Journal of the Operations Research Society of Japan ⃝

c The Operations Research Society of Japan

Vol. 60, No. 3, July 2017, pp. 337–352

PHASE-TYPE APPROXIMATION OF THE GERBER-SHIU FUNCTION

Kazutoshi Yamazaki
Kansai University

(Received September 9, 2016; Revised January 4, 2017)

Abstract The Gerber-Shiu function provides a way of measuring the risk of an insurance company. It
is given by the expected value of a function that depends on the ruin time, the deficit at ruin, and the
surplus prior to ruin. Its computation requires the evaluation of the overshoot/undershoot distributions of
the surplus process at ruin. In this paper, we use the recent developments of the fluctuation theory and
approximate it in a closed form by fitting the underlying process by phase-type Lévy processes. A sequence
of numerical results are given.

Keywords: Risk management, applied probability

1. Introduction
The fundamental objective of the actuarial ruin theory is to measure the vulnerability to
insolvency. Typically, the surplus of an insurance company is modeled by a stochastic
process, and ruin occurs at the first time it goes below a certain threshold. A most classical
and important quantity of interest is the ruin probability, and the Gerber-Shiu function is
its generalization; it is given as an expected discounted value of the cost function that is
dependent on the ruin time, the deficit at ruin, and the surplus prior to ruin. The evaluation
of the Gerber-Shiu function involves that of the overshoot and undershoot distributions at
the first down-crossing time that do not admit explicit expressions. Hence its computation
is in general a challenging task.
In ruin theory, the surplus process is commonly modeled by a stochastic process with
downward jumps. Due to the premiums received from the insured persons, the surplus
tends to increase constantly. On the other hand, it experiences sudden downward jumps
due to the insurance payments. The classical Cramér-Lundberg model uses a compound
Poisson process with downward jumps. Its generalization called the Sparre-Andersen model
modifies it by allowing the arrival of the claims to follow a general renewal process.
In the last decade, significant progress has been made in insurance mathematics and
ruin theory, thanks to the development of the theory of Lévy processes [9, 19]. In particular,
many results on the Cramér-Lundberg model have been generalized for a general spectrally
negative Lévy process, or the Lévy process with only downward jumps; see, e.g., [4, 6, 7, 24].
This generalization enables one to construct more realistic models; one can, for example,
introduce noise by including Brownian motion and/or infinitesimal jumps of infinite activ-
ity/variation. Using the so-called scale function, one can express concisely many quantities
of interest for a general spectrally negative Lévy process.
The objective of this paper is to give an approximation to the Gerber-Shiu function
using the theory of scale functions. By the compensation formula of the Lévy process,
the Gerber-Shiu function admits an expression as a (double) integral with respect to the
resolvent measure and the Lévy measure. Because the resolvent can be written using the

337
 338 K. Yamazaki

scale function, at least in theory the computation of the Gerber-Shiu function boils down
to that of the scale function.
However, a major obstacle still remains in putting these in practice because scale func-
tions are in general known only up to their Laplace transforms, and only a few cases admit
explicit expressions. The most straightforward approach of computing the scale function is
to apply numerical Laplace inversion as in [18, 33]. However, this approach is not suitable for
the computation of the Gerber-Shiu function because the numerically approximated scale
functions need to be further integrated with respect to the Lévy measure. In particular, the
undershoot density that is essential in the computation of the Gerber-Shiu function tends
to have a very peculiar form with a possible spike. The approximation hence requires a high
precision in computing the scale function and minimal discretization errors in numerical
integration.
In this paper, we adopt a phase-type fitting of [13] by using the scale function for the
class of spectrally negative phase-type Lévy processes, or Lévy processes with negative
phase-type-distributed jumps. Consider a continuous-time Markov chain with some initial
distribution and state space consisting of a single absorbing state and a finite number of
transient states. A phase-type distribution is that of the first entry time to the absorbing
state. As has been discussed in [13, 18], the scale function of this process becomes the sum
of (possibly complex) exponentials; it can be integrated with respect to the Lévy measure
analytically to obtain an explicit form of the Gerber-Shiu function. More importantly,
the class of phase-type distributions is dense in the class of all positive-valued distributions.
Consequently, the Gerber-Shiu function of any given spectrally negative Lévy process can be
approximated in a closed form by that of an approximating spectrally negative phase-type
Lévy process. Our aim is to evaluate numerically the practicability of this approach.
In our numerical results, we focus on the case where the Lévy measure is finite and has a
completely monotone density. In this case, the jump size distribution can be approximated
by a special class of phase-type distributions, called the hyperexponential distributions.
While fitting a phase-type distribution for a general distribution is often difficult, fitting
a hyperexponential distribution for the ones with completely monotone densities can be
efficiently done, for example by the algorithm by [14] that is guaranteed to converge to the
desired distribution. The class of Lévy processes with completely monotone Lévy densities
includes, for example, a subset of compound Poisson processes, variance gamma [26, 27],
CGMY [10], generalized hyperbolic [12] and normal inverse Gaussian [5] processes.
In order to evaluate our approach, we obtain, for the hyperexponential case, the closed-
form expressions of the (discounted) overshoot/undershoot distributions at the first down-
crossing time, and use it to approximate those for the processes with Weibull/Pareto-
distributed jumps. The obtained results are then compared with those obtained by Monte
Carlo simulation.
To our best knowledge, this is the first paper on the numerical evaluation of the Gerber-
Shiu function via phase-type fitting. As the Gerber-Shiu measure is sensitive to approxima-
tion errors, it is important to evaluate its numerical performance. Recently, the resolvents
of related extensions of the Lévy process have been developed, and it is reasonable to con-
jecture that the Gerber-Shiu function of these can be approximated precisely in the same
way.
The rest of the paper is organized as follows. Section 2 reviews the spectrally negative
Lévy process, the Gerber-Shiu function, and the scale function. Section 3 gives a summary
of [13] on the spectrally negative phase-type Lévy process and its scale function. Section
4 computes the Gerber-Shiu measure of the hyperexponential Lévy process as an approxi-

