Comparison of least-squares and stochastic gradient lattice predictor algorithms using two performance criteria

The least-squares (LS) and stochastic gradient (SG) lattice prediction algorithms are compared using two different performance criteria. These are a) output mean squared error and b) the accuracy of the autoregressive, spectral estimate obtained from the mean values of the lattice coefficients, assuming a stationayinput. It is found that the second performance criterion is more sensitive than the first. This "spectral" performance criterion is a measure of the accuracy of the estimatcd autoregressive model coefficients. Bias in the LS and SG coefficient estimates can cause significant deviation of the asymptotic spectral estimates from the actual input spectrum: The similarly between the LS and SG lattice algorithms enables comparative simulations with analogous initial conditions and exponential weighting constants. For both types of comparisons, the LS algorithm offers a modest performance improvement over the SG algorithms simulated. This improvement is more noticeable when the input is highly correlated. It is also found that slight changes in the SG lattice algorithm may significantly affect its performance.