9.6 The Metropolis–Hastings algorithm
9.6.1 Finding an invariant distribution
This whole section is largely based on the clear expository article by Chib and Greenberg (1995). A useful general review which covers Bayesian integration problems more generally can be found in Evans and Swartz (1995–1996; 2000).
A topic of major importance when studying the theory of Markov chains is the determination of conditions under which there exists a stationary distribution (sometimes called an invariant distribution), that is, a distribution such that
In the case where the state space is continuous, the sum is, of course, replaced by an integral.
In Markov Chain Monte Carlo methods, often abbreviated as MCMC methods, the opposite problem is encountered. In cases where such methods are to be used, we have a target distribution in mind from which we want to take samples, and we seek transition probabilities for which this target density is the invariant density. If we can find such probabilities, then we can start a Markov chain at an arbitrary starting point and let it run for a long time, after ...