This action might not be possible to undo. Are you sure you want to continue?
Ruin Probabilities
Seren Asmussen
World Scientific
Ruin Probabilities
ADVANCED SERIES ON STATISTICAL SCIENCE & APPLIED PROBABILITY
Editor: Ole E. BarndorffNielsen
Published Vol. 1: Random Walks of Infinitely Many Particles by P. Revesz Vol. 2: Ruin Probabilities by S. Asmussen Vol. 3: Essentials of Stochastic Finance : Facts, Models, Theory by Albert N. Shiryaev Vol. 4: Principles of Statistical Inference from a NeoFisherian Perspective by L. Pace and A. Salvan Vol. 5: Local Stereology by Eva B. Vedel Jensen Vol. 6: Elementary Stochastic Calculus  With Finance in View by T. Mikosch Vol. 7: Stochastic Methods in Hydrology: Rain, Landforms and Floods eds. O. E. Barndorff Nielsen et al. Vol. 8: Statistical Experiments and Decisions : Asymptotic Theory by A. N. Shiryaev and V. G. Spokoiny
Ruin P robabilities
Soren Asmussen
Mathematical Statistics Centre for Mathematical Sciences Lund University
Sweden
World Scientific
Singapore • NewJersey • London • Hong Kong
Published by World Scientific Publishing Co. Pte. Ltd. P O Box 128, Fatter Road , Singapore 912805 USA office: Suite 1B, 1060 Main Street, River Edge, NJ 07661 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Library of Congress CataloginginPublication Data Asmussen, Soren
Ruin probabilities / Soren Asmussen. p. cm.  (Advanced series on statistical science and applied probability ; vol. 2) Includes bibliographical references and index. ISBN 9810222939 (alk. paper) 1. InsuranceMathematics. 2. Risk. I. Tide. II. Advanced series on statistical science & applied probability ; vol. 2. HG8781 .A83 2000 368'.01dc2l 00038176
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
First published 2000 Reprinted 2001
Copyright ® 2000 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
Printed by Fulsland Offset Printing (S) Pte Ltd, Singapore
Contents
Preface I ix
Introduction 1 1 The risk process . . . . . . . . . . . . . .. . . . .. .. . . . . 1 2 Claim size distributions .. . . . . . . . .. . . . . . . . . . . . 5 3 The arrival process . . . . . . . . . . . . . . . . . . . . . . . . 11 4 A summary of main results and methods . . . . .. . . . . . . 13 5 Conventions . .. . .. .. . . . . . . . . . . . . . . . . . . . . 19
II Some general tools and results 23 1 Martingales . .. . .. .. . . . . . .. . . . . . . . . . . . . . 24 2 Likelihood ratios and change of measure . . .. . . . . . .. . 26 3 Duality with other applied probability models . . .. . . . . . 30 4 Random walks in discrete or continuous time . . . . . . . . . . 33 5 Markov additive processes . . . . . . . .. . . . . . . . . . . . 39 6 The ladder height distribution . . . .. . .. .. . . . . . . . . 47
III The compound Poisson model 57 1 Introduction . . . . . . . . .. .. .. . .. .. . . . . . . 58 . . . . . . . . . . . . . . . 61 3 Special cases of the PollaczeckKhinchine formula . . . . . . . 62 4 Change of measure via exponential families . . . .... . .. . 67 5 Lundberg conjugation . .. . . . . . . . . . . . . . . . . . . . . 69 6 Further topics related to the adjustment coefficient .. . . . . 75 7 Various approximations for the ruin probability . . . . . . . . 79 8 Comparing the risks of different claim size distributions . . . . 83 9 Sensitivity estimates . . . . . . . . . . . . . . . . . . . . . . . 10 Estimation of the adjustment coefficient . . . . . . . . . . . . 86 93 2 The PollaczeckKhinchine formula
v
vi
CONTENTS
IV The probability of ruin within finite time 97 1 Exponential claims . . . . . . . . . . . . . . . . . . . . . . . . 98 2 The ruin probability with no initial reserve . . . . . . . . . . . 103 3 Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . 108 4 When does ruin occur? . . . . . . . . . . . . . . . . . . . . . . 110 5 Diffusion approximations . . . . . . . . . . . . .. . . .. . . . 117 6 Corrected diffusion approximations . . . . . . . . . . .. . . . 121 7 How does ruin occur ? . . .. . . . . . . . . . . . . . . . . . . . 127 V Renewal arrivals 131 1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . 131 2 Exponential claims. The compound Poisson model with negative claims . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 3 Change of measure via exponential families . . . . . . . . . . . 137 4 The duality with queueing theory .. .. .. . . . .. . . . . . 141 VI Risk theory in a Markovian environment 145 1 Model and examples . . . . . . . . . . . .. . .. . . . . . . . 145 2 The ladder height distribution . . . . . . . . . .. . . . . . . . 152 3 Change of measure via exponential families ........... 160 4 Comparisons with the compound Poisson model ........ 168 5 The Markovian arrival process . . . . . . .. .. . . ... . . . 173 6 Risk theory in a periodic environment .. . . . .. . . . . . . . 176 7 Dual queueing models .... ... ................ 185 VII Premiums depending on the current reserve 189 1 Introduction . . . . . . . . . . . . . . . . . . . .. . . . . . . . 189 2 The model with interest . . . . . .. . . . . . . . . . .. . . . 196 3 The local adjustment coefficient. Logarithmic asymptotics . . 201 VIII Matrixanalytic methods 215 1 Definition and basic properties of phasetype distributions .. 215 2 Renewal theory . . . . . . . . . . . . . . . . . . . . . . . . . . 223 3 The compound Poisson model . . . . . . . . . .. . . . . . . . 227 4 The renewal model . . . . . . . . . . . . . . . .. . . . . . . . 229 5 Markovmodulated input . . .. . . . . . . . . . . . . . . . . . 234 6 Matrixexponential distributions . . . . . . . . . . . .. . . . 240 7 Reservedependent premiums . . . . .. . . . .. . . . . . . . 244
. .. . . . . . . 350 Bibliography Index 363 383 . . . . . . . . . . . . . . . . . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292 6 Sensitivity analysis . . . 336 A3 Matrixexponentials . . . . . . . . . . . . . . . . . .. . . . . .. . . . . . 326 Appendix 331 Al Renewal theory . . . . . . 344 AS Complements on phasetype distributions . . . . . . 271 6 Reservedependent premiums . . . . . . . . . . . . . .. . . . . . . .. . . . . . 259 3 The renewal model . . . . . 279 X Simulation methodology 281 1 Generalities .. . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions .. . . . . .. . . . . . . .. . . . . . . . . . . . . . . . . . . 287 4 Importance sampling for the finite horizon case . . . . . . . . . . . . . . . . . . . . 261 4 Models with dependent input . .. 340 A4 Some linear algebra . . . . . . .. 331 A2 WienerHopf factorization . . . 304 3 Large deviations . . . . . . . . . . . . . . . . . .. .. . . ... . . . . . . . . 316 5 Principles for premium calculation . .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .. . . . . . . . . . . . . . . . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . . . .. . . . . . . . . .. . . . . . . . . . . 290 5 Regenerative simulation . . . . . . . . . . . . . . . . . .. . The twobarrier ruin problem . . . . . 306 4 The distribution of the aggregate claims . . . . . . . . 251 2 The compound Poisson model . .. . . .. . . . . 285 3 Importance sampling via Lundberg conjugation . . . . . . . . . . . . . . . . . . . . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . 297 2 Further applications of martingales . 323 6 Reinsurance . . . . . .. .
This page is intentionally left blank .
University of Copenhagen. the book is basically mathematical in its flavour. this applies to longrange dependence which is intensely studied in the neighboring ix . Thus. The course was never realized. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. As an excuse: many of these projects were related to the book. and other projects absorbed my interest. But the pace was much slower than expected. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. and my belief was that this could be done rather quickly. Apart from these remarks. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. In particular. but the handouts were written and the book was started (even a contract was signed with a deadline I do not dare to write here!). that it can only say something about very simple models and questions. and the series editor Ole BarndorffNielsen for their patience. it is not by intention. and the result is now that the book is much more related to my own research than the initial outline. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. the idea was close to expand these to a short book on the subject. and has been an active area of research from the days of Lundberg all the way up to today. A similar thank goes to all colleagues who encouraged me to finish the project and continued to refer to the book by Asmussen which was to appear in a year which continued to be postponed. if the formulations occasionally give a different impression. I have deliberately stayed away from discussing the practical relevance of the theory. It has obviously not been possible to cover all subareas. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. However. Since I was to produce some handouts for the students anyway. which has in particular removed one of the standard criticisms of the area.
Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. Concerning ruin probabilities. see also Schmidli [325] and the references in Asmussen & Taksar [52]. an area which is becoming increasingly important.14. Willinger et al.89. the first part of 11. see in particular Michna [259].13 and IX. for the effects on tail probabilities. One is by model. A book like this can be organized in many ways. for which I apologize to the reader and the authors of the many papers who ought to have been on the list. The rest is up to your specific interests.lth.1. [381]). Here is a suggestion on how to get started with the book.3. it has not been possible to incorporate more numerical examples than the few there are. I intend to keep a list of misprints and remarks posted on my web page. Good luck! I have tried to be fairly exhaustive in citing references close to the text. Asmussen. VIII. The main motivation comes from statistical data for network traffic (e.g. another by method. In the classical setting of CramerLundberg models.g. e. Finally.13. I regret that due to time constraints.4a. Another interesting area which is not covered is dynamic control.2. incorporate 11. the standard stochastic control setting of diffusion models has been considered.x PREFACE field of queueing theory.13 and XI. VI. IX. IV. 111.2.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . The present book is in between these two possibilities. X. see e. 111. read Chapter I. In addition. VII.g. Chapters IXX then go in more depth with some of the special approaches for analyzing specific models and add a number of results on the models in Chapters IIIVII (also Chapter II is essentially methodological in its flavor). It is obvious that such a system involves a number of inconsistencies and omissions. IV. VII. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. For a brief orientation. http:// www. For a second reading.5.lth.2 more properly). Resnick & Samorodnitsky [303] and references therein.15. some papers not cited in the text but judged to be of interest are included in the Bibliography.45. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance.13. IV.se Lund February 2000 Soren Asmussen . More recently.6 (to understand the PollaczeckKhinchine formula in 111.maths . Hojgaard & Taksar [206].
6 by my 1999 simulation class in Lund. 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers. of which there are not many at this stage . IV.6 is reprinted from Asmussen & Schmidt [49] and parts of IX.1 by Bjarne Hojgaard and the table in Example 111. 5. not least the more complicated ones. More substantial remarks.PREFACE xi The second printing differs from the first only by minor corrections. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . Parts of II.8 .1 and X.4 from Asmussen. 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. many of which were pointed out by Hanspeter Schmidli . Section VIII. Aarhus. Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . A number of other figures were supplied by Christian Geisler Asmussen . Parts of X. . Section VII .2 by Rafal Kulik . Fig.5 from Asmussen [21] with permission from CRC Press.6. 111 . Fig. were produced by Lone Juul Hansen . supported by Center for Mathematical Physics and Stochastics (MaPhySto).3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . as well as some additional references continue to be at the web page.
This page is intentionally left blank .
Chapter I Introduction 1 The risk process In this chapter . The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero.3) sup St. M = (1. T) as ruin probabilities with infinite horizon and finite horizon . (1. as defined in broad terms . A risk reserve process { Rt}t>o. respectively. (1. and give a very brief summary of some of the models. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. They are the main topics of study of the present book. (1.Rt. MT = sup St. results and topics to be studied in the rest of the book. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t. we introduce some general notation and terminology. For mathematical purposes. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u .2) (O<t<T Ro=ul.4) O<t<oo O<t<T 1 . We denote throughout the initial reserve by u = Ro.T) = P inf Rt < 0 I .i(u.1) We also refer to t/) ( u) and 0(u. is a model for the time evolution of the reserves of an insurance company.
per unit time.1 .6) Sofar we have not imposed any assumptions on the risk reserve process. Putting things together.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un.5) i. Thus. . However. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals..i(u. (1. That is. t] is finite.E Uk..b(u) = P (r(u) < oo) = P(M > u). and T1 is the time of the first claim. St = E Uk . the time of arrival of the nth claim is an = T1 + • • • + Tn. (1. 1.2 CHAPTER I.1. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. we see that Nt Nt Rt = u + pt . say.pt. We denote the interarrival times of claims by T2. • Premiums flow in at rate p. the ruin probabilities can then alternatively be written as . respectively. (1.T) = F (MT > u) = P(r(u) < T). the number Nt of arrivals in [0. Figure 1. T3.
We shall discuss Brownian motion somewhat in Chapter IV.b(u) = 1 for all u. 1.. We shall not deal with this case either. If 77 < 0.1 the slope of {Rt} should depend also on the level). • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. we shall. then M = oo a. allowing a countable infinity of jumps on Fig.s.s. It would appear obvious. since any modeling involves some approximative assumptions. For the purpose of studying ruin probabilities this distinction is. and hence . 1. not discuss whether this actually corresponds to practice.(.1. say 10% . a basic references is Gerber [127]. • Brownian motion or more general diffusions. and hence O(u) < 1 for all sufficiently large u. . A further basic quantity is the safety loading (or the security loading) n defined as the relative amount by which the premium rate p exceeds p.1. but as an approximation to the risk process rather than as a model of intrinsic merit. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i.20%. that the insurance company should try to ensure 77 > 0. for example. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. However.8) holds. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small.) V 0. THE RISK PROCESS 3 Note that it is a matter of taste (or mathematical convenience) whether one allows {Rt} and/or {St} to continue its evolution after the time T(u) of ruin. though many results are straightforward to generalize from the compound Poisson model.. and the basic ruin probabilities are derived in XI.e. Thus.1. one could well replace Rt by Rtnr(u) or RtA. t * oo.1 Assume that (1. one may well argue that Brownian motion in itself could be a reasonable model. on Fig. however. of course. then M < oo a.8) The interpretation of p is as the average amount of claim per unit time. (1. We study this case in Ch. however. and in fact: Proposition 1. immaterial. VII. If 77 > 0.
oo t 0 J (provided the limit exists).. and that . Here it is easy to see that p = . (1.4 CHAPTER I.d. U2.10) hold with p constant.i. we obtain typically a somewhat stronger conclusion.6EU (on the average.Q claims arrive per unit time and the mean of a single claim is EU) and that also Nt t aoo t lira EEUk = p. zP(u . corresponding to the Pdlya process. (1.i(u. However..2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t.3 Assume p 54 1 and define Rt = Rt1p.. namely.11) . If u oo. M < oo a. this needs to be verified in each separate case.. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). t t p  p' t ^ oo. are i..d. U2. are i.10) Again. This case is referred to as the mixed Poisson process. (1. not all models considered in the literature have this feature: Example 1. 0 We shall only encounter a few instances of a Cox process. _ St __ k =1 Uk pt a4. Thus p may well be random for such processes. However. . tb(u) = 1 for all u holds also when rl = 0.10) is a property which we will typically encounter. then similarly limSt/t < 0.Q (say) and U1. and independent of {Nt}.T) = i. namely that M = oo a. INTRODUCTION Proof It follows from (1.v. 0(u.b(u) < 1 for all u when rl > 0. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. in connection with risk processes in a Markovian or periodic environment (Chapter VI). Proposition 1. with the most notable special case being V having a Gamma distribution. . . The simplest example is 3(t) = V where V is a r .s. rl > 0.8). St In concrete models.8) that F N. where {Nt} is a Poisson process with rate .s. and here (1.i. and independent of {(0(t). k=1 (1.8) is given by ^t p = EU • lim it (3(s) ds t. If 77 < 0. it is not too difficult to show that p as defined by (1.s. Nt)}. then this limit is > 0 which implies St a$ oo and hence M = oo a. if {(3(t)} is nonergodic.Tp). If U1.T) for {Rt} is given by V)(u) = t/i (u).
Gerber [159]) has a rather different flavour. Straub [353]. Schmidli. Schmidt & Teugels [307] and Seal [326]. Segerdahl [334] and Philipson [289]. [134]. Gerber [157]. Embrechts et al. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. lighttailed distributions (sometimes the term .. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. was largely initiated in Sweden in the first half of the century. Cox processes are treated extensively in Grandell [171]. Grandell [171]. often referred to as collective risk theory or just risk theory. Daykin. see also Chapter XI. another important early Swedish work is Tacklind [373]. and we do not get near to the topic anywhere in this book..2. Since { Rt } has premium rate 1.g. Sundt [354]. 2 Claim size distributions This section contains a brief survey of some of the most popular classes of distributions B which have been used to model the claims U1. in particular. in a number of models. Heilmann [191]. Buhlmann [82]. see e . while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. Taylor [364]. For mixed Poisson processes and Polya processes. the recent survey by Grandell [173] and references therein. Rolski. Some of the main general ideas were laid down by Lundberg [250]. Besides in standard journals in probability and applied probability. The Swedish school was pioneering not only in risk theory. Daykin et al. Note that when p = 1. the role of the result is to justify to take p = 1. the research literature is often published in journals like Astin Bulletin . [76].. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. the claim arrivals are Poisson or renewal at the same time). many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory.g. U2. We roughly classify these into two groups . Some early surveys are given in Cramer [91]. but in probability and applied probability as a whole.. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Note that life insurance (e. Hipp & Michel [198]. [330]. In the even more general area of nonlife insurance mathematics. Insurance: Mathematics and Economics. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. we shall be able to identify p with the traffic intensity of an associated queue. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). Some main later textbooks are (in alphabetical order) Buhlmann [82]. the assumption > 0 is equivalent to p < 1. Pentikainen & Pesonen [101]. De Vylder [110]. [101]. Notes and references The study of ruin probabilities.
and heavytailed distributions.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p.B(x) satisfies B(x) = O(e8x) for some s > 0. the m.g. 2a Lighttailed distributions Example 2. (2. then the conditional distribution of X .g.1) The parameter 6 is referred to as the rate or the intensity. In particular. for the compound Poisson model with exponential claim sizes the ruin probability . but different more restrictive definitions are often used: subexponential.f. where B(bo. the exponential distribution is by far the simplest to deal with in risk theory as well.O(u) can be found in closed form. For example in the compound Poisson model. B is heavytailed if b[s] = oo for all s > 0. B[s] is finite for some s > 0. 6 has density r(p)xPleax b(x) P and m. Here lighttailed means that the tail B(x) = 1 . INTRODUCTION 'Cramertype conditions' is used).8.6 CHAPTER I. regularly varying (see below) or even regularly varying with infinite variance.2) = 0.3) . In contrast. Equivalently. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . a fact which turns out to contain considerable information.f.u at the time of ruin given r(u) is again exponential u with rate 8. P B[s]= (8Is ) . i. As in a number of other applied probability areas. and can also be interpreted as the (constant) failure rate b(x)/B(x).x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). one could mention also the folklore in actuarial practice to consider B heavytailed if '20% of the claims account for more than 80% of the total claims'.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2.2 and /LB is the mean of B.e. On the more heuristical side. Example 2 . if 1 °O AB Jbos x B(dx) > 0. s<8. The crucial feature is the lack of memory: if X is exponential with rate 6.
) VarX1 (EX )2 p is < 1 for p > 1. 2. 0.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. P b(x) = r` aibiea. 0 < ai < 1. p. This special case is referred to as the Erlang distribution with p stages. where X1. . An important property of the hyperexponential distribution is that its s.. In particular..v. X2.2) can be considered as the pth power of the exponential density (2.d.1 Poisson events in [0. x] so that B(x) = r` e.v. the squared coefficient of variation (s.. The exact form of the tail B(x) is given by the incomplete Gamma function r(x. In particular.. the Gamma density (2.. is > 1. > 1 for p < 1 and = 1 for p = 1 (the exponential case). i = 1. and exponential with rate d. p) °° where r (x. .. among others. p) = J tPletdt. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. . if p is integer and X has the gamma distribution p. are i.c..y i=1 where >i ai = 1.2. Ruin probabilities for the general case has been studied. u Example 2 .1) (or the 1/pth root if p < 1). then X v Xl + • • • + X.c.). one has r(bx.i. we develop computationally tractable results mainly for the Erlang case (p = 1. p).ate (b2 ): L• i=o In the present text. by Grandell & Segerdahl [175] and Thorin [369]. B(x) = r(p) Asymptotically. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. u . or just the Erlang(p) distribution. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p .
but the current trend in applied probability is to restrict attention to the class of phasetype distributions. We give some theory for matrixu exponential distribution in VIII. equivalently.7) are possibly complexvalued but the parameters in (2.f.8) are realvalued. Example 2 .e. This class of distributions is popular in older literature on both risk theory and queues. then the claim size which is relevant from the point of view of the insurance company itself is U A xo rather than U u (the excess (U . . resp. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. The parameters of a phasetype distribution is the set E of transient states.8) j=1 j=1 j=1 where the parameters in (2. q2 q3 (2.1 and defer further details to u Chapter VIII. The couple (a. T) is called the representation. 1)' is the column vector with 1 at all entries.(2.6. The density and c. are b(x) = aeTxt.. a. B(x) = aeTxe where t = Te and e = (1 .4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. However.8 CHAPTER I.d. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution.f. Example 2 .xo)+ is covered by the reinsurer). We give a more comprehensive treatment in VIII. there exists a xo < oo such that B(x) = 0 for x > xo.. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix .6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i.g. T) or sometimes the triple (E.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. INTRODUCTION Example 2 . (or. the Erlang and the hyperexponential distributions. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities. Important special cases are the exponential. This class of distributions plays a major role in this book as the one within computationally tractable exact forms of the ruin probability z/)(u) can be obtained. which is slightly smaller but more amenable to probabilistic reasoning. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. of which one is absorbing and the rest transient. See XI.6.
N(0. a)/A)a+1' x > a. Writing c = d/r. or equivalently as the distribution of a°U+µ where U . a2).10) The loinormal distribution has moments of all orders. In particular. p is defined as the distribution of ev where V . (2. the tail is B (x ) 2 x.1).13) u .u l b(x) = d dx or J ax lor 1 exp Asymptotically. Example 2 .pl = 1 W (logx .11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. the mean u is eµ+a /2 and the second moment is e2µ+2o2. (2. x < a. There are various variants of the definition around.p a 1 (2. in practice one may observe that b(x) is either decreasing or increasing and may try to model smooth (incerasing or decreasing) deviations from constancy by 6(x) = dx''1 (0 < r < oo).1. (2. It follows that the density is 't (1ogX .8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2.9) which is heavytailed when 0 < r < I.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. and then b(x) = 0.N(p. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. u Example 2 .7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. b(x) = crx''le`xr. we obtain the Weibull distribution B(x) = eCx'. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a .2.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed. Here failure rates b(x) = b(x)/B(x) play an important role. the exponential distribution representing the simplest example since here b(x) is constant. However. All moments are finite.
examples of distributions with regularly varying tails are the Pareto distribution (2. in particular.10 CHAPTER I. A = 1 and X is standard exponential.(1 + 2x + 2x2)e2x) p = 2 (2. the loggamma distribution is a Pareto distribution. another standard example is (log x)').16) 11 Example 2. INTRODUCTION Example 2. Choudhury & Whitt [1] as the class of distributions of r.17) where L (x) is slowly varying. u Example 2 . B(x) = O(xP).13).e.1)/p. The motivation for this class is the fact that the Laplace transform is explicit (which is not the case for the Pareto or other standard heavytailed distributions). Thus. u . (2. x + 00.15) x2 + 16x3 ) a3x/2) 3 (1 .11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. the density is { 3 (1 .2). x 4 oo (any L having a limit in (0. (2.x6+lr(p) (2.v. The density is 8p(log x)pi b(x) .12 (DISTRIBUTIONS WITH REGULARLY VARYING TAILS) The tail B(x) of a distribution B is said to be regularly varying with exponent a if B(x) . The simplest examples correspond to p small and integervalued. oo) is slowly varying .10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. in particular. where Y is Pareto distributed with a = (p .14) The pth moment is finite if p < 5 and infinite if p > 5. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. 6 is defined as the distribution of et' where V has the gamma density (2.(1 + Zx + $ p = 3. i. satisfies L(xt)/L(x) 4 1. In general.L( x ). For p = 1.'s of the form YX.12) (here L (x) * 1) and ( 2. { s () 1s+3s29s3log(1+2s I p=3.
. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. the knowledge of the claim size distribution will typically be based upon statistical data. which each have a ( timehomogeneous) small rate of experiencing a . The reason is in part mathematical since this model is the easiest to analyze.1) that any distribution with a regularly varying tail is subexponential.. THE ARRIVAL PROCESS 11 Example 2.. and based upon such information it seems questionable to extrapolate to tail behaviour. one may argue that this difficulty is not resticted to ruin probability theory alone. At least as important is the specification of the structure of the point process {Nt } of claim arrivals and its possible dependence with the claims. However. Similar discussion applies to the distribution of the accumulated claims (XI. (2. but the model also admits a natural interpretation : a large portfolio of insurance holders . When studying ruin probabilities. for example the lognormal distribution is subexponential (but not regularly varying).4) or even to completely different applied probability areas like extreme value theory: if we are using a Gaussian process to predict extreme value behaviour. U2. though the proof of this is nontrivial.3.1. Namely. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. the claim size distribution represents of course only one aspect (though a major one). but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. and so is the Weibull distribution with 0 < r < 1. we may know that such a process (with a covariance function estimated from data) is a reasonable description of the behaviour of the system under study in typical conditions.18) B(x) It can be proved (see IX. We return to a closer study in IX. 3 The arrival process For the purpose of modeling a risk process . From a practical point of view.. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. Thus. Also. this phenomenon represents one of the true controversies of the area.
. The compound Poisson model is studied in detail in Chapters III. see 11.6. but with a general not necessarily exponential distribution ). e. getting away from the simple Poisson process seems a crucial step in making the model more realistic. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). not many detailed studies of the goodnessoffit of the Poisson model in insurance are available .i. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. such that 8(t) = . its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. the periodic and the Markov modulated models also have attractive features . found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . However . in particular to allow for certain inhomogeneities. the negative binomial distribution. In order to prove reasonably substantial and interesting results . it is more questionable whether it provides a model with a similar intuitive content as the Poisson model.. A more appealing way to allow for inhomogeneity is by means of an intensity . it may be used in a purely descriptive way when it is empirically observed that the claim arrivals are more bursty than allowed for by the simple Poisson process. The difficulty in such an approach lies in that it may be difficult or even impossible to imbed such a distribution into the continuous setup of {Nt } evolving over time . however. Mathematically. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate).. To the author 's knowledge . gives rise to an arrival process which is very close to a Poisson process. are i. we study this case in VI . with a common term {Nt} is a Markovmodulated Poisson process . with the extension to premiums depending on the reserve. 5. has some mathematically appealing random walk features . An obvious example is 3(t) depending on the time of the year (the season).3(t) fluctuating over time. and also that the ruin problem may be hard to analyze . when Jt = i.g.d. so that ..12 CHAPTER I. IV (and. which facilitate the analysis. INTRODUCTION claim . Historically. Nevertheless . Another one is Cox processes. too general and one neeed to specialize to more concrete assumptions . the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. In others. This model . T2. epidemics in life insurance etc. Cox processes are. I.8 (t) is a periodic function of t. to be studied in Chapter V. This applies also to the case where the claim size distribution depends on the time of the year or .e.(3. where {/3 (t)}too is an arbitrary stochastic process . in Chapter VII). Some of them have concentrated on the marginal distribution of NT (say T = one year ).
The study of the steady state is by far the most dominant topic of queueing and storage theory. stochastic differential equations. 4 A summary of main results and methods 4a Duality with other applied probability models Risk theory may be viewed as one of many applied probability areas.1) holds as well provided the risk process has a premium rule depending on the reserve. that quite often the emphasis is on computing expected values like EV. In fact. the classical result is that the ruin probabilities for the compound Poisson model are related to the workload (virtual waiting time) process {Vt}too of an initially empty M/G/1 queue by means of .1) where V is the limit in distribution of Vt as t + oo. and the limit t 4 oo is the steadystate limit. others are quite different. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. reliability. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps.'s like V is available. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. It should be noted. More generally. R = p(R) in between jumps. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models.4. and which seems well motivated from a practical point of view as well. this gives only f0 O°i (u)du which is of limited . with Poisson arrivals and constant release rule p(x) = 1. dam/storage processes. (4. The ones which appear most related to risk theory are queueing theory and dam/storage processes. Mathematically. 0(u) = P(V > u). point processes and so on. Similarly. others being branching processes.6) . time series and Gaussian processes. it is desirable to have a set of formulas like (4. interacting particle systems. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. Some of these have a certain resemblance in flavour and methodology. and here (4. and a lot of information on steadystate r.1) permitting to translate freely between risk theory and the queueing/storage setting.1). Markovmodulated or periodic can be related to queues with similar characteristics. methods or modeling ideas developed in one area often has relevance for the other one as well. this amounts to Vo having the stationary distribution of {Vt}).v. queueing theory. however.0 (u. In the setting of (4. extreme value theory. stochastic geometry. genetics models. ruin probabilities for risk processes with an input process which is renewal. Thus.T) = P(VT > u).
8. which gives a sample path version of (4. Similarly. B(x) = ebx.v.1.p(y) y^ Jo p(x) can be written in closed form. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. the functions w x f d 1 exdx () . see Boxma & Cohen [74] and Abate & Whitt [3]. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. much of the study of finite horizon problems (often referred to as transient behaviour) in queueing theory deals with busy period analysis which has no interpretation in risk theory at all.6.2). which can be expanded into a sum of exponential terms by diagonalization (see. INTRODUCTION intrinsic interest . • The compound Poisson model with some rather special heavytailed claim size distributions. as is typically the case. A prototype of the duality results in this book is Theorem 11. 3. have to some extent a different flavour.T). The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions. Example VIII. 3.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. Here ?P(u) is explicit provided that . though overlapping. the two areas. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .14 CHAPTER I.1).g.3. e .. 4b Exact solutions Of course . . • The compound Poisson model with a claim size distribution degenerate at one point. see Corollary III. Here O(u) = peryu where 3 is the arrival intensity. The cases where this is possible are basically the following for the infinite horizon ruin probability 0(u): • The compound Poisson model with constant premium rate p = 1 and exponential claim size distribution B. The fact that Theorem H.3. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u).1) in the setting of a general premium rule p(x): the events {VT > u} and {r (u) < T} coincide when the risk process and the storage process are coupled in a suitable way (via timereversion ). Vi(u. The infinite horizon (steady state ) case is covered by letting T oo.'s like the environmental process {Jt} in a Markovmodulated setting. see Corollary VII.3.1 .3. Thus . p = 0/8 and y = 8 .
say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model.7. For the finite horizon ruin probability 0(u.u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. T) can then be calculated numerically by some method for transform inversion. Embrechts. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI . A notable fact ( see again XI. relevant references are Grubel [179]. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. Here are some of the main approaches: Laplace transform inversion Often.4. it is easier to find the Laplace transforms = f e8 . see VIII. 191).2) is the natural scale. but are somewhat out of the mainstream of the area . We don't discuss Laplace transform inversion much.ff 2µ(y)/a2(y) dy} dx . T) themselves. f {eXp U LX 2. T). esuTb( u.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). Given this can be done. Abate & Whitt [2]. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). O(u. 1).Lef$er function. a2 (x): Ip (u) = where S(u) = f °O exp {. A SUMMARY OF MAIN RESULTS AND METHODS 15 • The compound Poisson model with a two step premium rule p(x) and B being phasetype with just a few phases.S(u) 1S(oo) f °D exp {. (u.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions.b(u)du . the formulas ( IV. Ab(u). However. • An astable Levy process with drift . the only example of something like an explicit expression is the compound Poisson model with constant premium rate p = 1 and exponential claim size distribution . .f f 2µ(y)/a2(y) dy} dx  (4. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. [s.
U is explicit in terms of the model parameters. which can equivalently be written as f3 [7] = 1 +13 .or integral equation. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. Differential.g. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. . T) as the solution to a differential. either as the iterative solution of a fixpoint problem or by finding the diagonal form in terms of the complex roots to certain transcendental equations. it states that i/i(u) . In the compound Poisson model with p = 1. An example where this idea can be carried through by means of a suitable choice of supplementary variables is the case of statedependent premium p(x) and phasetype claims.) B[s].4) 00['Y]1)'Y = 0.p)/(13B'[ry] .16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). However.and integral equations The idea is here to express 'O(u) or '(u. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). u * oo. and in quite a few cases (Chapter VIII). dt] most often leads to equations involving both differential and integral terms. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). and carry out the solution by some standard numerical method.3.3) where C = (1 . see VIII. (4. One example where this is feasible is the renewal equation for tl'(u) (Corollary III.f.1) and y > 0 is the solution of the Lundberg equation (4.7.Ce"u.3) in the compound Poisson model which is an integral equation of Volterra type. and in particular the naive idea of conditioning upon process behaviour in [0.
corrected diffusion approximations (see IV. some further possibilities are surveyed in 111 . incorporating correction terms may change the picture dramatically. Large claims approximations In order for the CramerLundberg approximation to be valid. In the case of heavytailed distributions. However. when the claim size distribution is of phasetype.2. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. Diffusion approximations Here the idea is simply to approximate the risk process by a Brownian motion (or a more general diffusion) by fitting the first and second moment. In fact. a Markovian environment or periodically varying parameters. This list of approximations does by no means exhaust the topic. T). (4. other approaches are thus required. Approximations for O(u) as well as for 1(u.7 and IV. u > oo. T) for large u are available in most of the models we discuss.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. It has generalizations to the models with renewal arrivals. A SUMMARY OF MAIN RESULTS AND METHODS 17 It is rather standard to call ry the adjustment coefficient but a variety of other terms are also frequently encountered. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. the claim size distribution should have an exponentially decreasing tail B(x). For example. Diffusion approximations are easy to calculate. for the compound Poisson model ^(u) p pu In fact . See Chapter IX. often for all u > 0 and not just for large u. In particular. in some cases the results are even more complete than for light tails. but typically not very precise in their first naive implementation. J B dx. However. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. . in such cases the evaluation of C is more cumbersome.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu.4.
lower bounds etc. and to plot the empirical mean residual life 1 N .x U > x] = B(x) f '(yx)B(dx) typically has a finite limit (possibly 0) in the lighttailed case and goes to oo in the heavytailed case.k (U(`) . empirical evidence shows that the general principle holds in a broad variety of settings. . . it splits up into the estimation of the Poisson intensity (the estimator is /l3 = NT/T) and of the parameter(s) of the claim size distribution.18 CHAPTER I. it has the advantage of not involving approximations and also. This procedure in itself is fairly straightforward.d.8. UNT may be viewed as an interpolation in or smoothing of the histogram). more importantly.i. UNT are i. in the compound Poisson model.g. the difficulty comes in when drawing inference about the ruin probabilities. . which is a standard statistical problem since the claim sizes Ui. This is proved for the compound Poisson model in 111. one expects a model with a deterministic claim size distribution B. INTRODUCTION Compared to the CramerLundberg approximation (4. .U(k)) i =k+ i . obtained say by observing the risk process in [0. .) at various places and in various settings. However. though not too many precise mathematical results have been obtained. For example. In practice. it is a general principle that adding random variation to a model increases the risk. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail.. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known.3). The standard suggestion is to observe that the mean residual life E[U . of being somewhat easier to generalize beyond the compound Poisson setting. e. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. T]. fitting a parametric model to U1. However. as a general rule.. For example. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. How do we produce a confidence interval? And. say degenerate at m. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. . When comparing different risk models. this is extrapolation from data due to the extreme sensitivity of the ruin probabilities to the tail of the claim size distribution in particular (in contrast. they have however to be estimated from data.. given NT.
in all other chapters than VI where we just write . and in fact methods from that area can often be used in risk theory as well . respectively.4.e. reference [14]. but is not very satisfying.3) or Section 3 of Chapter VI are referred to as Proposition VI.i. Klnppelberg & Mikosch [134].2. UN.v's) which can be generated by simulation. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. The chapter number is specified only when it is not the current one. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. The infinite horizon case presents a difficulty. Thus Proposition 4. formula VI. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory... Truncation to a finite horizon has been used.3). in this book referred to as [APQ]).. . See further Embrechts.3 (or just VI. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. < U(N) are the order statistics based upon N i. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. where U(1) < . Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. because it appears to require an infinitely long simulation. For example. CONVENTIONS 19 as function of U(k). Still.(5. to observe whether one or the other limiting behaviour is apparent in the tail. and of course the method is relevant in risk theory as well..d. claims U1.2.. We look at a variety of such methods in Chapter X. formula (5. However. A main problem is that ruin is typically a rare event (i.5. . (or a functional of the expectation of a set of r.3) and Section VI.v. this is a straightforward way to estimate finite horizon ruin probabilities. and look at them to see whether they exhibit the expected behaviour or some surprises come up.
or a more precise one like eh .g. r. mation.ceax. If. random variable s. independent identically distributed i.g. with respect to w.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. B(x) = P(X < x) = fx. as for typical claim size distributions. oo). b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. h + 0.p. cumulant generating function. B[s] the m. .s. . EX2/(EX)2.d.g.f. E expectation. log E[s] where b[s] is the m.o.g. (A.v.r. B(x) the tail 1 .c. n! 27r nn+1/2en. In particular.B(x) = P(X > x) of B. E. i. B(dy). say a heuristic approxi1 + h + h2/2. and for a defective probability distribution IIGII < 1.f.g.i.h.f. for a probability distribution IIGII = 1. with probability Mathematical notation P probability. (moment generating function) fm e82B(dx) of the distribution B.20 CHAPTER L INTRODUCTION Proposition 4. n i oo.4. References like Proposition A.3) or Section 3.f. B is concentrated on [0.h.29) refer to the Appendix.d. infinitely often l. moment generating function.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.s.e. A different type of asymptotics: less precise.d. if B(x) . Abbreviations c.2. right hand side (of equation) r. cumulative distribution function P(X < x) c. formula (5. see under b[s] below.f. IIGII the total mass (variation ) of a (signed ) measure G . The Laplace transform is b[s].f.v. i. w.t. squared coefficient of variation. then for 1s < 5). left hand side (of equation) m.
Matrices and vectors are denoted by bold letters. N(it.. (the dimension is usually clear from the context and left unspecified in the notation). i. . matrices have uppercase Roman or Greek letters like T. Usually. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B.. i. E[X. xa.5. In the Frenchinspired literature. Thus. intensity interpretation. R(s) the real part of a complex number s. I(A) the indicator function of the event A. an example or a remark. of numbers. Notation like f3i and 3(t) in Chapter VI has a similar . all stochastic processes considered in this book are assumed to have sample paths in this space.e.A] means E[XI(A)]. the value just before t. the processes we consider are piecewise continuous. and column vectors have lowercase Roman letters like t. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. a2) the normal distribution with mean p and variance oa2. F o r a given set x1. Xt_ the left limit limstt X8f i. Unless otherwise stated. In particular: I is the identity matrix e is the column vector with all entries equal to 1 ei is the ith unit column vector.e. the ith unit row vector is e'i. oo) the space of Rvalued functions which are rightcontionuous and have left limits. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). 21 D [0.oi denotes the column vector with the xi as components Special notation for risk processes /3 the arrival intensity (when the arrival process is Poisson). the ith entry is 1 and all other 0. Then the assumption of Dpaths just means that we use the convention that the value at each jump epoch is the right limit rather than the left limit. only have finitely many jumps in each finite interval. . A. a. 7r. row vectors have lowercase Greek letters like a. though slightly more complicated.e. . Usually. 0 marks the end of a proof.
cf. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. ry The adjustment coefficient. Notation like BE and B(t) in Chapter VI has a similar. I. FL. e.5. 'q the safety loading . interpretation. J the rate parameter of B for the exponential case B(x) = eby. or quantities with a similar time average interpretation.1. p the net amount /3pB of claims per unit time.5. .g. I.22 CHAPTER L INTRODUCTION B the claim size distribution. 111. cf. cf. VI. though slightly more complicated.1.
5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. When encountered for the first time in connection with the compound Poisson model in Chapter III. Sections 4. The duality results in Section 3 (and. somewhat more advanced than in the rest of the book. in part.Chapter II Some general tools and results The present chapter collects and surveys some topics which repeatedly show up in the study of ruin probabilities. strictly speaking. The topic is. More precisely. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. The general theory is. 5) are. however. in particular at a first reading of the book. Due to the generality of the theory. Sections 4. not crucial for the rest of the book. however. however. used in Chapter VI on risk processes in a Markovian (or periodic) environment. a parallel selfcontained treatment is given of the facts needed there. 23 . The likelihood ratio approach in Section 2 is basic for most of the models under study. All results are proved elsewhere . in most cases via likelihood ratio arguments. the level of the exposition is. The reader should therefore observe that it is possible to skip most of the chapter.
f1 .QµB < 1. We get 1 = Ee7So = E e'Y S(. T(u) > T] . Let e(u) = ST(u) . T(u) < oo] + 0 = eryuE [e7Vu). (b) St a$ oo on {T(u) = oo}.1 is basic for the study of the compound Poisson model in Chapter III. V) (u. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). Proposition 1. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem .)AT = E [e7ST(°).24 CHAPTER II.0. (1. {e'YS° }t>0 is a martingale. and the ruin probabilities are ip(u) = P (T(u) < oo).1 Assume that (a) for some ry > 0.2 Consider the compound Poisson model with Poisson arrival rate .T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. using r(u) A T invokes no problems because r(u) A T is bounded by T). the time to ruin r(u) is inf It > 0 : St > u}.2) As T > oo. Example 1 .2) takes the form 1 = E [e'ys(). T) = P(T(u) < T). Then e7u (u) = E[e74(u)j7(u) < oo] Proof We shall use optional stopping at time r(u)AT (we cannot use the stopping time T(u) directly because P(T(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem.(u). and in the limit (1. 1 Martingales We consider the claim surplus process {St} of a general risk process. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. claim size distribution B and p = . The more general Theorem 6.5 can be skipped.u denote the overshoot. As usual. Thus N. T(u) < T] + E [eryST . however. StUit..
Example 1 .1.Q(B[a] . Since {St} has stationary independent increments. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx . the martingale property now follows just as in Example 1. of the normal distribution.f. u Corollary 1.. u Corollary 1.i. O(u ) < e7".Q and the U.u + x is again just exponential with rate S. and (b) follows from p < 1 and the law of large numbers (see Proposition III.1) . the conditions of Proposition 1.Ft = a(S" : v < t).1. From this it is readily seen (see III. condition (a) of Proposition 1. with common distribution B (and independent of {Nt}). Thus.a. Thus.2. Eeas° = e"(°) where n(a) = a2a2/2 . 1. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.1 are satisfied. Under the conditions of Proposi Proof Just note that C(u) > 0. the ruin probability is O(u) = pe.3/6 < 1.1) . Since {St} has stationary independent increments. the conditional distribution of the overshoot e(u) = U . it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .2 and drift p > 0./3.ap. Now at the time r(u) of ruin {St} upcrosses level u by making a jump . are i.Q.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists.(„)_ = x is that of a claim size U given U > u .a it is immediately seen that y = S .g.1.3 Assume that {Rt} is Brownian motion with variance constant o. Proof Since c(a) = /3 (B[a] .4 (LUNDBERG ' S INEQUALITY ) tion 1 .2(c)).a = a . and thus Ee'rs° = 1. B(x) _ edx.1 is satisfied. and p =.d. and thus by the memoryless property of the exponential distribution . By standard formulas for the m.1) shows that Eels. The available information on this jump is that the distribution given r(u) = t and S.= e"(') where K(a) = . A simple calculation (see Proposition III.5 For the compound Poisson model with B exponential. MARTINGALES 25 where {Nt} is a Poisson process with rate .r" where y = S .x. and thus Ee7s° = 1. Then {St } is Brownian motion with variance constant o2 and drift p < 0.
then z/'(u) = e7" where 'y = 21A/a2. But if a $ ^ .Ft}too and the Borel afield F.. A]. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. A somewhat similar u argument gives singularity when B $ B. (2. then S and S are disjoint .v. However.1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. oo). [172]. Grandell [171]. the parameters of the two processes can be reconstructed from a single infinite path. .6 N S = { lim Nt I t +00 t gJ are both in F. Embrechts.3 below). we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. Example 2 . and by the law of large numbers for the Poisson process .6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. 2 Likelihood ratios and change of measure We consider stochastic processes {Xt} with a Polish state space E and paths in the Skorohod space DE = DE[0. P are then singular (concentrated on two disjoint measurable sets). Grandell & Schmidli [131]. P on (DE. Thus the sets S = I tlim +oot t =. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. as shown by the following example this setup is too restrictive: typically'. cf.e.F). P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3.1 Let F. Two such processes may be represented by probability measures F. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. Delbaen & Haezendonck [103] and Schmidli [320]. I. The number Nt F) of jumps > e before time t is a (measurable) r. Theorem 2. A E Ft. and F. F). and is further exploited in his book [157]. More recent references are Dassios & Embrechts [98]. on (DE. hence so is Nt = limfyo N2`i. B. 0 and claim size distributions B. F(S) = P(S) = 1. which we equip with the natural filtration {. Proof Just note that ^(u) = 0 by continuity of Brownian motion.
define P by Pt (A) = E[Lt. Ft).e. A]. A E Ft .2 Let {Ft}t>o be the natural filtration on DE.F8] = L8 and the martingale property. implies that E[LtI.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo.2. A E F. P) such that LLt = 1. G C {T < oo}.r.Y) such that P(A) = Pt(A). LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i.1) and nonnegativity by letting A = {Lt < 0}. Then Ft (A) = E[Lt. A].F). A E F8. Lt < 0] can only be nonnegative if P(A) = 0.t. (i) If {Lt}t> o is a nonnegative martingale w. if for some probability measure P and some {. 1 J (2. ({.r. then { 1 P(G) = EG . using the martingale property in the fourth step. u The following likelihood ratio identity (typically with r being the time r(u) to ruin) is a fundamental tool throughout the book: Theorem 2 . P be as in Proposition 2.Pt}adapted process {Lt}t>o (2. Proof Under the assumptions of (i). The truth of this for all A E Y. then {Lt} is a nonnegative martingale w. (ii) Conversely.1) holds. the restriction of P to (DE.r. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. _.t. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G).t. that the restriction of P to (DE. ({Ft} .F such that (2.Pt)) The following result gives the connection to martingales. Proposition 2.Ft}. ELt = 1 follows by taking A = DE in (2. Finally.Tt) is absolutely continuous w..2(i).3 Let {Lt}.1) holds. we have E [ LTIFT]1 = LT on {T < T}. G ] = E [LT . Hence E [_ . Conversely. If r is a stopping time and G E PT. This proves (i). By the martingale property. F the Borel o•field and P a given probability measure on (DE. Lets < t. . F) such that ELt = 1. . A] = E[Lt. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. Then P(A) = E[Lt. then there exists a unique probability measure Pon .
r.1: Corollary 2. (2. we need the concept of a multiplicative functional. Lr(u) 11 The advantage of (2. where {Rt} is the risk reserve process. 1 Since everything is non negative. The crucial step is to obtain information on the process evolving according to F.4) Proof Letting G = {r(u) < oo}. Xt = St. Consider a (timehomogeneous) Markov process {Xt} with state space E.28 CHAPTER II. the natural filtration {. In the context of ruin probabilities. St). we have F(G) = V )(u).1). Rt) or Xt = (Jt. say.4) compared to (1. applying (2. each F.t. we assume for simplicity that {Xt} has Dpaths. and letting T * oo. For the definition.1. T(u) < oo]. of (2. A change of measure is performed by finding a process {Lt} which is a martingale w. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1.r. Now just rewrite the r.Gn {r <T} . u From Theorem 2. To this end. SOME GENERAL TOOLS AND RESULTS In the general case .Ft} .Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. (2. 5) for processes with some randomwalklike structure. in continuous time (the discrete time case is parallel but slightly simpler).1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). Xt = (Jt. {St} = {u .3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1.4 Under condition (a) of Proposition 1. First we ask when the Markov property is preserved. one would typically have Xt = Rt. and this problem will now be studied.t.2) follows by monotone convergence. first in the Markov case and next (Sections 4. is nonnegative and has Ey Lt = 1 for all x.2) by noting that 1 = ersr(„) = e1'ue7Ou).. is Markov w.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 .h. t.s..O(u) = eryuE[e'YC(u). both sides are increasing in T.1) in Proposition 1. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. .
. Similarly.v.6) for any .'s of the form fl' f.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any . t] * [0. o 9tI. o 9t = V. where Ot is the shift operator. ({Xt+u} 0<u<8) Theorem 2. Zt. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].Ft }. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].v. for all x.YB] for any Ftmeasurable r. 0 .Ft]. A]. the natural filtration {Ft} on DE. and then L. The converse follows since the class of r. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.Ft] = Ex. oo).7).Pt+8measurable r. s.v.[Y. nonnegative and Lt+8 = Lt•(Lso9t) (2.s.5) Pxa. Indeed. t. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.2.8) which is the same as (2. Vt+e.(Xtitl) with all t(i) < t + s.r.Y8f t < s.'s of the form Zt • (Y8 o 0t) comprises all r. (2.v.(A) = Ex [Lt. Proof Since both sides of (2.6) implies (2. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.Ftmeasurable. t and let Px be the probability measure given by t. (2. since Zt • (Y8 o Ot ) is .7). Y8.v.. or. (2. The precise meaning of this is the following: being . By definition of Px. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2.5) are Tt+e measurable. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.t.F8measurable r.5 Let {Xt} be Markov w.7) for any Ftmeasurable Zt and any . the Markov property can be written E.Ft+8measurable. (2.T9measurable Y8.
. with a proof somewhat different from the present one. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process.At where At = k: vk <t U. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. t.. We work on a finite time interval [0. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. 3 Duality with other applied probability models In this section. the arrival epochs are Qi. .e. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. < aN < T. reflection at zero and initiar condition Vo = 0. More precisely .. 0 < vl < . } E[Lt+B I.. Ro = u (say).. R = p(R)). The result is a sample path relation. Thus R = Ro + f p(R8) ds .Ft] = LtE[L8 o 9t I. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. .. Indeed. The formulation has applications to virtually all the risk models studied in this book. then Remark 2. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. (using the Markov property in the second step) so that the martingale property is automatic. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. aN where or* = T UN_k+l. In between jumps. UN.6 For {u . {Vt} . The storage process {Vt }o<t<T is essentially defined by time reversion... t] = LtExt L8 = Lt. see Dynkin [128] and Kunita [239]. In between jumps. The corresponding claim sizes are Ul..5 can be found in Kuchler & Sorensen [240]... (3. . CN. A more elementary version along the lines of Theorem 2.1) The initial condition is arbitrary.30 CHAPTER H. the premium rate is p(r) when the reserve is r (i. ..
(3._.. Then rt°) > rt°) for all t when u > v. {Vt} remains at 0 until the next arrival). V)(u... 3. The sample path relation between these two processes is illustrated in Fig.e. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i. In particular.1) we have Vt = At  f P(Vs)ds where A= U= AT ..__. (3. That is..x..___ .1 The events {T(u) < T} and {VT > u} coincide._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3.____•_.. Theorem 3. V = p(V)). instead of (3.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u.1. :.. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous..__.T) = P(VT > u).11 4.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability. .3.AT_t.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u.
Hence if n satisfies VVN_n+1 = 0 (such a n exists. Theorem 3. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . and since ruin can only occur at the times of claims. Resnick & Tweedie [79]. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context.3). = r(VT) . with distribution B. 3. and so on. say of water in a dam though other interpretations like the amount of goods stored are also possible.1 with Ro = u = u2).3 and being i. this represents a model for storage. say V. one may feel that the interaction between the different areas has been surprisingly limited even up to today. Proof Let T ^ oo in (3. 3. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. the distinction between right. Corollary 3. see Siegmund [344]. We get: Corollary 3.2 from Harrison & Resnick [188]. if nothing else n = N).32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. we can repeat the argument and get VoN_1 > Ra2 and so on. If VaN > 0. Then Vo.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). The results can be viewed as special cases of Siegmund duality. and then '0 (u) = P(V > u).d. and a general premium rule p(r) when the reserve is r. Hence RQ„ > 0 for all n < N.Ul < roil  Ul = RQ„ Va1V_1 < RQ2. . Then similarly VVN = r0. Historically.1 with Ro = u = ul). Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo.2 Consider the compound Poisson risk model with a general premium rule p(r). Nevertheless. if and only if O(u) < 1 for all u. Then the storage process {Vt} has a proper limit in distribution. The arrival epochs correspond to rainfalls. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].i.T l .U1 > roil .1 and its proof is from Asmussen & Schock Petersen [50]. we have RQ„ < 0 so that indeed r(u) < T.U1 = Rol. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. Historically. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. we have r(u) > T.
. the Lindley process Wo. Then the events {r(u) < N} and {WN > u} coincide.YNn+1) n=0.. ...d. Z2 . Theorem 4.. there is an analogue of Theorem 3.. . 0). i.h. just verify that the r .. generated by Z1.1) Thus {Wn}n=o... as long as the random walk only takes nonnegative values... W1.. In particular..2) satisfies the same recursion as in (4. and is reset to 0 once the r...1))..1.. W1.1. . (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo.. .... Z2 = YN_1 i . . ...e. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN. I. .1. Z2. R valued sequence Z1. .. N min (Y1 + • • • + YNn) n=0..N From this the result immediately follows..w.2). Xo = 0.Y1 according to Wo = 0. = Xo + Y1 + • • • + Y.. Proof By (4..min n=0. can be viewed as the reflected version of the random walk with increments Z1. (c) The Lindley process {WN} generated by Zl = Y1.. RANDOM WALKS IN DISCRETE OR CONTINUOUS TIME 33 4 Random walks in discrete or continuous time A random walk in discrete time is defined as X..Yl min (YN . ZN = .. WN = YN .1. Most often.N (4. .1. evolves as a random walk with increments Z1i Z2. ..i.. n=0..1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}.. For discrete time random walks ... hits (oo. For a given i.. WN be the Lindley process generated by Z1 = YN. {Wn}n=0. with common distribution F (say). is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4.. of (4. Here F is a general probability distribution on R (the special case of F being concentrated on {1. 0 Corollary 4.d.4.s.1.2) (for a rigorous proof... has a proper limit W in distribution as n + oo.... Z2..1 in terms of Lindley processes .. Z2 = Y2.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0.. 1} is often referred to as simple random walk or Bernoulli random walk). where the Yi are i . W2. N min (Y1 + + Yn).. Let further N be fixed and let Wo...
1. For a random walk.s.s .=o.. equivalently.1 have the same distribution for n = 0.. 0 By the law of large numbers. the Y1.. .g.1 is actually not necessary . (Yi + • • • + Yn) > oo a.34 CHAPTER II. ZN or. . W v m and P(W > u) = P (m > u) = 0(u).N so that WN _P4 M = supra=0.. Proof Since (YN. (e).l. F has a strictly positive density and the Px corresponds to a Markov chain such that the density of X1 given Xo = x is also strictly positive.1.g.. Next consider change of measure via likelihood ratios.. Clearly... (d) #. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists.) sup n=0.the result is a sample path relation as is Theorem 3.2 and Theorem 4.. SOME GENERAL TOOLS AND RESULTS (d) m = inf. In that case .. N. YN in Theorem 4. a sufficient condition for (e) is that EY is welldefined and > 0. One then assumes Yn to be a stationary sequence. Combining these facts gives easily the equivalence of (a)(d).3 The i. . a Markovian change of measure as in Theorem 2. .. Remark 4 . 176) but appears to be rather intractable.ooa. + Z.5 does not necessarily lead to a random walk: if.1.. there is a more general version of Corollary 4.l..1 is equivalent to WN D MN = (Z1 + . The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. . ±1. or M < oo a. In general.. then the restrictions of Fx. By Kolmogorov's 01 law.) and defines Zn = Yn.. .s.d.1.o.. .. either M = oo a... . (e)Yi+•••+Yn 74 ..2. .. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). w. e. Similarly. assumption on the Z1..i.. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . .s.. the Lindley processes in Corollary 4. . doubly u infinite (n'= 0. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.. Thus the assertion of Theorem 4. YN). ±2. Y1) has the same distribution as (Y1..
this means E(g(x. cf. Y ) f (Y)] = E[g(O.g.. the random walk property implies Ex f (Y1) = Eo f (Y1 ). (a) = log F(a] is the c. For n = 2. for some function h with Eh(Y) = 1. We get: Corollary 4. the changed increment distribution is F(x ) = E[h(Y).3) holds. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4.. 100 ). of F).Y2) = h(1'i)h(I'a). . Y) = h(Y ) a....4. Y) f (Y)] for all f and x.f.4 Let {Ln} be a multiplicative functional of a random walk with E_L.nrc(a )} (4.4) ({Ln} is the familiar Wald martingale . Then the change of measure in Theorem 2. e..5 corresponds to a new random walk if and only if Ln = h(Y1) . = 1 for all n and x.4.g.s. then n n Ex [f f = Ex H fi a( YY) i=1 i_1 ( Y=) h(YY) H Ef=(Y=)h(Y=) = II J fi(Y )P( d) from which the random walk property is immediate with the asserted form of F. (4. Then the change of measure in Theorem 2.4). h(Yn) (4. Breiman [78] p. In that case. y ). Y1). implying g (x. and define Ln by (4.3) holds for n = 1. and so onforn =3.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) .3) 1Pxa. where h (y) = g(0. Conversely. In particular. u A particular important example is exponential change of measure (h(y) = e°y'(") where r. Y < x].5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite. Since L1 has the form g (Xo. Proof If (4. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1.s.
or imbedded into continuous time processes . A general jump process can be thought of as limit of compound Poisson processes with drift by considering a sequence v(n) of bounded measures with v(n) T v. In continuous time. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o).. however.5) Note that the structure of such a process admits a complete description. the tradition in the area is to use continuous time models. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . Roughly.. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time .6) More precisely. e f x:IxJ>e} v(dx) < oo (4. the claim surplus process for the compound Poisson risk model . . we are .1).t(i . Xt =Xo+pt+oBt+Mt. v2 = 0 and v = 3B).. In discrete time. a positive measure on R with the properties e J x2v(dx) < oo. i. with premium rate p.Xt)I. < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) ..). see Chapter V. they arise as models for the reserve or claim surplus at a discrete sequence of instants. (4. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). . corresponds to a process with stationary independent increments and u = p. this description needs some amendments. (4. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo.7) for all e > 0. but we omit the details ). a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. Xt(2)_t(l).e. In risk theory. However.36 CHAPTER II. {Xt} is a random walk.Ft] = Eof (X. say the beginning of each month or year . say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). The simplest case is 3 = JhvMM < oo. {Xt} can be written as the independent sum of a pure drift {pt}. the pure jump process is given by its Levy measure v(dx). Xt (n)t(n1) being independent whenever t(O) < t(1 ) < .Xn < x). given by a the increment distribution F(x) = P(Xn+l . which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular .
i.min Xt (4. Proposition 4. Now consider reflected versions of processes with stationary independent increments. then Ee'(xtxo) = Eoeaxt = etx(a). oo]. virtual waiting time refers to Vt being the amount of time a customer would have to wait before starting service if he arrived at time t (this interpretation requires FIFO = First In First Out queueing discipline: the customers are served in the order of arrival). where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f".4. RANDOM WALKS IN DISCRETE OR CONTINUOUS TIME 37 almost solely concerned with the compound Poisson case and shall therefore not treat the intricacies of unbounded Levy measures in detail. Corollary 4. Furthermore. has upwards jumps governed by B at the epochs of a Poisson process with rate . is easily seen to be f3pB < 1. cf. First assume in the setting of Section 3 that {Rt} is the risk reserve process for the compound Poisson risk model with constant premium rate p(r) = 1.s.1)v(dx) (4. Then the storage process {Vt} has constant release rate 1.10) . O(u.Q and distribution B of the service times of the arriving customers. V E [0.6 In the compound Poisson risk model with constant premium rate p(r) . T) = P(VT > u).2).8) O<t<T (assuming Wo = Xo = 0 for simplicity). VT + V for some r.6).3 and decreases linearly at rate 1 in between jumps. Here workload refers to the fact that we can interpret Vt as the amount of time the server will have to work until the system is empty provided no new customers arrive. [The condition for V < oo a.7 If {Xt} has stationary independent increments as in (4.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. Chapter III. (ex . where VT is the virtual waiting time at time T in an initially empty M/G/1 queue with the same arrival rate /3 and the service times having the same distribution B as the claims in the risk process.v. WT = XT . and b(u) = P(V > u). and the reflected version is then defined by means of the abstract reflection operator as in (4.e.1. having Poisson arrivals with rate . defined as a system with a single server working at a unit rate. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. jxJ v(dx) < oo.
e.6) are µ = µ + Oo2 .11) (eax . Eea(µt + QBt) = et{aµ +a2OZ/2}. X8) = Eof (X8)L8 = Eof (X8)• For the second.38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. By explicit calculation .f.Xt)I'Ftl = E [f(Xt+B . 5 has stationary independent increments as well. Theorem 4 . let e" (a ) = Eoeaxl. Chung [86]).g. t. use the representation as limit of compound Poisson processes.10) is the LevyKhinchine representation of the c. This is of course no coincidence since the distribution of Xl . .4 o) aµ + ((a + 0 ) 2  0 2 )o 2 /2+ r w J 00 (e (a + 9)x  a 9x )v(dx) 00 a(µ + O 2) + a2a2 / 2 + J (eax . then the changed parameters in the representation (4. of an infinitely divisible distribution (see.1 that E eaMt = exp fmoo In the general case . if Lt = e9(xt .g.5) and get E [f(Xt+B .xo)tk ( e). Then l e" (a) = Eo [ Li ea "] = eK (9)Eo {e ( a+9)x1 J I = er(a+o )K(B) R(a) = K(a + 0) . 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.Xt)I Ftl = Eof (X8)g(s.1 ) v(dx) .Xt)g(s. we show in the compound Poisson case ( IlvIl < oo) in Proposition III.1)eexv(dx). Xt +B . In particular. v(dx) = e9xv (dx).1. (4. u Note that (4. Xt Xo) with E2Lt = 1 for all x.Xt)L8 o 0tIFt] = E [f (Xt+s . Proof For the first statement ..Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. Q2 = v2. we use the characterization (4. Then the Markov process given by Theorem 2.
0 in the following.. the structure of MAP's is completely understood when E is finite: 2and only there . (5.Ft] = Ejt.. one reason is that in parts of the applied probability literature.1) For shorthand .8. where . Ei. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate ... Example 4 .9 If X0 = 0. 0.g.3 and claim size distribution B # B. Thus (since µ = p = 1. the corresponding claim sizes .0.3 =.(3B(dx). dB/dB = b/b when B.11 For an example of a likelihood ratio not covered by Theorem 4. Example 4 . abbreviated as MAP in this section2. MAP stands for the Markovian arrival process discussed below. u 5 Markov additive processes A Markov additive processes. B have densities b.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .3 and claim size distribution B. As for processes with stationary independent increments .2. <t whenever the RadonNikodym derivative dB/dB exists (e. U2. is defined as a bivariate Markov process {Xt} = {(Jt. Recalling that U1. b with b(x) > 0 for all x such that b(x) > 0). and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. are the arrival times and U1.(3 = . St)} where {Jt} is a Markov process with state space E (say) and the increments of {St} are governed by {Jt} in the sense that E [f (St+8 . then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4.3B[B]. MARKOV ADDITIVE PROCESSES 39 Remark 4. B(dx) = B[9] B(dx). we write Pi. Then we can write v(dx) _ /3eOxB(dx) = / (dx). let the given Markov process (specified by the Px) be the claim surplus process of a compound Poisson risk process with Poisson rate 0 and claim size distribution B.5.4).. corresponding to p = 1. . v(dx) _ . it is then easily seen that Lt = H dB(Ui) i:o.o[f (S8)g(J8)]. a = 0. .St)g(Jt+s)I. Ei instead of P2.l3 and claim u size distribution B.
Then a Markov additive process can be defined by letting t St = lim 1 I(IJB1 < e)ds E1o 2d o be the local time at 0 up to time t. a MAP is the same as a semiMarkov or Markov renewal process. which we omit and refer to Neveu [272] or cinlar [87]. oo). .[a) = (Ei[easl.o(Ji = j. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. by generating Yn according to Hij when J„_1 = i. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1. If all Fij are concentrated on (0. 1 J1 ='^])iJEE = (Fij[a])i . t+s) where Jt . In continuous time (assuming Dpaths)... In addition.J1=j)= Fij (dx) Pij In simulation language. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time. Proposition 5.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. Fn[a] = F[a]n where P[a] = P . As a generalization of the m..it = A. {Jt} is specified by its intensity matrix A = (Aij)i. let {Jt} be standard Brownian motion on the line. Jn = j. {St} evolves like a process with stationary independent increments and the parameters pi.9 EE = (iii&ij[a])i j EE . with the Y„ being interpreted as interarrival times. the distribution of which has some distribution Bij. v. As an example.Sr_1. consider the matrix Ft [a] with ijth element least Ei . (That a process with this description is a MAP is obvious. An alternative description is in terms of the transition matrix P = (piA.g. SOME GENERAL TOOLS AND RESULTS In discrete time.) If E is infinite a MAP may be much more complicated. vi(dx) in (4.6) depending on i.f.jEE• On an interval [t. Y1 E dx) where Y„ = S„ .40 CHAPTER H.1 For a MAP in discrete time and with E finite.i... the converse requires a proof. Y2.
Jt = k] { xk kEE j la] . 013 . Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . u Proposition 5. In matrix formulation .qkj + k?^j qkj Bkj [a] } = Ei [east.5. Jt = k] { 1 .1) } (recall that qjj = 0). J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. vi(dx). this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I.2 Let E be finite and consider a continuous time Markov additive process with parameters A. where K[a] = A+ (r.ijgij(Bij[a] .1 )v(dx). \ diag Ft[a] = Ft[a]K. MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn.1)) . aSt h = (1 + Ajjh) Ei [east . qij. up to o( h) terms. Sn ) yields Ei[easn+ '. kEE Jn = k]Ek[e"Y" . Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi .4c). assume that the Markov chain/process {Jt} is ergodic. a= . vi(dx) (i E E). Jt = j] Ejesh'^ + E Ak j hEi [ease . Then. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. u In the following. pi. 00 r(i) (a) = api + a2ot /2 + f (e° . Then the matrix Pt[a] with ijth element Ei [east. Jn+1 = A] = 41 Ei[ e 5„. Jt = j] is given by etK[a]. j E E) and So = 0.(')(a)) diag + (). Bij (i. By PerronFrobenius theory (see A. we infer that in the discrete time case the .
