certain strategies. We assume that an implementation
starts with a-priori parameters
Ito(x) E R, Mo(x) ~ 0, uo(x) ~ 0,
Ro(x) ~ TU(X) > O,to(x) ~ t > 0,

and uses certain numerical tests to accept or modify these values. Modifications will always
be by multiplication or division by a constant 1/ E (0,1) until the test is passed. We leave
the treatment of more sophisticated update formulae to the reader.
Furthermore, we restrict the control of parameters mainly to R(x), t(x), and u(x). Con-
trolling It(x) or M(x) is a common strategy only for Levenberg-Marquardt methods with
positive definite B(x) and implies a-posteriori bounds on s(x) like (13) or (14). We con-
sider such cases as indirect controls of R(x) and give them a separate treatment.
4 Basic Lemmas
The convergence theory of section 2 can be applied for ~(x, s) as in (10) or ~(x, s)
equal to
tp(x,s) = h(f(x) + "\7f(x)s) (15)
(see POWELL's two approaches in [8]). We use (15) just as an abbreviation and (10) for
both minimization and convergence analysis.
We start with some vector s(x) E R n giving a positive linearization gain
0< f(X,S(X)) 'S F(x) - :p(x,s(x)). (16)
The following lemmas can be applied for different choices of s(x):
Lemma 4.1 If
21Is(x)IILu(x) 'S f(X,S(X)), (17)
then
1
0< E(x,P(x)s(x)) := 2f{x,s(x)) 'S F(x) - tp (x, P(x)s(x)). (18)
193
Proof:
F(x) - <p(x, P(x)s(x))
> F(x) - h(f(x) + V!(x)s(x)) - LIIV!(x)(s(x) - P(x)s(x))11
> f(X,S(X)) - La(x)lls(x)11
> ~f(X,S(X)).
•
Lemma 4.2 If (16) and (17) hold, then for
. ( r(x) f(X, S(X)) ) (19)

r(x) := mm 1, Ils(x)II' 2M(x)lls(x)112
we have
F(x) -- 4>(x, s(x)) 2: ~ l(X, s(x))r(x). (20)
If
(1 - t(x))Il(X)p(x)2 ~ t(x) (F(x) - 4>(x,s(x))), (21)
then
F(x) - 4>(x, s(x)) 2: ~t(x) (F(x)- 4>(x, s(x))) 2: ~r(X)t(x)f(X' s(x)). (22)
Proof: We omit the argument x, where possible, and get
4>(x, s) ~ 4>(x, r Ps)
h(f(x) + V!(x)rPs) + ~r2(Psf BPs
< F(x) - n(x, Ps) + lr 2 MIIsl12
< F(x) - In(x, s) + ~n(x, s),
because r Ps is admissible. Furthermore
4>(x,s) = 4>(x,ts)
= h(f(x) + V!(x)ts) + ~t2sT Bs
< (1 - t)F(x) + t4>(x, s) - ~t(l - t)sT Bs
< F(x) - t(F(x) - 4>(x,s)) + ~(1- t)IlP2
< F(x) - ~t(F{x) - 4>(x,s)).
•
Lemma 4.3 The inequality
Lt(x)p{x)w(t(x)p(x)) + ~1l(x)t2{x)l(x) ~ (1 -,) (F(x) - 4>(x,s(x))) (23)
implies (5).
Proof:
(1 - ,)F{x) + ,4>(x,s)-- h(f(:r + s))
> (1 - ,)(F(x) - 4>(x, s)) + ~sT B.~ - Llis!lw(llsll)
> (1 -,){F(x) - 4>(x,s)) - !IlP 2t 2 - Ltpw(tp)
> o.
•
194
5 Global convergence
We now prove that any algorithm in our class will converge in the sense of Theorems
2.1 and 2.2, if parameters are controlled appropriately.
The analysis links F(x) - ~(x,s(x)) to 1,b(x) by using the vector s(x) maximizing 1,b(x), i.e.
1,b(X) = F(x) - h(J(x) + VJ(x)s(x)), Ils(x)11 = 1.

We take sJ(x) = s(x), S2(X) = P(x)sJ(x) and use Lemma 4.1 to get
1 1
E(X,S2(X)) ;::: zt(X,Sl(X)) = z1,b(x) (24)
for bounds
t(X, s;(x)) <:: F(x) - h(J(:r) + V J(:r)s;(x)),
if
2La(:r) < w(x) (25)
holds. Though L is unknown, a good choice of ao(x) would be proportional to 1jJ(x) for
small 1,b(x). Using (24) as a test and multiplying a(x) by !I in case of failure will give
a finite inner iteration that finally satisfies (24) and makes sure that the actually used
parameter is bounded by
a(x) ;::: min (ao(x), !I~i)) .
Now we apply Lemma 4.2 for S2(X) = P(x)s(x) and use Ils2(X)11 <:: 1 with
T(X) ;::: TJ(X) := mm

. ( 1,r(x), 2M
1,b(x)
(x) ) (26)
to get
_ 1,b(x)
F(x) - ~(x,s(x)) ;::: ~4~TJ(X) (27)
(see POWELL's Lemma 5 in ;S} Then
2 1,b(x)
(1- t(x))JJ(x)p (x) <:: ~4~TJ(X)t(x) (2S)
will be sufficient for (22) to hold, and we get
J/·(x)
F(x) - ~(x,s(x)) ~ ~S~Tdx)t(x). (29)
Methods with stepsize control that have no information about p(x) will have problems
with (2S) in case of {l(x) > 0 and t(x) < 1. This is the reason why genuine second-order
methods (i.e. with possibly indefinite matrices B(x)) require a trust- region strategy or
strong a-posteriori information about p(x).
195
Now we apply Lemma 4.3 and use

1 1-,),
Lp(x)w(t(x)p(x)) + 2J.t(x)t(X)p2(X) ~ -8-t/J(x)r1(x) (30)
as a sufficient condition for (23) and (5), which can easily be satisfied by control of t(x) or
R (x) after testing (5). Line-search methods wi11 decrease t (x) until
t(x) ) 1 t(x) 1 - ')'

Lp(x)w ( -p(x) + -J.t(x)_p2(X) > --t/J(x)rt{x) (31)
1/ 2 1/ 8
while trust-region methods guarantee R(x) = Ro{x) or
R(x) (R2(X)) 1 R (x)2 1 - ')'

L--w - - + -J.t(x)-- > --t/J(xh(x). (32)
1/ 1/ 2 2 1/ 8
The convergence of algorithms in the sense of Theorem 2.1 requires rl (x) and t(x) to be
bounded from below by positive expressions depending on t/J(x) (see (29)). We state the
following facts that can be read off the equations derived so far:
Theorem 5.1
1. Using proiections in any convergent algorithm is no problem provided that u(x) is

controlled to satisfy (25).
2. Trust region methods (using t(x) = I,R(x} = r(x} and controlling R{x} to ensure
(5)) will converge if M{x} and J.t(x) are globally bounded from above by constants or
singularities of type t/J-"'(x} , a > O.
3. Methods with absolutely no information on p(x) (e.g. "pure" Gauss-Newton methods
with S(x) = Rn) cannot be proven to be convergent.
4. Methods with stepsize control will converge, If Jl(x) = 0 and If M(x), R(x) are globally
bounded .
•
Levenberg-Marquardt type methods need a special treatment in this context. They assume
p(x) -::: 0 and therefore have no problems with (28) and (29). If they control J.t{x} < 0 by
testing (5) and dividing J.t(x) by 1/ in case of failure, they will guarantee (5) via (30) after
finitely many iterations, because J.t(x) tends to -00. If I/J.t(x) fails, sti11
in case of w(t) -::: Kt (e.g. if f E C 2 ). Since the classical Levenberg-Marquardt method

has -J.t(x) = M(x), we can therefore assume a-posteriori that M(x) and J.t(x) are globally
196
bounded away from infinity. Furthermore, the a-posteriori bounds (13) and (14) can be
used as p(x).
This gives a general version of the well-known result (see e.g. OSBORNE [7] or HA USSLER
[4]):
Theorem 5.2 Levenberg-Marquardt methods converge in the sense of Theorem 2.1 . •
A revision of these arguments in view of (7) with g(t) = t 2 and Theorem 2.2 will reveal
nearly the same facts for the quantitative convergence theory:
Theorem 5.3 The methods considered so far are quantitatively globally convergent in
the sense of
00
L ljJ2(xd < 00,

k
except that M(x) = 0 has to be assumed for line -search methods.
Proof: Examination of F(x)- <l>(x,s(x)) reveals a O(~,2(x)) behavior, except for the case
of line-search methods, where M(x) = 0 is needed to avoid TJ(X) ~ 1jJ(x) and t(x) ~ 1jJ2(X),
o
which would otherwise lead to (1jJ4(X)) . •
The argument above shows that the case of M(x) > 0 for line-search methods can still be
handled by using g(t) = t 4 :
Theorem 5.4 Well-controlled line-search methods with positive semidefinite second-order

terms and globally bounded control parameters will give at least a convergence of
•
This illustrates how to apply our theory to certain examples from our class of algorithms.
A general theory for linear convergence will be given in a forthcoming paper.
6 A specific method
We give an illustrative example of a "well-controlled" method whose convergence be-
havior can be analyzed easily, but which does not fall directly into the algorithm class
described above. We (theoretically) minimize
<l>(x, s) = :pix, s) = h(f(x) + \7 f(x)s)
subject to the "quadratic" constraint
K(x)llsI1 2 ::; (1 - 1) (F(x) - h(f(x) + \7 f(x)s)) . (33)

197
If K(x) is not smaller than the constant Ko appearing in
Ih(J(x + s)) - h(J(x) + V f(x)s)1 v s:

IIsl1 2 ~ no lor x,s ERn,s i= 0 (34)
for f E C2, then (33) is a sufficient condition for (5), and Axiom 2.1 is satisfied.
A very simple control of K(x) can therefore be implemented by testing (5) and multiplying
the current (estimated or "last") value k(x) for K(x) by v in case of failure. Estimates
for k(x) are cheaply available via quotients of the form arising in (34). Since results
from optimization theory imply that the solution of the minimization subject to (33) is
a solution of a generalized Levenberg-Marquardt regularization (in the sense of [9]), it
will suffice to produce a Levenberg-Marquardt solution satisfying (33) with equality, using
some constant K E [k(x),vk(x)] on the left-hand side. This solution can be obtained by
a simple finite inner iteration of the associated Levenberg-Marquardt parameter, and for
convergence analysis we can use the final value of K as K(x) in (33).
This method can be viewed as a trust-region method using t(x) 1,o(x) J.L(x)
M(x) = 0 and a dynamic control satisfying (5) directly; it gives
I-I
p2(X) ~ ]((F(x) - ~(x, s(x)), (35)
where K is a positive lower bound of K(x).
It is now easy to prove that
F(x) - ~(x,s(x)) ~ t/1 2 (x)(1 - I)K- 1 (36)

proceeding along the lines of the lemmas, and the global convergence theory implies
Theorem 6.1 The method has quantitative global convergence in the sense of Theorem 2.2.
•
The following result gives sufficient conditions for local linear convergence of this method:
Theorem 6.2 The method has local linear convergence in parameter space, if around a
limit point x' the linearization gain of the optimal increment x' - x is at least quadratic
and at least proportional to the total available error gain F(x) - F(x·):
F(x) - ~(x,x' - x) ~ calix - x'112, co> 0, (37)
F(x) - ~(x,x' - x) ~ cl(F(x) - F(x')), Cl > O. (38)
Proof: It is easy to see from (37) that for a stepsize t(x) E (0,1], bounded away from
zero, the increment s = t(x)(x' - x) is admissible. Then (38) and
F(x) - ~(x, s(x)) ~ t(x)(F(x) - ~(x, x' - x))

198
L 2 -Norm Leo-Norm
Example Deuflh GN-Stp Osb-LM Quad.C GN-Stp Quad.C
BI 24 22 27 21 25 24
BPR over 82 102 67 81 75
FR 106 44 153 304 130 124
JS fail 54 21 49 52 43
KO 75 65 75 104 70 56
MR4 44 25 28 21 fail 21
MR5 45 26 28 24 27 24
MR6 52 84 124 138 98 134
OL1 48 54 306 266 51 58
OL2 121 108 605 341 139 over
W fail fail 403 81 136 131
Table 1: Function evaluations
Legend for Table 1:
Deuflh DEUFLHARD's method NLSQN, [3]

