國立清華大學 數學系

Abstract:

Poisson inverse problems arise in many real-world applications, such as positron emission tomography and astronomical image deblurring. Expectation maximization (EM) is a standard---and perhaps the most popular---approach to solving a Poisson inverse problem. Vardi et al. proved EM asymptotically converges more than three decades ago; however, it was unclear how fast EM converges. In this talk, I will present a simple non-asymptotic convergence guarantee for EM. Our analysis exploits an interesting connection between EM and a portfolio selection method due to Cover