This action might not be possible to undo. Are you sure you want to continue?
https://www.scribd.com/doc/41948462/RuinProbAsmussen
11/06/2011
text
original
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
. . . . . . . . .. . . . . . . . . . . . . . . . . . 261 4 Models with dependent input . . . . . . . . . . . . . .. . . . . . .. . . . . . . . 297 2 Further applications of martingales . . .. 281 2 Simulation via the PollaczeckKhinchine formula . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . 292 6 Sensitivity analysis . . . . . . . . . . . . . . . .. . . . . . . . . .. . . . . . 287 4 Importance sampling for the finite horizon case . . 306 4 The distribution of the aggregate claims . . . . . . . . . . . . 323 6 Reinsurance . . 259 3 The renewal model . 271 6 Reservedependent premiums . . . . . . . . . .. 340 A4 Some linear algebra . . . . . . . 344 AS Complements on phasetype distributions . . . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . 290 5 Regenerative simulation . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . 304 3 Large deviations . . . . . . . . . . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . . . . . 279 X Simulation methodology 281 1 Generalities . . . 350 Bibliography Index 363 383 . . . . . . . . . . . . . . . 316 5 Principles for premium calculation . . . . . . . . . . .. . . . . . . . . . . . . 251 2 The compound Poisson model . . . . .. . . . The twobarrier ruin problem . . . . . 336 A3 Matrixexponentials . 326 Appendix 331 Al Renewal theory . . . .. . .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . 331 A2 WienerHopf factorization . . . . . . . . . . .. . 285 3 Importance sampling via Lundberg conjugation . . . . . . . . . . . .. . . . . ... . . .
This page is intentionally left blank .
which has in particular removed one of the standard criticisms of the area. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. and the series editor Ole BarndorffNielsen for their patience. 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. 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!). the idea was close to expand these to a short book on the subject. In particular. this applies to longrange dependence which is intensely studied in the neighboring ix . I have deliberately stayed away from discussing the practical relevance of the theory. The course was never realized. and other projects absorbed my interest. the book is basically mathematical in its flavour. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. and has been an active area of research from the days of Lundberg all the way up to today.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. Thus. However. and my belief was that this could be done rather quickly. that it can only say something about very simple models and questions. It has obviously not been possible to cover all subareas. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. and the result is now that the book is much more related to my own research than the initial outline. if the formulations occasionally give a different impression. University of Copenhagen. But the pace was much slower than expected. it is not by intention. As an excuse: many of these projects were related to the book. Apart from these remarks. Since I was to produce some handouts for the students anyway.
it has not been possible to incorporate more numerical examples than the few there are. the standard stochastic control setting of diffusion models has been considered.13. 111.g.45. see in particular Michna [259]. the first part of 11. Concerning ruin probabilities.4a. I intend to keep a list of misprints and remarks posted on my web page. In the classical setting of CramerLundberg models. For a second reading. IV. The present book is in between these two possibilities. for which I apologize to the reader and the authors of the many papers who ought to have been on the list. In addition.5. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. It is obvious that such a system involves a number of inconsistencies and omissions. 111.3. X. VI. see also Schmidli [325] and the references in Asmussen & Taksar [52]. The main motivation comes from statistical data for network traffic (e.89. Here is a suggestion on how to get started with the book. More recently. IX.lth.15. e. Hojgaard & Taksar [206].6 (to understand the PollaczeckKhinchine formula in 111. another by method. http:// www. for the effects on tail probabilities. Hojgaard & Taksar [35] and Paulsen & Gjessing [284].2.x PREFACE field of queueing theory. VII. Resnick & Samorodnitsky [303] and references therein.13 and IX. an area which is becoming increasingly important. IV. [381]). Finally.1. some papers not cited in the text but judged to be of interest are included in the Bibliography.2 more properly).g.maths . IV.2. 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). incorporate 11.14. VIII. For a brief orientation. Willinger et al.g. The rest is up to your specific interests.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths .lth. A book like this can be organized in many ways. Asmussen.se Lund February 2000 Soren Asmussen .13 and XI. I regret that due to time constraints. Another interesting area which is not covered is dynamic control. One is by model. see e. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. VII. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157].13. Good luck! I have tried to be fairly exhaustive in citing references close to the text. read Chapter I.
111 . Section VII . 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . 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 . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . were produced by Lone Juul Hansen .1 by Bjarne Hojgaard and the table in Example 111.6 is reprinted from Asmussen & Schmidt [49] and parts of IX. Fig. Aarhus.6.PREFACE xi The second printing differs from the first only by minor corrections. Fig.8 .2 by Rafal Kulik . .3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. More substantial remarks. not least the more complicated ones. as well as some additional references continue to be at the web page. supported by Center for Mathematical Physics and Stochastics (MaPhySto). IV.4 from Asmussen. Parts of X.5 from Asmussen [21] with permission from CRC Press.1 and X. many of which were pointed out by Hanspeter Schmidli . Parts of II.6 by my 1999 simulation class in Lund. Section VIII. 5. A number of other figures were supplied by Christian Geisler Asmussen .
This page is intentionally left blank .
t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t.4) O<t<oo O<t<T 1 . They are the main topics of study of the present book.Rt. respectively. (1.3) sup St. MT = sup St.1) We also refer to t/) ( u) and 0(u. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. as defined in broad terms . (1. is a model for the time evolution of the reserves of an insurance company. results and topics to be studied in the rest of the book. T) as ruin probabilities with infinite horizon and finite horizon . it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . M = (1.Chapter I Introduction 1 The risk process In this chapter . A risk reserve process { Rt}t>o.T) = P inf Rt < 0 I . we introduce some general notation and terminology. and give a very brief summary of some of the models. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. (1. We denote throughout the initial reserve by u = Ro.2) (O<t<T Ro=ul. For mathematical purposes.i(u.
• Premiums flow in at rate p. the time of arrival of the nth claim is an = T1 + • • • + Tn.6) Sofar we have not imposed any assumptions on the risk reserve process.. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. (1. (1. we see that Nt Nt Rt = u + pt . and T1 is the time of the first claim.b(u) = P (r(u) < oo) = P(M > u). Figure 1. That is.i(u.5) i. However. (1. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals.E Uk.pt.1 .2 CHAPTER I. t] is finite. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. per unit time. the ruin probabilities can then alternatively be written as . We denote the interarrival times of claims by T2. St = E Uk .7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. Thus. respectively.1.T) = F (MT > u) = P(r(u) < T). 1. the number Nt of arrivals in [0. Putting things together.. say. T3. .
s.(. one may well argue that Brownian motion in itself could be a reasonable model.20%.) V 0.1.1 Assume that (1. immaterial.e. of course.. since any modeling involves some approximative assumptions. t * oo. one could well replace Rt by Rtnr(u) or RtA. though many results are straightforward to generalize from the compound Poisson model. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. We shall not deal with this case either.1. We shall discuss Brownian motion somewhat in Chapter IV.1.. say 10% . For the purpose of studying ruin probabilities this distinction is. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. and the basic ruin probabilities are derived in XI. we shall. We study this case in Ch. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i.8) holds. on Fig. It would appear obvious. then M = oo a. not discuss whether this actually corresponds to practice.1 the slope of {Rt} should depend also on the level). for example. and hence .8) The interpretation of p is as the average amount of claim per unit time.b(u) = 1 for all u. a basic references is Gerber [127]. but as an approximation to the risk process rather than as a model of intrinsic merit. However. 1. Thus. If 77 < 0. • Brownian motion or more general diffusions. then M < oo a. 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. . VII. If 77 > 0. however. 1. and hence O(u) < 1 for all sufficiently large u.s. 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. and in fact: Proposition 1. (1. allowing a countable infinity of jumps on Fig. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. however. that the insurance company should try to ensure 77 > 0.
M < oo a. . rl > 0. Proposition 1. are i. zP(u . However.8) is given by ^t p = EU • lim it (3(s) ds t. (1. we obtain typically a somewhat stronger conclusion. U2.s. and independent of {(0(t). 0 We shall only encounter a few instances of a Cox process. not all models considered in the literature have this feature: Example 1. The simplest example is 3(t) = V where V is a r .4 CHAPTER I. then similarly limSt/t < 0. If 77 < 0. this needs to be verified in each separate case.T) for {Rt} is given by V)(u) = t/i (u).i. _ St __ k =1 Uk pt a4.3 Assume p 54 1 and define Rt = Rt1p.Q (say) and U1.d. corresponding to the Pdlya process. t t p  p' t ^ oo.Tp).oo t 0 J (provided the limit exists).8).. then this limit is > 0 which implies St a$ oo and hence M = oo a.s. 0(u.b(u) < 1 for all u when rl > 0.. However. .d. If u oo. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u).T) = i.10) hold with p constant.v. INTRODUCTION Proof It follows from (1.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. namely.10) is a property which we will typically encounter. and here (1. (1. Thus p may well be random for such processes. if {(3(t)} is nonergodic.10) Again.i. where {Nt} is a Poisson process with rate . St In concrete models.. (1. tb(u) = 1 for all u holds also when rl = 0. namely that M = oo a. This case is referred to as the mixed Poisson process.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t.8) that F N. U2. Nt)}. it is not too difficult to show that p as defined by (1. k=1 (1.6EU (on the average. and that .s.. are i. and independent of {Nt}. . Here it is easy to see that p = . with the most notable special case being V having a Gamma distribution.i(u.. If U1.11) . 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).
Segerdahl [334] and Philipson [289].. Note that life insurance (e. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. Insurance: Mathematics and Economics. Buhlmann [82]. Sundt [354]. Grandell [171].2. Gerber [159]) has a rather different flavour. Daykin et al. Notes and references The study of ruin probabilities. Some early surveys are given in Cramer [91]. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. In the even more general area of nonlife insurance mathematics. De Vylder [110]. Hipp & Michel [198].. Note that when p = 1. in a number of models. 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. Cox processes are treated extensively in Grandell [171]. U2. see e . and we do not get near to the topic anywhere in this book.g. Rolski. We roughly classify these into two groups . but in probability and applied probability as a whole. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. [134]. the recent survey by Grandell [173] and references therein. The Swedish school was pioneering not only in risk theory. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. Besides in standard journals in probability and applied probability. see also Chapter XI. Some of the main general ideas were laid down by Lundberg [250]. Some main later textbooks are (in alphabetical order) Buhlmann [82].. [330]. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. the claim arrivals are Poisson or renewal at the same time). Schmidt & Teugels [307] and Seal [326]. was largely initiated in Sweden in the first half of the century. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution).g. Heilmann [191]. Straub [353]. [76]. Pentikainen & Pesonen [101]. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. we shall be able to identify p with the traffic intensity of an associated queue. For mixed Poisson processes and Polya processes. Daykin. Taylor [364]. another important early Swedish work is Tacklind [373]. often referred to as collective risk theory or just risk theory. [101]. Schmidli. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. lighttailed distributions (sometimes the term . Embrechts et al.. the research literature is often published in journals like Astin Bulletin . the role of the result is to justify to take p = 1. in particular. Since { Rt } has premium rate 1. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Gerber [157]. the assumption > 0 is equivalent to p < 1.
1) The parameter 6 is referred to as the rate or the intensity.g. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) .O(u) can be found in closed form.B(x) satisfies B(x) = O(e8x) for some s > 0.8. s<8. B is heavytailed if b[s] = oo for all s > 0. As in a number of other applied probability areas. INTRODUCTION 'Cramertype conditions' is used). P B[s]= (8Is ) . Example 2 .2 and /LB is the mean of B. and can also be interpreted as the (constant) failure rate b(x)/B(x). B[s] is finite for some s > 0. if 1 °O AB Jbos x B(dx) > 0. then the conditional distribution of X . the m.6 CHAPTER I. a fact which turns out to contain considerable information. 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'.f. but different more restrictive definitions are often used: subexponential. Here lighttailed means that the tail B(x) = 1 .u at the time of ruin given r(u) is again exponential u with rate 8. On the more heuristical side.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. Equivalently. (2.e. In particular.f. regularly varying (see below) or even regularly varying with infinite variance. i. For example in the compound Poisson model. where B(bo. 6 has density r(p)xPleax b(x) P and m. the exponential distribution is by far the simplest to deal with in risk theory as well.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. In contrast.g.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). and heavytailed distributions. The crucial feature is the lack of memory: if X is exponential with rate 6. 2a Lighttailed distributions Example 2.2) = 0.3) . for the compound Poisson model with exponential claim sizes the ruin probability .
the Gamma density (2. 2.c. B(x) = r(p) Asymptotically. p) °° where r (x.v.. we develop computationally tractable results mainly for the Erlang case (p = 1. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p . This special case is referred to as the Erlang distribution with p stages. and exponential with rate d. . is > 1.) VarX1 (EX )2 p is < 1 for p > 1. p).. . 0 < ai < 1. In particular. > 1 for p < 1 and = 1 for p = 1 (the exponential case). by Grandell & Segerdahl [175] and Thorin [369].v. or just the Erlang(p) distribution..1) (or the 1/pth root if p < 1). i = 1. P b(x) = r` aibiea. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2...ate (b2 ): L• i=o In the present text. if p is integer and X has the gamma distribution p. are i. u .).d.. one has r(bx. X2. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. . the squared coefficient of variation (s.y i=1 where >i ai = 1.c.. where X1. among others. u Example 2 .i. An important property of the hyperexponential distribution is that its s. p) = J tPletdt. then X v Xl + • • • + X.1 Poisson events in [0. In particular. Ruin probabilities for the general case has been studied. The exact form of the tail B(x) is given by the incomplete Gamma function r(x.2) can be considered as the pth power of the exponential density (2. 0. p.2. x] so that B(x) = r` e.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions.
7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . Important special cases are the exponential.8 CHAPTER I.6. which is slightly smaller but more amenable to probabilistic reasoning. The couple (a. The density and c.7) are possibly complexvalued but the parameters in (2. However.. q2 q3 (2.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m.f.8) are realvalued..g. . Example 2 . We give some theory for matrixu exponential distribution in VIII. INTRODUCTION Example 2 .d. T) or sometimes the triple (E. of which one is absorbing and the rest transient. The parameters of a phasetype distribution is the set E of transient states.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i.1 and defer further details to u Chapter VIII.xo)+ is covered by the reinsurer). This class of distributions is popular in older literature on both risk theory and queues.8) j=1 j=1 j=1 where the parameters in (2. the Erlang and the hyperexponential distributions.e. a.6. See XI. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. equivalently. resp. 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. 1)' is the column vector with 1 at all entries. T) is called the representation. Example 2 .(2.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. are b(x) = aeTxt. We give a more comprehensive treatment in VIII. 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 . the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities.f. (or. B(x) = aeTxe where t = Te and e = (1 . there exists a xo < oo such that B(x) = 0 for x > xo.
p is defined as the distribution of ev where V . the mean u is eµ+a /2 and the second moment is e2µ+2o2. (2.pl = 1 W (logx . a2). It follows that the density is 't (1ogX .13) u .1. b(x) = crx''le`xr. or equivalently as the distribution of a°U+µ where U . There are various variants of the definition around.1). All moments are finite. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . Here failure rates b(x) = b(x)/B(x) play an important role.9) which is heavytailed when 0 < r < I. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. Example 2 .N(0. 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).N(p.u l b(x) = d dx or J ax lor 1 exp Asymptotically. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. In particular. (2. we obtain the Weibull distribution B(x) = eCx'. the tail is B (x ) 2 x. (2.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. u Example 2 .10) The loinormal distribution has moments of all orders.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. Writing c = d/r.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2.2.p a 1 (2. x < a.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed. and then b(x) = 0. However. the exponential distribution representing the simplest example since here b(x) is constant. a)/A)a+1' x > a.
examples of distributions with regularly varying tails are the Pareto distribution (2.L( x ).10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. in particular. 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). A = 1 and X is standard exponential.2). Thus.14) The pth moment is finite if p < 5 and infinite if p > 5.'s of the form YX. satisfies L(xt)/L(x) 4 1. i.x6+lr(p) (2.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. the loggamma distribution is a Pareto distribution. (2. another standard example is (log x)').16) 11 Example 2. u Example 2 . x 4 oo (any L having a limit in (0. The density is 8p(log x)pi b(x) .(1 + Zx + $ p = 3. u . { s () 1s+3s29s3log(1+2s I p=3. For p = 1.e. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. the density is { 3 (1 . oo) is slowly varying . 6 is defined as the distribution of et' where V has the gamma density (2. x + 00. where Y is Pareto distributed with a = (p . B(x) = O(xP).(1 + 2x + 2x2)e2x) p = 2 (2.v.15) x2 + 16x3 ) a3x/2) 3 (1 .12) (here L (x) * 1) and ( 2. The simplest examples correspond to p small and integervalued. (2. in particular. In general. INTRODUCTION Example 2. Choudhury & Whitt [1] as the class of distributions of r.10 CHAPTER I.1)/p.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) .17) where L (x) is slowly varying.13).
