Some observations on implementing various recursive least squares adaptive filtering algorithms

A combination of different approaches has enabled a comparison to be made between covariance and square-root information domain algorithms in a low-precision environment. Results have demonstrated the superior performance of square-root information domain algorithms over covariance domain algorithms in such an environment. This is due to the inherent numerical stability of the orthogonal rotations they employ. Of these algorithms, the QR-lattice algorithm proved the most beneficial in terms of its complexity, stability and sensitivity to such an environment. However, even its performance becomes increasingly unacceptable below the 5-bit level. If a longer mantissa is employed instead, the stabilized fast covariance domain algorithms will, in addition, directly provide the filter coefficients at a reduced complexity, although stability concern about such algorithms may remain.