Proposition 5. Jeast. SOME GENERAL TOOLS AND RESULTS matrix F[a] has a real eigenvalue ic(a) with maximal absolute value and that in the continuous time case K[a] has a real eigenvalue K(a) with maximal real part.c(a) (and h(")). h(") may be chosen with strictly positive components. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . Yrh(a ) = 1. Since v(").Eikjt = ttc'(0) + ki . Eie"sth^a) = e'Pt[a]h( a) = e. its derivatives are 'asymptotic cumulants'. Furthermore.2) where 7r = v(°) is the stationary distribution. Corollary 5.3 Ei [east. The function ic(a) plays in many respects the same role as the cumulant g. The corresponding left and right eigenvectors v("). Proof For the first assertion.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . Corollary 5. u Let k(a) denote the derivative of h() w.5 EiSt = tK'(0) + ki . cf. cf.t.Jt+v = easttK( a)E [ee (st+vst)vK(a)h(a) jt+v I ^tJ = easttt(a)EJt (easesvK(a )h^a)1 = easttK(a)h^a).r.4. In particular.7. and write k = k(°). of a random walk. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). Jt = j] .42 CHAPTER II. a.f. Proof By PerronFrobenius theory (see A. we are free to impose two normalizations. Then h(°) = e.4c). .h(a)vva)etw(a).4 Eie"sth(a) = h=a)et?("). We also get an analogue of the Wald martingale for random walks: Proposition 5. h(") are only given up to a constants. Corollary 5. and we shall take V(a)h(a) = 1. as will be seen from the following results.e=e°tk.tK(a)h(a) J jj it L o is a martingale. (5.
(5. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5..6 For any stopping time T with finite mean. [E. (5. In the same way. Squaring in Corollary 5. tam E tSt a (0).5 yields + W (a)k.Eikjr . MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5.+ k.e.7 No matter the initial distribution v of Jo. More precisely. subtraction yields Vary St = tic"(0) + O(1).4) . ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . one obtains a generalization of Wald's identity EST = Er • ES. for a random walk: Corollary 5.5.") }) . u The argument is slightly heuristic (e. the distribution of Jo).3) to get Ej [St a " st h i(a ) + 2Ste"st k(a) + e"st k^a) J etI(a) (kia )' + ttc (a)ki") + t {ic"(a)h.2ttc (0)Evkjt + 0(i). Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). Corollary 5.g.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. E=ST = tc'(0)E7. For the second .. . summing and letting a = 0 yields E„ [St + 2Stkj. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x). 8 Also for E being infinite (possibly uncountable ). we differentiate (5.5. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . t a oo. Remark 5 .St]2 = t2/c'(0 ) 2 + 2ttc (0)vk .a) + ttc (a)2hia ) Multiplying by v=.4. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.
gha(i. xEE . let ha(i.5 defines a new MAP. s) = ea8h(i). Remark 5. however.6) We shall not exploit this approach systematically.10 Let {(Jt. Usually.e. Then {Lt } is a multiplicative functional. From (5. 0) = n(a) h(i).5) is a martingale . (5. this leads to h(i) + tcha( i.f (x) tyo t provided the limit exists. St)} be a MAP and let 0 be such that h(Jt) OStt.4 that { h(Jt) easttK(a) L o (5.3b and Remark VI. 0 Proposition 5. 1) (i. h(Jo) Lo is a Px martingale for each x E E.1) one then ( at least heuristically) obtains lim Ex eaSv v a) K( v+oo nEx easttK(a)EJt east(vt)K(a) u[J = Ex easttk(a)h(Jt) It then follows as in the proof of Proposition 5.for the present purposes. u forsEE). Given a function h on E..s. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i). in particular that f is bounded. some extra conditions are imposed. this is.(9) {Lt}t>o = .. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. see. however. In view of this discussion . First.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. and the family {f LEE given by Theorem 2. 0) = h(i )( 1 + ttc(a)). For t small . we take the martingale property as our basic condition below (though this is automatic in the finite case).6.5.9 The condition that (5.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. V. An example beyond the finite case occurs for periodic risk processes in VI. where {Jt} is deterministic period motion on E = [0. Jt = (s+t) mod 1 P8a.6. St) } as follows. G is defined as Gf (x) = lim Exf (Xt) .
11 Consider the irreducible case with E finite. 1 + q(b . In the infinite case . 0<b<oo. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0]. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx).5. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. Bi.1) . one can directly verify that (5.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. In particular. 0 < qij < 1 and Bij [0] > 0. qij = r.11 below in the finite case. vi (dx) = f3 Bi(dx) with . ^i = of qij Bij [0] 1 + qij ( Bij [0] .7(dx) Bij [0] Bij(dx) in the continuous time case . Then the MAP in Proposition 5.tc(0)e = 0 .ic(0)e = ic(0)Oh e) h(e) .10 is given by P = eK(e) Oh e) F[e]Oh(').. Bi(dx) = Bi(dx). (5.1) eft ea' f ij (dx) = Hij (dx) Hij [0] . Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. Bi [0] Remark 5.c(0)e = tc(0)e . That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) . Ai = µi + 0Q.1) holds for the P.Qi < oo and Bi a probability measure.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. u Theorem 5. 0<q<1.7) In particular. . We omit the details. if vi(dx) is compound Poisson.(0)j.. That 0 < qij < 1 follows from the inequality qb <1. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . in the discrete time case.
46 CHAPTER II.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] . SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. are probability measures . v. . Jl = j] :(Yi E dx. F:j with a density proportional to eei . First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . it follows that indeed the normalizing constant is H1 [0].8) h(.r.. In matrix notation .e) Consider first the discrete time case . Jt = j] = hie) . (dx). .tc (') (0) corresponds to the stated parameters µ. Now we can write K[a] =A+A ) ( K[a + 0] . Yi E dx..t. Similarly.13) for matrixexponentials . Jt = A.11.Bay [0]) That k(') (a + 0) .tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . H1.8). Here the stated formula for P follows immediately by letting t = 1.8. (dx) of a process with stationary independent increments follows from Theorem 4. this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5.tc(0)I. This shows that F. since Hij. this implies k[a] = A 1 ) (K[a + 0] . Further Fib (dx ) = P=(YI E dx. Hence the same is true for H=j and H. in continuous time ( 5.. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. is absolutely continuous w. Letting a = 0 yields the stated expression for A.8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. v= . Jl = j) = Ei[Lt. a = 0 in (5.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) .
6. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. [261]. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S. oo). [226] and Miller [260]. . see also Fuh & Lai [149] and Moustakides [264].1) = Aij4ij(Bij[a] . Conditions for analogues of Corollary 5. and is typically defective.Bij[0]) = hjel)ijgijBij[0](Bij[a] . h. The closest reference on exponential families of random walks on a Markov chain we know of within the more statistical oriented literature is Hoglund [203]. i. there is. which.1).e. however. has no mass on (oo. however.6. 7+ < oo). Though the literature on MAP's is extensive. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. is slightly less general than the present setting. Note that G+ is concentrated on (0. [225]. the literature on the continuous time case tends more to deal with special cases.3 for an infinite E are given by Ney & Nummelin [266].7). IIG+II = G+(oo) = P(T+ < oo) = 0(0) < 1 when 77 > 0 (there is positive probability that {St} will never come above level 0). For the Wald identity in Corollary 5. hardly a single comprehensive treatment.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . 0]..)Ajjgij(Bij[a+0] . [262] in discrete time. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. < x) = 11 (S.. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265].+ < x.
at present we concentrate on the first ladder height. 0 f T+ (6. g(y)R+(dy) = E f g(St)dt.d.1. we shall first consider the compound Poisson model in the notation of Example 1. the ladder heights are i.2.00 ). 1 For the compound Poisson model with p = 01LB < 1..i. In any case. i. there are only finitely many). On Fig. Here bo(x) _ B(x)/µB.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. which gives an explicit expression for G+ in a very general setting. it follows that for g > 0 measurable. Recall that B(x) = 1 . Theorem 6 .(3B(x ) = pbo(x) on (0.T+ > t)dt = E f 0T+I(St E A) dt. In simple cases like the compound Poisson model. For the proof of Theorem 6. see Fig.. To illustrate the ideas. and the maximum M is the total height of the ladder. G+ is given by the defective density g + (x) =. R+ is concentrated on (oo.B(x) denotes the tail of B. oo). the dependence structure seems too complicated to be useful). where basically only stationarity is assumed. Thus. a fact which turns out to be extremely useful. the sum of all the ladder steps (if rl > 0.5 below.48 CHAPTER K. 6. they have a semiMarkov structure (but in complete generality. The first ladder step is precisely ST+. Also.1. o 00 (6. = ST+(1) Figure 6. In other cases like the Markovian environment model. by approximation with step functions .1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. The main result of this section is Theorem 6. i.1.ST+(1) and so on.2) . SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A.e. 6.e. has no mass on ( 0. 0]. the second ladder height (step) is ST+(2) .
P(STEA. That is. .ST<St. has the same distribution as {St}o<t<T.0<t<T) = F(ST E A.St<0.2(b): r+ < t Thus. 0 < t < T) P(STEA. 49 Proof Let T be fixed and define St = ST .6.ST_t. since the distribution of the Poisson process is invariant under time reversion.O<t<T) = P(STEA. THE LADDER HEIGHT DISTRIBUTION Lemma 6 . St S* t a Figure 6. 0]. see Fig.ST<St.T+>T) = P(STEA. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence.O<t<T).2. ST < ST_t. 0 < t < T.2 R+ is the restriction of Lebesgue measure to (00.2(a): T+ > t Figure 6. 6.
. E A} precisely when r+ > t.3 G+ is the restriction of /3R+*B to (0.T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .3 where the bold lines correspond to minimal values. it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time .y) (here we used the fact that the probability of a jump at u t is zero in the second step. G+(A) = Q f 0 B(A . cf.t dT. Figure 6..r. oo). oo). . and since the jump rate is /3. this is just the Lebesgue measure of A.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. SOME GENERAL TOOLS AND RESULTS Integrating w. U + St_ E A. we get G+ (A) = f 00 /3 dt E[B(A . The probability of this given { Su}u<t is B(A .3 Lemma 6 . for A C (0. But since St 4 oo a. s. 6. T+ > t] 0 _ /3 f E[B( A .2) in the last). Fig. and (6.St _).50 CHAPTER II.St).St _)I(r+ > t). That is.
The marked point process . this is equivalent to the risk process {St*} being stationary in the sense of (6. i.Q f r+(x . 4 (the points in the plane are (ak . Fig.z)B(dz) _ f I(x < z)B(dz) _ f (x). the first component representing time (the arrival time o.M o 08 shifted by s is defined the obvious way. as a point process on [0.. this does not depend on h). {St+8 . we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup. obviously.) where ak = Ti + • • • + Tk .* ) and the second the mark (the claim size Uk ). Uk) for those k for which ak . In the stationary case.S8 )t> o = {St }t>o for all s > 0.1 With r+(y) = I(y < 0) the density of R+. 0 Generalizing the setup. The sample path structure is assumed to be as for the compound Poisson case: {St*} is generated from interclaim times Tk and claim sizes Uk according to premium 1 per unit time. oo). 6 ..4). Nt St =>Uk k=1 t where Nt = max{k = O. . The points in the plane (marked by x on Fig. 2.:T1 +•••+Tk <t}. 6 . Uk) (k = 1.. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. . U k)} k=1 a is as a marked point process M *. h]} /h (by stationarity. i.s. we define the arrival rate as E# { k : ak E [0 . Lemma 6..1. assuming basically stationarity in time and space. The first ladder epoch r is defined as inf It > 0 : St > 0} and the corresponding ladder height distribution is * G+ (A) = P(S** E A) = P(ST+ E A.4) are (ak. oo) x (0.3 yields g+ (x) = .s > 0).6.e. cf.e. The traditional representation of the input sequence {(TT.T+ < oo)..
. o.QiBi(dx). the r. h] Eco(M*) = 1 E f co(M o Bt)dt. and let T = T2 denote the first proper interarrival time . Uk) k=1..g.. We represent M by the sequence (Tk. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . h. See. Section 5) which has pure jump structure corresponding to pi = a = 0.e. letting h J.5) does not depend on h. The two fundamental formulas connecting M* and M are Eco(M) = aE E. As above . This more or less gives a proof that indeed (6.. Example 6 . . i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. Sigman [348] for these and further aspects of Palm theory. where TI = 0. most often one takes h = 1). where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. Oh becomes the approximate probability F(ri < h) of an arrival in [0. vi(dx) = . i.5) represents the conditional distribution of M* given vi = 0. = 0 . of (6. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t.2. SOME GENERAL TOOLS AND RESULTS M* U.s. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists.4 Given a stationary marked point process M*. V(M* o eak ). 0. k: vk E [0.4 Consider a finite Markov additive process (cf.52 CHAPTER II. h] and the sum approximately ^o(M*)I(ul < h). e.
THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. we note: Corollary 6./.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 .s. and that p = 0EU0 < 1. the ruin probability . Jt = j is iri(3i /.p. dt A + E Aijgij j#i Thus the arrival rate for M* is 1] it A + E Aijgij iEE i#i Given that an arrival occurs at time t . we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i. If Jt_ = i.6 Under the assumptions of Theorem 6. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution . qij when {Jt} jumps from i to j and have mark distribution Bij.O for i = j and iriAijgij/.p.*(0) with initial reserve u = 0 is p = /3EU0.e. Before giving the proof. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. A stationary marked point process M* is obtained by assigning Jo distribution Tr. After that. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . .. let U0 be a r.oo a. It follows that we can describe the Palm version M as follows . Jo) w. Assume that St * . Then the ladder height distribution G+ is given by the (defective) density g+(x) = .OEU0. 5 Consider a general stationary claim surplus process {St }t>o. Note in particular that the Palm distribution of the mark size (i.O fo "o F(x)dx = . aij for (i.OF(x). and by some additional arrivals which occur w. an arrival for M* occurs before time t + dt w.O for i # j.p. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J. v.5.= i. the probability aij of Jt . let the arrivals and their marks be generated by {Jt} starting from Jo = j. First choose (Jo_.6.
. oo)). CHAPTER H. 0<u<t) = P(St EA. 2. has a very simple interpretation as the average amount of claims received per unit time . 105) shows that one can assume w.Su< 0.Su_ <0.Q_k and has size U_ k. The last property is referred to as insensitivity in the applied probability literature.o. h.Mt). oo) x (0 .s.l.g. 0).0<u<tIAt) = P(St EA. . A standard argument for stationary processes ([78] p.. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0.. { Su}0<u<t is distributed as a process {Su} .5.St*_ u.A.. in (0. that M* and M have doubly infinite time (i.e. .0<u<t) = P(StEA. Proof of Theorem 6.54 By (6. and the kth preceding claim arrives at time t ..St<Su. The sample path relation between { Su } and { Su } amounts to S„ = St .. which makes an upwards jump at time .. the mark at time Qk is denoted by Uk.1] here the r .o<u<t where a claim arrives at time t and has size Uo. are point processes on (oo . 6.).. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt. in (oo. oo) p(t) = P(St EA.. the arrival times 0 < 0'1 < Q2 < . Now conditionally upon At .St <.0<u<tIAt) = P(St EA. It follows that for A C (0. oo ) and the arrival times 0 > 0_1 > a_2 > .Su<0. (k = St}t>o 1. The result is notable by giving an explicit expression for a ruin in great generality and by only depending on the parameters of the model through the arrival rate 0 and the average ( in the Palm sense) claim size EU0. We then represent M by the mark (claim size ) Uo of the arrival at time 0.$St_ u.5.0<u<t) = P(St EA.(left limit) when 0 < it < t and is illustrated on Fig . Let p(t) be the conditional probability that ST+ E A.o. T+ = t given the event At that an arrival at t occurs .5). moves down linearly at a unit rate in between jumps and starts from S0 = U.
0 < u < t } is the event that { Su } has a relative minimum at t .5 where it = { St < Su. cf.. NIt)dt . the left endpoint of the support is oo. G' (A) = 3 f P(St E A. Fig. 6. Thus. the support of L has right endpoint U0. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo.6.s. and we let L(dy) be the random measure L(A) = fo°° I(St E A. A sample path inspection just as in the proof of Lemma 6 .5 where the boxes on the time axis correspond to time intervals where {St } is at a minimum belonging to A and split A into pieces corresponding to segments where {Su} is at a relative minimum. In Fig. and since by assumption St * oo a. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. Mt)dt = i3EL(A) o"o . Since So = U0. Uo]. t a oo.5. 6.
5 is due to Schmidt & coworkers [48]. SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6. A further relevant reference related to Corollary 6. [147].56 CHAPTER II.1).2.6 is Bjork & Grandell [67]. . [263] (a special case of the result appear in Proposition VI.
exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. are i. 4. i. U2. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability.e. and assume that • { Nt}t>o is a Poisson process with rate j3. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi.4 below . St = uRt = EUi t. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . see Chapter IV.Chapter III The compound Poisson model We consider throughout this chapter a risk reserve process {Rt } t>o in the terminology and notation of Chapter I. For finite horizon ruin probabilities . Some possibilities are numerical Laplace transform inversion via Corollary 3. . with common distribution B. Thus . It is worth mentioning that much of the analysis of this chapter can be carried over in a straightforward way to more general Levy processes . Panjer's recursion ( Corollary XI. • the premium rate is p = 1.d.. i. 3).6) and simulation methods ( Chapter X). and independent of {Nt}. A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . say. • the claim sizes U1..
g.t = t(p . Proof It was noted in Chapter I that p .. THE COMPOUND POISSON MODEL Poisson risk process. say stable Levy motion.s. (c) Ee8St = et" (8) where c(s) = f3(B[s] .1 is the expected claim surplus per unit time.)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. we get Ee8st = 00 e8t c` Ee8 (U1+. e . The same method yields also the variance as Nt Ne Nt Var St = Var E Uk = Var E ^ Uk Nt +EVar [ k=1 k=1 1 k=1 Uk Nt Var [Ntµs] + E[NtVar U] = 113µs + t13Var U = tf3pB2).+Uk)P(Nt = k) k=O e8t k=O B[s]k . (b) Var St = t. of the claim surplus St . and that B(k)[0] = Pak). For (c). 0 . Schmidli [319]. (d) The kth cumulant of St is tf3p(k) for k > 2.'s etc.6pBa).f.. Write pB^) = EUn' YB = Pali = EU. {Bt} a Brownian motion and {Rt} a pure jump process.58 CHAPTER III.Rt.1).g.1 (a) ESt = t(13µ$ . See e.1) . m. A more formal proof goes as follows: Nt r Nt ESt = E > U k . [324]. and Schlegel [316]. 1 Introduction For later reference. Dufresne & Gerber [126]. cumulants . We do not spell out in detail such generalizations.t = fltpB . Furrer [150]. we shall start by giving the basic formulas for moments. P = PAB = 1/(1 + rl) Proposition 1.1) = t(p .u . where K(k) (0) is the kth derivative of is at 0. and this immediately yields (a).t = E[Ntµs] . for (d) just note that the kth cumulant of St is tic(k) (0).t = E E [ U k k=1 k=1 Nt .1).
cf. if t = nh + v with 0 < v < h. u + v]. 2. then St> SnhV>Snhh.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo.V. where Tk is the time between the kth and the (k .. In this way.3EU01 = 1µs where rt is the safety loading. Here is one immediate application: Proposition 1. . Obviously.1 is the same as if {St} was a random walk indexed by t = 0. lim supt. then lien inft. (c) If 77 > 0. The right hand inequality in (1.h < St < S(n+1)h + h. (b) If 77 < 0.Tk.1)th claim. we get a discrete time random walk imbedded in the claim surplus process {St}. then Snh ..4. For the proof.3) is proved similarly. The point of view in the present chapter is. S„+V > S„ .d. II. meaning that the increments are stationary and independent.1. Sn+0 . 1. rather to view {St} directly as a random walk in continuous time. so that {Sok } is a random walk with mean EUET = EU. however.. then St 4 co. we need the following lemma: Lemma 1. Proof We first note that for u. which is often used in the literature for obtaining information about {St} and the ruin probabilities. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. Indeed.3 If nh < t < (n + 1)h. In particular. we have Sok . (a) No matter the value of 77. The connections to random walks are in fact fundamental. We return to this approach in Chapter V. obviously 0(u) = F(maxk Sok > u). INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. St = oo. For example. (d) If 17 = 0. and the value is then precisely v. then St 00. the Uk . v > 0.. St = oo.i.Tk are i.1 = .Sok_l = Uk .S„ attains its minimal value when there are no arrivals in (u.
..1(b)) that the assertion holds as t 4 oo through values of the form t = 0. Remark 1 . For any fixed h. and < 1 for all u when 77 > 0.4 The ruin probability 0(u) is 1 for all u when 77 < 0. h A similar argument for lim sup proves (a). Thus using Lemma 1. where the size of the portfolio at time t is M(t). Corollary 1.3.p.. and (b). or by a general result on discrete skeletons ([APQ] p. Snh/n a4' ESh = h(p . {Snh}n=o._.o.2.1). However.60 CHAPTER III. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1.1) as t 4 oo is normal vtwith mean zero and variance )3µsz) Proof Since {St}t>o is a Levy process (a random walk in continuous time). is a discretetime random walk for any h > 0. we get lim inf St t>oo t nroo nh<t<(n+1)h t = lim inf inf St h l++m of Sn 7t h = ESh = p . .. at least once. lim supn_. This contradicts u St400.t .2.. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time.2: Proposition 1. u 307). THE COMPOUND POISSON MODEL Proof of Proposition 1.. {Snh}n=o. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1. Part (d) follows by a (slightly more intricate) general random walk result ([APQ].. Considering the next downcrossing (which occurs w.3. p. 169) stating that lim infra. then {St} upcrosses level 0 a. There is also a central limit version of Proposition 1.s.. hence by induction i. If rl > 0.1.. 1 since St 4 oo) and repeating the argument.5 The limiting distribution of St . h. it is seen that upcrossing occurs at least twice.1. The general case now follows either by another easy application of Lemma 1. if P(M > 0) = 1.1. 2h.. Notes and references All material of the present section is standard. this case can be reduced to the compound Poisson model by an easy operational time transformation u T1(t) where T(s) = )3 fo M(t)dt. Snh u = 00 (the lemma is not needed for (d)). and hence by the strong law of large numbers. 0 Snh = 00. is a discrete time random walk. (c) are immediate consequences of (a). it suffices to prove 4'(0) = F(M > 0) < 1.6 Often it is of interest to consider size fluctuations.
IV. the formula for the distribution of M follows . cf. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. Thus .34 or A. Summing over n.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). nevertheless..1 provides a representation formula for 0(u). THE POLLACZECKKHINCHINE FORMULA 61 2 The PollaczeckKhinchine formula The time to ruin r(u) is defined as in Chapter I as inf It > 0: St > u}.e. d. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . Combined with i/i(u) = P ( M > u). the ladder heights are i. 1 The distribution of M is (1. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . i.1) is not entirely satisfying because of the infinite sum of convolution powers.6. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ).P) E PnBon(u) . p < 1.. equivalently.1) representing the distribution of M as a geometric compound. As a vehicle for computing tIi(u). Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 .6. but we shall be able to extract substantial information from the formula.1. 11. Theorem 2. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. 1e. It is crucial to note that for the compound Poisson model.IIG +II)EG+ . (2. B(x)/aB.2. We assume throughout rl > 0 or. oo ). and we further get information about the joint conditional distribution of the surplus and the deficit.just before ruin is again B0. Fig. The expression for g+ was proved in Theorem 11. 0 Alternatively. This follows simply by noting that the process repeats itself after reaching a relative maximum. Note that this . [APQ] Ch.1. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). n=0 (2. which we henceforth refer to as the PollaczeckKhinchine formula.
62 CHAPTER III. Theorem 2.6. Theorem A1.5. [62]. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. V is uniform on (0. where it requires slightly more calculation. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB.d. However. (a) 11 (ST(o)_ > x.2 and it gives an alternative derivation of the distribution of the deficit ST(o) Notes and references The PollaczeckKhinchine formula is standard in queueing theory.5. In the risk theory literature. and the conditional distribution of ST(o) given ST(o)_ = y is the overshoot distribution B(Y) given by Bov)(z) _ Bo (y + z )/Bo(y). ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. ST(o )) given r (0) < oo is the same as the distribution of (VW. Asmussen & Schmidt [49].2(a) is from Dufresne & Gerber [125].i. Theorem 2 . f +b (b) the joint distribution of (ST( o). . Again. Beekman [61]. 2 The joint distribution of (ST(o )_. Feller [143] or Wolff [384]. there is a general marked point process version. see for example [APQ]. (1 . see Schmidli [323] and references there. cf.V)W) where V. As shown in Theorem 11.1 is traditionally carried out for the imbedded discrete time random walk. We assume rt > 0 throughout. The proof of Theorem 11. in this setting there is no decomposition of M as a sum of i. W are independent. ST(o) > y. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. (c) the marginal distribution of ST(o)_ is Bo .6. cf. cf. (d) the marginal distribution of ST(o)_ is B0. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„).just after ruin. ST(o)) is given by the following four equivalent statements: B(z) dz. THE COMPOUND POISSON MODEL distribution is the same as the limiting joint distribution of the age and excess life in a renewal process governed by B.
B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 .p) = S . 1 .p)pSe a ( l v)x = p( S .O)e(b0)x. Integrating from u to oo. the result follows .1 e ax = n1 (n . The result can.6. Alternatively. SPECIAL CASES OF POLLACZECKKHINCHINE 3a The ruin probability when the initial reserve is zero 63 The case u = 0 is remarkable by giving a formula for V)(u) which depends only on the claim size distribution through its mean: Corollary 3.p) E pn S n x n. the formula for P(O) holds in a more general setting. For a failure at x.0(u) = pe(aA)" Proof The distribution Bo of the ascending ladder height ( given that it is defined ) is the distribution of the overshoot of {St} at time r+ over level 0. however . use Laplace transforms. But claims are exponential . also be seen probabilistically without summing infinite series .p.. 3b Exponential claims Corollary 3.1)1 00 ( 1 .1 0(0) = p = Nl2B = 1 1 +71 Proof Just note that (recall that T+ = r(0)) 00 z/^(0) = I' (r+ < oo) = IIG+II = )3 f(x)dx =l3LB• Notes and references The fact that tp(u) only depends on B through µB is often referred to as an insensitivity property. then. I.2 If B is exponential with rate S. hence without memory. and hence this overshoot has the same distribution as the claims themselves . Thus r(x) = S(1 .p. Let r ( x) be the failure rate of M at x > 0. Thus . As shown in 11.e. 0 . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w.3 so that the conditional distribution of M given M > 0 is exponential with rate S '3 and 0(u) = P(M > u) = P(M > 0)P(M > uIM > 0) = pe(6Mu. a further relevant reference is Bjork & Grandell [67].3.
2) Notes and references Corollary 3.3.+ <u. 0 Proof Write o (u) as P(M>u) = P(S.3 The ruin probability Vi(u) satisfies the defective renewal equation ik (u) = 6+ (u) + G+ * 0(u) = Q f B(y) dy + u 0 f u 0(u . u + oo. II.1. the survival probability Z(u) = 1 .g.2).y)/3B (y) dy.+ = y yields P(M>u.i(u) satisfies the defective renewal equation Z(u) = 1 . A variety of proofs are available .3)). (3. and weights 1/2 for each.S.h.3) below. (3. and conditioning upon S. (b) use stopped martingales .64 CHAPTER III. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7. u .T+ <oo).y)G+(dy) For the last identity in (3. cf.3. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).y)f3 (y) dy. 2 is one of the main classical early results in the area. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype. (3.1) For a heavytailed B.4) can be derived by elementary algebra from (3.S. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .4) zu P(M > u . We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3. The case of (3. just insert the explicit form of G+. T+ <00) (3.+ <U. then 24 1 V. (Example VIII. E.y)G+(dy ) = f U V(u .p + f u Z(u .p + G+ * Z(u) = 1 . is ?7+ ( u). THE COMPOUND POISSON MODEL In VIII.3).s.3) Equivalently. Corollary 3.T+ <oo)+P(M> u. Then the first term on the r.4) is similar (equivalently. (u) 35eu + 35e6u.+ >u. (3.1 p pBo(u).
3 is standard . Of course./3B[s] .Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 .p)2 3(1 .5) Proof We first find the m.8) Proof This can be shown.P)pB' (3.g.p)s s /3 .f. We omit the details (see. eau B(u) du = f PB 3PB SPB 0 o (3.3 .5 can be found in virtually any queueing book. by analytical manipulations (L'Hospital's rule) from (3. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used. Corollary 3.7) s +. g.1 Bo[s] = f oc. see e. 0 Notes and references Corollary 3./3B[s] which is the same as (3.(3B[s] 1 .. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3. for example. 191).5).4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . (3. 111112 or Feller [143].5). Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p.s .P)PB 2(1 . The approach there is to condition upon the first claim occuring at time t and having size x .(3 . Also (3.PPB2) EM2 = PPB) + QZPBl 2(1 . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.g. In view of (3.7).3 . Embrechts. [APQ] pp.s .p) E p"Bo[s]" = 1 . which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3. .5 The first two moments of M are 2 EM .p = (1 ..7) and Corollary 3. [APQ] pp. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. Bo of B0 as m e8u B(du) = B[s] .3. Griibel & Pitts [132]. 206207).6) 00 = (I .p)s .4) can be derived by elementary but tedious manipulations. In fact. e. Some relevant references are Abate & Whitt [2].Ps s(.)3B[s]) (3.pBo[s] no (1 .
For n < u < n + 1. .u) a)Qea" + (1 .Q (k 1 k= n  [O(k . differentiation yields Z'(u) _ /3Z(u) which together with the boundary condition Z(0) = 1/3 yields Z(u) _ (1/3) eAu so that (3.1).u)]k d 1 u) _ a) n ( du ( k! (1  .u)]k1 ku+1) [/3( k .Q) k=0 k! E e0( = /32(u) .66 CHAPTER III. Z^ =eR(k.u)]k k! (1 L3) 1: e_O(ku) NIN (k (k . of (3./32(u .3+ 18+ J0 Z(uy).Q) 3e.1)! k=1 u1 .1 < u < n and let Z(u ) denote the r.9).s.3I( 0<y<1)dy Z(y)/3I(0<u.9) shown for n . differentiation yields Z(u) _ /3Z(u) . THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.u) [N(k . Assume (3. then p) 1: ep(k u/.4) for Z( u) means f lhu Z(u) = 1.6 If B is degenerate at p. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .z/'(u) takes the form Z(u) L^J L.h.)3(1 .u)]k k! k0 The renewal equation (3.u) [p(k .y<1)dy 0<u<1 1 < u < oo uu ulhu 1a+/3 J0 uZ(y)dy U Z(y) dy 113+0 For 0 < u < 1./3Z(u .u + 1 )]k = QZ(u) .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.9) follows for 0 < u < 1.1). eO('u) [)3(k .
See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.g.r. The answer is yes: inserting in (4. but will now be repeated for the sake of selfcontainedness.2) shows that the solution is Ox [O]0].4.2) (Here 9 is any such number such that r. we just have to multiply (4. it follows that Z(u) = 2(u) for n<u<n+1. The question then naturally arises whether ie is the c. corresponding to a compound Poisson risk process in the sense that for a suitable arrival intensity 00 and a suitable claim size distribution BB we have no(a) = rc(a + 9) . in terms of the c.3B = .6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139].4) works as well.f. and thus (4. B9(dx) = B[9] B(dx). (4.g. and define rce by (4. of F9.rc(9) = . We could first tentatively consider the claim surplus X = St for a single t. (4.4) . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis.3B[9].g.1) or equivalently.1 that c(a) = /3(B[a] .f.f.d. say t = 1: recall from Proposition 1.a.1) .) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11. co(a) = rc(a + 9) . K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. (4. 0 Notes and references Corollary 3.4.1) .Qe(Bo[a] . 4 Change of measure via exponential families If X is a random variable with c. we set up .(9) is welldefined. Formalizing this for the purpose of studying the whole process {St}. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx).f.(9). or equivalently BB[a] = B[^+ Repeating for t 54 1.2).3) by t.a. F and c.
5) for the density of n i. and thus (4. Then the Xk are i.. BB by (4. v(Xi.0e and claim size distribution Be.S(k_1)Tln.d.FTn) = Q(SkTIn : k = 0.8) follows by discrete exponential family theory. replications from Fe (replace x by xi in (4.1. Ti(a)/n.2. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .5) for the density.9) Proof We first note that for any fixed t. (4.68 CHAPTER III. for G E FT. Proposition 4...t.8) By standard measure theory.r.r. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . t < T. . The identity (4.t. = exp {BST . and define 09. n) for a given n. EeeBSt + tk(B) = 1.(9)} (4. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.1) and multiply from 1 to n). oo) governing a given compound Poisson risk process with arrival intensity.i.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 .6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} . (4. The following result (Proposition 4.g. it suffices to consider the case where Z is measurable w. .i.f.10) . in particular the expression (4.1 Let P be the probability measure on D[0.2 For any fixed T. Xn).Tic (0)} . and PBT) the restriction of PB to FT. (4. . the corresponding expectation operator is E9.3 and claim size distribution B. G]. G]. with common c. . the PBT) are mutually equivalent on. and dP(T) dP^T) That is. .d. But let Xk = SkT/n .nr. (4. Z is measurable w.4). THE COMPOUND POISSON MODEL Definition 4. G C {T < oo}.3 Let T be any stopping time and let G E FT. . Let FT = o(St : t < T) denote the o•algebra spanned by the St. (4.7) Proof We must prove that if Z is FTmeasurable.FT. then EBZ = E [Ze9ST _T"(9)I .
Ee [exp { BST +Trc(9)} I(G) FT)] = 1.(Y) = 13(B['Y] . GT C_ Jr < T}. Thus.FT]] = EB [exp { BST + Trc(9)} I(G)] . Now consider a general G. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.1) . t = T . 77 Thus. . (0) + rc'(0)a = 0 + ES1 a = a (p .7 1 Some discussion further supporting this statement is given in the next section.r)rc(9)}I . c(a) is a convex function of a.ST) + (T . (4.9) holds with G replaced by GT.f.r is deterministic.g. The behaviour at zero is given by the first order Taylor expansion c(a) r.9) holds for G as well.1) _ 1 + a. Letting T t oo and using monotone convergence then shows u that (4.7) holds.1 It is seen that typically) a ry > 0 satisfying 0 = r. Then GT = G n Jr < T} satisfies GT E FT. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST .. subject to the basic assumption ij > 0 of a positive safety loading. according to what has just been proved.1(a).10). Then G E FT. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. 5. Thus by (4. Given FT. the typical shape of rc is as in Fig.5. and hence (4. 5 Lundberg conjugation Being a c.
3.2 is B(7) = 1 + ^.3) cf. 5.s). Equation (5.1) is known as the Lundberg equation and plays a fundamental role in risk theory .1(b). b[s] = 5/(b .QL instead of /37 and so on in the following .2) 7 Figure 5. we further note that ( 5. e. THE COMPOUND POISSON MODEL exists . It is then readily seen that the nonzero solution of (5. we write FL instead of F7.1 Consider the case of exponential claims.2)) is 7 = 5/3. Fig.4) yields /3L = b and that BL is again exponential with rate bL =. Taking T = r(u).1) (or (5. 5. G = {T(u) < oo} in Theorem 4. 5.4) ELS1 = #L(0) cf.g. Note that KL (a) = /L (BL [a] .1) .a = i(a + 7). An established terminology is to call y the adjustment coefficient but there are various alternatives around.3.2 s As support for memory. (5. the Lundberg exponent. Thus B[7] = 6/.3. Fig.70 CHAPTER III. Example 5 . Thus. u It is a crucial fact that when governed by FL. an equivalent version illustrated in Fig. . and (4. the claim surplus process has positive drift > 0. (5.1(b). (5. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes.1) is precisely what is needed for one of the terms in the exponent .
3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .5).G+L)(x)) dx ry^+L) J 00 f 0 (1 .6) Proof By renewal theory. T = T+. where C .(oo) (in the sense of weak convergence w. Proof Just note that e(u) > 0 in (5.Ce7u as u 4 oo. PL ) with density 1 .(u)} . LUNDBERG CONJUGATION 71 to vanish so that Theorem 4.1p . ST+ E A] .1 (5. see A . we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 .ascending ladder height distribution and µ+ its mean.1. (5. G = {S.1e. 0 Theorem 5 .r. (5.5. V)(u) < e7u.7) is the same as (5.3.+ E A} in Theorem 4. Since a7' is continuous and bounded. (5. we can rewrite this as 0(u) = e"ELe7^(u).G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.t. V) (u) = P(T(u) < oo) = EL [exp {ryS. T(u) < oo] .2 (LUNDBERG'S INEQUALITY) For all u > 0.8) .P Y j o' xeryxOB (x) dx /3k [Y] .5) Theorem 5 . take first 0 = ry.7) 0 and all that is needed to check is that ( 5.3 takes a particular simple form.u be the overshoot and noting that PL(T(u) < oo) = 1 by (5.4).e7x)G+(dx). which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx. Letting e(u) = ST(u) . To this end. e(u) has a limit i.6 ). Then P(ST+ E A) = EL [exp { 7S?+} .
72 CHAPTER III.7). that 7 = S .12) Example 5 .1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] . this solves the problem of evaluating (5.")G + (dx ) = 1  J0 00 3B(x) dx = 1p. Using (5. THE COMPOUND POISSON MODEL In principle.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5. or equivalently of how close the safety loading 77 is to zero.3 (this was found already in Example 5.1 . A direct proof of C = p is of course easy: B ['y] d S S (S7 )2 d7S y S 02' C 1p 1p _ 1p /3B' [7] 2 1 P1 p. u . of course. but some tedious (though elementary) calculations remain to bring the expressions on a final form.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.4). Then 0(u) = pe(a_Q)u where p = /3/S. we get L where 00 (1 . Noting that SIG(L)II = 1 because of (5.1) (5.10) VW = JI c* e° (x) dx = a (B[a] .4 Consider first the exponential case b(x) = Seax.8) yields +L) J0 xel'B ( x) dx (5.1 above) and that C = p. From this it follows. . (5.e.
5. LUNDBERG CONJUGATION Remark 5.5 Noting that PL  1 = ,3LIBL  1 = #ci (0 ) = k (ry) _ ,QB' ['Y]  1 ,
73
we can rewrite the CramerLundberg constant C in the nice symmetrical form G, _'(0)1  1  p K'(7) PL1
(5.13)
In Chapter IV, we shall need the following result which follows by a variant of the calculations in the proof of Theorem 5.3: 1  aB[ry  a]  1 Lemma 5 . 6 For a # ry, ELea^ (°°) = 7 aK'(7) 7  a Proof Replacing 7 by a in (5.7) and using ( 5.8), we obtain 1 (I 1  ^ e('ra) x,3 (x)dx) (L ) ELea^*) = a \\\ f
using integration by parts as in (3.6) in the last step . Inserting (5.12), the result follows. u
Notes and references The results of this section are classical, with Lundberg's inequality being given first in Lundberg [251] and the CramerLundberg approximation in Cramer [91]. Therefore, extensions and generalizations are main topics in the area of ruin probabilities, and in particular numerous such results can be found later in this book; in particular, see Sections IV.4, V.3, VI.3, VI.6.
The mathematical approach we have taken is less standard in risk theory (some of the classical ones can be found in the next subsection). The techniques are basically standard ones from sequential analysis, see for example Wald [376] and Siegmund [346].
5a Alternative proofs
For the sake of completeness, we shall here give some classical proofs, first one of Lundberg's inequality which is slightly longer but maybe also slightly more elementary:
74 CHAPTER III. THE COMPOUND POISSON MODEL
Alternative proof of Lundberg 's inequality Let X the value of {St} just after the first claim , F(x) = P(X < x). Then , since X is the independent difference U  T between an interarrival time T and a claim U, ,3+ry F'[7} = Ee7 ( UT) = Ee7U • Ee7T = B['Y] a = 1' where the last equality follows from c(ry) = 1. Let 0( n) (u) denote the probability of ruin after at most n claims. Conditioning upon the value x of X and considering the cases x > u and x < u separately yields
,0(n +1) (u) = F(u) +
Ju
0 (n) (u  x) F(dx).
We claim that this implies /,(n) (u) < e 7u, which completes the proof since Vi(u) = limniw 1/J(n) (u). Indeed , this is obvious for n = 0 since 00)(u) = 0. Assuming it proved for n, we get
„/, (n+1)(u) <
F(u) + e7u
00
Ju
e7(u=) F(dx)
00
<
f
e7x F(dx)
+ fu
e  7(u z) F(dx)
u
o0
= e 7uE[ 'Y] = e 7u.
Of further proofs of Lundberg's inequality, we mention in particular the martingale approach, see II.1. Next consider the CramerLundberg approximation. Here the most standard proof is via the renewal equation in Corollary 3.3 (however, as will be seen, the calculations needed to identify the constant C are precisely the same as above): Alternative proof of the CramerLundberg's approximation Recall from Corollary
3.3 that
(u) = )3
J OO B(x) dx + J U Vi(u  x)/3 (x) dx.
u 0
Multiplying by e7u and letting Z(u) = e7" O(u), we can rewrite this as
u Z(u) =
z(u) = e7u/
J
B(x)dx, F(dx) = e7x,QB(x)dx,
u
z(u)
f +
J
e7(ux ),Y' 1 • l•(u  x) • e7'/B(x) dx,
0
= z(u) +
J0 u Z(u  x)F(dx),
6. MORE ON THE ADJUSTMENT COEFFICIENT 75
i.e. Z = z+F*Z. Note that by (5.11) and the Lundberg equation, ry is precisely the correct exponent which will ensure that F is a proper distribution (IIFII = 1). It is then a matter of routine to verify the conditions of the key renewal theorem (Proposition A1.1) to conclude that Z (u) has the limit C = f z(x)dx/µF, so that it only remains to check that C reduces to the expression given above. However, µF is immediately seen to be the same as a+ calculated in (5.10), whereas
L
00
z(u) du =
f
J
/3e7udu "o
J "o B(x) dx = J "o B(x)dx J y,0eludu
u 0 0
B(x)^ (e7x  1) dx = ^' (B[7]  1)  As] [0 µs] = l y P^
using the Lundberg equation and the calculations in (5.11). Easy calculus now gives (5.6). u
6 Further topics related to the adjustment coefficient
6a On the existence of y
In order that the adjustment coefficient y exists, it is of course necessary that B is lighttailed in the sense of I.2a, i.e. that b[a] < oo for some a > 0. This excludes heavytailed distributions like the lognormal or Pareto, but may in many other cases not appear all that restrictive, and the following possibilities then occur: 1. B[a] < oo for all a < oo. 2. There exists a* < oo such that b[a] < oo for all a < a* and b[a] = 00 for all a > a*. 3. There exists a* < oo such that fl[a] < oo for all a < a* and b[a] = 00 for all a > a*. In particular , monotone convergence yields b[a] T oo as a T oo in case 1, and B[a] T oo as a f a* in case 2 (in exponential family theory , this is often referred to as the steep case). Thus the existence of y is automatic in cases 1 , 2; standard examples are distributions with finite support or tail satisfying B(x) = o(eax)
76 CHAPTER III. THE COMPOUND POISSON MODEL
for all a in case 1, and phasetype or Gamma distributions in case 2. Case 3 may be felt to be rather atypical, but some nonpathological examples exist, for example the inverse Gaussian distribution (see Example 9.7 below for details). In case 3, y exists provided B[a*] > 1+a*/,3 and not otherwise, that is, dependent on whether 0 is larger or smaller than the threshold value a*/(B[a*]  1). Notes and references Ruin probabilities in case 3 with y nonexistent are studied, e.g., by Borovkov [73] p. 132 and Embrechts & Veraverbeeke [136]. To the present authors mind, this is a somewhat special situation and therefore not treated in this book.
6b Bounds and approximations for 'y
Proposition 6.1 ry <
2(1  aps) 2µs
OMB PB)
Proof From U > 0 it follows that B[a] = Eea' > 1 + µsa + pB2)a2/2. Hence 1 = a(B[7]  1) > Q (YPB +72µs)/2) = 3µs + OYµa2) 2 (6.1) 7 'Y from which the results immediately follows. u
The upper bound in Proposition 6.1 is also an approximation for small safety loadings (heavy traffic, cf. Section 7c): Proposition 6.2 Let B be fixed but assume that 0 = ,3(77) varies with the safety loading such that 0 = 1 Then as 77 .0, µB(1 +rl) 2) Y = Y(77) 277 PB Further, the CramerLundberg constant satisfies C = C(r1)  1. Proof Since O(u) + 1 as r7 , 0, it follows from Lundberg's inequality that y * 0. Hence by Taylor expansion, the inequality in (6.1) is also an approximation so that OAY]  1) N Q (711s + 72µB2) /2) = p + 3,,,(2) B 'y 7 2 2(1  p) _ 271µB
QPB PB)
6. MORE ON THE ADJUSTMENT COEFFICIENT 77
That C 4 1 easily follows from y 4 0 and C = ELe7V°O) (in the limit, b(oo) is distributed as the overshoot corresponding to q = 0 ). For an alternative analytic proof, note that C  1P = rlµB 73B' [7]  1 B' [ry)  1/0 711µB µB +7µB2 )  µB(1 +77 ) 'l = 1. 277q
77
7PBIPB
 77
13 Obviously, the approximation (6.2) is easier to calculate than y itself. However, it needs to be used with caution say in Lundberg's inequality or the CramerLundberg approximation, in particular when u is large.
6c A refinement of Lundberg 's inequality
The following result gives a sharpening of Lundberg 's inequality (because obviously C+ < 1) as well as a supplementary lower bound:
Theorem 6 .3 C_eryu < ,)(u) < C+ eryu where
= B(x) = C_ x>o f °° e7( Yx)B(dy )' C+
B(x) xuo f 0 e'r( vx)B(dy)
Proof Let H(dt, dx ) be the PLdistribution of the time r(u) of ruin and the reserve u  S7(„)_ just before ruin . Given r(u) = t, u  ST (u) = x, a claim occurs at time t and has distribution BL(dy)/BL(x), y > x. Hence ELe7£(u) 0
J
°o
H(dt, dx)
fX
eY(Y x) 00 f°° B(dy) x
BL dy
BL(x)
o
f
f H(dt, dx)
L ^ H(dt, dx) f e7B( x)B(dy) Jo oc, < C+
J0 0 o" H(dt, dx) = C. o" J
The upper bound then follows from ik(u) = e7uELeVu), and the proof of the u lower bound is similar.
78 CHAPTER III. THE COMPOUND POISSON MODEL
Example 6.4 If B(x) = eax, then an explicit calculation shows easily that B(x) _ e6X fz ° e7(Yx)B(dy) f x' e(6,6)(Yx)8esydy = 5 = P. Hence C_ = C+ = p so that the bounds in Theorem 6.3 collapse and yield the exact expression pey" for O(u). u The following concluding example illustrates a variety of the topics discussed above (though from a general point of view the calculations are deceivingly simple: typically, 7 and other quantities will have to be calculated numerically). Example 6.5 Assume as for (3.1) that /3 = 3 and b(x) = 2 .3e3x + 2 .7e7x, and recall that the ruin probability is 24 5su 5eu + 3e *(u) = 3 Since the dominant term is 24/35 • e", it follows immediately that 7 = 1 and C = 24/35 = 0.686 (also, bounding aS" by a" confirms Lundberg's inequality). For a direct verification, note that the Lundberg equation is
7 = /3(B['Y]1)
= 3\
2.337
+2.7771
which after some elementary algebra leads to the cubic equation 273  1472 + 127 = 0 with roots 0, 1, 6. Thus indeed 7 = 1 (6 is not in the domain of convergence of B[7] and therefore excluded). Further, 1P = B [7] 181B = 13 2.3+2.71 = 1 3 1 7 I 7'
_ 17
2 (3 a )2 + 2 (7  a)2 «=7=1 2 1p _ 7 _ 24
36 '
3.171 35* 36 For Theorem 6.3, note that the function QB[Y]1 f°°{L 3e_3x+
u
• 7e7x 1 dx
J
3 + 3e4u
f 0c, ex .
I 2 . 3e3x + 2 . 7e7x l dx
l J
9/2 + 7/2e4u
7. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 79
attains its minimum C_ = 2/3 = 0.667 for u = oo and its maximum C+ = 3/4 = 0.750 for u = 0, so that 0.667 < C < 0.750 in accordance with C = 0.686.
Notes and references Theorem 6.3 is from Taylor [360]. Closely related results are given in a queueing setting in Kingman [231], Ross [308] and Rossberg & Siegel [309]. Some further references on variants and extensions of Lundberg's inequality are Kaas & Govaaerts [217], Willmot [382], Dickson [114] and Kalashnikov [218], [220], all of which also go into aspects of the heavytailed case.
7 Various approximations for the ruin probabil
ity
7a The BeekmanBowers approximation
The idea is to write i (u) as F(M > u), fit a gamma distribution with parameters A, 6 to the distribution of M by matching the two first moments and use the approximation
0(u)
f
u
Sa
r(A)
xa  leax dx.
According to Corollary 3.5, this means that A, 8 are given by A/S = a1, 2A/52 = a2 (2) PIB3) ^ZP(B)2 __ PPB a2 al 2(1  P)PB 3(1  P)µ8 + 2(1  p)2' i.e. S = 2a1 /a2, A = 2a2 1/a2.
Notes and references The approximation was introduced by Beekman [60], with the present version suggested by Bowers in the discussion of [60].
7b De Vylder's approximation
Given a risk process with parameters ,(3, B, p = 1, the idea is to approximate the ruin probability with the one for a different process with exponential claims, say with rate parameter S, arrival intensity a and premium rate p. In order to make the processes look so much as possible alike, we make the first three cumulants match, which according to Proposition 1.1 means p=AUB1=P1,
2N
(2) 6^= =OP
,
/3,4)
.
)3 )PBo µB  ./3)] 1 .p = (/3max 0)µB. but has an obvious interpretation also in risk theory: on the average. heavy traffic conditions mean that the safety loading q is positive but small.3 and Corollary 3. we shall represent this situation with a limit where /3 T fl but B is fixed.1. Notes and references The approximation (7. and hence the ruin probability approximation is b(u) e(bAln)u.PBo [s (/3max . we have according to the PollaczeckKhinchine formula in the form (3./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 . the premiums exceed only slightly the expected claims.1. Though of course it is based upon purely empirical grounds. Proposition 1.Ps(/3max . Proposition 7. [174]) shows that it may produce surprisingly good results. cf. That is. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory.2) was suggested by De Vylder [109]. the approximating risk process has ruin probability z. p* _ ./3)PBo PB .p + p { 1 1p ti 1 .8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) .(3)2 P PB 2µB 2µB Letting /3* = /3/P.3* /S.3 )1 } _ 1p 1 . Grandell [171] pp.1 As /3 f Nmax. or equivalently that /3 is only slightly smaller than /3max = 1/µ8.p .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . Letting Bo be the stationary excess life distribution. 1924. numerical evidence (e.80 CHAPTER III. Mathematically.P .g. (/3max . THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 .b(u) = p*e.s(/3max .(bA*)".
Numerical evidence shows that the fit of (7. VIII). .ryu . [APQ] Ch.ze a2unµB laB (7. Mathematically.g. Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. These results suggest the approximation Vi(u) e6(0_. It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce.4) suggested by the Cramer Lundberg approximation and Proposition 6. We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . light traffic is of some interest as a complement to heavy traffic ./3)u * v. 2 provides the better mathematical foundation. the first results of heavy traffic type seem to be due to Hadwiger [184].7.0)u. then P(u) 4 e6„ Proof Write z'(u) as P((/3max . In the setting of risk theory.p. but has an obvious interpretation also in risk theory: on the average .B AB ) 6()3max _'3) = However . in risk theory heavy traffic is most often argued to be the typical case rather than light traffic ./3)u).p _ 2rl11B PB p. light traffic conditions mean that the safety loading rl is positive and large .2 If . while the approximation may be far off for large u. the premiums are much larger than the expected claims . However . we shall represent this situation with a limit where 3 10 but B is fixed. That is . obviously Corollary 7.3) is reasonable for g being say 1020% and u being small or moderate. the term light traffic comes from queueing theory. u * oo in such a way that (3max . 7d The light traffic approximation As for heavy traffic . and hence 2µ2B 1 . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . or equivalently that 0 is small compared to µB .1 1 . VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7.2. as well as it is needed for the interpolation approximation to be studied in the next subsection.Q T /3max. This follows since rl = 1/p .l3)M > (/3max . Of course.
U > u] = /3iE(U . 0(u) /3 J B(x)dx = /3E[U . (7.u.= 1 aJ 1 a 0+ 1 = = p. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx.3 As . ( 3 J O B dx. Another way to understand that the present analysis is much simpler than in these references is the fact that in the queueing setting light traffic theory is much easier for virtual waiting times (the probability of the conditioning event {M > 0} is explicit) than for actual waiting times . Indeed.T > u). The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . THE COMPOUND POISSON MODEL Proposition 7. . En'=2 • • • = O(/32) so that only the first terms matters.82 CHAPTER III.5) u Proof According to the PollaczeckKhinchine formula. cf. i. [97]. the Poisson case is much easier than the renewal case. Sigman [347]. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. see Daley & Rolski [96].e.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. by monotone time T of the first claim .u)+. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. z/' (u) convergence P(U .T > u) = J o" B(x + u)/3eax dx . Light traffic does not appear to have been studied in risk theory. Again. For a more comprehensive treatment. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. Omax max m.(3 10.5) follow by integration by parts. u Note that heuristically the light traffic approximation in Proposition 7.3 is the same which comes out by saying that basically ruin can only occur at the F(U . 10 ( u The alternative expressions in (7. Asmussen [19] and references there.
(U). Instead.O0 M. no empirical study of the fit of (7.VHT) ( ax QmQ ) h (B) ( . Notes and references In the queueing setting .8. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). one may hope that some correction of the heavy traffic approximation has been obtained. z/i(E) (u) = pe(QmaxQ)u. however. we combine with our explicit knowledge of ip(u) for the exponential claim size distribution E whith the same mean PB as the given one B./3)) . the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ].3). _(E) (u).Wmax f(x ) dx + pee6mQ. we may ask which one carries the larger risk in the sense of larger values of the ruin probability V(') (u) for a fixed value of 0. with rate 1/µB = /3max. Thus . B(2). (7. that is. even if the safety loading is not very small. . Let OLT) (u) denote the light traffic approximation given by Proposition 7.3 and use similar notation for %(B) (u) = (u).6) is . to get nondegenerate limits . COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . Another main queueing paper is Whitt [380].x . Al . available. Substituting v = u(. The adaptation to risk theory is new. f / Qmax B(x)dx 00 eQmaxxdx 4/ Qmax 00 QmaxQ amaze" and the approximation we suggest is J B(x) dx = cLT(v) (say).O(E)(u) 1 (1 . "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm.6) (1p) The particular features of this approximation is that it is exact for the exponential distribution and asymptotically correct both in light and heavy traffic.3n. where further references can be found . ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). ) M. we see that the following limits HT) (u'). [84].
equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. this ordering measures difference in variability. then . or the existence of random variables U(l). Bill is said to be convexly smaller than B(2) (in symbols. Taking probabilities. In the literature on risk theory. B(') <i.6. Rather than measuring difference in size. In terms of the time to ruin. Proposition 8.1 is quite weak. THE COMPOUND POISSON MODEL To this end. u Of course. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B.2 If B(') <j. B(2)) if f fdB(1) < f fdB(2) for any convex function f. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. Recall that B(') is said to be stochastically smaller than B(2) (in symbols.' 1)(u) < V)(2) (U) for all u.s. U(2) distribution B(2) and U(1) < U(2) a. the proof is complete. B(' <. and a particular deficit is that we cannot compare the risks of claim size distributions with the same mean: if BM <d B(2) and µB«) = /IB(2). In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. XI. Proposition 8. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. Finally. we shall need various ordering properties of distributions.1 If B(') <d B(2). an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. we have the convex ordering.ill(u) < V)(2) (U) for all u. we can assume that 1) < St 2l for all t. . B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. Proof According to the above characterization of stochastical ordering. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. B(2) and PB(1) = µB(2). then Bill = B(2). whereas (consider x2) B(2) has the larger variance. then i. Here convex ordering is useful: Proposition 8. U(2) such that U(l) has distribution B('). cf.84 CHAPTER III.
.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1.3 provides another instance of this. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal. from which the result immediately follows. and here is one more result of the same flavor: Corollary 8. Example 8. with fixed mean. then B(1) <.(1) (.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. then /'(')(u) < 0(2)(u) for all u. Proof If f is convex. say to p.1 and µB at 1 so that the safety loading 11 is 10%. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB. B.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). larger variance is paramount to larger second moment. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. . Hence by the PollaczeckKhinchine formula . and consider the following claim size distributions: B1: the standard exponential distribution with density ay.8.1.6 Fix /3 at 1/1. This u implies that D <.5 If '0(1)(u) < p(2) (U) for all u and a. we have by Jensen 's inequality that E f (U) > f ( EU). B(2). (D) (u) < O(B) (U ) for all u.4) certainly supports this view: noting that.3 If B(1) <. Proof Consider the light traffic approximation in Proposition 7. A first attempt would of course be to identify 'variation' with variance. A partial converse to Proposition 8. A general picture that emerges from these results and numerical studies like in Example 8.4 Let D refer to the distribution degenerate at 'LB . Then V. The heavy traffic approximation (7. Corollary 8. The problem is to specify what 'variation' means. Bo1) <_d Bo2) which implies the same order relation for all convolution powers.2 is the following: Proposition 8.. B(2).e. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities.
1358.0' U0. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less.9A2e'2r where A. One then obtains the following table: U005 U0. and this is presumably a consequence of a heavier tail rather than larger variance. in comparison to B2 the effect on the ua does not show before a = 0. Let ua denote the a fractile of the ruin function. For B1i B2. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. Kluppelberg [234]. and consider a = 5%. i.4. the behaviour of which is governed by a parameter 9. B. 0.001 u0. sensitivity analysis (or pertubation analysis) deals with the calculation of the derivative (the gradient in higher dimensions) of a performance measure s(O) of a stochastic or deterministic system. However. Note to make the figures comparable.e. 1/)(u..e'\1x + 0.86 CHAPTER III. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!).) = a. 11 Notes and references Further relevant references are Goovaerts et al. 0. A standard example from queueing theory is . In terms of variances o2. B3: the Erlang distribution with density 4xe2x. We return to ordering of ruin probabilities in a special problem in VI. Pellerey [287] and (for the convex ordering) Makowski [ 252].lA. [166]. 32 50 75 100 B2 B3 B4 35 181 24 282 37 70 245 425 56 568 74 1100 (the table was produced using simulation and the numbers are therefore subject to statistical uncertainty). B3 the comparison is as expected from the intutition concerning the variability of these distributions. = 0. 1%. van Heerwarden [189].4142.01%. B4: the Pareto distribution with density 3/(1 + 2x)5/2.000. 9 Sensitivity estimates In a broad setting.1%. A2 = 3.01%. all distributions have mean 1.
In the present setting.(u) for large u. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV.2 Consider a risk process { Rt} with a general premium rate p.9. while /3 = j3 is an estimate. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . In particular . Similar conclusions will be found below. Assume for example that 8 is known. For example. i. where Q2 = fl ( l2 1113 / _ Ou2v)2.3. Thus.19P a/ .01/2u. the distribution of %3 0 is approximatively normal N(0„ Q/t). Then ib = Pe(613)u.. s(9) is of course the ruin probability t' = Vi(u) (with u fixed) and 0 a set of parameters determining the arrival rate 0. obtained say in the natural way as the empirical arrival rate Nt/t in [0. Thus at p = 1. where the partial derivatives are evaluated at p = 1. a2/t).e. a0 as ao 80 19P . Then if t is large . Proof This is an easy time transformation argument in a similar way as in Proposition 1. u Proposition 9.1. a/3 0 . the premium rate p and the claim size distribution B. if = a e(6A)u.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). we may be interested in a'/ap for assesing the effects of a small change in the premium.Ap). SENSITIVITY ESTIMATES 87 a queueing network. it follows that ' is approximatively normal N(0. t]. Then the arrival rate /3(P) for { R(P) } is )31p. Then a p ao = 00 Qa/. with 0 the vector of service rates at different nodes and routing probabilities. increasing in u. Let R(P) = Rtli. and s(9) the expected sojourn time of a customer in the network. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . say estimated from data. Example 9.
Differentiating w. Similar notation for partial derivatived are used below. namely that of a twoparameter exponential family of the form Bo. 3) ( 9 . and write yp = 8y/8/3 and so on . this intuition is indeed correct.w(O.3.Owe (9.(/3 + y)we(9 + 7.w(6. for the ruin probabilities . /3 yields w e(e + Y. various parametric families of claim size distributions could be considered. (9.^)] 1(/3+y)we(9+'y. (3+'y)PC (0+7.5) are similar. 4) (9 . ()} p(dx) . it suffices to fix the premium at p = 1 and consider only the effects of changing . so that heuristically we obtain '00 50ryu = Coe"u . mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. Viei '0(. Consider first the adjustment coefficient y as function of 3.3. we cannot expect in general to find explicit expressions like in Example 9.3 70 = 'Ye = = 7 /3(1 we(e +'y. x > 0 (9.2) (see Remark 9. Consider first the case of 8/8/3: .3) follows by straightforward algebra. u Now consider the ruin probability 0 = 0 (u) itself. Proposition 9.()^ 1 . () Proof According to (9.((dx ) = exp {Ox + (t(x) . 5) (Q+'Y)[we(0+7. (9.g. we can rewrite the Lundberg equation as w(9+ y. but must look for approximations for the sensitivities 0.0 = t/'(u) and the CramerLundberg constant C.()YC = 1 +y/ /3 \ Q2 From this (9.1 or Proposition 9. However . In the case of the claim size distribution B.4). The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation .t.()(0 +'0) ' (9 .10) below.()wC(e.r.uypCe7u urypO.3 or/and B.88 CHAPTER III. (. ^) . 9. but we shall concentrate on a special structure covering a number of important cases. e.6 below for some discussion of this assumption).6) As will be seen below. () = log(1 + y//3). THE COMPOUND POISSON MODEL As a consequence. Of course. and the proofs of (9.
QB(x) dx.we(9 .w(9.3) for z/'(u).x) F(dx ) f u J C F(dx) = C as u 4 oo. () exp {w(9 + a. By dominated convergence. () .x). we multiply by e7" and let Z(u) = elt" cp(u).3 (see in particular (5. we get p(u) = J "O B(x) dx + J U O(u . Z= zl + z2 where zl (u) = e7u J m B(x)dx. () .w(O. Combining these estimates . Z(u)/u a C//3PF where PF is the mean of F.4 As u oo. F(dx) = e'yy/3B(x)dx. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.x)B(x) dx + J U W(u .9.([a] = exp {w(9 + a. ()} .3(x) dx. ()} .9) (9. the proof is complete.2 of the Appendix ).g.8) Letting cp = e0/e/3 and differentiating (9.4t (U)e°`U = which are wellknown and easy to show (see e.12)).x)B(x)dx. 0(u) = /3 Ju"O B(x) dx + f 0 0(u . PF = (1 . SENSITIVITY ESTIMATES Proposition 9. 11 For the following.10) (9. BarndorffNielsen [58]). z2(u) _ 1 ^ e'ri`i7i( u . z2(U) = e7" J u b(u . we note the formulas Ee.St (U) Ee.w(9. u 0 Then Z = z + F * Z and F is a proper probability distribution .x). Further write de = [we (9 +'y. and alsoo zl(u) + 0 because of B['y ] < oo. Be. () . But from the proof of Theorem 5. O} (9.8).p)/C'y.C). Hence by a variant of the key renewal theorem (Proposition A1.(e"U = = wS(O.8) (Section 5). ()] exp {w (O + y. (9.11) Ee. () . w((9 + a.
2 z 07P N ue7u (3C de .5 Assume that (9. 0 x Multiplying by e7" and letting Z(u) = e"uV(u). THE COMPOUND POISSON MODEL [we(e+7. By dominated convergence and (9.x).x)f3 f ^[t(y) .wc(O. 01 (i+) do = +'Y.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . u Z2(U) = e7° f u ^/i(u .6e7u f "o f[t(y) . Then as u > oo. ()]e7vB(dy) 'fCd 7 c . ()} 1z(dy) = f [t(y) .1) B(dy) 'f '[t(y) .wc (O. ^)} [wc (0 + 7.12) f exp {O y + (t(y) .2) holds. ()]B(dy) dx.QB(x)dx. F(dx) = e7x.11). this implies Z = z + F * Z. ()](e7v .w(0.9)(9.6C do 89 1p 8( 1p Proof By straightforward differentiation.90 CHAPTER III. ()]B(dy) dx x 0 0C T ON O .w( (0.8) that cp(u) . z = zl + z2.wc(9.we (0. C)] (1 + 7 ) Proposition 9. oo z2 (u) f C . 8^ ue7u. ^) .w(e. 8 8() 8( (9.w( (0. )}B(dy)• Letting cp it thus follows from (9.e7x/3 f 00 [t(y) .w (9. ())B(dy) dx. C) .
t(x) = logx.6 Consider the gamma density b (x ) = Sa xa.a log S = log r(c) .18) (05 + 57 _'3_y .1 . ./35' a/i'y + aryl 625ry.Y)a+1 ' (9.yu/3C2do u86 89 1p' az/) = 8z/... () = C/9 = a/S. < = a.(log r(a) a log S)} • r(a) 1.1edz = 1 exp {Sx + a log x . by inserting in the above formulas. Example 9.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 . ue_Yu 'C2d( 8a 8( 1 p . It follows after some elementary calculus that p = a)3/5 and. we (9. We get w( (0.2) holds with p(dx) = xldx.ry) 5a1 cry (5 . SENSITIVITY ESTIMATES as u 3 oo. a /(S . U 7µF from which the second assertion of (9.C log(9). Z(u) /3C 91 o c'o e11(t (y) . w(e. () ='I'(t. ( 9.Sry a/32 + a/37 + /37 .16) (9.) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function. 9 = S.a/35a&y' ' (9.w((9.3ary tog('Finally.12) follows.12) takes the form y) alp a . ())B(dy) < oo. () = log r(a) .pa+1 .15) (9.13) (9.9. and also zj (u) 4 0 because of f Hence.QS 1 . Here (9. and the proof of the first one u is similar. that C = a..rye) S 5rya.17) (9.
