Statistical nonlinearity

The data itself often has noise bursts or gaps, and we will see later in Chapter that this leads us to readjusting the weighting function. In principle, we should fix the weighting function and solve the problem. Then we should revise the weighting function and solve the problem again. In practice we find it convenient to change the weighting function during the optimization descent. Failure is possible when the weighting function is changed too rapidly or drastically. (The proper way to solve this problem is with robust estimators. Unfortunately, I do not yet have an all-purpose robust solver. Thus we are (temporarily, I hope) reduced to using crude reweighted least-squares methods. Sometimes they work and sometimes they don't.)