Convergence properties of radial basis functions (Q1112273)

The multivariate interpolation problem is that of choosing a function s from \(R^ n\) that satisfies the interpolation conditions \(\{s(x_ i)=f_ i:\quad i=1,2,...,m\}\) at m different data points \(x_ i\). Employing the radial basis function \(\phi\) : \(R^+\to R\), means that s is from the linear space spanned by the functions \(\{\phi (\| x-x_ i\|_ 2):\quad x\in R^ n:\quad i=1,2,...,m\},\) except that some low-order polynomials are added and further conditions imposed on s taking up the extra degrees of freedom. Micchelli shows that for some choices of \(\phi\) these conditions uniquely define s. The more basic question of the suitability of the space is considered. Sufficient conditions are given for members of the linear space to converge locally uniformly to a prescribed continuous function on a bounded domain, as the data points \(\{x_ i:\) \(i=1,2,...\}\) become dense. For the case \(\phi (r)=r\) they are found to hold when n is odd but not when n is even.