The semidefinite program PhaseLift can solve
phase retrieval with corrupted data.
Paul
Hand, Rice University
Abstract:
Phase retrieval is the process of
recovering a vector from phaseless linear
measurements. It is a challenging mathematical task
that appears in X-ray crystallography and other
applications. The problem is nonconvex, and it can be
convexified into a well-known semidefinite rank-one matrix
recovery problem known as PhaseLift. In some cases,
the measurements can be corrupted. This can happen,
for example, when there are occlusions, sensor failures, or
sensor saturation. In this talk, we show that a
variant of PhaseLift can successfully tolerate corrupted
data. Specifically, we show that under a Gaussian
measurement model, any signal can be recovered with high
probability, provided there are enough measurements and
provided that at most a fixed fraction of the measurements
are arbitrarily corrupted.