By developing and applying a broad framework for rejection sampling using a
uxiliary randomness, we provide an extension of the perfect sampling algori
thm of Fill (1998a) to general chains on quite general state spaces, and de
scribe how use of bounding processes can ease computational burden. Along t
he way, we unearth a simple connection between the coupling from the past (
CFTP) algorithm originated by Propp and Wilson (1996) and our extension of
Fill's algorithm. (C) 2000 John Wiley & Sons, Inc.