On the convergence behavior of the LMS and the normalized LMS algorithms