Phase-type Distributed Research Articles

Phase-type representations of stochastic interest rates with applications to life insurance

The purpose of the present paper is to incorporate stochastic interest rates into a matrix-approach to multi-state life insurance, where formulas for reserves, moments of future payments and equivalence premiums can be obtained as explicit formulas in terms of product integrals or matrix exponentials. To this end we consider the Markovian interest model, where the rates are piecewise deterministic (or even constant) in the different states of a Markov jump process, and which is shown to integrate naturally into the matrix framework. The discounting factor then becomes the price of a zero-coupon bond which may or may not be correlated with the biometric insurance process. Another nice feature about the Markovian interest model is that the price of the bond coincides with the survival function of a phase-type distributed random variable. This, in particular, allows for calibrating the Markovian interest rate models using a maximum likelihood approach to observed data (prices) or to theoretical models like e.g. a Vasiček model. Due to the denseness of phase-type distributions, we can approximate the price behaviour of any zero-coupon bond with interest rates bounded from below by choosing the number of possible interest rate values sufficiently large. For observed data models with few data points, lower dimensions will usually suffice, while for theoretical models the dimensionality is only a computational issue.

The Shape of Phylogenies Under Phase-Type Distributed Times to Speciation and Extinction

Phylogenetic trees describe relationships between extant species, but beyond that their shape and their relative branch lengths can provide information on broader evolutionary processes of speciation and extinction. However, currently many of the most widely used macro-evolutionary models make predictions about the shapes of phylogenetic trees that differ considerably from what is observed in empirical phylogenies. Here, we propose a flexible and biologically plausible macroevolutionary model for phylogenetic trees where times to speciation or extinction events are drawn from a Coxian phase-type (PH) distribution. First, we show that different choices of parameters in our model lead to a range of tree balances as measured by Aldous’ beta statistic. In particular, we demonstrate that it is possible to find parameters that correspond well to empirical tree balance. Next, we provide a natural extension of the beta statistic to sets of trees. This extension produces less biased estimates of beta compared to using the median beta values from individual trees. Furthermore, we derive a likelihood expression for the probability of observing an edge-weighted tree under a model with speciation but no extinction. Finally, we illustrate the application of our model by performing both absolute and relative goodness-of-fit tests for two large empirical phylogenies (squamates and angiosperms) that compare models with Coxian PH distributed times to speciation with models that assume exponential or Weibull distributed waiting times. In our numerical analysis, we found that, in most cases, models assuming a Coxian PH distribution provided the best fit.

Continuous scaled phase-type distributions

Products between phase-type distributed random variables and any independent, positive and continuous random variable are studied. Their asymptotic properties are established, and an expectation-maximization algorithm for their effective statistical inference is derived and implemented using real-world datasets. In contrast to discrete scaling studied in earlier literature, in the present continuous case closed-form formulas for various functionals of the resulting distributions are obtained, which facilitates both their analysis and implementation. The resulting mixture distributions are very often heavy-tailed and yet retain various properties of phase-type distributions, such as being dense (in weak convergence) on the set of distributions with positive support.

Multivariate phase-type theory for the site frequency spectrum.

Linear functions of the site frequency spectrum (SFS) play a major role for understanding and investigating genetic diversity. Estimators of the mutation rate (e.g. based on the total number of segregating sites or average of the pairwise differences) and tests for neutrality (e.g. Tajima'sD) are perhaps the most well-known examples. The distribution of linear functions of the SFS is important for constructing confidence intervals for the estimators, and to determine significance thresholds for neutrality tests. These distributions are often approximated using simulation procedures. In this paper we use multivariate phase-type theory to specify, characterize and calculate the distribution of linear functions of the site frequency spectrum. In particular, we show that many of the classical estimators of the mutation rate are distributed according to a discrete phase-type distribution. Neutrality tests, however, are generally not discrete phase-type distributed. For neutrality tests we derive the probability generating function using continuous multivariate phase-type theory, and numerically invert the function to obtain the distribution. A main result is an analytically tractable formula for the probability generating function of the SFS. Software implementation of the phase-type methodology is available in the R package PhaseTypeR, and R code for the reproduction of our results is available as an accompanying vignette.

