Source-linked AI summary
Dynamic temperature selection for parallel-tempering in Markov chain Monte Carlo simulations
Will Vousden, Will M. Farr, Ilya Mandel
TL;DR
Parallel tempering improves sampling of complex multimodal posteriors, but choosing an efficient temperature ladder is difficult because it depends on the unknown target distribution. The paper dynamically adapts ladder spacing to equalize neighboring-chain exchanges, finding generally better and more consistent ACT performance than conventional geometric ladders. The method is simple to implement, but adaptive updates require suppression or diminishing adjustment to preserve long-run ergodicity.
Problem
Efficient parallel tempering requires a temperature ladder that minimizes cold-chain autocorrelation time, but the target distribution’s shape is generally unknown.
Method
The paper dynamically adapts an ensemble-based or single-walker parallel-tempering ladder so exchange rates between neighboring chains become uniform.
Results
Uniform-acceptance ladders generally outperform conventional geometric ladders, including a ∼4 ACT ratio in a five-dimensional egg-box case and a 26% average ACT reduction in the astrophysical application.
Takeaways & Limitations
A simple adaptive ladder provides more consistent sampling behavior across likelihood distributions and numbers of chains while requiring little tuning or intervention.
Takeaways & Limitations
Adaptive temperature adjustment violates detailed balance unless its amplitude diminishes, so the scheme suppresses adjustments to recover long-time Markovian behavior.
Abstract
from arXiv · showhide
Modern problems in astronomical Bayesian inference require efficient methods for sampling from complex, high-dimensional, often multi-modal probability distributions. Most popular methods, such as Markov chain Monte Carlo sampling, perform poorly on strongly multi-modal probability distributions, rarely jumping between modes or settling on just one mode without finding others. Parallel tempering addresses this problem by sampling simultaneously with separate Markov chains from tempered versions of the target distribution with reduced contrast levels. Gaps between modes can be traversed at higher temperatures, while individual modes can be efficiently explored at lower temperatures. In this paper, we investigate how one might choose the ladder of temperatures to achieve more efficient sampling, as measured by the autocorrelation time of the sampler. In particular, we present a simple, easily-implemented algorithm for dynamically adapting the temperature configuration of a sampler while sampling. This algorithm dynamically adjusts the temperature spacing to achieve a uniform rate of exchanges between chains at neighbouring temperatures. We compare the algorithm to conventional geometric temperature configurations on a number of test distributions and on an astrophysical inference problem, reporting efficiency gains by a factor of 1.2-2.5 over a well-chosen geometric temperature configuration and by a factor of 1.5-5 over a poorly chosen configuration. On all of these problems a sampler using the dynamical adaptations to achieve uniform acceptance ratios between neighbouring chains outperforms one that does not.
1 INTRODUCTION
Parallel tempering uses chains at different temperatures to explore multimodal posteriors, but selecting an efficient temperature ladder remains difficult. This paper proposes adaptive temperature selection and evaluates it on test distributions and gravitational-wave inference.
- Motivation: Parallel tempering runs multiple chains at different temperatures, enabling hot chains to traverse separated modes and cold chains to explore high-probability regions efficiently.Exchange of chain positions propagates exploration between temperature levels.
- Open problem: The ladder-selection problem is to minimize the cold chain’s autocorrelation time, with efficiency depending critically on position transfer between high and low temperatures.The appropriate ladder is difficult to determine because it depends on the target distribution.
- Approach: The paper presents a simple algorithm that dynamically adapts temperatures to produce a uniform exchange rate between neighboring chains.Implementations are provided for ensemble-based and traditional single-walker MCMC samplers.
- Results: > 1.2 improvement in ACT over conventional geometric spacing is reported across the test cases.The strategy is not necessarily ACT-optimal but improves on the conventional geometric ladder in the reported tests.
- Results: A gravitational-wave parameter-estimation application demonstrates an ACT reduction of as much as a factor of 2 over a geometric ladder.The experiment uses a single-walker MCMC sampler.
2 PARALLEL TEMPERING
Parallel tempering combines temperature-dependent exploration with swaps between chains, making ladder design central to cold-chain efficiency. The section motivates adaptive spacing by showing why geometric ladders can fail when target-distribution structure and prior limits are unknown.
- Parallel tempering: Parallel tempering samples tempered posterior distributions with chains ordered from the target temperature T = 1 to higher temperatures.Higher temperatures flatten likelihood peaks, while swaps exchange chain positions between levels.
