Optimal Control of Capital Injections by

Reinsurance and Investments

I NAUGURAL - D ISSERTATION
ZUR

E RLANGUNG DES D OKTORGRADES

DER M ATHEMATISCH -N ATURWISSENSCHAFTLICHEN FAKULT ÄT
DER U NIVERSIT ÄT ZU K ÖLN
VORGELEGT VON

Julia Eisenberg

aus Ašhabad
1. Berichterstatter: Prof. Dr. Hanspeter Schmidli, Universität zu Köln

2. Berichterstatter: Prof. Dr. Josef G. Steinebach, Universität zu Köln

Tag der mündlichen Prüfung: 04.02.2010

Acknowledgements
I would like to express my thanks to my supervisor Prof. Dr. Hanspeter Schmidli. I have
greatly profited from his hints and suggestions generously given during our conversations.
I am very grateful to Prof. Dr. Josef Steinebach who agreed to be coreferent.
I would like to acknowledge Markus Schulz for his listening ear and Mrs Anderka for
believing in me.
In particular, I would like to thank my family who has supported me all the time and
enabled a successful completion of this work.
Abstract
An insurance company, having an initial capital x, cashes premiums continuously and
pays claims of random sizes at random times. In addition to that, the company can buy
reinsurance or/and invest money into a riskless or risky assets. The company holders
are confronted with the problem of taking decisions on a business policy of the company.
Thus, a measure for the risk connected with an insurance portfolio is sorely needed.

The ruin probability, i.e. the probability that the surplus process becomes negative in
finite time, is typically the measure for an insurance company’s solvency. However, the
ruin probability approach has been criticised among other things for not considering the
severity of an insolvency and for ignoring the time value of money.

An alternative to measure the risk of a surplus process is to consider the value of ex-
pected discounted capital injections, which are necessary to keep the process above zero.
Naturally, it raises the question how to minimise this value. If the company holders
prefer (or are indifferent) investing tomorrow to investing today, it is optimal to inject
capital only when the surplus becomes negative and only as much as is necessary to keep
the process above zero.

In the first part of this work, we solve the problem of minimising the expected discounted
capital injections over all dynamic reinsurance strategies for the classical risk model and
its diffusion approximation. In the second part, we extend the concept by adding the
possibility of investing money, if the surplus remains positive, into a riskless asset. In
these two cases we are able to show the existence and uniqueness of the optimal rein-
surance strategy and the value function as the minimising value of expected discounted
capital injections.

In the third part, we consider the surplus process, where the company holders can invest
money into a risky asset modeled as a Black-Scholes model. The forth part extends
the setup of the third part by possibility of reinsurance. In the last two cases we solve
the problem explicitly only for the case of diffusion approximation. In the classical risk
model the concept of viscosity solutions introduced by Crandall and Lions has been used.

All the studies are illustrated by simulations, written in Java.

Zusammenfassung
Eine Versicherungsgesellschaft mit Startkapital x erhält fortlaufend eingehende Prämien
und zahlt an zufälligen Zeitpunkten Schadenbeträge von zufälliger Höhe aus. Zusätzlich
steht es dem Versicherungsunternehmen frei Rückversicherung zu kaufen, das Geld zu
einem risikolosen Zinssatz anzulegen oder in Aktien zu investieren. Die Inhaber müssen
also Entscheidungen in Hinsicht auf die Unternehmenspolitik treffen. Deshalb braucht
man ein Maß für die mit einem Versicherungsportfolio verbundenen Risiken.

Typischerweise wird die Ruinwahrscheinlichkeit, d.h. die Wahrscheinlichkeit, dass der

Überschussprozess in endlicher Zeit negativ wird, als Maß für die Solvenz eines Ver-
sicherungsunternehmens gewählt. Allerdings wird das Konzept der Ruinwahrschein-
lichkeiten unter anderem dafür kritisiert, dass die Ruinstärke und der Geldzeitwert
vollkommen ignoriert werden.

Als alternatives Risikomaß betrachtet man den Wert der erwarteten diskontierten Kapi-
talzuführungen, welche notwendig sind damit der Überschussprozess nichtnegativ bleibt.
Es stellt sich die Frage, wie man diesen Wert minimiert. Wir nehmen an, dass die Inhaber
der Versicherungsgesellschaft das Geld heute dem Geld morgen bevorzugen. Dann ist
es optimal das Kapital nur dann zuzuführen, wenn der Überschussprozess negativ wird,
und nur so viel, dass der Überschussprozess wieder auf 0 verschoben wird.

Im ersten Teil der Arbeit lösen wir das Problem der Minimierung der erwarteten diskon-
tierten Kapitalzuführungen über alle dynamischen Rückversicherungsstrategien für das
klassische Modell der Risikotheorie und für eine Diffusionsapproximation. Im zweiten
Teil erweitern wir das vorherige Konzept durch die Möglichkeit der Anlage des nicht-
negativen Überschusses zu einem festen risikolosen Zinssatz. In diesen Fällen werden
wir die Existenz und Eindeutigkeit der optimalen Rückversicherungsstrategie und der
Wertefunktion als Minimum der erwarteten diskontierten Kapitalzuführungen zeigen
können.

Im dritten Teil betrachten wir den Überschussprozess, bei dem die Versicherungsge-
sellschaft in die Aktien investiert, deren Preis durch Black-Scholes Modell beschrieben
wird. Teil vier erweitert den Aufbau des dritten Teils durch die Möglichkeit der Rückver-
sicherung. In den beiden zuletzt genannten Fällen konnten wir eine explizite Lösung nur
im Fall der Diffusionsapproximationen finden. Im klassischen Modell wird das Konzept
der Viskositätslösungen, eingeführt von Crandall and Lions, verwendet.

All die Betrachtungen werden durch Simulationen, geschrieben in Java, illustriert.

Contents

Preface 1

List of Principal Notation 9

1 Preliminaries 11
1.1 The Classical Risk Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.2 Premia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.1.3 Reinsurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2 Diffusion Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3 On the notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2 Optimal Control of Capital Injections by Reinsurance 23

2.1 Diffusion Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.1 Reflected Diffusion Processes . . . . . . . . . . . . . . . . . . . . . 23
2.1.2 Capital Injections and Reinsurance . . . . . . . . . . . . . . . . . . 26
2.1.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 The Classical Risk Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.2.1 Properties of V (x) . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2.3 The special case δ = 0 . . . . . . . . . . . . . . . . . . . . . . . . . 54

3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of

Interest 63
3.1 Proportional Reinsurance for a Diffusion Approximation . . . . . . . . . . 63
3.2 The Classical Risk Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.2 The special case δ = 0 . . . . . . . . . . . . . . . . . . . . . . . . . 83

4 Optimal Control of Capital Injections by Investments 93

4.1 Diffusion Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.2 The Classical Risk Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

ix
 4.2.1 Constant strategies. . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.2.2 General case δ ≥ 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.2.3 The special case δ = 0. . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.2.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

5 Optimal Control of Capital Injections by Reinsurance and Investments 135

5.1 Diffusion Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.2 The Classical Risk Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.2.1 General case δ ≥ 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
5.2.2 The special case δ = 0. . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.2.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

A Stochastic Processes and Martingales 163

A.1 Stochastic processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
A.2 Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

B Markov Processes and Infinitesimal Generators 167

B.1 Markov processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
B.2 Generators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
B.2.1 The classical risk model . . . . . . . . . . . . . . . . . . . . . . . . 168
B.2.2 Diffusion approximation . . . . . . . . . . . . . . . . . . . . . . . . 169

C Change of Measure and Change of Variables Formulae 171

C.1 Ito’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
C.2 Change of Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

D Stochastic Differential Equations and Local Times 175

D.1 Stochastic Differential Equations . . . . . . . . . . . . . . . . . . . . . . . 175
D.2 Local Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176

E Differential equations 179

F The Black-Scholes Model 183

References 185

Index 191
 Shylock:
If you repay me not on such a day
In such a place, such sum or sums as are
Express’d in the condition, let the forfeit
Be nominated for an equal pound
Of your fair flesh, to be cut off and taken
In what part of your body pleaseth me.
···
Antonio:
Come on, in this there can be no dismay,
My ships come home a month before the day.
“Merchant of Venice”, Shakespeare

Preface

Insurance and Reinsurance: A Couple of Historical Facts

Googling the word “insurance” one finds as one of the first references the following
definition: “Insurance is a promise of compensation for specific potential future losses in
exchange for a periodic payment. Insurance is designed to protect the financial well-being
of an individual, company or other entity in the case of unexpected loss.”. An expected
and somehow familiar explanation. We insure our cars, our lives or our belongings;
every day we are overwhelmed with information about “credit for reinsurance”, “risk
transfer”, “enterprise risk management” and the like: insurance is an integral part of
modern life, we cannot do without. At the same time the simplest mutual insurances
in form of mutual aid societies are well known from the humanity’s ancient past. The
nomadic people of ancient east, engaged in cattle-breeding and caravan trade, by loss
of cattle refund the expenses to the concerned tribal members. Also in the Bible one
can find examples of insurance dissemination. In the story of Joseph and pharaoh, Gen.
41, 1-8, Joseph creates an insurance fund storing up great abundance of corn during the
seven years of plenty. During the following seven years of famine “over all the face of
the earth” there was enough corn in Egypt, so that “all countries came into Egypt to
Joseph to buy corn”.
In the Middle Ages the Florentine, Genovian and Venetian merchants, who had in XII–
XV centuries very active maritime trade relations with the countries of Near East, were
the first to organise the mutual maritime transport insurances. At first, in the early 14th

1
century, the “verbal agreement” was the rule in commerce. Ships were very expensive to
build and maintain, the cargo, they carried was an accumulation of financial risks and
the perils of weather and piracy were beyond of individual control of the merchants, who
often could not even accompany their vessels. What was needed was a contract which
would protect merchants against the caprices of fortune. In special guarantee bonds
the merchants, who had sold their goods before transporting them over sea, committed
themselves to buy the goods back at a higher price if the ships arrive safely. The
difference in the price was a pay for the risk, which got the name premium. The first
records of such bonds are believed to date from the beginning of the XIV century. Policies
(derived from the Italian word polizza, meaning a promise or undertaking) were given
to those insuring marine risks. Venice became an insurance center.

However, to the end of the XVI century the center of maritime commercial insurance
moved to England. In 1559 Sir Nicholas Bacon, advisor to Queen Elizabeth I of England
and a member of the parliament, said: “Doth not the wise merchant, in every adventure
of danger, give part to have the rest assured?”.

The “Merchant of Venice”, a play by William Shakespeare, is believed to have been

written between 1596 and 1598. I.e. insurance, as business, was well known not only in
the scene of action, Venice, but also in the homeland of the author. But either for the
exciting plot or through ignorance, he was not an actuary, Shakespeare did not give his
protagonist, the merchant of Venice Antonio, the possibility to “insure” his ships. The
“Merchant of Venice” would have been a very different story if some clever insurance
agent had persuaded Antonio to insure, and Shylock would have fit in perfectly as a
contract lawyer for disputed claims.

It is beyond all question, that if the first insurer incautiously undertakes large risks,
it is better to transfer those risks to the other underwriters. Insurance for insurance
companies is called reinsurance. The first reinsurance contract was drawn up in Genua
in 1370. Thus the reinsurance business originates also from Northern Italian seaside
trading towns, which created a rudiment of profit-oriented insurance. In this initial
stage of development the whole risk was often shifted from insurer to reinsurer, whereas
the reinsurer’s premium was smaller than the insurer’s premium. One could earn money
without any risk! As a consequence any marine policy “without further proof of interest
than the policy, or by way of gaming or wagering, or without benefit of salvage to the
assurer” was prohibited in England from 1746 to 1864. Naturally the development of
reinsurance in England was slowed by this legal prohibition of reinsurance. On April 8,
1846 the first independent professional reinsurance company, Cologne Reinsurance Co,
was founded. The great fire in Hamburg made the need of such a company evident.
Nowadays insurance and reinsurance help to make the modern life possible. Hurricanes,
earthquakes, terrorism and the good old pirats present risks beyond the great fires of
the past.

2
Measuring Risks

Now we make a huge jump not only in the time but also in the dimension of dis-
cussion of risks and profits in the (re)insurance business. We leave the historical and
macroeconomic field and switch to microeconomic considerations.

We consider a non-life insurance company ABC, having an initial capital x. As it

is common in insurance business ABC receives premia and pays claims. The insurer
is also allowed to invest either in a risky asset or in a riskless one and to use dynamic
reinsurance. Whenever the company makes a decision that affects the surplus. The
company is termed to be solvent as long as the initial capital together with the return
of investments and already received premia exceed the claims to be paid.

Examples from the previous section reveal, that modern society relies on the stability
and strength of the insurance system. An insurance premium paid currently provides
coverage for losses that might arise many years in the future. For that reason, the
viability of an insurance company is very important. In recent years, a number of
insurance companies have become insolvent, leaving their policyholders with no coverage.
To guarantee the smooth functioning of the insurance system a proper risk management
is needed. Managing the risk means managing an insurance company so as to maintain
a comfortable surplus of premia and/or risky assets beyond liabilities. Thus an actuary
is interested in a convenient measure for the risk as a basis for the decisions. The
problem with describing the risk characteristics of an insurance portfolio is, that in
non-life insurance the number of claims, claim sizes and occurrence times are random.
In 1903 the Swedish actuary Filip Lundberg [52] introduced a simple model which is
capable of describing the basic dynamics of a homogeneous insurance portfolio. Harald
Cramér with his books [15, 16] contributed a lot to the understanding and dissemination
of Lundberg’s collective model, which is now called the Cramér-Lundberg model or the
classical risk model. This model is nowdays also one of the most popular in non-life
insurance mathematics. Another very popular approach in modeling the surplus consists
in replacing the random part of the surplus process in the classical risk model with a
diffusion process. The basic idea is to make the claim sizes in the classical model small
and simultaneously to let the number of claims grow in such a way that the risk process
converges weakly to a diffusion. The idea of application of weak convergence in risk
theory was first introduced by Iglehart [43] in 1969, see also Grandell [36] and Schmidli
[70].

The next step after modeling the surplus as a stochastic process is to find an appro-
priate measure for risk. In order to evaluate risks of an insurance portfolio several risk
measures can be constructed. For example in Solvency II risk measures as the Value–at–
Risk and Expected Shortfall are used. The classical risk measure used by actuaries is the
ruin probability, i.e. the probability, that the portfolio becomes negative in finite time.

3
One says, that ruin occurs when the surplus process, modeled as a stochastic process,
becomes negative for the first time.
Stochastic control has been used in finance since the papers of Merton [53, 54]. An
introduction to control theory one finds in the books by Fleming and Soner [25], by
Fleming and Rishel [26] or by Karatzas and Shreve [49]. In recent years there has been
a rapid development in the field of applications of control theory to different aspects of
insurance mathematics. One of the popular optimisation criterions in insurance mathe-
matics is to control the process in such a way, that the probability of ruin is minimised.
Numerous papers and books have been written on this topic, among others one finds
Hipp and Plum [41], Hipp and Schmidli [42], Schmidli [68, 69, 70, 71]. The probability
of ruin serves as a kind of litmus paper. It indicates the soundness of the insurer’s com-
bination of the income of an insurance company plus the initial capital on the one hand
and claims process on the other. Also we obtain a useful tool for portfolio comparison.
But despite these positive points, using ruin probabilities also elicited criticism. The
main objection issues are

• the ruin probability does not actually represent the probability, that the insurer
will go bankrupt in the near future;

• it might take centuries until the ruin will actually happen;

• risk surplus processes usually tend to infinity, which is not the case for a real
surplus of an insurer;

• the time and the severity of ruin are ignored.

Another classical control problem of actuarial mathematics is to find strategies for

optimal payment of dividends. In 1957 de Finetti [24] considered the optimal control
problem of finding the dividend payment strategy, that maximises the expected dis-
counted value of dividends, which are paid to the shareholders until the company is
ruined. He was the first to propose to measure the risk as the maximal value of expected
discounted dividends. Karl Borch [12] contributed a lot to the dissemination and devel-
opment of de Finetti’s idea. In the framework of the classical compound Poisson model
the problem was considered among others by Bühlmann [14], Gerber [28, 30], Gerber
and Shiu [33], Gerber et al. [34, 35], Dickson and Waters [18], Dickson and Drekic
[19]. In many later papers the problem was formulated and solved also for a diffusion
approximation, see Shreve at al. [72], Jeanblanc-Picqué and Shiryaev [46], Paulsen and
Gjessing [57], Højgaard and Taksar [38], Asmussen and Taksar [3], Asmussen et al. [4],
Paulsen [58].
Typically, by maximising the dividend income, the optimal dividend strategy would
be a band strategy leading to certain ruin; and if the ruin probability is minimised,

4
then obviously no dividend will be paid. Christian Hipp [40] considered the maximal
dividend income under a ruin constraint. But in real life a technical ruin does not
compulsory mean the bankruptcy. Of course, if the liabilities exceed the company value,
it is better to wind up the business. Otherwise the company holders have to raise new
money to proceed with the business. Therefore, capital injections could be added to
the surplus process, and the value of the discounted cash flow can be considered, see
Kulenko and Schmidli [50]. However if injecting additional money is not penalised, the
optimal strategy would be to pay the income as dividends and to finance the outflow by
capital injections. In other words the optimal strategy would keep the surplus at zero.

The idea of this work is to choose as a risk measure the minimal expected discounted
capital injections. Let X be the underlying surplus process with X0 = x. Let Y be
an increasing process with Y0− = 0. The process R ∞with capital injections is denoted by
XtY = Xt +Yt . We define the value V Y (x) = Ex [ 0 e−δt dYt ], where δ ≥ 0. The injection
process Y has to be chosen such that XtY ≥ 0 for all t (almost surely). The value function
is defined as V (x) = inf V Y (x), where the infimum is taken over all processes Y such
that XtY ≥ 0 for all t.
It should be noted that δ is not a financial discounting factor. We assume that the
process X is already discounted. Choosing δ > 0 then means that the investor prefers
investing tomorrow to investing today. If we choose δ = 0 the investor is indifferent
to investing tomorrow or today. We will see below that for a diffusion model δ = 0
corresponds to minimising the probability of ruin. Choosing δ < 0 would complicate the
problem. It could then be optimal to inject capital already before the process reaches
zero. If δ is too small, the value function will become infinite.

Measuring the risk as proposed above has following advantages. As the value of future
capital injections, the measure is economically motivated. It is clear, that it is not
optimal to inject capital before it is really necessary which means the sub-additivity of
the measure. For two underlying processes X and Z with corresponding value functions
VX (x) and VZ (x) we have for the value function of the sum process X + Z: VX+Z ≤
VX + VZ . Indeed,
R ∞ −δtit follows Rreadily that YtX+Z ≤ YtX + YtZ , so the sub-additivity
∞ −δt
follows from 0 e dYt = δ 0 e Yt dt. Compared with the valuation of dividends,
the advantage is, that we do not have to solve an optimisation problem in order to find
the value.

The problem we will be dealing with in this work is, how the insurer should control
the future capital injections by means of reinsurance and/or investments in the Cramér-
Lundberg model and in a diffusion approximation if δ ≥ 0.
In the first chapter we just give the basic definitions and a short description of the
models we will use later. The following four chapters introduce four different settings for
the classical risk model and its diffusion approximation. Every chapter opens with the
simplest case of a diffusion approximation. Then the classical risk model is considered

5
for the general case δ ≥ 0. The special case δ = 0, where we can always show the
existence and uniqueness of the solution, is considered in a final step. All considerations
are illustrated with examples.
In Chapter 2 we consider the models, in which the only available control is related
to dynamic reinsurance. We solve the problem explicitly in the case of a diffusion
approximation and give a closed expression for the value function. The optimal strategy
in this case is constant and the value function is an exponential function. In the classical
risk model we show the existence and the uniqueness of the value function and calculate
the value function and the optimal strategy numerically in the special case of proportional
reinsurance with exponentially and Pareto distributed claims. In both models we solve
the optimality problem with help of the Hamilton–Jacobi–Bellman (HJB) equation. In
the case of a diffusion approximation we will go through the following steps. At first
we give the HJB equation corresponding to the problem we consider and show that it
has a unique solution satisfying the boundary conditions derived from the optimisation
problem. Then using the verification theorem we show, that the solution obtained to
the HJB equation is the value function of the optimisation problem. In the classical risk
model we use another technique. We show directly that the value function solves the
HJB equation and that the solution is unique.
In Chapter 3 we change the setup of the second chapter by implementing interest at
a constant rate, i.e. in addition to possibility of reinsurance the insurer is allowed to
earn on the surplus, provided it remains positive. In this chapter we consider in the
case of a diffusion approximation proportional reinsurance only. Although the change
in the model was small, the optimal strategy and the value function change a lot. The
optimal strategy is not constant any more and the value function is composed of four
different functions. As for the classical risk model, the proof techniques are similar to
the techniques used in Chapter 2.
In Chapter 4 the insurer is allowed to invest in a risky asset which follows a geometric
Brownian motion. While in the diffusion approximation case we show again, that the
optimal strategy is constant and the value function is an exponential function, the case
of classical risk model becomes much more complicated as in Chapters 2 and 3. For
the general case δ ≥ 0 we can not show the existence of a classical solution to the
corresponding HJB equation and have to use the concept of viscosity solutions. For
introduction to viscosity solutions see for example Bardi and Capuzzo-Dolcetta [6] or
Crandall et al. [17]. The proof techniques used in the case δ ≥ 0, one finds in Benth et
al. [8], Azcue and Muler [5] and Albrecher and Thonhauser [1].
In the special case δ = 0 one can show the existence and uniqueness of the classical
solution to the HJB equation using the proof techniques from Schmidli [70]. In this case
we can also apply the result of Hipp and Plum [41], concerning the asymptotic behavior
of the optimal strategy.
In Chapter 5 we consider a combination of both investment and reinsurance. A risky
asset is also here modeled as a geometric Brownian motion. Diffusion approximation

6
yields a constant optimal strategy and an exponential function as value function. In the
case of the classical risk model we proceed as in Chapter 4. For simplicity we consider
in the special case δ = 0 only the proportional reinsurance. But also here we can show
the existence and uniqueness of the value function.
In the derivation and proofs of our results we have used a lot of stochastic analysis and
techniques from the martingale theory. Because we assume, that the reader of this work
is perfectly acquainted with the terms and definitions of stochastic calculus we just state
the theorems we have used in the Appendix. For detailed introduction to martingale
theory and stochastic analysis we refer to Protter [60], Bhattacharya and Waymire [9],
Karatzas and Shreve [49] or Ethier and Kurtz [23].

7
8
List of Principal Notation

N The natural numbers, N = {0, 1, 2, ..}

R The real numbers
R+ The non-negative real numbers, R+ = [0, ∞)
p∧q min{p, q} for p, q ∈ R
p∨q max{p, q} for p, q ∈ R
p+ max{p, 0} for p ∈ R
p− min{p, 0} for p ∈ R
Ē The closure of the set E ⊂ R
∂E The boundary of the set E ⊂ R
C1 Space of continuously differentiable functions

P The basic probability measure

Ω Event space on which probabilities are defined
F σ-Algebra of the probability space
F = {Ft } Filtration
E Expected value with respect to P
Ex [X] Expected value given initial value x
Var Variance with respect to P

{Zi }i∈N iid random variables, representing claim sizes

Z Generic random variable with the same distribution as Zi
MZ (y) Moment-generating function of the random variable Z
N = {Nt } Poisson process
λ Intensity of N
Ti i-th occurrence time
G Claim size distribution function
µ, µn The first and the n-th moment of the claim sizes
c Premium rate
η Safety coefficient of the first insurer
b Retention level b ∈ [0, b̃]
c(b) Premium rate function, depending on retention level b

9
 r(Z, b) Self-insurance function, depending on claim size Z
and retention level b
θ Safety coefficient of the reinsurer
δ Discounting factor

1I[.] Indicator function

X = {Xt } Stochastic process
W = {Wt } Standard Brownian motion
A Set of admissible investment strategies
U Set of admissible reinsurance strategies
B = {bt } Reinsurance strategy
B ∗ = {b∗t } The optimal reinsurance strategy
A = {at } Investment strategy
A∗ = {a∗t } The optimal investment strategy
Y = {Yt } Capital injection process
XtB , XtA , XtA,B Stochastic process under the strategies B, A,
or (A, B)
XtB,Y , XtA,Y , XtA,B,Y XtB , XtA , XtA,B with capital injections
V (x) Value function
V B (x), V A (x), V A,B (x) Return functions of the strategies B, A, (A, B)

10
1 Preliminaries
We begin with an overview of the models, which provide the motivation for our in-
vestigations. The goal of this chapter is just to give a rough review of the concepts we
use in the work. Therefore we will abandon the proofs of the theorems stated below and
just refer to the books and papers, where a detailed description can be found. At first
we consider the Cramér-Lundberg model, state its most important concepts and give
a short review of the premium calculation principles and reinsurance treaties. In the
second section we derive a diffusion approximation, which will be used throughout this
work. In the third section we make some agreements about terminology and notation.

In the following considerations we assume as given a complete probability space

(Ω, F, P), which is large enough to carry all the objects defined below. In addition
we are given a filtration {Ft }. By a filtration we mean a family of σ-algebras, that is
increasing, i.e. Fs ⊂ Ft for t ≥ s. For convenience, we will often write F for the filtration
{Ft }.

1.1 The Classical Risk Model

1.1.1 Introduction
The Cramér–Lundberg model, also called the classical risk model, is a risk process (or
surplus process) X defined as
Xt = x + ct − St , (1.1)

where
Nt
X
St = Zi (1.2)
i=1

is the corresponding loss process. The process {Nt } represents the number of claims
in the time interval [0, t], Zi for i ∈ N are the claim sizes, positive, independent and
identically distributed random variables with finite mean µ and finite second moment
µ2 . Claims occur at random instants of time 0 < T1 < T2 ... and interarrival times
T2 − T1 , T3 − T2 , T4 − T3 , ... are independent, exponentially distributed random variables
with finite mean λ−1 > 0. The processes {Zi } and {Ti+1 − Ti } are independent of each
other. c denotes the premium income in a time interval of length 1.

11
1 Preliminaries

The time of ruin

x




0 T1 T2 T3 T4 = τ

Figure 1.1: The Process {Xt }.

Given that the interarrival times are exponentially distributed, Nt is a homogeneous

Poisson process with intensity λ > 0, i.e.
(λt)k
P[Nt = k] = e−λt , k = 0, 1, 2, ...
k!
Typically, the surplus process has the form illustrated in Figure 1.1. X grows linearly
and at the random moments of claim arrivals drops by the amounts of the claims.

The concept of ruin is of central importance by modeling the surplus process. Figure
1.1 gives already the first intuitive idea of ruin. Below we give the precise definition.
Definition 1.1.1 (Ruin)
The event, that the process Xt , defined in (1.1), ever falls below zero is called ruin. The
time τ , when the process Xt falls below zero for the first time
τ = inf{t > 0 : Xt > 0}
is called ruin time. As severity of ruin one denotes the absolute value |Xτ |.
The probability of ruin for the initial capital x is then given by
ψ(x) = P[τ < ∞|X0 = x] .

12
 1.1 The Classical Risk Model

Lots of papers on the probability of ruin, on the severity of ruin and on surplus prior
to ruin in the different settings for the classical risk model have been written since the
pathbreaking paper of Gerber and Shiu [31]. A topic no less important, than the concept
of ruin, is the asymptotic behavior of the surplus process. A helpful tool in this regard
is the following proposition.

Proposition 1.1.2
The aggregate loss process {St }, defined in (1.2), and the homogeneous Poisson process
{Nt } satisfy the strong law of large numbers:

St St Nt t→∞
= · −→ λµ ,
t Nt t

Proof: Confer Mikosch [55, p. 82] 

This implies, lim Xt = ∞ or lim inf Xt = −∞ a.s. according to whether c > λµ or

t→∞ t→∞
c ≤ λµ. A result from the theory of random walks, see Rolski et al. [61], shows that in
the case c ≤ λµ the ruin occurs with probability one.

Definition 1.1.3 (Net profit condition)

One says, that the net profit condition is fulfilled if

c − λµ > 0 .

The constant
c − λµ
η=
λµ
is called the safety loading or safety coefficient of the insurer.

Thus, the net profit condition requires choosing the premium intensity larger than the
expected loss in a time interval of length 1.
In connection with the ruin probability one often speaks of adjustment coefficient.

Definition 1.1.4 (Adjustment or Lundberg’s coefficient)

Assume, that the moment generating function of Z, MZ (s) = E[esZ ], exists in some
neighborhood (−h0 , h0 ), h0 > 0. If the unique positive solution ρ to the equation

MZ (s) = E[esZ ] = 1

exists, it is called the adjustment or Lundberg coefficient.

One of the classical results in insurance mathematics is the following proposition.

13
1 Preliminaries

Proposition 1.1.5 (Lundberg’s inequality)

Assume, that the adjustment coefficient ρ exists. Then the following inequality holds
for all x > 0:
ψ(x) = P[τ < ∞|X0 = x] ≤ e−ρx .
Proof: Confer Rolski et al. [61, p. 416]. 

The exponential boundary of Lundberg’s inequality ensures that the probability of ruin
is very small if one starts with large initial capital x. The result tells us, that the smaller
ρ the more risky is the portfolio. For a detailed discussion of Lundberg’s coefficient we
refer to Rolski et al. [61].
The following operator which is applied to a suitable function f , see Proposition B.2.4
p. 168, is called the infinitesimal generator of the Markov process X,
Z ∞
′
Df (x) = cf (x) + λ f (x − z) dG(z) − λf (z) ,
0

which will be needed later on.

1.1.2 Premia
Most frequently the insurance premium is described as the “price” in exchange for
which the insurer takes over and bears the insured’s risk. Below we give three most
famous premium calculation principles. We refer to Heilmann [39], to Rolski et al. [61],
to Kaas [47] and to Rotar [65] for an explicit description of various premium principles
and their properties.
Let Z denote the random variable, which describes the risk of some insurance contract;
with c we will denote the premium. We assume, that Z is bounded and non-negative.
1. Expected value principle
The premium is
c = (1 + η)E[Z]
for some safety loading of the insurer η > 0.

2. Variance principle:
The premium is
c = E[Z] + αVar[Z]
for some α > 0.

3. Standard deviation principle: The premium is

p
c = E[Z] + α Var[Z]

for some α > 0.

14
 1.1 The Classical Risk Model

1.1.3 Reinsurance
In practice an insurer often transfers portions of risk portfolios to other parties in order
to reduce large obligations resulting from an insurance claim. An insurance company that
transfers a risk to another insurance company is called a cedent or a ceding company.
An insurance company that assumes all or part of an insurance policy written by a
ceding company is called a reinsurer. A reinsurance treaty is an agreement between
an insurer (cedent) and a reinsurer under which, claims that occur in some fixed period
of time are split between the insurer and reinsurer. There are three different types of
reinsurance treaties: reinsurance acting on individual claims, reinsurance acting on the
aggregate claim over a certain period, and reinsurance acting on the k largest claims
occurring during a certain period. For the sake of simplicity of notation we will restrict
our considerations only on the first type reinsurance treaties.
Below we discuss some functions of random variables which describe the portion of
risk, which the insurer will carry himself. Let Z again describe the loss amount and let r
be the function describing that part of the claim, that the insurer intends to pay himself.
Because r will depend not only on Z but also on reinsurance treaty-specific variables
we write r(Z, .). The function r(Z, .) is called self-insurance function. For a general
reinsurance treaty it holds 0 ≤ r(Z, .) ≤ Z.

The most famous reinsurance types acting on individual claims are

• r(Z, b) = b · Z, proportional reinsurance with retention level b;
• r(Z, b) = min(Z, b), excess of loss (XL) reinsurance with retention level b.
• r(Z, (a, b, γ)) = min(Z, a) + (Z − a − γ)+ + b min (Z − a)+ , γ , Proportional

reinsurance in a layer for a, γ > 0 and b ∈ (0, 1).

The proportional reinsurance in a layer has a multidimensional retention level (a, b, γ).
For the sake of simplicity we will consider in this work only one-dimensional reinsurance
treaties with some retention level b. The case of multidimensional treaties can be treated
similarly, if we let the retention level b be a vector.

The problem by buying reinsurance consists in determining for a given risk the portion,
which the insurer should carry himself. The decision is of course influenced by the
price, which reinsurer demands for taking over the ceded portion. Using the example of
proportional reinsurance, i.e. r(Z, b) = bZ, we calculate the premium remaining to the
insurer for the three premium calculation principles defined above. Let c(b) denote the
premium remaining to the insurer, if the retention level b was chosen.
• Expected value principle. Assume the safety loading of the insurer is given by η
and the safety loading of the reinsurer by θ.
c(b) = E[Z](1 + η) − E[Z − bZ](1 + θ) = E[Z](η − θ) + E[Z](1 + θ)b . (1.3)

15
1 Preliminaries

• Variance principle. Assume α is the premium parameter of the insurer and β of

the reinsurer.
c(b) = bE[Z] + (α − β(1 − b)2 )Var[Z] .

• Standard deviation principle. Assume α is the premium parameter of the insurer

and β of the reinsurer.

c(b) = bE[Z] + (α − β(1 − b))Var[Z] .

For our numerical calculations we will have to choose the distribution function of the
claim sizes. Exponential distribution Exp(γ) for some parameter γ > 0 with density
(
1 −x/γ
e : x≥0,
f (x) = γ
0 : x<0,

turns out to be very convenient for the numerical derivations. In insurance mathematics
the exponential distribution serves as a benchmark for characterisation of the heaviness of
the tails. As a tail of a distribution function F one denotes the function F̄ (x) = 1−F (x).
One says F is a light-tailed distribution if the tail function F̄ fulfils

F̄ (x)
lim sup <∞
x→∞ e−x/γ
for some γ > 0. If
F̄ (x)
lim sup =∞
x→∞ e−x/γ
for all γ > 0 we call F heavy-tailed.
Thus, Exp(γ) distribution is a light-tailed distribution and the large claims have com-
paratively small weight. Experience has shown, that the Pareto(2, µ) distribution, i.e.
the distribution with density
2µ2
f (x) = ,
(x + µ)3
is often appropriate for representing the tail of distributions, where large claims may
occur. More about heavy- and light-tailed distributions one finds for example in Mikosch
[55]. In our numerical examples we will consider both, the exponential and Pareto
distributions.

1.2 Diffusion Approximation

Despite the seeming simplicity of the classical risk model, it is sometimes hard to
obtain accurate numerical calculations. A useful method to derive an approximation is

16
 1.2 Diffusion Approximation

the “diffusion approximation”. The idea is to substitute the random part of the classical
risk model by a diffusion.
First of all we give a definition of a diffusion process.
Definition 1.2.1
A Markov process X = {Xt } is said to be a diffusion with infinitesimal drift function
m(x) and infinitesimal variance σ 2 (x) > 0, if

E[(Xs+t − Xs )1I[|Xt −x|≤ε] |Xs = x] = tm(x) + o(t) ,

E[(Xs+t − Xs )2 1I[|Xt −x|≤ε]|Xs = x] = tσ 2 (x) + o(t) ,
E[|Xs+t − Xs | > ε|Xs = x] = o(t)
o(t)
as t ↓ 0 for every x ∈ R and every ε > 0, where lim t = 0.
t→0
We find in Rolski et al. [61, p. 561], that a diffusion X is the unique solution to the
stochastic differential equation

dXt = m(Xt ) dt + σ(Xt ) dWt , (1.4)

which is then interpreted as stochastic integral equation

Z t Z t
Xt = X0 + m(Xs ) ds + σ(Xs ) dWs .
0 0

We say, that {Xt } is a (m(x), σ 2 (x))-diffusion.

Remark 1.2.2
The most famous examples for diffusion processes are Brownian motion and geometric
Brownian motion, see for instance Borodin and Salminen [13, p. 51, 77]. Throughout
this work we will denote the standard Brownian motion by W = {Wt } and a geometric
2
Brownian motion by Q = {Qt } = {Q0 e(m−σ /2)t+σWt } with some m, σ.
Next we give some sufficient conditions, which imply the existence of a unique solution
to stochastic differential equation (1.4).

Theorem 1.2.3
Suppose, that the coefficients m(x) and σ(x) satisfy the global Lipschitz and linear
growth conditions

|m(x) − m(y)| + |σ(x) − σ(y)| ≤ K|x − y| ,

|m(x)|2 + |σ(x)|2 ≤ K 2 (1 + x2 ) ,

for all x, y ∈ R, where K is a positive constant. Let further {Ft } be a filtration generated
by the standard Brownian motion W . Then there exists a continuous, adapted process
X = {Xt }, which is a strong unique solution to (1.4).

17
1 Preliminaries

Proof: Confer Karatzas and Shreve [49, p. 287, 289]. 

In the following chapters we will often use the concept of infinitesimal generator of a
diffusion in our proofs and derivations. Assume X is a diffusion, then its infinitesimal
generator is given by
σ 2 (x) ′′
Df (x) = m(x)f ′ (x) + f (x)
2
for some suitable function f . For proof see Proposition B.2.5 p. 170 in the Appendix.
Because the aim of the present section is to establish diffusion approximations, we will
show next how a diffusion approximation to the classical risk model can be constructed.
In order to do that we need the concept of weak convergence.

Definition 1.2.4 (Weak Convergence)

A sequence (X (n) ) of stochastic processes is said to converge weakly to a stochastic
process X if for every bounded continuous functional f it follows, that

lim E[f (X (n) )] = E[f (X)] .

n→∞

Then we write X (n) ⇒ X.

Definition 1.2.5
We say that a risk process X can be approximated by K if there exists a sequence of
(1)
stochastic processes (X (n) )n≥1 , such that X (n) ⇒ K and Xt = Xt .
(n)
If for a classical risk process {Xt } there exists a sequence {Ct } of classical risk processes
(1)
such that Ct = Xt and C (n) ⇒ W̃ = {mt + σW − t} for some m and σ 2 , then we call
W̃ a diffusion approximation to X.
The following basic construction is due to Schmidli. Denote the sequence of correspond-
(n) (n)
ing Poisson processes by N (n) = {Nt }, their intensities by λn , the claims by Zi with
distribution function Gn , the initial capitals by xn , the premium rates by cn and the loss
(n)
processes by {St }, i.e.
(n) (n)
Ct = xn + cn t − St .
Let further Z ∞ Z ∞
(n) (n)
µ := z dGn (z) and µ2 := z 2 dGn (z) ,
0 0
(n) cn −λn µ(n)
where we assume µ2 < ∞. We write for the safety loadings ηn := λn µ(n)
.
Note, that all the processes C (n) , being classical risk models, have independent and sta-
tionary increments. Let {Wt } be a standard Brownian motion and consider an (m, σ 2 )-
Brownian motion W̃ = {W̃t }, i.e. W̃t = W̃0 + mt + σWt .

18
 1.2 Diffusion Approximation

Theorem 1.2.6
In order to have C (n) ⇒ W̃ the parameters have to satisfy

lim λn = ∞ ;
n→∞
lim cn = ∞ ;
n→∞
lim ηn λn µn = m ;
n→∞
(n)

lim λn (µ(n) )2 + (µ2 )2 = σ 2 ;
n→∞
(n)
lim µ(n) = 0 and lim µ2 =0.
n→∞ n→∞

Proof: For proof see Schmidli [66, p. 74]. 

(n)
A very simple diffusion approximation can be obtained if we choose λn = nλ, Zi =
√ √ (n) √ √
Zi / n, ηn = η/ n and xn = x. Note, that Zi = Zi / n implies µ(n) = µ/ n and
√
Gn (z) = G( nz). Thus, it holds
(n)
NP
t
(n)
Nt λµnt − Zi
(n) √ √ X √ i=1
Ct = x + λµ n(1 + η/ n)t − Zi / n = x + λµηt + √ .
n
i=1

Let W = {Wt } be the standard Brownian motion again and µ2 the second moment of
claim sizes. With standard propositions from the theory of convergence of probability
measures, see for example Billingsley [10], we obtain as a diffusion approximation
p
x + λµηt + λµ2 Wt .

We have constructed a sequence of risk processes that converges in probability to the

risk process, whose random part is the well-known diffusion. The key assumption in our
construction was that both the interarrival times as well as the claim severities are in the
domain of attraction of stable distributions. This implies, that a diffusion approximation
is good in the case, when an individual claim is very small compared to the size of the
total reserve.

Assume now, that the insurer models his surplus as the classical risk model and
buys proportional reinsurance with retention level b ∈ [0, 1]. We assume also, that the
retention level can be chosen only at the beginning, i.e. b remains constant over time.
Then the surplus process C with initial capital x has the form
Nt
X
Ct = x + c(b)t − b Zi ,
i=1

19
1 Preliminaries

where c(b) denotes the premium remaining to the insurer. Using the expected value
principle, see (1.3), yields
Nt
X
Ct = x + λµ(η − θ + bθ + b)t − b Zi ,
i=1

where η is the safety loading of the insurer and θ the safety loading of the reinsurer.
Constructing a sequence of classical risk models in the way described above, we obtain
as a diffusion approximation
p
Xt = x + λµ(η − θ + bθ)t + b λµ2 Wt ,

where W is again the standard Brownian motion. If we assume, that the insurer can
change his retention level continuously, i.e. we consider a process B = {bt }, then we
obtain a diffusion approximation by means of the following proposition.
Proposition 1.2.7
Let m : R → R be a Lipschitz continuous function, (X (n) )n≥1 be a sequence of semi-
(n)
martingales, such that X0 = 0, and Z (n) be defined by
Z t
(n) (n)
Zt = x + Xt + m(Zs(n) ) ds .
0

Let further X be a semimartingale with X0 = 0 and Z be defined by

Z t
Zt = x + Xt + m(Zt ) ds .
0

Then Z (n) converges weakly to Z if and only if X (n) converges weakly to X.

Proof: For proof see Schmidli [67]. 

A diffusion approximation has then the form

Z t Z t
XtB = x + λµ (η − θ + bs θ) ds + λµ2
p
bs dWs .
0 0

Remark 1.2.8
In a diffusion approximation the event, that the diffusion approximation process Xt ever
falls below zero, is called ruin. The time τ , when the process Xt falls below zero for the
first time
τ = inf{t > 0 : Xt > 0}
is called ruin time. The probability of ruin for the initial capital x is

ψ(x) = P[τ < ∞|X0 = x] .

20
 1.3 On the notation

1.3 On the notation

The following conventions will be applied.

1. In the following we will always use a probability space (Ω, F, P) on which all
stochastic quantities are defined. We always assume that (Ω, F, P) is complete,
i.e. F contains all P-null sets. At the beginning of each section we give a short
description of the model we are going to consider. In each model we are given a
filtration F = {Ft }. In the following
T we always assume, that a filtration F = {Ft }
is right continuous, i.e. Ft = Fs for all 0 ≤ t < ∞. We do not assume that the
s>t
given filtration is complete because otherwise we could have problems with change
of measure techniques. To simplify the notation we will often omit the expressions
almost surely (a.s.) or with probability one.

2. In order to emphasise that the considered process has initial value x we write for
the corresponding expected value operator Ex [.].

3. For the convenience of reading a list of principal notation is given on page 9.

4. An introduction to the optimal control theory goes beyond the scope of this work.
Considering the models we simultaneously give an explanation of the optimal con-
trol concept concerning each special optimisation problem. To avoid misunder-
standings and ambiguity of the notation, we give here, without going into details,
some definitions we will use throughout the work.

The concept of control can be described as the process of influencing the behavior
of a dynamical system to achieve a desired goal. Consider some stochastic process
X = {Xt }. Let further U = {Ut } be a variable which models the decision process.
We will say that U is a feedback control if Ut = L(Xt ) for t ≥ 0, for a measurable
L : R → R.
We call a function, which is associated to some control strategy return function.
As a value function we will denote a function given as an infimum over all return
functions.

21
1 Preliminaries

22
2 Optimal Control of Capital Injections by
Reinsurance

2.1 Diffusion Approximation

Let (Ω, F, P) be a complete probability space, that is large enough to carry all the
stochastic objects defined in this section. By F = {Ft : t ≥ 0} we denote the natural
filtration of the standard Brownian motion W . The section is organised as follows. In
Subsection 2.1.1 we review some results on reflected diffusion processes and calculate
the value of an “uncontrolled” surplus process. In Subsection 2.1.2 we introduce the
diffusion approximation and add the possibility of reinsurance. We solve the problem
of minimising of the expected discounted capital injections. In order to illustrate the
results, we give in Subsection 2.1.3 some examples, where we can calculate the value
function explicitly.