Fractional inhomogeneous multi-state models in life insurance

In this paper, we demonstrate through the use of matrix calculus a transparent analysis of fractional inhomogeneous Markov models for life insurance where transition matrices commute. The resulting formulae are intuitive matrix generalizations of their single-state counterparts, and the absorption times are matrix versions of well-known scalar distributions. A further advantage of this approach is that it allows extending the analysis to the non-Markovian case where sojourns are Mittag-Leffler distributed, and where the absorption times are fractional phase-type distributed. Considering deterministic time transforms gives rise to fractional inhomogeneous phase-type distributions as absorption times. The latter underlying processes are an example of a regime where not only the present but also the history of a policyholder influences its future evolution. The sub-exponential nature of stable distributions translates into the multi-state insurance model as a random longevity risk at any given state of the chain.

Concomitants of order statistics from bivariate phase-type distributions with continuous density functions

Consider a random sample from a bivariate random variable. If the sample is ordered according to the Y-variates, then the X-variate paired with the r-th order statistic is called the concomitant of the r-th order statistic (or the r-th concomitant). Given a sample of independent, identically distributed and bivariate phase-type random variables with continuous density function, we provide a procedure to calculate the density function of concomitants of order statistics, and furthermore, we prove that concomitants from this sample are phase-type distributed.

Maximizing the expected net present value of a project with phase-type distributed activity durations: An efficient globally optimal solution procedure

We study projects with activities that have stochastic durations that are modeled using phase-type distributions. Intermediate cash flows are incurred during the execution of the project. Upon completion of all project activities a payoff is obtained. Because activity durations are stochastic, activity starting times cannot be defined at the start of the project. Instead, we have to rely on a policy to schedule activities during the execution of the project. The optimal policy schedules activities such that the expected net present value of the project is maximized. We determine the optimal policy using a new continuous-time Markov chain and a backward stochastic dynamic program. Although the new continuous-time Markov chain allows to drastically reduce memory requirements (when compared to existing methods), it also allows activities to be preempted; an assumption that is not always desirable. We prove, however, that it is globally optimal not to preempt activities if cash flows are incurred at the start of an activity. Moreover, this proof holds regardless of the duration distribution of the activities. A computational experiment shows that we significantly outperform current state-of-the-art procedures. On average, we improve computational efficiency by a factor of 600, and reduce memory requirements by a factor of 321.

Maximizing the Expected Net Present Value of a Project with Phase-Type Distributed Activity Durations: An Efficient Globally Optimal Solution Procedure

We study projects with activities that have stochastic durations that are modeled using phase-type distributions. Intermediate cash flows are incurred during the execution of the project. Upon completion of all project activities a payoff is obtained. Because activity durations are stochastic, activity starting times cannot be defined at the start of the project. Instead, we have to rely on a policy to schedule activities during the execution of the project. The optimal policy schedules activities such that the expected net present value of the project is maximized. We determine the optimal policy using a new continuous-time Markov chain and a backward stochastic dynamic program. Although the new continuous-time Markov chain allows to drastically reduce memory requirements (when compared to existing methods), it also allows activities to be preempted; an assumption that is not always desirable. We demonstrate, however, that it is globally optimal not to preempt activities if certain conditions are met. A computational experiment confirms this finding. The computational experiment also shows that we significantly outperform current state-of-the-art procedures. On average, we improve computational efficiency by a factor of 600, and reduce memory requirements by a factor of 321.

On dividends in the phase–type dual risk model

The dual risk model assumes that the surplus of a company decreases at a constant rate over time, and grows by means of upward jumps which occur at random times with random sizes. In the present work, we study the dual risk renewal model when the waiting times are phase-type distributed. Using the roots of the fundamental and the generalized Lundberg’s equations, we get expressions for the ruin probability and the Laplace transform of the time of ruin for an arbitrary single gain distribution. Then, we address the calculation of expected discounted future dividends particularly when the individual common gains follow a phase-type distribution. We further show that the optimal dividend barrier does not depend on the initial reserve. As far as the roots of the Lundberg equations and the time of ruin are concerned, we address the existing formulae in the corresponding Sparre-Andersen insurance risk model for the first hitting time, and we generalize them to cover also the situations where we have multiple roots. We do that working a new approach and technique, approach we also use for working the dividends, unlike others, it can be also applied for every situation.

