The Efficiency of Geometric Samplers for Exoplanet Transit Timing Variation Models

TL;DR

This paper evaluates the efficiency of various geometric MCMC samplers in modeling exoplanet transit timing variations, demonstrating that covariance-informed transformations significantly enhance sampling performance across different models.

Contribution

It systematically compares multiple geometric MCMC samplers for TTV models, highlighting the importance of covariance-based transformations for improved efficiency.

Findings

Transforming parameter space using covariance estimates improves sampler efficiency.

Differential Evolution Monte Carlo and Geometric adaptive Monte Carlo perform consistently well.

Hamiltonian Monte Carlo excels for near-Gaussian posteriors.

Abstract

Transit timing variations (TTVs) are a valuable tool to determine the masses and orbits of transiting planets in multi-planet systems. TTVs can be readily modeled given knowledge of the interacting planets' orbital configurations and planet-star mass ratios, but such models are highly nonlinear and difficult to invert. Markov chain Monte Carlo (MCMC) methods are often used to explore the posterior distribution for model parameters, but, due to the high correlations between parameters, nonlinearity, and potential multi-modality in the posterior, many samplers perform very inefficiently. Therefore, we assess the performance of several MCMC samplers that use varying degrees of geometric information about the target distribution. We generate synthetic datasets from multiple models, including the TTVFaster model and a simple sinusoidal model, and test the efficiencies of various MCMC samplers. We find that sampling efficiency can be greatly improved for all models by sampling from a parameter space transformed using an estimate of the covariance and means of the target distribution. No one sampler performs the best for all datasets, but several samplers, such as Differential Evolution Monte Carlo and Geometric adaptive Monte Carlo, have consistently efficient performance. For datasets with near Gaussian posteriors, Hamiltonian Monte Carlo samplers with 2 or 3 leapfrog steps obtained the highest efficiencies. Based on differences in effective sample sizes per time, we show that the right choice of sampler can improve sampling efficiencies by several orders of magnitude.