Greedy sparse linear approximations of functionals from nodal data

Abstract

In many numerical algorithms, integrals or derivatives of functions have to be approximated by linear combinations of function values at nodes. This ranges from numerical integration to meshless methods for solving partial differential equations. The approximations should use as few nodal values as possible and at the same time have a smallest possible error. For each fixed set of nodes and each fixed Hilbert space of functions with continuous point evaluation, e.g. a fixed Sobolev space, there is an error-optimal method available using the reproducing kernel of the space. But the choice of the nodes is usually left open. This paper shows how to select good nodes adaptively by a computationally cheap greedy method, keeping the error optimal in the above sense for each incremental step of the node selection. This is applied to interpolation, numerical integration, and numerical differentiation. The latter case is particularly important for the design of meshless methods with sparse generalized stiffness matrices. The greedy algorithm is described in detail, and numerical examples are provided. In contrast to the usual practice, the greedy method does not always use nearest neighbors for local approximations of function values and derivatives. Furthermore, it avoids multiple points from clusters and it is better conditioned than choosing nearest neighbors.