Approximation by radial bases and neural networks (Q5934364)

Approximants for \(f\) of the form \(n^{-s(1-r)}\sum_{{\mathbf j}\in A_n} L_{n{\mathbf j}}(f)\phi(n^r(\cdot-{\mathbf j}/n))\) are considered. \(n\) is the number of points, \(A_n\) is a subset of \({\mathbb{Z}}^s\), \(0<r\leq 1\), and \(\phi\) is some function to be chosen to give good approximations. The \(L_{n{\mathbf j}}\) are linear functionals of \(f\), for example function values. Under various conditions on \(\phi\), several theorems on the order of approximation are obtained. This is applied when \(\phi\) is chosen as a linear combination of translated and/or dilated versions of radial basis functions. Finally in the univariate case, \(\phi\) is chosen as a linear combination of translates of a sigmoid function. This is like a dilation equation to which a wavelet function can be associated. A class of sigmoidal functions is identified which allows to reconstruct a function from the corresponding integral wavelet transform. This construction is given a neural network interpretation.