A Fixed-Point Implementation of the Unnormalized Least-Squares Lattice Using Scaled-Integer Arithmetic

The unnormalized least-squares lattice algorithm is modified for fixed-point hardware by appropriate scaling of the crosscorrelation, covariance, and likelihood parameters. The scaling reveals the dependence of the unnormalized parameter magnitudes on the number of error signal observations used in their respective updates. Comparing fixed-point and floating-point simulations shows no significant loss of precision in the estimated PARCOR coefficients in the fixed-point environment with scaled parameters. Scaling requires seven more multiplies per lattice stage but provides a least-squares lattice algorithm suitable for a wide variety of fast low-power signal processing chips without square-root normalization.