Advances in Computational Mathematics 4( 1995)389-396 389

Piecewise polynomial, positive definite and compactly

supported radial functions of minimal degree
Holger Wendland
hzstitut fftr Numerische und Angewandte Mathematik, Universiti~t GSttingen, Lotzestrafle I6-18,
D-37083 GSttingen, Germany

Received 4 May 1995

Communicated by J.C. Mason

We construct a new class of positive definite and compactly supported radial functions
which consist of a univariate polynomial within their support. For given smoothness and
space dimension it is proved that they are of minimal degree and unique up to a constant
factor. Finally, we establish connections between already known functions of this kind.
Keywords: B-splines, positive definite functions, multivariate interpolation.
A M S subject classification: 41A05, 41A15, 41A30, 41A63, 65D07, 65D10.

1. I n t r o d u c t i o n

In translation-invariant spaces (cf. [12]) interpolants of the form

N

j=l

are natural tools to solve interpolation problems like Sf(X[) = yj, I <_j < N, with
pairwise distinct points xj E ~d. In many cases it is useful to assume • : lt~a ~ R
is a radial function if(x) = qS([[xl]) with a univariate function 4' :~>0 ~ ~ and
the Euclidean norm 11. H, and to assume ff to be positive definite, which means
that for all data sets {x~,..., xu} C_ R a of pairwise distinct points the interpolation
matrix
A =

is positive definite. A continuous function q~ : II~>0 ---' I~ is said to be positive definite

on IRd, q~ E PDd, if the induced function {(x) :=~b(tt x [[), x E R a, is positive definite.
This method of multivariate interpolation using radial basis functions yields
good numerical results (cf. [5]) and was studied extensively in recent years. Some
further applications of this theory are the construction of neural networks (cf.
[2]) and the construction and modelling of geometric objects (cf. [10]). If ff is one
© J.C. Baltzer AG, Science Publishers
390 H. Wendland / Radial functions of rain#hal degree

of the usual basis functions like multiquadrics or thin plate splines still some
problems exist, caused by a large number N of centers xj. Special methods of
computation (cf. [4]) and evaluation (cf. [9]) were developed to remedy this
defect. These problems could be avoided if the radial basis function ~ is compactly
supported. If • has compact support the interpolation matrices are sparse and for
the evaluation of the interpolants sf only few terms have to be considered. This
leads to efficient algorithms for the computation and the evaluation of the inter-
polants (cf. [10]). Another advantage of compactly supported basis functions is
the principle of locality which is well known from the classical B-splines. The
change of one center x i causes only a local change of sf.
In this paper we investigate functions ¢;(x) = ¢([[ x [[) of the following form:

¢(r) = { 0 ( r ) 0<r<I,
,.>1, (1)
with a univariate polynomial p(r) = Y~.~=l cj rj, CN 7£ O. We call N the degree of ¢5
or ¢. Note that due to Bochner a function of this kind is positive definite, if and
only if the d-variate Fourier transform

~(x) ~ 5ca¢(r) = (2rr) -a/2 ~ a ¢5(c°)e"r" dw

= r -(a-2)/2
/0 ¢(t)ta/2J(a_2)/2(rt )dt

is nonnegative and positive at least on an open subset. Here Jr denotes the Bessel
function of the first kind. Note that the operator 5ca acts in this way on univariate
functions. Up to now, only few positive definite functions of the form (1) are
known. These are for odd space dimension d = 2n + 1:
1. Euclid's Hat X t2"+~/which is generated by d-variate convolution of the char-
acteristic function of the unit ball with itself, and which is in PD2,,+j N C O
with a degree OX (2"+~) = 2n + 1
2. Wu'sfunctionsCk+,,,,, C PD2.+~ N C a with a degree 0¢k+,,,. = 2[d/2j + 4 k + 1.
Here, Ix] stands for the integer n satisfying n _< x < n + 1. In the first case it is
used that the d-variate convolution of radial functions is again radial. We denote
this with

(¢,a ¢)(llx II) = ¢(llyll)¢(llx yll) '/3'-

Both of these function types are handled in section 4, and a connection between
them is given.
Another known function is the truncated power function
= (1 -,.)e+, (2)
which is in PD a if g _> [d/2J + 1 (cf. [1,3,7]). Starting with this function we
construct in the following section a class of new functions Ce,k- In section 3 we
 H. Wendland / Radial functions of minimal degree 391

show that these functions are in PDd f-I C 2kand of degree O~be,k =~d/2] + 3k + 1. The
degree turns out to be minimal under the requirement q~ E C ~', and the minimal-
degree solution is unique up to a constant factor. Further, it will become obvious
why we always talk about even degrees of smoothness.

