Heavy-tailed random walks, buffered queues and hidden large deviations (Q2278655)

Let \((Z_i)_{i\in \mathbb{N}}\) be an independent and identically distributed sequence with \(\alpha\)-regularly varying tails (\(\alpha > 0\)) such that an appropriate tail-balance condition holds. When \(Z_1\) has finite mean assume additionally that the \(Z_i\) are centered. Consider the random walk \((S_n)_{n\in \mathbb{N}}\) defined by \(S_n = \sum_{k=1}^n Z_k\) and its càdlàg embedding \(X^{(n)}\) in the space \(\mathbb{D}\) of càdlàg functions on \([0,1]\), defined by \(X^{(n)}(t) = S_{\lfloor nt \rfloor}\). Let \((\lambda_n)_{n\in \mathbb{N}}\) be a \(\rho\)-regularly varying sequence of real numbers, where \(\rho>1/2\) when \(Z_1\) has finite variance and \(\rho > 1/\alpha\) when \(Z_1\) has infinite variance. For each \(j\in \mathbb{N}\), denote by \(\mathbb{D}_{\leq (j-1)}\) the set of càdlàg step functions on \([0,1]\) with at most \(j-1\) jumps, and let \(\gamma_n^{(j)} := [n P(|Z_1|>\lambda_n)]^{-j}\). The authors then show that \(\gamma_n^{(j)} P(X^{(n)}/\lambda_n \in \cdot)\) converges to a non-degenerate measure \(\mu_j\) as \(n\to\infty\), where the measure \(\mu_j\) is given explicitly. Here, the convergence concept is the so called \(\mathbb{M}_{\mathbb{O}}\)-convergence in the space \(\mathbb{M}_{\mathbb{D}\setminus \mathbb{D}_{\leq (j-1)}}\) of all Borel measures on \(\mathbb{D} \setminus \mathbb{D}_{\leq (j-1)}\) that are bounded away from \(\mathbb{D}_{\leq (j-1)}\). The used \(\mathbb{M}_{\mathbb{O}}\)-convergence concept, which is discussed in detail in [\textit{F. Lindskog} et al., Probab. Surv. 11, 270--314 (2014; Zbl 1317.60007)], is related to the usual concept of weak convergence of measures. The result described above is a hidden large deviation type limit result. For \(j=1\) it corresponds to a large deviation result on the space of càdlàg functions established by \textit{H. Hult} et al. [Ann. Appl. Probab. 15, No. 4, 2651--2680 (2005; Zbl 1166.60309)]. The key point of the proved theorem is that it also works for any \(j\in \mathbb{N}\), hence allowing for hidden large deviations of higher orders when \(j\geq 2\) is considered. The results are then applied in the context of queuing processes with heavy-tailed service times. They are also used to provide probability estimates of rare events governed by more than one large jump in the case of random walks with infinite mean.