Radial Basis Functions by Buhmann

Acta Numerica (2000), pp.
1–38 c Cambridge University Press, 2000

Radial basis functions

M. D. Buhmann
Mathematical Institute,
Justus Liebig University,
35392 Giessen, Germany
E-mail: Martin.Buhmann@math.uni-giessen.de
Radial basis function methods are modern ways to approximate multivariate

functions, especially in the absence of grid data. They have been known,
tested and analysed for several years now and many positive properties have
been identified. This paper gives a selective but up-to-date survey of several
recent developments that explains their usefulness from the theoretical point
of view and contributes useful new classes of radial basis function. We consider
particularly the new results on convergence rates of interpolation with radial
basis functions, as well as some of the various achievements on approximation
on spheres, and the efficient numerical computation of interpolants for very
large sets of data. Several examples of useful applications are stated at the
end of the paper.
CONTENTS
1 Introduction 1
2 Convergence rates 5
3 Compact support 11
4 Iterative methods for implementation 16
5 Interpolation on spheres 25
6 Applications 28
References 34
1. Introduction
There is a multitude of ways to approximate a function of many variables:
multivariate polynomials, splines, tensor product methods, local methods
and global methods. All of these approaches have many advantages and
some disadvantages, but if the dimensionality of the problem (the number
of variables) is large, which is often the case in many applications from stat-
istics to neural networks, our choice of methods is greatly reduced, unless
2 M. D. Buhmann
we resort solely to tensor product methods. In fact, sometimes we are given

scattered data, immediately excluding the use of tensor product methods.
Also, tensor product methods in high dimensions always require many, of-
ten too many data. In this situation, the method of choice is often a radial
basis function approach which is, incidentally, also highly useful in lower-
dimensional problems and as an alternative to (piecewise) polynomials, be-
cause of its excellent approximation properties; at any rate it is universally
applicable independent of dimension.
In order to formulate the problem as we will see it in this review, let Ξ
be a finite set of distinct points in Rn , which are traditionally called centres
in radial basis function jargon, because our basis functions will be radially
symmetric about these points. The goal of our work is to approximate an
unknown function that is only given at those centres via a set of real numbers
fξ , ξ ∈ Ξ. These are almost always interpreted as function evaluations of
some smooth function f : Rn ⊃ Ω → R, so that fξ = f (ξ). Here, Ω is a
domain in Rn . This point of view will allow us to measure conveniently the
uniform approximation error between f and its approximant s. This error
depends on the choice of the approximant, on Ξ, on f , and in particular on
its smoothness.
In order to approximate with s, which is usually by interpolation, we
take a univariate continuous function φ that is radialized by composition
with the Euclidean norm on Rn , or a suitable replacement thereof when
we are working on a sphere in n-dimensional Euclidean space, for instance.
This φ : R+ → R is the radial basis function. Additionally, we take the
given centres ξ from the given finite set Ξ of distinct points and use them
simultaneously for shifting the radial basis function and as interpolation
(collocation) points. Therefore, our standard radial function approximants
now have the form
X
s(x) = λξ φ(kx − ξk), x ∈ Rn , (1.1)
ξ∈Ξ
suitable adjustments being made when x is not from the whole space, and
the coefficient vector λ = (λξ )ξ∈Ξ is an element of RΞ . In many instances,
particularly those that will interest us in Section 3, the interpolation re-
quirements
s |Ξ = f |Ξ (1.2)
for given data f |Ξ lead to a positive definite interpolation matrix A =
{φ(kξ − ζk)}ξ,ζ∈Ξ . In that case, we call the radial basis function ‘positive
definite’ as well. If it is, the linear system of equations that comes from (1.1)
and (1.2) and uses precisely that matrix A yields a unique coefficient vector
λ ∈ RΞ for the interpolant (1.1).
All radial basis functions of Section 3 have this property of positive defin-
Radial basis functions 3
2 2
iteness, as does for instance the Gaussian radial basis function φ(r) = e−c r
for√all positive parameters c and the inverse multiquadric function φ(r) =
1/ r2 + c2 .
However, in some instances such as the so-called thin-plate spline radial
basis function, the radial function φ is only conditionally positive definite
of some order k on Rn , say, a notion that we shall explain and use in the
subsequent section. Now, in this event, polynomials p(x) ∈ Pk−1
n (x) of degree
k − 1 in n unknowns are augmented to the right-hand side of (1.1) so as to
render the interpolation problem again uniquely solvable. Consequently we
have as approximant
X
s(x) = λξ φ(kx − ξk) + p(x), x ∈ Rn . (1.3)
ξ∈Ξ
Then the extra degrees of freedom are taken up by requiring that the coeffi-
cient vector λ ∈ RΞ is orthogonal to the polynomial space Pk−1n (Ξ), that is,
all polynomials of total degree less than k in n variables restricted to Ξ:
X
RΞ 3 λ ⊥ Pk−1
n (Ξ) ⇐⇒ λξ q(ξ) = 0, ∀ q ∈ Pk−1
n . (1.4)
ξ∈Ξ
In order to retain uniqueness, Ξ has to contain a Pk−1

n -unisolvent subset in
this case. When k = 2, for instance, this means that the centre-set must
not be a subset of a straight line. This is a requirement that can easily be
met in most cases.
The two probably best-known and most often applied radial basis func-
tions are called multiquadrics and thin-plate splines,
√ respectively. The
former is, for a positive parameter c, φ(r) = r2 + c2 and the latter is
φ(r) = r2 log r, where in the second case (1.3) and (1.4) with k = 2 are
applied. The multiquadric is, subject to a sign change, conditionally posit-
ive definite of order k = 1, but it turns out that the original interpolation
problem without augmentation by constants is also nonsingular because of
special properties of the multiquadric function (Micchelli 1986).
One more well-known example comes up when we set c = 0 in the mul-
tiquadric example: then we have the so-called linear radial basis function
φ(r) = r which also gives a nonsingular interpolation problem without aug-
mentation by constants. All these examples are useful for various forms of
interpolation and approximation and they all allow this interpolation pro-
cedure in all dimensions n and all sets of distinct centres, independently of
the geometry of the points. In contrast to spline approximation in more
than one dimension, for instance, there is no triangulation or tessellation of
the data points required, nor is there any restriction on the dimensionality
of the problem. We note, however, that these nonsingularity properties are
strictly linked to the fact that we use Euclidean norms; for p-norms with
4 M. D. Buhmann
p > 2, including p = ∞, or p = 1, singularity can occur in any dimension

for unfortunate choices of data points ξ (Baxter 1991).
While it has been known for a long time (see Hardy (1990) for references
and the history of the approach) that approximants of the above form exist
and approximate well if the centres are sufficiently close together and the
function f is smooth enough, it took quite a while to underpin these empir-
ical results theoretically, that is, get existence, uniqueness and convergence
results, and extend the known classes of useful radial basis functions to fur-
ther examples. Now, however, research into radial basis functions is a very
active and fruitful area and it is timely to stand back and summarize its
new developments in this article. Among the plethora of new papers that
are published every year on radial basis functions we have made out and
selected five major directions which have had important new developments
and which we will review in this work.
One major feature, for example, that has been looked at recently (again)
and that we address, is the accuracy of approximation with radial basis
functions when the centres are scattered points in a domain, and, ultimately,
form a dense subset of the domain – a subject, incidentally, well fitted to
begin this review paper in the next section, because the whole theoretical de-
velopment started some 20 years ago in France with Duchon’s contributions
(1976, 1978, 1979) to exactly this question in a special context.
Apart from the fundamental question of unique solvability of the ensuing
linear system – which has by now been completely answered for large classes
of radial basis functions that are conditionally positive definite, mostly fol-
lowing Micchelli (1986) – an important question is that of convergence and
convergence rates of those interpolants to the function f that is being ap-
proximated by collocation to f |Ξ if f is in a suitable smoothness space.
Duchon (1976, 1978, 1979) gave answers to this question that address the
important special case of thin-plate splines φ(r) = r2 log r and its siblings,
for instance the odd powers of r (‘pseudo-cubics’ φ(r) = r3 , etc.), and recent
research has improved some of his 20 year-old results in various directions,
including inverse theorems, and theorems about optimality of convergence
orders, the sets of centres still always being finite. Several of these relevant
results we address in the next section. We do not comment further, how-
ever, on the nonsingularity results, as they have been reviewed often before
(Powell (1992a), for instance).
Next, some theorems are given on the new classes of radial functions
with compact support that are currently under investigation. They must,
in the author’s opinion, be seen as an alternative to the standard radial
functions of global support, like thin-plate
√ splines or the famous and ex-
2 2
tremely useful multiquadric φ(r) = r + c (Hardy 1990), but not as an
excluding alternative, because the approximation orders they give are much
less impressive than the dimension-dependent orders of the familiar radial
functions. Some possible applications, for instance, for numerical solutions

of partial differential equations come to mind when those radial functions
are used because they can act as finite elements, and they are being tested
at the moment by several colleagues for this very purpose. Results are
mildly encouraging, although the mathematical analysis is still lacking (see,
e.g., Fasshauer (1999)); some mathematical underpinning is given in Franke
and Schaback (1998) and Pollandt (1997). We will explain further and give
examples in Section 6.
In Section 4, several recent results about the efficient implementation of
the radial basis function interpolation are given, especially iterative methods
for the computation of interpolants when the number |Ξ| of centres is very
large. This work is called for when the radial functions are of global support
and increasing with increasing argument, as they often are, because no direct
or simple iterative methods will then work satisfactorily, the matrices being
ill-conditioned, and sometimes highly so, with exponentially increasing con-
dition numbers. This is particularly unfortunate because, in applications,
large numbers of data occur frequently.
Section 5 is devoted to radial basis functions on spheres. Indeed, several
radial basis functions were created to interpolate data given on the Earth’s
surface, for instance potential or temperature data (Hardy 1990). This has
inspired many researchers to consider the question of how to approximate
efficiently when the data are from a sphere and when the whole idea of
distance defining the above approximants is adjusted properly to distances
on spheres, that is, geodesic distances. Several groups are currently working
on these approximations, which need not be interpolants, although the main
focus is, as always with radial basis functions, on interpolation, and we give
a brief review of some important new results.
The final section is devoted to applications, mostly to some initial at-
tempts at the numerical solution of differential equations with radial basis
function methods.
At the end of this introduction we remind the reader that everything said
here is the result of a selection, not comprehensive, and explained, of course,
from the author’s personal point of view. It is certain that several relevant
theorems have been omitted as a consequence. On the other hand, there
will be a fairly comprehensive book by the author that gives more details
and proofs of many results that are merely stated here.
2. Convergence rates
As usual in the study of methods for the approximation of functions, one of
the central themes in the analysis of radial basis functions is their conver-
gence behaviour when the centres become dense in a domain or in the whole
underlying Euclidean space. This is highly relevant because it shows how
6 M. D. Buhmann
well we can approximate smooth functions even in the practical case when
the centres become close together but do not cover a whole domain. Several
of the results were initiated by the work of Duchon and this explains the
title of the following subsection.
2.1. Improvements and extensions to Duchon’s convergence theorems