3. for a < a* = z (.2 log (0.S[a] = exp {w (9 + a.9) (() .3Ee.log c = 2 In particular. which we omit in part . C) = B = Yc = de = do = . () = Cc .22."62 . t(x) _ . C) .l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2. further yield .1 16 +ry c C22ry 2( = + 70 We (e.1 = eXP {c(C . C = .([Y] . Be.w(9.2 .92 CHAPTER III.21og 2.CZ try)} 1 C C2 try . ()} = exp {c (C .2a) } Thus the condition B[a*] > 1 + a* /. Straightforward but tedious calculations . w(e.7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9. 9 = . THE COMPOUND POISSON MODEL Example 9.2) with µ(dx) = 2x3zrdx.
8 The specific form of (9. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x).7 and references there. then BT and hence ryT is undefined. if 1 PT = /3TNT(U1+.+UNT) > 1. (9.cue_7u)3C P Remark 9. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . Also. we have assumed k = 2 and ti (x) = x. To this end.12) takes the form a = a 93 ar. In general. in which case we can just let t(x) = 0. or Ct(x). sj=1 and let YT be defined by IKT('ryT) = 0. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3.. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. [379] consider a special problem related to reinsurance..a. in u which case the extension just described applies. to our knowledge. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. However. Thus. the results presented here are new. the models there (e.2 of the parameters.2) is motivated as follows. Comparatively less work seems to have been done in risk theory.. thus. the main tool is simulation. let NT 16T = ^T . Finally if k = 1.10. However . That it is no restriction to assume k < 2 follows since if k > 2. ae t 1lEY u S _ . Note that if NT = 0. B are assumed to be completely unknown.oo. we can just fix k .g. BT [a]= NT ^` e"U. queueing networks) are typically much more complicated than the one considered here. for which we refer to X.1) . . kT (a) = /T (BT [a] . and we estimate y by means of the empirical solution ryT to the Lundberg equation. Thus. then ryT < 0. and hence explicit or asymptotic estimates are in general not possible. That it is no restriction to assume one of the ti(x) to be linear follows since the whole setup requires exponential moments to be finite (thus we can always extend the family if necessary by adding a term Ox). Van Wouve et al. the exponent is either Ox.3C2de 1p' z a = c .
1)2 + E[27] . since NT /T .1 As T 4 oo.)vl+ N CO. vfoVFB[2y]./^ B[27] . a2 where a2 = /3r. B[2'Y]  /3T ) . Lemma 10 . 1) r.T y ..B[7]2 V2 .94 CHAPTER III.Q and Anscombe 's theorem.'s. V2 are independent N (0.(27)/K'(7)2. N ( n[7]. .2 As T * oo. THE COMPOUND POISSON MODEL Theorem 10 .: N ()3.B[7]2 }) ( T 0 .3T .3/ T).: N 0.B[7]2 n Hence ( 10.1) 'YT .2) follows from NT/T a4' .If . For the proof.2) rT(7) N N (0. Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . 7T a4' 7. we need a lemma.v.1) . More generally.B[7]) + B [7] . B [7]2 (10.B[7]) 0+ Iv/o(b[y]. then (10. If furthermore B[27] < oo. (10.1) . it is easy to see that we can write \ V1 1 l _ .a BT[7] I B[7] I + . 16T where V1.i3)(B[7] 1) + (3(BT[7]  . B[27] .'Y .b[Yp'V21 T { (E[7] .7 + (.3) Proof Since Var(eryU) = we have B[7].
E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r.4) + E). If ryT E (7  we have KT(7 .'T(a) = 1 E Uie°U' a$' EUe "u = B'[a]. where ryT is some point between ryT and ry. To this end . OT a 95 u 4 /3.3).'(y).4) and Lemma 10.KT(7) kT(7) K'(7) .e. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).1 By the law of large numbers. it follows that 7T7 KT(7T) . Proof of Theorem 10.(ry + e) and hence KT(7 . Combining ( 10. y + E) eventually.e) < 0 < r. NT i =1 n'(a) for all a so that for all sufficiently large T K7 .c'(7) N (0' T (2(7) / N (0. 7T E (y . Then r.E) < 0 < kT(7 + E) for all sufficiently large T .(ry ..1 can be used to obtain error bounds on the ruin probabilities when the parameters . lcT(a ) 4 /c(a). ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10. BT[a] 3 B[a]. NT BT [a] Hence r.10.E) < 4T(7T) < 4T(7 + E). Let 0 < E < ry. Theorem 10. and the truth of this for all e > 0 implies ryT at 'y.Q. first note that e7TU N (e7U u2e27Uo'2/T) 7 . By the law of large numbers.e. 6"Y (10. 0 are estimated from data .2. °7IT) . I.
and the known fact that the Y„ = max Vt tE[W„1..0) < 5.Wn) are i . A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. i.f. if B is exponential with rate 8 so that ry = 8 .. One (see Schmidli [321]) is to let {Vt} be the workload process of an M /G/1 queue with the same arrival epochs as the risk process and service times U1. with a tail of the form P(Y > y) .T VIT where r7ry. Herkenrath [192].96 CHAPTER III. Wn). Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. [197]. Notes and references Theorem 10. Vt = St .g.5%). Embrechts & Mikosch [133].. Csorgo & Teugels [95]. i .) = a (e.C1e"a ( see e.e.. = 1.1 : Vt = 0.e.ueryuU ". ft..96 if a = 2.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. Letting Wo = 0.i. THE COMPOUND POISSON MODEL Thus an asymptotic upper a confidence bound for a7' (and hence by Lundberg's inequality for 0(u)) is e"TU + f. Deheuvels & Steinebach [102]. > 0 for some t E [Wn_ 1..g. various alternatives have been developed.Q.. V. Mammitzsch [253] and Pitts. wn = inf{t > W.info< „< t S.. t]}. the nth busy cycle is then [Wn1. Asmussen [23]) can then be used to produce an estimate of ry. satisfies b(.1 is from Grandell [170].3 or equivalently p > 1/2 or 11 < 100%. U2. Frees [146]. This approach in fact applies also for many models more general than the compound Poisson one. For this reason . it means 2 (8 . . Griibel & Embrechts [292]. 6 < 2. For example .d. Hipp [196].
f. See Fig. 0. exists.s. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . 97 . generalizations to other models are either discussed in the Notes and References or in relevant chapters. Unless otherwise stated. B[•] and mean AB.1 (the role of ryy will be explained in Section 4b).g. Further let 'Yo be the unique point in (0. The safety loading is q = 1/p . the premium rate is 1. T) = P( /r(u) <T) \ = PI inf Rt <OIRo=u1 /\0<t<T PI sup St>ul 0<t<T Only the compound Poisson case is treated.1 where p = 13µB. The notation is essentially as in Chapter III. In particular.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. the Poisson intensity is 0 and the claim size distribution is B with m. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. 'y) where c(a) attains it minimum value.1) .
2 that E [T(u)k. 1 Exponential claims Proposition 1. using that the overshoot l.98 CHAPTER IV.1) (1. .(.(U) < 00] = ELT(u)ke'YS. PROBABILITY OF RUIN IN FINITE TIME Figure 0. E[T(u) I T(u) < 00 ] = ELT (U). Var[T(u) I T(u) < 00] = VarL T( U) . PL = 6/0 = 1/p > 1). we have for k = 1.9).r. the time of ruin is T(u) and ^(u) = ST(t&) . FL and independent of T(u). In particular. 1 FL.2) Proof Let as in Example 111.5 . By the likelihood identity III.) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k. (u) is exponential with rate 0 w.u is the overshoot. the conditional mean and variance of the time to ruin are given by E[r(u) I T (u) < oo] Var [T ( u) I T( u) < oo] /3u+1 J )3 _ 2/3Su+/3+S (S)3)3 (1.(4.t.. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .1 The claims surplus is {St}.1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. 7.
6a = 0 with solution 0 (the . .1//32 (6/)3 1)2 26(/3u + 1)/(6 .1.6 + a)0 .I (1.1)ELT(u).1)) ELST(u) ELT(u) (PL .(PL .B = a.0) .2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.3) B = 0(a) = + (6/3a)2+4a6 2 and hence that the value of ic(yo) Proof It is readily checked that yo = 6 .1) . we have by Wald's identity that (note that ELSt = t(pL .s.2).6. the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).1)T(u))2 = UL where = s.2) is aLELT( u) .1)T(u)) = VarLe(u) + (PL . 0 Proposition 1. T(u) < oo] fora > r. where = eBu I 1 .h. u + ELe(u) _ PL .V/ is as asserted.1)T(u) are independent with QL the same mean .(yo) = 2V ."(ry) = 26//32. Since Sr (u) and (PL .12 Thus the l.1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1. of (1. This means that /3(6/(6 . Wald's second moment identity yields 2 EL (Sr(u) . which leads to the quadratic 02 + (/3 . EXPONENTIAL CLAIMS For (1 .1)2VarLT(u) + 2 Ca 1I VarLT(u).s.1 (6)3)2 which is the same as the r.h. Let 0 > yo be determined by ^c(0 ) = a.s. the 1./3 . is V1rLSr( u) +VarL ((PL .h./3) . 1).
.100 CHAPTER IV.. (1. are the lengths of of the ladder segments 2. Using 5 = 6 ... Fig.. But by the fundamental likelihood ratio identity ( Theorem 111.4. Cf. T(u) < oo] = EB [exp {aT(u ) .3) we have E [e«T(u ).OuEee 04(u) = ee u be BB+B where we used that PB(T(u) < oo ) = 1 because 0 > ryo and hence E9S1 = K'(0) u > 0. Ti. St Ti F.v..9ST(u) +T(u)!c(0)} .3 that we can write EeaT( u) = eeuEe 017(o).v.T+ Ti a t U T I 1 a i F.1 .0.'s with rate 5. . Note that it follows from Proposition 1. .1 where Y1.3. Y(u) belonging to a convolution semigroup . the result follows. More precisely. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment .4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. T2 .Y1 Y2 Figure 1. 1. Y2. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0). are the ladder heights which form a terminating sequence of exponential r.. and M(u)+1 is the index of the ladder segment corresponding to T(u). T(u) < oo] = e..
Proof We use the formula . T) to be evaluated by numerical integration: Proposition 1.e.3 Assume that claims are exponential with rate b = 1. T.T are conditionally i. . 2.1(u. UN. and exponential with rate S = 1.T + • • • + UN.cos (u/..i. . including the customer being currently served).0. UN.1.ST). [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . For j = 0. Hence 00 F(VT > u ) P(QT = N)P(EN > u) N=1 00 N1 k F(QT = N) eu N=1 k=1 °O u k! k Ee k=0 1t P(QT . 1).. the conditional distribution of VT given QT = N is that of EN where the r. Since U1 . Let {Qt} be the queue length process of the queue (number in system.1 )!.v.T the service times of the customers awaiting service .1.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . let (cf. the following formula is convenient by allowing t. If QT = N > 0. U2.T) = P(VT > u) where {Vt } is the workload process in an initially empty M/M/1 queue with arrival rate 0 and service rate S = 1.I ex cos B cos j O dB fo " ..3 sin0 + 29) f3(0) = 1+/32/cos9.4.6. Corollary 11.. .. EN has an Erlang distribution with parameters (N. .T. where U1.k + 1). Then V(u.6(u) = Vfl/j l(Su..T. i. density xN lex/(N . .d.i (u. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V. then VT = U1. EXPONENTIAL CLAIMS 101 For numerical purposes .T) 1 I fl(O)h(0) fdO where (1.T is the residual service time of the customer being currently served and U2 . cf.T..
in particular equations (1. f3(0) . 912. PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.1 R [.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .1 00 ok+lR 00 j=k1 +1)/2e .cos((k + 1)0)] f3(0) 00 flk +1 > j=k1 3j/2 COS(jB) l)/2ei(k+1)e )3j/2eije = R)3(k+ (31 /2eie . Then (see Prabhu [294] pp.102 CHAPTER IV. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .ie . k k2 + $k+1 E bj 00 t j .i(k +1)e R [/3( klal/2e:0 (01 /2 e .112 l 1( k +1)/2 [ 31/ 2 cos(kO) .1)] L _112 /(k+1)/2 [. let I _ j (x) = Ij (x).(31/2eie .cos((k + 2)9)] d9.8 ) yields F(QT > k + 1) . similar formulas are in [APQ] pp.)3k +1 tj g'(QT >.3(k +1)/ 2ei(k + l)6 (. and define tj = e(1+R)Taj/2Ij(2vT T). (1.13(k +l)/2ei(k +1)9 R E .44).31 /2eie L 1)] 1 I/31/2eie .cos (( k + 1)0)] f3(9) Hence the integral expression in (1. 8789) 00 E aj j= 00 = 1.k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.38).(31/2 cos (( k + 2)9) .
equivalently. however. is numerically unstable for large T. or. k! k=O k0 i/z Co Uk ate" o'/z e . F(x. . We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p.7) that _ [^ au ak+l (30 k L. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. t )). Ui < x I / (note that P(St < x ) = F(x + t. T) in terms of F(. there are several misprints in the formula there. t) = P . k=0 103 Cu) A further application of Euler's formulas yields cc k =0 k 'ese)k __ U #kJ2 cos((k + 2)9) = R eNO ^` (u^1 L k= = eup i/z L OI = =ateU161/2 e '0+2iO COS a cos(u(31/2 sin 9 + 20). We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. it follows as in (1. from the accumulated claim distribution N.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59].e = e' COS a cos(uf31/2 sin 0). The rest of the proof is easy algebra.0(u.. the numerical examples in [12] are correct). T). and the next one (often called Seal's formula but originating from Prabhu [293]) shows how to reduce the case u 54 0 to this. Related formulas are in Takacs [359]. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. expresses V)(0.2. The first formula. Seal [327] gives a different numerical integration fomula for 1 . 2 The ruin probability with no initial reserve In this section .T) which. however. going back to Cramer. u Notes and references Proposition 1. oo (u)31/2e^e)k = )3k z cos(k9) = R k. E Fk.
T)dx.T)) 1 fT P(M(v. Stv^ _ Define M(v.T))dv E^T I(M(v. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 . 2. we define a new claim surplus process St StM NJ Figure 2.104 CHAPTER IV.T]. resp.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS.1.T))dv.0<w<t} St+v . Proof For any v E [0. T) = P(Tr(0) > T) = P(M(0. T T o where the second equality follows from II. 1 1 . and the third from the obvious fact (exchangeability properties of the Poisson process) that has the same distribution as St = { Si0)} so that P(M(v.t)= {Stv) < SM.T) T F(x. T].1 In formulas.T)) does not {Stv)} depend on v. co ).(.(0. f T lStv)} 0<t<T by a 'cyclic translation'. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. See Fig. [v.3) with A = (0.b (0.(6. ") } is at a minimum at time t. v]. Then 1 .i. .
T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .Sv. v) = M(0. w) for some small E. Obviously.T) = F(u+T. then i fT I(M(v.T) occurs. T)). letting w = inf It > 0 : St_ = mino<w<T Sw}.ST on M (0. T Theorem 2 . cf. then M(v. in which case there is a last time o where St downcrosses level u. Hence T TE f I( M(v.T)) dv f T I(M(0.t)dt. t) denote the density of F(•. 0<t<v} = {ST < St . v<t<T}n{ST<STSv+St. there exist v such that M(v. T) occurs.T) occurs or not as long as ST < 0. Fig 2. t).xdx. T) as {ST<St+ vS. this integral is 0 if STv) . v)) dv = ST T T o (note that the Lebesgue measure of the v for which {St} is at a minimum at v is exactly .Tt))f(u+t..T) and Sv < 0 on M(0. We claim that if M(0. v < t < T} n M(0. T) = M(0. v).. T) occurs. we can write M(v. ST > 0.2 10(u. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0.T)f(I z /)(0. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. . T T o i =1 Let f (•. we can take v E (w E. If ST < 0.v<t<T} = {ST<StSv. where the last equality follows from ST < St on M(0. or it occurs. For example.2. Indeed. 0<t <Tv}n{ST<ST Sv+St T+v. v). T]. v). T)) dv.2. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. T. It follows that if M(0 .
which occurs w.v.T)+ J0 T (1V. The proof is combined with the proof of Theorem 111.b(u. For a fixed T > 0.u+dt]).2.p.2 .T) = {St < 0. Proposition 2. define St = ST . PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. Hence P(ST<u) = 1 .(0.z. Proof of Theorem 111.T) = . t + dt] occurs if and only if St E [u. 0 < t <T.2 Here o.2. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •). {St > . E [t. ST_ _ z}. 0 < t < T. u which is the same as the assertion of the theorem. C*(z. {S t > z. z > 0.106 CHAPTER IV.ST_ t_ and let A(z. u + dt] and there is no upcrossing of level u after time t.t). which is independent of St and has the stationary excess distribution B0. ST_ _ z}.Tt))P(StE[u. The following representation of T(0) will be used in the next section. Let Z be a r. ST_ _ z} . O(T . 0 < t <T .3 Define r_ (z) = inf It > 0 : St = z}. 2.T) = C(z.
z + dz].2. Hence integrating (2. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. r(0) < oo) = 3R(z) dz JP(C(z.2. T + dT] I S7(o)_ E [z. u which is the assertion of Theorem 111. ST(o)_ E [z. we therefore have P(A(z. z + dz].1) z T . z + dz]) = P(A(z. z + dz]. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz.T) = C*(z. It follows by division by P(ST(o)_ E [z. Figure 2. Thus P(Sr(o)_>x.1) yields P(ST(o)_ E [z. 2. 7( 0) < oo) = P (C(z)) dT. {St }o<t<T have the same distribution . z + dz].T).3).3.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z.T)).2.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. T(0) < oo) = OR(z) dz in (2.ST(o) >y.3 But by sample path inspection (cf. (2. and since {St}o<t<T.T)) = P(Cx. A(z. Fig. Proof of Proposition 2.1) that P(T(0) E [T.T + dT].T))f3B(z) dz dT. .
2. z + dz]. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. cf.108 Hence CHAPTER IV. Let T_ (y) be defined as Proposition 2.(3(B[r( a)] . one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. who instead of the present direct proof gave two arguments. Notes and references For Theorems 2. Proof Optional stopping of the martingale I er (a) 9 t. 3 Laplace transforms Throughout in this section. see in addition to Prabhu [293] also Seal [326].3 was noted by Asmussen & Kl(ippelberg [36]. [329].5a). In the setting of general Levy processes. a martingale proof is in Delbaen & Haezendonck [103].5 and one upon excursion theory for Markov processes (see IX.1) .3. some relevant references are Shtatland [338] and Gusak & Korolyuk [181].1) where a > r.1. because of77>0. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.(yo). I L Let ga(x) be the density of the measure E[ear(°). PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T.T+dT]). Note that T_ (y) < oo a. Tak'ecs [359]. r(0) < oo.2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . Lemma 3. z + dz]. (3. Theorem 2.1 Eear( y) = eyr(a). Lemma 3 .1 and the present proof is in the spirit of Ballot theorems.r(a).2. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. T(0) < oo) = dTP(T_(Z) E [T.c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y).s.6. Proposition 2.
3.ga [b] 1 . u .ga [b] 0 TO Using Lemma 3. Corollary 3.5 f 00 o a/r(a) . E[ear (o) I T(0) < oo .2 P(Z E [y.2.3. Then by Proposition 2.f.°° ga(x)dx. £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v. the result follows after simple algebra. time T(u): u u Here is a classical result : the double m.r(a) b . Hence eb"du E[eaT(").r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .g. y + dy].x)ga (x) dx where za(u) = f. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).(v) = ev''(a).3. Z = y] = EeaT. r(u) < oo). LAPLACE TRANSFORMS 109 Proof Let Z be the surplus . Further by Theorem 111.ST(o)_ just before ruin .4 E[eaT (o). rr(0) < oo) = 1_ r(a) Proof Let b = 0.r(a) oo Q f ex(br(a))dx f00 eyr(a)B(dy) x 0 Q f evraB(dy) e(a))dx 0 Q cc ev(br (a)) .ic(b)/b x(b) + a eb"E[eaT(" ).r(a) = a [B[b] B[r(a)]] . It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.x)(a) B(dy)• Lemma 3 .3 ga[b] = c(b) Proof + b + a .1] evr(a)B(dy)[ b . b . (u . (Laplace transform) of the ruin Corollary 3.
the known results are even less explicit than for the exponential claims case. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S.1)Er(u) . uoo u using e. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 .3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). The first main result of the present section is that the value umL.h(u. P = /3µB > 1. T(u)/u mL as u + oo. By Proposition 111.1) i.mL > E T(u) < 00 ) 40.r(u) = Er(u) • ES. For the proof.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. mu ) ( 0 m < ML '(u) 1 m > rL.1. and take basically the form of approximations and inequalities..e. Later results then deal with more precise and refined versions of this statement.s. This proves the first assertion of (4. i.. note that by Wald's identity u + EC(u) = ES. That is. = (p .3).6.00 St = lim . Then given r(u) < 00. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) . and hence a. T(u) a.s. we need the following auxiliary result: Proposition 4.1 Assume 77 > 0.110 CHAPTER IV. for any c > 0 P( Further. (4.(u ) = o(u) a. for any m T(u) u . PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model.2 Assume ri < 0.mu D 2 4 N(0. Then as u * oo. Theorem 4 .w ) v/. For the second . cf. (u) t. Proposition A1. St/t 1 1/m.2.
1 (by considering 0(u. 1'r(U) .2 of [86]) and (4. and (4. According to Anscombe' s theorem (e. the result comes out not only by the present direct proof but also from any of the results in the following subsections. of (4.1). which may be viewed both as a refinement of Theorem 4. Notes and references Theorem 4.N(0.mL >E By Proposition 4.1). Tu) T( u) . Theorem 7.6.1. T(u) < oo f / 00) e7uE L [e_7 (t1). proving (4.g. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. implying T(u) .1 The l.mu (2) '• m3/2 µB 7 . cf. Thus.1 is standard. apB ) .mu m . 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. note first that ( Proposition 111.5) St .1) is T (u)  U mL P( T (u) < I > E.3.2. the same conclusion holds with t replaced by r(u).2) follows immediately from u (4. though it is not easy to attribute priority to any particular author.6µB2) Z v m (3µB2) Z. and as a timedependent version of the CramerLundberg approximation.mL U > E. For (4 .3). 4).s. T (u) < 00 J 0(u) e7'PL U \ I T u) .t/m D (2) 111 .r(u)/m T(u) ti µB2) Z. again Proposition A1. If Z .7 6  11 Proof of Theorem 4. .4. this can be rewritten as u + 1(u) . T) for T which are close to the critical value umL).1).^ N (o. PL (•)+ 0.h.
Then the distribution of T(u) .5) For the proof.6).L+YWLV'U) . letting Z be a N(0. Using ( 4.ul/4. with w2 as in (4. oo).ST( u') = u1/4 . and similarly as above we get E[f(^(u)) I Fr(u.4).e(u') oo w .6) whenever f.u1/4)I(S(u') > u1 /4) h(oo) + 0.T ( u')] = E[ T ( ul /4 . P because of ^(u') .T(u') given F. using that ul/4 . Let h(u) = E f (^(u)).l:(oo) (recall that rt < 0). Then for any y.w2) r.(u.f ( (oo)) . g are continuous and bounded on [0.3). Proof Define u' = u .mul h(oo)Eg(Z).4 (SIAM'S LEMMA) If 71 < 0. O . PROBABILITY OF RUIN IN FINITE TIME Corollary 4.112 CHAPTER IV. (oo. Hence Ef (Vu )) 9 (T(u.^(T(u')).) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u . we need the following auxiliary result: Proposition 4.v. Then h(u) 4 h(oo) = E f (6(oo)).)mu \ h(oo)Eg (r(ul) .a C4'(y )• ( 4.VU T. e'°'/b (u. we get E[ T (u) . S( u ) < ul/4] < ET(ul / 4) = O(ul/4). we can replace T(u) by r(u').um. oo ).r.3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1). then e(u) and r(u) are asymptotically independent in the sense that. E9(Z) (4. and thus in (4. one has 9 (r(u)_rnu) Ef (^(u)) * E.t.))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . resp .
7) To arrive at this . y u) < . define ay. just substitute T = umL + ywL in (4. T(u) < umL + ywL f. however . Cf. ELe7E (") . WHEN DOES RUIN OCCUR? Proof of Corollary 4.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . PL(T(u ) < umL + ywL) 113 4 C4(y).7) to be good.1.yK(ay)• (4.dependent version of Lundberg's inequality For y > 0. CL Fig.oo.(ay) = 17 7y = ay .z/)(u . also Hoglund [204].4. For practical purposes . where we used Stain's lemma in the third step and (4. in practice one would trust (4. Theorem 4.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML).7) to be valid is that T varies with u in such a way that y(T) has a limit in (.4) in the last.5) and solve for y = y(T). y u) < e 7v" . e7v" y < ^'(7) (4 . oo ) as u * oo.3 ery"z/i(u . that for the fit of (4. . The precise condition for (4.T) Ce7"4 (T . u needs to be very large). 10) '5(u) . 0. Segerdahl 's result suggests the approximation b(u. Notes and references Corollary 4 .5 '(u . y > k'(7) . see Asmussen [12] and Malinovskii [254]. For refinements of Corollary 4. umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). 4b Gerber's time.9) ( 4 . Thus . The present proof is basically that of Siegmund [342].umL wI V"U u (4.3 in terms of Edgeworth expansions . 3 is due to Segerdahl [333]. yy by 1 K. see also von Bahr [55 ] and Gut [182].
and generalizations to more general models are given in Chapter VI.1).5.ay4(u)+ T(u)K(ay ). yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . f Some urther discussion is given in XI.h(u.yu ) = eayuEav [e . the bound a7y° turns out to be rather crude . yu) is e 'Yyu/ . and hence t.3 yields easily the following sharpening of (4. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). a. Then ic(ay) > 0 (see Fig . Numerical comparisons are in Grandell [172 ]. 5 is due to Gerber [156 ]. if y > 1/ic'(y). who used a martingale argument.8).b (u. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). u Differentiating w. For a different proof. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. see MartinLM [257] .9): Proposition 4. the point is that we want to select an a which produces the largest possible exponent in the inequalities. 0. In view of Theorem 4.v"U.2. which shows that the correct rate of decay of tp(u. dy) Notes and references Theorem 4 .7 i. we arrive at the expression in (4. From the proof it is seen that this amounts to that a should maximize ayic(a).r. yy is sometimes called the timedependent Lundberg exponent.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4.8). PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y).114 CHAPTER IV. However.t.8 below . which may be understood from Theorem 4. An easy combination with the proof of Theorem 111.Y' (u. Hoglund [203] treats the renewal case. . we have rc(ay) < 0 and get (u) .6.
(4. if we want EaT(u) . (0) r1 (a) ' I. i.. the formula 0(u.11) ' If y > 1/ r . T(u) < yu] . This idea is precisely what characterizes the saddlepoint method. the choice of ay.ay y 'Yay  ay . (4. Proposition 4.e.^3 ]1/ Bay [lay . For any a > yo.ayuEay f eay^ ( u)+T(u)K(ay).2 yields EaT(u) u u r. yu) = e.z.i(u.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma. and b(u. The traditional application of the saddlepoint method is to derive approximations. As a motivation.: T.4. then the solution &y < ay of . Ea .6 with P replaced by Pay and FL by Pay.(u.'(y ). yu ) eaauEaye . then the relevant choice is precisely a = ay where y = T/u.c(&) = ic(ay) is < 0. We thereby obtain that T is 'in the center' of the Padistribution of T(u)..5.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00.e.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 .13) . (4. u 4 oo. [eT(u )K( ay). Using Lemma 111. we have ryas = ay . and in case of ruin probabilities the approach leads to the following result: Theorem 4 .ayC() .ay a.yu) c ay . T(u) suggests heuristically that l t/. it is instructive to reinspect the choice of the change of measure in the proof.. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations. not inequalities.8 If y < 1/ic'(ry). and ii(u) . then ay > 0.
a) .1) under Pay mation (4. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.1B[ay]1 ) y(ay .12) is 0 entirely similar.7ruw2 Inserting these estimates in (4. and the equation ic'(a) = 1/y is easily seen to have .9 Assume that B(x) = eay. Writing r(u) and W2 = I3ay{.ay + ayl /BLay] .13). a nr=.ay ) r.1)3 = (jB"[ay]l (Pay .c'(a) _ /3a/(8 .a)2 . (4. The proof of (4.3(5/(S .13) rigorously.I ay &y a ^c'(ay) a (1 +.c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.4). .1.13). Example 4. The difficulties in making the proof precise is in part to show (4. (ay) J0 1 K(ay )u 1 00 c2(x) dx /2 w 1 ezcp(z /( k(ay)u1 /2w)) dz /O° _ 1 1 J e Z .ay) ay +.116 CHAPTER IV. V < 01 Ir 00 er(ay)"1'2"'x eyur. i B[7ay . it seems tempting to apply the normal approxiyu + ul/2wV.11) follows. Then ic(a) = . PROBABILITY OF RUIN IN FINITE TIME ry I i . we get heuristically that Eay Ler (u)r(ay).1) .(j (1 . and in part that for the final calculation one needs a sharpened version of the CLT for t(u) (basically a local CLT with remainder term).(ay) _ y(ay . where V is normal(0.ay)K(ay) ay ayI&YI For the second term in (4.a.B[ay] /ay &y y(ay .l'B)y /(Pay .1)3 = y3/3B"[ay].
.tcp) Lo {Wo ( t)}t>0 .ay)3 0 3/2 and (4.f. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .because the c.g. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.i )( v s vc ('3 + s _2 / .= (s. is undefined for a > 5)./4 ^y for 1/i (u. then { __ .5.1) . yu) when y < 1/ic'('y) = p/1 . in discrete time: if p = ES.8 is from Arfwedson [9]. 0 Notes and references Theorem 4. A related result appears in BarndorffNielsen & Schmidli [59].. is the drift and o.1.11) gives the expression '31/4 ( .. and next to note that such an approximation in particular implies that the first passage probabilities are close. c a 00.p. 5 Diffusion approximations The idea behind the diffusion approximation is to first approximate the claim surplus process by a Brownian motion with drift by matching the two first moments.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . 2 = Var(Si ) the variance. (5.. The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o. It follows that 5^y =5ay = /«y =f3+ay=l3+d 1+1/y' V 1+^1/y /35 1+1/y /3' ay ay =Qay say =.
However.1) with S. and consider the limit p j p.2) t>o where p = pp = p .z } {W_1(t )}t>o (5. This is the regime of the diffusion approximation (note that this is just the same as for the heavy traffic approximation for infinite horizon ruin probabilities studied in III. We want an approximation of the claim surplus process itself.tp). PROBABILITY OF RUIN IN FINITE TIME where {W( (t)} is Brownian motion with drift S and variance (diffusion constant) 1 (here 2 refers to weak convergence in D = D[0. Mathematically. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes). for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.1 below). and this can be obtained under the assumption that the safety loading rt is small and positive.3) whenever c = cp f oo as p 1 p. of which a particular case is the claim surplus process (see the proof of Theorem 5.. Indeed .p. (5. p. n/c < t < (n + 1)/c. + {Wo(t ) . St = EN` U= .1.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.e. Lemma 111.1)) is inconvenient. where p is the critical premium rate APBTheorem 5 .3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}.t} _ {W_1(t)} .7c).. we have o {i!t s: . 0 . such that the claim size distribution B and the Poisson rate a are the same for all p (i. oo)).p/c < St(p) < S((n+l)/ c + Pp/c. cf.118 CHAPTER IV. Letting c = a2/pp. this is an easy consequence of (5. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p.a = Snp) and the inequalities Sn )C . a2 =/3µB2) Proof The first step is to note that { WC (St P) .1 As p J.3.
we obtain formally the approximation V.8 or [APQ] p. (5. and in fact some additional arguments are needed to justify (5. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x.6) from Theorem 5.5).f I \\\ J \ (5. 263) that the distribution IG(•. the continuity argument above does not generalize immediately. see Grandell [ 168]. C. is IG(T.Ta2 /p2).(u) ti IG(oo. 196.1..t. has a continuous distribution. . Corollary 5.h. Because of the direct argument in Chapter III.T) IG(Tp2/ a2).u). and the r. ulpI/a2).r. TS(u)=inf{t>0: WW(t)>u}. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}.1.1 I 7= .7c. we omit the details . Corollary 5 . w.h.6) This is the same as the heavy traffic approximation derived in III. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a.2 As p j p..s. However. [169] or [APQ] pp. (. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. 119 It is wellknown (Corollary XI.( ^ I + e2( \ I . u).1 . u) =PIT( (u) < x) = 1 . For practical purposes . 1.5.e. any probability measure concentrated on the continuous functions. is 1/ip (ua2 /IpI. ulpl /a2) = e2"1µl / or2.5) Note that letting T * oo in ( 5.2 suggests the approximation u 0(u. u) is defective when < 0. this implies P sup 0<t<T a 12 Stu2 /µ2 > u 4 P ( sup W_1( t) > u O<t<T But the l. (ua2 To2 op \ IPI > IG ( T . (5.h. since ti(u) has infinite horizon .s.s.4) Note that IG(.. 199. ('.
[169]. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. Theorem 5.5) and (5.g.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. . For claims with infinite variance. Then as 0 _+ 90. that 00 4090. Assume further that 039µB6 < pe. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. e.6 of [APQ]. In view of the excellent fit of the CramerLundberg approximation. we have ^A. on the premium rule involving interest.00µB6 + 0. The proof is a straightforward combination of the proof of Theorem 5. (5.t. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation. pe . 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe .6) are presented. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. in particular for large u. All material of this section can be found in these references. The picture which emerges is that the approximations are not terribly precise. pt? 4 peo. and two further standard references in the area are Grandell [168].120 CHAPTER IV. such that the Poisson rate Oe. as 0 * 00 and that the U2 are uniformly integrable w.r. We conclude this section by giving a more general triangular array version of Theorem 5. the B9. See for example Billingsley [64]. In contrast. in Asmussen [12]. in the next subsection we shall derive a refinement of (5.1. the simplicity of (5. as an example of such a generalization we mention the paper [129] by Emanuel et al. the claim size distribution B9 and the premium rate p9 depends on 0.5) combined with the fact that finite horizon ruin probabilities are so hard to deal with even for the compound Poisson model makes this approximation more appealing.Pe.1 and Section VIII. However. However.Po = 09µB6 . The first application in risk theory is Iglehart [207]. Furrer..6) therefore does not appear to of much practical relevance for the compound Poisson model. B0 * Boo.5) for the compound Poisson model which does not require much more computation. and which is much more precise.
9(s) = Ico ( s + 9) . and we are studying b(u. 9o T 0.Q (B[s] ./c(9 . CORRECTED DIFFUSION APPROXIMATIONS 121 6 Corrected diffusion approximations The idea behind the simple diffusion approximation is to replace the risk process by a Brownian motion (by fitting the two first moments ) and use the Brownian first passage probabilities as approximation for the ruin probabilities. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o].4. .90) and the given risk process corresponds to Poo where 90 = 'yo. Let PO refer to the risk process with parameters e9oz Qo = QB[90].ao (0) _ /c(s + 9 . this means the following: 1. P9(r (u) < oo) = 1 for 9 > 0.6. The objective of the corrected diffusion approximation is to take this and other deficits into consideration. which we have seen to play an important role for example for the CramerLundberg approximation . Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). 3.T) = Peo(r(u) < T) for 90 < 0. 77 = 1/p . claim size distribution B . Determine yo > 0 by r. However . this idea ignores (among other things) the presence of the overshoot e(u). B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . this is because in the regime of the diffusion approximation .1) . For each 9. Since Brownian motion is skip free. whereas there we let the given 3B.'(yo) = 0 and let 90 = 'Yo.90) .c(s) = .6. . In terms of the given risk process with Poisson intensity . Bo(dx) = B[eo]B(dx).90] B(dx). The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III.s and p = /3µB < 1. In this setup. it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). 77 is close to zero. PB('r(u ) < oo) < 1 for 9 < 0. risk process with safety loading 77 > 0 correspond to 9 = 0 .1 > 0. Then r. 2. 0(0) = 0. and we want to consider the limit 77 10 corresponding to Oo f 0.
PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. u) denotes the distribution function of the passage time of Brownian motion {W((t)} with unit variance and drift C from level 0 to level u > 0. tu2 ) i IG (t.3) this implies (take u = 1) Ego exp { . .. u) = euh(a . C) = 2A + (2 . (01. i. IGu+u2. 9otc0" (0) = 0061 = ul.. The first step in the derivation is to note that µ = k (0) = r0 (00) .1) . One has (6.1) IG(x. and Si = QoEoU2 = Q B"'['Yo Eo U3 ].T) 1+u2 (6.(y) = 0. (. 1) • Since L eatIG (dt.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process.e.122 CHAPTER IV. write r = T(u). The corrected diffusion approximation to be derived is (u. (. Vargo S. Theorem 5. ..S. means up to o(u1) terms): . S2 = 3E0U2 Bier [Yo] 3B"[Yo] Write the initial reserve u for the given risk process as u = C/Oo ( note that C < 0) and. () where h (A. C .u. (6. u) = IG(x/u2. (U. _ ^(u) = ST .() The idea of the proof is to improve upon this by an O (u1) term (in the following..C.3 applies and yields 1061 U61 Stdlu2/CZdi {W_1(t)}t>0 t>0 which easily leads to 1 StU2 {W( J(t)1t>0 { u S1 t>o Y'(u. bl IG(t81. 0o to. C.2) .7(u)/u2} eh(A.Varo S1 = f30Eo U2 = S1. for brevity. the solution of r.
that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . just replace t by Tb1/u2. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r.5) Once this is established . In ( 1) and (2). however .v. 6 .'yu /2)(1 + b2/u)} + Aug 1I J . p = 0.2).4.1 + 629. in (3) and (4). But the Laplace transform of such a r. of a (defective) r.z .f.6. 1. .1 + u2 I Indeed. .3 = 0. 9o T 0 in such as way that C = Sou is fixed. distributed as Z .exp { h(A. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.7.1 As u + oo. however.v. is the c. A numerical illustration is given in Fig. the r. which is based upon exponential claims with mean µB = 1. calculated using numerical integration and Proposition 1.d. . . The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3).ry2 . of (6. we have p =. bl I IG I t +2 .3). 1% in (2) and (4). it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} .s.5) according to (6. we get by formal Laplace transform inversion that C 2 u.1 below is exact.2 ). CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.3.h. The solid line represents the exact value . (6.52/u where Z has distribution IG (•. The justification for the procedure is the wonderful numerical fit which has been found in numerical examples and which for a small or moderate safety loading 77 is by far the best amoung the various available approximations [note. To arrive at (6. Note.h.s. and the dotted line the corrected diffusion approximation (6. that whereas the proof of Proposition 6.2) is indeed o(u1).
The proof of Proposition 6.011 L1 60 T IM 11.0 0.1 W IU.T1 00.1 It is seen that the numerical fit is extraordinary for p = 0.^) .1 proceeds in several steps.u2 2u3 (e . PROBABILITY OF RUIN IN FINITE TIME 0. For further numerical illustrations.05{ 0.00 0. see Asmussen [12].08 a.19)2 11 20 20 i0 T 1n0 Figure 6.() Lemma 6.2 e..02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.01 0.07 0. .199 0.7.EB 0 p ex p ( 7 S h ^)u .TI CHAPTER IV. A51 7(SAT 3 3 h(X. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.OOIi O. OM 0. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].114 0. (Inc 0s 0.(061 0.aa1 .T) 111 0.111 W(U. the fit at p = 0.124 0. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p..4 may not be outstanding but nevertheless.T) 0.7 or at values of Vi(u) like 1% is unsatisfying. and all of the numerical studies the author knows of indicate that its fit at p = 0. Similarly.08 0.W21 0.
co ((/u)) } Let 8 = (2a + (2)1/2 = h(). C) 1 1 + u2/ 111 + 2u CZ Z  (2A + ()1/2 J 1 Proof It follows by a suitable variant of Stam's lemma (Proposition 4.7) 2 2 .6) u U3 Lemma 6 .61a2T (B3 .co (e) .(3)Eea LauT exp i 3J .4) that the r. exp ue } al 1J 3 exP I [2). + a1b2 + ..2u (B3 .h. the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .1) h(A. () 62 Eeo exp u u2 J .00)(u +C)  'r (.. 1 = PB(T < oo) = Eo0 exp 125 {(B . (6.r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .(3) J t _ aa1T l + eh(A. CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.4 Ea.2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .6.C2 = 2). in Lemma 6... () + C and note that 2 KO (0) = 102.C)C/u .T (co (8/u) .+ h (A.s. 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration . the result follows. 1 / Po(C(0) > y) dy EoC(0) x k EDUk + 1 k Eo[(0)k+1 EoC(0) _ (k + 1)EoU' EoC(^) _ (k + 1) Eo£(0) Lemma 6 . (6.
h (A. yu/2). l Lemma 6 .4.6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4).h.() .\+ (2 (3 e 2u [ (2.x. C) ( 1+ u2 The result follows by combining Lemma 6 . yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. yields +90 62 0 + O(u 3) 2u2 +O(u 3). () . letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). yu/2) h(A.S) d e 62 .(2A + ()1/21 exp S h(A.126 CHAPTER IV.6  d h(A.1 (y/2 + Oo)u .2u [2A+ (2 3 .s.e h(aS)h (^^ 262 exp {_h(. Thus by Taylor expansion around ( = 90u. PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.2 (^/2 + 3y9o + 390) + O(u3). () by h(\. and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6.7) and using eh(a.() I 1 + u2 ) y . 5 exp { _h(A) (1 + / y u J)) exp 1. and inserting this and 9o 2 = S/u on the r. Thus a2 y = 290 + O (u2). 2 + 00 = . we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6. 2 and (6.2.2) for O(u) (indeed .\ + () 1 2 / . There are two reasons for this : in this way. 0 The last step is to replace h(A. [2+ (2 . we get h(A..
The adaptation to risk theory has not been carried out. the 'typical' value (say in sense of the conditional mean) was umL. ()} 3 h (A. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin.1: Just insert Lemma 6. with the translation to risk processes being carried out by the author [12]. In Siegmund's book [346]. () I 1 + u 2 ) } S 1 . . His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. ()} . The answer is similar: the process behaved as if it changed its whole distribution to FL. the approach to the finite horizon case is in part different and uses local central limit theorems. i. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. HOW DOES RUIN OCCUR? exp { h (x.4. that is. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting.(i+ 62 exP{ h(A. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. the analogous analysis of finite horizon ruin probabilities O(u. and to the Markovmodulated model of Chapter VI in Asmussen [16].1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. () (i+a ) 2A + (2 . Fuh [148] considers the closely related case of discrete time Markov additive processes.e. this case is in part simpler than the general random walk case because the ladder height distribution G+ can be found explicitly (as pBo) which avoids the numerical integration involving characteristic functions which was used in [345] to determine the constants.5 in Lemma 6. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs.T) has not been carried out and seems nontrivial. the same as for the unconditional Lundberg process. 'yu/2) 127 ( i+ M pz^ exP { h (A. 0 1 Proof of Proposition 6.7. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments.
. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . r(u) < oo) .F. {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u). we give a typical application of Theorem 7. u * oo. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL. ^(u) is exponential with rate 8 w.(u)_ is that i. + TMOO ).r. F(u)c] ti e' ru]PL (F(u)`) > 0.(u)_ and similarly the denominator is exactly equal to Ce7u.. then P(u) and FL coincide on .3L and the claim size distribution from B to BL.. Note that basically the difference between FT(u) and . so in the in the proof.(u)_ and ^(u) are independent .t. Proof P(u) (F(u)c) = F(flu)c.128 CHAPTER IV. .t. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7].T.F.TT(u) _measurable.vi(u) Ce'Yu Corollary 7. stating roughly that under F(u).ST(u)}t> o is just an independent copy of {St}t>o).EL[e7S. Recall that . In the exponential case. .1. the Poisson rate changes from . P(u) and rate = aL w. Theorem 7 . the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . F(u)c] P(r(u) < oo) ?P(U) < EL[e7u.FT(u) = o' (T(u ). (u) and satifying PL(F(u)) * 1. We are concerned with describing the F(u) distribution of {St}o<t<T(u) (note that the behaviour after rr(u) is trivial: by the strong Markov property.(u).(u) is not . Then also P(u)(F(u)) + 1. {ST(u)+t .2 If B is exponential. In fact. FL As example.FT(u) is the stopping time oalgebra carrying all relevant information about r(u) and {St}o<t<T(u)• Define P(u) = P(•IT(u) < oo) as the distribution of the risk process given ruin with initial reserve u.1 Let {F(u)}u>0 be any family of events with F(u) E F.3 to .r.
7.eaLx. the subject treated in this section leads into the area of large deviations theory.3. take I(Tk < x) . however. see further XI. From a mathematical point of view. A somewhat similar study was carried out in the queueing setting by Anantharam [6]. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). who also treated the heavytailed case. HOW DOES RUIN OCCUR? Corollary 7. Proof For the first assertion.(1 . the queueing results are of a somewhat different type because of the presence of reflection at 0.eALx) M(u) k=1 u The proof of the second is similar. This is currently a very active area of research.3 M(u) pcu) 1 . . Notes and references The results of the present section are part of a more general study carried out by the author [11].
This page is intentionally left blank .
. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). see A. U2. Proposition 1.Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. AA t*oo lim St = lim ESt t tioo t = p . In the socalled zerodelayed case. the one corresponding to the stationary case by 00)(u).1.. {St} is the claim surplus process given by I.7). A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. r(u) the time to ruin.d. with Nt = # {n: Un <t} the number of arrivals before t... and the one corresponding to T1 = s by 1/i8 (u).1 (T1 = a1). the Tn are independent. .. D'2. We use much of the same notation as in Chapter I. . the distribution Al of T1 is A as well. of the risk process form a renewal process: letting Tn = Qn . .1 Define p = !µ. are i.Q. the claim sizes U1. T3. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . Thus the premium rate is 1. with common distribution B.(1. with the same distribution A (say) for T2...i. . and M is the maximum of {St}.1). Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory.Then no matter the distribution Al of T1i B.
2 (DETERMINISTIC ARRIVALS) If A is degenerate. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U.1 gives the desired interpretation of the constant p as the expected claims per unit time. s + t µA PA 0 Of course. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). by the elementary renewal theorem (cf.St] = a(p . after E.Nt] * a/PA. 2). t 4oo Proof Obviously. CHAPTER V. Nt + EVar U. the definition 77 = 1/p . but the arrival rate in the ON state is .1 of the safety loading appears reasonable here as well.t. such that no arrivals occur in the off state. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. Nt ESt = E E UI Nt t = ENt•pB . 3) follows similarly by Blackwell 's renewal theorem. Example 1 . Thus. and (1 .0 > 0. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). Here are two special cases of the renewal model with a similar direct interpretation: Example 1. OFF.1) ENt/t + 1/µA. However . one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. say at a. If the environment is Markovian with transition rate A from on to off and u from OFF to ON. A.1).3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. For (1 . Proposition 1. RENEWAL ARRIVALS lim E [St+a .132 Furthermore for any a > 0. stating that E[Nt+a . The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. The renewal model is often referedd to as the Sparre Andersen process. From this ( 1. the .a is really the accumulated claims over a period u of length a.1) follows .
(1. Ao.4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0.1. . the fundamental connections to the theory of queues and random walks. the relevance of the model has been questioned repeatedly. However.s < u). the first term represents the probability F(U1 .t. and then the whole process repeats itself). initial vector (1 0) and phase generator 11 However. The values of the claim surplus process just after claims has the same distri bution as {Snd^ }• Since the claim surplus process {St} decreases in between max St = max ^d^. Therefore.r.y)B(dy). we have From this the result immediately follows.1.4) w. we feel it reasonable to present at least some basic features of the model. we note that the ruin probabilities for the delayed case T1 = s can be expressed as in terms of the ones for the zerodelayed case as u+8 z/i8(u) = B(u + s) + '( u + s .1.T between a claim U and an interarrival time T.. S o<t<oo n=0.s. integrate (1. if for nothing else then for the mathematical elegance of the subject...2. For the stationary case. as follows easily by noting that the evolution of the risk process after time s is that of a renewal risk model with initial reserve U1 . u For later use. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. Proof The essence of the argument is that ruin can only occur at claim times.4) with phase space {oN. and for historical reasons. A is phasetype (Example 1.. More precisely. in general the mechanism generating a renewal arrival process appears much harder to understand.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo..4) fo Indeed. arrival times. U1 . and the present author agrees to a large extent to this criticism.d.oFF}. INTRODUCTION 133 interarrival times become i. (an arrival occurs necessarily in the ON state..i.} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .
1) +ry. the claims and the premium rate are negative so that the risk reserve process .a*PB• > 1. (2.1) . That is . < 1. 2.3* pB. the remaining part of the prepayment (if any ) is made available to the company. A typical sample path of {Rt } is illustrated in Fig. 00). The initial reserve is obtained by prepayments from the policy holders. RENEWAL ARRIVALS 2 Exponential claims.3* (say ) and the U. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. the claim surplus process are given by Nt Nt Rt = u+^U.1 r* (u) One situation where this model is argued to be relevant is life annuities.Ut. then 0 * (u) = 1 for all u > 0.1 If. are independent of {Nt} and i. Theorem 2 . with common distribution B* (say) concentrated on (0. St = t . U Figure 2. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . i.1. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.d. each of which receive a payment at constant rate during the lifetime . resp . If . t. b=1 !=1 where {Nt } is a Poisson process with rate .134 CHAPTER V. At the time of death . Using Lundberg conjugation .0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) .
2.3*. > 1 . B* [7] and let {St} be a compound Poisson risk process with parameters . of {St} is c(a) = is*(a7). and thus 1 = P(T.f. T_ (u) = inf { t > 0 : St = u 'r* (u). Then the function k* is defined on the whole of (oo.(u ) < oo) = E {e7sr_ (u). cf.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. B.. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r.2(a).0. . Since ic'(0) < 0. 2.g.2(b). 0 Now return to the renewal model.UB. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *.s.2 Assume now . 2. Define T_ (u) = inf It > 0 : St = u} . Let B(dx) = ^e7x B*(dx).* (a) = log Ee'st I. If I3*pB* < 1.1. St=Rtu=St. then by Proposition 111. the safety loading of { St} is > 0.(a) 7 Figure 2. Then the c. B*. Hence T_(u) < oo a. Hence y exists and is unique.Rt. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). (a) is*(a) (b) . 0) and has typically the shape on Fig. and the Lundberg conjugate of {St} is { St } and vice versa. Proof Define 135 St =u . Fig.
. To + M(d) in the notation of Proposition 1. and from Fig .•. u Hence P(M(d) > u) _ 1r+e'r"..1.e..Tn} n=0...7r+ 7r EeTo b/(Sa) + +.4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value.)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[.. and hence the failure rate ..u+ and lr+. To + max {Ul+•••+UnTI.1 it is seen that ruin is equivalent to one of these values being > u. and (2 .f. 1) means that 8(A[ry] . However. with rate S (say).Y a I.+Tn U1 Un. A variant of the last part of the proof.Tr+. we get Ee'M(d) = Ee°M* _ Y/(.• • • . with the probability that a particular jump time is not followed by any later maximum values being 1 .Un } = max St = t>0 n=0. Taking m. According to Theorem 2.2). respectively. the failure rate of this process is y. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.1. T2 = U2. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 .. RENEWAL ARRIVALS Theorem 2 ..1) + ry = 0 which is easily seen to be the same as (2. Then B* = A. 3* = 6.2) 7 and7r+=1Proof We can couple the renewal model { St} and the compound Poisson model {St*} with negative claims in such a way the interarrival times of { St*} are To .2 If B is exponential...g.1. then ...136 CHAPTER V.a) = 1 . Hence M* max {To + Ti + • • • + Tn .1 means that M* is exponentially distributed with rate ry. Now the value of {St*} just before the nth claim is To +T1* +.Ui .'s and noting that V)*(u) = P(M* > u) so that Theorem 2. 2.Y] (2.4.. alternatively termination occurs at a jump time (having rate 8).Ti = U1. and 5PA > 1.
B(dx). Putting this equal to y. letting P(d) refer to the renewal risk model with these changed parameters . 0 3 Change of measure via exponential families We shall discuss two points of view.. B^d) where Aad> (dt) = ^[ a] A(dt). This follow since.5. 111.y/b.. the imbedded discrete time random walk and Markov additive processes. Furthermore. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +.7r+).. Bads (dx) = .3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. 3a The imbedded random walk The key steps have already been carried out in Corollary 11. which states that for a given a.T to F(d)(x) = eK^d^(«) ^x e"vFidi(dy) 00 K(d) (a) = log F(d) [a] = log B[a] + log A[a] .7r+) and hence r+ = 1. : S(d) > u} .3.2. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+.2. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. hence exponential with rate b. The probability that the first ladder step is finite is 7r+. It only remains to note that this change of measure can be achieved by changing the interarrival distribution A and the service time distribution B to Aad^.4. we see that ry = 6(1. a ladder step is the overshoot of a claim size.. resp.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U . However.6. Hence the failure rate of M(') is 6(1 .
T is nonlattice. (b) V)(u) . E(d)e 1' (u). For claim (b). ik. In fact. and claim (a) follows immediately from this and e (u) > 0.2 In the zerodelayed case.C8e7u where Cs = Ce78B[7]. Proof Proposition 3.4. just note that F7(d) is nonlattice when F is so . Corollary 3.138 CHAPTER V.. VIII. cf.p)/($B'[7] . (a) '(u) < eryu.1) is explicit given 7.2 p.3 For the delayed case Tl = s.t. 187) and thereby for ^(u) to be nonlattice w.1 For any a such that k(d)' (a) > 0. i . For the stationary case. It should be noted that the computation of the CramerLundberg constant C is much more complicated for the renewal case than for the compound Poisson case where C = (1 ..u the overshoot . O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . let 7 > 0 be the solution of r. Consider now the Lundberg case. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.Ce"u where C = limu.C(°)eryu where C(O) = C0[7] . in the easiest nonexponential case where B is phasetype. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.1). 7µA .r. provided the distribution F of U .(u) .e. 00)(u) .1 implies Cu) = e«uE ( 7d)e«^(u) . (d) (7) _ 0. we get: Proposition 3.. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3.
E8h0 (Jdt.(s. Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . 3b Markov additive processes We take the Markov additive point of view of II.St)} can be defined by taking Jt as the residual time until the next arrival. The expressions are slightly more complicated and we omit the details.a . 0) = tc(a)h(s).dt ) eadt = h ( s) . 0) = ah (s) . we look for a function h(s) and a k (both depending on a) such that Gh.3. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. According to Remark 11.1) = C(O).5.(°) ( Ce8B[7] Ao(ds) similar manner. another use of dominated convergence combined with Ao[s] = (A[s] 1)/SPA yields 00 u) e7u iP8(u) Ao(ds) + f 0 = CB['Y](A[y] . we invoke the behavior at the 1 = h«(0.h'(s)..y) B(dy) 0 For the stationary case. Here K. Sdt) = h(s . For s > 0. (s.4).9. (u + s . 0 0 . h(s) = e(a +x( a))8 (3. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v).. St)} and h. B(x) = o(e7x) and dominated convergence. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. where G is the infinitesimal generator of {Xt} = {(Jt. delayed version of Lundberg's inequality can be obtained in a e7u.dt(ah ( s) + h'(s)) so that Gha ( s. Let P8f E8 refer to the case Jo = s.5. y) = e°yh(s).Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds).0 ) = Eo[ha ( Jdt. To determine boundary 0.1) (normalizing by h(0) = 1)./c. IPA 0 Of course.
resp.4 The probability measure Pa. .e. .s governing {(Jt. . Further. Ba(dx) = B(dx).rc(a)] = B = Ba[13]Aa[5]. since JT. Proposition 3.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east . RENEWAL ARRIVALS B[a]A[a .( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. .140 CHAPTER V. we can now for each a define a new probability measure Pa. [a + /3] A[b . Aa as asserted .. .2) As in 11. An easy extension of the argument shows that U1. rc(a) = 0 (B[a] . T2. A[a .. ] = E.2). Remark 3 .. 5 For the compound Poisson case where A is exponential with rate .s(Jo = s) = 1 follows trivially from Lo = 1.13]A[b . (3.e(«+k(a))t esy A(dt).a .tK( a)e. [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +.rc(a)] B[a] A[a .a . . J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk.s and P(d).. i. T2 are independent with distributions Ba.. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same.c(a)] B[a] Proof Pa. = J8 = T2.8.2) means 1 = B[a]/3/(/3+a+rc (a)). (3.s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa.5.rc(a)] = 1.. U. An important exception is.c(a)] which shows that U1. however.1)a in agreement u with Chapter III. Note that the changed distributions of A and B are in general not the same for Pa. resp. Ba where Aa (dt) . Ease AU1+6T2 [ AU1+6T2 = Ea a LT. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J.
(u. for the zerodelayed case zp8(u.4.5 Proposition 3. [APQ] Ch. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. defined as the single server queue with first in first out (FIFO. the amount of . . which is the same as the asserted inequality for 0. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. Then J(rr(u)) = TM(u)+1 and hence Ws(u.)+1 e J j e(ay+w(ay))8 e . The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). Let M(u) be the number of claims leading to ruin .1 and n. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue.r. or FCFS = first come first served) queueing discipline and renewal interarrival times. .4.t. Using the Markov additive approach yields for example the following analogue of Theorem IV..ay+ray))TM(.5. For the approach via Markov additive processes.4. u The approach via the embedded random walk is standard. Then "^ e(ay+w(aY))8 Ys(u. The claim for the zerodelayed case follows by integration w. that is. that is. yu ) < e7yu. . XII.. yu) = F'ay. Proof As in the proof of Theorem IV.s eaysr(")+r(u ) K(ay) h (s) . 2. THE DUALITY WITH QUEUEING THEORY 141 The Markov additive point of view is relevant when studying problems which cannot be reduced to the imbedded random walk.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. yu). Label the customers 1. not after time T.g. and define yy = ay .yu ) e7vu A[ay . say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. see e. it is easily seen that ic(ay ) > 0. and assume that T„ is the time between the arrivals of customers n . A(ds).yx(ay). The virtual waiting time Vt at time t is the residual amount of work at time t. the time from he arrives to the queue till he starts service.y yuAa y [ay + K(ay) . and U„ the service time of customer n.
Tn)+. and combining with (4. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.2 Let Mnd) = maxk=o. (4.2. the waiting time a customer would have if he arrived at time t. on + Tn) the residual work decreases linearly until possibly zero is hit. Thus. RENEWAL ARRIVALS time the server will have to work until the system is empty provided no new customers arrive (for this reason often the term workload process is used) or.3.4. equivalently.2) (b) as t * oo. If W1 = 0. (u). Then: (a) as n + oo.. whereas in [On . Let the T there be the random time UN. then Wn v M.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ . we get: Corollary 4. but we shall present a slightly different proof via the duality result given in Theorem II. we have Wn = Van(left limit). (4. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims.. and we have P(W > u) = V. an+1) = [on.. equivalently. The next result summarizes the fundamental duality relations between the steadystate behaviour of the queue and the ruin probabilities (part (a) was essentially derived already in 11. Wn converges i n distribution to a random variable W. in which case {V} remains at zero until time on+1.1 and Corollary 11.142 CHAPTER V."^ Vi(N) (u). the proposition follows.1) The following result shows that {Wn} is a Lindley process in the sense of II. . It then jumps to VQ„ .4): Proposition 4. 0 Applying Theorem 11. The traffic intensity of the queue is p = EU/ET.1). Thus Vos}1 _ = (Wn + Un . and obviously z/'(u) = limN.4..• • • Tk ). (4.4: Proposition 4. since customer n arrives at time on.1 Wn+1 = (Wn + U.n1 (U1 +• • •+Uk Tl .3) Proof Part (a) is contained in Theorem 11.+ Un.1. but interchanging the set . Vt converges in distribution to a random variable V. and we have P(V > u) = ?/iiol(u).4. p < 1.3 Assume rl > 0 or.1.
Letting n oo in Corollary 4. which implies the convergence in distribution and (4.5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x).T* are independent and distributed as U1. Then the arrivals of {Rt} in [0..Ao in (b).4.4 The steadystate actual waiting time W has the same distribution as M(d). but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival.. and we get: .2). For part (b). Corollary 4. T1 .T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!).4) Tlim F(s) (VT > u) = limo P(s) (r(u) < T) = '+^io) (u)• 0 It should be noted that this argument only establishes the convergence in distribution subject to certain initial conditions. Ti) and similarly for the U. as WN. we obtain: Corollary 4.. THE DUALITY WITH QUEUEING THEORY 143 (T1. (4. Then K(x) = J x00K(x . cf. In fact .. conditioning upon U* .T*)+ < x) = P(W + U* . we let T be deterministic . A. K(x) = P(W < x)..e. T] form a stationary renewal process with interarrival distribution A.. T1.1. we get W = (W + U* .oo in Proposition 4.le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before .. namely W1 = 0 in (a) and Vo = 0. (4.. { Zt}o<t < T has the same distribution as the leftcontinuous version of the virtual waiting time process so that P(s)(VT > u) = P(s)(r(u) < T).(N)(u) has the limit tp(u) for all u.T* = y yields K(x) = P ((W + U* . resp . convergence in distribution hold for arbitrary initial conditions . However. where U*.T*)+. i.2. by an obvious reversibility argument this does not affect the distribution .5) Proof Letting n . x > 0. Hence for x > 0. hence (since the residual lifetime at 0 and the age at T have the same distribution . It follows that P(WN > u) =. Then the corresponding queue is M/G/1..y)F(dy). u Now return to the Poisson case . TN) with (TN.
g. Some early classical papers are Smith [350] and Lindley [246]. Proof For the Poisson case. Note that (4.6 For the M/G/1 queue with p < 1. The equation (4. VIII). Asmussen [24] and references there. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e .g. RENEWAL ARRIVALS Corollary 4.5) looks like the convolution equation K = F * K but is not the same (one would need (4.144 CHAPTER V. . the actual and the virtual waiting time have the same distribution in the steady state. the zerodelayed and the stationary renewal processes are identical. despite the fact that the extension from M/G/1 is of equally doubtful relevance as we argued in Section 1 to be the case in risk theory. That is. Hence '(u) = Ali(°)(u).5) is in fact a homogeneous WienerHopf equation. see e. Cohen [88] or [APQ] Ch. W v V. implying P(W > u) = P(V > u) for all u.5) to hold for all x E R and not just x > 0).
dv.f pi. i=1 0 and r(u) = inf It > 0: St > u}. N. 145 Oj( u. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). {Jt} describes the environmental conditions for the risk process. • The premium rate when Jt = i is pi. here it exists whenever A is irreducible which is assumed throughout. The intensity matrix governing {Jt} is denoted by A = (A. Thus.(3i when Jt = i.. t St = E Ui . Ire = 1. {St} denotes the claim surplus process. and can be computed as the positive solution of WA = 0.Chapter VI Risk theory in a Markovian environment 1 Model and examples We assume that arrivals are not homogeneous in time but determined by a Markov process {Jt}0<t<oo with a finite state space E as follows: • The arrival intensity is . • Claims arriving when Jt = i have distribution Bi.)iJEE and its stationary limiting distribution by lr. . As in Chapter I.T) = Pi (T(u) < T). M = supt>o St.
t(i) = T(')e are the exit rates. f3i and claim size distributions Bn. respectively. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E(').1 Consider car insurance. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. which is clearly unrealistic. Example 1. r^ = P (1.146 CHAPTER VI. leading to E having two states n.2. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. According to Theorem A5.a('). Example 1 . Cl The versatility of the model in terms of incorporating (or at least approximating) many phenomena which look very different or more complicated at a first sight goes in fact much further: Example 1. in blockpartitioned form. /3 = Nn when j E E(n).14. We let p Pi = /ji/AB. the operational time argument given in Example 1. u . a(i). we can approximate A(i) with a phasetype distribution (cf. we shall assume that pi = 1. Proposition 1. say. cf. this is no restriction when studying infinite horizon ruin probabilities. say. An example of how such a mechanism could be relevant in risk theory follows. For example. cf. Unless otherwise stated. we could distinguish between normal and icy road conditions. = iii when j E E(i). Bi.. and we have f3.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential.11 below. assume that the sojourn time in the icy state has a more general distribution A(i). i and corresponding arrival intensities Qn. Then the state space for the environment is the disjoint union of E(n) and E(i). T(=)). and assume that weather conditions play a major role for the occurence of accidents.4) with representation (E(i).1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. and p is the overall average amount of claims per unit time. Thus.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. one expects that 3i > on and presumably also that Bn # Bi.5 below.T(n)). P = E 7riPi. with rates Ani and Ain.
