On M-processes and M-estimation (Q914301)

Consider the usual linear regression model, \(Y_ j=x_ j'\theta +\sigma e_ j\), \(j=1,...,n\), where the dimension of the parameter vector \(\theta\) is p. This article is concerned with the asymptotic behaviour of M-estimators of \(\theta\) when p is allowed to increase with the sample size. This is done by using a stochastic equicontinuity argument. From this argument new results for one-step M-estimators, M-estimators and robust scale estimators are obtained. The objective of the article is to impose as weak conditions as possible on the criterion function and on the error distribution. If for example \(\psi\), in usual M-estimation notations, has a bounded derivative, it is required that \(p^ 2(\log n)^{2+\gamma}/n\to 0,\gamma >0\), for contrast estimation. If \(\psi\) is discontinuous, a somewhat stronger condition on p is required: \(p^ 3(\log n)^ 2/n\to 0.\) To obtain the distribution of quadratic forms it is required that \(p^ 3(\log n)^{2+\gamma}/n\to 0,\gamma >0\), if \(\psi\) has a bounded derivative, and that \(p^ 4(\log n)^ 2/n\to 0,\) when \(\psi\) is discontinuous.