Absolute and relative cut-off in adaptive approximation by wavelets (Q2568718)

The authors propose a class of relative cut-off methods which keep the local regularity of the function in the thresholding into account by comparing each wavelet coefficient with its neighbour coefficients. Neighbours are defined by their physical position and the scale level. The theoretical justification of this procedure is a characterization of the local Besov regularity of a function in terms of a subset of its wavelet coefficients. Two scales of Besov spaces are considered: vertical scale corresponding to Sobolev-type spaces and diagonal scale used by DeVore et al. for absolute cut-off operators, each scale yielding a family of relative cut-off operators. Relative cut-off is found to perform better than diagonal cut-off in the vertical scale provided the regularity index is not too small, otherwise the errors are asymptotically equivalent. In the diagonal scale, relative cut-off and optimal absolute cut-off are comparable.