Many of the theorems about convergence rates are related to the so-called
variational approach and founded on ideas of Duchon (1976, 1978, 1979)
who considered the – even today – important special cases of radial basis
functions of thin-plate spline type in n dimensions, namely

r2k−n log r, if 2k − n is an even integer,
φ(r) = (2.1)
r2k−n , if 2k − n is not an even integer,
where we admit nonintegral k but always demand k > 12 n. The typical case
we always think of is the thin-plate spline in two dimensions, namely n = k =
2 and therefore φ(r) = r2 log r. The aforementioned odd powers – linear or
cubic, for instance – also belong to this class. Several further important cases
such as multiquadrics√or shifted logarithms are in fact derived from the above
by composition with r2 + c2 : namely, φ(r) = r altered in this fashion √ gives
multiquadrics and φ(r) = log r provides the shifted logarithm log r2 + c2
(although in this case 2k = n). The reason for this transformation is to gain
infinite smoothness when c is positive (recall that all of the above are not
smooth at the origin when composed with the Euclidean norm).
In this context, the approximants are usually taken from the ‘native
spaces’ X of distributions (Jones (1982), for instance, for generalized func-
tions or distributions) in n unknowns whose total kth degree partial derivat-
ives are square-integrable: we call the space that depends on the choice of φ
the space X := D−k L2 (Rn ). The Sobolev embedding theorem tells us that
this space consists of continuous functions as long as k > n2 . This is the first
and perhaps most important example of the general notion of native spaces,
namely semi-Hilbert spaces X that ‘belong’ to the radial function and are
defined by all distributions f : Rn → R that render a certain φ-dependent
seminorm kf kφ finite. In the present case, the seminorm is the homogeneous
Sobolev norm of order k. We will return to this concept soon in somewhat
more generality.
A beautiful convergence result that holds on a very general domain Ω
and does not yet require explicitly native space seminorms (although they
are used implicitly in the proof, as we shall see) is the following one of
Powell (1994). So long as all domains as general as those in the statement
of the theorem are admitted, this is the best possible bound achievable for
this general class of domains, that is, the constant C on its right-hand side
cannot, for h → 0, be replaced by o(1). In order to state the result we let

h := sup inf kx − ξk. (2.2)
x∈Ω ξ∈Ξ
This notion is also used in the rest of this section, as is the notation k · k∞,Ω
for the Chebyshev norm restricted to Ω.
Theorem 1. Let φ be from the class (2.1) with k = n = 2 and Ω be
bounded and not contained in a straight line. Let s be the radial basis
function interpolant to f |Ξ satisfying (1.2) for Ξ ⊂ Ω where k keeps the
same meaning as in the introduction, that is, linear polynomials are added
to s and the appropriate side conditions (1.4) demanded. Then there is an
h- and Ξ-independent C such that
p
ks − f k∞,Ω ≤ Ch log(h−1 ), 0 < h < 1.
In fact we can be even more specific about the constant C in the above
error estimate. It is, of course, independent of h and Ξ, but its depend-
ence on f can be expressed by C = c̃kf kφ , where c̃ depends only on Ω and
kf kφ is the homogeneous Sobolev seminorm of order 2 of f , depending on
the aforementioned native space. The general approach to convergence es-
timates of this form is always to bound the error |s(x) − f (x)| by a fixed
constant multiple of
p
kf kφ Φ(α) (2.3)
that depends on the radial basis function (2.1) and the dimension n, where
α = (αξ ) ∈ RΞ are the coefficients of the representation
X
x= αξ ξ, x ∈ Ω,
ξ∈Ξ
and Φ is the so-called power functional

X X X
Φ(α) = φ(0) − 2 αξ φ(kx − ξk) + αξ αζ φ(kξ − ζk).
ξ∈Ξ ξ∈Ξ ζ∈Ξ
Consequently, the main work lies in bounding this power functional from
above. If this is done judiciously we can obtain optimal bounds, because
the bound of the error function by a suitable constant multiple of (2.3) is
best possible (Wu and Schaback 1993).
Bejancu (1997) has generalized this result to arbitrary k and n, and his
theorem includes the above result (see also Matveev (1997)). There are no
further restrictions on Ω except, in general, its Pk−1
n -unisolvency which we
demand for the following theorem.
Theorem 2. Let φ be from the class (2.1). Let Ω be bounded and contain
a Pk−1
n -unisolvent subset. Let s be the radial basis function interpolant
(1.3) to f |Ξ for Ξ ⊂ Ω as in the introduction where k keeps the same
8 M. D. Buhmann
meaning, that is, Pk−1

n polynomials p are added to s with the appropriate
side conditions (1.4) on the λ ∈ RΞ . Then there is an h-independent constant
C such that
 p
√ h log(1/h), if 2k − n = 2,
ks − f k∞,Ω ≤ C 0 < h < 1.
 h, if 2k − n = 1, and
h, in all other cases,
Johnson (1998b) has the following improved convergence orders that do,
however, require Ω to be a domain with Lipschitz continuous boundary and
satisfying an interior cone condition (see any of the cited papers by Duchon
for the standard definition of this concept). Now, (2.1) for general n and k
are admitted and H 2k denotes the usual Sobolev space.
Theorem 3. Let Ω̃ ⊂ Ω be compact and f ∈ H 2k (Ω) be supported in Ω̃.
Then we have for any Ξ ⊂ Ω that contains a Pk−1
n -unisolvent subset and an
h-independent constant C and the interpolant s on Ξ the error estimate in
the p-norm
ks − f kp,Ω ≤ Ch2k+min[n/p−n/2,0] , 0 < h < 1, (2.4)
where 1 ≤ p ≤ ∞.
Note that the best rate occurs in (2.4) when p = 2 but p = ∞ gives the
inferior rate 2k − 12 n.
2.2. Upper bounds on approximation orders and inverse theorems;

saturation orders
It is a remarkable new development to have upper bounds on the obtainable
convergence rate rather than the lower bounds thereon as above. Johnson
(1998a) shows that the rates of Theorem 3 are almost optimal. (As is well
known, the optimal ones in case of Ξ = hZn and p = ∞ are 2k (Buhmann
1990a, 1990b) – we will come back to this soon.) We still let the radial basis
function be from the above class (2.1), although many of the results cited
include multiquadric interpolation, for example.
Theorem 4. Let 1 ≤ p ≤ ∞ and let Ω be the unit ball. Suppose {Ξ =
Ξh ⊂ Ω}h>0 is a sequence of finite sets of centres with distance (2.2) for each.
Then there exists an infinitely smooth f such that, for the best Lp (Ω)-
approximation s to f of the form (1.1) with the appropriate polynomials
added, the error on the left-hand side of (2.4) is not o(h2k+1/p ) as h tends
to zero.
Therefore, in Theorem 3 we are not far from the optimal result. We do
get the optimal results of O(h2k ) for uniform convergence (p = ∞) on grids
Ξ = hZn as mentioned already, but also, as Bejancu (1999) shows us, on
finite grids Ξ = Ω ∩ hZn , where Ω is a cube with sides parallel to the axes.
This extension of the results on Ξ = hZn to Ξ = Ω∩hZn relies on the locality

of the cardinal interpolants on hZn which can be expressed conveniently in
Lagrange form
X X
s(x) = f (ξ)L h−1 (x − ξ) = f (hi)L h−1 x − i , x ∈ Rn .
ξ∈Ξ i∈Zn
In this expression, the Lagrange function L that satisfies L(k) = δ0k for all
multi-integers k, and may be expanded as
X
L(x) = λi φ(kx − ik), x ∈ Rn ,
i∈Zn
enjoys the remarkable property of the fast (exponential) decay of |L(x)|

(Madych and Nelson 1990), which renders the approximant local with re-
spect to the data. This, in turn, gives rise to the same convergence orders
when the domain of approximation is restricted from the whole of the Euc-
lidean space to a cube with sides parallel to the axes.
Theorem 5. Let Ω, Ξ = Ω ∩ hZn and the interpolant s be as above, and

let Ω̃ be a compact subset of the interior of Ω. Then, for f ∈ Lip2k+1 (Ω),
and for an h-independent constant C,
ks − f k∞,Ω̃ ≤ Ch2k , 0 < h < 1.
Precise upper bounds on the approximation order that can be identified for
Ξ = (hZ)n may be stated in a very general context even for h-dependent ra-
dial basis functions φh that are from a sequence of radial functions {φh }h>0 .
Therefore we study approximants from spaces

· n
Sh (φh ) = span φh −j |j ∈Z .
h
A special case is, of course, φh ≡ φ for all h, perhaps taken from one of
our (2.1). For the statement of the next theorem we let Ω be the unit ball
and σ : Rn → R a smooth (C ∞ ) cut-off function that is supported in Ω
and satisfies σ | 1 Ω = 1. We recall that the notation fˇ stands for the inverse
2
Fourier transform
Z
1
fˇ(x) = eix·t f (t) dt,
(2π)n
integrals being over Rn unless stated otherwise.

10 M. D. Buhmann
Theorem 6. Let 1 ≤ p ≤ ∞. Let φˆh : Rn → R be measurable and such

that, for some j0 ∈ Zn \{0} and some ε ∈ (0, 1):
(1) φˆh (x) 6= 0 for almost all x ∈ εΩ;
φˆh (·+2πj0 )
(2) the inverse Fourier transform σ ε· ˇ is absolutely integ-
φˆh
rable;
(3) for a measurable function ρe which is locally bounded and for all pos-
itive m,
φˆh (h · +2πj0 )
2k
h ρe − = o(h2k ), h → 0.
φˆh (h·) ∞,mΩ
Then the Lp -approximation order from Sh (φh ) cannot be more than 2k,
that is, the distance in the Lp -norm between the Lp -closure of Sh (φh ) and
the class of band-limited f whose Fourier transform is infinitely differentiable
cannot be o(h2k ) as h tends to zero.
We note that the class of all band-limited f whose Fourier transform is
infinitely differentiable is a class of very smooth local functions and if the
Lp -approximation order to such smooth functions cannot be more than h2k ,
it cannot be more than h2k to any general nontrivial larger set of less smooth
functions.
A typical example occurs when we again use radial basis functions of the
form (2.1) where φ̂(r) is a constant multiple of r−2k . Then condition (1) of
the above theorem is certainly true for φ̂h ≡ φ̂. Moreover, condition (2) is
true because the smoothness of the cut-off function, the smoothness of the
function φ̂(k · +2πj0 k) in a neighbourhood of the origin for nonzero j0 and
the fact that 2k > n imply by Lemma 2.7 of Buhmann and Micchelli (1992)
that condition (2) holds. Finally, we can take ρ̃ = φ̂−1 and get the required
result from condition (3), namely that O(h2k ) is best possible.
This is the obtainable (saturation) order and an inverse theorem of Scha-
back and Wendland (1998) tells us that all functions for which a better order
is obtainable must be trivial in the sense of polyharmonic functions. For its
statement we recall the standard notation ∆ for the Laplace operator.
Theorem 7. Let Ω be as in Theorem 2, Ξ ⊂ Ω a finite centre-set with
distance h as in (2.2) and φ as in (2.1). If, for any f ∈ C 2k (Ω), and all
compact Ω̃ ⊂ Ω,
ks − f k∞,Ω̃ = o(h2k ), 0 < h < 1,
then ∆k f = 0 on Ω.
A very similar, also inverse, but more general theorem from the same
paper is the following one with which we close this section. In order to
state it we come back to the notion of native spaces and recall that for a
radial basis function with positive distributional Fourier transform φ̂(k · k) :

