Learning a function from noisy samples at a finite sparse set of points (Q1048968)

The paper deals with the task to learn a function \(f\) defined on a domain \(\Omega\) if only the values at a sparse, discrete subset \(\omega\subset\Omega\) are available. In the process of generalizing this given information, a generalization error occurs. It is the aim of the paper to estimate the bounds of this error under restrictions which are fulfilled in many cases. In the appendix, two ways which meet the assumptions are discussed. The results are developed using a bound on the generalization error basing on the Koksma-Hlawka type. It is shown that the generalization error has a deterministic bound and tends to zero ``if the noise in the measurement tends to zero and the number of sampling points tends to infinity sufficiently fast.''