🤖 AI Summary
Bayesian inference for doubly-intractable models (e.g., Ising, Kent distributions) is hindered by intractable normalizing constants in both the likelihood and posterior, rendering exact posterior computation impossible; conventional pseudo-marginal MCMC further suffers from the restrictive non-negativity requirement on the likelihood estimator. To address this, we propose a signed pseudo-marginal Metropolis–Hastings algorithm. Its core innovation is the first unbiased block-Poisson estimator that admits negative values, integrated with importance sampling correction and correlated random number techniques to circumvent the non-negativity constraint. We also derive analytical heuristic guidelines for tuning its hyperparameters. Experiments on the Ising model and spherical Kent distribution demonstrate substantial gains in effective sample size per second (ESS/sec), while preserving posterior mean consistency. This work establishes a new paradigm for efficient Bayesian inference in doubly-intractable settings.
📝 Abstract
Doubly intractable models are encountered in a number of fields, e.g. social networks, ecology and epidemiology. Inference for such models requires the evaluation of a likelihood function, whose normalising function depends on the model parameters and is typically computationally intractable. The normalising constant of the posterior distribution and the additional normalising function of the likelihood function result in a so-called doubly intractable posterior, for which it is difficult to directly apply Markov chain Monte Carlo (MCMC) methods. We propose a signed pseudo-marginal Metropolis-Hastings (PMMH) algorithm with an unbiased block-Poisson estimator to sample from the posterior distribution of doubly intractable models. As the estimator can be negative, the algorithm targets the absolute value of the estimated posterior and uses an importance sampling correction to ensure simulation consistent estimates of the posterior mean of any function. The advantages of our estimator over previous approaches are that its form is ideal for correlated pseudo-marginal methods which are well known to dramatically increase sampling efficiency. Moreover, we develop analytically derived heuristic guidelines for optimally tuning the hyperparameters of the estimator. We demonstrate the algorithm on the Ising model and a Kent distribution model for spherical data.