- Ladder requirements: The cold chain’s efficiency is measured by its autocorrelation time, which depends on both reaching a temperature where modes mix and maintaining communication between neighboring chains.These are the two requirements for an effective ladder.
- Ladder requirements: The required maximum temperature depends on the modes’ size and separation, so different target distributions require different ladder endpoints.For two modes centered at ±10 with width σ = 1, the example chooses Tmax = 100 to broaden them substantially.
- Geometric spacing: With a finite prior volume, geometric ladders can add chains redundantly above Tprior because the unknown prior-transition temperature limits the useful range.Too few chains may fail to reach the prior-like regime, while too many can sample it redundantly.
- Adaptive motivation: Uniform acceptance-ratio ladders are proposed as an alternative, but the paper assumes rather than proves that they approximate ACT-optimal spacing.The authors explicitly identify this assumption as requiring further study.
- Geometric spacing: Geometric spacing can yield uniform acceptance ratios for an ideal unimodal Gaussian, but increasing dimension requires closer temperature spacing for the same acceptance ratio.For the ideal Gaussian, uniform acceptance ratios arise because acceptance depends on the temperature ratio; the required spacing nevertheless depends on dimension.
3 ADAPTIVE TEMPERATURE LADDERS
The algorithm dynamically adjusts chain temperatures so neighboring swap acceptance ratios become uniform, while its adjustment rate decays to support eventual convergence. The ladder keeps endpoint constraints and adapts internal spacings according to observed acceptance behavior.
- Adaptive objective: Uniform acceptance ratios provide the target for dynamically adapting the temperature ladder.The method assumes this approximates an optimal ladder and adjusts chain temperatures accordingly.
- Adaptive objective: Chains widen neighboring temperature gaps when swaps are accepted too often and narrow them when swaps are accepted too seldom.These rules drive internal chains toward evenly spaced acceptance ratios while fixing the cold and hottest endpoints.
- Temperature boundaries: The hottest chain must sample the prior so strongly separated modes become flattened and traversable.This is achieved by fixing its inverse temperature to β_N = 0, while the cold chain remains at T_1 = 1.
- Measurement and stability: Acceptance ratios are measured empirically from swap proposals, with fewer walkers requiring longer averaging to smooth measurement granularity.For ensemble samplers, A_i takes increments determined by the number of walkers; traditional one-sample chains yield binary measurements.
- Convergence and decay: Temperature adjustments use decaying dynamics because adaptation violates detailed balance and must diminish for convergence toward the target distribution.The paper notes a trade-off between ladder convergence and convergence of the sampler itself, using hyperbolic decay as a safeguard.
- Parameter choice: The dynamics should satisfy t_0 ≫ ν and νN_τ ≫ t_0 so temperatures equilibrate while responding to the sampler’s posterior distribution.For N_τ = 100, the authors report that t_0 = 10ν worked well in their test cases.
4 EXAMPLES
The examples evaluate dynamically adapted temperature ladders against geometric alternatives across truncated Gaussian, Rosenbrock, and egg-box problems, using cold-chain ACT as the primary efficiency measure. Uniform-acceptance adaptations generally improve ACT, especially when few chains or poorly chosen geometric endpoints create communication bottlenecks.
- Examples: The implementation applies temperature dynamics to ensemble and traditional single-walker parallel-tempered MCMC samplers.The method is implemented as a modification to emcee and also supports nonensemble samplers.
- 4.1 Truncated Gaussian: The truncated Gaussian tests show that prior truncation makes geometric spacing nonoptimal at high temperatures.The divergence from ideal-Gaussian behavior is caused by the tempered likelihood encountering prior boundaries.
- 4.2 Rosenbrock: An N-chain ladder adapted for uniform acceptance outperforms a same-sized geometric ladder for every tested N in the double Rosenbrock example.The benefit is strongest at low N, where larger temperature gaps can create severe communication bottlenecks near phase transitions.
- 4.2 Rosenbrock: For the double Rosenbrock distribution, reallocating temperatures reduces ACT by no more than 25% once the geometric ladder’s minimum acceptance exceeds approximately 10%, while overall improvement remains.This threshold occurs at approximately N = 7 in the reported example.
- 4.2 Rosenbrock: Rosenbrock autocorrelation functions separate intra-mode and inter-mode timescales, showing that removing low- or high-temperature chains affects different mixing processes.Removing the second chain increases intra-mode ACT, while removing higher-temperature chains increases inter-mode ACT.
