Gerber Shi Ur
Gerber Shi Ur
Gerber Shi Ur
Gerber-Shiu Functions
Hansjörg Albrecher a , Corina Constantinescu b , Gottlieb Pirsic b ,
Georg Regensburger b , Markus Rosenkranz b
a
Institute of Actuarial Science, Faculty HEC, University of Lausanne,
Extranef Building, CH-1015 Lausanne, Switzerland.
b
Johann Radon Institute for Computational and Applied Mathematics (RICAM),
Altenbergerstrasse 69, A-4040 Linz, Austria.
Abstract
We introduce an algebraic operator framework to study discounted penalty
functions in renewal risk models. For inter-arrival and claim size distributions
with rational Laplace transform, the usual integral equation is transformed
into a boundary value problem, which is solved by symbolic techniques. The
factorization of the differential operator can be lifted to the level of boundary
value problems, amounting to iteratively solving first-order problems. This
leads to an explicit expression for the Gerber-Shiu function in terms of the
penalty function.
1. Introduction
We consider the collective renewal risk model introduced by Sparre Andersen
(1957) that describes the amount of free capital U (t) at time t in an insurance
portfolio by
N (t)
X
U (t) = u + ct − Xk .
k=1
Here N (t) is a renewal process that counts the number of claims incurred
during the time interval (0, t], the constant c is the premium rate and the
random variables (Xk )k≥0 denote the claim sizes that occur at random times
1
(Tk )k≥0 , with τk = Tk − Tk−1 i.i.d. random variables denoting the k-th
interclaim (or inter-arrival) time (T0 = 0). The initial surplus (after the
claim at time 0 is paid) is given by u ≥ 0. Moreover, (Xk )k≥0 and (τk )k≥1 are
assumed to be independent. Ruin occurs when the surplus process becomes
negative for the first time, so the time of ruin is given by
The net profit condition cE(Tk ) > E(Xk ) is imposed to ensure that ψ(u) < 1
for all u ≥ 0.
Denoting by f (x, y, t | u) the joint probability density function of the sur-
plus immediately before ruin U (Tu −), the deficit at ruin |U (Tu )| and the time
of ruin Tu , we have
Z ∞Z ∞Z ∞
f (x, y, t | u) dx dy dt = ψ(u).
0 0 0
m(u) = E e−δTu w(U (Tu −), |U (Tu )|) 1Tu <∞ | U (0) = u
Z ∞Z ∞Z ∞
= e−δt w(x, y) f (x, y, t | u) dx dy dt,
0 0 0
2
The established methods for deriving explicit expressions for functions
arising in risk theory (e.g. ruin probability, Laplace transform of the time to
ruin, Gerber-Shiu function) are either based on defective renewal equations
or integral equations (Volterra of second kind). Specifically, starting with
the defective renewal equation satisfied by the Gerber-Shiu function, Lin
and Willmot (2000) propose a solution expressed in terms of the tail of a
compound geometric distribution. For particular claim sizes (combinations
of exponentials, mixture of Erlangs) they derive explicit analytic solutions for
this distribution. In Willmot (2007) this defective renewal equation method is
adapted to the analysis of renewal risk models with arbitrary distributions.
Another strategy, based on the defective renewal equation, was suggested
in the classical compound Poisson model by Drekic et al. (2004). They use
Mathematica to obtain the moments of the time to ruin, based on the system
of defective renewal difference equations derived by Lin and Willmot (2000).
In this paper, we introduce an algebraic operator approach with symbolic
techniques for deriving explicit expressions for Gerber-Shiu functions. These
techniques are easy to implement, and their further analysis can draw on the
full potential of current computer algebra systems.
In general renewal models, m(u) can alternatively be expressed as the
solution of a Volterra integral equation of the second kind and hence as a
Neumann series, see Gerber and Shiu (1998). Under the further assump-
tion that the interclaim times have rational Laplace transform, the integral
equation can be transformed into an integro-differential equation (IDE) with
suitable boundary conditions. For the solution of the IDE, due to its convo-
lution structure, Laplace transforms are often the key tool to derive explicit
solutions; see e.g. Cheng and Tang (2003), Albrecher and Boxma (2005)
and Li and Garrido (2005b). Landriault and Willmot (2008) obtain explicit
expressions for the Laplace transform that can be inverted back by partial
fractions, for arbitrary interclaim times and Coxian claim sizes. However,
explicitly inverting the Laplace transform is in general difficult. Li and Gar-
rido (2004) solved the IDE for Erlang(n) [E(n)] (sum of n independent ex-
ponential random variables) interclaim times by repeatedly integrating the
integro-differential equation satisfied by the Gerber-Shiu function.
