Asymptotic analysis of penalized likelihood and related estimators (Q2640278): Difference between revisions

The authors present a general approach for analyzing a certain class of curve estimators related to penalized likelihood estimation. Given a fit functional \(l_ n\) measuring the deviation of our function \(\theta\) from the data and a penalty functional J, being smaller for more desirable functions, the estimator \({\hat \theta}{}_ n\) in question is a minimizer of \(l_{n,\lambda}(\theta)=l_ n(\theta | data)+\lambda J(\theta),\) with some regularization parameter \(\lambda\). For certain penalty functionals and fit functionals it is possible to analyze the asymptotic behaviour of the estimates \({\hat \theta}{}_ n\) using linearization of the quantities involved. The main results give approximations of the systematic - and stochastic error and in particular for log-density estimation, log-hazard estimation and nonparametric logistic regression. Convergence rates are given for \({\hat \theta}{}_ n-\theta_ 0\) w.r. to spectral norms. The results of \textit{D. D. Cox} [ibid. 16, No.2, 694-712 (1988; Zbl 0671.62044)] are the basis of the analysis.