City Research Online

City, University of London Institutional Repository

Citation: Ignatov, Z. G. & Kaishev, V. K. (2016). First crossing time, overshoot and Appell-
Hessenberg type functions. Stochastics: An International Journal of Probability and
Stochastic Processes, 88(8), pp. 1240-1260. doi: 10.1080/17442508.2016.1230613

This is the accepted version of the paper.

This version of the publication may differ from the final published version.

Permanent repository link: https://openaccess.city.ac.uk/id/eprint/15137/

Link to published version: https://doi.org/10.1080/17442508.2016.1230613

Copyright: City Research Online aims to make research outputs of City,

University of London available to a wider audience. Copyright and Moral Rights
remain with the author(s) and/or copyright holders. URLs from City Research
Online may be freely distributed and linked to.

Reuse: Copies of full items can be used for personal research or study,
educational, or not-for-profit purposes without prior permission or charge.
Provided that the authors, title and full bibliographic details are credited, a
hyperlink and/or URL is given for the original metadata page and the content is
not changed in any way.
City Research Online: http://openaccess.city.ac.uk/ [email protected]
 First crossing time, overshoot and
Appell-Hessenberg type functions
Zvetan G. Ignatov1,* Vladimir K. Kaishev2,**
1 Faculty
of Economics and Business Administration, Sofia University “St Kliment
Ohridski”, 125 Tsarigradsko Shosse Blv., bl.3, Sofia 1113, Bulgaria. e-mail:
* [email protected]

2 Facultyof Actuarial Science and Insurance, Cass Business School, City University
London, 106 Bunhill Row, EC1Y 8TZ London,UK. e-mail: ** [email protected]

Abstract: We consider a general insurance risk model with extended flex-

ibility under which claims arrive according to a point process with inde-
pendent increments, their amounts may have any joint distribution and
the premium income is accumulated following any non-decreasing, possibly
discontinuous, real valued function. Point processes with independent in-
crements are in general non-stationary, allowing for an arbitrary (possibly
discontinuous) claim arrival cumulative intensity function which is appeal-
ing for insurance applications. Under these general assumptions, we derive a
closed form expression for the joint distribution of the time to ruin and the
deficit at ruin, which is remarkable, since as we show, it involves a new in-
teresting class of what we call Appell-Hessenberg type functions. The latter
are shown to coincide with the classical Appell polynomials in the Poisson
case and to yield a new class of the so called Appell-Hessenberg factorial
polynomials in the case of negative binomial claim arrivals. Corollaries of
our main result generalize previous ruin formulas e.g., those obtained for
the case of stationary Poisson claim arrivals.

MSC 2010 subject classifications: Primary 60K30; secondary 60K99.

Keywords and phrases: Appell polynomials, risk process, ruin probabil-
ity, first crossing time, overshoot, point process.

1. Introduction

The problem of first crossing of a boundary by a stochastic process has for long
time been of interest in applied probability. Such problems naturally arise in
many real-life applications, e.g. in insurance, finance, engineering, operations
research, reliability, queuing, chemistry, physics and biology. The literature on
first crossing is vast and various probabilistic models with different assumptions
on the boundary and the underlying stochastic process have been considered.
In risk theory the first crossing time and the overshoot of the process over the
boundary are interpreted as the ruin time and the deficit at ruin of an insurance
company. The boundary, most often an increasing linear function, represents the
accumulation of premiums over time, whereas the underlying stochastic process
is often assumed to be a compound Poisson process modelling the aggregate
claim amount. Ruin then occurs if the aggregate claim process exceeds the
premium income and the exceedance is the deficit at ruin. Studying the joint
1
 Ignatov and Kaishev/First crossing time and overshoot 2

distribution of the ruin time and deficit at ruin is important since both ruin
time and deficit are directly relevant in measuring and managing the solvency
and liquidity risk of an insurance company, applying e.g.,(Tail)-Value-at-Risk
type of analysis.
In this paper, we give explicitly this joint distribution in a risk model which
generalizes the classical one in several ways. We consider finite rather than infi-
nite time horizon. We relax the assumption of i.i.d claim amounts and consider
dependent claim sizes with any joint distribution. In our model, premium in-
come is accumulated following not just a linear function of time but following
any non-decreasing function, allowing jumps. Finally and most importantly, we
relax the classical assumption of Poisson claim arrivals and assume that claims
arrive according to a point process with independent increments. Somewhat sur-
prisingly, to the best of our knowledge, this reasonably general class of point
processes has not been considered in the literature on first crossing and ruin. As
we will demonstrate, it leads to a very elegant risk model which has significant
implications and allows to considerably extend the flexibility of modelling claim
arrivals. This is because point processes with independent increments are in
general non-stationary, allowing for an arbitrary (possibly discontinuous) claim
arrival cumulative intensity function. The latter feature is appealing for insur-
ance applications in which the intensity of claim arrivals can vary with time due
to, e.g., seasonal effects, environmental and climate changes and other reasons
related to economic slowdowns and speedups affecting insurance business. Fur-
thermore, the case of a cumulative intensity function with jumps corresponds
to the occurrence with non-zero probability of fixed points in the underlying
point process which is also relevant, e.g., in discrete time claim arrival models
of ruin. In the latter case a binomial claim arrival process naturally arises if a
finite-time ruin problem is considered. In the general case, point processes with
independent increments also include both stationary and non-stationary Pois-
son and negative binomial (NB) point processes as typical special cases of claim
arrivals. The latter processes allow for clustering of claims, including arrival of
two, three or more claims instantaneously, and/or clusters of arrivals at fixed
time instants. Clustering at both fixed and random time instants often occurs
in insurance portfolios. Therefore, such point process models have the potential
to capture better real claim arrival patterns.
Under these general risk model assumptions, in our main result given by
Theorem 2.3, we derive a closed form expression for the joint distribution of the
time to ruin and the deficit at ruin. The latter expression is remarkable, since
as we show, it involves a new interesting class of functions which are Appell
type functions and admit representation as Hessenberg determinants. For this
reason we refer to them as Appell-Hessenberg type functions. They generalize
the well known classical Appell polynomials introduced by Appell (1880). As
has first been shown by Ignatov and Kaishev (2000), classical Appell polynomi-
als naturally arise in ruin probability formulas in relation to the Poisson claim
arrivals. Different generalizations of the classical Appell polynomials have been
considered in the ruin context by Picard and Lefèvre (1997); Lefèvre and Pi-
card (2014b). It is worth noting however that these generalizations do not yield
 Ignatov and Kaishev/First crossing time and overshoot 3

the classical Appell polynomials as a special case and therefore differ from the
Appell-Hessenberg type functions considered here. For brevity in the sequel we
will refer to the latter simply as Appell-Hessenberg functions.
We show that in the case of Poisson claim arrivals the Appell-Hessenberg
functions coincide with the classical Appell polynomials, whereas when claims
arrive according to a negative binomial point process the Appell-Hessenberg
functions are expressed in terms of factorial functions. If the negative binomial
claim arrival process is stationary, the Appell-Hessenberg functions are shown to
yield a new class of polynomials which we call Appell factorial polynomials. Our
main result gives the (marginal) distribution of the time to ruin and therefore
generalizes the explicit ruin probability expressions, in terms of classical Appell
polynomials, obtained by Ignatov and Kaishev (2000, 2004, 2006), Ignatov et al.
(2001), and by Lefèvre and Loisel (2009) for the case of stationary Poisson claim
arrivals. Furthermore, it applies also to the special case of a stationary mixed
Poisson process (see Remark 3.8). This special case has recently been considered
by Lefèvre and Picard (2011, 2014a) within the context of processes with the
order statistics (OS) property. For more recent ruin-deficit formulas under OS
claim arrivals which cover and extend previous results by Lefèvre and Picard
(2011), see Dimitrova et al. (2014). It is worth mentioning that asymptotic ruin
probability results with respect to the initial capital, under some non-stationary
claim arrival processes (e.g. Hawkes and Cox processes with shot-noise intensity)
have recently been obtained by Stabile and Torrisi (2010) and Zhu (2013).
The paper is organized as follows. In section 2, we prove our main result
given by Theorem 2.3. For the purpose, we formulate and prove Lemmas 2.5,
2.6 and 2.7 (and also Proposition 2.2) which are of interest in their own right, es-
tablishing explicit and recurrent representations of the Appell-Hessenberg type
functions and sums of them. Corollaries 2.8, 2.9 of Theorem 2.3 give ruin for-
mulas for important special cases. In section 3, we specify the results of section
2 to the special cases of Poisson and negative binomial claim arrivals.

