Objective methods for comparing autoregressive order-determining criteria

Distance measures between a reference signal and the autoregressive estimate are used as an objective reference for comparing autoregressive order-determining criteria. The distance measures discussed are the rms log spectral deviation, its normalized cepstral distance approximation, the normalized autoregressive transfer function error, the equivalent Itakura distance, and the average squared prediction error. Using them it is found that the AIC criterion is one of the best criteria and performs better than the consistent Schwartz´ and Hannan and Quinn´s criteria.