Notes About Markov Chain CLTs.pdfVIP

Notes About Markov Chain CLTs [Rough notes by Jeffrey S. Rosenthal, February 2007, based on very helpful conversations with J.P. Hobert, N. Madras, G.O. Roberts, and T. Salisbury. For discussion and clarification only – not for publication. Comments appreciated.] 1. Introduction. These notes concern various issues surrounding central limit theorems (CLTs) for Markov chains, important notably for MCMC algorithms. A number of other papers have discussed related matters ([8], [13], [5], [3], [6], [7]), and probably much of the discussion below is already known, but we wanted to write it up for our own clarification. Let π(·) be a probability measure on a measurable space (X ,F). Let P be a Markov chain operator reversible with respect to π(·). Write 〈f, g〉 = ∫X f(x) g(x)π(dx); by reversibility, 〈f, Pg〉 = 〈Pf, g〉. Let h : X → R be measurable, with π(h2) ∞ and (say) π(h) = 0. Let {Xn}∞n=0 follow the transitions P in stationarity, so L(Xn) = π(·) and P[Xn+1 ∈ A |Xn] = P (Xn, A) for all A ∈ F , for n = 0, 1, 2, . . .. Let γk = E[h(X0)h(Xk)] = 〈h, P kh〉. Let r(x) = P[X1 = x |X0 = x] for x ∈ X . Let E be the spectral measure (e.g. [12]) associated with P , so that f(P ) = ∫ 1 ?1 f(λ) E(dλ) for “all” analytic functions f : R → R, and also E(R) = I. Let Eh be the induced measure for h, viz. Eh(S) = 〈h, E(S)h〉 , S ? [?1, 1] Borel a positive Borel measure (cf. [5], p. 1753), which is finite if π(h2) ∞ since then Eh(R) = 〈h, E(R)h〉 = 〈h, h〉 = π(h2) ∞. We are interested in the question of whether/when a root-n CLT exists for h, meaning that n?1/2 ∑n i=1 h(Xi) converges weakly to Normal(0, σ 2) for some σ2 ∞. 2. Representations of the Variance. There are a number of possible formulae for σ2 in the literature (e.g. [8], [5], [3]), including: A = lim n→∞n ?1Var ( n∑ i=1 h(Xi) ) ; 1 B = 1 + 2 ∞∑ k=1 γk = 1 + 2 lim n→∞ n∑ k=1 γk ; C = ∫ 1 ?1 1 + λ 1? λ Eh(dλ) . It is proved in [8] that if C ∞, then a CLT exists for h (with σ2 = C). And, it is proved in [9] that if limn→∞ nE[h2(X0) r(