2.1.1 Reflected Diffusion Processes

In this subsection we consider reflected one-dimensional diffusion processes. These
considerations will be used in the analysis of the control problems through out the
whole work. Consider the diffusion process X with the state space R and dynamics

dXt = m(Xt ) dt + σ(Xt ) dWt , (2.1)

where the functions m and σ are chosen in such a way that the above stochastic differ-
ential equation has a unique strong solution. This is for example the case if m and σ
are chosen like in Theorem 1.2.3.
We are interested in the minimal value of the expected discounted capital injections,
which are necessary to keep the process X above zero. The data for our problem are
the drift rate m, the volatility σ and the discounting rate δ ≥ 0. As it has been already
explained in Preface, the discounting factor δ ≥ 0 expresses the investment preferences
of the company holders. δ ≥ 0 implies that investing tomorrow is preferred to investing
today. It is clear that in this setup for initial value X0 < 0 it will be optimal just to
inject −X0 . Thus, our attention will be restricted to x ≥ 0.
It is clear that if X is at zero, we must apply control to stop the process entering
(−∞, 0).

23
2 Optimal Control of Capital Injections by Reinsurance

We interpret Yt as the cumulative capital injections up to time t and associate with

Y = {Yt } the controlled process
Z t Z t
Y
Xt = x + m(Xs ) ds + σ(Xs ) dWs + Yt .
0 0

Because the discounting factor δ is non-negative, we should inject capital only when the
process becomes negative and only as much, that the process is again shifted to zero. It
means Y changes only on the set [X Y = 0] and because x ≥ 0 it holds Y0 = 0. Thus,
the above equation is an equation of Skorokhod type. In Appendix D we find, that the
unique solution Y is the local time of X at zero and X Y is a reflected diffusion, which
follows the stochastic differential equation

dXtY = m(XtY ) dt + σ(XtY ) dWt + dYt .

After we have found the optimal strategy, we are interested in calculating the value
function V (x). It is clear that V (x) is a decreasing function. Because of the discounting
we also have that lim V (x) = 0. Because the capital injections start at time zero it
x→∞
holds V (x) = −x + V (0) for x < R0. We therefore can restrict to positive initial capital.
t
Consider now the martingale { 0 e−δs dYs + e−δt V (XtY )}
hZ ∞ i Z u
−δs
Ex e dYs Fu = e−δs dYs + e−δu V (XuY ) .
0 0

Suppose that V (x) is twice continuously differentiable. By Itô’s formula, see Proposition
C.1.1 p. 171,
Z t  σ 2 (X Y ) 
−δt Y s
e V (Xt ) = e−δs V ′′ (XsY ) + m(XsY )V ′ (XsY ) − δV (XsY ) ds
0 2
Z t Z t
−δs ′ Y Y
+ e V (Xs )σ(Xs ) dWs + e−δs V ′ (XsY ) dYs + V (x) .
0 0
Rt
Rearranging the terms and adding 0 e−δs dYs on both sides of the equation yields
Z t Z t
e−δt V (XtY ) + e−δs dYs − e−δs V ′ (XsY )σ(XsY ) dWs
0 0
Z t  σ 2 (X Y ) 
s
= e−δs V ′′ (XsY ) + m(XsY )V ′ (XsY ) − δV (XsY ) ds
0 2
Z t
e−δs 1 + V ′ (XsY ) dYs


+
0

Because we assume that X is the unique strong solution to Equation (2.1), see Chapter
D, and V ′ (XtY ) is locally bounded because of the continuity, we obtain that

24
 2.1 Diffusion Approximation

Rt
0e−δs V ′ (XsY )σ(XsY ) dWs is a local martingale. By means of optional stopping time
theorem we obtain that
nZ t
e−δs 1 + V ′ (XsY ) dYs


0
Z t  σ 2 (X Y )  o
s
+ e−δs V ′′ (XsY ) + m(XsY )V ′ (XsY ) − δV (XsY ) ds
0 2
is a local martingale. Note that since Y = {Ys } is an increasing process, it is of bounded
variation. Thus the above local martingale is of bounded variation and therefore con-
stant, say equal to C, see Proposition A.2.5 p. 166. Because for t = 0 the above local
martingale is zero, it follows C = 0. On the other hand {Ys } only increases on the set
{XtY = 0}. Define the stopping time τ = inf{t ≥ 0 : Xt < 0}. Then we obtain
h Z τ ∧t  σ 2 (X Y )  oi
s
E e−δs V ′′ (XsY ) + m(XsY )V ′ (XsY ) − δV (XsY ) ds = 0 .
0 2
Dividing by t and letting t go to zero we get by bounded convergence the equation
σ 2 (x) ′′
V (x) + m(x)V ′ (x) − δV (x) = 0 (2.2)
2
for x > 0. By continuity the equation also holds for x = 0, implying V ′ (0) = −1. In
particular, V (x) is at 0 continuously differentiable from the right.
Suppose lim f (x) = 0, f (x) is twice continuously differentiable and solves (2.2). Let
x→∞
Tn = inf{t : XtY > n}. Because as a continuous function, f ′ (x) is bounded on [0, n] the
R T ∧t
process { 0 n f ′ (XsY )σ(XsY ) dWs } is a martingale, see Proposition A.2.6 p. 166. Note
Rt Rt
further, that 0 e−δs f ′ (XsY ) dYs = f ′ (0) 0 e−δs dYs , because Yt only increases if XtY = 0.
Thus, Z Tn ∧t
n o
−δ(Tn ∧t) Y ′
e f (XTn ∧t ) − f (0) e−δs dYs
0
is a martingale by (2.2). In particular,
h Z Tn ∧t i
−δ(Tn ∧t)
f (x) = Ex e f (XTYn ∧t ) ′
− f (0) e−δs dYs .
0

By monotone and bounded convergence we can let n and t tend to infinity, giving
hZ ∞ i
f (x) = −f ′ (0)Ex e−δt dYt ,
0

This was first shown by Shreve et al. [72].

Shreve et al. [72] also found an explicit solution for a Brownian motion with drift. We
review this example, because we will use this result in the following.

25
2 Optimal Control of Capital Injections by Reinsurance

Example 2.1.1
Consider the Brownian motion with drift

Xt = x + mt + σWt .

Equation (2.2) reads

σ 2 ′′
V (x) + mV ′ (x) − δV (x) = 0 . (2.3)
2
The solution is of the form V (x) = c1 er1 x + c2 er2 x , where r1 > 0 > r2 are the roots of

σ 2 r 2 + 2mr − 2δ = 0 ,

see Proposition E.1.4. We have to select c1 , c2 such that

V (∞) = 0, V ′ (0) = −1 .
2
That is, c1 = 0 and c2 = √σ . Thus, the value function has in this case the
m+ m2 +2δσ2
form √
2 m2 +2δσ2
exp − m+
(
√σ


m+ m +2δσ2
2 σ2
x : x≥0,
V (x) = 2
√σ −x : x<0.
m+ m2 +2δσ2
Because this function is twice continuously differentiable on (0, ∞), it is the solution to
our problem by the above general considerations. 

2.1.2 Capital Injections and Reinsurance

We consider a surplus process where the mean number of claims in a time unit is λ and
the mean size of a claim is µ = E[Z], where Z is a generic claim size. We assume that
µ2 = E[Z 2 ] < ∞. The premium is c = (1 + η)λµ for some η > 0. The insurer can now
buy reinsurance. That is, a retention level b ∈ [0, b̃] has to be chosen, where b̃ ∈ (0, ∞].
The cedent pays then 0 ≤ r(Z, b) ≤ Z. The
reinsurance premium is calculated by an
expected value principle (1 + θ)λ µ − E[r(Z, b)] ; i.e. the premium rate remaining for the
insurer is c(b) = λ(1 + θ)E[r(Z, b)] − λµ(θ − η). Here b = 0 means “full reinsurance” and
b = b̃ means “no reinsurance”. We assume for simplicity that r(z, b) is increasing and
continuous in b. In order that the problem considered below does not have the trivial
solution V (x) = |x|1I[x<0] we assume that θ > η. Here 1IA denotes the indicator function.
For a constant strategy B ≡ b, the diffusion approximation to the surplus process is

Xtb = x + λ(θE[r(Z, b)] − (θ − η)µ)t + λE[r(Z, b)2 ]Wt ,

 p
(2.4)

see Section 1.2. As an extension the insurer can continuously change the reinsurance
leading to the process
Z t Z tp
B


Xt = x + λ θE[r(Z, bs )] − (θ − η)µ ds + λE[r(Z, bs )2 ] dWs . (2.5)
0 0

26
 2.1 Diffusion Approximation

Here B = {bt } is an admissible strategy, i.e. adapted reinsurance strategy with bt ∈ [0, b̃].
The set of admissible strategies we denote by U . In the following we will denote the
time of ruin by τxB = inf{t ≥ 0 : XtB < 0, X0B = x}. Adding the capital injections gives

XtB,Y = XtB + Yt .
R∞
The return function associated to a reinsurance strategy B is V B (x) = Ex [ 0 e−δt dYtB ],
where YtB denotes the local time at 0 of {XtB }. We minimise V B (x) over all admissible
reinsurance strategies B and let the value function be defined as

V (x) = inf V B (x) .

B∈U

Because B = {b̃ : 0 ≤ t} is an admissible strategy, we find from Example 2.1.1 that

0 ≤ lim V (x) ≤ lim V b̃ (x) = 0. From Subsection 2.1.1 we know, that the capi-
x→∞ x→∞
tal injection process Y B for the process (2.5) is given through the local time YtB =
− min{ inf XsB , 0}. Thus, we can write for the return function V B (x):
0≤s≤t

hZ ∞ i ′ B
V B (x) = Ex e−δt dYtB = V B (0)E e−δτx ,

τxB

where B ′ is the strategy Bs′ = BτxB +s . If we choose an ε-optimal strategy B ′ for initial
capital 0 we have
B B
V (0)E e−δτx ≤ V B (x) ≤ (V (0) + ε)E e−δτx .
 

B
In order to minimise V B (x) we have to minimise E[e−δτx ] over all admissible strategies
B B
B. Let LB (x) = E[e−δτx ] and L(x) = inf E[e−δτx ]. It is clear that L(x) is a decreasing
B∈U
function and lim L(x) = 0.
x→∞
We assume for the moment, that the optimal reinsurance strategy exists and denote
∗
it by B ∗ , i.e. L(x) = LB (x). Let x, y ∈ R+ . If we start with the initial value x + y we
B ∗ . This will happen at the time τ B ∗ , i.e. at the ruin
have to cross the level y before τx+y x
time of the process with initial value x. Because the paths are continuous the process
∗
with initial capital x + y will have almost surely the value y at the time τxB . Thus, we
can write for the value function:
∗ B∗ B∗ B∗ ) B∗ B∗ B∗ )
LB (x + y) = E[e−δτx e−δ(τx+y −τx ] = E[e−δτx ]E[e−δ(τx+y −τx ]
B∗ B∗
= L (x)L (y) .

Here we used that for an optimal strategy we can minimise both terms independently,
that the underlying process has independent increments and that τx+y B ∗ − τ B ∗ = τ B ′′ ,
x y
where Bt′′ = Bτ∗B∗ +t . From the above equation we find that L(x) = exp(−βx) for some
x

27
2 Optimal Control of Capital Injections by Reinsurance

β > 0. Dividing the interval [0, x] in n intervals of equal length we have to solve the same
optimisation problem in each interval [kx/n, (k + 1)x/n]. This implies that the optimal
strategy should be constant. The above argument shows that also for any constant
strategy Lb (x) = exp(−β(b)x) for some β(b). We therefore need to maximise β(b).
We now use the stochastic control approach and motivate the Hamilton–Jacobi–
Bellman equation for x ≥ 0. Let h > 0 be very small. Let b ∈ [0, b̃] and τ b be the
time of ruin of the process X b , defined by (2.4), with initial capital x. Suppose further
that for each capital xh at time h there is an admissible strategy B ε = {bεt } such that
ε
V (xh ) + ε > V B (xh ). Construct a strategy B̃ = {b̃t } as follows. Let b̃t = b for t ≤ τ b ∧ h
and b̃t+h∧τ b = bεt . Then we have
h ε i h i
B b
B̃ −δ(τ b ∧h) b −δ(τ b ∧h)
V (x) ≤ V (x) = E V (Xτ b ∧h )e ≤ E V (Xτ b ∧h )e +ε .

Because ε was arbitrary we can let it be equal to zero. Suppose now that the value
function V (x) is twice continuously differentiable. Then we can apply Ito’s formula, see
Chapter C, and obtain
Z τ b ∧h n
−δ(τ b ∧h)
(Xτbb ∧h ) e−δs λθE[r(Z, b)] − λµ(θ − η) V ′ (Xsb )


e V − V (x) =
0
λE[r(Z, b)2 ] ′′ b o
+ V (Xs ) − δV (Xsb ) ds
2
Z τ b ∧h p
+ λE[r(Z, b)2 ]V ′ (Xsb ) dWs .
0

Suppose the stochastic integral is a true martingale. Taking expectation yields

τ b ∧h
hZ n λE[r(Z, b)2 ] ′′ b
e−δs λθE[r(Z, b)] − λµ(θ − η) V ′ (Xsb ) +


0≤E V (Xs )
0 2
o i
−δV (Xsb ) ds .

Dividing by h and letting h go to zero yields, provided the limit and expectation can be
interchanged:

λE[r(Z, b)2 ] ′′
λθE[r(Z, b)] − λµ(θ − η) V ′ (x) +


V (x) − δV (x) ≥ 0 .
2

This equation must hold for all b ∈ [0, b̃]. thus

λE[r(Z, b)2 ] ′′
λθE[r(Z, b)] − λµ(θ − η) V ′ (x) +


inf V (x) − δV (x) ≥ 0 .
b∈[0,b̃] 2

28
 2.1 Diffusion Approximation

Suppose further there is an optimal strategy B ∗ = {b∗t } such that lim b∗t = b, then we
t→0
can conclude similarly as above
∗
τ B ∧h
hZ n ∗ λE[r(Z, b∗s )2 ] ′′ B ∗
e−δs λθE[r(Z, b∗s )] − λµ(θ − η) V ′ (XsB ) +


0=E V (Xs )
0 2
∗
o i
−δV (XsB ) ds .

Dividing by h and letting h go to zero we obtain

λE[r(Z, b)2 ] ′′
λθE[r(Z, b)] − λµ(θ − η) V ′ (x) +


V (x) − δV (x) = 0 .
2
Therefore, the Hamilton–Jacobi–Bellman (HJB) equation for V (x) for x ≥ 0 is,
λE[r(Z, b)2 ] ′′
V (x) + λ θE[r(Z, b)] − (θ − η)µ V ′ (x) − δV (x) = 0 .


inf (2.6)
b∈[0,b̃] 2

Plugging in the solution V (x) = V (0)e−βx yields

λE[r(Z, b)2 ] 2

inf β − λ θE[r(Z, b)] − (θ − η)µ β − δ = 0 . (2.7)
b∈[0,b̃] 2

The above expression is continuous in b and therefore there is a value b∗ where the
minimal value is attained.
Similarly, for a constant strategy b, the function V (x) = (β(b))−1 e−β(b)x plugged in
Equation (2.3) yields the equation
λE[r(Z, b)2 ]
β(b)2 − λ θE[r(Z, b)] − (θ − η)µ β(b) − δ = 0 ,


(2.8)
2
from which we can find β(b). We can then maximise β(b). One readily can see from
the convexity of (2.8) in β that (2.7) and (2.8) yield the same solution and the same
minimiser b∗ of (2.7) and the same maximiser b∗ of β(b), respectively.
Define now  
α(b) = λ θE[r(Z, b)] − (θ − η)µ .
Then we can express β(b) as
s
α(b)  α(b) 2 2δ
β(b) = + + >0.
λE[r(Z, b)2 ] λE[r(Z, b]2 ) λE[r(Z, b)2 ]

We see that there are two possibilities to calculate V (x) and b∗ : Either, we can solve
(2.7) directly or we can maximise β(b).
Part of our derivation of V (x) was heuristically. In order to show that our solution is
correct, we prove the following verification theorem.

29
2 Optimal Control of Capital Injections by Reinsurance

Theorem 2.1.2
The constant strategy b∗t = b∗ is an optimal reinsurance strategy, where b∗ is a maximum
point of the function β(b) on the set [0, b̃]. The value function is given by V (x) =
(β(b∗ ))−1 exp(−β(b∗ )x) for x ≥ 0 and V (x) = (β(b∗ ))−1 − x for x < 0.
Proof: We have shown in Example 2.1.1 that the constant strategy b∗t = b∗ yields the
∗
value V b (x) = V (x). Consider now an arbitrary reinsurance strategy B = {bt }. The
corresponding surplus process is
Z t Z tp
B


Xt = x + λ θE[r(Z, bs )] − (θ − η)µ ds + λE[r(Z, bs )2 ] dWs .
0 0

Because V (x) is twice continuously differentiable on (0, ∞) we can apply Ito’s formula
(note that XtB,Y ≥ 0 and therefore we can change V (x) on (−∞, 0) without changing
V (XtB,Y )):
Z t
−δt B,Y
e V (Xt ) = V (x) + e−δs V ′ (XsB,Y ) dYsB
0
Z t
+ e−δs [Ds,B V (XsB,Y ) − δV (XsB,Y )] ds
0
Z t
e−δs V ′ (XsB,Y ) λE[r(Z, bs )2 ] dWs
p
+
0
Z t
≥ V (x) − e−δs dYsB
0
Z t
e−δs V ′ (XsB,Y ) λE[r(Z, bs )2 ] dWs ,
p
+
0
where
1
Ds,B V (x) = λE[r(Z, bs )2 ]V ′′ (x) + λ θE[r(Z, bs )] − (θ − η)µ V ′ (x)
 
2
denotes the infinitesimal generator of the process Xtbs , compare Proposition B.2.5, p.
170. We have used (2.6) and the fact that YtB only increases on the set {XtB,Y = 0}.
Because |V ′ (x)| ≤ 1 we obtain that the last term is a martingale with mean value 0.
Thus hZ t i
−δt B,Y 
e−δs dYsB .

V (x) ≤ e Ex V (Xt ) + Ex
0

Because 0 ≤ V (XtB,Y ) ≤ V (0), we get by monotone convergence

hZ ∞ i
V (x) ≤ Ex e−δs dYsB .
0

This proves the result. 

30
 2.1 Diffusion Approximation

Remark 2.1.3
Define for a constant b the exponent
s
α(b)  α(b) 2 2δ
β̃(b) = − + <0.
λE[r(Z, b)2 ] 2
λE[r(Z, b] ) λE[r(Z, b)2 ]

Note that 21 λE[r(Z, b)2 ]β̃ 2 (b) − α(b)β̃(b) = δ. We consider the martingale

Mt = exp(−β̃(b)(Xtb − x) − δt) .
R
Note that E[Mt ] = 1. Define the new measure on Ft by ν[A] = A M dP for A ∈ Ft . The
measure ν can be extended to a measure on F. For the details see Chapter C, Rolski et
al. [61, p. 461] or Schmidli [70, p. 215]. By Girsanov’s theorem, see Chapter C, {Xtb }
is under ν again a Brownian motion with drift, but the drift to infinity is stronger than
under P. By changing the measure we obtain
b  β̃(b)X bb
Lb (x) = E[e−δτx 1I[τxb <∞] ] = Eν e I[τxb <∞] e−β̃(b)x

τx 1

= Eν [1I[τxb <∞] ]e−β̃(b)x = ν[τxb < ∞]e−β̃(b)x .

Thus, our problem could be considered as a minimisation of a penalised ruin probabil-

ity. It is therefore not surprising that the optimal strategy becomes constant as in the
diffusion problem in Schmidli [68].

2.1.3 Examples
We now illustrate our result by explicit examples.

Proportional Reinsurance
We consider the case of proportional reinsurance. The self-insurance function has the
form r(Z, b) = bZ for some b ∈ [0, 1]. In particular, E[r(Z, b)] = bµ and E[r(Z, b)2 ] =
b2 µ2 . The corresponding Brownian motion becomes

Xtb = x + λµ bθ − (θ − η) t + b λµ2 Wt .

p

Equation (2.7) reads

λµ2 2 2

inf b β − λµ θb − (θ − η) β − δ = 0 .
b∈[0,1] 2

The minimum is taken at

µθ
b∗ = ∧1.
µ2 β

31
2 Optimal Control of Capital Injections by Reinsurance

Suppose for the moment that b∗ < 1. In oder to get the optimal β we have to solve the
equation
λµ2 θ 2
λµ(θ − η)β − −δ =0 .
2µ2
The solution is
λµ2 θ 2 + 2δµ2
β= , (2.9)
2µ2 λµ(θ − η)
giving
2λµ2 θ(θ − η)
b∗ = ∧1.
λµ2 θ 2 + 2δµ2
This is smaller than 1 if
λµ2 θ 2 < 2(δµ2 + λµ2 θη) .
We could write the latter condition as
s
2δµ2
η<θ<η+ η2 + .
λµ2

For b∗ = 1 we have from Example 2.1.1 that

p
λµη + λ2 µ2 η 2 + 2λδµ2
β= .
λµ2

We now consider numerical examples.

Choose for example θ = 0.5, η = 0.3, λ = 1, δ = 0.04, µ = 1 and µ2 = 2. Then we
have b∗ = 0.4878 and β = 0.5125.
Choosing δ = 0 corresponds to minimising the ruin probability.
In this case our
findings coincide with the result in Schmidli [70, p. 37]: b∗ = 2 1 − ηθ ∧ 1, i.e. b∗ = 0.8
and β = 0.3125. This b∗ differs very much from the optimal b with δ = 0.04, and the
corresponding value function for δ = 0 lies above the value function for δ = 0.04. We
see that discounting has a large influence.
Figure 2.1 compares the value of the optimal strategy with the value in the case without
reinsurance, b = 1. We can see that buying reinsurance lowers the costs considerably.
In the above example we obtained a strategy b∗ such that the drift coefficient of the
corresponding Brownian motion was positive. But it may be possible that the drift
of the Brownian motion, corresponding to the optimal constant reinsurance strategy,
would be negative. It means, that the risk of the portfolio is so large that, because of
the discounting, the insurer would be willing to “throw away” money in order to prevent
early capital injections. Of course, the insurer should better sell the business in this case.
The result is possible because we do not allow the option “no business at all”, which of
course has the minimal costs 0.

32
 2.1 Diffusion Approximation

V b (x)

6
V 1 (x)
V 0.4878 (x)

x
−2 0 2 4 6 8

Figure 2.1: Value function with B ∗ ≡ 0.4878 and return function for B ≡ 1 for propor-
tional reinsurance.

With parameters θ = 0.5, η = 0.3, λ = 0.05, δ = 0.04, µ = 20 and µ2 = 800 we

obtain b∗ = 0.058 < 1 and β = 0.216. It is easy to verify, that the drift coefficient
is now smaller than zero. Figure 2.2 gives the value functions of the optimal strategy
b∗ = 0.058, for the strategy b = 1 − ηθ = 0.4 (the maximal reinsurance such that the
drift remains positive) and no reinsurance b = 1. We see that reinsurance lowers the
costs considerably. Optimal reinsurance moves the costs to the future by lowering the
diffusion coefficient. The problem with the non-controlled process is that quite likely
there will be early costs. This costs have a larger weight than the future costs in the
controlled process.
In the case of proportional reinsurance it was possible to find an optimal strategy b∗
without knowing explicitly the claim size distribution. We only needed to know the first
two moments. The next example shows, that it will not be always the case.

Excess of Loss Reinsurance

Rb
We choose now r(Z, b) = min{Z, b} for b ∈ [0, ∞]. Then we have E[r(Z, b)] = 0 (1 −
Rb
G(x)) dx and E[r(Z, b)2 ] = 0 2x(1 − G(x)) dx, where G(x) is the claim size distribution.
Equation (2.7) reads
Z b  Z b 
2
inf λ x(1 − G(x)) dxβ − λ θ (1 − G(x)) dx − (θ − η)µ β − δ = 0 .
b∈[0,∞] 0 0

33
2 Optimal Control of Capital Injections by Reinsurance

V b (x)
V 1 (x)
32 V 0.4 (x)
V 0.058 (x)

24

16

x
−10 0 10 20 30 40

Figure 2.2: Value function with B ∗ ≡ 0.058 and return functions for B ≡ 0.4 and B ≡ 1
for proportional reinsurance.

The derivative with respect to b is negative for b < θ/β and positive for b > θ/β provided
G(θ/β) < 1. In any case, the minimum is taken at b∗ = θ/β. For solving the equation
we can let β = θ/b and solve for b:
Z b Z b  δb2
λ x(1 − G(x)) dx − λb (1 − G(x)) dx − b0 µ − 2 = 0 , (2.10)
0 0 θ

where b0 = 1 − η/θ. The left hand side of the equation is concave in b with value 0 at
zero, derivative λb0 µ in zero, and tends to −∞ as b → ∞. Therefore, there is a unique
b∗ > 0 solving the equation. We thus found the solution
( ∗
b∗ θ −1 e−θx/b : if x ≥ 0,
V (x) =
b∗ θ −1 − x : if x < 0.

We now ask when it is optimal not to reinsure the portfolio. From the derivations above
we get that it is always optimal to buy reinsurance if the support of the claim size
distribution is unbounded. Let b̃ = sup{x : P(Z > x) > 0} and suppose b̃ < ∞. No
reinsurance is chosen if b∗ ≥ b̃. This implies by the concavity of the left hand side of
(2.10) that
µ2 δb̃2 µ2 b̃ b̃2
λ − λb̃µ(1 − b0 ) − 2 = λ − ληµ − δ 2 ≥ 0 .
2 θ 2 θ θ

34
 2.1 Diffusion Approximation

V (x)

3.2 Vp (x)
VXL (x)
2.4

1.6

0.8

x
−1 0 1 2 3 4 5
Figure 2.3: Value functions for proportional Vp (x) and Excess of loss VXL reinsurances.

This is equivalent to p
b̃ λ2 µ2 η 2 + 2λµ2 δ − λµη
≤ ,
θ 2δ
or p
θ λµη + λ2 µ2 η 2 + 2λµ2 δ
≥ .
b̃ λµ2
We see that we do not reinsure if reinsurance is too expensive or if the maximal claim
size is small. Let us now assume, that the claim sizes are exponentially distributed
Z ∼Exp( µ1 ).
Then the Equation (2.10) reads
b b
δb2
Z Z
−x/µ
λ xe dx − λb e−x/µ dx + λbb0 µ − = 0,
0 0 θ2

which is equivalent to

η δb2
−λµ2 e−b/µ − λµb + λµ2 − 2 = 0.
θ θ
Choosing as in Example 2.1.3, η = 0.3; θ = 0.5; δ = 0.04; µ = 1; λ = 1, the above
equation becomes
0 = − exp(−b) − 0.6b + 1 − 0.16b2 = 0 .

35
2 Optimal Control of Capital Injections by Reinsurance

Numerical solution yields b∗ = 0.715, i.e. β = 0.6993. According to this, the value
function is (
1
exp(−0.6993 · x) : x ≥ 0 ,
V 0.715 (x) = 0.6993
1
0.6993 − x : x < 0.
Gerber [29, p. 130] has shown that excess of loss reinsurance maximises the adjust-
ment coefficient. Because β(b) is a quantity similar to an adjustment coefficient it is
not surprising that also in our model excess of loss is favourable to proportional reinsur-
ance. Figure 2.3 shows the value functions for optimal proportional and excess of loss
reinsurance.

2.2 The Classical Risk Model

From now on we work on a probability space (Ω, F, P) that should be large enough to
carry all the stochastic quantities defined below. We start with introducing the model
and calculate the expected discounted value of the capital injections if no reinsurance
is taken. In Subsection 2.2.1 we will prove that the value function is Lipschitz continu-
ous and therefore absolutely continuous. We show that the value function satisfies the
Hamilton–Jacobi–Bellman equation. In Subsection 2.2.2 we illustrate the theory by some
specific examples. Finally, Subsection 2.2.3 treats the special case of no discounting.
Consider the classical risk model. The number of claims is assumed to follow a Poisson
process {Nt } with intensity λ. Claim sizes Z1 , Z2 , . . . are independent of {Nt }, positive
and iid with distribution function G(z) and first moment µ. By Z we denote a generic
random variable with the same distribution as Zi . The aggregate claim process is given
by
XNt
Xt = x + ct − Zi ,
i=1
where x is the initial capital, c = λµ(1+η) is the premium rate and η is the safety loading.
We do not assume that η > 0 if the discounting parameter δ introduced below is strictly
positive. The insurer can buy reinsurance. Choosing the level b ∈ [0, b̃], the insurer pays
r(Zi , b) for a claim of size Zi . Here b = 0 means full reinsurance, b = b̃ ∈ (0, ∞] means
no reinsurance. For example, for proportional reinsurance r(Z, b) = bZ and b ∈ [0, 1].
For excess of loss reinsurance we obtain r(Z, b) = min{Z, b} and b ∈ [0, ∞].
For the reinsurance cover the insurer pays a premium at rate c − c(b). That is,
the premium rate left for the cedent is c(b). For simplicity we assume that r(z, b) is
continuous and increasing in both z and b and that c(b) is continuous and increasing
with c(0) < 0 and c(b̃) = c. The assumption c(0) < 0 is needed in order that the problem
below is not trivial. If the reinsurer uses an expected value principle with safety loading
θ, we get
c(b) = c − (1 + θ)λE[Z − r(Z, b)] = (1 + θ)λE[r(Z, b)] − (θ − η)λµ .

36
 2.2 The Classical Risk Model

Figure 2.4: The process {XtB,Y }.

In order that c(0) < 0 we assume θ > η. The insurer can choose the level bt at any time
point t. Because no information on the future can be used the process {bt } is assumed
to be cadlag and adapted. The surplus of the insurer including reinsurance has then the
form
Z t Nt
X
XtB = x + c(bs ) ds − r(Zi , bTi − ) .
0 i=1

To prevent, that the surplus process becomes negative, the insurer has to inject additional
capital. We denote the accumulated capital injections until time t by {YtB }. The surplus
with capital injections has therefore the form
Z t Nt
XtB,Y
X
=x+ c(bs ) ds − r(Zi , bTi − ) + YtB .
0 i=1

One of the possible structures of the surplus process under some non-constant reinsurance
strategy is illustrated in Figure 2.4. We are interested in the expected discounted value
hZ ∞ i
V B (x) = Ex e−δt dYtB ,
0

where δ ≥ 0 is a discounting factor. In the case δ = 0 we assume that there is b such that
c(b) > λE[r(Z, b). The discounting expresses the investment preferences of the owners.

37
2 Optimal Control of Capital Injections by Reinsurance

δ > 0 means, that the investor prefers capital injections tomorrow to capital injections
today. Hence, δ can be seen as a parameter of the risk averseness. δ = 0 means that
the investor is indifferent to the time of the injections. Our goal is to find the minimal
return function V (x) = inf V B (x) and an optimal strategy B such that V (x) = V B (x).
B∈U
Here again U denotes all cadlag adapted processes {bt } with values in [0, b̃].
It is clear that capital injections only take place at times where the surplus would
be negative, and that after a capital injection the surplus is zero. This follows by the
discounting and the possibility that no injections are needed if x = 0.
We start by considering the case where no reinsurance is available.

Example 2.2.1
Suppose that there is no possibility to buy reinsurance. The surplus process has then
the form
XNt
XtY = x + ct − Zi + Yt .
i=1
We have to inject capital when the process falls below zero. Let
n Nt
X o
τ1 = inf t ≥ 0 : x + ct − Zi < 0
i=1

denote the time of the first ruin, compare Figure 1.1, and
N (τ1 )
X
Ỹ1 = Zi − x − cτ1
i=1

denote the first injection. At time τ1 the process starts with initial capital 0. Let for
i ≥ 2,
n XNt o
τi := inf t ≥ τi−1 : c(t − τi−1 ) − Zj < 0
j=N (τi−1 )+1

denote the time of the i-th injection. If τi−1 = ∞ we let τi = ∞. The size of the i-th
injection becomes, provided τi < ∞,
N (τi )
X
Ỹi = Zj − c(τi − τi−1 ) .
j=N (τi−1 )+1

The accumulated injections can then be described as

∞
X
Yt = Ỹi 1I[τi ≤t] .
i=1

38
 2.2 The Classical Risk Model

Denote by
φ(x) = Ex [Ỹ1 exp{−δτ1 }; τ1 < ∞]
the expected discounted penalty at ruin and by ψ(x) = E[exp{−δτ1 }; τ1 < ∞] the
discounted ruin probability. Then V b̃ (x) = φ(x) + ψ(x)V b̃ (0) for x ≥ 0. It follows
recursively that
∞
b̃
X ψ(x)φ(0)
V (x) = φ(x) + ψ(x)φ(0) ψ(0)k = φ(x) + .
1 − ψ(0)
k=0

Let ρ be the unique positive solution to the generalised Lundberg equation

Z ∞
δ + λ − cρ = λ e−ρx dG(x) . (2.11)
0

It follows from Gerber and Shiu [31], see also Schmidli [71],
λ ∞ −ρy λ ∞ δ
Z Z
ψ(0) = e (1 − G(y)) dy = (1 − e−ρx ) dG(x) = 1 − ,
c 0 cρ 0 cρ
λ ∞ δ − (c − λµ)ρ
Z
φ(0) = (1 − e−ρz )(1 − G(y)) dy = .
cρ 0 cρ2
This gives
 1 c − λµ 
V b̃ (x) = φ(x) + ψ(x) − .
ρ δ
In particular, V b̃ (0) = φ(0)/(1 − ψ(0)) = (δ − (c − λµ))/(δρ). Expressions for φ(x) and
ψ(x) we find in Gerber and Shiu [31]
∞
X ∞
X
φ(x) = g∗n ∗ h2 (x) , ψ(x) = g∗n ∗ h1 (x) ,
n=0 n=0

where
Z ∞
ω1 (u) = dG(z) = 1 − G(u) ,
u
Z ∞ Z ∞
ω2 (u) = (z − u) dG(z) = (1 − G(z)) dz ,
u u
λ ρx ∞ −ρz
Z
g(x) = e e dG(z) ,
c
Z x∞
λ ρx λ ρx ∞ −ρz
Z
−ρz
h1 (x) = e e ω1 (z) dz = e e (1 − G(z)) dz ,
c c
Z x∞ x
λ ρx λ 1
h2 (x) = e e−ρz ω2 (z) dz = ω2 (x) − h1 (x) .
c x ρc ρ

39
2 Optimal Control of Capital Injections by Reinsurance

The above functions are all differentiable for x > 0. It means, that the function V b̃ (x)
is differentiable as well.

More specifically, we consider now exponentially distributed claim sizes with mean
value µ. The two solutions to Equation (2.11) are
p
[δµ + λµ − c] + [δµ + λµ − c]2 + 4cµδ
ρ = ,
2cµ
p
[δµ + λµ − c] − [δµ + λµ − c]2 + 4cµδ
R = .
2cµ

Note that for R and ρ the following relations hold:

λµ
1 + Rµ = (2.12)
c(1 + ρµ)
−δ
ρR = . (2.13)
cµ

From Schmidli [71] it follows readily that

φ(x) = µ(1 + µR)eRx , ψ(x) = (1 + µR)eRx .

This yields the solution

1 + Rµ Rx
V b̃ (x) = e
−R
for x ≥ 0. If we would start with a negative initial capital we get

1 + Rµ
V b̃ (x) = V b̃ (0) − x = −x
−R
for x < 0.
For the parameters µ = 1, λ = 1, δ = 0.04, θ = 0.5 and η = 0.3 we obtain for x ≥ 0:
V (x) = 2.3e−0.3x . This function is plotted in the Figure 2.5.
b̃

The analogous calculations can be made for every constant strategy b with c(b) > 0.
Equation (2.11) reads then
Z ∞
δ + λ − c(b)ρ = λ e−ρr(z,b) dG(z) .
0

Representation of the functions V b (x) as a sum of two Gerber–Shiu functions has the
advantage, that we can approximate V b (x) corresponding to some complex individual
claims distribution by V b (x) with comparatively simple claim size distribution, for ex-
ample exponential distribution.

40
 2.2 The Classical Risk Model

V b̃ (x)

x
−2 0 2 4 6 8
Figure 2.5: Return function V b̃ (x).

Pitts and Politis [59] suggested to consider a functional Φ, which takes as input the
claim size distribution and produces as output the corresponding Gerber–Shiu penalty
function. In the case the functional Φ is continuous the desired Gerber–Shiu function will
be close to the one with simple claim size distribution if the two distribution functions
are close. This type of approximation is called zeroth-order approximations. A better
approximation, first-order approximation, can be obtained in the case the functional Φ
is Frechet differentiable. 

2.2.1 Properties of V (x)

Lemma 2.2.2
The function V (x) fulfils lim V (x) = 0 and is Lipschitz continuous with |V (x)−V (y)| ≤
x→∞
|x − y|.
Proof: It is clear that V (x) is decreasing. lim V (x) = 0 follows immediately from
x→∞
Example 2.2.1. Let z > x and B = {bt } be a reinsurance strategy for initial capital z
such that V B (z) ≤ V (z) + ε. For initial capital x choose the strategy B̃ (which is not
optimal): inject the capital z − x and then follow the strategy B. Thus

V (x) − V (z) ≤ V B̃ (x) − V B (z) + ε = z − x + ε .

Because ε is arbitrary we have |V (x) − V (z)| ≤ |x − z|, which proves the Lipschitz-
continuity. 

As a consequence, V (x) is absolutely continuous.

41
2 Optimal Control of Capital Injections by Reinsurance

We now only look for V (x) for x ≥ 0 and let V (x) = V (0) − x for x ≤ 0. The
Hamilton–Jacobi–Bellman equation is
Z ∞
V x − r(z, b) dG(z) + c(b)V ′ (x) − (δ + λ)V (x) = 0 .


inf λ (2.14)
b∈[0,b̃] 0

The derivation of the Hamilton–Jacobi–Bellman equation is described in the proof of

Theorem 2.2.3 below. Next we will show that the value function really solves the HJB
equation. Let b0 denote the unique argument at which c(b0 ) = 0.

Theorem 2.2.3
The function V (x) is continuously differentiable from the right and from the left at
all points where bt = b0 is not optimal. Its derivatives solve Equation (2.14) with the
interpretation c(b0 )V ′ (x) = 0, if the derivatives do not exist. Moreover, if there exists a
b such that c(b) ≥ λE[r(Z, b)], then any decreasing positive solution to (2.14) coincides
with V (x).

Proof: Assume x > 0. Let h > 0 and b ∈ [0, b̃] be fixed. We can assume that x+c(b)h ≥ 0,
i.e. the ruin does not occur because of the premium payments to the reinsurer. Let T1
be the time of the first claim and choose ε > 0. We further choose n ∈ N such that 2(x +
k
c(b)h)/n < ε. For each k there is a strategy B k = {bkt } such that V B (xk ) ≤ V (xk )+ε/2.
For initial capital xk ≤ x < xk+1 we choose the strategy B k . Thus, V B (x) ≤ V B (xk ) ≤
V (xk )+ ε/2 ≤ V (x)+ (x− xk )+ ε/2 < V (x)+ ε. This shows that for each x ∈ [0, x+ c(b)]
we can find in a measurable way a strategy B̂(x) such that V B̂(x) (x) < V (x)+ε. Consider
now the strategy bt = b1I[t<T1 ∧h] + b̂t−(T1 ∧h) (XT1 ∧h )1I[t≥T1 ∧h] . By conditioning on FT1 ∧h

V (x) ≤ V B (x)
 
= Ex 1I[T1 ≤h] V B̂ (x + c(b)T1 − r(Z1 , b))e−δT1 + 1I[T1 >h] V B̂ (x + c(b)h)e−δh
Z hZ ∞  
= λe−(δ+λ)t V B̂ x + c(b)t − r(z, b) dG(z) dt
0 0
Z ∞  
+ e−δh V B̂ x + c(b)h λe−λt dt
h
Z hZ ∞  
≤ λ e−(δ+λ)t V x + c(b)t − r(z, b) dG(z) dt
0 0
−(δ+λ)h
+e V (x + c(b)h) + ε .

Because ε was arbitrary we can let it be equal to zero. Rearranging the terms and

42
 2.2 The Classical Risk Model

dividing by h yields:


V x + c(b)h − V (x) −(δ+λ)h e−(δ+λ)h − 1
0 ≤ e + V (x)
h h
Z hZ ∞
1
λe−(δ+λ)t V x + c(b)t − r(z, b) dG(z) dt .


+ (2.15)
h 0 0

Now D 2
R t choose a strategy D(h) = {dt (h)}, such that V (x) ≤ V (x) + h . Let a(t, h) =
0 c(ds ) ds. In the same way as above we find, that

V (x + a(h, h)) − V (x) −(δ+λ)h e−(δ+λ)h − 1

0 ≥ e + V (x)
h h
Z hZ ∞
1
λe−(δ+λ)t V x + a(t, h) − r(z, dt ) dG(z) dt − h .


+
h 0 0

All terms with exception of the first one converge. We write

V (x + a(h, h)) − V (x) V (x + a(h, h)) − V (x) a(h, h)

= .
h a(h, h) h

The first term is in [−1, 0], the second term in [c(0), c]. Thus, there exists a sequence
hn → 0 such that

V (x + a(hn , hn )) − V (x) V (x + a(h, h)) − V (x)

lim = lim sup .
n→∞ hn h→0 h

Because V (x) is Lipschitz continuous, the above limit is finite. W.l.o.g. we assume, that
a(hn , hn )/hn converges to some c(b̌). This yields, using that {dt } is cadlag,

V (x + a(hn , hn )) − V (x)
lim − (λ + δ)V (x)
n→∞ hn
Z ∞


+λ V x − r(z, b̌) dG(z) dt ≥ 0 .
0

From (2.15) with b = b̌ we conclude that equality holds. In particular, the limit

V (x + c(b̌)h) − V (x)
lim
h→0 h
exists because the above limit does not depend on the subsequence chosen. Thus, if
c(b̌) > 0 then V (x) is differentiable at x from the right. If c(b̌) < 0 then V (x) is
differentiable at x from the left. In both cases we have shown (2.14) for the corresponding
derivative.

43
2 Optimal Control of Capital Injections by Reinsurance

Starting with initial capital x − c(b)h yields

V x − c(b)h − V (x) −(δ+λ)h e−(δ+λ)h − 1
0 ≤ e + V (x − c(b)h)
−h h
1 h ∞ −(δ+λ)t
Z Z


+ λe V x − c(b)(h − t) − r(z, b) dG(z) dt . (2.16)
h 0 0

Suppose now that V (x) is differentiable in x from the right; i.e., c(b̌) > 0. At each point
z < x we have seen that there is b̌(z) at which the minimum in (2.14) is taken. Suppose
there is a sequence xn tending to x from the left such that c(b̌n ) > 0. By taking a
subsequence we can assume that b̌n converges to some value b∗ with c(b∗ ) ≥ 0. We can
then find a sequence of hn such that (2.16) converges and by continuity (2.14) holds.
If c(b∗ ) > 0 differentiability from the left follows. If there is a sequence xn tending to
x from below such that c(b̌n ) < 0, then choose hn = −(x − xn )/c(b̌n ). If hn → 0 then
differentiability from the left follows. As for the case c(b̌n ) > 0 Equation (2.14) follows.
If hn does not converge to zero, then c(b̌n ) tends to zero, and (2.14) with the minimum
taken at b0 . If now b̌n = b0 , then in the limit (2.14) holds with the minimum taken at
b0 . This shows the result at all points where b̌ 6= b0 , that is, at all points where
Z ∞
λ V (x − r(z, b0 )) dG(z) > (δ + λ)V (x) .
0

At x = 0, we have to consider strategies with b ≤ b0 and b > b0 separately. For b > b0 ,

Equation (2.15) holds.
If b ≤ b0 , XtB = 0 and capital is injected at rate c(b) and the claims are paid by capital
injections. This yields
∞ ∞
λ c(b)
hX i Z
−δTi
V (0) = E r(Zi , b)e − c(b) e−δt dt = E[r(Z, b)] − . (2.17)
0 δ δ
i=1

Thus, equality in Equation (2.15) holds. If V (0) is given by (2.17), it follows from (2.15)
for b > b0 that the derivative from the right must be −1. If

λ c(b)
V (0) > E[r(Z, b)] − ,
δ δ

then the arguments above show that (2.14) holds with the infimum taken at some b̌ ≥ b0 .
Because b0 is not possible, we get differentiability from the right.
We show now uniqueness of the solution. Let f (x) be a decreasing and positive solution
to (2.14). We conclude that f ′ (0) ≥ −1. Because otherwise, the right hand side of (2.14)
would be strictly negative at x = 0 for any b for which c(b) > λE[r(Z, b)]. Consider an
arbitrary strategy B. From Proposition A.2.2, Chapter A, we know, that the process,

44
 2.2 The Classical Risk Model

using f (XTi − − r(Zi , bTi − )) = f (XTi ) + ∆YTi ,

Nt  
B,Y
X X
Mt = f (XτB,Y
i
) − f (Xτi − ) e−δτi
+ e−δs ∆Ys
i=1 s≤t
Z t  Z ∞ 
−δs
− e λ f (XsB,Y − r(z, bs )) dG(z) − f (XsB,Y ) ds
0 0

is a martingale. Because f (x) is absolutely continuous we also have

f (XτB,Y
i−
)e−δτi − f (XτBi−1 )e−δτi−1
Z τi −
1I[X B,Y >0] c(bs )f ′ (XsB,Y ) − δf (XsB,Y ) e−δs ds


=
s
τi−1
Z τi −
− 1I[X B,Y =0] δf (0) ds .
s
τi−1

We obtain, that
Z t
f (XtB,Y −δt
f (X0B,Y 1I[X B,Y >0] c(bs )f ′ (XsB,Y ) − δf (XsB,Y ) e−δs ds


)e − )−
s
0
Z t  Z ∞ 
− e−δs λ f (XsB,Y − r(z, bs )) dG(z) − f (XsB,Y ) ds
0 0
X
−δs
+ e ∆Ys
s≤t

is a martingale. Using the martingale property and taking the expectations yields
h
Ex f (XtB,Y )e−δt − f (X0B,Y ) +
X
e−δs ∆Ys
s≤t
Z t  Z ∞
−δs
− e λ f (XsB,Y − r(z, bs )) dG(z) − (δ + λ)f (XsB,Y )
0 0
 i
+ 1I[XsB,Y >0] c(bs )f ′ (XsB,Y ) ds = 0 .