Heuristic representation optimization for efficient generation of PH-distributed random variates

Phase-type (PH) distributions are being used to model a wide range of phenomena in performance and dependability evaluation. The resulting models may be employed in analytical as well as in simulation-driven approaches. Simulations require the efficient generation of random variates from PH distributions. PH distributions have different representations and different associated computational costs for pseudo random-variate generation (PRVG). In this paper we study the problem of efficient representation and efficient generation of PH distributed variates. We introduce various PH representations of different sizes and optimize them according to different cost functions associated with PRVG.

Two-layer symbolic representation for stochastic models with phase-type distributed events

Among the techniques that have been proposed for the analysis of non-Markovian models, the state space expansion approach showed great flexibility in terms of modelling capacities.The principal drawback is the explosion of the state space. This paper proposes a two-layer symbolic method for efficiently storing the expanded reachability graph of a non-Markovian model in the case in which continuous phase-type distributions are associated with the firing times of system events, and different memory policies are considered. At the lower layer, the reachability graph is symbolically represented in the form of a set of Kronecker matrices, while, at the higher layer, all the information needed to correctly manage event memory is stored in a multi-terminal multi-valued decision diagram. Such an information is collected by applying a symbolic algorithm, which is based on a couple of theorems. The efficiency of the proposed approach, in terms of memory occupation and execution time, is shown by applying it to a set of non-Markovian stochastic Petri nets and comparing it with a classical explicit expansion algorithm. Moreover, a comparison with a classical symbolic approach is performed whenever possible.

Moment distributions of phase-type (abstract only)

Both matrix-exponential and phase-type distributions have a number of important closure properties. Among those are the distributions of the age and residual life-time of a stationary renewal process with inter-arrivals of either type. In this talk we show that the spread, which is the sum of the age an residual life-time, is also phase-type distributed. Moreover, we give some explicit representations. The spread is known to have a first order moment distribution. If X is a positive random variable and ?i is its i'th moment, then the function fi(x) = xif(x)/?i is a density function, and the corresponding distribution is called the i'th order moment distribution. We prove that the classes of matrix-exponential or phase-type distributions are closed under the formation of moment distributions of any order. Other distributions which are closed under the formation of moment distributions are e.g. log-normal, Pareto and gamma distributions. We provide explicit representations for both the matrix-exponential class and for the phase-type distributions, where the latter class may also use the former representations, but for various reasons it is desirable to establish a phase-type representation when dealing with phase-type distributions. For the first order distribution we present an explicit formula for the related Lorenz curve and Gini index. Moment distributions of orders one, two and three have been extensively used in areas such as economy, physics, demography and civil engineering.

A complex discrete warm standby system with loss of units

A redundant complex discrete system is modelled through phase type distributions. The system is composed of a finite number of units, one online and the others in a warm standby arrangement. The units may undergo internal wear and/or accidental external failures. The latter may be repairable or non-repairable for the online unit, while the failures of the standby units are always repairable. The repairability of accidental failures for the online unit may be independent or not of the time elapsed up to their occurrence. The times up to failure of the online unit, the time up to accidental failure of the warm standby ones and the time needed for repair are assumed to be phase-type distributed. When a non-repairable failure occurs, the corresponding unit is removed. If all units are removed, the system is then reinitialized. The model is built and the transient and stationary distributions determined. Some measures of interest associated with the system, such as transition probabilities, availability and the conditional probability of failure are achieved in transient and stationary regimes. All measures are obtained in a matrix algebraic algorithmic form under which the model can be applied. The results in algorithmic form have been implemented computationally with Matlab. An optimization is performed when costs and rewards are present in the system. A numerical example illustrates the results and the CPU (Central Processing Unit) times for the computation are determined, showing the utility of the algorithms.

Correlated phase-type distributed random numbers as input models for simulations

The adequate modeling of correlated input processes is an important step in building simulation models. Modeling independent identically distributed data is well established in simulation whereas the integration of correlation is still a challenge. In this paper, ARTA processes which have been used several times for describing correlated input processes in simulation are extended by using ARMA instead of AR processes to realize the correlation and Acyclic Phase Type distributions to model the marginal distribution. For this new process type a fitting algorithm is presented. By means of some real network traces it is shown that the extended model allows a better fitting of the marginal distribution as well as the correlation structure and results in a compact process description that can be used in simulation models.

