MATH 590: Meshfree Methods

Chapter 11: Compactly Supported Radial Basis Functions

Greg Fasshauer

Department of Applied Mathematics

Illinois Institute of Technology

Fall 2010

fasshauer@iit.edu MATH 590 Chapter 11 1

Outline

1 Operators for Radial Functions and Dimension Walks

2 Wendlands Compactly Supported Functions

3 Wus Compactly Supported Functions

4 Oscillatory Compactly Supported Functions

5 Other Compactly Supported Radial Basis Functions

fasshauer@iit.edu MATH 590 Chapter 11 2

As we saw earlier, compactly supported functions that are truly
strictly conditionally positive definite of order m > 0 do not exist.
The compact support automatically ensures that is strictly positive
definite.

Another observation was that compactly supported radial functions

can be strictly positive definite on Rs only for a fixed maximal s-value.
It is not possible for a function to be strictly positive definite and radial
on Rs for all s and also have a compact support.

Therefore we focus our attention on the characterization and

construction of functions that are compactly supported, strictly positive
definite and radial on Rs for some fixed s.

fasshauer@iit.edu MATH 590 Chapter 11 3

According to Bochners theorem and generalizations thereof, a
function is strictly positive definite and radial on Rs if its s-variate
Fourier transform is non-negative.
From Appendix B:
Theorem
The Fourier transform of the radial function = (k k) is given by
another radial function
Z
1 s
(x) = Fs (kxk) = p (t)t 2 J s2 (tkxk)dt,
kxks2 0 2

where J is the Bessel function of the first kind of order .

Remark
A proof of this theorem can be found in [Wendland (2005a)].
This integral transform is also referred to as Fourier-Bessel
transform or Hankel transform.
fasshauer@iit.edu MATH 590 Chapter 11 4
Remark
The Hankel inversion theorem [Sneddon (1972)] ensures that the
Fourier transform for radial functions is its own inverse, i.e., for
radial functions we have

Fs [Fs ] = .

We used this earlier when we turned the Matrn functions upside

down to get the generalized inverse multiquadrics.

fasshauer@iit.edu MATH 590 Chapter 11 5

Operators for Radial Functions and Dimension Walks

A certain integral operator and its inverse differential operator

were defined in [Schaback and Wu (1996)].
In that paper an entire calculus was developed for how these
operators act on radial functions.

According to [Gneiting (2002)], these operators can be traced

back to [Matheron (1965)] who called the integral operator monte
and the differential operator descente motivated by an application
related to mining.

In the following we define these operators and show how they

facilitate the construction of compactly supported radial functions.

fasshauer@iit.edu MATH 590 Chapter 11 7

 Operators for Radial Functions and Dimension Walks

Definition
1 Let be such that t 7 t(t) L1 [0, ). Then we define the
integral operator I via
Z
(I)(r ) = t(t)dt, r 0.
r

2 For even C 2 (R) we define the differential operator D via

1
(D)(r ) = 0 (r ), r 0.
r
In both cases the resulting functions are to be interpreted as even
functions using even extensions.

Remark
Note that the integral operator I differs from the operator I introduced
earlier by a factor t in the integrand.
fasshauer@iit.edu MATH 590 Chapter 11 8
 Operators for Radial Functions and Dimension Walks

The most important properties of the monte and descente operators

are (see, e.g., [Schaback and Wu (1996)] or [Wendland (1995)]):
Theorem
(1) Both D and I preserve compact support, i.e., if has compact
support, then so do D and I.
(2) If C(R) and t 7 t(t) L1 [0, ), then DI = .
(3) If C 2 (R) ( 6 1) is even and 0 L1 [0, ), then ID = .
(4) If t 7 t s1 (t) L1 [0, ) and s 3, then Fs () = Fs2 (I).
(5) If C 2 (R) is even and t 7 t s 0 (t) L1 [0, ), then
Fs () = Fs+2 (D).

fasshauer@iit.edu MATH 590 Chapter 11 9

 Operators for Radial Functions and Dimension Walks

The operators I and D allow us to express s-variate Fourier transforms

as (s 2)- or (s + 2)-variate Fourier transforms, respectively.
In particular, a direct consequence of the above properties and the
characterization of strictly positive definite radial functions is
Theorem
(1) Suppose C(R). If t 7 t s1 (t) L1 [0, ) and s 3, then
is strictly positive definite and radial on Rs if and only if I is
strictly positive definite and radial on Rs2 .

