Gaussian conditional autoregressions have been widely used in spatial
statistics and Bayesian image analysis, where they are intended to des
cribe interactions between random variables at fixed sites in Euclidea
n space. The main appeal of these distributions is in the Markovian in
terpretation of their full conditionals. Intrinsic autoregressions are
limiting forms that retain the Markov property. Despite being imprope
r, they can have advantages over the standard autoregressions, both co
nceptually and in practice. For example, they often avoid difficulties
in parameter estimation, without apparent loss, or exhibit appealing
invariances, as in texture analysis. However, on small arrays and in n
onlattice applications, both forms of autoregression can lead to undes
irable second-order characteristics, either in the variables themselve
s or in contrasts among them. This paper discusses standard and intrin
sic autoregressions and describes how the problems that arise can be a
lleviated using Dempster's (1972) algorithm or an appropriate modifica
tion. The approach represents a partial synthesis of standard geostati
stical and Gaussian Markov random field formulations. Some nonspatial
applications are also mentioned.