Closing the Evidence Gap: reddemcee, a Fast Adaptive Parallel Tempering Sampler

1. Problem Statement

In astrophysics and Bayesian inference, Markov Chain Monte Carlo (MCMC) methods are the standard for sampling complex, high-dimensional posterior distributions and estimating parameter uncertainties. However, MCMC traditionally struggles with Bayesian evidence estimation (the marginal likelihood, $Z$), which is crucial for rigorous model comparison (e.g., determining the number of planets in a system).

Conversely, Nested Sampling (NS) and Dynamic Nested Sampling (DNS) excel at evidence estimation but often provide fewer posterior samples and less precise parameter estimates compared to well-tuned MCMC. There is a significant "evidence gap" where MCMC lacks the robust, low-bias evidence estimators found in NS, while NS lacks the rich posterior information density of MCMC.

2. Methodology: The reddemcee Framework

The authors introduce reddemcee, a Python-based adaptive parallel tempering (APT) ensemble sampler designed to close this gap. It integrates three core components:

A. Adaptive Temperature Ladder Strategies

Parallel tempering runs multiple chains at different temperatures ($\beta = 1/T$) to facilitate mixing across multimodal landscapes. reddemcee implements a unified differential update rule to dynamically adjust the temperature ladder ($\beta_i$) during the run. It offers five interchangeable adaptation objectives:

1. Uniform Swap-Acceptance Rate (SAR): A classic method aiming for equal swap probabilities between adjacent chains.
2. Gaussian Area Overlap (GAO): Adjusts spacing based on the overlap of likelihood distributions between chains.
3. Swap Mean Distance (SMD): Novel. Maximizes the average Euclidean distance of accepted swaps between chains to improve information transfer in high-dimensional spaces.
4. Small Gaussian Gap (SGG): Novel. Uses a Gaussian approximation for small temperature gaps to maintain uniform swap rates based on energy variance.
5. Equalised Thermodynamic Length (ETL): Novel. Distributes chains uniformly along the "thermodynamic length," effectively placing more chains where the specific heat ($C_\nu$) peaks (i.e., where the posterior shape changes most rapidly).

B. Robust Evidence Estimators

To overcome the limitations of standard estimators (which assume independent samples and uniform spacing), reddemcee introduces three enhanced estimators:

1. Curvature-Aware Thermodynamic Integration (TI+): Improves standard TI by using Piecewise Cubic Hermite Interpolating Polynomials (PCHIP) to model the energy curve. It estimates discretization error by comparing coarse and fine ladder grids, explicitly accounting for the curvature of the log-likelihood.
2. Geometric-Bridge Stepping Stones (SS+): Enhances the Stepping Stones method by using geometric-bridge sampling (Meng & Wong 1996). Instead of estimating ratios using only one distribution, it utilizes samples from both ends of the temperature gap, reducing variance and improving numerical stability.
3. Hybrid Algorithm (H+): A novel approach that splits the temperature ladder. It applies TI in the "prior-dominated" region (low $\beta$) where the distribution is broad, and SS in the "posterior-dominated" region (high $\beta$) where the distribution is sharply peaked. The transition point is determined by analyzing the dominance of energy slope vs. temperature gap.

C. Error Quantification

The sampler accounts for the autocorrelation inherent in MCMC and the coupling introduced by swap moves. It replaces standard IID-based error estimates with autocorrelation-adjusted Monte Carlo standard errors using Overlapping Batch Means (OBM), ensuring realistic uncertainty bars for the evidence.

3. Key Contributions

• Unified Framework: The first implementation to couple five distinct ladder adaptation strategies with three advanced evidence estimators in a single package.
• Novel Algorithms: Introduction of SMD, SGG, and ETL adaptation strategies, and the H+ hybrid evidence estimator.
• Bias Reduction: Development of TI+ and SS+ to correct for discretization errors and variance inflation common in standard TI and SS when using few temperatures.
• Automation: The system requires no manual tuning of the temperature ladder; hyper-parameters are fixed across diverse problems (from 2D to 21D).

4. Results and Benchmarks

The performance of reddemcee was benchmarked against dynesty (a leading DNS implementation) on synthetic and real-world datasets.

• Synthetic Benchmarks:

  • n-D Gaussian Shells (up to 15D): reddemcee achieved roughly 7 times the effective sampling speed (measured in "kenits" – thousands of effective samples per second) of the best DNS configuration. The SMD adaptation strategy proved superior in high dimensions (15D), outperforming SAR by a significant margin.
  • Gaussian Egg-Box (Multimodal): reddemcee successfully identified all modes and provided accurate evidence estimates, whereas some DNS configurations missed modes or required significantly more computational time.
  • Hybrid Rosenbrock (High Correlation): The sampler handled the narrow, curved valleys effectively, with the H+ estimator achieving the highest accuracy ($\Delta \ln Z < 0.02\%$).

• Evidence Accuracy:

  • Across all tests, the TI+, SS+, and H+ estimators recovered evidence values within $|\Delta \ln Z| \lesssim 3\%$ of the analytical truth.
  • Crucially, the estimated error bars closely tracked the actual run-to-run dispersion, unlike DNS methods which often underestimated uncertainty (ratios of estimated/empirical error $\ll 1$).

• Real-World Application (HD 20794):

  • Applied to radial velocity data of the exoplanet system HD 20794 (3 confirmed planets + 1 candidate).
  • reddemcee reproduced literature model rankings and yielded tighter yet consistent planetary parameters compared to dynesty.
  • It provided evidence uncertainties that were realistic (ratio $\approx 0.63$ for the 3-planet model), whereas dynesty severely underestimated uncertainty (ratio $\approx 0.07$).

5. Significance

The paper demonstrates that Adaptive Parallel Tempering (APT), when combined with modern adaptation strategies and robust evidence estimators, can match or surpass Dynamic Nested Sampling in both sampling throughput and evidence accuracy.

• Closing the Gap: reddemcee allows researchers to obtain the rich posterior information density of MCMC and the rigorous model comparison capabilities of Nested Sampling in a single workflow.
• Scalability: The SMD and ETL strategies show that APT scales exceptionally well to high-dimensional problems (15+ dimensions) where traditional geometric spacing fails.
• Practicality: The ability to use a single set of hyper-parameters for models ranging from simple noise-only to complex multi-planet systems makes it a highly practical tool for modern astrophysical inference.

In conclusion, reddemcee represents a significant advancement in Bayesian inference tools, offering a fast, automated, and statistically rigorous alternative to nested sampling for complex astrophysical problems.