In the present paper, we want to advocate an alternative approach to
derive explicit expressions for the Gerber-Shiu function in renewal models.
For interclaim time distributions with rational Laplace transform—or equiva-
lently if the interclaim density satisfies a linear ordinary differential equations
(LODE) with constant coefficients—we first use the systematic approach
of Constantinescu (2006) to transform the integral equation for m(u) into an
integro-differential equation. If the claim size distribution also has a ratio-
nal Laplace transform, the IDE can be further reduced to a linear boundary
3
value problem with appropriate boundary conditions (Section 2). Evaluat-
ing the IDE and its derivatives at 0 and imposing regularity conditions at
∞, we supplement the differential equation with sufficiently many boundary
conditions so that the Gerber-Shiu function is uniquely determined. This
program considerably extends the approach of Chen et al. (2007), who de-
rived a LODE for m(u) in a Poisson jump-diffusion process with phase-type
jumps and solved it explicitly for penalty functions that depend only on the
deficit at ruin.
Having arrived at a linear boundary problem, we employ the symbolic
method developed in Rosenkranz (2005) and Rosenkranz and Regensburger
(2008) for computing the integral operator (Green’s operator) that maps the
penalty function to the corresponding Gerber-Shiu function; see Section 3
for a brief description of this approach. Based on an algebraic operator
framework, this method uses noncommutative Gröbner bases for transform-
ing integro-differential and boundary operators to normal forms.
Whereas the classical version of this method works only for boundary
value problems on compact intervals, we extend the approach to problems on
the positive half-line in Section 4. There we consider operators on functions
vanishing at infinity, which is the appropriate setup for our purposes.
In Section 5 we present the solution of the boundary value problem in
terms of the Green’s operator. The method relies on the factorization of the
differential operator using the roots of the Lundberg fundamental equation.
This factorization is then lifted to the level of boundary value problems:
One can iteratively solve a sequence of first-order boundary value problems
with appropriate boundary conditions. It turns out that there is a crucial
difference between the roots with positive and negative real part and that
there are natural links to the so-called Dickson-Hipp operator. Altogether,
this approach allows to compute the Gerber-Shiu function up to quadratures.
In previous papers e.g. Li and Garrido (2004) and Chen et al. (2007), the
boundary conditions of the IDE are computed recursively in terms of deriva-
tives of m(u) at zero. In Section 6, we use an integrating factor method with
different integration bounds and exploit the Vandermonde-type structure of
the resulting matrix for directly deriving an explicit expression for each of
these boundary values. This in turn makes it possible to arrive at a fully
explicit formula for m(u) in terms of the penalty function. An illustration
of our method for E(n) interclaim times with E(m) claim sizes is given in
Section 7. The method also covers more general models like the case of re-
newal risk models perturbed by a Brownian motion treated in Section 8. We
conclude in Section 9 by discussing possible extensions of this approach.
4
2. Reduction to a Boundary Value Problem
Consider T1 to be the epoch of the first claim. Since ruin cannot occur in the
interval (0, T1 ), by the standard renewal argument of Feller (1971, p.183-184)
one has
for any claim size density fX and interclaim time density fτ . Due to the net
profit condition, the model satisfies the regularity condition
5
with almost homogenous initial conditions (5). In the special case of expo-
nentials with the same parameter λ, this is an Erlang(n) density fτ (t) =
1
(n−1)!
λn tn−1 e−λt , satisfying equation (4) with almost homogenous initial con-
ditions (5) and polynomial
Under assumption (4) one can now use the technique of integration by parts
as in Theorem 3 of Constantinescu (2006, Sec. 3.2) to obtain from (1) the
integro-differential equation
Z u
∗ d
pτ (c − δ) m(u) = a0 m(u − y) dFX (y) + a0 ω(u), (8)
du 0
where the
R ∞derivatives of m are assumed to exist and to be bounded. Here
ω(u) = u w(u, y − u) dFX (y) and
where p∗τ
( dtd ) denotes the adjoint operator of the Roperator pτ ( dtd ) defined
∞
through pτ ( dtd )f, g = f, p∗τ ( dtd )g with hf, gi = 0 f (x)g(x) dx together
with (5). In addition to the model regularity condition (2), we will derive in
Section 6 the initial values Mi (i = 0, . . . , n − 1) of the IDE (8) through a
variation of the classical integrating factor method of Gerber and Shiu (1998),
obtaining
m(0) = M0 , m0 (0) = M1 , . . . , m(n−1) (0) = Mn−1 . (9)
Together with (2), these boundary conditions make the boundary value prob-
lem regular.
