Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

MATH 590: Meshfree Methods

Chapter 7: Conditionally Positive Definite Functions

Greg Fasshauer

Department of Applied Mathematics


Illinois Institute of Technology

Fall 2010

fasshauer@iit.edu MATH 590 Chapter 7 1


Outline

1 Conditionally Positive Definite Functions Defined

2 CPD Functions and Generalized Fourier Transforms

fasshauer@iit.edu MATH 590 Chapter 7 2


Conditionally Positive Definite Functions Defined

In this chapter we generalize positive definite functions to


conditionally positive definite and strictly conditionally positive
definite functions of order m.
These functions provide a natural generalization of RBF
interpolation with polynomial reproduction.
Examples of strictly conditionally positive definite (radial) functions
are given in the next chapter.

fasshauer@iit.edu MATH 590 Chapter 7 4


Conditionally Positive Definite Functions Defined

Definition
A complex-valued continuous function is called conditionally positive
definite of order m on Rs if
N X
X N
cj ck (x j x k ) 0 (1)
j=1 k =1

for any N pairwise distinct points x 1 , . . . , x N Rs , and


c = [c1 , . . . , cN ]T CN satisfying
N
X
cj p(x j ) = 0,
j=1

for any complex-valued polynomial p of degree at most m 1.


The function is called strictly conditionally positive definite of order m
on Rs if the quadratic form (1) is zero only for c 0.

fasshauer@iit.edu MATH 590 Chapter 7 5


Conditionally Positive Definite Functions Defined

An immediate observation is
Lemma
A function that is (strictly) conditionally positive definite of order m on
Rs is also (strictly) conditionally positive definite of any higher order. In
particular, a (strictly) positive definite function is always (strictly)
conditionally positive definite of any order.

Proof.
The first statement follows immediately from the definition.
The second statement is true since (by convention) the case
m = 0 yields the class of (strictly) positive definite functions, i.e.,
(strictly) conditionally positive definite functions of order zero are
(strictly) positive definite.

fasshauer@iit.edu MATH 590 Chapter 7 6


Conditionally Positive Definite Functions Defined

As for positive definite functions we also have (see [Wendland (2005a)]


for more details)
Theorem
A real-valued continuous even function is called conditionally
positive definite of order m on Rs if
N X
X N
cj ck (x j x k ) 0 (2)
j=1 k =1
for any N pairwise distinct points x 1 , . . . , x N Rs , and
c = [c1 , . . . , cN ]T RN satisfying
N
X
cj p(x j ) = 0,
j=1
for any real-valued polynomial p of degree at most m 1.
The function is called strictly conditionally positive definite of order m
on Rs is zero only for c 0.
fasshauer@iit.edu MATH 590 Chapter 7 7
Conditionally Positive Definite Functions Defined

Remark
If the function is strictly conditionally positive definite of order m,
then the matrix A with entries Ajk = (x j x k ) can be interpreted as
being positive definite on the space of vectors c such that
N
X
cj p(x j ) = 0, p sm1 .
j=1

In this sense A is positive definite on the space of vectors c


perpendicular to s-variate polynomials of degree at most m 1.

fasshauer@iit.edu MATH 590 Chapter 7 8


Conditionally Positive Definite Functions Defined

We can now generalize the theorem we had in the previous chapter for
constant precision interpolation to the case of general polynomial
reproduction:
Theorem
If the real-valued even function is strictly conditionally positive
definite of order m on Rs and the points x 1 , . . . , x N form an
(m 1)-unisolvent set, then the system of linear equations
    
A P c y
= (3)
PT O d 0

is uniquely solvable.

fasshauer@iit.edu MATH 590 Chapter 7 9


Conditionally Positive Definite Functions Defined

Proof

The proof is almost identical to the proof of the earlier theorem for
constant reproduction.
Assume [c, d]T is a solution of the homogeneous linear system, i.e.,
with y = 0.
We show that [c, d]T = 0 is the only possible solution.
Multiplication of the top block of (3) by c T yields

c T Ac + c T Pd = 0.

From the bottom block of the system we know PT c = 0. This implies


c T P = 0T , and therefore
c T Ac = 0. (4)

fasshauer@iit.edu MATH 590 Chapter 7 10


Conditionally Positive Definite Functions Defined

Since the function is strictly conditionally positive definite of order m


by assumption we know that the above quadratic form of A (with
coefficients such that PT c = 0) is zero only for c = 0.

Therefore (4) tells us that c = 0.

The unisolvency of the data sites, i.e., the linear independence of the
columns of P (c.f. one of our earlier remarks), and the fact that c = 0
guarantee d = 0 from the top block

Ac + Pd = 0

of the homogeneous version of (3). 

fasshauer@iit.edu MATH 590 Chapter 7 11


CPD Functions and Generalized Fourier Transforms

As before, integral characterizations help us identify functions that are


strictly conditionally positive definite of order m on Rs .

An integral characterization of conditionally positive definite functions


of order m, i.e., a generalization of Bochners theorem, can be found in
the paper [Sun (1993b)].

However, since the subject matter is rather complicated, and since it


does not really help us solve the scattered data interpolation problem,
we do not mention any details here.

fasshauer@iit.edu MATH 590 Chapter 7 13


CPD Functions and Generalized Fourier Transforms The Schwartz Space and the Generalized Fourier Transform

In order to formulate the Fourier transform characterization of strictly


conditionally positive definite functions of order m on Rs we require
some advanced tools from analysis (see Appendix B).