GN-Stp GAUSS-NEWTON method with stepsize control, [9]
Osb-LM OSBORNE's implementation of the Levenberg-Marquardt
algorithm, [7]
Quad.C Algorithm with quadratic constraint (this paper)
fail No convergence within 150 iterations
over Overflow
BI Example of BEN-ISRAEL, [4]
BPR Example of BARRODALE, POWELL, and ROBERTS,
[1] exp(x) - pdX)/Q3(X), n = 5
FR Example of FREUDENSTEIN and ROTH, [10], n = 2
JS Example of JENNRICH and SAMPSON, [10], n = 2
KO Example of KOWALIK and OSBORNE, [4], [5], [10]
MRx Example of MEYER and ROTH No. x, [4], [6]
OLI Example of OSBORNE, [lOj, n = 5
OL2 Example of OSBORNE, [lOi, n = 12
W Example of WATSON, [10], n = 12
199
make sure that Axiom 2.4 is satisfied. Local linear convergence of F(Xk) will then follow
from Theorem 2.3. From (5) and (33) we deduce
to get local linear convergence of Ils(Xk)11 = Ilxk - xk+lll to zero. Standard arguments will
then imply local linear convergence of {Xk} to X'. •
7 Numerical Experience
The theoretical results derived so far (and those about local linear convergence to
appear in a forthcoming paper) suggest that all well-controlled methods using no second-
order terms should have about the same convergence behavior. This is supported by
experience for a series of test runs with different methods and approximation problems
(see Table 1 below).
All examples presented here have rather bad starting values (taken from the literature, if
available), and in general the local convergence was surprisingly good. Reliable values for
linear convergence factors were rare, and therefore a comparison by local linear convergence
factors was impossible. Thus we restrict ourselves to simply report the number of function
evaluations (gradient evaluations counted n-fold).
We used the same stopping criterion for all methods, namely tjJ(x) ~ 10- 7 (1 + F(x)),
implemented by a shared and independent subroutine. Calculations were done in double
precision on a VAX 11/780 under VMS-FORTRAN. The gradient of f was calculated by
numerical differentiation by another shared subroutine.
The numerical results should be interpreted with caution. All methods spend most of their
time far away from the optimum, and the variations in numerical effort depend to a large
extent on pure chance, because the methods may happen to avoid or encounter regions
where 1iJ(x) is small, leading to poor progress in error reduction.
Bibliography
[1] Barrodale, 1., Powell, M.J.D., and Roberts, F.D.K., The Differential Correction Al-
gorithm for Rationalloc-Approximation, SIAM J. Numer. Anal. 9, 493-504 (1972)
[21 Ben-Israel, A., A Newton-Raphson-Method for the solution of equations, J. Math.

Anal. Appl. 15, 243-252 (1966)
200
[3] Deuflhard, P., Apostolescu, V., A study of the Gauss- Newton Algorithm for the Solu-
tion of Nonlinear Least Squares Problems, in: Frehse, J., Pallaschke, D., Trottenberg,
U. (ed.): Special Topics of Applied Mathematics, 129-150, North Holland 1980
[4] Hiiufiler, W.M., A Local Convergence Analysis for the Gauss-Newton and Levenberg-
Morrison--Marquardt-Algorithms, Computing 31,231-244 (1983)
[5] Kowalik, J., Osborne, M. R., Methods for unconstrained optimization problems. New
York: American Elsevier 1968
[6] Meyer, R.R., and P.M. Roth, Modified damped least squares: an algorithm for non-
linear estimation. J. Inst. Maths. Applies. 9,218-233 (1972)
[7] Osborne, M.R., Nonlinear least squares: the Levenberg-Marquardt algorithm revis-
ited, J. Austr. Math. Soc., Ser. B, 19,343-357 (1972)
[8] Powell, M.J.D., General Algorithms for Discrete Nonlinear Approximation Calcu-
lations, in: Chui, Ch.K., Schumaker, L.L., and Ward, J.D. (eds): Approximation
Theory IV, 187-218, Academic Press 1983
[9] Schaback, R., Convergence Analysis of the General Gauss-Newton Algorithm, Numer.
Math. 46,281-309 (1985)
[10] Watson, G.A., The Minimax Solution of an Overdetermined System of Non-linear

Equations, J. Inst. Maths. Applies 23, 167-180 (1979)
Prof. Dr. R. Schaback

Institut fur Numerische
und Angewandte Mathematik
der Universitiit Gottingen
Lotzestrafie 16-18
D-3400-Gottingen
ORTHOGONAL POLYNOMIALS IN OPTO-ELECTRONICS:

FIBER OPTICAL COMMUNICATION SYSTEMS
Walter Schempp
Lehrstuhl fuer Mathematik I

University of Siegen, 0-5900 Siegen, Germany
ABSTRACT
It is the purpose of this paper to show that harmonic

analysis on the diamond solvable Lie group allows us to
compute explicitly the quantized transverse eigenmode
spectrum of optoelectronic devices used in fiber optical
communication systems. As a result, the coupling coeffi-
cients of optoelectronic devices are expressed in terms
of Krawtchouk polynomials.
1. Introduction
In recent years there has been a dramatic increase in the
physics and applications of what might be termed opto-electronic
devices. This has been brought about both by the development of
the laser, w~th its rapidly growing list of applications, and
the staggering growth in semiconductor device electronics and
photonic material technology. The narrow beams of highly coherent,
nearly monochromatic light with almost perfectly plane wavefronts
emitted by the laser eliminated a major obstacle to light-wave
communication: ordinary sources of light cannot be modulated
rapidly enough to carry large amounts of information. What was
lacking in the early years of optoelectronics was a practical set
of photonic materials that could meet its technical demands.
Because of these demands the optical losses in silica glass caused
by absorption have been reduced in the past 25 years by a factor
202
of 10.000. A window made by the best optical waveguide glass

3 kilometers thick would be more transparent to light than
an ordinary window only 3 millimeters thick. The advent of
optical fibers with losses considerably less than 20 dB km- l
and with high information carrying capacity (e.g., bandwidths
> 100 MHz) has meant that they have become viable alternatives
to coaxial cables in communication links.
In any practical application of optical waveguide tech-
nology, the hair-fine glass fibers need to be incorporated in
some type of cable structure. The information capacity pre-
viously undreamed of, which optoelectronic signal processing
made available due to the highly coherent, nearly monochromatic
optical frequency carriers of laser sources and the wideband
transmission media of fiber optical cables will be illustrated
by the following facts: In 1956 the first transatlantic wire
cable TAT-l connecting the United States and Europe had the
ability to transmit 36 telephone conversations simultaneously
whereas the latest coaxial cable TAT-7 which was laid in the
mid-1970's has a capacity of 10.000 conversations. The first
intercontinental fiber optical cable TAT-8 which has been in-
stalled in 1986 and which is scheduled to begin operating in
1988 provides a communication capacity of nearly 40.000 indivi-
dual telephone conversations. The estimated costs of TAT-8
will be about 335 million US-$ shared by 30 telephone companies
located in U.S.A., Canada, and West-Europe. Essentially all
new long-distance telephone cables in the U.S. are optical
ones. Light links are even beeing employed to connect adjacent
electronic machines or equipment frames within a machine. In
such applications they are desirable for their immunity from
electromagnetic interference as well as for their high pulse
rates and small size. However, engineers suggest that the biggest
potential market of wideband fiber optical communication systems
presently lies in the field of civil telecommunication.
A schematic diagram of an optical transmission system
is shown in a figure at the end of the paper. The emitter is
usual a light emitting diode (LED) or semi-conductor laser,
203
whilst the detector may be a PIN photo-detector or avalanche

photo diode (APO). Electro-optical repeater units may be necessary
over relative long links to counter the effects of fiber trans-
mission losses and dispersion. In these the signal is detected,
amplified and then re-emitted. A separate power supply line must
be provided for the repeater units.
If an eigenmode emerging from one of the components,

say a laser, is injected into another component of the trans-
mission system, a fiber optical waveguide, for example, a set
of the modes of the communication system is excited. The coupling
to the various eigenmodes as well as the power transfer can be
described by the harmonic analysis on the diamond solvable Lie
group O(R) and evaluated in terms of the Krawtchouk polynomials.
2. The Economic Power of Laser Systems
The list displayed below shows the sales of laser systems

predicted for the Western industrial countries in 1986 (Sources:
Institut der Oeutschen Wirtschaft, Cologne. Laser 87, Munich).
Laser systems (100%) 10.6 billions US-$

Laser systems for print processing (52%) 5.5 billions US-$
Laser systems for signal processing (20%) 2.1 billions US-$
Laser systems for medicine (3.7%) 0.4 billions US-$
*****
Leading nations in laser technology U. S. A. and Japan
West Europe 10% of the total sale
Federal Republic of Germany <3% of the total sale
3. The Lie Group O(R)
Let us identify the peripheral circle T of a round

optical fiber with the maximal compact subgroup SO(2,R) of
204
Sp(l,R) = SL(2,R) the elements of which act as automorphisms of

the three-dimensional real Heisenberg 2-step nilpotent Lie group
A(R) by rotating the first two coordinates of the elements
(x,y,z) EA(R) and leaving the third "central" coordinate fixed.
Then the external semi-direct product
D(R) = T IX A(R)
is called the diamond group. It forms a four-dimensional, real,

connected, non-exponential, solvable Lie group. Let
denote the Stone-von Neumann representation and HA the Hamil-

tonian of the harmonic oscillator with eigenvalues -27r(sign A).
(2n+i), n EN. Then Hi forms the infinitesimal generator of
SO(2,R) in Sp(l,R). Let
denote the double covering of the torus group T = I e 27ri8 I 8 E R I

and define a unitary linear representation SA of T x R by
SA(e 7ri8 ,z) = e i8H A
and the characters of T x R by the prescription
e 27ri ((2n+i)8+Az)id (n E Z) .
L 2 (R)
Then SAUAiiI X~ (nEZ) forms a family of topologically irreducible,

continuous, unitary, linear representations of D(R) having
UA as their restrictions to A(R). Conversely, each representation
of D(R) having these properties is unitarily isomorphic to
exactly one of the representations SA UAiii X~ (n E Z). See Lion [ 2].
4. Quantized Transverse Eigenmode Spectra
A wavefunction tf;E L2 (R) is called a quantized transverse

eigenmode of a round optical waveguide if the auto-correlation
205
H(~;x,y) = radial on the plane R ~ R. The

<U1(x,y,0)~I~>is
following theorems characterize the quantized transverse eigen-
modes and give the explicit forms of the associated cross-
correlation function H(~,~;x,y) = <U1(x,y,0)~I~>, respectively
(cf.[3] and [4]).
Theorem 1. Let (Hn)n>O denote the sequence of Hermite functions.

The waveform ~E L2 (R) is a quantized transverse eigenmode
of a circular optical waveguide if and only if
for an integer n ~ O.
Theorem 2. The wavefunction ~ E L2(R ~ R) is a quantized trans-

verse eigenmode of a rectangular optical waveguide if and only if
for integers n > 0, m > O.
Theorem 3. Let (L(P)) denote the sequence of Laguerre functions

n n>O
of order P >-1. Then Schwinger's formula
H(H m,H n ;x,y) = 1* 1 . I

m.j (V 7r (x+iy))
-
m-n (m-n)
Ln
2 2
(7r(x +y ))
holds for (x,y)E R ~ R.
5. Opto-Coupling
If the wave functions ~' ,~' and ~,~ belonging to L2 (R)

represent two quantized transverse eigenmodes of two coaxial
optical devices like laser resonators or optical fiber wave-
guides, then their coupling coefficient is defined according
to the prescription
206
C(1{;' ,<p' , 1{;, <p) f f H(1{;' ,<p' ;x,y) .H(1{;,<p;x,y)dxdy.
R Iil R
The integral has to be evaluated at the coupling plane R Iil R

transverse to the common beam axis of the optical devices. In
the present section we will calculate the coaxial coupling
coefficients explicitly in the circular as well as in the rec-
tangular case in terms of the Gaussian beam parameters a and
q (cf. Kogelnik [1 ]).
Recall the definition of the Krawtchouk polynomials Kn(x;p,N).

