Sparse linear algebra and geophysical migration: A review of direct and iterative methods

The pre-stack depth migration of reflection seismic data can be expressed, in the framework of waveform inversion, as a linear least squares problem. Together with the precise definition of this operator, we detail additional main characteristics of the forward model, like its huge size, its sparsity and the composition with convolution. It ends up with a so-called discrete ill-posed problem, whose acceptable solutions have to undergo a regularization procedure. Both direct and iterative methods have been implemented with specific attention to the convolution, and then applied to a given data set: a synthetic 2-dimensional profile of revealing size with some added noise. The efficiency with regard to computational effort and storage requirements is evaluated. The needed regularization of the solution is thoroughly studied in both cases. From the point of the global inverse problem, the extra feature of providing a solution that can be differentiated with respect to a parameter such as background velocity is also discussed.