Spline Functions and Approximation Theory: Proceedings of the Symposium held at the University of...
Transcript of Spline Functions and Approximation Theory: Proceedings of the Symposium held at the University of...
ISNM INTERNATIONAL SERIES OF NUMERICAL MATHEMATICS
INTERNATIONALE SCHRIFTENREIHE ZUR NUMERISCHEN MATHEMATIK
SERIE INTERNATIONALE D'ANALYSE NUMERIQUE
Editors:
eh. Blane, Lausanne; A. Ghizzetti, Roma; A. Ostrowski, Montagnola; J. Todd, Pasadena;
A. van Wijngaarden, Amsterdam
VOL. 21
Spline Functions and
Approximation Theory
Proceedings of the Symposium held at the University of A1berta, Edmonton
May 29 to June 1, 1972
Edited by A. Meir and A. Sharma
1973
Springer Basel AG
ISBN 978-3-0348-5980-6 ISBN 978-3-0348-5979-0 (eBook) DOI 10.1007/978-3-0348-5979-0
© Springer Basel AG 1973
Originally published by Birkhäuser Verlag Basel 1973.
Softcover reprint of the hardcover I st edition 1973
FOREWORV
Around the end of May 1972, a conference on
Approximation Theory was organized at the University
of Alberta, Edmonton. The participants came from all
parts of Canada, the United States and Europe. Since
the majority of talks were related to the theory of
spline functions, we decided to let this volume of the
Proceedings of the conference be entitled "Spline
Functions and Approximation Theory".
We take this opportunity to express our thanks to
all those who participated in the conference or contri
buted to this volume. Thanks are due to the University
of Alberta and to the National Research Council of
Canada for financial assistance and for the facilities
provided. Professor Ostrowski deserves our special
gratitude for accepting this volume for publication on
behalf of Birkhäuser Verlag. Finally we wish to express
our appreciation to the staff of Birkhäuser Verlag for
their courtesy and valuable co-operation.
LIST OF PARTICIPANTS
C. DeBoor (Lafayette)
R. Bojanic (Columbuc)
Q. Rahman (Montreal)
I.J. Schoenberg (Madison)
W.A. Al-Salam (Edmonton)
A.M. Ostrowski (Basel)
M.Z. Nashed (Madison)
J.M. Munteanu (Baitimore)
L.L. Schumaker (Austin)
E.W. Cheney (Austin)
J. Fields (Edmonton)
A. Meir (Edmonton)
D.W. Boyd (Vancouver)
M. Ismail (Edmonton)
D. Leeming (Victoria)
M. Marsden (Pittsburgh)
E. Schmidt (Calgary)
S.W. Jackson (Edmonton)
S. Riemenschneider (Edmonton)
B. Cairns (Edmonton)
R. Gopalan (Edmonton)
J.C. Fung
P. Kumar
H. Berens (Santa Barbara)
Z. Ditzian (Edmonton)
G.G. Lorentz (Austin)
T.N.E. Greville (Madison)
C. Davis (Toronto)
R.S. Varga (Parma)
J.W. Jerome (Evanston)
E.G. Straus (Los Angeles)
F. Richards (Edmonton)
R. DeVore (Edmonton)
P.M. Gauthier (Montreal)
A. Sharma (Edmonton)
A. Al-Hussaini (Edmonton)
A. Giroux (Montreal)
R. Gervais (Montreal)
R. Pierre (Montreal)
G. Votruba (Missoula)
C. Nasim (Calgary)
S. Cabay (Edmonton)
C.P. May (Edmonton)
R. Gaudet (Edmonton)
D.S. Goel
W.Y. Chum
CONTENTS
Berens, H.: Pointwise Saturation. • • 11
Davis, C.: A Combinatorial Problem In Best Uniform Approximation. • • • • • • • • • • 31
DeBoor, C.: Good Approximation By Splines With Variable Knots • • •• •.•• 57
DeVore, R. & Richards, F.: Saturation And Inverse Theorems For Sp1ine Approximation. • • 73
Ditzian, Z. & May, C.P.: Saturation Classes For Exponential Formulae Of Semi-Groups Of Operators 83
FieZds, J.L. & IsmaiZ, M.E.: On Some Conjeetures Of Askey Coneerning Completely Monotonie Funetions. 101
Gauthier, P.M.: Une Applieation De La Theorie De L'Approximation A L'Etude Des Fonetions Holomorphes 113
Jerome, J.W.: Linearization In Certain Noneonvex Minimization Problems And Generalized Sp1ine Projeetions. • . . • . . • • . . . 119
Lyche, T. & Schumaker, L.L.: On The Convergenee Of Cubie Interpolating Splines • 169
Motzkin, T.S., Sharma, A. & Straus, E.G.: Averaging Interpolation. • • • . • • • • • 191
MUnteanu, M.: On The Construetion Of Multidimen-sional Splines . • • • • • • . • • . . • • 235
Ostrowski, A.M.: On Error Estimates APosteriori In Iterative Proeedures. • • . • • • • • • • • • •• 267
Schoenberg, I.J.: Sp1ines And Histograms. . • •• 277
DeBoor, C.: Appendix To 'Sp1ines And Histograms' By I.J. Sehoenberg • •• •• • . . • • • • •• 329
Straus, E.G.: Real Ana1ytie Funetions As Ratios Of Abso1ute1y MOnotonie Funetions. • • • • • • •• 359
ABSTRACTS
DeVore, R.: Inverse Theorems For Approximation By Positive Linear Operators. • • • • • • • • • • 371
Meir, A. & Sh~a, A. : Laeunary Interpolation By Sp1ines. • • • • • 377
Morr-is, P.D. & Cheney, E.W.: Stabi1ity Properties Of Trigonometrie Interpolation Operators . . . •. 381
Vapga, R.S.: Chebyshev Semi-Diserete Approximation For Linear Parabo1ie Problems. • • • • • • • • •• 383
POINTWISE SATURATION
Hubert Berens
O. The phenomenon of saturation which is associated
with many approximation processes can be considered as
a form of a differentiation process, and the questions
asked about saturation are ana10gues of questions
asked about differentiation.
To make the claim more precise, let us look at an
examp1e instead of giving formal notations and
definitions. The examp1e to be considered are the
Bernstein polynomials.
1. Let e[O,l] denote the space of all rea1-va1ued,
continuous functions on [0,1] under the supremum
For f E e[O,l] the th Bernstein norm. an , n
po1ynomia1 B f(x), n = 1,2, ••• , is defined as n
11
12
where
B f(x) = n
n k I f(-)Pk (x) k n ,n
=0
H. Berens
S.N. Bernstein introduced these polynomials in 1912/13
and proved
1im B f(x) = f(x) on [0,1], n n-+oo
uniform1y in x for each f E C[O,l] •
In 1932 E.V. Voronovskaya estab1ished the fo110w
ing asymptotic relation
1im n{B f(x) - f(x)} n n-+oo
= x(l-x) f"(x) , 2
whenever the right-hand side exists.
This relation shows that the order of approxima
tion of a non-linear twice differentiab1e function f
by B f n
is bounded by O(l/n) independent of its
actua1 order of smoothness, and one says with
J. Favard that the approximation process towards the
identity I on C[O,l] given by the Bernstein
Pointwise Saturation 13
operators {B n
n > 1} is saturated with order O(l/n)
as n -r 00 •
On the other hand, we may consider the sequence of
operators {n[B -I] : n > 1} n -
on C[O,l] as a process
converging to the differential operator D as n -r 00 ,
where
Df(x) = x(l-x) f"(x) . 2
From this point of view it is natural to ask whether
Voronovskaya's relation is invertib1e. First, does
1im {B f(x) - f(x)} 0 n n-roo
pointwise on (a,b) c [0,1] for some f E C[O,l]
imp1y that f"(x) exists on (a,b) and equa1s zero,
i.e., f is linear on [a,b]?
This was conjectured by G.G. Lorentz in his mono
graph [14] on Bernstein polynomials in 1953 and proved
by B. Bajsanski-R-Bojanic in a note [3] in the BAMS
in 1964. 1t is a consequence of the fo11owing
THEOREM A. Let f E C[O,l] be such that for aZZ
x E (a,b) c [0,1]
14 H. Berens
° < lim n{B fex) - fex)} , - n n-?<X>
then f is convex on [a,b].
The result above is known as the pointwise 0-
theorem for the Bernstein polynomials, and it is now
meaningfu1 to ask the fo110wing more general question:
Let f E e[O,l] , and let g be a function on
(a,b) c [0,1] • Does
1im n{B fex) - fex)} n n-?<X>
= x(l-x) g(x) 2
on (a,b)
imply that f is twice differentiab1e on (a,b) and
f"(x) = g(x) ?
The answer is affirmative in the fo110wing sense:
THEOREM B. Let f E e[O,l], g E L(a,b) , finiteZy
vaZued, where (a,b) C [0,1] . If
1im n{B fex) - fex)} < x(l-x) g(x) < 1im n{B fex) - fex)}
-- n 2 - n n-?<X> n-?<X>
on (a,b), then
x t
fex) = Ax + B + f dt f g(u)du a a
Pointwise Saturation 15
on [a, b ]. A and B are two appropriate constants.
This is a pointwise saturation theorem and it
generalizes the already elassieal saturation theorem due
to G.G. Lorentz in 1963 [15, p. 104].
THEOREM C. Let f E C[O,l] , and let M be a positive
constant such that
nIBnf(x) - f(x)1 ~ x(~-x) M + 0x(l) on (a,b) c [O,lJ ,
then f is continuously differentiable on [a,b] and
I f' (x) - f' (x) I ~ M I x - xl , x, X E [a, b] ,
and vice versa.
Theorems A and Bare elosely related to a result of
H.A. Schwarz on the eharaeterizations of eonvex funetions
through seeond symmetrie derivatives and C. de la Vallee
Poussin's extension, a fundamental lemma in the theory of
trigonometrie series. The lemma reads:
Let f E C[a,b] ,let g E L(a,b) , finitely
valued, and let
f(x+t) + f(x-t) =
2 t" ° . If
16 H. Berens
1im ~ {Ltf(x) - fex)} t~O+ t
on (a,b) , then
x t fex) = Ax + B + f dt f g(u)du on [a,b].
a a
The proof is done by reducing it to the case
g(x) = 0 , which is essentia1ly Schwarz resu1t, and this
is obtained by a theorem in Lebesgue theory on majorant
and minorant functions due to de la Vallee-Poussin, see
G.H. Hardy-W.W. Rogosinski [10, p. 90].
Hence one may interpret the three theorems for the
Bernstein polynomials as analogues of the classical
results of Schwarz and de la Vallee-Poussin on second
symmetrie derivatives. These theorems can further be
extended to a whole class of approximation methods which
satisfy a Voronovskaya condition.
2. Let v,w be functions in C[a,b] , strictly positive
on (a,b) , and let
Pointwise Saturation
x x ~(x) = J v(t)dt I/I(x) = J w(t)dt
a a
x and ~(x) = J I/I(t)v(t)dt
a
For a function f in C[a,b] , define the differential
operation Df(x) at a point x E (a,b) by
1 {f' (x)}' Df(x) = DI/ID~f(x) = w(x) v(x) ,
17
whenever the right-hand side is meaningfu1, see S. Kar1in
W. Studden [13, eh. XI].
Let {L n > 1} be a sequence of positive linear n
transformations on C[a,b] into itse1f, let O. :n~l} n
be a sequence of positive rea1s tending to ~ as n ~ ~ ,
and let p(x) be a function in C[a,b] , strict1y positive
on (a, b) • We say {L : n > 1} n
satisfies a
Voronovskaya condition if
1im Ä {L fex) - fex)} = p(x)Df(x) , x E (a,b) , n~ n n
whenever Df(x) exists.
Under these conditions
18
1im L f(x) = f(x) n n-+<x>
(f e: C[a,b])
pointwise for each x e: (a,b) • We have
THEOREM AI. FoX' a funation f e: C[a,b]
o < 1im A {L f(x) - f(x)} on (a,b) - n n n-+<x>
H. Berens
if, and on"ly if, f is aonvex with X'espeat to {1,~} on
[a,b] i.e.,
{~(x1)-~(x )}f(x) < {~(x)-~(x )}f(x1) + {~(x1)-~(x)}f(x ) 0-0 0
x < x < xl • 0- -
COROLLARY. Let f e: Cla,b] be suah that
1im A {L f(x) - f(x)} = 0 on (a,b) , n n n-+<x>
then f is "lineax> with X'espeat to {1,~} , i.e.,
A + B~(x) , wheX'e A and B ax>e aonstants.
THEOREM BI. Let f e: C[a,b] , and "let g e: L(a,b)
finite"ly va"lued. If foX' eaah x e: (a,b)
f(x)
be
1im A {L f(x)-f(x)} < p(x)g(x) < 1im A {L f(x)-f(x)} , --nn - - nn n-+<x> n-+<x>
=
Pointwise Saturation 19
then
x t
fex) = A + B~(x) + J v(t)dt J g(u)w(u)du on [a,b] , a a
whepe A and B aPe aonstants.
These theorems are best possib1e in the fo1lowing
sense: If the limit relations are violated at even one
point in (a,b) then the conclusion does not hold.
THEOREM C'. Let f € C[a,b] ~ and Zet M be a positive
aonstant.
A IL fex) - f(x)1 < p(x)M + 0 (1) as n + ~ n n - x
if~ and onZy if~ D~f exists and beZongs to C[a,b] and
Theorem A' is independently due to several authors.
G. Mühlbach [17] used divided differences and results of
T. Popoviciu on generalized convexity, extending results
of V.A. Amel'kovic [1], in 1966. The proofs given by
J. Karamata-M. Vuilleumier [12] and G.G. Lorentz-
L.L. Schumaker [16] use in principal the arguments in the
proof of Schwarz' lemma, mentioned above. Theorem C' is
in the above form due to G.G. Lorentz-L.L. Schumaker, loc.
20 H. Berens
cit., who extended the versions of V.A. Amel'kovi~ and
G. Mühlbach. Theorem B' is due to the author r6].
The conditions imposed on the sequence of operators
{L : n > I} are satisfied for a large class of appro-n
ximation processes. In addition to the Bernstein opera
tors, we want to mention the Bernstein power series
introduced by W. Meyer-Konig and K. Zeller, the extension
due to E.W. Cheney and A. Sharma, the Szasz operators,
the Gauss-Weierstrass operators, etc.
A sequence of operators which does not belong to
this class are the Fejer-Hermite operators: Let
f E Cr-l,l] ,
T2 (x) n
H f (x) = --..:..:..,::--n 2
where T (x) n
is the
x. = cos(2i-l)~/2n • 1.
n
th n
n L f(x)
i=l
I-xx. 1.
Chebyshev polynomial and
{H : n > I} forms a sequence of positive linear n
interpolation operators which does not satisfy a
Voronovskaya condition, at least not pointwise.
Pointwise Saturation 21
3. The phenomenon of saturation is best studied for
summability methods of Fourier series and integrals, see
the monograph [7] of P.L. Butzer-R.J. Nessel or the
lecture notes [8] of R. DeVore.
Let fex) be a real-valued, 2~-periodic integrable
function on the real line. lts Fourier series is denoted
by
a ()O
fex) ~ ~ + I ~ (x) 2 k=l-K.
where and are its Fourier coefficients.
conjugate series of f is defined by
()O
Hf(x) ~ I Bk(x) k=l
The
Concerning pointwise saturation, a first result was given
by V.A. Andrienko [2] in 1968 for the Fej~r means. The
th n Fejer means of the F.s. of a function f, defined by
a 0" (f;x) = ~ + n 2
n k I (1 - n+l)~(x)
k=l n=1,2, ••• ,
form a positive summation methode Andrienko proved
22 H. Berens
THEOREM D. Let f e: L21T be finiteZy-vaZued and suah that
0n(f;x) aonve~es to fex) as n ~ ~ fop aZZ x in
some intepvaZ (a,b). If
(1) lim (n+l){o (f;x) - fex)} = 0 n n-+oo
fop aZZ x e: (a,b) ~ then~ fop aZmost aZZ x~ Hf(x) is
a aonstant funation on (a,b). Mopeovep~ if f e: C21T
then the aonaZusion pemains tpue even if (1) is vioZated
in a denumepabZe set of points.
To establish the connection to the previous sections,
let X21T be one of the spaces L~1T' 1 2 p < ~ , or C21T
endowed with the usual norm. The following asymptotic
relation is weIl known:
If fand [Hf]' belong to X21T , then
(2) lim (n+l){on(f) - f} = -[Hf]' in X21T-norm. n-+oo
The differential operator (d/dx)H on the right-hand
2 2 side of (2) and the operator d /dx are connected
through
Pointwise Saturation
d -H = dx
i.e., here we are dea1ing with an asymptotic relation
i d f i 1 f _d 2/dx2 , converg ng towar s a ract ona power 0
see [7].
23
The asymptotic relation (2) imp1ies the saturation
theorem for the Fejer means which is due to G. A1exits
for C2~' in 1941, and G. Sunouchi-C. Watari in general,
in 1958/59.
The pointwise saturation theorem for the Fejer means
reads:
THEOREM E. Let f € L2~ be such that
finitely for all X in some interval
1im ° (f;x) = fex) n
n~
(a,b) " and let
g € L(a,b) be finitely-valued and satisfy
(3) 1im (n+1){on+1(f;x) - fex)} = g(x) n~
00
pointwise on (a,b). Then ~1~(x)/k is the P.s. of
some F € L2n and for almost all x € (a,b)
x t
F(x) Ax + B + f dt f g(u)du a a
24 H. Berens
The proof of these two theorems rests on the
following observation: Let
(4) 00
l (-k)~(x) k=l
be the trigonometrie series assoeiated with [Hf]' • The
limit in (3) exists finitely if, and only if,
n k lim cr ([Hf]',x) = lim l (1 - n+l)(-k)~(x) n~ n n~ k=l
exists, and both limits are equal. This reduees the
theorem to a uniqueness problem of a Fejer summable
trigonometrie series. The uniqueness theorem ean be
obtained either from S. Verblunsky's uniqueness theorem
for Abel summable trigonometrie series (cf. A. Zygmund
[21, p. 352ff]), or from results due to F. Wolf [20]
about (C,A)-summable series.
As a eonsequenee of Theorem E we have the following
COROLLARY. Let f € C2~' g € L2~ , fineteZy-vaZued. If
(3) hoZdS tpue fop aZZ x e~aept possibZy on a denumep
abZe set, then Hf is absoZuteZy aontinuous and
[Hf]'(x) = -g(x) a.e.
Pointwise Saturation 25
The coro11ary substantia11y weakens the asymptotic
relation (2) for continuous functions.
The resu1ts stated for the Fej~r means can be ex
tended to the typica1 means of order y: For an
f E L21T
n y [T f](x) = L (1 - k )~ (x)
n,y k=l (n+1)y -K (y > 0, n = 1 2 ) , ,... .
In this case,
n~
1im (n+1)y{T f - f} n,y
See [4] and G. Sunouchi [19]; see also f5] for a point
wise saturation theorem for the Abe1 means.
R. DeVore [9] proved a general pointwise o-theorem
for approximation methods on C21T which commute with
translations and which have the saturation phenomenon.
To be precise, let f E C21T and let
a 00
~ ~ + L P ~ (x) ,. n = 1,2,... , 2 k=l k,n-K
26 H. Berens
where Iln is a positive even Borel measure on [-lT,lT] lT
with Pk = (l/lT) f cos k.x dll (x), Po,n = 1 , and such ,n n -lT that
l-p lim PI n = 1 and lim k,n = 1/Ik > 0 l-p n-+oo ' n-+oo l,n
for each k = 1,2, .••.
The last conditions assure that {L : n ~ I} is n
an approximation process saturated with order O(l-P l ) ,n as n -+ 00 •
THEOREM F. Let f be in C2lT suah that fo1' aZZ
IL fex) - fex) I = 0 (1 - P ) as n -+ 00 , n x l,n
then f is a aonstant-vaZued funation.
R. DeVore's proof is indirect and based on an
analysis of the support of Iln in [-lT,lT] as n-+ oo •
The associated pointwise saturation theorem is not known,
and it is unlikely to hold true in the given generality,
(see [8]). However, it seems to be of interest to find
sufficient conditions which guarantee a pointwise theorem
for classes of approximation processes. A first step in
this direction has been done by T. Hedberg [11], a young
Pointwise Saturation
Swedish mathematieian, and this even for Fourier series
and integrals in severa1 variables.
5. Riemann's theory of trigonometrie series and
integrals in m-dimensions was 1arge1y deve10ped by
27
V.L. Shapiro, (see [18]). Hedberg studied uniqueness
theorems for summab1e trigonometrie series and integrals,
where the summation method is given as a eonvo1ution
integral.
As an app1ieation he proved, e.g., the fo11owing
pointwise saturation theorem:
THEOREM G. Suppose that K is a positive radiaZ kerneZ
on Em whiah satisfies (iJ J Kdx = 1 and (iiJ
Em R + 00 , and set J Kdx = o(R-2)
Ixl~R as ~(x) = R~(Rx) •
If f is a bounded aontinuous funation in Em for whiah
2 1im R {f*~(x) - fex)} = g(x) , R+oo
at eaah point x E ~ , where g is finite and ZoaaZZy
integrabZe, then ~f = g •
~ is the Lap1aee operator. Hedberg also proves a
28 H. Berens
pointwise o-theorem for the Poisson integral on Em and
Tm , and gives an independent proof of the pointwise
saturation theorem for the Abe1 means of Fourier series
and integrals on the real 1ine.
REFERENCES
1. Ame1'kovic, V.G.: A theorem converse to a theorem of Voronovskaya type, Teor. Funkei!, Funkciona1 Anal. i Pri1ozen, Vyp 2 (1966), 67-74.
2. Andrienko, V.A.: Approximation of functions by Fejer means, Siberian Math. J. 9 (1968), 1-8.
3. Bajsanski, B. and Bojanic, R.: A note on approximation by Bernstein polynomials, Bu11. Amer. Math. Soc. 70 (1964), 675-677.
4. Berens, H.: On pointwise approximation of Fourier series by typica1 means, T6hoku Math. J. 23 (1971), 147-153.
5. Berens, H.: On the approximation of Fourier series by Abe1 means, J. Approximation Theory (accepted for pub1ication).
6. Berens, H.: Pointwise saturation of positive operators, J. Approximation Theory 5 (1972), 135-146.
Pointwise Saturation
7. Butzer, P.L. and Nessel, R.J.: Fourier Analysis and Approximation, Basel 1970.
29
8. DeVore, R.: Approximation of continuous functions by positive linear operators, Lecture Notes 1970.
9. DeVore, R.: A pointwise "0" saturation theorem for positive convo1ution operators, Proceedings of the Conference on Linear OperatoIS and Approximation, Oberwolfach 1971.
10. Hardy, G.H. and Rogosinski, W.W.: Fourier Series, Cambridge 1944.
11. Hedberg, T.: On the uniqueness of summab1e trigonometrie series and intergrals, Ark. Mat. 9 (1971), 223-241.
12. Karamata, J. and Vui11eumier, M.: On the degree of approximation of continuous functions by positive linear operators. Mathematics Research Center, U.S. Army, Madison, Wisconsin.
13. Kar1in, S. and Studden, W.: Tchebycheff Systems. New York 1966.
14. Lorentz, G.G.: Bernstein Polynomials. Toronto 1953.
15. Lorentz. G.G.: Approximation of Functions. New York 1968.
16. Lorentz, G.G. and Schumaker, L.L.: Saturation of positive operators, J. Approximation Theory 5 (1972), in printe
17. Mühlbach, G.: Operatoren vom Bersteinsehen Typ, J. Approximation Theory 3 (1970), 274-292.
30 H. Berens
18. Shapiro, V.L.: Fourier Series in Several Variables, Bull. Amer. Math. Soe. 70 (1964), 48-93.
19. Sunouehi, G.: Pointwise approximation of funetions by typieal means of Fourier series (to be pub1ished).
20. Wolf, F.: On summable trigonometrie series: an extension of uniqueness theorems, Proe. Land. Math. Soe. (2) 45 (1939), 328-356.
21. Zygmund, A.: Trigonometrie Series. Vol. I, Cambridge 1959.
A COMBINATORIAL PROBLEM IN BEST UNIFORM APPROXIMATION
Chandler Davis
PROBLEM: Given a funation f of one vaPiable, to
minimize
Ilf - gll= sup If(t) - g(t)1 t
among all g whiah aPe monotonia on at most n sub
intervals.
1. ROW TO POSE TRE PROBLEM AND WHY
31
Sometimes all we really want to know about a
function is the sequence of its increases and decreases.
When it is given empirically or by an imperfect formula,
the imperfection we most regret may be the occurrence of
small spurious ups and downs of the graph. Indeed some
of them may not be so small that we feel safe and easy
about modifying the function so as to flatten them out.
32 C. Davis
Let us put the matter quantitatively.
DEFINITION~ A function g defined on areal interval
[a,b] will be said to have 'oscillation order' n in
case there exists a partition [a,b] = [to,t1 ] U [t1 ,t2]
U ••• U [t l,t] such that the restrictions of g to n- n
[t. l,t.] (j = l, ••. ,n) are, in alternation, non-con-J- J
stant increasing and non-constant decreasing functions.
Por a function defined on a subset (in particular, a
discrete subset) of an interval, the terminology is the
scune, wi th the taci t understanding that a symbo l for an
interval denotes the intersection of that interval wi th
the domain.
Thus there are two kinds of functions of oscil
lation order n : those which begin on the first sub
interval with an increase, and those which begin with a
decrease; but if any of the n restrictions were con
stant, we would be obliged to change the partition to
one with fewer subintervals and say the oscillation
order was properly less than n. Note that the
definition does cover the extreme situation in which the
domain of g is just {t , t 1 , •.• , t } . o n
Assume the given function f has some finite
oscillation order N. (No interesting cases are there-
A Combinatoria1 Problem in Best Approximation
by exc1uded.) We ask whether any g with osci11ation
order ~ n is reasonab1y c10se to f But by what
metric shou1d we judge what is "c1ose"?
One reasonab1e definition wou1d be by an LP(~)
norm, for some measure ~ on the domain, and
1 ~ p < ~ For p = 2, n = 1 this is c10se to a
33
problem which has been studied by statisticians [1]. A
second possibi1ity, which appears the most natural in
case the functions in question are frequency functions,
is the Kantorovic-Rubinstein metric [3]. In this metric
the distance from f to g is the solution of a trans
portation problem: to move a unit mass distributed
according to f unti1 it is distributed according to
g , at minimum total cost (mass times distance). Here
I fo11ow a third alternative definition, the supremum
norm. It is especia11y suited to those situations which
are unaffeated by order-ppeserving ahanges of the inde
pendent variabLe. Thus this work is descended, in
spirit though not in detail, from the qualitative theory
of Cebysev systems, especia11y [4], [5], [2].
By assumption, there are N subinterva1s,
exhausting [a,b] , on each of which f is monotonie,
with alternation between increase and decrease on
successive interva1s. Let [te 1,t.] be one of the J- J
34 C. Davis
subintervals, whose endpoints do lie in the domain, and
let g be any proposed approximant to f. Now suppose
g to be replaced by that function g' which has the
same values as g except that at points of the domain
lying in ]t. l,t.[ J- J
(if any) it is redefined so that
g' is, on [t. l,t.] , a linear transform of f J- J
The
change can not increase the distance away from f, nor
can it increase the oscillation order of the approximant.
Accordingly there was no need to consider the irrelevant
added structure of g in the first place; that is to
say, the whole problem under study is determined by the
sequence of values of f at the points of its domain
where it reverses the sense of monotonicity. Since the
parametrization is also without effect, we are free to
declare these points to be equally spaced. Summing up,
there is no loss in generality in confining ourselves to
the following discrete version:
PROBLEM: Given a function f on {O,l, •.. ,N} , such
that the sequence (f(j) - f(j-l)~=l is alternating
in sign (strictly)~ to minimize
max If(j) - g(j)1 j
among all functions g on {O,l, ... ,N} which are of
oscillation order at most n (n < N) •
A Combinatorial Problem in Best Approximation
These assumptions and notations will be retained
throughout the rest of the paper.
35
The main features of the problem appear already in
the simplest ease.
Example 1; N = 2, n
j
f(j)
1
o
o
1
3
Define f by
2
1
We are to find its best approximation by a monotonie
funetion (funetion of oseillation order < 1). It is
easy to see that the best approximating g must have
gel) = g(2) = 2 ; but there is no uniqueness, g(O) ean
be anything subjeet to -1 ~ g(O) < 1 .
Apparently the main objeetive should be an effieient
algorithm leading to some best approximant. I believe
this is aehieved satisfaetorily by the analysis in
Seetion 3,4. First I state in Seetion 2 some of the
neater results of the analysis, not beeause they are its
whole aim, but in order to show at onee how the be
haviour observed in Example 1 generalizes.
36 C. Davis
2. SOME OF THE RESULTS
DEFINITION:. A ' subfunction' of f is the restriction
of f to a subset {i ,il, ..• ,i} of {O,l, ••. ,N} o m
(m ~ N; i k_l < i k) ~ such that the sequence
(f(ik) - f(ik_l»~=l is alternating in sign (strictly).
(The subset deterrnines the subfunction~ and it wiU be
harmless to say the subset is the subfunction~ f being
fixed.) The ' discY'epancy' of the subfunction
{io ,il ,··· ,im} is mink !f(ik) - f(ik_l )! •
THEOREM: Let the maximal discY'epancy~ among aU sub
functions of osciUation oY'deY' exactZy n+ 1 ~ be 28.
Then the minimum of 11 f - g 11 ~ among aU g of
oscillation oY'deY' ~ n ~ is 8.
Half of this can be proved at once. Let
{io,il , •.. ,in+1} be a subfunction of discrepancy 20 ,
and let g be any function with IIf - gll = 8' < 8 •
Then because each
each g(ik) - g(ik_1) has to be non-zero and of the
same sign as f(ik) - f(ik_1). Therefore g has
A Combinatorial Problem in Best Approximation 37
oscillation order > n + 1 •
The inequality in the other direction will be
proved along with the construction of certain special
best approximants. Part of the details will be stated
in this Section and proved in Section 3.
DEFINITION: The subfunction {i ,il, ... ,i} will be o m
called 'critical' in case (for each k 0,1, ... ,mJ
f(i ) = m~x {f(·) k ml.m J
the signs "max" and "min" being chosen in alternation as
k increases. Here the otherwise undefined symbol i_I
is to be interpreted as o ~ whether or not i = 0 ; o
similarly~ i mtl is to be interpreted as N.
Convention: To say a subfunction {io,il , .•• } 'has
maximal discrepancy' implies not only that
mink /f(ik ) - f(ik_l )/ has the largest possible value,
but also that this minimum is attained for the smallest
possible number of different k.
PROPOSITION: Among those subfunctions of order m
having maximal discrepancy~ at least one is critical.
38 C. Davis
THEOREM: Let {io,il, ••• ,in+l } be a critiaaZ sub
funation having maximaZ disapepanay, and Zet
mi~ If(ik) - f(ik_1)I be attained fop k = k' • Then
thepe is a funation g minimizing 11 f - g 11 among
funations of osaiZZation oPdeP < n and having the =
fupthep pPOpepty that g(ik) = f(ik) fop aZZ
k ~ k'-l,k' ; exaept that k = 0 must aZso be exaZuded
if i ~ 0 , and k = n + 1 must aZso be exaZuded o
if i n+l ~ N •
ExampZe 2: N = S •
j 0 1 2 3 4 S
f(j) 0 3 2 S 0 2
Then the subfunction {0,1,4,S} has maximal discrepancy
but is not critical. Similarly for {2,3,4,S} •
3. FROM CRITICAL SUBFUNCTION TO BEST APPROXlMANT
The first aim is to prove the Proposition stated
in Section 2.
If there is some k' (0 < k' < m) such that = = f(ik ,) is greater than f(ik'_l) or f(ik'+l) but
A Combinatorial Problem in Best Approximation
max {f(j)
then redefine i k , to equal a j where that maximum
is assumed; similarly for minima. Each such change
increases either one or two of the If(ik - f(ik_l )I
39
and leaves the others unchanged, so their minimum, the
discrepancy, is if anything increased. For the same
reason we note that, in any step which does not alter
that minimum, the number of different k at which it is
assumed is if anything decreased. After a finite number
of iterations the process can not be continued, and the
subfunction is then critical.
Nowassume {io,il, .•. ,in+l } is a critical sub
function having maximal discrepancy 28. I will con-
struct a function g of oscillation order m < n such =
that Ilf - gll = 8
First I will choose a set of points {h ,hl, •.• ,h } o m
from among {O,l, ••. ,N} (hk_ l < hk). They will serve
as guides in the construction of g; hl, .•• ,hm_l are
to be local extrema of g , while
mere navigational aids.
h o
and h m
Fix k' such that If(ik ,) - f(ik'_l)1 = 28
are
40 C. Davis
(1 ~ k' ~ n + 1) •
Gase A: k' = 1. In this case {il ,i2 , ••• ,in+l } will
be chosen as the hk , and the function g we get will
have oscillation order n.
The symmetrical case with k' = n + 1 is given
corresponding treatment and will also be called Case A.
Gase B: j < i with o
f(j) < f(io) - 28. Let jo be some j which minimizes
f (j) among all j < i o
In this case, the hk will
be {jo,io,il , .•• ,in+l}\{ik'_l,ik } , and we will again
get a function g of oscillation order n.
There are symmetrical variants of this case too.
(Apparently a subfunction could fall under Case A
as regards the left end of the interval and also fall
under Case B as regards the right end. Such an ambiguity
is actually prohibited by the hypotheses, but we won't
have need of that fact. If a subfunction is in both
Case A and B as regards the same end, then folIoweither
rule, they will give the same function.)
Gase G: We are not in Case A or B. In this case,
A Combinatoria1 Problem in Best Approximation
{io, ..• ,ik'_2,ik'+1, ... ,in+1} will be chosen as the
hk ' so that we will get a function g of oscia11ation
order n - 1 •
In the rest of the Section, let {h ,h1 , •.• ,h } o m
(m = n or n - 1) be a sequence obtained by one of
the ru1es A, B, or C, and take without 10ss of
genera1ity f(ho) < f(h1) •
Construation of the approximant: Define g(hk) to be
f(hk) for 1 < k < m - 1. For j' E ]h1 ,h2 [ , define
2g (j ') = min {f (j) : h1 < j < j'}
+ max {f(j) j' < j < h 2 }
Simi1ar1y for Jh2,h3[, •.• ,Jhm_2,hm_1[. For
j' E [O,h1 [ , define
2g(j') = max {f(j) : 0 ~ j ~ j'}
+ min {f(j) ., < . < h } J = J 1
Simi1ar1y for ]hm_1 ,N] •
There are severa1 verifications to make.
41
42 C. Davis
First, how many intervals of monotonicity does this
g have? It is (non-strictly) increasing and (non
strictly) decreasing, alternately, in the intervals
[O,hl],[hl,h2], ••• ,[hm_2,hm_l]' [hm_l,N]. To see
this requires two observations: (i) g is non-increasing
on (say) ]hl ,h2 [ because each term in the above
definition is non-increasing, and (ii) f(hl ) ~ f(j)
for hl < j ~ h2 by a short argument using the
derivation of {ho,hl , .•. } from a critical subfunction;
similarly for the other subintervals. Furthermore, on
each of these m intervals the value of g changes by
at least 0 (as you may check from the definitions) and
so is surely not constant. Therefore g really does
have oscillation order m.
Next, the value of Ilf - gll is really < o. It
is convenient to prove this in terms of an auxiliary
notion.
DEFINITION: On one of the intervaZs [O,hl ],[h2 ,h3],
[h4 ,hS]' •.. where g is inareasing, the 'retrogression'
of f is defined to be ° if f is aZso (non-striatZy)
inareasing there; othe~ise, it is defined to be
max {f (j ') - f (j ") } taken over pairs j' < j" in the
intervaZ. SimiZarZy, on one of the intervaZs [hl ,h2],
A Combinatoria1 Problem in Best Approximation 43
[h3,h4], .•. where g is deareasing, the 'retrogression'
of f is defined to be 0 if f is aZso (non-striatZy)
deareasing there; otherwise, it is defined to be
max {f(jlt) - f(j')} taken over pairs j' < jlt in the
intervaZ.
It is easy to see that the gwhich has been de
fined will be within 8 of f if and on1y if f has
retrogression ~ 28 on each of the m interva1s.
According1y we want to estab1ish the 1atter fact.
To this end, consider first an interval of mono
tonicity obtained (by any of Cases A, B, C above) as
[ik"_l,iklt ] , and suppose if possib1e that f has
retrogression > 28 there: say, that f(ik"_l) > f(ik,,)
and that f(j") - f(j') > 28 for ° <Jo'<Jo"<i 1 k"_1 = = k"
Actua11y, the subfunction being critica1, f(ik"_l)
f(j ') and f(j") - f(ik,,) are both ~ f(j") - f(j')
(which forces, incidenta11y, and Jo" ..L ° ) T 1 k " .
Let, as before, min If(ik ) - f(ik_1) I be attained for
k = k' • Of course f(ik"_l) - f(ik,,) > 28 , so k":; k'.
Th en {i 0 ' i 1 ' • . • , i k" -1 ' j , , j " , i k", i k" + 1 ' . . • , i n+ 1 } \ Ü k ' -1 '
i k ,} is a subfunction which, as compared to {io ,i1 , •.. ,
44 C. Davis
i n+l } , has lost at least one difference If(ik )
f(ik_l ) I which is exactly 28 and has not gained any.
In view of the Convention of Section 2, this contradicts
the choice of {io,il, .•. ,in+l } as having maximal
discrepancy. (Remember to verify the assertions of
this paragraph even in the exceptional cases that
k' is 1 or n + 1 , and that ktf is k' - 1 or
k' + 1 .)
A very similar argument disposes of an interval of
monotonicity obtained by Case B or C above as
[ik '_2,ik, +l ]
Still to be excluded are the possibilities that
f has retrogression > 28 on IO,hl ] or Ihm_l,N]
Suppose if possible that f(j') - f(j") > 28 for
o < j' < j" ~ hl . The reasoning already given covers
the case 'I > h 'd J = 0' so cons~ er j 1 < h o
loss of generality assume that max {f(j)
attained at j = j' , and that min {f(j)
Without
, < '''} J = J is
is attained at j j". Similar conventions in treating
the alternative h <j'<j". m-l
We can now rule out the possibility that j" < h o
A Gombinatorial Problem in Best Approximation
(and its symmetrical variant
j" < h o
with f(j") > fCh ) o
j' > h ). m
Indeed, if
, then we are free to re-
45
define j" to be h o
(possibly redefining j' as well,
but so what?). If j" < h with f(j") < feh ) , then o 0
h can not be i , and neither can it be the j of 000
Gase B, else the criticality of the initial subfunction
would be violated; the remaining possibility is Gase A
with ho = i l ' and then {j',j",i2,i3, ... ,in+l } shows
that {io,il, ••• ,in+l } could not have had minimal
discrepancy.