(2) If C 2 (R) is even and t 7 t s 0 (t) L1 [0, ), then is strictly

positive definite and radial on Rs if and only if D is strictly
positive definite and radial on Rs+2 .

fasshauer@iit.edu MATH 590 Chapter 11 10

 Operators for Radial Functions and Dimension Walks

Remark
This allows us to construct new strictly positive definite radial
functions from given ones by a dimension-walk technique that
steps through multivariate Euclidean space in even increments.

The examples presented in the following sections illustrate this

technique.

fasshauer@iit.edu MATH 590 Chapter 11 11

 Wendlands Compactly Supported Functions

Probably the most popular family of compactly supported radial

functions presently in use was constructed in [Wendland (1995)].

Wendland starts with the truncated power function

` (r ) = (1 r )`+ .

which we know to be strictly positive definite and radial on Rs for

integer ` b 2s c + 1.

Then he walks through dimensions by repeatedly applying the integral

operator I.

fasshauer@iit.edu MATH 590 Chapter 11 13

 Wendlands Compactly Supported Functions

Definition
With ` (r ) = (1 r )`+ we define

s,k = I k bs/2c+k +1 .

Remark
Note the use of a single subscript for the truncated power function,
and double subscript for the Wendland functions.
It turns out that the functions s,k are all supported on [0, 1] and
have a polynomial representation there.

fasshauer@iit.edu MATH 590 Chapter 11 14

 Wendlands Compactly Supported Functions

More precisely,
Theorem
The functions s,k are strictly positive definite and radial on Rs and are
of the form 
ps,k (r ), r [0, 1],
s,k (r ) =
0, r > 1,
with a univariate polynomial ps,k of degree bs/2c + 3k + 1.

Moreover, s,k C 2k (R) are unique up to a constant factor, and the

polynomial degree is minimal for given space dimension s and
smoothness 2k .

fasshauer@iit.edu MATH 590 Chapter 11 15

 Wendlands Compactly Supported Functions

Remark
This theorem states that any other compactly supported
polynomial function that globally C 2k and strictly positive definite
and radial on Rs will not have a smaller polynomial degree.
Our other examples below (Wus functions, Gneitings functions)
illustrate this fact.
The strict positive definiteness of Wendlands functions s,k
starting with non-integer values of ` was established in
[Gneiting (1999)].
Note, however, that then the functions are no longer guaranteed to
be polynomials on their support.

fasshauer@iit.edu MATH 590 Chapter 11 16

 Wendlands Compactly Supported Functions

Wendland gave recursive formulas for the functions s,k for all s, k .

We instead list the explicit formulas of [Fasshauer (1999a)].

Theorem
The functions s,k , k = 0, 1, 2, 3, have the form

s,0 (r ) = (1 r )`+ ,
.
s,1 (r ) = (1 r )`+1
+ h [(` + 1)r + 1] ,
.
i
s,2 (r ) = (1 r )`+2
+ (`2 + 4` + 3)r 2 + (3` + 6)r + 3 ,
.
h
s,3 (r ) = (1 r )`+3
+ (`3 + 9`2 + 23` + 15)r 3 + (6`2 + 36` + 45)r 2
+(15` + 45)r + 15] ,
.
where ` = bs/2c + k + 1, and the symbol = denotes equality up to a
multiplicative positive constant.

fasshauer@iit.edu MATH 590 Chapter 11 17

 Wendlands Compactly Supported Functions

Proof.
The case k = 0 follows directly from the definition.
Application of the definition for the case k = 1 yields
Z
s,1 (r ) = (I` )(r ) = t` (t)dt
Z r

= t(1 t)`+ dt
Zr 1
= t(1 t)` dt
r
1
= (1 r )`+1 [(` + 1)r + 1] ,
(` + 1)(` + 2)

where the compact support of ` reduces the improper integral to a

definite integral which can be evaluated using integration by parts.
The other two cases are obtained similarly by repeated application of
I.

fasshauer@iit.edu MATH 590 Chapter 11 18

 Wendlands Compactly Supported Functions

Example

k 3,k (r ) smoothness

0 (1 r )2+ C0
1 (1 r )4+ (4r + 1) C2
r )6+ 2
C4


2 (1 35r + 18r + 3
(1 r )8+ 32r 3 + 25r 2 + 8r + 1 C6


3

Table: Wendlands compactly supported radial functions s,k for various

choices of k and s = 3.

