Hostname: page-component-848d4c4894-jbqgn Total loading time: 0 Render date: 2024-07-08T02:22:07.573Z Has data issue: false hasContentIssue false

ON THE CONVERGENCE OF METROPOLIS-TYPE RELAXATION AND ANNEALING WITH CONSTRAINTS

Published online by Cambridge University Press:  10 October 2002

Marc C. Robini
Affiliation:
CREATIS, UMR CNRS 5515, INSA Lyon, Villeurbanne, France, E-mail: marc.robini@creatis.insa-lyon.fr
Yoram Bresler
Affiliation:
Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, Urbana, IL 61801, E-mail: ybresler@uiuc.edu
Isabelle E. Magnin
Affiliation:
CREATIS, UMR CNRS 5515, INSA Lyon, Villeurbanne, France

Abstract

We discuss the asymptotic behavior of time-inhomogeneous Metropolis chains for solving constrained sampling and optimization problems. In addition to the usual inverse temperature schedule (βn)n∈[hollow N]*, the type of Markov processes under consideration is controlled by a divergent sequence (θn)n∈[hollow N]* of parameters acting as Lagrange multipliers. The associated transition probability matrices (Pβn,θn)n∈[hollow N]* are defined by Pβ,θ = q(x, y)exp(−β(Wθ(y) − Wθ(x))+) for all pairs (x, y) of distinct elements of a finite set Ω, where q is an irreducible and reversible Markov kernel and the energy function Wθ is of the form Wθ = U + θV for some functions U,V : Ω → [hollow R]. Our approach, which is based on a comparison of the distribution of the chain at time n with the invariant measure of Pβn,θn, requires the computation of an upper bound for the second largest eigenvalue in absolute value of Pβn,θn. We extend the geometric bounds derived by Ingrassia and we give new sufficient conditions on the control sequences for the algorithm to simulate a Gibbs distribution with energy U on the constrained set [Ω with tilde above] = {x ∈ Ω : V(x) = minz∈ΩV(z)} and to minimize U over [Ω with tilde above].

Type
Research Article
Copyright
© 2002 Cambridge University Press

Access options

Get access to the full version of this content by using one of the access options below. (Log in options will check for institutional or personal access. Content may require purchase if you do not have access.)