Similarly, and more simply, we rule out the
possibility that ho < j" ~ hl (and its symmetrical
variant hm_l ~ j' < hm).
Gase G is now disposed of: if, say, f(j') - 28 >
f(io ) < feil) then the conditions for Gase B would pre-
vail, and hence Gase G would be excluded by definition.
All possibilities under Gase Aare obtainable by
symmetry from the following two:
j" = h = i • The manner of fixing j' and j" , o 1
together with the criticality of the given subfunction,
46 C. Davis
entail f(j') = f(i ). But h could not have been o 0
chosen as i l unless Ifeil) - feio) I = 20 , contradic-
ting f(j') - f(j") > 20
A2 •• h • o = 1.1 ,
is a subfunction which, as compared to {io,il, ••• ,in+l}'
has lost one difference If(ik ) - f(ik_l )I which is
exactly 20 and has not gained any - a contradiction.
In Case B, considering by symmetry only the
alternative h < h = i there are still two o 1 0'
possibili ties :
BI: j' < j" = h o
Then {j',h , •.• ,h} is a subfunc-o n
tion which is readily seen to violate the assumed
maximal discrepancy of {io, •.• ,in+l}.
j" > j' =
same role.
h n
Then {h , ••• , h ,j 11 } o n
will fill the
This completes the proof that IIf - gll ~ 0 •
Together with the Proposition, it supplies all that
was lacking to establish the first Theorem. Furthermore,
the function g constructed has all the special
A Gombinatorial Problem in Best Approxim&tion 47
properties prescribed in the second Theorem. (If, say,
i = 0 f k' - 1 , then h o 0
o , and gei ) o
does indeed
equal f (i ).) o
The structure of the argument may be brought out
by some examples.
It emerged in the proof that the oscillation order
of the best approximant g can always be taken to be
n - 1 or n. It can not always be taken to be n .
ExampZe 3: N = 3, n = 2
j o 1 2 3
f (j) o 2 1 3
Then every best approximant of order < 2 must have
order 1.
Thus Gase G is indispensable, and so, obviously,
is Gase A. As to Gase B, consider
Examp Ze 4: N = 5, n = 3
j o 1 2 3 4 5
f (j) 1 3 1 2 o 2
The only critical subfunctions of order 4 are of course
48 C. Davis
{O,1,2,3,4} and {1,2,3,4,S}. App1ying the ru1e of
Case C to either of these yie1ds an approximant at
distance 1 from f. We need the ru1e of Case B to
get {ho ,h1 ,h2,h3 } = {O,1,4,S} and IIf - gll = 1/2
Let me comment also on the Convention concerning
"maximal diserepancy". It recommends itse1f on general
grounds: it makes the choice of {io , ... ,in+1 } insensi-
tive to sma11 perturbations in the va1ues of f. But
it is also essential for the truth of the second Theorem!
ExampZe 5: N = 8, n = 4
j
f (j)
o
1
1
2
2
1
3
2
4
1
S
3
6
1
7
2
8
o
The subfunction {O,1,2,3,4} satisfies all requirements
of the Theorem except for the Convention, yet it gives no
e1ue to the best approximation. Less drastic is
ExampZe 6: N = 4, n = 2 .
j
f (j)
o
o
1
2
2
1
3
3
4
2
The subfunction {1,2,3,4} satisfies all requirements
of the Theorem except for the Convention, yet it leads to
A Combinatorial Problem in Best Approximation
a best approximant only if the choice of k' is for
tunate.
4. THE SEARCH FOR THE SUBFUNCTION
Do the ideas introduced really help solve the
Problem? How would one actually set about finding 0
for a non-trivial instance with N around 100 and n
around 8?
49
One can write an algorithm which starts with an
arbitrary g of the required oscillation order and
progressively diminishes Ilf - gll , ending after a finite
string of changes with a best approximating g like that
constructed in the last Section. Inevitably, it uses the
same ideas as I have been using. There are no numerical
experiments of such a size as to be interesting; however,
I would guess it is ordinarily better to approach the
problem from the other side, as follows.
Setting o(n) for the 0 of Section 2, to make
explicit its dependence on n we note that 20(N-l) =
min. If(j) - f(j-l) land ask for a simple way of going J
from any o(n) to o(n-l) •
LEMMA: If {i , ... ,i } o m
has maximaZ disarepanay 20 "
50 C. Davis
then f has retrogression < 28 on eaah intervat ==
[ik_1,ikJ (k == 1, ••• ,m) •
Though we did not need this fact in Section 3, its
proof was essentia11y given there, so I will say no more
about it.
DEFINITION: If i o < i 1 < ••• < im 3 then to 'throwout'
a pair (ik_1 ,ik) will mean to replaae the set
{i , ... ,i} by o m
{i, ... ,i 1}3 if k=m; o m-
DEFINITION: Calt a subfunation {i , ... ,i } o m 'basia'
in aase it has maximal disarepanay 28 and3 beside
being aritiaal3 satisfies
1 f (j) - f (i )1 < 2n o =
(j < i ) o
(j > i ) m
('the endpoint aonditions').
A Combinatorial Problem in Best Approximation
THEOREM: From a basic subfunction {io '· •• ,in+l } with
discrepancy 28, throw out a pair (ik'_l,ik ,) such
that If(ik ,) - f(ik'_l)I = 28. If this is not enough
to determine k' , choose it so as to reduce by as much
as possible the number of steps If(ik ) - f(ik_l ) I
equal to 28. Then the resulting subfunction will
again be basic.
51
PROOF: It is evident that it is a critical subfunction.
The endpoint conditions are also easily checked. In
studying the alleged maximal discrepancy, we have to
compare with an arbitrary competing critical subfunction
{h ,hl, •.. ,h} of the same order m ; here m = n if o m
k' is 1 or n + 1 , and otherwise m = n - 1 •
When can there be an hk which is not among the
i l ? As a representative case, assume hk and i l
both maxima of the respective subfunctions, and hk_l <
Because both subfunctions are critical,
i l lies in an interval where f assumes its maximum
at ~ and also hk lies in an interval where f
assumes its maximum at i l . Therefore f(hk) = feil)
To assign to hk the new value i l would therefore not
52 C. Davis
change the discrepancy we are investigating. Nor wou1d
it spoi1 the relation
needed for critica1ity; because if any j( E ]il,~I
satisfied f(j') < f hk+1) then the retrogression of f
on Ii1 ,il +1] wou1d be greater than
which wou1d contradict the Lemma.
Another case, with hk and i 1 both maxima, is
i 1 ~ hk_2 < hk ~ i 1+1 • The Lemma shows that it can
occur on1y if
= 20
and f(hk) - f(hk_1) = 20. And the preceding paragraph
shows that if it does occur then we are free to suppose
that are among h , •.• ,h • o m
A Combinatoria1 Problem in Best Approximation 53
as
The above reasoning does not e1iminate such a case
h < h = i . but the endpoint conditions with the o 10'
Lemma show that it can occur if
discrepancy {h , •.• ,h } = 25 o m
Summing up, there is no 10ss of genera1ity in
confining attention to competing subfunctions
{ho,.·.,hm} which are obtained from {io , ••. ,in+1 }
by throwing out some pairs of its elements, and then
(in case the discrepancy can be kept at the va1ue 25)
inserting new elements into some of the remaining
interva1s, inc1uding rO,iol and ]in+1 ,N] •
This proves the Theorem in the special case that
min If(ik) - f(ik_1) I is attained for on1y one va1ue of
k. In the contrary case, one more observation remains
to be made: Suppose that, in the process described, the
throwing out of a pair (other than the pair (ik '_l,ik ,» reduces the number of steps ]f(ik ) - f(ik_1)] = 25 by
more than one. Then each of the corresponding insertions
of new hk in the process must have increased it by at
least the same number, otherwise the subfunction initia11y
54 C. Davis
given would not have had maximal discrepancy. I hope
tbe idea here is sufficiently clear that the reader can
check this observation. From it, we see that throwing
out (ik'_l,ik) gives a subfunction satisfying the
Convention.
The Theorem comes close to answering the call for
a simple way of passing from any o(n) to o(n-l) •
Wbat it provides directly is a simple way of passing
from o(n) to either o (n-l) or o (n-2) • The most
pleasing feature is that at each step one refers only
to the subfunction at hand, not to values of f at
any other points. One simply works down from basic
subfunction to basic subfunction, throwing out one
interval at each step, until n is small enough
or until o(n) is too large, if you prefer. At the
end, if it is desired to specify a best approximant,
there is a short computation on the whole domain
{O,I, ... ,N} .
A natural additional assumption (resembling the
hypo thesis of general position in geometry) is that a
non-zero value If(jl) - f(j2)! can equal !f(j3)-
f(j4)! only if the pairs {jl,j2} and {j3,j4} are
the same. Under this assumption, things are still
simpler, and a still more complete description can be
A Combinatorial Problem in Best Approximation
given:
(1) There is abasie subfunetion of a given order if
and only if the eritieal subfunetion of maximal dis
erepaney is unique.
(2) Non-uniqueness arises only in the following way:
55
If {io, .•. ,in+l } is basic, and mink If(ik) - f(ik_l ) I is attained for k r l,n+l , then the eritieal subfunc
tions of order n having maximal discrepancy are
{io,···,in } and {il, .•• ,in+l }.
Proofs are similar to the foregoing.
Given a function not satisfying the assumption,
one way to proceed would be to modify it slightly so
that it did, then deal with this simpler situation.
5. A PROPOSEn EXTENSION
The following generalization of the problem treated
here seems much harder.
PROBLEM: Given v functions fl, .•• ,fv on {O,l, ••• ,N},
to find v functions which wiZZ minimize
56 C. Davis
Hf - g//= max /f (j) - g (j)/ j K K
,K
subjeat to the aondition that no Zinear aombination of
the gK have osaiZZation opdep > n •
REFERENCES
1. Bartho1omew, D.J.: A test of homogeneity of means under restricted alternatives, J. Royal Statist. Soc. Sero B23(1961) , 239-273.
2. Davis, C.: Mapping properties of some Cebysev systems, Dok1. Akad. Nauk SSSR 175(1967), 280-283 = Soviet Math. Dok1. 7(1966), 1395-1398.
3. Kantorovic, L.V. and Rubinstein, G.S.: Aspace of comp1ete1y additive functions, Vestnik Leningrad. Gos. Univ. 13(1958), no. 7, 52-59.
4. Krein, M.G.: The ideas of P.L. Cebysevand A.A. Markov in the theory of 1imiting va1ues of integrals and their further deve1opment, Uspehi Mat. Nauk 6(1951), no. 4(44), 3-120 = Amer. Math. Soc. Trans1. (2) 12(1959), 1-121.
5. Videnskii, V.S.: An existence theorem for the po1ynomia1 with a given sequence of extrema, Dok1. Akad. Nauk SSSR 171(1966), 17-20 = Soviet Math. Dok1. 7(1966), 1395-1398.
GOOD APPROXIMATION BY SPLINES WITR VARIABLE KNOTS
* Carl de Boor
57
Consider approximation of a given funetion f, on
[0,1] say, by elements of Sk , i.e., by poZynomiaZ 1T
spZines 0/ order k (or, degree < k) on some partition
(t.)N+l of [0,1], ~ 0
Rere,
o = = 1 .
tl, ... ,tN are the knots or joints of k
SES , 1T
and their multiplieity, i.e., equality among two or
more of these, indieates redueed smoothness at that
knot loeation in the usual way.
Best approximation to f by elements of
* This work was supported in part by NSF grant GP-07l63
58 C. de Boor
quite we11 understood for a variety of norms since,
after all, Sk is a Zinear space. It seems practica11y 1T
more important and theoretica11y more interesting to
investigate the approximation to f by spZines of opdep
k with N knots, Le., by elements of
where the union is taken over all partitions 1T of
[0,1] with N interior points. For, the approximation
power of sp1ine functions seems to 1ie preciae1y in the
possibi1ity of p1acing the knots in a usua11y quite non
uniform way to suit the pecu1iarities of the given f •
Yet the straightforward approach, vize the construction
k of a best approximation to f in SN' has turned out
to be beset with difficu1ties. It being a somewhat
nasty nonZinear minimization problem, no satisfactory
characterization of a best approximation can be found
in general, see e.g., [4] for the case of Chebyshev
approximation. Consequent1y, any computationa1 scheme
has to be content to find, by some descent method, a
ZoaaZZy best approximation, and even that seems to be
computationa11y quite expensive. Also, the function
f may be "given" in a way that makes the ca1cu1ation
of best knots impossible simp1y because 11 f - s 11
Splines with Variable Knots
cannot be calculated. E.g., f may be the unique
solution of some differential equation
m m-l D f(t) = F(t,f(t), •.• ,D f(t», for t E rO,lJ
with side conditions S.f 1
i=l, .•• ,m
where F, the linear functionals SI' ..• ,Sm and the
numbers cl, .•. ,cm are known, but the value of f at
t is not.
59
For these and other reasons, it becomes important
to search for methods which will produce relatively
cheaply good, if not best, knots for the approximation
of a given function from a variety of information about
this function. And the literature concerning bounds on
seems to be a good place to start such a search.
One approach, taken, e.g., by Freud and Popov [7],
[8], and by Sendov and Popov 116], has been to reduce
the problem of estimating dist(f,S~) to the simpler
problem of estimating dist(g,S~) for given g E eID,l]
60 e. de Boor
making use of a fact such as the fo110wing
LEMMA 1. Fop evepy f € e(k-2)[0,1] , and evepy papti
tion ~ = (t.)N+1 fop [0,1], ~ 0
(1)
with max.ßt. ~ ~
A simple proof of this 1emma goes as fo110ws: On
e[O,l], define the linear map P by
P e[O,l] + Sk f + ~ f( )N l.i T. • k ~ ~ ~,
B-sp1ine basis for with (Ni,k) the norma1ized
(see, e.g., [3]). Since the Ni,k are nonnegative and
add up to 1 at any particu1ar point, it then fo110ws
that
I(f - Pf)(t)1 = ILi(f(t) - f(Ti»Ni,k(t) I
~ max {If(t) - f(Ti) I INi,k(t) # O} •
On the other hand, since Ni,k is nonzero on1y on
(ti,t i+k) , it is possib1e to choose T. ~
in [0,1] so
Sp1ines with Variable Knots
that
for all t E [0,1] , N. k(t) :f 0 1.,
With such a choice, one then obtains
Wf being the modu1us of continuity of f; hence, for
f E C(l)[O,l] , and arbitrary s E Sk ,
therefore
making use of the facts that
and repeated app1ication of this last inequa1ity gives
the estimate (1).
Choosing now, in particu1ar, TI so that
61
62 C. de Boor
and then augmenting TI by at most N - 1 points to
insure that
ITII ~ l/N ,
one obtains from (1) the estimate, valid for
f E c(k-2)rO,1]
The simpler problem of best approximation by
broken 1ines, i.e. in 2 SN ' is taken care of by a
resu1t such as the fo11owing
LEMMA 2. For every g E AC with g' E BV
This can be found, e.g., in [17] as a special case
of a much more general resu1t, but can also be proved
direct1y as fo11ows: If the straight 1ine s inter
po1ates f at the points a < b , then
Splines with Variable Knots
(f - s)(t) = f[a,b,t](t-a) (t-b)
= flb,t]-f[a,t](t_a)(t_b) b - a
63
with f[ro, ••• ,rk] denoting the k-th divided difference
of f at It follows that
sup ](f-s)(t)] < (b-a)/4 Osc[a,b]f' a<t<b
b < (b-a)/4 f ]df']
a
if f E AC and f' E BV , where
Osc[a,b]g = ess.sup g - ess.inf g • [a,b] [a,b]
Hence, if such f is approximated by the broken
2 which interpolates f o = t s E SN_l at < t l 0
li ne
< ... < t 1 < N- t = 1 N , and the t i 's are chosen so that
for some a, then
64
while
(1 - 0) Var(f') N2
C. de Boor
by Jensen's Inequality, sinee I/x is eonvex for
x > o. This proves the lemma.
The two lemmata have the desired
COROLLARY. If f e C(k-2) [0,1] ~ with f(k-2) e AC
and f(k-l) e BV , then
(3)
This is to be eompared with the eustomary statement
that
(4)
in ease f e C(k-l) [0,1] with f(k-l) e AC and
Splines with Variable Knots 65
But, although this improvement of (3) over (4) was
achieved by a particular choice of knots, the argument
has to be suspect since it relies on choosing the knots
so as to produce a good approximation to f(k-2) rather
than to f. Even the more sophisticated argument of
Subbotinand Chernykh [17] (who obtain (3) by construc
ting an approximation to f in the spirit of Birkhoff's
local spline approximation by moments [1], [2], followed
by an appropriate choice of knots so as to make the
error small) excludes consideration of such practically
interesting functions as
f(t) = t a , some a E (0,1)
and therefore does not give, e.g., Rice's startling
result [10] that
(5) for f(t) = t a with 0 < a ,
k -k dist (f,SN) < const kN
00 a,
Rice's argument is a direct verification that for
a certain set of knots selected according to a rule de
pending on a everything works out. In an attempt to
66 c. de Boor
genera1ize Rice's resu1t, H. Burchard [5] proved the
fo11owing intriguing
THEOREM 1. Por f E c(k)[O,1J , and N ~ Nf ' and tor
where
er = er p,k 1/(k + 1/p) •
Simi1ar results have been obtained, for the special
case p = 2 , by Sacks and Ylvisaker [12-15], and more
or 1ess by McC1ure [9], again dealing only with
f E C(k) or even f E C(k+l) , and therefore not giving
Rice's resu1t (5). Nevertheless, these considerations
bring out the importance of the er-norm of f(k) for
a < 1 in the discussion of distp(f,S~) and suggest
that, e.g., (5) ho1ds because, for a f(t) = t
is finite. This is confirmed by the fo11owing
8p1ines with Variable Knots 67
THEOREM 2. Ir f E C[O,l] n c(k)(O,l] , and If(k)(t) I
is monotone deopeasing3 then
This can be proved as fo11ows: Consider approximation
to f k
in 8rr ' where rr has N - 1 distinot points
in (0,1)
say, but each repeated k times. Then
with Tf the piecewise po1ynomia1 function of order k
which, on (ti ,ti+1) , agrees with the Tay10r series for
f around up to terms of order < k , hence, for
1 1f t (k) k-1 I !f(t) - Tf(t)! = (k-1)! f (r)(t-r) dr t i +1
68
1 < -:( =-k -""';l'-:-)-:"'"!
1 .2k!
C. de Boor
The last inequality is easy to prove if, as we assume,
!f(k)! decreases monotonely (- consider both sides as
a function of t i+l and differentiate -) but impossible
to find in the literature. In any event, choose now
the t i 's so that the ßi 's defined above are all
constant,
hence
ß. = ß for all i. Then 1
1 1 1 -k 1 (k) 11 IIf - Tf L.2k! ß = k! N If l/k
which proves the theorem in view of the fact that
distoo(f,S~) decreases with increasing N
A similar result holds for distp(f,S~) with
p < 00 , as proved by D.S. Dodson in [6], where on can
also find the following
Splines with Variable Knots 69
THEOREM 3. If~ foT' some oonveT'ging net (7T) of par>ti
tions of [0,1] ~ the oOT'T'esponding ZoweT' Riemann sums
foT'
1 J If(k)(r)10dr with ° l/(k+l/p)
o
oonveT'ge to A ~ then
foT' some positive oonstant constk independent of (7T)
and f •
These facts and arguments suggest that in approxi-
mating f by elements of k SN ' one should choose the
N knots tl, •.. ,tN so as to make
approximately constant as a function of i. This has
been tried by Dodson 16] in ascheme for the adaptive
solution of an ordinary differential equation. From
a current piecewise polynomial approximation of order
< k to the solution f , he guesses a piecewise
70 C. de Boor
constant approximation g to f(k) , and then se1ects
a new knot set so as to equa1ize over
subinterva1s. To give an examp1e, Russe11 and Shampine
[11] solve the problem
ei"(t) - (2-t2)f(t) = -1 on [-1,1]
with f(-l) = f(l) = 0
for E = 10-8 by co110cation, using sp1ines of order
6 with 47 distinct knots, each of multiplicity 3. The
knots are p1aced on an ad hoc basis so as to pile up
near + 1. They obtain an approximation with error
-6 -4 of 10 near zero, deteriorating to an error of 5.10
near the boundary. Dodson obtains the same accuracy
with 19 distinct knots, and obtains, with 47 knots,
-6 an accuracy of 2.10 even near the boundary (and an
10-8 error in the midd1e of the interval).
REFERENCES
1. Birkhoff, G.: J. Math. Mech. 16(1967), 987-990.
Sp1ines with Variable Knots 71
2. de Boor, C.: J. Math. Mech. 17(1968), 729-736.
3. de Boor, C.: J. Approx. Thy 6(1972), 50-62.
4. Braess, D.: Numer. Math. 17(1971), 357-366.
5. Burchard, H.: "Sp1ines (with optimal knots) are better" , to appear in J. App1icab1e Math. 1 (1972).
6. Dodson, D.S.: Ph.D. Thesis, Comp. Sei. Dpt., Purdue University, Lafayette, Ind. (1972).
7. Freud, G. and Popov, V.A.: Studia Scient. Math. Hungar. 5(1970), 161-171
8. Freud, G. and Popov, V.A. : Proc. Conf. Constr. Thy. Fctns, Hungar. Acad. Sei. (1970) , 163-172.
9. McC1ure, D.E.: Ph.D. Thesis, Div. Appl. Math., Brown University, Providence, R.!. (1970).
10. Rice, J .R.: in "Approximations with special emphasis on sp1ine functions", !.J. Schoenberg ed., Acad. Press, New York (1969), 349-365.
11. Russe11 , R.D. and Shampine, L.F.: "A co11ocation method for boundary va1ue problems", to appear in Numer. Math. (1972/73).
12. Sacks, J. and Y1visaker, D. : Ann. Math. Stat. 37 (1966) , 66-89.
13. Sacks, J. and Y1visaker, D. : Ann. Math. Stat. 39 (1968), 49-69.
14. Sacks, J. and Y1visaker, D. : Ann. Math. Stat. 41 (1970) , 2057-2074.
15. Sacks, J. and Y1visaker, D. : Proc. 12th Bienn. Sem. Canad. Math. Conga 115-136.
72 C. de Boor
16. Sendov, B1. and Popov, V.A.: C.R. Acad. Bulgare Sei. 23 (1970), 755-758.
17. Subbotin, Yu.N. and Chernykh, N.I.: Matern. Zametki 7 (1970), 31-42.
73
SATURATION AND INVERSE THEOREMS FOR SPLINE APPROXIMATION
R. DeVore & F. Richards
The purpose of this note is to examine the con
nections between the smoothness of a function and its
degree of approximation by algebraic polynomial splines
of a fixed degree. Results of this type are known,
usually in the form of an estimate for the degree of
approximation for a certain method of spline approxi
mation in terms of the smoothness of the function.
Estimates like this are customarily called direct
theorems of approximation. Our main interest lies in
the opposite direction, i.e., what inferences can be
made about the smoothness of a function when its degree
of approximation is known.
We say S is a spline of degree k - I if there
are points 0 = Xo < Xl < ••• < xm = I such that on
each interval [xi_l,xi ), i = 1,2, ••• ,m, S is an
74 R. DeVore & F. Richards
algebraic polynomial of degree at most k - 1. The
points X. 1.
are called the knots of the spline. For
generality, we make no restrietion on the continuity of
S at the knots.
If 0 = {O = Xo < xl < ••• < xm = I} , let S(o)
denote the collection of all splines of degree k - 1
with knots contained in 0 Define the error in
approximating f by S(o) as
E 0 (f) = inf" f - s 11 , SES (0)
where 11.11 denotes the supremum norm on [0,1].
o n
Now suppose (0) n
= {O = x(n) < (n) o xl
is a sequence of sets of knots,
< ••• < x(n) = 1} • m
n We let
max l<i<m --n
I (n) (n) I xi - xi _l
and assume that 110 11 -+ O. This guarantees that n
Eo (f) -+ 0 for each f € C[O,l] • n
By ßk we denote the t
th k-- difference operator
Saturation and Inverse Theorems 75
th so that the corresponding ~ order modu1us of continuity
of f is given by
sup Illlk(f,x)II [0,1-kt] o<t<h t
The notation 11.II[a,b] is used to indicate that the
norm is taken over [a,b] When
the norm is understood to be over
[a,b] is omitted,
[0,1] •
A proof of the f0110wing direct estimate for
Eo (f) in terms of wk(f,h) can be found in [3] • n
THEOREM 1. Suppose 0 < a ~ k, f E C[0,1] , and
wk(f,h) = O(ha ) as h + O. Then
(1) Eo (f) = 0(11 on lIa) (n + (0) • n
We shou1d note that the estimate (1) can actua11y be
obtained by using sp1ines
k-2 SEC [0,1]. n
S E S(o ) , with n n
76 R. DeVore & F. Richards
Our main concern is in what sense are the estimates
of (1) the best possib1e? We ask the fo11owing two
questions: When does Eo (f) = O(lIon"a), (n -+00) imp1y n
that a
wk(f,h) = O(h ), (h -+ 0) , Le., does the inverse
theorem to (1) hold? Second1y, is it possib1e to
improve (1) if we assurne higher smoothness for f?
It is not possib1e to answer these questions with
out some additional restrictions on the sets of knots.
The easiest way to see this is when a fixed point, say
1 2 ' appears in each set o n
Then any sp1ine S
which has a single knot at 1 2
will sa tis fy Eo (S) = 0, n
n = 1,2, ••• , but S need not even be continuous. More
generally, the same phenomenon manifests itse1f when a
fixed point on1y falls in sma11 interva1s, in comparison
to 110 11. In order to avoid this, we will require that n
(0) satisfies the fo11owing mixing condition: n
(2) There is a constant p > 0 with the property that
whenever
such that
n > 0 and 1 < i < m -1 ~ there - n (n') (n)
x. < x. (n' )
< x j +1 with J ~
is an n' > n
Saturation and Inverse Theorems 77
min (I x (n) _ x (n ') I I x (n) _ x (n ') I) > pli 0 11 i j , i j+1 n •
It is easy to see that (2) guarantees that the fo11owing
must hold:
(3) Thero is a constant p > 0 with the property that
whenever n > 0 and x E: [O,l-pllö 11] .J then thero is n
(n') (n') an n' > n such that x j < x < x j +1 with
Note that equa11y spaced knots (i.e., ö = (i)n) n n 0
satisfy the mixing condition.
If the mixing condition ho1ds then we can show
that the estimate (1) is the best possib1e in the sense
we have asked.
THEOREM 2. Let (0) be a sequenae of sets of knots n
whiah satisfy the mixing aondition (2) and 11 0 11 -I- 0 • n
If 0 < a ~ k and f E: C[O,l] .J then
(4) E 0 (f) = ° (11 on W), (n + 00) n
if and onZy if
78 R. DeVore & F. Richards
and
(5) Eo (f) = O(llonllk), (n -+ (0) , n
if and onZy if f is a poZynomiaZ of degree < k - 1 •
Remark: The theorem ho1ds without the restriction that
lIonll tends to 0 monotonica11y but the proof becomes
somewhat more cumbersome and hides the essential ideas
invo1ved.
The equiva1ence in (4) is the inverse theorem to
(1). The equiva1ence in (5) estab1ishes the saturation
phenomenon for sp1ines and shows that the estimates (1)
can not be improved by assuming higher smoothness for
the function. Since Theorem 2 is proved with no con
tinuity requirement at the knots, it app1ies to any
sp1ine approximation method provided the mixing con
dition on the knots ho1ds.
Theorem 2 is a1ready known for approximation by
sp1ines with equa11y spaced knots. K. Scherer [6] has
given a proof of this under the additional assumption
that the sp1ines are smooth (i.e., in Ck- 2 [O,1]).
Scherer's proof is based on the general method for
obtaining inverse theorems deve10ped by P.L. Butzer
and K. Scherer [1]. For the saturation parts of
Saturation and Inverse Theorems 79
Theorem 2 with equally spaced knots, independent proofs
have been given by D. Gaier [4] (a "0" theorem) and
F. Richards [5] ("0" and "0" theorems). Our proof of
Theorem 2 is new and quite simple and of course has the
additional advantage of handling non-equa11y spaced
knots. Also, our technique can be genera1ized to give
inverse and saturation theorems for Chebyshevian sp1ines
(see [2]).
PROOF OF THEOREM 2. Because of Theorem 1, we need only
estab1ish the necessity in (4) and (5). We will on1y
consider (4) since the proof for (5) is almost identica1.
First observe that because of the mixing condition (2)
and the assumption that (11 on 11) is monotone, we must
have
(6) 1,2, ...
Now, suppose that S E::S(o), n n
o < CI. < k , and
(7)
with K a constant. We want to show that wk(f,h) =
o (hCl.) , (h + 0) . Choose n so that
80 R. DeVore & F. Richards
(8)
We will on1y consider those h 's for which the index n
1 in (8) satisfies 11 on 11 < 4. This covers all
sufficient1y sma11 h •
Let 0 < t < h • 3 Then for any x E [0'4] , (8)
gives that
(9) [x,X+kt] =- [x,X+kh] =- [x,x+pllon ll] =- [0,1] •
Since (3) ho1ds, there is an n' > n satisfying
(10)
Now, S, n
is a po1ynomia1 of degree at most k - 1 on
(n') (n') [xi ,xi +1 ) Therefore,
Using (7) in the last expression, we find that there is
a constant K1 such that
Saturation and Inverse Theorems
3 x E: [0 '4] , 0 < t < h •
where in the second to last inequa1ity we used (6) and
in the last inequa1ity we used (8).
3 To get the inequa1ity (11) for x E: [4,1-kt] , we
consider the function g(x) = f(l - x). The sp1ines
T (x) = S (1 - x) have their knots contained in n n
6' = {l_x~n) x~n) E: 6} and satisfy n 1 1 n
81
The sequence of sets of knots (6') also satisfies the n
mixing condition (2). Hence, arguing as we have in
obtaining (11), we find that
This shows that (11) is valid for
o < t < h. Therefore, wk(f,h)
necessity in (4) is estab1ished.
o < t < h •
3 x E: [4,1-kt],
< K2h<X and the
82 R. DeVore & F. Richards
REFERENCES
1. Butzer, P.L. and Scherer, K.: Approximationsprozesse und Interpolations-methoden. (Hochschulskripten 826/826a) Bibliograph. Inst. 1968, 172 pp.
2. DeVore, R. and Richards, F.: The degree of approximation by Chebyshevian sp1ines. (to appear).
3. Freud, G. and Popov, v.: On approximation by sp1ine functions, Proc. of the Conference on Constructive Theory of Functions, held in Budapest, 1969, 163-172.
4. Gaier, D.: Saturation bei Sp1ine-Approximation und Quadratur, Numer. Math. 16 (1970), 129-140.
5. Richards, F.: On the Saturation c1ass for sp1ine functions, Proc. Amer. Math. Soc., 33 (1972), 471-476.
6. Scherer, K.: On the best approximation of continuous functions by sp1ines, SIAM J. of Num. Analysis, 7 (1970), 418-423.
SATURATION CLASSES FOR EXPONENTIAL FORMULAE OF SEMI
GROUPS OF OPERATORS
z. Ditzian & C.P. May
83
A C semi-group of operators T(t) on a Banach o
space B into itself can be approximated by formulae
known as exponential formulae [5, p. 359]. The rate
of convergence of some of the exponential formulae in
terms of the moduli of continuity of T(t)f and
AT(t)f ,where A is the infinitesimal generator, was
investigated in [1], [3], and [4]. In this paper we
shall find the optimal rate of convergence, that is,
the saturation sequence, and the class of functions
on which it is achieved, that is the saturation class
for exponential formulae satisfying certain conditions.
These conditions are satisfied by the exponential
This research was supported partly by NRC grant A48l6.
84 z. Ditzian & C.P. May
formu1ae of Rille, Kenda11, Post-Widder and Phi11ips.
Also, it is important to note that the resu1ts depend
on1y on points in (t,t+O) for some 0 (no matter how
small) •
2. PRELINIMARIES AND DEFINITIONS.
Many exponentia1 formu1ae, and actua11y all those
treated in this paper, can be written as:
(2.1) 00
S(T,t)f(·) = f W(T,t,u)T(u)f(·)du o
where T has either the va1ues in {Ti} (Ti + 0+) or
all va1ues in (O,n) , and W(T,t,U) is a positive
kerne1 satisfying the fo11owing:
00
(2.2) Um T-1{! W(T,t,u)du-1} = o , T+O+ 0
00
(2.3) 1im -1 f T W(T,t,U) (u-t)du 0 , T+O+ 0
Saturation C1asses 85
(2.4) 1im .-1 !W(.,t,u) (u-t)2du = pet) I 0, t > 0 , .+0+
and for any positive ö and real a
(2.5) 1im .-1 ! W(.,t,u)eaudu = 0 . • +0+ I u-t I~o
Remark 2.1. It is easy to see that the assumption
CX>
(2.6) ! W(.,t,u) (t_u)4eaudu = 0(.) o
• + 0+ for any real a
imp1ies (2.5). For our particu1ar purposes (2.6) is
eas1iear to verify.
We define the c1ass of functions B2 (t) as
(2.7) B2 (t) :: {f E B, f E V(A2T(t+c:) for all E > O} •
Obvious1y we have for n > 0
(2.8)
86 z. Ditzian & C.P. May
~
n V(A2T(t+11» where 11>0
We define ~
V(A2T(t+11» is defined as the c10sure with respect to
B of V(A2T(t+11» , that is f E B such that for
every 11 > 0 a sequence {f }, n
exists satisfying 1im f = f n
in B and
~
Whenever B is reflexive, B2(t) = B2(t) (as
can be seen fo110wing [2, p. 373]).
3. THE SATURATION RESULT.
We sha11 state and prove the f0110wing saturation
resu1t which we will app1y in section 4 to various
exponentia1 formu1ae.
THEOREM 3.1. Let S«,t)f be defined by C2.1) and
satisfy (2.2) - (2.5), then for f E B
group T(t) of operators, we have:
and a C semio
(A) IIS«,t)f - T(t)fll = 0«)< -+- 0+ for aZZ t > t o
if and onZy if f E V(A2T(to» and A2T(to)f = 0
Saturation Classes
(B) IIS(T,t)f - T(t)fll = O(T)T + 0+ tor alZ t > t
PROOF. To prove sufficiency we recall that if
f E V(A2T(t1» , then for t 2 > t 1 T"(t )f 2 exists
and is equal to A2T(t2) . Therefore for any
t = t 2 > t 1 > t , and sorne o = o(t) > 0 , we have
for It-ul < 0
1 2 2 T(u)f - T(t)f = (u-t)AT(t)f + 2(u-t) A T(t)f
2 + E1(u-t) ,
where IIElll2.E. This irnplies for fE V(A2T(t1»
using (2.2),
2.IIT-1 f W(T,t,u){(u-t)AT(t)f lu-tl<o
122 2 + 2 (u-t) A T(t)f+E(U-t) }du-
- } P(t)A2T(t)fll
87
o
88 z. Ditzian & C.P. May
+ 11 T -1 f W(T, t,u){T(u)f-T(t)f}dull lu-tl >0
Recalling (2.5) and IIT(u) 11 ~ MeClu , 12 = 0(1) •
Combining (~.3) with (2.4) and the above, we get
11 = 0(1) •
Therefore one can easi1y see that f E B2(to)
imp1ies IIS(T,t)f - T(t)fll = O(T) for t > t ~ 0
1f f E V(A T(t1» , then a sequence {fn} exists
fow which fn E V(A2T(t» and 11 A2T (t1)fn 11 ~ M •
Therefore, for t > t 1 T"(t)fn exists for all n
and 11 T" (t) fn 11 ~ M1 ' which imp1ies
Since S(T,t)f (for a fixed T) and T(t)f are
continuous as operators on B, the sufficiency of
the condition in (B) is proved whi1e that of the
condition in (A) is simpler.
Saturation C1asses
To prove that eonditions in (A) and (B) are
neeessary too, we introduee first the operator
2 2 2 l/n l/n fami1y J • J = n f f T(u+V)dudv (see also
n n o 0
[2, p. 502-506]). J2 eommutes with T(t) and n
therefore with S(T,t) •
Fo110wing known eonsiderations [2, p. 505],
J2 f E V(A2), J2 f E V(A2T(t», n n
A2J 2f = n2(T(1)_I)2f n 2
89
and (JL)2T(t)J2f = A2T(t)J2f . dt n n Therefore using (3.1)
on J2f (instead of f) for t > t n 0
1 2 2 -1 2 2 -2 p(t)A T(t)Jnf = s - 1im T [S(T,t)J f - T(t)J f] T~ n n
= s - 1im J 2{T-1 (S(T,t)f - T(t)f} • T~+ n
2 Sinee A is a e10sed operator (see [1, p. 11] for
examp1e) and s - 1im J2f = f , we obtain (in ease n
(A», fo110wing [2, p. 505], f E V(A2T(t» and
A2T(t)f = 0 for all t > t However, this implies o
90 z. Ditzian & C.P. May
T(t )f + T(t +2h)f - 2T(t +h)f for all h ,0 0 2 0 = A2T(~)f
h
= 0 or A2T(t)f = 0 . o
For case (B) we obtain, fo11owing 12] again,
f E A2T(t) , but since this is shown for all
the proof is comp1eted.
t > t o
4. APPLICATIONS TO VARIOUS EXPONENTIAL FORMULAE.
In this section we sha11 state the resu1ts that
fo11ow Theorem 3.1 for various exponentia1 formu1ae.