- Examples: All tested ladders yield lower ACT than the default geometric ladder, even when that geometric ladder uses prior knowledge of Tprior.This comparison reinforces the empirical importance of evaluating the cold-chain ACT rather than relying only on ladder construction criteria.
- 4.3 Egg-box: Uniform-acceptance ladders maintain lower ACT than geometric ladders in the egg-box problem, with τgeo > 2τacc for N ≥ 12.Adding temperatures reduces ACT for the adapted ladder, whereas geometric additions above Tprior do not reduce ACT for N ≥ 7.
5 GRAVITATIONAL WAVE SIGNALS FROM COMPACT BINARY COALESCENCES
The paper applies dynamically adapted parallel tempering to gravitational-wave parameter estimation, where structured posteriors with modes and degeneracies challenge conventional MCMC. Across synthetic GW problems, uniform-acceptance temperature ladders reduce autocorrelation time relative to geometric ladders, while single-walker communication barriers limit the improvement.
- The target is the Bayesian posterior for parameters of a merging binary system, inferred from detector data, priors, and a model waveform.
- Gravitational-wave parameter estimation is difficult because its likelihood and parameter space produce highly structured posteriors with many modes and degeneracies.
- The implemented method adapts the temperature ladder of LALInference's single-walker parallel-tempered MCMC sampler for this inference problem.LALInference uses jump proposals tuned to the posterior structure generated by a compact-binary-coalescence signal.
- The dynamical ladder shifts its structure toward higher temperatures as signal-to-noise ratio increases, while its equilibrium chain density decreases.This follows the reported increase of the prior-transition temperature with signal-to-noise ratio under a fixed prior.
- 26% average ACT reduction was achieved across the tested GW systems and signal-to-noise ratios.The uniform-A ladder was at least as effective as the geometric ladder in all tested cases within error bars.
- Single-walker dynamics partition the temperatures into two non-communicating groups, so the observed improvement comes from reallocating temperatures below the phase transition.Chains above the critical temperature remain isolated while sampling where the noise-only model is preferred.
6 DISCUSSION
The uniform-A scheme addresses temperature-ladder selection by fixing Tmax = ∞ and dynamically allocating intermediate temperatures for uniform swap acceptance. It generally improves sampling efficiency over geometric ladders, but its gains depend on the target distribution, ladder choice, and sampler implementation.
- Temperature-ladder design: Tmax = ∞ provides a target-distribution-independent hot-chain temperature that samples from the prior.This removes the need to choose a finite maximum temperature in advance.
- Temperature-ladder design: A fixed number of intermediate temperatures is allocated between Tmin and Tmax to maintain communication and efficient sampling.The dynamical scheme targets uniform acceptance ratios between neighboring chains.
- Efficiency: Uniform-A dynamics are simple to implement and generally produce lower ACTs and more consistent behavior than conventional geometric ladders.The scheme is not necessarily ACT-optimal, but tests show improvements across likelihood distributions and chain counts.
- Scope and limitations: Uniform-A gains depend strongly on the likelihood distribution and geometric baseline, while τ_geo/τ_acc approaches 1 as N increases.Distributions with broad log-likelihood ranges or sharp phase transitions benefit most from many chains.
- Efficiency: ∼4 is the largest reported τ_geo/τ_acc ratio for the 5-dimensional egg-box distribution under a poor geometric Tmax choice.Uniform-A CPU time continued decreasing with N in the explored range while geometric-ladder CPU time increased.
- Astrophysical application: 26% is the average ACT reduction in the astrophysical application relative to LALInference’s default geometric ladder.The single-walker tests also revealed instability in the uniform-A dynamics.
APPENDIX A: AUTOCORRELATION TIME ESTIMATION
The appendix defines and estimates integrated autocorrelation time from a normalized autocorrelation function, truncating the estimate to limit noise and enforce a sample-size condition.
- The paper's ACT is the integrated autocorrelation time described by Sokal (1997).
- For a time series x(t) with normalized autocorrelation ρ(t) and ρ(0) = 1, integrated ACT is defined from ρ(t).
- Because ρ(t) ≈ 0 at lags much larger than τ, large-lag contributions are dominated mainly by autocorrelation-estimation noise.
- The ACT is estimated over a window M = 5 ACTs long, requiring Mτ < N/2 for a trustworthy result.