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.1. U2. 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. From a practical point of view. (2. 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... but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. We return to a closer study in IX. though the proof of this is nontrivial..13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. one may argue that this difficulty is not resticted to ruin probability theory alone. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. this phenomenon represents one of the true controversies of the area. for example the lognormal distribution is subexponential (but not regularly varying). 3 The arrival process For the purpose of modeling a risk process . and so is the Weibull distribution with 0 < r < 1.18) B(x) It can be proved (see IX. but the model also admits a natural interpretation : a large portfolio of insurance holders . Also.1) that any distribution with a regularly varying tail is subexponential. When studying ruin probabilities. Thus. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. and based upon such information it seems questionable to extrapolate to tail behaviour. Namely. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. However. which each have a ( timehomogeneous) small rate of experiencing a . Similar discussion applies to the distribution of the accumulated claims (XI.3. the claim size distribution represents of course only one aspect (though a major one).. THE ARRIVAL PROCESS 11 Example 2. The reason is in part mathematical since this model is the easiest to analyze. the knowledge of the claim size distribution will typically be based upon statistical data.
5. epidemics in life insurance etc.i. The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too.8 (t) is a periodic function of t. when Jt = i. to be studied in Chapter V. In others. it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. the negative binomial distribution. in particular to allow for certain inhomogeneities.. Mathematically. This model . not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . 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). The compound Poisson model is studied in detail in Chapters III.(3. we study this case in VI .3(t) fluctuating over time. gives rise to an arrival process which is very close to a Poisson process. and also that the ruin problem may be hard to analyze .d. with a common term {Nt} is a Markovmodulated Poisson process . such that 8(t) = . however. Nevertheless . too general and one neeed to specialize to more concrete assumptions . In order to prove reasonably substantial and interesting results . To the author 's knowledge . Historically. getting away from the simple Poisson process seems a crucial step in making the model more realistic.g. This applies also to the case where the claim size distribution depends on the time of the year or . Another one is Cox processes. with the extension to premiums depending on the reserve. INTRODUCTION claim . the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. Some of them have concentrated on the marginal distribution of NT (say T = one year ). T2.e. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . where {/3 (t)}too is an arbitrary stochastic process .. are i. I. e. An obvious example is 3(t) depending on the time of the year (the season). the periodic and the Markov modulated models also have attractive features . but with a general not necessarily exponential distribution ).12 CHAPTER I.. 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 . its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). which facilitate the analysis. However . Cox processes are. 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. see 11. IV (and.. has some mathematically appealing random walk features .6. in Chapter VII). A more appealing way to allow for inhomogeneity is by means of an intensity . so that .
T) = P(VT > u). reliability. this amounts to Vo having the stationary distribution of {Vt}).1) where V is the limit in distribution of Vt as t + oo. Some of these have a certain resemblance in flavour and methodology. Mathematically. Similarly. extreme value theory. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. The study of the steady state is by far the most dominant topic of queueing and storage theory. methods or modeling ideas developed in one area often has relevance for the other one as well. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. and the limit t 4 oo is the steadystate limit.1) holds as well provided the risk process has a premium rule depending on the reserve. ruin probabilities for risk processes with an input process which is renewal. however. The ones which appear most related to risk theory are queueing theory and dam/storage processes. others being branching processes. In fact. with Poisson arrivals and constant release rule p(x) = 1. time series and Gaussian processes. queueing theory. interacting particle systems. genetics models. stochastic geometry. point processes and so on.1) permitting to translate freely between risk theory and the queueing/storage setting. and which seems well motivated from a practical point of view as well. 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 . (4. it is desirable to have a set of formulas like (4.0 (u. More generally.4. In the setting of (4. and here (4.1). The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models.6) . Thus. dam/storage processes. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. 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. this gives only f0 O°i (u)du which is of limited . that quite often the emphasis is on computing expected values like EV. R = p(R) in between jumps. and a lot of information on steadystate r.'s like V is available. 0(u) = P(V > u).v. others are quite different. Markovmodulated or periodic can be related to queues with similar characteristics. It should be noted. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. stochastic differential equations.
B(x) = ebx. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases . 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. see Corollary VII.1). 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.6. 3.2). though overlapping. 3.3. have to some extent a different flavour.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 ).T). which gives a sample path version of (4. The infinite horizon (steady state ) case is covered by letting T oo. Here O(u) = peryu where 3 is the arrival intensity. • The compound Poisson model with some rather special heavytailed claim size distributions. see Corollary III..'s like the environmental process {Jt} in a Markovmodulated setting. as is typically the case. 4b Exact solutions Of course . • The compound Poisson model with a claim size distribution degenerate at one point.p(y) y^ Jo p(x) can be written in closed form.g.3. Thus .1 . see Boxma & Cohen [74] and Abate & Whitt [3]. which can be expanded into a sum of exponential terms by diagonalization (see. p = 0/8 and y = 8 . Vi(u.8.3. INTRODUCTION intrinsic interest . .v. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. Similarly.14 CHAPTER I. Here ?P(u) is explicit provided that . the two areas. 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. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. A prototype of the duality results in this book is Theorem 11.3.1. The fact that Theorem H. e . the functions w x f d 1 exdx () .
Abate & Whitt [2]. T) can then be calculated numerically by some method for transform inversion. • An astable Levy process with drift . [s.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). a2 (x): Ip (u) = where S(u) = f °O exp {. We don't discuss Laplace transform inversion much. O(u. However.b(u)du .S(u) 1S(oo) f °D exp {.2) is the natural scale. For the finite horizon ruin probability 0(u.4. T) themselves. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. see VIII. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI . say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. .f f 2µ(y)/a2(y) dy} dx  (4. 191). relevant references are Grubel [179]. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). Ab(u). it is easier to find the Laplace transforms = f e8 . Here are some of the main approaches: Laplace transform inversion Often.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. esuTb( u. f {eXp U LX 2.u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. Given this can be done. 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 .ff 2µ(y)/a2(y) dy} dx . the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. but are somewhat out of the mainstream of the area . (u.Lef$er function. Embrechts. 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. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p.7. T). 1). A notable fact ( see again XI. the formulas ( IV.
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. U is explicit in terms of the model parameters.g.and integral equations The idea is here to express 'O(u) or '(u. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m.4) 00['Y]1)'Y = 0. In the compound Poisson model with p = 1.) B[s]. However. . whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!).3) where C = (1 . most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight.7. and in particular the naive idea of conditioning upon process behaviour in [0. T) as the solution to a differential.3. 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. u * oo. and carry out the solution by some standard numerical method.1) and y > 0 is the solution of the Lundberg equation (4. which can equivalently be written as f3 [7] = 1 +13 .Ce"u.or integral equation.3) in the compound Poisson model which is an integral equation of Volterra type. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential).p)/(13B'[ry] . (4. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). and in quite a few cases (Chapter VIII). see VIII.f. One example where this is feasible is the renewal equation for tl'(u) (Corollary III. dt] most often leads to equations involving both differential and integral terms. it states that i/i(u) .
T). In the case of heavytailed distributions. T) for large u are available in most of the models we discuss. J B dx. incorporating correction terms may change the picture dramatically. In particular. some further possibilities are surveyed in 111 . the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. in some cases the results are even more complete than for light tails. For example. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. often for all u > 0 and not just for large u. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself.7 and IV. 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. In fact. in such cases the evaluation of C is more cumbersome. u > oo. This list of approximations does by no means exhaust the topic. . the claim size distribution should have an exponentially decreasing tail B(x).4. See Chapter IX. However. for the compound Poisson model ^(u) p pu In fact .6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. a Markovian environment or periodically varying parameters. Diffusion approximations are easy to calculate. other approaches are thus required. It has generalizations to the models with renewal arrivals. However. when the claim size distribution is of phasetype. Approximations for O(u) as well as for 1(u. but typically not very precise in their first naive implementation.2. corrected diffusion approximations (see IV. (4. 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.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. Large claims approximations In order for the CramerLundberg approximation to be valid.
the difficulty comes in when drawing inference about the ruin probabilities.18 CHAPTER I. which is a standard statistical problem since the claim sizes Ui. . This is proved for the compound Poisson model in 111. empirical evidence shows that the general principle holds in a broad variety of settings. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. as a general rule. e. . UNT may be viewed as an interpolation in or smoothing of the histogram). more importantly. For example.) at various places and in various settings. For example. it has the advantage of not involving approximations and also.. . INTRODUCTION Compared to the CramerLundberg approximation (4. of being somewhat easier to generalize beyond the compound Poisson setting. they have however to be estimated from data.8.g.U(k)) i =k+ i . This procedure in itself is fairly straightforward. 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. lower bounds etc..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. When comparing different risk models. 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. fitting a parametric model to U1. obtained say by observing the risk process in [0. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. in the compound Poisson model. . it is a general principle that adding random variation to a model increases the risk.. In practice. UNT are i.k (U(`) . one expects a model with a deterministic claim size distribution B. How do we produce a confidence interval? And. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion.i. However.3). given NT. and to plot the empirical mean residual life 1 N . The standard suggestion is to observe that the mean residual life E[U .d. say degenerate at m. . . to have smaller ruin probabilities than when B is nondegenerate with the same mean m. T]. though not too many precise mathematical results have been obtained. However.
3) or Section 3 of Chapter VI are referred to as Proposition VI.3 (or just VI.2.i. Thus Proposition 4. and look at them to see whether they exhibit the expected behaviour or some surprises come up..5.3). < U(N) are the order statistics based upon N i. formula VI. However.4.v. in this book referred to as [APQ]). Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r.. where U(1) < .d.3) and Section VI. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. in all other chapters than VI where we just write .(5. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available.. CONVENTIONS 19 as function of U(k). 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. UN..e. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. For example. The chapter number is specified only when it is not the current one.2. See further Embrechts. (or a functional of the expectation of a set of r. claims U1.v's) which can be generated by simulation. . but is not very satisfying. . and in fact methods from that area can often be used in risk theory as well . Klnppelberg & Mikosch [134]. formula (5. The infinite horizon case presents a difficulty. reference [14]. this is a straightforward way to estimate finite horizon ruin probabilities. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. and of course the method is relevant in risk theory as well. because it appears to require an infinitely long simulation. We look at a variety of such methods in Chapter X.. Truncation to a finite horizon has been used. A main problem is that ruin is typically a rare event (i. respectively. to observe whether one or the other limiting behaviour is apparent in the tail. Still.
s. or a more precise one like eh . w. If.c.f. mation. as for typical claim size distributions. References like Proposition A.d.ceax. cumulative distribution function P(X < x) c.h.v. B(dy).d.s.v.g. see under b[s] below.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. E expectation.o. with respect to w.h.f.r. with probability Mathematical notation P probability. (A. random variable s. h + 0.g.20 CHAPTER L INTRODUCTION Proposition 4. The Laplace transform is b[s].d.g. formula (5. A different type of asymptotics: less precise. B(x) the tail 1 .2.f.B(x) = P(X > x) of B. left hand side (of equation) m.4. (moment generating function) fm e82B(dx) of the distribution B. independent identically distributed i. i.f.t. infinitely often l. i. moment generating function. oo). cumulant generating function.g. .g.f. and for a defective probability distribution IIGII < 1. IIGII the total mass (variation ) of a (signed ) measure G . n! 27r nn+1/2en.29) refer to the Appendix. EX2/(EX)2. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. In particular. n i oo. squared coefficient of variation.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit. B is concentrated on [0. . B[s] the m.f. B(x) = P(X < x) = fx. for a probability distribution IIGII = 1.i.3) or Section 3. if B(x) .e. say a heuristic approxi1 + h + h2/2. right hand side (of equation) r. Abbreviations c. r. E. then for 1s < 5).p. log E[s] where b[s] is the m.
7r. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. E[X.5. Matrices and vectors are denoted by bold letters. an example or a remark. . row vectors have lowercase Greek letters like a. though slightly more complicated. matrices have uppercase Roman or Greek letters like T. all stochastic processes considered in this book are assumed to have sample paths in this space.e. intensity interpretation. the value just before t. . CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. (the dimension is usually clear from the context and left unspecified in the notation). . xa. Thus. Notation like f3i and 3(t) in Chapter VI has a similar .e. oo) the space of Rvalued functions which are rightcontionuous and have left limits. i. 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.. the processes we consider are piecewise continuous. only have finitely many jumps in each finite interval. the ith unit row vector is e'i. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi).A] means E[XI(A)]. R(s) the real part of a complex number s. a2) the normal distribution with mean p and variance oa2. i. F o r a given set x1. I(A) the indicator function of the event A. and column vectors have lowercase Roman letters like t.e. Usually. Usually. the ith entry is 1 and all other 0.. In the Frenchinspired literature. 0 marks the end of a proof. N(it. a.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). Xt_ the left limit limstt X8f i. Unless otherwise stated. A. 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. of numbers. 21 D [0.
cf.5. or quantities with a similar time average interpretation. FL.22 CHAPTER L INTRODUCTION B the claim size distribution. ry The adjustment coefficient. I. Notation like BE and B(t) in Chapter VI has a similar.1. cf. though slightly more complicated. VI. I. cf. J the rate parameter of B for the exponential case B(x) = eby. p the net amount /3pB of claims per unit time. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. interpretation. 111. e. 'q the safety loading . .g.1.5.
in part. The general theory is. The duality results in Section 3 (and. All results are proved elsewhere . however. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. however. used in Chapter VI on risk processes in a Markovian (or periodic) environment. 23 .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. Sections 4. a parallel selfcontained treatment is given of the facts needed there. More precisely. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. the level of the exposition is. Sections 4. The topic is. in most cases via likelihood ratio arguments. The reader should therefore observe that it is possible to skip most of the chapter. somewhat more advanced than in the rest of the book. The likelihood ratio approach in Section 2 is basic for most of the models under study. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. in particular at a first reading of the book. 5) are. Due to the generality of the theory. When encountered for the first time in connection with the compound Poisson model in Chapter III. strictly speaking. not crucial for the rest of the book. however.
V) (u. StUit. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. however.5 can be skipped. As usual. Let e(u) = ST(u) . the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). claim size distribution B and p = . 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.0.QµB < 1. T(u) < oo] + 0 = eryuE [e7Vu).T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. The more general Theorem 6.2) takes the form 1 = E [e'ys(). f1 . Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem . Proposition 1. 1 Martingales We consider the claim surplus process {St} of a general risk process.1 is basic for the study of the compound Poisson model in Chapter III. T(u) < T] + E [eryST . the time to ruin r(u) is inf It > 0 : St > u}. {e'YS° }t>0 is a martingale. using r(u) A T invokes no problems because r(u) A T is bounded by T). T(u) > T] .1 Assume that (a) for some ry > 0. We get 1 = Ee7So = E e'Y S(. (b) St a$ oo on {T(u) = oo}..(u).2 Consider the compound Poisson model with Poisson arrival rate .)AT = E [e7ST(°). (1. Thus N. and in the limit (1. Example 1 . T) = P(T(u) < T).u denote the overshoot. and the ruin probabilities are ip(u) = P (T(u) < oo).2) As T > oo.24 CHAPTER II.
6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists.1. B(x) _ edx. with common distribution B (and independent of {Nt}).(„)_ = x is that of a claim size U given U > u .2.3/6 < 1./3. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.r" where y = S . u Corollary 1.3 Assume that {Rt} is Brownian motion with variance constant o.ap.1) shows that Eels. 1. are i.g. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .a.1. and thus Ee7s° = 1.4 (LUNDBERG ' S INEQUALITY ) tion 1 .= e"(') where K(a) = . Proof Since c(a) = /3 (B[a] . A simple calculation (see Proposition III.x. MARTINGALES 25 where {Nt} is a Poisson process with rate . Now at the time r(u) of ruin {St} upcrosses level u by making a jump . By standard formulas for the m. From this it is readily seen (see III. condition (a) of Proposition 1.. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx . the martingale property now follows just as in Example 1.1.Q(B[a] . Since {St} has stationary independent increments.1) . of the normal distribution. the conditions of Proposition 1. and thus by the memoryless property of the exponential distribution .5 For the compound Poisson model with B exponential.f.1 is satisfied.a = a .i. and p =. Since {St} has stationary independent increments. and (b) follows from p < 1 and the law of large numbers (see Proposition III.u + x is again just exponential with rate S. Eeas° = e"(°) where n(a) = a2a2/2 .2(c)).Ft = a(S" : v < t).1) . Thus. Example 1 .d.1 are satisfied.2 and drift p > 0. Then {St } is Brownian motion with variance constant o2 and drift p < 0. O(u ) < e7". u Corollary 1. The available information on this jump is that the distribution given r(u) = t and S. and thus Ee'rs° = 1. Thus.a it is immediately seen that y = S .Q. the conditional distribution of the overshoot e(u) = U .Q and the U. Under the conditions of Proposi Proof Just note that C(u) > 0. the ruin probability is O(u) = pe.
