Matrix Completion with Nonuniform Sampling: Theories and Methods.

Prevalent matrix completion theories reply on an assumption that the locations of missing data are distributed independently and randomly (i.e., uniform sampling). Nevertheless, the reason for an observation being missing often depends on the unseen observations themselves, and thus the locations of the missing data in practice usually occur in a correlated fashion (i.e., nonuniform sampling) rather than independently. To break through the limits of uniform sampling, we introduce in this work a new hypothesis called isomeric condition, which is provably weaker than the assumption of uniform sampling. Equipped with this new tool, we prove a collection of theorems for missing data recovery as well as matrix completion. In particular, we prove that the exact solutions that identify the target matrix are included as critical points by the commonly used bilinear programs. Even more, when an extra condition called relative well-conditionedness is obeyed as well, we prove that the local optimality of the exact solutions is guaranteed in a deterministic fashion. Among other things, we study in detail a Schatten quasi-norm induced method termed isomeric dictionary pursuit (IsoDP), and we show that IsoDP exhibits some distinct behaviors absent in the traditional bilinear programs.