JOURNAL OF APPROXIMATION ‘THEORY 69, 84-109 (1992)

Norm Estimates for the inverses of a General Class

of Scattered- Data Radial- Function Interpolation Matrices
FRANCIS J. NARCCIWICH AND JOSEPH D. WARD*

Department of Mathematics, Texas A&M University,

College Station, Texas 77843-3368

Communicated by Nira Dyn

Received February i2, 1990; revised February 18, 1991

In this paper we investigate those radial basis functions h associated with func-
tions whose mth derivative (modulo a scalar multiple) is completely monotonic.
Our results apply both to interpolation problems that require polynomial repro-
duction and to those that do not. In the case where polynomial reproduction is not
required and the order M is 0 or 1, we obtain estimates on the norms of inverses of
scattered-data interpolation matrices. These estimates depend only on the minimal-
separation distance for the data and on the dimension of the ambient space, Iw”.
When the order 112 satisfies m 2 2, we show that there exist parameters n,, __.,a,
such that the function h(x) +a, +a, , r2 + + ar rZm-’ gives rise to an invertible
interpolation matrix, and we obtain bounds on the norm of the inverse of this
matrix. For interpolation methods in which one wishes to reproduce polynomials
of totd degree IYI - 1 or less, bounds for the norm of the inverse of the interpolation
matrix are obtained, provided the data contains a s,,,_ r(P) unisolvent subset.
These results apply, in particular, to Duchon’s “thin-plate spline” results. 0 1992
Academic Press, Inc.

I. INTRODUCTION

Over the last few years, there has been considerable progress concerning
the theoretical development of data fitting in two or more dimensions. In
particular, the method of thin-plate splines, as developed by Duchon [2],
and, more recently, the results of Micchelli [9] and Madych and Nelson
[7, 81 concerning radial basis functions are notable examples. In each of
these papers, certain classes of interpolation matrices associated with
scattered data in W were shown to be invertible. These results established
the “well-poisedness” of the scattered data interpolation problem relative to
the families of either the thin-plate splines or certain radial basis functions,
such as the Hardy multiquadrics [6].

* Supported by the National Science Foundation under Grant DMS-8901345.

84
0021.9045192 $3.00
Copyright 0 1992 by Academic Press, Inc.
All rights of reproduction in any form reserved.
 ESTIMATES FOR INTERPOLATION MATRICES ES

More recently, progress has been made in “quantifying” these interpola-

tion methods, in the sense of estimating the norms of the inverses of these
interpolation matrices as well as their condition numbers. For example, in
[I] such estimates were determined for matrices associated with the func-
tion h(x) = /(x\(~. In [lo], a general approach, using Fourier transform
techniques, was developed for making such estimates for interpolation
matrices coming from conditionally negative definite radial (CNDR) func-
tions of order one. Previous to these papers, there were some useful
numerica results on condition numbers obtained by yn, Levin, and
Rippa [3]+
In this paper, we employ the methods introduced in [IlO] to estimate
certain quadratic forms associated with the family of Gaussians, ePr2’, t >
We first use these estimates to get an upper bound on (iA -’ (/ for interpol
tion matrices A determined by CNDR functions that are of order zero or
one and that are generated by completely monotonic functions with cone-
sponding orders. The estimates for jjA-‘]\ are ven at the end of Section II.
In addition to the dependence on the CND interpolation function, the
estimates for (1Avlll depend only on the minimal separation distance for the
data set and the dimension s of the ambient space, (w”. For interpolation
methods employing CNDR functions that are of order ppz> 2 and which are
associated with functions whose mth derivative is completely monotonic,
we have results for two different interpolation probl
These results greatly extend the results given in [ in that the results
obtained in this paper apply to a much broader class of frictions and t
hold for arbitrary dimension.
Let h be a CNDR function generated by a completely monotonic func-
tion of order m 22. For interpolation methods in which one wishes to
reproduce polynomials of total degree m - 1 or less and where the inter-
polating function has the form
N
f cih(x-xj)+ c k,x”, with
J=1 /aI <m

bounds for the norm of the inverse of the iuter~o~ation matrix are
obtained, provided the data contains a E,-,(W) unisolvent subset. These
results are given in Section IV. In particular, they are relevant to Ducho
“thin-plate spline” interpolation [2].
In Section V we deal with the problem in which h is a CN
of order m > 1 and where polynomial reproduction is not
show that there exist parameters a,, .... CI, such that the shifts to the inter-
polation points of the function

h(x)fa,+a,_,r2+ .” -+a,r’“-*
86 NARCOWICH AND WARD

give rise to an invertible interpolation matrix, and we obtain bounds on the

norm of the inverse of this matrix. We also show that if slightly worsened
bounds are tolerated, the parameters may be chosen so that the resulting
interpolation matrix is negative definite. We mention that the parameters
and bounds appear to depend on the details of the distribution of the data
when m 3 2.
In the closing section of our paper we apply our results to certain radial
basis functions and to thin-plate splines. We now turn to a precise discus-
sion of the interpolation problems mentioned above.
Background. Throughout the remainder of this paper, we will be
considering two types of interpolation problems that we now proceed to
describe.
Given a continuous function h: R” -+ R, vectors {xi} ;” in R”, and scalars
fYj>Py one version of the scattered data interpolation problem consists of
finding a functionf such that the system of equations

f txj) =Yj, j= 1, .... N

has a solution of the form

f(x)= ; qh(x-xj). (1.1)

j=l

Equivalently, one wishes to know when the N x N matrix A with entries

A,, = h(xj - xk) is invertible.
In the second version of the scattered data interpolation problem, the
interpolant is required to have the form

s(x) = 2 cjh(x-xj) + c k,x”, (1.2)

j=t lal -=zm

where the constants cj and k, must satisfy

$ Cjh(xi-xj)+ C kaXg=yi, i= 1, .... N (1.2a)

j=l kl <m
N
c cjx; = 0, JctJ<m. (1.2b)
j=t

This second method guarantees polynomial reproduction in case the data

contains a fl, _ i( R”) unisolvent subset.
The following class of functions has played a prominent role in the study
of both scattered-data problems [3-5, 7-121.
 ESTIMATES FOR INTERPOLATION MATRICES 87

EFINITION 1.1. Let h: EP’+ @ be continuous. We say that h is condi”

tionally negative definite of order m tffor every finite set (x,jy of distinct
points in !R” andfor every set of complex scalars (c,)? satisfying

we have C;” Fjckh(xk -xi) ~0. We denote by JV~ the class of these
functions.
In what follows, we will assume that R” is endowed with a. norm // . /I.
define the function V: R” + R+ by v(x) = /Ix/j.