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. Embrechts. then z/'(u) = e7" where 'y = 21A/a2. as shown by the following example this setup is too restrictive: typically'. But if a $ ^ .26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. Theorem 2. Thus the sets S = I tlim +oot t =. oo). and by the law of large numbers for the Poisson process . A]. More recent references are Dassios & Embrechts [98]. 0 and claim size distributions B.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0.v..F). P on (DE. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3.Ft}too and the Borel afield F.3 below). the parameters of the two processes can be reconstructed from a single infinite path. Example 2 . then S and S are disjoint . F(S) = P(S) = 1. I. Delbaen & Haezendonck [103] and Schmidli [320]. and F. A somewhat similar u argument gives singularity when B $ B. A E Ft. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. F). The number Nt F) of jumps > e before time t is a (measurable) r. However. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. which we equip with the natural filtration {. Grandell & Schmidli [131]. Proof Just note that ^(u) = 0 by continuity of Brownian motion. on (DE. cf. hence so is Nt = limfyo N2`i. Grandell [171]. . B. (2.e. Two such processes may be represented by probability measures F. [172]. and is further exploited in his book [157]. P are then singular (concentrated on two disjoint measurable sets). The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random.6 N S = { lim Nt I t +00 t gJ are both in F.1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt.1 Let F.
. then {Lt} is a nonnegative martingale w. By the martingale property. This proves (i). that the restriction of P to (DE. Finally. Then Ft (A) = E[Lt.t. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. F the Borel o•field and P a given probability measure on (DE. implies that E[LtI. then { 1 P(G) = EG . If r is a stopping time and G E PT. A] = E[Lt. P be as in Proposition 2. _.3 Let {Lt}.F8] = L8 and the martingale property. G C {T < oo}. .1) holds. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). Lt < 0] can only be nonnegative if P(A) = 0. if for some probability measure P and some {. define P by Pt (A) = E[Lt.Tt) is absolutely continuous w. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE.1) and nonnegativity by letting A = {Lt < 0}. we have E [ LTIFT]1 = LT on {T < T}. Lets < t. Proof Under the assumptions of (i). Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. the restriction of P to (DE.2 Let {Ft}t>o be the natural filtration on DE.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). Proposition 2. F) such that ELt = 1. Conversely.Y) such that P(A) = Pt(A).Pt)) The following result gives the connection to martingales. (i) If {Lt}t> o is a nonnegative martingale w.1) holds.e.t. G ] = E [LT .2(i). Ft). A E Ft . A].2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. P) such that LLt = 1.r. .r.Pt}adapted process {Lt}t>o (2. ELt = 1 follows by taking A = DE in (2. 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 . Then P(A) = E[Lt. A]. ({. The truth of this for all A E Y.Ft}. (ii) Conversely.F).F such that (2. A E F.r. using the martingale property in the fourth step.t. 1 J (2. Hence E [_ .2. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. ({Ft} . A E F8. then there exists a unique probability measure Pon .
say.1.1: Corollary 2. (2.t. we assume for simplicity that {Xt} has Dpaths. is Markov w. we need the concept of a multiplicative functional. and letting T * oo. where {Rt} is the risk reserve process.4) compared to (1. Consider a (timehomogeneous) Markov process {Xt} with state space E.4 Under condition (a) of Proposition 1.Ft} .3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1.1).r. each F. Rt) or Xt = (Jt. (2. we have F(G) = V )(u). t.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). Now just rewrite the r. 1 Since everything is non negative. . one would typically have Xt = Rt. SOME GENERAL TOOLS AND RESULTS In the general case . and this problem will now be studied.s.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. T(u) < oo]. in continuous time (the discrete time case is parallel but slightly simpler).Gn {r <T} . The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure.1) in Proposition 1. In the context of ruin probabilities. u From Theorem 2. Xt = (Jt.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . 5) for processes with some randomwalklike structure. is nonnegative and has Ey Lt = 1 for all x. First we ask when the Markov property is preserved. A change of measure is performed by finding a process {Lt} which is a martingale w. Lr(u) 11 The advantage of (2. of (2. first in the Markov case and next (Sections 4. For the definition. Xt = St. To this end.4) Proof Letting G = {r(u) < oo}.2) follows by monotone convergence..r. {St} = {u . both sides are increasing in T. applying (2.2) by noting that 1 = ersr(„) = e1'ue7Ou).28 CHAPTER II.O(u) = eryuE[e'YC(u). The crucial step is to obtain information on the process evolving according to F.t.h.. the natural filtration {. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. St).
let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x. (2. t and let Px be the probability measure given by t. The converse follows since the class of r.F8measurable r.. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional. or.v.5) Pxa.Ft] = Ex.v.T9measurable Y8. A].'s of the form fl' f.Ftmeasurable. o 9t = V.Ft].s.5 Let {Xt} be Markov w.r. for all x. By definition of Px. Vt+e. s.7). the natural filtration {Ft} on DE. (2.[Y. since Zt • (Y8 o Ot ) is . and then L. Similarly.Ft+8measurable.v.v.v. The precise meaning of this is the following: being . t.. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2.t. oo).7) for any Ftmeasurable Zt and any . Zt.(Xtitl) with all t(i) < t + s. t] * [0. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.YB] for any Ftmeasurable r.Y8f t < s. 0 . which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. (2. the Markov property can be written E. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].6) for any .5) are Tt+e measurable. Proof Since both sides of (2.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any . ({Xt+u} 0<u<8) Theorem 2. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)]. o 9tI. (2. nonnegative and Lt+8 = Lt•(Lso9t) (2.(A) = Ex [Lt. Y8. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.Pt+8measurable r.8) which is the same as (2.'s of the form Zt • (Y8 o 0t) comprises all r.Ft }.7).6) implies (2. Indeed. where Ot is the shift operator.2.
6 For {u .1) The initial condition is arbitrary.5 can be found in Kuchler & Sorensen [240]. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. We work on a finite time interval [0. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process.e. the premium rate is p(r) when the reserve is r (i. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. Indeed..Ft] = LtE[L8 o 9t I. . t] = LtExt L8 = Lt. {Vt} . 3 Duality with other applied probability models In this section. Thus R = Ro + f p(R8) ds . 0 < vl < . then Remark 2. (using the Markov property in the second step) so that the martingale property is automatic. In between jumps. Ro = u (say). t. In between jumps.. < aN < T. reflection at zero and initiar condition Vo = 0. the arrival epochs are Qi. ... } E[Lt+B I. CN. A more elementary version along the lines of Theorem 2... The result is a sample path relation. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. (3. u Notes and references The results of the present section are essentially known in a very general Markov process formulation.. The corresponding claim sizes are Ul. The formulation has applications to virtually all the risk models studied in this book. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. . UN. see Dynkin [128] and Kunita [239]. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. .. aN where or* = T UN_k+l.. with a proof somewhat different from the present one. The storage process {Vt }o<t<T is essentially defined by time reversion..At where At = k: vk <t U.30 CHAPTER H. R = p(R)). it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x.. More precisely . .
3. Theorem 3.11 4..AT_t._.T) = P(VT > u)...___ . Then rt°) > rt°) for all t when u > v. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous. The sample path relation between these two processes is illustrated in Fig.e. (3.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0...__. :. In particular.__.. . DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.1 The events {T(u) < T} and {VT > u} coincide...x. V)(u. instead of (3.1) we have Vt = At  f P(Vs)ds where A= U= AT . V = p(V)).1. That is. {Vt} remains at 0 until the next arrival).3.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability. (3..____•_.1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u.
3). and since ruin can only occur at the times of claims. one may feel that the interaction between the different areas has been surprisingly limited even up to today. the distinction between right. Theorem 3. this represents a model for storage. we have RQ„ < 0 so that indeed r(u) < T. and a general premium rule p(r) when the reserve is r. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. The results can be viewed as special cases of Siegmund duality. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. Hence RQ„ > 0 for all n < N.d. Then the storage process {Vt} has a proper limit in distribution. with distribution B.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Then Vo. Hence if n satisfies VVN_n+1 = 0 (such a n exists. if nothing else n = N).U1 = Rol. see Siegmund [344]. . Historically. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig.1 and its proof is from Asmussen & Schock Petersen [50].2 from Harrison & Resnick [188].i. Proof Let T ^ oo in (3. Then similarly VVN = r0. The arrival epochs correspond to rainfalls. Historically. Corollary 3. = r(VT) .3 and being i. If VaN > 0. say of water in a dam though other interpretations like the amount of goods stored are also possible.U1 > roil . We get: Corollary 3.Ul < roil  Ul = RQ„ Va1V_1 < RQ2.T l . say V. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell.2 Consider the compound Poisson risk model with a general premium rule p(r). 3. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . we can repeat the argument and get VoN_1 > Ra2 and so on. if and only if O(u) < 1 for all u.1 with Ro = u = u2). and so on. we have r(u) > T. Resnick & Tweedie [79]. 3. Nevertheless.1 with Ro = u = ul). and then '0 (u) = P(V > u). Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions.
if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN..2) satisfies the same recursion as in (4. Z2 = Y2. where the Yi are i . ...h... ZN = .. For discrete time random walks . . . the Lindley process Wo.N From this the result immediately follows. 0 Corollary 4. i. N min (Y1 + • • • + YNn) n=0. W1. W2. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo. n=0. ..... ..2) (for a rigorous proof.1.. For a given i. = Xo + Y1 + • • • + Y.1)).d. N min (Y1 + + Yn).. 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.. Proof By (4.min n=0. WN = YN .YNn+1) n=0.1.. has a proper limit W in distribution as n + oo..Y1 according to Wo = 0.. 0). Z2. . is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4. Z2.. {Wn}n=0.. Here F is a general probability distribution on R (the special case of F being concentrated on {1. Z2 .. can be viewed as the reflected version of the random walk with increments Z1. there is an analogue of Theorem 3.1. Xo = 0.1.w. Theorem 4.. generated by Z1..i. Then the events {r(u) < N} and {WN > u} coincide.2)..d. hits (oo. with common distribution F (say).s.N (4. . as long as the random walk only takes nonnegative values.. .e. just verify that the r .1) Thus {Wn}n=o.1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}.Yl min (YN . and is reset to 0 once the r. WN be the Lindley process generated by Z1 = YN.. I..1. .. Most often. (c) The Lindley process {WN} generated by Zl = Y1....1 in terms of Lindley processes .... W1. Z2 = YN_1 i ... . Let further N be fixed and let Wo... In particular. R valued sequence Z1. evolves as a random walk with increments Z1i Z2.2 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)...1..4....... of (4..
(Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u )...1 have the same distribution for n = 0. . By Kolmogorov's 01 law. (Yi + • • • + Yn) > oo a. (d) #. (e)Yi+•••+Yn 74 .2 and Theorem 4. N.=o.5 does not necessarily lead to a random walk: if. a Markovian change of measure as in Theorem 2. . Y1) has the same distribution as (Y1. equivalently.1.s. + Z. Similarly. Thus the assertion of Theorem 4.the result is a sample path relation as is Theorem 3.. 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. .N so that WN _P4 M = supra=0. Combining these facts gives easily the equivalence of (a)(d).ooa. then the restrictions of Fx..i.1.s.. .. YN). . doubly u infinite (n'= 0. W v m and P(W > u) = P (m > u) = 0(u). assumption on the Z1.1 is equivalent to WN D MN = (Z1 + ..... ±2. Clearly. . SOME GENERAL TOOLS AND RESULTS (d) m = inf. . the Lindley processes in Corollary 4. (e).. w. 176) but appears to be rather intractable. In general.34 CHAPTER II..) and defines Zn = Yn.... .1 is actually not necessary . ±1.. ZN or.l.. a sufficient condition for (e) is that EY is welldefined and > 0. In that case ..1. or M < oo a.) sup n=0. the Y1.. Proof Since (YN.g. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.o.. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists.. Next consider change of measure via likelihood ratios. For a random walk.1.s.2. there is a more general version of Corollary 4.. 0 By the law of large numbers. ..d.s . The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . ..l.3 The i.. e.. One then assumes Yn to be a stationary sequence.g. either M = oo a. YN in Theorem 4. Remark 4 . .
f.3) 1Pxa. y ). and define Ln by (4. the changed increment distribution is F(x ) = E[h(Y). Then the change of measure in Theorem 2.. Conversely. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. Then the change of measure in Theorem 2. Y1). e. Y) = h(Y ) a. In particular. 100 ). for some function h with Eh(Y) = 1.4) ({Ln} is the familiar Wald martingale ..4).. cf. 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. this means E(g(x.Y2) = h(1'i)h(I'a).4.g. u A particular important example is exponential change of measure (h(y) = e°y'(") where r. Y) f (Y)] for all f and x.3) holds for n = 1. . and so onforn =3. In that case... (a) = log F(a] is the c.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) . of F)... the random walk property implies Ex f (Y1) = Eo f (Y1 ).4. implying g (x. We get: Corollary 4.5 corresponds to a new random walk if and only if Ln = h(Y1) . we get L2 = L1 (L1 o91 ) = h(Y1)g(X1.4 Let {Ln} be a multiplicative functional of a random walk with E_L. Y < x].g. Breiman [78] p. Since L1 has the form g (Xo.s.s. = 1 for all n and x. (4.3) holds. where h (y) = g(0. For n = 2.nrc(a )} (4. h(Yn) (4. Y ) f (Y)] = E[g(O.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . Proof If (4.
Ft] = Eof (X. this description needs some amendments. the claim surplus process for the compound Poisson risk model . which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . or imbedded into continuous time processes . Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . Xt(2)_t(l). i.36 CHAPTER II.t(i . The simplest case is 3 = JhvMM < oo..1). Xt =Xo+pt+oBt+Mt. In continuous time. (4. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}.Xt)I. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . see Chapter V. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). given by a the increment distribution F(x) = P(Xn+l . the pure jump process is given by its Levy measure v(dx). a positive measure on R with the properties e J x2v(dx) < oo. say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples).6) More precisely. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). v2 = 0 and v = 3B). corresponds to a process with stationary independent increments and u = p. {Xt} can be written as the independent sum of a pure drift {pt}. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. e f x:IxJ>e} v(dx) < oo (4. 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. (4. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . however.. we are . with premium rate p. In discrete time.). < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . they arise as models for the reserve or claim surplus at a discrete sequence of instants. In risk theory. However. . the tradition in the area is to use continuous time models.7) for all e > 0.. but we omit the details ). say the beginning of each month or year .. {Xt} is a random walk.Xn < x). Roughly.e. .5) Note that the structure of such a process admits a complete description.
is easily seen to be f3pB < 1. oo].8) O<t<T (assuming Wo = Xo = 0 for simplicity). where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f".6 In the compound Poisson risk model with constant premium rate p(r) .1)v(dx) (4.1. then Ee'(xtxo) = Eoeaxt = etx(a). and b(u) = P(V > u).min Xt (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.s. WT = XT . 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) = 1.3 and decreases linearly at rate 1 in between jumps.7 If {Xt} has stationary independent increments as in (4.4. V E [0. having Poisson arrivals with rate . Furthermore.10) . i. (ex . 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. Then the storage process {Vt} has constant release rate 1. 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. 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).v. Now consider reflected versions of processes with stationary independent increments.Q and distribution B of the service times of the arriving customers. O(u. and the reflected version is then defined by means of the abstract reflection operator as in (4.6). jxJ v(dx) < oo.e. cf. defined as a system with a single server working at a unit rate. Proposition 4. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue.] 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. has upwards jumps governed by B at the epochs of a Poisson process with rate . [The condition for V < oo a. T) = P(VT > u). VT + V for some r. Corollary 4.2).
if Lt = e9(xt . of an infinitely divisible distribution (see.1 ) v(dx) . Xt +B . v(dx) = e9xv (dx).38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. (4.1. Q2 = v2. e. we use the characterization (4.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. X8) = Eof (X8)L8 = Eof (X8)• For the second. By explicit calculation . Then the Markov process given by Theorem 2.f. let e" (a ) = Eoeaxl. 5 has stationary independent increments as well. t. we show in the compound Poisson case ( IlvIl < oo) in Proposition III.1 that E eaMt = exp fmoo In the general case . use the representation as limit of compound Poisson processes.Xt)I'Ftl = E [f(Xt+B ..6) are µ = µ + Oo2 . Proof For the first statement . This is of course no coincidence since the distribution of Xl . 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. u Note that (4.1)eexv(dx). Eea(µt + QBt) = et{aµ +a2OZ/2}.xo)tk ( e). .10) is the LevyKhinchine representation of the c.Xt)I Ftl = Eof (X8)g(s.Xt)L8 o 0tIFt] = E [f (Xt+s . then the changed parameters in the representation (4. Theorem 4 .g.g. Chung [86]).11) (eax .Xt)g(s. Xt Xo) with E2Lt = 1 for all x.5) and get E [f(Xt+B . In particular.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 .