For all integers n > 0 these hypergeometric polynomials are
given by
where N ~ 0, 0 ~ x ~ N, PE]O,l]. In terms of shifted raising

factorials, Kn(x;p,N) admits the following expression
(0 < n < N).
The Krawtchouk polynomials are orthogonal on the set I 0,1, ... , NI

with respect to the binomial distribution. Thus their ortho-
gonality relations take the discrete form
L..J
' " K (x;p,N)K (x;p,N) (N) p x (l-p) N-x o (n '" m)
O<x<N n m x
for n < N, m < Nand pE ]0,1[.
Theorem 4. Let m > n > 0, m' > n' > 0 be integers. Keep to
the preceding notations.
a) In the circular case set m p, m - n 1, m' p', m' - n'

= 1'. Then
Cp, 1 , p ' 1
, ' = 0 for I * 1',
i.e., there is no coupling between quantized transverse eigen-
207
modes of different angular moments. Moreover,
2 )1 +1 ( P+P I +1) ! (l-~)P(l-~)Pl

(
Cp,l,p',l = ww'q ..JP!p'!(p+l)!(p'+l)! q q
K (p,.(q-a)(q-a' ) p+p'+l)
p 'q(q-a-a')'
b) In the rectangular case we have
m+m' - 1 mod 2,
Cm,n,m " ,n a for
n+n' - 1 mod 2,
i.e., there is no coupling between even and odd quantized trans-

verse eigenmodes. In other words, the parity is preserved under
the coupling of quantized eigenmodes. In the case
mI = 2~', m = 2~, nI 2 v', n 2v
we get in terms of the Krawtchouk polynomials
Cm, n, m" , n
( l_~)~+v (l_~)~I+vl.K (II,.(q-a)(q-a' ) II+II'-'!')

q q ~ .. , q(q-a-a') ' .... 2
K (v'· (q - a) (q - a I ). V + V I _~)
V ' q(q-a-a ' ) , t:.
A similar result holds in the case
m' 2~'+l, m 2~+1, n' 2 v' +l, n 2v+1.
The fact that the weight function associated with the Krawtchouk
polynomials is discrete reflects the fact that the quantized
transverse eigenmode spectra of the circular and rectangular
208
fiber optical waveguides form a discrete series. For details

see [4] and [5].
Opto-couplers are used as repeaters in optoelectronic

telecommunication systems. Moreover they are used to separate
electrically analog and digital circuits. The figure at the
end of the paper displays an example of an opto-coupling device.
References
1. Kogelnik, H., "Coupling and conversion coefficients for

optical modes". Proc. of the Symposium on Quasi-Optics,
pp. 333-347. Brooklyn, N.Y.: Polytechnic Press 1964
2. Lion, G., "Extensions de repr~sentations de groupes de
Lie nilpotents et indices de Maslov". C.R. Acad. Sc. Paris
288 (1979), S~rie A, pp. 615 - 618
3. Schempp, W., "Harmonic analysis on the Heisenberg nilpotent

Lie group, with applications to signal theory". Pitman
Research Notes in Mathematics Series, Vol. 147. Harlow,
Essex: Longman Scientific & Technical. New York: John Wiley
& Sons 1986
4. Schempp, W., "The oscillator representation of the meta-
plectic group applied to quantum electronic and computerized
tomography". Dordrecht-Boston-Lancaster-Tokyo: D.Reidel
(in print)
5. Schempp, W., "Laseroptik und nicht-kommutative harmonische
Analyse: Optoelektronische SignalObertragung". Notes of
Lectures given at the University of Vienna and the Technical
University of Vienna (to appear)
I _
r----- ----,I r -POwei1udline-1 r-------,
I _ I
I ~-----~I
e I P+k 11) ~out

I ~
~
~um)!II
_____
I Cl !\
Emitter
_ _ ...JI
l!ho~~~c~~ _-' I!~~~~~~.J
Transmitter Optical fiber Repeater unit Emitter Receiver
~
\0
Schematic diagram of the main components of a fiber optical com-
munication system.
210
An opto-coupler
211
... ............
'•. ,.""",..•••...••• -.
............. ..... .
........
. . . . . . . . . . . . . . . . . . . It • • • • •
. . ..
..
~
~
~ .... ........... .... ..... ........ . .....
.. .. --
.. ......
..
.
-
.... ........... ., ............• ...... ..
~ ~
. .................
~
........
.... .. .
................
,
..
~ ~
. -.
....................•
.....................
-
...
...... ...... .... ....
.. .....................•
~ ~
•......... ,...........•
~
-_ ..............""".
~ ~
Higher transverse eigenmodes of a rectangular laser .

212
e 10111811DIIOO
0000000 000000
cell 000000000000.
cc:. OOOOOOOOOOOOOD CD:>
<c. 000000800000.~
~oaOOOOOOQOOOOOOO~
ell.
~ooooooooo~ooo.~o~
~o oooooooooooooo~
<C:a CO 01001000000000 0 0 0 c:D;l
<a::D <00) 0 0 10 00 eOl 0 0 0 0 0 c::IID
~OOOOOGO~ooo~ooo.~
<S:a 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 10 ~
~ oooooO>oooo~oooo 0> c==:>
<CCD 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 ~
<a:3 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 CD c:z:>
~ '000'0'0'00'000000'000 ~
<!ID OOOOOOOOOOOOOO«!J 0 ~
<C:DOOOOOOOOOOOOOOOO~
<C:It • • 000 OOGOOOOOO • • c::D:>
cca OOODQOOOOOOOOOc::a>
0000000000000 _
000000 0000000
00811118101"'"
Experiment and computer plot (according Theorem 2)
ON THE CONVEX CUBIC C2-SPLlNE INTERPOLATION
Jochen W.Schmidt
1. INTRODUCTION
In this note a necessary and sufficient criterion is

given under which the problem of convex interpolation has at
least one C2-solution. This condition applies to cubic splines
and also to some types of related splines. Further, recently
developed results concerning the convex interpolation with
cubic C1-splines are reviewed.
Let 6 be a partition of the interval [0,1],
The data-points (xo'Yo), ••• ,(xn'Yn) associated with 6 are said

to be in convex (strictly convex) position if
't'1 ~ 't'a ~ ••• ~ 't'n ('t'1 < 't 2 < ••• < 't'n)
where = (Yi-Yi-1)/hi , hi = x i -xi_1. Now it is of interest

't i
whether an interpolating spline s, i.e. a spline s with

(1) s(xi ) = Yi (i=O, ••• ,n)
is convex on [0,1] or not.

214
For some types of splines this problem of convex

interpolation can be equivalently formulated by a finite set
of conditions with a finite number of parameters, see Neuman
[8], He~/Schmidt [6], Schmidt [11] for the convex C1-interpola-
tion and Fritsch/Carlson [4] for the monotone C1-interpolation.
In the case of the C2-interpolation these convexity preserving
conditions are often of the form
(2) PiMi-1+(O'i + 0' i+1 )lIi + Pi+1 Mi+1 ='t' i+1- 't'i (i=1, •• ,n-1),
(3) Mi ~ 0 (i=O, •• ,n)
where the constants Pi and 0' i satisfy
Thus, the problem of the nonnegative solvability of a tridia-

gonal system of n-1 equations and n+1 parameters Mo,M1' •• '~
arises. The aim of the present note is mainly to derive a
necessary and sufficient condition under which a solution of
this system (2),(3) exists. For conditions which are only suf-
ficient see Miroshnichenko L7].
2. CONVEXITY PRESERVING CONDITIONS FOR SOME C2-SPLINES
2.1 CUBIC SPLINES
Let SP3(2) the set of cubic C2-splines on d. The

cubic spline s defined by
wi th x = x i - 1 + hit (i=1, •• ,n) belongs, as it is well-known,

to Spj(2) if and only if
215
The interpolation condition (1) holds, and it follows
Further, because of SIl(X) = (1-t)Mi _ 1 +tMi for x i - 1 S x:li xi'

s is convex on [0,1] if and only if (3) is valid. Thus, one is
led to a problem (2),(3) for which (4) holds. This problem (2),
(3) may be not solvable, i.e. there are data sets being in con-
vex position for which all cubic C2-interpolants are not convex
on [0,1].
2.a EXPONENTIAL SPLINES
Let ESp(aiA) be the set of exponential Ca-splines s,
where A1 > O, ••• ,A n > °are given parameters, and A = (A 1 , .. .A n )·

For x i - 1 :5 x S Xi' with x = x i - 1 + ~t and ~i =hiAi' set
Then it follows (1) and (7). Further, s E ESp(a;A) if and only

if (2) holds, with
sinh~-J.L.i ~icosh~-sinhJ.L.i
(9) Pi = 2. hi' a i = h. ,
~i s~nhllt ~~sinh IJ:i. ~
see e.g. Pruess [9], Rentrop [10]. Obviously, (4) is true and
216
furthermore s is convex on [0,1J if and only if condition (3)

is valid. Thus, for the convex C2-interpolation with exponen-
tial splines a problem (2),(3) arises.
Because of
system (2) tends to a diagonal system, and in the case of

strictly convex data sets the solution is positive. Thus, for
sufficiently large parameters A1 , •• ,An the problem of convex
interpolation is solvable in ESp(2;A)i see also [9J.
2.3 RATIONAL SPLINES
The rational splines introduced by Gregory [5J,

Miroshnichenko [7J are used here in the following form:
(10) sex) = Yi_1+'f i h i t -
t(1-t)hf (2+ri-(1+ri)t)Mi_1+(1+(1+ri)t)Mi
(1+ri )(6+2ri ) 1+ri t(1-t)
o S t ~ 1
with x = x i _ 1 + hit (i=1, •• ,n), and with parameters

r 1 > 0, •• ,rn > 0, r = (r1 , •• ,rn ). One gets (1) and (7), and s
is a C2-spline if and only if (2) is valid, now with
(11)
In this case the corresponding set of splines is denoted by

RSP3(2 jr).
217
Because of
(1+ri t(1-t»
3
s"(x) •
(X(t)"i_1 + ~(t)Mi
(1+ri)(6+~ri )
for ° S t s 1
with aCt) :~t)+(~+ri)y(1-t),~(t) : (~+ri)y(t)+Y(1-t),
yet) : -~+6t+~ritJ and aCt) ~ 0, Set) ~ ° for 0 S t s 1, the

spline s is convex on (0,1] if and only if (3) is satisfied
[11]. Thus, the convexity of interpolants s E RSP3(~;r) is
also determined by (~),(3).
The limits
°< a f+OPi i+1 < 0iPi : ~

,,+ri
... °for r i ... + 00
assure that there exist convex interpolants from RSPJ(~;r) if

the data sets are in strictly convex position and if the para-
~ters r 1 , ••• ,rn are sufficiently large.
3. AN ALGORITHM FOR SOLVING (2),(3)

It is easily seen that the solution set of system
(~) can be described as follows:
where the numbers (Xi'~i'Yi are recursively given by
(Xi+1: -{Pia i-1+(Oi+o i+1)ai}/Pi+1' (Xo: 1'(X1: 0;
(13) f3 i +1 : -{Pi~i-1+(0 i+ai+1)~i}/Pi+1' 90 :0,9 1 = 1;
Yi+1 = {'l:i+1-'ti-PiYi-1-(Oi+ai+1)Yi}/Pi+1'YO= 0'Y1 =0.

These numbers possess the following properties valid for
218
i = 1, •• ,n-1:
(14) sign cxi+1 = (-1) i ,lcxi+11 ~ IUil;
(15) sign a i +1 = (-1)i, lai+1 1 ~ lai l ;

cxi+1 cxi
(16) -~-
ai +1 ai
Proof: Obviously (14),(15),(16) are valid for i = 1. If

Icxil ~ Icxi-11 is assumed then, in view of (4), it follows
Pi cxi-1 ai ai+1
cx =-cx (--+-+-),
i+1 i Pi+1 cxi Pi+1 Pi+1
implying sign cxi+1 = -sign cxi and Icxi+11 ~ ICXil. Analogously

(15) can by verified.
and then, because of -a i +1 ai > O. in finishing the proof
CXi+1 cxi
~~
Pi+1 ai
Now, using (12) and (15), inequalities (3) read
M1 ~ -(UiMo+vi)/ai for i odd, 1 • i • n,
M1 • -(CXiMo+vi)/ai for i even, 1 • i • n.

