@conference {li2012SEGspmamp, title = {Sparsity-promoting migration accelerated by message passing}, booktitle = {SEG Technical Program Expanded Abstracts}, volume = {31}, year = {2012}, month = {11}, pages = {1-6}, publisher = {SEG}, organization = {SEG}, abstract = {Seismic imaging via linearized inversion requires multiple iterations to minimize the least-squares misfit as a function of the medium perturbation. Unfortunately, the cost for these iterations are prohibitive because each iteration requires many wave-equation simulations, which without direct solvers require an expensive separate solve for each source. To overcome this problem, we use dimensionality-reduction to decrease the size of seismic imaging problem by turning the large number of sequential shots into a much small number of simultaneous shots. In our approach, we take advantage of sparsifying transforms to remove source crosstalk resulting from randomly weighting and stacking sequential shots into a few super shots. We also take advantage of the fact that the convergence of large-scale sparsity-promoting solvers can be improved significantly by borrowing ideas from message passing, which are designed to break correlation built up between the linear system and the model iterate. In this way, we arrive at a formulation where we run the sparsity-promoting solver for a relatively large number of very iterations. Aside from leading to a significant speed up, our approach had the advantage of greatly reducing the memory imprint and IO requirements. We demonstrate this feature by solving a sparsity-promoting imaging problem with operators of reverse-time migration, which is computationally infeasible without the dimensionality reduction.}, keywords = {Imaging, inversion, SEG}, doi = {10.1190/segam2012-1500.1}, url = {https://slim.gatech.edu/Publications/Public/Conferences/SEG/2012/li2012SEGspmamp/li2012SEGspmamp.pdf}, presentation = {https://slim.gatech.edu/Publications/Public/Conferences/SEG/2012/li2012SEGspmamp/li2012SEGspmamp_pres.pdf}, author = {Xiang Li and Felix J. Herrmann} }