. u Example 1. let T 9(T) = f pi.Q. One way to model this would be to take A(') to be Coxian (cf. say. The simplest model for the arrival intensity amounts to .j = .3i/pi.>. w.. i ) : n E H. dt.4) with states i1. Indeed. one could for example have H = {i1.n.. is the probability that a long icy period is followed by a short normal one. Approximating each A('?) by a phasetype distribution with representation (E('l). 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. . u From now on. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L).3.. depending only on 77.1. In the car insurance example..T(n)). T(1) +w11t(1)a(1) w12t (1)a(2) w21t(2)a(1) w1gt(1)a(9) w2gt ( 2)a(q) T(2) +w22t( 2)a(2) A = wg1t(9)a(1) wg2t(9)a(2) .. St = SB=(t).a(n). Qi = . T(9) +wggt(9)0. 1 .tEH is a transition matrix.. iq (visited in that order) and letfOil >. but assume now that the arrival intensity changes during the icy period.J017. say it is larger initially. (9) where q = CHI.3 Consider again the alternating renewal model for car insurance in Example 1. t(n) = T("i)e.3i. i8f n1. A('^). MODEL AND EXAMPLES 147 Example 1 . This amounts to a family (A(")) ?CH Of sojourn time distributions. u Example 1 ..5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. the state space E for the environment is { ('q. resp.1. and . Example VIII. and 1/ii(u) = t/ii(u). the parameters are ^ij = aid/pi. . where W = (w. i E E(n) }. it = Jel(t)..2. such that a sojourn time of type rt is followed by one of type c w... n8}.p.. 4 (SEMIMARKOVIAN ENVIRONMENT) Dependence between the length of an icy period and the following normal one (and vice versa) can be modelled by semiMarkov structure. and similarly for the normal period. Then for example wi.
o = 0. JT1 = j) = Qj • e. the empirical distribution of the claims is B*. one can associate in a natural way a standard Poisson one by averaging over the environment.(Qi)diag)• More precisely. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1. dx. we put )3* = E 7fi/3i. Proof The result immediately follows by noting that T1 is obtained as the lifelength of {Jt} killed at the time of the first arrival and that the exit rate obviu ously is f3j in state j. t l=1 Note that the last statement of the proposition just means that in the limit. B* = 1 /^* Bi. Note also that (as the proof shows) 7ri/3i//3* gives the proportion of the claims which are of type i (arrive in state i). .6 The claim surplus process {St} of a risk process in a Markovian environment is a Markov additive process corresponding to the parameters µi = pi. Nt Nt a .(3iBi(dx). vi(dx) = . )3*.A . More precisely. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. A remark which is fundamental for much of the intuition on the model consists in noting that to each risk process in a Markovian environment. qij = 0 in the notation of Chapter 11.5.e(A(Oi)d'sg)xe. Next we note a semiMarkov structure of the arrival process: Proposition 1. Pi (Ti E dx. . In particular.7 The Pidistribution of T1 is phasetype with representation (ei. The key property for much of the analysis presented below is the following immediate observation: Proposition 1.8 As t oo.148 CHAPTER VI. N > 1(Ul < x) a4 B*(x). MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion.
Proof According to Proposition 1. oo) as a 4 oo. Conditioning . Nt a' t t iEE Also. i. In particular. Proposition 1.(/3i)aiag). N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. By Proposition A5. MODEL AND EXAMPLES 149 Proof Let ti = f1 I(JJ = i) ds be the time spent in state i up to time t and Nti) the number of claim arrivals in state i . the limiting distribution of the first claim size U1 is B*.2.7. and let {St °i} refer to the one with parameters Pi. {St} to the compound Poisson model with parameters 0 *.6. The next result shows that we can think of the averaged compound Poisson risk model as the limit of the Markovmodulated one obtained by speeding up the Markovmodulation. given {Jt}0<t<0. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. ^j 7riNi. zli( (u) .9 Consider a Markovmodulated risk process {St} with param eters Ni.. this converges to the exponential distribution with rate 0* as a * oo. cf. Bi.4. Then {St)} + {St*} in D[0. y Ni) i Nti) t a. aA. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti... Then it is standard that ti lt '4' iri as t > oo. Bi(x).aA .* (u) for all u. Hence Nt'> a .x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). Bi. In particular. has distribution (7ri)3i //3*)iEE and is independent of Ti.. A. and furthermore in the limit JT. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size.1. denoting the sizes of the claims arriving in state i by U(') 1 standard law of large numbers yields U(') the N 1: I(Ukik < x) k=1 N a$. However . B*.. Example 11. e.
. and (at least when a is small such that state changes of the environment are infrequent). U2) are independent of . we may imagine that we have two types of claims such that the claim size distributions are E3 and E7. T) + ?P* (u. 1. oo). Claims of type E3 arrive with intensity 2 .2. thick.=1. with T2 being exponential with rate .l3* and U2 having distribution B*.s = o in state 1 and with intensity 1 .150 CHAPTER VI. is as in {St }. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier)... The fact that indeed 0(a) (U) 3 0* (u) follows. B1=3E3+2E7. e. from Theorem 3.1 with periods with positive drift alternating with periods with negative drift.. s 5 in state 2. U. which also yield O(a) (u. 9 .10 Let E_{1..1 of [145]. That is.g. there are p = 2 background states of {Jt}.2 +2 2 = 3. resp.2}. lines in the path of {St}. shows similarly that in the limit (T2.2. MARKOVIAN ENVIRONMENT upon FT.). > 1. Continuing in this manner shows that the limiting distribution of (T. 1. 0 Example 1.. marked by thin. state 1 appears as more dangerous than state 2. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. and in fact P1 = 31AB1 = 9 3 1 2 (5 3 3 1 1 2 1 5 7 1 81 70 ' _ 19 4 5 P2 = . the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. the company even suffers an average loss. T) for all u and T.s = 1o in state 2.31µB 2 = 2 5 3 7 70 Thus in state 1 where p. From this the convergence in distribution follows by general facts on weak convergence in D[0. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. Computing the parameters of the averaged compound Poisson model. we first get that 3 (3* = 2.1. those of type E7 with intensity z s = 5 in state 1 and with intensity z . B2=1E3+4E7. 132=2. On Fig.FT.. Thus... the paths of the surplus process will exhibit the type of behaviour in Fig.
8.1 a.11 (a ) ESt/t * p .. MODEL AND EXAMPLES 151 Figure 1.1). 0 The definition (1. iEE . Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. (b) St/t * p . a fraction r.1) of the safety loading is (as for the renewal model in Chapter V) based upon an asymptotic consideration given by the following result: Proposition 1. For (b). That is.1.s. t + oo.(3. Proof In the notation of Proposition 1. the averaged compound Poisson model is the same as in III. 01 /.1.1 Thus. note first that EN Uk')/N a4' µgi.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. t * oo. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7. = P. we have E[St + t I (t(i))iE EI = E t(i)OW = iEE t(i)Pi• iEE Taking expectations and using the well known fact Et(i)/t * 7ri yields (a).
[302]. Proof The case 77 < 0 is trivial since then the a. X3. 136 or A. Now obviously the w. s.. and hence 1/ii(u ) = 1 for all i and u. MARKOVIAN ENVIRONMENT Corollary 1. and a more comprehensive treatment in Asmussen [16]. n n Thus {SWn l is a discrete time random walk with mean zero. and hence M = 00.0i(u) < 1 for all i and u. Since the X„ are independent .. and so on. see the Notes to Section 7. PB. There seems still to be more to be done in this area. and hence oscillates between 0o and oo so that also here M = oo. and involves a version of the .1 and the Corollary are standard. limit p . In risk theory. X2 =SW2 So. X 1 =Sty. Eiw. 38) Eiw1 = 1/ir. and .. Then by standard Markov process formulas (e. [212]. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. and hence wn /n a4.a form a renewal process ... EiX = 0. .12 If 77 < 0. Proposition 1. w2=inf {t>w1:Jt_#i. Now let r) = 0. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.. then M < oo a.1 of St / t is > 0.s.Jt=i}.. [APQ]. [315].1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes.g. see [APQ] p. The case 77 > 0 is similarly easy. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. Theorem II. The proof of Proposition 1..Jt=i}.Eiw o'o Eiw • E ^ifjµs. also + . + Xn SWn ](1 a . If 77 > 0. Statistical aspects are not treated here. then M = 00 a. with X2.. dt ..s. having the Pidistribution of X. See Meier [258] and Ryden [314].1)Eiw = 0.\ i and EiX1 Ei f 13 J.ld.1 jEE = (p .2(a) p. let some state i be fixed and define w=wl=inf{t >0:Jt_#i.4. The mainstream of the present chapter follows [16]. .152 CHAPTER VI. some early studies are in Janssen & Reinhard [211].
j. oo) = J ao 0 G+(i.g.IIG +II)e.x). 00 (2. by specializing results for general stationary risk processes (Theorem II . the definition of .5.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. 6.Jr+ =j. IIG+ II denotes the matrix with ijth element IIG+(i.i. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. Thus. we define the convolution operation by the same rule as for multiplication of realvalued matrices. That is.A) =ZI(St E. see also Example II. G+ is the matrix whose ijth element is E G +(i.j. dx)/jBj(y . oo)) = f R(i. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0).. B* in Section 1. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. j. let G+(i.(u) = Pi(M < u) = e' E G+ (u)(I .4) we obtain the following result . j.6.3*B *(y)dy. j.x. k.a/i. •) II = JG+(i. •).2) R(dx)S((y . oo)). T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i.j.A) = Pt(ST+ E A. only with the product of real numbers replaced by convolution of measures.1 irG+(dy)e =. but is substantially more involved than for the compound Poisson case . •) * G +(k. . However . we get the same ladder height distribution as for the averaged compound Poisson model. •)• kEE Also. n=0 (2.j E E. For measurevalued matrices.dx). Proposition 2. j.2 (a) The distribution of M is given by 00 1 . e. (y. for i. The form of G+ turns out to be explicit (or at least computable). T R(i.6*.EA. cf. Let further R denote the preT+ occupation kernel. Proposition 2.Jt=j)dt.1) 0 (b) G+ (y.2.
154 CHAPTER VI. JJ = j. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. hence uniquely specified by its intensity matrix Q (say).IIG+II)e. {mx} is a non terminating Markov process on E. we need as in Chapters II. see Figure 2.2 useful . the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. we need to invoke the timereversed version {Jt } of {Jt} .3 When q > 0.3.1 for an illustration in the case of p = 2 environmental states of {Jt}. lines in the path of {St}. To make Proposition 2. From this (2.1) follows by summing over n and j. III to bring R and G+ on a more explicit form . To this end . 0  x Figure 2. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . resp.3) We let {St*} be defined as {St}. mx = j when for some (necessarily unique) t we have St = x. and that the environment is j at the nth when we start from i is e . The u proof of (2. thick. That is. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value.2) is just the same as the proof of Lemma 11. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). .6. marked by thin.1 The following observation is immediate: Proposition 2. St < S* for u < t.
0.and a jump (claim arrival) occurs at time t. Q( n+l) _ ^. the sequence {Q(n)} A* defined by Q(O) = . we say that the excursion has depth 0. Proof The argument relies on an interpretation in terms of excursions. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal.2. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. Figure 2. = x}.*. The definitions are illustrated on Fig. For example the excursion of depth 2 has one subexcursion which is of depth 1. In general. ( Q( n)) converges monotonically to Q. } is a minimum value at v = t.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* . Otherwise each jump at a minimum level during the excursion starts a subexcursion. s]. 2. 0 mms1   ^O \ T. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2.(/3i)diag + T S(dx) eQx. If there are no jumps in (t. and the excursion is said to have depth 1 if each of these subexcursions have depth 0.2 where there are three excursions of depth 1.2 . and the excursion ends at time s = inf {v > t : S. corresponding to two subexcursions of depth 0. Note that the integral in the definition of W(Q) is the matrix whose ith row is the ith row of _ 3 f e2Bi(dx).(/3i) diag.2. {S. An excursion of {St*} above level x starts at time t if St = x.. Furthermore.
Then a jump to j (i. e. 7rE Proof We shall show that Fi(Jt=j. = j.4). StEA .. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. mx+dx = j) occurs in two ways .6) . Theorem 2 .4) To show Q = cp(Q).156 CHAPTER VI. A) = f Pi(mx = j) dx eie4xej dx A u (2.5 R(i. (2.Qi + )%pij) Now just note that t pij and insert (2.u< t). It follows that qij = A. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . A) = L' U(j. i.s. Similarly..T+>t) _ ^iF 7ri (JJ =i. h. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. p1^) Define a further kernel U by f U(i.(01)diag = Q. Similarly by induction . Suppose mx = i. By considering minimum values within the excursion.5) A (note that we use A = {x : x E Al on the r. 0)).j. Q = W(Q) follows. of the definition to make U be concentrated on (co.j +/3ipij. A). or through an arrival starting an excursion terminating with J. j.St EA. MARKOVIAN ENVIRONMENT Let p=7) be the probability that an excursion starting from Jt = i has depth at most n and terminates at J8 = j and pij the probability that an excursion starting from Jt = i terminates at J8 = j. Writing out in matrix notation .St <S*. we first compute qij for i $ j. either due to a jump of {Jt } which occurs with intensity A= j.
and get irPi(Jt =j. St E A... 0<u<t). THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. consider stationary versions of {Jt}.S„<0. and we let k be the corresponding right eigenvector normalized by Irk = 1. oo))dx. K( n (d) the sequence converges monotonically to K.Jo=i. (b) for z > 0.0<u<t. G+((z. we shall see that nevertheless we have enough information to derive.. Remark 2.0<u<t) = P. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). x < 0. {Jt }.Qi)diag. St < St U.St <Su.=StSt. 0 < u < t) = 7rjPj(Jt =i. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx)..g. dt.4].. and to obtain a simple solution in the .t.r. e. To this end.z+>t) = P.6).(Jt=j. We may then assume Ju=Jtu. oo)) = f o' eIXS((x + z. and this immediately yields (2. (Jo = j. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A. Jt = i.StEA. the CramerLundberg approximation (Section 3).6 is hardly all that explicit in general.2. 0 +1) = cp (K( n)) defined by K(o) = A .St EA.6 (a) R(dx) = eKxdx. St EA.6(b): from 7rK = 0 we get 7rG+(dy)e = J W 7reKx(fiiBi(dy + x))diag dx • e 0 w(fiiB1(dy + x))col dx f 0 EirifiiBi(y)dy = fi*B*(y)dy• iEE 0 Though maybe Corollary 2. S.(.7 It is instructive to see how Proposition 2.1 can be rederived using the more detailed form of G+ in Corollary 2. where A is the diagonal matrix with 7r on the diagonal: Corollary 2. From Qe = 0. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2.
158
CHAPTER VI. MARKOVIAN ENVIRONMENT
special case of phasetype claims (Chapter VIII). As preparation, we shall give at this place some simple consequences of Corollary 2.6. Lemma 2 .8 (I  IIG+II)e = (1  p)k. Proof Using Corollary 2.6(b) with z = 0, we get
IIG+II = feIxsux, oo dx.
In particular, multiplying by K and integrating by parts yields
0
(2.7)
I)S(dx) KIIG+II =  (eKx
T
= K  A + (,13i)diag 
Z
S(dx) = K A.
2.8)
0 OO
Let L = (kir  K)'. Then (k7r  K) k = k implies Lk = k. Now using (2.7), (2.8) and ireKx = ir, we get
kirIIG +IIe =
ao k f
7rS((x , oo))e = k (lri(3ips, ) rowe = pk,
0 KIIG+IIe = Ke,
(kirK)(I  IIG+II)e = kKepk+Ke = ( 1p)k.
Multiplying by L to the left, the proof is complete. u
Here is an alternative algorithm to the iteration scheme in Corollary 2.6 for computing K. Let IAI denote the determinant of the matrix A and d the number of states in E. Proposition 2.9 The following assertions are equivalent: (a) all d eigenvalues of K are distinct; (b) there exist d distinct solutions 8 1 , .. , sd E {s E C : its < 0} of (A + (131(Bi[s]  1))diag  sIl = 0. (2.9) I n that case , then Si, ... , sd are precisely the eigenvalues of K, and the corresponding left row eigenvectors al, ... , ad can be computed by
ai (A 
(fi(Bi[Si]

1))d iag  siI) = 0.
(2.10)
2. THE LADDER HEIGHT DISTRIBUTION
Thus, al seal K=
159
(2.11)
ad sdad Proof Since K is similar to the subintensity matrix Q, all eigenvalues must indeed be in Is E C : 2s < 0}.
Assume aK = sa. Then multiplying K = W(K) by a to the left, we get sa = a
f A It follows that if (a) holds, then so does (b), and the eigenvalues and eigenvectors
(
 (f3i)diag +
eS(dx)
= a (A  (/3i) diag + (/3iEi[s])diag)
can be computed as asserted. The proof that (b) implies (a) is more involved and omitted; see Asmussen u [16]. In the computation of the CramerLundberg constant C, we shall also need some formulas which are only valid if p > 1 instead of (as up to now) p < 1. Let M+ denote the matrix with ijth entry M+(i,j) = xG+(i,j;dx). 0 Lemma 2 .10 Assume p > 1. Then IIG+II is stochastic with invariant probability vector C+ (say) proportional to irK, S+ _ 7rK/(7rKe). Furthermore, irKM+e = p  1. Proof From p > 1 it follows that St a4' oo and hence IIG+II is stochastic. That 7rK = e'Q'0 is nonzero and has nonnegative components follows since Qe has the same property for p > 1. Thus the formula for C+ follows immediately by multiplying (2.8) by 7r, which yields irKIIG+II = irK. Further M+ = fdzfeS(( x+z oo)) dx f 00 dy fy eKx dx S((y, oo)) 0 0 m K' f (eKy  I) S((y, oo))dy, 0 00
7rKM+e = 7r f d y(I  eKy) S((y, oo))e
= lr(/3ipB;) diage 
irII G +Ile
=p1
160
CHAPTER VI. MARKOVIAN ENVIRONMENT
u
(since IIG+II being stochastic implies IIG+ IIe = e).
Notes and references The exposition follows Asmussen [17] closely (the proof of Proposition 2.4 is different). The problem of computing G+ may be viewed as a special case of WienerHopf factorization for continuoustime random walks with Markovdependent increments (Markov additive processes ); the discretetime case is surveyed in Asmussen [15] and references given there.
3 Change of measure via exponential families
We first recall some notation and some results which were given in Chapter II
in a more general Markov additive process context. Define Ft as the measurevalued matrix with ijth entry Ft(i, j; x) = Pi[St < x; Jt = j], and Ft[s] as the matrix with ijth entry Ft[i, j; s] = Ei[e8St; Jt = j] (thus, F[s] may be viewed as the matrix m.g.f. of Ft defined by entrywise integration). Define further
K[a] = A + ((3i(Bi[a]  1))  aI
diag
(the matrix function K[a] is of course not related to the matrix K of the preceding section]. Then (Proposition 11.5.2):
Proposition 3.1 Ft[a] = etK[a] It follows from II.5 that K[a] has a simple and unique eigenvalue x(a) with maximal real part, such that the corresponding left and right eigenvectors VW, h(a) may be taken with strictly positive components. We shall use the normalization v(a)e = v(a)hi') = 1. Note that since K[0] = A, we have vi°> = 7r, h(°) = e. The function x(a) plays the role of an appropriate generalization of the c.g.f., see Theorem 11.5.7. Now consider some 9 such that all Bi[9] and hence ic(9), v(8), h(e) etc. are welldefined. The aim is to define governing parameters f3e;i, Be;i, Ae = 0!^1)i,jEE for a risk process, such that one can obtain suitable generalizations of the likelihood ratio identitites of Chapter II and thereby of Lundberg's inequality, the CramerLundberg approximation etc. According to Theorem 11.5.11, the appropriate choice is
e9x
09;i =13ihi[9], Bo;i (dx) = Bt[B]Bi(dx),
Ae = AB 1K[9]De  r.(9)I oB 1 ADe + (i3i(Bi[9] 
1))diag  (#c(9) + 9)I
3. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES
161
where AB is the diagonal matrix with h(e) as ith diagonal element . That is,
hie) DEB) _ ^Y' Me)
iii
i#j i=j
+ /i(Bi[9] 1)  r. (9)  0
We recall that it was shown in II . 5 that Ae is an intensity matrix, that Eie°St h(o) = etK(e)hEe ) and that { eest  t(e)h(9 ) } is a martingale. t>o We let Pe;i be the governing probability measure for a risk process with parameters ,69;i, B9; i, A9 and initial environment Jo = i. Recall that if PBT) is ]p(T) the restriction of Pe ;i to YT = a {(St, Jt) : t < T} and PET) = PoT), then and PET) are equivalent for T < oo. More generally, allowing T to be a stopping time, Theorem II.2.3 takes the following form: Proposition 3.2 Let r be any stopping time and let G E Pr, G C {r < oo}. Then
PiG = Po;iG = hE°) Ee;i lh
1 j,)
exp {BST + rrc(0 ) }; G .
J
(3.1)
Let F9;t[s], ice ( s) and pe be defined the same way as Ft[s], c (s) and p, only with the original risk process replaced by the one with changed parameters. Lemma 3.3 Fe;t [s] = et"(B) 0 1 Ft[s + O]0. Proof By II.( 5.8). u
Lemma 3.4 rte ( s) = rc(s+B )  rc(O). In particular, pe > 1 whenever ic'(s) > 0. Proof The first formula follows by Lemma 3.3 and the second from Pe = rc'' (s).
Notes and references The exposition here and in the next two subsections (on likelihood ratio identities and Lundberg conjugation) follows Asmussen [16] closely (but is somewhat more selfcontained).
3a Lundberg conjugation
Since the definition of c( s) is a direct extension of the definition for the classical Poisson model, the Lundberg equation is r. (y) = 0. We assume that a solution
162
CHAPTER VI. MARKOVIAN ENVIRONMENT
y > 0 exists and use notation like PL;i instead of P7;i; also, for brevity we write h = h(7) and v = v(7).
Substituting 0 = y, T = T(u), G = {T(u) < oo} in Proposition 3.2, letting ^(u) = S7(u)  u be the overshoot and noting that PL;i(T(u) < oo) = 1 by Lemma 3.4, we obtain: Corollary 3.5
V)i(u,
T) =
h ie 7uE L,i
e 7{(u)
h =(u)
e WO
; T(u) < T ,
(3 . 2) (3.3)
ioi(u)
= h ie 7u E
hj,(„)
.
Noting that 6(u) > 0, (3.3) yields
Corollary 3.6 (LUNDBERG'S INEQUALITY) Oi(u)  < hi efu. min2EE h9
Assuming it has been shown that C = limo, 0 EL;i[e7^(u)/hj,(„j exists and is independent of i (which is not too difficult, cf. the proof of Lemma 3.8 below), it also follows immediately that 0j(u)  hiCe7u. However, the calculation of C is nontrivial. Recall the definition of G+, K, k from Section 2.
Theorem 3 .7 (THE CRAMERLUNDBERG APPROXIMATION) In the lighttailed case, 0j(u)  hiCe7u, where
C (PL 1) "Lk.
(3.4)
To calculate C, we need two lemmas . For the first, recall the definition of (+, M+ in Lemma 2.10. Lemma 3 .8 As u 4 oo, (^(u), JT(u)) converges in distribution w.r.t. PL;i, with the density gj(x) (say) of the limit (e(oo), JT(,,,,)) at b(oo) = x, JT(oo) = j being independent of i and given by
gi (x) = L 1 L E CL;'GL (e,.1; (x, oo)) S+M+e LEE
Proof We shall need to invoke the concept of semiregeneration , see A.1f. Interpreting the ladder points as semiregeneration points (the types being the environmental states in which they occur), {e(u),JJ(u))} is semiregenerative with the first semiregeneration point being (^(0), JT(o)) _ (S,+, J,+). The formula for gj (x) now follows immediately from Proposition A1.7, noting that the u nonlattice property is obvious because all GL (j, j; •) have densities.
3. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES
Lemma 3 .9 KL = 01K0  ryI, G+[ry] _
163
111G+IIA, G+['y]h = h.
Proof Appealing to the occupation measure interpretation of K, cf. Corollary 2.6, we get for x < 0 that eteKxej dx =
fPs(StE dx,J =j,r > t)dt
= hie7x f O PL;i(St E dx, Jt = j, T+ > t) dt hj o
= ht e7xe^eK`xej dx,
which is equivalent to the first statement of the lemma. The proof of the second is a similar but easier application of the basic likelihood ratio identity Proposition 3.2. In the same way we get G+['y] = AIIG+IIT1, and since IIG+ IIe = e, it follows that
G +[ry l h
= oIIG+IIo 1h = AIIG+ IIe =
De
= h.
Proof of Theorem 3.7 Using Lemma 3.8, we get EL (e'W ); JT(.) = jl = f 00 e 7xgj (x) dx L J o 1 °°
f e7^G+( t, j; (x, oo)) dx S+M+e LEE °

1 (+;l f S +M +e LEE 0
rr ry S +M +e LEE
0 1(1  e7 x ) G+(1,j; dx)

1
E(+(IIG+(e,j)IIG+[t,j;
In matrix formulation, this means that
C =
E L;i
e7f()
hj,r(_) L
 L
ryC M e
L
c+
(IIG+II  G +[ 7]) 0le
1
L
YC+M+e
'y(PL  1)
(ir KL) (I  G+[ y]) 0le,
164
CHAPTER VI. MARKOVIAN ENVIRONMENT
using Lemma 2.10 for the two last equalities. Inserting first Lemma 3.9 and next Lemma 2.8, this becomes 1 7r LA 1(YI  K)(I  IIG+II)e 'Y(PL  1) = 1 P 7r LA 1(yI  K) k = 1P 7rLO1k. Y(PL  1) (PL  1 ) Thus, to complete the proof it only remains to check that irL = vL A. The normalization vLhL = 1 ensures vLOe = 1. Finally, VLOAL = vLAA'K['Y]A = 0
since by definition vLK[y] = k(y)vL = 0.
u
3b Ramifications of Lundberg 's inequality
We consider first the timedependent version of Lundberg 's inequality, cf. IV.4. The idea is as there to substitute T = yu in 'Pi (u, T) and to replace the Lundberg exponent y by yy = ay  yk(ay ), where ay is the unique solution of rc(ay)= 1 Y Graphically, the situation is just as in Fig. 0.1 of Chapter IV. Thus, one has always yy > y, whereas ay > y, k( ay) > 0 when y < 1/k'(y), and ay < y, k(ay) < 0 when y > 1/k'(y). Theorem 3 .10 Let C+°) (y) _ 1
miniEE hiav)
Then 1 y< (y)
y>
Vi(u,yu)
Pi(u) 
C+°)(y)hiav)
e7vu,
(3.6)
V,i(u,yu)
< C+)(y)hiar )e 'Yvu,
(y) (3.7)
Proof Consider first the case y <
Then, since k (ay) > 0, (3 .1) yields
'12(u,yu)
hiav)]E'iav,i
h(ay ) J*(u)
exp {ayST(,L ) +r(u)k( ay)}; T(u) < yu
8 ) Then for all i E E and all u > 0. (3.7. if y > 1lk'(ry).11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy). we have ic(ay) < 0 and get 'i(u) . yu) f h(av)e v avuE«v.(ay)}. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av). av 'i [h. as in the classical case (3.i [eT(u)K(av ). Our next objective is to improve upon the constant in front of a7u in Lundberg's inequality as well as to supplement with a lower bound: Theorem 3.i [e*(u)K(av). oo)) and.y)G+(z. r(u) yu o)(y)eavuEav.j) * coj)(u) _ f u ^Pj(u .5).V)i(u.3.9) For the proof. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav . (u. we shall need the matrices G+ and R of Section 2.00 su e7( ( 3. 1 Similarly. Chie ryu < Vi(u ) < C+hie 7u. hj P .5) will produce the maximal ryy for which the argument works.j.. yu < r(u) < 00] < hiav)C+o)( y)eavu+yuw(av) 0 Note that the proof appears to use less information than is inherent in the definition (3.i I (a) exp {aye(u) + r(u)r. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. exp {e() + r(u))} . for a vector <p(u) = (cpi (u))iEE of functions . However. dy)• o iEE jEE . We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. we let G+ * W(u) be the vector with ith component E(G+(i.
jEE u 0 j. _ To see that the ith component of U * G(u) equals ?Pi (u). dy) = aj f Bj(dy . dy) : 1(u) < C+ > hj u e(1tL)G+(i.ery(&u+x)Bj (dy) Bj(u Bj (u . just note that the recursion <p(n+1) = G + G+ * (p(n) holds for the particular case where cpin)(u) is the probability of ruin after at most n ladder steps and that then obviously u cp2n) (u) + t. °O . dx) f e7( vu)Bj (dy . dy) 00 C+ ijhj f R(i. 00 f C_ hj f e(Y)G+(i.x) jEE 00 u 0 //^^ C+E. j. Proof Write UN = EN G+ .& (u). dx). dx) 100 C .x ) = Gi(u).u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. we have G *(N +1) * ^. 0 G+(i. j. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. j. if r+ (n) is the nth ladder epoch.166 CHAPTER VI. = Eo G+ G. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u).x) x) jEE 0 E Qj f jEE R(i. 00 Thus C+ > hj f"o e7(Yu)G +(i.j. MARKOVIAN ENVIRONMENT Lemma 3 . U = U".x ) R(i. Then cpin)(u) sit (u) as n + oo.j.3jhj // f 00 R(i.u IMP:°) (u) I < oo. n > oo.(0) ] (u) < sup Jt t. dx ) Bj (u .12 Assume sup1. Hence lim cp(n) exists and equals U * G.7. Lemma 3 .13 For all i and u.dy). j.
T) < C+(')' o)hi7u)e7ou8T .y)G+(i. y]hj = C_ e7uhi. taking cps°) (u) = 0. dy) (3.M > u) = Pi(ST<u.10 ) by Lemma 3 . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.3. let C+(yo) be as in (3. from which the lower inequality follows by letting n * oo. MT < u.11) C_e7u 57 O+[i. 14 Let yo > 0 be the solution of 'c'(yo ) = 0. T) = Pi (7. j. 167 u Proof of Theorem 3. +i . (3.10: Theorem 3 . dy) jEE o (3.13 Let first cp=°)(u) = C_ hie"u in Lemma 3.10) C_ 1 f hje7(y.n) ( u .12) Proof We first note that just as in the proof of Theorem 3. Indeed.MT<u. dy) jEE u U +C_ hje7( yu)G jEE"" +(i.13) Hence. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T .8) with y replaced by yo and hi by h=7o ). and assuming it shown for n. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. j.Pi(MT > u) = Pi(MT < u.M>u) = Ei [VGJT (u .13. (3. 13 and the second by the induction hypothesis . jEE estimating the first term in (3. u The proof of the upper inequality is similar . Then 0< Vi (u )  0i(u. and the proof of the lower one is similar. we have Vii (u) .ST).(u) < T ) to 0i (u) which is different from Theorem 3. and let 8 = e'(70). j. letting MT = maxo<t<T St.u)G+(i.T) = Pi(M > u) . 9 for the last equality in (3. it follows that Vi(u) < C_(yo) h=70)e7ou. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. and using Lemma 3 . We claim by induction that then cpin) (u) > C_ hie7u for all n. j.11).11.tpi(u. this is obvious if n = 0.
[177]. B2 <_s. and finally in part from queueing theory. u > 0. It was long conjectured that 0* Vi. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u).3p.3). The results to be presented show that quite often this is so. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained.o. we also assume that there exist i # j such that either /3i <. we refer to . The Markov process {Jt} is stochastically monotone (4.33 or Bi 0 Bj.. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u). (4. The motivation that such a result should be true came in part from numerical studies.3) Bl <_s. Bp... 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'..o. ".2) (4.31:5)32 . but that in general the picture is more diverse.5) Note that whereas (4. in part from the folklore principle that any added stochastic variation increases the risk..2) alone just amounts to an ordering of the states.168 CHAPTER VI.3) to B = Bi does not depend on i.o. (4. this is not the case for (4.1) Obviously.. this correponds to the usual stochastic ordering of the maxima M'. Further related discussion is given in Grigelionis [176]. < . (4. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44].4) To avoid trivialities. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). <s. Occasionally we strengthen (4. The conditions which play a role in the following are: .o. V)" if z/i'(u) <'c"" (u)..0. we define the stochastic ordering by 0' < s. For the notion of monotone Markov processes. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and ..
< a.x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.8) ^j Tri/iBd(x) .6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB.r (JT(o) = i. then P P P 7rjbj.x) dx u o i =1 i=1 (4. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.. the second follows from an extension of Theorem I1. = aP or b1 = . . it follows by a standard .9) (4.3* f uB(x) z/^. Lemma 4 .9 ) below). Comparing (4. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. dx (4..r (Sr(o) E dx Jr(o) = i.2. ^i 7ri = 1.2)(4..x) of i and using Lemma 4.10) and (4.. we need two lemmas.4) is automatic in some simple examples like birthdeath processes or p = 2 . then j is the more risky state . COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]. Conditioning upon the first ladder epoch.. E 7r i Wi(u . 0 Here (4. note that (4.1 Assume that conditions (4. .3 for the second) *(u) _ /3 *B* (u) +. Then V..1. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = .6.r(u x)dx. (b) P. 2 If al < .7) and Lemma 4.9) follows by considering the increasing functions 3iBi (x) and Oi (u .2)(4.. 7(0) < oo) = pirf+). where 7r2+) = QiµBilri/p.13* J0 u 0*(u .3iBi(x)YPi(u . (4. Proposition 2.. also Proposition 2.7) 7ri. Lemma 4 . = b.1 for the first term in (4... < bp and 7ri > 0 (i = 1. 3 (a) P.10) Q*B*(u)+. p). Proof of Theorem 4.* For the proof.1) which with basically the same proof can be found in Asmussen & Schmidt [49].2.5 (cf. T(0) < oo) = Bi(x) dx/tcai . Theorem 4 ..4. Section 4.. Conditions (4. Proschan & Walkup [140]. b1 < ..4) hold.. we obtain (cf.4) say basically that if i < j .x)B*(x) dx.6).
Q2 = 1.1 and Proposition 4.. Rolski & Schmidt [32]. Proof Since 0. µB2 = 104.r (u ) fails for all sufficiently small e > 0. u To see that Proposition 4. = 102. Recall that . Then i/i*(u) < . dominates the solution 0* to the renewal equation (4.4 is not vacuous. Using (4.0*• i=1 But it is intuitively clear (see Theorem 3.4 Assume that .2. of (4.6). Notes and references The results are from Asmussen.s. 0. µB. For u = 0. of order 101. let = ( 1/2 1/2 ) . 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. (u) may fail for some u. this ruin probability is /3iPBi.* (0).11) i=1` and that A has the form eAo for some fixed intensity matrix A0. except possibly for a very special situation .. it is sufficient to show that 0'. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4.h. it will hold for all sufficiently large u.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters .h.(0) < b *'(0) for e small enough. they are at present not quite complete.3µi < 1 for all i. MARKOVIAN ENVIRONMENT argument from renewal theory that tk. 01 = 103. Bi as e J... As is seen. Frey.4) is essential (the present author conjectures it is).170 CHAPTER VI.4 is the understanding of whether the stochastic monotonicity condition (4. u Here is a counterexample showing that the inequality tp* (u) < V). What is missing in relation to Theorem 4.6).s.(0) = V./3*. i=1 i=1 7'r(0) _ EFioiwi(0) . and from this the claim follows. Then the l.8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi . (u) is not in general true: Proposition 4.0.3i.11) is of order 104 and the r.. (4.
Xt)} is a Markov additive process (a socalled Markovian fluid model. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] . (4. with strict inequality unless rci (y*) does not depend on iEE. is nondegenerate unless bi does not depend on i E E. cf. Hence if 5i 54 0 for some i E E. e. It is clear that the distribution of X.1) .Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.14) is nonzero so that A"(0) > 0. and by Proposition II. This implies that A is strictly convex.)a. Now we can view {Xt} as a cumulative process (see A.4. COMPARISONS WITH THE COMPOUND POISSON MODEL 171 the adjustment coefficient for the Markovmodulated model is defined as the solution y > 0 of ic(y) = 0 where c(a) is the eigenvalue with maximal real part of the matrix A + (rci(a))diag where rci(a) = ai(Bi[a] .1) . it follows by Proposition A1.5..5. Jt = i])' EE = vA+n(6. (4.5 y < ry*. Asmussen [20]) as discussed in 11.2 we have (Ei[e"X'.(a) > 0 for all a 0 0.12) iEE Theorem 4. Further (see Corollary 11.ld) with generic cycle w = inf{t>0: Jt_54 k. in particular .a = E irirci(a). with strict inequality unless a = 0 or bi = 0 for all i E E.13) (4.g.13) implies A(a) > 0 for all a.6 Let (di)iEE be a given set of constants satisfying EiEE iribi = 0 and define A(a) as the eigenvalue with maximal real part of the matrix A + a(bi)diag• Then )t(a) > 0.14) A„(O) iioo varXt t t By convexity.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. Then {(Jt.g. Lemma 4.4(b) that the limit in (4. 0 . which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E. (4.5.a. Proof Define X= f &ids..
6. improving upon more incomplete results from Asmussen. If rci(y* ) is not a constant function of i E E.15) Normalizing h by 7rh = 0. Hence letting e = 0 in (4. (4. a = 1 in Lemma 4.172 CHAPTER VI. the basic equation is (A + (rci(y))diag)h = 0.Qi and Bi are fixed .. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. Since ic is convex with rc'(0) < 0 . we have 7rh' = 0. we get rc (y*) > 0 which in a similar manner implies that u y < y*..p yi and compute 8y 8a a=0 In both cases.e7r)1 (Ici(Y*))diage. The corresponding adjustment coefficient is denoted by ry(e). Frey. Rolski & Schmidt [32]. Here we put a = 1/e.) and rc (•). h'(0) = (Ao . where A. h(0) = e.15) once more and letting e = 0 we get .. whereas the . Notes and references Theorem 4. (4.5 is from Asmussen & O'Cinneide [40]. h depend on the parameter (e or a). this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'..12) and rc*(y*) = 0. In the case of e. Let bi = rci(y*). y.5. Then > risi = 0 because of (4. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. Thus y(e) * y* as e 10.16) Differentiating (4.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . Hence rc (y*) > 0. note that y(a) + mins=1. MARKOVIAN ENVIRONMENT Proof of Theorem 4. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo.eir)h'(0). Further a(1) = rc(y*) by definition of A(.
the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . The additional feature of the model is the following: • Certain transitions of {Jt} from state i to state j are accompanied by a claim with distribution Bid.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. (4.18).19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector).20) Letting a = 0 in (4. (4. Frey. which has recently received much attention in the queueing literature.i(7' *))diagh'(0). (4. We get 0 = (aAo + ( lc&Y))diag)h. then 8a a=o All rci (0) Notes and references The results are from Asmussen.19) holds. and we have proved: Proposition 4. The analogue of Proposition 4.17) (4.8 If (4. We assume that 0 < y < 7i.16) yields Proposition 4. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) .17) by 7r to the left to get (4..18) 0 = 27'(0)p+27r(rs.20) and multiplying by el to the left we get 0 = All + 7'(0)rci (0) + 0 (here we used icl (ry(0)) = 0 to infer that the first component of K[7(0)]h'( 0) is 0). . i = 2. and may have some relevance in risk theory as well (though this still remains to be implemented). 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. 5 The Markovian arrival process We shall here briefly survey an extension of the model. . p.5. Rolski & Schmidt [32].8 when ryi < 0 for some i is open.. Inserting (4. multiplying (4.
2) A(1) = A(' 1) ® A(1. A(1) = A .^) etc. j(2) } be two independent environmental processes and let E(k). refer to notation) { Jt k) }. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . In the above setting. the definition of Bi is redundant because of f3i = 0. We then let (see the Appendix for the Kronecker E = E(1) x E(2). Thus .i. A(1'k) A(2 k1). the definition of Bij is redundant for i i4 j. u Example 5 .6i ) diag. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. Jt2)) (2. where qij is the probability that a transition i * j is accompanied by a claim with distribution. II.(13i )diag.4).174 CHAPTER VI. Indeed. that Bii = Bi . we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . A(l) = tv. with common distribution B. Here are some main examples: Example 5 . For i = j. let { Jt 1) }. This is the only way in which arrivals can occur.d. the claim surplus is a Markov additive process (cf. Again .1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. A(l) = T. and the marked transitions are then the ones corresponding to arrivals. Jt = (Jtl). and thus 1i = 0. Bii = Bi . but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. Note that the case that 0 < qij < 1.2). . and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0.2 for details). T). Bij = B. we use the convention that a1i = f3i where 3i is the Poisson rate in state i. B. A ( 2) = A (2`1 ) ® A. the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (.
say. more recently.iN are zero and all Bi are redundant.3 (AN INDIVIDUAL MODEL) In contrast to the collective assumptions (which underly most of the topics treated sofar in this book and lead to Poisson arrivals). Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. Thus. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired.g. THE MARKOVIAN ARRIVAL PROCESS Bij.. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. DIVORCED.. 11...kl is redundant).4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. u Notes and references The point process of arrivals was studied in detail by Neuts [267] and is often referred to in the queueing literature as Neuts ' versatile point process ... claims occur only at state transitions for the environment so that AN2.. the kth policy enters a recovering state.iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. iN. The individual pays at rate pi when in state i and receives an amount having distribution Bij when his/her state changes from i to j. as the Markovian arrival process ( MAP). and that the policy then expires.. RETIRED...kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. assume that there is a finite number N of policies. superpositions of renewal processes.}. In this way we can model. E 10. with rate ai.. Example 5 . Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . Similarly.1i2 .iN. iN = all BOi2.iN. u Example 5 . DEAD etc.. all Al i2.5. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. MARRIED. iN. INVALIDIZED. WIDOWED.. This means that the environmental states are of the form i1i2 • • • iN with il.iil.iN C17 AilO.. However . i2i . or. In fact .. The versatility of the setup is even greater than for the Markovmodulated model. e.. Bilo.. E = { WORKING.. • upon a claim. possibly having a general phasetype sojourn time.1i2.iN = a2. after which it starts afresh...iil..
for s E E = [0. • Claims arriving at time t of the year have distribution B(t). • The premium rate at time t of the year is p(t). p * = 0 p(t) dt. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . For the Markovmodulated model. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. let the period be 1.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* .3*µs • p = f /3(v) dv 0 0 (6. we may assume that the functions /3(t). 1). Sengupta [336]. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. 1). from an application point of view. Obviously. MARKOVIAN ENVIRONMENT sense any arrival stream to a risk process can be approximated by a model of the type studied in this section : any marked point process is the weak limit of a sequence of such models . Thus at time t the premium rate is p(s + t). p(t) and B(t) are defined also for t t [0. Let 1 1 /3* _ f /3(t) dt. By periodic extension. Lucantoni et at. . [248]. B* = J f B(t) ((*) dt. 0 < t < 1. 6 Risk theory in a periodic environment 6a The model We assume as in the previous part of the chapter that the arrival mechanism has a certain timeinhomogeneity. we talk of s as the 'time of the year'. Neuts [271] and Asmussen & Perry [42]. Without loss of generality. continuity would hold in presumably all reasonable examples.p)/p. one needs to assume also (as a minimum) that they are measurable in t. Some main queueing references using the MAP are Ramaswami [298]. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). where i f00 xB(°) (dx) _ .2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size.176 CHAPTER VI. Lucantoni [248]. )3 t 1 J (6.
B*.t. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . in agreement with the general principle of added variation increasing the risk (cf. Example 6 . p* = A whereas B* is a mixture of exponential distributions with intensities 3 and 7 and weights 1/2 for each (1/2 = ff w(t)dt = f o (1. the average compound Poisson model is the same as in III. Section 4b). In contrast.8. equivalently. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf.(3. It is easily seen that . The behaviour of the periodic model needs not to be seen as a violation of this principle. and thus the averaged standard compound Poisson models have the same risk for all A.10.3) Note that A enters just as a scaling factor of the time axis. let .9).1. Thus.2 Define T 6(T) = p(t ) dt. St = SeI(t). (6. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. the discussion in 111.3(t) = 3A(1 + sin 27rt). The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . the conditional distribution .1 As an example to be used for numerical illustration throughout this section. In contrast. not random. u Remark 6 . it turns out that they have the same adjustment coefficient. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities.6.w(t). Thus . p* as an averaged version of the periodic model. Many of the results given below indicate that the averaged and the periodic model share a number of main features. since the added variation is deterministic. We u assume in the rest of this section that p(t) . and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u.w(t)) dt). In particular. one may think of the standard compound Poisson model with parameters 3*. of the periodic model as arising from the compound Poisson model by adding some extra variability.3* = 3A. we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations.1) and Example 1. respectively. p(t) = A and let B(t) be a mixture of two exponential distributions with intensities 3 and 7 and weights w(t) _ (1 +cos27rt)/2 and 1 . or.
but it turns out to have obvious benefits in terms of guidelining the analysis of the model as a parallel of the analysis for the Markovian environment risk process.5.a) Proof Conditioning upon whether a claim occurs in [t. [101] .3(s + t)dt[B(8+t)[a] .178 CHAPTER VL MARKOVIAN ENVIRONMENT of U.3(v)(B(vl [a] .a. with some variants in the proofs. i. we obtain E.8)..a) = exp { .T) = P(8)(r(u) <T).adt +. 6b Lundberg conjugation Motivated by the discussion in Chapter II. let f 8+1 tc *(a) _ (B* [a] . of the averaged compound Poisson model (the last expression is independent of s by periodicity).1) a = J8 . see the Notes to Section 7).. Dassios & Embrechts [98] and Asmussen & Rolski [43].(3(s + t)dt)e«St adt + /3(s + t)dt .Q(v) (B(„) [a] . 0 (5)(u.f.1) dv . . given that the ith claim occurs at time t is B(8+t).4) At a first sight this point of view may appear quite artificial.g. (6. t + dt] or not.. and define h(s. r(u) _ inf It > 0 : St > u} is the time to ruin . The exposition of the present chapter is basically an extract from [44].a be the c. a) etw*(a) h(s+t.s .1]) .g.5 (see in particular Remark 11. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00). 3 E(8)eaSt = h(s.tc* (a)] dv then h (. Daykin et. Jt = (s + t) mod 1 P(8) . we start by deriving formulas giving the m.g. e.e.^8 [. As usual. 1). J Theorem 6 . a) is periodic on R. The claim surplus process {St} may be seen as a Markov additive process.east B(8+t) [a] east .f.a . To this end.al.(8) [eaSt+dt I7t] = = (1 . with the underlying Markov process {Jt} being deterministic period motion on E = [0. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.(1 . Notes and references The model has been studied in risk theory by. of the claim surplus process.1) .
1]) .1)dv l og E(8) et where atetk•(a) h(t. h(s + t.log h(s.0(s + t)dt[B(8+t)[a] .3.6 . 0) exist and are finite. St)} and . According to Remark 11.t = 1 by Theorem 6.s. a) = h(s. 9) is a P ( 8)martingale with mean one.5.4 For each 0 such that the integrals in the definition of h(t . a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11. dt log E(8)east a + f3(s + t) [B(8+t) [a] . a) = exp I f t3(v)(kv)[a) .1]) . at + f log h(s + t.. Proof In the Markov additive sense of (6. a) .* (a) h(s. so that obviously {Lo. we can write Lo Jt.(e) Let = h( h(Jo.1)dv  o h(t. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.4). + v)(B([a] . St)} . a).5 The formula for h(s) = h(s. it then suffices to note that E(8)Le.c* (e) {Le. 0) P(8)a. B) eoSt t.3(s + t)[D(8 +t)[a] .2.9) eastt. u Remark 6. a) Thus E(8)east = h(s + t.1]. With g the infinitesimal generator of {Xt} = {(Jt.t} is a multiplicative functional for the Markov process { (Jt. a) h(s + t. E (8)east (a +.(8)east 179 = = = = = E(8)east (1 .6.t.9 as follows. a) Corollary 6.adt +.t}t>o = h(s. a) et.
of St is as for the asserted periodic risk model.3(v)( Bi"i [a] .y) = eayh(s). Proof (i) Check that m. P(s) (T(u) < oo) = 1 for all u > 0.'y).f. MARKOVIAN ENVIRONMENT ha(s.tc] dv} (normalizing by h(0) = 1).1. Sdt) = h(s + dt) eadt (1 . ( iii) use approximations with piecewiese constant /3(s).2. Now define 'y as the positive solution of the Lundberg equation for the averaged model. That is. see [44] for 11 a formal proof. B(s). Proposition 6. such that for any s and T < oo. St)} with governing probability measures Fes). J s [. Lemma 6 . A further important constant is the value yo (located in (0. as above E (s) ha(Jdt.1) .180 CHAPTER VI. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + .3(s)B(s) [a]h(s). yo is determined by 0 = k* (70) = QB*. Proposition 6. say.6 ( s ) exp { 0( s )&s) [a] + tc .g.3(s)dt • B(s)[a]h(s) = gha(s. . correspond to a new periodic risk model with parameters ex .4. we put for short h(s) = h(s. (iv) finally.3. it follows by Theorem II. When a = y. cf. y solves n* (y) = 0.0) = Kh(s). the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity.7 When a > yo. For each 0 satisfying the conditions of Corollary 6.a .3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +.6 The P(s).5 that we can define a new Markov process {(Jt.3(s)h(s) + h'(s) +.(3(s)dt) +. the requirement is cha(i. [70] . Bet)(dx) = ^ B(t ) (dx). However.60(t) = a(t)B(t)[0]. 0 < s < 1. 0) = h(s) + dt {ah(s) . ry)) at which n* (a) attains its minimum. That is.T. (ii) use Markovmodulated approximations (Section 6c).
The proof involves machinery from the ergodic theory of Markov chains on a general state space. J C R+ such that the B(8).g. the Markov process {(^(u). 1).9) and noting that weak convergence entails convergence of E f (^(u).6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. and we refer to [44]. ^(u) = ST(u) . a) a > ry0 (6.9) 0(')(u) = h(s. T(u) < (6.1. 9(u)) for any bounded continuous function (e. a) e«uE(8 ) e «^ .10) Then for each a. Corollary 6.4.1) the distribution of (l: (oo). a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11.7) h(B(u). xEJ 0 (s)b(8)(x) > 0. q) = eryx/h(q)). B(oo)). (6. and no matter what is the initial season s. the mean number of claims per unit time is p« 181 = Jo 1.2). we need the following auxiliary result . u which is > 1 by convexity.9(u))} u>0. s E I.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u. Lemma 6 .8) (6. which is not used elsewhere in the book. RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6.9 Assume that there exist open intervals I C [0. has a unique stationary distribution.2. f (x. a) TI h(9(u). 0(u)) * (b(oo).u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. T) = h(s. Wu). e(cc)) Letting u > oo in (6. considered with governing probability measures { E(8) }E[ . say s0.6. have components with densities b(8)(x) satisfying inf sEI. Here and in the following. we get: .
182 CHAPTER VI. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3.16. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs . it does not seem within the range of our methods to compute C explicitly.11) Note that ( 6. where C(o) = 1 + info < t<i h(t) . For our basic Example 6 . 10 shows that certainly ry is the correct Lundberg exponent.10 Under the condition (6.6 for the Markovmodulated model: Theorem 6 .1. At this stage . this provides an algorithm for computing C as a limit. 6.W. Vi(8) (u) . Noting that ^(u) > 0 in ( 6. MARKOVIAN ENVIRONMENT Theorem 6. A=1/4 A=1 A=4 0 Figure 6.10) of Lemma 3. Among other things. (6. 1. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c. illustrating that the effect of seasonality increases with A.Ch(s)ery".) C = E1 h(B(oo)) u + oo.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.1. Theorem 6 .9).1 In contrast to h.ir) } Plots of h for different values of A are given in Fig. 11 7/'O (u) < C+°)h(s) ery". where e.
g. 1 (6. 1 ) and all u > 0. ay) • (6.4. we first note that the function fu° exu {w • 3e . C_h(s)e7u < V. Theorem 6.6.13 Let = 1 B(t) C o<tf i h(t) 2no f °O e'r(Yx)B( t) (dy)' (x) x 1 B(t) (x) C+ = sup sup o<t<i h ( t) xo J. . r.42 • exp {J_ cos 27rs . we obtain Co) = 1.(ay) > 0 when y < 1/ic' (7).13 to our basic example. #c( ay) < 0 when y > 1/tc'('y). Just as in IV. RISK THEORY IN A PERIODIC ENVIRONMENT Thus.w)e4u dx 9w + 7(1 .(s)(u) < C+h(s)e7".13) Elementary convexity arguments show that we always have ryy > Y and ay > ry.12 Let 00)(y) 1 Then info < t<i h(t. T) and replace the Lundberg exponent ry by ryy = ay . Theorem 6 . e.17) (6. yu) 000 (u) .12) As for the Markovian environment model. (6..47r sin 27rs + 167r cos 47rs .w ) • 7e u{w • 3e3x + ( 1 .3x + (1 . where ay is the unique solution of W(ay) =y• (6.7x j dx _7x } _ 6w + 6(1 . we substitute T = yu in 0(u.(8) (u. Consider first the timedependent version of Lundberg's inequality.42 so that 183 tp(8) (u) < 1.167r I Cu.7e . e7 ( yx)B(t)(dy) > Then for all $ E [0.15) The next result improves upon the constant C+) in front of eryu in Theorem 6.11 as well as it supplements with a lower bound. whereas ay < y.yr. Lundberg's inequality can be con siderably sharpened and extended. the proofs are basically the same as in Section 3 and we refer to [44] for details. We state the results below. (ay). in our basic example with A = 1.w) .w)e4u .0(8) (u+ yu) (6.14) < C+)(y)h(s) e7yu.16 In order to apply Theorem 6.
. 1) for the environment). The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'. exp 2^ cos 21rs . completing a cycle .013. we have the following result: Theorem 6 .19 I eu.66. C+ = 1.I eu.181 s(u) < 1. Some of the present proofs are more elementary by avoiding the general point process machinery of [44]. n}.T) < C+('Yo)h( s.4^ sin 2irs + 16^ cos 41rs .cos 27rs . 1).g. where the environment at time t is (s + t) mod 1 E [0.184 CHAPTER VI. much of the analysis of the preceding section is modelled after the techniques developed in the preceding sections for the case of a finite E.18) Notes and references The material is from Asmussen & Rolski [44]. and in fact. but thereby also slightly longer.20). 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. for A = 1 (where 3 e0. .013.'Yo)e (6. Finally.19 } 0 <8<1 8 + cos 21rs Thus e.16) with 'y replaced by yo and h(t) by h(t. Then 7oudT . yo).66. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment. . 14 Let C+('yo) be as in (6.1 sin 2irs + 16_ cos 47rs . . 0 <'p(8)(u ) .L sin 27rs + 1 I cos 47rs . Thus. Thus C_ = 2 inf ex cos 2irs . such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0.\ 3 C+ = sup 6 exp { A (.\ = 0 .16. the nth Markovian environmental process {Jt} moves cyclically on {1. and let 8 = er' (Y0).(8)(u. 1/i18 1 s (u) > 0.. Of course.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.0.20 •exp { 2n cos 27rs . with s the initial season.1 sin 27rs + 1 cos 47rs .
19) n 0 0 ••• n Arrivals occur at rate /3ni and their claim sizes are distributed according to Bni if the governing Markov process is in state i. one simple choice is Oni = 0( i . AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. DUAL QUEUEING MODELS 185 within one unit of time on the average . We let {Stn)} (6. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model.jEE of the risk process. . To this end. Notes and references See Rolski [306]. Let 0j.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). since the settings are equivalent from a mathematical point of view. T) can be expressed in a simple way in terms of the waiting time probabilities of a queueing system with the input being the timereversed input of the risk process. M(n) = Supt>o Stn).1 ((i 1)/n) ) and Bni = B . Bi. (6. This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade.7. it is desirable to have formulas permitting freely to translate from one setting into the other. but others are also possible.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). A be the parameters defining the risk process in a random environment and consider a queueing system governed by a Markov process {Jt } ('Markovmodulated') as follows: • The intensity matrix for {Jt } is the timereversed intensity matrix At _ A ())i. z/'i (u. Thus.
Proposition 7. Now let In denote the environment when customer n arrives and I* the steadystate limit. J*) is the steadystate limit of (Vt. The first conclusion of that result then states that the events {T(u) < T. Jo = i. {Jt }o<t<T• Then we may assume that Jt = JTt. . JT = i) = 'P. (7. The actual waiting time process 1W1.1) over j. In particular. .3). (7. (VT > u I JT = 2). Jt ). (7. (7. Proof Consider stationary versions of {Jt}o<t<T. JT = i) = P(V > u.2). For (7. T) = 7ri 1 P.3. J* = i) for all j.1) 7ri In particular. JT = Z). • The queueing discipline is FIFO. ii (u) = it /3 P(W > u.P(V > u. JT = j) = 7rjPj(VT > u. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V.186 CHAPTER VI.. Proposition 7. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi.1) follows. Taking probabilities and using the stationarity yields 7riPi(T(u) < T. JT = j} and {VT > u. and (7. Then Pi(T(u) < T.=1 .oo in u (7.2) and use that limF (VT > u. JT = j) = LjPj (VT > u.1 Assume V0 = 0. Jo = j.3) 7ri where (V. JJ = i). I* )3i P(V > u. 2 .2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. I* = i). let T . 0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11. JT = i} coincide.. J* = i) = P.0i (u . just sum (7.1).. J* = i).2) Oi(u) = 1.T(V > u I J* = i).4) where 0* = >jEE 7rj/3j.n(VT > u. and for (7.
7) (7.T)(T(u) <T) = P(8)(VT > u). Proposition 7. In the setting of the periodic model of Section 6.4) and (7. and of these. >u..5) follows from (7. and Rolski [306].=i) a4. we have 1: I(W.o.7. [243]. (7.6) (7. and further references (to which we add Prabhu & Zhu [296]) can be found there. see Regterschot & van Doorn [123]. N * oo. the dual queueing model is a periodic M/G/1 queue with arrival rate 0(t) and service time distribution B(') at time t of the year (assuming w. on average /32TP(V > u.3) improving somewhat upon (2.8) For treatments of periodic M/G/1 queue. I* = i. with (7. on average 0*T customers arrive in [0. if T is large. T]. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. that /3(t).3).I *=i). and (7.4). .7) of that paper. and one has PI'>(rr(u) < T) = P('_T)(VT > u). I*) with the timeaverage . The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301].I. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. Lemoine [242]. B(t) have been periodically extended to negative t). P(1')(r(u) < oo) = P(')(00) > u). With {Vt} denoting the workload process of the periodic queue. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. n=1 N However.1 is from Asmussen [16].g. a general formalism allowing this type of conclusion is 'conditional PASTA'. P(W >u. see in particular Harrison & Lemoine [186]. Taking the ratio yields (7. J* = i) see W > u.4) can be found in Regterschot & de Smit [301]. The relation (7. A more probabilistic treatment was given by Asmussen [17]. a paper relying heavily on classical complex plane methods.l. P(.
This page is intentionally left blank .
1) (other terms are accumulated claims or total claims). Thus. However . and the evolution of the reserve may be described by the equation Rt = u . {Rt} moves according to the differential equation R = p(R). and that the claim sizes U1. U2. t] are Nt At = Ui (1.Chapter VII Premiums depending on the current reserve 1 Introduction We assume as in Chapter III that the claim arrival process {Nt} is Poisson with rate .2) tk(u. resp . .. the aggregate claims in [0.d. are i. finite horizon.6. Zt As earlier.. 189 . and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo).T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite. the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r.i. i&(u. z/i(u) = F IinffRt< 0IRo=u 1 (1.T) = F(T(u) < T). with common distribution B and independent of {Nt}. Thus in between jumps.At + p(R8) ds.
say e. However. 1 . the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. Now return to the general model. there is positive probability. No tractable necessary and sufficient condition is known in complete generality of the model. Assume 0(u) < 1 for some u. when x > p/S. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . but assume now that the company borrows the deficit in the bank when the reserve goes negative. If Ro = v < u. we can put Rt = Rt + p/S.p2. Proposition 1. i. Proof Obviously '(u) < ilb(v) when u > v.i(u) = 1 for all u.4 Either i.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v.'(u)) > 0 so that V'(v) < 1. Hence in terms of survival probabilities. it seems reasonable to assume monotonicity (p(r) is u .p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. Example 1. oo) is given by i (u + p/S). rather than when the reserve itself becomes negative. we get p(r) = p + er.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e.190 CHAPTER VII. where one would try to attract new customers as soon as the business has become reasonably safe.e. P(r) _ p + e(r . say at interest rate b. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r.Vi(v) u > e(1 . Example 1. and the probability of absolute ruin with initial reserve u E [p/S. That is. or o(u) < 1 for all u. that {Rt} will reach level u before the first claim arrives.2. Thus at deficit x > 0 (meaning Rt = x). but when the reserve comes above v. pi > p2 and p(r) = One reason could be competition. In this situation. dividends are paid out at rate pi .
However. B and (constant) premium rate p. Then 0(u) = P(V > u). let uo be chosen such that p(r) < p = /3µB for r > uo.1.uo) and. that u zPp(u . (1.2(d)). In particular.2) we have t Vt = At . Then if u > no. This is basically covered by the following result (but note that the case p(r) .6) ..2 once more.2(d). which was proved in 11.uo) < 1. . In case (b). [APQ] pp.+ p(r) exists. and hence by a geometric trials argument.3. V = p(V)).3µB for all sufficiently large r.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. if and only if V)(u) < 1 for all u. then l/i(u) < 1 for all u. appealing to Proposition 111.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. In case (a).T) = P(VT > u). Starting from Ro = uo. obviously infu<uo z/'(u) > 0.4. { Vt} remains at 0 until the next arrival). In between jumps.6 For any T < oo. (1.1. We next recall the following results. Hence ik(u) < 1 for all u by Proposition III.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.1 and increasing in Example 1. Proof This follows by a simple comparison with the compound Poisson model. one can couple the risk process and the storage process on [0. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. say V. cf. That is.5) and the process {Vt} has a proper limit in distribution .1. we have z/i(u) <p(u .1. Proposition I1I.2) for r sufficiently large so that p(oo) = limr. INTRODUCTION 191 decreasing in Example 1. let uo be chosen such that p(r) > p = 0ILB + e for r > uo.b(u. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. (1. and P(Rt + oo) > 0. Theorem 1.f p(Vs) ds. hence Rt < uo also for a whole sequence of is converging to oo. {Vt} decreases at rate p(v) when Vt = v (i. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0.e.5 (a) If p(r) < /. instead of (1. then ?(u) = 1 for all u. 296297): Theorem 1. Let Op(u) refer to the compound Poisson model with the same 0.
x] to (x. say if p(r) goes to 0 at rate 1 /r or faster as r j 0.y)g(y) dy. (1. oo). RESERVEDEPENDENT PREMIUMS In order to make Theorem 1.Sx} dx.Qw(x) . It follows in particular that 0(u) = fg(Y)dy. say when {Vt} is in state y.y. t + dt] if and only if Vt E [x. the l. one having an atom at 0 of size 'yo.s. Note that it may happen that w (x) = oo for all x > 0. x + p(x)dt]).7) Proposition 1. of (1.h. Oeax f x e'Yg (y) dy } = p) eaxa(x) . Then the ruin probability is tp (u) = f' g(y)dy.192 CHAPTER VII.7 p(x)g(x) = tofB (x) + a f (x . the flow of mass from [0. and the other being given by a density g(x) on (0. Now obviously. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. Considering the cases y = 0 and 0 < y < x separately.9) Proof We may rewrite (1. In view of the path structure of {V t }. u Define ^x 1 w(x) Jo p(t) dt. Then w(x) is the time it takes for the reserve to reach level x provided it starts with Ro = 0 and no claims arrive. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. (1. say. say.h. we thus need to look more into the stationary distribution G. of (1.8) as g(x) = p 1 {yo13e_6x +. for the storage process {Vt}. Corollary 1.Sx}. and is succesful if the jump size is larger than x .8) Proof In stationarity. oo) must be the same as the flow the other way.6 applicable.6w(x) .8) as the rate of upcrossings. yo ^ 1 + oo Q exp {. Jo AX) (1.8) is the rate of downcrossings (the event of an arrival in [t. An attempt of an upcrossing occurs as result of an arrival. we arrive at the desired interpretation of the r.s.6x and that w(x) < oo for all x > 0. B(x) = e. where g(x) = p( ^ exp {. It is intuitively obvious and not too hard to prove that G is a mixture of two components.8 Assume that B is exponential with rate b.
1. INTRODUCTION
where c(x) = 1o + fo elyg(y) dy so that (x) = eaxg(x) _
193
1
p(x)
nkx).
Thus log rc(x) = log rc(0) + Jo X L dt = log rc(0) + /3w(x), p(t) c(x) = rc (0)em"lxl = Yoes"lxl, g(x) = eaxK' (x) = e6x ,Yo)3w'(x)e'6"lxl which is the same as the expression in (1.9). That 'Yo has the asserted value is u a consequence of 1 = I I G I I = yo + f g• Remark 1.9 The exponential case in Corollary 1.8 is the only one in which explicit formulas are known (or almost so; see further the notes to Section 2), and thus it becomes important to develop algorithms for computing the ruin probabilities. We next outline one possible approach based upon the integral equation (1.8) (another one is based upon numerical solution of a system of differential equations which can be derived under phasetype assumptions, see further VIII.7). A Volterra integral equation has the general form x g(x) = h(x) + f K(x, y)9(y) dy, 0 (1.10)
where g(x) is an unknown function (x > 0), h(x) is known and K(x,y) is a suitable kernel. Dividing (1.8) by p(x) and letting K(x, y) _ ,QB(x  y) _ 'YoIB(x) p(x) , h(x) p(x) we see that for fixed to, the function g(x) in (1.8) satisfies (1.10). For the purpose of explicit computation of g(x) (and thereby %(u)), the general theory of Volterra equations does not seem to lead beyond the exponential case already treated in Corollary 1.8. However, one might try instead a numerical solution. We consider the simplest possible approach based upon the most basic numerical integration procedure, the trapezoidal rule hfxN() dx = 2 [f ( xo) + 2f (xi) + 2f ( x2) + ... + 2f (XN1) + f (xN)1
p
194
CHAPTER VII. RESERVEDEPENDENT PREMIUMS
where xk = x0 + kh. Fixing h > 0, letting x0 = 0 (i.e. xk = kh) and writing 9k = 9(xk ), Kk,e = K(xk, xe), this leads to h 9N = hN + 2 {KN,09o+KN,N9N}+h{KN,191+'''+KN,N19N1},
i.e. 9 N=
hN+ ZKN ,ogo +h{KN,lgl+•••+KN,N19N1} 1  ZKNN
(
1.11
)
In the case of (1.8), the unknown yo is involved. However, (1.11) is easily seen to be linear in yo. One therefore first makes a trial solution g*(x) corresponding to yo = 1, i.e. h(x) = h*(x) = (3B(x)/p(x), and computes f o' g*(x)dx numerically (by truncation and using the gk). Then g(x) = yog*(x), and IIGII = 1 then yields f 00 g*(x)dx (1.12) 1= 1+ 'Yo from which yo and hence g(x) and z/'(u) can be computed. u
la Twostep premium functions
We now assume the premium function to be constant in two levels as in Example 1.1, p(r) _ J 1'1 r < v P2 r > v. (1.13)
We may think of the risk reserve process Rt as pieced together of two risk reserve processes R' and Rt with constant premiums p1, P2, such that Rt coincide with Rt under level v and with above level v. For an example of a sample path, Rt see Fig. 1.1.