DEFINITION 1.2. We will say that a continuous function F: [S, 00) -+ R is

a conditionally negative definite radial function of order m if Fo v is in ML.
We will denote the set of all suchfunctions by Xhf~. For the special case
m=O, we say that GE~M’~ if F:= -G is ipa %Af;.

Note that if the norm used is 11.II 2s the Hilbert space no

$I < s2 < co, then one has the inclusions &V; 3 gN;+ 3 $ATAf~, The class
&N~ includes those functions F which are continuous on [O, co) and
which (-1) m+‘(d”/do”)F(J-) fl is . completely monotonic on (0, cc ) [ 1
we will denote the class of such I; by .6&V:,,. The inchrsion
LAW-~ 3 Bw,;, is due to Schoenberg [k2] in the m = 0 case and to
Micchelli [9] m the case where m 2 1. The reverse inchrsion is also know
for m = 0 (Schoenberg [12]) and for m = 1 (Micchelh [9])
reverse inclusion is apparently true due to a recent result of
and X. Sun.
When h is conditionally negative definite of er zero, --A is a positive
definite function (in the sense of Bochner), a the matrix -A is non.-
negative definite. It will be invertible if and o if the quadratic form
associated with -A is a positive definite. Moreover, lIA - r/j is precisely t
reciprocal of the minimum of Q over all unit vectors. For the case of
CNDR functions that are of order one, there is again a connection between

what follows all norm symbols will refer to the I, nor

LEMMA 1.3 (K. Ball Cl]). Let (xi);” be distinct points in R” and let
FESKNS be nonnegative and supposethat h(x) = F(‘(iix\l) is a strictly condi-
tionally negative definite function of order 1. Also, let A be the matrix with
entries AiSk= h(x, - xk). If the inequality
88 NARCOWICHANDWARD

is satisfied whenever the complex numbers tj satisfy c,!= 1 tj = 0, then

JIA-lJI d 0-l.

The proof of this result involves elementary matrix theory and will be
omitted. The result itself is important because it is useful for obtaining
estimates on /(A-‘(( in the ti = 1 case and because it suggests the connec-
tion between norm estimates and quadratic form estimates in cases where
m 3 2, a connection that we exploit in obtaining our results.
We close this section with a result that will be of use later.

LEMMA 1.4 [9]. Let k= 1,2, 3, ... . If Crc,p(x,) =0 for all

p~z~-~([W~), then

(4)“~~CiCj (lxi-xjl12%0, (1.3)

1 1

where equality holds in (1.3) if and only if

5 c,p(x,) = 0, P E V&J0 (1.4)

II. BASIC ESTIMATES

In this section we derive an upper estimate on certain quadratic forms.

In addition, we obtain estimates on jIA-‘lj in the m = 0 and m = 1 cases.
The quadratic form estimates will also be used later, in connection with the
various interpolation problems mentioned in Section I. We begin with the
following lemma, which is found in [9]. For completeness, we include a
short proof here.

LEMMA 2.1. Let g: [O, co) + R be continuous on [0, co) and let E> 0 be
arbitrary. If ( - 1)” g’“’ is completely monotonic on (0, co), then on [O, co)
there exists a nonnegative Bore1 measure dy(t) for which

m- 1gW(E)(o- E)i
g(o)= c j!
j=O

+Iff~{e-.‘-(m~1(-1)~,~-&)‘tj)e-&r)4(r).
j=O
 ESTIMATES FOR INTERPOLATION MATRIGES 89

ProojI Since (- 1)” g’“) .is completely monotone, there exists [I$] a
unique nonnegative Bore1 measure &(t) such that

(2.1)

Next, let G be an arbitrary point in (0, KI). Using Taylor’s Theorem wit
remainder about the point (r = E, we have
m- 1 g(i)(E) 1
&T(a) = 1 - (“-E)‘+(m- ___ 1) i Eu g’“‘(o’)(a’ - CT)“- 1 L&J’.
j=. j!
Substituting (2.1) into the integral expression above yields the desired
result.
In what follows, V,, k = 0, 1, 2, .... will be the subspace

Vk := (u E: RN : vj =p(xj), p E Tlk(lRS), j = 1, ...) Iv> (2.Z)

while Vi will be the orthogonal complement of Vk relative to RN.

COROLLARY 2.2. Let FE BJY~~, and let be the associated quadratic

form below. Then, provided 5 = (Sj,’ .... 5,) E Vi- 1, it follows that

j,k=I

where Qr is the quadratic form given by

Q, := 5 <j~ke-IIXI-XkllZ’. (2.4)
j,k= 1

ProoJ Note that F(r) = -g(r’) for some function g for which
(- 1)” g’“” is completely monotonic. Note also that by Lemma 1.4, the
quadratic form involving the polynomial part of F is zero and hence our
claim is validated. 1
We wish to obtain a lower bound on Q restricted to the subspace VA _ 1.
Corollary 2.2 allows us to obtain such a bound by first getting exphcit,
positive, lower bounds on the quadratic form Q, defined in (2.4), then
stituting them into (2.3), and finally integrating. To describe the 1
bound on Q,, we need to introduce some notation. As in [IO], we let 4
half the smallest distance between any two points in our data set
90 NARCOWICHANDWARD

ix1>...>
x,}; we will call q the separation radius. We assume that the data
set comprises N distinct points, so that q > 0. The lower bound on Ql can
now be stated.

