No label defined (Q4527097)

In the introduction some general remarks are presented. Markov chain Monte Carlo (MCMC) simulates a probability distribution \(\pi\) by running a Markov chain \(X_1,X_2,\dots\) having \(\pi\) as its unique stationary distributions. An expectation \[ \mu=E_\pi \bigl\{f(x) \bigr\}=\int f(x)\pi(dx) \tag{1} \] is approximated by the sample mean over the path of the Markov chain \[ \widehat\mu_n= {1\over n}\bigl(f(x_1) +\cdots+f (x_n)\bigr).\tag{2} \] This is similar to the classical independent sample Monte Carlo (ISMC), which uses a sample mean of the same form as (2) but with \(X_1,X_2,\dots\) independent and identically distributed \(\pi\). The law of large numbers (LLN) and the central limit theorem (CLT) hold. They guarantee that \(\widehat\mu_n\) converges almost surely to \(\mu\) and \(\sqrt n(\widehat\mu_n -\mu)\) converges in law to some normal random variable. This justifies the use of \(\widehat\mu_n\) as an approximation of \(\mu\) and provides an estimate of the approximation error, just as in ISMC. The difference between MCMC and ISMC is that the variance in CLT is typically larger, sometimes much larger in MCMC than in ISMC. An MCMC algorithm is characterized by its transition probability kernel \(P\) defined by \[ P(x,A)= P(X_{n+1}\in A\mid X_n=x). \] A kernel \(P\) has a probability measure \(\pi\) as its stationary distribution if \(\pi(A)=\int P(x,A)\pi(dx)\) holds for all measurable sets \(A\). The authors use Harris recurrent kernels. In the second part of the note preliminary Hilbert space theory is presented and in the third part a necessary and sufficient condition for efficiency is given. In the fourth part the authors construct the inverse. So, they consider Markov chains on finite state spaces and explain how to construct a generalized inverse that represents the inverse Laplacian. In the fifth section the non-reversible Markov chain introduced by P. Diaconis, S. Holmes and R. M. Neal (1997) is analyzed. In the sixth section comparing the performance of reversible and non-reversible kernels is presented. In the last part entitled ``Discussion'' interesting conclusions for non-reversible Markov chains are considered. In this sense a particular example of Diaconis et al. (1997) applying the criterion of asymptotic efficiency is reanalyzed. For other details see the authors' references. For the entire collection see [Zbl 0951.00044].