We consider a queuing system with multitype customers and nonhomogeneous flexible servers, in the heavy traffic asymptotic regime and under a complete resource pooling (CRP) condition. For the input-queued (IQ) version of such a system (with customers being queued at the system “entrance,” one queue per each type), it was shown in the work of Mandelbaum and Stolyar that a simple parsimonious Gcμ scheduling rule is optimal in that it asymptotically minimizes the system customer workload and some strictly convex queuing costs. In this article, we consider a different—output-queued (OQ)—version of the model, where each arriving customer must be assigned to one of the servers immediately upon arrival. (This constraint can be interpreted as immediate routing of each customer to one of the “output queues,” one queue per each server.) Consequently, the space of controls allowed for an OQ system is a subset of that for the corresponding IQ system.

We introduce the MinDrift routing rule for OQ systems (which is as simple and parsimonious as Gcμ) and show that this rule, in conjunction with arbitrary work-conserving disciplines at the servers, has asymptotic optimality properties analogous to those Gcμ rule has for IQ systems. A key element of the analysis is the notion of system server workload, which, in particular, majorizes customer workload. We show that (1) the MinDrift rule asymptotically minimizes server workload process among all OQ-system disciplines and (2) this minimal process matches the minimal possible customer workload process in the corresponding IQ system. As a corollary, MinDrift asymptotically minimizes customer workload among all disciplines in either the OQ or IQ system.

We consider a generic continuous-time system in which events of random magnitudes occur stochastically and study the system's extreme-value statistics. An event is described by a pair (t,x) of coordinates, where t is the time at which the event took place and x is the magnitude of the event. The stochastic occurrence of the events is assumed to be governed by a Poisson point process.

We study various issues regarding the system's extreme-value statistics, including (i) the distribution of the largest-magnitude event, the distribution of the nth “runner-up” event, and the multidimensional distribution of the “top n” extreme events, (ii) the internal hierarchy of the extreme-value events—how large are their magnitudes when measured relative to each other, and (iii) the occurrence of record times and record values. Furthermore, we unveil a hidden Poissonian structure underlying the system's sequence of order statistics (the largest-magnitude event, the second largest event, etc.). This structure provides us with a markedly simple simulation algorithm for the entire sequence of order statistics.

We consider a system consisting of a server alternating between two service points. At both service points, there is an infinite queue of customers that have to undergo a preparation phase before being served. We are interested in the waiting time of the server. The waiting time of the server satisfies an equation very similar to Lindley's equation for the waiting time in the GI/G/1 queue. We will analyze this Lindley-type equation under the assumptions that the preparation phase follows a phase-type distribution, whereas the service times have a general distribution. If we relax the condition that the server alternates between the service points, then the model turns out to be the machine repair problem. Although the latter is a well-known problem, the distribution of the waiting time of the server has not been studied yet. We derive this distribution under the same setting and we compare the two models numerically. As expected, the waiting time of the server is, on average, smaller in the machine repair problem than in the alternating service system, but they are not stochastically ordered.

We consider a storage model that can be on or off. When on, the content increases at some state-dependent rate and the system can switch to the off state at a state-dependent rate as well. When off, the content decreases at some state-dependent rate (unless it is at zero) and the system can switch to the on position at a state-dependent rate. This process is a special case of a piecewise deterministic Markov process. We identify the stationary distribution and conditions for its existence and uniqueness.

To compare two multivariate random vectors of the same dimension, we define a new stochastic order called upper orthant dispersive ordering and study its properties. We study its relationship with positive dependence and multivariate hazard rate ordering as defined by Hu, Khaledi, and Shaked (Journal of Multivariate Analysis, 2002). It is shown that if two random vectors have a common copula and if their marginal distributions are ordered according to dispersive ordering in the same direction, then the two random vectors are ordered according to this new upper orthant dispersive ordering. Also, it is shown that the marginal distributions of two upper orthant dispersive ordered random vectors are also dispersive ordered. Examples and applications are given.

We consider control policies for perishable inventory systems with random input whose purpose is to mitigate the effects of unavailability. In the basic uncontrolled system, the arrival times of the items to be stored and the ones of the demands for those items form independent Poisson processes. The shelf lifetime of every item is finite and deterministic. Every demand is for a single item and is satisfied by the oldest item on the shelf, if available. The first controlled model excludes the possibility of unsatisfied demands by introducing a second source of fresh items that is completely reliable and delivers without delay whenever the system becomes empty. In the second model, there is no additional ordering option by outsourcing. However, to avoid the most adverse effects of unavailability, the demands are classified into different categories of urgency. An incoming demand is satisfied or not according to its category and the current state of the system. For both models, we determine the steady-state distribution of the virtual outdating process, which is then used to derive the relevant cost functionals: the steady-state distribution and expected value of the number of items in the system, the rate of outdatings, as well as, for model 1, the rate of special orders from the external source and, for model 2, the rate of unsatisfied demands.

We propose an analytically tractable model of loading-dependent cascading failure that captures some of the salient features of large blackouts of electric power transmission systems. This leads to a new application and derivation of the quasibinomial distribution and its generalization to a saturating form with an extended parameter range. The saturating quasibinomial distribution of the number of failed components has a power-law region at a critical loading and a significant probability of total failure at higher loadings.

We consider a Jackson network with two nodes, with no exogenous input, but instead an infinite supply of work at each of the nodes: Whenever a node is empty, it processes a job from this infinite supply. We obtain an explicit expression for the steady state distribution of this system, as an infinite sum of product forms.