Thus, the following statement is justified.
219
PROPOSITION: Under the assumption (4), problem (2),(3) is sol-

vable if and only if
(18) a(x) = max{ -(cxix+y i )/a i : i odd, 1 ~ i ~ n),
(19) b(x)=min{-(uix+Yi)/13i: i even, 1 ~ i ~ n},
(20) A =0, B =-Y2/cx2.
The determination of the essential set P can be per-

formed successively. Let P j be the set which results if only
the first j constraints are considered,
Pj = {(Mo ,M1 ): aj(Mo ) ~ M1 ~bj(Mo)' Aj ~ Mo~ Bj }

with
aj(x) =max{-(cxix+Yi)/ai: i E Kj }
bj(X) =min{-(cxix+Yi)/13i: i E Lj }.
Here Kj ~ {i : i odd, 1 ~ i
j}, Lj C {i : i even. 1
~ ~ i ~ j}
denote the index sets of the relevant constraints.
At the beginning set K2 ={1}. L2 ={2}. A2 =O. B2 =-Y2/cx2"

Now. P j +1 shall be constructed starting from P j " To this end
denote by (xi'Yi) the intersection pOint of the two lines
gi: Y =-(cx i X +Yi)/13i' gj+1: Y = -(cx j+1 X +Y j+1 )/a j +1 "

At first. let j be odd. Then. if (16) is taken into account.
the procedure is as follows, compare with figure 1:
220
Case 1: If xi < Aj for i = max Kj then P j +1 =0 implying P = 0.
Case 2s If xi iI Bj for i = min Kj then P j +1 = P j , i.e. set
Kj+1 =Kj' Lj +1 = Lj' Aj+1 = Aj' Bj+1 = Bj •

Case )s Otherwise, let ~ € K j be the largest index with
Y~ = max{Yi: i € K j };
set Aj+1=Aj,Bj+1=x~, Kj +1 =K j \{i€K j S i<~};
let v € L j be the smallest index with
Yv • min{Yi: i € L~};
if Xv > Aj set Lj +1 = Lj\{i € L j S i >,,} v {j+1} other-
wise set Lj +1 = {j+1}.
Next, let j be even.

221
Case 1: If xi > Bj for i = min Kj then P j +1 = 0.
Case 3: Otherwise, let ~ E Lj be the largest index with
Y~ = min{Yi: i E Lj };
set Aj+1 = max{Aj,x~},Bj+1 = Bj ,L j +1 =L j \{iE Lj : i<~};
let v E Kj be the smallest index with
Yv = max{Yi: i E Kj };
set Kj +1 = Kj\{iE Kj : i > v} v{j+1}.
In both cases the determination of the indices ~ and v should

be done by using property (16).
4. OPTIMAL CONVEX C2-SPLINES
For Pn = P *
0 every point (M o ,M 1 ) E P leads to a
convex interpolating spline s. Using (12) one gets M2 , •• ,Mn
and, e.g. for cubic C2-splines, formula (5) applies immediate-
ly. In general there exist an infinite number of convex inter-
polants. For selecting one of them an objective function is to
be nominated, e.g. in SP3(2)
n Xi n
E Wi S sll(x)2dx = E
i=1 %i-1 i=1
see l2]. With regard to Holladays theorem the weights may be

Wi : 1 (i=1, •• ,n) while an approximation of the geometric cur-
vature leads to Wi = 1/(1~f)3 (i=1, •• ,n). In the spaces
222
ESp(2;h) and RSP3(2;r) the function f2 should be somewhat

modified, see e.g. [10],[6]. Moreover, an approximation of the
integrals by a quadrature formula should be taken into account.
In view of formula (12), f2 becomes a quadratic fUnc-

tion depending only on Mo and M1 , say f 2 (S) = F(Mo ,M 1 ). Thus,
the two-dimensional quadratic optimization problem
(22) F(Mo ,M1 ) ... Min!
subject to O:iMo+l3 i M1 + Yi ~ ° (i E Kn U Ln)
arises. Such a program can be solved by applying well-known

a~gorithms or even by simple geometric considerations.
5. CONVEX CUBIC C1-SPLINES
Let SP3(1) the set of cubic C1-splines on the grid 6. Then the
spline
(23) sex) = y i-1+mi-1 hi t+O''r i -2mi _ 1-m i )hi t 2 +
+ (m i _ 1 +m i -2r i)~ t 3 , 0" t" 1
with x = x i _ 1 +hi t (i::;;1, •• ,n) is from SP3(1). Moreover, inter-

polation condition (1) is satisfied, and
(24) s'(xi) ::;; mi (i::;;O, •• ,n).
This spline interpolant is convex on [0,1] i f and only if
see Neuman [8]. A necessary and sufficient criterion for the

solvability of (25) reads
223
(26) ai _1 ~ ~i (i=1, •• ,n)
where the ai's are recursively defined by
(27 )
a i = max{~i,(Jri-bi_1)/2},bi = Jr i -2a i _ 1 (i=1, •• ,n),
see Schmidt/He~ [12J.
The determination of the optimal interpolan~ with

respect to the objective function (21) leads to the weak~
coupled program
[2(8)
(28)
with
(29)
Generally, to (28) belongs the dual program of the form

n
()O) - E H1(Pi-1,-Pi) ~ max! with Po = P = 0
i=1 n
where the so-called Fenchel's conjugates are to be computed by
In the case (29) one obtains

224
Ifi (t '11) = ~ i (t +1')) + 1;;i (t 2-('fI+fl2) for t ~ 0, fill 0
= ~ i (t +1')) + 1;; (~- fI)2 for 0 ~ t ~ 2t)

i
(32)
and furthermore, then return-formula
holds, see Burmeister/He~/Schmidt [2], Dietze/Schmidt [3].
For computing optimal convex SP3(1)-interpolants

these results are used as follows. At first the existence is
tested by (26). Then, i~ (26) is valid the unconstrained pro-
gram (30),(32) is solved by the usual Newton method starting
with Pi = ~i~i+1 (i=1, •• ,n-1). Finally, return-formula (33)
gives the quantities mo ,m 1 , •• ,mn , and the optimal spline s is
determined by (23). This procedure turns out to be very effec-
tive. The total amount of arithmetical operations is equal to
O(n) because Newton's method is observed to terminate after a
small number of steps, see [2].
The described technique also applies to the monotone

SP3(1)-interpolation [3], the ESp(1;A)-interpolation [6] and
the RSP3(1;r)-interpolation [11]. Furthermore, two-dimensional
problems leading to a program
225
m+1 n m n+1
i;1 j;1 Fij(mij,mij+1)+ 1;1 j;1Gij(mij,mi+1j)~minl
(34) subject to (mij ,m ij +1 ) E Wij (i=1, •• ,m+1, j=1, •• ,n),
(m ij ,mi +1j ) E Vij (i=1, •• ,m, j=1, •• ,n+1)
can be treated by extending the dual program (30),(31).
6. APPENDIX. A STABLE VERSION OF THE ALGORITHM OF CHAPTER 3
If the dimension n is small the algorithm proposed in

chapter 3 works well. But for larger n the instability of for-
mula (13) caused by (Oi+Oi+1)/Pi+1 > 1 makes it necessary to
somewhat modify this algorithm. To this end, let now Mo and ~
be the parameters on which the solution set of system (2) 1s

depending. Then by means of (12) it follows immediately
(5) Mi = aiMo +biMn +c i (i:::O, ••• ,n)

with
(36) a i = (ai~n-anai)/~n' b i = ~i/an'
c i = (Yian-Yn~i)/an'
Let, e.g., n be even. Then, because of an < 0, condition (3)
leads to
Therefore, problem (2),(3) is solvable if and only if
where
226
a(x) = max{-(aix+ci)/b i : i even, 1 :Ii i ~ n},
b(x) = min{-(aix+ci)/b i : i odd, 1 :Ii i ;a n},
A = 0, B = -c 1/b 1 •
Because of (16) and ()6) it follows
a
O:li i+1
bi +1
=
-
and hence the determination of the set P can be done succes-
sively in an analogous manner as shown for P in chapter ).
Now, it is essential not to compute the numbers ai,b i and ci
by means of formula (36), but directly by solving system (2)
via Cholesky factorization. In more detail, write system (2) as
with the diagonal dominant tridiagonal matrix
s = 0' 1+0'2 P2
P2 0' 2+0'3 P3
Pn - 2 O'n-2+O'n-1 Pn- 1
Pn-1 O'n_1+ O'n
and the vectors
227
Then the vectors A,B and C with the components ai,b i and ci'
satisfy the linear systems
SA =E1 , SB =En- 1 , SC =D
which can be solved by using the Cholesky factorization S =L LT
of the matrix S. Following this line a stable method for
computing the coefficients in (35) is established.
REFERENCES
[1] W.Burmeister, S.Dietze, W.He~ and J.W.Schmidts Solution of
a class of weakly coupled programs via dualization, and
applications. Proceed.ISAM Wei~ig 1986, Akademie-Verlag
Berlin (to appear).
[2] W.Burmeister, W.He~ and J.W.Schmidt: Convex spline inter-
polants with minimal curvature. Computing 35(1985),
219-229.
[3] S.Dietze and J.W.Schmidt: Determination or shape preserving
spline interpolants with minimal curvature via dual pro-
grams. TU Dresden-Informationen 07-06-1985 and J.Approx.
Theory (to appear).
[4] F.N.Fritsch and R.E.Carlson: Monotone piecewise cubic
interpolation. SIAM J.Numer.Anal. 17(1980), 238-246.
[5] J.A.Gregor.y: Shape preserving spline interpolation.
Computer-aided design 18(1986), 53-57.
[6] W.He~ and J.W.Schmidt: Convexity preserving interpolation
with exponential splines. Computing 36(1986), 335-342.
228
[7] V.L.Mirosbnichenko: Convex and monotone spline interpola-

tion. Proceed.Constr.Theory of Funct. Sofia 84(1985),
610-620.
[8] E.Neuman: Uniform approximation by some Hermite interpola-
ting splines. J.Comput.Appl.Math.4(1978), 7-9.
[9] S.Pruess: Properties of splines in tension. J.Approx.
Theory 17(1976), 86-96.
[10] P.Rentrop: An algorithm for the computation of the ex-
ponential splines. Numer.Math. 35(1980), 81-93.
[11] J.W.Schmidt: On shape preserving spline interpolation:
existence theorems and determination of optimal splines.
Banach Center Publ. XXVII, Approx.Theory Funct.Spaces 86
(to appear).
[12] J.W.Schmidt and W.He~: Schwach verkoppelte Ungleichungs-
systeme und konvexe Spline-Interpolation. Elem.Math.
39(1984), 85-95.
Jochen W.Schmidt
Technical University of Dresden
Mommsenstra~e 13
Dresden, 8027, German Democratic Republic
RAYLEIGH-QUOTIENT-MINIMIERUNG MIT VORKONDITIONIERUNG
Hans-Rudolf Schwarz
The smallest eigenvalues and corresponding eigenvec-

tors of Ax = ABx, where A and B are symmetric, positive defi-
nite and sparse matrices of high order, can be computed by
minimizing the Rayleigh quotient by means of the method of
conjugate gradients. An appropriate preconditioning of the
eigenvalue problem results in an essential improvement of the
convergence. It is interesting that an implicit realization of
the preconditioning does not increase the required computational
effort per iteration step. The algorithm is highly vectorizable
and is indeed a very efficient method with respect to computa-
tional work and to storage requirements.
1. Einleitung
Wir betrachten das allgemeine Eigenwertproblem
Ax
- = ABx
- ( 1)
mit symmetrischen, schwach besetzten Matrizen A und B hoher

Ordnung n, und wir setzen ohne Einschrankung A und B posi-
tiv definit voraus. Gesucht sind die m kleinsten Eigenwerte
(2 )
und die zugehorigen Eigenvektoren ~1' ~2' ... , ~ von (1),

so dass
Az.
-)
= A.Bz.,
) -)
(j=I,2, ••• ,m) (3 )
gilt. Die Anzahl m der gewtinschten Eigenpaare (A.,Z.)