Remark 2. Note that the same analysis also works for the case in which the
boundary conditions are not of homogeneous type (as for instance would be
the case for a mixtures of Erlangs). In that case the Laplace transform of
fτ has a polynomial numerator of lower degree than of the polynomial in the
denominator. As a consequence, one obtains further integral terms on the
right-hand side of (8), leading to a slightly more cumbersome procedure.
Define the polynomial
If moreover the claim size density fX satisfies a LODE with constant coeffi-
cients
d
pX ( )fX (y) = 0,
dy
6
and (for simplicity) almost homogeneous boundary conditions
(k)
fX (0) = 0 (k = 0, . . . , n − 2),
(n−1)
fX (0) = b0 ,
of (11) is the Lundberg fundamental equation of this model. Since both the
claim sizes and the inter-arrival times have rational Laplace transforms, we
know by the results in Li and Garrido (2005a) and Landriault and Willmot
(2008) that this equation has exactly n roots with positive and m roots with
negative real part as long as δ > 0. Note that we exclude the limiting case
δ = 0, which is equivalent to having 0 as a solution of the Lundberg equation;
see Section 5 for a brief discussion of this case.
7
simple boundary value problem with inhomogeneous boundary conditions
but f = 0.
The boundary value problem (13) is called regular if for every f there
exists a unique g or equivalently if the associated homogeneous problem only
has the trivial solution. This can be checked by testing whether the matrix
formed by evaluating the boundary conditions on a fundamental system is
regular; for details see Kamke (1967, p. 184). In this case, there is a well-
defined operator G : C[a, b] → C n [a, b] mapping f 7→ g, known as the Green’s
operator of (13). While G is usually represented by its associated Green’s
function (Stakgold, 2000), the operator formulation is more practical in the
present setting.
An essential feature of the symbolic operator calculus is that it allows to
compose two boundary value problems (in particular those of the form (13))
such that the composite Green’s operator is given by the composition of the
constituent Green’s operators. For solving boundary value problems, the
other direction is more important: Any factorization of the underlying dif-
ferential operator can be lifted to a factorization of boundary value problems.
Since we are dealing with differential operators with constant coefficients, we
can actually achieve a factorization into first-order boundary value problems.
For more details on composing and factoring boundary value problems for
LODE, we refer again to Rosenkranz and Regensburger (2008). The theory is
developed in an abstract algebraic setting, including in principle also bound-
ary value problems for linear partial differential equations, in Regensburger
and Rosenkranz (2009).
In the present setting, we can describe the first-order Green’s operators
as follows. Writing
rx rb rb
A = a , B = x , and F = a = A + B,
and
Aσ = eσx Ae−σx , Bρ = eρx Be−ρx , and Fσρ = eσx F e−ρx
for ρ, σ ∈ C, the basic first-order boundary value problems, with respect
to each of the end points of the interval, (D − σ) g = f, g(a) = 0 and
(D − ρ) g = −f, g(b) = 0, have respectively Aσ and Bρ as their Green’s
operators as one can see by the fundamental theorem of calculus. Written as
operator identities, this means in particular that
(
(D − σ)Aσ = 1,
(14)
(D − ρ)Bρ = −1,
8
By Rosenkranz (2005,Table 1), we obtain furthermore for any ρ̃, σ̃ ∈ C
(σ − σ̃) Aσ Aσ̃ = Aσ − Aσ̃
(ρ̃ − ρ) Bρ Bρ̃ = Bρ − Bρ̃ (15)
(ρ − σ) Aσ Bρ = Aσ + Bρ − Fσρ
on C[a, b]; the first two are called resolvent identities (Yosida, 1995). For the
extension to non-compact intervals in Section 4 we mention an alternative,
purely algebraic, way to derive (15), namely as a consequence of conditions
that will be simpler to establish in the more general case:
Lemma 1. The identities (15) are algebraic consequences of
Aσ (D − σ)Aσ̃ = Aσ̃
Bρ (D − ρ)Bρ̃ = −Bρ̃ (16)
Aσ (D − σ)Bρ = Bρ − Fσρ
9
for all f ∈ C0 and x ≥ y ≥ 0. Fixing ε > 0, the first summand is smaller than
ε/2 for x ≥ x0 (ε, y) because η > 0. Since f ∈ C0 , we have |f (ξ)| < εη/2 for
all ξ ≥ y0 (ε), so the second summand is smaller than ε/2 for x ≥ y0 (ε) and
y = y0 (ε). Thus we obtain |A−ρ f (x)| < ε for all x ≥ max {y0 (ε), x0 (ε, y0 (ε))}.
