Iterative Truncated SVD for Inverse Scattering Problems

Abstract

We propose an iterative truncated SVD (iTSVD) algorithm based on a nonlinear least squares formulation for Newton's method for inverse scattering problems. The key idea is to estimate the effective rank of the corresponding linearized operator (or the Jacobian matrix after discretization), which defines the dimension of the signal space of the associated forward operator. Once the effective rank is estimated, truncated SVD (TSVD) reconstruction in the signal space can be computed efficiently based on fast random SVD (RSVD) -- only requiring matrix-vector multiplication which can be implemented in a matrix-free formulation by solving multiple forward and adjoint problems of the same type. Our analysis and numerical experiments show that iTSVD can markedly improve computational efficiency by reducing the number of iterations and enhance stability by physics/information based inversion. Moreover, the effective rank estimate is naturally model- and data-based. Extensive experiments, including far- or near-field, limited aperture, and random measurement data without or with noise, are presented.