First we denote the exponentia1 formu1ae as fo11ows:
(4.1) Sl(T,t)f 00 k
= e-t / T L ~, T(kT)f k=o .
for t > 0;
Saturation Classes 91
for t > 0
00 00 2 n n-l = e-At{f e-Au l (A t) U T(u)fdu+f}
o n=1 nl (n-l) I
for t > 0 ;
for t > 0 ;
00 n (4.6) S6(L,t)f = r (t-~) ~nT(a)f
n. L n=o
00 n n = l (t-a) l (_I)n-k(n)T(a+kL)f
n k n=o nIL k=o
for t > a > 0 •
92 Z. Ditzian & C.P. May
We sha11 denote our exponentia1 forrnu1ae Si(T,t)
i = 1,2, ••. ,6 understanding that T for i = 2 or
3 attains on1y va1ues 1 n
(or a subsequence of l) n
THEOREM 4.1. Let f E Band T(t) be a C semi o
gpoup on a Banaoh spaoe B, then fop t > 0 , 0-
A: 11 Sk (T , t) f - T (t)f 11 = 0 (T) fop t < t < t+o 0 0
fop some k if and onZy if f E V (A2T (t )) 0
A2T(t )f = 0 • 0
B: 11 Sk (T , t) f - T (t)f 11 = 0 (T) fop t < t < t +cS 0 0 ~
fop some k if and onZy if f E B2 (to) •
PROOF. The proof is actua11y simp1y verifying (2.2) ,
(2.3) , (2.4) and (2.6). For 1 S3 (u' t) and 1
S4 (-, t)
the estimates are to be found in [4]; other estimates
are also readi1y avai1ab1e or can be ca1cu1ated in a
re1ative1y simple, but somewhat tedious, way. The
function pet) in (2.4) which we will denote as
Pi(t) for Si(T,t) is given by:
Saturation C1asses
(In fact, after some ca1cu1ations, one obtains
S6(T,t) = exp «t-a)AT)T(a) which reduces S6(T,t)
to S1(T,t) operated on g = T(a)f .)
Remark 4.2. In our theorem we actua11y proved that
IISk(T,t)f - T(t)fll = O(T) (or O(T»
for t < t < t +0 o 0
are equiva1ent for various k for which Sk(T,t)
is defined for t E (t ,t +0) • o 0
Remark 4.3. If {T.} is a sequence of rea1s ~
is enough to obtain the necessary and sufficient
condition in (A) (or (B» since we did not use in
Theorem 3.1 more than the va1ues of S(T,t) on a
sequence T. + 0+. ~
93
94 Z. Ditzian & C.P. May
5. COROLLARIES AND REMARKS.
(I) It is tempting to try and follow the proof in
[2, p. 50] and prove Theorem 3.1 or 4.1 using the
estimate of "Sk(T, t ) - T(t )f 11 on1y, that is, at o 0
one point t o
The proof, however, does not fo1low
since neither
T"(t ) o
nor S(t) = s - 1im JL [T(t -h)-2T(t )+T(t +h)] h2 0 0 0
which are 1ike1y to rep1ace A in [5, p. 505], are
c10sed operators. This can be seen using the
examp1e: T(t)f(x) = f(x+t),
f(x) = {~ - x o < x < 1 x > 1
and
{ 1-~-X f (x) = n 0
o < x < 1
x > 1 1-n
f E C , o
1 n , for which
S(1)f (x) = T"(1)f (x) = 0 while S(l)f and n n
T" (1) f do not exis t.
(11) In fact, the difficu1ty expressed in (I) is
not on1y in the proof since an example for which
Saturation C1asses
but 1im" 12 {T(t-h)f-2T(t)f+T(t+h)f}II = co h
can be given as fo11ows:
Let S(. ,t) r
B = C [O,co), o
T(t)f(x) = f(x+t) and f(x) be given by
o ,
00
and f(x) = l f (x) m=3 2m
otherwise
2r 1 Obvious1y, when h = 2- and t = 2'
95
T(t)f = T(i + h)f = 0 and ,,~ (T(t-h)f-2T(t)frT(f+h)fll h
2r+1 1 2r > 2 -- = 2 which is not bounded. We have
2 _2r
96 z. Ditzian & C.P. May
where 11 , 1 2 and 13 are the sums on 3 ~ m < r,
m = rand m > r respective1y.
k1 2r+1 r+1
11 1 I \ (2 k ) 1 22 (k '\ r~l I 11 1 ~suxp k~o (2)·T r+1J l.. f m(x)
22 m=3 2
-1 _2 r - 1 -3 2r - 1 2r+1 where k1 = max {klk ~ (2 -2 +2· )2 }
or, since IIrI1
f m(x) 11 ~ 1 , we have, fo11owing [6; m=32
p.1S (8)] for 1
IX = 4'
11 1 3 11 ~ SUpIS2(S-2r+1,~) I f m/ ~ M sup I If ml x m=r+1 2 x m=r+1 2
2r+1 ~2M2- ,(M= sup IIT(t)lO •
t<l
Saturation C1asses
-1 _2r -3 2r 2r+1 where k2 = max {klk ~ (2 -2 +2 ·)2 } and
supp f ( ~1 + x) are disjoint, and therefore 2r 2
2
2r+1 2r+1 r r 2 "I~I~ e k ) (t)2 ·2-2 2. C(2-2) •
2
(111) For ho10morphic semi-groups we a1ways have
"Sk(T ,t)f - T(t)fll = O(T). In this case
IIS(T,t )f - T(t )fll = O(T) imp1ies, using (3.1), o 0
that A2T(t)f = 0 . o
(IV) If T(t) is a group of operators such that
s - 1im T(t+h) = T(t) , then A = s - 1im T(h)-I h+O h+O h
is a c10sed operator and so is A2 • In this case
11 Sk (T ,t )f - T (t )f" = 0 (T ) if and on1 y if o 0
2 = 0 or A f = 0 ,
97
98 z. Ditzian & C.P. May
~d
Ilsk("to)f - T(to)fll = 0(,) if and on1y if
f E V(A2T(t » = V(A2) • o
Such a situation occurs when B = Co(_~,m) ~d
T(t)f(x) = f(t+x) •
REFERENCES
1. Butzer, P.L. and Berens, H.: Semi-groups of operators ~d approximation, Srpinger-Ver1ag, 1967.
2. Butzer, P.L. ~d Nessel, R.J.: Fourier analysis and approximation, Vo1. I, Birkhauser-Ver1ag, 1971.
3. Ditzi~, z.: On Hi11e's first exponentia1 formu1a, Proc. Amer. Math. Soc., 22 (1969), pp. 351-355.
4. Ditzian, Z.: Exponentia1 formu1ae for semi-group of operators in terms of the reso1vent, Israel Jour. of Math., Vo1. 9, #4, 1971, pp. 541-553.
Saturation C1asses 99
5. Hi11e, E. and Phi11ips, R.S.: Functiona1 analysis and semi-groups, Americam Mathematica1 Society Co11oquium Pub1ications, Vo1. 31, 1957.
6. Lorentz, G.G.: Bernstein Polynomials, University of Toronto press, 1953.
101
ON SOME CONJECTURES OF ASKEY CONCERNING COMPLETELY
MONOTONIC FUNCTIONS
J.L. Fields & M.E. Ismail
INTRODUCTION.
In [1], Richard Askey analysed the LP conver
gence of the Lagrange interpolation polynomials when
the zeros of the classical Jacobi polynomials,
p(a,ß)(z) , are used as the points of interpolation. n
His analysis was complete, except for some results
concerning the positivity of the Cesaro means of some
order y, (C,y) , for the Poisson KerneI,
1 K = f
n -1
1 a,ß ~ - I' 0 < r < 1 •
102 J.L. Fields & M.E. Ismail
Taking into account iso la ted results of this nature,
(e.g. Fejer [4] showed that for
(C,2) means of P (x,y) r
are positive, while
the
Kogbetliantz [10] showed that for 0. = ß , the (C,2a.+2)
means of P (x,y) are positive) Askey [2] made the r
following conjecture.
CONJECTURE 1.
positive for
The (C,a. +ß+2)
-1 a.,ß ~ '2 .
means of P (x,y) r are
See [2] for a more complete list of when this and the
following conjectures are true.
Using some results of Gasper [8, 9] on convolution
structures, Askey pointed out that it is actually
sufficient to prove Conjecture 1 with y = 1. Thus,
Conjecture 1 is equivalent to the following.
CONJECTURE 1'. If A = a.+ß+l , and
G(t, z) -().+1) ( ),+1, ),+2
= (l_t2) F 2 2 2 1
ß + 1
()C)
= l: g (z)tn n n=o
4zt ) (l+t) 2
Completely Monotonie Funetions
then
g (z) > 0 n
o < z < 1
103
Assuming the validity of Conjeeture 1 (or 1'),
Askey [2] dedueed several other related results.
Stated as eonjeetures, they are as foliows:
CONJECTURE 2. If
H(t,z) = (1_t)-2Ä[1+2t(2z_l)+t2]-Ä ,
then
co
= L h (z)tn n n=o
h (z) > 0 n
o < z < 1 Ä > 0 •
Remark: This eonjeeture is known to be true for
Ä = t and follows from Kogbetliantz's for Ä > 1 •
CONJECTURE 3. If 4y = 2a+3 , then fop x > 0
104 J.L. Fie1ds & M.E. Ismai1
1 --4y () 22 r (2y+1)x6y- 1 2y I 2
r(n+1)r(6y) 1F2 1 -~ 3y,3Y+2
~ 0, 2n > -1 or 2y ~ 1 •
Remark: This eonjeeture is known to be valid for
2y = 1,2, ••••
CONJECTURE 4. If 2y = 2n+1 , then fop x > 0
o
( y I 2) -x 1'4 '
2y, 2y~
Remark: Quite reeent1y, Askey and Po11ard [3] have
estab1ished this eonjeeture for y > 0 •
By Bernstein's Theorem on eomp1ete1y monotonie
funetions, Conjeetures 3 and 4 are equiva1ent,
respeetive1y, to the fo11owing:
CONJECTURE 3'. x-2Y (x2+1)-2y is eompZeteZy monotonie
Comp1ete1y Monotonie Funetions
fop 2y ~ 1 and x > 0 •
CONJECTURE 4'. x- 2Y (x2+1)-Y is aompZeteZy monotonia
fop y > 0 and x > 0 •
In Theorem 1, we prove a resu1t whieh ine1udes
Conjeetures 2 and 4. Conjeeture 4 is further genera-
1ized in Theorem 2. Conjeetures 3 (and 4) ean be
proved using asymptotie methods deve10ped in [5], but
the proofs are 10ng, and will appear e1sewhere.
THEOREM 1. If
then fop aZZ n,
p (z) n
r(n+2pA) (-n,n+2PA,AI ) = -r"""'(n-+;:""l-:-)-r'-:-( 2""'::P-A"'-) 3 F 2 1 I z ,
PA ,PA+z
> 0 P > 2, A > 0, 0 < z < 1 .
and
105
106 J.L. Fields & M.E. Ismail
lim nl-2PApn(~21 = ----,-1_.,- F r(2pA) 1 2 n-+o:l n
( A ) 1 -z , PA 'PHzl
> 0 p > 2 A > 0 z reaZ.
PROOF. The identification of p (z) follows from the n
general formula ([7])
A A+l
(2'-2-' a l , ... ,ap )
(l-t)-A F ! -4tw p+2 q 2
bl, ••• ,bq (l-t)
CX>
r(n+A)tn F (-n,n+A,al, ••• ,ap ! ) r(A)n! P+2 q b b w
1'· .. , q = L
n=o
q ~ p+l, Itl < 1 ,
and the identification
F(t,z) = (1_t)-2PA F (AI -4tz ) 1 0 (1_t)2
A trivial modification of the Askey and Pollard
proof [3] yields the non-negativeness of the p (z) • n
Completely Monotonie Funetions 107
Set
q(t) = -2(p-l) log (l-t) .- log [1-2t eos 8+t2] ,
l-2z = eos 8 •
Then
q' (t) 2(p-l) + 2 eos 8-2t = , l-t i8 -i8 (t-e )(t-e )
2(e-l ) -i8 i8
+ e + e = l-t -i8 i8 I-te I-te
00
= L {2(p-l) + 2 eos [ (n+I)8]}tn > 0 p > 2 • n=o
Thus q(t) is absolutely monotonie (q(O) = 0) ,
so is F(t,z) = eAq(t) , whieh implies p (z) > 0 n -
all n. The last statement of the theorem then
follows from the non-negativeness of the p (z) n
a result in [6] eoneerning eonfluent limits.
Remark 1: One might be tempted to try to prove
Theorem I for
from [7], that
3 p = - and 2
A > I , but it follows
and
for
and
108 J.L. Fie1ds & M.E. Ismai1
p (1) = 2-2AN2PA-2A-1{ 1 + O(N-2)} n r(2pA-2A)
+ (_1)n22A-2PAN2A-1 { 1 + O(N-2)} , r(2A)
which can be negative for p < 2 , imp1ying that p ~ 2
is a necessary condition for
on 0 < z < 1 .
p (z) n
to be non-negative
Remark 2: In [3], Askey and Po11ard show that the
p = 2 resu1t leads to a simple proof of the
Kobet1iantz resu1ts concerning the positivity of the
(C,2a+2) means of the Poisson kerneI.
A final genera1ization of Conjectures 2 and 4 is the
following.
THEOREM 2.
( -n,n+2pA ,p-a I )
3F2 z pA+b,PA+C
> 0 o < z < 1 .
F (A-a \_112) > 0 1 2 pA+b,PA+C 4
11 reaZ"
o < a < A, 0 < b 1 "2 2. c, p > 2 A > 0 •
Comp1ete1y MOnotonie Funetions
PROOF. Consider the Beta transform
g(t) (-n.n+2PA ;AI 2) = 3F2 zt
C,D o.+1,ß+1 > 0
With
g (t) rn .... 20A .A I 2) = 3F2 zt 0 1 p). ,p).~
o 2 t,z ~ 1
0.+1 = 2p)., ß+1 - b, b > 0 ,
one obtains
> 0
App1ying the above transform to
0. = 2p). ß + 0.+3 = p).+e , 2
o ~ t,z ~ 1 .
with
1 e >-2
109
110 J.L. Fields & M.E. lsmail
one obtains,
_ ( -n, n+ 2p ;\. ,;\. I 2) - 3F2 zt
p;\.+b,p;\.+c :> 0 , o ~ t,z ~ 1 •
Finally, if the Beta transform is applied to g2(t)
with
a + 1 = 2;\' - 2ß - 2 = 2;\' - 2a, 0 < a < ;\. ,
one obtains the first statement of the theorem with
strict inequalities. The general theorem is then
obtained by taking limits. The same sequence of
transformations yields the statement concerning the
lF2 'so This result can also be obtained by taking
the confluent limit of the 3F2 polynomials, see [6].
REFERENCES
1. Askey, R.: Mean convergence of orthogonal series and Lagrange interpolation, Acta Math. (Budapest), to appear.
2. Askey, R.: Summability of Jacobi series, to appear.
Comp1ete1y Monotonie Funetions
3. Askey R. and Po11ard, H.: Some abso1ute1y monotonie and eomp1ete1y monotonie funetions, SIAM Journal Math. Anal. to appear.
4. Fej~r, L.: Neue Eigenschaften der Mittelwerte bei den Fourierreihen, J. London Math. Soe. 8 (1933), 53-62, Gesammelte Arbeiten 11, 493-501.
5. Fie1ds, J.: A uniform treatment of Darboux's method, Areh. Rat. Meeh. and Anal., 27 (1968), 289-305.
6. Fie1ds, J.: Conf1uent expansions, Math. Comp. 21 (1967), 189-197.
7. Fie1ds, J.: Asymptotie Expansions of a C1ass of Hypergeometrie Polynomials with Respeet to the Order 111, J. Math. Anal. App1., 12 (1965), 593-601.
8. Gasper, G.: Positivity and the eonvo1ution strueture for Jaeobi series, Anna1s of Math. 93 (1971), 112-118.
111
9. Gasper, G.: Banach a1gebras for Jaeobi series and positivity of a kerne1, Anna1s of Mathematies 95 (1972), 261-280.
ID. Kogbet1iantz, E. Recherehes sur 1a sommabi1it~ des series ultra spherique par 1a m~thode des moyennes arithmetiques, Jour. de Math. pures et app1iques (9) 3 (1924), 107-187.
UNE APPLlCATION DE LA THEORIE DE L'APPROXlMATION A
L'ETUDE DES FONCTIONS HOLOMORPHES
* P.M. Gauthier
INTRODUCTION.
Recemment J. Clunie [3] a d~montr~ Ie resultat
remarquable qu'il existe des fonctions meromorphes
distinctes f et g dans Ie disque unite satisfai
santes a
(1) Iim X(f(z),g(z» = 0 , Izl~I
ou X designe Ia distance sur Ia sphere de Riemann.
Depuis P. Lappan [5] et H. Allen [1] ont obtenu fonc
tions holomorphes jouissantes des memes proprietes.
* Subventionne par Ie C.N.R. du Canada et par Ie Gouvernement du Quebec.
113
114 P.M. Gauthier
Dans cette note, en nous appuyant sur 1e th~oreme
dtArak~lian sur 1 tapproximation par fonctions ho10mor
phes, nous donnons une demonstration extremement courte
de ces r~su1tats. Aussi nous a110ns r~pondre A une
autre question pos~ par M. C1unie 10rs d'une visite A
Montr~a1 en 1971.
Designons par D 1e disque unite dans 1e plan
* comp1exe, par D 1a compactification Dt A1exandrov,
par E un (relativement) ferme dans D, par C(E)
1es fonctions continues (A valeurs comp1exes) sur E
par CH(E) 1es fonctions continues sur E et ho10-
morphes Altinterieur EO de E, par H(D) 1es
fonctions ho10morphes dans D, et par H(D) 1es
1imites uniformes sur E de fonctions dans H(D) •
A10rs evidemment on a
1e ce1ebre theoreme d tArak€lian nous dit:
THEORtME A. (Voip 12]J. H(D) = CH(E) si et seuZement
* si D \E est connexe et ZocaZement connexe.
Approximation des Fonctions Holomorphes
1. COMPORTEMENT A LA FRONTIERE
Par des techniques astucieuses du genre Wiman-
Valiron, M. Clunie demontre:
THEOREME B. Il existe des fonations h et k
morphes dans D telles que
(2) lim Ih(z) I + Ik(z) I = 00
Izl~l-
De la, il deduit en quelques lignes:
THEOREME C. Si h et k sont les fonations du
Th~or~me B, alors les fonations
f = k/h et g = (k-l)/h
satisfont ~ (1).
holo-
115
Notre raccourcissement consiste a donner une courte
demonstration du Theoreme B. Pour cela nous avons
besoin d'un theoreme ayant un certain interet en lui
m~me. Par un domaine spirale E nous entendons un
ensemble de D dont la frontiere (dans la topologie de
D) est une courbe simple a(t), _00 < t < +00 , satis
faisant au conditions:
la(t) I ~ 1 et arg a(t) ~ +00 ,
lorsque Itl ~ 00 •
116 P.M. Gauthier
~ , THEOREME 1. Soit E un domaine spiraZe. AZors iZ
existe une fonation G 3 hoZomorphe dans D 3 teZZe que
(3) Re G(z) + +00, lorsque Izl + 1 dans E.
Ce theor~me entratne 1e Theor~me B. En fait, soit
EI et E2 deux domaines spira1es dont 1a reunion est
D. Soient G et k 1es fonctions correspondantes a EI et E2 respectivement d'apr~s 1e Theoreme 1, et
posons h = exp G. A10rs h et k satisfont a (2)
et en plus h n'a pas de zeros. Donc pour ce choix 1e
Theoreme C nous donne deux fonctions ho1omorphes
satisfaisantes a (1).
Nous remarquons que 1e Theoreme 1 genera1ise un
theor~me de W. Schneider [6] qui dit qu'i1 existe une
fonction G ho1omorphe et non-bornee dans D, mais
bornee dans D\E.
Demonstration du Theoreme 1. Soit F une repre
sentation conforme de EO sur 1a demi-bande
Re w > 0, 11m wl < 1 ,
teIle que
F(z) + 00, lorsque Izl + 1 dans E.
Approximation des Fonctions Ho1omorphes
Par 1e Th~oreme A d'Arake1ian, i1 existe une
fonction G ho1omorphe dans D et te11e que
IG(z) - F(z) I < 1, z E E •
117
A10rs G satisfait ~ (3) et 1e Th€or~me 1 est d~montr~.
"-
2. UN PROBLEME DE CLUNIE
Nous rappe10ns un th€or~me de R. Remmert (non
pub1i€) •
~ "-
THEOREME D. IZ existe trois fonctions g,h,k hoZo-
morphes dans D dont une injective, teZZes que
Ig(z) I + Ih(z) I + Ik(z) I ~ 00, Zorsque Izl ~ 1 .
En vertu des Theoremes B et D, M. C1unie posa 1e
probleme a savoir si on peut supposer h injective dans
1e Theoreme B. La reponse est non. En fait supposons
par contradiction que h est injective. A10rs si f
et g sont 1es fonctions du Theoreme C, f - g = l/h
est injective et donc ades 1imites angu1aires non
nu11es presque partout sur 1e cercle unite (voir [4,
p. 56]). Soit p un tel point du cercle unite et
{z} une suite qui tend vers p dans un angle de n
Stolz. Puisque f et g satisfont a (1) i1 en suit
118 P.M. Gauthier
que si fest borne sur {z } n
a10rs f - g tend vers
zero sur {z} ce qui est exc1u. On en conc1ut que n
f a 1a limite angu1aire infinie en p et donc f a
limite angu1aire infinie presque partout ce qui est
absurde [4, p. 146].
, , REFERENCES
1. Allen, H.: Distinct ho10morphic functions with identica1 boundary va1ues (a paraitre).
2. Arake1ian, N. U.: Approximation comp1exe et proprietes des fonctions ana1ytiques. Actes, Congres intern. Math. , 1970 , Tome 2, 595-600.
3. C1unie, J.: On a problem of Gauthier, Mathematica 18 (1971), 126-139.
4. Co11ingwood, E.F. et Lohwater, A.J.: The theory of Cluster sets. Cambridge University Press, Cambridge, 1966, MR 38 #325.
5. Lappan, P.: A note on a problem of Gauthier. Mathematika 18 (1971), 274-275.
6. Schneider, W.J.: An e1ementary Proof and extension of an examp1e of Va1iron (a paraitre).
LINEARIZATION IN CERTAIN NONCONVEX MINIMIZATION
PROBLEMS AND GENERALIZED SPLINE PROJECTIONS
Joseph W. Jerome
1. INTRODUCTION.
The problem of minimizing the curvature, in the
L2 norm, of smooth functions f with square inte
grable second derivatives, subject to certain inter
polatory constraints on f, leads directly to the
consideration of constrained nonconvex minimization
problems in a Hilbert space and, more generally, in
a Banach space. In [6] and [7] an existence and
convergence theory was developed for the solutions of
extremal problems of the form
Research supported by National Science Foundation Grant GP-32116.
119
120 J.W. Jerome
(1.1) IITs 11 = min IITul1 'p UEU p
where T is a (possibly nonlinear) mapping of the
real Sobolev space wrn,P(a,b) into LP(a,b),
1 <p.::."" and U is a closed convex subset of
wrn,PCa,b) these spaces will have their usual meaning
throughout. In 16], a thorough analysis of the problem
of minimum curvature was made and a necessary condition
was given for solutions s in terms of the non
negativity of the Gateaux derivative of the objective
functional at s on the cone {u: u + s E U} from
which the nonlinear differential equation satisfied
locally by solutions and the global smoothness
properties of solutions were deduced. Such results
are valid in the more general setting as we describe
in §2 of the sequel for 1 < p < "" and yield conditions
which are also sufficient if the objective functional
is convex as is the case, for example, when T is a
linear operator, a case treated by Golomb [3]. In §2,
we also discuss certain convex minimization problems
"" in Land the spline-type analytical character of
solutions. However, in §3, our results go beyond this
to include a characterization of Kuhn-Tucker type for
the solutions of classes of nonconvex minimization
Linearization and Spline Projections 121
problems, including problems of the form (1.1), in
Banach spaces. The nonnegative objective functionals
considered are of the form ~.~ where ~ is Gäteaux
differentiable and ~ is a seminorm on a suitable
space. Dur characterization theorem, which includes
both necessary and sufficient conditions for a
solution, makes fundamental use of results contained in
the book of Pshenichny; [11]. This analysis is carried
out in §3 and an application to problems of the form
(1.1) for 1 ~ p ~ 00 is made at the close of this
section in Corollary 3.3. This analysis is intended
as a preliminary step to the problem of obtaining an
algorithm for the determination of solutions of (1.1).
In §4, we consider the problem of constructing
generalized interpolating spline functions which are
obtained as projections via abilinear form
(1.2) B(u,v) =
on the Sobolev space wrn,2(a,b) , where
00
b _ 1 mm and
b .. E L (a,b) . No assumptions whatsoever are made 1J
concerning the nonnegativity of B(u,u) or the
symmetry of B(u,v). In fact, the spline solution
122 J.W. Jerome
may fail to minimize B(u,u) in any suitable way.
However, we obtain a characterization theorem, valid
in all cases, which leads to the interesting result
that any nonsingular linear differential Euler operator
A has associated interpolating spline functions,
locally in the null space of A. Previous results of
this type required either direct or indirect (e.g.
A = L*L) assumptions on the nonnegativity of B.
Essential use here is made of elegant results of Aubin
[1], particularly a generalized integration by parts
formula, which yields the characterization theorem.
Although the notion of a spline as a projection is
quite old, the consideration of general bilinear
forms as the context for such projections is quite
recent (cf. [8] and [10]). The present work should
be viewed simply as an extension of the work of [8]
and [10] (the former in the case of nonsingular
problems) in which the universal existence of genera
lized spline functions is deduced for any bilinear
form of the type (1.2), thereby eliminating nonnega
tivity assumptions. This result is contained in
Theorem 4.5 and may be considered independently of the
remainder of the paper. On the other hand, it is
related to the expansion, through quadratic terms, of
the nonconvex objective functionals mentioned earlier.
Linearization and Sp1ine Projections
The quadratic term in such express ions is in general
not globa11y nonnegative, as in the convex case.
Another interesting aspect of the projection approach
is that we also obtain a c1ass of sp1ines which is,
in a sense, dual to the usua1 c1ass, i.e., we obtain,
a1ternate1y, sp1ines satisfying inhomogeneous forced
boundary conditions and homogeneous natural boundary
conditions on the one hand and sp1ines satisfying
homogeneous forced boundary conditions and inhomo
geneous natural boundary conditions on the other.
Such sp1ines are re1ated by so-ca11ed comp1ementarity
princip1es. MOre generally, we are ab1e to define
sp1ines which satisfy nonhomogeneous forced and
natural boundary conditions. In the particu1ar
123
examp1e of piecewise cubics, this means that the sign
of the jumps in the second derivative can be contro11ed
at noda1 points, thus producing 10ca11y convex or con
cave interpo1ants, if desired. Theorem 4.6 contains
the precise statement of this in general setting.
Notice fina11y, that by producing sp1ines with
specified jumps in higher order derivatives, we can
direct1y construct Green's functions for multipoint
boundary va1ue problems.
124 J.W. Jerome
2. ANALYTICAL CHARACTERIZATIONS.
In this section we examine the structure of
solutions of certain minimization problems in
LP(a,b), 1 < P ~ 00. The results discussed here for
00
the L (a,b) case are expository in nature and are
quoted from [2]. An interesting consequence of these
results (cf. Theorem 2.5) is that the minimization
problem
= inf {IIDmf:f E tf1,oo(a,b):f(x.) = r. 1 1
o < i < n}
for prescribed a = x < ••• < x = b o n
and r , ... , r o n
has a unique spline solution of degree n (there may
be other solutions however). The results discussed
for LP(a,b), 1 < p < 00 , are new, however, and
enable one to deduce, in certain cases, that the
solutions of (1.1) satisfy locally a nonlinear
differential equation of the form (cf. Theorem 2.1)
(2.2)
where T is a nonlinear differential operator of
order m, DT(s;') is a linear differential operator
of order m which is the Gateaux derivative of T at
Linearization and Spline Projections
* sand DT(s;·) is its formal adjoint. For the
case T = Dm this result was obtained by Golomb [4].
We consider first the case of p for 1 < p < 00
Let T be a mapping of wm,P(a,b) into LP(a,b)
given by
(2.3) m-l m Tf = X ( • ,f ( • ) , .•. ,D f ( • ) ) D f
m-l + w(·,f(·), ... ,D f(·»
125
where X and ware continuous real-valued functions
on [a,b] x ~. We have the following
THEOREM 2.1. Let U be any alosed aonvex subset of
Wm,P(a,b). Let T: wm,P(a,b) + LP(a,b) be given by
(2.3). Then the minimization problem
(2.4) IITSlb = inf IITfl1 = Cl
fEU
has a solution s E U provided there is a bounded
minimizing sequenae in wm,P(a,b) , i.e., a bounded
sequenae {f} c U suah that IITf 11 + Cl. If, n n
moreover, the funations X(T,E; , ••• ,E; 1) o m- and
w(T,E; , ••• ,E; 1) have partial derivatives with respeat o m-
126 J.W. Jerome
to ~ , .•. ,~ 1 whiah are aontinuous on [a,b] x ~ o m-
then the mapping T has a Gateaux diffepential
DT(f ;.) at evepy f € wm,P(a,b) whiah is a o 0
bounded linear diffepential opepatop mapping wm,P(a,b)
into LP(a,b) given by
m-1 m + x(·,f (·), •.. ,D f (·»D f . o 0
The funational
has a GG.teaux diffepential D8 (f ;.) at every o
f € Wm,P(a,b) whiah is a aontinuous linear funational o
on wm,P(a,b) given by
(2.6) D8(f ;f) o
fop eaah f € wm,P(a,b) FinaZZy, if s is any
solution of the extpemal ppoblem (2.4), then
Linearization and Sp1ine Projections 127
ns(s;e) ~ 0 for aZZ e in the convex support cone
(2.7) K = {e e = A(f-s) for some A > 0 and some f € U}
i.e . .J b -1
(2.8) f ITsl P signum TS·nT(s;e) > 0 for atz e € K • a
ConverseZy.J if slip is convex.J as is true when T
is Zinear.J then (2.8) is sufficient for s to be a
soZution of (2.4).
PROOF. The existence resu1t is a consequence of
Theorem 2 of [7]. (2.5) is a consequence of the
multidimensional Tay10r theorem [3, p. 57]; indeed,
for each fixed T € [a,b] if we set
(2.9) s (~ , ... ,~ ) = X(T,~ , ... ,~ 1)~ Tom 0 m- m
+ W(T,~ , ... ,~ 1) , o m-
then we have
(2.10) s (~+tn , ••. ,~ +tn ) = s (~ , ... ,~ ) TO 0 m m TO m
m + o(t .I Inil) as t ~ 0
1.=0
128 J.W. Jerome
where the order expression is uniform in T E [a,b]
Thus, choosing and i n. = D f(T) . and 1
using (2.9) and (2.10), we deduce that DT(f •• ) 0'
exists as a bounded linear operator from wm,P(a,b)
into LP(a,b) given by (2.5). (2.6) results from an
application of the chain rule to the composition
mapping 8 = 11· IIP • T Now if s is any solution
(2.4) and e E K = K , then, for all sufficiently s
small A > 0 we have s + Ae E U and, hence,
Letting A tend to zero yields
D8(s;e) > 0 for all e E K .
The converse implication, when is convex,
follows from well-known results of convexity [11,
Theorem 2.1] if Ts I 0 and is trivial if Ts = 0 •
COROLLARY 2.2. Let T be a mapping from Wm,P(a,b)
into LP(a,b) given by (2.3) suah that X and w
have aontinuoUB partiaZ deriviatives with respeat to
~ , ... ,~ 1 on [a,b] x Rm • Let U be a nonempty o m-
of
Linearization and Sp1ine Projections
aonvex subset of wm,P(a,b) aonsisting of funations
f satisfying arbitp~ affine inequaZity aonstpaints
at nodaZ points a=x < ••• <x =b on linear o n
aombinations of depivatives of f thpough opdep
m - 1. If s is a soZution of the minimization
ppobZem (2.4), and, if the fopmaZ adJoint [DT(s;.)]*
of DT(s;o) exists as a Zinear diffepentiaZ opepatop
of opdep m with integpabZe aoeffiaients, and, if
(2.11) (m-1) x(·,s(o), ••• ,s (.» > C > 0
on [a,b], then ITsl P- 1signum Ts is in
m-1 C (xi ,xi +1) with (m-1)th dePivative absoZuteZy
aontinuous fop eaah i = 0,1, ••• ,n-1 and
(2.12) a. e. .
Mopeovep, aeptain highep opdep aontinuity aonditions
ape satisfied by ITsl P- 1signum Ts aaposs intepiop
nodaZ points x1 ' ... ,xn_1 ; speaifiaaZZy if s
satisfies affine HePmite-type aonstpaints of the foPm
129
j = 0, ... ,k.-1, ~
o < i < n ,
then
130
(2.14) j =
J.W. Jerome
m, ••. ,2m-k.-1 • 1
PROOF. The assumption (2.11) ins ures that DT(S;·) is
a nonsingular linear differential operator of order m
with continuous coefficients on Ia,b] . Now, if
and Xi+1 are adjacent nodes and if ep is an
infinite1y differentiab1e function with compact
support in (xi ,xi +1) then ep and -ep are in the
convex support cone K and it fo110ws from (2.8)
that
x i +1 -1 J {ITsl p signum Ts}DT(s;ep) = 0
xi
so that ITsl P- 1signum Ts is a solution, in the
sense of distributions, of
(2.15)
x . 1
* Since [DT(s;·)] is a nonsingular
linear differential operator of order m with
integrab1e coefficients, every distribution solution
is a c1assica1 solution in the sense described in the
statement of the theorem [5, Chapter 8]. The final
statement is a consequence of an argument given in
Linearization and Spline Projections 131
even more general form in 19, Chapter 7].
We remark that the Euler equation (2.12) for the
special case T = Dm was obtained in the case of
Lagrange-Hermite interpolation by Golomb [4] who also
obtained higher order continuity conditions across the
nodes and lower degree of the extremal solution at
infinity. Although he stated these only as necessary
conditions it is clear that they are also sufficient
since they lead to (2.8) for a convex minimization
problem.
In the remainder of this section we consider the
minimization of 11 Lfll, L a nonsingular linear p
differential operator of order m for 1< p ~oo ,
where f is subject to so-called extended Hermite
Birkhoff constraints. Consider then a mesh
••• < x = b n
of [a,b] and, associated with each of the points xi'
consider the continuous linear functionals L.. on 1.J
wrn,P(a,b) defined by
(2.16) m-l
L .. f = I a~~)f(\))(x.) 1.J \) =0 1.J 1.
j
i=O, ... ,n,
132 J.W. Jerome
(v) for prescribed real numbers a.. such that, for each
~J
(0) (m-l) i , the k. m-tuples (a .. , .•. ,a. j ) are linearly ~ ~J ~
independent; here 1 < k < m for i- i = O, .•. ,n and,
at x and x , the derivatives are taken in the o n
limiting sense.
Let L be a nonsingular linear differential
operator of the form
(2.17) m-l .
L = Dm + l c.DJ j-o J
where c. E C[a,b] • Now let r .. , j J ~J
O, ••• ,k.-l, ~
i O, ... ,n be prescribed real numbers. Consider,
for 1 < P 2 00 , the minimization problem
(2.18) IILs I~ = a = inf {"Lf I~ fE~,P(a,b)
L .. f ~J
r .. ~J
j = O, ... ,k.-l, i = O, ... ,n} . ~
THEOREM 2.3. The minimization probtem (2.18) has a
sotution for 1 < P 2 00. Por fixed interpotation
vatues r .. the ctass S of sotutions is a convex ~J
Linearization and Sp1ine Projections
set. Among aU soZutions in S there exists a
soZution s* with the property that on each sub-
intervaZ ° ~ i ~ n-1 , the L P
norm of
is minimaZ in the foZZowing sense: Let S = S 1
133
and, for 2 < i < n , Zet S. consist of aZZ soZutions - ~
to the minimization probZem
SES. 1} ~-
then S n
is nonempty, i.e., there is an s* in S n
PROOF. For p = 00 , this is Theorem 1 of [2]; the
proof there, however, i8 valid for 1 < P < 00 also.
Now for each fixed i=O, ... ,n let A. ~
~x m matrix
A. = (a~~» ~ ~J
where j denotes row and v co1umn indices.
be the
Let Ä. ~
be any nonsingular m x m argmentation of Ai' Let
H. ~
be the inverse of the transpose of A
A. ~
134 J.W. Jerome
If operators o are defined on suitab1y smooth v
functions by
(2.19) 0 f = v
for v = O, .•• ,m-1 and if operators Rij
by
are defined
(2.20) j = 0, •.• , m-l , i = 0, ••• , n
then [cf. Lemma 3.1], if the notation
by
[.]. is defined 1
and
[~]. = ~(X.+) - ~(x.-), for 0 < i < n , 111
[ ~ ] = Hx +), [ ~ ] = Hx -) , o 0 n n
we have, for i = O,l, .•• ,n ,
(2.21) m-1 m-1 l njg(x.)[Ojf]. = l L.jgIR .. f]i
j=o 1 1 j=o 1 1J
b * = f [LfLg - L Lf·g]
a
Linearization and Sp1ine Projections 135
for all f,g for which (2.21) is meaningfu1. Notice
that A, ~
induces operators L .. ~J
for k < • < m-1 • i - J
m THEOREM 2.4. Suppose cj E C Ia,b] and 1 < p < 00 •
Then s is a solution of the minimization problem
(2.8) if and only if s E wrn,P(a,b) and
(i)
(2.22) (ii)
(iii)
on
i = 0, ... ,n-1 ,
L, ,s ~J
r, , ~J
j = 0, ... ,k.-1, i = O ... ,n ~
[R .. s]. = ° if j ~J 1
k. , ... ,m-1, ~
i = O, ••• ,n ,
PROOF. The direct imp1ications are a consequence of
(2.8), (2.12) and the integration by parts formu1a
(2.21). Converse1y, if (2.22 i, ii, iii) are satisfied,
then, using (2.21), we deduce that
b -1 J rlLsl P signum Ls]Le °
a
for all e such that L. ,e ~J
0, j O, ... ,k,-l, ~
136 J.W. Jerome
i = O, ... ,n. We conc1ude from Theorem 2.1 that s
solves (2.18).
Before stating the final theorem of this section,
we require a 1exicographic ordering of the Lij , i.e.,
i if N = 0 and N. L k i = 0, ... , n with 0 1. \)'
\)=0
N N , we define n
= L .. 1.J
° < j < k.-1, i = O, .•. ,n • - - 1.
We make the fo110wing assumption:
(I) N ~ m + 1 ; for each \) = 1, .•. ,N-m+1 the
functiona1s L, .•. ,L +m 1 are consistent with respect \) \)-
to the null space of L, i.e., for any prescribed
nmnbers r\), .•. , rv+m_1 there is a unique function u
in the null space of L satisfying L.u=r., 1. 1.
