Optimality in the estimation of a MA system from a long AR model for simulation studies

The determination of moving average (MA) models from a prior autoregressive (AR) approximation of a specified (target) spectral matrix is addressed; this is done in context with the need to simulate ground shaking and other natural phenomena as multivariate random processes. First, an existing technique based on a direct modeling of the target expression is revisited. In this regard, the influence of the order of the prior AR approximation, and the number of its harmonics used in the determination of the MA model, is described. Further, a simple selection technique of these parameters is presented that leads to an optimum MA approximation. Next, the relationship between a method based on the Cholesky factorization of the coveriance matrix and the present technique is investigated to derive additional insight into its convergence properties. Finally, an alternative modeling technique based on an AR representation of the inverse of the target spectral matrix is presented.