Using a similar argument as for the second summand, we obtain Bρ f ∈ C0 .
One immediately checks that e−ρx A and Be−ρx map even bounded functions
into C0 .
Next we verify that the operators areRcontinuous. The norm R x ηξ bound for
−ηx x ηξ −ηx
A−ρ follows from |A−ρ f (x)| ≤ e kf k∞ 0 e dξ and e 0
e dξ ≤ 1/η;
similarly for e−ρx A and Be−ρx . For Bρ we use the representation
Z ∞
Bρ f (x) = e−ρξ f (ξ + x) dξ (18)
0
R∞
and the fact that 0 e−ηξ dξ = 1/η.
Now we turn to differentiability and identities (14). For A−ρ this follows
immediately from the fundamental theorem of calculus. Using representa-
tion (18), the difference quotient (Bρ f (x + h) − Bρ f (x))/h is given by
∞ h
eρh − 1
Z Z
−ρξ 1
e f (ξ + x) dξ − e−ρξ f (ξ + x) dξ,
h h h 0
for all f ∈ C01 . The identity for Aσ carries over from the bounded case and is
even valid on C 1 [0, ∞), the one for Bρ follows from the representation (18)
and integration by parts.
10
Proposition 3. The subalgebra E0 is dense in C0 , and the operators (17)
map E0 into itself.
Proof. Density follows from the Stone-Weierstrass Theorem for locally com-
pact spaces (Conway, 1990, p. 147). For proving that the operators (17) map
E0 into itself, one uses induction on j and integration by parts.
Note that—by the same reasoning—the operators Aρ and Bρ also map
E0 into itself if Re(ρ) = 0 but they are no longer continuous.
This proposition provides an alternative approach to proving the iden-
tities (14), (15): Since E0 is dense in C0 and the operators are continuous,
it suffices to prove them for exponential polynomials—this can be done by
an elementary computation and induction on j. Density arguments of this
type could also be useful for generalizing to larger function spaces like Lp or
spaces based on regular variation (Bingham et al., 1987).
Tρ = (D − ρ1 ) · · · (D − ρn ) and Tσ = (D − σ1 ) · · · (D − σm ).
11
Note that in order to have a regular boundary value problem, it is sufficient
to prescribe m initial conditions even though the order of T is m + n. This
is due to the regularity condition g ∈ C0 : The general solution g of the
associated homogeneous differential equation T g = 0 is a linear combination
of eρj x and eσi x , where all terms with positive roots must vanish and the
remaining m coefficients are determined by the m conditions at zero.
The crucial point is that it is possible to factor this boundary value prob-
lem along T = Tρ Tσ into the regular boundary value problems
Tσ g = h, Tρ h = f,
and (21)
g(0) = . . . = g (m−1) (0) = 0 h ∈ C0
with m n
Y Y
−1
ai = (σi − σk ) and bj = − (ρj − ρk )−1
k=1,k6=i k=1,k6=j
Q1
as their Green’s operators, so g = Gσ h and h = Gρ f, where k=1,k6=1 = 1.
Proof. Let us first prove the identity for Gσ by induction (the case for Gρ is
analogous). The base case m = 1 is trivial, so assume the identity for m − 1.
Then (15) yields
m−1
X m−1
X Y m
−1
Aσ1 · · · Aσm−1 Aσm = ai Aσi − (σi − σk ) Aσm
i=1 i=1 k=1,k6=i
and we are done since the parenthesis is equal to −am by the well-known
partial fraction formula.
By Proposition 2, the Green’s operators Gρ and Gσ map C0 to C0m and C0n ,
respectively, and (14) yields Tσ Gσ = 1 and Tρ Gρ = 1. It remains to check
that Gσ f satisfies the initial conditions. For that we prove for all i < m the
identity
X i
i
D Gσ = hi−l (σ1 , . . . , σl+1 ) Aσl+1 · · · Aσm , (22)
l=0
12
follows because Aσ1 f (0), . . . , Aσm f (0) = 0. The base case i = 0 is trivial, so
assume (22) for i − 1. Using DAσl+1 = 1 + σl+1 Aσl+1 from (14), this gives
i−1
X
i
D Gσ = hi−l−1 (σ1 , . . . , σl+1 ) DAσl+1 · · · Aσm
l=0
i−1
X
= hi−l (σ1 , . . . , σl ) + σl+1 hi−l−1 (σ1 , . . . , σl+1 ) Aσl+1 · · · Aσm
l=1
+ σ1i Aσ1 · · · Aσm + Aσi+1 · · · Aσm
after a little rearrangement. But the parenthesized factor in the sum sim-
plifies to hi−l (σ1 , . . . , σl+1 ), while the outlying summands also have the right
factors hi−0 (σ1 ) = σ1i and hi−i (σ1 , . . . , σj+1 ) = 1, respectively.