\) < i < v+m-1 . We now quote the fo110wing theorem
from [2] for p = 00
THOEREM 2.5. Suppose that (I) is satisfied and that
L* exists with continuous coefficients and has the
property that its nuZZ space is spanned by a
Linearization and Spline Projections
Tahebyaheff system. Then thepe is a fundamental,
intepval, J = [x ,x J with 0 ~ vI < v2 ~ n and vI v2
v2 l ki ~ m + I suah that any two sol,utions of (2.18)
:i:= vI
fop p=
ILs 1 = Cl
'00 agpee on J. Mopeovep, i f s € S then
a. e. on J, whepe S denotes the aonvex
sol,ution set of (2.18). If s* € S is ahosen as in
Theopem 2.3 then s* is unique in S. Mopeovep,
137
s* satisfies the ppopepty that ILs*LI is equival,ent
to a step funation on (x ,x) with disaontinuities o n
pestpiated to xl'.· .,xn_l and, on (xi ,xi +1),
i = 0,1, ••• ,n-l, Ls* is equival,ent to a step funation
with at most. m - 1 disaontinuities.
3. A THEOREM OF KUHN-TUKER TYPE.
Let X be a Banach space and let ,j, be a '1'0
(possib1y nonconvex) objective functional defined on
X. For i = -k, .•• ,~l and i = i, ••• ,l , let <Pi
be so-ca1led constraint functionals and let M c X
138 J.W. Jerome
be defined by
(3.1) M = {x E X ~i(x) < 0 for i < 0 and ~i(x) = 0
for i > O} •
We are interested in obtaining necessary and sufficient
conditions for x E M to be a solution of the problem: o
(3.2) = inf{~ (x) o
X E M} •
Now a functiona1 ~ on X is said to be
quasi-differentiabZe at a point x E X if there exists
a convex weak* c10sed subset M(x) c X' (the dual of
X) such that the directiona1 derivative
(3.3) ~(x) ae = 1im ~(X+te) - ~(x)
t-+O+ t
exists for each direction e E X and satisfies
(3.4) l.!!(x) = ae sup A(e) AEM(x)
The notion of quasi-differentiabi1ity is thus a
genera1ization of convexity; indeed, if ~ is convex
on X then ~ is quasi-differentiab1e on X and
Linearization and Spline Projections 139
M(x) may be taken to be the subdifferential of ~ at
x , i.e., the set of all A E X' satisfying
~(y) - u(x) ~ A(y - x) for all y E X
and in this case M(x) is bounded in X' as weIl as
convex and weak* closed and M(x) satisfies (3.4).
In the sequel, we shall be interested in quasi
differentiable functionals on X such that, for
certain x E X, M(x) is bounded in the Banach space
X' . Indeed, we have
LEMMA 3.1. Suppose that X E M is fixed. o Let the
functionals ~i' I < i < l , be linear and, for
-k < i < 0 , suppose that the ~i are quasi-differen-
tiable on X with the corresponding sets M. (x ) 1. 0
bounded in X' •
convex functional
Moreover, suppose that there is a
h. such that 1.
(3.5) hi(e) > sup A(e) , e EX, AEM. (x )
1. 0
and that there is a privileged convex cone ~ such
that, for each e E ~ , and eaah sufficiently smaU
t > 0 ,
140
x o
x(t) = x + te E M • o
J.W. Jerome
is a solution to (3.2)3 there exist
numbers t. 3 not all zero3 and functionals l.
A. E Mi(x) such that l. 0
for a11 e E ~
and such that t i ~ 0 for i < 0 and ti~.(x) = 0 l. 0
for i < 0 .
PROOF. Simp1y combine Theorem 4.1, p. 83, the
Coro11ary of p. 88 and Lemma 4.2, p. 89 in reference
[11] .
We are interested in special choices of ~o as
remarked in the introduction. Thus, let Y be a
Banach space, ~ a seminorm on Y and let W be a " mapping from X into Y which is Gateaux differen-
tiab1e, i.e., for each x E X there exists a con
tinuous linear mapping DW(x;·) from X into Y
such that, for each y EX,
w(x+ty) - W(x) = tDW(X;y) + o(t) , as t ~ 0 .
Linearization and Spline Projections
The nonnegative real-valued functional ~ = ~.t/J on o
X is quasi-differentiable 111, p. 69] and, if
IDt/J(xo ;·)], denotes the transpose of Dt/J(xo ;·) and
141
M(t/J(x » o denotes the subdifferential of ~ at t/J(x )
o
then the set
M (x ) = [Dt/J(x ;·)]'M(t/J(x » 000 0
satisfies (3.4) with x = x and, moreover, M (x ) 000
is convex, weak* closed and bounded in X' [11, p. 69],
the boundedness following since M(t/J(x» is bounded o
in Y' •
THEOREM 3.2. Let the (aonvex) set M in (3.1) be
defined by aonvex funationaZs ~i for -k ~ i ~ -1
and Zinear funationaZs for 1 ~ i ~ l. Let
~ = ~.t/J be a (quasi-differentiabZe) objeative o
funationaZ on X where t/J is a ch.teaux differentiabZe
mapping of X into a Banaah spaae Y and ~ is a
semi-norm on Y. If x € M is a soZution of the o
extremaZ probZem (3.2) then there exist numbers t i ,
not aZZ zero, and funationaZs Ai € Mi(Xo) suah that,
142 J.W. Jerome
(3.6)
for all e e: ~ ,
and suoh that t i ~ 0 for i ~ 0 and ti~i(xo) = 0
for i < o. Here ~ is the oonvex oone defined by
(3.7) ~ = {t(x-xo) : t > 0, x e: M} •
Conversely, if x e: M and (J.B) is satisfied for o
i < 0, t > 0, then x is a solution of (J.2) o 0
provided the ~teaux differential D$(x ;.) has the o
property that $(M) c D$(x ;.)~ + $(x) and o -N 0
provided
-1 (3.8) -.L tiAi(e) > 0
~=-k
for all e e: ~ •
PROOF. We first prove the converse. If (3.6) and
(3.8) hold, then, 1etting e e: ~ be arbitrary, we
have,
Linearization and Sp1ine Projections 143
-1 ~ (x +te) - ~o(xo) o 2. - l tiA.; (e) < t A (e) < t Um --.;;;..o---=o::...-. __ --.,;:;_~
i=-k • - 0 0 - 0 t~ t
~($(xo) + tD$(x je»~ - ~ (x ) + o(t) < t 1im 0 0 0 - 0 t
t~
~($(xo) + tD$(x je»~ - ~($(x » = t 1im 0 0
o t~ t
< t [~($(x ) + D$(x je»~ - ~($(x »] - 0 0 0 0
the 1atter inequa1ity a consequence of the convexity
of ~. The resu1t is now a consequence of the
hypotheses t > 0 and $(M) c D$(x ;·)K_ + $(x ) • o 0 -N 0
rhe necessity is simp1y a consequence of Lemma 3.1.
Indeed, the cone ~ defined by (3.7) satisfies
x(t) = x + te € M o
for all o ~ t 2. l/t'
if e = t'(x - xo) € ~. MOreover, if i = 0 , then
the functiona1 h can be defined by o
h (e) = o
sup A(e) A€M (x )
o 0
e € X ,
144 J.W. Jerome
and the convexity of h is a consequence of the o
fo11owing inequa1ities.
~ (x +t(ae1+ße2))-~ (x ) 1im 0 0 0 0
t~ t
(~(a$(x )+atD$(x ;e1)) 1 +~(~$(X )+ßt~$(X ;e2))-~ (x )+o(t)
< 1im 0 0 0 0
- t~ t
(a[~($(X )+tD$(x ;e1)-~ (x )] 1 o 000
+ß[~($(x )+tD$(x ;e2)-~ (x )] = 1im 0 0 0 0
t~ t
For -k ~ i ~ -1 , we define
and (3.5) is a consequence of (3.4), the convexity of
~. , and the increasing property of 1
Linearization and SpIine Projections
for t > o. This concIudes the proof.
We are now prepared to state a major consequence
of Theorem 3.2. Let X = Wm,P(a,b) for m > 1 and
1 ~ p ~ 00. Let T be a mapping of wrn,P(a,b) into
LP(a,b) of the form
(3.8') Tf = (m-I) m X(·,f(·), .•. ,f (.»n f
(m-I) + w(·,f(.), ... ,f (.».
We have
COROLLARY 3.3. Let T be a mapping of Wm,P(a,b)
145
into LP(a,b) given by (3.8~ for 1 ~ p ~ 00. Suppose
that X(T,~ , ... ,~ 1) and W(T,~ , ... ,~ 1) have o m- 0 m-
continuous partiaZ derivatives with respect to
~ , ... ,~ 1 on [a,b] x Rm and that convex func-o m-
tionaZs ~-k' ... '~-I and continuous affine functionaZs
~I' ... '~l are prescribed on Wm,P(a,b). Let U be
defined by
146 J.W. Jerome
U = {f € ~,P(a,b) ~i(f) ~ 0, -k ~ i ~ -1 and
~i(f) = 0, 1 ~ i ~ L} .
Then, if f € U satisfies o
(3.9) IITf P = inf IITf I1 , o P f€U 1>
therae exist numberas t i , not aZZ zerao, and functionaZs
Ai € M.(f) such that ~ 0
fora aZZ e in the convex supporat cone
{t (f - f ) : t > 0, f € U} • o
Herae, Mi(fo) fora -k ~ i ~ -1 is the subdifferaentiaZ
of ~. at f , and ~ 0
M (f ) = [DT(f ;·)]'M(T(f » 000 0
wherae M(T(fo» is the subdifferaentiaZ of 11· I~ at
T(fo) and DT{fo ;·) is the cateaux differaentiaZ of
T at f o . AZso, t i ~ 0 fora i < ° and
Linearization and Sp1ine Projections
PROOF. The Gateaux differentiabi1ity of T fo110ws
as in Theorem 2.1 and the coro11ary is a consequence
of Theorem 2.2.
147
We remark that Theorem 2.1 contains sufficient
conditions for (3.9) to have a solution for 1 < p < 00
For p = 00 , sufficient conditions are given in [7,
Theorem 4]. Fina11y, the necessity of Theorem 3.2
and Coro11ary 3.3 is tru1y meaningfu1 on1y when no
affine equa1ity constraints are present.
4. THE UNIVERSAL EXISTENCE OF GENERALIZED INTER
POLATING SPLINE FUNCTIONS.
Fo110wing Aubin [1], we sha11 describe the abstract
Hi1bert space framework for our proeb1m and the resu1ts
in general form, before proceeding to the app1ication
of interest. To this end, let Hand V be Hi1bert
spaces and B(u,v) abilinear form on V such that
(i) I: V ~ H is continuous ,
(4.1) (ii) B(u,v) is continuous on V,
(iii) V is dense in H.
148 J.W. Jerome
Let W be a Hi1bert space and r a linear mapping
of V into W such that
(i) r is a continuous mapping of V onto W,
(4.2)
and (ii) the kerne1 V of r is dense in H. o
We have
LEMMA 4.1. Let (4.1) and (4.2) be satisfied. There
exists a Zinear operator A with domain DA C V
dense in H satisfying
(4.3) B(u,v) = (Au,v)H for aZZ u E DA' V E V o
A is the restriction of a continuous Zinear operator
A from V into the duaZ V' of V and DA is a o 0 II
HiZbert space under the graph norm
(4.4)
A is continuous as a Zinear operator from DA into
H. Moreover, there exists a continuous uniqueZy
detePmined Zinear operator ~ mapping DA into the
duaZ W' of W such that
Linearization and Spline Projections
(4.5) B(u,v) = (Au,v)H + (Qu,rv)
for all u € DA' v € V
where (.J.) represents the duality pairing on
W' x W •
PROOF. We give only an outline and refer the reader
to [1, Chapter 6] for complete details. For each ....
u € V we define Au € V' by o
(Au,V) = B(u,v) , for all v € V o
DA is defined to be the space of u € V such that ....
149
Au € H ,where H is he re identified as a dense linear
subspace of V' • o
Since the graph
{(u,Au) U € D}
is closed in V' x H it follows that DA is complete
under the norm induced by (4.4). Now the bilinear form
B(u,v) - (Au,v)H
is continuous on DA x V and there exists a continuous
linear operator B from DA J.
into V c V' • o
150
The transpose r' of
range of
r' is
r 1
V o
is c1osed.
since V o
r
J.W. Jerome
has c10sed range, since the
It fo11ows that the range of
is the kerne1 of r . If M 1
V o
is a continuous right inverse of r' mapping
onto W' ,then n = MB satisfies the Lemma.
Remark: By identifying W' with W it is possib1e
to assert the existence of a unique continuous operator
n from DA into W satisfying
(4.6) B(u,v) = (Au,v)H + (nu,rv)W
for all u E DA' v E V •
Our next resu1t describes the equiva1ence of a
variationa1 problem with that of a genera1ized boun
dary va1ue problem. Let P be an orthogonal pro
jection of W into itse1f and set Q = I - P. Let
M be any continuous right inverse of r. We have
LEMMA 4.2. Let (4.1) and (4.2) be satisfied and let
u denote the kernel of pr. Then the boundary o
value problem
Linearization and Sp1ine Projections
(4.7)
(i) Au = 0
(ii) Pfu = t 1
(iii) QQu = t 2
151
has a soZution u € ~ whe1'e t 1 € PW and t 2 € QW
if and onZy if the1'e exists
(4.8)
U € U satisfying o 0
fo1' atz V € U • o
In this aase~ u and u are 1'eZated by the equation o
(4.9)
PROOF. Suppose u € DA is a solution of (4.7). Then,
by (4.5), which we rewirte in the form,
(4.10) B(u,v) = (Au,v)H + (PQu,Pfv)W + (QQU,Qrv)W '
for v € V , it fo110ws that
solution of (4.8). Indeed,
u = u - Mt o 1 is a
Pfu = 0 so that ·0
U € U and (4.8) fo110ws from (4.7) and (4.10) (with o 0
u rep1aced by u). o Converse1y, if U € U satisfies o 0
152 J.W. Jerome
(4.8) then u given by (4.9) satisfies (4.7 ii) and
since V E U we have, o 0
B(u,v) = 0 = (O,v)H for all V E V o
Thus, u E DA and Au = O. Fina11y, to deduce
(4.7 iii) we use (4.8), (4.9) and (4.10) to conc1ude
that
1. e.,
(4.11)
Now Qf maps
for all v E U o
U onto QW so that (4.11) imp1ies o
that t 2 - Q~u = O. This comp1etes the proof of the
lemma.
This lemma, of course, does not guarantee the
existence of a solution of (4.7); it mere1y states
the simu1taneous existence of solutions of (4.7) and
(4.8). To obtain the existence of solutions we make
the fo110wing assumptions.
Linearization and Sp1ine Projections
(i) I: V ~ H is compact, and,
(ii) there exist positive constraints C
(4.12) and a such that B(u,u) + C(u,u)H ~
a(u,u)V for all u € V •
As previous1y, let u o
denote the kerne1 of pr
and let B be the unique1y determined continuous
linear operator from its dense domain DB C Uo which
is comp1ete under the usua1 graph norm, into H
satisfying
(4.13) B(u,v) = (Bu,v)H for all u € DB, v € Uo •
B is a restrietion of the operator A defined
ear1ier. By (4.12 ii) and the Lax-Mi1gram theorem
it fo11ows that B + CI maps DB onto Hand by
(4.12 i) the Riesz-Fredho1m-Schauder theory is
app1icab1e to B + CI + AI and hence to B + AI •
Specifica11y, we have that for real A, B + AI is
a continuous linear injective mapping of DB onto
H except when A be10ngs to a countab1e subset E
153
of R with no finite accumu1ation points. If A € E
the kerne1 of B + AI is a finite-dimensional subspace
154 J.W. Jerome
with dimension equal to the dimension of the kernel of
its transpose B' + Aland the range of B + Aland
B' + A I are closed. Finally, B' + AI is a con-
tinuous linear injective mapping of H' onto
and only if AlE •
D' B
if
Now if we identify H with H' and view B as
a closed linear operator with dense domain DB in H,
then the transpose mapping B' is an extension of the
* usual adjoint mapping B whose domain D * is B
defined as the set of all h € H such that there
* exists B h € H satisfying
(4.14) * (Bu,h)H = (u,B h)H for all u € DB •
* A mayaIso be characterized as the operator induced
by
(4.15) * B (u,v) = B(v,u)
in the same manner that A is defined by (4.5). * A
is not the adjoint of A • It is easily seen that the
kernel of B' + Al coincides with the kernel of B
* B +Al and is contained in DA* for all real A •
Linearization and Sp1ine Projections
We may thus app1y the Fredho1m-Riesz-Schauder theory
using the 1anguage of the adjoint operator. If w
is the operator such that
(4.16) * B (u, v) * (A u,v)H + (wu,rv)w for all
we have the fo11owing existence theorem. Reca11 that
A E E if and on1y if -A is an eigenva1ue of B •
155
LEMMA 4.3. Suppose that (4.1), (4.2) and (4.12) are
satisfied. Then if 0 I E , i.e., if the only solution
u E DA to the boundary va lue prob lem
(i) Au = 0
(4.17) (ii) Pru = 0
(iii) Qnu = 0
is the zero solution, then (4.7) has a solution
U E DA for every t 1 E PW and every t 2 E QW • If
o E E , then (4.7) has a solution if and only if
t 1 and t 2 satisfy the oompatibility oondition
* (4.18) (t2,Qrv)w = (t1 ,wPv)w for atz v E N
156
where
* of B
J.W. Jerome
* N = {u E DA * : pr u = Qw u = O} is the nutt space
PROOF. If 0' Ethen by the Riesz-Fredho1m-Schauder
theory there exists Uo E DA satisfying
(4.19) (Bu ,v) = l(v) for all v E U o 0
where lEW' is given by
(4.20)
Thus u is a solution of (4.8) and hence (4.7). The o
characterization of 0 E E given by a non-zero
solution of (4.17) is immediate from Lemma 4.2. If
o E E , then (4.18), via the integration by parts
*1 formu1a (4.16), imp1ies that l E N where l is
defined by (4.20). The Riesz-Fredho1m-Schauder theory
again yie1ds a solution u to (4.19) and hence to o
(4.7). The converse is simi1ar.
m >
We proceed now to the app1ications. For a fixed
1 , let ~,2(a,b) be the usua1 real Sobo1ev
Hi1bert space and let B(u,v) be the continuous
bi1inear form on Wm,2(a,b) given by
Linearization and Sp1ine Projections
(4.21) B(u,v) =
where bmm = 1 on (a,b) and bij € L~(a,b),
1 ~ i,j ~ m •
LEMMA 4.4. B(u,u) satisfies the aoepaiveness in
equaZity (4.12 iiJ fop positive aonstants C and
with V = Wm,2(a,b) and H = L2(a,b) .
PROOF. We sha11 first estab1ish the inequa1ity, for
u € Wm,2(a,b) ,
where C is given by
2 C = max [(m max II bij II~) ,
o<i,j<m-l L
Now by (4.21),
157
158 J.W. Jerorne
Using the elernentary inequalities, valid for E > 0 ,
(4.23)
We have,
(4.24)
122 2(u,v) 2 ~ - -ll u ll 2 - Ilvll 2
L E L L
1 2 rn-I. 2 B(u,u) ~ (1 - -)11 Dmull 2 - ~" L b .DJuli 2
E L j=o rnJ L
rn-I . 2 rn-I i. - III ): birnD~j" 2 + ( 4 bijD u,D1u) 2
~=o L i,J=o L
and choosing E = 2 and using the inequality,
rn-I i _i 2 I( 2 bi·D u,~u) 21 ~ (rn
i,j=o J L
we obtain (4.22) from (4.24).
Now the coerciveness inequality (4.12 ii) follows
frorn (4.22) and the well-known interpolation in
equality, valid for 0 < E < 1, and 0 < j < rn ,
Linearization and Spline Projections
(4.25)
This completes the proof of Lemma 4.4.
Let ro, ••• ,rJ be J + 1 linearly independent
continuous linear functionals on wm,2(a,b) • Let
B(u,v) be given by (4.21). Then, by Lemma 4.1, with
159
H the closure in 2 L (a,b) of the intersection of the
kerneis of ro " •• ,rJ , w= RJ+l and ff = (r of, ••• ,r I),
we deduce the existence of a closed linear operator
A 2 in L (a,b) and continuous linear functionals
J (4.26) B(u,v) = (Au,v) 2 + I niuriv
L i=o
for all u € DA and v € wm,2(a,b) • Now let
O 2 i l < i 2 < ••• < iN 2 J be a subset of N < J + 1
of the integers O, ••• ,J. Let P be an orthogonal
nJ+l projection in K defined by
160 J.W. Jerome
0 if i :I ~, k = 1, ••• ,N
(4.27) [P (r 0' ••• , r J) ] i = r i if i = i k
k
The fo11owing theorem is a direct consequence of Lemmas
4.2, 4.3 and 4.4 with P defined by (4.27) and rand
Q the Cartesian products of r i and Qi •
THEOREM 4.5. Let B(u,v) be a biZineaP fopm on
Wm,2(a,b) defined by (4.21) and Zet r1 , ••• ,rn be
n ZineaPZy independent aontinuous ZineaP funationaZs
on Wm,2(a,b). Let I = {i1 , .•• ,iN} be a nonempty
subset of 0, ... ,J . Let A and Q. satisfy (4.26). ~
Then, if the boundary vaZue ppobZem
(i) Au = 0
(4.28) (ii) r u = i 0 i I. I
(iU) Q u = i 0 i E I
has onZy the identiaaUy zepo soZution in DA' the
boundaPy vaZue ppobZem
Linearization and Sp1ine Projections
(i) Au = 0
(4.29) i € I
i /. I
has a sol,ution u € DA for arbitrary real, numbers
u satisfies the identity
(4.30) B(u,v) = L tiQiv for al,l, v such that il1
riv = 0, i € I.
If (4.28) has nontrivi~l, sol,utions, then (4.29) has a
sol,ution if and onl,y if to, ••• ,tJ satisfy the
compatibil,ity condition
(4.31)
for aU v in the nontrivial, l,inear space
Here wo, ••. ,wJ satisfy the genera1ized adjoint
integration by parts re1ationship
161
162 J.W. Jerome
* * J B (u,v) = (A u,v) + .I wiuriv 1.=0
* for all u E DA* and v E Wm,2(a,b), A derives from
* (4.16) and B (u,v) is defined by (4.15).
Our final theorem presents an app1ication of
Theorem 4.5 to the case where the operators r. are 1.
extended Hermite-Birkhoff functiona1s. Specifica11y,
we use the notation of §2 and we consider as given a
mesh a = x < ... < x = b and k. functiona1s Lij 0 n 1.
associated with each point x. of the form (2.16), 1.
where the corresponding matrix Ai of coefficients is
of fu11 rank. As in section two we augment Ai to
A A(\) obtain an invertib1e m x m matrix Ai = (aij ) • This
augmentation then yei1ds additional functiona1s L .. 1.J
at each xi ; more precise1y, we define ro, ••. ,rm(n+~l
as fo110ws:
(4.32) j = 0, ... ,m-1,
i = 0, ... ,n
The sub set I of the previous theorem is then
Linearization of Sp1ine Projections 163
defined by
(4.33) I = {O, •.• ,k -1,m, ••• ,m+k1-1, ... ,mn, ..• ,mn+k -1}. o n
Let H. = (hi(~» be the inverse of the transpose of ~ J
" Ai. Define, on suitab1y smooth functions, operators
o by v
(4.34) 0 = v
m-v-1 . . m I (_1)J+1nJ( I b DJ +v+i -m+1) i,j+v+1 •
j=o i=m-j-v-1
Here we have assumed that m oo bij E W ' (a,b) for all
i,j. Fina11y, as in (2.20) we define operators Rij •
Proper1y ordered, as ro, ••• ,rJ in (4.32)
(J = m(n+1)-1) , the operators [Rijf]i satisfy (2.21)
and, moreover, if we re1abe1 these operators no, ••• ,nJ
we have,
for all f,g for which the express ions are meaningfu1.
Here we have defined
164 J.W. Jerome
(4.36) A = m 'i' (_l)JDJ (b Di ) L iJ' i,j=o
The fo11owing theorem shows that the operator A of
(4.36) agrees with that of the previous theorem. By
the set W we mean
W o
o
THEOREM 4.6. Suppose that the coefficients bij of
B(u,v) are in ~,oo(a,b) and that k. extended ~
Hermite-Birkhoff functionals
i = O, ... ,n, are specified.
defined by (4.36) with domain
Lij at the points xi'
Then the operator A
n-l (4.37) D = Wm,2(a,b) n W2m,2(.u (xi,xi +1» n Wo
~=o
is a closed linear operator in
its domain, is characterized by
2 L (a,b) which, with
(4.38) (Af,g) 2 = B(f,g) for all f E DA' L
g E U o
Linearization of Spline Projections
where U is the intersection of the kernels of the o
operators Lij , The boundary value problem
165
(4,39)
(i) As = ° (ii) Lij = r ij j = O"",ki -1, i = O"",n
(iii) [R'jS]' = r .. 1. 1. 1.J
i=O",.,n,
has a solution for any specified numbers r" provided 1.J
the auxiliary boundary value problem (4.39) with each
r ij = ° has only the zero solution. In the event
that the auxiliary homogeneous problem has nontrivial
solutionB 3 let NA* denote the linear subspace of
DA* of solutions of the (purely) homogeneous adJoint
boundary value problem. Then (4.39) has a solution if
and only if the compatibility condition (4.31) holds
for all v E NA* .
PROOF. If (4.38) is satisfied for A and DA the
theory of distributions and the integration by parts
formu1a (4.35) imp1y that DA has the asserted form
(4.37). Converse1y, if DA satisfies (4.37), then
166 J.W. Jerome
(4.35) implies that (4.38) holds. The result is now
a consequence of Theorem 4.5 and earlier results.
b Remark: If B(u,u) = f (Lu)2 , where the coefficients
a m= of L are in W' (a,b) , it follows that (4.39) always
has a solution if the conditions in (iii) are homo
geneous (cf. [9]).
REFERENCES
1. Aubin, J.P.: Approximation of Elliptic Boundary Value Problems, Wiley Interscience, New York, 1972.
2. Fisher, S.D. and Jerome, J.W.: Existence, characterization and essential uniqueness of solutions of = L minimization problems, Acta Mathematica (Sweden),
submitted.
3. Goffman, C.: Calculus of Several Variables, Harper and Row, New York, 1965.
4. Golomb, M.: Hm,P extensions by Hm,p splines, J. Approx. Th. 7 (1972), 238-275.
5. Halperin, I.: Theory of Distributions, University of Toronto Press, Toronto, 1952.
Linearization of Sp1ine Projeetions
6. Jerome, J.W.: Minimization problems and linear and nonlinear sp1ine funetions. I : Existenee, SIAM J. Numer. Anal., to appear.
7. Jerome, J.W.: Minimization problems and linear and nonlinear sp1ine funetions. 11 Convergenee, SIAM J. Numer. Anal. (submitted).
8. Jerome, J.W. and Pieree, J.: On sp1ine funetions determined by singular se1f-adjoint differential operators, J. Approx. Th. 5 (1972), 15-40.
167
9. Jerome, J.W. and Sehumaker, L.L.: On Lg-sp1ines, J. Approx. Th. 2 (1969), 29-49.
ID. Lueas, T.R.: M-sp1ines, J. Approx. Th. 5 (1972), 1-14.
11. Psheniehnyi, B.N.: Neeessary Conditions for an Extremum, Maree1 Dekker, Ine., New York, 1971. (Translation).
ON THE CONVERGENCE OF CUBIC INTERPOLATING SPLINES
Tom Lyche & Larry L. Schumaker
1. INTRODUCTION.
Given n > 2 , a partition /J. = {o = x < xl < n 0
< x = 1} of [0,1] , and a function n 0
C[O,l] g(O) = g(l)} f E C[O,l] = {g E , let P f n
be the periodic cubic sp1ine interpo1ating f at
(For a precise definition of P f , see §2). n
The fo11owing question has received considerab1e
attention recent1y (see e.g., [1 - 4, 6 - 9, 11 - 12]
and references therein): Given a sequence </J.) of n
partitions of [0,1] with
Research supported in part by AFOSR-69-1812-D.
169
170 T. Lyche & L.L. Schumaker
(1.1) 116 11 = n as n + 00 ,
what further conditions on (6) are needed to guarann
tee that the sequence of spIine interpoIants
converges uniformIy to f as n + oo? The study of
this question was stimuIated by the discovery by
Nord [9] of an exampIe of a sequence (6) satisfying n
o (1.1) and a function f E C[O,I] such that
IIf-Pflifo as n+ oo • n
Sufficient conditions on (6) n
to assure
11 f - P f JI + 0 as n + 00 have been given in terms of n
various measures of the spacing of the partition. One
such measure is
(1.2)
where h = i
In terms of
THEOREM A.
[0,1] with
m = n
xi - xi_I'
h max i/h O<i,~ <n j li-jl=I
i = 1, ••• ,n
m Marsden [7] proved n
If <6 ) is a sequenae n
m < m < 2 • 439+ , then n-
and h = h 0 n
o f parti tions of
Iim sup IIp 11 < 00
n n+oo ;
Convergence of Cubic Splines 171
o i.e. foT' evepy f € C[O,l].J IIf - Pnfll-+ 0 aB n -+ 00 •
Earlier vers ions of Theorem A were obtained by
MeirjSharma [8] (with m< 1:2), by CheneyjSchurer [3]
(with m < 2), and by Hall [6] (with m < 1 + 12). (See [7] for details of the history.)
The purpose of this paper is to establish analogs
of Theorem A for cubic natural spline interpolation
and for cubic type-I spline interpolation (see §2 for
the definitions). We relate our positive results to
certain negative results in §5. There are several
remarks in §6.
2. CUBIC SPLINES.
Given n > 2 and a partition tJ. = {O = x < x n 0 i
< < x = I} of [0,1], we denote the class of n
cubic splines with knots
(2.1)
tJ. by n
is a cubic polynomial
172 T. Lyche & L.L. Schumaker
We single out the fo11owing three subc1asses of S3:
(2.2) S3(6n ) = {s E S3(6n) : s(v)(O) = s(v)(l),
v = 0,1,2} ,
s"(O) = s"(l) = O} ,
These classes are the fami1iar periodic, natural, and
type I cubic splines.
(2.5) P f(x.) = f(x.) , i = O,l, ••• ,n n 1 1
Similar1y, given f E C[O,l] , we define N f E NS 3 (6 ) n n
(2.6) N f(xi ) = ° f(x.) = f(xi ) , i = O,l, ••• ,n n n 1
o 0
Pn defines a projection of C[O,l] onto S3(6n) ,
while N and 0 are projections of C[O,l] onto n n
Convergence of Cubic Sp1ines
In view of the e1ementary inequa1ities
and the we11-known fact that d(f,NS3 (ön» =
inf Ilf - sll and d(f,OS3(ön » converge to zero sENS 3 (ön )
for any f E C[O,l] provided ö satisfies (1.1), n
(e.g., this fo11ows from the resu1ts of [10] for
smooth functions or the results of [12]) it suffices
for us to s tudy "N 11 and "0 11. We dis cuss the n n
type-I sp1ines in §3 and the natural sp1ines in §4.
3. TYPE-I SPLINES.
The main resu1t of this section is
173
THEOREM 3.1. If <ö ) is a sequenae of partitions of n
[0,1] satisfying
174 T. Lyehe & L.L. Sehumaker
(3.1) m < m < 2.439+ for all n, n-
then 1im sup 11 0 11 < 00. Thus for any sequenae of n
n~
partitions \Ö > of [0,1] satisfying (3.1) and for n
any fEe [0 ,1], 11 f - 0 fll + ° aB n + 00 • n
Our proof of this resu1t is simi1ar to that used
by Marsden [7] for Theorem A. First we need a loea1
3 Let M(x,y) = (y - x)+ and let
M[X;~l' ••• '~r] denote the r - Ith divided differenee
of M as a funetion of y taken over ~1 < ~2 <
< ~r· Setting x_i =-xi and xn+i = 2 - xn_i '
i = 1,2,3 we eonsider the usua1 norma1ized B-sp1ines
(3.2)
We reea11 ° ~ NMi(x) ~ 1
X E [0,1] . We define
i = -1, ••. ,n+1 .
n+1 and L NMi(X) - 1
i=-l for
Convergence of Cubic Sp1ines
NMi(X) , i = 0,2,3, .•• ,n-2,n
(3.3) ~i(x) = NM1 (x) + NM_1(x), i = 1 ,
NMn_1 (X) + NMn+1 (x) , i = n-l •
By the construction of the ~. 's it is not difficu1t 1.
175
to see that they form a basis for OS3(ßn ) (e.g., the
symmetry of the knots guarantees ~'(O) = ~'(O) = o 1
~' (1) = ~'(1) = 0). Moreover, ° < ~ (x) < 2, n-1 n - i -
n i O,l, ••• ,n and L ~i(x) = 1 + NM_1 (X) + NMn+1 (x) ~ 3.
i=o
The estimate for 110 11 is based on the fol10wing n
lemma.
LEMMA 3.2. Then
n = 3 max I IA .. I
O,$j~n i=o 1.J
PROOF. For i = O,l, •.• ,n let si(x) E OS3(ßn ) be
the sp1ines satisfying si(xj ) = 0ij' i,j = O,l, ••• ,n •
176 T. Lyche & L.L. Schumaker
n n Then 0 f(x) = L f(xi)si(x) and 11 On 11 ~ L Isi(x)1 •
n i=o i=o
n Suppose ß = (ßij)i,j=o is such that si(x) =
n .L ßij~.(x). Then J=o J
But I = Aß and the resu1t f011ows.
PROOF OF THEOREM 3.1: The matrix A is exp1icitly
b c 0 0 0 0
(al + c_l ) bl cl
0 a2 b2 c2
A= an_2 b n-2 c n-2
0 0
0 an_1 b n-l (cn_l + an+l )
0 0 a b n n
Convergence of Cubic Splines
where for i = -l, ••• ,n+l
2 h i +2
ci = NMi (xi +1) = ~(h:-i-+-l~+-:h~i-+-2-=-) -::(h-i"':;;"'-'+-:h:-i-+-l-+~h-i-+-2-=-) ,
To estimate IIA-lill we follow Marsden [7] and use
(3.4)
where D is any matrix. We take D to be the
diagonal matrix with entries It is shown
2 in [7] that b-l < (m + m + l)(m + 2) if ß
i = (2m + 1) (m + 1) n
satisfies (3.1). Now
177
178 T. Lyche & L.L. Schumaker
DA - I ""
0 c /b 0 0 o 0
(a1+c_1)/b1 O cl/bI 0
a2/b2 0 c2/b2
a /b n-2 n-2 0 c /b n-2 n-2 0
o 0 a /b 0 n n
It is shown in [7] that ci/bi + ai +2/b i +2 < 1 for
i = O, ••• ,n-2 under the assumption (3.1). It remains
to show that (al + c_1)/b1 < 1 and (cn_1 + an+1)/
b n-1 < 1 . It suffices to consider the first of these.
By the symmetry in the choice of knots x_3 , x_2 ,
x_I' a1 = c_1 , so we need on1y show 2a1 < b1 . It
is an exercise in ca1cu1us to see that the maximum of
occurs for and (In
doing this, reca11 ho = h1 .) For this choice,
Convergence of Cubic Splines 179
2 2a /b < (m + ~)(m + m + 1) < 1 for m satisfying
1 1 - 4m l + Sm + 3
(3.1) .
The spline 0 f given above is not quite the n
usual type I-interpolating spline of the literature
(see e.g., [10]). Given f with f'(O) and f'(I)
defined, the type I interpolating cubic spline 0 f n
is defined by
(0 f)'(x.) = f'(x.), i = O,n • n 1. 1.
THEOREM 3.3. Let
f'(I) are defined.
f E C[O,I]
If (& > n
and suppose f' (0)
is a sequence of
parti tions wi th m < m < 2.439+ , then n-
lim 11 0 f - f 11 = ° . n n400
and
PROOF. Given f, choose a cubic polynomial p such
that p' (0) = f' (0) , p' (1) = f' (1) . For functions
g with g'(O) = g'(I) = 0, ° g = 0 g. Thus n n
180 T. Lyche & L.L. Schumaker
Ilonf - fll ~ IIOn(f - p) - (f - p) 11 + lIonP - pli
= 11 0 (f - p) - f - p) 11 + 11 ° p - p 11 • n n
The first term on the right converges to 0 as n + 00
by Theorem 3.1 whi1e the second converges to 0 by
weIl known resut1s for smooth functions (see e.g. [10]).
4. CUBIC NATURAL SPLINES.
In this section we prove
THEOREM 4.1. If (ßn> is a sequenae of partitions of
[0,1] with
(4.1) m < (1 + 1:13)/2 = 2.30+ and m < m for aU n-
then lim sup 11 N 11 < 00. Thus for any sequenae of n n~
partitions satisfying (4.1) and any f € C[O,l],
11 f - N fll + 0 aB n + 00 • n
n ,
The basic out1ine of the proof is the same as in
§3. First we need a basis for NS3 (ßn). With M(x,y)
as in §3 we let
Convergence of Cubic Sp1ines 181
(xi +2 - xi_2)Mlx;xi_2,···,xi+2]'
(4.2) ~i(x) = i = 2, ••. ,n-2
The
Mlx;x 3'x 2'x l'x], i = n-1 n- n- n- n
M[x;x 2'x l'x ]/(h 1 + 2h ), i = n . n- n- n n- n
{~}n is (except for a norma1ization) the basis i 0
constructed by Grevi11e [5] for NS 3(6n ). It is easi1y
n-2 verified that 0 < ~.(x) < 1,
- 1 -I ~i(x) ~ 1 , and
i=2
~ (x ) = ~l(x ) = ~ 1(x) = ~ (x ) = 1 • o 0 0 n- n n n
Before proving Theorem 2.1 we need a lemma whose
proof is on1y a slight variant of that used for Lemma
3.2.