) -)
sei
klein im Vergleich zur Ordnung n der Matrizen. Bei den typi-
schen Anwendungen in der Methode der finiten Elemente besitzen
die Matrizen A und Beine identische, jedoch unregelmassige
Besetzungsstruktur.
230
Zur numerischen Behandlung der Aufgabe existieren

effiziente Verfahren wie der Lanczos-Algorithmus [6,10,12], die
Bisektionsmethode [11,12] oder die Vektoriteration [11,12],
welche eine Cholesky-Zerlegung oder allgemeiner eine LDU-Zerle-
gung einer bestimmten symmetrischen Matrix erfordern, wobei die
schwache Besetzung der zu zerlegenden Matrix zumindest innerhalb
der Hlille verloren geht. Dieser Teilschritt kann deshalb sehr
speicheraufwendig seine Die schwache Besetzung der Matrizen A
und B kann nur von Algorithmen zur Minimierung des Rayleigh-
schen Quotienten voll ausgenlitzt werden, da nur sie die Multi-
plikation der Matrizen mit einem Vektor benotigen. lm folgenden
betrachten wir die Methode der konjugierten Gradienten zur Mini-
mierung des Rayleighschen Quotienten zur Berechnung des Eigen-
paares (A1'~1). Durch eine Vorkonditionierung der Aufgabe kann
die Konvergenzeigenschaft so entscheidend verbessert werden,
dass ein sehr effizientes Rechenverfahren resultiert, sukzessive
die kleinsten Eigenwerte zu berechnen [7].
2. Minimierung des Rayleighschen Quotienten

Das Minimum des Rayleighschen Quotienten R[~] zu (1)
~st gleich A1 und wird von ~1 angenommen:
xTAx
min R[~] = min = A1 = (4 )
~;tQ ~;tQ ~TBX
Urn das Minimum von R[~] iterativ mit der Methode der konjugier-
ten Gradienten [4,5,8,9] zu bestimmen, wird mit dem Gradienten
von R[~] zu einem lterationsvektor ~k
gk = g(~k) = grad R[~k] = ~{A~k - R[~k]B~k} (5)

~kB~k
eine Suchrichtung Ek+1 festgelegt, in welcher R[~] minimiert
wird. Bedeutet -0
x den Startvektor, dann definiere ich die
Suchrichtung im Gegensatz zur liblichen Festlegung [3,5,7,8] im
folgenden durch
(k=1,2, ••• ) . (6 )
231
Mit den Skalarprodukten im Koeffizienten von Ek trage ich der

dem Problem (1) angepassten Metrik Rechnung, und die Definition
(6) wird sich bei der Vorkonditionierung vorteilhaft herausstel-
len.
Der nachste Iterationsvektor
(k=O, 1, 2 , ••• (7 )
bestimmt sich fUr 0k+1 = ° aus der Forderung

T
~kA~k +
= Min! (8)
Mit den Grossen

T T T
a := ~kA~k' ~ := Ek+1A~k' y := Ek+1AEk+1 ( 9)
T T T
p := ~kB~k' a := Ek+1B~k' 1 := Ek+1BEk+1 (10)
erhalt man aus (8) nach einfacher Rechnung [7] mit
~ := (al - yp)2 - 4(ya - ~l)(~p - aa) > 0 (11)
den gesuchten Wert
0k+1 = [al - yp + 1~]/[2(ya - ~l)]. (12)
Damit kann der grundlegende Algorithmus zur Minimie-

rung des Rayleighschen Quotienten mit der Methode der konjugier-
ten Gradienten wie folgtzusammengefasst werden, falls fUr die
Vektoren Yk := A~k' Yk:= B~k eine rekursive Berechnung ver-
wendet wird, zudem die Vektoren ~k := AEk und ~k := BEk ein-
gefUhrt werden und schliesslich eine Indexsubstitution vorgenom-
men wird.
Start: Wahl von x Q:

-0
'"
v = T
v
-0
= A~, -0 B~i a = ~Ya' P
T~
= ~Yai
q = a/pi ~ 2(Ya - qYa)/p.
232
Iterationsschritt (k=1,2,3, ... ):

falls k = 1:
falls k > = nT ~k-1

Bn
~k-1
InT
~k-2
Bn
~k-2
(13 )
~k = A£k , ~k = B£k ;
T
= 2k-1~k' y
T xT w
cr = -k-1-k'
~ = £k~k'
fj, und ok gemass (11) und (12) ;
2k = 2k-1 + °k£k
Yk = Yk-1 + °k~k
T T~
a = 2kYk' P = 2kYk q = alp
Test auf Konvergenz
Der wesentliche Rechenaufwand eines Iterationsschrit-

tes setzt sich zusammen aus drei Matrix-Vektor-Multiplikationen
Bgk - 1 , A£k' B£k' aus sieben Skalarprodukten und sechs Multi-
plikationen eines Vektors mit einem Skalar.
3. Konvergenz und Vorkonditionierung

Die Konvergenz der Folge 2k des Algorithmus (13)
gegen die Richtung von ~1 wird im wesentlichen bestimmt durch
die Konditionszahl der Hesseschen Matrix H(2) des Rayleighschen
Quotienten, ausgewertet fUr den Eigenvektor ~1. Es gilt
H(x) = _2___ {A - R[x]B - g(2)T B2 - (B2 )Tg (2)}. (14)

- xTBX -
FUr den B-normierten Eigenvektor ~1 ist wegen g(~l) = 0

H(~l) = 2{A - AlB} • (15)
233
FUr die Eigenvektoren z.

-)
der Eigenwertaufgabe (1) gilt aber
Folglich ist die Spektralnorm von H(~l)' zugeharig zur Vektor-

norm 11~112 , B = (~TB~) 1/2, gegeben durch
(17)
Da H(~l) singular, aber positiv semidefinit ist, ist die mass-

gebende Konditionszahl K2,B(H(~1» durch den Quotienten des
grassten zum kleinsten positiven Eigenwert gegeben [8].
FUr die notwendige Anzahl ~S von Iterationsschritten,

die Lasung eines linearen Gleichungssystems mit Hilfe der Methode
der konjugierten Gradienten mit einer relativen Genauigkeit £
zu berechnen, existiert eine Schatzung [1], die wir sinngemass
Ubernehmen. SolI ~1 mit der relativen Genauigkeit £ berechnet
werden, gilt die qualitative Aussage
~S = ~ I ( 19)
In Analogie zur Vorkonditionierung bei linearen Glei-

chungssystemen [11] ist es maglich, die Konvergenz der Methode
der konjugierten Gradienten zur Minimierung des Rayleighschen
Quotienten dadurch wesentlich zu verbessern, dass die Konditions-
zahl der Hesseschen Matrix durch gleiche Massnahmen stark ver-
kleinert wird. Zu diesem Zweck set zen wir
C regular. (20)
Substitution im Rayleighschen Quotienten (4) ergibt
yTc-lAC-Ty
R[~] = = =: R[y] (21)
yTc- 1BC- Ty
mit
(22 )
234
Die Matrizen
-
A und
-
B sind kongruent zu A und B und folg-
lich symmetrisch und positiv definit. Der Rayleighsche Quotient
R[y] entspricht dem allgemeinen Eigenwertproblem
-
Ay = ABy,
-
dessen Eigenvektoren y. = cTz., (j=1,2, ... ,n) zu den Eigenwer-
J -J
ten Aj sind. Die zugehorige Hessesche Matrix zum Eigenvektor
Y1 ist gemass (15)
H(Y1) = 2{A - AlB} = 2C- 1 {A - A1 B}C- T = C-1H(~1)C-T. (23)
Sie ist ahnlich zur im allgemeinen unsymmetrischen Matrix
-
Die Matrix K liefert im Spezialfall, dass C die Linksdrei-
ecksmatrix der Cholesky-Zerlegung von A = cc T darstellt, einen
Anhaltspunkt Uber die gUnstige Wahl von c. Denn in diesem Fall
gilt fUr
(25)
und somit fUr die Eigenvektoren von (1)
(26)
-
Wegen der Ahnlichkeit der Matrizen_ K und H(Y1) sind die
Eigenwerte der Hesseschen Matrix H(Y1) in aufsteigender Reihen-
folge
( 27)
Folglich ist jetzt die massgebende Konditionszahl fUr die norma Ie

Spektralnorm gegeben durch
- 1 - AllAn A - A1.A2
n
K 2 ,I(H(Y1)) = = (28)
1 - A1/A2 A2 - A1 A1
Im Spezialfall A = cc T gilt also
(29)
235
so dass immer dann eine wesentliche Verkleinerung der Konditions-

zahl erzielt wird, falls A2 « An ist.
~~!§E!~b. Das tridiagonale Matrizenpaar
2 1 4 1
1 2 1 1 4 1
121 141
A= B =
1 2 1 4
der Ordnung n hat die Eigenwerte
Fur grosses n sind die Eigenwerte naherungsweise
Deshalb gilt nach (18) approximativ
Bei vorgegebener relativer Genauigkeit £ ist zu erwarten, dass

die Zahl der Iterationsschritte ~ linear mit n ansteigt. Im
spezialfall cc T = A wird die Konditionszahl des vorkonditio-
nierten Problems
so dass die Zahl der notigen Iterationsschritte unabhangig von

der Ordnung n
wird! Eine Testrechnung bestatigt diese qualita-
tive Aussage vollauf, denn fur E = 10- 6 sind die festgestellten
Iterationszahlen ~ ohne, bzw. ~ mit Vorkonditionierung in
Tab. 1 zusammengestellt.
Tab. 1 Konvergenzverhalten fur Testbeispiel
n = 10 20 40 80 160
~ = 30 46 87 164 314
-
~ = 8 7 7 7 7
236
Die Cholesky-Zerlegung der Matrix A kommt fur prak-

tische Zwecke nicht in Frage, falls die schwache Besetzung nicht
zerstort werden solI. Urn aber dennoch eine moglichst gute Vorkon-
ditionierung zu erzielen, muss C so gewahlt werden, dass cc T
eine Naherung fur A darstellt. Im folgenden sei C die Links-
dreiecksmatrix, welche bei der partiellen Cholesky-Zerlegung von
A resultiert, wobei jedes Fill-in unterdruckt wird, so dass C
die gleiche Besetzungsstruktur wie A aufweist (vgl. dazu [11]).
Es sind naturlich auch andere Varianten'anwendbar.
4. Der Algorithmus mit impliziter Vorkonditionierung

Im Algorithmus (13) kann das Matrizenpaar (A,B) for-
mal durch (A,B) und ~k durch Yk ersetzt werden. Es ist aber
vorteilhafter, die Vorkonditionierung implizit durchzufuhren. In
Anlehnung an das Vorgehen bei linearen Gleichungen [11] werden
einige Umformungnen vorgenommen und geeignete neue Vektoren ein-
gefuhrt. Dazu seien in (13) in Gedanken die auftretenden Vektoren
und Skalare durch ein versehen. Sodann bearbeiten wir die
einzelnen Zeilen von (13). Aus (20) und (21) folgen die Beziehun-
gen
-1
=c Yo
T::: T
P -vv
- "'-0-0 =-ox-ov
- - -- -
~ = 2(Ya - qYa)/p = C-1 {2(Ya --
- qYa)/p} = C-1 ~ •
Weiter fuhren wir folgende Vektoren ein:
(CCT)hk = .9:k !!k := Ahk ; (30)
C- 1 BC- TC- 1.9.k C- 1 B(CC T )-1.9.k = C- 1 (Bhk ) -lA

=: C !!k'
Damit gilt fur das Skalarprodukt wegen (30)

237
= ~k-1C
T -T -l Bh
C -k-1
= -k-1
hT Bh
-k-1
Aus der Definitionsgleichung fUr Ek erhalten wir
-T- -T- -T- -T -1 -T-
C Ek = -C gk-1 + £k-1 C Ek-1 -C C gk-1 + £k-1 C Ek-1
T -1 -T
-(CC) gk-1 + £k-1(C Ek-1)· ( 31)
Mit der Definition

-T- ( 32)
§.k := C Ek
ergibt sich aus (31) und wegen (30) die Relation
( 33 )
Weiter erkennt man aus den Definitionen
~k AEk C- 1 AC-TEk C- 1 (A§.k) =: C-1 ~k

