Optimal adaptive inference in random design binary regression (Q2405200)

Certain smoothness classes both for the regression and design density under suitable assumptions on the boundedness are considered. Estimators of the underlying regression and design density (apart from jointly adapting to desired regimes of smoothness) having the additional property of boundedness in the \(L^2\) sense (both in point-wise and Besov type norm sense) are presented. Confidence sets for the regression function in nonparametric binary regression with the design density as a nuisance parameter are constructed. These confidence sets are adaptive in the \(L^2\) loss over a continuous class of Sobolev type spaces. Adaptation consists in the smoothness of the regression function, over the maximal parameter spaces where adaptation is possible, provided the design density is smooth enough. Two key regimes are identified -- one where adaptation is possible, and the other where some critical regimes must be removed. The complete solution (lower and upper bounds) to the problem of asymptotic minimax goodness of fit testing with both simple and composite null hypotheses and unknown design density is provided. All the presented procedures are based on second order U-statistics constructed from projection kernels of suitable wavelet bases.