Local approximation of operators (Q2689140)

Many current problems of interest such as deep learning, neural networks, inverse problems and solutions of partial differential equation can be expressed as the approximation of operators on compact sets. Concretely, we are given finitely many data that describe the operator and we wish to study the order of approximation that can be achieved by encoding/decoding of the data. The main problem in this task is the so-called curse of dimensionality, because the number of data and the dimensions of the underlying spaces, respectively, can be prohibitively large. In this article, the approximation orders of these approximation operators are studied with a particular emphasis to having efficient algorithms, precise constants in the error estimates and especially useful upper bounds. The article's theme should in particular contrasted to the classical questions of approximation theory where (nonlinear) functions are approximated, where in this paper we deal now with nonlinear operators and functionals. In this context, it is very relevant that its special usefulness also comes from the uniform error estimates that are important in practice, as compared to the standard estimates in Hilbert space norm which are good in theory, seen often in other contributions, but less useful in applications. A particular emphasis is placed on examples of operators defined on high-dimensional spheres.