Nonlinearly Bandlimited Signals
For researchers in inverse scattering and signal processing, this work offers a rigorous sampling-theoretic framework and efficient numerical algorithms for solving a class of nonlinear inverse problems.
This paper addresses the inverse scattering problem for signals with compactly supported reflection coefficients, providing a constructive proof of existence for solutions of the Gelfand-Levitan-Marchenko equations under restrictions on the nonlinear impulse response. It develops algorithms with algebraic convergence rates, achieving O(N^{-1/2}) using Whittaker-Kotelnikov-Shannon sampling and O(N^{-m-1/2}) using the Helms-Thomas version, with complexities ranging from O(N_iter N^2) to O(N_iter N log N) per sample.
In this paper, we study the inverse scattering problem for a class of signals that have a compactly supported reflection coefficient. The problem boils down to the solution of the Gelfand-Levitan-Marchenko (GLM) integral equations with a kernel that is bandlimited. By adopting a sampling theory approach to the associated Hankel operators in the Bernstein spaces, a constructive proof of existence of a solution of the GLM equations is obtained under various restrictions on the nonlinear impulse response (NIR). The formalism developed in this article also lends itself well to numerical computations yielding algorithms that are shown to have algebraic rates of convergence. In particular, the use Whittaker-Kotelnikov-Shannon sampling series yields an algorithm that converges as $\mathscr{O}\left(N^{-1/2}\right)$ whereas the use of Helms and Thomas (HT) version of the sampling expansion yields an algorithm that converges as $\mathscr{O}\left(N^{-m-1/2}\right)$ for any $m>0$ provided the regularity conditions are fulfilled. The complexity of the algorithms depend on the linear solver used. The use of conjugate-gradient (CG) method yields an algorithm of complexity $\mathscr{O}\left(N_{\text{iter.}}N^2\right)$ per sample of the signal where $N$ is the number of sampling basis functions used and $N_{\text{iter.}}$ is the number of CG iterations involved. The HT version of the sampling expansions facilitates the development of algorithms of complexity $\mathscr{O}\left(N_{\text{iter.}}N\log N\right)$ (per sample of the signal) by exploiting the special structure as well as the (approximate) sparsity of the matrices involved.