High-dimensional signal estimation plays fundamental roles in various engineering and science applications, such as medical imaging, video and network surveillance. Estimation procedures that maintain both statistical and computational efficacy are of great practical value, which translate into desiderata such as less time patients need to spend in a medical scanner, faster response to cyber attacks, and capabilities to handle very large datasets. While a lot of signal estimation tasks are naturally formulated as nonconvex optimization problems, existing results for nonconvex methods have several fundamental limitations, and the current state of the art is still limited in terms of when, why and which nonconvex approaches are effective for a given problem. The goal of this research program is to significantly deepen and broaden the understanding and applications of nonconvex optimization for high-dimensional signal estimation.
In this project, the investigators will study high-dimensional signal estimation via direct optimization of nonconvex, and potentially nonsmooth, loss functions, without resorting to convex relaxation. This research will explore geometric structures shared by nonconvex functions commonly encountered in signal estimation, and study the fundamental roles these structures play in determining the algorithmic convergence. These results will then be exploited as guidelines to develop fast and provably correct algorithms for estimating high-dimensional signals with physically induced structures and under streaming data observations. Specifically, the research program consists of three major thrusts: (1) understanding the geometric structures of important classes of nonconvex loss surfaces, and characterizing their impact on the convergence of optimization algorithms; (2) developing fast algorithms and the associated theory for the recovery of structured low-rank matrices; (3) designing new online algorithms that are time and space efficient under a streaming setting, with the capability of detecting and tracking the time-varying signals of interest.
