Variational analysis of constrained M-estimators (Q2215758)

For computing nonparametric \(M\)-estimators, more general versions of the optimization problem, as used for parametric \(M\)-estimators, are considered. Here, the unknown quantity is a certain function \(f\), lying in a given set \(F\) of functions on \(\mathbb{R}^d\) with dimension \(d\). An example is the probability density to be estimated in case of nonparametric maximum likelihood (ML)-estimators. With a certain loss function \(\psi = \psi(x,f)\), the objective function \(\Psi=\Psi(f)\) to be minimized with respect to \(f \in F\) is defined by the average of \(\psi(x_i,f)\), \(i=1,\dots,n\), with samples \(x_i\) of random variables \(X_i\), \(i=1,\dots,n\); moreover, \(\Psi(f)\) involves still an additive penalty term \(\pi = \pi(f)\) for possible smoothing or regulation. Under certain assumptions, results on the existence and consistency of the resulting solutions of the constrained \(M\)-estimator problem are given. Moreover, an algorithm for an approximate solution of the \(M\)-estimator problem is suggested and examples are given.