If XsB,Y = 0 and c(bs ) < 0, then c(bs )f ′ (XsB,Y ) ≤ −c(bs ). If c(bs ) ≥ 0, then
c(bs )f ′ (XsB,Y ) ≤ 0. Thus
h Z t
Ex f (XtB,Y )e−δt − f (x) + e−δs dYs
0
Z t  Z ∞
− e−δs λ f (XsB,Y − r(z, bs )) dG(z) − (δ + λ)f (XsB,Y )
0 0
 i
′ B,Y
+ c(bs )f (Xs ) ds ≥ 0 .

45
2 Optimal Control of Capital Injections by Reinsurance

Because B was arbitrary and using (2.14) we obtain from the above equation
h Z t i
f (x) ≤ Ex f (XtB,Y )e−δt + e−δs dYs .
0

B ∗ ,Y
Equality holds if B is the strategy b∗t = b(Xt ), where b(x) is an argument at which
B,Y
the minimum in (2.14) is taken. Because f (Xt ) ≤ f (0) < ∞, so that we obtain
lim Ex [f (XtB,Y )e−δt ] = 0 by bounded convergence. This proves that f (x) = V (x). 
t→∞

Suppose now that

c(b) = λ(1 + θ)E[r(Z, b)] − λµ(θ − η) ; (2.18)

i.e., the reinsurer uses an expected value principle. Consider the capital x = 0. Consider
first the case that a strategy with c(b) ≤ 0 is optimal at x = 0. Then, compare with
(2.17) in the Appendix,
λ  
V (0) = E[r(Z, b)] − (1 + θ)E[r(Z, b)] − (θ − η)µ
δ
λ 
= (θ − η)µ − θE[r(Z, b)] .
δ
This is decreasing in b, hence b = b0 would be optimal.
In particular, V (0) = λE[r(Z, b0 )]/δ. Let κ, ε > 0 such that κ > c(b̃)ε. Consider the
strategy bt = b0 1I[t≥T1 ∧ε] . This strategy has the return function bounded by
λ λ 
e−(λ+δ)ε E[r(Z, b0 )] − c(b̃)ε(1 − G(κ))
δ Z λ+δ
ε
λ 
+ λµ + E[r(Z, b0 )] − c(b̃)t(1 − G(κ)) λe−(λ+δ)t dt .
0 δ
Taking the derivative with respect to ε shows that the function is decreasing in ε, with a
derivative bounded away from zero. Thus for κ and ε small enough the return function
of the above strategy is smaller than the return function of the strategy bt = b0 . This
shows that b0 cannot be optimal.
Equation (2.14) at x = 0 reads

inf λE[r(Z, b)] 1 + (1 + θ)V ′ (0) − λµ(θ − η)V ′ (0) − δV (0) = 0 .

 
b∈[0,b̃]

We see that the minimum is taken either at b = 0 or b = b̃. Because b = 0 is not

optimal, we conclude that b = b̃. Because a level with b(x) ≤ b0 cannot be crossed, this
means that no reinsurance will be taken for capital close to zero. In particular we obtain
1 1
V ′ (0) < − 1+θ and V ′ (x) < − 1+θ for x ∈ [0, y) for some y > 0.

46
 2.2 The Classical Risk Model

Remark 2.2.4
Consider now the function
Z ∞
gx (b) := c(b)V ′ (x) − (δ + λ)V (x) + λ


V x − r(z, b) dG(z) .
0

Let further b1 , b2 ∈ [0, b̃] and b1 > b2 then we obtain with Lipschitz continuity of V (x)
for every x ∈ [0, ∞):
Z ∞



gx (b1 ) − gx (b2 ) = λ V x − r(z, b1 ) − V x − r(z, b2 ) dG(z)
0
+ (c(b1 ) − c(b2 ))V ′ (x)
Z ∞



= λ V x − r(z, b1 ) − V x − r(z, b2 )
0
n o
+ r(z, b1 ) − r(z, b2 ) (1 + θ)V ′ (x) dG(z)
Z ∞n o
≤ λ r(z, b1 ) − r(z, b2 ) · [1 + (1 + θ)V ′ (x)] dG(z) .
0
1
V ′ (x) ≤ − 1+θ implies, that gx (b) is decreasing, so that the minimum is taken in b = b̃,
which is then the optimal strategy if V (x) is differentiable in x. On the other hand, if
1
b0 6= b < b̃ is optimal for some x ∈ [0, ∞), then it must hold V ′ (x) > − 1+θ .

Lemma 2.2.5
If the value function V (x) is convex, then it is continuously differentiable.
Proof: From Theorem 2.2.3 we know, that V (x) is continuously differentiable in all
x ∈ R+ , where b0 is not optimal. Let V ′ (x−) and V ′ (x+) denote the derivatives from
the right and from the left respectively. Assume further that there is x̃ ∈ R+ with
V ′ (x̃−) < V ′ (x̃+). Define
Z ∞
f (x) := λ V (x − r(b0 , z)) dG(z) − (δ + λ)V (x)
0

Note that it holds f (x) ≥ 0. Let further x̃ := inf{x ∈ R+ : f (x) = 0}. Because the
optimal strategy in 0 is b∗ = b̃, it holds x̃ > 0. By Theorem 2.2.3 we obtain f (x̃) = 0.
W.o.l.g we can assume that f (x) is differentiable on (0, x̃).
Because f (x̃) = 0 there exist sequences (hn )n≥0 ∈ [0, x̃], lim hn = x̃, with f ′ (hn ) < 0
n→∞
and (xn )n≥0 ∈ (x̃, ∞), lim hn = x̃, with f ′ (xn ) ≥ 0. Letting n go to infinity we obtain
n→∞
Z ∞
λ V ′ (x̃ − r(b0 , z)) dG(z) − (δ + λ)V ′ (x̃−) ≤ 0 ,
0
Z ∞
λ V ′ (x̃ − r(b0 , z)) dG(z) − (δ + λ)V ′ (x̃+) ≥ 0 .
0

47
2 Optimal Control of Capital Injections by Reinsurance

Thus, V ′ (x̃+) ≤ V ′ (x̃−), which is a contradiction. 

We will see in the next section that V (x) is convex in the case of proportional reinsurance.

2.2.2 Examples
Example 2.2.6 (Proportional Reinsurance and Z ∼Exp(1/µ).)
For proportional reinsurance r(Z, b) = bZ and the expected value principle, we get that
V (x) is a convex function. Indeed, let x, z ≥ 0, α ∈ (0, 1) and y = αx + (1 − α)z. Let
{bxt } be the optimal strategy for initial capital x and {bzt } the optimal strategy for initial
capital z. Define the new strategy byt = αbxt + (1 − α)bzt . Then for the expected value
principle c(byt ) = αc(bxt ) + (1 − α)c(bzt ). Then

Xty − Yty + αYtx + (1 − α)Ytz

Z t Nt
byTi − Zi + αYtx + (1 − α)Ytz
X
y
= y+ c(bs ) ds −
0 i=1
= αXtx + (1 − α)Xtz ≥ 0 .

This implies that Yty ≤ αYtx + (1 − α)Ytz . Thus

y
V (y) ≤ V B (y) ≤ αV (x) + (1 − α)V (z) ,

which proves the convexity. In particular, all derivatives from the left and from the right
exist and therefore solve the HJB equation.
Note, that the convexity in the case of proportional reinsurance can be also shown for
all premium calculation principles, where c(b) is concave in b. This is for example the
case for the standard deviation and variation principles. For definition see Subsection
1.1.2, p. 14.
Let us assume Z ∼Exp(1/µ). The HJB equation reads now:
∞
λ
Z
inf V (x − bz) exp(−z/µ) dz + c(b)V ′ (x) − (λ + δ)V (x) = 0 .
b∈[0,1] µ 0

From Example 2.2.1 we know, that the return function for the constant strategy b = 1
is given by (
1+Rµ
1
V (x) = 1+Rµ−R exp(Rx) : x ≥ 0 ,

−R − x : x<0,

where p
δµ + λµ − c − [δµ + λµ − c]2 + 4cµδ
R= .
2cµ

48
 2.2 The Classical Risk Model

Figure 2.6: Dependence of critical θ on chosen η.

Consider now the function

λ ∞ 1
Z
−z
gx (b) = V (x − bz)e µ dz + c(b)(V 1 )′ (x) − (δ + λ)V 1 (x) .
µ 0

The function gx (b) is convex in b for each x ∈ [0, ∞). Hence it holds for the first
derivative of gx (b) with respect to b: gx′ (b) ≤ gx′ (1). Therefore, V 1 (x) is the value
function if gx′ (1) ≤ 0 for all x ∈ [0, ∞), which is equivalent to the inequality
∞
1 (V 1 )′ (x − z)
Z
z
· z · e− µ dz ≤ (1 + θ) .
µ2 0
1 ′
(V ) (x)
(V 1 )′ (x−z) (V 1 )′ (x−z)
The function (V 1 )′ (x) is increasing in x with lim 1 ′ = e−Rz . Letting x go to
x→∞ (V ) (x)
1
infinity in the left hand side of the above inequality we obtain θ ≥ (Rµ+1)2 − 1.

For the parameters δ = 0.04, µ = 1, λ = 1 and η = 0.1 we obtain θ ≥ 813. In this

case the value function is
(
3.55 exp(−0.22x) : x ≥ 0 ,
V (x) =
−x + 3.55 : x<0.

In Figure 2.6 the maximal value of θ for which reinsurance should be bouhgt is given as
a function of η. The parameters are δ = 0.04, µ = 1 and λ = 1.

49
2 Optimal Control of Capital Injections by Reinsurance

50

40

30

x
2
20

1
10

K2
0

0 2 4 6 8 10 0,42 0,44 0,46 0,48 0,50

x b

Figure 2.7: Functions Ṽ (x) (solid line) and V 1 (x) (dashed line) and the corresponding
“optimal” level b(x).

1
The optimal strategy for the case θ < (Rµ+1) 2 − 1 is more complicated. In this case,

the optimal strategy is not constant.

Consider for the moment constant strategies only. It is straight forward to verify, that
values corresponding to the constant strategies b ∈ (b0 , b̃], b− ∈ (0, b0 ), 0 and b0 , with
c(b0 ) = 0, are given by the functions
( 1+R(b)µb
b
− R(b) eR(b)x : x ≥ 0 ,
V (x) =
−x − 1+R(b)µb
R(b) : x<0,
(
−c(b)+λbµ ρ(b)x
e : x≥0,
V b− (x) = δ
−c(b)+λbµ
−x + δ : x<0,
( δ
−c(0) − c(0) x
V 0 (x) = δ e : x≥0,
−c(0)
−x + δ : x<0,
( δ
λb0 µ − b0 µ(δ+λ) x
V b0 (x) = δ e : x≥0,
λb0 µ
−x + δ : x<0
respectively, where
p
[δµb + λµb − c(b)] − [δµb + λµb − c(b)]2 + 4c(b)µbδ
R(b) = ,
2c(b)µb
p
[δµb + λµb − c(b)] + [δµb + λµb − c(b)]2 + 4c(b)µbδ
ρ(b) = .
2c(b)µb

50
 2.2 The Classical Risk Model

The calculations are similar to the calculations in Example 2.2.1. It is easy to see, that
V b (x) < V b− (x) for every choice of b ∈ [b0 , 1] and b− ∈ [0, b0 ). So we will consider only
the constant strategies with positive premium income.
It is easy to see, that the optimal b ∈ [b0 , 1] minimising V b (x) depends on x. This
shows that the optimal strategy cannot be constant. For parameters given above the
function Ṽ (x) = min V b (x) is plotted in Figure 2.7
b∈[b0 ,1]
Next we will calculate the value function numerically for exponentially and Pareto
distributed claim sizes. Note, that the start value V (0) should be calculated separately.
Denote the function, which result from choosing some start value a > V (0) by f (x) and
define
g(x) := f (x) − V (x) .

g(x) is continuously differentiable and g(0) > 0 by definition. Let further b∗ (x) denote
the optimal strategy for V (x), then we have
Z ∞
c(b∗ (x))g′ (x) ≥ λ g(x − b∗ (x)z) dG(z) − (δ + λ)g(x) .
0

Because g(0) > 0 and b∗ (0) = R1 it follows g′ (0) > 0. Let x̂ = inf{x : g′ (x) = 0}. Since
∞
g′ (x) > 0 on [0, x̂) it holds λ 0 g(x − b∗ (x)z) dG(z) − (δ + λ)g(x) > 0, which is a
contradiction. Therefore, numerically calculated functions with start values bigger than
V (0) do not cut the curve V (x), i.e. they do not cut the x axis and begin to increase.
In order to specify V (0) we begin with some start value V 1 (0) and calculate the
corresponding function. If the function begin to increase, the chosen value V0 was too
big. Proceeding in that way we obtain after a while the right value V (0).
In Figure 2.8 one can see the value function numerically calculated under assumption,
that the claims are exponentially distributed. The initial value calculated in the above
described way is V (0) = 1.7. 

Example 2.2.7 (Proportional Reinsurance and Z ∼Pareto(2, µ).)

µ2
Consider now Pareto(2, µ) distributed claim sizes, i.e. G(x) = 1− (µ+x) 2 . It is impossible
1
to give a closed form expression for the function V (x) in this case. But we can calculate
it numerically using the method of Gerber and Shiu [31]. Consider again the parameters
δ = 0.04, λ = µ = 1, θ = 0.5 and η = 0.3. Then it is easy to verify with help
of some mathematical programs: ρ = 0.08331, φ(0) = 1.66343 and V 1 (0) = 4.50372.
Numerically calculated value function is given in Figure 2.10. The initial value is in this
case V (0) = 2.3847712. 

Example 2.2.8 (Excess of Loss Reinsurance and Z ∼Exp(1/µ).)

For excess of loss reinsurance r(z, b) = min(z, b) we cannot determine whether the value

51
2 Optimal Control of Capital Injections by Reinsurance

3,5

3,0

2,5

2,0

1,5

1,0

0,5

0 2 4 6 8 10

Figure 2.8: Numerically calculated value function V (x) with initial value V (0) = 1.7
(solid line) and V 1 (x) (dashed line) for exponentially distributed claims.

1,0

0,9

0,8

0,7

0,6

0,5

0,4

0,3

0,2

0 1 2 3 4 5 6 7 8 9

Figure 2.9: Optimal strategy b(x) for exponentially distributed claims.

52
 2.2 The Classical Risk Model

2,5

2,0

1,5

1,0

0,5

0 1 2 3 4 5 6

Figure 2.10: Numerically calculated V (x) for Pareto distributed claims with initial value
2.3847.

1,0

0,9

0,8

0,7

0,6

0,5

0,4

0,3

0,2

0,0 0,5 1,0 1,5 2,0 2,5

Figure 2.11: Optimal strategy b(x) for Pareto distributed claims.

53
2 Optimal Control of Capital Injections by Reinsurance

function is convex. For the HJB equation we get

Z b
V (x − z) dG(z) + c(b)V ′ (x)

inf λV (x − b)(1 − G(b)) + λ
b∈[0,∞] 0
− (λ + δ)V (x) = 0 ,
 Rb 
where c(b) = λµ(η − θ) + λ(1 + θ) b(1 − G(b)) + 0 z dG(z) .
Consider again
Z b
gx (b) = λV (x − b)(1 − G(b)) + λ V (x − z) dG(z) + c(b)V ′ (x) − (λ + δ)V (x)
0
Z b
= λ (1 − G(u))[(1 + θ)V ′ (x) − V ′ (x − u)] du − λµ(θ − η)V ′ (x) .
0

Then gx (b) is differentiable at all points where V (y) is differentiable with y = x − b.

Derivation with respect to b yields

gx′ (b) = −λ(1 − G(b)) V ′ (x − b) − (1 + θ)V ′ (x) .

 

1
It is clear, that the optimal strategy is b∗ (x) = ∞ in the case V ′ (x) ≤ − 1+θ . If
′ 1 ∗
V (x) > − 1+θ it must hold 0 < b (x) ≤ x. The following relationship holds

V ′ ({x − b∗ (x)}+) ≤ (1 + θ)V ′ (x) ≤ V ′ ({x − b∗ (x)}−) .

In particular, if we choose an ε > 0 very small, we obtain for x large enough

Z ∞
h(x) = λ V (x − z) dG(z) − (λ + δ)V (x) < ε .
0

1
If V ′ (x) ≤ − 1+θ , we conclude h(x) + cV ′ (x) < 0, which is a contradiction to (2.14).
Thus, V (x) tends to zero and b∗ (x) < ∞ for x large enough.
′ 

2.2.3 The special case δ = 0

We consider now the special case δ = 0. We assume the net profit condition η > 0. In
this case the Hamilton–Jacobi–Bellman equation has the form
Z ∞
inf λ V (x − r(z, b)) dG(z) + c(b)V ′ (x) − λV (x) = 0 . (2.19)
b∈[0,b̃] 0

It is immediately clear, that a strategy b with c(b) ≤ 0 can not be optimal for some
x ∈ R+ . Indeed, a level with c(b) ≤ 0 cannot be crossed, implying that V (0) = ∞. But
the strategy bt = b̃ yields a finite value.

54
 2.2 The Classical Risk Model

We transform the HJB equation using the fact, that for the minimiser b we have
c(b) > 0. Define for this purpose

s(x, b) = inf{z : r(z, b) > x}.

Then we can write the equation (2.19) as follows:

n Z s(x,b)


inf λ V x − r(z, b) dG(z) + λ 1 − G(s(x, b)) V (0) − λV (x)
b∈[0,b̃] 0
Z ∞
o
+λ r(z, b) dG(z) + c(b)V ′ (x) − λ 1 − G(s(x, b)) x = 0
s(x,b)

With Fubini’s Theorem we obtain:

Z s(x,b)



λ V x − r(z, b) dG(z) + λ 1 − G(s(x, b)) V (0) − λV (x)
0
Z x
1 − G(s(y, b)) V ′ (x − y) dy


= −λ
0

and
Z ∞

Z ∞
λ r(z, b) dG(z) − λ 1 − G(s(x, b)) x = λ 1 − G(s(y, b)) dy.
s(x,b) x

Because for each x there exists a b, we can write for the optimal b(x):
Z x
′
1 − G(s(y, b(x))) V ′ (x − y) dy


c(b(x))V (x) = λ
0
Z ∞
−λ 1 − G(s(y, b)) dy , (2.20)
x

and (2.19) transforms to

λ h x
Z
′
1 − G(s(y, b)) f ′ (x − y) dy


f (x) = sup
b∈(b0 ,b̃] c(b) 0
Z ∞ i
− 1 − G(s(y, b)) dy (2.21)
x

From Theorem 2.2.3 we know that the solution to (2.19) and therefore to (2.21) is unique.
For x = 0 we obtain
λE[r(Z, b)]
V ′ (0) = sup .
b∈(b0 ,b̃] c(b)
1
In particular, for an expected value principle (2.18) we find V ′ (0) = − 1+η .

55
2 Optimal Control of Capital Injections by Reinsurance

Proposition 2.2.9
There is a unique solution f (x) to (2.19), x ≥ 0, with f (∞) = 0.
Proof: Define an operator F , acting on negative functions w(x) by
λ h x
Z


F (w(x)) = sup 1 − G(s(y, b)) w(x − y) dy
b∈(b0 ,b̃] c(b) 0
Z ∞ i
− 1 − G(s(y, b)) dy (2.22)
x

We have already seen, that the value function V (x) solves the equation (2.19). Let
now w0 (x) := (V b̃ )′ (x), the solution of Example 2.2.1 with bt = b̃. Define recursively
wn (x) = F (wn−1 (x)). We show at first, that the sequence wn is monotone increasing in
n. It is clear, that w0 (x) ≤ w1 (x) because (V b̃ )′ (x) solves the right hand side of (2.20)
with b = b̃ instead of the sup. Assume wn−1 (x) ≤ wn (x). Because the right side of (2.22)
is continuous in b, there is a maximum point bn ∈ [b0 , b̃] for which wn (x) = F (wn−1 (x))
attains its maximum. So we have
wn+1 (x) − wn (x) = F (wn (x)) − F (wn−1 (x))
λ h x
Z


= F (wn (x)) − 1 − G(s(y, bn )) wn−1 (x − y) dy
c(bn ) 0
Z ∞ i
+ 1 − G(s(y, bn )) dy
x
λ h x
Z i


≥ 1 − G(s(y, bn )) [wn (x − y) − wn−1 (x − y)] dy
c(bn ) 0
≥ 0.
So the sequence wn (x) is increasing in n and wn (x) < 0. It means we obtain that
w(x) = lim wn (x) exists pointwise. By the proposition of Lebesgue we have then
n→∞
Z x Z x



lim 1 − G(s(y, b)) wn (x − y)dy = 1 − G(s(y, b)) w(x − y) dy
n→∞ 0 0
for all x and b.
Let b be a maximal point of F (w(x)). Then we have
λ h x
Z


wn (x) = 1 − G(s(y, bn )) wn−1 (x − y) dy
c(bn ) 0
Z ∞ i
− 1 − G(s(y, bn ))dy
x
λ h x
Z


≥ 1 − G(s(y, b)) wn−1 (x − y) dy
c(b) 0
Z ∞ i
− 1 − G(s(y, b)) dy ,
x

56
 2.2 The Classical Risk Model

which means w(x) ≥ F (g(x)). On the other hand wn (x) are increasing in n, i.e. wn (x) ≤
w(x):

λ h x
Z


wn (x) = 1 − G(s(y, bn )) wn−1 (x − y) dy
c(bn ) 0
Z ∞ i
− 1 − G(s(y, bn )) dy
x
λ h x
Z


≤ 1 − G(s(y, bn )) w(x − y) dy
c(bn ) 0
Z ∞ i
− 1 − G(s(y, bn )) dy
x
λ h x
Z


≤ 1 − G(s(y, b)) w(x − y) dy
c(b) 0
Z ∞ i
− 1 − G(s(y, b)) dy ,
x

which means w(x) ≤ F (w(x)). We have therefore w(x) = F (w(x)), and w(x) is contin-
uous.
Because wn (x) is increasing, we can define
Z ∞ Z ∞
f (x) = − w(y) dy ≤ − w0 (y) dy = V b̃ (x) .
x x

f (x) fulfils (2.22) with f (∞) = 0. f (x) is also decreasing, continuously differentiable
and bounded 0 < f (x) ≤ V b̃ (x).

Suppose now, that f1 (x) and f2 (x) are solutions to (2.19) with f1 (∞) = f2 (∞) = 0.
Denote further by gi (x) = fi′ (x) the derivatives and by bi (x) the value, for which the
minimum is obtained. Choose now some x∗ > 0. Because the right-hand side of Equation
(2.21) is continuous in b and tends to infinity as c(b) tends to zero, we conclude, that
c(bi (x)) is bounded away from zero on [0, x∗ ]. Let x1 = inf{c(b1 (x)) ∧ c(b2 (x)) : 0 ≤
x ≤ x∗ } and xn = nx1 ∧ x∗ . W.l.o.g. we assume, that x1 ≤ x∗ . Suppose we have
already proved f1 (x) = f2 (x) on the interval [0, xn ]. Then for x ∈ [xn , xn+1 ], with
m= sup |g1 (x) − g2 (x)| it holds
xn ≤x≤xn+1

g1 (x) − g2 (x) = F (g1 (x)) − F (g2 (x))

Z x
λ 
≤ g1 (z) − g2 (z) [1 − G(s(x − z))] dz
c(b2 (x)) xn
λmx1 m
≤ ≤
c(b2 (x)) 2

57
2 Optimal Control of Capital Injections by Reinsurance

Reversing the roles of g1 (x) and g2 (x), it follows that |g1 (x) − g2 (x)| ≤ m
2 . This is only
possible for all x ∈ [xn , xn+1 ] if m = 0. this shows that f1 (x) = f2 (x) on [0, xn+1 ]. So
f1 (x) = f2 (x) on [0, x∗ ]. Because x∗ was arbitrary, uniqueness follows. 

Next we illustrate the result by a couple of examples.

Examples
Note, that because we know, that the value function is differentiable we will use the
modified HJB equation
Z x Z ∞
V ′ (z) 1 − G((x − z)/b) dz + c(b)V ′ (x) + λ


inf −λ 1 − G((x + z)/b) dz = 0 .
b∈[0,b̃] 0 x
(2.23)
for numerical calculation of the value function. Furthermore, all the considerations
concerning the function V b̃ (x) in the case δ > 0 hold also in the case δ = 0.
In the following examples we will use the expected value principle, i.e. c(b) = −λµ(θ −
η) + λ(1 + θ)E[r(Z, b)].

Example 2.2.10 (Proportional Reinsurance and Zi ∼Exp( µ1 ).)

For the exponentially distributed claim sizes we obtain as the function V 1 (x):
(µ η
1 η exp(− µ(1+η) · x) : x ≥ 0 ,
V (x) = µ
η −x : x<0.

η(c+λµ)
It is easy to verify, that V 1 (x) is the value function if θ ≥ λµ . Assume now θ <
η(c+λµ)
λµ . Equation (2.23) becomes
Z x x−z x
inf −λ V ′ (z)e− µb dz + c(b)V ′ (x) + λµbe− µb = 0 .
b∈[0,b̃] 0

The above equation differs very little from the HJB equation for the ruin probability as a
function of initial capital, compare Schmidli [70, p. 47]. In the case of ruin probability we
−x −x
have λµe µb instead of the term λµbe µb in our case. But this difference is responsible
for the behavior of the optimal strategy. While in the ruin probability case the optimal
strategy jumps down from the 1-level, in our case it falls down continuously. For the
parameters µ = λ = 1, θ = 0.5 and η = 0.3 we can calculate the x̂ := inf{x ∈ R+ :
b∗ (x) < 1} exactly and obtain x̂ = 0.935. V (x) and the optimal strategy are plotted in
1
Figures 2.12 and 2.13 respectively. Note, that because V ′ (0) = − 1+η we can calculate
the initial value explicitly: V (0) = 3.11. 

58
 2.2 The Classical Risk Model

Example 2.2.11 (Proportional Reinsurance and Zi ∼Pareto(2, µ).)

Consider the case, where the claim sizes are Pareto(2, µ)–distributed, i.e. 1 − G(x) =
µ2
(µ+x)2
. In this case we are not able to give a closed expression for the function V 1 (x),
but we can calculate it numerically with the method of Gerber and Shiu. The value
function and the optimal strategy are plotted in Figures 2.14 and 2.15 respectively. The
initial value is V (0) = 18.27. We see, that the initial value V (0) is considerably larger,
than in the case of exponential distribution. 

59
2 Optimal Control of Capital Injections by Reinsurance

0 5 10 15 20

Figure 2.12: Numerically calculated value function for Zi ∼Exp( µ1 ).

1,0

0,9

0,8

0,7

0 1 2 3 4 5 6 7 8 9

Figure 2.13: Optimal strategy for Zi ∼Exp( µ1 ).

60
 2.2 The Classical Risk Model

30

20

10

0 100 200 300

Figure 2.14: Numerically calculated value function for Zi ∼Pareto(2, µ).

1,0

0,9

0,8

0 20 40 60 80

Figure 2.15: Optimal strategy for Zi ∼Pareto(2, µ).

61
2 Optimal Control of Capital Injections by Reinsurance

62
3 Optimal Control of Capital Injections by
Reinsurance with Riskless Rate of
Interest
In this chapter we consider the case, where the reinsurer is allowed to invest his
positive surplus into a riskless asset with constant interest rate. In the case of diffusion
approximation we consider only the case of proportional reinsurance.

3.1 Proportional Reinsurance for a Diffusion Approximation

Consider the surplus process of an insurance company, where the time horizon is
infinite
Nt
X
Ct = x + ct − Zi .
i=1

We have a probability space (Ω, F, P) and on this space there is the Poisson process {Nt }
with intensity λ > 0 and a sequence of iid random variables {Zi }i∈N . Zi are assumed to
have a distribution G with µ = E[Zi ], µ2 = E[Zi2 ] < ∞ and be independent of {Nt }. The
premium income of the insurer is c = (1 + η)λµ for some η > 0. Further the insurer can
buy proportional reinsurance. That is the insurer has to choose a retention level b ∈ [0, 1]
and the reinsurer carries (1 − b)Zi from each claim Zi . The premium rate remaining to
the insurer calculated by an expected value principle is c(b) = −λµ(θ − η) + λµb(1 + θ),
where θ is the safety loading of the reinsurer. In order to avoid, that the insurer can
make a riskless profit, buying full reinsurance and still receiving a positive premium, we
assume θ > η. As an extension the insurer can change his retention level continuously.
A diffusion approximation to the above classical risk model fulfils then the differential
equation
dXtB = λµ bt θ − (θ − η) dt + bt λµ2 dWt ,
   p

where B = {bt } is some admissible reinsurance strategy with bt ∈ [0, 1]. We call a
reinsurance strategy admissible if it is adapted and cadlag; the set of all reinsurance
strategies we denote by U .
Usually it is supposed, that because of inflation, the original risk process is discounted,
and the riskless interest rate is set to 0. Now we offset this assumption and allow the

63
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

insurer to earn interest on positive surplus with a constant force of interest. That is, the
controlled process with capital injections Y B = {YtB } is then
dXtB,Y,m = mXtB,Y,m + λµ bt θ − (θ − η) dt + bt λµ2 dWt + dYtB ,
   p
(3.1)
for a riskless interest rate m > 0.
We want to measure the risk, connected to R ∞some reinsurance strategy B, by expected
discounted capital injections V B (x) := Ex [ 0 e−δt dYt ]. Our goal is to find the value
function by minimising V B (x) over all admissible reinsurance strategies
V (x) := inf V B (x) .
B∈U

Here δ expresses just the investing preferences of the insurer and is not a financial
discounting factor. In the case δ < m the insurer prefers to invest all his money into a
riskless asset, so that it makes no sense for him to carry on an insurance business. But
here we do not restrict to δ ≥ m.
It is clear, that because of the discounting, the value function V (x) is decreasing. In
particular we obtain for the constant strategy B ≡ 0 before the ruin occurs
Z t
Xt0,m = x − λµ(θ − η)t + m Xs0,m ds = (x − λµ(θ − η)m−1 )emt + λµ(θ − η)m−1 .
0

Since it holdsXt0,m> 0 for all t if x ≥ λµ(θ − η), we conclude {Yt0 } ≡ 0 and accordingly
V (x) = 0 for x ≥ λµ(θ − η). Thus we have to consider only 0 ≤ x < λµ(θ − η)m−1 .
Remark 3.1.1
Let {Xt } be a diffusion process with values in R, fulfilling the stochastic differential
equation
dXt = a(Xt ) dt + σ(Xt ) dWt ,
where {Wt } is a standard Brownian motion and a, σ are functions, such that the above
equation has a unique strong solution. The reflected process fulfils then
dXtY = a(XtY ) dt + σ(XtY ) dWt + dYt ,
whereas Y is the local time of the process at zero.
We find in
R ∞Remark 2.1.3, that the corresponding return function
V (x) = Ex [ 0 e−δt dYt ] solves the differential equation
σ 2 (x) ′′
V (x) + a(x)V ′ (x) − δV (x) = 0 (3.2)
2
for x ≥ 0 and fulfils V ′ (0) = −1, lim V (x) = 0. From Subsection 2.1.1 we know, that
x→∞
every solution f (x) to the above differential equation, vanishing at infinity, has the form
 ∞ −δt
Z
′

f (x) = f (0)Ex e dYt .
0

64
 3.1 Proportional Reinsurance for a Diffusion Approximation

K1 0 1 2 3 4 5 6 7 8
x

0.5 (x) (solid line), V 0.5 (x) (dotted line) and V 1 (x) dashed
Figure 3.1: Return functions V0.5 0.8
line.

Now equipped with the knowledge, how to calculate the return function for a given
reinsurance strategy B, we illustrate the method by an example.

Example 3.1.2
Consider now a constant strategy B ≡ b ∈ [0, 1]. The process {Xtb,Y,m } solves the
stochastic differential equation

dXtb,Y,m = mXtb,Y,m + λµ bθ − (θ − η) dt + b λµ2 dWt + dYtb .

   p

Due to Remark 3.1.1 the corresponding return function V b (x) solves then the differential
equation

b2 λµ2 ′′
f (x) + mx + λµ(bθ − (θ − η) f ′ (x) − δf (x) = 0 .


2

With the power series method, see Remark E.1.5 p. 179, we find that solutions to the

65
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

above differential equation are given by

n
Q δ
 ∞ (m + 2 − 2k)  
X k=1 2m n
C1 1 + (x + λµ(bθ − θ + η)m−1 )2n
n=1
(2n)! λµ2 b2
n
Q δ
∞
X (m + 1 − 2k) 
k=1 2m n
+ C2 (x + λµ(bθ − θ + η)m−1 )2n+1
n=1
(2n + 1)! λµ2 b2

−1
+ x + λµ(bθ − θ + η)m .

Using the initial conditions lim V b (x) = 0 and (V b )′ (0) = −1 one can calculate the
x→∞
coefficients C1 and C2 .
Let for example b = 0.5, λ = µ = 1, µ2 = 2, θ = 0.5, η = 0.3, δ = 0.04 and m = 0.03.
Then we obtain C1 = 4.084921164 and C2 = −1.947322694. Letting the parameter
θ = 0.8 yields C1 = 0.9686572638 and C2 = −0.4617685869. The return functions for
the constant strategy B ≡ 0.5 V0.50.5 (x) for θ = 0.5 (solid line), V 0.5 (x) for θ = 0.8 (dotted
0.8
line) and the return function for B ≡ 1 V 1 (x) (dashed line) are plotted in Figure 3.1.
We see, that for θ = 0.5 the return function corresponding to B ≡ 0.5 lies below V 1 (x);
and for θ = 0.8 above V 1 (x). We will see later, that for some θ it holds V 1 (x) = V (x)
on some intervals. 

We have found out, that the value function V (x) fulfils V (x) = 0 for x ≥ λµ(θ − η)m−1 .
Consider now x < λµ(θ − η)m−1 . For such x the Hamilton–Jacobi–Bellman equation
(compare Section 2.1) is given by

1
inf λµ2 b2 V ′′ (x) + {mx − λµ(θ − η) + λµbθ}V ′ (x) − δV (x) = 0 . (3.3)
b∈[0,1] 2

We abandon the explicit derivation of the HJB equation since it is very similar to the
motivation in Subsection 2.1.2. Note that if the value function exists, is twice continu-
ously differentiable and solves the HJB equation above, it must be convex. In fact, if we
choose b̂ = 1 − ηθ − λµθ
mx
for some x ∈ [0, λµ(θ − η)m−1 ) (note that b̂ ∈ [0, 1]) we obtain

1
λµ2 b̂2 V ′′ (x) − δV (x) ≥ 0 .
2

We make the ansatz

κ
V (x) = C m−1 λµ(θ − η) − x

66
 3.1 Proportional Reinsurance for a Diffusion Approximation

for some C, κ > 0. Then (3.3) reads

2

κ−2
κ(κ − 1) m−1 λµ(θ − η) − x
1
0 = inf λµ2 b
b∈[0,1] 2

κ−1
− {mx + λµ[bθ − (θ − η)]}κ m−1 λµ(θ − η) − x

κ
− δ m−1 λµ(θ − η) − x .

The optimal b is then given by

θµ(m−1 λµ(θ − η) − x)
b(x) = , (3.4)
µ2 (κ − 1)

provided κ > 1 and b(x) ≤ 1; i.e., x is close enough to m−1 λµ(θ−η). For κ < 1 we should
choose b(x) = 0. Then we would obtain that V is concave, which is a contradiction. If
κ = 1 then b(x) = 1 has to be chosen and our ansatz would not lead to a solution. If
b(x) > 1 no reinsurance has to be chosen.
Plugging in the optimal b(x) and dividing by (m−1 λµ(θ − η) − x)κ we find

λκθ 2 µ2
mκ − −δ =0. (3.5)
2µ2 (κ − 1)

Solving for κ yields the solution

q
δµ2 + mµ2 + 21 λθ 2 µ2 + (δµ2 + mµ2 + 21 λθ 2 µ2 )2 − 4mµ22 δ
κ= . (3.6)
2mµ2

Note that the other solution is smaller than one.

Remark 3.1.3
Let X ∗ = X b(x),Y,m with initial value 0 ≤ x < λµ(θ − η)m−1 , where the reinsurance
strategy b(x) is given in (3.4). Consider the process Zt = m−1 λµ(θ − η) − Xt∗ before the
process X ∗ ruins. Then
√  λµ2 θ 2
λµ2 θµ 
dZt = − Zt dWt − − m Zt dt .
µ2 (κ − 1) µ2 (κ − 1)

That means that {Zt } is a geometric Brownian motion. Taking the logarithm gives
√
λµ2 θµ  λµ2 θ 2 (2κ − 1) 
d(log(Zt )) = − dWt + m − dt .
µ2 (κ − 1) 2µ2 (κ − 1)2

In particular, the surplus X ∗ will never reach the value m−1 λµ(θ − η), where full rein-
surance would be bought.

67
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

The considerations we used in deriving (3.4) are of heuristic nature. Hence, it remains
to prove the verification theorem.

Theorem 3.1.4
n (Verification theorem)o
Define x̃ := m−1 λµ(θ − η) − µ2 (κ−1)
θµ ∨ 0. Then the strategy


0
 : x ≥ m−1 λµ(θ − η) ,
b∗ (x) = b(x) : x̃ < x < m−1 λµ(θ − η) ,

1 : x ≤ x̃ ,


where b(x) is given in (3.4), is an optimal reinsurance strategy. The function f (x), given
by 


 0 : x ≥ m−1 λµ(θ − η) ,
: x̃ ≤ x < m−1 λµ(θ − η) ,

f (x)
2
f (x) =


 f1 (x) : 0 < x < x̃ ,

f (0) − x : x ≤ 0 ,

with

n
Q δ
 ∞ (m + 2 − 2k)  
X k=1 2m n
f1 (x) = C1 1 + (x + λµηm−1 )2n
n=1
(2n)! λµ2
n
Q δ
 ∞ (m + 1 − 2k)  
−1
X k=1 2m n
+ C2 x + λµηm + (x + λµηm−1 )2n+1
(2n + 1)! λµ2
n=1

and

κ
f2 (x) = C3 r −1 λµ(θ − η) − x ,

where κ is given in (3.6), is twice continuously differentiable, solves the HJB equation
(3.3) and f (x) = V (x). If x̃ > 0 the coefficients C1 , C2 and C3 are uniquely given by
the system of equations

f1′ (0) = −1 ,
f1′ (x̃) = f2′ (x̃) ,
f1′′ (x̃) = f2′′ (x̃) ;

if x̃ ≤ 0, C3 is given by f2′ (0) = −1.

68
 3.1 Proportional Reinsurance for a Diffusion Approximation

Proof: We show at first for each interval, that the strategy b∗ (x) yields the function
f (x) and that the function f (x) solves the HJB equation. In the second part we prove
f (x) = V (x).

1) Consider at first the interval [λµ(θ − η)m−1 , ∞). We have already seen, that the
strategy B = 0 yields V 0 (x) = 0 for x ≥ λµ(θ − η)m−1 .

2) On the interval [x̃, λµ(θ − η)m−1 ) we let {Xt∗ } be the underlying process and V ∗ (x)
the return function for the strategy b∗ (x). Let further τ ∗ = inf{t ≥ 0 : Xt∗ < x̃}, i.e. τ ∗
is the ruin time of the process Xt∗ − x̃. From Remark 3.1.1 we know, that the capital
injection process Y ∗ for the process X ∗ is given through the local time

Yt∗ = − min{ inf Xs∗ , 0} .

0≤s≤t

Thus, we can write because the process Xt∗ has continuous paths
∗
V ∗ (x) = V ∗ (x̃)Ex [e−δτ ] .

Note, that λµ(θ − η)m−1 − x̃ ≥ 0. From Remark 3.1.3 we know, that it holds Xt∗ =
λµ(θ − η)m−1 − Zt , where Zt
√  λµ2 θ 2
λµ2 θµ 
dZt = − Zt dWt − − m Zt dt ,
µ2 (κ − 1) µ2 (κ − 1)
i.e. a geometric Brownian motion. This implies

τ ∗ = inf{t ≥ 0 : log λµ(θ − η)m−1 − x̃ − log(Zt ) < 0} ,

i.e. τ ∗ is the ruin time of a Brownian motion W̃t

√
−1 λµ2 θµ
− x̃ − log λµ(θ − η)m−1 − x +



W̃t = log λµ(θ − η)m Wt
µ2 (κ − 1)
− αt
2 θ 2 (2κ−1)
 
with α = m − λµ 2µ2 (κ−1) 2 . Using the change of measure technique, see Chapter C,
and Remark 2.1.3 we obtain
∗
n
β
−β o
Ex [e−δτ ] = exp log λµ(θ − η)m−1 − x + log λµ(θ − η)m−1 − x̃

β
−β
= λµ(θ − η)m−1 − x · λµ(θ − η)m−1 − x̃ ,

where β is given by
s
 α 2 µ2 (κ − 1)2 α
β= 2 2
µ 2 (κ − 1)2 + 2δ − 2 2 µ2 (κ − 1)2 .
λθ µ λµ2 θ 2 λθ µ

69
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

λµ2 θ 2 (2u−1)
 
Define the functions χ(u) := m − 2µ2 (u−1)2 and
s
 χ(u) 2 µ2 (u − 1)2 χ(u)
γ(u) := 2 2
µ 2 (u − 1)2 + 2δ 2 2
− 2 2 µ2 (u − 1)2 .
λθ µ λµ θ λθ µ

It is easy to verify, that the function γ(u) has a unique fixed point u2 = γ(u2 ) and it
holds
q
δµ2 + mµ2 + 21 λθ 2 µ2 + (δµ2 + mµ2 + 21 λθ 2 µ2 )2 − 4mµ22 δ
γ(u2 ) = u2 = .
2mµ2

From (3.6) we then obtain u2 = κ and according

κ to this β = γ(κ) = κ.
∗ −1
Thus V (x) = C3 λµ(θ − η)m − x = f3 (x). Plugging in f3 (x) into the HJB
equation yields b(x), given in (3.4).

