Constrained Adaptive Linear Multiuser Detection Schemes

By using a fair comparison method we show that contrary to the general belief the conventional LMS, when in training mode, does not necessarily outperform the popular blind LMS (BLMS). With the help of a constrained MMSE criterion we identify the correct trained version which is guaranteed to have uniformly superior performance over BLMS since it maximizes the SIR over an algorithmic class containing BLMS. Because the proposed optimum trained version requires knowledge of the amplitude of the user of interest we also present simple and efficient techniques that estimate the amplitude in question. The resulting algorithm in both modes, training and decision directed, is significantly superior to BLMS.