On the Gibbs phenomenon for wavelet expansions (Q1907514)

It is shown that for a wide class of orthonormal wavelet expansions a Gibb's phenomenon occurs for the wavelet approximation at a certain scale, analogous to the classical Gibb's phenomenon for Fourier series of functions with a jump discontinuity. The Gibb's phenomenon for wavelets follows from considerations similar to those in the classical situation, i.e., by representing the wavelet approximation by an integral operator, the kernel of which oscillates in a similar fashion as the Dirichlet kernel. However, the analysis is somewhat more involved, since the kernel is not explicitly known. The main result (Theorem 3.5) says a Gibb's phenomenon occurs at 0 (left or right) if the scaling function is continuous, at some dyadic point is differentiable with nonvanishing derivative, and decays as \((1+ |x|)^{-\beta}\), \(\beta> 3\) at \(\pm\infty\). It seems likely that this result is not best possible, although it is clear that some regularity is needed, since there is no Gibb's phenomenon for the Haar wavelet. A more detailed analysis is performed in the special and simpler case of spline wavelets, and, finally, it is shown how the Gibb's phenomenon can be avoided by replacing the \(L^2\) projection by an integral operator with a positive kernel, in the same way as the Dirichlet kernel is replaced by the Féjer kernel in the classical situation.