Rt
V
Figure 1.1
1. INTRODUCTION
195
Proposition 1.10 Let V)' (u) denote the ruin probability of {Rt}, define a = inf It > 0 : Rt < v}, let pi ( u) be the probability of ruin between a and the next upcrossing of v (including ruin possibly at a), and let q(u) = 1  V" (u) Then
1  q(u) + q ( u)z,b(v) p1(v) u = 0<u<v v
0 < u < v. (1.14)
1 + pi (v )  '02 (0) pi (u) + (0, (u  v)  pi (u)) z/i(v ) v < u < oo.
Proof Let w = inf{ t > 0 1 Rt= v or Rt < 0} and let Q1 (u) = Pu(RC,, = v) be the probability of upcrossing level v before ruin given the process starts at u < v. If we for a moment consider the process under level v, Rt , only, we get Vil (u ) = 1  q, (u ) + g1(u),O1( v). Solving for ql (u), it follows that q1 (u) = q(u). With this interpretation of q(u) is follows that if u < v then the probability of ruin will be the sum of the probability of being ruined before upcrossing v, 1  q(u), and the probability of ruin given we hit v first , q(u)z'(v). Similarly, if u > v then the probability of ruin is the sum of being ruined between a and the next upcrossing of v which is pl (u), and the probability of ruin given the process hits v before ( oo, 0) again after a, (Pu(a < oo )  p1(u))''(v) = (Vi2(u  v)  p1 (u))''(v)• This yields the expression for u > v, and the one for u = v then immediately follows. u Example 1 .11 Assume that B is exponential, B(x) = e62. Then
01 (u)
_
0 e .yiu ,,2 (u) = )3 e 72u p1S P2S
1  ~ ery1u p1S 1  Q eryly P1S
where ry; = S  ,Q/p;, so that
q

Furthermore , for u > v P(a < oo ) = 02(u  v) and the conditional distribution of v  Ro given a < oo is exponential with rate S . If v  Ro < 0, ruin occurs at time a . If v  R, = x E [0, v], the probability of ruin before the next upcrossing of v is 1  q(v  x). Hence
196
CHAPTER VII. RESERVEDEPENDENT PREMIUMS
( pi(u) _ 02 ( u  v){ aav + J (1  q(v  x))bedxdx 0 I
1 a e 7i(v x)
eP2,e 7z(uv)
1
_
P16 0 1  a e7iv P16
Se6xdx
1  e 6V Qbe72(uv)
P2 1 
a
e 71v (e(71 6)v  1)
1  p1(71  b)
Ie71v P16
p2be 7z(uv) 1 _
1  e71v a
1  e 7iv P '6
0
Also for general phasetype distributions, all quantities in Proposition 1.10 can be found explicitly, see VIII.7.
Notes and references Some early references drawing attention to the model are Dawidson [100] and Segerdahl [332]. For the absolute ruin problem, see Gerber [155] and Dassios & Embrechts [98]. Equation (1.6) was derived by Harrison & Resnick [186] by a different approach, whereas (1.5) is from Asmussen & Schock Petersen [50]; see further the notes to II.3. One would think that it should be possible to derive the representations (1.7), (1.8) of the ruin probabilities without reference to storage processes. No such direct derivation is, however, known to the author. For some explicit solutions beyond Corollary 1.8, see the notes to Section 2 Remark 1.9 is based upon Schock Petersen [288]; for complexity and accuracy aspects, see the Notes to VIII.7. Extensive discussion of the numerical solution of Volterra equations can be found in Baker [57]; see also Jagerman [209], [210].
2 The model with interest
In this section, we assume that p(x) = p + Ex. This example is of particular application relevance because of the interpretation of f as interest rate. However, it also turns out to have nice mathematical features.
2. THE MODEL WITH INTEREST
197
A basic tool is a representation of the ruin probability in terms of a discounted stochastic integral Z =  f eEtdSt 0 (2.1)
w.r.t. the claim surplus process St = At  pt = EN` U;  pt of the associated compound Poisson model without interest . Write Rt") when Ro = u. We first note that: Proposition 2.1 Rt") = eetu + Rt°) Proof The result is obvious if one thinks in economic terms and represents the reserve at time t as the initial reserve u with added interest plus the gains/deficit from the claims and incoming premiums. For a more formal mathematical proof, note that
dR(u) = p + eR(u)  dAt,
d [R(")  eetu] = p + e [R(u)  eEtu]  dAt . Since R( ;u)  eE'0u = 0 for all u, Rt")  eEtu must therefore be independent of u which yields the result. 0 Let
Zt = eetR(0) = eet (ft (p + eR(°)) ds  At I
Then dZt = e Et (_edt
f t (p + eR°) ds + (p + eR°)) dt + e dt A dA
v Z,, =  eetdSt,
= e_et (pdt  dAt) = eEtdSt. / Thus 0 where the last integral exists pathwise because {St} is of locally bounded variation. Proposition 2.2 The r.v. Z in (2.1) is welldefined and finite, with distribution H(z) = P(Z < z) given by the m.g.f.
H[a] = Ee" = exp
where k(a) _
(aeEt) dt} = exp {f °° k
k
{fa
(y) dy}
13(B[a]  1)  pa. Further Zt a ' Z
as t + oo.
198
CHAPTER VII. RESERVEDEPENDENT PREMIUMS
Proof Let Mt =At tAUB. Then St = Mt+t(/3pBp) and {M„} is a martingale. eEtdMt} From this it follows immediately that {fo is again a martingale. The mean is 0 and (since Var(dMt) = /3PB2)dt)
Var (
Z
'
e'tdMt )
J e eft/3p(B)dt = a2B (1  e2ev). o
/' v
(2)
Hence the limit as v 3 oo exists by the convergence theorem for L2bounded martingales, and we have v
Zv =
v
eEtdSt = f et(dMt + (,3pB  p)dt)
o o

0  f0"
J
a'
0  f 0 oo
eEt
(dMt + (3p$ 
p)dt)
eEtdSt = Z.
Now if X1i X2, ... are i.i.d. with c.g.f. 0 and p < 1, we obtain the c .g.f. of E0° p'Xn at c as
00
00
00
log E fl ea°n X„
n=1
= log 11 e0(av ") _
n=1
E 0(apn). n=1
Letting p = eEh, Xn = Snh  S( n+1)h, we have q5(a) = hic( a), and obtain the c.g.f. of Z =  f0,30 e'tdSt as 00 00 00 lim E 0(apn ) = li h E rc(ae Fnh) = f tc (aet) dt;
n=1 1 n=1 0
the last expression for H[a] follows by the substitution y = aeEt Theorem 2.3 z/'(u) = H(u) E [H(RT(u)) I r(u) < oo] .
u
Proof Write r = r(u) for brevity. On {r < oo }, we have

u + Z =
(u + Zr ) + ( Z  Zr) = e
ET {e
(u + Zr)  f '* eE(tT )dSt] T
e
ET [
R( u)
+ Z`],
2. THE MODEL WITH INTEREST
199
where Z* =  K* eE(tT)dSt is independent of F, and distributed as Z. The last equality followed from Rt") = eEt(Zt + u), cf. Proposition 2.1, which also yields r < oo on {Z < u}. Hence H(u) = P(u + Z < 0) = P(RT + Z* < 0; r < oo) zb(u)E [P(RT + Z* < 0 I)7T, r < oo)] _ O(u)E [H(RT(")) I r(u) < oo] .
Corollary 2.4 Assume that B is exponential, B(x) = e6', and that p(x) _ p + Ex with p > 0. Then
. o€Q/E Ir, (8(p + cu);
V) (u)
aA/Epal Ee 6n1 E +^3E1 / E
1\ E E
1r
Cbp;
E El al
where 1'(x; i) = f 2°° tnletdt is the incomplete Gamma function. Proof 1 We use Corollary 1.8 and get
w(x) fo P + Etdt = g(x) = p +0x
e log(p + Ex)  e loge,
exp {  log(p + Ex)   log p  6x }
pal(p + ex)plE1e6^ J ryo)3 70 = 1 + J p) exp {Ow(x)  Sx} dx x r^ = 1+ ' /E (p + Ex)01'leax dx + 0
f J
= 1+
a
Epo/ E
f yI/ E 1e 6(Y P)/E dy
P (
1+ OEA/E 1e6 P /Er
60/e po/ e
,;,3 )
E E
lp(u) = to foo a exp {w(x)  bx} AX)
acO/E" 1 ePE l
Yo
50 1epolE
(
+ cu); 0)
5(p
E E
pa. RESERVEDEPENDENT PREMIUMS u from which (2.5 The analysis leading to Theorem 2. r (j3/E) In particular.3/E) By the memoryless property of the exponential distribution.3a/ (5 . 13/E). i.V.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments.2) follows by elementary algebra. where V is Gamma(b. of Z is IogH[a] = f ytc(y)dy = e fa (0. The process {St} corresponds to {Wt} so that c(a) or2a2/2 . RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .x) dx e. As an example.e.pa./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E .a) . and the c.f. . From ic(a) = .V < x)]0 + f P(V > p/E ) + eby fv (p/E . it follows that logH[a] = f 1 c(y)dy = 1 f '(pa/(a +y))dy f 0 0 Ey R/E 1 [pa + )3log 8 . /^ u Example 2 .b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.2y +µ ) dy . then {Rt} is the diffusion with drift function p+Ex and constant variance a2.g.200 CHAPTER VII.3.2) follows by elementary algebra. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2. Proof 2 We use Theorem 2. assume that {Wt} is Brownian motion with drift µ and variance v2.13 /E) r (.
. [129] and Harrison [185]. write y* for the solution of the Lundberg equation f3(B[ry *] . Q2/2E). it is also used as basis for a diffusion approximation by these authors. as in the proof of Proposition 2. that the analysis does not seem to carry over to general phasetype distributions. It must be noted. not even Erlang(3) or H3. [357]. or to nonlinear premium rules p(•). Paulsen & Gjessing [286] and Sundt & Teugels [356]. see e. Corollary 2. write Vi* (u) for the ruin probability etc. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. for a martingale proof. The formula (2. se e.g.3.e. Gerber [155]. Emanuel et at. Gerber [157] p. [283]. and since RT = 0 by the continuity of Brownian motion. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) .g. Paulsen [281].3) was derived by Emanuel et at..2 is a special case of a perpetuity.Y*p* W*(u) < ery*u = 0. [282]. 3 The local adjustment coefficient. The solution is in terms of Bessel functions for an Erlang(2) B and in terms of confluent hypergeometric functions for a H2 B (a mixture of two exponentials). and recall Lundberg 's inequality .p*. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1.v. Goldie & Griibel [167]. [129]. Delbaen & Haezendonck [104]. however. Some of these references also go into a stochastic interest rate.4 is classical. Z is normal (p/E. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. A r.d. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case).3 is from Harrison [185].8.i. of the form Ei° p"X" with the X„ i.1) . Further studies of the model with interest can be found in Boogaert & Crijns [71].
let p* be a in (3. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V.3) When trying to extend these results to the model of this chapter where p(x) depends on x. a) = f3(B[a] . If 60 s f 6o.e..C*ef*".'y ( x)) = 0 where r.1. Then we have the following lower bound for the time for the reserve to go from level u to level u + v without a claim: w(u + v) .E). then log u (u) In the proof as well as in the remaining part of the section .w (u) J dt > c(3)eeu v 1 p(u+ t) . x>0 (3.i)eex. the function y(x) of the reserve x obtained by for a fixed x to define y(x) as the adjustment coefficient of the classical risk model with p* = p(x).1.>o 7(x) > 0.ap(x).*(u) .1 Assume that for some 0 < 5o < oo. (3.1 ). it holds that f3[s] T oo.2) such that p(x) < c(. Then lim sup u>oo u and e E''p(r) + 0.1) .202 CHAPTER VII.e.log '(u)/u < ry*(1 . i. and (for simplicity) that inf p(x) > (3µs . Let y* < So. oo for all E > 0. (x. (3.4) we assume existence of y(x) for all x. The intuitive idea behind introducing local adjustment coefficients is that the classical risk model with premium rate p* = p(x) serves as a 'local approximation ' at level x for the general model when the reserve is close to x. e(1o+e)2 (x ) u > 00. choose uo such that p( x) > p* when x > u0E. 1) and for a given E > 0. When u > uo. as solution of the equation n(x. For the last asssertion . i. The steepness assumption and p(x) + oo ensure 'y(x) * So. B(x) > C(2)e(ao+f)x for all x. log ?i(u) < < 00 JO . we will use the local adjustment coefficient 'y(x). choose c(. c(. a first step is the following: Theorem 3 . x * oo.5) which implies inf. as will hold under the steepness assumption of Theorem 3. and that p(x) * oo. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. obviously O(u) can be bounded with the probability that the Cramer Lundberg compound Poisson model with premium rate p* downcrosses level uE starting from u .. Proof of Theorem 3. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.
(3. I. The rest of this section deals with tail estimates involving the local adjustment coefficient. Theorem 3 . . by cU2. or (b) Condition 3. If p(x) = pis constant . then Rte) = CRtie for all t so that V). Condition 3.2 is also an approximation under appropriate conditions.13 is a technical condition on the claim size distribution B. (u) = O(u/e).0 are the same. 2.3 Assume that either (a) p(r) is a non decreasing function of r.1 only presents a first step. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. Then . (3. UJU > x cannot have a much heavier tail than the claim U itself.2). 3.(u) > so. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(.3 to be reasonably precise and use e` (u) as approximation to 0 (u). the asymptotics u * oo and c . However. which essentially says that an overshoot r. Theorem 3. 3) = (1 .. Bucklew [81]). noting that in many cases the constant C is close to 1. ruin will occur if the claim is at least u + v. Then limelog l/ie (u) = I(u). the result is not very informative if bo = oo.7) CIO Remarks: 1. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e.' (u) < eI("). and in particular. {Rte)} defined as in (1. The form of the result is superficially similar to the CramerLundberg approximation.v.13 below holds. For e > 0. u Obviously.6) The second main result to be derived states that the bound in Theorem 3.e.ea°/(ecf1)).4)eE" Given such an arrival.3. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U.g. one can then assume that e = 1 is small enough for Theorem 3. The slow Markov walk limit is appropriate if p(x) does not vary too much compared to the given mean interarrival time 1/0 and the size U of the claims.
7) is only captures 'the main term in the exponent' but is not precise to describe the asymptotic form of O(u) in terms of ratio limit theorems (the precise asymptotics could be logI(u)e1(U) or I(u)"e_I(u).bx} dx 1+0.bu}. RESERVEDEPENDENT PREMIUMS 4.4 Consider again the exponential case B(x) = eax as in Corollary 1. Then y(x) = b . 5. it is formally needed only for Theorem 3.204 CHAPTER VII. we show how to rewrite the explicit solution for ti(u) in Corollary 1.(3/p(x).(x) dx.bx}]o + b /' oo exp low (x) .exp {/33w(u) . As typical in large deviations theory.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) . 3a Examples Before giving the proofs of Theorems 3.3. rather than eI(u)).8 in terms of I(u) when the claims are exponential: Example 3 . However. 3.1 + b f e.2. the logaritmic form of (3.bx} dx = 1 + J0 dodx(x) exp {.3. u . J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) . and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.8.(iw(x) .bx} dx fo 00 1 + [exp {/(3w(x) .6). we get = 1+ J" AX) exp {(3w (x) . One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3. say.bx} dx . we consider some simple examples. First.
(3. note first that the appropriate slow Markov walk assumption amounts to u.fa 7(x+u)dx/Edy o The analogue of (3.I ( v )dy fo +u) dxdy .0.e. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e. IE(u) = I(u)/e. Be = e log U000 e.10) where AE = e log 000 e.7) follows. The appropriate definition of the local adjustment coefficient 7(x) is then as the one 2p(x)la2(x) for the locally approximating Brownian motion. It is well known that (see Theorem XI.fo 7(x)dx/Edy f . Similarly.fo 7(x) dx /E dy > av 'yo /Edy = E (1 .. we get r 00 e.5 Assume that {Rt} is a diffusion on [0.2. oo) with drift µ(x) and variance a2 (x) > 0 at x.2(X) = ev2(x) so that 7e(x) = 7(x)/e.3.3 in the particularly simple case of diffusions: Example 3. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.9) yields e log .1.BE. 70 > 0 such that 7(x) < 7o for y < yo.1/8 . BE * 0.f y(x)dxd y If 7(x) is increasing .ev 0 O /E) J0 70 70 Yo This implies lim inf A. > lime log e = 0 and AE * 0.3. (u) = I(u) + AE . THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . and (3.2.10 or Karlin & Taylor [222] pp. Choosing yo.5) is infx>o 7(x) > 0 which implies that f °O . u .fory(x+u)dxdy ( 3. In particular. For Theorem 3. in the definition of AE converges to 0. and (3.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. (3..I ( u) fool. 3. ry(x /b I u o e f0 °° e  e. 0. (X) = µ(x). applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3.9 ) 11000 eI(v)dy f000 e.
E+o e*O By (3. . the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large.7o C 15 I I.+1 (u) = o( 1). this leads to (3. Ignoring 1/5 in the formula there. we have 5 > 7o and get lim inf AE > lime log e . 7(x) is typically not explicit./3 1 AX dx.0/e.206 CHAPTER VII.e. > . . .. + Gq(u) + o(G9(u))• Gi (u) It should be noted . G. G..Q/p*.(u) oo. .0. Nevertheless .6 Assume that B is exponential with rate S.6) exactly as in Example 3. Then the solution of the Lundberg equation is y* = b . (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo.1 3. Thus 7e(x) _7(x)/e and (3. the slow Markov walk assumption means 5E = b/c.6/p* so that u 1 I (U) = bu . ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3. the results are suggestive in their form and much more explicit than anything else in the literature. 0.0) = 0. As in Example 3. . 0 Now (3. Further.7) follows just as in Example We next investigate what the upper bound / approximation aI (°) looks like in the case p(x) = a + bx (interest) subject to various forms of the tail B(x) of B.5) and 7* = 5 . lim sup Af < lim sup c log(1 ..5.5. _ .5 for risk processes with exponential claims is as follows: Example 3 . I.. I(u ) = G1(u) + .10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. Of course. that the interchange of the slow Markov walk oo is not justified and in fact. however . G.4. RESERVEDEPENDENT PREMIUMS The analogue of Example 3.5. so our approach is to determine standard functions Gl (u).
:. I(u) Pt. .1) leads to .12) with y > 1.s)C' f "o o as s T S. c4 = c2b 1/'/(1 . (3.y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . and hence (3.11) that b[s] * co as s f S and hence 7* T S as p* + oo. 1.4) or gamma distributions. .1) leads to (S7T N Ocp a. phasetype distributions (Example 1..c2 Su a dx ) Su a<1 Su . u(logu + r7loglogu). Here B[s] is defined for all s and B[s] . This covers mixtures or convolutions of exponentials or.11) with a > 0.3cse7*I7(77) .Y . y = 2 if B is uniform on (0.ry*°p*.1/a). x T 1. ry* loge*+ g7loglogp*.C2p* C2 = (3clr( a))11'. More precisely. For example. if the phase generator is irreducible ( Proposition VIII.x)n1. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . the typical case is a = 1 which holds .7 Assume that B(x) .clxale5x 207 (3. Hence (3.8 Assume next that B has bounded support. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b.g. 2. It follows from (3.3.1/k). say 1 is the upper limit and B(x) . fu I(u) Su . u Example 3 . e.cs(1 . 77 = 1 if B is degenerate at 1. in the phasetype case .8). 1) and 17 = k + 1 if B is the convolution of k uniforms on (0. more generally.
I (u) c8u log u 0 where c8 = 2/c7.c8 log ..g.sp(u).14) for the m .4). of U1 + v . This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u . The assumption implies that ru(t) .r„(Ti). By convexity of the m . assume that B(x) CO x2/2c7. Hence for u<V.13) We get b[s] .4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3.15) Proposition 3.•.ru(TI)) .f. of the increment in a small time interval [0. h]. (3. 1 = E.(t))dt.e7o ( u)(ul+u r.12).Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u). 7 * .r^. 1 0 3e.(T1)) > Ee7o(u)(ul+vr»(Ti)). x f oo .10 Assume that p(x) is a nondecreasing function of x.208 CHAPTER VII.9 As a case intermediate between (3. 3b Proof of Theorem 3. one could also have considered the increment ru (T1) . (3.1) . (3.11) and (3. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.3 (B[s] . RESERVEDEPENDENT PREMIUMS Example 3 . h 10.css 2%rc7eC782/2. (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3.u . .4) is the formula h logEues ( Rhu) . this is only possible if 7o(v) 2 7o(u)• .2 We first remark that the definition (3.Ote7o( u)(u.B[7o (u)] . g.log p*.f.u is a nondecreasing function of u. ec78)2/2c7 dx C7 .1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .
THE LOCAL ADJUSTMENT COEFFICIENT For (b).(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .2 in terms of 7o. Separating after whether ruin occurs at the first claim or not.17) shown for n and let Fu(x) = P(U1 + u .u[70(u)] fo eyo(x)dx .1) .ru(T1) < x). Then (u) < efo Yo(x)dx. Assume (3. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.3.u > tp(u). note that the assumption implies that ru(t) .11 Assume that p(x) is a nondecreasing function of x. u We prove Theorem 3.17) from which the theorem follows by letting n + oo.Fu(u ) + J ^(n)(u . it is easily seen that fu x7o(y)dy < xyo (u). we obtain „I. Hence 1 = EeYo(u)(U1+uru(T1)) < E. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3. this is only possible if yo(u) > 7(u).7o (u)p(u)• Since (3. the case of 7 then follows immediately by Proposition 3. fa 7o(y)dy < u7o(u) < xyo (u) for x > u.10(b): Theorem 3.e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] .es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.4) considered as function of 7 is convex and 0 for y = 0.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn). Hence „/.(n+l) (u) 1 . (3.x)Fu(dx) 00 U efo J = o (y) dYF (dx) )+f I 11 /' / 00 e f oFu fu dx) + of u :7o(Y)dYFu(dx) 00 J u 1 l` Considering the cases x > 0 and x < 0 separately. Also.
C*e where the first equality follows by an easy scaling argument and the approximation by (3.3/e and U. the probability that ruin occurs in the CramerLundberg model with p* = pn.E (u/n). op*. 3.n = sup p(x). the value of {R(E)} at the time of downcrossing is < unl.n inf n uk1. P k. by €U=. Let Ck.2). pk n = uk_l.10(a) for some of the inequalities.210 CHAPTER VII.3 is required. To this end. (un2.nbe C*.n u k}1. For these reasons.n) > k =1 II v ^k n.n) pn niE (u /n) n n_1 n. ryk. Also. 0.: y(u).E (u/ n) Y'E (un . 3c Proof of Theorem 3.n.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. 0. 0 It follows from Proposition 3.n.2.n (starting from u/n) without that 2u/n is upcrossed before ruin.12 lim sup4^o f log O. yo(u) appears more difficult to evaluate than y(u). W O .3). Further.I.e (u) = v'.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x . (3.. and here it is easily seen that yo(u) . The probability of this is at least n n. in accordance with the notation i/iE (u).E ( u/n) ^•e.2. However. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3. (u) < I(u)..E (u) denote the ruin probability for the classical model with 0 replaced by . we have chosen to work with y(u) as the fundamental local adjustment coefficient. y* evaluated for p* = Pk.10(b ) that the bound provided by Theorem 3.n. Lemma 3.n AX). in .. define uk.11 is sharper than the one given by Theorem 3.15). x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3.3). (u)... for either of Theorems 3.E (u/n) Now as e . Y*u /E..x/n.n. we used also Proposition 3. let Op*.n = ku. resp.n <Z auk}l. Proof For ruin to occur. given downcrossing occurs.n) must first downcross unl. {RtE)} (starting from u = un. and.n so that n.
V < oo such that (i) for any u < oo there exist Cu < oo and a (u) > supy <„ 7(x) such that P(V > x) < Cue. we need the following condition: Condition 3. B(x) (3. /' (u/n) 'T nk.n = sup ?'(x).n .E (u/n) OP +^p•. since ry' is an increasing function of p'.. 11 Theorem 3. 40 Combining with the upper bound of Lemma 3.n <X<Uk. v. ne7k.a( u)z.! (u/n) n n m 7k.nk=1 limsupelogv).n.12 completes the proof.E (urn) < \ *I.log Ck.3. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). so that Theorem 3.19) .. in obvious notation one has tC (x) = y(x)/e.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).nu /fn( 1 Ck  e. . (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof.13 There exists a r.n cE (2u/n) Ck ne7k. Indeed . 211 Clearly.n + 0(1). for all x .E (u/n) Op•. uk_1.F (2u/n).e.7k. (3. THE LOCAL ADJUSTMENT COEFFICIENT particular. k=1 k=1 n u _ nE7 k.nu /En) o(1)). also ryk. 0 with n and u fixed.i. 3 now follows easily in case (a). i. In case (b). *p•..18) (ii) the family of claim overshoot distributions is stochastically dominated by V..nu/en(1 + where o(1) refers to the limit e . It follows that n log V'C (u) k =1 log Ypk.
n V.E(E) (u..212 CHAPTER VII. u/n) < oo] . The probability of ruin in between two downcrossings is bounded by Epp .nu /EnE [e71.( . For E2. let v < u and define T(E) (u. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. Then the standard Lundberg inequality yields El < E?. P (T(E) (u.nu/En0(1) . v ) = v . Write EO.^(E) (u. u /n) < oo] l = = < E [OE (u/n . Then Y'E (u) ^(E) (u.V) = e71 nu/Eno(l) (using (3. u /en 0(i) _n so that E2 < e2ryl nu/En0(1).v. N with EN < 1 = infx>2u/nA(x) = 0(1).18) for the last equality).EV) = e. .^'' = E [ .QEU 1 . u/n)) I T(E) (u. u/n) < oo) .EV) = EiI 1 . u/n) < oo) EV). v ) = inf { t > 0 : R(c ) < v R) = u } .R<) (u v).E (2u/n . ) (u u /n)) . (u/n . infx>2u /n P(x) . RESERVEDEPENDENT PREMIUMS To complete the proof. T() (u. u/n) < oo] E [OE (u/n . V < u/En] + P(V > u/En) (u/En . (u/n . u/n)) . where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest.of:>2 in n(x).5) and the standard formula for b(0).2y 1 ' .eV) • P (T(E) (u. (3.. (R.n < ery1.E (u/n ..EV) = El + E2. T(E) (u. Ei + E2 < e71.E (0) cf..1 n.
[89].J y(Rs)dR.13. one can in fact arrive at the optimal path by showing that the approximation for 0(u.1.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof. s) as in (3. Typically. the results are from Asmussen & Nielsen [39]. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. Similarly. Comparing these references with the present work shows that in the slow Markov walk setup. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.)ds + Y(R2.7) then comes out (at least heuristically) by analytical manipulations with the action integral.) = exp .)Ui } .20) (with ic(x.T) = P „(info<t <T Rt < 0) via related large deviations techniques.=1 J An approximation similar to (3. T) is maximized over T by taking T as the time for (3.t.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. the approximation (3.4) and the prime meaning differentiation w. 0 and b T 00 are interchangeable in the setting of [89]. u Notes and references With the exception of Theorem 3.7(x)) (3.3EU) (3. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S .21) to pass from u to 0. 0 ) (= p(x) . the rigorous implementation of these ideas via large deviations techniques would require slightly stronger smoothness conditions on p(x) than ours and conditions somewhat different from Condition 3. Whereas the result of [122] is given in terms of an action integral which does not look very explicit. whereas the most probable path leading to ruin is the solution of r(x) _ k (x. s). l o JJJ o .3. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e .J r(Rs)p(R. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali. . Djehiche [122] gives an approximation for tp(u.21) (the initial condition is r(0) = u in both cases).g. it might be possible to show that the limits e .r. the risk process itself is close to the solution of the differential equation r(x) _ r (x. Bucklew [81]).
For different types of applications of large deviations to ruin probabilities .214 CHAPTER VII. We should like. however..3. see XI. e.g. to point out as a maybe much more important fact that the present approach is far more elementary and selfcontained than that using large deviations theory. the exponential distribution ). RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . .
This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 .Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. on Eo = E U {A} where A is some extra state which is absorbing. Typically. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. oo) is said to be of phasetype if B is the distribution of the lifetime of a terminating Markov process {Jt}t>o with finitely many states and time homogeneous transition rates. a terminating Markov process {Jt} with state space E and intensity matrix T is defined as the restriction to E of a Markov process {Jt}o<t<. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . A distribution B on (0. P. and not in other cases. then the problem may admit an algorithmic solution involving a reasonable degree of computational effort if one allows for the more general assumption of phasetype structure. refers to the case Jo = i. that is. if a problem can be solved explicitly when the relevant distributions are exponentials. We often write p for the number of elements of E. if v = (vi)iEE is a probability distribution. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory.1) is 'Here as usual . More precisely. Note that since (1.
and the phasetype distribution is the lifetime of a particle with constant failure rate /3. Equivalently. E = {i. an exponential distribution with rate parameter . tij > 0 for i 54 j and EjEE tij < 0 . and we have t = Te. t1 = /3. j.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. Then a = a1 = 1.1 Suppose that p = 1 and write .3. 0 2this means that tii < 0. the rows sum to one which in matrix notation can be rewritten as t + Te = 0 where e is the column Evector with all components equal to one. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. a. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. k}. i. T is a subintensity matrix2.216 CHAPTER VIII. that is.1 The phase diagram of a phasetype distribution with 3 phases. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. Here are some important special cases: Example 1 . (1. B(t) = Fa(^ < t ).0 = t11.e. In particular. The initial vector a is written as a row vector. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. i.e. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. We now say that B is of phasetype with representation (E. T) (or sometimes just (a. C is the lifetime sup It > 0 : Jt E E} of {Jt}.2) The interpretation of the column vector t is as the exit rate vector.
.. 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •. 0 S 6 Example 1.2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p. .. 00)) S s o . PHASETYPE DISTRIBUTIONS 217 Example 1. .3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51. 0 0 0 0 S 6 ... a = (1 0 0 ..1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. 0 •...1. 6..x i=1 Thus E _ Si 0 T 0 S2 0 0 ... 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 . . the EP distribution may be represented by the phase diagram (p = 3) Figure 1.. so that the density is P E ai6ie6. 0 SP 0 and the phase diagram is (p = 2) ..2 corresponding to E = {1. .. p}.. .
1 tP1 1 Figure 1.aeTxe. Then for i .t2 yt bP. the Erlang distribution is a special case of a Coxian distribution. p:. the backwards equation for {Jt} (e. see A. j E E. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. .3 .1)"n! aT"e.4 For example. 5 Let B be phasetype with representation (E.d. [APQ ] p. a. (c) the m.f. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. 36) yields s d.g. MATRIXANALYTIC METHODS Figure 1. i.3 0 Example 1 .4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature.f is B (x) = 1 . ds^ = ds' = ttlaj + tikpkj. (b) the density is b(x ) = B'(x) = aeTxt. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects .218 CHAPTER VIII. Then: (a) the c.e.g. T). Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . the restriction of P8 to E. The basic analytical properties of phase type distributions are given by the following result . dp. Theorem 1 .
tij / . we i w. Then h tit ti + ti3 h j .5) as hi(tii + s) = ti  t ij hj. i. j#i jEE tijhj + his = ti. After that. in which case the time to absorption is 0 with m . the solution is P8 = eT8.g. (1) n+1n!aT .p.. ti/ . . Rewriting ( 1.T) 't = (.f.f.s) is the m . we arrive once more at the stated expression for B[s]. define hi = Eie8S. Part (d) follows by differentiating the m. 1. B(n)[0] = _ Alternatively.B(x) = 1'a (( > x) = P.tii is the rate of the exponential holding time of state i and hence (tii)/(tii .f. and since obviously P° = I. of the initial sojourn in state i.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute).s I .6) . the rule (A. Since 1 . h = (T + sI)1t. i. this means in vector notation that (T + sI)h = t. or w. for n = 1 we may put ki = Ei( and get as in (1.T) 1t.5) ki = 1 + tii L jj:Ai tii (1.tii tii .12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . d" dsn a (.g. = aPxe. hj .f.tii we go to A..p.n lt .g. Alternatively. PHASETYPE DISTRIBUTIONS 219 That is.5) Indeed . (Jx E E) = this proves (a).e.1 ) n +l n ! a (s I + T ) . and since b[s] = ah.g.s j# tii i (1.1. and (b) then follows from 1: aipF.n1t = (1)nn!aTn1Te (1)nn! aTne. For (c). d8 P8 = TP8.tii and have an additional time to absorption either go to state j which has m .
Consider for example 3 9 a= (2 2).h.220 CHAPTER VIII. there are some examples where it is appealing to write T on diagonal form. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x . One obvious instance is the hyperexponential distribution.7) the diagonal form of T is 9 9 1 9 T 10 7 10 70 1 10 6 10 7 0 70 9 1 10 where the two matrices on the r. making the problem trivial.s. This implies that we can compute the nth moment as (1)"n! aT "e 1"n! 1 1 22 9 9 10 70 7 1 10 10 1 9 +6. are idempotent."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly. another the case p = 2 where explicit diagonalization formulas are always available. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. Example A3. T= 2 111 so that 2 2 Then (cf.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer. 0 Example 1. see the Appendix.
4.T). then the matrix m. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I). < 1.e 11BIJ = 1laDD < 1. i. • The phasetype distribution B is zeromodified.e a mixture of a phasetype distribution with representation (a/llall. i. This is the traditional choice in the literature.7) Proof According to (A.7 If B is phasetype with representation (v.T) with weight hall and an atom at zero with weight 1 .29) and Proposition A4. hail = E=EE a.hall. where the initial vector a is substochastic. and in fact one also most often there allows a to have a component ao at A. 5 and serves at this stage to introduce Kronecker notation and calculus (see A. 00 B[Q] = J0 f veTxteQx dx = (v ® I) ( f° eT x edx I (t I) (v (& I) ( (T ®Q)xdx f o" e o )( t ® I) _ (v ® I)(T ® Q)1(t ® I).11aDD. 0 Sometimes it is relevant also to consider phasetype distributions.g. T) is then defined to be oo on a set of probability 1. There are two ways to interpret this: • The phasetype distribution B is defective.1. PHASETYPE DISTRIBUTIONS 1 10 7 10 221 9 6 70 7 9 10 2 +e 6x (1 11 2 2 35ex + 18e6x 35 The following result becomes basic in Sections 4. (1. a random variable U having a defective phasetype distribution with representation (a. .f.4b for definitions and basic rules): Proposition 1. or one just lets U be undefined on this additional set.
(1. T). cf. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve .. the conditions of Proposition 1. v. the Erlang case). but in many practical cases. cf.g. All material of the present section is standard. Here is a sufficient condition: Proposition 1. assume that T is irreducible . here k = p1). MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. where C.4c). Schmidt & Teugels [307] and Wolff [384]. . let v.F. No satisfying .f. In older literature. the result follows (with C = (ah)(ve)).g. distributions with a rational m. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. Then the tail B(x) is asymptotically exponential. 0 Of course.8). See in particular the notes to Section 6. 2. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. but todays interest in the topic was largely initiated by M. x * oo. B(x) .222 CHAPTER VIII. Using B(x) = aeTxe . Schmidli. The Erlang distribution gives an example where k > 0 (in fact.1 of the Appendix. one has k = 0. not only in the tail but in the whole distribution. Lipsky [247]. Neuts. 1. the text is essentially identical to Section 2 of Asmussen [26]. In Proposition A5. and we have eTx . i is real and positive.hve7x.. B[s] = p(s)/q(s) to be phasetype: the density b(x) should be strictly positive for x > 0 and the root of q(s) with the smallest real part should be unique (not necessarily simple. see his book [269] (a historical important intermediate step is Jensen [214]). Rolski. but the relevant T is not irreducible. (or Laplace transform) are often used where one would now work instead with phasetype distributions.Ce7'.8 Let B be phasetype with representation (a. h can be chosen with strictly positive component.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail.8) Proof By PerronFrobenius theory (A.q be the eigenvalue of largest real part of T. we give a criterion for asymptotical exponentiality of a phasetype distribution B. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m.f. Notes and references The idea behind using phasetype distributions goes back to Erlang. let . Example A5. 77 > 0 and k = 0.
.r. but is in part repeated below. with common distribution B and define4 U(A) = E# {n = 0. the jumps of the j(k) and the it } k) to the next J( k+l) A jump jumps corresponding to a transition from one Jt 4Here the empty sum U1 +.1 Consider a renewal process with interarrivals which are phasetype with representation (cr. known. as the lifetimes of items (say electrical bulbs) which are replaced upon failure. if U is absolutely continuous on (0... however..+UnEA).. and U(A) is then the expected number of replacements (renewals) in A. +UnEA} 00 = EEI(U1 +. or the density is available ) is.2. we denote the density by u(x) and refer to u as the renewal density. .1 of the Appendix. we refer to U as the renewal measure. .. For this reason. U2.i.. but nevertheless.g. If B is exponential with rate 0. Jt={Jt?ul}. Lebesgue measure. JtJt1) Then { 0<t<U1 . is Markov and has two types of jumps . what is the smallest possible dimension of the phase space E? 2 Renewal theory A summary of the renewal theory in general is given in A. A related important unsolved problem deals with minimal representations: given a phasetype distribution . oo) w.... the renewals form a Poisson process and we have u(x) = 0..f. n=O We may think of the U. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2.. + U0 is 0 . Let U1. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. be i.: U1 + .d. (2..1.T). . The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions.1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together. U1<t < U1+U2. Then the renewal density exists and is given by u(x) = ae(T+ta)xt.t.
B is defective .1. Equivalently. the density is veTxt = B(x)/µB. 2.224 CHAPTER VIII. Proof Just note that { it } is a governing phase process for the lifetime..U1 U3 U2 U3 U4 Figure 2.1) remains valid for that case. However..IIBII which is > 0 in the defective case. see Fig. and let µB = aTle be the mean of B.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. Hence ( 2. u Returning to nonterminating renewal processes . and the jumps of the first type are governed by T. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter .e. . define the excess life e(t) at time t as the time until the next renewal following t. fi(t) U2 U1 . Then: (a) the excess life t(t) at time t is phasetype with representation ( vt.1 Corollary 2. that is. This is defined as U1 + .3 Consider a renewal process with interarrivals which are phasetype with representation (a. Hence the intensity matrix is T + ta. the lifetime of the renewal process. i. and the distribution of Jx is ae ( T+t«)x. i. which is phase type with representation (v. IIafl < 1.T + ta). The renewal density at x is now just the rate of jumps of the second type.1) follows by the law of total probability.T) where v = aT1 /µB. as the time of the last renewal. is the first k with Uk = 00.T) where vt = ae (T+ta)t .e. . since Uk = oo with probability 1 . and hence ( 2. + Uit_1 where s. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj . Corollary 2. Then the lifetime is zeromodified phase type with representation (a. T). which is ti in state i. this is welldefined. (b) £(t) has a limiting distribution as t * oo.T). u The argument goes through without change if the renewal process is terminating.
= qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. cf. The renewal density is then aeQtt = (al a2) ( 7i 7"2. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB.T) where vt is the distribution of it which is obviously given by the expression in (a). u Example 2 . (ii) First check the asserted identity for the density: since T.e.1. the unique positive solution of ve = 1.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. we first compute the stationary distribution of Q. The time of the next renewal after t is the time of the next jump of the second type. (2. The formulas involve the matrixexponential of the intensity matrix Q = T + to = ( tll + tlal t12 + t2al tlz + tlaz _ q1 ql t22 + t2a2 q2 q2 (say).2) v(T + ta) = 0. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it.6.4 Consider a nonterminating renewal process with two phases. i. T1 and eTx commute. hence e(t) is phasetype with representation (vt. According to Example A3. Al.2. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2.q2.) ( t2 ) .e.
4 yields the renewal density as u(t) = 2 (1 . A = 25. t1B 0 Example 2 .t2) .4 yields the renewal density as u(t) = 5152 e.6 Let B be hyperexponential. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .5 Let B be Erlang(2).t2) 1 + eat (a17r2 .tl) 7r2t2 + eat (a17r2 . and Example 2. MATRIXANALYTIC METHODS e. Then Q= 0 55 )+(1o)=( j ad ).a27rl) (tl .52a1. )t (51 .226 CHAPTER VIII.52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221]. .a27r1) (t1 . The present treatment is somewhat more probabilistic.(biaz + aza.`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . and Example 2. Hence 7r = (1/2 1/2).e2bt) 13 Example 2 .
we see that the ladder height Sr+ is just the residual lifetime of the Markov process corresponding to the claim causing upcrossing of level 0.3. with 0 denoting the Poisson intensity. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a).2. G+(. T(0) < oo) the ladder height distribution and M = supt>o St. use the phasetype representation of Bo. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. a+j.e. Thus the total rate is tip + tia+. add a more selfcontained explanation of why of the phasetype structure is preserved. . Then: (a) G+ is defective phasetype with representation (a+.3. Corollary 3.) = F(ST(o) E •.1 on the next page. Considering the first. For (b). Here we have taken the terminating Markov process underlying B with two states. and M is zeromodified phasetype with representation (a+. T) where a+ is given by a+ = . i. THE COMPOUND POISSON MODEL 227 3 The compound Poisson model 3a Phasetype claims Consider the compound Poisson (CramerLundberg) model in the notation of Section 1. r(u) the time of ruin with initial reserve u.f3aT1. T + to+). Next. Since the results is so basic.1 Assume that the claim size distribution B is phasetype with representation (a. Within ladder steps. and rewriting in matrix form yields the phase generator of {my} as T + ta+. We asssume that B is phasetype with representation (a. (b) V. the transitions are governed by T whereas termination of ladder steps may lead to some additional ones: a transition from i to j occurs if the ladder step terminates in state i. however. marked by thin and thick lines on the figure.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. {St} the claim surplus process. The stars represent the ladder points ST+(k). itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. T). Corollary 2. which occurs at rate ti. T). we shall.p. the Markov processes representing ladder steps can be pieced together to one {my}. and if there is a subsequent ladder step starting in j whic occurs w. B the claim size distribution. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M.i. The essence is contained in Fig. 3. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. cf.
. 3e3x + ..228 CHAPTER VIII.1 This derivation is a complete proof except for the identification of a+ with .M {mx} ST+(2)  S . MATRIXANALYTIC METHODS t .QaT1.Q = 3 and b(x) = . 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).3.2 Assume that . This is in fact a simple consequence of the form of the excess distribution B0.7)diag so that a+ = QaT 1 = 3 ( 3 2 2) 0 3 9 2 14 7 2 11 2 T+ta+ = 3 0 07/+( 7I \ 2 14 .1 . T = (3 . see Corollary 2..t t d kkt S. 0 Example 3. Figure 3.
4.and Markovmodulated models. For further more or less explicit computations of ruin probabilities. (a) G+ is of phasetype with representation (a+. In the next sections.j). the discussion around Fig. see Section 6. and the argument for the renewal case starts in just the same way (cf. with A denoting the interarrival distribution and B the service time distribution. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). The result carries over to B being matrixexponential.2 are taken from Gerber [157].6. For an attempt. see Stanford & Stroinski [351] . T) for some vector a+ = (a+. 3. the duality result given in Corollary 11. That is. cf. 4 The renewal model We consider the renewal model in the notation of Chapter V. We assume p = PB/µA < 1 and that B is phasetype with representation (a. Fig.1 can be found in Neuts [269] (in the setting of M/G/1 queues.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3. T). this was obtained in Section 3. his derivation of +'(u) is different. so that as there 229 9 9 e(T+ta+)u 1 9 e_u 10 70 10 70 7 10 Thus 1 7 9 10 ) + e6'4 ( 10 10 . We shall derive phasetype representations of the ruin probabilities V) (u).1 which does not use that A is exponential) by noting that the distribution G+ of the ascending ladder height ST+ is necessarily (defective) phasetype with representation (a+. we encounter similar expressions for the ruin probabilities in the renewal. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. see Shin [340]. THE RENEWAL MODEL This is the same matrix as is Example 1.6).T). It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially.1 In the zerodelayed case. 3. where a+ is the (defective) . The parameters of Example 3.1): Proposition 4. if we define {mz} just as for the Poisson case (cf. For the compound Poisson model. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration).4.
with intensity matrix Q given by Q = T + to+. Since the conditional distribution of my given T1 = y is ae4y. the calculation of the first ladder height is simple in the stationary case: Proposition 4. G(') = pBo.230 distribution of mo. 4.T). Then {m.3 a+ satisfies a+ = V(a+).*} from {St+y . the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). obviously mo = m. Hence by Theorem 11.Sy} in the same way as {mx} is defined from {St}. Then . CHAPTER VIII.1). Nevertheless.*'} is Markov with the same transition intensities as {mx}. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. B0 is phasetype with representation (aT1/µa. but with initial distribution a rather than a+.1) Proof We condition upon T1 = y and define {m. the Palm distribution of the claim size is just B. cf. where B0 is the stationary excess life distribution corresponding to B.3.6.2 The distribution G(s) of the first ladder height of the claim surplus process {Ste) } for the stationary case is phase type with representation (a(8). MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. But by Corollary 2.1. which for numerical purposes can be solved by iteration. In fact.T). (4.T)• Proposition 4. We have now almost collected all pieces of the main result of this section: Theorem 4 . (c) {mx } is a (terminating) Markov process on E. where a(8) = aT1/PA. Fig. Also. Proof Obviously.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a.5. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). The key difference from the Poisson case is that it is more difficult to evaluate a+.
. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. THE RENEWAL MODEL 231 .1) and a(8) _ aT.^(8)(u) = a ( 8)e(T+ta +) xe.. the maximum claim surplus for the stationary case has a similar representation as in Proposition 4.1 by noting that the distribution of mo is a+..3) Proof The first expression in (4.3).1/pA.•.3) (defined on the domain of subprobability vectors . It remains to prove convergence of the iteration scheme (4.0. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . i. Furthermore . Hence ^p(. only with initial distribution a(*) for mo. i y ^ T1= y `•r Figure 4. (4.. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case.1 . a+ can be computed by iteration of (4. a+2) = ^p (a+l)) .2 ) follows from Proposition 4.1(b).M. and that this is given by Proposition 4.^(u) = a+e ( T+ta+)xe. by a+ = lim a +n) where a+°) . thus .4.2. I {mx} . a+l ) = cp (a+°)) . . (4. .2) where a+ satisfies (4.0) is an increasing function of /3.1). In particular .e.
a+ ) exists . s ¢ sp(T).5) Since s $ sp(T).5) yields h = (sI . In that case. Similarly. To prove the converse inequality.g. Then (4. and hence we may assume that h has been normalized such that ahA[s] = 1. with B[s]. and let &+". MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn.T1. . 0 0 We next give an alternative algorithm. Obviously.4.4) whenever EeR(S)U < oo. n) &+n) T a+. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. Proof Suppose first Qh = sh. It follows that n1) so that on Fn the feedback to {mz} after each ladder step cannot exceed &+ a+ n) < a f ^ e(T+ t&+ 1))YA(dy) o < a is e(T+t«+1')YA(dy) _ w (a+1 )) = a+n).5 Let s be some complex number with k(s) > 0. To this end.f. the corresponding right eigenvector may be taken as (sI .} can contain at most n . Then F[s] = a(sI . For n = 0. the normalization is equivalent to F(s) = 1. Assume the assertion shown for n .1. Then each subexcursion of {St+Tl .ST. limn4oo a ) < a+. However..T)1t.T)It. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes).) = P(mTl = i. F[s] being interpreted in the sense of the analytical continuation of the m. (4. Fn ).2.4).4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). both quantities are just 0 . Theorem 4. 7+ ].1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). let F be the distribution of U1 . Thus by (4. (4. Thus .232 CHAPTER VIII. this implies that ahA[s] # 0. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. we use an argument similar to the proof of Proposition VI.P[s] = A[s]B[s].T)'t • A[s] (4. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . so to complete the proof it suffices to show that &+ < a+) for all n.
. D that with columns p1 hl. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . the matrix Q in Theorem 2. . Q = CD1 where C is the matrix with columns hl.4.type with representation (a+.T)It. ..... explicit expressions for the ruin/ queueing probabilities are most often derived under the slightly more general assumption that b is rational (say with degree d of the polynomial in the denominator) as discussed in Section 6. The roots are counted and located by Rouche' s theorem (a classical result from complex analysis giving a criterion for two complex functions to have the same number of zeros within the unit circle ). As in Corollary 4.. (4..T)lt = sh. Q has diagonal form d d Q = dpivi®hi = dpihivi..1 has the d distinct eigenvalues . This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . .6 Suppose u < 0.. In older literature . Since R(s) > 0 and G _ is concentrated on (oo. and hence by the WienerHopf factorization identity (A.. in turn. Given T has been computed.lt we get Qh = (T + to+)h = T(sI . . 0). Hence with h = (sI T). the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. we get at a(Q .5.p1i .T) = 1 ata+ = a+. Further. T) with a+ = a(QT)/at. THE RENEWAL MODEL 233 Suppose next F(s) = 1.. This gives d roots 'y. hd. . Pd in the domain ER(s) > 0 . . W v M(d) in the notation of Chapter V).. Corollary 4. . yd satisfying R(ryi) > 0.' that the equation F(s) = 1 has d distinct roots p1... Pd with corresponding eigenvectors hl. hd.5(c) means that a+(sI T)1t = 1.. we have IG_ [s] I < 1 . pdhd. Notes and references Results like those of the present section have a long history. t(ry) > 0.6) i=1 i=1 Proof Appealing to Theorem 4..T)lt + t = s(sI .. and the topic is classic both in risk theory and queueing theory (recall that we can identify 0(u) with the tail P(W > u) of the GI/PH /1 waiting time W.6. and define hi = (piI . and the solution is .9) we have G+[s] = 1 which according to Theorem 1. Then G+ is phase. Let d denote the number of phases.
where R is an unknown matrix.. an alternative approach (the matrixgeometric method ) has been developed largely by M. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. with representation say (a(' ). the background Markov process with p states is {Jt}. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. whereas the approach was introduced in queueing theory by Smith [350]. We assume that each B. [270] and Latouche & Ramaswami [241]. see Dickson & Hipp [118]. For surveys . That is . a pioneering paper in this direction is Tacklind [373]. and the distribution of an arrival claim is B. E(t)).) d (see. The arrival rate in background state i is a. For further explicit computations of ruin probabilities in the phasetype renewal case . Neuts and his students. It turns out that subject to the phase.4. e. T('). In risk theory. the intensity matrix is A and the stationary row vector is ir .exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. The exposition here is based upon [18]. see Neuts [269]. is phasetype.g. In queueing theory.234 then in transform terms CHAPTER VIII. involving . but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes )..contained derivation). which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. and appears already in some early work by Wallace [377]. Here phase..type assumption . the ruin probability can be found in matrixexponential form just as for the renewal model. The matrix. The solutions are based upon iterations schemes like in Theorem 4. starting around in 1975. the fixpoint problems look like R=Ao+RAI+R2A2+ . The distribution of W comes out from the approach but in a rather complicated form .type assumptions are basic. Asmussen & O'Cinneide [ 41] for a short self. This complex plane approach has been met with substantial criticism for a number of reasons like being lacking probabilistic interpretation and not giving the waiting time distribution / ruin probability itself but only the transform. [119]. The number of elements of El=> is denoted by q.F. Numerical examples appear in Asmussen & Rolski [43]. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI.
This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. 5a Calculations via fluid models.1 In Fig. Diagonalization Consider a process {(It. The key unknown is the matrix K.6. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. states . 5.Vt)} obtained by time reversing the I component. p = ql = Q2 = 2.5. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. the phase space E(°) for B. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. However. and the one E(•) for B. •. O. The version of the process obtained by imposing reflection on the V component is denoted a Markovian fluid and is of considerable interest in telecommunications engineering as model for an ATM (Asynchronuous Transfer Mode) switch. say with slope r(i) on intervals where It = i. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4.1. the analysis involves new features like an equivalence with first passage problems for Markovian fluids and the use of martingales (these ideas also apply to phasetype renewal models though we have not given the details).2. The two environmental states are denoted o. 5.1. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. has states o.4. Section 5b then gives a representation along the lines of Theorem 4. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5.
Let E denote the matrix . resp.1 A complex number s satisfies 'A+ (f3i(Bi[s] . Recall that in the phasetype case. we have more martingales at our disposal. r(i) _ 1. i E E. 5. First. MATRIXANALYTIC METHODS 4. a E E(i) } . Bi[s] = a(i)(T(i) + sI)it('). The intensity matrix for { It} is (taking p = 3 for simplicity) I A . V. F = E U { (i. This implies that in the fluid context. in the fluid model Eel'. a) : i E E. Eli) + Proposition 5. 5. Thus F = {o. of E into components indexed by E. j = 1. 4.(Ni)diag r(i. a) = 1.Vt)} is then obtained by changing the vertical jumps to segments with slope 1. •. corresponding to the partitioning + Epp). < oo for all s. o. whereas Ee8s' = oo for all t and all s > so where so < oo. 4}.31a(l) (/3i)diag .1) if and only if s is an eigenvalue of E. the probability in the Markovmodulated model of upcrossing level u in state i of {Jt} and phase a E Eli) is the same as the probability that the fluid model upcrosses level u in state (i.236 CHAPTER VIII. '31a(1) 0 0 f32a(2) 0 0 AI = t(1) 0 0 0 t(2) 0 0 0 t(3) 0 T1 0 0 0 0 T(2) 0 '33a(3) 0 0 T(3) The reasons for using the fluid representation are twofold. a) of {It}.1))diag ) a = sa and the eigenvector b = .92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. In the general formulation .1))diag + sII = 0 (5. F is the disjoint union of E and the Eli). consider the vector a satisfying (A + (13i(Bi[ s] . 2. Second. t.1(a).1(b) {(It . 4. The fluid model on Fig . If s is such a number. A claim in state i can then be represented by an E()valued Markov process as on Fig.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 .
c = a.E22)1 E21) a = 0. d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .E12E22 E21 I . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I . t(1) 0 0 then also 0 t(2) 0 . assume that a is chosen as asserted which means (Ell .sI 0 0 0 T(3) .sI.sI)1t)) iag I = 0 which is the same as (5. resp .sI ()3i)diag . where c.1).5. E(1) + + E(P). it follows that if Qla(1) 0 0 .sI+ ((3ia(i)(T(i) . iEE (a> of 0* 1 AI. with Eii replaced by Eii .sI + E12 (sI . Then (up to a constant) c = a.Nla(1) 0 0 T 1.E22)1 E21a.sI 0 0 0 T(2) .32a(2) (/3i)diag . Noting that E11c + E12d = se by definition.(sI .E22)1 E21a E21a . 0 . Then E21c+E22d = E21a .sI 0 0 t(3) 0 0 = 0. d = (sI E22)1E21a = E ai(sI .E21a + sd = sd.A .sI) (sI .E22 . For the assertions on the eigenvectors.T('))1t(i) .A . and let d = (sI . MARKOVMODULATED INPUT 237 indexed by E.
/' u = e' (esiuc ( 1) . j. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 .v) = j). pi(u. v. a) and noting that i1 (u) = >I j.v}.v) = = p i( u . Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u. j) pi( u . j.sv)b(v) = 0. j.v.. v) yields C{V) = e8 . < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s.a)d^ ). v.pi(u.v) = Optional stopping at time w (u. . . We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. d("))1 e.4 that {e"1b(v) is a martingale .. v > 0. . v = 1.. Example 5 .a Solving for the pi(u. c j. a) = (j. Iw(u. B2 are both exponential with rates 51 i b2. the result u follows. w(u)=inf{t >O:Vtu}. sq with $2s. j... a) = Pi (Vw(u. e89uc(e)) (d(1) . u) Iw(u. To determine 0 (u). define w(u. Example 5 ...j. MATRIXANALYTIC METHODS Theorem 5. it follows by Proposition II.. Then . Proof Writing Or'Alb( v) = svb( v) as (AI . I' i( V P2 (w (u) < oo.4 Assume that E has two states and that B1.O.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. v. a).( u.. w(u...5.v)=inf{t >0:Vtu orVt=.. . .v) = v) I. Thus 0(u) = esu/d = pe7 ° as u should be. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b .3 Consider the Poisson model with exponential claims with rate 5. For u. j.Q. a)). . a)).v) = (j.j)c v .. a )d(a + e8 °vpi (u ..238 CHAPTER VIII. Here E has one state only. Then we get V)i (u) as sum of two exponential terms where the rates s1.upi(u. q.
Proof We must show that G+ (i.b (u) = Pi(M > u) = 9(i)euue.6 For i E E.5. (5. is 0 /3 f R(i . according to VI.y = to B k7 j # k In particular.Qj eie 0 f e (j) T(') x T(j)y ej a e dx e e 00 00 eKx ® e T(')' dx (ej (& I)e T(')ye eKa®T(')x dx (ej (9 I)eT(') Ye e(i)eT(')ye. (') a T( However .x) 00 f ° (') (j) eT (yy)edx .33(e = 0 a(j))(K ®T ( j))(ej (9 I). the Pidistribution of M is phasetype with representation (E(1) + + E(P). 8^')IT(j)) where e 3^') =. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI. j. •) is phasetype with representation (E(i).2) the l. we get the following phasetype representation for the ladder heights (see the Appendix for the definition of the Kronecker product 0 and the Kronecker sum ®): Proposition 5.h. j. oo)) j)ye. 0 Theorem 5 . U) where t(j) + t(j)O(j j = k uja.xxej • a 00 oo el .3) . j.2.3j eye. 9('). In terms of K.k. i.( 2. dx)Bj(y .s. (y.5 G+(i.
e. +bn0i1 0n +a10n1 +. For a transition from (j.f. say.240 CHAPTER VIII. However.k y. that the density b(x) can be written as aeTxt for some row vector a.) which is rational.. This yields the asserted form of uja... bn1 bn). u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. 0 1)'.2) . if b* [0] = b1 +b20+b302 +.. we have the additional possibility of a phase change from a to ry within the ladder step.. the initial value of (i. some square matrix T and some column vector t (the triple (a. a) to (k. i.e.. and lifelength M. . 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. a) is obviously chosen according to e(`). intensity matrix U. which occurs at rate t(i). which occurs w.2. T. For j = k. see Example 1. Bk7 . Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). a m. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps .1 Let b(x) be an integrable function on [0..g.5). which occurs at rate t^^7. the current ladder step of type j must terminate. (6. t = (0 0 . we have sofar concentrated on a claim size distribution B of phasetype. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). Then b*[0] is rational if and only b(x) is matrixexponential. equivalently. t) is the representation of the matrixexponential distribution/density): Proposition 6. Numerical illustrations are given in Asmussen & Rolski [43]. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . Associated with each ladder step is a phase process. Starting from Jo = i.y) to occur when j # k. and it just remains to check that U has the asserted form. in many cases where such expressions are available there are classical results from the prephasetypeera which give alternative solutions under the slightly more general assumption that B has a Laplace transform (or. An alternative characterization is that such a distribution is matrixexponential. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform.. the ratio between two polynomials (for the form of the density. i.p. Furthermore. and a new ladder step of type k must start in phase y.
bn of the denominator to be distinct and expand the r. of (6./(0 + bi). S = f c/2 0 21ri .h.T)1 is so.. since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.s. .1. where c = 1 + 1/47r 2. s = c/ 2 . .3) 0 0 0 0 0 .2).1 0 0 )3 = (111).e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. see Asmussen & Bladt [29] (the representation (6. Namely. matrixexponentiality implies a rational transform. t). For a proof. namely to asssume the roots 6l.. shows that the distribution B with density b(x) = c(1 cos(21r x))ex. . u giving b(x) = E 1 ciebiz/bY. MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . . T.47r2 3 .3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. Thus. then b*[0] = a(0I T)1t which is rational since each element of (01 . b(x) > 0 for x > 0..2 A remarkable feature of Proposition 6. t= 0 . Writing b(x) = c(e( 2ni1 ) y/2 .1 0 .. 1 . (6..4) 0 0 1 c This representation is complex.2).an_3 an _ 4 . S. a2 a1 Proof If b(x) = aeTxt.1 is that it gives an explicit Laplace tranform inversion which may appear more appealing than the first attempt to invert b* [0] one would do.. we can always obtain a real one (a.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). (6. 0 1 an an1 an _2 . T= 0 0 1 . but as follows from Proposition 6.6. u Remark 6.. (6. ... Example 6 .1) as E 1 c.3) was suggested by Colm O'Cinneide. The converse follows from the last statement of the theorem. One of his elementary criteria.(6. 0 0 0 0 1 0 0 . personal communication). s) is given by 27r i . cannot be phasetype.47x2 3 1 0 . 0 0 .
leading to matrix calculus in high dimensions when b is small.5 (6.4) the Laplace transform of the ruin probability is /g(e)PO 0*[e] _ /' eeu^G(u)dU = 0 9(/3a0p(9)ap (9)/q(9)) . (6. .1)2 + 6). that despite that the proof of (6. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). For the second algorithm. We recall (see Section 3.6) The remarkable fact is.5) Thus. MATRIXANALYTIC METHODS Example 6 . and that the minimal number of phases in a phasetype representation increases to 0o as 5 . T. For the first.242 CHAPTER VIII. Then (cf. then: Proposition 6. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. T. As for the role of matrixexponential distributions in ruin probability calculations. (6.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . 0. and can use this to invert by the method of Proposition 6. t) of b(x). Consider the distribution with density = 15 ((2e2x . q are polynomials without common roots. t) a phasetype representation with a the initial vector. recall that t = Te) that if B is phasetype and (a. But since 15(1 +6)02 + 1205 0 + 2255 + 105 b* [9] _ (7 + 155)03 + (1355 + 63)92 + (161 + 3455)9 + 2256 + 105 Proposition 6.1 to get i (u) = f3esus. Corollary 111. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1.6) holds true also in the matrixexponential case.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. we use a representation (a.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. T the phase generator and t = Te.3. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. and present two algorithms for calculating '(u) in that setting. we take as starting point a representation of b* [0] as p( O)/q(9) where p.
(6.7) 9( cf.1UB(B + BVA1UB).T)1 + 1 ib* (91.6.T)1 + (6I .'t.5 ).T)1T1t. this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .1 + 82 (9I .to+)1 = (BI .1 + b+ = b++ 1 .1t = f3a (0I T)1T1t . xb(x) dx = aT2t.T).B=land V=a+. (6. we get (91. with A = 91T.T)1t ( l .b* (6. Presumably. (91.A . b+ = a+(9I .T .T)1T 2 = and 1 = AB IT2 + 82T . 519) (A + UBV ). b+ = a +(BI .1 = A1 .1BVA1. since (91T)1T .6b* .6).T .to+)1T .1t du = . From the general matrix identity ([331] p.1t = b* .T . Then in Laplace transform formulation . .t. U =. the assertion is equivalent to a+(BI .T)1t)1a +(9I .T)1 (91.T)1 J0 00 b(x) dx = f aT1t. MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .a+(9I . we get b+ = 0aT1(9I T).T)1t.T)1ta+(OI . but we shall give an algebraic proof.1 = ^(T1 + ( 91T)1).T)1 so that b* b** b** a+(9I .to+)1T . Now.
but the argument of [286] does not apply in any reasonable generality). 3.T)1T2t .T)1T. In Corollary VII.244 CHAPTER VIII. cf. a.h.T)1)t = 8 (1 . T).b*). /3aT1(0I . Much of the flavor of this classical approach and many examples are in Cohen [88].1) is the same as the r. The proof of Proposition 6. a key early paper is Cox [90] (from where the distribution in Example 6.7). premium rate p(r) at level r of the reserve {Rt} and claim size distribution B which we assume to be of phasetype with representation (E. VII.1. .3 is taken).1t = /3a (9I .la. 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4.1. the ruin probability(u) was found in explicit form for the case of B being exponential.8. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). For expositions on the general theory of matrixexponential distributions. 7.s.5 is similar to arguments used in [29] for formulas in renewal theory.1. From this it is straightforward to check that b**/(b+ . see Asmussen & Bladt [29]. which is selfexplanatory given Fig.8 a(T1 + (01.82b*. 0 Notes and references As noted in the references to section 4. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. see the Notes to VII. of (6.1. Lipsky [247] and Asmussen & O'Cinneide [41]. See Fig.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. MATRIXANALYTIC METHODS . We present here first a computational approach for the general phasetype case (Section 7a) and next (Section 7b) a set of formulas covering the case of a twostep premium rule.
is no longer timehomogeneous. Figure 7. In fact. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}.1 The difference from the case p(r) = p is that {m2}. O<. t + s) . Given the v(t) have been computed. Proof The first statement is clear by definition.u)e = A(u)e (7. Let P(tl.tl < t2 < u. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. By general results on timeinhomogeneous Markov processes. P(tl. Define further vi(u) as the probability that the risk process starting from RD = u downcrosses level u for the first time in phase i.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . we obtain V)(u) = P(m„ E E) = v(u)P(0. Note that in general >iEE Vi (U) < 1.1z I. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . t2) = exp where Q(t) = ds [P(t. i.1) where A(t) = v(u)P(0.I] I 80 { tq f Q(v) dvl t1 1 . the definition of {m8} depends on the initial reserve u = Ro. 0 < t < u. though still Markov.7. in contrast to Section 3.t)). Also. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). t) is the vector of state probabilities for mt.t2) be the matrix with ijth element P (mt2 =j I mtl = i). Ai(t) = P(mt = i).e.
given A. Proposition 7. Given this occurs.t).t) for the second.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. the interpretation of Q(t) as the intensity matrix of {my} at time t shows that Q(t) is made up of two terms: obviously. Given A'. Hence Q(t) _ T + tv(u . from a computational point of view the remaining problem is to evaluate the v(t). (7. A'(t) = A(t)Q(t) = A(t)(T + tv(u . Thus.246 CHAPTER VIII.Sj i)p(u)dt • tji = Sji + p(u)tji dt. 0 < t < u. two things can happen: either the current jump continues from u + p(u)dt to u. whereas in the second case the probability is p(u)dt • tjvi(u).(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u).t and being followed by a downcrossing. the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). we get vi(u) = aidt + (1 .(u) p ( u) = . the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 .t)). Given A. the probability of which is 1 . vi.2 For i E E. In the first case. the probability that level u is downcrossed for the first time in phase i is ai. dt]. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . or it stops between level u + p(u)dt and u. MATRIXANALYTIC METHODS However.3dt. 0 Thus. {mx} has jumps of two types. the probability of downcrossing level u in phase i for the first time is E vj (u + p(u)dt) (Sji + p( u)dt • tji + p(u)dt • tjvi(u)) jEE vi(u) + vi' (u)p(u)dt + p(u) dt E {tji + tjvi(u)} jEE Collecting terms. jEE .Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}.
after a certain level v.00 Proof Let A be the event that the process downcrosses level u in phase i given that it starts at u and let B" be the event By={o. When solving the differential equation in Proposition 7.4) backwards for {va (t)}v>t>o. Rt .s.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. we have p(r) = p = vi (u) 0aTe.) + 0 as v + oo.. To deal with this. This yields v.2. Let p" (t). say. vi (U) = lim v= (u). Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. V . RESERVEDEPENDENT PREMIUMS 247 Subtracting v. F" etc. then P(A n Bv) _ P"(A n BV').^ 0. .h.. P u which implies that v. (v) is given by the r.3 For any fixed u > 0. Then P(B. Since the processes Rt and Rt coincide under level B..7. consider a modification of the original process {Rt} by linearizing the process with some rate p. Now since both P(A n Bv) 3 0 and P"(A n Bv) . and similarly P"(Bv) .i7rT1/p. From Section 3.) P"(AnBv) = P(AnB. supRt>v l t<7 I where o. say. refer to the modified process. we can first for a given v solve (7.. of (7.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. starting from v"(v) = . we face the difficulty that no boundary conditions is immediately available.)P"(AnB. (u) on both side and dividing by dt yields the asserted differential u equation. (u) for any values of u and v such that u < v. <oo. Thus. denotes the time of downcrossing level u .5).
