Skip to content

Monte Carlo sampling methods using Markov chains and their applications

Why this mattered

Hastings’s 1970 paper turned the Metropolis algorithm from a specialized computational trick into a general method for sampling from complex probability distributions. The key shift was the acceptance rule for asymmetric proposal distributions, which made it possible to design Markov chains whose stationary distribution was a chosen target distribution even when direct sampling was impractical. This moved Monte Carlo from relying mainly on independent samples from tractable distributions toward a broader paradigm: construct a stochastic process that eventually visits states in the right proportions.

That generalization was especially important for statistics, where posterior distributions, likelihood-based models, and high-dimensional integrals often cannot be normalized or sampled from directly. Hastings showed that useful computation could proceed using only ratios of target densities, avoiding the need to know the normalizing constant. This made a large class of Bayesian and likelihood problems computationally approachable, not by simplifying the model, but by changing what counted as a feasible numerical method.

The paper became foundational for Markov chain Monte Carlo. Later breakthroughs such as the Gibbs sampler’s adoption in mainstream Bayesian statistics, simulated annealing, hierarchical Bayesian modeling, and modern probabilistic computation all depend on the same central idea: inference can be performed by building a Markov chain with the desired distribution as its equilibrium law. Hastings did not solve all practical difficulties, and the paper explicitly treated error assessment and convergence as hard issues, but it supplied the general framework that made those difficulties worth studying.

Abstract

A generalization of the sampling method introduced by Metropolis et al. (1953) is presented along with an exposition of the relevant theory, techniques of application and methods and difficulties of assessing the error in Monte Carlo estimates. Examples of the methods, including the generation of random orthogonal matrices and potential applications of the methods to numerical problems arising in statistics, are discussed.

Sources