no. 14 · Friday 13 November 2026
Swapping and Switching for Markov Processes: Symmetrisation and Fast-Exchange Limits
Abstract
In parallel tempering, exchanging the positions of the replicas (swapping) and exchanging the assignment of dynamics to the replicas (switching) are often used interchangeably. We study the two descriptions as stochastic processes, for Markov species (the dynamics that are exchanged) and bounded measurable exchange rates. To guarantee the Markovianity of each process under the exchanges, switching records both the positions and the label that tracks which species each replica carries; swapping records only the positions. Driving both with one exchange clock places them on a common probability space and lets us write swapping as switching with the label forgotten. This relation holds pathwise at every time and, once decomposed over the label, yields an exact identity between their transition kernels. This characterisation lets us relate, in a precise way, where the two processes agree and where they differ. We describe mainly the relation between their decay rates at finite exchange rate, and between their distributions and observables at finite and infinite exchange rate. The relation between the Markovianity of the two is particularly interesting. At finite rate, switching needs the label to be Markov, while swapping is Markov on its own. In the infinite-rate limit this reverses: swapping becomes non-tight, while switching becomes Markov on the physical coordinates alone. We illustrate all these relations on a solvable benchmark of two Brownian motions with drift under a constant exchange rate. The rich behaviour already visible in this simple case, and developed throughout the paper, suggests that the swapping mechanism is of interest in its own right, beyond the balance conditions needed in parallel tempering.
- fields
- Probability