2. A formula for P (T < x, Y > y)

We assume that the amounts of consecutive claims to an insurance company are

modelled by the random variables W1 , W2 , . . . with partial sums Y1 = W1 , Y2 =
W1 + W2 , . . . having joint density f (y1 , . . . , yk ).
We will further assume that claims arrive according to a point process, ξ,
defined on (0, ∞), whose consecutive points, i.e., claim arrival times, can be
represented by a sequence of random variables 0 < T1 ≤ T2 ≤ . . ., independent
of W1 , W2 , . . ., with liml→+∞ Tl = +∞ a.s.. Such point processes exclude infinite
aggregation of points within finite time domains and are therefore well suited
for modelling arrivals of insurance claims and events in other applications. If
B + is the Borel σ-algebra on (0, ∞) and if B ∈ B + , then by ξ(B) we will denote
the number of points of ξ in B and in particular, ξ(0, t] will denote the number
of points (claim arrivals) on (0, t].
The cumulative premium income of the insurance company up to time t
 Ignatov and Kaishev/First crossing time and overshoot 4

is modelled by the function h(t) which is assumed a non-negative and non-

decreasing real-valued function, defined on [0, +∞), such that limt→∞ h(t) =
+∞. Let us also note that the function h(t) does not need to be necessarily
continuous and can therefore model arrivals of lump sum premium amounts.
If h(t) is discontinuous, we define h−1 (y) = inf{z : h(z) ≥ y}. The insurance
company’s surplus process is then expressed as Rt = h(t)−St , where St = Yξ(0,t]
is the aggregate claim amount process, and the instant of ruin, T , is defined as
T := inf {t : t > 0, Rt < 0} or T = ∞, if Rt ≥ 0 for all t. Given ruin occurs, i.e.
T < ∞, the deficit at ruin Y is defined as Y = −RT .
We consider a finite-time interval [0, x] and denote by P (T > x) and P (T <
x, Y > y) the probability of non-ruin in [0, x] and the probability that ruin
occurs before time x with a deficit, Y , exceeding y ≥ 0. In what follows, we
will give explicit expressions for these and other related probabilities under the
assumption that the process of claim arrivals, ξ, belongs to the class of point
processes with independent increments. Before elaborating on this class, let us
introduce some further notation. Denote by P (ξ(0, z] = i) the probability of i
points of the process ξ occurring in an interval (0, z], where P (ξ(0, z] = 0) 6= 0 for
any z > 0. Denote also by Λ(B) ≡ EξB, the average number of points in B, i.e.,
Λ(B) is the intensity measure of the process ξ. We will also assume that Λ((0, z])
is a possibly discontinuous function of z ∈ (0, ∞) and limz%+∞ Λ((0, z]) = +∞.
For brevity, we will further denote Λ((0, z]) = Λ(z) and R refer to it as cumulative
function of the intensity measure Λ(B), i.e. Λ(B) = B dΛ((0, z]), which more
concisely will be referred to as the cumulative intensity function. Let us note that
the case of jump discontinuous cumulative intensity function Λ(z) corresponds
to the presence of fixed points in ξ which is also covered within the class of point
processes with independent increments.
We will call ξ a point process with independent increments if for any 0 < s ≤ t,
the number of points, ξ(0, s], in the interval (0, s] and the number of points,
ξ(s, t], in the interval (s, t] are independent random variables. As known, a
point process with independent increments is in general non-stationary and
admits representation as a marked point process. Furthermore, every point pro-
cess with independent increments can be represented as a sum of a deterministic
component, a fixed points component and a compound Poisson point process
component (see e.g. Kallenberg (2002), Corollary 12.11 and Karr (1991), The-
orem 1.34 and Definition 1.37). In what follows we consider processes without
a deterministic component as the latter does not have a relevant interpretation
for insurance claim arrivals. Two important members of the class are the Pois-
son point process and the negative binomial point process. Since these processes
are in general non-stationary, they have been widely used in developing point
process models in various applications in queuing theory, physics, risk theory,
operations research, astronomy etc. Other compound Poisson processes such as
the Polya-Aeppli process have also been considered. For further properties of
point processes with independent increments we refer to e.g., section 1.5 of Last
and Brandt (1995), Chapter 1 of Karr (1991), and Chapter 10 of Daley and
Vere-Jones (2008). It should be noted that such processes are also considered
within the class of integer-valued Lévy processes.
 Ignatov and Kaishev/First crossing time and overshoot 5

In order to formulate our main result, we will need to introduce a special

class of functions which we call Appell-Hessenberg (type) functions.
Definition 2.1. For a fixed non-negative integer j, let 0 ≡ z0 < z1 < z2 <
. . . < zj < z be an arbitrary increasing sequence of positive real numbers. Define
the function Hj (z; z1 , . . . , zj ), z ∈ (zj , ∞) , as
Hj (z; z1 , . . . , zj ) = (−1)j det (∆ (z1 , . . . , zj , z)) , (1)
where
 P (ξ(0,z1 ]=1) 
P (ξ(0,z1 ]=0) 1 0 0 ··· 0
 P (ξ(0,z2 ]=2) P (ξ(0,z2 ]=1)


 P (ξ(0,z2 ]=0) P (ξ(0,z2 ]=0) 1 0 ··· 0 


∆ (z1 , . . . , zj , z) =  .. .. .. .. 
.
 . . . . 
 P (ξ(0,zj ]=j) P (ξ(0,zj ]=j−1) 

 P (ξ(0,zj ]=0) P (ξ(0,zj ]=0) ··· 1 

P (ξ(0,z]=j) P (ξ(0,z]=j−1)
P (ξ(0,z]=0) P (ξ(0,z]=0) ··· 1

In Definition 2.1, for j = 0, we assume ∆(z) ≡ (1), H0 (z) ≡ (−1)0 det(∆(z)) ≡

1.
As will be demonstrated, the functions Hj (z; z1 , . . . , zj ) of the variable z ∈
(zj , ∞), j = 0, 1, 2, . . ., defined by the parameters 0 ≡ z0 < z1 < z2 < . . . <
zj , can be viewed as generalizations of the classical Appell polynomials. The
latter have been first considered by Appell (1880), (see also Kaz’min (1995)
and Vein and Dale (1999)). Since the function Hj (z; z1 , . . . , zj ) is defined as a
determinant of a Hessenberg matrix, we call these functions Appell-Hessenberg
(type) functions. A matrix whose elements above or below the first subdiagonal
are equal to zero (i.e., all elements aij = 0 if j − i > 1 or if i − j > 1) are
called Hessenberg matrixes. For properties of Hessenberg matrixes and their
determinants we refer to e.g. Vein and Dale (1999). In what follows, it will some
times be convenient to interchangeably use the notation 0 ≡ z0 < z1 < z2 <
. . . < zj < zj+1 , with zj+1 ≡ z, for the sequence 0 ≡ z0 < z1 < z2 < . . . < zj <
z. Let us also note that in Definition 2.1 we have implicitly assumed that the
sequence z0 < z1 < z2 < . . . < zj < zj+1 is such that liml→+∞ zl = +∞. The
following recurrence formula facilitates the numerical evaluation of the Appell-
Hessenberg functions Hj (z; z1 , . . . , zj ).
Proposition 2.2. For a fixed non-negative integer j, let 0 ≡ z0 < z1 < z2 <
. . . < zj < z be an arbitrary increasing sequence of positive real numbers. For
the Appell-Hessenberg functions, Hj (z; z1 , . . . , zj ), defined in (1), we have
j
X P (ξ(0, z] = j − i)
Hj (z; z1 , . . . , zj ) = Hi (0; z1 , . . . , zi ) , j ≥ 1, (2)
i=0
P (ξ(0, z] = 0)

where H0 (z) ≡ 1, for z ≥ 0 and

i−1
X P (ξ(0, zi ] = i − k)
Hi (0; z1 , . . . , zi ) = − Hk (0; z1 , . . . , zk ) , i ≥ 1.
P (ξ(0, zi ] = 0)
k=0
 Ignatov and Kaishev/First crossing time and overshoot 6