3) Now we assume x̃ > 0, otherwise it does not make any sense to consider the interval
2δ
µ2 (κ−1) 1−
[0, x̃]. Note, that λµm−1 (θ − η) − θµ ≥ 0 holds iff η ≤ θ 2(1−mκδ ) .
mκ
That the constant strategy B ≡ 1 yields the function f1 (x) we have already seen in
Example 3.1.2. It remains to show, that plugging in f1 (x) into the HJB equation we
obtain b∗ (x) = 1 for x ∈ [0, x̃]. Note, that the coefficients C1 , C2 and C3 are such, that
b∗ (x̃) = 1. Consider the HJB Equation (3.3) with the function f1 (x). It is easy to see,
that the optimal b∗ (x) is given by

−µθf1′ (x)
b∗ (x) = ∧ 1.
µ2 f1′′ (x)

−µθf ′ (x)
We need to show, that g(x) := µ2 f ′′1(x) ≥ 1 for all x ∈ [0, x̃].
1
Assume for the moment that there exists some x ∈ [0, x̃] with g(x) < 1. Because
g(x̃) = 1 and g is continuous, there exist some interval [a, b] ⊂ [0, x̃] and x∗ ∈ [a, b] such,
that g′ (x) > 0 on [a, b] and g(x∗ ) < 1. So we know, that

θµ θµ f ′ (x∗ )f ′′′ (x∗ )

g′ (x∗ ) = − + > 0.
µ2 µ2 f ′′ (x∗ )2

It follows readily

f1′ (x∗ )f1′′′ (x∗ ) −θµf1′ (x∗ ) −µ2 f1′′′ (x∗ ) ∗ −µ2 f1′′′ (x∗ )
1< = · = g(x ) · .
f1′′ (x∗ )2 µ2 f1′′ (x∗ ) θµf1′′ (x∗ ) θµf1′′(x∗ )

−µ2 f1′′′ (x∗ )

Because g(x∗ ) was assumed to be smaller than one, θµf1′′ (x∗ ) has to be larger than one.

70
 3.1 Proportional Reinsurance for a Diffusion Approximation

The function f1 (x) is smooth. From Example 3.1.2 we know, that f1 (x) fulfils the
differential equation

λµ2 ′′
f1 (x) + mx + λµη f1′ (x) − δf1 (x) = 0 ,


2

from which we obtain the following representation:

λµ2 ′′′ ∗
f (x ) + {mx∗ + λµη}f1′′ (x∗ ) − (δ − m)f1′ (x∗ ) = 0 .
2 1

Rearranging the terms, dividing by f1′′ (x∗ ) and using (3.5) yields

−µ2 f1′′′ (x∗ ) 2f1′ (x∗ ) 2

= −(δ − m) + {mx∗ + λµη}
θµf1′′ (x∗ ) ′′ ∗
λθµf1 (x ) λθµ
2µ2 2
< |δ − m| 2 2 + {mx̃ + λµη}
λθ µ λθµ
2µ2 2 n mµ2 (κ − 1) o
= |δ − m| 2 2 + λµθ −
λθ µ λθµ θµ
κ mκ
= |δ − m| +2−
(κ − 1)(mκ − δ) mκ − δ
κ−2
= <1,
κ−1

which is a contradiction.
Now we will show f (x) = V (x). Consider an arbitrary admissible reinsurance strategy
B = {bt } and denote X̂t = X B,Y,m . Then X̂ is given by the differential equation

dX̂t = mX̂t + λµ bt θ − (θ − η) dt + bt λµ2 dWt + dYtB .

   p

Because f (x) is two times continuously differentiable, X̂t ≥ 0, f ′ (0) = −1 and using

71
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

(3.3) we apply Ito’s formula on the function e−δt f (x) and obtain:
Z t
−δt
e f (X̂t ) = f (x) + e−δs f ′ (X̂s ) dYsB
0
Z t n o
+ e−δs Ds,B f (X̂s ) − δf (X̂s ) ds
0
Z t
e−δs f ′ (X̂s )bs λµ2 dWs
p
+
0
Z t
≥ f (x) + e−δs f ′ (X̂s ) dYsB
0
Z t
e−δs f ′ (X̂s )bs λµ2 dWs
p
+
0
Z t
= f (x) − e−δs dYsB
0
Z t
e−δs f ′ (X̂s )bs λµ2 dWs ,
p
+
0

where
λµ2 b2s ′′ n o
Ds,B f (x) = f (x) + mx + λµ(bs θ − θ + η) f ′ (x)
2
is the infinitesimal generator of the process Xtbs ,m . Because the derivative of the value
function is bounded, we can conclude, that the stochastic integral is a martingale with
zero-expectation. Thus applying the expectations on the both sides of the above in-
equality we have
hZ t i
−δt
e−δs dYsB .
 
f (x) ≤ e Ex f (X̂t ) + Ex
0

Because X̂t ≥ 0 and f is decreasing, we get 0 ≤ f (X̂t ) ≤ f (0), from which it follows
with monotone convergence
hZ ∞ i
f (x) ≤ Ex e−δs dYsB .
0

Since the strategy B was arbitrary, this proves our claim f (x) = V (x). 

Example 3.1.5
Consider now the parameters η = 0.3, λ = µ = 1, µ2 = 2 and δ = 0.04. In Figure 3.2 we
see the curve, which describes the dependence of x̃ on parameters θ and m. The pairs
(θ, m) on the curve yield x̃ = 0; the pairs to the left from the curve yield x̃ < 0 and
finally the pairs to the right yield x̃ > 0. Note, that for example for θ = 0.5 we would

72
 3.1 Proportional Reinsurance for a Diffusion Approximation

µ2 (κ−1)
Figure 3.2: x̃ = m−1 λµ(θ − η) − θµ = 0 in dependence from m and θ.

n o
obtain x̃ = m−1 λµ(θ − η) − µ2 (κ−1)
θµ < 0 for all m ≤ δ, so that V (x) = f2 (x) in this
case. We also see in Figure 3.1, that for θ = 0.5 even the return function corresponding
to the constant strategy B ≡ 0.5 (solid line) lies below f1 (x) (dashed line). Choose
θ = 0.8 and m = 0.03. It is easy to verify, that κ = 7.49 and the optimal strategy on
[x̃, λµ(θ − η)m−1 ] and value function are given by:

13.3 − 0.8x
b(x) =
12.98

 ∞ 
X 1.33 · · · (1.33 − 2n + 2)
f1 (x) = 64.28 1 + 0.03n (x + 10)2n
n=1
(2n)!
 ∞ 
X (1.33 − 1) · · · (1.33 − 2n + 1)
− 15.32 x + 10 + 0.03n (x + 10)2n+1
(2n + 1)!
n=1
f2 (x) = 1.575819495 · 10−9 (6.7 − x)7.49 .

The optimal strategy is mapped in Figures 3.4. In Figure 3.3 we can see the value
function, composed of 4 functions. The dashed line corresponds to f (0) − x, the dotted
line to f1 (x), the solid line to f2 (x) and the line with wide dots to 0. 

73
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

Figure 3.3: Value function for the proportional reinsurance V (x).

Figure 3.4: Optimal strategy b∗ (x).

74
 3.2 The Classical Risk Model

3.2 The Classical Risk Model

In this section we consider the classical risk model. The number of claims is assumed
to follow a Poisson process {Nt } with intensity λ. Claim sizes Z1 , Z2 , . . . are independent
of {Nt }, positive and iid with distribution function G(z) and first moment µ. By Z we
denote a generic random variable with the same distribution as Zi . The aggregate claim
process is given by
Nt
X
Xt = x + ct − Zi ,
i=1
where x is the initial capital, c = λµ(1 + η) is the premium rate and η is the safety
loading.
The insurer can buy reinsurance. Choosing the level b ∈ [0, b̃], the insurer pays
r(Zi , b) for a claim of size Zi . Here b = 0 means full reinsurance, b = b̃ ∈ (0, ∞] means
no reinsurance. For example, for proportional reinsurance r(Z, b) = bZ and b ∈ [0, 1].
For excess of loss reinsurance we obtain r(Z, b) = min{Z, b} and b ∈ [0, ∞].
For the reinsurance cover the insurer pays a premium at rate c − c(b). That is,
the premium rate left for the cedent is c(b). For simplicity we assume that r(z, b) is
continuous and increasing in both z and b and that c(b) is continuous and increasing
with c(0) < 0 and c(b̃) = c. The assumption c(0) < 0 is needed in order that the problem
below is not trivial. If the reinsurer uses an expected value principle with safety loading
θ, we get
c(b) = c − (1 + θ)λE[Z − r(Z, b)] = (1 + θ)λE[r(Z, b)] − (θ − η)λµ .
In order that c(0) < 0 we assume θ > η. The insurer can choose the level bt at any time
point t. Because no information on the future can be used the process {bt } is assumed
to be cadlag and adapted. The surplus of the insurer including reinsurance has then the
form
Z t Nt
X
XtB = x + c(bs ) ds − r(Zi , bTi − ) .
0 i=1
To prevent that the surplus process becomes negative, the insurer has to inject addi-
tional capital. We denote the accumulated capital injections until time t by {YtB }. The
surplus with capital injections has therefore the form
Z t Nt
B,Y
X
Xt =x+ c(bs ) ds − r(Zi , bTi − ) + YtB .
0 i=1
Assume in addition to this concept, that the insurer can invest his money, if his surplus
is positive, with a fixed rate of interest m > 0. I.e. his surplus has to fulfil
Z t Nt
B,m,Y
X
c(bs ) + mXsB,m,Y ds − r(Zi , bTi − ) + YtB .


Xt =x+
0 i=1

75
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

R∞
We are interested in the minimal value V (x) = inf E[ 0 e−δt dYtB ], where δ is a
B∈U
discounting factor. The discounting expresses the investment preferences of the company
holders. If m ≥ δ, there would be no sense in holding an insurance company, because
investing money in a riskless asset with interest rate m would be more profitable. Here
we do not assume δ > m, although m ≥ δ seems to be economically sinless. We also
abandon the assumption η > 0 for the safety loading η.
Note, that our process remains deterministic before the first claim occurs. Let τ
denote the ruin time of the process X B,m . Then we have for t < τ
Z t Z t
XtB,m =x+ c(bs ) + m XsB,m ds ,
0 0

so that we can write

Z t 
XtB,m = c(bs )e−ms ds + x emt . (3.7)
0

Lemma 3.2.1
The function V (x) is decreasing with V (x) = 0 for x ≥ λµ(θ − η)m−1 . V (x) is Lipschitz
continuous with |V (x) − V (y)| ≤ |x − y|.

Proof: It is clear that V (x) is decreasing. Assume now x ≥ c(0) = λµ(θ − η)m−1 .
Consider the process with the constant strategy B ≡ 0. Then we obtain from (3.7)

Xt0,m = emt x + c(0)m−1 (1 − e−mt ) .



Because x ≥ c(0) = λµ(θ − η)m−1 it holds Xt0 ≥ 0 for all t ≥ 0. Thus for initial capital
x ≥ λµ(θ − η)m−1 it holds {Yt0 } ≡ 0 and accordingly 0 ≤ V (x) ≤ V 0 (x) = 0.
Let now x < λµ(θ − η)m−1 and choose z > x. Let further B = {bt } be a reinsurance
strategy for initial capital z such that V B (z) ≤ V (z) + ε. For initial capital x choose the
strategy B̃ (which is not optimal): inject the capital z − x and then follow the strategy
B. Thus
V (x) − V (z) ≤ V B̃ (x) − V B (z) + ε = z − x + ε .

Because ε is arbitrary we have |V (x) − V (z)| ≤ |x − z|, which proves the Lipschitz-
continuity. As a consequence, V (x) is absolutely continuous. 

We conjecture, that the value function V (x) solves the Hamilton–Jacobi–Bellman equa-
tion
Z ∞
inf λ V (x − r(z, b)) dG(z) + (c(b) + mx)V ′ (x) − (δ + λ)V (x) = 0 . (3.8)
b∈[0,b̃] 0

76
 3.2 The Classical Risk Model

Theorem 3.2.2
The function V (x) is continuously differentiable from the right and from the left at all
points where b0 (x) with c(b0 (x)) = −mx is not optimal. Its derivatives solve Equation
(3.8) with the interpretation (c(b) + mx)V ′ (x) = 0, if the derivatives do not exist.
Moreover, if there exists a b such that c(b) ≥ λE[r(Z, b)], then any decreasing positive
solution to (3.8) coincides with V (x).

Proof: Assume x > 0. Let h > 0 and b ∈ [0, b̃] be fixed. We can assume that emt (x +
c(b)m−1 )−c(b)m−1 ≥ 0, i.e. the ruin does not occur because of the premium payments to
the reinsurer. Let T1 be the time of the first claim and choose ε > 0. We further choose
n ∈ N such that 2(emh (x + c(b)m−1 ) − c(b)m−1 )/n < ε. For each k there is a strategy
k
B k = {bkt } such that V B (xk ) ≤ V (xk )+ε/2. For initial capital xk ≤ x < xk+1 we choose
the strategy B k . Thus V B (x) ≤ V B (xk ) ≤ V (xk )+ε/2 ≤ V (x)+(x−xk )+ε/2 < V (x)+ε
by Lemma 3.2.1. This shows that for each x ∈ [0, emh (x + c(b)m−1 ) − c(b)m−1 ] we can
find in a measurable way a strategy B̂(x) such that V B̂(x) (x) < V (x) + ε.
Consider now the strategy bt = b1I[t<T1 ∧h] + b̂t−(T1 ∧h) (XT1 ∧h )1I[t≥T1 ∧h] and define the
deterministic process kt := −c(b)m−1 + (x + c(b)m−1 )emt . Then we obtain by condition-
ing on FT1 ∧h and using (3.7)

V (x) ≤ V B (x)
h  
= Ex 1I[T1 ≤h] V B̂ kT1 − r(Z1 , b) e−δT1
i
+ 1I[T1 >h] V B̂ kh e−δh

Z hZ ∞
λe−(δ+λ)t V B̂ kt − r(z, b) dG(z) dt


=
0 0
Z ∞
+ e−δh V B̂ (kh )λe−λt dt
h
Z hZ ∞
e−(δ+λ)t V kt − r(z, b) dG(z) dt


≤ λ
0 0
+ e−(δ+λ)h V (kh ) + ε .

Because ε was arbitrary we can let it be equal to zero. Rearranging the terms and
dividing by h yields:

hZ ∞
1
Z
λe−(δ+λ)t V kt − r(z, b) dG(z) dt


0 ≤
h 0 0


V kh − V (x) −(δ+λ)h e−(δ+λ)h − 1
+ e + V (x) .
h h
(3.9)

77
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

RConsider a strategy D(h) = {dt (h)} with V D (x) ≤ V (x) + h2 . Let further a(t, h) =
t −ms
0 c(ds )e ds. Then we obtain in the same way as above
V emh {x + a(h, h)} − V (x) −(δ+λ)h e−(δ+λ) − 1


0 ≥ e + V (x)
h h
1 h ∞ −(δ+λ)t
Z Z
V emt {x + a(t, h)} − r(z, dt ) dG(z) dt − h .


λe
h 0 0
All the terms with exception of the first one converge. And for the first one we obtain
V emh {x + a(h, h)} − V (x) V emh {x + a(h, h)} − V (x)



=
h emh {x + a(h, h)} − x
emh {x + a(h, h)} − x
× .
h
The first term on the right hand side of the above equation has the values in the interval

[−1, 0], because
 of the Lipschitz continuity; the second one has his values in c(0) +
xm, c + xm . Thus there exists a sequence hn → 0 such that
V emhn {x + a(hn , hn )} − V (x) V emh {x + a(h, h)} − V (x)



lim = lim sup .
n→∞ hn h→0 h
Because of the Lipschitz continuity of the value function V (x) the above limit is finite.
mhn {x+a(h ,h )}−x
Assume now, that e hn
n n
goes to c(b̌) + mx as n goes to infinity. Using, that
the strategy D is cadlag, we obtain
V emh {x + a(hn , hn )} − V (x)

Z ∞


lim +λ V x − r(z, b̌) dG(z)
n→∞ hn 0
−(λ + δ)V (x) ≤ 0 .
Together with (3.9) and b = b̌ equality follows. The above limit does not depend on the
chosen subsequence, i.e. the limit
V − c(b̌)m−1 + {x + c(b̌)m−1 }emh − V (x)


lim .
h→0 h
exists. It means in particular, that in the case c(b̌) + mx > 0 V (x) is differentiable in x
from the right; in the case c(b) + mx < 0 from the left. In both cases (3.8) was shown
for the corresponding derivatives.
Consider now the initial capital ph = −c(b̌)m−1 1 − e−mh + xe−mh . It holds with the

same arguments as by the derivation of (3.9):

1 h ∞ −(δ+λ)t  mt
Z Z


λe V e {xe−mh + c(b̌)m−1 e−mh − e−mt } dG(z) dt
h 0 0
V (ph ) − V (x) e−(δ+λ)h − 1

+ + V ph ≥ 0 . (3.10)
h h

78
 3.2 The Classical Risk Model

Assume further, that V (x) is differentiable in x from the right, i.e. c(b̌) + mx > 0. For
all z < x there exists some b̌(z), where the minimum in (3.8) is taken. Suppose there is
a sequence xn → x with xn ≤ x and c(b̌n ) + mxn = c(b̌(xn )) + mxn > 0. By taking a
subsequence we can assume, that b̌n converges to some b∗ with c(b∗ ) + mx ≥ 0. Then
we can find a sequence hn such that (3.10) converges; (3.8) holds then by continuity.
If in addition to the above assumptions c(b∗ ) > 0 we can conclude, that V (x) is also
differentiable from the left.
If there is a sequence xn → x with xn ≤ x and c(b̌n ) + mxn < 0, choose hn =
− c(b̌x−x n
)+mx
. If hn → 0, then differentiability from the left follows and Equation (3.8)
n n
holds. If hn does not converge to 0, then c(b̌n )+mxn → 0 and (3.10) holds with minimum
taken at b̌ with c(b̌) + mx = 0. In the case c(b̌n ) = −mxn (3.10) holds and the minimum
is taken in b̌ with c(b̌) + mx = 0.
At x = 0 we distinguish between b ≤ b0 (0) and b > b0 (0). For b > b0 (0) (3.9) holds.
For b ≤ b0 (0) it holds XtB,m,Y = 0 and
∞ i c(b)
hX λE[r(Z, b)] − c(b)
V (0) = r(Zi , b)e−δTi − = .
δ δ
i=1

It means that equality in Equation (3.9) holds. However, it follows from (3.9) for b >
b0 (0) that the derivative from the right must be −1. If it holds V (0) > λE[r(Z,b)]−c(b)
δ ,
then the infimum is taken at some b ≥ b0 (0). Because b0 (0) is not possible, it follows
the differentiability from the right.

The proof of uniqueness goes perfectly similar to the proof of Theorem 2.2.3. 

The optimal strategy for x ∈ [0, ε) for some ε > 0 is given by b∗ (x) = b̃. The proof
goes perfectly similar to the explanations before and in Remark 2.2.4.
We have seen, that the derivative of the value function may not exist everywhere and
if it exists, it is not necessarily continuous. However, there is a condition under which
the value function is actually continuously differentiable.

Lemma 3.2.3
If the value function V (x) is convex, then V (x) is continuously differentiable.
Proof: Let b0 (x) denote the root of the equation c(b) + mx for x ≥ 0 and define
Z ∞
f (x) := λ V (x − r(b0 (x), z)) dG(z) − (δ + λ)V (x) .
0

By (3.8), f (x) ≥ 0. Let V ′ (x−) and V ′ (x+) denote the derivatives from the right and
from the left, respectively. Assume now, there exists x̃ ∈ R+ with V ′ (x̃−) < V ′ (x̃+).
By Theorem 3.2.2, f (x̃) = 0. Note, that because the optimal strategy in 0 is b∗ = 1,

79
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

it holds x̃ > 0. W.l.o.g. we can assume, that V (x) is continuously differentiable on

(0, x̃), which means f (x) is continuously differentiable on (0, x̃). There exist sequences
(hn )n≥0 ∈ (0, x̃), lim hn = x̃, with f ′ (hn ) ≤ 0 and (xn )n≥0 ∈ (x̃, ∞), lim xn = x̃, with
n→∞ n→∞
f ′ (xn ) ≥ 0. Letting n tend to infinity we obtain
Z ∞
λ V ′ (x̃ − r(b0 (x), z)) dG(z) − (δ + λ)V ′ (x̃−) ≤ 0 ,
0
Z ∞
λ V ′ (x̃ − r(b0 (x), z)) dG(z) − (δ + λ)V ′ (x̃+) ≥ 0 .
0

Thus V ′ (x̃+) ≤ V ′ (x̃−), which is a contradiction. 

Next we consider some examples in the special case of proportional reinsurance.

3.2.1 Examples
Here we will consider only the proportional reinsurance.
Let us first notice that in the case of proportional reinsurance we obtain as in the case
without interest rate, that the value function is convex, if c(b) is concave. Let namely
x, z ≥ 0, α ∈ (0, 1) and y = αx + (1 − α)z. Let {bxt } be the optimal strategy for initial
capital x and {bzt } the optimal strategy for initial capital z. Define the new strategy
byt = αbxt + (1 − α)bzt . Then for the expected value principle c(byt ) = αc(bxt ) + (1 − α)c(bzt ).
Hence,

Xty − Yty + αYtx + (1 − α)Ytz

Z t Nt
byTi − Zi + αYtx + (1 − α)Ytz
X
y
= y+ c(bs ) ds −
0 i=1
= αXtx + (1 − α)Xtz ≥ 0 .

This implies that Yty ≤ αYtx + (1 − α)Ytz . Thus,

y
V (y) ≤ V B (y) ≤ αV (x) + (1 − α)V (z) ,

which proves the convexity. Like in the case without interest rate, the convexity can be
shown for every premium calculation principle, where c(b) is concave in b.
From the convexity we can conclude, that the value function is continuously differen-
tiable, for proof see Lemma 3.2.3.
The next problem in numerical calculation of the value function is, that we have no
conditions to calculate the initial value V (0) exactly. But since we know V (x) = 0 for
x ≥ λµ(θ − η)m−1 the calculation of initial value becomes less complicated as in the
case without interest rate. We have just to find V (0) with V (λµ(θ − η)m−1 ) = 0. If the
chosen V (0) was too large, we would obtain V (λµ(θ − η)m−1 ) > 0.

80
 3.2 The Classical Risk Model

Proof: Define
g(x) := f (x) − V (x) ,
where f (x) is numerically calculated value function with some fixed initial value f (0)
and g(0) > 0. Let b∗ (x) denote the optimal strategy for V (x) and b0 (x) the root of the
equation c(b) + mx = 0. Replacing the optimal b for f (x) by b∗ (x) yields
Z ∞
∗

′
c(b (x)) + mx g (x) + λ g(x − b∗ (x)z) dG(z) − (δ + λ)g(x) ≥ 0 . (3.11)
0

Note that g(x) = g(0) for x ≤ 0. Because g(0) > 0 and b∗ (0) = 1 it follows g′ (0) > 0.
Let x̂ = inf{x : g′ (x) ≤ 0}. Because g(x) is increasing we conclude that b∗ (x) > b0 (x)
on [0, x̂). From (3.8) and Lemma 3.2.1 we conclude that also b∗ (x̂)
> b0 (x). Because
g(x) is increasing on [0, x̂], it follows from (3.11) that c(b∗ (x̂)) + mx g′ (x̂) > 0, which is
a contradiction. So the function g(x) is strictly increasing on R+ . Therefore, f (x) will
ultimately be increasing. 

First we show, that in the special case of exponentially distributed claim sizes, it is
possible to give a closed expression for the return function corresponding to the constant
strategy B ≡ 1.

Example 3.2.4 (No reinsurance)

In the case, we would not buy reinsurance at all, one can show with usual arguments,
that the corresponding return function V b̃ (x) solves the integro-differential equation
Z ∞
λ f (x − z) dG(z) + (c + mx)f ′ (x) − (δ + λ)f (x) = 0 .
0

Consider now the special case, where the claim sizes are exponentially distributed. Then
the above integro-differential equation becomes

λ ∞
Z
y
f (x − y)e− µ dy + (c + mx)f ′ (x) − (δ + λ)f (x) = 0 .
µ 0

Let
∞
X (a)n n
M (a, b, x) = x ,
(b)n · n!
n=0
π  M (a, b, x) M (1 + a − b, 2 − b, x) 
U (a, b, x) = − x1−b
sin(πb) Γ(1 + a − b)Γ(b) Γ(a)Γ(2 − b)

denote the Kummer’s functions, where (a)n = a(a + 1) · · · (a + n − 1). Then the solutions
to the above integro-differential equation are given by a sum of Kummer’s functions

81
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

Figure 3.5: V b̃ (x) (solid line) and VCl

b̃ (x) for exponentially distributed claim sizes.

multiplied by an exponential function.

  m + δ 2m + δ + λ c + mx 
−x δ+λ
e µ C1 · M , , · (c + mx) m (m + δ + λ)
m m µm
 m + δ 2m + δ + λ c + mx  δ+λ
+ C2 · U , , · (c + mx) m (m + δ + λ)
m m µm
m+δ  2m + δ 3m + δ + λ c + mx  m+δ+λ
+ C1 · M , , · (c + mx) m
(m + δ + λ)µ m m µm

m+δ  2m + δ 3m + δ + λ c + mx  m+δ+λ
− C2 · U , , · (c + mx) m .
mµ m m µm

The coefficients C1 and C2 can be verified from the equations δV b̃ (0) = λµ + c(V b̃ )′ (0)
and lim V b̃ (x) = 0.
x→∞
For the parameters c = 1.3, µ = λ = 1, δ = 0.04 and m = 0.03 we obtain C1 = 0 and
C2 = 0.1076089638. In particular we find V b̃ (0) = 2.047949073. V b̃ (0) yields an upper
boundary for the value V (0).
The function V b̃ (x) together with the return function VCl
b̃ of the strategy B ≡ b̃ in the

classical risk model without interest rate, are plotted in Figure 3.5. One sees, that the

82
 3.2 The Classical Risk Model

curve VClb̃ (x) lies considerably above the curve V b̃ (x). Thus one can minimise the cost

just investing the excess in some riskless asset. 

Example 3.2.5 (Exp( µ1 ) and Pareto distributed claims.)

For the exponentially distributed claim sizes we have to consider

λ ∞
Z
−z
inf V (x − bz)e µ dz + (λµ(bθ − η) + mx)V ′ (x) − (δ + λ)V (x) = 0 .
b∈[0,1] µ 0

Numerically calculated optimal strategy and value function are plotted in Figures 3.6
and 3.7 respectively. The initial value is 1.37. For Pareto distributed claims the HJB
equation becomes
Z ∞
2µ2
inf λ V (x − bz) dz + (λµ(bθ − η) + mx)V ′ (x) − (δ + λ)V (x) = 0 .
b∈[0,1] 0 (µ + z)3

Numerically calculated optimal strategy and value function are plotted in Figures 3.8
and 3.9 respectively. The initial value is 1.91963. 

3.2.2 The special case δ = 0

We consider now the special case δ = 0. The methods, we use in this subsection, differ
very little from the methods of subsection 2.2.3. But there are some interesting nuances
in the case we will consider below. In particular we will see, that investing money into
a riskless asset lowers the cost also in the case δ = 0.
Here we have again to assume the net profit condition η > 0. The Hamilton–Jacobi–
Bellman equation has the form
Z ∞
inf λ V (x − r(z, b)) dG(z) + (c(b) + mx)V ′ (x) − λV (x) = 0 . (3.12)
b∈[0,b̃] 0

From Lemma 3.2.1 we know, that the value function is Lipschitz continuous with |V (x)−
V (y)| ≤ |x − y| and V (x) ≡ 0 for x ≥ λµ(θ − η)m−1 .
We also obtain due to representation (3.7) of the underlying process, that a constant
strategy b with c(b) + mx ≤ 0 for some x ∈ R+ cannot be optimal, because in this case
V b (x) will have an infinite value. Because the strategy bt = b̃ yields a finite value, the
value function, provided it exists, will also be finite.
Let in the following b0 (x) denote the root of the equation c(b)+mx. Note, that b ≤ b0 (x)
will never be the optimal strategy for x, because for such b it would hold
Z ∞
λ V (x − r(z, b)) dG(z) + (c(b) + mx)V ′ (x) − λV (x) > 0 .
0

83
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

1,0

0,9

0,8

0,7

0,6

0,5

0,4

0,3

0,2

0,1

0 1 2 3 4 5 6 7

Figure 3.6: Optimal reinsurance strategy for exponentially distributed claim sizes.

2,0

1,5

1,0

0,5

K1 0 1 2 3 4 5 6

Figure 3.7: Value function for exponentially distributed claim sizes.

84
 3.2 The Classical Risk Model

Figure 3.8: Optimal reinsurance strategy for Pareto distributed claim sizes.

Figure 3.9: Value function for Pareto distributed claim sizes.

85
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

Next we transform the HJB equation using the fact, that for the minimiser b we have
c(b) + mx > 0. Define for this purpose

s(x, b) = inf{z : r(z, b) > x}.

Then we can rewrite the equation (3.12) as follows:

n Z s(x,b)


inf λ V x − r(z, b) dG(z) + λ 1 − G(s(x, b)) V (0) − λV (x)
b∈[0,b̃] 0
Z ∞

o
+λ r(z, b) dG(z) + (c(b) + mx)V ′ (x) − λ 1 − G(s(x, b)) x = 0 .
s(x,b)

With Fubini’s Theorem we obtain:

Z s(x,b)



λ V x − r(z, b) dG(z) + λ 1 − G(s(x, b)) V (0) − λV (x)
0
Z x
1 − G(s(y, b)) V ′ (x − y) dy


= −λ
0

and
Z ∞

Z ∞
λ r(z, b) dG(z) − λ 1 − G(s(x, b)) x = λ 1 − G(s(y, b)) dy.
s(x,b) x

Because for each x there exists an optimal b, we can write for the optimal b(x):
Z x
′
1 − G(s(y, b(x))) V ′ (x − y) dy


(c(b(x)) + mx)V (x) = λ
0
Z ∞
−λ 1 − G(s(y, b(x))) dy , (3.13)
x

and (3.12) transforms to

λ hZ x
′
1 − G(s(y, b)) f ′ (x − y) dy


f (x) = sup
b∈(b0 (x),b̃] c(b) + mx 0
Z ∞ i
− 1 − G(s(y, b)) dy (3.14)
x

From Theorem 3.2.2 we know that the solution to (3.12) and therefore to (3.14) is unique.
For x = 0 we obtain
λE[r(Z, b)]
V ′ (0) = sup .
b∈(b0 (0),b̃] c(b)
1
In particular, for an expected value principle (2.18) we find V ′ (0) = − 1+η .

86
 3.2 The Classical Risk Model

Proposition 3.2.6
There is a unique solution f (x) to (3.12), x ≥ 0, with f (∞) = 0.
Proof: Define an operator F , acting on negative functions w(x) by
λ hZ x

F (w(x)) = sup 1 − G(s(y, b)) w(x − y) dy
b∈(b0 (x),b̃] c(b) + mx 0
Z ∞ i
− 1 − G(s(y, b)) dy (3.15)
x
We have already seen, that the value function V (x) solves the equation (3.12). Let
now w0 (x) := (V 1 )′ (x), the solution of Example 2.2.1 with bt = b̃. Define recursively
wn (x) = F (wn−1 (x)). We show at first, that the sequence wn is monotone increasing
in n. It is clear, that w0 (x) ≤ w1 (x) because (V b̃ )′ (x) solves the right hand side of
(3.13) with b = b̃ instead of the sup. Assume wn−1 (x) ≤ wn (x). Because the right
side of (3.15) is continuous in b, there is a maximum point bn ∈ [b0 (x), b̃] for which
wn (x) = F (wn−1 (x)) attains its maximum. So we have
wn+1 (x) − wn (x) = F (wn (x)) − F (wn−1 (x))
λ hZ x

= F (wn (x)) − 1 − G(s(y, bn )) wn−1 (x − y) dy
c(bn ) + mx 0
Z ∞ i
+ 1 − G(s(y, bn )) dy
x
λ hZ x
i
≥ 1 − G(s(y, bn )) [wn (x − y) − wn−1 (x − y)] dy
c(bn ) + mx 0
≥ 0.
So the sequence wn (x) is increasing in n and wn (x) < 0. It means we obtain that
w(x) = lim wn (x) exists pointwise. By the monotone convergence theorem we have
n→∞
then
Z x Z x



lim 1 − G(s(y, b)) wn (x − y)dy = 1 − G(s(y, b)) w(x − y) dy
n→∞ 0 0
for all x and b.
Let b be a maximal point of F (w(x)). Then we have
λ hZ x

wn (x) = 1 − G(s(y, bn )) wn−1 (x − y) dy
c(bn ) + mx 0
Z ∞ i
− 1 − G(s(y, bn ))dy
x
λ hZ x

≥ 1 − G(s(y, b)) wn−1 (x − y) dy
c(b) + mx 0
Z ∞ i
− 1 − G(s(y, b)) dy ,
x

87
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

which means w(x) ≥ F (g(x)). On the other hand wn (x) are increasing in n, i.e. wn (x) ≤
w(x):

λ hZ x

wn (x) = 1 − G(s(y, bn )) wn−1 (x − y) dy
c(bn ) + mx 0
Z ∞ i
− 1 − G(s(y, bn )) dy
x
λ hZ x

≤ 1 − G(s(y, bn )) w(x − y) dy
c(bn ) + mx 0
Z ∞ i
− 1 − G(s(y, bn )) dy
x
λ hZ x

≤ 1 − G(s(y, b)) w(x − y) dy
c(b) + mx 0
Z ∞ i
− 1 − G(s(y, b)) dy ,
x

which means w(x) ≤ F (w(x)). We have therefore w(x) = F (w(x)), and w(x) is contin-
uous.
Because wn (x) is increasing, we can define
Z ∞ Z ∞
f (x) = − w(y) dy ≤ − w0 (y) dy = V b̃ (x) .
x x

f (x) fulfils (3.15) with f (∞) = 0. f (x) is also decreasing, continuously differentiable
and bounded 0 < f (x) ≤ V 1 (x).

Suppose now, that f1 (x) and f2 (x) are solutions to (3.12) with f1 (∞) = f2 (∞) = 0.
Denote further by gi (x) = fi′ (x) the derivatives and by bi (x) the value, for which the
minimum is obtained. Choose now some x∗ > 0. Because the right-hand side of Equation
(3.14) is continuous in b and tends to infinity as c(b) tends to zero, we conclude, that
c(bi (x)) is bounded away from zero on [0, x∗ ]. Let x1 = inf{c(b1 (x)) ∧ c(b2 (x)) : 0 ≤
x ≤ x∗ } and xn = nx1 ∧ x∗ . W.l.o.g. we assume, that x1 ≤ x∗ . Suppose we have
already proved f1 (x) = f2 (x) on the interval [0, xn ]. Then for x ∈ [xn , xn+1 ], with
l= sup |g1 (x) − g2 (x)| it holds
xn ≤x≤xn+1

g1 (x) − g2 (x) = F (g1 (x)) − F (g2 (x))

Z x
λ 
≤ g1 (z) − g2 (z) [1 − G(s(x − z))] dz
c(b2 (x)) + mx xn
λlx1 l
≤ ≤
c(b2 (x)) + mx 2

88
 3.2 The Classical Risk Model

Reversing the roles of g1 (x) and g2 (x), it follows that |g1 (x) − g2 (x)| ≤ 2l . This is only
possible for all x ∈ [xn , xn+1 ] if l = 0. this shows that f1 (x) = f2 (x) on [0, xn+1 ]. So
f1 (x) = f2 (x) on [0, x∗ ]. Because x∗ was arbitrary, uniqueness follows. 

Example 3.2.7 (Proportional Reinsurance Zi ∼Exp( µ1 ) and Zi ∼Pareto(2, µ).)

We consider here only the proportional reinsurance, i.e. r(z, b) = zb.
All the considerations concerning the function V 1 (x) in the case δ > 0 hold also in the
case δ = 0. But here it is easier to consider the derivative (V 1 )′ . For the exponentially
distributed claim sizes, Zi ∼Exp( µ1 ) we have to solve the integro-differential equation
Z x x−y x
−λ (V 1 )′ (y)e− µ dy + c(V 1 )′ (x) + λµe− µ = 0 .
0

This equation is easy to solve, and we obtain as the derivative function (V 1 )′ (x):

λµ  mx λ/m−1 − µx
(V 1 )′ (x) = − 1+ e
c c
for x ≥ 0. Choose µ = λ = 1, m = 0.03 and θ = 0.5 and η = 0.3. We calculate the
value function numerically. The value function and the optimal strategy for Zi ∼Exp( µ1 )
are plotted in Figures 3.10 and 3.11 respectively. The value function and the optimal
strategy for Zi ∼Pareto(2, µ) are plotted in Figures 3.12 and 3.13 respectively. 

89
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

2,5

2,0

1,5

1,0

0,5

0 1 2 3 4 5 6

Figure 3.10: Numerically calculated value function for Zi ∼Exp( µ1 ).

1,0

0,8

0,6

0,4

0,2

0 1 2 3 4 5 6

Figure 3.11: The optimal strategy for Zi ∼Exp( µ1 ).

90
 3.2 The Classical Risk Model

3,5

3,0

2,5

2,0

1,5

1,0

0,5

0 1 2 3 4 5 6

Figure 3.12: Numerically calculated value function for Zi ∼ Pareto(2, µ).

1,0

0,9

0,8

0,7

0,6

0,5

0,4

0,3

0,2

0,1

0 1 2 3 4 5 6

Figure 3.13: The optimal strategy for Zi ∼Pareto(2, µ).

91
3 Optimal Control of Capital Injections by Reinsurance with Riskless Rate of Interest

92
4 Optimal Control of Capital Injections by
Investments, modeled as a Black-Scholes
Model
In the following we will consider a classical risk model and a diffusion approximation
under the special constraint, which allows the insurer in addition to the reinsurance
also to invest money into a risky asset. We also assume, that the scarcity of funds is
impossible. That is the money whenever needed can be put up for example by company
holders. We start with a diffusion approximation to a classical risk model. Because in
this case we will be able to give an explicit expression for the value function and for the
optimal strategy.

4.1 Diffusion Approximation

Let (Ω, F, P) be a complete probability space. We consider a surplus process, where
the mean number of claims in a time unit is λ and the mean size of a claim is µ = E[Z],
where Z is a generic claim size. We assume that µ2 = E[Z 2 ] < ∞. The premium is
c = (1 + η)λµ for some η > 0.
p
Xt = x + λµηt + λµ2 Ws

In addition to the above setup assume now, that the insurer can invest money in a risky
asset, modeled as a Black-Scholes model
n σ2
o
Qt = exp m− t + σ W̃t ,
2
where m, σ > 0. The return of such a process is then the stochastic process {Q′t } given
by the stochastic differential equation

dQ′t = m dt + σ dW̃t .

We assume further, that the Brownian motions {Wt } and {W̃t } are independent and
denote in following the filtration generated by the couple of Brownian motions {(Wt , W̃t )}
by F = {Ft }. We assume also, that the insurer can change the amount at ∈ R invested
at time t continuously and call every strategy A = {at }, at ∈ R, admissible if it is

93
4 Optimal Control of Capital Injections by Investments

cadlag and F adapted. In the following the set of admissible investment strategies will
be denoted by A . The surplus process under the investment strategy A = {at } fulfils
the stochastic differential equation:

dXtA = (λµη + at m) dt + λµ2 dWt + at σ dW̃t .

p
(4.1)

If the surplus falls below zero, the company holders have to inject capital to keep the
process nonnegative. Our goal is to minimise the expected discounted capital injec-
tions over all admissible investment strategies. Denote by {YtA } the process of cu-
mulated capital injections corresponding to the investment strategy A. The value of
the expected discounted
R ∞ −δt capital injections corresponding to the strategy A is given by
A

V (x) = Ex 0 e dYt , where δ ≥ 0. δ expresses the readiness of the company
holders to part from money injecting it into the surplus process. Here we assume δ ≥ 0,
which means, that “injecting money as late as possible” is preferred to “injecting money
now”, if δ > 0; and “injecting money now” is equally attractive as “injecting money
later” if δ = 0. Our goal is to find the value function V (x) = inf V A (x) and the
A∈A
optimal strategy.
Note, that in the case we would not invest at all, i.e. A ≡ 0, the surplus process will
have the form
Xt0 = x + λµηt + λµ2 Wt ,
p

which is a Brownian motion. This case was explicitly treated in Example 2.1.1. The
return function for this case is given by V 0 (x) = β1 exp(−βx) for x ≥ 0 and V 0 (x) = β1 −x
√
λµη+ λ2 µ2 η2 +2λµ2 δ
for x < 0. Where β = λµ2 . Because A ≡ 0 is an admissible strategy we
obtain lim V (x) = 0.
x→∞

Remark 4.1.1
Let {Xt } be a diffusion process with values in R, fulfilling the stochastic differential
equation
dXt = m(Xt ) dt + σ1 (Xt ) dWt1 + σ2 (Xt ) dWt2 ,
where {Wt1 } and {Wt2 } are standard Brownian motions, m, σ1 and σ2 are functions,
such that the above equation has a unique strong solution. The reflected process fulfils
then
dXtY = m(XtY ) dt + σ1 (XtY ) dWt1 + σ2 (XtY ) dWt2 + Yt ,
whereas Y = {Yt } is the local time of the process at zero.
We find in
R ∞Subsection 2.1.1, that the corresponding return function
V (x) = Ex [ 0 e−δt dYt ] solves the differential equation

σ12 (x) + σ22 (x) ′′

V (x) + m(x)V ′ (x) − δV (x) = 0 (4.2)
2

94
 4.1 Diffusion Approximation

for x ≥ 0 and fulfils V ′ (0) = −1, lim V (x) = 0. From Subsection 2.1.1, we know, that
x→∞
every solution f (x) to the above differential equation, vanishing at infinity, has the form
 ∞ −δt
Z
′

f (x) = f (0)Ex e dYt .
0

Example 4.1.2
Now we consider a special case of constant strategies, i.e. we let A ≡ a ∈ R. The
underlying process solves then the stochastic differential equation

dXta = (λµη + am) dt + λµ2 dWt + aσ dW̃t ,

p

and from Remark 4.1.1 it follows, that the corresponding return function V a (x) solves
the differential equation

λµ2 + a2 σ 2 a ′′
(V ) (x) + (λµη + am)(V a )′ (x) − δV a (x) = 0 .
2
The unique solution to the above equation with initial conditions
lim V a (x) = 0 and (V a )′ (0) = −1 is given by
x→∞
√
(λµη+am)2 +2δ(λµ2 +a2 σ 2 )
λµ2 + a2 σ 2 −
λµη+am+
λµ2 +a2 σ 2
x
p e .
λµη + am + (λµη + am)2 + 2δ(λµ2 + a2 σ 2 )

The natural question is, whether there is a constant value a, which minimises V a (x) for
all x. Later we will give a detailed answer to this question. 
Thus, the capital injection YtA for the process (4.1) is given as a local time: YtA =
− min{ inf XsA , 0}. Denoting by τxA the time of ruin of the process {XtA } with initial
0≤s≤t
capital x we can write for the value V A (x):
 ∞ −δt
Z
A ′ A
dYt = V A (0)E e−δτx ,
 
V (x) = E e
τxA

where A′t = AτxA +t . Choose now a strategy A′ in such a way, that V (x) + ε > V A (x) for
some ε > 0, then we can write for initial capital 0:
A A
V (0)E e−δτx ≤ V A (x) ≤ (V (0) + ε)E e−δτx .
 

A
Thus, in order to minimise V A (x) we have to minimise E e−δτx over all admissi-

A
ble strategies A ∈ A . For this purpose we define LA (x) := E e−δτx and L(x) :=


inf LA (x). Because V (x) is a decreasing function and lim V (x) = 0 we can transfer
A∈A x→∞
these characteristics to the function L(x).

95
4 Optimal Control of Capital Injections by Investments

Now we assume, that the optimal strategy exists and denote it by A∗ = {a∗t }, which
∗
means V (x) = V A (x). Note, that if we start with initial capital x + y for x, y ≥ 0 we
will cross the level y before the ruin at τx+yA∗ . In particular we will cross the level y at
∗ ∗
the time τxA , i.e. at the ruin time of the process {XtA } with initial value x. Now we
take advantage from the fact, that the process {Xt } has independent increments and
A∗ − τ A∗ = τ A′′ , where a′′ = a∗
τx+y . Using, that for an optimal strategy we can
x y t t+τxA∗
 −δτ A∗   −δ(τ A∗ −τ A∗ ) 
minimise the terms E e x and E e x+y x independently we can decompose
the function L(x) as follows:
∗ A∗  A∗ A∗ A∗ 
L(x + y) = LA (x + y) = E e−δτx+y = E e−δτx · e−δ(τx+y −τx )
 

A∗  A∗ A∗ 
= E e−δτx · E e−δ(τx+y −τx )
 
∗ ∗
= LA (x) · LA (y) .

