Norm estimates for the inverses of a general class of scattered-data radial-function interpolation matrices (Q1185257)

The class \({\mathcal RN}^ \infty_{m,c}\) of completely monotonic conditionally negative definite radial functions \(F\) of order \(m\) (cf. the authors [J. Approximation Theory 64, No. 1, 69-94 (1991; Zbl 0724.41004)]) is considered. The main tool is a given lower estimate on the quadratic form \(Q\) associated with \(F\) and scattered data \(\{x_ j\}^ N_ 1\subset{\mathbf R}^ s\). Norm estimates for the inverses of interpolation matrices of the functions \(F\in{\mathcal RN}^ s_ m\), \(m=0,1\), wich have the Borel measure \(d\alpha(u)=u^{s+1}\rho(u)du\) in their standard integral representation (cf. loc. cit.) are given. Here \(\rho(u)\) is strictly positive, decreasing, continuous on \((0,\infty)\). Similar estimates are given for the data set \(\{x_ j\}\) which contains a \(\pi_{m-1}({\mathbf R}^ s)\) unisolvent set. There are provided two ways to select a polynomial \(p\in\pi_{m-1}({\mathbf R}^ s)\) sucht hat \((F(r)+p(r^ 2))\in{\mathcal RN}^ \infty_{m,c}\) yields an invertible interpolation matrix. Estimates on inverses of such matrices are obtained. The results apply to the thin-plate splines. For a survey of the problems considered here see \textit{N. Dyn} [Approximation Theory VI, Proc. 6th Int. Symp., College Station/TX (USA) 1989, Vol. I, 211-234 (1989; Zbl 0705.41006)].