Coordinate selection rules for Gibbs sampling
Fishman, George S.
Ann. Appl. Probab., Tome 6 (1996) no. 1, p. 444-465 / Harvested from Project Euclid
This paper studies several different plans for selecting coordinates for updating via Gibbs sampling. It exploits the inherent features of the Gibbs sampling formulation, most notably its neighborhood structure, to characterize and compare the plans with regard to convergence to equilibrium and variance of the sample mean. Some of the plans rely completely or almost completely on random coordinate selection. Others use completely or almost completely deterministic coordinate selection rules. We show that neighborhood structure induces idempotency for the individual coordinate transition matrices and commutativity among subsets of these matrices. These properties lead to bounds on eigenvalues for the Gibbs sampling transition matrices corresponding to several of the plans. For a frequently encountered neighborhood structure, we give necessary and sufficient conditions for a commonly employed deterministic coordinate selection plan to induce faster convergence to equilibrium than the random coordinate selection plans. When these conditions hold, we also show that this deterministic selection rule achieves the same worst-case bound on the variance of the sample mean as that arising from the random selection rules when the number of coordinates grows without bound. This last result encourages the belief that faster convergence for the deterministic selection rule may also imply a variance of the sample mean no larger than that arising for random selection rules.
Publié le : 1996-05-14
Classification:  Coordinate selection,  Gibbs sampling,  Markov random field,  Monte Carlo,  60J05,  62E25
@article{1034968139,
     author = {Fishman, George S.},
     title = {Coordinate selection rules for Gibbs sampling},
     journal = {Ann. Appl. Probab.},
     volume = {6},
     number = {1},
     year = {1996},
     pages = { 444-465},
     language = {en},
     url = {http://dml.mathdoc.fr/item/1034968139}
}
Fishman, George S. Coordinate selection rules for Gibbs sampling. Ann. Appl. Probab., Tome 6 (1996) no. 1, pp.  444-465. http://gdmltest.u-ga.fr/item/1034968139/