Abstract
By mixing the target posterior distribution with a surrogate distribution, of which the normalizing constant is tractable, we propose a method for estimating the marginal likelihood using the Wang-Landau algorithm. We show that a faster convergence of the proposed method can be achieved via the momentum acceleration. Two implementation strategies are detailed: (i) facilitating global jumps between the posterior and surrogate distributions via the Multiple-try Metropolis; (ii) constructing the surrogate via the variational approximation. When a surrogate is difficult to come by, we describe a new jumping mechanism for general reversible jump Markov chain Monte Carlo algorithms, which combines the Multiple-try Metropolis and a directional sampling algorithm. We illustrate the proposed methods on several statistical models, including the Log-Gaussian Cox process, the Bayesian Lasso, the logistic regression, and the g-prior Bayesian variable selection.
Introduction
In Bayesian statistics, the Bayes factor serves as the primary tool for model comparison and selection. However, its calculation requires the marginal likelihood—the integral of the likelihood and prior over the entire parameter space. For complex, high-dimensional models, this integral is analytically intractable and numerically challenging. Traditional methods like importance sampling and bridge sampling often struggle when the proposal distribution does not overlap significantly with the target posterior. While path sampling and Sequential Monte Carlo (SMC) offer alternatives, they frequently require intensive tuning of tempering schedules.
This paper introduces a robust, general-purpose framework for estimating marginal likelihoods by mixing the target posterior with a tractable surrogate distribution. By leveraging an accelerated version of the Wang-Landau (WL) algorithm, the proposed method effectively navigates high-dimensional spaces and overcomes the "overlap" limitations inherent in earlier techniques.
The Wang-Landau Mixture Method
The Mixture Framework
The core innovation lies in the construction of a mixture distribution: $$\pi(\theta) = \gamma(\theta) + q(\theta)$$ where $\gamma(\theta)$ is the unnormalized target posterior and $q(\theta)$ is a surrogate distribution with a known normalizing constant. The goal is to estimate the ratio of the normalizing constants of these two components. Once this ratio is known, the marginal likelihood of the target posterior is easily recovered.
Adaptive Estimation via Wang-Landau
To estimate the relative weights of the mixture components, the authors employ the Wang-Landau algorithm, a technique originally developed in statistical physics. The algorithm adaptively penalizes the component currently being visited by the Markov chain. This iterative process eventually forces the chain to spend equal time in both the target and the surrogate, at which point the accumulated penalties reveal the ratio of their normalizing constants.
Momentum Acceleration
A significant contribution of this work is the reframing of the WL update as a stochastic gradient descent problem. By introducing a "momentum" term—a concept widely used in deep learning optimization—the authors significantly accelerate the convergence of the normalizing constant estimate. This "Accelerated Wang-Landau" (AWL) approach reduces the oscillations typically seen in standard WL implementations.
Implementation Strategies
Automated Surrogates via Variational Inference
For the mixture method to be efficient, the surrogate $q(\theta)$ should ideally reside in the same region of the parameter space as the target $\gamma(\theta)$. The authors propose using Variational Inference (VI), specifically Coordinate Ascent Variational Inference (CAVI), to automatically construct a surrogate that approximates the target. This reduces the need for manual tuning and makes the method more accessible for diverse statistical models.
Multiple-Try Metropolis (MTM)
To ensure the Markov chain can transition effectively between the target and surrogate distributions—especially when they are distinct or high-dimensional—the authors utilize the Multiple-Try Metropolis (MTM) algorithm. MTM proposes several candidate states and selects the most promising one, which facilitates "global jumps" and prevents the sampler from becoming trapped in a single component of the mixture.
Experimental Evaluation
The authors evaluated the AWL mixture method across several challenging benchmarks:
- Log-Gaussian Cox Process: In a spatial point pattern analysis involving 900 parameters, the WL mixture method demonstrated significantly lower variance in its estimates compared to traditional importance sampling.
- Bayesian Lasso: Using simulated data, the AWL reached stable estimates in approximately 2,000 iterations, whereas the standard WL algorithm failed to converge even after 10,000 iterations.
- Logistic Regression: The method was applied to the German Credit dataset (301 parameters), successfully estimating marginal likelihoods where simpler methods often falter due to dimensionality.
- Variable Selection: Using a g-prior model on the Pollution dataset, the authors' MTM-based Reversible Jump MCMC (RJMCMC) extension correctly identified true predictors, outperforming standard birth-and-death RJMCMC which tended to get stuck in local modes.
Conclusion
The Wang-Landau Mixture Method provides a powerful alternative for Bayesian model selection. By combining the adaptive nature of the WL algorithm with momentum acceleration and variational surrogates, the method offers a robust solution for high-dimensional marginal likelihood estimation. Its ability to handle distributions with minimal overlap and its improved convergence speed make it a valuable tool for researchers working with complex Bayesian models. Future research may expand this framework to include multi-component mixtures to further bridge the gap in highly multi-modal posterior landscapes.
Summary updated