First we define the Schwartz space of rapidly decreasing test functions

S = { C (Rs ) : lim x (D )(x) = 0, , Ns0 },


kxk

where we use the multi-index notation


s
|| X
D = , || = i .
x11 xss i=1

fasshauer@iit.edu MATH 590 Chapter 7 14


CPD Functions and Generalized Fourier Transforms The Schwartz Space and the Generalized Fourier Transform

Some properties of the Schwartz space

S consists of all those functions C (Rs ) which, together with


all their derivatives, decay faster than any power of 1/kxk.
S contains the space C0 (Rs ), the space of all infinitely
differentiable functions on Rs with compact support.
C0 (Rs ) is a true subspace of S since, e.g., the function
2
(x) = ekxk belongs to S but not to C0 (Rs ).
S has a classical Fourier transform which is also in S.

fasshauer@iit.edu MATH 590 Chapter 7 15


CPD Functions and Generalized Fourier Transforms The Schwartz Space and the Generalized Fourier Transform

Of particular importance are the following subspaces Sm of S

Sm = { S : (x) = O(kxkm ) for kxk 0, m N0 }.

Furthermore, the set V of slowly increasing functions is given by

V = {f C(Rs ) : |f (x)| |p(x)| for some polynomial p s }.

fasshauer@iit.edu MATH 590 Chapter 7 16


CPD Functions and Generalized Fourier Transforms The Schwartz Space and the Generalized Fourier Transform

The generalized Fourier transform is now given by


Definition
Let f V be complex-valued. A continuous function f : Rs \ {0} C is
called the generalized Fourier transform of f if there exists an integer
m N0 such that
Z Z
f (x)
(x)dx = f (x)(x)dx
Rs Rs

is satisfied for all S2m .


The smallest such integer m is called the order of f .

Remark
Various definitions of the generalized Fourier transform exist in the
literature (see, e.g., [Gelfand and Vilenkin (1964)]).

fasshauer@iit.edu MATH 590 Chapter 7 17


CPD Functions and Generalized Fourier Transforms The Schwartz Space and the Generalized Fourier Transform

Since one can show that the generalized Fourier transform of an


s-variate polynomial of degree at most 2m is zero, it follows that
the inverse generalized Fourier transform is only unique up to
addition of such a polynomial.
The order of the generalized Fourier transform is nothing but the
order of the singularity at the origin of the generalized Fourier
transform.
For functions in L1 (Rs ) the generalized Fourier transform
coincides with the classical Fourier transform.
For functions in L2 (Rs ) it coincides with the distributional Fourier
transform.

fasshauer@iit.edu MATH 590 Chapter 7 18


CPD Functions and Generalized Fourier Transforms Fourier Transform Characterization

This general approach originated in the manuscript


[Madych and Nelson (1983)]. Many more details can be found in the
original literature as well as in [Wendland (2005a)].
The following result is due to [Iske (1994)].
Theorem
Suppose the complex-valued function V possesses a generalized
Fourier transform of order m which is continuous on Rs \ {0}. Then
is
is strictly conditionally positive definite of order m if and only if
non-negative and non-vanishing.

Remark
This theorem states that strictly conditionally positive definite functions
on Rs are characterized by the order of the singularity of their
generalized Fourier transform at the origin, provided that this
generalized Fourier transform is non-negative and non-zero.

fasshauer@iit.edu MATH 590 Chapter 7 19


CPD Functions and Generalized Fourier Transforms Fourier Transform Characterization

Since integral characterizations similar to our earlier theorems of


Schoenberg for positive definite radial functions are so complicated in
the conditionally positive definite case we do not pursue the concept of
a conditionally positive definite radial function here.
Such theorems can be found in [Guo et al. (1993a)].

Examples of radial functions via the Fourier transform approach


are given in the next chapter.
In Chapter 9 we will explore the connection between completely
and multiply monotone functions and conditionally positive definite
radial functions.

fasshauer@iit.edu MATH 590 Chapter 7 20


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.
Gelfand, I. M. and Vilenkin, N. Ya. (1964).
Generalized Functions Vol. 4.
Academic Press (New York).
Iske, A. (2004).
Multiresolution Methods in Scattered Data Modelling.
Lecture Notes in Computational Science and Engineering 37, Springer Verlag
(Berlin).
Wendland, H. (2005a).
Scattered Data Approximation.
Cambridge University Press (Cambridge).

fasshauer@iit.edu MATH 590 Chapter 7 21


Appendix References

References II

Guo, K., Hu, S. and Sun, X. (1993a).


Conditionally positive definite functions and Laplace-Stieltjes integrals.
J. Approx. Theory 74, pp. 249265.
Iske, A. (1994).
Charakterisierung bedingt positiv definiter Funktionen fr multivariate
Interpolationsmethoden mit radial Basisfunktionen.
Ph.D. Dissertation, Universitt Gttingen.
Madych, W. R. and Nelson, S. A. (1983).
Multivariate interpolation: a variational theory.
manuscript.
Sun, X. (1993b).
Conditionally positive definite functions and their application to multivariate
interpolation.
J. Approx. Theory 74, pp. 159180.

fasshauer@iit.edu MATH 590 Chapter 7 22

You might also like