THEOREM 2.3. For every [ E RN and every t 3 0, we have that

Q,34(t) Iltll*? where d(t) := Cst--sf2q--se--62q-Zf-‘, (2.5)

where C, and 6 are constants given by
is2
6 := 12 and
i c”:=2s+11-((s+2)/2)~

The proof of this theorem is technical and is deferred to Section III.

Using the lower estimate for Q stated above, we easily obtain the
following lower estimate on (2.

THEOREM 2.4. Let FE&~:,,, let 5= (tji>j”=lE Vi-1, andset

where I(t) appears in Theorem 2.3 and dy(t) is the measure for the com-
pletely monotonic function ( - 1)” g’“‘(t), g(t) being the function -F(&).
Then, provided g’“’ IS
. nonconstant, the following holds:

i, j = 1

Prooj By Theorem 2.3, it follows that for each fixed t,

Substitution of this inequality in (2.3) yields the result. 1

In case m = 0 or m = 1, we can use this theorem to obtain the following
norm-estimates on inverses of the interpolation matrices associated with
radial functions in ~%?,Jlrz.~.

COROLLARY 2.5. Let FE~J& be nonconstant. If 8 is given by (2.6)

with m = 0, then, for any finite subset {x1, .... x,} E KY, with N and s
arbitrary positive integers, the N x N interpolation matrix A, with entries
Ajk = F( j/xi - xkll ), has an inverse that satisfies
/IA-‘11 <8-l.
 ESTIMATES FOR INTERPOLATION MATRICES 9

PronJ F being nonconstant implies that the corresponding completest

monotonic function g(a) := -F(&) is generated by a measure dq that is
supported on some Bore1 set not containing (r = 0.
defining 9 in (2.6) is strictly positive, which in turn
positive definite and that 8 is a lower bound on the smallest eigenvalue of
-A. The norm estimate then follows from elementary matrix theory.

COROLLARY 2.6. Let FE L@?JV~~ = 9&V:. Suppose that r-IF’ = 2g’(rZ)

is non-constant on (0, co) and that F(0) 2 0. If0 is given by (2.6) with m = 1,
then, for any finite subset {x1, .... x,,,} E R”, with N and s arbitrary positive
integers, the N x N interpolation matrix A, with entries A, = F( Ii-u, - xkj/ ),
has an inverse that satisfies

Proof. The fact that r-‘F’ = 2g’(r2) is non-constant guarantees that the
measure dq appearing in (2.1) is supported on a ore1 subset of (0, cc
Consequently, 6 in (2.6) (with nz = 1) is again sitive. Since F(0) >
implies that FE C!.LNF is nonnegative, an application of Lemma 1.3 yields
the result. h

III. A LOWER BOUND FOR THE QUADRATIC FORM (3,

The quadratic form Qt defined in (2.4) is associated with the Gaussian

function

Ft(r) := e-“‘, (3.1)

where t > 0. Our aim is to find a positive lower bound for

Indeed, we will show that e(t) has the lower bound given in Theorem 2.3.
To find this lower bound, we will use an adaptation of the method intro-
duced in [ 10, Sect. IV].
At first glance, it appears that the techniques developed in [lo] for
estimating lower bounds on quadratic forms such as Q, cannot be directly
employed. The lower estimates there were obtained at the expense of
constraining the components of {E R”; they had to satisfy Cy= 1 <,$==0.
Fortunately, this constraint was used only to get a certain integral
representation for the quadratic form being estimated. For the function
F,(r), this integral representation can be obtained without employing the
92 NARCOWICH AND WARD

constraint. Consequently, the results from [lo] can be used to obtain

lower estimates for Qt.
To obtain the appropriate integral representation for Q,, first write
.10,(r) = ecrzt in terms of its well-known Fourier transform over R” [14,
p. 91. (The variable r is, for this purpose, regarded as the Euclidean norm
in R”.) Once this is done, replace the integral over the angle variables by
the function 52, [14, p. 261,

where Ss-I, w,-~, and do,- r are, respectively, the unit sphere in R”, its
volume, and the usual measure on it. (For future reference, we note that 52,
and o,- 1 have the following explicit formulas [ 14, pp. 26-271:

Q,(z) = T(s/2)(2/z)(“-2)‘2J~,-*),2(z) (3.3)

and

-- 2?zs12 (3.4)
os-1 - T(s/2)’

Here, r( .) denotes the Gamma function and J,( .) denotes the order p
Bessel function of the first kind.) The result of this replacement is the radial
representation
co
e --r% _-
s QAur)
0
4,(u), (3.5)

where dyl(u) = o,_ 1(4~t)-S’2e-u*‘4t~s-1 du. Finally, insert (3.5) into (2.4)
and interchange the finite sum and the integral; this gives the following
representation for Qt.

LEMMA 3.1. The quadratic form Qr given in (2.4) has the form

(3.6)

with dol, given by

da,(u) = co- 1(4~)-“‘2t-“‘2e-“2~4’~“C ’ du. (3.7)

This is the integral representation that we are seeking. The minimization

procedure from [lo, Sect. IV] may be directly applied to it, without the
 ESTIMATES FOR INTERPOLATION MATRICES 93

imposition of any constraint. Also, to avoid confusion, be aware that

representation arises in connection with F, being a positive definite ra
function; it is not directly connected with the integral rc~rese~tatio~s
discussed in Section II.
According to the method introduced in [lo, Sect. V], finding the lower
estimate 0(t), as given by [lo, (4.12)], begins with finding a functio
x: 6%”-+ R having Fourier transform f that satisfies

(i) i>O
(ii) 9 is a radial function and
(iii) &zda,(u)/u22 f(u)u’-r du.

