Institutes outside Greece
Large deviations for semi-exponential distributions: theory and applications
Abstract
dc:descriptionDeveloped in Chapter 2, is the sample path large deviation principle for Lévy processes and random walks with heavy-tailed Weibull (semi-exponential) increments. This result holds in the Skorokhod space with respect to the M1’ topology. In addition, we prove the extended sample path LDP in the Skorokhod space with the finer J1 topology. Furthermore, we develop theoretical tools for extended LDP and show that the standard LDP cannot be satisfied for the Lévy processes with heavy-tailed Weibull increments. This suggests that the extended LDP is the optimal result one can achieve with respect to the J1 topology. We illustrate this by constructing a counterexample; showing that the LDP in the J1 topology is not possible. These large deviations results have been extended to multidimensional settings in the case of independent Lévy processes and random walks. To enhance the applicability of the extended LDP, we have also developed a form of contraction principle. In particular, we study ruin probabilities in a reinsurance example. That is, we consider level crossing probabilities of Lévy processes where the jump sizes are conditioned to be moderate. In conclusion, for the random processes treated in this chapter, our large deviation analysis demonstrates that associated rare events are caused by big discontinuities of their sample paths; this phenomenon has been characterized as the principle of multiple big jumps. The third chapter centers on sample path large deviations for Markov additive processes. More precisely, we prove the sample path LDP for unbounded additive functionals of processes with light-tailed increments that are induced by the Lindley recursion. The LDP holds in the Skorokhod space equipped with the M1’ topology and with sub-linear speed. Although the process under consideration is constructed by light-tailed increments, rare events are caused by “big jumps”. This result establishes that the structure of light-tailed random processes can induce (asymptotically) a heavy-tailed behavior. Our technique hinges on a suitable decomposition of the Markov chain in terms of regeneration cycles. At each regeneration cycle, we study the accumulated area of the Lindley process. Consequently, the area displays heavy-tailed behavior, and it satisfies an LDP. To derive tail asymptotics for the area, we use sample path analysis; as a by-product of our LDP we show that large areas are caused by concave trajectories of our process. In the fourth chapter, we focus on, the multiple server queue. An important question is the likelihood of a large queue length or waiting time in such systems. To exemplify, consider d parallel servers, each working at a certain speed and suppose that the service time distribution is heavy-tailed. If k large jobs appear in the system simultaneously, then they reduce the capacity of the system, which is not detrimental if the remaining service capacity exceeds the system load ρ. One expects that k* large jobs are required to make the system behave poorly, where k* is the minimum number of big jobs needed to cause instability—in the sense of congestion—in the system. The main results in Chapter 4 provide an estimate for the probability of large queue lengths as well as detailed answers on how large queue lengths occur. For the latter part, we determine the number of big jobs and their sizes that lead to congestion. In Chapter 5, we apply our fundamental results of Chapter 2 to study a stochastic fluid network model with heavy-tailed input (compound Poisson processes with semi-exponential increments). Our results include the continuity of the multidimensional reflection map on certain subspaces of the Skorokhod space under the product J1 topology. Based on the continuity of the multidimensional reflection map we prove large deviation bounds for the multidimensional buffer content process of the stochastic fluid network. Furthermore, we use the large deviation bounds of the buffer content process to estimate overflow probabilities for a subset of the stochastic fluid network nodes. We associate the overflow probabilities with a simplified optimization problem. Lastly, we perform explicit computations in the case of a certain network that relates to—w.r.t. its network topology—the multiple on-off sources model.
Degree
thesis:*- Grantor dc:publisher
- Institutes outside Greece
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Bazmpas, Michail
- Μπάζμπας, Μιχαήλ
Subjects
dc:subject × 30- Θεωρία μεγάλων αποκλίσεων
- Λογαριθμικά ασυμπτωτικά
- Τυχαίοι περίπατοι
- Διαδικασίες Lévy
- Διαδικασίες Markov
- Tοπολογία J1
- Τοπολογία M1'
- Ουρά πολλαπλών εξυπηρετητών
- Στοχαστικό δίκτυο
- Ανακλαστική απεικόνιση
- Βαριές ουρές
- Κατανομή Weibull
- Large deviations
- Logarithmic asymptotics
- Random walks
- Lévy processes
- Markov processes
- J1 topology
- M1' topology
- Multiple server queue
- Stochastic networks
- Reflection map
- Heavy tails
- Weibull distribution
- Μαθηματικά
- Φυσικές Επιστήμες
- Στατιστική και Πιθανότητες
- Mathematics
- Natural Sciences
- Statistics and Probability
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/57324