Almost surely consistent nonparametric regression from recursive partitioning schemes (Q795439)

The authors investigate the almost sure convergence of within box means estimators of a regression function \(h(x)={\mathbb{E}}(Y| X=x)\) under a weak moment assumption on Y. For sample size n, the boxes form a finite partition \(Q^{(n)}\) of the X-space generated by a learning sample \((X_ 1,Y_ 1),...,(X_ n,Y_ n)\) such that \(Q^{(n+1)}\) is a refinement of \(Q^{(n)}\). The method of proof uses martingale arguments and sharp bounds for the deviation between the empirical and the true distribution uniformly over boxes.