Dynamic Sampling from Graphical Models

In this paper, we study the problem of sampling from a graphical model when itself is changing dynamically with time. This derives its interest variety inference, learning, and settings in machine computer vision, statistical physics, theoretical science. While static has received considerable attention, works for dynamic variants have been largely lacking. The main contribution paper an algorithm that can sample broad class models over discrete random variables. Our parallel Las Vegas: it knows to stop, outputs samples exact distribution. We also provide sufficient conditions under which runs time proportional size update on general as well well-studied specific spin systems. particular obtain, Ising (ferromagnetic or antiferromagnetic) hardcore first algorithms handle both edge vertex updates (addition, deletion, change functions). these are efficient within regimes close respective uniqueness regimes, beyond which, even approximate sampling, no local were known intractable. relies resampling new “equilibrium" property shown be satisfied by our at each step enables us prove correctness. equilibrium robust enough guarantee correctness algorithm, helps improve bounds fast convergence models, should independent interest.