Proof. The proof is similar to the proof given in Ignatov and Kaishev (2000) for
the case of classical Appell polynomials (see Lemma 1 therein) and is therefore
omitted.
Next, we state our main result which shows that the joint distribution of the
time to ruin and the deficit at ruin in the risk model with claim arrivals following
an arbitrary point process with independent increments, ξ, can be expressed in
terms of the Appell-Hessenberg functions, Hj (z; z1 , . . . , zj ), j = 0, 1, 2, . . . .
Theorem 2.3. The probability P (T < x, Y > y), x > 0, y ≥ 0, is given by
Z +∞ Z h(x)+y
P ξ(0, h−1 (y1 − y)] = 0 f (y1 ) dy1


P (T < x, Y > y) = f (y1 ) dy1 −
y y
Z +∞
− P (ξ(0, x] = 0) f (y1 ) dy1
h(x)+y
∞ Z
X Z
Bk−2 h−1 (yk−1 ); h−1 (y1 ), . . . , h−1 (yk−2 )


+ ...
k=2 Ck

−Bk−1 h−1 (yk − y); h−1 (y1 ), . . . , h−1 (yk−1 ) f (y1 , . . . , yk ) dyk . . . dy1


∞ Z
X Z
Bk−2 h−1 (yk−1 ); h−1 (y1 ), . . . , h−1 (yk−2 )


+ ...
k=2 Dk

−Bk−1 x; h−1 (y1 ), . . . , h−1 (yk−1 ) f (y1 , . . . , yk ) dyk . . . dy1 ,

(3)

where Ck = {(y1 , . . . , yk ) : 0 < y1 < . . . < yk−1 ≤ yk−1 + y < yk < h(x) + y},
Dk = {(y1 , . . . , yk ) : 0 < y1 < . . . < yk−1 < h(x) ≤ h(x) + y ≤ yk < +∞} and

Bj (z; z1 , . . . , zj ) = P (ξ(0, z] = 0) [H0 (z) + H1 (z; z1 ) + . . . + Hj (z; z1 , . . . , zj )] ,

(4)
for j = 0, 1, 2 . . . and Hj (z; z1 , . . . , zj ) are defined as in (1) with z1 = h−1 (y1 ) , . . . , zj =
h−1 (yj ).
The structure of formula (3) is reasonably straightforward: the first three
summands represent the probability of ruin at the first claim with deficit at least
y; the two sums correspond to the probability of survival after the first k − 1
claims and ruin at the k-th claim with deficit at least y for k = 2, 3, . . . (cf. the
proof). Practical applications (with numerical illustrations) of P (T < x, Y > y)
for the special case of stationary Poisson and Erlang claim arrivals in various
risk analysis problems, e.g. insurance solvency, systems reliability, flood risk
management, inventory management, are extensively discussed in Dimitrova
et al. (2015).
Remark 2.4. For the efficient numerical evaluation of (3), it is essential to: 1)
appropriately truncate the infinite summations; 2) compute the underlying mul-
tiple integrals; 3) efficiently compute the integrand functions Bj (z; z1 , . . . , zj ).
The latter can be done using recurrence formula (2) to compute each of the
 Ignatov and Kaishev/First crossing time and overshoot 7

functions Hl (z; z1 , . . . , zj ), l = 0, 1, . . . , j in (4). Methods for solving 1) and 2)

developed in Dimitrova et al. (2016) for the special case of stationary Poisson
claim arrivals could be generalized to the case of claim arrivals following a point
process with independent increments and in particular to the case of negative
binomial claim arrivals. Details of how this could be done are outside the scope
of the present paper and will be considered separately.
In order to prove Theorem 2.3 and some related corollaries, we will need the
following lemmas.
Lemma 2.5. For the real sequence 0 ≡ z0 < z1 < z2 < . . . < zj < z, and
Hj (z; z1 , . . . , zj ) defined as in (1), we have

j−1
!
X Y
P (ξ(0, z] = 0) Hj (z; z1 , . . . , zj ) = P (ξ(zl , zl+1 ] = gl ) P (ξ(zj , z] = gj ) ,
(g0 ,...,gj )∈E(0,j) l=0
(5)
Q−1
for j = 0, 1, 2, . . ., where it is assumed that l=0 (·) = 1, E(0, m) is the set of
(m + 1)-tuples of non-negative integers such that

E(0, m) = {(g0 , . . . , gm ) : g0 ≤ 0, g0 + g1 ≤ 1, . . . , g0 + . . . + gm−1 ≤ m − 1, g0 + . . . + gm = m} ,

(6)
where m is a non-negative integer.
Proof. Expressing Hj (z; z1 , . . . , zj ) from (1), we have
 
P (ξ(0, z] = 0) Hj (z; z1 , . . . , zj ) = (−1)j det ∆˜ (z1 , . . . , zj , z) , (7)

where ∆˜ (z1 , . . . , zj , z) is equal to ∆ (z1 , . . . , zj , z), introduced in Definition 2.1,

with its last row multiplied by P (ξ(0, z] = 0).
First, we verify that Lemma 2.5 holds in the cases j = 0 and j = 1. When
j = 0, and 0 ≡ z0 < z, for the left-hand side of (5), we have

P (ξ(0, z] = 0) H0 (z) ≡ P (ξ(0, z] = 0) ,

and for the right-hand side, we have

X
P (ξ(0, z] = g0 ) = P (ξ(0, z] = 0)
g0 ∈E(0,0)

and therefore, Lemma 2.5 holds. When j = 1 and 0 ≡ z0 < z1 < z, for the
left-hand side of (5), we have
P (ξ(0,z1 ]=1)
!
P (ξ(0,z1 ]=0) 1
P (ξ(0, z] = 0) H1 (z; z1 ) = (−1) det
P (ξ(0, z] = 1) P (ξ(0, z] = 0)
 Ignatov and Kaishev/First crossing time and overshoot 8

P (ξ(0, z1 ] = 1) P (ξ(0, z] = 0)
= P (ξ(0, z] = 1) × 1 −
P (ξ(0, z1 ] = 0)


= P (ξ(0, z1 ] = 0) ∩ (ξ(z1 , z] = 1) ∪ (ξ(0, z1 ] = 1) ∩ (ξ(z1 , z] = 0)

P (ξ(0, z1 ] = 1) P ((ξ(0, z1 ] = 0) ∩ (ξ(z1 , z] = 0))

−
P (ξ(0, z1 ] = 0)

=P (ξ(0, z1 ] = 0) P (ξ(z1 , z] = 1) + P (ξ(0, z1 ] = 1) P (ξ(z1 , z] = 0)

P (ξ(0, z1 ] = 1) P (ξ(0, z1 ] = 0) P (ξ(z1 , z] = 0)

− ,
P (ξ(0, z1 ] = 0)
where in the last equality we have used the fact that ξ is a process with in-
dependent increments and also that some of the events are mutually exclusive.
Finally, for the left-hand side of (5), after some trivial algebra, we obtain

P (ξ(0, z] = 0)H1 (z; z1 ) = P (ξ(0, z1 ] = 0) P (ξ(z1 , z] = 1) ,

where it can be directly verified that the right-hand side coincides with
X
P (ξ(0, z1 ] = g0 ) P (ξ(z1 , z] = g1 )
(g0 ,g1 )∈E(0,1)

and therefore, equality (5) is again valid.

