On interpolation with products of positive definite functions (Q1373131)

The goal of this paper is to solve the following problem of multivariate interpolation. Given points \(x_1, \dots, x_n\in \mathbb{R}^d\), \(d\geq 1\), and \(z_1, \dots, z_n\in \mathbb{R}\); find a continuous function \(f:\mathbb{R}^d \to\mathbb{R}\) satisfying \(f(x_i)= z_i\), \(i=1, \dots,n\). The theory of reproducing kernels is used to obtain general results on interpolation for scattered data, where the interpolating functions are found as solutions of variational problems. To obtain the solution of the above problem the author uses linear combinations of products of positive definite kernel functions \(K_1\), \(K_2\). The main result is the following. Theorem. Let \(S_i\subset \mathbb{R}^d\) and positive definite kernels \(K_i: S_i\otimes S_i\to \mathbb{R}\), \(i=1,2\), be given. Let \(S^*= S_1\otimes S_2\) and assume that \(K:S^* \otimes S^*\to \mathbb{R}\) is defined by \(K(s,t;x,y)= K_1(s,x) K_2(t,y)\) for all \(((s,t),(x,y)) \subset S^*\otimes S^*\). Assume that \(K_i (s,u_1), \dots, K_i(s,u_k)\) are linearly independent for all choices of distinct points \(u_1, \dots, u_k\) in \(S_i\), \(i=1,2\). Let distinct points \((x_i,y_i) \subseteq S^*\), \(i=1, \dots,n\), be given. Then, (a) The linear system \(\sum^n_{i=1} a_iK(x_j,y_j; x_i,y_i)=z_j\), \(j=1,\dots,n\), has a unique solution \((a^*_i)^n_{i=1}\) for all choices of real numbers \(z_1, \dots,z_n\). (b) There exists a Hilbert space \(H\) defined on \(S^*\) with inner product \(( , )\) such that \(K(s,t;x,y)\) is a reproducing kernel of \(H\). Assume that \(h\in H\) is an arbitrary function satisfying \(h(x_i,y_i)= z_i\), \(i=1,2, \dots,n\). Then \(h^*=\sum^n_{i=1} a^*_iK(\cdot; x_i,y_i)\), satisfies \(|h|\geq |h^*|\), where \((a^*_i)^n_{i=1}\) are the solutions of the linear system in (a) and \(|f|= (f,f)^{1\over 2}\). Chebychevian splines and \(L\)-splines are here included in a list of examples of positive definite kernels, and reproducing kernels. The former are used in giving an extension of well-known results for spline interpolation in the univariate case to scattered data interpolation for multivariate splines. In particular, the minimal properties of interpolating natural splines are extended to the multivariate case.