Transient behavior of fixed point LMS adaptation

We relate the distinguishing features of the fixed point power-of-two step size LMS algorithm´s learning curve to the precision of its data and coefficient variables. In particular, we show that the increase in the steady state MSE floor due to finite precision effects is determined primarily by data quantization while the decrease in convergence rate due to finite precision is determined by both data and coefficient quantization. We also derive a condition under which the slowdown phenomenon can be eliminated, given the reference variance and lower bounds on the minimum MSE and optimal weight vector magnitude