Stochastic

Processes with
Applications
Edited by
Antonio Di Crescenzo, Claudio Macci and Barbara Martinucci
Printed Edition of the Special Issue Published in Mathematics

www.mdpi.com/journal/mathematics
Stochastic Processes with Applications
Stochastic Processes with Applications

Special Issue Editors

Antonio Di Crescenzo
Claudio Macci
Barbara Martinucci

MDPI • Basel • Beijing • Wuhan • Barcelona • Belgrade

Special Issue Editors
Antonio Di Crescenzo Claudio Macci Barbara Martinucci
University of Salerno Università di Roma Tor Vergata University of Salerno
Italy Italy Italy

Editorial Office
MDPI
St. Alban-Anlage 66
4052 Basel, Switzerland

This is a reprint of articles from the Special Issue published online in the open access journal
Mathematics (ISSN 2227-7390) from 2017 to 2018 (available at: https://www.mdpi.com/journal/
mathematics/special issues/Stochastic Processes Applications).

For citation purposes, cite each article independently as indicated on the article page online and as
indicated below:

LastName, A.A.; LastName, B.B.; LastName, C.C. Article Title. Journal Name Year, Article Number,
Page Range.

ISBN 978-3-03921-728-1 (Pbk)

ISBN 978-3-03921-729-8 (PDF)

Cover image courtesy of Antonio Di Crescenzo.


c 2019 by the authors. Articles in this book are Open Access and distributed under the Creative
Commons Attribution (CC BY) license, which allows users to download, copy and build upon
published articles, as long as the author and publisher are properly credited, which ensures maximum
dissemination and a wider impact of our publications.
The book as a whole is distributed by MDPI under the terms and conditions of the Creative Commons
license CC BY-NC-ND.
Contents

About the Special Issue Editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Preface to ”Stochastic Processes with Applications” . . . . . . . . . . . . . . . . . . . . . . . . . ix

Mario Abundo
The Randomized First-Hitting Problem of Continuously Time-Changed Brownian Motion
Reprinted from: Mathematics 2018, 6, 91, doi:10.3390/math6060091 . . . . . . . . . . . . . . . . . . 1

Ravi Agarwal, Snezhana Hristova, Donal O’Regan and Peter Kopanov

Stability Analysis of Cohen–Grossberg Neural Networks with Random Impulses
Reprinted from: Mathematics 2018, 6, 144, doi:10.3390/math6090144 . . . . . . . . . . . . . . . . . 11

Giuseppina Albano and Virginia Giorno

On Short-Term Loan Interest Rate Models: A First Passage Time Approach
Reprinted from: Mathematics 2018, 6, 70, doi:10.3390/math6050070 . . . . . . . . . . . . . . . . . . 23

Giacomo Ascione, Nikolai Leonenko and Enrica Pirozzi

Fractional Queues with Catastrophes and Their Transient Behaviour
Reprinted from: Mathematics 2018, 6, 159, doi:10.3390/math6090159 . . . . . . . . . . . . . . . . . 35

G. Ayyappan and S. Karpagam

An M [X] /G(a, b)/1 Queueing System with Breakdown and Repair, Stand-By Server,
Multiple Vacation and Control Policy on Request for Re-Service
Reprinted from: Mathematics 2018, 6, 101, doi:10.3390/math6060101 . . . . . . . . . . . . . . . . . 61

Raquel Caballero-Águila, Aurora Hermoso-Carazo and Josefa Linares-Pérez

Fusion Estimation from Multisensor Observations with Multiplicative Noises and Correlated
Random Delays in Transmission
Reprinted from: Mathematics 2017, 5, 45, doi:10.3390/math5030045 . . . . . . . . . . . . . . . . . . 79

Davide De Gaetano
Forecast Combinations in the Presence of Structural Breaks: Evidence from U.S. Equity Markets
Reprinted from: Mathematics 2018, 6, 34, doi:10.3390/math6030034 . . . . . . . . . . . . . . . . . . 99

Antonio Di Crescenzo and Patrizia Di Gironimo

Stochastic Comparisons and Dynamic Information of Random Lifetimes in a Replacement
Model
Reprinted from: Mathematics 2018, 6, 204, doi:10.3390/math6100204 . . . . . . . . . . . . . . . . . 118

Antonio Di Crescenzo, Virginia Giorno, Balasubramanian Krishna Kumar and

Amelia G. Nobile
A Time-Non-Homogeneous Double-Ended Queue with Failures and Repairs and Its
Continuous Approximation
Reprinted from: Mathematics 2018, 6, 81, doi:10.3390/math6050081 . . . . . . . . . . . . . . . . . . 131

Maria Gamboa and Maria Jesus Lopez-Herrero

On the Number of Periodic Inspections During Outbreaks of Discrete-Time Stochastic SIS
Epidemic Models
Reprinted from: Mathematics 2018, 6, 128, doi:10.3390/math6080128 . . . . . . . . . . . . . . . . . 154

v
Rita Giuliano and Claudio Macci
Large Deviation Results and Applications to the Generalized Cramér Model
Reprinted from: Mathematics 2018, 6, 49, doi:10.3390/math6040049 . . . . . . . . . . . . . . . . . . 167

Antonio Gómez-Corral, Martı́n López-Garcı́a

A Within-Host Stochastic Model for Nematode Infection
Reprinted from: Mathematics 2018, 6, 143, doi:10.3390/math6090143 . . . . . . . . . . . . . . . . . 179

Alexander I. Nazarov and Yakov Yu. Nikitin

On Small Deviation Asymptotics In L2 of Some Mixed Gaussian Processes
Reprinted from: Mathematics 2018, 6, 55, doi:10.3390/math6040055 . . . . . . . . . . . . . . . . . . 203

Luca Pratelli and Pietro Rigo

Convergence in Total Variation to a Mixture of Gaussian Laws
Reprinted from: Mathematics 2018, 6, 99, doi:10.3390/math6060099 . . . . . . . . . . . . . . . . . . 212

Patricia Román-Román, Juan José Serrano-Pérez and Francisco Torres-Ruiz

Some Notes about Inference for the Lognormal Diffusion Process with Exogenous Factors
Reprinted from: Mathematics 2018, 6, 85, doi:10.3390/math6050085 . . . . . . . . . . . . . . . . . . 226

Anna Sinitcina, Yacov Satin, Alexander Zeifman, Galina Shilova, Alexander Sipin, Ksenia
Kiseleva, Tatyana Panfilova, Anastasia Kryukova, Irina Gudkova and Elena Fokicheva
On the Bounds for a Two-Dimensional Birth-Death Processwith Catastrophes
Reprinted from: Mathematics 2018, 6, 80, doi:10.3390/math6050080 . . . . . . . . . . . . . . . . . . 239

Somayeh Zarezadeh, Somayeh Ashrafi and Majid Asadi

Network Reliability Modeling Based on a Geometric Counting Process
Reprinted from: Mathematics 2018, 6, 197, doi:10.3390/math6100197 . . . . . . . . . . . . . . . . . 256

vi
About the Special Issue Editors
Antonio Di Crescenzo, Full Professor of Probability and Mathematical Statistics at the University
of Salerno, received his degree in Information Science in 1988 from University of Salerno, and Ph.D.
degree in Applied Mathematics and Informatics in 1996 from University of Naples Federico II.
He has served as Research Associate at the University of Naples Federico II (1991–1998) and
Associate Professor firstly at the University of Basilicata, Potenza (1998–2001) followed by the
University of Salerno (2001–2018). He has published more than 90 research papers in probability and
stochastic models. He co-organized several international meetings and participated in more than
60 international scientific meetings, especially regarding applications of probability. He is member
of the Editorial Board of various journals, including Mathematics (MDPI) and Methodology and
Computing in Applied Probability (Springer). His research interests include stochastic modeling, theory
and simulation of stochastic processes with applications to biomathematics, queueing systems,
reliability theory, aging notions, and information measures.

Claudio Macci, Associate Professor, received his M.S. degree in Mathematics in 1993 from Sapienza
University of Rome, and Ph.D. degree in Mathematics from Tor Vergata University of Rome in 1998.
He served as Assistant Professor at University of Turin from 1999 to 2003 before moving to Tor
Vergata University of Rome, where has served as Associate Professor since 2013. He has published
more than 80 papers in probability and statistics. His main research interest is in large deviations
theory. Other research interests include Bayesian statistics, ruin probabilities in insurance, and
fractional processes.

Barbara Martinucci Assistant Professor, received her M.S. degree in Mathematics in 2000 and earned
her Ph.D. in Applied Mathematics and Informatics in 2005 from University of Naples Federico
II. She has been serving as Assistant Professor of Probability and Mathematical Statistics at the
Department of Mathematics of the University of Salerno since her appointment in November 2008.
Her research interests lie in the general area of stochastic processes with a focus on random evolutions
with finite velocity.

vii
Preface to ”Stochastic Processes with Applications”
The aim of the Special Issue “Stochastic Processes with Applications” is to present a collection
of original papers covering recent advances in the theory and application of stochastic processes.
The focus is especially on applications of stochastic processes as models of dynamic phenomena in
various research areas, such as economics, statistical physics, queuing theory, biology, theoretical
neurobiology, and reliability theory.
The volume contains 17 articles collected from June 2017 to September 2018. We appreciate
that the contributions originate from three continents. Indeed, the geographical distribution of the
46 authors is as follows: Italy (13), Russia (12), Spain (9), India (3), Iran (3), Bulgaria (2), UK (2),
Ireland (1), and USA (1). It is a pleasure to note that some of the authors that contributed to this
volume subsequently served as Editors of other Special Issues of Mathematics on similar topics.
The stochastic processes treated in this book range within quite wide areas, such as diffusion
and Gaussian processes, stochastic volatility models, epidemic models, neural networks, counting
processes, fractional processes, Markov chains, and birth–death processes. Some investigations
also involve stochastic processes describing phenomena subject to catastrophes and repairs, and
related first-passage-time problems. Asymptotic results are also treated, as related to large and small
deviations and convergence problems for mixed Gaussian laws. Various contributions are based on
simulation tools or on statistical methods, such as maximum likelihood estimation and hypothesis
testing. Computational methods are also adopted and are mainly based on recursive algorithms for
estimation or forecasting. Some articles of the volume deal with results concerning applied fields,
with special reference to mathematical finance, biomathematics, queueing, and reliability theory.
We are confident that the research presented herein will attract the interest of Mathematics readers
and of the large community of scholars active in the realm of stochastic processes and applications.
Finally, we would like to express our deep gratitude to the researchers who contributed to this
volume. Moreover, our warm thanks go to those who reviewed the manuscripts. The refereeing
activity was essential to the realization of this Special Issue.

Antonio Di Crescenzo, Claudio Macci, Barbara Martinucci

Special Issue Editors

ix
 mathematics
Article
The Randomized First-Hitting Problem of
Continuously Time-Changed Brownian Motion
Mario Abundo
Dipartimento di Matematica, Università Tor Vergata, 00133 Rome, Italy; [email protected];
Tel.: +390672594627

Received: 4 April 2018; Accepted: 25 May 2018; Published: 28 May 2018

Abstract: Let X (t) be a continuously time-changed Brownian motion starting from a random position
η, S(t) a given continuous, increasing boundary, with S(0) ≥ 0, P(η ≥ S(0)) = 1, and F an assigned
distribution function. We study the inverse first-passage time problem for X (t), which consists in
finding the distribution of η such that the first-passage time of X (t) below S(t) has distribution
F, generalizing the results, valid in the case when S(t) is a straight line. Some explicit examples
are reported.

Keywords: first-passage time; inverse first-passage problem; diffusion

1. Introduction
This brief note is a continuation of [1,2]. Let σ(t) be a regular enough non random function,
t
and let X (t) = η + 0 σ (s)dBs , where Bt is standard Brownian motion (BM) and the initial position
t
η is a random variable, independent of Bt . Suppose that the quadratic variation ρ(t) = 0 σ2 (s)ds is
 
increasing and ρ(+∞) = ∞, then there exists a standard BM B such that X (t) = η + B(ρ(t)), namely
X (t) is a continuously time-changed BM (see e.g., [3]). For a continuous, increasing boundary S(t),
such that P(η ≥ S(0)) = 1, let

τ = τS = inf{t > 0 : X (t) ≤ S(t)} (1)

be the first-passage time (FPT) of X (t) below S. We assume that τ is finite with probability one and
dF (t)
that it possesses a density f (t) = dt , where F (t) = P(τ ≤ t). Actually, the FPT of continuously
time-changed BM is a well studied problem for constant or linear boundary and a non-random initial
value (see e.g., [4–6]).
Assuming that S(t) is increasing, and F (t) is a continuous distribution function, we study the
following inverse first-passage-time (IFPT) problem:

given a distribution F, find the density g of η (if it exists) for which it results P(τ ≤ t) = F (t).

The function g is called a solution to the IFPT problem. This problem, also known as the
generalized Shiryaev problem, was studied in [1,2,7,8], essentially in the case when X (t) is BM and
S(t) is a straight line; note that the question of the existence of the solution is not a trivial matter (see
e.g., [2,7]). In this paper, by using the properties of the exponential martingale, we extend the results
to more general boundaries S.
The IFPT problem has interesting applications in mathematical finance , in particular in credit
risk modeling, where the FPT represents a default event of an obligor (see [7]) and in diffusion models
for neural activity ([9]).
Notice, however, that another type of inverse first-passage problem can be considered: it consists
in determining the boundary shape S, when the FPT distribution F and the starting point η are assigned
(see e.g., [10–13]).

Mathematics 2018, 6, 91; doi:10.3390/math6060091 1 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 91

The paper is organized as follows: Section 2 contains the main results, in Section 3 some explicit
examples are reported; Section 4 is devoted to conclusions and final remarks.

2. Main Results
The following holds:

Theorem 1. Let be S(t) a continuous, increasing boundary with S(0) ≥ 0, σ(t) a bounded, non random
t
continuous function of t > 0, and let X (t) = η + 0 σ(s)dBs be the integral process starting from the random
t 2
position η ≥ S(0); we assume that ρ(t) = 0 σ (s)ds is increasing and satisfies ρ(+∞) = +∞. Let F be
the probability distribution of the FPT τS of X below the boundary S (τS is a.s. finite by virtue of Remark 3).
We suppose that the r.v. η admits a density g( x ); for θ > 0, we denote by g(θ ) = E(e−θη ) the Laplace transform
of g.
Then, if there exists a solution to the IFPT problem for X, the following relation holds:
 +∞
θ 2 ρ(t)
g(θ ) = e−θS(t)− 2 dF (t). (2)
0

Proof. The process X (t) is a martingale, we denote by Ft its natural filtration. Thanking to the
hypothesis, by using the Dambis, Dubins–Schwarz theorem (see e.g., [3]), it follows that the process
(t) = X (ρ−1 (t)) is a Brownian motion with respect to the filtration F −1 ; so the process X (t) can
B ρ (t)
be written as X (t) = η + B(ρ(t)) and the FPT τ can be written as τ = inf{t > 0 : η + B
(ρ(t)) ≤ S(t)}.
For θ > 0, let us consider the process Zt = e−θX (t)− 2 θ ρ(t) ; as easily seen, Zt is a positive martingale;
1 2

t
indeed, it can be represented as Zt = e−θX (0) − θ 0 Zs σ(s)dBs (see e.g., Theorem 5.2 of [14]).
We observe that, for t ≤ τ the martingale Zt is bounded, because X (t) is non negative and
therefore 0 < Zt ≤ e−θX (t) ≤ 1. Then, by using the fact that, for any finite stopping time τ one
has E[ Z0 ] = E[ Zτ ∧t ] (see e.g., Formula (7.7) in [14]), and the dominated convergence theorem,
we obtain that
E[ Z0 ] = E[e−θX (0) ] = E[e−θη ] = lim E[e−θX (τ ∧t)− 2 θ ρ(τ ∧t) ]
1 2

t→∞

−θX (τ ∧t)− 12 θ 2 ρ(τ ∧t)

] = E[e−θS(τ )− 2 θ
1 2 ρ(τ )
= E[ lim e ]. (3)
t→∞

Thus, if g(θ ) = E(e−θη ) is the Laplace transform of the density of the initial position η,
we finally get  
θ2
g(θ ) = E e−θS(τ )− 2 ρ(τ ) , (4)

that is Equation (2).

Remark 1. If one takes in place of X (t) a process of the form X  (t) = ηS(t) + S(t) B(ρ(t)), with η ≥ 1,
that is, a special case of continuous Gauss-Markov process ([15]) with mean ηS(t), then X  (t)/S(t) is still a
continuously time-changed BM, and so the IFPT problem for X  (t) and S(t) is reduced to that of continuously
time-changed BM and a constant barrier, for which results are available (see e.g., [4–6]).

Remark 2. By using Laplace transform inversion (when it is possible), Equation (4) allows to find the solution
g to the IFPT problem for X, the continuous increasing boundary S, and the distribution F of the FPT τ. Indeed,
some care has to be used to exclude that the found distribution of η has atoms together with a density. However,
as already noted in [2,7], the function g may not be the Laplace transform of some probability density function,
so in that case the IFPT problem has no solution; really, it may admit more than one solution, since the right-hand
member of Equation (4) essentially furnishes the moments of η of any order n, but this is not always sufficient
to uniquely determine the density g of η. In line of principle, the right-hand member of Equation (4) can be
expressed in terms of the Laplace transform of f (t) = F  (t), though it is not always possible to do this explicitly.

2
Mathematics 2018, 6, 91

A simple case is when S(t) = a + bt, with a, b ≥ 0, and ρ(t) = t, that is, X (t) = Bt (σ(t) = 1); in fact,
one obtains   

θ2 θ θ (θ + 2b)
g(θ ) = E e−θ ( a+bτ )− 2 τ = e−θa E e−θ (b+ 2 )τ = e−θa f , (5)
2
which coincides with Equation (2.2) of [2], and it provides a relation between the Laplace transform of the density
of the initial position η and the Laplace transform of the density of the FPT τ.

Remark 3. Let S(t) be increasing and S(0) ≥ 0, then τ is a.s. finite; in fact τ = ρ(τ ) = inf{t > 0 :
η+B t ≤ S(t)} ≤ τ1 , where S(t) = S(ρ−1 (t)) is increasing and τ1 is the first hitting time to S(0) of BM

B starting at η; since τ1 is a.s. finite, also τ is so. Next, from the finiteness of τ it follows that τ = ρ−1 (τ)
is finite, too. Moreover, if one seeks that E(τ ) < ∞, a sufficient condition for this is that ρ(t) and S(t) are
both convex functions; indeed, τ ≤ τ2 , where τ2 is the FPT of BM B  starting from η below the straight line
a + bt ( a = S(0) ≥ 0, b = S (0) ≥ 0) which is tangent to the graph of S(t) at t = 0. Thus, since E(τ2 ) < ∞,
it follows that E(τ) is finite, too; finally, being ρ−1 concave, Jensen’s inequality for concave functions implies
that E(τ ) = E(ρ−1 (τ)) ≤ ρ−1 ( E(τ)) and therefore E(τ ) < ∞.

Remark 4. Theorem 1 allows to solve also the so called Skorokhod embedding (SE) problem:
Given a distribution H, find an integrable stopping time τ ∗ , such that the distribution of X (τ ∗ ) is H,
namely P( X (τ ∗ ) ≤ x ) = H ( x ).
In fact, let be S(t) increasing, with S(0) = 0; first suppose that the support of H is [0, +∞); then,
from Equation (4) it follows that
θ 2 ρ ( S−1 ( X ( τ )))
g(θ ) = E[e−θX (τ )− 2 ], (6)

and this solves the SE problem with τ∗= τ; it suffices to take the random initial point X (0) = η > 0 in such a
way that its Laplace transform g satisfies
 S(+∞)
θ 2 ρ ( S−1 ( x ))
g(θ ) = e−θx− 2 dH ( x ). (7)
0

In the special case when S(t) = a + bt ( a, b > 0) and ρ(t) = t, Equation (7) becomes (cf. the result in [8]
for a = 0) :

aθ 2 θ (θ + 2b)
g(θ ) = e 2b  h , (8)
2b

where h( x ) = H  ( x ) and 
h denotes the Laplace transform of h.
In analogous way, the SE problem can be solved if the support of H is (−∞, 0]; now, the FPT is understood
as τ − = inf{t > 0 : η + B(ρ(t)) > −S(t)} (η < 0), that is, the first hitting time to the boundary
S− (t) = −S(t) from below.
Therefore, the solution to the general SE problem, namely without restrictions on the support of the
distribution H, can be obtained as follows (see [8], for the case when S(t) is a straight line).
The r.v. X (τ ) can be represented as a mixture of the r.v. X + > 0 and X − < 0 :

X+ with probability p+ = P( X (τ ) ≥ 0)
X (τ ) = (9)
X− with probability p− = 1 − p+ .

Suppose that the SE problem for the r.v. X + and X − can be solved by S+ (t) = S(t) and η + = η > 0, and
S− (t) = −S(t) and η − = −η < 0, respectively. Then, we get that the r.v.

η+ with probability p+
η± = (10)
η− with probability p−

and the boundary S± (t) = S+ (t) ∪ S− (t) solve the SE problem for the r.v. X (τ ).

3
Mathematics 2018, 6, 91

If g is analytic in a neighbor of θ = 0, then the moments of order n of η, E(η n ), exist finite, and
dn 

they are given by E(η n ) = (−1)n dθ n g θ =0 . By taking the first derivative in Equation (4) and calculating

it at θ = 0, we obtain
E(η ) = − g (0) = E(S(τ )). (11)

By calculating the second derivative of g at θ = 0, we get

E(η 2 ) = g (0) = E(S2 (τ ) − ρ(τ ))), (12)

and so
Var (η ) = E(η 2 ) − E2 (η ) = Var (S(τ )) − E(ρ(τ )). (13)

Thus, we obtain the compatibility conditions


E(η ) = E(S(τ ))
(14)
Var (S(τ )) ≥ E(ρ(τ )).

If Var (S(τ )) < E(ρ(τ )), a solution to the IFPT problem does not exist. In the special case
when S(t) = a + bt ( a, b ≥ 0) and ρ(t) = t, Equation (11) becomes E(η ) = a + bE(τ ) and
Equation (13) becomes Var (η ) = b2 Var (τ ) − E(τ ), while Equation (14) coincides with Equation
(2.3) of [2]. By writing the Taylor’s expansions at θ = 0 of both members of Equation (4), and equaling
the terms with the same order in θ, one gets the successive derivatives of g(θ ) at θ = 0; thus, one can
write any moment of η in terms of the expectation of a function of τ; for instance, it is easy to see that

E(η 3 ) = E[(S(τ ))3 ] − 3E[S(τ )ρ(τ )], (15)

E(η 4 ) = E[(S(τ )4 ] − 6E[(S(τ )2 ρ(τ )] + 3E[(ρ(τ )2 ], (16)

E(η ) = E[15S(τ )ρ (τ ) − 240S (τ )ρ(τ ) + S (τ )].

5 2 3 5
(17)

2.1. The Special Case S(t) = α + βρ(t)

If S(t) = α + βρ(t), with α, β ≥ 0, from Equation (4) we get

θ 2 ρ(τ )
g(θ ) = E[e−θ (α+ βρ(τ ))− 2 ] = e−θα E[e−θρ(τ )( β+θ/2) ]. (18)

Thus, setting τ = ρ(τ ), we obtain (see Equation (5)):


g(θ ) = e−θα E[e−θ ( β+θ/2)τ ] = e−θα f(θ ( β + θ/2)), (19)

having denoted by f the density of τ. In this way, we reduce the IFPT problem of X (t) = η + B(ρ(t))
below the boundary S(t) = α + βρ(t) to that of BM below the linear boundary α + βt. For instance,
β
taking ρ(t) = t3 /3, the solution to the IFPT problem of X (t) through the cubic boundary S(t) = α + 3 t3 ,
and the FPT density f , is nothing but the solution to the IFPT problem of BM through the linear
boundary α + βt, and the FPT density f.
Under the assumption that S(t) = α + βρ(t), with α, β ≥ 0, a number of explicit results can be
obtained, by using the analogous ones which are valid for BM and a linear boundary (see [2]). As for
the question of the existence of solutions to the IFPT problem, we have:

Proposition 1. Let be S(t) = α + βρ(t), with α, β ≥ 0; for γ, λ > 0, suppose that the FPT density f = F 
is given by ⎧
⎨ λγ ρ(t)γ−1 e−λρ(t) ρ (t) i f t > 0
f ( t ) = Γ(γ) (20)
⎩0 otherwise

4
Mathematics 2018, 6, 91

(namely the density√f of τ is the Gamma density with parameters (γ, λ)). Then, the IFPT problem has solution,
provided that β ≥ 2λ, and the Laplace transform of the density g of the initial position η is given by:
     
( β − β2 − 2λ)γ ( β + β2 − 2λ)γ
g(θ ) = e−αθ/2  · e−αθ/2  , (21)
(θ + β − β2 − 2λ)γ (θ + β + β2 − 2λ)γ

 Z2 , such that Zi − α/2

which is the Laplace transform of the sum of two independent random variables, Z1 and
has
 distribution Gamma of parameters γ and λ i ( i = 1, 2 ) , where λ 1 = β − β2 − 2λ and λ2 = β +
β2 − 2λ.

Remark 5. If f is given by√Equation (20), that is f is the Gamma density, √ the compatibility condition in
Equation (14) becomes β ≥ λ, which is satisfied under the assumption β ≥ 2λ required by Proposition 1.
In the special case when γ = 1, then η has the same distribution as α + Z1 + Z2 , where Zi are independent and
exponential with parameter λi , i = 1, 2.

The following result also follows from Proposition 2.5 of [2].

Proposition 2. Let be S(t) = α + βρ(t), with α, β ≥ 0; for β > 0, suppose that the Laplace transform of f
has the form:
N
 Ak
f (θ ) = ∑ , (22)
k =1
( θ + Bk )ck

for some ck > 0, Ak , Bk > 0, k = 1, . . . , N. Then, there exists a value β∗ > 0 such that the solution to the
IFPT problem exists, provided that β ≥ b∗ .
If β = 0 and the Laplace transform of f has the form:

N
 Ak
f (θ ) = ∑ √
( 2θ + Bk )ck
, (23)
k =1

then, the solution to the IFPT problem exists.

2.2. Approximate Solution to the IFPT Problem for Non Linear Boundaries
Now, we suppose that there exist α1 , α2 , β 1 , β 2 with 0 ≤ α1 ≤ α2 and β 2 ≥ β 1 ≥ 0, such that, for
every t ≥ 0 :
α1 + β 1 ρ ( t ) ≤ S ( t ) ≤ α2 + β 2 ρ ( t ), (24)

namely S(t) is enveloped from above and below by the functions Sα2 ,β2 (t) = α2 + β 2 ρ(t) and
Sα1 ,β1 (t) = α1 + β 1 ρ(t).
Then, by using Proposition (3.13) of [16] (see also [1]), we obtain the following:

Proposition 3. Let S(t) a continuous, increasing boundary satisfying Equation (24) and suppose that the
FPT τ of X (t) = η + B(ρ(t)) (η > S(0)) below the boundary S(t) has an assigned probability density f and
that there exists a density g with support (S(0), +∞), which is solution to the IFPT problem for X (t) and the

boundary S(t); as before, denote by f(t) the density of ρ(τ ) and by f(θ ) its Laplace transform, for θ > 0. Then:

(i) If α2 > α1 and the function g ∈ L p (S(0), α2 ) for some p > 1, its Laplace transform g(θ ) must satisfy:


 α 1/p 
 θ (θ + 2β 2 ) p −1 2
e−α2 (θ +2( β2 − β1 )) f − (α2 − S(0)) p g p ( x )dx ≤ g(θ )
2 S (0)


θ (θ + 2β 1 )
≤ e−α1 θ f ; (25)
2

5
Mathematics 2018, 6, 91

(ii) If α1 = α2 = S(0), then Equation (25) holds without any further assumption on g (and the term
p −1   1/p
α2
(α2 − S(0)) p p
S(0) g ( x ) dx vanishes).

Remark 6. The smaller α2 − α1 and β 2 − β 1 , the better the approximation to the Laplace transform of g. Notice
p −1   1/p
α2
that, if g is bounded, then the term (α2 − S(0)) p p
S(0) g ( x ) dx can be replaced with (α2 − S(0))|| g||∞ .

2.3. The IFPT Problem for X (t) = η + B(ρ(t))+ Large Jumps

As an application of the previous results, we consider now the piecewise-continuous process X (t),
obtained by superimposing to X (t) a jump process, namely we set X (t) = η + B(ρ(t)) for t < T, where
T is an exponential distributed time with parameter μ > 0; we suppose that, for t = T the process X (t)
makes a downward jump and it crosses the continuous increasing boundary S, irrespective of its state
before the occurrence of the jump. This kind of behavior is observed e.g. in the presence of a so called
catastrophes (see e.g., [17]). For η ≥ S(0), we denote by τ S = inf{t > 0 : X (t) ≤ S(t)} the FPT of X (t)
below the boundary S(t). The following holds:

Proposition 4. If there exists a solution g to the IFPT problem of X (t) below S(t) with X (0) = η ≥ S(0),
then its Laplace transform is given by
   +∞

 +∞

θ2 θ 2 ρ ( t )− μt
g(θ ) = E e−θS(τ )− 2 ρ(τ )−μτ + μ
 e−θS(t)− 2 f (s)ds dt. (26)
0 t

Proof. For t > 0, one has:

P(τ S ≤ t) = P(τ S ≤ t|t < T ) P(t < T ) + 1 · P(t ≥ T ) = P(τS ≤ t)e−μt + (1 − e−μt ). (27)

Taking the derivative, one obtains the FPT density of τ :

 +∞
f (t) = e−μt f (t) + μe−μt f (s)ds, (28)
t

where f is the density of τ. Then, by the same arguments used in the proof of Theorem 1, we obtain
 
θ2
g(θ ) = E e−θS(τ )− 2 ρ(τ )


 ∞
θ 2 ρ(t)
= e−θS(t)− 2 f (t)dt
0
 ∞   ∞ 
θ 2 ρ(t)
= e−θS(t)− 2 e−μt f (t) + μe−μt f (s)ds dt
0 t
 ∞ 2
 ∞ 2

 ∞

−θS(t)− θ2 ρ(t)−μt −θS(t)− θ2 ρ(t)−μt
= e f (t)dt + μ e f (s)ds dt
0 0 t

that is Equation (26).

Remark 7. (i) For μ = 0, namely when no jump occurs, Equation (26) becomes Equation (4).
(ii) If τ is exponentially distributed with parameter λ, then Equation (26) provides:
 
 λ+μ θ2
g(θ ) = E e−θS(τ )− 2 ρ(τ )−μτ . (29)
λ

6
Mathematics 2018, 6, 91

(iii) In the special case when S(t) = α + βρ(t) (α, β ≥ 0), we can reduce to the FPT τ of BM + large jumps
below the linear boundary α + βt; then, it is possible to write 
g in terms of the Laplace transform of τ.
Really, by using Proposition 3.10 of [16] one gets
  −1

 2μ f θ (θ + 2β) 2μ
g(θ ) = e−αθ 1− −μ − ,
θ (θ + 2β) 2 θ (θ + 2β) − 2μ

where, for simplicity of notation we have denoted again with f the Laplace transform of τ; of course,
if ρ(t) = t, then f is the Laplace transform of τ. Notice that, if μ = 0 the last equation is nothing but
Equation (5) with α, β in place of a, b.

3. Some Examples
Example 1. If S(t) = a + bt, with a, b ≥ 0, and X (t) = Bt (ρ(t) = 1), examples of solution to the IFPT
problem, for X (t) and various FPT densities f , can be found in [2].
Example 2. Let be S(t) = α + βρ(t), with α, β ≥ 0, and suppose that τ has density f (t) =
λe−ρ(t) ρ (t)1(0,+∞) (t) (that is, the density f of τ = ρ(τ ) is exponential with parameter λ). By using
Proposition 1 we get that η = α + Z1 + Z2 , where Zi are independent random − α/2 has
 variable, such that Zi 
exponential distribution with parameter λi (i = 1, 2), where λ1 = β − β2 − 2λ and λ2 = β + β2 − 2λ.
Then, the solution g to the IFPT problem for X (t) = η + B(ρ(t)), the boundary S and the exponential FPT
distribution, is:  λλ √
1 2
e−λ1 ( x−α) − e−λ2 ( x−α) , if b > 2λ
g ( x ) = λ2 − λ1 √ √ ( x ≥ α) (30)
2λ( x − α)e− 2λ( x− a) , if b = 2λ.

In general, for a given continuous increasing boundary S(t) and an assigned distribution of
τ, it is difficult to calculate explicitly the expectation on the right-hand member of Equation (4) to
get the Laplace transform of η. Thus, a heuristic solution to the IFPT problem can be achieved
by using Equation (4) to calculate the moments of η (those up to the fifth order are given by
Equations (11), (12) and (15)–(17)). Of course, even if one was able to find the moments of η of
any order, this would not determinate the distribution of η. However, this procedure is useful to study
the properties of the distribution of η, provided that the solution to the IFPT problem exists.

Example 3. Let be S(t) = t2 , ρ(t) = t and suppose that τ is exponentially distributed with parameter
λ; we search for a solution η > 0 to the IFPT problem by using the method of moments, described above.
The compatibility condition in Equation (14) requires that λ3 < 20 (for instance, one can take λ = 1).
From Equations (11), (12) and (15)–(17), and calculating the moments of τ up to the eighth order, we obtain:

2 24 − λ3 2 20 − λ3
E(η ) = E(τ 2 ) = ; E(η 2 ) = E(τ 4 ) − E(τ ) = ; σ (η ) = Var (η ) = ;
λ2 λ4 λ4

720 − 18λ3 8! − 36λ5 + 6λ6

E(η 3 ) = E(τ 6 ) − 3E(τ 3 ) = ; E(η 4 ) = E(τ 8 ) − 6E(τ 3 ) + 3E(τ 2 ) = .
λ6 λ8
Notice that, under the condition λ3 < 20 the first four moments of η are positive, as it must be. However,
they do not match those of a Gamma distribution.
An information about the asymmetry is given by the skewness value

E(η − E(η ))3 24 − λ3

= −12 < 0,
σ(η ) 3 (20 − λ3 )3/2

meaning that the candidate η has an asymmetric distribution with a tail toward the left.

7
Mathematics 2018, 6, 91

4. Conclusions and Final Remarks

We have dealt with the IFPT problem for a continuously time-changed Brownian motion X (t)
starting from a random position η. For a given continuous, increasing boundary S(t) with η ≥ S(0) ≥
0, and an assigned continuous distribution function F, the IFPT problem consists in finding the
distribution, or the density g of η, such that the first-passage time τ of X (t) below S(t) has distribution
F. In this note, we have provided some extensions of the results, already known in the case when X (t)
is BM and S(t) is a straight line, and we have reported some explicit examples. Really, the process we
t
considered has the form X (t) = η + 0 σ (s)dBs , where Bt is standard Brownian motion, and σ(t) is a
t
non random continuous function of time t ≥ 0, such that the function ρ(t) = 0 σ2 (s)ds is increasing
and it satisfies the condition ρ(+∞) = +∞. Thus, a standard BM B  exists such that X (t) = η + B(ρ(t)).
Our main result states that  
θ 2 ρ(τ )
g(θ ) = E e−θS(τ )− 2 , (31)

where, for θ > 0, g(θ ) denotes the Laplace transform of the solution g to the IFPT problem.
Notice that the above result can be extended to diffusions which are more general than the process
X (t) considered, for instance to a process of the form

(ρ(t)) + w(η )),

U ( t ) = w −1 ( B (32)

where w is a regular enough, increasing function; such a process U is obtained from BM by a space
transformation and a continuous time-change (see e.g., the discussion in [2]). Since w(U (t)) =
w(η ) + B (ρ(t)), the IFPT problem for the process U, the boundary S(t) and the FPT distribution
F, is reduced to the analogous IFPT problem for X (t) = η1 + B (ρ(t)), starting from η1 = w(η ),
instead of η, the boundary S1 (t) = w(S(t)) and the same FPT distribution F. When σ(t) = 1, i.e.
ρ(t) = t, the process U (t) is conjugated to BM, according to the definition given in [2]; two examples
of diffusions conjugated to BM are the Feller process, and the Wright–Fisher like (or CIR) process,
(see e.g., [2]). The process U (t) given by Equation (32) is indeed a weak solution of the SDE:

ρ (t)w (U (t)) ρ (t)
dU (t) = − dt + dBt , (33)
2(w (U (t)))3 w (U (t))

where w ( x ) and w ( x ) denote first and second derivative of w( x ).

Provided that the deterministic function ρ(t) is replaced with a random function, the representation
in Equation (32) is valid also for a time homogeneous one-dimensional diffusion driven by the SDE

dU (t) = μ(U (t))dt + σ(U (t))dBt , U (0) = η, (34)

where the drift (μ) and diffusion coefficients (σ ) satisfy the usual conditions (see e.g., [18]) for existence
and uniqueness of the solution of Equation (34). In fact, let w( x ) be the scale function associated to the
diffusion U (t) driven by the SDE Equation (34), that is, the solution of Lw( x ) = 0, w(0) = 0, w (0) = 1,
d2 h
where L is the infinitesimal generator of U given by Lh = 12 σ2 ( x ) dx dh
2 + μ ( x ) dx . As easily seen, if the
 t 2μ(z)
integral 0 σ2 (z) dz converges, the scale function is explicitly given by

 x
 t

2μ(z)
w( x ) = exp − dz dt. (35)
0 0 σ (z)
2

If ζ (t) := w(U (t)), by Itô’s formula one obtains

 t
ζ (t) = w(η ) + w (w−1 (ζ (s)))σ(w−1 (ζ (s)))dBs , (36)
0

8
Mathematics 2018, 6, 91

that is, the process ζ (t) is a local martingale, whose quadratic variation is
 t
.
ρ(t) = ζ t = [w (U (s))σ(U (s))]2 ds, t ≥ 0. (37)
0

The (random) function ρ(t) is differentiable and ρ(0) = 0; if it is increasing to ρ(+∞) = +∞,
by the Dambis, Dubins–Schwarz theorem (see e.g., [3]) one gets that there exists a standard BM B  such
(ρ(t)) + w(η ). Thus, since w is invertible, one obtains the representation in Equation (32).
that ζ (t) = B
Notice, however, that the IFPT problem for the process U given by Equation (32) cannot be
addressed as in the case when ρ is a deterministic function. In fact, if ρ(t) given by Equation (37) is
random, it results that ρ(t) and the FPT τ are dependent. Thus, in line of principle it would be possible
to obtain information about the Laplace transform of g, only in the case when the joint distribution of
(ρ(t), τ ) was explicitly known.

Funding: This research was funded by the MIUR Excellence Department Project awarded tothe Department of
Mathematics, University of Rome Tor Vergata, CUP E83C18000100006.
Acknowledgments: I would like to express particular thanks to the anonymous referees for their constructive
comments and suggestions leading to improvements of the paper.
Conflicts of Interest: The author declares no conflict of interest.

References
1. Abundo, M. Some randomized first-passage problems for one-dimensional diffusion processes. Sci. Math. Jpn.
2013, 76, 33–46.
2. Abundo, M. An inverse first-passage problem for one-dimensional diffusion with random starting point.
Stat. Probab. Lett. 2012, 82, 7–14. [CrossRef]
3. Revuz, D.; Yor, M. Continous Martingales and Brownian Motion; Springer: Berlin/Heidelberg, Germany, 1991.
4. Darling, D.A.; Siegert, A.J.F. The first passage problem for a continuous Markov process. Ann. Math. Stat.
1953, 24, 264–639. [CrossRef]
5. Hieber, P.; Scherer, M. A note on first-passage times of continuously time-changed Brownian motion.
Stat. Probab. Lett. 2012, 82, 165–172. [CrossRef]
6. Hurd, T.R. Credit risk modeling using time-changed Brownian motion. Int. J. Theor. Appl. Financ. 2009, 12,
1213–1230. [CrossRef]
7. Jackson, K.; Kreinin, A.; Zhang, W. Randomization in the first hitting problem. Stat. Probab. Lett. 2009, 79,
2422–2428. [CrossRef]
8. Jaimungal, S.; Kreinin, A.; Valov, A. The generalized Shiryaev problem and Skorokhod embedding.
Theory Probab. Appl. 2014, 58, 493–502. [CrossRef]
9. Lanska, V.; Smiths, C.E. The effect of a random initial value in neural first-passage-time models. Math. Biosci.
1989, 93, 191–215. [CrossRef]
10. Peskir, G. On integral equations arising in the first-passage problem for Brownian motion. J. Integral
Equat. Appl. 2002, 14, 397–423. [CrossRef]
11. Sacerdote, L.; Zucca, C. Threshold shape corresponding to a Gamma firing distribution in an
Ornstein-Uhlenbeck neuronal model. Sci. Math. Jpn. 2003, 19, 1319–1346.
12. Zucca, C.; Sacerdote, L. On the inverse first-passage-time problem for a Wiener process. Ann. Appl. Probab.
2009, 8, 375–385. [CrossRef]
13. Abundo, M. Limit at zero of the first-passage time density and the inverse problem for
one-dimensional diffusions. Stoch. Anal. Appl. 2006, 24, 1119–1145. [CrossRef]
14. Klebaner, F.C. Introduction to Stochastic Calculus With Applications; Imperial College Press: London, UK, 2005.
15. Di Nardo, E.; Nobile, A.G.; Pirozzi, E.; Ricciardi, L.M. A computational approach to first-passage-time
problems for Gauss-Markov processes. Adv. Appl. Probab. 2001, 33, 453–482. [CrossRef]
16. Abundo, M. An overview on inverse first-passage-time problems for one-dimensional diffusion processes.
Lect. Notes Semin. Interdiscip. Matematica 2015, 12, 1–44. Available online: http://dimie.unibas.it/site/home/
info/documento3012448.html (accessed on 28 May 2018)

9
Mathematics 2018, 6, 91

17. Di Crescenzo, A.; Giorno, V.; Nobile, A.G.; Ricciardi, L.M. On the M/M/1 queue with catastrophes and its
continuous approximation. Queueing Syst. 2003, 43, 329–347. [CrossRef]
18. Ikeda, N.; Watanabe, S. Stochastic Differential Equations and Diffusion Processes; North-Holland Publishing
Company: Amsterdam, The Netherlands, 1981.

c 2018 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

10
 mathematics
Article
Stability Analysis of Cohen–Grossberg Neural
Networks with Random Impulses
Ravi Agarwal 1,2 , Snezhana Hristova 3, *, Donal O’Regan 4 and Peter Kopanov 3
1 Department of Mathematics, Texas A&M University-Kingsville, Kingsville, TX 78363, USA;
[email protected]
2 Florida Institute of Technology, Melbourne, FL 32901, USA
3 Faculty of Mathematics, Plovdiv University, Tzar Asen 24, 4000 Plovdiv, Bulgaria; [email protected]
4 School of Mathematics, Statistics and Applied Mathematics, National University of Ireland,
H91 CF50 Galway, Ireland; [email protected]
* Correspondence: [email protected]

Received: 27 July 2018; Accepted: 17 August 2018; Published: 21 August 2018

Abstract: The Cohen and Grossberg neural networks model is studied in the case when the neurons
are subject to a certain impulsive state displacement at random exponentially-distributed moments.
These types of impulses significantly change the behavior of the solutions from a deterministic one
to a stochastic process. We examine the stability of the equilibrium of the model. Some sufficient
conditions for the mean-square exponential stability and mean exponential stability of the equilibrium
of general neural networks are obtained in the case of the time-varying potential (or voltage) of the
cells, with time-dependent amplification functions and behaved functions, as well as time-varying
strengths of connectivity between cells and variable external bias or input from outside the network to
the units. These sufficient conditions are explicitly expressed in terms of the parameters of the system,
and hence, they are easily verifiable. The theory relies on a modification of the direct Lyapunov
method. We illustrate our theory on a particular nonlinear neural network.

Keywords: Cohen and Grossberg neural networks; random impulses; mean square stability

1. Introduction
Artificial neural networks are important technical tools for solving a variety of problems in
various scientific disciplines. Cohen and Grossberg [1] introduced and studied in 1983 a new model
of neural networks. This model was extensively studied and applied in many different fields such
as associative memory, signal processing and optimization problems. Several authors generalized
this model [2] by including delays [3,4], impulses at fixed points [5,6] and discontinuous activation
functions [7]. Furthermore, a stochastic generalization of this model was studied in [8]. The included
impulses model the presence of the noise in artificial neural networks. Note that in some cases in the
artificial neural network, the chaos improves the noise (see, for example, [9]).
To the best of our knowledge, there is only one published paper studying neural networks
with impulses at random times [10]. However, in [10], random variables are incorrectly mixed with
deterministic variables; for example I[ξ k ,ξ k+1 ) (t) for the random variables ξ k , ξ k+1 is not a deterministic
index function (it is a stochastic process), and it has an expected value labeled by E, which has to be
taken into account on page 13 of [10]; in addition, in [10], one has to be careful since the expected
value of a product of random variables is equal to the product of expected values only for independent
random variables. We define the generalization of Cohen and Grossberg neural network with impulses
at random times, briefly giving an explanation of the solutions being stochastic processes, and we
study stability properties. Note that a brief overview of randomness in neural networks and some
methods for their investigations are given in [11] where the models are stochastic ones. Impulsive

Mathematics 2018, 6, 144; doi:10.3390/math6090144 11 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 144

perturbation is a common phenomenon in real-world systems, so it is also important to consider

impulsive systems. Note that the stability of deterministic models with impulses for neural networks
was studied in [12–18]. However, the occurrence of impulses at random times needs to be considered
in real-world systems. The stability problem for the differential equation with impulses at random
times was studied in [19–21]. In this paper, we study the general case of the time-varying potential
(or voltage) of the cells, with the time-dependent amplification functions and behaved functions, as
well as time-varying strengths of connectivity between cells and variable external bias or input from
outside the network to the units. The study is based on an application of the Lyapunov method. Using
Lyapunov functions, some stability sufficient criteria are provided and illustrated with examples.

2. System Description
We consider the model proposed by Cohen and Grossberg [1] in the case when the neurons are
subject to a certain impulsive state displacement at random moments.
Let T0 ≥ 0 be a fixed point and the probability space (Ω, F , P) be given. Let a sequence of
independent exponentially-distributed random variables {τk }∞ k=1 with the same parameter λ > 0
defined on the sample space Ω be given. Define the sequence of random variables {ξ k }∞
k =0 by:

k
ξ k = T0 + ∑ τi , k = 0, 1, 2, . . . . (1)
i =1

The random variable τk measures the waiting time of the k-th impulse after the (k − 1)-th impulse
occurs, and the random variable ξ k denotes the length of time until k impulses occur for t ≥ T0 .

(λt)k−1
Remark 1. The random variable Ξ = ∑ik=1 τi is Erlang distributed, and it has a pdf f Ξ (t) = λe−λt ( k −1) !
and
−1 (λt) j
a cdf F (t) = P(Ξ < t) = 1 − e−λt ∑kj= 0 j! .

Consider the general model of the Cohen–Grossberg neural networks with impulses occurring at
random times (RINN):
 n 
xi (t) = − ai ( xi (t)) b( xi (t)) − ∑ cij (t) f j (x j (t)) + Ii (t)
j =1

for t ≥ T0 , ξ k < t < ξ k+1 , k = 0, 1, . . . , i = 1, 2, . . . n, (2)

xi (ξ k + 0) = Φk,i ( xi (ξ k − 0)) for k = 1, 2, . . . ,
xi ( T0 ) = xi0 ,

where n corresponds to the number of units in a neural network; xi (t) denotes the potential (or voltage)
of cell i at time t, x (t) = ( x1 (t), x2 (t), . . . , xn (t)) ∈ Rn , f j ( x j (t)) denotes the activation functions
of the neurons at time t and represents the response of the j-th neuron to its membrane potential
and f ( x ) = ( f 1 ( x1 ), f 2 ( x2 ), . . . , f n ( xn )). Now, ai (.) > 0 represents an amplification function; bi (.)
represents an appropriately behaved function; the n × n connection matrix C (t) = (cij (t)) denotes the
strengths of connectivity between cells at time t; and if the output from neuron j excites (resp., inhibits)
neuron i, then cij (t) ≥ 0 (resp., cij (t) ≤ 0), and the functions Ii (t), I (t) = ( I1 (t), I2 (t), . . . , In (t)) ∈ Rn
correspond to the external bias or input from outside the network to the unit i at time t.
We list some assumptions, which will be used in the main results:
(H1) For all i = 1, 2, . . . , n, the functions ai ∈ C (R, (0, ∞)), and there exist constants Ai , Bi > 0
such that 0 < Ai ≤ ai (u) ≤ Bi for u ∈ R.
(H2) There exist positive numbers Mi,j , i, j = 1, 2, . . . , n such that |ci,j (t)| ≤ Mi,j for t ≥ 0.

Remark 2. In the case when the strengths of connectivity between cells are constants, then Assumption (H2) is
satisfied.

12
Mathematics 2018, 6, 144

For the activation functions, we assume:

(H3) The neuron activation functions are Lipschitz, i.e., there exist positive numbers Li , i =
1, 2, . . . , n, such that | f i (u) − f i (v)| ≤ Li |u − v| for u, v ∈ R.

Remark 3. Note that the activation functions satisfying Condition (H3) are more general than the usual sigmoid
activation functions.

2.1. Description of the Solutions of Model (2)

Consider the sequence of points {tk }∞ k=1 where the point tk is an arbitrary value of the
corresponding random variable τk , k = 1, 2, . . . . Define the increasing sequence of points { Tk }∞
k=1 by:

k
Tk = T0 + ∑ tk . (3)
i =1

Note that Tk are values of the random variables ξ k , k = 1, 2, . . . .

Consider the corresponding RINN (2) initial value problem for the system of differential equations
with fixed points of impulses { Tk }∞k=1 (INN):

 n 
xi (t) = − ai ( xi (t)) b( xi (t)) − ∑ cij (t) f j (x j (t)) + Ii (t)
j =1

for t ≥ T0 , t = Tk , k = 0, 1, . . . , i = 1, 2, . . . n, (4)
xi ( Tk + 0) = Φk,i ( xi ( Tk − 0)) for k = 1, 2, . . . ,
xi ( T0 ) = xi0 .

The solution of the differential equation with fixed moments of impulses (4) depends not only on
the initial point ( T0 , x0 ), but on the moments of impulses Tk , k = 1, 2, . . . , i.e., the solution depends on
the chosen arbitrary values tk of the random variables τk , k = 1, 2, . . . . We denote the solution of the
initial value problem (4) by x (t; T0 , x0 , { Tk }). We will assume that:

x ( Tk ; T0 , x0 , { Tk }) = lim x (t; T0 , x0 , { Tk }) for any k = 1, 2, . . . . (5)

t→ Tk −0

Remark 4. Note that the limit (5) is well defined since Tk , k = 1, 2 . . . , are points from R. This is different than
limt→ξ k −0 x (t) because ξ k is a random variable (see its incorrect use by the authors in [10]).

The set of all solutions x (t; T0 , x0 , { Tk }) of the initial value problem for the impulsive fractional
differential Equation (4) for any values tk of the random variables τk , k = 1, 2, . . . generates a stochastic
process with state space Rn . We denote it by x (t; T0 , x0 , {τk }), and we will say that it is a solution of
RINN (2).

Remark 5. Note that x(t; T0 , x0 , {Tk }) is a deterministic function, but x(t; T0 , x0 , {τk }) is a stochastic process.

Definition 1. For any given values tk of the random variables τk , k = 1, 2, 3, . . . , respectively, the solution
x (t; T0 , x0 , { Tk }) of the corresponding initial value problem (IVP) for the INN (4) is called a sample path
solution of the IVP for RINN (2).

Definition 2. A stochastic process x (t; T0 , x0 , {τk }) with an uncountable state space Rn is said to be a solution
of the IVP for the system of RINN (2) if for any values tk of the random variables τk , k = 1, 2, . . . , the
corresponding function x (t; T0 , x0 , { Tk }) is a sample path solution of the IVP for RINN (2).

13
Mathematics 2018, 6, 144

2.2. Equilibrium of Model (2)

We define an equilibrium of the model (2) assuming Condition (H1) is satisfied:

Definition 3. A vector x∗ ∈ Rn , x∗ = (x1∗ , x2∗ , . . . , xn∗ ) is an equilibrium point of RINN (2), if the equalities:

n
0 = b( xi∗ ) − ∑ cij (t) f j (xi∗ ) + Ii (t) for i = 1, 2, . . . , n (6)
j =1

and
xi∗ = Φk,i ( xi∗ ) for t ≥ 0, k = 1, 2, . . . , i = 1, 2, . . . , n (7)

hold.

We assume the following:

(H4) Let RINN (2) have an equilibrium vector x ∗ ∈ Rn .

If Assumption (H4) is satisfied, then we can shift the equilibrium point x ∗ of System (2) to the
origin. The transformation y(t) = x (t) − x ∗ is used to put System (2) in the following form:
 n 
yi (t) = − pi (yi (t)) q(yi (t)) − ∑ cij (t) Fj (y j (t))
j =1

for t ≥ T0 , ξ k < t < ξ k+1 , k = 0, 1, . . . , i = 1, 2, . . . n, (8)

yi (ξ k + 0) = φk,i (y(ξ k − 0)) for k = 1, 2, . . . ,
yi ( T0 ) = y0i ,

where pi (u) = ai (u + xi∗ ),qi (u) = bi (u + xi∗ ) − bi ( xi∗ ), Fj (u) = f j (u + x ∗j ) − f j ( x ∗j ), j = 1, 2, . . . , n and

φk,i (u) = Φk,i (u + xi∗ ) − Φk,i ( xi∗ ), i = 1, 2, . . . , n, k = 1, 2, . . . , y0i = xi0 − xi∗ .

Remark 6. If Assumption (H3) is fulfilled, then the function F in RINN (8) satisfies | Fj (u)| ≤ L j |u|, j =
1, 2, . . . , n, for u ∈ R.

Note that if the point x ∗ ∈ Rn is an equilibrium of RINN (2), then the point y∗ = 0 is an
equilibrium of RINN (8). This allows us to study the stability properties of the zero equilibrium of
RINN (8).

3. Some Stability Results for Differential Equations with Impulses at Random Times
Consider the general type of initial value problem (IVP) for a system of nonlinear random
impulsive differential equations (RIDE):

x  (t) = g(t, x (t)) for t ≥ T0 , ξ k < t < ξ k+1 ,

x (ξ k + 0) = Ψk ( x (ξ k − 0)) for k = 1, 2, . . . , (9)
0
x ( T0 ) = x ;

with x0 ∈ Rn , random variables ξ k , k = 1, 2, . . . are defined by (1), g ∈ C ([ T0 , ∞) × Rn , Rn ) and

Ψ k : Rn → Rn .

Definition 4. Let p > 0. Then, the trivial solution (x0 = 0) of RIDE (9) is said to be p-moment
exponentially stable if for any initial point ( T0 , y0 ) ∈ R+ × Rn , there exist constants α, μ > 0 such that
E[||y(t; T0 , y0 , {τk )})|| p ] < α||y0 || p e−μ(t−T0 ) for all t > T0 , where y(t; T0 , x0 , {τk )} is the solution of the IVP
for RIDE (9).

14
Mathematics 2018, 6, 144

Definition 5. Let p > 0. Then, the equilibrium x ∗ of RINN (2) is said to be p-moment exponentially stable
if for any initial point ( T0 , x0 ) ∈ R+ × Rn , there exist constants α, μ > 0 such that E[|| x (t; T0 , x0 , {τk )}) −
x ∗ || p ] < α|| x0 − x ∗ || p e−μ(t−T0 ) for all t > T0 , where x (t; T0 , x0 , {τk )} is the solution of the IVP for RINN (2).

Remark 7. We note that the two-moment exponential stability for stochastic equations is known as the mean
square exponential stability, and in the case of p = 1, it is called mean exponential stability.

Note that the p-moment exponential stability of RIDE (9) was studied in [20] by an application of
Lyapunov functions from the class Λ( J, Δ), J ⊂ R+ , Δ ⊂ Rn , 0 ∈ Δ with:

Λ( J, Δ) = {V (t, x ) ∈ C ( J × Δ, R+ ) : V (t, 0) ≡ 0,
V (t, x ) is locally Lipschitzian with respect to x}.

We will use the Dini derivative of the Lyapunov function V (t, x ) ∈ Λ( J, Δ) given by:
 
1
(9) D+ V ( t, x ) = lim sup V (t, x ) − V (t − h, x − hg(t, x ))
h →0+ h (10)
for t ∈ J, x ∈ Δ.

Now, we will give a sufficient condition result:

Theorem 1 ([20]). Let the following conditions be satisfied:

1. For t ≥ 0 : g(t, 0) ≡ 0 and Ψk (0) = 0, k = 1, 2, . . . and for any initial values ( T0 , x0 ), the corresponding
IVP for the ordinary differential equation x  (t) = g(t, x (t)) has a unique solution.
2. The function V ∈ Λ([ T0 , ∞), Rn ), and there exist positive constants a, b such that:

(i) a|| x || p ≤ V (t, x ) ≤ b|| x || p for t ≥ T0 , x ∈ Rn ;

(ii) there exists a function m ∈ C (R+ , R+ ) : inft≥0 m(t) = L ≥ 0, and the inequality:

(9) D+ V ( t, x ) ≤ −m(t)V (t, x ), for t ≥ 0, x ∈ Rn

holds;
(iii) for any k = 1, 2, . . . , there exist constants wk : 0 ≤ wk < 1 + L
λ for t ≥ 0 such that:

V (t, Ik (t, x )) ≤ wk V (t, x ) for t ≥ 0, x ∈ Rn . (11)

Then, the trivial solution of RIDE (9) is p-moment exponentially stable.

4. Stability Analysis of Neural Networks with Random Impulses

We will introduce the following assumptions:
 (H5) For i = 1, 2,. . . , n, the functions bi ∈ C (R, R), and there exist constants β i > 0 such that
u bi (u + xi∗ ) − bi ( xi∗ ) ≥ β i u2 for any u ∈ R where x ∗ ∈ Rn , x ∗ = ( x1∗ , x2∗ , . . . , xn∗ ), is the equilibrium
from Condition (H4).

Remark 8. If Condition (H5) is satisfied, then the inequality uq(u) ≥ β i u2 , u ∈ R holds for RINN (8).

(H6) The inequality:

 n n 
ν =2 min Ai β i − max Bi max ∑ Mij L j + ( ∑ max Mij L j ) > 0 (12)
i =1,n i =1,n i =1,n j=1 i =1 j=1,n

15
Mathematics 2018, 6, 144

holds.
ν
(H7) For any k = 1, 2, . . . , there exists positive number Kk < 1 + λ such that the inequalities:

n  2 n
∑ Φk,i ( xi ) − Φk,i ( xi∗ ) ≤ Kk ∑ (xi − xi∗ )2 , xi ∈ R, i = 1, 2, . . . , n,
i =1 i =1

hold where x ∗ ∈ Rn , x ∗ = ( x1∗ , x2∗ , . . . , xn∗ ), is the equilibrium from Condition (H4).

Remark 9. If Assumption (H7) is fulfilled, then the impulsive functions φk , k = 1, 2, . . . in RINN (8) satisfy
2 (u ) ≤ K
the inequalities ∑in=1 φk,i n 2
i k ∑ i =1 u i .

Theorem 2. Let Assumptions (H1)–(H7) be satisfied. Then, the equilibrium point x ∗ of RINN (2) is mean
square exponentially stable.

Proof. Consider the quadratic Lyapunov function V (t, x ) = x T x, x ∈ Rn . From Remarks 6, 8 and
inequality 2|uv| ≤ u2 + v2 , we get:
n   n 
(8) D+ V ( t, y ) ≤ 2 ∑ yi − pi (yi ) q(yi ) − ∑ cij (t) Fj (y j )
i =1 j =1
n n n
= −2 ∑ yi pi (yi )q(yi ) + 2 ∑ yi pi (yi ) ∑ cij (t) Fj (y j )
i =1 i =1 j =1
n n n
≤ −2 ∑ Ai β i y2i + 2 ∑ |yi | Bi ∑ Mij L j |y j |
i =1 i =1 j =1
n n n
≤ −2 ∑ Ai β i y2i + ∑ Bi ∑ Mij L j (y2i + y2j )
i =1 i =1 j =1
(13)
n n n n n
≤ −2 ∑ Ai β i y2i + ∑ Bi y2i ∑ Mij L j + ∑ Bi ∑ Mij L j y2j
i =1 i =1 j =1 i =1 j =1
n
≤ −2 min Ai β i
i =1,n
∑ y2i
i =1
 n n  n
+ max Bi max ∑ Mij L j + ( ∑ max Mij L j ) ∑ y2i
i =1,n i =1,n j=1 i =1 j=1,n i =1
n
= −ν ∑ y2i .
i =1

where the positive constant ν is defined by (12). Therefore, Condition 2(ii) of Theorem 1 is satisfied.
Furthermore, from (H7), it follows that Condition 2(iii) of Theorem 1 is satisfied.
From Theorem 1, the zero solution of the system (9) is mean square exponentially stable,
and therefore, the equilibrium point x ∗ of RINN (2) is mean square exponentially stable.

Example 1. Let n = 3, t0 = 0.1, and the random variables τk , k = 1, 2, . . . are exponentially distributed with
λ = 1. Consider the following special case of RINN (2):

 3 
xi (t) = − ai ( xi (t)) 2xi (t) + ∑ cij (t) f j (x j (t)) − π
j =1

for t ≥ 0 ξ k < t < ξ k+1 , i = 1, 2, 3, (14)

xi (ξ k + 0) = Φk,i ( xi (ξ k − 0)) for k = 1, 2, . . . ,
xi (0.1) = xi0 ,

16
Mathematics 2018, 6, 144

|u|
with ai (u) = 2 + 1 + |u| ∈ [2, 3), i = 1, 2, 3, f i (u) = αi cos(u), α1 = 0.1, α2 = 0.01, α2 = 2, Φk,i (u) =
u sin k + (1 − sin k)0.5π, and C = cij (t) is given by:
⎛ ⎞
−0.1 sin t 0.4 0.3
⎜ 2 ⎟
C (t) = ⎝ − 5t2t +1 0.3 t
5t+1 ⎠. (15)
t
10t+1 −0.2 cos t −0.1 sin t

The point x ∗ = (0.5π, 0.5π, 0.5π ) is the equilibrium point of RINN (14), i.e., Condition (H4) is
satisfied. Now, Assumption (H1) is satisfied with Ai = 2, Bi = 3, i = 1, 2, 3. In addition, Assumption
(H5) is satisfied with β i = 2, i = 1, 2, 3.
Furthermore, |cij | ≤ Mij , i, j = 1, 2, 3, t ≥ 0 where M = { Mij }, is given by:
⎛ ⎞
0.1 0.4 0.3
⎜ ⎟
M = ⎝0.2 0.3 0.2⎠ . (16)
0.1 0.2 0.1

Therefore, Assumption (H2) is satisfied. Note that Assumption (H3) is satisfied with Lipschitz
constants L1 = 0.1, L2 = 0.01, L3 = 2.
Then, the constant ν defined by (12) is ν = 8 − 3(1.814) = 2.558 > 0. Next, Assumption (H7) is
fulfilled with Kk = 1 because:

3  2 3  2
∑ Φk,i ( xi ) − Φk,i ( xi∗ ) = ∑ xi sin k + (1 − sin k )0.5π − 0.5π
i =1 i =1
(17)
3  2 3  2
= ∑ ( xi − 0.5π ) sin k ≤ ∑ xi − 0.5π , k = 1, 2, . . . .
i =1 i =1

Therefore, according to Theorem 1, the equilibrium of RINN (14) is mean square exponentially stable.
Consider the system (14) without any kind of impulses. The equilibrium x ∗ = (0.5π, 0.5π, 0.5π )
is asymptotically stable (see Figures 1 and 2). Therefore, an appropriate perturbation of the neural
networks by impulses at random times can keep the stability properties of the equilibrium.

tx
2.0

1.8

x1 HtL
x2 HtL
1.6

1.4 x3 HtL

1.2

t
2 4 6 8

Figure 1. Example 1. Graph of the solution of the system ODE corresponding to (14) with x10 = 1, x20 =
2, x30 = 1.4.

17
Mathematics 2018, 6, 144

tx

1.5

x1 HtL
1.0

x2 HtL
0.5
x3 HtL

t
1 2 3 4

Figure 2. Example 1. Graph of the solution of the system ODE corresponding to (14) with x10 =
−0.1, x20 = 0.2, x30 = −0.4.

Remark 10. Note that Condition (H7) is weaker than Condition (3.6) in Theorem 3.2 [16], and as a special
case of Theorem 2, we obtain weaker conditions for exponential stability of the Cohen and Grossberg model
without any type of impulses. For example, if we consider (14) according to Condition (3.6) [16], the inequality
δ = 2|| M ||2 34 = 1.0374 < 1 is not satisfied, and Theorem 3.2 [16] does not give us any result about stability
(compare with Example 1).

Now, consider the following assumption:

(H8) The inequality:
n
ν = min γi − ∑ max Mij > 0 (18)
i =1,n i =1 j=1,n

holds.

Theorem 3. Let Assumptions (H1)–(H5), (H7) and (H8) be satisfied. Then, the equilibrium point x ∗ of RINN
(2) is mean exponentially stable.
 ui sign(s)
Proof. For any u ∈ Rn , we define V (u) = ∑in=1 0 ai ( s )
ds. Then:

n  ui n
sign(s) 1
V (u) ≤ ∑ Ai
ds = ∑
A i
|u | ≤ A||u||
i =1 0 i =1 i

and:
n  ui n
sign(s) 1
V (u) ≥ ∑ Bi
ds = ∑ |ui | ≥ B||u||
B
i =1 0 i =1 i
1 1
where A = maxi=1,n Ai , B = mini=1,n Bi .

18
Mathematics 2018, 6, 144

Then, for t ≥ 0 and y ∈ Rn according to Remarks 6 and 8, we obtain:

n  n 
( 8 ) D+ V ( y ) ≤ ∑ −sgn(yi ) q ( yi ) − ∑ cij (t) Fj (y j )
i =1 j =1
n  n  n  n 
≤ ∑ − β i |yi | + ∑ Mij | Fj (y j )| ≤ ∑ − β i |yi | + ∑ Mij |y j |
i =1 j =1 i =1 j =1
(19)
n n n n  n  n
= − ∑ β i | yi | + ∑ ∑ Mij |y j | ≤ − min β i ∑ | yi | + ∑ max Mij ∑ |y j |
i =1 i =1 j =1 i =1,n i =1 i =1 j=1,n j =1
n
ν
≤ − ν ∑ | y i | ≤ − V ( u ).
i =1
B

Furthermore, from (H7) and Remark 9, it follows that Condition 2(iii) of Theorem 1 is satisfied.
From Theorem 1, we have that Theorem 3 is true.

Example 2. Let n = 3, t0 = 0.1, and the random variables τk , k = 1, 2, . . . are exponentially distributed with
λ = 1. Consider the following special case of RINN (2):

 3 
xi (t) = − ai ( xi (t)) 2xi (t) + ∑ cij (t) f j (x j (t)) − 1
j =1

for t ≥ 0 ξ k < t < ξ k+1 , i = 1, 2, 3 (20)

xi (ξ k + 0) = Φk,i ( xi (ξ k − 0)) for k = 1, 2, . . .
xi (0.1) = xi0 ,

|u|
with ai (u) = 2 + 1 + |u| ∈ [2, 3), i = 1, 2, 3, f i (u) = log( 1 −u u ), Φk,i (u) = u sin k + (1 − sin k)0.5,
and C = cij (t) is given by (15).

The point x ∗ = (0.5, 0.5, 0.5) is the equilibrium point of RINN (20), i.e., Condition (H4) is satisfied.
Now, Assumption (H5) is satisfied with β i = 2, i = 1, 2, 3.
Furthermore, |cij | ≤ Mij , i, j = 1, 2, 3, t ≥ 0 where M = { Mij }, is given by (16). Therefore,
Assumption (H2) is satisfied. Then, the inequality mini=1,n β i = 2 > ∑3i=1 max j=1,3 Mij = 0.4 + 0.3 +
0.2 = 0.9 holds.
According to Theorem 3, the equilibrium of (20) is mean exponentially stable.
Consider the system (20) without any kind of impulses. The equilibrium x ∗ = (0.5, 0.5, 0.5)
is asymptotically stable (see Figures 3 and 4). Therefore, an appropriate perturbation of the neural
networks by impulses at random times can keep the stability properties of the equilibrium.

19
Mathematics 2018, 6, 144

tx
0.8

0.7

x1 HtL
0.6

x2 HtL
0.5

0.4 x3 HtL
0.3

0.2

t
1 2 3 4 5

Figure 3. Example 2. Graph of the solution of the system ODE corresponding to (20) with x10 =
0.55, x20 = 0.8, x30 = 0.1.

tx

0.6

0.5

x1 HtL
x2 HtL
0.4

0.3 x3 HtL

0.2

t
1 2 3 4 5

Figure 4. Example 2. Graph of the solution of the system ODE corresponding to (20) with x10 =
0.4, x20 = 0.3, x30 = 0.1.

5. Conclusions
In this paper, we study stability properties of the equilibrium point of a generalization of the
Cohen–Grossberg model of neural networks in the case when:

- the potential (or voltage) of any cell is perturbed instantaneously at random moments, i.e.,
the neural network is modeled by a deterministic differential equation with impulses at random
times. This presence of randomness in the differential equation totally changes the behavior of
the solutions (they are not deterministic functions, but stochastic processes).
- the random moments of the impulsive state displacements of neurons are exponentially distributed.

20
Mathematics 2018, 6, 144

- the connection matrix C = (cij ) is not a constant matrix which is usually the case in the literature
(it is a matrix depending on time since the strengths of connectivity between cells could be
changed in time).
- the external bias or input from outside the network to any unit is not a constant (it is variable
in time).
- sufficient conditions for mean-square exponential stability and for mean exponential stability of
the equilibrium are obtained.

Author Contributions: All authors contributed equally to the writing of this paper. All four authors read and
approved the final manuscript.
Funding: The research was partially supported by Fund MU17-FMI-007, University of Plovdiv “Paisii
Hilendarski”.
Conflicts of Interest: The authors declare that they have no competing interests.

References
1. Cohen, M.; Grossberg, S. Stability and global pattern formation and memory storage by competitive neural
networks . IEEE Trans. Syst. Man Cyber 1983, 13, 815–826. [CrossRef]
2. Guo, S.; Huang, L. Stability analysis of Cohen-Grossberg neural networks. IEEE Trans. Neural Netw. 2006, 17,
106–117. [CrossRef] [PubMed]
3. Bai, C. Stability analysis of Cohen–Grossberg BAM neural networks with delays and impulses. Chaos Solitons
Fractals 2008, 35, 263–267. [CrossRef]
4. Cao, J.; Liang, J. Boundedness and stability for Cohen–Grossberg neural network with time-varying delays.
J. Math. Anal. Appl. 2004, 296, 665–685. [CrossRef]
5. Aouiti, Ch.; Dridi, F. New results on impulsive Cohen–Grossberg neural networks. Neural Process. Lett. 2018,
48, 1–25. [CrossRef]
6. Liu, M.; Jiang, H.; Hu, C. Exponential Stability of Cohen-Grossberg Neural Networks with Impulse Time
Window. Discret. Dyn. Nat. Soc. 2016, 2016, 2762960. [CrossRef]
7. Meng, Y.; Huang, L.; Guo, Z.; Hu, Q. Stability analysis of Cohen–Grossberg neural networks with
discontinuous neuron activations. Appl. Math. Model. 2010, 34, 358–365. [CrossRef]
8. Huang, C.; Huang, L.; He, Y. Mean Square Exponential Stability of Stochastic Cohen-Grossberg Neural
Networks with Unbounded Distributed Delays. Discret. Dyn. Nat. Soc. 2010, 2010, 513218. [CrossRef]
9. Bucolo, M.; Caponetto, R.; Fortuna, L.; Frasca, M.; Rizzo, A. Does chaos work better than noise? IEEE Circuits
Syst. Mag. 2002, 2, 4–19 . [CrossRef]
10. Vinodkumar, A.; Rakkiyappan, R. Exponential stability results for fixed and random type impulsive Hopfield
neural networks. Int. J. Comput. Sci. Math. 2016, 7, 1–19. [CrossRef]
11. Scardapane, S.; Wang, D. Randomness in neural networks: an overview. WIREs Data Min. Knowl. Discov.
2017, 7, 1–18. [CrossRef]
12. Gopalsamy, K. Stability of artificial neural networks with impulses. Appl. Math. Comput. 2004, 154, 783–813.
[CrossRef]
13. Rakkiyappan, R.; Balasubramaiam, P.; Cao, J. Global exponential stability of neutral-type impulsive neural
networks. Nonlinear Anal. Real World Appl. 2010, 11, 122–130. [CrossRef]
14. Song, X.; Zhao, P.; Xing, Z.; Peng, J. Global asymptotic stability of CNNs with impulses and
multi-proportional delays. Math. Methods Appl. Sci. 2016, 39, 722–733.
15. Wu, Z.; Li, C. Exponential stability analysis of delayed neural networks with impulsive time window.
In Proceedings of the 2017 Ninth International Conference on Advanced Computational Intelligence (ICACI),
Doha, Qatar, 4–6 February 2017; pp. 37–42.
16. Wang, L.; Zou, X. Exponential stability of Cohen-Grossberg neural networks. Neural Netw. 2002, 15, 415–422.
[CrossRef]
17. Yang, Z.; Xu, D. Stability analysis of delay neural networks with impulsive effects. IEEE Trans. Circuits Syst.
Express Briefs 2015, 52, 517–521. [CrossRef]
18. Zhou, Q. Global exponential stability of BAM neural networks with distributed delays and impulses.
Nonlinear Anal. Real World Appl. 2009, 10, 144–153. [CrossRef]

21
Mathematics 2018, 6, 144

19. Agarwal, R.; Hristova, S.; O’Regan, D.; Kopanov, P. P-moment exponential stability of differential equations
with random impulses and the Erlang distribution. Mem. Differ. Equ. Math. Phys. 2017, 70, 99–106. [CrossRef]
20. Agarwal, R.; Hristova, S.; O’Regan, D. Exponential stability for differential equations with random impulses
at random times. Adv. Differ. Equ. 2013, 372, 12. [CrossRef]
21. Agarwal, R.; Hristova, S.; O’Regan, D.; Kopanov, P. Impulsive differential equations with Gamma distributed
moments of impulses and p-moment exponential stability. Acta Math. Sci. 2017, 37, 985–997. [CrossRef]

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

22
 mathematics
Article
On Short-Term Loan Interest Rate Models: A First
Passage Time Approach
Giuseppina Albano 1, * and Virginia Giorno 2
1 Dipartimento di Scienze Economiche e Statistiche, University of Salerno, 84084 Fisciano, SA, Italy
2 Dipartimento di Informatica, University of Salerno, 84084 Fisciano, SA, Italy; [email protected]
* Correspondence: [email protected]; Tel.: +39-089-962-645

Received: 2 February 2018; Accepted: 25 April 2018; Published: 3 May 2018

Abstract: In this paper, we consider a stochastic diffusion process able to model the interest rate
evolving with respect to time and propose a first passage time (FPT) approach through a boundary,
defined as the “alert threshold”, in order to evaluate the risk of a proposed loan. Above this alert
threshold, the rate is considered at the risk of usury, so new monetary policies have been adopted.
Moreover, the mean FPT can be used as an indicator of the “goodness” of a loan; i.e., when an
applicant is to choose between two loan offers, s/he will choose the one with a higher mean exit time
from the alert boundary. An application to real data is considered by analyzing the Italian average
effect global rate by means of two widely used models in finance, the Ornstein-Uhlenbeck (Vasicek)
and Feller (Cox-Ingersoll-Ross) models.

Keywords: loan interest rate regulation; diffusion model; first passage time (FPT)

1. Introduction
In recent decades, increasing attention has been paid to the study of the dynamics underlying the
interest rates. The intrinsically stochastic nature of the interest rates has suggested the formulation
of various models often based on stochastic differential equations (SDEs) (see, for example, [1,2]
and references therein). More recently, further stochastic representations of non-usurious interest
rates have been provided in order to obtain information concerning costs of loans. Most of them
are simple and convenient time-homogeneous parametric models, attempting to capture certain
features of observed dynamic movements, such as heteroschedasticity, long-run equilibrium, and other
peculiarities (see, for example, [3–5]).
An interest rate is “usurious” if it is markedly above current market rates. France was the first
European country to introduce an anti-usury law in 1966. In Italy, the first law of this nature (Law
No. 108) was introduced in 1996. An inventory of interest rate restrictions against usury in the EU
Member States was achieved at the end of 2010. In particular, the EU authorities’ attention focused
on the interest rate restrictions established on precise legal rules restricting credit price, both directly
by fixed thresholds as well as indirectly by intervening on the calculation of compound interest
(Directorate-General of the European Commission, 2011).
Since May 2011, the Italian law has governed interest rates in loans with new regulations, fixing a
threshold above which interest rates applied in loans are considered usurious. The threshold rate is
based on the actual global average rate of interest (TEGM) that is quarterly determined by the Italian
Ministry of Economy and Finance (Ministero dell’Economia e delle Finanze), and it is a function of
various types of homogeneous transactions. Specifically, the threshold rate is calculated as 125% of the
reference TEGM plus 4%. Therefore,

Threshold rate = 1.25 TEGM + 0.04.

Mathematics 2018, 6, 70; doi:10.3390/math6050070 23 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 70

Moreover, the difference between the TEGM and the usury threshold cannot exceed 8%, so the
maximum value admissible for TEGM cannot exceed 16%.
Note that the penal code (art. 644, comma 4, c.p.) establishes that the scheduling of the usury
interest rate takes into account errands, wages, and costs, but not taxes related to the loan supply,
but, to compute the TEGM, the Bank of Italy does not consider these items. Therefore, this difference
between the principle stated by the legislature and the instructions of the Bank of Italy decreases
both the average rates and the threshold rates. Therefore, another boundary that is lower than that
established by the Bank of Italy should be introduced. This case has also been extended to other
European countries.
The basic idea of the present work is to investigate the (random) time in which an interest rate
reaches an “alert boundary”, that is near the admitted limit of 0.16. To do this, we start with two
classical models in the literature: Vasicek and Cox-Ingersoll-Ross (CIR) ([6,7]) since they provide good
characterization of the short-term real rate process. In particular, the CIR model is able to capture the
dependence of volatility on the level of interest rates ([8]).
We then investigated the first passage time (FPT) through a boundary generally depending on
time. This approach is useful in economy since it suggests the time in which the trend of a loan interest
rate can be considered at risk of usury, so it has to be modified from the owner of the loan service.
Moreover, the mean first exit time through the alert boundary could be adopted as an indicator of the
“goodness” of the loan, in the sense that an applicant choosing between two loan offers will choose the
one with a higher mean exit time from the alert boundary. For the FPT analysis, we consider a constant
boundary; clearly this kind of approach is applicable to other underlying models that are different
from the Vasicek and CIR models and to boundaries generally depending on time, which is the case of
time-dependent loan interest rate.
The layout of the paper is as follows. In Section 2, a brief review of diffusion models describing
the dynamics of the interest rate is discussed. The FPT problem through a time-dependent threshold
S(t) is analyzed. In Section 3, we consider data of the TEGM published by Bank of Italy. In particular,
we compare the Vasicek and CIR models in order to establish which model better fits our data.
Moreover, a Chow test shows the presence of structural breaks. In Section 4, the FPT problem through
a constant “alert boundary” is analyzed. Concluding remarks follow.

2. Mathematical Background
We denote by { X (t), t ≥ t0 } the stochastic process describing the dynamics of a loan interest
rate. We assume that X (t) is a time-homogeneous diffusion process defined in I = (r1 , r2 ) by the
following SDE:
dX (t) = A1 [ X (t)]dt + A2 [ X (t)] dW (t), X (t0 ) = x0 a.s., (1)

where A1 ( x ) and A2 ( x ) > 0 denote the drift and the infinitesimal variance of X (t) and W (t) is
a standard Wiener process. The instantaneous drift A1 ( x ) represents a force that keeps pulling
the process towards its long-term mean, whereas A2 ( x ) represents the amplitude of the random
fluctuations. Let   x 
A1 ( z ) 2
h( x ) = exp −2 dz , s( x ) = (2)
A2 ( z ) A2 ( x ) h ( x )
be the scale function and speed density of X (t), respectively. The transition probability density function
(pdf) of X (t), denoted by f ( x, t|y, τ ), is a solution of the Kolmogorov equation,

∂ f ( x, t|y, τ ) ∂ f ( x, t|y, τ ) A (y) ∂2 f ( x, t|y, τ )

+ A1 ( y ) + 2 =0
∂τ ∂y 2 ∂y2

and of the Fokker–Planck equation,

24
Mathematics 2018, 6, 70

∂ f ( x, t|y, τ ) ∂  ∂2  A2 ( x )
=− A ( x ) f ( x, t|y, τ ) + 2 f ( x, t|y, τ ) ,
∂t ∂x 1 ∂x 2
with the delta initial conditions:

lim f ( x, t|y, τ ) = lim f ( x, t|y, τ ) = δ( x − y).

t↓τ τ ↑t

The above conditions assure the uniqueness of the transition pdf only when the endpoints of
the diffusion interval are natural; otherwise, suitable boundary conditions may have to be imposed
(cf., for istance, [9]).
Further, if X (t) admits a steady-state behavior, then the steady-state pdf is

s( x )
W ( x ) ≡ lim f ( x, t| x0 , t0 ) =  ∞ .
t→∞
−∞ s ( z ) dz

Let
TX0 = inf {t : X (t) > S(t) | X (t0 ) = x0 }
t > t0

be the FPT variable of X (t) through a time-dependent boundary S(t) starting from x0 , and let
g[S(t), t| x0 , t0 ] = dP( Tx0 < t)/dt be its pdf. In the following, we assume that x0 < S(t0 ) since
in our context x0 represents the initial observed value of the interest rate. The FPT problem has
far-reaching implications (see, for instance, [10,11]).
As shown in [12,13], if S(t) is in C2 [t0 , ∞), g can be obtained as a solution of the following
second-kind Volterra integral equation:
 t
g[S(t), t|y, τ ] = −2Ψ[S(t), t|y, τ ] + 2 g[S(ϑ ), ϑ |y, τ ] Ψ[S(t), t|S(ϑ ), ϑ ] dϑ (3)
t0

where
1 ! 3 "
Ψ[S(t), t|y, τ ] = f ( x, t|y, τ ) S (t) − A1 [S(t)] + A2 [S(t)]
2 4
A [S(t)] ∂ f ( x, t|y, τ )
+ 2 .
2 ∂x x =S(t)

If A1 ( x ) and A2 ( x ) are known, i.e., if the process is fixed, some closed form solution of (3) can be
obtained for particular choices of the boundary S(t). Further results have been obtained in [14–16].
Alternatively, a numerical algorithm can be successfully used; for example, the R package fptdApprox
is also a useful instrument for the numerical evaluation of the FPT pdf (see [17,18]).
Further, if the FPT is a sure event and if S(t) = S is time-independent, the moments of the FPT
can be evaluated via a recursive Siegert-type formula (see, for instance, [9]):
 ∞  S  z
t n ( S | x0 ) = tn g(S, t| x0 ) dt = n dz h(z) s(u) tn−1 (S|u) du
0 x0 −∞
n = 1, 2, . . . (4)

where t0 (S| x0 ) = P( Tx0 < ∞) = 1 and h( x ) and s( x ) given in Equation (2).

3. Modeling the Italian Loans

In this section, we consider two stochastic processes widely used in the financial literature, the
Vasicek and CIR models (see [1,19]), for describing a historical series of Italian average rates on loans.
We use the Akaike information criterion (AIC) as an indicator of the goodness of fit of the two models.
Moreover, the presence of structural breaks is verified by means of a Chow test applied to the Euler

25
Mathematics 2018, 6, 70

discretization of the corresponding SDE. More precisely, the Chow test is sequentially applied for
each instant in order to evaluate whether the coefficients of the Euler discretization made on each
subinterval are equal to those including all observed time intervals.
TEGM values are quarterly settled and published by the Bank of Italy (see https://www.
bancaditalia.it) for different types of credit transactions. We refer to the TEGM values for a particular
credit transaction, “one-fifth of salary transfer”, in the period from 1 July 1997 to 31 March 2015
(data are quarterly observed, so the number of observations is 72). Moreover, two amount classes
are analyzed:

• Dataset A: up to 10 million lira (until 31 December 2001) and up to e 5000 (after 2002);
• Dataset B: above 10 million lire (until 31 December 2001) and above e 5000 (after 2002).

In Figure 1, Dataset A is shown on the left and Dataset B is on the right.







7(*0

7(*0






       

'DWH 'DWH

Figure 1. TEGM for one-fifth of salary transfer up to e 5000 (on the left) and above e 5000 (on the right).

We estimate the parameters for the Vasiceck and CIR models, maximizing the conditional
likelihood function. Specifically, we assume that the process X (t) is observed at n discrete time
instants t1 , . . . , tn with ti ≥ t0 and denote by x1 , . . . , xn the corresponding observations.
Let θ be the vector of the unknown parameters and let us assume P[ X (t1 ) = x1 ] = 1.
The likelihood function is
n
L( x1 , . . . , xn ; θ) = ∏ f ( x i , t i | x i −1 , t i −1 ).
i =2

3.1. The Vasiceck Model

The Vasiceck model describes the short rate’s dynamics. It can be used in the evaluation of interest
rate derivatives and is more suitable for credit markets. It is specified by the following SDE:

dX (t) = [θ1 − θ2 X (t)]dt + θ3 dW (t), (5)

where θ1 , θ2 , θ3 are positive constants. The model (5) with θ1 = 0 was originally proposed by Ornstein
and Uhlenbeck in 1930 in the physical context to describe the velocity of a particle moving in a fluid
under the influence of friction and it was then generalized by Vasicek in 1977 to model loan interest
rates. It is also used as a model and in physical and biological contexts (see, for instance, [20–23]).

26
Mathematics 2018, 6, 70

We note that, for θ2 > 0, the process X (t) is mean reverting oscillating around the equilibrium
point θ1 /θ2 . The process is defined in R and the boundaries ±∞ are natural. The transition pdf of X (t)
is given by
 
1 [ x − M(t| x0 , t0 )]2
f ( x, t| x0 , t0 ) =  exp − , (6)
2πV (t|t0 ) 2V (t|t0 )
where

θ1 # $ θ32 # $
M ( t | x0 , t0 ) = 1 − e − θ2 ( t − t0 ) + x 0 e − θ2 ( t − t0 ) , V ( t | t0 ) = 1 − e−2θ2 (t−t0 )
θ2 2θ2

represent the mean and the variance of X (t) with the condition that X (t0 ) = x0 , respectively.
Further, X (t) has the following steady-state density:
% 
2 &
s( x ) θ2 θ2 θ1
W (x) =  ∞ = exp − x− ,
−∞ s ( z ) dz πθ32 θ32 θ2

which describes a Gaussian distribution with mean θ1 /θ2 and variance θ32 /2 θ2 .
Let θ = (θ1 , θ2 , θ3 ) be the vector of the unknown parameters. The maximum likelihood estimate is
obtained as θ = arg maxθ log L( x1 , . . . , xn ; θ). Implementing this method, making use of the R package
sde (see [24,25]), the procedure produces the results shown in Table 1. In the last row of this table, the
AIC, i.e.,
AIC = 6 − 2 log L( x1 , . . . , xn ; 
θ),

is shown for the two datasets.

Table 1. ML estimates of Model (5) for Dataset A (on the left) and for Dataset B (on the right). The last
row shows the AIC.

Vasicek Model
Dataset A Dataset B
estimate standard error estimate standard error
θ1 0.9473455 0.62502358 1.9016919 0.41732799
θ2 0.0658379 0.03621181 0.1675858 0.03403613
θ3 1.0355084 0.08881145 0.5610075 0.04793382
AIC 207.8207 113.8432

For Datasets A and B, the Chow test applied to the Euler discretization of Model (5) shows a
structural break at time t = 42, corresponding to 1 January 2008 (p-value = 0.002726) for Dataset A
and at time t = 47 corresponding to 1 January 2009 (p-value = 0.006231) for Dataset B. In Table 2, the
ML estimates for Datasets A and B are shown considering separately the series before and after these
dates. Precisely, we consider for Dataset A the following sub-periods:

• first period: 1 July 1997–1 October 2007;

• second period: 1 January 2008–31 March 2015;

for Dataset B, the sub-intervals are as follows:

• first period: 1 July 1997–1 October 2008;

• second period: 1 January 2009–31 March 2015.

The existence of a structural break is quite clear just looking at the data in Figure 1, but the Chow
test permits us to establish the time at which the break verifies, and the AIC values confirm that the
estimations evaluated in the two periods work better then the estimates on the whole dataset.

27
Mathematics 2018, 6, 70

Table 2. ML estimates of Model (5) and the corresponding AIC for the periods indicated by Chow test
for Dataset A (on the top) and for Dataset B (on the bottom).

The Vasicek Model

Dataset A
First Period Second Period
1 July 1997–1 October 2007 1 January 2008–31 March 2015
estimate standard error estimate standard error
θ1 5.3865862 2.4412428 4.3464435 1.6780736
θ2 0.2862057 0.1245413 0.3411949 0.1264711
θ3 1.2012565 0.1489351 0.8262836 0.1179829
AIC 126.2129 67.89682
Dataset B
First Period Second Period
1 July 1997–1 October 2008 1 January 2009–31 March 2015
estimate standard error estimate standard error
θ1 1.5003276 0.37401583 7.9272605 3.1400451
θ2 0.1380841 0.02919803 0.6949408 0.2786503
θ3 0.4386253 0.04613525 0.7898723 0.1424101
AIC 54.51816 48.00966

In Figure 2, the steady state pdf are plotted for the two datasets, making use of the estimates of
the parameter θ given in Table 1 for the whole period and in Table 2 for the sub-intervals.


 







 

   

 


 


 

   

 






































 







Figure 2. The Vasicek steady-state densities for Datasets A (on the left) and B (on the right) evaluated
by using θ given in Table 1 for the whole period: 1 July 1997–31 March 2015 and by using the parameters
given in Table 2 for the sub-intervals identified by the Chow test.

3.2. The CIR Model

The CIR model, originally introduced by Feller as a model for population growth in 1951,
was proposed by John C. Cox, Jonathan E. Ingersoll, and Stephen A. Ross as an extension of the
valuation of interest rate derivatives. It describes the evolution of interest rates, and it is characterized
by the following SDE:
dX (t) = [θ1 − θ2 X (t)]dt + θ3 X (t) dW (t). (7)

We point out that Model (7) has widely been used in the literature in the context of neuronal
modeling (see, for example, [26–28]).

28
Mathematics 2018, 6, 70

The process X (t) in (7) is defined in I = (0, +∞). The nature of the boundaries 0 and +∞ depends
on the parameters of the process and establishes the conditions associated with the Kolmogorov and
Fokker–Planck equations to determine the transition pdf. In particular, the lower boundary 0 is exit
if θ1 ≤ 0, regular if 0 < θ1 < θ32 /2, and entrance if θ1 ≥ θ32 /2, whereas the endpoint +∞ is natural
(see [29]). In the following, we assume that θ1 , θ2 , θ3 are positive constants and that θ1 ≥ θ32 /2. This last
condition assures that X (t) is strictly positive so that the zero state is unattainable. In this case, the 0
state is an entrance boundary, so that the transition pdf can be obtained solving the Kolmogorov and
Fokker–Planck equations with the initial delta condition and a reflecting condition on the zero state.
Specifically, denoting
h−1 (y) = y2θ1 /θ3 e−2θ2 y/θ3
2 2

the inverse of the scale function defined in (2), the reflecting condition for the Kolmogorov equation is
∂
lim h−1 (y) f ( x, t|y, τ ) = 0,
y →0 ∂y
whereas, for the Fokker–Planck equation, it is

! ∂  θ2 x "
lim 3
f ( x, t|y, τ ) − (θ1 − θ2 x ) f ( x, t|y, τ ) = 0.
x →0 ∂x 2

Therefore, for θ1 ≥ θ32 /2, one obtains

 &

x −θ2 t θ1 /θ3 −1/2
2
2θ2 2θ2 ( x + x0 e−θ2 t )
f ( x, t| x0 , t0 ) = exp − e
θ32 [1 − e−θ2 t ] θ32 [1 − e−θ2 t ] x0
 
θ t
4θ2 (e x x0 )
2 1/2
× I2θ1 /θ 2 −1 (8)
3 θ32 (eθ2 t − 1)

where Iν (z) denotes the modified Bessel function of the first kind:

∞
(z/2)2k+ν
Iν (z) = ∑ k!Γ ( ν + k + 1)
k =0

and Γ is the Euler Gamma function:

 +∞
Γ(z) = tz−1 e−t dt.
0

The steady-state pdf for X (t) is a Gamma distribution with shape parameter 2θ1 /θ32 and scale
parameter θ32 /2θ2 , i.e.,

1  2θ 2θ1 /θ32 ! 2θ "

2 2
W (x) = x exp − 2 x .
x Γ(2θ1 /θ3 ) θ3
2 2 θ3

For Model (7), in Table 3, the maximum likelihood estimates of the parameters and the standard
errors and the AIC values are shown for Datasets A and B. Moreover, in the last row of this table, the
AIC is shown for the two datasets.
Note that the Chow test applied to the Euler discretization of Model (7) produces the same results
that the Vasiceck model does. Indeed, Models (5) and (7) show the same trend, but in the CIR model
one assumes residuals heteroschedasticity that does not bias the parameter estimates; it only makes
the standard errors incorrect. Moreover, in Table 4, the estimates of the parameters, the standard errors,
and the AIC values are shown before and after the structural breaks indicated by the Chow test. In
addition, in this case, the estimates for the two separated periods work better than the estimates using
only one model for the whole period shown from the AIC values.

29
Mathematics 2018, 6, 70

Table 3. ML estimates of Model (7) for Dataset A (on the left) and for Dataset B (on the right). Last row
shows the AIC.

CIR Model
Dataset A Dataset B
estimate standard error estimate standard error
θ1 0.87234371 0.57486263 0.71000000 0.69208186
θ2 0.06140606 0.03471271 0.06913925 0.05744829
θ3 0.24781675 0.02122343 0.16565230 0.01524389
AIC 204.3006 123.7508

In Figure 3, the steady state pdf are plotted for the two datasets making use of the estimates of the
parameter θ given in Table 3 for the whole period and in Table 4 for the sub-intervals.

Table 4. ML estimates of Model (7) and the corresponding AIC for the periods indicated by the Chow
test for Dataset A (on the top) and for Dataset B (on the bottom).

CIR Model
Dataset A
before 1 January 2008 after 1 January 2008
estimate standard error estimate standard error
θ1 0.70000000 1.96946686 0.50000000 1.5237235
θ2 0.04645403 0.10114005 0.05015628 0.1157965
θ3 0.25705245 0.02838514 0.21809718 0.0293771
AIC 130.8434 73.32189
Dataset B
before 1 January 2009 after 1 January 2009
estimate standard error estimate standard error
θ1 0.55000000 0.7378984 0.53234000 2.02364571
θ2 0.06276713 0.0587019 0.03746054 0.18050398
θ3 0.12443707 0.0159996 0.21651927 0.03194828
AIC 58.07491 75.83153


 







 

   

 


 


 

   

 


































 

 





 

 

Figure 3. CIR steady-state densities for Datasets A (on the left) and B (on the right) evaluated by using θ given
in Table 3 for the whole period: 1 July 1997–31 March 2015 and by using the parameters given in Table 4 for the
sub-intervals identified by the Chow test.

30
Mathematics 2018, 6, 70

4. FPT Analysis for TEGM

In this section, we consider the FPT analysis for the Vasicek model. This choice is motivated by
the results of Section 3. Indeed, comparing the two models by looking at the AIC values, we can see
that the Vasicek model better fits our datasets in all cases (only for Dataset A does the CIR model work
better than the Vasicek model).
For Datasets A and B, due to the Markovianity of the process, we consider the estimates relative
to the second periods (1 January 2008–31 March 2015 for Dataset A and 1 January 2009–31 March 2015
for Dataset B) shown in Table 2.
By using the recursive Equation (4), we obtain the estimates of FPT moments for Dataset A on the
period 1 January 2008–31 March 2015. In Table 5, these estimates are shown for various values of S (on
the top), with x0 = 13.28 corresponding to the mean of the data in the considered period, and various
values of the initial point x0 (on the bottom) fixing the alert boundary S = 15.
Table 6 shows the analogous analysis of Table 5 for Dataset B, with x0 = 12.1636 and S = 14.
We note that, by increasing the distance between S and x0 , the mean FPT increases. From an
economic point of view, the choice of such a distance can be interpreted as a choice of “propensity of
risk” of an available loan. Figure 4 shows the mean FPT (quarters starting from the loan deposit) as a
function of the alert boundary (up) and as a function of the initial point x0 . Clearly, each applicant
knows the initial point x0 and can choose the “alert boundary” S.

Table 5. For Dataset A, second period, mean, second order moment, and variance of the random
variable Tx0 through various values of the threshold S (on the top) and for various values of x0
(on the bottom).

x0 = 13.28 S t1 ( S | x 0 ) t2 ( S | x 0 ) Var (S| x0 )

14.0 6.780026 1.312067 × 102 8.5238 × 101
14.2 1.018325 × 10 2.612374 × 102 1.575387 × 102
14.4 1.488234 × 10 5.135466 × 102 2.920625 × 102
14.6 2.157931 × 10 1.021095 × 103 5.554285 × 102
14.8 3.144937 × 10 2.089585 × 103 1.100522 × 103
15 4.651822 × 10 4.462957 × 103 2.299013 × 103
15.2 7.038275 × 10 1.00658 × 104 5.112068 × 103
15.4 1.096303 × 102 2.421257 × 104 1.219378 × 104
15.6 1.76713 × 102 6.262414 × 104 3.139667 × 104
15.8 2.959467 × 102 1.752716 × 105 8.768709 × 104
16 5.16416 × 102 5.332541 × 105 2.665686 × 105
S = 15 x0 t1 ( S | x 0 ) t2 ( S | x 0 ) Var (S| x0 )
12.0 5.116115 × 10 4.937052 × 103 2.319589 × 103
12.2 1.808844 × 102 6.413534 × 104 6.413534 × 104
12.4 1.803493 × 102 6.394063 × 104 3.141475 × 104
12.6 1.797354 × 102 6.371754 × 104 3.141272 × 104
12.8 1.79022 × 102 6.345886 × 104 3.140997 × 104
13 1.781812 × 102 6.315425 × 104 3.140572 × 104
13.2 1.77174 × 102 6.279051 × 104 3.139988 × 104
13.4 1.759456 × 102 6.234752 × 104 3.139065 × 104
13.6 1.74417 × 102 6.179734 × 104 3.137605 × 104
13.8 1.724714 × 102 6.109852 × 104 3.135214 × 104
14 1.699329 × 102 6.018864 × 104 3.131143 × 104

31
Mathematics 2018, 6, 70

Table 6. For Dataset B, second period, mean, second order moment, and variance of the random
variable Tx0 through the threshold S for various values of S and of x0 .

x0 = 12.1636 S t1 ( S | x 0 ) t2 ( S | x 0 ) Var (S| x0 )

13 0.2653719 × 102 1.568046 × 103 8.638236 × 102
13.2 0.5332441 × 102 5.985332 × 103 3.141839 × 103
13.4 1.134217 × 102 2.631623 × 104 1.345175 × 104
13.6 2.606306 × 102 1.371027 × 105 6.917443 × 104
13.8 6.547357 × 102 8.602519 × 105 4.315731 × 105
14 1.808475 × 103 6.548544 × 106 3.277963 × 106
14.2 5.502628 × 103 6.05788 × 107 3.029989 × 107
14.4 1.844072 × 104 6.801859 × 108 3.401257 × 108
14.6 6.800683 × 104 9.250082 × 109 4.625154 × 109
14.8 2.75718 × 105 1.520417 × 1011 7.602126 × 1010
15 1.227843 × 106 3.015198 × 1012 1.507601 × 1012
S = 14 x0 t1 ( S | x 0 ) t2 ( S | x 0 ) Var (S| x0 )
11 1.813076 × 103 6.565252 × 103 3.278009 × 106
11.2 1.812688 × 103 6.563835 × 106 3.277998 × 106
11.4 1.81221 × 103 6.562115 × 106 3.27801 × 106
11.6 1.811604 × 103 6.559909 × 106 3.277999 × 106
11.8 1.810807 × 103 6.557028 × 106 3.278004 × 106
12. 1.809714 × 103 6.553067 × 106 3.278 × 106
12.2 1.80814 × 103 6.547343 × 106 3.277975 × 106
12.4 1.805734 × 103 6.538643 × 106 3.277966 × 106
12.6 1.801816 × 103 6.52444 × 106 3.2779 × 106
12.8 1.794949 × 103 6.499556 × 106 3.277714 × 106
13 1.781937 × 103 6.452422 × 106 3.277121 × 106

Dataset A Dataset B
500

Mean FPT (*10)

100 200
Mean FPT

200
0

14.0 14.5 15.0 15.5 16.0 14.0 14.5 15.0 15.5

Alert boundary Alert boundary

Mean FPT (*10)
180.5

180.5
Mean FPT

178.5

178.5

11.5 12.0 12.5 13.0 11.5 12.0 12.5 13.0

Initial value Initial value

Figure 4. Mean FPT versus the alert boundary (x0 = 13.28) and the initial value x0 (S = 15) for Dataset
A, second period (left), and for Dataset B, second period (right).