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 Phase-Type Approximation 339

mation for the case the jump size distribution admits a completely monotone density. We
evaluate the performance using numerical results in Section 5. Section 6 concludes with
remarks on the cases of other variants of spectrally negative Lévy processes.

2. Gerber-Shiu Functions for Spectrally Negative Lévy Processes

We let (Ω, F, P) be a probability space hosting a spectrally negative Lévy process X =
{Xt ; t ≥ 0} that models the surplus of a company. Let Px be the conditional probability
under which X0 = x (and also P ≡ P0 ), and F := {Ft : t ≥ 0} the filtration generated by
X. The process X is uniquely characterized by its Laplace exponent
∫
[ sX1 ] 1 2 2
ψ(s) := log E e = cs + σ s + (esz − 1 − sz1{z>−1} )Π(dz), s ≥ 0, (2.1)
2 (−∞,0)

where σ ≥ 0 is the diffusion (Brownian motion) coefficient and ∫ Π is a Lévy measure with
the support (−∞, 0) that satisfies the integrability condition (−∞,0) (1 ∧ |z| )Π(dz) < ∞. It
2

has paths of bounded variation if and only if

∫
σ = 0 and (1 ∧ |z|)Π(dz) < ∞;
(−∞,0)

see, for example, Lemma 2.12 of [19]. In this case, we can rewrite the Laplace exponent
(2.1) by
∫
ψ(s) = µs + (esz − 1)Π(dz),
(−∞,0)

with
∫
µ := c − zΠ(dz).
(−1,0)

We disregard the case when X is the negative of a subordinator (or decreasing a.s.).
2.1. Gerber-Shiu functions
Define the ruin time as the first time the surplus goes below zero:

τ0− := inf {t ≥ 0 : Xt < 0} .

Here and throughout the paper, we use the convention that inf ∅ = ∞. On the event
{τ0− < ∞}, the random variables Xτ0− and Xτ0− − model, respectively, the deficit at ruin and
surplus immediately before ruin.
Fix f : (−∞, 0] × [0, ∞) → [0, ∞) bounded and measurable. We define the Gerber-Shiu
function
[ − ( ) ]
GSf (x, q) := Ex e−qτ0 f Xτ0− , Xτ0− − ; τ0− < ∞ .

With the Gerber-Shiu measure

[ −
]
K (q) (x, dy, dz) := Ex e−qτ0 ; Xτ0− ∈ dy, Xτ0− − ∈ dz, τ0− < ∞ , x, z > 0, y < 0,

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 340 K. Yamazaki

we can write
∫ ∫
GSf (x, q) = f (y, z)K (q) (x, dy, dz);
(0,∞) (−∞,0)

see pages 4 and 5 of [20].

Using the compensation formula (see Theorem 4.4 of [19]), this can be written in terms
of the q-resolvent measure of X killed on exiting [0, ∞):
∫ ∞
(q)
R (x, dz) := e−qt Px {Xt ∈ dz, τ0− > t}dt, x, z > 0, (2.2)
0

which is known to admit a density r(q) for the case of spectrally negative Lévy processes
such that

R(q) (x, dz) = r(q) (x, z)dz, x, z > 0; (2.3)

see (2.10) below for the form of r(q) . As in Section 1.3 of [20], we can write

K (q) (x, dy, dz) = Π(dy − z)r(q) (x, z)dz. (2.4)

Hence the computation of the Gerber-Shiu measure boils down to that of the resolvent
measure.
2.2. Scale functions
In order to compute the resolvent measure (2.2), we shall introduce the scale function.
Fix q ≥ 0. The scale function W (q) : R → [0, ∞) of X is a function whose Laplace
transform is given by
∫ ∞
1
e−sx W (q) (x)dx = , s > Φ(q) (2.5)
0 ψ(s) − q

where

Φ(q) := sup{s ≥ 0 : ψ(s) = q}, q ≥ 0. (2.6)

On the negative half line, it is assumed that W (q) (x) = 0.

