Tutorial on Markov Chain Monte Carlo.pdf

Tutorial on Markov Chain Monte Carlo Kenneth M. Hanson Los Alamos National Laboratory Presented at the 29th International Workshop on Bayesian Inference and Maximum Entropy Methods in Science and Technology, Gif-sur-Yvette, France, July 8 – 13, 2000 This presentation available at /kmh/talks/ July, 2000 - Revised 14/05/08 Bayesian and MaxEnt Workshop 1 LA-UR-05-5680 Acknowledgements • MCMC experts – Julian Besag, Jim Guberantus, John Skilling, Malvin Kalos • General discussions – Greg Cunningham, Richard Silver July, 2000 Bayesian and MaxEnt Workshop 2 Problem statement • Parameter space of n dimensions represented by vector x • Given an “arbitrary” target probability density function (pdf), q(x), draw a set of samples {xk } from it • Only requirement typically is that, given x, one be able to evaluate Cq(x), where C is an unknown constant – MCMC algorithms do not typically require knowledge of the normalization constant of the target pdf; from now on the multiplicative constant C will not be made explicit • Although focus here is on continuous variables, MCMC can be applied to discrete variables as well July, 2000 Bayesian and MaxEnt Workshop 3 Uses of MCMC • Permits evaluation of the expectation values of functions of x, e.g., 〈f (x)〉 = ∫f (x) q(x) dx ≅ (1/K) Σ f (x ) k k 2 – typical use is to calculate mean 〈x〉 and variance 〈(x - 〈x〉) 〉 • Also useful for evaluating integrals, such as the partition function for