Exponential bounds of mean error for the nearest neighbor estimates of regression functions (Q1091060): Difference between revisions

Considering an i.i.d. \(({\mathbb{R}}^ d\times {\mathbb{R}})\)-valued sample \((X_ 1,Y_ 1),...,(X_ n,Y_ n)\) with bounded random variables \(Y_ j\), the author studies N.N.-estimators \(\hat m_ n(x)= \sum^{n}_{j=1} V_{nj}Y^ x_ j\) for the regression function \(m(x)=E(Y| X=x)\) where \((V_{nj})\) are given probability vectors and \((Y^ x_ j)\) denotes a rearrangement of the variables \((Y_ j)\) satisfying \(\| X^ x_ 1-x\| \leq...\leq \| X^ x_ n-x\|\); see also e.g. \textit{L. Devroye}, Z. Wahrscheinlichkeitstheor. Verw. Geb. 61, 467-481 (1982; Zbl 0483.62029) and the literature cited there. The aim of this paper is to prove for \(D_ n=\int | \hat m_ n(x)- m(x)| dF_ X(x)\) the following exponential inequality: \(P(D_ n>\epsilon)<e^{-c_{\epsilon}n}\), with some constant \(c_{\epsilon}>0\), provided some assumptions are satisfied. This generalizes a result of \textit{J. Beck}, Probl. Control. Inf. Theory 8, 303-311 (1979; Zbl 0497.62050), giving a corresponding inequality for the case \(V_{nj}=1\) for \(1\leq j\leq k_ n\) and \(V_{nj}=0\) otherwise under stronger assumptions on \(F_ X\) and m.