Large deviations for stochastic fluid networks with Weibullian tails (Q2095026)

The paper considers a single-class open stochastic fluid network with \(d\) nodes. Denote by \({J_i}(t) \overset{\Delta}{=} \sum_{j = 1}^{N_i(t)}J_i^{(j)}\) the total amount of external work that arrives at station \(i\) which is a compound Poisson process where \(\{ J_i^{(j)}\} _{j = 1,2,\dots}\) is an iid jump size sequence for each \(i = 1,\dots,d\). Denote also by J the subset of nodes that have an exogenous input. The processes \(\{ {J_1}(t):t \ge 0\} ,\dots,\{ {J_d}(t):t \ge 0\} \) are assumed to be independent. The key assumption on the distribution of the jump sizes \(J_i^{(1)}\) for \(i \in J\) is that they are semi-exponential: for each \(i \in J \subseteq \{ 1,\dots,d\} \), \(P\{ J_1^{(1)} \ge x\} = \exp \{ - {c_i}L(x){x^\alpha }\} \), where \(\alpha \in (0,1)\), \({c_i} \in (0,\infty)\), and \(L\) is a slowly varying function such that \(L(x)/{x^{1 - \alpha}}\) is non-increasing for sufficiently large \(x\). The paper gives large deviations estimates for the buffer content process in the vector-valued Skorokhod space which is endowed with the product \(J_1\) topology. The main results contain upper and lower large deviation bounds for the buffer content process, logarithmic asymptotics for overflow probabilities of the buffer content process over fixed times, and an explicit analysis of the two-node tandem network.