LEMMA 4.2. Then
182 T. Lyche & L.L. Schumaker
PROOF OF THEOREM 4.1. The matrix A in Lemma 4. 2 has
the following form
o o A =
c n-2 o
where the Ci are as in §3 and
2 -1 2h )-1 . ~ (x 1) = h l(h 1 + h) (h 1 + ~n n- n- n- n n- n
We estimate /lA-1 111 by the formula (3.4), where
Convergence of Cubic Sp1ines
now we choose
a i = 1
ß , i = 2
Dii = 1 i = 3, ••• ,n-1
y , i = n
15 i = n+1 •
with a, 15 > 1 and 0 < ß, y < 1 to be chosen
1ater. Then
0 a2 b2_1 c2
0 0 b -1 3
DA - I =
a b -1 n-2 n-2 c n-2
0 ya 1 y-yc -1 n- n-2
0
y
0 0 öep (x 1) 15-1 n n-
183
Marsden showed (by e1ementary Ca1cu1us in a pre1iminary
version of [7]) that ci + ai +2 < bi +1 , i = 1,2, ••• ,11-3 ,
184 T. Lyche & L.L. Schumaker
under the assumption (4.1). It remains to show the
ll-norms of the first and last two columns do not
exceed 1. We concentrate on the first two, and
choose a,ß such that the norms are 1 - E, E > 0
sma11. Thus we require
(4.3)
The solution of this system is
a = 2-3a2+Ea2-2E
(1-a2+~o(x1» , ß =
a2+E+2~o(xl)-E~o(x1)
(1-a2+~o(xl»
For 1 < a and 0 < ß < 1 we need
(4.4)
The minimum of 1 - a2 + ~o(xl) and the maximum of
2a2 + ~o(xl) both occur for partitions with
hi +1 = hi/m, i = 1,2. In this case
Convergence of Cubic Sp1ines 185
3 m + 1
(m2+m+1) (m+1) (m+1) (2m+1)
It is easi1y checked that if (4.1) ho1ds then (4.4)
does too (if we take E sufficient1y sma11).
5. NEGATIVE RESULTS.
There is a definite limit to the extent to which
the ear1ier resu1ts can be improved.
THEOREM 5.1. For eaah fixed m> (3 + 15)/2 there.
exists a sequenae of partitions (ÄJ of [0,1]
satisfying (1.1) and mn ~ m, aZZ n, suah that
1im sup 11 p 11 = W'<'" n
= 00 •
By the uniform boundedness princip1e, Theorem 5.1
asserts that if m> (3 + 15) /2 there exists a sequence
of partitions with mn ~ m and IIÄnll + ° and a func-o
tion f € C[O,l] such the P f does not converge n
186 T. Lyche & L.L. Schumaker
uniforrnly to f. Similarly, there exist continuous
functions and partitions with m < m n- and
such that N fand 0 f do not converge to the n n
functions.
Aversion of Theorem 5.1 without the assertion
that "L\ 11 -+ 0 as n -+ 00 was proved by Marsden [7]. n
Without this property the theorem would not be a true
negative convergence result. The proof of Theorem 5.1
re lies on sign regularity properties of the matrix A,
and is deferred to a later paper.
6. REMARKS.
1. The periodic cubic spline has received consider
ably more attention than the natural and type-I
splines. There are, of course, general results which
apply for smoother functions fand there are results
for continuous f when the partitions are quasi
uniform (see e.g. [10, 12]). Hall [6] obtained a
version of Theorem 3.1 under the hypothesis m <m<oo n-
coupled with an additional hypothesis on another
measure of partition behavior. Cheney (unpublished
lecture notes) proved aversion of Theorem 4.1 assuming
Convergence of Cubic Splines
m < m < 2 (by entirely different methods). n-
187
2. The methods employed here can be applied in other
cases as suggested by Marsden. For example, they can
be used to show that for equally spaced knots periodic,
natural, or type-I splines of any order converge as
n + 00 for arbitrary continuous functions. (This is
of course also a direct consequence of the results of
Swartz/Varga [12].) However, for higher-order splines
it is increasingly difficult to estimate IIA-l"l. A
useful tool for this purpose is a lemma of de Boor [4]
which requires some sign-regularity properties of A.
Such sign regularity properties of spline bases can be
derived conveniently from the total-positivity of the
B-splines; one need only study the properties of
certain transformation matrices relating the desired
basis functions to the B-splines. We hope to make
further application of these ideas to higher-order
spline convergence.
3. The gap between the constants in the positive
results (2.439+ in Theorems A and 3.1 and 2.30+
in Theorem 4.1) and the constant (3 + 15)/2 = 2.62-
in Theorem 5.1 is annoying. There is hope that the
methods used here could be employed to close the gap.
188 T. Lyche & L.L. Schumaker
REFERENCES
1. Birkhoff, G.D. and de Boor, C.: Error bounds for sp1ine interpolation, J. Math. Mech. 13 (1964), 827-835.
2. Cheney, E.W. and Schurer, F.: A note on the operators arising in sp1ine approximation, J. Approx. Th. 1 (1968), 94-102.
3. Cheney, E.W. and Schurer, F.: Convergence of Cubic sp1ine interpo1ants, J. Approx. Th. 3 (1970), 114-116.
4. de Boor, C.: On the convergence of odd-degree sp1ine interpolation, J. Approx. Th. 1 (1968), 452-563.
5. Grevi11e, T.N.E.: Introduction to sp1ine functions, in Theory and App1ication of Sp1ine Functions, Academic Press, New York (1969), 1-35.
6. Hall, C.A.: Uniform convergence of cubic sp1ine interpo1ants, J. Approx. Th., to appear.
7. Marsden, M.: Cubic sp1ine interpolation of continuous functions, J. Approx. Th., to appear.
8. Meir, A. and Sharma, A.: On uniform approximation by cubic sp1ines, J. Approx. Th. 2 (1969), 270-274.
9. Nord, S.: Approximation properties of the sp1ine fit, BIT 7 (1967), 132-144.
10. Schultz, M. and Varga, R.S.: L-sp1ines, Numer. Math. 10 (1967), 345-369.
Convergence of Cubic Sp1ines 189
11. Sharma, A. and Meir, A.: Degree of approximation of sp1ine interpolation, J. Math. Mech. 15 (1966), 759-768.
12. Swartz, B.K. and Varga, R.S.: Error bounds for sp1ine and L-sp1ine interpolation, to appear.
191
AVERAGING INTERPOLATION
T.S. Motzkin & A. Sharma & E.G. Straus
1. INTRODUCTION.
Given a set X of N(~ n) real numbers there
exists a unique polynomial P of degree n - 1 n-l,f,X
or less that approximates best on X to a given real
function, that is, that minimizes the deviation o(f,P)
defined by the th power metric (1 < p < 00) with p
positive weights or (p = 00) by max If - pi on X
and there are at least n + 1 points on X , where
the difference f - P takes alternatingly positive
and negative values. When N = n the polynomial
P is the polynomial of interpolation, and for n-l,f,X
N = n + 1 and p = 00 we obtain the next-to-inter-
Research supported in part by NSF Grant GP 28696.
192 T.S. MOtzkin & A. Sharma & E.G. Straus
po1atory polynomial.
In arecent paper [3] we have estab1ished a
relation between interpo1atory and next-to-interpo1atory
polynomials on a finite set X and extended it to sets
X of multiple points. We have shown that for a given
p, 1 < p ~ m and N = n + 1 every positive1y
weighted mean of interpo1ators is a next-to-interpo1ator,
Le., it minimizes eS for a suitab1e choice of weights,
and in order to bring out this resu1t in its genera1ity,
we have introduced the concept of uniso1vence relative
to a given set of functiona1s on an-parameter fami1y
on n + 1 points. The object of the present paper is
twofo1d. First in §2 and §3, we extend the concept of
interpolation by considering "average interpo1ators"
that share with the given function certain va1ue
averages, instead of the va1ues themse1ves. Second1y
we study ac1ose1y connected minimization problem. This
enab1es us to genera1ize resu1ts of [4].
In §2 we formu1ate the general interpolation
problem. Our principa1 resu1t is proved in §3 and a
special case is discussed in §4. The trigonometric
ana10gue of Theorem 2 is formu1ated in §5. The special
case when the points of X are such that
ßXi + Y for all i is treated in some detail in §6.
Averaging Interpolation 193
In this ease we show how to obtain the polynomial of
interpolation explieitly. In §7 we generalize the
eoneept of relative unisolvenee introdueed earlier in
13]. §8 and §9 deal with possible applieations of
our prineipal results to extend the results of Motzkin
and Walsh [4].
2. A GENERAL INTERPOLATION PROBLEM.
Let X = (xl< ..• < xn) be n given real
numbers and let m be a given integer, 0 < m < n-l •
Let a. > 0, r = 1, .•. ,m. Set r
(2.1) A(z)
and
A (z) r
m = II(l+a. z)
1 r
= A(z)/(l+a. z) = r
m-l
a o
l a zll o ll,r
If l = (ll, ••• ,l ) are integers with n-m
1 .
o < l. < n-m-l , eonsider the funetionals Ll, •.• ,L - J - n-m
given by
194 T.S. Motzkin & A. Sharma & E.G. Straus
(2.2) m (I. )
L. (f) = I a f J (x +.) J p=o P P J
j = 1, ..• , n-m •
The problem is to find a po1ynomia1 p(x) of degree
< n-m-1 which satisfies
(2.3) L.(P) = L.(f) J J
j = 1, ... ,n-m •
We note the fo11owing special cases:
1. m = 0, I = (0, ••• ,0) . The po1ynomia1 P is
the Lagrange interpolation po1ynomia1 of degree
< n-1 •
2. m = 0, I = (0,1, .•• ,n-1). The po1ynomia1 P
is the Abe1-Gontcharoff po1ynomia1 of interpolation.
3. m = 1, I = (0, ... ,0) The po1ynomia1 P is
the next-to-interpo1atory po1ynomia1 of degree < n-2 •
If A(x) = 1 + u1x, u1 > ° then the po1ynomia1 p(x)
minimizes
j-1 1 I max u1 O. '-1 J J- , •.. ,n
where O. = P(x.) - f(x.) . J J J
4. m = n-1, I 1 = 0. Then the po1ynomial P is a
constant C given by
Averaging Interpolation
C = l a f +1 La, n-l /n-l
o ~ ~ 0 ~
where, as in the sequel, f j = f(xj ) .
The functionals (2.2) can be considered in terms
of operators on the n-vectors of functional values
f = (f1 , ... ,fn) , composed of the following three
basic operators:
195
(i) The differentiation operator nf = (fi, ••• ,f~) •
(ii) The truncating identity operator
Jf = (f1 , ... ,fn_1) .
(iii) The shift operator
Then if
operator
given by
(2.4)
A(z) is the polynomial (2.1), we have the
L = (L1 , ... ,L ) with n-m
Lf = (L1 (f), ••• ,L (f» n-m
196 T.S. MOtzkin & A. Sharma & E.G. Straus
which maps n-vectors into (n-m)-vectors. With on1y
a slight abuse of 1anguage, we can write
A(E) = (J+a1E) ••• (J+a E) = (J+a E)A (E) m r r
The conditions (2.3) can now be written as
(2.5)
where P = (P1 , .•• ,Pn) and L is given by (2.4).
3. TRE AX-POLYNOMIAL.
We sha11 restriet ourse1ves to the case when all
l. are zero. p(x) is then ca11ed an AX-po1ynomia1. J
First we prove
THEOREM 1. If A(z) of (2.1) has onZy negative ze~o,
then the~e exists a unique AX-poZynomiaZ p(x) of
deg~ee ~ n-m-1 , that is, a poZynomiaZ such that
m (3.1) La P(x +.)
0]..1 ]..I J
m
= La f +j o ]..I ]..I j = 1, ... , n-m .
Averaging Interpolation 197
For m > 1 and arbitrary r = 1, ... ,m, P(x) is the
poZynomiaZ which minimizes
(3.2)
. 1 m-1 max
j=l, .•. ,n-m-1 aJ - I \ a 0 I
r ~ ll,r ll+j
o = Q(x ) - f v v v v=l, •.. ,n
over aZZ poZynomiaZs Q(x) 01 degree < n-m-1 •
The po1ynomia1 P(x) satisfying (3.1) is thus a
solution of m (in general different) minimization
problems .
PROOF: In order to prove that (3.1) has a unique
solution, it is enough to consider f = 0, v
v = 1, ... ,n and show that this entai1s p(x) _ 0 .
The equations (3.1) can then be written as
(3.3) m
A(E)P:: Ir (J+a.E)P = 0, P = (P1 , ... ,Pn) . j=l J
Since a. > 0, j = 1, .•• ,m , it is easy to see that J
the operator n-vectors into
(n-1)-vectors is variation diminishing. After m - 1
steps, we arrive at the (n-m-1)-vector
198 T.S. Motzkin & A. Sharma & E.G. Straus
given by A (E)P . r
Since
(3.4) (J+a E)R = (J+a E)A (E)P = A(E)P = 0 r r r
it follows that if Rn-m+l ~ 0 , then veR) , the
-+ number of strong sign changes in R is n - m and so
v(p) ~ n-m , which is impossible for a polynomial of
are zero. degree ~ n-m-l. This proves that all Rj
Repeating the argument m times, we see that all Pj
are zero, i.e., P(x) = 0
give
If the f are not all zero, the conditions (3.1) v
which is equivalent to
o +a 0 = 0 +a 0 = ••• =0 +a 0 = 0 r,l r r,2 r,2 r r,3 r,n-m r r,n-m+l
where & = (0 1,·.·,0 m+l) = A (E)!. That is r r, r,n- r
n-m o I = -a 0 2 = •.. = +(-a) 0 m+l· r, r r, r r,n-
This proves that P(x) minimizes (3.2).
Averaging Interpolation
Remark: In order to find an exp1icit expression for
P(x) , we set
m
gj = La f +j 0 1111
j = 1, ••. ,n-m
and first solve the linear system of equations:
(3.5) m
La P +" 0 1111 J
g j , j = 1,..., n-m •
We introduce m parameters gn-m+1, ••• ,gn by the
equations:
(3.6) n-j L a11P 11+j = gj , j = n-m+1, •.• ,n • o
The two systems (3.5) and (3.6) in the n unknowns
P1 ",.,Pn can be solved exp1icit1y if we set
(3.7) b = 1 • o
It fo11ows by e1ementary computation that
(3.8) P = \I
\I 1, ... n •
199
200 T.S. Motzkin & A. Sharma & E.G. Straus
Since P(x) is a po1ynomia1 of degree ~ n-m-1 , the
divided differences of order n - m of the numbers
If we set w (x) = (x-x ) v v
(x-x +n ) , we can write these conditions as m v -m
equations:
(3.9) n-m PV+k l = 0, v = 1, ... ,m • k=o w~(xv+k)
Using (3.8), we get from (3.9) after some simp1ifi
cation
(3.10)
with
c P,v =
n-v ~ c g = 0 v = 1, .•. ,m L. p v ll+V '
p=o '
ntm bp_k L. w' (x )'
k=o v v+k
~ b}l-k L. w'(x )' o v v+k
p = n-m+1, ••• ,n-v
1.1 = 0,1, ••• ,n-m
Solving (3.10) by Cramer's ru1e for g m+1,···,g n- n
and using (3.8) we determine all the P v and then
P(x) is determined by Lagrange interpolation.
Averaging Interpolation 201
Theorem 1 can be further genera1ized.
* * THEOREM 2. If A1(z), ... ,A (z) aPe n - m poZy-n-m
nomiaZs of degree ~ m eaoh having onZy negative zeros~
then there exists a unique AX-poZynomiaZ P(x) of
degree < n-m-1 ~ that is a poZynomiaZ suoh that
(3.11)
where
m * la jP(x +.) oll llJ
m * = la .f +j
o llJ II
* m * A. (z) = la .zp J 0 pJ
j = 1, ... ,n-m
For m ~ 1 ~ and aPbitrary r = 1, ... ,m~ P(x) is the
poZynomiaZ whioh minimizes
(3.12)
(3.13)
m-1 (j) max w./la o+j/' _ J pr P j-1, ... ,n-m+1 0
m-1 w. = II Cl
J 1 rll w = 1
o
(v = 1, ... ,n)
over aZZ poZynomiaZs Q(x) of degree < n-m+1 ~ where
202 T.S. MOtzkin & A. Sharma & E.G. Straus
m = II(l+o. ,z) =
1 rJ
The proof of Theorem 2 fo11ows the same 1ines as
that of Theorem 1 and is omitted.
Note that the weights wj in (3.12) are com
p1ete1y arbitrary positive weights, thereby removing
the rather artificia1 restrietion to weights of the , 1
form o.J - in (3.2).
4. THE CASE m = 2 .
The case m = 2 is of sufficient interest and
i11ustrates Theorem 1 fair1y we11. Suppose A(z) = (1+o.1z) (1+a.2z), 0.1 ,0.2 > 0 and we seek to find the
minimum of
(4.1) j-1 1 I max 0.1 0, + o. 2oJ'+1 '-1 1 J J- , ..• ,n-
where 0v = Q(xv) - f v ' v = 1, ••• ,n over all
polynomials Q(x) of degree < n-3. If this minimum
A is taken by P(x) , then
Averaging Interpolation 203
(4.2)
. Pn- l + a 2Pn = f n_l + a 2f n + (_1)n-2A/a~-2
This system of equations can be easily solved. Indeed,
we have for v = l, ••• ,n-l
(4.3)
P = (-a )n-v(p -f ) + f v 2 n n v
1- r.a2)n-V
(-1) v-I laI + v-I
a l
v-I = (-a2)n-v(p -f ) + f + (-1) (n-v)A
n n v v-I a 2
Then by Lagrange interpolation
p(x) = ~ w(x) P LI (x-x )w'(x) v
v v w(x)
n = L(x-x.)
1 J
204 T.S. Motzkin & A. Sharma & E.G. Straus
Since p(x) is a polynomial of degree ~ n-3 , the
coefficients of n-l n-2 must vanish. This x x
requirement yields
n p n x P (4.4) I w' (~ ) = 0 , I v v o . w' (x ) =
1 v 1 v
From (4.3) and (4.4) we then have for a1 ~ a2 :
whence we have
n x f = \ v v
- l. w' (x ) 1 v
where D1 , D2 are determinants given by
Averaging Interpolation
n f n (-a )n-v
I w' (~ ) I 2 w' (x ) 1 v 1 v
Dl = x f n-v n n (-a2) Xv I v v I 1 w f(X) w' (x ) 1 v
n (-a )n-v n {-a )n-v
I 1 I 2 w' (x ) w'(x) 1 v 1
D2 = n-v n-v n (-al) Xv n (-(2) Xv
I w'(x) I w' (x ) 1 1 v
5. TRIGONOMETRIe POLYNOMIALS •
Theorem 2 has a trigonometrie analogue. We
suppose for this purpose that X has 2n + m + 1
points, {O ~ xl < ••• < x2n+m+l < 2~}. Suppose
205
206 T.S. Motzkin & A. Sharma & E.G. Straus
(5.1) * A. (z) J
m * = La .z].l = o 11J
m TI (1 +a. . z)
r=l rJ j = 1, •.. , 2n+ 1
* (5.2) A • (z) = rJ
=
Then the argument used in the proof of Theorem 2 can
give simi1ar1y
* THEOREM 3. If A.(z), (j = 1, .•. ,2n+1) are the J
polynomials given by (5.1), then there is a unique
AX-trigonomitrie polynomial T(x) of order n, such
that
(5.3) m * La .T(x +.) = o].lJ ].l J
m * La .f +. o ].lJ ].l J
j 1, .•. , 2n+ 1 •
For m > 1 , and arbitrary r (1 2 r 2 m), T(x) is
the unique trigonometrie polynomial of order n whieh
minimizes
(5.4) m-1
max w. I L a (j ) 0 . I j=1, ..• ,2n+2 J 0 ].lr l1+J
over all trigonometrie polynomials of order n, where
w. are given by (3.13), J
Averaging Interpolation
6. A SPECIAL POINT SET X.
In this section we restrict attention to sets of
points X (not necessarily real) which satisfy a
linear recurrence relation X. + y ~
(i = l, ... ,n-l) . In order that the points X =
207
{xl, .•• ,Xn } be distinct, we must impose the conditions:
(ß-l)xl + Y f. ° (6.0) and
if ß f. l, then ßj = 1 for j = 2, ... ,n-l
In return for this relatively special choice of the
sequence X, we can now get results analogous to those
of Theorem 1 with only minor restrictions on the zeros
of the polynomial A(z) of Theorem 1 instead of the
requirement that the zero be negative. At the same
time, we can combine information on the vectors
+ +(k) A(E) f, ..• ,A(E) f· to get a more general interpolation
result.
THEOREM 4. Let X= {xl'···xn } with xi +l = ßXi + Y
m (i = l, ... ,n-l) satisfy (6.0). Let A(z) = TI (l+a z)
1 r
be a po ZynormaZ wi th A(ß j ) f. 0., j O,l, ..• ,n-m-l.
208 T.S. MOtzkin & A. Sharma & E.G. Straus
Then thepe is a unique poZynomiaZ P(x) of degpee
< n-m-1 such that
(6.1) A(E)P = A(E)f
PROOF: It suffices to consider the case where
A(E)f = o. Then A(E)P = 0 leads to the successive
equations:
(6.2) (J+a E) m
=
-+ (J+a1E)P = (J+amE)
= (J+a E)Q = Q = 0 m m-1 m
where Qv(x) = Qv_1(x) + avQv_1(ßx+y), v = 1, ••• ,m,
Q (x) = P(x). Since Q (~) is a po1ynomia1 of degree o m -+
< n-m-1 and since 0 = (0 (x1), ••• ,Q (x » = 0 it 1n 1n m n-m
fo11ows that Qm(x) = o. If Qm_1(x) ~ 0 and if
then from (6.2), we have
Averaging Interpolation 209
so that a ßk + 1 = 0 , i.e., A(ßk) = 0 , contrary to m
hypothesis. Hence Q l(x) = 0 • m-Simi1ar reasoning
gives Qm_2(x) = ... = Q1(x) = P(x) = o. This
comp1etes the proof of the theorem.
THEOREM 5. x = ßXi + y i+1
(i = 1, ••. ,n-1) satisfying (6.0). suppose
A(ßj ) I 0, j = O,l, ••• ,k(n-m)-l
where k is a given integer ~ 1. Then there exists
a unique poZynomiaZ P(x) of degree ~ k(n-m)-l such
that
A(E)P(j) = A(E)1(j) , j = 0,1, .•. ,k-1
where
= (f (j ) f (j ) ) 1 ' ... , n
210 T.S. Motzkin & A. Sharma & E.G. Straus
PROOF: As in the proof of Theorem 4, it suffices to
consider the case f = 0 which leads to Q (j ) (x.) = 0, m 1
i = 1, ••• ,n-m; j = 0,1, ••• ,k-1 Since ~(x) is a
po1ynomia1 of degree < k(n-m) with k-fo1d zeros
at it follows that Q (x) = 0 . m
By
the same argument as in the proof of Theorem 4, this
imp1ies that
Qm_1(x) - - Q1(x) = P(x) - 0 •
Remark 1: If the condition A(ß j ) # 0 is violated
for some j € {0,1, ••• ,n-m-1} in Theorem 4 or for
j E {O,l, ••• ,k(n-m)-l} in Theorem 5, say a = 1 -j
-ß
then the condition Q1 - 0 does not imp1y that
p(x) = 0
arbitrary
If A(z)
then
Indeed P(x) = c{(ß-1)x + y}j with
c satisfies Q1(x) = P(x) +a1P(ßX+Y~ = 0 • j1 Jk
has severa1 distinct zeros ß , .•• ,ß ,
P(x) k . J v = lC {(ß-1)x+y} 1 v
Multiple zeros of A(z) of the form ßj do not lead
to additional free parameters.
Averaging Interpolation 211
Remark 2: We can ca1cu1ate the po1ynomia1 P(x) of
Theorem 4 exp1icit1y by a simple device. Let
l(x) =
be the po1ynomia1 determined by the conditions
Then
(6.3) A(E)P(x) = l(x) .
If ß # 1 ,put Y = x + ~ and set Q(y) =
P(y - ß~l) , where
n-m-1 Q(y) = L
o
Then (6.3) yie1ds
that is,
212 T.S. Motzkin & A. Sharma & E.G. Straus
n-m-1 c A(l) + \ c A(ßv)yV = t(y - -1-)
o L v ß-1 o
=
n-m-1 t (v) (-1-) \ 1-8 v Lv! Y o
Hence
c = t(V)(y/(l-ß» v v!A(ßV)
v = O,1, •.• ,n-m-1 •
If ß = 1 , the problem of finding P (x) becomes
simpler. Indeed, putting E = 1 + ß , where
ßf = f(x+1) - fex) and ßV = ßßv-1 , and setting
A(l+x)
we have
P(x) 1
= A(l+ß) tex) .
If
00
-1 \ v (A(l+x» = Lbvx , o
we have
Averaging Interpolation 213
p(x} =
7. RELATIVE UNISOLVENCE.
In this section we discuss the concept of relative
unisolvence introduced in [3] and extend it to finite
dimensional subspaces which are not necessarily of
co-dimension one.
DEFINITION. Let F be a Zinear spaae and Zet
L = {L.li € I} where I is some index set, be a 1
maximaZ system of ZinearZy independent funationaZs on
F so that f € F is determined by its ao-ordinates
i € I . An n-dimensionaZ subspaae F of F n
is unisoZvent reZative to L if an eZement f € F
is determined by any n of its ao-ordinates
Li f, ... ,Li f where {il, ... ,in} cl. 1 n
The examples given in [3] illustrate this general
definition when the finite dimensional subspace is of
co-dimension one.
214 T.S. Motzkin & A. Sharma & E.G. Straus
Example: Let F be the space of functions from a field
A to itself and let L = {L la E A} where L f = f(a) a a
for all a E A and all f E F • If 'JT n-l is the
subspace of polynomials of degree < n-l in A[x] -then by Lagrange interpolation, we know that f E 'JT n-l
is uniquely determined by L f, ••• ,L f al an
for any
Thus is unisolvent relative
to L.
On the other hand if we let F be the space of
formal power-series A[[x]] and let L = {Lo,Ll , ••• }
where L f = coefficient of n
n x in f , then the
subspace 'JT n-l
of polynomials of degree < n-1 is not
unisolvent relative to L since Lnf = Ln+1f = = 0
for all f E 'JT 1. n-
LEMMA 1. The space F is unisolvent relative to n
if and only if for any n + 1 functionals
L ,Ll, .•• ,L E L ~ there is a unique linear relation o n
(7.1) (L + alL 1 + ... + a L )f o n n o
L
Averaging Interpolation 215
for aZZ f E F ~ where the n
aroe non-zero saaZaros.
PROOF. It is c1ear that (7.1) imp1ies uniso1vence
relative to L. For, let L1 , ••• ,Ln be any n-tup1e
of functiona1s of L, then (7.1) shows that L1f, ••• ,
L f n
determine L f o
for any L E L o
and hence
determine f.
Converse1y, assume F unisolvent relative to n
Since F is n-dimensiona1, any n + 1 functiona1s n
L •
L , ••• ,L restricted to F satisfy a linear relation o n n
b L + ... + b L = 0 Now if one of the b. were o 0 n n l.
zero (say b = 0) o '
then there wou1d be a linear
dependence among L1 , .•. ,Ln on Fand if, say, n
wou1d determine L fand n
hence f for all f E F contrary to the fact that n
Fn is n-dimensiona1. Hence all the bi are non-zero.
Since b ~ 0 , we may assume b = 1. If there o 0
were two different relations of the form (7.1), then
e1iminating L o among them wou1d lead to a linear
dependence among L1 , ..• ,Ln again contradicting the fact
216 T.S. Motzkin & A. Sharma & E.G. Straus
that F is n-dimensiona1. This comp1etes the proof n
of the lemma.
THEOREM 6. The space nk_1 of poZynomiaZs of degree
< k < n-m is unisoZvent reZative to the functionaZs
L1 , •.• ,L of Theorem 2. n-m
PROOF. For the sake of simp1icity we prove the
theorem for the functiona1s of Theorem 1. We need to
show that for any k functiona1s
conditions
L. , •.. ,Li the 1.1 k
(7.2) = L. P = 0 1.k
app1ied to the vector with
P = P(x), v = 1, .•. ,n imp1y p(x) = 0 • \! V
As in the proof of Theorem 1 we write
A(E) = (J+u1E)A1 (E) and use the fact that the
-+ app1ication of A1 (E) to P is variation diminishing.
-+ -+ Set A1 (E)P = (Q1, •.• ,Qn-m+1) = Q. If Q is not
zero then the conditions (7.2) imp1y
Averaging Interpolation 217
= 0
= 0
-+ which in turn imp1ies that P has at least k sign-
changes so P(x) = 0 • -+
If Q = 0 , then we write
and proceed as before.
COROLLARY. L. , .•• ,L i are any 1.1 k
Suppose k func-
tionals of Theorem 2. Then there exist non-zero con-
stants (unique except for a constant factor) c1 , ... ,ck such that
(7.3) 0, f E: 1Tk_2 .
PROOF. From Theorem 6 the space 1Tk_2 is unisolvent
relative to L1 , ••• ,L and from Lemma 1, we know n-m
that there exist non-zero constants c satisfying v
the coro11ary.
218 T.S. Motzkin & A. Sharma & E.G. Straus
In order to find the exp1icit va1ue of c , we \)
observe that the po1ynomia1 P(x) € ~k-1 which
satisfies the interpo1atory conditions
L. (P) = Li (f) , \) = 1, ••. ,k, 1v \)
is given by the fo11owing determinantal equation:
1 2 k-1 P(x) x x x
Li (1) Li (x) 2 Li (xk- 1) L. (f) Li (x )
1 1 1 1 ~1
Li (1) L. (x) 2 Li (xk- 1) Li (f) L~ (x )
k ~ k k
Then
(7.4)
1 k-1 x x
= 0
0
(_1)k+1M•P(x) = Li (1) L. (x) Li (xk- 1) Li (f) 1 ~1 1 1
Li (1) Li (x) L. (xk- 1) L. (f) k k ~k ~k
Averaging Interpolation
where
M ,. ~ 0 by Theorem 6.
If f is a polynomial of degree ~ k-2 , it is
clear that P(x) = 0 , hence the coefficient of
219
k-l x in the determinant on the right in (7.4) vanishes
when f E nk_2 • Thus
Li (1) 1
Li (x) 1
Li (xk- 2) 1
Li (f) 1
(7.5) = 0
L~ (1) L~ (x) Li (xk- 2) k
L~ (f)
Since the determinant M ~ 0 the minors of its last
column are not all zero, so that (7.5) is a non-trivial
linear r ela tion among the Li (f) , i = 1, ••• , k •
Therefore by the corollary all the minors are non-zero
220 T.S. Motzkin & A. Sharma & E.G. Straus
and so the relation (7.5) coincides with (7.3). Hence
L, (1) 1 1
L, (x) ~l
L (1) L. (x) L, (xk- 2) i, 1 1, 1 1, 1 J- J- J-
e, = J L. (1) L (x) L (xk- 2)
1 j +l i j +l i j +l
L, (x) 1 k
k L, (x )
1 k
On the basis of Lemma 1, we see that if F is of
dimension (N > n) that is; F = {(fl , ... ,fN)} , and
if L = {LI"" ,LN} with
for all f ~ F , then F n
L,f = f., (i = 1, .•• ,N) 1 1
is unisolvent relative to
L if and only if it intersects each of the co-ordinate
(N-n)-subspaces only at the origin.
The best approximation problem settled in (I3],
§§2 and 3) can now be extended to the more general
situation discussed here.
Averaging Interpolation
PROBLEM: Let A be a valued field and let F be a
spaae of dimension N with aoordinates given by the
maximal system of linearly independent funationals
L = {Ll, ... ,LN}. Suppose Fn is a subspaae of
dimension n « N) unisolvent relative to L. Por
any cP E: F " find the
distanae funation
f E: F whiah minimizes a n
(7.6) Ilcp - fll = H(IL1(CP-f)I , ... ,I~(CP-f)l) •
221
where H is some (aonvex) funation of N non-negative
variables.
To illustrate the method we restriet attention to
the case when A = R , the real field and
the Euclidean distance.
By unisolvence there exist unique elements
f(l) , ... ,f(n+l) E: F so that n
L f(i) = L j j
j f i, j = l, ... ,n .
222 T.S. Motzkin & A. Sharma & E.G. Straus
If ~ e F , our problem is trivia11y solved by f = ~. n
If ~ t F , then by uniso1vence every f e F has a n n
unique expression
(7.7)
Hence
n+1 f = L A f(i)
1 i
n+1 L A. = 1 • 1 ~
We can now extremize (7.3) subject to the condition
to get
(7.9)
n+1 L A = 1 1 i
(i = 1, ••• ,n+1)
Averaging Interpolation 223
where ~ is the Lagrange multiplier determined so that
and
(7.10) ~ij =
n+l I Ai 1, 1
The equations (7.9) determine the Ai and hence f
uniquely.
8. APPLICATION.
Theorem 1 shows that if A(z) of (2.1) is a
polynomial having only negative zeros, then polynomials
of degree ~ n-m-2 are unisolvent relative to the n - m
functionals
L (f) \)
m - 1a fex + )
o ~ ~ \) \) = 1, ... ,n-m •
It follows from Lemma 1 (§7) that there exist non-zero
scalars ll, ... ,l such that n-m
224
(8.1)
T.S. Motzkin & A. Sharma & E.G. Straus
n-m ( I l L )P = 0 ,
1 v v P E 'IT 2. n-m-
We can now app1y Lemma 1 in [3] (p. 1200) and obtain
the fo11owing theorems:
THEOREM 7. Fop any funation CP.t the poZynomiaZ P(x)
of degpee < n-m-2 whiah minimizes
n-m (8.2) I ~kILk(cp-p)la, ~k > 0, a > 1
1
is unique and aan be wr'itten as
(8.3) n-m
P(x) = I AkPk(x) 1
whepe Pk(x) aY'e the poZynomiaZs dete~ined by the
interpoZatopy aonditions
(8.4) L. (Pk ) = L. (cp) , j '" k, j = 1, ••• ,n-m J J
and
Averaging Interpolation 225
where the lk are given by (8.1).
THEOREM 8. For any funation <I> ~ the poZynomiaZ p(x)
of degree n - m - 2 whiah minimizes
(8.5) max ~kILk(<I>-p)1 , ~k > 0, k = 1, ••• ,n-m • k
can be wri tten as
(8.6) n-m
P(x) = I AkPk(x) 1
where Pk(x) are the poZynomiaZs given by (8.4) and
(8.7)
9. WEAI< AND STRONG A-SIGN CHANGES.
If A(z) is a po1ynomia1 of degree m given by
(2.1) having all negative zeros, then we sha11 say that
a function fex) has n - m - 1 weak A-sign changes
on X if there exist n points < x n
in X
226 T.S. Motzkin & A. Sharma & E.G. Straus
such that for € = 1 or -1,
(9.1) m
€(-l)j La f{x +j) > 0 o II II
j = 1,2, ••• ,n-m •
If (9.1) ho1ds with > instead of ~, f{x) will be
said to have n - m - 1 strong A-sign changes. For
m = 0 , the A-sign changes coincide with the usua1
sign changes. F 1 d (m) 1 or m > an a = , et II II
f{x ) = (_1)r-1, 1 h r = , ••• ,n; t en r
m
La f{x +j) o II II
Thus the sequences
j = 1, ..• ,n-m •
{(_1)r-1}n· has n - m - 1 weak 1
A-sign changes for any m > 1 , where m A{z) = (l+z) ,
but for m = 0 it has n - 1 strong sign changes.
We sha11 say that P{x) weak1y (or strong1y)
AX-interpo1ates f{x) if there exist points
x1 , ••• ,xn € X such that P{x) - f{x) has n - m - 1
weak (or strong) A-sign changes on X.
Let n-m-1 = b 1 kX + ... , n-m- , k = 1, ••• ,n-m
Averaging Interpolation 227
be the po1ynomia1 of degree n - m - 1 such that
(9.2) k
= (-1) j = k
as j runs through 1, .•• ,n-m. The polynomials Bk(x)
are unique1y determined as is seen from Theorem 1. We
now formu1ate
LEMMA 2. AZZ poZynomiaZs T(x) of degree n - m - 1
whiah have n - m - 1 weak A-sign ahanges on X are
given by
(9.3)
whex>e
(9.4) n-m l Ak = 1 • 1
Fox> T(x) to have n - m - 1 stx>ong A-sign ahanges
on X ~ (9.4) is to be x>epZaaed by
(9.5) Akb 1 k > 0 , n-m- ,
n-m l Ak = 1 • 1
228 T.S. Motzkin & A. Sharma & E.G. Straus
The poZynomiaZs Bk (X)/bn_m_1 ,k are the onZy ones
which AX-interpoZate zero exactZy n - m - 1 times.
LEMMA 3. If a poZynomiaZ P(x) of degree n - m - 2
n-m-1 weakZy AX-interpoZates -x " then there exists
a set of positive numbers w1 ""'w such that n-m
P(x) minimizes the A-norm (p = 1)
(9.6) n-m m " I w. I a P (x +.)
j=l J ~=o ~ J
where P(x) = xn- m- 1 + p(x) •
PROOF. By Lemma 2, since P(x) weak1y A-interpo1ates
zero on X, we have
* where the Ak satisfy (9.4). Set
!bn- m- 1 ,j! * wj = if A. ;: 0 J
2!bn_m_1 ,j! * = if A = 0 j
Averaging Interpolation
Since
* -1 j =Ajb 1.(-1) n-m- ,J
we have
n-m m L wjl L aP(x+.)1
j=l ~=o ~ ~ J
For any other po1ynomia1
we have
n-m m LW. I L a T (x +.) I =
j=l J p=o P ~ J
=
where }:' extends over those
* f Aj :/: 0 and extends over
This proves the lemma.
n-m * = L A. = 1 •
j=l J
with
n-m
n-m L A = 1 , 1 k
IAjl L w
j=l j Ibn- m- 1 ,jl
, " L A. + 2 L Aj > 1
J
indices j for which
* those for which A = j
229
o .
230 T.S. Motzkin & A. Sharma & E.G. Straus
LEMMA 4. Let p > 1 be given. If a poZynomiaZ P(x)
of degree n - m - 2 strongZy A-inter-poZates
n-m-1 -x on X then there exist positive weights
w1 , ... ,wn_m such that P{x) minimizes the A-noPm:
where A n-m-1 P{x) = x + P{x) •
PROOF. By Lenuna 2,
A n-m * -1 * P{x) = L Akb _ -1 kBk{x) , Akb 1 k > 0 , k=l n m , n-m- ,
n-m * L Ak = 1. 1
Set
_ I IP * 1-p w. - b 1. (A.) • J n-m-,J J
Then 11 P IIA = 1. For any other po1ynomia1 T (x) , ,p
as in (9.3), we have
n-m L A. = 1 1 J
Averaging Interpolation 231
* which will have its minimum if A. =A ,which comp1etes J j
proof of the lemma.