5. Conclusions
This paper addresses stochastic modeling of loan interest rate dynamics according to the current
laws against usury. Such modeling states an upper bound, above which an interest rate is considered a
usury rate and illegal. Here we focus on the Italian case and consider two models commonly used in
short-term loan rates, i.e., the Vasicek and CIR models. We propose a strategy based on FPT through
an alert boundary, above which the rate is considered at the risk of usury and hence has to be kept
under control. Moreover, the mean first exit time through the alert boundary can be an indicator of

32
Mathematics 2018, 6, 70

the “goodness” of the loan, in the sense that an applicant, when he/she is choosing between two loan
offers, should choose the one with a higher mean exit time from the alert boundary.
The procedure was applied to a historical series of Italian average rates on loans in the period
from 1 July 1997 to 31 March 2015. We considered “one-fifth of salary transfer” and two amount
classes were analyzed: (a) up to 10 million lira (until 31 December 2001) and up to e 5000 (after 2002);
and (b) above 10 million lire (until 31 December 2001) and above e 5000 (after 2002). The model
parameters were estimated by MLE, and a Chow test was applied to detect the presence of structural
breaks in our datasets.
The model and proposed strategy are apt for further development. Indeed, we can extend the
analysis to more general processes in which some parameters are time-dependent, or we can consider
time-dependent thresholds to model varying loan interest rates. Further generalization can include
analysis of FPT through two boundaries: the upper one describing an alert threshold and the lower
one representing a favorable interest rate.

Author Contributions: Giuseppina Albano and Virginia Giorno contributed equally in the writing of this work.
Acknowledgments: This paper was supported partially by INDAM-GNCS.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Brigo, B.; Mercurio, F. Interest Rate Models-Theory and Practice with Smile, Inflation and Credit; Springer: Berlin,
Germany, 2007.
2. Buetow, G.W., Jr.; Hanke, B.; Fabozzi, F.J. Impact of different interest rate models on bound value measures.
J. Fixed Income 2001, 3, 41–53. [CrossRef]
3. Fan, J.; Jiang, J.; Zhang, C.; Zhou, Z. Time-dependent diffusion models for term structure dynamics. Stat. Sin.
2003, 13, 965–992.
4. Chen, W.; Xu, L.; Zhu, S.P. Stock loan valuation under a stochastic interest rate model. Comput. Math. Appl.
2015, 70, 1757–1771. [CrossRef]
5. Di Lorenzo, E.; Orlando, A.; Sibillo, M. A Stochastic model fro loan interest rates. Banks Bank Syst. 2013,
8, 94–99.
6. Cox, J.C.; Ingersoll, J.E.; Ross, S.A. A Theory of the Term Structure of Interest Rates. Econometrica 1985, 53,
385–407. [CrossRef]
7. Vasicek, O. An equilibrium characterization of the term structure. J. Financ. Econ. 1977, 5, 177–188. [CrossRef]
8. Khramov, V. Estimating Parameters of Short-Term Real Interest Rate Models; IMF Working Papers; International
Monetary Fund: Washington, DC, USA, 2012.
9. Ricciardi, L.M.; Di Crescenzo, A.; Giorno, V.; Nobile, A.G. An outline of theoretical and algorithmic
approaches to first passage time problems with applications to biological modeling. Math. Jpn. 1999, 50,
247–322.
10. Albano, G.; Giorno, V.; Román-Román, P.; Torres-Ruiz, F. On the effect of a therapy able to modify both the
growth rates in a Gompertz stochastic model. Math. Biosci. 2013, 245, 12–21. [CrossRef] [PubMed]
11. Albano, G.; Giorno, V.; Román-Román, P.; Torres-Ruiz, F. On a Non-homogeneous Gompertz-Type Diffusion
Process: Inference and First Passage Time. In Computer Aided Systems Theory-EUROCAST; Moreno Diaz, R.,
Pichler, F., Quesada Arencibia, A., Eds.; Lecture Notes in Computer Science; Springer: Cham, Switzerland,
2017; Volume 10672, pp. 47–54.
12. Buonocore, A.; Nobile, A.G.; Ricciardi, L.M. A new integral equation for the evaluation of first-passage-time
probability densitie. Adv. Appl. Prob. 1987, 19, 784–800. [CrossRef]
13. Giorno, V.; Nobile, A.G.; Ricciardi, L.M., Sato, S. On the evaluation of firtst-passage-time probability densities
via non-singular integral equations. Adv. Appl. Prob. 1989, 21, 20–36. [CrossRef]
14. Buonocore, A.; Caputo, L.; D’Onofrio, G.; Pirozzi, E. Closed-form solutions for the first-passage-time problem
and neuronal modeling. Ricerche di Matematica 2015, 64, 421–439. [CrossRef]
15. Di Crescenzo, A.; Giorno, V.; Nobile, A.G. Analysis of reflected diffusions via an exponential time-based
transformation. J. Stat. Phys. 2016, 163, 1425–1453. [CrossRef]

33
Mathematics 2018, 6, 70

16. Abundo, M. The mean of the running maximum of an integrated Gauss–Markov process and the connection
with its first-passage time. Stoch. Anal. Appl. 2017, 35, 499–510. [CrossRef]
17. Román-Román, P.; Serrano-Pérez, J.; Torres-Ruiz, F. More general problems on first-passage times for
diffusion processes: A new version of the fptdapprox r package. Appl. Math. Comput. 2014, 244, 432–446.
[CrossRef]
18. Román-Román, P.; Serrano-Pérez, J.; Torres-Ruiz, F. fptdapprox: Approximation of First-Passage-Time
Densities for Diffusion Processes, R Package Version 2.1, 2015. Available online: http://cran.r-project.org/
package=fptdApprox (accessed on 23 November 2017).
19. Linetsky, V. Computing hitting time densities for CIR and OU diffusions: Applications to mean reverting
models. J. Comput. Financ. 2004, 4, 1–22. [CrossRef]
20. Albano, G.; Giorno, V. A stochastic model in tumor growth. J. Theor. Biol. 2006, 242, 229–236. [CrossRef]
[PubMed]
21. Buonocore, A.; Caputo, L.; Nobile, A.G.; Pirozzi, E. Restricted Ornstein–Uhlenbeck process and applications
in neuronal models with periodic input signals. J. Comp. Appl. Math. 2015, 285, 59–71. [CrossRef]
22. Dharmaraja, S.; Di Crescenzo, A.; Giorno, V.; Nobile, A.G. A continuous-time Ehrenfest model with
catastrophes and its jump-diffusion approximation. J. Stat. Phys. 2015, 161, 326–345. [CrossRef]
23. Giorno, V.; Spina, S. On the return process with refractoriness for non-homogeneous Ornstein–Uhlenbeck
neuronal model. Math. Biosci. Eng. 2014, 11, 285–302. [PubMed]
24. Iacus, S.M. Simulation and Inference for Stochastic Differential Equations with R Examples; Springer Series in
Statistics: Berlin, Germany, 2008.
25. Iacus, S.M. sde-Manual and Help on Using the sde R Package. 2009. Available online: http://www.
rdocumentation.org/packages/sde/versions/2.0.9 (accessed on 24 October 2017).
26. Giorno, V.; Lanský, P.; Nobile, A.G.; Ricciardi, L.M. Diffusion approximation and first-passage-time problem
for a model neuron. III. A birth-and-death process approach. Biol. Cybern. 1988, 58, 387–404. [CrossRef]
[PubMed]
27. Giorno, V.; Nobile, A.G.; Ricciardi, L.M. Single neuron’s activity: On certain problems of modeling and
interpretation. BioSystems 1997, 40, 65–74. [CrossRef]
28. Nobile, A.G.; Pirozzi, E. On time non-homogeneous Feller-type diffusion process in neuronal modeling.
In Computer Aided Systems Theory-EUROCAST 2017; Moreno Diaz, R., Pichler, F., Quesada Arencibia, A., Eds.;
Lecture Notes in Computer Science, Springer: Cham, Switzerland, 2015; Volume 9520, pp. 183–191.
29. Karlin, S.; Taylor, H.W. A Second Course in Stochastic Processes; Academic Press: New York, NY, USA, 1981.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

34
 mathematics
Article
Fractional Queues with Catastrophes and Their
Transient Behaviour
Giacomo Ascione 1 , Nikolai Leonenko 2 and Enrica Pirozzi 1, *
1 Dipartimento di Matematica e Applicazioni “Renato Caccioppoli”, Università degli Studi di Napoli Federico
II, 80126 Napoli, Italy; [email protected]
2 School of Mathematics, Cardiff University, Cardiff CF24 4AG, UK; [email protected]
* Correspondence: [email protected]

Received: 27 July 2018; Accepted: 31 August 2018; Published: 6 September 2018

Abstract: Starting from the definition of fractional M/M/1 queue given in the reference by
Cahoy et al. in 2015 and M/M/1 queue with catastrophes given in the reference by Di Crescenzo et al.
in 2003, we define and study a fractional M/M/1 queue with catastrophes. In particular, we focus our
attention on the transient behaviour, in which the time-change plays a key role. We first specify the
conditions for the global uniqueness of solutions of the corresponding linear fractional differential
problem. Then, we provide an alternative expression for the transient distribution of the fractional
M/M/1 model, the state probabilities for the fractional queue with catastrophes, the distributions of
the busy period for fractional queues without and with catastrophes and, finally, the distribution of
the time of the first occurrence of a catastrophe.

Keywords: fractional differential-difference equations; fractional queues; fractional birth-death

processes; busy period

MSC: 60K25; 60J80

1. Introduction
Stochastic models for queueing systems have a wide range of applications in computer systems,
sales points, telephone or telematic systems and also in several areas of science including biology,
medicine and many others. The well known M/M/1 queueing model [1–5] constitutes the theoretical
basis for building many other refined models for service systems.
Due to the Markov nature of its inter-arrival times of the customers and of its service times,
the model can be mathematically treated in a simple manner, and, for this reason, it is widely used in
many modeling contexts. Nevertheless, in the past few decades, the advent of fractional operators,
such as fractional derivatives and integrals (see, for instance, [6] and [7] and references therein), has
made it clear that different time scales, themselves random, that preserve memory (therefore not
Markovian), allow the construction of more realistic stochastic models.
The introduction of the fractional Caputo derivative into the system of differential-difference
equations for an M/M/1-type queue was done in [8], where, for a fractional M/M/1 queue, the state
probabilities were determined. In this kind of queue model, the inter-arrival times and service times
are characterized by Mittag–Leffler distributions [9]; in this case, the model does not have the property
of memory loss that is typical of the exponential distributed times of the classical M/M/1 model.
Indeed, a time-changed birth-death process [10,11], by means of an inverse stable subordinator [12],
solves the corresponding fractional system of differential-difference equations and fractional Poisson
processes [13] characterize the inter-arrival and service times.
The fractional M/M/1 model in [8] is an interesting and powerful model, not only because it is
a generalization of the classical one, where the fractional order is set to 1, but also because its range of

Mathematics 2018, 6, 159; doi:10.3390/math6090159 35 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 159

applications is extremely wide. Its importance can be further augmented by including in the model
the occurrence of catastrophes, as it was considered in [14] for the classical M/M/1.
The catastrophe is a particular event that occurs in a random time leading to the instantaneous
emptying of the system, or to a momentary inactivation of the system, as, for example, the action
of a virus program that can make a computer system inactive [15]; other applications of models
with catastrophes can be found in population dynamics and reliability contexts (see [16] and
references therein).
Motivated by the mathematical need to enrich the fractional M/M/1 model of [8] with the
inclusion of catastrophes, we study in this paper the above model; specifically, we determine the
transient distribution, the distribution of the busy period (including that of the fractional M/M/1
queue of [8]) and the probability distribution of the time of the first occurrence of the catastrophe.
For these purposes, we need to guarantee the global uniqueness of the solution of the considered
linear fractional Cauchy problem on Banach spaces. After recalling the definitions and known results in
Section 2, we address the problem of uniqueness in Section 3. In Section 2, we also provide the transient
distribution of the fractional M/M/1 model in an alternative form to that given in [8]. In Section 4,
the distribution of the busy period for the fractional M/M/1 queue (without catastrophes) is obtained.
Here, the time-changed birth-death process plays a key role to derive the results. In Section 5, we define
the fractional queue with catastrophes; we are able to obtain the distribution of the transient state
probabilities by following a strategy similar to that in [14]. We also found the distribution of the
busy period and of the time of the first occurrence of the catastrophe starting from the empty system.
Some special operators and functions used in this paper are specified in the Appendices A and B.

2. Definition of a Fractional Process Related to M/M/1 Queues

The classical M/M/1 queue process N (t), t ≥ 0 can be described as continuous time Markov
chain whose state space is the set {0, 1, 2, . . . } and the state probabilities

pn (t) = P( N (t) = n| N (0) = 0), n = 0, 1, 2 . . . (1)

satisfy the following differential-difference equations:

⎧
⎪
⎨ Dt pn (t) = −(α + β) pn (t) + αpn−1 (t) + βpn+1 (t),
⎪ n ≥ 1,
Dt p0 (t) = −αp0 (t) + βp1 (t), (2)
⎪
⎪
⎩ p (0) = δ , n ≥ 0,
n n,0

where δn,0 is the Kroeneker delta symbol, Dt = d

dt and α, β > 0 are the entrance and service
rates, respectively.

Let Sν (t), t ≥ 0, ν ∈ (0, 1) be the Lévy ν-stable subordinator with Laplace exponent given by:

log E e−zSν (t) = −tzν , z > 0.

Consider the inverse ν-stable subordinator

Lν (t) = inf{u ≥ 0 : Sν (u) > t}, t ≥ 0.

For 0 < ν < 1, the fractional M/M/1 queue process N ν (t), t ≥ 0 is defined by a non-Markovian
time change Lν (t) independent of N (t), t ≥ 0, i.e.,

N ν (t) = N ( Lν (t)), t ≥ 0. (3)

36
Mathematics 2018, 6, 159

This process was defined in [8] and it is non-Markovian with non-stationary and non-independent
increments. For ν = 1, by definition, N1 (t) = N (t), t ≥ 0. Then, for a fixed ν ∈ (0, 1], the state probabilities

pνn (t) = P{ N ν (t) = n| N ν (0) = 0}, n = 0, 1, . . . (4)

of the number of customers in the system at time t in the fractional M/M/1 queue are characterized
by arrivals and services determined by fractional Poisson processes of order ν ∈ (0, 1] [13] with
parameters α and β. They are solutions of the following system of differential-difference equations
⎧
⎪ C ν ν ν ν ν
⎨0 Dt pn (t) = −(α + β) pn (t) + αpn−1 (t) + βpn+1 (t),
⎪ n ≥ 1,
C D ν pν ( t ) = − αpν ( t ) + βp1ν (t), (5)
⎪0 t 0
⎪
0
⎩ p ν (0) = δ , n ≥ 0,
n n,0

where 0C Dtν is the Caputo fractional derivative (see Appendix A).

Using Equation (5) and representation (3), the state probabilities are obtained in [8]:


n
α α
pνn (t) = 1−
β β

n +∞ n+k

α k−m k+m k (6)
+
β ∑ ∑ k+m k
α
k =0 m =0
× βm−1 tν(k+m)−ν Eν,ν
k+m
(k+m)−ν+1
(−(α + β)tν ),

as well as its Laplace transform

 +∞ 
n

α 1 α
πnν (z) = e−zt pνn (t)dt = + 1−
0 β z β

n +∞ n+k

α k−m k+m k
+
β ∑ ∑ k+m k
α
k =0 m =0
z ν −1
× β m −1 , z > 0.
(zν + α + β )k+m
ρ
In Equation (6), the functions Eν,μ are generalized Mittag–Leffler functions (see Appendix B).
Note that pνn (t) ≥ 0 ∀n ≥ 0 and ∑n+=∞0 pνn (t) = 1.

d
Alternatively, let hν (t, x ) = P{ Lν (t) ≤ x }, x ≥ 0, be the density of Lν (t); then it is known
dx
(see, i.e., [17]) that  +∞
e−sx hν (t, x )dx = Eν (−stν ), s ≥ 0, (7)
0
and (see, i.e., [18], Proposition 4.1)
 +∞
1 ν
hν (t, x ) = uν−1 e−tu− xu cos(νπ )
sin (πν − xuν sin(πν))) du, x ≥ 0. (8)
π 0

37
Mathematics 2018, 6, 159

Using (7) and an analytical expression for pνn (t) given in [19], we can write down an alternative
expression for (6) as

n ∞
p (α + β)r rν (r)
pνn (t) = ∑ t Eν (−(α + β)tν )
q r =n r!
[ r−2 n ]
 (9)
r + 1 − 2k r + 1 k r−k
× ∑ r+1 k
p q ,
r =0

α β (r )
where p = , q = , and Eν (−(α + β)tν ) is the r −th derivative of the function Eν (z)
α+β α+β
evaluated at z = −(α + β)tν .
Actually, it is easy to see from (7) that
 +∞
e−sx xr hν (t, x )dx = Eν (−stν )tνr ;
(r )
0

thus, using [19] and (3), we have

 +∞
pνn (t) = pn (s)hν (t, s)ds
0

n ∞ 
p (α + β)r +∞ −(α+ β)s ν
=
q ∑ r! 0
e s hν (t, s)ds
r =n
[ r−2 n ]

r + 1 − 2k r + 1 k r−k
× ∑ r+1 k
p q ,
r =0

and formula (9) follows. On the other hand, using (8), we have
 +∞
pνn (t) = pn (s)hν (t, s)ds
0

n ∞ 
p (α + β)r +∞ ν−1 −tu
=
q ∑ r! 0
u e Fν,r (u)du
r =n
[ r−2 n ]

r + 1 − 2k r + 1 k r−k
× ∑ r+1 k
p q ,
r =0

where  +∞
1
Fν,r (u) = exp {−(α + β) x − xuν cos(νπ )} xr sin (πν − xuν sin(πν)) dx.
π 0

3. Linear Fractional Cauchy Problems on Banach Spaces

In order to describe the transient probabilities for our queues, we will need some uniqueness
results for solutions of linear fractional Cauchy problems defined on Banach spaces. To do that, let us
recall the following Theorem (Theorem 3.19 from [20]):

Theorem 1. Let ( X, | · |) be a Banach space and J = [0, T ] for some T > 0. Consider the ball BR = { x ∈ X :
| x | ≤ R}. Let ν ∈ (0, 1) and f : J × BR → X and consider the following Cauchy problem:

C ν
0 Dt x ( t ) = f (t, x (t)),
(10)
x (0) = x0 ,

where 0C Dtν is the Caputo derivative operator (see Appendix A).

Suppose that:

38
Mathematics 2018, 6, 159

• f ∈ C ( J × BR , X );
• There exists a constant M( R) > 0 such that

| f (t, x (t))| ≤ M( R)

for all x ∈ BR and t ∈ J and such that

M( R) T ν
R ≥ | x0 | + ;
Γ ( ν + 1)
2M( R)
• There exists a constant L > 0 such that L ≥ Γ ( ν +1)
;
• There exists a constant L0 > 0 such that

| f (t, x1 ) − f (t, x2 )| ≤ L0 | x1 − x2 |

for all x1 , x2 ∈ BR and t ∈ J;

• There exist constants ν1 ∈ (0, ν) and τ > 0 such that

L0 T (1+ β)(1−ν1 )  ν1 ν1

LA = < 1,
Γ(ν) (1 + β)1−ν1 τ

ν −1
where β = 1−ν1 .
Then, if x0 ∈ BR , the problem (10) admits a unique solution x ∗ ∈ Cν ( J, BR ).

The previous theorem can be easily adapted to the case in which J = [t0 , T + t0 ] and the starting
point of the derivative is t0 . Since we are interested in linear (eventually non-homogeneous) equations,
let us show how the previous theorem can be adapted in such a case.

Corollary 1. Consider the system (10) and suppose f (t, x ) = Ax + ξ where A : X → X is a linear and
continuous operator and ξ ∈ X. Then, there exists a R > | x0 | and T > 0 such that the system admits a unique
solution x ∗ ∈ Cν ( J, BR ).

Proof. Observe that, if | x | ≤ R, then

| f ( x )| ≤  A| x | + |ξ | ≤  A R + |ξ |.

Let us choose T such that the conditions of Theorem 1 are verified. To do that, consider M( R) =
 A R + |ξ |. Fix R ≥ | x0 | and define R̃ = R + ε for some ε > 0. Define then
 1
εΓ(ν + 1) ν
T=
M ( R̃)

and observe that

M( R̃) T ν
| x0 | + = | x0 | + ε ≤ R + ε = R̃.
Γ ( ν + 1)
2M( R̃)
Thus, one can fix L = Γ ( ν +1)
and L0 =  A. Moreover, since for fixed ν1 ∈ (0, ν) the function
τ → L A (τ ) is decreasing and limτ →0 L A (τ ) = 0, then one can easily find a τ > 0 such that L A (τ ) < 1.
Since we are under the hypotheses of Theorem 1, then we have shown the local existence and
uniqueness of a solution x ∗ ∈ Cν ( J, BR̃ ).

However, using such corollary, we can only afford local uniqueness. Global uniqueness of the
solution of the Cauchy problem (10) can be obtained with the additional hypothesis that such solution
is uniformly bounded:

39
Mathematics 2018, 6, 159

Corollary 2. Suppose we are under the hypotheses of Corollary 1. If there exists a solution x ∗ ∈ C ([0, +∞[, X )
and a constant k > 0 such that for any t ≥ 0 we have | x ∗ (t)| ≤ k, then such solution is unique.

Proof. Observe that | x0 | ≤ k and then fix R̃ = k + ε. Define

 1
εΓ(ν + 1) ν
ΔT = .
M( R̃)

Fix T1 = ΔT and observe that, by using Corollary 1, there exists a unique solution in [0, T1 ].
Since x ∗ is a solution of such problem, we have that x ∗ is unique. Suppose we have defined Tn−1 such
that x ∗ is the unique solution of system (10) in [0, Tn−1 ]. Consider the problem

C ν
Tn−1 Dt x ( t ) = f ( x (t)),
(11)
x ( Tn−1 ) = x ∗ ( Tn−1 ).

Define then Tn = Tn−1 + ΔT and observe that, since | x ∗ ( Tn−1 )| ≤ k, by using Corollary 1,
there exists a unique solution in [ Tn−1 , Tn ].
By using a change of variables, it is easy to show that

ν
C
Tn−1 Dt x = 0C Dtν−Tn−1 x̃,

where x̃ : t → x (t + Tn−1 ). By using such relation, we have that system (11) is equivalent to

C ν
0 Dt x̃ ( t ) = f ( x̃ (t)),
x̃ (0) = x ∗ ( Tn−1 ),

whose unique solution is x̃ (t) = x ∗ (t + Tn−1 ) so that x (t) = x ∗ (t) and x ∗ (t) is the unique solution of
system (10) in [0, Tn ]. Since Tn → +∞ as n → +∞, we have global uniqueness of limited solutions.

4. The Fractional M/M/1 Queue

Let us consider again the fractional M/M/1 process N ν (t), t ≥ 0 defined by (3) with state
probabilites in (6).
Consider the Hilbert space (l 2 (R), | · |2 ) with the norm | x |22 = ∑k+=∞0 xk2 and let Cν ([0, T ], l 2 (R)) be
the space of the ν-Hölder continuous functions from [0, T ] to l 2 (R). One can rewrite the system (5) in
l 2 (R) as follows:

C ν ν ν
0 Dt p ( t ) = A 0 p ( t ) ,
(12)
pν (0) = (δn,0 )n≥0 ,

where pν (t) = ( pνn (t))n≥0 ∈ C ([0, T ], l 2 (R)) and

⎛ ⎞
−α β 0 0 0 ···
⎜ α −(α + β) β · · ·⎟
⎜ 0 0 ⎟
⎜ ⎟
A0 = ⎜ 0 α −(α + β) β 0 · · ·⎟
⎜ ⎟
⎜ 0 0 α −( α + β) β · · ·⎟
⎝ ⎠
.. .. .. .. .. ..
. . . . . .

is an infinite tridiagonal matrix with A0 = ( ai,j )i,j≥0 . Let us show the following:

Lemma 1. The linear operator A0 is continuous and  A0  ≤ 2(α + β).

40
Mathematics 2018, 6, 159

Proof. To show that A0 is continuous, let us use Schur’s test (Theorem 5.2 in [21]). Observe that

+∞ +∞
∑ |ak,0 | = 2α, ∑ |ak,j | = 2(α + β) for j = 0
k =0 k =0

so that, in general,
+∞
∑ |ak,j | ≤ 2(α + β).
k =0

Moreover,
+∞ +∞
∑ |a0,k | = α + β, ∑ |a j,k | = 2(α + β) for j = 0,
k =0 k =0

so that, in general,
+∞
∑ |a j,k | ≤ 2(α + β).
k =0

By Schur’s test, we have that A0 is a bounded operator on l 2 and

 A0  ≤ 2( α + β ).

Thus, by Corollary 1, we obtain local existence and uniqueness of the solution of system (5).
Global uniqueness can be obtained a posteriori, since the solutions of such system are known.
Let us also observe that the distributions of the inter-arrival times are Mittag–Leffler distributions.
To do that, consider the system, for fixed n ≥ 0
⎧
⎪ C ν ν ν
⎪0 Dt bn (t) = −αbn (t),
⎪
⎪
⎪
⎨C D ν bν (t) = αbν (t),
0 t n +1 n
⎪
⎪ bnν (0) = 1,
⎪
⎪
⎪ ν
⎩ bn+1 (0) = 0,

which are the state probabilities of a queue with null death rate, fixed birth rate, starting with n
customers and with an absorbent state n + 1. Under such assumptions, bnν +1 (t) is the probability that
a customer arrives before t. Moreover, the normalizing condition becomes

bnν (t) + bnν +1 (t) = 1.

One can solve the first equation (see Appendix A) to obtain

bnν (t) = Eν (−αtν ),

where Eν is the one-parameter Mittag–Leffler function (see Appendix B), and then, by using the
normalizing condition, we have
bnν +1 (t) = 1 − Eν (−αtν ).

41
Mathematics 2018, 6, 159

In a similar way, let us show that the distributions of the service times are Mittag–Leffler
distributions. To show that, consider the system, for fixed n ≥ 0,
⎧
⎪
⎪0C Dtν dνn (t) = βdνn+1 (t),
⎪
⎪
⎪
⎨C D ν dν (t) = − βdν (t),
0 t n +1 n +1
⎪
⎪dνn (t) = 0,
⎪
⎪
⎪
⎩dν (t) = 1,
n +1

which are the state probabilities of a queue with null birth rate, fixed death rate, starting with n + 1
customers with an absorbent state n. Under such assumption, dνn (t) is the probability that a customer
is served before t. Moreover, the normalizing condition becomes

dνn (t) + dνn+1 (t) = 1.

One can solve the second equation to obtain

dνn+1 (t) = Eν (− βtν ), t ≥ 0

and then, by using the normalizing condition, we have

dνn (t) = 1 − Eν (− βtν ), t ≥ 0.

Moreover, since we know that ∀t ≥ 0 pνn (t) ≥ 0 and ∑∞ ν

n=0 pn ( t ) = 1, by the continuous inclusion
⊆ l (R) (see [22]), ( pn (t))n≥0 is uniformly bounded in l (R) and then, by Corollary 2, it is the
l 1 (R) 2 2

(global) unique solution of system (5).

Distribution of the Busy Period

We want to determine the probability distribution K ν (t) of the busy period K ν of a fractional
M/M/1 queue. To do this, we will follow the lines of the proof given in [1] and [4].

Theorem 2. Let K ν be the random variable describing the duration of the busy period of a fractional M/M/1
queue N ν (t) and consider K ν (t) = P(K ν ≤ t). Then,

+∞ +∞
K ν (t) = 1 − ∑ ∑ Cn,m tν(n+2m−1) Eν,ν
n+2m
(n+2m−1)+1
(−(α + β)tν ), (13)
n =1 m =0

where

n + 2m n
Cn,m = α n + m −1 β m . (14)
m n + 2m
ν ν ν
Proof. Let us first define a queue N (t) such that P( N (0) = 1) = 1 and N (t) behaves like N ν (t)
except for the state 0 being an absorbent state. Thus, state probability functions are solution of the
following system
⎧
⎪ C ν ν ν
⎪0 Dt p0 = βp1 (t),
⎪
⎪
⎪
⎨C D ν pν = −(α + β) pν (t) + βpν (t),
0 t 1 1 2
(15)
⎪
⎪0C Dtν pνn = −(α + β) pνn (t) + αpνn−1 (t) + βpνn+1 (t), n ≥ 2,
⎪
⎪
⎪
⎩ p ν (0) = δ ,
n n,1 n ≥ 0.

42
Mathematics 2018, 6, 159

First, we want to show that, if we consider Lν (t), the inverse of a ν-stable subordinator that is
1 ν d 1
independent from N (t), then N (t) = N ( Lν (t)). To do that, consider the probability generating
ν
function G ν (z, t) of N (t) defined as

+∞
G ν (z, t) = ∑ zk pνk (t). (16)
k =0

From system (15), we know that G ν (z, t) solves the following fractional Cauchy problem:

z0C Dtν G ν (z, t) = [αz2 − (α + β)z + β][ G ν (z, t) − p0ν (t)],
(17)
G ν (z, 0) = z,

which, for ν = 1, becomes


d 1
z dt G (z, t) = [αz2 − (α + β)z + β][ G1 (z, t) − p10 (t)],
(18)
G1 (z, 0) = z.

Taking the Laplace transform in Equation (17) and using Equation (A1), we have

ν (z, s) − zsν−1 ] = [αz2 − (α + β)z + β][ G

z[sν G ν (z, s) − π ν (s)], (19)
0

ν (z, s) and π ν (s) are Laplace transforms of G ν (z, t) and pν (t).

where G 0 0
ν d 1
We know that N (t) = N ( Lν (t)) if and only if
 +∞
ν 1
pνn (t) = P( N (t) = n) = P( N ( Lν (t)) = n) = p1n (y) P( Lν (t) ∈ dy) (20)
0

and then if and only if, by Equation (16),

 +∞
G ν (z, t) = G1 (z, y) P( Lν (t) ∈ dy). (21)
0

Taking the Laplace transform in Equations (20) and (21) for n = 0 and by using (see, i.e.,
Equation (10) in [12])
L[P( Lν (t) ∈ dy)](s) = sν−1 e−ys dy, (22)

we know we have to show that

 +∞
ν
π 0ν (s) = p1n (y)sν−1 e−ys dy (23)
0

and  +∞
ν
ν (z, s) =
G G1 (z, y)sν−1 e−ys dy. (24)
0
Since Equation (17) admits a unique solution, then we only need to show that the right-hand sides
of Equations (23) and (24) solve Equation (19), that is to say that we have to verify
  +∞

ν
z sν G1 (z, y)e−ys dy − z
0
 +∞  +∞ 
ν ν
= [αz2 − (α + β)z + β] G1 (z, y)e−ys dy − p10 (y)e−ys dy . (25)
0 0

43
Mathematics 2018, 6, 159

To do that, consider the right-hand side of the previous equation and, recalling that G1 (z, t) is
solution of Equation (18):
 +∞  +∞

ν d ν
[αz2 − (α + β)z + β][ G1 (z, y) − p10 (y)]e−ys dy = G1 (z, y) e−ys dy
0 0 dy

and then, by integrating by parts, we have Equation (25).

Now remark that p0ν (t) = Bν (t). Thus, we want to determine p0ν (t). To do that, let us recall,
from [1,4] that 
p1n (t) = nt−1 α 2 −1 β− 2 e−(α+ β)t In (2 αβt) for n ≥ 1
n n


from which, explicitly writing In (2 αβt), we have

+∞

n + 2m n 1
p1n (t) = ∑ αn+m−1 βm tn+2m−1 e−(α+ β)t for n ≥ 1.
m =0
m n + 2m (n + 2m − 1)!

Posing Cn,m = (n+m2m) n+n2m αn+m−1 βm , we have

+∞
Cn,m
p1n (t) = ∑ tn+2m−1 e−(α+ β)t for n ≥ 1
m =0
( n + 2m − 1)!

and then
+∞ +∞
Cn,m
p10 (t) = 1 − ∑ ∑ tn+2m−1 e−(α+ β)t . (26)
n =1 m =0
(n + 2m − 1)!
ν 1
Since N (t) = N ( Lν (t)), we have
 +∞
p0ν (t) = p10 (y) P( Lν (t) ∈ dy)
0

and then, using Equation (26), we have

+∞ +∞  +∞
Cn,m
p0ν (t) = 1 − ∑ ∑ yn+2m−1 e−(α+ β)y P( Lν (t) ∈ dy). (27)
n =1 m =0
(n + 2m − 1)! 0

Taking the Laplace transform in Equation (27), using Equation (22), we have
 +∞
1 +∞ +∞ Cn,m ν )y
π 0ν (s) = s ν −1 yn+2m−1 e−(α+ β+s
s n∑ ∑
− dy
=1 m=0 ( n + 2m − 1) ! 0

and then integrating

1 +∞ +∞ s ν −1
π 0ν (s) = − ∑ ∑ Cn,m .
s n =1 m =0 (α + β + sν )n+2m
Finally, using formula (A2), we have

+∞ +∞
p0ν (s) = 1 − ∑ ∑ Cn,m tν(n+2m−1) Eν,ν
n+2m
(n+2m−1)+1
(−(α + β)tν )
n =1 m =0

Remark 1. As ν → 1 we obtain, by using

n+2m e−(α+ β)t

E1,n +2m (−( α + β ) t ) = (n + 2m − 1)!
,

44
Mathematics 2018, 6, 159

that p0ν (t) → p10 (t) and then K ν (t) → K1 (t).

5. The Fractional M/M/1 Queue with Catastrophes

Let us consider a classical M/M/1 queue with FIFO discipline and subject to catastrophes whose
effect is to instantaneously empty the queue [14] and let Nξ1 (t) be the number of customers in the
system at time t with state probabilities
1,ξ
pn (t) = P( Nξ1 (t) = n| Nξ1 (0) = 0), n = 0, 1, . . .

1,ξ
Then, the function pn satisfy the following differential-difference equations:
⎧
⎪ 1,ξ 1,ξ 1,ξ
⎨ Dt p0 (t) = −(α + ξ ) p0 (t) + βp1 (t) + ξ,
⎪
1,ξ 1,ξ 1,ξ 1,ξ
Dt pn (t) = −(α + β + ξ ) pn (t) + αpn−1 (t) + βpn+1 (t), n ≥ 1, (28)
⎪
⎪
⎩ p1,ξ (0) = δ , n ≥ 0,
n n,0

where δn,0 is the Kroeneker delta symbol, Dt = dt d

, α, β > 0 are the entrance and service rates,
respectively, and ξ > 0 is the rate of the catastrophes when the system is not empty.

For ν ∈ (0, 1), we define the fractional M/M/1 queue process with catastrophes as

Nξν (t) = Nξ1 ( Lν (t)), t ≥ 0,

where Lν is an inverse ν-stable subordinator that is independent of Nξ1 (t), t ≥ 0 (see Section 2).
We will show that the state probabilities
ν,ξ
pn := P( Nξν (t) = n| Nξν (0) = 0)

satisfy the following differential-difference fractional equations:

⎧
⎪ C ν ν,ξ ν,ξ ν,ξ
⎨0 Dt p0 (t) = −(α + ξ ) p0 (t) + βp1 (t) + ξ,
⎪
C ν ν,ξ ν,ξ ν,ξ ν,ξ
0 Dt pn ( t ) = −( α + β + ξ ) pn ( t ) + αpn−1 ( t ) + βpn+1 ( t ), n ≥ 1, (29)
⎪
⎪
⎩ pν,ξ (0) = δ , n ≥ 0,
n n,0

where 0C Dtν is the Caputo fractional derivative (see Appendix A).

In the classical case, catastrophes occur according to a Poisson process with rate ξ if the system is
not empty. In our case, consider for a fixed n > 0,
⎧
⎪ C ν ν ν
⎪0 Dt c0 (t) = ξ (1 − c0 (t)),
⎪
⎪
⎪
⎨C D ν cν (t) = −ξcν (t),
0 t n n
⎪
⎪ c0ν (0) = 0,
⎪
⎪
⎪
⎩cν (0) = 1,
n

which describes the state probabilities of an initially non empty system with null birth and death rate
but positive catastrophe rate. In such case, c0ν (t) is the probability a catastrophe occurs before time t.
Moreover, the normalization property becomes

c0ν (t) + cνn (t) = 1.

45
Mathematics 2018, 6, 159

In such case, we can solve the second equation to obtain

cνn (t) = Eν (−ξtν ), t ≥ 0.

Using the normalization property, we finally obtain

c0ν (t) = 1 − Eν (−ξtν ), t ≥ 0 (30)

and then the distributions of the inter-occurrence of the catastrophes are Mittag–Leffler distributions.
We can conclude that, in the fractional case, catastrophes occur according to a fractional Poisson
process ([10,11,13]) with rate ξ if the system is not empty. Since the operators 0C Dtν are Caputo
fractional derivatives, we expect the stationary behaviour of the queue to be the same as the classic
one. Denoting with Nξ1 the number of customers in the system at the steady state of a classical M/M/1
with catastrophes and defining the state probabilities

qn = P( Nξ1 = n), n ≥ 0,

we can use the results obtained in [15] to observe that


n
1 1
qn = 1− , n ≥ 0, (31)
z1 z1

where z1 is the solution of

αz2 − (α + β + ξ )z + β = 0 (32)

such that z1 > 1. Let us call z2 the other solution of Equation (32) and observe that 0 < z2 < 1 < z1 .
Some properties coming from such equations that will be useful hereafter are

β
α + β + ξ = αzi + (33)
zi

and
αz2i = (α + β + ξ )zi − β (34)

with i = 1, 2.

5.1. Alternative Representation of the Fractional M/M/1 Queue with Catastrophes

We want to obtain an alternative representation of the fractional M/M/1 queue with catastrophes
in a way which is similar to Lemma 2.1 in [14]. To do that, we firstly need to assure that system (29)
admits a unique uniformly bounded solution. To do that, let us write system (29) in the form

C ν ν,ξ ν,ξ
0 Dt p ( t ) = f ( p ( t )),
(35)
ν,ξ
p (t) = (δn,0 )n≥0 ,

ν,ξ
where pν,ξ (t) = ( pn (t))n≥0 ∈ C ([0, T ], l 2 (R)), f ( x) = Aξ x + ξ, ξ = (ξ, 0, . . . , 0, . . . ) and
⎛ ⎞
−(α + ξ ) β 0 0 0 ···
⎜ α −(α + β + ξ ) β · · ·⎟
⎜ 0 0 ⎟
⎜ ⎟
Aξ = ⎜ 0 α −(α + β + ξ ) β 0 · · ·⎟
⎜ ⎟
⎜ 0 0 α −(α + β + ξ ) β · · ·⎟
⎝ ⎠
.. .. .. .. .. ..
. . . . . .

is an infinite tridiagonal matrix with Aξ = ( ai,j )i,j≥0 . We need to show the following:

46
Mathematics 2018, 6, 159

Lemma 2. The linear operator Aξ is continuous and  Aξ  ≤ 2(α + β) + ξ.

Proof. To obtain an estimate of the norm of Aξ , let us use Schur’s test. Observe that

+∞ +∞
∑ |ak,0 | = 2α + ξ, ∑ |ak,j | = 2α + 2β + ξ with j = 0,
k =0 k =0

so that, in general,
+∞
∑ |ak,j | ≤ 2α + 2β + ξ.
k =0

Moreover,
+∞ +∞
∑ |a0,k | = α + β + ξ, ∑ |a j,k | = 2α + 2β + ξ for j = 0
k =0 k =0

so that, in general,
+∞
∑ |a j,k | ≤ 2α + 2β + ξ.
k =0

By Schur’s test, we have that Aξ is a bounded operator on l 2 and

 Aξ  ≤ 2(α + β) + ξ.

Observe that, if ξ = 0, the operator A0 is the same of system (12). Let us also observe that by
Corollary 1 there locally exists a unique solution. Moreover, if we show that a solution is uniformly
bounded, such solution is unique. Now, we are ready to adapt Lemma 2.1 of [14] to the fractional case.

Theorem 3. Let N  ν (t) be the number of customers in a fractional M/M/1 queue with arrival rate αz1 and
β  ν (0) = 0) = 1 and consider N a random variable independent from N
 ν (t) whose
service rate z such that P( N
1
state probabilities qn are defined in Equation (31). Define

 ν (t), N }, t ≥ 0.
Mν (t) := min{ N

Then, the state probabilities of Mν (t) are the unique solutions of (29).

d d ν,ξ
Moreover, Mν (t) = Nξν (t), where = is the equality in distribution, and then pn (t), n = 0, 1, . . . are the
unique solutions of (29).

Proof. Define pn∗,ν (t) = P( Mν (t) = n) and   ν (t) = n). Since N

pνn (t) = P(N  ν (t) and N are independent, then

 ν (t) ≥ n) + P( N
p∗n,ν (t) = P( N = n) P( N  ν (t) = n) P( N > n),

which, by using the definitions of pνn (t) and qn , becomes

( )
+∞ +∞
p∗n,ν (t) = qn ∑ pνk (t) + ∑ qn pνn (t). (36)
k=n k = n +1

Moreover, by using Equation (31), we have

+∞
 +∞
k

 n +1 + ∞
k
 n +1
1 1 1 1 1 1
∑ qn = 1−
z1 ∑ z1
= 1−
z1 z1 ∑ z1
=
z1
(37)
k = n +1 k = n +1 k =0

47
Mathematics 2018, 6, 159

and then, substituting Equation (37) in (36), we obtain

+∞
 n +1
1
pn∗,ν (t) = qn ∑ pνk (t) + pνn (t). (38)
k=n
z1

We want to show that Mν (t) = N ν (t). Since, by definition, p∗n,ν (t) are non-negative and
∗,ν
∑n+=∞0 pn (t) = 1, they are uniformly
bounded in l 2 (R). Thus, we only need to show that p∗,ν (t) =
( p∗n,ν (t))n≥0 solves system (35).
The initial conditions are easily verified, so we only need to verify the differential relations. Observe that

1 ν
p0∗,ν (t) = q0 + p (t)
z1 0

and then, applying the Caputo derivative operator, we obtain

C ν ∗,ν 1C ν ν
0 Dt p 0 ( t ) = D p (t).
z1 0 t 0
β
Since p0ν (t) is a solution of system (5) with rates αz1 and z1 , we have

C ν ∗,ν β ν
0 Dt p 0 ( t ) = −α p0ν (t) + p (t).
z21 1

Observe also that

2
1
p1∗,ν (t) = q1 (1 − p0ν (t)) + p1ν (t)
z1
so we have

− (α + ξ ) p0∗,ν (t) + βp1∗,ν (t) + ξ

  
2 
1 1
= −(α + ξ ) q0 + p0ν (t) + β q1 (1 − p0ν (t)) + p1ν (t) + ξ.
z1 z1

After some calculations, we obtain

α+ξ ν β
−(α + ξ ) p0∗,ν (t) + βp1∗,ν (t) + ξ = −(α + ξ )q0 − p (t) + βq1 − βq1 p0ν (t) + 2 p1ν (t) + ξ.
z1 0 z1

Let us remark that



1 1 1
q0 = 1 − , q1 = 1− ,
z1 z1 z1

so we have

− (α + ξ ) p0∗,ν (t) + βp1∗,ν (t) + ξ

−αz21 + (α + β + ξ )z1 − β (−(α + β + ξ )z1 + β) p0ν (t) β
= + + 2 p1ν (t).
z21 z21 z1

By using Equations (32) and (34), we obtain

β ν
−(α + ξ ) p0∗,ν (t) + βp1∗,ν (t) + ξ = −α p0ν (t) + p (t) = 0C Dtν p0∗,ν (t).
z21 1

48
Mathematics 2018, 6, 159

Rewrite now Equation (38) in the form

( )
 n +1
+∞
1
pn∗,ν (t) = qn 1− ∑ pνk (t) + pνn (t) (39)
k =0
z1

and then apply a Caputo derivative operator to obtain

+∞
 n +1
C ν ∗,ν C ν ν 1
0 Dt p n ( t ) = −qn ∑ k (t) − qn 0C Dtν p0ν (t) +
0 Dt p
C ν ν
n (t).
0 Dt p
k =1
z1

β
Since pνn (t) is a solution of system (5) with birth rate αz1 and death rate z1 , then we have


C ν ∗,ν β n −1 ν n −2
β n
0 Dt p n ( t ) = qn αz1 + ∑ pk − qn αz1 ∑ pνk (t) − qn ∑ pνk (t)
z1 k =1 k =0
z1 k =2

 n +1

β 1 β
+ αz1 qn p0 (t) − qn p1ν − αz1 + pνn (t)
z1 z1 z1

n
 n +2
1 1
+α pνn−1 (t) + β pνn+1 (t).
z1 z1

Remark that, by using Equation (39),

−(α + β + ξ ) pn∗,ν (t) + αpn∗,ν ∗,ν

−1 ( t ) + βpn+1 ( t ) =
( ( )
 n +1 )
n −1
1
− (α + β + ξ ) qn 1− ∑ pνk (t) + pνn (t)
k =0
z1
( ( )
n )
n −2
1
+ α q n −1 1 − ∑ pνk (t) + pνn−1 (t)
k =0
z1
( ( )
 n +2 )
n
1
+ β q n +1 1 − ∑ pνk (t) + pνn+1 (t) .
k =0
z1

qn
Then, recalling that by definition qn−1 = z1 qn and qn+1 = z1 and doing some calculations,
we have

−(α + β + ξ ) p∗n,ν (t) + αp∗n,ν ∗,ν

−1 ( t ) + βpn+1 ( t ) =
n −1 n −2n
β
(α + β + ξ )qn ∑ pνk (t) − αz1 qn ∑ pνk (t) −
qn ∑ pνk (t)
k =1 k =0
z1 k =2

 n +1
( α + β + ξ ) z1 − β β 1
+ qn p0ν (t) − qn p1ν (t) − (α + β + ξ ) pνn (t)
z1 z1 z1

n
 n +2
1 1 αz2 − (α + β + ξ )z1 + β
+α pνn−1 (t) + β pνn+1 (t) + 1 qn .
z1 z1 z1

49
Mathematics 2018, 6, 159

Finally, by using Equations (32), (33) and (34), we have

−(α + β + ξ ) p∗n,ν (t) + αp∗n,ν ∗,ν

−1 ( t ) + βpn+1 ( t ) =

 n −1 n −2 n
β β
αz1 + qn ∑ pνk (t) − αz1 qn ∑ pνk (t) − qn ∑ pνk (t)
z1 k =1 k =0
z 1 k =2


 n +1
β β 1
+ αz1 qn p0ν (t) − qn p1ν (t) − αz1 + pνn (t)
z1 z1 z1

n
 n +2
1 1
+α pνn−1 (t) + β pνn+1 (t) = 0C Dtν p∗n,ν (t).
z1 z1

We have shown that the state probabilities pn∗,ν (t) of Mν (t) are the unique solutions of system (29).
d
Now, we need to show that Mν (t) = N ν (t). To do this, consider N  1 (t) a classical M/M/1 queue with
ξ
β  ν (t) and N
 1 (t) with
arrival rate αz1 and service rate z , N a random variable independent from N
1
probability masses qn and finally Lν (t) the inverse of a ν-stable subordinator which is independent from
N and N 1 (t). Define also M1 (t) = min{ N 1 (t), N }. By Lemma 2.1 of [14], we know that M1 (t) = d
N 1 ( t ), ξ
d d  1 ( Lν (t)) =
so M1 ( Lν (t)) = Nξ1 ( Lν (t)) = Nξν (t). However, by definition, we know that N
d ν
N ( t ),
thus finally
d d d
Mν (t) = M1 ( Lν (t)) = Nξ1 ( Lν (t)) = Nξν (t).

5.2. State Probabilities for the Fractional M/M/1 with Catastrophes

d
Since we have defined Nξν (t) = Nξ1 ( Lν (t)), where Lν (t) is the inverse of a ν-stable subordinator,
which is independent from Nξ1 (t), we can use such definition and Theorem 3 with the results obtained
in [14] to study the state probabilities of Nξν (t). In particular, we refer to the formula

+∞ m+n 1
Cn,m,r
∑ ∑ tm+r−1 e−(α+ β+ξ )t
1,ξ
pn (t) = qn +
m =0 r =0 ( m + r − 1) !
+∞ +∞ 2
(40)
Cn,m,r
+ ∑ ∑ (m + r − 1)! tm+r−1 e−(α+β+ξ )t ,
m =0 r = m + n +1

where


1 z1 − 1 m + r m − r m r −1
Cn,m,r = β α ,
(z1 − z2 )z1n+m+1−r r m+r

 (41)
2 1 − z2 m + r r − m m r −1
Cn,m,r = n + m +1−r
β α .
( z1 − z2 ) z2 r r+m

By using such formula, we can show the following:

Theorem 4. For any t > 0 and n = 0, 1, . . . , we have

+∞ m+n
ν,ξ
pn (t) = qn + ∑ ∑ 1
Cn,m,r tν(m+r−1) Eν,ν
m +r
(m+r −1)+1
(−(α + β + ξ )tν )
m =0 r =0
+∞ +∞
(42)
+ ∑ ∑ 2
Cn,m,r tν(m+r−1) Eν,ν
m +r
(m+r −1)+1
(−(α + ν
β + ξ ) t ),
m =0 r = m + n +1

i
where Cn,m,r are defined in (41).

50
Mathematics 2018, 6, 159

d
Proof. From Nξν (t) = Nξ1 ( Lν (t)), we have
 +∞
ν,ξ 1,ξ
pn (t) = pn (y) P( Lν (t) ∈ dy)
0

and then, by using formula (40),

+∞ m+n 1  +∞
ν,ξ Cn,m,r
pn (t) = qn + ∑ ∑ ym+r−1 e−(α+ β+ξ )y P( Lν (t) ∈ dy)
m =0 r =0
( m + r − 1) ! 0
+∞ +∞ 2  +∞
Cn,m,r
+ ∑ ∑ ym+r−1 e−(α+ β+ξ )y P( Lν (t) ∈ dy).
m =0 r = m + n +1
( m + r − 1) ! 0

Taking the Laplace transform and using Equation (22), we obtain

+∞ m+n 1  +∞
ν,ξ Cn,m,r ν
πn (s) = qn + ∑ ∑ s ν −1 ym+r−1 e−(α+ β+ξ +s )y dy
m =0 r =0
( m + r − 1) ! 0
+∞ +∞ 2  +∞
Cn,m,r ν
+ ∑ ∑ s ν −1 ym+r−1 e−(α+ β+ξ +s )y dy
m =0 r = m + n +1
( m + r − 1) ! 0

and then, integrating

+∞ m+n
ν,ξ s ν −1
πn (s) = qn + ∑ ∑ 1
Cn,m,r
( α + β + ξ + s ν ) m +r
m =0 r =0
+∞ +∞
s ν −1
+ ∑ ∑ 2
Cn,m,r
( α + β + ξ + s ν ) m +r
.
m =0 r = m + n +1

Finally, by using Equation (A2), we obtain

+∞ m+n
ν,ξ
pn (t) = qn + ∑ ∑ 1
Cn,m,r tν(m+r−1) Eν,ν
m +r
(m+r −1)+1
(−(α + β + ξ )tν )
m =0 r =0
+∞ +∞
+ ∑ ∑ 2
Cn,m,r tν(m+r−1) Eν,ν
m +r
(m+r −1)+1
(−(α + β + ξ )tν ).
m =0 r = m + n +1

ν,ξ
Remark 2. From formula (42), we can easily see that limt→+∞ pn (t) = qn so, as we expected, the steady-state
probabilities are the same as the classical ones. For such reason, we can say that the fractional behaviour is
influential only in the transient state of the queue.

Remark 3. As ν → 1, by using

m +r e−(α+ β+ξ )t
E1,m +r (−( α + β + ξ ) t ) = ( m + r − 1) !
,

ν,ξ 1,ξ
we obtain that limν→1 pn (t) = pn (t).
   n
β
Remark 4. If α < β and ξ = 0, then z1 = α and z2 = 1. For such reason, qn = 1 − αβ α
β ,
 n
m −r ν,ξ
1
Cn,m,r = αβ m m + r m r −1 and C2
+r ( m ) α β n,m,r = 0. Then, we have that pn ( t ) of Equation (42) has the form of
Equation (6).

51
Mathematics 2018, 6, 159

β
If α > β and ξ → 0 then z1 = 1 and z2 = α.
1
In such case, qn = 0, Cn,m,r 2
= 0 and Cn,m,r =
αn+m βr−n−1 (mm+r) rm−+mr . For such case, we have

n +∞ +∞

ν,ξ α m r −1 m + r r − m ν ( m +r −1) m +r
lim pn (t) =
ξ →0 β ∑ ∑ α β
m m+r
t Eν,ν(m+r−1)+1 ,
m =0 r = m + n +1

which is not recognizable as a previously obtained formula. This is due to the fact that the formula

n +∞
 r
e−(α+ β)t α β 2 
∑
1,ξ
lim pn (t) = r Ir (2 αβt) (43)
ξ →0 βt β r = n +1
α

(which is the one that is obtained from (42) as ν = 1 and α > β, as done in [14]) has no known equivalent in
the fractional case. It is also interesting to observe that in [8] another representation of the Laplace transform
of pνn (t) is given in formula 2.40, which is not easily invertible, but has been obtained by using (43) instead of
Sharma’s representation of p1n (t) ([2])


n
n +∞
α α α (αt)r k +r
( βt)m−1
p1n (t) = 1− + e−(α+ β)t ∑ ∑ (r − m ) .
β β β r =0 r! m =0 m!

5.3. Distribution of the Busy Period

Let Bν denote the duration of the busy period and Bν (t) = P( Bν ≤ t) be its probability distribution
function. Let us observe that, if we pose N ν (0) = 1, then the queue empties within t if and only if
a catastrophe occurs within t or otherwise the queue empties without catastrophes within t. Let us
remark that, if there is no occurrence of catastrophes, the queue behaves as a fractional M/M/1. Let us
define K ν as the duration of a busy period for a fractional M/M/1 queue without catastrophes, Ξν
the time of first occurrence of a catastrophe for a non empty queue and K ν (t) = P(K ν ≤ t) and
Ξν (t) = P(Ξν ≤ t) their probability distribution functions. Thus, we have

Bν (t) = Ξν (t) + (1 − Ξν (t))K ν (t). (44)

Remark 5. If we denote with bν (t), ξ ν (t) and kν (t) the probability density functions of Bν , Ξν and K ν , we have,
by deriving formula (44),
bν (t) = ξ ν (t)(1 − K ν (t)) + (1 − Ξν (t))kν (t),

which, for ν = 1, is formula (17) of [14].

By using formula (44), we can finally show:

Theorem 5. Let Bν be the duration of the busy period of a fractional M/M/1 queue with catastrophes and
Bν (t) = P( Bν ≤ t). Then,

+∞ +∞
Bν (t) = 1 − Eν (−ξtν ) ∑ ∑ Cn,m tν(n+2m−1) Eν,ν
n+2m
(n+2m−1)+1
(−(α + β)tν ), (45)
n =1 m =0

where Cn,m is given in (14).

Proof. Observe that, by formula (30), we have

Ξν (t) = c0ν (t) = 1 − Eν (−ξtν )

and by formula (13) we also have a closed form of K ν (t). Thus, by using formula (44), we obtain
Equation (45).

52
Mathematics 2018, 6, 159

5.4. Distribution of the Time of the First Occurrence of a Catastrophe

We already know that if the queue starts from a non-empty state, then the occurrence of the
catastrophes is a Mittag–Leffler distribution. However, we are interested in such distribution as the
queue starts being empty. To do that, we will need some auxiliary discrete processes.

Theorem 6. Let D ν be the time of first occurrence of a catastrophe as P( N ν (0) = 0) = 1 and let D ν (t) =
P(D ν ≤ t). Then,
+∞ +∞
D ν (t) = 1 − Cm,j tν(2m+ j−1) Eν,ν(2m+ j−1)+1 [−(α + β + ξ )tν ],
2m+ j
∑∑ (46)
j =1 m =0

where

j ( β + ξ ) j − α j 2m + j
Cm,j = (αβ)m .
2m + j β + ξ − α m
ν
Proof. Following the lines of [14], let us consider the process N (t) with state space {−1, 0, 1, 2, . . . }
ν ν
such that P( N (t) = 0) = 1 and posing rn (t) = P( N (t) = n), n ≥ −1 as its state probability, we have
⎧
⎪ C ν ν ν ν
⎪0 Dt r−1 (t) = ξ [1 − r−1 (t) − r0 (t)],
⎪
⎪
⎪
⎨C D ν r ν (t) = −αr ν (t) + βr ν (t),
0 t 0 0 1
(47)
⎪
⎪0C Dtν rnν (t) = −(α + β + ξ )rnν (t) + αrnν −1 (t) + βrnν +1 (t), n ≥ 1,
⎪
⎪
⎪ ν
⎩ rn (0) = δn,0 , n ≥ −1.

Let us remark that such process represents our queue until a catastrophe occurs: in such case,
instead of emptying the queue, the state of the process becomes −1, which is an absorbent state.
With such interpretation, we can easily observe that D ν (t) = r−
ν ( t ).
1
ν d 1
In order to determine rnν (t), we will first show that N (t) = N ( Lν (t)) where Lν (t) is the inverse
1 ν
of a ν-stable subordinator which is independent from N . To do that, let us consider N (t) + 1 instead
ν ν ν ν +∞ n ν
of N (t). Let us remark that P( N (t) + 1 = n) = rn−1 (t). Let G (z, t) = ∑n=0 z rn−1 (t) be the
ν
probability generating function of N (t) + 1. Multiplying the third sequence of equations in (47) with
zn+1 and then, summing all these equations, we have
( )
+∞ +∞ +∞ +∞
C ν
0 Dt ∑ zn rnν −1 (t) = −(α + β + ξ ) ∑ zn rnν −1 (t) + α ∑ zn rnν −2 (t) + β ∑ zn rnν (t). (48)
n =2 n =2 n =2 n =2

Now observe that

+∞ +∞
∑ zn rnν −1 (t) = ∑ zn rnν −1 (t) − r−ν 1 (t) − zr0ν (t) = Gν (z, t) − r−ν 1 (t) − zr0ν (t); (49)
n =2 n =0

moreover,

+∞ +∞ +∞
∑ zn rnν −2 (t) = ∑ zn+1 rnν −1 (t) = z ∑ zn rnν −1 (t)
n =2 n =1 n =1
= z[ G ν (z, t) − r−
ν ν ν ν 2 ν
1 ( t )] = z [ G ( z, t ) − r−1 ( t ) − zr0 ( t )] + z r0 ( t ); (50)

53
Mathematics 2018, 6, 159

finally,

+∞ +∞ +∞
1
∑ zn rnν (t) = ∑ zn−1 rnν −1 (t) = z ∑ zn rnν −1 (t)
n =2 n =3 n =3
1
= [ G ν (z, t) − r−
ν ν 2 ν
1 ( t ) − zr0 ( t ) − z r1 ( t )]
z
1
= [ G ν (z, t) − r− ν ν ν
1 ( t ) − zr0 ( t )] − zr1 ( t ). (51)
z
Using Equations (49),(50) and (51) in Equation (48), we obtain

C ν ν ν ν
0 Dt [ G ( z, t ) − r−1 ( t ) − zr0 ( t )]
 
β
= αz − (α + β + ξ ) + [ G ν (z, t) − r−
ν ν
1 ( t ) − zr0 ( t )]
z
+ αz2 r0ν (t) − βzr1ν (t). (52)

Finally, by using the first and the second equation of Equation (47) in Equation (52), we obtain
 
C ν ν β
0 Dt G ( z, t ) = αz − (α + β + ξ ) + [ G ν (z, t) − r−
ν ν
1 ( t ) − zr0 ( t )]
z
+ αz(z − 1)r0ν (t) + ξ [1 − r−
ν ν
1 ( t ) − r0 ( t )].

ν
We have obtained that the probability generating function G ν (z, t) of N (t) + 1 solves the
Cauchy problem
⎧ 
⎪ C ν ν ν ν ν
⎨z0 Dt G (z, t) = αz − (α + β + ξ )z + β [ G (z, t) − r−1 (t) − zr0 (t)]
2
⎪

⎪ +αz2 (z − 1)r0ν (t) + ξz[1 − r−

ν ν
1 ( t ) − r0 ( t )],
(53)
⎪
⎩ ν
G (z, 0) = z,

that, for ν = 1, becomes

⎧ 
⎪
⎪ d 1
⎨z dt G (z, t) = αz − (α + β + ξ )z + β [ G (z, t) − r−1 (t) − zr0 (t)]
2 1 1 1
⎪
+αz2 (z − 1)r01 (t) + ξz[1 − r− (54)
1 ( t ) − r0 ( t )],
1 1
⎪
⎪
⎪
⎩ G1 (z, 0) = z.

Let Gν (z, s), 

r0ν (s) and ν ( s ) be the Laplace transforms of G ν ( z, t ), r ν ( t ) and r ν ( t ) and let us
r− 1 0 −1
take the Laplace transform in Equation (53) to obtain

ν (z, s) − sν−1 z] = αz2 − (α + β + ξ )z + β [ G
z[sν G ν (z, s) − ν
r0ν (s)]
r− 1 ( s ) − z
  (55)
1
+αz2 (z − 1)r0ν (s) + ξz ν
− r− 1 (s) − r0ν (s) .
s

ν d 1
Now, let us remark that N (t) + 1 = N ( Lν (t)) + 1 if and only if for all n ≥ 0:
 +∞
rnν −1 (t) = rn1 −1 (y) P( Lν (t) ∈ dy) (56)
0

that is to say if and only if

 +∞
G ν (z, t) = G1 (z, y) P( Lν (t) ∈ dy).
0

54
Mathematics 2018, 6, 159

Taking Laplace transform and using Equation (22), we obtain

 +∞
ν
ν (z, s) = sν−1
G G1 (z, y)e−ys dy,
0
 +∞
ν ν −1 1 −ysν

r− 1 (s) = s r− 1 (y)e dy, (57)
0
 +∞
ν
r0ν (s) = sν−1
 r01 (y)e−ys dy.
0

Thus, by substituting the formulas (57) in (55), we obtain

  +∞ 
ν
z s ν s ν −1 G1 (z, y)e−ys dy − sν−1 z = [αz2 − (α + β + ξ )z + β]
0
  +∞  +∞  +∞

ν −1 ν −ysν ν
× s G1 (z, y)e−ys dy − sν−1 1
r− 1 (y)e dy − zsν−1 r01 (y)e−ys dy
0 0 0
 +∞
ν −1 ν
+ αz (z − 1)s
2
r01 (y)e−ys dy
0
  +∞  +∞ 
1 ν −1 −ysν ν −1 −ysν
+ ξz −s 1
r− 1 ( y ) e dy − s r 1
0 ( y ) e dy .
s 0 0

1
Finally, multiplying with s ν −1
, we have
  +∞ 
ν
z sν G1 (z, y)e−ys dy − z = [αz2 − (α + β + ξ )z + β]
0
 +∞  +∞  +∞ 
ν −ysν ν
× G1 (z, y)e−ys dy − 1
r− 1 (y)e dy − z r01 (y)e−ys dy
0 0 0
 +∞ (58)
ν
+ αz (z − 1)
2
r01 (y)e−ys dy
0
  +∞  +∞ 
1 −ysν ν
+ ξz ν − 1
r− 1 (y)e dy − r01 (y)e−ys dy .
s 0 0

ν d 1
Now we know that N (t) = N ( Lν (t)) if and only if Equation (58) is verified. For this reason,
 +∞ ν
we only need to show such equation. To do that, remarking that 0 e−ys dy = s1ν , consider the
right-hand side of Equation (58) and observe that
 +∞  +∞  +∞ 
ν ν ν
[αz2 − (α + β + ξ )z + β] G1 (z, y)e−ys dy − r−1
1 (y)e
−ys
dy − z r01 (y)e−ys dy
0 0 0
 +∞   +∞  +∞ 
ν 1 −ysν −ysν
+ αz2 (z − 1) r01 (y)e−ys dy + ξz ν − 1
r− 1 ( y ) e dy − r 1
0 ( y ) e dy
0 s 0 0
 +∞
= ([αz2 − (α + β + ξ )z + β][ G1 (z, y) − r−
1
1 ( y ) − zr0 ( y )]
1
0
−ys ν
+ αz (z − 1)r01 (y) + ξz[1 − r−
2 1
1 ( y ) − r0 ( y )]) e
1
dy.

55
Mathematics 2018, 6, 159

Thus, by using Equation (54), we have

 +∞  +∞  +∞ 
ν ν ν
[αz2 − (α + β + ξ )z + β] G1 (z, y)e−ys dy − r−1
1 (y)e
−ys
dy − z r01 (y)e−ys dy
0 0 0
 +∞   +∞  +∞ 
ν 1 −ysν −ysν
+ αz2 (z − 1) r01 (y)e−ys dy + ξz ν − 1
r− 1 ( y ) e dy − r 1
0 ( y ) e dy
0 s 0 0
 +∞

d 1 ν
=z G (z, y) e−ys dy
0 dt
 +∞ 
ν
=z G1 (z, y)e−ys dy − z ,
0

concluding the proof of our first claim.

From Theorem 3.1 of [14], we know that

+∞ +∞ Cm,j
1
r− 1 (t) = 1 − ∑∑ t2m+ j−1 e−(α+ β+ξ )t (59)
j =1 m =0
( 2m + j − 1) !

ν d 1
and, since we know that N (t) = N ( Lν (t)), we can use (59) in (56) with n = 0 to obtain:

+∞ +∞  +∞
Cm,j
ν
r− 1 (t) = 1 − ∑∑ y2m+ j−1 e−(α+ β+ξ )y P( Lν (t) ∈ dy). (60)
j =1 m =0
(2m + j − 1)! 0

Taking the Laplace transform in (60) and using formula (22), we obtain
 +∞
1 +∞ +∞ Cm,j ν
ν
r− 1 (s) = −∑ ∑ s ν −1 y2m+ j−1 e−(α+ β+ξ +s )y dy
s j=1 m=0 (2m + j − 1)! 0

and then integrate

ν 1 +∞ +∞ s ν −1
s j∑ ∑ Cm,j (α + β + ξ + sν )2m+ j .

r− 1 (s) = − (61)
=1 m =0

Finally, applying the inverse Laplace transform on Equation (61) and using formula (A2), we
complete the proof.

6. Conclusions
Our work focused on the transient behaviour of a fractional M/M/1 queue with catastrophes,
deriving formulas for the state probabilities, the distribution of the busy period and the distribution of
the time of the first occurrence of a catastrophe. This is a non-Markov generalization of the classical
M/M/1 queue with catastrophes, obtained through a time-change. The introduction of fractional
dynamics in the equations that master the behaviour of the queue led to a sort of transformation
of the time scale. Fractional derivatives are global operators, so the state probabilities preserve
memory of their past, eventually slowing down the entire dynamics. Indeed, we can see how
Mittag–Leffler functions take place where in the classical case we expected to see exponentials.
However, such fractional dynamic seems to affect only the transient behaviour, since we have shown
in Remark 2 that the limit behaviour is the same.
The main difficulty that is linked with fractional queues (or in general time-changed queues)
is the fact that one has to deal with non-local derivative operators, such as the Caputo derivative,
losing Markov property. However, fractional dynamics and fractional processes are gaining attention,
due to their wide range of applicability, from physics to finance, from computer science to biology.
Moreover, time-changed processes have formed a thriving field of application in mathematical finance.

56
Mathematics 2018, 6, 159

Future works will focus on an extension of such results to Ek /M/1 and M/Ek /1 queues, or even to a
generalization of fractional M/M/1 queue to a time-changed M/M/1 queue by using the inverse of
any subordinator.

Author Contributions: The three authors have participated equally in the development of this work. The paper
was also written and reviewed cooperatively.
Funding: This research was partially funded by GNCS and MANM. Nikolai Leonenko was supported in parts by
the project MTM2015–71839–P(cofounded with Feder funds), of the DGI, MINECO, and the Australian Research
Council’s Discovery Projects funding scheme (project number DP160101366).
Acknowledgments: We want to thank V. V. Ahn of QUT, Brisbane, Australia for his support.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A. Fractional Integrals and Derivatives

Let us recall the definition of fractional integral [7]. Given a function x : [t0 , t1 ] ⊆ R → R,
its fractional integral of order ν > 0 in [t0 , t] for t0 ≤ t ≤ t1 is given by:
 t
1
t0 I νt x = (t − τ )ν−1 x (τ )dτ.
Γ(ν) t0

The Riemann–Liouville fractional derivative operator is defined as:

RL ν dm m−ν
t 0 Dt = t I
dtm 0 t
while the Caputo fractional derivative operator is defined as:

C ν −ν dm
t 0 Dt = t0 I m
t
dtm
whenever m − 1 < ν < m. Obviously, such operators are linear. It is interesting to remark that

RL ν t−ν C ν
0 Dt 1 = , 0 Dt 1 = 0.
Γ (1 − ν )

Note that, for a function x (t), t ≥ 0 and ν ∈ (0, 1), the Caputo fractional derivative is defined as:
 t
C ν 1 d ds
0 Dt x = x (t − s) ν
Γ (1 − ν ) dt 0 s
ν x (0)
=0RL Dt x − ,
Γ (1 − ν ) t ν

where  t
RL ν 1 d ds
0 Dt x = x (t − s) ,
Γ(1 − ν) dt 0 sν
and for its Laplace transform, denoting by x(z) the Laplace transform of x,

L[0C Dtν x ](z) = zν x(z) − zν−1 x (0). (A1)

Moreover, for ν ∈ (0, 1), a well-posed fractional Cauchy problem with Riemann–Liouville
derivatives is given in the form ⎧
⎨ RL D ν x = f (t, x (t)),
t0 t
⎩ t0 It1−ν x = x0 ,
| t = t0

57
Mathematics 2018, 6, 159

in which the initial condition is given in terms of fractional integrals, while if we use Caputo derivatives
we have: 
C D ν x = f ( t, x ( t )),
t0 t
x ( t0 ) = x0 ,
in which the initial condition is related only with the initial value of the function. For such reason, we
will prefer adopting Caputo derivatives as fractional derivatives in this paper.
Finally, let us remark that the definition of fractional integral and derivative can be also considered
for functions x : [t0 , t1 ] ⊆ R → B where B is a Banach space and all the involved integrals are Bochner
integrals ([23]).

Appendix B. Some Special Functions

We recall the definitions of some special functions we use in such text.

Gamma funcion
The Gamma function is defined as:
 ∞
Γ(z) := tz−1 e−t dt.
0

In particular, we have Γ(z + 1) = zΓ(z) and, for z = n ∈ N, Γ(n + 1) = n!.

The modified Bessel function ([24]) of the first kind can be defined by its power series expansion as:

+∞
 x 2m+r
Ir ( x ) = ∑ 2
m!Γ(m + r + 1)
.
m =0

One-parameter Mittag–Leffler functions ([6]) are defined by their power series expansion as:

∞
zk
Eν (z) = ∑ Γ(νk + 1)
, ν > 0, z ∈ C.
k =0

Two-parameters Mittag–Leffler functions are also defined by their power series expansion as:

∞
zk
Eν,μ (z) = ∑ Γ ( νk + μ)
, ν > 0, μ > 0, z ∈ C.
k =0

Remark that Eν,1 (t) = Eν (t).

Generalized Mittag–Leffler functions are defined by their power series expansion as:

+∞
ρ (ρ)k zk
Eν,μ (z) = ∑ Γ ( νk + μ) k!
, ν > 0, μ > 0, ρ > 0, z ∈ C,
k =0

where (ρ)k is the Pochhammer symbol

(ρ)k = ρ(ρ + 1)(ρ + 2) · · · (ρ + k − 1).

An alternative way to define the Generalized Mittag–Leffler function is:

+∞
ρ zk Γ(ρ + k )
Eα,β (z) = ∑ k!Γ (αk + β)Γ(ρ)
, z ∈ C.
k =0

58
Mathematics 2018, 6, 159

1
Remark also that Eα,β = Eα,β . Functions with similar series expansions are also involved in
the study of the asymptotic behaviour of some integrals, which arise from a Feynman path integral
approach to some financial problems (see, i.e., [25] Section 4).
Recall also the following Laplace transform formula [9]

sνδ−γ
L[zγ−1 Eν,γ
δ
(wzν )](s) = , γ, ν, δ, w ∈ C, (γ), (ν), (δ) > 0, s ∈ C, |wsν | < 1. (A2)
(sν− w)δ

Finally let us remark, for ν ∈ (0, 1) that the Cauchy problem


C ν = λx,
0 Dt x
x (0) = x0 ,

admits as unique solution x (t) = x0 Eν (λtν ) ([6], p. 295).

References
1. Conolly, B.W. Lecture Notes on Queueing Systems; E. Horwood Limited: Ann Arbor, MI, USA, 1975;
ISBN 0470168579, 9780470168578.
2. Conolly, B.W.; Langaris, C. On a new formula for the transient state probabilities for M/M/1 queues and
computational implications. J. Appl. Probab. 1993, 30, 237–246. [CrossRef]
3. Kleinrock, L. Queueing Systems: Theory; Wiley: Hoboken, NJ, USA, 1975; ISBN 0471491101.
4. Lakatos, L.; Szeidl, L.; Telek, M. Introduction to Queueing Systems with Telecommunication Applications; Springer
Science & Business Media: Berlin, Germany, 2012; ISBN 978-1-4614-5316-1.
5. Parthasarathy, P.R. A transient solution to an M/M/1 queue: A simple approach. Adv. Appl. Probab. 1987, 19,
997–998, doi:10.2307/1427113. [CrossRef]
6. Kilbas, A.A.; Srivastava, H.M.; Trujillo, J.J. Theory and Applications of Fractional Differential Equations.
North-Holland Math. Stud. 2006, 204, 7–10.
7. Li, C.; Qian, D.; Chen, Y.Q. On Riemann-Liouville and Caputo derivatives. Discrete Dyn. Nat. Soc. 2011, 2011.
[CrossRef]
8. Cahoy, D.O.; Polito, F.; Phoha,V. Transient behavior of fractional queues and related processes.
Methodol. Comput. Appl. 2015, 17, 739–759. [CrossRef]
9. Haubold, H.J.; Mathai, A.M.; Saxena, R.K. Mittag–Leffler functions and their applications. J. Appl. Math.
2011, 2011. [CrossRef]
10. Meerschaert, M.M.; Nane, E.; Vellaisamy, P. The fractional Poisson process and the inverse stable subordinator.
Electron. J. Probab. 2011, 16, 1600–1620. [CrossRef]
11. Laskin, N. Fractional Poisson process. Commun. Nonlinear Sci. 2003, 8, 201–213. [CrossRef]
12. Meerschaert, M.M.; Straka, P. Inverse stable subordinators. Math. Model. Nat. Phenom. 2013, 8, 1–16.
[CrossRef] [PubMed]
13. Aletti, G.; Leonenko, N.; Merzbach, E. Fractional Poisson fields and martingales. J. Stat. Phys. 2018, 170,
700–730. [CrossRef]
14. Di Crescenzo, A.; Giorno, V.; NobileL, A.; Ricciardi, G.M. On the M/M/1 queue with catastrophes and its
continuous approximation. Queueing Syst. 2003, 43, 329–347.:1023261830362. [CrossRef]
15. Krishna Kumar, B.; Arivudainambi, D. Transient solution of an M/M/1 queue with catastrophes. Comput.
Math. Appl. 2000, 40, 1233–1240. [CrossRef]
16. Giorno, V.; Nobile, A.; Pirozzi, E. A state-dependent queueing system with asymptotic logarithmic
distribution. J. Math. Anal. Appl. 2018, 458, 949–966. [CrossRef]
17. Bingham, N.H. Limit theorems for occupation times of Markov processes. Zeitschrift für
Wahrscheinlichkeitstheorie und Verwandte Gebiete 1971, 17, 1–22. [CrossRef]
18. Kataria, K.K.; Vellaisamy, P. On densities of the product, quotient and power of independent subordinators.
J. Math. Anal. Appl. 2018, 462, 1627–1643. [CrossRef]
19. Leguesdron, P.; Pellaumail, J.; Sericola, B. Transient analysis of the M/M/1 queue. Adv. Appl. Probab. 1993,
25, 702–713. [CrossRef]

59
Mathematics 2018, 6, 159

20. Zhou, Y. Basic Theory of Fractional Differential Equations; World Scientific: Singapore, 2014; ISBN 978-981-3148-16-1.
21. Halmos, P.R.; Sunder, V.S. Bounded Integral Operators on L2 Spaces; Springer Science & Business Media: Berlin,
Germany, 2012; ISBN:978-3-642-67018-3.
22. Villani, A. Another note on the inclusion L p (μ) ⊂ Lq (μ). Am. Math. Mon. 1985, 92, 485–C76. [CrossRef]
23. Yosida, K. Functional Analysis; Springer: Berlin, Germany, 1978; ISBN 978-3-642-61859-8.
24. Korenev, B.G. Bessel Functions and Their Applications; CRC Press: Boca raton, FL, USA, 2003; ISBN 9780415281300.
25. Issaka, A.; SenGupta, I. Feynman path integrals and asymptotic expansions for transition probability
densities of some Lévy driven financial markets. J. Appl. Math. Comput. 2017, 54, 159–182. [CrossRef]

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

60
 mathematics
Article
An M [ X ] /G ( a, b)/1 Queueing System with Breakdown
and Repair, Stand-By Server, Multiple Vacation and
Control Policy on Request for Re-Service
G. Ayyappan and S. Karpagam *
Department of Mathematics, Pondicherry Engineering College, Puducherry 605014, India; [email protected]
* Correspondence: [email protected]

Received: 29 March 2018; Accepted: 29 May 2018; Published: 14 June 2018

Abstract: In this paper, we discuss a non-Markovian batch arrival general bulk service single-server
queueing system with server breakdown and repair, a stand-by server, multiple vacation and
re-service. The main server’s regular service time, re-service time, vacation time and stand-by
server’s service time are followed by general distributions and breakdown and repair times of the
main server with exponential distributions. There is a stand-by server which is employed during the
period in which the regular server remains under repair. The probability generating function of the
queue size at an arbitrary time and some performance measures of the system are derived. Extensive
numerical results are also illustrated.

Keywords: non-Markovian queue; general bulk service; multiple vacation; breakdown and repair;
stand-by server; re-service

Mathematics Subject Classification: 60K25; 90B22; 68M20

1. Introduction
Queueing systems with general bulk service and vacations have been studied by many researchers
because they deal with effective utilization of the server’s idle time for secondary jobs. Such queueing
systems have a wide range of application in many real-life situations such as production line systems,
inventory systems, digital communications and computer networks. Doshi [1] and Takagi [2] have
made a comprehensive survey of queueing systems with vacations. A batch arrival M[ X ] /G/1
queueing system with multiple vacations was first studied by Baba [3]. Krishna Reddy et al. [4] have
discussed an M[ X ] /G ( a, b)/1 model with an N-policy, multiple vacations and setup times. Jeyakumar
and Senthilnathan [5] analyzed the bulk service queueing system with multiple working vacations
and server breakdown.
The first work on re-service was done by Madan [6]. He consider an M/G /1 queueing model,
in which the server performs the first essential service for all arriving customers. As soon as the
first service is executed, they may leave the system with probability (1 − θ ), and the second optional
service is provided with θ. Madan et al. [7] considered a bulk arrival queue with optional re-service.
Jeyakumar and Arumuganathan [8] discussed a bulk queue with multiple vacation and a control
policy on request for re-service. Recently, Haridass and Arumuganathan [9] analyzed a batch service
queueing system with multiple vacations, setup times and server choice of admitting re-service.
No system is found to be perfect in the real world, since all the devices fail more or less
frequently. Thus, the random failures and systematic repair of components of a machining system
have a significant impact on the output and the productivity of the machining system. A detailed
survey on queues with interruptions was undertaken by Krishnamoorthy et al. [10]. Ayyappan and
Shyamala [11] derived the transient solution to an M[ X ] /G/1 queueing system with feedback, random

Mathematics 2018, 6, 101; doi:10.3390/math6060101 61 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 101

breakdowns, Bernoulli schedule server vacation and random setup time. An M/G/1 queue with two
phases of service subject to random breakdown and delayed repair was examined by Choudhury
and Tadj [12]. Senthilnathan and Jeyakumar [13] studied the behavior of the server breakdown
without interruption in an M[ X ] /G ( a, b)/1 queueing system with multiple vacations and closedown
time. An M/G/1 two-phase multi-optional retrial queue with Bernoulli feedback, non-persistent
customers and breakdown and repair was analyzed by Lakshmi and Ramanath [14]. Recently, a discrete
time queueing system with server breakdowns and changes in the repair times was investigated by
Atencia [15].
The operating machine may fail in some cases, but due to the standby machines of the queueing
machining system, it remains operative and continues to perform the assigned job. The provision
of stand-by and repairmen support to the queueing system maintains the smooth functioning of the
system. In the field of computer and communications systems, distribution and service systems,
production/manufacturing systems, etc., the applications of queueing models with standby support
may be noted.
This paper is organized as follows. A literature survey is given in Section 2. In Section 3, the
queuing problem is defined. The system equations are developed in Sections 4. The Probability
Generating Function (PGF) of the queue length distribution in the steady state is obtained in Section 5.
Various performance measures of the queuing system are derived in Section 6. A computational study
is illustrated in Section 7. Conclusions are given in Section 8.

2. Literature Survey
Various authors have analyzed queueing problems of server vacation with several combinations.
A batch arrival queue with a vacation time under a single vacation policy was analyzed by
Choudhury [16]. Jeyakumar and Arumuganathan [17] have discussed steady state analysis of an
M[ X ] ]/G/1 queue with two service modes and multiple vacation, in which they obtained PGF of
the queue size and some performance measures. Balasubramanian et al. [18] discussed steady state
analysis of a non-Markovian bulk queueing system with overloading and multiple vacations. Haridass
and Arumuganathan [19] discussed a batch arrival general bulk service queueing system with a variant
threshold policy for secondary jobs. Recently, Choudhury and Deka [20] discussed a batch arrival
queue with an unreliable server and delayed repair, with two phases of service and Bernoulli vacation
under multiple vacation policy.
Queueing systems, where the service discipline involves more than one service, have been
receiving much attention recently. They are said to have an additional service channel, or to
have feedback, or to have optional re-service, or to have two phases of heterogeneous service.
Madan [21] analyzed a queueing system with feedback. Madan [22], generalized his previous model
by incorporating server vacation. Medhi [23] discussed a single server Poisson input queue with
a second optional channel. Arumugananathan and Maliga [24] also examined a bulk queue with
re-service of the service station and setup time. Baruah et al. [25] studied a batch arrival queue
with two types of service, balking, re-service and vacation. Ayyappan and Sathiya [11] derived the
PGF of the non-Markovian queue with two types of service and optional re-service with a general
vacation distribution.
One can find an enormous amount of work done on queueing systems with breakdowns.
For some papers on random breakdowns in queueing systems, the reader may see Aissani et al. [26],
Maraghi et al. [27] and Fadhil et al. [28]. Rajadurai et al. [29] analyzed an M[ X ] /G/1 retrial queue with
two phases of service under Bernoulli vacation and random breakdown. Jiang et al. [30] have made a
computational analysis of a queue with working breakdown and delayed repair.
The operating system may fail in some cases, but due to stand-by machines, it remains operative
and continuous to perform the assigned job. Madan [31] studied the steady state behavior of a queuing
system with a stand-by server to serve customers only during the repair period. In that work, repair
times were assumed to follow an exponential distribution. Khalaf [32] examined the queueing system

62
Mathematics 2018, 6, 101

with four different main servers’ interruption and a stand-by server. Jain et al. [33] have made a
cost analysis of the machine repair problem with standby, working vacation and server breakdown.
Kumar et al. [34] discussed a bi-level control of a degraded machining system with two unreliable
servers, multiple standbys, startup and vacation. Murugeswari et al. [35] analyzed the bulk arrival
queueing model with a stand-by server and compulsory server vacation. Recently, we provided an
excellent survey on standby by Kolledath et al. [36].

3. Model Description
This paper deals with a queueing model whose arrival follows a compound Poisson process with
intensity rate λ. The main server and stand-by servers serve the customers under the general bulk
service rule. The general bulk service rule was first introduced by Neuts [37]. The general bulk service
rule states that the server will start to provide service only when at least ‘a’ units are present in the
queue, and the maximum service capacity is ‘b’ (b > a). On completion of a batch service, if less than
‘a’ customers are present in the queue, then the server has to wait until the queue length reaches the
value ‘a’. If less than or equal to ‘b’ and greater than or equal to ‘a’ customers are in the queue, then all
the existing customers are taken into service. If greater than or equal to ‘b’ customers are in the queue,
then ‘b’ customers are taken into service. The main server may breakdown at any time during regular
service with exponential rate α, and in such cases, the main server immediately goes for a repair, which
follows an exponential distribution with rate η, while the service to the current batch is interrupted.
Such a batch of customers is changed to the stand-by server, which starts service to that batch afresh.
The stand-by server remains in the system until the main server’s repair is completed. At the instant
of repair completion, if the stand-by server is busy, then the current batch of customers is exchanged to
the main server, which starts that batch service afresh. At the completion of a regular service (by the
main server), the leaving batch may request for a re-service with probability π. However, the re-service
is rendered only when the number of customers waiting in the queue is less than a. If no request
for re-service is made after the completion of a regular service and the number of customers in the
queue is less than a, then the server will avail itself of a vacation of a random length. The server
takes a sequence of vacations until the queue size reaches at least a. In addition, we assume that the
service time of the main server and stand-by server, re-service and vacation time of the main server are
independent of each other and follow a general (arbitrary) distribution.

Notations
Let X be the group size random variable of arrival, gk be the probability of ‘k’ customers arriving in
a batch and X (z) be its PGF. Sb (.), R(.), Ss (.) and V (.) represent the Cumulative Distribution Functions
(CDF) of the regular service and re-service time of the main server, the service time of the stand-by
server and the vacation time of the main server with corresponding probability density functions
of sb ( x ), r ( x ), ss ( x ) and v( x ), respectively. Sb0 (t), R0 (t), Ss0 (t) and V (0) (t) represent the remaining
regular service and re-service time of service given by the main server, the remaining service time of
service given by the stand-by server and the remaining vacation time of the main server at time ‘t’,
respectively. S̃b (θ ), R̃(θ ), S̃s (θ ) and Ṽ (θ ) represent the Laplace–Stieltjes Transform (LST) of Sb , R, Ss
and V, respectively.
For further development of the queueing system, let us define the following:
ε(t) = 1, 2, 3, 4, 5 and 6 at time t; the main server is in regular service, re-service and vacation, and at
time t, the stand-by server is in service and idle, respectively.
Z (t) = j, if the server is on the j-th vacation.
Ns (t) = number of customers in service station at time t.
Nq (t) = number of customers in the queue at time t.
Define the probabilities:

63
Mathematics 2018, 6, 101

Tn (t)Δt = Pr { Nq (t) = n, ε(t) = 5}, 0 ≤ n ≤ a − 1,

Pm,n ( x, t)Δt = Pr { Ns (t) = m, Nq (t) = n, x ≤ Sb0 (t) ≤ x + Δt, ε(t) = 1},
a ≤ m ≤ b, n ≥ 0,
Rn ( x, t)Δt = Pr { Ns (t) = m, Nq (t) = n, x ≤ R (t) ≤ x + Δt, ε(t) = 2},
0

a ≤ m ≤ b, n ≥ 0,
Bm,n ( x, t)Δt = Pr { Ns (t) = m, Nq (t) = n, x ≤ Ss0 (t) ≤ x + Δt, ε(t) = 4},
a ≤ m ≤ b, n ≥ 0,
Ql,j ( x, t)Δt = Pr { Z (t) = l, Nq (t) = j, x ≤ V 0 (t) ≤ x + Δt, ε(t) = 3},
l ≥ 1, j ≥ 0.

4. Queue Size Distribution

From the above-defined probabilities, we can easily construct the following steady state equations:

b
(λ + η ) T0 = ∑ Bm,0 (0), (1)
m= a
b n
(λ + η ) Tn = ∑ Bm,n (0) + ∑ Tn−k λgk , 1 ≤ n ≤ a − 1, (2)
m= a k =1
b  ∞

− Pi,0 ( x ) = −(λ + α) Pi,0 ( x ) + ∑ Pm,i (0)sb ( x ) + η
0
Bi,0 (y) dysb ( x )
m= a
∞
+ Ri (0)sb ( x ) + ∑ Ql,i (0)sb ( x ), a ≤ i ≤ b, (3)
l =1
 ∞ j

− Pi,j ( x ) = −(λ + α) Pi,j ( x ) + η
0
Bi,j (y) dysb ( x ) + ∑ Pi,j−k (x)λgk ,
k =1
j ≥ 1, a ≤ i ≤ b − 1, (4)
b  ∞

− Pb,j ( x ) = −(λ + α) Pb,j ( x ) + ∑ Pm,b+ j (0)sb ( x ) + η
0
Bb,j (y) dysb ( x )
m= a
∞ j
+ Rb+ j (0)sb ( x ) + ∑ Ql,b+ j (0)sb ( x ) + ∑ Pb,j−k (x)λgk , j ≥ 1, (5)
l =1 k =1
b  ∞

− Bi,0 ( x ) = −(λ + η ) Bi,0 ( x ) + ∑ Bm,i (0)ss ( x ) + α
0
Pi,0 (y) dyss ( x )
m= a
a −1
+ ∑ Tk λgi−k ss (x), a ≤ i ≤ b, (6)
k =0
 ∞ j

− Bi,j ( x ) = −(λ + η ) Bi,j ( x ) + α
0
Pi,j (y) dyss ( x ) + ∑ Bi,j−k (x)λgk ,
k =1
j ≥ 1, a ≤ i ≤ b − 1, (7)
b j

− Bb,j ( x ) = −(λ + η ) Bb,j ( x ) + ∑ Bm,b+ j (0)ss ( x ) + ∑ Bb,j−k (x)λgk
m= a k =1
 ∞ a −1
+α
0
Pb,j (y) dy ss ( x ) + ∑ Tk λgb+ j−k ss (x), j ≥ 1, (8)
k =0

64
Mathematics 2018, 6, 101

b

− R0 ( x ) = −λR0 ( x ) + π ∑ Pm,0 (0)r ( x ), (9)
m= a
b n

− Rn ( x ) = −λRn ( x ) + π ∑ Pm,n (0)r ( x ) + ∑ Rn−k (x)λgk , 1 ≤ n ≤ a − 1, (10)
m= a k =1
n

− Rn ( x ) = −λRn ( x ) + ∑ Rn−k (x)λgk , n ≥ a, (11)
k =1
b

− Q1,0 ( x ) = −λQ1,0 ( x ) + (1 − π ) ∑ Pm,0 (0)v( x ) + R0 (0)v( x ) + ηT0 v( x ), (12)
m= a
b

− Q1,n ( x ) = −λQ1,n ( x ) + (1 − π ) ∑ Pm,n (0)v( x ) + Rn (0)v( x ) + ηTn v( x )
m= a
n
+ ∑ Q1,n−k (x)λgk , 1 ≤ n ≤ a − 1, (13)
k =1
n

− Q1,n ( x ) = −λQ1,n ( x ) + ∑ Q1,n−k (x)λgk , n ≥ a, (14)
k =1

− Q j,0 ( x ) = −λQ j,0 ( x ) + Q j−1,0 (0)v( x ), j ≥ 2, (15)
n

− Q j,n ( x ) = −λQ j,n ( x ) + Q j−1,n (0)v( x ) + ∑ Q j,n−k (x)λgk , j ≥ 2,
k =1
1 ≤ n ≤ a − 1, (16)
n

− Q j,n ( x ) = −λQ j,n ( x ) + ∑ Q j,n−k (x)λgk , j ≥ 2, n ≥ a. (17)
k =1

Taking the LST on both sides of Equations (3)–(17), we get,

b  ∞
θ P̃i,0 (θ ) − Pi,0 (0) = (λ + α) P̃i,0 (θ ) − ∑ Pm,i (0)S̃b (θ ) − η
0
Bi,0 (y) dyS̃b (θ )
m= a
∞
− Ri (0)S̃b (θ ) − ∑ Ql,i (0)S̃b (θ ), a ≤ i ≤ b, (18)
l =1
 ∞ j
θ P̃i,j (θ ) − Pi,j (0) = (λ + α) P̃i,j (θ ) − η
0
Bi,j (y) dy S̃b (θ ) − ∑ P̃i,j−k (θ )λgk ,
k =1
a ≤ i ≤ b − 1, j ≥ 1, (19)
b  ∞
θ P̃b,j (θ ) − Pb,j (0) = (λ + α) P̃b,j (θ ) − ∑ Pm,b+ j (0)S̃b (θ ) − η
0
Bb,j (y) dyS̃b (θ )
m= a
∞ j
− Rb+ j (0)S̃b (θ ) − ∑ Ql,b+ j (0)S̃b (θ ) − ∑ P̃b,j−k (θ )λgk , j ≥ 1, (20)
l =1 k =1
b  ∞
θ B̃i,0 (θ ) − Bi,0 (0) = (λ + η ) B̃i,0 (θ ) − ∑ Bm,i (0)S̃s (θ ) − α
0
Pi,0 (y) dyS̃s (θ )
m= a
a −1
− ∑ Tk λgi−k S̃s (θ ), a ≤ i ≤ b, (21)
k =0
 ∞ j
θ B̃i,j (θ ) − Bi,j (0) = (λ + η ) B̃i,j (θ ) − α
0
Pi,j (y) dy S̃s (θ ) − ∑ B̃i,j−k (θ )λgk ,
k =1
j ≥ 1, a ≤ i ≤ b − 1, (22)

65
Mathematics 2018, 6, 101

b j
θ B̃b,j (θ ) − Bb,j (0) = (λ + η ) B̃b,j (θ ) − ∑ Bm,b+ j (0)S̃s (θ ) − ∑ B̃b,j−k (θ )λgk
m= a k =1
 ∞ a −1
−α
0
Pb,j (y) dy S̃s (θ ) − ∑ Tk λgb+ j−k S̃s (θ ), j ≥ 1, (23)
k =0
b
θ R̃0 (θ ) − R0 (0) = λ R̃0 (θ ) − π ∑ Pm,0 (0) R̃(θ ), (24)
m= a
b n
θ R̃n (θ ) − Rn (0) = λ R̃n (θ ) − π ∑ Pm,n (0) R̃(θ ) − ∑ R̃n−k (θ )λgk ,
m= a k =1
1 ≤ n ≤ a − 1, (25)
n
θ R̃n (θ ) − Rn (0) = λ R̃n (θ ) − ∑ R̃n−k (θ )λgk , n ≥ a, (26)
k =1
b
θ Q̃1,0 (θ ) − Q1,0 (0) = λ Q̃1,0 (θ ) − (1 − π ) ∑ Pm,0 (0)Ṽ (θ ) − R0 (0)Ṽ (θ )
m= a
− ηT0 Ṽ (θ ), (27)
b
θ Q̃1,n (θ ) − Q1,n (0) = λ Q̃1,n (θ ) − (1 − π ) ∑ Pm,n (0)Ṽ (θ ) − Rn (0)Ṽ (θ )
m= a
n
− ηTn Ṽ (θ ) − ∑ Q̃1,n−k (θ )λgk , 1 ≤ n ≤ a − 1, (28)
k =1
n
θ Q̃1,n (θ ) − Q1,n (0) = λ Q̃1,n (θ ) − ∑ Q̃1,n−k (θ )λgk , n ≥ a, (29)
k =1
θ Q̃ j,0 (θ ) − Q j,0 (0) = λ Q̃ j,0 (θ ) − Q j−1,0 (0)Ṽ (θ ), j ≥ 2, (30)
n
θ Q̃ j,n (θ ) − Q j,n (0) = λ Q̃ j,n (θ ) − Q j−1,n (0)Ṽ (θ ) − ∑ Q̃ j,n−k (θ )λgk , j ≥ 2,
k =1
1 ≤ n ≤ a − 1, (31)
n
θ Q̃ j,n (θ ) − Q j,n (0) = λ Q̃ j,n (θ ) − ∑ Q̃ j,n−k (θ )λgk , j ≥ 2, n ≥ a. (32)
k =1

To find the Probability Generating Function (PGF) for the queue size, we define the
following PGFs:
∞ ∞
P̃i (z, θ ) = ∑ P̃i,j (θ )z j , Pi (z, 0) = ∑ Pi,j (0)z j , a ≤ i ≤ b,
j =0 j =0
∞ ∞
R̃(z, θ ) = ∑ R̃ j (θ )z j , R(z, 0) = ∑ R j (0) z j ,
j =0 j =0
∞ ∞
B̃i (z, θ ) = ∑ B̃i,j (θ )z j , Bi (z, 0) = ∑ Bi,j (0)z j , a ≤ i ≤ b, (33)
j =0 j =0
∞ ∞
Q̃l (z, θ ) = ∑ Q̃l,j (θ )z j Ql (z, 0) = ∑ Ql,j (0)z j , l ≥ 1.
j =0 j =0

66
Mathematics 2018, 6, 101

By multiplying Equations (18)–(32) with suitable power of zn and summing over n (n = 0 to ∞) and
using Equation (33), we get:

 b ∞
(θ − u(z)) P̃i (z, θ ) = Pi (z, 0) − S̃b (θ ) ∑ Pm,i (0) + Ri (0) + ∑ Ql,i (0) + η B̃i (z, 0) ,
m= a j =0

a ≤ i ≤ b − 1, (34)
z (θ − u(z)) P̃b (z, θ ) = (z − S̃b (θ )) Pb (z, 0)
b b

 b −1 ∞
− S̃b (θ ) ∑ Pm (z, 0) + R(z, 0) + ∑ Ql (z, 0) + zb η B̃b (z, 0)
m= a l =1
b −1  b ∞ 
− ∑ ∑ Pm,j (0)z j + R j (0)z j + ∑ Ql,j (0)z j , (35)
j =0 m= a j =0
 b a −1
(θ − v(z)) B̃i (z, θ ) = Bi (z, 0) − S̃s (θ ) α P̃i (z, 0) + ∑ Bm,i (0) + ∑ Tk λgi−k ,
m= a k =0
a ≤ i ≤ b − 1, (36)
 b −1
zb (θ − v(z)) B̃b (z, θ ) = (zb − S̃s (θ )) Bb (z, 0) − S̃s (θ ) ∑ Bm (z, 0) + zb α P̃b (z, 0)
m= a
a −1 ∞ b −1 b
+λ ∑ ∑ Tk zk gj−k z j−k − ∑ ∑ Bm,j (0)z j , (37)
k =0 j = b j =0 m = a
a −1 b
(θ − w(z)) R̃(z, θ ) = R(z, 0) − π R̃(θ ) ∑ ∑ Pm,n (0)zn , (38)
n =0 m = a
a −1  b
(θ − w(z)) Q̃1 (z, θ ) = Q1 (z, 0) − Ṽ (θ ) ∑ (1 − π ) ∑ Pm,n (0)zn + Rn (0)zn
n =0 m= a

+ ηTn zn , (39)
a −1
(θ − w(z)) Q̃ j (z, θ ) = Q j (z, 0) − Ṽ (θ ) ∑ Q j−1,n (0)zn , j ≥ 2, (40)
n =0
where
u(z) = λ + α − λX (z), v(z) = λ + η − λX (z), w(z) = λ − λX (z).
Substitute θ = u(z) in (34) and (35), we get,
 b ∞
Pi (z, 0) = S̃b (u(z)) ∑ Pm,i (0) + Ri (0) + ∑ Ql,i (0) + η B̃i (z, 0) ,
m= a j =0

a ≤ i ≤ b − 1, (41)
S̃b (u(z))  b −1 ∞
Pb (z, 0) =
(zb − S̃b (u(z)))
∑ Pm (z, 0) + R(z, 0) + ∑ Ql (z, 0)
m= a l =1
b −1  b ∞ 
+ zb η B̃b (z, 0) − ∑ ∑ Pm,j (0)z j + R j (0)z j + ∑ Ql,j (0)z j , (42)
j =0 m= a j =0

67
Mathematics 2018, 6, 101

substitute θ = v(z) in (36) and (37), we get,

 b a −1
Bi (z, 0) = S̃s (v(z)) α P̃i (z, 0) + ∑ Bm,i (0) + ∑ Tk λgi−k , a ≤ i ≤ b − 1, (43)
m= a k =0

S̃s (v(z))  b −1
Bb (z, 0) = ∑ Bm (z, 0) + zb α P̃b (z, 0)
(zb − S̃s (v(z))) m=a
a −1 ∞ b −1 b
+λ ∑ ∑ Tk zk gj−k z j−k − ∑ ∑ Bm,j (0)z j , (44)
k =0 j = b j =0 m = a

substitute θ = w(z) in (38) to (40), we get

a −1 b
R(z, 0) = π R̃(w(z)) ∑ ∑ Pm,n (0)zn , (45)
n =0 m = a
a −1  b
Q1 (z, 0) = Ṽ (w(z)) ∑ (1 − π ) ∑ Pm,n (0)zn + Rn (0)zn + ηTn zn , (46)
n =0 m= a
a −1
Q j (z, 0) = Ṽ (w(z)) ∑ Q j−1,n (0)zn , j ≥ 2. (47)
n =0

Substitute Equations (41)–(47) in Equations (34)–(40) after simplification, and we get,

 b ∞
(θ − u(z)) P̃i (z, θ ) = (S̃b (u(z)) − S̃b (θ )) ∑ Pm,i (0) + Ri (0) + ∑ Ql,i (0)
m= a j =0

+ η B̃i (z, 0) , a ≤ i ≤ b − 1, (48)

(S̃ (u(z)) − S̃b (θ ))  b−1 ∞

(θ − u(z)) P̃b (z, θ ) = b b
(z − S̃b (u(z))) m=a
∑ Pm (z, 0) + R(z, 0) + ∑ Ql (z, 0)
l =1
b −1  b ∞ 
+ zb η B̃b (z, 0) − ∑ ∑ Pm,j (0)z j + R j (0)z j + ∑ Ql,j (0)z j , (49)
j =0 m= a j =0

 b a −1
(θ − v(z)) B̃i (z, θ ) = (S̃s (v(z)) − S̃s (θ )) α P̃i (z, 0) + ∑ Bm,i (0) + ∑ Tk λgi−k ,
m= a k =0
a ≤ i ≤ b − 1, (50)
(S̃s (v(z)) − S̃s (θ ))  b−1
(θ − v(z)) B̃b (z, θ ) =
(zb − S̃s (v(z))) m=a
∑ Bm (z, 0) + zb α P̃b (z, 0)
a −1 ∞ b −1 b
+ λ ∑ ∑ Tk zk g j−k z j−k − ∑ ∑ Bm,j (0)z j , (51)
k =0 j = b j =0 m = a
a −1 b
(θ − w(z)) R̃(z, θ ) = ( R̃(w(z)) − R̃(θ ))π ∑ ∑ Pm,n (0)zn , (52)
n =0 m = a
a −1  b
(θ − w(z)) Q̃1 (z, θ ) = (Ṽ (w(z)) − Ṽ (θ )) ∑ (1 − π ) ∑ Pm,n (0)zn
n =0 m= a

+ Rn (0)zn + ηTn zn , (53)
a −1
(θ − w(z)) Q̃ j (z, θ ) = (Ṽ (w(z)) − Ṽ (θ )) ∑ Q j−1,n (0)zn , j ≥ 2. (54)
n =0

68
Mathematics 2018, 6, 101

5. Probability Generating Function of the Queue Size

5.1. The PGF of the Queue Size at an Arbitrary Time Epoch

Let P(z) be the PGF of the queue size at an arbitrary time epoch. Then,

b b ∞
P(z) = ∑ P̃i (z, 0) + ∑ B̃i (z, 0) + R̃(z, 0) + ∑ Q̃l (z, 0) + T (z). (55)
i=a i=a l =1

By substituting θ = 0 in Equations (48)–(54), then Equation (55) becomes:

b −1 a −1
K1 ( z ) ∑ (zb − zi )ci + (1 − Ṽ (w(z)))K3 (z) ∑ cn zn
i=a n =0
b −1 a −1
+ K2 ( z ) ∑ (zb − zi )di + (Ṽ (w(z)) − R̃(w(z)))K3 (z) ∑ π pn zn (56)
i=a n =0
 a −1
+ η [Y1 (z) − Ṽ (w(z))K3 (z)] − v(z)K2 (z) + w(z)Y1 (z) ∑ Tk zk
k =0
P(z) =
w(z)Y1 (z)

where pi = ∑bm= a Pm,i (0), vi = ∑∞ b

l =1 Ql,i (0), qi = ∑m= a Bm,i (0), Ri (0) = ri , ci = pi + vi + ri and
a −1
di = qi + ∑k=0 Tk λgi−k and the expressions for K1 (z), K2 (z), K3 (z) and Y1 (z) are defined in
Appendix A.