Then, 0(t) satisfies the inequality 0(t) 3 c,(x(Q) - c2) > 0. The precise
constants c1 and c2 are given in (3.15).
There are infinitely many functions that satisfy these criteria. Selecting a
useful x hinges on the convergence of a certain series that depends on x.
(See [IO, Sect. IV, (4.12)].) We will produce a family of such functions,
and select the one that best serves our purposes. To do that, we need to
look at the function $p, the characteristic function of the &P-ball centered
at the origin and having radius /I.
This function is radial and has a radial (inverse) Fourier transform given
bY

t+!+(x) = (27c-” JR,-e-‘<“qg~)n~.

Integrating over the angle variables and using (3.2) yield

Substituting (3.3) and (3.4) into (3.8) and changing the variable of integra-
tion form 24to 2)= /Ix/I 2.4result in

1,6&x)= (2~)-“‘~ llxll -’ j;x” ’ Jc,,2;_l(v)vs’2 dv.

The integral on the right above can be evaluated explicitly ClS, p. 3501; t
result is

(3.9)
94 NARCOWICH AND WARD

Using (3.9) and the series expansion for J, found in [15, p. 3591, one
easily sees that I/~ is analytic in (1x(1#O and that it has a removable
singularity at l/x/l = 0, with

P” (3.10)
h40) = (4?T)“V((s + 2)/2) ’ O.

The function that wil be used for x will be a multiple of I+$. This will
ensure that (i) wil be satisfied, since f will be the convolution of two
positive functions. Let us therefore define the function

x(x) = K+;(x) = K & ’ J:,,Wll Ph (3.11)

( >

We will select /3 later. To determine K, observe that the Convolution

Theorem implies that

f(t) = (27t) -‘K&s * &dt).

Since $ is the characteristic function for the ball of radius /I, center 0, the
convolution product above is nonnegative and has its support contained in
a ball centered at 0 and having radius 2/I. Also, because it is the Fourier
transform of a radial function, K@;(x), it is itself radial. Thus 2 satisfies
two of the three criteria imposed on it. The last of the three,

will determine K. (Writing f(u) is an abuse of notation. It should not cause

any difficulty, though, for f(t) is radial and therefore is constant for
Ili”ll = u = constant. With identical justification, we will also write x(r).)
From the definition of 2 and standard theorems concerning the convolu-
tion, one has that

ii(t) d P~n)-“K Il$,ll;.

The square of the L2-norm of 1,6~is just the volume of the sphere in R” with
radius /I and center 0, and so (3.4) and the last equation yield

j(C)< (2n)-“I‘L-‘~“o,-l

28” K
= s(4n)“*l-(s/2) ’
 ESTIMATES FOR INTERPOLATIQN MATRICES 95

From this inequality, (3.7) and the support of 2 being contained in the ball
with center 0 and radius 26, it is clear that the third criterion will be
satisfied by choosing

(3.12)

Combining (3.11) and (3.12) yields the following result.

LEMMA 3.2. Let fl> 0. A function x: R” -+ R having Fourier tran~fQr~~ f

that satisfies the crieria (i), (ii), and (iii) is given by

x(x) =s(4t~ I/XI/2)~s’2e-P”‘rJ~,2(IlX!I (3.13)

Having found a family of appropriate x’s, we y now proceed with the

next step in the procedure. From [lo, Sect. IVj, s lower bound 6(t) may
be estimated as follows.
First, recall that in Section II we defined the paration radius q of the
data as half the minimum distance between two ta points. Second, jet

K, := sup(Ix(x)l : ng < //XII < (n i l)q). (3.14)

The estimate on O(t), as given in [ 101 and described in the discussion

following (3.7) is then given by

Q(t)>@&(O)-3”Z), where Z= ns-SC,. (3.15)

n=l
From (3.10), (3.11), and (3.12), we have

x(O) = Kl#O) = s (&)‘;’ e-p+ (I-(?

Th.e K,‘S are somewhat harder to estimate in a useful way. To do so

requires an inequality involving Bessel functions; this inequality is stated
below.

LEMMA 3.3. For s = 1, 2, .... and for all z > 0, J&(z) < Yi2Jz7t.
Proof When s = 1, from [13, p. 2971 we have J&(z) d 2./zr~< 2’+ ‘Jzz-r.
When s > 2; Weber’s “crude” inequality 113, p. 211-J for W$(z), the order
s/2 Mankel function, implies that
96 NARCOWICH AND WARD

For z real, JS,*(z) = ‘%i(H$(z)}. Thus from the last inequality we get

with 4z>s- 1. (3.17)

Now, it is easy to show that the maximum of zJf,,(z) must occur for z
greater than the first positive root of J&. One can estimate this root from
below by z/@/2)(.9 -I 4)/2 > (s- 1)/2. (See [13, p. 4871.) This fact and
(3.17) imply that, for all z > 0,
p+2
~J~/~(z)<~(l-(l/2))-~-~=- (3.18)
n ’

which immediately yields the desired inequality. 1

We can use the inequality in the lemma to estimate 1x(x)1. From
Lemma 3.3 and (3.13), we obtain the inequality
4s
I-&)J <-t--“2n.-‘-“* llXll-l-Se-m:
P
and, consequently, that
K, := sup{ Ix(x)1 : nq < llxll G (n + l)q}

Our next goal is to use the estimate for IC, in (3.19) to estimate from
above the sum Z appearing in (3.15). From the expression given for .X in
(3.15) and the inequality in (3.19), it follows that

==Izn”-l ‘8”
n=l
K,<- t-“2n.-1-“q-‘-“e-P~/’
nzl n-2.

