An \(\epsilon\)-median polish algorithm (Q578824)

L\({}_ 1\) estimation is a useful tool for outlier location in linear models, especially in the case of factorial equireplicated designs where all the observations have the same influence. In this paper a method of \(L_ 1\) estimation is proposed based on the approximation of the piecewise linear convex objective function whose minimization leads to \(L_ 1\) estimation by a smooth convex function. A procedure similar to the median polish is obtained. The procedure is approximate, but the connection between its result and the optimal residual \(L_ 1\) norm, as well as the convergence of the algorithm, is rigorously established.