A Kolmogorov inequality for \(U\)-statistics based on Bernoulli kernels (Q1341367)

Let \(X_ 1, \dots, X_ n\) be independent and identically distributed random variables with distribution function \(F\). Let \(h(x_ 1, \dots, x_ m)\) be a symmetric function with \(P_ F(h(X_ 1, \dots, X_ m) = 1) = p\) and \(P_ F (h(X_ 1, \dots, X_ m) = 0) = 1 - p\). The function \(h\) is called a Bernoulli kernel. Consider the \(U\)-statistic \(U_ n\) defined by \[ U_ n = {n \choose m}^{-1} {\textstyle {\sum_ c}} h(X_{i_ 1}, \dots, X_{i_ m}), \] where \(\sum_ c\) denotes the summation over all \({n \choose m}\) combinations from \(\{1, \dots, n\}\). \(U_ n\) is called a \(U\)-statistic based on the Bernoulli kernel \(h\). It is obvious that \(E_ F (h(X_ 1, \dots, X_ m)) = p\). Let \(k=[n/m]\) and \(\varepsilon > 0\). Assume that either \(\varepsilon + p < {1 \over 2}\) or \({1 \over 2} + {1 \over 3} \varepsilon \leq p \leq 1\). The author proves that \(P(U_ n - p \geq \varepsilon) \leq (1 - 4 \varepsilon^ 2)^{k/2}\). The result generalizes an inequality derived by \textit{A. N. Kolmogorov} [Sankhyā, Ser. A. 25, 369-376 (1963; Zbl 0126.332)] and \textit{D. W. Turner}, \textit{D. M. Young} and \textit{J. W. Seaman jun.} [Stat. Probab. Lett. 13, No. 3, 223-227 (1992; Zbl 0747.60020)] for simple averages of Bernoulli random variables and gives a bound sharper than that obtained by the author [Stat. Probab. Lett. 12, No. 3, 257-261 (1991; Zbl 0742.60018)].