0 in the following. Recalling that U1.. MARKOV ADDITIVE PROCESSES 39 Remark 4. 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. As for processes with stationary independent increments .9 If X0 = 0.2.5. MAP stands for the Markovian arrival process discussed below.3 =. corresponding to p = 1. 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. are the arrival times and U1. . a = 0. B(dx) = B[9] B(dx). the corresponding claim sizes . b with b(x) > 0 for all x such that b(x) > 0). Example 4 . dB/dB = b/b when B. Then we can write v(dx) _ /3eOxB(dx) = / (dx). 0.3 and claim size distribution B # B.St)g(Jt+s)I. we write Pi..(3 = . it is then easily seen that Lt = H dB(Ui) i:o.Ft] = Ejt. where .. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. B have densities b.. .3B[B].g. <t whenever the RadonNikodym derivative dB/dB exists (e. v(dx) _ . Ei.3 and claim size distribution B.o[f (S8)g(J8)].1) For shorthand .11 For an example of a likelihood ratio not covered by Theorem 4. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . abbreviated as MAP in this section2. u 5 Markov additive processes A Markov additive processes.. Example 4 .4).(3B(dx). U2.8.0. one reason is that in parts of the applied probability literature. is defined as a bivariate Markov process {Xt} = {(Jt. Ei instead of P2. Thus (since µ = p = 1.l3 and claim u size distribution B. the structure of MAP's is completely understood when E is finite: 2and only there .10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .
If all Fij are concentrated on (0. (That a process with this description is a MAP is obvious. which we omit and refer to Neveu [272] or cinlar [87]. the distribution of which has some distribution Bij.[a) = (Ei[easl. {Jt} is specified by its intensity matrix A = (Aij)i. {St} evolves like a process with stationary independent increments and the parameters pi..Sr_1. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.it = A. In addition. An alternative description is in terms of the transition matrix P = (piA. Proposition 5. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. SOME GENERAL TOOLS AND RESULTS In discrete time.) If E is infinite a MAP may be much more complicated.i.g. by generating Yn according to Hij when J„_1 = i. As an example. 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. In continuous time (assuming Dpaths).jEE• On an interval [t. Y2.40 CHAPTER H.6) depending on i.f.1 For a MAP in discrete time and with E finite.9 EE = (iii&ij[a])i j EE .. with the Y„ being interpreted as interarrival times. let {Jt} be standard Brownian motion on the line. vi(dx) in (4.o(Ji = j. v. t+s) where Jt . Jn = j. As a generalization of the m.. oo).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 . the converse requires a proof.. . 1 J1 ='^])iJEE = (Fij[a])i . Y1 E dx) where Y„ = S„ . a MAP is the same as a semiMarkov or Markov renewal process..J1=j)= Fij (dx) Pij In simulation language. consider the matrix Ft [a] with ijth element least Ei .
4c). Then the matrix Pt[a] with ijth element Ei [east. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. pi. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. \ diag Ft[a] = Ft[a]K.(')(a)) diag + (). Jt = k] { xk kEE j la] . Then. Jn+1 = A] = 41 Ei[ e 5„. kEE Jn = k]Ek[e"Y" . By PerronFrobenius theory (see A. Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . u In the following. In matrix formulation . vi(dx) (i E E). j E E) and So = 0. where K[a] = A+ (r. Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. Jt = j] is given by etK[a].1)) . Bij (i. up to o( h) terms. Jt = j] Ejesh'^ + E Ak j hEi [ease . vi(dx).1) } (recall that qjj = 0). 013 . 00 r(i) (a) = api + a2ot /2 + f (e° .1 )v(dx). aSt h = (1 + Ajjh) Ei [east . assume that the Markov chain/process {Jt} is ergodic.5. a= .ijgij(Bij[a] . MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. u Proposition 5. Sn ) yields Ei[easn+ '. qij. we infer that in the discrete time case the .qkj + k?^j qkj Bkj [a] } = Ei [east.2 Let E be finite and consider a continuous time Markov additive process with parameters A. Jt = k] { 1 .
and write k = k(°). Jeast. and we shall take V(a)h(a) = 1.2) where 7r = v(°) is the stationary distribution.tK(a)h(a) J jj it L o is a martingale. Corollary 5. Then h(°) = e. Proof For the first assertion.f.r. h(") may be chosen with strictly positive components.Eikjt = ttc'(0) + ki . just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a).42 CHAPTER II.5 EiSt = tK'(0) + ki . Since v("). Jt = j] . Furthermore. We also get an analogue of the Wald martingale for random walks: Proposition 5. cf.4 Eie"sth(a) = h=a)et?("). u Let k(a) denote the derivative of h() w.h(a)vva)etw(a). its derivatives are 'asymptotic cumulants'.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . Proof By PerronFrobenius theory (see A. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . . cf. Proposition 5. we are free to impose two normalizations.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). Yrh(a ) = 1. (5.t. Eie"sth^a) = e'Pt[a]h( a) = e. In particular. a.c(a) (and h(")). of a random walk.7. h(") are only given up to a constants.3 Ei [east. as will be seen from the following results.4. 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. The function ic(a) plays in many respects the same role as the cumulant g. Corollary 5. The corresponding left and right eigenvectors v(").e=e°tk. Corollary 5.4c).
") }) .5. (5. (5. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . For the second . E=ST = tc'(0)E7. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x).+ k.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. Corollary 5.. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.2ttc (0)Evkjt + 0(i). 8 Also for E being infinite (possibly uncountable ). Squaring in Corollary 5.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk .e. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . tam E tSt a (0). Remark 5 . for a random walk: Corollary 5.a) + ttc (a)2hia ) Multiplying by v=. we differentiate (5.6 For any stopping time T with finite mean. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5.Eikjr ..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. subtraction yields Vary St = tic"(0) + O(1).7 No matter the initial distribution v of Jo.5. MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. u The argument is slightly heuristic (e. one obtains a generalization of Wald's identity EST = Er • ES. [E.4) . the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions. t a oo.5 yields + W (a)k.4. .g. In the same way. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). summing and letting a = 0 yields E„ [St + 2Stkj. the distribution of Jo). More precisely.
h(Jo) Lo is a Px martingale for each x E E. First.s. in particular that f is bounded. 0) = h(i )( 1 + ttc(a)). St) } as follows.f (x) tyo t provided the limit exists. xEE .5) is a martingale . gha(i.(9) {Lt}t>o = . and the family {f LEE given by Theorem 2.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. 0 Proposition 5. u forsEE).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. Given a function h on E. we take the martingale property as our basic condition below (though this is automatic in the finite case). G is defined as Gf (x) = lim Exf (Xt) .44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. let ha(i. s) = ea8h(i). however.5 defines a new MAP. For t small . Remark 5. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i)..6.10 Let {(Jt. 1) (i.e.3b and Remark VI. some extra conditions are imposed. From (5. where {Jt} is deterministic period motion on E = [0. Usually. Then {Lt } is a multiplicative functional. In view of this discussion .. Jt = (s+t) mod 1 P8a. this is.9 The condition that (5. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. however. 0) = n(a) h(i).5. this leads to h(i) + tcha( i. V.6. St)} be a MAP and let 0 be such that h(Jt) OStt. An example beyond the finite case occurs for periodic risk processes in VI. see. (5.4 that { h(Jt) easttK(a) L o (5.6) We shall not exploit this approach systematically.
In particular. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) . ^i = of qij Bij [0] 1 + qij ( Bij [0] .7(dx) Bij [0] Bij(dx) in the continuous time case .7) In particular. Bi(dx) = Bi(dx).1) . We omit the details. 0 < qij < 1 and Bij [0] > 0. (5. in the discrete time case. if vi(dx) is compound Poisson.1) holds for the P. Bi. vi (dx) = f3 Bi(dx) with .5.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. That 0 < qij < 1 follows from the inequality qb <1. Bi [0] Remark 5. In the infinite case .12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. one can directly verify that (5.c(0)e = tc(0)e . .. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0]. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ .1) eft ea' f ij (dx) = Hij (dx) Hij [0] .tc(0)e = 0 ..10 is given by P = eK(e) Oh e) F[e]Oh('). Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. qij = r. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. u Theorem 5.11 below in the finite case. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). 1 + q(b .11 Consider the irreducible case with E finite.ic(0)e = ic(0)Oh e) h(e) .(0)j. 0<b<oo. 0<q<1.Qi < oo and Bi a probability measure. Ai = µi + 0Q. Then the MAP in Proposition 5.
.11.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . since Hij. in continuous time ( 5. Jt = A. Now we can write K[a] =A+A ) ( K[a + 0] . Further Fib (dx ) = P=(YI E dx.tc (') (0) corresponds to the stated parameters µ. F:j with a density proportional to eei . Jl = j) = Ei[Lt. a = 0 in (5. is absolutely continuous w.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) . SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. In matrix notation . are probability measures . Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. Jl = j] :(Yi E dx. Hence the same is true for H=j and H.8) h(.13) for matrixexponentials . this implies k[a] = A 1 ) (K[a + 0] .e) Consider first the discrete time case .tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] .46 CHAPTER II. . (dx) of a process with stationary independent increments follows from Theorem 4.. Yi E dx.t. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' .tc(0)I. it follows that indeed the normalizing constant is H1 [0].8. v. Letting a = 0 yields the stated expression for A.Bay [0]) That k(') (a + 0) . . This shows that F. v= . Similarly. H1.. Jt = j] = hie) . (dx).8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5.r.8). Here the stated formula for P follows immediately by letting t = 1.
[261].Bij[0]) = hjel)ijgijBij[0](Bij[a] . [226] and Miller [260].1). h. has no mass on (oo. Note that G+ is concentrated on (0. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. however. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. oo)..(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . there is. is slightly less general than the present setting. and is typically defective. 0]. see also Fuh & Lai [149] and Moustakides [264]. the literature on the continuous time case tends more to deal with special cases. 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).3 for an infinite E are given by Ney & Nummelin [266].6. [262] in discrete time. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S. hardly a single comprehensive treatment. i. [225]. .. 7+ < oo).7).+ < x. Conditions for analogues of Corollary 5. < x) = 11 (S. Though the literature on MAP's is extensive.e.6. For the Wald identity in Corollary 5. which.)Ajjgij(Bij[a+0] . Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. 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]. however.1) = Aij4ij(Bij[a] .
Recall that B(x) = 1 . the second ladder height (step) is ST+(2) . they have a semiMarkov structure (but in complete generality. In any case. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. 6. 1 For the compound Poisson model with p = 01LB < 1. In other cases like the Markovian environment model.2.d.2) . g(y)R+(dy) = E f g(St)dt. by approximation with step functions . see Fig.5 below.48 CHAPTER K. o 00 (6. 0]. and the maximum M is the total height of the ladder. The main result of this section is Theorem 6.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. where basically only stationarity is assumed. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr.00 ). = ST+(1) Figure 6. which gives an explicit expression for G+ in a very general setting. there are only finitely many).1. On Fig.e. it follows that for g > 0 measurable. Here bo(x) _ B(x)/µB. the dependence structure seems too complicated to be useful). 6.ST+(1) and so on.. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A.1.. oo). G+ is given by the defective density g + (x) =. i.(3B(x ) = pbo(x) on (0. a fact which turns out to be extremely useful.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. the sum of all the ladder steps (if rl > 0. The first ladder step is precisely ST+. at present we concentrate on the first ladder height. i. the ladder heights are i.B(x) denotes the tail of B.i.1.T+ > t)dt = E f 0T+I(St E A) dt. For the proof of Theorem 6. Also. Theorem 6 . To illustrate the ideas. Thus. In simple cases like the compound Poisson model. 0 f T+ (6. R+ is concentrated on (oo. has no mass on ( 0.e. we shall first consider the compound Poisson model in the notation of Example 1.
49 Proof Let T be fixed and define St = ST . 0]. 0 < t < T. has the same distribution as {St}o<t<T. St S* t a Figure 6. That is. . ST < ST_t.ST<St.T+>T) = P(STEA.2(a): T+ > t Figure 6. see Fig.2(b): r+ < t Thus.2.2 R+ is the restriction of Lebesgue measure to (00.ST_t.O<t<T) = P(STEA.ST<St.0<t<T) = F(ST E A. P(STEA.6. 6. THE LADDER HEIGHT DISTRIBUTION Lemma 6 .St<0. since the distribution of the Poisson process is invariant under time reversion.O<t<T). 0 < t < T) P(STEA. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence.
50 CHAPTER II. But since St 4 oo a.2) in the last). and since the jump rate is /3. oo).y) (here we used the fact that the probability of a jump at u t is zero in the second step.. for A C (0.t dT..3 where the bold lines correspond to minimal values.r.St). SOME GENERAL TOOLS AND RESULTS Integrating w. That is. U + St_ E A. and (6.St _). G+(A) = Q f 0 B(A . s. Fig.St _)I(r+ > t). E A} precisely when r+ > t. 6.T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .3 G+ is the restriction of /3R+*B to (0. . it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time . Figure 6. T+ > t] 0 _ /3 f E[B( A . we get G+ (A) = f 00 /3 dt E[B(A . 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. oo). The probability of this given { Su}u<t is B(A . cf.3 Lemma 6 .
assuming basically stationarity in time and space. i.. cf. the first component representing time (the arrival time o.. i. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup.. as a point process on [0.6. In the stationary case.3 yields g+ (x) = . {St+8 .e.z)B(dz) _ f I(x < z)B(dz) _ f (x).:T1 +•••+Tk <t}. 6 . oo).s. oo) x (0. 6 . ..4). Nt St =>Uk k=1 t where Nt = max{k = O. The marked point process . 2.4) are (ak. 4 (the points in the plane are (ak . The points in the plane (marked by x on Fig.S8 )t> o = {St }t>o for all s > 0. Uk) (k = 1. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. Uk) for those k for which ak . this is equivalent to the risk process {St*} being stationary in the sense of (6.M o 08 shifted by s is defined the obvious way. . We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. 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.Q f r+(x . Fig. 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. U k)} k=1 a is as a marked point process M *. this does not depend on h). Lemma 6..1.) where ak = Ti + • • • + Tk . obviously. The traditional representation of the input sequence {(TT.T+ < oo).1 With r+(y) = I(y < 0) the density of R+. 0 Generalizing the setup.e.* ) and the second the mark (the claim size Uk ). h]} /h (by stationarity. we define the arrival rate as E# { k : ak E [0 .s > 0).
where TI = 0. = 0 . i. h] and the sum approximately ^o(M*)I(ul < h). SOME GENERAL TOOLS AND RESULTS M* U.. Section 5) which has pure jump structure corresponding to pi = a = 0.e. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. e. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 .52 CHAPTER II. h] Eco(M*) = 1 E f co(M o Bt)dt. letting h J.. We represent M by the sequence (Tk. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. The two fundamental formulas connecting M* and M are Eco(M) = aE E. Example 6 . Oh becomes the approximate probability F(ri < h) of an arrival in [0.5) does not depend on h.g. Sigman [348] for these and further aspects of Palm theory.s. See. As above . the r.. . k: vk E [0. most often one takes h = 1).. vi(dx) = . Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. of (6. o. This more or less gives a proof that indeed (6.5) represents the conditional distribution of M* given vi = 0. Uk) k=1.2. V(M* o eak ). and let T = T2 denote the first proper interarrival time .4 Consider a finite Markov additive process (cf.QiBi(dx).4 Given a stationary marked point process M*. h. 0.
we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i. Assume that St * . j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise.O for i # j.OF(x). If Jt_ = i. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J. let U0 be a r. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . 5 Consider a general stationary claim surplus process {St }t>o. an arrival for M* occurs before time t + dt w. the ruin probability .6 Under the assumptions of Theorem 6. aij for (i.OEU0. and by some additional arrivals which occur w. and that p = 0EU0 < 1. After that.oo a. First choose (Jo_.O fo "o F(x)dx = .6.p.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 .= i. It follows that we can describe the Palm version M as follows . Jt = j is iri(3i /. Note in particular that the Palm distribution of the mark size (i. we note: Corollary 6. Before giving the proof.e.p.*(0) with initial reserve u = 0 is p = /3EU0.s. let the arrivals and their marks be generated by {Jt} starting from Jo = j. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution .5.. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = .O for i = j and iriAijgij/. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. qij when {Jt} jumps from i to j and have mark distribution Bij. 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 . . A stationary marked point process M* is obtained by assigning Jo distribution Tr./. the probability aij of Jt . v. Jo) w.p.
g.0<u<tIAt) = P(St EA. h.(left limit) when 0 < it < t and is illustrated on Fig .. . 6. in (oo... Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt.Su_ <0..0<u<tIAt) = P(St EA. . oo ) and the arrival times 0 > 0_1 > a_2 > .A. . Let p(t) be the conditional probability that ST+ E A.$St_ u. has a very simple interpretation as the average amount of claims received per unit time . the mark at time Qk is denoted by Uk. oo)).)..l.Q_k and has size U_ k.Mt). 105) shows that one can assume w.St <. oo) p(t) = P(St EA. that M* and M have doubly infinite time (i.o. The last property is referred to as insensitivity in the applied probability literature. and the kth preceding claim arrives at time t .. in (0. The sample path relation between { Su } and { Su } amounts to S„ = St . oo) x (0 .0<u<t) = P(St EA. It follows that for A C (0. moves down linearly at a unit rate in between jumps and starts from S0 = U.0<u<t) = P(StEA. (k = St}t>o 1.54 By (6. Proof of Theorem 6.1] here the r ..o.5. Now conditionally upon At .Su< 0. { Su}0<u<t is distributed as a process {Su} .s. the arrival times 0 < 0'1 < Q2 < . 0).5.St*_ u.o<u<t where a claim arrives at time t and has size Uo. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0. T+ = t given the event At that an arrival at t occurs . 2.St<Su.e.. We then represent M by the mark (claim size ) Uo of the arrival at time 0.5). 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.Su<0. A standard argument for stationary processes ([78] p. which makes an upwards jump at time . are point processes on (oo . CHAPTER H. 0<u<t) = P(St EA.