fasshauer@iit.edu MATH 590 Chapter 11 19

 Wendlands Compactly Supported Functions

Example (cont.)
All functions in the table are strictly positive definite and radial on
Rs for s 3.
Their degree of smoothness 2k is specified.
The functions were determined using the direct formulas from the
above theorem and thus match the definition only up to a positive
constant factor.
Note that (x)`+ is to be interpreted as ((x)+ )` , i.e., we first apply
the cutoff function, and then the power.

fasshauer@iit.edu MATH 590 Chapter 11 20

 Wendlands Compactly Supported Functions

Figure: Plot of Wendlands functions s,k for various choices of k and s = 3.

fasshauer@iit.edu MATH 590 Chapter 11 21

 Wendlands Compactly Supported Functions

For the M ATLAB implementation in the next chapter it is better to

express the compactly supported functions in a shifted form.
es,k = s,k (1 ) so that
We list the appropriate functions
es,k (1 r ) = s,k (r ).

k 3,k (r )
e3,k (r )

0 (1 r )2+ r+2
1 (1 r )4+ (4r + 1) r+4 (5 4r )
r )6+ 2
r+6 56 88r + 35r 2



2 (1 35r + 18r + 3
(1 r )8+ 32r 3 + 25r 2 + 8r + 1 r+8 66 154r + 121r 2 32r 3



3

Table: Wendlands compactly supported radial functions s,k and

es,k = s,k (1 ) for various choices of k and s = 3.

fasshauer@iit.edu MATH 590 Chapter 11 22

 Wus Compactly Supported Functions

In [Wu (1995b)] we find another way to construct strictly positive

definite radial functions with compact support.
Wu starts with the function
(r ) = (1 r 2 )`+ , ` N,
which in itself is not positive definite (see the discussion at the end of
Chapter 5).
However, Wu then uses convolution to construct another function that
is strictly positive definite and radial on R, i.e.,

Z )(2r )
` (r ) = (
= (1 t 2 )`+ (1 (2r t)2 )`+ dt

Z 1
= (1 t 2 )` (1 (2r t)2 )`+ dt.
1
Remark
This function is strictly positive definite since its Fourier transform is
essentially the square of the Fourier transform of and therefore
non-negative.
fasshauer@iit.edu MATH 590 Chapter 11 24
 Wus Compactly Supported Functions

Just like the Wendland functions, the Wu function ` is a

polynomial on its support.
The degree of the polynomial is 4` + 1, and ` C 2` (R).
Now, a family of strictly positive definite radial functions is
constructed by a dimension walk using the differential operator D.

Definition
With ` (r ) = ((1 2 )`+ (1 2 )`+ )(2r ) we define

k ,` = Dk ` .

fasshauer@iit.edu MATH 590 Chapter 11 25

 Wus Compactly Supported Functions

The functions k ,`

are strictly positive definite and radial on Rs for s 2k + 1,

are polynomials of degree 4` 2k + 1 on their support

and in C 2(`k ) in the interior of the support.

On the boundary the smoothness increases to C 2`k .

fasshauer@iit.edu MATH 590 Chapter 11 26

 Wus Compactly Supported Functions

Example

For ` = 3 we can compute the four functions

k ,3 (r ) = Dk 3 (r ) = Dk ((1 2 )3+ (1 2 )3+ )(2r ), k = 0, 1, 2, 3.

They are listed on the next slide along with their smoothness.

The maximal space dimension s for which these functions are strictly
positive definite and radial on Rs is also listed.

Just as with the Wendland functions, the functions in Table 12 match

the definition only up to a positive multiplicative constant.

fasshauer@iit.edu MATH 590 Chapter 11 27

 Wus Compactly Supported Functions

Example (cont.)

k k ,3 (r ) smoothness s

0 (1 r )7+ (5 + 35r + 101r 2 + 147r 3 + 101r 4 + 35r 5 + 5r 6 ) C6 1

1 (1 r )6+ (6 + 36r + 82r 2 + 72r 3 + 30r 4 + 5r 5 ) C4 3
2 (1 r )5+ (8 + 40r + 48r 2 + 25r 3 + 5r 4 ) C2 5
3 (1 r )4+ (16 + 29r + 20r 2 + 5r 3 ) C0 7