Using the well-known formula C,“= 1 n-’ = 71’16, we arrive at

&E (tn)-S/2q-l-Se-8Z/‘.
38
From this inequality, (3.15), and (3.16), it follows that

s(t)~~(X(O)-3’1;)
s 1

1_71~*((3+w @q/12)-“-’ .
18 I
 ESTIMATES FOR INTERPOLATION MATRICES 97

By using (3.4), (3.16), and the identity &I + 1) =pF(p), we may transform
the lower estimate for 0(t) given above into the form

P” -p/t 1 _ nrZ((s + 2)/2)

(3.20)
o(t) > 2VI-( (s + 2)/2) e i 1%

Recall that B > 0 was left unspecified. Let us choose it to satisfy

Solving this equation yields

B:==wq, where 6 := 12

Inserting (3.21) into (3.20) results in our final Lower estimate for Q(t):

(j(t) 3 Cs*--s12q~se~62q-2r-‘, 6”
where C, := (3.22)
2”+‘I-((s -t 2)/Z)’

A5 a corollary to the construction used to obtain the estimate on 6(t), we

have these results.

THEOREM 3.4. Supposethat F&LV~ has the representation

F(r) = jOm‘9 da(u), da(u) = u”+“p(u) du,

where p(u) is strictly positive, decreasing, and continuous OYE(0, CO).

as given in (3.21), one has

2sq”
b-‘Il G pp(26,q)’

ProojI Follow the construction of x up to the point where the constant

M appearing in (3.11) is determined. Note that in this case the argument
used to get (3.12) results in

28
KG Pcm
s(4n)“‘2T(s/2)

so that one may choose

98 NARCOWICHANDWARD

To obtain an estimate from below for the 8 that is the minimum for the
quadratic form Q associated with F, one may now use steps identical to the
ones leading up to the estimate (3.22) to arrive at

from which the desired norm estimate can be obtained as an immediate

consequence of elementary matrix theory. 1
Similar reasoning used in conjunction with Lemma 1.3 and the estimates
from [ 10, Sect. IV] yield a similar result for the m = 1 case:

THEOREM 3.5. Suppose that FE~LV; has the representation

F(r) = F(0) + JF ’ -$(ur) dol(u), dor(z.4)= u”+lp(u) du,

where p(u) is strictly positive, decreasing, and continuous on (0, co) and
F(0) 3 0. With 6 as given in (3.21), one has

IV. INTERPOLATION WITH POLYNOMIAL REPRODUCTION

In this section, we apply the results of the previous section to obtain

invertibility criteria for interpolation matrices arising from functions
FE -@J,“,,, in the case of interpolation with polynomial reproduction.
Consider the interpolation problem as described by (1.2), (1.2a), and
(1.2b). In matrix notation the interpolation problem converts to finding,
for given y E UP, vectors c E RN and k E R”’ such that

Ac+Bk=y
(4-l )
BTc = 0,

where A is an NxN matrix and B is Nxm’ with m’=dim~,P,(R”). We

further assume that the data set {xi};” contains a rc,- r(F) unisolvent set
so that the matrix B has rank m’. In this case, the system (4.1) can be
reduced to the single matrix equation

Gz := AP’z + Pz, G: RN --+ RN,

 ESTIMATES FOR INTERPOLATION MATRICES 99

where P denotes the orthogonal projection onto the space V, ~ 1 defined In

(2.2).
We will now proceed both to show that G is invertible and to ~bt~~~
bounds on jj GP ’ 11.Recall that if U and W are subspaces of WN, then t
angle x E [O, z/2] between them is defined to be that angle for which cos a
is largest and

holds when u E U and w E W. Using compactness and duality, one may

show that there exists u0 E U such that IIuJ = I and such that

sina=sup{/(u,, w’)l : W’E W’and IIw’II = 1). (4.3)

This formula yields the following result:

LEMMA 4.1. Let c( be the angle between U= ange(AP’) and

W= Range P = V. Then, it follows that

where t3 is as in Theorem 2.4 and A = (/API /I.

ProoJ From our assumption, the vectors u0 and w’ appearing in (4.3)

satisfy

u. = AP’z, and w’ = Plz

IMl = 1 and IjPLzll = 1.

Since l(uOll = 1, u,#O and so Plz,#O. From (4.3),

sin ~ ~ I (AP% P%)l.

lIf%ll
In this inequality, we can remove the restriction on the norm of AP’z, by
simply dividing and multiplying by an appropriate factor, Doing this an
also using the self-adjointness of P’ yields

sin cI> I<PLAP*z,> zo>l 11%/12

/
11%/12 llA~iz,ll ll~‘-4 ’
The inequality (4.4) then follows from the last inequality an
Theorem 2.4. 1

64016911.8
100 NARCOWICH AND WARD

The estimate on sin CI in Lemma 4.1 and the bound on the quadratic
form appearing in Theorem 2.4 suffice to prove the invertibility of G and to
estimate IIG-iIJ.

THEOREM 4.2. The matrix G is invertible and satisfies

(4.5)

where 8 is as in Theorem 2.4 and A = 11

API (I.

ProoJ Let U and W be as in Lemma 4.1. In addition, let u = APlz and

w = Pz, where z E RN; clearly, Gz = u + w. From the definition of cx,

IlW123 ll4*+ llwl12-2 ~0sa 1141Ilwll.

Since 2ab < a2 + b2 for all real a, b, we have

IlW23 (1 --OS ~Wl12+ llwl12).

From 1 - cos CI> (l/2) sin’ a and Lemma 4.1, we arrive at

(4.6)

Theorem 2.4 provides us with an estimate on l/ull:

~IUI~~= ((pAPLz(12 + I\pLApLzl12

B (IPLAP’zl12
> e* J(PLzl12.