Rn \ {0} → R, the square of the native space norm is
Z
1 1
2
kf kφ := |fˆ(t)|2 dt (2.5)
(2π)n φ̂(ktk)
and the native space X is the space of all distributions f on Rn for which
(2.5) is finite. In the case of the thin-plate splines or, more generally, (2.1),
X agrees with D−k L2 (Rn ), because φ̂(ktk)−1 is a constant multiple of ktk2k ,
and by the Parseval–Plancherel theorem (Stein and Weiss 1971); this being
a special case, other positive φ̂ are admitted as well. For the statement of
the following theorem we recall that the radial basis function is conditionally
positive definite of order k in Rn if the matrix A with centres from Ξ ⊂ Rn
is nonnegative definite on the subspace of coefficient vectors λ ∈ RΞ that
are orthogonal to Pk−1n (Ξ). The functions (2.1) are all conditionally positive
definite of order k subject to a possibly needed sign change. We use the
notation τ (x) ∼ t(x) if both τ (x)/t(x) and t(x)/τ (x) are uniformly bounded
for the appropriate range of x.
Theorem 8. Let Ω be a bounded domain containing a Pk−1 n -unisolvent
subset. Let φ be conditionally positive definite of order k and satisfy
φ̂(r) ∼ r−2k , r > 0.
If f ∈ C(Ω) and there exists µ > k such that
ks − f k∞,Ω ≤ Chµ , h → 0,
for the radial basis function interpolants s defined on all finite sets of centres
Ξ ⊂ Ω with h given by (2.2), then f is an element of X. (The constant C
depends on f but not on h.)
An example for the application of this result is the radial basis function
(2.1) with k there and in Theorem 8 being the same.
3. Compact support
This section deals with radial basis functions that are compactly supported,
quite in contrast to everything else we have seen before in this article. All of
the radial basis functions that we have considered so far have global support,
and in fact many of them, such as multiquadrics, do not even have isolated
zeros (thin-plate splines do, by contrast). Moreover, the radial basis func-
tions φ(r) are usually increasing with growing argument r → ∞, so that
square-integrability, for example, and especially absolute integrability are
immediately ruled out. In most cases, this poses no severe restrictions; we
can, in particular, interpolate with these functions and get good convergence
rates nonetheless, as we have seen in the previous section. There are, how-
ever, practical applications that demand local support of the basis functions
12 M. D. Buhmann
such as finite element applications (but note one of the approaches to partial
differential equations in the last section which works specifically with glob-
ally supported φ) or applications with very quickly growing data, or where
frequent evaluations with substantial amounts of centres are required, etc.
Therefore there is now a theory of radial basis functions of compact sup-
port where the entire class of φs gives rise to positive definite interpolation
matrices for distinct centres.
We note that, however, in the approximation theory community there
is an ongoing discussion about the usefulness of radial basis functions of
compact support because of their inferior convergence properties (unless we
wish to again forego the advantages of compact support – see our discussion
about scaling in the text below), and in comparison to standard, piecewise
polynomial finite elements. The latter, however, are harder to use in grid-
free environments and n > 3, because they require triangulations to be found
first.
As we shall see in this section, there exist at present essentially two ap-
proaches to constructing univariate, compactly supported φ : R+ → R such
that the interpolation problem is uniquely solvable with a positive definite
collocation matrix A = {φ(kξ − ζk)}ξ,ζ∈Ξ . There will be no restriction on
the geometry of the set Ξ of centres, but there are – in fact, there must be
– bounds on the maximal spatial dimension n which is admitted for each
radial function φ so that positive definiteness is retained. So they are still
useful in grid-free and high-dimensional applications.
In contrast to the radial basis functions of the previous section, the ap-
proximation orders we state, unfortunately, are not nearly as good as the
maximal ones available of the well-known globally supported radial func-
tions such as (2.1). This indeed puts a stricter limit to the usefulness of
compactly supported radial basis functions than we are used to for the glob-
ally supported ones. Be that as it may, it is nonetheless interesting to study
the question of when compactly supported radial functions give nonsingular
and convergent interpolation schemes.
3.1. Wendland’s functions

Initially, there were the approaches by Wu and by Wendland, where the ra-
dial basis functions consist of only one polynomial piece on the unit interval
[0, 1] and are otherwise zero. Although by this means they are piecewise
polynomial seen as a univariate function, the resulting approximants are, of
course, not. The whole idea is based on the earlier work by Askey (1973)
who observed by considering Fourier transforms that the truncated power
function φ(r) = φ0 (r) = (1 − r)`+ gives rise to positive definite interpola-
tion matrices A for ` ≥ [n/2] + 1. Already here we see, incidentally, an
upper bound on the spatial dimension n if we fix the degree of the piecewise
polynomial.
In order to derive a large class of compactly supported radial basis func-
tions starting from Askey’s result, Schaback and Wu (1995) introduced the
two so-called operators on radial functions
1
Df (x) := − f 0 (x), x > 0,
x
and Z ∞
If (x) := rf (r) dr, x > 0,
x
that are defined for suitably differentiable or asymptotically decaying f , re-
spectively, and are inverse to each other. Additive polynomial terms do
not arise on integration since we always restrict ourselves to compactly sup-
ported functions. Next, Wendland (1995) and Wu (1995) use the fact that
the said interpolation matrix A for the truncated power φ0 remains positive
definite if the basis function
φ(r) = φn,k (r) = I k φ0 (r), r ≥ 0, (3.1)
is used when ` = k+[n/2]+1. The way to establish that fact is by considering
the Fourier transform of the n-variate radially symmetric function φ(k · k),
which is also radially symmetric and computed by the univariate Hankel
transform
Z ∞
n/2 1−n/2
φ̂(r) = (2π) r sn/2 Jn/2−1 (rs)φ(s) ds, r = kxk ≥ 0, (3.2)
0
where Jn/2−1 is a Bessel function. This is the radial part of the radially
symmetric Fourier transform of φ(k · k). Hence we have to show that pos-
itivity of the Fourier transform, which is necessary and sufficient for the
positive definiteness of the matrix A by Bochner’s theorem, prevails when
the operator I above is applied, as long as the restrictions on dimension
are observed. This proof is carried out by studying the action of I on the
Hankel transform (3.2) and by direct computation and use of identities of
Bessel functions.
Starting from this, Wendland (1995, 1998), in particular, developed an
entire theory of the radial basis functions of compact support which are
piecewise polynomial and are positive definite. This theory encompasses
recursions for their coefficients when they are expanded in linear combin-
ations of powers and truncated powers of lower order, convergence results,
and minimality of polynomial degree for given dimension and smoothness.
Two results below serve as examples for the whole theory. The first states
the minimality of degree for given smoothness and dimension and the second
is a convergence result.
14 M. D. Buhmann
Theorem 9. The radial function (3.1) gives rise to a positive definite in-
terpolation matrix A with radial basis function φ = φn,k , and with distinct
centres Ξ in Rn . Further, among all such functions for dimension n and
smoothness C 2k , it is of minimal polynomial degree.
Theorem 10. Let φ be defined by (3.1), let f be a function in the Sobolev
space H k+(n+1)/2 (Rn ), and let k be at least 1 when n = 1 or n = 2. Then,
for a compact domain Ω with centres Ξ ⊂ Ω, the interpolant (1.1) satisfies

ks − f k∞,Ω = O hk+1/2 , h → 0,
where h is given by (2.2).

Examples are, for ` ≥ [n/2] + 1,
φ(r) = (1 − r)`+1
+ ((` + 1)r + 1),
φ(r) = (1 − r)`+2 2 2
+ ((` + 4` + 3)r + (3` + 6)r + 3),
in a form proposed by Fasshauer.
3.2. Further contributions

The second class of radial basis functions of compact support (Buhmann
1998, 2000) are reminiscent of the famous thin-plate splines, albeit truncated
in a suitable way, and of a certain convolution form. It contains, for instance,
the following cases. (We state the value of the function only on the unit
interval; elsewhere it is zero. It can, of course, be scaled suitably depending
on the distances between the centres.)
Namely, two examples that give twice and three-times continuously differ-
entiable functions, respectively, in three dimensions and two dimensions are
as follows. We state them first because they are useful to illustrate the goal
of our later result. The parameter choices for the theorem below α = δ = 12 ,
ρ = 1, and λ = 2, give, for n = 3,
7 16 1
φ(r) = 2r4 log r − r4 + r3 − 2r2 + , 0 ≤ r ≤ 1,
2 3 6
while the choices α = 34 , δ = 12 , ρ = 1, and λ = 2, provide, for n = 2,
112 9 16 7 14 1
φ(r) = r 2 + r 2 − 7r4 − r2 + , 0 ≤ r ≤ 1.
45 3 15 9
Another two-dimensional (n = 2) example which is twice times continuously
differentiable is
1 4 1 4
φ(r) = − r2 + r3 + r4 − r3 log r, 0 ≤ r ≤ 1.
18 9 2 3
Theorem 11. Let 0 < δ ≤ 12 , ρ ≥ 1 be reals, and suppose λ 6= 0 and

α > −1 are also real quantities with

 (− 12 , ∞), α ≤ min[ 12 , λ − 12 ], if n = 1, or





 [1, ∞), − 12 < α ≤ 12 λ, if n = 1, and

λ ∈ (− 12 , ∞), α ≤ min[ 12 (λ − 12 ), λ − 12 ], if n = 2, and



 (0, ∞), α ≤ 12 (λ − 1), if n = 3, and



 1 1 1
( 2 (n − 5), ∞), α ≤ 2 (λ − 2 (n − 1)), if n > 3.
Then the radial basis function
Z ∞ λ
φ(r) = 1 − r2 /β β α (1 − β δ )ρ+ dβ, r ≥ 0, (3.3)
0 +
has a positive Fourier transform and therefore, by Bochner’s theorem, gives

rise to positive definite interpolation matrices A with centres Ξ from Rn .
Moreover, φ(k · k) ∈ C 1+d2αe (Rn ).
There is also a convergence estimate for the above radial functions avail-
able which includes scaling of the radial basis function, because, as the
distances between the centres become smaller, we wish to decrease the sup-
port of the radial function as well. Otherwise we would lose the advantages
of compact support since the support relative to the distance between the
centres would increase. Note that, in the statement of the following theorem,
the approximand f is continuous by the Sobolev embedding theorem pre-
cisely as long as 1 + α is positive by the conditions in the previous theorem
(this is therefore a necessary condition).
Theorem 12. Let φ be as in the previous theorem and suppose addition-
ally ρ > 1, 2α ≤ λ − n/2 − 3 + bρc. Let Ξ be a finite set in a compact domain
Ω. Let s be the scaled interpolant (1.1)
X
s(x) = λξ φ(η −1 kx − ξk), x ∈ Rn ,
ξ∈Ξ
to f ∈ L2 (Rn ) ∩ D−n/2−1−α L2 (Rn ), with the interpolation conditions s|Ξ =

f |Ξ satisfied. Then the uniform convergence estimate
kf − sk∞,Ω ≤ Ch1+α η −n/2−1−α
holds for h → 0 and positive bounded η, the positive constant C being
independent of both h and η.
It is interesting to observe that the function classes of the first subsec-
tion of this section can be integrated into the more general class discussed
in this subsection. Namely, when the operator D is applied to our radial
16 M. D. Buhmann
functions (3.3), it gives

