Convergence of non-stationary semi-discrete RBF schemes for the heat and wave equation (Q6052455)

Approximation in general and interpolation in particular using radial basis functions is a successful and often used technique to perform numerical solutions of partial differential equations and other problems. In particular non-stationary radial basis function approximations (non-stationary kernels being employed) turn out to be extremely useful for solving such problems in numerical analysis. When performing such calculations, a great choice of nodes for the kernels (``centres'') is available, a popular feature of radial basis function being that they are available for scattered centres, but also infinite, equally spaced meshes of points are useful for approximations. In this paper, the so-called method of lines is used in order to approximate the solutions of heat- and wave-equations. The method of lines consists of a semi-discrete form of collocation where collocation is used only in one direction. (Collocation is the method of choice when we wish to use interpolation with radial basis functions to approximate data.) The class of radial basis functions which is admitted to be used is defined via their generalised Fourier transforms. This is reasonable because most of the results on polynomial reproduction and convergence properties of kernel methods depend on features of their generalised Fourier transform (positivity, decay, smoothness, ...) and are not features of the real domain. Also, using these Fourier transforms (that are indeed required to be positive), weighted \(L^2\) norms can be defined which are in turn essential to identify so-called native spaces. Those are then required to perform convergence analyses. A very useful selection of examples of suitable radial basis functions is presented. The article begins with identifying convergence properties of such approximations in the realm of Sobolev spaces. (When the approximand is band-limited, the approximation orders are proved to be optimal.) Also, the existence of cardinal interpolants using the fairly general classes of radial basis functions which are defined by the author is derived. Those are then used for error estimates first of the method of lines for the heat equation and finally for the more general wave equation. In all cases, the properties of the Fourier transforms of the approximants are essential for the proofs.