Exact asymptotic behaviour of the distribution of the supremum (Q5937053)

Let \(X_1,X_2,\dots\) be i.i.d. random variables with common nonarithmetic distribution \(F\). Put \(S_0= 0\), \(S_n= X_1+\cdots+X_n\), \(n\geq 1\), and let \(S_n\to-\infty\) a.s. so that \(M= \max\{S_n, n\geq 0\}< \infty\) a.s. The paper studies the tail behaviour of the distribution \(D\) of \(M\) under the assumption that there is a distribution \(G\in{\mathcal S}(\gamma)\) with \(\gamma>0\) such that \[ \lim F(x,\infty)/G(x,\infty)= \lambda< \infty.\tag{\(*\)} \] The class \({\mathcal S}(\gamma)\) studied by \textit{M. S. Sgibnev} [Sib. Math. J. 29, No. 4, 647-655 (1988); translation from Sib. Mat. Zh. 29, No. 4(170), 162-171 (1988; Zbl 0653.46027)] generalizes the class of subexponential distributions \((\gamma= 0)\). One has \(H\in{\mathcal S}(\gamma)\) if \(H(-\infty, 0)= 0\), \(H(x,\infty)> 0\), \(x> 0\), \(\lim H(x+ y,\infty)/H(x, \infty)= \exp(-\gamma y)\), \(y> 0\), as \(x\to\infty\) and \(\lim(H* H)(x,\infty)/H(x,\infty)= 2\int\exp(\gamma x) H(dx)\). When \(\varphi(s)= \int\exp(sx) F(dx)\neq 1\), \(\text{Re }s= \gamma\), then \(D\) is equal to a sum of distributions determined explicitly by the zeros of \(1-\varphi(s)\) in \(0< \text{Re }s< \gamma\) and a remainder with the same tail behaviour as \(G\). When \(\lambda> 0\), the distribution of \(\max(0,X_1)\) also belongs to \({\mathcal S}(\gamma)\). When \(\varphi(\gamma)= 1\), the condition \((*)\) has to be replaced by a more complicated one but the tail behaviour of \(D\) is simpler. Proofs apply Banach algebra methods and derive some new results on \({\mathcal S}(\gamma)\).