NIt)dt . THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. cf. and since by assumption St * oo a. Mt)dt = i3EL(A) o"o .6. Thus. t a oo. the support of L has right endpoint U0. 0 < u < t } is the event that { Su } has a relative minimum at t . In Fig.s.5 where it = { St < Su. Uo]. Fig. and we let L(dy) be the random measure L(A) = fo°° I(St E A.5. 6. G' (A) = 3 f P(St E A. A sample path inspection just as in the proof of Lemma 6 . the left endpoint of the support is oo. 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. Since So = U0..
56 CHAPTER II.1). [263] (a special case of the result appear in Proposition VI.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. [147]. .2.6 is Bjork & Grandell [67]. A further relevant reference related to Corollary 6.
• the premium rate is p = 1.d. 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 . Some possibilities are numerical Laplace transform inversion via Corollary 3.4 below . with common distribution B.. 4. exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. For finite horizon ruin probabilities . see Chapter IV. St = uRt = EUi t. 3). say. and assume that • { Nt}t>o is a Poisson process with rate j3. Panjer's recursion ( Corollary XI. and independent of {Nt}.e. • the claim sizes U1.6) and simulation methods ( Chapter X).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. are i.. Thus . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . U2. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. i. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. . i. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 .
m. 0 . and that B(k)[0] = Pak). A more formal proof goes as follows: Nt r Nt ESt = E > U k .6pBa).g.)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. of the claim surplus St .f. [324].58 CHAPTER III.t = E[Ntµs] . cumulants . Schmidli [319].s.1)..t = t(p . and this immediately yields (a).+Uk)P(Nt = k) k=O e8t k=O B[s]k .g. (c) Ee8St = et" (8) where c(s) = f3(B[s] . 1 Introduction For later reference. and Schlegel [316]. (d) The kth cumulant of St is tf3p(k) for k > 2.1) = t(p . See e. we shall start by giving the basic formulas for moments.1) .Rt. where K(k) (0) is the kth derivative of is at 0. We do not spell out in detail such generalizations. For (c). for (d) just note that the kth cumulant of St is tic(k) (0). THE COMPOUND POISSON MODEL Poisson risk process. we get Ee8st = 00 e8t c` Ee8 (U1+. {Bt} a Brownian motion and {Rt} a pure jump process. Proof It was noted in Chapter I that p .1 (a) ESt = t(13µ$ . Furrer [150].'s etc.t = E E [ U k k=1 k=1 Nt .1 is the expected claim surplus per unit time. Write pB^) = EUn' YB = Pali = EU. e . P = PAB = 1/(1 + rl) Proposition 1. (b) Var St = t.t = fltpB . say stable Levy motion. Dufresne & Gerber [126]. 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).1)..u .
d. u + v].. meaning that the increments are stationary and independent. then St 4 co. then Snh .h < St < S(n+1)h + h. 2.3) is proved similarly. II.S„ attains its minimal value when there are no arrivals in (u.Tk are i. obviously 0(u) = F(maxk Sok > u). lim supt. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. St = oo. Obviously.1 = . . For the proof.V. which is often used in the literature for obtaining information about {St} and the ruin probabilities.1)th claim. we need the following lemma: Lemma 1.1. For example. Indeed.. then lien inft. rather to view {St} directly as a random walk in continuous time. The point of view in the present chapter is. the Uk . and the value is then precisely v. In particular. cf. where Tk is the time between the kth and the (k . S„+V > S„ . INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. Here is one immediate application: Proposition 1. (b) If 77 < 0.Tk. v > 0. so that {Sok } is a random walk with mean EUET = EU. if t = nh + v with 0 < v < h. we have Sok . The right hand inequality in (1. then St> SnhV>Snhh...Sok_l = Uk .3EU01 = 1µs where rt is the safety loading. Sn+0 . Proof We first note that for u. (c) If 77 > 0. (a) No matter the value of 77. we get a discrete time random walk imbedded in the claim surplus process {St}.1 is the same as if {St} was a random walk indexed by t = 0. (d) If 17 = 0. We return to this approach in Chapter V.4. In this way.3 If nh < t < (n + 1)h.i. 1. then St 00. however. The connections to random walks are in fact fundamental.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. St = oo.
lim supn_.60 CHAPTER III. is a discretetime random walk for any h > 0.1. If rl > 0. However.4 The ruin probability 0(u) is 1 for all u when 77 < 0.2. and < 1 for all u when 77 > 0. hence by induction i. p. Remark 1 .3. it is seen that upcrossing occurs at least twice.o.1(b)) that the assertion holds as t 4 oo through values of the form t = 0. it suffices to prove 4'(0) = F(M > 0) < 1. 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. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time.. where the size of the portfolio at time t is M(t). at least once...t . THE COMPOUND POISSON MODEL Proof of Proposition 1.2.2: 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). For any fixed h. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1._. There is also a central limit version of Proposition 1. The general case now follows either by another easy application of Lemma 1. {Snh}n=o. 0 Snh = 00. u 307). {Snh}n=o.p. Considering the next downcrossing (which occurs w. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1.5 The limiting distribution of St .6 Often it is of interest to consider size fluctuations. Corollary 1. Snh/n a4' ESh = h(p . This contradicts u St400. h. or by a general result on discrete skeletons ([APQ] p..... 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 .1). (c) are immediate consequences of (a). Part (d) follows by a (slightly more intricate) general random walk result ([APQ]. then {St} upcrosses level 0 a. 2h. h A similar argument for lim sup proves (a).. is a discrete time random walk. and hence by the strong law of large numbers. . 169) stating that lim infra.1.. if P(M > 0) = 1. and (b). 1 since St 4 oo) and repeating the argument. Thus using Lemma 1.s.. Snh u = 00 (the lemma is not needed for (d)).1. Notes and references All material of the present section is standard.3.
1) is not entirely satisfying because of the infinite sum of convolution powers.e. The expression for g+ was proved in Theorem 11. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o).1. the formula for the distribution of M follows . equivalently. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. Fig. 11. cf. 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}.6.1 provides a representation formula for 0(u).1) representing the distribution of M as a geometric compound. which we henceforth refer to as the PollaczeckKhinchine formula. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). As a vehicle for computing tIi(u). It is crucial to note that for the compound Poisson model.1.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ).6. Summing over n.34 or A. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 .just before ruin is again B0. We assume throughout rl > 0 or. oo ).2. and we shall here exploit the decomposition of the maximum M as sum of ladder heights.. the ladder heights are i. (2.IIG +II)EG+ . nevertheless.P) E PnBon(u) . and we further get information about the joint conditional distribution of the surplus and the deficit. d. Thus . 1e. B(x)/aB.. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. p < 1. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . but we shall be able to extract substantial information from the formula. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. Combined with i/i(u) = P ( M > u). i. This follows simply by noting that the process repeats itself after reaching a relative maximum. [APQ] Ch. Theorem 2. 1 The distribution of M is (1. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . n=0 (2. 0 Alternatively. Note that this . IV.
cf.2(a) is from Dufresne & Gerber [125]. ST(o)) is given by the following four equivalent statements: B(z) dz. 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). W are independent. where it requires slightly more calculation. In the risk theory literature.d. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„).6.6. (c) the marginal distribution of ST(o)_ is Bo . 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. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. 2 The joint distribution of (ST(o )_. V is uniform on (0. cf. f +b (b) the joint distribution of (ST( o). 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. ST(o )) given r (0) < oo is the same as the distribution of (VW.V)W) where V. (1 . Asmussen & Schmidt [49].62 CHAPTER III. (a) 11 (ST(o)_ > x. (d) the marginal distribution of ST(o)_ is B0. see for example [APQ]. The proof of Theorem 11. Theorem 2. As shown in Theorem 11. there is a general marked point process version.5.2 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.i. . Again. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. Beekman [61]. However. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. [62]. ST(o) > y. in this setting there is no decomposition of M as a sum of i. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0.1 is traditionally carried out for the imbedded discrete time random walk.5. Theorem A1. see Schmidli [323] and references there. Feller [143] or Wolff [384]. We assume rt > 0 throughout. Theorem 2 .just after ruin. cf.
however . also be seen probabilistically without summing infinite series .p.O)e(b0)x. 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. 0 . Alternatively.p)pSe a ( l v)x = p( S .p) = S . Integrating from u to oo.3.e. As shown in 11. use Laplace transforms. The result can. 3b Exponential claims Corollary 3. For a failure at x. B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations .6. then. a further relevant reference is Bjork & Grandell [67]..1)1 00 ( 1 . Thus . Thus r(x) = S(1 . the formula for P(O) holds in a more general setting. the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w.2 If B is exponential with rate S. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 . hence without memory.1 e ax = n1 (n . 1 .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. the result follows .p. Let r ( x) be the failure rate of M at x > 0.p) E pn S n x n. I.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. and hence this overshoot has the same distribution as the claims themselves . But claims are exponential .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.
cf.+ <u. A variety of proofs are available .1.y)G+(dy) For the last identity in (3. then 24 1 V.g. (Example VIII.s. 2 is one of the main classical early results in the area. and weights 1/2 for each. (3.y)/3B (y) dy. 0 Proof Write o (u) as P(M>u) = P(S.T+ <oo). (3. (3.2) Notes and references Corollary 3. The case of (3. u . and conditioning upon S. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.+ >u.3.y)f3 (y) dy.S. (b) use stopped martingales . II. Then the first term on the r. T+ <00) (3.3) Equivalently.S. the survival probability Z(u) = 1 .4) zu P(M > u .3) below. Corollary 3.i(u) satisfies the defective renewal equation Z(u) = 1 .+ <U.1) For a heavytailed B.3)).1 p pBo(u). We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .3. is ?7+ ( u). we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype.h.4) is similar (equivalently. THE COMPOUND POISSON MODEL In VIII. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).T+ <oo)+P(M> u. (3.64 CHAPTER III.2).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 .4) can be derived by elementary algebra from (3. (u) 35eu + 35e6u.p + f u Z(u .+ = y yields P(M>u.y)G+(dy ) = f U V(u . u + oo. E. just insert the explicit form of G+.3).p + G+ * Z(u) = 1 .
4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. Griibel & Pitts [132]./3B[s] . 206207).p)2 3(1 . 111112 or Feller [143].P)PB 2(1 . it is not surprising that such arguments are more cumbersome since the ladder height representation is not used. see e.4) can be derived by elementary but tedious manipulations.5) Proof We first find the m.Ps s(.3 .5). for example. Also (3. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3. [APQ] pp. (3. g. Bo of B0 as m e8u B(du) = B[s] .. 191). Embrechts. by analytical manipulations (L'Hospital's rule) from (3. Corollary 3.p)s s /3 .f. The approach there is to condition upon the first claim occuring at time t and having size x . 0 Notes and references Corollary 3. We omit the details (see. In view of (3. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3.pBo[s] no (1 .P)pB' (3.3 is standard . e.8) Proof This can be shown.5 The first two moments of M are 2 EM .p) E p"Bo[s]" = 1 .p)s .5 can be found in virtually any queueing book.7) s +.(3 .6) 00 = (I .Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities. Some relevant references are Abate & Whitt [2]. In fact.7).)3B[s]) (3.5). [APQ] pp.(3B[s] 1 .7) and Corollary 3.3.s .s . .3 . Of course./3B[s] which is the same as (3.g.g.. eau B(u) du = f PB 3PB SPB 0 o (3.p = (1 .1 Bo[s] = f oc.PPB2) EM2 = PPB) + QZPBl 2(1 . Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p.
u) a)Qea" + (1 .4) for Z( u) means f lhu Z(u) = 1.u)]k1 ku+1) [/3( k .1)! k=1 u1 .h.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.u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.)3(1 .9) follows for 0 < u < 1.u) [p(k . Assume (3.u + 1 )]k = QZ(u) .1 < u < n and let Z(u ) denote the r.Q) k=0 k! E e0( = /32(u) .3I( 0<y<1)dy Z(y)/3I(0<u.1).u)]k d 1 u) _ a) n ( du ( k! (1  .6 If B is degenerate at p.s.9).Q (k 1 k= n  [O(k . differentiation yields Z(u) _ /3Z(u) .3+ 18+ J0 Z(uy). THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3. eO('u) [)3(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. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .Q) 3e.66 CHAPTER III./3Z(u .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k . then p) 1: ep(k u/. Z^ =eR(k. For n < u < n + 1.u) [N(k . of (3.9) shown for n . .1).u)]k k! k0 The renewal equation (3./32(u .z/'(u) takes the form Z(u) L^J L.
it follows that Z(u) = 2(u) for n<u<n+1. Formalizing this for the purpose of studying the whole process {St}.3B = . (4.4.g. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. We could first tentatively consider the claim surplus X = St for a single t. F and c. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.2). K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. say t = 1: recall from Proposition 1. The answer is yes: inserting in (4.2) shows that the solution is Ox [O]0].g. 4 Change of measure via exponential families If X is a random variable with c. but will now be repeated for the sake of selfcontainedness. and define rce by (4. we just have to multiply (4. 0 Notes and references Corollary 3.Qe(Bo[a] .f.g.(9) is welldefined.1) or equivalently. The question then naturally arises whether ie is the c. or equivalently BB[a] = B[^+ Repeating for t 54 1.f.1) .3B[9].2) (Here 9 is any such number such that r. (4. B9(dx) = B[9] B(dx).4) works as well. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). in terms of the c. (4.(9).4) . 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) .f.4.3) by t. we set up .a.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11.r.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. and thus (4.d.1) .a.rc(9) = . co(a) = rc(a + 9) .1 that c(a) = /3(B[a] . of F9.f.
7) Proof We must prove that if Z is FTmeasurable. Xn).. v(Xi.1 Let P be the probability measure on D[0.r. and define 09. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .d.r.FTn) = Q(SkTIn : k = 0. n) for a given n.Tic (0)} . .7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 . = exp {BST . Ti(a)/n.t.3 Let T be any stopping time and let G E FT.g. (4. Z is measurable w. BB by (4. t < T. .8) By standard measure theory.t. for G E FT. and thus (4.5) for the density of n i. and PBT) the restriction of PB to FT.3 and claim size distribution B. .8) follows by discrete exponential family theory. (4. But let Xk = SkT/n . G]. G C {T < oo}. Then the Xk are i. and dP(T) dP^T) That is.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} . oo) governing a given compound Poisson risk process with arrival intensity. the PBT) are mutually equivalent on.f. in particular the expression (4. (4. it suffices to consider the case where Z is measurable w. Proposition 4.. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity. Let FT = o(St : t < T) denote the o•algebra spanned by the St. .2.2 For any fixed T.10) . the corresponding expectation operator is E9.nr.S(k_1)Tln. The following result (Proposition 4. then EBZ = E [Ze9ST _T"(9)I . (4. EeeBSt + tk(B) = 1. with common c.0e and claim size distribution Be.i.i.4).5) for the density.9) Proof We first note that for any fixed t.1) and multiply from 1 to n). G]. The identity (4. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} .68 CHAPTER III. . THE COMPOUND POISSON MODEL Definition 4.d. replications from Fe (replace x by xi in (4.(9)} (4.FT.. . (4.1.