248 CHAPTER VIII. > 0} and the last term the contribu tion from {R.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. where v = inf It > 0 : Rt < v}.zp1(u)/(1 . p2.7) equals 01 (v . 2/n. v = u. We recall from Propositon VII.e.1. say. p(r) P. However. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. such that Rt coincide with RI under level v and with Rt above level v. 2u.7) f o (the integral is the contribution from {R. which is available since the z/i'(. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. i. r<v r > v. The f iin in (7.10 that in addition to the O'(•)..v v(u)eTat 1 1 . assuming u > v for the moment. for u > v the distribution of v . 3u etc. (u)}. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII. (7. Corollary 3. cf. 1/n.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. let v(u) = a+2ieiT +ta+>)(uv). Thus we obtain a convergent sequence of solutions that converges to {vi(t)}u>t>o• Notes and references The exposition is based upon Asmussen & Bladt [30] which also contains numerical illustrations.1a. 0 < u < v.. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u..) are so.. The precision depends on the particular quadrature rule being employed. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). where. Therefore u pl(vvueTa t 1.z51(v)).RQ (defined for or < oo only) is defective phasetype with representation (v(u). < 0}). Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v..1. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. T).9. The trapezoidal rule used in [288] gives a precision of 0(n 3). as well p1(u).x) dx f v(u)eT xt dx .V" M 0 .1. To evaluate p1(u). the probability of ruin between a and the next upcrossing of v. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v. the evaluation of Vi(u) requires q(u) = 1 .
jl (t ®e) Thus.e. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e . Since µB = 5/21.e6v Let Al = 3 + 2V'2.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .v(u)eTve).be the eigenvalues of T + to( 2 ). RESERVEDEPENDENT PREMIUMS 1 .v) + (2^ + 3v2 ea'(u " .21 = ? yields 0(u) = 1.x) dx 1 ^(v) ( 1 . p2 < 3. Example 7.(7 The arrival rate is (i = 3.8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t. 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .1 from which we see that pl (u) = 1 + 1 249  1 . (7.4 Let {Rt } be as in Example 3. all quantities involved in the computation of b(u) have been found in matrix form.2.24e.7.v(u)eTVe .to+))11 {e{T®(Ttoy+ ))}„ .e6u 35 .v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.u . I..v(u ) eTV e J v(u)eTxtz/)l (v .8) The integral in (7.2V"2.01 (v . so we consider the nontrivial case example p2 = 4 and p1 = 1.and A2 = 3 . From Example 3.24ev .x) dx} V 1 1(v) f V v(u) eTxt.^1(v) 1 .2.
7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v./2) ea 1(u .24es" .1 Thus..24e5v . ) e sv + ( 2v/2./2 ea1(u") .250 CHAPTER VIII.v)esv + 7 4_ 2. MATRIXANALYTIC METHODS From (7.1)' ?.+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.21(35e6v . and one gets 12e5" .2 35e6v .1.24es" . The analysis and the example are from Asmussen & Bladt .b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v . 21 3 In particular.1 V2 = 4e5"+6 35e6v . Notes and references [30].24e5v . 192esv + 8 P1 .
III. we require that B is concentrated on (0. see I. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. and instead we shall work within the class S of subexponential distributions . x 2iror2 (c) the Weibull distribution with decreasing failure rate . the exponential change of measure techniques discussed in II.g. For the definition . The definition b[s] = oo for all s > 0 of heavy tails is too general to allow for a general nontrivial results on ruin probabilities. x 4 oo.N(µ. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. for all t > 0.4. A rough distinction between light and heavy tails is that the m. B[s] = f e8x B(dx) is finite for some s > 0 in the lighttailed case and infinite for all s > 0 in the heavytailed case. For further examples. oo ) and say then that B is subexponential (B E S) if 251 .2b.46 and at numerous later occasions require a light tail. L(tx)/L(x) 4 1. B(x) = ex0 with 0<0<1.Chapter IX Ruin probabilities in the presence of heavy tails 1 Subexponential distributions We are concerned with distributions B with a heavy right tail B(x) = 1.B(x). For example. (b) the lognormal distribution (the distribution of eu where U .f. a2)) with density 1 e(logyFh) 2/2az .
the r. then (with high probau bility) so are both of X1. proving (a). the distribution of independent r. (1. B(x) ax.F(X1 > x. P(Xi <yI Xi+X2>x) 1B(y). . given X1 + X2 > x.252 CHAPTER IX. As contrast to Proposition 1. x 3 00. Then: (a) P(max(Xi.1 Let B be any distribution on (0.p. we have {max(Xi. Proof By the inclusionexclusion formula. HEAVY TAILS B*2\ 2.'s X1.v. In contrast. X2) > x}.v. if X1 + X2 is large .p.1(b) is oo.2 If B E S. That is. oo). In terms of r. X2) > x) is P(X1 > x) + P(X2 > x) . X2 > x) = 2B(x) . X2) > u x)/B(x) = 2.'s.2B(x). that is. X2 with distribution B. X2) > x} C {X1 + X2 > x}. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex.v. To capture the intuition behind this definition. 1). B(x) Here B*2 is the convolution square. one can check that x x where U is uniform on (0. the behaviour in the lighttailed case is illustrated in the following example: Example 1. The proof shows that the condition for B E S is that the probability of the set {X1 + X2 > x} is asymptotically the same as the probability of its subset {max(Xi. X1 is w. Thus . in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. and thus the lim inf in (b) is at least lim inf P(max(Xi. oo). 1/2 it has the distribution of X1I X1 > x. then P(X1>xI X1+X2>x)* 2. (b) liminf BB(() ) > 2. Since B is concentrated on (0. note first the following fact: Proposition 1.1) then means P(X1 +X2 > x) 2P(Xi > x). X2 but none of them exceeds x. P(max(Xi. X2) > x) ^' 2B(x).2. We later show: Proposition 1.3 Consider the standard exponential distribution. Thus the liminf in Proposition 1. That is.B(x)2 . 1/2 'typical' (with distribution B) and w.
we get limsupB*2(x)/B(x) < 2. and combining with Proposition u 1. Let 0 < 5 < 1/2.B*(n ) B(dz) (1. a contradiction. B( 0 . 1 < B(x ) B( x) Y) < B( 0).5 If B E S.B*n(x . This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. yo] as X + 00.'s: if X . 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1.B E S. Hence lim sup a+oo B*2(x) 2B((1 . we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 .y)/B(x) > 1. Finally lim inf B(x .S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . then either one of the Xi exceeds (1 .1.S)x.1(b) we get B*2(x)/B(x) * 2.v. x]. If X1 + X2 > x.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10. We now turn to the mathematical theory of subexponential distributions. Proposition 1.] Proof Consider first a fixed y.yo].B(y) = 2. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.4 Any B with a regularly varying tail is subexponential.xIX > x converges in distribution tooo. If lim sup B(x . Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. then the overshoot X .z B(x) .B(y)) . then B(B(x)y) * 1 uniformly in y E [0.6)x)/((1 . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . or they both exceed Sx.y)/B(x) > 1 since y > 0. y] and (y. [In terms of r.
nI < e for x > y. Proof We use induction. then e"R(x) * oo. This implies B(x) > c2e5x for all x. Proposition 1.. choose y such that IB*n(x)/B(x) . its intuitive content is the same as discussed in the case n = 2 above.2). and this immediately yields the desired conclusions.5 and the induction hypothesis.5 that B(n) > e6B(n .5 and dominated convergence.z ) (x ) = 1 + (^ B(x . x oo. so assume the proposition has been shown for n.z) B(dz).B(x .7 If B E S. HEAVY TAILS Corollary 1. Given e > 0. The case n = 2 is just the definition. Proof For 0 < 5 < e. B(x) \Jo _ B(x . 0 Proof of Proposition 1.z) B(dz) _y B(x) 111 Lx B .254 CHAPTER IX.z) B(x) Here the second integral can be bounded by B*n(y) B(x) .2. then for any n B*n(x)/B(x) * n. b[c] = oo for all e > 0.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.B*2 (x) B(x) (x . O The following result is extremely important and is often taken as definition of the class S.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . The first integral is y B(x . P(X1 > xIX1 + X2 > x) _ P(Xi > x) _ B(x) 1 P(X1 + X2 > x) B2(x) 2 1 y P(X1<y X1 + X2 > x) B(x . B*(n+1) (x I xy + Jxx y) W. we have by Proposition 1. Then by (1.1) for all large n so that B(n) > cle6n for all n.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1.6 If B E 8.
5 easily yields P(X1 + X2 > x. Xi <= v Ai (x .ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 . Proof Define 5 > 0 by (1+5)2 = 1+e. For any fixed v.X1 > xv. A2 be distributions on (0.y)B(dy) = B(x)ov (1)• v (1. an = supx>o B*n(x)/B(x).v. choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). oo) such that Ai (x) _ aiB(x) for some B E S and some constants al. Then Al * A2 (x) .z) B(dz) x .4) . SUBEXPONENTIAL DISTRIBUTIONS 255 Here the first term in {•} converges to 1 (by the definition of B E S) and the second to 0 since it is bounded by (B(x) . Proof Let X1.z) B(x . it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .3) Using the necessity part in the case Al = A2 = B yields f xv B(x . Then Al * A2(x) = P(X1 + X2 > x).z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .y)Ai(dy) = (x)o(1) (1. a2 with a1 + a2 > 0.z) B(dz) < 1 + A + an(1 + d) . X2 be independent r.'s such that Xi has distribution Ai. Since P(X1+X2 > x. 0 Lemma 1.8 If B E S. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.z) B(x) < 1 + A + an sup f x B(x . Then by (1. an+1 fX B*n( *n(x .B(x .ala2B(x)2 which can be neglected.9 Let A1.0 completes the proof. Proposition 1.(al + a2)B(x). 0 Proposition 1.y)Ai(dy) v) f o .y))/B(x). e > 0. x>T o B(x) The truth of this for all n together with al = 1 implies an < K(1 + 5)2n where K = (1 + A)/e.X2 > xv) < A1(xv)A2(x v) .1.2).i). Combining these estimates and letting a 4.
u Corollary 1.B(x) Proof Take Al = A.y)Ai(dy) = B(x)o„(1). then A E S. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. it is easy to see that if L1.2aB(x) .(x) is decreasing for x > x0 with limit 0 at oo.13 Let B have density b and failure rate A(x) such that .Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1.Bl (x) + B2 (x) when B1. V (1.aiB(v)) = B(x)o„(1). then so is L = L1 + L2. However. A(x) = o(B(x)).10 The class S is closed under tailequivalence. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1. a2 = 1. i = 1.11 Let B E S and let A be any distribution with a ligther tail. it should hold that B1 * B2 E S and B1 * B2 (x) . with a > 0 and L1. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown. u Corollary 1. A2 = B so that a1 = 0.5) Here approximately the last term is B(x)o„(1) by ( 1. if q(x) aB(x) for some B E S and some constant a > 0. Proof Taking Al = A2 = A. In the regularly varying case.h. L2 slowly varying.v)B(v) + _'U Aq(x .2. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290].s. B2 E S. Hence Corollary 1. the l. B1 * B2 (x) .Ai(x .y)B(dy).4).256 Now (1. HEAVY TAILS 'VV B(x . That is.3) follows if CHAPTER LX. Then A * B E S and A * B(x) .5) becomes x B(x . a1 = a2 = a yields A*2(x) .2A(x). f " By a change of variables.v)Ai(v) . whereas the two first yield B(x)(Ai(v) . L2 are slowly varying. u It is tempting to conjecture that S is closed under convolution.12 Assume that Bi(x) = Li(x)lxa. Then B E S provided fo "O exA(x) b(x) dx < oo. of (1. That is. .9).
y) < yA(x .y) * 0..A(y)\(y) dy + fox/ 2 eA(x ). SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x .1 has limit 1 + 0.U) /or) v 2x This yields easily that ex. Since ) (x . replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). x .12. Thus. we can use the same domination for the second integral but now the integrand has limit 0 .y) dy. the first integral has limit 1 .y) < A (y) for y < x/2.15 In the lognormal distribution.`(x)b(x) is integrable. Thus. In the regularly varying case. Thus A(x) is everywhere decreasing. the DFR Weibull distriu bution is subexponential. Example 1. Then b(x) = Ox0lexp. The middle bound shows that it converges to b(y) for any fixed y since \ (x .1)xcl1 . f ' L(y) dy . To illustrate how Proposition 1.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ).e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx .(y) dy.A(xy)A ( y). Then B(x) = eA(x). Define A(x) = fo .(x . Goldie & Teugels [66]): Proposition 1. the u lognormal distribution is subexponential. proving B E S. Thus B*2(x )/ B(x) .2). 0 A(x) . B*2(x) . subexponentiality has alrady been proved in Corollary 1. L(x) y° (a . we first quote Karamata's theorem (Bingham. By (1. Further. an integrable function by assumption..1. Example 1.14 Consider the DFR Weibull case B(x) = ex0 with 0 <. and exa(x)b(x) = (3x01e(10)x9 is integrable. a(x) = ax01.16 For L(x) slowly varying and a > 1. Jo For y < x/2.3 < 1.A(y)a(y ) = ev'(y) b(y).A(x . Thus by dominated convergence .13 works in this setting. elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing.
18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.13 may present a problem in some cases so that the direct proof in Proposition 1.E(X . (c) Under the assumptions of either ( a) or (b).4 is necessary in full generality. HEAVY TAILS Proposition 1. f O B(y) dy .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. we get (1.ea b(x) is integrable. Proof ( a): Using Karamata's theorem.17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.1)])a 1 1 . 'y(x) = EXix>. More precisely. Then 7(x) . .L(x)/x" and )t(x) .1/A(x) and P(X ixil'Y (x) > y) * e'.xjX > x.)/Y(x) > y) (1 + y/(a .1))a .x)+ _ 1 °° P(X > x) P(X>x )J L PX >y)dy 1 x L(y)/ydy L(x)/((a1)x'1) x )l ° J °° ( ()l a x a1 Further P ((a . yo] .y(x)B(x).a/x. then 7(x) x/(a . let X W = X . Thus exa(x)b(x) .258 From this we get CHAPTER IX.1)] I X > x) L(x[1 + y/(a .6) EX(x) . the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull.1)]) xa L(x) (x[1 + y/(a . However. the monotonicity condition in Proposition 1. We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions. (1 + y/(a . Then: Proposition 1.1) and P(X (.1)X(x)/x > y) = P(X > x[1 + y/(a . then B(x) .
2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B.n0 1•P(K= n)•n = EK. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). d. Bo(x) = f0 B(y) dy / µB. Notes and references A good general reference for subexponential distribution is Embrechts. Recall that B0 denotes the stationary excess distribution. We get p(yl+.1/.EK G(u). Lemma 2. cf. Let St = Ei ` Ui .v.+YK> u) = ^•P(K = n)G* n(u ) . . then Vi(u) P Bo(u)...7) is referred to as 1/A(x) being selfneglecting. with EzK < oo for some z > 1. 1. u a oo.A(x) I X > x) = exp {A(x) . with common distribution G E S and let K be an independent integervalued r.2 Let Y1. The remaining statement (1.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. Y2. r(u) = inf it > 0. . Then P(Y1 + • • • + YK > u) .A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P . St > u}.nn. Kliippelberg & Mikosch [134]. It is trivially verified to hold for the Weibull.t be the claim surplus at time t and M = sups>0 St.and lognormal distributions . i. nG(u). and that for each Proof Recall from Section 1 that G*n (u) z > 1 there is a D < oo such that G*n(u) < G(u)Dzn for all u..8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. Examples 1. P The proof is based upon the following lemma (stated slightly more generally than needed at present). 0 G(u) L G(u) .(x).14.2.1 If Bo E S.. be i. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . Theorem 2 .15.
Bo E S.. Proof of Theorem 2. Note that in these examples .260 CHAPTER IX. a]. u The condition Bo E S is for all practical purposes equivalent to B E S. u x+a Notes and references Theorem 2. see Abate.18. P(K = k) = (1. Borovkov [73] and Pakes [280].x400 PBB(x) PB Leta+oo. Since EK = p/(1. (2. Bo is more heavytailed than B . HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant.p) and EzK < oo whenever pz < 1. r(1/Q) xlQexp B(x) = ex' From this . we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . Bo E S is immediate in the regularly varying case. Bo ¢ S.13). _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) .x^ ) B(x) _ f or ( lox .. See also Embrechts & Veraverbeeke [136]. In general: Proposition 2.1 is notoriously not very accurate. lognormal . The problem is a very slow rate of convergence as u ^ oo.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. The approximation in Theorem 2. then Bo(x)/B(x) + 00.3 If B E S. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0.1) In particular . x 4 00. in our three main examples (regular variation .1 is essentially due to von Bahr [56]. the result follows immediately from Lemma 2. as well as examples where B ¢ S. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1.µ J ) . mathematically one must note that there exist (quite intricate) examples where B E S. However.µB(01 . The PollaczeckKhinchine formula states that (in the setup of Lemma 2.?(xµ 8 (x).1)xa1' vxe(109x11)2/202 2 +° /2 µB = eµ Bo(x) eµ+O2/2(log x)2 27r' = µB = F(1/0 ) Bo(x 1 ) . Weibull) one has Bo(x ( B(x) .p)p'.1.2. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. For some numerical studies. .
T1 the ith interarrival time and Xi = U.9+ < oo) = P(S. Let U= be the ith claim . E. i=1 .. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA).. i. This shows that even the approximation is asymptotically correct in the tail. (3.3. [279]. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. Define further 0 = IIG+II = P(r9+ < oo).. . there is a representation of M similar to the PollaczeckKhinchine formula. {n= 0.} Then ik(u) = F ( M > u) = P(i9 (u) < oo). Kalashnikov [219] and Asmussen & Binswanger [27].1.1) this end .g. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. in [219] p. + Xn. THE RENEWAL MODEL 261 Choudhury & Whitt [1].1 gives 1010.] The proof is based upon the observation that also in the renewal setting. Then K M=EY. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. We assume positive safety loading. Based upon ideas of Hogan [200].e. Somewhat related work is in Omey & Willekens [278]. t9(u) = inf {n : Snd> > u} ..Ti. also a second order term is introduced but unfortunately it does not present a great improvement. p = iB /µA < 1.y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. M = sup s$ . In [1]. The main result is: Theorem 3 . G+ (A) = P(Sq+ E A. Snd) = Xl +.. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x .1 when u is small or moderately large. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. T+ < oo) where r+ = T1 + • • • + T.+ E A.y)/B (x) > 1 uniformly on compact y internals... To Bo(u) u + 00.
B(x) _ J O° B(B(x)y) A(dy) f 1 . we will use the fact that the proof of (3. x * oo.Y2. A(dy) = 1. A.B(x). and hence FI(x) . 0] to the integral is O(F(x)) = o(FI(x)).2 F(x) . d+ < oo).. G_(A) = P(S..3 G+ (x) . The heuristics is now that because of (b). cf. FI (x) _ fz ° F(y) dy. 0 The lemma implies that (3. x > 0. the contribution from the interval (. (3.i. with distribution G+/9 (the distribution of S.g+ > x.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. Let F denote the distribution of the Xi and F1 the integrated tail. Lemma 3 .9)9'' and Y1.oo.Ti). P(K = k) = (1 .PBBo(x). this representation will be our basic vehicle to derive tail asymptotics of M but we face the added difficulties that neither the constant 9 nor the distribution of the Yi are explicit. Proof Let R+(A) = E E'+ ' I(S.3) and we will prove it in this form (in the next Section.(. (b) and does not rely on the structure Xi = Ui .d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.1 IPG_ I / F(x . Lemma 3 . oo) = F(S.1) holds for a general random walk satisfying the analogues of (a). whereas for large y .y) dy = 1 Pi (X) oo IPG_ I . HEAVY TAILS where K is geometric with parameter 9. 0] normalized by IPG_ I so that we should have to G+(x) .2). As for the compound Poisson model. u a 00. Write G+( x) = G+ ( x.FI(x) /IPG_I. Proof By dominated convergence and (b). are independent of K and i.y) R+(dy ) _ j (x_y)U_(dY) G+ (x) = J 00 00 (the first identity is obvious and the second follows since an easy time reversion argument shows that R+ = U_. x + oo.262 CHAPTER IX. Then 0 0 F( x .1) is equivalent to P(M > u) " .d. Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch.N.FI(u). U_ (dy) is close to Lebesgue measure on (..y+ given r+ < oo).
and that U_(n .2). n] < (1 + e)/1µc_ I for n > N.G+[s]) . n] + 1/I µG_ I. In the lattice case.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N.1.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10. We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x . THE RENEWAL MODEL 263 We now make this precise. Similarly.(dy) fN FI ( x) + lim sup ZY00 N F(x .O[s])(1 . > (1 . and in the last that FI is asymptotically proportional to Bo E S. By Lemma 3. If G_ is nonlattice.3.2) yields 00 F F I (u) P(M > u) _ E(1 . we can assume that the span is 1 and then the same conclusion holds since then U(n .UG_ I x.y) U. (3.=1 BIp G_ I (1.1. Hence using dominated convergence precisely as for the compound Poisson model. F(Y= > x) FI(x)/(OIp _ 1).1. the proof is complete. n] is just the probability of a renewal at n.0)0k k I(u) A.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .1.1 I .oo Fj(x) N J (1 +6)2 I {IC_ I lim sup X400 FI(x + N) _ (1 + e)z (x) I Pi µ G_ I Here in the third step we used that (b) implies B(x)/Bo(x) + 0 and hence F(x)/FI(x) 4 0. then by Blackwell 's renewal theorem U_ (n . choose N such that F(n .F[s] = (1 .9) 1 . n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. u Proof of Theorem 3.3. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .1.9)IpG_ I Differentiating the WienerHopf factorization identity (A. Given e.
HEAVY TAILS µF = (1 .IIG+II)µc_ = (1 .1 is due to Embrechts & Veraverbeke [136]. u)) > P(w(u) < oo)(i lp (0))• On the other hand. S+9(u) .4 For any a < oo.1)6+[0] . .a) N P(M > u).(u)+n .a..2. Note that substantially sharper statements than Lemma 3.4 on the joint distribution of (S.a. see Asmussen & Kliippelberg [36].Se(u)_1 < a) = o(Fj(u)). Notes and references Theorem 3. on the set {M > u.. Mn < u}.a.So(u)) are available. with roots in von Bahr [56] and Pakes [280]. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .u)) = o(P (M > u)) = o(FI(u)). In view of the `one large claim' heuristics it seems reasonable to expect that similar results as for the compound Poisson and renewal models should hold in great generality even when allowing for such dependence.yiui_1. Then P(M E (u ..Sty(u)_I < a} we have w(u) < oo. Sty(u) . must attain a maximum > 0 so that P(M > u.(1 . 10(0) But since P(M > u . 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times. u).264 and letting s = 0 yields CHAPTER IX.a. S+q(u) .So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . we have P(M E (u .l.0)ua_ . allowing also for possible dependence between the arrival process and the claim sizes. and {Su. Proof Let w(u) = inf {n : Sid) E (u . u)). Therefore by Lemma 3. P(M > u.AB iP We conclude by a lemma needed in the next section: Lemma 3 .SS(u)}n=o.
M* = max S. examples and counterexamples.1 based upon a regenerative assumption. 0o(u) etc. assume pp.. G(x) (4.1 except for the first one) is a random walk.F*(X) = P0(Si > x) .n n=0. We let F* denote the Podistribution of Si. The idea is now to observe that in the zerodelayed case. For further approaches.i. We give here one of them. 4. t>0 S. Theorem 4. Figure 4. < 0 and EoX < oo where X = X2 . Schmidli & Schmidt [47]. such that {SXo+t  SXo}0<t< X 1Xo . We return to this point in Example 4. +1. {Sn}n=o....Sxi}0<t<x2Xl .Sxk}o<t<xk+1xk is the same for all k = 1.. 4... The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.1 = max k=0. M = sup St.4.. See Fig.. and apply it to the Markovmodulated model of Chapter VI. 2.1 where the filled circles symbolize a regeneration in the path.1. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.. (corresponding to the filled circles on Fig....1. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i.1 Note that no specific sample path structure of {St} (like in Fig.1) .1) is assumed.. and the distribution of {Sxk+t . ..X2 < . see [47]. Assume that the claim surplus process {St}t>o has a regenerative structure in the sense that there exists a renewal process Xo = 0 < Xl <. M. 4. Define S..4 below..... {SX1+t . Thus the assumption .X1 is the generic cycle. E0.d. = Sx..
* i o<t<xn+1x.2) to show F(M* > u) > 1.3) hold.266 CHAPTER IX. Then '00 (u) = Fo(M > u) . See Fig. Sxn +t .2.2 Theorem 4.3) applicable so that F(M* > u) 141 F*(u). (4. HEAVY TAILS for some G such that both G E S and Go E S makes (3. Proof Since M > M*. N N Xi=0 N Figure 4. (4. jF11 F* (U).1) and (4. 4. it suffices by (4.S.. u p 00.1 Assume that (4... the assumption means that Mix) and Sl are not too far away. Since clearly M(x) > Sl . (4.3) where Mnx) = sup o<t<xn +1 X.2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent..Sxn = sup Sxn+t . Fo(Si > X). The one we focus on is Fo (Mix) > x) .4) liminf u>oo F(M > u) .
Mn+l > a V (u .: Sn > u} .e)Po (M > U).(u) . /3(u) = inf{n=1. Theorem 4. u)} < P(M* E (u .Sn 0<t<x„+j ( 1 ...4. E (u ... MW O(u)+1 < a) IN ( U n=1 A1. We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 .+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). )) > (1 .Mn +1 >aV(u n=1 00 S. choose a such that Po(Si > x ) > (1 . Under suitable conditions . Letting first u + oo and next e .S. 2.a.a. .4)..4.. assume the path structure Nt St = EUit+Zt i=1 .Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u. u))/P(M* = 0) = o(Po(M* > u)).: S.( u)1 > a) 00 1: Po(Mn<u.Po (M* > u. Given e > 0. Po(M* > u) .1 can be rewritten as 00 (U) (4. To this end.e)Po (M > u. Let a > 0 be fixed. 0 yields (4.5) which follows since Po (M > u. S.. Then by Lemma 3.6) 1 p pBo(u) u where B is the Palm distribution of claims and p .E) Po ( n max St u.1 = limti00 St/t. x > a.2.(1 .e)Po (MMX> > x). M^xu)+l > a) .
6) holds with p = .4). i=1 (4.X both have tails of sup Zt. HEAVY TAILS N` U. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . Mix) < > UE + i=1 o<t<x Thus Theorem 4. cf. the proof of Lemma 4. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4. and also for Mix) since Nx FNX U. (ii) EozNX < oo for some z > 1. oX (see Proposition A1. (iii) For some o field Y. Assume further that (i) both B and Bo are subexponential. Corollary 4.I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .2 Assume that {St} is regenerative and satisfies (4.3PB.p) Ju P Bo(u) 1p 0 . and ENX Ui .6 below. X and N.Q = EoNx/EoX. independent of {> CHAPTER IX. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .'s order EoNx • B(x).} and satisfying Zt/t N.268 with {Zt} continuous.1 is in force.7).8) x Write . we get 00 (u) 1 IPF.v. The same is true for Sl. Proof It is easily seen that the r. a4' 0. since the tail of Zx is lighter than B(x) by (iv).
Example 4 . We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. > 0.9). consider the periodic model of VI.4..(NX). Zt . we assume that B E S. The arrival rate is /3i and the claim size distribution Bi when Jt = i.0 (thus (iv) is trivial).. Bo E S. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. The key step of the proof is the following lemma. and taking F = o. X2 = 1. The regenerative assumption is satisfied if we take Xo = Xi = 0.t + EN'I Ui where {>N`1 Ui . In particular. We consider the case where one or more of the claim size distributions Bi are heavytailed. MODELS WITH DEPENDENT INPUT 269 Example 4 . (iii) is obvious. The number N. in particular lighttailed. . i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. Then (4. < 1..3 As a first quick application. note that the asymptotics of i/io( u) is the same irrespective of whether the Brownian term Zt u in St is present or not. More precisely. X3 = 2.6 with arrival rate /3(t) at time t (periodic with period 1) and claims with distribution B (independent of the time at which they arrive).6) holds. we conclude just as in Example 4. Assume that B E S.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. . i. (i) holds.6) holds. X2 = 1.e. Bo E S. Again . X3 = 2.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2.6) u holds. Thus we conclude that (4... Taking again Xo = Xi = 0. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. . . of claims arriving in [0.3 that (4. we will assume that lim B2(x) = ci x+oo G(x) for some distribution G such that both G and the integrated tail fx°O G(y) dy are subexponential .4 Assume that St = Zt . and for some constants ci < oo such that cl + • • • + c. Theorem 4.
F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 ...270 CHAPTER IX. . 6 Let (N1. as x a oo. i=1 P(Yx > x ^) < P(Y0 > x I. If Jo = i... Thus dominated convergence yields ( P(Yo>x P(Yo>x ."+Np .Fmeasurable. and that for some + cp distribution G on [0. 1. .2. i1 = E\ G(x) In the general case..5.ciG(x). NP ) be a random vector in {0..F) < CG(x)zn'1+. and F a aalgebra such that (N1. An easy conditioning argument then yields the result when Jo is u random. . cp with cl + > 0 it holds that Fi(x) .^•) G(x) P ^ E ciNi = C. For lighttailed distributions. X > 0 a r. we can define the regenerations points as the times of returns to i. oo) and define p Ni Yx = EEX'i . i=1 Proof Consider first the case X = 0. 2 . . oo) such that G E S and some c1.v. Let {Fi}t=1 P be a family of distributions on [0.}P. The same dominated convergence argument completes the proof. and the rest of the argument is then just as the proof of Corollary 4. are independent with distribution Fi for Xij. P P P(YX and > x I... HEAVY TAILS Lemma 4 . . Assume EzN1+"'+Np < oo for some z > 1 and all i. . It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni.G( x ) > ciNi . .X i=1 j=1 where conditionally upon F the Xi. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin . u Proof of Theorem 4. Then P P(Yx > x) .c'(x) where c = ciENi . NP ) and X are . i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.
It follows from Theorem 4..4. m is a (orfinite) .i.e.5 shows that basically only the tail dominant claim size distributions (those with c. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. 5 was first proved by Asmussen. this is applied for example to risk processes with Poisson cluster arrivals. Notes and references Theorem 4. IV. there exist constants Y(u) such that the F(u)distribution of r(u)/y(u) has a limit which is either Pareto (when B is regularly varying) or exponential (for B's such as the lognormal or DFR Weibull). 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. in particular Proposition 2. I T(u) < oo). see Schlegel [316]. Then O(u) .2 and Example 4. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u.4.4.4. i. i.5.. We start by reviewing some general facts which are fundamental for the analysis. the discussion provides an alternative point of view to some results in Chapter IV.pl(1 . r(u) is the time of ruin and as in IV.7. Schmidli & Schmidt [47]. Theorem 4. Theorem 2. cf. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5.1..T2. cf.T2. Theorem 5..6) to hold in a situation where the interclaim times (T1. ) form a general stationary sequence and the U. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). Combined with the approximation for O(u)... for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. 5a Excursion theory for Markov processes Let until further notice {St} be an arbitrary Markov process with state space E (we write Px when So = x) and m a stationary measure. In contrast. Within the class of risk processes in a Markovian environment. VI. as well as a condition for (4.1 is from Asmussen. this should be compared with the normal limit for the lighttailed case.7).5 that the effect of Markovmodulation is in some sense less dramatical for heavytailed distributions: the order of magnitude of the ruin probabilities remains ft°° B(x) dx. The present approach via Theorem 4. For further studies of perturbations like in Corollary 4. The main result of this section. As usual. states that under mild additional conditions.3. cf. we let PN"N = P(. Essentially. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. and the final reduction by Jelenkovic & Lazar [213].d.p)Bo(u). and independent of (T1. An improvement was given in Asmussen & Hojgaard [33]. ).
z) dx G(dz) = ffh(y + z) k(y)dy G(dz).2) with t = 1 means m. The equality of the l. where we can take h. m. {Rt}.s. but the example of relevance for us is the following: Proposition 5. to consider only the case Px(w(F`) = 0) 0. y = 0).). however . k as indicator functions. j.h. .= y.s=j are the transition probabilities for {St}.rij = mjsji where r13. Lebesgue measure. u For F C E. oo). k on E. and (5.t.t + EI U. a familiar case is time reversion (here m is the stationary distribution).s.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . r. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. t. Then there is a Markov process {Rt} on E such that fE m(dx)h(x)Exk(Rt) = Lm(dy)k(y)Eyh(St) (5.y = Qx (. for states i. the whole of R and not as usual impose the restrictions x > 0.)k(x . x = 0+ and F = (0. Say {St} is reflected Brownian motion on [0. Proof Starting from Ro = x. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. y to vary in. Rt is distributed as x + t . Then (5. The simplest example is a discrete time discrete state space chain.h. Sw(F.r. {St} and {Rt} are in classical duality w.00). St is distributed as y . in the terminology of general Markov process theory. (note that we allow x. .272 CHAPTER IX. Thus. an excursion in F starting from x E F is the (typically finite) piece of sample path' {St}o<t<w(F°) I So = x where w(Fc) = inf It > 0: St 0 F} . resp.2) means ffh(a.>N` Ui. For the present purposes it suffices . and starting from So = y. We let QS be the corresponding distribution and Qx. w(Fc) < oo ) 'In general Markov process theory. follows by the substitution y = x .t. to the r.z. Let G denote the distribution of ENt U. say.2) for all bounded measurable functions h.
0].1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0.. z > 0. Sn+1 E Fc) nx. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. in with i0.itt) = P Px(w(Fc) < 00. We can then view Qy.). The theorem is illustrated in Fig .y(2p21 . and we let Qy y refer to the time reversed excursion .13AB < 1] Proof of Theorem 5. Sw(Fo)_ should be interpreted as Sw(F^)_1).(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. i1. oo) = r(0) x= St y (a) Figure 5. when p = .y as a measure on all strings of the form i0i1 . That is. S. .= y) Theorem 5 . .5.s. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. We consider the discrete time discrete state space case only (wellbehaved cases such as the risk process example can then easily be handled by discrete approximations). Sn = in = y.1 for the case F = (oo.2. io = x. QR and QRy are defined similarly. /^s x (S1 = Z1. in E F.. Sw(F)1 = y) .... x = 0. . this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. [note that w(z) < oo a.2 Qy.y() = P ({SW(F`)t} 0<t<w(F °) E So = x. The theorem states that the path in (b) has the same distribution as an excursion of {Rt} conditioned to start in y < 0 and to end in x = 0.3 The distribution of r(0) given r(0) < oo.. Thus.. But in the risk theory example (corresponding to which the sample paths are drawn).y = Qy Q. 5. In particular: Corollary 5.SS(F. in = y. the one in (b) is the time reversed path. . Qx.. w(0.
J (i. . in = x... ..... Si1y k=1 i1 .... t' y and Qy x are measures on all strings of the form ipi l ...i„_iEF Similarly.... 2p) when 20.. MY Thus Qx(ioii .... i0 = y. in)  Pt' (R1 = ii. HEAVY TAILS E E Px (Si = 21i . Sn+1 E Fc) n=1 i1.in1 ..... .. 2n) = Qx. in) = oo jEF^ Sxin1 ..... 20 = y.y(inin _ E SYj jEF` 00 Sxik _1 ....ii ..274 note that Fx(w(Fc) < 00. in = x... Rn = in = x....TI( 2n2n _1 .. note first that Pt' (R l = il.. . Si1y 00 jEF° E E 5xik_ 1 .rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 ..in E F. Silt' E SO k=1 i1.. R ... = in = y.. .. To show Q y x (i0 i 1 .... . Rn+1 E F`) F (w(Fc) < 00.ik_1EF Sxin_1 .. i0) Q x. . 2p)... S. Si l io E mjSjx. Silt' E Sxik_1 .. 21 .. ..(F<)1 = Y) S S and Qx y( ipil . S. Si11 S 1 ... in with 20.. Rn+1 E FC) TioilTili2.ik_1EF . (Fc)1 = y) 00 CHAPTER IX. Rn = in = x.ik1EF Similarly but easier Sxin_1 ... R Qy x(2p21 . in) = Qx.gilt' k=1 ii .. . in E F.
To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . Y > u). U T(O) = T (u) Y Figure 5.t. That is.(o) > y} = {T(0) < oo.v. 7(0) < oo. Z) is described in Theorem 111.2. FINITEHORIZON RUIN PROBABILITIES 275 5b The time to ruin Our approach to the study of the asymptotic distribution of the ruin time is to decompose the path of { St} in ladder segments .t.')density of Y is B(y)/[. S. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. that is.p. P(") = P(. that is. the case r (O) < oo. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). the P(u. 1 w .'s are defined w.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . Bo") is also the P(u.2. Y > y} .2. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo.r.UBBo(u)].2 The distribution of (Y. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . ST(o) > y.B(a) +a PBBo(u) . Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r. y > u. 5.')distribution of Yu is Bo").5. P(o) ). Now the P(u.r. see Fig. the distribution w.
Zk be defined similarly as Y = Y1. Z1).. the random vectors (YI. cf. Since the conditional distribution of Z is known (viz.p) in Pi"'')distribution. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. However.. It is straightforward that under the conditions of Proposition 1. a slight rewriting may be more appealing. conditionally upon r+ (n) < oo. and YI.p).p) in F(u) distribution.. denote the ladder epochs and let Yk.. Z = ZI but relative to the kth ladder segment. r(u)/Z 4 1/(1 . Then Corollary 5. In the proof. . in particular of Z. Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 . Then..d. We let K(u) = inf In = 1. Yn_1 'typical'.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). Z/'y(u) * W in Pi "' ')distribution ..3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). Fig.. That is . Zn). . i. Recall the definition of the auxiliary function y(x) in Section 1. Z). are i. Since w(z)/z a$. The idea is now to observe that if K(u) = n. P(Z < a I Y > u) 3 0. z ^ oo. then by the subexponential property Yn must be large.i. K(u) = n). must be large and Z1. the duration T+ (n) . more precisely.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. (Y.. .p). let r+(1) = T(0). HEAVY TAILS Let {w(z)}Z^. this in principle determines the asymptotic behaviour of r(u). > u with high probability. and distributed as (Y.1. . 5. Then 7(u)/y(u) ^ W/(1 .e.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. and since its dominates the first n . we get the same asymptotics as when n = 1. .: r+ (n) < oo. i. 2. ..r+ (n .e. 1/(1 .. Bo") ). ..276 CHAPTER IX. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long... it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 .18(c) Bo")(yY (u)) + P(W > y) ( 5.T+(2).3..p) then yields the final result T(u)/y(u) + W/(1 . Hence Z.3) holds.. 4 Assume that Bo E S and that (5.
FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5. I A"(u ))II + 0. Yn .u) II 0. A"(u) _ {K(u)=n} = {Y1+ P(. +Yn1<u.. .. > u}..u) E •) . Y„1... A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events). suitably adapted). .3 In the following. Proof We shall use the easily proved fact that if A'(u). . P(. then IIP( I A'(u)) Taking A'(u) = {Y. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1..Bo (ri1) ®B( .5. . I A'(u)) = P(u. II ' II denotes the total variation norm between probability measures and ® product measure.5 Ilp(u.Yn1iYn .n).n) (y1.Yl+ +Yf1>u}.u) E • I A'(u)) = Bo (n1) ®Bou) . Lemma 5.2. Further. P (Yj..
.i..p) < y).. . +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 .. 2.1 and Y„ ..' = y is BM. ...1 P PBo(u) • P(W/(1 . Y") u etc. Then according to Section 5a.4). Zn). y > u. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +.. For Theorem 5.4.d. .P) > y) Corollary 5.. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit.P) Bo(u) for n = 1..P(Z' E •)II > 0 (here Y..t.. HEAVY TAILS ((Z1'. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou).6 IIPIu'n ) CHAPTER IX. . n.P(Y' E •)II * 0. P(u) since by Theorem 2. . n . {wn(z)} be i..1).2. and clearly Zi. . the marginal distribution of Zk is Bo for k < n.y(u)T) ...6.. see Fitzsimmons [144]). Zn are independent..Bo (n1) ®Bo' 0.u has distribution Bout That is..1.. Similarly (replace u by 0). wk(Zk) has a proper limit distribution as u + oo for k < n. Proof of Theorem 5. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + . .+y 1 p"F(Yn > u) P)Pn1 P/(1 . then 11P(Z E •) . By Lemma 5.. . Z' are arbitrary random vectors.r.. the discussion just before the statement of Theorem 5. in particular his Proposition (2.. the density of Yn is B(y)/[IBBO(u)].. n_1 < u. Z11).. Y1 +. The first step is to observe that K(u) has a proper limit distribution w. Zn) E •) . copies of {w(z)}. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). Proof Let (Y11. Z. Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) .278 Lemma 5 . whereas wn(Zn) has the same limit behaviour as when n = 1 (cf.7 O (u. Y'. (Y.. k = 1. Let {wl(z)}.). Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic... in our example Y = (Y1.. be independent random vectors such that the conditional distribution of Zk given Y. . .. and that Yk has marginal distribution B0 for k = 1. . .
1. p(Y) and the result follows. that MQ becomes large as consequence of one big jump. The rigorous proof is. and define the cycle as a = inf{t>0: Vt=0. x > oo. . V.2 Define M.2.(u) = P(V > u) = f f (y) dy .y) . u (6.1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. the results only cover the regularly varying case. Then P(MT > u) . however.B(u). one expects the level y form which the big jump occurs to be 0(1). Asmussen & Teugels [53] studied approximations of i (u. nontrivial and we refer to Asmussen [22]. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. that fo p(x)1 dx < oo. and premium rate p(x) at level x of the reserve. claim size distribution B. 3. the probability that is exceeds u is then B(u . max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 .. i.6. Proof of Theorem 6. Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. RESERVEDEPENDENT PREMIUMS 279 The results of Section 5b are from Asmussen & Kluppelberg [36] who also treated the renewal model and gave a sharp total variation limit result . 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. cf. Then 0 (u) Qf "O ^) dy. Corollary II. The form of the result then follows by noting that the process has mean time Ea to make this big jump and that it then occurs with intensity /3B(u)./3Ea B(u). = supo<t<0.1 Assume that B is subexponential and that p(x) > 00. The heuristic motivation is the usual in the heavytailed area.(3 u u J B(y) dy . Assume for simplicity that {Vt} regenerates in state 0 .e. T) when T + oo with u fixed. More precisely. Theorem 6 .
there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. Further the conditional distribution of the number of downcrossings of u during a cycle given Mo > u is geometric with parameter q(u) = P(Mo > u I Vo = u).q ( u)) 1 .P(MT > u) $B(u) Ft µ(1 . It is also shown in that paper that typically.280 CHAPTER IX. HEAVY TAILS Define D(u) as the steadystate rate of downcrossings of {Vt} of level u and Da (u) as the expected number of downcrossings of level u during a cycle. u Notes and references The results are from Asmussen [22]. D(u) = DQ(u)/µ.q(u) Now just use that p(x) * oo implies q (x) + 0. Hence f (u)r(u) = D(u) = Do(u) . where also the (easier) case of p(x) having a finite limit is treated . Then D(u) = f(u)p(u) and. . by regenerative process theory.
estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. topics of direct relevance for the study of ruin probabilities are treated in more depth.. vrN(z .96s z f (1.2) is an asymptotic 95% confidence interval . Ripley [304]. Fox & Schrage [77].i. We shall be brief concerning general aspects and refer to standard textbooks like Bratley. replicates Zl. The crude Monte Carlo ( CMC) method then amounts to simulating i. 281 . and this is the form in which the result of the simulation experiment is commonly reported. .d. where a2 = Var(Z )..Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . z) 2 = Zit NE ii ii According to standard central limit theory . Rubinstein [310] or Rubinstein & Melamed [311] for more detail . Hence 1.z) 4 N(0.. ZN. la The crude Monte Carlo method Let Z be some random variable and assume that we want to evaluate z = EZ in a situation where z is not available analytically but Z can be simulated. 4Z).
However. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. conditional Monte Carlo and importance sampling. variance reduction is hardly worthwhile. Say that Var(Z') = Var(Z)/2. Letting Z' = E[Z I Y]. generated at the same time as Z. and in most cases this modest increase of N is totally unproblematic. Further. v. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. Therefore. Then replacing the number of replications N by 2N will give the same precision for the CMC method as when simulating N' = N replications of Z'. This is a classical area of the simulation literature. and many sophisticated ideas have been developed.282 CHAPTER X. an added programming effort. one can argue that unless Var(Z') is considerable smaller than Var(Z). writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . so that Z' is a candidate for a Monte Carlo estimator of z. we then have EZ = EZ = z. there are others which are widely used in other areas and potentially useful also for ruin probabilities. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). The difficulty in the naive choice Z = I(T(u) < oo) is that Z can not be simulated in finite time: no finite segment of {St} can tell whether ruin will ultimately occur or not. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). We survey two methods which are used below to study ruin probabilities. and a longer CPU time to produce one replication. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r .b(u. SIMULATION METHODOLOGY In the setting of ruin probabilities. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). T): just simulate the risk process {Rt} up to time T (or T n 7(u)) and let Z be the indicator that ruin has occurred. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. We mention in particular ( regression adjusted) control variates and common random numbers. The situation is more intricate for the infinite horizon ruin probability 0(u).
it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. . even if the optimal change of measure is not practical. it may often be impossible to describe P in such a way that it is straightforward to simulate from P).1. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction.z2 = 0. To this end. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio.. However. This may also be difficult to assess . L such that z = EZ = E[LZ]. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). (ZN. but tentatively. using the CMC method one generates (Z1. Then z Var(LZ) = E(LZ)2 .v.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi .e. Thus we cannot compute L = Z/z (further.3). Importance sampling The idea is to compute z = EZ by simulating from a probability measure P different from the given probability measure F and having the property that there exists a r. .E [Z Z]2 = z2 . L1).. one would try to choose P to make large values of Z more likely. In order to achieve (1. . (1. Nevertheless. it appears that we have produced an estimator with variance zero.96 sis v^ N 2 1 where srs = N j(LiZi . and the problem is to make an efficient choice. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. i. the argument cheats because we are simulating since z is not avaliable analytically. Thus.[E(LZ)] = E Z2 Zz . LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1.3) Thus.zrs.
just the same problem as for importance sampling in general comes up: we do not know z which is needed to compute the likelihood ratio and thereby the importance sampling estimator. To introduce these. I. in terms of the halfwidth of the confidence interval. This leads to the equation 1. but if the point estimate z is of the order 105. i.e.. the optimal P is the conditional distribution given A. we may try to make P look as much like P(•IA) as possible. However.z) 1001.0.100 .z) which tends to zero as z ^ 0. In ruin probability theory. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . For each u. as is the case of typical interest. let z(u) = P(A(u)).e. N . The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . We then . Thus. and further it is usually not practicable to simulate from P(•IA).. it does not help telling whether z is of the magnitude 104.e. assume that the A(u) are rare in the sense that z(u) * 0.96oz /(zV) = 0.96 2Z ( 1 . say 10%. u + oo. z I.5 or even much smaller . SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . However. An example where this works out nicely is given in Section 3. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). 10 . say of the order 103 or less. and let Z(u) be a Monte Carlo estimator of z(u). a confidence interval of width 10 4 may look small. Again. The CMC method leads to a variance of oZ = z(1 . We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem.z) 1 > 00.284 CHAPTER X. Z z V5 In other words . if z is small.96 2 z2 z increases like z1 as z .B = iP(AB) = P(BIA).1. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. large sample sizes are required. Z = I(A) and A is a rare event.1.
F(K = k) = (1 .4) for any e > 0. are i.log Var(Z(u)) lim inf > 2 u+oo .1. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1.i. SIMULATION VIA THE POLLACZECKKHINCHINE FORMULA 285 say that {Z(u)} has bounded relative error if Var(Z(u))/z(u)2 remains bounded as u 3 oo. Logarithmic efficiency is defined by the slightly weaker requirement that one can get as close to the power 2 as desired: Var(Z(u)) should go to 0 as least as fast as z(u)2E. P(K = k) = (1 . logarithmic efficiency is almost as good as bounded relative error. it is appealing to combine with some variance reduction method . If M > u. which gives a logarithmically efficient estimator . Notes and references For surveys on rare events simulation. but as a CMC method .log z(u) of (1. However. . Therefore . let Z +. where M = X1 + • • • + XK. According to the above discussion.X1 + + XK. where X1. see Asmussen & Rubinstein [45] and Heidelberger [190]. The algorithm gives a solution to the infinite horizon problem . XK from the density bo(x). where Z = I(M > u) may be generated as follows: 1.4). let Z +. O (u) = z = EZ. and in practice. i.0.. Let M .2. X2. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. the PollaczeckKhinchine formula III. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. 3.e. the mathematical definition puts certain restrictions on this growth rate.(2. 2..1) may be written as V) (u) = P(M > u).p)pk. . .p)pk.. with common density bo(x) = B(x)/µB and K is geometric with parameter p.. Generate X1. Generate K as geometric. so that NE (u) may go to infinity. Thus.d. The term logarithmic comes from the equivalent form . This allows Var(Z(u)) to decrease slightly slower than z(u)2. it is not efficient for large u . Otherwise. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded.
K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X. So.X(K1)) .. This calculation shows that the reason that this algorithm does not work well is that the probability of one single Xi to become large is too big. To see this... XK1. note first that To check the formula for the P(X(n) > x I X(1).. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail. The idea of [27] is to avoid this problem by discarding the largest X.X(2). and let Z(2)(u) = _ P (SK B0((u > u I X(l).b(u) = P (Xl +•••+XK>u) = EF[Xl + . Z(1)(u) is defined as 0). assume in the following that Bo(x) .. As a conditional Monte Carlo estimator . For the simulation.. Then (cf. we generate only X1...X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) ...286 CHAPTER X.. . form the order statistics X(1) < X(2) < .2. However.XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. X1 + + XK_ 1 > x when X1 > x. we thus generate K and X1i .SK1)2.. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x). XK.1) V)(u) . and that Bo(y) = 1.. and the problem is to produce an estimator Z(u) with a variance going to zero not slower (in the logarithmic sense ) than Bo(u)2.p/(l . compute Y = u ... Xl > x.. y < 0).. just note that EZ(1)(u ) 2 > E[Bo (x .X(2).+XK > uIXl..X1 . A first obvious idea is to use conditional Monte Carlo: write i.XK_1). and considering only the remaining ones.L(x)/x`' with a > 0 and L(x) slowly varying. Thus. .....S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1). Theorem IX.. Z(1) (u) has a smaller variance than Zl (x).Xl .. conditional probability... . .XK1] = EBo(uX1 ..p)Bo(x).. < X(K) throw away the largest one X(K).
1) . and we refer to [27]. BL instead of 0._1) > P(X(n) > _ X X(1). and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. and define )3L. that is. BL by I3L = /3B[y]. Asmussen . X(2)..modulated model P(r+ < oo) and G+ are not explicit ). . l)) . .. X(2). in the renewal or Markov. X(2). X(n1)) Bo((x . . Also in other respects the findings of [28] are quite negative: the large deviations ideas which are the main approach to rare events simulation in the lighttailed case do not seem to work for heavy tails.. for the purpose of recording Z(u) = erysr(u).5). Thus. . ..u is the representation 0(u) = e7sr(u) overshoot (cf.y. Notes and references The proof of Theorem 2.1 is elementary but lengty. 111.S(n_1) I X(1). .3. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. BL(dx) = e7sB(dx)/B[y]. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") .Khinchine formula and importance sampling .S (n1)) V X (. X(n1)) P(X(TZ) + S(. For practical purposes.. X (. Then the algorithm given by { Z (2) (u) } is logarithmically efficient. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . using 13L. it must be noted that a main restriction of both algorithms is that they are so intimately tied up with the compound Poisson model because the explicit form of the PollaczeckKhinchine formula is crucial (say. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". X . However . l)) BO(X(n1)) Theorem 2 . The algorithm is sofar the only efficient one which has been developed for the heavytailed case. B. and simulate from FL. the algorithm for generating Z = Z(u) is: 1. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] ..
SIMULATION METHODOLOGY 3. In detail .. In fact: Theorem 3. . and the change of measure F r FL corresponds to B > BL. It resolves the infinite horizon problem since FL(.F. be i. b different from . Generate T as being exponential with parameter . cf. P'[y] < oo for some ry > 0. to deal with the infinite horizon problem .T. There are various intuitive reasons that this should be a good algorithm. 4. and we have: Theorem 3. M(u) = inf {n : S„ > u}. let Z F e_'s.d. one must restrict attention to the case 4µB > 1. Proof Just note that EZ(u)2 < e . and assume that µF < 0 and that F[y] = 1. return to 3. namely ELery£("). with distribution F. The estimator is then M(u) /3eQT' dB Z(u) (Ui) j=1 )3 e $Ti dB where M(u) is the number of claims leading to ruin. Let Sf0 CHAPTER X. let S. and avoid simulating the known part e7".2ryu _ z (u)2/C2.. X2. the discussion at the end of Section 1b. We formulate this in a slightly more general random walk setting '. Let FL (dx) = 'For the renewal model. b) # (/3L. = X1 + .i.l3 and U from B. Let S .QL.Q. BL). If S > u.S+U . A > AL as in Chapter V. u It is tempting to ask whether choosing importance sampling parameters . The answer is no. The algorithm generalizes easily to the renewal model . BL could improve the variance of the estimator . so that changing the measure to FL is close to the optimal scheme for importance sampling . r(u) < oo) and FL (both measures restricted to. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.r(u) < oo) = 1... Let X1.... + X.288 2.2 The estimator (3.7 tell that P(.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error.3. .(u)) are asymptotically coincide on {r(u)} < oo. The proof is given below as a corollary to Theorem 3. Xi = U. Otherwise. B) is not logarithmically efficient when (/3. More precisely.1) (simulated with parameters ^3. the results of IV. Ti.
. K2. EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.2 > 0.2'X1 .2ryELXi)} . where e' = EL Iog dFL (Xi) > 0 by the information inequality. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . where Kl og (X) (j) 2 ) = log dFL (Xi) . When F # FL. Proof The first statement is proved exactly as Theorem 3 . F(XM(u)).. + KM(u))} = exp {ELM(u)(E . it thus follows that for 0 < e < e'/ELXi.3. Here ELK. By the chain rule for RadonNikodym derivatives... 1. (3... write W(F IF) _ F(XI). .yu+elu u +oo etry' 1 > lim up C2e2.. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) .2) (simulated with distribution F of the X3 has bounded relative error when .+KM(u)}.yu = G. ..2ryELXi. Since K1.i.3 The estimator (3.2) dF Theorem 3.P = FL. are i. Since ELM(u)/u + 1/ELXi. is not logarithmically efficient.. For the second.d. More generally. = c'. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). The importance sampling estimator is then Z( u) = e'rSM( ). EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.
2.T' = U" . we write T = yu. T) with T < oo. yu) is close to zk(u).T' has a left exponential tail with rate /3' and U" . In [13]. so that one would expect the change of measure F 4 FL to produce close to optimal results. Then according to Theorem 3. claim size distributions B'. with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. U' . we conclude by differentiation that Bo(x)=B' (x)forallx > 0. U' .'(y).4. u Notes and references The importance sampling method was suggested by Siegmund [343] for discrete time random walks and further studied by Asmussen [ 13] in the setting of compound Poisson risk models . Further discussion is in Lehtonen & Nyrhinen [245]. As in IV. The optimality result Theorem 3.T". B" and generic claim sizes U'.1 If y > 1/ic'('y).T' D U" . U". SIMULATION METHODOLOGY u Proof of Theorem 3.4 indicate that we can expect a major difference according to whether y < 1/r. /3". then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. This immediately yields / = 3".i. generic interarrival times T' . First by the memoryless distribution of the exponential distribution .e.'(y) or y > 1/r.g. Next. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. . all that needs to be shown is that if U' . 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. the references in Asmussen & Rubinstein [45] and Heidelberger [190].3.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". The queueing literature on related algorithms is extensive . T".T" > x) J /3"e0 yB (x + y) dy = .290 which completes the proof. /3'eQ'YR'( x + y) dy = . In fact: Proposition 4.B'=B". The easy case is y > 1/k'(y) where O(u.3'eO'x f f P (U" .T". see e. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo.1 is from Lehtonen & Nyrhinen [244].T" has a left exponential tail with rate /3'. The results of IV. BL) has bounded relative error. Consider compound Poisson risk process with intensities /3'. CHAPTER X. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. then /3' B' = B".
Further . T(u) < yu] . Bounding u ELZ(u)2 above by a7u.log Var(Z(u)) l im of .yy> 2 .(u) * 1 (Theorem IV. the result follows as in the proof of Theorem 3. and in fact.8).4.4. lim inf uoo 27yu . 7y (4.4. that ryy = ay .'(7). yu)/z. We next consider the case y < 1/r.1).2) Since the definition of ay is equivalent to Eay r(u) . T( u) < yu] e2ryyuEay le 2ay^(u).1.1) so that z(u) = zP(u. 3 Theorem IV. and that ryy > ry. Proof Since ryy > y. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.(ay).4.8 has a stronger conclusion than (4.log 4')u) 4 u (Theorem IV.5) follows. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu).log Var(Z(u)) _ . one would expect that the change of measure F Pay is in some sense optimal. (4. (4. .yu. yu) is of order of magnitude a71.2 The estimator (4. Bay) is logarithmically efficient.3) (simulated with parameters /gay.yk(ay) determines the order of magnitude of z'(u. yu) in the sense that . Let Qy2 = . we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r.to g x ( u ) u u so that (1. Remark 4 .1) which is all that is needed here can be showed much easier .1). T(u) < yu] e Hence by (4. yu) = eayu Eay Leay^(u)+r(u)K(ay).O(u. We recall that ay is defined as the solution of a'(a) = 1/y.3) and we have: Theorem 4.
and (5.ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV.uaoo U That lim sup < follows similarly but easier as when estimating En.1) (see Proposition IV. related discussion is given in a heavy traffic limit q J. we believe that there are examples also in risk theory where (5.yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu .4. One main example is {Vt} being regenerative (see A.2).292 CHAPTER X.. 0 Notes and references The results of the present section are new.u1/2 < r(u) < yu Le ] l = e. zi(u) = INV.7y x(u) > hm inf u+Oo U . 0 rather than when u 3 oo. However. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ). SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .1) where the identity for Vi(u) requires that Vt has a limit in distribution V. In most of the simulation literature (say in queueing applications).(av)Eav l e.a vt(u). '%(u) = P I info Rt < 0) = P(VV > u).1): then by Proposition A1.3.4. N(0.4). (5. yu .yu)/(uyu1/2) .T) = P O<t<T inf Rt < 0 = P(VT > u).3: for many risk processes {Rt }.b(u. Hence lira inf log ryyu + vyu 1/2 tc(ay) . the object of interest is {Vt} rather than {Rt}. there exists a dual process { V t} such that i.Qyu1/2 < T(u) C yu e. Z (u)2 above. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W).1) may be useful.2) . 5 Regenerative simulation Our starting point is the duality representations in 11.1) is used to study Voo by simulating {Rt} (for example. > u) = E f I(VV > u) dt 0 (5. yu .o.o . In Asmussen [13].
EZ1'i = z1 = Ew.. The method of regenerative simulation.+Z(N) z 1. For details . + Z1N>) . z2)) > N(O. 2. the regenerative estimator z%(u) is consistent. provides estimates for F ( V. For the ith cycle..h (zl. j = 1. is the cycle length.. record Zi'i = (Z1'). .. Zl the LLN yields Z1 a$' Z(1) +.. Then (Z1z1i Z2z2 ) 4 N2(0.. 02) (5. REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. i.. > u) (and more general expectations Eg(V. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi.E). Then Z(1).. . To derive confidence intervals .t(u)) 4 N(0. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. z2) z2/z1 yields Vh = (z2/z2 1/zl). + Z2N)) . and Z2'>) where Zi'i = w. which we survey below . let E denote the 2 x 2 covariance matrix of Z(').d. i (^(u) . Z2 .5.. oh) for h : R2 ^ R and Ch = VhEVh. EZ2'i = z2 = E Thus. Taking h(zl. Z2 = N (X21' + . letting J0 'o I (Vt > u) dt . Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. Vh = (8h/8z1 8h/ 8z2).3) . Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ').. consider first the case of independent cycles .. Z(N) are i . Z1 = (Zl1i +. Z2 a4* z2. Thus. Therefore . )).
Z) ^Z(=) .g S12 (5. with distribution depending on a parameter (. The regenerative method is not likely to be efficient for large u but rather a brute force one. to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process .9. () dx f Ax) (dl d()f (x' () f ( z.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . Let X have a density f (x. Then z(() = f cp(x) f (x. Z of the form Z = ^p(X) where X is a r . () dx = E[SZ] f(X. However . the expectation z = EZ of a single r. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims .96s/v"N. We here consider simulation algorithms which have the potential of applying to substantially more complex situations.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . asymptotic estimates were derived using the renewal equation for z /i(u). consider an extremely simple example .5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. in some situations it may be the only one resolving the infinite horizon problem . 6 Sensitivity analysis We return to the problem of 111 . 9. Notes and references The literature on regenerative simulation is extensive.2.294 where 01 2 CHAPTER X. There is potential also for combining with some variance reduction method. see e. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .C)dx = f w(x) d( f ( x. () depending on C. v. Before going into the complications of ruin probabilities .0 . () dx so that differentiation yields zS d( fco(x)f(x.g. Rubinstein [310] and Rubinstein & Melamed [311]. In 111.v.
()) h((U. ( where h( (u. () where U is uniform(0. IPA will estimate zS by 0 which is obviously not correct. For the SF method.r. one. r(u) = Tl + • • • +TM(u)).log U/(. For example . p. cp(h(U. A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. /3o is M(u) Oe (3T: < oo) . ()) d( hc(U. for some Co = (o(U). The following example demonstrates how the SF method handles this situation.1). = E [`d (h(U. Thus . zc = E [d co(h(U. C) f(X. /3 is 0.6. () is an unbiased Monte Carlo estimator of zS. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. The likelihood ratio up to r(u) for two Poisson processes with rates /3. ()) is 0 w . () = d log f (X. So assume that a r. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. () = log U/(2. if f (x. To see this. ()). The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. () _ (eSx.t.() d( is the score function familiar from statistics . just take cp as an indicator function . I(r(u) . Infinitesimal perturbation analysis (IPA) uses sample path derivatives. For IPA there are. this is usually unproblematic and involves some application of dominated convergence .t.1 Consider the sensitivity tka(u) w. () = . 11 /3oeOoT. C)). ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. SZ is an unbiased Monte Carlo estimator of z(. giving h( (U.r. cp' (h(U.v. with density f (x. say W(x) = I(x > xo) and assume that h(U. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . () Thus. one can take h (U. () is increasing in C. the Poisson rate /3 in the compound Poisson model. however . C). Let M(u) be the number of claims up to the time r(u) of ruin (thus. () can be generated as h(U. Then . In the setting of ruin probabilities . Thus. Example 6 . () = (8/8()h (u. Then z(() = Ecp(h(U.
296 CHAPTER X. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) .3L. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). differentiating w. To resolve the infinite horizon problem . We recall (Proposition 111. However. the risk process should be simulated with parameters .t. Example 6.T(u)) e7uerVu) for ?P. There have been much work on resolving the difficulties associated with IPA pointed out above. whereas for the SF method we refer to Rubinstein & Shapiro [312]. in part for different measures of risk than ruin probabilities. We then arrive at the estimator ZZ(u) = (M(u) . In the setting of ruin probabilities.9 .T(u)) I(T(u) < co) ] . .r. BL). SIMULATION METHODOLOGY Taking expectation. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. Thus. different parameters.3 (u) (to generate Zp (u). ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . a relevant reference is VazquezAbad [374].1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. change the measure to FL as when simulating tp(u).0(1) so that in fact the estimator Zf(u) has bounded relative error. 4) that V5.3 (u) is of the order of magnitude ue7u.r. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . j3 and letting flo = 0. for different models and for the sensitivities w.t.
