An SVD-free Pareto curve approach to rank minimization

Recent SVD-free matrix factorization formulations have enabled rank optimization for extremely large-scale systems (millions of rows and columns). In this paper, we consider rank-regularized formulations that only require a target data-fitting error level, and propose an algorithm for the corresponding problem. We illustrate the advantages of the new approach using the Netflix problem, and use it to obtain high quality results for seismic trace interpolation, a key application in exploration geophysics. We show that factor rank can be easily adjusted as the inversion proceeds, and propose a weighted extension that allows known subspace information to improve the results of matrix completion formulations. Using these methods, we obtain high-quality reconstructions for large scale seismic interpolation problems with real data.