Large deviations for sums of lattice random variables under the Cramer conditions (Q4522774)

Let \(\xi, \xi_1, \xi_2, \ldots \) be i.i.d. random variables with a lattice distribution, i.e. taking values \(a + kh\) with \(k\) integer. Assume that \[ 0 = \inf\{s: E[\exp(s\xi)] < \infty\} < \sup\{s: E[\exp(s\xi)] < \infty\} = \lambda < \infty. \] Let \(f(s) = E[\exp(s\xi)]\), \(M(s) = f'(s)/ f(s)\). Note that \(M(0) = E[\xi]\) (\(M(0) = -\infty\) is possible). Let \(M = \lim_{ s \to \lambda}M(s)\) (\(M = \infty\) is possible) and \(\rho(x) = \inf_{0\leq s < \lambda}f(s)\exp(-sx)\). Let \(\delta\) and \(\gamma\) be small positive numbers and set \(M_\delta = M-\delta\) if \(M > \infty\), \(M_\delta = 1/\delta\) else, and \(m_\gamma = M(0) + \gamma\), if \(M(0) < \infty\), \(m_\gamma = - 1/\gamma\) else. Then, a theorem of \textit{V. V. Petrov} [``Sums of independent random variables'' (1975; Zbl 0322.60042), Sect. VIII, 4.2, p. 249] says that for \(S_n : = \xi_1 + \ldots +\xi_n\), \[ P[S_n =x] = B_1(x/n) n^{-1/2}\rho(x/n)^n(1+o(1))\tag{1} \] and \[ P[S_n \geq x] = B_2(x/n) n^{-1/2}\rho(x/n)^n(1+o(1))\tag{2} \] uniformly in \(x\), \(x = na+kh\), \(m_\gamma \leq x \leq M_\delta\), with \(s(x) = M^{-1}(x)\) and \(B_1(x) = h(2\pi M'(s(x))^{-1/2}\), \(B_2(x) = h/(1 - \exp(-s(x)h))(2\pi M'(s(x))^{-1/2}\). The author investigates the tail behaviour of \(S_n\) in the case \(x/n > M_\delta\). He gives conditions under which (1) and (2) hold true uniformly in \(x\), \(x = na+kh\), \(m_\gamma < x\).