The use of vector-valued martingales in risk theory

A completely dependent risk process with perturbation and phase-type distributed claim sizes is analyzed. Claim arrivals are modeled by a Markovian arrival process. Using a vector-valued martingale, the Laplace transform of the time to ruin is derived algorithmically. The conditional memoryless property of the phase-type distribution yields the distribution of the deficit at ruin as a corollary.

On an (s, S) Inventory Policy with Service Time, Vacation to Server and Correlated Lead Time

We consider an inventory with random positive service time. The control policy is of (s, S) type. Customer demands constitute a Markovian arrival process (MAP). Service times have phase type distribution. Lead time for replenishment of inventory follows a correlated process similar to the customer arrival process. At a service completion epoch if no customer is waiting or the inventory level is zero then the server goes on a vacation of random duration. The vacation time is also phase type distributed. On return from vacation if the server finds no customer waiting or inventory level is zero, the server goes on another vacation of random duration having the same phase type distribution as the earlier one; we assume that the successive vacation times are independent of each other. This goes on until on return he finds at least one customer and at least one item in the inventory at which he starts serving. The customer arrival process is subject to balking. While waiting for service, customers may become impatient and leave the system (renege). Under these assumptions the underlying level dependent quasi birth-death process (LDQBD) is analysed. Several particular cases are considered in which the system is exhaustively studied. Numerical illustrations are provided. A cost function is introduced to numerically evaluate the optimal values of s and S.

Shock and wear models under policy N using phase-type distributions

We consider a device submitted to shocks and wear. The shocks occur with interarrival times phase-type distributed, and can induce the device failure. The device also can fail due to ageing, and the lifetime follows a phase-type distribution. The interarrival times between shocks and the lifetime depend on the number of cumulated shocks. The number of shocks that the device can support is limited. For this model, the survival probability is calculated. Other models are considered under these assumptions, and the survival functions are determined and expressed in a well-structured form. A numerical application illustrates the methods introduced in the paper.

Replacement times and costs in a degrading system with several types of failure: The case of phase-type holding times

A reliability system submitted to external and internal failures, that can be repairable or non-repairable, with degradation levels, and with sojourn times phase-type distributed, is considered. Repair is not as good as new, and the repair of internal failure follows policy N, that is, after N completed repairs the system is replaced by a new one to the following failure, repairable or not. For this system, a Markov model is constructed, and the stationary probability vector is calculated. It is shown that the distribution of the time between two consecutive replacements follows a phase-type distribution, whose representation is determined. The costs of these periods are calculated. An optimization problem involving the costs, the availability, and the number of internal repairs is illustrated by a numerical example.

A deteriorating two-system with two repair modes and sojourn times phase-type distributed

We study a two-unit cold standby system in steady-state. The online unit goes through a finite number of stages of successive degradation preceding the failure. The units are reparable, there is a repairman and two types of maintenance are considered, preventive and corrective. The preventive repair aims to improve the degradation of a unit being operative. The corrective repair is necessary when the unit fails. We will assume that the preventive repair will be interrupted in favour of a corrective repair in order to increase the availability of the system. The random operational and repair times follow phase-type distributions. For this system, the stationary probability vector, the replacement times, and the involved costs are calculated. An optimisation problem is illustrated by a numerical example. In this, the optimal degradation stage for the preventive repair of the online unit is determined by taking into account the system availability and the incurred costs.

Analysis of R-out-of-N repairable systems: the case of phase-type distributions

An R-out-of-N repairable system, consisting of N independent components, is operating if at least R components are functioning. The system fails whenever the number of good components decreases from R to R-1. A failed component is sent to a repair facility. After a failed component has been repaired it is as good as new. Formulae for the availability of the system using Markov renewal and semi-regenerative processes are derived. We assume that either the repair times of the components are generally distributed and the components' lifetimes are phase-type distributed or vice versa. Some duality results between the two systems are obtained. Numerical examples are given for several distributions of lifetimes and of repair times.