The Potts model is a mathematical framework for studying how particles (like tiny magnets) arranged on a network can align into ordered phases. Imagine a grid of voters, each able to choose among q political parties, where neighbors tend to influence each other toward the same choice. The "mean-field" version simplifies the geometry by assuming every voter is equally connected to every other voter. Researchers study how efficiently computer simulations of such models reach their equilibrium, or stationary, behavior. The standard way to simulate these systems is called Glauber dynamics, which updates one randomly chosen voter at a time. This paper instead analyzes "systematic scan" dynamics, where voters are updated one by one in a fixed, predetermined order, cycling through all of them repeatedly. This approach is widely used in practice because it tends to perform well empirically, but its theoretical behavior has been much harder to pin down mathematically.
The central result of the paper is that, below a critical temperature threshold, systematic scan dynamics mixes in a time proportional to the logarithm of the number of particles, and moreover exhibits a sharp "cutoff" phenomenon. Cutoff means the system behaves in a striking all-or-nothing way: for a long time the simulation looks nothing like equilibrium, and then within a very brief window of time it suddenly snaps into equilibrium. This is analogous to a glass of water with a drop of dye that remains swirled and unmixed for a while and then abruptly becomes uniformly colored. The authors precisely identify the mixing time and show that above the critical temperature threshold the system takes exponentially longer to mix, confirming the result is as sharp as possible.
This matters for several reasons. First, it provides the first rigorous cutoff result for systematic scan dynamics applied to any spin system, closing a significant gap between theory and practice in the study of Markov chain Monte Carlo algorithms, which are workhorse tools in statistics, physics, and machine learning. Second, systematic scan dynamics has two features, being global (each full sweep updates every particle) and non-reversible (the fixed ordering breaks time-reversal symmetry), that make cutoff theory especially difficult to develop. Most existing cutoff theory applies to simpler, reversible chains. By handling both challenges simultaneously, the paper opens a path toward understanding a broader and more practically relevant class of simulation algorithms.