We will continue the proof by induction. We showed that Lemma 2.5 holds
for j = 0 and j = 1. Assume it is true for all indexes up to j − 1. Lemma
2.5 will be proved if we show that (5) is true also for the index j. Expanding
the determinant on the right-hand side of equality (7) with respect to the first
column of ∆˜ (z1 , . . . , zj , z) and after some matrix algebra we obtain
 
P (ξ(0, z] = 0) Hj (z; z1 , . . . , zj ) = (−1)j det ∆˜ (z1 , . . . , zj , z)

P (ξ(0, z1 ] = 1) P (ξ(0, z2 ] = 2)
=− P (ξ(0, z) = 0)Hj−1 (z; z2 , . . . , zj ) − P (ξ(0, z) = 0)
P (ξ(0, z1 ] = 0) P (ξ(0, z2 ] = 0)
P (ξ(0, zj ] = j)
× Hj−2 (z; z3 , . . . , zj ) − . . . − P (ξ(0, z) = 0)H0 (z) + P (ξ(0, z] = j) .
P (ξ(0, zj ] = 0)
(8)
Let us denote by
j−1
!
[ \
Gj−n (zn+1 , . . . , zj , z) = (ξ(zl+1 , zl+2 ] = gl−n+1 ) ,
(g1 ,...,gj−n )∈E(1,j−n) l=n
(9)
for n = 0, . . . , j − 1, where on the right-hand side we assume that zj+1 ≡ z,
E(l, m), 1 ≤ l ≤ m ≤ j, is the set of (m − l + 1)-tuples of non-negative integers
such that
 Ignatov and Kaishev/First crossing time and overshoot 9

E(l, m) = {(gl , . . . , gm ) : gl ≤ 1, gl + gl+1 ≤ 2, . . . , gl + . . . + gm−1 ≤ m − l, gl + . . . + gm = m − l + 1} ,

(10)
and where for n = j, G0 (z) ≡ Ω.
Assume that Lemma 2.5 is true for any index n = 1, . . . , j − 1, i.e. apply (5)
noting that the right-hand side is equal to P ((ξ(0, zn+1 ] = 0) ∩ Gj−n (zn+1 , . . . , zj , z)),
evaluated at n = 0. Then, for the n-th term, n = 1, . . . , j, on the right-hand
side of (8), we have

P (ξ(0, zn ] = n)
− P (ξ(0, z) = 0)Hj−n (z; zn+1 , . . . , zj )
P (ξ(0, zn ] = 0)
P (ξ(0, zn ] = n)
= P ((ξ(z0 , zn+1 ] = g0 ) ∩ (Gj−n (zn+1 , . . . , zj , z)))
P (ξ(0, zn ] = 0)
P (ξ(0, zn ] = n)
= P (ξ(0, zn ] = 0) P (ξ(zn , zn+1 ] = 0) P (Gj−n (zn+1 , . . . , zj , z))
P (ξ(0, zn ] = 0)

= P (ξ(0, zn ] = n) P (ξ(zn , zn+1 ] = 0) P (Gj−n (zn+1 , . . . , zj , z)) = P (G(n)),

(11)
where we assume that zj+1 ≡ z and where in the last equality in (11)

G(n) = (ξ(0, zn ] = n) ∩ (ξ(zn , zn+1 ] = 0) ∩ Gj−n (zn+1 , . . . , zj , z) , (12)

n = 1, . . . , j. When n = 0, definition (12) shall be interpreted as

G(0) = (ξ(0, z0 ] = 0) ∩ (ξ(z0 , z1 ] = 0) ∩ Gj (z1 , . . . , zj , z)

(13)
= (ξ(0, z1 ] = 0) ∩ Gj (z1 , . . . , zj , z)

where we have assumed that (ξ(0, z0 ] = 0) is the sure event, i.e. (ξ(0, z0 ] = 0) ≡
Ω. We also recall that the probability of the event on the right-hand side of the
last equality in (13) is equal to the expression on the right-hand side of (5).
Now, applying (11), one can then rewrite (8) as
 
P (ξ(0, z] = 0)Hj (z; z1 , . . . , zj ) = P (ξ(0, z] = j) − P (G(1)) + . . . + P (G(j)) .
(14)
Let us note that for the events G(n), n = 1, . . . , j, appearing on the right hand
side of (14), the following statements are true
 
G(n) ⊂ ξ(0, z] = j , n = 0, . . . , j, (15)

G(n) ∩ G(m) = ∅, if 1 ≤ n 6= m ≤ j, (16)

where ∅ is the impossible event, and
j
!
[
(ξ(0, z] = j) \ G(l) ≡ G(0). (17)
l=1
 Ignatov and Kaishev/First crossing time and overshoot 10

Verification of (15), (16) and (17) is technically involved and will be omitted.
Now, in view of (16), (17) and (13), one can rewrite (14) as
 
P (ξ(0, z] = 0) Hj (z; z1 , . . . , zj ) = P (ξ(0, z] = j) − P (G(1) ∪ . . . ∪ G(j))
   
=P (ξ(0, z] = j) \ (G(1) ∪ . . . ∪ G(j)) = P (G(0)) = P (ξ(0, z1 ] = 0) ∩ Gj (z1 , . . . , zj , z)
j−1
!
X Y
= P (ξ(zl , zl+1 ] = gl ) P (ξ(zj , z) = gj ) ,
(g0 ,...,gj )∈E(0,j) l=0

where in the last equality, E(0, j) is defined in (6) for m = j and we have used
definitions (13), and (9) with n = 0, the properties of the probability measure
P , and the fact that ξ is a process with independent increments. This completes
the proof of Lemma 2.5.
Lemma 2.6. For the real sequence, 0 ≡ z0 < z1 < z2 < . . . < zk < z, we have
k−1
!
X Y
Bk (z; z1 , . . . , zk ) = P (ξ(zl , zl+1 ] = gl ) P (ξ(zk , z) = gk ) ,
(g0 ,...,gk )∈I(0,k) l=0
(18)
Q−1
where Bk (z; z1 , . . . , zk ) is defined as in (4) with j = k, l=0 (·) = 1, and where
I(0, k) is the set of (k + 1)-tuples of non-negative integers such that

I(0, m) = {(g0 , . . . , gm ) : g0 ≤ 0, g0 + g1 ≤ 1, . . . , g0 + . . . + gm−1 ≤ m − 1, g0 + . . . + gm ≤ m} .

Proof. Applying an appropriate mapping which relates each element (g0 , . . . , gj ) ∈

k
∪ E(0, j), where E(0, j) is defined in (6), to a set of elements from I(0, k), it
j=0
can be shown that, for the right-hand side of (18), we have
k−1
!
X Y
P (ξ(zl , zl+1 ] = gl ) P (ξ(zk , z) = gk )
(g0 ,...,gk )∈I(0,k) l=0
j−1
k
!
X X Y
= P (ξ(zl , zl+1 ] = gl ) P (ξ(zj , z) = gj ) .
j=0 (g0 ,...,gj )∈E(0,j) l=0
(19)

The assertion of Lemma 2.6 follows, applying Lemma 2.5 to the second sum
on the right-hand side of (19) which gives
j−1
k
! k
X X Y X
P (ξ(zl , zl+1 ] = gl ) P (ξ(zj , z] = gj ) = P (ξ(0, z) = 0) Hj (z; z1 , . . . , zj )
j=0 (g0 ,...,gj )∈E(0,j) l=0 j=0

≡ Bk (z; z1 , . . . , zk ) .
 Ignatov and Kaishev/First crossing time and overshoot 11

Lemma 2.7. Let 0 < T1 ≤ T2 ≤ . . . ≤ Tk ≤ Tk+1 ≤ . . . be the consecutive

points of a point process with independent increments, ξ, and let 0 ≡ z0 < z1 <
z2 < . . . < zk < z be a sequence of positive real numbers. For a fixed k, we have

P (T1 > z1 , . . . , Tk > zk , Tk+1 > z) = Bk (z; z1 , . . . , zk ) ,

where Bk (z; z1 , . . . , zk ) is defined as in (4) with j = k.

