BVG-MC: Do Markov chain simulation to sample from a bivariate Gaussian.
The bvg-mc program is the specialization of xxx-mc to the task of
sampling from a bivariate Gaussian. See xxx-mc.doc for the generic
features of this program.
The following applications-specific sampling procedures are implemented:
gibbs Does a Gibbs sampling update for all replications
of the bivariate Gaussian. A update consists of
sampling first from the conditional distribution for
the first component, and then from the conditional
distribution for the second component.
gibbs0 alpha Does an Adler-style overrelaxed update for the first
component (separately for all replications). The
new value of the offset of the component from its
conditional mean is alpha times the old offset plus
Gaussian noise of variance 1-alpha^2 times the
conditional variance.
gibbs1 alpha Does an Adler-style overrelaxed update for the second
component.
To do Adler-style successive overrelaxation, gibbs0 and gibbs1 should
be done alternately.
The inverse temperature used in tempering methods is interpreted in
the standard way, as a power to raise the (unnormalized) probability
density to, or equivalently, a factor to multiply the energy by.
The default dynamical stepsizes are all set to one, except that they
are appropriately scaled during tempering.
Copyright (c) 1995-2003 by Radford M. Neal