Estimation and annealing for Gibbsian fields (Q1106598)

Consider a Gibbs distribution \(\pi_{\theta}(x)=\exp (-<\alpha (x),\theta >)Z(\theta)^{-1}\) on a finite configuration space. The maximum likelihood estimator \(\theta_*\) is given as the solution of \(E_{\theta}[\alpha]=\alpha (x)\). A direct evaluation of \(E_{\theta}[\alpha]\) or a calculation of \(\theta_*\) by the Robbins- Monro algorithm is not feasible here because no simple simulation method of \(\pi_{\theta}\) is available. The very nice main idea of the paper is to use a stochastic gradient algorithm with the so-called Gibbs-sampler which is an inhomogeneous Markov chain \(P_{\theta}^{n,n+1}\) converging weakly to \(\pi_{\theta}\). It is shown hat the sequence of iterates \(\theta\) (n) of this algorithm converges asymptotically to \(\theta_*\) if the step size is not too small. Practical aspects of implementing this algorithm are also discussed. Finally it is shown that the annealing algorithm continues to work with \(\theta_*\) replaced by \(\theta\) (n), i.e. \(P^{n,n+1}_{\theta (n)/T(n)}\) converges weakly to the uniform distribution on \(\{x| <\alpha (x),\theta_*>\) is minimal\(\}\) if T(n)\(\to 0\) slowly. These results have applications in image processing.