摘要：Markov chain Monte Carlo methods have become ubiquitous across science and engineering to model dynamics and explore large sets of configurations. The idea is to perform a random walk among the configurations so that even though only a very small part of the space is visited, samples will be drawn from a desirable distribution. Over the last 20 years there have been tremendous advances in the design and analysis of efficient sampling algorithms for this purpose, building on insights from statistical physics. One of the striking discoveries has been the realization that many natural Markov chains undergo phase transitions, whereby they change from being efficient to inefficient as some parameter of the system is modified, also revealing interesting properties of the underlying random structures. We will explore how phase transitions can provide valuable insights in three settings. First, they allow us to understand the limitations of certain classes of sampling algorithms, potentially leading to faster alternative approaches. Second, they reveal statistical properties of stationary distributions, giving insight into various interacting models. Example include colloids, or binary mixtures of molecules, segregation models, where individuals are more likely move when they are unhappy with their local demographics, and interacting particle systems from statistical physics. Last, they predict emergent phenomena that can be harnessed for the design of distributed algorithms for certain asynchronous models of programmable active matter. We will see how these three research threads are closely interrelated and inform one another. The talk will take a random walk through some of the results included in the references.
