Maximum Likelihood Identification of Gaussian Autoregressive Moving Average Models

Closed form representations of the gradients and an approximation to the Hessian are given for an asymptotic approximation to the log likelihood function of a multidimensional autoregressive moving average Gaussian process. Their use for the numerical maximization of the likelihood function is discussed. It is shown that the procedure described by Hannan (1969) for the estimation of the parameters of one-dimensional autoregressive moving average processes is equivalent to a three-stage realization of one step of the Newton-Raphson procedure for the numerical maximization of the likelihood function, using the gradient and the approximate Hessian. This makes it straightforward to extend the procedure to the multidimensional case. The use of the block Toeplitz type characteristic of the approximate Hessian is pointed out.