Z 1 p
λ λ
D φ(r) = Γ(λ + 1)2 β α−λ (1 − β)ρ dβ, 0 ≤ r ≤ 1,
r2
for δ = 12 and any integral nonnegative λ. Therefore, by one further applic-

ation of the differentiation operator and explicit evaluation,
Dλ+1 φ(r) = Γ(λ + 1)2λ+1 r2α−2λ (1 − r)ρ+ = Γ(λ + 1)2λ+1 (1 − r)ρ+ , r ≥ 0,
for α = λ. Now let λ = k − 1, k ≥ 1,ρ = [ 12 n] + k + 1 and recall that, on the
other hand, we know that the radial functions φn,k of Wendland are such
that
[ 1 n]+k+1
Dk φn,k (r) = (1 − r)+2 , r ≥ 0.
Therefore, the functions generated in this fashion can also be derived from
(3.3), albeit with a parameter α that does not fulfil the conditions of The-
orem 11, so it is a type of continuation. The functions from Wu (1995) also
belong to that class as they were shown by Wendland to be special cases
of his functions. Therefore, since the class of Wendland functions described
here contains those introduced by Wu (1995), our class covers both. Finally,
because we know that all those functions have positive Fourier transforms,
there follow immediately new results about the positivity of Hankel trans-
forms (3.2) which extend the work by Misiewicz and Richards (1994), and
this is in spite of the fact that the parameter α here does not yield the
conditions of Theorem 11.
4. Iterative methods for implementation

Among the most useful radial basis functions which provide good, accurate
approximations are the thin-plate splines, for instance in two dimensions,
and multiquadrics. However, these two, like many other examples of radial
basis functions with global support, lead to linear interpolation systems that
require special considerations if |Ξ| is more than a few hundred when we
calculate the coefficients λξ . Otherwise the computational cost of O(|Ξ|3 )
is prohibitively expensive for direct methods, storage being another major
obstacle if the size of the set of centres is too large – even with the fastest
and biggest workstations currently available.
There are several approaches currently under intense investigation, mainly
in Cambridge and in Christchurch, New Zealand, for dealing with this prob-
lem, two of which we shall explain here to exemplify what can be done today
when we have 50000 centres, say. This is the BGP (Beatson–Goodsell–
Powell) method of local Lagrange functions, another highly relevant and
successful class of methods being the fast multipole approaches (Beatson
and Newsam (1992); see also Beatson and Light (1997) and Beatson, Cher-
rie and Mouat (1999)). Like the fast multipole methods, the BGP algorithm
(Beatson, Goodsell and Powell 1995) is iterative and depends on an initial

structuring of the centres Ξ before the start of the iteration, although that
one is less complicated than the fairly sophisticated hierarchical structure
demanded for the multipole schemes, especially when n > 2.
4.1. The BGP method

We explain the method for n = 2, but we remark that the case n = 3
is under investigation and will probably be admissible and have suitable
software soon. The special further complication for n = 3 is the important
three-dimensional structuring of the data; conceptually there are no changes
in the algorithm we describe below. We also take thin-plate splines only,
although the method has actually been already tested successfully on multi-
quadrics as well (Faul and Powell 1999), and once again there are almost no
changes except for a different k which plays an important rôle below. This is
because the method makes, as we shall see, extensive use of the variational
properties explained in Section 2 of this paper, and those can be extended
to multiquadrics, for instance, although they were first found in connection
with thin-plate splines.
The interpolant we wish to compute still has the same form as in the
introduction for k = n = 2 and φ(r) = r2 log r but we denote the actual,
sought interpolant by s∗ , whereas s denotes in this section only the active
approximation to s∗ at each stage of the algorithm. This is useful for de-
scribing our iterative scheme. The basic idea of the algorithm is derived
from a Lagrange formulation of the interpolant
X
s∗ (x) = f (ξ)Lξ (x), x ∈ Rn , (4.1)
ξ∈Ξ
instead of (1.1), where each Lagrange function Lξ satisfies the Lagrange

conditions
Lξ (ζ) = δζξ , ξ, ζ ∈ Ξ, (4.2)
and is of the form
X
Lξ (x) = λζξ φ(kx − ζk) + pξ (x), x ∈ Rn . (4.3)
ζ∈Ξ
Here, pξ ∈ P12 and λ·,ξ ⊥ P12 (Ξ) for each ξ as usual. Clearly, the computation
of such full Lagrange functions would be just as expensive as solving the
full usual linear interpolation system of equations. Therefore the idea is to
replace (4.2) by local Lagrange conditions which require for each ξ only that
the identity holds for some q = 30 points, say, ζ that are nearby ξ. Hence
we take Lagrange functions that are still of the form (4.3) but with at most
q nonzero coefficients. We end up with an approach to our interpolation
18 M. D. Buhmann
problem that resembles a domain decomposition approach, because we di-

vide the set of centres into subsets and regard the interpolation method as
a local method on those subsets.
So that we can associate with each ξ the Lagrange function Lξ and its
‘active’ point set of centres, we call the latter Lξ ⊂ Ξ, |Lξ | = q. In addition
to having q elements, Lξ must contain a unisolvent subset with respect
to Pk−1
n as required before. We also extract another set of centres with a
unisolvent subset Σ of approximately the same size q from Ξ for which no
local Lagrange functions are computed because that set is sufficiently small
to allow direct solution of the interpolation problem.
In order that an iterative method can be applied, we order the centres
Ξ \ Σ = {ξi }mi=1 for which local Lagrange functions are computed and have
the final requirement that, for all i = 1, 2, . . . , m,
ξi ∈ Lξi ⊂ Ξ \ {ξ1 , ξ2 , . . . , ξi−1 }.
We require that the q points in the set above are those among the centres
that follow ξi which are the closest ones to ξi . There may be ties in the
necessary ordering procedure that can be broken randomly. The Lagrange
conditions that must be satisfied by the local Lagrange functions are now
the locally restricted conditions
Lξ (ζ) = δζξ , ζ ∈ Lξ . (4.4)
We employ the same notation for the local Lagrange functions as for the
full ones, as the latter will no longer occur in this section. Using these
local Lagrange functions, (4.1) is naturally no longer a representation of
the exact approximant s∗ , but only an approximation thereof, and here is
where the iteration and iterative correction come in. Of course the Lagrange
functions, that is, their coefficients, are computed in advance once and for
all and stored before the beginning of the iterations. (This is an O(|Ξ|)
process.) A useful comparison of this approach with the Newton formulation
of univariate polynomial interpolants is made by Powell (1999).
In those iterations, we make an iterative refinement of the approximation
at each step of the algorithm by correcting its residual through updates
s(x) −→ s(x) + Lξ (x) × cξ (x),
where
1 X
cξ (x) = λζξ (f (ζ) − s(ζ)). (4.5)
λξξ
ζ∈Lξ
We shall see in the proof of the next theorem why λξξ is positive and we
may therefore divide by λξξ . The correction (4.5) is added for all ξ ∈ Ξ \ Σ
for each sweep of the algorithm. The final stage of each sweep consists of
the correction
s(x) −→ s(x) + σ(x),
where σ is the full standard thin-plate spline solution of the interpolation
problem with centres Σ computed with a direct method
σ(ξ) = f (ξ) − s(ξ), ξ ∈ Σ.
Here also, s denotes the current approximation to s∗ after all intermediate
steps. This finishes the sweep of the algorithm, and, if we started with a
trivial approximation s = s0 = 0 to s∗ , the sweep replaces sm by sm+1 that
goes into the next sweep. The stopping criterion can be, for instance, that
we terminate the algorithm if all the residuals are sufficiently small. The
main work is clearly the computation of the residuals, the coefficients of the
local Lagrange made available before the start. For that, a method such as
the one described in the next subsection about multipole methods can be
used.
It is quite a novelty that there is a convergence proof of this algorithm,
although it had been known for some time that the method performs ex-
ceptionally well in practical computations even if Ξ contains as many as
50000 points. It turns out that it is not unusual to have an increase of ac-
curacy of one digit per sweep of the algorithm, that is, each sweep reduces
max |s(ξ) − f (ξ)| by a factor of ten, which indicates very fast convergence.
We provide the essentials of the proof of this theorem (Faul and Powell
1999) – which is in fact not very long – because it is highly instructive about
the working of the algorithm and involves many concepts that are typical
for the use and analysis of radial basis functions, such as native spaces and
their seminorms and semi-inner products.
Theorem 13. Let {sj }∞ ∗
j=0 be a sequence of approximations to s gener-
ated by s0 = 0 and the above algorithm. Let the radial function be the
thin-plate spline function and n = 2. Then limj→∞ sj = s∗ .
Proof. We show first that the native space norm ks∗ − sj kφ is monotonic-
ally decreasing with increasing j. Let (·, ·) be the semi-inner product that
corresponds to the native space norm, namely
Z
1 1
(u, v) = n
û(t)v̂(t) dt. (4.6)
(2π) φ̂(ktk)
Here, the requirements kukφ < ∞, kvkφ < ∞ are sufficient for the above
integral to exist, by the Cauchy–Schwarz inequality. In the thin-plate spline
case and in two dimensions, this is by Parseval–Plancherel a multiple of
Z 2
∂ u(x, y) ∂ 2 v(x, y) ∂ 2 u(x, y) ∂ 2 v(x, y) ∂ 2 u(x, y) ∂ 2 v(x, y)
+ 2 + dx dy
∂x2 ∂x2 ∂x∂y ∂x∂y ∂y 2 ∂y 2
and has as kernel the space of linear polynomials.
20 M. D. Buhmann
We start with s = s0 = 0. Each full sweep of the algorithm replaces sj

by sj+1 , but within each sweep there are further m + 1 updates . By those
updates, sj,0 = sj is replaced by sj,1 , and so on until sj,m is replaced by
sj+1 = sj,m+1 = sj,m + σ where σ is defined above. At the (j − 1)st stage,
for each index i, ks∗ − sj−1,i−1 kφ is replaced by
ks∗ − sj−1,i−1 − θξi Lξi kφ ,
where θξi = cξi . It is elementary that we get
ks∗ − sj−1,i−1 − θLξi k2φ = min!
when the linear parameter θ assumes the value
(s∗ − sj−1,i−1 , Lξi )
θ= . (4.7)
kLξi k2φ
We claim that (4.7) is the same as θξi and in the proof of this fact we shall
also show as a by-product that the denominator in (4.7) is positive. Indeed,
recalling that the additional polynomials used are in the kernel of the inner
product (4.6), we get from the reproducing kernel identity
!
X X
kLξi k2φ = λζξi φ(k · −ζk), λτ ξi φ(k · −τ k)
ζ∈Lξi τ ∈Lξi
X X
= λζξi λτ ξi φ(kζ − τ k)
ζ∈Lξi τ ∈Lξi
= λ ξi ξi ,
by (4.4) and because pζ is annihilated by the side conditions on the coeffi-
cients λζξi . This also shows, incidentally, that the important inequality
λ ξi ξi > 0 (4.8)
holds because kLξi k2φ is not zero. Otherwise Lξi would be in the kernel of
our semi-inner product and not able to satisfy the cardinality conditions.
Moreover, by the same token, we get from the reproducing kernel properties
X
(s∗ − sj−1,i−1 , Lξi ) = λζξi (s∗ − sj−1,i−1 , φ(k · −ζk))
ζ∈Lξi
X
= λζξi (s∗ (ζ) − sj−1,i−1 (ζ))
ζ∈Lξi
X
= λζξi (f (ζ) − sj−1,i−1 (ζ)).
ζ∈Lξi
Next, we have to consider the alteration to ks∗ − sj−1,m kφ as soon as σ is

added to sj−1,m . In order to prove the monotonic decrease of ks∗ − sj−1,m kφ
when this happens, we need to prove that the following inner product van-
ishes !
X
∗
s − sj−1,m − σ, λ̂τ φ(k · −τ k) + p̂
τ ∈Σ
!
X
= s∗ − sj−1,m − σ, λ̂τ φ(k · −τ k)
τ ∈Σ
= 0.
In the above, the coefficients λ̂τ are real and p̂ ∈ Pk−1 n . This orthogonality
relation is established using the same facts as those we required for showing
kLξi k2φ = λξi ξi and therefore we do not repeat the arguments.
In summary, we now know that ksj,i − s∗ kφ tends to a limit for all fixed
i, and j increasing, because it is bounded below by zero and monotonically
decreasing, as it is a subsequence of the monotonically decreasing doubly
indexed sequence {ksj,i − s∗ kφ }∞ j=0,0≤i≤m . Moreover, by the definition of our
semi-inner product
(s∗ − sj−1,i−1 , Lξi )2
ks∗ − sj−1,i−1 − θξi Lξi k2φ = ksj−1,i−1 − s∗ k2φ − .
kLξi k2φ
Therefore, for all centres
(s∗ − sj−1,i−1 , Lξi ) → 0, j → ∞, ξi ∈ Ξ \ Σ, (4.9)
because ksj,i − s∗ kφ converges. In particular, it follows that
(s∗ − sj , Lξ1 ) → 0, j → ∞. (4.10)
Moreover, sj,1 = sj + (s∗ − sj , Lξ1 )Lξ1 /kLξ1 k2φ , so that we also get

