An estimate about multiple stochastic integrals with respect to a normalized empirical measure

Abstract

Let a sequence of iid. random variables $\xi_1,...,\xi_n$ be given on a measurable space $(X,\cal X)$ with distribution $\mu$ together with a function $f(x_1,...,x_k)$ on the product space $(X^k,{\cal X}^k)$. Let $\mu_n$ denote the empirical measure defined by these random variables and consider the random integral $$ J_{n,k}(f)={{n^{k/2}}\over{k!}}\int' f(u_1,...,u_k) (\mu_n(du_1)-\mu(du_1))...(\mu_n(du_k)-\mu(du_k)), $$ where prime means that the diagonals are omitted from the domain of integration. In this work a good bound is given on the probability $P(|J_{n,k}(f)|>x)$ for all $x>0$. This result shows that the tail behaviour of the distribution funtcion of the random integral $J_{n,k}(f)$ and that of the integral of the function $f$ with respect to a Gaussian random field show a similar behaviour. The proof is based on an adaptation of some methods of the theory of Wiener--Ito integrals. In particular, a sort of diagram formula is proved for the random integrals $J_{n,k}(f)$ together with some of its important properties, a result which may be interesting in itself. The relation of this estimate to some results about $U$-statistics is also discussed.