On the stability of robust filter-cleaners (Q1111302)

Let \(x_ t\) be a stationary AR process. Assume that \(y_ t=x_ t+u_ t\) where \(u_ t\) are additive outliers which are equal to zero in the majority. A filter for cleaning \(y_ t\) is based on the idea to put \(\hat x_ t=y_ t\) if \(y_ t\) is close to the prediction of \(x_ t\) and to replace \(y_ t\) otherwise by some value closer to this prediction. The authors show that the function in the recursive equation is in general not contractive, but that nevertheless two iterations with arbitrary initial conditions coincide after some random time \(T_ 0\).