On the correlation structure of unilateral AR processes on the plane (Q4507948)

The stationary Gaussian process \(\mathbf X\) indexed on \(\mathbb Z^2\) is a quarter-plane AR\((p,q)\) process [denoted \(\text{QPAR}(p,q)\) process] if it results from filtering a Gaussian white noise through a quarter-plane filter of order \((p,q)\). Let \(\text{Rect}_{ij}^{kl}\) denote the rectangle in \(\mathbb Z^2\) delimited by \((i,j)\) and \((k,l)\). Let \(\mathbf r=(r_{mn})_{-p,-q}^{pq}\) be a positive definite correlation sequence. Then \(\mathbf r\) fulfills a Pickard \((p,q)\) property if the density of a stationary Gaussian process \(\mathbf X\) indexed on \(\text{Rect}_{00}^{pq}\) satisfies \(p(\mathbf x_B\mid \mathbf x_A, \mathbf x_C)=p(\mathbf x_B\mid \mathbf x_A)\), where \(A= \text{Rect}_{00}^{p-1,q-1}\), \(B=\text{Rect}_{0q}^{p-1,q}\), \(C= \text{Rect}_{p0}^{p,q-1}\). The authors show that QPARs form a subclass of stationary Markov random fields on a finite lattice. Then they prove that QPARs enjoy an extended Pickard property. This implies that QPARs fulfill the correlation matching property and the maximum entropy property. In a two-dimensional case, Yule-Walker equations contain more correlation parameters than available equations, but the extended Pickard property provides the missing equations and this leads to a one-to-one mapping between AR and correlation parameters. Two numerical methods for computing correlations are presented. Finally, the results are extended to unilateral AR processes.