Theorem 5. The boundary value problem (20) has the Green’s operator
m X
X n
Gσ Gρ = cij (Aσi + Bρj − Fσi ρj )
i=1 j=1
m X
X n
= cij eσi x A(e−σi x − e−ρj x ) + (eρj x − eσi x )Be−ρj x
i=1 j=1
13
with the corresponding Lundberg fundamental equation
where fˆX (z) = E(e−zX ) is the Laplace transform of fX (u). Equation (24)
has exactly n solutions ρi (i = 1, . . . , n) with positive real part, according
to Li and Garrido (2004).
We will use a similar integrating factors technique as the one proposed
in Gerber and Shiu (1998) and arrive at a system of linear equations in the
initial values that we can solve explicitly. A different choice of the integra-
tion bounds will simplify some steps compared to a related approach of Li
and Garrido (2004). The change of variables and order of integration used
in Gerber and Shiu (1998) is then not necessary here. Let us multiply equa-
tion (23) by e−ρi u for each i = 1, . . . , n, and then integrate from u = ∞ to
u = x to arrive at
n Z x
X n j
(−c) (λ + δ) (n−j)
e−ρi u m(j) (u) du
j=0
j ∞
Z x Z u Z x
−ρi u
=λ n
e m(u − y) dFX (y) du + λ n
e−ρi u ω(u) du.
∞ 0 ∞
to obtain
Z x j−1
X
e −ρi u (j)
m (u)du = e−ρi x ρki m(j−k−1) (x) + ρji Ii (x),
∞ k=0
Z x
where Ii (x) = e−ρi u m(u)du.
∞
Then evaluating each equation at x = 0, we note that the left-hand side and
the right-hand side terms pertaining to Ii (0) cancel due to (24) evaluated at
z = ρi . Also we see that in the right-hand side the second integral is actually
−ω̂(ρi ), the Laplace transform of ω evaluated at ρi . We obtain a system of
n equations in n unknown variables m(k) (0)
n j−1
X n j n−j
X
(−c) (λ + δ) ρki m(j−k−1) (0) = −λn ω̂(ρi )
j=1
j k=0
14
for k = 0, . . . , n − 1. Collecting and rearranging the terms, we get
n−1 n−k−1
X j n
X
(k) n λ+δ (n−k−1)−j λ
m (0) − ρi =− − ω̂(ρi ), (25)
k=0 j=0
j c c
| {z }
pn−k−1 (ρi )
15
Proof. According to Krattenthaler (1999), the determinant of the matrix A
is the same as the Vandermonde determinant Vn = Vn (ρ1 , . . . , ρn ) so
Y
det(A) = (ρj − ρi ).
1≤i<j≤n
we obtain
k
X
det(Ai,n−k ) = Vn−1 (ρ0i ) dj ek−j (ρ0i ),
j=0
where
n j+1
(−1)j + (1 − λ+δ x)n − 1 + − λ+δ x
d j = xj c
λ+δ λ+δ
nc
(1 − c x)n − − c x
and [xj ]f (x) = f (j) (0)/j! denotes the coefficient of xj of a power series f (x).
We will show below that
j
λ+δ n−1+j
dj = − . (29)
c j
Inserting the resulting formula for the determinant Ai,k into the expansion
of det(Bk ) in Cramer’s rule, we get
n k
λ X Vn−1 (ρ0i ) X
m(k) (0) = (− )n (−1)i+n−k+1 ω̂(ρi ) dj ek−j (ρ0i ),
c i=1 Vn j=0
which after cancellation of the Vandermode terms leads to the result stated.
16
It
Premains to show equation (29). From equation (34) we get that (−1)j dj =
j
[x ] m=0 (−q(x))m . Since j < n we can safely add terms of order at least
j
Finally, the last sum vanishes due to Graham et al. (1989, 5.42) since it is the
(j + 1)-th difference of n(l−1)
j
as a polynomial in l, which is only of degree
j.
Since the Gerber-Shiu function is the unique solution of (19), it has the
form
m(u) = Gσ Gρ f (u) + mp (u),
where Gσ Gρ is given in Theorem 5 and mp (u) is the particular solution
obtained as a linear combination of the eσi u , with factors determined by the
initial values from Proposition 6.