Regarding the smoothness of the scale function, if the Lévy measure has no atoms or X
is of unbounded variation, then W (q) ∈ C 1 (0, ∞); if it has a Gaussian component (σ > 0),
then W (q) ∈ C 2 (0, ∞). See [11] for other known results on the smoothness.
The behavior in the neighborhood of zero is given as follows. As in Lemmas 4.3 and 4.4
of [23], for every q ≥ 0,
{ }
(q) 0, if X is of unbounded variation
W (0) = 1 ,
µ
, if X is of bounded variation
 2 
 σ2 , if σ > 0  (2.7)
(q)′ ∞, if σ = 0 and Π(−∞, 0) = ∞ .
W (0+) =
 q+Π(−∞,0) 
µ2
, if X is compound Poisson

We plot in Figure 1 sample plots of the scale function and its derivative for the bounded
variation case with σ = 0 and the unbounded variation case with σ > 0.

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 Phase-Type Approximation 341

3.5 2
σ = 0 σ = 0
σ = 1 σ = 1
3

1.5
2.5

2
scale function

derivative
1

1.5

1
0.5

0.5

0 0
0 5 10 15 20 0 5 10 15 20
x x

Figure 1: Sample plots of the scale function (left) and its derivative (right). The ones
indicated in red (resp. blue) are those for the case of bounded (resp. unbounded) variation.
As in (2.7), it is confirmed that it vanishes at zero if and only if it is of unbounded variation.

The most well-known application of the scale function can be found in the two-sided exit
identities. Let us define the first down- and up-crossing times, respectively, of X by

τb− := inf {t > 0 : Xt < b} and τb+ := inf {t > 0 : Xt > b} , b ∈ R. (2.8)

Then, for any b > 0 and x ≤ b,

[ ] W (q) (x)
−qτb+
E e
x
1{τ + <τ − } = ,
b 0 W (q) (b)
[ −
] W (q) (x)
Ex e−qτ0 1{τ + >τ − } = Z (q) (x) − Z (q) (b) (q) , (2.9)
b 0 W (b)
[ −
] q
Ex e−qτ0 = Z (q) (x) − W (q) (x),
Φ(q)
where
∫ x
Z (q)
(x) := 1 + q W (q) (y)dy, x ∈ R.
0

For a comprehensive account on the scale function, see among others [18, 19].
2.3. Resolvent via the scale function
Recall, as in our discussion in Section 2.1, that the Gerber-Shiu function can be written in
terms of the resolvent measure. This can be concisely written in terms of the scale function
as follows: by Corollary 8.8 of [19], the resolvent density (2.3) can be written

r(q) (x, z) = e−Φ(q)z W (q) (z) − W (q) (x − z), x, z > 0. (2.10)

Now in view of the identity (2.4), the computation of the Gerber-Shiu measure boils down
to that of the scale function.

3. Scale Functions for Spectrally Negative Phase-type Lévy Processes

As discussed in the previous section, the scale function is defined by its Laplace transform
and for its computation the Laplace transform (2.5) must be inverted. Here, we review the
results on [13] for a special class of Lévy processes where it can be inverted analytically.

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 342 K. Yamazaki

3.1. Phase-type distribution

Consider a continuous-time Markov chain

Y = {Yt ; t ≥ 0}

with a finite state space {1, . . . , m} ∪ {∆} where 1, . . . , m are transient and ∆ is absorbing.
Its initial distribution is given by a simplex α = [α1 , . . . , αm ] such that αi = P {Y0 = i} for
every i = 1, . . . , m. The intensity matrix Q is partitioned into the m transient states and
the absorbing state ∆, and is given by
[ ]
T t
Q := .
0 0

Here T is an m × m-matrix called the phase-type generator, and t = −T 1 where 1 =

[1, . . . , 1]′ . A distribution is called phase-type with representation (m, α, T ) if it is the
distribution of the absorption time to ∆ in the Markov chain described above. It is known
that T is non-singular and thus invertible; see [2]. Its distribution and density functions are
given, respectively, by

F (z; α, T ) = 1 − αeT z 1 and f (z; α, T ) = αeT z t, z > 0.

3.2. Phase-type Lévy processes

Let X = {Xt ; t ≥ 0} be a spectrally negative Lévy process of the form

∑
Nt
Xt − X0 = µt + σBt − Zn , 0 ≤ t < ∞, (3.1)
n=1

for some µ ∈ R and σ ≥ 0 (with µ > 0 when σ = 0 so that it is not the negative of a
subordinator). Here B = {Bt ; t ≥ 0} is a standard Brownian motion, N = {Nt ; t ≥ 0}
is a Poisson process with arrival rate λ, and Z = {Zn ; n = 1, 2, . . .} is an i.i.d. sequence of
phase-type-distributed random variables with representation (m, α, T ). These processes are
assumed mutually independent. Its Laplace exponent is then
1 ( )
ψ(s) = µs + σ 2 s2 + λ α(sI − T )−1 t − 1 , (3.2)
2
which is analytic for every s ∈ C except at the eigenvalues of T .
Let Iq be the set of (the sign-changed) negative roots:

Iq := {i : ψ(−ξi,q ) = q and R(ξi,q ) > 0} , (3.3)

where R(z) is the real part of z ∈ C. Let n denote the number of different roots in Iq and
mi the multiplicity of a root ξi,q for i = 1, . . . , n.
Because the Laplace exponent (3.2) has a rational form, it can be inverted analytically by
partial fraction decomposition. Hence, in view of (2.5), the scale function can be obtained.
Proposition 3.1 (Section 5.4 of [18] and Proposition 2.1 of [13]). Suppose q ≥ 0 and
ψ ′ (0+) < 0 if q = 0. Then the scale function is written

eΦ(q)x ∑∑ n
i m
(k) x
k−1
W (q)
(x) = ′ − Ci,q e−ξi,q x , x ≥ 0, (3.4)
ψ (Φ(q)) i=1 k=1 (k − 1)!

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 Phase-Type Approximation 343

where

(k) 1 ∂ mi −k (s + ξi,q )mi 
Ci,q := , 1 ≤ k ≤ mi and 1 ≤ i ≤ n.
(mi − k)! ∂smi −k q − ψ(s) s=−ξi,q

In particular, if all the roots in Iq are distinct, then

eΦ(q)x ∑ n
W (q) (x) = − Ci,q e−ξi,q x , x ≥ 0, (3.5)
ψ ′ (Φ(q)) i=1

where

s + ξi,q  1
Ci,q :=  =− ′ .
q − ψ(s) s=−ξi,q ψ (−ξi,q )

3.3. Approximation results

It is known that the class of phase-type distributions is dense in the class of all positive-
valued distributions. Using this, Proposition 1 of [3] shows that, there exists, for any spec-
trally negative Lévy process X, a sequence of spectrally negative phase-type Lévy processes
(n)
X (n) converging to X in D[0, ∞). In other words, X1 → X1 in distribution by Corollary
VII 3.6 of [15]; see also [32].
Using these results, Egami and Yamazaki [13] study the convergence of the corresponding
scale function, and show that the approximation is in most cases very accurate. On the other
hand, as they also point out, no existing algorithm is guaranteed to construct a converging
sequences and fitting phase-type distributions can get difficult; see, e.g., the case of fitting
for the uniform distributed jump size in [13].
On the other hand, as we shall discuss next, it is guaranteed to work for the case the
jump size admits a completely monotone density.
3.4. Hyperexponential case
As an important example where all the roots in Iq are distinct and real, we consider the
case where Z has a hyperexponential distribution with a density function

∑
m
f (z) = αj ηj e−ηj z , z > 0,
j=1

for some 0 < η1 < · · · < ηm < ∞ and αj > 0 for 1 ≤ j ≤ m such that α1 + · · · + αm = 1;
this is the phase-type distribution with its Markov chain such that the m transient states
are connected only to ∆. Its Laplace exponent (2.1) is then

1 2 2 ∑m
s
ψ(s) = µs + σ s − λ αj .
2 j=1
ηj + s

Notice in this case that −η1 , . . . , −ηm are the poles of the Laplace exponent. Furthermore,
all the roots in Iq are distinct and real and satisfy the following interlacing condition for
every q > 0:
1. for σ > 0, there are m + 1 roots −ξ1,q , . . . , −ξm+1,q such that

0 < ξ1,q < η1 < ξ2,q < · · · < ηm < ξm+1,q < ∞; (3.6)

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 344 K. Yamazaki

2. for σ = 0, there are m roots −ξ1,q , . . . , −ξm,q such that

0 < ξ1,q < η1 < ξ2,q < · · · < ξm,q < ηm < ∞. (3.7)

Because all roots are real and distinct, the scale function can be written as (3.5).
Recall that a density function f of a positive valued random variable is called completely
monotone if all the derivatives exist and, for every n ≥ 1,

(−1)n f (n) (x) ≥ 0, x ≥ 0,

where f (n) denotes the nth derivative of f .

Feldmann and Whitt [14] showed that if a density function is completely monotone,
then it can be approximated by those of hyperexponential distributions. As shown by [8],
every completely monotone density function is a mixture of exponential density functions,
and this implies that, for any distribution with a completely monotone density, there exists
a sequence of hyperexponential distributions converging to it. The class of distributions
with completely monotone densities contains a number of distributions such as the (subset
of) Pareto distribution, the Weibull distribution, and the gamma distribution. Feldmann
and Whitt [14] took advantage of this fact and proposed a recursive algorithm for fitting
hyperexponential distributions to these distributions. We refer the reader to [1, 17] for other
approximation methods.

