Multivariate interpolation by polynomials and radial basis functions (Q2386423)

For a fixed set \(X=\{x_1,\ldots,x_N\}\subset {\mathbb R}^d\) of scattered nodes of interpolation, the behavior of the Lagrange interpolation basis \(u_1^c(x),\ldots,u_M^c(x)\) \((x\in {\mathbb R}^d)\) for \(c\to 0\) is considered. This basis is the solution of the system \[ \sum_{j=1}^M \varphi_c(\| x_j-x_k\| _2)\,u_j^c(x) = \varphi_c(\| x-x_k\| _2) \] for \(k=1,\ldots,M\), where \(\varphi_c(r):=\exp(-c^2r^2)\) is the scaled Gaussian radial basis function. In many cases, the limits of \(u_j^c(x)\) for \(c\to 0\) are multivariate polynomials. It is known that the convergence may depend on the geometry of \(X\) and on the radial basis function [see also \textit{E. Larsson} and \textit{B. Fornberg}, Comput. Math. Appl. 49, No. 1, 103--130 (2005; Zbl 1074.41012)]. In this interesting paper, the author proves that multivariate interpolation by scaled Gaussians converges toward ``least'' polynomial interpolant in the sense of \textit{C. de Boor} and \textit{A. Ron} [Math. Comput. 58, No. 198, 705--727 (1992; Zbl 0767.41003) and Math. Z. 210, No. 3, 347--378 (1992; Zbl 0735.41001)]. First, the author studies the polynomial spaces spanned by linear combinations of \(\| x-x_j\| _2^{2l}\) and introduces a discrete moment basis to define a new well-posed multivariate polynomial interpolation process which is of minimal degree, ``least'' and ``degree-reducing'' in the sense of C. de Boor and A. Ron. Then, the author generalizes the de Boor/Ron interpolation process and shows that it occurs as the limit of multivariate interpolation by scaled Gaussians \(\varphi_c(r)\) for \(c\to 0\). Further, a stable method for preconditioning the matrices arising with interpolation by smooth radial basis functions is proposed.