Error analysis of algorithms in linear prediction

Errors caused by the finite wordlength of the computer in several popular algorithms in linear prediction are studied. Both the classical algorithms and the split cases of the classical algorithms are superior to other algorithms since the errors of computation at each stage propagate more slowly than in the Levinson and Schur cases. The results of the Levinson and Schur algorithms are almost the same. For the split cases, the errors accumulate much faster than in the classical situations since the three-term recurrence is used.<
>