4. Computation of the Gerber-Shiu Function

We shall now consider the approximation of the Gerber-Shiu function using the fitted scale
functions. Here, we consider the hyperexponential case as discussed in Section 3.4; in
this case, because the scale function and Lévy measure are both written as mixtures of
exponential functions, we attain closed-form expressions.
4.1. Overshoot and undershoot distributions
Recall from (2.4) regarding the equivalence of the Gerber-Shiu measure and the product
of the Lévy measure and the resolvent measure. If we define for all A ∈ B(−∞, 0) and
B ∈ B(0, ∞),
[ ] ∫
−qτ0− −
hq (x; A, B) := E e x
; Xτ0− ∈ A, Xτ0− − ∈ B, τ0 < ∞ = K (q) (x, dy, dz). (4.1)
A×B

Combining (2.4) and (2.10), we can write

∫ ∞ { ∫ ∫ }
−Φ(q)y
hq (x; A, B) = W (q)
(x) e dy − W (q)
(x − y)dy Π(du),
0 B∩(A+u) B∩(A+u)

where Π is the Lévy measure of the dual process −X.

When X is a phase-type Lévy process, as we have studied in the previous section W (q) (x)
can be written as a sum of (possibly∑complex) exponentials. In particular, if it is hyperex-
−ηj u
ponential, we can write Π(du) = λ m j=1 αj ηj e du for all u ∈ (0, ∞), and hence hq can
be obtained analytically.
Here we assume that X is a hyperexponential Lévy process and q > 0. The following
results are immediate by straightforward integration.

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 Phase-Type Approximation 345

Proposition 4.1. 1. Suppose B = (b, b) and A = (−a, −a) for some 0 ≤ a ≤ a and
0 ≤ b ≤ b. Then
∑
m
hq (x; A, B) = λ αj (e−ηj a − e−ηj a )κj,q (x; B)
j=1

where, for each 1 ≤ j ≤ m,

eΦ(q)x ( )
−(ηj +Φ(q))(b∨x) −(ηj +Φ(q))(b∨x)
κj,q (x; B) := ′ e −e
ψ (Φ(q))(ηj + Φ(q))
∑ [ ( )
−ξi,q x 1
+ Ci,q e e−(ηj −ξi,q )(b∧x) − e−(ηj −ξi,q )(b∧x)
i∈Iq
ηj − ξi,q
1 ( )]
−(ηj +Φ(q))b −(ηj +Φ(q))b
− e −e .
ηj + Φ(q)
2. We have
[ ] ∑
m
−qτ0− −
E e
x
; −Xτ0− ∈ da, Xτ0− − ∈ B, τ0 < ∞ = λ αj ηj e−ηj a κj,q (x; B),
j=1
[ −
] ∑
m
Ex e−qτ0 ; Xτ0− ∈ A, Xτ0− − −
∈ db, τ0 < ∞ = λ αj (e−ηj a − e−ηj a )
j=1
{ ∑ ( −(η −ξ )b ) }
−ξi,q x −(ηj +Φ(q))b
C e e j i,q
− e , b < x
× i∈Iq i,q ∑ .
1
ψ ′ (Φ(q))
eΦ(q)x−(ηj +Φ(q))b − i∈Iq Ci,q e−(ξi,q x+(ηj +Φ(q))b) , b ≥ x

In particular, by setting B = (0, ∞) (A = (−∞, 0)),

[ ] ∑
m
−qτ0− −
E e
x
; −Xτ0− ∈ da, τ0 < ∞ = λ αj ηj e−ηj a κj,q (x; (0, ∞)),
j=1
[ −
]
Ex e−qτ0 ; Xτ0− − ∈ db, τ0− < ∞
{ ∑m ∑ ( )
λ j=1 αj i∈Iq Ci,q e−ξi,q x e−(ηj −ξi,q )b − e−(ηj +Φ(q))b , b < x,
= ∑m [ ∑ ] (4.2)
1
λ j=1 αj ψ′ (Φ(q)) eΦ(q)x−(ηj +Φ(q))b − i∈Iq Ci,q e−(ξi,q x+(ηj +Φ(q))b) , b ≥ x,

where
1
κj,q (x; (0, ∞)) = e−ηj x
ψ ′ (Φ(q))(η + Φ(q))
j
∑ [ ]
1 ( −ξi,q x −ηj x
) −ξi,q x 1
+ Ci,q e −e −e .
ηj − ξi,q ηj + Φ(q)
i∈I q

5. Numerical Results
Using the identities obtained in Proposition 4.1, we shall evaluate the efficiency of the
phase-type fitting approach of the Gerber-Shiu function.
Here we consider the spectrally negative Lévy processes X (weibull) and X (pareto) in the
form (3.1) where Z is

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 346 K. Yamazaki

(i) Weibull(0.6, 0.665), and

(ii) Pareto(1.2, 5)
respectively. Recall that the Weibull distribution with parameters c and a, denoted by
Weibull(c, a), is given by F (t) = 1 − e−(t/a) , t ≥ 0, and the Pareto distribution with
c

positive parameters a and b, denoted by Pareto(a, b), is given by F (t) = 1−(1+bt)−a , t ≥ 0.

