Optimal convergence rates in signal recovery (Q921713)

The aim of the paper is to investigate the performance of truncated Fourier deconvolution methods as far as the rate of convergence is concerned. For each n let us consider a signal on the periodic lattice of \({\mathbb{Z}}^ d\) with \(n^ d\) sites, whose Fourier transform \(\hat t(\xi)\) is bounded in absolute value by \(An^ d(1+\| n\xi \|)^{-a},\) for \(a>1\) and \(A>0\). Such a signal is blurred by a properly scaled point- spread function which is either of geometric type (out-of-focus blur) or uniform in an interval (one-dimensional motion blur). Independent, identically distributed noise with finite variance \(\eta^ 2\) is added, scaled in such a way that the necessary condition for consistency \(\xi =n^{-d}\eta^ 2\) holds. Next consider restorations of the type \[ \hat t(j)=(2\pi)^{- d}Re\int_{\Theta}e^{-ijr\vartheta}\zeta (\vartheta)\{\chi (\vartheta)\}^{-1}d\vartheta \] where \(\zeta\) is the Fourier transform of the recorded signal and \(\chi\) is the Fourier transform of the PSF. Intuitively the smoothing set \(\Theta\) should be designed in order to exclude regions in which \(\chi\) is too small. If an appropriate choice of the smoothing set is performed, the order of convergence of the worst case mean square estimation error, which depends only on \(\xi\), is exactly computed, and if the error is Gaussian, it is not possible to improve it. In the case of out-of-focus blur, the rate of convergence improves as a increases, that is as the image becomes smoother. This does not happen for motion blur since the rate is insensitive to higher-order derivatives, due to the fact that the restoration is made difficult at low frequencies rather than at high frequency components.