Professional Documents
Culture Documents
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
Edited by
Herausgegeben von
L. CoUatz, Hamburg
G. Meinardus, Mannheim
G. Niirnberger, Erlangen-Niimberg
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.
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.
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
°
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
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».
and when r :. 227 for k = 6, Wr agrees with its limiting value limr~ Wr to
s. PRACTICAL ALGORITHMS.
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}.
6. CONCLUSIONS
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.
International Series of
Numerical Mathematics, Vol. 81
© 1987 Birkhiiuser Verlag Basel
in Matrix Spaces
by
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
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-
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
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
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
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 ,
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
\
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
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
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,
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
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
where
Moreover,
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). •
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
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}.
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.
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.
International Series of
Numerical Mathematics, Vol. 81
© 1987 Birkhauser Verlag Basel
C. K. Chui 1 and T. X. He 2
Texas A&M University
Department of Mathematics
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
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
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
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
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
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 •
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
( 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
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
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
- 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,
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:
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
.
( 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
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:
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.
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)
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,
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.
b) Ein anderes Beispiel ist die stromung urn eine Ecke, Fig. 5,
vgl. Collatz [gO].
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.
(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.).
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.
Franz-Jurgen De1vos
o. Introduction
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
ke2Z) •
j e 2Z )
n-1
L g(t-t,) eiktj (O<k<n)
j =0 J
and assume
( 0 < k < n )
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
(1 .2 )
°0 ,J. O:;;j<n)
projector Sn on C2TI
n-l
(1 • 3 ) L f(t.) M(t-t.)
j =0 J J
thearticles [4,5,6,8J.
2. Periodic Histopolation
tionals
t.J+ 1
a. (f)
J
f f(t) dt (O:;;j<n)
tj
conditions
o :ii j < n
dic histopolation.
Theorem 1
polation condition
(O:s;j<n)
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
(2.3) u(t)
Lemma
(2.4) a. (u)
]
= a. (f)
]
(O::ij<n)
(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
Lemma 2
space V~ (h) •
which implies
o :0 k < n )
n-1
(2.7) u(t) cOIf) + L J(f) (to) N(t-to)
j =1 J J
Lemma 3
n-1
1
L
0
Then we have
(2.9) °0 ,J 0
(O~j<n)
cO(P)
1
J(P)(to) = 1 - i ( 0 < j < n )
21T J n
n-1
(2.10) L a. (f)P(t-t.)
j =0 J J
O:;;j<n)
3. Trigonometric Histopolation
( 3 .2) T
m
Corollary 1
For any f e L2~ there is a unique trigonometric polyno-
conditions
tj +1 tj +1
( 3.3 ) fH(f)(t)dt f f(t) dt (O:oj<n)
t. n t.
J J
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
Corollary 2
conditions
(3•7 ) (O:oj<n)
62
(4.2) Sp(q-1,n,a)
(4.3) B (t)
q
L (ik) -q exp (ikt)
k~O
This implies
( see also [3, 10J ) that the assumptions for 9 and h = g'
Corollary 3
matching conditions
tj +1
(4.6) f H (f)(t) dt ( 0 ;:;; j < n )
t. n
J
63
(4.10)
( b > 0 )
64
References
Franz-Jurgen Del v 0 s
Lehrstuhl fur Mathematik I
Universitat Siegen
Holderlin-Str. 3
D-5900 Siegen (w.-Germany)
International Series of
Numerical Mathematics, Vol. 81
©1987 Birkhauser Verlag Basel
1. Introduction
2. Metric Selections
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
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.
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
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.
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
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
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.
(2) There exists a closed subset N of ker PM such that (QI N)-l is
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) ~ _.)
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
7. References
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).
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
n = 1.2.3•......
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.
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
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
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),
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
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 .
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
, 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
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
-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
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
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" .
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
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.
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
I - li-j-Ilh ~ j-k+1
{
I - 1i-j+2klh , j-k
f(si,ti+k)=I-kh
tk I-kh
Figure 4
1;{S,tJ-- 1
2
T Figure 5
2
S
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,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) &
j=i+k
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
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
Then
Thus again -I '" z&(si,tj) - W&(si,9 '" 1. Finally, we examine > k, '" k. In this
case
93
°, 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
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.
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.