Proof. Let us consider the intersection of events

(T1 > z1 ) ∩ . . . ∩ (Tk > zk ) ∩ (Tk+1 > z) .

For k = 0 the event (T1 > z) coincides with the event (ξ(0, z] = 0), i.e. (T1 > z) =
(ξ(0, z] = 0), since if the realization T1 (ω) of T1 of the process ξ has occurred
within the interval (z, ∞), then the interval (0, z] remains empty, since the se-
quence T1 (ω) ≤ T2 (ω) ≤ . . . is non-decreasing.
Let us now sequentially transform the intersections
   
(T1 > z) ; (T1 > z1 )∩(T2 > z) ; . . . ; (T1 > z1 )∩. . .∩(Tk > zk )∩(Tk+1 > z) .

We have [
(T1 > z) = (ξ(0, z] = 0) ≡ (ξ(0, z] = g0 ) ,
g0 ∈I(0,0)

since I(0, 0) = {g0 = 0}, and

(T1 > z1 ) ∩ (T2 > z) = (ξ(0, z1 ] = 0) ∩ (ξ(z1 , z] = 0) ∪ (ξ(0, z1 ] = 0) ∩ (ξ(z1 , z] = 1)

 
= (ξ(0, z1 ] = 0) ∩ (ξ(z1 , z] = 0) ∪ (ξ(z1 , z] = 1)
   
[ [
= (ξ(0, z1 ] = g0 ) ∩  ((ξ(z0 , z1 ] = g0 ) ∩ (ξ(z1 , z] = g1 ))
g0 ∈I(0,0) (g0 ,g1 )∈J(1)

[
= ((ξ(z0 , z1 ] = g0 ) ∩ (ξ(z1 , z] = g1 )) ,
(g0 ,g1 )∈I(0,1)

where J(k) = {(g0 , g1 , . . . , gk ) : g0 + g1 + . . . + gk ≤ k}. Similarly, we also have

(T1 > z1 ) ∩ (T2 > z2 ) ∩ (T3 > z)

 !  !
[ \1 [ 1
\
= (ξ(zl , zl+1 ] = gl )  ∩  (ξ(zl , zl+1 ] = gl ) ∩ (ξ(z2 , z] = g2 ) 
(g0 ,g1 )∈I(0,1) l=0 (g0 ,g1 ,g2 )∈J(2) l=0
1
!
[ \
= (ξ(zl , zl+1 ] = gl ) ∩ (ξ(z2 , z] = g2 ) .
(g0 ,g1 ,g2 )∈I(0,2) l=0
 Ignatov and Kaishev/First crossing time and overshoot 12

By straightforward induction, one can write

(T1 > z1 ) ∩ (T2 > z2 ) ∩ . . . ∩ (Tk > zk ) ∩ (Tk+1 > z)

k−1
!
[ \ (20)
= (ξ(zl , zl+1 ] = gl ) ∩ (ξ(zk , z] = gk ) .
(g0 ,...,gk )∈I(0,k) l=0

Therefore (20) gives

 !
[ k−1
\
P (T1 > z1 , . . . , Tk > zk , Tk+1 > z) = P  (ξ(zl , zl+1 ] = gl ) ∩ (ξ(zk , z] = gk ) 
(g0 ,...,gk )∈I(0,k) l=0
k−1
!
X Y
= P (ξ(zl , zl+1 ] = gl ) P (ξ(zk , z] = gk ) .
(g0 ,...,gk )∈I(0,k) l=0
(21)

The assertion of Lemma 2.7 now follows applying Lemma 2.6 to the sum on
the right-hand side of (21).

We are now in position to prove Theorem 2.3.

Proof of Theorem 2.3: It is not difficult to see that

T1 < h−1 (Y1 − y) ∩ (T1 < x)

P (T < x, Y > y) = P
∞ k−1
! !
X \
−1 −1



+ P Tl > h (Yl ) ∩ Tk < h (Yk − y) ∩ (Tk < x) ,
k=2 l=1
(22)

where T1 < h−1 (Y1 − y) ∩ (T1 < x), is the event of ruin at the first claim with


k−1 
Tl > h (Yl ) ∩ Tk < h−1 (Yk − y) ∩ (Tk < x)
−1
T


deficit at least y, and
l=1
is the event of survival after the first k − 1 claims have arrived and ruin at the
k-th claim with deficit at least y.
Let us now transform the probabilities in (22). By means of conditional prob-
abilities, and after appropriate transformation of the domain of integration, it
is easy to show that

P T1 < h−1 (Y1 − y) ∩ (T1 < x)

Z +∞ Z h(x)+y
P ξ(0, h−1 (y1 − y)] = 0) f (y1 ) dy1


= f (y1 ) dy1 −
y y (23)
Z +∞
− P (ξ(0, x] = 0) f (y1 ) dy1 .
h(x)+y
 Ignatov and Kaishev/First crossing time and overshoot 13

It is straightforward to show that the remaining probabilities in (22), can be

expressed as
k−1
! ! Z Z ( "k−1 #
\ \
Tl > h−1 (Yl ) ∩ Tk < h−1 (Yk − y) ∩ (Tk < x) = . . . Tl > h−1 (yl )




P P
l=1 Ck l=1
" k−1
! #)
\
Tl > h−1 (yl ) ∩ Tk > h−1 (yk − y)



−P f (y1 , . . . , yk ) dy1 . . . dyk
l=1
Z Z ( "k−1
\
#
−1


+ ... P Tl > h (yl )
Dk l=1
" k−1
! #)
\
−1


−P Tl > h (yl ) ∩ (Tk > x) f (y1 , . . . , yk ) dy1 . . . dyk ,
l=1
(24)
where we have set Ck = {(y1 , . . . , yk ) : 0 ≤ y1 ≤ . . . ≤ yk−1 ≤ yk−1 + y ≤ yk ≤ h(x) + y}
and Dk = {(y1 , . . . , yk ) : 0 ≤ y1 ≤ . . . ≤ yk−1 ≤ h(x) ≤ h(x) + y ≤ h(x) < yk < +∞}.
From (22),(23) and (24), applying Lemma 2.7 to the probabilities on the
right-hand side of (24) we obtain the asserted formula (3).

The following two corollaries of Theorem 2.3 give explicitly formulas for the
joint distribution of the ruin time and deficit and for the finite and infinite time
probability of ruin, under a claim arrival process with independent increments,
which generalize previous results of Ignatov and Kaishev (2000, 2004) obtained
for the Poisson case.
Corollary 2.8. In the case of discrete claim amounts W1 , W2 , . . . with joint
probability mass function Pw1 ,...,wk = P (W1 = w1 , . . . , Wk = wk ), k = 1, 2, . . .,
we have
m−1
X l
X
Pw1 × P ξ(0, h−1 (w1 − y)) = 0


P (T < x, Y > y) = 1 − Pw 1 −
w1 =1 w1 =m
l−1
!
X
− 1− Pw1 × P (ξ(0, x) = 0)
w1 =1

l
X X
Pw1 ,...,wk Bk−2 h−1 (w1 + . . . + wk−1 ) ; h−1 (w1 ) , . . . , h−1 (w1 + . . . + wk−2 )


+
k=2 (w1 ,...,wk )∈C̃k

−Bk−1 h−1 (w1 + . . . + wk − y) ; h−1 (w1 ) , . . . , h−1 (w1 + . . . + wk−1 )

n+1
X X
Pw1 ,...,wk Bk−2 h−1 (w1 + . . . + wk−1 ) ; h−1 (w1 ) , . . . , h−1 (w1 + . . . + wk−2 )


+
k=2 (w1 ,...,wk )∈D̃k

−Bk−1 x; h−1 (w1 ) , . . . , h−1 (w1 + . . . + wk−1 ) ,

 Ignatov and Kaishev/First crossing time and overshoot 14

where m = [y] + 1, n = [h(x)], l = [h(x) + y], with [·] denoting the integer part,
C̃k = {(w1 , . . . , wk ) : 1 ≤ wi , i = 1, . . . , k, y < wk , w1 + . . . + wk < h(x) + y},
D̃k = {(w1 , . . . , wk ) : 1 ≤ wi , i = 1, . . . , k, w1 + . . . + wk ≤ h(x) ≤
h(x) + y ≤
w1 + . . . + wk < +∞}, and Bj z; h−1 (w1
) , . . . , h−1 (w1 + . . . + wj )
= P (ξ(0, z) = 0) H0 (z) + H1 z; h−1 (w1 ) + . . . + Hj z; h−1 (w1 ) , . . . , h−1 (w1 + . . . + wj ) .



