Abstract not found

A critical issue for users of Markov Chain Monte Carlo (MCMC) methods in applications is how to determine when it is safe to stop sampling and use the samples to estimate characteristics of the distribution of interest. Research into methods of computing theoretical convergence bounds holds promise for the future but currently has yielded relatively little that is of practical use in applied work. Consequently, most MCMC users address the convergence problem by applying diagnostic tools to the output produced by running their samplers. After giving a brief overview of the area, we provide an expository review of thirteen convergence diagnostics, describing the theoretical basis and practical implementation of each. We then compare their performance in two simple models and conclude that all the methods can fail to detect the sorts of convergence failure they were designed to identify. We thus recommend a combination of strategies aimed at evaluating and accelerating MCMC sampler conver...

this paper we investigate the convergence properties of the Gibbs sampler as applied to a particular hierarchical Bayes model. The model is related to James-Stein estimators (James and Stein, 1961; Efron and Morris, 1973, 1975; Morris, 1983). Briefly, James-Stein estimators may be defined as the mean of a certain empirical Bayes posterior distribution (as discussed in the next section). We consider the problem of using the Gibbs sampler as a way of sampling from a richer posterior distribution, as suggested by Jun Liu (personal communication). Such a technique would eliminate the need to estimate a certain parameter empirically and to provide a "guess" at another one, and would give additional information about the distribution of the parameters involved. We consider, in particular, the convergence properties of this Gibbs sampler. For a certain range of prior distributions, we establish (Section 3) rigorous, numerical, reasonable rates of convergence. The bounds are obtained using the methods of Rosenthal (1995b). We thus rigorously bound the running time for this Gibbs sampler to converge to the posterior distribution, within a specified accuracy (as measured by total variation distance). We provide a general formula for this bound, which is of reasonable size, in terms of the prior distribution and the data. This Gibbs sampler is perhaps the most complicated example to date for which reasonable quantitative convergence rates have been obtained. We apply our bounds to the numerical baseball data of Efron and Morris (1975) and Morris (1983), based on batting averages of baseball players, and show that approximately 140 iterations are sufficient to achieve convergence in this case. For a different range of prior distributions, we use the Submartingale Convergence Theo...

This paper analyzes the Gibbs sampler applied to a standard variance component model, and considers the question of how many iterations are required for convergence. It is proved that for K location parameters, with J observations each, the number of iterations required for convergence (for large K and J) is a constant times

. I describe a simple procedure for investigating the convergence properties of Markov Chain Monte Carlo sampling schemes. The procedure employs multiple runs from a sampler, using the same random deviates for each run. When the sample paths from all sequences converge, it is argued that approximate equilibrium conditions hold. The procedure also provides a simple diagnostic for detecting modes in multimodal posteriors. Several examples of the procedure are provided. In Ising models, the relation between the correlation parameter and the convergence rate of rudimentary Gibbs samplers is investigated. In another example, the effects of multiple modes on the convergence of coupled paths are explored using mixtures of bivariate normal distributions. The technique is also used to evaluate the convergence properties of a Gibbs sampling scheme applied to a model for rat growth rates (Gelfand et al 1990). Acknowledgements I would like to thank Steve MacEachern, Julian Besag, Donald Rubin, A...

Markov chain Monte Carlo (MCMC) methods, including the Gibbs sampler and the Metropolis-Hastings algorithm, are very commonly used in Bayesian statistics for sampling from complicated, high-dimensional posterior distributions. A continuing source of uncertainty is how long such a sampler must be run in order to converge approximately to its target stationary distribution. Rosenthal (1995b) presents a method to compute rigorous theoretical upper bounds on the number of iterations required to achieve a specified degree of convergence in total variation distance by verifying drift and minorization conditions. We propose the use of auxiliary simulations to estimate the numerical values needed in Rosenthal's theorem. Our simulation method makes it possible to compute quantitative convergence bounds for models for which the requisite analytical computations would be prohibitively difficult or impossible. On the other hand, although our method appears to perform well in our example problems...

This paper is organised as follows. In Section 2, we present an over-simplified version of a convergence diagnostic, and study analytically its performance on certain simple Markov chains. We restrict ourselves primarily to chains which in fact produce i.i.d. samples from

. A standard method for handling Bayesian models is to use Markov chain Monte Carlo methods to draw samples from the posterior. We demonstrate this method on two core problems in computer vision---structure from motion and colour constancy. These examples illustrate a samplers producing useful representations for very large problems. We demonstrate that the sampled representations are trustworthy, using consistency checks in the experimental design. The sampling solution to structure from motion is strictly better than the factorisation approach, because: it reports uncertainty on structure and position measurements in a direct way; it can identify tracking errors; and its estimates of covariance in marginal point position are reliable. Our colour constancy solution is strictly better than competing approaches, because: it reports uncertainty on surface colour and illuminant measurements in a direct way; it incorporates all available constraints on surface reflectance and on illumination in a direct way; and it integrates a spatial model of reflectance and illumination distribution with a rendering model in a natural way. One advantage of a sampled representation is that it can be resampled to take into account other information. We demonstrate the effect of knowing that, in our colour constancy example, a surface viewed in two different images is in fact the same object. We conclude with a general discussion of the strengths and weaknesses of the sampling paradigm as a tool for computer vision. Keywords: Markov chain Monte Carlo, colour constancy, structure from motion 1.

Quantitative science requires the assessment of uncertainty, and this means that measurements and inferences should be described as probability distributions. This is done by building data into a probabilistic likelihood function which produces a posterior “answer ” by modulating a prior “question”. Probability calculus is the only way of doing this consistently, so that data can be included gradually or all at once while the answer remains the same. But probability calculus is only a language: it does not restrict the questions one can ask by setting one’s prior. We discuss how to set sensible priors, in particular for a large problem like image reconstruction. We also introduce practical modern algorithms (Gibbs sampling, Metropolis algorithm, genetic algorithms, and simulated annealing) for computing probabilistic inference.

this paper of writing moment conditions as a sequence of conditional moments was motivated by the work of Michael Keane on discrete panel data. The step of utilizing increasing sequences of conditional moments was suggested by work of Paul Ruud on adaptive search algorithms. Key elements in the proof of the asymptotic efficiency of the partitioning algorithm were provided by Peter Bickel, Whitney Newey, and Keunkwan Ryu. A number of the results collected here were first presented at the Rotterdam Conference on Simulation Estimators, June 1991