Gaussian radial basis function interpolant for the different data sites and basis centers (Q2363677): Difference between revisions

Radial basis function interpolation (and, more generally, approximation from the linear function spaces generated by them) is one of the fundamental approximation tools whenever functions are to be approximated in numerical analysis, especially in (arbitrarily) high base dimensions. The methods depend on mostly interpolating a provided set of function values by a linear combination of shifts of a radially symmetric function (the ``radial basis function''), the shifts usually being the same points in multiple dimensions as the points at which we wish to interpolate. This leads to a symmetric matrix that is often non-singular for a large class of radial basis functions (that as such, are univariate, as the radial part of a radially symmetric function); more specifically, the eigenstructure of such matrices is well-understood, leading to e.g.\ positive definite matrices, conditionally positive definite functions (giving matrices which are positive definite on a large space with small co-dimension). The present article analyses the numerical properties of such approximation schemes for a specific choice of radial basis function, i.e. the Gauß-kernel. The authors propose numerical methods to choose the aforementioned data points and study again the non-singularity of the interpolation matrix.