Corollary 2.9. In the case of an infinite time horizon, i.e. when x = ∞,

formula (3) simplifies to
Z +∞ Z +∞
P ξ(0, h−1 (y1 − y)) = 0 f (y1 ) dy1


P (T < ∞, Y > y) = f (y1 ) dy1 −
y y
∞ Z
X Z
Bk−2 h−1 (yk−1 ) ; h−1 (y1 ) , . . . , h−1 (yk−2 )


+ ...
k=2 Ck

−Bk−1 h−1 (yk − y) ; h−1 (y1 ) , . . . , h−1 (yk−1 ) f (y1 , . . . , yk ) dyk . . . dy1 ,

where Ck = {(y1 , . . . , yk ) : 0 < y1 < . . . < yk−1 ≤ yk−1 + y < yk < +∞}, and
Bk is defined as in Theorem 2.3.
Let us note that further Corollaries which generalize previous ruin probability
formulas of Ignatov and Kaishev (2000, 2004, 2006) obtained for the Poisson case
can be easily obtained by directly substituting in (3), y = 0 and y = 0, x = ∞.

3. P (T < x, Y > y) for some special cases of the claim arrival

process ξ

In this section we provide some further corollaries of our main result given
by Theorem 2.3, which cover important special cases of claim arrival point
processes with independent increments, namely, the (non-)stationary Poisson
and negative binomial cases. To the best of our knowledge such models have not
been extensively considered in the ruin probability literature (with the exception
of the stationary Poisson claim arrival case).

3.1. Non-stationary Poisson claim arrivals

Let G(z) be the cumulative function of the measure Λ(·), i.e. G(z) = Λ((0, z]),
such that limz→∞ G(z) = ∞. If the process of claim arrivals, ξ, is a Poisson
point process with cumulative intensity function G(z) = Eξ((0, z]) then it is
not difficult to see that the definition of the Appell-Hessenberg functions, given
in (1) specifies to
 Ignatov and Kaishev/First crossing time and overshoot 15

G(z1 )1
 
1! 1 0 0 ... 0
 
 G(z2 )2 G(z2 )1 
 2! 1! 1 0 ... 0 
.. ..
 
Φj (z; z1 , . . . , zj ) = (−1)j det  ..
, (25)
 
. . .
G(zj )j G(zj )j−1
 

 j! (j−1)! ... 1 

 
j j−1
G(z) G(z)
j! (j−1)! ... 1

where j = 0, 1, 2, . . . and Φ0 (z) ≡ 1. The latter can be efficiently evaluated using

the following corollary of Proposition 2.2.
Corollary 3.1. For the Appell-Hessenberg functions, Φj (z; z1 , . . . , zj ), defined
in (25), we have
j
X G(z)j−i
Φj (z; z1 , . . . , zj ) = Φi (0; z1 , . . . , zi ) , j ≥ 0,
i=0
(j − i)!

where Φ0 (z) ≡ 1,
i−1 i−k
X G (zi )
Φi (0; z1 , . . . , zi ) = − Φk (0; z1 , . . . , zk ) , i ≥ 1,
(i − k)!
k=0

with Φ0 (0) ≡ 1.
From Lemma 2.7, for the Poisson case, we have
Lemma 3.2. Let 0 < T1 ≤ T2 ≤ . . . ≤ Tk ≤ Tk+1 ≤ . . . be the consecutive
points of a Poisson point process, ξ, and let 0 ≡ z0 < z1 < z2 < . . . < zk < z be
a sequence of positive real numbers. For a fixed k, we have
P (T1 > z1 , . . . , Tk > zk , Tk+1 > z) = βk (z; z1 , . . . , zk ) ,
where
βk (z; z1 , . . . , zk ) = e−G(z) [Φ0 (z) + Φ1 (z; z1 ) + . . . + Φk (z; z1 , . . . , zk )]
is the particular Poisson case version of the function Bk (z; z1 , . . . , zk ), given in
Theorem 2.3.
For an arbitrary G(z), the Appell-Hessenberg type function Φj (z; z1 , . . . , zj )
coincides with a classical Appell polynomial, as established by the following.
Corollary 3.3. For a non-stationary Poisson point process, ξ, with cumu-
lative intensity G(z), the Appell-Hessenberg type functions Φj (z; z1 , . . . , zj ),
j = 0, 1, 2, . . ., defined in (25), coincide with the classical Appell polynomials
Aj (G(z); G (z1 ) , . . . , G (zj )) ≡ Aj (G(z)) of degree j with a coefficient in front
of G(z)j equal to 1/j!, i.e.,
Φj (z; z1 , . . . , zj ) ≡ Aj (G(z); G (z1 ) , . . . , G (zj ))
 Ignatov and Kaishev/First crossing time and overshoot 16

where

A0 (G(z)) = 1,
A0j (G(z)) = Aj−1 (G(z)), and (26)
Aj (G (zj )) = 0,

j = 1, 2, . . ., with 0 ≤ z1 ≤ . . . ≤ zj , zj ∈ R.
The following corollary is a direct consequence of Corollary 3.3 and Lemma
3.2.
Corollary 3.4. With the notation of Lemma 3.2, we have

P (T1 > z1 , . . . , Tk > zk , Tk+1 > z) = bk (z; z1 , . . . , zk ) ,

where

bk (z; z1 , . . . , zk ) = e−G(z) [A0 (G(z)) + A1 (G(z); G (z1 )) + . . . + Ak (G(z); G (z1 ) , . . . , G (zk ))] ,
(27)
and A0 (G(z)), A1 (G(z); G (z1 )), . . ., Ak (G(z); G (z1 ) , . . . , G (zk )) are the clas-
sical Appell polynomials defined as in (26) with j = k, i.e. evaluated at G(z)
and defined by the sequence G (z1 ) , . . . , G (zk ).
Lemma 2.5 in the non-stationary Poisson case can be reformulated as follows.
Lemma 3.5. We have
g1
X (G (z2 ) − G (z1 ))
e−G(z) Φj (z; z1 , . . . , zj ) = e−G(z1 ) e−(G(z2 )−G(z1 )) × ...
g1 !
(g1 ,...,gj )∈E(1,j)

gj−1 g
(G (zj ) − G (zj−1 )) (G(z) − G (zj )) j
× e−(G(zj )−G(zj−1 )) e−(G(z)−G(zj )) ,
gj−1 ! gj !

where E(1, j) is defined as in (10), with l = 1 and m = j.

It can directly be seen from Lemma 2.6 that for the non-stationary Poisson
case we have
Lemma 3.6. For bk (z; z1 , . . . , zk ), defined as in (27) we have
X (G (z2 ) − G (z1 )) g1
bk (z; z1 , . . . , zk ) = e−G(z1 ) e−(G(z2 )−G(z1 )) × ...
g1 !
(g1 ,...,gk )∈I(1,k)

gk−1
(G (zk ) − G (zk−1 )) (G(z) − G (zk )) gk
× e−(G(zk )−G(zk−1 )) × e−(G(z)−G(zk )) ,
gk−1 ! gk !

where I(1, k) = {(g1 , . . . , gk ) : g1 ≤ 1, g1 + g2 ≤ 2, . . . , g1 + . . . + gk ≤ k}.

