Replica-exchange Monte Carlo is a popular computational technique for sampling complex probability distributions. The core idea is to run multiple copies of a system simultaneously, occasionally swapping their configurations, so that copies stuck in difficult regions of the landscape can escape by exchanging positions with more mobile copies. Despite widespread practical use, a rigorous mathematical foundation for this "swapping" mechanism in continuous time has been lacking. This paper fills that gap by carefully constructing the mathematical framework for two interacting particles whose positions are randomly exchanged, proving that the resulting process is well-defined and unique, and deriving the equation governing how probability spreads over time.
A central contribution is clarifying the relationship between swapping and a superficially similar mechanism called regime-switching, where a particle changes its dynamical rules rather than its position. The authors show that these two mechanisms are deeply connected through a symmetry identity: any observable that treats the two particles as interchangeable will give identical results under either mechanism, while observables sensitive to which particle is which will differ. They also prove that swapping never makes the system's mixing speed worse compared to independent particles, and in certain symmetric settings it strictly improves it. Crucially, for large collections of particles, swapping stays tractable because you only need to track particle positions, whereas switching requires keeping track of an exponentially growing number of label assignments, representing a fundamental practical advantage of the swapping approach.
To ground the theory in something concrete, the authors work out a fully explicit example: two Brownian particles with constant drift that randomly swap positions at a fixed rate. For this model, they derive exact formulas for the transition probabilities (expressed using special functions called modified Bessel functions), the moments of the distribution, and how particle positions are correlated across time. Their analysis reveals some subtle effects: the particles behave as if they have a slightly altered effective diffusivity, lingering asymmetries persist even at long times, and when swaps happen very frequently, the system converges to its limiting behavior at a rate proportional to the inverse square root of the swap rate, meaning faster swapping helps but with diminishing returns.