Discussion of Random Rates in Anisotropic Regression by Homann and Lepski.pdfVIP

Discussion of “Random Rates in Anisotropic Regression” by Hoffmann and Lepski Lawrence D. Brown and Yi Lin University of Pennsylvania and University of Wisconsin, Madison 1 Introduction. We congratulate the authors for a stimulating paper (referred to as HL in the following). As the authors correctly stated, the number of variables does not affect the optimal rate of convergence in a regular parametric model, but it does affect the optimal rate of convergence in nonparametric models. To be more precise, the optimal rate of convergence in a nonpara- metric function estimation problem depends on the “effective” nonparametric dimension of the model. For example, the “effective” dimension of a nonparametric additive model is one, no matter how many variables are in the model. Therefore, variable selection in the additive model does not affect the optimal rate of convergence of the model. In light of this, we would like to consider the problem of adaptive estimation in the nonparametric functional ANOVA setup. It is readily seen that the variable selection problem considered in HL is a special case of the problem we consider. Following the example of HL, we will concentrate on the white noise model setting. This can be motivated by the results in Brown and Low (1996), and Nussbaum (1996). 2 Rates of convergence Before turning to our main topic we would like to point out a feature of the result in HL that we found surprising. We hope that they will be able to comment on this, and perhaps provide some additional background and a heuristic explanation. The feature that concerns us first appears in the formula for z (i ) near the end of Section 1, and is repeated in various n s forms later on, including in Theorem 1. To focus on this feature, let us consider the case of on