This preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: ISyE8843A, Brani Vidakovic Handout 10 1 MCMC Methodology. Independence of X 1 ,...,X n is not critical for an approximation of the form E θ  x h ( X ) = 1 n ∑ n i =1 h ( X i ) , X i ∼ π ( θ  x ) . In fact, when X ’s are dependent, the ergodic theorems describe the approximation. An easy and convenient form of dependence is Markov chain dependence. The Markov dependence is perfect for computer simulations since for producing a future realization of the chain, only the current state is needed. 1.1 Theoretical Background and Notation Random variables X 1 ,X 2 ,...,X n ,... constitute a Markov Chain on continuous state space if they possess a Markov property, P ( X n +1 ∈ A  X 1 ,...,X n ) = P ( X n +1 ∈ A  X 1 ,...,X n ) = Q ( X n ,A ) = Q ( A  X n ) , for some probability distribution Q. Typically, Q is assumed a timehomogeneous, i.e., independent on n (“time”). The transition (from the state n to the state n +1 ) kernel defines a probability measure on the state space and we will assume that the density q exists, i.e., Q ( A  X n = x ) = Z A q ( x,y ) dy = Z A q ( y  x ) dy. Distribution Π is invariant, if for all measurable sets A Π( A ) = Z Q ( A  x )Π( dx ) . If the transition density π exists, it is stationary if q ( x  y ) π ( y ) = q ( y  x ) π ( x ) . Here and in the sequel we assume that the density for Π exists, Π( A ) = R A π ( x ) dx. A distribution Π is an equilibrium distribution if for Q n ( A  x ) = P ( X n ∈ A  X = x ) , lim n →∞ Q n ( A  x ) = Π( A ) . In plain terms, the Markov chain will forget the initial distribution and will converge to the stationary distri bution. The Markov Chain is irreducible if for each A for which Π( A ) > , and for each x , one can find n , so that Q n ( A  x ) > . The Markov Chain X 1 ,...,X n ,... is recurrent if for each B such that Π( B ) > , P ( X n ∈ B i.o.  X = x ) = 1 , a.s. ( in distribution of X ) It is Harris recurrent if P ( X n ∈ B i.o.  X = x ) = 1 , ( ∀ x ) . The acronym i.o. stands for infinitely often. 1 Figure 1: Nicholas Constantine Metropolis, 19151999 1.2 Metropolis Algorithm Metropolis algorithm is the fundamental to MCMC development. Assume that the target distribution is known up to a normalizing constant. We would like to construct a chain with π as its stationary distribution. As in ARM, we take a proposal distribution q ( x,y ) = q ( y  x ) , where the proposal for a new value of a chain is y , given that the chain is at value x . Thus q defines transition kernel Q ( A,x ) = R A q ( y  x ) dx which is the probability of transition to some y ∈ A. Detailed Balance Equation. A Markov Chain with transition density q ( x,y ) = q ( y  x ) satisfies detailed balance equation if there exists a distribution f such that q ( y  x ) f ( x ) = q ( x  y ) f ( y ) . (1) The distribution f is stationary (invariant) and the chain is reversible....
View
Full Document
 Spring '11
 VIDAKOVIC
 Markov chain, Markov chain Monte Carlo, Gibbs Sampling, Gibbs sampler

Click to edit the document details