5.2. Steady State Condition

The probability generating function has to satisfy P(1) = 1. In order to satisfy this condition,
applying L’Hopital’s rule and evaluating lim P(z), then equating the expression to one, we have,
z →1
H = (−λX1 ) F1 , where the expressions H and F1 are defined in Appendix B.
Since pi , ci , di and Ti are probabilities of  i customers being in the queue, it follows that H must
be positive. Thus, P(1) = 1 is satisfied iff (−λX1 ) F1 > 0. If:

λX1 (α + η )(1 − S̃b (α))(1 − S̃s (η ))

ρ=
bαη [S̃b (α)(1 − S̃s (η )) + S̃s (η )(1 − S̃b (α))]
then ρ < 1 is the condition for the existence of the steady state for the model under consideration.

5.3. Computational Aspects

Equation (56) has 2b + a unknowns c0 , c1 , ..., cb−1 , d a , ..., db−1 , p0 , p1 , ..., p a−1 and T0 , T1 , ..., Ta−1 .
Now, Equation (56) gives the PGF of the number of customers involving only 2b + a unknowns.
We can express ci (0 ≤ i ≤ a − 1) in terms of pi and Ti in such a way that the numerator has only 2b
constants. Now, Equation (56) gives the PGF of the number of customers involving only 2b unknowns.
By Rouche’s theorem, it can be proven that Y1 (z) has 2b − 1 zeros inside and one on the unit circle
|z| = 1. Since P(z) is analytic within and on the unit circle, the numerator must vanish at these
points, which gives 2b equations in 2b unknowns. We can solve these equations by any suitable
numerical technique.

5.4. Result 1
The probability that n(0 ≤ n ≤ a − 1) customers are in queue during the main server’s re-service
completion rn can be expressed as the probability of n customers in the queue during the main server’s
regular busy period pn as,

69
Mathematics 2018, 6, 101

n
rn = π ∑ γn − k p k , n = 0, 1, 2..., a − 1 (57)
k =0

where γn are the probabilities of n customers arriving during the main server’s re-service time.

5.5. Result 2
The probability that n(0 ≤ n ≤ a − 1) customers are in queue cn can be expressed as the sum of
the probability of n customers in the queue during the main server’s busy period and the stand-by
server’s idle time pn and Tn as,
n
∑ τn−k pk + τn−k Tk , 0 ≤ n ≤ a − 1.
(1) (2)
cn = (58)
k =0

where:
(1)
(1) π (γn − β n ) + ∑nk=1 β k τn−k
τn =
1 − β0
(2)
(2) ηβ n + ∑nk=1 β k τn−k
τn = (59)
1 − β0

γn , β n are the probabilities of n customers arriving during the main server’s re-service and vacation
time, respectively.

5.6. Particular Case

Case 1:
When there is no breakdown and re-service, then Equation (65) reduces to:

(S̃b (w(z)) − 1) ∑ib=−a1 (zb − zi )ci + (zb − 1)(Ṽ (w(z)) − 1) ∑na− 1

=0 c n z
n
P(z) = (60)
(−w(z))(z − S̃b (w(z)))
b

which coincides with the PGF of Senthilnathan et al. [19] without closedown.
Case 2:
When there is no breakdown, then Equation (65) reduces to:

b −1 a −1
(1 − S̃b (w(z))) ∑ (zb − zi )ci + (zb − 1)(1 − Ṽ (w(z))) ∑ cn zn
i=a n =0
a −1 (61)
+ (zb − 1)(Ṽ (w(z)) − R̃(w(z))) ∑ π pn zn
n =0
P(z) =
(w(z))(zb − S̃b (w(z)))

which is the PGF of Jeyakumar et al. [38].

5.7. PGF of the Queue Size in Various Epochs

5.7.1. PGF of the Queue Size in the Main Server’s Service Completion Epoch
The probability generating function of the main server’s service completion epoch M (z) is
obtained from Equations (48) and (49):

70
Mathematics 2018, 6, 101

  b −1 a −1
(1 − S̃b (u(z))) v(z)(zb − S̃s (v(z))) ∑ (zb − zi )ci − ∑ ck zk
i=a k =0
a −1 a −1 
+ π R̃(w(z)) ∑ pn zn + Ṽ (w(z)) ∑ ((1 − π ) pn + rn + ηTn + vn )zn (62)
n =0 n =0
 b −1 a −1 
∑ (zb − zi )(qi + ri ∑ Tk λgi−k )
(2)
+ zb η (1 − S̃s (v(z))) + − v(z) T (z)
i=a k =0
M(z) =
Y1 (z)

5.7.2. PGF of the Queue Size in the Vacation Completion Epoch

The PGF of the main server’s vacation completion epoch V (z) is obtained from Equations (53)
and (54); we get,

(1 − Ṽ (w(z))) ∑na−
=0 ((1 − π ) pn + rn + ηTn + vn ) z
1 n
V (z) = (63)
w(z)

5.7.3. PGF of the Queue Size in the Main Server’s Re-Service Completion Epoch
The PGF of the main server’s re-service completion epoch R(z) is obtained from Equation (52);
we get,

(1 − R̃(w(z))) ∑na− 1
=0 π p n z
n
R(z) = (64)
w(z)

5.7.4. PGF of the Queue Size in the Stand-by Server’s Service Completion Epoch
The probability generating function of the stand-by server’s service completion epoch N (z) is
derived from Equations (50) and (51); we get,

  b −1 a −1
(1 − S̃s (v(z))) zb α(1 − S̃b (u(z))) ∑ (zb − zi )ci − ∑ ck zk
i=a k =0
a −1 a −1 
+ π R̃(w(z)) ∑ n
pn z + Ṽ (w(z)) ∑ ((1 − π ) pn + rn + ηTn + vn )zn (65)
n =0 n =0
 b −1 a −1 
+ u(z)(zb − S̃b (u(z))) ∑ (zb − zi )(qi + ∑ Tk λgi−k ) − v(z)T (z)
i=a k =0
N (z) =
Y1 (z)

6. Some Performance Measures

6.1. The Main Server’s Expected Length of Idle Period

Let K be the random variable denoting the ‘idle period due to multiple vacation processes’.
Let Y be the random variable defined by:

0 if the server finds at least ‘a’ customers after the first vacation
Y=
1 if the server finds less than ‘a’ customers after the first vacation

Now,

E(K ) = E(K/Y = 0) P(Y = 0) + E(K/Y = 1) P(Y = 1)

= E(V ) P(Y = 0) + ( E(V ) + E(K )) P(Y = 1),

71
Mathematics 2018, 6, 101

Solving for E(K ), we get:

E (V ) E (V )
E(K ) = =   
(1 − P(Y = 1)) a −1 n
1 − ∑n=0 ∑i=0 β i [ pn + rn + ηTn ]

6.2. Expected Queue Length

The mean number of customers waiting in the queue E( Q) in an arbitrary time epoch is obtained
by differentiating P(z) at z = 1 and is given by:

 b −1
f 1 ( X, Sb , Ss ) ∑ [b(b − 1) − i(i − 1)]ci
i=a
 b −1
+ f 1 ( X, Sb , Ss ) ∑ (b(b − 1) − i(i − 1))di
i=a
 b −1 b −1
+ f 2 ( X, Sb , Ss ) ∑ ( b − i ) ci + f 3 ( X, Sb , Ss ) ∑ ( b − i ) di
i=a i=a
a −1 a −1
+ f 4 ( X, Sb , Ss , V ) ∑ ncn + f5 (X, Sb , Ss , V ) ∑ cn
n =0 n =0
a −1 a −1
+ f 6 ( X, Sb , Ss , R, V ) ∑ πnpn + f7 (X, Sb , Ss , R, V ) ∑ π pn
n =0 n =0
a −1 a −1
+ f 8 ( X, Sb , Ss , V ) ∑ nTn + f9 (X, Sb , Ss , V ) ∑ Tn
n =0 n =0
E( Q) = , (66)
3( F12 )2

the expressions for f i (i = 1, 2, ..., 9) are defined in Appendix B.

6.3. Expected Waiting Time

The expected waiting time is obtained using Little’s formula as:

E( Q)
E (W ) = (67)
λE( X )

where E( Q) is given in Equation (66).

7. Numerical Example
A numerical example of our model is analyzed for a particular case with the following assumptions:

1. The batch size distribution of the arrival is geometric with mean 2.

2. Take a = 5 and b = 8, and the service time distribution is Erlang-2 (both servers).
3. The vacation and re-service time of the main server follow an exponential distribution with
parameter ω = 5,  = 3, respectively.
4. Let m1 be the service rate for the main server.
5. Let m2 be the service rate for the stand-by server.

The unknown probabilities of the queue size distribution are computed using numerical
techniques. The zeros of the function Y1 (z) are obtained (see Figure 1) , and simultaneous equations
are solved by using MATLAB. The values which are satisfies the stability condition (see Figure 2) are
used for calculating the table values.
The expected queue length E( Q) and the expected waiting time E(W ) are calculated for various
arrival rate sand service rates, and the results are tabulated.
From Tables 1–4, the following observations can be made.

72
Mathematics 2018, 6, 101

1. As arrival rate λ increases, the expected queue size and expected waiting time are also increase.
2. When the main server’s and stand-by server’s service rate increases, the expected queue size and
expected waiting time decrease.
3. When the main server’s vacation rate increases, the expected queue size increases.

Table 1. Arrival rate vs. expected queue length and expected waiting time for the values
m1 = 10, m2 = 9.5, α = 1, η = 2, π = 0.3,  = 3, and ω = 5.

λ ρ E( Q) E (W )
5.00 0.131407 8.657374 0.865737
5.25 0.137978 9.539724 0.908545
5.50 0.144548 10.375816 0.943256
5.75 0.151119 11.149076 0.969485
6.00 0.157689 11.843026 0.986919
6.25 0.164259 12.441064 0.995285
6.50 0.170830 12.927026 0.994387
6.75 0.177400 13.285663 0.984123
7.00 0.183970 13.502456 0.964461
7.25 0.190541 13.563622 0.935422
7.50 0.197111 13.456834 0.897122

Table 2. Main server’s service rate vs. expected queue length and expected waiting time for the
values λ = 5, m2 = 5, α = 1, η = 2, π = 0.3,  = 3, and ω = 5.

m1 ρ E( Q) E (W )
5.25 0.257410 26.824821 2.682482
5.50 0.248885 26.082729 2.608273
5.75 0.240905 25.388555 2.538855
6.00 0.233420 24.735574 2.473557
6.25 0.226385 24.117976 2.411798
6.50 0.219761 23.531054 2.353105
6.75 0.213513 22.971202 2.297120
7.00 0.207610 22.435611 2.243561
7.25 0.202024 21.921511 2.192151
7.50 0.196731 21.427009 2.142701
7.75 0.191707 20.950354 2.095035
8.00 0.186934 20.490051 2.049005

Table 3. Stand-by server’s service rate vs. expected queue length and expected waiting time for the
values λ = 5, m1 = 10, α = 1, η = 2, π = 0.3,  = 3, and ω = 5.

m2 ρ E( Q) E (W )
4.0 0.260103 65.007246 4.062953
4.5 0.254645 59.415276 3.713455
5.0 0.249401 53.808260 3.363016
5.5 0.244361 48.320593 3.020037
6.0 0.239515 43.028238 2.689265
6.5 0.234853 37.970425 2.373152
7.0 0.230366 33.162455 2.072653
7.5 0.226045 28.605406 1.787838
8.0 0.22188 24.292435 1.518277
8.5 0.217865 20.212075 1.263255
9.0 0.213991 16.351005 1.021938

73
Mathematics 2018, 6, 101

Table 4. The effect of the main server’s vacation rate on expected queue length for the values
λ = 5, m1 = 10, m2 = 9.5, α = 1, η = 2, π = 0.3, and  = 3.

ω Erlang Exponential
5.00 8.657374 8.279153
5.25 8.808004 8.448114
5.50 8.950939 8.607757
5.75 9.086640 8.758748
6.00 9.215535 8.901685
6.25 9.338068 9.037158
6.50 9.454625 9.165674
6.75 9.565595 9.287730
7.00 9.671330 9.403767
7.25 9.772165 9.514200
7.50 9.868410 9.619407
7.75 9.960351 9.719734
8.00 10.048247 9.815494

Figure 1. MATLAB code for finding the roots.

74
Mathematics 2018, 6, 101

Figure 2. MATLAB code for finding the rho value.

8. Conclusions
In this paper, a batch arrival general bulk service queueing system with breakdown and
repair, stand-by server, multiple vacation and control policy on request for re-service is analyzed.
The probability generating function of the queue size distribution at an arbitrary time is obtained.
Some performance measures are calculated. The particular cases of the model are also deduced.
From the numerical results, it is observed that when the arrival rate increases, the expected queue
length and waiting time of the customers are also increase; if the service rate increases (for both
server’s), then the expected queue length and expected waiting time decrease. It is also observed that,
if the main server’s vacation rate increases, then the expected queue length increases.
Author Contributions: G.A.: To describe the model. S.K.: Convert the theoretical model into mathematical model
and solving.
Conflicts of Interest: There is no conflict of interest by the author to publish this paper.

Appendix A
The expressions used in Equation (56) are defined as follows:

K1 (z) = w(z)(1 − S̃b (u(z))) A1 (z),

K2 (z) = w(z)(1 − S̃s (v(z))) A2 (z),
K3 (z) = Y1 (z) − w(z) A1 (z)(1 − S̃b (u(z))),
where
A1 (z) = v(z)(zb − S̃s (v(z))) + zb α(1 − S̃s (v(z)))
A2 (z) = u(z)(zb − S̃b (u(z))) + zb η (1 − S̃b (u(z)))
Y1 (z) = u(z)v(z)(zb − S̃b (u(z)))(zb − S̃s (v(z))) − z2b αη (1 − S̃b (u(z)))(1 − S̃s (v(z))).

75
Mathematics 2018, 6, 101

Appendix B
The expressions for f i ’s in (66) are defined as follows:

f 1 ( X, Sb , Ss ) = 3E1 F12 ,
f 2 ( X, Sb , Ss ) = 3F7 F12 − 2E1 F13 ,
f 3 ( X, Sb , Ss ) = 3F9 F12 − 2E1 F13 ,
f 4 ( X, Sb , Ss , V ) = 6bαηV1 E2 F12 ,
f 5 ( X, Sb , Ss , V ) = 3F12 [bαηE2 V2 + V1 ( F7 − F2 )] − 2bαηE2 V1 F13 ,
f 6 ( X, Sb , Ss , R1 , V ) = 6bαη ( R1 − V1 ) E2 F12 ,
f 7 ( X, Sb , Ss , R1 , V ) = 3F12 [bαηE2 ( R2 − V2 ) + ( R1 − V1 )( F7 − F2 )] − 2bαηE2 ( R1 + V1 ) F13 ,
f 8 ( X, Sb , Ss , V ) = 6F11 F12 ,
f 9 ( X, Sb , Ss , V ) = 3E10 F12 − 2F11 F13 ,

where:

E1 = −λX1 (α + η )(1 − S̃b (α))(1 − S̃s (η )),

E2 = S̃b (α)(S̃s (η ) − 1) + S̃s (η )(S̃b (α) − 1),
E3 = Sb1 (1 − S̃s (η )) + Ss1 (1 − S̃b (α)),
E4 = −λX1 (1 − S̃s (η )),
E5 = −λX2 (1 − S̃s (η )) + 2λX1 Ss1 ,
E6 = −λX3 (1 − S̃s (η )) + 3λX2 Ss1 + 3λX1 Ss2 ,
E7 = −λX1 (1 − S̃b (α)),
E8 = −λX2 (1 − S̃b (α)) + 2λX1 Sb1 ,
E9 = −λX3 (1 − S̃b (α)) + 3λX2 Sb1 + 3λX1 Sb2 ,
E10 = η [bαηE2 V2 − V1 ( F2 − F7 )] + λX2 ( E1 − F1 ) + λX1 ( F9 − F2 ) − (η/3)( F8 + F10 ),
F1 = E1 − bαηE2 ,
F2 = (1 − S̃b (α))(1 − S̃s (η ))[2(λX1 )2 − λX2 (α + η ) − 2b(2b − 1)αη ]
+ [b(b − 1)αη − 2bλX1 (α + η )][(1 − S̃b (α)) + (1 − S̃s (η ))]
+ 2[2bαη + λX1 (α + η )] E3 + 2bαη (b − Sb1 − Ss1 )
F3 = (α + η )(b − Ss1 ) + E4 − bαS̃s (η ),
F4 = (α + η )(b(b − 1) − Ss2 ) − bα[(b − 1)S̃s (η ) + 2Ss1 ] − 2bλX1 + E5 ,
F5 = (α + η )(b − Sb1 ) + E7 − bη S̃b (α),
F6 = (α + η )(b(b − 1) − Sb2 ) − bη [(b − 1)S̃b (α) + 2Sb1 ] − 2bλX1 + E8 ,
F7 = E8 (α + η )(1 − S̃s (η )) + 2E7 F3 ,
F8 = E9 (α + η )(1 − S̃s (η )) + 3E8 F3 + 3E7 F4 ,
F9 = E5 (α + η )(1 − S̃b (α)) + 2E4 F5 ,
F10 = E6 (α + η )(1 − S̃b (α)) + 3E5 F5 + 3E4 F6 ,
F11 = bαηE2 (ηV1 + λX1 ) + ηλX1 [(α + η ) E3 − (1 − S̃b (α)) F3 + (1 − S̃s (η )) F5 ],
F12 = −2λX1 F1 ,
F13 = −3[λX1 F2 + λX2 F1 ].

2E1 ∑ib=−a1 (b − i )ci + 2E1 ∑ib=−a1 (b − i )di + 2bαηE2 V1 ∑na− 1
=0 c n
H=
+2bαηE2 ( R1 + V1 ) ∑na− 1 a −1
=0 π pn + 2F11 ∑n=0 Tn ,

76
Mathematics 2018, 6, 101

where:
  
Sb1 = −λX1 S̃b (α), Sb2 = S̃b (α)(−λX1 )2 − λX2 S̃b (α),
  
Ss1 = −λX1 S̃s (η ), Ss2 = S̃s (η )(−λX1 )2 − λX2 S̃s (η ),
R1 = λX1 E( R), R2 = λX2 E( R) + λ2 X12 E( R2 )
V1 = λX1 E(V ), V2 = λX2 E(V ) + λ2 X12 E(V 2 ), X1 = E( X ), X2 = E( X 2 ).

References
1. Doshi, B.T. Queueing systems with vacations—A survey. Queueing Syst. 1986, 1, 29–66. [CrossRef]
2. Takagi, H. Vacation and Priority Systems. Queuing Analysis: A Foundation of Performance Evaluation;
North-Holland: Amsterdam, The Netherlands, 1991; Volume I.
3. Baba, Y. On the M[ X ] /G/1 queue with vacation time. Oper. Res. Lett. 1986, 5, 93–98. [CrossRef]
4. Krishna Reddy, G.V.; Nadarajan, R.; Arumuganathan, R. Analysis of a bulk queue with N policy multiple
vacations and setup times. Comput. Oper. Res. 1998, 25, 957–967. [CrossRef]
5. Jeyakumar, S.; Senthilnathan, B. Modelling and analysis of a bulk service queueing model with multiple
working vacations and server breakdown. RAIRO-Oper. Res. 2017, 51, 485–508. [CrossRef]
6. Madan, K.C. An M/G/1 queue with second optional service. Queueing Syst. 2000, 34, 37–46. [CrossRef]
7. Madan, K.C. On M[ X ] /( G1, G2)/1 queue with optional re-service. Appl. Math. Comput. 2004, 152, 71–88.
8. Jeyakumar, S.; Arumuganathan, R. A Non-Markovian Bulk Queue with Multiple Vacations and Control
Policy on Request for Re-Service. Qual. Technol. Quant. Manag. 2011, 8, 253–269. [CrossRef]
9. Haridass, M.; Arumuganathan, R. A batch service queueing system with multiple vacations, setup time and
server’s choice of admitting re-service. Int. J. Oper. Res. 2012, 14, 156–186. [CrossRef]
10. Krishnamoorthy, A.; Pramod, P.K.; Chakravarthy, S.R. Queues with interruptions: A survey. Oper. Res.
Decis. Theory 2014, 22, 290–320. [CrossRef]
11. Ayyappan, G.; Shyamala, S. Transient solution of an M[ X ] /G/1 queueing model with feedback, random
breakdowns, Bernoulli schedule server vacation and random setup time. Int. J. Oper. Res. 2016, 25, 196–211.
[CrossRef]
12. Choudhury, G.; Tadj, L. An M/G/1 queue with two phases of service subject to the server breakdown and
delayed repair. Appl. Math. Model. 2009, 33, 2699–2709. [CrossRef]
13. Senthilnathan, B.; Jeyakumar, S. A study on the behaviour of the server breakdown without interruption in a
M[ X ] /G ( a, b)/1 queueing system with multiple vacations and closedown time. Appl. Math. Comput. 2012,
219, 2618–2633.
14. Lakshmi, K.; Kasturi Ramanath, K. An M/G/1 two phase multi-optional retrial queue with Bernoulli
feedback, non-persistent customers and breakdown and repair. Int. J. Oper. Res. 2014, 19, 78–95. [CrossRef]
15. Atencia, I. A discrete-time queueing system with server breakdowns and changes in the repair times.
Ann. Oper. Res. 2015, 235, 37–49. [CrossRef]
16. Choudhury, G. A batch arrival queue with a vacation time under single vacation policy. Comput. Oper. Res.
2002, 29, 1941–1955. [CrossRef]
17. Jeyakumar, S.; Arumuganathan, R. Analysis of Single Server Retrial Queue with Batch Arrivals, Two Phases
of Heterogeneous Service and Multiple Vacations with N-Policy. Int. J. Oper. Res. 2008, 5, 213–224.
18. Balasubramanian, M.; Arumuganathan. R.; Senthil Vadivu. A. Steady state analysis of a non-Markovian
bulk queueing system with overloading and multiple vacations. Int. J. Oper. Res. 2010, 9, 82–103. [CrossRef]
19. Haridass, M.; Arumuganathan, R. Analysis of a batch arrival general bulk service queueing system with
variant threshold policy for secondary jobs. Int. J. Math. Oper. Res. 2011, 3, 56–77. [CrossRef]
20. Choudhury, G.; Deka, M. A batch arrival unreliable server delaying repair queue with two phases of service
and Bernoulli vacation under multiple vacation policy. Qual. Technol. Quant. Manag. 2018, 15, 157–186.
[CrossRef]
21. Madan, K.C. A cyclic queueing system with three servers and optional two-way feedback.
Microelectron. Reliabil. 1988, 28, 873–875. [CrossRef]
22. Madan, K.C. On a single server queue with two-stage heterogeneous service and deterministic server
vacations. Int. J. Syst. Sci. 2001, 32, 837–844. [CrossRef]

77
Mathematics 2018, 6, 101

23. Medhi, J. A Single Server Poisson Input Queue with a Second Optional Channel. Queueing Syst. 2002,
42, 239–242. [CrossRef]
24. Arumuganathan, R.; Malliga, T.J. Analysis of a bulk queue with repair of service station and setup time.
Int. J. Can. Appl. Math. Quart. 2006, 13, 19–42.
25. Baruah, M.; Madan, K.C.; Eldabi, T. A Two Stage Batch Arrival Queue with Reneging during Vacation and
Breakdown Periods. Am. J. Oper. Res. 2013, 3, 570–580. [CrossRef]
26. Aissani, A.; Artalejo, J. On the single server retrial queue subject to breakdowns. Queueing Syst. 1988,
30, 309–321. [CrossRef]
27. Maraghi, F.A.; Madan, K.C.; Darby-Dowman, K. Bernoulli schedule vacation queue with batch arrivals
and random system breakdowns having general repair time distribution. Int. J. Oper. Res. 2010, 7, 240–256.
[CrossRef]
28. Fadhil, R.; Madan, K.C.; Lukas, A.C. An M(X)/G/1 Queue with Bernoulli Schedule General Vacation Times,
Random Breakdowns, General Delay Times and General Repair Times. Appl. Math. Sci. 2011, 5, 35–51.
29. Rajadurai, P.; Varalakshmi, M.; Saravanarajan, M.C.; Chandrasekaran, V.M. Analysis of M[ X ] /G/1 retrial
queue with two phase service under Bernoulli vacation schedule and random breakdown. Int. J. Math.
Oper. Res. 2015, 7, 19–41. [CrossRef]
30. Jiang, T.; Xin, B. Computational analysis of the queue with working breakdowns and delaying repair under
a Bernoulli-schedule-controlled policy. J. Commun. Stat. Theory Methods 2018, 1–16. [CrossRef]
31. Madan, K.C. A bulk input queue with a stand-by. S. Afr. Stat. J. 1995, 29, 1–7.
32. Khalaf, R.F. Queueing Systems With Four Different Main Server’s Interruptions and a Stand-By Server. Int. J.
Oper. Res. 2014, 3, 49–54. [CrossRef]
33. Jain, M.; Preeti. Cost analysis of a machine repair problem with standby, working vacation and server
breakdown. Int. J. Math. Oper. Res. 2014, 6, 437–451. [CrossRef]
34. Kumar, K.; Jain, M. Bi-level control of degraded machining system with two unreliable servers, multiple
standbys, startup and vacation. Int. J. Oper. Res. 2014, 21, 123–142. [CrossRef]
35. Murugeswari, N.; Maragatha Sundari, S. A Standby server bulk arrival Queuing model of Compulsory
server Vacation. Int. J. Eng. Dev. Res. 2017, 5, 337–341.
36. Kolledath, S.; Kumar, K.; Pippal, S. Survey on queueing models with standbys support. Yugoslav J. Oper. Res.
2018, 28, 3–20. [CrossRef]
37. Neuts, M.F. A general class of bulk queues with poisson input. The Annals of Mathematical Statistics, 1967,
38, 759–770. [CrossRef]
38. Ayyappan, G.; Sathiya, S. Non Markovian Queue with Two Types service Optional Re-service and General
Vacation Distribution. Appl. Appl. Math. Int. J. 2016, 11, 504–526.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

78
 mathematics
Article
Fusion Estimation from Multisensor Observations
with Multiplicative Noises and Correlated Random
Delays in Transmission
Raquel Caballero-Águila 1, *, Aurora Hermoso-Carazo 2 and Josefa Linares-Pérez 2
1 Department of Statistics, University of Jaén, Paraje Las Lagunillas, 23071 Jaén, Spain
2 Department of Statistics, University of Granada, Avda. Fuentenueva, 18071 Granada, Spain;
[email protected] (A.H.-C.); [email protected] (J.L.-P.)
* Correspondence: [email protected]; Tel.: +34-953-212-926

Received: 20 July 2017; Accepted: 29 August 2017; Published: 4 September2017

Abstract: In this paper, the information fusion estimation problem is investigated for a class of
multisensor linear systems affected by different kinds of stochastic uncertainties, using both the
distributed and the centralized fusion methodologies. It is assumed that the measured outputs
are perturbed by one-step autocorrelated and cross-correlated additive noises, and also stochastic
uncertainties caused by multiplicative noises and randomly missing measurements in the sensor
outputs are considered. At each sampling time, every sensor output is sent to a local processor
and, due to some kind of transmission failures, one-step correlated random delays may occur.
Using only covariance information, without requiring the evolution model of the signal process,
a local least-squares (LS) filter based on the measurements received from each sensor is designed
by an innovation approach. All these local filters are then fused to generate an optimal distributed
fusion filter by a matrix-weighted linear combination, using the LS optimality criterion. Moreover,
a recursive algorithm for the centralized fusion filter is also proposed and the accuracy of the
proposed estimators, which is measured by the estimation error covariances, is analyzed by a
simulation example.

Keywords: fusion estimation; sensor networks; random parameter matrices; multiplicative noises;
random delays

1. Introduction
Over the past decades, the use of sensor networks has experienced a fast development encouraged
by the wide range of potential applications in many areas, since they usually provide more information
than traditional single-sensor communication systems. So, important advances have been achieved
concerning the estimation problem in networked stochastic systems and the design of multisensor
fusion techniques [1]. Many of the existing fusion estimation algorithms are related to conventional
systems (see e.g., [2–5], and the references therein), where the sensor measured outputs are affected only
by additive noises and each sensor transmits its outputs to the fusion center over perfect connections.
However, in a network context, usually the restrictions of the physical equipment or the
uncertainties in the external environment, inevitably cause problems in both the sensor outputs
and the transmission of such outputs, that can worsen dramatically the quality of the fusion estimators
designed without considering these drawbacks [6]. Multiplicative noise uncertainties and missing
measurements are some of the random phenomena that usually arise in the sensor measured outputs
and motivate the design of new estimation algorithms (see e.g., [7–11], and references therein).
Furthermore, when the sensors send their measurements to the processing center via a
communication network some additional network-induced phenomena, such as random delays

Mathematics 2017, 5, 45; doi:10.3390/math5030045 79 www.mdpi.com/journal/mathematics

Mathematics 2017, 5, 45

or measurement losses, inevitably arise during this transmission process, which can spoil the fusion
estimators performance and motivate the design of fusion estimation algorithms for systems with
one (or even several) of the aforementioned uncertainties (see e.g., [12–24], and references therein).
All the above cited papers on signal estimation with random transmission delays assume independent
random delays at each sensor and mutually independent delays between the different sensors; in [25]
this restriction was weakened and random delays featuring correlation at consecutive sampling times
were considered, thus allowing to deal with some common practical situations (e.g., those in which
two consecutive observations cannot be delayed).
It should be also noted that, in many real-world problems, the measurement noises are usually
correlated; this occurs, for example, when all the sensors operate in the same noisy environment or
when the sensor noises are state-dependent. For this reason, the fairly conservative assumption that the
measurement noises are uncorrelated is commonly weakened in many of the aforementioned research
papers on signal estimation. Namely, the optimal Kalman filtering fusion problem in systems with
noise cross-correlation at consecutive sampling times is addressed, for example, in [19]; also, under
different types of noise correlation, centralized and distributed fusion algorithms for systems with
multiplicative noise are obtained in [11,20], and for systems where the measurements might have
partial information about the signal in [7].
In this paper, covariance information is used to address the distributed and centralized fusion
estimation problems for a class of linear networked stochastic systems with multiplicative noises
and missing measurements in the sensor measured outputs, subject to transmission random one-step
delays. It is assumed that the sensor measurement additive noises are one-step autocorrelated and
cross-correlated, and the Bernoulli variables describing the measurement delays at the different
sensors are correlated at the same and consecutive sampling times. As in [25], correlated random
delays in the transmission are assumed to exist, with different delay rates at each sensor; however,
the proposed observation model is more general than that considered in [25] since, besides the random
delays in the transmission, multiplicative noises and missing phenomena in the measured outputs are
considered; also cross-correlation between the different sensor additive noises is taken into account.
Unlike [7–11] where multiplicative noise uncertainties and/or missing measurements are considered
in the sensor measured outputs, in this paper random delays in the transmission are also assumed to
exist. Hence, a unified framework is provided for dealing simultaneously with missing measurements
and uncertainties caused by multiplicative noises, along with random delays in the transmission
and, hence, the proposed fusion estimators have wide applicability. Recursive algorithms for the
optimal linear distributed and centralized filters under the least-squares (LS) criterion are derived
by an innovation approach. Firstly, local estimators based on the measurements received from each
sensor are obtained and then the distributed fusion filter is generated as the LS matrix-weighted linear
combination of the local estimators. Also, a recursive algorithm for the optimal linear centralized
filter is proposed. Finally, it is important to note that, even though the state augmentation method
has been largely used in the literature to deal with the measurement delays, such method leads to a
significant rise of the computational burden, due to the increase of the state dimension. In contrast to
such approach, the fusion estimators proposed in the current paper are obtained without needing the
state augmentation; so, the dimension of the designed estimators is the same as that of the original
state, thus reducing the computational cost compared with the existing algorithms based on the
augmentation method.
The rest of the paper is organized as follows. The multisensor measured output model with
multiplicative noises and missing measurements, along with the transmission random one-step delay
model, are presented in Section 2. The distributed fusion estimation algorithm is derived in Section 3,
and a recursive algorithm for the centralized LS linear filtering estimator is proposed in Section 4.
The effectiveness of the proposed estimation algorithms is analyzed in Section 5 by a simulation
example and some conclusions are drawn in Section 6.

80
Mathematics 2017, 5, 45

Notation: The notation throughout the paper is standard. Rn and Rm×n denote the n-dimensional
Euclidean space and the set of all m × n real matrices, respectively. For a matrix A, the symbols A T
and A−1 denote its transpose and inverse, respectively; the notation A ⊗ B represents the Kronecker
product of the matrices A, B. If the dimensions of vectors or matrices are not explicitly stated, they
are assumed to be compatible with algebraic operations. In particular, I denotes the identity matrix
of appropriate dimensions. The notation a ∧ b indicates the minimum value of two real numbers a, b.
For any function Gk,s , depending on the time instants k and s, we will write Gk = Gk,k for simplicity;
analogously, F (i) = F (ii) will be written for any function F (ij) , depending on the
 sensors
i and j.
(i ) (i ) (i )
Moreover, for an arbitrary random vector αk , we will use the notation αk ≡ E αk , where E[.] is
the mathematical expectation operator. Finally, δk,s denotes the Kronecker delta function.

2. Problem Formulation and Model Description

This paper is concerned with the LS linear filtering estimation problem of discrete-time stochastic
signals from randomly delayed observations coming from networked sensors using the distributed
and centralized fusion methods. The signal measurements at the different sensors are affected by
multiplicative and additive noises, and the additive sensor noises are assumed to be correlated and
cross-correlated at the same and consecutive sampling times. Each sensor output is transmitted to a
local processor over imperfect network connections and, due to network congestion or some other
causes, random one-step delays may occur during this transmission process; in order to model different
delay rates in the transmission from each sensor to the local processor, different sequences of correlated
Bernoulli random variables with known probability distributions are used.
In the distributed fusion method, each local processor produces the LS linear filter based on
the measurements received from the sensor itself; afterwards, these local estimators are transmitted
to the fusion center over perfect connections, and the distributed fusion filter is generated by a
matrix-weighted linear combination of the local LS linear filtering estimators using the mean squared
error as optimality criterion. In the centralized fusion method, all measurement data of the local
processors are transmitted to the fusion center, also over perfect connections, and the LS linear filter
based on all the measurements received is obtained by a recursive algorithm.
Next, we present the observation model and the hypotheses on the signal and noise processes
necessary to address the estimation problem.

2.1. Signal Process

The distributed and centralized fusion filtering estimators will be obtained under the assumption
that the evolution model of the signal to be estimated is unknown and only information about its mean
and covariance functions is available; specifically, the following hypothesis is required:
* +
Hypothesis 1. The n x -dimensional signal process xk k≥1 has zero mean and its autocovariance function is
expressed in a separable form, E[ xk xsT ] = Ak BsT , s ≤ k, where Ak , Bs ∈ Rnx ×n are known matrices.

Note that, when the system matrix Φ in the state-space model of a stationary signal is available,
the signal autocovariance function is E[ xk xsT ] = Φk−s E[ xs xsT ], s ≤ k, and Hypothesis 1 is clearly
satisfied taking, for example, Ak = Φk and Bs = E[ xs xsT ](Φ−s ) T . Similarly, if xk = Φk−1 xk−1 + wk−1 ,
the covariance function can be expressed as E[ xk xsT ] = Φk,s E[ xs xsT ], s ≤ k, where Φk,s = Φk−1 · · · Φs ,
−1 T
and Hypothesis 1 is also satisfied taking Ak = Φk,0 and Bs = E[ xs xsT ](Φs,0 ) . Furthermore,
Hypothesis 1 covers even situations where the system matrix in the state-space model is singular,
although a different factorization must be used in those cases (see e.g., [21]). Hence, Hypothesis 1
on the signal autocovariance function covers both stationary and non-stationary signals, providing
a unified context to deal with a large number of different situations and avoiding the derivation of
specific algorithms for each of them.

81
Mathematics 2017, 5, 45

2.2. Multisensor Measured Outputs

Consider m sensors, whose measurements obey the following equations:
 
(i ) (i ) (i ) (i ) (i ) (i )
zk = θk Hk + ε k Ck xk + vk , k ≥ 1, i = 1, . . . , m (1)

(i )
where zk ∈ Rnz is the measured output of the i-th sensor at time k, which is transmitted to a local
(i ) (i )
processor by unreliable network connections, and Hk , Ck are known time-varying matrices of suitable
* (i ) +
dimensions. For each sensor i = 1, . . . , m, θk k≥1 is a Bernoulli process describing the missing
* (i ) + * (i ) +
phenomenon, ε k k≥1 is a scalar multiplicative noise, and vk k≥1 is the measurement noise.

The following hypotheses on the observation model given by Equation (1) are required:
* (i ) +
Hypothesis 2. The processes θk k≥1 , i = 1, . . . , m, are independent sequences of independent Bernoulli
 (i )  (i )
random variables with know probabilities P θk = 1 = θ k , k ≥ 1.

* (i ) +
Hypothesis 3. The multiplicative noises ε k k≥1 , i = 1, . . . , m, are independent sequences of
independent
 scalar random variables with zero means and known second-order moments; we will denote
(i ) (i ) 2
σk ≡ E ε k , k ≥ 1.

* (i ) +
Hypothesis 4. The sensor measurement noises vk k≥1 , i = 1, . . . , m, are zero-mean sequences with known
second-order moments defined by:
# (i ) ( j ) T $ (ij) (ij)
E vk vs = Rk δk,s + Rk,k−1 δk−1,s , s ≤ k; i, j = 1, . . . , m

From Hypothesis 2, different sequences of independent Bernoulli random variables with known
probabilities are used to model the phenomenon of missing measurements at each sensor; so, when
(i ) (i ) (i )
θk = 1, which occurs with known probability θ k , the state xk is present in the measurement zk
(i )
coming from the i-th sensor at time k; otherwise, θk
= 0 and the state is missing in the measured
output from the i-th sensor at time k, which means that such observation only contains additive
(i ) (i )
noise vk with probability 1 − θ k . Although these variables are assumed to be independent from
sensor to sensor, such condition is not necessary to deduce either the centralized estimators or the
local estimators, but only to obtain the cross-covariance matrices of the local estimation errors, which
are necessary to determine the matrix weights of the distributed fusion estimators. Concerning
Hypothesis 3, it should be noted that the multiplicative noises involved in uncertain systems are
usually gaussian noises. Finally, note that the conservative hypothesis of independence between
different sensor measurement noises has been weakened in Hypothesis 4, since such independence
assumption may be a limitation in many real-world problems; for example, when all the sensors
operate in the same noisy environment, the noises are usually correlated, or even some sensors may
have the same measurement noises.

2.3. Observation Model with Random One-Step Delays

(i )
For each k ≥ 1, assume that the measured outputs of the different sensors, zk , i = 1, . . . , m,
are transmitted to the local processors through unreliable communication channels and, due to network
congestion or some other causes, random one-step delays with different rates are supposed to exist
in these transmissions. Assuming that the first measurement is always available and considering

82
Mathematics 2017, 5, 45

* (i ) +
different sequences of Bernoulli random variables, γk k≥2 , i = 1, . . . , m, to model the random
delays, the observations used in the estimation are described by:

(i ) (i ) (i ) (i ) (i ) (i ) (i )
yk = (1 − γk )zk + γk zk−1 , k ≥ 2; y1 = z1 ; i = 1, . . . , m (2)

(i ) (i ) (i )
From Equation (2) it is clear that γk = 0 means that yk = zk ; that is, the local processor receives
(i ) (i ) (i )
the data from the i-th sensor at the sampling time k. When γk = 1, then yk = zk−1 , meaning that
(i )
the measured output at time k is delayed and the previous one is used for the estimation. z k −1
These Bernoulli random variables modelling the delays are assumed to be one-step correlated,
thus covering many practical situations; for example, those in which consecutive observations
transmitted through the same channel cannot be delayed, or situations where there are some sort
of links between the different communications channels. Specifically, the following hypothesis
is assumed:
* (i ) +
Hypothesis 5. γk k≥2 , i = 1, . . . , m, are sequences of Bernoulli random variables with known means,
(i ) # (i ) $ (i ) ( j)
γk ≡ E γk , k ≥ 2. It is assumed that γk and γs are independent for |k − s| ≥ 2, and the second-order
(i,j) # (i ) ( j ) $
moments, γk,s ≡ E γk γs , s = k − 1, k, and i, j = 1, . . . , m, are also known.

Finally, the following independence hypothesis is also required:

* + * (i ) + * (i ) + * (i ) + * (i ) +
Hypothesis 6. For i = 1, . . . , m, the processes xk k≥1 , θk k≥1 , ε k k≥1 , vk k≥1 and γk k≥2 are
mutually independent.

In the following proposition, explicit expressions for the autocovariance functions of the
transmitted and received measurements, that will be necessary for the distributed fusion estimation
algorithm, are derived.

(ij) (i ) ( j ) T y (i ) (i ) (i ) T
Proposition 1. For i, j = 1, . . . , m, the autocovariance functions Σzk,s ≡ E[zk zs ] and Σk,s ≡ E[yk ys ]
are given by:

(i ) (i )
(i ) (i ) T (i ) (i ) (i ) T (i )
Σzk = θ k Hk Ak BkT Hk + σk Ck Ak BkT Ck + Rk , k ≥ 1
(ij) (i ) ( j ) (i ) ( j) T (ij) (ij)
Σzk,s = θ k θ s Hk Ak Bs Hs + Rk,k−1 δk−1,s + Rk δk,s , i =
T j, or s < k
y(ij)  (i ) ( j) (i,j)  (ij)  ( j) (i,j)  (ij)
Σk,s = 1 − γk − γs + γk,s Σzk,s + γs − γk,s Σzk,s−1 (3)
 (i ) (i,j)  (ij) (i,j) (ij)
+ γk − γk,s Σzk−1,s + γk,s Σzk−1,s−1 , k, s ≥ 2
y ( ij )  (i )  z(ij) + γ(i ) Σz(ij) ; y(ij) (ij)
Σ2,1 = 1 − γ2 Σ2,1 2 1 Σ1 = Σ1z

Proof. From Equations (1) and (2), taking into account Hypotheses 1–6, the expressions given in
Equation (3) are easily obtained.

3. Distributed Fusion Linear Filter

In this section, we address the distributed fusion linear filtering problem of the signal from the
randomly delayed observations defined by Equations (1) and (2), using the LS optimality criterion.
In the distributed fusion method, each local processor provides the LS linear filter of the signal xk
(i )
based on the measurements from the corresponding sensor, which will be denoted by xk/k ; afterwards,
(D)
these local filters are transmitted to the fusion center where the distributed filter, is designed as xk/k ,
a matrix-weighted linear combination of such local filters. First, in Section 3.1, for each i = 1, . . . , m,
(i )
a recursive algorithm for the local LS linear filter, xk/k , will be deduced. Then, in Section 3.2,

83
Mathematics 2017, 5, 45

# $
the derivation of the cross-correlation matrices between any two local filters, Σ  (ij) = E x(i) x( j)T ,
k/k k/k k/k
i, j = 1, . . . , m, will be detailed. Finally, in Section 3.3, the distributed fusion filter weighted by matrices,
(D)
xk/k , will be generated from the local filters by applying the LS optimality criterion.

3.1. Local LS Linear Filtering Recursive Algorithm

To obtain the signal LS linear filters based on the available observations from each sensor, we will
use an innovation approach. For each sensor i = 1, . . . , m, the innovation at time k, which represents
(i ) (i ) (i )
the new information provided by the k-th observation, is defined by μk = yk − yk/k−1 , k ≥ 1,
(i ) (i )
where yk/k−1 is the LS linear estimator of yk based on the previous observations, ys , s ≤ k − 1,
(i ) (i )
with y1/0 = E[y1 ] = 0.
* (i ) +
As it is known (see e.g., [26]), the innovations, μk k≥1 , constitute a zero-mean white process,
(i ) (i )
and the LS linear estimator of any random vector αk based on the observations y1 , . . . , y L , denoted by
(i ) (i ) (i )

αk/L , can be calculated as a linear combination of the corresponding innovations, μ1 , . . . , μ L ; namely,

L # (i ) T $
∑E
(i ) (i )−1 (i )

αk/L = αk μh Πh μh (4)
h =1

(i ) # (i ) (i ) T $ (i )
where Πh ≡ E μh μh denotes the covariance matrix of μh .
This general expression for the LS linear estimators along with the Orthogonal Projection Lemma
(OPL), which guarantees that the estimation error is uncorrelated with all the observations or,
equivalently, that it is uncorrelated with all the innovations, are the essential keys to derive the
proposed recursive local filtering algorithm.
Taking into account Equation (4), the first step to obtain the signal estimators is to find an explicit
(i ) (i )
formula for the innovation μh or, equivalently, for the observation predictor yh/h−1 .
(i )
Using the following alternative expression for the observations yk given by Equation (2),

(i )  (i )  (i )  (i ) (i ) (i )  (i ) (i ) (i ) (i )
yk = 1 − γk θk Hk + ε k Ck xk + γk θ k−1 Hk−1 xk−1 + wk , k ≥ 2
(i ) 
(i ) (i ) (i )  (i ) (i ) (i ) (i ) (i )
wk = γk θk−1 − θ k−1 Hk−1 xk−1 + γk θk−1 ε k−1 Ck−1 xk−1 (5)
 (i )  (i ) (i ) (i )  (i ) (i )  (i ) (i ) 
+ 1 − γ k v k + γ k v k − 1 − γk − γ k z k − z k −1 , k ≥ 2

and taking into account the independence hypotheses on the model, it is easy to see that:

(i )  (i )  (i ) (i ) (i ) (i ) (i ) (i ) (i ) (i )
yk/k−1 = 1 − γk θ k Hk xk/k−1 + γk θ k−1 Hk−1 xk−1/k−1 + w
k/k−1 , k ≥ 2

(i ) (i )
Now, taking into account that wh is uncorrelated with ys for s ≤ h − 2, and using Equation (4)
(i )
for k/k−1 ,
w we obtain that:

(i )  (i )  (i ) (i ) (i ) (i ) (i ) (i ) (i )
yk/k−1 = 1 − γk θ k Hk xk/k−1 + γk θ k−1 Hk−1 xk−1/k−1
(k−1)∧2 (6)
∑
(i ) (i )−1 (i )
+ Wk,k−h Πk−h μk−h , k ≥ 2
h =1

(i ) # (i ) (i ) T $
where Wk,k−h ≡ E wk μk−h , h = 1, 2.
Equation (6) for the one-stage observation predictor is the starting point to derive the local
recursive filtering algorithm presented in Theorem 1; this algorithm provide also the filtering error
(i ) # (i ) (i ) $
covariance matrices, Pk/k ≡ E ( xk − xk/k )( xk − xk/k ) T , which measure the accuracy of the estimators
(i )
xk/k when the LS optimality criterion is used.

84
Mathematics 2017, 5, 45

(i )
Theorem 1. Under Hypotheses 1–6, for each single sensor node i = 1, . . . , m, the local LS linear filter, xk/k ,
(i )
and the corresponding error covariance matrix, Pk/k , are given by:

(i ) (i )
xk/k = Ak Ok , k ≥ 1 (7)
and:
 
(i ) (i ) T
Pk/k = Ak Bk − Ak rk , k≥1 (8)

(i ) (i ) (i ) (i ) T
where the vectors Ok and the matrices rk = E[Ok Ok ] are recursively obtained from:

(i ) (i ) (i ) (i )−1 (i ) (i )
Ok = Ok−1 + Jk Πk μk , k ≥ 1; O0 = 0 (9)

(i ) (i ) (i ) (i )−1 (i ) T (i )
rk = rk−1 + Jk Πk Jk , k ≥ 1; r0 = 0 (10)
(i ) (i ) (i ) T
and the matrices Jk = E[Ok μk ] satisfy:

(k−1)∧2
∑
(i ) (i ) T (i ) (i ) T (i ) (i )−1 (i ) T (i ) (i ) T
Jk = H B − r k −1 H A − Jk−h Πk−h Wk,k−h , k ≥ 2; J1 = H B1 (11)
k k
h =1

(i ) (i )
The innovations μk , and their covariance matrices, Πk , are given by:

(k−1)∧2
∑
(i ) (i ) (i ) (i ) (i ) (i )−1 (i ) (i ) (i )
μ k = y k − H A Ok − 1 − Wk,k−h Πk−h μk−h , k ≥ 2; μ1 = y1 (12)
k
h =1

and:
(k−1)∧2
y (i ) (i )  (i ) T (i )  (i )−1  (i ) (i ) (i )  T
Πk = Σk − H A H B − Jk − ∑ Wk,k−h Πk−h H A Jk−h + Wk,k−h , k ≥ 2
(i ) (i )
k k
h =1
k (13)
(i ) y (i )
Π1 = Σ1

(i ) (i ) (i ) T
The coefficients Wk,k−h = E[wk μk−h ], h = 1, 2, are calculated as:

(i ) y (i ) (i ) (i ) T (i ) (i )−1  (i ) (i ) (i ) T
Wk,k−1 = Σk,k−1 − H A H Bk−1 − Wk,k−2 Πk−2 HA Jk−2 + Wk−1,k−2 , k≥3
k k −1
(i ) y (i ) (i ) (i ) T (14)
W2,1 = Σ2,1 − H A2 H B1
(i ) (i ) (i ) (i ) (i ) (i ) (i )
Wk,k−2 = γk (1 − γk−2 ) Rk−1,k−2 , k ≥ 4; W3,1 = γ3 R2,1

y (i ) (i )
Finally, the matrices Σk,s are given in Equation (3) and HΨs , Ψ = A, B, s = k − 1, k, are obtained by:

(i )  (i )  (i ) (i ) (i ) (i ) (i ) (i ) (i ) (i )
HΨs = 1 − γs θ s Hs Ψs + γs θ s−1 Hs−1 Ψs−1 , s ≥ 2; HΨ1 = θ 1 H1 Ψ1 (15)

(i )
Proof. The local filter xk/k will be obtained from the general expression given in Equation (4), starting
from the computation of the coefficients:

(i ) # (i ) T $ # (i ) T $ # (i ) T $
Xk,h = E xk μh = E xk yh − E xk yh/h−1 , 1 ≤ h ≤ k

85
Mathematics 2017, 5, 45

The independence hypotheses and the separable structure of the signal covariance assumed in
(i ) T (i ) T (i )
Hypothesis 1 lead to E[ xk yh ] = Ak H Bh , with H Bh given by Equation (15). From Equation (6) for
(i )
yh/h−1 , h ≥ 2, we have:

# (i ) T $  (i )  (i ) # (i ) T $ (i ) T (i ) (i ) # (i ) T $ (i ) T
E xk yh/h−1 = 1 − γh θ h E xk xh/h−1 Hh + γh θ h−1 E xk xh−1/h−1 Hh−1
(h−1)∧2
∑
(i ) (i )−1 (i ) T
+ Xk,h− j Πh− j Wh,h− j
j =1

(i ) (i )
Hence, using now Equation (4) for xh/h−1 and xh−1/h−1 , the filter coefficients are expressed as:

h −1  
(i )  (i ) (i ) T (i ) T (i ) (i )
∑ Xk,j Π j
(i ) (i ) T (i ) (i )−1 (i ) T (i ) T
Xk,h = Ak H Bh − 1 − γh θ h Xh,j Hh + γh θ h−1 Xh−1,j Hh−1
j =1
(h−1)∧2
∑
(i ) (i )−1 (i ) T
− Xk,h− j Πh− j Wh,h− j , 2 ≤ h ≤ k
j =1
(i ) (i ) T
Xk,1 = Ak H B1

(i ) (i ) (i )
which guarantees that Xk,h = Ak Jh , 1 ≤ h ≤ k, with Jh given by:

h −1 (h−1)∧2
(i ) T (i )−1 (i ) T (i ) T
∑ Jj ∑
(i ) (i ) (i ) (i )−1 (i ) T
Jh = H B h − Πj Jj H Ah − Jh− j Πh− j Wh,h− j , h ≥ 2
j =1 j =1 (16)
(i ) (i ) T
J1 = H B1

k # (i ) (i ) T $
∑ Jh
(i ) (i ) (i )−1 (i ) (i )
Therefore, by defining Ok = Πh μh and rk = E Ok Ok , Equation (7) for the filter
h =1
follows immediately from Equation (4), and Equation (8) is obtained by using the OPL to express
(i ) # $ # (i ) (i ) T $
Pk/k = E xk xkT − E xk/k xk/k , and applying Hypothesis 1 and Equation (7).
The recursive Equations (9) and (10) are directly obtained from the corresponding definitions,
k
∑ Jh
(i ) (i ) (i )−1 (i ) T
taking into account that rk = Πh Jh which, in turn, from Equation (16), leads to
h =1
(i )
Equation (11) for Jk .
(i ) (i ) (i ) (i )
From now on, using that xk/k−1 = Ak Ok−1 , xk−1/k−1 = Ak−1 Ok−1 and Equation (15),
the expression for the observation predictor given by Equation (6) will be rewritten as follows:

(k−1)∧2
∑
(i ) (i ) (i ) (i ) (i )−1 (i )
yk/k−1 = H A Ok−1 + Wk,k−h Πk−h μk−h , k ≥ 2 (17)
k
h =1

From Equation (17), Equation (12) for the innovation is directly obtained and, applying the OPL
(i ) # (i ) (i ) T $ # (i ) (i ) T $
to express its covariance matrix as Πk = E yk yk − E yk/k−1 yk/k−1 , the following identity holds:

(i ) y (i ) (i ) # (i ) (i ) T $ (k−1)∧2 (i ) (i )−1 # (i ) (i ) T $
Πk = Σk − H A E Ok−1 yk/k−1 − ∑h=1 Wk,k−h Πk−h E μk−h yk/k−1 , k ≥ 2
k
(i ) y (i )
Π1 = Σ1

Now, using again Equation (17), and taking Equation (11) into account, it is deduced
# (i ) (i ) T $ (i ) T (i ) # (i ) (i ) T $ (i ) # (i ) (i ) T $ (i )
that E Ok−1 yk/k−1 = H B − Jk and, since E yk/k−1 μk−h = H A E Ok−1 μk−h + Wk,k−h and
# (i ) (i ) T $ (i )
k
(i )
k

E Ok−1 μk−h = Jk−h , h = 1, 2, Equation (13) for Πk is obtained.

86
Mathematics 2017, 5, 45

(i ) # (i ) (i ) T $ (i )
To complete the proof, the expressions for Wk,k−h = E wk μk−h , h = 1, 2, with wk given
(i ) (i )
in Equation (5), are derived using that wk is uncorrelated with yh , h ≤ k − 3. Consequently,
(i ) # (i ) (i ) T $ (i )
Wk,k−2 = E wk yk−2 , and Equation (14) for Wk,k−2 is directly obtained from Equations (1), (2) and (5),
using the hypotheses stated on the model.
(i ) (i ) # (i ) (i ) T $ # (i ) (i ) T $
Next, using Equation (4) for yk−1/k−2 in Wk,k−1 = E wk yk−1 − E wk yk−1/k−2 , we have:

# (i ) (i ) T $  # 
(i ) (i ) (i )−1 (i ) (i ) T $ T
Wk,k−1 = E wk yk−1 − Wk,k−2 Πk−2 E yk−1 μk−2 (18)

To compute the first expectation involved in this formula, we write:

(i ) (i ) (i ) (i )  (i ) (i ) (i )  (i ) (i ) (i )
w k = y k − (1 − γ k ) θ k Hk + ε k Ck xk − γk θ k−1 Hk−1 xk−1
# (i ) T $ # (i ) (i ) T $
and we apply the OPL to rewrite E xs yk−1 = E xs/k−1 yk−1 , s = k, k − 1,
# (i ) (i ) T $ y (i ) (i ) # (i ) (i ) T $
thus obtaining that E wk yk−1 = Σk,k−1 − H A E Ok−1 yk−1 ; then, by expressing
# (i ) (i ) T $ # (i ) (i ) T $ # (i ) (i ) T $ k

E Ok−1 yk−1 = E Ok−1 μk−1 + E Ok−1 yk/k−1 and using Equations (11) and (17), it follows that
# (i ) (i ) T $ y (i ) (i ) (i ) T
E wk yk−1 = Σk,k−1 − H A H B .
k k −1
The second expectation in Equation (18) is easily computed taking into account that, from the
# (i ) (i ) T $
OPL, it is equal to E yk−1/k−2 μk−2 and using Equation (17).
So the proof of Theorem 1 is completed.

3.2. Cross-Correlation Matrices between Any Two Local Filters

To obtain the distributed filtering estimator, the cross-correlation matrices between any pair of
local filters must be calculated; a recursive formula for such matrices is derived in the following
theorem (the notation in this theorem is the same as that used in Theorem 1).

(ij)
Theorem 2. Under Hypotheses 1–6, the cross-correlation matrices between two local filters, Σk/k =
# (i ) ( j ) T $
E xk/k xk/k , i, j = 1, . . . , m, are calculated by:

 (ij) = Ak r (ij) A T , k ≥ 1
Σ (19)
k/k k k

(ij) # (i ) ( j ) T $
with rk = E Ok Ok satisfying:

(ij) (ij) (ij) ( j)−1 ( j) T (i ) (i )−1 ( ji ) T (ij)

rk = rk−1 + Jk−1,k Πk Jk + Jk Πk Jk , k ≥ 1; r0 =0 (20)

(ij) # (i ) ( j ) T $
where Jk−1,k = E Ok−1 μk are given by:

 (i ) (k−1)∧2
(ij) (ij)  ( j) T ( ji ) T
Jk−1,k = rk−1 − rk−1 H A + ∑ Jk−h Πk−h Wk,k−h
(i ) (i )−1
k
h =1
(k−1)∧2 (21)
(ij) ( j)−1 ( j) T
− ∑ Jk−1,k−h Πk−h Wk,k−h , k≥2
h =1
(ij)
J0,1 = 0

(ij) # (i ) ( j ) T $
and Jk,s = E Ok μs , for s = k − 1, k, satisfy:

(ij) (ij) (i ) (i )−1 (ij) (ij) (i ) (i )−1 (ij)

Jk,s = Jk−1,s + Jk Πk Πk,s , k ≥ 2; J1 = J1 Π1 Π1 (22)

87
Mathematics 2017, 5, 45

(ij) # (i ) ( j ) T $
The innovation cross-covariance matrices Πk = E μk μk are obtained as:

(k−1)∧2
(ij) y(ij) (i )  ( j ) T ( j) (ij)  (i )−1 (ij)
− H A H Bk − Jk − Jk−1,k − ∑ Wk,k−h Πk−h Πk−h,k
(i )
Πk = Σk
k
h =1
(k−1)∧2
(ij) ( j)−1  ( j) ( j) ( j)  T (23)
− ∑ Wk,k−h Πk−h H Ak Jk−h + Wk,k−h , k≥2
h =1
(ij) y(ij)
Π1 = Σ1

(ij) # (i ) ( j ) T $
where Πk,s = E μk μs , s = k − 2, k − 1, are given by:

(k−1)∧2
(ij) (i )  ( j ) (ij)  (ij) (i )−1 (ij)
Πk,s = H A Js − Jk−1,s + Wk,s − ∑ Wk,k−h Πk−h Πk−h,s , k ≥ 2
(i )
(24)
k
h =1

(ij) # (i ) ( j ) T $
The coefficients Wk,k−h = E wk μk−h , h = 1, 2, are computed by:

(ij) y(ij) (i ) ( j) T (ij) ( j)−1  ( j) ( j) ( j) T

Wk,k−1 = Σk,k−1 − H A H Bk−1 − Wk,k−2 Πk−2 HA Jk−2 + Wk−1,k−2 , k≥3
k k −1
(ij) y(ij) (i ) ( j ) T (25)
W2,1 = Σ2,1 − H A2 H B1
(ij) (i )  ( j)  (ij) (ij) (i ) (ij)
Wk,k−2 = γk 1 − γk−2 Rk−1,k−2 , k ≥ 4; W3,1 = γ3 R2,1

y(ij) (l ) (l )
Finally, the matrices Σk,s , and H As , H Bs , s = k − 1, k, l = i, j, are given in Equations (3) and (15),
respectively.

Proof. Equation (19) for Σ  (ij) is directly obtained using Equation (7) for the local filters and defining
k/k
(ij) # (i ) ( j ) T $
r k = E Ok Ok .
(ij)
Next, we derive the recursive formulas to obtain the matrices rk , which clearly satisfy
(ij) # (i ) ( j ) T $
Equation (20) just by using Equation (9) and defining Js,k = E Os μk , s = k − 1, k.
( j)
For later derivations, the following expression of the one-stage predictor of yk based on the
observations of sensor i will be used; this expression is obtained from Equation (5), taking into account
(i ) (i ) ( ji ) # ( j ) (i ) T $
that xk/s = Ak Os , s = k − 1, k, and defining Wk,k−h = E wk μk−h , h = 1, 2:

(k−1)∧2
( j/i ) ( j) ( ji )
∑
(i ) (i )−1 (i )
yk/k−1 = H A Ok−1 + Wk,k−h Πk−h μk−h , k ≥ 2 (26)
k
h =1

As Equation (17) is a particular case of Equation (26), for i = j, hereafter we will also refer to it for
(i )
the local predictors yk/k−1 , k ≥ 2.
# (i ) ( j ) T $ # (i) ( j/i)T $
By applying the OPL, it is clear that E Ok−1 yk = E Ok−1 yk/k−1 and, consequently,
(ij) # (i )  ( j/i ) ( j)  T $
we can rewrite Jk−1,k = E Ok−1 yk/k−1 − yk/k−1 ; then, using Equation (26) for both predictors,
(ij)
Equation (21) is easily obtained. Also, Equation (22) for Jk,s , s = k − 1, k, is immediately deduced
(ij) # (i ) ( j ) T $
from Equation (9), just defining Πk,s = E μk μs .
(ij) y(ij) # (i/j) ( j)T $
To obtain Equation (23), first we apply the OPL to express Πk = Σk − E yk/k−1 yk/k−1
# (i ) ( j) T $ (i/j) (i ) (ij)
− E yk/k−1 μk . Then, using Equation (26) for yk/k−1 and yk/k−1 , and the definitions of Jk−1,k and
(ij)
Πk−h,k , we have:

88
Mathematics 2017, 5, 45

# (i/j) ( j)T $ (k−1)∧2

(i ) # ( j ) ( j ) T $ (ij) ( j)−1 # ( j) ( j) T $
E yk/k−1 yk/k−1 = H A E Ok−1 yk/k−1 + ∑ Wk,k−h Πk−h E μk−h yk/k−1
k
h =1
# (k−1)∧2
( j) T $ (i ) (ij) (i )−1 (ij)
∑
(i ) (i )
E yk/k−1 μk = H Ak Jk−1,k + Wk,k−h Πk−h Πk−h,k
h =1
# ( j) ( j) T $ ( j) T ( j)
so Equation (23) is obtained taking into account that E Ok−1 yk/k−1 = H Bk − Jk and
# ( j) ( j) T $ ( j) ( j) ( j)
E yk/k−1 μk−h = H A Jk−h + Wk,k−h , as it has been shown in the proof of Theorem 1.
k
(ij) # (i ) ( j ) T $ # (i ) ( j) T $
Equation (24) for Πk,s = E yk μs − E yk/k−1 μs , with s = k − 2, k − 1, is obtained from
# (i ) ( j ) T $ (i ) ( j ) (ij) # (i ) ( j) T $
E yk μs = H A Js + Wk,s , and using Equation (26) in E yk/k−1 μs .
k
(ij) # (i ) ( j ) T $
Finally, the reasoning to obtain Equation (25) for the coefficients Wk,k−h = E wk μk−h , h = 1, 2,
(i )
is also similar to that used to derive Wk,k−h in Theorem 1, so it is omitted and the proof of Theorem 2
is then completed.

3.3. Derivation of the Distributed LS Fusion Linear Filter

As it has been mentioned previously, a matrix-weighted fusion linear filter is now generated
from the local filters by applying the LS optimality criterion. The distributed fusion filter at any time
 
k is hence designed as a product, Fk X k/k , where X
 k/k = x(1)T , . . . , x(m)T T is the vector constituted
k/k k/k
by the local filters, and Fk ∈ Rnx ×mnx is the matrix obtained by minimizing the mean squared error,
#   $
E xk − Fk X  k/k T xk − Fk X
 k/k .
# $ # $  −1
opt T
As it is known, the solution of this problem is given by Fk = E xk X E X k/k X
T and,
k/k k/k
consequently, the proposed distributed filter is expressed as:
# $ −1
(D) T Σ
xk/k = E xk X  X  (27)
k/k k/k k/k

# $  
 k/k ≡ E X
with Σ  k/k X
T  (ij)  (ij) are the cross-correlation matrices between
k/k = Σk/k i,j=1, ...,m
, where Σ k/k
any two local filters given in Theorem 2.
The distributed fusion linear filter weighted by matrices is presented in the following theorem.
 (1) T (m) T  T
Theorem 3. Let X  k/k = xk/k , . . . , xk/k denote the vector constituted by the local LS filters given in
  # (i ) ( j ) T $

Theorem 1, and Σk/k =  (ij)
Σk/k  (ij)
, with Σk/k = E xk/k xk/k given in Theorem 2. Then, the distributed
i,j=1, ...,m
(D) (D)
filtering estimator, xk/k , and the error covariance matrix, Pk/k , are given by:
 
(D)  (1) , . . . , Σ
 (m) Σ 
 −1 X
xk/k = Σ k/k k/k k/k k/k , k ≥ 1 (28)

and:    T
(D)
Pk/k = Ak BkT − Σ (1) , . . . , Σ
 (m) Σ  −1 Σ  (1) , . . . , Σ
 (m) , k ≥ 1 (29)
k/k k/k k/k k/k k/k

Proof. As it has been discussed previously, Equation (28) is immediately derived from Equation (27),
# $  # $ # $   (1) 
 T = E x(1) x(1)T , . . . , E x(m) x(m)T = Σ   (m)
since the OPL guarantees that E xk X k/k k/k k/k k/k k/k k/k , . . . , Σk/k .
(D) # $ # (D) (D)T $
Equation (29) is obtained from Pk/k = E xk xkT − E xk/k xk/k , using Hypothesis 1 and Equation (28).
Then, Theorem 3 is proved.

4. Centralized LS Fusion Linear Filter

In this section, using an innovation approach, a recursive algorithm is designed for the LS linear
(C )
centralized fusion filter of the signal, xk , which will be denoted by xk/k .

89
Mathematics 2017, 5, 45

4.1. Stacked Observation Model

In the centralized fusion filtering, the observations of the different sensors are jointly processed
(C )
at each sampling time to yield the filter xk/k . To carry out this process, at each sampling time k ≥ 1
 (1) T (m) T  T
we will deal with the vector constituted by the observations from all sensors, yk = yk , . . . yk ,
which, from Equation (2), can be expressed as:
 
yk = I − Γk zk + Γk zk−1 , k ≥ 2; y1 = z1 (30)
 (1) T (m) T  T
where zk = zk , . . . zk is the vector constituted by the sensor measured outputs given in
 (1) (m) 
Equation (1), and Γk = Diag γk , . . . , γk ⊗ I.
 (1) (m) 
Let us note that the stacked vector zk is affected by random matrices Θk = Diag θk , . . . , θk ⊗I
 (1) (m)   (1) T (m) T  T
and Ek = Diag ε k , . . . , ε k ⊗ I, and by a measurement additive noise vk = vk , . . . vk ;
 (1) T (m) T  T  (1) T (m) T  T
so, denoting Hk = Hk , . . . Hk and Ck = Ck , . . . Ck , we have:

zk = Θk ( Hk + Ek Ck ) xk + vk , k ≥ 1 (31)

Hence, the problem is to obtain the LS linear estimator of the signal, xk , based on the observations
y1 , . . . , yk , and this problem requires the statistical properties of the processes involved in Equations (30)
and (31), which are easily inferred from the model Hypotheses 1–6:
* +
Property 1. Θk k ≥1
is a sequence of independent random parameter matrices whit known means
 (1) (m) 
Θk ≡ E[Θk ] = Diag θk , . . . , θk ⊗ I.
* +
Property 2. Ek k≥1 is a sequence of independent random parameter matrices whose entries have zero means
and known second-order moments.
* +
Property 3. The noise vk k≥1 is a zero-mean sequence with known second-order moments defined by the
 (ij) 
matrices Rk,s ≡ Rk,s i,j=1, ...,m .

* +  (1) (m) 
Property 4. The matrices Γk k≥2 have known means, Γk ≡ E[Γk ] = Diag γk , . . . , γk ⊗ I , k ≥ 2,
and Γk and Γs are independent for |k − s| ≥ 2.
* + * + * + * + * +
Property 5. The processes xk k≥1 , Θk k≥1 , Ek k≥1 , vk k≥1 and Γk k≥2 are mutually independent.

4.2. Recursive Filtering Algorithm

In view of Equations (30) and (31) and the above properties, the study of the LS linear filtering
problem based on the stacked observations is completely similar to that of the local filtering problem
carried out in Section 3. Therefore, the centralized filtering algorithm described in the following
theorem is derived by an analogous reasoning to that used in Theorem 1 and its proof is omitted.

(C )
Theorem 4. The centralized LS linear filter, xk/k , is given by:

(C )
xk/k = Ak Ok , k ≥ 1
# $
where the vectors Ok and the matrices rk = E Ok OkT are recursively obtained from:

Ok = Ok−1 + Jk Π−
k μk , k ≥ 1; O0 = 0
1

rk = rk−1 + Jk Π−
k Jk , k ≥ 1; r0 = 0
1 T

90
Mathematics 2017, 5, 45

# $
The matrices Jk = E Ok μkT satisfy:

(k−1)∧2
Jk = H BTk − rk−1 H TAk − ∑ Jk−h Π−
k − h Wk,k− h , k ≥ 2; J1 = H B1 .
1 T T
h =1

The innovations, μk , and their covariance matrices, Πk , are given by:

(k−1)∧2
μ k = y k − H A k Ok − 1 − ∑ Wk,k−h Πk−−1h μk−h , k ≥ 2; μ1 = y1 ,
h =1

and:

  (k−1)∧2  T
Πk = Σk − H Ak H BT − Jk − ∑ Wk,k−h Π−
y y
k − h H Ak Jk − h + Wk,k− h , k ≥ 2; Π1 = Σ1 ,
1
k
h =1
# $
respectively, and the coefficients Wk,k−h = E wk μkT−h , h = 1, 2, satisfy:
 T
Wk,k−1 = Σk,k−1 − H Ak H BTk−1 − Wk,k−2 Π−
y
k −2 H Ak−1 Jk −2 + Wk−1,k−2 , k ≥ 3
1
y
W2,1 = Σ2,1 − H A2 H BT1
Wk,k−2 = Γk Rk−1,k−2 ( I − Γk−2 ), k ≥ 4; W3,1 = Γ3 R2,1 .

y y y  y(ij) 
In the above formulas, the matrices Σk and Σk,k−1 are computed by Σk,s = Σk,s i,j=1,...,m , s = k, k − 1,
y(ij)  (1) T (m) T  T (i )
with Σk,s given in Equation (3), and HΨk = HΨ , . . . , HΨ , with HΨ defined in Equation (15).
k k k

(C )
The performance of the LS linear filters xk/k , k ≥ 1, is measured by the error covariance matrices
# $ 
(C ) (C ) (C ) T
Pk/k = E xk xkT − E xk/k xk/k , whose computation, not included in Theorem 3, is immediate from
(C )
Hypothesis 1 and expression xk/k = Ak Ok of the filter:

(C )
Pk/k = Ak ( Bk − Ak rk ) T , k ≥ 1

Note that these matrices only depend on the matrices Ak and Bk , which are known, and the
matrices rk , which are recursively calculated and do not depend on the current set of observations.
Hence, the filtering error covariance matrices provide a measure of the estimators performance even
before we get any observed data.

5. Numerical Simulation Example

In this section, a numerical example is shown to examine the performance of the proposed
distributed and centralized filtering algorithms and how the estimation accuracy is influenced by the
missing and delay probabilities. Let us consider that the system signal to be estimated is a zero-mean
scalar process, { xk }k≥1 , with autocovariance function E[ xk x j ] = 1.025641 × 0.95k− j , j ≤ k, which is
factorizable according to Hypothesis 1 just taking, for example, Ak = 1.025641 × 0.95k and Bk = 0.95−k .
Sensor measured outputs. The measured outputs of this signal are assumed to be provided by
three different sensors and described by Equation (1):
 
(i ) (i ) (i ) (i ) (i ) (i )
zk = θk Hk + ε k Ck xk + vk , k ≥ 1, i = 1, 2, 3,

where
(1) (2) (3) (1) (2) (3)
• Hk = Hk = 1, Hk = 0.75 and Ck = Ck = 0, Ck = 0.95.

91
Mathematics 2017, 5, 45

* (i ) +
• The processes θk k≥1 , i = 1, 2, 3, are independent sequences of independent Bernoulli random
 (i ) 
variables with constant and identical probabilities for the three sensors P θk = 1 = θ.
* (3) +
• ε k k≥1 is a zero-mean Gaussian white process with unit variance.
(i ) + (i )
• The additive noises {vk k≥1 , i = 1, 2, 3, are defined as vk = ci (ηk + ηk+1 ), i = 1, 2, 3, where
* +
c1 = 0.75, c2 = 1, c3 = 0.5, and ηk k≥1 is a zero-mean Gaussian white process with unit variance.

Note that there are only missing measurements in sensors 1 and 2, and both missing measurements
* (i ) +
and multiplicative noise in sensor 3. Also, it is clear that the additive noises vk k≥1 , i = 1, 2, 3, are only
(ij) (ij)
correlated at the same and consecutive sampling times, with Rk = 2ci c j , Rk,k−1 = ci c j , i, j = 1, 2, 3.
Observations with transmission random one-step delays. Next, according to our theoretical
observation model, it is supposed that, at any sampling time k ≥ 2, the data transmissions are subject
to random one-step delays with different rates and such delays are correlated at consecutive sampling
(i )
times. More precisely, let us assume that the available measurements yk are given by:

(i ) (i ) (i ) (i ) (i )
yk = (1 − γk )zk + γk zk−1 , k ≥ 2, i = 1, 2, 3
* (i ) +
where the variables γk k≥2 modeling this type of correlated random delays are defined using two
* (i ) +
independent sequences of independent Bernoulli random variables, λk k≥1 , i = 1, 2, with constant
(i ) (i ) (i ) (i ) (i )
probabilities, P[λk = 1] = λ , for all k ≥ 1; specifically, we define γk = λk+1 (1 − λk ), for i = 1, 2,
(3) (1) (1)
and γk = λ k (1 − λ k +1 ).
(i ) (i )
It is clear that the sensor delay probabilities are time-invariant: γ(i) = λ (1 − λ ), for i = 1, 2,
* (i ) +
and γ(3) = γ(1) . Moreover, the independence of the sequences λk k≥1 , i = 1, 2, together with
(i )
the independence of the variables within each sequence, guarantee that the random variables γk
( j)
and γs are independent if |k − s| ≥ 2, for any i, j = 1, 2, 3. Also, it is clear that, at each sensor,
* (i ) + (i,i )
the variables γk k≥2 are correlated at consecutive sampling times and γk,s = 0, for i = 1, 2, 3 and
* (3) + * (2) + * (1) +
|k − s| = 1. Finally, we have that γk k≥2 is independent of γk k≥2 , but correlated with γk k≥2
(1,3) (1) (3,1) (1)
at consecutive sampling times, with γk,k−1 = γ(1) λ and γk,k−1 = γ(1) (1 − λ ).
(i ) (i )
Let us observe that, for each sensor i = 1, 2, 3, if = 1, then γk
= 0; this fact guarantees that,γk + 1
when the measurement at time k is delayed, the available measurement at time k + 1 is well-timed.
Therefore, this correlation model covers those situations where the possibility of consecutive delayed
observations at the same sensor is avoided.
To illustrate the feasibility and analyze the effectiveness of the proposed filtering estimators,
the algorithms were implemented in MATLAB, and a hundred iterations were run. In order to measure
the estimation accuracy, the error variances of both distributed and centralized fusion estimators were
calculated for different values of the probability θ of the Bernoulli random variables which model the
missing measurements phenomena, and for several values of the delay probabilities, γ(i) , i = 1, 2, 3,
(i ) (i ) (i )
obtained from several values of λ . Let us observe that the delay probabilities, γ(i) = λ (1 − λ ),
(i ) (i ) (i )
for i = 1, 2, are the same if 1 − λ is used instead of λ ; for this reason, only the case λ ≤ 0.5
was analyzed.
Performance of the local and fusion filtering algorithms. Let us assume that θ = 0.5, and consider
(i )
the same delay probabilities, γ(i) = 0.21, for the three sensors obtained when λ = 0.3, i = 1, 2.
In Figure 1, the error variances of the local, distributed and centralized filters are compared; this figure
shows that the error variances of the distributed fusion filtering estimator are lower than those of
every local estimator, but slightly greater than those of the centralized one. However, this slight
difference is compensated by the fact that the distributed fusion structure reduces the computational

92
Mathematics 2017, 5, 45

cost and has better robustness and fault tolerance. Analogous results are obtained for other values of
the probabilities θ and γ(i) .

0.95
Local filtering error variances (sensor 1)
Local filtering error variances (sensor 2)
Local filtering error variances (sensor 3)
0.9
Distributed filtering error variances
Centralized filtering error variances

0.85

0.8

0.75

0.7

0.65

0.6

0.55
0 10 20 30 40 50 60 70 80 90 100
Time k

Figure 1. Filtering error variances for θ = 0.5, and γ(i) = 0.21, i = 1, 2, 3.

Influence of the missing measurements. Considering again γ(i) = 0.21, i = 1, 2, 3, in order to

show the effect of the missing measurements phenomena, the distributed and centralized filtering
error variances are displayed in Figure 2 for different values of the probability θ; specifically, when θ is
varied from 0.1 to 0.9. In this figure, both graphs (corresponding to the distributed and centralized
fusion filters, respectively) show that the performance of the filters becomes poorer as θ decrease,
which means that, as expected, the performance of both filters improves as the probability of missing
measurements, 1 − θ, decreases. This figure also confirms that both methods, distributed and
centralized, have approximately the same accuracy for the different values of the missing probabilities,
thus corroborating the previous comments.

(a)
1
distributed error variances

0.8

0.6 θ = 0.1

θ = 0.2
0.4 θ = 0.3
θ = 0.4
0 20 40 60 80 100
Time k θ = 0.5
(b)
θ = 0.6
1
centralized error variances

θ = 0.7

0.8 θ = 0.8

θ = 0.9
0.6

0.4

0 20 40 60 80 100
Time k

Figure 2. (a) Distributed and (b) centralized filtering error variances for different values of θ, when
γ(i) = 0.21, i = 1, 2, 3.

93
Mathematics 2017, 5, 45

Influence of the transmission delays. For θ = 0.5, different values for the probabilities γ(i) ,
i = 1, 2, 3, of the Bernoulli variables modelling the one-step delay phenomenon in the transmissions
from the sensors to the local processors have been considered to analyze its influence on the
performance of the distributed and centralized fusion filters. Since the behavior of the error variances
is analogous for all the iterations, only the results at a specific iteration (k = 100) are displayed
here. Specifically, Figure 3 shows a comparison of the filtering error variances at k = 100 in the
following cases:
(1) (2) (1)
(I) Error variances versus λ , when λ = 0.5. In this case, the values λ = 0.1, 0.2, 0.3, 0.4 and
0.5, lead to the values γ(1) = γ(3) = 0.09, 0.16, 0.21, 0.24 and 0.25, respectively, for the delay
probabilities of sensors 1 and 3, whereas the delay probability of sensor 2 is constant and equal
to 0.25.
(1) (2) (1)
(II) Error variances versus λ , when λ = λ . Now, as in Figure 2, the delay probabilities of the
three sensors are equal, and they all take the aforementioned values.

Figure 3 shows that the performance of the distributed and centralized estimators is indeed
(i ) (i )
influenced by the probability λ and, as expected, better estimations are obtained as λ becomes
(i )
smaller, due to the fact that the delay probabilities, γ(i) , decrease with
λ . Moreover, this figure shows
that the error variances in case (II) are less than those of case (I). This is due to the fact that, while
the delay probabilities of the three sensors are varied in case (II), only two sensors vary their delay
probabilities in case (I); since the constant delay probability of the other sensor is assumed to take
its greatest possible value, this figure confirms that the estimation accuracy improves as the delay
probabilities decrease.

0.6 0.6
centralized error variances
distributed error variances

0.595 0.595

0.59 0.59

0.585 0.585

(2) (2)
0.58 λ = 0.5 0.58 λ = 0.5
(2) (1) (2) (1)
λ =λ λ =λ
0.1 0.2 0.3 0.4 0.5 0.1 0.2 0.3 0.4 0.5
(1) (1)
probability λ probability λ

(a) (b)
(1)
Figure 3. (a) Distributed and (b) centralized filtering filtering error variances at k = 100, versus λ .

Comparison. Next, we present a comparative analysis of the proposed centralized filter and the
following ones:

- The centralized Kalman-type filter [4] for systems without uncertainties.

- The centralized filter [8] for systems with missing measurements.
- The centralized filter [25] for systems with correlated random delays.

94
Mathematics 2017, 5, 45

Assuming the same probabilities θ = 0.5 and γ(i) = 0.21 as in Figure 1, and using one thousand
independent simulations, the different centralized filtering estimates are compared using the mean
1 1000  (s) 
(s) 2
square error (MSE) at each sampling time k, which is calculated as MSEk = ∑
1000 s=1
xk − xk/k ,
! "
(s) (s)
where xk ; 1 ≤ k ≤ 100 denotes the s-th set of artificially simulated data and xk/k is the filter
at the sampling time k in the s-th simulation run. The results are displayed in Figure 4, which
shows that: (a) the proposed centralized filtering algorithm provides better estimations than the
other filtering algorithms since the possibility of different simultaneous uncertainties in the different
sensors is considered; (b) the centralized filter [8] outperforms the filter [25] since, even though the
latter accommodates the effect of the delays during transmission, it does not take into account the
missing measurement phenomenon in the sensors; (c) the filtering algorithm in [4] provides the worst
estimations, a fact that was expected since neither the uncertainties in the measured outputs nor the
delays during transmission are taken into account.

1.6

1.4

1.2

0.8

0.6

MSE centralized filter [4]

MSE centralized filter [8]
0.4 MSE centralized filter [25]
MSE proposed centralized filter

10 20 30 40 50 60 70 80 90 100
Iteration k

Figure 4. Filtering mean square errors when θ = 0.5 and γ(i) = 0.21.

Six-sensor network. Finally, according to the anonymous reviewers suggestion, the feasibility
of the proposed estimation algorithms is tested for a larger number of sensors. More specifically,
three additional sensors are considered with the same characteristics as the previous ones, but a
 (i ) 
probability θ ∗ = P θk = 1 = 0.75, i = 4, 5, 6, for the Bernoulli random variables modelling the
missing measurements phenomena. The results are shown in Figure 5, from which similar conclusions
to those from Figure 1 are deduced.

95
 Mathematics 2017, 5, 45

1
Local filtering error variances (sensor 1)
Local filtering error variances (sensor 2)
Local filtering error variances (sensor 3)
0.9 Local filtering error variances (sensor 4)
Local filtering error variances (sensor 5)
Local filtering error variances (sensor 6)
Distributed filtering error variances
0.8
Centralized filtering estimates

0.7

0.6

0.5

0.4

0 10 20 30 40 50 60 70 80 90 100
Time k

Figure 5. Filtering error variances when θ = 0.5, θ ∗ = 0.75 and γ(i) = 0.21.

6. Conclusions
In this paper, distributed and centralized fusion filtering algorithms have been designed in
multi-sensor systems from measured outputs with both multiplicative and additive noises, assuming
correlated random delays in transmissions. The main outcomes and results can be summarized
as follows:

• Covariance information approach. The evolution model generating the signal process is not required
to design the proposed distributed and centralized fusion filtering algorithms; nonetheless, they
are also applicable to the conventional formulation using the state-space model.
• Measured outputs with multiplicative and additive noises. The sensor measured outputs are assumed to
be affected by different stochastic uncertainties (namely, missing measurements and multiplicative
noises), besides cross-correlation between the different sensor additive noises.
• Random one-step transmission delays. The fusion estimation problems are addressed assuming
that random one-step delays may occur during the transmission of the sensor outputs through
the network communication channels; the delays have different characteristics at the different
sensors and they are assumed to be correlated and cross-correlated at consecutive sampling times.
This correlation assumption covers many situations where the common assumption of independent
delays is not realistic; for example, networked systems with stand-by sensors for the immediate
replacement of a failed unit, thus avoiding the possibility of two successive delayed observations.
• Distributed and centralized fusion filtering algorithms. As a first step, a recursive algorithm for the
local LS linear signal filter based on the measured output data coming from each sensor has been
designed by an innovation approach; the computational procedure of the local algorithms is very
simple and suitable for online applications. After that, the matrix-weighted sum that minimizes the
mean-squared estimation error is proposed as distributed fusion estimator. Also, using covariance
information, a recursive centralized LS linear filtering algorithm, with an analogous structure to
that of the local algorithms, is proposed. The accuracy of the proposed fusion estimators, obtained
under the LS optimality criterion, is measured by the error covariance matrices, which can be
calculated offline as they do not depend on the current observed data set.

Acknowledgments: This research is supported by Ministerio de Economía y Competitividad and Fondo Europeo
de Desarrollo Regional FEDER (grant No. MTM2014-52291-P).

96
Mathematics 2017, 5, 45

Author Contributions: All the authors contributed equally to this work. Raquel Caballero-Águila,
Aurora Hermoso-Carazo and Josefa Linares-Pérez provided original ideas for the proposed model and collaborated
in the derivation of the estimation algorithms; they participated equally in the design and analysis of the simulation
results; and the paper was also written and reviewed cooperatively.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Li, W.; Wang, Z.; Wei, G.; Ma, L.; Hu, J.; Ding, D. A Survey on multisensor fusion and consensus filtering for
sensor networks. Discret. Dyn. Nat. Soc. 2015, 2015, 683701.
2. Ran, C.; Deng, Z. Self-tuning weighted measurement fusion Kalman filtering algorithm. Comput. Stat.
Data Anal. 2012, 56, 2112–2128.
3. Feng, J.; Zeng, M. Optimal distributed Kalman filtering fusion for a linear dynamic system with
cross-correlated noises. Int. J. Syst. Sci. 2012, 43, 385–398.
4. Yan, L.; Li, X.R.; Xia, Y.; Fu, M. Optimal sequential and distributed fusion for state estimation in
cross-correlated noise. Automatica 2013, 49, 3607–3612.
5. Feng, J.; Wang, Z.; Zeng, M. Distributed weighted robust Kalman filter fusion for uncertain systems with
autocorrelated and cross-correlated noises. Inf. Fusion 2013, 14, 78–86.
6. Hu, J.; Wang, Z.; Chen, D.; Alsaadi, F.E. Estimation, filtering and fusion for networked systems with
network-induced phenomena: New progress and prospects. Inf. Fusion 2016, 31, 65–75.
7. Liu, Y.; He, X.; Wang, Z.; Zhou, D. Optimal filtering for networked systems with stochastic sensor gain
degradation. Automatica 2014, 50, 1521–1525.
8. Caballero-Águila, R.; García-Garrido, I.; Linares-Pérez, J. Information fusion algorithms for state estimation
in multi-sensor systems with correlated missing measurements. Appl. Math. Comput. 2014, 226, 548–563.
9. Peng, F.; Sun, S. Distributed fusion estimation for multisensor multirate systems with stochastic observation
multiplicative noises. Math. Probl. Eng. 2014, 2014, 373270.
10. Pang, C.; Sun, S. Fusion predictors for multi-sensor stochastic uncertain systems with missing measurements
and unknown measurement disturbances. IEEE Sens. J. 2015, 15, 4346–4354.
11. Tian, T.; Sun, S.; Li, N. Multi-sensor information fusion estimators for stochastic uncertain systems with
correlated noises. Inf. Fusion 2016, 27, 126–137.
12. Shi, Y.; Fang, H. Kalman filter based identification for systems with randomly missing measurements in a
network environment. Int. J. Control 2010, 83, 538–551.
13. Li, H.; Shi, Y. Robust H∞ filtering for nonlinear stochastic systems with uncertainties and random delays
modeled by Markov chains. Automatica 2012, 48, 159–166.
14. Li, N.; Sun, S.; Ma, J. Multi-sensor distributed fusion filtering for networked systems with different delay
and loss rates. Digit. Signal Process. 2014, 34, 29–38.
15. Sun, S.; Ma, J. Linear estimation for networked control systems with random transmission delays and packet
dropouts. Inf. Sci. 2014, 269, 349–365.
16. Caballero-Águila, R.; Hermoso-Carazo, A.; Linares-Pérez, J. Covariance-based estimation from multisensor
delayed measurements with random parameter matrices and correlated noises. Math. Probl. Eng. 2014,
2014, 958474.
17. Chen, B.; Zhang, W.; Yu, L. Networked fusion Kalman filtering with multiple uncertainties. IEEE Trans.
Aerosp. Electron. Syst. 2015, 51, 2332–2349.
18. Caballero-Águila, R.; Hermoso-Carazo, A.; Linares-Pérez, J. Optimal state estimation for networked
systems with random parameter matrices, correlated noises and delayed measurements. Int. J. Gen. Syst.
2015, 44, 142–154.
19. Wang, S.; Fang, H.; Tian, X. Recursive estimation for nonlinear stochastic systems with multi-step
transmission delays, multiple packet dropouts and correlated noises. Signal Process. 2015, 115, 164–175.
20. Chen, D.; Yu, Y.; Xu, L.; Liu, X. Kalman filtering for discrete stochastic systems with multiplicative noises
and random two-step sensor delays. Discret. Dyn. Nat. Soc. 2015, 2015, 809734.
21. Caballero-Águila, R.; Hermoso-Carazo, A.; Linares-Pérez, J. Fusion estimation using measured outputs with
random parameter matrices subject to random delays and packet dropouts. Signal Process. 2016, 127, 12–23.

97
Mathematics 2017, 5, 45

22. Chen, D.; Xu, L.; Du, J. Optimal filtering for systems with finite-step autocorrelated process noises, random
one-step sensor delay and missing measurements. Commun. Nonlinear Sci. Numer. Simul. 2016, 32, 211–224.
23. Wang, S.; Fang, H.; Tian, X. Minimum variance estimation for linear uncertain systems with one-step
correlated noises and incomplete measurements. Digit. Signal Process. 2016, 49, 126–136.
24. Gao, S.; Chen, P.; Huang, D.; Niu, Q. Stability analysis of multi-sensor Kalman filtering over lossy networks.
Sensors 2016, 16, 566.
25. Caballero-Águila, R.; Hermoso-Carazo, A.; Linares-Pérez, J. Linear estimation based on covariances for
networked systems featuring sensor correlated random delays. Int. J. Syst. Sci. 2013, 44, 1233–1244.
26. Kailath, T.; Sayed, A.H.; Hassibi, B. Linear Estimation; Prentice Hall: Upper Saddle River, NJ, USA, 2000.

c 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

98
 mathematics
Article
Forecast Combinations in the Presence of Structural
Breaks: Evidence from U.S. Equity Markets
Davide De Gaetano 1,2
1 University of Roma Tre, Via Silvio D’Amico, 77–00145 Rome, Italy
2 SOSE—Soluzioni per il Sistema Economico S.p.A., Via Mentore Maggini, 48/C–00143 Rome, Italy;
[email protected]

Received: 24 January 2018; Accepted: 20 February 2018; Published: 1 March 2018

Abstract: Realized volatility, building on the theory of a simple continuous time process, has recently
received attention as a nonparametric ex-post estimate of the return variation. This paper addresses
the problem of parameter instability due to the presence of structural breaks in realized volatility in
the context of three HAR-type models. The analysis is conducted on four major U.S. equity indices.
More specifically, a recursive testing methodology is performed to evaluate the null hypothesis of
constant parameters, and then, the performance of several forecast combinations based on different
weighting schemes is compared in an out-of-sample variance forecasting exercise. The main findings
are the following: (i) the hypothesis of constant model parameters is rejected for all markets under
consideration; (ii) in all cases, the recursive forecasting approach, which is appropriate in the absence
of structural changes, is outperformed by forecast combination schemes; and (iii) weighting schemes
that assign more weight in most recent observations are superior in the majority of cases.

Keywords: realized volatility; forecast combinations; structural breaks

1. Introduction
Modeling and forecasting volatility comprise an important issue in empirical finance.
Traditional approaches are based on the univariate GARCH class of models or stochastic
volatility models. Realized Volatility (RV) has lately become very popular; it uses improved measures
of ex-post volatility constructed from high frequency data and provides an efficient estimate of
the unobserved volatility of financial markets. In contrast with the GARCH approach, in which
the volatility is treated as a latent variable, RV can be considered as an observable proxy, and as a
consequence, it can be used in time series models to generate forecasts.
Many authors, staring from [1], have highlighted the importance of structural breaks in RV.
Their presence in the data-generating process can induce instability in the model parameters.
Ignoring structural breaks and wrongly assuming that the structure of a model remains fixed
over time have clear adverse implications. The first finding is the inconsistency of the parameter
estimates. Moreover, structural changes are likely to be responsible for most major forecast failures of
time-invariant series models. Recently, Kumar [2] has found that volatility transmission from crude
oil to equity sectors is structurally unstable and exhibits structural breaks; Gong and Lin [3] have
examined whether structural breaks contain incremental information for forecasting the volatility of
copper futures, and they have argued that considering structural breaks can improve the performance
of most of the existing heterogeneous autoregressive-type models; Ma et al. [4] have introduced
Markov regime switching to forecast the realized volatility of the crude oil futures market; in the same
context, Wang et al. [5] have found that time-varying parameter models can significantly outperform
their constant-coefficient counterparts for longer forecasting horizons.
In this paper, three different model specifications of the log-RV have been considered. The first is
the Heterogeneous Autoregressive model (HAR-RV) proposed in [6], which is able to capture many of

Mathematics 2018, 6, 34; doi:10.3390/math6030034 99 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 34

the features of volatility including long memory, fat tails and self-similarity. The second is the Leverage
Heterogeneous Autoregressive model (LHAR-RV) proposed in [7], which is able to approximate
both long-range dependence and the leverage effect. The last is the Asymmetric Heterogeneous
Autoregressive model (AHAR-RV), which is a simplified version of the model proposed in [1]. In the
spirit of the EGARCH model, the AHAR-RV allows for asymmetric effects from positive and negative
returns. These models, which have become very popular in the econometric literature on RV, have
very parsimonious linear structures, and as a consequence, they are extremely easy to implement and
to estimate. Moreover, they have good performance in approximating many features that characterize
the dynamics of RV [8], and in the forecasting context, they seem to provide results that are at least as
good as more sophisticated models that consider additional components of the realized variance, such
as semivariance and jumps [9,10].
The aim of this paper is to empirically investigate the relevance of structural breaks for forecasting
RV of a financial time series. The presence of structural breaks in the considered RV-representations
has been investigated and verified by resorting to a fluctuation test for parameter instability in a
regression context. In particular, attention has been focused on the recursive estimates test [11].
This choice is particularly motivated in those cases where no particular pattern of the deviation from
the null hypothesis of constant parameters is assumed. Furthermore, the proposal does not require the
specification of the locations of the break points.
In order to handle parameter instability, some specific forecast combinations have been introduced
and discussed. They are based on different estimation windows with alternative weighting schemes.
These forecast combinations, proposed in a regression setting, are employed in financial time series,
highlighting, also in this context, their usefulness in the presence of structural breaks. Moreover, all of
them are feasible for a high sample size; they do not explicitly incorporate the estimation of the break
dates; and as shown by [12], they do not suffer from this estimation uncertainty.
The forecasting performance of the proposed forecast combinations for the three different
specifications of RV models has been compared in terms of two loss functions, the Mean Squared Error
(MSE) and the Quasi-Likelihood (QLIKE) described below. These are the loss functions most widely
used to compare volatility forecasting performance, and according to [13], they provide robust ranking
of the models.
In order to statistically assess if the differences in the forecasting performance of the considered
forecast combinations are relevant, the model confidence set, proposed in [14], has been used.
The empirical analysis has been conducted on four U.S. stock market indices: S&P 500, Dow Jones
Industrial Average, Russell 2000 and Nasdaq 100. For all the series, the 5-min RV has been considered;
it is one of the most used proxies of volatility, and as shown in [15], it favorably compares to more
sophisticated alternatives in terms of estimation accuracy of asset price variation.
The structure of this paper is as follows. Section 2 introduces the empirical models for RV and
briefly illustrates the problem of structural breaks. In Section 3, some of the most used procedures to
test parameters’ instability in the regression framework are reviewed. Attention has been focused on
the class of fluctuation tests, and in particular, the recursive estimates test has been discussed. Section 4
introduces the problem of forecasting in the presence of structural breaks and discusses some forecast
combinations able to take into account parameters’ instability. In Section 5, the empirical results on the
four U.S. stock market indices are reported and discussed. Some final remarks close the paper.

