CAUSAL AR MODELING USING A LINEAR COMBINATION OF CUMULANT SLICES

We address the problem of estimating the parameters of a pure AR causa l model excited with non-Gaussian, ergodic, unobservable process. Outp ut samples may be corrupted with colored Gaussian noise. Other types o f noise are allowed provided that they have a set of mth order statist ics whose value is zero, if the same set of statistics are different f rom zero for the signal. It is shown that each sample of the impulse r esponse of the AR system that generates the process may be expressed a s a linear combination of cumulant slices of any order, thus providing a new framework to combine cumulants of different orders. The resulti ng algorithm is shown to be well behaved and to provide consistent est imates while reducing the complexity significantly with respect to oth er approaches.