17
with discount rate δ > 0. From the previous section we get n boundary con-
ditions. As described in Section 5 one in fact needs m boundary conditions,
so we assume m ≤ n (otherwise, one can derive the remaining conditions by
evaluating higher derivatives of the integro-differential equation (23)). We
obtain a boundary value problem for the differential equation T m = f with
d
D = du , where
and boundary conditions (2) and (9). To apply the results from Section 5,
we can choose any sufficiently smooth penalty function w(x, y) such that
limu→∞ f (u) = 0. By Proposition 3 this includes all bivariate exponential
polynomials whose terms xi y j eαx eβy satisfy α < β < µ.
Since the characteristic equation for T is the Lundberg fundamental equa-
tion, we know from the general results mentioned in Section 2 that it has n
roots ρ1 , . . . , ρn with positive real part and m roots σ1 , . . . , σm with negative
real part. So we have the factorization
T = Tρ Tσ = (D − ρ1 ) · · · (D − ρn )(D − σ1 ) · · · (D − σm ),
and Theorem 5 gives us the Green’s operator for the corresponding homoge-
neous boundary value problem.
Writing fˆ for the Laplace transform of f and using the definition of the
corresponding operators, we obtain from Theorem 5 the explicit form of the
Gerber-Shiu function
m X n Z ∞ !
X Z u
m(u) = cij eσi (u−ξ) + eρj (u−ξ) f (ξ) dξ − fˆ(ρj )eσi u +mp (u)
i=1 j=1 0 u
(31)
with m n
Y Y
cij = − (σi − σk )−1 (ρj − ρk )−1 (ρj − σi )−1 .
k=1,k6=i k=1,k6=j
With the initial values from formula (28) the computation of the particu-
lar solution mp satisfying the inhomogeneous boundary conditions reduces
to solving a system of linear equations, obtained from imposing the condi-
tion that the particular solution satisfies these given initial conditions, i.e.
(mp )(i) (0) = Mi . As remarked in Section 5, formula (31) remains valid for
suitable f also in the limiting case δ = 0, which is equivalent to having 0
among the ρ1 , . . . , ρn .
18
So the problem of computing the Gerber-Shiu function for a given penalty
function is reduced to quadratures: Since symbolic algorithms for evaluat-
ing one-dimensional integrals are very powerful (Bronstein, 2005) and easily
accessible in current computer algebra systems, one will often obtain an ex-
plicit expression for the Gerber-Shiu function. Otherwise one can resort to
standard numerical methods for obtaining approximations.
In the particular case n = 2, m = 1 one has
T = (D + µ) (−c D + λ + δ)2 − λ2 µ,
Z ∞
2
f (u) = λ µ (D + µ) w(u, y − u) e−µy dy.
u
After calculating the particular solution using the initial value from Propo-
sition 6, we obtain the Gerber-Shiu function in the explicit form
2 !
eσu fˆ(ρ1 ) fˆ(ρ2 ) λ
m(u) = − − (ω̂(ρ1 ) − ω̂(ρ2 ))
ρ1 − ρ2 ρ1 − σ ρ2 − σ c
Z ∞
1 1 ρ1 (u−ξ) 1 ρ2 (u−ξ)
− e − e f (ξ) dξ
ρ1 − ρ2 u ρ1 − σ ρ2 − σ
Z u
1 1
+ eσ(u−ξ) f (ξ) dξ,
ρ1 − σ ρ2 − σ 0
where one should recall that ρ1 , ρ2 are the positive roots and σ is the negative
root of the fundamental Lundberg equation. For example, when w(x, y) =
xj y k with j and k positive integers, one obtains
∆µk
ρ2 − σ ρ1 u j! ρ1 − σ 2
σu
m(u) = − j Γ j, (ρ1 + µ)u e + 2 −c e
k!λ2 (ρ1 + µ)j c ρ1 + µ
ρ1 − σ ρ2 u j! ρ2 − σ 2
σu
+ j Γ j, (ρ2 + µ)u e + 2 −c e
(ρ2 + µ)j c ρ2 + µ
ρ1 − ρ2
j Γ j, (σ + µ)u − j! eσu ,
−
(σ + µ)j
where ∆ = (ρ1 −ρ2 )(ρ1 −σ)(ρ2 −σ) is the square root of the discriminant
R ∞ a−1 asso-
ciated to the fundamental Lunderberg equation and Γ(a, x) = x t e−t dt
is the incomplete Gamma function. This formula extends Equation (3.8) of
Cheng and Tang (2003) and similar examples with n = 2 from Li and Garrido
(2004, 2005b) and Gerber and Shiu (2005).