P(•r(u... as e.. T+(a) = inf It > 0 : Rt > al. where X1.. That is. }). in most cases . are i. R„ = u+X. either this makes no difference (P(R. a) = r(u)).g. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo. .Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion. defined as Ro = u (with u E {0. Consider first a Bernoulli random walk. T(u. with P(Xk = 1) = 9. wherel T(u) = inf {t > 0 : Rt < 0} ... and {1.. in the Bernoulli random walk example below.(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. X2.1.1}valued .. Besides its intrinsic interest .(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u.. 'Note that in the definition of r(u ) differs from the rest of the book where we use r(u) = inf {t > 0 : Rt < 0} ( two sharp inequalities ). 297 . a) = r(u) A T+(a). The twobarrier ruin problem The twobarrier ruin probability 0.i.+• • •+X.d.. Y'a(U) = P(T (u) = r+(a)) = 1 .
.2). and the other more advanced but applicable also in some other settings.1) is solution.. C1_0\a.2) Oa(a . zu = EzRO = EzRT(u. one elementary but difficult to generalize to other models.o» = z°P (RT ( u.. the solution of F[.. tba(2) _ (1 .4) by ea(u+Xl+.1) o If 0 = 1/ 2. 7/la(a .. i..o)'t/1a(a . The Lundberg equation becomes 1=F[ry]=(19)+9z. and in view of the discrete nature of a Bernoulli random walk we write z = e7.(4.a) Y.+Xn) F[ a]n n=0.y] = 1.. The martingale is then {zuzXl+•••+X„ } = {zR° }.(u) I\ e = 1 oa ' ()i a = u. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). Wald's exponential martingale is defined as in 11.. u Proof 2..298 CHAPTER XI. (1. In a general random walk setting .e. We choose a = ry where ry is the Lundberg exponent.1. By optional stopping.o)T/la (1) + 8z/'u(3).1) = (19)4/'0(a3)+9ba(a1). MISCELLANEOUS TOPICS Proposition 1.0)/0. = z°Va(u) + za(1  . where a is any number such that Ee°X = F[a] <oo. u + 1.a(u)). = (1 .a) = 0) + zap ( R. Proof 1. then 'Oa(u) _ au a We give two proofs .(1B)u oJ 0.r(u. and insertion shows that ( 1. z and the solution is z = (1 .1 For a Bernoulli random walk with 0 0 1/2.
TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . BROWNIAN MOTION. If p<0.4 For a Brownian motion with drift u > 0. Corollary 1. then Proof Since 'Oa (U)  au a Eea(R°. . i1(u) = e211 .zu)/(za .1) for p # 0.1. 1h (u) = a el u \1 If 9 < 1/ 2. (1.5) . pa( u) _ u Corollary 1. If 9 = 1/2. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .1 If p = 0. (1. thenz1 (u)=1.2 For a Bernoulli random walk with 9 > 1/2.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . Then for p 0 0.2) is trivial (z = 1). and solving for 9/la(u) yields Z/)a(u) = (e 76 .1 yields 't/la(u) = (a . However.• a2µa e2µu .} is then itself a martingale and we get in a similar manner u = ER° = ER ra( u) = 0 • Y'a (u) + all  au Y'a( u)).e7u)/(e7° .. then Vi(u) = 1.1). Proof Let a+ oo in (1.u)/u.0a(u)). RANDOM WALK. Applying optional stopping to the exponential martingale {e7R.1). {R.ba(u) = e2µa .u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. u Proposition 1. If p = 0. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1.
say.4).a) = a) = 5 y = P (R (u. passing to even more general cases the method quickly becomes unfeasible (see. .3. Here is one more case where this is feasible: Example 1. and thus one encounters the problem of controlling the undershoot under level 0. 1 .e7a (u) = 6 /0 .0 (u) (where u p =. Here the undershoot under 0 is exponential with rate 5.vi(a) Proof By the upwards skipfree property.a) = r+ (a)} and similarly for the boundary 0. 0. the paths are upwards skipfree but not downwards.7). (1. However. CHAPTER XI.a) < 0) + e7°P (R(u. (u) _ O(u) .e7a Again .300 Proof Let a * oo in (1. MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths. VIII.. It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1.7/la(u)).6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1.7) . 7/'(u) = 1).5 Consider the compound Poisson model with exponential claims (with rate. Ic 5ry 'pa(u) Using y = 6 . implying R(u. 5). valid if p < 1 (otherwise . For most standard risk processes . we obtain 'Oa a7u . and hence e7u = Ee7Ro E [e7R(.616).a ) < 0) + e 7aF ( R (u.5a). this immediately yields (1.+^a(u))^(a) If 7k(a) < 1. 7O(u) = 7/la(u) + (1 . a) I R(u a ) < 0] P (R(u . however.a) = a on {r (u.a) = a ) + e ' ° ( 1 .0(a) 0 < u < a. letting a * oo yields the standard expression pe7u for .
µ so that {St} is Brownian motion with drift µ .7 For Brownian motion with drift 0. in particular symmetric so that from time r(u) (where the level is level u) it is equally likely to go to levels < u and levels > u in time T . (1. Hence P(MT>u. Then the density and c.ST<u) = P(MT>u. (i).. MT > U) = P(ST > u) + P(ST > u) (1.1a for computing ruin probabilities for a twostep premium function.T) P(MT > u) where MT = maxo<t<T St.d. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .. MT > u) = P (ST > u) + P (ST > u. Here {St } is Brownian motion with drift 0 (starting from 0).. T(u) E dT.4) I = . RANDOM WALK.Rt}.8 ).11) VIT ) Proof For p = 0. T ) = P(T(u) < T ) = 241.f.11 ) is the same as (1.8 Let {Rt} be Brownian motion with drift .9) = 2P(ST > u). the density dPµ / dP0 of St is eµsttµ2/2. P(MT > u) = P(ST > u) + P(ST < u. Corollary 1. TWO BARRIERS 301 Note thas this argument has already been used in VII. BROWNIAN MOTION. ( 1. 10) follows then by straightforward differentiation.r(u). 0(u.10) Pµ (T(u) < T) !.. and hence Pµ('r(u) E dT) = Eo [e µsr(. For µ # 0.ST>U). For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. + µ2T) } . = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T . we have ili(u. (1.µ%T (1.µ T I + e2µ"4) ( . We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities.. and (1 . = 1 .1.2 .8) Proof In terms of the claim surplus process { St} = {u .)_ _( u)µ2 /2.
that 0(u).11) then follows by checking that the derivative of the r. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). as defined above as the probability of actually hitting 0. whenever u.10 Consider a diffusion process {Rt} on [0.8 also applies to the case 9 54 1/2.10) and that the value at 0 is 0.. Here {2T( (v}TT)/2) v=T. T are integervalued and nonnegative... MISCELLANEOUS TOPICS which is the same as (1.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x).9 For Bernoulli random walk with 9 = 1/2. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. see e. Breiman [78] or Karlin & Taylor [222] p. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. Proof The argument leading to ( 1.g.T2.s.T+2.T (1.13) with 0 as lower limit of integration. Let s(y) = ef0 ry(.12) is the same as ( 1. We finally consider a general diffusion {Rt} on [0.T)dx. oo).302 CHAPTER XI. and in a similar spirit as in VII.9) goes through unchanged.12) P(ST = v) = 0 otherwise. (1. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 . If this assumption fails.T) = P(ST = u) + 2P (ST > u). 0 0 (1.g. 226). oo) with drift µ(x) and variance a2 (x) at x. and (1. The same argument as used for Corollary 1.. is finite for all x > 0. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.10). u Small modifications also apply to Bernoulli random walks: Proposition 1. as defined in (1. but we omit the details. is (1.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion. e. S(x) = f x s(y)dy. S(oo) = f c s(y)dy. Thus. the behaviour at the boundary 0 is more complicated and it may happen.h.9). Theorem 1. Vi(u.
Letting a T oo and considering the cases S(oo) = oo. In view of (1.b(a) = 0 then yield the result. If b < u < a.ba.b(u) + L. then 0 < 2l. 15) i. if (1.6 to Markovmodulated models . A classical reference for further aspects of Bernoulli random walks is Feller [142].S(u)/S(a).b(u) be the probability that {Rt} hits b before a starting from u. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. O. BROWNIAN MOTION. see in particular pp.b = 0 implies that VQ b/s is constant.14) fails. Lemma 1. TWO BARRIERS 303 for x > 0. i.b('u) = Eu &0.b (Rdt) = Oa.e.16) yields 4b (u) = 1 . elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] . the function S(x) is .0(u) = 1 for all u > 0.b(b) = 1.16). Assume further that S (x) as defined in (1. (1 .13) is finite for all x > 0.S(u) (1.b = a+/3S. Then YIa. (1.14) S(oo) < 00. RANDOM WALK. so that Y)n.10.17) Hence L.b(u)dt. Further references on twobarrier ruin problems include Dickson & Gray [116]. The obvious boundary conditions '0a. For generalizations of Proposition 1.16) S(a) . . Notes and references All material of the present section is standard.b(u) = S(a) .11 Let 0 < b < u < a and let t&0. Using s'/ s = 2p/a2. Wa.ba.e LVa.1.b(Rdt).b('u) = Eu . we can ignore the possibility of ruin or hitting the upper barrier a before dt. [117]. and we get Wo. Letting b J. 0 in (1. A good introduction to diffusions is in Karlin & Taylor [222]. E„ q(Rdt) = q(u)+Lq(u)dt. b = 0. S(oo) < oo separately u completes the proof. 0 Proof of Theorem 1. 191195 for material related to Theorem 1.(u) < 1 for all u > 0 and ^ S^ Conversely. 1'.S(b) Proof Recall that under mild conditions on q.10. then. see Asmussen & Perry [42]. If (1.
(.304 CHAPTER XI.4) I.ytc (ay).aR. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u. IV. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. equivalently.6) . Remark 11. defined by the density 1/va(u)s(u) showing up in (1.1.aRo . variance 0. yu) where W (ay) = y. yielding eau = Ee.3) < e 7yu. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications.t&(u. Markovmodulated Brownian models . Another basic quantity is the speed measure M . (2.1) (2.13)). correponding to piecewise linear paths or .3. much of the literature dels with the pure drift case.2) C_e7u < t(u) < C+e _7u. 1 y < k (y).5) A martingale proof of (2. one works instead with a lower limit 5 > 0 of integration in (1. Lo is a martingale (cf.4.o•K(a) = Ee . 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.4.17). (2..9 ) and optional stopping applied to the stopping time r(u) A T. 111 . with the drift and the variance depending on an underlying Markov process .6. y > .(a) (2. The emphasis is often on stationary distributions . (2.5. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature. is currently an extremely active area of research.)AT . information on ruin probabilities can be obtained . but by duality.1 ) was given already in II. and here are alternative martingale proofs of the rest . (2. 7y = ay . They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu.(7) .5): _ z/'(u) < e 7u.(T(u)AT) r. Lo I. yu) '+/1(u) .2.
4): We take a = ay in (2.yuk (ay)(u&(u. y > r. For (2.3).2. when Rt_ = r.f. Let H(dt.4). eyuk (ay) = e7yu e > eyu"(ay ) ij(u. we have tc(ay) > 0 and we can bound (2. (2.. it follows easily from (2. FURTHER APPLICATIONS OF MARTINGALES 305 (we cannot use the stopping time r(u) directly because P(r(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem). (B(y) . Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt. so that i/1(uL yu) < eayu . Equivalently.1.T)  V. u Proof of (2. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.(u. Rt has distribution B(r + dy)/B(r). dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.yu))• b(u.1.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0).3). yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II. and the proof of the lower inequality is similar. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).1 . dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows.yu) Y Similarly for (2.E [e. dr) denote the conditional distribution of (T(u).2): As noted in Proposition II. A claim leading to ruin at time t has c.6) below by 1 E Le7Rr(.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e.6) with = 'y that eyu . .d.7R.( u ) I T(U) < 00] .)r(u)r.6). Proof of ( 2. RT(u)_) given r(u) < oo.B(r))/B(r).
large deviations results have usually a weaker form. For example.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. (3. + X. logarithmic asymptotics . gn 4 0. Thus.gn if nioo lim 109 fn = 1 log gn (later in this section.1) is an example of sharp asymptotics : . such that the cumulant generating function r. however . However .g. if x > EX1.1 We will go into some more detail concerning (3. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . 1) but only the dominant exponential term . The advantage of the large deviations approach is. large deviations results been.3na with a < 1.1) does not capture the \ in (3. cle .(B) = log EeOX 1 is defined for sufficiently many 0.1).means (as at other places in the book) that the ratio is one in the limit (here n * oo). . The classical result in the area is Cramer's theorem.. og For sequences fn.the correct sharp asymptotics might as well have +. we will write fn 1./n E I) for intervals I C R. then P C S. (3. gn with fn + 0 . not quite so much in insurance risk.. and gave sharp asymptotics for probabilities of the form P (S. v2 later..^ e nn 1 > x n 0o 2xn (3. Example 3. Cramer considered a random walk Sn = X1 + . Thus . the parameter will be u rather than n). . ri. in being capable of treating many models beyond simple random walks which are not easily treated by other models . e.1) where we return to the values of 0.2). Accordingly. MISCELLANEOUS TOPICS 3 Large deviations The area of large deviations is a set of asymptotic results on rare event probabilities and a set of methods to derive such results.nn or C2e. and that a considerable body of theory has been developed.. The last decades have seen a boom in the area and a considerable body of applications in queueing theory. which in the setting of (3.306 CHAPTER XI. its generality. nroo n n /// Note in particular that (3.2) can be rewritten as F (Sn/n > x) 1g a'fin.?n typically only give the dominant term in an asymptotic expression . The limit result (3.
3.960/) * 0. rc*(x) = sup(Ox . if we replace Sn by nx + o / V where V is N(0. V > 0 e. 2 where o2 = o2(x) = rc"(0). LARGE DEVIATIONS Define rc* as the convex conjugate of rc.r.2).tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn .4) n Next. since Sn is asymptotically normal w. More precisely.(0)) e 307 (other names are the entropy.q = rc* (x). Most often.425.4 enn +1.t.e.r. i.(e)i XI E dx]. Define . the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .the mean rc'(0) of the distribution of X1 exponentially tilted with 0. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function). nx < Sn < nx + 1. S rtn > x 1. we get P(Sn/n > x) E [e9nx +nK(9)9" '. of P(X1 E dx) = E[e9X1K.4) immediately yields (3. In fact. (3.96o /] > 0. and hence for large n P(Sn/n > x) > E [e.1). Since P nn > x) = E {e_8 ' ( 9). exponential change of measure is a key tool in large deviations methods.9S„+n' ( 9).sseo f which in conjunction with (3. P with mean nx and variance no.3) is put equal to x. we have P(nx < Sn < nx + 1. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3.rc(0) where 0 = 0(x) is the solution of x = rc'(0). which is a saddlepoint equation .
.h... which is of similar spirit as the dicussion in VII. (ii) lim supn.3. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y . r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo).e < 8 < y + e. Xn given by Fn(dxl. The substitution by V needs. (iv) tc(ry) = 0 and r. Sanov's theorem which give rare events asymptotics for empirical distributions. .. 1) and no such that Sn .2 (GLYNN & WHITT [163]) Let X1. . that is.'s. In the application of large deviations to ruin probabilities.v.1)... . Pn Sn1 .. be a sequence of r. MISCELLANEOUS TOPICS which is the same as (3.dxn) = 05nKn(7)Fn(dx1. Further main results in large deviations theory are the GartnerEllis theorem... see Jensen u [215] or [APQ] p.. .. For the proof. Assume that there exists 'y. X2.. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 . . (iii) #c (8) = limn..e < 8 < y + e. to be made rigorous. however.308 CHAPTER XI. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. Then i/./^ >7 < zn n for n n0.. We shall need: Lemma 3 . 1].p > 7 < zn. we introduce a change of measure for X1.dxn) where Fn is the distribution of (X1i . there exists z E (0. integrates to 1 by the definition of Icn)..'(u) )Ng a"u. Xn) and sn = x1 + • • • + xn (note that the r. n Icn(0) exists and is finite for ry . Ee9X n < oo for e < 0 < e. Mogulskii's theorem which gives path asymptotics. We further write µ = tc'(ry).3 For each i > 0. and the WentzellFreidlin theory of slow Markov walks. is differentiable at ry with 0 < K'(y) < 00. and write Sn = X1 + • • • + Xn... 260 for details.s. commonly denoted as is the saddlepoint approximation.o log Ee9Sn /n. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0.
in particular the r. Since I EeqOX „ ] 1/q is bounded for large n by (ii).n m µ 1 + rl . u Proof of Theorem 3. This establishes the first claim of the lemma . can be chosen strictly negative by taking 9 small enough. for Sn. ( U) P(S.Kn(7)e'n (p(O +7))/p I Ee geXn]1/q where we used Holder's inequality with 1/p+ 1/q = 1 and p chosen so close to 1 and 0 so close to 0 that j p(0 +.91) + o(O ) as 0 J. Clearly. > 1 +17] m(7).n > u ) = [ Em [em Em 1e. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1. S.m(7)...+r. The rest of the argument is as before.Bµ . LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3.W.s.s..s.71 < e and jq9j < e.. The corresponding claim for Pn(Sn/n < µ .]1/q = e.> .r (7) n = e.2.3. 0.YS. h.ne(p+ 17). can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0. Then V. This proves the existence of z < 1 and no such that Pn (Sn/n > µ. h.2. We first show that lim inf„_.77) follows by symmetry (note that the argument did not use µ > 0).077 n^oo n and by Taylor expansion and (iv ). log zl'(u)/u > 'y. mµ Sm > u] km e7Sm+n. is of order .+r7) < zn for n > no.y) .. we get lim sup 1 log Pn (Sn1 /n > µ + r7) < 0(1i + r7) + i(p(0 +'Y))/p n+oo n and by Taylor expansion.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . S. it is easy to see that the r. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.h. For Sn1i we have Fn(Sn 1/n > µ+r7) < ene(µ+ 1?)EneeS„1 = ene ( µ+n)EneeSneX„ eno(µ +n) Ee(e+7)Sn ex„ wn (7) < e.µ?7 . the r .
(•) goes to 1 by Lemma 3. logO(u)/u > ry. Pn \ > la+ 8 I < zn (3. n=1 .YS +^c CHAPTER XI. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. 3. P(T(u) = n) < P(Sn > u) = En [e7S. and since Ic. 0 yields liminfu __.6) for some z < 1 and all n > n(E)..n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. we write P(T(u) = n) = Il + I2 + I3 + I4 'i/I(u) _ E00 n=1 where n(b) Lu(10/µJ Ii = 1: F(T(u) = n).+wn(7).n(ry)/u 4 0andm/u* (1 + r7)/µ. For lim supu... we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J.. I2 = F(T(u) = n). n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).0 log i'(u )/u < 'y. MISCELLANEOUS TOPICS (7). Obviously.3. Sn > u] < eYu+Kn(7)pn(Sn > u) (3.7) so that n(b) I1 < e'Yu E en. this is possible by (iii).(Y)... (iv) and Lemma 3.I < µl1 1+77 I M 1_ 1+277 S.log z) /2 and Sn Fn\ n >lb+S) <Zn. I > IL exp `S..310 ]Em I e.
S.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 . Sn1 C U.3. u .11) [u(1+6)/µJ+1 1  Thus an upper bound for z/'(u) is n(6) e'Yu n=1 eKn (7) + 2 + (28U + 1) e6u(1+6)/µ Fi 1 zl /2 and using (i). > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) . LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^.10) 00 I4 < E F(Sn_1 < u. Sn > U] [ e(u(1+6)/µJ+l < eYu (u(1+6)/µJ+1 7u r 0 0 e L^ en('Y ) fPn (I Sn 1 . we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0. eryu en logz/2p n nt n. C 26u `p / +1 I e6u(1+6)/µ (3.zl/z en6 [u(1 +6)/µJ 1u (1 +6) /µJ ekn(7) < e' 13 < C" E Yu l u(16)/lij+1 Lu(16)/µJ+l1 < e7U Finally.
7' a"ju. For 14. For 12.4/3rc'(y) > 0. this is straightforward since the last inequality in (3.4 Under the assumptions of Theorem 3.(u) = I1+I2+I3+I4'^ ery( u). MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs.Q is so small that w = 1 . Then for n large. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 . 2. 4 there is an aj > 0 and a cj < oo such that Ij < c3e. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7).b)/i(7). 13 = P(T (u) E (u(1 b)l^ (7). e'.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . cf.u(1+b)/rc'(7)).11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 .2. ryue«iu .312 CHAPTER XI. it suffices to show that for j = 1.z 1/z For I1. u(1 + b)/i(7)) Proof Since V. (7 + a) < 2arc'(7). For I.('+'Y). I2.3ui where . the typical time is u/rc'(7) just as for the compound Poisson model. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r.8) by P(S. we replace the bound P(Sn > u ) < 1 used in (3.. say n n1.. we get Lou] E exp {( 7 + a)u + Kn(a +7)} n=1 Il Lou] exp {(y + a)u} { 111 + exp {4narc'(7)} n=1 exp {('y + a)u} c1 exp {4/3uarc'(7)} = clewhere a1 = aw. Letting c11 = maxn<n. we need to redefine n(b) as L. the last steps of (3.4.xl/2 to give the desired conclusion. Corollary 3. u . IV.
.2 shows that the discrete time structure is used in an essential way.1.'(y) > 0. It is then wellknown and easy to prove that Sn has a normal distribution with mean np and a variance wn satisfying i lim wn = wz = Var(X1 ) + 2 E Cov(Xl. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate . Obviously many of the most interesting examples have a continuous time scale.2 then immediately yields the estimate log F( sup Skh > u) a7u (3. and we conclude that Theorem 3 .12) k=0... To verify these in concrete examples may well present considerable difficulties.3(s) at time s. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3. V(s) with m.1. the key condition similar to (iii). If {St}t> 0 is the claims surplus process.. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0. whether P ( sup St > u ltg a ^" 0<t<oo // (3.14) is needed in both examples . (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0. An event occuring at time s is rewarded by a r. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process. Assuming that the further regularity conditions can be verified. Xk+l) k=1 00 naoo n provided the sum converges absolutely. Thus the total reward in the interval [0. 09(9). criteria are given in Duffield & O'Connell [124].13) One would expect this to hold in considerable generality. Theorem 3. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * ... and in fact.g.e. r.(O) = 9µ+02 for all 9 E R.LARGE DEVIATIONS 313 Example 3 . 11 Inspection of the proof of Theorem 3. t] is Rt = E V (Un) n: o„ <t .5 Assume the Xn form a stationary Gaussian sequence with mean p < 0.. i. we shall give two continuous time examples and tacitly assume that this can be done.f.v. but nevertheless.. 2 is in force with y = 2p/wz.3.
. Kt (0) t (Ee9U"it8i J0 .14).v. if the nth claim arrives at time a.9t = /3 J t (Ee8U° i8l . is At . nondecreasing and with finite limits Un as s T oo ( thus. the best estimator of /3µB based upon Ft.1) ds . We further assume that the processes {U1(s)}8>0 are i.'`1 U. Then logEeOR° = J0 /3(s)(^8(9) . MISCELLANEOUS TOPICS are the event times. Un represents the total payment for the nth claim).. e.. 0 Example 3 .Q„) . (3. assuming a continuous premium inflow at unit rate. this is not realistic .d. where Ft = a(A8 : 0 < s < t). n: o. O'n +S] is a r .It. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics .Lundberg model has the larger ruin probability.s). we conclude that Cu) log e7 u (cf. Most obviously. derive . It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot .15) ..1) ds rt (3.1) .14) (to see this . it contributes to St by the amount Un(t .314 where the an CHAPTER XI. e > 0 such that ic('y) = 0 and that r. . Un(s). Thus by (3. More precisely. <t which is a shotnoise process. Of course. leading to St = At(1+77) Joo t S8 ds. If the nth claim arrives at time Qn = s. 7 Given the safety loading 77. we have S.1) ds . (9) < oo for 9 < 'y + C. a differential equation in t). we have rct (9)/t 4 ic (9). but that a claim is not settled immediately.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . 0 and since EeOUn(8) + Ee°U^ as s * oo.9t. Example 3. then the payments from the company in [on. i.t. Of course . one would take p(t) = (1 + rt)At/ t.2 are trivial to verify. Since the remaining conditions of Theorem 3. the Cramer. We let ic (9) = 3(EeWU° . Thus. Thus. the above discussion of discrete skeletons).0 and assume there are y. At = .g.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. = U„ ( t .
e. the Vi = .b(u) IN a'Yu (cf. which yields eau f 1 t(1+n )audtl = E r Ee°Y = E [O(1+n)aueaul = E [eau J L Jo J L1+(l+r))aUJ .log Oi are i. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 . To see this .i.d.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.i. It then follows from (3.(1 + 17)0µB = 0.2 hold. again the above discussion of discrete skeletons) where y solves ic('y) = 0 It is interesting to compare the adjustment coefficient y with the one y* of the CramerLundberg model.(1 +i) f > i= 1 s ds = E Ui 1 . LARGE DEVIATIONS With the Qi the arrival times.d. uniform (0. we conclude that t. (3. standard exponential .1) .19) with equality if and only if U is degenerate.1) or .20) (3. one has y > y' (3. and since the remaining conditions are trivial to verify.21) This follows from the probabilistic interpretation Si EN '1 Yi where Yi = Ui( 1+(1 +r7)log ©i) = Ui(1(1 +17)Vi) where the Oi are i . Thus. the solution of /3(Eelu . typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model . equivalently.18) Thus (iii) of Theorem 3. Indeed./3. (3.(1 + r7) log t (3. we have Nt t N. Ui Nt / t 01i 315 St = Ui .16) i=1 o i=1 Let ict (a) = log Eeast . i.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.3.
this in turn yields y > y*. say one year. a* (s) are convex with tc'(0) < 0 . This is a topic of practical importance in the insurance business for assessing the probability of a great loss in a period of length t. the function k(x) = e7*x . see also Nyrhinen [275] for Theorem 3. though we do not always spell this out. Further applications of large deviations idea in risk theory occur in Djehiche [122].(1 + ri)y*x is convex with k(oo) = 00.20) is due to Tatyana Turova. rc*' (0 ) < 0. For notational simplicity. The main example is Nt being Poisson with rate fit. are i. and since tc(s).d. 0 < x < x0. and k(x) < 0. using that Ek(U) = 0 because of (3. y = y* can only occur if U . the study is motivated from the formulas in IV.1 E [1+(1+77)y*U] 0 k (+ *y B(+ 1 + (1(+71)y*y B(dy) L xa 1 + f + (1 + rl) Y* xo jJxo k(y) B(dy ) + f' k(y) B(dy) } = 0. This implies n(y*) < 0. Therefore e7'U _ k(U) E [1+(1+77)y*U] . Lehtonen & Nyrhinen [244]. MartinL6f [256]. In particular. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. see Nyrhinen [275] and Asmussen [25]. For Example 3.2. [257] and Nyrhinen [275].2 expressing the finite horizon ruin probabilities in terms of the distribution of A.xo. MISCELLANEOUS TOPICS Next. much of the analysis carries over to more general cases. Further. k(0) = 0. we are interested in estimating P(A > x) for large x. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U.19). we then take t = 1 so that p. = P(N = n) = e(3an However. In addition to Glynn & Whitt [163]. [245].i. assuming that the U.1 . the proof of (3.7.316 CHAPTER XI. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. x > x0. with common distribution B and independent of Nt. at time t.. k'(0) < 0. Further. .
A > x) eex+K(e ) ee AB°[ely 1 ev2/2 dy 0 2^ 00 9x+p(e) e ezez2/(2BZpB „[9)) dz 9 27r/3B" [9] fo eex+w ( e) oo z x)] ] 0 27r /3B" [9] o e 9 2 /3B" [9] J eex+w(B) dz . A > x)] = eex+K( e)E9 [e .e. Proposition 4."(0) = . The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). B"[s] = oo.2) implies that the limiting Pedistribution of (A . In particular. K'(0) _ ic'(9) = x. no(a) = logE9e'A = rc(a + 9) . i.3B"[9].3e(bo[a] . Hence P(A > x) = E e [e9A+ ic(9).1).9(Ax).3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx).1). e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x.4. we define the saddlepoint 9 = 9(x) by EBA = x. only with 0 replaced by a9 and B by B9. (4. Then as x * oo. Then Ee"A = e'(") where x(a) _ 0(B[a] .[s])3/2 = 0. 818' where s' = sup{s : B[s] < oo}.ic(9) = . The analysis largely follows Example 3.x)//3B"[9] is standard normal. Vare(A) = s.1) where )30 = . A E dx] .1.1 Assume that lim8T8. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. B"' [s] lim (B". For a given x. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution.
under the Poisson assumption (4.x') where x' = sup {x : b(x) > 0}. A covers the exponential distribution and phasetype distributions.ycix °ie6x B. 1 . it is quite questionable to use (4. In fact. 3 A word of warning should be said right away : the CLT (and the Edgeworth expansion) can only be expected to provide a good fit in the center of the distribution . b is gammalike.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. see Embrechts et al. For example. the distribution of A is approximately normal . MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward.2 If B is subexponential and EzN < oo for some z > 1. more generally. then P(A > x) . Jensen [215] and references therein. leading to P(A > x) :. Var(A) _ ^3p.2) is often referred to as the Esscher approximation. B covers distributions with finite support or with a density not too far from ax° with a > 1. b is logconcave. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r.v. The present proof is somewhat heuristical in the CLT steps.1). 2). and (4. i. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. For details.3) and related results u for the case of main interest .2i and that (A . For a rigorous proof.l3pB.318 CHAPTER XI. some regularity of the density b(x) of B is required. In particular. bounded with b(x) .(D X . either of the following is sufficient: A. Notes and references Proposition 4.EN B(x). Furthermore 00 b(x)Sdx < oo for some ( E (1. Thus . large x. just the same dominated convergence argument as in the proof of Theorem 2. [138]. 4b The NP approximation In many cases .1 yields: Proposition 4. or. For example. it holds that EA = .Q{AB (4. b(x) = q(x)eh(z). Remark 4 .4) .e.1 goes all the way back to Esscher [141].(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo.
defined as the the solution of P(A < yle) = 1 . A particular case is a... (4.5) follows by integration.99.2 2 .. in particular.equantile in the distribution of Y. one expects the u3 term to dominate the terms of order u4. so that 1(u) 3 exp { . zl_e be the 1 . are the cumulants . (4.6) . K2 = Var (Y). In concrete examples . one needs to show that 163. then P(Y < y) 4(y) .5)..6(1 .y2)^P(y)• 319 Note as a further warning that the r.. Heuristically. the CLT for Y = Y6 is usually derived via expanding the ch.f.. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) . .e.1). as u2 u3 u4 9(u) = Ee'uY = exp {iuci .i 6 r 1 3 so that we should take b = ic3/6 in (4.s..4. where Kl .5) is obtained by noting that by Fourier inversion. the density of Y is 1 °° _ eiuy f(u) du 2x _.h.5) may be negative and is not necessarily an increasing function of y for jyj large. and from this (4. are small. and so as a first approximation we obtain a1_E = EA + yle Var(A) .2X2 . u5. Let Y = (A . s. Var(Y) = 1 as above .i 3 K3 } Pt^ exp .: EA + zl_E Var(A) ..c2i. however. of (4. Thus if EY = 0. . Rather than with the tail probabilities F(A > x). .l = EY. ylE should be close to zl_E (cf.3& (y). If the distribution of Y is close to N(0.5 (y3 .EY)3.EA)/ Var(A) and let yl_E. K4 . If this holds .3!). the standard normal distribution.2K3 + 4i 64 + .2 ^ \1 . THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter.. resp. the NP (normal power) approximation deals with the quantile al_E. Remark 4. which is often denoted VaR (the Value at Risk). K3 = E(Y .
[101]. In particular .E(/3PB^1 )1^2 + s(z1E . this yields the NP approximation 6(Z1 _E . Using Y = (A . b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .7) as 1 (3) a1E = Qµa +z1 .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4. We can rewrite (4.y2)cp( y) term...zl E)^o(zl E) . k3 is small for large /3 but dominates 1c4. Note. however. .5) by noting that the 4.6pBki) d/2.E )Azl E) 4(z1E) + ( ylE .E)A1 l E) 1 E 4)(yl E) ^' .6 (1 .1)EY3.EA ) / Var(A).E . For example.(y) terms dominate the S(1 .zlE)W(zlE) 1 . and assume that there exist n ) Pn_i . let pn Pn = (a+ = P(N = n). n = 1. MISCELLANEOUS TOPICS A correction term may be computed from (4. that [101] distinguishes between the NP and Edgeworth approximations.E = z1E + S(zi_E . the kth cumulant of A is /3PBk' and so s.1)^ 2) µ'E Notes and references We have followed largely Sundt [354].1)! n ^eQ ..E + (yl..yi.S(1 . b such that EN 1 U%...3ni /3 . this holds with a = 0. K5 .320 CHAPTER XI. This leads to t( yl E) .1) E (A . 4c Panjer 's recursion Consider A = constants a.k = /3µB^1 / (.S(1 .1). 21 .zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 .. as required .5(1 . Another main reference is Daykin et at.zi.zl E)V(zl_E) .
2. (4..11) Remark 4. j1 g. fj = E (a+ b k =1 )9kfi_k . then j (a + b!) 1ag k_1 3 gkfj. E[a +bU=I >Ui =j l i=1 J (4.4 Assume that B is concentrated on {0. . j = 0..10) f o = po. the complexity (number of arithmetic operations required) is O(j3) for (4. and calculating the gj*n recursively by 9*1 = 9j.4 is that the algorithm is much faster than the naive method. j = 1. j = 1. . fj = P(A = j).. ...13) Namely. .14) is therefore a + b/n.13) but only O(j2) for Proposition 4. the value of (4.} and write gj = 2 .4. . The expression for fo is obvious. (4.1.12) where g*n is the nth convolution power of g. 1.. 2. . Since the sum over i is na + b. .5 The crux of Proposition 4.. u Proof of Proposition 4. Then fo = >20 9onpn and fi = 1 E In particular. (4.12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 . n.14) is independent of i = 1.k . By symmetry.4. . (4.12). THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4...... n = k=n1 9k(n1 )9j k • (4.9). Hence by (4. 2. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4. . if go = 0.4.
322
00 J
CHAPTER XI. MISCELLANEOUS TOPICS
EE (a + bk I gkg3 _ k lieni n=ik=0 (a+bk l gkE g j'`kpn = E (a+b!)9kfi_k n=0 k=0 k=0 ^I 1 E(a+b. agofj+ k Jgkfjk, k=i /
and and (4.9) follows . (4.11) is a trivial special case.
u
If the distribution B of the Ui is nonlattice , it is natural to use a discrete approximation . To this end, let U(;+, U(h) be U; rounded upwards, resp. downwards , to the nearest multiple of h and let A}h) = EN U. An obvious modification of Proposition 4.4 applies to evaluate the distribution F(h) of A(h) letting f( ) = P(A() = jh) and
g(h) gkh+
= P (U(h2 = kh) = B((k + 1)h)  B(kh ), k = 0, 1, 2, ... , = P (U4;+ = kh) = B(kh)  B (( k  1)h) = gk  l,, k = 1, 2, ... .
Then the error on the tail probabilities (which can be taken arbitrarily small by choosing h small enough ) can be evaluated by
00 00
< P(A > x ) f (h) j=Lx/hl j=Lx/hl
Further examples ( and in fact the only ones , cf. Sundt & Jewell [355]) where (4.9) holds are the binomial distribution and the negative binomial (in particular, geometric ) distribution . The geometric case is of particular importance because of the following result which immediately follows from by combining Proposition 4.4 and the PollaczeckKhinchine representation: Corollary 4.6 Consider a compound Poisson risk process with Poisson rate 0 and claim size distribution B. Then for any h > 0, the ruin probability zb(u) satisfies 00 00
f^,h) Cu) < E ff,+, j=Lu/hJ j=Lu/hJ (4.15)
f! h)
5. PRINCIPLES FOR PREMIUM CALCULATION
where f^ +, f^ h) are given by the recursions
(h) 3 (h) (h)
323
fj,+ = P 9k fjk,+ ' I = 17 2, .. .
k=1 3 (h)
(h)
=
P
(h)
f9,  (h) gk,fAk, e 1  ago, k=1
j = 1+2,
starting from fo + = 1  p, f(h) = (1  p)/(1  pgoh) and using 07
g(kh) 1 (k+1)h
=
Bo((k + 1 ) h)  Bo(kh ) =  f
AB
kh
B(x) dx, k = 0, 1, 2, ... , k = 1,2 .....
gkh+
Bo(kh )  Bo((k  1 ) h) = 9kh)1 ,
Notes and references The literature on recursive algorithms related to Panjer's recursion is extensive, see e.g. Dickson [115] and references therein.
5 Principles for premium calculation
The standard setting for discussing premium calculation in the actuarial literature does not involve stochastic processes, but only a single risk X > 0. By this we mean that X is a r.v. representing the random payment to be made (possibly 0). A premium rule is then a [0, oo)valued function H of the distribution of X, often written H(X), such that H(X) is the premium to be paid, i.e. the amount for which the company is willing to insure the given risk. The standard premium rules discussed in the literature (not necessarily the same which are used in practice!) are the following: The net premium principle H(X) = EX (also called the equivalence principle). As follows from the fluctuation theory of r.v.'s with mean, this principle will lead to ruin if many independent risks are insured. This motivates the next principle, The expected value principle H(X) = (1 + 77)EX where 77 is a specified safety loading. For 77 = 0, we are back to the net premium principle. A criticism of the expected value principle is that it does not take into account the variability of X which leads to The variance principle H(X) = EX+77Var(X). A modification (motivated from EX and Var(X) not having the same dimension) is
324
CHAPTER XI. MISCELLANEOUS TOPICS
Var(X).
The standard deviation principle H(X) = EX +rl
The principle of zero utility. Here v(x) is a given utility function, assumed to be concave and increasing with (w.lo.g) v(O) = 0; v(x) represents the utility of a capital of size x . The zero utility principle then means v(0) = Ev (H(X)  X); (5.1)
a generalization v(u) = Ev (u + H(X)  X ) takes into account the initial reserve u of the company. By Jensen 's inequality, v(H(X)  EX) > Ev(H(X)  X) = 0 so that H(X) > EX. For v(x) = x, we have equality and are back to the net premium principle. There is also an approximate argument leading to the variance principle as follows. Assuming that the Taylor approximation
v(H(X)  X) ^ 0 +v'(0)(H (X)  X) + v 0 (H(X)  X)2 ,/2
is reasonable , taking expectations leads to the quadratic v"H(X )2 + H(X) (2v'  2v"EX) + v"EX2  2v'EX = 0 (with v', v" evaluated at 0) with solution
H(X)=EXv^±V( ^ )2Var(X).
Write
( vI ) 2 \
Var(X) v^  2v^Var(X)/ I  (
, Var(X) )2
If v"/v' is small, we can ignore the last term. Taking +f then yields H(X) ,:: EX 
2v'(0) VarX;
since v"(0) < 0 by concavity, this is approximately the variance principle. The most important special case of the principle of zero utility is The exponential principle which corresponds to v(x) = (1  e6x)/a for some a > 0. Here (5.1) is equivalent to 0 = 1  e0H(X)EeaX, and we get
H(X) = 1 log Ee 0X .
a
5. PRINCIPLES FOR PREMIUM CALCULATION
325
Since m.g.f.'s are logconcave, it follows that H,, (X) = H(X) is increasing as function of a. Further, limQyo Ha (X) = EX (the net premium princiHa (X) = b (the premium ple) and, provided b = ess supX < oo, lim,, H(X) = b is called the maximal loss principle but is clearly not principle very realistic). In view of this, a is called the risk aversion The percentile principle Here one chooses a (small ) number a, say 0.05 or 0.01, and determines H(X) by P(X < H(X)) = 1  a (assuming a continuous distribution for simplicity). Some standard criteria for evaluating the merits of premium rules are 1. 77 > 0, i .e. H(X) > EX. 2. H(X) < b when b (the ess sup above ) is finite 3. H(X + c) = H(X) + c for any constant c
4. H(X + Y) = H(X) + H(Y) when X, Y are independent
5. H(X) = H(H(XIY)). For example , if X = EN U= is a random sum with the U; independent of N, this yields
H
C^
U; I = H(H(U)N)
(where, of course, H(U) is a constant). Note that H(cX) = cH(X) is not on the list! Considering the examples above, the net premium principle and the exponential principle can be seen to the only ones satisfying all five properties. The expected value principle fails to satisy, e.g., 3), whereas (at least) 4) is violated for the variance principle, the standard deviation principle, and the zero utility principle (unless it is the exponential or net premium principle). For more detail, see e.g. Gerber [157] or Sundt [354]. Proposition 5.1 Consider the compound Poisson case and assume that the premium p is calculated using the exponential principle with time horizon h > 0. That is,
N,,
Ev I P  E U;
i =1
= 0 where
v(x) = 1(1  e°x
a
Then ry = a, i.e. the adjustment coefficient 'y coincides with the risk aversion a.
326
Proof The assumption means
CHAPTER XI. MISCELLANEOUS TOPICS
0 a (1  eareo (B[a11)
l
i.e. /3(B[a]  1)  ap = 0 which is the same as saying that a solves the Lundberg u equation. Notes and references The theory exposed is standard and can be found in many texts on insurance mathematics, e.g. Gerber [157], Heilman [191] and Sundt [354]. For an extensive treatment, see Goovaerts et al. [165].
6 Reinsurance
Reinsurance means that the company (the cedent) insures a part of the risk at another insurance company (the reinsurer). Again, we start by formulation the basic concepts within the framework of a single risk X _> 0. A reinsurance arrangement is then defined in terms of a function h(x) with the property h(x) < x. Here h(x) is the amount of the claim x to be paid by the reinsurer and x  h(x) by the the amount to be paid by the cedent. The function x  h(x) is referred to as the retention function. The most common examples are the following two: Proportional reinsurance h(x) = Ox for some 0 E (0, 1). Also called quota share reinsurance. Stoploss reinsurance h(x) = (x  b)+ for some b E (0, oo), referred to as the retention limit. Note that the retention function is x A b. Concerning terminology, note that in the actuarial literature the stoploss transform of F(x) = P(X < x) (or, equivalently, of X), is defined as the function
b * E(X  b)+ =
f
(s  b)F(dx) _ f
6 00
(x) dx.
An arrangement closely related to stoploss reinsurance is excessofloss reinsurance, see below.
Stoploss reinsurance and excessofloss reinsurance have a number of nice optimality properties. The first we prove is in terms of maximal utility: Proposition 6.1 Let X be a given risk, v a given concave nondecreasing utility function and h a given retention function. Let further b be determined by E(X b)+ = Eh(X). Then for any x,
Ev(x  {X  h(X)}) < Ev(x  X A b).
6. REINSURANCE
327
Remark 6 .2 Proposition 6.1 can be interpreted as follows. Assume that the cedent charges a premium P > EX for the risk X and is willing to pay P1 < P for reinsurance. If the reinsurer applies the expected value principle with safety loading q, this implies that the cedent is looking for retention functions with Eh(X) = P2 = P1/(1 + 77). The expected utility after settling the risk is thus
Ev(u + P  P1  {X  h(X)})
where u is the initial reserve . Letting x = u + P  P1, Proposition 6.1 shows that the stoploss rule h (X) = (X  b)+ with b chosen such that E(X  b)+ u = P2 maximizes the expected utility. For the proof of Proposition 6.1, we shall need the following lemma: Lemma 6 .3 (OHLIN'S LEMMA) Let X1, X2 be two risks with the same mean, such that Fj(x) < F2 (x), x < b, Fi(x) ? F2(x), x > b for some b where Fi(x) = P(Xi < x). Then Eg(X1) < g(X2) for any convex function g. Proof Let Yi=XiAb, Zi=Xivb.
Then
P(Yl < x) _ Fi(x) <_ F2 (x) = P(Y2 < x) x < b 1=P(Y2<x) x>b so that Y1 is larger than Y2 in the sense of stochastical ordering . Similarly, P(Zl < x) _ 0 = P(Z2 < x) x < b Fi(x) > F2(x) = P(Z2 < x) x > b
so that Z2 is larger than Zl in stochastical ordering. Since by convexity, v(x) = g(x)  g(b)  g'(b)(x  b) is nonincreasing on [0, b] and nondecreasing on [b, oo), it follows that Ev(Y1) < Ev(Y2), Ev(Zi) < Ev(Z2). Using v(Yi) + v(Zi) = v(Xi), it follows that
0 < Ev(X2)  Ev(Xi) = Eg(X2)  Eg(X1),
using EX1 = EX2 in the last step. u
Proof of Proposition 6.1. It is easily seen that the asssumptions of Ohlin' s lemma hold when X1 = X A b, X2 = X  h(X); in particular, the requirement EX1
328
CHAPTER XI. MISCELLANEOUS TOPICS
= EX2 is then equivalent to E(X  b)+ = Eh(X). Now just note that v is convex. u
We now turn to the case where the risk can be written as N
X = Ui
i=1
with the Ui independent; N may be random but should then be independent of the Ui. Typically, N could be the number of claims in a given period, say a year, and the Ui the corresponding claim sizes. A reinsurance arrangement of the form h(X) as above is called global; if instead h is applied to the individual claims so that the reinsurer pays the amount EN h(Ui), the arrangement is called local (more generally, one could consider EN hi(Ui) but we shall not discuss this). The following discussion will focus on maximizing the adjustment coefficient. For a global rule with retention function h* (x) and a given premium P* charged for X  h* (X), the cedents adjustment coefficient y* is determined by
1 = Eexp {ry*[X  h*(X)  P*]},
for a local rule corresponding to h(u) and premium P for X look instead for the ry solving
J _f
(6.2) N 1 h (Ui), we
[ X_P_^
1 = Eexp
[ Ei  h(Ui)] P [U
= Eexp{ry
h(Ui)]
l (6.3) This definition of the adjustment coefficients is motivated by considering ruin at a sequence of equally spaced time points, say consecutive years, such that N is the generic number of claims in a year and P, P* the total premiums charged in a year, and referring to the results of V.3a. The following result shows that if we compare only arrangements with P = P*, a global rule if preferable to a local one. Proposition 6.4 To any local rule with retention function h(u) and any
N
J}
P > E X  N h(Ui)
4 =1
(6.4)
there is a global rule with retention function h* (x) such that
N
Eh*(X) = Eh(U1)
i=1
and 'y* > ry where ry* is evaluated with P* = P in (6.3).
b)+ = Eh(U) (and the same P) satisfies 71 > ry. N E X . (6.4).b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. ' ii (6. Eexp 7 [E [Ui .6 Assume the Ui are i.P]}.h(Ui)] . Local reinsurance with h(u) = (u . Proof As in the proof of Proposition 6.h(u) and any P satisfying (6.h(Ui)P JJJ l:='l {ry ] or. however.6) u where C[ry] = Ee'r(u4(u)). Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.5) reduce quite a lot. The arrangement used in practice is.. appealing to (6.4.P } < 1 = Eexp E[Ui.3). expectations like those in (6.4).P I = EC [7]N.h(Ui)] . .h * (X) .P.d.h(U)]. we get EX = EN • EU. (6. then (6.6. This follows by taking Xl = U A b. Then for any local retention function u .4). the excess ofloss rule hl (u) = (u . and so on. it suffices to show that Eexp {ry ii 'UiAb. X2 = U .5) holds trivially.b)+ with b determined by E(U .d. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') . Remark 6.h(U) (as in the proof of Proposition 6.6).5 Because of the independence assumptions . Assuming for simplicity that the Ui are i.P > EexP{7[X . i. u But since ry > 0.h(Ui)] .i.h( UU) = EN • E[U . ry* > 0 because of (6. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x . this implies 7* > 7. y = Ei [Ui . we get N 1 = Eexp ry E[Ui ii . as often local as global.4) and u g(x) = e7x in Ohlin's lemma.
The present proof is from van Dawen [99].many texts on insurance mathematics. See further Hesselager [194] and Dickson & Waters [120]. . The original reference for Ohlin's lemma is Ohlin [277]. [76]. e. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. see also Sundt [354]. Bowers et at. Heilman [191] and Sundt [354].g.330 CHAPTER XI.
U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. t 00 (A. Technically. then Stone 's decomposition holds : U = U.. The point process is called a renewal process if Yo.U(t) is the expected number of renewals in (t. i.} for any h > 0.. of epochs or the set Y1.. when t is large. Y2. stating that U(t+a)U (t) ^ a. Y. t]) so that U(t + a) . = T„ .T„_1). note in particular that U({0}) = 1. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . Lebesgue measure dt normalized by the mean to of F. The number max k : Tk_j < t of renewals in [0.e.. 2h. all have the same distribution. not concentrated on {h. Then Blackwell 's renewal theorem holds. of interarrival times and the time Yo = To of the first arrival (that is.t.. The renewal theorem asserts that U(dt) is close to dt/µ..Appendix Al Renewal theory la Renewal processes and the renewal theorem By a simple point process on the line we understand a random collection of time epochs without accumulation points and without multiple points. some condition is needed: that F is nonlattice. The mathematical representation is either the ordered set 0 < To < T1 < . Lebesgue measure for some n > 1).. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. .... If Yo = 0.1) (here U(t) = U([0.. That is. t +a]). the renewal process is called zerodelayed. Y1. are independent and Y1. . denoted by F in the following and referred to as the interarrival distribution. the distribution of Yo is called the delay distribution.r. Y2. . t] is denoted by Nt. .
then Z(u) i f0 z(x)dx . u u PF 4 00. (A. µF (A. ENt 4 1 lb Renewal equations and the key renewal theorem The renewal equation is the convolution equation Z(u) = z(u) + f where Z(u) is an unknown function of u E [0 . oo).3) Further. see [APQ] Ch. Then Z(u) 4 z(oo). i. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. in convolution notation Z = z + F * Z. U Z(u .332 APPENDIX u(t) has limit 1/µ as t 4 oo.2) Z(u) = J0 u z(x)U(dx).4) If F is spread. and F(dx) a known probability measure .x)F(dx). Note in particular that F is spreadout if F has a density f. then it suffices for (A. Both result are valid for delayed renewal processes. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. In 111. z(u) a known function..2) has the unique solution Z = U * z. and that F has a bounded density2. stating that U(t)/t > 1/p.4) that z is Lebesgue integrable with limZ.5) 2This condition can be weakened considerably .1 if F is nonlattice and z (u) is directly Riemann integrable (d. Equivalently.out.9.EN(t) . IV). Under weak regularity conditions (see [APQJ Ch.R.e. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1.i". (A. the statements being EN(t + a) . (A.2 Assume that Z solves the renewal equation (A. (A. but suffices for the present purposes . IV).2).i. z(x) = 0. resp.a. that z(u) has a limit z(oo) (say) as u 4 oo.
A regenerative process converges in distribution under very mild conditions: . . However. . multiply (A.5a. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i.r. Y1 . A stochastic process {Xt}t>0 with a general state space E is called regenerative w. Tk (or. or many queueing processes..(3. Tk and {Xt }o<t<Tk • For example. We let FO. To this end.} be a renewal process. Hence by dominated convergence..x)u(x) dx = z(u( 1 . however.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. Y2. refer to the zerodelayed case. T1.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). i. cycles.3) satisfied by the ruin probability for the compound Poisson model. Eo etc.. . results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. The kth cycle is defined as {XTk+t}o<t<Yk . that F is a probability measure. Here the relevant F does not have mass one (F is defective)... where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). This program has been carried out in III. 1c Regenerative processes Let {T. • . T1.. {Tn} if for any k. a basic reason that renewal theory is relevant is the renewal equation II.k+t }t>o is independent of To.. . Z(u) U = 1 u 1 u f z(u . equivalently. . The simplest case is when {Xt} has i. the present more general definition is needed to deal with say Harris recurrent Markov chains. However. Note. of Yo. asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. z(x) = e7xz(x). that the existence of y may fail for heavytailed F. is called the cycle length distribution and as before.e. Yk ).i.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO).. . 0 PF µF 11 In risk theory. and its distribution does not depend on k. the postTk process {XT. Assuming that y can be chosen such that f °° Ox F(dx) = 1. we let µ denote its mean. F(dx) = e7xF(dx).d.t. The distribution F of Y1.. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths.
.ZT Then: (a) If E sup I ZTo+t .. [0. This is the case considered in [APQ] V. but in fact.6) id Cumulative processes Let {Tn} be a renewal process with i.t : t < Tk}.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1. {Tn}..e. If p = oo. C).+ X. then (Zt .oo (i.ZTOI < 00. Then {e(t)}.4 Let {Zt}t^.r. (b) If in addition Var(Ul ) < oo.. for n = 1.e. {Tn}.. oo). e(t )) .t. Then it (ii. in total variation.0 is called cumulative w. under the condition of Blackwell's renewal theorem. We denote the limiting r. assume that p < 00 and define Un = ZT}1 .ZT }0<t<Y„+.. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt.t. and q(t) = sup It .t. fi (t) = inf {Tk . P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo..r. just the same proof as there carries over to show: Proposition A1.. r.3. µ 0 If F is spreadout.'s by e..i. Then Xt Di X.r. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p. Then {Zt}t^.v. 2.d. cycles (we allow a different distribution of the first cycle). resp .3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. oo). {Tn} if the processes {ZT +t . then Xt ..Tk : t < Tk} as the age. then e (t) .0 be cumulative w. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t.d.i. An example is Zt = fo f (X8) ds where {Xt} is regenerative w. i. where the distribution of X. {i7(t)} are Markov with state spaces (0..334 APPENDIX Proposition A1. 0<t<Yi then Zt /t a$• EU1/µ. and we have: holds more generally that (rl(t). are i.. Otherwise . (A.
V is uniform on (0.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli).5 Under the condition of Blackwell's renewal theorem.(t).6 Consider a renewal process with µ < oo. and the equivalence of (a) with (b)(d) is an easy exercise. Y1i Y2.. l:) is the same as the distribution of (VW. Since the maximum Mn of n i.i. Proof The number Nt of renewal before t satisfies Nt/t a4' p.^(t))} as a regenerative process.d. Hence t t lt ) = f U(dy)z(t . . 1) and W has distribution Fw given by dFw/dF(x) = x/pF.APPENDIX 335 Theorem A1. For the second. if in addition EYo < oo.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. the joint distribution of (rl. and the conditional distribution of ri given l. we used: Proposition A1. ^) is given by the following four equivalent statements: (a) P (77 > x. (b) the joint distribution of (ri. Hence for t large enough... (c) the marginal distribution of q is FO. r. use t E^(t)/t = E[Yo .U(x) (c < oo because it is easily seen that U(x + 1) .y)P(Yo E dy) . Y1 > t] 4 0. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 . Yl > t].y) = f U(t .U(x) < U( 1)). (1 V)W) where V. In the general case. ^ > y) = 1 f +Y (z)dz. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y).t. U(x + 1) . Yo > 0] + f Eo^ (t . (d) the marginal distribution of ^ is FO.d.4.t. the sum is o(t) so that Eo£(t)/t + 0 .i. are not i. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.v. EC(t)/t + 0. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . W are independent. = z is Foz) The proof of (a) is straightforward by viewing {(r. In IV. Then fi(t)/t a4' 0 and. the first statement follows. assume first the renewal process is zerodelayed. but governed by a Markov chain {Jn} (we . Since z ( k) < E[Yi .
These facts allow many definitions and results to be reduced to ordinary renewal.. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. IT. the Markov renewal process if for any n. Further: Proposition A1.r.t.}. oo). Jn_1.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y.i . G_(x) = P(ST_ < x. r+ < oo). . J1 i . X2. .. distribution ofjXt}t>o itself where Pi refers to the case Jo = i. . the semiregenerative process is called nonlattice if {T. . and define r+=inf{n>0: Sn>0}. . .t..and regenerative processes.g.. be i. < yIJ) = Fij( y) on {Jn= i..+ < x. Jo. Alsmeyer [5] and Thorisson [372]. Jn +1=j} where J = a(JO. where the distribution of X. Let X1. e.. Jn = i is the same as the P. Then Xt 4 Xo. = io for some arbitrary but fixed reference state io E E. The semiregenerative process is then regenerative w. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. 0] or (0 . in [APQ]. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE.d. . .7 Consider a nonlattice semiregenerative process.jEE is a family of distributions on (0. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. A Markov renewal process {Tn} contains an imbedded renewal process. Sn = X1 + • • • + Xn the associated random walk.. the conditional distribution of {XT„+t}t>o given Yo.T_ < oo). namely {Twk } where {Wk } is the sequence of instants w where Jo.. T_=inf{n>0: Sn<0}.r..} is nonlattice (it is easily seen that this definition does not depend on i). Y1.. Yn.) and (Fij )i. with common distribution F. For example.. A stochastic process {Xt}t>o is called semiregenerative w... oo). . . G+(x) = P(S... Notes and references Renewal theory and regenerative processes are treated.
S.. U. A C (0. n=0 n=0 00 00 and the T+. m<j<n}. On {T_ > 2}. In (A. >0.G+ * G_: (b) G_ (A) = f °° F(A . we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A).1 .=EGn. 0). Proof Considering the restrictions of measures to (oc. Sr_ _1 is at its minimum .1 (a) F = G+ + G_ . More rigorously.S. we consider the last such time (to make w unique) so that {w=m.7) follows since G+(A) = 0 when A C (oo. (e) R_ = U+. (d) R+ = U_. oo) (A.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). oo).. (A. oo). define w as the time where the preT_ path S1.7) (A.. 0].x)R+(dx).r. A C (0. G+.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. n=0 The basic identities are the following: Theorem A2. F(A . F(A) + (G+ * G_)(A). 0<j<m.=n w=m i Figure A. 0] and (0. the renewal measures U+=>G+.8) (e. . u . .g. n 0 R_(A) = E I(Sn E A). 0]). A C (oo. S. G_.>0.7). F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. A C (oo.T_=n} = {S. (c) G+(A) = f °.x)R_ (dx).
0 < k < n. m it follows (see Fig. ST+Edu).. (b) follows from 00 G+ (A) _ E F(Sn E A.u) f0m m=1 n=m+1 00 J0 OO P(S.>0. .. Sr_ E Adu) (s ee again Fig .. r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. S. ST_ E A) P(T+ = m. m=1 f S mming over n = 2.F(r_n_mSrEA_u). SnEAIS. SmEdu) = P(T+=m.3._ = n ._ E A .7) follows...0<k<ri .+ E du)P(S. m < j <n. and the proof of (A.1) that P(Sj Sn.du) (G+ * G)(A)• C llecting terms. clearly (Sj Sm>0. ST_ E A . It follows that for n > 2 F (7.1. Aso.+ E du) E P(S.. 0<j<m.8) is similar.= n. (A.m. A. Sn1 E dx) n=1  F(A .x)P(Sk < 0. S._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu).3 8 APPENDIX Reversing the time points 0. and reversing the order of summation yields P(T_ > 2.Sn_1Edx. .. E du) = P(T_=nm.x)R+(dx). . A.1).XnEAx) 00 f 0 f 0 00 00 1: F(A ..
a number of related identities can be derived.1(a) is from Kennedy [228]. is based upon representing G+ as in (b). and similarly H_ (s) = 1 .Sn_k.O<k<n. Summing over n yields R+ (A) = U_ (A). In continuous time. if {St} is Brownian motion. Since G+ is concentrated on (0.6.9) whenever F[s].G_[s]) (A.T+> n) = P(Sk < O. In discrete time.SnEA) = P(Sn<Sk. u Notes and references In its above discrete time version. WienerHopf theory is only used at a few places in this book. and G+.s. 11. such developments motivate the approach in Chapter VI on the Markovian environment model.SnEA) = P(SnSn_ k. E. which is basic for the PollaczeckKhinchine formula.1). G_ are trivial. . consider a fixed n and let Xk = Xn_k+l. and sometimes in a larger strip. cf. there are direct analogues of Theorem A2..1.0<k<n.APPENDIX 339 and the proof of (c) is similar. The classical analytical form of the WienerHopf problem is to write 1 .G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. Another main extension of the theory deals with Markov dependence.g. we can rewrite (a) as 1 .g.SnEA) is the probability that n is a weak descending ladder point with Sn E A. However.f. the derivation of the form of G+ for the compound Poisson model (Theorem 11. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1).O<k<n. this holds always on the line its = 0. Nevertheless. then T+ = inf It > 0 : St = 0} is 0 a.g. The present proof of Theorem A2. In this generality of. see e. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). oo).'s. G_ [s] are defined at the same time..0<k<n.2 In terms of m. Then for A C (oo. and the proof of (e) is similar.F[s] = (1 . Again. u Remark A2. see for example Bingham [65]. P(SnEA . the survey [15] by the author and the extensive list of references there.0+[s])(1 . For (d). it serves as model and motivation for a number of results and arguments in continuous time.P as a product H+H_ of functions with such properties. the analogue of a random walk is a process with stationary independent increments (a Levy process. 6+ [s]. there is no direct analogue of Theorem A2.1. For example. Sk = X1 + • • • + Xk = Sn . being concentrated at 0.4).SnEA) = P(Sn<Sk. 0].
Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A.1 (SCALING AND SQUARING) The difficulty in directly applying t e series expansion eQ = Eo Q"/n! arises when the elements of Q are large. Thus. Eo Kn/n! converges rapidly and can be evaluated without p oblems. ere A is the eigenvalue of largest absolute value.12) eA'AO = Ale AA (A. if m is s fficiently large.5 that when handling phase type distributi ons. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step). _I 0 (A. It is seen from Theorem VIII.11) A f eAtdt = eA. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . however . hen the elements of Q"/n! do not decrease very rapidly to zero and may contribute a nonnegligible amount to eQ even when n is quite large and very any terms of the series may be needed (one may even experience floating point overflow when computing Qn). three of the c rrently most widely used ones: xample A3. To circumvent this. 0 .340 APPENDIX 3 Matrixexponentials T e exponential eA of a p x p matrix A is defined by the usual series expansion 00 An eA n=0 n! he series is always convergent because A' = O(nk Ialn) for some integer k < p. whereas there is no similar single established a proach in the case of matrix exponentials. and eQ can then be computed as the mth power (by squaring if = 2).10) d dteAt = AeAt = eAtA (A.13) henever A is a diagonal matrix with all diagonal elements nonzero. 1. Here are. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ).
15) Then it is easily checked that P is a transition matrix . Let vi. i.e. the intensity matrix Q is the same as the one Q for {Xt} since a jump from i to j 11 i occurs at rate qij = 77pij = q22. . Here is a further method which appears quite appealing at a first sight: Example A3 .2 (UNIFORMIZATION) Formally. p different eigenvalues Aj i . However . some jumps are dummy in the sense that no state transition occurs ).4 (DIAGONALIZATION) Assume that Q has diagonal form. In practice.e. i. . and we may consider a new Markov process {Xt} which has jumps governed by P and occuring at epochs of {Nt} only (note that since pii is typically nonzero .3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt.]t)n (A. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1.14) E n n=0 which is easily seen to be valid as a consequence of eqt = en(Pr)t = entenpt The idea which lies behind is uniformization of a Markov process {Xt}. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. Zo = a (Z = QZ.. we have k = QK (or KQ) which is a system of p2 linear differential equations which can be solved numerically by standard algorithms (say the RungeKutta method) subject to the boundary condition Ko = I. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. the procedure consists in choosing some suitable i > 0.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (. Zo = h). One then can reduce to p linear differential equations by noting that k = ZQ.. Ap. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. vp be the corresponding left .7t) n=0 n! u °O n Pn (to see this. To this end. condition upon the number n of Poisson events in [Olt])  Example A3.. The approach is in particular convenient if one wants eQt for many different u values of t.APPENDIX 341 Example A3..3 i (A. The probabilistic reason that (A..
hp..342 APPENDIX (row) eigenvectors and hl. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close. say Al. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. There are. hp the corresponding right (column) eigenvectors. Then P P Q = > Aihivi = E Aihi (9 vi. we have an explicit formula for eQt once the A j. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. D = ) 2 2 . Then vihj = 0.. v5Q = Aivi. Complex calculus : Typically.. and hence A2 is so because of A2 = tr(Q). i= 1 i=1 P P (A... In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. (A.16) (A.g H1.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. i # j. Qhi = vihi. (A. vi. the eigenvalue. not all ai are real. under the conditions of the PerronFrobenius theorem). and vihi ¢ 0. Example A3. we can take H as the matrix with columns hl. say A = (Ai)diag.17) eQt = E e\`thivi = E ea:thi ® vi. Nevertheless.. and we may adapt some normalization convention ensuring vihi = 1. of largest real part is often real (say. hi have been computed. however..18) contains terms which almost cancel and the loss of digits may be disasterous.18) Namely.. i=1 i=1 Thus. The phenomenon occurs not least when the dimension p is large. and writing eQt as eQt = He°tH1 = H (e\it)di. some cases remain where diagonalization may still be appealing.
20) ir = q2 ql qi +q 2 9l +q2 (A. replacing ai by A2.e. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. h2 = Thus. u Example A3. The other eigenvalue is A = A2 = q1 . where (A.7 Let 3 9 2 14 7 11 2 2 . v2 and h2 can be computed in just the same way. Then 7r = (ir1 7r2 ) = a (q21 Al .q. b are any constants ensuring//Irk = 1. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . 1) . l ab (g12g21 + (A1  411) 2) = 1. However. Of course.19) Example A3 .k1). k  C k2 ) =b ( A1 q 1 Q11 / where a .APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. i.
(A. and a generalized inverse may not unique.344 Then D= 2+ 11)' 7 T4 2 =52. e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. They are most often constructed by imposing some additional properties . 2 2 1=ab(142+(1+2)2 ) = tab. Generalized inverses play an important role in statistics. for example AA+A = A. A+AA+ = A+.11/2 . (A.6. APPENDIX x1 3/2 . but only that dimensions match . (AA+)' = AA+.11/2 + 5 1.5 .satisfying AAA = A. (A+A)' = A+A. ir =a(2 9 9 14 2 1 3 2 2)' k=b 14 =b 1+ 2 ir1 k1 ir2 k1 _ 9 2 10 5 7 9 70 1 ' 7r1 k2 7r2 k2 10 9 9 10 10 + 7 1 10 10 10 1 10 7 10 9 70 9 10 0 e4" = e_...22) Note that in this generality it is not assumed that A is necessarily square.23) . A2 = 3/2 .