(34)
~k = BEk = C- 1 BC-TEk = C- 1 (B§.k) =: c- L~k
wegen (33) und (30) die neuen Rekursionsformeln fUr
( 35)
FUr die vier Skalarprodukte resultieren nach (34) die Darstellun-

gen
T -1 T
~k_1CC ~k = ~k-1~k' = x T 'II
-k-1-k'
T -1 T =
T~
§.k CC ~k = §.k~k T §.k~k
Schliesslich resultieren aus den Rekursionsformeln fUr

und Yk nach entsprechenden Multiplikationen
-T -T ~ -T-
C Yk = ~k = C Yk-1 + 0k C Ek = ~k-1 + 0k§.k
-
= Yk = CYk-1 + °kC~k = Yk - 1 + °k~k
Yk = CYk-1 + 6kC~k v
= -k-1 + 6k~k
Der vorkonditionierte Algorithmus zur Minimierung des

Rayleighschen Quotienten mit der Methode der konjugierten Gra-
dienten lautet mit einer leicht geanderten Anordnung der Anwei-
238
sungen und nach einer Ersetzung von skalaren Grassen wie a

durch a wie folgt:
Wahl von e, so dass eeT ::: A

Start: Wahl von ~
;f= O·
-' Yo = A~, v = B~;
-0
T T-
a o ==
~Yo' Po = ~Yo; go = ao/po;
Iterationsschritt (k=1,2,3, ... l:
gk-1 = 2(Yk-1 - gk-1Yk-1 l / Pk-1

T
(ee lhk - 1 = gk-1;
falls k = 1: ~1 = -~,
falls k > 1: 7;
= -k-1-k-1'
hT 11 .
§.k == -h
-k-1 + E: k - 1§.k-1
~k == -u
-k-1 + E:k-1~k-1 (36)
~k = -11-k-1 + E: k-1-k-1
Vi
T T_
13 xT w
== -k-1-k' Y = §.k~k' x T Vi
a = -k-1-k' 1: == §.k~k;
~ und ok gemass (11) und (12) berechnen;
~k = ~k-1 + °k§.k'
Yk = Yk-1 + °k~k' Yk v
= -k-1 + °k~k;
T T_
a k = ~kYk' Pk = ~kYk; gk == ak/P k ; 7;a = 7;;
Test auf Konvergenz
Jeder Iterationsschritt des Algorithmus (36) erfordert

T
die Lasung des Gleichungssystems (ee )hk - 1 = gk-1 nach hk - 1
239
und die Multiplikation des Vektors hk - 1 mit den Matrizen A

und B. Infolge unserer Annahme Uber C entsprechen die drei
Operationen etwa drei Multiplikationen eines Vektors mit einer
der schwach besetzten Matrizen. Ferner werden sieben Skalarpro-
dukte und acht Multiplikationen von Vektoren mit einem Skalar
benotigt. Folglich ist der Rechenaufwand fUr einen Iterations-
schritt mit Vorkonditionierung praktisch gleich demjenigen ohne
Vorkonditionierung! Zudem ist die hohe Vektorisierbarkeit des
Verfahrens ersichtlich.
5. Berechnung der hoheren Eigenwerte

Es seien bereits die (~-1) kleinsten Eigenwerte A1 <
A2 ... ~ A~_l mit den B-normierten Eigenvektoren ~1' ~2'···'
~
~~-1 naherungsweise berechnet. Urn das nachste Eigenpaar (A~,~~)
als Minimum eines Rayleighschen Quotienten bestimmen zu konnen,
verwendet man eine Deflation mit partieller Spektralverschiebung.
Anstelle von A betrachten wir die Matrix [7]
I
~-1
A~ := A + (d - A )(Bz )(Bz )T.

v -v -v (37)
v=l
FUr die Eigenwerte und Eigenvektoren von A~~ = AB~ gilt wegen
der Orthonormiertheit der Eigenvektoren ~j von (1)
~-1
AnJC-]
z. = Az.
-] + 'L (d - A )(Bz )(Bz )T z .
v -v -v -]
v=l
=1 d(Bz.)
-]
fUr j = 1,2, ... ,~-1
A. (Bz.)
] -1
fUr j = ~,~+l, ... ,n
Die bekannten Eigenwerte Aj gehen Uber in den (~-l)-fachen
Eigenwert d, wahrend die Ubrigen unverandert bleiben. Damit A~
der kleinste Eigenwert von A~~ = AB~ ist, muss d > A~ sein.
Unter BerUcksichtigung der Darstellung der Konditionszahl (18)
ist es besser, d > A~+l zu wahlen. Die Erfahrung lehrt ander-
seits, dass d auch nicht zu gross gewahlt werden darf, da sonst
240
die Vorkonditionierung an Wirkung verliert, denn sie wird ja mit

der festen Matrix C fUr das Paar (A~,B) angewandt.
Im Algorithmus (36) ist A durch A~ zu ersetzen. Zur
effizienten AusfUhrung von A~~ = Ax + L(d - Av )(Bz
-v
)(Bz )Tx
-v -
sind zweckmassigerweise die Vektoren (Bz),
-v (v=1,2, •.. ,~-1) ab-
zuspeichern. Der Rechenaufwand pro Iterationsschritt erhoht sich
urn je (~-1) Skalarprodukte und Multiplikationen von Vektoren
mit Skalaren.
6. Beispiele
Die Effizienz der Rayleigh-Quotient-Minimierung mit
Vorkonditionierung solI an drei Beispielen aus verschiedenen In-
genieurbereichen dargelegt werden. Die Steuerung der Verschiebung
d erfolgte so, dass sie zu Beginn der Berechnung von A~ (~ > 1)
mindestens gleich einem Vielfachen f von A~_l ist. Der Faktor
f wird vom Startwert 2 sukzessive verkleinert bis auf etwa
1.5, urn dann konstant gehalten zu werden .. Zudem wird nach zehn
Iterationsschritten geprUft, ob der momentane Rayleighsche QUo-
tient q die Bedingung fq < d erfUllt. Andernfalls wird d
mit f multipliziert, f allenfalls verkleinert und die Itera-
tion neu gestartet. Die Iteration wird gestoppt, sobald sich der
Rayleighsche Quotient urn weniger als 10- 14 andert bei skalierten
Matrizen A und B, fUr die a ii = 1 und max(b ii ) = 1 gilt.
Das Abbruchkriteriurn entspricht einer absoluten Toleranz.
~~!~E!~!_!. Die Berechnung der akustischen Eigenfre-

quenzen eines Autoinnenraurnes fUhrt auf die Eigenwertaufgabe
llu + AU =0 in G
unter Neurnannschen Randbedingungen
au
an
=0 auf aG,
wo G das Gebiet eines Langsschnittes eines Autos ist (Fig. 1).

FUr die Triangulierung der Fig. 1 in 525 Dreieckelemente mit ku-
bischen Ansatzen zu neun Freiheitsgraden [11] ergibt sich eine
Eigenwertaufgabe (1) der Ordnung n = 925. Da die Matrix A singu-
241
lar ist, muss eine Regularisierung vermittels einer Spektralver-

schiebung vorgenommen werden:
(A + cB)~ = (A + c)B~, c = 0.0001.
Fig. 1 Autolangsschnitt mit Triangulierung
In Tab. 2 ist der Rechenablauf zur Berechnung der 18

kleinsten Eigenwerte dargestellt. Es bedeuten Ai die Eigen-
werte, d die verwendeten Verschiebungen und ~ die Zahl der
Iterationsschritte.
Tab. 2 Rechenablauf fur Autoproblem
A.:!: 0 0.012666 0.044374 0.056555, 0.116217 0.137304

l.
d :!: - 0.050 0.100 0.174 0.271 0.271
~ = 57 36 64 37 48 48
A.:!: 0.142429 0.200007 0.270795 0.290647 0.345755 0.374871

l.
d :!: 0.271 0.387 0.387 0.387 0.514 0.514
~ = 29 32 47 37 58 80
A.:!: 0.385334 0.396457 0.465918 0.559286 0.615272 0.682350

l.
d :!: 0.514 0.514 0.645 0.774 0.774 0.895
~ = 55 33 37 42 52 42
242
Die Vorkonditionierung reduziert den Rechenaufwand etwa

auf den vierten Teil im Vergleich zur Methode der konjugierten
Gradienten ohne Vorkonditionierung, fUr welche die Iterations-
zahlen fUr die ersten fUnf Eigenwerte ~1 = 277, ~2 = 170, ~3 =
251, ~4 = 141, ~5 = 225 betragen. Mit A915 845.84 und =
K2,B(H(~1)) =
6.678.10 4 resultiert nach (19) mit £ = 10- 6 der
zu grosse Schatzwert ~S ~ 1800. Die Rechenzeit ist vergleichbar
mit derjenigen des Lanczos-Verfahrens, jedoch bei bedeutend
kleinerem Speicherbedarf [10].
Eine rechteckige Platte der Lange L = 8m

~~!~E!~!_g.
und der Breite B = 4m ist in 24 x 12 = 288 Quadratelemente
eingeteilt, so dass 325 Knotenpunkte resultieren. Zur Berechnung
der Eigenfrequenzen der Platte sind in den Elementen die nicht-
konformen kubischen Ansatze [11] verwendet worden. Die Platte
sei am linken Rand eingespannt, am rechten Rand gelenkig gelagert
und am Ubrigen Rand frei. Die Materialkonstanten sind E = 40(!),
v = 0.1667, p = 8.25, und die Dicke der Platte ist h = 0.15m.
Nach BerUcksichtigung der Randbedingungen ergibt sich eine
Eigenwertaufgabe (1) der Ordnung n = 910, fUr welche die ersten
20 Eigenwerte berechnet wurden (vgl Tab. 3).
Tab. 3 Rechenablauf fUr rechteckige Platte
L= ~
0.66281 3.1460 6.9871 16.398 30.509 33.561 50.887
d= - 7.885 12.07 35.74 51. 31 51. 31 73.69
~ = 87 98 73 68 96 43 53
A.= ~
69.404 89.411 124.61 139.67 195.17 208.53 260.90
d = 105.8 151. 9 218.2 218.2 313.3 313.3 449.9
~ = 49 45 49 33 69 40 172
A.= ~
262.23 267.48 397.13 419.41 456.66 495.09
d = 449.9 449.9 646.0 646.0 927.6 927.6
~ = 82 31 77 65 90 39
Eigenwertaufgaben fUr die Analyse von Plattenschwingun-
gen besitzen bekanntlich schlechte Konditionen. In der Tat ist
243
mit A910 ~ 4.852.10 6 die Konditionszahl K2,B(H(~1»~1.954.106,

so dass der Schatzwert (19) fUr ~ S ~ 10 4 eine ausserst langsame
Konvergenz erwarten lasst. Diese Zahl ist etwas Uberschatzt, doch
bringt die Vorkonditionierung in diesem Beispiel eine ganz ent-
scheidende Beschleunigung der Konvergenz.
~~!~E!~!_~. Gesucht seien die 20 kleinsten Eigenfre-

quenzen einer Eisenbalkenkonstruktion in der Gestalt eines neun-
stockigen Hochhauses mit den Abmessungen nach Fig. 2. Die Balken
haben quadratischen Querschnitt mit einer Seitenlange von 8cm.
Die Materialkonstanten sind E = 2.10 11 , v = 0.3, p = 8.25.10 3 .
Die Ordnung des Eigenwertproblems ist n = 648.
L ./ ./ ./
L ./ ./ ./
L ./ .. / ./
L / / 7
L / ./ :/
.t:. ./ 1/ V
l..::::
L / '" ./
/
"/
7
IL
~
/ '" /
1/
7
7
L ./ / 17
L / ./ V
L ./ / ./
L /' / ./
L /
L ./
V 7
./ :7
L ./
'L /
./ 7
./ i7
......_ _ _ ..c5/V't1 ____ ;-:Olffl

Fig. 2 Hochhauskonstruktion
244
Tab. 4 Rechenablauf fUr Hochhauskonstruktion
A.=
l.
1.0135 1.1124 1. 4330 10.808 11.716 14.533 23.982
d = - 4.054 7.307 18.46 26.50 36.04 46.81
II = 232 263 96 217 149 86 79
A.=
l.
36.883 38.166 39.333 41.920 49.604 53.663 71. 804
d = 58.46 58.46 58.46 70.62 70.62 82.91 95.04
II = 209 180 118 82 99 64 169
A.=
l.
72.066 90.348 91. 029 92.183 103.96 109.20
d = 106.7 117.8 128.1 128.1 128.1 128.1
II = 72 253 107 75 87 195
Die in diesem Beispiel beobachteten, teilweise recht

hohe Zahl von Iterationsschritten ist durch die eng benachbarten
Eigenwerte bedingt, welche eine sehr grosse Konditionszahl der
Hesseschen Matrix zur Folge haben. Es ist A648 = 9.650.10 6 , und
somit sind K2,B(H(~1)) = 9.76.10 7 und llS ~ 71'600. Die Vorkon-
ditionierung bringt auch hier entscheidende Verbesserungen.
Literaturangaben
[1] Axelsson, 0.: Solution of linear systems of equations:

