OUTPUT ERROR CONVERGENCE OF ADAPTIVE FILTERS WITH COMPENSATION FOR OUTPUT NONLINEARITIES

Output error convergence of a Wiener model-based nonlinear stochastic gradient algorithm is analyzed. The normalized scheme estimates the pa rameters of a linear finite impulse response (FLR) model in cascade wi th a known output nonlinearity. The algorithm can be interpreted as a normalized least mean square (NLMS) algorithm with compensation for an output nonlinearity. Linearizing inversion of the nonlinearity is not utilized. Global output error convergence is then proved, provided th at the nonlinearity is monotone (not strictly monotone), and provided that a previously observed mechanism resulting in deadlock does not oc cur. The algorithm and the analysis include important practical eases like sensor saturation and deadzones that must be excluded when global parametric convergence is studied.