AR and ARMA identification algorithms of Levinson type: An innovations approach

Fast recursive-in-time identification procedures for both AR and ARMA processes (e.g., Chandrasekhar, square root algorithms,... ) have been available for a few years. These algorithms realize the desired transfer functions in the classical polynomial or rational form. On the other hand, the synthesis of polynomial and rational transfer functions in lattice and ladder form has fostered great interest in network theory by virtue of its pleasant properties. This type of synthesis is strongly related to the theory of orthogonal polynomials on the unit circle. An identification procedure with the realization of the desired whitening filter in lattice form was available for AR processes. We give here a simple approach for obtaining such algorithms, investigating furthermore the connections between the so obtained algorithms and the classical ones (least squares procedure). In the same way, we obtain identification procedures with realization of the desired filter in lattice and ladder form for ARMA processes, together with the connection with the classical extended least squares procedure. The method is based upon a fairly general Levinson orthogonalization lemma in Hilbert space, involving innovation techniques. We extend the method to various other estimation problems. The algorithms we obtain are fast (even somewhat faster than the previous fast ones) and seem to be well conditioned.