We consider the problem of determining when to exit an investment whose cumulative return follows a Brownian motion with drift μ and volatility σ2. After an unobserved exponential amount of time, the drift drops from μH > 0 to μL < 0. Using results from stochastic differential equations, we are able to show that it is optimal to exit the first time the posterior probability of being in the low state falls to p*, where the value of p* is given implicitly. We effect a complete comparative statics analysis; one surprising result is that a decrease in μL is beneficial when |μL| is large.

Two well-known orders that have been introduced and studied in reliability theory are defined via stochastic comparison of inactivity time: the reversed hazard rate order and the mean inactivity time order. In this article, some characterization results of those orders are given. We prove that, under suitable conditions, the reversed hazard rate order is equivalent to the mean inactivity time order. We also provide new characterizations of the decreasing reversed hazard rate (increasing mean inactivity time) classes based on variability orderings of the inactivity time of k-out-of-n system given that the time of the (n − k + 1)st failure occurs at or sometimes before time t ≥ 0. Similar conclusions based on the inactivity time of the component that fails first are presented as well. Finally, some useful inequalities and relations for weighted distributions related to reversed hazard rate (mean inactivity time) functions are obtained.

In this article, we study comparison theorems for stochastic functionals like V(∞;C) = sup0≤t {M(t) − C(t)} or V(T;C) = sup0≤t≤T {M(t) − C(t)}, where {M(t)} and {C(t)} are two independent nondecreasing processes with stationary increments. We will tacitly assume that the considered stochastic functionals are proper random variables. We prove that V(T;C′) ≤icxV(T;C) ≤icxV(T;C′′), where and C′′(dt) = c(0) dt, provided dC(t) is absolute continuous with density c(t). Similarly, we show that V(∞;C′) ≤icxV(∞;C) ≤icxV(∞;C′′). For proofs, we develop the theory of the ≤idcv ordering defined by increasing directionally concave functions. We apply the theory to M/G/1 priority queues and M/G/1 queues with positive and negative customers.

We present some tips for generating multiple sequences of two- and three-valued random variables.

A surprisingly simple and explicit expression for the waiting time distribution of the MX/D/c batch arrival queue is derived by a full probabilistic analysis, requiring neither generating functions nor Laplace transforms. Unlike the solutions known so far, this expression presents no numerical complications, not even for high traffic intensities.

The distribution of the linear combination αX + βY is derived when X and Y are independent Laplace random variables. Extensive tabulations of the associated percentage points are also given. The work is motivated by examples in automation, control, fuzzy sets, neurocomputing, and other areas of informational sciences.

For simulating GI/G/1 queues, we investigate estimators of stationary delay-in-queue moments that were suggested but not investigated in our recent article and we develop new ones that are even more efficient. Among them are direct spread estimators that are functions of a generated sequence of spread idle periods and are combinations of estimators. We also develop corresponding conditional estimators of equilibrium idle-period moments and delay moments. We show that conditional estimators are the most efficient; in fact, for Poisson arrivals, they are exact. In simulation runs with both Erlang and hyperexponential arrivals, conditional estimators of mean delay are more efficient than a published method that estimates idle-period moments by factors well over 100 and by factors of over 800 to several thousand for estimating stationary delay variance.

Consider a single-commodity inventory system in which the demand is modeled by a sequence of independent and identically distributed random variables that can take negative values. Such problems have been studied in the literature under the name cash management and relate to the variations of the on-hand cash balances of financial institutions. The possibility of a negative demand also models product returns in inventory systems. This article studies a model in which, in addition to standard ordering and scrapping decisions seen in the cash management models, the decision-maker can borrow and store some inventory for one period of time. For problems with back orders, zero setup costs, and linear ordering, scrapping, borrowing, and storage costs, we show that an optimal policy has a simple four-threshold structure. These thresholds, in a nondecreasing order, are order-up-to, borrow-up-to, store-down-to, and scrap-down-to levels; that is, if the inventory position is too low, an optimal policy is to order up to a certain level and then borrow up to a higher level. Analogously, if the inventory position is too high, the optimal decision is to reduce the inventory to a certain point, after which one should store some of the inventory down to a lower threshold. This structure holds for the finite and infinite horizon discounted expected cost criteria and for the average cost per unit time criterion. We also provide sufficient conditions when the borrowing and storage options should not be used. In order to prove our results for average costs per unit time, we establish sufficient conditions when the optimality equations hold for a Markov decision process with an uncountable state space, noncompact action sets, and unbounded costs.

The asymptotic decay rate of the sojourn time of a customer in the stationary M/G/1 queue under the foreground–background (FB) service discipline is studied. The FB discipline gives service to those customers that have received the least service so far. We prove that for light-tailed service times, the decay rate of the sojourn time is equal to the decay rate of the busy period. It is shown that FB minimizes the decay rate in the class of work-conserving disciplines.

The problem of controlling transmission rate over a randomly varying channel is cast as a Markov decision process wherein the channel is modeled as a Markov chain. The objective is to minimize a cost that penalizes both buffer occupancy (equivalently, delay) and power. The nature of the optimal policy is characterized using techniques adapted from classical inventory control.

We consider the loss probability of a customer in a single-server queue with finite buffer and partial rejection and show that it can be identified with the tail distribution of the cycle maximum of the associated infinite-buffer queue. This equivalence is shown to hold for the GI/G/1 queue and for dams with state-dependent release rates. To prove this equivalence, we use a duality for stochastically monotone recursions, developed by Asmussen and Sigman (1996). As an application, we obtain several exact and asymptotic results for the loss probability and extend Takács' formula for the cycle maximum in the M/G/1 queue to dams with variable release rate.

In the article published earlier this year in Probability in the Engineering and Informational Sciences 19(1), Theorem 2 on p. 81 was stated incorrectly.