Adaptive estimation under single-index constraint in a regression model (Q2448721)

The paper under review deals with adaptive multivariate function estimation under single-index constraint in a regression model with random design and weak assumptions on the noise. The authors propose an estimation method that adapts simultaneously to an unknown index vector and to the link function smoothness by employing a special family of kernel estimators. They derive the so-called local oracle inequality that allows obtaining minimax adaptive results for estimating a function at a given point. Next, this inequality is used for proving the so-called global oracle inequality, which is shown to be handy in assessing the quality of the entire function estimation. The results are applied to pointwise adaptive estimation over a collection of Hölder functional classes and global adaptive estimation over a collection of Nikol'skii functional classes.