From Theorem 2.3, Lemma 3.2 and Corollary 3.4, we have
 Ignatov and Kaishev/First crossing time and overshoot 17

Corollary 3.7. In the case of Poisson claim arrivals with cumulative intensity
function G(z), the probability P (T < x, Y > y), x > 0, y ≥ 0, is given by
Z +∞ Z h(x)+y
−1
P (T < x, Y > y) = f (y1 ) dy1 − e−G(h (y1 −y)) f (y1 ) dy1
y y
Z +∞
−e−G(x) f (y1 ) dy1
h(x)+y

∞ Z
X Z
bk−2 h−1 (yk−1 ) ; h−1 (y1 ) , . . . , h−1 (yk−2 )


+ ...
k=2 Ck

−bk−1 h−1 (yk − y) ; h−1 (y1 ) , . . . , h−1 (yk−1 ) f (y1 , . . . , yk ) dyk . . . dy1

∞ Z
X Z
bk−2 h−1 (yk−1 ) ; h−1 (y1 ) , . . . , h−1 (yk−2 )


+ ...
k=2 Dk

−bk−1 x; h−1 (y1 ) , . . . , h−1 (yk−1 )

f (y1 , . . . , yk ) dyk . . . dy1 ,
(28)
where Ck , and Dk , are defined as in Theorem 2.3 and

bj z; h−1 (y1 ) , . . . , h−1 (yj )

= e−G(z) A0 (z) + A1 G(z); G(h−1 (y1 )) + . . . + Aj G(z); G(h−1 (y1 )), . . . , G(h−1 (yj )) .




(29)
for j = 0, 1, 2 . . . are the functions defined as in (27) (see Corollary 3.4).
Consider the case of claims arriving according to a stationary Poisson point
process, ξ, with intensity 1, i.e. with cumulative intensity G(z) = z. All results
from Section 3.1 directly hold for the stationary Poisson case replacing G(z)
with z.
Remark 3.8. In the case when claim arrivals follow a stationary Poisson pro-
cess with intensity λ, formula (28) for the probability P (T < x, Y > y) holds
with G(z) = λz in (29), i.e. with

bj z; h−1 (y1 ), . . . , h−1 (yj )

= e−λz A0 (z) + A1 λz; λh−1 (y1 ) + . . . + Aj λz; λh−1 (y1 ), . . . , λh−1 (yj ) .




Furthermore, if λ is a positive random variable, say Λ, formula (28) holds true

with

bj z; h−1 (y1 ), . . . , h−1 (yj )

h
 i
= E e−Λz A0 (z) + A1 Λz; Λh−1 (y1 ) + . . . + Aj Λz; Λh−1 (y1 ), . . . , Λh−1 (yj )


.

Let us note that this case corresponds to the mixed Poisson process of claim
arrivals considered also by Lefèvre and Picard (2011) as an order statistics point
process, i.e., formula (28) with P (T < x, Y > 0) covers formula (4.1) therein.
 Ignatov and Kaishev/First crossing time and overshoot 18

3.2. Negative binomial claim arrivals

Let the claim arrival process, ξ, be a negative binomial point process. Let G(z) be
the cumulative function of its intensity measure Λ(·), i.e. G(z) = Λ((0, z]), such
that limz→∞ G(z) = ∞. In other words, we assume that the random variable
ξ((0, z]) has a negative binomial distribution with parameters q and r(z) =
p
q G(z), (p = 1 − q), i.e.
 
−r(z)
P (ξ(0, z] = k) = pr(z) (−q)k , k = 1, 2, . . . .
k
Clearly, the process of claim arrivals, ξ, is a non-stationary negative binomial
point process with independent increments. Then, for the process ξ the definition
of the Appell-Hessenberg functions given in (1) specifies as

Ψj (z; z1 , . . . , zj )
   
r (z1 )
1 0 0 ... 0 0 

  1    
 r (z2 ) + 1 r (z2 ) 
 1 0 ... 0 0 

 2 1 

= (−1)j q j det  .. .. .. ,
 
   .  .  .   
 r (zj ) + j − 1 r (zj ) + j − 2 r (zj ) 
 ... 1 

  j   j−1   1 


 r(z) + j − 1) r(z) + j − 2 r(z) 
... 1
j j−1 1
(30)
where, j = 0, 1, 2, . . . and Ψ0 (z) ≡ 1. As seen from (30), Ψj (z; z1 , . . . , zj ) are
expressed in terms of factorial functions. The following corollary is a direct
consequence of Proposition 2.2.
Corollary 3.9. For the Appell-Hessenberg functions, Ψj (z; z1 , . . . , zj ), defined
in (30), we have
j  
X r(z) + j − i − 1
Ψj (z; z1 , . . . , zj ) = Ψi (0; z1 , . . . , zi ) , j ≥ 0,
j−i
i=0

where Ψ0 (z) ≡ 1,
i−1  
X r (zi ) + i − k − 1
Ψi (0; z1 , . . . , zi ) = − Ψk (0; z1 , . . . , zk ) , i ≥ 1,
i−k
k=0

with Ψ0 (0) ≡ 1.
In the case when r(z) = z, i.e. when ξ is a stationary NB process, from (30)
we obtain that
Ψj (z; z1 , . . . , zj )
Aj (z; z1 , . . . , zj ) = (31)
qj
 Ignatov and Kaishev/First crossing time and overshoot 19

are polynomials which we can view as a certain generalization of the clas-

sical Appell polynomials. In view of (30), the polynomials Aj (z; z1 , . . . , zj ),
j = 0, 1, 2, . . . can be referred to as Appell factorial polynomials, noting that
  
z+j−1 
= (z + j − 1)(z + j − 2) . . . (z + 0) /j!. (32)
j

The Appell factorial polynomials, Aj (z; z1 , . . . , zj ), j = 0, 1, 2, . . ., can be

recurrently computed by the following recurrence expression which follows from
Corollary 3.9.
Corollary 3.10. For the Appell-Hessenberg functions, Aj (z; z1 , . . . , zj ), defined
in (31), we have
j  
X z+j−i−1
Aj (z; z1 , . . . , zj ) = Ai (0; z1 , . . . , zi ) , j ≥ 0,
j−i
i=0

where A0 (z) ≡ 1,
i−1  
X zi + i − k − 1
Ai (0; z1 , . . . , zi ) = − Ak (0; z1 , . . . , zk ) , i ≥ 1,
i−k
k=0

with A0 (0) ≡ 1.
Let us recall that the classical Appell polynomials admit the Hessenberg
determinant representation (25) with G(z) replaced by z. So, comparing (25)
and (30), with (32) in mind, one can see that Aj (z; z1 , . . . , zj ) can formally be
obtained from Aj (z; z1 , . . . , zj ), j = 0, 1, 2, . . ., by replacing multiplication in
(25) with factorial multiplication, as in (32).
In the general case when r(z) 6= z we have that
Ψj (z; z1 , . . . , zj )
Aj (r(z); r (z1 ) , . . . , r (zj )) = ,
qj
i.e. the Appell-Hessenberg type functions, Ψj (z; z1 , . . . , zj ), j = 0, 1, 2, . . ., are
expressed through the Appell factorial polynomials, Aj (r(z); r (z1 ) , . . . , r (zj )).