See [16] for more details about these distributions. With the choice of our parameters, the
corresponding Lévy densities are thus completely monotone.
As has been noted in Section 3.4, any spectrally negative Lévy process with a completely
monotone Lévy density can be approximated arbitrarily closely by fitting hyperexponential
distributions. Here, we use the fitted data computed by [14] to approximate the scale
function for X (weibull) and X (pareto) (with or without a Brownian motion component). Tables
3 and 9, respectively, of [14] show the parameters of the hyperexponential distributions
obtained by [14] fitted to (i) with m = 6 and to (ii) with m = 14. We use these parameters
to construct hyperexponential Lévy processes X e (weibull) and X
e (pareto) (see Section 3.4) that
(weibull) (pareto)
will be used to approximate X and X , respectively.

Table 1: Parameters of the hyperexponential distribution fitted to Weibull(0.6,0.665) and

Pareto(1.2,5) (taken from Tables 3 of [14]).
i αi ηi i αi ηi i αi ηi
1 0.029931 676.178 1 8.37E-11 8.3E-09 8 0.000147 0.0020
2 0.093283 38.7090 2 7.18E-10 6.8E-08 9 0.001122 0.0100
3 0.332195 4.27400 3 5.56E-09 3.9E-07 10 0.008462 0.0570
4 0.476233 0.76100 4 4.27E-08 2.2E-06 11 0.059768 0.3060
5 0.068340 0.24800 5 3.27E-07 1.2E-05 12 0.307218 1.5460
6 0.000018 0.09700 6 2.50E-06 6.5E-05 13 0.533823 6.5160
– – – 7 1.92E-05 3.5E-04 14 0.089437 23.304
(i) Weibull(0.6, 0.665) (ii) Pareto(1.2, 5)

In order to evaluate the errors associated with the phase-type-fitting of the Gerber-Shiu
function, we shall consider approximating the overshoot/undershoot density
[ ] [ ]
−qτ0− − −qτ0− −
E e
x
; −Xτ0 ∈ da, τ0 < ∞
− and E ex
; Xτ0 − ∈ db, τ0 < ∞ ,
−

for X (weibull) and X (pareto) . The phase-type-fitting approach approximates them by comput-
ing those for the approximating hyperexponential Lévy processes X e (weibull) and X
e (pareto) .
These can be done analytically by the identity (4.2).
We evaluate these results by comparing with the simulated results. For X (weibull) and
X (pareto) , we simulate
[ −
]
Ex e−qτ0 ; −Xτ0− ∈ (a − ∆a/2, a + ∆a/2), τ0− < ∞ /∆a, and
[ ]
−qτ0− −
E e
x
; Xτ0− − ∈ (b − ∆b/2, b + ∆b/2), τ0 < ∞ /∆b

with ∆a = ∆b = 0.1 by Monte Carlo simulation with 500, 000 samples.

In order to confirm the accuracy of the simulated results, we also compute the results
using these two methods for X (exp) corresponding to the process where the jump size is

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 Phase-Type Approximation 347

exponential with parameter 1; this is a special case of the phase-type (and hyperexponential)
Lévy process and hence obtained results are exact.
Figures 2 and 3 show the results for the cases σ = 1 and σ = 0 with common parameters
x = 5, µ = 1, λ = 10 and q = 0.05. From the results on the exponential case, we can confirm
that the Monte Carlo simulated results are accurate. In Figure 3, the density has a jump
at the initial position x = 5 for the case σ = 0 (while it is continuous for the case σ = 1)
due to the fact that x = 5 is irregular for (−∞, 5) (see Definition 6.4 of [19]). As can be
seen from these figures, the approximation accurately captures the overshoot/undershoot
densities for X (weibull) and X (pareto) . The spike at the initial position in Figure 3 is precisely
realized thanks to the closed-form expression (4.2); this would be difficult to realize if the
scale function is approximated via numerical Laplace inversion.

6. Concluding Remarks
In this paper, we studied and evaluated the performance of the phase-type fitting approach
(motivated by [13]) for computing the Gerber-Shiu function for the spectrally negative
Lévy process. The method is overall accurate and is powerful in that it can obtain an
approximation in a closed form.
While this paper focused on the case of a spectrally negative Lévy process, the proposed
method can easily be generalized. By the compensation formula, the decomposition (2.4)
holds for a more general class of stochastic processes. Therefore, by simply replacing the
resolvent measure r(q) , the Gerber-Shiu function can be computed in the same way.
Thanks to the recent developments of the fluctuation theories, the resolvents of various
extensions of the spectrally negative Lévy process are now available in terms of the scale
function. Here, we list several known examples of the resolvent.
1. The fluctuation theories of reflected spectrally negative Lévy processes are well-developed.
In the optimal dividend problem, where one wants to maximize the expected net present
value of dividends until ruin, it is in many cases shown to be optimal to reflect the surplus
process at a suitable boundary. It is therefore of interest to investigate the Gerber-Shiu
function of the reflected process to evaluate the risk of an dividend paying company. As
given in [31], the resolvent requires the derivative or the integral of the scale function,
depending on whether the reflection barrier is upper or lower. For the doubly reflected
case with both upper and lower barriers, see [30].
2. As a variant of the reflected process, the refracted spectrally negative Lévy process of [21]
changes its drift by δ > 0 whenever it is above a threshold b – it is the unique strong
solution U to the stochastic differential equation

dUt = dXt − δ1{Ut >b} dt, t ≥ 0.