The above equation is the functional equation of the exponential function, i.e. L(x) =
exp(−βx) for some β > 0. In particular we obtain, that for arbitrary constant strategy
A ≡ a it holds La (x) = exp(−β(a)x) for some β(a) > 0. Choose now x > 0 arbitrary
and n ∈ N. Dividing the interval [0, x] in n equidistant intervals we must solve the same
optimisation problem in each interval, which means, that the optimal strategy should
be constant. From La (x) = exp(−β(a)x) it follows, that in order to minimise La (x)
(and accordingly V a (x)) over a we have to maximise β(a). The other way would be
to consider the Hamilton–Jacobi–Bellman (HJB) equation for the function V (x) and to
find the optimal a directly. An explicit derivation of the HJB equation, corresponding
to the problem we consider, can be found in Subsection 2.1.2.
We obtain for x ≥ 0:

σ 2 a2 + λµ2 ′′
inf V (x) + (λµη + am)V ′ (x) − δV (x) = 0 . (4.3)
a∈R 2
In particular we find in Subsection 2.1.1, that V ′ (0) = −1. The above expression is
continuously differentiable in a, so that we obtain by differentiating it with respect to a,
′ (x)
that the optimal strategy should be given by a = − σmV 2 V ′′ (x) . Plugging this ansatz into

the HJB equation (4.3) yields the differential equation

m2 V ′ (x)2 λµ2 ′′
− + V (x) + λµηV ′ (x) − δV (x) = 0 .
2σ 2 V ′′ (x) 2

Because we conjecture, that the optimal strategy is constant, we make the ansatz V (x) =
V (0) exp(−βx) and obtain the equation

m2 λµ2 2
− 2 + β − λµηβ − δ = 0 .
2σ2 2

96
 4.1 Diffusion Approximation

Now we can find β. The above equation is merely a quadratic equation in β, so that the
unique positive root is given by
q 2
m
λµη + λ2 µ2 η 2 + 2λµ2 ( 2σ 2 + δ))
2
β= , (4.4)
λµ2

and according to this

m mλµ2 /σ 2
a= = . (4.5)
σ2 β
q
m2
λµη + λ2 µ2 η 2 + 2λµ2 ( 2σ 2 + δ))

In order to show, that our claim, the optimal strategy is a constant one, is correct, we
prove the verification theorem.

Theorem 4.1.3
Let f (x) be a solution to Equation (4.3), then f (x) = V (x) = β1 exp(βx) for x ≥ 0 with
β given in (4.4) and f (x) = V (x) = β1 − x for x < 0. The constant strategy given in
(4.5) is optimal.

Proof: We know from Subsection 2.1.1, that the return function corresponding to some
constant strategy A ≡ a ∈ R is given by
(
1
β(a) exp(β(a)x) : x≥0,
V a (x) = 1
β(a) −x : x<0,

where the factor β(a) depends on a and is given by

p
λµη + am + (λµη + am)2 + (a2 σ 2 + λµ2 )δ
β(a) = .
a2 σ 2 + λµ2

Maximising β(a) with respect to a yields the values in (4.4) and (4.5). Note, that the
∗
function V a (x), a∗ given in (4.5), is the unique solution to the HJB equation with
∗
initial constraint (V a )′ (0) = −1. Denote by A∗ = {a∗t } the optimal strategy and by
X ∗ the corresponding process. Consider further an arbitrary strategy A = {at }. The
corresponding process XtA fulfils then
Z t Z t
XtA
p
= x + λµηt + m as ds + λµ2 Wt + σ as dW̃s .
0 0

∗
Consider now V a , which is a twice continuously differentiable solution to the HJB

97
4 Optimal Control of Capital Injections by Investments

equation (4.3). We apply Ito’s formula and obtain with (4.3):

Z t
e−δt V a (XtA,Y ) = f (x) +
∗ ∗
e−δs (V a )′ (XsA,Y ) dYsA
0
Z t h i
∗ ∗
+ e−δs Ds,A V a (XsA,Y ) − δV a (XsA,Y ) ds
0
Z t
∗
+σ e−δt (V a )′ (XsA,Y )as dW̃s
0
Z t
∗
e−δt (V a )′ (XsA,Y ) dWs
p
+ λµ2
0
Z t
∗
≥ V a (x) − e−δs dYsA
0
Z t
∗
+σ e−δt (V a )′ (XsA,Y )as dW̃s
0
Z t
∗
e−δt (V a )′ (XsA,Y ) dWs
p
+ λµ2
0

a∗ a∗ ∗
where Ds,A V (x) = 21 (a2s σ 2 + λµ2 )(V )′′ (x) + (λµη + mas )(V a )′ (x). Note, that for
∗
AR= A∗ the equality holds. Because √ the derivative (V a )′ (x) is bounded, the processes
t −δt t
σ 0 e (V a )′ (Xs )as dW̃s and λµ2 0 e−δt (V a )′ (XsA,Y ) dWs are martingales with
A,Y
∗ R ∗

expectation equal to 0. Applying the expectations on the both sides of the above in-
equality yields
 t −δs
Z
E e−δt V a (XtA,Y ) ≥ V a (x) − E
∗ ∗
dYsA .
  
e
0

Because it holds V a∗ (XtA,Y )≤V a∗

(0) we obtain by bounded convergence
 ∞ −δs
Z
a∗
dYsA .

V (x) ≤ E e
0
∗
For X = X ∗ equality holds and thus V a (x) = V (x). 

98
 4.2 The Classical Risk Model

V b (x)
V 6.2778 (x)
V 0 (x)
6
V 43.06 (x)

x
0 2 4 6 8

Figure 4.1: Return functions for optimal investment a∗ = 6.2778, a = 0 and a = 43.06.

Example 4.1.4
Choose for example η = 0.3, λ = 1, δ = 0.04, µ = 1 and µ2 = 2. Then one can easily
calculate, that β = 0.4778719262 and a∗ = 6.277832691. Figure 4.1 compares the value
of the optimal strategy with the return function in the case without investments, a = 0,
and return function for the constant strategy A ≡ 43.06. We can see that investments can
lower the costs. However, the choice of a “wrong” strategy can also cause considerable
costs, see the dotted curve in Figure 4.1.
Choosing δ = 0 corresponds to minimising the ruin probability. In this case our
findings coincide with the result in Schmidli [70, p. 41]. With the same parameters
and δ = 0 we obtain β = 0.4098076211 and a∗ = 7.320508077. Also in this case the
discounting has a big influence on the optimal strategy and accordingly on the value
function. 

4.2 The Classical Risk Model

Now we switch to the classical risk model. This case is considerably more complicated
than the case of diffusion approximations. Here we will not have the comfort of knowing
that the value function is two times continuously differentiable, and viscosity solutions
come up to take the place of exact solutions. At first we will consider the concept
of viscosity solutions for the general case δ ≥ 0. Then we show under some special
constraints the existence and uniqueness of the exact solution in the special case δ = 0.

99
4 Optimal Control of Capital Injections by Investments

Consider now a classical risk model with investments and without the possibility to
Nt
P
buy reinsurance. We assume as usual that the aggregate claims amount Zi obeys
i=1
compound Poisson distribution, i.e. the distribution where the number of claims is a
Poisson process Nt with intensity λ, the individual claims amounts Zi are iid with mean
value µ. By c we denote the premium income in a time unit. I.e. the surplus process
has the form
XNt
Xt = x + ct − Zi .
i=1

In addition to the classical setup we allow the insurer to invest into a risky asset, modeled
as a Black-Scholes model
σ2
Qt = exp{(m − )t + σWt } .
2
where {Wt } is a standard Brownian motion and m, σ > 0. The return of such a process
is then the stochastic process {Q′t } given by the stochastic differential equation

dQ′t = m dt + σ dWt .
Nt
P
We also assume, that the processes { Zi } and {Wt } are independent and consider in
i=1
Nt
P
following the filtration {Ft }, generated by the two dimensional process ( Zi , Wt ).
i=1
The insurer can change the amount at ∈ R, which should be invested at time t ≥ 0,
continuously. If the insurer chooses some admissible ({Ft }-adapted and cadlag) strategy
A = {at }, then his surplus process fulfils the stochastic differential equation
Nt
X
dXtA = c + mat dt − d Zi + σat dWt .
i=1
Rt
Cadlag processes are locally bounded, so that the integral 0 as dWs is well defined.
To prevent, that the surplus process becomes negative, the insurer has to inject capital.
The process of cumulated capital injections corresponding to some admissible investment
strategy A we denote by {YtA }. The process with capital injections has then the form
Z t Nt Z t
XtA,Y
X
= x + ct + m as ds − Zi + σ as dWt + YtA .
0 i=1 0

R∞
We are interested in the value function V (x) = inf Ex [ 0 e−δt dYtA ], where A denotes
A∈A
the set of admissible strategies. We also target to find the optimal strategy A∗ , leading

100
 4.2 The Classical Risk Model

∗
to the value function V (x) = V A (x). It is clear, that it makes sense to inject capital
only if the surplus becomes negative; and directly after the capital injection the surplus
is equal to zero. Using this fact we will calculate the return functions for the constant
strategies. In general it will be impossible to give a closed expression for the value
function. All the more it is important to give as far as possible the closed expressions for
return functions corresponding to the constant strategies. Thus, we start by considering
constant strategies.

4.2.1 Constant strategies.

In this section we assume, that an investment strategy can be chosen only at the
beginning. For some fixed a ∈ R the surplus process has then the form
Nt
X
Xta = x + (c + am)t + σaWt − Zi .
i=1

The simplest case, where we can write down the return function without problems is the
Special case A ≡ a = 0.

The surplus process without investments and capital injections has the form
Nt
X
Xt0 = x + ct − Zi ,
i=1

i.e. we have now a classical risk model. Now we distinguish between the case δ = 0 and
δ > 0. For δ = 0 we know from Example 2.2.1, that if the net profit condition c > λµ
is fulfilled, the corresponding return function V 0 (x) can be expressed as a sum of two
Gerber-Shiu penalty functions. For the special case of exponentially distributed claims
one can actually give a closed expression for V 0 (x):
µ − µ(1+η)
η
x
V 0 (x) = e ,
η

for x ≥ 0. For the negative initial capital we have V 0 (x) = V 0 (0) − x = c−λµ
λ − x.
If δ > 0 we do not need to assume c > λµ. We know from Example 2.2.1, that the
return function corresponding to the constant strategy A = 0 is like in the case δ = 0 just
a sum of Gerber-Shiu penalty functions; in the special case of exponentially distributed
claim sizes we have
1 + Rµ
V 0 (x) = − exp(Rx)
R
for x ≥ 0. Where R is the unique negative root of the equation
Z ∞
δ + λ + cR = λ eRx dG(x) . (4.6)
0

101
4 Optimal Control of Capital Injections by Investments

Figure 4.2: Ruin due to the diffusion.

The case A = 0 was explicitly treated in Example 2.2.1, thus we skip all further partic-
ulars and switch to the
General case A ≡ a ∈ R.

The underlying process has now the form

Nt
X
a
X = x + (c + ma)t + σaWt − Zi .
i=1

Note at first, that in this case the ruin can occur in two ways: by a jump, caused by
PNt
the process Zi ; or by oscillation of the Brownian motion. Figure 4.2 illustrates the
i=1
situation, where the ruin occurs by oscillation. In Figure 4.3 one can see the possible
changes in the scenario, if the ruin occurs by a jump: after the ruin, occurring in this
example directly at the first claim arrival time T1 , we shift the process to zero by injecting
additional money.
In particular we obtain for an arbitrary constant strategy A ≡ a ∈ R
N
Xa − x Wt 1 Xt
t→∞
= c + ma + σa − Zi −→ c + ma − λµ .
t t t
i=1

Hence if δ = 0 it holds V a (0) = ∞ for all a ∈ R with a ≤ λµ−c

m , because in the case
a < λµ−c
m the process Xt
a goes to −∞ and in the case a = λµ−c X a oscillates as t goes
m t
to ∞. Conversely in the case a > λµ−c
m X a goes to ∞ as t → ∞. Thus, if the net profit
t

102
 4.2 The Classical Risk Model

120

100

80

60

40

20

0,5 1,0 1,5 2,0

K20

K40
T1

Figure 4.3: Ruin due to a claim.

condition is not fulfilled c ≤ λµ , choose a > λµ−c a

m . For δ > 0 we will have V (0) < ∞
for all a ∈ R.
In following we show, how to calculate the return function of a constant strategy for the
case δ > 0. All considerations can be repeated also for the case δ = 0 restricted on a ∈ R
with V a (0) < ∞.
For the constant strategies A = a with a 6= 0 arbitrary it is easier to calculate at first
V a (x) in dependence of an unknown initial value V a (0). In the second step we derive
V a (0) by pointing out some special constraints on it.
Fix a ∈ R and let τ define the time of ruin of the process {Xta }. Then we can write
for the corresponding return function

V a (x) = V a (0)Ex [e−δτ 1I[Xτa =0] ; τ < ∞]

+ Ex [e−δτ (V a (0) − Xτa )1I[Xτa <0] ; τ < ∞] .

It means we distinguish between the case, where ruin occurs due to the oscillation
Xτa = 0, and the case, when ruin occurs by a jump Xτa < 0. Let further ρa be the
unique positive root of the extended Lundberg equation

∞
σ 2 a2 ρ2a
Z
λ + δ − (c + ma)ρa − =λ e−ρa y dG(y) .
2 0

Gerber and Landry showed in [32], that the function V a (x) fulfils the defective renewal

103
4 Optimal Control of Capital Injections by Investments

equation
Z x Z ∞
a a
V (x) = V (x − y)ga (y) dy + (V a (0) + (y − x))ga (y) dy
0 x
Z ∞
a
+ V (0)Ha (x) − Ha (x) (V a (0) + y)ga (y) dy , (4.7)
0

where
σ 2 a2
Da = ,
2
c + ma
βa = + ρa ,
Da
Ha (x) = e−βa x ,
c + ma −βa s
ha (s) = ·e ,
Da
Z ∞
λ
γa (s) = eρa s e−ρa x dG(x) ,
c + ma s
Z y
ga (y) = ha (y − s) · γa (s) ds .
0

Using standard methods from renewal theory one can find the unique solution V a (x) to
the above renewal equation, which depends on the still unknown value V a (0).
Now we are able to specify the initial value V a (0). We assume a > 0, the case a < 0
goes similar. V a (0) can be decomposed into the expected discounted capital injections
due to the Brownian motion up to T1 ; the expected discounted capital injection at T1 due
to the first jump and finally the expected discounted capital injections from T1 forward,
i.e.
h Z T1 i
a
e−δt dYta + E e−δT1 V a (XTa,Y
 
V (0) = E 1−
− Z1 ) . (4.8)
0
Note, that until T1 − the capital injections can be caused only by the oscillation of the
Brownian motion, see Figure 4.4. Thus consider the Brownian motion

Ct = (c + ma)t + aσWt .

The process CtY with capital injections Yt fulfils then

CtY = (c + ma)t + aσWt + Yt ,

Consider now the process x + Ct and let U (x) denote the expected discounted capital
injections due to the process x + Ct up to the time T1 :
h Z T1 i
U (x) := Ex e−δt dYtx .
0

104
 4.2 The Classical Risk Model

60

50

40

30

20

10

0,5 1,0 1,5

KK Z
K10 a, Y
XT 1

T 1

K20
1

Figure 4.4: Brownian motion up to T1 and the first claim.

where Ytx are the local times and fulfil

n o n o
Ytx = − min inf (x+(c+ma)t+aσWt), 0 = max sup (−x−(c+ma)t−aσWt), 0 .
0≤s≤t 0≤s≤t

It is clear, that U (x) is a decreasing function and because of the discounting it follows
lim U (x) = 0. For negative capital we let U (x) = U (0) − x. Like in Subsection 4.1 the
x→∞
function U (x) satisfies the differential equation

a2 σ 2 ′′
U (x) + (c + ma)U ′ (x) − (δ + λ)U (x) = 0 .
2
with initial constraints U ′ (0) = −1 and lim U (x) = 0. Every solution to the above
x→∞
equation is given by a sum of exponential functions multiplied with some constants K1
and K2 , see Proposition E.1.4,
n (c + ma) + p(c + ma)2 + 2a2 σ 2 (δ + λ) o
K1 exp −
a2 σ 2
n (c + ma) − p(c + ma)2 + 2a2 σ 2 (δ + λ) o
+ K2 exp − .
a2 σ 2
Because U (x) goes to 0 as x goes to infinity we √
obtain K2 = 0. On the other hand it
(c+ma)2 +2a2 σ2 (δ+λ) −1
n o
(c+ma)+
holds U ′ (0) = −1, which yields K1 = 2
a σ 2 . In particular

105
4 Optimal Control of Capital Injections by Investments

we have U (0) = K1 . Thus, it remains to calculate

E e−δT1 V a (CTY1 − Z1 ) . If we denote by ft (x) the density function of CtY − Z1 we obtain


Z ∞
E e−δT1 V a (CTY1 − Z1 ) = λe−(δ+λ)t E[V a (CtY − Z1 )] dt
 
0
Z ∞ Z ∞
−(δ+λ)t
= λe V a (y)ft (y) dy dt
0 0
Z ∞ Z 0
−(δ+λ)t
V a (0) − y ft (y) dy dt .


+ λe
0 −∞

Thus we have to find ft (x). We assume aσ > 0, then we can rewrite the process CtY in
the following way
c + ma c + ma
CtY = −aσ{− t − Wt } + aσ sup (− s − Ws ) ,
aσ 0≤s≤t aσ

i.e. CtY is just the difference of a Brownian motion −(c + ma)t − aσWt and its running
maximum. For definition of running maximum see Remark A.1.6, p. 165. With change
of measure techniques, see Proposition C.2.4 p. 172, it is straight forward to calculate
the density function of the process CtY , which is given by

2 1 n (x − (c + ma)t)2 o
ht (x) = √ exp −
aσ 2πt a2 σ 2 t2
c + ma  x c + ma √
 n (c + ma)x o
−2 2 2 1−Φ √ + t exp 2 ,
a σ aσ t aσ a2 σ 2

where Φ denotes the distribution function of the standard normal distribution. Let G̃
be the distribution function of −Z1 , then we obtain for the desired density ft (x):
x
(R
ht (x − y) dG̃(y) : x ≤ 0 ,
ft (x) = R−∞0
−∞ ht (x − y) dG̃(y) : x > 0 .

Now we are able to calculate E e−δT1 V a (CTY1 − Z1 ) using the knowledge of the function
 

V a (x) for x > 0. Finally we can find V a (0) solving the Equation (4.8). Next we illustrate
the method by an example.

Example 4.2.1
−x
Assume, that the claim sizes are exponentially distributed G(x) = 1 − e µ , then we can
give a closed expression for the function ga (y):

λ µ y
ga (y) = · · (e− µ − e−βa y ) .
Da (ρa µ + 1) (βa µ − 1)

106
 4.2 The Classical Risk Model

λ
Let Ja := Da (ρa µ+1) . The renewal Equation (4.7) becomes
x
µ
Z
−y
a
V (x) = Ja V a (x − y) (e µ − e−βa y ) dy + V a (0)e−βa x
0 βa µ − 1
µ −x
− e−βa x ) Ja · µ2 + Ja · V a (0)µ .


+ (e µ
βa µ − 1
x x
µ
Note, that βa µ−1 (e− µ −e−βa x ) is equal to the convolution of the functions e− µ and e−βa x .
It is straight forward to calculate, that the solution to the above renewal equation is given
by the function
Z x
x 1
2 −µ −y(βa − µ ) 2


f (x) = µ Ja e e 0 F1 (1, Ja xy − Ja y ) − 1 dy
0
x
nZ x 1
−µ
a
+ V (0)Ja e e−y(βa − µ ) y · 0 F1 (2, Ja xy − Ja y 2 )) dy
0
Z x
1
o
−y(βa − µ ) 2
+µ e F
0 1 (1, Ja xy − Ja y ) − 1 dy
0
x
−µ
Ja µ(e − e−βa x )(µ2 V a (0) + µ)
+ + V a (0)e−βa x ,
(βa µ − 1)
∞
P xn
where 0 F1 (α, x) = Γ(n+α)n! , i.e. a hypergeometric function and Γ(x) the Gamma
n=0
function. For the parameters m = 0.03, σ 2 = 0.01, µ = λ = 1, c = 1.3 and a = 12 we
obtain V 12 (0) = 1.824317721. The function V 12 (x) is plotted as a solid line in Figure 4.5.
We see, that the return function corresponding to the strategy A ≡ 0, no investment,
(dashed line in Figure 4.5) lies above the function V 12 (x). Thus even a constant strategy,
which is not optimal, may reduce the cost. 

4.2.2 General case δ ≥ 0.

For the general case δ ≥ 0 we will not be able either to indicate the value function
explicitly or to show the existence of the optimal strategy. The following lemma states
a couple of useful properties of the value function, which hold for δ ≥ 0.

Lemma 4.2.2
The value function V (x) has the following properties

1. V (x) is decreasing with lim V (x) = 0.

x→∞

2. V (x) is Lipschitz continuous on [0, ∞) with |V (x) − V (y)| ≤ |x − y|.

3. V (x) is convex.

107
4 Optimal Control of Capital Injections by Investments

Figure 4.5: Return functions for the constant strategies A ≡ 0 (dashed line) and A ≡ 12
(solid line) for exponentially distributed claim sizes.

Proof: It is clear, that the value function is decreasing. Consider now x, z ≥ 0, α ∈ (0, 1)
and define y = αx + (1 − α)z. Let further Ax = {axt } be the optimal strategy for the
initial capital x and Az = {azt } - for initial capital z. Denote by Ay = {ayt } the optimal
strategy for the initial capital y, then it holds
y ,Y y x z
XtA − YtA + αYtA + (1 − α)YtA
Z t Z t Nt
x z
X
y y
= x + ct + m as ds + σ as dWs − Zi + αYtA + (1 − α)YtA
0 0 i=1
Ax ,Y Az ,Y
= αXt + (1 − α)Xt ≥0.
x z y
Thus αYtA +(1−α)YtA ≥ YtA . Since the capital injection processes {Yt } are increasing,
they are of bounded variation. With integration by parts we obtain
Z ∞ Z ∞
−δt
e dYt = δ e−δt Yt dt.
0 0

This yields together with the above result

V (αx + (1 − α)z) = V (y) ≤ αV (x) + (1 − α)V (z) .

Let now z > x and A = {at } be an investment strategy for initial capital z such that
V A (z) ≤ V (z) + ε. For initial capital x choose the strategy à (which is not optimal):

108
 4.2 The Classical Risk Model

inject the capital z − x and then follow the strategy A. Thus

V (x) − V (z) ≤ V Ã (x) − V A (z) + ε = z − x + ε .

Because ε is arbitrary we have |V (x) − V (z)| ≤ |x − z|, which proves the Lipschitz-
continuity. That is V (x) is absolutely continuous. Because of the Lipschitz-continuity
and convexity the value function is differentiable almost everywhere and satisfies −1 ≤
V ′ (x) ≤ 0. At the points, where V (x) is not differentiable the derivatives from the left
and from the right exist.
From Subsection 4.2.1 it follows directly, that lim V (x) = 0. 
x→∞

We conjecture, that the value function V (x) solves the Hamilton–Jacobi–Bellman equa-
tion (HJB)
Z ∞
σ 2 a2 ′′
inf V (x) + λ V (x − z) dG(z) + (c + am)V ′ (x) − (δ + λ)V (x) = 0 . (4.9)
a∈R 2 0

For explicit derivation of the HJB equation we refer to the proof of Theorem 4.2.6 or
alternatively to Schmidli [70, p. 55]. Because we do not even know, whether the value
function is once continuously differentiable, we will be looking for a viscosity solution.
Note, that for every twice continuously differentiable function f and continuous function
u the minimum of
Z ∞
σ 2 a2 ′′
f (x) + λ u(x − z) dG(z) + (c + am)f ′ (x) − (δ + λ)u(x)
2 0
−f ′ (x)m
in a is attained at a = σ2 f ′′ (x)
. Using this fact we give the precise definition of viscosity
solutions.
Definition 4.2.3
We say that a continuous function u : [0, ∞) → R+ is a viscosity subsolution to (4.9)
at x ∈ (0, ∞) if any twice continuously differentiable function ψ : (0, ∞) → R with
ψ(x) = u(x) such that u − ψ reaches the maximum at x satisfies
Z ∞
m2 ψ ′ (x)2
− 2 ′′ +λ u(x − z) dG(z) + cψ ′ (x) − (δ + λ)u(x) ≥ 0 , (4.10)
2σ ψ (x) 0

and we say that a continuous function ū : [0, ∞) → R+ is a viscosity supersolution

to (4.9) at x ∈ (0, ∞) if any twice continuously differentiable function φ : (0, ∞) → R
with φ(x) = ū(x) such that ū − φ reaches the minimum at x satisfies
Z ∞
m2 φ′ (x)2
− 2 ′′ +λ ū(x − z) dG(z) + cφ′ (x) − (δ + λ)ū(x) ≤ 0 .
2σ φ (x) 0

A viscosity solution to (4.9) is a continuous function u : [0, ∞) → R+ if it is both a

viscosity subsolution and a viscosity supersolution at any x ∈ (0, ∞).

109
4 Optimal Control of Capital Injections by Investments

There is an equivalent formulation of viscosity solutions.

Definition 4.2.4
A continuous function u : [0, ∞) → R+ is a viscosity subsolution to (4.9) at x ∈ (0, ∞)
if any twice continuously differentiable function ψ : (0, ∞) → R with ψ(x) = u(x), such
that u − ψ reaches the maximum at x, satisfies
∞
m2 ψ ′ (x)2
Z
− +λ ψ(x − z) dG(z) + cψ ′ (x) − (δ + λ)ψ(x) ≥ 0 , (4.11)
2σ 2 ψ ′′ (x) 0

where ψ(−x) = ψ(0) − x for x ∈ (0, ∞).

A continuous function ū : [0, ∞) → R+ is a viscosity supersolution to (4.9) at x ∈ (0, ∞)
if any twice continuously differentiable function φ : (0, ∞) → R with φ(x) = ū(x), such
that ū − φ reaches the minimum at x, satisfies
∞
m2 φ′ (x)2
Z
− 2 ′′ +λ φ(x − z) dG(z) + cφ′ (x) − (δ + λ)φ(x) ≤ 0 ,
2σ φ (x) 0

where φ(−x) = φ(0) − x for x ∈ (0, ∞).

Given a twice continuously differentiable function f and a continuous function u, we

denote in the following
Z ∞
m2 f ′ (x)2
L(u, f )(x) = − 2 ′′ +λ u(x − z) dG(z) + cf ′ (x) − (δ + λ)u(x) ,
2σ f (x) 0
Z ∞
m2 f ′ (x)2
L(f )(x) = − 2 ′′ +λ f (x − z) dG(z) + cf ′ (x) − (δ + λ)f (x) .
2σ f (x) 0

Proof of equivalence:
The proof technique we use below can be found for example in Benth et al. [8].
We prove the statement for the subsolutions, the supersolution case can be proven sim-
ilarly.
Let u be a viscosity subsolution and ψ a twice continuously differentiable function,
fulfilling the conditions from Definition 4.2.3. Then it holds because ψ(y) ≥ u(y) for all
y ∈ (0, ∞) and ψ(x) = u(x):

L(ψ)(x) ≥ L(u, ψ)(x) ≥ 0 ,

which yields Definition 4.2.4.

Conversely, let u be a subsolution, satisfying the conditions 1., 2. and 3. from Lemma
4.2.2, and ψ̃ a twice continuously differentiable function in the sense of Definition 4.2.4.

110
 4.2 The Classical Risk Model

Let further h ∈ [0, x] be fixed and gn be a smooth function satisfying 0 ≤ gn ≤ 1,

gn (y) = 1 for y ∈ (x − h + n1 , x + h − n1 ) and g(y) = 0 for y ∈
/ [x − h, x + h]. Define now
(
1
exp(− 1−y 2) : |y| < 1 ,
f (y) =
0 : |y| ≥ 1 .

Then f (x) is an even, nonnegative, smooth function with support in (−1, 1). Let
1
1 s+1
Z
fn (y) = R 1 u(y − )f (s) ds .
n
−1 f (s) ds −1

It is obvious, that fn are smooth functions and fn converges uniformly to u with fn ≥ u.

Then define the test functions

ψn (y) = gn (y)ψ̃(y) + (1 − gn (y))fn (y) .

Apparently ψn are twice continuously differentiable, ψn (x) = ψ̃(x) = u(x), ψn (y) ≥ u(y)
for all y ∈ (0, ∞) and ψn converges uniformly to u. Therefore, it holds due to Definition
4.2.4
n→∞
0 ≤ L(ψn )(x) −→ L(u, ψ̃)(x) .
Thus Definition 4.2.3 follows.

Remark 4.2.5
Note, that if V is a viscosity solution to (4.9) and is twice continuously differentiable,
then V is a classical solution. The optimal investment strategy is then given by a(x) =
′ (x)
− σmV
2 V ′′ (x) .

Note, that for V ′′ (x) = 0 the HJB equation would not have any solution, because
otherwise there would exist a > 0 with
Z ∞
λ V (x − z) dG(z) + (c + am)V ′ (x) − (δ + λ)V (x) < 0 .
0

The function V (x) is in this case strictly convex.

We show at first, that the value function is a viscosity solution. In the proof of the
theorem below we used the technique proposed by Benth et al. [8], which was also
successfully applied by Albrecher and Thonhauser [1] and Azcue and Muler [5].

Theorem 4.2.6
V (x) is a viscosity solution to (4.9)

111
4 Optimal Control of Capital Injections by Investments

Proof: Assume x > 0. Let h > 0 and a ∈ R be fixed. We choose n ∈ N such that
|V (x) − V (y)| < ε/2 for |x − y| ≤ nx for some ε > 0.
Define further xk = k nx and τ0 := inf{t ≥ 0 : |Xta − x| > nx }. Then τ1 := τ0 ∧ T1 ∧ h
is a stopping time. In particular we obtain Xτa1 ∈ − ∞, x(n+1)

n . For each k there is
k k A k
a measurable strategy A = {at } such that V (xk ) ≤ V (xk ) + ε/2. For the capital
k k
xk ≤ Xτa1 < xk+1 , we choose the strategy Ak . Thus V A (Xτa1 ) ≤ V A (xk ) ≤ V (xk ) +
ε/2 ≤ V (Xτa1 ) + |Xτa1 − xk | + ε/2 < V (Xτa1 ) + ε by Lemma 4.2.2. If Xτa1 < 0 consider
k
V A (0)−Xτa1 . This shows that for each a ∈ R we can find in a measurable way a strategy
 such that V  (Xτa1 ) < V (Xτa1 ) + ε.
We show at first, that V (x) is a viscosity subsolution. Construct a strategy à = {ãt }
in following way. Define τ a := inf{t ≥ 0 : Xta < 0} and T = T1 ∧ τ a ∧ h for some h > 0
very small. Let further ãt = a for t ≤ T and a˜t = at−T for t > T and a strategy A = {at }
with V (x) + ε ≥ V A (X) with ε > 0. Then we obtain

V (x) ≤ V Ã (x) = E[e−δT V A (XTa )]

h a
= E e−δT1 V A (XTa1 )1I[T =T1 ] + e−δτ V A (0)1I[T =τ a ]
i
+ e−δh V (Xha )1I[T =h]

Rearranging the terms and diving by h yields

h1 1 1 i
E e−δT1 V (Xta )1I[T =T1 ] + e−δh V (Xha )1I[T =h] − V (x)
h h h
1 a
+ V (0)E[e−δτ 1I[T =τ a ] ] ≥ 0 .
h
1 a a
Consider at first the term hV (0)E[e−δτ 1I[T =τ a ] ]. Note that e−δτ is bounded and
P[T = τ a ] = P[τ a ≤ h] − P[T1 < τ a ≤ h] .
It holds h−1 P[τ a ≤ h] → 0 as h → 0, from which it follows, that
1 −δτ a 1I
h V (0)E[e [T =τ a ] ] → 0 as h → 0. So in following we can skip considering this term.
Consider further
h1 1 1 i
E e−δT1 V (Xta )1I[T =T1 ] + V (Xha )1I[T =h] − V (x)
h h h
h1 V (X a ) − V (x)
= E e−δT1 V (Xta )1I[T =T1 ] + h
e−δh 1I[T =h]
h h
1 − 1I[T =h] e−δh i
− V (x)
h
h1
−δT1 a V (Xha ) − V (x) −δh i
= E e V (Xt )1I[T =T1 ] + e 1I[T =h]
h h
1 − P[T = h]e −δh
− V (x) .
h

112
 4.2 The Classical Risk Model

On the set {T = h} it holds P[T = h] = P[τ a > h]. Thus we obtain

1 − P[T = h]e−δh
lim V (x) = (δ + λ)V (x) .
h→∞ h
For the first summand h1 e−δT1 V (Xta )1I[T =T1 ] we obtain
h1 i h1 i
E e−δT1 V (Xta )1I[T =T1 ] = E e−δT1 V (Xta )1I[T1 ≤h]
h h
h1 i
− E e−δT1 V (Xta )1I[τ a ≤T1 ≤h] .
h
V (XTa1 )1I[T =T1 ] e−δT1 is bounded and P[τ a < T1 ≤ h]h−1 → 0 as h → 0. so that we obtain
E[ h1 e−δT1 V (Xta )1I[τ a ≤T1 ≤h] ] → 0 as h → 0. We will skip considering this term in the
following proof. For the remaining term we obtain
h1 i 1 Z hZ ∞
−δT1 a
E e V (Xt )1I[T1 ≤h] = E[V (Xta )] dG(z) dt .
h h 0 0
Because V (Xta ) is bounded on the interval [0, h] we have by bounded convergence

1 h ∞
Z Z Z ∞
lim E[V (Xta )] dG(z) dt = V (x − z) dG(z) .
h→0 h 0 0 0

Let further ψ be a twice continuously differentiable function such that V − ψ reaches

the maximum at x and ψ(x) = V (x). We have then:

1 h ∞ 1 − P[T = h]e−δh
Z Z
0 ≤ E[V (Xta )] dG(z) dt − V (x)
h 0 0 h
h V (X a ) − V (x) i
h
+E 1I[T =h]
h
1 h ∞ 1 − P[T = h]e−δh
Z Z
≤ E[V (Xta )] dG(z) dt − ψ(x)
h 0 0 h
h ψ(X a ) − ψ(x) i
h
+E 1I[T =h] . (4.12)
h
Because ψ is two times continuously differentiable we obtain for the generator of the
process Xha on the set T = h:
h ψ(X a ) − ψ(x) i σ 2 a2 ′′
lim E h
1I[T =h] = (c + am)ψ ′ (x) + ψ (x) .
h→0 h 2
So letting h go to 0 in (4.12) we obtain
Z ∞
σ 2 a2 ′′
0≤λ V (x − z) dG(z) − (δ + λ)V (x) + (c + am)ψ ′ (x) + ψ (x)
0 2

113
4 Optimal Control of Capital Injections by Investments

Because a ∈ R was arbitrary we have

Z ∞
σ 2 a2 ′′
inf ψ (x) + λ V (x − z) dG(z) + (c + am)ψ ′ (x) − (δ + λ)V (x) ≥ 0 ,
a∈R 2 0
which yields the result.
It remains to show, that V (x) is a viscosity supersolution at any x > 0. Arguing by
contradiction we assume, that V (x) is not a supersolution at x. Then due to Definition
4.2.4 there exist ξ > 0 and a twice continuously differentiable function φ0 : (0, ∞) → R
with φ0 (x) = V (x) and V (y) − φ0 (y) ≥ 0 for all y > 0, such that
L(φ0 )(x) > 2ξ .
Note, that since −1 < lim inf V ′ (x) ≤ φ′0 (x) ≤ lim sup V ′ (x), we obtain φ′0 (x) > −1.
The case lim inf V ′ (x) = −1 is impossible because of  Remark
 4.2.5.
4 (x−y)π
Consider now the function φ1 (y) := φ0 (y) − sin 2x · λξ . It is obvious, that φ1 is
 
4 (x−y)π ξ
twice continuously differentiable, φ1 (x) = V (x) and φ1 (y) ≤ V (y) − sin 2x λ . We
also have
L(φ1 )(x) > ξ ,
R∞  
because φ′1 (x) = φ′0 (x), φ′′1 (x) = φ′′0 (x) and 0 sin4 yπ 2x dG(y) ≤ 1 . Since L(φ1 )(x)
is continuous in x there is h̃ ∈ (0, x2 ) such that for all y ∈ [x − 2h̃, x + 2h̃] it holds
L(φ1 )(y) > 2ξ .
Like above we conclude φ′1 (x) > −1. Because φ′1 is continuous there is k > 0 such
that φ′1 (y) ≥ −1 and φ(y) > 0 for y ∈ [x − k, x + k]. We let in following [x − 2h̃, x +
2h̃] ∩ [x − k, x + k] = [x − 2h, x + 2h].
Let now (
1
exp(− 1−y 2) : |y| < 1 ,
f (y) =
0 : |y| ≥ 1 .
Then f (y) is an even, nonnegative, smooth function with support in (−1, 1). Define
Z 1/n n
1 hπ
ξ o
fn (y) = R 1 V (y − s) − sin4 nf (ns) ds .
2x 2λ
−1 f (s) ds −1/n

It is obvious, that

fn are smooth functions and the sequence (fn )n≥1 converges uniformly
4 hπ ξ
to V − sin 2x 2λ . Thus, there is n0 ∈ N such that
hπ
ξ hπ
ξ
V (y) − sin4 ≤ fn0 (y) ≤ V (y) − sin4
2x λ 2x 4λ
′
for all y ≥ 0, also we obtain 0 ≥ fn0 (y) ≥ −1 on [0, x − h] because of the construction.
In particular it holds fn0 (y) ≥ φ1 (y) > 0 on [x − 2h, x − h]. Then there exists κ > 1, such
that φ1 (y)κ ≥ fn0 (y) for all y ∈ [x − 2h, x − h]. Let further g be a twice continuously
differentiable function satisfying

114
 4.2 The Classical Risk Model

π
• 2 ≤ g(y) ≤ π
• g(y) = π for y ∈ [x − h, x + h]
π
• g(y) = 2 for y ∈
/ (x − 2h, x + 2h)
• g′ (y) ≥ 0 for y ∈ [x − 2h, x − h].
We define  hπ  ξ
n ξ o
ε = min sin4 ,
2x 12λ 4δ
and the function φ by

fn0 (y)
φ(y) = cos2 (g(y))φ1 (y) + sin2 g(y) .
κ
Obviously φ(y) = φ1 (y) on [x − h, x + h], from which it follows L(φ)(y) > 2δε on
[x − h, x + h]. Since
fn (y)
φ′ (y) = −2g′ (y) cos(g(y)) sin(g(y)) φ1 (y) − 0

κ
f ′ (y)
+ cos2 (g(y))φ′1 (y) + sin2 g(y) n0


.
κ
we obtain φ′ (y) ≥ −1 for y ∈ [0, x + h] and φ(y) ≤ V (y) − 3ε for y ∈ [0, x − h] ∪ {x + h}.
Let A be an arbitrary admissible strategy with {X̂t } = {XtA,Y } and τ ∗ be the exit
time from [x − h, x + h]. It is clear, that X̂τ ∗ ∈ [0, x − h] ∪ {x + h}, because the paths are
continuous between the jumps and the jumps are always downwards. Thus, we obtain
V (X̂τ ∗ ) ≥ φ(X̂τ ∗ ) + 3ε .
Since the function φ(x) is twice continuously differentiable, we have from Proposition
A.2.2, p. 165, that the process
Z τ ∗ ∧t  Z ∞ 
−δs
Mt = − e λ φ(X̂s − z) dG(z) − λφ(X̂s ) ds
0 0
NX
τ ∗ ∧t

e−δTi φ(X̂Ti ) − φ(X̂Ti − )

+
i=1

is a martingale with zero-expectation. Rearranging the terms we obtain

∗ ∧t) ∗
e−δ(τ φ(X̂τ ∗ ∧t ) = e−δ(τ ∧t) φ(X̂τ ∗ ∧t ) − e−δTNτ ∗ ∧t φ(X̂TN ∗ ) + φ(x)
τ ∧t
Z τ ∗ ∧t Z ∞ 
+λ e−δs φ(X̂s − z) dG(z) − φ(X̂s ) ds
0 0
Nτ ∗ ∧t
X
e−δTi φ(X̂Ti − ) − e−δTi−1 φ(X̂Ti−1 ) + Mt .


+ (4.13)
i=1

115
4 Optimal Control of Capital Injections by Investments

On the other hand it holds with Ito’s formula

Z Ti
−δTi −δTi−1
e φ(X̂Ti − ) − e φ(X̂Ti−1 ) = σas e−δs φ′ (X̂s ) dWs
Ti−1
Z Ti
+ e−δs φ′ (X̂s ) dYsA
Ti−1
Z Ti n σ2 o
+ e−δs a2s φ′′ (X̂s ) + (c + mas )φ′ (X̂s ) − δφ(X̂s ) ds
Ti−1 2

Z τ ∗ ∧t
∗ ∧t)
e−δ(τ φ(X̂τ ∗ ∧t ) − e−δTNτ ∗ ∧t φ(X̂TN )= e−δs φ′ (X̂s ) dYsA
τ ∗ ∧t
TNτ ∗ ∧t
Z τ ∗ ∧t
+ e−δs σas φ′ (X̂s ) dWs
TNτ ∗ ∧t
Z τ ∗ ∧t n σ2 o
+ e−δs a2s φ′′ (X̂s ) + (c + mas )φ′ (X̂s ) − δφ(X̂s ) ds .
TNτ ∗ ∧t 2

Plugging in these results into Expression (4.13) we obtain

Z τ ∗ ∧t  σ 2 a2
−δ(τ ∗ ∧t) s
e φ(X̂ τ ∗ ∧t ) − φ(x) = e−δs φ′′ (X̂s ) + (c + mas )φ′ (X̂s )
0 2
Z ∞ 
φ(X̂s − z) dG(z) − (δ + λ)φ(X̂s ) ds
+λ
0
Z τ ∗ ∧t Z τ ∗ ∧t
−δs ′
+ e σas φ (X̂s ) dWs + Mt + e−δs φ′ (X̂s ) dYsA .
0 0

Using
∞
σ 2 a2s ′′
Z
L(φ)(X̂s ) ≤ φ (X̂s ) + (c + mas )φ′ (X̂s ) + λ φ(X̂s − z) dG(z)
2 0
−(δ + λ)φ(X̂s ) ,

and φ′ (X̂s ) ≥ −1 on [x − h, x + h] we obtain

Z τ ∗ ∧t
∗ ∧t)
e−δ(τ φ(X̂τ ∗ ∧t ) − φ(x) ≥e−δs L(φ)(X̂s ) ds + Mt
0
Z τ ∗ ∧t Z τ ∗ ∧t
−δs ′
+ e σas φ (X̂s ) dWs − e−δs dYsA .
0 0

116
 4.2 The Classical Risk Model

Rearranging the terms and using φ(X̂τ ∗ ) ≤ V (X̂τ ∗ ) − 3ε and L(φ)(X̂s ) ≥ 2δε yields
Z τ ∗ ∧t Z τ ∗ ∧t
−δ(τ ∗ ∧t) −δs
e V (X̂τ ∗ ∧t ) + e dYsA ≥ φ(x) + 2δε e−δs ds
0 0
Z τ ∗ ∧t
+ σas φ′ (X̂s ) dWs + Mt
0
∗ ∧t)
+ 3εe−δ(τ . (4.14)
R τ ∗ ∧t
The stochastic integral 0 σas φ′ (X̂s ) dWs is a local martingale. Let Sn be a lo-
calisation sequence for it and Jn = Sn ∧ τ ∗ . Then we have from (4.14) applying the
expectations
h Z Jn i
−δJn
e−δs dYsA ≥ φ(x) + 2ε + Ex [e−δJn ] 3ε − 2ε .


Ex e V (X̂Jn ) +
0

Because the strategy A was arbitrary it follows

h Z Jn i
V (x) = inf Ex e−δJn V (X̂Jn ) + e−δs dYsA ≥ φ(x) + 2ε ,
A∈A 0

which contradicts the assumption V (x) = φ(x). 

We have shown, that V (x) is a viscosity solution to (4.9). The uniqueness of the
value function follows from the comparison principle, which we show in the next propo-
sition. More about comparison principle one can find for example in Bardi and Capuzzo-
Dolcetta [6, p. 82]. The proof technique for δ > 0 used below can be found for example
in Azcue and Muler [5].

Proposition 4.2.7
Let now v(x) be a super- and u(x) a subsolution to (4.9), satisfying conditions 1.–3.
from Lemma 4.2.2. If it holds u(0) ≤ v(0), then u(x) ≤ v(x) on [0, ∞).

