The performance of per-survivor processing (PSP) algorithms with stochastic gradient estimators is investigated for frequency-selective channels and array measurements. The focus is on the degree to which the underlying channel structure can be reliably exploited by such algorithms in tracking the time-varying channel and detecting the data.

The problem of performing adaptive maximum likelihood sequence detection (MLSD) for a slowly time-varying, frequency-selective, multipath channel based on array measurements is considered. Two adaptive MLSD algorithms with LMS estimation of the overall channel (i.e., all filtering and physical channel effects) are compared through simulations: conventional adaptive MLSD (CA-MLSD) and per survivor processing (PSP) based adaptive MLSD. The results suggest that PSP based adaptive MLSD is particularly effective and robust to variations in the channel structure, level of dynamics, delay spread profile, and timing offsets. Analytical characterization of the PSP based algorithm is developed. This algorithm has the additional advantage that it may be fully initialized with the same training overhead used in current single-antenna systems for channel identification.