In insurance, this can be used to model the surplus of a dividend paying company when
the dividend rate must be bounded from above by δ (see [22]). The resolvent is given
in [21]. The resolvents for the cases with additional classical reflection have recently
been obtained by [28, 29].
3. Given two levels s and S, the classical (s, S)-policy controls the process by pushing
the process to S immediately when it goes above or below s. It is, under a suitable
condition, an optimal strategy in the optimal dividend problem in the presence of a
fixed cost (see [7, 25]).The resolvent is obtained in [35]. For its two-sided cases with
four-parameter (d, D, U, u) policy, see [34].

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
348 K. Yamazaki

via scale function via scale function

via simulation via simulation
1 1

0.8 0.8
density

density
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
overshoot overshoot

Exp(1) with σ = 1 Exp(1) with σ = 0

via scale function via scale function

via simulation via simulation
1 1

0.8 0.8
density

density

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
overshoot overshoot

Weibull(0.6,0.665) with σ = 1 Weibull(0.6,0.665) with σ = 0

via scale function via scale function

via simulation via simulation
1 1

0.8 0.8
density

density

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 2 4 6 8 10 0 2 4 6 8 10
overshoot overshoot

Pareto(1.2,5) with σ = 1 Pareto(1.2,5) with σ = 0

−
Figure 2: Computation of the overshoot density Ex [e−qτ0 ; −Xτ0− ∈ da, τ0− < ∞]. The solid
lines indicate the fitted density functions and red marks indicate the values obtained from
simulation.

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 Phase-Type Approximation 349

0.8 0.8
via scale function via scale function
0.7 via simulation 0.7 via simulation

0.6 0.6

0.5 0.5
density

density
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 2 4 6 8 10 0 2 4 6 8 10
undershoot undershoot

Exp(1) with σ = 1 Exp(1) with σ = 0

0.8 0.8
via scale function via scale function
0.7 via simulation 0.7 via simulation

0.6 0.6

0.5 0.5
density

density

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 2 4 6 8 10 0 2 4 6 8 10
undershoot undershoot

Weibull(0.6,0.665) with σ = 1 Weibull(0.6,0.665) with σ = 0

0.8 0.8
via scale function via scale function
0.7 via simulation 0.7 via simulation

0.6 0.6

0.5 0.5
density

density

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 2 4 6 8 10 0 2 4 6 8 10
undershoot undershoot

Pareto(1.2,5) with σ = 1 Pareto(1.2,5) with σ = 0

−
Figure 3: Computation of the undershoot density Ex [e−qτ0 ; Xτ0− − ∈ db, τ0− < ∞]. The solid
lines indicate the fitted density functions and red marks indicate the values obtained from
simulation.

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 350 K. Yamazaki

In these examples, the resolvent r(q) is written in terms of the scale function, and hence
as in the cases considered in this paper, those for the phase-type case can be written as
(a linear combination of) exponential forms. Consequently, the Gerber-Shiu function can
be analytically computed in the same way. In view of the results obtained in this paper,
the phase-type error is minimal and the same procedure is expected to give an accurate
approximation of the Gerber-Shiu function.

References
[1] H. Albrecher, F. Avram, and D. Kortschak: On the efficient evaluation of ruin prob-
abilities for completely monotone claim distributions. Journal of Computational and
Applied Mathematics, 233-10 (2010), 2724–2736.
[2] S. Asmussen: Fitting phase-type distributions via the EM algorithm. Scandinavian
Journal of Statistics, 23 (1996), 419–441.
[3] S. Asmussen, F. Avram, and M.R. Pistorius: Russian and American put options under
exponential phase-type Lévy models. Stochastic Processes and their Applications, 109-
1 (2004), 79–111.
[4] F. Avram, Z. Palmowski, and M.R. Pistorius: On the optimal dividend problem for
a spectrally negative Lévy process. The Annals of Applied Probability, 17-1 (2007),
156–180.
[5] O.E. Barndorff-Nielsen: Processes of normal inverse Gaussian type. Finance and
Stochastics, 2-1 (1998), 41–68.
[6] E. Bayraktar, A.E. Kyprianou, and K. Yamazaki: On optimal dividends in the dual
model. Astin Bulletin, 43-3 (2013), 359–372.
[7] E. Bayraktar, A.E. Kyprianou, and K. Yamazaki: Optimal dividends in the dual model
under transaction costs. Insurance: Mathematics Economics, 54 (2014), 133–143.
[8] S. Bernstein: Sur les fonctions absolument monotones. Acta Mathematica, 52-1 (1929),
1–66.
[9] J. Bertoin: Lévy Processes (Cambridge University Press, Cambridge, 1996).
[10] P. Carr, H. Geman, D.B. Madan, and M. Yor: The structure of asset returns: An
empirical investigation. Journal of Business, 75 (2002), 305–332.
[11] T. Chan, A.E. Kyprianou, and M. Savov: Smoothness of scale functions for spectrally
negative Lévy processes. Probability Theory and Related Fields, 150-3 (2011), 691–708.
[12] E. Eberlein, U. Keller, and K. Prause: New insights into smile, mispricing and value
at risk: The hyperbolic model. Journal of Business, 71 (1998), 371–405.
[13] M. Egami and K. Yamazaki: Phase-type fitting of scale functions for spectrally negative
Lévy processes. Journal of Computational and Applied Mathematics, 264 (2014), 1–22.
[14] A. Feldmann and W. Whitt: Fitting mixtures of exponentials to long-tail distributions
to analyze network performance models. Performance Evaluation, 31 (1998), 245–279.
[15] J. Jacod and A.N. Shiryaev: Limit Theorems for Stochastic Processes (Springer-Verlag,
Berlin, 2003).
[16] N. Johnson and S. Kotz: Distributions in Statistics: Continuous Multivariate Distri-
butions (John Wiley & Sons Inc., New York, 1972).
[17] D.W. Kammler: Chebyshev approximation of completely monotonic functions by sums
of exponentials. SIAM Journal on Numerical Analysis, 13-5 (1976), 761–774.