Using this and w = Pz in (4.6) results in

llW12>~ (0’ lPz(l*+ IIPZII’), (4.7)

which yields both the invertibility of G and (4.5) as immediate conse-

quences. 1

Several remarks are in order:

(1) In the case of thin-plate splines, the invertibility of G was estab-
lished by Duchon [2], who used methods much different from ours. The
invertibility of G in the general case is a result of Madych and Nelson [7]
and Micchelli [9].
 ESTIMATESFOR INTERPOLATION ~AT~~~~S 101

(2) To obtain true polynomial reproduction, we assumed that the

rank of P was the dimension of the appropriate polynomial space. This
requires that the underlying data set have a 7~~~ I(RS) ~nisolv~nt subset. If
the data set fails to have such a subset, then ~oly~ornia~ reprod~~~~~
becomes impossible, because some nonzero polynomials of degree m - 1 or
less will vanish on the data. Equivalently, rank(B) will fall below the
dimension of rc, ~ I( R”).

V. INTERPOLATION wrrH0u-r POLYNOMIAL REPRODUCTION

We now turn our attention to the interpolation that arises from

using FE L&N,“,,,, but that no longer demands po reproducibility.
Thus, we wish to investigate the invertibility of the i~ter~olat~o~ matrix
itself.
For nontrivial FE C%W1”) the interpolation matrix A was shown to be
invertible, provided F # a + br2 and F(O) 2 0 [7, 91. Estimates on IIA ~-! j/
W obtained in [l, lo].
course, not all FE LJLN~ satisfy F(O) > 0. owever, given an F, one
may choose a constant c so that H(r) := F(r) + c (which is still in !2RN~)
satisfies H(0) 20. We will show that something similar happens when
m > 1. When FE&?JY~,,, rn > 1, we will provide two ways to select a
polynomial p(t), with degree m - 1, such that

H(r) := F(r) +p(r2) E BN,“,,,

yields an invertible interpolation matrix. We will also obtain estimates on

inverses of the corresponding interpolation matrices. However, if pn > 1, our
choice of p(r2) does depend on the number of data points.
We begin with two lemmas. In what follows, e(A) wilI denote the
spectrum of A.

LEMMA 5.1. Let A be a seu-adjoint matrix and let P, deenste the

orthogonal projection onto some k dimensional subspace V c RN. If
o(P,APx-)c [E, co), E>O, and o(P:APt)c(--CD, -e], then A has k
positive eigenvalues, N-k negative eigenvahes, and CT(A)c (- CXI~-&I u
[E, co ). In particular,

ProoJ It suffices to show that A has k positive eigenvalues contained in

[E, 00 ), since the same type of argument shows that B := -A has N-k
positive eigenvalues contained in [E, co).
102 NARCOWICH AND WARD

Let e1 denote the maximum eigenvalue of A with associated eigenvector

or. Then

g1 = max (Ax, x) 2 max (Ay, y) = max (PkAPk y, y) 3~.

ILlI= 1 p 71 pII=_
kY-Y kY-Y

For the case k = 1, we are done. Let k 3 2. If cr2(rr2d a,) denotes the next
largest eigenvalue of A, then

Note that W: is an N- 1 dimensional space while the range of Pk is k

dimensional. Provided that N - 1 + k > N, there exists a nontrivial y in the
intersection of these vector spaces so,

CQ= max (Ax, x)2 max (Ay, ~)a&.

&J,~ L 0 Il.!4 = 1
‘$ “:‘;O
k

In this way, one may continue finding positive eigenvalues up to ok. 1

Remarks. (1) In the m = 1 case, where A is the interpolation matrix
associated with FE 9f.N ;” , one can use Lemma 5.1 to recover Ball’s obser-
vation [ 1] that if F(0) 3 0, and if

6:=inf -(Ac,c):~cj=O >O,

{ 1

then I/A-‘/j d l/6. To do this, first note that such an F is nonnegative, and
so the entries of A are nonnegative, too. Consequently,

0
1
. ..l)A i >O.
1

Taking V=Span((l...l)T), with PI being the projection onto V, we

see that a(P, AP,) c [S,, cc). Since by assumption we also have
o(Pf APf ) c (- co, -61, it is clear that A has one positive eigenvalue, A+,
and N - 1 negative eigenvalues. If we label these -A,- 1 < -A,- 2 < . . . <
--A,, then from Lemma5.1, 1,>6. Also, IidA+, for TraceA=i+-
1, - ... - ;1,- r 2 0. Consequently, IIA-‘ll = l/l1 < l/6.
(2) In Lemma 5.1, it is clear that if o(PkAPk) c [s,, co) and
o(P:APi)c(-co, --s-l, with E, #E-, but both positive, then
a(A) c (-co, --E-I u [a, co) and IIA-‘l/ dmax{E;i, .s:r}.
 ESTIMATES FOR INTERPOLATION MATRICES HO3

LEMMA 5.2. Let W be a real vector space with inner product ( , >, and
let U G W be a subspace. In addition, let S and T be self-anoint linear
transformations that satisfy these conditions:
(i) (SW, w> 20, for all wE W.
(ii) (Su,u)>o I)u[/~, cr>O,for all uEU.
(iii) (Su’, u’ ) = 0 and ( Tu’, u’ ) 3 z /lu’/l 2, where z > 0, f~or all u’ E U’.
Then, one may choose y so that, for all w E W, yS+ T satisfies

(b’s+T)w,w>2; llw/12.
Proof Since W= U@ Ul, we can write w E W a5 M;= au + pa’, where
U, u’ are unit vectors in U, U’, respectively, and I/w//’ = cc2+ 8’. If y is any
real number, then

Since S is nonnegative in W,

I(Su, ul>l <J?z-zJm,

and so, by (iii), (Su, U’ ) = 0. Hence,

<W+ T)w w> =~2(~Vu, u> + <Tu, G>

+ 2@( Tu, u’) + p’( Tu’, u’).

