A0354
Title: Dynamic matrix recovery
Authors: Ziyuan Chen - Peking University (China)
Ying Yang - Fudan University (China) [presenting]
Fang Yao - Peking University (China)
Abstract: Matrix recovery from sparse observations is an extensively studied topic emerging in various applications, such as recommendation systems and signal processing, which includes matrix completion and compressed sensing models as special cases. A general framework is proposed for dynamic matrix recovery of low-rank matrices that evolve smoothly over time. Starting from the case that the observations are independent across time, it is extended to the setting that both the design matrix and noise possess certain temporal correlations by modified concentration inequalities. By combining neighboring observations, sharp estimation error bounds of both settings are obtained, showing the influence of the underlying smoothness, the dependence and effective samples. A dynamic, fast iterative shrinkage thresholding algorithm that is computationally efficient and characterizes the interplay between algorithmic and statistical convergence is proposed. Simulated and real data examples are provided to support such findings.
