Design variables for bias distribution in transfer function estimation

Estimation of transfer functions of linear systems is one of the most common system identification problems. Several different design variables, chosen by the user for the identification procedure, affect the properties of the resulting estimate. In this paper it is investigated how the choices of prefilters, noise models, sampling interval, and prediction horizon (i.e., the use ofk-step ahead prediction methods) influence the estimate. An important aspect is thai the true system is not assumed to be exactly represented within the chosen model set. The estimate will thus be biased. It is shown how the distribution of bias in the frequency domain is governed by a weighting function, which emphasizes different frequency bands. The weighting function, in turn, is a result of the previously listed design variables. It is shown, e.g., thai the common least-squares method has a tendency to emphasize high frequencies, and that this can be counteracted by prefiltering. It is also shown that, asymptotically, it is only the prediction horizon itself, and not how it is split up into sampling interval times number of predicted sampling instants, that affects this weighting function.