19
u
σ̃ 2 d2
Z
d
(− 2
− c + λ + δ)n m(u) = λn m(u − x)fX (x)dx + λn ω(u), (32)
2 du du 0
20
as an explicit expression for the Gerber-Shiu function. This formula gener-
alizes equation (4.6) of Chen et al. (2007) for the case of exponential claim
sizes and Example 1 of Li and Garrido (2005a) for exponential inter-arrival
times.
9. Conclusion
We have shown that the link between symbolic computation and risk theory
can be mutually fruitful and can be utilized to identify fully explicit expres-
sions for the Gerber-Shiu function in general renewal models in terms of the
employed penalty function. In the presented approach, Laplace transforms
only enter in a very restricted form:
21
Appendix: A Generalized Vandermonde Determinant
For computing the initial values in Proposition 6 we are led to consider the
n × n alternant matrix
p0 (x1 ) · · · pn−1 (x1 )
A = ... .. ..
. .
p0 (xn ) · · · pn−1 (xn )
with polynomials pi (x) = ai,i xi + . . . + ai,0 with ai,i = 1. In the special case
pi (x) = xi this is the usual Vandermonde matrix with the determinant Vn
in the indeterminates x1 , . . . , xn , but det A = Vn holds in general (Kratten-
thaler, 1999, Prop. 1).
We want to compute the (k, l) minor of A, the determinant of the (n −
1)×(n−1) matrix Ak,l obtained by deleting the k-th row and the l-th column.
It suffices to consider
p0 (x1 ) · · · pl−1 (x1 ) pl+1 (x1 ) · · · pn−1 (x1 )
An,l =
.. ... .. .. ... ..
. . . .
p0 (xn−1 ) · · · pl−1 (xn−1 ) pl+1 (xn−1 ) · · · pn−1 (xn−1 )
since
A(x1 , . . . , xn )k,l = A(x1 , . . . , xk−1 , xk+1 , . . . , xn , xk )n,l .
For pi (x) = xi it is known (Heineman, 1929) that det An,l /Vn−1 yields the
elementary symmetric polynomial en−1−l in x1 , . . . , xn−1 .
Proposition A1. We have
n−1
!
det An,l X X
= en−1−l + (−1)j+l+m aj1 ,j2 aj2 ,j3 · · · ajm ,jm+1 en−1−j ,
Vn−1 j=l+1 J
where the inner sum ranges over J = (j1 , . . . , jm+1 ) such that m ≥ 1 and
j = j1 > . . . > jm+1 = l.
Proof. Writing xj for the column vector (xj1 , . . . , xjn−1 )T , the determinant of
the matrix
0 l−1 l+1 n−1
r r r r
P P P P
An,l = a0,r x · · · al−1,r x al+1,r x · · · an−1,r x
r=0 r=0 r=0 r=0
is given by multilinearity as
0
X l−1
X l+1
X n−1
X
··· ··· cr det(xr0 , . . . , xrl−1 , xrl+1 , . . . , xrn−1 )
r0 =0 rl−1 =0 rl+1 =0 rn−1 =0
22
with cr = a0,r0 · · · al−1,rl−1 al+1,rl+1 · · · an−1,rn−1 . Observe that for the first l
indices r0 , . . . , rl−1 there always exist some i < j < l such that ri = rj unless
ri = i for all i < l. Since the determinant vanishes for the cases ri = rj and
the pi are monic (i.e., ai,i = 1), the determinant reduces to
l+1
X n−1
X
··· al+1,rl+1 · · · an−1,rn−1 det(x0 , . . . , xl−1 , xrl+1 , . . . , xrn−1 )
rl+1 =l rn−1 =l
23
Corollary A2. If pi (x) = a0 xi + . . . + ai , a0 = 1, the formula in Proposi-
tion A1 simplifies to
n−1 X
det An,l X
j−l m
X
= en−1−l + (−1) (−1) ad1 . . . adm en−1−j ,
Vn−1 j=l+1 m≥1 d ,...,d 1 m
where the inner sum ranges over d1 , . . . , dm > 0 such that d1 +. . .+dm = j −l.
Using generating functions, we have also
n−1−l
X (−1)j + q(x)j+1
det An,l
= xj en−1−l−j ,
Vn−1 j=0
1 + q(x)
for m ≤ j − l; for m > j − l the sum over the composition is empty. Note
that this even covers the cases m = 0, for which the term is zero except for
j − l = 0, when it becomes one. The stated formula then follows by
j−l
X 1 − (−q(x))j−l+1
(−1)j−l [xj−l ](−q(x))m = (−1)j−l [xj−l ] (34)
m=0
1 + q(x)
j−l (−1)j−l + q(x)j−l+1
= [x ] .