∗ (s∗ − sj , Lξ1 )Lξ1
s − sj − , Lξ2 → 0, j → ∞.
kLξ1 k2φ
Therefore
(s∗ − sj , Lξ2 ) → 0, j → ∞,
and indeed
(s∗ − sj , Lξ ) → 0, j → ∞, ξ ∈ Ξ \ Σ. (4.11)
Finally, we observe
X
(s∗ − sj , Lξi ) = λζξi (s∗ (ζ) − sj (ζ)) → 0, j → ∞. (4.12)
ζ∈Lξi
Recalling that s∗ −sj restricted to Σ vanishes anyway, and recalling that λξi ξi
is positive, we now go backwards and start from i = m with an induction
argument, whereupon (4.8), (4.9) and (4.12) imply sj (ζ) → s∗ (ζ) as j → ∞
22 M. D. Buhmann
for all ζ ∈ Ξ. This implies sj → s∗ for j → ∞, as demanded, because

the space spanned by the translates φ(k · −ξk) plus the polynomial space is
finite-dimensional. 2
4.2. Fast multipole methods

Another approach for approximation and iterative refinement to the ra-
dial basis function interpolants is that of fast multipole methods. These
algorithms are based on analytic expansions of the underlying radial func-
tions for large argument. We see that this is possible for all non-compactly
supported radial basis functions that have been mentioned, because they are
smooth away from their centres. For the compactly supported ones, the ap-
proach for fast evaluation and solution of the linear system is, of course, not
required. The salient ideas date back to a paper of Greengard and Rokhlin
(1987) where the methods are used to solve numerically integral equations –
this is related to our radial basis function approximations because the sums
and coefficients (‘weights’) can be viewed as discretizations of integrals.
The methods require structuring the data in a hierarchical way before the
onset of the iteration, and computing so-called far-field expansions (Laurent
series expansions for large arguments, as already alluded to) in advance.
The goal is to reduce the cost of a single approximate evaluation of the
radial function sum down to O(1) and storage to O(|Ξ|). The far-field
expansions exploit the fact that radial basis functions of the form (2.1)
are analytic except at the origin, even when made multivariate through
composition with Euclidean norms. Therefore, they can be approximated
well away from the origin by a truncated Laurent expansion. The accuracy
(that is, the length of the truncated expansion) to this can be preset and
made to match any chosen accuracy ε, for instance a small positive multiple
of the machine precision of the computer in use, but of course the cost of
the method, that is, the multiplier in the operation count, rises with larger
accuracy. In tandem with the hierarchical structure of the centre-set, which
we shall explain in some detail below, this allows approximative evaluation
of many radial basis function terms whose centres are close to each other
simultaneously by a single finite Laurent expansion that is inexpensive, as
long as the argument x is far from the said cloud of centres. By contrast, all
radial functions whose translates are close to x are computed exactly and
explicitly. We shall call those centres the near field below.
The hierarchical structure of the centres is fundamental to the algorithm
because it decides and orders what is far and what is near any given x
where we wish to evaluate. It is built up as follows. We assume for the sake
of an easy exposition that Ξ ⊂ [0, 1]2 and that the ξ are fairly uniformly
distributed in that unit square so that uniform subdivisions are acceptable.
We form a quad-tree of centres which contains as a root Ξ and as the next
children the intersections of the four quarter squares of the unit square with
Ξ. They are in turn divided into four grandchildren each in the same fashion.
We stop this process at a predetermined level that in part determines the
accuracy of the calculation in the end. Each member of this family is called
a panel. We point out that there are implementations where the divisions
are only by two and not by four (Powell 1993). There is no principal reason
for making this choice that forms a binary tree instead of a quad-tree and
both approaches have been tried successfully.
Now we must define the far field and the near field for each x where s(x) is
computed. Given x, all centres ξ that are in the near field give rise to explicit
evaluations of φ(kx − ξk). The near field consists simply of contributions
from all points which are not ‘far’, according to the following definition: we
say x is far away from a panel T and therefore from all centres in that panel,
if there is at least one more panel between x and T , and if this panel is on
the same level of parenthood as T itself.
Next, we have to decide how to group the far points. All panels Q are
in the evaluation list of a panel T if Q is either at the same or a coarser
(higher) level than T and every point in T is far away from Q, and if, finally,
T contains a point that is not far away from the parent of Q. Thus the far
field of an x whose closest centre is in a panel T is the sum of all
X
sQ (x) = λξ φ(kx − ξk) (4.13)
ξ∈Q
such that Q is in the evaluation list of T . For each (4.13), a common Laurent
series is computed. Since we do not know the value of x that is to be inserted
into (4.13), we compute the coefficients of the Laurent series and insert x
later on.
In a set-up process, we compute these expansions of the radial function for
large argument, that is, their coefficients, and store them; when x is provided
at the evaluation stage, we combine those expansions for all centres from
each evaluation list, and in an additional, final step we can simplify further
by approximating the whole far field by one Taylor series.
A typical Laurent series expansion to thin-plate spline terms, which we
still use as a paradigm for algorithms with more general classes of radial basis
functions, is as follows. To this end, we work in two dimensions n = 2 and
identify two-dimensional real space with one-dimensional complex space.
Lemma 1. Let z and t be complex numbers, and let
φt (z) := kt − zk2 log kt − zk.
Then, for all kzk > ktk, and denoting the real part by Re ,
( ∞ !)
X 1 t k
φt (z) = Re (z̄ − t̄)(z − t) log z − ,
k z
k=1
24 M. D. Buhmann
which is the same as

( ∞
)
X
(kzk2 − 2Re (t̄z) + ktk2 ) log kzk + Re (ak z̄ + bk )z −k ,
k=0
where bk = −t̄ak and a0 = −t and ak = tk+1 /[k(k + 1)] for positive k.

Moreover, if the above series is truncated after p + 1 terms, the remainder
is bounded above by

ktk2 c+1 1 p
(p + 1)(p + 2) c − 1 c
with c = kz/tk.
In summary, the principal steps of the whole algorithm are as follows.
Set-up
(1) Perform the repeated subdivision of the square down to log |Ξ| levels
and sort the elements of Ξ into the finest level panels.
(2) Form the Laurent series expansions (i.e., compute their coefficients) for
all fine level panels R.
(3) Translate centres of expansions so that the expansions can be re-used
for other centres and, by working up the tree towards coarser levels,
form analogous expansions for all less refined levels.
(4) Working down the tree from the coarsest level, compute Taylor expan-
sions of the whole far field for each panel Q.
Evaluation at x
(1) Locate the finest level panel Q containing the centre closest to x.
(2) Evaluate s(x) by computing near field contributions explicitly and by
using the Taylor approximation of the entire far field. For the far field,
we use all panels that are far away from x and are not subsets of any
coarser panels already considered.
The computational cost without set-up is O(|Ξ|) because of the hierarchical
structure. The set-up cost is O(log |Ξ|), because of the tree structure, but the
constant contained in this estimate may be large because of the computation
of the various expansions and the design of the tree. This is so although each
expansion is an O(1) procedure. In practice it turns out that the method is
superior to direct computation if |Ξ| is at least of the order of 200 points.
In which way is this algorithm now related to the computation of inter-
polation coefficients? It is related in one way because the efficient evaluation
of the linear combination of thin-plate spline translates is required in the
first algorithm presented in this section. There the residuals fξ − s(ξ) played
an important rôle, s(ξ) being the current linear combination of thin-plate
splines at a centre and the current approximation to the solution we re-

quire. Therefore, to make the BGP algorithm efficient, fast evaluation of s
is needed, partly because our radial basis functions are conditionally posit-
ive definite (the preconditioning making the matrices positive (semi-)definite
matrices) and partly because the condition numbers severely influence the
convergence properties of conjugate gradient methods.
However, the importance of fast availability of residuals is not only re-
stricted to the BGP method. Other iterative methods for computing radial
basis function interpolants such as conjugate gradient methods are also in
need of these residuals. In order to apply those, however, a preconditioning
method is usually needed.
There are various possible improvements to the algorithm of this subsec-
tion. For instance, we can use an adaptive method for subdividing into the
hierarchical structure of panels, so that the panels may be of different sizes,
but always contain about the same number of centres. This is particularly
advantageous when the data are highly non-uniformly distributed.
5. Interpolation on spheres
Because of the many applications that suit radial basis functions in geodesy,
there is already a host of papers that specialize radial basis function approx-
imation and interpolation to spheres. Freeden and co-workers (1981, 1986,
1995) have made a very large number of contributions to this aspect of ap-
proximation theory of radial basis functions. There are excellent and long
review papers available from the work of this group (see the cited references)
and we will therefore be relatively brief in this section. Of course, we no
longer use the conventional Euclidean norm in connection with a univariate
radial function when we approximate on the (n − 1) sphere S n−1 within
Rn but apply so-called geodesic distances. Therefore the standard notions
of positive definite functions and conditional positive definiteness no longer
apply, and one has to study new concepts of (conditionally) positive defin-
ite functions on the (n − 1) sphere. This started with Schoenberg (1942),
who characterized positive definite functions on spheres as those ones whose
expansions in series of Gegenbauer polynomials always have nonnegative
coefficients. Xu and Cheney (1992) studied strict positive definiteness on
spheres and gave necessary and sufficient conditions. This was further gen-
eralized by Ron and Sun in 1996
Recent papers by Jetter, Stöckler and Ward (1999) and Levesley, Light,
Ragozin and Sun (1999) use native spaces, (semi-)inner products and repro-
ducing kernels (cf. Saitoh (1988) for a treatise on the theory of reproducing
kernels) to derive approximation orders in a very similar fashion to the work
(`) `
summarized in Section 2. They all apply the spherical harmonics {Yk }dk=1
that form an orthonormal basis of the d` -dimensional space of polynomials
26 M. D. Buhmann
on the sphere in P`n (S n−1 ) ∩ P`−1

n (S
n−1 )⊥ . They are called spherical har-
monics because they are the restrictions of polynomials of total degree ` to

the sphere and are in the kernel of the Laplace operator ∆. The dimension
d` is computable, and this is not difficult, but it is not important to us at
this stage.
Then, a native space X is defined for all expansions of functions on the
sphere in spherical harmonics: namely, the native space’s elements are
d
∞ X̀
X (`)
f (x) = fˆ`k Yk (x), x ∈ S n−1 , (5.1)
`=0 k=1
whose coefficients fˆ`k satisfy certain square summability conditions with

