Model-building problem of periodically correlated \(m\)-variate moving average processes (Q1268004)

An \(m\)-dimensional process \(X_t\) satisfying a linear stochastic difference equation of order \(q\), \(X_t= \sum^q_{j=0} \Theta_{t,j} \varepsilon_{t-j}\), where \(\varepsilon_t\) is an uncorrelated process with mean zero and matrix variance \(\Sigma_t\), is called a periodic (of period \(d)\) \(m\)-dimensional moving average model of order \(q\) if for some \(d\) it holds that \(\Sigma_{t+kd}= \Sigma_t\) and \(\Theta_{t+kd,j} =\Theta_{t,j}\) \(j=0,1, \dots,q\). The paper shows that for a given periodic autocovariance function of such a process, there are two classes (that may reduce to one class) of periodic (equivalent) models, and any other model is not periodic, but is asymptotically periodic. The matrix coefficients of the periodic models are given in terms of limits of some periodic matrix continued fractions; they are particular solutions of some prospective or retrospective recursion equations, and a procedure is given to calculate these limits. Numerical examples are given for periodically correlated univariate one- and two-dependent processes.