Study of nonlinear variants of the least mean square (LMS) algorithm

The class of LMS-type algorithms utilizing zero-memory limiters (on one or both correlation multiplier inputs) is examined. A theorem by R. Price (1958) is utilized to simplify the analysis, which otherwise might be extremely difficult. The correlator limiters may be included to simplify the hardware implementation or may arise owing to the effect of limited dynamic range in the loop components. The analysis provides closed-form expressions for the mean weight convergence behavior, permitting establishment of convergence and performance bounds for this class of algorithms.<
>