Bases for kernel-based spaces (Q651113): Difference between revisions

When computing interpolants with radial basis functions or other kernel methods, using translates of the kernels directly is most often a bad idea, since the linear systems that give rise to the interpolation coefficients are ill conditioned. This does not mean that the problem as such is not well-solvable, but a new basis within the space of translates of the kernel is needed. In this paper, many important special choices of such changes of bases are considered, such as orthonormal bases, Lagrange bases, Newton bases and several others. All these bases depend on the choices of data points as well. An adaptive choice of Newton bases is also considered, as well as numerical examples given. The paper is a most useful and comprehensive contribution to the class of kernel and radial basis function approximation schemes.