On Probabilities of Large Deviations for Random Walks. I. Regularly Varying Distribution Tails (Q3147221)

The authors are interested in large deviation problems of special type for the sequence of sums \((S_k)_{k\geq 1},\;S_k=X_1+\cdots+X_k\) (called random walk), of i.i.d. r.v. \(X, X_1, X_2,\dots\) Namely it is known that under the very weak conditions \[ EX=0,\quad EX^2=1,\quad E|X|^b<\infty,\;b\geq 2,\qquad V(t)\equiv P(X>t)=t^{-\beta}L(t),\;\beta>2, \] (\(L(t)\) is a slowly varying function, \(\beta\geq b\)) we have the asymptotic relation \[ V_n(x)\equiv P(S_n>x)=[1-\Phi(x\sqrt{n})+nV(x)](1+o(1)),\quad n\to \infty, \] uniformly in \(x>\sqrt{n}\), \(\Phi\) being the standard normal distribution function. And what is more the last term in the brackets becomes dominant in the so-called large deviation region \[ V_n(x)\sim nV(x),\qquad x\geq c(n\log n)^{1/2},\qquad c>(\beta-2)^{1/2}. \tag{1} \] Thus in that region the distribution of the sum is close to that of the maximum of the \(X\)'s: \(P(S_n>x)\sim P(\max _{k\leq n} X_k>x).\) That's why the main purpose of the paper is to derive an analogue of (1) for the probabilities of crossing the so-called arbitrary boundaries (considered instead of \(x\)) by the random walk \((S_k)\) until time \(n,\) and also asymptotic expansions (in powers of \(\sqrt{n}/x\)), refining both (1) and similar relations for the above-mentioned probabilities, in particular, for the distribution tail \(\overline V_n(x)=P(\overline S_n>x)\) of the maximum \(\overline S_n=\max_{k\leq n} S_k.\) In part II of the paper the ``regular exponential case'' is considered in the same manner when \[ V(t)=e^{-\mu t}t^{-\beta-1}L(t),\quad \mu>0,\quad \beta>2. \]