2. Realized Volatility Models

Let p(s) be the log-price of a financial asset at time s, σ2 (s) the instantaneous or spot volatility
and w(s) the standard Brownian motion. Define a simple continuous time process:

dp(s) = σ(s)dw(s) (1)

100
Mathematics 2018, 6, 34

and assume that σ2 (s) has locally square integrable sample paths, stochastically independent of w(s).
The integrated volatility for day t is defined as the integral of σ2 (s) over the interval (t, t + 1):
 t +1
IVt = σ2 (s) ds (2)
t

where a full twenty four-hour day is represented by Time Interval 1. The integrated volatility is not
observable, but it can be estimated using high frequency asset returns.
If m intraday returns are available for each day t, {rt,i } i = 1, . . . , m, it is possible to define a
precise volatility measure, called a realized volatility, as the squared sum of them over day t:
m
RVt = ∑ rt,i2 (3)
i =1

If there were no market microstructure noise, the realized volatility would provide a consistent
estimator of the integrated volatility, that is as the time interval approaches zero or equivalently m
goes to infinity:
RVt → IVt (4)

In this paper, we focus on 5-min realized volatility; this choice is justified on the grounds of past
empirical findings that show that at this frequency, there is no evidence of micro-structure noise [16].
Moreover, as shown in [15], 5-min RV favorably compares to more sophisticated alternatives in terms
of estimation accuracy.
In the econometric literature, many approaches have been developed to model and forecast
realized volatility with the aim of reproducing the main empirical features of financial time series such
as long memory, fat tails and self-similarity. In this paper, attention has been focused on the classic
Heterogeneous Autoregressive model of Realized Volatility (HAR-RV) and on some of its extensions.
The HAR-RV model, proposed in [6], has a very simple and parsimonious structure; moreover,
empirical analysis [8] shows remarkably good forecasting performance. In this model, lags of RV are
used at daily, weekly and monthly aggregated periods.
More precisely, let vt = log( RVt ) where RVt is the realized volatility at time t = 1, 2, . . . , T.
The logarithmic version of the HAR-RV similar to that implemented by [17] is defined as:

(5) (22)
v t = β 0 + β 1 v t −1 + β 2 v t + β 3 vt + t (5)

(5) (22)
where t ∼ N ID (0, σ2 ) and vt and vt are defined, respectively, as:

(5) v t −1 + v t −2 + . . . + v t −5
vt = (6)
5
(22) vt−1 + vt−2 + . . . + vt−22
vt = (7)
22
The HAR-RV model is able to capture some well-known features of financial returns such as long
memory and fat tails [8].
The first extension of this model is the Leverage Heterogeneous Autoregressive model of Realized
Volatility (LHAR-RV) proposed in [7]. This model is defined as:

+ β 4 rt−−1 + β 5 rt
(5) (22) (5)−
v t = β 0 + β 1 v t −1 + β 2 v t + β 3 vt +
(8)
(22)− (5)+ (22)+
+ β 6 rt + β 7 rt+−1 + β 8 rt + β 9 rt + t

101
Mathematics 2018, 6, 34

where t ∼ N ID (0, σ2 ), rt are the daily returns and:

(5)− r t −1 + r t −2 + . . . + r t −5
rt = I{(rt−1 +rt−2 +...+rt−5 )<0} (9)
5
(5)+ r t −1 + r t −2 + . . . + r t −5
rt = I{(rt−1 +rt−2 +...+rt−5 )>0} (10)
5
r + rt−2 + . . . + rt−22
= t −1
(22)−
rt I{(rt−1 +rt−2 +...+rt−22 )<0} (11)
22
r + rt−2 + . . . + rt−22
= t −1
(22)+
rt I{(rt−1 +rt−2 +...+rt−22 )>0} (12)
22
where I is the indicator function. The LHAR-RV model approximates both long-range dependence
and the leverage effect. Some authors ([18]) suggest including only the negative part of heterogeneous
returns since the estimates of the coefficients of the positive ones are usually not significant.
The second extension is the Asymmetric Heterogeneous Autoregressive model of Realized
Volatility (AHAR-RV), which is a simplified version of the model proposed in [1]. It is defined as:

|r | |r |
+ β 4 √ t−1 + β 5 √ t−1 I{rt−1 <0} + t
(5) (22)
v t = β 0 + β 1 v t −1 + β 2 v t + β 3 vt (13)
RVt−1 RVt−1

The last two terms allow for asymmetric effects from positive and negative returns in the spirit of
the EGARCH model.
All the considered models can be rewritten in a standard regression framework:

yt = xt β + t (14)

where yt = vt , xt is the p × 1 vector of the regressors at time t and β is the p × 1 vector of the
corresponding coefficients. Of course, the number p and the specification of the vector xt are different
for each model.
Many studies (see, for example, [1]) agree on the existence of structural breaks in RV. If structural
breaks are present in the data-generating process, they could induce instability in the model parameters.
Ignoring them in the specification of the model could provide the wrong modeling and forecasting for
the RV.
To deal with structural breaks, the linear regression model (14) is assumed to have time-varying
coefficients, and so, it may be expressed as:

yt = xt βt + t t = 1, 2, . . . T (15)

In many applications, it is reasonable to assume that there are m breakpoints at the date
τ1 , τ2 , · · · , τm in which the coefficients shift from one stable regression relationship to a different one.
Thus, there are m + 1 segments in which the regression coefficients are constant. Model (15) can be
rewritten as:
yt = xt βτj−1 +1:τj + t t = 1, 2, . . . T j = 1, 2, . . . m + 1 (16)

and, by convention, τ0 = 1 and τm+1 = T.

3. Testing for Structural Changes

The presence of structural breaks can be tested through the null hypothesis that the regression
coefficients remain constant over time, that is:

H0 : βt = β t = 1, 2, . . . T (17)

against the alternative that at least one coefficient varies over time.

102
Mathematics 2018, 6, 34

In the statistical and econometric literature, testing for parameters’ instability in a regression
framework has been treated using different approaches. The classical test for structural change is the
well-known Chow test [19]. This testing procedure splits the sample into two sub-samples, estimates
the parameters for each sub-sample, and then, using a classic F statistic, a test on the equality of the
two sets of parameters is performed. For a review, which includes also some extensions in different
contexts, see [20]. The principal issue of the Chow test is the assumption that the break-date must be
known a priori. Generally, this procedure is used by fixing an arbitrary candidate break-date or by
selecting it on the basis of some known feature of the data. However, the results can be highly sensitive
to these arbitrary choices; in the first case, the Chow test may be uninformative, and in the second case,
it can be misleading [21].
More recently, the literature has focused on a more realistic problem in which the number of
break points and their locations are supposed to be unknown (see [22], for a survey). In this context,
one of the major contributions is the strategy proposed in Bai and Perron ([23–25]) who developed
an iterative procedure that allows consistent estimation of the number and the location of the break
points together with the unknown regression coefficients in each regime. In their procedure, the breaks
are considered deterministic parameters, and so, the specification of their underlying generating
process is not required. The number of breaks can be sequentially determined by testing for q + 1
against q or using a global approach of testing for q against no breaks. However, the procedure
needs the specification of some restrictions such as the minimum distance between breaks and their
maximum number.
Another approach to change point testing is based on the generalized fluctuation tests (for a
survey, see [26]). Such an approach has the advantage of not assuming a particular pattern of deviation
from the null hypothesis. Moreover, although it is possible in principal to carry out the location of
the break points, this method is commonly used only to verify their presence; with this aim, the
fluctuation tests will be used in this paper. The general idea is to fit a regression model to the data
and derive the empirical process that captures the fluctuation in the residuals or in the parameter
estimates. Under the null hypothesis of constant regression coefficients, fluctuations are governed by
functional central limit theorems ([27]), and therefore, boundaries can be found that are crossed by
the corresponding limiting processes with fixed probability α. When the fluctuation of the empirical
process increases, there is evidence of structural changes in the parameters. Moreover, its trajectory
may also highlight the type of deviation from the null hypothesis, as well as the dating of the structural
breaks. As previously pointed out, the generalized fluctuation tests can be based on the residuals or on
the parameter estimates of the regression model. The first class includes the classical CUSUM based
on Cumulative Sums of recursive residuals [28], the CUSUM test based on OLS residuals [29] and
the Moving Sums (MOSUM) tests based on the recursive and OLS residuals [30]. The second class
includes the Recursive Estimates (RE) test [11] and the Moving Estimates (ME) test [31]. In both, the
vector of unknown parameters is estimated recursively with a growing number of observations, in the
RE test, or with a moving data window, in the ME test, and then compared to the estimates obtained
by using the whole sample.
Define:

y1:t = ( y1 , y2 , . . . , y t ) (18)

X1:t = ( x1 , x2 , . . . , x t )  (19)

and let:
 = (X X1:t )−1 X y1:t
β t = p, p + 1, . . . , T (20)
1:t 1:t 1:t

be the Ordinary Least Squares (OLS) estimate of the regression coefficients based on the observations
up to t.
The basic idea is to reject the null hypothesis of parameter constancy if these estimates fluctuate
too much. Formally, the test statistic is defined as:

103
Mathematics 2018, 6, 34

, ,
, ,
S(T ) = sup ,B(T ) (z), (21)
0≤ z ≤1 ∞

with:
φ(z)   
B(T ) (z) =

σT
(X1:T X1:T )1/2 ( β 1:φ(z) − β1:T ) (22)

where:  1/2
T
1

σ=
T−p ∑ (yt − xt β1:T ) (23)
t =1

and φ(z) is the largest integer less than or equal to p + z( T − p) and .∞ the maximum norm.
As proven in [11], B(T ) (z) is a p-dimensional stochastic process such that

D
B(T ) (z) → B(z) (24)

where B(z) is a Brownian bridge. The distribution of sup0≤z≤1  B(z)∞ is given in [32]; in particular,
it is: ( )  p
∞
= 1 + 2 ∑ (−1)i e−2i
2 x2
P sup  B(z)∞ ≤ x x≥0 (25)
0≤ z ≤1 i =1

4. Forecasting Methods in the Presence of Structural Breaks

Once the parameter instability due to the presence of structural breaks has been detected,
the problem is how to account for it when generating forecasts. Indeed, parameter instability could
cause forecast failures in macroeconomic and financial time series (for a survey, see [33]).
When it is possible to identify the exact date of the last break, the standard solution is to use
only observations over the post-break period. In practice, the dates of the break points are not known
a priori, and an estimation procedure has to be used. It could produce imprecise values, which
negatively affect the specification of the forecasting model and, as a consequence, poor performance of
the forecasts. Furthermore, even if the last break date is correctly estimated, the forecasts generated
by this scheme are likely to be unbiased and may not minimize the mean square forecast error [12].
Moreover, if the last break is detected close to the boundaries of the data sample, the parameters of
the forecasting model are estimated with a relatively short sample, and the estimation uncertainty
may be large.
However, as pointed out in [12], the pre-break observations could be informative for forecasting
even after the break. More specifically, it is appropriate to choose a high fraction of the pre-break
observations especially when the break size is small, the variance parameter increases at the break
point and the number of post break observations is small. Furthermore, the forecasting performance is
sensitive to the choice of the observation window. A relatively long estimation window reduces the
forecast error variance, but increases its bias; on the other hand, a short estimation window produces an
increase in the forecast error variance although the bias decreases. Therefore, an optimal window size
should balance the trade-off between an accurate estimate of the parameters and the possibility that the
data come from different regimes. In this context, Pesaran and Timmermann [12] have proposed some
methods to select the window size in the case of multiple discrete breaks when the errors of the model
are serially uncorrelated and the regressors are strictly exogenous; Pesaran et al. [34] have derived
optimal weights under continuous and discrete breaks in the case of independent errors and exogenous
regressors; Giraitis et al. [35] have proposed to select a tuning parameter to downweight older data by
using a cross-validation based method in the case of models without regressors; Inoue et al. [36] have
suggested to choose the optimal window size that minimizes the conditional Mean Square Forecast
Error (MSFE). However, in practice, the selection of a single best estimation window is not an easy
task, and in many empirical studies, it is arbitrarily determined.

104
Mathematics 2018, 6, 34

Alternatively, in order to deal with the uncertainty over the size of the estimation window, it is
possible to combine forecasts generated from the same model, but over different estimation windows.
This strategy is in the spirit of forecast combinations obtained by estimating a number of alternative
models over the same sample period (for a review, see [37]). It has the advantage of avoiding the
direct estimation of breakpoint parameters, and it is applicable to general dynamic models and for
different estimation methods. In this context, Pesaran and Timmermann [12] have proposed forecast
combinations formed by averaging across forecasts generated by using all possible window sizes
subject to a minimum length requirement. Based on the same idea, more complex forecasting schemes
have been proposed (see, for example, [34,38]).
The idea of forecast averaging over estimation windows has been fruitfully applied also
in macroeconomic forecasting, in particular in the context of vector autoregressive models with
weakly-exogenous regressors ([39,40] and in the context of GDP growth on the yield curve ([41]).
Pesaran and Pick [42] have discussed the theoretical advantages of using such combinations
considering random walks with breaks in the drift and volatility and a linear regression model with
a break in the slope parameter. They have shown that averaging forecasts over different estimation
windows leads to a lower bias and root mean square forecast error than forecasts based on a single
estimation window for all but the smallest breaks. Similar results are reported in [43]; they have
highlighted that, in the presence of structural breaks, averaging forecasts obtained by using all the
observations in the sample and forecasts obtained by using a window can be useful for forecasting. In
this case, forecasts from only two different windows have been combined, and so, this procedure can
be seen as a limited version of that proposed in [12].
In view of the above considerations, in this paper, attention has been focused on forecast schemes
generated from the same model, but over different estimation windows. In particular, for each of the
considered realized volatility models, different forecast combinations have been considered focusing
on those that are feasible for financial time series and that do not explicitly incorporate the estimation
of the break dates. Moreover, in the analysis, one-step ahead forecasts have been considered, and so, it
is assumed that no structural breaks occur in the forecast period (for forecasting with structural breaks
over the forecast period, see [44,45]).

4.1. Forecast Combination With Equal Weights

As previously pointed out, the forecast combination with equal weights is the simplest
combination, but it is robust to structural breaks of unknown break dates and sizes. Moreover,
it performs quite well especially when the break is of moderate magnitude and it is located close to the
boundaries of the data sample [42].
Let ω be the minimum acceptable estimation window size. The forecast combination with equal
weights is defined by:
1 T −ω    
T − ω τ∑
yT +1 = x T +1 βτ +1:T (26)
=1

Many research works have highlighted the advantages of this scheme; it has good performance
also when there is uncertainty about the presence of structural breaks in the data. This approach also
avoids any estimation procedure for the weights.

4.2. Forecast Combination With Location Weights

By looking at Equation (26), it is evident that the weights in the equally-weighted combination
can be converted into weights on the sample observations xt . As discussed in [38], the ω most recent
observations are used in all of the forecasts, whereas the older observations are used less. Furthermore,
the influence of each observation is inversely proportional to its distance from the forecasting origin:
the most recent data are usually more relevant especially if the regression parameters have significant
changes close to the end of the sample.

105
Mathematics 2018, 6, 34

A way to place heavier weights on the forecasts that are based on more recent data much more
than under the equally-weighted forecast combination is to use constant weights proportional to the
location of τ in the sample.
More precisely, this combination, known as the forecast combination with location weights,
is defined by:

1 T −ω  
yT +1 = ∑ 
τ xT +1 β (27)
−ω τ +1:T
∑τT= 1 τ τ =1

Also in this case, no estimation of the weights is needed.

4.3. Forecast Combination With MSFE Weights

This approach, proposed in [12], is based on the idea that the weights of the forecasters obtained
with different estimation windows should be proportional to the inverse of the associated out-of-sample
MSFE values. To this aim, a cross-validation approach is used.
To better understand, let m be the generic start point of the estimation window and assume that ω̃
is the number of observations used in the cross-validation set, that is the observations used to measure
pseudo out-of-sample forecasting performance. The recursive pseudo out-of-sample MSFE value is
computed as:
T −1  2
MSFE(m| T, ω̃ ) = ω̃ −1 ∑ 
yτ +1 − xτ β (28)
m:τ
τ = T −ω̃

The forecast combination with MSFE weights is then defined as:

 
T −ω −ω̃ 
xT β −1
∑m =1 m:T ( MSFE ( m | T, ω̃ ))
yT +1 = T −ω −ω̃
(29)
∑m =1 ( MSFE(m| T, ω̃ ))−1
Together with the parameter ω, the length of the minimal estimation window, this method also
requires the choice of the parameter ω̃ and the length of the evaluation window. If this parameter is set
too large, too much smoothing may result, and as a consequence, in the combination, the forecasting
based on older data will be preferred. On the other hand, if ω̃ is set too short, although a more precise
estimation of the MSFE can be obtained, the ranking of the forecasting methods is more affected
by noise. Of course, the selection of this parameter depends on the problem at hand and on the length
of the series.

4.4. Forecast Combination With ROC Weights

This approach, proposed in [38], is based on by-products of the Reverse Ordered CUSUM (ROC)
structural break test considered in [46].
It is a two-stage forecasting strategy. In the first step, a sequence of ROC test statistics, starting from
the most recent observations and going backwards in time, is calculated. Each point in the sample is
considered as a possible most recent break point.
This test is related to the classical CUSUM test, but in this case, the test sequence is made in
reverse chronological order. In particular, the time series observations are placed in reverse order, and
the standard CUSUM test is performed on the rearranged dataset.
In the paper [46], the test statistics are used to perform a formal structural break test and to
estimate the last breakpoint in the sample.
In the second step, the ROC statistics are used to weight the associated post break forecast,
developing a forecast combination. Moreover, the weights do not depend on finding and dating a
structural break, but they are constructed in order to give more weights to observations subsequent to
a potential structural break.

106
Mathematics 2018, 6, 34

In the first step of the procedure, for τ = T − ω + 1, T − ω, . . . , 2, 1, let:

yT:τ = (y T , y T −1 , . . . , yτ +1 , yτ ) (30)

XT:τ = (x T , x T −1 , . . . , xτ +1 , xτ ) (31)

be the observation matrices, and let:

 ( R) = (X X T:τ )−1 X y T:τ

β (32)
T:τ T:τ T:τ

be a sequence of least squares estimates of β associated with the reverse-ordered datasets.

The ROC test statistics st are defined as:

∑tT=−τω ξ t2
sτ = for τ = T − ω, T − ω − 1, . . . , 2, 1 (33)
∑tT=−1ω ξ t2

where ξ t are the standardized one-step-ahead recursive residuals defined as:

 ( R)
yt − xt β T:t+1
ξt =   (34)
(1 + xt (X T:t+1 X T:t+1 )−1 xt )1/2

In the second step of the procedure, all dates τ are considered as possible choices for the last
breakpoint. The combination weight on each τ is constructed as:
 

sτ − T −Tω−−ωτ +1
cwτ =   τ = 1, 2, . . . , T − ω (35)
−ω T − ω − τ +1
∑τT= 1 st − T −ω

Since, under the null hypothesis of no structural break in τ, it is:

T−ω−τ+1
E(sτ ) = (36)
T−ω

the combination weights vary according to the absolute distances between sτ and its expected value.
As a consequence, cwτ is larger if this distance is large, that is if the evidence of a structural break is
stronger. On the contrary, if in τ, there is no evidence of substantial breakpoint, the associated weight
is small.
Moreover, the weights do not depend on finding and dating a structural break. However, if the
absolute values of the difference between the ROC statistics and their expectation, under the null
hypothesis, start to grow (giving evidence of a potential structural break), the weights cwτ increase
giving more weights to the observations on data, subsequent to τ.
The one-step-ahead forecast based on ROC statistics is defined as:

T −ω  
yT +1 = ∑ 
cwτ (xT +1 β τ +1:T ) (37)
τ =1

4.5. Forecast Combination With ROC Location Weights

In order to take into account a prior belief on the probability that a time τ could be the most
recent break point, it is possible, in the definition of ROC weights, to incorporate an additional weight
function lτ .
Following [38], the new weights are defined as:

107
Mathematics 2018, 6, 34

 

sτ − T −Tω−−ωτ +1 lτ
cwτ =   τ = 1, 2, . . . , T − ω (38)
−ω T − ω − τ +1
∑τT= 1 st − T −ω lτ

For example, if a single break point seems to be equally likely at each time point, the natural
choice is lτ = 1 for τ = 1, 2, . . . , T − ω. In this case, the weights depend only on the magnitude of the
ROC statistics, and the combination defined in (37) is obtained.
However, in the forecasting context, where the identification of the most recent break is essential,
the prior weight lτ could be chosen as an increasing function of the location of time τ in the full sample.
In the spirit of forecast combination with location weights, the most natural choice is lτ = τ. Of course,
different specifications are also allowed.

5. Empirical Application
The data were obtained from the Oxford-Man Institute’s Realised library. It consists of 5-min
realized volatility and daily returns of four U.S. stock market indices: S&P 500, Dow Jones Industrial
Average, Russell 2000 and Nasdaq 100. The sample covers the period from 1 January 2012–4 February
2016; the plots of the RV series and the log-RV series are reported in Figure 1.
In order to investigate the constancy of the regression coefficients in all the considered models
(HAR-RV, LHAR-RV and AHAR-RV), an analysis based on the recursive estimates test has been
employed for all four series. The results are reported in Table 1. The rejection of the null hypothesis of
the constancy of the regression parameters for all the models and for all the series is evident.

Table 1. Recursive estimates test. LHAR, Leverage Heterogeneous Autoregressive model; AHAR,
Asymmetric Heterogeneous Autoregressive model; RV, Realized Volatility.

S&P 500 DJIA

Model statistic p-value Model statistic p-value
HAR-RV 2.369 0.0001 HAR-RV 2.167 0.0006
LHAR-RV 2.198 0.0010 LHAR-RV 1.943 0.0084
AHAR-RV 2.156 0.0011 AHAR-RV 2.089 0.0019
RUSSELL 2000 NASDAQ 100
Model statistic p-value Model statistic p-value
HAR-RV 2.216 0.0004 HAR-RV 1.848 0.0086
LHAR-RV 1.997 0.0048 LHAR-RV 1.659 0.0104
AHAR-RV 2.044 0.0028 AHAR-RV 1.782 0.0093

Moreover, in Figure 2, the fluctuation process, defined in Equation (21), is reported for each model
specification and for each series, along with the boundaries obtained by its limiting process at level
α = 0.05.
The paths of the empirical fluctuation process confirm the parameters’ instability: the boundaries
are crossed for all the series and for all the models.
The above analysis seems to confirm the effectiveness of parameter instability in all three
RV-models for all the considered indices. Moreover, it supports the use of forecasting methods
that take into account the presence of structural breaks.
In order to evaluate and to compare the forecasting performance of the proposed forecast
combinations for each class of model, an initial sub-sample, composed of the data from t = 1 to
t = R, is used to estimate the model, and the one-step-ahead out-of-sample forecast is produced.
The sample is then increased by one; the model is re-estimated using data from t = 1 to t = R + 1;
and the one-step-ahead forecast is produced. The procedure continues until the end of the available
out-of-sample period. In the following examples, R has been fixed such that the number of out of

108
Mathematics 2018, 6, 34

sample observations is 300. To generate the one-step-ahead forecasts, the five competing forecast
combinations, defined in Section 4, have been considered together with a benchmark method.

S&P500 S&P500

−6
0.003

−8
0.002

−10
0.001

−12
0.000

2012 2013 2014 2015 2016 2012 2013 2014 2015 2016

DOW JONES INDUSTRIAL AVERAGE DOW JONES INDUSTRIAL AVERAGE

0.006

−6
0.005
0.004

−8
0.003

−10
0.002
0.001

−12
0.000

2012 2013 2014 2015 2016 2012 2013 2014 2015 2016

RUSSELL 2000 RUSSELL 2000

0e+00 1e−04 2e−04 3e−04 4e−04 5e−04 6e−04

−8
−9
−10
−11
−12

2012 2013 2014 2015 2016 2012 2013 2014 2015 2016

NASDAQ 100 NASDAQ 100

−6
0.0020

−7
0.0015

−8
−9
0.0010

−10
0.0005

−11
−12
0.0000

2012 2013 2014 2015 2016 2012 2013 2014 2015 2016

Figure 1. RV (left) and log-RV (right) for S&P 500, Dow Jones Industrial Average, Russell 2000 and
Nasdaq 100.

As a natural benchmark, we refer to the expanding window method, which ignores the presence
of structural breaks. In fact, it uses all the available observations. As pointed out in [12], this choice is
optimal in situations with no breaks, and it is appropriate for forecasting when the data are generated

109
Mathematics 2018, 6, 34

by a stable model. For each class of models, this method produces out-of-sample forecasts using a
recursive expanding estimation window.
S&P 500 HAR S&P 500 LHAR S&P 500 AHAR

2.0
2.0
2.0
Empirical fluctuation process

Empirical fluctuation process

Empirical fluctuation process

1.5
1.5
1.5

1.0
1.0
1.0

0.5
0.5
0.5

0.0
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Time Time Time

DOW JONES INDUSTRIAL AVERAGE HAR DOW JONES INDUSTRIAL AVERAGE LHAR DOW JONES INDUSTRIAL AVERAGE AHAR
2.0

2.0
2.0

1.5

1.5
Empirical fluctuation process

Empirical fluctuation process

Empirical fluctuation process

1.5

1.0

1.0
1.0

0.5

0.5
0.5

0.0
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Time Time Time

RUSSELL 2000 HAR RUSSELL 2000 LHAR RUSSELL 2000 AHAR

2.0

2.0
2.0

1.5

1.5
Empirical fluctuation process

Empirical fluctuation process

Empirical fluctuation process
1.5

1.0

1.0
1.0

0.5

0.5
0.5

0.0
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Time Time Time

NASDAQ 100 HAR NASDAQ 100 LHAR NASDAQ 100 AHAR

2.0
2.0
1.5

1.5
1.5
Empirical fluctuation process

Empirical fluctuation process

Empirical fluctuation process
1.0

1.0
1.0
0.5

0.5
0.5

0.0
0.0

0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Time Time Time

Figure 2. Empirical fluctuation process of the HAR-RV: (left) the LHAR-RV; (middle) the AHAR-RV;
(right) model parameters for S&P 500, Dow Jones Industrial Average, Russell 2000 and Nasdaq 100.
The red line refers to the boundary with significance level α = 0.05.

110
Mathematics 2018, 6, 34

Common to all the considered forecast combinations is the specification of the minimum
acceptable estimation window size ω. It should not be smaller than the number of regressors plus one;
however, as pointed out by [12] , to account for the very large effect of parameter estimation error, it
should be at least 3 times the number of unknown parameters. For simplicity, the parameter ω has
been held fixed at 40 for all the RV model specifications.
Moreover, for the MSFE weighted average combination, the length of the evaluation window
has been held fixed at 100. This value allows a good estimation of the MSFE and, at the same time,
a non-excessive loss of data at the end of the sample where a change point could be very influential for
the forecasting.
In order to evaluate the quality of the volatility forecasts, the MSE and the QLIKE loss functions
have been considered. They are defined as:

MSEt = (yt − yt )2 (39)


yt yt
QLIKEt = − log −1 (40)
yt yt

where yt is the actual value of the 5-min RV at time t and yt is the corresponding RV forecast. They are
the most widely-used loss functions, and they provide robust ranking of the models in the context of
volatility forecasts [13].
The QLIKE loss is a simple modification of the Gaussian log-likelihood in such a way that the
minimum distance of zero is obtained when yt = yt+1 . Moreover, according to [13], it is able to better
discriminate among models and is less affected by the most extreme observations in the sample.
These loss functions have been used to rank the six competing forecasting methods for each RV
model specification. To this aim, for every method, the average loss and the ratio between its value
and the average loss of the benchmark method have been computed. Obviously, a value of the ratio
below the unit indicates that the forecasting method “beats” the benchmark according to the loss
function metric.
Moreover, to statistically assess if the differences in the performance are relevant, the model
confidence set procedure has been used [14].
This procedure is able to construct a set of models from a specified collection, which consists of
the best models in terms of a loss function with a given level of confidence, and it does not require the
specification of a benchmark. Moreover, it is a stepwise method based on a sequence of significance
tests in which the null hypothesis is that the two models under comparison have the same forecasting
ability against the alternative that they are not equivalent. The test stops when the first hypothesis is
not rejected, and therefore, the procedure does not accumulate Type I error ([14]). The critical values of
the test, as well as the estimation of the variance useful to construct the test statistic are determined by
using the block bootstrap. This re-sampling technique preserves the dependence structure of the series,
and it works reasonably well under very weak conditions on the dependency structure of the data.
In the following, the results obtained for the three different model specifications for the realized
volatility of the four considered series are reported and discussed.
Tables 2–4 provide the results of the analysis for the HAR-RV, LHAR-RV and AHAR-RV model
specification respectively for the four considered series. In particular, they report the average MSE and
the average QLIKE for each forecasting method, as well as their ratio for an individual forecasting
method to the benchmark expanding window method and the ranking of the considered methods
with respect to each loss function. Moreover, the value of the test statistic of the model confidence set
approach and the associated p-value are also reported.
In the case of the HAR model, for all the considered series and for both loss functions, there is
significant evidence that all the forecast combinations have better forecasting ability with respect to the
expanding window procedure; the ratio values are all less than one. Moreover, for both loss functions,
the best method is the forecast combination with ROC location weights for S&P 500 and Dow Jones

111
Mathematics 2018, 6, 34

Industrial Average and with location weights for Russell 2000 and Nasdaq 100; this result confirms the
importance of placing heavier weights on the forecast based on more recent data.

Table 2. Out of sample forecasting result for the HAR-RV model. ROC, Reverse Ordered
Cumulative Sum.

S&P 500
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.5166 1.0000 6 2.154 (0.00) 0.4082 1.0000 6 2.035 (0.00)
MSFE weights 0.5016 0.9710 5 1.010 (1.00) 0.3879 0.9500 4 0.610 (1.00)
ROC weights 0.4986 0.9653 2 −0.956 (1.00) 0.3825 0.9370 2 −0.753 (1.00)
ROC location weights 0.4980 0.9639 1 −1.410 (1.00) 0.3794 0.9294 1 −1.534 (1.00)
Equal weights 0.5015 0.9708 4 0.917 (1.00) 0.3901 0.9557 5 1.191 (0.99)
Location weights 0.5007 0.9694 3 0.437 (1.00) 0.3874 0.9489 3 0.492 (1.00)
DJIA
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.6182 1.0000 6 2.112 (0.00) 0.5610 1.0000 6 2.007 (0.00)
MSFE weights 0.6088 0.9849 5 1.344 (0.97) 0.5410 0.9643 4 0.441 (1.00)
ROC weights 0.6066 0.9813 3 −0.213 (1.00) 0.5388 0.9603 2 −0.102 (1.00)
ROC location weights 0.6046 0.9781 1 −1.599 (1.00) 0.5319 0.9480 1 −1.789 (1.00)
Equal weights 0.6079 0.9834 4 0.690 (1.00) 0.5441 0.9699 5 1.208 (1.00)
Location weights 0.6066 0.9813 2 −0.220 (1.00) 0.5402 0.9629 3 0.252 (1.00)
RUSSELL 2000
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.3491 1.0000 6 1.904 (0.00) 0.2066 1.0000 6 1.649 (0.00)
MSFE weights 0.3450 0.9881 5 1.235 (0.98) 0.2039 0.9868 4 1.018 (1.00)
ROC weights 0.3445 0.9868 4 0.925 (1.00) 0.2040 0.9872 5 1.071 (0.65)
ROC location weights 0.3424 0.9808 2 −0.548 (1.00) 0.2011 0.9733 2 −0.702 (1.00)
Equal weights 0.3430 0.9826 3 −0.106 (1.00) 0.2025 0.9798 3 0.123 (1.00)
Location weights 0.3410 0.9768 1 −1.500 (1.00) 0.1998 0.9670 1 −1.507 (1.00)
NASDAQ 100
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.3509 1.0000 6 2.061 (0.00) 0.2397 1.0000 6 1.897 (0.00)
MSFE weights 0.3448 0.9825 4 1.007 (1.00) 0.2307 0.9697 5 1.106 (1.00)
ROC weights 0.3449 0.9829 5 1.130 (1.00) 0.2319 0.9676 4 0.907 (1.00)
ROC location weights 0.3433 0.9784 3 −0.229 (1.00) 0.2285 0.9535 2 −0.484 (1.00)
Equal weights 0.3432 0.9781 2 −0.319 (1.00) 0.2300 0.9595 3 0.108 (1.00)
Location weights 0.3417 0.9738 1 −1.578 (1.00) 0.2257 0.9417 1 −1.631 (1.00)
Note: For both the loss functions (MSE and QLIKE), the entries are: the values of the average loss; the ratio of
the average loss to that of the expanding window method; the rank (rk) according to the average loss function;
the statistic and the p-value of the Model Confidence Set (MCS) procedure with α = 0.10. A bold entry denotes
the value of the smallest average loss.

The model confidence set has the same structure for all four series and for both loss functions;
it excludes only the forecast generated by the expanding window procedure and includes all the
forecast combinations.
For the LHAR model, focusing on the results of S&P 500 and Dow Jones Industrial Average,
which have very similar behavior, the forecast combination with MSFE weights offers the best
improvement in forecasting accuracy according to the MSE metric, while the forecast combination
with ROC weights according to the QLIKE metric.
For Russell 2000, the forecast combination with ROC location weights beats all the competing
models according to MSE loss function, while, under the QLIKE, the best method is the forecast

112
Mathematics 2018, 6, 34

combination with location weights. This last method has better performance with respect to both of
the loss function metrics for Nasdaq 100.

Table 3. Out of sample forecasting result for the LHAR-RV model.

S&P 500
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.4247 1.0000 5 1.776 (0.00) 0.2858 1.0000 6 2.143 (0.00)
MSFE weights 0.4176 0.9834 1 −1.258 (1.00) 0.2666 0.9328 3 1.316 (0.01)
ROC weights 0.4183 0.9851 2 −0.617 (1.00) 0.2657 0.9298 1 −0.632 (1.00)
ROC location weights 0.4204 0.9899 4 1.275 (0.13) 0.2658 0.9301 2 0.618 (0.55)
Equal weights 0.4197 0.9882 3 0.599 (1.00) 0.2678 0.9370 4 1.557 (0.00)
Location weights 0.4260 1.0031 6 1.536 (0.00) 0.2695 0.9430 5 1.673 (0.00)
DJIA
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.5291 1.0000 5 1.191 (0.09) 0.4236 1.0000 6 1.998 (0.00)
MSFE weights 0.5255 0.9932 1 −1.181 (1.00) 0.4012 0.9471 3 1.368 (0.00)
ROC weights 0.5260 0.9942 2 −0.844 (1.00) 0.3996 0.9433 1 −0.913 (1.00)
ROC location weights 0.5290 0.9998 4 1.092 (0.57) 0.3999 0.9440 2 0.912 (0.44)
Equal weights 0.5268 0.9957 3 −0.311 (1.00) 0.4063 0.9592 4 1.633 (0.00)
Location weights 0.5325 1.0063 6 1.783 (0.00) 0.4091 0.9658 5 1.522 (0.00)
RUSSELL 2000
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.3096 1.0000 6 2.082 (0.00) 0.1828 1.0000 6 1.564 (0.00)
MSFE weights 0.3058 0.9877 5 1.783 (0.00) 0.1821 0.9964 5 1.464 (0.00)
ROC weights 0.3044 0.9833 4 0.777 (1.00) 0.1809 0.9895 4 1.334 (0.00)
ROC location weights 0.3035 0.9805 1 −1.676 (1.00) 0.1794 0.9812 2 −0.989 (1.00)
Equal weights 0.3044 0.9832 3 0.686 (1.00) 0.1804 0.9867 3 1.268 (0.00)
Location weights 0.3042 0.9827 2 0.210 (1.00) 0.1789 0.9786 1 −0.993 (1.00)
NASDAQ 100
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.3180 1.0000 6 2.083 (0.00) 0.2070 1.0000 6 1.984 (0.00)
MSFE weights 0.3112 0.9787 4 1.058 (0.99) 0.1979 0.9562 5 1.066 (1.00)
ROC weights 0.3113 0.9788 5 1.086 (0.99) 0.1977 0.9551 4 0.970 (1.00)
ROC location weights 0.3099 0.9747 3 −0.034 (1.00) 0.1947 0.9406 2 −0.315 (1.00)
Equal weights 0.3093 0.9726 2 −0.586 (1.00) 0.1953 0.9436 3 −0.047 (1.00)
Location weights 0.3082 0.9691 1 −1.516 (1.00) 0.1915 0.9252 1 −1.665 (1.00)
Note: For both loss functions (MSE and QLIKE), the entries are: the values of the average loss; the ratio of the
average loss to that of the expanding window method; the rank (rk) according to the average loss function; the
statistic and the p-value of the Model Confidence Set procedure with α = 0.10. A bold entry denotes the value
of the smallest average loss.

By looking at the MSE ratios for S&P 500 and Dow Jones Industrial Average Index, it is evident
that the forecast combination with location weights is unable to beat the expanding window procedure;
for all the others, the combinations are able to outperform it consistently.
In the model confidence set, when the MSE loss function is used, the expanding window is
eliminated from the model confidence set for all the series. However, the forecast combination with
location weights for S&P 500 and Dow Jones Industrial Average and that with MSFE weights for Russell
2000 are also eliminated. A quite different situation arises when the loss function QLIKE is considered.
In this case, the only surviving models in the model confidence set are the two combinations based
on ROC statistics for S&P 500 and Dow Jones Industrial Average and those based on ROC location
weights and on location weights for Russell 2000. For Nasdaq 100, all the combinations have the

113
Mathematics 2018, 6, 34

same forecasting accuracy. Excluding the last case, the QLIKE loss function, as pointed out previously,
seems to better discriminate among forecasting methods.

Table 4. Out of sample forecasting result for the AHAR-RV model.

S&P 500
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.4576 1.0000 6 2.186 (0.00) 0.3315 1.0000 6 2.151 (0.00)
MSFE weights 0.4397 0.9608 5 1.230 (0.94) 0.3072 0.9267 5 1.268 (0.00)
ROC weights 0.4371 0.9552 2 −0.860 (1.00) 0.3043 0.9179 3 1.393 (0.00)
ROC location weights 0.4364 0.9537 1 −1.416 (1.00) 0.3012 0.9085 1 −0.973 (1.00)
Equal weights 0.4391 0.9596 4 0.781 (1.00) 0.3063 0.9240 4 1.413 (0.00)
Location weights 0.4385 0.9582 3 0.252 (1.00) 0.3016 0.9097 2 −0.977 (0.38)
DJIA
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.5669 1.0000 6 2.189 (0.00) 0.4718 1.0000 6 2.091 (0.00)
MSFE weights 0.5534 0.9762 5 1.015 (0.99) 0.4418 0.9365 5 0.986 (1.00)
ROC weights 0.5529 0.9754 3 0.291 (1.00) 0.4404 0.9335 3 0.688 (1.00)
ROC location weights 0.5516 0.9731 1 −1.744 (1.00) 0.4308 0.9132 1 −1.337 (1.00)
Equal weights 0.5532 0.9758 4 0.679 (0.99) 0.4407 0.9341 4 0.749 (1.00)
Location weights 0.5526 0.9748 2 −0.219 (1.00) 0.4320 0.9158 2 −1.083 (1.00)
RUSSELL 2000
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.3193 1.0000 6 1.964 (0.00) 0.1898 1.0000 6 1.706 (0.00)
MSFE weights 0.3132 0.9811 5 1.498 (0.00) 0.1868 0.9842 5 1.348 (0.00)
ROC weights 0.3113 0.9750 4 1.183 (0.19) 0.1854 0.9769 4 1.385 (0.00)
ROC location weights 0.3086 0.9665 2 −0.735 (1.00) 0.1824 0.9610 2 0.995 (0.20)
Equal weights 0.3107 0.9731 4 0.760 (1.00) 0.1849 0.9741 3 1.385 (0.00)
Location weights 0.3079 0.9644 1 −1.204 (1.00) 0.1818 0.9579 1 −0.995 (1.00)
NASDAQ 100
MSE Ratio rk MCS QLIKE Ratio rk MCS
Expanding window 0.3299 1.0000 6 2.078 (0.00) 0.2223 1.0000 6 1.992 (0.00)
MSFE weights 0.3207 0.9719 4 0.825 (1.00) 0.2122 0.9545 4 0.936 (1.00)
ROC weights 0.3213 0.9738 5 1.212 (1.00) 0.2125 0.9560 5 1.073 (1.00)
ROC location weights 0.3192 0.9675 3 −0.092 (1.00) 0.2088 0.9394 2 −0.429 (1.00)
Equal weights 0.3189 0.9665 2 −0.298 (1.00) 0.2101 0.9451 3 0.085 (1.00)
Location weights 0.3167 0.9599 1 −1.643 (1.00) 0.2058 0.9257 1 −1.651 (1.00)
Note: For both loss functions (MSE and QLIKE), the entries are: the values of the average loss; the ratio of the
average loss to that of the expanding window method; the rank (rk) according to the average loss function; the
statistic and the p-value of the Model Confidence Set (MCS) procedure with α = 0.10. A bold entry denotes the
value of the smallest average loss.

Finally, in the case of the AHAR model, in line with the previous results, the expanding window
appears to offer the worst forecasting performance overall.
Moreover, for both MSE and QLIKE loss functions, the method that offers the major improvement
in forecasting accuracy is the forecast combination with ROC location weights for S&P 500 and Dow
Jones Industrial Average Index and the the forecast combination with location weights for Russell 2000
and Nasdaq 100.
Focusing on the model confidence set, for the MSE loss function, the expanding window is
always eliminated for all the series together with the forecast combination with MSFE weights for
Russell 2000. With respect to the QLIKE loss function, for Dow Jones Industrial Average Index and
Nasdaq 100, the only excluded method is the expanding window procedure. For the other series, the

114
Mathematics 2018, 6, 34

results confirm the better discriminative property of the QLIKE metric; the only surviving methods are
forecast combination with ROC location weights and with location weights.
In conclusion, even if it is not clear which combination has the best forecasting performance,
the forecast combination with ROC location weights and that with location weights seems to be always
among the best methods. However, the forecast combination with ROC location weights always
outperforms the expanding window method, and it is always in the top position with respect to the
loss function ratio and is never excluded by the model confidence set.

6. Concluding Remarks
This paper has explored the relevance of taking into account the presence of structural breaks
in forecasting realized volatility. The analysis has been based on 5-min realized volatility of four U.S.
stock market indices: S&P 500, Dow Jones Industrial Average, Russell 2000 and Nasdaq 100. Three
different model specifications of the log-realized volatility have been considered. For all the considered
market indices, the instability in the parameters of the RV models has been verified through the
recursive estimates test. In order to handle this problem, five forecast combinations, based on different
estimation windows with alternative weighting schemes, have been introduced and compared with
the expanding window method, a natural choice when the data are generated by a stable model. The
forecasting performance has been evaluated, for each RV model specification, through two of the most
relevant loss functions, the MSE and the QLIKE. Moreover, to this aim, the average loss function has
been calculated, and in order to statistically assess if the differences in the performance are relevant,
the model confidence set approach has been considered.
The analysis, repeated for each class of RV models separately, has highlighted the importance
of taking into account structural breaks; in fact, the expanding window appears to offer the worst
forecasting performance overall. In particular, in almost all the considered cases, the two combinations
that make adjustments for accounting for the most recent possible break point (the forecast combination
with location weights and with ROC location weights) are placed in first position and, as a consequence,
have better forecasting performance. Nevertheless, the forecast combination with ROC location weights
always outperforms the expanding window method; it is always in the top position with respect to the
loss function ratio, and it is never excluded by the model confidence set.

Conflicts of Interest: The author declares no conflict of interest.

References
1. Liu, C.; Maheu, J.M. Are there structural breaks in Realized Volatility? J. Financ. Econom. 2008, 6, 326–360.
2. Kumar, D. Realized volatility transmission from crude oil to equity sectors: A study with economic
significance analysis. Int. Rev. Econ. Financ. 2017, 49, 149–167.
3. Gong, X.; Lin, B. Structural breaks and volatility forecasting in the copper futures market. J. Future Mark.
2017, doi:10.1002/fut.21867.
4. Ma, F.; Wahab, M.I.M.; Huang, D.; Xu, W. Forecasting the realized volatility of the oil futures market:
A regime switching approach. Energy Econ. 2017, 67, 136–145.
5. Wang, Y.; Pan, Z.; Wu, C. Time-Varying Parameter Realized Volatility Models. J. Forecast. 2017, 36, 566–580.
6. Corsi, F. A Simple Long Memory Model of Realized Volatility. Woking Paper, University of Southern
Switzerland, Manno, Switzerland, 2004.
7. Corsi, F.; Renó, R. HAR Volatility Modelling with Heterogeneous Leverage and Jumps; University of Siena: Siena,
Italy, 2010.
8. Corsi, F. A Simple Approximate Long-Memory Model of Realized Volatility. J. Financ. Econom. 2009, 7,
174–196.
9. Sevi, B. Forecasting the volatility of crude oil futures using intraday data. Eur. J. Oper. Res. 2014, 235, 643–659.
10. Prokopczuk, M.; Symeonidis, L.; Wese Simen, C. Do jumps matter for volatility forecasting? evidence from
energy markets. J. Future Mark. 2016, 36, 758–792.

115
Mathematics 2018, 6, 34

11. Ploberger, W.; Krämer, W.; Kontrus, K. A new test for structural stability in the linear regression model.
J. Econom. 1989, 40, 307–318.
12. Pesaran, M.H.; Timmermann, A. Selection of estimation windows in the presence of breaks. J. Econom. 2007,
137, 134–161.
13. Patton, A.J. Volatility forecast comparison using imperfect volatility proxies. J. Econom. 2011, 160, 246–256.
14. Hansen, P.R.; Lunde, A.; Nason, J.M. The model confidence set. Econometrica 2011, 79, 453–497.
15. Liu, L.Y.; Patton, A.J.; Sheppard, K. Does anything beat 5-min RV? A comparison of realized measures across
multiple asset classes. J. Econom. 2011, 187, 293–311.
16. Andreou, E.; Ghysels, E.; Kourouyiannis, C. Robust Volatility Forecasts in the Presence of Structural Breaks.
Woking Paper, University of Cyprus, Nicosia, Cyprus, 2012.
17. Andersen, T.G.; Bollerslev, T.; Diebold, F.X. Roughing It Up: Including Jump Components in the
Measurement, Modeling anf Forecasting of Returns Volatility. Rev. Econ. Stat. 2007, 89, 701–720.
18. Asai, M.; McAleer, M.; Medeiros, M.C. Asymmetry and long memory in volatility modeling. J. Financ. Econom.
2012, 10, 495–512.
19. Chow, G.C. Tests of Equality Between Sets of Coefficients in Two Linear Regressions. Econometrica 1960, 28,
591–605.
20. Andrews, D.W.K.; Fair, R.C. Inference in Nonlinear Econometric Models with Structural Change.
Rev. Econ. Stud. 1988, 55, 615–639.
21. Hansen, B.E. The new econometrics of structural change: Dating breaks in US labor productivity.
J. Econ. Perspect. 2001, 15, 117–128.
22. Perron, P. Dealing with structural breaks. In Palgrave Handbook of Econometrics, Vol. 1: Econometric Theory;
Patterson, K., Mills, T.C., Eds.; Palgrave Macmillan: Basingstote, UK, 2006; pp. 278–352.
23. Bai, J.; Perron, P. Estimating and testing linear models with multiple structural changes. Econometrica 1998,
66, 47–78.
24. Bai, J.; Perron, P. Computation and analysis of multiple structural change models. J. Appl. Econom. 2003, 18,
1–22.
25. Bai, J.; Perron, P. Critical values for multiple structural change tests. Econom. J. 2003, 6, 72–78.
26. Zeileis, A.; Kleiber, C.; Kramer, W.; Hornik, K. Testing and dating of structural changes in practice.
Comput. Stat. Data Anal. 2003, 44, 109–123.
27. Kuan, C.M.; Hornik, K. The generalized fluctuation test: A unifying view. Econom. Rev. 1995, 14, 135–161.
28. Brown, R.L.; Durbin, J.; Evans, J.M. Techniques for Testing the Constancy of Regression Relationships over
Time. J. R. Stat. Soc. 1975, 37, 149–163.
29. Ploberger, W.; Krämer, W. The CUSUM Test With OLS Residuals. Econometrica 1992, 60, 271–285.
30. Chu, C.S.J.; Hornik, K.; Kuan, C.M. MOSUM Tests for Parameter Constancy. Biometrika 1995, 82, 603–617.
31. Chu, C.S.J.; Hornik, K.; Kuan, C.M. The Moving-Estimates Test for Parameter Stability. Econom. Theory 1995,
11, 669–720.
32. Billingsley, P. Convergence of Probability Measures, 2nd ed.; John Wiley & Sons: New York, NY, USA, 2013.
33. Clements, M.P.; Hendry, D.F. Forecasting with breaks. In Handbook of Economic Forecasting; Elliot, G.,
Granger, C.W.J., Timmermann, A., Eds.; Elsevier: Amsterdam, The Netherlands 2006, 605–657.
34. Pesaran, M.H.; Pick, A.; Pranovich, M. Optimal forecasts in the presence of structural breaks. J. Econom. 2013,
177, 134–152.
35. Giraitis, L.; Kapetanios, G.; Price, S. Adaptive forecasting in the presence of recent and ongoing structural
change. J. Econom. 2013, 177, 153–170.
36. Inoue, A.; Jin, L.; Rossi, B. Rolling window selection for out-of-sample forecasting with time-varying
parameters. J. Econom. 2017, 196, 55–67.
37. Timmermann, A. Forecast combinations. In Handbook of Economic Forecasting; Elliot, G., Granger, C.W.J.,
Timmermann, A., Eds.; Elsevier: Amsterdam, The Netherlands, 2006; pp. 135–196.
38. Tian, J.; Anderson, H.M. Forecast Combinations under Structural Break Uncertainty. Int. J. Forecast. 2014, 30,
161–175.
39. Assenmacher-Wesche, K.; Pesaran, M.H. Forecasting the Swiss economy using VECX models: An exercise in
forecast combination across models and observation windows. Natl. Inst. Econ. Rev. 2008, 203, 91–108.
40. Pesaran, M.H.; Schuermann, T.; Smith, L.V. Forecasting economic and financial variables with global VARs.
Int. J. Forecast. 2009, 25, 642–675.

116
Mathematics 2018, 6, 34

41. Schrimpf, A.; Wang, Q. A reappraisal of the leading indicator properties of the yield curve under structural
instability. Int. J. Forecast. 2010, 26, 836–857.
42. Pesaran, M.H.; Pick, A. Forecast combination across estimation windows. J. Bus. Econ. Stat. 2011, 29, 307–318.
43. Clark, T.E.; McCracken, M.W. Improving forecast accuracy by combining recursive and rolling forecasts.
Int. Econ. Rev. 2009, 50, 363–395.
44. Pesaran, M.H.; Pettenuzzo, D.; Timmermann, A. Forecasting time series subject to multiple structural breaks.
Rev. Econ. Stud. 2006, 73, 1057–1084.
45. Maheu, J.M.; Gordon, S. Learning, forecasting and structural breaks. J. Appl. Econom. 2008, 23, 553–583.
46. Pesaran, M.H.; Timmermann, A. Market Timing and Return Prediction under Model Instability.
J. Empir. Financ. 2002, 9, 495–510.

c 2018 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

117
 mathematics
Article
Stochastic Comparisons and Dynamic Information of
Random Lifetimes in a Replacement Model
Antonio Di Crescenzo and Patrizia Di Gironimo *
Dipartimento di Matematica, Università degli Studi di Salerno, Via Giovanni Paolo II n. 132,
84084 Fisciano (SA), Italy; [email protected]
* Correspondence: [email protected]

Received: 27 September 2018; Accepted: 14 October 2018; Published: 16 October 2018

Abstract: We consider a suitable replacement model for random lifetimes, in which at a fixed
time an item is replaced by another one having the same age but different lifetime distribution.
We focus first on stochastic comparisons between the involved random lifetimes, in order to assess
conditions leading to an improvement of the system. Attention is also given to the relative ratio
of improvement, which is proposed as a suitable index finalized to measure the goodness of the
replacement procedure. Finally, we provide various results on the dynamic differential entropy of the
lifetime of the improved system.

Keywords: reliability; stochastic orders; scale family of distributions; proportional hazard rates;
differential entropy

MSC: 62N05 ; 60E15 ; 94A17

1. Introduction
In reliability theory, various stochastic models have been proposed in the past in order to describe
replacement policies of system components. A classical model in this area is the relevation transform,
which describes the overall lifetime of a component which is replaced at its (random) failure time by
another component of the same age, whose lifetime distribution is possibly different. See the paper by
Krakowski [1] that introduced this topic, and the further contributions by Baxter [2], Belzunce et al. [3],
Chukova et al. [4], Shanthikumar and Baxter [5], Sordo and Psarrakos [6], for instance. A similar
model, named reversed relevation transform, has been considered in Di Crescenzo and Toomaj [7]
in order to describe the total lifetime of an item given that it is less than an independent random
inspection time. Such transforms deserve a large interest since they can be employed in restoration
models of failed units, or in the determination of optimal redundancy policy in coherent systems.
In both cases, the above models involve a random replacement (or inspection) time. In this paper,
we aim to consider a different stochastic model dealing with replacement occurring at deterministic
arbitrary instants. Specifically, we assume that an item having random lifetime X is planned to be
replaced at time t by another item having the same age but possibly different random lifetime Y.
The main aim is to investigate the effect of the replacement, with emphasis on criteria leading to attain
better performance for the overall system.
The tools adopted in our investigation are based on stochastic orders and other typical notions of
reliability theory. Specifically, we study the consequence of suitable assumptions by which the initial
lifetime X is smaller than the lifetime Y of the replacing item according to some stochastic criteria.
We also propose a suitable index finalized to assess the effective improvement gained by the system
due to the replacement. In addition, we aim to propose the residual differential entropy as a dynamic
measure of the information content of the replacement model.

Mathematics 2018, 6, 204; doi:10.3390/math6100204 118 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 204

This is the plan of the paper: in Section 2, we introduce the stochastic model and the main
quantities of interest. Section 3 is devoted to establish some stochastic comparisons concerning the
proposed model. We also deal with the case when the relevant random variables belong to the same
scale family of distributions. In Section 4, we introduce the relative ratio of improvement for the
considered model, and investigate its behavior in some examples. Section 5 is centered on results on
two dynamic versions of the differential entropy for the proposed model, with reference to the residual
and past entropy. Finally, some concluding remarks are given in Section 6.
Throughout the paper, as usual, we denote by [ X | B] a random variable having the same
distribution of X conditional on B. The expectation of X is denoted by E[ X ]. Moreover, 1 A is the
sgn
indicator function of A, i.e., 1 A = 1 if A is true, and 1 A = 0, otherwise. Furthermore, a = b means
d
that a and b have the same sign, and = means equality in distribution.

2. The Stochastic Model

Let X be an absolutely continuous nonnegative random variable with cumulative distribution
function (CDF) F (t) = P( X ≤ t), probability density function (PDF) f (t), and survival function
F (t) = 1 − F (t). Bearing in mind possible applications to reliability theory and survival analysis,
we assume that X describes the random lifetime of an item or a living organism. Let us now recall two
functions of interest; as usual, we denote by

d f (t)
λ X (t) = − log F (t) = , t ∈ R+ , F ( t ) > 0 (1)
dt F (t)

the hazard rate (or failure rate) of X, and by

d f (t)
τX (t) = log F (t) = , t ∈ R+ , F ( t ) > 0 (2)
dt F (t)

the reversed hazard rate function of X. See Barlow and Proschan [8] and Block et al. [9] for some
illustrative results on these notions. Denote by Y another absolutely continuous nonnegative random
variable with CDF G (t), PDF g(t), survival function G (t), hazard rate λY (t) and reversed hazard
rate τY (t).
We assume that X and Y are independent lifetimes of systems or items, both starting to work at
time 0. A replacement of the first item by the second one (having the same age) is planned to occur at
time t, provided that the first item is not failed before. Let us now define It = 1{0≤ X ≤t} , so that It is
a Bernoulli random variable with parameter P( It = 1) = F (t). Hence, denoting by XtY the random
duration of the (eventually replaced) system, it can be expressed as follows:

[ X | X ≤ t] if 0 ≤ X ≤ t
XtY = [ X | X ≤ t] It + [Y |Y > t](1 − It ) = ( t ∈ R+ ). (3)
[Y |Y > t] if X > t

In classical minimal repair models, an item, upon failure, is replaced by another item having
the same failure distribution, and the same age of the previous item at the failure time. The present
model also presumes that the item is replaced by another one having the same age at the failure time.
The difference is that the replacement occurs at a preassigned deterministic time t, and that the new
item possesses a possibly different failure distribution.
By (3), for any Borel set B, the following mixture holds:

P( XtY ∈ B) = P( X ∈ B| X ≤ t) F (t) + P(Y ∈ B|Y > t) F (t), t ∈ R+ . (4)

119
Mathematics 2018, 6, 204

Then, the CDF and the PDF of XtY are respectively


F ( x ), if 0 ≤ x ≤ t,
FtY ( x ) = P( XtY ≤ x) = F (t) (5)
F (t) + G (t)
[G(x) − G (t)], if x > t,

and 
d Y f ( x ), if 0 ≤ x ≤ t
f tY ( x ) = F (x) = F (t) (6)
dx t G (t)
g ( x ), if x > t,

so that the survival function of XtY can be expressed as


Y F ( x ), if 0 ≤ x ≤ t,
Ft ( x) = F (t) (7)
F (t) + G (t)
[G(x) − G (t)], if x > t.

For instance, the special case when Y is uniformly distributed on [0, 1] is treated in Example 4.18
of Santacroce et al. [10] concerning the analysis of some exponential models.
It should be pointed out that the replacement model given in (3) can be easily extended to
integer-valued random variables. In this case, the model when Y is uniformly distributed on a set of
integers is of interest in information theory (see, for instance, the operator considered in Equation (3)
of Cicalese et al. [11]).
A relevant issue about model (3) is the following: to assess if the replacement planned at time
t is beneficial to the system. This can be attained in several ways. As a first step, hereafter, we face
the problem of establishing if the duration of the replaced system is larger than that of the originating
lifetime (or smaller than that of the replaced lifetime) in some stochastic sense. To this aim, in the
following section, we provide various useful comparisons based on stochastic orders.

3. Stochastic Comparisons

3.1. Definitions and Main Comparisons

In order to compare X and Y with XtY , let us now recall some well-known definitions of partial
stochastic orders, which involve the notions treated in Section 2. As a reference, see Shaked and
Shanthikumar [12] or Belzunce et al. [13].

Definition 1. Let X be an absolutely continuous random variable with support (lX , u X ), CDF F, and PDF f .
Similarly, let Y be an absolutely continuous random variable with support (lY , uY ), CDF G, and PDF g. We
say that X is smaller than Y in the

(a) usual stochastic order (X ≤st Y) if F (t) ≤ G (t) ∀ t ∈ R or, equivalently, if F (t) ≥ G (t) ∀ t ∈ R;
(b) hazard rate order (X ≤hr Y) if G (t)/F (t) increases in t ∈ (−∞, max(u X , uY )) or, equivalently, if
λ X (t) ≥ λY (t) for all t ∈ R, where λ X (t) = f (t)/F (t) and λY (t) = g(t)/G (t) are respectively the
hazard rates of X and Y, or equivalently if f ( x ) G (y) ≥ g( x ) F (y) ∀ x ≤ y;
(c) likelihood ratio order (X ≤lr Y) if f ( x ) g(y) ≥ f (y) g( x ) for all x ≤ y, with x, y ∈ (lX , u X ) ∪ (lY , uY ) or,
equivalently, g(t)/ f (t) increases in t over the union of supports of X and Y;
(d) reversed hazard rate order (X ≤rh Y) if G (t)/F (t) increases in t ∈ (min(lX , lY ), +∞) or, equivalently,
if τX (t) ≤ τY (t) for all t ∈ R, where τX (t) = f (t)/F (t) and τY (t) = g(t)/G (t) are respectively the
reversed hazard rates of X and Y.

We recall the following relations among the above-mentioned stochastic orders:

X ≤lr Y ⇒ X ≤hr Y ⇒ X ≤st Y, X ≤lr Y ⇒ X ≤rh Y ⇒ X ≤st Y. (8)

120
Mathematics 2018, 6, 204

With reference to (3), we now investigate the effect of the replacement when the lifetime of the
first item is stochastically smaller than the second in the sense of the criteria given in Definition 1.

Theorem 1. Let X and Y be absolutely continuous nonnegative random variables. Then,

(i) X ≤hr Y ⇒ X ≤hr XtY ≤hr Y ∀t > 0;

(ii) X ≤lr Y ⇒ XtY ≤lr Y ∀t > 0;
(iii) X ≤rh Y ⇒ XtY ≤rh Y ∀t > 0;
(iv) X ≤st Y ⇔ XtY ≤st Y ∀t > 0.

Proof. From (6) and (7), the hazard rate of XtY is given by

f tY ( x ) λ X ( x ), if 0 ≤ x ≤ t,
λYt ( x ) = Y
= (9)
Ft ( x) λY ( x ) , if x > t.

We observe that, if X ≤hr Y, from (9), we immediately deduce λ X ( x ) ≥ λYt ( x ) ≥ λY ( x ) for all
x ≥ 0, and so we obtain X ≤hr XtY ≤hr Y ∀t > 0.
Now, by taking into account Equation (6), we get
⎧ g( x )
g( x ) ⎨ f (x)
, if 0 ≤ x ≤ t,
=
Y
f t (x) ⎩ G (t)
, if x > t.
F (t)

g(t)
Hence, since assumption X ≤lr Y implies that g( x )/ f ( x ) is increasing in x > 0, and that f (t)
≤
G (t)
F (t)
for all t > 0 by the first of (8), we finally obtain XtY ≤lr Y, this completing the proof of (ii).
Note that ⎧ F(x)
FtY ( x ) ⎨ G(x)
, if 0 ≤ x ≤ t,
=
G(x) ⎩ F (t)
+ F (t) [ G ( x )− G (t)]
, if x > t.
G(x) G (t) G(x)

Hence, for 0 ≤ x ≤ t, we have that FtY ( x )/G ( x ) is decreasing in x if and only if X ≤rh Y.
Moreover FtY ( x )/G ( x ) is continuous in x = t. Finally, it is not hard to see that the derivative of
FtY ( x )/G ( x ) is nonpositive if G (t) ≤ F (t) for all t ≥ 0, i.e., X ≤st Y, this being ensured by assumption
X ≤rh Y. The proof of (iii) is thus completed.
The proof of (iv) can be easily checked from (5), by seeing that FtY ( x ) ≥ G ( x ) for all x ≥ 0 and
t ≥ 0, if and only if assumption X ≤st Y holds.

Differently from case (i) of Theorem 1, condition X ≤lr Y does not imply that X ≤lr XtY ∀t > 0.
This can be easily checked, for instance, when X and Y are exponentially distributed with rates λ X
and λY , with λ X > λY . In this case, one has X ≤lr Y, whereas, recalling (6), the ratio f tY ( x )/ f ( x ) is not
increasing for all x > 0, and thus X ≤lr XtY is not true. A similar conclusion holds for the cases (ii) and
(iii). Indeed, in the following counterexample, we see that

X ≤st Y ⇒ X ≤st XtY ∀t > 0,

X ≤rh Y ⇒ X ≤rh XtY ∀t > 0.

Counterexample 1. Let X be exponentially distributed with parameter 1, and let Y = max{ X, Z },

where Z is Erlang distributed with parameters (2, 2) and is independent from X. Hence, since
F ( x ) = 1 − e− x , x ≥ 0, and G ( x ) = F ( x ) H ( x ), with H ( x ) = 1 − (1 + 2x )e−2x , x ≥ 0, we immediately
have that X ≤rh Y, and thus X ≤st Y. However, recalling (5), it is not hard to see that FtY ( x )/F ( x ) is

121
Mathematics 2018, 6, 204

not monotonic and is not smaller than one for suitable choices of t, as shown in Figure 1. Hence, both
the conditions X ≤st XtY ∀t > 0 and X ≤rh XtY ∀t > 0 are not true.

Figure 1. Plot of FtY ( x )/F ( x ) for t = 1, with reference to Counterexample 1.

Let us now prove another result concerning stochastic orderings similar to those of Theorem 1.

Theorem 2. Let X, Y be random lifetimes. If X ≤st XtY ∀t > 0, then X ≤hr Y.

Proof. From assumption X ≤st XtY ∀t > 0, we have

F (t)
F ( x ) ≥ F (t) + [ G ( x ) − G (t)], x > t > 0,
G (t)

so that
F ( x ) − F (t) 1 G ( x ) − G (t) 1
≥ .
x−t F (t) x−t G (t)
f (t) g(t)
In the limit as x ↓ t, we have F (t)
≥ G (t)
for all t > 0, thus X ≤hr Y.

3.2. Scale Family of Distributions

Engineers in the manufacturing industries have used accelerated test experiments for many
decades (see Arnold et al. [14], Escobar and Meeker [15], for instance). Various models for accelerated
test involve time-transformations of suitable functions. The simplest case is based on linear
transformations and on distribution functions. Then, let us now adapt the model (3) to the instance in
which the distributions of X and Y belong to the same scale family.
Given the random lifetimes X and Y, having distribution functions F ( x ) and G ( x ) respectively,
we assume that X and Y belong to the same scale family of distributions, i.e.,

G ( x ) = F (αx ) ∀ x ∈ R, 0 < α < 1. (10)

Hence, for 0 < α < 1, one has X ≤st Y. We recall that the quantile function of X is given by

Q X (u) = inf{ x ∈ R| F ( x ) ≥ u}, 0 < u < 1.

Assumption (10) means that X and Y satisfy the proportional quantile functions model (see
Section 4.1 of Di Crescenzo et al. [16]) expressed by Q X (u) = αQY (u) ∀u ∈ (0, 1), where QY (u) is
similarly defined as Q X (u). From (5), under the assumption (10) the distribution function of XtY is

F ( x ), if 0 ≤ x ≤ t,
FtY ( x ) = F (t)
F (t) + F (αt)
[ F (αx ) − F (αt)], if x > t.

122
Mathematics 2018, 6, 204

Among the quantities of interest in reliability theory, wide attention is devoted to the residual
lifetime of a given random lifetime X, defined as

Xt : = [ X − t | X > t ], t ≥ 0. (11)

The residual lifetime defined in (11) is involved in the following well-known notion of
positive ageing.

Definition 2. We say that X is IFR (increasing failure rate) if Xt ≤st Xs for all t ≥ s ≥ 0, that is, if F ( x ) is
logconcave, or equivalently the failure rate λ X (t) is increasing in t ≥ 0.

Remark 1. If X and Y satisfy condition (10) and if X is IFR, then, for 0 < α < 1,

λY (t) = αλ X (αt) < λ X (αt) ≤ λ X (t), t ≥ 0,

so that X ≤hr Y, and hence X ≤hr XtY ≤hr Y due to point (i) of Theorem 1.

3.3. Further Results

From (7), the expected value of XtY can be expressed as
 t  ∞
F (t)
E[ XtY ] = F ( x )dx + F (t) + [ G ( x ) − G (t)] dx
0 t G (t)
 ∞
1
= E[ X ] + [ F (t) G ( x ) − F ( x ) G (t)]dx.
G (t) t

Hence, recalling that the mean residual life of the random lifetime X is
 ∞
1
m X (t) = E[ X − t| X > t] = F ( x )dx, t ∈ R+ , F (t) > 0,
F (t) t

with mY (t) similarly defined, we have

E[ XtY ] = E[ X ] + F (t)[mY (t) − m X (t)]. (12)

Let us now recall the mean residual life order (see Section 2.A of Shaked and Shanthikumar [12]).

Definition 3. Let X and Y be absolutely continuous random variables with CDFs F (t) and G (t), and with
finite mean residual lives m X (t) and mY (t), respectively. We say that X is smaller than Y in the mean residual
life order (X ≤mrl Y) if m X (t) ≤ mY (t) for all t or, equivalently, if
∞  ∞
G ( x )dx
t∞ is decreasing over {t : F ( x )dx > 0}.
t F ( x )dx t

Consequently, recalling (12), we immediately have the forthcoming result.

Proposition 1. The relation E[ X ] ≤ E[ XtY ] holds for all t if and only if X ≤mrl Y.

We can now come to a probabilistic analogue of the mean value theorem.

Theorem 3. Let X and Y be non-negative random variables satisfying X ≤hr Y and E[ X ] < E[ XtY ] < ∞ and
let Zt = Ψ( X, Y ). Let g be a measurable and differentiable function such that E[ g( X )] and E[ g(Y )] are finite,

123
Mathematics 2018, 6, 204

and let its derivative g be measurable and Riemann-integrable on the interval [ x, y] for all y ≥ x ≥ 0. Then,
E[ g ( Z )] is finite and
E[ g( XtY )] − E[ g( X )] = E[ g ( Zt )]{E[ XtY ] − E[ X ]}, (13)

where Zt is the absolutely continuous random variable having PDF

Y
 
F (x) − F(x) 1 G(x) F(x)
f Zt ( x ) = t Y = − , x ≥ t.
E[ Xt ] − E[ X ] mY ( t ) − m X ( t ) G (t) F (t)

Proof. The proof follows from the Theorem 4.1 of Di Crescenzo [17].

It is interesting to point out that the relation (13) can be used in various applied contexts.
For instance, if g is an utility function, then E[ g( X )] can be viewed as the expected utility granted by
an item having lifetime X. Accordingly, Equation (13) expresses the variation of the expected utility
when such an item is subject to the replacement procedure described in Section 2. Clearly, it can be
used to construct useful measures able to evaluate the goodness of the procedure. This specific task is
not undertaken here, whereas in the following section we propose a different approach to assess the
effectiveness of the replacement.

4. Relative Ratio of Improvement

Consider a system having random lifetime X, which is replaced by Y at time t. If X is smaller
than Y according to some stochastic order, we expect that the reliability of the system at time x > t > 0
is improved. In order to measure the usefulness of replacing the lifetime X with Y at time t, let us now
introduce the relative ratio of improvement evaluated at x > t > 0. It is defined in terms of (7) as

Y
 
F (x) − F(x) F (t) F (t) G(x) G (t)
Rt ( x ) := t = −1+ − . (14)
F(x) F(x) G (t) F(x) F(x)

Clearly, if X ≤hr Y, then, from point (i) of Theorem 1, it follows that X ≤hr XtY and, in turn,
X ≤st XtY so that Rt ( x ) ≥ 0 for all x > t > 0.

Example 1. Let { Z (t), t ≥ 0} be an iterated Poisson process with parameters (μ, λ). In other
terms, such process can be expressed as Z (t) = M[ N (t)], where M(t) and N (t) are independent
Poisson processes with parameters μ, λ ∈ R+ , respectively (see Section 6 of Di Crescenzo et al. [18]).
Denoting by
Tk = inf{t > 0 : Z (t) ≥ k} (15)

the first crossing time (from below) of Z (t) through the constant level k ∈ N, the corresponding
survival function is (cf. Section 7 of [18])

k −1
μj
P[ Tk > t] = exp{−λ(1 − e−μ )t} ∑ B (λe−μ t), t ≥ 0, (16)
j =0
j! j

where Bj (·) is the j-th Bell polynomial. We consider a system subject to replacement policy as described
in (3), where the relevant random lifetimes are given by the first-crossing times defined in (15),
d d
with X = T1 and Y = Tk . The relative ratio of improvement of this system is then evaluated by means
of (14). Figure 2 provides some plots of Rt ( x ), showing that the relative ratio of improvement is
increasing in x and k.

124
Mathematics 2018, 6, 204

Figure 2. With reference to Example 1, left: plot of Rt ( x ) for 1 < x < 10, with t = 1 and k = 2, 3, 4, 5
(from bottom to top); right: contour plot of Rt ( x ) for 1 < x < 10 and 0 < μ < 5, with k = 2, t = 1 and
λ = 1.

In the remaining part of this section, we restrict our attention to the special case in which X and Y
satisfy the proportional hazard rates model (see Cox [19] or, for instance, the more recent contributions
by Nanda and Das [20], and Ng et al. [21]). Hence, assuming that G (t) = [ F (t)]θ , ∀t ≥ 0, for θ > 0,
θ = 1, the relative ratio defined in (14) becomes
 
F (t) [ F (t)]θ
Rt ( x ) = − 1 + [ F (t)]1−θ [ F ( x )]θ −1 − , x > t > 0. (17)
F(x) F(x)

Here, the most interesting case is for 0 < θ < 1, since this assumption ensures that X ≤hr Y.

Example 2. Let X and Y be exponentially distributed with parameters 1 and θ, respectively, with
0 < θ < 1. Since F (t) = e−t , G (t) = e−θt , t ≥ 0, from (17), we have

Rt ( x ) = e−(t− x)(1−θ ) − 1, x > t > 0.

Some plots of Rt ( x ) are given in Figure 3, confirming that the relative ratio of improvement is
increasing in x − t > 0, and is decreasing in θ ∈ (0, 1).

Figure 3. With reference to Example 2, left: plot of Rt ( x ) for 5 < x < 10, with t = 5 and θ = 0.3, 0.4,
0.5, 0.6, 0.7 (from top to bottom); right: contour plot of Rt ( x ) for 0.5 < θ < 1 and 5 < t < 10, with
x = 10.

125
Mathematics 2018, 6, 204

5. Results on Dynamic Differential Entropies

In this section, we investigate some informational properties of the replacement model considered
in Section 2.
A classical measure of uncertainty for an absolutely continuous random variable X is the
differential entropy, defined as 
HX = − f ( x ) log f ( x )dx, (18)
R
with 0 log 0 = 0 by convention. This measure has some analogies with the entropy of discrete random
variables, even though the differential entropy lacks a number of properties that the Shannon discrete
entropy possesses (see, for instance, Cover and Thomas [22] for details).
In the context of lifetimes truncated over intervals of the form [0, t] or (t, ∞), specific forms of
the differential entropy have been investigated in the recent decades (see the initial contributions
by Muliere et al. [23]). Specifically, the following dynamic measure (named residual entropy of the
lifetime X) has been extensively investigated:
 ∞
f (x) f (x)
HX ( t ) = − log dx, t ∈ R+ , F (t) > 0. (19)
t F (t) F (t)

This quantity is suitable for measuring the uncertainty in residual lifetimes defined as in (11).
The residual entropy of X has been studied by Ebrahimi [24], Ebrahimi and Pellerey [25]; see also
Asadi and Ebrahimi [26], Ebrahimi et al. [27] on this topic. A similar reasoning leads to the past entropy
of X, defined as the differential entropy of the past lifetime [ X | X ≤ t], t > 0, i.e.,
 t
f (x) f (x)
H X (t) = − log dx, t ∈ R+ , F (t) > 0. (20)
0 F (t) F (t)

This measure is also named ‘past entropy’ of X; it has been investigated in Di Crescenzo and
Longobardi [28], Nanda and Paul [29], Kundu et al. [30]. Other results and applications of these
dynamic information measures can be found in Sachlas and Papaioannou [31], Kundu and Nanda [32],
and Ahmadi et al. [33].
Hereafter, we denote by

H(t) = − F (t) log F (t) − F (t) log F (t), t>0 (21)

the partition entropy of X at time t (see Bowden [34]), which measures the information (in the sense of
Shannon entropy) about the value of the random lifetime X derived from knowing whether X ≤ t
or X > t.
Under the conditions specified in Section 2, let us now provide a decomposition result for the
differential entropy of (3).

Proposition 2. For all t > 0, we have

HXY = H(t) + F (t) H X (t) + F (t) HY (t). (22)

t

Proof. From (6) and (18), we have that, for t > 0,

 ∞  t  ∞  F (t)
F (t)
HX Y = − f tY ( x ) log f tY ( x )dx = − f ( x ) log f ( x )dx − g( x ) log g( x ) dx.
t 0 0 t G (t) G (t)

126
Mathematics 2018, 6, 204

Recalling the alternative expression of the residual entropy (19) given in (2.2) of [24], and the
alternative expression of the past entropy (20) shown in (2.1) of [28], we have
 t
− f ( x ) log f ( x )dx = F (t)[ H X (t) − log F (t)],
0

and  ∞
− g( x ) log g( x )dx = G (t)[ HY (t) − log G (t)].
t
Hence, we obtain
 
F (t) F (t)
HX Y = F (t)[ H X (t) − log F (t)] + G (t)[ HY (t) − log G (t)] − G (t) log
t G (t) G (t)
= − F (t) log F (t) − F (t) log F (t) + F (t) H X (t) + F (t) HY (t).

This completes the proof of (22), due to (21).

We note that Equation (22) can be interpreted as follows. The uncertainty about the failure time of
an item having lifetime XtY can be decomposed into three parts: (i) the uncertainty on whether the
item has failed before or after time t, (ii) the uncertainty about the failure time in (0, t) given that the
item has failed before t, and (iii) the uncertainty about the failure time in (t, +∞) given that the item
has failed after t, and thus the failure time is distributed as Y since the replacement occurred at time t.
Clearly, if X and Y are identically distributed, then Equation (22) becomes the identity given in
Proposition 2.1 of [28], i.e.,

HX = H(t) + F (t) H X (t) + F (t) HX (t), t > 0. (23)

The given results allow us to perform some comparisons involving the above entropies. To this
aim, we recall a suitable stochastic order (see Definition 2.1 of Ebrahimi and Pellerey [25]):

Definition 4. Let X and Y be random lifetimes; X is said to have less uncertainty than Y, and write X ≤LU Y, if

HX (t) ≤ HY (t) for all t ≥ 0.

From Proposition 2, we can now infer the following result.

Corollary 1. If X and Y are random lifetimes such that X ≤LU Y, then

HX ≤ HX Y ∀t > 0.
t

Proof. By comparing Equations (22) and (23), we obtain

HXY − HX = F (t)[ HY (t) − HX (t)], t > 0. (24)

t

Thus, from Definition 4, we obtain the proof immediately.

Let us now investigate some sufficient conditions leading to the monotonicity of the differential
entropy of (3). To this aim, we recall the following notion, which was first considered in [24].

Definition 5. Assume that the residual entropy of the random lifetime X is finite. If HX (t) is decreasing
(increasing) in t ≥ 0, we say that X has decreasing (increasing) uncertainty of residual life, i.e., X is
DURL (IURL).

127
Mathematics 2018, 6, 204

Proposition 3.

(i) Let X ≤LU Y. If X is IURL and Y is DURL, then XtY is DURL.

(ii) Let Y ≤LU X. If X is DURL and Y is IURL, then XtY is IURL.
(iii) If λ X (t) ≤ e ≤ λY (t) for all t ≥ 0, and HY (t) ≥ 0 for all t ≥ 0, then XtY is IURL.

Proof.

(i) Under the given assumptions, we have that the right-hand-side of (24) is nonnegative and
decreasing, so that HXY is decreasing.
t
(ii) Differentiating both sides of (24), we obtain:

d * +
H Y = F (t) λY (t)[ HY (t) − 1 + log λY (t)] − λ X (t)[ HY (t) − 1 + log λ X (t)] . (25)
dt Xt

From Theorem 3.2 of Ebrahimi [24], we deduce that, if Y is IURL, then HY (t) − 1 + log λY (t) ≥ 0.
Moreover, since Y ≤LU X and X is DURL, we have HY (t) − 1 + log λ X (t) ≤ 0. Hence, from (25),
it follows that HXY is increasing.
t
(iii) By expressing the derivative of HXY in an alternative form, we obtain
t

d * +
H Y = F (t) [λY (t) − λ X (t)] HY (t) + λY (t)[−1 + log λY (t)] − λ X (t)[−1 + log λ X (t)] .
dt Xt
Hence, thanks to the given hypothesis, the right-hand-side of the above identity is nonnegative.

We remark that the assumption λ X (t) ≤ e ≤ λY (t) for all t ≥ 0, considered in point (iii)
of Proposition 3, implies that Y is larger than X in the ‘up hazard rate order’, i.e., Y ≤hr↑ X
(see Theorem 6.21 of Lillo et al. [35]). This condition, in turn, implies that Y ≤hr X.
Classical studies in reliability theory show that the random lifetimes of items or systems follow
suitable Weibull distributions. As an illustrative instance, in the forthcoming example, we investigate
the effect of replacement for Weibull distributed lifetimes in terms of dynamic differential entropy.

Example 3. Assume that X and Y have Weibull distribution, with F (t) = 1 − e−(t/λ) , t ≥ 0, and
k

G (t) = 1 − e−(t/μ) , t ≥ 0, with λ, k, μ, h > 0. The differential entropy of the lifetime (3) can be obtained
h

by means of (22). However, we omit its expression since it is quite cumbersome. Some plots of the
dynamic differential entropy HXY are given in Figure 4, in order to show the effect of the replacement
t
at time t. Specifically, for the considered parameters, we have that HXY grows when μ increases and h
t
decreases. Moreover, HXY has a reversed bathtub shape, with a single maximum attained for positive
t
values of t. Clearly, such maximum provides useful information in order to chose optimal values (in
the differential entropy sense) of the replacement instant.

128
Mathematics 2018, 6, 204

Figure 4. With reference to Example 3, plot of HXY for λ = 1 and k = 2. Left: h = 2 and μ = 1.5, 2, 3, 4
t
(from bottom to top). Right: μ = 2 and h = 2, 3, 4, 5 (from top to bottom).

6. Conclusions
Certain typical replacement models in reliability theory involve minimal repair instances, in
which—upon failure—an item is replaced by another one having the same reliability of the failed item
at the failure instant. The model considered in this paper deals with a different scenario, in which the
replacement is planned in advance, the replaced item possessing a different failure distribution and
having the same age of the replaced item.
The investigation has been centered first on the stochastic comparison of the resulting random
lifetimes. We have proposed measuring the goodness of the replacement criteria by means of the
relative ratio of improvement. The information amount provided by the dynamic version of the system
lifetime differential entropy has also been considered as a relevant tool in this respect.
Possible future developments of the given results can be finalized to the extension of the model
to more specific instances in which the replacement instant is constrained by operational guidelines,
which can be implemented through suitable weight functions.

Author Contributions: All of the authors contributed equally to this work.

Funding: This research received no external funding.
Acknowledgments: A.D.C. and P.D.G. are members of the groups GNCS and GNAMPA of INdAM (Istituto
Nazionale di Alta Matematica), respectively.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Krakowski, M. The relevation transform and a generalization of the gamma distribution function. Rev. Fr.
Autom. Inform. Rech. Opér. 1973, 7, 107–120. [CrossRef]
2. Baxter, L.A. Reliability applications of the relevation transform. J. Appl. Probab. 1982, 29, 323–329. [CrossRef]
3. Belzunce, F.; Martínez-Riquelme, C.; Ruiz, J.M. Allocation of a relevation in redundancy problems. Appl. Stoch.
Model Bus. Ind. 2018, online first. [CrossRef]
4. Chukova, S.; Dimitrov, B.; Dion, J.P. On relevation transforms that characterize probability distributions.
J. Appl. Math. Stoch. Anal. 1993, 6, 345–357. [CrossRef]
5. Shanthikumar, J.G.; Baxter, L.A. Closure properties of the relevation transform. Nav. Res. Logist. Q.
1985, 32, 185–189. [CrossRef]
6. Sordo, M.A.; Psarrakos, G. Stochastic comparisons of interfailure times under a relevation replacement policy.
J. Appl. Probab. 2017, 54, 134–145. [CrossRef]
7. Di Crescenzo, A.; Toomaj, A. Extension of the past lifetime and its connection to the cumulative entropy.
J. Appl. Probab. 2015, 52, 1156–1174. [CrossRef]
8. Barlow, R.; Proschan, F. Mathematical Theory of Reliability; Classics in Applied Mathematics (Book 17);
With Contributions by Larry C. Hunter; SIAM: Philadelphia, PA, USA, 1996.

129
Mathematics 2018, 6, 204

9. Block, H.W.; Savits, T.H.; Singh, H. The reversed hazard rate function. Probab. Eng. Inf. Sci. 1998, 12, 69–90.
[CrossRef]
10. Santacroce, M.; Siri, P.; Trivellato, B. New results on mixture and exponential models by Orlicz spaces.
Bernoulli 2016, 22, 1431–1447. [CrossRef]
11. Cicalese, F.; Gargano, L.; Vaccaro, U. Bounds on the entropy of a function of a random variable and their
applications. IEEE Trans. Inf. Theory 2018, 64, 2220–2230. [CrossRef]
12. Shaked, M.; Shanthikumar, J.G. Stochastic Orders; Springer Series in Statistics; Springer: New York, NY, USA, 2007.
13. Belzunce, F.; Martínez-Riquelme, C.; Mulero, J. An Introduction to Stochastic Orders; Elsevier/Academic Press:
Amsterdam, The Netherlands, 2016.
14. Arnold, B.C.; Castillo, E.; Sarabia, J.M. Conditional Specification of Statistical Models; Springer:
New York, NY, USA, 1999.
15. Escobar, L.A.; Meeker, W.Q. A review of accelerated test models. Stat. Sci. 2006, 21, 552–577. [CrossRef]
16. Di Crescenzo, A.; Martinucci, B.; Mulero, J. A quantile-based probabilistic mean value theorem. Probab. Eng.
Inf. Sci. 2016, 30, 261–280. [CrossRef]
17. Di Crescenzo, A. A probabilistic analogue of the mean value theorem and its applications to reliability theory.
J. Appl. Probab. 1999, 39, 706–719. [CrossRef]
18. Di Crescenzo, A.; Martinucci, B.; Zacks, S. Compound Poisson process with a Poisson subordinator.
J. Appl. Probab. 2015, 52, 360–374. [CrossRef]
19. Cox, D.R. Regression models and life tables (with Discussion). J. R. Stat. Soc. Ser. B 1972, 34, 187–220.
20. Nanda, A.K.; Das, S. Dynamic proportional hazard rate and reversed hazard rate models. J. Stat. Plan. Inference
2011, 141, 2108–2119. [CrossRef]
21. Ng, H.K.T.; Navarro, J.; Balakrishnan, N. Parametric inference from system lifetime data under a proportional
hazard rate model. Metrika 2012, 75, 367–388. [CrossRef]
22. Cover, T.M.; Thomas, J.A. Elements of Information Theory, 2nd ed.; Wiley-Interscience: Hoboken, NJ, USA, 2006.
23. Muliere, P.; Parmigiani, G.; Polson, N.G. A note on the residual entropy function. Probab. Eng. Inf. Sci.
1993, 7, 413–420. [CrossRef]
24. Ebrahimi, N. How to measure uncertainty in the residual life time distribution. Sankhyā Ser. A 1996, 58, 48–56.
25. Ebrahimi, N.; Pellerey, F. New partial ordering of survival functions based on the notion of uncertainty.
J. Appl. Probab. 1995, 32, 202–211. [CrossRef]
26. Asadi, M.; Ebrahimi, N. Residual entropy and its characterizations in terms of hazard function and mean
residual life function. Stat. Probab. Lett. 2000, 49, 263–269. [CrossRef]
27. Ebrahimi, N.; Kirmani, S.N.U.A.; Soofi, E.S. Multivariate dynamic information. J. Multivar. Anal. 2007,
98, 328–349. [CrossRef]
28. Di Crescenzo, A.; Longobardi, M. Entropy-based measure of uncertainty in past lifetime distributions.
J. Appl. Probab. 2002, 39, 434–440. [CrossRef]
29. Nanda, A.K.; Paul, P. Some properties of past entropy and their applications. Metrika 2006, 64, 47–61.
[CrossRef]
30. Kundu, C.; Nanda, A.K.; Maiti, S.S. Some distributional results through past entropy. J. Stat. Plan. Inference
2010, 140, 1280–1291. [CrossRef]
31. Sachlas, A.; Papaioannou, T. Residual and past entropy in actuarial science and survival models.
Methodol. Comput. Appl. Probab. 2014, 16, 79–99. [CrossRef]
32. Kundu, A.; Nanda, A.K. On study of dynamic survival and cumulative past entropies. Commun. Stat.
Theory Methods 2016, 45, 104–122. [CrossRef]
33. Ahmadi, J.; Di Crescenzo, A.; Longobardi, M. On dynamic mutual information for bivariate lifetimes.
Adv. Appl. Probab. 2015, 47, 1157–1174. [CrossRef]
34. Bowden, R.J. Information, measure shifts and distribution metrics. Statistics 2012, 46, 249–262. [CrossRef]
35. Lillo, R.E.; Nanda, A.K.; Shaked, M. Some shifted stochastic orders. In Recent Advances in Reliability Theory;
Statistics for Industry and Technology; Birkhäuser: Boston, MA, USA, 2000; pp. 85–103.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

130
 mathematics
Article
A Time-Non-Homogeneous Double-Ended
Queue with Failures and Repairs and Its
Continuous Approximation
Antonio Di Crescenzo 1 , Virginia Giorno 2 , Balasubramanian Krishna Kumar 3 and
Amelia G. Nobile 2, *
1 Dipartimento di Matematica, Università degli Studi di Salerno, Via Giovanni Paolo II n. 132,
84084 Fisciano (SA), Italy; [email protected]
2 Dipartimento di Informatica, Università degli Studi di Salerno, Via Giovanni Paolo II n. 132,
84084 Fisciano (SA), Italy; [email protected]
3 Department of Mathematics, Anna University, Chennai 600 025, India; [email protected]
* Correspondence: [email protected]

Received: 6 April 2018; Accepted: 9 May 2018; Published: 11 May 2018

Abstract: We consider a time-non-homogeneous double-ended queue subject to catastrophes and

repairs. The catastrophes occur according to a non-homogeneous Poisson process and lead the system
into a state of failure. Instantaneously, the system is put under repair, such that repair time is governed
by a time-varying intensity function. We analyze the transient and the asymptotic behavior of the
queueing system. Moreover, we derive a heavy-traffic approximation that allows approximating
the state of the systems by a time-non-homogeneous Wiener process subject to jumps to a spurious
state (due to catastrophes) and random returns to the zero state (due to repairs). Special attention
is devoted to the case of periodic catastrophe and repair intensity functions. The first-passage-time
problem through constant levels is also treated both for the queueing model and the approximating
diffusion process. Finally, the goodness of the diffusive approximating procedure is discussed.

Keywords: double-ended queues; time-non-homogeneous birth-death processes; catastrophes;

repairs; transient probabilities; periodic intensity functions; time-non-homogeneous jump-diffusion
processes; transition densities; first-passage-time

MSC: 60J80; 60J25; 60J75; 60K25

1. Introduction
Double-ended queues are often adopted as stochastic models for queueing systems characterized
by two flows of agents, i.e., customers and servers/resources. When there are a customer and a server
in the system, the match between the request and service occurs immediately, and then, both agents
leave the system. As a consequence, there cannot be simultaneously customers and servers in the
system. Namely, denoting by N (t) the state of the system at time t, it is assumed that N (t) = n, n ∈ N,
if there are n customers waiting for available servers, whereas N (t) = −n, n ∈ N, if there are n servers
waiting for new customers, and N (t) = 0, if the system is empty. Hence, typical models for N (t) are
bilateral continuous-time Markov chains or similar stochastic processes.
Double-ended queueing systems can be applied to model numerous situations in real-world
scenarios. A classical example is provided by taxi-passenger systems, where the role of customers
and servers is played by passengers and taxis, respectively. We recall the first papers on this
topic by Kashyap [1,2] and the subsequent contributions by Sharma and Nair [3], Tarabia [4]
and Conolly et al. [5]. Other examples are provided by the dynamical allocation of live organs

Mathematics 2018, 6, 81; doi:10.3390/math6050081 131 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 81

(servers) to candidates (customers) needing transplantation (cf. Elalouf et al. [6] and the references
therein). Double-ended queues are suitable also to describe different streams arriving at a system (see
Takahashi et al. [7]).
In this area, the interest is typically in the determination of the transient distribution and the
asymptotic distribution of the system state, the busy period density, the waiting time density and
related indices such as means and variances. The difficulties related to the resolution of the birth-death
processes describing the length of the queue, in some cases, can be overcome by means of suitable
transformations as those presented in Di Crescenzo et al. [8]. Such a transformation-based approach
has been successfully exploited also for diffusion processes (see Di Crescenzo et al. [9]), this being of
interest for the analysis of customary diffusion approximations of queue-length processes.
Attention is given often also to variants of the relevant stochastic processes that are adopted
to describe more complex situations, such as bulk arrivals, truncated queues, the occurrence of
disasters and repairs, and so on. In this respect, we recall the recent paper by Di Crescenzo et al. [10],
which is centered on the analysis (i) of a continuous-time stochastic process describing the state of a
double-ended queue subject to disasters and repairs and (ii) of the Wiener process with jumps arising
as a heavy-traffic approximation to the previous model.
In many queueing models of manufacturing systems, it is assumed that the times to failure and
the times to repair of each machine are exponentially distributed. However, exponential distributions
do not always accurately represent distributions encountered in real manufacturing systems. Some
of these models adopt the phase-type distributions for failure and repair times (see, for instance,
Altiok [11–13] and Dallery [14]).
In this paper, we propose and analyze an extension of the queueing model treated in [10] to a
time-non-homogeneous setting in which the intensities of arrivals, services, disasters and repairs
are suitably time dependent. Similarly, we investigate the related heavy-traffic jump-diffusion
approximation, as well. The key features of our analysis and the motivations of the proposed study
are based mainly on the following issues:

• Queueing systems subject to disasters are appropriate to model more realistic situations in
which the number of customers is subject to an abrupt decrease by the effect of catastrophes
occurring randomly in time and due to external causes. The literature on the area of
stochastic systems evolving in the presence of catastrophes is very broad. We restrict
ourselves to mentioning the papers by Economou and Fakinos [15,16], Kyriakidis and
Dimitrakos [17], Krishna Kumar et al. [18], Di Crescenzo et al. [19], Zeifman and Korotysheva [20],
Zeifman et al. [21] and Giorno et al. [22]. The analysis of some time-dependent queueing
models with catastrophes has been performed in Di Crescenzo et al. [23] and, more recently,
in Giorno et al. [24], with special attention to the M(t)/M (t)/1 and M (t)/M (t)/∞ queues.
• We include a repair mechanism in the queueing system under investigation, since it is essential to
model instances when the (random) repair times are not negligible. We remark that the interest in
this feature is increasing in the recent literature on queueing theory (see, for instance, Dimou and
Economou [25]).
• Heavy-traffic approximations are very often proposed in order to describe the queueing systems
under proper limit conditions of the parameters involved. This allows one to come to more
manageable formulas for the description of the queue content. Typically, a customary rescaling
procedure allows one to approximate the queue length process by a diffusion process, as indicated
in Giorno et al. [26]. Examples of diffusion models arising from heavy-traffic approximations
of double-ended queues and of similar matching systems can be found in Liu et al. [27] and
Büke and Chen [28], respectively. In the case of queueing systems subject to catastrophes,
a customary approach leads to jump-diffusion approximating processes (see, for instance,
Di Crescenzo et al. [29] and Dharmaraja et al. [30]).

132
Mathematics 2018, 6, 81

Plan of the Paper

In Section 2, we consider a non-homogeneous double-ended queue, whose arrivals and departures
occur with time-varying intensity functions λ(t) > 0 and μ(t) > 0, respectively. We discuss various
features of such a model, including the first-passage time through a constant level.
In Section 3, we consider the non-homogeneous double-ended queue subject to disasters and
repairs, both occurring with time-varying rates. Specifically, we assume that catastrophes occur
according to a non-homogeneous Poisson process with intensity function ν(t) > 0. The effect of
catastrophes moves the system into a spurious failure state, say F. The completion of a system’s repair
occurs with time-varying intensity function η (t) > 0. After any repair, the system starts afresh from
the zero state. Our first aim is to determine the probability q(t|t0 ) that the system at time t is in the
failure state and the probability p0,n (t|t0 ) that the system at time t is in the state n ∈ Z (working state).
In Section 4, we study the asymptotic behavior of the state probabilities in two different cases:
(i) when the rates λ(t), μ(t), ν(t), η (t) admit finite positive limits as t tends to infinity and (ii) when
the double-ended queue is time-homogeneous, the catastrophe intensity function ν(t) and the repair
intensity function η (t) being periodic functions with common period Q.
In Section 5, we consider a diffusion approximation, under a heavy-traffic regime, of the
non-homogeneous double-ended queue discussed in Section 2. In this case, the approximating process
is a time-non-homogeneous Wiener process. We discuss various results on this model, including a
first-passage-time problem through a constant level.
In Section 6, we deal with the heavy-traffic jump-diffusion approximation for the discrete model
with catastrophes and repairs. Various results shown for the basic diffusion process treated in the
previous section are thus extended to the present case characterized by jumps. In both Sections 5 and 6,
the goodness of the approximating procedure is discussed, as well.
In Section 7, we finally consider the asymptotic behavior of the densities in the same cases
considered in Section 4. In conclusion, we perform some comparisons between the relevant quantities
of the queueing system and of the approximating diffusion process under the heavy-traffic regime.

2. The Underlying Non-Homogeneous Double-Ended Queue

This section is devoted to the analysis of the basic time-non-homogeneous double-ended queue.
Let { N  (t), t ≥ t0 }, with t0 ≥ 0, be a continuous-time Markov chain describing the
number of customers in a time-non-homogeneous double-ended queueing system, with state-space
Z = {. . . , −1, 0, 1, . . .}. We assume that arrivals (upward jumps) and departures (downward jumps)
at time t occur with intensity functions λ(t) > 0 and μ(t) > 0, respectively, where λ(t) and μ(t) are
 +∞  +∞
bounded and continuous functions for t ≥ t0 , such that t λ(t) dt = +∞ and t μ(t) dt = +∞.
0 0
The given assumptions ensure that the eventual transitions from any state occur w.p. 1. The state
diagram of N  (t) is shown in Figure 1.

λ(t) λ(t) λ(t) λ(t) λ(t) λ(t) λ(t) λ(t)

··· −3 −2 −1 0 1 2 3 ···
μ(t) μ(t) μ(t) μ(t) μ(t) μ(t) μ(t) μ(t)

Figure 1. State diagram of the non-homogeneous double-ended queueing system.