2. C o n s t r u c t i o n

W e start with some notation. F o r a E IR\{-1}, v E R and g E N we define

bracket-operators
I
[~]_, := [a]0:= 1, [~l~:=~(~- 1)..- ( ~ - e + 1), ~_>e-1
a+l'
and
(~)0 := 1, (~)e :=-(~ + 1 ) . . . (~ + e - 1).
Then we have simple relations such as [c~]e= ( a - e + 1)e, [c~]e[c~-g+ 1]. =
(a - g + 1)[o~]e+,,_l and (u)e(u + g),, = (U)e+,,. After this we introduce the operators

I(f)(r) :=
F sf(s) ds, (3)

I
D(f)(r) := - r d f ( r ) ' (4)

for all f u n c t i o n s f for which the terms on the right-hand side are well defined. The
operators I, D, .Ta and *a are connected in the following way.

Lemma 2.1
(1) If ~b E C 2k then Dklkq5 = IkDkqb =- qS.
(2) IfI~b E Ll(R d) then ~-a(I~b) = )rd+2(4~).
(3) If D~b E L, (R d) then Jrd(DC~) = .Td_2(fb).
(4) If I~b, I~b E L~ (N d) then I(~b *d+2 ~b) = (Iq~) *d (I~b).
(5) If Dq~, D~/' E L~ (R d) then D(q5 *d-2 ~b) = (D~) *d (D~).
A p r o o f of these facts can be found in [13].
Note that assertions (2) and (3) give rise to construction methods of compactly
supported positive definite functions via integration or differentiation. Actually,
Wu used statement (3) in [15] to construct such functions. He started with very
smooth, positive definite functions in R ~ and gained less s m o o t h functions that
are positive definite in higher dimensional spaces. Here, we take the opposite
direction. We start with the function in (2) and construct
~be.k := IkCe, (5)
or, more general, %,k := Ik¢~ with u > 0 and ¢~(r) = (1 - r)+. Therefore, we need
first of all a lemma about the incomplete Beta function with one natural argument,
which can be proved by simple induction.
392 H. Wendland / Radial Jimctions of minimal degree

Lemma 2.2
For k E N and u E R>o we define

Then we obtain the identity

Mk,~(r) :=
f sk~b~(s)ds.

k [kle k-t
Mk.~(r) = Z (u + 1)e+, r ~b~+e+,(r).
g=0

Using this lemma we get the following representation of ~b,.k, which can again be
proved by induction.

Theorem 2.3
We have the representation
k
~b,,k(r) = Z &k"" ' " (6)

The coefficients satisfy the recursion

fl0.0 = 1,

k fl,,.k [n + 1],,_j+1 ,
0<j_<k+l,
n=j-- I (u + 2k - n + 1),,_j+2

if the term for n = - 1 for j = 0 is ignored.

If we now choose u = g E N we get

Lemma 2.4
1. The function ~be,k has support in [0, 1], and consists there o f a polynomial o f
degree O~bt.k = g + 2k.
2. It possesses 2k continuous derivatives around zero.
3. It possesses k + g - 1 continuous derivatives around 1.

Proof
The proof can be given either directly by theorem 2.3 or by means of induction,
taking into account that a function of the form (1) is in C 2k around zero iff the
first k odd coefficients o f p vanish. []

3. Main results

In this section we shall prove our main assertions by choosing g for ~br.k
appropriately. The shortness of the proofs demonstrates the strength of the
operators I and D.
 H. Wendland / Radial /unctions of minimal degree 393

Theorem 3.5
For every space dimension d E N and every k E N there exists a function ¢ of the
form (1) with degree 0¢ = [d/2J + 3k + 1 and ¢ c PDa M C 2k.

