Within the theoretical literature on stationary random variables, pure Moving Average models and pure Autoregressive models have a rich body of work, but the corresponding literature on Autoregressive Moving Average (ARMA) models is very sparse. We attempt to fill certain gaps in this sparse line of work. Central to our observations is the spectral lemma connecting supnorm based function approximation on the unit circle to random variable approximation. This method allows us to provide quantitative approximation bounds in contrast with the qualitative boundedness and stability guarantees associated with unit root tests. Using the spectral lemma we first identify a class of stationary processes where approximation guarantees are feasible. This turns a known heuristic argument motivating ARMA models based on rational approximations into a rigorous result. Second, we identify an idealized stationary random process for which we conjecture that a good ARMA approximation is not possible. Third, we calculate exact approximation bounds for an example process, and a constructive proof that, for a given order, Pad\'e approximations do not always correspond to the best ARMA approximation. Unlike prior literature, our approach uses the generating function of the random process rather than the spectral measure, and further our results focus on approximation error of the random variable rather than the prediction error as in some classical infimum results by Szego, Kolmogorov, and Wiener.

翻译：暂无翻译