WaRIance: wavefield reconstruction inversion with stochastic variable projection

Abstract

We propose a variation on wavefield reconstruction inversion for seismic inversion, which takes advantage of randomized linear algebra as a way to overcome the typical limitations of conventional inversion techniques. Consequently, we can aim both to robustness towards convergence stagnation and large-sized 3D applications. The central idea hinges on approximating the optimal slack variables involved in wavefield reconstruction inversion via a low-rank stochastic approximation of the wave-equation error covariance. As a result, we obtain a family of inversion methods parameterized by a given model covariance (suited for the problem at hand) and the rank of the related stochastic approximation sketch. The challenges and advantages of our proposal are demonstrated with some numerical experiments.