These lemmas lead to the fo110wing theorems. (We
omit the proofs as they run parallel to those of
Theorem 13.10 and 13.11 in I51, p. 286-288.)
THEOREM 9. If P(x) is a poZynomiaZ of degree
n - m - 1 whioh weakZy A-interpoZates f(x) on X
(having M points), then there is a set w. J
(j = 1, ••• ,M-m) of positive numbers suoh that P(x)
is a best approximation to f(x) in the weighted
A-nomz (p = 1)
M-m m L w.1 L a {p (x +j) - f (x]1+j)} 1 •
j=l J ]1=0 ]1 ]1
THEOREM 10. If in Theorem 9, the "weakZy" is repZaoed
by "strongZy", then for any p > 1 , there is a set
w > 0 j
(j = 1, ... , M-m) suoh that P(x) is a
best approximation to f(x) in the weighted A-nomz:
lM-m m J1/P L wj l I a {P(x +.) - f(x +j)}I P , j=l ]1=0]1 ]1 J ]1
p > 1 .
232 T.S. Motzkin & A. Sharma & E.G. Straus
10. CONCLUSION.
the
It may be interesting to remark that if X is
th n row of a triangular matrix with all the
in [a,b] , then for any given m and fixed polynomials
Ai(z) a Faber-Bernstein type result holds. In fact
using the argument in [1], it is easy to see that if
IDfx (xi +l - Xi) tends to zero as n + ~ , there exists
a continuous function f such that the polynomials
P(x) of Theorem 2 do not converge to f(x) • It
would be interesting to see if other results for
Lagrange interpolation hold for the polynomials of
averaging interpolation as weIl.
1. Curtis, P.C.: polynomials. pp. 385-387.
REFERENCES
Convergence of approximating Proc. Amer. Math. Soc. 13 (1962),
2. Malozemov, V.N.: On the method of equal sums (Russian) Vestnik Leningrad Univ. 13 (1967), pp. 167-170.
Averaging Interpolation 233
3. Motzkin, T.S. and Sharma, A.: Next-to-interpo1atory approximation on sets with mu1tip1icities. Can. J. Math. 18 (1966), pp. 1196-1211.
4. Motzkin, T.S. and Wa1sh, J.L.: Least pth power polynomials on a real finite point set. Trans. Am. Math. ·Soc. 78 (1) (1955), pp. 67-71; 83 (2) (1956), pp. 371-396.
5. Rice, J.R.: The approximations of functions Vo1. 11, Addision-Wes1ey (1969).
ON THE CONSTRUCTION OF MULTIDIMENSIONAL SPLINES
Marie-Jeanne Munteanu
INTRODUCTION
Smoothing polynomial splines have been in
troduced by I.J. Schoenberg [17]. Generalizations
of this notion have been given for example by
M. Atteia [2], [3], G.M. Nielson [15].
In a previous paper we have given a general
abstract definition of splines for the combined case
of interpolation and smoothing (see [14]).
The purpose of this paper is to give several
methods of construction for some important classes of
splines which are particular cases of the general
definition mentioned above.
235
236 M. Munteanu
1. GENERAL DEFINITION.
Let X,Z be Banach spaces, Zi i = 1,m+n
Hi1bert spaces.
We want to approximate Gx, x E X where G is
continuous linear operator on X into Z, using
m + n observations h i , i = 1,m+n. Suppose
j h. = Fx, j = m+1,m+n are interpo1ating da ta and
J
the elements h i , i = 1,m represent experimental
va1ues which are approximations of the quantities
F~, i = 1,m where Fi , i = 1,m+n are continuous
linear operators on X into i = 1,m+n •
Consider U a continuous, linear and surjective
operator from X onto a Hi1bert space Y. Let us
define an operator V on X into n, where n is
the cartesian product of the spaces
12 m Vx = [F x,F x, •.. ,Fx] .
1 m Z , ••• ,Z •
Denote by ZO = Yxn the cartesian product of the
spaces Y and n and we define the operator L on
X into ZO as fo110ws
Multidimensional Splines 237
Lx = [Ux, Vx] •
We note also by hO = [O,h] € ZO , where 0 is the
origin of Y and h = [hl ,h2 , .•• ,hm] € n. As Atteia,
Anselone, Laurent Il] [2] [3] did for the case of
smoothing spline for functionals we introduce in ZO
the following norm
lI[y,Zl' ••• 'Z ]11 2 m ZO
p > 0 •
DEFINITION: We wiZZ aaZZ a generaZized smoothing
spZine tor operators, any eZement s € X whiah
minimizes the quantity:
under the aondi tions i = m+l, ••• ,m+n •
Remark:
10 • This definition combines the smoothing and
interpolating case.
20 • For the case p = ~ , we obtain the generalized
smoothing splines for operators, which we treated in
238 M. Munteanu
a previous paper [13].
3°. The mentioned authors Atteia, Anse10ne, Laurent
did not treat the eombined ease. They studied
genera1ized sp1ines for the ease when Fi are
funetiona1s.
4°. The degree of genera1ity chosen permits us to
treat severa1 interesting eases for app1ieations.
We find again some sp1ines treated a1ready by Nie1son
in [15], Atteia [2], [3] and find severa1 new types of
sp1ines.
THEOREM. (Existenae and DniquenessJ
Suppose:
whe1'e i
= 1,m} N1 = {x E XIFx = 0, i
j N2 = {x E XIFx = 0, j = m+1,m+n} .
2°. U(N2) is a aZosed bounded set.
Multidimensional Sp1ines 239
Then there exists a unique SEX such that
= Min i i
Fx=h ,i=m+ 1, m+n
For the proof of the theorem see the paper I presented
at the meeting on Approximation Theory, Michigan State
University, 22-24 March 1972. ([14]).
2. METHODS OF CONSTRUCTION.
The purpose of this paper is to present three
different methods for the construction of multidi
mensional sp1ines.
I. Variationa1 method. In a preivous paper [11],
we genera1ized for the one dimensional case a
variationa1 method given by Carasso in his doctora1
thesis [4]. The intention is to adapt this method
to the multidimensional case.
ll. The method based on Gordon's interpolation
resu1ts [5] [6], of forming minimal and maximal
projectors. for the case of more than two variables,
I studied in my doctora1 thesis intermediate projectors
240 M. Munteanu
[12]. Ana1ogous projectors will be considered in this
case. We will present these methods for classes of
splines in severa1 variables, which are particu1ar
cases of the general definition.
111. The method of Anse10ne and Laruent. We will
app1y the projection method given by the mentioned
authors for certain c1asses of multidimensional sp1ines.
3. METHOD 1.
Let us consider X the space of rea1-va1ued
functions f, which are defined on R = I'xI",
1'=[0 1] I" = [0,1] such that the fo11owing , x' y
derivatives exist almost everywhere respective1y on
R, I', Irr and have the fol1owing properties
D(i,j)f(x,y) € C(R), i < p, j < q
(p-1,j) Df(x,O) is abs. cont. ,
(p ,j) Df(x,O) € L2(I') j = 0,q-1
(i,q-1) Df(O,y) is abs. cont. ,
(i,q) 2 Df(O,y) € L (I") i =O,p-l
(p-1,q-l) Df(x,y) is abs. cont. ,
(p ,q) Df(x,y) € L2[R] •
Multidimensional Sp1ines
For all f be10nging to this space we have the
representation
f(x,y) p-1 q-1 i j (i,j)
= L L ~, ~ f(O,O) i=o j=o ~. J.
q-1 j 1 p-1 (p,j) + L ~ f ~x-~~r Df(t,O)dt
j=o J. 0 p-
p-1 i 1 q-1 (i,q) + \ x f (y-u)+ Df(O,u)du
L iT (q-1)! i=o 0
p-1 1 1 (x-t)+
+ f f (p-1)! o 0
q-1 (p.q) (y-u)+ Df(t,u)dtdu (q-1) !
241
One can prove that this space is comp1ete with respect
to the norm furnished by the inner product:
p-1 q-1 (i,j) (i,j) (f,g)x = L L f(O,O)g(O,O)
i=o j=o
p-1 1 (i,q) (i,q) + L f Df(O,u)Dg(O,u)du
i=o 0
242
q-1 1 (p,j) (p,j) + L f Df(t,O)Dg(t,O)dt
j=o 0
1 1 (p,q) (p,q) + f f Df(t,u)Dg(t,u)dtdu
o 0
M. Munteanu
p-1 q-1 (i,j) (i,j) = L I f(O,O)g(O,O) + [f,g]x •
i=o j=o
We noted by [f,f]x a seminorm with the null space
{x~yj}p-1 q-1. For the proof of these assertions i=o j=o
see A. Sard [16] and L. Mansfie1d [10]. Indeed X
admits the fo11owing decomposition
Let U be the operator
Uf = {(p,O) (p,l) (p,q-1) (p,q) Df(t,O),Df(t,O) .•. Df(t,O) ; Df(t,u);
(O,q) (l,q) (p-1,q)} Df(O,u),Df(O,u), ••• ,Df(O,u)
and Y the Hi1bert space
Multidimensional Sp1ines
The kerne1 of the operator U is the set
{xi i}p-1 q-1 • Y i=o j=o
Remark: I have chosen the Tay10rian functiona1s in
order to simp1ify the presentation.
The Fi are functiona1s linear and continuous
on X of the form
1 1 L(f) = I J J
i<p 0 0 j<q
(i ,j) ij f(t,u)dll (t,u)
243
+ I 1 (i,j) "j
J f(t,ß)d1l 1 (t) + I 1 (i,j) i"
J f(a,u)dll J(u) •
where each
i<p 0
i,j II
j<q 0
is of bounded variation on its
domain. It is a natural way to deduct from the
variationa1 definition, using ca1cu1us of variations,
the ana1ytica1 expression of the sp1ine function s.
App1ying the method which I treated in one dimensional
case in [11], we obtain in a simi1ar way
244 M. Munteanu
s(x,y) = p-1 q-1 . . i. m+n . l l a~,Jx yJ + l A.F~ [K(x,y,t,u)]
i=o j=o i=l ~ u
where the coefficients fo11ow from the conditions
i = 1,m
j = m+1,m+n
where
K(x,y,t,u) = (G(x,y,~,n),G(t,u,~,n»y •
G(x,y,t,u) is defined as:
p-1 q-1 i j (i,j) f(x,y) = l l ~, ~ f(O,O) + (G(x,y,t,u),Uf(t,u»y.
i=o j=o . J
The function K(x,y,t,u) is the reproducing kerne1
of the orthogonal comp1ement of the set {xi yi}p-1,q-1 i=o j=o
which is the null space of the operator U.
Multidimensional Splines
q-l p-l i t i 1 (y-n)+
K(x,y,t,u) = L ~! i! f (q-l)! i=o 0
q-l (u-n)+ (q-l) ! dn
p-l q-l j j 1 (x-S)+
+ L ~ ~, f (p-l)!
p-l (t-S;)+ dS; (p-l) ! j=o J. J. 0
245
1 1 p-l p-l q-l q-l + f f (x-O+ (t-S;)+ (y-n)+ (u-n)+ cl;dn
(p-l) ! (p-l)! (q-l) (q-l) I o 0
One can verify that the resulting spline function s
minimizes p(f) under the restrietions F~ = Fl = Ljf
= hj j = m+l,m+n. Indeed one finds that
p(f) - pes) ~ 0 , where
o 0 i=o 0
See for proof. G. Nielson [15].
The uniqueness is assured by the hypo thesis
246 M. Munteanu
Ker U N2) = 0 X ' and we suppose also that
i i i {F [K(x,y,t,u)], i = 1,m+n; x Y , i = 0,p-1, j tu
is a 1inear1y independent set.
4. METHOn 2.
0,q-1}
A very important case will be when the operators
Fij = L iM j are product of linear continuous funcx y
tiona1s L i defined on the one dimensional space x
defined on the space
We note by S the genera1ized sp1ine in one x
variable relative to the operator nP and the
functiona1s Li. x
We set x = HP[O 1] , x' 2 Y = L [0,1] , x
U = nP , and we suppose that the hypothesis which assures
the existence and unicity of S is valid (see the x
general theorem of the section 1).
Similarly we introduce the operator S Y
The
Multidimensional 8plines
operators 8 and 8 are projections. x y Using
Gordon's procedure (see {5] [6]) we can form the
analogue of minimal approximation for the case of
two variables.
81 = 8 8 • x Y
We consider the same spaces as in §3. The spline
function 81 will minimize the quantity ~(f) under
247
the conditions Fjs = hj j = m+l,m+n. (~(f) defined
as in section 3).
Let us form now the analogue of maximal
approximation:
We choose the operator U being n(p,q) , the space
being
R = [0,1] x [0,1] , the operators F x y
(a,O) f(x,y) + f(xi,y)
(O,ß) f(x,y) + f(x,y.)
J
a < p
ß < q
248
(a ,ß) f(x,y) ~ f(xi'Yj)
Corresponding to the operators
M. Munteanu
a < p, ß < q •
ia F we choose the x
spaces Zi being the space of functions fex) defined
on [0,1] such that the p-th derivative exists almost x
everywhere on [0,1] ,is an element of x
and such that the Tay10r formu1a
fex)
p-1 p-1 i (i) 1 (x-t)+ (p)
= L ~, f(O) + f (p-1)! f(t)dt, i=o . 0
2 L [0,1] , x
X E [0,1] , x
is valid. This is the Sobo1ev space HP [O,l] • x
Corresponding to the operators ·ß
FJ we choose the space y
zj being the Sobo1ev spaces Hq [O,l] • Obvious1y y
corresponding to the functiona1s FiaFjS Zij will x Y ,
be the euc1idian space R.
The quantity which is minimized by S2 is the
following
Multidimensional 8plines
1 1 1(f) = f f (D(P,q)f)2dtdu + l l (FiaFjßf
o 0 ia jß
1 + l f [Fia(Dqf) - hia]2du
ia 0
+ l fl[~ß(DPf) - h jß ]2dt • jß 0
One can easily prove that ~(f) - ~(82) ~ 0 • «see
G. Nielson [15]). For the three dimensional case we
can form for example
888 x y z
corresponding to the minimal approximation,
82 = 8 + 8 + 8 - 8 8 - 8 8 - 8 8 + 8 8 8 x y z x y x z y z x y z
corresponding to themaximal approximation,
83 = 8 8 + 8 8 + 8 8 - 28 8 8 x Y x z Y z x Y z
corresponding to an intermediate approximation.
In my doctoral thesis [12] I treated the inter
mediate approximations including the limiting cases
of minimal and maximal approximations, for the n
249
250 M. Munteanu
dimensional case. Analogously we can form the
corresponding spline functions choosing in an
i appropriate way the spaces Y, Z , the quantity
~(f) , as we did for the two dimensional case.
5. PROJECTION METHOD.
Anselone and Laurent presented the projection
method for the case of interpolation [1], and for the
case of smoothing [8]. The purpose of this section is
to adapt the projection method for the case of
smoothing splines in several variables.
We will present first the general definition
of smoothing splines given by Atteia in [2] [3] and
describe the projection method of construction as
given by Laurent in [8].
Let X and Y be two Hilbert spaces and T a
linear continuous operator on X onto Y. Denote by
N the null space of the operator T; we suppose
dirn Ker T = n
Let Li' 1 < i 2 m , where m ~ n , be con
tinuous linearly independent functionals defined on
Multidimensional Sp1ines
X. We denote by K the subspace spanned by the
Riesz representers of these functiona1s.
We designate by z = Y x E the cartesian m
product space, endowed with the sca1ar product
251
(p > 0) •
We define the operator L on X into Z
Lf = ITf,Af]
We put a = [O,e] E Z, 0 being the null element of
Y •
We sha11 refer to a genepaZized smoothing spZine
as an element s of X which minimizes the quantity
L(s) = IITs 11; + pllAs - eil; = IILs - all~ • m
We denote by N~ and K~ the orthogonal comp1ements
of N and K.
Under the hypothesis ~ N n K = 0 , one can
demonstrate the existence and unicity of s (see
252 M. Munteanu
Atteia [2], [3]).
Now we give Laurent's resu1ts concerning the con
struction of the genera1ized smoothing sp1ine [8].
LEMMA 1. The suhspaae H = K n N.L is of dimension
m - n •
LEMMA 2. The kernel * G of the adJoint operator L
is a suhspaae of Z of dimension m - n •
LEMMA 3. If h i -, 1 2 i 2 m - n -' form a basis of H-,
we can introduce in G the following basis
1 < i < m - n
where
1 < i < m - n •
THEOREM. Under the hypo thesis N n K.L = 0 -' there
exists a unique spline s whiah minimizes L(s) ; viz.
m-n Ls a + I Aig.
i=l l.
where the coefficients A. are the solutions of the l.
Multidimensional Splines 253
linear system
The steps of the construction are the following:
1. One seeks a base hi , l<i<m- n , of
H = K n rI m
hi = I bikj l<i<m- n • j=l
2. One deduces the base ~i
1 < i < m - n •
3. Then follows the base
4. Then one can write
254 M. Munteanu
m-n Ls - a = ): \gi
l.=1
the coefficients Ai are given by the system (*)
5. From Ls = [Ts,As] OIle has fina11y
m-n Ts = L Ai~·
i=l l.
As - e
We will app1y this method to the fo110wing c1ass of
smoothing sp1ines in two variables. Let us consider
X the space of rea1-va1ued functions, defined on
R = I'xI", I' = [0 1] , x' I" = [0 1] such that , y'
the fo110wing derivatives exist almost everywhere
respective1y on R, I' , I" , and
(x,y) ER,
D(2m-j ,j)f(x,O) E L2(I') , x E I', j < m ,
Y EI", i < m ,
Multidimensional Splines
and such that the Taylor formula
f(x,y) \' xi l. n(i,j)f(O 0) L .,. , '
i+j<2m 1.. J.
1· 1 ( ) 2m-i-I y-u + (i 2m-i)
+ .L ~! 1 (2m-i-I)! n' f(O,u)du 1<m 0
. 1 ( )2m- j -I ~ x-t + (2m-j j)
+ L ., 1 (2m-j-I)! n 'f(t,O)dt j<m J. 0
m-I m-I (x-t)+ (y-u)+ (m m)
+ 1I (m-I)! (m-I)! n ' f(t,u)dtdu R
holds for any (x,y) ER.
With the inner product
(f,g)X = I n(i,j)g(o,o)n(i,j)g(O,O) i+j<2m
1 + I 1 n(i,2m-i)f(0,u)n(i,2m-i)g(0,u)du
i<m 0
+ I IIn(2m-j ,j)f(t,0)n(2m-j ,j)g(t,0)dt j<m 0
+ Iln(m,m)f(t,u)n(m,m)g(t,u)dtdu R
255
256 M. Munteanu
X is a Hilbert space.
Thespace X admits in fact the following
decomposition
where ~ designates the number of functionals
D(i,j)f(O,O), i + j < 2m •
Let T be the operator defined by
Tf = {D(2m,O)f(t,O);D(2m-l,1)f(t,O); ••• ;
D(m+l,m-l)f(t,O);D(m,m)f(t,U);D(O,2m)f(O,u);
D(l,2m-l)f(O,u); ••• ;D(m-l,m+l)f(O,u)}
and Y the Hilbert space
We see easily that the kernel of T is the set
of polynomials of degree less than or equal to 2m - 1
Multidimensional Splines 257
in x and y Then Ker T and
dim Ker T = ~ < 00
We consider n(n ~W continuous, linearly independent
functionals of the same form as in the Section 3. We
will suppose that the subspace spanned by the Riesz
representers of these functionals contains the kernel
of the operators T. Denote by e = [el ,e2,··· ,en] € En
the vector symbolizing the experimental values of the
given functionals for f ( X •
Let us denote with s(x,y) the corresponding
smoothing spline.
In order to employ the projection method
previously explained, we must choose n - ~ func
tionals 0., 1 ~ i ~ n - ~ , linearly independent, 1
defined on X, and having the following properties
where
the functionals
~ . I Aj Nj
1 (f) j=l
p = l,~,
to the set of given functionals,
q = l,n-~, belong
258 M. Munteanu
20 the n - ~ chosen functionals exhaust the n
given functionals.
One verifies immediately that if we choose
o (f) = q
q q q q q i j q 2m-I) q" N (I)N (x)N (xy)N (y) ••• N (x Y ) "+" 2 ••• N (y N.l.f) ~ ~ p p P 1 J< m ~ p
we have
o (f) = 0, Vf E Ker T, q q I,n-p •
Choosing these functionals we used Jerome and
Schumaker's idea for the case of univariate splines.
(see [7]).
If we appIy the functionals o , q
to the representation of f we obtain
q I,n-p,
Multidimensional Splines
o (f) q
[ , 2m-i-I~ 1 x1 (y-u)+ (i 2m-i)
L f 0 7f (2 -'-1)1 D' f(O,u)du i<m 0 q 1, m 1 •
~ , 2m-j-l~ 1 J (x-t) + (2' ') + L f 0 ~ (2 -'-1)1 D m-J,J f(t,O)dt , qJ. mJ . J<m 0
259
(y-u)+ (m m) m-I] (rn-I)! D ' f(t,u)dtdu.
We deduce the base q = I,n-)l
(Y_U)m-I~ ~(X_t)m-Il (Y_U)m-I~ -,---..,+,..-- 0 + +
(m-I) ! ' q (m-I) ! (m-I)! '
r m-I (x-t) m~ } 8 Y + q t(m-I) ! m!
E Y , q = I,n-)l
and the scaIar products
260 M. Munteanu
(x-t)+ (y-u)+ ~ m-l m-~ + ffR öq (m-l) I (m-l) I
u
Concerning the base
(x-t)+ (y-u)+ ~ m-l m-l~ öl (m-l)! (m-l)! dtdu.
b , q
q = l,n-~ , we can
obtain the components of the vectors by expanding the
functionals Ö, q = l,n-~ according to the elements q
of the last column.
Having the bases ~l' l = l,n-~ and
b l , l = l,n-~ we can write the base gl' l = l,n-~
Then we have
n-~
Ls - a = lAg q=l q q
Mu1tidimensiona1 Sp1ines 261
where Lf = [Tf,Af] € Z, Z= Y xE, Af € E being
the vector having the n components {L kM . .eJk€J' ,l€J"o x J
The coefficients Aare the so1utions of the
system
a = [O,e] € Z ,
We deduce
and
q
n-J.\ Ts = L A <P
q=1 q q
l = 1,n-J.\
262 M. Munteanu
Fina11y in order to find s(x,y) we use the
representation (1)
s(x,y) =
Remarks:
i j (. .) L .;-~ D 1.,J s(O,O) .+. 2 1. J. 1. J< m
i 1 + L ~J
i<m 1.. °
. 1 ~ + L . ,1
j<m J. °
T)-].l + L JJ
q=l R u
(y-u) 2m-i-\_].l ~ i ( lm-i~l + \' A 0 ~ -..,.y_-_u'-":+---'-l-
(2m-i-1)! q:1 q q i! (2m-i-1)! du
2m-j-1 ~ (x-t) T)-].l j + 0 L
(2m-j-1)! I q j! q-1 -
m-1 m-1 (x-t)+ (y-u)+
(m-1) ! (m-1) !
(x-t) 2m-j -ll
(2m-~-1) ! j dt
~x_t)1l1.-1 o +
·q_(m-1)!
m-1~ (y-u)+ (m-1) ! dtdu.
1°. This method can be app1ied also for the case of
smoothing sp1ines described in the section 3.
2°. The construction of the genera1ized sp1ines in
severa1 variables treated by Atteia in his doctora1
Multidimensional Sp1ines
thesis [3], can be treated in an ana10gue manner.
It is c1ear that the sp1ines introduced by Atteia
263
are a particu1ar case of the general definition given
in §1.
30 . We can also app1y to the c1ass of smoothing sp1ines
treated in this section the variationa1 method of Section
3. In fact in this case we can consider the more general
case of interpolation plus smoothing. We obtain again
the ana1ytica1 expression of the sp1ine function given
by Nie1son in his doctora1 thesis [15].
REFERENCES
1. Anse10ne P.M. and Laurent, P.J.: A general method for the construction of interpo1ating or smoothing sp1ines-functions, Num. Math., 12,1968, 68-82.
2. Atteia, M.: Fonctions-sp1ine genera1isees, C.R. Acad. Sei. Paris, t. 261, 1965, 2149-2152.
3. Atteia, M.: Theorie et app1ications des fonctionssp1ines en analyse numerique, These, Grenob1e, 1966.
4. Carasso, C.: Methodes numeriques pour l'obtention des fonctions-sp1ine, these de 3-eme cyc1e, Universite de Grenob1e, 1966.
264 M. Munteanu
5. Gordon, W.J.: Sp1ine-b1ended interpolation through curve networks, J. Math. Mech. 18, (1969), 931-952.
6. Gordon, W.J.: Distributive 1attices and the approximation of mu1tivariate functions, in Approximation with special emphasis on sp1ine functions, I.J. Schoenberg, ed. Academic Press, N.Y., 1969, 223-277.
7. Jerome, J. and Schumaker L.: A note on obtaining sp1ine functions by the abstract approach of Laurent, MRC Technica1 Summary Report #776, August 1967, Madison, Wisconsin.
8. Laurent, P.J.: Representation de donnees experimentales a l'aide de fonctions sp1ine d'ajustement et evaluation optimale de fonctionne11es 1ineaires continues, Ap1ikace Math., 13, 1968, 154-162.
9. Laurent, P.J.: Cours de theorie de l'approximation, Fascicu1e 3, Facu1te des Sciences de Grenob1e, 1967-1968.
10. Mansfie1d, L.E.: On the optimal approximation of linear functiona1s in spaces of bivariate functions, SIAM J. Num. Anal. 8 (1971), 115-126.
11. Munteanu, M.J.: Observatii asupra solutiei optimale a unei probleme diferentia1e ne1iniare cu va10ri 1a 1imita, pe subspatii de functii sp1ine genera1izate, Bu11. Sei. lnst. Po1ytech. C1uj, 1 (1968), 47-56.
Multidimensional Sp1ines 265
12. Munteanu, M.J.: Contributions a 1a theorie des fonctions sp1ines a une et a p1usieurs variables, Doctora1 thesis, Univ. de Louvain, Be1gium, January 1971.
13. Munteanu, M.J.: Genera1ized smoothing sp1ine functions for operators, SIAM J. Numer. Anal. Vo1. 10, No. 1, March 1972.
14. Munteanu, M.J.: Multi-dimensional smoothing sp1ine functions, Symposium on approximation theory and its app1ications, March 22-24, 1972, Michigan State University.
15. Nie1son, G.M.: Surface approximation and data smoothing using genera1ized sp1ine functions, Doctora1 thesis, Univ. of Utah, June, 1970.
16. Sard, A.: Linear Approximation, Am. Math. Soc., Providence 1963.
17. Schoenberg, I.J.: Sp1ine functions and the problem of graduation, Proc. Net. Acad. Sei., 52, 1964, 947-950.
ON ERROR ESTIMATES APOSTERIORI IN ITERATIVE
PROCEDURES
A.M. Ostrowski
1. 1939, in a note in the C.R. of the Paris
Academy, 209, 777-779, I introduced the concepts of
the computation errors apriori and aposteriori, the
first being estimated be fore the beginning of the
essential computation and the second being deduced
after severa1 steps of the computation have been
comp1eted.
If in ametrie space the sequence x tends to v ~ the estimate aposteriori of the error, /xv'~/'
has to be deduced using the values a1ready computed
of x1 , ••• ,xv and, if possib1e, of the correction
/xv ,xv+1 / .
267
268 A.M. Ostrowski
Usua11y we have the situation where for a q,
o < q < 1 ,
(1) (v + (0) •
However, this eannot be used for the estimate a pos
teriori of Ix ,si . v
The situation is eomp1ete1y different if we have
the recurrent estimate,
In this ease we have
(3) 1 ---<
l+q-Ixv,sl 1
Ixv,xv+1 1 ~ 1 - q
(3a)
and the estimate (3) gives obviously a pretty e10se
evaluation of Ix ,si . v
Error Estimates aPosteriori
2. We have a more general situation if (2) is
rep1aced with
(4) (\! = 0,1, ... )
where ~(x) is positive and < 1 with x > 0 and
non-increasing with x ~ o. In this case we have
(5)
269
where ~(Ix\!,sl) has to be rep1aced in each case with
a convenient majorant.
An important special case is if for a sequence
00
~\! with 1 > ~\! > 0 , TI ~ \! = 0 , we have \!=1
(6) (\! = 1,2, .•. ) .
Then it fo11ows
(7) 1 Ix\!,sl
1 + ~\! < IX\!+l,xl 1
2. -1---~- . \!
For instance we cou1d use the sequence 1
~\! = 1 - v
270 A.M. Ostrowski
3. In the case tha t x and 1;; are n-dimenv
siona1 vectors, n > 1 , and generally xV +1 = Av Xv
with quadratic matrices A , the fo110wing estimate v
can be found in the literature *):
(8)
which can be used if the norm "Av " corresponding
to the chosen vector norm, is < 1 .
4. In the above case we have for the error
vectors ~ : = X - 1;; the relation v v
(9) t: v+1 = A ~ v v (v 0,1, •.• ) .
If A = A does not depend on v the iteration v
is convergent for any choice of ~o iff AA' the
*) Cf. for instance, J. Weissinger, Ueber das Iterationverfahren, ZAMM 31 (1951), p. 245.
Error Estimates aPosteriori
speetral radius of A, is < 1. And in this ease
we have *)
(v -+ 00) •
But this again eannot be used for estimates a
posteriori.
However, it follows from (9)
C = (I _ A )-l(~ C ) ~v v v - ~v+l '
assuming that I - A is non-singular, and therefore,
using the euelidian norms,
(10) 1 11 - A I v e
5. In order to use (10) we have to obtain
eonvenient estimates for 11 - AI, I (I - A)-ll e e
*) Cf. for instanee, A.M. Ostrowski, Ueber Normen von Matrizen, Math. Z. 63 (1955), p. 5, formula (11).
271
272 A.M. Ostrowski
for a general matrix A. To obtain such estimates
assume that the matrix A = (a ) llV
eigenva1ues AV so that AA = M~x
use the so-ca11ed Frobenius norm of
(11)
of order n has
I AI. We will v
A
n Here we have a1ways lAI; ~ L IAv I2 so that we can
v=l put
(12) t:,.A
where t:,.A i9 a "measure for the norma1ity of the
matrix A" and in particular t:,. A = 0 iff A is normal.
In this case lAie = AA •
6. In the general case we have
(13)
Error Estimates aPosteriori
(14)
The last formu1a can on1y be used if a convenient
estimate of Idet (I - A)I can be found.
7. Better estimates can be obtained if we
assume, instead, that ÄA is known and is < 1 •
273
Then we have, using convenient1y a resu1t by Henrici*),
(16)
(17)
I (I - A)-ll < 1 e-1-t. -Ä A A
*) P. Henrici, Bounds for iterates, inverses, spectra1 variation and fie1ds of va1ues of non-normal matrices, Numer. Math. 4 (1962), p. 30, theorem 3.
274
(18)
(19) I (I _ A)-ll < n ( ~A )n e - 1 - A 1 - A A A
A.M. Ostrowski
In the relations (16) - (19) AA and ~A can be
rep1aced (simultaneous1y, both in the conditions and
assertions) by arbitrary majorants as long as the
majorant of AA remains < 1 •
The simp1est majorant of ~A is of course IAI F
On the other hand, a c10se estimate of ~A in terms
of * * AA - A A due to Henrici*) is known:
(20) H3 n - n * * ~ A 2. 12 ;I A A - AA I F •
*) 1.c.p. 27, formula (1.6)
Error Estimates aPosteriori 275
8. It may be of interest to observe that the
argument of sec. 1 and 2 can be genera1ized to more
general situations. If we have, for instance, instead
of (2) the so-ca11ed weakZy ZineaP convergence *),
(21) (v = 0,1, ••• )
where N is an integer > 1 , we have, instead of (3):
(22)
*) Cf. A.M. Ostrowski, Solution of Equations and Systems of Equations, 2d. edition (1966), p. 204.
277
SPLINES AND HISTOGRAMS
I.J. Schoenberg
INTRODUCTION.
In [3] Boneva, Kendall and Stefanov (B.K.S.) have
effectively rediscovered the essential features of what
I like to call cardinal cubic spline interpolation.
Moreover, and this is an important point, the data are
not the usual function values that are to be inter
polated, but rather approximations of the derivative
(i.e. the unknown density function) in the form of a
histogram. This (pershaps only apparent) difference is
bridged by the ingenious area-matching condition.
In [10] I carried out a suggestion of J.F.C. Kingman
Sponsored by the United States Army under Contract No. DA-3l-l24-ARO-D-462.
278 I.J. Schoenberg
(see [3, 55]) and applied variation diminishing cubic
spline approximations to histograms. Now I believe that
this approach smoothes the data too strongly and the
MRC Report #1222 is not going to appear elsewhere.
Actually, the formation of a histogram is already a
strong form of smoothing (or fluctuation-reducing) and
further smoothing should be done with care. Here we do
no further smoothing and area-matching (or volume
matching in the bivariate case) is done exactly.
In the present paper I am describing the application
of finite spline interpolation (S.I.) to histograms in
one and in two dimensions. The results used from uni
variate S.I. are weIl known. The results concerning
bivariate S.I. seem to be new (Theorems 3 and 4 below)
in spite of the lively activity in this field (see [1]
and [6], also for further references). This is perhaps
not surprising in view of the novel statistical setting
of these problems on the one hand, and the number of
possible variations on the theme of bivariate S.I. on
the other.
In the present paper probability considerations
and criteria are conspicuous by their absence. A
mathematical analyst can provide the statistician with
new tools of approximation. The statisticians must
decide on their usefulness and their reliability at
Splines and Histograms
different levels of probability.
I. THE UNIVARIATE CASE
1. THE MAIN RESULTS.
Let
(1.1) H = (h.), (j = 1,2, •.• ,m) J
279
be a histogram, where h. denotes the frequency in the J
interval (j-l,j), and where observations that fall on
the common boundary of adjacent class-intervals count
1. h f h as 2 1n eac 0 t ese. The following crucial defini-
tion is due to B.K.S.
DEFINITION 1. We say that the integrable funation
f(x) , defined in [O,m], enjoys the area-matahing
property for the histogram H, provided that
(1.2) / j-l
f(x)dx = h. J
(j = 1, ... , m) •
Let the symbol AM(H) denote the alass of funations
satisfying this aondition.
In [3] B.K.S. extend the definition of h. J
to all
280 I. J. Schoenberg
integers j by setting
h = 0 if j < 1 or j > m , j
and construct a cardinal quadratic spline function a(x)
that satisfies the area-matching condition (1.2) for all
integer values of j •
There is some advantage in restricting ourselves
to the class of functions AM(H) , whose elements are
defined in IO,m] only. A ready source of potentially
useful elements of AM(H) is afforded by finite S.I.
as folIows. We associate with Hits corresponding
cumulative sequence
(1. 3)
defined by
(1.4)
F = F = (F) (j = 0,1, ... ,m) , H j
1, ... , m) •
Furthermore, let
(1. 5)
denote the class of spline functions (S.F.) of degree
k , defined in IO,m] , and having the points
x = 1,2, .•. ,m-l as simple knots. This means that the
restrietion of Sex) to [j-l,j] is a polynomial of
Splines and Histograms 281
degree not exceeding k , while k-l
Sex) € C IO,m].
Observe that if k = 1 and Sl(X) is the linear
S.F. (or continuous piece-wise linear function) such
that
(1. 6) Sl(j) = Fj , (j = O, ••• ,m) ,
then
(1. 7)
is easily seen to be a step-function whose graph is
identical with the geometric representation of H by
rectangles Rj of area h j (or height h j ) and basis
[j-l,j]. It is weIl known that the interpolatory
conditions (1.6) can also be met by elements of
Sk[O,m] , provided that k is odd (= 3,5, ••• ) , and
that appropriate boundary conditions (B.C.) are pre
scribed for Sex) • We single out the two most useful
cases, k = 3 and k = 5 , and state the known results
as lemmas (see e.g. [8, §13]).
LEMMA 1. Thepe is a unique aubia spZine S3(x) suah
that
(1. 8) S 3 (j) = F j' (j = 0, ••• , m) ,
and
282 I.J. Schoenberg
(1.9) S'(O) = F' S'(m) = F' 3 0' 3 m '
whepe F' and F' have ppeassigned vaZues. o m
LEMMA 2. Thepe is a unique quintic spZine S5(x) such
that
(1.10) S 5 (j) = F j' (j = 0, ••• , m) ,
and
(1.11) S;(O) = F~ S"(O) = F" 5 0
S'(m) = F' 5 m
S" (m) = F" 5 m
whepe F', ... ,F" aPe ppeassigned. o m
In the present paper we discuss on1y histograms H
that may be polymodal, but are assumed to be "be11-
shaped" in the sense that they have thin tai1s. Accor
ding1y, we shaZZ assume that h1 and hm aPe smaZZ
compaped to max h.. For this reason we sha11 assume l.
in our present app1ications that all boundary va1ues
become
(1.12)
F(s) , are taken to be zero, and (1.9), (1.11), m
S3(0) = 0, Sj(m) = 0
8p1ines and Histograms 283
and
(1.13) 8;(0) = 8S(0) = 0, 8;(m) = 8S(m) = 0 .
The reason for this choice of vanishing boundary va1ues
will become c1ear in our discussion fo11owing Theorem 2
be1ow.
If the sp1ine function 8(x) satisfies (1.8), or
perhaps (1.10), it fo11ows that its derivative
(1.14) er (x) = 8' (x)
is an element of the c1ass AM(H) , for
(1.15) / er(x)dx = / 8'(x)dx 8 (j) - 8 (j -1) j-1 j-1
F. J
F. 1 = h. J- J
(j 1, ... ,m) ,
by the very definition (1.4) of the cumu1ative sequence
(Fj ). From Lemmas 1 and 2 we therefore get the
fo11owing theorems.
THEOREM 1. There is a unique
such that
284 LJ. Schoenberg
(1.17)
THEOREM 2. There is a unique
suah that
(1.19)
PROOFS: Invoking Lemmas 1 and 2 it suffices to set
(1. 20)
in \-!ew of (1.15). It is also c1ear that (1.17) and
(1.19) are imp1ied by (1.12) and (1.13), respective1y,
in view of the definitions (1.20).
Fo11owing B.K.S. we ca11 02(x) and 04(x) the
histosp1ines of degrees 2 and 4, respective1y. Observe
that the B.C. (1.17), (1.19) tend to give their graphs
the required beZZ-shaped ahaY'aater.
It might not be irrelevant to mention some optimal
properties enjoyed by the histosplines. They fo11ow
from known properties of S.I. and may be stated as
Sp1ines and Histograms
folIows.