 (t) = n| N
For all j, n ∈ Z and t > t0 ≥ 0, the transition probabilities pj,n (t|t0 ) = P{ N  (t0 ) = j} are
solutions of the system of Kolmogorov forward equations:

d pj,n (t|t0 )
= λ(t) pj,n−1 (t|t0 ) − [λ(t) + μ(t)] pj,n (t|t0 ) + μ(t) pj,n+1 (t|t0 ), j, n ∈ Z, (1)
dt

133
Mathematics 2018, 6, 81

with the initial condition limt↓t0 pj,n (t|t0 ) = δj,n , where δj,n is the Kronecker delta function. For t ≥ t0
and 0 ≤ z ≤ 1, let:
 +∞
j (z, t|t0 ) = E z N (t) N
G  ( t0 ) = j = ∑ zn pj,n (t|t0 ), j∈Z (2)
n=−∞

 (t). For any t ≥ t0 , we denote the cumulative arrival and

be the probability generating function of N
service intensity functions by:
 t  t
Λ ( t | t0 ) = λ(τ ) dτ, M ( t | t0 ) = μ(τ ) dτ. (3)
t0 t0

Due to (1), for t ≥ t0 , the probability generating Function (2) is the solution of the partial
differential equation:

∂  ! μ(t) " 
Gj (z, t|t0 ) = −[λ(t) + μ(t)] + λ(t) z + Gj (z, t|t0 ), j∈Z
∂t z
to be solved with the initial condition limt↓t0 Gj (z, t|t0 ) = z j . Hence, (2) can be expressed in terms of (3)
as follows:
 
# $
Gj (z, t|t0 ) = z j exp − Λ(t|t0 ) + M (t|t0 ) + Λ(t|t0 ) z + M(t|t0 ) , j ∈ Z. (4)
z

Recalling that (cf. Abramowitz [31], p. 376, n. 9.6.33):

!s  1 " +∞
exp
2
r+
r
= ∑ r n In (s) (r = 0), (5)
n=−∞

where:
∞
(z/2)ν+2m
Iν (z) = ∑ (ν ∈ R)
m=0 m! Γ ( ν + m + 1)

denotes the modified Bessel function of first kind and by setting:

%
Λ ( t | t0 )
s=2 Λ ( t | t0 ) M ( t | t0 ), r=z
M ( t | t0 )

in (5), from (4), one has:

# $ +∞  Λ(t|t ) k/2 
j (z, t|t0 ) = e−
G Λ(t|t0 )+ M (t|t0 )
∑ z j+k
0
Ik 2 Λ ( t | t0 ) M ( t | t0 ) , j ∈ Z.
k =−∞
M ( t | t0 )

Hence, recalling (2), one obtains the transition probabilities:

# $ (n− j)/2 
Λ ( t | t0 )
pj,n (t|t0 ) = e− Λ(t|t0 )+ M (t|t0 )
In− j 2 Λ ( t | t0 ) M ( t | t0 ) , j, n ∈ Z. (6)
M ( t | t0 )

We remark that, since In (z) = I−n (z) for n ∈ N, the following symmetry relation holds:
 Λ ( t | t ) n− j
0
pj,n (t|t0 ) = pj,2j−n (t|t0 ) j, n ∈ Z.
M ( t | t0 )

 (t), for t ≥ t0 and j ∈ Z:

Moreover, from (6), we recover the conditional mean and variance of N

 (t)| N
E[ N  ( t0 ) = j ] = j + Λ ( t | t0 ) − M ( t | t0 ),  (t)| N
Var[ N  ( t0 ) = j ] = Λ ( t | t0 ) + M ( t | t0 ). (7)

134
Mathematics 2018, 6, 81

We point out that the transition probabilities given in (6) constitute the probability distribution of
the difference of two independent non-homogeneous Poisson processes with intensities λ(t) and μ(t),
respectively, originated at zero (cf. Irwin [32] or Skellam [33] for the homogeneous case).
Let us now consider the first-passage-time (FPT) of N  (t) through the state n ∈ Z, starting from
the initial state j ∈ Z. Such a random variable will be denoted as:

Tj,n (t0 ) = inf{t ≥ t0 : N

 ( t ) = n },  (t0 ) = j, j = n,
N

where gj,n (t|t0 ) is its probability density function (pdf). Special interest is given to Tj,0 (t0 ), which
represents the busy period of the double-ended queue, with initial state N  (t0 ) = j. As is well-known,
due to the Markov property, gj,n (t|t0 ) satisfies the integral equation:
 t
pj,n (t|t0 ) = gj,n (τ |t0 ) pn,n (t|τ ) dτ, j, n ∈ Z, j = n. (8)
t0

Hereafter, we consider the special case in which the arrival and departure intensity functions
are proportional.

Remark 1. Let λ(t) = λϕ(t) and μ(t) = μϕ(t), with λ, μ positive constants, where ϕ(t) is a positive,
∞
bounded and continuous function of t ≥ t0 , such that t ϕ(t) dt = +∞. By setting  = λ/μ and:
0

 t
Φ ( t | t0 ) = ϕ(τ ) dτ, t ≥ t0 , (9)
t0

then the transition probabilities (6) of the non-homogeneous double-ended queueing system N  (t) can be
expressed as:
# $
pj,n (t|t0 ) = e−(λ+μ)Φ(t|t0 ) (n− j)/2 In− j 2λμΦ(t|t0 ) , j, n ∈ Z. (10)

Hence, from the results given in Section 5 of Giorno et al. [24], we have:

|n − j| ϕ(t)
gj,n (t|t0 ) = pj,n (t|t0 ), j, n ∈ Z, j = n. (11)
Φ ( t | t0 )

Furthermore, the FPT ultimate probability is given by:

 +∞

1, (λ − μ)(n − j) ≥ 0,
P{Tj,n (t0 ) < +∞} = gj,n (t|t0 ) dt =
t0 n− j , (λ − μ)(n − j) < 0.

3. The Queueing System with Catastrophes and Repairs

This section deals with the analysis of the queueing system with catastrophes and repairs.
Let { N (t), t ≥ t0 }, with t0 ≥ 0, be a continuous-time Markov chain that describes the number of
customers of a time-non-homogeneous double-ended queueing system subject to disasters and repairs.
The state-space of { N (t), t ≥ t0 } is denoted by S = { F } ∪ Z = { F, 0, ±1, ±2, . . .}, where F denotes the
failure state. We assume that the catastrophes occur according to a non-homogeneous Poisson process
with intensity function ν(t). If a catastrophe occurs, then the system goes instantaneously into the
failure state F, and further, the completion of a repair occurs according to the intensity function η (t)
(cf. the diagram shown in Figure 2). We assume that the rates ν(t) and η (t) are positive, bounded and
∞ ∞
continuous functions for t ≥ t0 , such that t ν(t) dt = +∞ and t η (t) dt = +∞. After every repair,
0 0
the system starts again from the zero state.

135
Mathematics 2018, 6, 81

λ(t) λ(t) λ(t) λ(t) λ(t) λ(t) λ(t) λ(t)

··· −3 −2 −1 0 1 2 3 ···
μ(t) μ(t) μ(t) μ(t) μ(t) μ(t) μ(t) μ(t)

η (t) ν(t)
ν(t) ν(t)
ν(t) ν(t)
ν(t) F ν(t)

Figure 2. State diagram of the time-non-homogeneous double-ended queueing system with

catastrophes and repairs.

For any n ∈ Z and t > t0 ≥ 0, we set:

p0,n (t|t0 ) = P{ N (t) = n| N (t0 ) = 0}, q ( t | t0 ) = P { N ( t ) = F | N ( t0 ) = 0}. (12)

Hence, p0,n (t|t0 ) is the transition probability from zero, at time t0 , to state n, at time t, when
the system is active (in this case, we say that the system is in the “on” state), whereas q(t|t0 ) is the
probability that the queueing system is in the state F (called the “failure” state) at time t starting from
zero at time t0 . The probabilities given in (12) are the solution of the forward Kolmogorov system of
differential equations:

dq(t|t0 )
= −η (t) q(t|t0 ) + ν(t) [1 − q(t|t0 )], (13)
dt
dp0,0 (t|t0 )
= −[λ(t) + μ(t) + ν(t)] p0,0 (t|t0 ) + λ(t) p0,−1 (t|t0 ) + μ(t) p0,1 (t|t0 ) + η (t) q(t|t0 ), (14)
dt
dp0,n (t|t0 )
= −[λ(t) + μ(t) + ν(t)] p0,n (t|t0 ) + λ(t) p0,n−1 (t|t0 ) + μ(t) p0,n+1 (t|t0 ), n ∈ Z \ {0}, (15)
dt

to be solved with the following initial conditions, based on the Kronecker delta function:

lim pn,0 (t|t0 ) = δn,0 , lim q(t|t0 ) = 0. (16)

t ↓ t0 t ↓ t0

Conditions (16) imply that at initial time t0 , the system is active and it starts from the zero state.
In order to determine the transient probabilities of N (t), similarly as in (3), in the following, we denote
the cumulative catastrophe and repair intensity functions by:
 t  t
V ( t | t0 ) = ν(τ ) dτ, H ( t | t0 ) = η (τ ) dτ, t ≥ t0 , (17)
t0 t0

respectively.

Transient Probabilities
We first determine the probability that the system is under repair at time t. By solving Equation (13)
with the second of the initial conditions (16), recalling (17), one obtains the probability that the process
N (t) is in the state F (“failure” state) at time t, starting from zero at time t0 :
 t
q ( t | t0 ) = ν(τ ) e−[V (t|τ )+ H (t|τ )] dτ, t ≥ t0 . (18)
t0

The transient analysis of the process N (t) can be performed by relating the transient probabilities
to those of the same process in the absence of catastrophes. Indeed, by conditioning on the time of the

136
Mathematics 2018, 6, 81

last catastrophe of N (t) before t, the probabilities p0,n (t|t0 ) can be expressed in terms of p0,n (t|t0 ) as
follows, for n ∈ Z and t ≥ t0 (cf. [10,15,16]):
 t
p0,n (t|t0 ) = e−V (t|t0 ) p0,n (t|t0 ) + q(τ |t0 ) η (τ ) e−V (t|τ ) p0,n (t|τ ) dτ. (19)
t0

We note that the first term on the right-hand side of (19) expresses the probability that process
N (t) occupies state n at time t and that no catastrophes occurred in [0, t]. Similarly, the second term
gives the probability that process N (t) occupies state n at time t and that at least one catastrophe (with
successive repair) occurred in [0, t], i.e.,

- starting from zero at time t0 , at least a catastrophe and the subsequent repair occur before t; let
τ ∈ (0, t) be the instant at which the last repair occurs, so that a transition entering in the zero
state occurs at time τ with intensity η (τ );
- no catastrophe occurs in the interval (τ, t); then the system, starting from the zero state at time τ,
reaches the state n at time t.

Note that Equation (19) is the suitable extension of (2.7) of [10], which refers to the case of constant
rates. Furthermore, we remark that from (18) and (19), one obtains:

+∞
∑ p0,n (t|t0 ) + q(t|t0 ) = 1, t ≥ t0 . (20)
n=−∞

Making use of (6) and (18) in (19), for t ≥ t0 and n ∈ Z, one has the following expression for the
transition probabilities of N (t):
 Λ(t|t ) n/2 
p0,n (t|t0 ) = e−[Λ(t|t0 )+ M(t|t0 )+V (t|t0 )]
0
In 2 Λ(t|t0 ) M(t|t0 )
M ( t | t0 )
 t  n/2   τ
−[Λ(t|τ )+ M(t|τ )+V (t|τ )] Λ ( t | τ )
+ dτ η (τ )e In 2 Λ(t|τ ) M(t|τ ) ν(ϑ )e−[V (τ |ϑ)+ H (τ |ϑ)] dϑ.
t0 M(t|τ ) t0

Let us now introduce the r-th conditional moment of N (t), for r ∈ N:

+∞
1
Mr (t|t0 ) := E[ N r (t)| N (t) ∈ Z, N (t0 ) = 0] =
1 − q ( t | t0 ) ∑ nr p0,n (t|t0 ). (21)
n=−∞

From (19), it is not hard to see that the moments (21) can be expressed in terms of the conditional
-r (t|t0 ) := E[ N
moments M  r (t)| N
 (t0 ) = 0] as follows, for r ∈ N and t ≥ t0 :
  t 
1 -r (t|t0 ) + -r (t|τ ) dτ .
Mr ( t | t0 ) = e −V ( t | t0 ) M q ( τ | t0 ) η ( τ ) e −V ( t | τ ) M (22)
1 − q ( t | t0 ) t0

Hence, by virtue of (7), from (22), the conditional mean and second order moment of N (t) can be
evaluated based on the knowledge of the relevant intensity functions.
Hereafter, we see that if the arrival and departure rates are constant, then some simplifications
hold for the transition probabilities and conditional moments.

Theorem 1. For the queueing system with catastrophes and repairs, having constant arrival rates λ(t) = λ
and departure rates μ(t) = μ, for t ≥ t0 and n ∈ Z, one has:
 t − t0  x
p0,n (t|t0 ) = e−V (t|t0 ) p0,n (t − t0 |0) + dx ν(t − x ) e−V (t|t− x) η (t − u) e− H (t−u|t− x) p0,n (u|0) du (23)
0 0

and, for r ∈ N,

137
Mathematics 2018, 6, 81

  t − t0 
1 -r (t − t0 |0) + -r ( x |0) dx .
Mr ( t | t0 ) = e −V ( t | t0 ) M q ( t − x | t0 ) η ( t − x ) e −V ( t | t − x ) M (24)
1 − q ( t | t0 ) 0

Furthermore, it results:
 t
p0,n (t|t0 ) = p0,0 (τ |t0 ) e−V (t|τ ) g0,n (t|τ ) dτ, n ∈ Z \ {0}, t ≥ t0 . (25)
t0

Proof. Since λ(t) = λ and μ(t) = μ, by virtue of (18), Relation (23) follows from (19), whereas
Equation (24) derives from (22). Moreover, making use of (19) in the right-hand side of (25), one has:
 t  t
p0,0 (u|t0 ) e−V (t|u) g0,n (t|u) du = e−V (t|t0 ) p0,0 (u|t0 ) g0,n (t|u) du
t0 t0
 t  t
+ dτ e−V (t|τ ) η (τ ) q(τ |t0 ) p0,0 (u|τ ) g0,n (t|u) du. (26)
t0 τ

By virtue of (6), we note that p0,0 (t|t0 ) = pn,n (t|t0 ) for n ∈ Z and t ≥ t0 . Moreover, since λ(t) = λ
and μ(t) = μ, we obtain:
 t  t
p0,0 (u|t0 ) g0,n (t|u) du = pn,n (u|t0 ) g0,n (t|u) du
t0 t0
 t − t0
= pn,n (t − t0 |τ ) g0,n (τ |0) dτ = p0,n (t − t0 |0) = p0,n (t|t0 ). (27)
0

Substituting (27) in (26), by virtue of (19), Relation (25) immediately follows.

The integrand on the right-hand side of Equation (25) refers to the sample-paths of N (t) that start
from zero at time t0 , then reach the state zero at time τ ∈ (t0 , t) and, finally, go from zero at time τ to n
at time t for the first time, without the occurrence of catastrophes in (τ, t).

4. Asymptotic Probabilities
In this section, we analyze the asymptotic behavior of the probabilities q(t|t0 ) and p j,n (t|t0 ) of the
process N (t) in two different cases:

(i) the intensity functions λ(t), μ(t), ν(t) and η (t) admit finite positive limits as t → +∞,
(ii) the intensity functions λ(t) and μ(t) are constant, whereas the rates ν(t) and η (t) are periodic
functions with common period Q.

4.1. Asymptotically-Constant Intensity Functions

In the following theorem, we determine the steady-state probabilities and the asymptotic failure
probability of the process N (t) when the intensity functions λ(t), μ(t), ν(t) and η (t) admit finite
positive limits as t tends to +∞.

Theorem 2. If:

lim λ(t) = λ, lim μ(t) = μ, lim ν(t) = ν, lim η (t) = η, (28)

t→+∞ t→+∞ t→+∞ t→+∞

with λ, μ, ν, η positive constants, then the steady-state probabilities and the asymptotic failure probability of the
process N (t) are:

138
Mathematics 2018, 6, 81

ν
q∗ := lim q(t|t0 ) = , (29)
t→+∞ ν+η
ν (1 − q )
p0∗ := lim p0,0 (t|t0 ) =  ,
t→+∞ (λ + μ + ν)2 − 4λμ
 λ + μ + ν − (λ + μ + ν)2 − 4λμ n
p∗n := lim p0,n (t|t0 ) = p0∗ , n ∈ N, (30)
t→+∞ 2μ

 λ + μ + ν − (λ + μ + ν)2 − 4λμ n
p∗−n := lim p0,−n (t|t0 ) = p0∗ , n ∈ N.
t→+∞ 2λ

Furthermore, the asymptotic conditional mean, second order moment and variance are:

λ−μ 2( λ − μ )2 (λ + μ)
M1∗ = lim M1 (t|t0 ) = , M2∗ = lim M2 (t|t0 ) =+ ,
t→+∞ ν t→+∞ ν2 ν
(31)
(λ − μ) 2 λ+μ
V ∗ = lim Var(t|t0 ) = lim {M2 (t|t0 ) − [M1 (t|t0 )]2 } = + .
t→+∞ t→+∞ ν2 ν

Proof. The steady-state probabilities and the asymptotic failure probability of N (t) can be obtained
by taking the limit as t → +∞ in Equations (13)–(15), and then solving the corresponding balance
equations. From (21), making use of (29) and (30), the asymptotic conditional mean and variance (31)
immediately follow.

4.2. Periodic Catastrophe and Repair Intensity Functions

Let us assume that the arrival and departure intensity functions are constant, whereas the
catastrophe intensity function ν(t) and the repair intensity function η (t) are periodic, such that
ν(t + kQ) = ν(t) and η (t + kQ) = η (t) for all k ∈ N, t ≥ t0 , for a given constant period Q > 0.
We denote by:
 
1 Q 1 Q
ν∗ = ν(u) du, η∗ = η (u) du, (32)
Q 0 Q 0
the average catastrophe and repair rates over the period Q. Since ν(t) and η (t) are periodic functions,
from (17), we have, for t ≥ t0 :
 t+kQ  t+kQ
V (t + kQ) = ν(u) du = kQν∗ , H (t + kQ) = η (u) du = kQη ∗ , k ∈ N. (33)
t t

Let us now investigate the asymptotic distribution for the process N (t), which can be defined as
follows, for t ≥ t0 :

q∗ (t) = lim q(t + kQ|t0 ), ∗

p0,n (t) = lim p0,n (t + kQ|t0 ), n ∈ Z. (34)
k →+∞ k →+∞

Theorem 3. For the queueing system with catastrophes and repairs, having constant arrival rates λ(t) = λ > 0
and departure rates μ(t) = μ > 0, with ν(t) and η (t) continuous, positive and periodic functions, with period
Q, for t ≥ t0 , one has:
 +∞  x
∗
p0,n (t) = dx ν(t − x )e−V (t|t− x) η (t − u)e− H (t−u|t− x) p0,n (u|0) du, n ∈ Z. (35)
0 0
 +∞
q∗ (t) = ν(t − x )e−[V (t|t− x)+ H (t|t− x)] dx. (36)
0

Furthermore, an alternative expression for the failure asymptotic probability is:

139
Mathematics 2018, 6, 81

 Q
1
q∗ (t) = Q(ν∗ +η ∗ )
ν(t + u)e[V (t+u|t)+ H (t+u|t)] du, (37)
e −1 0

with ν∗ and η ∗ given in (32).

Proof. Since λ(t) = λ and μ(t) = μ, from (23), for k ∈ N0 and t ≥ t0 , one has:
 t−t0 +kQ
p0,n (t + kQ|t0 ) = e−V (t+kQ|t0 ) p0,n (t − t0 + kQ|0) + dx ν(t − x ) e−V (t+kQ|t− x+kQ)
0
 x
× η (t − u) e− H (t−u+kQ|t− x+kQ) p0,n (u|0) du. (38)
0

Due to the periodicity of ν(t) and η (t), the following equalities hold:

V (t + kQ|t − x + kQ) = V (t|t − x ), H (t − u + kQ|t − x + kQ) = H (t − u|t − x ).

Hence, from (38), it follows:

 t−t0 +kQ
p0,n (t + kQ|t0 ) = e−V (t+kQ|t0 ) p0,n (t − t0 + kQ|0) + dx ν(t − x ) e−V (t+|t− x)
0
 x
× η (t − u) e− H (t−u|t− x) p0,n (u|0) du. (39)
0

Then, taking the limit as k → +∞ in (39) and recalling the second of (34), one obtains (35). Hence,
we note that:
+∞  +∞  x
∗
∑ p0,n (t) = dx ν(t − x )e−V (t|t− x) η (t − u)e− H (t−u|t− x) du
n=−∞ 0 0
 +∞
= 1− ν(t − x ) e−[V (t|t− x)+ H (t|t− x)] dx. (40)
0

Consequently, by virtue of (20), Equation (36) immediately follows. To prove Equation (37),
we first consider (18), which implies:
 t  t+kQ 
∗ +η ∗ )
q(t + kQ|t0 ) = e−kQ(ν ν(τ )e−[V (t|τ )+ H (t|τ )] dτ + ν(τ )e[V (t|τ )+ H (t|τ )] dτ . (41)
t0 t

Since ν(t) and η (t) are periodic functions, one has:

 t+kQ  k −1  Q
ν(τ )e[V (t|τ )+ H (t|τ )] dτ = ∑ ν(t + x )e[V (t+rQ+ x|t)+ H (t+rQ+ x|t)] dx
t r =0 0
 Q
 k −1
∗ ∗
= ν(t + x )e[V (t+ x|t)+ H (t+ x|t)] dx ∑ erQ(ν +η )
0 r =0
∗ ∗  
ekQ(ν +η ) − 1 Q
[V (t+ x |t)+ H (t+ x |t)]
= Q(ν∗ +η ∗ ) ν(t + x )e dx . (42)
e −1 0

Substituting (42) in (41) and taking the limit as k → +∞, one finally is led to (37).

Under the assumptions of Theorem 3, by virtue of the periodicity of ν(t) and η (t), from (35)–(37),
∗ ( t ) and q∗ ( t ) are periodic functions with period Q. From (21), making use of (35), the
one has that p0,n
asymptotic conditional moments are expressed as:

140
Mathematics 2018, 6, 81

Mr∗ (t) := lim Mr (t + kQ|t0 )

k →+∞
 +∞  x
1 -r (u|0) du,
= dx ν(t − x )e−V (t|t− x) η ( t − u ) e− H (t−u|t− x ) M (43)
1 − q∗ (t) 0 0

-r (t|0) := E[ N
with M  r (t)| N
 (0) = 0] and q∗ (t) given in (36) or (37).

Example 1. Assume that N (t) has constant arrival rates λ(t) = λ > 0 and departure rates μ(t) = μ > 0.
Furthermore, let the periodic catastrophe and repair intensity functions be given by:

πa  2πt  πb  2πt 
ν(t) = ν + sin , η (t) = η + cos , t ≥ 0, (44)
Q Q Q Q
with a > 0, b > 0, ν > πa/Q and η > πb/Q. Clearly, from (32) and (44), we have that the averages of ν(t)
and η (t) in the period Q are equal to ν and η, respectively. In Figures 3–5, the relevant parameters are taken as:

λ = 0.2, μ = 0.1, Q = 1, ν = 0.5, a = 0.1, η = 0.6, b = 0.15.

On the left of Figure 3, the catastrophe intensity function ν(t) (black curve) is plotted with its average
ν = 0.5 (black dashed line). The repair intensity function η (t) (red curve) is plotted, as well, with its average
η = 0.6 (red dashed line). On the right of Figure 3, the failure probability q(t|0), given in (18), is plotted and
is compared with the asymptotic failure probability q∗ = ν/(ν + η ) = 0.454545. The latter is obtained by
considering constant intensity functions ν(t) = ν and η (t) = η. As proved in Theorem 3, q(t|0) admits an
asymptotic periodic behavior, which is highlighted on the right of Figure 3. Instead, in Figure 4, we plot the
probability p0 (t|0) (magenta curve), on the left. Moreover, on the right of Figure 4, we show the probabilities
p−1 (t|0) (blue curve) and p1 (t|0) (red curve), given in (23). The dashed lines indicate the steady-state
probabilities p0∗ = 0.364447 (magenta dashed line), p∗−1 = 0.0470761 (blue dashed line) and p1∗ = 0.0941522
(red dashed line), obtained by considering constant intensity functions ν(t) = ν and η (t) = η. As shown
in Figure 4, the probabilities admit an asymptotic periodic behavior, with period Q = 1. Finally, in Figure 5,
the mean M1 (t|0) and the variance Var(t|0) of the process N (t), obtained via (24), are plotted and compared
with the asymptotic values (dashed lines) M1∗ = 0.2 and V ∗ = 0.64, given in (31). Figures 3–5 show that the
relevant quantities reflect the periodic nature of the rates and illustrate the limiting behavior as t grows.


















 

  
  

Figure 3. On the left: the periodic catastrophe intensity function (black curve) and repair intensity
function (red curve), with their averages (dashed lines). On the right: the failure probability q(t|t0 ),
given in (18), and the limit q∗ (dashed line), given in (29). The parameters are specified in Example 1.

141
Mathematics 2018, 6, 81








































 

   
   

Figure 4. On the left: p0 (t|0) and steady-state probability p0∗ (dashed line). On the right: p−1 (t|0) (blue
curve) and p1 (t|0) (red curve), with steady-state probabilities p∗−1 and p1∗ (dashed lines), given in (30).
The parameters are specified in Example 1.



  































    

    

Figure 5. Plots of the mean M1 (t|0) (left) and the variance of Var(t|0) (right) of the process N (t),
obtained by means of (24). The dashed lines indicate the asymptotic values M1∗ and V ∗ . The parameters
are specified in Example 1.

5. Diffusion Approximation of the Double-Ended Queueing System

With reference to the time-non-homogeneous double-ended queueing system discussed in
Section 2, hereafter, we consider a heavy-traffic diffusion approximation of the queue-length process.
This is finalized to obtain a more manageable description of the queueing system under a heavy-traffic
regime. To this purpose, we shall adopt a suitable scaling procedure based on a scaling parameter ε.
We first rename the intensity functions related to the double-ended process N  (t), by setting:


λ(t) ω 2 (t) (t) ω 2 (t)
μ
λ(t) = + , μ(t) = + , n ∈ Z. (45)
ε 2 ε2 ε 2 ε2

Here, functions  (t) and ω 2 (t) are positive, bounded and continuous for t ≥ t0 and
λ ( t ), μ
 +∞  +∞  +∞
satisfy the conditions t  λ(t) dt = +∞, t μ (t) dt = +∞ and t ω 2 (t) dt = +∞. Furthermore,
0 0 0
the constant ε in the right-hand sides of (45) is positive and plays a relevant role in the following
approximating procedure.
Let us consider the Markov process { N  ε (t), t ≥ t0 }, having state-space {0, ±ε, ±2ε, . . .}. Namely,
it is defined as N ε (t) =  N  (t), provided that the intensity functions are modified as in (45). By a
customary scaling procedure similar to that adopted in [10,30], under suitable limit conditions and
for ε ↓ 0, the scaled process N  ε (t) converges weakly to a diffusion process { X  (t), t ≥ t0 } having
state-space R and transition probability density function (pdf):

∂
f( x, t| x0 , t0 ) =  (t) ≤ x | X
P{ X  ( t0 ) = x0 }, x, x0 ∈ R, t ≥ t0 .
∂x

142
Mathematics 2018, 6, 81

Indeed, with reference to System (1), substituting pj,n (t|t0 ) with f(nε, t| jε, t0 )ε in the
 ε (t), we have:
Chapman–Kolmogorov forward differential-difference equation for N

∂ f(nε, t| jε, t0 ) 
λ(t) ω 2 (t)  
λ(t) μ (t) ω 2 (t) 
= + f [(n − 1)ε, t| jε, t0 ] − + + 2 f (nε, t| jε, t0 )
∂t ε 2 ε2 ε ε ε
μ
(t) ω (t) 
2
+ + f [(n + 1)ε, t| jε, t0 ], j, n ∈ Z.
ε 2 ε2

After setting x = nε and x0 = jε, expanding f as Taylor series and taking the limit as ε ↓ 0,
we obtain the following partial differential equation:

∂  ∂  ω 2 ( t ) ∂2 
f ( x, t| x0 , t0 ) = −[
λ(t) − μ
(t)] f ( x, t| x0 , t0 ) + f ( x, t| x0 , t0 ), x, x0 ∈ R. (46)
∂t ∂x 2 ∂x2

The associated initial condition is limt↓t0 f( x, t| x0 , t0 ) = δ( x − x0 ), where δ( x ) is the Dirac

delta-function. We remark that, due to (45), the limit ε ↓ 0 leads to a heavy-traffic condition about the
 (t). Hence, X
rates λ(t) and μ(t) of process N  (t) is a time-non-homogeneous Wiener process with drift

λ(t) − μ
(t) and infinitesimal variance ω 2 (t), with initial state x0 at time t0 . For t ≥ t0 and s ∈ R, let:
 +∞
  ( t0 ) = x0 ] =
H (s, t| x0 , t0 ) = E[eis X (t) | X eisx f( x, t| x0 , t0 ) dx, x0 ∈ R (47)
−∞

 (t). Due to (46), the characteristic function (47) is the solution of the
be the characteristic function of X
partial differential equation:

∂ ! s2 "
H (s, t| x0 , t0 ) = is [λ(t) − μ(t)] − ω 2 (t) H (s, t| x0 , t0 ), x 0 ∈ R,
∂t 2

to be solved with the initial condition limt↓t0 H (s, t| x0 , t0 ) = eisx0 . Hence, for t ≥ t0 , one has:

!  2 "
H (s, t| x0 , t0 ) = exp is x0 + Λ  ( t | t0 ) − s Ω ( t | t0 ) ,
 ( t | t0 ) − M x 0 ∈ R, (48)
2
where we have set:
 t  t  t
 ( t | t0 ) =
Λ 
λ(τ ) dτ,  ( t | t0 ) =
M (τ ) dτ,
μ Ω ( t | t0 ) = ω 2 (τ ) dτ, t ≥ t0 . (49)
t0 t0 t0

Clearly, (48) is a normal characteristic function, so that the solution of (46) is the Gaussian pdf:
 # $
 ( t | t0 ) 2 
 ( t | t0 ) + M
1 x − x0 − Λ
f( x, t| x0 , t0 ) =  exp − , x, x0 ∈ R, t ≥ t0 . (50)
2 π Ω ( t | t0 ) 2 Ω ( t | t0 )

Then, the conditional mean and variance are:

 (t)| X
E[ X  ( t | t0 ) − M
 ( t0 ) = x0 ] = x0 + Λ  ( t | t0 ),  (t)| X
Var[ X  ( t0 ) = x0 ] = Ω ( t | t0 ), t ≥ t0 . (51)

Let us now consider a first-passage-time problem for X  (t). We denote by Tx ,x (t0 ) the random
0
variable describing the FPT of X (t) trough state x ∈ R, starting from x0 at time t0 , with x0 = x.
In analogy to (8), the Markov property yields:
 t
f( x, t| x0 , t0 ) = g( x, τ | x0 , t0 ) f( x, t| x, τ ) dτ, x 0 , x ∈ R, x = x 0 , (52)
t0

143
Mathematics 2018, 6, 81

x ,x (t0 ).
where g( x, t| x0 , t0 ) is the pdf of T 0
Hereafter, we deal with a special case, in which the functions  (t) and ω 2 (t) are proportional.
λ ( t ), μ

Remark 2. Let  λ(t) =   ϕ(t) and ω 2 (t) = ω 2 ϕ(t), where 

(t) = μ
λϕ(t), μ , ω are positive constants
λ, μ
∞
and ϕ(t) is a positive, bounded and continuous function for t ≥ t0 , such that t ϕ(t) dt = +∞. Then,
0
 (t) becomes:
the transition pdf of X
 # $2 
1 x − x0 − ( 
λ−μ ) Φ(t|t0 )
f( x, t| x0 , t0 ) =  exp − , x, x0 ∈ R, t ≥ t0 ,
2 π ω 2 Φ ( t | t0 ) 2 ω 2 Φ ( t | t0 )

x ,x (t0 ) can be expressed as (see, for instance, [26]):

where Φ(t|t0 ) is defined in (9). Moreover, the FPT pdf of T 0

| x − x0 | ϕ ( t ) 
g( x, t| x0 , t0 ) = f ( x, t| x0 , t0 ), x 0 , x ∈ R, x = x 0 .
Φ ( t | t0 )

The corresponding FPT ultimate probability is given by:

 +∞

1, (
λ−μ
)( x − x0 ) ≥ 0,
x ,x (t0 ) < +∞} =
P{ T g( x, t| x0 , t0 ) dt =
2( 
λ−μ
)( x − x0 )/ω 2
(
0
t0 e , λ−μ
)( x − x0 ) < 0.

x ,0 (t0 ) is a suitable approximation of the busy period Tj,0 (t0 ) considered in Section 2.
Clearly, T 0

Goodness of the Approximating Procedure

Thanks to the above heavy-traffic approximation, the state of the time-non-homogeneous
double-ended queue N  (t) has been approximated by the non-homogeneous Wiener process X  (t), with
the transition pdf given in (50).
A first confirmation of the goodness of the approximating procedure can be obtained by the
comparing mean and variance of N  (t) with those of X
 (t)/ε, for λ(t) and μ(t) chosen as in (45).
Recalling (7) and (51), the means satisfy the following identity, for all ε > 0:

 (t)| X
E[ X  (t0 ) = jε] = εE[ N
 (t)| N
 ( t0 ) = j ]. (53)

Moreover, for the variances, we have:

 (t)| N
Var[ N  ( t0 ) = j ]  (t)| N
ε2 Var[ N  ( t0 ) = j ] ε2 [Λ(t|t0 ) + M (t|t0 )]
lim   = lim = lim = 1,
ε ↓0 X (t) X (t ) ε↓0 Var[ X (t)| X
 (t0 ) = jε] ε ↓0 Ω ( t | t0 )
Var ε ε 0 = j

so that for ε close to zero, one has:

 (t)| X
Var[ X  (t0 ) = jε]  ε2 Var[ N
 (t)| N
 ( t0 ) = j ]. (54)

The discussion of the goodness of the heavy-traffic approximation involves also the probability
(ε)  (t), for λ(t) and μ(t)
distributions. Let us denote by pj,n (t) the transition probabilities of the process N
given in (45), and for n = x/ε, j = x0 /ε. The following theorem holds.

Theorem 4. For t ≥ t0 , one has:

(ε)
pj,n (t|t0 )
lim = f( x, t| x0 , t0 ), (55)
ε↓0, ε
nε= x,jε= x0

with f( x, t| x0 , t0 ) given in (50).

144
Mathematics 2018, 6, 81

Proof. To prove (55), we consider separately the following cases: (i) n = j and (ii) n = j, with j, n ∈ Z.

(i) For n = j, from (6), one has:

%
! Λ  
(ε)  ( t | t0 ) + M
 ( t | t0 ) Ω ( t | t0 ) " Λ
 ( t | t0 ) Ω ( t | t0 )  M
 ( t | t0 ) Ω ( t | t0 )
pn,n (t|t0 ) = exp − − I0 2 + + . (56)
ε ε 2 ε 2ε 2 ε 2ε 2

√
We recall that Iν (z)  ez / 2πz (cf. [31], p. 377, n. 9.71) when |z| is large, for ν fixed. Hence,
from (56) as ε is close to zero, one has:

 (t|t0 ) Ω(t|t0 ) −1/4  M

1 Λ  (t|t0 ) Ω(t|t0 ) −1/4
(ε)
pn,n (t|t0 )
 √ + +
ε 2ε π ε 2ε2 ε 2ε2
 % % 2 &
 ( t | t0 ) Ω ( t | t0 )
Λ M ( t | t0 ) Ω ( t | t0 )
× exp − + − + , n ∈ Z. (57)
ε 2ε2 ε 2ε2

We note that:

1 Λ (t|t0 ) Ω(t|t0 ) −1/4  M  (t|t0 ) Ω(t|t0 ) −1/4

lim √ + +
ε ↓0 2ε π ε 2ε 2 ε 2ε2
  −1/4
 ( t | t0 ) + Ω ( t | t0 )  ( t | t0 ) + Ω ( t | t0 )
1 − 1/4 1
= √ lim εΛ εM = 
2 π ε ↓ 0 2 2 2 π Ω ( t | t0 )

and:

 % % 2 &
 ( t | t0 ) Ω ( t | t0 )
Λ M ( t | t0 ) Ω ( t | t0 )
lim exp − + − +
ε ↓0 ε 2ε2 ε 2ε2
  . .  −2 &
# $2 Ω ( t | t0 ) Ω ( t | t0 )

= lim exp − Λ(t|t0 ) − M(t|t0 ) 
ε Λ ( t | t0 ) + 
+ ε M ( t | t0 ) +
ε ↓0 2 2
 #  ( t | t0 )
$2 
Λ ( t | t0 ) − M
= exp − ,
2 Ω ( t | t0 )

so that, taking the limit as ε ↓ 0 in (57), Equation (55) follows for n = j and x = x0 .

(ii) For n = j, recalling that In (z) = I−n (z), from (6), one has:

! Λ   ( x− x0 )/(2ε)
(ε)  ( t | t0 ) + M  (t|t0 ) Ω(t|t0 ) " 2εΛ ( t | t0 ) + Ω ( t | t0 )
pj,n (t|t0 ) = exp − −
ε ε 2
2ε M  ( t | t0 ) + Ω ( t | t0 )
%
 Λ ( t | t0 ) Ω ( t | t0 )  M 
 ( t | t0 ) Ω ( t | t0 )
× I| x− x0 |/ε 2 + + . (58)
ε 2ε2 ε 2ε2

Making use of the asymptotic result (cf. [31], p. 378, n. 9.7.7):

1 !  z "
Iν (ν z)  √ exp ν 1 + z2 + ln √ , ν → +∞, 0 < z < +∞,
2πν (1 + z2 )1/4 1 + 1 + z2

from (58), we obtain:

(ε)
pj,n (t|t0 ) 4
(ε)
 ∏ A j ( x, t| x0 , t0 ), (59)
ε j =1

145
Mathematics 2018, 6, 81

where:
  (t|t0 ) + Ω(t|t0 ) ( x− x0 )/(2ε)
(ε) 2εΛ
A1 ( x, t| x0 , t0 ) =
2ε M  ( t | t0 ) + Ω ( t | t0 )
   −1/4
(ε) 1  (t|t0 ) + Ω(t|t0 ) 2ε M  ( t | t0 ) + Ω ( t | t0 )
A2 ( x, t| x0 , t0 ) = √ ε2 ( x − x0 )2 + 2εΛ
2π
 # $# $ & |x−εx0 |
2εΛ  (t|t0 ) + Ω(t|t0 ) 2ε M  ( t | t0 ) + Ω ( t | t0 )
(ε)
A3 ( x, t| x0 , t0 ) = # $# $
ε | x − x0 | + ε2 ( x − x0 )2 + 2εΛ  (t|t0 ) + Ω(t|t0 ) 2ε M  ( t | t0 ) + Ω ( t | t0 )
! Λ  ( t | t0 ) + M ( t | t0 ) Ω ( t | t0 ) "
(ε)
A4 ( x, t| x0 , t0 ) = exp − −
ε ε2
!1 # $# $"

× exp 2 ε ( x − x0 ) + 2εΛ(t|t0 ) + Ω(t|t0 ) 2ε M
2 2  ( t | t0 ) + Ω ( t | t0 ) .
ε
Since:
! Λ ( t | t0 ) − M  ( t | t0 ) "
(ε)
lim A1 ( x, t| x0 , t0 ) = exp ( x − x0 ) ,
ε ↓0 Ω ( t | t0 )
(ε) 1
lim A2 ( x, t| x0 , t0 ) =  ,
ε ↓0 2 π Ω ( t | t0 )
! ( x − x )2 "
(ε) 0
lim A3 ( x, t| x0 , t0 ) = exp − ,
ε ↓0 Ω ( t | t0 )
! ( x − x )2 " ! [Λ  (t|t0 )]2 "
 ( t | t0 ) − M
(ε) 0
lim A4 ( x, t| x0 , t0 ) = exp exp − ,
ε ↓0 2Ω(t|t0 ) 2Ω(t|t0 )

by taking the limit as ε ↓ 0 in (59), Equation (55) follows for n = j and x = x0 .

Finally, the goodness of the heavy-traffic approximation is confirmed by the approximation:

(ε)
pj,n (t|t0 )  ε f( x, t| x0 , t0 ),

which is a consequence of Equation (55) and is valid for ε close to zero.

6. Diffusion Approximation of the Double-Ended Queueing System with Catastrophes

and Repairs
In this section, we consider a heavy-traffic approximation of the time-non-homogeneous
double-ended queueing system subject to disasters and repairs, discussed in Section 3. The continuous
approximation of the discrete model leads to a jump-diffusion process and is similar to the scaling
procedure employed in Section 5. The relevant difference is that the state-space of the process N (t)
presents also a spurious state F.
Let us now consider the continuous-time Markov process { Nε (t), t ≥ t0 }, having state-space
{ F, 0, ±ε, ±2ε, . . .}. Under suitable limit conditions, as ε ↓ 0, the scaled process Nε (t) converges
weakly to a jump-diffusion process { X (t), t ≥ t0 } having state-space { F } ∪ R. The limiting procedure
is analogous to that used in Buonocore et al. [34], which involves spurious states, as well. As in
the previous section, for the approximating procedure, we first assume that the rates λ(t) and
μ(t) are modified as in (45). Hence, the limit ε ↓ 0 leads to a heavy-traffic condition about such
intensity functions. Instead, the catastrophe rate ν(t) and the repair rate η (t) are not affected by the
scaling procedure.
We note that X (t) describes the motion of a particle, which starts at the origin at time t0
and then behaves as a non-homogeneous Wiener process, with drift  λ(t) − μ(t) and infinitesimal
variance ω 2 (t), until a catastrophe occurs. We remark that the catastrophes arrive according to a

146
Mathematics 2018, 6, 81

time-non-homogeneous Poisson process with intensity function ν(t). As soon as a catastrophe occurs,
the process enters into the failure state F and remains therein for a random time (corresponding to the
repair time) that ends according to the time-dependent intensity function η (t). Clearly, catastrophes
are not allowed during a repair period. The effect of a repair is the instantaneous transition of the
process X (t) to the state zero. After that, the motion is subject again to diffusion and proceeds as
before. We recall that ν(t) and η (t) are positive, bounded and continuous functions for t ≥ t0 , such
 +∞  +∞
that t ν(t) dt = +∞ and t η (t) dt = +∞. We denote by:
0 0

∂
f ( x, t|0, t0 ) = P { X ( t ) ≤ x | X ( t0 ) = 0}, x ∈ R, t ≥ 0 (60)
∂x

the transition density of X (t) and by q(t|t0 ) = P{ X (t) = F | X (t0 ) = 0} the probability that the process
is in the failure-state at time t starting from zero at time t0 . We point out that the adopted scaling
procedure does not affect the spurious state, so that q(t|t0 ) is identical to the analogous probability of
the process N (t) and is given in (18). From (15), proceeding similarly as for (46), one obtains that (60)
is the solution of the following partial differential equation, for t > t0 :

∂ ∂ ω 2 ( t ) ∂2
f ( x, t|0, t0 ) = −ν(t) f ( x, t|0, t0 ) − [
λ(t) − μ
(t)] f ( x, t|0, t0 ) + f ( x, t|0, t0 ), x ∈ R \ {0}, (61)
∂t ∂x 2 ∂x2

to be solved with the initial condition limt↓t0 f ( x, t|0, t0 ) = δ( x ) and, in analogy to (20), with the
compatibility condition:
 +∞
f ( x, t|0, t0 ) dx + q(t|t0 ) = 1, t ≥ t0 . (62)
−∞

6.1. Transient Distribution

Similarly to the discrete model discussed in Section 3, the pdf (60) can be expressed as follows, in
terms of the transition pdf of the time-non-homogeneous Wiener process X  (t) treated in Section 5:
 t
f ( x, t|0, t0 ) = e−V (t|t0 ) f( x, t|0, t0 ) + q(τ |t0 ) η (τ ) e−V (t|τ ) f( x, t|0, τ ) dτ, x ∈ R, t ≥ 0, (63)
t0

with q(t|t0 ) and f( x, t| x0 , t0 ) given in (18) and (50), respectively. Making use of (18) and (50) in (63),
for t ≥ t0 and x ∈ R, one has:
 #  ( t | t0 ) + M
$
 ( t | t0 ) 2 
e −V ( t | t0 ) x−Λ
f ( x, t|0, t0 ) =  exp −
2 π Ω ( t | t0 ) 2 Ω ( t | t0 )
 # $
 t
e −V ( t | τ ) x−Λ  (t|τ ) + M  (t|τ ) 2   τ
+ dτ η (τ )  exp − ν(ϑ )e−[V (τ |ϑ)+ H (τ |ϑ)] dϑ. (64)
t0 2 π Ω(t|τ ) 2 Ω(t|τ ) t0

For r ∈ N, let us now consider the r-th conditional moment of X (t):

 +∞
1
Mr (t|t0 ) := E[ X r (t)| X (t) ∈ R, X (t0 ) = 0] = xr f ( x, t|0, t0 ) dx. (65)
1 − q ( t | t0 ) −∞

From (63), for r ∈ N, it results:

  t 
1 - r ( t | t0 ) + -r (t|τ ) dτ ,
Mr ( t | t 0 ) = e −V ( t | t0 ) M q ( τ | t0 ) η ( τ ) e −V ( t | τ ) M (66)
1 − q ( t | t0 ) t0

where M  r (t)| X
- r ( t | t0 ) : = E[ X  (t0 ) = 0] is the r-th conditional moment of X
 (t). Hence, by virtue of (51),
from (66), we obtain the conditional moments Mr (t|t0 ).

147
Mathematics 2018, 6, 81

In the following theorem, we discuss the special case when the functions 
λ(t) − μ
(t) and ω 2 (t)
are constant.

Theorem 5. Consider the process X (t) such that  (t) = 

λ(t) − μ λ−μ
 and ω 2 (t) = ω 2 for all t ≥ t0 . Then,
for t ≥ t0 and x ∈ R, one has:
 t − t0  x
f ( x, t|0, t0 ) = e−V (t|t0 ) f( x, t − t0 |0, 0) + dx ν(t − x )e−V (t|t− x) η (t − u)e− H (t−u|t− x) f( x, u|0, 0) du (67)
0 0

and, for r ∈ N,
  t − t0 
1 - r ( t − t0 |0) + -r ( x |0) dx .
Mr ( t | t 0 ) = e −V ( t | t0 ) M q ( t − x | t0 ) η ( t − x ) e −V ( t | t − x ) M (68)
1 − q ( t | t0 ) 0

Furthermore, it results:
 t
f ( x, t|0, t0 ) = f (0, τ |0, t0 ) e−V (t|τ ) g( x, t|0, τ ) dτ, x ∈ R \ {0}, t ≥ t0 , (69)
t0

0,x (τ ), introduced in Section 5.

where g( x, t|0, τ ) is the FPT pdf of T

Proof. It proceeds similarly to the proof of Theorem 1.

6.2. Goodness of the Approximating Procedure

Let us now analyze the goodness of the heavy-traffic approximation considered above. The
time-non-homogeneous process describing the state of the double-ended queueing system with
catastrophes and repairs has been approximated by the diffusion process X (t), whose transition pdf is
given in (63).
First of all, we compare the mean, second order moment and variance of N (t) with those of
X (t)/ε, when λ(t) and μ(t) are chosen as in (45). By virtue of (53) and (54), one has:

M  (t)| X
- 1 ( t | t0 ) = E[ X  (t0 ) = 0] = εE[ N (t)| N
 ( t0 ) = 0] = ε M1 ( t | t0 ),
M  (t)| X
- 2 ( t | t0 ) = E[ X 2  ( t0 ) = 0]  ε E[ N
2  (t)| N
2  ( t0 ) = 0] = M2 ( t | t0 ) as ε ↓ 0.

Hence, recalling (22) and (66), one has:

M1 (t|t0 ) ≡ E[ X (t)| X (t0 ) = 0] = εE[ N (t)| N (t0 ) = 0] ≡ εM1 (t|t0 ).

Moreover,
 
-2 (t|t0 ) + t q(τ |t0 ) η (τ )e−V (t|τ ) M
ε 2 e −V ( t | t0 ) M -2 (t|τ ) dτ
E[ N 2 (t)| N (t0 ) = 0] t0
lim  2 = lim  = 1,
ε ↓0 E X ( t ) X ( t0 ) = 0 ε ↓0 - 2 ( t | t0 ) + t q ( τ | t0 ) η ( τ ) e −V ( t | τ ) M
e −V ( t | t0 ) M -2 (t|τ ) dτ
ε2
ε t0

so that the variances satisfy the following relation, for ε close to zero:

Var[ X (t)| X (t0 ) = 0]  ε2 Var[ N (t)| N (t0 ) = 0].

(ε)
Furthermore, we denote by p j,n (t) the transition probabilities of the process N (t), when n = x/ε
and the intensity functions λ(t) and μ(t) are given in (45). The following theorem holds.

148
Mathematics 2018, 6, 81

Theorem 6. For t ≥ t0 , one has:

(ε)
p0,n (t|t0 )
lim = f ( x, t|0, t0 ), (70)
ε↓0, nε= x ε

with f ( x, t|, 0, t0 ) given in (63).

Proof. From (19), one obtains:

(ε) (ε)  t (ε)

p0,n (t|t0 ) p0,n (t|t0 ) p0,n (t|τ )
= e −V ( t | t0 ) + q ( τ | t0 ) η ( τ ) e −V ( t | τ ) dτ. (71)
ε ε t0 ε

Taking the limit as ε ↓ 0 on both sides of (71) and recalling (55), for t ≥ t0 , one has:

(ε)  t
p0,n (t|t0 )
lim = e−V (t|t0 ) f( x, t|0, t0 ) + q(τ |t0 ) η (τ ) e−V (t|τ ) f( x, t|0, τ ) dτ,
ε↓0, nε= x ε t0

so that (70) immediately follows by using (63).

As a consequence of Theorem 6, for λ(t) and μ(t) chosen as in (45) and under heavy-traffic
(ε)
conditions, the probability p0,n (t|t0 ) of the discrete process N (t) is close to ε f (nε, t|0, t0 ) for ε near
to zero.

7. Asymptotic Distributions
Similar to the analysis performed in Section 4, in this section, we consider the asymptotic behavior
of the density f ( x, t|0, t0 ) of the process X (t) in two different cases:

(i) the functions  (t), ω 2 (t), ν(t) and η (t) admit finite positive limits as t → +∞,
λ ( t ), μ

(ii) the functions λ(t), μ (t) and ω 2 (t) are constant, and the rates ν(t) and η (t) are periodic functions
with common period Q.

7.1. Asymptotically-Constant Intensity Functions

We assume that the functions  (t), ω 2 (t), ν(t) and η (t) admit finite positive limits as t tends
λ ( t ), μ
to +∞. In this case, the failure asymptotic probability q∗ = limt→+∞ q(t|t0 ) of the process X (t) is
provided in (29). Moreover, the steady-state density of the process X (t) is an asymmetric bilateral
exponential density, as given in the following theorem.

Theorem 7. Assuming that:

lim λ(t) = λ, lim μ(t) = μ, lim ω 2 (t) = ω 2 , lim ν(t) = ν, lim η (t) = η, (72)
t→+∞ t→+∞ t→+∞ t→+∞ t→+∞

with  , ω 2 , ν, η positive constants, then the steady-state pdf of the process X (t) is, for x ∈ R,
λ, μ

ην 1 ! (
λ−μ
) (
λ−μ
)2 + 2ω 2 ν "
f ∗ ( x ) := lim f ( x, t|0, t0 ) = exp x− | x | . (73)
t→+∞ η+ν ω2 ω2
(
λ−μ
)2 + 2ω 2 ν

Furthermore, the asymptotic conditional mean, second order moment and variance are:


λ−μ  2( 
λ−μ )2 ω2
lim M1 (t|t0 ) = , lim M2 (t|t0 ) =
+ ,
t→+∞ t→+∞ν ν 2 ν
(74)
(λ−μ )2 ω2
lim Var(t|t0 ) = lim {M2 (t|t0 ) − [M1 (t|t0 )]2 } = + .
t→+∞ t→+∞ ν 2 ν

149
Mathematics 2018, 6, 81

Proof. The steady-state density can be obtained by taking the limit as t → +∞ in Equations (61) and
(62) and recalling (29). Moreover, the asymptotic conditional mean and variance (74) follow from (65),
making use of (29) and (73).

7.2. Periodic Intensity Functions

Let us assume that the functions  (t) and ω 2 (t) are constant and that the catastrophe
λ ( t ), μ
intensity function ν(t) and the repair intensity function η (t) are periodic, so that ν(t + kQ) = ν(t) and
η (t + kQ) = η (t) for k ∈ N and t ≥ t0 . The average catastrophe and repair rates in the period Q are
defined in (32). The asymptotic distribution of the process X (t) is described by the following functions,
for t ≥ t0 ,

q∗ (t) := lim q(t + kQ|t0 ), f ∗ ( x, t) := lim f ( x, t + kQ|0, t0 ), x ∈ R. (75)

k →+∞ k →+∞

Note that the asymptotic failure probability q∗ (t) is given in (36) or, alternatively, in (37). Moreover,
the asymptotic density f ∗ ( x, t) is determined in the following theorem.

Theorem 8. Consider the stochastic process X (t) and assume  λ(t) − μ(t) = λ−μ  and ω 2 (t) = ω 2 , and that
the intensities ν(t) and η (t) are continuous, positive and periodic functions with period Q. Then, one has:
 +∞  x
f ∗ ( x, t) = dx ν(t − x )e−V (t|t− x) η (t − u)e− H (t−u|t− x) f( x, u|0, 0) du, x ∈ R. (76)
0 0

Proof. It proceeds similarly to the proof of Theorem 3, by starting from Equation (67).

By virtue of the periodicity of ν(t) and η (t), from (76), one has that f ∗ ( x, t) is a periodic function
with period Q. From (65), making use of (76), the asymptotic conditional moments are:

Mr∗ (t) := lim Mr (t + kQ|t0 )

k →+∞
 +∞  x
1 -r (u|0) du,
= dx ν(t − x )e−V (t|t− x) η ( t − u ) e− H (t−u|t− x ) M (77)
1 − q∗ (t) 0 0

where M  ( t )r | X
- r ( t |0) = E[ X  (0) = 0] and where q∗ (t) is given in (36) or in (37).
The following illustrative example concludes the section.

Example 2. Let X (t) be the approximating jump-diffusion process, subject to disasters and repairs, with drift

λ−μ  and infinitesimal variance ω 2 , where  λ = 2.0, μ  = 1.0 and ω 2 = 0.2 and with periodic catastrophe
intensity function ν(t) and repair intensity function η (t) given by (44). The parameters ν, a, η, b, Q are taken
as in Example 1. For these choices, probability q(t|0) is identical as for the discrete model. It is plotted in
Figure 3, on the right.
We now consider the two choices ε = 0.05 and ε = 0.025. Then, the parameters λ and μ are determined
according to (45), so that for ε = 0.05, we have λ = 80 and μ = 60, whereas for ε = 0.025, we have λ = 240
and μ = 200. To show the validity of the approximating procedure, we compare the quantity ε f (ε n, t|0, 0)
with the probability p0,n (t|0), for n = 0, −1, 1, in Figures 6–8, respectively. The case ε = 0.05 is shown on
the left, and ε = 0.025 is on the right. Recall that f ( x, t|0, 0) is given in (67), whereas p0,n (t|0) is given in
(23). We note that the goodness of the continuous approximation for p0,n (t|0) improves as ε decreases, this
corresponding to an increase of traffic in the double-ended queue with catastrophes and repairs, due to (45).

150
Mathematics 2018, 6, 81



































 

    
    

Figure 6. For λ = 2.0, μ  = 1.0, ω 2 = 0.2, the function ε f (0, t|0, 0) (red curve) is shown with the
probability p0,0 (t|0) (black dashed curve) for ε = 0.05 (left) and ε = 0.025 (right). The parameters λ
and μ are shown in Example 2, according to (45).












































    
    

Figure 7. For the same choices of parameters of Figure 6, the function ε f (−ε, t|0, 0) (red curve) is
compared with the probability p0,−1 (t|0) (black dashed curve) for ε = 0.05 (left) and ε = 0.025 (right).
































 

    
    

Figure 8. For the same choices of parameters of Figure 6, the function ε f (ε, t|0, 0) (red curve) is
compared with the probability p0,1 (t|0) (black dashed curve) for ε = 0.05 (left) and ε = 0.025 (right).

8. Conclusions
We analyzed a continuous-time stochastic process that describes the state of a double-ended queue
subject to disasters and repairs. The system is time-non-homogeneous, since arrivals, services, disasters
and repairs are governed by time-varying intensity functions. This model is a suitable generalization
of the queueing system investigated in [10]. Indeed, the previous model is characterized by constant
rates of arrivals, services, catastrophes and repairs. However, motivated by the need to describe more
realistic situations in which the system evolution reflects daily or seasonal fluctuations, in this paper,
we investigated the case where all such rates are time-dependent. Whereas in the previous model, the
approach involved the Laplace transforms, in the present case, the analysis cannot be based on such
a method, but rather on a direct analysis of the relevant equations. Our analysis involved also the
heavy-traffic approximation of the system, which leads to a time-non-homogeneous diffusion process

151
Mathematics 2018, 6, 81

useful to describe the queue-length dynamics via more manageable formulas. Future developments of
the present investigation will be centered on the inclusion of multiple types of customers and more
general forms of catastrophe/repair mechanisms.

Author Contributions: All the authors contributed equally to this work.

Acknowledgments: A.D.C., V.G. and A.G.N. are members of the group GNCS of INdAM.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Kashyap, B.R.K. A double-ended queueing system with limited waiting space. Proc. Natl. Inst. Sci. India A
1965, 31, 559–570. doi:10.1007/BF02613588. [CrossRef]
2. Kashyap, B.R.K. The double-ended queue with bulk service and limited waiting space. Oper. Res. 1966, 14,
822–834. doi:10.1287/opre.14.5.822. [CrossRef]
3. Sharma, O.P.; Nair, N.S.K. Transient behaviour of a double ended Markovian queue. Stoch. Anal. Appl. 1991,
9, 71–83. doi:10.1080/07362999108809226. [CrossRef]
4. Tarabia, A.M.K. On the transient behaviour of a double ended Markovian queue. J. Combin. Inf. Syst. Sci.
2001, 26, 125–134.
5. Conolly, B.W.; Parthasarathy, P.R.; Selvaraju, N. Doubled-ended queues with impatience. Comput. Oper. Res.
2002, 29, 2053–2072. doi:10.1016/S0305-0548(01)00075-2. [CrossRef]
6. Elalouf, A.; Perlman, Y.; Yechiali, U. A double-ended queueing model for dynamic allocation of live organs
based on a best-fit criterion. Appl. Math. Model. 2018, in press. doi:10.1016/j.apm.2018.03.022. [CrossRef]
7. Takahashi, M.; Ōsawa, H.; Fujisawa, T. On a synchronization queue with two finite buffers. Queueing Syst.
2000, 36, 107–123. doi:10.1023/A:1019127002333. [CrossRef]
8. Di Crescenzo, A.; Giorno, V.; Nobile, A.G. Constructing transient birth-death processes by means of suitable
transformations. Appl. Math. Comput. 2016, 281, 152–171. doi:10.1016/j.amc.2016.01.058. [CrossRef]
9. Di Crescenzo, A.; Giorno, V.; Nobile, A.G. Analysis of reflected diffusions via an exponential time-based
transformation. J. Stat. Phys. 2016, 163, 1425–1453. doi:10.1007/s10955-016-1525-9. [CrossRef]
10. Di Crescenzo A.; Giorno V.; Krishna Kumar B.; Nobile, A.G. A double-ended queue with catastrophes
and repairs, and a jump-diffusion approximation. Methodol. Comput. Appl. Probab. 2012, 14, 937–954.
doi:10.1007/s11009-011-9214-2. [CrossRef]
11. Altiok, T. On the phase-type approximations of general distributions. IIE Trans. 1985, 17, 110–116. [CrossRef]
12. Altiok, T. Queueing modeling of a single processor with failures. Perform. Eval. 1989, 9, 93–102. [CrossRef]
13. Altiok, T. Performance analysis of manufacturing systems. Springer Series in Operations Research; Springer:
New York, NY, USA, 1997.
14. Dallery, Y. On modeling failure and repair times in stochastic models of manufacturing systems using
generalized exponential distributions. Queueing Syst. 1994, 15, 199–209. [CrossRef]
15. Economou, A.; Fakinos, D. A continuous-time Markov chain under the influence of a regulating point
process and applications in stochastic models with catastrophes. Eur. J. Oper. Res. 2003, 149, 625–640.
doi:10.1016/S0377-2217(02)00465-4. [CrossRef]
16. Economou, A.; Fakinos, D. Alternative approaches for the transient analysis of Markov chains with
catastrophes. J. Stat. Theory Pract. 2008, 2, 183–197. doi:10.1080/15598608.2008.10411870. [CrossRef]
17. Kyriakidis, E.; Dimitrakos, T. Computation of the optimal policy for the control of a compound
immigration process through total catastrophes. Methodol. Comput. Appl. Probab. 2005, 7, 97–118.
doi:10.1007/s11009-005-6657-3. [CrossRef]
18. Krishna Kumar, B.; Vijayakumar, A.; Sophia, S. Transient analysis for state-dependent queues with
catastrophes. Stoch. Anal. Appl. 2008, 26, 1201–1217. doi:10.1080/07362990802405786. [CrossRef]
19. Di Crescenzo, A.; Giorno, V.; Nobile, A.G.; Ricciardi, L.M. A note on birth-death processes with catastrophes.
Stat. Probab. Lett. 2008, 78, 2248–2257. doi:10.1016/j.spl.2008.01.093. [CrossRef]
20. Zeifman, A.; Korotysheva, A. Perturbation bounds for Mt /Mt /N queue with catastrophes. Stoch. Models
2012, 28, 49–62. doi:10.1080/15326349.2011.614900. [CrossRef]
21. Zeifman, A.; Satin, Y.; Panfilova, T. Limiting characteristics for finite birth-death-catastrophe processes.
Math. Biosci. 2013, 245, 96–102. doi:10.1016/j.mbs.2013.02.009. [CrossRef] [PubMed]

152
Mathematics 2018, 6, 81

22. Giorno, V.; Nobile, A.G.; Pirozzi, E. A state-dependent queueing system with asymptotic logarithmic
distribution. J. Math. Anal. Appl. 2018, 458, 949–966. doi:10.1016/j.jmaa.2017.10.004. [CrossRef]
23. Di Crescenzo, A.; Giorno, V.; Nobile, A.G.; Ricciardi, L.M. On time non-homogeneous stochastic processes
with catastrophes. In Cybernetics and Systems 2010, Proceedings of the Austrian Society for Cybernetics Studies
(EMCSR 2010), Vienna, Austria, 6–9 April 2010; Trappl, R., Ed.; Austrian Society for Cybernetic Studies: Vienna,
Austria, 2010; pp. 169–174.
24. Giorno, V.; Nobile, A.G.; Spina, S. On some time non-homogeneous queueing systems with catastrophes.
Appl. Math. Comput. 2014, 245, 220–234. doi:10.1016/j.amc.2014.07.076. [CrossRef]
25. Dimou, S.; Economou, A. The single server queue with catastrophes and geometric reneging. Methodol.
Comput. Appl. Probab. 2013, 15, 595–621. doi:10.1007/s11009-011-9271-6. [CrossRef]
26. Giorno, V.; Nobile, A.G.; Ricciardi, L.M. On some time-non-homogeneous diffusion approximations to
queueing systems. Adv. Appl. Probab. 1987, 19, 974–994. doi:10.2307/1427111. [CrossRef]
27. Liu, X.; Gong, Q.; Kulkarni, V.G. Diffusion models for double-ended queues with renewal arrival processes.
Stoch. Syst. 2015, 5, 1–61. doi:10.1214/13-SSY113. [CrossRef]
28. Büke, B.; Chen, H. Fluid and diffusion approximations of probabilistic matching systems. Queueing Syst. 2017
86, 1–33. doi:10.1007/s11134-017-9516-3. [CrossRef]
29. Di Crescenzo, A.; Giorno, V.; Nobile, A.G.; Ricciardi, L.M. On the M/M/1 queue with catastrophes and its
continuous approximation. Queueing Syst. 2003, 43, 329–347. doi:10.1023/A:1023261830362. [CrossRef]
30. Dharmaraja, S.; Di Crescenzo, A.; Giorno, V.; Nobile, A.G. A continuous-time Ehrenfest model
with catastrophes and its jump-diffusion approximation. J. Stat. Phys. 2015, 161, 326–345.
doi:10.1007/s10955-015-1336-4. [CrossRef]
31. Abramowitz, M.; Stegun, I. Handbook of Mathematical Functions with Formulas, Graphs and Mathematical Tables;
Dover Publications, Inc.: New York, NY, USA, 1972.
32. Irwin, J.O. The frequency-distribution of the difference between two independent variates following the same
Poisson distribution. J. R. Stat. Soc. 1937, 100, 415–416. [CrossRef]
33. Skellam, J.G. The frequency-distribution of the difference between two Poisson variates belonging to different
populations distribution. J. R. Stat. Soc. 1946, 109, 296. [CrossRef]
34. Buonocore, A.; Di Crescenzo, A.; Giorno, V.; Nobile, A.G.; Ricciardi, L.M. A Markov chain-based model for
actomyosin dynamics. Sci. Math. Jpn. 2009, 70, 159–174.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

153
 mathematics
Article
On the Number of Periodic Inspections During
Outbreaks of Discrete-Time Stochastic SIS
Epidemic Models
Maria Gamboa and Maria Jesus Lopez-Herrero *
Department of Statistics and Data Science, Faculty of Statistical Studies, Complutense University of Madrid,
28040 Madrid, Spain; [email protected]
* Correspondence: [email protected]

Received: 4 July 2018; Accepted: 20 July 2018; Published: 24 July 2018

Abstract: This paper deals with an infective process of type SIS, taking place in a closed population
of moderate size that is inspected periodically. Our aim is to study the number of inspections that
find the epidemic process still in progress. As the underlying mathematical model involves a discrete
time Markov chain (DTMC) with a single absorbing state, the number of inspections in an outbreak
is a first-passage time into this absorbing state. Cumulative probabilities are numerically determined
from a recursive algorithm and expected values came from explicit expressions.

Keywords: discrete time stochastic model; first-passage time; time between inspections

1. Introduction
The spread of infectious diseases is a major concern for human populations. Disease control
by any therapy measure lies on the understanding of the disease itself. Epidemic modeling is an
interdisciplinary subject that can be addressed from deterministic applied mathematical models and
stochastic processes theory [1,2] to quantitative social and biological science and empirical analysis of
data [3]. Epidemic models are also used to describe spreading processes connected with technology,
business, marketing or sociology, where the interest is related to dissemination of news, rumors or
ideas among different groups [4–8]. Mathematical models provide an essential tool for understanding
and forecasting the spread of infectious diseases and also to suggest control policies.
There is a large variety of models for describing the spread of an infective process [1,2].
An essential distinction is done between deterministic and stochastic models. Deterministic models
constitute a vast majority of the existing literature and are formulated in terms of ordinary differential
equations (ODE); consequently they predict the same dynamic for an infective process given the same
initial conditions. However this is not what it is expected to happen in real world diseases, outbreaks
do not involve the same people becoming infected at the same time and uncertainty should be included
when modeling diseases. The stochastic models [2], analogous to those defined by ODE, take into
account the random nature of the events and, hence, they are mainly used to measure probabilities
of major outbreaks, disease extinction and, in general, to make statistical analysis of some relevant
epidemic descriptors.
In any case, both deterministic and stochastic frameworks are important but stochastic models
seems to be more appropriate to describe the evolution of an infective process evolving in a small
community, rather than their deterministic counterparts (e.g., [2,9]). In deterministic models,
persistence or extinction of the epidemic process is determined by the basic reproduction number,
R0 (see for instance [10,11]). Usually, stochastic models inherited the basic reproduction number from
their deterministic counterparts however, in stochastic epidemic where Markov chains model disease
spread (see [12,13]), two alternative random variables (namely, the exact reproduction number and the

Mathematics 2018, 6, 128; doi:10.3390/math6080128 154 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 128

population transmission number) provide a real measurement of the spread of the disease at the initial
time or at any time during the epidemic process.
Most of the research works, both deterministic and stochastic, deal with continuous-time models.
However, one of the earliest works is the model studied by Reed and Frost in 1928, who formulate
an SIR model using a discrete-time Markov chain (DTMC) [1]. In recent years, literature shows an
increasing interest in using discrete-time models. Emmert and Allen, in [14], consider a discrete-time
model to investigate the spread of a disease within a structured host population. Allen and van den
Driessche [15] applied the next generation approach for calculating the basic reproduction number to
several models related to hantavirus and chytridiomycosis wildlife diseases. In [16], authors introduce
probabilities to formulate death, recovery and incidence rates of epidemic diseases on discrete-time SI
and SIS epidemic models with vital dynamics. Bistability of discrete-time SI and SIS epidemic models
with vertical transmission is the subject matter of [17]. An SIS model with a temporary vaccination
program is studied in [18]. D’Onofrio et al. [19] analyze a discrete-time SIR model where vaccination
coverage depends on the risk of infection. In [20,21], the Reed–Frost model is generalized by Markovian
models of SIR and SEIR types, where transition probabilities depend on the total number of diseased,
the number of daily encounters and the probability of transmission per contact. Under demographic
population dynamics, van den Driessche and Yakubu [22] use the next generation method to compute
R0 and to investigate disease extinction and persistence. An approximation of the deterministic
multiple infectious model by a Branching process is employed in [23] to extract information about
disease extinction. Accuracy of discrete-time approaches for studying continuous time contagion
dynamics is the topic developed in [24], who show potential limitations of this approach depending
on the time-step considered.
In this paper we deal with a stochastic SIS epidemic model, that describes diseases such as
tuberculosis, meningitis or some sexually transmitted diseases, in which infected individuals do not
present an exposed period and are recovered with no immunity. Hence, individuals have reoccurring
infections. The host population is divided into two groups: susceptible (S) or infected (I). We assume
that disease transmission depends on the number of infective individuals and also on a contact rate, α;
in addition individual recoveries depend on the recovery rate γ. For any event (in our model, either
contact or recovery), the event rate or intensity provides the mean number of events per unit time.
Hence, in case of recoveries 1/γ denotes the mean infectious time. To control the epidemic spread,
the population is observed at a fixed time interval.
The aim of this paper is to analyze, for a discrete-time SIS model, the number of inspections that
find an active epidemic. We remark that this quantity is the discrete-time analogous of the extinction
time that describes the length of the epidemic process.
The extinction time has been the subject of interest of many papers. Many of them focus on the
determination of the moments and a few also on whole distribution. In that sense, assuming a finite
birth-death process, Norden [25] first obtained an explicit expression for the mean time to extinction
and established that the extinction time, when the initial distribution equals the quasi-stationary
distribution, follows a simple exponential distribution. Allen and Burgin [26], for SIS and SIR models
in discrete-time, investigated numerically the expected duration. Stone et al. [27], for an SIS model
with external infection, determine also expressions for higher order moments. Artalejo et al. [28],
for general birth-death and SIR models, develop algorithmic schemes to analyze Laplace transforms
and moments of the extinction time and other continuous measures.
We model the evolution of the epidemics in terms of an absorbing DTMC providing the amount
of infective individuals at each stage or inspection point, that introduces in the model individual
variations coming from chance circumstances. As the extinction of the epidemic process is certain, we
will investigate both the distribution and expected values of the number of inspections taking place
prior the epidemic end, conditioned to the initial number of infected.
The rest of the paper is organized as follows. In Section 2, we introduce the discrete-time SIS
model and the DTMC describing the evolution of the epidemic. In Section 3, we present recursive

155
Mathematics 2018, 6, 128

results and develop algorithmic schemes for the distribution of the random variable representing
the number of inspections that find an active epidemic process. For any number of initial infected
individuals, the expected number of inspections will be explicitly determined. Finally, numerical
results regarding the effect of the model parameters are displayed in Section 4. That also includes an
application to evaluate cost and benefit per outbreak in such epidemics.

2. Model Description
Let us consider a closed population of N individuals that is affected by a communicable disease
transmitted by direct contact. In addition, it is assumed that there are no births or deaths during disease
outbreaks, therefore population size remains constant. Individuals in the population are classified
as susceptible, S, or infective, I, according to their health state regarding the disease. Transitions
from susceptible to infective depend on the contact rate between individuals and also on the quantity
of infective present in the population. Once recovered, individuals become again susceptible to the
disease. Consequently, individuals can be infected several times during an epidemic process but the
epidemic stops as soon as there are no infective individuals to transmit the contagious disease.
In a discrete-time study, time is discretized into time steps Δt and transitions from states occur
during this time interval with certain probabilities. In chain-Binomial models [1,3] the time step
corresponds to the length of the incubation period, contact process depend on the Binomial distribution
and during a fixed time interval zero, one or even more infections may happen. In [16] time step is
one and probabilities depend on effective transmission through the time. In [23] a branching process
describes transitions and survival probabilities during any stage.
In our model, the population is observed periodically at time points n ∗ Δt, with n ≥ 0, where
interval’s length Δt is chosen as to guarantee that between consecutive inspections at most one
change—either an infection or a recovery—occurs. The underlying mathematical model is the
discrete-time SIS model described, for instance, in [26] assuming zero births or deaths per individual
in the time interval Δt.
Due to the constant population hypothesis, the evolution of the disease can be represented by
a one-dimensional Markov chain, { In ; n ≥ 0} where In is a random variable giving the number of
infective individuals in the population at the n−th inspection. State space is finite and contains a single
absorbing state, the state zero.
Non-negative transition probabilities depend on the time interval and have the following form,
for 0 ≤ i ≤ N:

P{ In+1 = i − 1| In = i } = pi,i−1 = γiΔt,

α
P{ In+1 = i | In = i } = pi,i = 1 − γiΔt − i ( N − i )Δt, (1)
N
α
P{ In+1 = i + 1| In = i } = pi,i+1 = i ( N − i )Δt,
N
where α represents the contact rate and γ represents the recovery rate.
We need to fix an interval length, Δt, providing that probabilities in (1) are well defined and
therefore that the chosen time step guarantees that at most one change occurs between successive
inspections. In particular, for any choice on the model parameters it is required that,
α
1 − γiΔt − i ( N − i )Δt ≥ 0, for 0 ≤ i ≤ N.
N
Which, after some algebra, provides a bound for time step. Hence, in what follows Δt will be
chosen small enough as to satisfy
 1
γN , if α ≤ γ,
Δt ≤ (2)
4α
( α + γ )2 N
, if α > γ.

156
Mathematics 2018, 6, 128

Notice that the bound given in (2) can be written in terms of the basic reproduction number,
R0 = α/γ, as in [26].
As { In : n ≥ 0} is a reducible aperiodic DTMC, with a single absorbing state, the standard theory
of Markov chains (see for instance [29]) gives that

lim P{ In = j | I0 = i } = δj0 , for every 1 ≤ i, j ≤ N, (3)

n→∞

where δab is the Kronecker’s delta, defined as one for a = b and 0, otherwise.
The limiting behavior result (3) states that in the long term, for any choice on model parameters,
there will be no infective individuals in the population. Hence, the end of any outbreak of the disease
occurs almost surely, but it may take a long number of inspections until the disease disappears from
the population; as it was observed in [27,28] for continuous-time models. Thus, a theoretical study of
this random variable is well supported.

3. Analysis of the Number of Inspections

We consider the random variable T that counts the number of inspections of the population that
find an active epidemic process; i.e., T is the number of steps that it takes, to the DTMC { In : n ≥ 0},
to reach the state zero. Thus, T can be seen as a first-passage time and we define it as

T = min{n ≥ 0 : In = 0}.

In this section we describe its probabilistic behavior in terms of distribution functions and expected
values. Theoretical discussion is based on the conditional first-passage times Ti , for 1 ≤ i ≤ N, defined
as the number of inspections that take place during an outbreak, given that at present population
contains I0 = i infected. Notice that, even for a finite population, Ti = ( T | I0 = i ) is a discrete random
variable with countable infinite mass points.
Next we introduced some notation for point and cumulative probabilities, and expected values
regarding random variables Ti , for 1 ≤ i ≤ N.

αi ( n ) = P{ Ti = n} = P{ T = n | I0 = i }, for n ≥ 0,
ui ( n ) = P{ Ti ≤ n} = P{ T ≤ n | I0 = i }, for n ≥ 0,
mi = E[ Ti ] = E[ T | I0 = i ].

We want to point out some trivial facts. Notice that

α i (0) = ui (0) = 0, for 1 ≤ i ≤ N, (4)

ui ( n ) = 0, whenever 0 ≤ n < i, (5)

(4) is trivially true due to the definition of Ti as a first-passage time. On the other hand, condition (5)
follows from the fact that, starting with i infective individuals and as at each inspection we observe at
most one event, we need at least i inspections in order to observe that all initial infected have been
recovered. Because point probabilities, αi (n), can be determined from cumulative probabilities, ui (n),
with the help of the well-known relationship

αi (n) = ui (n) − ui (n − 1), for n ≥ 1 and 1 ≤ i ≤ N,

we present results in order to deal with the cumulative ones.

Next theorem provides a recursive scheme for determining cumulative probabilities associated to
an initial number of infective individuals.

157
Mathematics 2018, 6, 128

Theorem 1. For any initial number of infective individuals, I0 = i with 1 ≤ i ≤ N, the set of cumulative
probabilities satisfies the following recursive conditions, for n ≥ 1 :

ui ( n ) = pi,0 + (1 − δi1 ) pi,i−1 ui−1 (n − 1) (6)

+ pi,i ui (n − 1) + (1 − δiN ) pi,i+1 ui+1 (n − 1).

Proof. The proof is an application of a first-step analysis by conditioning on the first transition out of
the current state.

Remark 1. Notice that P{ Ti < ∞} = 1, for 1 ≤ i ≤ N, because states {1, 2, ..., N } are a non-decomposable
set of states. Consequently, Ti is a non-defective random variable and limn→∞ ui (n) = 1, for 1 ≤ i ≤ N.

Remark 2. The use of the iterative Equation (6) produces a sequence of increasing probabilities converging to one
but, for computational purposes, a stopping criteria should be provided in order to avoid longer computation runs.

For each number of inspections, n ≥ 0, Equation (6) is solved recursively, with the help of the
boundary conditions (4) and the trivial result (5). Numerical results appearing in the following section
have been obtained with the help of a recursive algorithm, that stops as soon as a certain percentile
value of the distribution is accumulated. For each initial number of infective individuals, I0 = i0 ,
cumulative probabilities are computed up to the q-th percentile, using the following pseudo-code.

Remark 3. For any appropriate time interval Δt satisfying (2), point and cumulative probabilities are
determined numerically from Algorithm 1. Moreover, for n = 1, 2 inspections they present the following
explicit forms:

P{ Ti = 1} = δi1 γΔt, for 1 ≤ i ≤ N,

⎧  
2α( N −2)
⎪
⎨ γΔt 2 − N Δt − γΔt , for i = 1,
P{ Ti ≤ 2} = 2 (γΔt)2 , for i = 2,
⎪
⎩
0, for 3 ≤ i ≤ N.

Explicit values displayed in Remark 3 indicate that distribution of the random variable Ti depends
on rates α and γ not only through its ratio. Consequently, models sharing the same basic reproductive
number, R0 , present different probabilistic characteristics.

Algorithm 1:
The sequence of cumulative probabilities conditioned to the current number of infective
individuals, {ui (n) : n ≥ 0}, for 1 ≤ i ≤ N, are determined as follows:

Step 0: Set q ∈ (0, 1) and i0 ≥ 1.

Step 1: Set n = 0 and ui (n) = 0, for 1 ≤ i ≤ N.
Step 2: Set n = n + 1 and i = 0.
Step 2a: Set i = i + 1 and ui (n) = 0.
Step 2b: If i ≤ n compute ui (n) using Equation (6).
Step 2c: If i = i0 and ui (n) ≥ q then Stop.
Step 2d: If i < N go to step 2a.
Step 3: Go to step 2.
Expected values mi , for 1 ≤ i ≤ N, provide the long-run average value of inspections prior to
the epidemic end, given that the outbreak started with i infected. Typically, expected values can be
computed from mass distribution functions but, in our case, the use of the recursive procedure given in

158
Mathematics 2018, 6, 128

Algorithm 1 produces a lower approximation of the true value. Instead of that, next theorem provides
a closed form expression for mi , given any initial number of infective individuals.

Theorem 2. Expected values mi = E[ Ti ], present the following form:

i i
 N −k
1 1 α
mi =
γΔt ∑ Ak + NγΔt ∑ ( N − k)! Nγ
, for 1 ≤ i ≤ N − 1, (7)
k =1 k =1


1 N −1 1 N −1 α k
mN = ∑
γΔt k=1
Ak + ∑
NγΔt k=0
k!
Nγ
. (8)

j −1
(α/Nγ) j−k
where Ak = 1
k + ∑ jN=−k+
1
1 j ∏ ( N − s), for 1 ≤ k ≤ N − 1.
s=k

Proof. The proof is based on a first-step argument. By conditioning on the state visited by the
underlying Markov chain after first transition, we get the following set of equations that involve
expected values mi , for 1 ≤ i ≤ N initial infective.

mi = E[ T | I0 = i ] = E[ T | I0 = i, I1 = i − 1 ] pi,i−1 (9)
+ E[ T | I0 = i, I1 = i ] pi,i + (1 − δiN ) E[ T | I0 = i, I1 = i + 1 ] pi,i+1 .

But E[ T | I0 = 1, I1 = 0 ] = 1 and E[ T | I0 = i, I1 = j ] = 1 + E[ T | I0 = j ] = 1 + m j , when j = 0.

Substituting in (9) yields

m1 = p1,0 + (1 + m1 ) p1,1 + (1 + m2 ) p1,2 ,

mi = (1 + mi−1 ) pi,i−1 + (1 + mi ) pi,i + (1 − δiN )(1 + mi,i+1 ) pi,i+1 , for 1 < i ≤ N,

that is in accordance with results appearing in [26] (see Section 2.2.5).

Using the normalization condition pi,i−1 + pi,i + (1 − δiN ) pi.i+1 = 1, we get that conditioned
moments satisfy the following tridiagonal system:

(1 − pi,i )mi = 1 + (1 − δi0 ) pi,i−1 mi−1 + (1 − δiN ) pi,i+1 mi+1 , (10)

that can be solved explicitly. Note that Equation (10) can be rewritten as

p1,0 m1 = 1 + p1,2 (m2 − m1 ), (11)

pi,i−1 (mi − mi−1 ) = 1 + pi,i+1 (mi+1 − mi ), for 1 < i < N, (12)
p N,N −1 (m N − m N −1 ) = 1. (13)

Now we use a method of finite difference equations. First, we introduce differences di defined
as follows
di = mi − (1 − δi1 )mi−1 , for 1 ≤ i ≤ N, (14)

and then substitute (14) in Equations (11)–(13). The tridiagonal system is reduced to a bidiagonal one

pi,i−1 di = 1 + pi,i+1 di , for 1 ≤ i < N, (15)

p N,N −1 d N = 1. (16)

Moreover, Equation (16) provides a closed expression for d N :

1 1
dN = = (17)
p N,N −1 NγΔt

159
Mathematics 2018, 6, 128

and remainder differences can be expressed in terms of (17) by noticing that

1 p
di = + i,i+1 di+1 , for 1 ≤ i < N. (18)
pi,i−1 pi,i−1

Using backward substitution and the expressions for transition probabilities (1), we get that

1
di = A + Bi d N , for 1 ≤ i < N, (19)
γΔt i

j −1
(α/Nγ) j−i
where Ai = 1
i + ∑ jN=−i+1 1 j ∏ ( N − s) and Bi = ( N − i )!(α/Nγ) N −i .
s =i
On the other hand, from definition (14) and by noticing that mi = di + (1 − δi1 )mi−1 , we can write

i
mi = ∑ dk , for 1 ≤ i ≤ N. (20)
k =1

Finally, using (19) repeatedly in combination with (20) gives the explicit expressions (7)–(8).

Remark 4. Notice that, from (7) and (8), expected values mi depend on contact and recovery rates not only
through its ratio R0 . Hence, SIS models sharing the basic reproduction number can present different long run
average values for the number of inspections prior to the end of the infective process.

4. Numerical Results
The objective of this section is to reveal the main insights of the mathematical characteristic that is
the subject matter of this paper. In the previous section we have derived theoretical and algorithmic
results regarding the probabilistic behavior of the random variables Ti , for i ≤ i ≤ N. Probability
distribution, conditioned on the initial number of infected, is obtained as the solution of a system
of linear equations. But unfortunately, we have not reached a well-known, or even a closed form,
distribution for Ti and, in addition, the model relies on a group of parameters that varies over a fairly
broad range. Hence, we are going to examine and quantify the effect of changing one or more of the
parameter’s value in the possible outcomes of the number of inspections. In more details, numerical
results come from the application of Theorem 1 and Algorithm 1, when we focus on probabilities of
different outcomes, and from the explicit Equations (7)–(8) when the interest are expected values mi .
Our aim is two-fold: to investigate the influence of the model parameters in the probabilistic
behavior of these random variables and show a possible application in evaluating benefits associated
to the quantity of inspections conducted over an outbreak of a discrete-time SIS epidemic model.

4.1. Influence of Model Parameters

First we assume that we are able to detect the epidemics as soon as first infection appears, that is
I0 = 1. The objective is to characterize the random variable T1 that counts, from the very beginning
of the epidemic, the total number of periodic inspections taking place prior to the epidemics end.
We choose a contact rate α = 2.0 and a recovery rate γ = 1.0.
Figure 1 is a bar chart of P{ T1 = n}, for n ≥ 1. We consider a population of N = 100 individuals
and time interval between inspections is Δt = 0.01 time units. Mass function presents a decreasing
shape, with a single mode for n = 1. From numerical results, we get that 1% of the outbreaks end by
the time of the first inspection, but also 1% of the outbreaks will last more than 30,000 inspections.

160
Mathematics 2018, 6, 128

0.01

0.009

0.008

0.007

0.006

0.005

0.004

0.003

0.002

0.001

0
0 50 100 150

Figure 1. Mass function for N = 50.

Table 1 displays several cumulative probabilities up to 150 inspections and the expected value m1 ,
for a population of N = 5, 25, 50 and 75 individuals. We keep rates α = 2.0, γ = 1.0 and time interval
length as Δt = 0.01. For each population, only 1% of the outbreaks are inspected once before extinction.
For a fixed number of inspections, cumulative probabilities are smaller when population size is larger.
During outbreaks, large populations are inspected more times, in average, than smaller ones. Even for
a small population of 5 individuals, we observe that the 48% of the outbreaks are still in progress at
the 150-th inspection and an average of around 295 inspections take place while the epidemic is active.

Table 1. Cumulative probabilities for different population sizes.

N=5 N = 25 N = 50 N = 75
P{ T1 ≤ 1} 0.01 0.01 0.01 0.01
P{ T1 ≤ 10} 0.0894 0.0882 0.0880 0.0880
P{ T1 ≤ 20} 0.1599 0.1558 0.1553 0.1552
P{ T1 ≤ 30} 0.2172 0.2093 0.2083 0.2080
P{ T1 ≤ 50} 0.3047 0.2879 0.2858 0.2852
P{ T1 ≤ 75} 0.3827 0.3536 0.3501 0.3489
P{ T1 ≤ 100} 0.4400 0.3983 0.3932 0.3916
P{ T1 ≤ 125} 0.4856 0.4302 0.4236 0.4215
P{ T1 ≤ 150} 0.5226 0.4539 0.4458 0.4431
m1 294.66 7094.24 582627.67 58366793.08

Next, we describe the distribution of T1 using a Box-Whiskers plot diagram. The objective is to
compare the patterns of the number of inspections when we vary the contact rate. The box encloses the
middle central part of the distribution, lower and upper edges of the box correspond to the lower and
the upper quartile, respectively, and the line drawn across the box shows the median of the distribution.
Finally, whiskers below start at 1 and whiskers above the boxes reach up to the 99% of the distribution.
Figure 2 shows the distribution of T1 for α = 0.5, 1.0 and 2.0. We consider a recovery rate γ = 2.0,
a population of N = 20 individuals and a time interval of Δt = 0.01 units length between successive
inspections. The distribution of the random variable is skewed to the right and longer right tails are
observed for larger contact rates. Additional numerical results for γ = 5.0 show a similar shape for
box plot diagrams, with 99th quantile under 200 inspections. This fact is according to the intuition,
because large recovery rates give more chance to recoveries than to new infections and, consequently,

161
Mathematics 2018, 6, 128

outbreaks will involve lesser infective individuals and present shorter extinction times in comparison
with Figure 2’s scenario.

800

700

600

500

400

300

200

100

1 2 3

Figure 2. Box plot for T1 .

In Table 2, we display the expected number of inspections for outbreaks starting from a single
infective case, we consider several values for contact and recovery rates. Population contains
20 individuals and time interval between inspections is Δt = 0.01, for every pair of rates. As was stated
on Remark 4, results show that expected values are not a function of the basic reproductive number
R0 = α/γ. The expected number of inspections prior the end of the infective outbreak increases as a
function of contact rate α and decreases as a function of recovery rate. That remark is according to the
intuition too, because the epidemic length enlarges when contacts between individuals occur more
often and/or when individuals need longer times to be recovered.

Table 2. Expected inspections before extinction time.

α = 0.5 α = 1.0 α = 2.0 α = 5.0

γ = 0.5 444.73 6242.14 1.294 × 107 1.917 × 1013
γ = 1.0 134.19 222.36 3121.07 1.536 × 108
γ = 2.0 56.99 67.09 111.18 11, 325.60
γ = 5.0 21.01 22.16 25.02 44.47

Next we focus on outbreaks that are first observed when the epidemic process involves i, not
necessarily one, infected. Our aim is to describe the expected values mi when varying the initial
number of infective individuals. Notice that Equation (20) guarantees that mi ≥ mi−1 , for 1 < i ≤ N.
Thus, the expected number of inspections taking place prior to the epidemics end is a non-decreasing
function regarding the amount of initial infective.
Figure 3 provides a numerical illustration for a population of N = 20 , contact rate α = 1.8,
recovery rate γ = 0.8 and time interval between periodic inspections Δt = 0.01. Pictured graph agrees
with theoretical results, it quantifies the growth of the number of inspections when infective rises and
it shows the importance of an early detection of such a epidemic process. More specifically, outbreaks
detected at the first infection will be active, in mean terms, about 4000 inspections while if the outbreak
is first checked when two infected are present in the population then, the expected inspections will
rise up to 6000 times or up to 8000 inspections when first checking shows five infected.

162
Mathematics 2018, 6, 128

9000

8000

7000

6000

5000

4000

3000

2000

1000

0
0 2 4 6 8 10 12 14 16 18 20

Figure 3. Expected number of inspections versus initial infective.

Additional numerical results, not included here, report that when we choose intervals between
inspections with decreasing length Δt, the mass distribution function of any Ti , for 1 ≤ i ≤ N, provides
an aproximation to the density function of the extinction time of an epidemic process starting with i
infected [28], that is the continuous counterpart of the number of periodic inspections taking place
while the epidemic process is active.
In the following section we present a possible use of the probabilistic behavior of Ti = ( T | I0 = i ).

4.2. Application to Evaluate Outbreak Benefits

Let us assume that every inspection has a travel or approaching cost c1 and, whenever there is
a change in the population regarding the immediate previous inspection, we get a profit in terms
of information that depends on the type of event. Let gR and g I represent recovery and infection
detection’s gain, respectively.
Associated to every outbreak, the random variable Ti provides the total number of inspections
conducted during an outbreak that starts from I0 = i infective individuals. On the other hand, for
outbreaks starting with I0 = i infective individuals, Artalejo et al. introduced in [30] the random
variables NiR and NiI defined as the number of recoveries and infections per outbreak, respectively.
By noticing that the number of recoveries in an outbreak agrees with the total number of infections in
the same outbreak, we get
NiR = i + NiI , for i ≥ 1. (21)

With the help of the above random variables and its relationship (21), we can determine outbreak’s
benefit, for instance, just by defining a benefit function conditioned to the initial number of infective,
as follows
B( T, R| I0 = i ) = ( gR + g I ) NiR − ig I − c1 Ti . (22)

The expected benefit per outbreak will depend on the mean values of NiR and Ti , but also on the
choice of travel cost and information profits.
Figure 4 represents expected benefit when the initial number of infective varies in 1 ≤ i ≤ 11,
for a population of N = 25, with contact rate α = 2.5, recovery rate of γ = 1.0 and a time interval
of Δt = 0.01 units. We fixed a unitary travel cost per the inspection; i.e. c1 = 1, and gain values
per recovery or infection have been chosen as gR = g I = c2 . Several graphs are drawn by fixing
c2 > c1 . We notice first that, for a fixed number of infective and in order to obtain a positive benefit,
recovery gain c2 must satisfy c2 > c1 E[ Ti ]/(2E[ NiR ] − i ). Hence, the trivial restriction c2 > c1 does

163
Mathematics 2018, 6, 128

not guarantee a positive expected benefit per outbreak. In any case, the expected benefit is a linear
increasing function of the gain value c2 .

4000

c2=1.100
c2=1.125
3000
c2=1.150
c2=1.175

2000 c2=1.200

1000

-1000

-2000

-3000

-4000
1 2 3 4 5 6 7 8 9 10 11

Figure 4. Expected Benefit versus initial infective.

As we can see in Figure 4, depending on the choice for c2 , expected benefit functions present
different shape as a function of the initial number of infective individuals. Numerical results show
that, for c2 ≤ 1.125, expected benefit decreases as the initial infected increases. For c2 = 1.15 we obtain
a minimum expected benefit at I0 = 9 infective. When we set c2 = 1.175 the minimum corresponds
to the value I0 = 2 but we obtain expected values close to 0. For c2 ≥ 1.2, expected benefit remains
almost constant for I0 ≥ 5. These facts illustrate that a deep knowledge on E[ NiR ] or E[ Ti ] will help in
decision making process regarding travel cost and gain values.

5. Discussion and Conclusions

Literature in mathematical modelization of epidemics includes both continuous and discrete-time
models. Continuous-time models are more accurate but more difficult to implement than discrete-time
ones. On the other hand, discrete-time models fit better with real information; data related to real-world
epidemic processes are often given by unit time, so it is natural to preserve dynamic features by
modeling a dynamical system from observations at discrete times which are adapted to time step.
This paper focuses on the discrete-time SIS model, where transition probabilities for event
occurring during time-steps are described in terms of an absorbing DTMC. The population is observed
at periodic time points assuming that at most one event takes in a time-step. The discrete-time
stochastic epidemic SIS models are formulated as DTMC which may be considered approximations to
the continuous-time Markov jump processes. The size of the time step must be controlled to assure
that the model gives genuine probability distributions.
Our purpose is to study the extinction time counterpart in discrete-time, that is the random
variable that counts the total number of inspections that find an active epidemic process.
Subject to the initial number of infective individuals, mass probability function of the number
of inspections, Ti , is numerically determined through a recursive scheme; complementing the
probabilistic knowledge of this variable provided by its expected value, that comes directly from
an explicit expression.
A really interesting extension of this work arises when considering equidistant time inspections
relaxing the requirement about the maximum number of events observed during inspections.
This problem, that appears to be analytically intractable, is the aim of the paper [31], where authors

164
Mathematics 2018, 6, 128

tackle this difficulty via the total area between the sample paths of the numbers of infective individuals
in the continuous-time process and its discrete-time counterpart.

Author Contributions: These authors contributed equally to this work.

Funding: This work was supported by the Government of Spain (Department of Science, Technology and
Innovation) and the European Commission through project MTM 2014-58091.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Bailey, N.T.J. The Mathematical Theory of Infectious Diseases; Griffin & Co.: New York, NY, USA, 1975;
ISBN 0852642318, 9780852642313.
2. Andersson, H.; Britton, T. Stochastic Epidemic Models And Their Statistical Analysis; Lectures Notes in Statistics;
Springer: New York, NY, USA, 2000; Volume 151, ISBN 978-1-4612-1158-7.
3. Becker, N.G. Analysis of Infectious Disease Data; Chapman & Hall/CRC: Boca Raton, FL, USA, 2000;
ISBN 0-412-30990-4.
4. Kawachi, K. Deterministic models for rumor transmission. Nonlinear Anal. Real World Appl. 2008, 9,
1989–2028. [CrossRef]
5. Isham, V.; Harden, S.; Nekovee, M. Stochastic epidemics and rumours on finite random networks. Phys. A
Stat. Mech. Appl. 2010, 389, 561–576. [CrossRef]
6. Amador, J.; Artalejo, J.R. Stochastic modeling of computer virus spreading with warning signals.
J. Frankl. Inst. 2013, 350, 1112–1138. [CrossRef]
7. De Martino, G.; Spina, S. Exploiting the time-dynamics of news diffusion on the Internet through a
generalized Susceptible-Infected model. Phys. A Stat. Mech. Appl. 2015, 438, 634–644. [CrossRef]
8. Giorno, V.; Spina, S. Rumor spreading models with random denials. Phys. A Stat. Mech. Appl. 2016, 461,
569–576. [CrossRef]
9. Britton, T. Stochastic epidemic models: A survey. Math. Biosci. 2010, 225, 24–35. [CrossRef] [PubMed]
10. Diekman, O.; Heesterbeek, H.; Britton, T. Mathematical for Understanding Infectious Disease Dynamics; Princeton
University Press: Princeton, NJ, USA, 2012; ISBN 9781400845620.
11. Van den Driessche, P. Reproduction numbers of infectious disease models. Infect. Dis. Model 2017, 2, 288–303.
[CrossRef] [PubMed]
12. Artalejo, J.R.; Lopez-Herrero, M.J. On the exact measure of disease spread in stochastic epidemic models.
Bull. Math. Biol. 2013, 75, 1031–1050. [CrossRef] [PubMed]
13. Lopez-Herrero, M.J. Epidemic transmission on SEIR stochastic models with nonlinear incidence rate.
Math. Methods Appl. Sci. 2016, 40, 2532–2541. [CrossRef]
14. Emmert, K.E.; Allen, L.J.S. Population persistence and extinction in a discrete-time, stage-structured epidemic
model. J. Differ. Equ. Appl. 2004, 10, 1177–1199. DOI 10.1080/10236190410001654151. [CrossRef]
15. Allen, L.J.S.; van den Driessche, P. The basic reproduction number in some discrete-time epidemic models.
J. Differ. Equ. Appl. 2008, 14, 1127–1147. [CrossRef]
16. Li, J.Q.; Lou, J.; Lou, M.Z. Some discrete SI and SIS epidemic models. Appl. Math. Mech. 2008, 29, 113–119.
[CrossRef]
17. Zhang, J.; Jin, Z. Discrete time SI and SIS epidemic models with vertical transmission. J. Biol. Syst. 2009, 17,
201–212. [CrossRef]
18. Farnoosh, R.; Parsamanesh, M. Disease extinction and persistence in a discrete-time SIS epidemic model
with vaccination and varying population size. Filomat 2017, 31, 4735–4747. [CrossRef]
19. d’Onofrio, A.; Manfredi, P.; Salinelli, E. Dynamic behaviour of a discrete-time SIR model with information
dependent vaccine uptake. J. Differ. Equ. Appl. 2016, 22, 485–512. [CrossRef]
20. Tuckell, H.C.; Williams, R.J. Some properties of a simple stochastic epidemic model of SIR type. Math. Biosci.
2007, 208, 76–97. [CrossRef] [PubMed]
21. Ferrante, M.; Ferraris, E.; Rovira, C. On a stochastic epidemic SEIHR model and its diffusion approximation.
TEST 2016, 25, 482–502. [CrossRef]
22. Van den Driessche, P.; Yakubu, A. Disease extinction versus persistence in discrete-time epidemic models.
Bull. Math. Biol. 2018. [CrossRef] [PubMed]

165
Mathematics 2018, 6, 128

23. Allen, L.J.S.; van den Driessche, P. Relations between deterministic and stochastic thresholds for disease
extinction in continuous- and discrete-time infectious disease models. Math. Biosci. 2013, 243, 99–108.
[CrossRef] [PubMed]
24. Fennell, P.G.; Melnik, S.; Gleeson, J.P. Limitations of discrete-time approaches to continuous-time contagion
dynamics. Phys. Rev. E 2016, 94, 052125. [CrossRef] [PubMed]
25. Norden, R.H. On the distribution of the time to extinction in the stochastic logistic population model.
Adv. Appl. Probab. 1982, 14, 687–708. [CrossRef]
26. Allen, L.J.S.; Burgin, A.M. Comparison of deterministic and stochastic SIS and SIR models in discrete time.
Math. Biosci. 2000, 163, 1–33. [CrossRef]
27. Stone, P.; Wilkinson-Herbots, H.; Isham, V. A stochastic model for head lice infections. J. Math. Biol. 2008, 56,
743–763. [CrossRef] [PubMed]
28. Artalejo, J.R.; Economou, A.; Lopez-Herrero, M.J. Stochastic epidemic models revisited: Analysis of some
continuous performance measures. J. Biol. Dyn. 2012, 6, 189–211. [CrossRef] [PubMed]
29. Kulkarni, V.G. Modeling and Analysis of Stochastic Systems; Chapman & Hall: London, UK, 1995;
ISBN 0-412-04991-0.
30. Artalejo, J.R.; Economou, A.; Lopez-Herrero, M.J. On the number of recovered individuals in the SIS and SIR
stochastic epidemic models. Math. Biosci. 2010, 228, 45–55. [CrossRef] [PubMed]
31. Gómez-Corral, A.; López-García, M.; Rodríguez-Bernal, M.T. On time-discretized versions of SIS epidemic
models. 2018, in press.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

166
 mathematics
Article
Large Deviation Results and Applications to the
Generalized Cramér Model †
Rita Giuliano 1 and Claudio Macci 2,∗
1 Dipartimento di Matematica, Università di Pisa, Largo Bruno Pontecorvo 5, I-56127 Pisa, Italy;
[email protected]
2 Dipartimento di Matematica, Università di Roma Tor Vergata, Via della Ricerca Scientifica, I-00133 Rome, Italy
* Correspondence: [email protected]
† The support of INdAM (Fondi GNAMPA) and Università di Pisa (Fondi di Ateneo) is acknowledged. The first
version of this paper was written during the staying of the first author at the University Jean Monnet (St. Etienne).

