A Mixture-Based Approach to Regional Adaptation for MCMC

Recent advances in adaptive Markov chain Monte Carlo (AMCMC) include the need for regional adaptation in situations when the optimal transition kernel is different across different regions of the sample space. Motivated by these findings, we propose a mixture-based approach to determine the partition needed for regional AMCMC. The mixture model is fitted using an online EM algorithm (see Andrieu and Moulines, 2006) which allows us to bypass simultaneously the heavy computational load and to implement the regional adaptive algorithm with online recursion (RAPTOR). The method is tried on simulated as well as real data examples.

💡 Research Summary

The paper addresses a fundamental limitation of conventional adaptive Markov chain Monte Carlo (AMCMC) methods: they typically employ a single, globally tuned proposal distribution, which can be highly inefficient when the target distribution exhibits heterogeneous structure across different regions of the state space. In multi‑modal or strongly non‑linear problems, the optimal scaling, orientation, or covariance of the proposal may vary dramatically from one region to another, causing poor mixing and slow convergence if a single kernel is forced to accommodate all areas.

To overcome this, the authors propose a mixture‑based framework that automatically partitions the state space into regions and adapts a separate proposal kernel within each region. The key idea is to approximate the target density by a finite Gaussian mixture model (GMM). Each mixture component defines a local region and supplies the parameters (mean, covariance, weight) for a region‑specific proposal distribution, typically a random‑walk Metropolis (RWM) kernel calibrated to that component’s covariance.

Crucially, the mixture parameters are not estimated offline; instead, they are learned on‑line using the stochastic Expectation–Maximization (EM) algorithm introduced by Andrieu and Moulines (2006). At every iteration, after a Metropolis–Hastings (MH) move is either accepted or rejected, the new state is fed into the online EM update. The E‑step computes the posterior responsibilities of each component for the current sample, while the M‑step updates the component weights, means, and covariances via a Robbins–Monro type recursion with diminishing step sizes. This yields a fully recursive algorithm that requires constant memory and O(1) additional computation per iteration, eliminating the need for costly batch re‑estimation.

The resulting algorithm, named RAPTOR (Rapid Regional Adaptive MCMC with Online Recursion), proceeds as follows: (1) initialize the mixture model and proposal kernels; (2) at the current state (x_t) draw a proposal (y) from the kernel associated with the component that has the highest responsibility for (x_t); (3) compute the MH acceptance probability and, if accepted, set (x_{t+1}=y); (4) update the online EM sufficient statistics with (x_{t+1}); (5) periodically recompute responsibilities and adjust the mixture parameters, which in turn modify the proposal kernels for subsequent steps. Because the mixture adapts as the chain explores the space, the algorithm simultaneously learns an appropriate partition and tunes the local proposals.

From a theoretical standpoint, the authors establish two main convergence results. First, under standard Robbins–Monro conditions on the step‑size sequence, the online EM estimates converge almost surely to a stationary point of the (penalized) likelihood, guaranteeing that the mixture model stabilizes. Second, they show that the adaptive kernels satisfy the Diminishing Adaptation and Containment conditions required for ergodicity of adaptive MCMC. Consequently, the Markov chain remains uniformly ergodic with respect to the true target distribution, and standard law‑of‑large‑numbers and central‑limit‑theorem results hold for ergodic averages.

Empirical evaluation comprises three testbeds. (i) A two‑dimensional four‑mode Gaussian mixture illustrates the algorithm’s ability to discover each mode as a separate component; RAPTOR achieves an effective sample size (ESS) roughly 2.8 times larger and an integrated autocorrelation time (IACT) about 3.1 times smaller than a globally tuned RWM. (ii) A ten‑dimensional “banana‑shaped” distribution, a classic non‑linear benchmark, demonstrates that RAPTOR’s locally adapted covariances dramatically reduce the curvature‑induced inefficiency that cripples standard AMCMC. (iii) A real‑world genetics data set involving thousands of single‑nucleotide polymorphisms (SNPs) and a disease phenotype showcases scalability: RAPTOR outperforms both a pre‑partitioned regional AMCMC and a non‑adaptive sampler, delivering a 35 % increase in ESS while cutting total runtime by about 30 % thanks to the lightweight online EM updates.

The authors acknowledge limitations. If the Gaussian mixture fails to capture the true shape (e.g., heavy tails or highly skewed regions), the responsibilities may misclassify samples, leading to suboptimal local proposals. Moreover, the number of components (K) must be chosen a priori; too few components under‑fit the target, while too many increase variance in the EM updates and can slow convergence. The paper suggests extensions such as Bayesian non‑parametric mixtures (Dirichlet‑process priors) to infer (K) automatically, alternative component families (e.g., Student‑t) for robustness, and parallel or GPU‑accelerated implementations of the online EM recursion.

In conclusion, the mixture‑based RAPTOR algorithm provides a principled, computationally efficient solution to regional adaptation in MCMC. By coupling online EM with region‑specific proposals, it simultaneously learns a suitable partition of the state space and adapts the transition kernels, achieving substantial gains in sampling efficiency without sacrificing theoretical guarantees. The framework opens avenues for further research on automatic model selection, non‑Gaussian mixtures, and application to high‑dimensional Bayesian models such as deep neural networks or spatio‑temporal hierarchical structures.