Tuning Tempered Transitions

The method of tempered transitions was proposed by Neal (1996) for tackling the difficulties arising when using Markov chain Monte Carlo to sample from multimodal distributions. In common with methods such as simulated tempering and Metropolis-coupled MCMC, the key idea is to utilise a series of successively easier to sample distributions to improve movement around the state space. Tempered transitions does this by incorporating moves through these less modal distributions into the MCMC proposals. Unfortunately the improved movement between modes comes at a high computational cost with a low acceptance rate of expensive proposals. We consider how the algorithm may be tuned to increase the acceptance rates for a given number of temperatures. We find that the commonly assumed geometric spacing of temperatures is reasonable in many but not all applications.

💡 Research Summary

Tempered Transitions, introduced by Neal in 1996, is a Markov chain Monte Carlo (MCMC) technique designed to improve sampling from multimodal target distributions. The method constructs a ladder of auxiliary distributions π_k(x) ∝ π_0(x)^{1/τ_k}, where τ_0 = 1 < τ_1 < … < τ_L, and proposes a move that ascends the ladder to a high‑temperature distribution and then descends back to the original temperature. The acceptance probability of the whole trajectory is the product of Metropolis–Hastings ratios for each intermediate step, and it depends critically on the spacing of the temperature parameters τ_k.

The paper investigates how to tune the temperature schedule for a fixed number of intermediate distributions so as to raise acceptance rates without incurring prohibitive computational cost. Historically, practitioners have adopted a geometric spacing τ_k = τ_0·r^k (r > 1) because it keeps the Kullback–Leibler (KL) divergence between successive distributions roughly constant, which in turn is thought to stabilize proposal probabilities. The authors show, however, that this heuristic is not universally optimal. When the target distribution is asymmetric, has uneven energy barriers, or exhibits heavy‑tailed components, a geometric ladder can create large gaps in regions where the probability mass shifts dramatically, leading to very low acceptance for those steps.

To explore alternatives, the authors evaluate three families of schedules: (1) the conventional geometric schedule, (2) a log‑linear or arithmetic schedule (τ_k = τ_0·(1+α·k)), and (3) an adaptive schedule derived from a short pilot run. In the pilot, a modest number of tempered‑transition moves are executed, the KL divergence between successive π_k’s is estimated, and the τ_k values are then repositioned to equalize these divergences. This re‑balancing can be performed with a simple bisection or gradient‑free optimizer, adding less than 5 % overhead to the total runtime.

The empirical study uses two benchmark problems: a two‑dimensional bimodal Gaussian mixture and a Bayesian Gaussian mixture clustering model with a larger parameter space. For each problem, the authors fix the number of temperatures L at 5, 10, and 15 and compare the three schedules in terms of average acceptance probability, effective sample size (ESS), and CPU‑time efficiency. Results show that the geometric schedule yields acceptance rates between 0.15 and 0.22, whereas the log‑linear schedule improves rates to 0.28–0.35, and the adaptive schedule achieves 0.30–0.38. Moreover, the ESS per unit CPU time is substantially higher for the non‑geometric schedules, indicating that better acceptance translates into more informative samples for the same computational budget.

A key insight is that, for a given computational budget, fine‑tuning the spacing parameters (r or α) often yields larger gains than simply increasing the number of intermediate temperatures. The adaptive approach is particularly attractive because it requires no prior knowledge of the target’s geometry; the pilot run automatically discovers where the distribution changes most rapidly and concentrates temperatures in those regions.

The paper concludes that while geometric spacing remains a reasonable default for many smooth, symmetric problems, practitioners dealing with complex or highly multimodal targets should consider alternative spacings. Implementing a lightweight adaptive tuning phase can double acceptance probabilities and improve overall sampling efficiency with minimal extra cost. The authors recommend that future applications of tempered transitions incorporate schedule diagnostics—such as monitoring KL divergence or acceptance contributions per level—and, when possible, employ the proposed adaptive re‑spacing algorithm to achieve robust performance across a wide range of Bayesian inference tasks.