ON THE SENSITIVITY OF RADIAL BASIS INTERPOLATION TO MINIMAL DATA SEPARATION DISTANCE
Academic Article

Overview

Research

Identity

Additional Document Info

View All

Overview

abstract

Motivated by the problem of multivariate scattered data interpolation, much interest has centered on interpolation by functions of the form {Mathematical expression} where g:R + R is some prescribed function. For a wide range of functions g, it is known that the interpolation matrices A=g(x i -x j ) i,j=1N are invertible for given distinct data points x 1 , x 2 ,..., x N . More recently, progress has been made in quantifying these interpolation methods, in the sense of estimating the (l 2 ) norms of the inverses of these interpolation matrices as well as their condition numbers. In particular, given a suitable function g:R + R, and data in R s having minimal separation q, there exists a function h s :R + R + , which depends only on g and s, and a constant C s , which depends only on s, such that the inverse of the associated interpolation matrix A satisfies the estimate {norm of matrix}A -1 {norm of matrix}C s h s (q). The present paper seeks "converse" results to the inequality given above. That is, given a suitable function g, a spatial dimension s, and a parameter q>0 (which is usually assumed to be small), it is shown that there exists a data set in R s having minimal separation q, a constant {Mathematical expression} depending only on s, and a function k s (q), such that the inverse of the interpolation matrix A associated with this data set satisfies {Mathematical expression}. In some cases, it is seen that h s (q)=k s (q), so the bounds are optimal up to constants. In certain others, k s (q) is less than h s (q), but nevertheless exhibits a behavior comparable to that of h s (q). That is, even in these cases, the bounds are close to being optimal. 1992 Springer-Verlag New York Inc.