Proof: Let u be a sub- and v a supersolution, which fulfil conditions 1.-3. from Lemma
4.2.2. Assume there is x0 ∈ (0, ∞) with u(x0 ) − v(x0 ) > 0. Define vk (x) = kv(x) for k >
1. Then vk (x) is also a supersolution. Choose now k > 1 such that u(x0 )−vk (x0 ) > 0. It
is trivial, that u(0) ≤ vk (0). Because u and vk are decreasing and Lipschitz continuous
we obtain the following estimation

u(x) − vk (x) ≤ u(0) − vk (x) ≤ v(0) − vk (x)

= k(v(0) − v(x)) + (1 − k)v(0)
≤ kx + (1 − k)v(0) .

117
4 Optimal Control of Capital Injections by Investments

Thus it holds u(x) − vk (x) ≤ 0 for x ≤ v(0)(k−1)

k . So we will consider only the crucial
v(0)(k−1)
interval (d, ∞) with d = k .
Define further
M := sup(u(x) − vk (x)) .
x≥0

It follows readily

0 < u(x0 ) − vk (x0 ) ≤ M = sup (u(x) − vk (x)) .

x∈(d,∞)

Let x∗ be such, that M = u(x∗ ) − vk (x∗ ).

Define H := {(x, y) : d < y < ∞, d < x ≤ y} and for ξ > 0

ξ 2k
fξ (x, y) := u(x) − vk (y) − (x − y)2 − 2 .
2 ξ (y − x) + ξ

Let Mξ := sup fξ (x, y). Because fξ is continuous, there exists (xξ , yξ ) ∈ H̄, where H̄
(x,y)∈H
denotes the closure of H, with Mξ = fξ (xξ , yξ ). Because x∗ > d it holds (x∗ , x∗ ) ∈ H.
Thus we have
2k 2k
Mξ ≥ fξ (x∗ , x∗ ) = u(x∗ ) − vk (x∗ ) − =M− .
ξ ξ
4k
We can therefore conclude, that Mξ > 0 for ξ > M and lim inf Mξ ≥ M .
ξ→∞
On the other hand we know, that (xξ , yξ ) ∈ H̄, from which it follows yξ ≥ xξ .
Now we will show, that there exists ξ0 such that for all ξ ≥ ξ0 it holds (xξ , yξ ) ∈
/ ∂H.
The boundary ∂H is the union
  
(d, y) : y ∈ (d, ∞] ∪ (x, x) : x ∈ [d, ∞] ∪ (x, ∞) : x ∈ (d, ∞) .

It is clear, that for fixed x ∈ [d, ∞) it holds lim fξ (x, y) = −∞. Consider the set
y→∞

(x, x) : x ∈ [d, ∞] . Note, that it holds

2k 2k
lim fξ (x, x) = lim u(x) − vk (x) − =− <0
x→∞ x→∞ ξ ξ

and fξ (d, d) = u(d) − vk (d) − 2k

ξ < 0. For d < x1 ≤ x2 < ∞ we have because of the
properties 1. and 2. from Lemma 4.2.2

vk (x1 ) − vk (x2 )
0≤ ≤k. (4.15)
x2 − x1

118
 4.2 The Classical Risk Model

So it holds

fξ (x, x) − fξ (x, x + h) u(x) − vk (x) − 2k

ξ − u(x) + vk (x + h) + 2ξ h2 + 2k
ξ 2 h+ξ
=
−h −h
vk (x) − vk (x + h) ξ 2k
= − − h+
−h 2 ξh + 1
ξ 2k
≥ −k − h + .
2 ξh + 1
fξ (x,x)−fξ (x,x+h)
Thus lim inf −h ≥ k, which means fξ (x, x) ≤ f (x, x + h). It remains to
h→0 
consider the set (d, y) : y ∈ (d, ∞] . For y > d it holds

fξ (d, y) − fξ (d, y − h) vk (y − h) − vk (y)

=
h h
− 2ξ (d + h − y)2 − 2k
ξ 2 (y−d−h)+ξ + 2ξ (d − y)2 + 2k
ξ 2 (y−d)+ξ
+
h
vk (y − h) − vk (y) ξ
= + (2y − h − 2d)
h 2
2k
−
(ξ(y − d − h) + 1)(ξ(y − d) + 1)
ξ
≥ (2y − h − 2d) − 2k .
2
Thus, we obtain

fξ (d, y) − fξ (d, y − h)
lim inf ≥ ξ(y − d) − 2k .
h→0 h

And on the other hand because fξ (d, d) < 0 there is ε > 0 such that for y ∈ [d, d + ε]
it holds fξ (d, y) < 0. For ξ ≥ 2k
ε , y ∈ [d + ε, ∞) we obtain fξ (d, y) ≤ x→∞
lim fξ (d, x) < 0.
 2k 4k
Letting ξ0 := max ε , M we obtain (xξ , yξ ) ∈ / ∂H for ξ ≥ ξ0 .
Consider now the functions
ξ 2k
ψ(x) = vk (yξ ) + (x − yξ )2 + 2 +C
2 ξ (yξ − x) + ξ
ξ 2k
φ(y) = u(xξ ) − (xξ − y)2 − 2 −C ,
2 ξ (y − xξ ) + ξ

where
ξ 2k
C = u(xξ ) − vk (yξ ) − (xξ − yξ )2 − 2 .
2 ξ (yξ − xξ ) + ξ

119
4 Optimal Control of Capital Injections by Investments

The functions ψ(x), φ(y) are twice continuously differentiable. Furthermore, u(x)− ψ(x)
attains its maximum at xξ ; vk (y) − φ(y) attains its minimum at yξ . Also it holds

ψ ′ (xξ ) = φ′ (yξ ) and 0 < ψ ′′ (xξ ) = −φ′′ (yξ ) . (4.16)

Therefore, it holds per Definition 4.2.4 of viscosity sub- and supersolutions

Z ∞
m2 ψ ′ (xξ )2
− 2 ′′ +λ u(xξ − z) dG(z) + cψ ′ (xξ ) − (δ + λ)u(xξ ) ≥ 0
2σ ψ (xξ ) 0
Z ∞
m2 φ′ (yξ )2
− 2 ′′ +λ vk (yξ − z) dG(z) + cφ′ (yξ ) − (δ + λ)vk (yξ ) ≤ 0 .
2σ φ (yξ ) 0

From above and from (4.16) it follows

Z ∞
m2 ψ ′ (xξ )2 λ+δ
− 2 ′′ + u(xξ − z) − vk (yξ − z) dG(z) ≥ (u(xξ ) − vk (yξ )) (4.17)
λσ ψ (xξ ) 0 λ

It holds (xξ , xξ ), (yξ , yξ ) ∈ H and

fξ (xξ , xξ ) + fξ (yξ , yξ ) ≤ 2fξ (xξ , yξ ) .

Thus, we obtain

ξ(xξ − yξ )2 ≤ u(xξ ) − u(yξ ) + vk (xξ ) − vk (yξ ) + 4k(yξ − xξ ) .

From (4.15) we conclude, that

ξ(xξ − yξ )2 ≤ 6k|xξ − yξ | . (4.18)

Choose a sequence ξn → ∞ as n → ∞ such that (xξn , yξn ) → (x̄, ȳ) ∈ H. From (4.18)
we obtain |xξn − yξn | ≤ 6k 2
ξn , from which it follows x̄ = ȳ; and ξn (xξn − yξn ) → 0. It holds
as ξn → ∞:

2 2
ψ ′ (xξn )2

ξn (xξn − yξn ) · ξn (yξn − xξn ) + 1 + 2k
= →0.
ψ ′′ (xξn ) ξn (ξn (yξn − xξn ) + 1)3 + 4kξn
Thus (4.17) becomes
Z ∞
λ+δ
M≥ u(x̄ − z) − vk (ȳ − z) dG(z) ≥ (u(x̄) − vk (ȳ)) .
0 λ
Together with lim inf Mξ ≥ M we obtain
ξ→∞

λ
M ≤ lim inf Mξ ≤ lim Mξn = u(x̄) − vk (ȳ) ≤ M,
ξ→∞ ξn →∞ λ+δ

120
 4.2 The Classical Risk Model

which is an obvious contradiction if δ > 0.

Suppose now, that δ = 0. From above it follows, that lim inf Mξ = M and there is some
ξ→∞
sequence (ξn )n≥0 such that ξn → ∞ and (xξn , yξn ) → (x∗ , x∗ ) as n → ∞. Note, that
for every n ∈ N it holds fξn (x∗ , x∗ ) < 0, i.e. there is ε > 0 with fξn (x, y) ≤ 0 for all
(x, y) ∈ Uε (x∗ , x∗ ), where
n o
Uε (x∗ , x∗ ) = (x, y) ∈ H : ||(x − x∗ , y − x∗ )||2 < ε

denotes the open ball with center in (x∗ , x∗ ), radius ε and ||.|| denotes the Euclidean
norm. But on the other hand for every ε > 0 we can find n0 ∈ N such that

||(xξn − x∗ , yξn − x∗ )||2 < ε

for all n ≥ n0 . This contradicts the fact Mξ > 0 for all ξ > ξ0 . Thus, we have shown the
uniqueness of the value function also in the case δ = 0. 

A direct consequence of the previous propositions is

Corollary 4.2.8
There is a unique viscosity solution to (4.9) with initial condition V (0) = v(0).

4.2.3 The special case δ = 0.

Let us now consider the special case δ = 0, i.e. the insurer is indifferent to investing
today or tomorrow.
The Hamilton–Jacobi–Bellman equation has the form
Z ∞
σ 2 a2 ′′
inf V (x) + λ V (x − z) dG(z) + (c + am)V ′ (x) − λV (x) = 0 , (4.19)
a∈R 2 0

or equivalently applying integration by parts:

Z ∞
σ 2 a2 ′′
inf V (x) − λ V ′ (x − z)(1 − G(z)) dz + (c + am)V ′ (x) = 0 .
a∈R 2 0

Under some additional constraints we will be able to show, that the value function V (x)
is two times continuously differentiable, solves the HJB equation and is unique. Note,
that for V ′′ (x) = 0 the HJB equation would not have any solution, because otherwise
there would exist a > 0 with
Z ∞
λ V (x − z) dG(z) + (c + am)V ′ (x) − λV (x) < 0 .
0

Thus we are searching for the function V (x), which is strictly convex.

121
4 Optimal Control of Capital Injections by Investments

As for the optimal strategy, it is clear, that the optimal strategy A∗ should be given
′ (x)
by a∗ (x) = − σmV
2 V ′′ (x) . Because we are searching for V (x), which is decreasing, strictly

convex and two times continuously differentiable on (0, ∞), it follows readily a∗ (x) ≥ 0
and a∗ (x) is continuous on (0, ∞). Plugging in this ansatz into the HJB equation yields
the integro-differential equation
Z ∞
m2 V ′ (x)2
− 2 ′′ +λ V (x − z) dG(z) + cV ′ (x) − λV (x) = 0 . (4.20)
2σ V (x) 0
Next we prove the verification theorem.
Theorem 4.2.9 (Verification theorem)
Let f (x) be a decreasing, vanishing at infinity, twice continuously differentiable solution
to (4.19), then f (x) = V (x).
Proof: Let A = {at } be an arbitrary admissible strategy. Let further X ∗ denote the
surplus process for the optimal strategy and X̂ the process X A,Y . Choose n > x and
define Sn = inf{t ≥ 0 : X̂t ∈ / [0, n]}, then it holds X̂Sn ∧t ∈ [0, n], because X̂t has
downward jumps only; in particular we obtain X̂Sn = n. We know from Proposition
A.2.2, p. 165, that the process
NSn ∧t Z Sn ∧t  Z ∞ 
X

Mt = f (X̂Ti ) − f (X̂Ti − ) − λ f (X̂s − z) dG(z) − f (X̂s ) ds
i=1 0 0

is a martingale. Rearranging the terms we obtain

NSn ∧t
X

f (X̂Sn ∧t ) = f (X̂Sn ∧t ) − f (X̂TNS ) + f (x) + f (X̂Ti − ) − f (X̂Ti−1 )
n ∧t
i=1
Z Sn ∧t  Z ∞ 
+ Mt + λ f (X̂s − z) dG(z) − f (X̂s ) ds . (4.21)
0 0
On the other hand it holds with Ito’s formula
Z Ti 2
σ 2 ′′
f (X̂Ti − ) − f (X̂Ti−1 ) = as f (X̂s ) + (c + mas )f ′ (X̂s ) ds
Ti−1 2
Z Ti Z Ti
′
+ σas f (X̂s ) dWs + f ′ (X̂s ) dYsA
Ti−1 Ti−1
Sn ∧t
σ2
Z
f (X̂Sn ∧t ) − f (X̂TNS ) = a2s f ′′ (X̂s ) + (c + mas )f ′ (X̂s ) ds
n ∧t
TNS 2
n ∧t
Z Sn ∧t
+ σas f ′ (X̂s ) dWs
TNS
n ∧t
Z Sn ∧t
+ f ′ (X̂s ) dYsA .
TNS
n ∧t

122
 4.2 The Classical Risk Model

Plugging in these results into the expression (4.21) we obtain

Z Sn ∧t  2 2
σ as ′′
f (X̂Sn ∧t ) = f (x) + f (X̂s ) + (c + mas )f ′ (X̂s )
0 2
Z ∞ 
+λ f (X̂s − z) dG(z) − λf (X̂s ) ds
0
Z Sn ∧t Z Sn ∧t
+ σas f ′ (X̂s ) dWs + Mt + f ′ (X̂s ) dYsA .
0 0

By HJB Equation (4.19) we obtain

Z Sn ∧t Z Sn ∧t
′
f (X̂Sn ∧t ) ≥ f (x) + σas f (X̂s ) dWs + Mt + f ′ (X̂s ) dYsA ,
0 0
R S ∧t
and for X ∗ holds the equality. The stochastic integral 0 n σas f ′ (X̂s ) dWs is a local
martingale. Let {Km } be a localisation sequence for it and let Jnm = Sn ∧ Km . Then we
have for the expectations
h Z Jnm ∧t i
Ex [f (X̂Jnm ∧t )] − Ex f ′ (X̂s ) dYsA ≥ f (x) .
0

Letting m and t go to infinity we obtain by the proposition of Lebesgue

h Z Sn i h Z Sn i
Ex [f (X̂Sn )] − Ex f ′ (X̂s ) dYsA = f (n) − Ex f ′ (X̂s ) dYsA ≥ f (x) .
0 0

Because f (x) is Lipschitz continuous with |f (x) − f (y)| ≤ |x − y|, it holds −f ′ (x) ≤ 1
and we obtain
h Z Sn i
f (n) + Ex dYsA ≥ f (x)
0
R∞
dYsA ≥ f (x). For X ∗ holds the equality, which

Letting n go to infinity yields Ex 0
means V (x) = f (x). 

It remains to show, that there exists a decreasing and twice continuously differentiable
solution to (4.19). We will see, that it is possible to show the existence only under the
constraint, that the distribution function of the claims amounts is absolutely continuous
and the density function is bounded.
But first of all we consider the question of optimal strategy on (−∞, 0]. It is clear,
that for x < 0 the optimal strategy should be given by a∗ (x) = 0. In the same model
by minimising the ruin probability, compare Schmidli [70, p. 56], it was obvious, that
the optimal strategy for x = 0 was a∗ (0) = 0, because otherwise the ruin would occur
immediately due to the oscillation of Brownian motion. But in our case it is not self-
explanatory, what is the optimal strategy at x = 0.

123
4 Optimal Control of Capital Injections by Investments

Lemma 4.2.10
Assume the value function V (x) exists, the net profit condition c > λµ is fulfilled and
the claim size distribution G(x) has a bounded density, then the optimal strategy at
x = 0 is a∗ = 0.
Proof: Arguing by contradiction we assume a∗ > 0 and denote by V (x) the value
function.
Denote the function, which results from choosing the start strategy a(0) = 0 and solving
the HJB equation (4.19) by f (x). It is clear, that f ′′ (0) = ∞ and f ′ (0) = − λµ
c . With
′′
the same arguments as in Remark 4.2.5 we obtain f (x) > 0 for all x ∈ R+ . Let further
V 0 (x) define the return function corresponding to the constant strategy A ≡ 0. From
Example 2.2.1 we know, that if the claim size distribution G(x) is absolutely continuous
and the density function is bounded, V 0 (x) is twice continuously differentiable with
(V 0 )′ (0) = − λµ ′ 0 ′′ c−λµ
c = f (0), (V ) (0) = λ c2 < ∞ and solves the integro-differential
equation Z ∞
c(V 0 )′ (x) + λ V 0 (x − z) dG(z) − λV 0 (x) = 0 .
0

Thus we obtain for h(x) := f (x) − V 0 (x)

Z ∞
′
ch (x) ≥ −λ h(x − z) dG(z) + λh(x)
0



Z x
= λ h(x) − h(0) 1 − G(x) + λ h(x) − h(x − z) dG(z) .
0

Note, that it holds h′ (0) = 0 and h′′ (0) = ∞, from which it follows h′ (x) > 0 for x
small enough. Let x̂ = inf{x > 0 : h′ (x) < 0}. Then it holds h′ (x) > 0 on (0, x̂), which
implies ch′ (x) ≥ λ(h(x) − h(0)) on (0, x̂). We can conclude h(x) > 0 for all x > 0.
Thus we obtain lim inf f (x) − V 0 (x) ≥ f (0) − V 0 (0). Because |f (0) − V 0 (0)| is finite and
x→∞
lim V 0 (x) = 0 we conclude, that lim f (x) = −∞ is impossible.
x→∞ x→∞
Define further
g(x) := V (x) − f (x) .
It is obvious that it holds
λµ λµ
V ′ (0) = − >− = f ′ (0)
c + a∗ m/2 c

as well as 0 < V ′′ (0) < ∞. Thus g(x) is twice continuously differentiable, g′ (0) > 0 and
g′′ (0) = −∞ by definition. Let further a(x) denote the optimal strategy for f (x), then
we obtain from the HJB equation
Z x
a(x)2 σ 2 ′′ ′
g (x) + (c + ma(x))g (x) − λ g′ (x − z)(1 − G(z)) dz ≥ 0 .
2 0

124
 4.2 The Classical Risk Model

From f ′′ (0) = ∞ and from

∞
ma(x) ′
Z
− f (x) = cf ′ (x) − λ f ′ (x − z)(1 − G(z)) dz
2 0
we obtain, that a(x) is strictly increasing close to zero, i.e. a(x) > 0 for x ∈ [0, ε) for
some ε > 0. Let x̌ = inf{x : g′′ (x) > 0} ∧ inf{x > ε : a(x) = 0}, then
Z x
a(x)2 σ 2 ′′
0> g (x) ≥ λ g′ (x − z)(1 − G(z)) dz − (c + ma(x))g′ (x)
2 0
≥ (λµ − c − ma(x))g′ (x) .
on [0, x̌). Thus it holds g′ (x) ≥ 0 for all x ∈ R+ . We conclude, that the function g(x)
is increasing, which implies f (x) is decreasing. In particular we have either lim f (x) =
x→∞
−∞ or lim f (x) = d, with |d| < ∞. We have already shown above, that the first case is
x→∞
impossible. In the second case consider the function f1 (x) = f (x) − d. f1 (x) solves the
HJB equation for the strategy a(x), is twice continuously differentiable, decreasing with
lim f1 (x) = 0 and satisfies f1′ (0) = − λµ
c , which is a contradiction to Theorem 4.2.9. 
x→∞

Thus we can assume a∗ (0) = 0, i.e. V ′ (0) = − λµ

c . Using this we show the existence and
uniqueness of the value function.
Lemma 4.2.11
If G(x) has a bounded density, then there exists ξ > 0, such that there is a solution f (x)
√ λµm √ √
o( x)
to (4.19) on [0, ξ] and f ′ (x) = − λµ
c + x σc3/2 + o( x) as x ↓ 0, where lim
√
x
= 0.
x↓0

Proof: The proof follows closely the proof in Hipp and Plum [41] see also Schmidli [70, p.
59] and is organised as follows. At first we construct a contraction on the interval [0, ξ]
and show the existence of the value function on the half-open interval [0, ξ) by Banach
fixed point theorem. Then we extend the solution to the closed interval [0, ξ].
Rewrite at first Equation (4.20) as follows using integration by parts
Z ∞
m2 V ′ (x)2
− 2 ′′ −λ V ′ (x − z)(1 − G(z)) dz + cV ′ (x) = 0 .
2σ V (x) 0
Let now h(x) = V ′ (x) and rewrite the above equation in order to get an expression for
h′ (x): Z ∞
m2 h(x)2 h i−1
h′ (x) = − λ h(x − z)(1 − G(z)) dz + ch(x) .
2σ 2 0
The ansatz p(x) = h(x2 ) yields:
Z x
′ m2 xp(x)2 n
zp(z) 1 − G(x2 − z 2 ) dz


p (x) = 2
− 2λ
σ 0
Z ∞ o−1
+λ 1 − G(z) dz + cp(x) . (4.22)
x2

125
4 Optimal Control of Capital Injections by Investments

|f ′ (x)−f ′ (0)|
Let further K(f ) := sup x and L := {f ∈ C 1 [0, ξ] : K(f ) < ∞}. Define the
0<x<ξ
norm
||f || := max{||f ||∞ , |f ′ (0)|, ξK(f )} ,
where ||f ||∞ is the supremum norm. With this norm the space L is complete. Suppose
further V ′ (0) = − λµ
c and let
n λµ mλµ λµ λµ o
Dl := f ∈ L : f (0) = − , f ′ (0) = 3/2 , || + f ||∞ ≤ , K(f ) ≤ l ;
c σc c 3c
the set Dl is a closed set. Define an operator F acting on the set Dl :
Z x 2 Z y
m yf (y)2 n 2 2


F (f (x)) : = − 2λ zf (z) 1 − G(y − z ) dz
0 σ2 0
Z ∞ o−1 λµ
+λ 1 − G(z) dz + cf (y) dy − .
y2 c

One can find pairs (l, ξ) (where l is large enough and ξ small enough), such that Dl is
mapped into itself and F is a contraction. By Banach fixed point theorem it follows,
that there exists a fixed point p̃(x). Because p̃ ∈ Dl it follows p̃(x) ≤ − 2λµ
3c , from which
it follows, that the derivative is bounded. The derivative p̃′ (x) can become negative
only if m2 xp̃(x)2 = 0 for some x > 0. But it is impossible because the derivative is
√
bounded. Thus p̃′ (x) > 0 on (0, ξ]. Moreover the function h(x) = p̃( x) is then the
√ √ √
desired solution and it holds f ′ (x) = p̃( x) = − λµ λµm
c + x σc3/2 + o( x) as x ↓ 0.
Consider now again Equation (4.20) and let h(x) = V ′ (x). Rearranging the terms and
building the reciprocal yields
Z x Z ∞
h′ (x) m2 n o−1
− = λ h(x − z)(1 − G(z)) dz − λ 1 − G(z) dz − ch(x) ,
h(x)2 2σ 2 0 x2

from which it follows

x Z y
1 c m2 n
Z
=− + 2
λ h(y − z)(1 − G(z)) dz
h(x) λµ 0 2σ 0
Z ∞ o−1
−λ 1 − G(z) dz − ch(y) dy .
y2

Thus
x y
m2 n

c
Z Z
h(x) = − + λ h(y − z)(1 − G(z)) dz
λµ 0 2σ 2 0
Z ∞ o−1 −1
−λ 1 − G(z) dz − ch(y) dy .
y2

126
 4.2 The Classical Risk Model

Every solution to Equation (4.22) solves also the above equation. Thus there is a strictly
negative, increasing solution to the above equation on some interval [0, ξ) with ξ > 0.
For this solution, say p(x), it holds p(x) < 0 on [0, ξ). If we can show, that
Z ξ Z ∞
λ p(ξ − z)(1 − G(z)) dz − λ 1 − G(z) dz − cp(ξ) < 0 ,
0 ξ2

then p(ξ) < 0; and the solution can be extended to [0, ξ]. Suppose now, that
Z ξ Z ∞
λ p(ξ − z)(1 − G(z)) dz − λ 1 − G(z) dz − cp(ξ) = 0 .
0 ξ2

Then we obtain p′ (ξ) = ∞. On the other hand it holds for all x ∈ [0, ξ):
Rx R∞
λ 0 p(z)(1 − G(x − z)) dz − λ x2 1 − G(z) dz − cp(x)
0 >
ξ−x
Rξ R∞
λ 0 p(z)(1 − G(ξ − z)) dz − λ ξ 2 1 − G(z) dz − cp(ξ)
−
ξ−x
Z ξ
(G(ξ − z) − G(x − z)) dz p(ξ) − p(x)
≥ λ p(z) +c
0 ξ−x ξ−x
Z ξ2
1 − G(z)
+λ .
x2 ξ−x

Because G(x) has a bounded density and p(x) is bounded, the above expression goes to
∞ as x → ξ. Thus we obtain 0 > ∞, which is a contradiction and we can extend the
solution to [0, ξ]. 

Theorem 4.2.12
If G(x) has a bounded density, then there is a twice continuously differentiable solution
to (4.19).

Proof: Let ξ be the largest value such that there is a solution p(x) to (4.19) on [0, ξ].
From Lemma 4.2.11 we know, that ξ > 0. Assume now ξ < ∞. Let then

ξ m2
c
Z
2σ2
d1 := − + Ry R∞ dy
λµ 0 λ 0 p(y − z)(1 − G(z)) dz − λ y2 1 − G(z) dz − cp(y)

and Z ξ Z ∞
d2 := λ p(ξ − z)(1 − G(z)) dz − λ 1 − G(z) dz − cp(ξ) < 0 .
0 ξ2

127
4 Optimal Control of Capital Injections by Investments

Define an operator F on the set of continuous, negative and increasing functions h(x)
m2 x−ξ −1


with domain [0, ξ + ζ), where h(x) = p(x) on [0, ξ] and h(x) ≥ 2 d1 + 2σ 2 d
2
on
[ξ, ξ + ζ):
1
F (p(x)) = Rx n o−1 ,
m2
d1 + ξ 2σ2 U (y) ∨ d2 dy

where Z y Z ∞
U (y) = λ p(y − z)(1 − G(z)) dz − λ 1 − G(z) dz − cp(y) .
0 y2

The parameter ζ we will be chosen later.

Let further h1 and h2 be two functions in the domain of the operator F with
h Z x Z ∞ i
Hi (x) = λ hi (x − z)(1 − G(z)) dz − λ 1 − G(z) dz − chi (x) ∨ d2 ,
0 x

i ∈ {1, 2}. For x > ξ we obtain

m2 x −1 − H (y)−1 | dy
R
2σ 2 ξ |H1 (y) 2
|F (h1 (x)) − F (h2 (x))| ≤ 2   .
R x m2
−1 dy
Q
d1 + ξ 2σ 2 Hi (y)
i=1

We can estimate the integral in the nominator as follows

Z x Z x
|H1 (y)−1 − H2 (y)−1 | dy ≤ d−2
2 |H1 (y) − H2 (y)| dy
ξ ξ
1
≤ d−2 2
2 [(x − ξ) + (x − ξ) ] sup |h1 (y) − h2 (y)| .
2 ξ≤y≤ξ+ζ

Thus for ζ small enough F is a contraction and there is a fixed point h(x). By continuity
one can choose ζ so small, that
Z x Z ∞
λ h(x − z)(1 − G(z)) dz − λ 1 − G(z) dz − ch(x) < d2
0 x2

for all x ∈ [ξ, ξ + ζ]. Therefore, there is a solution to (4.19) also on [ξ, ξ + ζ), which is a
contradiction to the assumption, that [0, ξ] is the largest interval. This proves our claim
ξ = ∞. 

Next we consider the asymptotic behavior of the value function. In order to characterise
V (x) for x → ∞ we need the concept of moment generating function, defined as

MZ (r) = E[erZ ]

128
 4.2 The Classical Risk Model

for r > 0. We distinguish between the case, where MZ (r) < ∞ for some r > 0, and the
case, where MZ (r) = ∞ for all r > 0. The characteristics of the value function stated
below have been already proven for the ruin probability as a function of initial capital
x. Since the proofs will be perfectly similar we will skip them and just refer to Schmildi
[70] and Hipp and Plum [41].

Remark 4.2.13
Consider once again the return functions V a (x), corresponding to the constant strategies
A ≡ a. In Gerber and Landry [32] it was shown, that in the case there is R > 0 such
that Z ∞
MZ (R) = eRy dG(y) = 1 ,
0
√
mR+ m2 R2 +2σ2 R2 (δ+cR)
i.e. for a = σ2 R2

∞
a2 σ 2 2
Z
λ + δ + (c + ma)R − R =λ eRy dG(y)
2 0

is fulfilled, and it holds V a (x) ∼ C1 e−Rx with some constant C1 > 0. Thus, we obtain
V (x) . C1 e−Rx . Due to Example 2.2.1 the return function corresponding to the constant
strategy A ≡ 0 is given by the function 1+ −R̃
R̃ R̃x
e , where R̃ is the unique positive solution
to the equation Z ∞
λ − cR̃ = λ e−R̃x dG(x) . (4.23)
0

1+R̃ −R̃x
Thus, V (x) ≤ −R̃
e .

The following lemma deduces the behavior of the value function and the optimal strategy
in the case MZ (r) = ∞ for all r > 0.

Lemma 4.2.14
Assume MZ (r) = ∞ for all r > 0. Then for any r > 0 it holds

lim sup V (x)erx = ∞ .

x→∞

For proof see Schmidli [70, p. 181]. Like in Hipp and Plum [41] we obtain

Proposition 4.2.15
Assume MZ (r) = ∞ for all r > 0. Then the optimal strategy a∗ (x) is unbounded, i.e.
lim sup a∗ (x) = ∞.
x→∞

For more special constraints we obtain

129
4 Optimal Control of Capital Injections by Investments

Proposition 4.2.16
If the hazard rate G′ (y)/(1 − G(y)) goes to 0 as y → ∞, µ = E[Z] < ∞ and

x
(1 − G(x − y))(1 − G(y))
Z
lim dy = 2µ ,
x→∞ 0 1 − G(x)

then lim a∗ (x) = ∞.

x→∞

For the proof we also refer to Schmidli [70, p. 183].

4.2.4 Examples
Example 4.2.17 (Exponentially distributed claim sizes)
Consider the special case, where the claims are exponentially distributed, G(x) = 1 −
ex/µ . In this case our findings coincide with the considering of ruin probability as a
function of initial capital x.
Let now µ = λ = 1, m = 0.03, σ 2 = 0.01 and c = 1.3. Then we obtain from the Equation
(4.20), letting V ′ (x) = h(x), the following integro-differential equation:

x
m2 h(x)2
Z
y
−x x
−µ
− 2 ′ − λe µ h(y)e µ dy + ch(x) + λµe =0,
2σ h (x) 0

which is also the HJB equation for the ruin probability, compare formula (2.19) in
Schmidli [70, p. 58]. The numerical solution to the above equation, integrated from x to
infinity, is given in Figure 4.6. The optimal investment strategy is given in Figure 4.7.
We see, that the optimal investment behaves like stated in Lemma 4.2.11 near zero.
Let
1
C− = inf , (4.24)
z E[eR(Z−z)] |Z > z]

where the infimum is taken over the set {z : P[Z > z] > 0}, R = sup R(a) with R(a) the
a
unique positive solution to the equation
∞
a2 m2 R(a)2
Z
λ eR(a)z dG(z) − λ − (c + am)R(a) + =0.
0 2σ 2

It was shown in Schmidli [70, p. 170], that in the case C− > 0, the optimal strategy
A∗ (x) fulfils lim a∗ (x) = â with some â ≥ 0. We see in Figure 4.7, that the optimal
x→∞
strategy converges approximately to 9.5 for x → ∞. Note, that because V ′ (0) = − λµ
c
we can calculate the initial value V (0) exactly. For the above parameters we have
V (0) = 2.2787. 

130
 4.2 The Classical Risk Model

Example 4.2.18 (Pareto(2, µ)-distributed claim sizes)

µ2
Consider now Pareto(2, µ) distributed claim sizes, i.e. G(x) = 1 − (µ+x) 2 . We have to

solve the integro-differential equation

Z x
m2 h(x)2 µ2 µ2
− 2 ′ −λ h(y) dy + ch(x) + λ =0,
2σ h (x) 0 (µ + (x − y))2 µ+x

where again h(x) = V ′ (x). For the parameters µ = λ = 1, m = 0.03, σ 2 = 0.01 and
c = 1.3 numerical calculated V (x) and the optimal strategy are given in Figures 4.8 and
4.9 respectively. Note, that Pareto distribution is in the class of distributions, described
in Proposition 4.2.16. It means the optimal investment strategy goes to infinity as
x → ∞. In Figure 4.9 we can see, that a∗ (40) ≈ 70. Using V ′ (0) = − λµ c we obtain
V (0) = 4.4271 for initial value. 

131
4 Optimal Control of Capital Injections by Investments

0 5 10 15 20 25

Figure 4.6: V (x) for exponentially distributed claims.

0
0 1 2 3 4

Figure 4.7: Optimal Strategy A∗ for exponentially distributed claim sizes.

132
 4.2 The Classical Risk Model

Figure 4.8: V (x) for Pareto(2, µ)-distributed claim sizes.

60

50

40

30

20

10

0
0 10 20 30

Figure 4.9: The optimal Strategy A for Pareto(2, µ)-distributed claim sizes.

133
4 Optimal Control of Capital Injections by Investments

134
5 Optimal Control of Capital Injections by
Reinsurance and Investments, modeled
as a Black-Scholes Model
In following we complexify the previous model by adding the possibility of reinsurance.
In the examples we will show, that an optimal reinsurance treaty can lower the costs
considerably. As in the previous section we start with the simple case of a diffusion
approximation to the classical risk model.

5.1 Diffusion Approximation

We consider a surplus process where the mean number of claims in a time unit is λ and
the mean size of a claim is µ = E[Z], where Z is a generic random variable describing the
claim size. We assume that µ2 = E[Z 2 ] < ∞. The premium is c = (1 + η)λµ for some
η ∈ R. The insurer can now buy reinsurance. That is, a retention level b ∈ [0, b̃] has to
be chosen, where b̃ ∈ (0, ∞]. The cedent pays then 0 ≤ r(Z, b) ≤ Z. The reinsurance


premium is calculated by an expected value principle (1 + θ)λ µ − E[r(Z, b)] ; i.e. the
premium rate remaining for the insurer is c(b) = λ(1+θ)E[r(Z, b)]−λµ(θ−η). Here b = 0
means “full reinsurance” and b = b̃ means “no reinsurance”. We assume for simplicity
that r(z, b) is increasing and continuous in b and z. In order that the problem does not
have the trivial solution V (x) = |x|1I[x<0] we assume that θ > η. Here 1IA denotes the
indicator function.
For a constant strategy, the diffusion approximation to the surplus process is
 p
x + λ(θE[r(Z, b)] − (θ − η)µ)t + λE[r(Z, b)2 ]Wt ,

compare Section 1.2. As an extension, the insurer can continuously change the reinsur-
ance leading to the process
Z t Z tp
B


Xt = x + λ θE[r(Z, bs )] − (θ − η)µ ds + λE[r(Z, bs )2 ] dWs . (5.1)
0 0

For the invetsments, the risky asset is again modeled as a Black-Scholes model
n σ2
o
Qt = exp m− t + σ W̃t ,
2

135
5 Optimal Control of Capital Injections by Reinsurance and Investments

where m, σ > 0. We consider the return of such a process, which is then the stochastic
process {Q′t } given by the stochastic differential equation

dQ′t = m dt + σ dW̃t .

We assume further, that the Brownian motions Wt and W̃t are independent and denote
the filtration generated by the couple {(Wt , W̃t )} by F = {Ft }.
The amount at ∈ R invested at time t can be changed continuously. We call every
strategy (A, B), where A = {at } and B = {bt }, admissible if A and B are cadlag and F
adapted. The surplus process under the reinsurance strategy B = {bt } and investment
strategy A = {at } fulfils the stochastic differential equation:
n o
dXtA,B = λµ(η − θ) + λE[r(Z, bt )]θ + m · at dt + λE[r(Z, bt )2 ] dWt
p

+σat dW̃t .

Note, that here we do not need to assume the net profit condition. We can always find
an investment strategy A = {at }, such that the underlying process X A,B has a positive
drift.
By Y A,B = {YtA,B } we denote in following the process of cumulated capital injections
corresponding to the strategy (A, B). The underlying process {XtA,B } with capital in-
jections fulfils then
dXtA,B,Y = dXtA,B + dYt .
Our goal is to minimise the expected discounted capital injections over all admissible
strategies (A, B).
 R ∞We define the
 value associated to the strategy pair (A, B) by
V A,B (x) = Ex 0 e−δt dYtA,B , where δ ≥ 0 is the sense of Subsection 4.1, and we are
looking for the function
V (x) = inf V A,B (x)
(A,B)

and for the optimal strategy (A∗ , B ∗ ).

Assume for the moment, that the optimal strategy and the value function exist. Then
it is clear, that V (x) should be decreasing. Because the constant strategy (0, b̃) is
admissible and the corresponding surplus process has the form

Xt0,b̃ = x + λµηt + λµ2 Wt

p

we obtain from Example 2.1.1:

1
0 ≤ lim V (x) ≤ lim V 0,b̃ (x) = lim exp(−βx) = 0 ,
x→∞ x→∞ x→∞ β
where p
λµη + λ2 µ2 η 2 + 2λµ2 δ
β= .
λµ2

136
 5.1 Diffusion Approximation

We find in Subsection 2.1.1, that for every constant strategy (a, b) the corresponding
return function V a,b (x) solves the differential equation

a2 σ 2 + λE[r(Z, b)2 ] a,b ′′

(V ) (x) + (λµ(η − θ) + λE[r(Z, b)]θ + ma)(V a,b )′ (x) − δV a,b (x) = 0
2
(5.2)
′ a,b
for x ≥ 0, fulfils V (0) = −1 and lim V (x) = 0.
x→∞
1
It means V a,b (x) = β(a,b) exp(−β(a, b)x) for some β(a, b) > 0.
The Hamilton–Jacobi–Bellman equation for x ≥ 0 is
n σ 2 a2 + λE[r(Z, b)2 ] o
inf V ′′ (x) + (λµ(η − θ) + λE[r(Z, b)]θ + ma)V ′ (x) − δV (x) = 0 .
a∈R 2
b∈[0,b̃]

For the explicit derivation of HJB equation we refer to Section 2.1.2. Note, that from
above we obtain, that the optimal investment strategy in x ≥ 0 is given by a∗ (x) =
′ (x)
− σmV
2 V ′′ (x) . Using this we can simplify the HJB equation

m2 V ′ (x)2 λE[r(Z, b)2 ] ′′

n
′
o
inf − + V (x)+ (λµ(η − θ)+ λE[r(Z, b)]θ)V (x)− δV (x) =0.
b∈[0,b̃] 2σ 2 V ′′ (x) 2
(5.3)
Make now the ansatz V (x) = β1 exp(−βx), where β does not depend on x. I.e. we
conjecture, that the optimal strategies for both reinsurance and investment are constant.
Plugging in the function β1 exp(−βx) into the Hamilton–Jacobi–Bellman equation (5.3),
we obtain

m2 λE[r(Z, b)2 ] 2
inf − + β − (λµ(η − θ) + λE[r(Z, b)]θ)β − δ = 0 . (5.4)
b∈[0,b̃] 2σ 2 2

The above expression is a continuous function in b and therefore there is a b∗ , where the
minimum is attained. Note, that plugging in the return function
1
V a,b (x) = β(a,b) exp(−β(a, b)x) corresponding to the constant strategy (a, b) into Equa-
tion (5.2) yields

a2 σ 2 + λE[r(Z, b)2 ]
β(a, b)2 − (λµ(η − θ) + λE[r(Z, b)]θ + ma)β(a, b) − δ = 0 .
2
The above equation is just a quadratic equation in β(a, b), such that we can easily find
β(a, b). Define α(a, b) := λθE[r(Z, b)] − λµ(θ − η) + ma, then we obtain for β(a, b):
p
α(a, b) + α(a, b)2 + 2δ(λE[r(Z, b)2 ] + a2 σ 2 )
β(a, b) = . (5.5)
a2 σ 2 + λE[r(Z, b)2 ]

137
5 Optimal Control of Capital Injections by Reinsurance and Investments

We can maximise β(a, b) with respect to (a, b). Maximising β(a, b) with respect to a
m
yields immediatelly a = β(a,b)σ 2 . Plugging in this result into the above equation yields

λE[r(Z, b)2 ] m2
β(a, b)2 − (λµ(η − θ) + λE[r(Z, b)]θ)β(a, b) − δ − 2 = 0 . (5.6)
2 σ
Because the above equation is convex in β we can conclude, that (5.4) and (5.6) yield the
same solution and the same minimiser b∗ of (5.4) and the same maximiser b∗ of β(a, b),
respectively.
Next we prove our claim, that the optimal strategy is constant and the value function
is an exponential function in the verification theorem.
Theorem 5.1.1 (Verification theorem)
Optimal strategies are given by B ∗ ≡ b∗ and A∗ ≡ a∗ , where (a∗ , b∗ ) is a maximum point
of the function β(a, b) given in (5.5). The corresponding value function is
(
1
β(a∗ ,b∗ ) − x : x<0,
V (x) = 1 ∗ ∗
β(a∗ ,b∗ ) exp(−β(a , b )x) : x ≥ 0 .

Proof: From Subsection 2.1.1, we know, that the return function corresponding to the
constant strategy pair (a∗ , b∗ ) a∗ ∈ R, b∗ ∈ [0, b̃] is given by
(
1
a∗ ,b∗ β(a∗ ,b∗ ) − x : x<0,
V (x) = 1 ∗ ∗
β(a∗ ,b∗ ) exp(β(a , b )x) : x ≥ 0 ,
∗ ∗
where (a∗ , b∗ ) maximises β(a, b) given in (5.5). Note, that the function V a ,b (x) is
∗ ∗
the unique solution to the HJB equation with initial constraints (V a ,b )′ (0) = −1 and
∗ ∗
lim V a ,b (x) = 0. Consider further an arbitrary strategy (A, B) with A = {at } and
x→∞
B = {bt } and let X̂t = XtA,B,Y . The process X̂t fulfils then
 
dX̂t = λµ(η − θ) + λE[r(Z, bt )]θ + m · at dt + λµ2 dWt + σat dW̃t + dYtA,B .
p

Now we apply the Ito’s formula on V a,b (x) and obtain using the HJB equation:
Z t
−δt a∗ ,b∗ a∗ ,b∗ ∗ ∗
e V (X̂t ) = V (x) + e−δs (V a ,b )′ (X̂s ) dYsA,B
0
Z t h i
∗ ∗ ∗ ∗
+ e−δs Ds,A,B V a ,b (X̂s ) − δV a ,b (X̂s ) ds
0
Z t Z t
−δt a∗ ,b∗ ′ ∗ ∗
e−δt (V a ,b )′ (X̂s ) dWs
p
+σ e (V ) (X̂s )as dW̃s + λµ2
0 0
Z t
∗ ∗
≥ V a ,b (x) − e−δs dYsA,B
0
Z t Z t
−δt a∗ ,b∗ ′ ∗ ∗
e−δt (V a ,b )′ (X̂s ) dWs
p
+σ e (V ) (X̂s )as dW̃s + λµ2
0 0

138
 5.1 Diffusion Approximation

∗ ∗ ∗ ∗ ∗ ∗
where Ds,A,B V a ,b (x) = 12 (a2s σ 2 + λµ2 )(V a ,b )′′ (x) + (λµη + mas )(V a ,b )′ (x). Note,
∗ ∗
that for (A, B) = (a∗ , b∗R) the equality holds. Because√the derivative (V a ,b )′ (x) is
t ∗ ∗ t ∗ ∗
bounded, the processes σ 0 e−δt (V a ,b )′ (X̂s )as dW̃s and λµ2 0 e−δt (V a ,b )′ (X̂s ) dWs
R

are martingales with zero-expectations. Applying the expectations on the both sides of
the above inequality yields
 t −δs
Z
∗ ∗ ∗ ∗
E e−δt V a ,b (X̂t ) ≥ V a ,b (x) − E dYsA,B .
  
e
0

Because it holds V a∗ ,b∗ (X̂t ) ≤ V a∗ ,b∗

(0) we obtain by bounded convergence
 ∞ −δs
Z
a∗ ,b∗
dYsA,B .

V (x) ≤ E e
0

For X = X ∗ equality holds and thus V (x) = V (x). a∗ ,b∗

Note, that the uniqueness of the optimal strategy should be shown in every concrete
case. 

Next we illustrate the result by two examples.