Proof
We choose g = [(d + 2k)/2j + I = [d/2j + k + 1. Then we get ~e E PDa+2k N C O
by the remarks in the introduction. Using lemma 2.1 leads us to Ce,k E PDa. By
lemma 2.4 we achieve that Ce,k possesses 2k smooth derivatives around zero and
2k + [d/2J smooth derivatives around 1. This completes the proof. []

Before we prove the optimality and the uniqueness o f these functions, we wish
to emphasize some additional features. If we take space dimension d = 1 our C 2-
function ¢2,1 has degree 0¢2,1 = 4 which is just one degree higher than the degree
of the classical cubic B-splines.
The d-variate Fourier transform is computable by means of lemma 2.1 as

a~a Ce, k ( r) = .~a l k Ce( r) = .Td+ 2k Ce( r )

= r - d - ~ - e l r ( r - t)et(d/2)+kJ(d/2)+k_ l ( t ) d t ,
JO
which is, except for ¢~,0, strictly positive (see Gasper [6]), so that the techniques of
Narcowich and Ward [8] and Schaback [11] can be used to obtain bounds for the
condition of the interpolation matrix. The existence of zeros of the Fourier
transforms was a very unpleasant feature o f Wu's functions and the Euclidean Hat.
Figure 1 shows the function ~b3,1(r )--" (l -- r)4+(1 + 4r) E P D 3 f') C 3. Table I
contains a table of some o f the new functions. Here, - means equality up to a
constant.

Theorem 3.6
There exists no function ¢ of the form (1) with C EPDaMC ~k and
0¢ < [d/2J + 3k + 1.

L5

1,5

Figure t, '~3.).
394 H. Wend~and~Radialfunctions of m#1hnal degree

Table 1
Table of functions.

d= 1 ~*.o=(1-r)+ CO
~,, --"(1 - r)~+(3~ + 1) C2
~3.-~ "-" (1 - ,')5+(8,'2 + 5 r + 1) C4

d : 3 ~'2.o = (1 - r)2+ CO
~3,, =" (1 - r)4+(4r + 1) C2
@4._, - (1 - r)6+(35r 2 + 18r + 3) Ca
tbs.3 =" (1 - r)8+(32r 3 + 25r z + 8r + 1) C6

d=5 %.o = (I - ,.)3+ Co

~4., - (1 - ,')~(Sr + l) c-'
~95.2 - (1 - r)7(161 ~ + 7 r + 1) C4

Proof
Suppose there exists such a function qS. On account of the C2k-smoothness of q~we
can write p(r) = q(r 2) + r2k+lh(r) with polynomials h, q, Oq = k. But then Dkp is
again a polynomial because D operates on q(r 2) as d/dr on q. So ~ := Dkq~ is a
function of the form (1) with ~b E PDa+2k A C O and 0~b < [(d + 2k)/2j + 1. This
is a contradiction to the following lemma. []

Lemma 3.7
If q~ is a continuous, on IRa positive definite function of the form (1) then ~b has
degree greater than or equal to kd/2J + 1.

Proof
A p r o o f of this was given by Chanysheva [3] and uses the existence of [d/2J s m o o t h
derivatives, which again follows in odd dimensional spaces easily from the integral
representation for positive definite radial functions. In even dimensional spaces the
special form of q~ has to be considered. []

Note that there is no difference between the optimal function for d = 2n

and d = 2n + 1. This optimal function is in the continuous case unique if it is
normalized by q~(0) = 1, namely ~b(r) = (1 - r) td/2j+l . This generalizes to

Theorem 3.8
U p to a constant factor there exists exactly one positive definite function
• (x) = (llxll) on ~ ' with 4~ of the form (1), i.e.
fp(r) 0<r<l,
0 r>t,
consisting of a univariate polynomial p of degree Op = &b = kd/2] + 3k + 1 and
satisfying 4~ c C 2k.
 H. Wendland / Radial functions of minimal degree 395

Proof
Assume there exist two such functions ¢, ¢ with ¢ # c~b, c # 0. Then we must also
have ¢ : = D k ¢ # c D k ¢ = : ¢ because of ¢ = I k $ and ¢ = I k ~ by lemma 2.1. But
this contradicts the just mentioned uniqueness in the continuous case. []