Next, by (ii), (iii), and /lull = llu’ll = 1,

~~~~+~~~,~)~~2(y~-ll~ll)-~l~l WI W,0fP2~.
Since JIuII = \Iu’II = 1, 1(Tu, u’)l < II TII. This and the inequality
2ab < ~~~41~+ cb2 imply that

((yS+ T)w, w)>,cr2 +/3’(r-E IIT//).

Choose E= 2/(2 I(TII ) to get

((YS+T)W, w)2cr2 ay- IITIl -___

Finally, pick y = (l/o)[ II Tlj + 22-l I\TII 2 + ~/2] to complete the proof.
104 NARCOWICH AND WARD

We are now ready to prove the main results of this section. Both
theorems we are about to state concern the invertibility of the interpolation
matrix coming from an FEgHz,, that has been modified by adding a
polynomial in r2. One of the theorems is the direct analog of results proved
in the case m = 1, while the other shows that F may be modified so that the
resulting interpolation matri.x will be negative definite.

THEOREM 5.3. Let FERN,,, be such that (d”/dt”)[F(&)] is non-

constant. Then there exist scalars a,, a,, .... a, such that

H(r) :=F(r)+a,+a,-,r*+ ... +a,r*“-* (5.1)

is in .@?.J”& and the interpolation matrix B corresponding to H is invertible,
has y positive eigenvalues, N - u negative eigenvalues, and satisfies
IIB-‘II GO-‘, (5.2)
where % is as in Theorem 2.4 and p = dim V,,_ 1, where V,,, _ 1 is defined in
(2.2).
Proof: Note that HE ~JY:,, since (d”‘ldr”)(H(&) - F(A)) = 0. As
before, let

It is easy to see that V,_, can be decomposed in the following way:

v,-l=(v;-, 0 v;_,,o(v;_, 0 vi-,,o ..’ 0 v,.

The proof proceeds by induction. First, let IV, = V$- z 0 Vi- 1 = U,, so
relative to IV,, U: = (0). Let S, be the matrix of (- l)“- ’ rZm-*, and let
T be the interpolation matrix of F. By Lemma 1.4,

(SIT w> 20, lIwl12, WE w,.

Since U: = {0}, the r1 that appears in Lemma 5.2 is arbitrary, so choose

it to be z = z1 = 2’9. Thus, we can find y1 so that

((YIS, + T,)w, w>>2”-‘0 llwl12, WE w,.

Choose aI = (- l)“-‘7,.
Second, define the following:

W,=(JL 0 v~~,>ocv~-, 0 VA-1,

u,= vi-, 0 vi-,
S, = interpolation matrix of (- l)m-2r2m-4
T,=y,Sl+T,.
 ESTIMATES FOR INTERPOLATION MATRICES 105

Note that relative to W,, U: = W, . Consequently, the inequality above

implies that on Uk, T2 is bounded below by r2 = 2”-‘8. From Lemma 1.4,
it is clear that (S,u& u;) = 0, u; E Vi = W,, and that there is a constant
cr2 such that

<S2%, u2) a Q2 11412~ U2E u,.

Applyig Lemma 5.2 again gives us a constant y2 such that

((y2%+ T2h w> 32m-2@ 11~112, WE w2.

Choose a, = (- 1)“-2y,.
We may proceed in this way until all the coefficients have been deter-
mined. The result is then a function HE %?Jf~,, with the property that its
interpolation matrix B satisfies

<Bw w> 26’ I!wl12, WE v,-1.

On the other hand, if we let A be the interpolation matrix associate
F, then Lemma 1.4 implies

(Bw, w> = (Aw, w>, WE VA-,.

From Theorem 2.4, we then obtain

<BY w> G -0 11~112, WE V;+

Applying Lemma 5.1 then yields the theorem. 1

THEOREM 5.4. Let F’E~JV’~,, be such that (dm/dt”)[F(&)] is non-

coltstunt. Then there exist scalars b,, .... b, such that

K(r):=F(r)+b,+b,-,r2+ ..~ +blr2m-2 (5.3)

is in .S?N~, , and the interpolation matrix C corresponding to K is invertible,

negative definite, and satisfies

(5.4)

where 0 is as in Theorem 2.4.

ProoJ The proof again follows by induction. Start by observing that if

W,=(V$_, 0 vi-,)O v;-1,

106 NARCOWICH AND WARD

then, with U, = Vi-, 0 Vi-i and U: its orthogonal complement in W,

(i.e., 17: = Vi _ 1), the matrices

S, := interpolation matrix of (- l)m-1r2m-2

T, := -A

will satisfy the conditions of Lemma 5.2, provided

z,=e and 0,=iinf{(S,u, V) : 2)~ Vie2 0 Vk_,}.

That this is so is a consequence of Lemma 1.4 and Theorem 2.4. Lemma 5.2
then implies that there is a y1 such that

((Y,S, - TI)W w> 2; IIwl12, WE w,.

Choose b, = -(-l)“-‘y,.
In the next step, we work with

w, := (Vi-3 0 v;-,,o vk-2

S, := interpolation matrix of (- 1)m-2r2m-4
i T2 :=ylSl-A.
After noting that W, = Vi- 2, the argument is virtually identical to that
above. The inequality analogous to the one above is

((~2S2 + 7’2)~ w>2; Ilw11*, WEw,.

Choose b, = (- 1)“-2y2.
Continuing in this way, one may choose the b’s in (5.3). The interpola-
tion matrix C obviously satisfies

8
(Cw, w) d -- lIwl12, W&RN.
2”

Consequently, C is negative definite and invertible, and I(CV1l( satisfies

(5.4). I
Remark. As far as we know, no one has interpolated using the func-
tions H and K constructed above. We would certainly be interested in
numerical tests that used these functions to interpolate scattered data. We
close by pointing out that although norm estimates in B-’ are better than
on C’, in the sense of being smaller, C’ can be computed using steepest
descent methods-because C is negative definite.
 ESTIMATES FOR INTERPOLATION MATRICES 104

VI. THIN-PLATE SPIJNES

Thin-plate splines are radial functions that are in %QVz,, and that mini-
mize certain Sobolev norms. Because these functions have been exte~s~v~l~