Table: Wus compactly supported radial functions k ,` for various choices of k

and ` = 3.

fasshauer@iit.edu MATH 590 Chapter 11 28

 Wus Compactly Supported Functions

Figure: Plot of Wus functions k ,` for various choices of k and ` = 3.

fasshauer@iit.edu MATH 590 Chapter 11 29

 Wus Compactly Supported Functions

Example (cont.)
Again, we also list the functions ek ,` = k ,` (1 ) used in our M ATLAB
implementation in Chapter 12.

k
ek ,3 (r ) smoothness s

0 7 (429 1287r + 1573r 2 1001r 3 + 351r 4 65r 5 + 5r 6 )

r+ C6 1
1 6 (231 561r + 528r 2 242r 3 + 55r 4 5r 5 )
r+ C4 3
2 5 (126 231r + 153r 2 45r 3 + 5r 4 )
r+ C2 5
3 4 (70 84r + 35r 2 5r 3 )
r+ C0 7

Table: Shifted version ek ,` of Wus compactly supported radial functions k ,`

for various choices of k and ` = 3.

fasshauer@iit.edu MATH 590 Chapter 11 30

 Wus Compactly Supported Functions

Remark
As predicted by the theorem about the Wendland functions , for a prescribed space
dimension s and smoothness the polynomial degree of
Wendlands functions is lower than that of Wus functions.
For example, both Wendlands function 3,2 and Wus function 1,3
are C 4 smooth and strictly positive definite and radial on R3 .
However, the polynomial degree of Wendlands function is 8,
whereas that of Wus function is 11.
Another comparable function is Gneitings oscillatory function 2
(see below), which is a C 4 polynomial of degree 9 that is strictly
positive definite and radial on R3 .

While the two families of strictly positive definite compactly

supported functions discussed above are both constructed via
dimension walk, Wendland uses integration (and thus obtains a
family of increasingly smoother functions), whereas Wu needs to
start with a function of sufficient smoothness, and then obtains
successively less smooth functions (via differentiation).
fasshauer@iit.edu MATH 590 Chapter 11 31
 Oscillatory Compactly Supported Functions

Other strictly positive definite compactly supported radial functions

have been proposed by Gneiting (see, e.g., [Gneiting (2002)]).

He showed that a family of oscillatory compactly supported functions

can be constructed using the so-called turning bands operator of
[Matheron (1973)].

Starting with a function s that is strictly positive definite and radial on

Rs for s 3 the turning bands operator produces

r 0s (r )
s2 (r ) = s (r ) + (1)
s2

which is strictly positive definite and radial on Rs2 .

fasshauer@iit.edu MATH 590 Chapter 11 33

 Oscillatory Compactly Supported Functions

Example

One such family of functions is generated if we start with the Wendland

functions
s+2,1 (r ) = (1 r )`+1
+ [(` + 1)r + 1]

(` non-integer allowed).

Application of the turning bands operator results in the functions

 
` (` + 1)(` + 2 + s) 2
s,` (r ) = (1 r )+ 1 + `r r ,
s
s+5
which are strictly positive definite and radial on Rs provided ` 2
(see [Gneiting (2002)]).

fasshauer@iit.edu MATH 590 Chapter 11 34

 Oscillatory Compactly Supported Functions

Example (cont.)

` 2,` (r ) smoothness

7/2
1 + 72 r 135 2
C2


7/2 (1 r )+ 8 r
(1 r )5+ 1 + 5r 27r 2
C2


5
15/2 15 391 2
C2


15/2 (1 r )+ 1+ 2 r 8 r
(1 r )12 2
C2


12 + 1 + 12r 104r

Table: Gneitings compactly supported radial functions s,` for various choices
of ` and s = 2.