7) holds.1) _ 1 + a. GT C_ Jr < T}. 5 Lundberg conjugation Being a c.7 1 Some discussion further supporting this statement is given in the next section. Given FT.1) .ST) + (T .g. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. c(a) is a convex function of a. The behaviour at zero is given by the first order Taylor expansion c(a) r.. and hence (4. t = T .1 It is seen that typically) a ry > 0 satisfying 0 = r.FT]] = EB [exp { BST + Trc(9)} I(G)] . 5. Now consider a general G. Ee [exp { BST +Trc(9)} I(G) FT)] = 1. 77 Thus. Thus. Thus by (4.(Y) = 13(B['Y] . LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.9) holds with G replaced by GT. Letting T t oo and using monotone convergence then shows u that (4. .f. Then G E FT. the typical shape of rc is as in Fig.r is deterministic.10). (4.r)rc(9)}I . according to what has just been proved. subject to the basic assumption ij > 0 of a positive safety loading. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST .9) holds for G as well.1(a).5. (0) + rc'(0)a = 0 + ES1 a = a (p . Then GT = G n Jr < T} satisfies GT E FT.
QL instead of /37 and so on in the following . Thus B[7] = 6/. Example 5 . and (4.2 is B(7) = 1 + ^. we write FL instead of F7. (5. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes. (5. 5. an equivalent version illustrated in Fig.3.1) (or (5. Thus. Fig. 5.s). 5.4) yields /3L = b and that BL is again exponential with rate bL =. Note that KL (a) = /L (BL [a] .1(b).1) is known as the Lundberg equation and plays a fundamental role in risk theory . e.3) cf.2)) is 7 = 5/3.1) .a = i(a + 7). An established terminology is to call y the adjustment coefficient but there are various alternatives around. THE COMPOUND POISSON MODEL exists . u It is a crucial fact that when governed by FL. we further note that ( 5. (5. It is then readily seen that the nonzero solution of (5. the claim surplus process has positive drift > 0. .4) ELS1 = #L(0) cf.1(b).2) 7 Figure 5.g.1 Consider the case of exponential claims.1) is precisely what is needed for one of the terms in the exponent .70 CHAPTER III. Fig. b[s] = 5/(b .3. G = {T(u) < oo} in Theorem 4.3. the Lundberg exponent. Taking T = r(u).2 s As support for memory. Equation (5.
LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 .u be the overshoot and noting that PL(T(u) < oo) = 1 by (5. Proof Just note that e(u) > 0 in (5. take first 0 = ry.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.1p .1 (5. To this end.P Y j o' xeryxOB (x) dx /3k [Y] .5) Theorem 5 . (5.(oo) (in the sense of weak convergence w. V)(u) < e7u.ascending ladder height distribution and µ+ its mean.(u)} . which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.+ E A} in Theorem 4.3.6 ). (5. PL ) with density 1 . Then P(ST+ E A) = EL [exp { 7S?+} . ST+ E A] .5).r.e7x)G+(dx).1e. (5.3 takes a particular simple form.8) . where C .t. T(u) < oo] . T = T+. Letting e(u) = ST(u) .5.7) 0 and all that is needed to check is that ( 5.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .2 (LUNDBERG'S INEQUALITY) For all u > 0. G = {S. Since a7' is continuous and bounded. see A . we can rewrite this as 0(u) = e"ELe7^(u).Ce7u as u 4 oo. V) (u) = P(T(u) < oo) = EL [exp {ryS.4).7) is the same as (5.1. 0 Theorem 5 . e(u) has a limit i.6) Proof By renewal theory.G+L)(x)) dx ry^+L) J 00 f 0 (1 .
1 above) and that C = p. but some tedious (though elementary) calculations remain to bring the expressions on a final form. (5.4 Consider first the exponential case b(x) = Seax.4).11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.7). From this it follows.10) VW = JI c* e° (x) dx = a (B[a] . this solves the problem of evaluating (5. of course.1) (5. Then 0(u) = pe(a_Q)u where p = /3/S.")G + (dx ) = 1  J0 00 3B(x) dx = 1p. Using (5. Noting that SIG(L)II = 1 because of (5.8) yields +L) J0 xel'B ( x) dx (5.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one. 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. that 7 = S .72 CHAPTER III.e. we get L where 00 (1 .1 . u . THE COMPOUND POISSON MODEL In principle. .3 (this was found already in Example 5. or equivalently of how close the safety loading 77 is to zero.12) Example 5 .
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)
.
p* _ .1.Ps(/3max . Notes and references The approximation (7.3 )1 } _ 1p 1 ./3)PBo PB .(3)2 P PB 2µB 2µB Letting /3* = /3/P. [174]) shows that it may produce surprisingly good results.p + p { 1 1p ti 1 .b(u) = p*e. we shall represent this situation with a limit where /3 T fl but B is fixed. and hence the ruin probability approximation is b(u) e(bAln)u.3* /S. Proposition 7.80 CHAPTER III. the approximating risk process has ruin probability z. we have according to the PollaczeckKhinchine formula in the form (3. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. but has an obvious interpretation also in risk theory: on the average.p . Proposition 1. Letting Bo be the stationary excess life distribution.(bA*)".)3 )PBo µB  .2) was suggested by De Vylder [109]. (/3max .P .1 As /3 f Nmax. 1924./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 .g. the premiums exceed only slightly the expected claims. Though of course it is based upon purely empirical grounds. Grandell [171] pp.3 and Corollary 3.1. Mathematically.s(/3max . That is.7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 .PBo [s (/3max .p = (/3max 0)µB. cf. numerical evidence (e./3)] 1 . THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 .8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . heavy traffic conditions mean that the safety loading q is positive but small. or equivalently that /3 is only slightly smaller than /3max = 1/µ8.
/3)u * v. obviously Corollary 7. Of course. The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e .3) is reasonable for g being say 1020% and u being small or moderate. the term light traffic comes from queueing theory. Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. the premiums are much larger than the expected claims . These results suggest the approximation Vi(u) e6(0_. 7d The light traffic approximation As for heavy traffic ./3)u).p _ 2rl11B PB p.1 1 .g. in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . VIII). light traffic is of some interest as a complement to heavy traffic . u * oo in such a way that (3max . This follows since rl = 1/p .ze a2unµB laB (7. the first results of heavy traffic type seem to be due to Hadwiger [184]. or equivalently that 0 is small compared to µB . We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . Mathematically.2 If . [APQ] Ch. as well as it is needed for the interpolation approximation to be studied in the next subsection. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. However . Numerical evidence shows that the fit of (7. That is . In the setting of risk theory. light traffic conditions mean that the safety loading rl is positive and large .B AB ) 6()3max _'3) = However .l3)M > (/3max .2.7.0)u. while the approximation may be far off for large u.Q T /3max. but has an obvious interpretation also in risk theory: on the average . . It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. 2 provides the better mathematical foundation. we shall represent this situation with a limit where 3 10 but B is fixed.p. then P(u) 4 e6„ Proof Write z'(u) as P((/3max . and hence 2µ2B 1 .4) suggested by the Cramer Lundberg approximation and Proposition 6.ryu .
e.T > u). ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. 0(u) /3 J B(x)dx = /3E[U .Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. (7. see Daley & Rolski [96]. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . ( 3 J O B dx. u Note that heuristically the light traffic approximation in Proposition 7. 10 ( u The alternative expressions in (7. cf. Indeed. the Poisson case is much easier than the renewal case.3 As . Omax max m. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx.82 CHAPTER III. Sigman [347]. For a more comprehensive treatment.5) u Proof According to the PollaczeckKhinchine formula. 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 . Again.u)+. Light traffic does not appear to have been studied in risk theory. by monotone time T of the first claim .3 is the same which comes out by saying that basically ruin can only occur at the F(U .= 1 aJ 1 a 0+ 1 = = p.u. .5) follow by integration by parts. z/' (u) convergence P(U . i. THE COMPOUND POISSON MODEL Proposition 7. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69].(3 10. Asmussen [19] and references there.T > u) = J o" B(x + u)/3eax dx . [97]. En'=2 • • • = O(/32) so that only the first terms matters. U > u] = /3iE(U .
The adaptation to risk theory is new. . available. z/i(E) (u) = pe(QmaxQ)u. that is. [84]. we see that the following limits HT) (u'). where further references can be found .6) is . Substituting v = u(. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless .VHT) ( ax QmQ ) h (B) ( . ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). Let OLT) (u) denote the light traffic approximation given by Proposition 7. Al . 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. . with rate 1/µB = /3max.O(E)(u) 1 (1 . Another main queueing paper is Whitt [380]. 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. (7. no empirical study of the fit of (7.3n. _(E) (u). Instead.3 and use similar notation for %(B) (u) = (u). Notes and references In the queueing setting . 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)./3)) .3).8. Thus .Wmax f(x ) dx + pee6mQ. ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. ) M. B(2).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. however. even if the safety loading is not very small.O0 M. (U). 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.x . the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. to get nondegenerate limits . "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz .
ill(u) < V)(2) (U) for all u. 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). U(2) distribution B(2) and U(1) < U(2) a. Proposition 8. we have the convex ordering. XI. . B(') <i. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. then Bill = B(2). Proof According to the above characterization of stochastical ordering.6. Proposition 8. then i. In terms of the time to ruin. whereas (consider x2) B(2) has the larger variance. we shall need various ordering properties of distributions.2 If B(') <j. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. 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. THE COMPOUND POISSON MODEL To this end. this ordering measures difference in variability. U(2) such that U(l) has distribution B('). Rather than measuring difference in size. the proof is complete. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. In the literature on risk theory.84 CHAPTER III.s. Taking probabilities. 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]. Here convex ordering is useful: Proposition 8. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. cf. Bill is said to be convexly smaller than B(2) (in symbols.' 1)(u) < V)(2) (U) for all u. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f.1 If B(') <d B(2). u Of course. Finally. or the existence of random variables U(l). this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. B(' <. then .1 is quite weak. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. B(2) and PB(1) = µB(2). we can assume that 1) < St 2l for all t.
and here is one more result of the same flavor: Corollary 8.3 If B(1) <. (D) (u) < O(B) (U ) for all u. The problem is to specify what 'variation' means. Corollary 8. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. Hence by the PollaczeckKhinchine formula .1. Bo1) <_d Bo2) which implies the same order relation for all convolution powers. then /'(')(u) < 0(2)(u) for all u. we have by Jensen 's inequality that E f (U) > f ( EU). B(2).4 Let D refer to the distribution degenerate at 'LB .2 is the following: Proposition 8. A partial converse to Proposition 8.3 provides another instance of this.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. B(2). Proof If f is convex. say to p. Example 8. and consider the following claim size distributions: B1: the standard exponential distribution with density ay.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. .5 If '0(1)(u) < p(2) (U) for all u and a. with fixed mean.4) certainly supports this view: noting that.e.8. 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. A general picture that emerges from these results and numerical studies like in Example 8.1 and µB at 1 so that the safety loading 11 is 10%.6 Fix /3 at 1/1. B. then B(1) <. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal.(1) (. The heavy traffic approximation (7. This u implies that D <. A first attempt would of course be to identify 'variation' with variance. Then V. larger variance is paramount to larger second moment.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). Proof Consider the light traffic approximation in Proposition 7. from which the result immediately follows...
Let ua denote the a fractile of the ruin function. 11 Notes and references Further relevant references are Goovaerts et al. 9 Sensitivity estimates In a broad setting. in comparison to B2 the effect on the ua does not show before a = 0. 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.01%.4142. B. 0.lA. 1%. [166]. all distributions have mean 1. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.9A2e'2r where A.4. However. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable.1358. B3 the comparison is as expected from the intutition concerning the variability of these distributions. Pellerey [287] and (for the convex ordering) Makowski [ 252].01%.000. i. and consider a = 5%.0' U0.e. B4: the Pareto distribution with density 3/(1 + 2x)5/2.86 CHAPTER III. Note to make the figures comparable. 0. 1/)(u. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less.) = a.1%. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!).e'\1x + 0. A standard example from queueing theory is . Kluppelberg [234]. 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).001 u0. A2 = 3. We return to ordering of ruin probabilities in a special problem in VI. and this is presumably a consequence of a heavier tail rather than larger variance. the behaviour of which is governed by a parameter 9. = 0.. For B1i B2. One then obtains the following table: U005 U0. B3: the Erlang distribution with density 4xe2x. van Heerwarden [189]. In terms of variances o2.
where Q2 = fl ( l2 1113 / _ Ou2v)2. Then a p ao = 00 Qa/. increasing in u. with 0 the vector of service rates at different nodes and routing probabilities. say estimated from data.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. where the partial derivatives are evaluated at p = 1. Assume for example that 8 is known. the premium rate p and the claim size distribution B. i. Let R(P) = Rtli.e.2 Consider a risk process { Rt} with a general premium rate p. Similar conclusions will be found below. it follows that ' is approximatively normal N(0. and s(9) the expected sojourn time of a customer in the network.(u) for large u.01/2u. a/3 0 . obtained say in the natural way as the empirical arrival rate Nt/t in [0. u Proposition 9. 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. SENSITIVITY ESTIMATES 87 a queueing network. the distribution of %3 0 is approximatively normal N(0„ Q/t). Then ib = Pe(613)u.1. a2/t). or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. For example.. t]. In the present setting. In particular .19P a/ . while /3 = j3 is an estimate.Ap). Proof This is an easy time transformation argument in a similar way as in Proposition 1. Example 9.3. Then if t is large . if = a e(6A)u. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 .9. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . Thus. a0 as ao 80 19P . Thus at p = 1. we may be interested in a'/ap for assesing the effects of a small change in the premium. Then the arrival rate /3(P) for { R(P) } is )31p.
((dx ) = exp {Ox + (t(x) . namely that of a twoparameter exponential family of the form Bo. and the proofs of (9.w(O.t. (3+'y)PC (0+7.1 or Proposition 9.3.()YC = 1 +y/ /3 \ Q2 From this (9. various parametric families of claim size distributions could be considered.()wC(e. Viei '0(. Consider first the case of 8/8/3: . Consider first the adjustment coefficient y as function of 3.r. we cannot expect in general to find explicit expressions like in Example 9.^)] 1(/3+y)we(9+'y.5) are similar. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. 9.uypCe7u urypO. we can rewrite the Lundberg equation as w(9+ y.2) (see Remark 9.88 CHAPTER III. /3 yields w e(e + Y. so that heuristically we obtain '00 50ryu = Coe"u . this intuition is indeed correct. 4) (9 . (9.3 or/and B.w(6. (. Of course.()(0 +'0) ' (9 . x > 0 (9.g. In the case of the claim size distribution B. However . u Now consider the ruin probability 0 = 0 (u) itself.6) As will be seen below. THE COMPOUND POISSON MODEL As a consequence.0 = t/'(u) and the CramerLundberg constant C. Similar notation for partial derivatived are used below. e. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . it suffices to fix the premium at p = 1 and consider only the effects of changing .Owe (9. and write yp = 8y/8/3 and so on . for the ruin probabilities . 5) (Q+'Y)[we(0+7.(/3 + y)we(9 + 7. Differentiating w. but must look for approximations for the sensitivities 0.3 70 = 'Ye = = 7 /3(1 we(e +'y. ^) .10) below.4). () Proof According to (9. Proposition 9.6 below for some discussion of this assumption).3. (9. but we shall concentrate on a special structure covering a number of important cases. 3) ( 9 . ()} p(dx) .3) follows by straightforward algebra. () = log(1 + y//3).()^ 1 .
() . F(dx) = e'yy/3B(x)dx.8) (Section 5).g.12)).QB(x) dx.2 of the Appendix ). ()} . () .we(9 . Combining these estimates . we note the formulas Ee.([a] = exp {w(9 + a.x)B(x) dx + J U W(u .C).8) Letting cp = e0/e/3 and differentiating (9.8). we multiply by e7" and let Z(u) = elt" cp(u). 11 For the following. But from the proof of Theorem 5. ()] exp {w (O + y.x). z2(u) _ 1 ^ e'ri`i7i( u . () exp {w(9 + a. Z(u)/u a C//3PF where PF is the mean of F.x).w(9. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9.3(x) dx. Z= zl + z2 where zl (u) = e7u J m B(x)dx. By dominated convergence.9) (9.x)B(x)dx. we get p(u) = J "O B(x) dx + J U O(u .(e"U = = wS(O. 0(u) = /3 Ju"O B(x) dx + f 0 0(u .St (U) Ee.9. Be. (9.3) for z/'(u). ()} .w(O.x) F(dx ) f u J C F(dx) = C as u 4 oo. PF = (1 . z2(U) = e7" J u b(u . and alsoo zl(u) + 0 because of B['y ] < oo. () . Further write de = [we (9 +'y.4t (U)e°`U = which are wellknown and easy to show (see e.11) Ee. O} (9.4 As u oo. the proof is complete. w((9 + a. Hence by a variant of the key renewal theorem (Proposition A1.3 (see in particular (5. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.p)/C'y. u 0 Then Z = z + F * Z and F is a proper probability distribution .w(9. BarndorffNielsen [58]). () .10) (9. SENSITIVITY ESTIMATES Proposition 9.