Now from Lemmas 2.5, 2.6 and 2.7, for the case when ξ is non-stationary
negative binomial point process, we have the corollaries
Corollary 3.11. For the consecutive points T1 , T2 , . . . , of a negative binomial
point process, ξ, we have

P (T1 > z1 , . . . , Tk > zk , Tk+1 > z) = γk (z; z1 , . . . , zk ) ,

where 0 ≡ z0 < z1 < z2 < . . . < zk < z is a sequence of positive real numbers,

γk (z; z1 , . . . , zk ) = pr(z) [Ψ0 (z) + Ψ1 (z; z1 ) + . . . + Ψk−1 (z; z1 , . . . , zk )]

= pr(z) A0 (r(z)) + qA1 (r(z); r(z1 )) + . . . + q k−1 Ak (r(z); r(z1 ), . . . , r(zk ))
 
 Ignatov and Kaishev/First crossing time and overshoot 20

is the negative binomial special case of the function Bk (z; z1 , . . . , zk ) given in

Theorem 2.3, Ψk (z; z1 , . . . , zk ), k = 0, 1, 2, . . ., are the Appell-Hessenberg func-
tions defined in (30), and Ak (r(z); r(z1 ), . . . , r(zk )) are the Appell factorial poly-
nomials evaluated at r(z) and defined as in (31) by the sequence r(z1 ), . . . , r(zk ).
Corollary 3.12. For the Appell-Hessenberg functions defined in (30), we have

pr(z) Ψj (z; z1 , . . . , zj )
X  r(z2 ) − r(z1 ) + g1 − 1   r(z3 ) − r(z2 ) + g2 − 1  
r(z) − r(zj ) + gj − 1

j
=q × ×. . .× .
g1 g2 gj
E(1,j)

Corollary 3.13. We have

 
r(z)
X r(z2 ) − r(z1 ) + g1 − 1
g1 +...+gk
γk (z; z1 , . . . , zk ) = p q ×
g1
I(1,k)
     
r(z3 ) − r(z2 ) + g2 − 1 r(zk ) − r(zk−1 ) + gk−1 − 1 r(z) − r(zk ) + gk − 1
×. . .× × ,
g2 gk−1 gk
where I(1, k) is defined as in Lemma 3.6.
From Theorem 2.3 and Corollary 3.11 and Corollary 3.12, we have
Corollary 3.14. In the case when claim arrivals follow a negative binomial
point process with intensity function G(z), (i.e. NB parameters q and r(z) =
p
q G(z)), the probability P (T < x, Y > y), x > 0, y ≥ 0, is given by
Z +∞ Z h(x)+y
−1
P (T < x, Y > y) = f (y1 ) dy1 − p r (h (y1 −y))
f (y1 ) dy1
y y
Z +∞
−pr(x) f (y1 ) dy1
h(x)+y

∞ Z
X Z
γk−2 h−1 (yk−1 ); h−1 (y1 ), . . . , h−1 (yk−2 )


+ ... (33)
k=2 Ck

−γk−1 h−1 (yk − y); h−1 (y1 ), . . . , h−1 (yk−1 )

f (y1 , . . . , yk ) dyk . . . dy1
∞ Z
X Z
γk−2 h−1 (yk−1 ); h−1 (y1 ), . . . , h−1 (yk−2 )


+ ...
k=2 Dk

−γk−1 x; h−1 (y1 ), . . . , h−1 (yk−1 ) f (y1 , . . . , yk ) dyk . . . dy1 ,

where Ck and Dk are defined as in Theorem 2.3, and

γj z; h−1 (y1 ), . . . , h−1 (yj )

= pr(z) A0 (r(z)) + qA1 r(z); r(h−1 (y1 )) + . . . + q j Aj r(z); r(h−1 (y1 )), . . . , r(h−1 (yj )) ,




for j = 0, 1, 2 . . ., are the functions defined as in Corollary 3.11.

 Ignatov and Kaishev/First crossing time and overshoot 21

Remark 3.15. Let us note that the NB process with independent increments
considered here is in general different from the NB process considered by Lefèvre
and Picard (2011) in the context of OS risk processes. While both NB processes
are non-stationary and the distribution of the number of points in a fixed time
interval is negative binomial, the NB-process considered by Lefèvre and Picard
(2011), is a linear birth process with immigration which is Markovian and is
not a process with independent increments. Another difference is that the lat-
ter NB-process does not allow for the instantaneous arrival of clusters of points
which is possible under the definition adopted here. In conclusion, the two defi-
nitions are underpinned by different stochastic constructions, e.g. a NB process
with independent increments can be constructed by taking a compound Poisson
process with logarithmically distributed summands, whereas the NB process with
the OS property cannot (see Kozubowski and Podgórski 2009), leading to their
different properties and in particular different clustering of the points.

Acknowledgements

The authors would like to thank the Editor, the Associate Editor and the two
referees for the thorough review and for the valuable comments and suggestions
made, which helped us to significantly improve the paper.

References

Appell, P. 1880. Sur une classe de polynômes. Annales Scientifiques de l’École

Normale Suprieure 9, 119–144.
Daley, D.J., Vere-Jones, D. 2008. An Introduction to the Theory of Point Pro-
cesses. Volume II: General Theory and Structure. 2nd ed. Series in Probability
and Its Applications. Springer, New York.
Dimitrova, D.S., V.K. Kaishev, S. Zhao. 2016. On the evaluation of finite-time
ruin probabilities in a dependent risk model. Applied Mathematics and Com-
putation, 275, 268–286.
Dimitrova, D.S., V.K. Kaishev, and S. Zhao. 2015. Modelling finite-time failure
probabilities in risk analysis applications, Risk Analysis, 35(10), 1919-1939.
Dimitrova, D.S., Z.G. Ignatov, V.K. Kaishev. 2014. Ruin and deficit under claim
arrivals with the order statistics property. Submitted.
Ignatov, Z.G., V.K. Kaishev. 2000. Two-sided bounds for the finite-time prob-
ability of ruin. Scandinavian Actuarial Journal, 2000(1), 46–62.
Ignatov, Z.G., V.K. Kaishev. 2004. A finite-time ruin probability formula for
continuous claim severites. Journal of Applied Probability, 41(2), 570–578.
Ignatov, Z.G., V.K. Kaishev. 2006. On the infinite-horizon probability of
(non)ruin for integer-valued claims. Journal of Applied Probability, 43, 535–
551.
Ignatov, Z.G., V.K. Kaishev, R.S. Krachunov. 2001. An improved finite-time
ruin probability formula and its ”Mathematica” implementation. Insurance:
Mathematics and Economics, 29(3), 375–386.
 Ignatov and Kaishev/First crossing time and overshoot 22

Kallenberg, O. 2002. Foundations of Modern Probability. 2nd ed. Springer Series

in Probability and Its Applications. Springer-Verlag, New York.
Karr, A.F. 1991. Point Processes and Their Statistical Inference. 2nd ed. Series
in Probability: Pure and Applied. CRC/Marcel Dekker, New York.
Kaz’min, Y.A. 1995. Appell polynomials. In Encyclopedia of Mathematics, Vol.
1, ed. M. Hazewinkel, Kluwer Academic Publishers, Dordrecht, pp. 209–210.
Kozubowski, T., Podgórski, K. 2009. Distributional properties of the negative
binomial Lévy process. Probability and Mathematical Statistics, 29(1), 43–71.
Last, G., Brandt, A. 1995. Marked Point Processes on the Real Line. The Dy-
namical Approach. Series in Probability and Its Applications. Springer-Verlag,
New York.
Lefèvre, C., S. Loisel. 2009. Finite-time ruin probabilities for discrete, possibly
dependent, claim severities. Methodology and Computing in Applied Probabil-
ity, 11(3), 425–441.
Lefèvre, C., Picard, P. 2011. A new look at the homogeneous risk model. Insur-
ance: Mathematics and Economics, 49, 512–519.
Lefèvre, C., Picard, P. 2014a. Ruin probabilities for risk models with the ordered
claim arrivals. Methodology and Comp. in Applied Probability, 16, 885–905.
Lefèvre, C., Picard, P. 2014b. Appell pseudopolynomials and Erlang-type risk
models, Stochastics An International Journal of Probability and Stochastic
Processes, 86(4), 676-695.
Picard, P., Lefèvre, C. 1997. The probability of ruin in finite time with discrete
claim size distribution. Scandinavian Actuarial Journal, 1, 58–69.
Stabile, G., Torrisi, G.L. 2010. Risk Processes with non-stationary Hawkes
claims arrivals. Methodology and Comp. in Applied Probability, 12, 415–429.
Vein, R., P. Dale. 1999. Determinants and Their Applications in Mathematical
Physics. Springer-Verlag, New York.
Zhu, L. 2013. Ruin probabilities for risk processes with non-stationary arrivals
and subexponential claims. Insurance: Math. and Economics, 53, 544–550.