employed in interpolation problems, we have chosen them to illustrate our
results.
Let d= 1, 2, 3, .... and let m > d/2, where m is an integer.
tions
(-l)~-Cd’2171G-W
d odd
r(m+l-d/2) ’
M + 1~ d/2 cm ~ d/2 ln ~
d even.
T(m+ 1--d/2) ’

The thin-plate spline associated with the pair m, d is

Fm,&) := -gm,d(r2).

Using induction, one may verify that the m th derivative of gm,d is given by

A standard Laplace transform formula then yields

which of course shows that F,,,E~ZJV~& and implies that the measure
dq(t) in Theorem 2.4 is given by

dr/(t) = td’2- 1 dt.

Using this measure allows us to calculate the quantity 0 that appears in

Theorem 2.4. We have

e=c,q-frnt 0
(d--s)/2---11,-629-2,-1 dt.

If we substitute u = J2qp2t ~ ’ above, the integral that results is standar

Doing it yields
108 NARCOWICH AND WARD

where C, and 6 are given by

d2
and c, := (6.2)
2”+lr((s + 2)/2)’

The first example that we wish to look at is the case in which m = 1 and
d= 1, and the CNDR function is F,,i(r) = 27r”*r. We chose this case
because, when s = 1, the best estimate on /iA -’ I/is known [ 11. If a constant
factor is accounted for, the result from Cl] is that /A-‘ll ~n-l’~q-‘,
which is sharp. To get an estimate using our results, first note that with
m = d = s = 1, we have that 6 = 27~ and that C, = 2rc3’*. From (6.1), it
follows that (3= q/2 &. Applying Corollary 2.6 then yields the estimate

Thus our results yield an estimate that is a factor of 271larger than the best
possible estimate, but that has the same q dependence.
The two most important cases occur when m = 2 and s = d = 2 or s = 3,
d= 1. In these cases, we have the following &values:

when s=d=m=2;
when m=2 and s=3, d=l. (6.3)

We dealt with three interpolation matrices: G defined in Theorem 4.2; B,

defined in Theorem 5.3; and C, defined in Theorem 5.4. For the function
F2,2 = -r* In r2 and F2,1 = (47~‘/~/3)y~, only (6.3) is required to estimate
IIB-‘jl and IIC-‘ll. To estimate IIG-‘ll, we also need

(6.4)

Combining the theorems mentioned above with (6.3) and (6.4) results in
the table below:
IWill IP’II IW’II
-r2 In r2 (808D’NIn D)q-’ max{ 1,57lq-‘} 5719-2 2,284q -’
4 & r’/3 (4.54 x 105q-‘iVD3) max(1, q-3} 1.36 x 105q-’ 5.44 x 105q-’

In the table above, D, which is the diameter of the data set, was assumed
to be larger than 2. Also, although N, D, and q are independent, one can
show (see [lo, Sect. 71) that

N<\ Df&? 2
( 2q > .
 ESTIMATES FOR INTERPOLATION MATRICES PO9

ACKNOWLEDGMENT

We thank the referee for several helpful comments.

REFERENCES

1. K. BALL, Invertibility of Euclidean distance matrices and radial basis interpolation,

preprint.
2. J. Docrrolrl, Sphnes minimizing rotation invariant semi-norms in Sobolev spaces, in “Con-
structive Theory of Functions of Several Variables” (W. Schempp and K. Zeller, Eds.),
pp. 85-100, Springer-Verlag, Berlin, 1977.
3. N. DYN, D. LEVIN, AND S. RIPPA, Numerical procedures for global surface fitting of scat-
tered data by radial functions, SIAM J. Sci. Statist. Comput. 7 (1986), 639-659; N. Dyn,
“Interpolation and Approximation by Radial and Related Functions,” in “Approximation
Theory, VI” (C. K. Chui, L. L. Schumaker, and 4. D. Ward, Eds.), Vol. 1: pp. 211-234,
Academic Press, Boston, 1989.
4. N. DYN, W. A. LIGHT, AND E. W. CKENEY, Interpolation by piece-wise linear radial basis
functions, I, J. Approx. Theory 59 (1989), 202-223.
5. 1. M. GELFAND AND N. YA, Vilenkin, “Generalized Functions,” Vol. 4, Academic Press,
New York, 1964.
6. R. L. HARDY, Multiquadric equations of topography and other irregular surfaces,
J. Geophys. Res. 76 (1971), 1905-1915.
7. W. R. MADYCH AND S. A. NELSON, Multivariate interpolation and conditionally positive
definite functions, Approx. Theory Appl. 4 (1988), 77-89
8. W. R. MADYCH AND S. A. NELSON, Multivariate interpolation and conditionally positive
definite functions, II, Math. Comp. 54 (1990), 211-230.
9. C. A. MICCHELLI, Interpolation of scattered data: Distances, matrices, and conditionally
positive definite functions, Constr. Approx. 2 (1986) 11-22.
10. F. J. NARCOWICH AND J. D. WARD, Norms of inverses and condition numbers for
matrices associated with scattered data, J. Approx. Theory 64 (1991), 69-94.
11. M. J. D. POWELL, Radial basis functions for multivariable approximation, in “Aigorithms
for Approximation” (J. C. Mason and M. 6. Cox, Eds.), Oxford Univ. Press, Oxford,
1987.
12. I. J. SCHOENBERG,Metric spaces and completely monotone functions, Ann. of Mutk
(1938), 811-841.
13. 6. pd. WATSON, “Theory of Bessel Functions,” Cambridge Univ. Press, London, 1966.
14. J. II. WELLS AND R. L. WILLIAMS, “Embeddings and Extensions in Analysis,” Ergebnisse;
Vol. 84, Springer-Verlag, Berlin, 1975.
15. E. T. WHITTAKER AND G. N. WATSON, “A Course of Modern Analysis,” 4th ed.,
Cambridge Univ. Press, London, 1965.
16. D. V. WIDDER, “The Laplace Transform,” Princeton Univ. Fress, Princeton, 1941.