M. v. Golitschek
Institut ftir Angewandte Mathematik und Statistik
der Universitat Wtirzburg
F. Schardt
Medizinische Poliklinik
der Universitat Wtirzburg
M. Wiegand
Medizinische Poliklinik
der Universitat Wtirzburg
1. Einleitung
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
n+2
(2. 1 ) u ( t) = B c k Uk ( t )
k=o
m
2
(2.2) B (y r - u ( tr) ) + dt
r=l
m
2
L) (Y
r - f ( tr» +
r=1
m n+2
F(c ,oo.,c 2) L) (y r - L)
o n+
r=1 k=o
(J.1 ) (B + ~ C)c = b
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
n+2
(3. 2 ) u*(t) = B c~ s( (t - xk)/h)
k=o
Beispiel: Atemminutenvolumen
~~
. #.
I ',I
o 2.000
101
(iii) Knickpunkte
bestimmt werden.
oder falls
lu*"(Xk-l ) I
Literatur
David Handscomb
certain subrange, but finds that the knots originally laid down
When one does so, however, one needs to change the B-spline basis
... } S[j]
knots {zi' i = 6 •• , -1. O. 1. 2. and let
n
denote
S
n
M . (x) x] € S ( 1. 1)
nl n
where
n-l
g [z; x] := (z - x)+ . (1. 2)
n
express a B-spline of
105
M[~](x)
nl
.= g n [z.1-n , ... , z. l' z. l'
J- J+
... , Zl'; x] (1. 3)
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)
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)
[(z.-z.
J
. l(x) + (z.-z.)
l-n- 1) Mn,l- 1 J Mnl.(x}]/(z.-z.
1 l-n- 1)'
(1. 7)
i < j. Then
(1. 8)
where
s(x) = (1.11)
s(x) (1.11)
by taking
107
(1.12)
inserting a knot.
2. Knot Removal
from the original spline over a finite interval only; say over
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).
[.] [j]
aK = b n ,K+l,j a Kil + c nKj a K + T , (2.2)
unique.
spline of order n > 1 . having knots at zl < z2 < ... < zk and
(3. 1 )
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
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)
4. References
1. Introduction
2. Notation
m+1 m+2
z 1-bz az +
H (z)
1_b 2 1-b/z
IIHII
sup 1 H (z) 1 resp. IIhll := sup Ih(x} I.
Izl=1 -1 ~ x ~ 1
3. Basic Estimates
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.
II vII
IIv-pll
4. Compensation
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
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
6. Products
In addition to b and a we introduce -1 <d < 1 and
c ._ -d/(1-d 2 ) and put
-1 2 -2 -1 2-2
(az+1+bz +bz + •.. )(cz+1+dz +dz + ... )
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
which we normalize by w = 1.
1
~(a,c) .- 1+ad+bc'(ac,a+c) -. (u, v)
_1_ . T c ~(S)
1 +2t2
8. Summability
AP* Pu
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 Krabs
Fachbereich Mathematik der TH Darmstadt
1. The Problem
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
it follows that
2r kn 2
v(r,t) = R r: [c k exp( - a(T) t)
00
k=l
o
We assume for the following that
8(r,t) = 2r r: (-1,
k=I n r R (1.9)
cr ( R) ra
,t =""l-U 6 r: 1 exp «-a 1{
kn) 2t )8 - u( t )
{:2" ~ 0
00
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
For every T>O let UT = C[O,T] be equipped with the maximum norm
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.
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
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) } .
gT (u) = u (T) ,
126
- 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
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
(2.9)
and
(2.10)
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 )
(KT - KT,N)u(t) 1:
k=N+1
3 0f exp(-a(~ )2(t-S))u(s)ds
R
t
implies
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.
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
Author's address:
O.
1.
In order to have a well-posed problem the consistency condition r < k+l has
to be satisfied.
2.
W: = _1 (( ))N-1 21Ti
~ ~
11\1
11,\1 =0 ~ = exp N
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)
In [2] it has been shown that the common determinant of the linear systems
(4) can be written in the form
and
2k+1 (1)
(l-z) H2k +1,2(1,z) h2k +1,2(t,z) =
and (6)
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)
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.
3.
q (t)
...:.p = (ql ,p (t), q2 ,p (t), ... ,qN ,p (t)) T , P = O(I)r-l
dO
- q. (t)1 opO 0'1 p,O O(I)r-l, j 1(I)N .
dt O J,P J
t=O
By observing that
z N-l 1
res ~ =N
z =~ . . z
11 -1
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
-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
(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
\!
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.
and
respectively.
If we now define for p = 0,1
(12)
i.e.
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
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)
From (13) together with (16) - (18) we now conclude in accordance with
(14)
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:
, a = O(l)2k-r+l
~
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
The components of the respective Hermite fundamental splines are now easily
available according to (8).
143
References
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
[5] MERZ, G. und W. SIPPEL (to appear) Zur Konstruktion periodischer Hermite-
Interpolationssplines bei aquidistanter Knotenverteilung. J. Approxima-
tion Theory
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
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.)