01 (i+) do = +'Y.90 CHAPTER III. z = zl + z2.x)f3 f ^[t(y) . THE COMPOUND POISSON MODEL [we(e+7. )}B(dy)• Letting cp it thus follows from (9.wc(O.QB(x)dx. C)] (1 + 7 ) Proposition 9.wc (O.w( (0.w(e.6C do 89 1p 8( 1p Proof By straightforward differentiation. ()]B(dy) dx. 0 x Multiplying by e7" and letting Z(u) = e"uV(u). F(dx) = e7x. 8^ ue7u. ()} 1z(dy) = f [t(y) . ^) . ()]e7vB(dy) 'fCd 7 c .11). this implies Z = z + F * Z. oo z2 (u) f C .6e7u f "o f[t(y) .e7x/3 f 00 [t(y) .8) that cp(u) .x).w( (0.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . ()]B(dy) dx x 0 0C T ON O .9)(9. 8 8() 8( (9. u Z2(U) = e7° f u ^/i(u . By dominated convergence and (9.we (0.12) f exp {O y + (t(y) . ()](e7v .1) B(dy) 'f '[t(y) . Then as u > oo. ())B(dy) dx. 2 z 07P N ue7u (3C de . ^)} [wc (0 + 7.5 Assume that (9.w (9. C) .2) holds.wc(9.w(0.
16) (9. () = C/9 = a/S.15) (9. () = log r(a) .rye) S 5rya. that C = a.w((9.12) takes the form y) alp a . w(e.Y)a+1 ' (9. SENSITIVITY ESTIMATES as u 3 oo. Z(u) /3C 91 o c'o e11(t (y) ..C log(9). t(x) = logx.pa+1 .a log S = log r(c) . ue_Yu 'C2d( 8a 8( 1 p .QS 1 .17) (9.6 Consider the gamma density b (x ) = Sa xa. ( 9. by inserting in the above formulas. we (9. and the proof of the first one u is similar. 9 = S.1edz = 1 exp {Sx + a log x . Example 9.) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.ry) 5a1 cry (5 .2) holds with p(dx) = xldx.18) (05 + 57 _'3_y . We get w( (0./35' a/i'y + aryl 625ry.12) follows. .a/35a&y' ' (9.yu/3C2do u86 89 1p' az/) = 8z/. a /(S .1 .9.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 .. () ='I'(t. ())B(dy) < oo. U 7µF from which the second assertion of (9. < = a. Here (9.. It follows after some elementary calculus that p = a)3/5 and.Sry a/32 + a/37 + /37 . and also zj (u) 4 0 because of f Hence.3ary tog('Finally.(log r(a) a log S)} • r(a) 1.13) (9.
9 = .3Ee. which we omit in part .3.22. C) . Be. t(x) _ . w(e. Straightforward but tedious calculations .2) with µ(dx) = 2x3zrdx. for a < a* = z (. C = .2 log (0.log c = 2 In particular.1 = eXP {c(C . ()} = exp {c (C . C) = B = Yc = de = do = . further yield .2 . THE COMPOUND POISSON MODEL Example 9."62 .92 CHAPTER III.9) (() .21og 2.7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9.CZ try)} 1 C C2 try .1 16 +ry c C22ry 2( = + 70 We (e.S[a] = exp {w (9 + a. () = Cc .2a) } Thus the condition B[a*] > 1 + a* /.l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.([Y] .w(9.
the results presented here are new.2) is motivated as follows. That it is no restriction to assume k < 2 follows since if k > 2. for which we refer to X.10. BT [a]= NT ^` e"U. and hence explicit or asymptotic estimates are in general not possible. However . to our knowledge. Thus. then BT and hence ryT is undefined.8 The specific form of (9. if 1 PT = /3TNT(U1+.1) . Finally if k = 1. Comparatively less work seems to have been done in risk theory. the main tool is simulation. Thus. we can just fix k . let NT 16T = ^T .2 of the parameters. sj=1 and let YT be defined by IKT('ryT) = 0.. ae t 1lEY u S _ . in u which case the extension just described applies. Note that if NT = 0. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. To this end. In general.12) takes the form a = a 93 ar. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally.. Also. B are assumed to be completely unknown. queueing networks) are typically much more complicated than the one considered here. Van Wouve et al. we have assumed k = 2 and ti (x) = x. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). and we estimate y by means of the empirical solution ryT to the Lundberg equation. then ryT < 0. 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).cue_7u)3C P Remark 9. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . the exponent is either Ox.7 and references there. thus. [379] consider a special problem related to reinsurance.g. However. or Ct(x).. .oo. the models there (e. (9.3C2de 1p' z a = c .a. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3.+UNT) > 1. in which case we can just let t(x) = 0. kT (a) = /T (BT [a] .
2) follows from NT/T a4' . vfoVFB[2y].B[7]) 0+ Iv/o(b[y]. 16T where V1.1) .(27)/K'(7)2. More generally. 7T a4' 7.i3)(B[7] 1) + (3(BT[7]  . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . since NT /T . N ( n[7]. (10. THE COMPOUND POISSON MODEL Theorem 10 .a BT[7] I B[7] I + . For the proof.3T . we need a lemma.: N 0.1) 'YT ./^ B[27] . If furthermore B[27] < oo.7 + (.If .B[7]2 V2 . Lemma 10 . B[2'Y]  /3T ) . B [7]2 (10. B[27] .3/ T).'s.1) .2) rT(7) N N (0. V2 are independent N (0.Q and Anscombe 's theorem.: N ()3.B[7]2 }) ( T 0 .1 As T 4 oo.v.b[Yp'V21 T { (E[7] . .B[7]2 n Hence ( 10. it is easy to see that we can write \ V1 1 l _ . 1) r.94 CHAPTER III..B[7]) + B [7] .2 As T * oo.T y . then (10.)vl+ N CO. a2 where a2 = /3r.3) Proof Since Var(eryU) = we have B[7].'Y .1)2 + E[27] .
lcT(a ) 4 /c(a). Combining ( 10.4) and Lemma 10. 7T E (y .(ry + e) and hence KT(7 . NT i =1 n'(a) for all a so that for all sufficiently large T K7 . I. Theorem 10. where ryT is some point between ryT and ry. OT a 95 u 4 /3. BT[a] 3 B[a]. y + E) eventually. 6"Y (10. Let 0 < E < ry. To this end . Proof of Theorem 10.e. it follows that 7T7 KT(7T) . ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.(ry . If ryT E (7  we have KT(7 . By the law of large numbers.e) < 0 < r.'(y).'T(a) = 1 E Uie°U' a$' EUe "u = B'[a]. °7IT) .2. and the truth of this for all e > 0 implies ryT at 'y..Q. 0 are estimated from data . first note that e7TU N (e7U u2e27Uo'2/T) 7 . NT BT [a] Hence r.KT(7) kT(7) K'(7) .3).E) < 0 < kT(7 + E) for all sufficiently large T .4) + E).10.1 By the law of large numbers.e.c'(7) N (0' T (2(7) / N (0.1 can be used to obtain error bounds on the ruin probabilities when the parameters . Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).E) < 4T(7T) < 4T(7 + E). Then r.E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r.
it means 2 (8 . > 0 for some t E [Wn_ 1.. Letting Wo = 0.g. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. Notes and references Theorem 10. Griibel & Embrechts [292]..e. if B is exponential with rate 8 so that ry = 8 .f.3 or equivalently p > 1/2 or 11 < 100%. Mammitzsch [253] and Pitts. satisfies b(..5%). V. t]}. Embrechts & Mikosch [133].g.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. . Deheuvels & Steinebach [102]..i.. Hipp [196]. [197]. the nth busy cycle is then [Wn1.) = a (e.. 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 U1.0) < 5. Csorgo & Teugels [95]..d. 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. i.e. Herkenrath [192]. Asmussen [23]) can then be used to produce an estimate of ry. i .96 CHAPTER III.1 is from Grandell [170]. Wn). This approach in fact applies also for many models more general than the compound Poisson one. wn = inf{t > W.Wn) are i .Q. 6 < 2.info< „< t S. and the known fact that the Y„ = max Vt tE[W„1. with a tail of the form P(Y > y) . Vt = St . various alternatives have been developed.T VIT where r7ry. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. ft. U2. For example . Frees [146]. For this reason .1 : Vt = 0.96 if a = 2. = 1..ueryuU ".C1e"a ( see e.
Further let 'Yo be the unique point in (0.1) . generalizations to other models are either discussed in the Notes and References or in relevant chapters. 'y) where c(a) attains it minimum value. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. 97 .Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. 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. the premium rate is 1.g. 0. Unless otherwise stated.1 (the role of ryy will be explained in Section 4b). The notation is essentially as in Chapter III.1 where p = 13µB. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . the Poisson intensity is 0 and the claim size distribution is B with m. exists. The safety loading is q = 1/p . In particular.f. See Fig. B[•] and mean AB.s.
.1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. 1 FL.(. using that the overshoot l. 7.9). In particular.r.(U) < 00] = ELT(u)ke'YS. E[T(u) I T(u) < 00 ] = ELT (U).t.) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k.5 .1 The claims surplus is {St}. PROBABILITY OF RUIN IN FINITE TIME Figure 0. Var[T(u) I T(u) < 00] = VarL T( U) .u is the overshoot. . (u) is exponential with rate 0 w. 2 that E [T(u)k. 1 Exponential claims Proposition 1. 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 (1.98 CHAPTER IV. the time of ruin is T(u) and ^(u) = ST(t&) . By the likelihood identity III.2) Proof Let as in Example 111. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .1) (1. PL = 6/0 = 1/p > 1). we have for k = 1.(4. FL and independent of T(u).
1)ELT(u). 1).(PL .I (1. 0 Proposition 1.2) is aLELT( u) .h.1)T(u) are independent with QL the same mean . the 1.0) . we have by Wald's identity that (note that ELSt = t(pL . Let 0 > yo be determined by ^c(0 ) = a. EXPONENTIAL CLAIMS For (1 .1//32 (6/)3 1)2 26(/3u + 1)/(6 . Wald's second moment identity yields 2 EL (Sr(u) .h. u + ELe(u) _ PL .1.1)2VarLT(u) + 2 Ca 1I VarLT(u).s.6 + a)0 .s.1)T(u))2 = UL where = s.1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.3) 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) . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u). is V1rLSr( u) +VarL ((PL . T(u) < oo] fora > r.1)T(u)) = VarLe(u) + (PL . of (1.1 (6)3)2 which is the same as the r.12 Thus the l.s.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0. where = eBu I 1 ./3 .V/ is as asserted.6. which leads to the quadratic 02 + (/3 . Since Sr (u) and (PL .(yo) = 2V ."(ry) = 26//32.h.2)./3) . This means that /3(6/(6 .6a = 0 with solution 0 (the .B = a. .1)) ELST(u) ELT(u) (PL .
T(u) < oo] = e.9ST(u) +T(u)!c(0)} ..Y1 Y2 Figure 1. Y(u) belonging to a convolution semigroup .'s with rate 5. Fig. (1. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).T+ Ti a t U T I 1 a i F.v. the result follows. Using 5 = 6 .0.1 where Y1.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. are the ladder heights which form a terminating sequence of exponential r. .1 .3) we have E [e«T(u ). and M(u)+1 is the index of the ladder segment corresponding to T(u).. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment .100 CHAPTER IV. St Ti F. Note that it follows from Proposition 1. 1. More precisely. T(u) < oo] = EB [exp {aT(u ) ..4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. But by the fundamental likelihood ratio identity ( Theorem 111. Ti.3.4.3 that we can write EeaT( u) = eeuEe 017(o). T2 . Cf. are the lengths of of the ladder segments 2. Y2.v.. .....
I ex cos B cos j O dB fo " . 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 .T are conditionally i. Let {Qt} be the queue length process of the queue (number in system.i.T.k + 1).cos (u/..d. where U1.. 1).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. density xN lex/(N . If QT = N > 0.1(u. UN.0. EXPONENTIAL CLAIMS 101 For numerical purposes .T.1 )!.v.T is the residual service time of the customer being currently served and U2 . .. the following formula is convenient by allowing t. For j = 0.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . . . then VT = U1.T the service times of the customers awaiting service . and exponential with rate S = 1. Proof We use the formula . Then V(u. EN has an Erlang distribution with parameters (N. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . . UN.3 sin0 + 29) f3(0) = 1+/32/cos9.6(u) = Vfl/j l(Su. T.. 2.ST).1.3 Assume that claims are exponential with rate b = 1. T) to be evaluated by numerical integration: Proposition 1. let (cf. i. U2.e. cf..4. Corollary 11. Since U1 . Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V.T. ... the conditional distribution of VT given QT = N is that of EN where the r.T + • • • + UN. including the customer being currently served).T) 1 I fl(O)h(0) fdO where (1.1.6.i (u.
44).(31/2eie . similar formulas are in [APQ] pp.1 00 ok+lR 00 j=k1 +1)/2e .1)] L _112 /(k+1)/2 [.)3k +1 tj g'(QT >.1 R [. in particular equations (1.(31/2 cos (( k + 2)9) .i(k +1)e R [/3( klal/2e:0 (01 /2 e . f3(0) .38). PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j. and define tj = e(1+R)Taj/2Ij(2vT T). 912. let I _ j (x) = Ij (x).13(k +l)/2ei(k +1)9 R E .cos((k + 2)9)] d9.3(k +1)/ 2ei(k + l)6 (.cos (( k + 1)0)] f3(9) Hence the integral expression in (1.112 l 1( k +1)/2 [ 31/ 2 cos(kO) . (1.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) . 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .ie . Then (see Prabhu [294] pp. 8789) 00 E aj j= 00 = 1.102 CHAPTER IV.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 .8 ) yields F(QT > k + 1) .k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.31 /2eie L 1)] 1 I/31/2eie . k k2 + $k+1 E bj 00 t j .
oo (u)31/2e^e)k = )3k z cos(k9) = R k. 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]. . THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. the numerical examples in [12] are correct).e = e' COS a cos(uf31/2 sin 0). equivalently. or.T) which. T). however. expresses V)(0. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. F(x. Ui < x I / (note that P(St < x ) = F(x + t. going back to Cramer. E Fk. it follows as in (1. t) = P . there are several misprints in the formula there. k! k=O k0 i/z Co Uk ate" o'/z e .3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. is numerically unstable for large T. T) in terms of F(. however. u Notes and references Proposition 1. The first formula. 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).7) that _ [^ au ak+l (30 k L. 2 The ruin probability with no initial reserve In this section .2.. t )). Seal [327] gives a different numerical integration fomula for 1 . we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. from the accumulated claim distribution N. The rest of the proof is easy algebra.0(u.
T].0<w<t} St+v .(6. . meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. f T lStv)} 0<t<T by a 'cyclic translation'.1. T) = P(Tr(0) > T) = P(M(0. 2. we define a new claim surplus process St StM NJ Figure 2.T))dv.i. T]. resp.104 CHAPTER IV.(.T)) does not {Stv)} depend on v.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS.T))dv E^T I(M(v. Proof For any v E [0. v]. See Fig.(0. 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.b (0.1 In formulas. ") } is at a minimum at time t. co ).T)dx. [v. Stv^ _ Define M(v.T) T F(x.3) with A = (0.T)) 1 fT P(M(v.t)= {Stv) < SM. T T o where the second equality follows from II. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 . Then 1 .
T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .T) and Sv < 0 on M(0.T)f(I z /)(0. where the last equality follows from ST < St on M(0. then i fT I(M(v. If ST < 0. v < t < T} n M(0.Tt))f(u+t. Fig 2. It follows that if M(0 . It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0.T) occurs. For example. 0<t <Tv}n{ST<ST Sv+St T+v.. we can write M(v. v). there exist v such that M(v.Sv. T)) dv. we can take v E (w E.2. T) as {ST<St+ vS. Indeed. t) denote the density of F(•. letting w = inf It > 0 : St_ = mino<w<T Sw}. Obviously. w) for some small E. T T o i =1 Let f (•. or it occurs. v) = M(0. T.T) = F(u+T.2 10(u. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v.. T) occurs. T].t)dt. T)). We claim that if M(0. t). cf. 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 . T Theorem 2 . .xdx. then M(v. 0<t<v} = {ST < St .T)) dv f T I(M(0. ST > 0. v). this integral is 0 if STv) .T) occurs or not as long as ST < 0. T) = M(0.v<t<T} = {ST<StSv. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0.2. in which case there is a last time o where St downcrosses level u. T) occurs.ST on M (0. v<t<T}n{ST<STSv+St. v). Hence T TE f I( M(v.