Iterative methods. In Barker, V.A. (ed.): Sparse matrix
techniques, Lecture Notes Math. 572, springer, Berlin-
Heidelberg-New York, 1977, 1-51.
[2] Bathe, K.-J.: Finite-Elemente-Methoden. Springer, Berlin-
Heidelberg-New York-Tokyo, 1986.
[3] Bradbury, W.W. and Fletcher, R.: New iterative methods for
the solution of the eigenproblem. Numer. Math. ~ (1966),
259-267.
[4] Hestenes, M.R. and Karush, W.: Solutions of Ax = ABx.
J. Res. Nat. Bur. Standards 49 (1951), 471-478.
[5] Longsine, D.E. and McCormick, S.F.: Simultaneous Rayleigh
quotient minimization methods for Ax = ABx. Lin. Alg. Appl.
11 (1980), 195-234.
[6] Parlett, B.N.: The symmetric eigenvalue problem. Englewood
Cliffs, Prentice-Hall, 1980.
[7] Perdon, A. and Gambolati, G.: Extreme eigenvalues of large
sparse matrices by Rayleigh quotient and modified conjugate
gradients. Comp.Meth.Appl.Mech.Eng. 56 (1986), 251-264.
245
[8] Ruhe, A.: Computation of eigenvalues and eigenvectors. In

Barker, V.A. (ed.): Sparse matrix techniques, Lecture Notes
Math. 572, Springer, Berlin-Heidelberg-New York, 1977,
130-184.
[9] Schwarz, H.R.: Simultane Iterationsverfahren fUr grosse
allgemeine Eigenwertprobleme. Ingenieur-Archiv 50 (1981),
329-338. ---
[10] Schwarz, H.R.: Eine Variante des Lanczos-Verfahrens. ISNM
69 (1983), 161-180.
[11] Schwarz, H.R.: Methode der finiten Elemente, 2. Aufl.,
Teubner, Stuttgart, 1984.
[12] Waldvogel, P.: Numerische Behandlung von allgemeinen Eigen-
wertproblemen. Dissertation, Universitat ZUrich, 1985.
Prof. Dr. H.R. Schwarz

Institut fUr angew. Mathematik
Universitat ZUrich
Ramistrasse 74
CH - 8001 Z Uri c h
DATA FITTING BY SUMS OF EXPONENTIALS

USING THE ~l NORM
G. Alistair Watson
Department of Mathematical Sciences
University of Dundee
1. Introduction
This paper is concerned with some aspects of the problem of

fitting data (x., y.) , i = l,2, ... ,m, by a sum of exponentials
1. 1.
n -A lC
j
1: a e Most attention has been given to the minimization of the least
j=l j
m .th
squares norm of !:;, where!:; E: R h as 1. component
n -A x
j i
r i = Yi - 1: a.e i 1,2, ... ,m, ( 1.1)
j=l J
and a number of algorithms have been studied. The problem is widely

recognised as being difficult computationally, and it is often the case that
widely varying parameter values can give nearly optimal results. Of interest
here is the calculation of parameters by minimizing the ~1 norm of!:, a
criterion which is particularly relevant if there are gross errors or wild
points in the data as it is not sensitive to these. The intention is to
consider some computational aspects of this problem, and also to examine
questions of parameter sensitivity primarily by means of some examples. In
what follows, unadorned norm symbols will denote ~1 norms.
m
For A E: R, define ~(A) E: R by
T
e(A) = (exp(-Ax ), .... , exp(-Ax »
~ 1 m
Then if a* E: Rn, A* E: Rn minimize 1 Irl I, it follows from standard analysis
~ ~ T ~ *
that there exists f(A) = ~ ~(A) with ~ E: 81 I!:; II such that
0, j 1,2, .. ,n, (1.2)

247
1,2, •• ,n,
° if j (1.3)
where the superscript * on ! implies evaluation at ~ * ,~ * ,and the dash

denotes differentiation with respect to A. The usual notation a is used
for the subdifferential of a convex function; in this particular case
aII! II
The above conditions define ~*, ~* to be a stationary point of the 11

problem. It is clear that ~*, ~* will remain a stationary point if, for
any i for which r: f 0, perturbations are introduced into Yi which do not
change sign (r i ).
2. The positive coefficient case
An important special case of the above problem arises when the

coefficients a. are constrained to be positive and the exponents are
J
distinct and lie in an interval [a,B] of the real line: some applications
are given in [15]. In this case, n need not be specified in advance, but
may be regarded as a parameter of the problem, which may therefore be
stated as
find ~ E: K to minimize Ill: - ~II (2.1)
where K is the convex cone in Rm generated by ~(A)' A E: [a,S]. The

existence of a best approximation is now guaranteed, and necessary and
sufficient conditions for a*, A* to be a solution are that there exists
f(A) = ! T~(A) , ! E: al Ir* I -I such
- that
0, j 1,2, •• ,n, (2.2)
f(A) ~ 0, for all A E: [a,6], (2.3)
(see for example [19]). It follows from (2.3) that

248
f 11 (A~) 0, A* E: (a,S), (2.4)

J j
showing the connection with (1.3). An algorithm for (2.1) is given in [19]
based on earlier work by Ruhe [15] and Gustafson [6]. It consists of two
phases, the second of which is the efficient solution of the system of
equations (2.2), (2.4) by a method which is asymptotically Newton's method.
The first phase determines n and the structure of al Ir*1 I , and is based
on initially solving the constrained linear R.l problem
m t -A.X.
min Ely. E a.e J ~I
a~O i=l ~ j=l J (2.5)
for fixed A., j 1,2, .. ,t in [a,S]. This may be formulated as the linear
-- J
programming problem
subject to
[I -I (2.6)
!;!, y, ~ ~ Q,
T
where E: (1,1, ... ,1) and E(~) is the mxt matrix
[e(AI),···,e(A t )]
~ ~
At an optimal basic feasible solution, let ~ denote the simplex multiplier

vector. Then the optimality conditions correspond to
T
II!! 1100 :$ I , 11
~
e(A)
~ J
:$ 0, j 1,2, .. , t,
with equality holding for j such that e(A.) is in the basis matrix.
~ J
It is easily verified that ~ E: al lEI I , so that if additionally
T
!! ~(A) :$ 0 for all A E: [a,S], the original problem is solved.
249
T -
Otherwise it is possible to identify some A E [a.B] with ~ ~(A) > 0 and a
new column ~(A) may be brought into the basis matrix by a simplex step.
Optimality may be regained for the new (enlarged) problem and we may
continue in this way to make systematic progress towards the solution until
a stage is reached where the second phase may be entered; for full details
of an algorithm. see [19].
3. The general case
The assumptions of the previous section introduce certain pleasing

features into the theory and computation of best approximations. In the
general case. however. as is well known. there may be local minima of the
problem; in addition existence is no longer guaranteed.
Example 1
Let (x i •y i ) = (i - l.i). i = 1.2 •...• m ; n 2.
It is easily verified that if A2 = O. a l + a 2 = 1.
a l = -l/A l • Al .... 0 then IIEII .... O.
Therefore the best approximation (for any norm) does not exist.
Example 2
Now fix m 5. n 2. with data as in Example 1. and define the
function
min Ilrll
e
Then computation of values of I shows that descent towards a point
satisfying (1.2). (1.3) is achieved by allowing Al and A2 to coalesce.
For example
1(-0.5849. -0.5810) 0.419 013.
with a l =-53.261 725. a2 54.588 529.

250
In fact it may be verified that a local minimum corresponds to the

confluent case when Al = A = -0.5829348 (the common value A* satisfies
* 2
A = ~n(x), where x is a root of 5x 3 - 6x 2 + 1 = 0) and so does not exist
for the original problem. The sum of 2 exponentials degenerates to the
expression
(3.1)
and the optimal coefficient values are
1. 326809, a2 -0.2102939,
with I lEI I = 0.419 006. The first order necessary conditions for
approximation by the function (3.1) correspond to the existence of
f(A) = 'fT~(A) , 'f e; a liE * II such that
the graph of f(A) is shown in Fig. 1.

Assuming existence, a second phase procedure virtually identical
to that used for the problem of the previous section may be applied. In
the absence of a first phase (or if only poor quality approximations are
available at the end of a first phase) more robust methods may be used,
for example one of the modern methods available for general nonlinear ~l
approximation ([1], [3], [4], [7], [10]). However the nature of the
problem is such that a good initial approximation is important, and a
natural question to ask is whether a first phase procedure similar to that
described in the previous section may be developed. Assume that an interval
* A* , ... ,A *
[a,S] of the realline is i~entified in which the exponents AI' 2 n
lie. Let t > n and ~ e; Rt be a vector whose components are distinct
values lying in [a,S]. Then a problem analogous to (2.6) is that problem
with the restriction that at most n columns of E(~) can be present in the
optimal basis matrix. An appropriate formulation is
251
subject to
l~l ,
[I -I E(J,.) ] }; (3.2)
-o.K
1-
:> a.
1-
~ o.K,
1-
i 1,2, •. , t,
&T§ n ,
o or 1, i 1,2, •• ,t
where K is a constant such that K ~ max la~l. Thus the discretized

1-
l:>i:>t
problem (with values of J,. restricted to a finite set) can be posed as a
zero-one linear programming problem. This problem is finite, but i f t is
large, it can be computationally very expensive. In addition, from an
optimal solution, there is no obvious way in which new J,. values can be
introduced in a satisfactory manner so as to make progress towards the
solution of the original problem.
Another possible approach exploits the fact that the exponential
sum (or a degenerate form) is the solution of a linear constant coefficient
ordinary differential equation, and thus a reparameterization of the problem
can be effected. For the approximation of an n-times continuously
differentiable function y(x), Robitzsch and Schaback [14] suggest a method
for the calculation of a suboptimal approximation (essentially a first phase
procedure). See also Schaback [17]. However, whether an approach of this
kind can be exploited in the present situation remains to be established.
Finally, we remark that the fact that thevariables ~ and ~ readily
separate has been used in the 12 case: for example the methods of [5], [8],
[12], [16] may all be applied. For norms which are not smooth, exploiting
separability in the same sort of way is no longer straightforward.
Certainly if ~ € Rn is given, it is easy to compute ~ minimizing I Irl I. It
is therefore attractive to consider the possibility of an iterative procedure
in which ~ € Rn is modified at each step, and the corresponding ~ is then
252
calculated. However possible limitations of this kind of iteration are

illustrated by the following example.
Example 3
m 5, n 1.
o 0.05 0.1 0.15 0.2
2.51 2.04 1. 67 1.37 1.12
Let Al = 3.997474 , a l 2.491347 with IIEII 0.021276. Then it may be

verified that
min min
a A
However, this is not a minimum of I lEI I, which occurs for

Al = 4.036480, a l = 2.51 with Ilrll = 0.018029.
4. Sensitivity
In this section we examine questions of sensitivity of best ~l
approximations, and illustrate typical situations by means of some examples.