Received: 2 March 2018; Accepted: 27 March 2018; Published: 2 April 2018

Abstract: In this paper, we prove large deviation results for some sequences of weighted sums of random
variables. These sequences have applications to the probabilistic generalized Cramér model for products of
primes in arithmetic progressions; they could lead to new conjectures concerning the (non-random) set of
products of primes in arithmetic progressions, a relevant topic in number theory.

Keywords: arithmetic progressions; first Chebyshev function; products of primes; regularly varying
functions; slowly varying functions

1. Introduction
The aim of this paper is to prove asymptotic results for a class of sequences of random variables, i.e.,
 
∑nk=1 Lk Xk
:n≥1 (1)
bn

for suitable sequences of real numbers {bn : n ≥ 1} and { Ln : n ≥ 1} (see Condition 1 in Section 3) and
suitable random independent variables { Xn : n ≥ 1} defined on the same probability space (Ω, F , P).
We also present analogue results for the slightly different sequence
 
Ln ∑nk=1 Xk
:n≥1 . (2)
bn

More precisely we refer to the theory of large deviations, which gives an asymptotic computation of
small probabilities on an exponential scale (see, e.g., [1] as a reference on this topic). We recall [2] as a recent
reference on large deviations for models of interest in number theory.
The origin and the motivation of our research rely on the study of some random models similar in nature
to the celebrated Cramér model for prime numbers: i.e., what we have called the generalized model (for products of
prime numbers in arithmetic progressions). We are not aware of any work where these probabilistic models are
studied. Details on these structures will be given in Section 2. Here we only point out that, as the classical
probabilistic model invented by Cramér has been used to formulate conjectures on the (non-random) set
of primes (see [3] for details), in a similar way we can draw out conjectures also for the non-random sets
of products of primes or products of primes in arithmetic progressions. The large deviation results for the
sequences concerning these structures will be given in Corollary 1.
We also remark that the particular form of the sequence (1) is motivated by analogy with the first
Chebyshev function, as will be explained in Section 2.

Mathematics 2018, 6, 49; doi:10.3390/math6040049 www.mdpi.com/journal/mathematics

167
Mathematics 2018, 6, 49

It is worth noting that also some moderate deviation properties can be proved (in terms of suitable
bounds on cumulants and central moments) for the centered sequences
   
∑nk=1 Lk ( Xk − E[ Xk ]) Ln ∑nk=1 ( Xk − E[ Xk ])
:n≥1 and :n≥1 .
bn bn

Such propositions will not be dealt with in the sequel since, though some specific assumptions must be
made in the present setting, these results are in the same direction as those of the paper [4], where moderate
deviations from the point of view of cumulants and central moments are fully investigated.
It should be noted that our results are a contribution to the recent literature on limit theorems of interest
in probability and number theory; here, we recall [5], where the results are formulated in terms of the mod-ϕ
convergence (see also [6] where the simpler mod-Gaussian convergence is studied).
We here introduce some terminology and notation. We always set 0 log 0 = 0, ∞c = 0 for c = 0, and
 x  := max{k ∈ Z : k ≤ x < k + 1} for all x ∈ R. Moreover, we write

• an ∼ bn to mean that limn→∞ an

bn = 1;
law
• Z ∼ B( p), for p ∈ [0, 1], to mean that P( Z = 1) = p = 1 − P( Z = 0);
law λk −λ
• Z ∼ P (λ), for λ > 0, to mean that P( Z = k ) = k! e for all integers k ≥ 0.

The outline of this paper is as follows: We start with some preliminaries in Section 2, and we present
the results in Section 3. The results for the generalized Cramér model (for products of primes in arithmetic
progressions) are presented in Corollary 1.

2. Preliminaries

On large deviations.
We refer to [1] (pages 4–5). Let Z be a topological space equipped with its completed Borel σ-field.
A sequence of Z -valued random variables { Zn : n ≥ 1} satisfies the large deviation principle (LDP) with
speed function vn and rate function I if the following is true: limn→∞ vn = ∞, and the function I : Z → [0, ∞]
is lower semi-continuous.
1
lim sup log P( Zn ∈ F ) ≤ − inf I (z) for all closed sets F
n→∞ vn z∈ F

1
lim inf log P( Zn ∈ G ) ≥ − inf I (z) for all open sets G.
n→∞ vn z∈ G

A rate function I is said to be good if its level sets {{z ∈ Z : I (z) ≤ η } : η ≥ 0} are compact.
Throughout this paper, we prove LDPs with Z = R. We recall the following known result for future use.

Theorem 1 (Gärtner–Ellis Theorem). Let { Zn : n ≥ 1} be a sequence of real valued random variables. Assume that
the function Λ : R → (−∞, ∞] defined by

1 
Λ(θ ) := lim log E evn θZn (for all θ ∈ R) (3)
n→∞ vn

exists; assume, moreover, that Λ is essentially smooth (see e.g., Definition 2.3.5 in [1]) and lower semi-continuous.
Then { Zn : n ≥ 1} satisfies the LDP with speed function vn and good rate function Λ∗ : R → [0, ∞] defined by

Λ∗ (z) := sup{θz − Λ(θ )}.

θ ∈R

Proof. See, e.g., Theorem 2.3.6 in [1].

168
Mathematics 2018, 6, 49

The main application of Theorem 1 in this paper concerns Theorem 2, where we have

θ ∗ x log x − x + 1 if x ≥ 0
Λ(θ ) = e − 1, which yields Λ ( x ) = (4)
∞ if x < 0.

The LDP in Theorem 3 will instead be proved by combining Theorem 4.2.13 in [1] with Theorem 2, i.e.,
by checking the exponential equivalence (see, e.g., Definition 4.2.10 in [1]) of the involved sequences.

On the generalized Cramér model (for products of primes in arithmetic progressions).

The Cramér model for prime numbers consists in a sequence of independent random variables
{ Xn : n ≥ 1} such that, for every n ≥ 2,

law
Xn ∼ B(1/ log n). (5)

This model can be justified by the prime numbers theorem (PNT), which roughly asserts that the
expected density of primes around x is log1 x : the cardinality of prime numbers ≤ n is
 n
1
π (n) := ∑ 1 ∼ li(n) := 2 log t
dt,
p≤n

and, with the words of [7] (see footnote on p. 6), “the quantity log1 n appears here naturally as the derivative
n 1
t dt ∼ log n , another way of stating the PNT is
n
of li( x ) evaluated at x = n”. Since 2 log

π (n) 1
∼ . (6)
n log n

A first extension of this formula concerns the case of integers n which are products of exactly r prime
factors (r ≥ 2). More precisely, we consider the sets

Ar (n) := {k ≤ n : Ω(k) = r } and Br (n) := {k ≤ n : ω (k ) = r }

where ω (n) is the number of distinct prime factors of n, and Ω(n) counts the number of prime factors of n
ω (n) α
(with multiplicity); this means that, letting (by the canonical prime factorization of n) n = ∏i=1 pi i , where
p1 , . . . , pn are the distinct prime factors of n, we have

ω (n)
Ω(n) := ∑ αi .
i =1

A result proved by Landau in 1909 (see, e.g., [8]) states that the cardinalities τr (n) and πr (n) of Ar (n)
and Br (n) respectively verify

n(log log n)r−1 n(log log n)r−1

τr (n) := ∑ 1∼
(r − 1)! log n
and πr (n) := ∑ 1∼
(r − 1)! log n
;
k ∈ Ar ( n ) k∈ Br (n)

see also, e.g., Theorem 437 in [9] (Section 22.18, page 368) or [10] (II.6, Theorems 4 and 5). Note that this
formula for πr (n) reduces to Equation (6) when r = 1.
Going a little further, for fixed integers a and q, we can consider the sets of products of primes in
arithmetic progressions

(q) (q)
Ar (n) =: {k ≤ n : Ω(k ) = r, k ≡ a mod q} and Br (n) =: {k ≤ n : ω (k ) = r, k ≡ a mod q}.

169
Mathematics 2018, 6, 49

One can prove (by similar methods as in [10,11]) that, for any a and q with ( a, q) = 1, the cardinalities
(q) (q) (q) (q)
τr (n) and πr (n) of Ar (n) and Br (n) respectively verify

(q) 1 n(log log n)r−1 (q) 1 n(log log n)r−1

τr (n) := ∑ 1∼ ·
φ(q) (r − 1)! log n
and πr (n) := ∑ 1∼ ·
φ(q) (r − 1)! log n
,
(q) (q)
k ∈ Ar ( n ) k ∈ Br (n)

where φ is Euler’s totient function. Notice that, for r = 1, we recover the sets of primes in arithmetic
progressions, considered for instance in [8,10] II.8, or [11]; the case r = 2 is studied in [12]; the general case
r ≥ 1 is considered in the recent preprint [13]; for q = 1, we recover the sets and the formulas for the model
described above.
Therefore, following Cramér’s heuristic, Equation (5), we can define the generalized Cramér model for
products of r prime numbers (or products of r prime numbers in arithmetic progression) as a sequence of
independent random variables { Xn : n ≥ 1} such that

law n 1 (log log n)r−1

Xn ∼ B(λn ), where λn := and n := · . (7)
log n φ(q) (r − 1) !

Obviously in Equation (7) we take n ≥ n0 , where n0 is an integer, such that λn ∈ (0, 1] for n ≥ n0 ; the
definition of λn for n < n0 is arbitrary.
Large deviation results for this model will be presented in Corollary 1 as a consequence of Theorem 3
and Remark 2, with
Ln := log n and bn := nn ; (8)

thus, the sequences in Equations (1) and (2) become

∑nk=1 (log k) Xk (log n) ∑nk=1 Xk

and (9)
nn nn

respectively. Moreover, by taking into account Remark 3 presented below, the sequences in Equation (9)
converge almost surely to 1 (as n → ∞).

On the first Chebyshev function.

The first Chebyshev function is defined by

θ ( x ) := ∑ log p,
p≤ x

where the sum is extended over all prime numbers p ≤ x.

Therefore, when considering the classical Cramér model, this function is naturally modeled with
∑nk=1 (log k) Xk (and we obtain the numerator of the first fraction in Equation (9)).
It must be noted that T. Tao, in his blog (see [14]), considers the same random variable ∑k≤ x (log k) Xk
and proves that almost surely one has

∑ (log k)Xk = x + Oε (x1/2+ε )

k≤ x

for all ε > 0 (where the implied constant in the Oε (·) notation is allowed to be random). In particular, almost
surely one has
∑ (log k) Xk
lim k≤n = 1.
n→∞ n
It appears clearly that in this setting we have a sequence of the form of Equation (1), with the particular
choices Ln = log n and bn = n. What we are going to investigate in the sequel is how the sequence of

170
Mathematics 2018, 6, 49

random variables { Xn : n ≥ 1} and the two sequences of numbers { Ln : n ≥ 1} and {bn : n ≥ 1} must be
connected in order to obtain large deviations and convergence results (see also Equations (8) and (9) above).

On slowly and regularly varying functions (at infinity).

Here we recall the following basic definitions. A positive measurable function H defined on some
neighborhood of [ x0 , ∞) of infinity is said to be slowly varying at infinity (see, e.g., [15], page 6) if

H (tx )
lim = 1 for all x > 0.
t→∞ H (t)

Similarly, a positive measurable function M defined on some neighborhood of [ x0 , ∞) of infinity is said

to be regularly varying at infinity of index ρ (see, e.g., [15], page 18) if

M (tx )
lim = x ρ for all x > 0.
t→∞ M(t)

Obviously, we recover the slowly varying case if ρ = 0. Recall the following well-known result for
slowly varying functions.

Lemma 1 (Karamata’s representation of slowly varying functions). A function H is slowly varying at infinity if
and only if
 x 
φ(t)
H ( x ) = c( x ) exp dt
x0 t
where φ( x ) → 0 and c( x ) → c∞ for some c∞ > 0 (as x → ∞).

Proof. See, e.g., Theorem 1.3.1 in [15].

In view of what follows we also present the following results. They are more or less known; but we
prefer to give detailed proofs in order to ensure that the paper is self-contained.

Lemma 2. Let M be a regularly varying function (at infinity) of index ρ ≥ 0. Then,

M (tx )
lim = x ρ for all x > 0.
t→∞ M(t)

Proof. It is well-known (see, e.g., Theorem 1.4.1 in [15]) that we have M( x ) = x ρ H ( x ) for a suitable slowly
varying function H. Thus, it is easy to check that it suffices to prove the result for the case ρ = 0 (namely for
a slowly varying function H), i.e.,

H (tx )
lim = 1 for all x > 0. (10)
t→∞ H (t)

By Lemma 1, for all x > 0, we have

 tx 

H (tx ) c(tx ) φ(v)
= exp dv
H (t) c(t) t v

c(tx )
for t > 0. Obviously, c(t)
→ 1 (as t → ∞). Moreover, for all ε > 0, we have
 tx
φ(v)
dv ≤ ε| log(tx /t)|
t v

171
Mathematics 2018, 6, 49

for t > 0, and log(tx /t) → log x (as t → ∞); thus,

 tx
φ(v)
dv → 0 (as t → ∞)
t v

by the arbitrariness of ε > 0. Thus, Equation (10) holds, and the proof is complete.

Lemma 3. Let H be a slowly varying function (at infinity). Then,

xH ( x )
lim x
= 1.
x →∞
∑ k =1 H ( k )

Proof. By the representation of H in Lemma 1, for all ε > 0 there is an integer n0 ≥ 1 such that, for all
x > n0 , we have c∞ − ε < c( x ) < c∞ + ε and −ε < φ( x ) < ε. Then, we take x ≥ n0 + 1, and

x n x
∑ k =1 H ( k ) ∑ 0 H ( k ) ∑ k = n0 +1 H ( k )
= k =1 + .
xH ( x ) xH ( x ) xH ( x )

The first summand in the right hand side can be ignored since, if we take ε ∈ (0, 1), for a sufficient high
x, we have
 x 

c∞ 1 c∞ x −ε
H (x) > exp −ε dt = ,
2 x0 t 2 x0
which yields xH ( x ) > c1 x1−ε for a suitable constant c1 > 0 (and x1−ε → ∞ as x → ∞). Therefore, we
concentrate our attention on the second summand and, by taking into account again the representation of H
in Lemma 1, for a sufficiently high x, we have
    
x x k φ(t) x c(k) x φ(t)
∑ k = n0 +1 H ( k ) ∑k=n0 +1 c(k) exp x0 t dt ∑ k = n0 +1 c( x )
exp − k t dt
=   = .
xH ( x ) x φ(t) x
xc( x ) exp x t dt
0

Moreover,
  
x c(k) x φ(t) x −ε
∑ k = n0 +1 c( x )
exp − k t dt c ∞ + ε ∑ k = n0 +1 k c∞ + ε 1
≤ → (as x → ∞)
x c∞ − ε x 1− ε c∞ − ε 1 − ε

and   
x c(k) x φ(t) x
∑ k = n0 +1 c( x )
exp − k t dt c ∞ − ε ∑ k = n0 +1 k
ε
c∞ − ε 1
≥ → (as x → ∞),
x c∞ + ε x 1+ ε c∞ + ε 1 + ε
and the proof is complete by the arbitrariness of ε.

3. Results
In this section we present large deviation results for Equations (1) and (2). We start with the case of
Poisson distributed random variables (see Theorem 2 and Remark 1), and later we consider the case of
Bernoulli distributed random variables (see Theorem 3 and Remark 2). Our large deviation results yield
the almost sure convergence to 1 (as n → ∞) of the involved random variables (see Remark 3 for details).
In particular, the results for Bernoulli distributed random variables can be applied to the sequences of the
generalized Cramér model in Equation (9) (see Corollary 1).
In all our results, we assume the following condition.

Condition 1. The sequence {bn : n ≥ 1} is eventually positive; { Ln : n ≥ 1} is eventually positive and

non-decreasing.

172
Mathematics 2018, 6, 49

In general, we can ignore the definition of {bn : n ≥ 1} and { Ln : n ≥ 1} for a finite number of indices;
therefore, in order to simplify the proofs, we assume that {bn : n ≥ 1} and { Ln : n ≥ 1} are positive
sequences and that { Ln : n ≥ 1} is non-decreasing.
We start with the case where { Xn : n ≥ 1} are (independent) Poisson distributed random variables.

Theorem 2 (the Poisson case; the sequence in Equation (1)). Let {bn : n ≥ 1} and { Ln : n ≥ 1} be two
sequences as in Condition 1. Assume that

{ Ln : n ≥ 1} is the restriction (on N) of a slowly varying function (at infinity). (11)

bcn
For all c ∈ (0, 1), α(c) := lim exists, and lim α(c) = 0. (12)
n→∞ bn c ↓0

Ln
lim = 0. (13)
n→∞ bn
law
Moreover, assume that { Xn : n ≥ 1} are independent and Xn ∼ P (λn ) for all n ≥ 1, where {λn : n ≥ 1} are
positive numbers such that
n
bn
∑ λk ∼ Ln . (14)
k =1

The sequence in Equation (1) then satisfies the LDP with speed function vn = bn
Ln and good rate function Λ∗
defined by Equation (4).

We point out that Equation (12) is satisfied if the sequence {bn : n ≥ 1} is nondecreasing and is the
restriction (on N) of a regularly varying function with positive index (at infinity); this is a consequence of
Lemma 2.
∑n L X
Proof. We apply Theorem 1, i.e., we check that Equation (3) holds with Zn = k=1bn k k , vn = Lbnn , and Λ as
in Equation (4) (in fact, Equation (3) holds even without assuming (13); however, Equation (13) must be
required in order that vn = Lbnn be a speed function). We remark that

 n   ∑n L X  n 
Ln bn ∑ k =1 L k X k Ln k =1 k k Ln
log E e Ln θ bn = log E eθ Ln = ∑ log E e(θLk /Ln ) Xk
bn bn bn k =1
n n
Ln θLk /Ln
−1) Ln
=
bn ∑ log(eλk (e )=
bn ∑ λk (eθLk /Ln − 1) for all θ ∈ R.
k =1 k =1

Equation (3) trivially holds for θ = 0. The proof is divided in two parts: the proof of the upper bound,
 n 
Ln bn ∑ k =1 L k X k
lim sup log E e Ln θ bn ≤ eθ − 1 for all θ ∈ R, (15)
n→∞ bn

and that of the lower bound,

 n 
Ln bn ∑ k = 1 L k X k
lim inf log E e Ln θ bn ≥ eθ − 1 for all θ ∈ R. (16)
n→∞ bn

We start with the proof of Equation (15). For θ > 0, we have

 n  n n
Ln bn ∑ k =1 L k X k Ln Ln
log E e Ln θ bn = ∑ λk (eθLk /Ln − 1) ≤ ∑ λ k ( e θ − 1)
bn bn k =1
bn k =1

173
Mathematics 2018, 6, 49

since { Ln : n ≥ 1} is nondecreasing, and we obtain Equation (15) by letting n go to infinity and by taking
into account Equation (14). For θ < 0, we take c ∈ (0, 1) and

γ := sup{ Ln : n ≥ 1}

Lcn
(possibly infinite). Recalling that { Ln : n ≥ 1} is nondecreasing and that Ln → 1 (it is a consequence of
Lemma 2), we have

 n  n
Ln bn ∑ k =1 L k X k Ln
log E e Ln θ bn = ∑ λk (eθLk /Ln − 1)
bn bn k =1
cn n
Ln Ln
≤ ∑ λk (eθL1 /γ − 1) + bn ∑ λk (eθLcn /Ln − 1)
bn k =1 k =cn+1
 &
cn
Ln bcn Lcn
= ∑ λk (eθL1 /γ − 1)
Lcn bn bcn k =1
( &  &)
n cn
Ln Ln bcn Lcn
+ ∑ λk − ∑ λk (eθLcn /Ln − 1).
bn k =1
Lcn bn bcn k =1

Then, by Equation (11) (and Lemma 2 with ρ = 0), (12) and (14), we obtain
 n 
Ln bn ∑ k =1 L k X k
lim sup log E e Ln θ bn ≤ α(c)(eθL1 /γ − 1) + (1 − α(c))(eθ − 1).
n→∞ bn

Using Equation (12), we conclude by letting c ↓ 0.

The proof of Equation (16) is similar with reversed inequalities; hence, we only sketch it here. For θ < 0,
we have  n 
Ln bn ∑ k =1 L k X k Ln n Ln n
log E e Ln θ bn ∑ λk (eθLk /Ln − 1) ≥ λ k ( e θ − 1),
bn k∑
=
bn bn k = 1 =1

and we obtain Equation (16) by letting n go to infinity and by taking into account (14). For θ ≥ 0, we take
c ∈ (0, 1) and, for γ defined as above, after some manipulations, we obtain
 n 
Ln bn ∑ k =1 L k X k
lim inf log E e Ln θ bn ≥ α(c)(eθL1 /γ − 1) + (1 − α(c))(eθ − 1).
n→∞ bn

We conclude by letting c ↓ 0 (by Equation (12)).

Remark 1 (The Poisson case; the sequence in Equation (2)). The LDP in Theorem 2 holds also for the sequence
in Equation (2) in place of the sequence in Equation (1). In this case we only need to use Condition 1 and to assume
Equations (13) and (14), whereas Equations (11) and (12) (which were required in the proof of Theorem 2) can be
Ln ∑nk=1 Xk
ignored. For the proof, we still apply Theorem 1, so we have to check that Equation (3) holds with Zn = bn ,
bn
vn = Ln , and Λ as in Equation (4). This can be easily checked noting that

 n   n 
Ln bn L n ∑ k =1 X k Ln Ln
log E e Ln θ bn
n
= log E eθ ∑k=1 Xk = ∑ log E eθXk
bn bn bn k =1
n n
Ln λ k ( e θ −1) Ln
=
bn ∑ log(e )=
bn ∑ λk (eθ − 1) → eθ − 1 for all θ ∈ R
k =1 k =1

where the limit relation holds by Equation (14).

174
Mathematics 2018, 6, 49

The next result is for Bernoulli distributed random variables { Xn : n ≥ 1}. Here we shall use the
concept of exponential equivalence (see, e.g., Definition 4.2.10 in [1]). The proof is similar to the one of
Proposition 3.5 in [16] (see also Remark 3.6 in the same reference). We point out that it is not unusual to
prove a convergence result for Bernoulli random variables { Xn : n ≥ 1} starting from a similar one for
Poisson random variables {Yn : n ≥ 1} and by setting Xn := Yn ∧ 1 for all n ≥ 1; see, for instance, Lemmas 1
and 2 in [17].

Theorem 3 (The Bernoulli case; the sequence in Equation (1)). Let {bn : n ≥ 1} and { Ln : n ≥ 1} be as in
Theorem 2 (thus, Condition 1 together with Equations (11)–(13) hold). Moreover, assume that { Xn : n ≥ 1} are
law
independent and Xn ∼ B(λn ) for all n ≥ 1 and that Equation (14) and limn→∞ λn = 0 hold. The sequence in
Equation (1) satisfies the LDP with speed function vn = Lbnn and the good rate function Λ∗ defined by Equation (4).

Proof. Let n0 such that λn ∈ [0, 1) for all n ≥ n0 (recall that λn → 0 as n → ∞), and let { Xn∗ : n ≥ 1} be
law
independent random variables such that Xn∗ ∼ P (λ̂n ) (for all n ≥ 1), where λ̂n := log 1−1λn for n ≥ n0
(the definition of λ̂n for n < n0 is arbitrary). Notice that
n n
∑ λ̂k ∼ ∑ λk
k =1 k =1

because ∑nk=1 λk → ∞ (as n → ∞) by Equations (13) and (14) and, by the Cesaro theorem,

1
∑nk=1 λ̂k λ̂n log 1−λn
lim = lim = lim = 1.
n→∞ ∑nk=1 λk n→∞ λn n→∞ λn

Hence, the assumption of Equation (14) and Theorem 2 are in force for the sequence { Xn∗ : n ≥ 1} (in
fact, we have Equation (14) with {λ̂n : n ≥ 1} in place of {λn : n ≥ 1}) and, if we set Xn := Xn∗ ∧ 1 (for all
n ≥ 1), the sequence { Xn : n ≥ 1} is indeed an instance of the sequence appearing in the statement of the
law
present theorem since, by construction, Xn ∼ B(1 − e−λ̂n ) and 1 − e−λ̂n = λn .
The statement will be proved by combining Theorem 4.2.13 in [1] and Theorem 2 (for the sequence
{ Xn∗ : n ≥ 1}). This means that we have to check the exponential equivalence condition

Ln
lim sup log P(Δn > δ) = −∞ (for all δ > 0) (17)
n→∞ bn

where
1 n n
1 ∗
Δn :=
bn ∑ L k Xk −
bn ∑ k k .
L X (18)
k =1 k =1

We remark that
n
Ln
Δn ≤
bn ∑ |Xk − Xk∗ | (19)
k =1

by the monotonicity and the nonnegativeness of { Ln : n ≥ 1}; therefore, if we combine Equation (19) and
the Chernoff bound, for each arbitrarily fixed θ ≥ 0, we obtain
( ) 
θ n |X −X∗ |
Ln n E e ∑ k =1 j j
P(Δn > δ) ≤ P ∑ | Xk − Xk∗ | >δ ≤ .
bn k =1 eθδbn /Ln

Therefore,
Ln Ln n  ∗

log P(Δn > δ) ≤ ∑ log E eθ | Xk − Xk | − θδ.
bn bn k =1

175
Mathematics 2018, 6, 49

Moreover, if we set
θ
(θ ) eλk e − 1
ρk := ,
λk eθ
we have
 ∗
∞
E eθ |Xk −Xk | = P( Xk∗ = 0) + P( Xk∗ = 1) + ∑ eθ|1−h| P(Xk∗ = h)
h =2
∞  
θ ( h −1) λ k − λ k
h
−λk −λk θ
=e + λk e + ∑e e = e−λk + λk e−λk + e−θ e−λk eλk e − 1 − λk eθ
h =2
h!
      
−λk −θ −λk θ θ (θ )
=e +e e eλk e − 1 = e −λk 1 + e −θ eλk e − 1 = e−λk 1 + λk ρ k .

Therefore,
Ln Ln n
Ln n  
(θ )
bn
log P(Δn > δ) ≤ −
bn ∑ λk + bn ∑ log 1 + λk ρk − θδ. (20)
k =1 k =1

The proof will be complete if we show that, for all θ > 0,

Ln n  
(θ )
lim sup
n→∞ bn ∑ log 1 + λk ρk ≤ 1. (21)
k =1

In fact, by Equations (14) and (21), we deduce from Equation (20) that

Ln
lim sup log P(Δn > δ) ≤ −θδ,
n→∞ bn

and we obtain Equation (17) by letting θ go to infinity.

(θ )
Thus, we prove Equation (21). We remark that ρn → 1 because λn → 0 (as n → ∞). Hence, for all
(θ )
ε ∈ (0, 1), there exists n0 such that, for all n > n0 , we have ρn < 1 + ε and

Ln n   Ln n0   Ln n  
(θ ) (θ ) (θ )
bn ∑ log 1 + λk ρk =
bn ∑ log 1 + λk ρk +
bn ∑ log 1 + λk ρk
k =1 k =1 k = n0 +1

Ln n0   Ln n
(θ )
≤
bn ∑ log 1 + λk ρk +
bn ∑ log (1 + λk (1 + ε)) .
k =1 k = n0 +1

 
n0 (θ )
Moreover, Ln
bn ∑k= 1 log 1 + λk ρk → 0 (as n → ∞) by Equation (13) and
( )
n n n n0
Ln Ln Ln Ln
bn ∑ log (1 + λk (1 + ε)) ≤ (1 + ε)
bn ∑ λ k = (1 + ε )
bn ∑ λ k − bn ∑ λ k .
k = n0 +1 k = n0 +1 k =1 k =1

Hence, Equation (21) follows from Equations (13) and (14), and the arbitrariness of ε.

Remark 2 (The Bernoulli case; the sequence in Equation (2)). The LDP in Theorem 3 holds also for the sequence
in Equation (2) in place of the sequence in Equation (1). The proof is almost identical to the one of Theorem 3: in this
case, we have
L n Ln n ∗
n
bn k∑ bn k∑
Δn := X k − X k
=1 =1

in place of Equation (18), and Inequality (19) still holds (even without the monotonicity of { Ln : n ≥ 1}).

Remark 3 (Almost sure convergence to 1 of the sequences in Theorems 2 and 3). Let { Zn : n ≥ 1} be either the
sequence in Equation (1) or the sequence in Equation (2), where { Xn : n ≥ 1} is as in Theorem 2 or as in Theorem 3

176
Mathematics 2018, 6, 49

(so we also consider Remarks 1 and 2). Then, by a straightforward consequence of the Borel–Cantelli lemma, the
sequence { Zn : n ≥ 1} converges to 1 almost surely (as n → ∞) if

∑n≥1 P( Zn ∈ C ) < ∞ for closed set C such that 1 ∈

/ C.

Obviously this condition holds if C ⊂ (−∞, 0) because { Zn : n ≥ 1} are nonnegative random variables. On the
other hand, if C ∩ [0, ∞) is not empty, Λ∗ (C ) := infx∈C Λ∗ ( x ) is finite; moreover, Λ∗ (C ) ∈ (0, ∞) because 1 ∈
/ C.
Then, by the upper bound of the closed set, for all δ > 0, there exists nδ such that, for all n > nδ , we have
∗ (C )− δ ) b
P( Zn ∈ C ) ≤ e−(Λ n /Ln .

Thus, again by the Borel–Cantelli lemma, { Zn : n ≥ 1} converges almost surely to 1 (as n → ∞) if, for all κ > 0,
we have

∑ e−κbn /Ln < ∞. (22)

n ≥1

Then, by the Cauchy condensation test, Equation (22) holds if and only if ∑n≥1 2n e−κb2n /L2n < ∞ and, as we see
below, the convergence of the condensed series is a consequence of the ratio test and of some hypotheses of Theorems 2
and 3. In fact,



2n+1 e−κb2n+1 /L2n+1 b2n+1 b2n L2n+1
= 2 exp − κ 1 − · → 0 (as n → ∞)
2n e−κb2n /L2n L 2n +1 b2n+1 L2n

b2n L2n +1 b2n+1

because b2n+1 → α(1/2) by Equation (12), L2n → 1 by Equation (11) and L2n +1 → +∞ by Equation (13).

We conclude with the results for the generalized Cramér model (the sequences in Equation (9)).

Corollary 1 (Application to the sequences in Equation (9)). Let { Xn : n ≥ 1} be the random ! variables in Equation
"
∑nk=1 (log k ) Xk
(7), and let {bn : n ≥ 1} and { Ln : n ≥ 1} be defined by Equation (8). Then, the sequences nn :n≥1
! "
(log n) ∑nk=1 Xk nn
and nn : n ≥ 1 in Equation (9) satisfy the LDP with speed function vn = Lbnn = log n and the good rate
∗
function Λ defined by Equation (4).

Proof. In this proof, the sequences in Equation (9) play the roles of the sequences in Equations (1) and (2) in
Theorem 3 and Remark 2, respectively. Therefore, we have to check that the hypotheses of Theorem 3 are
satisfied. Condition 1 and Equations (11) and (13) and limn→∞ λn = 0 can be easily checked. Moreover, one
can also check Equation (12) with α(c) = c; note that in this case, we have a regularly varying function with
index ρ = 1 (as n → ∞), and {bn : n ≥ 1} is eventually nondecreasing. Finally, Equation (14), which is

k
(log n) ∑nk=1 log k
lim = 1,
n→∞ nn

can be obtained as a consequence of Lemma 3; in fact, {n : n ≥ 1} and {n /(log n) : n ≥ 1} are restrictions
(on N) of slowly varying functions at infinity.

In conclusion, we can say that, roughly speaking, for any Borel set A such that 1 ∈ / Ā (where Ā is the
 n   
∑k=1 (log k) Xk (log n) ∑nk=1 Xk − nn inf Λ∗ ( x )
closure of A), the probabilities P nn and P nn decay exponentially as e log n x∈ A
(as n → ∞). Thus, in the spirit of Tao’s remark, we are able to suggest estimations concerning a sort of
∑ p1 ··· pr ≤ x log( p1 ··· pr ) (log x ) ∑ p1 ··· pr ≤ x 1
“generalized” Chebychev function defined by xx or by xx . To our knowledge,
such estimations are not available for r > 1.

Author Contributions: Rita Giuliano and Claudio Macci equally contributed to the proofs of the results. The paper was
also written and reviewed cooperatively.

177
Mathematics 2018, 6, 49

Conflicts of Interest: The authors declare no conflict of interest.

References
1. Dembo, A.; Zeitouni, O. Large Deviations Techniques and Applications, 2nd ed.; Springer: New York, NY, USA, 1998.
2. Fang, L. Large and moderate deviation principles for alternating Engel expansions. J. Number. Theory 2015,
156, 263–276.
3. Granville, A. Harald Cramér and the distribution of prime numbers. Scand. Actuar. J. 1995, 1995, 12–28.
4. Döring, H.; Eichelsbacher, P. Moderate deviations via cumulants. J. Theor. Probab. 2013, 26, 360–385.
5. Féray, V.; Méliot, P.L.; Nikeghbali, A. Mod-ϕ Convergence, I: Normality Zones and Precise Deviations.
Unpublished Manuscript, 2015. Available online: http://arxiv.org/pdf/1304.2934.pdf (accessed on
23 November 2015).
6. Jacod, J.; Kowalski, E.; Nikeghbali, A. Mod-Gaussian convergence: new limit theorems in probability and number
theory. Forum Math. 2011, 23, 835–873.
7. Tenenbaum, G.; Mendès France, M. The Prime Numbers and Their Distribution; (Translated from the 1997 French
original by P.G. Spain); American Mathematical Society: Providence, RI, USA, 2000.
8. Landau, E. Handbuch der Lehre von der Verteilung der Primzahlen (2 Volumes), 3rd ed.; Chelsea Publishing: New York,
NY, USA, 1974.
9. Hardy, G.H.; Wright, E.M. An Introduction to the Theory of Numbers, 4th ed.; Oxford University Press: London,
UK, 1975.
10. Tenenbaum, G. Introduction to Analytic and Probabilistic Number Theory, 3rd ed.; (Translated from the 2008 French
Edition by P.D.F. Ion); American Mathematical Society: Providence, RI, USA, 2015.
11. Davenport, H. Multiplicative Number Theory, 3rd ed.; Springer: New York, NY, USA; Berlin, Germany, 2000.
12. Ford, K.; Sneed, J. Chebyshev’s bias for products of two primes. Exp. Math. 2010, 19, 385–398.
13. Meng, X. Chebyshev’s Bias for Products of k Primes. Unpublished Manuscript, 2016. Available online:
http://arxiv.org/pdf/1606.04877v2.pdf (accessed on 16 August 2016).
14. Tao, T. Probabilistic Models and Heuristics for the Primes (Optional). In Terence Tao Blog. 2015.
Available online: https://terrytao.wordpress.com/2015/01/04/254a-supplement-4-probabilistic-models-and-
heuristics-for-the-primes-optional/ (accessed on 4 January 2015).
15. Bingham, N.H.; Goldie, C.M.; Teugels, J.L. Regular variation. In Encyclopedia of Mathematics and its Applications;
Cambridge University Press: Cambridge, UK, 1987; Volume 27.
16. Giuliano, R.; Macci, C. Asymptotic results for weighted means of random variables which converge to a Dickman
distribution, and some number theoretical applications. ESAIM Probab. Stat. 2015, 19, 395–413.
17. Arratia, R.; Tavaré, S. Independent processes approximations for random combinatorial structures. Adv. Math.
1994, 104, 90–154.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article
distributed under the terms and conditions of the Creative Commons Attribution (CC BY)
license (http://creativecommons.org/licenses/by/4.0/).

178
 mathematics
Article
A Within-Host Stochastic Model for
Nematode Infection
Antonio Gómez-Corral 1, * and Martín López-García 2
1 Instituto de Ciencias Matemáticas CSIC-UAM-UC3M-UCM, Calle Nicolás Cabrera 13-15,
28049 Madrid, Spain
2 Department of Applied Mathematics, School of Mathematics, University of Leeds, Leeds LS2 9JT, UK;
[email protected]
* Correspondence: [email protected]

Received: 25 June 2018; Accepted: 11 August 2018; Published: 21 August 2018

Abstract: We propose a stochastic model for the development of gastrointestinal nematode infection in
growing lambs under the assumption that nonhomogeneous Poisson processes govern the acquisition
of parasites, the parasite-induced host mortality, the natural (no parasite-induced) host mortality
and the death of parasites within the host. By means of considering a number of age-dependent
birth and death processes with killing, we analyse the impact of grazing strategies that are defined
in terms of an intervention instant t0 , which might imply a move of the host to safe pasture and/or
anthelmintic treatment. The efficacy and cost of each grazing strategy are defined in terms of the
transient probabilities of the underlying stochastic processes, which are computed by means of
Strang–Marchuk splitting techniques. Our model, calibrated with empirical data from Uriarte et al
and Nasreen et al., regarding the seasonal presence of nematodes on pasture in temperate zones
and anthelmintic efficacy, supports the use of dose-and-move strategies in temperate zones during
summer and provides stochastic criteria for selecting the exact optimum time instant t0 when these
strategies should be applied.

Keywords: host-parasite interaction; nematode infection; nonhomogeneous Poisson process; seasonal

environment; Strang–Marchuk splitting approach

1. Introduction
Gastrointestinal (GI) nematodes are arguably (see [1,2]) the major cause of ill health and poor
productivity in grazing sheep worldwide, especially in young stock. Productivity losses result from
both parasite challenge and parasitism, while regular treatment of the infections is costly in terms
of chemicals and labour. The relative cost of GI parasitism has become greater in recent decades
as the availability of effective broad-spectrum anthelmintics (see Chapter 5 of [1]) has enabled the
intensification of pastoral agriculture. To an extent, it appears the success of the various anthelmintic
products developed since the 1960s has created a rod for our own backs, particularly as resistance has
arisen to each active family in turn (see, for example, [3,4]). Options for the control of GI nematode
infections (which do not rely uniquely on the use of anthelmintics) include management procedures
(involving intervention with anthelmintics, grazing management, level of nutrition and bioactive
forages), biological control (with nematophagous fungi), selection for genetic resistance in sheep
(within breed/use of selected breeds) and vaccination. The article by Stear et al. [5] gives an overview
of alternatives to anthelmintics for the control of nematodes in livestock, and it complements and
extends other review articles by Hein et al. [6], Knox [7], Sayers and Sweeney [8] and Waller and
Thamsborg [9]. Moreover, we refer the reader to the article by Smith et al. [10] for stochastic and
deterministic models of anthelmintic resistance emergence and spread.

Mathematics 2018, 6, 143; doi:10.3390/math6090143 179 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 143

The aim of this paper is to present a stochastic model for quantitatively comparing among various
grazing strategies involving isolation, movement or treatment of the host, but without incorporating the
risk of selecting for resistance. This amounts to the assumption that the nematodes in our model have
not been previously exposed to the anthelmintic treatments under consideration; see, for example, [11].
We point out that the effect of the resistance in the dynamics is usually limited by the rotation of
different anthelmintic classes on an annual basis (see [12,13]).
A wide range of mathematical models can be found in the literature for modelling the infection
dynamics of nematodes in ruminants. Originally, simple deterministic models were proposed in terms
of systems of ordinary differential equations describing the population dynamics of infected ruminants
and nematodes on pasture. By describing these dynamics in a deterministic way, the resulting
models were tractable from a mathematical and analytical point of view [14]. However, efforts
were soon redirected towards stochastic approaches given the importance of stochastic effects in
these systems [15]. These stochastic effects are related to, among others, spatial dynamics, clumped
infection events or individual heterogeneities related to the host’s immune response to infection [16].
Without any aim of an exhaustive enumeration, we refer the reader to [15,17,18] for deterministic
and stochastic models of nematode infection in ruminants for a population of hosts maintaining a
fixed density.
In this paper, we develop a mathematical model for the within-host GI nematode infection
dynamics, to compare the effectiveness and cost of various worm control strategies, which are related
to pasture management practices and/or strategic treatments based on the use of a single anthelmintic
drug. Control criteria are applied to the development of GI nematode parasitism in growing lambs.
Specifically, the interest is in the parasite Nematodirus spp. with Nematodirus battus, Nematodirus
filicollis and Nematodirus spathiger as the main species. The resulting grazing management strategies
are specified in terms of an intervention instant t0 that, under certain specifications, implies moving
animals to safe pastures and/or anthelmintic treatment. For a suitable selection of t0 , we present two
control criteria that provide a suitable balance between the efficacy and cost of intervention. Our
methodology is based on simple stochastic principles and time-dependent continuous-time Markov
chains; see the book by Allen [19] for a review of the main results for deterministic and stochastic
models of interacting biological populations.
Our work in this paper is directly related to that in [20], where we examine stochastic models
for the parasite load of a single host and where the interest is in analysing the efficacy of various
grazing management strategies. In [20], we defined control strategies based on isolation and treatment
of the host at a certain age t0 . This means that the host is free living in a seasonal environment,
and it is transferred to an uninfected area at age t0 . In the uninfected area, the host does not acquire
new parasites, undergoes an anthelmintic treatment to decrease the parasite load and varies in its
susceptibility to parasite-induced mortality and natural (no parasite-induced) mortality. From a
mathematical point of view, an important feature of the analysis in [20] is that the underlying processes,
recording the number of parasites infesting the host at an arbitrary time t, can be thought of as
age-dependent versions of a pure birth process with killing and a pure death process with killing,
which are both defined on a finite state space.
Here, we complement the treatment of control strategies applied to GI nematode burden that
we started in [20] by focusing on strategies that are not based on isolation of the host; to be concrete,
our interest is in three grazing strategies that reflect the use of a paddock with safe pasture and/or
the efficacy of an anthelmintic drug. Seasonal fluctuations in the acquisition of parasites, the death of
parasites within the host and the natural and parasite-induced host mortality are incorporated into our
model by using nonhomogeneous Poisson processes. Contrary to [20], grazing management strategies
considered in this work lead to, instead of pure birth/death processes with killing, the analysis
of several age-dependent birth and death processes with killing. The efficacy and cost of each
grazing strategy are defined in terms of the transient probabilities of each of the underlying stochastic
processes; that is, the probability that the parasite load of the infected host is at any given level at each

180
Mathematics 2018, 6, 143

time instant, given that a particular control strategy has been applied at the intervention instant t0 .
In order to compute these probabilities, we apply Strang–Marchuk splitting techniques for solving the
corresponding systems of differential equations.
The paper is organized as follows. In Section 2, we define the mathematical model used in various
grazing management strategies, derive the analytical solution of the underlying time-dependent
systems of linear differential equations and present two criteria allowing us to find the instant t0 that
appropriately balances effectiveness and cost of intervention in these grazing strategies. In Section 3,
we examine seasonal changes of GI nematode burden in growing lambs. Finally, concluding remarks
are given in Section 4.

2. Stochastic Within-Host Model and Control Criteria

In this section, we first propose a mathematical stochastic model for the within-host infection
dynamics by GI nematodes in growing lambs, define grazing management strategies and set down a
set of equations governing the dynamics of the underlying processes. We then present control criteria
based on stochastic principles. For the sake of brevity, we refer the reader to Appendix A where we
comment on the equivalence used in Table 1 of [20] in the identification of the degree of infestation,
level of infection, eggs per gram (EPG) value, number of L3 infective larvae in the small intestine
and the points system. Further details on the taxonomy and morphology of the parasite Nematodirus
spp. and the treatment and control of parasite gastroenteritis in sheep can be found in [1,2,21].

2.1. Grazing Management Strategies: A Stochastic Within-Host Model

We define the mathematical model in terms of levels of infection and let the random variable
M (t) record the infection level of the host at time t. From Table 1 of [20], this means that the degree
of infestation is null if M (t) = 0, light if M(t) = m with m ∈ {1, 2, 3}, moderate if M(t) = m with
m ∈ {4, 5, 6, 7}, high if M(t) = m with m ∈ {8, 9, 10, 11} and heavy if M(t) = −1. In the setting of GI
nematode parasitism, the value M0 = 11 amounts to a critical level that does not permit the host to
develop immunity to the nematode infection, in such a way that an eventual intervention is assumed
to be ineffective as the degree of infestation is heavy. Therefore, we let M (t) = −1 be equivalent to
the degree heavy of infestation (i.e., the number of L3 infective larvae in the small intestine is greater
than 12,000 worms) or the death of the host. Let S denote the set {0, 1, . . . , M0 } of infection levels,
with M0 = 11.
We consider individual-based grazing strategies, which are related to a single lamb (host) that
is born, parasite-free, at time t = 0 and, over its lifetime, is exposed to parasites at times that form a
nonhomogeneous Poisson process of rate λ(t). At every exposure instant, the host acquires parasites,
independently of one exposure to another. It is assumed that the number of acquired parasites does
not allow the level M(t) of infection to increase more than one unit at any acquisition instant, which
is a plausible assumption in our examples where increments in the number of L3 infective larvae in
the small intestine are registered at daily intervals. Let δ(t) be the death rate of the host at age t in
the absence of any parasite burden, and assume that this rate is increased by an amount γm (t), which
is related to the parasite-induced host mortality as the infection level equals m at age t. For later
use, we define the functions λm (t) = (1 − 1−1,m )λ(t) and δm (t) = (1 − 1−1,m )(δ(t) + γm (t)) for levels
m ∈ {−1} ∪ S , where 1k,m denotes Kronecker’s delta.
At age τ, the interest is in the level M (τ ) of infection as, under certain grazing assumptions,
intervention is prescribed at a certain age t0 < τ. Note that the host at age t0 can be dead or its degree
of infestation can be heavy (M(t0 ) = −1), and it can be alive and parasite-free (M (t0 ) = 0), or it can be
alive and infected (M(t0 ) = m with m ∈ {1, 2, . . . , M0 }).
In analysing the process Z = { M(t) : 0 ≤ t ≤ τ }, we distinguish between the free-living interval
[0, t0 ) and the post-intervention interval [t0 , τ ]; for ease of presentation, we first digress to briefly
recall the analytic solution for ages t ∈ [0, t0 ) given in [20]. For a host that has survived at age t
with t < t0 , the infection dynamics within the host are analysed in terms of transient probabilities

181
Mathematics 2018, 6, 143

πm (t) = P ( M(t) = m| M(0) = 0), for levels m ∈ {−1} ∪ S . That is, πm (t) represents the probability
of the host being at infection level m at time t. These dynamics lead us to a pure birth process with
killing on the state space {−1} ∪ S (see Figure 1 in [20]), the age-dependent birth and killing rates of
which are given by λm (t) = λ(t) and δm (t) = δ(t) + γm (t), respectively, for m ∈ S , and where −1 is
an absorbing state. Expressions for πm (t) can be then evaluated following our arguments in Section 2.2
of [20].
Next, we focus on three grazing strategies that are defined in terms of the intervention instant t0 .
This implies that, at post-intervention ages t ∈ (t0 , τ ], the rates λ(t), δ(t) and γm (t) are replaced by
functions λ (t), δ (t) and γm
 ( t ), respectively, allowing us to show concrete effects of an intervention

on the lamb and its environmental conditions. To be concrete, the functions λ (t), δ (t) and γm  (t)

appropriately reflect the use of a paddock with safe pasture and/or the efficacy of an anthelmintic
treatment, in accordance with the following grazing strategies:

Strategy UM: The host is left untreated, but moved to a paddock with safe pasture at age t0 .
The resulting process Z can be thought of as an age-dependent pure birth process with
killing, the birth rates of which are given by λm (t) = λ(t) if t ∈ [0, t0 ), and λ (t)
if t ∈ [t0 , τ ], and killing rates are defined by δm (t) = δ(t) + γm (t) if t ∈ [0, t0 )
and δ (t) + γm  ( t ) if t ∈ [ t , τ ], for m ∈ S .
0
Strategy TS: The host is treated with anthelmintics and set-stocked at age t0 . Let ηm  ( t ) be the death

rate of parasites when the infection level of the host is m ∈ S at time t with t > t0 .
In this case, Z can be seen as an age-dependent birth and death process with killing.
The birth and death rates are defined by λm (t) = λ(t) if t ∈ [0, τ ], ηm (t) = 0 if t ∈ [0, t0 )
and ηm  ( t ) if t ∈ [ t , τ ], for m ∈ S , respectively. Killing rates are defined identically to
0
the rates δm (t) in strategy UM.
Strategy TM: The host is treated with anthelmintics and moved to safe pasture at age t0 . In a similar
manner to strategy TS, the process Z may be formulated as an age-dependent birth
and death process with killing. Birth, death and killing rates are identical to those in
strategy TS with the exception of λm (t) for time instants t ∈ [t0 , τ ], which has the form
λ m ( t ) = λ  ( t ).

In strategies TS and TM, a single anthelmintic drug is used. In accordance with the empirical
data in [22], resistance is not incorporated into modelling aspects since τ = 1 year is assumed in
Section 3. The resulting models are thus related to the rotation of different anthelmintic classes on an
annual basis, which has been widely promoted and adopted as a strategy to delay the development of
anthelmintic resistance in nematode parasites; see, e.g., [12,13].
For the sake of completeness, we introduce the term scenario US to reflect no intervention, that is
the host is left untreated and set-stocked. Note that, in such a case, the process Z is an age-dependent
pure birth process with killing, and its birth and killing rates are specified by λm (t) = λ(t) and
δm (t) = δ(t) + γm (t) if t ∈ [0, τ ], for m ∈ S . It follows then that the transient distribution of Z is
readily derived from [20] for time instants t ∈ (0, τ ].
A slight modification of our arguments in solving Equations (1) and (2) of [20] allows us to derive
explicit expressions for the transient solution at post-intervention instants t ∈ (t0 , τ ] in grazing strategy
UM. For time instants t ∈ [t0 , τ ], we introduce the probability πm U M ( t ; t ) = P ( M ( t ) = m ) of the
0
process being at infection level m at time t, given that strategy UM is implemented at the intervention
instant t0 , and initial conditions πm UM ( t ; t ) = π ( t ), for m ∈ {−1} ∪ S . Then, the transient solution
0 0 m 0
at time instants t ∈ (t0 , τ ] can be readily expressed as:
( )
m −1
−Λ(t0 ;t)−Δm (t0 ;t)
∑
UM,j
UM
πm ( t0 ; t ) =e UM
πm (t0 ; t0 ) + (1 − 10,m ) π UM
j ( t0 ; t0 ) K m −1 ( t0 ; t ) , (1)
j =0

t t
where Λ(t0 ; t) = t0 λ (u)du and Δm (t0 ; t) = t0 ( δ
 ( u ) + γ ( u )) du.
m Starting from:

182
Mathematics 2018, 6, 143

 t
UM,m−1
Km −1 ( t0 ; t ) = λ (u)eΔ̃m−1 (t0 ;u) du,
t0

UM,j
the functions Km−1 (t0 ; t), for values 0 ≤ j ≤ m − 2, are iteratively computed as:
 t
λ (u)eΔ̃m−1 (t0 ;u) Km−2 (t0 ; u)du,
UM,j UM,j
K m −1 ( t0 ; t ) =
t0

with Δ̃m−1 (t0 ; t) = Δm (t0 ; t) − Δm−1 (t0 ; t).

2.2. Splitting Techniques

For grazing strategies TS and TM, the transient solution at time instants t ∈ (t0 , τ ] can be
numerically derived by using splitting techniques; see [23]. In a unifying manner, we may observe
that, for a host that has survived at age t with t0 < t < τ and M(t) = m ∈ S , the possible transitions
(in both strategies TS and TM) are as follows (Figure 1):

(i) m → m + 1 at rate λm (t), for levels m ∈ {0, 1, . . . , M0 − 1};

(ii) m → m − 1 at rate ηm (t), for levels m ∈ {1, 2, . . . , M0 };
(iii) m → −1 at rate δm (t), for levels m ∈ {0, 1, . . . , M0 − 1};
(iv) M0 → −1 at rate δM0 (t) + λ M0 (t).

λ0 (t) λ1 (t) λm−1 (t) λm (t) λm+1 (t) λM0 −1 (t)

- -
R
 R
R
R
 R
R

−1 0 1 ··· m−1 m m+1 ··· M0 − 1 M0

     
6 I I  I I I  I
η1 (t) η2 (t) ηm (t) ηm+1 (t) ηm+2 (t) ηM0 (t)


δ0 (t) 

δ1 (t) 

δm−1 (t)

δm (t)

δm+1 (t)


δM0 −1 (t) ?δM0 (t) + λM0 (t)


Figure 1. State space and transitions at post-intervention instants t ∈ [t0 , τ ]. Grazing strategies TS
and TM.

Then, if we select a certain grazing strategy s with s ∈ { TS, TM}, the resulting probabilities
s ( t ; t ) = P ( M ( t ) = m ), for m ∈ {−1} ∪ S and time instants t ∈ [ t , τ ], satisfy the equality:
πm 0 0

M0
s
π− 1 ( t0 ; t ) = 1− ∑ s
πm ( t0 ; t ),
m =0

and the time-dependent linear system of differential equations:

d s
Π ( t0 ; t ) = B ( t ) Π s ( t0 ; t ), (2)
dt

where Πs (t0 ; t) = (π0s (t0 ; t), π1s (t0 ; t), . . . , π sM0 (t0 ; t)) T , and B(t) is a tridiagonal matrix with entries:
⎧
⎪
⎪ −(λi (t) + δi (t) + (1 − 10,i )ηi (t)), if 0 ≤ i ≤ M0 , j = i,
⎪
⎨ η ( t ),
i +1 if 0 ≤ i ≤ M0 − 1, j = i + 1,
(B(t))i,j =
⎪
⎪ λ i −1 ( t ), if 1 ≤ i ≤ M0 , j = i − 1,
⎪
⎩
0, otherwise.

183
Mathematics 2018, 6, 143

Needless to say, initial conditions in Equation (2) are given by Πs (t0 ; t0 ) =

(π0 (t0 ), π1 (t0 ), . . . , π M0 (t0 )) T where the values for πm (t0 ) with m ∈ S do not depend on the
grazing strategy under consideration.
In principle, the system (2) of differential equations could be solved in many ways,
but Strang–Marchuk splitting techniques are concretely used in Section 3 to derive its solution.
Following the approach in [23], the original problem given by Equation (2) is first split into several
subsystems that are then solved cyclically one after the other. This procedure is particularly advisable
when tailor-made numerical methods can be applied for each split subsystem or when, as occurs in
our case, explicit solutions for the subsystems can be derived.
The approach in Section 1.3 of [23] is of particular interest when, for a certain splitting B(t) =
U(t) + V(t), the time-dependent linear systems of differential equations:

d s
Π ( t0 ; t ) = U ( t ) Π s ( t0 ; t ), t0 ≤ t ≤ τ,
dt
d s
Π ( t0 ; t ) = V ( t ) Π s ( t0 ; t ), t0 ≤ t ≤ τ,
dt
can be accurately and efficiently solved, which is our case here. In our examples in Section 3,
we consider the splitting B(t) = U(t) + V(t) with:
⎛ ⎞
−(λ0 (t) + δ0 (t))
⎜ ⎟
⎜ λ0 ( t ) −(λ1 (t) + δ1 (t)) ⎟
U(t) = ⎜
⎜ .. ..
⎟,
⎟
⎝ . . ⎠
λ M0 −1 (t) −(λ M0 (t) + δM0 (t))
⎛ ⎞
0 η1 ( t )
⎜ − η1 ( t ) η2 ( t ) ⎟
⎜ ⎟
⎜ ⎟
V(t) = ⎜ .. .. ⎟,
⎜ . . ⎟
⎜ ⎟
⎝ −η M0 −1 (t) η M0 (t) ⎠
−η M0 (t)

and we evaluate numerically the transient solution πm s ( t ; t ) at time instants t ∈ { t , t + 1, . . . , τ } by

0 0 0
solving a sequence of four time-dependent linear subsystems of differential equations.
s ( t ; t + 1) for levels m ∈ S and a certain grazing
In order to determine the probabilities πm 0 0
strategy s with s ∈ { TS, TM}, we first select the splitting time-step as Δt = N −1 with N = 103 ,
and introduce the notation:

an = t0 + (n − 1)Δt, n ∈ {1, 2, . . . , N + 1}, (3)

bn = t0 + (n − 0.5)Δt, n ∈ {1, 2, . . . , N }. (4)

At step n with n ∈ {1, 2, . . . , N }, we solve the subsystems (S1 )n , (S2 )n , (S3 )n and (S4 )n cyclically
on successive intervals of length Δt, using the solution of one subsystem as the initial condition of the
other one as follows:
⎧ d s
dt Π1 ( an ; t ) = U ( t ) Π1 ( an ; t ), an ≤ t ≤ bn ,
s
⎪
⎪
⎪
⎨
Subsystem (S1 )n ≡ 
⎪
⎪ Π s ( t0 ; t0 ), if n = 1,
⎪
⎩ Π1 ( a n ; a n ) =
s
Π4s (bn−1 ; an ), if 2 ≤ n ≤ N.

184
Mathematics 2018, 6, 143

⎧
⎪
⎨
d s
dt Π2 ( an ; t ) = V(t)Π2s ( an ; t), an ≤ t ≤ bn ,
Subsystem (S2 )n ≡
⎪
⎩ Π s ( a ; a ) = Π s ( a ; b ).
2 n n 1 n n
⎧
⎪
⎨ dt Π3 (bn ; t) = V(t)Π3 (bn ; t), bn ≤ t ≤ an+1 ,
d s s

Subsystem (S3 )n ≡
⎪
⎩ Π s ( b ; b ) = Π s ( a ; b ).
3 n n 2 n n
⎧
⎪
⎨ dt Π4 (bn ; t) = U(t)Π4 (bn ; t), bn ≤ t ≤ an+1 ,
d s s

Subsystem (S4 )n ≡
⎪
⎩ Πs (b ; b ) = Πs (b ; a
4 n n 3 n n +1 ).

This procedure results in the solution at t = t0 + 1, which is given by Πs (t0 ; t) = Π4s (b N ; a N +1 )

since a N +1 = t0 + 1. Then, we may proceed similarly in the numerical evaluation of the transient
solution at subsequent time instants t = t0 + k with k ≥ 2 and t0 + k ≤ τ, by replacing t0 by t0 + k in
(3) and (4), so that the solution of the previous subsystems at time instant t = t0 + k − 1 is now used as
the initial condition in the subsystem (S1 )n at step n = 1. We refer the reader to [23] for qualitative
properties of the operator splitting approach and convergence order.
For grazing strategy s ∈ { TS, TM}, the entries πm s ( a ; t ), for levels m ∈ S , of the vector Πs ( a ; t )
n 1 n
are given by Equation (1) for time instants t ∈ [ an , bn ], with t0 replaced by an , and the function λ (t)
replaced by λ(t) in the case TS.
The solution Π2s ( an ; t) at time instants t ∈ [ an , bn ] has entries:
( )
M0
− Hm ( an ;t) s,M − j
s
πm ( an ; t) =e s
πm ( an ; an ) + (1 − 1m,M0 ) ∑ π sj ( an ; an )Km+10 ( an ; t) , (5)
j = m +1

t  ( u ) du and, starting from:

where Hm ( an ; t) = (1 − 10,m ) an ηm
 t
s,M −(m+1)  H̃m+1 ( an ;u)
Km+10 ( an ; t) = ηm +1 ( u ) e du,
an

s,M − j
the functions Km+10 ( an ; t), for values m + 2 ≤ j ≤ M0 , can be iteratively evaluated as:
 t
s,M − j  H̃m+1 ( an ;u) s,M − j
Km+10 ( an ; t) = ηm +1 ( u ) e Km+20 ( an ; u)du,
an

with H̃m ( an ; t) = Hm−1 ( an ; t) − Hm ( an ; t).

In a similar manner, the solution Π3s (bn ; t) at time instants t ∈ [bn , an+1 ] has the form (5), with an
replaced by bn . The entries πm s ( b ; t ), for levels m ∈ S , of the solution Πs ( b ; t ) are given by Equation (1)
n 4 n
for time instants t ∈ [bn , an+1 ], with t0 replaced by bn and λ (t) replaced by λ(t) in the case TS.

2.3. Control Criteria Based on Stochastic Principles

For grazing strategies UM, TS and TM, we define a control strategy by means of an age t0
and an intervention rule, which is related to a concrete infection level m ∈ {1, 2, . . . , M0 } and the
resulting probability:

M0
P≥m (t) = ∑  π m ( t ).
m=m

This age-dependent probability allows us to determine a set I≥m of potential intervention instants
t ∈ (0, τ ) satisfying the inequality P≥m (t) ≥ p for a predetermined index p ∈ (0, 1); note that
I≥m = (0, τ ) in the case p = 0 regardless of the threshold m .

185
Mathematics 2018, 6, 143

It should be pointed out that, in grazing strategies UM and TM, maintaining safe-pasture
conditions in a paddock for the whole year does not seem feasible in practice. Moreover, treating the
host with anthelmintic drugs (cases TS and TM) within early days of the year will not yield optimal
results, since profits of treatment cannot be obtained before host exposure to infection; see Figure A1
in Appendix A. Thus, we focus on values p > 0 in such a way that, for a fixed pair (m , p) with
p > 0, those instants t ∈ / I≥m can be seen as either low-risk (states m ∈ {0, 1, . . . , m − 1}) or
extreme-risk (m = −1) intervention instants, and consequently, they are not taken into account in our
next arguments.
In carrying out our examples, we select the threshold m = 4 yielding a moderate degree
of infestation (according to Table 1 of [20]) and the index p ∈ {0.1, 0.2, . . . , 0.7}. Then, for each
resulting set I≥m of potential intervention instants, the problem is to find a single instant t0 ∈ I≥m
that appropriately balances the effectiveness and cost of intervention in the grazing strategy under
consideration. In our approach, the effectiveness and cost functions can be seen as alternative measures
of the efficacy of an intervention, with a negative significance in the case of the cost function. To be
concrete, in an attempt to reflect the effect of the parasite burden on the lamb weight at age τ,
effectiveness is measured in terms of:
3
e f f s ( t0 ; τ ) = ∑ s
πm ( t0 ; τ ),
m =0

which corresponds to the probability that the degree of infestation at age τ is null or light as the
intervention is prescribed at age t0 in accordance with the grazing strategy s with s ∈ {U M, TS, TM }.
In contrast, we make the cost of intervention depend on the probability:

11
costs (t0 ; τ ) = ∑ s
πm s
( t0 ; τ ) + π − 1 ( t0 ; τ )
m =8

that either the host does not survive or its degree of infestation is high at age τ. It is worth noting that
operational (financial) costs are not considered within the modelling framework, which will allow us
to derive a single intervention instant t0 regardless of concrete specifications for the cost of maintaining
safe-pasture conditions, or the cost of purchasing the anthelmintic drugs. Then, the proposed cost
function costs (t0 ; τ ), which can be seen as a negative measure of efficacy, is related to productivity
losses corresponding to high levels of infection, and it may be advisable when the financial fluctuations
(in comparing various drugs, how prices of anthelmintics change) over time are not known in advance.
For a suitable choice of t0 , the following control criteria are suggested:
! "
Criterion 1: We select the intervention instant t0 verifying costs (t0 ; τ ) = inf costs (t; τ ) : t ∈ J≥
1
m ,

m consists of those potential intervention instants t ∈ I≥m satisfying

where the subset J≥ 1

the inequality e f f s (t; τ ) ≥ p1 , for a certain probability p1 ∈ (0, 1).! "

Criterion 2: We select the intervention instant t0 such that e f f s (t0 ; τ ) = sup e f f s (t; τ ) : t ∈ J≥
2
m ,

m is defined by those time instants t ∈ I≥m verifying cost ( t; τ ) ≤ p2 ,

where the subset J≥2 s

for a certain probability p2 ∈ (0, 1).

Our objective in Criterion 1 is thus to minimize the cost of intervention and to maintain a minimum
level of effectiveness, which is translated into the probability p1 ∈ (0, 1). In Criterion 2, the objective
is to maximize the effectiveness and to set an upper bound p2 ∈ (0, 1) to the cost of intervention.
An alternative manner to measure the effectiveness and cost of intervention at a certain age t0 < τ is
given by the respective values:

186
Mathematics 2018, 6, 143

 τ 3
τ −1 E s ( t0 ; τ ) = τ −1 ∑ s
π̃m (t0 ; u)du,
0 m =0
 τ
( )
11
−1 s −1
τ C ( t0 ; τ ) = τ
0
∑ s
π̃m s
(t0 ; u) + π̃− 1 ( t0 ; u ) du,
m =8

where π̃m s ( t ; u ) = π US ( u ) if u ∈ (0, τ ) in scenario US and π̃ s ( t ; u ) = π s ( u ) if u ∈ (0, t ),

0 m m 0 m 0
and πms ( t ; u ) if u ∈ [ t , τ ) in grazing strategy s with s ∈ {UM, TS, TM }; then, values for τ −1 Es ( t ; τ )
0 0 0
and τ −1 C s (t0 ; τ ) are related to the expected proportions of time that the degree of infestation is either
null or light, and either high or heavy, respectively.

3. Empirical Data, Age-Dependent Rates and Results

Age-dependent patterns are from now on specified to reflect that the parasite-induced death of
the host is negligible, and death rates in the absence of any parasite burden at free-living instants and
post-intervention instants are identical, that is δm (t) = δ(t) = δ (t) for levels m ∈ S . Nevertheless,
we point out that, in a general setting, the analytical solution in Equations (1) and (2) allows δ (t)
and γm  ( t ) to be potentially different from δ ( t ) and γ ( t ), respectively, and it can be therefore applied
m
when, among other circumstances, maintaining identical environmental conditions at free-living
and post-intervention instants is not possible (i.e., different rates δ(t) and δ (t)) and/or anthelmintic
resistance must be considered within the modelling framework (i.e., different functions γm (t) and
γm ( t )). In our examples, we select δ ( t ) = δ ( t )=e−10.0t , from which it follows that the probability that,

in absence of any parasite burden, the host dies in the interval [0, τ ] with τ = 1 year equals 9.5162%,
and the conditional probability that the host death occurs within the first 24 hours, given that it dies in
the interval [0, τ ], equals 99.9995%.
In Section 3.1, the age-dependent rates λ(t) and ηm (t) defining grazing strategies UM, TS and
TM are inherently connected to the empirical data in [24] and Figure 2 of [22]. To be concrete, we first
use the results in Section 3.2 of [20] to specify the function λ(t) for time instants t ∈ [0, τ ] in scenario
US and for time instants t ∈ [0, t0 ] in grazing strategies UM, TS and TM. Concrete specifications
for age-dependent patterns at time instants t ∈ (t0 , τ ] are then derived by suitably modifying these
functions under the distributional assumptions in the cases UM, TS and TM. Results yielding scenario
US are related to the study conducted by Uriate et al. [22], which was designed to describe monthly
fluctuations of nematode burden in sheep (Rasa Aragonesa female lambs) raised under irrigated
conditions in Ebro Valley, Spain, by using worm-free tracer lambs and monitoring the faecal excretion
of eggs by ewes. Specifically, the age-dependent rate λ(t) for ages t ∈ [0, τ ] in scenario US and grazing
strategy TS and ages t ∈ [0, t0 ] in the cases UM and TM is obtained by following our arguments in
Section 3.2 of [20]. Therefore, the function λ(t) is related to increments in the number of L3 infective
larvae in the small intestine of the lamb (Figures 2–4, shaded area), and it is computed as a function
of the infection levels in Table 1 of [20]. To reflect the use of safe pasture in grazing strategies UM
and TM, it is assumed that λ (t) = 0.25λ(t) for ages t ∈ (t0 , τ ], where λ(t) denotes the previously
specified function, which is linked to the original paddock. In grazing strategies TS and TM, the
empirical data in [22] are appropriately combined with those data in [24] on the clinical efficacy
assessment of ivermectin, fenbendazole and albendazole in lambs parasited with nematode infective
larvae; similarly to Section 3.2 in [20], the death rates of parasites in the cases TS and TM are then given
by ηm  ( t ) = mη ( t ), for levels m ∈ S , where η ( t ) reflects the chemotherapeutic efficacy of a concrete

anthelmintic over time. More details on the specific form of λ(t) and η (t) can be found in Appendix A.

3.1. Preliminary Analysis

Because of seasonal conditions, a preliminary analysis of the probabilities e f f s (t0 ; τ ) and
costs (t0 ; τ ) in the cases UM, TS and TM is usually required to determine values p1 and p2 in such a
way that Criteria 1 and 2 lead us to non-empty subsets J≥ 1 2
m and J≥m of potential intervention instants

187
Mathematics 2018, 6, 143

t0 ∈ I≥m , for a predetermined threshold m . A graphical representation of e f f s (t0 ; τ ) and costs (t0 ; τ )
can help in measuring allowable values for the minimum value of effectiveness and the maximum
cost of intervention in terms of concrete values for p1 and p2 , respectively. Figures 2 and 3 show how
e f f s (t0 ; τ ) and costs (t0 ; τ ) behave in terms of t0 for grazing strategies UM, TS and TM. We remark
here that, in scenario US, the effectiveness (respectively, cost of intervention) is given by ∑3m=0 πm US ( τ )

(respectively, ∑11 π
m =8 m
US ( τ ) + π US ( τ )), which is a constant as a function of t . It is worth noting that
−1 0
the value ∑3m=0 πm US ( τ ) (respectively, 11 π US ( τ ) + π US ( τ )) results in a lower bound (respectively,
∑ m =8 m −1
upper bound) to the corresponding values of effectiveness (respectively, cost of intervention) in grazing
strategies UM, TS and TM.

Figure 2. Effectiveness e f f s (t0 ; τ ) as a function of the intervention age t0 for τ = 1 year and increments
in the number of L3 infective larvae in the small intestine (shaded area, right vertical axis). Scenario
US, and grazing strategies UM, TS and TM with the anthelmintics ivermectin, fenbendazole and
albendazole (from top to bottom).

188
Mathematics 2018, 6, 143

The effectiveness and cost functions for strategies UM and TM are monotonic in one direction,
while the corresponding curves for strategy TS are largely in the opposite direction. This corroborates
that an early movement of the host to safe pasture results in a more effective (Figure 2) and less
expensive (Figure 3) solution regardless of other actions on the use of anthelmintic drugs, which is
related to the safe-pasture conditions having 75% less free-living L3 than the original paddock. On the
contrary, set-stocking conditions made an early intervention seem inadvisable and, due to the effect of
the therapeutic period (28 days), intervention should be prescribed by the end of November in the
case TS.

Figure 3. Cost costs (t0 ; τ ) of intervention as a function of the intervention age t0 for τ = 1 year
and increments in the number of L3 infective larvae in the small intestine (shaded area, right vertical
axis). Scenario US, and grazing strategies UM, TS and TM with the anthelmintics ivermectin,
fenbendazole and albendazole (from top to bottom).

189
Mathematics 2018, 6, 143

Figure 4. Expected proportions τ −1 Es (t0 ; τ ) (top) and τ −1 C s (t0 ; τ ) (bottom) versus the intervention
age t0 for τ = 1 year and increments in the number of L3 infective larvae in the small intestine
(shaded area, right vertical axis). Scenario US, and grazing strategies UM, TS and TM with the
anthelmintic fenbendazole.

As intuition tells us, grazing strategy TM results in the most effective procedure for every
time instant t0 , regardless of the anthelmintic treatment. In Figures 2 and 3, it is also seen that
grazing strategy UM is preferred to grazing strategy TS when intervention is prescribed at ages
t0 < 293 (21 October), 285 (13 October) and 286 (14 October) as the respective anthelmintics ivermectin,
fenbendazole and albendazole are used in the case TS; on the contrary, the latter is preferred to the
former at intervention instants t0 > 293, 285 and 286. This behaviour is also noted in Figure 4, where we
make the effectiveness and cost of intervention amount to τ −1 Es (t0 ; τ ) and τ −1 C s (t0 ; τ ), respectively;
in such a case, grazing strategy UM is preferred to grazing strategy TS for intervention instants
t0 < 278 (6 October) as the host is treated with fenbendazole, and the latter is preferred to the former
in the case of intervention instants t0 > 281 (9 October). For grazing strategy TS, it is seen in Figure 2
(respectively, Figure 3) that the effectiveness function e f f TS (t0 ; τ ) (respectively, the cost function
cost TS (t0 ; τ )) appears to behave as an increasing (respectively, decreasing) function of the intervention
instant t0 as t0 < 346 (13 December) and 338 (5 December) if the anthelmintic ivermectin and the
anthelmintics fenbendazole and albendazole are administered to the host (respectively, t0 < 309
(6 November), 308 (5 November) and 339 (6 December) if anthelmintics ivermectin, fenbendazole and
albendazole are used); moreover, its variation over time seems to be more apparent, in agreement
with three periods of maximum pasture contamination, with 42.0 L3 kg−1 DM (by mid-February), 68.0
L3 kg−1 DM (by 2 June) and 80.0 L3 kg−1 DM (between October and November) as maximum values
of infective larvae on herbage. Figure 2 (respectively, Figure 3) allows us to remark that, in comparison
with the case TS, these periods of maximum pasture contamination influence in an opposite manner
the effectiveness (respectively, cost of intervention) in grazing strategies UM and TM.

190
Mathematics 2018, 6, 143

3.2. Intervention Instants t0

In Table 1, we list the value of the effectiveness e f f s (t0 ; τ ) and the cost costs (t0 ; τ ) of intervention
for certain intervention instants t0 derived by applying Criteria 1 and 2 in grazing strategies UM,
TS and TM, for probabilities p1 ∈ {0.50, 0.60, 0.70} and p2 ∈ {0.15, 0.20, 0.25} and a variety of values
of the index p; in scenario US, effectiveness and cost are replaced by the probabilities ∑3m=0 πm US ( τ )
11
and ∑m=8 πm (τ ) + π−1 (τ ), respectively. A detailed discussion on the instants t0 in Table 1 and some
US US

related consequences can be found in Appendix A.2. It can be noticed that the selection t0 = 273
(1 October), which is related to the index p = 0.1 in the case TM with the anthelmintic fenbendazole,
results in the minimum cost of intervention (0.09589, instead of 0.49951 in scenario US) and the
maximum effectiveness (0.79086, instead of 0.06072 in scenario US), and it can be thus taken as optimal
for our purposes. Moreover, the anthelmintic fenbendazole is found the most effective drug since the
highest values of e f f s (t0 ; τ ) and the smallest values of costs (t0 ; τ ) are observed in Table 1 for every
grazing strategy s ∈ { TS, TM } and fixed intervention instant t0 .

Table 1. Effectiveness and cost of intervention. Scenario US and grazing strategies UM, TS and TM
with the anthelmintics ivermectin, fenbendazole and albendazole.

Strategy (s) Anthelmintic t0 Criteria e f f s ( t0 ; τ ) cost s (t0 ; τ ) τ − 1 E s ( t0 ; τ ) τ − 1 C s ( t0 ; τ )

US — — — 0.06072 0.49951 0.68645 0.14746
UM 170 1&2 0.54431 0.11049 0.79996 0.09726
274 2 0.45540 0.12524 0.76629 0.09983
281 2 0.38981 0.14216 0.74973 0.10267
286 2 0.32115 0.16811 0.73306 0.10715
290 2 0.26634 0.19763 0.72023 0.11233
298 2 0.20886 0.24130 0.70769 0.11984
TS ivermectin 358 2 0.41766 0.16608 0.69160 0.14217
fenbendazole 308 1 0.50340 0.12350 0.75871 0.10433
336 1 0.60161 0.13144 0.71941 0.12421
338 2 0.60604 0.13209 0.71613 0.12578
albendazole 313 1 0.50240 0.12842 0.74908 0.10793
338 2 0.57385 0.13407 0.71312 0.12626
TM ivermectin 170 1&2 0.73224 0.09721 0.86987 0.09525
274 1&2 0.71025 0.09797 0.82480 0.09580
281 1&2 0.69119 0.09877 0.81634 0.09602
286 1&2 0.66653 0.10011 0.80686 0.09644
290 1&2 0.64110 0.10197 0.79743 0.09713
298 1&2 0.61142 0.10528 0.78209 0.09911
308 1&2 0.56977 0.11374 0.76202 0.10372
fenbendazole 273 1&2 0.79086 0.09589 0.83891 0.09557
274 1&2 0.79080 0.09589 0.83820 0.09558
281 1&2 0.78559 0.09601 0.83107 0.09573
286 1&2 0.77604 0.09636 0.82304 0.09605
290 1&2 0.76467 0.09707 0.81476 0.09662
298 1&2 0.75182 0.09895 0.79922 0.09852
308 1&2 0.72721 0.10573 0.77734 0.10310
albendazole 272 1&2 0.78128 0.09605 0.83749 0.09558
274 1&2 0.78102 0.09606 0.83605 0.09560
281 1&2 0.77361 0.09623 0.82838 0.09576
286 1&2 0.76132 0.09666 0.81971 0.09610
290 1&2 0.74737 0.09747 0.81089 0.09671
298 1&2 0.73134 0.09945 0.79492 0.09867
308 1&2 0.70211 0.10641 0.77271 0.10336

Values for τ −1 Es (t0 ; τ ) and τ −1 C s (t0 ; τ ) in Table 1 correspond to the expected proportions of time
that the host infection level M(t) remains in the subsets of levels {0, 1, 2, 3} and {8, 9, 10, 11} ∪ {−1},
respectively. It is remarkable to note that the maximum effectiveness τ −1 Es (t0 ; τ ) = 0.86987 (instead
of 0.68645 in scenario US) and the minimum cost of intervention τ −1 C s (t0 ; τ ) = 0.09525 (instead of
0.14746 in scenario US) are both related to the selection t0 = 170 (19 June) in grazing strategy TM with

191
Mathematics 2018, 6, 143

the anthelmintic ivermectin. It should be noted that t0 = 170 results in the longest post-intervention
interval [t0 , τ ] in our examples; similarly to the case of control strategies based on isolation and
anthelmintic treatment of the host (see Section 3.3 in [20]), the maintenance of stable safe-pasture
conditions for a long period of time may often be difficult and highly expensive, so that the choice
t0 = 170 might be unsustainable for practical use.
An interesting question concerns the comparative analysis between the mass functions {πm s (t ; τ ) :
0
m ∈ {−1} ∪ S} of the parasite burden at age τ = 1 year in grazing strategies UM, TS and TM and the
corresponding mass function {πm US ( τ ) : m ∈ {−1} ∪ S} in the case of no intervention. In Figure 5,

we first focus on this question as intervention is prescribed at age t0 = 170 in grazing strategies
UM, TS and TM, with the anthelmintic drug ivermectin in the cases TS and TM. The movement of
the host to safe pasture (strategies UM and TM) at day t0 = 170 yields a significant decrease in the
probability that the host does not survive at age τ = 1 year (0.09528 and 0.09516 in the cases UM and
TM, respectively, instead of 0.15708 in scenario US), as well as an important decrease in the expected
degree of infestation in the case of survival; more particularly, the degree of infestation is expected
to be light as either anthelmintic drugs are used (E TM,t0 [ M(τ )| M (τ ) = −1] = 2.25085) or the host is
transferred to a paddock with safe pasture (EUM,t0 [ M (τ )| M(τ ) = −1] = 3.20312), instead of moderate
and nearly high in the case US (EUS [ M(τ )| M (τ ) = −1] = 6.88878). Set-stocking conditions are not as
effective as the movement of the host to safe pasture since the expected degree of infestation amounts
to a moderate degree in the case of survival (E TS,t0 [ M (τ )| M(τ ) = −1] = 6.13509); moreover, for
grazing strategy TS, the decrease in the probability of no-survival is apparent, but it is not as notable as
for strategies UM and TM. In Figure 6, we plot the mass function of the parasite burden M(τ ) at age
τ = 1 year in scenario US versus its counterpart in grazing strategy TM, when animals are treated with
ivermectin, fenbendazole and albendazole at ages t0 = 170, 273 and 272, respectively. By Tables A2
and A3 in Appendix A, ages t0 = 170, 273 and 272 are all feasible intervention instants, which leads
us to mass functions that are essentially comparable in magnitude. On the contrary, the shape and
magnitudes of the mass function in grazing strategy TM are dramatically different from the shape
and magnitudes in scenario US, where no intervention is prescribed, irrespective of the anthelmintic
product.

Figure 5. The mass function of the parasite burden M(τ ) at age τ = 1 year. Scenario US and grazing
strategies UM, TS and TM (from left to right) with the anthelmintic ivermectin as the intervention
prescribed at age t0 = 170.

192
Mathematics 2018, 6, 143

Figure 6. The mass function of the parasite burden M(τ ) at age τ = 1 year. Scenario US and grazing
strategy TM with the anthelmintics ivermectin, fenbendazole and albendazole (from left to right) as the
intervention prescribed at ages t0 = 170, 273 and 272, respectively.

4. Conclusions
It is of fundamental importance in the development of GI nematode infection in sheep to
understand the role of grazing management in reducing anthelmintic use and improving helminth
control. With empirical data of [22,24], we present a valuable modelling framework for better
understanding the host-parasite interaction under fluctuations in time, which arguably represents the
most realistic setting for assessing the impact of seasonal changes in the parasite burden of a growing
lamb. Grazing strategies UM, TS and TM in Section 2.1 are defined in terms of an eventual movement
to safe pasture and/or chemotherapeutic treatment of the host at a certain age t0 ∈ (0, τ ). For a suitable
choice of t0 , we suggest to use two control criteria that adequately balance the effectiveness and cost of
intervention at age t0 by using simple stochastic principles. Specifically, each intervention instant t0
in Table 1 yields an individual-based grazing strategy for a lamb that is born, parasite-free, at time
t = 0 (1 January, in our examples). The individual-based grazing strategies UM, TS and TM can be
also thought of as group-based grazing strategies in the case of a flock consisting of young lambs,
essentially homogeneous in age. In such a case, intervention at age t0 is prescribed (in accordance with
a predetermined grazing strategy) by applying our methodology to a typical lamb that is assumed to
be born, parasite-free, at a certain average day t . Then, results may be routinely derived by handling
the set of empirical data in Figure 2 in [22] starting from day t , instead of Day 0, since intervention at
time instant t + t0 amounts to age t0 of the typical lamb in the paddock. From an applied perspective,
the descriptive model in Section 2 becomes a prescriptive model as the set of empirical data in Figure 2
in [22] is appropriately replaced by a set of data derived by taking the average of annual empirical
data from historical records.
For practical use, the profits of applying Criteria 1 and 2 in grazing strategies UM, TS and TM
should be appropriately compared with experimental results. To that end, we first comment on general
guidelines (see Part II of [25]) for control of GI nematode infection. From an experimental perspective,
the dose-and-move strategy (termed TM) is usually recommended in mid-July, this recommendation
being applicable in temperate zones where the maximum numbers of L3 infective larvae do not occur
before midsummer, which is our case (Figures 2–4, shaded area). As stated in [25], midsummer
movement to safe pasture without deworming (strategy UM) is thought of as a low cost control
measure; it can even be effective at moderate levels of pasture infectivity, and it has the advantage

193
Mathematics 2018, 6, 143

of creating no anthelmintic resistance by drug selection. We may translate these specifications into
an intervention at day t0 = 195 (15 July) in strategies TM and UM. Guidelines for the application
of an anthelmintic drug without movement (strategy TS) are not so clearly available, and various
alternatives (based mainly on the several-dose approach) are applied in a variety of circumstances that
strongly depend on the geographical region, climate and farming system. For comparative purposes
in strategy TS, we compare our results (derived by applying Criteria 1 and 2) with an intervention at
day t0 = 287 (15 October), when the maximum number of infective larvae L3 on herbage is observed
(Figures 2–4, shaded area).
We present in Table 2 a sample of our results when the anthelmintic fenbendazole is used in
grazing strategies TS and TM. Table 2 lists values of the reduction in the mean infection level (RMIL)
at age τ = 1 year, under the taboo that the host survives at age τ = 1 year, and the reduction in the
total lost probability (RTLP) when, instead of scenario US, intervention is prescribed at day t0 < τ by a
certain grazing strategy s with s ∈ {UM, TS, TM}. The indexes RMIL and RTLP for grazing strategy s,
with s ∈ {UM, TS, TM}, are defined by:


Es,t0 [ M (τ )| M(τ ) = −1]
RMILs,t0 = 100 × 1 − US %,
E [ M (τ )| M(τ ) = −1]


Ps,t0 ( M(τ ) = −1)
RTLPs,t0 = 100 × 1 − US %,
P ( M ( τ ) = −1)

where Es,t0 [ M (τ )| M(τ ) = −1] denotes the conditional expected infection level of the host at age
τ = 1 year, given that it survives at age τ = 1 year, and Ps,t0 ( M(τ ) = −1) is the probability that
the host does not survive at age τ = 1 year, when intervention is prescribed at day t0 according to
grazing strategy s. The values EUS [ M(τ )| M(τ ) = −1] and PUS ( M(τ ) = −1) are related to scenario
US, and they reflect no intervention.

Table 2. Indexes reduction in the mean infection level (RMIL) and reduction in the total lost probability
(RTLP) for strategies UM, TS and TM with the anthelmintic fenbendazole.

Strategy (s) Criteria t0 RM ILs,t0 RT LPs,t0

UM 1&2 170 53.50% 39.34%
2 274 46.98% 39.17%
2 281 41.85% 38.88%
2 286 36.00% 38.21%
2 290 30.78% 37.12%
2 298 24.54% 34.86%
Midsummer 195 50.28% 39.28%
TS 1 308 51.01% 33.27%
1 336 59.19% 19.77%
2 338 59.56% 19.10%
Maximum pasture contamination 287 37.78% 38.17%
TM 1&2 273 72.36% 39.41%
1&2 274 72.36% 39.41%
1&2 281 71.88% 39.39%
1&2 286 71.03% 39.26%
1&2 290 70.07% 38.92%
1&2 298 69.09% 37.82%
1&2 308 67.47% 33.63%
Midsummer 195 70.47% 39.42%

In grazing strategies UM and TM, the experimental selection t0 = 195 (midsummer) is found
to be near an optimal solution, and Table 2 permits us to analyse the effects of the stochastic control
criteria in a more detailed manner. Based on the decreasing monotonic behaviour of RMIL and RTLP
with respect to the intervention instant t0 , it is noticed that the later we apply grazing strategies UM

194
Mathematics 2018, 6, 143

and TM, the worse the results we obtain. This is closely related to the important role played in the
cases UM and TM by the use of safe pasture, which is reflected in the 75% contamination reduction
with respect to the original paddock. Therefore, the movement of the host to safe pasture appears to be
dominant in the use of anthelmintics, so that the sooner the host is moved, the safer it is for the host.
The maintenance of stable safe-pasture conditions for a long period of time may be difficult and/or
highly expensive, whence additional considerations should be taken into account when selecting
the time instant t0 for moving the host. In grazing strategy UM, the intervention instant t0 = 170
should be considered as optimal for our purposes, and it yields a reduction of 53.50% in the mean
infection level at the end of the year, as well as a reduction of 39.34% in the probability of no-survival.
However, an intervention at day t0 = 274 (i.e., moving the host to safe pasture more than one hundred
days later) would result in significantly lower operational costs, but predicted reductions are still
around high levels (RMILUM,274 = 46.98%, RTLPUM,274 = 39.17%). It is clear that a balance between
operational costs and the magnitudes of the indexes RMIL and RTLP should be made. It is seen
that the experimental selection t0 = 195 seems to implicitly incorporate this balance, delaying the
movement of the host almost a month with respect to t0 = 170, at the expense of losing 3.22% and
0.06% of efficiency in the indexes RMIL and RTLP, respectively. Although the selection of t0 may
depend on external factors, the movement of the host to safe pasture before day t0 = 287 (maximum
pasture contamination) is highly recommendable, and intervention instants t0 = 290 and 298 should
be discarded in the light of these results.
Similar comments can be made for grazing strategy TM. In this case, the experimental selection
t0 = 195 allows us to achieve a good index RMIL in comparison with those time instants t0 obtained
by applying Criteria 1 and 2, while obtaining the highest index RTLP. The intervention at day t0 = 195
is more than two months advanced with respect to the day t0 = 273, which is derived by applying
Criteria 1 and 2. The experimental selection t0 = 195 results in higher operational costs due to an early
movement, and it amounts to a minor improvement of 0.01% in the index RTLP; it is also seen that the
option t0 = 273 yields the value RMIL TM,273 = 72.36%, which is higher than the corresponding value
for the experimental choice. Thus, when comparing grazing strategy TM with strategy UM, the use
of an anthelmintic drug seems to permit delaying the movement of the host to safe pasture, while
maintaining good indexes RMIL and RTLP; note that it is still possible to have values of RMIL and
RTPL above 70% and 39%, respectively, if the intervention is delayed at day t0 = 286.
Under set-stocking conditions, the use of an anthelmintic drug at day t0 = 287 (15 October) may
be seen as optimal in terms of the index RTLP, but at the expense of an unacceptable value 37.78% of
RMIL. Note that an application of Criteria 1 and 2 leads us to intervention instants t0 = 308, 336 and
338, with values RMIL TS,t0 varying between 50% and 60%. In particular, the time instant t0 = 308
permits us to achieve a significant improvement of the index RMIL (51.01% instead of 37.78%) and
maintain the index RTLP above 33%, which is comparable with the value 38.17% resulting from an
intervention when maximum values of L3 on herbage are observed.
One of the simplifying assumptions in Section 3 (see also Appendix A) is related to the effect
that the infestation degree of the lambs might have on the pasture infection level itself. We deal
with a non-infectious assumption, and specifically, the empirical data in Figure 2 in [22] allow us
to partially incorporate this effect into the age-dependent patterns in terms of the infection level of
a standard paddock during the year. The analytical solution in Section 2.1 can be however used to
examine the infectious nature of the parasite in a more explicit manner. Although it is an additional
topic for further study, we stress that the infectious nature of the parasite appears to be a relevant
feature in grazing strategy UM, where the force-of-infection in a field seeded with untreated lambs
would likely increase back up to a similar level to the original paddock. In an attempt to address this
question, various variants of the age-dependent rate λ (t) can be conjectured, such as the function
λ (t) = (0.25 + 0.75h−1 (t − t0 ))λ(t) at post-intervention instants, with h > 0. Then, under proper data
availability, the selection λ (t) reflects the use of a paddock with safe pasture at initial post-intervention

195
Mathematics 2018, 6, 143

instants (λ (t0 ) = 0.25λ(t0 )) and how the pasture infection level reaches the pre-intervention level,
represented by λ (t0 + h) = λ(t0 + h), after a period consisting of h days.

Author Contributions: A.G.-C. and M.L.-G. conceived of the model, carried out the stochastic analysis and
designed the numerical results. M.L.-G. programmed the computer codes and obtained results for all figures
and tables. A.G.C. wrote the first version of the draft. A.G.-C. and M.L.-G. finalised the paper and revised
the literature.
Funding: This research was funded by the Ministry of Economy, Industry and Competitiveness (Government of
Spain), project MTM2014-58091-P and grant BES-2009-018747.
Acknowledgments: The authors thank two anonymous referees whose comments and suggestions led to
improvements in the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A. GI Nematode Infection in Growing Lambs

In its complete life cycle, the parasitic phase of Nematodirus spp. commences when worms in
the larval stage L3 encounter the host, which is a largely passive process with the grazing animal
inadvertently ingesting larvae with herbage as it feeds. As a result, infection occurs by ingestion of
the free-living L3 , with an establishment proportion (i.e., the proportion of ingested free-living L3 that
become established in the small intestine of the host) ranging between 45% and 60%; see, e.g., [26–28].
Various external factors (moisture levels, temperature and the availability of oxygen) are key drivers
that affect how quickly eggs hatch and larvae develop and how long larvae and eggs survive on pasture.
Therefore, the occurrence of nematode infections in sheep is inherently linked to seasonal conditions,
and it is therefore connected to a diversity of physiographic and climatic conditions; see [22,29,30],
among others. The adverse effects of GI nematode parasites on productivity are diverse, and reductions
of live weight gain in growing stock have been recorded as being as high as 60–100%. Anthelmintics,
such as ivermectin, fenbendazole and albendazole, are drugs that are effective in removing existing
burdens or that prevent establishment of ingested L3 .
Faecal examination for the presence of worm eggs or larvae is the most common routine aid to
diagnosis employed. In the faecal egg count (FEC) reduction test, animals are allocated to groups of
ten based on pre-treatment FEC, with one group of ten for each anthelmintic treatment tested and a
further untreated control group. For instance, this requires the use of forty animals in [24], where the
efficacy of three anthelmintics (ivermectin, fenbendazole and albendazole) against GI nematodes is
investigated. A full FEC reduction test is understandably expensive and takes a significant length
of time before farmers are presented with the results; in addition, accurate larval differentiation also
demands a high degree of skill. As an alternative test, a points system (see [21]) may serve as a crude
guide to interpreting worm counts, which is based on the fact that one point is equivalent to the
presence of 4000 worms, a total of two points in a young sheep is likely to be causing measurable
losses of productivity and clinical signs and deaths are unlikely unless the total exceeds three points.
Based on the above comments, Table 1 in [20] presents an equivalence in the identification of
the degree of infestation, level of infection, eggs per gram (EPG) value, number of L3 infective larvae
in the small intestine and the points system, which can be used to study the parasite load of a lamb
in a unified manner. We refer the reader to [1,2,21] for further details on nematode taxonomy and
morphology and the treatment and control of parasite gastroenteritis in sheep.

Appendix A.1. Empirical Data and Age-Dependent Rates

In this section, we first use the results in Section 3.2 of [20] to specify the functions λ(t) and ηm (t)
for time instants t ∈ [0, τ ] in scenario US and for time instants t ∈ [0, t0 ] in grazing strategies UM,
TS and TM. Concrete specifications for age-dependent patterns at time instants t ∈ (t0 , τ ] are then
derived by suitably modifying these functions under the distributional assumptions in the cases UM,
TS and TM. Results yielding scenario US are related to the study conducted by Uriate et al. [22], which
is designed to describe monthly fluctuations of nematode burden in sheep (Rasa Aragonesa female

196
Mathematics 2018, 6, 143

lambs) raised under irrigated conditions in Ebro Valley, Spain, by using worm-free tracer lambs and
monitoring the faecal excretion of eggs by ewes. Specifically, we use the set of empirical data in Figure 2
of [22] recording the number of L3 infective larvae on herbage samples at weekly intervals from a fixed
paddock of the farm. In grazing strategies TS and TM, the empirical data in [22] are appropriately
combined with those data in [24] on the clinical efficacy assessment of ivermectin, fenbendazole and
albendazole in lambs parasited with nematode infective larvae.
In Figure 2 of [22], results are expressed as infective larvae per kilogram of dry matter (L3 kg−1
DM) after drying the herbage overnight at 60◦ C, and the numbers of L3 infective larvae on herbage
samples correspond to Chabertia ovina and Haemonchus spp. (9.6%), Nematodirus spp. (4.0%), Ostertagia
spp. (71.4%) and Trichostrongylus spp. (15.0%). In our work, the increments in the number of L3
infective larvae in the small intestine (Figure 1, shaded area) are estimated by fixing the value 55% as
the establishment proportion and incorporating concrete specifications for the lamb growth pre- and
post-weaning. To be concrete, it is assumed that, for a host that is born on 1 January (Day 0), the lamb
birth weight equals 5 kg, the pre-weaning period consists of four weeks and the lamb growth rate
from birth to weaning is given by 0.3 kg per day. The lamb growth rate on pasture post-weaning is
assumed to be equal to 0.15 kg per day, and the daily DM intake is given by the 6% of body weight
(BW); see [31] for details on lamb growth rates on pasture.
These specifications determine the age-dependent rate λm (t) = λ(t) for ages t ∈ [0, τ ] in scenario
US and grazing strategy TS, and for ages t ∈ [0, t0 ] in grazing strategies UM and TM, with τ = 1 year.
More concretely, the function λ(t) is defined to be the piecewise linear function formed by connecting
the points (n, λ(n)) in order by segments, where the value λ(n) at the n-th day is determined in [20]
as a function of the number of L3 infective larvae of Nematodirus spp. on pasture, from Figure 2
of [22], the DM intake at the n-th day, the establishment proportion and the interval length l = 103
used in Table 1 of [20] to define infection levels m ∈ S in terms of numbers of infective larvae in the
small intestine. To reflect the use of safe pasture in grazing strategies UM and TM, it is assumed that
λ (t) = 0.25λ(t) for ages t ∈ (t0 , τ ] where λ(t) denotes the previously specified function, which is
related to the original paddock.
Similarly to Section 3.2 in [20], the death rates of parasites in grazing strategies TS and TM
are given by ηm  ( t ) = mη ( t ) for levels m ∈ S , where η ( t ) reflects the chemotherapeutic efficacy of

a concrete anthelmintic over time. We use the empirical data of [24], where the efficacy of three
anthelmintic products against GI nematodes is investigated. In the FEC reduction test of [24], animals
were allocated to four groups termed A, B, C and D. Animals of Group A served as the control, whereas
animals of Groups B, C and D were orally administered ivermectin (0.2 mg·kg−1 ·BW), fenbendazole
(5.0 mg·kg−1 ·BW) and albendazole (7.5 mg·kg−1 ·BW), respectively. Animals were sampled for FEC at
Day 0 immediately before administering the drug and thereafter on Days 3, 7, 14, 21 and 28. Then,
the function η (t) associated with each anthelmintic is defined as the polyline connecting the points
(tn , η (tn )), where the instants tn are given by t0 , t1 = t0 + 3, t2 = t0 + 7, t3 = t0 + 14, t4 = t0 + 21 and
t5 = t0 + 28. The length of the therapeutic period is assumed to be equal to 28 days, so that η (t) = 0
for instants t > t5 . Values η (tn ) are determined in Table 1 of [20] from the EPG value and the infection
level at time tn , as well as the length l  = 50 used to define levels of infection in terms of EPG values.

Appendix A.2. Intervention Instants t0

Values of t0 are listed in Table A1 for grazing strategy UM and denoted by t10 and t20 as they are
derived by applying Criteria 1 and 2, respectively. In Tables A2 and A3, values of t0 are listed for
grazing strategies TS and TM and the anthelmintics ivermectin (Group B), fenbendazole (Group C)
and albendazole (Group D), which are denoted by t0B , t0C and t0D , respectively. The selection m = 4
in Tables A1–A3 amounts to a degree of infestation that is moderate (Figure A1), and consequently,
a measurable presence of worms is observed.

