Sample convergence of the normed LMS algorithm with feedback delay

When a delay of even one iteration is introduced in the coefficient update loop, the projection properties of the NLMS (normed least mean square) algorithm are lost, allowing the error vector to increase as well as decrease in any given update. This makes the analysis of the algorithm with delay much more difficult. An exact analysis of the delayed update algorithm, on a sample function basis, is developed. It is shown that for any delay, the gain parameter, can be chosen sufficiently small to guarantee exponential convergence, assuming only that the input satisfies the standard mixing condition