COROLLARY 1.
1. The quadratic histospline a2(x) of Theorem 1
minimizes the integral
m (1. 21) J
o
2 (a' (x» dx
among all functions a(x) that belong to AM(H) and
satisfy the B.C. (1.1?).
2. The quartic histospline a4(x) of Theorem 2
minimizes the integral
m (1. 22) J
o
2 (a"(x» dx
among atz functions a(x) that belong to AM(H) and
satisfy the B.C. (1.19).
Of course, we a110w for competition on1y such
a(x) for which the corresponding integrals (1.21),
(1.22), make sense.
285
286 I.J. Schoenberg
2. A FIRST CONSTRUCTION OF TRE QUARTIC HISTOSPLINE
G4(x) •
Mrs. Julia Gray, of the MRC Computing Staff, wrote
a computer program, called spZint, that furnishes the
quintic spline Ss(x) and G4(x) = S;(x) of Theorem
2. The input data are the integer m (~4) and the
values of the m + 5 quantities F , •.• ,F , o m
F' , ... ,F" . o m
For statistical applications to bell-shaped distribu
tions we choose F = 0 ,and F' = ••• = F" = 0 • o 0 m
The program uses the quintic B-spline (see [9, 70-
71]
(2.1)
where u+ = max (O,u). In terms of this function, a
general element of SS[O,m] may be written uniquely in
the form m+2
(2.2) L -2
and the program computes the coefficients c. J
for the
solution of the interpolation problem (1.10), (1.11).
Splines and Histograms
Sex) represnets an approximation of the (cumulative)
distribution function of the statistic that furnished
H. An approximation of the density function is the
histospline
m+l (2.3) °4(x) Ss(x) = L
-2
where
287
(2.4) E-oo < x < (0) •
Thus Splint solves numerically any problem (1.10),
(1.11), of so-called complete quintic S.l. for equi
distant data. The examples given in §3 below were
computed by means of Splint. In §4 we present an
alternative numerical approach that furnishes 04(x)
direc tly, and will help to clarify our proc edure in
dealing with the bivariate case.
3. TWO EXAMPLES.
1. The age distribution of Bulgarian mothers of 1963.
From [3, 21] we take the following table giving the
age distribution of 50226 Bulgarian mothers during the
year 1963
288 1.J. Schoenberg
j 1 2 3 4
Age group
h. 7442 19261 14385 6547 J
j 5 6 7
Age group
h. 2123 451 17 J
The corresponding histogram is shown in Figure 1 where
we have changed sca1e and origin by setting
age = 15 + 5x, (0 < x ; 7) .
The frequencies h1 , .•. ,h7 are the areas (and there
fore also the heights) of the seven rectangles of
Figure 1, of which the last does not show at our
sca1e or ordinates due to its sma11 height.
The curve shown in Figure 1 is the quartic
histosp1ine 04(x) of Theorem 2. The graph shows
c1ear1y its area-matching property. The curve owes its
nice1y ba1anced shape perhaps to the optimal property
Splines and Histograms
of minimizing the integral (1.22).
We are now going to subject our histospline to a
severe test. In [3, 23] we also find the histogram
with 7 x 5 = 35 entries corresponding to observed
*
289
* H
annuaZ frequencies. We have reproduced H graphically
also in Figure 1. How weIl does our 04(x) match the
* areas of H ? As we see from Figure 1, some of the
* areas of H are matched weIl, some less so, especially
those in the interval .8 < x < 2.2. Dur histospline
underestimates the observed high annual frequencies in
the interval .8 < x < 1.4 (ages 19-22). These obser
vations suggest the following comments.
Dur results (Theorems 1 and 2) remain valid, up
to notational changes, if we choose unequaZ class
intervals. Let the lengths of the class-intervals
(x. l'x.) be denoted by l. = x. - x. 1 ' while the J- J J J J-
corresponding frequencies are again h. , as before. J
The area-matching requirement now amounts to the
relations
(3.1)
x f j
f(x)dx h. J
(j = 1, ... , m) ,
290
o o o Q
c-(
1. J. Schoenberg
~~~~~--------------------~~
~--~~----------------~------------~~
o o ~
o o o -
Splines and Histograms
while the height of the rectangle R. J
is now = h.II. J J
The program Sptint is no longer applicable, and must
be replaced by the efficient programs developed by
C. de Boor (see I4] for references). These are also
based on the use of B-splines, but they use B-splines
with unequat spacing of knots.
291
We should regard our equidistant G4 (x) as purely
diagnostic. In fact Figure 1 shows that the interval
[.8, 1.4] should be chosen as one of the class-inter
vals, in order to trap within this interval the high
frequencies peculiar to that age group. Would perhaps
a choice of class-intervals resulting in frequencies
h j that do not vary (essentially) with h , be
advantageous? More experimentation seems called for.
A last remark suggested by this example deals with
fact that the histospline G4 (x) may assume negative
values. Clearly all hj ~ 0 ; from (1.4) it follows
that (Fj ) is a nondecreasing sequence. From (1.10)
we conclude that the sequence (SS(j)) is non
decreasing. Unfortunately, this does not imply that
SS(x) is a non-decreasing function in [O,m] In
such cases the derivative G4 (x) will assume also
292 1.J. Schoenberg
negative va1ues. In our case of the Bu1garian mothers
we find that a(1.6) = 20184.305 , whi1e at the right
hand tai1 we have the tab1e of va1ues
x 6.0 6.2 6.4 6.6 6.8 7.0
a 4(x) 107.792 42.512 6.093 -6.513 -4.259 .000
This is due to the sma11 va1ue of h 7 = 17. Because
of the area-matching condition, this is bound to happen
in an interval (j-1,j) where hj is sufficient1y
sma11 compared to max h .• J
2. The distpibution of Zengths of eaps of cOPn. We
consider the data from I1, 93] giving the distribution
of the 1engths of 578 ears of a certain lewa variety of
corn, grouped into 14 histogram ce11s of equa1 widths,
the first ce11 representing a 1ength range of 10.5 -
11.5 cm. and the last a range of 23.5 - 24.5 cm. The
data are as fo11ows
j 123456 7 8 9 10 11 12 13 14
(3.2) 1 5 20 38 50 110 110 104 66 44 18 10 1 1
Splines and Histograms 293
B.K.S. show in [3, 13, Figure 2] the graph of the
cardinal quadratic histospline fitted to the data (3.2).
Their curve shows four relative extremes (polymodality).
Our Figure 2 shows the histogram H = (h.) and the J
corresponding quartic histospline 04(x) of Theorem 2
is (of the two curves shown) the one that is also
endowed with 4 extreme points. In fact our curve and
B.K.S.'s are not very different.
At this point the author decided to try some
further smoothing by the simple device of doubling
the length of all cZass intervals. In this way we
obtain 7 class intervals, all of length 2, and (3.2)
gives the new histogram
described by the table
j 1 2 3
(3.3)
* H
4 5
(j = 1, ... , 7)
6 7
* h. 6 58 160 214 110 28 2 J
Notice that the heights of the new rectangles are
* * h./2 • J
The corresponding histospline 04(x) was
computed by splint and its graph is also shown in
* Figure 2, as weIl as H * The appearance of 04(x)
294 1.J. Schoenberg
seems to indicate that the doub1ing of the c1ass
intervals was just about the correct amount of smoothing
that the data (3.2) required. Even so, we did not quite
* escape the b1emish of negative va1ues of 04(x) . Whi1e
* 04(3.4) = 113.312 is c10se to the maximal va1ue, we
found that
x 6.5 6.6 6.8 7.0
.344 -.085 -.055 .000 * ° 4 (x)
* We fina11y remark that the graph of 04(x) is rather
c10se to the curve obtained by I.J. Good (see [3, 54,
Figure 6]) for the data (3.2) by an entire1y different
approach.
4. A SECOND CONSTRUCTION OF 04(x) .
Let
(4.1) {cr (x) }
denote the subspace of S4[O,m] of those elements
o(x) that satisfy the B.C.
Splines and Histograms 295
Figure 2
" L - - - - - ~"',~--~--t--------f
o o - o r-
....
296 I.J. Schoenberg
(4.2) a(O) = a'(O) = 0, a(m) = a'(m) = 0 .
The main idea is to work within this subspace (4.1) and
to find a convenient basis for it.
A basis meeting all requirements is furnished by
I7, Theorem S, 81-82] as fo110ws. The general4th
degree B-sp1ine is obtained from
(4.3)
in the form
M(x;t) 4 = 5(t - x)+
This is the fifth-order divided difference of (4.3),
with respect to the variable t, and based on the
six points
(4.5)
x. 1
satisfying the conditions
The function (4.4) is defined for all real x, it is
positive in (xo'xS) and vanished everywhere in its
comp1ement. M(x) E C3(R) as 10ng as the knots (4.5)
are all distinct. However, if x. 1
4-v mu1 tiplicity v, then M(x) E C
is a knot of
near Finally,
Splines and Histograms
M(x) is so normalized that
00
(4.6) J M(x)dx = 1 . _00
For simplicity, let us assume that
(4.7) m > = 5 .
From [7, Theorem 5] it follows that we obtain a
basis of (4.1) as folIows: We consider the knots
x = ° and x = m as tripZe knots and write accor
dingly
(4.8) ° , ° , ° , 1, 2, . . • , m-l , m , m , m •
We group these m + 5 elements in all possible ways
into 6 aonseautive ones (call one such group
xo ,xl , ••• ,x5) and form the corresponding B-splines
(4.4). The m B-splines so obtained form a basis of
* S4[O,m] • We repeat this statement as
LEMMA 3. If UJe UJY'ite
~l(x) M(x;O,O,O,1,2,3)
~2(x) = M(x;O,O,1,2,3,4)
297
298 I.J. Schoenberg
(4.9)
~3(x) = M(x;0,1,2,3,4,5)
~4(x) = M(x;1,2,3,4,5,6)
~m_2(x) = M(x;m-5,m-4,m-3,m-2,m-1,m) ,
~m_1(x) = M(x;m-4,m-3,m-2,m-1,m,m) ,
~ (x) = M(x;m-3,m-2,m-1,m,m,m) , m
then these m funations form a basis for the spaae
defined by (4.1), (4.2).
The notations used in (4.9) assumed (4.7). If
m < 5 then obvious changes are necessary. Thus in the
extreme case that m = 1 , then the basis (4.9) reduces
to the single element
122 = M(x;O,O,O,l,l,l) = 30 x (1 - x) in [0,1] •
Every element of (4.1) having a unique represen
tation
(4.10) m
a(x) = L y~~~(x) 1
we may now inforce the area-matching property
Sp1ines and Histograms
i (4.11) f cr(x)dx = h.
i-I ~
as folIows. Writing
(4.12) A. ~a
i
= f i-I
(i=l, ... ,m)
4> (x)dx a
and substituting (4.10) into (4.11) we obtain the
relations
(4.13) h. ~
(i = 1, ... ,m) •
299
This is a non-singular system since it defines cr(x)
unique1y. Due to the structure of the functions (4.9),
the system (4.13) has a 5-diagona1 matrix.
Solving the system (4.13), we obtain the Ya and
then cr4 (x) = cr(x) is expressed by (4.10). To use this,
we need to know the 4> (x). These are given by the a
fo110wing exp1icit formu1ae:
(4.14) 5 4 27 4
4> (x) = -- {(3 - x) - -- (2 - x) + 27(1 1 54 + 4 +
if x ~ 0 ,
300
(4.15)
I.J. Schoenberg
~ ( ) = ~ {(4 - x)4 _ 16 (3 _ x)4 '+'2 x 96 + 3 +
4 4 + 12(2 - x)+ - 16(1 - x)+}
if x ~ 0 ,
whi1e both these functions are o if x < 0 .
Moreover
(4.16) (-00 < X < 00) ,
~4(x) ~3(x - 1)
~5(x) ~3(x - 2)
(4.17)
~m_2(x) ~3(x - m + 5) ,
and fina11y by symmetry
(4.18) ~ (x) = ~ (m -m-1 2 x) , ~ (x) ~1 (m - x) . m
The elements (4.12) of the matrix IIA. 11 are 1(1.
obtained from (4.12) and (4.14) - (4.18) by direct
Splines and Histograms 301
integration. These elements are rational numbers.
Reducing them to their least common denominator, which
is 4320 = 25335 we obtain that
(4.19) 4320 IIA.all = 1
2300 525 36 o o
1940 2595 936 36
80 1155 2376
o 45 936
o 0 36
o o
45 0
1155 80
2595 1940
o 525 2300
This matrix is symmetrie with respect to its center.
The elements of the columns 3,4, •.• ,~2, are all
identical with the numbers
(4.20) 36 xl, 36 x 26 , 36 x 66 ,
36 x 26, 36 xl,
302 I.J. Schoenberg
appropriate1y shifted. Thus, if m = 5 , then (4.13)
becomes
2300 525 36 0 0 Y1 hJ.
1940 2595 936 45 0 Y2 h2
80 1155 2376 1155 80 Y3 = 4320 h3
0 45 936 2595 1940 Y4 h4
0 0 36 525 2300 Y5 h5
We may use (4.19) even if m = 4 , when the co1umns
with elements (4.20) are missing a1together.
11. TRE BIVARIATE GASE
5. TRE MAIN RESULTS.
Para11e1ing the deve10pments of §1 we discuss
the approximation of bivariate density functions by
means of biquadratic and biquartic sp1ine functions.
Let us exp1ain the problem and our termino1ogy.
Let
(5.1) R = [O,m] x [O,n]
Splines and Histograms 303
be a rectangle in the (x,y)-plane of dimensions m
and n, where m and n are natural numbers. We
think of it as dissected into mn unit squares
(5.2) Rij = Ii-l,j] x Ij-l,j] •
A bivapiate histogpam
(5.3) H = (hij ), (i = l, ••. ,m; j = l, ••• ,n) ,
is a matrix of observed frequencies, where h.. is 1J
the number of observations falling within the square
Rij •
DEFINITION 2. An integpabZe funation f(x,y) ~ defined
in R ~ is said to possess the voZume-matahing ppopepty
with pespeat to H ~ ppovided that
(5.4) 11 f(x,y)dxdy = hij for all (i,j). Rij
We denote by the syrriboZ VM(H) the aZass of funations
that matah the mn voZumes of H •
DEFINITION 3. Let k be a natuPaZ nurribep. We denote
by
(5.5) Sk,k(R) = {S(x,y)}
304 1.J. Schoenberg
the class of functions S (x,y) " defined in R" and
having the following properties:
(5.6) In each Rij , S (x,y) is of the form
k k a(i,j)xayß S (x,y) = L L
11.=0 ß=o aß '
and hence depends on (k + 1)2 parameters.
(5.7) The k2 partial derivatives
are continuous in R.
Our main resu1ts are the fo110wing two theorems.
THEOREM 3.
1. There is a unique G2(x,y) such that
(5.8)
and satisfying the B.C.
(5.9)
along the entire boundary of R.
Sp1ines and Histograms
2. Among aZZ functions f(x,y), defined in R, and
satisfying the three conditions:
(i) f (x,y) E VM(H) ,
(5.10) (ii) f(x,y) = 0 if (x,Y) E aR ,
305
(iii) f(x,y) is absoZuteZy continuous in the
sense of [5, 653],
the biquadPatic spZine cr 2 (x,y) has the optimaZ
property
(5.11) f f (f (x,y»2dxdy > f f (cr2 (x,y»2dxdy , R xy R ,xy
unZess f(x,y) = cr(x,y) throughout R.
THEOREM 4.
1. There is a unique cr 4(x,y) such that
(5.12)
and satisfying the B. C.
(5.13) cr4 (x,y) = cr 4 (x,y) = cr4 (x,y) = 0 ,x ,y
if (x,y) E aR •
306 1. J. Schoenberg
2. Among all functions f(x,y), defined in R, and
satisfying the thpee conditions:
(5.14)
(i) f(x,y) E VM(H) ,
(ii) f(x,y) = f (x,y) x
= f (x,y) = 0 y
if (x,y) E aR ,
(iii) The mixed paPtial depivative f (x,y) xy
absolutely continuous in the sense of
[5, 653],
is
the biquaptic sp Une °4 (x, y) has the op tima l ppopepty
(5.15) f f (f (x,y»2dxdy > f f (04 (x,y»2dxdy , R xxyy R' xxyy
unless f(x,y) = 0 4 (X,y) thpoughout R.
According to [5, 654, Satz 1 und Satz 2] f(x,y) -
is abso1ute1y continuous in R if and on1y if it
admits in R a representation
x y x y f(x,y) = f f g(x,y)dxdy + f gl (x)dx + f g2(y)dy + C ,
o 0 0 0
where g(x,y), gl (x) , and g2(x) , are summab1e
functions.
Sp1ines and Ristograms 307
6. ON TRE REPRESENTATION OF BIVARIATE SPLINES.
In the proofs that are to fo110w we are going to
concentrate main1y on the biquartic case of Theorem 4.
Let
(6.1) * S4,4(R) = {o(x,y)}
denote the subspace of S4,4(R) of those elements
o(x,y) that satisfy the B.C.
(6.2) o(x,y) = 0 (x y) = 0 (x,y) = ° if (x,y) € eR , x' y
our immediate aim being to state an ana10gue of Lemma
3 of §4. For this purpose we need the m B-sp1ines
(4.9) and also n further B-sp1ines, functions of
y , defined by
(6.3)
1/I1 (Y) = M(y;0,0,0,1,2,3)
1/I2(Y) = M(y;0,0,1,2,3,4)
1/I3(Y) = M(y;0,1,2,3,4,S)
1/In-2(y) = M(y;n-S,n-4,n-3,n-2,n-1,n)
1/In-1(y) = M(y;n-4,n-3,n-2,n-1,n,n)
1/In (Y) = M(y;n-3,n-2,n-1,n,n,n) •
308 I.J. Schoenberg
* LEMMA 4. The spaae S4,4(R) has 4the dimension mn
and
* (6.4) a (x, y) € S 4 , 4 (R)
impZies a unique representation in R of the form
(6.5)
where
a(x,y) = m n I I y .. 4>. (x)1/Jj (y)
i=l j=l 1J 1
are appropriate aonstants.
* That (6.5) furnishes on1y elements of S4,4(R)
is c1ear, because all functions 4>i(x)~j(Y) be10ng to
this space. To save space we omit the e1ementary but
long proof of the converse statement.
Remark: earl de Boor adds the fo11owing interesting
observations: 1. The first part of Theorem 3 fo11ows
from de Boor's 1962 resu1t concerning bicubic sp1ine
interpolation (for reference see [1, 278]). 2. On
the basis of Lemma 4 the first part of Theorem 4
fo11ows from Theorem 2 by an appeal to general pro
perties of the tensor product. For details see
de Boor's Appendix to the present paper.
Sp1ines and Histograms
7. PROOFS OF THE FIRST PARTS OF THEOREUS 3 AND 4.
PROOF FOR THEOREM 4: In terms of the histogram (5.3)
we define the cumu1ative matrix
(7.1) F = (F ij) (i = 0, ... ,m; j 0, ... ,n)
by setting
(7.2) F. = F . 1.,0 O,J
= 0 F .. = 1.,J
if i > 0, j > 0
In terms of cr(x,y) , defined by (6.5), we define
x y (7.3) S(x,y) = J J cr(u,v)dudv, (x,y) ER.
o 0
This is a biquintic sp1ine function, and (7.3) shows
that
(7.4) cr(x,y) = S (x,y). xy
From (7.3), (7.4), and (6.2), we see that S(x,y)
satisfies the B.C.
(7.5) S(x,O) = 0, S(O,y) = 0
if 0 < x ~ m, 0 ~ y < n ,
309
310 I.J. Schoenberg
(7.6) S (x,y) = S (x,y) = S (x,y) = 0 xy xxy xyy
if (x,y) € aR •
Finally, from (7.3) and (6.5), we find that
m n x y (7.7) S(x,y) = I I Yij (! ~i(u)du)(! wj(v)dv)
i=l j=l 0 0
and that it depends also on mn parameters.
LEMMA 5. The paroameteps Y ij can be uniquely detep
mined such that
(7.8) S(i,j) = Fij , (I! i ~ m, 1 ~ j ~ n) .
PROOF OF LEMMA 5. It suffices to show that if
(7.9) S(i,j) = 0 fop alZ (i,j)
then
(7.10) Yij = 0 fop alZ (i,j).
Assuming (7.9) to hold, we conclude that
Sp1ines and Histograms
X \I (7.11) S(x,\I) = ~ lj Yij (! ~i(u)du)(J ~j(v)dv)
1 0 0
(\I fixed integer)
is a quinticsp1ine vanishing if x = O,l, ..• ,m .
Moreover, its first and second derivatives vanish
at both ends: From (7.11)
(7.12) Sx (x, \I) ~. (v)dv , J
\I (7.13) Sxx(x,\I) = L L Yij~~(x) J ~j(v)dv,
o
and both vanish if x = 0 , or x = m , because all
311
~i(x) have the property that ~i (0) = ~! (0) = ~i(m) = 1
~i(m) = 0 . We conc1ude from Lenuna 2 ( §l) that
S (x, \I) = 0 if o < x < m •
Therefore also S (x,\I) = 0 vanishes identica11y, for x
each \I . Now (7.12) imp1ies that
m n \I
L ~i(x) L Y .. J ~. (v)dv = 0 if 0 < x < m , j=l 1J J = i=l 0
312 1.J. Schoenberg
and from the linear independence of the ~i(x) we
conc1ude that
n \I
(7.14 ) L Yij J n=l 0
l/Jo (v)dv = 0 ~
for each
Keeping i fixed we observe that
(7.15) I YiJo JY l/JJo(v)dv
j=l 0
i and each \I •
is a quintic sp1ine in Y, vanishing for Y = O, ... ,m
by (7.14), and having the first two derivatives
which also vanish if y = 0 or y = n 0 Again by
Lemma 2 we conc1ude that the quintic sp1ine (7.15)
vanishes identica11y. Therefore also its derivative
vanishes identica11y so that
if o < y < n •
From the linear independence of the l/Jj(Y)
that (7.10) indeed ho1ds.
we see
Sp1ines and Histograms 313
We return to the proof of Theorem 4. By Lemma 5
we know that (7.8) are satisfied by a unique S(x,y) •
It is now easy to show that the quartic sp1ine a(x,y) ,
defined by (6.5), or equiva1ent1y, by (7.4), satisfies
all the conditions of the first part of Theorem 4.
Since * a(x,y) E S4,4(R) we know that the B.C. (5.13)
are satisfied. We are yet to show that
(7.16) a(x,y) E VM(H) •
To show this we use (7.3), or (7.4), and observe that
JJR a(x,y)dxdy = S(i,j) - S(i-1,j) - S(i,j-1) ij
+ S(i-1,j-1)
F .. - F. 1 . - F. j 1 + F. 1 . 1 1J 1-, J 1, - 1- , J -
= h .. 1J
in view of the definition (7.2) of the F ..• 1J
A proof of the first part of Theorem 3 is entire1y
simi1ar and not any shorter. It is based on ana10gues
of Lemmas 3 and 4 for quadratic and biquadratic sp1ines,
respective1y. The ro1e p1ayed by Lemma 2 is taken over
by Lemma 1 on cubic sp1ine interpolation. Fina11y, the
314
partial derivatives o , x
1.J. Schoenberg
o ,do not appear and therey
fore the B.C. (7.6) are to be rep1aced by
S (x,y) = 0 if (x,y) E aR . xy
8. PROOFS OF THE SECOND PARTS OF THEOREMS 3 AND 4.
PROOF OF THEOREM 3. Let 02(x,y) = o(x,y) be the
biquadratic sp1ine of the first part of Theorem 3,
and let f(x,y) satisfy the three conditions (5.10).
We first note the identity
JJ (f -0 )2dxdy R xy xy
2 JJ (0 ) dxdy R xy
- 2JJ (f -0 )0 dxdy R xy xy xy
and wish to show that the last integral vanishes.
In view of (5.9), and (5.10)(ii), we see that the
difference
(8.2) ~(x,y) = f(x,y) - o(x,y)
satisfies the B.C.
Sp1ines and Histograms
(8.3) ~(x,y) = 0 if (x,y) E aR •
In any case we may write
ff (f - er )er dxdy RXY xyxy
n m
ff ~ er dxdy R xy xy
n m = / {f ~ er dx}dy =
o 0 xy xy / {j er d ~ }dy o oxyxy
315
However, from (8.3) we conc1ude that ~ = 0 on the two y
vertica1 sides of dR. On integrating by parts the
last inside integral, we therefore find the last
repeated integral to be
n m -/ {J ~ er dx}dy
o 0 y xxy
m n -J {J ~ er dy}dx yxxx
o 0
and by (8.3) this is
m n = J {J ~er dy}dx
o 0 xxyy
We have just shwon that
-J J ~ er dxdy R y xxy
m n -J {j er d ~ J dx
o 0 xxy y
JJ ~er dxdy. R xxyy
316 I.J. Schoenberg
(8.4) ff (f - 0 )0 dxdy RXY xyxy
1f (f - 0)0 dxdy. R xxyy
Observe thet 0 (x,y) is a step-function that has a xxyy
constant value inside eaah square We
may therefore write the last integral as
L c.j(ff fdxdy - JJ odxdy) i,j 1 R. . R ..
1J 1J
and all terms of this sum clearly vanish because f as
weIl as 0 belong to VM(H) •
Therefore (8.1) reduces to
J J (f ) 2dxdy = R xy
However, the last integral vanishes if and only if
f = 0 holds almost everywhere in R, so that xy xy
everywhere
f (x, y) x y
J J f (u,v)dudv o 0 xy
for all (x,y) ER.
x y J J 0 (u,v)dudv = o(x,y) , o 0 xy
Splines and Histograms 317
PROOF OF THEOREM 4. The proof is very similar and only
slightly more elaborate. Let a4(x,y) = a(x,y) be
the biquartic spline of the first part of Theorem 4.
It satisfies the B.C.
(8.5) a = a = a = 0 along 3R , x Y
while the Itarbitrarylt function f(x,y) satisfies the
simila r B. C.
(8.6) f = f = f = 0 along dR. x y
In order to establish the inequality (5.15) we write
an identity similar to (8.1), with f , xy
a xy replaced
by f xxyy' a respectively. xxyy , Again we wish to
show that its last integral vanishes. Writing
(8.7) ~(x,y) = f(x,y) - a(x,y) ,
this amounts to showing that
(8.8)
We first write
318 1. J. Schoenberg
n m ff ~ cr dxdy
R xxyy xxyy f {f ~ cr dx}dy
o 0 xxyy xxyy
n m f If cr d ~ }dy o 0 xxyy x xyy
However, ~ = f x x
cr = 0 a10ng aR and therefore x
,j, - 0 '+'xyy
a10ng the two vertiaat sides of aR.
The last repeated integral is therefore
n m -f {f cp cr dx}dy =
o 0 xyy xxxyy -ff cp cr dxdy xyy xxxyy
R
m n m n -f {f cp cr dy}dx xyy xxxyy
o 0
-f {f cr d cp }dx xxxyy Y xy o 0
Since a10ng aR we conc1ude that
o
a10ng the two horizontat sides of aR.
Sp1ines and Histograms
We may therefore integrate the last integral by parts
and find it to be
m n = f {f ~ cr dy}dx =
o 0 xy xxxyyy ff ~ cr dxdy •
R xy xxxyyy
We repeat the entire operation once more as fo110ws:
n m
319
= -J {J ~ cr dx}dy = o 0 y xXXXYYY
-Jf ~ cr dxdy y xxxxyyy R
m n m n = -f {f cr d ~}dx
o 0 xxxxyyy y = f {f ~cr dy}dx
o 0 xxxxyyyy
= ff ~cr dxdy • xxxxyyyy R
Observe that the last integration by parts was 1egiti-
mate even though cr xxxxyyy is not a continuous functio~
The reason: The integration by parts was performed with
respect to y, for a fixed vaZue of x , so that
cr is a aontinuous funation of y • xxxxyyy
Fina11y the last integral vanishes for the same
320 1. J. Schoenberg
reason as in the proof of Theorem 3: a is a xxxxyyyy
step-function, while ~ matches the volumes of the
"zero"-histogram. Hence (8.8) is established and the
analogue of (8.1), that we started from, establishes
the inequality (5.15), possibly with the equality sign.
In fact we do have equality in (5.15) if and only if
(8.9) f (x,y) = a (x,y) xxyy xxyy
almost everywhere in R.
If we integrate both sides of (8.9) over the rectangle
[O,x] x [O,y] we find that
f (x,y) - f (x,O) - f (O,y) + f (0,0) xy xy xy xy
(8.10) = a (x y) - a (x 0) - a (O,y) + a (0,0) xy' xy' xy xy
if (x,y) ER.
On the other hand, from (8.5) and (8.6) we find, by
partial differentiation, that f and a vanish xy xy
on aR, so that (8.10) reduces to
f (x,y) = a (x,y) in R. xy xy
Splines and Histograms 321
One more integration of both sides over IO,x] x IO,y]
shows that indeed f(x,y) = o(x,y) everywhere in R.
9. A CONSTRUCTION OF THE BIQUARTIC HISTOSPLINE 04(x,y).
Not only did the B-spline representation of Lemma
4 furnish the foundation of a proof of the first part
of Theorem 4, but it also allows to solve conveniently
the numerical problem involved, as folIows. We know by
lemma 4 that
(9.1) o(x,y)
represents the most general biquartic spline satisfying
the B.C. (5.13). In order to satisfy also the volume-
matching conditions
(9.2) ffR o(x,y)dxdy ij
h .. 1J
We substitute the expression (9.1) into (9.2). Using,
as in (4.12), the symbols
(9.3) i
f ~a(x)dx i-I
j J 1)JS(y)dy j-l
322 I.J. Schoenberg
we obtain the system
m n (9.4) I I A. BjßY ß = h .. for all (i,j).
a=l ß=l ~a a ~J
This, then, is the linear system of mn equations in as
many unknowns Yaß that has to be solved.
In (4.19) we have given the numerical values of the
Aia The matrix IIBj ß 11 has the same structure and in
fact the same elements appear in both, since
All = Bll = 2300/4320, A12 = B12 = 525/4320 a.s.f.
Of course, they differ in their sizes.
How sparse is the matrix of the system (9.4) ?
This question is easily answered: The matrix IIAia 11
has, by (4.19), exactly m + 2(m-l) + 2(m-2) = Sm - 6
positive elements, and similarly IIBjßl1 has Sn - 6
positive elements. It follows that the number of non
vanishing (actually positive) coefficients of the
system (9.4) is exactly equal to
(9.5) (Sm - 6)(5n - 6) •
Sp1ines and Histograms
2 2 This is considerab1y sma11er than mn x mn - m n
especia11y if m and n are not sma11. Thus if
323
m = n = 7 we find that among the 49 2 = 2401 possib1e
elements of the matrix of (9.4) exact1y 841, or 35%,
are positive.
Mrs. J. Gray, who wrote the program Splint of §2,
will hopefu11y soon also write a computer program for
the solution of the system (9.4).
Dur last comment concerns the choice of the
coordinate axes. This requires some exp1anations.
Suppose that we are given a scatter diagram (S.D.),
or set of observed points, indicated in Figure 3 by an
ova1-shaped figure. Let xOy be the original
coordinate axes. In the case of Figure 3 it wou1d not
be advisab1e to work within a rectang1e.
R = {O ~ x < m, 0 < y < n} ,
for many of the ce11s R .• 1J
wou1d carry vanishing
frequencies h.. . 1J
Dver all such ce11s the histosp1ine
G4 (X,y) wou1d assume sma11 va1ues of both signs and
to no good purpose. It seems reasonab1e in this case
to change to the new coordinate system x'o'y' and to
324
enclose the S.D. within a rectangle
I.J. Schoenberg
R' = {O ~ x' < m' = '
o ~ y' ~ n'} that would clearly show fewer cells with
vanishing frequencies. If convenient, the frame
x'o'y' need not even be orthogonal, since all results
remain valid for oblique axes. This approach is clearly
only then possible (expecting translation ofaxes) when
all the original observations are available and not
merely a histogram.
Splines and Histograms 325
( ?n,', ",,' )
o
Figure 3 •
326 I.J. Sehoenberg
REFERENCES
1. Ah1berg, J.H., Nilson, E.N. and Wa1sh, J.L.: The theory of sp1ines and their app1ieations, Aeademie Press, New York/London, 1967.
2. B1iss, C.I.: Statist1es in Bio1ogy, Vo1. 1, Me Graw-Hi 11 , New York, 1967.
3. Boneva, L.I., Kenda11, D.G. and Stefanov, I.: Sp1ine transformations: Three new diagnostie aids for the statistiea1 data-ana1yst, J. of the Royal Statistiea1 Soe., Series B, 33 (1971), 1-70.
4. de Boor, C.: On ea1eu1ating with B-sp1ines, J. of Approximation Theory, 6 (1972), 50-62.
5. Caratheodory, C.: Vorlesungen über reelle Funktionen, Seeond Edition, B.G. Teubner, Leipzig-Ber1in, 1927.
6. Car1son, R.E. and Hall, C.A.: On pieeewise po1ynomia1 interpolation in rectangular polygons, J. of Approx. Theory, 4 (1971), 37-53.
7. Curry, H.B. and Sehoenberg, I.J.: On Po1ya frequeney funetions IV. The fundamental sp1ine funetions and their limits. J. d'Ana1yse Math. (Jerusa1em), 17 (1966), 71-107.
8. Grevi11e, T.N.E.: Introduetion to sp1ine funetions, 1-35 in Theory and app1ieations to sp1ine funetions (T.N.E. Grevi11e, Ed.), Aeademie Press, New York/ London, 1969.
Sp1ines and Histograms
9. Schoenberg, I.J.: Contributions to the problem of approximation of equidistant data by ana1ytic functions, Quart. App1. Math. 4 (1946), 45-99, 112-141.
10. Schoenberg,I.J.: Notes on sp1ine functions 11. On the smoothing of histograms, MRC Tech. Sumo Report #1222, March 1972, Madison, Wisconsin.
327
, , APPENDIX TO SPLINES AND HISTOGRAMS BY I.J. SCHOENBERG
Car1 de Boor
It is the purpose of this appendix to point out
the tensor product structure of the construction in
part 11, thus re1ating it to known resu1ts concerning
interpolation by tensor product sp1ines. The addi
tional a1gebraic machinery required seems worth
knowing since one may use it to advantage for the
efficient ca1cu1ation of bivariate histosp1ines.
1. ODD DEGREE SPLINE INTERPOLATION AND RELATED
INTERPOLATION PROBLEMS.
329
We begin with a recapitu1ation of known resu1ts
(see e.g., I8]) concerning odd degree interpolation on
arbitrary meshes.
330 C. de Boor
Let
m ~ = (x.)
1 0
be a partition for the interval Ia,b] , i.e.,
We denot~by Sk(~) the class of spline functions of
degree k, defined on Ia,b] and having the points
x1"",xm_1 as simpZe knots. This means that Sk(~)
is contained in cCk- l ) [a,b] and that the restrietion
of each sex) E Sk(~) to the interval (xj_l,Xj ) is
a polynomial of degree not exceeding k, (j = l, ••• ,m).
Sk(~) is a linear space of dimension k + m , and,
according to [7, Theorem 5], a basis for Sk(~) is
given by the sequence
of B-splines, with
and
k = (k + l)(t - x)+
Appendix to Splines and Histograms
f a , for i < 0
xi -1 b , for i > 0 .
LEMMA Al. POP given f € C(k)Ia,b] 3 thepe exists
exactty one s € S2k+l(~) such that
331
s(a) f(a), SI (a) = f' (a), ..• ,s (k) (a) = f(k) (a) ,
s (x.) = f (x.), i = 1, ... , m-l 1. 1.
s (b) = f (b) s'(b) = f'(b), .•• ,s(k)(b)
Denote this s by P f Then ~
a a
top att g € C(k)[a,b] with absotutety continuous k-th
and squaPe-integpabte (k+l)st depivative othep than
g = P f top which P g = P f . ~ ~ ~
Now observe that two functions fand g agree
at xo, .•. ,xm if and only if fand g agree at one
of these points, say at x a, and o
332
f(x.) - fex. 1) = g(x.) - g(x. 1)' i ~ ~- ~ ~-
If f is abso1ute1y continuous, then
hence
x fex) f(a) + ! f' (y)dy
f(x.) - fex. 1) ~ ~- f
a
x. ~
f' (y)dy
C. de Boor
1, ... ,m •
Two sbso1ute1y continuous functions fand g agree,
therefore, at xo, ... ,xm if and on1y if
x. ~
x. ~
f(a) g(a);! f' (y)dy f g' (y)dy i 1, .. . ,m .
Since
Lemma Al has therefore the fo11owing
CORROLLARY. POP evepy
exaat Zy one
f s C(k-1) [a,b] ~ thepe exists
~ denoted by p (1) f ~ fop UJhiah 'IT
Appendix to Splines and Histograms
X. l.
J Xi
s(x)dx = J f (x) dx
333
j = 0, ... , k-I ,
i=I, ... ,m.
Furthep, odd-degpee spZine inteppoZation is peZated to
this apppoximation scheme by
(P (1) f) (x) Tr
x = (p J f(y)dy) (1) (x)
Tr a
Schoenberg's Theorems 1 and 2 and their corollaries
follow from this. More generally, one has
THOEREM Al.
Zet AM(H)
m Fop a given peaZ sequence H = (hi)l '
denote the cZass of functions
g E c(k-l)[a,b] satisfying
(i) g(k-l) is absoZuteZy continuous and
g(k) E L2 [a,b]
334
(H)
X. 1
f X. 1 1-
C. de Boor
g(x)dx = h. 1
i = 1, ... ,m
(Hi) g(1) (a) = gO) (b) = 0, j = 0, ..• ,k-1 •
Then, S2k(~) and AM(H) have exaatly one element in
aommon, denoted by
A (H) , ~
and this element A (H) also uniquely minimizes ~
b (k) 2 f [g (x)] dx a
over g E AM(H) •
In order to compute A (H) , consider the basis ~
for S2k(~)' One checks that
~(j)(a) ~ 0 if and on1y if i
~~j)(b) ~ 0 if and on1y if 1
i - j-k+1l, i = m+k-jJ
j = 0, ... ,k-1
Appendix to Splines and Histograms
Hence, after leaving off ~l-k'···'~o and
~ nrl-l' ... , P nrl-k , the remaining sequence
~l'···'~m
is a basis for
* {sES 2k (1T) IsO) (a) s (j) (b) S 2k (1T) = 0,
Let now A be the m x m matrix given by
x. ~
j
A(i,j) J ~. (x)dx , J
i,j l, ... ,m. x i - l
Then
A (H) 1T
m L Y.~.