prescribed positive real weights a`k . In other words, the native space is
defined by
( ∞ d )
X X̀ |fˆ |2
`k
X = f <∞ . (5.2)
a`k
`=0 k=1
The native space X given in (5.2) and functions (5.1) will give rise to a
reproducing kernel that is positive definite, but if we enlarge the space by
starting the first sum in (5.2) only at ` = κ and thereby weakening condi-
tions, we can also get conditionally positive definite (reproducing) kernels
(Levesley et al. 1999) for the ensuing spaces Xκ . Then the native space will
be a semi-Hilbert space, that is, the inner product that we shall describe
shortly has a nullspace
( )
d
κ−1 X̀
X
(`)
K = f f = fˆ`k Y k .
`=0 k=1
Now, a standard choice for the positive weights for defining the space X
which are often independent of k is a`k = (1 + λ` )−s . This gives rise to the
Sobolev space H s (S n−1 ). Here the λ` = `(` + n − 2) are eigenvalues of the
Laplace–Beltrami operator.
The inner product that the native space is equipped with can be described
by
d
∞ X̀
X 1 ˆ
hf, gi = f`k ĝ`k ,
a`k
`=0 k=1
where the coefficients are defined through (5.1); they are still assumed to
be positive. The reproducing kernel that results from this Hilbert space X
with the above inner product and that corresponds to the function of our
previous radial basis functions in the native space is, when x and y are on
the sphere,
d
∞ X̀
X (`) (`)
φ(x, y) = a`k Yk (x)Yk (y), x, y ∈ S n−1 . (5.3)
`=0 k=1
This can be simplified by the famous addition theorem (Stein and Weiss
1971) to φ(x, y) = φ(xT y), where
∞
1 X
φ(t) = d` a`k P` (t), (5.4)
ωn−1
`=0
ωn−1 being the measure of the unit sphere, if the coefficients are constant
with respect to k. Here, the d` are as above and P` is a Gegenbauer poly-
nomial (Abramowitz and Stegun 1972) normalized by P` (1) = 1. Therefore,
we now use (5.3) or (5.4) for interpolation on the sphere, in the same place
and with the same centres Ξ as before, but they are from the sphere them-
selves of course. Convergence estimates are available from all three sources
mentioned above that vary in approaching the convergence question. Using
the mesh norm
h = sup inf arccos(xT ξ),
x∈S n−1 ξ∈Ξ
Jetter et al. prove the following theorem. The notation |Ξ| is for the cardin-
ality of the set Ξ as before.
Theorem 14. Let X and Ξ be as above with the given mesh norm h. Let
κ be a positive integer such that h ≤ 1/(2κ). Then, for any f ∈ X, there is
a unique interpolant s in
span {φ(ξ, ·) | ξ ∈ Ξ}
that interpolates f on Ξ and satisfies the error estimate
X∞
2 5(|Ξ| + 1) 2
ks − f k∞ ≤ kf kφ d` max a`k . (5.5)
ωn−1 k=1,...,d`
`=κ+1
Corollary 1. Let the assumptions of the previous theorem hold and sup-
pose further that |Ξ| + 1 ≤ C1 κn−1 and
C2 1
≤h≤ .
1+κ 2κ
Then the said interpolant s provides
(α−n)/2 !
h
ks − f k∞ = O
C2
or
exp(−αC2 /2h)
ks − f k∞ =O ,
h(n−1)/2
28 M. D. Buhmann
respectively, if d` ×maxdk=1
`
a`k is bounded by a constant multiple of (1+`)−α
for an α > n or by a constant multiple of exp(−α(1 + `)) for a positive α,
respectively.
An error estimate of Levesley et al., which includes conditionally positive
definite kernels for κ = 1 or κ = 2, is as follows, for n = 2 (see also Freeden
and Hermann (1986) for a similar, albeit slightly weaker result).
Theorem 15. Let X, Ξ, h and κ be as above, let s be the minimal norm
interpolants to f ∈ Xκ on Ξ. When φ is twice continuously differentiable
on [1 − ε, ε] for some ε ∈ (0, 1), then
ks − f k∞ ≤ Ch2 kf kφ ,
so that, in particular, a polynomial p ∈ Pκ−1
3 (S 2 ) is added to the s used in
Theorem 14.
6. Applications
6.1. Numerical solution of partial differential equations
Given that radial basis functions are known to be useful to approximate
multivariate functions efficiently, it is suitable to apply them to approxim-
ate solutions of partial differential equations numerically. Three approaches
have been tried and tested in this direction, namely collocation techniques,
variational formulations and boundary element methods, all in order to solve
elliptic partial differential equations with boundary values given. There are
various reasons why radial basis functions are useful for these three ap-
proaches. The first of these is useful because we know much about existence
and accuracy of radial basis function interpolants, especially when the data
are scattered, which is useful for non-grid collocation. The second resembles
typical finite element applications, where usually radial basis functions of
compact support are used to mimic the standard finite element approach
with multivariate piecewise polynomials. Finally, boundary element meth-
ods are suitable in several cases when radial basis functions are known to
be fundamental solutions (Green’s functions) of elliptic partial differential
operators, most notably powers of the Laplace operator. An example is the
thin-plate spline radial basis function and the bi-Laplacian operator. After
all, in boundary element methods, explicit solutions of the associated homo-
geneous problem are required in advance, for which it is immensely helpful
to have Green’s functions to work with.
Naturally, an important decision is the choice of radial basis function,
especially whether globally or locally supported ones should be used. In
a Galerkin approach, locally supported elements are almost always em-
ployed. Further, the use of radial basis functions becomes particularly in-
teresting when nonlinear partial differential equations are solved or non-grid
approaches are used, for instance because of non-smooth domain boundar-

ies, where non-uniform knot placement is always important to modelling the
solution to good accuracy.
We begin with a description of the collocation approach, which is the
first approach one is tempted to try because of our knowledge of interpol-
ation properties of the radial basis functions. The first important decision
is whether to use the well-known, standard globally supported radial func-
tions such as multiquadrics or the new compactly supported ones that are
described earlier in this review. Since the approximation properties of the
latter are not as good as the former ones, we have a trade-off between ac-
curacy on one hand and sparsity of the collocation matrix on the other
hand. Compactly supported ones, if scaled suitably, give banded colloca-
tion matrices, while the globally supported ones give no sparsity to speak
of. When we use the compactly supported radial functions we have, in fact,
another trade-off which we have not mentioned so far, because even their
scaling pits accuracy against sparseness of the matrix. If we impose no scal-
ing to the radial basis function as in Theorem 12, we do have satisfactory
convergence, as shown in Theorem 10, but basically the radial basis function
behaves as a globally supported one, with essentially full matrices, since the
centres become dense in each of the radial functions’ supports. In the other
extreme case, when we scale so that there is always a uniformly bounded
number of centres inside each support, we run the risk of losing convergence
altogether – but the interpolation matrix may be diagonal (and nonsingular,
of course). There are several approaches to fixing this problem, and we will
mention two of them while describing algorithms.
One typical linear partial differential equation problem suitable for col-
location techniques reads
Lu(x) = f (x), x ∈ Ω ⊂ Rn , (6.1)

Bu|∂Ω = q, (6.2)
where Ω is a domain with suitably smooth – at least Lipschitz-continuous –

boundary ∂Ω and f , q are prescribed functions. Here L is a linear differential
operator and B a boundary operator. We will soon come to some specific
nonlinear examples in the context of boundary element techniques.
The usual approach to collocation is then for centres Ξ that are partitioned
in two disjoint sets Ξ1 and Ξ2 , the former from the domain, the latter from
its boundary, to solve the Hermite–Birkhoff interpolation system
Λξ uΞ = f (ξ), ξ ∈ Ξ1 ,
Λζ uΞ = q(ζ), ζ ∈ Ξ2 . (6.3)
30 M. D. Buhmann
The approximants uΞ are defined by the sums

X X
uΞ (x) = cξ Λξ φ(kx − ξk) + dζ Λζ φ(kx − ζk).
ξ∈Ξ1 ζ∈Ξ2
The Λξ and Λζ are suitable discrete functionals to describe our operators L

and B on the discrete set of centres. In the above display the operators are
applied with respect to the variable x. Examples for such discrete function-
als to replace the operators L and B are obtained, for example, by replacing
derivatives by symmetric differences or one-sided differences for the bound-
ary in case of Neumann problems. Thus we end up with a square symmetric
system of linear equations whose collocation matrix is nonsingular if, for
instance, the radial basis function is positive definite, smooth enough for
application of the operators Λ, and the discrete linear functionals are lin-
early independent functionals in the dual space of the native space of the
radial basis functions (see Section 2 and Wu (1992) for the details). An error
estimate is given in Franke and Schaback (1998). For those error estimates,
it has been noted that more smoothness of the radial basis function is re-
quired than for a comparable finite element setting in order to get the same
approximation orders, but clearly, the radial basis function setting has the
distinct advantage of availability in any dimension and the absence of grids
or triangulations.
If a compactly supported radial basis function is used, the necessary scal-
ing leads to the aforementioned trade-off between accuracy and bandwidth
of the matrix. In fact, the conditioning of the collocation matrix is also af-
fected, becoming worse with smaller η, with φ(·/η) being used, although the
matrix is positive definite. A Jacobi preconditioning by the diagonal p values
helps here, so the matrix A is replaced by P −1 AP −1 where P = diag(A),
the diagonal elements of the matrix being positive. Moreover, one can
use a multilevel method (Narcowich, Schaback and Ward (1999), Fasshauer
(1999)) where numerical approximations {uk }N k=0 are computed on nested
sets of centres Ξk ⊃ Ξk−1 , k = 1, 2, . . . , N , ΞN = Ξ, and, within each sweep
of the algorithm, a new approximation to the desired solution is computed
as follows. For instance, to solve Lu = f on the domain Ω with Dirich-
let boundary conditions only, at each step k of one sweep one computes,
starting with u0 = 0,
Lũk = (f − Luk−1 )
and sets uk = ũk + uk−1 . Unfortunately, little is known about the conver-
gence behaviour of such a multilevel method.
In the event that a Galerkin method is applied, for instance, to the Helm-
∂
holtz equation with Neumann conditions when L = −∆ + I and B = ∂n ,
we end up with a square system of linear equations, the stiffness equations
for uΞ ,
a(uΞ , φ(k · −ξk)) = (f, φ(k · −ξk))L2 (Ω) , ξ ∈ Ξ,
with Z
a(u, v) = (∇u)T (∇v) + uv.
Ω
If φ is a radial basis function of compact support such that its Fourier
transform satisfies the decay estimate |φ̂(r)| = O(r−2k ), then Franke and
Schaback (1998) establish the convergence estimate
ku − uΞ kH 1 (Ω) ≤ Chσ−1 kukH σ (Ω) ,
where h is given by (2.2) and k ≥ σ > n/2 + 1.
We now outline the third method, that is, a boundary element (BEM)
method, following Pollandt (1997). The dual reciprocity method uses the
second Green’s formula and a fundamental solution φ(k · k) of the Laplace
operator ∆, in order to reformulate a boundary value problem as a boundary
integral problem over a space of one dimension lower. This will then lead
through discretization to a linear system with a full matrix for collocation by
radial basis functions, in the way familiar from other applications of radial
basis function interpolation. The radial basis function that occurs in that
context is this fundamental solution, and, naturally, it is highly relevant in
this case that the Laplace operator is rotationally invariant and has radial
functions as Green’s functions. We give a concrete example. Namely, for a
nonlinear problem on a domain Ω ⊂ Rn with Dirichlet boundary conditions
such as the following one with a nonlinear right-hand side
∆u(x) = u2 (x), x ∈ Ω ⊂ Rn , (6.4)
u|∂Ω = q, (6.5)
the goal is to approximate the solution u of the elliptic partial differential
equation on the domain by g plus a boundary term r̃ that satisfies ∆r̃ ≡ 0
on the domain. To this end, one gets after an application of Green’s formula
the equation on the boundary that u(x) is the same as
Z Z
∂ ∂
u(y)2 φ(kx − yk) dy − φ(kx − yk) u(y) − u(y) φ(kx − yk) dΓy
Ω ∂Ω ∂ny ∂ny
(6.6)
for x ∈ Ω, where ∂n∂ y is the normal derivative with respect to y on Γ = ∂Ω.
The radially symmetric φ is still the fundamental solution of the Laplace
operator used in the formulation of the differential equation above. Further,
one gets after two applications of Green’s formula the equation
Z
1 ∂
(u(x) − g(x)) + φ(kx − yk) × (u(y) − g(y)) −
2 ∂Ω ∂n y
∂
(q(y) − g(y)) φ(kx − yk) dΓy = 0, x ∈ ∂Ω.
∂ny
We will later use this equation to approximate the boundary part of the
32 M. D. Buhmann
solution, that is, the part satisfying the boundary conditions. Now we as-
sume that there are real coefficients λξ such that the infinite expansion
(which will be truncated later on)
X
u2 (y) = λξ φe (ky − ξk), y ∈ Ω,
ξ
holds, and set

