Nayantara Bhatnagar and
Dana Randall
Torpid Mixing of Simulated Tempering
on the Potts Model
Abstract:
Simulated tempering and swapping are two families of sampling algorithms
in which a parameter corresponding to temperature is varied during the
simulation. The hope is that we will be able to overcome bottlenecks
that cause Glauber (local) dynamics to be prohibitively slow at low
temperatures, due to the existence of bad cuts in the state space of
configurations.
The Ising model, and its generalization the Potts model, are models
from statistical physics for magnetism. Glauber dynamics and the
Swendsen-Wang algorithm have been shown to be slowly mixing for the
Potts model. On the other hand, it has been shown that the swapping and
tempering algorithms allow efficient sampling at low temperatures for
the mean-field Ising model. It is
reasonable therefore to ask if swapping or tempering can speed up
sampling for the Potts model.
In this paper, we show that in fact there are temperatures at which
tempering (and hence swapping) for the 3-state ferromagnetic Potts model
on the complete graph convereges slowly. Using this insight, we are
able to define a variant of the swapping algorithm that samples
efficiently from a class of bimodal distributions including the q-state
Potts model on the complete graph.