Pseudolikelihood for exponential family models of spatial point processes (Q1175969)

Let \(N\) be a finite point process on a base measure space \((S,\delta,\lambda)\), where \(\lambda\) is a diffuse measure, and let \(P\) be the probability under which \(N\) is a Poisson process with mean measure \(\lambda\). The authors consider a parametric statistical model for \(N\), with densities \(f_ \theta(\mu)\), \(\theta\in\Theta\) satisfying suitable regularity conditions. In a number of situations, these densities have the form \(f_ \theta(x)=g(x,\theta)/Z(\theta)\), where the function \(g\) is tractable, but the partition function (normalizing constant) \(Z(\theta)\) is impossible to calculate. The authors propose as an alternative the pseudolikelihood \[ PL_ T(\theta,\mu)=e^{- \lambda(T)}\lim_{i\to\infty}\prod_{j=1}^{m_ i} f_ \theta(\mu_{A_{ij}}\mid\mu_{S\backslash A_{ij}}), \] where \(T\) is a fixed subset of \(S\), the \(A_{ij}\) are a null array of partitions of \(T\) and the conditional density \(f_ \theta(\mu_ A\mid\mu_{S\backslash A})\) is defined by \[ f_ \theta(\mu_ A\mid\mu_{S\backslash A})=f_ \theta(\mu_ A+\mu_{S\backslash A})/\int f_ \theta(\eta+\mu_{S\backslash A})P_ A(d\eta). \] The main attraction of the pseudolikelihood is that, as a consequence of the conditioning, the partition function does not appear in it. The pseudolikelihood function is calculated for a general class of spatial point processes, and some of its properties established. In particular, maximum pseudolikelihood estimators are shown to be consistent for the class of finite range Markov models on \(\mathbb{R}^ d\), where the sense of the consistency is that of observation over bounded sets increasing to \(\mathbb{R}^ d\).