Controlling linearization errors with rerandomization

Least squares migration aims to fit the observed seismic data with data predicted by linearized modelling, by solving an PDE-constrained optimization problem. This problem is challenging mostly because of its prohibitive computational cost. To address the issue, dimensionality reduction techniques were proposed in the literature. However, the solution of the reduced problem can deviate from that of the full problem when there are components in the observed data that cannot be explained by linearized modelling. We solve the problem by rerandomizing our $\ell_1$ regularized inversion. In this presentation, I will explain the method and demonstrate what we can achieve with rerandomization, especially for resolving fine sub-salt structures.