197
Mathematics 2018, 6, 143

Table A1. Intervention instants t0 versus the index p and the lower bound p1 for effectiveness
(Criterion 1) and the upper bound p2 for the cost of intervention (Criterion 2) for m = 4. Grazing
strategy UM.

p I≥4 p1 1
J≥ t01 p2 2
J≥ t02
4 4

0.1 [170, 365) 0.70 —– —– 0.25 [170, 299] 170

0.60 —– —– 0.20 [170, 290] 170
0.50 [170, 194] 170 0.15 [170, 282] 170
0.2 [274, 365) 0.70 —– —– 0.25 [274, 299] 274
0.60 —– —– 0.20 [274, 290] 274
0.50 —– —– 0.15 [274, 282] 274
0.3 [281, 365) 0.70 —– —– 0.25 [281, 299] 281
0.60 —– —– 0.20 [281, 290] 281
0.50 —– —– 0.15 [281, 282] 281
0.4 [286, 365) 0.70 —– —– 0.25 [286, 299] 286
0.60 —– —– 0.20 [286, 290] 286
0.50 —– —– 0.15 —– —–
0.5 [290, 365) 0.70 —– —– 0.25 [290, 299] 290
0.60 —– —– 0.20 [290, 290] 290
0.50 —– —– 0.15 —– —–
0.6 [298, 365) 0.70 —– —– 0.25 [298, 299] 298
0.60 —– —– 0.20 —– —–
0.50 —– —– 0.15 —– —–
0.7 [308, 365) 0.70 —– —– 0.25 —– —–
0.60 —– —– 0.20 —– —–
0.50 —– —– 0.15 —– —–

Figure A1. The age-dependent probability P≥m (t) as a function of the age t ∈ (0, τ ) with τ = 1 year,
for m = 1 (broken line), 4 (dotted line) and 8 (solid line), and increments in the number of L3 infective
larvae on the small intestine (shaded area, right vertical axis).

198
Mathematics 2018, 6, 143

Table A2. Intervention instants t0 versus the index p and the lower bound p1 for effectiveness
(Criterion 1) for m = 4. Grazing strategies TS and TM with the anthelmintics ivermectin (B), fenbendazole
(C) and albendazole (D).

1,B 1,C 1,D

p I≥4 p1 J≥ 4 t0B J≥ 4 t0C J≥ 4 t0D
0.1 [170, 365) 0.70 TS —– —– —– —– —– —–
TM [170, 278] 170 [170, 319] 273 [170, 308] 272
0.60 TS —– —– [336, 339] 336 —– —–
TM [170 ,301] 170 [170, 344] 273 [170, 342] 272
0.50 TS —– —– [308, 344] 308 [313, 343] 313
TM [170, 343] 170 [170, 348] 273 [170, 346] 272
0.2 [274, 365) 0.70 TS —– —– —– —– —– —–
TM [274, 278] 274 [274, 319] 274 [274, 308] 274
0.60 TS —– —– [336, 339] 336 —– —–
TM [274, 301] 274 [274, 344] 274 [274, 342] 274
0.50 TS —– —– [308, 344] 308 [313, 343] 313
TM [274, 343] 274 [274, 348] 274 [274, 346] 274
0.3 [281, 365) 0.70 TS —– —– —– —– —– —–
TM —– —– [281, 319] 281 [281, 308] 281
0.60 TS —– —– [336, 339] 336 —– —–
TM [281, 301] 281 [281, 344] 281 [281, 342] 281
0.50 TS —– —– [308, 344] 308 [313, 343] 313
TM [281, 343] 281 [281, 348] 281 [281, 346] 281
0.4 [286, 365) 0.70 TS —– —– —– —– —– —–
TM —– —– [286, 319] 286 [286, 308] 286
0.60 TS —– —– [336, 339] 336 —– —–
TM [286, 301] 286 [286, 344] 286 [286, 342] 286
0.50 TS —– —– [308, 344] 308 [313, 343] 313
TM [286, 343] 286 [286, 348] 286 [286, 346] 286
0.5 [290, 365) 0.70 TS —– —– —– —– —– —–
TM —– —– [290, 319] 290 [290, 308] 290
0.60 TS —– —– [336, 339] 336 —– —–
TM [290, 301] 290 [290, 344] 290 [290, 342] 290
0.50 TS —– —– [308, 344] 308 [313, 343] 313
TM [290, 343] 290 [290, 348] 290 [290, 346] 290
0.6 [298, 365) 0.70 TS —– —– —– —– —– —–
TM —– —– [298, 319] 298 [298, 308] 298
0.60 TS —– —– [336, 339] 336 —– —–
TM [298, 301] 298 [298, 344] 298 [298, 342] 298
0.50 TS —– —– [308, 344] 308 [313, 343] 313
TM [298, 343] 298 [298, 348] 298 [298, 346] 298
0.7 [308, 365) 0.70 TS —– —– —– —– —– —–
TM —– —– [308, 319] 308 [308, 308] 308
0.60 TS —– —– [336, 339] 336 —– —–
TM —– —– [308, 344] 308 [308, 342] 308
0.50 TS —– —– [308, 344] 308 [313, 343] 313
TM [308, 343] 308 [308, 348] 308 [308, 346] 308

199
Mathematics 2018, 6, 143

Table A3. Intervention instants t0 versus the index p and the upper bound p2 for the cost of intervention
(Criterion 2) for m = 4. Grazing strategies TS and TM with the anthelmintics ivermectin (B), fenbendazole
(C) and albendazole (D).

2,B 2,C 2,D

p I≥4 p2 J≥ 4 t0B J≥ 4 t0C J≥ 4 t0D
0.1 [170, 365) 0.25 TS [286, 363] 358 [268, 363] 338 [270, 363] 338
TM [170, 363] 170 [170, 363] 273 [170, 363] 272
0.20 TS [299, 362] 358 [279, 362] 338 [281, 361] 338
TM [170, 362] 170 [170, 362] 273 [170, 362] 272
0.15 TS —– —– [290, 346] 338 [292, 344] 338
TM [170, 350] 170 [170, 350] 273 [170, 348] 272
0.2 [274, 365) 0.25 TS [286, 363] 358 [274, 363] 338 [274, 363] 338
TM [274, 363] 274 [274, 363] 274 [274, 363] 274
0.20 TS [299, 362] 358 [279, 362] 338 [281, 361] 338
TM [274, 362] 274 [274, 362] 274 [274, 362] 274
0.15 TS —– —– [290, 346] 338 [292, 344] 338
TM [274, 350] 274 [274, 350] 274 [274, 348] 274
0.3 [281, 365) 0.25 TS [286, 363] 358 [281, 363] 338 [281, 363] 338
TM [281, 363] 281 [281, 363] 281 [281, 363] 281
0.20 TS [299, 362] 358 [281, 362] 338 [281, 361] 338
TM [281, 362] 281 [281, 362] 281 [281, 362] 281
0.15 TS —– —– [290, 346] 338 [292, 344] 338
TM [281, 350] 281 [281, 350] 281 [281, 348] 281
0.4 [286, 365) 0.25 TS [286, 363] 358 [286, 363] 338 [286, 363] 338
TM [286, 363] 286 [286, 363] 286 [286, 363] 286
0.20 TS [299, 362] 358 [286, 362] 338 [286, 361] 338
TM [286, 362] 286 [286, 362] 286 [286, 362] 286
0.15 TS —– —– [290, 346] 338 [292, 344] 338
TM [286, 350] 286 [286, 350] 286 [286, 348] 286
0.5 [290, 365) 0.25 TS [290, 363] 358 [290, 363] 338 [290, 363] 338
TM [290, 363] 290 [290, 363] 290 [290, 363] 290
0.20 TS [299, 362] 358 [290, 362] 338 [290, 361] 338
TM [290, 362] 290 [290, 362] 290 [290, 362] 290
0.15 TS —– —– [290, 346] 338 [292, 344] 338
TM [290, 350] 290 [290, 350] 290 [290, 348] 290
0.6 [298, 365) 0.25 TS [298, 363] 358 [298, 363] 338 [298, 363] 338
TM [298, 363] 298 [298, 363] 298 [298, 363] 298
0.20 TS [299, 362] 358 [298, 362] 338 [298, 361] 338
TM [298, 362] 298 [298, 362] 298 [298, 362] 298
0.15 TS —– —– [298, 346] 338 [298, 344] 338
TM [298, 350] 298 [298, 350] 298 [298, 348] 298
0.7 [308, 365) 0.25 TS [308, 363] 358 [308, 363] 338 [308, 363] 338
TM [308, 363] 308 [308, 363] 308 [308, 363] 308
0.20 TS [308, 362] 358 [308, 362] 338 [308, 361] 338
TM [308, 362] 308 [308, 362] 308 [308, 362] 308
0.15 TS —– —– [308, 346] 338 [308, 344] 338
TM [308, 350] 308 [308, 350] 308 [308, 348] 308

An examination of the resulting instants t0 in Tables A1–A3 reveals the following

important consequences:

(i) In applying Criterion 1 (respectively, Criterion 2) to grazing strategy TM, values of the lower
bound p1 ∈ {0.5, 0.6, 0.7} for effectiveness (respectively, the upper bound p2 ∈ {0.15, 0.2, 0.25}
for the cost of intervention) result in identical intervention instants t0 , irrespective of the
anthelmintic drug, with the exception of the case p = 0.1. More concretely, we observe that,
in the case p = 0.1, identical intervention instants t0 are derived for each fixed anthelmintic
drug, but a replacement of the predetermined drug by another anthelmintic results in different
intervention instants.

200
Mathematics 2018, 6, 143

(ii) For every anthelmintic drug and fixed index p, Criteria 1 and 2 applied to grazing strategy
TM yield identical intervention instants t0 , with the exception of those pairs ( p, p1 ) for the
1,B
anthelmintic ivermectin leading us to empty subsets J≥ 4 . In order to maintain high values of
the minimum level of effectiveness (Criterion 1), we have therefore to handle smaller values
of the index p (0.1 and 0.2 in Table A2) for grazing strategy TM, which means that low-risk
intervention instants should become potential intervention instants.
(iii) For every anthelmintic, the intervention instant t0 derived in grazing strategy TM behaves as an
increasing function of the index p, regardless of the control criterion.
(iv) For every anthelmintic and fixed value p1 , the intervention instant t0 in grazing strategy TS
appears to be constant as a function of the index p. This is in agreement with the fact that the
maximum levels of effectiveness (Figure 2) and the minimum costs of intervention (Figure 3) are
observed at the end of the year (November–December), in such a way that this period of time
always consists of potential intervention instants (Figure A1) for the index p ranging between
0.1 and 0.7.
(v) In contrast to grazing strategies TS and TM, the values p1 ∈ {0.5, 0.6, 0.7} for grazing strategy
UM lead us to empty subsets J≥ 1 of potential intervention instants, with the exception of the
4
pair ( p, p1 ) = (0.1, 0.5). This observation is closely related to the monotonic behaviour of the
effectiveness (Figure 2) and cost (Figure 3) functions, which links the first months of the year to
the highest effectiveness and the minimum cost of intervention.
(vi) The upper limit of the set I≥4 in Tables A1–A3 is always at Day 365, which can be readily
explained from the monotone behaviour (Figure A1) of the age-dependent probability P≥m (t)
in the case m = 4. It is clear that other thresholds m will not necessarily yield Day 365;
for example, I≥m = (280, 360) in the case m = 1 with p = 0.85.
(vii) For strategies UM (Table A1) and TM (Tables A2 and A3), the lower limits of the resulting sets
1 and J 2 always coincide with the lower limit of the set I
J≥ 4 ≥4 ≥4 of potential intervention instants
t0 , but this is not the case for strategy TS. This means that an early movement of the host to safe
pasture should lead to feasible intervention instants.

The values of the effectiveness e f f s (t0 ; τ ) and the cost costs (t0 ; τ ) of intervention for instants t0 in
Tables A1–A3 are listed in Table 1 and analysed in more detail in Section 3.2.

References
1. Sutherland, I.; Scott, I. Gastrointestinal Nematodes of Sheep and Cattle. Biology and Control; Wiley-Blackwell:
Chichester, UK, 2010.
2. Taylor, M.A.; Coop, R.L.; Wall, R.L. Veterinary Parasitology, 3rd ed.; Blackwell: Oxford, UK, 2007.
3. Bjørn, H.; Monrad, J.; Nansen, P. Anthelmintic resistance in nematode parasites of sheep in Denmark with
special emphasis on levamisole resistance in Ostertagia circumcincta. Acta Vet. Scand. 1991, 32, 145–154.
[PubMed]
4. Entrocasso, C.; Alvarez, L.; Manazza, J.; Lifschitz, A.; Borda, B.; Virkel, G.; Mottier, L.; Lanusse, C. Clinical
efficacy assessment of the albendazole-ivermectin combination in lambs parasited with resistant nematodes.
Vet. Parasitol. 2008, 155, 249–256. [CrossRef] [PubMed]
5. Stear, M.J.; Doligalska, M.; Donskow-Schmelter, K. Alternatives to anthelmintics for the control of nematodes
in livestock. Parasitology 2007, 134, 139–151. [CrossRef] [PubMed]
6. Hein, W.R.; Shoemaker, C.B.; Heath, A.C.G. Future technologies for control of nematodes of sheep. N. Z.
Vet. J. 2001, 49, 247–251. [CrossRef] [PubMed]
7. Knox, D.P. Technological advances and genomics in metazoan parasites. Int. J. Parasitol. 2004, 34, 139–152.
[CrossRef] [PubMed]
8. Sayers, G.; Sweeney, T. Gastrointestinal nematode infection in sheep—A review of the alternatives to
anthelmintics in parasite control. Anim. Health Res. Rev. 2005, 6, 159–171. [CrossRef] [PubMed]
9. Waller, P.J.; Thamsborg, S.M. Nematode control in ‘green’ ruminant production systems. Trends Parasitol.
2004, 20, 493–497. [CrossRef] [PubMed]

201
Mathematics 2018, 6, 143

10. Smith, G.; Grenfell, B.T.; Isham, V.; Cornell, S. Anthelmintic resistance revisited: Under-dosing, chemoprophylactic
strategies, and mating probabilities. Int. J. Parasitol. 1999, 29, 77–91. [CrossRef]
11. Praslička, J.; Bjørn, H.; Várady, M.; Nansen, P.; Hennessy, D.R.; Talvik, H. An in vivo dose-response
study of fenbendazole against Oesophagostomum dentatum and Oesophagostomum quadrispinulatum in pigs.
Int. J. Parasitol. 1997, 27, 403–409. [CrossRef]
12. Coles, G.C.; Roush, R.T. Slowing the spread of anthelmintic resistant nematodes of sheep and goats in the
United Kingdom. Vet. Res. 1992, 130, 505–510. [CrossRef]
13. Prichard, R.K.; Hall, C.A.; Kelly, J.D.; Martin, I.C.A.; Donald, A.D. The problem of anthelmintic resistance in
nematodes. Aust. Vet. J. 1980, 56, 239–250. [CrossRef] [PubMed]
14. Anderson, R.M.; May, R.M. Infectious Diseases of Humans: Dynamics and Control; Oxford University Press:
Oxford, UK, 1992.
15. Marion, G.; Renshaw, E.; Gibson, G. Stochastic effects in a model of nematode infection in ruminants. IMA J.
Math. Appl. Med. Biol. 1998, 15, 97–116. [CrossRef] [PubMed]
16. Cornell, S.J.; Isham, V.S.; Grenfell, B.T. Stochastic and spatial dynamics of nematode parasites in farmed
ruminants. Proc. R. Soc. B Biol. Sci. 2004, 271, 1243–1250. [CrossRef] [PubMed]
17. Roberts, M.G.; Grenfell, B.T. The population dynamics of nematode infections of ruminants: Periodic
perturbations as a model for management. IMA J. Math. Appl. Med. Biol. 1991, 8, 83–93. [CrossRef] [PubMed]
18. Roberts, M.G.; Grenfell, B.T. The population dynamics of nematode infections of ruminants: The effect of
seasonality in the free-living stages. IMA J. Math. Appl. Med. Biol. 1992, 9, 29–41. [CrossRef] [PubMed]
19. Allen, L.J.S. An Introduction to Stochastic Processes with Applications to Biology; Pearson Education: Hoboken,
NJ, USA, 2003.
20. Gómez-Corral, A.; López García, M. Control strategies for a stochastic model of host-parasite interaction in a
seasonal environment. J. Theor. Biol. 2014, 354, 1–11. [CrossRef] [PubMed]
21. Abbott, K.A.; Taylor, M.; Stubbings, L.A. Sustainable Worm Control Strategies for Sheep, 4th ed.;
A Technical Manual for Veterinary Surgeons and Advisers; SCOPS: Worcestershire, UK, 2012.
Available online: http://www.scops.org.uk/workspace/pdfs/scops-technical-manual-4th-edition-updated-
september-2013.pdf (accessed on 1 June 2018).
22. Uriarte, J.; Llorente, M.M.; Valderrábano, J. Seasonal changes of gastrointestinal nematode burden in sheep
under an intensive grazing system. Vet. Parasitol. 2003, 118, 79–92. [CrossRef] [PubMed]
23. Faragó, I.; Havasi, A.; Horváth, R. On the order of operator splitting methods for time-dependent linear
systems of differential equations. Int. J. Numer. Anal. Model. Ser. B 2011, 2, 142–154.
24. Nasreen, S.; Jeelani, G.; Sheikh, F.D. Efficacy of different anthelmintics against gastro-intestinal nematodes of
sheep in Kashmir Valley. VetScan 2007, 2, 1.
25. Kassai, T. Veterinary Helminthology; Butterworth-Heinemann: Oxford, UK, 1999.
26. Barger, I.A. Genetic resistance of hosts and its influence on epidemiology. Vet. Parasitol. 1989, 32, 21–35.
[CrossRef]
27. Barger, I.A.; Le Jambre, L.F.; Georgi, J.R.; Davies, H.I. Regulation of Haemonchus contortus populations in sheep
exposed to continuous infection. Int. J. Parasitol. 1985, 15, 529–533. [CrossRef]
28. Dobson, R.J.; Waller, P.J.; Donald, A.D. Population dynamics of Trichostrongylus colubriformis in sheep:
The effect of infection rate on the establishment of infective larvae and parasite fecundity. Int. J. Parasitol.
1990, 20, 347–352. [CrossRef]
29. Bailey, J.N.; Kahn, L.P.; Walkden-Brown, S.W. Availability of gastro-intestinal nematode larvae to sheep
following winter contamination of pasture with six nematode species on the Northern Tablelands of New
South Wales. Vet. Parasitol. 2009, 160, 89–99. [CrossRef] [PubMed]
30. Valderrábano, J.; Delfa, R.; Uriarte, J. Effect of level of feed intake on the development of gastrointestinal
parasitism in growing lambs. Vet. Parasitol. 2002, 104, 327–338. [CrossRef]
31. Grennan, E.J. Lamb Growth Rate on Pasture: Effect of Grazing Management, Sward Type and Supplementation;
Teagasc Research Centre: Athenry, Ireland, 1999.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

202
 mathematics
Article
On Small Deviation Asymptotics In L2 of Some
Mixed Gaussian Processes
Alexander I. Nazarov 1,2 and Yakov Yu. Nikitin 2,3, *
1 St. Petersburg Department of the Steklov Mathematical Institute, Fontanka 27, 191023 St. Petersburg, Russia;
[email protected]
2 Saint-Petersburg State University, Universitetskaya nab. 7/9, 199034 St. Petersburg, Russia
3 National Research University, Higher School of Economics, Souza Pechatnikov 16, 190008 St. Petersburg, Russia
* Correspondence: [email protected]; Tel.: +7-921-5832100

Received: 23 March 2018; Accepted: 3 April 2018; Published: 5 April 2018

Abstract: We study the exact small deviation asymptotics with respect to the Hilbert norm for some
mixed Gaussian processes. The simplest example here is the linear combination of the Wiener process
and the Brownian bridge. We get the precise final result in this case and in some examples of more
complicated processes of similar structure. The proof is based on Karhunen–Loève expansion together
with spectral asymptotics of differential operators and complex analysis methods.

Keywords: mixed Gaussian process; small deviations; exact asymptotics

MSC: 60G15, 60J65, 62J05

1. Introduction
The problem of small deviation asymptotics for Gaussian processes was intensively studied in
last years. Such a development was stimulated by numerous links between the small deviation theory
and such mathematical problems as the accuracy of discrete approximation for random processes,
the calculation of the metric entropy for functional sets, and the law of the iterated logarithm in the
Chung form. It is also known that the small deviation theory is related to the functional data analysis
and nonparametric Bayesian estimation.
The history of the question is described in the surveys [1,2], see also [3] for recent results. The most
explored is the case of L2 -norm. For an arbitrary square-integrable random process X on [0, 1] put

⎛ ⎞ 12
1
|| X ||2 = ⎝ X (t)dt⎠ .
2

We are interested in the exact asymptotics as ε → 0 of the probability P{|| X ||2 ≤ ε}.
Usually one studies the logarithmic asymptotics while the exact asymptotics was found only for
several special processes. Most of them are so-called Green Gaussian processes. This means that the
covariance function GX is the Green function for the ordinary differential operator (ODO)
 ()  (−1)
Lu ≡ (−1) p (t)u() + p−1 (t)u(−1) + · · · + p0 (t)u, (1)

(p (t) > 0) subject to proper homogeneous boundary conditions. This class of processes contains, e.g.,
the integrated Brownian motion, the Slepian process and the Ornstein–Uhlenbeck process, see [4–10].
Notice that some strong and interesting results were obtained recently for non-Green processes by
Kleptsyna et al., see [11] and references therein.

Mathematics 2018, 6, 55; doi:10.3390/math6040055 203 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 55

In the present paper, we are interested in small deviations of so-called mixed Gaussian processes
which are the sum (or the linear combination) of two independent Gaussian processes, usually with
zero mean values. Mixed random processes arise quite naturally in the mathematical theory of finances
and engineering applications and are known long ago.
Cheredito [12] considered the linear combination of the standard Wiener process W and the
fractional Brownian motion (fBm) W H with the Hurst index H, namely the process

Y(Hβ) (t) = W (t) + βW H (t),

where β = 0 is a real constant. It is assumed that the processes W and W H are independent.
The covariance function of this process is min(s, t) + β2 GW H (s, t), where the covariance function
of the fBm is given by the well-known formula

1 2H
GW H (s, t) = (s + t2H − |s − t|2H ),
2

and H ∈ (0, 1) is the so-called Hurst index. For H = 1/2 the fBm process turns into the usual
Wiener process.
This paper strongly stimulated the probabilistic study of such process and its generalizations
concerning the regularity of its trajectories, its martingale properties, the innovation representations,
etc. The papers [13–15] are the typical examples.
The small deviations of the process Y(Hβ) were studied at the logarithmical level in [16], where the
following result was obtained. We cite it in the simplified form (without the weight function).

Proposition 1. As ε → 0 the following asymptotics holds


ln P{||W ||2 ≤ ε}, if H > 1/2;
ln P{||Y(Hβ) ||2 ≤ ε} ∼
β1/H ln P{||W H ||2 ≤ ε}, if H < 1/2.

From [17] we know that as ε → 0

( ) 1
2H
H Γ(2H + 1) sin(πH )
ln P{||W H || ≤ ε} ∼ − 2H +1   π 2H +1 ε−1/H ,
(2H + 1) 2H sin 2H +1

and the exact small deviation asymptotics of W is given below, see (3).
However, the exact small deviations of mixed processes have not been explored. In general case it
looks like a very complicated problem. First steps were made in a special case when a Gaussian process
is mixed with some finite-dimensional “perturbation”. The general theory was built in [18], later some
refined results were obtained in the case of Durbin processes (limiting processes for empirical processes
with estimated parameters), see [19] as a typical example.
We can give the solution in two cases. In Section 2 we consider the linear combination of two
processes whose covariance functions are Green functions for two different boundary value problems
to the same differential equation. The simplest example here is given by the standard Wiener process
W (t) and the Brownian bridge B(t). Also we provide the exact small deviation asymptotics for more
complicated mixtures containing the Ornstein–Uhlenbeck processes.
In Section 3 we deal with pairs of processes whose covariance functions are kernels of integral
operators which are powers (or, more general, polynomials) of the same integral operator. The basic
example here is the Brownian bridge and the integrated centered Wiener process
 t
 1 
W (t) = W (s) − W (u)du ds. (2)
0 0

204
Mathematics 2018, 6, 55

Another series of examples is given by the Wiener process and the so-called Euler integrated
Wiener process.

2. Mixed Green Processes Related to the Same Ordinary Differential Operator (ODO)
Let X1 and X2 be independent zero mean Gaussian processes on [0, 1]. We assume that their
covariance functions G1 (s, t) and G2 (s, t) are the Green functions for the same ODO (1) with different
boundary conditions. This means they satisfy the equation

L Gi (s, t) = δ(s − t), i = 1, 2.

in the sense of distributions and satisfy corresponding boundary conditions.

We consider the mixed process

Z β (t) = X1 (t) + βX2 (t), t ∈ [0, 1].

Since X1 and X2 are independent, it is easy to see that its covariance function equals

GZ β (s, t) = G1 (s, t) + β2 G2 (s, t)

and satisfies the equation

L GZ β (s, t) = (1 + β2 )δ(s − t)
in the sense of distributions. Therefore, it is the Green function for the ODO 1+1β2 L subject to some
(in general, more complicated) boundary conditions. This allows us to apply general results of [6,8] on
the small ball behavior of the Green Gaussian processes and to obtain the asymptotics of P{|| Z β ||2 ≤ ε}
as ε → 0 up to a constant. Then the sharp constant can be found by the complex variable method as
shown in [7], see also [20].
To illustrate this algorithm we begin with the simplest mixed process

β
Z1 (t) = B(t) + βW (t), t ∈ [0, 1].

The covariance function G β is given by (1 + β2 ) min(s, t) − st, and the integral equation for
Z1
eigenvalues is equivalent to the boundary value problem

1 + β2
− f  (t) = f ( t ), f (0) = 0, f (1) + β2 f  (1) = 0.
λ
β
It is easy to see that the process √ 1 Z1 (t) coincides in distribution with the process W ( β) ,
1+ β2
so-called “elongated” Brownian bridge from zero to zero with length 1 + β2 , see ([21], Section 4.4.20).
Therefore, we obtain, as ε → 0,
! ε " (∗) 1 + β2 ! ε "
β
P{|| Z1 ||2 ≤ ε} = P ||W ( β) ||2 ≤  ∼ · P ||W ||2 ≤ 
1 + β2 | β| 1 + β2

(the relation (∗) was derived in ([7], Proposition 1.9), see also ([18], Example 6)).
The last asymptotics was obtained long ago:

4  1 
P{||W ||2 ≤ ε} ∼ √ · ε · exp − ε−2 , ε → 0, (3)
π 8

and we arrive at the following Theorem:

Theorem 1. The following asymptotic relation holds as ε → 0:

205
Mathematics 2018, 6, 55

4 ε  1 + β2 
P{|| B + βW ||2 ≤ ε} ∼ √ · · exp − ε −2 .
π | β| 8

The next process we consider is

β
Z2 (t) = Ů(α) (t) + βU(α) (t), t ∈ [0, 1].

Here Ů(α) (t) is the Ornstein–Uhlenbeck process starting at the origin and U(α) (t) is the stationary
Ornstein–Uhlenbeck process. Both them are Gaussian processes with zero mean-value. Their covariance
functions are, respectively,

GŮ (s, t) = (e−α|t−s| − e−α(t+s) )/(2α); GU(α) (s, t) = e−α|t−s| /(2α).

(α)

β
Direct calculation shows that the integral equation for eigenvalues of Z2 is equivalent to the
boundary value problem

1 + β2  
− f  (t) + α2 f (t) = f ( t ); f  − α(1 + 2β−2 ) f (0) = ( f  + α f )(1) = 0.
λ
By standard method we derive that if r1 < r2 < ... are the positive roots of transcendental equation

sin(ζ )
F1 (ζ ) := (ζ 2 − α2 (1 + 2β−2 )) − 2α(1 + β−2 ) cos(ζ ) = 0
ζ

β 1+ β2
then λn ( Z2 ) = r2 +α2 , n ≥ 1.
n
Recall that the eigenvalues of the stationary Ornstein–Uhlenbeck process were derived in [22].
 1+ β2
By rescaling we obtain λn ( 1 + β2 U(α) ) = ρ2 +α2 , n ≥ 1, where ρ1 < ρ2 < ... are the positive roots of
n
transcendental equation
sin(ζ )
F2 (ζ ) := (ζ 2 − α2 ) − 2α cos(ζ ) = 0.
ζ
β 
We claim that λn ( Z2 ) and λn ( 1 + β2 U(α) ) are asymptotically close, and therefore, using the
Wenbo Li comparison theorem, see [22], we can write
! ε "
β
P{|| Z2 ||2 ≤ ε} ∼ Cdist · P ||U(α) ||2 ≤  , ε → 0, (4)
1 + β2

where the distortion constant is given by

(  ) 12 ( )1
∞ λn ( 1 + β2 U(α) ) ∞ 2
r 2 + α2
Cdist = ∏ β
= ∏ ρn2 + α2 .
n =1 λn ( Z2 ) n =1 n

To justify (4) we should prove the convergence of the last infinite product. As in [7], we use the
complex variable method.
For large N in the disk |ζ | < π ( N − 12 ) there are exactly 2N zeros ±r j , j = 1, ..., N, of F1 (ζ ),
and exactly 2N zeros ±ρ j , j = 1, ..., N, of F2 (ζ ). By the Jensen theorem, see ([23], Section 3.6.1), we have
( ) 2π
| F1 (0)| N ρ2n 1 | F1 (π ( N − 12 ) exp(iϕ))|
| F2 (0)| n∏
ln · = ln dϕ.
2
=1 r n 2π | F2 (π ( N − 12 ) exp(iϕ))|
0

It is easy to see that if we take |ζ | = π ( N − 12 ) then

206
Mathematics 2018, 6, 55

| F1 (ζ )|
⇒ 1, N → ∞.
| F2 (ζ )|

Therefore,
∞
r2 | F1 (0)|
∏ ρn2 =
| F2 (0)|
. (5)
n =1 n

Now we use Hadamard’s theorem on canonical product, see ([23], Section 8.24):
∞
 ∞

ζ2 ζ2
F1 (ζ ) ≡ F1 (0) · ∏ 1−
rn2
; F2 (ζ ) ≡ F2 (0) · ∏ 1−
ρ2n
.
n =1 n =1

In view of (5) this gives

∞
/ ∞

rn2 + α2 | F1 (0)| ∞ α2 α2 | F (iα)|
2
Cdist =∏ 2 = · ∏ 1+ 2 ∏ 1+ 2 = 1 = 1 + β −2 .
ρ
n =1 n + α 2 | F2 ( 0 )| n =1 r n n =1 ρ n | F2 (iα)|

Thus, (4) is proved. Since the small deviation asymptotics of U(α) is known, see ([7], Proposition 2.1)
and ([20], Corollary 3), we obtain the following Theorem:

Theorem 2. The following asymptotic relation holds as ε → 0:

.

αeα 8ε2 1 + β2 −2
P{||Ů(α) + βU(α) ||2 ≤ ε} ∼ ·  · exp − ε .
π | β | 1 + β2 8

Finally, we consider the stationary process

β
Z3 (t) = B(α) (t) + βU(α) (t), t ∈ [0, 1],

where B(α) is the Bogoliubov periodic process ([24–26]) with zero mean and covariance function

1  α
GB(α) (s, t) = cosh α|t − s| − .
2α sinh(α/2) 2

β
A portion of tedious calculations gives the boundary value problem for eigenvalues of Z3 :

1 + β2
− f  (t) + α2 f (t) = f ( t ); f  (0) − A1 f (0) + A2 f (1) = f  (1) + A1 f (1) − A2 f (0) = 0.
λ
Here
(1 + β2 γ )2 + 1 1 + β2 γ
A1 = α ; A2 = 2α ; γ = 1 − e−α .
(1 + β2 γ )2 − 1 (1 + β2 γ )2 − 1
β 1+ β2
Just as in the previous example we obtain λn ( Z3 ) = r2n +α2
, n ≥ 1, where r1 < r2 < ... are the
positive roots of transcendental equation

sin(ζ )
F3 (ζ ) := (ζ 2 − α2 ) − 2A1 cos(ζ ) + 2A2 = 0.
ζ

Arguing as before, we derive

! ε "
β
P{|| Z3 ||2 ≤ ε} ∼ Cdist · P ||U(α) ||2 ≤  , ε → 0,
1 + β2

where

207
Mathematics 2018, 6, 55

∞
2 = r2 + α 2 | F3 (iα)| γ (1 + β2 )2
Cdist ∏ ρn2 + α2 =
| F2 (iα)|
= 2
β (2 + β2 γ )
,
n =1 n

and thus we obtain the following Theorem:

Theorem 3. The following asymptotic relation holds as ε → 0:

.

α ( e α − 1) 8ε2 1 + β2 −2
P{||B(α) + βU(α) ||2 ≤ ε} ∼ ·  · exp − ε .
π | β | 2 + β2 (1 − e − α ) 8

3. Mixed Processes Related to Polynomials of Covariance Operator

Recall that the covariance operator GX related to the Gaussian process X is the integral operator
with kernel GX .

Lemma 1. Let covariance operators GX and GZ are linked by relation GZ = P (GX ), where P is a polynomial

P ( x ) : = x + a2 x 2 + · · · + a k −1 x k −1 + a k x k .

Then the following asymptotic relation holds as ε → 0:

P{|| Z ||2 ≤ ε} ∼ Cdist · P{|| X ||2 ≤ ε}.

Proof. By the Wenbo Li comparison theorem, we should prove that the following infinite
product converges:
∞
2 = ∏ λn ( X ) .
Cdist
λ (Z)
n =1 n

It is well known that the set of eigenvalues of P (GX ) coincides with the set P ({λn ( X )}n∈N ).
Moreover, since P increases in a neighborhood of the origin, for sufficiently large n we have just
λn ( Z ) = P (λn ( X )). Thus,
∞ ∞
2 = λn ( X )
Cdist ∏ P (λn (X )) = ∏ (1 + O(λn (X ))).
n =1 n =1

Since X is square integrable, the series ∑ λn ( X ) converges. Therefore, the infinite product also
n
converges, and the lemma follows. 
The first example is the mixed process

β
Z4 (t) = B(t) + βW (t), t ∈ [0, 1],

where the integrated centered Wiener process W is defined in (2).

The integral equation for the eigenvalues of W is equivalent to the boundary value problem [4]

1
y( IV ) = y, y(0) = y(1) = y (0) = y (1) = 0. (6)
λ
It is easy to see that the operator of the problem (6) is just the square of the operator of the
boundary value problem
1
− y = y, y(0) = y(1) = 0, (7)
λ
which corresponds to the Brownian bridge. Therefore, we have the relation GW = G2B (surely, this can
be checked directly). Thus,
G β = GB + β2 G2B .
Z4

208
Mathematics 2018, 6, 55

Therefore, we can apply Lemma 1. Since the small ball asymptotics for the Brownian bridge was
obtained long ago .
8  1 
P{|| B||2 ≤ ε} ∼ · exp − ε−2 ,
π 8
it remains to calculate
∞ ∞
2 = λn ( B) (πn)2
Cdist ∏ λ n ( B ) + β2 λ2 ( B ) = ∏ (πn)2 + β2 .
n =1 n n =1

sin(ζ )
The application of Hadamard’s theorem to the function F4 (ζ ) = ζ gives

∞  β2  F (iβ) sinh( β)
∏ 1+
(πn)2
= 4
F4 (0)
=
β
,
n =1

and we arrive at the following Theorem:

Theorem 4. The following asymptotic relation holds as ε → 0:

%
8β  1 
P{|| B + βW ||2 ≤ ε} ∼ · exp − ε−2 .
π sinh( β) 8

Now we consider a family of mixed processes (m ∈ N)

 β (t) = W (t) + βW2m

E β E
Z2m ( t ); Z2m−1 ( t ) = W (1 − t ) + βW2m−1 ( t ), t ∈ [0, 1],

where WmE is so-called Euler integrated Brownian motion, see [27,28]:

 1  t
W0E (t) = W (t); E
W2m −1 ( t ) =
E
W2m −1 ( s ) ds;
E
W2m (t) = E
W2m −1 ( s ) ds.
t 0

It was shown in [28], see also ([6], Proposition 5.1), that the covariance operator of WmE can be
expressed as
GW E = GW 2m+1
; GW E = GW 2m
 ,
2m 2m−1

 (t) = W (1 − t). Obviously, the small ball asymptotics for W

where W  and for W coincide.
Thus, we can apply Lemma 1, and it remains to calculate

∞ ∞ (π (n − 1 ))2k
2 = λ n (W )
Cdist ∏λ β2 λnk+1 (W )
= ∏ (π (n − 1 ))22k + β2
n =1 n (W ) + n =1 2

(here k = 2m or k = 2m − 1).
Application of Hadamard’s theorem to the function cos(ζ ) gives
∞  ζ2 
∏ 1−
(π (n − 2 ))
1 2
= cos(ζ ). (8)
n =1

 iπ  1 1 1
Put z = exp 2k and multiply relations (8) for ζ = β k z, ζ = β k z3 ,. . . , ζ = β k z2k−1 . This gives
( ) −1
k  
2 =
C ∏ cos
1
β z
k 2j−1
.
dist
j =1

We take into account that

209
Mathematics 2018, 6, 55

 1   1   1 
cos β k z2j−1 · cos β k z2k−2j+1 = | cos β k z2j−1 |2
 1   1 
= sinh2 β k sin( π (2j2k−1) ) + cos2 β k cos( π (2j2k−1) )

and obtain the following Theorem:

Theorem 5. For m ∈ N, the following asymptotic relations hold as ε → 0:

E 4
P{||W + βW2m ||2 ≤ ε} ∼ √
π
1  1 
×% · ε · exp − ε−2 ;
m   1 π (2j−1)   1 π (2j−1)  8
∏ sinh2 β 2m sin( 4m ) + cos2 β 2m cos( 4m )
j =1

 + βW E 4
P{||W 2m−1 ||2 ≤ ε } ∼  1 
π cosh β 2m−1
1  1 
×% · ε · exp − ε−2 .
m −1   1 π (2j−1)   1 π (2j−1)  8
∏ sinh2 β 2m−1 sin( 4m−2 ) + cos2 β 2m−1 cos( 4m−2 )
j =1

4. Discussion
We have initiated the study of the complicated problem of exact small deviations asymptotics
in L2 for mixed Gaussian processes with independent components. After the survey of the problem,
we consider the linear combination of two processes whose covariance functions are Green functions
for two different boundary value problems to the same differential equation. The simplest example here
is given by the standard Wiener process W (t) and the Brownian bridge B(t). Also we provide the exact
small deviation asymptotics for more complicated mixtures containing the Ornstein–Uhlenbeck processes.
Next, we deal with pairs of processes whose covariance functions are kernels of integral operators
which are powers (or, more general, polynomials) of of the same integral operator. The basic example
here is the Brownian bridge and the integrated centered Wiener process
 t
 1 
W (t) = W (s) − W (u)du ds.
0 0

Another series of examples is given by the Wiener process and the so-called Euler integrated
Wiener process.
It would be interesting to understand the genesis of boundary conditions and integral operators
in the more general cases of mixed processes.

Acknowledgments: This work was supported by the grant of RFBR 16-01-00258 and by the grant SPbGU-DFG
6.65.37.2017.
Author Contributions: Both authors contributed equally in the writing of this article.
Conflicts of Interest: The authors declare that there is no conflict of interests.

References
1. Li, W.V.; Shao, Q.M. Gaussian processes: Inequalities, Small Ball Probabilities and Applications. In Stochastic
Processes: Theory and Methods, Handbook of Statistics; Rao, C.R., Shanbhag, D., Eds.; Elsevier: Amsterdam,
The Netherlands, 2001; Volume 19, pp. 533–597.
2. Lifshits, M.A. Asymptotic behavior of small ball probabilities. In Probability Theory and Mathematical Statistics,
Proceedings of the Seventh Vilnius Conference (1998), Vilnius, Lithuania, 12–18 August 1998; Grigelionis, B., Ed.;
VSP/TEV: Rancho Cordova, CA, USA, 1999; pp. 453–468.

210
Mathematics 2018, 6, 55

3. Lifshits, M.A. Bibliography of Small Deviation Probabilities, on the Small Deviation Website. Available
online: http://www.proba.jussieu.fr/pageperso/smalldev/biblio.pdf (accessed on 22 March 2018).
4. Beghin, L.; Nikitin, Y.Y.; Orsingher, E. Exact small ball constants for some Gaussian processes under the
L2 -norm. J. Mathem. Sci. 2005, 128, 2493–2502.
5. Kharinski, P.A.; Nikitin, Y.Y. Sharp small deviation asymptotics in L2 -norm for a class of Gaussian processes.
J. Math. Sci. 2006, 133, 1328–1332.
6. Nazarov, A.I.; Nikitin, Y.Y. Exact small ball behavior of integrated Gaussian processes under L2 -norm and
spectral asymptotics of boundary value problems. Probab. Theory Relat. Fields 2004, 129, 469–494.
7. Nazarov, A.I. On the sharp constant in the small ball asymptotics of some Gaussian processes under L2 -norm.
J. Math. Sci. 2003, 117, 4185–4210.
8. Nazarov, A.I. Exact L2 -small ball asymptotics of Gaussian processes and the spectrum of boundary-value
problems. J. Theor. Prob. 2009, 22, 640–665.
9. Nikitin, Y.Y.; Orsingher, E. Exact small deviation asymptotics for the Slepian and Watson processes in the
Hilbert norm. J. Math. Sci. 2006, 137, 4555–4560.
10. Nikitin, Y.Y.; Pusev, R.S. Exact small deviation asymptotics for some Brownian functionals. Theor. Probab. Appl.
2013, 57, 60–81.
11. Chigansky, P.; Kleptsyna, M.; Marushkevych, D. Exact spectral asymptotics of fractional processes. arXiv
2018, arXiv:1802.09045. Available online: https://arxiv.org/abs/1802.09045 (accessed on 22 March 2018).
12. Cheridito, P. Mixed fractional Brownian motion. Bernoulli 2001, 7, 913–934.
13. El-Nouty, C. The fractional mixed fractional Brownian motion. Stat. Probab. Lett. 2003, 65, 111–120.
14. Cai, C.; Chigansky, P.; Kleptsyna, M. Mixed Gaussian processes: A filtering approach. Ann. Prob. 2016, 44,
3032–3075.
15. Yor, M. A Gaussian martingale which is the sum of two independent Gaussian non-semimartingales.
Electron. Commun. Probab. 2015, 20, 1–5.
16. Nazarov, A.I.; Nikitin, Y.Y. Logarithmic L2-small ball asymptotics for some fractional Gaussian processes.
Theor. Probab. Appl. 2005, 49, 645–658.
17. Bronski, J.C. Small ball constants and tight eigenvalue asymptotics for fractional Brownian motions.
J. Theor. Probab. 2003, 16, 87–100.
18. Nazarov, A.I. On a set of transformations of Gaussian random functions. Theor. Probab. Appl. 2010, 54, 203–216.
19. Nazarov, A.I.; Petrova, Y.P. The small ball asymptotics in Hilbertian norm for the Kac–Kiefer–Wolfowitz
processes. Theor. Probab. Appl. 2016, 60, 460–480.
20. Gao, F.; Hannig, J.; Lee, T.-Y.; Torcaso, F. Laplace transforms via Hadamard factorization with applications to
Small Ball probabilities. Electr. J. Probab. 2003, 8, 1–20.
21. Borodin, A.N.; Salminen, P. Handbook of Brownian Motion: Facts and Formulae; Birkhäuser: Basel, Switzerland,
1996.
22. Li, W.V. Comparison results for the lower tail of Gaussian seminorms. J. Theor. Probab. 1992, 5, 1–31.
23. Titchmarsh, E. Theory of Functions, 2nd ed.; Oxford University Press: Oxford, UK, 1975.
24. Sankovich, D.P. Gaussian functional integrals and Gibbs equilibrium averages. Theor. Math. Phys. 1999, 119,
670–675.
25. Sankovich, D.P. Some properties of functional integrals with respect to the Bogoliubov measure.
Theor. Math. Phys. 2001, 126, 121–135.
26. Pusev, R.S. Asymptotics of small deviations of the Bogoliubov processes with respect to a quadratic norm.
Theor. Math. Phys. 2010, 165, 1348–1357.
27. Chang, C.-H.; Ha, C.-W. The Greens functions of some boundary value problems via the Bernoulli and Euler
polynomials. Arch. Math. 2001, 76, 360–365.
28. Gao, F.; Hannig, J.; Lee, T.-Y.; Torcaso, F. Integrated Brownian motions and exact L2 -small balls. Ann. Probab.
2003, 31, 1320–1337.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

211
 mathematics
Article
Convergence in Total Variation to a Mixture of
Gaussian Laws
Luca Pratelli 1 and Pietro Rigo 2, *
1 Accademia Navale, Viale Italia 72, 57100 Livorno, Italy; [email protected]
2 Dipartimento di Matematica “F. Casorati”, Universita’ di Pavia, via Ferrata 1, 27100 Pavia, Italy
* Correspondence: [email protected]

Received: 29 April 2018; Accepted: 5 June 2018; Published: 11 June 2018

dist
Abstract: It is not unusual that Xn −→ VZ where Xn , V, Z are real random ! variables, "V is
 
independent of Z and Z ∼ N (0, 1). An intriguing feature is that P VZ ∈ A = E N (0, V 2 )( A) for
each Borel set A ⊂ R, namely, the probability distribution of the limit VZ is a mixture of centered
Gaussian laws with (random) variance V 2 . In this paper, conditions for d TV ( Xn , VZ ) → 0 are given,
where d TV ( Xn , VZ ) is the total variation distance between the probability distributions of Xn and VZ.
To estimate the rate of convergence, a few upper bounds for d TV ( Xn , VZ ) are given as well. Special
attention is paid to the following two cases: (i) Xn is a linear combination of the squares of Gaussian
random variables; and (ii) Xn is related to the weighted quadratic variations of two independent
Brownian motions.

Keywords: mixture of Gaussian laws; rate of convergence; total variation distance; Wasserstein
distance; weighted quadratic variation

MSC: 60B10; 60F05

1. Introduction
All random elements involved in the sequel are defined on a common probability space (Ω, F , P).
We let B denote the Borel σ-field on R and N ( a, b) the Gaussian law on B with mean a and variance b,
where a ∈ R, b ≥ 0, and N ( a, 0) = δa . Moreover, Z always denotes a real random variable such that:

Z ∼ N (0, 1).

In plenty of frameworks, it happens that:

dist
Xn −→ VZ, (1)

where Xn and V are real random variables and V is independent of Z. Condition (1) actually occurs
in the CLT, both in its classical form (with V = 1) and in its exchangeable and martingale versions
(Examples 3 and 4). In addition, condition (1) arises in several recent papers with various distributions
for V. See, e.g., [1–8].
An intriguing feature of condition (1) is that the probability distribution of the limit:
  
P VZ ∈ A = N (0, V 2 )( A) dP, A ∈ B,

is a mixture of centered Gaussian laws with (random) variance V 2 . Moreover, condition (1) can be
often strengthened into:

Mathematics 2018, 6, 99; doi:10.3390/math6060099 212 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 99

dW ( Xn , VZ ) → 0, (2)

where dW ( Xn , VZ ) is the Wasserstein distance between the probability distributions of Xn and VZ.
In fact, condition (2) amounts to (1) provided the sequence ( Xn ) is uniformly integrable; see Section 2.1.
A few (engaging) problems are suggested by conditions (1) and (2). One is:

(*) Give conditions for d TV ( Xn , VZ ) → 0, where

d TV ( Xn , VZ ) = sup | P( Xn ∈ A) − P(VZ ∈ A)|.

A∈B

Under such (or stronger) conditions, estimate the rate of convergence, i.e., find quantitative
bounds for d TV ( Xn , VZ ).

Problem (*) is addressed in this paper. Before turning to results, however, we mention an example.

Example 1. Let B be a fractional Brownian motion with Hurst parameter H and

 1
n 1+ H
Xn = tn−1 ( B12 − Bt2 ) dt.
2 0

The asymptotics of Xn and other analogous functionals of the B-paths (such as weighted power variations)
is investigated in various papers. See, e.g., [5,7–10] and references therein. We note also that:
 1
Xn H
tn Bt dBt = − for each H ≥ 1/4,
0 nH 2H + n

where the stochastic integral is meant in Skorohod’s

 sense (it reduces to an Ito integral if H = 1/2).

Let a( H ) = 1/2 − |1/2 − H | and V = H Γ(2H ) B1 ∼ N 0, H Γ(2H ) . In [8], it is shown that, for
every β ∈ (0, 1), there is a constant k (depending on H and β only) such that:

d TV ( Xn , VZ ) ≤ k n− β a( H ) for all n ≥ 1,

where Z is a standard normal random variable independent of V. Furthermore, the rate n− β a( H ) is quite close to
be optimal; see condition (2) of [8].

In Example 1, problem (*) admits a reasonable solution. In fact, in a sense, Example 1 is our
motivating example.
This paper includes two main results.

The first (Theorem 1) is of the general type. Suppose ln := |t φn (t)| dt < ∞, where φn is
the characteristic function of Xn . (In particular, Xn has an absolutely continuous distribution). Then,
an upper bound for d TV ( Xn , VZ ) is provided in terms of ln and dW ( Xn , VZ ). In some cases, this bound
allows to prove d TV ( Xn , VZ ) → 0 and to estimate the convergence rate. In Example 5, for instance,
such a bound improves on the existing ones; see Theorem 3.1 of [6] and Remark 3.5 of [7]. However,
for the upper bound to work, one needs information on ln and dW ( Xn , VZ ), which is not always
available. Thus, it is convenient to have some further tools.
In the second result (Theorem 2), the ideas underlying Example 1 are adapted to weighted
quadratic variations; see [5,8,9]. Let B and B be independent standard Brownian motions and

n −1  * +
 
Xn = n1/2 ∑ f Bk/n − Bk/n (ΔBk/n )2 − (ΔBk/n )2 ,
k =0


where f : R → R is a suitable function, ΔBk/n = B(k+1)/n − Bk/n and ΔBk/n = B( k+1)/n − Bk/n
 . Under

some assumptions on f (weaker than those usually requested in similar problems), it is shown that

213
Mathematics 2018, 6, 99

 1 √ 
d TV ( Xn , VZ ) =O(n−1/4 ), where V = 2 0 f
2 2Bt dt. Furthermore, d TV ( Xn , VZ ) =O(n−1/2 ) if
one also assumes inf| f | > 0. (We recall that, if an and bn are non-negative numbers, the notation
an = O(bn ) means that there is a constant c such that an ≤ c bn for all n).

2. Preliminaries

2.1. Distances between Probability Measures

In this subsection, we recall a few known facts on distances between probability measures.
We denote by (S, E ) a measurable space and by μ and ν two probability measures on E .
The total variation distance between μ and ν is:

μ − ν = sup |μ( A) − ν( A)|.

A∈E

If X and Y are (S, E )-valued random variables, we also write:

d TV ( X, Y ) =  P( X ∈ ·) − P(Y ∈ ·) = sup| P( X ∈ A) − P(Y ∈ A)|

A∈E

to denote the total variation distance between the probability distributions of X and Y.
Next, suppose S is a separable metric space, E the Borel σ-field and
 
d( x, x0 ) μ(dx ) + d( x, x0 ) ν(dx ) < ∞ for some x0 ∈ S,

where d is the distance on S. The Wasserstein distance between μ and ν is:

W (μ, ν) = inf E[d( X, Y )],

X ∼μ,Y ∼ν

where inf is over the pairs ( X, Y ) of (S, E )-valued random variables such that X ∼ μ and Y ∼ ν. By a
duality theorem, W (μ, ν) admits the representation:
 

W (μ, ν) = sup f dμ − f dν ,
f

where sup is over those functions f : S → R such that | f ( x ) − f (y)| ≤ d( x, y) for all x, y ∈ S; see, e.g.,
Section 11.8 of [11]. Again, if X and Y are (S, E )-valued random variables, we write:
# $
dW ( X, Y ) = W P( X ∈ ·), P(Y ∈ ·)

to mean the Wasserstein distance between the probability distributions of X and Y.

Finally, we make precise the connections between convergence in distribution and convergence
according to Wasserstein distance in the case S = R. Let Xn and X be real random variables such that
E| Xn | + E| X | < ∞ for each n. Then, the following statements are equivalent:

- limn dW ( Xn , X ) = 0;
dist
- Xn −→ X and E| Xn | → E| X |;
dist
- Xn −→ X and the sequence ( Xn ) is uniformly integrable.

2.2. Two Technical Lemmas

The following simple lemma is fundamental for our purposes.

214
Mathematics 2018, 6, 99

Lemma 1. If a1 , a2 ∈ R, 0 ≤ b1 ≤ b2 and b2 > 0, then:

%
b1 | a − a2 |
N ( a1 , b1 ) − N ( a2 , b2 ) ≤ 1 − + √1 .
b2 2 π b2

Lemma 1 is well known; see e.g. Proposition 3.6.1 of [12] and Lemma 3 of [8].
Note also that, if a1 = a2 = a, Lemma 1 yields:

|b1 − b2 |
N ( a, b1 ) − N ( a, b2 ) ≤ for each i such that bi > 0.
bi

The next result, needed in Section 4, is just a consequence of Lemma 1. In such a result, X and
Y are separable metric spaces, gn : X × Y → R and g : X × Y → R Borel functions, and X and Y
random variables with values in X and Y , respectively.

Lemma 2. Let ν be the probability distribution of Y. If X is independent of Y and

   
gn ( X, y) ∼ N 0, σn2 (y) , g( X, y) ∼ N 0, σ2 (y) , σ2 ( y ) > 0

for ν-almost all y ∈ Y , then:

  !  |σ (Y ) − σ(Y )|   |σ2 (Y ) − σ2 (Y )|  "
n
d TV gn ( X, Y ), g( X, Y ) ≤ min E , E n
.
σ (Y ) σ 2 (Y )

Proof. Since X is independent of Y,

       

d TV gn ( X, Y ), g( X, Y ) = sup P gn ( X, y) ∈ A − P g( X, y) ∈ A ν(dy)
A∈B
    
≤  P gn ( X, y) ∈ · − P g( X, y) ∈ ·  ν(dy).

Thus, since gn ( X, y) and g( X, y) have centered Gaussian laws and g( X, y) has strictly positive
variance, for ν-almost all y ∈ Y , Lemma 1 yields:
   |σ (y) − σ(y)|  |σ (Y ) − σ(Y )| 
n n
d TV gn ( X, Y ), g( X, Y ) ≤ ν(dy) = E
σ(y) σ (Y )
   |σ2 (y) − σ2 (y)|  |σ2 (Y ) − σ2 (Y )| 
and d TV gn ( X, Y ), g( X, Y ) ≤ n
ν(dy) = E n
.
σ2 ( y ) σ 2 (Y )

3. A General Result
As in Section 1, let Xn , V and Z be real random variables, with Z ∼ N (0, 1) and V independent
of Z. Since |V | Z ∼ VZ, it can be assumed V ≥ 0. We also assume E| Xn | + E|VZ | < ∞, so that we
can define:

dn = dW Xn , VZ ).

In addition, we let:

Xn = Xn + d1/2
n U,

where U is a standard normal random variable independent of ( Xn , V, Z : n ≥ 1).

215
Mathematics 2018, 6, 99

We aim to estimate d TV ( Xn , VZ ). Under some conditions, however, the latter quantity can be
replaced by d TV ( Xn , Xn ).

Lemma 3. For each α < 1/2,

|d TV ( Xn , VZ ) − d TV ( Xn , Xn )| ≤ d1/2 1/2−α

n + dn + P(V < dαn ).

In addition, if E(1/V ) < ∞, then:

* +
|d TV ( Xn , VZ ) − d TV ( Xn , Xn )| ≤ d1/2
n 1 + E(1/V ) .

Proof. The Lemma is trivially true if dn = 0. Hence, it can be assumed dn > 0. Define
Xn = VZ + d1/2
n U and note that:

|d TV ( Xn , VZ ) − d TV ( Xn , Xn )| ≤ d TV ( Xn , Xn ) + d TV ( Xn , VZ ).

For each A ∈ B ,
 
P( Xn ∈ A) = N ( Xn , dn )( A) dP and P( Xn ∈ A) = N (VZ, dn )( A) dP.

Hence, Lemma 1 yields:

  

d TV ( Xn , Xn ) = sup N ( Xn , dn )( A) − N (VZ, dn )( A) dP
A∈B

E| Xn − VZ |
≤ N ( Xn , dn ) − N (VZ, dn ) dP ≤ .
d1/2
n

On the other hand, the probability distribution of Xn can also be written as:

P( Xn ∈ A) = N (0, V 2 + dn )( A) dP.

Arguing as above, Lemma 1 implies again:


d TV ( Xn , VZ ) ≤ N (0, V 2 + dn ) − N (0, V 2 ) dP
 V   d1/2  d1/2
≤ E 1−  ≤E  n ≤ n + P (V <  )
2
V + dn 2
V + dn 

for each  > 0. Letting  = dαn with α < 1/2, it follows that:

E| Xn − VZ |
|d TV ( Xn , VZ ) − d TV ( Xn , Xn )| ≤ + d1/2
n
−α
+ P(V < dαn ). (3)
d1/2
n

Inequality (3) holds true for every joint distribution for the pair ( Xn , VZ ). In particular, inequality
(3) holds if such a joint distribution is taken to be one that realizes the Wasserstein distance, namely,
one such that E| Xn − VZ | = dn . In this case, one obtains:

|d TV ( Xn , VZ ) − d TV ( Xn , Xn )| ≤ d1/2 1/2−α

n + dn + P(V < dαn ).

Finally, if E(1/V ) < ∞, it suffices to note that:

 d1/2 
d TV ( Xn , VZ ) ≤ E  n ≤ d1/2
n E (1/V ).
2
V + dn

216
Mathematics 2018, 6, 99

For Lemma 3 to be useful, d TV ( Xn , Xn ) should be kept under control. This can be achieved under
various assumptions. One is to ask Xn to admit a Lipschitz density with respect to Lebesgue measure.

Theorem 1. Let φn be the characteristic function of Xn and

  ∞
ln = |t φn (t)| dt = 2 t |φn (t)| dt.
0

Given β ≥ 1, suppose supn E| Xn | β < ∞ and dn → 0. Then, there is a constant k, independent of n,

such that:
  β/( β+1)
d TV ( Xn , Xn ) ≤ k ln d1/2
n .

In particular,
  β/( β+1)
1/2−α
d TV ( Xn , VZ ) ≤ d1/2
n + dn + P(V < dαn ) + k ln d1/2
n

for each α < 1/2, and

* +   β/( β+1)
d TV ( Xn , VZ ) ≤ d1/2
n 1 + E(1/V ) + k ln d1/2
n if E(1/V ) < ∞.

It is worth noting that, if β = 1, the condition supn E| Xn | < ∞ follows from dn → 0. On the
dist
other hand, dn → 0 can be weakened into Xn −→ VZ whenever supn E| Xn | β < ∞ for some β > 1; see
Section 2.1.

Proof of Theorem 1. If ln = ∞, the Theorem is trivially true. Thus, it can be assumed ln < ∞.
Since φn is integrable, the probability distribution of Xn admits a density f n with respect to
Lebesgue measure. In addition,

| f n ( x ) − f n (y)| = (1/2π ) | (e−itx − e−ity ) φn (t) dt|

| x − y| ln | x − y |
≤ |t φn (t)| dt = .
2π 2π

Given t > 0, it follows that:


2 d TV ( Xn , Xn ) ≤ 2  P( Xn ∈ ·) − P( Xn + d1/2
n u ∈ ·) N (0, 1)( du )
 
= | f n ( x ) − f n ( x − d1/2
n u )| dx N (0, 1)( du )
 t 
≤ P(| Xn | > t) + P(| Xn | > t) + | f n ( x ) − f n ( x − d1/2
n u )| N (0, 1)( du ) dx.
−t

Since supn E| Xn | β < ∞ and dn → 0, one obtains:

P(| Xn | > t) + P(| Xn | > t) ≤ P(| Xn | > t) + P(| Xn | > t/2) + P(d1/2
n |U | > t/2)
β/2
d n E |U | β
≤ 2 P(| Xn | > t/2) +
(t/2) β
β/2
2 E | Xn | β + d n E |U | β k∗
≤ β
≤ β
(t/2) t

217
Mathematics 2018, 6, 99

for some constant k∗ . Hence,

 
k∗ ln d1/2 t
2 d TV ( Xn , Xn ) ≤ + n
|u| N (0, 1)(du) dx
tβ 2π −t
k∗ 1/2
ln dn
≤ β + t for each t > 0.
t π

Minimizing over t, one finally obtains:

 l d1/2  β/( β+1)

2 d TV ( Xn , Xn ) ≤ c( β) (k∗ )1/( β+1)
n n
,
π

where c( β) is a constant that depends on β only. This concludes the proof.

Theorem 1 provides upper bounds for d TV ( Xn , VZ ) in terms of ln and dn . It is connected to

Proposition 4.1 of [4], where d TV is replaced by the Kolmogorov distance.
In particular, Theorem 1 implies that d TV ( Xn , VZ ) → 0 provided V > 0 a.s. and

lim d1/2
n ln = lim dW ( Xn , VZ )
1/2
|t φn (t)| dt = 0.
n n

In addition, Theorem 1 allows to estimate the convergence rate. As an extreme example, if dn → 0,

* +
E(1/V ) < ∞ and supn ln + E| Xn | β < ∞ for all β ≥ 1, then:

d TV ( Xn , VZ ) = O(dαn ) for every α < 1/2.

We next turn to examples. In each such examples, Z is a standard normal random variable
independent of all other random elements.
√
Example 2. (Classical CLT). Let V = 1 and Xn = (1/ n) ∑in=1 ξ i , where ξ 1 , ξ 2 , . . . is an i.i.d. sequence
of real random variables such that E(ξ 1 ) = 0 and E(ξ 12 ) = 1. In this case, dn = O(n−1/2 ); see Theorem 2.1
of [13]. Suppose now that E|ξ 1 | β < ∞ for all β ≥ 1 and ξ 1 has a density f (with respect to Lebesgue measure)
 * +
such that | f  ( x )| dx < ∞. Then, supn ln + E| Xn | β < ∞ for all β ≥ 1, and Theorem 1 yields:

d TV ( Xn , Z ) = O(n−α ) for each α < 1/4.

This rate, however, is quite far from optimal. Under the present assumptions on ξ 1 , in fact,
d TV ( Xn , Z ) = O(n−1/2 ); see Theorem 1 of [14].
* +
We finally prove supn ln + E| Xn | β < ∞. It is well known that E|ξ 1 | β < ∞ for all β implies
β
supn E| Xn | < ∞ for all β. Hence, it suffices to prove supn ln < ∞. Let φ be the characteristic function of

ξ 1 and q = | f  ( x )| dx. An integration by parts yields |φ(t)| ≤ q/|t| for each t = 0. By Lemma 1.4 of [15],
one also obtains |φ(t)| ≤ 1 − (1/43)(t/q)2 for |t| < 2q (just let b = 2q and c = 1/2 in Lemma 1.4 of [15]).
√
Since φn (t) = φ(t/ n)n for each t ∈ R,
 q √n n √
|φn (t)| ≤ for |t| ≥ q n and
|t|
 t2  n √
|φn (t)| ≤ 1 − for |t| < q n.
43 q2 n

Using these inequalities, supn ln < ∞ follows from a direct calculation.

As noted above, the rate provided by Theorem 1 in the classical CLT is not optimal. While not
exciting, this fact could be expected. Indeed, Theorem 1 is a general result, applying to arbitrary Xn ,
and should not be requested to give optimal bounds in a very special case (such as the classical CLT).

218
Mathematics 2018, 6, 99

Example 3. (Exchangeable CLT). Suppose now that (ξ n ) is an exchangeable sequence of real random variables
with E(ξ 12 ) < ∞. Define

∑in=1 {ξ i − E(ξ 1 | T )}
V= E(ξ 12 | T ) − E(ξ 1 | T )2 and Xn = √ ,
n

where T is the tail σ-field of (ξ n ). By de Finetti’s theorem,

 
d TV ( Xn , VZ ) ≤ E  P( Xn ∈ · | T ) − N (0, V 2 ) .

P
Hence, d TV ( Xn , VZ ) → 0 provided  P( Xn ∈ · | T ) − N (0, V 2 ) −→ 0. As to Theorem 1, note that
dist
Xn −→ VZ (see e.g. Theorem 3.1 of [16]) and

* + !  ∑n (ξ − E(ξ | T ))  "
i =1 i 1 2
E( Xn2 ) = E E( Xn2 | T ) = n E E |T
n
= n E(V 2 /n) = E(V 2 ) < ∞.
!   "
Furthermore, ln ≤ E |t| | E eitXn | T | dt . Thus, by Theorem 1, d TV ( Xn , VZ ) → 0 whenever
!   "
E(ξ 12 | T ) > E(ξ 1 | T )2 a.s. and lim d1/2
n E |t| | E eitXn | T | dt = 0.
n

Example 4. (Martingale CLT). Let

kn
Xn = ∑ ξ n,j ,
j =1

where (ξ n,j : n ≥ 1, j = 1, . . . , k n ) is an array of real square integrable random variables and k n ↑ ∞. For each
n ≥ 1, let:

Fn,0 ⊂ Fn,1 ⊂ . . . ⊂ Fn,kn

be sub-σ-fields of F with Fn,0 = {∅, Ω}. A well known version of the CLT (see e.g. Theorem 3.2 of [17]) states
dist
that Xn −→ VZ provided:
 
(i) ξ n,j is Fn,j -measurable and E ξ n,j | Fn,j−1 = 0 a.s.;
2 −→ P P  
(ii) ∑ j ξ n,j V 2 , max j |ξ n,j | −→ 0, supn E max j ξ n,j
2 < ∞;
(iii) Fn,j ⊂ Fn+1,j .

Condition (iii) can be replaced by:

 
(iv) V is measurable with respect to the σ-field generated by N ∪ ∩n,j Fn,j where N = { A ∈ F : P( A) = 0}.

Note also that, under (i), one obtains E( Xn2 ) = ∑kj=

n
1 E ( ξ n,j ).
2

Now, in addition to (i)–(ii)–(iii) or (i)–(ii)–(iv), suppose supn ∑kj=

n
1 E ( ξ n,j ) < ∞. Then, Theorem 1 (applied
2

with β = 2) implies d TV ( Xn , VZ ) → 0 whenever V > 0 a.s. and limn d1/2

n ln = 0. Moreover,
 2/3 
d TV ( Xn , VZ ) = O ln d1/2
n if E(1/V ) + ln < ∞ for each n.

Our last example is connected to the second order Wiener chaos. We first note a simple fact as
a lemma.

219
Mathematics 2018, 6, 99

Lemma 4. Let ξ = (ξ 1 , . . . , ξ k ) be a centered Gaussian random vector. Define:

k * +
Y= ∑ aj ξ 2j − γ2j ,
j =1

where a j ∈ R and γ = (γ1 , . . . , γk ) is an independent copy of ξ. Then, the characteristic function ψ of Y can be
written as:
* 2 +    
ψ(t) = E e−t S , t ∈ R, where S = ∑ ai a j E ξi ξ j ξ i + γi ξ j + γj .
i,j

  √ √
Proof. Let σi,j = E ξ i ξ j , ξ ∗ = (ξ + γ)/ 2 and γ∗ = (ξ − γ)/ 2. Then,

(ξ ∗ , γ∗ ) ∼ (ξ, γ), Y=2 ∑ a j ξ ∗j γ∗j , S=2 ∑ ai a j σi,j ξ i∗ ξ ∗j .

j i,j

Therefore,
*  + * −2t2 ∑i,j ai a j σi,j ξ ∗ ξ ∗ + * 2 +
ψ(t) = E E eitY | ξ ∗ = E e i j = E e−t S .

Example 5. (Squares of Gaussian random variables). For each n ≥ 1, let (ξ n,1 , . . . , ξ n,kn ) be a centered
Gaussian random vector and
kn * +
Xn = ∑ an,j 2
ξ n,j − E(ξ n,j
2
) where an,j ∈ R.
j =1

Take an independent copy (γn,1 , . . . , γn,kn ) of (ξ n,1 , . . . , ξ n,kn ) and define:

kn * +
Yn = ∑ an,j 2
ξ n,j − γn,j
2
,
j =1
kn kn    
Sn = ∑ ∑ an,i an,j E ξ n,i ξ n,j ξ n,i + γn,i ξ n,j + γn,j .
i =1 j =1

   
Note that Sn is a (random) quadratic form of the covariance matrix E ξ n,i ξ n,j : 1 ≤ i, j ≤ k n . Therefore,
Sn ≥ 0.
Since |φn |2 agrees with the characteristic function of Yn , Lemma 4 yields:
* 2 +
|φn (t)|2 = E e−t Sn .

Being Sn ≥ 0, it follows that:

! 2 " ! 2 "
|φn (t)|2 = E e−t Sn 1  −4 + E e − t Sn 1  −4
{ Sn ≥ t 2 } { Sn < t 2 }
/2  −4 /2   (2−  ) 
≤ e−t + P(Sn < t 2 ) = e−t + P Sn−2− > t4+ 2
/2    (2−  )
≤ e−t + E Sn−2− t−4− 2 for all  > 0 and t > 0.

220
Mathematics 2018, 6, 99

Hence,
 ∞  ∞
ln
= t |φn (t)| dt ≤ 1 + t |φn (t)| dt
2 0 1
 ∞    ∞ − 1−  (2−  )
−t/2
≤ 1+ te 2 dt + E Sn−2− t 4 dt,
1 1
 −2− 

so that supn ln < ∞ whenever supn E Sn < ∞ for some  ∈ (0, 2).
To summarize, applying Theorem 1 with β = 2, one obtains:

d TV ( Xn , VZ ) = O(d1/3
n ) (4)

dist
provided Xn −→ VZ, for some V independent of Z, and
*   +
E(1/V ) + sup E Sn−2− + E( Xn2 ) < ∞ for some  ∈ (0, 2).
n

The bound (4) requires strong conditions, which may be not easily verifiable in real problems. However,
the above result is sometimes helpful, possibly in connection with the martingale CLT of Example 4. As an
example, the conditions for (4) are not hard to be checked when ξ n,1 , . . . , ξ n,kn are independent for fixed n. We
also note that, to our knowledge, the bound (4) improves on the existing ones. In fact, letting p = 2 in Theorem
3.1 of [6] (see also Remark 3.5 of [7]) one only obtains d TV ( Xn , VZ ) = O(d1/5 n ).

4. Weighted Quadratic Variations

Theorem 1 works nicely if one is able to estimate dn and ln , which is usually quite hard. Thus, it is
convenient to have some further tools. In this section, d TV ( Xn , VZ ) is upper bounded via Lemma 2.
We focus on a special case, but the underlying ideas are easily adapted to more general situations.
The results in [8], for instance, arise from a version of such ideas.
For any function x : [0, 1] → R, denote:

Δx (k/n) = x ((k + 1)/n) − x (k/n) where n ≥ 1 and k = 0, 1, . . . , n − 1.

Let q ≥ 2 be an integer, f : R → R a Borel function, and J = { Jt : 0 ≤ t ≤ 1} a real process.

The weighted q-variation of J on {0, 1/n, 2/n, . . . , 1} is:

n −1  q
Jn∗ = ∑ f ( Jk/n ) ΔJk/n .
k =0

As noted in [5], to fix the asymptotic behavior of Jn∗ is useful to determine the rate of convergence
of some approximation schemes of stochastic differential equations driven by J. Moreover, the study
of Jn∗ is also motivated by parameter estimation and by the analysis of single-path behaviour of J.
See [5,9,18–21] and references therein.
More generally, given an R2 -valued process:

( I, J ) = {( It , Jt ) : 0 ≤ t ≤ 1},

one could define:

n −1  q
( I, J )∗n = ∑ f ( Ik/n ) ΔJk/n .
k =0
 q
The weight f ( Ik/n ) of ΔJk/n depends now on I. Thus, in a sense, ( I, J )∗n can be regarded as the
weighted q-variation of J relative to I.

221
Mathematics 2018, 6, 99

Here, we focus on:

n −1  * +
 
Xn = n1/2 ∑ f Bk/n − Bk/n (ΔBk/n )2 − (ΔBk/n )2 , (5)
k =0

where B and B are independent standard Brownian motions. Note that, letting q = 2 and I = B − B ,
one obtains:
* +
Xn = n1/2 ( I, B)∗n − ( I, B )∗n .

Thus, n−1/2 Xn can be seen as the difference between the quadratic variations of B and B relative
to I = B − B .
We aim to show that, under mild assumptions on f , the probability distributions of Xn converge
in total variation to a certain mixture of Gaussian laws. We also estimate the rate of convergence.
The smoothness assumptions on f are weaker than those usually requested in similar problems;
see, e.g., [5].

Theorem 2. Let B and B be independent standard Brownian motions and Z a standard normal random variable
independent of ( B, B ). Define Xn by Equation (5) and
%
 1 √ 
V=2 f2 2Bt dt .
0

Suppose E(1/V 2 ) < ∞ and

| f ( x ) − f (y)| ≤ c | x − y| e| x|+|y|

for some constant c and all x, y ∈ R. Then, there is a constant k independent of n satisfying:

d TV ( Xn , VZ ) ≤ k n−1/4 .

Moreover, if inf| f | > 0, one also obtains d TV ( Xn , VZ ) ≤ k n−1/2 .

To understand better the spirit of Theorem 2, think of the trivial case f = 1. Then, the asymptotic
−1
* 
+
behavior of Xn = n1/2 ∑kn= 0 ( ΔBk/n ) − ( ΔBk/n )
2 2 can be deduced by classical results. In fact,

d TV ( Xn , 2Z ) = O(n−1/2 ) and this rate is optimal; see Theorem 1 of [14]. On the other hand, since
V = 2, the same conclusion can be drawn from Theorem 2.
We finally prove Theorem 2.
√ √
Proof of Theorem 2. First note that T = ( B + B )/ 2 and Y = ( B − B )/ 2 are independent standard
Brownian motions and
n −1 √ 
Xn = 2 n1/2 ∑ f 2 Yk/n ΔTk/n ΔYk/n .
k =0

Note also that:

%
 1 √ 
VZ ∼ 2 T1 f2 2Yt dt.
0

222
Mathematics 2018, 6, 99

Thus, in order to apply Lemma 2, it suffices to let X = Y = C [0, 1], X = T, and

n −1 √ 
gn ( x, y) = 2 n1/2 ∑ f 2 y(k/n) Δx (k/n) Δy(k/n),
k =0
%
 1 √ 
g( x, y) = 2 x (1) f2 2y(t) dt.
0

For fixed y ∈ Y , gn ( T, y) and g( T, y) are centered Gaussian random variables. Since

* 2 +
E ΔTk/n = 1/n,

* + n −1 √  2
σn2 (y) = E gn ( T, y)2 = 4 ∑ f2 2 y(k/n) Δy(k/n)
k =0
* +  1 √ 
and σ2 (y) = E g( T, y)2 = 4 f2 2y(t) dt.
0

On noting that σ2 (Y ) ∼ V 2 , one also obtains:

* +
σ2 (Y ) > 0 a.s. and E 1/σ2 (Y ) = E(1/V 2 ) < ∞.

Next, define:
* +
an = (1/4) E |σn2 (Y ) − σ2 (Y )|
! n−1 √
 2  1 √  "
= E ∑ f 2 2 Yk/n ΔYk/n − f2 2Yt dt .
k =0 0

By Lemma 2 and the Cauchy–Schwarz inequality,

 2  |σ (Y ) − σ(Y )| 2
n
d TV ( Xn , VZ )2 = d TV gn ( T, Y ), g( T, Y ) ≤ E
σ (Y )
* + * 2 +
≤ E 1/σ (Y ) E σn (Y ) − σ(Y )
2
* +
≤ E(1/V 2 ) E |σn2 (Y ) − σ2 (Y )| = 4 E(1/V 2 ) an .

If inf| f | > 0, since σ2 (Y ) ≥ 4 inf f 2 , Lemma 2 implies again:

 
 |σ2 (Y ) − σ2 (Y )|  E |σn2 (Y ) − σ2 (Y )| an
d TV ( Xn , VZ ) ≤ E n
≤ = .
σ 2 (Y ) 4 inf f 2 inf f 2

Thus, to conclude the proof, it suffices to show that an =O(n−1/2 ).

Define c∗ = max(c, | f (0)|) and note that:

| f (s)| ≤ c∗ e2|s| and | f (s)2 − f (t)2 | ≤ 2 c c∗ |s − t| e3(|s|+|t|) for all s, t ∈ R.

Define also:
! n−1 √
  2  "
an = E ∑ f 2 2 Yk/n
(1)
ΔYk/n − 1/n and
k =0
! n −1 √   1 √  "

∑
(2)
an = E (1/n) f2 2 Yk/n − f2 2Yt dt .
k =0 0

223
Mathematics 2018, 6, 99

Since an ≤ an + an , it suffices to see that an =O(n−1/2 ) for each i. Since Y has independent
(1) (2) (i )
* 2 2 +
increments and E ΔYk/n − 1/n = 2/n2 ,

 (1)  2
! n−1 √
  2  "2
an = E ∑ f 2 2 Yk/n ΔYk/n − 1/n
k =0
n −1 ! √   2 2 "
≤ ∑ E f4 2 Yk/n ΔYk/n − 1/n
k =0
n −1 * √ + * 2 2 +
= ∑ E f 4 2 Yk/n E ΔYk/n − 1/n
k =0
n −1 * √ +
= (2/n2 ) ∑ E f 4 2 Yk/n
k =0
* √ +
2 (c∗ )4 E e8 2M
≤ where M = sup |Yt |.
n 0≤ t ≤1

Similarly,

! n −1 √   1 √  "

∑
(2)
an = E (1/n) f2 2 Yk/n − f2 2Yt dt
k =0 0
n−1  (k +1)/n * √  √ +
≤ ∑ k/n
E |f 2
2 Yk/n − f 2 2 Yt | dt
k =0
√ n−1  (k +1)/n * √ +
≤2 2 c c∗ ∑ E |Yk/n − Yt | e6 2M dt
k =0 k/n

√ * √ + n−1  (k +1)/n * +
≤2 2 c c∗ E e12 2M
∑ E (Yk/n − Yt )2 dt
k =0 k/n
√ ∗
* √ + 1
≤2 2cc E e12 2M √ .
n

Therefore, an =O(n−1/2 ) for each i, and this concludes the proof.

(i )

Author Contributions: Each author contributed in exactly the same way to each part of this paper.
Funding: This research was supported by the Italian Ministry of Education, University and Research (MIUR):
Dipartimenti di Eccellenza Program (2018-2022) - Dept. of Mathematics “F. Casorati”, University of Pavia.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Azmoodeh, E.; Gasbarra, D. New moments criteria for convergence towards normal product/tetilla laws.
arXiv 2017, arXiv:1708.07681.
2. Eichelsbacher, P.; Thäle, C. Malliavin-Stein method for Variance-Gamma approximation on Wiener space.
Electron. J. Probab. 2015, 20, 1–28. [CrossRef]
3. Gaunt, R.E. On Stein’s method for products of normal random variables and zero bias couplings. Bernoulli
2017, 23, 3311–3345. [CrossRef]
4. Gaunt, R.E. Wasserstein and Kolmogorov error bounds for variance-gamma approximation via Stein’s
method I. arXiv 2017, arXiv:1711.07379.
5. Nourdin, I.; Nualart, D.; Tudor, C.A. Central and non-central limit theorems for weighted power variations
of fractional Brownian motion. Ann. I.H.P. 2010, 46, 1055–1079. [CrossRef]
6. Nourdin, I.; Poly, G. Convergence in total variation on Wiener chaos. Stoch. Proc. Appl. 2013, 123, 651–674.
[CrossRef]

224
Mathematics 2018, 6, 99

7. Nourdin, I.; Nualart, D.; Peccati, G. Quantitative stable limit theorems on the Wiener space. Ann. Probab.
2016, 44, 1–41. [CrossRef]
8. Pratelli, L.; Rigo, P. Total Variation Bounds for Gaussian Functionals. 2018, Submitted. Available online:
http://www-dimat.unipv.it/rigo/frac.pdf (accessed on 10 April 2018)
9. Nourdin, I.; Peccati, G. Weighted power variations of iterated Brownian motion. Electron. J. Probab. 2008, 13,
1229–1256. [CrossRef]
10. Peccati, G.; Yor, M. Four limit theorems for quadratic functionals of Brownian motion and Brownian bridge.
In Asymptotic Methods in Stochastics, AMS, Fields Institute Communication Series; Amer. Math. Soc.: Providence,
RI, USA, 2004; pp. 75–87.
11. Dudley R.M. Real Analysis and Probability; Cambridge University Press: Cambridge, UK, 2004.
12. Nourdin, I.; Peccati, G. Normal Approximations with Malliavin Calculus: From Stein’s Method to Universality;
Cambridge University Press: Cambridge, UK, 2012.
13. Goldstein, L. L1 bounds in normal approximation. Ann. Probab. 2007, 35, 1888–1930. [CrossRef]
14. Sirazhdinov, S.K.H.; Mamatov, M. On convergence in the mean for densities. Theory Probab. Appl. 1962, 7,
424–428. [CrossRef]
15. Petrov, V.V. Limit Theorems of Probability Theory: Sequences of Independent Random Variables; Clarendon Press:
Oxford, UK, 1995.
16. Berti, P.; Pratelli, L.; Rigo, P. Limit theorems for a class of identically distributed random variables.
Ann. Probab. 2004, 32, 2029–2052.
17. Hall, P.; Heyde, C.C. Martingale Limit Theory and Its Applications; Academic Press: New York, NY, USA, 1980.
18. Barndorff-Nielsen, O.E.; Graversen, S.E.; Shepard, N. Power variation and stochastic volatility: A review
and some new results. J. Appl. Probab. 2004, 44, 133–143. [CrossRef]
19. Gradinaru, M.; Nourdin, I. Milstein’s type schemes for fractional SDEs. Ann. I.H.P. 2009, 45, 1058–1098.
[CrossRef]
20. Neuenkirch, A.; Nourdin, I. Exact rate of convergence of some approximation schemes associated to SDEs
driven by a fractional Brownian motion. J. Theor. Probab. 2007, 20, 871–899. [CrossRef]
21. Nourdin, I. A simple theory for the study of SDEs driven by a fractional Brownian motion in dimension one.
In Seminaire de Probabilites; Springer: Berlin, Germany, 2008; Volume XLI, pp. 181–197.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

225
 mathematics
Article
Some Notes about Inference for the Lognormal
Diffusion Process with Exogenous Factors
Patricia Román-Román † , Juan José Serrano-Pérez † and Francisco Torres-Ruiz *,†
Departamento de Estadística e Investigación Operativa, Facultad de Ciencias, Universidad de Granada,
Avenida Fuente Nueva, 18071 Granada, Spain; [email protected] (P.R.-R.); [email protected] (J.J.S.-P.)
* Correspondence: [email protected]; Tel.: +34-9582-41000 (ext. 20056)
† These authors contributed equally to this work.

Received: 16 April 2018; Accepted: 15 May 2018; Published: 21 May 2018

Abstract: Different versions of the lognormal diffusion process with exogenous factors have been
used in recent years to model and study the behavior of phenomena following a given growth curve.
In each case considered, the estimation of the model has been addressed, generally by maximum
likelihood (ML), as has been the study of several characteristics associated with the type of curve
considered. For this process, a unified version of the ML estimation problem is presented, including
how to obtain estimation errors and asymptotic confidence intervals for parametric functions when no
explicit expression is available for the estimators of the parameters of the model. The Gompertz-type
diffusion process is used here to illustrate the application of the methodology.

Keywords: lognormal diffusion process; exogenous factors; growth curves; maximum likelihood
estimation; asymptotic distribution

1. Introduction
The lognormal diffusion process has been widely used as a probabilistic model in several
scientific fields in which the variable under consideration exhibits an exponential trend.
Originally, the lognormal diffusion process was mainly applied to modeling dynamic variables in the
field of economy and finance. Important contributions have been made in this direction by Cox and
Ross [1], Markus and Shaked [2], and Merton [3], showing the theoretical and practical importance of
the process in that environment. For example, this process is associated with the Black and Scholes
model [4] and appears in later extensions as terminal swap-rate models (Hunt and Kennedy [5],
Lamberton and Lapeyre [6]).
In 1972, Tintner and Sengupta [7] introduced a modification of the process by including a linear
combination of time functions in the infinitesimal mean of the process. The motivation for this was
the introduction of external influences on the interest variable (endogenous variable), influences that
could contribute to a better explanation of the phenomenon under study. For this reason, these time
functions are known as exogenous factors, whose time behavior is assumed to be known or partially
known. By using these time functions we can model situations wherein the observed trend shows
deviations from the theoretical shape of the trend during certain time intervals, and can therefore use
them to help describe the evolution of the process. Furthermore, a suitable choice of the exogenous
factors can contribute to the external control of the process for forecasting purposes. Note that the
methodology derived from the inclusion of exogenous factors has been applied to several contexts
other than the lognormal process (see, for example, Buonocore et al. [8]).
The lognormal diffusion process with exogenous factors has been widely studied in relation to
some aspects of inference and first-passage times. It has been applied to the modeling of time variables
in several fields (see, for example [9,10]). On occasion, the endogenous variable itself helps identify the
exogenous factors. However, there are situations in which external variables to the process that have

Mathematics 2018, 6, 85; doi:10.3390/math6050085 226 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 85

an influence on the system are not available, or situations in which their functional expressions are
unknown. In such cases, Gutiérrez et al. [11] suggested approaching the exogenous factors by means
of polynomial functions.
The ability to control the endogenous variable using exogenous factors makes this process
particularly useful for forecasting purposes. Some of its main features, such as the mean, mode and
quantile functions (that can be expressed as parametric functions of the parameters of the process),
can be used for prediction purposes. Therefore, the inference of these functions has been the subject
of considerable study, both from the perspective of point estimation and of estimation by confidence
intervals. With respect to the former, in [10] a more general study was carried out to obtain maximum
likelihood (ML) estimators. In that case, the exact distribution of the estimators was found, and then
used to obtain the uniformly minimum variance unbiased (UMVU) estimators. In addition, expressions
for the relative efficiency of ML estimators, with respect to UMVU estimators, were obtained. This last
study was extended for a class of parametric functions which include the mean and mode functions
(together with their conditional versions) as special cases. Concerning estimation by confidence bands,
in this paper the authors extended the results obtained by Land [12] on exact confidence intervals
for the mean of a lognormal distribution, thus obtaining confidence bands for the mean and mode
functions of the lognormal process with exogenous factors and expressing these functions in a more
general form.
In most of the works cited, inference has been approached from the ML point of view, considering
discrete sampling of the trajectories. To this end, it is essential to have the exact form of the transition
density functions from which the likelihood function associated with the sample is constructed.
However, alternatives are available for a range of situations. For example, approximating the transition
density function using Euler-type schemes derived from the discretization of the stochastic differential
equation that models the behavior of the phenomenon under study (sometimes this approach is known
as naive ML approach). Other possible alternatives to ML are those derived, for example, from the
use of the concept of estimating functions (Bibby et al. [13]) and the generalized method of moments
(Hansen [14]). Fuchs in [15] presents a good review of these and other procedures. The Bayesian
approach is also present in the study of diffusion processes, as suggested by Tang and Heron in [16].
On the other hand, considering particular choices of the time functions that define the exogenous
factors has enabled researchers to define diffusion processes associated to alternative expressions of
already-known growth curves. Along these lines, we may cite a Gompertz-type process [17] (applied to the
study of rabbit growth), a generalized Von Bertalanffy diffusion process [18] (with an application to the
growth of fish species), a logistic-type process [19] (applied to the growth of a microorganism culture),
and a Richards-type diffusion process [20]. In [21], a joint analysis of the procedure for obtaining
these processes is shown. More recently, Da Luz-Sant’Ana et al. [22] have established, following a
similar methodology, a Hubbert diffusion process for studying oil production, while Barrera et al. [23]
introduced a process linked to the hyperbolastic type-I curve and applied it in the context of the
quantitative polymerase chain reaction (qPCR) technique.
In these last cases, obtaining the ML estimators was a rather laborious task. In fact, the resulting
system of equations is exceedingly complex and does not have an explicit solution, and numerical
procedures must be employed instead, with the subsequent problem of finding initial solutions
(see, for instance [18,19,22]). However, it is impossible to carry out a general study of the system of
equations in order to check the conditions of convergence of the chosen numerical method, since it
is dependent on sample data. One alternative is then to use stochastic optimization procedures like
simulated annealing, variable neighborhood search, and the firefly algorithm [20,23,24]. In any case,
the exact distribution of the estimators cannot be obtained. Recently, the asymptotic distribution
of the MLestimators and delta method have been used in order to obtain estimation errors, as well
as confidence intervals, for the parameters and parametric functions in the context of the Hubbert
diffusion model [25].

227
Mathematics 2018, 6, 85

The main objective of this paper is to provide a unified view of the estimation problem by means
of discrete sampling of trajectories, and to cover all the diffusion processes mentioned above. To this
end, we will consider the generic expression of the lognormal diffusion process with exogenous factors.
In Section 2, a brief summary of the main characteristics of the process is presented. Sections 3 and 4
address the problem of estimation by ML by using discrete sampling. In Section 3, the distribution of the
sample is obtained, while in Section 4 the generic form adopted by the system of likelihood equations
is derived in terms of the exogenous factor included in the model. Section 5 deals with obtaining the
asymptotic distribution of the estimators, after calculating the Fisher information matrix, for which
the results of Section 3 are fundamental. Finally, and as an application of the previous developments,
Section 6 deals with the particular case of the Gompertz-type process introduced in [17].

2. The Lognormal Diffusion Process With Exogenous Factors

Let I = [t0 , +∞) be a real interval (t0 ≥ 0), Θ ⊆ Rk an open set, and hθ (t) a continuous, bounded
and differentiable function on I depending on θ ∈ Θ.
The univariate lognormal diffusion process with exogenous factors is a diffusion process
{ X (t); t ∈ I }, taking values on R+ , with infinitesimal moments

A1 ( x, t) = hθ (t) x
(1)
A2 ( x ) = σ 2 x 2 , σ>0

and with a lognormal or degenerate initial distribution. This process is the solution to the stochastic
differential equation

dX (t) = hθ (t) X (t)dt + σX (t)dW (t), X ( t 0 ) = X0 ,

where W (t) is a standard Wiener process independent on X0 = X (t0 ), t ≥ t0 , being this solution
 
X (t) = X0 exp Hξ (t0 , t) + σ(W (t) − W (t0 )) , t ≥ t0

with  t
σ2
Hξ (t0 , t) = hθ (u)du − ( t − t0 ), ξ = (θT , σ2 ) T .
t0 2
An explanation of the main features of the process can be found in [21], where the authors
carried out a detailed theoretical analysis. As regards the distribution of the process, if X0
# $
is distributed according to a lognormal distribution Λ1 μ0 ; σ02 , or X0 is a degenerate variable
(P[ X0 = x0 ] = 1), all the finite-dimensional distributions of the process are lognormal. Concretely,
∀n ∈ N and t1 < · · · < tn , vector ( X (t1 ), . . . , X (tn )) T has a n-dimensional lognormal distribution
Λn [ε, Σ], where the components of vector ε and matrix Σ are

ε i = μ0 + Hξ (t0 , ti ), i = 1, . . . , n

and
σij = σ02 + σ2 (min(ti , t j ) − t0 ), i, j = 1, . . . , n,

respectively. The transition probability density function can be obtained from the distribution of
( X (s), X (t)) T , s < t, being
( # $2 )
1 ln( x/y) − Hξ (s, t)
f ( x, t|y, s) =  exp − , (2)
x 2πσ2 (t − s) 2σ2 (t − s)

228
Mathematics 2018, 6, 85

that is, X (t)| X (s) = y follows a lognormal distribution

 
X (t) | X (s) = y  Λ1 ln y + Hξ (s, t), σ2 (t − s) , s < t.

From the previous distributions, one can obtain the characteristics most commonly employed for
practical fitting and forecasting purposes. These characteristics can be expressed jointly as

  λ4 
Gξλ (t|y, τ ) = Mξ (t|y, τ )λ1 exp λ2 λ3 σ02 + σ2 (t − τ ) , (3)

 
with λ = (λ1 , λ2 , λ3 , λ4 ) T and where Mξ (t|y, τ ) = exp y + Hξ (τ, t) . Table 1 includes some of these
characteristics (the n−th moment, and the mode and quantile functions as well as their conditional
versions) according to the values of λ, τ and y.

Table 1. Values used to obtain the n-th moment and the mode and quantile functions from Gξλ (t|z, τ ).
zα is the α-quantile of a standard normal distribution.

Function Expression z τ λ
n-th moment E[ X (t)n ] μ0 t0 (n, n2 /2, 1, 1) T
n-th conditional moment E[ X (t)n | X (s) = y] ln y s (n, n2 /2, 0, 1) T
mode Mode[ X (t)] μ0 t0 (1, −1, 1, 1) T
conditional mode Mode[ X (t)| X (s) = y] ln y s (1, −1, 0, 1) T
α-quantile Cα [ X (t)] μ0 t0 (1, zα , 1, 1/2) T
α-conditional quantile Cα [ X (t)| X (s) = y] ln y s (1, zα , 0, 1/2) T

3. Joint Distribution of d Sample-Paths of the Process

Let us consider a discrete sampling of the process, based on d sample paths, at times tij ,
 T
(i = 1,. . . , d, j = 1, . . . , ni ) with ti1 = t0 , i = 1, . . . , d. Denote by X = X1T | · · · |XdT the vector containing
T
the random variables of the sample, where Xi includes the variables of the i-th sample-path, that is
Xi = ( X (ti1 ), . . . , X (ti,ni )) T , i = 1, . . . , d.
From Equation (2), and if the distribution of X (t1 ) is assumed lognormal Λ1 (μ1 , σ12 ), the probability
density function of X is
(  )

 ln( xi,j+1 /xij )−mξ
i,j,j+1 2

exp i1 − μ1 ]
− [ln x2σ
2
n −1
exp − j+1,j
d 2 i 2σ2 Δi
∏ ∏
1
f X (x) = √
i =1 xi1 σ1 2π j =1 xij σ 2πΔi
j+1,j

i,j+1,j j+1,j
where mξ = Hξ (tij , ti,j+1 ) and Δi = ti,j+1 − tij .
# $T  T
Now, we consider vector V = V0T |V1T | · · · |VdT = V0T |V(T1) , built from X by means of the
following change of variables:

V0i = Xi1 , i = 1, . . . , d
j+1,j −1/2 Xi,j+1 (4)
Vij = (Δi ) ln , i = 1, . . . , d; j = 1, . . . , ni − 1.
Xij

Taking into account this change of variables, the density of V becomes

229
Mathematics 2018, 6, 85


 T  
exp − 2σ1 2 (ln v0 − μ1 1d ) T (ln v0 − μ1 1d ) exp − 2σ1 2 v(1) − γξ v(1) − γ ξ
1
f V (v) = (5)
d  d (2πσ2 ) 2
n

∏ v0i
2
2πσ12
i =1

with ln v0 = (ln v01 , . . . , ln v0d ) T , n = ∑id=1 (ni − 1). Here, 1d represents the d-dimensional vector
whose components are all equal to one, while γξ is a vector of dimension n with components
ξ j+1,j −1/2 i,j,j+1
γij = (Δi ) mξ , i = 1, . . . , d; j = 1, . . . , ni − 1.
From Equation (5) it is deduced that:

• V0 and V(1) are independents,

# $
• the distribution of V0 is lognormal Λd μ1 1d ; σ12 Id ,
# $
• V(1) is distributed as an n-variate normal distribution Nn γξ ; σ2 In ,

being Id and In the identity matrices of order d and n, respectively.

4. Maximum Likelihood Estimation of the Parameters of the Process

Consider a discrete sample of the process in the sense described in the previous section, including
the transformation of it given by Equation (4). Denote by η = (μ1 , σ12 ) T and suppose that η and ξ are
functionally independent. Then, for a fixed value v of the sample, the log-likelihood function is

d ln σ12 d ∑ [ln v0i − μ1 ]2 Z1 + Φξ − 2Γξ

(n + d) ln(2π ) n ln σ2
Lv (η, ξ ) = − − − ∑ ln v0i − i =1
− − (6)
2 2 i =1 2σ12 2 2σ2

where  
i,j+1,j 2 i,j+1,j
d n i −1 d n i −1 mξ d ni −1 vij mξ
Z1 = ∑∑ v2ij , Φξ = ∑∑ j+1,j
, Γξ = ∑∑ j+1,j 1/2
.
i =1 j =1 i =1 j =1 Δi i =1 j =1 ( Δ i )
Taking into account Equation (6), and since η and ξ are functionally independent, the ML
∂f
estimation of η is obtained from the system of equations (Given a function f : Rk → R, ∂xT =
 
∂f ∂f ∂f
∂x , . . . , ∂x . Notation ∂xT indicates that the result is a row vector).
1 k

( )
∂Lv (η, ξ ) ∂Lv (η, ξ ) ∂Lv (η, ξ )
= , =0
∂ηT ∂μ1 ∂σ12

resulting in
d d
1 1
1 =
μ
d ∑ ln v0i σ12 =
and 
d ∑ (ln v0i − μ1 )2 .
i =1 i =1

On the other hand, by denoting

i,j+1,j i,j+1,j i,j+1,j

d n i −1 m ∂mξ d n i −1 ∂mξ
1 ∂Φξ ξ 1 ∂Γξ vij
Ωξ = =∑ ∑ , Ψθ = =∑ ∑
2 ∂θT
i =1 j =1 Δ
j + 1,j ∂θT 2 ∂θT
i =1 j =1 ( Δ
j + 1,j
)1/2 ∂θ
T
i i
(7)
∂Φξ d ∂Γξ d n i −1
j+1,j 1/2
Υξ = − 2 = ∑ mξ i , Z2 = −2 2 = ∑ ∑ vij (Δi
i,n ,1
)
∂σ i =1
∂σ i =1 j =1

230
Mathematics 2018, 6, 85

we have

∂Lv (η, ξ ) 1 # $
= 2 Ψθ − Ωξ
∂θT σ
∂Lv (η, ξ ) n Z1 + Φξ − 2Γξ Z2 − Υξ
=− 2+ − .
∂σ2 2σ 2σ4 2σ2
Thus, the ML estimation of ξ is obtained as the solution of the following system of k + 1 equations:

Ψθ − Ωξ = 0 (8)
Z1 + Φξ − 2Γξ − σ2 Z2 + σ2 Υξ = nσ2 (9)

In the case where hθ is a linear function in θ, it is possible to determine an explicit solution for
this system of equations (see [10,26]). In other cases, the existence of a closed-form solution can not
be guaranteed, and it is therefore necessary to use numerical procedures for its resolution. The fact
that these methods require initial solutions has motivated the construction of ad hoc procedures which
depend on the process derived according to the function hθ considered (see [18,19,22]). However, it is
impossible to carry out a general study of the system of equations in order to check the conditions of
convergence of the chosen numerical method, since the system is dependent on sample data and this
may lead to unforeseeable behavior. One alternative would be using stochastic optimization procedures
like simulated annealing, variable neighborhood search and the firefly algorithm. These algorithms
are often more appropriate than classical numerical methods since they impose fewer restrictions on
the space of solutions and on the analytical properties of the function to be optimized. Some examples
of the application of these procedures in the context of diffusion processes can be seen in [19,21,23,25].

