An iterated quasi-interpolation approach for derivative approximation (Q2192583)

Quasi-interpolation is a well-established excellent tool for function approximation e.g. on equally-spaced data (grids, meshes) or, sometimes in advanced settings, on scattered data. It consists of forming sums of (usually shifts of) so-called kernel functions times function evaluation of the approximand (other evaluations rather than point values are possible, averages for instance when the approximand is not continuous or is noisy). The advantages of quasi-interpolation are manifold: simplicity of formulation, no need for interpolation and solution of interpolation equations, usually very good accuracy and asymptotic convergence obtained from polynomial precision (reproduction), very large classes of suitable kernels that provide successful results, relations to many other useful mathematical approaches (sampling theorems for example or minimisation properties). The authors of this paper wish to approximate not only functions but their derivatives too. The usual method would be to compute the quasi-interpolant and then differentiate it, as one would normally compute derivatives of approximands. This is standard practice in approximation theory (e.g. polynomial interpolation or otherwise), but it leads to loosing the asymptotic convergence orders one by one. This article overcomes this disadvantage by computing quasi-interpolants \textit{iteratively\/} that is, compute a quasi-interpolant, then differentiate it, and compute a quasi-interpolant again. The idea is of striking simplicity and usefulness; the authors establish the corresponding convergence orders which now have no loss anymore and offer numerical examples to underline the success of the algorithm.