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.
j = 1, ... , n + I, j #= i. (1.2)
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
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)
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
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)
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.
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
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).
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.
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
arguments (cp. SINGER [10], p.178, Theorem 1.3, Proof of (2) ~ (4) ) there exists a
D'E {-1,1} such that
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
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
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
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
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.
n
w(x) := n(x-x ) and
v=1 v
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
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
n k-1
P{x) L L p{h){x)t h{x)
v=1 h=o v v
157
n
TIn,k(x) := r-T(x-x vk )
v=1
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
v
x=x--
v
{Sin
E:
E: (A
(A -x ~
v
dx, (V =0, ± 1 , ± 2, .•. ) .
J
00
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
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>
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
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
(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
-00
]!,
a
_00 -00
i.e.
f
00
[ f (0) I 2 ~ ~ [f (x) [ 2 dx
_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
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
x -2 1f(x) 12 dx
so(z) minimizes the integral
J
_00
over all entire
00 k-1
f(z) = L L f(h) (A)Lvh(Z)
V=-OO h=o
where
where 00
wvh : =
f_00Lvh(x)dx (-00<v<00,0~h~k-1)
where 1jJ(z):=
(k-1)/2
r-r1
j =
(1 + ~2).
J
2
If f is an entire function of
00
k-1
J f(x)dx
1T
a j~O (2a)j a j ,k-1
-00
j even
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
1. Introduction
II p II : = max Ip (x) I •
-Hx:;;1
(1) II p II :; 1,
170
then
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.
1
Ilpll~:= sup ~(x) Ip(x) I·
-1<x<1
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
>../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
provided that
(8) j ~ A?
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
2
(9) Ip(x)I:>1-x for -1:> x :> 1.
Then
Proof. Let
This yields
which is (10).
174
Remark 2. It follows from [12, Theorem 1'] that under the hypo-
thesis of Proposition 1
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
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)
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
(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)
we conclude that
(x+1) R,~ (x) 1+xV
R,~(x) = -~-"2 R,n (x) ?: 0
(x-xv)
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
p(x)
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
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
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)
n 2 (1 + ~ G)
1T2
71 n-2 2 2
p*(x) :="2 . -2- (l-x )Pm(x)
where as n -+ 00 •
( ') dj 2
(15 ) II p J II 5; - , (x -1) Un - 2 (x) I
- dx J x=l
-L
2j-l
(1 )
x1
(n) •
x1 ••••
of nodes with
as n tends to infinity.
Theorem 4. Let
2v-1 7T
cos(n:::-:!. 2)' v= 1,2, •.• ,n-1.
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
References
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).
M. Reimer
FB Mathematik, Universitat Dortmund
Dekomposition
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
P(xy) = { P(xt)P(ty)dt
It =1
\l \l
= L L cvc\ { Pv(xt)P\(ty)dt.
v=o \=0 It =1
\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
~
mit --4-
7T 3/4
<0.85.
Literatur
R. Schaback
The author gratefully dedicates this paper to the memory of his academic
teacher, Prof. Dr. Helmut Werner.
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,
with the modulus w(.) of continuity of V f is frequently used (see section 2 of [8]).
The first axiom (see [8] and [9], for instance) concerns the descent of the error:
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
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
(8)
Proof: By summation of
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.
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
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.
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
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
• projection methods
• Levenberg-Marquardt methods
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
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.
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
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
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 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
to get
_ 1,b(x)
F(x) - ~(x,s(x)) ;::: ~4~TJ(X) (27)
2 1,b(x)
(1- t(x))JJ(x)p (x) <:: ~4~TJ(X)t(x) (2S)
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
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
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
•
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
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
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 :
•
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
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)
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·):
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
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
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)
[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)
Walter Schempp
ABSTRACT
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
D(R) = T IX A(R)
e 27ri ((2n+i)8+Az)id (n E Z) .
L 2 (R)
for an integer n ~ O.
5. Opto-Coupling
R Iil R
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
Theorem 4. Let m > n > 0, m' > n' > 0 be integers. Keep to
the preceding notations.
Cp, 1 , p ' 1
, ' = 0 for I * 1',
i.e., there is no coupling between quantized transverse eigen-
207
K (p,.(q-a)(q-a' ) p+p'+l)
p 'q(q-a-a')'
m+m' - 1 mod 2,
Cm,n,m " ,n a for
n+n' - 1 mod 2,
Cm, n, m" , n
K (v'· (q - a) (q - a I ). V + V I _~)
V ' q(q-a-a ' ) , t:.
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
References
An opto-coupler
211
... ............
'•. ,.""",..•••...••• -.
............. ..... .
........