1 + q(x)
24
As a final remark note that the determinant takes an even simpler form
if the pi are not ‘reversed’ as they are in the previous corollary.
Proof. The proof proceeds in a similar way as for the previous corollary. Here
we have the case ai,j = aj , so the inner sum in Proposition A1 evaluates to
j−1
X Y
m
(−1) aj2 · · · ajm ajm+1 = al (ak − 1),
{j2 ,...,jm }⊆{l+1,...,j−1}, k=l+1
j>j2 >···>jm >l
Acknowledgements
We would like to thank Jose Garrido for his encouraging feedback at an
early stage of this project. We would also like to thank Christian Kratten-
thaler for valuable hints on simplifying the determinant of Proposition A1
and Peter Paule and Christoph Koutschan for help with proving the ex-
plicit form of dj in the proof of Proposition 6. Furthermore we thank the
anonymous referees for helpful comments to improve the presentation of the
manuscript. Hansjörg Albrecher was partly supported by the Austrian Sci-
ence Fund Project P18392, Gottlieb Pirsic was partly supported by the Aus-
trian Science Fund Project P19004-N18.
References
Albrecher, H., Boxma, O. J., 2005. On the discounted penalty function in
a Markov-dependent risk model. Insurance: Mathematics & Economics
37 (3), 650–672.
25
Butzer, P. L., Berens, H., 1967. Semi-groups of operators and approxima-
tion. Die Grundlehren der mathematischen Wissenschaften, Band 145.
New York: Springer-Verlag New York Inc.
Chen, Y.-T., Lee, C.-F., Sheu, Y.-C., 2007. An ODE approach for the ex-
pected discounted penalty at ruin in a jump-diffusion model. Finance and
Stochastics 11 (3), 323–355.
Cheng, Y., Tang, Q., 2003. Moments of the surplus before ruin and the deficit
at ruin in the Erlang(2) risk process. North American Actuarial Journal
7 (1), 1–12.
Dickson, D. C. M., Hipp, C., 2001. On the time to ruin for Erlang(2) risk
processes. Insurance: Mathematics & Economics 29 (3), 333–344.
Drekic, S., Stafford, J. E., Willmot, G. E., 2004. Symbolic calculation of the
moments of the time of ruin. Insurance: Mathematics & Economics 34 (1),
109–120.
Gerber, H. U., Shiu, E. S. W., 1997. The joint distribution of the time of ruin,
the surplus immediately before ruin, and the deficit at ruin. Insurance:
Mathematics & Economics 21 (2), 129–137.
Gerber, H. U., Shiu, E. S. W., 1998. On the time value of ruin. North Amer-
ican Actuarial Journal 2 (1), 48–78.
Gerber, H. U., Shiu, E. S. W., 2005. The time value of ruin in a Sparre
Andersen model. North American Actuarial Journal 9 (2), 49–84.
26
Kamke, E., 1967. Differentialgleichungen. Lösungsmethoden und Lösungen.
Teil I: Gewöhnliche Differentialgleichungen, eighth Edition. Vol. 18 of
Mathematik und ihre Anwendungen in Physik und Technik A. Leipzig:
Akademische Verlagsgesellschaft.
Li, S., Garrido, J., 2004. On ruin for the Erlang(n) risk process. Insurance:
Mathematics & Economics 34 (3), 391–408.
Li, S., Garrido, J., 2005a. The Gerber-Shiu function in a Sparre Andersen
risk process perturbed by diffusion. Scandinavian Actuarial Journal (3),
161–186.
Li, S., Garrido, J., 2005b. On a general class of renewal risk process: analysis
of the Gerber-Shiu function. Advances in Applied Probability 37 (3), 836–
856.
Lin, X. S., Willmot, G. E., 2000. The moments of the time of ruin, the surplus
before ruin, and the deficit at ruin. Insurance: Mathematics & Economics
27 (1), 19–44.
Rosenkranz, M., 2005. A new symbolic method for solving linear two-point
boundary value problems on the level of operators. Journal of Symbolic
Computation 39 (2), 171–199.
Sparre Andersen, E., 1957. On the collective theory of risk in case of con-
tagion between claims. Bulletin of the Institute of Mathematics and its
Applications 12, 275–279.
27
Stakgold, I., 2000. Boundary Value Problems of Mathematical Physics. Vol.
I, II. Vol. 29 of Classics in Applied Mathematics. Philadelphia, PA: Society
for Industrial and Applied Mathematics (SIAM), corrected reprint of the
1967–68 original.
Willmot, G. E., 2007. On the discounted penalty function in the renewal risk
model with general interclaim times. Insurance: Mathematics & Economics
41 (1), 17–31.
28