PhD Student, Central South University

Talk Title

Perturbation Analysis on Hellinger Distance for Graphical Models

Abstract

For complicated graph models with large degrees, the performance of the classical Gibbs sampler may be limited. A natural approach is to replace it with a computationally-cheaper proxy. This raises a question of how much difference can exist between the classical and the proxy. By using an approach based on the decay-of-correlation property and subadditivity of Hellinger distance, we provide a weaker perturbation condition for Markov random fields, which ensures the performance of the Gibbs sampler is robust. Furthermore, we illustrate our results in two different application settings, the minibatch Gibbs samplers and the hierarchical models.

Bio

My name is Na Lin, and I am a PhD student in Statistics at Central South University under the supervision of Professor Yuanyuan Liu. Currently, I am a visiting research student at the University of Ottawa, working with Professor Aaron Smith, and I will be staying here for two years. My research interests primarily focus on Markov chains and the related algorithms (e.g. Markov chain Monte Carlo), particularly their stability analysis.