No label defined (Q3973919)

[For the entire collection see Zbl 0722.00032.] In the context of statistical estimation of functions, a list of parametric models is provided from which one is to be automatically selected from the data. These lists of parametric models have the property that for essentially any target function there exists a convergent sequence of approximating functions of the list. In the absence of knowledge of the true function, the problem is to estimate an appropriate size model that provides a good tradeoff between the approximation error, which is best for larger models, and the parameter estimation error, which is best for smaller models. An index of resolvability is presented that defines the best tradeoff between approximation error and the complexity of models. General complexity regularization criteria are defined and bounds on the statistical risk of the estimated functions are established. These bounds establish consistency, yield rates of convergence, and demonstrate the near asymptotic optimality of the model selection criterion in both parametric and nonparametric cases. Applications are given to polynomial regression and artificial neural networks.