From Stein identities to moderate deviations (Q1942117)

The authors apply Stein's method to obtain a Cramer-type moderate deviation result for dependent random variables whose dependence is defined in terms of an identity, called Stein identity, which plays a central role in Stein's method. A corolary for zero-bias coupling is deduced. The result is then applied to four examples: (i) a combinatorial central limit theorem; (ii) the anti-voter model; (iii) a general system of binary codes; (iv) the Curie-Weiss model. A general Cramer-type moderate deviation result (Theorem 3.1.). Let \(W\) be a random variable of interest. Assume that there exist a deterministic positive constant \(\delta\), a random positive measure \(\hat{\mu}\) with support \([-\delta,\delta]\) and a random variable \(R\) such that \[ \operatorname{E}(Wf(W))=\int_{|t|\leq \delta}f'(W+t)d\hat{\mu}(t)+ \operatorname{E}(Rf(W)) \] for all absolutely continuous function \(f\) for which the expectation of either side exists. Let \[ D:=\int_{|t|\leq \delta}d\hat{\mu}(t). \] Theorem. Suppose that there exist constants \(\delta_1\), \(\delta_2\) and \(\theta>0\) such that \[ |\operatorname{E}(D|W)-1|\leq \delta_1(1+|W|), \] \[ |\operatorname{E}(R|W)|\leq \delta_2(1+|W|) \] or \[ |\operatorname{E}(R|W)|\leq \delta_2(1+W^2)\quad\text{and}\quad\delta_2|W|\leq \alpha<1 \] and \[ \operatorname{E}(D|W)\leq \theta. \] Then \[ \frac{\operatorname{P}(W>x)}{1-\Phi(x)}= 1+O_\alpha(1)\theta^3(1+x^3)(\delta+\delta_1+\delta_2) \] for \(0\leq x\leq \theta^{-1} \min\{\delta^{-1/3},~\delta_1^{-1/3},~\delta_2^{-1/3}\}\), where \(O_\alpha(1)\) denotes a quantity whose absolute value is bounded by a universal constant which depends on \(\alpha\) only, \(\Phi(x)\) is the standard normal distribution function.