. 1 ~ ~ ~=
335
O, ••• ,k-l}.
with r (Yi ) the solution of the linear system
Ar = H .
Before deriving the corresponding results for
bivariate splines, we introduce some language concerning
linear interpolation problems, which we have found
336 C. de Boor
convenient in the discussion of tensor product schemes.
If ~l""'~m is a sequence in the linear space
F , and Al" .. ,An is a sequence of linear functionals
on F, then we may consider the Linear InterpoZation
ProbZem (LIP) given by (A.) : to find, for J
given f E F , a linear combination g = L.a.~. 111
of
the ~. 's so that 1
A.g 1
A.f 1
i l, ... ,n.
Actually, the problem does not depend on the
individual
span
cfJ i 's and A. 's, but only on the linear J
of ~. 's and the linear span 1
of the Ai 'so This is quite clear for the ~i 's,
since "a linear combination g = L.a.~. of the ~1' 's" 1 1 1
Appendix to Splines and Histograms 337
is equivalent to "a g E: Fl " which makes no reference
to the individual Pi 'so But, also,
if and only if
A.g A.f i = l, ... ,n 1. 1.
E.a.A.g 1. 1. 1.
Hence, the LIP can also be stated: to find, for
given f E: F ,a g E: Fl so that
Ag Af, for all A E: A •
We say tht the LIP given by Fl and A is
correct if it has exactly one solution for every f E: F.
The following lemma is not difficult to prove:
LEMMA A2. Let
(Aj)~ be a basis for A. Then
Ci) the LIP given by Fl and A is correct if and
only if the Gramian matrix
338 c. de Boor
is invertibZe; in partiauZar~ n = m is a neaessary
(but not a suffiaient) aondition for the aorreatness
of the LIP.
(ii) if the LIP given by FI and A is aorreat~
then the interpoZant Pf for given f E F aan be
aomputed as
Pf with
In these terms, Lemma Al states that the LIP
given by FI = S2k+I(n) and
A2k+l , (k) (1) (k)
= [8 , ... ,8 ,8 , ... ,8 ,8b ,8b , ... ,8b ] a a xl xm_l
is correct, where, by definition, is the linear
functionalon c(k)[a,b] given by the rule
and 8 is short for 8(0) a a
Appendix to Splines and Histograms
2. TENSOR PRODUCTS OF LINEAR SPACES OF FUNCTIONS.
The tensor product of two (or more) algebraic
structures is a weIl understood construct of Algebra.
But since we only need a few notions concerning the
tensor product of two linear spaces of functions, we
give a short discussion of this special case in order
to spare the reader an excursion into an abstract
algebra text.
Let F be alinear space of functions, all
defined on some set X into the reals, and let G
be, similarly, a linear space of functions defined on
some set Y into R. For each f E Fand each
g E G , the rule
h(x,y) = f(x)g(y), all (x,y) E X x Y
defines a function on X x Y , called the tensor
produat of f with g and denoted by
f ® g •
Further, the set of all finite linear combinations of
functions on X x Y of the form f ® g for some
f E Fand some g E G is called the tensor product
of F with G and is denoted by F ® G. Thus,
339
340 C. de Boor
n F ® G = { La. f. ® g. la. E: R, f. E: F, g. E: G,
i=l ~ ~ ~ ~ ~ ~
i = 1, .•. , n; some n}
and F ® G is a linear space.
A simple and important example is provided by
polynomials in two variables. Taking F Ph ' the
linear space of polynomials of degree < h , as
functions on X = R and similarly G = P as k
functions on Y = R we easily recognize F ® G as
the linear space Ph,k of all polynomials in two
variables of degree < h in the first and of degree
< k in the second variable, considered as functions
on the plane 2 xxy=R.
A second simple example arises with the choice
Rm F = , the linear space of real m-vectors considered
as functions on X = {1,2, •.. ,m} , and, similarly,
Rn G considered as a linear space of functions on
y {l,2, ... ,n} In this case, F ® G is the linear
space of all m x n matrices, considered as functions
on X x Y = {(i,j) li = l, ... ,m; j = l, ... ,n}
Appendix to Sp1ines and Histograms 341
One verifies that the tensor product is biZinear,
i.e., the map
F x G + F ® G : (f,g) I + f ® g
is linear in each argument:
In particu1ar,
F ® G = {E.f. ® g. If. E F, g. E G, i ~ ~ ~ ~ ~
1, ••• ,n; some n}
which saves a 1itt1e writing.
Let now A and ~ be linear functionals on
Fand G, respective1y. One defines A ® ~ by the
ru1e
= E.(Afi)(~g·) , ~ ~
all E.f. ® gi. ~ ~
C1ear1y, if A ® ~ is a map on F ® G satisfying
(2.1), then A ® ~ is a linear functiona1 on F ® G •
But, (2.1) requires some discussion before we can
accept it as defining a map on F ® G. For, (2.1)
342 C. de Boor
makes use of the particu1ar form of Lifi ® gi ' i.e.,
the particu1ar f i 's and gi 's, to define A ® ~ on
Lifi ® gi. On the other hand, an element Z E F ® G
may be written in many different ways. 1f, e.g.,
Z = f ® g and f = f 1 + f 2 ,and g = 3g1 ' then
we can write Z as
f ® g or f 1 ® g + f 2 ® g or even as
Corresponding1y, the ru1e (2.1) wou1d give
(among others) for "the" va1ue of (A ® ~)z •
The doubts just raised can be dispe11ed as
fo11ows. 1f Z is any function X x Y , and y is
a particu1ar point in Y , then
Z (x) = z(x,y), all x E X Y
defines a function z on X, the y-seation of z • y
Appendix to Splines and Histograms
If, in particular,
Z = E.f. ® g. , 1. 1. 1.
for some f 's i
343
in Fand
gi 's in G,
then, by the definition of f. ® g. , we can compute 1. 1.
Z (x) as y
i.e.,
Z = Eig. (y)f. . y 1. 1.
This shows that Z E F , hence allows us to compute y
the number
(2.2)
Let now Z
function on
AZ ,and to compute it as y
AZ = E.g.(y)(Af.) . Y 1. 1. 1.
be the A-section of Z , Y defined by
ZA(y) AZ for all y E Y Y
i. e. , the
.
The notation is correct, ZA depends only on Z
A (and not on the particular f. 's and g. ' s) , 1. 1.
and
but
344 c. de Boor
can be computed by (2.2) as
z = E.(Af.)g. whenever z = E.f. ® g. 11111 1
This shows that ZA E: G , hence a110ws us to compute the
number ~ZA' and to compute it as
~Z~ = ~(E.(Af.)g.) = E.(Af.)(~g.) ~ 111 111
thus showing that E. (H . ) (~g. ) 111
depends only on A,~
and the function Z = E.f. ® g .• 111
We conc1ude that, for every linear functiona1
A on Fand every linear functiona1 ~ on G, (2.1)
defines a linear functiona1 on F ® G , and that this
functiona1 satisfies
(2.3) (A ® ~)z = A(Z~) = ~(ZA)' for all Z E: F ~ G •
, Here, Z is the ~-se()tion of Z ,
~
the resu1t of applying ~ to z(x,y)
of y for each fixed x E: X .
To give a simple examp1e, let F
and let
i. e. , Z (x) is ~
as a function
Appendix to Splines and Histograms
for some a and ß, and some integers rand
s < k , so that, e.g.,
Ai = f(r) (a) •
345
Then F ~ G is contained in C(k,k)(R2) , the space
of bivariate functions with k continuous derivatives
in each variable. Further (on F ~ G), A ~ ~ agrees
with the linear functional
v = ü(r,s) a,ß
since, for every f, g E c(k)(R) ,
v (f ~ g) f(x) g(y) I = x=a y=ß
With ~ changed to
y ~g = J g(y)dy
ß
f(r) (a)g(s) (ß)
= (H) Ülg) •
A ® ~ agrees with the linear functional v given by
the rule
346 C. de Boor
Y r r AZ = f (a lax )z(a,y)dy , 11 C(k,k)
a z E •
ß
THOEREM A2. Suppose that the GY'amian A = (A i 4>i) foY'
the sequenae 4>1, ••• ,4>m in F and the sequenae
A1, ••• ,Am of Zinear funationaZs on F is inveY'tibZe,
so that the LIP given by
is aOY'Y'eat. SimiZarZy, a8sume that B = (~i$j) is
inveY'tibZe, with $1, •. ·,$n E G and ~1'···'~n
Zinea:l' funationaZs on G, and set
FinaZZy, a8sume that (vij ) is a matY'ix (oY' doubZe
sequenae) of Zinear funationaZs on some ZineaY' spaae
H aontaining F ® G so that
Vijf ® g = (Aif)(~jg) , foY' aZZ i,j; atz
f ® g E F ® G .
Appendix to Spiines and Histograms
(ii) the LIP on H given by Fi ® Gi and
[(vij)ij] is aorreat, and
347
(iii) for given h EH, the interpoZant Rh aan
be aomputed as
(2.4) Rh = L. jr(i,j)~. x ~J' ~, ~
with
r = r = A-iL (BT)-i h h
where
Lh(i,j) = v .. h aU i,j . ~J
Remark. Here and beiow, we write D(i,j) rather than
Dij or dij for the (i,j)-th entry of a matrix D.
PROOF OF THE THEOREM. If h E Fi ® Gi ' then
(2.5) h = L . . r(i,j)~i ® ~. ~,J J
for some matrix r. But then
348 c. de Boor
In.(r,s) = Ar ~ ~ h = E. jf(i,j)(A $i)(~ $.) s 1, r s J
= (AfBT)(r,s) , all r,s
or
Since both A and Bare invertible by assunption,
and since ~ does not depend on the particular
representation (2.5) for h but only on h, this
implies the uniqueness of the expansion (2.5) for h,
therefore showing (i).
It follows further that, for a given matrix L
and a given h E F1 ~ GI ' we have ~ = L if and
only if the coefficient matrix f for H (with
respect to the basis ($i ~ $j)i,j of Fl ~ GI)
satisfies
proving (ii) and (iii);
The significance of (iii) for computations is
Appendix to Splines and Histograms 349
clear. Instead of having to solve the linear system
(2.6) L • • (A 4>.)(11 ljJj)r(i,j) = L (r,s) , ~,J r ~ s K
all r,s
of m x n equations in order to compute the coeffi
cient matrix r for given h, one only needs to
solve two systems, of size m and n, respectively,
involving, respectively, n and m right sides.
More explicitly, having obtained by Gaussian
elimination a triangular factorization A A = A :t u
for A and B respectively in about
operations, one then computes
as
and
in about 2 2 O(n m + m n) operations. Straightforward
application of Gauss elimination would take
3 O«m x n» operations instead. The savings are
even more significant if (as in the applications below)
A and B are band matrices, a fact difficult to
exploit in a direct attack on (2.6) whatever the
actual ordering of r into a vector might be.
We close this section with aremark concerning
350 C. de Boor
the A.-sections and the llj-SectionS of the interpo-1.
lant Rh defined in the theorem. By (2.3) ,
A. (Rh) = (A. (1g 1l.)Rh A. (1g ll.h 1. ll. 1. J 1. J
J
while, as argued earlier, (Rh) E: FI since llj
Rh E: FI (1g GI . Hence, the ll.-section (Rh) J llj
Rh is the unique element s of FI for which
A.S 1.
i=l, ... ,m.
of
In words: The ll.-section of the interpolant to h is J
the interpolant to the ll.-section of h. J
This
establishes the
COROLLARY. Let h E: F I (1g GI . Then
h = 0 if and onZy if A. (1g ll.h = 0 for i 1, ... ,m. llj 1. 1.
SimiZarZy,
hA. = 0 if and onZy if A. ® ll.h 0 for j 1, ... ,n. 1. J 1.
Appendix to Sp1ines and Histograms 351
3. THE BIVARIATE CASE.
We consider the c1ass
C(k,k)(R)
of functions defined on the rectangle
R = [a,b] x [a' ,b']
and k times continuous1y differentiab1e in each
variable, i.e., having each of the (k + 1)2 partial
derivatives
continuous on R.
With ~ = (x.)m and 1. 0
i,j = 0, .•. ,k
n ~' = (y.) partitions
J 0
for [a,b] and [a',b'], respecitve1y, we define
as the c1ass of sp1ines of degree k (in each
variable) having the 1ines x = xi' i = 1, ••• ,m-1 ,
and j = 1, •.• ,n-1 , as simple mesh 1ines.
35Z c. de Boor
This means that Sk,k (7f x 7f') ~ C(k-1,k-1)(R) , and
that the restriction of each s e Sk k(7f x 7f') to , each of the mn rectang1e
is a po1ynomia1 of degree < k in each variable.
Even degree area matching sp1ine interpolation
to f e c(k-1)[a,b] by element of SZk(7f) (as
described in the coro11ary to Lemma Al) invo1ves
the Zk + m interpolation conditions
by
(3.1) A. f = ~
Xi
J f(x)dx l xi_1
f(i-m-1)(b) ,
m+k (A i )l-k given
i = 1, ... ,m
i > m
The corresponding scheme for SZk(7f') invo1ves
agreement at the Zk + n linear functiona1s
Appendix to Splines and Histograms
lll_k' •.• , II m+k given by the rules
f (-j) (a') j < 0 -
(3.2) lljf = jYj
f(y)dy j = 1, ... ,n y. 1 J-
f(j-n-l) (b ') j > n
The tensor product of these two schemes in the
spirit of Theorem A2 would then associate with each
f E C(k-l) [a,b] x C(k-l)[a',b'] the one element
s E S2k(TI) ® S2k(TI') for which
(A. ® 1l.)S ~ J
(A. ® ll.)f, i ~ J
l-k, •.. ,m+k
j l-k, ••. ,n+k
But this result is unsatisfactory for the reason
that we would prefer to interpolate by elements of
S 2k 2k (TI x TI') . ,
To overcome this objection, we prove that, for
all k,
353
354 C. de Boor
(3.3)
a special case of which is more or less the content
of Schoenberg's Lemma 4. For the proof, observe that
hence
= (k + m)(k + n)
so that (3.3) holds provided we can show that
(3.4) dim Sk, k ('If x 'If') ~ (k + m) (k + n) •
This we show by induction on k: For k = 0,
Sk k('If x 'If') consists of all functions on R which , are constant on each of the mn rectangles Rij ,
hence has dimension mn. Assuming (3.4) to hold for
given k = h , we make use of the fact that then
f E Sh+l,h+l('If X 'If') is absolutely continuous, hence
satisfies
Appendix to Splines and Histograms 355
x y f(x,y) = f(a,a') + f gl(r)dr + f g2(s)ds
a a'
x x + f f g(r,s)dsdr
a a'
with
g(x,y) = fxy(x,y) E Sh,h(~ x ~') •
Therefore
< 1 + h + m + h + n + (h+m)(h+n)
= (h+l+m)(h+l+n)
showing that (3.4) holds then for k = h + 1 , too.
This proves (3.3) and establishes, with Theorem A2,
the following
LEMMA A4. POP every f E C(k-l) [a,b] ® C(k-l) [a',b'] ,
thepe exists exaatZy one s E S2k,2k(~ x ~') top
whiah
356 C. de Boor
i = 1-k, ••• ,m+k; j = 1-k, ••• ,n+k •
We will denote this interpo1ating s by Rf •
Next, we define, in ana10gy to (6.1), (6.2), the
linear space
as the subspace of S ( .... x .... ') 2k,2k" " consisting of
those f for which
(3.5)
for (x,y) € 3R, j = O, ••• ,k-1 •
In the terms of the preceding section and of the
particu1ar Ai 's and ~j 's defined in (3.1) and (3.2)
* S2k,2k(7f X 7f') consists of those
for which
Appendix to Splines and Histograms
and
f = 0, for i = l-k, ••. ,O, m+l, ..• ,m+k Ai
f = 0, for j = l-k, ••. ,O, n+l, ••. ,n+k • llj
357
By the corollary to Theorem A2, we therefore conclude
* that S2k,2k(TI x TI') consists of those
fES 2k 2k (TI x TI') for which ,
Ai ~ lli f = 0, for all (i,j) i {l, ••• ,m} x {l, .•• ,n}.
On combining this statement with Lemma A4, we
obtain the following general version of the first
parts of Schoenberg's Theorems 3 and 4.
THOEREM A3. For a given reaZ m x n matrix
H = (hij ) ~ there exists exactZy one
f V(H) E S2k,2k(TI x TI') which satisfies (3.5)
and Yj xi
=! ! f(x,y)dxdy = hij , y. 1 x. 1 J- 1.-
i = 1, ... ,m; j l, ... , n
358 C. de Boor
This vo Zume matching sp Une V (H) can be computed
as
m n V(H) = I ly··cj>i 01j1 •
i=1 j=1 1J J
with
Here, A= C\cj>j) , B = (lliljlj) , with the cj>i 's
defined in (1.1), and the ljIi 's defined, correspon-
dingly, by
PROOF. It suffices to prove that
But this is clear since, certainly, the left hand
side contains the right hand side, while, on the
other hand, the dimension of the two spaces agree.
The existence and uniqueness of the interpolating
* V(H) implies, by Lemma A2, that dim S2k,2k(TI x TI')
* * ( , = mn, while mn = dim S2k(TI) x dim S2k TI )
REAL ANALYTIC FUNCTIONS AS RATlOS OF ABSOLUTELY
MONOTONIC FUNCTIONS
E.G. Straus
359
The problem we consider here is under what condi
tions analytic functions which are positive on a segment
of the real axis can be expressed as ratios of two
absolutely monotonie functions, that is, functions all
of whose derivatives are non-negative on the given
segment.
The motivation for this question comes from the use
of generating functions of the form
where A is a given set of (non-negative) integers and
m is a set of (non-negative) multiplicities. It is a
clear that whatever functional equation we may get for
360 E.G. Straus
f A we can use the hoped for combinatorial interpretation
only if the Taylor coefficients turn out to be non-nega
tive integers for all a E A. Let me illustrate this
with an example from a problem raised and solved by
Leo Maser:
PROBLEM: Divide the natural numbeps N = {0,1,2, ••• }
into two disjoint sets A3 B so that the sets of sums
of two distinct elements of A is the same as the set
of sums of two distinct elements of B (counting multi
p Uai ties) .
The solution by Lambek and Maser [2] is brief and
elegant:
Set and assume
without loss of generality that 0 E A so that f A (0) = 1, 00
f (0) 0 N b h th i fA + fB -- \ xn -- (l_x)-l B = • ow y ypo es s l n=o
and
(1)
=
Real Analytic Functions as Ratios 361
If we write G(x) = fA(x) - fB(x) then (1) yields
or, by iteration
(3) 2 4 G(x) = (I-x) (l-x ) (l-x )
__ ~ q2(n) n L (-1) x
2n (l-x )
where q2(n) is the sum of the digits of n written to
the base 2. Thus we get the unique solution
1 1 fA(x) = 2 (G(x) + I-x) =
even
a x
so that A consists of those numbers with even sums of
digits to the base 2 ,and B of those with odd sums
of digits.
To see the intimate connection with the problem of
this talk, let us consider the Moser problem where A u B
is not necessarily the set of all natural numbers, A n B
is not necessarily empty and the elements may have multi
plicity greater than 1. However for the sake of
362
analytic simplicity we assume m = O(eEa), a
E.G. Straus
Eb ~ = O(e )
for all a E A, bEB, E > O. If we set F = f A + f B
the equation (2) becomes
(2') G(x)F(x) = G(x2)
and (3) becomes (if we assume 0 E A, o '- B, m = 1) o
(3' ) 1 G(x) = ------=-----2n 2 F(x)F(x ) F(x ) ..•
which constitutes a representation of G(x) on the
interval [0,1) as the ratio of two absolutely monotonic
functions (the numerator being rather simple). Of course
the combinatorial problem is by no means settled by (3')
since m > 0, n ~ 0 will hold only provided a - 0
(4) n = 0,1,2, ....
We can sum this up as folIows.
THEOREM. The Lambek-Moser problem of dividing a given
set C of natural numbers with m = 1, m = O(eEC ) o c
for all c E C, E > 0 , into two sets A, B so that
the sums of two distinctly labelled elements are the
same sets with the same muUipUcities for A and B
Real Analytic Functions as Ratios 363
has at most one soZution given by
co n fA(x) = ~ (fc(x) + TTfc(x2 )-1)
n=o
co n f B (x) = ~ (fc(x) - 1T f c (x2 )-1)
n=o
co n If G(x) = TI f C(x2 ) -1 violates (4) then the Lambek-
n=o
Moser problem has no solution. In particular, if fC(x)
has a zero inside the unit disk then G(x) has poles in
the unit disk and the Lambek-Moser problem has no
solution. On the other hand, whenever
(5)
where Yi , ~i are positive natural numbers we get
and condition (4) is satisfied with A consisting of
the sums of even numbers of Y 's; B the sums of odd
numbers of y 's; and C of all finite sums of y 'so
364 E.G. Straus
There are also known examp1es [1] where some of the
exponents ~i in (5) are negative and (4) remains
satisfied. It wou1d be interesting to know whether all
solutions of the Lambek-Moser problem are of this form.
The answer is yes for finite sets [1] as we sha11 now
show.
My eoworkers (A. Fraenkel, B. Gordon, J. Se1fridge
[5], [1]) and I have examined the Lambek-Moser problem
for finite sets. Here equation (2') gives us a good
deal of information sinee F, Gare polynomials and
F(O) = 1. We see immediate1y that whenever G(S) = 0
then G(S2) = G(S)F(S) ~ O. Sinee S ~ 0 this is
possib1e on1y if S is a root of unity and henee both
G(x) and F(x) are eye1otomie polynomials. If we write
G(x) k Cli.e. ß.
= TT (1 - x ) / 1T (1 - x J) i=l j=l
we get
(6) F(x) = G(x2)/G(x) k Cl..e. ß.
= TI (1 + x ~) / TT (1 + x J) i=l j=l
as the ratio of two abso1ute1y monotonie funetions. One
entertaining eonsequenee of (6) is that whenever two
distinet sets A, B of order lAI = IBI = n have the
Real Analytie Funetions as Ratios 365
same sums of pairs of elements we get 2n = F(l) k-l 2 •
In other words the Lambek-Moser problem ean have
solutions in finite sets only if the number of elements,
n , is apower of 2. Equation (5) shows that all sueh
values of n are indeed possible. The finite Lambek
Moser problem has thus been " redueed" to that of deter
mining those eyelotomie polynomials (6) whieh themselves
are absolutely monotonie, that is have only non-negative
eoeffieients.
It is this problem whieh led T.S. Motzkin and me to
rediseover and reprove a theorem of G. Polya [4].
THEOREM (Polya). If P(x) = P(xl, ••• ,xn) is a poZyno
miaZ such that p(x) _> 0 whenevep x. > 0, i = l, ... ,n , ~-
then thepe exist poZynomiaZs Q(x), R(x) with positive
coefficients so that P(x) = Q(x)/R(x) .
Fortunately Motzkin and I were not aware of this
theorem and as a result obtained more preeise quanti
tative results. For polynomials in one variable with
P(x) > 0 for x > 0 it suffiees to eonsider the
irredueible real faetors whieh are either x + r, r > 0
or 2 2
x - (2r eos 8) + r with in whieh
eases the faetors are already absolutely monotonie; or
366 E.G. Straus
finally 2 2
Q(x) = x - (2r cos 8) + r ,
in which case we have
TI o < 8 < 2' r > 0
THEOREM (Motzkin-Straus [3]). Let n be the integer so
that (n-l)8 < TI < n8. Then a positive polynomial
divisible by Q(x) must be of degree d > n. If
n8 = TI then
xn + r n Q (x) = ----::--~--"'-----
n-2 L (rn- 2- ksin(n_l_k)8)xk
k=o
1 sin 8
If n8 > TI then for each m3 0 < m < n we have
where
p (x) m
R (x) m
Q(x) p (x)
m R (x)
m
1 (sin n-2 + sin 28rxn- 3 + sin 8 8X ... n-2 sin n8 n-m (sin 8xm-2 + + sin(n-l)8r ) sin 8 sin m8 r
+ sin 28rxm- 3 + m-2 ... + sin(m-l)8r )
Real Analytic Functions as Ratios 367
Moreover if cos a ~ m/ (m+2) then m Q(x)(x+r) has
positive coefficients so that every polynomial which is
positive on the positive axis can even be expressed as
a positive polynomial divided by a totally positive
polynomial, that is a polynomial whose zeros are
positive.
This raises a number of questions
1. For every polynomial F(x)
e:cists a minimal p01.Vep nF of
of the foPm (6) there
(X+l) so that nF
F(x) (X+l) has non-negative aoeffiaients. If deg F = n what aan we say about nF ?
We can answer this in part.
LEMMA: Let ~2m (x) be the irreduaible ayalotomia
polynomial of order 2m (degree ~(m») then
2 n~ ~ 3m /4 •
2m
Thus for F = ~2 ~2 ••• ~2 we get ml m2 m,e.
3 2 222 nF ~ '4 (mI + .•• + m,e.) ~ cn (log log n) • On the other
hand if F(x) = x2n _ xn + 1 = x3n - 1 xn + 1
then
2 nF > log 2 n(n-l). Thus the upper bound for nF seems
368
to be of the order of magnitude of
faetor of magnitude 2 (log log n) •
2 n
E.G. Straus
up to a possible
It would be possible
to eonvert these estimates into rough estimates of the
number of pairs of sets A, B of natural numbers with
max e < n whieh have the same sums two at a time. eEAuB
11. To what extent ean the resuZts of PoZya-Motzkin
Straus be extended to generaZ power series?
The results do not extend to all funetions, f,
analytie on an interval [a,b] of the real axis with
f(x) > 0 for a < x < b. This follows from a result of
S. Bernstein to the effeet that an absolutely monotonie
funetion whieh is analytie on [a,b] is in faet analytie
in the disk lx-al< b - a. Thus, if we eould write
f = g/h where g and h are absolutely monotonie then
f must be meromorphie in the disk lx-al< b - a. Sinee
a funetion, f, meromorphie in, say, lxi < 1 with
f(x) > 0 for 0 < x < 1 ean be expressed as the ratio
of two funetions g(x)/h(x) holomorphie in the unit disk
with g(x) > 0, h(x) > 0 for 0 < x < 1 it suffiees to
eonsider the following question.
111. Let f(x) be hoZomorphie in lxi < l. Can we
express f as the ratio g/h of two absoZuteZy mono
tonie funetions g, h whieh are hoZomorphie in
Real Ana1ytic Functions as Ratios
lxi< I? Can we insist that, in addition, h have
onZy negative zeros in lxi< 1 ?
The answer is yes to both questions if f has
a finite number of zeros in lxi < 1 and hence yes
we look for the representation g/h on1y in a disk
lxi < r < 1 •
369
on1y
if
In its comp1ete genera1ity this question as weIl as
its genera1izations to ana1ytic functions of severa1
variables remain unanswered.
REFERENCES
1. Fraenkel, A.S. and Gordon, B. and Straus, E.G.: On the determination of sets by sets of sums of a certain order, Pacific J. Math. 12 (1962), 187-196.
2. Lambek, J. and Moser, L.: On some two way c1assifications of integers, Can. Math. Bu11. 2 (1959), 85-89.
3. Motzkin, T.S. and Straus, E.G.: Divisors of polynomials and power series with positive coefficients, Pacific J. Math. 29 (1969), 641-652.
4. Po1ya, G.: Über positive Darstellung von Polynomen, Vierte1jahvsschrift Zürich 73 (1928), 141-145.
370 E.G. Straus
5. Se1fridge, J.L. and Straus, E.G.: On the determination of numbers by their sums of a fixed order, Pacific J. Math. 8 (1958), 847-856.
INVERSE THEOREMS FOR APPROXIMATION BY POSITIVE
LINEAR OPERATORS
Ron DeVore
371
We are interested in studying the relation
between the smoothness of a function and its degree of
approximation by means of a sequence (L ) n
of positive
linear operators defined on aspace of continuous
* functions C[a,b] or C [-n,n]. Our main interest
is in what inferences can be made about the smoothness
of a function f when we assume something about the
rate of decrease of 11 f - L (f)11 . Such a result is n
customarily called an inverse theorem of approximation
while a result which estimates 11 f - L (f)11 in terms n
of the smoothness of f is called a direct theorem.
Direct theorems are relatively easy to obtain and
are known for the classical examples. On the other
hand, inverse theorems are much more difficult to prove
372 R. DeVore
and indeed may not even hold. The eustomary way of
proving inverse theorems is to use the ideas used by
s. Bernstein in his proof of the inverse theorems for
approximation by trigonometrie polynomials.
Bernstein's teehnique relies on knowing estimates
for suitable derivatives of L (f) (the analogue cr n
Bernstein's inequality). This preeludes the handling
of general sequenees (L ) n
sinee L (f) need not even n
be differentiable in the general ease.
In this work, we replaee the eonditions on the
derivatives of L (f) by suitable eonditions on the n
eoneentration of "mass" of L n
For an example,
suppose (L ) n
is a sequenee of positive eonvolution
operators, i. e.
L (f,x) n
I 1T = - f f(X+t)d~ (t)
1T n -1T
with d~ a non-negative, even Borel measure on n
with unit masse Let
<1>2 1T
t2d~ (t) = f n n -1T
[-1T ,1T]
Inverse Theorems 373
then we can show
THEOREM. Let 0 < a < 2. If there is an M > 0 with
(1)
and
(2)
~n < M < +00, n = 1,2, •..
~n+1 -
TI
f -TI
then * f E Lip a if and only if "f - L (f)" = O(~a) • n n
The assumption (2) is the restrietion on the con
centration of mass. The assumption (1) is a1ways
needed for general inverse theorems and it essentia11y
guarantees that the sequence (L ) n
is not to sparse,
i.e. there are sufficient1y many L n
As an examp1e of this theorem, let t '" 0 and n
Each L can be written as convo1ution with the n
measure d~ which is pure1y atomic with masses n
at each of the points -t and t n n
The theorem
TI
2
374
shows that if
(3)
then
(4)
imp1ies
(5)
t _n_ < M < -t<x> t n+1 -
1I1I 2 (f , x) 11 t n
n = 1,2, •••
O(ta ) n
where 2 lI t (f,x) = f(x+t) + f(x-t) - 2f(x) •
R. DeVore
It can
also be shown that (3) is a necessary condition for (4)
to imp1y (5) in the sense that if (3) does not hold
then there is a function f which s~tisfies (4) but
not (5).
We can also use our technique to prove inverse
theorems for operators that are not given by con-
vo1ution. In this case,
and (2) is rep1aced by
~2 is replaced by n
Inverse Theorems
4 L «t-x) ,x) n
0(4) \x» n
This gives, for example, the inverse theorems for
Bernstein polynomials which were given by H. Berens
and G.G. Lorentz. Namely, a necessary and sufficient
* condition for f to be in Lip a is that
If(x) - L (f,x) I < M {x(1-x)}a/2 n n
for some constant M > 0 •
375
377
LACUNARY INTERPOLATION BY SPLINES
A. Meir & A. Sharma
1. In 1955, J. Sur~nyi and P. Turan commenced the
study of what they called (0,2) interpolation. By
(0,2) interpolation we mean the problem of finding the
algebraic polynomial of degree ~ 2n-l , if it exists,
whose values and second derivatives are prescribed on
n given nodes.
From the above article of Suranyi and Turan and
from the subsequent articles of Balazs and Turan it is
transparent that the problem of explicit construction
of the (0,2) interpolatory polynomials is a difficult
one. Moreover, their methods apply only to special sets
of nodes, which do not include the equidistant case.
More recently Schoenberg has initiated the study
of the so-called g-splines in connection with the pro
blem of lacunary interpolation by splines. He showed
378 A. Meir & A. Sharma
that under certain conditions the interpo1atory g
sp1ines exist and are unique.
In this paper we obtain error bounds for some
c1asses of quintic sp1ines which interpo1ate to (0,2)
data on equidistant knots. Natura11y, such quintic
sp1ines are deficient sp1ines. According to the demands
of continuity and the end conditions required, we obtain
different c1asses of quintic sp1ines. Our method of
proof leads to an a1gorithm for the numerica1 evaluation
of the interpo1atory sp1ines on the basis of the given
data.
2. DEFICIENT QUINTIC SPLINE INTERPOLATION.
For n = 2,3,4 ... we sha11 denote by the
c1ass of quintic sp1ines S(x) on [0,1] having the
fo11owing two properties:
(i) 3 S(x) E C [0,1]
(ii) S(x) is a quintic in [~v+1] ° 1 1 n' n ' v = " ••• ,n- .
It is c1ear that S(35) is a 2n + 4 n,
linear subspace of C3 [0,1] and if
dimensional
S(x) E S(35) , then n,
Lacunary Interpolation by Sp1ines 379
S(x) n-1 v 4 v S
= q(x) + L {c (x - -) + d (x - -) } v=l v n + v n +
where q(x) is a quintic and cv ' dv are constants.
THEOREM 1. Por every odd integer n and for every
given set of 2n + 4 reaZ numbers
{f f f 'f" f" f"·f'" f"'} '1'···" '1'···" , o non 0 n
there exists a unique S(x) € S(3) such that n,S
(1) v
S(-) = f n
v S" (-) = n
(2)
(3) S'" (0) =
f" v
f'" o
v
v
= O,l, ... ,n
= O,l, .•• ,n
S"'(l) = f'" n
THEOREM 2. Let f E C4 [0,11 and n
Then for the unique quintic spZine
an odd integer.
S (x) satisfying n
v (1), (2) and (3) with f v = f(n)'
v f" = f" (-) v n '
v = 0,1, ... ,n; f'" = f'" (0), o
f'" = f'" (1) , we have n
(4)lls(r)-f(r)11 < 7Snr - 3w (!.) + 8nr-41If(4)~ r = 0,1,2,3 n 00 - 4 n 00'
380
where w (.) 4
A. Meir & A. Sharma
denotes the moduZus oi continuity oi f(4).
THEOREM 3. Let f E C4rO,I] and n any positive
integer., h = n-l Then there exists a unique quintic
spZine S*(x) E S(3) n n,S such that
(1' )
(2')
(3' )
Moreover.,
* S (vh) = f(vh) , n
*" S (vh) = f"(vh) , v = O,l, .•• ,n n
*' S (0) = f' (0) , n
*'" S (0) = f"'(O) • n
r = 0,1,2,3 •
Remark: In general there is no unique quintic periodic
spline S (x) E S(3) which satisfies (1) and (2), even n n,S
if the data {f} and {f"} are periodic. v v
This paper will appear in full in the S.I.A.M.
Journal of Numerical Analysis.
STABILITY PROPERTIES OF TRIGONOMETRIe INTERPOLATION
OPERATORS
P.D. Morris & E.W. Cheney
381
Consider the spaee C of all 2n-periodie
eontinuous real funetions, and the subspaee TI of all
n-th order trigonometrie polynomials. The index n is
held fixed, and the spaees are endowed with the usual
supremum norm. Any operator L : C + TI whieh ean be
m written in the form Lx = L x(sk)Yk with ° 2 sk < 2n
1
and Yk E TI is said to be carried by the point set
If Lx = x for all x E TI , then L
is a projection of C onto TI.
defined to be the set of points
for k = O, ••• ,2n .
The uniform grid is
-1 t k = kn(2n + 1)
THEOREM 1. Let S be a set of 2n + 2 points con-
382 P.D. Morris & E.W. Cheney
taining the uniform grid. Among all the projections
from C onto TI carried by s, the interpolating
projection carried by the uniform grid has least norm.
THEOREM 2. Let S be a set of points containing the
uniform grid and containing at least one point between
each two points of the uniform grad. Then the inter
polating projection on the uniform grid is not minimal
in norm among the projections carried by s.
THEOREM 3. There exists a set S of 2n + 4 points
containing the uniform grid with the property that the
interpolating projection carried by the uniform grid
is not minimal among the projections carried by s.
CHEBYSHEV SEMI-DISCRETE APPROXIMATIONS FOR LINEAR
PARABOLIC PROBLEMS
Richard S. Varga
Consider the approximate solution of the linear
system of ordinary differential equations
J d!!.( t) A!!.(t) + r Vt > 0 , = -dt (1)
1 !!.(O) ~
= u
383
where A is a given n x n Hermitian and positive
definite matrix, and where rand u are given
n-vectors. Such linear systems arise, for examp1e, in
the so-ca11ed semi-discrete numerica1 approximations to
linear heat-conduction problems in which the spatia1
variables are suitab1y differenced, but the time
variable is 1eft continuous. The sQ1ution of (1) is
c1ear1y given by
384 R.S. Varga
(2) ~(t) -1 1
= A ~ + exp(-tA){~ - A- r} Vt ~ 0
In contrast with the usual Pade methods, based on
Pade rational approximations of -x e in the neighbor-
hood of x = 0 , we consider here approximations of
w(t) of (2) based on Chebyshev rational approximations
of -x e on [0,+00) , defined as folIows. If 1f
m
denotes all real polynomials of degree at most m
and 1f analogously denotes all real rational m,n
functions r (x) = p(x)/q(x) m,n
define
(3) A m,n
with q E 1f n
r E 1f } m,n m,n
for all nonnegative integers with 0 < m < n , and let
f = ß /4 E 1f be such that m,n m,n m,n m,n
(4) A m,n
Then, the Chebyshev semi-discrete approximation,
w (t), of the solution ~(t) of (1) is defined by -nl,n
Vt > 0 .
Chebyshev Semi-Discrete Approximation
Using i. e. ,
shown that
n I Iv1 12 , it 1s then
i=1
and, as (6) is valid for all t ~ 0 , the Chebyshev
sem1-d1screte approximation w (t) 1Il,n
can be regarded
385
as a one-step approximation of ~(t) for any t > 0 .
The accuracy of the Chebyshev semi-discrete method
depends, from (6), on A , and it is obvious from m,n
(3) that
(7) o < A < A < n,n - n-l,n-< A
o,n
Moreover, it is known that the A 's have goemetric o,n
convergence to zero, i.e.,
(8) 11m (A )l/n = 1 o,n 3 n--)<lO
Finally, it is natural to ask which entire func
tions fex) (having nonnegative coefficients in its
Taylor's expansion about x = 0) possess the property
that, if
386 R.S. Varga
(9) A - inf {ll f (l) - r (x)II L [0 co] : r € 7f } m,n x m,n co' m,n m,n
then there exists a q > 1 such that
(10) 1im (A )l/n < 1 < 1 • o,n - q
n~
Sufficient conditions are given for this, and a re
stricted converse is also estab1ished.