Remark
All functions are in C 2 (R).
If we want smoother functions, then we need to start with a
smoother Wendland family as described in the next example.
fasshauer@iit.edu MATH 590 Chapter 11 35
 Oscillatory Compactly Supported Functions

Figure: First family of oscillatory functions by Gneiting.

fasshauer@iit.edu MATH 590 Chapter 11 36

 Oscillatory Compactly Supported Functions

Example

We get a set of oscillatory functions that are strictly positive definite

and radial on R3 by applying the turning bands operator to the
Wendland functions 5,k which are strictly positive definite and radial
on R5 for different choices of k .
Then the resulting functions k will have the same degree of
smoothness 2k as the original functions and they will be strictly
positive definite and radial on R3 .

k k (r ) smoothness

(1 r )4+ 1 + 4r 15r 2 C2


1
(1 r )6+ 3 + 18r + 3r 2 192r 3
C4


2
r )8+ 2 3 4
C6


3 (1 15 + 120r + 210r 840r 3465r

Table: Oscillatory compactly supported functions that are strictly positive

definite and radial on R3 parametrized by smoothness.
fasshauer@iit.edu MATH 590 Chapter 11 37
 Oscillatory Compactly Supported Functions

Figure: Second family of oscillatory functions by Gneiting.

fasshauer@iit.edu MATH 590 Chapter 11 38

 Oscillatory Compactly Supported Functions

Remark
Gneiting also suggests the construction of strictly positive definite
radial functions by taking the product of the (appropriately scaled)
Poisson functions s (see Chapter 4) with a certain compactly
supported non-negative function (see [Gneiting (2002)] for more
details).

Since the product of strictly positive definite functions is strictly positive

definite (see Chapter 3) the resulting function will be strictly positive
definite.

This will also yield oscillatory compactly supported functions.

fasshauer@iit.edu MATH 590 Chapter 11 39

 Other Compactly Supported Radial Basis Functions

There are many other ways in which one can construct compactly
supported functions that are strictly positive definite and radial on Rs .
In [Schaback (1995a)] several such possibilities are described.

Euclids hat functions are constructed in analogy to B-splines.

Example

It is well known that the univariate function

(r ) = (1 |r |)+

is a second-order B-spline with knots at 1, 0, 1, and it is obtained as

the convolution of the characteristic function of the interval [ 12 , 12 ] with
itself.

Euclids hat functions are now obtained by convolving the characteristic

function of the s-dimensional Euclidean unit ball with itself.

fasshauer@iit.edu MATH 590 Chapter 11 41

 Other Compactly Supported Radial Basis Functions

Euclids hat
The resulting functions can be written for r [0, 1] in the form
22k 1 (2r )r (1r 2 )k
(
2k +1 k = 1, 2, 3, . . . ,
2k +1 (2r ) =
2(1 r ) k = 0,

for odd space dimensions s = 2k + 1, and as

( 2 k
2
22k (2r )r (1r )(1r )
2k +2 k = 1, 2, 3, . . . ,
2k +2 (2r ) =
2(arccos r r 1 r 2 ) k = 0,

for even space dimensions s = 2k .

Note that these functions are defined to be zero outside the interval
[0, 2].

fasshauer@iit.edu MATH 590 Chapter 11 42

 Other Compactly Supported Radial Basis Functions

Euclids hat (cont.)

In the table below we have employed a substitution 2r r and a
normalization factor such that the functions all have a value of one at
the origin.

s s (r ) smoothness

1 1 2r C0
1

r

C0


2
2 2 4 arccos 2 r 4 r
1
(4 + 16)r r 3 C0


3 1 32
2

arccos 2r 321
4 r 2 20r + r 3 C0



4
1
(12 + 8 + 32 2 )r (3 + 2)r 3
C0


5 1 64 2

Table: Euclids hat functions (defined for 0 r 2) for different values of s.

fasshauer@iit.edu MATH 590 Chapter 11 43

Other Compactly Supported Radial Basis Functions

Figure: Euclids hat functions.

fasshauer@iit.edu MATH 590 Chapter 11 44

 Other Compactly Supported Radial Basis Functions

Remark
Another construction described in [Schaback (1995a)] is the
radialization of the s-fold tensor product of univariate B-splines of
even order 2m with uniform knots.
These functions do not seem to have a simple representation that
lends itself to numerical computations.
As can be seen from its radialized Fourier transform, the
radialized B-spline itself is not strictly positive definite and radial
on any Rs with s > 1.
For s = 1 only the B-splines of even order are strictly positive
definite (see, e.g., [Schlkopf and Smola (2002)]).

fasshauer@iit.edu MATH 590 Chapter 11 45

 Other Compactly Supported Radial Basis Functions

The last family of compactly supported strictly positive definite radial

functions we would like to mention is due to [Buhmann (1998)].

Buhmanns functions contain a logarithmic term in addition to a

polynomial.