X
g(y) = e (ky − ξk),
λξ Φ y ∈ Ω,
ξ
so that ∆g = u2 everywhere with no boundary conditions. Therefore we

are first solving a homogeneous problem. Here φe is a suitable radial basis
function, which is to be distinguished from the fundamental solution φ,
with ∆Φ e · k), and the centres ξ are from Ω. We now replace the
e (k · k) = φ(k
infinite sums by finite ones (i.e., we approximate the infinite expansions by
finite sums for a suitable Ξ), so that exchanging summation and the Laplace
operator and term-by-term differentiation in the sum is admitted:
X
u2 (y) = λξ φe (ky − ξk), y ∈ Ω, (6.7)
ξ∈Ξ
and
X
g(y) = e (ky − ξk),
λξ Φ y ∈ Ω.
ξ∈Ξ
We require that the equation in the display after (6.6) holds for finitely many
points x = ζj ∈ ∂Ω, j = 1, 2, . . . , t, only. Then we solve for the coefficients
λξ by requiring that (6.7) holds for y = ξ, for all ξ ∈ Ξ. This fixes the λξ by
interpolation (collocation in the language of differential equations), whereas
the equation after (6.6) determines the normal derivative ∂n∂ y u(y) on ∂Ω,
where we are replacing ∂n∂ y u(y) by another approximant, a polynomial
spline τ (y), for instance. Thus the spline is found by requiring the above
equation for all x = ζj ∈ ∂Ω, j = 1, 2, . . . , t, and choosing a suitable t.
Finally, an approximation u e(x) to u(x) is determined on Ω by the identity
Z
∂
u
e(x) := g(x) + (q(y) − g(y)) φ(kx − yk) dΓy −
∂Ω ∂n y
Z
∂g(y)
φ(kx − yk) τ (y) − dΓy , x ∈ Ω, (6.8)
∂Ω ∂ny
where the boundary term r̃ corresponds to the second term on the right-hand
side of the display.
Now, all expressions on the right-hand side are known. This is an outline
of the approach but we have skipped several important details. Nonethe-
less, one can clearly see how radial basis functions appear in this algorithm;
indeed, it is most natural to use them here, since many of them are funda-
mental solutions of Laplace operators or their iterates in certain dimensions.
In the above example and n = 2, φ(r) = (2π)−1 log r, φ(r)e = r2 log r (thin-
e
plate splines) and Φ(r) 1 4 1 4
= 16 r log r − 32 r are the correct choices. As
the collocation matrices that appear in the boundary integral equation we
are left with in the method are dense and as interpolation is not absolutely
necessary in this approach, it can be substituted by quasi-interpolation.
Quasi-interpolation has asymptotically essentially the same approximation
behaviour (i.e., convergence speed) as interpolation but does not require
any data-dependent linear systems to be solved; see also Buhmann (1990a).
Convergence theorems for the method are available in the paper by Pollandt.
6.2. Other applications

We outline a few further practical applications of radial basis functions.
They include, for example, mappings of two- or three-dimensional images
such as portraits or underwater sonar scans into other images for comparison.
Here interpolation comes into play because some special features of an image
may have to be preserved while others need not be mapped exactly, thus
enabling a comparison of some features that may differ while at the same
time retaining others. Such so-called ‘markers’ can be, for example, certain
points of the skeleton in an x-ray that has to be compared to another one,
taken at another time. The same structure appears if we wish to compare
sonar scans of a harbour at different times, the rocks being suitable as
markers this time. Thin-plate splines turned out to be excellent for such
very practical applications (Barrodale and Zala 1997, 1999). Work of this
kind led to the invention of the methods for fast evaluation of thin-plate
splines and other radial basis functions discussed above, because, after the
interpolation, the computed interpolant had to be evaluated on a very fine
square grid in two dimensions for analysis or display (Powell 1993).
Measurements of gravitational potential or temperature on the Earth’s
surface at ‘scattered’ meteorological stations or measurements on other mul-
tidimensional objects, may give rise to interpolation problems that require
scattered data. Multiquadric approximations are performing well for this
type of use (Hardy 1990). Much work on radial basis functions on spheres
which we have only touched upon in this article originates from those ap-
plications of geophysics.
Many applications involve high-dimensional interpolation or approxima-
tion problems when data are coming through many ‘channels’, for instance
electrodes measuring brain activity from nerve cells. Typical applications
from neural physics produce 50–100 dimensional data taken from those elec-
trodes recording measurements from the brain and require post-processing
for smoothing, for example. Radial basis functions have been used extens-
34 M. D. Buhmann
ively for this (e.g., Eckhorn (1999), Hochreiter and Schmidhuber (1999),
Anderson, Das and Keller (1998)).
The approximation to so-called learning situations by neural networks
usually leads to very high-dimensional interpolation problems with scattered
data. Girosi (1992) mentions radial basis functions as a very suitable ap-
proach to this, partly because of their availability in arbitrary dimensions,
and their smoothness. A typical application is in fire detectors. An advanced
type of fire detector has to consider several measured parameters, such as
colour, spectrum, intensity, movement of an observed object from which it
must decide whether, for instance, it is looking at a fire in the room or not,
because the apparent fire is reflected sunlight. There is a learning procedure
before the implementation of the device, where several prescribed situations
(these are the data) are tested and the values zero (no fire) and one (fire)
are interpolated, so that the device can ‘learn’ to interpolate between these
standard situations for general cases later when it is used in real life. Radial
basis function methods have been tried very successfully for this application
because they are excellent tools for high-dimensional problems that will
undoubtedly find many more applications in real life, such as polynomial
splines have done in at least the last 30 years and still do now.
REFERENCES
M. Abramowitz and I. Stegun, eds (1972), Handbook of Mathematical Functions,
National Bureau of Standards, Washington.
R. W. Anderson, S. Das and E. L. Keller (1998), ‘Estimation of spatiotemporal
neural activity using radial basis function networks’, J. Comput. Neuroscience
5, 421–441.
N. Arad, N. Dyn and D. Reisfeld (1994), ‘Image warping by radial basis functions:
applications to facial expressions’, Graphical Models and Image Processing
56, 161–172.
R. Askey (1973), ‘Radial characteristic functions’, MRC Report 1262, University
of Wisconsin, Madison.
I. Barrodale and C. A. Zala (1997), ‘MJDP–BCS industrial liaison: applications
to defence science’, in Approximation and Optimization: Tributes to M. J. D.
Powell (M. D. Buhmann and A. Iserles, eds), Cambridge University Press,
Cambridge, pp. 31–46.
I. Barrodale and C. A. Zala (1999), ‘Warping aerial photographs to orthomaps
using thin plate splines’, Adv. Comput. Math. 11, 211–227.
B. J. C. Baxter (1991), ‘Conditionally positive functions and p-norm distance
matrices’, Constr. Approx. 7, 427–440.
R. K. Beatson and W. A. Light (1992), ‘Quasi-interpolation in the absence of
polynomial reproduction’, in Numerical Methods of Approximation Theory
(D. Braess and L. L. Schumaker, eds), Birkhäuser, Basel, pp. 21–39.
R. K. Beatson and W. A. Light (1997), ‘Fast evaluation of radial basis functions:

methods for 2-dimensional polyharmonic splines’, IMA J. Numer. Anal. 17,
343–372.
R. K. Beatson and G. N. Newsam (1992), ‘Fast evaluation of radial basis fun-
ctions: I’, Comput. Math. Appl. 24, 7–19.
R. K. Beatson, G. Goodsell and M. J. D. Powell (1995), ‘On multigrid techniques
for thin plate spline interpolation in two dimensions’, in Vol. 32 of Lectures
in Applied Mathematics, pp. 77–97.
R. K. Beatson, J. B. Cherrie and C. T. Mouat (1999), ‘Fast fitting of radial basis
functions: methods based on preconditioned GMRES iteration’, Adv. Comput.
Math. 11, 253–270.
A. Bejancu (1997), ‘The uniform convergence of multivariate natural splines’,
DAMTP Technical Report, University of Cambridge.
A. Bejancu (1999), ‘Local accuracy for radial basis function interpolation on finite
uniform grids’, J. Approx. Theory 99 242–257.
M. D. Buhmann (1990a), ‘Multivariate interpolation in odd-dimensional Euclidean
spaces using multiquadrics’, Constr. Approx. 6, 21–34.
M. D. Buhmann (1990b), ‘Multivariate cardinal-interpolation with radial-basis fun-
ctions’, Constr. Approx. 6, 225–255.
M. D. Buhmann (1998), ‘Radial functions on compact support’, Proceedings of the
Edinburgh Mathematical Society 41, 33–46.
M. D. Buhmann (2000), ‘A new class of radial basis functions with compact sup-
port’, to appear in Math. Comput.
M. D. Buhmann and C. A. Micchelli (1991), ‘Multiply monotone functions for
cardinal interpolation’, Adv. Appl. Math. 12, 359–386.
M. D. Buhmann and C. A. Micchelli (1992), ‘On radial basis approximations on
periodic grids’, Mathematical Proceedings of the Cambridge Philosophical So-
ciety 112, 317–334.
J. Duchon (1976), ‘Interpolation des fonctions de deux variables suivant le principe
de la flexion des plaques minces’, Rev. Française Automat. Informat. Rech.
Opér. Anal. Numer. 10, 5–12.
J. Duchon (1978), ‘Sur l’erreur d’interpolation des fonctions de plusieurs variables
pars les Dm -splines’, Rev. Française Automat. Informat. Rech. Opér. Anal.
Numer. 12, 325–334.
J. Duchon (1979), ‘Splines minimizing rotation-invariant semi-norms in So-
bolev spaces’, in Constructive Theory of Functions of Several Variables
(W. Schempp and K. Zeller, eds), Springer, Berlin/Heidelberg, pp. 85–100.
N. Dyn, F. J. Narcowich and J. D. Ward (1997), ‘A framework for interpolation
and approximation on Riemannian manifolds’, in Approximation and Optim-
ization: Tributes to M. J. D. Powell (M. D. Buhmann and A. Iserles, eds),
Cambridge University Press, Cambridge, pp. 133–144.
R. Eckhorn (1999), ‘Neural mechanisms of scene segmentation: recordings from
the visual cortex suggest basic circuits for linking field models’, IEEE Trans.
Neural Net. 10, 1–16.
R. Estrada (1998), ‘Regularization of distributions’, Int. J. Math. Math. Sci. 21,
625–636.
36 M. D. Buhmann
G. Fasshauer (1999), ‘Solving differential equations with radial basis functions:

multilevel methods and smoothing’, Adv. Comput. Math. 11, 139–159.
A. C. Faul and M. J. D. Powell (1999), ‘Proof of convergence of an iterative tech-
nique for thin plate spline interpolation in two dimensions’, Adv. Comput.
Math. 11, 183–192.
M. Floater and A. Iske (1996), ‘Multistep scattered data interpolation using com-
pactly supported radial basis functions’, J. Comput. Appl. Math. 73, 65–78.
C. Franke and R. Schaback (1998), ‘Solving partial differential equations by colloc-
ation using radial basis functions’, Comput. Math. Appl. 93, 72–83.
W. Freeden (1981), ‘On spherical spline interpolation and approximation’, Math.
Meth. Appl. Sci. 3, 551–575.
W. Freeden and P. Hermann (1986), ‘Uniform approximation by harmonic splines’,
Math. Z. 193, 265–275.
W. Freeden, M. Schreiner and R. Franke (1995), ‘A survey on spherical spline
approximation’, Technical Report 95–157, University of Kaiserslautern.
F. Girosi (1992), ‘Some extensions of radial basis functions and their applications
in artificial intelligence’, Comput. Math. Appl. 24, 61–80.
M. von Golitschek and W. A. Light (2000), ‘Interpolation by polynomials and radial
basis functions on spheres’, to appear in Constr. Approx.
L. Greengard and V. Rokhlin (1987), ‘A fast algorithm for particle simulations’, J.
Comput. Phys. 73, 325–348.
R. L. Hardy (1990), ‘Theory and applications of the multiquadric-biharmonic me-
thod’, Comput. Math. Appl. 19, 163–208.
S. Hochreiter and J. Schmidhuber (1999), ‘Feature extraction through Lococode’,
Neural Computation 11, 679–714.
K. Jetter, J. Stöckler and J. D. Ward (1999), ‘Error estimates for scattered data
interpolation on spheres’, Math. Comput. 68, 733–747.
M. J. Johnson (1997), ‘An upper bound on the approximation power of principal
shift-invariant spaces’, Constr. Approx. 13, 155–176.
M. J. Johnson (1998a), ‘A bound on the approximation order of surface splines’,
Constr. Approx. 14, 429–438.
M. J. Johnson (1998b), ‘On the error in surface spline interpolation of a compactly
supported function’, manuscript, University of Kuwait.
M. J. Johnson (1999), ‘Approximation in Lp (Rd ) from spaces spanned by the
perturbed integer translates of a radial function’, manuscript, University of
Kuwait.
D. S. Jones (1982), The Theory of Generalised Functions, Cambridge University
Press, Cambridge.
J. Levesley, W. A. Light, D. Ragozin and X. Sun (1999), ‘A simple approach to the
variational theory for interpolation on spheres’, in New Developments in Ap-
proximation Theory (M. D. Buhmann, M. Felten, D. Mache and M. W. Müller,
eds), Birkhäuser, Basel, pp. 117–143.
W. R. Madych and S. A. Nelson (1990), ‘Polyharmonic cardinal splines’, J. Approx.
Theory 60, 141–156.
O. V. Matveev (1997), ‘On a method for interpolating functions on chaotic nets’,
Math. Notes 62, 339–349. Translated from Mat. Zametki (1997) 62, 404–417.
C. A. Micchelli (1986), ‘Interpolation of scattered data: distance matrices and

conditionally positive definite functions’, Constr. Approx. 1, 11–22.
J. K. Misiewicz and D. St. P. Richards (1994), ‘Positivity of integrals of Bessel
functions’, SIAM J. Math. Anal. 25, 596–601.
F. J. Narcowich, R. Schaback and J. D. Ward (1999), Multilevel interpolation and
approximation, Appl. Comput. Harm. Analysis 7, 243–261.
R. Pollandt (1997), ‘Solving nonlinear equations of mechanics with the boundary
element method and radial basis functions’, Internat. J. Numer. Methods En-
grg 40, 61–73.
M. J. D. Powell (1992a), ‘The theory of radial basis function approximation in 1990’,
in Advances in Numerical Analysis II: Wavelets, Subdivision, and Radial Fun-
ctions (W. A. Light, ed.), Oxford University Press, Oxford, pp. 105–210.
M. J. D. Powell (1992b), ‘Tabulation of thin plate splines on a very fine two-dimen-
sional grid’, in Numerical Methods of Approximation Theory (D. Braess and
L. L. Schumaker, eds), Birkhäuser, Basel, pp. 221–244.
M. J. D. Powell (1993), ‘Truncated Laurent expansions for the fast evaluation of
thin plate splines’, Numer. Alg. 5, 99–120.
M. J. D. Powell (1994), ‘The uniform convergence of thin-plate spline interpolation
in two dimensions’, Numer. Math. 67, 107–128.
M. J. D. Powell (1999), ‘Recent research at Cambridge on radial basis functions’,
in New Developments in Approximation Theory (M. D. Buhmann, M. Felten,
D. Mache and M. W. Müller, eds), Birkhäuser, Basel, pp. 215–232.
A. Ron and X. Sun (1996), ‘Strictly positive definite functions on spheres in Euc-
lidean spaces’, Math. Comput. 65, 1513–1530.
S. Saitoh (1988), Theory of Reproducing Kernels and its Applications, Longman,
Harlow.
R. Schaback and H. Wendland (1998), ‘Inverse and saturation theorems for radial
basis function interpolation’, Technical Report, University of Göttingen.
R. Schaback and Z. Wu (1995), ‘Operators on radial basis functions’, J. Comput.
Appl. Math. 73, 257–270.
I. J. Schoenberg (1942), ‘Positive definite functions on spheres’, Duke Math. J. 9,
96–108.
R. Sibson and G. Stone (1991), ‘Computation of thin plate splines’, SIAM J. Sci.
Statist. Comput. 12, 1304–1313.
E. M. Stein and G. Weiss (1971), Introduction to Fourier Analysis on Euclidean
Spaces, Princeton University Press, Princeton.
H. Wendland (1995), ‘Piecewise polynomial, positive definite and compactly sup-
ported radial functions of minimal degree’, Adv. Comput. Math. 4, 389–396.
H. Wendland (1998), ‘Error estimates for interpolation by radial basis functions of
minimal degree’, J. Approx. Theory 93, 258–272.
H. Wendland (1997), ‘Sobolev-type error estimates for interpolation by radial basis
functions’, in Surface Fitting and Multiresolution Methods (A. LeMéhauté and
L. L. Schumaker, eds), Vanderbilt University Press, Nashville, pp. 337–344.
Z. Wu (1992), ‘Hermite–Birkhoff interpolation of scattered data by radial basis
functions’, Approx. Theory Appl. 8, 1–10.
Z. Wu (1995), ‘Multivariate compactly supported positive definite radial functions’,
Adv. Comput. Math. 4, 283–292.
38 M. D. Buhmann
Z. Wu and R. Schaback (1993), ‘Local error estimates for radial basis function
interpolation of scattered data’, IMA J. Numer. Anal. 13, 13–27.
Y. Xu and E. W. Cheney (1992), ‘Strictly positive definite functions on spheres’,
Proc. Amer. Math. Soc. 116, 977–981.

Radial Basis Functions by Buhmann

Uploaded by

Copyright:

Available Formats

Radial Basis Functions by Buhmann

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Radial Basis Functions by Buhmann

Uploaded by

Copyright:

Available Formats

Acta Numerica (2000), pp.

1–38 c Cambridge University Press, 2000

Radial basis functions

Radial basis function methods are modern ways to approximate multivariate

we resort solely to tensor product methods. In fact, sometimes we are given

In order to retain uniqueness, Ξ has to contain a Pk−1

p > 2, including p = ∞, or p = 1, singularity can occur in any dimension

functions. Some possible applications, for instance, for numerical solutions

2.1. Improvements and extensions to Duchon’s convergence theorems

cannot, for h → 0, be replaced by o(1). In order to state the result we let

and Φ is the so-called power functional

meaning, that is, Pk−1

2.2. Upper bounds on approximation orders and inverse theorems;

This extension of the results on Ξ = hZn to Ξ = Ω∩hZn relies on the locality

enjoys the remarkable property of the fast (exponential) decay of |L(x)|

Theorem 5. Let Ω, Ξ = Ω ∩ hZn and the interpolant s be as above, and

ks − f k∞,Ω̃ ≤ Ch2k , 0 < h < 1.

integrals being over Rn unless stated otherwise.

Theorem 6. Let 1 ≤ p ≤ ∞. Let φˆh : Rn → R be measurable and such

radial basis function with positive distributional Fourier transform φ̂(k · k) :

3.1. Wendland’s functions

where h is given by (2.2).

in a form proposed by Fasshauer.

3.2. Further contributions

Theorem 11. Let 0 < δ ≤ 12 , ρ ≥ 1 be reals, and suppose λ 6= 0 and

has a positive Fourier transform and therefore, by Bochner’s theorem, gives

to f ∈ L2 (Rn ) ∩ D−n/2−1−α L2 (Rn ), with the interpolation conditions s|Ξ =

functions (3.3), it gives

for δ = 12 and any integral nonnegative λ. Therefore, by one further applic-

4. Iterative methods for implementation

(Beatson, Goodsell and Powell 1995) is iterative and depends on an initial

4.1. The BGP method

instead of (1.1), where each Lagrange function Lξ satisfies the Lagrange

problem that resembles a domain decomposition approach, because we di-

We start with s = s0 = 0. Each full sweep of the algorithm replaces sj

Next, we have to consider the alteration to ks∗ − sj−1,m kφ as soon as σ is

for all ζ ∈ Ξ. This implies sj → s∗ for j → ∞, as demanded, because

4.2. Fast multipole methods

which is the same as

where bk = −t̄ak and a0 = −t and ak = tk+1 /[k(k + 1)] for positive k.

splines at a centre and the current approximation to the solution we re-

on the sphere in P`n (S n−1 ) ∩ P`−1

monics because they are the restrictions of polynomials of total degree ` to

whose coefficients fˆ`k satisfy certain square summability conditions with

approaches are used, for instance because of non-smooth domain boundar-

Lu(x) = f (x), x ∈ Ω ⊂ Rn , (6.1)

where Ω is a domain with suitably smooth – at least Lipschitz-continuous –

The approximants uΞ are defined by the sums

The Λξ and Λζ are suitable discrete functionals to describe our operators L

holds, and set

so that ∆g = u2 everywhere with no boundary conditions. Therefore we

6.2. Other applications

R. K. Beatson and W. A. Light (1997), ‘Fast evaluation of radial basis functions:

G. Fasshauer (1999), ‘Solving differential equations with radial basis functions:

C. A. Micchelli (1986), ‘Interpolation of scattered data: distance matrices and

You might also like