. . . . . . . . . . . . . . . . . . . It • • • • •
. . ..
..
~
~
~ .... ........... .... ..... ........ . .....
.. .. --
.. ......
..
.
-
.... ........... ., ............• ...... ..
~ ~
. .................
~
........
.... .. .
................
,
..
~ ~
. -.
....................•
.....................
-
...
...... ...... .... ....
.. .....................•
~ ~
•......... ,...........•
~
-_ ..............""".
~ ~
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)
International Series of
Numerical Mathematics, Vol. 81
©1987 Birkhauser Verlag Basel
Jochen W.Schmidt
1. INTRODUCTION
't'1 ~ 't'a ~ ••• ~ 't'n ('t'1 < 't 2 < ••• < 't'n)
(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)
vex position for which all cubic C2-interpolants are not convex
on [0,1].
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
Because of
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
(11)
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
The limits
i = 1, •• ,n-1:
(14) sign cxi+1 = (-1) i ,lcxi+11 ~ IUil;
CXi+1 cxi
~~
Pi+1 ai
Now, using (12) and (15), inequalities (3) read
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.
Y~ = max{Yi: i € K j };
Yv • min{Yi: i € L~};
if Xv > Aj set Lj +1 = Lj\{i € L j S i >,,} v {j+1} other-
Y~ = min{Yi: i E Lj };
Yv = max{Yi: i E Kj };
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
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 +
(27 )
a i = max{~i,(Jri-bi_1)/2},bi = Jr i -2a i _ 1 (i=1, •• ,n),
coupled program
[2(8)
(28)
with
(29)
m+1 n m n+1
i;1 j;1 Fij(mij,mij+1)+ 1;1 j;1Gij(mij,mi+1j)~minl
c i = (Yian-Yn~i)/an'
Let, e.g., n be even. Then, because of an < 0, condition (3)
leads to
where
226
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
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'
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
Jochen W.Schmidt
Department of Mathematics
Technical University of Dresden
Mommsenstra~e 13
Dresden, 8027, German Democratic Republic
International Series of
Numerical Mathematics, Vol. 81
© 1987 Birkhauser Verlag Basel
Hans-Rudolf Schwarz
1. Einleitung
Wir betrachten das allgemeine Eigenwertproblem
Ax
- = ABx
- ( 1)
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
(k=1,2, ••• ) . (6 )
231
(k=O, 1, 2 , ••• (7 )
T T T
p := ~kB~k' a := Ek+1B~k' 1 := Ek+1BEk+1 (10)
~k = A£k , ~k = B£k ;
T
= 2k-1~k' y
T xT w
cr = -k-1-k'
~ = £k~k'
2k = 2k-1 + °k£k
Yk = Yk-1 + °k~k
T T~
a = 2kYk' P = 2kYk q = alp
(17)
~S = ~ I ( 19)
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)
-
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)
(26)
-
Wegen der Ahnlichkeit der Matrizen_ K und H(Y1) sind die
Eigenwerte der Hesseschen Matrix H(Y1) in aufsteigender Reihen-
folge
( 27)
- 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
2 1 4 1
1 2 1 1 4 1
121 141
A= B =
1 2 1 4
der Ordnung n hat die Eigenwerte
n = 10 20 40 80 160
~ = 30 46 87 164 314
-
~ = 8 7 7 7 7
236
T::: T
P -vv
- "'-0-0 =-ox-ov
- - -- -
~ = 2(Ya - qYa)/p = C-1 {2(Ya --
- qYa)/p} = C-1 ~ •
= ~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)
( 33 )
( 35)
Yk = CYk-1 + 6kC~k v
= -k-1 + 6k~k
T T-
a o ==
~Yo' Po = ~Yo; go = ao/po;
Iterationsschritt (k=1,2,3, ... l:
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;
~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;;
I
~-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
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.
au
an
=0 auf aG,
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
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
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
Literaturangaben
G. Alistair Watson
Department of Mathematical Sciences
University of Dundee
1. Introduction
n -A x
j i
r i = Yi - 1: a.e i 1,2, ... ,m, ( 1.1)
j=l J
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
1,2, •• ,n,
° if j (1.3)
aII! II
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 )]
~ ~
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].
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
(3.1)
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
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
Example 3
m 5, n 1.
min min
a A
4. Sensitivity
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*
~
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
_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
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
0.008
0.006
0.004
0.002
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
References
12. Osborne, M.R. (1975) Some special nonlinear least squares problems ,
SIAM J. Num. Anal. ~, pp.57l-592.
16. Ruhe, A. and Wedin, P-A. (1980) Algorithms for separable nonlinear
least squares problems, SIAM Rev. 22, pp. 318-339.