Numerical performances of autoregressive spectrum estimators based on three-term recurrences

The numerical performance of two fast algorithms for estimating the power spectral density of an autoregressive process is studied. The algorithms perform similarly to the Burg algorithm, but require only two-thirds as many multiplications as the most efficient implementation of the Burg algorithm. This allows the high resolution associated with the Burg algorithm to be obtained using many fewer computations. One algorithm is the deterministic form of the split lattice algorithm adjoined to the split Levinson recursions; however, its resolution is relatively poor. The other algorithm corrects a bias in the first algorithm, and has resolution similar to the Burg algorithm