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
 Phase-Type Approximation 351

[18] A. Kuznetsov, A.E. Kyprianou, and V. Rivero: The theory of scale functions for spec-
trally negative Lévy processes. In Lévy Matters II, Springer Lecture Notes in Mathe-
matics (Springer-Verlag, Berlin, 2013), 97–186.
[19] A.E. Kyprianou: Fluctuations of Lévy Processes with Applications, 2nd Edition
(Springer-Verlag, Berlin, 2006).
[20] A.E. Kyprianou: Gerber-Shiu Risk Theory (Springer-Verlag, Berlin, 2013).
[21] A.E. Kyprianou and R. Loeffen: Refracted Lévy processes. Annales de l’Institut Henri
Poincaré, Probabilités et Statistiques, 46-1 (2010), 24–44.
[22] A.E. Kyprianou, R. Loeffen, and J.-L. Pérez: Optimal control with absolutely contin-
uous strategies for spectrally negative Lévy processes. Journal of Applied Probability,
49-1 (2012), 150–166.
[23] A.E. Kyprianou and B. Surya: Principles of smooth and continuous fit in the de-
termination of endogenous bankruptcy levels. Finance and Stochastics, 11-1 (2007),
131–152.
[24] R.L. Loeffen: On optimality of the barrier strategy in de Finetti’s dividend problem
for spectrally negative Lévy processes. The Annals of Applied Probability, 18-5 (2008),
1669–1680.
[25] R.L. Loeffen: An optimal dividends problem with transaction costs for spectrally neg-
ative Lévy processes. Insurance: Mathematics and Economics, 45-1 (2009), 41–48.
[26] D.B. Madan, P. Carr, and E.C. Chang: The variance gamma processes and option
pricing. European Finance Review, 2 (1998), 79–105.
[27] D.B. Madan and F. Milne: Option pricing with VG martingale components. Mathe-
matical Finance, 1-4 (1991), 39–55.
[28] J.-L. Pérez and K. Yamazaki: On the refracted-reflected spectrally negative Lévy pro-
cesses. To appear in Stochastic Processes and their Applications.
[29] J.-L. Pérez and K. Yamazaki: Refraction-reflection strategies in the dual model. Astin
Bulletin, 47-1 (2017), 199–238.
[30] M.R. Pistorius: On doubly reflected completely asymmetric Lévy process. Stochastic
Processes and their Applications, 107-1 (2003), 131–143.
[31] M.R. Pistorius: On exit and Ergodicity of the spectrally one-sided Lévy process re-
flected at its infimum. Journal of Theoretical Probability, 17-1 (2004), 183–220.
[32] M.R. Pistorius: On maxima and ladder processes for a dense class of Lévy process.
Journal of Applied Probability, 43-1 (2006), 208–220.
[33] B. Surya: Evaluating scale functions of spectrally negative Lévy processes. Journal of
Applied Probability, 45-1 (2008), 135–149.
[34] K. Yamazaki: Cash management and control band policies for spectrally one-sided Lévy
processes. In M. Kijima, Y. Muromachi, and T. Shibata (eds.): Recent Advances in
Financial Engineering 2014: Proceedings of the TMU Finance Workshop 2014 (World
Scientific, 2016), 199–215.
[35] K. Yamazaki: Inventory control for spectrally positive Lévy demand processes. Math-
ematics of Operations Research, 42-1 (2017), 212–237.

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.
352 K. Yamazaki

Kazutoshi Yamazaki
Department of Mathematics
Faculty of Engineering Science
Kansai University
3-3-35 Yamate-cho, Suita-shi,
Osaka 564-8680, Japan
E-mail: kyamazak@kansai-u.ac.jp

Copyright ⃝
c by ORSJ. Unauthorized reproduction of this article is prohibited.