Bounds for non-Gaussian approximations of U-statistics (Q1610504): Difference between revisions

Let \(X,\overline X,X_1,\dots, X_n\) be i.i.d. random variables taking values in a measurable space \(({\mathbf X},{\mathbf B})\). Let \(\Phi:{\mathbf X}^2\to\mathbb{R}\) be a symmetric degenerate kernel such that \(E\Phi(X,\overline X)< \infty\). Define U-statistics of degree two by \[ U_n= n^{-1} \sum_{1\leq i< j\leq n}\Phi(X_i, X_j). \] Let \(\{q_j\}\) \((|q_1|\geq|q_2|\geq\cdots)\) denote the eigenvalues of the Hilbert-Schmidt operator, \(Q\), associated with the kernel \[ Qf(x)= \int\Phi(x,y) f(y) P(dy),\quad f\in L_2({\mathbf X},{\mathbf B},P) \] with \(P={\mathcal L}(X)\). Let \(\{e_j\}\) denote a corresponding orthogonal, complete system of eigenfunctions of the operator \(Q\). Further, let \(\{\tau_j\}\) be a sequence of i.i.d. standard normal random variables and put \[ U_\infty= {1\over 2} \sum^\infty_{j=1} q_j(\tau^2_j- 1). \] It is known that \(U_n\) converges weakly in distribution of \(U_\infty\). The authors obtain an upper bound, expressed by terms containing tail and truncated moments, for \[ \sup_x|{\mathbf P}(U_n\leq x)-{\mathbf P}(U_\infty\leq x)| \] assuming that at least \(q_{13}\neq 0\). The proof depends on the method developed by \textit{V. Bentkus} and \textit{F. Götze} [Ann. Probab. 27, No. 1, 454-521 (1999)].