4. C o n n e c t i o n b e t w e e n W u ' s f u n c t i o n s a n d E u c l i d ' s H a t

In the introduction we mentioned that in the odd dimensional case two

additional instances o f positive definite functions of the form (1) exist. We will
now show that these functions are connected by the same trick we used to
construct our optimal functions. Let us suppose in this section that d = 2g + 1 is
odd. As both function types are constructed via convolution, let us suppose for
simplicity now that their support is {x : 0 _< [[x H <- 2}.
The Euclidean Hat X (~+0 is constructed by 2g + 1-variate convolution of the
characteristic function X o f the unit ball with itself. Explicit formulas are given
in [14] and the proof is sketched in [10]. These functions are of the form (1)
and of degree OX 12e+° = 2g + 1. Wu constructed his functions by starting with
fe(r) = (1 -r2)+,e Ce = f e *fe with univariate convolution and ended with Ce,k =
DkCe, 0 < k < g. This led him to functions Ce.k C PDzk+l ffl C 2e-zk.
If we m a r k equality up to a constant again with - we get

Theorem 4.9
Wu's functions Ce,k and Euclid's Hat X (2e+l) are connected by

Cg,k -- I e-kX(2~+~).

Proof
We use lemma 2. t to get

Ce,k = Dk(fe *l re)

= (Dkfe) *2k+l (Dkfe)

= (Ie-kDtfe) *2k+, (Ie-kDefe)

= le-k((Deft),2e+t (Defe))
__
• ie-kX(2e+l),

if we take into account that Deft(r) = 2eg!x(r). []

Note that this leads to explicit formulas for Wu's functions, but it is obvious that
these formulas cannot be as simple as the formulas given in theorem 2.3.
396 H. Wendtand / Radial functions of minhnal degree

Acknowledgement

T h e a u t h o r w o u l d like to t h a n k Prof. R. S c h a b a c k , w h o helped to bring this

p a p e r into shape.

References

[1] R. Askey, Radial characteristic functions, MRC Technical Sum: Report no. 1262, University of
Wisconsin (1973).
[2] D.S. Broomhead and D. Lowe, Multivariable functional interpolation and adaptive networks,
Complex Syst. 2 (1988) 321-355.
[3] A.Y. Chanysheva, Positive definite functions of a special form, Vestnik Moskovskogo
Universiteta Matematika UDC 517.5; 519.2 (1989). English translation in: Moscow University
Math. Bull. 45 (1990) 57-59.
[4] N. Dyn, D. Levin and S. Rippa, Numerical procedures for surface fitting of scattered data by
radial functions, SIAM J. Sci. Stat. Comp. 7 (1986) 639-659.
[5] R. Franke, Scattered data interpolation: Tests of some methods, Math. Comp. 38 (1982)
181-200.
[6] G. Gasper, Positive integrals of Bessel functions, SIAM J. Math. Anal. 6 (1975) 868-881.
[7] C.A. Micchelli, Interpolation of scattered data: Distance matrices and conditionally positive
definite functions, Constr. Approx. 2 (1986) 11-22.
[8] F.J. Narcowich and J.D. Ward, Norm estimates for the inverse of a general class of scattered-
data radial-function interpolation matrices, J. Approx. Theory 69 (1992) 84-109.
[9] M.J.D. Powell, Truncated Laurent expansions for the fast evaluation of thin plate splines,
DAMTP/1992/NA 10, University of Cambridge (1992).
[10] R. Schaback, Creating surfaces from scattered data using radial basis functions, in:
Mathematical Methods in CAGD III, eds. M. Daehlen, T. Lyche and L.L. Schumaker (1994).
[11] R. Schaback, Error estimates and condition numbers for radial basis function interpolation,
Adv. Comp. Math. 3 (1995) 251-264.
[12] R. Schaback, Multivariate interpolation and approximation by translates of a basis function, in:
Approximation Theory VIII, eds. C.K. Chui and L.L. Schumaker (1995).
[13] R. Schaback and Z. Wu, Operators on radial functions, preprint (1994),
[14] H. Wendland, Ein Beitrag zur Interpolation mit radialen Basisfunktionen, Diplomarbeit,
G6ttingen (1994).
[15] Z. Wu, Multivariate compactly supported positive definite radial functions, Adv. Comp. Math. 4
(1995) 283-292.