Example 5.1.2 (Proportional reinsurance)

Consider at first proportional reinsurance, i.e. r(Z, b) = bZ. We have to minimise the
function
p
α(a, b) + α(a, b)2 + 2δ(σ 2 a2 + λµ2 b2 )
β(a, b) = ,
σ 2 a2 + λµ2 b2
where α(a, b) = λθµb + ma − λµ(θ − η).
m
The optimal investment should be given by a∗ = σ2 β
. Equation (5.6) reads then

λµ2 b2 2 m2
β − (λµ(η − θ) + λµbθ)β − δ − 2 = 0 .
2 σ
µθ
Minimising with respect to b yields β = µ2 b∗ . Plugging in this result into the above
equation gives the following:
λµ2 θ 2 m2 µθ
− − δ − 2 + λµ(θ − η) =0.
2µ2 σ µ2 b∗
Thus,
2θσ 2 λµ2 (θ − η)
b∗ = ∧1.
2σ 2 δµ2 + λθ 2 σ 2 µ2 + m2 µ2
It is straight forward to verify, that the optimal investment strategy is given by
mµ2 b∗
a∗ = .
σ 2 µθ

139
5 Optimal Control of Capital Injections by Reinsurance and Investments

V (a,b) (x)
V (4.0678,0.339) (x)
V (0,1) (x)
6
V (25,0.8) (x)

x
0 2 4 6 8
Figure 5.1: Return functions for proportional reinsurance for the optimal strategy
(a∗ , b∗ ) = (4.0678, 0.339), (a, b) = (0, 1) and (a, b) = (25, 0.8).

Note, that in the case of proportional reinsurance the function β(a, b) has a unique
maximum point (a∗ , b∗ ), which means the uniqueness of the optimal strategy.
Let λ = µ = 1, θ = 0.5, η = 0.3, m = 0.03, σ 2 = 0.01 and δ = 0.04. Then we have
β = 0.7374999999 and (a∗ , b∗ ) = (4.0678, 0.339). Figure 5.1 compares the value function,
the return function for (a, b) = (0, 1) and return function for (a, b) = (25, 0.8). We can
see that investments as well as reinsurance lower the costs considerably. But the choice
of a wrong strategy, see the dotted curve in Figure 5.1, can cause higher costs, than in the
case without reinsurance and investments. Thus the principle √ “It is in any case better
ληµσ+ λ2 η2 µ2 σ2 +2λσ2 δµ +λm2 µ
2 2
to reinsure and invest” does not work here. For θ ≥ λσµ the
a,1
function V (x) is the value function. The reinsurance is in this case too expensive. 
x
Example 5.1.3 (Excess of Loss Reinsurance and G(x) = 1 − e− µ .)
For the Excess of Loss reinsurance it holds r(z, b) = min{z, b}, where b ∈ [0, ∞]. Then
we obtain
Z b Z b
2
E[r(Z, b)] = (1 − G(z)) dz and E[r(Z, b) ] = 2 z(1 − G(z)) dz .
0 0
Equation (5.6) becomes
Z b Z b
2 m2
λ z(1 − G(z)) dz · β − (λµ(η − θ) + λθ (1 − G(z)) dz) · β − δ − 2 = 0 .
0 0 σ

140
 5.1 Diffusion Approximation

V (a,b) (x)
(3.1456,0.5243)
VXL (x)
V (0,∞) (x)
6
(4.0678,0.339)
VP (x)

x
0 2 4 6 8
Figure 5.2: Return functions for XL-strategy (a∗ , b∗ ) = (3.1456, 0.5243), proportional
strategy (a, b) = (4.0678, 0.339) and (a, b) = (0, ∞).

The derivative with respect to b is equal to zero for β = θb , provided 1 − G(θ/β) > 0.
Plugging in this result into the above equation gives
Z b Z b
η m 2 b2
λ z(1 − G(z)) dz − (λµ( − 1) + λ (1 − G(z)) dz) · b − (δ + 2 ) 2 = 0 . (5.7)
0 θ 0 σ θ
The left hand side of the above equation is concave in b, is equal to 0 for b = 0, tends to
−∞ for b → ∞ and the derivative at b = 0 is λµ(1 − ηθ ) > 0. Thus, there is a unique b∗
∗
solving the equation. The optimal investment strategy is then given by a∗ = mb σ2 θ .
In the case of proportional reinsurance we have found an upper boundary for the safety
loading θ of the reinsurer. If θ violates this level, it is optimal not to buy reinsurance,
Now we want to analyse, whether we can find such a boundary for the Excess of Loss rein-
surance. From the derivations above we get that it is always optimal to buy reinsurance
if the support of the claim size distribution is unbounded. Let b̃ = sup{x : P(Z > x) > 0}
and suppose b̃ < ∞. No reinsurance is chosen if b∗ ≥ b̃. This implies by the concavity
of the left hand side of (5.7) that
µ2 b̃ m2 b̃2
λ − ληµ − (δ + 2 ) 2 ≥ 0 .
2 θ σ θ
This is equivalent to
q
m2
b̃ λ2 µ2 η 2 + 2λµ2 (δ + σ2
)− λµη
≤ m2
,
θ 2(δ + σ2 )

141
5 Optimal Control of Capital Injections by Reinsurance and Investments

or q
2
θ λµη + λ2 µ2 η 2 + 2λµ2 (δ + m
σ2
)
≥ .
b̃ λµ2
We see that we do not reinsure if reinsurance is too expensive or if the maximal claim
size is small.
x
−µ
Assume now, that the claim sizes Zi are exponentially distributed: G(x) = 1 − e .
Then we have to maximise the function
q
b b

α(a, b) + α(a, b)2 + 2δ σ 2 a2 + λ(2µ2 − 2µ2 e− µ − 2µbe− µ )
β(a, b) = ,
−b −b
σ 2 a2 + λ(2µ2 − 2µ2 e µ − 2µbe µ )
b
where α(a, b) = λµθ(1 − e− µ ) + ma − λµ(θ − η).
Let again λ = µ = 1, θ = 0.5, η = 0.3, m = 0.03, σ 2 = 0.01 and δ = 0.04.
With help of mathematical programs one finds easily (a∗ , b∗ ) = (3.1456, 0.5243) and
β(3.1456, 0.5243) = 0.9537134001. Figure 5.2 compares the value function for Excess of
Loss (XL) reinsurance, the value function for proportional (P) reinsurance and the return
function corresponding to the constant strategy (0, ∞). One can see, that the choice
of Excess of Loss reinsurance reduces the costs more, than the choice of proportional
reinsurance. 

5.2 The Classical Risk Model.

Consider again the classical risk model, where the surplus process of the insurer is
given by
Nt
X
Xt = x + ct − Zi .
i=1

x denotes the initial capital, c - the premium income in a time unit, Nt is a Poisson
process with intensity λ and Zi the i-th claim amount. We let the insurer buy reinsurance
and invest money into a risky asset, modeled as a Black-Scholes model, simultaneously.
For the investments we use the setup of the Section 4.2.3, i.e. the surplus process under
some investment strategy A fulfils
Nt
X
dXtA = (c + mat ) dt − d Zi + at σ dWt .
i=1

As for the reinsurance we use the setup of Section 5.1 with the only difference, that
the premium rate remaining for the insurer is c(b) = c − λ(1 + θ)E[Z − r(Z, b)] =

142
 5.2 The Classical Risk Model.

λµ(η − θ) + λ(1 + θ)E[r(Z, b)]. We denote the filtration generated by the two dimensional
PNt
process ( Zi , Wt ) by {Ft }. Let B = {bt }t≥0 , b ∈ [0, b̃] be a reinsurance strategy and
i=1
A = {at }t≥0 , a ∈ R, an investment strategy. We call a strategy (A, B) admissible if it is
cadlag and {Ft } measurable and denote the set of admissible reinsurance strategies by
U and admissible investment strategies by A. The surplus process under reinsurance
and investments fulfils
Nt
dXtA,B
X
= (c(bt ) + mat ) dt − d r(Zi , bTi − ) + at σ dWt .
i=1

Our goal is to minimise the expected capital injections over all admissible reinsurance and
investment strategies. The surplus process {XtA,B } with corresponding capital injections
{YtA,B } solves then the stochastic differential equation

Nt
dXtA,B,Y = (c(bs ) + at m) dt − d r(Zi , bTi − ) + at σ dWt + dYtA,B .
X

i=1

R∞
We define V A,B (x) = E 0 e−δt dYtA,B and target to calculate the value function


V (x) = inf V A,B (x) and to find the optimal strategy pair (A∗ , B ∗ ). Like in the case
U
without reinsurance we state a couple of useful properties of the value function.

Lemma 5.2.1
The function V (x) satisfies
(1) V (x) is decreasing and fulfils lim V (x) = 0;
x→∞
(2) V (x) is Lipschitz continuous with |V (x) − V (y)| ≤ |x − y|.
(3)′ If we choose the proportional reinsurance, i.e. r(z, b) = bz the corresponding value
function is convex, provided the premium calculation principle is chosen such that c(b)
is concave in b.

For the proof see the proof of Lemma 4.2.2. Note, that in general the value function
would not be convex.

5.2.1 General case δ ≥ 0.

From Lemma 5.2.1 we know, that the value function V (x) is decreasing with
lim V (x) = 0 and Lipschitz continuous on [0, ∞) with |V (x) − V (y)| ≤ |x − y|. For
x→∞
the same reasons as in the case without reinsurance we will consider here the concept of

143
5 Optimal Control of Capital Injections by Reinsurance and Investments

viscosity solutions. We conjecture, that the value function V (x) solves the Hamilton–
Jacobi–Bellman equation
Z ∞
σ 2 a2 ′′
inf V (x)+λ V (x−r(z, b)) dG(z)+(c(b)+am)V ′ (x)−(δ+λ)V (x) = 0 . (5.8)
a∈R 2 0
b∈[0,b̃]

For explicit derivation of the HJB equation we refer to Subsections 2.2.1 and 4.2.2, see
also Schmidli [70, p. 55]. Note, that for every twice continuously differentiable function
f and continuous function u the minimum of
Z ∞
σ 2 a2 ′′
f (x) + λ u(x − r(z, b)) dG(z) + (c(b) + am)f ′ (x) − (δ + λ)u(x)
2 0

−f ′ (x)m
in a is attained at a = σ2 f ′′ (x)
. Using this fact we define like in Subsection 4.2.2

Definition 5.2.2
We say that a continuous function u : [0, ∞) → R+ is a viscosity subsolution to (5.8)
at x ∈ (0, ∞) if any twice continuously differentiable function ψ : (0, ∞) → R with
ψ(x) = u(x) such that u − ψ reaches the maximum at x satisfies

m2 ψ ′ (x)2 n Z ∞ o
− 2 ′′ + inf λ u(x − r(z, b)) dG(z) + c(b)ψ ′ (x) − (δ + λ)u(x) ≥ 0 , (5.9)
2σ ψ (x) b∈[0,b̃] 0

and we say that a continuous function ū : [0, ∞) → R+ is a viscosity supersolution

to (5.8) at x ∈ (0, ∞) if any twice continuously differentiable function φ : (0, ∞) → R
with φ(x) = ū(x) such that ū − φ reaches the minimum at x satisfies

m2 φ′ (x)2 n Z ∞ o
− 2 ′′ + inf λ ū(x − r(z, b)) dG(z) + c(b)φ′ (x) − (δ + λ)ū(x) ≤ 0 .
2σ φ (x) b∈[0,b̃] 0

A viscosity solution to (5.8) is a continuous function u : [0, ∞) → R+ if it is both a

viscosity subsolution and a viscosity supersolution at any x ∈ (0, ∞).
And the equivalent formulation

Definition 5.2.3
A continuous function u : [0, ∞) → R+ is a viscosity subsolution to (5.8) at x ∈ (0, ∞)
if any twice continuously differentiable function ψ : (0, ∞) → R with ψ(x) = u(x), such
that u − ψ reaches the maximum at x, satisfies

m2 ψ ′ (x)2 n Z ∞ o
− 2 ′′ + inf λ ψ(x − r(z, b)) dG(z) + c(b)ψ ′ (x) − (δ + λ)ψ(x) ≥ 0 , (5.10)
2σ ψ (x) b∈[0,b̃] 0

where ψ(−x) = ψ(0) − x for x ∈ (0, ∞).

144
 5.2 The Classical Risk Model.

A continuous function ū : [0, ∞) → R+ is a viscosity supersolution to (5.8) at x ∈ (0, ∞)

if any twice continuously differentiable function φ : (0, ∞) → R with φ(x) = ū(x), such
that ū − φ reaches the minimum at x, satisfies

m2 φ′ (x)2 n Z ∞ o
− 2 ′′ + inf λ φ(x − r(z, b)) dG(z) + c(b)φ′ (x) − (δ + λ)φ(x) ≤ 0 ,
2σ φ (x) b∈[0,b̃] 0

where φ(−x) = φ(0) − x for x ∈ (0, ∞).

The proof of equivalence is similar to the proof on page 110 and originates from Benth
et al. [8].
Given a twice continuously differentiable function f and a continuous function u, we
denote in following

m2 f ′ (x)2 n Z ∞ o
Lre (u, f )(x) = − 2 ′′ + inf λ u(x − r(z, b)) dG(z) + c(b)f ′ (x)
2σ f (x) b∈[0,b̃] 0
− (δ + λ)u(x) ,
m2 f ′ (x)2 n Z ∞ o
Lre (f )(x) = − 2 ′′ + inf λ f (x − r(z, b)) dG(z) + c(b)f ′ (x)
2σ f (x) b∈[0,b̃] 0
− (δ + λ)f (x) .

Next we will prove, that the value function is a viscosity solution. We note here again,
that the proof originates from Benth et al. [8].

Proposition 5.2.4
V is a viscosity solution to (5.8).
Proof: Here we only prove, that the value function is a supersolution.
For the proof, that the value function is a subsolution see Proposition 4.2.6.
Arguing by contradiction we assume, that V (x) is not a supersolution at x. Then due
to Definition 5.2.3 there exist ξ > 0 and a twice continuously differentiable function
φ0 : (0, ∞) → R with φ0 (x) = V (x) and V (y) − φ0 (y) ≥ 0 for all y > 0, such that

Lre (φ0 )(x) > 2ξ .

 
Consider now the function φ1 (y) := φ0 (y) − sin4 (x−y)π 2x · λξ . It is obvious, that φ1 is
 
twice continuously differentiable, φ1 (x) = V (x) and φ1 (y) ≤ V (y) − sin4 (x−y)π 2x
ξ
λ . We
also have
Lre (φ1 )(x) > ξ ,
R∞  
because φ′1 (x) = φ′0 (x), φ′′1 (x) = φ′′0 (x) and 0 sin4 yπ 2x dG(y) ≤ 1 .
In the proof of Proposition 4.2.6, we have used, that the value function is convex and

145
5 Optimal Control of Capital Injections by Reinsurance and Investments

according to this lim inf V ′ (x) = −1 was impossible, see Remark 4.2.5. Thus in the case
of proportional reinsurance we can just use the proof of Proposition 4.2.6. But if the
value function is not convex it may happen for some x ∈ R+ , that lim inf V ′ (x) = −1.
If φ1 (x) > −1 the proof of Proposition 4.2.6 holds.
Assume now φ1 (x) = −1. By construction of φ1 we obtain for a = −c(0)/m and b = 0:

a2 σ 2 ′′
φ (x) ≥ δφ1 (x) + ξ > 0 .
2 1
Thus φ′′1 (x) > 0, from which it follows φ′1 (y) < −1 for y < x small enough. Since
Lre (φ1 )(y) is continuous in y, c(b) and r(z, b) are continuous in b, there are h̃ ∈ (0, x2 )
and j ∈ N, j > 2, such that for all y ∈ [x − 2h̃, x + 2h̃] it holds
∞
m2 φ′ (y)2 n c(b)(j − 1) r(z, b)(j − 1) 
Z  o
− 2 1′′ + inf φ′1 (y) + λ φ1 y − dG(z)
2σ φ1 (y) b∈[0,b̃] j 0 j
ξ
−(λ + δ)φ1 (y) > . (5.11)
2
Because φ1 is twice continuously differentiable there is 0 < k < x such that 0 ≥ φ′1 (y) ≥
j
− j−1 and φ1 (y) > 0 on [x − k, x + k]. We let in following [x − 2h̃, x + 2h̃] ∩ [x − k, x + k] =
[x − 2h, x + 2h].
Define like in the proof of Proposition 4.2.6
1/n
1  hπ  ξ o
Z n
fn (y) = R 1 V (y − s) − sin4 nf (ns) ds .
2x 2λ
−1 f (s) ds −1/n

It is obvious, that

fn are smooth functions and the sequence (fn )n≥1 converges uniformly
4 hπ ξ
to V − sin 2x 2λ . Thus there is n0 ∈ N such that
 hπ  ξ  hπ  ξ
V (y) − sin4 ≤ fn0 (y) ≤ V (y) − sin4
2x λ 2x 4λ
for all y ≥ 0, also we obtain 0 ≥ fn′ 0 (y) ≥ −1 because of the Lipschitz continuity of
V and fn0 (y) ≥ 0 for y ∈ [0, x − h]. In particular it holds fn0 (y) ≥ φ1 (y) > 0 on
[x − 2h, x − h].
Let further g be a twice continuously differentiable function satisfying
π
• 2 ≤ g(y) ≤ π

• g(y) = π for y ∈ [x − h, x + h]
π
• g(y) = 2 for y ∈
/ (x − 2h, x + 2h)

• g′ (y) ≥ 0 for y ∈ [x − 2h, x − h].

146
 5.2 The Classical Risk Model.

and define l(y) = y(j−1)+x

j for y ≥ 0. The function l(y) has the following properties

• l [0, x + h] ⊂ [ xj , x + h(j−1) = [x − h(j−1) , x + h(j−1)

j ] and l [x − h, x + h] j j ]
• l is twice continuously differentiable
• l(x) = x
j−1
• l′ (y) = j .
It means there is κ > 1 such that φ1 (l(y))κ ≥ fn0 (y) on [x − 2h, x − h].
We define  hπ  ξ
n ξ o
ε = min sin4 ,
2x 12λ 4δ
and the function φ by
fn (y)
φ(y) = cos2 (g(y))φ1 l(y) + sin2 (g(y)) 0


.
κ
Obviously it holds φ(x) = φ1 (x) and for each y ∈ [x − h, x + h] we have φ(y) = φ1 (l(y)).
From (5.11) it follows readily Lre (φ)(y) > 2δε on [x − h, x + h]. Also it holds because l
is increasing and φ1 is strictly decreasing on [x − 2h, x]: φ1 (l(y)) ≤ φ1 (y) on [x − 2h, x].
Since

fn (y)
φ′ (y) = −2g′ (y) cos(g(y)) sin(g(y)) φ1 l(y) − 0 + cos2 (g(y))l′ (y)φ′1 l(y)


κ
f ′ (y)
+ sin2 (g(y)) n0
κ
we obtain φ′ (y) ≥ −1 for y ∈ [0, x + h]. For y ∈ [x − 2h, x − h] it holds
fn (y)
φ(y) = cos2 (s(g(y)))φ1 l(y) + sin2 (g(y)) 0


κ
2 2 fn0 (y)
≤ cos (g(y))φ1 (y) + sin (g(y))
κ
≤ V (y) − 3ε .
And for y ∈ [0, x − 2h]
fn0 (y)
φ(y) = ≤ V (y) − 3ε .
κ
Thus, we obtain
φ(y) ≤ V (y) − 3ε
for y ∈ [0, x − h] ∪ {x + h}.
The remaining part of the proof goes similar to the proof of Proposition 4.2.6. 

We have shown, that V (x) is a viscosity solution to (5.8). Next we show the uniqueness
of the value function with the technique one finds in Azcue and Muler [5].

147
5 Optimal Control of Capital Injections by Reinsurance and Investments

Proposition 5.2.5
Let now v(x) be a super- and u(x) a subsolution to (4.9), satisfying conditions (1) and
(2) from Lemma 5.2.1. If it holds u(0) ≤ v(0), then u(x) ≤ v(x) on [0, ∞).
Proof: Let u be a sub- and v a supersolution, which fulfil conditions (1) and (2) from
Lemma 5.2.1, i.e. u and v are decreasing, Lipschitz continuous and vanish at infinity.
Assume there is x0 ∈ (0, ∞) with u(x0 ) − v(x0 ) > 0. Define vk (x) = kv(x) for k > 1.
Then vk (x) is also a supersolution. Choose now k > 1 such that u(x0 ) − vk (x0 ) > 0.
Clearly u(0) ≤ vk (0). Because u and vk are decreasing and Lipschitz continuous we
obtain the following estimation
u(x) − vk (x) ≤ u(0) − vk (x) ≤ v(0) − vk (x)
= k(v(0) − v(x)) + (1 − k)v(0)
≤ kx + (1 − k)v(0) .

Thus it holds u(x) − vk (x) ≤ 0 for x ≤ v(0)(k−1)

k . So we will consider only the crucial
v(0)(k−1)
interval (d, ∞) with d = k .
Define further
M := sup(u(x) − vk (x)) .
x≥0

It follows readily
0 < u(x0 ) − vk (x0 ) ≤ M = sup (u(x) − vk (x)) .
x∈(d,∞)

Let x∗ be such, that M = u(x∗ ) − vk (x∗ ).

Define H := {(x, y) : d < y < ∞, d < x ≤ y} and for ξ > 0
ξ 2k
fξ (x, y) := u(x) − vk (y) − (x − y)2 − 2 .
2 ξ (y − x) + ξ
Let Mξ := sup fξ (x, y). Because fξ is continuous, there exists (xξ , yξ ) with Mξ =
(x,y)∈H
fξ (xξ , yξ ). Because x∗ > d it holds (x∗ , x∗ ) ∈ H. Thus, we have
2k 2k
Mξ ≥ fξ (x∗ , x∗ ) = u(x∗ ) − vk (x∗ ) − =M− .
ξ ξ
4k
We can therefore conclude, that Mξ > 0 for ξ > M and lim inf Mξ ≥ M .
ξ→∞
On the other hand we know, that (xξ , yξ ) ∈ H̄, where H̄ denotes the closure of H, from
which it follows yξ ≥ xξ .
Now we will show, that there exists ξ0 such that for all ξ ≥ ξ0 it holds (xξ , yξ ) ∈
/ ∂H.
The boundary ∂H is the union
  
(d, y) : y ∈ (d, ∞] ∪ (x, x) : x ∈ [d, ∞] ∪ (x, ∞) : x ∈ (d, ∞) .

148
 5.2 The Classical Risk Model.

It is clear, that for fixed x ∈ [d, ∞) it holds lim fξ (x, y) = −∞. Consider the set
y→∞

(x, x) : x ∈ [d, ∞] . Note, that it holds
2k 2k
lim fξ (x, x) = lim u(x) − vk (x) − =− <0
x→∞ x→∞ ξ ξ

and fξ (d, d) = u(d) − vk (d) − 2k

ξ < 0. For d < x1 ≤ x2 < ∞ we have because of the
properties (1) and (2) in Lemma 5.2.1
vk (x1 ) − vk (x2 )
0≤ ≤k. (5.12)
x2 − x1
So it holds

fξ (x, x) − fξ (x, x + h) u(x) − vk (x) − 2k

ξ − u(x) + vk (x + h) + 2ξ h2 + 2k
ξ 2 h+ξ
=
−h −h
vk (x) − vk (x + h) ξ 2k
= − − h+
−h 2 ξh + 1
ξ 2k
≥ −k − h + .
2 ξh + 1
fξ (x,x)−fξ (x,x−h)
Thus lim inf h > k, which means fξ (x, x) ≤ fξ (x, x + h) for some small
h→0 
h > 0. It remains to consider the set (d, y) : y ∈ (d, ∞] . For y > d it holds

fξ (d, y) − fξ (d, y − h) vk (y − h) − vk (y)

=
h h
ξ
− 2 (d + h − y)2 − 2k
ξ 2 (y−d−h)+ξ
+ 2ξ (d − y)2 + 2k
ξ 2 (y−d)+ξ
+
h
vk (y − h) − vk (y) ξ
= + (2y − h − 2d)
h 2
2k
−
(ξ(y − d − h) + 1)(ξ(y − d) + 1)
ξ
≥ (2y − h − 2d) − 2k .
2
Thus, we obtain
fξ (d, y) − fξ (d, y − h)
lim inf ≥ ξ(y − d) − 2k .
h→0 h
And on the other hand because fξ (d, d) < 0 there is ε > 0 such that for y ∈ [d, d + ε] it
holds fξ (d, y) < 0. For ξ ≥ 2k
ε , y ∈ [d + ε, ∞) we obtain fξ (d, y) ≤ x→∞
lim fξ (d, ∞) < 0.
 2k 4k
Letting ξ0 := max ε , M we obtain (xξ , yξ ) ∈ / ∂H for ξ ≥ ξ0 .

149
5 Optimal Control of Capital Injections by Reinsurance and Investments

Consider now the functions

ξ 2k
ψ(x) = vk (yξ ) + (x − yξ )2 + 2 +C
2 ξ (yξ − x) + ξ
ξ 2k
φ(y) = u(xξ ) − (xξ − y)2 − 2 −C ,
2 ξ (y − xξ ) + ξ

where
ξ 2k
C = u(xξ ) − vk (yξ ) − (xξ − yξ )2 − 2 .
2 ξ (yξ − xξ ) + ξ
The functions ψ(x), φ(y) are twice continuously differentiable. Furthermore, u(x)− ψ(x)
attains its maximum at xξ ; vk (y) − φ(y) attains its minimum at yξ . Also it holds

ψ ′ (xξ ) = φ′ (yξ ) and 0 ≤ ψ ′′ (xξ ) = −φ′′ (yξ ) . (5.13)

Therefore, it holds per Definition 5.2.3 of viscosity sub- and supersolutions

m2 ψ ′ (xξ )2 n Z ∞ o
− 2 ′′ + inf λ u(xξ − r(z, b)) dG(z) + c(b)ψ ′ (xξ )
2σ ψ (xξ ) b∈[0,b̃] 0
−(δ + λ)u(xξ ) ≥ 0 (5.14)
∞
m2 φ′ (y ξ )2 n Z o
− + inf λ vk (yξ − r(z, b)) dG(z) + c(b)φ′ (yξ )
2σ 2 φ′′ (yξ ) b∈[0,b̃] 0
−(δ + λ)vk (yξ ) ≤ 0 . (5.15)

Because the functions u, vk , r and c are continuous, there exist optimal b1 and b2 , at
which the minimum of (5.14) and (5.15) respectively is attained. Plugging in b1 into
Equation (5.15) and using (5.13) we obtain
Z ∞
m2 ψ ′ (xξ )2 λ+δ
− 2 ′′ + u(xξ − r(z, b1 )) − vk (yξ − r(z, b1 )) dG(z) ≥ (u(xξ ) − vk (yξ ))
λσ ψ (xξ ) 0 λ
(5.16)
It holds (xξ , xξ ), (yξ , yξ ) ∈ H and

fξ (xξ , xξ ) + fξ (yξ , yξ ) ≤ 2fξ (xξ , yξ ) .

Thus we obtain

ξ(xξ − yξ )2 ≤ u(xξ ) − u(yξ ) + vk (xξ ) − vk (yξ ) + 4k(yξ − xξ ) .

From above we conclude, that

ξ(xξ − yξ )2 ≤ 6k|xξ − yξ | . (5.17)

150
 5.2 The Classical Risk Model.

Choose a sequence ξn → ∞ as n → ∞ such that (xξn , yξn ) → (x̄, ȳ) ∈ H. From (5.17)
we obtain |xξn − yξn | ≤ 6k 2
ξn , from which it follows x̄ = ȳ; and ξn (xξn − yξn ) → 0. It holds
as ξn → ∞:

2 2
ψ ′ (xξn )2

ξn (xξn − yξn ) · ξn (yξn − xξn ) + 1 + 2k
= →0.
ψ ′′ (xξn ) ξn (ξn (yξn − xξn ) + 1)3 + 4kξn

Thus (5.16) becomes

Z ∞
λ+δ
M≥ u(x̄ − r(z, b1 )) − vk (ȳ − r(z, b1 )) dG(z) ≥ (u(x̄) − vk (ȳ)) .
0 λ
Together with lim inf Mξ ≥ M we obtain
ξ→∞

λ
M ≤ lim inf Mξ ≤ lim Mξn = u(x̄) − vk (ȳ) ≤ M,
ξ→∞ ξn →∞ λ+δ
which is an obvious contradiction if δ > 0.
Suppose now, that δ = 0. From above it follows, that lim inf Mξ = M and there is some
ξ→∞
sequence (ξn )n≥0 such that ξn → ∞ and (xξn , yξn ) → (x∗ , x∗ ) as n → ∞. Note, that
for every n ∈ N it holds fξn (x∗ , x∗ ) < 0, i.e. there is ε > 0 with fξn (x, y) ≤ 0 for all
(x, y) ∈ Uε (x∗ , x∗ ), where
n o
Uε (x∗ , x∗ ) = (x, y) ∈ H : ||(x − x∗ , y − x∗ )||2 < ε

denotes the open ball with center in (x∗ , x∗ ), radius ε and ||.|| denotes the Euclidean
norm. But on the other hand for every ε > 0 we can find n0 ∈ N such that

||(xξn − x∗ , yξn − x∗ )||2 < ε

for all n ≥ n0 . This contradicts the fact Mξ > 0 for all ξ > ξ0 . Thus we have shown the
uniqueness of the value function also in the case δ = 0. 

A direct consequence of the previous propositions is

Corollary 5.2.6
There is a unique viscosity solution to (5.8) with initial condition V (0) = v(0).

5.2.2 The special case δ = 0.

Consider the special case δ = 0. For simplicity we consider only the proportional
reinsurance, i.e. the self-insurance function is given by r(z, b) = zb for b ∈ [0, 1]. Like
in the case without reinsurance we will show the existence and uniqueness of the value

151
5 Optimal Control of Capital Injections by Reinsurance and Investments

function and of the optimal strategy. Finally the method will be illustrated by two
examples.
The Hamilton–Jacobi–Bellman equation is
∞
σ 2 a2 ′′
Z
inf V (x) + λ V (x − bz) dG(z) + (c(b) + am)V ′ (x) − λV (x) = 0 .
a∈R 2 0
b∈[0,1]
(5.18)

Like in the case without reinsurance it is clear, that we must have V ′′ (x) > 0. From
Lemma 5.2.1 it follows that the value function is convex in the case of proportional
reinsurance, i.e. we have V ′′ (x) ≥ 0. In the case V ′′ (x) = 0 and V ′ (x) < 0 on some
interval we would obtain for the strategies a = 1−c(0)
m , b = 0:

σ 2 a2 ′′
V (x) + (c(0) + am)V ′ (x) < 0 .
2
In the case V ′′ (x) = V ′ (x) = 0 we would inevitable obtain V (x) = 0. Thus, the HJB
equation would not have any solution in the case V ′′ (x) = 0.
We will see later, that in order to show the existence of the value function we will
need the optimal strategy (a∗ (0), b∗ (0)) at x = 0. The verification theorem below will
help to specify (a∗ (0), b∗ (0)).

Theorem 5.2.7 (Verification theorem)

Let f (x) be a decreasing, twice continuously differentiable solution to (5.18) with
lim f (∞) = 0. Then f (x) = V (x) and the optimal strategy is the strategy of the
x→∞
feedback form (A∗ (Xt ), B ∗ (Xt )).

Since the proof is perfectly similar to the proof of verification theorem in the case without
reinsurance, see Proposition 4.2.9, we will skip it.
′ (x)
It is clear, that for the minimising strategy A∗ it holds a∗ (x) = − σmV 2 V ′′ (x) . Together

with V ′′ (x) > 0 and V ′ (x) ≤ 0 we get a∗t ≥ 0 for the optimal investment strategy
A∗ = {a∗t }. The knowledge of the minimiser a∗ gives us the possibility to simplify the
HJB equation:
∞
m2 V ′ (x)2 n Z o
− 2 ′′ + inf λ V (x − bz) dG(z) + c(b)V ′ (x) − λV (x) = 0 . (5.19)
2σ V (x) b∈[0,1] 0

Using integration by parts (5.19) becomes

∞
m2 V ′ (x)2 n Z o
− 2 ′′ + inf − λb V ′ (x − bz)(1 − G(z)) dz + c(b)V ′ (x) = 0 .
2σ V (x) b∈[0,1] 0

152
 5.2 The Classical Risk Model.

For the initial capital x = 0 the Equation (5.19) becomes

m2 V ′ (0)2
+ λE[r(Z, b)] 1 + (1 + θ)V ′ (0) + λµ(η − θ)V ′ (0) = 0 .

inf − 2 ′′
b∈[0,1] 2σ V (0)
1
Thus, if it holds V ′ (0) < − 1+θ the optimal reinsurance strategy in 0 will be b∗ (0) = 1.
1
In the case V ′ (0) = − 1+θ we have
m2 V ′ (0)2
− + c(0)V ′ (0) = 0 ,
2σ 2 V ′′ (0)
from which it follows, that the optimal investment strategy in 0 should be given by
a∗ (0) = 2λµ(θ−η)
m and the reinsurance strategy b∗ (0) is arbitrary.
1
If we assume 0 > V ′ (0) > − 1+θ , then the optimal strategy in 0 would be b∗ (0) = 0,
from which we would again obtain a∗ (0) = 2λµ(θ−η)
m .
V ′ (0) = 0 would imply b = 0 and a = 0 as optimal strategies in 0, from which
it would follow because V is decreasing: V (x) ≡ 0. But for initial capital 0 and
(a∗ (0), b∗ (0)) = (0, 0) the ruin would occur immediately because of the premium pay-
ments and accordingly Yt > 0. I.e. the case (a∗ (0), b∗ (0)) = (0, 0) is impossible.
Thus, we have three candidates for the optimal strategy in 0:
 2λµ(θ−η)
′ 1
( m , 0) : V (0) ∈ (− 1+θ , 0) ,

∗ ∗ 1
(a , b ) = (a∗ , 1) : V ′ (0) ∈ (−1, − 1+θ ),
 2λµ(θ−η)

′ 1
( m , b) : V (0) = − 1+θ ,
where b ∈ [0, 1]. The following remark deduces the behavior of the optimal reinsurance
strategy B ∗ for small initial values x ≥ 0.
Remark 5.2.8
Consider the function
Z ∞
′


gx (b) := c(b)V (x) − λV (x) + λ V x − r(z, b) dG(z) .
0
Let further b1 , b2 ∈ [0, 1] and b1 > b2 . Then we obtain with the Lipschitz continuity of
V (x) for every x ∈ [0, ∞):
Z ∞



gx (b1 ) − gx (b2 ) = λ V x − r(z, b1 ) − V x − r(z, b2 ) dG(z)
0
+ (c(b1 ) − c(b2 ))V ′ (x)
Z ∞



= λ V x − r(z, b1 ) − V x − r(z, b2 )
0
n o
+ r(z, b1 ) − r(z, b2 ) (1 + θ)V ′ (x) dG(z)
Z ∞n o
≤ λ r(z, b1 ) − r(z, b2 ) · [1 + (1 + θ)V ′ (x)] dG(z) .
0

153
5 Optimal Control of Capital Injections by Reinsurance and Investments

1
V ′ (x) ≤ − 1+θ implies, that gx (b) is decreasing, so that the minimum is taken in b = 1.

Lemma 5.2.9
Assume the value function V (x) is the unique, twice continuously differentiable, vanish-
ing at infinity solution to the HJB equation (5.18); and the net profit condition c > λµ
is fulfilled, then the optimal investment strategy at x = 0 is a∗ = 0.
Proof: Arguing by contradiction we assume a∗ > 0 and denote by V (x) the value
function.
Denote the function, which results from choosing the start strategy (a(0), b(0)) = (0, 1)
and solving the HJB equation (5.18) by f (x). It is clear, that f ′′ (0) = ∞ and f ′ (0) =
− λµ ′′
c . With the same arguments as above we obtain f (x) > 0 for all x ∈ R+ . Let
further V 0,1 (x) define the return function of the constant strategy (A, B) ≡ (0, 1). From
Example 2.2.1 we know, that if the claim size distribution G(x) is absolutely continuous
and the density function is bounded, V 0,1 (x) is twice continuously differentiable with
(V 0,1 )′ (0) = − λµ ′
c = f (0) and solves the integro-differential equation
Z ∞
0,1 ′
c(V ) (x) + λ V 0,1 (x − z) dG(z) − λV 0,1 (x) = 0 .
0

Thus, we obtain for h(x) := f (x) − V 0,1 (x)

Z ∞
′
ch (x) ≥ −λ h(x − z) dG(z) + λh(x) .
0

Note, that it holds h′ (0) = 0 and h′′ (0) = ∞, from which it follows h′ (x) > 0 for all
x > 0. Thus, we obtain lim f (x) − V 0,1 (x) ≥ f (0) − V 0,1 (0). Because f (0) − V 0,1 (0) is
x→∞
finite and lim V 0,1 (x) = 0 we conclude, that lim f (x) = −∞ is impossible.
x→∞ x→∞
Define further
g(x) := V (x) − f (x) .
Because a∗ > 0 it must hold either
1 1 λµ
V ′ (0) ≥ − >− =− = f ′ (0)
1+θ 1+η c
or
λµ λµ
V ′ (0) = − >− = f ′ (0)
c + am/2 c
as well as 0 < V ′′ (0) < ∞. Thus g(x) is twice continuously differentiable, g(0) < 0,
g′ (0) > 0 and g′′ (0) = −∞ by definition. Let further (a(x), b(x)) denote the optimal
strategy for f (x), then we obtain from the HJB equation
Z x
a2 σ 2 ′′ ′
g (x) + (c(b) + ma)g (x) − λ g′ (x − bz)(1 − G(z)) dz ≥ 0 .
2 0

154
 5.2 The Classical Risk Model.

Note, that because b = 1 and

∞
ma ′
Z
− f (x) = cf ′ (x) − λ f ′ (x − z)(1 − G(z)) dz
2 0

close to zero, it follows from f ′′ (0) = ∞, that a(x) is strictly increasing near zero, i.e.
a(x) > 0 on (0, ε) for some ε > 0. Let x̌ = inf{x : g′′ (x) > 0} ∧ {x > ε : a(x) = 0}, then
Z x
a(x)2 σ 2 ′′
0> g (x) ≥ λ g′ (x − b(x)z)(1 − G(z)) dz − (c(b(x)) + ma(x))g′ (x)
2 0
≥ (λµ − c(b(x)) − ma(x))g′ (x) .
on (0, x̌). Because g′ (x) is continuous, it holds g′ (x) > 0 and (λµ − c(b(x)) − ma(x)) < 0
on [0, x̌). And we can conclude g′ (x) > 0 for x ≥ 0. Thus, the function g(x) is increasing,
f (x) is decreasing. In particular we have either lim f (x) = −∞ or lim f (x) = d, with
x→∞ x→∞
|d| < ∞. We have already shown above, that the first case is impossible. In the second
case consider the function f1 (x) = f (x) − d. f1 (x) solves the HJB equation for the
strategy (a(x), b(x)), satisfies the conditions from Lemma 5.2.1 and is convex, which
contradicts the result of Theorem 5.2.7. 

Thus the optimal strategy in 0 is given by (0, 1). Because we are looking for V (x), which
1
is twice continuously differentiable on (0, ∞), it holds V ′ (x) < − 1+θ for x small enough.
Due to Lemma 5.2.8 the optimal reinsurance strategy for those x is then also given by
1.
It remains to show, that the strong solution exists. Also in the case with reinsurance
the strong solution can exist only under some special constrains, which are specified in
the next proposition.
Proposition 5.2.10
There is a unique decreasing, twice continuously differentiable solution to (5.18), if the
c−c(b)
claims distribution function G(x) has a bounded density and lim > 0.
b→1 1−b

Proof: Note at first, that the optimal reinsurance strategy for x small enough is given
by b∗ (x) = 1, see Remark 5.2.8. Thus, we can conclude from Lemma 4.2.11, that there
is a twice continuously differentiable solution to 5.19 on some interval [0, ξ) with ξ > 0.
Next we show, that the solution can be extended to [0, ξ].
Let h(x) = V ′ (x) and
Z x Z ∞


U (x, b) = −λ h(z) 1 − G((x − z)/b) dz + c(b)h(x) + λ 1 − G(y/b) dy .
0 x

Then we obtain from HJB equation (5.18)

m2 h(x)2
= inf U (x, b) ,
2σ 2 h′ (x) b∈[0,1]

155
5 Optimal Control of Capital Injections by Reinsurance and Investments

which implies
d 1  m2  −1
= − 2 inf U (x, b) .
dx h(x) 2σ b∈[0,1]
Integrating the both sides of the above equation and taking the reciprocal yields
1
h(x) = − Rx −1 . (5.20)
c m2
λµ + 2σ2 0 inf U (y, b) dy
b∈[0,1]

It is clear, that h(ξ) is in either case well defined. Next we show, that the solution can
be extended to [0, ξ] and inf U (ξ, b) > 0. Suppose now inf U (ξ, b) = 0. Let xn be a
b∈[0,1] b∈[0,1]
sequence such that xn ↑ ξ as n → ∞ and denote by bn the optimal strategy at xn , i.e.
b∗ (xn ) = bn . We can find a converging subsequence bnk → b̂ as k → ∞. By continuity it
holds U (ξ, b̂) = inf U (ξ, b) = 0.
b∈[0,1]
Rx
Assume b̂ > 0. Then it must hold c(b̂) > 0 and h(ξ) < 0, because −λ 0 h(z) 1 −

R∞
G((x − z)/b) dz + λ x 1 − G(y/b) dy > 0. Since U (ξ, b̂) = 0 and (5.20) holds we obtain
h′ (ξ) = ∞. On the other hand it holds for all x ∈ [0, ξ):
U (ξ, b̂) − U (x, b̂)
0 >
ξ−x
Z ξ
G((ξ − z)/b̂) − G((x − z)/b̂) h(ξ) − h(x)
≥ λ h(z) dz + c(b̂)
0 ξ−x ξ−x
Z ξ
1 − G(z/b̂)
+λ .
x ξ−x
Because G(x) has a bounded density and h(x) is bounded, the above expression goes to
∞ as x → ξ. Thus, we obtain 0 > ∞, which is a contradiction.
Thus we can conclude lim b∗ (x) = 0 and h(ξ) = 0. Let further b0 = inf{b : c(b) ≥ 0}.
x↑ξ
Because lim b∗ (x) = 0 there is x̃ such that b∗ (x̃) < b0 /2 for all x > x̃. Then we obtain
x↑ξ
for x > x̃: c(b∗ ) < 0 and
Z x
1 c m2  −1
= − − inf U (x, b)
h(x) λµ 0 2σ 2 b∈[0,1]
Z x̃ x
c m2  m2 
Z
−1 −1
= − − inf U (y, b) dy − inf U (y, b) dy
λµ 0 2σ 2 b∈[0,1] x̃ 2σ 2 b∈[0,1]
Z x̃ x
c m2  m2
Z
−1
> − − inf U (y, b) dy − dy .
λµ 0 2σ 2 b∈[0,1] x̃ 2σ 2 c(b0 /2)h(y)
Let
m2
β=− ≥0
2σ 2 c(b0 /2)

156
 5.2 The Classical Risk Model.

and
x̃
c m2 
Z
−1
C= + inf U (y, b) dy
λµ 0 2σ 2 b∈[0,1]
Then we obtain from Gronwall’s inequality, see Proposition E.1.6 page 181:
1
− < Ceβ(x−x̃) .
h(x)

This contradicts h(ξ) = 0. The remaining part of the proof goes similar to the proof of
Theorem 4.2.12. 

5.2.3 Examples
−x
Example 5.2.11 (Proportional Reinsurance and G(x) = 1 − e µ .) x
Consider the proportional reinsurance, i.e. r(z, b) = zb and G = 1 − e− µ . Like in the
case without reinsurance, compare Example 4.2.17, the HJB equation differs from the
HJB equation
x
for the ruin probability as a function of initial capital only in the term
λµbe− bµ Letting V ′ (x) = h(x) we obtain
Z x
m2 h(x)2 n x
− bµ y
−x
o
− 2 ′ + inf − λe h(y)e bµ dy + c(b)h(x) + λµbe bµ = 0 ,
2σ h (x) b∈[0,1] 0

For the parameters σ 2 = 0.01, m = 0.03, δ = 0.04, µ = λ = 1, η = 0.3 and θ = 0.5

we can solve the problem numerically. The value function is plotted in Figure 5.3. The
optimal strategies A and B are shown in Figures 5.5 and 5.6 respectively. Also in this
case we can specify the value V (0) exactly: V (0) = 1.5718. 