His functions have the general form

 
r2
Z
(r ) = 1 t (1 t )+ dt.
0 t +

Here 0 < 21 , 1, and in order to obtain functions that are strictly

positive definite and radial on Rs for s 3 the constraints for the
remaining parameters are 0, and 1 < 1 2 .

fasshauer@iit.edu MATH 590 Chapter 11 46

 Other Compactly Supported Radial Basis Functions

Example

Let = = 21 , = 1 and = 2 (see [Buhmann (2000)]):

.
(r ) = 12r 4 log r 21r 4 + 32r 3 12r 2 + 1, 0 r 1.

This function is in C 2 (R) and strictly positive definite and radial on Rs

for s 3.

fasshauer@iit.edu MATH 590 Chapter 11 47

 Other Compactly Supported Radial Basis Functions

Remark
In [Buhmann (2000)] it is stated that his construction
encompasses both Wendlands and Wus functions.
An even more general theorem that shows that integration of a
positive function f L1 [0, ) against a strictly positive definite
kernel K results in a strictly positive definite function can be found
in [Wendland (2005a)] (see also Chapter 4).
More specifically, Z
(r ) = K (t, r )f (t)dt
0

is strictly positive definite.

Buhmanns construction then corresponds to choosing
 2

f (t) = t (1 t )+ and K (t, r ) = 1 rt .
+

fasshauer@iit.edu MATH 590 Chapter 11 48

 Appendix References

References I

Buhmann, M. D. (2003).
Radial Basis Functions: Theory and Implementations.
Cambridge University Press.
Fasshauer, G. E. (2007).
Meshfree Approximation Methods with M ATLAB.
World Scientific Publishers.
Iske, A. (2004).
Multiresolution Methods in Scattered Data Modelling.
Lecture Notes in Computational Science and Engineering 37, Springer Verlag
(Berlin).
Matheron, G. (1965).
Les variables rgionalises et leur estimation.
Masson (Paris).

fasshauer@iit.edu MATH 590 Chapter 11 49

 Appendix References

References II

Schlkopf, B. and Smola, A. J. (2002).

Learning with Kernels: Support Vector Machines, Regularization, Optimization,
and Beyond.
MIT Press (Cambridge, MA).
Wendland, H. (2005a).
Scattered Data Approximation.
Cambridge University Press (Cambridge).
Buhmann, M. D. (1998).
Radial functions on compact support.
Proc. Edin. Math. Soc. II 41, pp. 33-46.
Buhmann, M. D. (2000).
Radial basis functions.
Acta Numerica 2000 9, pp. 138.

fasshauer@iit.edu MATH 590 Chapter 11 50

 Appendix References

References III

Fasshauer, G. E. (1999a).
On smoothing for multilevel approximation with radial basis functions.
in Approximation Theory IX, Vol.II: Computational Aspects, Charles K. Chui, and
L. L. Schumaker (eds.), Vanderbilt University Press, pp. 5562.
Gneiting, T. (1999).
Correlation functions for atmospheric data analysis.
Quart. J. Meteorol. Soc. 125, pp. 24492464.
Gneiting, T. (2002).
Compactly supported correlation functions.
J. Multivariate Analysis 83, pp. 493508.
Matheron, G. (1973).
The intrinsic random functions and their applications.
Adv. Appl. Prob. 5, pp. 439468.

fasshauer@iit.edu MATH 590 Chapter 11 51

 Appendix References

References IV

Schaback, R. (1995a).
Creating surfaces from scattered data using radial basis functions.
in Mathematical Methods for Curves and Surfaces, M. Dhlen, T. Lyche, and L.
Schumaker (eds.), Vanderbilt University Press (Nashville), pp. 477496.
Schaback, R. and Wu, Z. (1996).
Operators on radial functions.
J. Comput. Appl. Math. 73, pp. 257270.
Sneddon, I. H. (1972).
The Use of Integral Transforms.
McGraw-Hill (New York).
Wendland, H. (1995).
Piecewise polynomial, positive definite and compactly supported radial functions
of minimal degree.
Adv. in Comput. Math. 4, pp. 389396.

fasshauer@iit.edu MATH 590 Chapter 11 52

 Appendix References

References V

Wu, Z. (1995b).
Compactly supported positive definite radial functions.
Adv. in Comput. Math. 4, pp. 283292.

fasshauer@iit.edu MATH 590 Chapter 11 53