Fast implementation of the LMS algorithm

LMS algorithms are next to their numerically robust structure well suited for hardware implementation. In high data rate systems, however, their recursive formulation limits their performance. An architecture with more parallelism is desirable in which speed can be traded against chip area. This paper shows under which conditions such a parallel LMS structure is possible by means of particular training sequences. A new design method is proposed to generate training sequences with very distinct properties.