A Cramér type theorem for weighted random variables (Q1858648)

Let \((Z_i)_{i\in N}\) be a sequence of \({\mathbb R}^d\)-valued i.i.d. random vectors satisfying \({\mathbb E}e^{\alpha|Z_1|} <\infty\) for some \(\alpha>0.\) Let \({\mathcal X}\) be a topological vector space endowed with its Borel \(\sigma\)-field \({\mathcal B}\). Let \((x^n_i,\;1\leq i\leq n,\;n\geq 1)\) be an \({\mathcal X}\)-valued sequence of random elements satisfying \(\frac{1}{n}\sum_{i=1}^n \delta_{x^n_i}\to R\) weakly, \(n\to \infty.\) Here, \(R\) is assumed to be a strictly positive probability measure, that is \(R(U)>0\) whenever \(U\) is a nonempty open subset of \({\mathcal X}\). Let \({\mathbf f}\) be an \(m\times d\) matrix and let \(Z\in {\mathbb R}^d\), \(\theta\in {\mathbb R}^m\). We denote by \(\theta {\mathbf f}\in {\mathbb R}^d,\) \({\mathbf f}Z^T\in {\mathbb R}^m\) the usual matrix products. We introduce the weighted empirical mean \[ \langle L_n,{\mathbf f}\rangle = \frac{1}{n} \sum_{i=1}^n {\mathbf f(x^n_i)}Z^T_i\in {\mathbb R}^m, \] where \({\mathbf f}={\mathbf f}(x):~{\mathcal X}\to {\mathbb R}^{m\times d}\) is a matrix-valued bounded continuous function. The main result of the article: The family \(\langle L_n,{\mathbf f}\rangle \) satisfies the large deviation principle (LDP) in \({\mathbb R}^m\) with the good rate function \[ I_{{\mathbf f}}(y)=\sup_{\theta\in {\mathbb R}^m}\Bigl\{\theta y^T-\int_{{\mathcal X}} \Lambda(\theta {\mathbf f}(x))R(dx)\Bigr\},\quad y\in {\mathbb R}^m, \] where \(\Lambda(\lambda)=\log {\mathbb E}e^{\lambda Z^T_1},\) \(\lambda\in {\mathbb R}^d\). As applications, the author derives a LDP for the family of empirical measures and for the empirical processes. Various examples illustrate the scope of the results.