5. Distribution of the ML Estimators of the Parameters and Related Parametric Functions

In this section we will discuss some aspects related to the distribution of the estimators of the
parameters of the model, and their repercussions in the corresponding distributions of parametric
functions, which can be of interest for several applications.
With regard to the distribution of the estimators of η, it is immediate to verify that

d σ12
01  N1 [μ1 ; σ12 /d]
μ and  χ2d−1 .
σ12

If hθ is linear, it is then possible to calculate exact distributions associated with the estimators of ξ,
which allows us to establish confidence regions for the parameters as well as UMVU estimators and
confidence intervals for linear combinations of θ and σ2 (see [10,26]). However, in the non-linear case,
the fact that an explicit expression for the estimators of ξ is not always readily available precludes
obtaining, in general, exact distributions for them. In that case, asymptotic distributions can be used
instead. In fact, on the basis of the properties of the ML estimators, it is known that ξ is asymptotically
distributed as a normal distribution with mean ξ and covariance matrix I (ξ )−1 , where I (ξ ) is the
Fisher’s information matrix associated with the full sample (in this case, ignoring the data of the
initial distribution).
First we calculate the associated Hessian matrix: (we have adopted the usual expression for the
∂2 f
Hessian matrix of f : Rk → R using vectorial notation, that is ).
∂x∂xT

231
Mathematics 2018, 6, 85

⎛
T ⎞
∂2 Lv (η, ξ ) ∂2 Lv (η, ξ )
⎜ ⎟
∂ Lv (η, ξ )
2 ⎜ ∂θ∂θT ∂σ2 ∂θT ⎟
⎜ ⎟
H (ξ ) = =⎜ ⎟
∂ξ∂ξ T ⎜ 2 ⎟
⎝ ∂ Lv (η, ξ ) ∂ Lv (η, ξ )
2 ⎠
∂σ2 ∂θT ∂ ( σ 2 )2

⎛ 1
∂Υ  T ⎞
1  T ξ
⎜ Πξ − Ξξ − Ψ θ − Ω T
ξ + ⎟
1 ⎜ σ2 2 ∂θT ⎟
= ⎜ ⎟
σ2 ⎜ ⎟
⎝ 1 # $ 1 ∂Υξ n Z1 + Φξ − 2Γξ Z2 − Υξ Z ⎠
− Ψθ − Ωξ + − + − 3
σ2 2 ∂θT 2σ2 σ4 σ2 4
where
d n i −1
i,j+1,j
∂2 m ξ  
j+1,j −1/2 j+1,j −1/2 i,j+1,j
Πξ = ∑∑ ∂θ∂θT
( Δi ) vij − (Δi ) mξ
i =1 j =1

and ⎛ ⎞T
i,j+1,j i,j+1,j
d ni −1 ∂mξ ∂mξ d
j+1,j −1 ⎝ ⎠
∑∑ ∑ Δi i
n ,1
Ξξ = ( Δi ) , Z3 = .
i =1 j =1
∂θT ∂θT i =1

Taking into account the distribution of the sample (see Section 3), we have

E[Πξ ] = 0, E[ Z1 ] = nσ2 + Φξ , E[ Z2 ] = Υξ , E[Ψθ ] = Ωξ , E[Γξ ] = Φξ

so, the Fisher’s information matrix is given by

⎛
T ⎞
1 ∂Υξ
⎜ Ξξ − ⎟
1 ⎜ 2 ∂θT ⎟
I (ξ ) = − E[ H (ξ )] = 2 ⎜ ⎟,
σ ⎜⎝
⎟
⎠
1 ∂Υξ n Z
− + 3
2 ∂θT 2σ2 4
# $
from where it is concluded that ξ → Nk+1 ξ; I (ξ )−1 . In addition, and by applying the delta method,
D

for a q−parametric function g(ξ ) (q ≤ k + 1) it is verified that


g(ξ) → Nq g(ξ ); ∇ g(ξ ) T I (ξ )−1 ∇ g(ξ )
D

where ∇ g(ξ ) represents the vector of partial derivatives of g(ξ ) with respect to ξ.
The elements in the diagonal of matrix I (ξ )−1 provide asymptotic variances for the estimations
of the parameters, while the delta method provides the asymptotic covariance matrix for g(ξ) (and
consequently the elements of the diagonal are the asymptotic variances for the estimation of each
parametric function of g(ξ )). For example, if we consider g(ξ ) = Gξλ (t|y, τ ), that is the general
expression for the main characteristics of the process given by Equation (3), then

  λ4 −1 
∂Hξ (τ, t) λ1
∇ g ( ξ ) = g ( ξ ) λ1 , ( t − τ ) − + λ 2 λ 4 λ 3 σ0
2
+ σ 2
( t − τ ) .
∂θT 2

232
Mathematics 2018, 6, 85

6. Application: The Gompertz-Type Diffusion Process

In this section we focus on the Gompertz-type diffusion process introduced in [17] with the aim
of obtaining a continuous stochastic model associated with the Gompertz curve whose limit value
depends on the initial value. Concretely


m  −β t
f (t) = x0 exp − e − e − β t0 , t ≥ t0 ≥ 0, m, β > 0 and x0 > 0.
β

To this end, the non-homogeneous lognormal diffusion process with infinitesimal moments

A1 ( x, t) = me− β t x (10)
A2 ( x ) = σ2 x2

was considered.
In order to apply the general scheme developed in the preceding sections, we consider the
following reparameterization θ = (δ, α) T = (m/β, e− β ) T , which leads to expressing the Gompertz
curve as
  
f θ (t) = x0 exp −δ αt − αt0 (11)

whereas the infinitesimal moments (10) are written in the form of Equation (1), with hθ (t) = −δαt ln α.
α ti,j+1 α i,j+1,j α
Denoting ϕi,j +1,j = α − αti,j and ωi,j +1,j = ti,j+1 α
ti,j+1
− tij αtij , one has mξ = −δϕi,j +1,j −
σ2 j+1,j
2 Δi and
∂mξ
i,j+1,j  
α α
= − ϕi,j +1,j , δωi,j+1,j ,
∂θT
so, from Equation (8), and by taking into account of Equation (7), the following system of
equations appears

σ2 α
X1α + δX2α + X =0
2 3
σ 2
X4α + δX5α + X6α = 0
2
where  2
d ni −1 α d n i −1 α
ϕi,j
vij ϕi,j +1,j +1,j d
X1α = ∑∑ j+1,j 1/2
, X2α = ∑∑ j+1,j
, X3α = α
∑ ϕi,n i ,1
i =1 j =1 ( Δ i ) i =1 j =1 Δi i =1

d ni −1 vij ω α d ni −1 ϕ α α
i,j+1,j ωi,j+1,j
d
i,j+1,j
X4α = ∑∑ j+1,j 1/2
, X5α = ∑∑ j+1,j
, X6α = α
∑ ωi,n i ,1
.
i =1 j =1 ( Δ i ) i =1 j =1 Δi i =1

After some algebra, one obtains

X3α X4α − X1α X6α X α X α − X2α X4α

δα = and σα2 = 2Sα , where Sα = 1α 5α .
X2α X6α − X3α X5α X2 X6 − X3α X5α

On the other hand, and since

σ4 σ2 σ2
Φξ = δ2 X2α + Z3 + δσ2 X3α , Γξ = −δ X1α − Z2 , Υξ = −δ X3α − Z3 ,
4 2 2
Equation (9) results in

Sα [2n + Sα ] − δα [2X1α + δα X2α ] − Z1 = 0 (12)

233
Mathematics 2018, 6, 85

The solution of this equation provides the estimation of α, whereas those of the other parameters
are given by δα and σα2 .
 it is a trivariate normal distribution with mean ξ and
As regards the asymptotic distribution of ξ,
covariance matrix given by I (ξ )−1 , being
⎛ ⎞
X2α δ X5α − X3α
1 ⎜ δ X5α δ2 X7α −δ X6α ⎟
I (ξ ) = 2 ⎜ ⎟
σ ⎝ n Z3 ⎠
− X3α −δ X6α +
2σ2 4
with  2
d ni −1 α
ωi,j +1,j
X7α = ∑∑ j+1,j
.
i =1 j =1 Δi
This distribution can be used to obtain the asymptotic standard errors for the estimation of the
parameters as well as for some parametric functions of interest (see the last comment of the previous
section). In particular, we focus on the inflection time and the corresponding expected value of
the process at this instant, conditioned on X (t0 ) = x0 . Another important parametric function in
this context is the upper bound that determines the carrying capacity of the system modeled by the
process. Concretely:
 
• Upper bound, conditioned on X (t0 ) = x0 , g1 (θ) = x0 exp δ αt0 .
• Inflection time, g2 (θ) = − ln δ/ ln α.
• Value of the process at the time of inflection, conditioned on X (t0 ) = x0 , g3 (θ) = g1 (θ)/e.

On the other hand, when using the model for predictive purposes some of the parametric functions
of Table 1 can be used. In particular, the conditioned mean function adopts the expression
  
E[ X (t)| X (τ ) = y] = g4 (θ) = y exp −δ αt − ατ .

Note that this curve is of the type of Equation (11). For this reason, this function is useful for
forecasting purposes. In this case, it is of interest to provide not only the value of the function at each
time instant, but also the standard error of the prediction and a confidence interval determining a
range of values that includes, with a given confidence level, the true real value of the forecast.

Application to Real Data

The following example is based on a study developed in [27] on some aspects related to the
growth of a population of rabbits. Figure 1 shows the weight (in grams) of 29 rabbits over 30 weeks.
The sample paths begin at different initial values, thus showing a sigmoidal behavior, and their bounds
are dependent on the initial values. These two aspects suggest that using the Gompertz-type model
proposed above would be appropriate.
This data set has been used in various papers to illustrate some aspects of the Gompertz-type
process, such as the estimation of the parameters and the study of some time variables that may
be of interest in the analysis of growth phenomena of this nature. As regards the estimation of
the parameters, in [17] the authors designed an iterative method for solving the likelihood system
of equations, while in [24] the maximization of the likelihood function was directly addressed by
simulated annealing. In addition, in [28] two time variables of interest for this type of data were
analyzed: concretely the inflection time and the time instant in which the process reaches a certain
percentage of total growth. Both cases were modeled as first-passage time problems.
In this paper the estimation of the parameters has been carried out from the resolution of
Equation (12) by means of the bisection method (see Figure 2) and then by using expressions δα
and σα2 .

234
Mathematics 2018, 6, 85




:HLJKWJUDPV





      

7LPHZHHNV

Figure 1. Weight of 29 rabbits over 30 weeks.

Figure 2. Graph of equation for α.

Table 2 contains the estimated values for the parameters and the inflection time, as well as the
asymptotic estimation error and 95% confidence intervals by applying the delta method.

Table 2. Estimated values, standard errors and 95% confidence intervals of the parameters and the
inflection time.

Parametric Function δ α σ g2 (θ)

Estimated value 4.1020 0.8301 0.0708 7.5803
Standard error 0.0556 0.0021 0.0002 0.1053
Confidence interval (3.9929, 4.1063) (0.8258, 0.8343) (0.0704, 0.0713) (7.3738, 7.7869)

235
Mathematics 2018, 6, 85

As regards the weight value at the inflection time and the upper bound, remember that these
values depend on the one observed at the initial instant. Taking into account the range of observed
weight values at the initial instant of observation, several values have been considered within this
range. For these values, the expected weight of a rabbit at the moment of inflection has been studied,
as well as the possible value of the maximum weight (upper bound). Table 3 contains the estimated
values, the asymptotic standard errors, and the 95% confidence intervals.
Function E[ X (t)| X (t0 ) = x0 ] can be used to provide forecasts of the weight of a rabbit that
presents an initial weight x0 . Figure 3 shows, for a selection of four of the rabbits used in the study,
the estimated mean function together with the 95% asymptotic confidence intervals obtained for each
value of this function. Additionally, the observed values are included to check the quality of the
adjustment made by the model under consideration. Obviously, this type of representation can also be
obtained by considering any value of x0 in the range of the initial distribution of the weight. Note that
the estimated mean functions for each rabbit depend on the initial value, and so do the corresponding
confidence intervals for the mean at each time instant. Therefore, the graphs in the figure are different
for each rabbit although the estimation of the parameters is unique.

Table 3. Estimated values, standard errors, and 95% confidence intervals of the upper bound and value
at the inflection time for several values of the initial weight.

Upper Bound Value at Inflection Time

Initial Weight

g3 (θ) St. Error 95% Interval 
g1 (θ) St. Error 95% Interval
145 1772.836 70.546 (1634.568, 1911.104) 4819.068 191.764 (4443.215, 5194.920)
155 1772.836 75.411 (1625.032, 1920.640) 4819.068 204.990 (4417.295, 5220.841)
165 1883.638 80.276 (1726.298, 2040.978) 5120.260 218.215 (4692.566, 5547.954)
175 2105.243 85.142 (1938.367, 2272.118) 5722.643 231.440 (5269.028, 6176.258)
185 2216.045 90.007 (2039.634, 2392.456) 6023.835 244.665 (5544.299, 6503.371)
195 2216.045 94.872 (2030.098, 2401.992) 6023.835 257.890 (5518.378, 6529.291)
205 2105.243 99.737 (1909.760, 2300.726) 5722.643 271.115 (5191.266, 6254.020)
215 1883.638 104.603 (1678.620, 2088.657) 5120.260 284.341 (4562.961, 5677.558)

Figure 3. Cont.

236
Mathematics 2018, 6, 85

Figure 3. Observed values, estimated mean function, and confidence intervals for a choice of rabbits.

7. Conclusions
The present paper deals with some topics about inference for the non-homogeneous lognormal
process (or with exogenous factors). Starting from the general form of the process, we studied the
ML estimation of the parameters by using discrete sampling. This general overview enabled us to
provide a unified method for several diffusion processes which can be built from particular cases of
the non-homogeneous lognormal process for several choices of exogenous factors. In addition, we also
looked into the asymptotic distribution of estimators, through which we can calculate the estimation
errors and confidence intervals for the estimators of a wide range of parametric functions of interest
in many fields. Finally, the process here described is applied to the Gompertz-type diffusion process
introduced in [17].

Author Contributions: The three authors have participated equally in the development of this work, either in the
theoretical developments or in the applied aspects. The paper was also written and reviewed cooperatively.
Acknowledgments: This work was supported in part by the Ministerio de Economía, Industria y Competitividad,
Spain, under Grants MTM2014-58061-P and MTM2017-85568-P.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Cox, J.C.; Ross, S.A. The evaluation of options for alternative stochastic processes. J. Financ. Econ. 1976, 3,
145–166. [CrossRef]
2. Marcus, A.; Shaked, I. The relationship between accounting measures and prospective probabilities of insolvency:
An application to the banking industry. Financ. Rev. 1984, 19, 67–83. [CrossRef]
3. Merton, R.C. Option pricing when underlying stock returns are discontinuous. J. Financ. Econ. 1976, 3,
125–144. [CrossRef]
4. Black, F.; Scholes, M. The pricing of options and corporate liabilities. J. Political Econ. 1973, 81, 637–654.
[CrossRef]
5. Hunt, P.J.; Kennedy, J.G. Financial Derivatives in Theory and Practice; Revised Edition; John Wiley and Sons:
Chichester, UK, 2004; ISBN 978-0-470-86359-6.
6. Lamberton, D.; Lapeyre, B. Introduction to Stochastic Calculus Applied to Finance, 2nd ed.; Chapman and Hall:
New York, NY, USA, 2007; ISBN 9781584886266.

237
Mathematics 2018, 6, 85

7. Tintner, G.; Sengupta, J.K. Stochastic Economics; Academic Press: New York, NY, USA, 1972; ISBN 9781483274027.
8. Buonocore, A.; Caputo, L.; Pirozzi, E.; Nobile, A.G. A Non-Autonomous Stochastic Predator-Prey Model.
Math. Biosci. Eng. 2014, 11, 167–188. [CrossRef] [PubMed]
9. D’Onofrio, G.; Lansky, P.; Pirozzi, E. On two diffusion neuronal models with multiplicative noise: The mean
first-passage time properties. Chaos 2018, 28. [CrossRef]
10. Gutiérrez, R.; Román, P.; Romero, D.; Torres, F. Forecasting for the univariate lognormal diffusion process
with exogenous factors. Cybern. Syst. 2003, 34, 709–724. [CrossRef]
11. Gutiérrez, R.; Rico, N.; Román, P.; Romero, D.; Serrano, J.J.; Torres, F. Lognormal diffusion process with
polynomial exogenous factors. Cybern. Syst. 2006, 37, 293–309. [CrossRef]
12. Land, C.E. Hypothesis tests and interval estimates. In Lognormal Distributions, Theory and Applications;
Crow, E.L., Shimizu, K., Eds.; Marcel Dekker: New York, NY, USA, 1988; pp. 87–112, ISBN 0-8247-7803-0.
13. Bibby, B.; Jacobsen, M.; Sørensen, M. Estimating functions for discretely sampled diffusion type models.
In Handbook of Financial Econometrics; Aït-Sahalia, Y., Hansen, L., Eds.; North-Holland: Amsterdam,
The Netherlands; 2009; pp. 203–268, ISBN 978-0-444-50897-3.
14. Hansen, L. Large sample properties of generalized method of moments estimators. Econometrica 1982, 50,
1029–1054. [CrossRef]
15. Fuchs, C. Inference for Diffusion Processes; Springer: Heidelberg, Germany, 2013; ISBN 978-3-642-25968-5.
16. Tang, S.; Heron, E. Bayesian inference for a stochastic logistic model with switching points. Ecol. Model. 2008,
219, 153–169. [CrossRef]
17. Gutiérrez, R.; Román, P.; Romero, D.; Serrano, J.J.; Torres, F. A new gompertz-type diffusion process with
application to random growth. Math. Biosci. 2007, 208, 147–165. [CrossRef] [PubMed]
18. Román-Román, P.; Romero, D.; Torres-Ruiz, F. A diffusion process to model generalized von Bertalanffy
growth patterns: Fitting to real data. J. Theor. Biol. 2010, 263, 59–69. [CrossRef] [PubMed]
19. Román-Román, P.; Torres-Ruiz, F. Modelling logistic growth by a new diffusion process: Application to
biological system. BioSystems 2012, 110, 9–21. [CrossRef] [PubMed]
20. Román-Román, P.; Torres-Ruiz, F. A stochastic model related to the Richards-type growth curve. Estimation
by means of Simulated Annealing and Variable Neighborhood Search. App. Math. Comput. 2015, 266,
579–598. [CrossRef]
21. Román-Román, P.; Torres-Ruiz, F. The nonhomogeneous lognormal diffusion process as a general process
to model particular types of growth patterns. In Lecture Notes of Seminario Interdisciplinare di Matematica;
Università degli Studi della Basilicata: Potenza, Italy, 2015; Volume XII, pp. 201–219.
22. Da Luz Sant’Ana, I.; Román-Román, P.; Torres-Ruiz, F. Modeling oil production and its peak by means of a
stochastic diffusion process based on the Hubbert curve. Energy 2017, 133, 455–470. [CrossRef]
23. Barrera, A.; Román-Román, P.; Torres-Ruiz, F. A hyperbolastic type-I diffusion process: Parameter estimation
by means of the firefly algorithm. Biosystems 2018, 163, 11–22. [CrossRef] [PubMed]
24. Román-Román, P.; Romero, D.; Rubio, M.A.; Torres-Ruiz, F. Estimating the parameters of a Gompertz-type
diffusion process by means of simulated annealing. Appl. Math. Comput. 2012, 218, 5121–5131. [CrossRef]
25. Da Luz Sant’Ana, I.; Román-Román, P.; Torres-Ruiz, F. The Hubbert diffusion process: Estimation via
simulated annealing and variable neighborhood search procedures. Application to forecasting peak oil
production. Appl. Stoch. Models Bus. 2018. [CrossRef]
26. Gutiérrez, R.; Román, P.; Torres, F. Inference on some parametric functions in the univariate lognormal
diffusion process with exogenous factors. Test 2001, 10, 357-373. [CrossRef]
27. Blasco, A.; Piles, M.; Varona, L. A Bayesian analysis of the effect of selection for growth rate on growth
curves in rabbits. Genet. Sel. Evol. 2003, 35, 21–41. [CrossRef] [PubMed]
28. Gutiérrez-Jáimez, R.; Román, P.; Romero, D.; Serrano, J.J.; Torres, F. Some time random variables related to a
Gompertz-type diffusion process. Cybern. Syst. 2008, 39, 467–479. [CrossRef]

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

238
 mathematics
Article
On the Bounds for a Two-Dimensional Birth-Death
Process with Catastrophes
Anna Sinitcina 1,† , Yacov Satin 1,† , Alexander Zeifman 2, *,† , Galina Shilova 1,† , Alexander Sipin 1,† ,
Ksenia Kiseleva 1,† , Tatyana Panfilova 1,† , Anastasia Kryukova 1,† , Irina Gudkova 3,†
and Elena Fokicheva 1,†
1 Faculty of Applied Mathematics, Computer Technologies and Physics, Vologda State University,
160000 Vologda, Russia; [email protected] (An.S.); [email protected] (Y.S.); [email protected] (G.S.);
[email protected] (Al.S.); [email protected] (K.K.); [email protected] (T.P.);
[email protected] (A.K.); [email protected] (E.F.)
2 Faculty of Applied Mathematics, Computer Technologies and Physics, Vologda State University,
Institute of Informatics Problems, Federal Research Center “Computer Science and Control” of the Russian
Academy of Sciences, Vologda Research Center of the Russian Academy of SciencesSciences,
160000 Vologda, Russia
3 Applied Probability and Informatics Department, Peoples’ Friendship University of Russia
(RUDN University), Institute of Informatics Problems, Federal Research Center “Computer Science and
Control” of the Russian Academy of Sciences, 117198 Moskva, Russia; [email protected]
* Correspondence: [email protected]
† These authors contributed equally to this work.

Received: 19 April 2018; Accepted: 8 May 2018; Published: 11 May 2018

Abstract: The model of a two-dimensional birth-death process with possible catastrophes is studied.
The upper bounds on the rate of convergence in some weighted norms and the corresponding
perturbation bounds are obtained. In addition, we consider the detailed description of two examples
with 1-periodic intensities and various types of death (service) rates. The bounds on the rate of
convergence and the behavior of the corresponding mathematical expectations are obtained for
each example.

Keywords: continuous-time Markov chains; catastrophes; bounds; birth-death process; rate

of convergence

1. Introduction
There is a large number of papers devoted to the research of continuous-time Markov chains and
models with possible catastrophes, see for instance [1–21], and the references therein. Such models
are widely used in queueing theory and biology, particularly, for simulations in hight-performance
computing. In some recent papers, the authors deal with more or less special birth-death processes
with additional transitions from and to origin [9–13,18–20]. In [22], a general class of Markovian
queueing models with possible catastrophes is analyzed and some bounds on the rate of convergence
are obtained. Here we consider a more specific but important model of a two-dimensional birth-death
process with possible catastrophes and obtain the upper bounds on the rate of convergence in some
weighted norms and the corresponding perturbation bounds.
Ergodicity bounds in l-1 norm (associated with total variation) for such processes can be obtained
quite easily due to the possibility of catastrophes, i.e., transitions to zero from any other state.
Obtaining the estimates in weighted norms that guarantee the convergence of the corresponding
mathematical expectations as well as the construction of the corresponding limiting characteristics are
more complex problems.

Mathematics 2018, 6, 80; doi:10.3390/math6050080 239 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 80

In addition, we consider in detail two examples with 1-periodic intensities and various types of
death (service) rates. The bounds on the rate of convergence and the behavior of the corresponding
mathematical expectations are obtained for each example.
Our results seem to be interesting for both queueing theory and biology applications.
Let X(t) = ( X1 (t), X2 (t)) be two-dimensional birth-death-catastrophe process (where Xi (t) is
the corresponding number of particles of type i, i = 1, 2) such that in the interval (t, t + h) the
following transitions are possible with order h: birth of a particle of type i, death of a particle of type i,
and catastrophe (or transition to the zero state 0 = (0, 0)).
Denote by λ1,i,j (t), λ2,i,j (t), μ1,i,j (t), μ2,i,j (t), and by ξ i,j (t) corresponding birth, death,
and catastrophe rates for the process. Namely, λ1,i,j (t) is the rate of transition from state (i, j) to
state (i + 1, j) at the moment t, λ2,i,j (t) is the rate of transition from state (i, j) to state (i, j + 1), μ1,i,j (t)
is the rate of transition from state (i, j) to state (i − 1, j), μ2,i,j (t) is the rate of transition from state
(i, j) to state (i, j − 1), and finally, ξ i,j (t) is the rate of transition from state (i, j) to state (0, 0) at the
moment t.
The transition rate diagram associated with the process is presented in Figure 1.

Figure 1. Transition rate diagram.

Suppose that all intensities are nonnegative and locally integrable on [0; ∞) as functions of t.
Moreover, we also suppose that the condition of boundedness

λ1,i,j (t) + λ2,i,j (t) + μ1,i,j (t) + μ2,i,j (t) + ξ i,j (t) ≤ L < ∞, (1)

hold for any i, j and almost all t ≥ 0.

We renumber the states of two-dimensional process X(t) = ( X1 (t), X2 (t)) (0,0), (0,1), (1,0), (0,2),
(1,1), (2,0), . . . by increasing the sum of coordinates, and in the case of the same sum, by increasing
the first coordinate. Hence we obtain one-dimensional vector p(t) = ( p0 (t), p1 (t), . . . ) T of state

240
Mathematics 2018, 6, 80

probabilities in a new numeration, and therefore, we can rewrite the forward Kolmogorov system in
the following form:
dp
= A (t) p, t ≥ 0, (2)
dt
 
where A(t) = aij (t) is the corresponding transposed intensity matrix:
⎛ ⎞
a00 μ1,0,1 + ξ 0,1 μ2,1,0 + ξ 1,0 ξ 0,2 ξ 1,1 ξ 2,0 ξ 0,3 ξ 1,2 ξ 2,1 ξ 3,0 ···
⎜ λ1,0,0 μ1,0,2 μ2,1,1 ··· ⎟
⎜ a11 0 0 0 0 0 0 ⎟
⎜ ⎟
⎜ λ2,0,0 0 a22 0 μ1,1,1 μ2,2,0 0 0 0 0 ··· ⎟
⎜ ⎟
⎜ 0 λ1,0,1 0 a33 0 0 μ1,0,3 μ2,1,2 0 0 ··· ⎟
A(t) = ⎜
⎜
⎟,
⎟
⎜ 0 λ2,0,1 λ1,1,0 0 a44 0 0 μ1,1,2 μ2,2,1 0 ··· ⎟
⎜ ⎟
⎜ 0 0 λ2,1,0 0 0 a55 0 0 μ1,2,1 μ2,3,0 ··· ⎟
⎜ ⎟
⎝ 0 0 0 λ1,0,2 0 0 a66 0 0 0 ··· ⎠
··· ··· ··· ··· ··· ··· ··· ··· ···

and aii (t) = − ∑i ai,j (t).

Throughout the paper by  ·  we denote the l1 -norm, i. e., x = ∑ | xi |, and  B = sup j ∑i |bij |
∞ .
for B = (bij )i,j =0
Let Ω be a set all stochastic vectors, i.e., l1 vectors are with nonnegative coordinates and unit
norm. Hence the assumption (1) implies the bound A(t) ≤ 2L for almost all t ≥ 0. Therefore,
the operator function A(t) from l1 into itself is bounded for almost all t ≥ 0 and locally integrable on
[0; ∞). Therefore, we can consider the forward Kolmogorov system as a differential equation in the
space l1 with bounded operator.
It is well known, see [23], that the Cauchy problem for such a differential equation has a unique
solution for an arbitrary initial condition, and p(s) ∈ Ω implies p(t) ∈ Ω for t ≥ s ≥ 0.
We have
p(t) = U (t, s)p(s), (3)

where U (t, s) is the Cauchy operator of Equation (2).

Note that the vector of state probabilities can be written in ’two-dimensional form’ as
p(t) = ( p00 (t), p01 (t), p10 (t), p02 (t), p11 (t), . . . ) T .

2. Bounds in l1 Norm
Consider the first equation in forward Kolmogorov system and rewrite it in the following form:

dp0
= − ( a00 + ξ (t)) p0 + ∑ ( a0i (t) − ξ (t)) pi + ξ (t), (4)
dt i ≥1

where ξ (t) = infi,j ξ i,j (t).

Then we have from Equation (2) the following system:

dp
= B ( t ) p + f ( t ), (5)
dt

where f(t) = (ξ (t), 0, . . .) T and

⎛ ⎞
a00 − ξ μ1,0,1 + ξ 0,1 − ξ μ2,1,0 + ξ 1,0 − ξ ξ 0,2 − ξ ξ 1,1 − ξ ξ 2,0 − ξ ξ 0,3 − ξ ξ 1,2 − ξ ξ 2,1 − ξ ξ 3,0 − ξ ···
⎜ ⎟
⎜ ⎟
⎜ λ1,0,0 a11 0 μ1,0,2 μ2,1,1 0 0 0 0 0 ··· ⎟
⎜ ⎟
⎜ ⎟
⎜ λ2,0,0 0 a22 0 μ1,1,1 μ2,2,0 0 0 0 0 ··· ⎟
⎜ ⎟
⎜ ⎟
⎜ 0 λ1,0,1 0 a33 0 0 μ1,0,3 μ2,1,2 0 0 ··· ⎟
⎜ ⎟
B(t) = ⎜ ⎟.
⎜ λ2,0,1 λ1,1,0 μ1,1,2 μ2,2,1 ··· ⎟
⎜ 0 0 a44 0 0 0 ⎟
⎜ ⎟
⎜ ⎟
⎜ 0 0 λ2,1,0 0 0 a55 0 0 μ1,2,1 μ2,3,0 ··· ⎟
⎜ ⎟
⎜ ⎟
⎜ 0 0 0 λ1,0,2 0 0 a66 0 0 0 ··· ⎟
⎝ ⎠
··· ··· ··· ··· ··· ··· ··· ··· ···

241
Mathematics 2018, 6, 80

We have  t
p ( t ) = V ( t ) p (0) + V (t, τ )f(τ ) dτ, (6)
0

where V (t, τ ) is the Cauchy operator of Equation (5).

Then one can estimate the logarithmic norm of B(t) in the space of sequences l1 (see [24]):
(
γ ( B(t))1 = max a00 (t) − ξ (t) + ∑ ai0 (t),
i ≥1
( )) (7)
sup aii (t) + a0i (t) − ξ (t) + ∑ a ji (t) = − ξ ( t ).
i ≥1 j =i,j≥1

Then for all 0 ≤ s ≤ t we have

t
− ξ (τ ) dτ
V (t, s) ≤ e s . (8)

Therefore, the following statement is correct (see details in [16,18]).

Theorem 1. Let the intensities of catastrophes be essential, that is

 ∞
ξ (t) dt = ∞. (9)
0

Then the process X(t) is weakly ergodic and the following bound of the rate of convergence holds:

t
− ξ (τ ) dτ
∗ ∗∗
p (t) − p (t) ≤ 2e 0 , (10)

for all initial conditions p∗ (0), p∗∗ (0) and any t ≥ 0.

Consider now the ”perturbed” process X̄ = X̄(t), t ≥ 0, adding a dash on top for all
corresponding characteristics.
Put Â(t) = A(t) − Ā(t), and assume that the perturbations are “uniformly small”, i.e., for almost
all t ≥ 0 the following inequality is correct

Â(t) ≤ ε. (11)

Consider the stability bounds of the process X(t) under these perturbations. In addition,
we assume that the process is exponentially ergodic, that is, that for some positive M, a and for all s, t,
0 ≤ s ≤ t the following inequality holds
t
e− s ξ (u) du
≤ Me−a(t−s) . (12)

Then from Theorem 1:

t t
− ξ (τ ) dτ − ξ (τ ) dτ
∗ ∗∗ ∗ ∗∗
p (t) − p (t) ≤ e s p (s) − p (s) ≤ 2e s ≤ 2Me−a(t−s) . (13)

Here we apply the approach proposed in [25] for a stationary case and generalized for a
nonstationary situation in [15,16].
We have
 t
p(t) − p̄(t) ≤ Me−a(t−s) p(s) − p̄(s) + M  Â(u)e−a(u−s) du. (14)
s

242
Mathematics 2018, 6, 80

Therefore we obtain

p(s) − p̄(s) + (t − s)ε, 0 < t < a−1 log M,
p(t) − p̄(t) ≤ (15)
a−1 (log M + 1 − Me− a(t−s) )ε + Me− a(t−s) p(s) − p̄(s), t ≥ a−1 log M.

It implies the following statement.

Theorem 2. If the condition (12) is fulfilled and the perturbations are uniformly small:

 Â(t) ≤ ε, (16)

for almost all t ≥ 0. Then the following bound holds:

ε (1 + log M )
lim sup p(t) − p̄(t) ≤ , (17)
t→∞ a

for any initial conditions p(0), p̄(0).

Corollary 1. Let the intensities of the process be 1-periodic and instead of (12) we have the following inequality
 1
ξ (t)dt ≥ θ > 0. (18)
0

Then (17) is correct for

M = eK , a = θ, (19)
t
where K = sup|t−s|≤1 s ξ (τ )dτ < ∞.

3. Bounds in Weighted Norms

Consider the diagonal matrix D = diag(d0 , d1 , d2 , d3 , · · · ), with entries of the increasing sequence
* +
{dn }, where d0 = 1, and the corresponding space of sequences l1D = z = ( p0 , p1 , p2 , . . .) T such that
z1D =  Dz1 < ∞.
Then one can estimate the logarithmic norm of operator B(t) in l1D space.
According to the general approach, we obtain the matrix
⎛ ⎞
d d d d d d d
⎜ a00 − ξ (μ1,0,1 + ξ 0,1 − ξ ) 0 (μ2,1,0 + ξ 1,0 − ξ ) 0 (ξ 0,2 − ξ ) 0 (ξ 1,1 − ξ ) 0 (ξ 2,0 − ξ ) 0 (ξ 0,3 − ξ ) 0 (ξ 1,2 − ξ ) 0 ··· ⎟
⎜ d1 d2 d3 d4 d5 d6 d7 ⎟
⎜ ⎟
⎜ d d d ⎟
⎜ λ1,0,0 1 a11 0 μ1,0,2 1 μ2,1,1 1 0 0 0 ··· ⎟
⎜ d0 d3 d4 ⎟
⎜ ⎟
⎜ ⎟
⎜ d d d ⎟
⎜ λ2,0,0 2 0 a22 0 μ1,1,1 2 μ2,2,0 2 0 0 ··· ⎟
⎜ d0 d4 d5 ⎟
⎜ ⎟
⎜ d d d ⎟
⎜ 0 λ1,0,1 3 0 a33 0 0 μ1,0,3 3 μ2,1,2 3 ··· ⎟
⎜ ⎟
DB(t) D −1 = ⎜
⎜
⎜
d1 d6 d7 ⎟
⎟
⎟
,
d d d4
⎜ 0 λ2,0,1 4 λ1,1,0 4 0 a44 0 0 μ1,1,2 ··· ⎟
⎜ d1 d2 d7 ⎟
⎜ ⎟
⎜ ⎟
⎜ d ⎟
⎜ 0 0 λ2,1,0 5 0 0 a55 0 0 ··· ⎟
⎜ d2 ⎟
⎜ ⎟
⎜ d ⎟
⎜ 0 0 0 λ1,0,2 6 0 0 a66 0 ··· ⎟
⎜ d3 ⎟
⎜ ⎟
⎝ ⎠
··· ··· ··· ··· ··· ··· ···

where aii (t) = − ∑i ai,j (t).

Consider now the logarithmic norm
 
γ ( B(t))1D = γ DB(t) D −1 . (20)
1

243
Mathematics 2018, 6, 80

Let us make the correspondence between the column number of matrix DB(t) D −1 and the number
of zeros under the main diagonal in this column (till the first nonzero element). Then we obtain the
arithmetic progression { ai }:

a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11 a12 a13 a14 a15 a16 a17 a18 a19 a20 a21 ···
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 ···
1 1 2 2 2 3 3 3 3 4 4 4 4 4 5 5 5 5 5 5 6 ···

We compose the sequence {bi } of the number of identical entries of the third line: 2, 3, 4, 5, 6, · · · .
Note that ∑iN=1 bi is equal to the last ak , corresponding to the number of zeros N in the k-th column.
Then the sum of the first N elements of sequence {bi } is approximately equal to the number of
the column an = n:
(2b1 + ( N − 1)) N ≈ 2an = 2n.
Knowing the column number n, one can find the formula for the number of zeros N under the
main diagonal in this column till the first nonzero element. We note that the number of zeros over the
diagonal till μ1,i,j is one less.
If N is not an integer, we must take the nearest right to N an integer.
j
One can see that columns 2, 5, 9, 14, ... (these columns correspond to sums ∑i=1 bi and integer N)
contain death rates μ2,i,j (t) only, and columns 3, 6, 10, 15, ... contain death rates μ1,i,j (t) only, and all
other columns contain the both death intensities.
Consider the following quantities:
for n = 0
αn (t) = λ1,0,0 (t) + λ2,0,0 (t) + ξ (t) − λ1,0,0 (t) dd1 − λ2,0,0 (t) dd2 ,
0 0

for n = 1

αn (t) = λ1,0,1 (t) + λ2,0,1 (t) + μ1,0,1 (t) + ξ 1 (t) − λ1,0,1 (t) dd3 − λ2,0,1 (t) dd4 − (μ1,0,1 (t) + ξ 1 (t) − ξ (t)) dd0 ,
1 1 1

for n = 2

αn (t) = λ1,1,0 (t) + λ2,1,0 (t) + μ2,1,0 (t) + ξ 2 (t) − λ1,1,0 (t) dd4 − λ2,1,0 (t) dd5 − (μ2,1,0 (t) + ξ 2 (t) − ξ (t)) dd0 ,
2 2 2

√
−3+ 9+8n
for integer 2 :

d n + N +1
αn (t) = λ1,s−1 (n) (t) + λ2,s−1 (n) (t) + μ2,s−1 (n) (t) + ξ s(i,j) (t) − λ1,s−1 (n) (t) dn − λ2,s−1 (n) (t) dn+dNn +2
√
−3+ 9+8n
−μ2,s−1 (n) (t) dn−dNn −1 − (ξ s(i,j) (t) − ξ (t)) ddn0 , N= 2 ,
√
−3+ 9+8( n −1)
for integer 2 :

d n + N +1
αn (t) = λ1,s−1 (n) (t) + λ2,s−1 (n) (t) + μ1,s−1 (n) (t) + ξ s(i,j) (t) − λ1,s−1 (n) (t) − λ2,s−1 (n) (t) dn+dNn +2
1 √ 2 dn
−μ1,s−1 (n) (t) dnd−nN − (ξ s(i,j) (t) − ξ (t)) ddn0 , N = −3+ 2 9+8n ,

in other cases:
d n + N +1
αn (t) = λ1,s−1 (n) (t) + λ2,s−1 (n) (t) + μ1,s−1 (n) (t) + μ2,s−1 (n) (t) + ξ s(i,j) (t) − λ1,s−1 (n) (t) dn
−λ2,s−1 (n) (t) dn+dNn +2 − μ1,s−1 (n) (t) dnd−nN − μ2,s−1 (n) (t) dn−dNn −1
1 √ 2
−(ξ s(i,j) (t) − ξ (t)) ddn0 , N = −3+ 2 9+8n .

Then the following algorithm helps us to correlate the number n and pair (i, j):
(1) n1 = n − 1,

244
Mathematics 2018, 6, 80

(2) n2 = n1 − 2,
···
(k) nk = nk−1 − k, while nk > 0,
(k+1) i = nk , j = k − nk .

We obtain
γ ( B(t))1D = − inf αn (t) = −α(t). (21)
t

Therefore, for all 0 ≤ s ≤ t we have the bound for the corresponding Cauchy operator:

t
− α(τ ) dτ
V (t, s)1D ≤ e s , (22)

and the following statement.

Theorem 3. Let for some sequence {di } we have the condition

 ∞
α(t) dt = ∞. (23)
0

Then the process X(t) is weakly ergodic and the following bound of the rate of convergence is correct:

t
− α(τ ) dτ
p∗ (t) − p∗∗ (t)1D ≤ e 0 p∗ (0) − p∗∗ (0)1D , (24)

for any initial conditions p∗ (0), p∗∗ (0) and for all t ≥ 0.

Mathematical expectations for both processes X1 (t) and X2 (t) can be obtained using formulas:

E1 (t) = 1( p2 + p4 + p7 + . . .) + 2( p5 + p8 + p12 + · · · ) + . . . (25)

= 1( p10 + p11 + p12 + . . .) + 2( p20 + p21 + p22 + · · · ) + . . .

and

E2 (t) = 1( p1 + p4 + p8 + . . .) + 2( p3 + p7 + p12 + · · · ) + . . . (26)

= 1( p01 + p11 + p21 + . . .) + 2( p02 + p12 + p22 + · · · ) + . . . .

Let us now introduce a process N (t) = |X(t)| = X1 (t) + X2 (t), that is the number of all particles
at the moment t.
Then one has for the mathematical expectation (the mean) of this process the following equality:

EN (t) = 1( p01 + p10 ) + 2( p02 + p11 + p20 ) + 3( p03 + p12 + p21 + p30 ) + · · · ) + . . . (27)
= 1( p1 + p2 ) + 2( p3 + p4 + p5 ) + 3( p6 + p7 + p8 + p9 ) + . . . = E1 (t) + E2 (t).
di
We note that for W = infi≥0 i the next inequality holds

1  p(t)1D
E N ( t ) = 1( p1 + p2 ) + 2( p3 + p4 + p5 ) + 3( p6 + p7 + p8 + p9 ) + . . . ≤ ∑ ipi ≤ W ∑ di pi = W
.
i ≥1 i ≥1

Denote by EN (t, k ) = E(| X (t)|/| X (0)| = k) the conditional expected number of all particles in
the system at instant t, provided that initially (at instant t = 0) k particles of both types were present in
the system.

245
Mathematics 2018, 6, 80

Corollary 2. Let the condition (23) hold and there is a sequence {di } such that W > 0, then the process
N (t) has the limiting mean φN (t) = EN (t, 0), and for any j and all t ≥ 0 the following bound of the rate of
convergence is correct:
t
1 + dj − α(τ ) dτ
| EN (t, j) − EN (t, 0)| ≤ e 0 . (28)
W

Applying in addition the condition that the perturbations of the intensity matrix are small enough
in the corresponding norm, that is  Â(t)1D ≤ ε, one can also obtain  B̂(t)1D ≤ ε.
We assume here that the process X(t) is exponentially ergodic in l1D -norm, that is for some
positive M1 , a1 and for all s, t, 0 ≤ s ≤ t the following inequality holds:
t
e− s α(u) du
≤ M1 e−a1 (t−s) . (29)

Here we apply the approach from [18].

One can rewrite the original system for the unperturbed process in the form:

dp
= B̄(t)p(t) + f̄(t) + B̂(t)p(t) + f̂(t). (30)
dt
Then  t  t
p(t) = V̄ (t, 0)p(0) + V̄ (t, τ )f̄(τ ) dτ + V̄ (t, τ ) B̂(τ ) p(τ ) (31)
0 0
and  t
p̄(t) = V̄ (t, 0)p̄(0) + V̄ (t, τ )f̄(τ ) dτ. (32)
0
Therefore, in any norm for any initial conditions we have the correct bound:
 t  
p(t) − p̄(t) ≤ V̄ (t, τ )  B̂(τ )p(τ ) + f̂(τ ) dτ. (33)
0

Then we have the following inequality for the logarithmic norm:

γ( B̄(t))1D ≤ γ( DB(t) D −1 )1 +  B̂(t)1D ≤ −α(t) + ε. (34)

On the other hand, one can obtain the estimation using inequality (29):

t
p(t)1D ≤ V (t)p(0)1D + V (t, τ )f(τ ) dτ 1D ≤
0
LM1
≤ M1 e−a1 t p(0)1D + (35)
a1

for any initial condition p(0). Moreover, f̂(τ )1D ≤ ε.

Then using bound (33), we have
 t 

 
t LM1
p(t) − p̄(t)1D ≤ e− τ (α(u)−ε)du ε M1 e− a1 τ p(0)1D + + ε dτ ≤
0 a1
εM1 (1 + LM1 /a1 )
≤ o (1) + .
a1 − ε

Therefore, the following statement is correct.

246
Mathematics 2018, 6, 80

Theorem 4. Let inequalities (12) and (29) hold for any initial condition p(0) ∈ l1D and for all t ≥ 0,
then we have

LM1 + a1
p(t) − p̄(t)1D ≤ M1 ε + M1 te−(a1 −ε)t p(0)1D , (36)
a1 ( a1 − ε )
and
M1 ε( LM1 + a1 )
lim sup p(t) − p̄(t)1D ≤ . (37)
t→∞ a1 ( a1 − ε )

Corollary 3. Let in addition the sequence be increasing fast enough, such that W > 0, then for any j, t ≥ 0
we have
M ε( LM1 + a1 )
lim sup | EN (t) − ĒN (t)| ≤ 1 . (38)
t→∞ Wa1 ( a1 − ε)

4. Examples
Example 1. Let λ1,i,j (t) = λ2,i,j (t) = 2 + sin 2πt, i, j ≥ 0, μ2,0,1 (t) = 1 + cos 2πt, μ1,1,0 (t) = 2(1 + cos 2πt),
and other μ1,i,j (t) = μ2,i,j (t) = 3(1 + cos 2πt), and let catastrophe intensities be ξi,j (t) = 5. Put ε = 10−6.

n
Choose dn = 1 + 10 , then α(t) = α0 (t) − 2
3 = 56
10 − 3
10 sin 2πt, a = 5, M = 1, a1 = 3.8, M1 = 1.1,
W = 1/10.
We obtain now the following bounds

| EN (t, 1) − EN (t, 0)| ≤ 3.4 · 10−4 , t ≥ 4, (39)

lim sup p(t) − p̄(t) ≤ 2 · 10−4 , (40)

t→∞

lim sup | EN (t, 0) − ĒN (t, 0)| ≤ 0.023. (41)

t→∞

The values of αn (t) are shown in Figure 2. The mean for the process N (t) on the interval t ∈ [0, 3]
for different initial conditions are shown in Figures 3–6, and the bounds for the limiting perturbed
mean is shown in Figure 7.

Figure 2. The values of several αn (t) for Example 1.

247
Mathematics 2018, 6, 80










      

Figure 3. The mean EN (t, 0) on the interval t ∈ [0, 3] with initial condition X(t) = (0, 0) for Example 1.












      

Figure 4. The mean EN (t, 29) on the interval t ∈ [0, 3] with initial condition X(t) = (0, 29) for
Example 1.

248
Mathematics 2018, 6, 80












      

Figure 5. The mean EN (t, 29) on the interval t ∈ [0, 3] with initial condition X(t) = (14, 15) for
Example 1.












      

Figure 6. The mean EN (t, 29) on the interval t ∈ [0, 3] with initial condition X(t) = (29, 0) for
Example 1.

249
Mathematics 2018, 6, 80














     

Figure 7. The limiting perturbed mean for t ∈ [2, 3] for Example 1.

Example 2. Let now λ1,i,j (t) = λ2,i,j (t) = 2 + cos 2πt, i, j ≥ 0, μ1,i,j (t) = min(1 + i · j, 3)(1 + cos 2πt),
i ≥ 1, j ≥ 0, μ2,i,j (t) = min(1 + i · j, 3)(1 + cos 2πt), i ≥ 0, j ≥ 1, and let the catastrophe rates be ξ i,j (t) = 5.
Let ε = 10−3 .

n
Put dn = 1 + 10 , then α(t) = α9 (t) − 17 = 548
133 + 19 sin 2πt − 19 cos 2πt, a = 5, M = 1, a1 = 4.12,
4 9

M1 = 1.2, W = 1/10.
Then we obtain
| EN (t, 1) − EN (t, 0)| ≤ 9.4 · 10−4 , t ≥ 2, (42)

lim sup p(t) − p̄(t) ≤ 2 · 10−4 , (43)

t→∞

lim sup | EN (t, 0) − ĒN (t, 0)| ≤ 0.021. (44)

t→∞

The values of αn (t) are shown in Figure 8.

The mean for the process N (t) on the interval t ∈ [0, 3] for different initial conditions are shown
in Figures 9–12, and the bounds for the limiting perturbed mean is shown in Figure 13.

250
Mathematics 2018, 6, 80

Figure 8. The values of several αn (t) for Example 2.
















      

Figure 9. The mean EN (t, 0) on the interval t ∈ [0, 3] with initial condition X(t) = (0, 0) for Example 2.

251
Mathematics 2018, 6, 80












      

Figure 10. The mean EN (t, 29) on the interval t ∈ [0, 3] with initial condition X(t) = (0, 29) for
Example 2.












      

Figure 11. The mean EN (t, 29) on the interval t ∈ [0, 3] with initial condition X(t) = (14, 15) for
Example 2.

252
Mathematics 2018, 6, 80












      

Figure 12. The mean EN (t, 29) on the interval t ∈ [0, 3] with initial condition X(t) = (0, 29) for
Example 2.












     

Figure 13. The limiting perturbed mean for t ∈ [2, 3] for Example 2.

253
Mathematics 2018, 6, 80

Remark 1. These graphs give us the additional information on the considered examples. Namely, Figures 2 and 8
show the bounding on the rate of convergence, see Equation (21) for Examples 1, 2 respectively, in Figures 3–6
and 9–12 one can see the mathematical expectation of the number of all particles at the moment t until the
stationary behaviour. Finally, the limiting behaviour of the limiting mathematical expectation of the number of
all particles for the perturbed process is shown in Figures 7 and 13.

Author Contributions: Conceptualization, A.Z., Al.S. and G.S.; Methodology, A.Z., K.K., T.P, E.F.; Software, Y.S.;
Validation, A.Z., An.S, K.K., A.K. and I.G.; Investigation, An.S., A.Z.; Writing-Original Draft Preparation, An.S,
A.Z.; Writing-Review and Editing, K.K., I.G.; Supervision, A.Z.; Project Administration, K.K.
Acknowledgments: The publication was supported by the Ministry of Education and Science of the Russian
Federation (project No. 2.882.2017/4.6).
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Di Crescenzo, A.; Giorno, V.; Nobile, A.G.; Ricciardi, L.M. On the M/M/1 queue with catastrophes and its
continuous approximation. Queueing Syst. 2003, 43, 329–347. [CrossRef]
2. Di Crescenzo, A.; Giorno, V.; Nobile, A.G.; Ricciardi, L.M. A note on birth-death processes with catastrophes.
Stat. Probab. Lett. 2008, 78, 2248–2257. [CrossRef]
3. Di Crescenzo, A.; Giorno, V.; Kumar, B.K.; Nobile, A.G. A double-ended queue with catastrophes and repairs,
and a jump-diffusion approximation. Methodol. Comput. Appl. Probab. 2012, 14, 937–954. [CrossRef]
4. Dharmaraja, S.; Di Crescenzo, A.; Giorno, V.; Nobile, A.G. A continuous-time Ehrenfest model with
catastrophes and its jump-diffusion approximation. J. Stat. Phys. 2015, 161, 326–345. [CrossRef]
5. Dudin, A.; Nishimura, S. A BMAP/SM/1 queueing system with Markovian arrival input of disasters.
J. Appl. Probab. 1999, 36, 868–881. [CrossRef]
6. Dudin, A.; Karolik, A. BMAP/SM/1 queue with Markovian input of disasters and non-instantaneous
recovery. Perform. Eval. 2001, 45, 19–32. [CrossRef]
7. Dudin, A.; Semenova, O. A stable algorithm for stationary distribution calculation for a BMAP/SM/1
queueing system with Markovian arrival input of disasters. J. Appl. Probab. 2004, 41, 547–556. [CrossRef]
8. Do, T.V.; Papp, D.; Chakka, R.; Sztrik, J.; Wang, J. M/M/1 retrial queue with working vacations and negative
customer arrivals. Int. J. Adv. Intell. Paradig. 2014, 6, 52–65. [CrossRef]
9. Chen, A.Y.; Renshaw, E. The M/M/1 queue with mass exodus and mass arrives when empty. J. Appl. Prob.
1997, 34, 192–207. [CrossRef]
10. Chen, A.Y.; Renshaw, E. Markov bulk-arriving queues with state-dependent control at idle time.
Adv. Appl. Prob. 2004, 36, 499–524. [CrossRef]
11. Chen, A.Y.; Pollett, P.; Li, J.P.; Zhang, H.J. Markovian bulk-arrival and bulk-service queues with
state-dependent control. Queueing Syst. 2010, 64, 267–304. [CrossRef]
12. Li, J.; Chen, A. The Decay Parameter and Invariant Measures for Markovian Bulk-Arrival Queues with
Control at Idle Time. Methodol. Comput. Appl. Probab. 2013, 15, 467–484.
13. Zhang, L.; Li, J. The M/M/c queue with mass exodus and mass arrivals when empty. J. Appl. Probab. 2015,
52, 990–1002. [CrossRef]
14. Zeifman, A.; Korotysheva, A.; Satin, Y.; Shorgin, S. On stability for nonstationary queueing systemswith
catastrophes. Inform. Appl. 2010, 4, 9–15.
15. Zeifman, A.; Korotysheva, A.; Panfilova, T.; Shorgin, S. Stability bounds for some queueing systems with
catastrophes. Inform. Appl. 2011, 5, 27–33.
16. Zeifman, A.; Korotysheva, A. Perturbation Bounds for Mt/Mt/N Queue with Catastrophes. Stoch. Model.
2012, 28, 49–62. [CrossRef]
17. Zeifman, A.; Satin, Y.; Panfilova, T. Limiting characteristics for finite birth-death-catastrophe processes.
Mater. Biosci. 2013, 245, 96–102. [CrossRef] [PubMed]
18. Zeifman, A.; Satin, Y.; Korotysheva, A.; Korolev, V.; Shorgin, S.; Razumchik, R. Ergodicity and perturbation
bounds for inhomogeneous birth and death processes with additional transitions from and to origin. Int. J. Appl.
Math. Comput. Sci. 2015, 25, 787–802. [CrossRef]

254
Mathematics 2018, 6, 80

19. Zeifman, A.; Satin, Y.; Korotysheva, A.; Korolev, V.; Bening, V. On a class of Markovian queuing systems
described by inhomogeneous birth-and-death processes with additional transitions. Dokl. Math. 2016, 94,
502–505. [CrossRef]
20. Zeifman, A.; Korotysheva, A.; Satin, Y.; Kiseleva, K.; Korolev, V.; Shorgin, S. Ergodicity and uniform in time
truncation bounds for inhomogeneous birth and death processes with additional transitions from and to
origin. Stoch. Model. 2017, 33, 598–616. [CrossRef]
21. Giorno, V.; Nobile, A.; Spina, S. On some time non-homogeneous queueing systems with catastrophes.
Appl. Math. 2014, 245, 220–234. [CrossRef]
22. Zeifman, A.; Korotysheva, A.; Satin, Y.; Kiseleva, K.; Korolev, V.; Shorgin, S. Bounds for Markovian Queues
with Possible Catastrophes. In Proceedings of the 31th European Conference on Modeling and Simulation,
Budapest, Hungary, 23–26 May 2017; pp. 628–634.
23. Daleckii, J.L.; Krein, M.G. Stability of solutions of differential equations in Banach space (No. 43). Am. Math. Soc.
2002, 53, 34–36.
24. Zeifman, A.; Satin, Y.; Chegodaev, A. On nonstationary queueing systems with catastrophes. Inform. Appl.
2009, 3, 47–54.
25. Mitrophanov, A.Y. Stability and exponential convergence of continuous-time Markov chains. J. Appl. Prob.
2003, 40, 970–979. [CrossRef]

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

255
 mathematics
Article
Network Reliability Modeling Based on a Geometric
Counting Process
Somayeh Zarezadeh 1 , Somayeh Ashrafi 2 and Majid Asadi 2,3, *
1 Department of Statistics, Shiraz University, Shiraz 71454, Iran; [email protected]
2 Department of Statistics, University of Isfahan, Isfahan 81744, Iran; s.ashrafi@sci.ui.ac.ir
3 School of Mathematics, Institute of Research in Fundamental Sciences (IPM),
P.O. Box 19395-5746, Tehran, Iran
* Correspondence: [email protected]

Received: 4 August 2018 ; Accepted: 8 October 2018; Published: 11 October 2018

Abstract: In this paper, we investigate the reliability and stochastic properties of an n-component
network under the assumption that the components of the network fail according to a counting
process called a geometric counting process (GCP). The paper has two parts. In the first part, we consider
a two-state network (with states up and down) and we assume that its components are subjected to
failure based on a GCP. Some mixture representations for the network reliability are obtained in terms
of signature of the network and the reliability function of the arrival times of the GCP. Several aging
and stochastic properties of the network are investigated. The reliabilities of two different networks
subjected to the same or different GCPs are compared based on the stochastic order between their
signature vectors. The residual lifetime of the network is also assessed where the components fail
based on a GCP. The second part of the paper is concerned with three-state networks. We consider a
network made up of n components which starts operating at time t = 0. It is assumed that, at any
time t > 0, the network can be in one of three states up, partial performance or down. The components
of the network are subjected to failure on the basis of a GCP, which leads to change of network states.
Under these scenarios, we obtain several stochastic and dependency characteristics of the network
lifetime. Some illustrative examples and plots are also provided throughout the article.

Keywords: two-dimensional signature; multi-state network; totally positive of order 2; stochastic

order; stochastic process

1. Introduction
In recent years, there has been a great growth in the use of networks (systems), such as
communication networks and computer networks, in human life. The networks are a set of nodes
that are connected by a set of links to exchange data through the links, where some particular nodes
in the network are called terminals. Usually, a network can be modeled mathematically as a graph
G(V, E, T) in which V shows the collection of nodes, E shows the set of links and T denotes the set
of terminals. Depending on the purpose of designing a network, the states of the network can be
defined in terms of the connections between the terminals. In the simplest case, the networks have two
states: up and down. However, in some applications, the networks may have several states which are
known, in reliability engineering, as the multi-state networks. Multi-state networks have extensive
applications in various areas of science and technology. From a mathematical point of view, the states
of multi-state networks are usually shown by, K = 0, 1, . . . , M, in which K = 0 shows the complete
failure of the network and K = M shows the perfect functioning of the network. A large number of
research works have been published in literature on the reliability and aging properties of multi-state
networks and systems under different scenarios. For the recent works on various applications and
reliability properties of networks, we refer to [1–10].

Mathematics 2018, 6, 197; doi:10.3390/math6100197 256 www.mdpi.com/journal/mathematics

Mathematics 2018, 6, 197

When a network is operating during its mission, its states may change over the time according
to the change of the states of its components. From the reliability viewpoint, the change in the states
of the components may occur based on a specific stochastic mechanism. In a recent book, Gertsbakh
and Shpungin [11] have proposed a new reliability model for a two-state network under the condition
that components (with two states) fail according to a renewal process. Motivated by this, Zarezadeh
and Asadi [12] and Zarezadeh et al. [13] studied the reliability of networks under the assumption
that the components are subject to failure according to a counting process. Under the special case that
the process of the components’ failure is a nonhomogeneous Poisson process (NHPP), they arrived
at some mixture representations for the reliability function of the network lifetime and explored its
stochastic and aging properties under different conditions.
The aim of the present paper is to assess the network reliability under the condition that the
failure of the components appear according to a recently proposed stochastic process called geometric
counting process (GCP). We assume that the nodes of the network are absolutely reliable and, throughout
the paper, whenever we say that the components of the network fail, we mean that the links of the
network fail. Let {ξ (t), t ≥ 0} be a counting process where ξ (t) denotes the number of events in [0, t].
A GCP, introduced in [14], is a subclass of counting process {ξ (t), t ≥ 0}, which satisfies the following
necessary conditions (for the sufficiently small Δ(t))

1. ξ (0) = 0,
2. P(ξ (t + Δ(t)) − ξ (t) = 1) = λ(t)Δ(t) + o (Δ(t)),
3. P(ξ (t + Δ(t)) − ξ (t) ≥ 2) = o (Δ(t)).

To be more precise, a GCP is a counting process ξ (t), with ξ (0) = 0 such that, for any interval
( t1 , t2 ],

1  Λ ( t2 ) − Λ ( t1 )  k
P ( ξ ( t2 ) − ξ ( t1 ) = k ) = , k = 0, 1, . . . , (1)
1 + Λ ( t2 ) − Λ ( t1 ) 1 + Λ ( t2 ) − Λ ( t1 )

where Λ(t) = E(ξ (t)) is the mean value function (MVF) of the process. It is usually assumed that
Λ(t) is a smooth function in the sense that there exists a function λ(t) such that λ(t) = dΛ(t)/dt. The
function λ(t) is called the intensity function of the process. We have to mention here that, as noted
by Cha and Finkelstien [14], the NHPP also lies in the class of counting process satisfying (i)–(iii),
with an additional property that the increments of the process are independent. The motivation of
using the GCP, in comparison with NHPP, is natural in some practical situations as we mention in the
following. The GCP model, like the NHPP model, has a simple form and easy to handle mathematical
characteristics. In an NHPP model, the increments of the process are independent, while, in the GCP
model, the increments of the process have positive dependence. In practice, there are situations in
which there is positive dependence of increments in a process that occurs naturally. For instance,
assume that the components of a railway network destroyed by an earthquake that occurs according
to a counting process. Then, the probability of the next earthquake is often higher if the previous
earthquake has happened recently, compared with the situation that it happened earlier (see [14]).
Furthermore, the NHPP has a limitation that the mean and the variance of the process are equal,
i.e., E(ξ (t)) = Var (ξ (t)), while, in GCP, the variance of the process is always greater than the mean,
i.e., Var (ξ (t)) > E(ξ (t)). This property of the GCP makes it cover many situations that can not be
described and covered by the NHPP. For more details on recent mathematical developments and
applications of the GCP model, see [14,15].
The reminder of the paper is arranged as follows. In Section 2, we first give the well-known
concept of signature of a network. Then, we consider a two-state network that consists of n components.
We assume that the components of the network fail according to a GCP. We obtain some mixture
representations for the reliability of the network based on the signatures. Several aging and stochastic
properties of the network are explored. Among others, conditions are investigated under which the
monotonicity of the intensity function of the process of component failure implies the monotonicity of

257
Mathematics 2018, 6, 197

the network hazard rate. The reliabilities of the lifetimes of the different networks, subjected to the
same or different GCPs, are compared based on the stochastic order between the associated signature
vectors. We also study the stochastic properties of the residual lifetime of the network where the
components fail based on a GCP. Section 3 is devoted to the reliability assessment of the single-step
three-state networks. Recall that a network is said to be single-step if the failure of one component
changes the network state at most by one. First, we give the notion of a two-dimensional signature
associated with three-state networks. Then, we consider an n-component network and assume that the
network has three states up, partial performance and down. We again assume that the components of the
network are subjected to failure on the basis of GCP, which results in the change of network states.
Under these conditions, we obtain several stochastic and dependency characteristics of the networks
based on the two-dimensional signature. Several examples and plots are also provided throughout the
article for illustration purposes.
Before giving the main results of the paper, we give the following definitions that are useful
throughout the paper. For more details, see [16].

Definition 1. Let X and Y be two random variables (RVs) with survival functions F̄X and F̄Y , probability
density functions (PDFs) f X and f Y , hazard rates h X and hY , and reversed hazard rates r X and rY , respectively:

• X is said to be smaller than Y in the usual stochastic order (denoted by X ≤st Y) if F̄X ( x ) ≤ F̄Y ( x ) for
all x.
• X is said to be smaller than Y in the hazard rate order (denoted by X ≤hr Y) if h X ( x ) ≥ hY ( x ) for all x.
• X is said to be smaller than Y in the reversed hazard rate order (denoted by X ≤rh Y) if r X ( x ) ≤ rY ( x ) for
all x.
• X is said to be smaller than Y in the mean residual life order (denoted by X ≤mrl Y) if E( X − x | X > x ) ≤
E(Y − x |Y > x ) for all x.
• X is said to be smaller than Y in likelihood ratio order (denoted by X ≤lr Y) if f Y ( x )/ f X ( x ) is an
increasing function of x.

It can be shown that, if X ≤lr Y, then X ≤hr Y and X ≤rh Y. In addition, X ≤hr Y implies
X ≤mrl Y and X ≤st Y.

Definition 2. Let X and Y be two random vectors with survival functions F̄X and F̄Y , respectively.

• X is said to be smaller than Y in the upper orthant order (denoted by X ≤uo Y) if F̄X (x) ≤
F̄Y (x) for all x ∈ Rn .
• X is said to be smaller than Y in the usual stochastic order (denoted by X ≤st Y) if E(ρ(X)) ≤ E(ρ(Y))
for every increasing function ρ(·) for which the expectations exist.

Definition 3.

• The nonnegative function g(x) is called multivariate totally positive of order 2 (MTP2 ) if g(x) g(y) ≤
g(x ∨ y) g(x ∧ y), for all x, y ∈ Rn , where x ∧ y = (min{ x1 , y1 }, . . . , min{ xn , yn }) and x ∨ y =
(max{ x1 , y1 }, . . . , max{ xn , yn }).
• The RVs X and Y are said to be positively quadrant dependent (PQD) if, for every pair of increasing
functions ψ1 ( x ) and ψ2 ( x ),
Cov(ψ1 ( X ), ψ2 (Y )) ≥ 0.
• The RVs X and Y are said to be associated if for every pair of increasing functions ψ1 ( x, y) and ψ2 ( x, y),
Cov(ψ1 ( X, Y ), ψ2 ( X, Y )) ≥ 0.

In a special case when n = 2, the MTP2 is known as totally positive of order 2 (TP2 ).

258
Mathematics 2018, 6, 197

2. Two-State Networks under GCP of Component Failure

In the reliability engineering literature, several approaches have been employed to assess the
reliability of networks and systems. Among various ways that are considered to explore the reliability
and aging properties of the networks, an approach is based on the notion of signature (or D-spectrum).
The concept of signature, which depends only on the network design, has proven very useful in the
analysis of the networks performance particularly for comparisons between networks with different
structures. Consider a network (system) that consists of n components. The signature associated with
the network is a vector s = (s1 , s2 , . . . , sn ), in which the ith element shows the probability that the ith
component failure in the network causes the network failure, under the condition that all permutations
of order of components failure are equally likely. In other words, the ith element si is equal to si = ni /n!,
i = 1, . . . , n, where ni is the number of permutations in which the ith component failure changes
the network state from up to down. For more details on signatures and their applications in the
study of system reliability, see, for example, Refs. [17–20] and references therein. In this section, we
give a signature-based mixture representation for the reliability of the network under the condition
that the components of the network fail according to a GCP {ξ (t), t ≥ 0} with MVF Λ(t). We have
from Equation (1)

1  Λ(t) k
P(ξ (t) = k) = , k = 0, 1, . . . .
1 + Λ(t) 1 + Λ(t)

Then, the survival function of the kth arrival time of process, ϑk , is given as

F̄ϑk (t) := P(ϑk > t) = P(ξ (t) < k)

 Λ(t) k
=1 − , k = 0, 1, . . .
1 + Λ(t)

and the PDF of the kth arrival time ϑk is achieved as

λ(t)  Λ(t) k
f ϑk ( t ) = k , k = 0, 1, . . . .
Λ(t)(1 + Λ(t)) 1 + Λ(t)

Let T denote the lifetime of a network with n components. The components of network are
subjected to failure based on a GCP with MVF Λ(t). From the reliability modeling proposed by
Zarezadeh and Asadi [12], the reliability of the network lifetime, denoted by F̄T , is represented as
n   Λ ( t ) i 
F̄T (t) = ∑ si 1 − , t > 0, (2)
i =1
1 + Λ(t)

or equivalently as

n −1 
1 Λ ( t ) i
F̄T (t) =
1 + Λ(t) ∑ S̄i 1 + Λ(t)
, t > 0, (3)
i =0

where S̄i = ∑nk=i+1 sk is the survival signature of the network. Then, the PDF of T is obtained as

λ(t) n  Λ ( t )  i −1
f T (t) =
(1 + Λ(t))2 ∑ isi 1 + Λ(t)
, t > 0, (4)
i =1

259
Mathematics 2018, 6, 197

where λ(t) = dΛ(t)/dt. In addition, the hazard rate of network lifetime is given as follows:
 Λ(t) i
λ(t) ∑in=1 isi 1+Λ(t)
hT (t) =
Λ(t)(1 + Λ(t)) ∑n s #1 −  Λ(t) i $
.
i =1 i 1+ Λ ( t )
Λ(t) i
λ(t) ∑in=1 isi 1+Λ(t)
=  Λ(t) i , t > 0.
Λ(t) ∑in=−01 S̄i 1+Λ(t)

With hϑk (t) = f ϑk (t)/ F̄ϑk (t) as the hazard rate of the kth arrival time of the GCP, it can be seen
that the hazard rate of network lifetime can be also written as

∑nk=1 sk f ϑk (t) n
hT (t) =
∑nk=1 sk F̄ϑk (t)
= ∑ s k ( t ) h ϑk ( t ) , (5)
k =1

which is a mixture representation with mixing probability vector s(t) = (s1 (t), . . . , sn (t)) where,
for k = 1, . . . , n,
# Λ(t) $
s k 1 − ( 1+ Λ ( t ) ) k
sk (t) = # Λ(t) $ (6)
∑nk=1 sk 1 − ( 1+Λ(t) )k
# Λ(t) $
s k 1 − ( 1+ Λ ( t ) ) k
= Λ(t)
. (7)
∑in=−01 S̄i (1 − )( Λ(t) )i
1+ Λ ( t ) 1+ Λ ( t )

One can easily show that sk (t) is the probability that the lifetime of system is equal to the kth
arrival time of the process given that the network lifetime is greater than t.
Let us look at the following example.

Example 1. Consider a network that consists of six nodes and 10 links with the graph depicted in Figure 1.

9 a
c

3 1

4 10
6 b
2 5
d f
7 8
e

Figure 1. A network with 10 links, and six nodes

The network is assumed to work if there is a connection between some of nodes which we consider them as
terminals. We consider two different sets of terminals for the network:

• First, we consider all nodes as terminals, T = V (all-terminal connectivity). In this case, we can show that
the corresponding signature vector of a network is as follows:

1 9 29 65
s = (0, 0, , , , , 0, 0, 0, 0).
30 70 90 126

260
Mathematics 2018, 6, 197

• Second, assume that the network is working if there is a connection between nodes c and f . That is, the
terminals set is T = {c, f }. In this case, the signature vector is obtained as

1 37 179 379 19 1 1
s = (0, 0, , , , , , , , 0).
120 840 1260 1260 70 6 15
An algorithm for calculating these signatures is available from the authors upon the request.
Assume that, in each case, the network is subjected to failure based on the GCPs with the same MVFs.
Denote the lifetimes of the network corresponding to (a) and (b) by T and T  , respectively. Comparing the
corresponding survival signatures of network for two cases shows that S̄i ≤ S¯i , for i = 0, 1, . . . , 9, and
hence based on (3) we have T ≤st T ∗ implying that the network with two-terminal connectivity is more
reliable than the network with all-terminal connectivity, as expected intuitively.

Figure 2a gives the plot of network reliability in the case of all-terminal connectivity, T = V, and when
Λ(t) = t a , for different values of a. As seen, the reliability function of network does not order with respect to a
for all t > 0. Of course, this is true in any network when MVF Λ(t) = t a , a > 0. This is so using the fact that
  t a k
P(ϑk > t; a) = 1 −
1 + ta

is increasing (decreasing) in a for 0 < t < 1 (t > 1), for a general signature vector s the reliability of the network
n
P( T > t; a) = ∑ sk P(ϑk > t; a),
k =1

is also increasing (decreasing) in a for 0 < t < 1 (t > 1). Figure 2b represents the hazard rate of network when
T = V for a = 0.5, 1, 1.5. Figure 3a,b shows the plots of reliability function and hazard rate of the network
lifetime when the terminals set is considered as T = {c, f }.

(a) (b)

Figure 2. The plots of (a) the reliability function; (b) the hazard rate of network lifetime when the
terminals set is T = V.

261
Mathematics 2018, 6, 197

(a) (b)

Figure 3. The plots of (a) the reliability function; (b) the hazard rate of network lifetime when the
terminals set is T = {c, f }.

It is interesting to compare the network reliability when the failure of components appear
according to a GCP and the network reliability when the failure of components occur based on
an NHPP. In the sequel, we show that, if the network has a series structure, then the reliability of the
network in the GCP model dominates the reliability of the network in an NHPP model. Consider a
two-state series network with the property that the first and the last components are considered to be
terminals. We assume that the network fails if the linkage between the two terminals are disconnected.
This occurs at the time of the first component failure. Let TNP and TGP denote the lifetimes of the
network when the component failure appears according to NHPP and GCP with the same MVF Λ(t),
respectively. If ϑ1,NP and ϑ1,GP denote the arrival times of the first component failure based on NHPP
and GCP, respectively, then, from inequality e x > (1 + x ), x > 0, we can write

1
P(ϑ1,NP > t) = e−Λ(t) < = P(ϑ1,GP > t).
1 + Λ(t)

Hence, based on the fact that, for a series network s = (1, 0, . . . , 0), relation (2) implies that
TNP ≤st TGP .
The following example reveals that the above result, proved for the series network, is not
necessarily true for any network.

Example 2. Consider the network described in Example 1. Figure 4 shows that the reliability functions of the
network for part (a). As the plots show the reliability functions are not ordered in NHPP and GCP models with
the same MVFs Λ(t) = t. The reliability of the network in the NHPP model is higher than the GCP model for
the early times of operating of the network. However, when the time goes ahead, the network reliability in NHPP
declines rapidly and stays below the reliability of the GCP model.

262
Mathematics 2018, 6, 197

Figure 4. The reliability of network in Example 2.

The next theorem explores the monotonicity relation of the intensity function of the process and
the hazard rate of the network.

Theorem 1. Let the components of a network fail based on a GCP with increasing intensity function λ(t).
Then, the hazard rate of network is increasing if and only if ψ(u) is increasing in u where

−1
∑kn= 0 ( k + 1) s k +1 u
k
ψ(u) = −1
. (8)
∑kn= 0 S̄k u
k

Proof. From (3) and (4), the hazard rate of network can be written as

f T (t)
hT (t) = = λ(t)ψ(μ(t)), (9)
F̄T (t)

where ψ(·) is defined in (8) and μ(t) = Λ(t)/(1 + Λ(t)) is increasing in t. If ψ(t) is increasing, from (9),
it can be easily seen that h T (t) is increasing. This completes the ‘if’ part of the theorem. To prove ‘only
if’ part of the theorem, let h T (t) be increasing and ψ(t) be decreasing in the interval ( a, b). For MVF
Λ(t) = ct, c > 0, and hence λ(t) = c as an increasing function, we conclude that h T (t) is decreasing
on interval ( a, b), which contradicts with the assumption that the hazard rate of network is increasing
for all t.

Theorem 2. Let T and T  denote the lifetimes of two networks with signature vectors s = (s1 , . . . , sn ) and
s = (s1 , . . . , sn ), respectively. Suppose that the components of networks fail based on GCPs with MVFs Λ(t)
and Λ (t), respectively. If s ≤st s and Λ(t) ≥ Λ (t) for all t ≥ 0, then T ≤st T  .

Proof. Let ϑk and ϑk denote the kth arrival times of the two processes, k = 1, ..., n. Since P(ϑk ≤
ϑk+1 ) = 1, then, from Theorem 1.A.1 of [16], we can write ϑk ≤st ϑk+1 . Hence,
n n
P( T > t) = ∑ sk F̄ϑk (t) ≤ ∑ sk F̄ϑk (t). (10)
k =1 k =1

263
Mathematics 2018, 6, 197

In addition, the condition Λ(t) ≥ Λ (t) for all t ≥ 0 implies that ϑk ≤st ϑk , k = 1, ..., n. Then,
we get
n n
∑ sk F̄ϑk (t) ≤ ∑ sk F̄ϑk (t) = P(T  > t). (11)
k =1 k =1

Hence, the result follows from (10) and (11).

Theorem 3. For two networks as described in Theorem 2, assume that the failure of components of both networks
appear according to the same GCPs:
1. If s ≤st s , then T ≤st T  ,
2. If s ≤hr s , then T ≤hr T  ,
3. If s ≤rh s , then T ≤rh T  ,
4. If s ≤lr s , then T ≤lr T  .

Proof. It can be easily shown that ϑk ≤lr ϑk+1 k = 1, 2, . . . . Since lr-ordering implies hr-, rh- and
st-ordering, parts (i), (ii), (iii), and (iv) are proved, by using (2), from Theorems 1.A.6, 1.B.14, 1.B.52,
and 1.C.17 of [16], respectively.

From part (ii) of Theorem 3, since hr-ordering implies mrl-ordering, we conclude that if s ≤hr s
then T ≤mrl T  . However, the following example shows that the assumption s ≤hr s can not be
replaced with s ≤mrl s to have T ≤mrl T  .

Example 3. Consider two networks with signature vectors s = (0, 2/3, 1/3) and s = (1/3, 0, 2/3). It is
easy to see that s ≤mrl s . However, s st s and hence s hr s . Assume that the components of both
networks fail based on the same GCPs with MVF Λ(t) = t2 . Then, a straightforward calculation gives
E( T ) = 2.5525 > E( T  ) = 2.4871, which, in turn, implies that T mrl T  .

2.1. Residual Lifetime of a Working Network

Let T denote the lifetime of a network whose components are subjected to failure based on a GCP
with MVF Λ(t). If the network is up at time t, then the residual lifetime of the network is presented by
the conditional RV ( T − t| T > t) with conditional reliability function given as
n
1
P( T − t > x | T > t) =
P( T > t) ∑ s k P ( ϑk > t + x )
k =1
n
= ∑ s k ( t ) P ( ϑk − t > x | ϑk > t ) ,
k =1

where sk (t) is the kth element of vector sk (t) as defined in (6). This shows that the reliability function
of the residual lifetime of the network is a mixture of the reliability functions of residual lifetimes of
the first n arrival times of GCP, where the mixing probability vector is s(t) = (s1 (t), . . . , sn (t)). As
we have already mentioned, sk (t) is in fact the probability that the kth component failure causes the
failure of the network, given that the lifetime of the network is more than t; that is,

sk (t) = P( T = ϑk | T > t), k = 1, 2, ..., n.

In what follows, we call the vector s(t) as the conditional signature of the network. In the sequel,
we give some stochastic properties of the conditional signature of network under the condition that
the components of the network fail based on GCP model.

Theorem 4. Consider a network with signature vector s = (s1 , s2 , . . . , sn ). With M = max{i |si > 0},
limt→0 s(t) = s and limt→∞ s(t) = s̃ where s̃ = (s̃1 , . . . , s̃ M ) and s̃i = ∑n isi ks , i = 1, . . . , n.
k =1 k

264
Mathematics 2018, 6, 197

Proof. From (6), we can write

# Λ(t) $
s i 1 − ( 1+ Λ ( t ) ) i
si ( t ) = # Λ(t) $, i = 1, . . . , n.
∑kM=1 sk 1 − ( 1+Λ(t) )k

Then, it is easily seen that, for any i, limt→0 si (t) = si and hence

lim s(t) = s.
t →0

On the other hand, for i = 1, . . . , n,

s i (1 − u i )
lim si (t) = lim ,
t→∞ u →1 ∑nk=1 sk (1 − uk )
isi ui−1 is
= lim = n i ,
u →1 ∑nk=1 ksk uk−1 ∑k=1 ksk

and consequently the result follows.

Example 4. For the network in Example 1, with Λ(t) = t a , we have

akt ak
h ϑi ( t ) = # $, i = 1, 2, . . . .
(1 + t a ) (1 + t a )k − t ak

Hence, it is easily seen that limt→∞ hϑi (t) = 0. Thus, based on (5) and Theorem 4, we have
limt→∞ h T (t) = 0 for any network structure.

Theorem 5. Consider a network whose components fail based on two different GCPs with MVFs Λ(t)
and Λ (t), respectively. Denote by s(t) and s (t) the corresponding conditional signatures of the two
networks. Then,

• s(t) and s (t) are increasing in the sense of st-ordering with respect to t;
• Λ(t) ≤ Λ (t) implies that s(t) ≤st s (t) for all t ≥ 0.

Proof. With δ(i, u) = 1 − ui , from (6), for each MVF Λ(t), si (t) can be written as

si δ(i, μ(Λ(t)))
si ( t ) = , i = 1, . . . , n, (12)
∑in=1 si δ(i, μ(Λ(t))

where μ(v) = v/(1 + v). Then, we have

n
∑ si (t) = βk (μ(Λ(t))),
i=k

in which

∑in=k si [1 − ui ]
β k (u) = .
∑in=1 si [1 − ui ]

265
Mathematics 2018, 6, 197

For i1 ≤ i2 ,
i −1
δ ( i 2 , u ) 1 − u i2 ∑ j2=0 u j
= = i −1
δ ( i1 , u ) 1 − u 1i
∑ j1=0 u j
i −1 i −1
∑ j1=0 u j + ∑ j2=i1 u j
= i −1
∑ j1=0 u j
i2 −1
1
=1 + ∑ i1 −1 
(13)
j = i1 ∑ j =0
 u j− j

is increasing in u and hence δ(i, u) is TP2 in i and u. Using this fact and Lemma 2.4 of [13]:

• Since μ(u) is increasing in u, then, for an arbitrary MVF Λ(·), we have

n n
∑ s i ( t ) ≤ ∑ s i ( t  ), t ≤ t .
i=k i=k

• Since μ(u) is increasing in u, and Λ(t) ≤ Λ (t) for all t ≥ 0, then

n n
∑ si (t) ≤ ∑ si (t), t ≥ 0.
i=k i=k

Therefore, the proof of the theorem is complete.

The following lemma from [13] is useful to get some stochastic properties of conditional signature
expressed in (6). Before expressing the lemma, we recall that a non-negative function f ( x ), x ≥ 0, is
said to be upside-down bathtub-shaped if it is increasing on [0, a], is constant on [ a, b] and is decreasing
on [b, ∞) where 0 ≤ a ≤ b ≤ ∞.

Lemma 1. Let α(.) and β(.) be non-negative discrete functions and γ be positive and real-valued. Define

γ ∑ik=l  α(i )ui
τ (u) = , u > 0, (14)
∑ik=l α(i ) β(i )ui

where l ≤ l  , l  < k and k ≤ k. Assume that β(.) is a non-constant decreasing (increasing) function on
{l  , l  + 1, ..., k }. Then, for l = l  and k < k(k = k and l  > l), we have

1. τ (u) is upside-down bathtub-shaped with a single change-point;

2. τ (u) is bounded above by γ/β(k ) (γ/β(l  )).

Now, we have the following theorem.

Theorem 6. For a network with signature vector s = (s1 , . . . , sn ),

1. sm (t) is decreasing in t and s M (t) is increasing in t where m = min{i |si > 0} and M = max{i |si > 0};
2. s j (t), m ≤ j ≤ M, is upside-down bathtub-shaped with a single change-point;
3. s j (t), m ≤ j ≤ M, is bounded above by s j /S̄ j−1 ;
4. The maximum value of s j (t), m ≤ j ≤ M, does not depend on the MVF Λ(t).

Proof. Assume that μ(t) = t/(1 + t). From (12), we can write

 M
δ(i, μ(Λ(t))) −1  M
δ(i, μ(Λ(t))) −1
sm (t) = sm ∑ si δ(m, μ(Λ(t))) s M (t) = s M ∑ si δ( M, μ(Λ(t))) . (15)
i=m i=m

266
Mathematics 2018, 6, 197

As seen in (13), δ(i2 , x )/δ(i1 , x ), for i2 ≥ i1 , is increasing in x. Since μ(Λ(t)) is increasing in t, then
it can be concluded that δ(i2 , μ(Λ(t)))/δ(i1 , μ(Λ(t))) is also increasing in t, for i2 ≥ i1 . Based on this
fact and (15), we observe that sm (t) is a decreasing function of t and s M (t) is an increasing function of
t. This completes the proof of part (a).
From relation (7), we have

j −1  i
s j ∑i=0 μ(Λ(t))
s j (t) =  i .
∑in=−01 S̄i μ(Λ(t))

With γ = s j , α(i ) = 1, β(i ) = S̄i , l = l  = 0, k = j − 1 and k = n − 1 in (14), define ω j (u) := τ (u).

Then, we can write
  
s j (t) = ω j μ Λ(t) .
 
Since μ Λ(t) is increasing in t, parts (b) and (c) follow from parts (i) and (ii) of Lemma 1.
Part (d) can be proved from the fact that
  
max s j (t) = max ω j μ Λ(t) = max ω j (t).
t >0 t >0 t >0

The following theorem compares the performance of two used networks based on their
conditional signatures.

Theorem 7. Let s(t) and s (t) be the conditional signatures of two networks with lifetimes T and T  ,
respectively. Suppose that the component failure in both networks appear based on the same GCPs.

1. If s(t) ≤st s (t), then ( T − t| T > t) ≤st ( T  − t| T  > t);

2. If s(t) ≤hr s (t), then ( T − t| T > t) ≤hr ( T  − t| T  > t);
3. If s(t) ≤rh s (t), then ( T − t| T > t) ≤rh ( T  − t| T  > t);
4. If s(t) ≤lr s (t), then ( T − t| T > t) ≤lr ( T  − t| T  > t).

Proof. It can be easily seen that ϑk ≤lr ϑk+1 , k = 1, . . . , n − 1. Then, from Theorem 1.C.6 of [16], we
have, for k = 1, . . . , n − 1 and t ≥ 0,

(ϑk − t|ϑk > t) ≤lr (ϑk+1 − t|ϑk+1 > t).

Hence, these residual lifetimes are also hr-, rh- and st-ordered. Since s(t) ≤st s (t), from
Theorem 1.A.6 of [16], we have, for all x > 0,
n
P( T − t > x | T > t) = ∑ s k ( t ) P ( ϑk − t > x | ϑk > t )
k =1
n
≤ ∑ sk (t) P(ϑk − t > x|ϑk > t) = P(T  − t > x|T  > t).
k =1

This establishes part (a). The proof of parts (b), (c) and (d) are obtained similarly by using
Theorems 1.B.14, 1.B.52, and 1.C.17 of [16], respectively.

3. Three-State Networks under GCP of Component Failure

In this section, we study the reliability of the lifetimes of the networks with three states under
the condition that the components fail according to a GCP with MVF Λ(t) = λt. In order to develop
the results, we need the notion of two-dimensional signature that has been defined for single-step
three-state networks by Gertsbakh and Shpungin [11]. Throughout this section, we are dealing with a
single-step three-state network consisting of n binary components where we assume that the network

267
Mathematics 2018, 6, 197

has three states: up (denoted by K = 2), partial performance (denoted by K = 1) and down (denoted by
K = 0). Suppose that the network starts to operate at time t = 0 where it is in state K = 2. Denote by
T1 the time that the network remains in state K = 2 and by T2 the network lifetime i.e., the entrance
time into state K = 0. Let I (J) be the number of failed components when the network enters into state
K = 1 (K = 0). Gertsbakh and Shpunging [11] introduced the notion of two-dimensional signature as
ni,j
si,j = P( I = i, J = j) = , 1 ≤ i < j ≤ n, (16)
n
where ni,j represents the number of permutations in which the ith and the jth components failure
change the network states from K = 2 to K = 1 and from K = 1 to K = 0, respectively. We denote by
matrix S the two-dimensional signature with elements defined in (16). In the following, we first obtain
the joint reliability function of (T1 , T2 ). Under the assumption that all orders of components failure are
equally probable, we have
n n
P(T1 > t1 , T2 > t2 ) = ∑ ∑ P( I = i, J = j) P(ξ (t1 ) < i, ξ (t2 ) < j| I = i, J = j)
i =0 j = i +1
n n
= ∑ ∑ si,j P(ξ (t1 ) < i, ξ (t2 ) < j),
i =0 j = i +1

in which the second equality follows from the fact that the event { I = i, J = j} depends only on the
network structure and does not depend on the mechanism of the components failure. In addition, it
can be shown, by changing the order of summations, that
n n
P(T1 > t1 , T2 > t) = ∑ ∑ S̄i,j P(ξ (t1 ) = i, ξ (t2 ) = j), (17)
i =0 j = i

−1
where S̄i,j = ∑kn= n
i +1 ∑ j=max{k,j}+1 sk,l .
Suppose that the component failures occur at random times ϑ1 , . . . , ϑn that are corresponding to
the first n arrival times of the GCP {ξ (t), t ≥ 0}. Using the fact that the event (ξ (t) = i ) occurs if and
only if (ϑi ≤ t < ϑi+1 ), it can be shown that
n n
P(T1 > t1 , T2 > t2 ) = ∑ ∑ si,j P(ϑi > t1 , ϑ j > t). (18)
i =0 j = i +1

Assuming that the MVF of the GCP is Λ(t) = λt, Di Crescenzo and Pellerey [15] obtained the
PDF of (ϑ1 , . . . , ϑn ) as

n!λn
f ϑ1 ,...,ϑn (t1 , . . . , tn ) = , 0 < t1 < · · · < t n . (19)
(1 + λtn )n+1

Using (19), the joint PDF of ϑi and ϑ j is achieved as

j!λ j tii−1 (t j − ti ) j−i−1

f ϑi ,ϑj (ti , t j ) = , 0 < ti < t j . (20)
(1 + λt j ) j + 1 ( i − 1) ! ( j − i − 1) !

In the following, we present an example of a three-state network whose components fail according
to a GCP with MVF Λ(t) = λt.

Example 5. Consider a network with a graph as depicted in Figure 5. The network has 14 links and eight nodes
in which the dark nodes are considered to be terminals. Suppose that the nodes are absolutely reliable and the
links are subjected to failure.

268
Mathematics 2018, 6, 197

Figure 5. The network with eight nodes, and 14 links.

Assume that the network is in up state if all terminals are connected, in partial performance state if
two terminals are connected and in down state if all terminals are disconnected. Let the network components
fail according to a GCP with intensity function λ(t) = 1 and all orders of links failure are equally likely.
Figure 6 presents the plot of joint reliability function of the network lifetimes (T1 , T2 ). The elements si;j of the
two-dimensional signature are calculated using an algorithm by the authors, which can be provided to the readers
upon the request.

Figure 6. The joint reliability function of (T1 , T2 ) in Example 5.

In the following theorem, we compare the state lifetimes of two three-state networks. In order to
do this, we need the following Lemma.

Lemma 2. Assume that {ξ 1 (t), t ≥ 0} and {ξ 2 (t), t ≥ 0} are two GCPs with intensity functions λ1 (t) =
λ1 > 0, and λ2 (t) = λ2 > 0, respectively. Let ϑ1,1 , ϑ1,2 , . . . and ϑ2,1 , ϑ2,2 , . . . denote the arrival times
corresponding to the two processes, respectively. If λ1 ≥ λ2 , then (ϑ1,1 , . . . , ϑ1,n ) ≤st (ϑ2,1 , . . . , ϑ2,n ) for every
n ≥ 1.

Proof. Using relation (19), it can be seen that f ϑi,1 ,...,ϑi,n is MTP2 , which implies that ϑi,1 , . . . , ϑi,n are
associated, i = 1, 2. In addition, we have

f ϑ2,1 ,...,ϑ2,n (t1 , . . . , tn ) λ2 n 1 + λ1 t n n +1

=( ) ( ) , 0 < t1 < · · · < t n ,
f ϑ1,1 ,...,ϑ1,n (t1 , . . . , tn ) λ1 1 + λ2 t n

which is increasing in (t1 , . . . , tn ). Therefore, the required result is concluded from Theorem 6.B.8
of [16].

269
Mathematics 2018, 6, 197

Theorem 8. Consider two three-state networks that each consist of n components having signature matrices
S1 and S2 , respectively. Let the components of ith network fail according to GCP {ξ i (t), t ≥ 0}, i = 1, 2 with
intensity function Λi (t) = λi . Let ϑi,1 , ϑi,2 , ... denote the arrival times corresponding to {ξ i (t), t ≥ 0}. Suppose
(1) (1) (2) (2)
that (T1 , T2 ) and (T1 , T2 ) are the corresponding state lifetimes of the two networks, respectively:
(1) (1) (2) (2)
• If λ1 ≥ λ2 and S1 ≤uo S2 , then (T1 , T2 ) ≤uo (T1 , T2 ),
(1) (1) (2) (2)
• If λ1 ≥ λ2 and S1 ≤st S2 , then (T1 , T2 ) ≤st (T1 , T2 ).

Proof. From Lemma 2, if λ1 ≥ λ2 , then (ϑ1,1 , . . . , ϑ1,n ) ≤st (ϑ2,1 , . . . , ϑ2,n ), which implies
(ϑ1,i , ϑ1,j ) ≤st(uo) (ϑ2,i , ϑ2,j ) for all 1 ≤ i < j ≤ n.

• Using representation (17), we have

n n
∑ ∑ S̄1,i,j P(ξ 1 (t1 ) = i, ξ 1 (t2 ) = j)
(1) (1)
P(T1 > t1 , T2 > t2 ) =
i =0 j = i
n n
≤ ∑ ∑ S̄2,i,j P(ξ 1 (t1 ) = i, ξ 1 (t2 ) = j)
i =0 j = i
n n
= ∑ ∑ s2,i,j P(ϑ1,i > t1 , ϑ1,j > t2 )
i =0 j = i +1
n n
≤ ∑ ∑ s2,i,j P(ϑ2,i > t1 , ϑ2,j > t2 )
i =0 j = i +1
(2) (2)
= P(T1 > t1 , T2 > t2 ),

where the first inequality follows from the assumption that S1 ≤uo S2 and the second inequality
follows from (ϑ1,i , ϑ1,j ) ≤uo (ϑ2,i , ϑ2,j ) for 1 ≤ i < j ≤ n.
• Using the fact that (ϑ1,i , ϑ1,j ) ≤st (ϑ2,i , ϑ2,j ) for all i < j and the assumption that S1 ≤st S2 , the
required result is concluded from Theorem 3.3 of [21].

In the sequel, we investigate the dependency between T1 and T2 based on the dependency
between RVs I and J. In fact, we show that, if I and J are PQD (associated), then T1 and T2 are also
PQD (associated). Before that, let
(1) (2)
si = P ( I = i ), sj = P ( J = j ),

(k) (k)
and S̄i = ∑nl=i+1 sl , k = 1, 2.

Theorem 9. Let T1 be the lifetime of a three-state network in state K = 2 and T2 be the lifetime of the network.
Let the components failure of the network appear according to the GCP {ξ (t), t ≥ 0} with arrival times
ϑ1 , . . . , ϑn .

• If I and J are PQD, then T1 and T2 are PQD.

• If I and J are associated, then T1 and T2 are associated.

Proof. From representation (20), one can show that f ϑi ,ϑj (ti , t j ) is TP2 , which implies that ϑi and ϑ j are
associated and PQD.

270
 Mathematics 2018, 6, 197

(a) Let φ(·) and ψ(·) be two increasing functions. From representation (18), we have

n −1 n
E(φ(T1 )ψ(T2 )) = ∑ ∑ si,j E(φ(ϑi )ψ(ϑ j ))
i =1 j = i +1
n −1 n
≥ ∑ ∑ si,j E(φ(ϑi )) E(ψ(ϑ j ))
i =1 j = i +1
n n
∑ si E(φ(ϑi )) ∑ s j E(ψ(ϑ j ))
(1) (2)
≥
i =1 j =1
= E(φ(T1 )) E(ψ(T2 )),

where the first inequality follows from the fact that ϑi and ϑ j are PQD and the second inequality
follows from the assumption that I and J are PQD.
(b) Proof of part (b) is the same as the proof of part (a) using the fact that, for every two-variate
increasing functions φ (·, ·) and ψ (·, ·),

n −1 n
E(φ (T1 , T2 )ψ (T1 , T2 )) = ∑ ∑ si,j E(φ (ϑi , ϑ j )ψ (ϑi , ϑ j )).
i =1 j = i +1

The results of the theorem are interesting in the sense that the PQD (associated) property of I and
J, which is non-aging and depends only on the structure of the network, is transferred to the PQD
(associated) property of T1 and T2 , which is the aging characteristic of the network.

Example 6. Consider again the network presented in Example 5. It can be seen that, for every i, j = 1, . . . , 14,
(1) (2)
S̄i,j ≥ S̄i S̄ j . This implies that I and J are PQD. Hence, if the components fail according to a GCP, then T1
and T2 are PQD.

4. Conclusions
In this article, we studied the reliability, aging and stochastic characteristics of an n-component
network whose components were subjected to failure according to a geometric counting process
(GCP). We first considered the case that the network has two states (up and down). Some mixture
representations of the network reliability were obtained in terms of signature of the network and
the reliability functions of the arrival times of the GCP. We studied the conditions under which the
hazard rate of the network is increasing in the case that intensity function of the process is increasing.
Stochastic comparisons were made between the lifetimes of different networks, subjected to GCPs,
based on the stochastic comparisons between their signatures. The residual lifetime of the network
was also explored. In the second part of the paper, we considered the networks with three states: up,
partial performance, and down. The components of the network were assumed to fail based on a
GCP with mean value function Λ(t) = λt, which leads to the change of the network states. Under
these circumstances, we arrived at several stochastic and dependency properties of the networks with
the same and different structures. The results of Section 3 were obtained under the special case that
the MVF of the GCP is Λ(t) = λt. The developments of the paper were mainly dependent on the
notions of signature and two-dimensional signature. The results of the paper were illustrated by
several examples.

Author Contributions: The collaboration of the authors in conceptualization, methodology and formal analysis
of the paper are in the order of names appeared in the article.
Funding: Asadi’s research work was carried out in the IPM Isfahan branch and was in part supported by a grant
from IPM (No. 96620411).

271
Mathematics 2018, 6, 197

Acknowledgments: The authors would like to thank the editor and two anonymous reviewers for their comments
and suggestions that led to improving the exposition of the article.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Ashrafi, S.; Asadi, M. On the stochastic and dependence properties of the three-state systems. Metrika 2015,
78, 261–281. [CrossRef]
2. Ashrafi, S.; Asadi, M. Dynamic reliability modeling of three-state networks. J. Appl. Prob. 2014, 51, 999–1020.
[CrossRef]
3. Ashrafi, S.; Zarezadeh, S. A Shock-Based Model for the Reliability of Three-State Networks. IEEE Trans.
Reliab. 2018, 67, 274–284. [CrossRef]
4. Eryilmaz, S. Mean residual and mean past lifetime of multi-state systems with identical components. IEEE
Trans. Reliab. 2010, 59, 644–649. [CrossRef]
5. Eryilmaz, S.; Xie, M. Dynamic modeling of general three-state k-out-of-n:G systems: Permanent based
computational results. J. Comput. Appl. Math. 2014, 272, 97–106. [CrossRef]
6. Huang, J.; Zuo, M.J.; Wu, Y. Generalized multi-state k-out-of-n:G systems. IEEE Trans. Reliab. 2000, 49,
105–111. [CrossRef]
7. Lisnianski, A.; Levitin, G. Multi-State System Reliability: Assessment, Optimizationand Applications;
World Scientific Publishing: Singapore, 2003.
8. Tian, Z.; Yam, R.C.M.; Zuo, M.J.; Huang, H.Z. Reliability bounds for multi-state k-out-of-n systems. IEEE
Trans. Reliab. 2008, 57, 53–58. [CrossRef]
9. Zhao, X.; Cui, L. Reliability evaluation of generalized multi-state k-out-of-n systems based on FMCI approach.
Int. J. Syst. Sci. 2010, 41, 1437–1443. [CrossRef]
10. Zuo, M.J.; Tian, Z. Performance evaluation for generalized multi-state k-out-of-n systems. IEEE Trans. Reliab.
2006, 55, 319–327. [CrossRef]
11. Gertsbakh, I.; Shpungin, Y. Network Reliability and Resilience; Springer: Berlin, Germany, 2011.
12. Zarezadeh, S.; Asadi, M. Network reliability modeling under stochastic process of component failures. IEEE
Trans. Reliab. 2013, 62, 917–929. [CrossRef]
13. Zarezadeh, S.; Asadi, M.; Balakrishnan, N. Dynamic network reliability modeling under nonhomogeneous
Poisson processes. Eur. J. Oper. Res. 2014, 232, 561–571. [CrossRef]
14. Cha, J.H.; Finkelstein, M. A note on the class of geometric counting processes. Prob. Engin. Inform. Sci. 2013,
27, 177–185. [CrossRef]
15. Di Crescenzo, A.; Pellerey, F. Some results and applications of geometric counting processes. Methodol.
Comput. Appl. Probab. 2018.10.1007/s11009-018-9649-9. [CrossRef]
16. Shaked M.; Shanthikumar, J.G. Stochastic Orders; Springer: Berlin, Germany, 2007.
17. Lindqvist, B.H.; Samaniego, F.J.; Huseby, A.B. On the equivalence of systems of different sizes, with
applications to system comparisons. Adv. Appl. Prob. 2016, 48, 332–348. [CrossRef]
18. Navarro, J.; Samaniego, F.J.; Balakrishnan, N. Signature-based representations for the reliability of systems
with heterogeneous components. J. Appl. Prob. 2011, 48, 856–867. [CrossRef]
19. Samaniego, F.J. System Signatures and their Applications in Engineering Reliability; Springer: Berlin,
Germany, 2007.
20. Samaniego, F.J.; Navarro, J. On comparing coherent systems with heterogeneous components. Adv. Appl.
Prob. 2016, 48, 88–111. [CrossRef]
21. Belzunce, F.; Mercader, J.A.; Ruiz, J.M.; Spizzichino, F. Stochastic comparisons of multivariate mixture
models. J. Multivariate Anal. 2009, 100, 1657–1669. [CrossRef]

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

272
 MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel. +41 61 683 77 34
Fax +41 61 302 89 18
www.mdpi.com

Mathematics Editorial Office

E-mail: [email protected]
www.mdpi.com/journal/mathematics
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel: +41 61 683 77 34
Fax: +41 61 302 89 18
www.mdpi.com ISBN 978-3-03921-729-8