![where Ac[-1,1] and B>1. The cdf is graphically presented in Figure 2.](https://figures.academia-assets.com/94358666/figure_003.jpg)
580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can parallel and adaptive strategies improve exploration and convergence in Metropolis–Hastings algorithms?
This research direction investigates modifications to the classical Metropolis–Hastings (MH) algorithm aimed at enhancing sampling efficiency and mixing properties, particularly in challenging settings like high-dimensional or multimodal target distributions. Many works study the design and analysis of parallel MCMC schemes, adaptive proposal mechanisms, and multiple-try variants to balance local exploration and global traversal of the state space while maintaining ergodicity guarantees. Efficient use of computational resources via parallelism and adaptation helps overcoming slow mixing and convergence bottlenecks inherent in standard MH, especially when naive or fixed proposals are used.
Key finding: This paper systematically categorizes acceleration strategies for MCMC, differentiating between exploration-level methods (e.g., tempering, Hamiltonian Monte Carlo) and exploitation-level improvements (e.g.,... Read more
Key finding: Introduces orthogonal MCMC (O-MCMC), a novel parallel MCMC framework that combines multiple independent 'vertical' random-walk chains with 'horizontal' independent proposal-based MCMC techniques applied jointly over the... Read more
Key finding: Proposes an adaptive MCMC algorithm that overcomes incomplete adaptation issues observed in Adaptive Rejection Metropolis Sampling (ARMS) within Gibbs sampling. By improving the proposal adaptation to ensure convergence to... Read more
Key finding: Provides a theoretical framework demonstrating the flexibility in designing Multiple Try Metropolis (MTM) algorithms satisfying detailed balance, allowing multiple candidates per iteration from different proposals with... Read more
Key finding: Identifies scenarios where increasing the number of tries in standard MTM does not improve or even degrades performance, especially when using a single random-walk proposal. The work analyzes these mixing issues linking them... Read more
2. How can variance reduction through control variates and diffusion approximations enhance Metropolis–Hastings efficiency?
Research in this theme focuses on reducing estimator variance in MH algorithms and related gradient-based MCMC schemes by constructing control variates informed by solutions to Poisson equations, diffusion approximations, or related functional equations. Approaches include solving or approximating Poisson equations associated with Markov chains to build zero or low variance estimators, deploying diffusion limits (e.g., Langevin diffusion) for variance reduction, and exploiting gradient information while balancing robustness and computational cost. These methods improve efficiency by decreasing sample correlation and estimator variance without additional expensive sampling.
Key finding: Introduces a practical variance reduction framework for Random Walk Metropolis and Metropolis-adjusted Langevin algorithms based on approximate solutions of the Poisson equation associated with the target distribution.... Read more
Key finding: Develops control variates constructed via minimizing asymptotic variance in connection to the Langevin diffusion Poisson equation, enabling variance reduction in discrete-time MCMC estimators. The approach leverages... Read more
Key finding: Shows that spectral gaps of classical gradient-based MCMC methods like Langevin and Hamiltonian Monte Carlo decay exponentially fast with mismatches in scaling of proposal and target scales, indicating high sensitivity to... Read more
Key finding: Analyzes a broad class of first-order locally-balanced MH algorithms, including Langevin and Barker proposals, deriving an explicit limiting acceptance rate (~57%) and optimal scaling (n^{-1/3}) in high dimension under... Read more
3. How can Bayesian and scalable approaches adapt Metropolis–Hastings for big data and complex models?
This theme explores methodologies extending MH algorithms to handle modern computational challenges posed by big data and complex models. Strategies include replacing full data likelihood evaluations with stochastic or bootstrap approximations, employing parallel and distributed computing architectures, and incorporating specialized MH variants within broader scalable frameworks. The goal is to retain theoretical guarantees (e.g., ergodicity, unbiasedness) while making computations practically feasible and efficient over large datasets or intricate inference problems.
Key finding: Proposes the Bootstrap Metropolis-Hastings (BMH) algorithm that replaces full data log-likelihood evaluations with Monte Carlo averages computed from multiple bootstrap samples, enabling embarrassingly parallel computations... Read more
Key finding: Besides acceleration techniques at the algorithmic level, this work highlights scalable methods designed to reduce computational burden on large datasets by modifying proposals or exploiting Rao-Blackwellization, outlining... Read more
Key finding: Develops a Monte Carlo method yielding unbiased and finite-variance estimates of filtering distributions for discretely observed diffusion processes via a double randomization scheme incorporating multilevel particle filter... Read more