The robustness of the ~l solution with respect to data changes has already
been remarked on, and we are concerned here with the effect on I Irl I of
parameter inaccuracies. Let ~* c Rn , ~* c Rn solve the problem, and let
s* = {i : r~
1.
o}, e~1. sign (r *)
i , 1.' ..L S*.
Then if ~
-1<'
= rl~:l
~ ,and d c R2n is arbitrary, it is readily seen by direct
calculation that if y > 0 is small enough
253
where ~ to Clllr*11 satisfies (1.2), (1.3). If IS*I < 2n, then there exists
2 ~ Q such that the coefficient of y is zero, and so the sensitivity of
I lEI I with respect to changesof the parameters from their optimal values in
these directions depends on second derivatives. This is the situation in
differentiable cases, for example approximation with respect to the £2
norm, when the eigenvalues of the Hessian matrix of I lEI I can be used
as a measure of local sensitivity (see, for example Varah [18]). For
other directions (and all directions if lsi ~ n) then the coefficient of
y is positive and sensitivity depends primarily on the condition of the
matrix with columns Vr~, i to S*
~
From the point of view of exponent sensitivity, a readily

available measure is provided by the size of the quantities
f"(A~) , j = 1,2, .• ,n. These numbers play an important role in the £2
J
analysis of Varah [18] (where ~ is a .
mult~ple of E* ): see also [2], [13].
We illustrate by taking 2 well-known data sets considered also by
Varah [18], and due to Lanczos [9] (m = 24) and Osborne [11] (m = 33).
Example 4
First of all, we fit a sum of 2 exponentials to the first 5
data points of the Lanczos data (these are the data used in Example 3).
The solution is such that S* = {1,2,3,5} with A; 3.9947407,
A2* = 63.503704 , a *l = 2.4899852 , a *2 = 0.0200148 , and I IE * II = 0.002387.
An indication of the behaviour of I Irl I is given by the 3-dimensional plot
of 1(~) shown in Fig. 2, for the range 3.5 ~ Al ~ 4.4, 35 ~ A2 ~ 80; the
vertical scale is such that 1(~*) ~ 1(~) ~ 1(3.5,80) = 0.136757. For
this example, f"(A~) = -0.001178 , f"(A~) = -0.000003, and the graph of
f(A) is shown in Fig. 3. The insensitivity of I Irl I to A2 is further
illustrated by the fact that there is a 'local minimum' for this problem
obtained by letting A2 + 00; it is given by Al = 3.997474, al = 0.018653,
a2 = 2.491347, IIEII 0.002623 (cf. Example 3).
Example 5
For the full Lanczos data set (and n = 2) the solution is given by
Al*= 1. 796832, A*2 = 4.564162, a *1 = 0.398913, a *2 = 2.111087,
IIE*II = 0.040423. A plot of 1(~) is shown in Fig. 4 for the range
254
0.5
0.4
0.3
0.2
0.1
-1.0 -0.8 -0.6 -0.2
_0.1
Figure
">
"
Figure 2
255
0.004
0.002
0.000
0 20 40 60
-0.002
-0.004
-0.006
-0.008
-0.010
-0.012
-0.014
-0.016
-0.018
-0.020
Figure 3
Figure 4
256
1.4 ~ Al ~ 2.3, 4.1 ~ A2 ~ 5.0, with vertical scale such that

I(A * ) ~ I(~) ~ For this example, fll(A*)
1(1.4,5.0) = 0.464200. 0.012095,
1
fll(A *
2 ) = 0.003844 and the graph of f(A) is shown in Fig. 5. For
comparison the graph of f(A) for the solution of this problem with
positive coefficien~in [0,10] is given in Fig. 6 : here n = 3.
Example 6
For the Osborne data set, and n=2, the solution is given by
Al* = 1.274774, A2* 2.246652, a *l = 1.868593, a *
2 = -1.399993,
III*II = 0.035117. A plot of I(A) is shown in Fig. 7 for the range
0.9 ~ Al ~ 1.8, 1.9 ~ A2 ~ 2.8 with vertical scale such that
I(~ *) ~ I(~) ~ 1(1.8,2.8) = 1.424885. * = 0.010756,
Here fll(Al)
f "( A2*) = 0.001996, and the graph of f(A) is shown in Fig. 8. For
comparison" the graph of f(A) for the solution of this problem with
positive coefficients is shown in Fig. 9 here n = 1, with Al 0.847610.
For both these data sets, I S* I 2n, and so locally a
satisfactory performance is achieved by methods which do not require
second derivative information, for example those of Gauss-Newton or
Levenberg-Marquardt type. In general, however, second derivative
information will be necessary for a good local convergence rate.
5. Concluding Remarks
A number of examples have been given to demonstrate various

aspects of the calculation of best ~l approximations by sums of
exponentials. In particular, two well known data sets, neither regarded
as pathological, have been used to illustrate the sensitivity of IIIII
to changes in exponent values. The examples exhibit the steep sided
valley characteristics which are typical of exponential fitting problems,
and more generally of difficult minimization problems, and the
availability of good starting values is important. For the positive
coefficient case, it is possible to make good use of the problem structure
in the calculation of best approximations, and in particular good
starting values for input to a fast locally convergent method can readily
be obtained. For the general case, the provision of a satisfactory
257
0.008
0.006
0.004
0.002
2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
Figure 5
-0.0005 -
-0.0010
-0.0015
-0.0020
-0.0025
-0.0030
-0.0035
-0.0040
-0.0045
-0.0050
-0.0055
-0.0060
Figure 6
258
Figure 7
0.0010
O. 0008
0.0006
0.0004
~L~L
0.0002 -
0.0000
1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0
Figure ,8
259
10
~ 1
~3
~4
Figure 9
first phase process remains an open problem, and it has yet to be

established whether special methods, tailored for the problem, can
be successfully developed.
References
1. Bartels, R.H. and Conn, A.R. (1982) An approach to nonlinear

~l data fitting, in Hennart, J.P. (ed.), Numerical Analysis,
Cocoyoc 1981, Springer Verlag, pp.45-58.
2. Bates, D.M. and Watts, D.G. (1980) Relative curvature measures

of nonlinearity, J. Royal Stat. Soc., Ser. B, 42, pp.1-25.
3. Fletcher, R. (1981) Numerical experiments with an exact ~1

penalty function method, in Mangasarian, O.L., Meyer, R.R.
and Robinson, S.M. (eds.), Nonlinear Programming 4, Academic
Press, New York, pp.99-129.
4. Fletcher, R. (1985) An ~l penalty method for nonlinear constraints,
in Boggs, P.T.,Byrd, R.H. and Schnabel, R.B. (Eds), Numerical
Optimization 1984, SIAM Publications, Philadelphia, pp.26-40.
5. Golub, G.H. and Pereyra, V. (1973) The differentiation of

pseudoinverses and nonlinear least squares problems whose variables
separate, SIAM J. Num. Anal. lQ, pp.4l3-432.
6. Gustafson, S-A (1981) A computational scheme for exponential

approximation, ZAMM~, pp.284-287.
7. Hald, J. and Madsen, K. (1985) Combined LP and quasi-Newton methods

for nonlinear ~l optimization, SIAM J. Num. Anal. 22, pp.68-80.
8. Kaufman, L. (1978) A program for solving separable nonlinear least

squarffiproblems, Bell Labs. Tech. Memo. 78-1274-7.
9. Lanczos, C. (1956) Applied Analysis, Prentice Hall, Englewood

Cliffs, N.J.
10. Murray, W. and Overton, M.L. (1981) A projected Lagrangian algorithm

for nonlinear ~l optimization, SIAM J. Sci. Stat. Compo ~,
pp.207-224.
11. Osborne, M.R. (1972) Some aspects of nonlinear least squares

calculations, in Lootsma, F.A. (ed.), Numerical Methods for Nonlinear
Optimization, Academic Press, New York.
12. Osborne, M.R. (1975) Some special nonlinear least squares problems ,
SIAM J. Num. Anal. ~, pp.57l-592.
13. Ramsin, H. and Wedin, P-A. (1977) A comparison of some algorithms

for the nonlinear least squares problem, BIT ll, pp.72-90.
14. Robitzsch, H. and Schaback, R. (1978) Die numerische Berechnung von

Startnaherungen bei der Exponentialapproximation, in Collatz, L.,
Meinardus, G. and Werner, H. (eds.), ISNM 42, Numerische Methoden
der Approximationstheorie IV, Birkhauser, Basel-Stuttgart.
15. Ruhe, A. (1980) Fitting empirical data by positive sums of

exponentials, SIAM J. Sci. Stat. Compo !, pp. 481-498.
16. Ruhe, A. and Wedin, P-A. (1980) Algorithms for separable nonlinear
least squares problems, SIAM Rev. 22, pp. 318-339.
17. Schaback, R. (1979) Suboptimal exponential approximation, SIAM J.

Num. Anal. ~, pp.1007-l0l8.
18. Varah, J.M. (1985) On fitting exponentials by nonlinear least

squares, SIAM J. Sci. Stat. Compo ~, pp.30-44.
261
19. Watson, G.A. (1986) Data fitting by positive sums of exponentials,

in Cox, M.G. and Mason, J.C. (eds.), Algorithms for the Approximation
of Functions and Data, Proceedings Shrivenham, 1985, Oxford
University Press.
Dr G.A. Watson, Department of Mathematical Sciences, University of Dundee,

Dundee, DDl 4HN, Scotland.

(International Series of Numerical Mathematics

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(International Series of Numerical Mathematics

Uploaded by

Copyright:

Available Formats

ISNM81:

International Series of Numerical Mathematics

Numerische Methoden der

1987 Springer Basel AG

CIP-Kurztitelaufnahme der Deutscben Bibliothek

© Springer Basel AG 1987

ISBN 978-3-0348-6657-6 ISBN 978-3-0348-6656-9 (eBook)

Die Herausgeber dieses symposiumbandes gedenken ihres Kollegen

Professor Dr.rer.nat. Helmut Werner,

Wir widmen daher den vorliegenden Band seinem Andenken.

Lothar Collatz GUnther Meinardus GUnther NUrnberger

This international conference was convened under the joint

The relaxed and pleasant working atmosphere facilitated inten-

Lothar Collatz Gunther Meinardus Gunther Nurnberger

Die internationale Tagung stand unter der Leitung von L.

Die freie und harmonisch Atmosphare fuhrte zu intensiven wis-

Lothar Collatz Gunther Meinardus Gunther Nurnberger

BAKER, T.H.C. - DERAKHSHAN, Mir S., Computed Approximations to

COMPUTED APPROXIMATIONS TO SOME POWER SERIES

Christopher T H Baker and Mir S Derakhshan

Department of Mathematics, The University of Manchester.

1. The Problem and its Applications

where (for all n > k)

The 'natural' periodic extension of {ar}~;;- ~ of period M = 2N denotes the

We shall use a non -standard variant of this result later.

Direct Di rect FFT

128 0.001 0.001 0.036

256 0.002 0.002 0.068

512 0.003 0.004 0.143

Table I (Timings in seconds on the CDC 7600)

4. Iterations for the Computation of {~}.

Implicit in our application of the FFT in §3 is the observation that the

then d r + 1](p.) = [n(P.)hN where N = 2r , for r = 0,1,2,3, ....

The precise description of an algorithm requires the use of a programming

[1If(J.I)]m. where either n = m or n = ~m. evaluate the Newton improvement

for i := m ~ 1 12 2m-1 do Qj := - Pi;

compute the first 2R +, coefficients nr of neJL) = {a o + a, JL + a 2 JL 2 + ... OkJLk) -1·

in place of this comment. the following code (cf. [1]):

N HAIRER et al. [1] Modified Version HERON [including

128 0.034 0.018 0.047

Table IT (Comparative timings. in seconds. of codes implemented on the CDC7600)

The above descriptions of algorithms allow us to make a count of various

A Continuous Selection of the Metric Projection

H. BERENS AND M. FINZEL

It is well· known that

then A~ is well·deflned and e nxn 3 A H A~ is a continuous, sunny selection of ooPR. ••

A = S + Un W and W· S + S·W is positive definite,

Un being the smallest singular value of A.

Vz,y E E (y , z)s = 1->0+

where (.,.) : E· X E -+ R is the bilinear form on E· X E and 4l : E -+ E· the duality map, or

~. The proof of Theorem A

respectively, such that X = US and Y = VT and

S = diag(W1 ,. • • ,W.,So) and T = diag(Wlo' .. ,W.,To),

where So,To, and WI, 1 ~ l ~ r, are unitary matrices in cmoxmo, mo + ranle(A) = m, in

If m = n and if A is nonsingular then UV· = XY·.

values of a matrix A E c nx n are continuous functionals on C nx n, the selection is continuous.

As a consequence of Theorem A we have

being endowed with the spectral norm.

S. The semi-inner product

where U1:V· is the SVD of A.

{We c nxn :!R (W,A) = 1 and IIWIlI = I}.

Because of the invariance of the sesqui·linear form w. r. t. similarity, it suffices to characterize

Lemma. Let 1: = diag(u1!'''' Un), 1= Ul = ... = Up > Up+l ~ ... ~ Un ~ 0.

T=diag(T1! ... ,Tp),Tl~ ... Tp~O, and LTj=I}.

a; +k j = _1_ [ 4f t +k-l j - Ai +k_ 1a; +k-2 j - Bj +1a; +k-l j+l