Small deviation probabilities for sums of independent positive random variables whose density has a power decay at zero (Q357248)

The author improves one of his previous results concerning small deviation probabilities of a sum of i.i.d. positive random variables whose density has a power decrease at zero. Let these variables be \((X_j)\) with distribution \(V(x)=\operatorname{P}(X_j<x)\) absolutely continuous and with density \(p(x)\) that satisfies \(p(x)\leq CV(x)/x\), \(0<x\leq x_0\), for some constant \(C>0\) (i.e. only in a vicinity of zero instead of the whole line), \(\mathrm{ess}\,\inf{V}=0\), and there exist constants \(b\in (0,1)\), \(c_1,c_2>1\), and \(r_0>0\) such that \[ c_1V(br)\leq V(r)\leq c_2V(br) \] for any \(r\), \(0<r\leq r_0\) (and this condition essentially relaxes the former assumption that the distribution \(V\) is regularly varying at zero). Then the main result of the paper (Theorem 2) argues that the distribution of the sum \(S_n=\sum_1^nX_j\) is absolutely continuous in the region \(0<x\leq x_0n\) with density \(f_n(x)\) such that \[ f_n(x)=\frac 1{\sigma(h)\sqrt{n}}L^n(h)e^{hx} (\phi(\beta)+\sum_{\nu=1}^{s-1}q_{\nu h}(\beta)n^{-\nu/2}+O(n^{-s/2})) ,\;\;\;n\to \infty, \] for any integer \(s\geq 1,\) some \(\beta\), any fixed positive \(h_0\), and \(a<x_0\) uniformly in \(0<x\leq an\) and \(h\geq h_0\); \(\phi(\, .\, )\) is the standard normal density, \(L(h)=\operatorname{E}e^{hX_j}\), the functions \(q_{\nu h}(y)\) are uniformly bounded in \(y\in \mathbb{R}\).