1 Let A be an irreducible intensity matrix with stationary row vector it.25) . Rather than with generalized inverses .1Q = Q(Q .eir)1 = I .D + O(ebt).. (A.23) is called the MoorePenrose inverse of A. = 0 where m < p is the rank of A..ew.P).e. . Assume that a unique stationary distribution w exists .I) (A.. lt o eAx dx = te7r + D(eAt .24) = te7r . then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A ..g. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .P + e7r ). _ A.e ® 7r)1. and define D = (A . 0 01 In applied probability. if A is a possibly singular covariance matrix (nonnegative definite).eir ). and exists and is unique (see for example Rao [300]).eir )1. Here is a typical result on the role of such matrices in applied probability: Proposition A4. .. most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. one is also faced with singular matrices . These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . (I . Am+1 = . are ordered such that Al > 0.= (I .g. E. Am > 0. Then for some b > 0. ( Q .1 goes under the name fundamental matrix of the Markov chain)..P + e7r)1 (here ( I .APPENDIX 345 A matrix A+ satisfying (A. one then works with Q = (Q .
s.J {xe^r + D(e .e. it follows that h ® it is the k x m matrix with ijth element hi7rj .h. of (A. . .I)}.27) Proof Let A(t).. the rows are proportional to it.D + D2 + O(ebt). u 4b The Kronecker product ® and the Kronecker sum We recall that if A(1) is a k1 x ml and A(2) a k2 x m2 matrix.2 Let it be a row vector with m components and h a column vector with k components. and in fact any rank 1 matrix can be written on this form. I. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4. resp.346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . Then A(O) _ B(O) = 0. the formulas involving O(e6t) follow by PerronFrobenius theory. Note that h ® it has rank 1. (A.2e7r .26) 2 = 2 e7r + tD . o Finally.DZ(ent . h ® it reduces to hit in standard matrix notation.24). B'(t) = e7r + DAeAt = eir + (I . Interpreting 7r. (A. see below. then the Kronecker (tensor) product A(') ®A(2) is the (k1 x k2) x (ml x m2) matrix with (il i2) (jl j2)th entry a. For example. h as 1 x m and k x 1 matrices. the r. Equivalently.I) (A.eir)eAt = eAt = A'(t).91a(2) . respectively.h.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ).26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx . in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4.I) .I)} dx. B(t) denote the l.s. and the columns to h.
then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).3f 4v/.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.A9. each of which is A ® I or I ® B.3V8. then v1B1h1 and v2B2h2 are real numbers.4vf. C2 = h2 are column vectors.(A.3v'6. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors. (A. such a factor is Ak (&B 1k according to (A.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2).29) If A and B are both square (k1 = ml and k2 = m2).4 eA® B = eA ®eB.29).APPENDIX 347 Then A®B = 2 f 20. if A ® I occurs k times.50 6 7 6 4f 4.31) Indeed. and the number of such factors is precisely given by the relevant binomial coefficient. Using (A. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) . if Al = vi.31).3vV/72f 20. (A B)' = eA®B e! L 1=0 0 .28) In particular. (A.5v/.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I . A2 = v2 are row vectors and C1 = h1. Proof We shall use the binomial formula A crucial property is the fact that the functional equation for the exponential t / l (A ®B)t = I k Ak 0 B1k k=0 (A.
where transition matrix of the bivariate Markov chain {X n1). From what has been said about matrices of {Yt( 1).I)(h ® k). X ) }. A special case of Proposition A4. Q(2). Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }.5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). Yt(2 ) }. Thus . in the definition (A. independent Markov chains.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). the same time.4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . resp .33) . { On the other hand. and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . first term on the r .348 APPENDIX Remark A4. Let P8f P(Sl). (A. P(2). { 1't(1) }.s. h. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .32). {Yt(1).3 < 0 Lemma A4 . v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. P(t) Yt(2) }. p = P(1) ® {X }. we have P8 = Pal) ® p(2). and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. Let further it. represents ces Q( 1).6 Suppose that A and of B. k any column vectors. the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . P8 = Pal ) ® P82) exp {Q ( 1) ® Q(2)1 = eXp {Q( 1) } ® exp {Q(2) } Also the following formula is basic: B are both square such that a +.
. Similarly.. p there should exist io.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. = j and atk_li. j = 1. h = e and v = 7r (the stationary row vector).3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. . . (A. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. and the corresponding left and right eigenvectors v.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. . f o r each i. n.. h can be chosen with 3By this. and if we normalize v.. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u.12). we have AO = 1. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. in such that io = i. then IN < Ao for all A E sp(A). .34) Note that for a transition matrix. Here is the PerronFrobenius theorem. ao). which can be found in a great number of books. .g. . the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). il. Now note that the eigenvalues of A ® B are of the form a +. see e. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements.29). E (0. Then the eigenvalue Ao with largest real part is simple and real. h such that vh = 1..7 Let A be a p x pmatrix with nonnegative elements. > 0 for k = 1. (b) if in addition A is aperiodic. i. .. so that by asssumption A ® B is u invertible. [APQ] X. and the corresponding left and right eigenvectors v. That is. h can be chosen with strictly positive elements. . . and appeal to (A.APPENDIX 349 Proof According to (A.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements..
if we normalize v. but is an easy consequence of the PerronFrobenius theorem. relate the eigenvalues of B to those of B via (A.8 is most often not stated explicitly in textbooks. we have A0 = 0. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. Ao). Proposition A5.35) for some p E (oo. the condition is that t is small compared to Q.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. let {Yti°i } be a Markov process with initial distribution a and intensity .(3. not only in the tail but in the whole distribution. The next result gives a condition for asymptotical exponentiality. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. Furthermore.2). let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . the phasetype distribution B(a) with representation (. Example A3. Corollary A4. Then for any (3. h = e and v = 7r (the stationary row vector).. h such that vh = 1. Note that for an intensity matrix.8. the analogy of this procedure with unformization. note that we can write the phase generator T as Q . To this end. 10) and use the formula me at e Bt = e 00 Antn = e .n t AL n=0 n! (cf.1. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII.e. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). one can consider A = 77I + B where rl > 0 is so large that all diagonal elements of A are strictly positive (then A is irreducible and aperiodic). I.(ti)ding. For example. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength.350 APPENDIX strictly positive elements.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0).
Hence we can represent ( (a) as ((a) = inf { t > O : f tY( )dv=V } ^l = inf { t > O : t adv = V } l jat inf{t > 0: tydv =aV} = JJJ a J J where o (x) = inf {t >0: fo tY dv = x}. = YQ(x).g. it states that the state.1. J^O)_ = j) Pi (v(aaV) > x. and this easily yields a(x)/x a' 1/t*. Let further V be exponential with intensity V and independent of everything else.(a) > x . Then a(a'V)/a (aV) a' 1. We shall . we get dx F (Idx = j) = (1 + qij t )Sij + qij dt.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. Hence O ((a) aa.bij) Hence the intensity matrix of { Ix} is (qij/ti)i. prove a somewhat more general result which was used in the proof of Proposition VI. a' = a . fo tY dv/t a$' t*. v/ t. and that Yt(a) = Yat for all t.a' + oo (e. a .9. a'/a + 1. Conditioning upon whether { Yt} changes state in [0.2 Pi (c(a) > x. Since JJ(. and write Yt = Yt(1). t < (a). dx/ti] or not.APPENDIX 351 ((1) etc.Yj(av) = j f . from which it is easily checked that the limiting stationary distribution is (aiti/t*)iEE• Now let a' 4 oo with a in such a way that a' < a. Proof Assume first ti > 0 for all i and let I. from which the phase process is terminated . J(()) _ = i) + at•x t tt' .x (1 . By the law of large numbers for Markov processes . in fact . Then {Ix} is a Markov process with to = Yo.YQ(av) = j) Pi ( ci(a'V) > x. In addition to the asymptotic exponentiality. has a limit distribution: Proposition A5.jEE. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ . We can assume that Jta) = Yt(°). {t Y( a) } v>0 .aE where 0 < e < 1).
j) and initial distribution a. so we shall be brief.. ' pk 0 k>1 11 Theorem A5. let E and Pkj j=k1. zkbk is za(I . Penev & Turbin [238]. (c) the nth moment k 1 k"bkis 1)"n!aP"p. Et II I a(a^V) > x) at' . P. . = 0 for one or more i.g. these results are in the spirit of rare events theory for regenerative processes (e. 1 k=1 1 0 otherwise. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]). a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. Then P is substochastic and the vector of exit probabilities is p = e .+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. Example A5.} is said to be discrete phasetype with representation (E. See also Korolyuk. k = 1. Then: (a) The point probabilities are bk = aPklp. . a = b = (bk)k=1.4 Any discrete distribution B with finite support. Example A5..5 Let B be discrete phasetype with representation (P. the simplest discrete phasetype distribution: here E has only one element. an easy modification of the argument yields finally the result for the case where t. (b) the generating function b[z] _ E' . However.2 do not appear to be in the literature.. K}. 2. . a).352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) ..x k > K. with point probabilities bk = (1 . A distribution B on {1... say bk = 0.zP)'p.. 2.p)k1 p. > 0}... Keilson [223].. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p.1 and A5. u Notes and references Propositions A5.3 As the exponential distribution is the simplest continuous phasetype distribution.Pe. so is the geometric distribution.. Indeed. k>1. is discrete phasetype.
36) in blockpartitioned notation (where we could also write a as (a (1) 0)). resp.1 This corresponds to a convolution of r geometric distributions with the same parameter p. and a=1). { Jt 2) } with lifetimes U1 .a(2). resp. Jt t > U1 + U2. a' . U2.. and hence the negative binomial distribution is discrete phaseu type.6 (CONVOLUTIONS) Let B1. _ i E E(1) T(1) t(1)a(2) i E E(2) .T(2)). (E(2). a. B2 be phasetype with representations (E(1). A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.{ 0.APPENDIX 353 5c Closure properties Example A5.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. Then {Jt} has lifetime U1 + U2 . A. initial distribution a and phase generator T.. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2)..a(1).6. Then the convolution B = B1 * B2 is phasetype with representation (E. 11 Example A5.2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U.6 is the Erlang distribution Er which is the convolution of r exponential distributions.. The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r.r + 1. r . . T= ( 0 T(2) ) (A. as is seen by minor modifications of Example A5.T(1)).
Thus. this means that a = (Oa(1) (1 . In risk theory. B2 be phasetype with representations (E(1).0)a(2) E(2) Figure A.E) where a(°) = fAa(a)v(da).8 (FINITE MIXTURES) Let B1. U2.O)B2 (0 < 0 < 1) is phasetype with representation (E. (E(2). a. Example A5. a reduced phase diagram is f a E t Figure A.3 In exactly the same way.354 APPENDIX Example A5..37) (1) (1 . Then it is trivial to see that B(") is u phasetype with representation (a(").T(2)).. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims. Let B(") be the corresponding phasetype distribution. and o'i Oa. are i. we need to restart the phase process for B w. i E E(1) T 0 I (A. a. A reduced phase diagram is 0a(1) E(1) A . T) and C = EO°_1(1 . resp. i E E(2) 0 T(2) =IT (in blockpartitioned notation.p)pn1.T.T(1)).a(1). P(N = n) = (1 . then C is the distribution of Ul + • • • + UN. p at each termination.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2)..i.0)ai2). a mixture of more than two phasetype distributions is seen to be phasetype.4 .0)a(2))). if U1. Example A5.9 (INFINITE MIXTURES WITH T FIXED) Assume that a = a(°) depends on a parameter a E A whereas E and T are the same for all a.d.a(2). To obtain a phase process for C. and consider B(") = fA B(a) v(da) where v is a probability measure on A.p)pn1B*n. Then the mixture B = 9B1 + (1 . with common distribution and N is independent of the Uk and geometrically distributed with parameter p.p. Equivalently.').
9) that (U .a(1).2. Indeed. (E(2). 13 (MINIMA AND MAXIMA ) Let U1. cf. T) and C = F.7. U2. Example A5 .X)+ is zeromodified phasetype with representation (E..x)+.T) if U is phasetype with representation (E. B2 of phasetype with representations (E('). a.aeTx. v.f.. then C is the distribution of U1 + • • • + UN. To obtain a phase representation for C . with common distribution B and N is independent of the Uk with P(N = n) = f.TWWW). 12 (PHASETYPE COMPOUNDS ) Let fl. f2. E). are i.°.v. say v. Thus the representation is (E(1) x E(2). It is zeromodified phasetype with representation (E. a(1) ® a(2 ). but the same T. then U1 +• is phasetype with representation (E. If U1 has a different initial vector.. let B be a continuous phasetype distribution with representation (F. let {Jtl)}.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . Example A5. a(2). Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. of F. if {Jt} is a phase process for U. then Jy has distribution aeTx.APPENDIX 355 and C is phasetype with representation (E. Corollary VIII.T + pta). T + ta. say with distribution F. Example A5 .1. +UN 2. Proposition VIII. T).. T(1) ® T(2)). X independent of U.g. Note that this was exactly the structure of the lifetime of a terminating renewal u process. it follows by mixing (Example A5. resp. { Jt2) } be independent with lifetimes U1. .°_1 f„ B*?l. T + pta). Minor modifications of the argument show that 1. if B is defective and N + 1 is the first n with U„ = oo.d. U2. i. P). we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }..2. if U1. Equivalently... a. v. be the point probabilities of a discrete phasetype distribution with representation (E. a. T(2) ). If we replace x by a r.aF[T]. cf. resp.. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). . U2 be random variables with distributions B1.T) where F[T] = J0 "o eTx F(dx) u is the matrix m. To see this. j E F}. For U1 A U2. let the phase space be E x F = {i j : i E E. then U1 + • • + UN is zeromodified phasetype with representation (a. { 4 } as exit of {Jt}.
Example A5. and the phase generator is T(1) ®T(2) T(1) ®t(2) t(1) ® T(2) 0 T(1) 0 0 0 T(2) Notes and references The results of the present section are standard . By the diagonal argument (subsequent thinnings). however. we need to allow { Jt.n = I:pi(n)Er v ( __ ) n) ) a= 1 .356 APPENDIX For U1 V U2. and let Bn be the Erlang distribution E. Here are the details at two somewhat different levels of abstraction: (diagonal argument . Let the support of Dn be {xl(n). Then from above. say degenerate at b.B(bk) I < 1/n for n > k. relies more on matrix algebra than the probabilistic interpretation exploited here). we can assume that ID. Then we must find phasetype distributions Bn with B. Hence it is immediate that Bn 4 B. Proof Assume first that B is a onepoint distribution.. i= 1 C.(bk) + B(bk) for all k.. The general case now follows easily from this. That is.8.14 To a given distribution B on (0. and vice versa.xq(n)(n)}.2) } to go on (on E(2)) when { i 1) } exits. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5.. elementary) Let {bk} be any dense sequence of continuity points for B(x). r # oo.. see Neuts [269] (where the proof.(Sn) with Sn = n/b. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness .(bk) + B(bk) for all k as n * oo. Now we can find first a sequence {Dm} of distributions with finite support such that D.} of phasetype distributions such that Bn 3 B as n + oo.(n) = D. cf. oo)... any distribution B on (0. the initial vector is (a(1) (& a (2) 0 0). Thus the state space is E(1 ) x E(2) U E(1) U E( 2). with weight pi(n) for xi(n). and the closedness of the class of phasetype distributions under the formation of finite mixtures. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. there is a sequence {B. q(n) q(n) pi(n)a .(bk)'... the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support.
. Corollary A5.n( b k ) . u Theorem A5. oo) and any fl. It should be noted..(x)Bf.D(bk)I < n. oo) such that f (x) = O(e«x).n (bk) . But To is the class G of all distributions on [0. one would use the B given by some statistical fitting procedure (see below). f2.B(bk )I < . in at least two ways: insensitivity Suppose we are able to verify a specific result when B is of phasetype say that two functionals Cpl (B) and W2 (B) coincide. If Cpl (B) and ^02(B) are weakly continuous. and we can take Bn = Cr(n). k < n. i = 1. Let E be the class of functions f : [0. Since PET is closed under the continuous operation of formation of finite mixtures. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f..t..APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n).( dx) * f r f{(x)B(dx). 2. replications). x 4 oo. k < n. oo) * [0.r.. the class CO of all discrete distributions. oo) approximation Assume that we can compute a functional W(B) when B is phasetype. Then ICr( n ). compute W(B) and use this quantity as an approximation to cp(B0)..n. however.d.i. E E. then it is immediate that WI(B) = p2(B) for all distributions B on [0.. Hence G C PET and L = PIT. say on the claim size distribution B in risk theory. i.. oo). we can then approximate Bo by a phasetype B. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. PIT contains all finite mixtures of onepoint distributions. for some a < oo. . In particular.14 is fundamental and can motivate phasetype assumptions. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. For a general Bo. and that cp is known to be continuous.15 To a given distribution B on (0 . if information on Bo is given in terms of observations (i.e. u 2 (abstract topological ) The essence of the argument above is that the closure (w.
n(dx) < 1+. liminf B. .. 2. f° xtBn(dx ) * f °° x`B( dx). Now returning to the proof of (A. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge..f ' f (x)B(dx). TO (A. then cc f (x)Bn ( dx) = (?!c ) e'= . if f (x ) = e°x.2 .. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A...39) Indeed.. and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ). for each i. n. i = 1.14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx).  APPENDIX B implies that 00 o o 00 n. By (A. we may assume that in the proof of Theorem A5. f00 fi(x)Cr..16 To a given distribution B on (0 . n.. i = 1.f (x)B(dx).358 Proof By Fatou' s lemma.38) We first show that for each f E E. n B=az. i=1. . and the case of a general f then follows from the definition of the class E and a uniform integrability argument.38 ). and hence we may choose r(n) such that L 9l) f (x)Cr(n).. ...oo J fi(x)B.(dx) > J fi(x)B(dx).39).n(dx) + f 0 fi(x)Dn(dx). oo). \\ 0 Corollary A5.f (z) = f = 1 1 1 1n/ o . .. i = 1.. .
. However.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions. e ) and ei J. from a more conceptual . the adjustment coefficient 'y = 7(B. . one can obtain 7(Bn. The present section is a survey of some of the available approaches and software for inplementing this. lim inf > is proved similarly. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5./3) is defined as the unique solution > 0 of B[y] = l+y/j3. . . . there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n .16. O We state without proof the following result: Corollary A5. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal.. (N.> y for some sequence {ei} with ei E (0..18 In the setting of Corollary A5. (N or a given distribution Bo. The adjustment coefficient is a fundamental quantity. lim sup ryn < 7. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. I. /3) = ry for all n.17 To a given /3 > 0 and a given distribution B on (0.e. oo) with B[y +e] < oo for some e > y = 7(B. the remaining results may be slightly stronger than those given in the literature. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . Notes and references Theorem A5.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn. .l3µb < 1. For practical purposes./3). Proof Let fi(x) = el'r+E. 5e Phasetype fitting As has been mentioned a number of times already. If ei > 0. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1..14 is classical. there is a sequence {B. but are certainly not unexpected. .3). 0 as i * oo.
[317] ) has considered an extension of this setup. cf.g. Schmickler (the MEDA package. and used a nonlinear programming approach .. three for a mixture of two Erlangs ). [216] ). defined by the absence of loops in the phase diagram . The characteristics of all of these methods is that even the number of parameters may be low (e. [70]) restrict attention to acyclic phase type distributions . g.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa .d. the L1 distance between the c . [202].'s). . for some suitable large n. where more than two Erlangs are allowed and in addition to the exact matching of the first three moments a more general deviation measure is minimized (e.. Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . .g. The constraints were the exact fit of the two first moments and the objective function to be minimized involved the deviation of the empirical and fitted c. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). e . and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. risk theory. (N is the empirical distribution Be. at a a number of selected points . Of course. the number of phases required for a good fit will typically be much larger.f. B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly.} of phasetype distribution such that Bo. giving mass 1 /N to each S=. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. It seems therefore a key issue to develop methods allowing for a more general phase diagram. we do not not want to perform matrix calculus in hundreds or thousands dimensions). A method developed by Bobbio and coworkers (see e..f. and as fitted distribution we may take B. and in practice this sets a limitation to the usefulness (the curse of dimensionality . Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . reliability or queueing theory. a program package written in C for the SUN workstation or the PC is available as shareware. d. we have constructed a sequence { B.g. In a series of papers (e. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood.g ..
x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations. . one is lead to an iterative scheme. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations.. Thus. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. In fact. e.... N Ti = I(J= i) dt.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j). (n+1) _ Ea (n). EN where ai = N 1 I ((k) = i) tii=i iEE. E.g. . .g..T(n) (Ti ^^ 1.(N) = E Ea(n)..T(n) (Nik IC1.T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei .T (n)(TiI(1. . it is easy to see that N (k Ea(n).. then the estimators would be of simple occurenceexposure type.. Nii = = . eieT(n)((k..APPENDIX 361 tion Sk was available. = j) f k=1 k =1 tE[0. jEEA. since this is parameterdependent. the methods of [70] and [38] appear to produce almost identical results. it seems open whether the restriction to the acyclic case is a severe loss of generality. (N) tJk Ea ( n). In practice. .
This page is intentionally left blank .
Abate. Chichester New York.Bibliography [1] J. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. New York. Preprint. [7] E. pp. Alsmeyer (1991) Erneuerungstheorie. [13] S. 345368. The case of equal risk sums. with applications to risk reserve processes and the GI /G/1 queue. 57. Dover. 20. Queueing Systems 5 . Teubner. 213229. Stuttgart. Aktuar Tidskr. Borokov. J. In: Limit Theorems and Related Problems (A. The Mathematical Scientist 14. 1985. Appl. Optimizations Software. Adv. [6] V. 363 . 219229. 3157 . New York. [10] K.). B. [2] J. Choudhury & W. II. [15] S. 53100. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions .A. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Asmussen (1987) Applied Probability and Queues. Proc. [3] J. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . 101116. John Wiley & Sons. Transactions XVth International Congress of Actuaries. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Abate & W. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. 143170. G. 587. Skand. Asmussen (1984) Approximations for the probability of ruin within finite time . Probab.G. Scand. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions .L. Abate & W. 311338. Arfwedson ( 1955) Research in collective risk theory. [4] M. Queueing Systems 16. Asmussen (1985) Conjugate processes and the simulation of ruin problems. Appl. ed. [12] S. Aktuar Tidskr. AT&T. 191223. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . ibid.). 14. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. [14] S. Arfwedson ( 1954) Research in collective risk theory. [11) S. 1984 . Stoch. [9] G. [5] G. Abramowitz & I. Act. 253267. Skand. Queueing Systems 10. 37. [8] G. 38. New York.
[28] S. 297318.S.K. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. [30] S. L. Asmussen & M. Asmussen & M. Bernoulli 6. Ann. [31] S. Probab.). 137168. Floe Henriksen & C. Boca Raton. Stoch. [18] S. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull.364 BIBLIOGRAPHY [16] S. 2149. CRC Press. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. [21] S. stationary distributions and first passage probabilities. Methods V Problems (J. [23] S. J. 37. 354374. [25] S. Kalashnikov & A. 20. Asmussen (1999) On the ruin problem for some adapted premium rules. Asmussen (2000) Matrixanalytic models and their analysis. 54. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Asmussen & B. Asmussen. [22] S. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. 96107. Ann. [17] S. Chakravarty. Stochastic Models 11. 772789. Proc. 4966. Riga Aviation University. Proc. 1989 . Extremes 1 . Andronov.M. 69100. Appl. Asmussen (1992b) Light traffic equivalence in single server queues. K. [20] S.). SIAM Review 40. Statist. Probabilistic Analysis of Rare Events (V. New York. [34] S. Astin Bulletin 27. Appl. Asmussen (1995) Stationary distributions via first passage times. Act. Th. Act. J. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. 79102. Binswanger & B. Appl. Scand. [29] S. [19] S. 2943. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. 193226. 555574. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. Frey. Act. Florida. A. eds. Scand. T. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Random Processes 2. 189201.). Asmussen & K. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Scand. [26] S. [27] S. 2. 1999 . 313326. 106119. Dshalalow ed. Asmussen. Probab. 8. . Asmussen (1998c) A probabilistic look at the WienerHopf equation. MatrixAnalytic Methods in Stochastic Models (A. 313341. Hojgaard (2000) Rare events simulation for heavytailed distributions. Marcel Dekker. 1996. Ann. [24] S. eds. [32] S. [33] S. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. Alfa & S. 27. Advances in Queueing: Models. 315. Rolski & V. J. 303322. Asmussen & B. Stoch. Appl. 1936. Asmussen. 25. Scand. Probab. J. Asmussen (1989b) Risk theory in a Markovian environment. Asmussen (1992a) Phasetype representations in random walk and queueing problems.
Oper. Asmussen & S. Stochastic Models 8. Asmussen & D. Proc. Asmussen & C. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. [43] S.Y. [47] S. Appl. Th. Schmidli & V. 259274. (to appear).BIBLIOGRAPHY 365 [35] S. Appl. Boca Raton. Asmussen & K.A. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Wiley. [49] S. 913916. Asmussen & R. Opns. 429466. J. Appl. B. Rubinstein (1999) Sensitivity analysis of insurance risk models. 365372.). Prob 32 . Supplementary Volume (Kotz. 299324. Schmidt (1995) Ladder height distributions with marks. with applications to insurance risk Stoch. 736755. 103125. Asmussen. 10. 64. Probab. Advances in Queueing: Models. Finance and Stochastics 4. 120. [42] S. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Hojgaard & M. Perry (1992) On cycle maxima. 30. [46] S. 419441. Adv. Asmussen & C. Sigman (1996) Monotone stochastic recursions and their duals. [37] S. Inf. [36] S. Probab. Methods 8 Problems (J. Asmussen. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. Asmussen & R. Asmussen & M. [51] S. Taksar (2000) Optimal risk control and dividend distribution policies . H. Res. Asmussen & C. Asmussen & H. 421458. [39] S. [48] S. Probab. Florida. [40] S. J. Probab. Scand. CRC Press. Koole (1993). Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Stoch. Johnson. Sc. Proc. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. first passage problems and extreme value theory for queues. Read eds. Appl. Dshalalow ed. [52] S. [38] S. J. Insurance: Mathematics and Economics 20. 58. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. [41] S. 31. Asmussen & V. Statistica Neerlandica 47. . Insurance: Mathematics and Economics 10. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks.M. 105119. O.Y. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Asmussen & T. Eng. [44] S. 422447. Asmussen . Olsson (1996). 11251141. Math. 115. Asmussen & T. [45] S. Asmussen & V. 19.). Adv. [50] S. 20. Nerman & M. Fitting phasetype distributions via the EM algorithm. Asmussen & G. Res. Kliippelberg (1996) Large deviations results for subexponential tails. Marked point processes as limits of Markovian arrival streams. Statist. 23. Appl. Taksar (1997) Controlled diffusion models for optimal dividend payout.A. Example of excessoffloss reinsurance for an insurance corporation. Management Science 45. 410433. Appl.
Berman & R.A. 610. Probab. Boogaert & A. 11811190.L. BarndorffNielsen & H. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. New York. J. [71] P. J. [58] O. Act. [62] J. J. Trans. Insurance: Mathematics and Economics 9. Cambridge University Press. 1985 .H. Adv. [63] A. [69] P. Bloomfield & D. Soc. [68] T. Halsted Press.T. Cox (1972) A low traffic approximation for queues. 148156. Beekman (1974) Two Stochastic Processes. Cambridge. [65] N. [54] K. 9599.L. Bobbio & M. C. Actuaries 21. Asmussen & J. J.R. Appl. [64] P. New York. Act. 7. [66] N. J. [57] C. Appl. Beekman (1985) A series for infinite time ruin probabilities. Bingham (1975) Fluctuation theory in continuous time. Probab. Act. Appl. Scand. [70] A. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. 275279. [67] T. 4148. Athreya & P. 832840. J. 661667. [72] P. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. de Waegenaere (1990) Simulation of ruin probabilities. 1995. 705766. Oxford. Bjork & J. 1975. Bingham. [61] J. New York.H. Goldie & J. Berlin. Scand.A. 33. 190204. SIAM. Bjork & J. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Wiley. 129134. Beekman (1969) A ruin function approximation. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. J. [59] O.366 BIBLIOGRAPHY [53] S. Teugels (1987) Regular Variation.B .J. SpringerVerlag. 221232. Scand. Act.M. Grandell (1985) An insensitivity property of the ruin probability. 1988 .H. Probab. Borovkov (1976) Asymptotic Methods in Queueing Theory. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Insurance: Mathematics and Economics 6. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. Scand. Act. [55] B. 9. . [60] J. Springer . 1974. Ney ( 1972) Branching Processes . Wiley. Boogaert & V. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . [56] B. Stochastic Models 10. Billingsley (1968) Convergence of Probability Measures. 77111. [73] A. Baker (1977) The Numerical Solution of Integral Equations. 169186. Clarendon Press. Insurance: Mathematics and Economics 4. Scand.
Gerber.Y. 907920. [76] N. S. B.J. Tech. Computer Performance (K. H.I. J. [84] D.W. Chung (1974) A Course in Probability Theory (2nd ed. Qinlar (1972) Markov additive processes. D. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. II. [86] K. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions .C. . 34. Springer.M. [77] P. Stockholm. The Jubilee volume of Forsakringsbolaget Skandia. Simulation and Estimation. verve. SpringerVerlag.W. Z.).A.. [78] L. Berlin. 14. Veraverbeke (1990). Breiman ( 1968) Probability. IEEE Trans.) NorthHolland.J. Academic Press.A. Philos. Bratley. [82] H. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Cohen (1982) The Single Server Queue (2nd ed. Aut. J. 127130. 313319. Syst. New York.L. Illinois. Croux & N. AddisonWesley. [87] E. Control AC28. Fox & L. Amsterdam. IEEE J. [93] K. 14331453. Broeck . 93121. Schrage (1987) A Guide to Simulation. Nesbitt (1986) Actuarial Mathematics . Bowers . 3539. The Society of Actuaries . [89] M.). [85] W. Nonparametric estimators for the probability of ruin. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. J. Bux & U. Hickman. 392433. Oper. Reading. [79] P. [92] H.R. Boxma & J. Wiley.U. Cottrell.R.Y . Tweedie (1982) Storage processes with general release rule and additive inputs . Probab. Cramer (1955) Collective risk theory. [88] J. NorthHolland.R. Appl. Geb. Sel. Insurance : Mathematics and Economics 5. [90] D. 16. Resnick & R. 177204. Res. [80] F. Chandy & M. Reiser eds. Boxma & J. Brockwell . Herzog (1977) The phase concept : approximations of measured data and performance analysis. Bell. 749763. Wahrscheinlichkeitsth. 51. Goovaerts & F. Skandia Jubilee Volume. 105119. Jones & C. Queueing Systems 33. Stockholm. Amsterdam. Insurance : Mathematics and Economics 9. [91] H. New York San Francisco London. Biihlmann ( 1970) Mathematical Methods in Risk Theory. Area Commun. De Vylder (1986) Ordering of risks and ruin probabilities. Burman & D .W. [81] J.L. Jr.BIBLIOGRAPHY 367 [74] O. New York. M. Fort & G. 24. Cambr. Bucklew ( 1990) Large Deviation Techniques in Decision.L. Burman & D. Adv. [75] O. Itasca.L. Proc. [83] D. 2338. Cramer (1930) On the Mathematical Theory of Risk.C. Soc. Cox (1955) Use of complex probabilities in the theory of stochastic processes. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . 62.
Daley & T. [105] A. 3. Delbaen & J. [95] M. 16. Statist. Boston. Dassios & P. Csorgo & J. 121131. [102] P. [100] A. 17. Chapman & Hall. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII.P. Daykin. Actuarial J. New York. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . Suppl. 583602. Appl. Scand. Insurance: Mathematics and Economics 4. Insurance: Mathematics and Economics 6. De Vylder & M. Oper.368 BIBLIOGRAPHY [94] M. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . De Vylder & M. Haezendonck (1987) Classical risk theory in an economic environment. 27. Devroye (1986) Nonuniform Random Variate Generation. Rolski (1991) Light traffic approximations in queues.D. Jones and Bartlett. [109] F. [106] A. Csorgo & J. [97] D. Res. Pentikainen & E.. Martingales and insurance risk. De Vylder (1977) A new proof of a known result in risk theory. Insurance: Mathematics and Economics 10. J. Rolski (1984) A light traffic approximation for a single server queue. 135159. Aktuar. Appl. Probab. [101] C. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Springer.B. Insurance: Mathematics and Economics 7. Laird & D. Roy. [107] L. 624628. [108] F. Act. Comp. Oper. 3750. 1990 . Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Pesonen (1994) Practical Risk Theory for Actuaries. J. Res. De Vylder (1996) Advanced Risk Theory. Editions de 1'Universite de Bruxelles.. English version published in Skand. Math. Insurance: Mathematics and Economics 3. Math. [103] F. Embrechts (1989). 114119. Haezendonck (1985) Inversed martingales in risk theory. 85116. 227229. Soc. Delbaen & J. J. [110] F. .L. 7083 (1969). 201206. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Math. Deheuvels & J. J. Stochastic Models 5. [99] R. Dembo & O. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. N. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. [111] F. T. Tidskr. 88101. Daley & T. [112] F.M. Stockholm. Goovaerts (1984) Bounds for classical ruin probabilities. 435436. [98] A. Scand. [96] D. 181217. 22. Dempster. Zeitouni (1992) Large Deviations Techniques and Applications. [104] F. 5771. 9 .
Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. 363374. 1741. Hipp (1998) Ruin probabilities for Erlang (2) risk processes. 193199.C.U. Camb. [128] E. [130] P. [127] F. [116] D. Dickson & C.J.M. 4259. Dickson & H. [119] D.R. [114] D. Dickson & C. 1975 . J. Res.C. Regterschot (1988) Conditional PASTA. Dickson & H. [121] D.BIBLIOGRAPHY 369 [113] D. 131138.M. Dufresne & H. 1994 . Dynkin (1965) Markov Processes I. and the amount of claim causing ruin. [122] B.J.U. Act. Duffield & N. 3745. [118] D. Act. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier.B. Dickson (1994) An upper bound for the probability of ultimate ruin. 1.U. [124] N. 269274. J.M.S. Embrechts .W.G. 229232. Letters 7. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. J.M. Hipp (1999) Ruin problems for phasetype(2) risk processes. J. 1984 . Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. Dufresne & H. 1993 . Insurance: Mathematics and Economics 19. 147167. with applications. Insurance : Mathematics and Economics 10. Berlin Gottingen Heidelberg. Harrison & A. J. Scand. Dickson (1995) A review of Panjer's recursion formula and its applications. [125] F. Grandell & H.M. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. [131] P. Proc.C.R.M.C. Astin Bulletin 21. Waters (1996 ) Reinsurance and ruin . Soc. . Emanuel .C. 6180. J. Scand. Insurance: Mathematics and Economics 7. Springer . Insurance : Mathematics and Economics 7. Insurance : Mathematics and Economics 11. Scand. [115] D. Dufresne. Philos. Gerber (1988) The surpluses immediately before and at ruin.C. 107124.R.C. Scand. Ruin estimates for large claims . 105115. Shiu (1991) Risk theory with the Gamma process. Embrechts ( 1988 ). Act. 4962. [123] E. Act. Insurance: Mathematics and Economics 25. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case .C. Dickson & J. Scand. 251262. Scand. [117] D. 191207. van Doorn & G. 1993 . 2000 . H. Scand.M. J. Insurance : Mathematics and Economics 22.R. [129] D . Dickson (1992) On the distribution of the surplus prior to ruin . J.C. Waters (1999) Ruin probabilities wih compounding.C. J.M.M. [126] F. Act. 1984 . Dickson & J. Math. Gerber & E.K.M. Act. Djehiche (1993) A large deviation estimate for ruin probabilities. Act. British Actuarial J. [120] D. Oper. 118. 177192. 174186. 5159.
566580. Appl. Probab. Griibel & S. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . 38. Schmidt (1992) An insensitivity property of ladder height distributions. New York. Probab. 35. 404422. [146] E. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics .L. Appl. 5974. [139] A.J. 623637. [144] P. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Frees (1986) Nonparametric estimation of the probability of ruin. Appl. [140] J. Embrechts. 181190.D. Embrechts & J. Fitzsimmons (1987). Embrechts. Scand.A. Franken. Heidelberg. Appl. Kliippelberg & T. [147] M. 5975. Proschan & D. 159178.D. 8190. U. Adv. Fuh & T. Akt. Schmidli (1994) Ruin estimation for a general insurance risk model. J. Frenz & V. . Esary. [145] P. Embrechts & N.K. Lai (1998) Wald's equations. [138] P. [141] F. [143] W. Embrechts. Embrechts & H. Insurance: Mathematics and Economics 7. Walkup (1967) Association of random variables. Erlang 2. Fields 75. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Nyt Tidsskrift for Matematik B20. Villasenor (1988) Ruin estimates for large claims. 269274. C. 131137. [134] P. [150] H. Embrechts & T. Probab. New York. Math. Springer.W. 1998 . Wiley. Appl. 17.). [136] P. Teugels (1985). [137] P. Arndt & V. Trans. Insurance : Mathematics and Economics 1. R. Rel. Approximations for compound Poisson and Polya processes. with applications.D. 29. Danish Academy Tech. Schmidt (1982) Queues and Point Processes. Maejima & J. Konig. J. Wiley. [148] C. Skand. D. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. Tidsskr. F. Astin Bulletin 16. Adv. Probab.). Adv. 5572. [135] P. 695712. [142] W. Wiley. Probab. Ann. Statistica Neerlandica 47. 175195.L.W. 14661474. 616624.M. Insurance: Mathematics and Economics 10. On the excursions of Markov processes in classical duality. [133] P. Furrer (1998) Risk processes perturbed by astable Levy motion. first passage times and moments of ladder variables in Markov random walks. Th. 3339. Sci. Probab. J. Act. A bootstrap procedure for estimating the adjustment coefficient.K.370 BIBLIOGRAPHY [132] P. Mikosch (1991). M. Jensen. Esscher (1932) On the probability function in the collective theory of risk. Statist. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. 29. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. [149] C. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. 26. J.
28.J. Probab. Weron (1997) Stable Levy motion approximation in collective risk theory.J. Ver. Kluwer. Dordrecht. Probab. [168] J.71. Haezendonck (1984) Insurance Premiums. Scand. 31A. van Heerwarden & T. [156] H. Math. Glasserman & S. Appl. 1523. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. Glynn & W. J.E. 131156. Vers. 424445. S. [161] P.N. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. Gerber (1986). 6370. [166] M. Goldie & R. 2336.).U. Mitt.U. Goovaerts. 5. Huebner Foundation Monographs. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'.W. 1977. Galambos & J. 97114. Gnedenko & I. Insurance: Mathematics and Economics 15. Michna & A. Grandell (1977) A class of approximations of ruin probabilities. J. eds. Goovaerts. 1978 . ETH Zurich. R. 7778. Suppl. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Math. Gerber (1988) Mathematical fun with ruin theory.U. 1981 . Birkhauser . A.. Ver.U. Boston. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. 73. NorthHolland. In Studies in Applied Probability (J. [157] H. Scand. Mitt. Appl. Furrer & H. Act. Amsterdam. contained in the authors PhD thesis. Springer. [159] H. Life Insurance Mathematics. Schweiz. Furrer. [153] H. Skand. . Ann. Amsterdam. [158] H. Unpublished. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Gani.U. Kou (1995) Limits of first passage times to rare sets in regenerative processes. Appl. Probab. 105115. [154] H. [167] C. 205210. J. [163] P. [160] H. F. Z. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. University of Pennsylvania. Gerber (1973) Martingales in risk theory.M. Insurance: Mathematics and Economics 20. Act.. Scand. [162] P.V. [152] H. [164] B.).BIBLIOGRAPHY 371 [151] H. Griibel (1996) Perpetuities with thin tails. J.S. NorthHolland.U. 3752. Adv. Bauwelinckx (1990) Effective Actuarial Methods. 205216. [165] M. Schweiz.U. Vers. Aktuarietidskrift 1970 . de Vylder & J. Gerber (1979) An Introduction to Mathematical Risk Theory. [155] H. [169] J. Basel. Kaas. London. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. 463480. Insurance : Mathematics and Economics 7.G. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. Act.
8. Probab.). Proc . [173] J. Grnbel (1991) G/G/1 via FFT. Lemoine (1977) Limit theorems for periodic queues. [185] J. 14. Harrison (1977) Ruin problems with compounding assets . [176] B.372 BIBLIOGRAPHY [170] J. Grandell & C. [177] B. [187] J. Informatica 2. mathematische Wirtschaft . M. Tinbergen Institute Research Series 20. KTH. Math.. 7590. 54. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . 332. Appl. [172] J. J. Resnick (1977) The recurrence classification of risk and storage processes . Segerdahl (1971) A comparison of some approximations of ruin probabilities. Grandell (1999) Simple approximations of ruin functions. Statistical Algorithm 265. Proc. Grandell (1990) Aspects of Risk Theory. 1. Grandell (1979) Empirical bounds for ruin probabilities.S. Matem. Opns . 3 . 5766. [188] J. Preprint.V. Amsterdam. 167176. [179] R. [183] M. Aktuar Tidskr. [189] A. [181] D. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . [174] J. Winter School on Stochastic Analysis and Appl . Probability Theory and Mathematical Statistics (I.131135. Grandell (1997) Mixed Poisson Processes. 400409. Appl.I. Grandell ( 1992) Finite time ruin probabilities and martingales . SpringerVerlag. 566576.I. Rink. Harrison & S. Res. Appl. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. Stoch. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . Astin Bulletin 29. Theory and Applications. 347358. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. M.O. 3041. Gusak & V. Berlin. [178] B.243255. Gut (1988 ) Stopped Random Walks . Ibragimov. [175] J. Chapman & Hall. Liet. [171] J. [182] A. E. 355365. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. Probab. Res. AkademieVerlag. Griibel & R.and Sozialforschung 6. SpringerVerlag. . 5 .J. M. 6779. Proc.A. Harrison & S. [180] R. M. Th. [184] H . 14. Harrison & A. Grigelionis (1996) Lundbergtype stochastic processes. 33. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. New York. Math. [186] J. Archiv v. Stoch. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . 143158. Gyllenberg & D. Opns . Skand. Applied Statistics 40. Insurance: Mathematics and Economics 26. 197214. ed. Appl.
12981310. 259268. Hill (1975) A simple general approach to inference about the tail of a distribution.A. Hoglund (1974) Central limit theorems and statistical inference for Markov chains.B. Verlag Versicherungswirtschaft e. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. [204] T. Karlsruhe. Studies in Statistical Quality Control and Reliability 1992 :4. [198] C. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. Huskova. 19. Statist. 29. [194] 0. 8996. J. [192] U. Insurance: Mathematics and Economics 5. Iversen & L.L. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. [200] M. ACM TOMACS 6. [195] B. Chalmers University of Technology and the University of Goteborg. Haggstrom. 3. J. Hipp (1989a) Efficient estimators for ruin probabilities. [207] D.M. verve. Probab. Foruth Prague Symp. Scand. eds. Horvath & E. 1990 . [191] W. Klugman (1984) Loss Distributions. 378389.a program for fitting phasetype distributions. [197] C. 123151.V. 23. Iglehart (1969) Diffusion approximations in collective risk theory. [208] V. Act. Hoglund (1991). Prob. 18. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. Scand. Hogg & S.). 11631174. 6.V. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden.BIBLIOGRAPHY 373 [190] P. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. 285293. [202] 0. Appl. [193] U. Act. 285292. 377381. Astin Bulletin 19. 4385.L. Math. Ann. Nachrichten 30. [203] T. 25.. 5770. Proc. 166180. Mandl & M. Geb. The ruin problem for finite Markov chains. Versicherungswirtschaft. 8095. Wiley. Hipp & R. 305313. Electronic Letters 35.R. . [206] B. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. J. 1998 . Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. Probab. Appl. Ann. [205] T. Mathematical Statistics. on Asymptotic Statistics (P. Nerman (1992) EMPHT . [201] L. S. Probab. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. New York. J. Karlsruhe. Insurance: Mathematics and Economics 5. Heilmann (1987) Grundbegriffe der Risikotheorie. Z. Ann. [199] R. Asmussen & 0. Hojgaard & M. No. Wahrscheinlichkeitsth. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. [196] C. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time.
Soc. 11. Kalashnikov (1996) Twosided bounds of ruin probabilities. Kluwer. 239256. Taaffe (1989/90) Matching moments to phase distributions.M. Keilson & D.J. J.H.C. Jelenkovic & A. Astin Bull. 561563. 4151. Taylor (1981) A Second Course in Stochastic Processes. [214] A. [224] J. 283306. Soc. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Janssen & J. J. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. Amer.M. 1996. 187193. [219] V. Astin Bull.B. Jensen (1995) Saddle Point Approximations. Proc.M.J.M. New York. 6. Karlin & H. 325247. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. Academic Press. Keilson & D. Johnson & M. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. 866870. Probab. ibid. [220] V. [212] J.L. Keilson & D. [228] J. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. Kemperman (1961) The Passage Problem for a Markov Chain.G. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. [222] S. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. University of Chicago Press. Scand. [216] M.374 BIBLIOGRAPHY [209] D. 63. Cambridge Philos. [221] E. Addenda to for processes defined on a finite Markov chain. Math. Jensen (1953) A Distribution Model Applicable to Economics. 711743. NEC Research Institute. 31. [218] V. Jagerman (1985) Certain Volterra integral equations arising in queueing. 116129. Oxford. [217] R. [213] P. N. Clarendon Press. Stochastic Models 5. 123133. Technometrics 30. 547567. 165167. Kao (1988) Computing the phasetype renewal and related functions. . Appl. Statist. [227] J. [226] J. Goovaerts (1986) General bound on ruin probabilities. Kaas & M. Copenhagen. [211] J. Wishart (1964). 37. J. Insurance: Mathematics and Economics 5. Chicago. Cambridge Philos.A. Soc. ibid.P. 35. Act.G. 3. Cambridge Philos. Manuscript. 8793. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. [225] J. Princeton.R. [210] D. 259281. Proc. Munksgaard. 118. 60. Ann. 61. 15. Probab. J.G.M. 173190. Appl. Stochastic Models 1. [215] J. Proc. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. 6. Act. N. [223] J.
Insurance : Mathematics and Economics 12. Math. J. 132141. (235] C.J.F. New York. Kuchler & M. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. Lehtonen & H. 1995 . Oxford 12. 283284. [236] C. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input.BIBLIOGRAPHY 375 [229] J. SIAM. Kingman (1962) On queues in heavy traffic. Kingman (1964) A martingale inequality in the theory of queues. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities.. Bernoulli 1 . Act. Lemoine (1981) On queues with periodic Poisson input. 279285. 18. Lecture Notes in Mathematics 511. 26.F. Klnppelberg & T. [246] D. Act. Proc. Proc. [241] G. Cambr. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. [245] T. Appl. Sorensen (1997) Exponential Families of Stochastic Processes. 889900. [243] A. Philos. SpringerVerlag. [231] J. 4958.F. Adv. [242] A. Kingman (1961) A convexity property of positive matrices . Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Mikosch (1995) Modelling delay in claim settlement. 4477. Lipsky (1992) Queueing Theory . I. 154168. Probab. Soc. Lehtonen & H. J. SpringerVerlag. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. Kunita (1976) Absolute continuity of Markov processes. [230] J. Insurance: Mathematics and Economics 8. Scand. Probab. Klnppelberg & T. 133135 (in Russian). 24. Statist. Scand. Seminaire de Probabilties X . [244] T. 383392. J. Prob. 277289. 390397. Cybernetica 4. [248] D.P. [240] U. J.J. Macmillan. Roy. Quart. [247] L. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. 48. 125147.C. [234] C. .C. [238] V. 6075. Lindley (1952) The theory of queues with a single server. Kliippelberg & U. Kliippelberg (1988) Subexponential distributions and integrated tails. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. Stochastic Models 7. [239] H. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Actuarial J. Penev & A. B24.S. J.a Linear Algebraic Approach. 359361. J. 60 . Soc. Quart. J.C.F. [233] C. Scand. 146. [232] C. Latouche & V. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . 858874. Appl. [237] C. Soc. Probab. 1998 . Korolyuk. 25. Appl. Camb. 259264. Appl. Philos.
1996 . Nummelin (1987) Markov additive processes I. [257] A. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. 286298. Nagaev (1957) Some limit theorems for stationary Markov chains. [266] P. MeierHellstern & M . [252] A. Ney & E. Ann.M. Cambridge Philos. [251] F. Probab. [255] V. [260] H. Soc. 147149. Meier ( 1984). 36. Soc. Almqvist & Wiksell. Stockholm. with an application . Stoch. Appl. . Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. [263] M . 561592. Opns. MartinL6f (1983) Entropy estimates for ruin probabilities . A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . 32. 834841. 1261 (0?)1270. Lucantoni .376 BIBLIOGRAPHY [249] D. . Ann. Probab . Europ.K. Miyazawa & V. 11 .S. Probab. Th. Uppsala. 124147. Proc.).D.S. [253] V. 2939. 161174. Res. Hoist eds. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. 676705. Appl. Appl. Michna ( 1998) Selfsimilar processes in collective risk theory. Schmidt ( 1993) On ladder height distributions of general risk processes . Math. Act. 15. 429448. Gut & J. Englunds Boktryckeri AB. F. Appl. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes .D. [254] V . J. [264] G. J. J.V. 31. MartinL3f (1986) Entropy.D. K. Cambridge Philos. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. 22. Scand. [250] F. J. 4859. Adv.F. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. [265] S. Probability and Mathematical Statistics (A. Makowski ( 1994) On an elementary characterization of the increasing convex order .K. J. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . 58 . 370377. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. Scand. Eigenvalue properties and limit theorems . Proc. 763776. [256] A. Anal. 2. [258] K . Probab. Probab. Ann. a useful concept in risk theory. 29 . II Aterforsdkring av Kollektivrisker . 378406. [262] H .V. Probab. [259] Z . Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. Act. 223235. Insurance : Mathematics and Economics 5. Scand. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . 1994. J. 268285. [261] H. Statist. Act. 58. Appl. Appl. Math. 3. J. 1986 .
10081026.F. New York. Johns Hopkins University Press. Stochastic Models 3 . [280] A. Naval Research Logistics Quarterly 25. Omey & E.K. Appl. Stoch. [284] J. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. 445454. O'Cinneide (1990) Characterization of phasetype distributions. 555564. 157. 764779. Appl. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. London. J. Stoch. 71. Nyrhinen (1999) Large deviations for the time of ruin. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. Pakes (1975) On the tail of waiting time distributions. Hamburg. Proc. 36. [285] J. Commun. 733746. Abh. [272] J . [268] M . Proc. Appl. E75B . 249266. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Models 6. Appl. 316. [276] C. IEICE Trans. Proc. 327361.F.a survey. Probab. Probab. [279] E. 12551265. 46. Adv. [278] E. 75. [269] M . J. 273299. Appl. Baltimore . 123144. Neuts (1977) A versatile Markovian point process. Sem. J. Paulsen (1993) Risk theory in a stochastic economic environment. [281] J. Appl. 215223. 21. [277] J.F. Paulsen & H. 135148. Proc. Norberg ( 1990) Risk theory and its statistics environment . [282] J . Probab. 30. Paulsen & H.F. Paulsen (1998) Ruin theory with compounding assets . 12. Neuts ( 1978) Renewal processes of phase type. Stoch.BIBLIOGRAPHY 377 [267] M . [273] R. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. 339353. 311342. Stoch. Probab.A. Stoch.F. 16. [274] H . Math. Appl. Omey & E. [271] M. Neuts (1992) Models based on the Markovian arrival process. Insurance: Mathematics and Economics 22. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. Marcel Dekker. Insurance: Mathematics and Economics 20. [270] M . [283] J. Statistics 21.K. Astin Bulletin 5 . Appl. . Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. [275] H.G. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations.
1989 . Insurance: Mathematics and Economics 16. Philipson (1968) A review of the collective theory of risk. 4568. Astin Bulletin XIV. Ann. R. [297] R. [303] S. Wiley. 32. 2343. [306] T. Appl. Probab. 691707. Prabhu (1980) Stochastic Storage Processes. [293] N. Scand.M. Adv. Wiley. J. Appl. [294] N.L. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Springer. 537555. J. 19. Embrechts (1996) Confidence bounds for the adjustment coefficient. 46. On the ruin problem of collective risk theory.A. Math. 12. Oper. [302] J. Act. Appl.U. Austr. Griibel & P. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Pitts. Queues. .H. 29A. Prabhu & Zhu (1989) Markovmodulated queueing systems. [301] G. 757764. 11. 465483. [296] N. [304] B. Probab. 215246. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. Berlin. Insurance: Mathematics and Economics 3. 117133. 61. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Res. Heidelberg. Probab . 30. 2330.G. 29. [290] E.M. Statist.J. 235243. [300] C. Statist.U. Opns. 28.U. [292] S. Gjessing (1997c) Ruin theory with stochastic return on investments. Res. 4. Prabhu (1961). Resnick & G. Math. Rao (1965 ) Linear Statistical Inference and Its Applications. [295] N.U. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. 337347. Queueing Systems 5. 222261. Math. Wiley. [298] V. 965985. Probab. Aktuar. Math. J. 390413. Ripley (1987) Stochastic Simulation . New York. [291] S. [299] C. Tidskr. and Dams. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. [305] C.. Ann.K. [289] C. 820827. Regterschot & J. Appl. 139143. 568576. Ann. Ann. [287] F. Prabhu (1965) Queues and Inventories. Paulsen & H. Pyke (1959) The supremum and infimum of the Poisson process. Appl. New York. Adv. 147159. Insurance Risk. Probab. Skand.J.R. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. Pellerey (1995) On the preservation of some orderings of risks under convolution.M. Stat. [288] S. Adv.G. Math. Inst.K.378 BIBLIOGRAPHY [286] J. Rolski (1987) Approximation of periodic queues. Soc. Pitman (1980) Subexponential distribution functions.
[325] H . [311] R. Rubinstein & B. J. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). 4068. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. Probab. Adv. Wiley. Rubinstein & A. J. Schmidli ( 1996) Martingales and insurance risk. 21. [322] H. Wiley. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . Math. Operationsforsch. Proc. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . Rossberg & GSiegel (1974) Die Bedeutung von Kingmans Integralgleichungen bei der Approximation der stationiiren Wartezeitverteilung im Modell GI/C/1 mit and ohne Verzogerung beim Beginn einer Beschiiftigungsperiode. Stochastic Models 6. Appi. Seal ( 1969) The Stochastic Theory of a Risk Business. 121133. Act. Appl. [312] R. [323] H. J. Astin Bulletin 29. Stochastic Models 10 . 687699. Teugels (1999) Stochastic Processes for Insurance and Finance. [324] H . 431447 [316] S. Comp. Insurance: Mathematics and Economics 16. [309] H. Ann.L. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . [310] R. 5. Stoch. Insurance: Mathematics and Economics 22. Scand. Probab. Singapore. Statist. Wiley. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory.Y. 145165. 131156. [321] H. 4857. 262286. [319] H. Stochastic Models 10. [326] H. [314] T. 5. [318] H. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Statist.M. V. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. Act. Melamed (1998) Modern Simulation and Modelling. Data Anal. [317] L. 417421. 2001 . Appl. Schmidli ( 1999b) Perturbed risk processes: a review . Probab . Science Culture Technology Publishing .Y. Scand. [313] M. 93104.J. 1997. [315] T. 7. 11. 5. 135149. 795829. Th. H. [320] H. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting .BIBLIOGRAPHY 379 [307] T. 227244. Wiley. 155188. Rolski. Schmidli. . [308] S. Wiley. Rudemo (1973) Point processes generated by transitions of a Markov chain.Y. 365388. Rubinstein (1981) Simulation and the Monte Carlo Method. Schlegel ( 1998) Ruin probabilities in perturbed risk models. Schmidt & J.
Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Skand. 4383. [342] D. Appl. Chapman and Hall. Almqvist & Wiksell. Verein Schweiz. Chapman & Hall. 279299. 171178. 191197. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. Scand. [329] H. Seal (1974) The numerical calculation of U(w. [343] D. [336] B. 383413. In Probability and statistics . Shin (1989) Ruin probability by operational calculus. [345] D. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. Probab . 159180. 1955. Appl. Th. Verein Schweiz. Probab. [344] D. Sengupta (1990) The semiMarkov queue: theory and applications.A. [335] B . Math. [332] C : O. 11. Probab.S. [347] K. Seal (1978) Survival Probabilities. Statist. [337] M. Shaked & J. Aktuar Tidsskr.W. 483487. Insurance: Mathematics and Economics 8. 21. New York. Grenander). Seal ( 1972). Adv. 72. 673684. .S.L.F. 121139. Risk teory and the single server queue . Sigman (1992) Light traffic for workload in queues.L. Versich. 75.380 BIBLIOGRAPHY [327] H. [330] H. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. 157166. [334] C: O. [341] E. Shin (1987) Convolution of uniform distributions and ruin probability. Probab. Act. t). 429442. 4. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. 11. L. Ann. Stochastic Models 6. [328] H. Stockholm. Math. Seber (1984) Multivariate Observations. Shwartz & A. Segerdahl (1959) A survey of results in the collective theory of risk. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Verein Schweiz Versich. Siegmund (1975) The time until ruin in collective risk theory. 77100. Academic Press. [348] K . 4. Aktuar Tidsskr. [338] E. Weiss (1995) Large Deviations for Performance Analysis. Adv. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. Mitt.L. 2236. Scand. Math. pp. [339] A. J. the probability of nonruin in an interval (0. [346) D . [340] E. 72. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. 243249.S. 914924. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen .W.the Harald Cramer volume (ed. Versich. 61. 1987. SpringerVerlag. Wiley. Mitt. Ann. Siegmund ( 1985) Sequential Analysis . Queueing Systems 11.G. Shantikumar (1993) Stochastic Orders and Their Applications. Wiley. J. 701719. Mitteil. . 1974. Appl. t). [333] C: O. [331] G. Act.
[361] G. Appl. Act. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. Stanford & K. [369] O. 65102. [350] W. 21. Act. [366] O. [358] R. Actuarial J. 197208. Proceeedings of the IEEE 77. Insurance: Mathematics and Economics 1. 1974.. 114137. [356] B . [351] D . 1980 . [363] G. 1982 . Astin Bulletin 24. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed.J.).C. Taylor (1979) Probability of ruin with variable premium rate. 1976. [353] E.C. Wiley. 1978. Thorin (1982) Probabilities of ruin. [359] L. [362] G. New York. NorthHolland. Daley ed. J. Scand.C. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . Insurance: Mathematics and Economics 19. [357] B.A. Scand. New York. Scand. Scand. Act. Proc. Astin Bulletin 16 . Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes.L. 235254. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. 118. Scand. John Wiley & Sons. 722. Smith ( 1953) Distribution of queueing times. [355] B .L. Cambridge Philos. Teugels ( 1995 ) Ruin estimates under interest force . [360] G. Teugels (1982) Estimation of ruin probabilities . Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. Probab. 163175. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . 5776. Slud & C. [364] G. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . Karlsruhe. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue.C. J. Soc 49. Astin Bulletin 12.J. 2739. Sundt & J. Ruin probabilities prepared for numerical calculations.S. 65102. Sundt & J. J. [368] O. [352] D.and largedeviations probabilities in actuarial risk theory. 725741. J. J. Scand. Straub (1988) NonLife Insurance Mathematics. 449461.BIBLIOGRAPHY 381 [349] E. [354] B. Springer. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. Act. Sundt & W.V.).. . 8594. 149162. Unpublished manuscript. 8199. Taylor (1986) Claims reserving in NonLife Insurance.L. [365] J. Act. Insurance: Mathematics and Economics 16.L. Hoesman ( 1989) Moderate.C. Adv. Verlag Versicherungswirtschaft e. Thorin (1977). [367] O.
Reidel. De Vylder. Willmot & X. [377] V. 148164. Submitted.R. Astin Bulletin IX. D. 5762. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. In: Premium Calculation in Insurance (F. Willinger. 4963. [371] 0. J. Act. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Lin (1994) Lundberg bounds on the tails of compound distributions. 142. M. Tidskr. Stationarity and Regeneration. [375] N. Skand. Waters (1983) Probability of ruin for a risk process with claims cost inflation. Statistical Science 10. 37. Appl. J. Wolff (1990) Stochastic Modeling and the Theory of Queues. Res. J. SpringerVerlag. [374] F. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. [372] H. Wiley. F.382 BIBLIOGRAPHY [370] 0. Tacklind (1942) Sur le risque dans les jeux inequitables. [380] W.E. 743756. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue.E. Unpublished Ph. 936952. R. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Taqqu. 31. M. [381] W. Thorin & N. thesis. Probab. Thorisson (2000) Coupling. [382] G. [376] A. Goovaerts. 1983 . Haezendonck eds.). Sherman & D. 6785. Van Wouve. Thorin & N. 1942. Dordrecht Boston Lancaster. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. [373] S. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Insurance : Mathematics and Economics 13. Scand. 137153. Insurance: Mathematics and Economics 15. . Astin Bulletin VII. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. [378] H. [384] R. Wald (1947) Sequential Analysis. 231246. PrenticeHall. University of Michigan. Oper. [379] M. De Vylder & M. [383] G.W. Aktuar.
318320 change of measure 2630. 97.314316. 2526. 318319 Erlang distribution 7.242.251280 heavy traffic 76.4447. 226. 79. 5. 1112.6779. 37.301 central limit theorem 60 . 341.359 aggregate claims 103106.217. 17. 141144.and sum 221. 91.135. 301 Kronecker product.9899. 271274. 278 gamma distribution 67. 89.346349 383 . 3436. 1415.100.272. 239. 138139.292293 Edgeworth expansion 113. 86.328330. 180182. 119. 360 excursion 155156.185187. 7879.Index adjustment coefficient 17. 97129. 9293.86. 302303 diffusion approximation 17. 4851. 248 WienerHopf 144 interest rate 190. 196201 inverse Gaussian distribution 76. 245248. 111117. 308.178184. 12 CramerLundberg approximation 1617. 9496.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.121129. 5796.285292. 332333 Volterra 192194. 2425.299. 316323 Bessel function 102.281.203. 361 diffusion 3. 205.203.160167. 162164. 40. 117127 corrected 121127 duality 1314. 3334. 39.293294.249. 110113. 8283 hyperexponential distribution 7. 122.228229. 17. 7475.287292. 14.249250 integral equation 16 Lindley 143 renewal 64.137141. 135.269. 9396. 323 Coxian distribution 147. 80 81.182.259261. 189. 1819. 3032. 117128. 227229. 201 Brownian motion 3 .226. 170173. 7079.307312 compound Poisson model 4. 218 Cox process 4. 283.200201.150. 3839. 201214. 15. 207 heavytailed distribution 6. 217. 7179.
44. 179 NP approximation 318320 Palm distribution 5253.297299. 203204. 175 light traffic 8183 Lindley integral equation 143 process 3334. 2730.350361 Poisson process Markovmodulated 12 periodic 12.340350 multiplicative functional 2830. 44. 3947. 35. 162.128129. 138. 37.259261.161.287291 INDEX matrix equation . 25. 144.238. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 157. 7179. 108 life insurance 5. 142 likelihood ratio : see change of measure lognormal distribution 9.161. 251.201. 3947. 3639. 9899. 257.304 process 2830.134135. 176185. 132133.152160. 5758.148. 306316 Levy process 3.240244.146148. 42. 106108. 271274.123. 295. 171. 245 M/G/1 13. 52 53. 15. 149.174. 145187. 96.348 terminating 215216. 134. 16. 59.285287 queue 14 . 133. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167.234240. 141144. 39. 71. 7576.160161. 108109. 178 modulation 12. 16. 260 Lundberg conjugation 6979 . 227228. 4446. 133.315 inequality 1718. 100.234.298299. 269 PerronFrobenius theory 4142.139141. 112113. 38. 14. 86 periodicity 12.302. 113114.215250.336339 Laplace transform 15.178182.218221. 213214.180. 137139. 6970.287. 304305 random walk 3336.161164.349 350 perturbation 172173. 134135. 261264.275278. 41.384 ladder heights 4756. 230. 234 matrixexponential distribution 240244 matrixexponentials 14.339 large deviations 129. 203 Markov additive process 12. nonlinear 155. 99. 267269 Panjer's recursion 320323 Pareto distribution 910. 154. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 25.227230. 65.288290.261264. 32. 38. see also sensitivity analysis phasetype distribution 8. 35. 80.108.269271. 6162.336339 .
89. 174. 87. 280. 261264 reservedependent premiums 14. 186187 virtual: see workload rational Laplace transform 8. 4950. 141144. 37. 177 timereversion 14. 257. 213. 327 .INDEX 385 waiting time 141. 168172 storage process 13. 281296 stable process 15.336339 workload 13. see also matrixexponential distribution regenerative process 264 268. 160. 256258.186. 294296 shotnoise process 314 simulation 19. 251. 317318 semiMarkov 147.314. 338 utility 324. 233. 279280 subexponential distribution 11. 1819. 260 reinsurance 8. 251. 8386. 7475. 189214. 186187 renewal process 131.154157. 120 statistics x. 335336 sensitivity analysis 8693. 332333 model 12. 123. 244. 147. 233234. 107. 131144. 172173. 331336 equation 64. 12. 191192. 152.359361 stochastic control x stochastic ordering 18. 238 saddlepoint method 115117. 333334 regular variation 10. 292294. 3032. 229234. 307308. 326330 Weibull distribution 9. 253. 162.279280 Rouche roots 158.244250. 31. 251280 time change 4. 11. 223226.262263. 60.273274. 9693. 240. 5455. 146. 222. 260 WienerHopf theory 144.
for heavytailed claim size distributions). Special features of the book are the emphasis on change of measure techniques.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. y finite horizon ruin probabilities. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. Some i (l I JL I J r of the topics are Lundberg's inequality." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. the ^W A l \ i l ' ''' CramerLundberg approximation.g. 2 A I 11 JjVb l' i  i Yj .. It is a comprehensive treatment of the known results on ruin probabilities.. extensions of the classical compound Poisson model to allow f o r reservedependent premiums.Vol.com 2779 he 9 "789810ll22293211 .Advanced Series on Statistical Science & Applied Probability . I 1! Ruin Probabilities . worldscientific. exact solutions. P'i yfliother approximations (e. "This book is a must for anybody working in applied probability. Markovmodulation or periodicity.