Higher order normalizing transformations of asymptotic \(U\)-statistics for removing bias, skewness and kurtosis (Q1969137)

Let \(X_1, X_2,\ldots , X_n\) be independent and identically distributed random vectors (possibly taking values in any measurable space) with distribution function \(F_{\theta }(x)\), and \(\widehat \theta =\widehat\theta (X_1,\ldots , X_n)\) be an estimator of the parameter \(\theta\). The studentized statistic \(\widehat s = \sqrt n (\widehat \theta - \theta)/\widehat \sigma\) is considered, where \(\widehat\sigma^2\) is a consistent estimator of the variance \(\sigma^2=\text{var} (\sqrt n(\widehat \theta - \theta))\). A normalizing transformation \(\pi (s)\), which removes bias, skewness and kurtosis simultaneously, is proposed . The main result states that \[ P(\pi (\widehat s)\leq z) = \Phi(z) + o(n^{-1}). \] Here \(\Phi (z)\) is the standard normal distribution function. The transformation \(\pi \) is polynomial with degree 5 and monotone. It is assumed that \(\widehat s\) has the asymptotic representation \[ \widehat s = n^{-1/2}\delta + V_n + R_n, \] where \(V_n\) are the first members as in Hoeffding's decomposition and \[ P(|R_n|\geq n^{-1}(\log n)^{-3/2}) = o(n^{-1}). \] It is also shown, that the studentized \(U\)- statistic satisfies the conditions imposed on \(\widehat s\).