µ2
Example 5.2.12 (Proportional reinsurance and G(x) = 1 − (µ+x)2
.)
Consider now the case of Pareto(µ, 2)-distributed claims. For the parameters from the
Example with exponentially distributed claim sizes the value function is plotted in Figure
5.4. The optimal strategies A and B are shown in Figures 5.7 and 5.8 respectively.
We see, that the optimal investment strategy converges approximately to 11.2 and the
optimal reinsurance strategy converges to zero. The initial value is given by V (0) = 2.5.


157
5 Optimal Control of Capital Injections by Reinsurance and Investments

x
Figure 5.3: Value function V (x) for G(x) = 1 − e− µ .

0 5 10 15 20
x
µ2
Figure 5.4: Value function V (x) for G(x) = 1 − (µ+x)2
.

158
 5.2 The Classical Risk Model.

x
−µ
Figure 5.5: Optimal investment strategy A for G(x) = 1 − e .

x
−µ
Figure 5.6: Optimal reinsurance strategy B for G(x) = 1 − e .

159
5 Optimal Control of Capital Injections by Reinsurance and Investments

12

10

0
0 2 4 6 8 10 12 14
µ2
Figure 5.7: Optimal investment strategy A for G(x) = 1 − (µ+x)2 .

1,0

0,8

0,6

0,4

0,2

0 5 10 15 20 25

µ2
Figure 5.8: Optimal reinsurance strategy B for G(x) = 1 − (µ+x)2
.

160
Appendix

161
A Stochastic Processes and Martingales
Martingales have been an important tool in the proofs of existence of the value func-
tion. In this appendix we consider some well known results, which state the processes to
be martingales under some special constrains. For detailed introduction to martingales
we refer for example to Protter [60], to Bhattacharya and Waymire [9] or to Rogers and
Williams [63, 64].

A.1 Stochastic processes

We assume as given a complete probability space (Ω, F, P). Let I = N or I = R, then
the stochastic process on I with state space R, is a family of R-valued random variables
X = {Xt : t ∈ I}, or in the short notation X = {Xt }. X is said to be cadlag, if its
sample paths are right continuous with left limits. We define the variation of X over
(0, t] by
nX n o
FX (t) = sup |Xti − Xti−1 | ,
i=1

where the supremum is taken over all n ≥ 1 and all partitions 0 = t0 < t1 < · · · < tn = t.
The process X is said to be of bounded variation if FX (t) < ∞ for all t > 0.

Definition A.1.1
A filtration F = {Ft } is a non-decreasing family of sub-σ-algebras. For any process
Y = {Yt }, the natural filtration FY is given by FtY = σ{Ys ; s ≤ t}. Thus FtY is the
σ-algebra generated by Y up t and represents the history of Y up to time t.
We say Y is adapted to F if Yt is Ft measurable for all t ≥ 0.

In theTfollowing we always assume, that a filtration F = {Ft } is right continuous, i.e.

Ft = Fs for all 0 ≤ t < ∞.
s>t
A σ-algebra P on Ω × [0, ∞) is called predictable if it is generated by all processes
{Xt− }, where {Xt } is {Ft } is adapted. A process X is called predictable if it is P
measurable.
Definition A.1.2
A random variable T : Ω → [0, ∞] is a stopping time, if the event {T ≤ t} ∈ Ft for
every 0 ≤ t ≤ ∞.

163
A Stochastic Processes and Martingales

In the classical risk model and its diffusion approximation the concepts of Poisson process
and Brownian motion are of central importance.

Definition A.1.3 (Poisson Process)

An {Ft }- adapted process N = {Nt } defined by
X
Nt = 1I[t≥Tn ] ,
n≥1

where Tn are stopping times, is a Poisson process if

1. for any s, t with 0 ≤ s < t < ∞, Nt − Ns is independent of Fs ;

2. for any s, t, u, v with 0 ≤ s < t < ∞, 0 ≤ u < v < ∞ and t − s = v − u the

distribution of Nt − Ns is the same as that of Nv − Nu .

3. It holds
e−λt (λt)n
P[Nt = n] = .
n!

{Tn } are called the arrival times.

Definition A.1.4 (Brownian Motion)

An {Ft }-adapted continuous process W = {Wt } taking values in R is called standard
Brownian motion if

1. for 0 ≤ s < t < ∞, Wt − Ws is independent of Fs ;

2. for 0 < s < t, Wt − Ws is a Gaussian random variable with mean zero and variance
(t − s).

A standard Brownian motion starts at 0, P[W0 = 0] = 1.

A process {mt + σWt } is called an (m, σ 2 )-Brownian motion.

Proposition A.1.5 (Strong law of large numbers)

Let W = {Wt } be a standard Brownian motion. Then it holds

Wt
lim =0.
t→∞ t
Proof: See Rolski et al. [61, p. 431]. 

Among other processes the running maximum of the standard Brownian motion played
an important role in our considerations.

164
 A.2 Martingales

Remark A.1.6 (Brownian Motion and Its Running Maximum)

Let W be a standard Brownian motion and define the running maximum

Mt = max Wt . (A.1)
0≤s≤t

The density function of the pair (Wt , Mt ) is given by

2(2b − a) n (2b − a)2 o

f (a, b) = √ exp − 1I[a≤b] 1I[b≥0] .
2πt3 2t

Let further X be an (m, σ 2 )-Brownian motion with m > 0. Define τ x = inf{t ≥ 0 : Xt =

−x}, x > 0. Then τ x is a stopping time and
 mt + x  2
 mt − x 
P[τ x ≤ t] = Φ − √ + e−2mx/σ Φ √ ,
σ t σ t
where Φ denotes the distribution function of the standard normal distribution.
For proofs see Karatzas and Shreve [49, p. 95].

A.2 Martingales
In this section we give the concepts of martingales and local martingales and the
propositions, we have used in the work. In some of the propositions below stochastic
integrals play a central role. The concept of stochastic integral goes beyond the scope of
this appendix. We abandon any definitions and explanations and just refer for example
to Roger and Williams [64].

Definition A.2.1
A real-valued, adapted process X = {Xt } is called a martingale with respect to the
filtration F = {Ft } if

1. E[|Xt |] < ∞ for all t ≥ 0;

2. E[Xt |Fs ] = Xs for all s ≤ t.

The next two propositions state very significant for this work properties of the Brownian
motion and Poisson process.

Proposition A.2.2
Let N = {Nt } be a Poisson process with intensity λ and arrival times {Tn }. Then
Mt = Nt −λt, the compensated
R t Poisson process, is a martingale. For
R t an {Ft } predictable
process H = {Ht } with E[ 0 |Hs | ds] < ∞, the process It = 0 Hs dMs is an {Ft }
martingale.

165
A Stochastic Processes and Martingales

Proposition A.2.3
Let W = {Wt } be a standard Brownian motion and l > 0. Then the following processes
are martingales

{Wt } ,
{Wt2 − t} ,
2 t/2
{elWt −l }.

Definition A.2.4
An adapted process X is a local martingale, if there exists a sequence of increasing
stopping times, Tn with lim Tn = ∞ a.s., such that Xt∧Tn 1I[Tn >0] is a martingale for
n→∞
each n.
Proposition A.2.5
A continuous local martingale of bounded variation is constant.
A predictable local martingale of bounded variation is constant.
For proof see Rolski et al. [61, p. 566].

Proposition A.2.6
Let X = {Xt } be an arbitrary stochastic process and let W = {Wt } be a standard
Brownian motion.
Rt Rt
1. If 0 Xs2 ds < ∞, then the stochastic integral 0 Xs dWs is a local martingale.
Rt Rt
2. If 0 E[Xs2 ] ds < ∞, then the stochastic integral 0 Xs dWs is a true martingale
with zero expectation.

In particular, if X is a bounded process, then the stochastic integral is a martingale.

Definition A.2.7
A semimartingale is an adapted real-valued process X = {Xt } of the form

Xt = Mt + At ,

where M = {Mt } is a local martingale and A = {At } is a process of bounded variation.

166
B Markov Processes and Infinitesimal
Generators
In this work we have often dealt with Markov processes, and used their properties in
several proofs. Thus, we start with the definition of a Markov process.

B.1 Markov processes

Definition B.1.1
Let ν be a probability measure on the space (R, B(R)), where B(R) is the Borel σ-
Algebra on R. An F adapted real-valued one-dimensional process X = (Xt )t≥0 on
(Ω, F, P) is said to be a Markov process with initial distribution ν if
1. P[X0 ∈ B] = ν[B] for all B ∈ B(R),

2. for s, t ≥ 0 and B ∈ B(R)

P[Xt+s ∈ B|Fs ] = P[Xt+s ∈ B|Xs ] , P-a.s.

We define the transition function of the process X by

Pt (s, x, B) = P[Xt+s ∈ B|Xt = x] .

If the transition function does not depend on t, we call the Markov process homogeneous,
and we will omit the index t.

In the following we let Cb (E), E ⊂ R denote the set of all measurable bounded real
functions on E. We endow Cb (E) with the supremum norm ||f ||∞ = sup |f (x)|.
x∈E

Definition B.1.2
Let {T (h) : h ≥ 0} be a family of bounded linear operators from Cb (E) to Cb (E). Then
{T (h)} is called a contraction semigroup on Cb (E) if
1. T (0) = I, where I is the identity function I : Cb (E) → Cb (E), f 7→ f ;

2. T (s + t) = T (s)T (t) for all s, t ≥ 0;

3. ||T (h)f ||∞ ≤ ||f ||∞ .

167
B Markov Processes and Infinitesimal Generators

Proposition B.1.3
Let {Xt } be a Markov process with transition functions P (t, x, B). The mapping family
Z
T (t)f (x) = f (y)P (t, x, dy) = Ex [f (Xt )] , (B.1)

where f ∈ Cb (E), is a contraction semigroup.

For proof see Rolski et al. [61, p. 440].

B.2 Generators
Definition B.2.1
The infinitesimal generator of a Markov process {Xt } or of its corresponding semigroup
{T (t)}, defined in (B.1), is the linear operator D defined by
T (t)f (x) − f (x)
Df (x) = lim ,
t↓0 t
provided the limit exists. We call the class of functions, where the right side converges
to some function uniformly in x, the domain D(D) of generator D.

Theorem B.2.2
Assume that X = {Xt } is an E-valued Markov process, E ⊂ R with transition functions
P (h, x, B). Let further {T (h)} denote the corresponding semigroup, defined by (B.1),
and let D be its generator. Then for each f ∈ D(D) the stochastic process {Mt } is an
{FtX }-martingale, where
Z t
Mt = f (Xt ) − f (X0 ) − Df (Xs ) ds .
0

For proof see Rolski et al. [61, p. 442].

B.2.1 The classical risk model

Lemma B.2.3
Every stochastic process X with stationary and independent increments is a Markov
process.
The process defined in (1.1), i.e. the classical risk model, is a Markov process since it
has stationary and independent increments.
Proposition B.2.4
The process X defined in (1.1) has the infinitesimal generator
Z ∞
′
Df (x) = cf (x) + λ f (x − z) dG(z) − λf (z) ,
0

168
 B.2 Generators

where f is a bounded real-valued, differentiable function and G is the distribution func-

tion of the claim sizes Zi .
Proof: Consider the semigroup T (h) associated to Xt = x + ct − St . It is given by
∞
X (λh)k
T (h)f (x) = e−λh E[f (x + ch − Sh )|Nh = k]
k!
k=0

for x ∈ R and f ∈ Cb (R). Thus, we obtain

T (h)f (x) = e−λh f (x + ch) + λhe−λh E[f (x + ch − Z1 )|Nh = 1]
∞
2
X λk+2 hk −λh
+h e E[f (x + ch − Sh )|Nh = k + 2]; .
(k + 2)!
k=0

Because f is bounded and differentiable we have

T (h)f (x) − f (x) n f (x + ch) − f (x)
lim = lim e−λh
h→0 h h→0 h
e−λh − 1
+ f (x)
h
+ λe−λh E[f (x + ch − Z1 )|Nh = 1]
∞
X λk+2 hk −λh o
+h e E[f (x + ch − Sh )|Nh = k + 2]
(k + 2)!
k=0
Z ∞
′
= cf (x) − λf (x) + λ f (x − z) dG(z) .
0


B.2.2 Diffusion approximation

Next we will give the infinitesimal generator of a diffusion process. The concept of
a diffusion we have already explained in Definition 1.2.1 p. 17. In Bhattacharya and
Waymire [9, p. 367] one finds that a diffusion X, solving the stochastic differential
equation
dXt = m(x) dt + σ(x) dWt ,
fulfils
E[(Xs+t − Xs )1I[|Xt −x|≤ε] |Xs = x] = tm(x) + o(t) ,
E[(Xs+t − Xs )2 1I[|Xt −x|≤ε]|Xs = x] = tσ 2 (x) + o(t) , (B.2)
E[|Xs+t − Xs | > ε|Xs = x] = o(t)
o(t)
as t ↓ 0 for every x ∈ R and every ε > 0, where lim t = 0.
t→0

169
B Markov Processes and Infinitesimal Generators

Proposition B.2.5
Let {Xt } be a diffusion defined above. Then for all bounded twice continuously diffe-
rentiable real valued f the infinitesimal generator of {Xt } is given by

σ 2 (x) ′′
Df (x) = m(x)f ′ (x) + f (x) .
2
Proof: Let x ∈ R be fixed and δ > 0. Because f is twice continuously differentiable,
there is ε > 0 such that |f ′′ (x) − f ′′ (y)| < δ for all |x − y| ≤ ε. Further it holds

T (t)f (x) = Ex [f (Xt )1I[|Xt −x|≤ε] ] + Ex [f (Xt )1I[|Xt −x|>ε] ] .

Taking a Taylor expansion on f (Xt ) in a neighborhood of x

hn (Xt − x)2 ′′
Ex [f (Xt )1I[|Xt −x|≤ε]] = E f (x) + (Xt − x)f ′ (x) + f (x)
2
(Xt − x)2 ′′ o i
+ (f (ξt ) − f ′′ (x)) 1I[|Xt −x|≤ε] .
2
where ξt is a number between Xt and x. Using (B.2) the expectations of the first three
terms add up to

tσ 2 (x) ′′
f (x) + tm(x)f ′ (x) + f (x) + o(t) ,
2
t ↓ 0. Thus, we obtain
h (X − x)2 i δσ 2 (x) o(t)
t
Ex (f ′′ (ξt ) − f ′′ (x))1I[|Xt −x|≤ε] ≤ + lim sup
2 2 t
for t ↓ 0. Because f is bounded and using (B.2) we obtain

Ex [|f (Xt )|1I[|Xt −x|>ε]] ≤ ||f ||E[1I[|Xt −x|>ε] ] = o(t)

for t ↓ 0. It means lim sup Ex [f (Xt )1I[|Xt −x|>ε] ] = 0. Thus we have

t↓0

T (t)f (x) − f (x) σ 2 (x) ′′ δσ 2 (x)

lim sup − m(x)f ′ (x) − f (x) ≤ .
t↓0 t 2 2

Because δ > 0 was arbitrary it follows

T (t)f (x) − f (x) σ 2 (x) ′′

lim = m(x)f ′ (x) + f (x) .
t↓0 t 2


170
C Change of Measure and Change of
Variables Formulae
In this appendix we give the important change of measure and change of variables
formulae.

C.1 Ito’s Lemma

We start with the change of variables formula, also called Ito’s lemma. Here we give
just a one-dimensional version for the processes of the form
Z t Z t Z t
(3)
Xt = x + Ks(1) ds + Ks(2) dWs + Ks− dYs , (C.1)
0 0 0

where W is the standard Brownian motion and Y is a process of bounded variation.

Such process, if it is well defined, is called Ito process. A general form of Ito’s result can
be found for example in Protter [60].

Proposition C.1.1 (Ito’s Lemma)

Let X be an Ito process defined in (C.1) and f (t, x) be a function which is continuously
differentiable with respect to t and twice continuously differentiable with respect to x.
Then X̃t = f (t, Xt ) fulfils
Z tn (2)
∂f ∂f (Ks )2 ∂ 2 f o
X̃t = (s, Xs ) + Ks(1) (s, Xs ) + (s, Xs ) ds
0 ∂s ∂x 2 ∂x2
Z t Z t
(2) ∂f ∂f (3)
+ Ks (s, Xs ) dWs + (s, Xs )Ks− dYs .
0 ∂x 0 ∂x

Proof: For proof see Rolski et al. [61, p. 559]. 

C.2 Change of Measure

In order to introduce the change of measure formula we need some preliminary re-
marks. Let W = {Wt } be an F = {Ft } adapted standard Brownian motion living

171
C Change of Measure and Change of Variables Formulae

Ront (Ω, F, P). Let further {ls } be an adapted process such that the stochastic integral
0 ls dWs is well defined. Define further
nZ t 1 t 2 o
Z
Kt = exp ls dWs − l ds .
0 2 0 s

Theorem C.2.1 (Novikov)

If for each t ≥ 0
h 1 Z t i
E exp ls2 ds <∞,
2 0
then it holds E[Kt ] = 1 for all t ≥ 0. In this case the process {Kt } is a positive
martingale.
For the change of measure formula we do not really need Novikov’s theorem. But it
gives us a useful tool for constructing a strictly positive martingale with expected value
1.
Definition C.2.2
Let ν and κ be two measures ν and κ on the same measurable space, κ is said to be
absolutely continuous with respect to ν, or dominated by ν if κ(A) = 0 for every set
A for which ν(A) = 0. This is written as κ ≪ ν .

Remark C.2.3
Let M be a strictly positive martingale with respect to the filtration {Ft } and expected
value 1. Then Z
P̃t [A] = Mt dP = EP [1IA Mt ]
A
is a probability measure on Ft . If T is a given time, then for any A ∈ Ft ⊂ FT it holds

P̃ [A] = EP [1IA MT ] = P̃t [A] ,

and P̃ is a probability measure on (Ω, FT ). Moreover P̃ is absolutely continuous with

respect to P and Mt is a version of its Radon-Nikodym derivative when the measures
are restricted to Ft , 0 ≤ t ≤ T

dP̃
= Mt P − a.s.
dP Ft

Proposition C.2.4 (Girsanov’s theorem) R

t
Let J = {Jt } be an adapted process satisfying 0 Js2 ds < ∞ a.s. and such that the
process M = {Mt } defined by
 Z t 1 t 2 
Z
Mt = exp − Js dWs − J ds
0 2 0 s

172
 C.2 Change of Measure

is an {Ft } martingale.
Rt Then, under the measure P̃ , defined as in Remark C.2.3, the
process {Wt + 0 Js2 ds : 0 ≤ t ≤ T } is an {Ft }-standard Brownian motion.
Proof: For the proof see Rogers and Williams [63, p. 82], where uncompleted filtrations
are used and the result is formulated in R. 

Remark C.2.5
Consider an (m, σ 2 )-Brownian motion Xt = x+mt+σWt on (Ω, F, P), where W = {Wt }
is again the standard Brownian motion.
(i) Let J be a constant process J = {Jt } = { σm2 }. Then, due to Proposition A.2.3
 m m2 
Mt = exp − 2 Wt − 4 t
σ 2σ

is a martingale with expectation 1. Under the measure P̃ , defined in Remark C.2.3, the
process Xt is a (0, σ 2 )-Brownian motion with initial value x. Note, that because the
(0, σ 2 )-Brownian motion exists, the measure P̃ can be extended to F.
(ii) From change of measure formula we also obtain for every positive β and stopping
time τ = inf{t ≥ 0 : Xt < 0}:
√ 2
√ 2
2 2 2 2
EP [e−βτ 1I[τ <∞]] = EP̃ [eXτ (m+ m +2βσ )/σ ] · e−x(m+ m +2βσ )/σ
√ 2
2 2
= e−x(m+ m +2βσ )/σ .

173
C Change of Measure and Change of Variables Formulae

174
D Stochastic Differential Equations and
Local Times
In this appendix we make a short digression into the theory of diffusions. Our con-
siderations are restricted to the one-dimensional case. For a detailed introduction to
stochastic differential equations we refer to Karatzas and Shreve [49], to Protter [60] or
to Revuz and Yor [62].

D.1 Stochastic Differential Equations

Let (Ω, F, P) be a probability space on which a one-dimensional standard Brownian
motion W = {Wt } is given. Let further {Ft } be the natural filtration of the Brownian
motion W . In Preliminaries we have already given a definition of a diffusion and in Sec-
tion 2.1 we consider a diffusion, which is a strong solution to some stochastic differential
equation (SDE). Next we give a definition, what a strong solution is.

Definition D.1.1
Let m : R → R and σ : R → R be Borel-measurable functions. A process X, X0 = x
taking values in R is said to be a strong solution to the stochastic differential equation

dXt = m(Xt ) dt + σ(Xt ) dWt (D.1)

if

1. X is {Ft }-adapted;

2. P[X0 = x] = 1;
Rt
3. 0 |σ(Xt )|2 + |m(Xt )| ds < ∞ for all t ≥ 0 a.s.;

Rt Rt
4. Xt = x + 0 m(Xs ) ds + 0 σ(Xs ) dWs for all t ≥ 0 a.s.

If X and X̃ are two strong solutions to (D.1) with P[Xt = X̃t , 0 ≤ t < ∞] = 1, then we
say that Equation (D.1) has a unique strong solution.

Note that due to 3. the integrals in 4. are well defined. It is also clear from 4. that X is
a continuous semimartingale.

175
D Stochastic Differential Equations and Local Times

D.2 Local Times

We start with the most simple diffusion, namely with the standard Brownian motion.
Let W = {Wt } be the standard Brownian motion.

Definition D.2.1
By the local time of W we denote a family of non-negative random variables Lx =
{Lxt , x ∈ R} such that with probability one, the following holds
1. (t, x) 7→ Lxt is continuous;
2. for every Borel subset A ⊂ R and t ∈ [0, ∞)
Z t Z
1IA (Ws ) ds = Lxt dx .
0 A

If such a family exists, then it is unique and is given by

Z t
x 1
Lt = lim 1I[(x−ε,x+ε)](Ws ) ds .
ε↓0 2ε 0

Theorem D.2.2
The local time {Lxt } of W exists.
Proof: Confer Ikeda and Watanabe [44, p. 113] 

The local time at zero {L0t } has been of particular interest for our considerations. From
now on, speaking of local time, we will mean the local time at zero. Itô’s formula for
twice continuously differentiable functions f can be extended to functions, satisfying less
restrictive assumptions. The first result in this direction is known as Tanaka’s formula.
Let 
1 , : x≥0,
sgn(x) =
−1 , : x < 0 .
Then it holds Z t
|Wt | = sgn(Ws ) dWs + L0t .
0
The process |W | = {|Wt |} we call the reflected standard Brownian motion.
Proof: See for example Protter [60, p. 217]. 

Next theorem states that the reflected Brownian motion can be characterised by the
Skorokhod equation.

Theorem D.2.3 (Skorokhod)

Let x : R+ → R be a continuous function with x(0) ≥ 0. Then there exists a unique pair
(a, y) of continuous functions such that

176
 D.2 Local Times

1. a(t) = x(t) + y(t),

2. a(t) ≥ 0, y is increasing, y(0) = 0,

Rt
3. 0 1I[a(s)6=0] dy(s) = 0, ∀t. Moreover, the function y(t) is explicitly given by

y(t) = − min{ inf x(s), 0} .

0≤s≤t

Proof: See Rogers and Williams [64, p. 117]. 

In particular, we obtain that the Brownian local time at zero L0 starts at zero, is
continuous, non-decreasing and dL is almost surely supported by {t ≥ 0 : |Wt | = 0}.
Moreover, the local time L0 is given by L0t = − min{ inf Ws , 0}. Also it holds
0≤s≤t

|Wt | = Wt + L0t .

Note that |W | is a diffusion process solving the stochastic differential equation dWt +
dL0t . With change of measure and time-change formulae the results for the standard
Brownian motion can be transfered, under some assumptions, to strong solutions to
SDEs (see Rogers and Williams [64, p. 289]).

177
D Stochastic Differential Equations and Local Times

178
E Differential equations
In this appendix we review some results from the theory of differential equations,
which we have used in the work. We start with an equation, which we have had to solve
in every diffusion approximation model with the exception of the model with constant
interest rate, Section 3.1. We find in Kamke [48] for linear homogeneous second-order
differential equations with constant coefficients:
Proposition E.1.4
Every solution to the differential equation of the form
y ′′ + by ′ + cy = 0
is given by
 −b + √b2 − 4c  b+
√
b2 − 4c
y = C1 · exp · x + C2 · exp(− · x) ,
2 2
provided b2 − 4c ≥ 0.

In Section 3.1 we have had to solve the differential equations of the form
y ′′ (x) + p(x)y ′ (x) + qy(x) = 0 ,
where p(x) is a linear function and q is a constant. This is a linear second-order homo-
geneous differential equation. In remark below we explain, how to solve this differential
equation with the so-called power series method.
Remark E.1.5
Consider then
y ′′ (x) + p(x)y ′ (x) + qy(x) = 0 ,
where p(x) = J1 x + J2 and q is a constant. Let φ(p(x)) = y(x), then it holds y ′ (x) =
J1 φ′ (p(x)) and y ′′ (x) = J12 φ′′ (p(x)). Thus, the above differential equation can be trans-
formed to
J12 φ′′ (p) + pJ1 φ′ (p) + qφ(p) = 0 .
The power series method calls for the construction of a power series solution
∞
X
φ(p) = an pn
n=0

179
E Differential equations

∞ ∞
with φ′ (p) = nan pn−1 and φ′′ (p) = n(n − 1)an pn−2 . Subsetting these into the
P P
n=1 n=2
differential equation yields

∞
X ∞
X ∞
X
0 = J12 n(n − 1)an pn−2 + pJ1 nan pn−1 + q an pn
n=2 n=1 n=0
∞
X n o
= pn J12 (n + 2)(n + 1)an+2 + (J1 n + q)an .
n=0

Now, if this series is a solution, all the coefficients must be zero, so:

an (−J1 n − q)
an+2 = 2 .
J1 (n + 1)(n + 2)

Solving this relation recursively, we obtain an expression for every an , n ≥ 2 in terms of

a0 and a1 :

−a0 q
a2 = ,
J12 2
−J1 − q
a3 = a1 ,
J12 2 · 3
..
.
n−1
Q
(−2k − q/J1 )
k=0
a2n = a0 ,
J1n (2n)!
n
Q
(−(2k − 1) − q/J1 )
k=1
a2n+1 = a1 .
J1n (2n + 1)!

Thus, it holds

n−1
Q
∞  ∞ (−2k − q/J1 ) 
k=0
X X
n
φ(p) = an p = a0 1 + p2n
J1n (2n)!
n=0 n=1
n
Q
 ∞ (−(2k − 1) − q/J1 ) 
k=1
X
+ a1 p + p2n+1
n=1
J1n (2n + 1)!

180
Plugging in p(x) = J1 x + J2 yields
n−1
Q
 ∞ (−2k − q/J1 ) 
k=0
X
y(x) = a0 1 + J1n (x + J2 /J1 )2n
(2n)!
n=1
n
Q
 ∞ (−(2k − 1) − q/J1 ) 
k=1
X
+ a1 J1 x + J2 /J1 + J1n (x + J2 /J1 )2n+1 .
n=1
(2n + 1)!

2m 2λµη 2δ
For example for J1 = λµ2 , J2 = λµ2 and q = − λµ 2
we obtain
∞ δ δ
 X
m · · · (m − 2n + 2)  2m n λµη 2n 
y(x) = a0 1 + (x + )
(2n)! λµ2 m
n=1
δ ∞ δ
2m  λµη X ( m − 1) · · · ( m − 2n + 1)  2m n λµη 2n+1 
+ a1 x+ + (x + ) .
λµ2 m n=1
(2n + 1)! λµ2 m

Gronwall’s lemma, which we will consider below, has been an important tool to obtain
the existence of a solution in Section 5.2. Gronwall’s lemma allows one to bound a
function, that is known to satisfy a certain differential or integral inequality, by the
solution of the corresponding differential or integral equation. There are two forms of
the lemma, a differential form and an integral form. We will consider here the integral
form.
Proposition E.1.6
Let I = [a, b] ⊂ R and u, C : I → R and β : I → [0, ∞) be continuous functions.
Suppose
Z b
u(x) ≤ C(x) + β(y)u(y) dy
a
for all x ∈ I. Then it holds
Z b Rx
β(s) ds
u(x) ≤ C(x) + C(y)β(y)e y dy .
a

for all x ∈ I.
If C and β are constants with β ≥ 0, then we have

u(x) ≤ Ceβ(x−a) .

Proof: For proof see Amann and Escher [2, p. 131]. 

181
E Differential equations

182
F The Black-Scholes Model
The Black Scholes model is the most widely used model for pricing options. The
model and associated call and put option formulas remained the industry standard in
equity and currency markets for the last thirty years. That is why, giving the insurer
the possibility to invest his money, we have used the Black-Scholes setup for modeling
of a risky asset.
In this section we give a short introduction of the Black-Scholes model. For detailed
introduction we refer to Müller-Möhl [56], Lamberton and Lapeyre [51] and Rolski et al.
[61].

We have a probability space (Ω, F, P), which is large enough to carry all the stochastic
quantities defined below. With {Ft } we denote the smallest right-continuous filtration
such that the standard Brownian motion W = {Wt } is adapted with respect to {Ft }.
The Black-Scholes model is a continuous-time model, which describes the behavior of
the price of a risky asset with price Qt at time t and a riskless asset with price Q0t at
time t. We suppose that the behavior of Q0 follows the differential equation

dQ0t = rQ0t dt ,

where r is a non-negative constant, which is called risk-free rate. Further we assume

that the behavior of the stock price is determined by the stochastic differential equation

dQt = Qt (m dt + σ dWt ) Q0 > 0 ,

where m and σ are constants.

We see, that the process Q = {Qt } is a solution to the above differential equation if and
only if the process {log(Qt )} is an (m, σ)- Brownian motion; and it holds
 
Qt = Q0 exp (m − σ 2 /2)t + σWt .

The return of the process Q is given by

dQt
Q′t = = m dt + σ dWt .
Qt

m is called the expected rate of return and σ > 0 the volatility.

183
F The Black-Scholes Model

An European Option assumes that it can be exercised only at expiration. The Black-
Scholes model requires that the risk-free rate r, the expected rate of return m and the
volatility of the underlying stock price σ 2 remain constant over the period of analysis.
The model also assumes that the underlying stock does not pay dividends. The Black-
Scholes formula calculates the price of an European call option to be:

C = SΦ(d1 ) − Xe−rT Φ(d2 ) ,

where

C = price of the call option

X = option exercise price
T = current time until expiration
Φ = standard Normal distribution
ln(Q/X)+(r+σ2 /2)T
d1 = σT 1/2
d2 = d1 − σT 1/2

Put-call parity requires that:

P = C − Q + Xe−rT

Then the price of an European put option is:

P = Xe−rT Φ(−d2 ) − QΦ(−d1 ) .

184
Bibliography

[1] Albrecher, H. and Thonhauser, S. (2007). Dividend maximization under considera-

tion of the time value of ruin. Insurance: Mathematics and Economics 41, 163–184.

[2] Amann, H. and Escher, J. (1999). Analysis II. Birkhäuser Verlag, Basel.

[3] Asmussen, S. and M. Taksar (1997). Controlled diffusion models for dividend pay-
out. Insurance: Mathematics and Economics 20, 1–15.

[4] Asmussen, S., Højgaard B. and Taksar M. (2000). Optimal risk control and div-
idend distribution policies. Example of excess-of-loss reinsurance for an insurance
corporation. Finance Stoch. 4, 299–324.

[5] Azcue, P. and Muler, N. (2005). Optimal reinsurance and dividend distribution
policies in the Cramér–Lundberg model. Math. Finance 15, 261–308.

[6] Bardi, M. and Capuzzo-Dolcetta, I. (1997). Optimal Control and Viscosity Solutions
of Hamilton–Jacobi–Bellman Equations. Birkhäuser, Boston.

[7] Beard, R.E., Pentikäinen, T. and Pesonen, E. (1984). Risk Theory. Chapman and
Hall, London.

[8] Benth, F.E., Karlsen, K.H. and Reikvam, K. (2001). Optimal portfolio selection with
consumption and nonlinear integro-differential equations with gradient constraint:
A viscosity solution approach. Finance Stoch. 5(3), 275–303.

[9] Bhattacharya, R. N. and Waymire, E. C. (1990). Stochastic Processes with Appli-

cations. Wiley, New York.

[10] Billingsley, P. (1999) Convergence of Probability Measures. Wiley, New York.

[11] Bremaud, P. (1981). Point Processes and Queues, Martingale Dynamics. Springer-
Verlag, New York.

[12] Borch, K.H. (1974). The Mathematical Theory of Insurance. Cambridge, MA: Lex-
ington Books, D.C. Heath.

185
Bibliography

[13] Borodin, A.N. and Salminen P. (2002). Handbook of Brownian Motion – Facts and
Formulae. Birkhäuser Verlag, Basel.

[14] Bühlmann, H. (1970). Mathematical methods in risk theory. Springer-Verlag, Berlin.

[15] Cramér, H. (1930) On the Mathematical Theory of Risk. Scandia Jubilee Volume,
Stockholm.

[16] Cramér, H. (1955) Collective risk theory. Scandia Jubilee Volume, Stockholm.

[17] Crandall, M.G., Ishii, H. and Lions, P.L. (1992). User’s guide to viscosity solutions
of second order partial differential equations. Bull. Amer. Math. Soc 27, 1–67.

[18] Dickson, D.C.M and Waters, H.R. (2004). Some optimal dividend problems. ASTIN
Bulletin 34, 49–74.

[19] Dickson, D.C.M. and Drekic, S. (2006). Optimal dividends under a ruin probability
constraint. Annals of Actuarial Science 1 (2), 291–306.

[20] Durrett, R. (1996). Stochastic Calculus: A Practical Introduction. CRC Press, Boca
Raton.

[21] Eisenberg, J. and Schmidli, H. (2008). Optimal control of capital injections by

reinsurance in a diffusion approximation. Blätter der DGVFM 30, 1–13

[22] Eisenberg, J. and Schmidli, H. (2009). Minimising Expected Discounted Capital

Injections by Reinsurance in a Classical Risk Model. To appear in Scandinavian
Actuarial Journal.

[23] Ethier, S.N. and Kurtz, T.G. (1986). Markov Processes. Characterization and Con-
vergence. Wiley, New York.

[24] de Finetti, B. (1957). Su un’impostazione alternativa della teoria collettiva del ris-
chio. Transactions of the XVth International Congress of Actuaries 2, 433–443.

[25] Fleming, W.H. and Soner, H.M. (1993). Controlled Markov Processes and Viscosity
Solutions. Springer-Verlag, New York.

[26] Fleming, W.H. and Rishel, R.W. (1976). Deterministic and Stochastic Optimal Con-
trol. Springer-Verlag, New York.

[27] Gerber, H.U. (1969). Entscheidungskriterien für den zusammengesetzten Poisson-

Prozess. Schweiz. Verein. Versicherungsmath. Mitt. 69, 185–228.

[28] Gerber, H.U. (1974). The dilemma between dividends and safety and a generalisa-
tion of the Lundberg-Cramer formulas. Scandinavian Actuarial Journal 46–57.

186
 Bibliography

[29] Gerber, H.U. (1979). An Introduction to Mathematical Risk Theory. Huebner Foun-
dation Monographs, Philadelphia.
[30] Gerber, H.U. (1981). On the probability of ruin in the presence of a linear dividend
barrier. Scandinavian Actuarial Journal 105–115.
[31] Gerber, H.U. and Shiu E.S.W. (1997). On the time value of ruin. Insurance: Math-
ematics and Economics 69, 145–199.
[32] Gerber, H.U. and Landry B. (1998). On the discounted penalty at ruin in a jump-
diffusion and the perpetual put option. Insurance: Mathematics and Economics 22,
263–276.
[33] Gerber, H.U. and Shiu, E.S.W. (2006) On the optimal dividend stategies in the
compound Poisson model. North American Actuarial Journal 10(2), 76–93.
[34] Gerber, H.U., Lin, X.S. and Yang, H. (2006) A note on the dividends-penalty
identity and the optimal dividend barrier. Astin Bulletin 36(2), 489–503.
[35] Gerber, H.U., Shiu, E.S.W. and Smith, N. (2008) Methods for estimating the op-
timal dividend barrier and the probability of ruin. Insurance: Mathematics and
Economics 42, 243–254
[36] Grandell, J. (1977) A class of approximations of ruin probabilities. Scandinavian
Actuarial Journal, 37–52.
[37] Grandell, J. (1991). Aspects of Risk Theory. Springer-Verlag, New York.
[38] Højgaard, B. and M. Taksar, (1999). Controlling risk exposure and dividends pay-
out schemes: Insurance company example. Mathematical Finance 9, 153–182.
[39] Heilmann, W. (1987). Grundbegriffe der Risikotheorie. VVW, Karlsruhe.
[40] Hipp, C. (2003). Optimal dividend payment under a ruin constraint: Discrete time
and state space. Blätter der DGVFM 26, 255–264.
[41] Hipp, C. and Plum, M. (2000). Optimal investment for insurers. Insurance: Math-
ematics and Economics 27, 215–228.
[42] Hipp, C. and Schmidli, H. (2004). Asymptotics of ruin probabilities for controlled
risk processes in the small claims case. Scandinavian Actuarial Journal, 321–335.
[43] Iglehart, D.L. (1969). Diffusion approximations in collective risk theory. Journal of
Applied Probability 6, 285–292.
[44] Ikeda, N. and Watanabe, S. (1989) Stochastic Differential Equations and Diffusion
Processes. Elsevier Science Publishers B.V., Amsterdam.

187
Bibliography

[45] Itô, K. and McKean, H.P.,Jr. (1974) Diffusion Processes and Their Sample Paths.
Springer-Verlag, Berlin.

[46] Jeanblanc-Picqué, M. and Shiryaev, A.N. (1995). Optimisation of the flow of divi-
dends. Russian Math. Surveys 50, 257–277.

[47] Kaas, R., Goovaerts, M., Dhaene, J. and Denuit M. (2008) Modern Actuarial Risk
Theory. Springer-Verlag, Berlin.

[48] Kamke, E. (1944) Differentialgleichungen Lösungsmethoden und Lösun-

gen. Akademische Verlagsgesellschaft Becker and Erler Kom.-Ges., Leipzig.

[49] Karatzas, I. and Shreve, S.E. (2000). Brownian Motion and Stochastic Calculus.
Springer-Verlag, New York.

[50] Kulenko, N. and Schmidli, H. (2008). Optimal dividend strategies in a Cramér–

Lundberg model with capital injections. Insurance: Mathematics and Economics
43, 270–278

[51] Lamberton, D. and Lapeyre, B. (2008). Introduction to Stochastic Calculus Applied

to Finance. Chapman and Hall, London.

[52] Lundberg, F. (1909). On the theory of Risk. Transactions of the Sixth International
Congress of Actuaries 1, p. 877

[53] Merton, R.C. (1969) Lifetime portfolio selection under uncertainty: The continuous-
time case. Review of Economics and Statistics 51, 247–257.

[54] Merton, R.C. (1971) Optimum consumption and portfolio rules in a continuous-time
model. Journal of Economic Theory 3, 373–413.

[55] Mikosch, T. (2004). Non-Life Insurance Mathematics. Springer-Verlag, Berlin.

[56] Müller-Möhl, E. (1999). Optionen und Futures. Schäffer-Poeschel Verlag,

Stuttgart.

[57] Paulsen, J. and Gjessing H.K., (1997). Optimal choice of dividend barriers for a
risk process with stochastic return on investments. Insurance: Mathematics and
Economics 20, 215–223.

[58] Paulsen, J. (2003). Optimal dividend payouts for diffusions with solvency con-
straints. Finance Stochast. 7, 457–473.

[59] Pitts, S.M. and Politis, K. (2007). Approximations for the Gerber-Shiu expected
discounted penalty function in the compound Poisson risk Model. Advances in Ap-
plied Probability 39, 385–406

188
 Bibliography

[60] Protter, P.E. (2004). Stochastic Integration and Differential Equations. Springer-
Verlag, Berlin.

[61] Rolski, T., Schmidli, H., Schmidt, V. and Teugels, J. (1999). Stochastic Processes
for Insurance and Finance. Wiley, Chichester.

[62] Revuz, D. and Yor, M. (1991). Continuous Martingales and Brownian Motion.
Springer, Berlin.

[63] Rogers, L.C.G. and Williams, D. (2000). Diffusions, Markov Processes and Martin-
gales 1. Foundations. Cambridge University Press, Cambridge.

[64] Rogers, L.C.G. and Williams, D. (2000). Diffusions, Markov Processes and Martin-
gales 2. Itô Calculus. Cambridge University Press, Cambridge.

[65] Rotar, V. (2000). Risk Processes. Cambridge University Press, Cambridge.

[66] Schmidli, H. (1992). A General Insurance Risk Model. Diss. ETH Nr. 9881, ETH
Zürich.

[67] Schmidli, H. (1994). Diffusion approximation for a risk process with the possibility of
borrowing and investment. Communications in statistics. Stochastic models 10(2),
365–388.

[68] Schmidli, H. (2001). Optimal proportional reinsurance policies in a dynamic setting.

Scand. Actuarial J., 55–68.

[69] Schmidli, H. (2002). On minimising the ruin probability by investment and reinsur-
ance. Annals of Applied Probability 12, 890–907.

[70] Schmidli, H. (2008). Stochastic Control in Insurance. Springer-Verlag, London.

[71] Schmidli, H. (2008). On the Gerber–Shiu Function and Change of Measure.

Preprint, University of Cologne.

[72] Shreve, S.E., Lehoczky, J.P. and Gaver, D.P. (1984). Optimal consumption for
general diffusions with absorbing and reflecting barriers. SIAM J. Control and Op-
timization 22, 55–75.

189
Bibliography

190
Index

adjustment coefficient, 13 light-tailed, 16

aggregate loss process, 13 local time, 176
loss process, 11
Black-Scholes model, 183 Lundberg’s inequality, 14
Brownian motion, 164
running maximum, 165 Markov process, 167
homogeneous, 167
cadlag, 163 martingale, 165
cedent, 15 local, 166
change of measure, 172 semi-, 166
classical risk model, 11 moment generating function, 13
contraction semigroup, 167
control net profit condition, 13
feedback, 21
Cramér-Lundberg model, 11 Poisson process, 164
predictable, 163
diffusion approximation, 17, 18
diffusion process, 17 reinsurance
excess of loss (XL), 15
expected value principle, 14 proportional, 15
reinsurer, 15
filtration, 163 retention level, 15
natural, 163 return function, 21
Gronwall’s lemma, 181 ruin, 12
severity of, 12
Hamilton–Jacobi–Bellman equation, 28, time, 12
42, 66, 76, 96, 109, 137, 144
heavy-tailed, 16 safety loading, 13
self-insurance function, 15
infinitesimal generator, 168 Skorokhod equation, 176
domain of, 168 standard deviation principle, 14
interarrival times, 11 stochastic differential equation, 175
Ito’s formula, 171 strong solution to, 175

191
Index

supremum norm, 167

transition function, 167

value function, 21, 27

variance principle, 14
variation, 163
bounded, 163
viscosity
solution, 109, 144
subsolution, 109, 144
supersolution, 109, 144

weak convergence, 18

192
Ich versichere, dass ich die von mir vorgelegte Dissertation selbständig angefertigt, die
benutzten Quellen und Hilfsmittel vollständig angegeben und die Stellen der Arbeit,
die den anderen Werken im Wortlaut oder dem Sinn nach entnommen sind, in jedem
Fall als Entlehnung kenntlich gemacht habe; dass diese Dissertation noch keiner anderen
Fakultät oder Universität zur Prüfung vorgelegen hat; dass sie - abgesehen von unten
angegebenen Teilpublikationen - noch nicht veröffentlicht worden ist sowie, dass ich
eine solche Veröffentlichung vor Abschluss des Promotionsverfahrens nicht vornehmen
werde. Die von mir vorgelegte Dissertation ist von Prof. Dr. Hanspeter Schmidli betreut
worden.

Köln, im Dezember 2009

Julia Eisenberg

List of Publications:

1. Optimal control of capital injections by reinsurance in a diffusion approximation.

(2009) Blätter der DGVFM 30(1), 1–13. (with H. Schmidli)

2. Minimising expected discounted capital injections by reinsurance in a classical risk

model. (2009) to appear in Scandianvian Actuarial Journal. (with H. Schmidli)

3. Minimising expected discounted capital injections by reinsurance with riskless rate

of interest. (In Vorbereitung)

4. Minimising expected discounted capital injections by reinsurance and investments.

(In Vorbereitung)