Central limit theorems of partial sums for large segmental values (Q2366185)

This paper concerns the asymptotic conditional distributional behavior, for large \(y>0\), of random sums \(\sum^{T(0,y)}_{i=1}U_ i\) given the event \[ {\mathcal E}_ y\equiv\left[\sum_{i=1}^{T(0,y)}X_ i\geq y\right],\quad\text{where}\quad T(0,y)=\inf\left\{k\geq 1:\sum^ k_{i=1}X_ i\notin(0,y)\right\}, \] and where \((X_ i,U_ i)\) are bounded \(\mathbb{R}^{1+d}\) vectors which either are i.i.d. or are i.i.d. conditionally given the transitions \(\beta_ i\equiv(\alpha_ i,\alpha_{i+1})\) of an underlying ergodic finite-state Markov chain \(\{\alpha_ i\}\), under the assumptions that \(E(X_ i)<0\), \(P(X_ i>0)>0\). Such conditional limiting questions arise naturally in queueing theory, insurance, and (the motivating application for the authors) DNA sequence analysis in molecular biology. Because of the negative expectation of the \(X_ i\) variables, the conditioning event \({\mathcal E}_ y\) is exponentially unlikely, and a conditional central limit theorem is stated and proved by using the techniques of large deviation theory, essentially the Wald martingale family \[ \exp\left(\vartheta\sum^ n_{i=1}X_ i+\Bigl\langle t,\sum^ n_{i=1}U_ i\Bigr\rangle\right)\left/\bigl(E(e^{\vartheta X_ 1+\langle t,U_ 1\rangle})\bigr)^ n\right. \] together with Taylor expansion in \(\vartheta\) about the point \(\vartheta^*>0\) where \(E\exp(\vartheta^*X_ 1)=0\). Although the Theorem in the case of i.i.d. (non-lattice) \((X_ i,U_ i)\) was anticipated by \textit{S. Asmussen} [Adv. Appl. Probab. 14, 143-170 (1982; Zbl 0501.60076)] the large-deviation setting and proof given here lead to the natural Markov- chain generalization.