T)+ J0 T (1V.t). 0 < t <T. O(T .T) = .p.b(u.u+dt]).2 Here o. E [t. which occurs w.T) = {St < 0. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2.2. C*(z. Hence P(ST<u) = 1 . {S t > z. ST_ _ z} . t + dt] occurs if and only if St E [u. u which is the same as the assertion of the theorem. which is independent of St and has the stationary excess distribution B0.2 . Proof of Theorem 111. Let Z be a r. For a fixed T > 0.106 CHAPTER IV. The following representation of T(0) will be used in the next section.v. Proposition 2. 2. 0 < t <T .2. ST_ _ z}. {St > .ST_ t_ and let A(z. The proof is combined with the proof of Theorem 111. ST_ _ z}.Tt))P(StE[u. 0 < t < T. define St = ST .z.T) = C(z. u + dt] and there is no upcrossing of level u after time t. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).3 Define r_ (z) = inf It > 0 : St = z}. z > 0.(0.
3). Hence integrating (2. T(0) < oo) = OR(z) dz in (2.1) that P(T(0) E [T. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. we therefore have P(A(z. Proof of Proposition 2. {St }o<t<T have the same distribution .T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. A(z. T + dT] I S7(o)_ E [z.2. z + dz].T))f3B(z) dz dT. r(0) < oo) = 3R(z) dz JP(C(z. It follows by division by P(ST(o)_ E [z.2.T).T + dT].3 But by sample path inspection (cf. z + dz]) = P(A(z.3. Fig. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz.1) yields P(ST(o)_ E [z. z + dz]. z + dz].1) z T . Figure 2.2. and since {St}o<t<T. u which is the assertion of Theorem 111. 2.ST(o) >y. Thus P(Sr(o)_>x.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. 7( 0) < oo) = P (C(z)) dT. .T) = C*(z. z + dz].T)) = P(Cx. ST(o)_ E [z.T)). (2.
some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. Note that T_ (y) < oo a. a martingale proof is in Delbaen & Haezendonck [103]. r(0) < oo. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. In the setting of general Levy processes.5a).T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z.1) . Tak'ecs [359]. because of77>0. Proposition 2.3 was noted by Asmussen & Kl(ippelberg [36]. cf. 2.1 Eear( y) = eyr(a).1. Proof Optional stopping of the martingale I er (a) 9 t. see in addition to Prabhu [293] also Seal [326].r(a).T+dT]).(3(B[r( a)] .5 and one upon excursion theory for Markov processes (see IX.6.(yo).1 and the present proof is in the spirit of Ballot theorems.3. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.s.2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . z + dz]. Theorem 2.2. (3. z + dz]. Let T_ (y) be defined as Proposition 2. I L Let ga(x) be the density of the measure E[ear(°).108 Hence CHAPTER IV. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. who instead of the present direct proof gave two arguments.1) where a > r. T(0) < oo) = dTP(T_(Z) E [T.c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). Lemma 3 . PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. Notes and references For Theorems 2. 3 Laplace transforms Throughout in this section. Lemma 3. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = . [329].
r(u) < oo). LAPLACE TRANSFORMS 109 Proof Let Z be the surplus . time T(u): u u Here is a classical result : the double m.r(a) b . y + dy]. the result follows after simple algebra.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .2 P(Z E [y.3. (Laplace transform) of the ruin Corollary 3.5 f 00 o a/r(a) .ga [b] 0 TO Using Lemma 3.(v) = ev''(a). (u .f.3.ic(b)/b x(b) + a eb"E[eaT(" ).x)ga (x) dx where za(u) = f.2. Corollary 3.g. E[ear (o) I T(0) < oo . b . Further by Theorem 111. Then by Proposition 2.r(a) = a [B[b] B[r(a)]] . Z = y] = EeaT. £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).x)(a) B(dy)• Lemma 3 .°° ga(x)dx. 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)) . It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.3.3 ga[b] = c(b) Proof + b + a .4 E[eaT (o).ga [b] 1 .ST(o)_ just before ruin .r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a. u . Hence eb"du E[eaT(").1] evr(a)B(dy)[ b .
Later results then deal with more precise and refined versions of this statement. By Proposition 111. note that by Wald's identity u + EC(u) = ES.1 Assume 77 > 0.mL > E T(u) < 00 ) 40.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant)..s. and take basically the form of approximations and inequalities. That is.1)Er(u) .3). the known results are even less explicit than for the exponential claims case. = (p .110 CHAPTER IV. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. and hence a. Then given r(u) < 00. cf.w ) v/. For the proof.2.1) i. T(u)/u mL as u + oo.1. Then as u * oo. St/t 1 1/m.00 St = lim . i. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) . we need the following auxiliary result: Proposition 4. Theorem 4 .r(u) = Er(u) • ES.mu D 2 4 N(0. for any c > 0 P( Further. (4. uoo u using e.6.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. for any m T(u) u . The first main result of the present section is that the value umL.e.s. (u) t. This proves the first assertion of (4. T(u) a.. For the second . P = /3µB > 1. Proposition A1.(u ) = o(u) a.2 Assume ri < 0. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model.h(u. mu ) ( 0 m < ML '(u) 1 m > rL. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 .
1).mL U > E.1 is standard. For (4 . T(u) < oo f / 00) e7uE L [e_7 (t1).N(0.1).3). Notes and references Theorem 4. note first that ( Proposition 111.1). this can be rewritten as u + 1(u) .7 6  11 Proof of Theorem 4. PL (•)+ 0.6µB2) Z v m (3µB2) Z. T) for T which are close to the critical value umL).1. the result comes out not only by the present direct proof but also from any of the results in the following subsections. Theorem 7.2 of [86]) and (4. of (4. 4). proving (4. again Proposition A1. implying T(u) . and as a timedependent version of the CramerLundberg approximation. Tu) T( u) .mu (2) '• m3/2 µB 7 .6. and (4.2) follows immediately from u (4.t/m D (2) 111 .1 The l. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. cf.4. apB ) .3. Thus.mu m . T (u) < 00 J 0(u) e7'PL U \ I T u) .1 (by considering 0(u. though it is not easy to attribute priority to any particular author.s. the same conclusion holds with t replaced by r(u).g.2.h. According to Anscombe' s theorem (e.^ N (o.mL >E By Proposition 4.5) St . 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. 1'r(U) .1) is T (u)  U mL P( T (u) < I > E. . If Z .r(u)/m T(u) ti µB2) Z. which may be viewed both as a refinement of Theorem 4.
(oo. resp .6) whenever f. e'°'/b (u. Using ( 4.4 (SIAM'S LEMMA) If 71 < 0. g are continuous and bounded on [0. we get E[ T (u) . we need the following auxiliary result: Proposition 4.e(u') oo w .VU T. then e(u) and r(u) are asymptotically independent in the sense that.5) For the proof.T ( u')] = E[ T ( ul /4 .a C4'(y )• ( 4.112 CHAPTER IV. Hence Ef (Vu )) 9 (T(u. one has 9 (r(u)_rnu) Ef (^(u)) * E.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u . Proof Define u' = u . P because of ^(u') .))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . using that ul/4 .^(T(u')).3). PROBABILITY OF RUIN IN FINITE TIME Corollary 4.w2) r.4). E9(Z) (4.r.t.6). oo).)mu \ h(oo)Eg (r(ul) . O . we can replace T(u) by r(u').L+YWLV'U) . Let h(u) = E f (^(u)).v.mul h(oo)Eg(Z).um. S( u ) < ul/4] < ET(ul / 4) = O(ul/4).T(u') given F. Then the distribution of T(u) . Then for any y.3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).f ( (oo)) .ST( u') = u1/4 . letting Z be a N(0.(u.l:(oo) (recall that rt < 0).ul/4. and similarly as above we get E[f(^(u)) I Fr(u. and thus in (4. oo ). Then h(u) 4 h(oo) = E f (6(oo)). with w2 as in (4.u1/4)I(S(u') > u1 /4) h(oo) + 0.
For refinements of Corollary 4. however . that for the fit of (4. y u) < e 7v" . For practical purposes . e7v" y < ^'(7) (4 . T(u) < umL + ywL f. Cf. see also von Bahr [55 ] and Gut [182]. yy by 1 K.7) to be good. 4b Gerber's time. 3 is due to Segerdahl [333].5 '(u .1. The present proof is basically that of Siegmund [342]. PL(T(u ) < umL + ywL) 113 4 C4(y). in practice one would trust (4.umL wI V"U u (4.7) to be valid is that T varies with u in such a way that y(T) has a limit in (.dependent version of Lundberg's inequality For y > 0. .3 ery"z/i(u .4. umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V ").T) Ce7"4 (T . Notes and references Corollary 4 . y > k'(7) .z/)(u . 0. ELe7E (") . Segerdahl 's result suggests the approximation b(u.5) and solve for y = y(T). The precise condition for (4.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). oo ) as u * oo. y u) < .3 in terms of Edgeworth expansions . define ay. 10) '5(u) .oo. u needs to be very large).yK(ay)• (4.9) ( 4 . where we used Stain's lemma in the third step and (4.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . see Asmussen [12] and Malinovskii [254].7) To arrive at this .(ay) = 17 7y = ay .4) in the last. Theorem 4. Thus . just substitute T = umL + ywL in (4. CL Fig. also Hoglund [204]. WHEN DOES RUIN OCCUR? Proof of Corollary 4.
the bound a7y° turns out to be rather crude .ay4(u)+ T(u)K(ay ). yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . a. In view of Theorem 4.5.8). yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). and hence t. if y > 1/ic'(y). Numerical comparisons are in Grandell [172 ]. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. which shows that the correct rate of decay of tp(u.r. yy is sometimes called the timedependent Lundberg exponent. 0. Hoglund [203] treats the renewal case.8 below . Then ic(ay) > 0 (see Fig . which may be understood from Theorem 4. However.h(u. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). dy) Notes and references Theorem 4 . u Differentiating w. and generalizations to more general models are given in Chapter VI.yu ) = eayuEav [e .1). f Some urther discussion is given in XI. yu) is e 'Yyu/ .v"U.8). who used a martingale argument. see MartinLM [257] . we arrive at the expression in (4.Y' (u. For a different proof.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. An easy combination with the proof of Theorem 111. 5 is due to Gerber [156 ].2.114 CHAPTER IV.6. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). we have rc(ay) < 0 and get (u) .7 i.b (u. From the proof it is seen that this amounts to that a should maximize ayic(a). the point is that we want to select an a which produces the largest possible exponent in the inequalities.t.3 yields easily the following sharpening of (4.9): Proposition 4. .
and b(u. the formula 0(u. and in case of ruin probabilities the approach leads to the following result: Theorem 4 . the choice of ay.4.ayC() . (4. and ii(u) .e. yu) = e.e.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 . then the relevant choice is precisely a = ay where y = T/u.yu) c ay . T(u) < yu] . we have ryas = ay .2 yields EaT(u) u u r. if we want EaT(u) . (4..: T. Using Lemma 111. [eT(u )K( ay). not inequalities. yu ) eaauEaye . yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma. it is instructive to reinspect the choice of the change of measure in the proof. u 4 oo. As a motivation.ay y 'Yay  ay . T(u) suggests heuristically that l t/.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00.ay a. We thereby obtain that T is 'in the center' of the Padistribution of T(u). This idea is precisely what characterizes the saddlepoint method.(u.6 with P replaced by Pay and FL by Pay. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations.i(u.c(&) = ic(ay) is < 0. For any a > yo.13) .ayuEay f eay^ ( u)+T(u)K(ay)..5. then ay > 0. Ea .^3 ]1/ Bay [lay .12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. Proposition 4. i. (4.8 If y < 1/ic'(ry).z. (0) r1 (a) ' I.11) ' If y > 1/ r . The traditional application of the saddlepoint method is to derive approximations..'(y ). then the solution &y < ay of .
a.ay) ay +. we get heuristically that Eay Ler (u)r(ay).116 CHAPTER IV.1)3 = (jB"[ay]l (Pay .ay ) r. and the equation ic'(a) = 1/y is easily seen to have . V < 01 Ir 00 er(ay)"1'2"'x eyur.13).c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.1.1) .3(5/(S .13).a) . it seems tempting to apply the normal approxiyu + ul/2wV. Example 4.ay)K(ay) ay ayI&YI For the second term in (4.(ay) _ y(ay . .1)3 = y3/3B"[ay]. a nr=.a)2 . T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV. 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) 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 . PROBABILITY OF RUIN IN FINITE TIME ry I i .1) under Pay mation (4. Writing r(u) and W2 = I3ay{. (4.13) rigorously.12) is 0 entirely similar.B[ay] /ay &y y(ay .4).l'B)y /(Pay .I ay &y a ^c'(ay) a (1 +.ay + ayl /BLay] .7ruw2 Inserting these estimates in (4.c'(a) _ /3a/(8 .1B[ay]1 ) y(ay . i B[7ay . The proof of (4. The difficulties in making the proof precise is in part to show (4. Then ic(a) = .(j (1 .9 Assume that B(x) = eay.11) follows. where V is normal(0.
is undefined for a > 5). DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .= (s. (5. 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 =..8 is from Arfwedson [9].5. A related result appears in BarndorffNielsen & Schmidli [59]./4 ^y for 1/i (u. 0 Notes and references Theorem 4. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.f. yu) when y < 1/ic'('y) = p/1 . then { __ . in discrete time: if p = ES. 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.ay)3 0 3/2 and (4.tcp) Lo {Wo ( t)}t>0 . is the drift and o. . and next to note that such an approximation in particular implies that the first passage probabilities are close. c a 00..1.1) .i )( v s vc ('3 + s _2 / .because the c.11) gives the expression '31/4 ( .. 2 = Var(Si ) the variance. The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.p.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 .g.
tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5. (5. where p is the critical premium rate APBTheorem 5 .e. a2 =/3µB2) Proof The first step is to note that { WC (St P) .7c). we have o {i!t s: . and consider the limit p j p. and this can be obtained under the assumption that the safety loading rt is small and positive.t} _ {W_1(t)} . this is an easy consequence of (5.tp). However.p/c < St(p) < S((n+l)/ c + Pp/c. Indeed . of which a particular case is the claim surplus process (see the proof of Theorem 5. cf.p. 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. 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. Lemma 111.1)) is inconvenient.a = Snp) and the inequalities Sn )C .1) with S. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes).1 below). Letting c = a2/pp. such that the claim size distribution B and the Poisson rate a are the same for all p (i. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p.118 CHAPTER IV.1 As p J. St = EN` U= .. 0 .1. n/c < t < (n + 1)/c.3) whenever c = cp f oo as p 1 p.3. + {Wo(t ) . for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5. p. We want an approximation of the claim surplus process itself.z } {W_1(t )}t>o (5.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. Mathematically..2) t>o where p = pp = p . oo)).
any probability measure concentrated on the continuous functions. . Corollary 5 .2 suggests the approximation u 0(u. is 1/ip (ua2 /IpI.u). and the r. C. [169] or [APQ] pp. 263) that the distribution IG(•. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x.s. w. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. For practical purposes .1 I 7= . we obtain formally the approximation V. ulpl /a2) = e2"1µl / or2. ('.e.T) IG(Tp2/ a2).h.f I \\\ J \ (5. 196.6) This is the same as the heavy traffic approximation derived in III.5) Note that letting T * oo in ( 5. is IG(T.t. (5. Corollary 5.s. (. 199. has a continuous distribution.Ta2 /p2). ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. we omit the details . the continuity argument above does not generalize immediately. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. u).( ^ I + e2( \ I .7c.. TS(u)=inf{t>0: WW(t)>u}.6) from Theorem 5. 119 It is wellknown (Corollary XI.(u) ti IG(oo. ulpI/a2).8 or [APQ] p. However. (ua2 To2 op \ IPI > IG ( T . 1. since ti(u) has infinite horizon . 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.5. Because of the direct argument in Chapter III.1 . see Grandell [ 168].. and in fact some additional arguments are needed to justify (5.4) Note that IG(.h.5).r.2 As p j p. u) =PIT( (u) < x) = 1 ..1.h.s. u) is defective when < 0.1. (5.
3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. as an example of such a generalization we mention the paper [129] by Emanuel et al. . The first application in risk theory is Iglehart [207].6) are presented. Assume further that 039µB6 < pe. in the next subsection we shall derive a refinement of (5. that 00 4090.1. See for example Billingsley [64].g. Theorem 5. in particular for large u. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. The proof is a straightforward combination of the proof of Theorem 5.t.00µB6 + 0. We conclude this section by giving a more general triangular array version of Theorem 5.r. pe . in Asmussen [12].6) therefore does not appear to of much practical relevance for the compound Poisson model. pt? 4 peo. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . For claims with infinite variance. such that the Poisson rate Oe.5) for the compound Poisson model which does not require much more computation. The picture which emerges is that the approximations are not terribly precise. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation.5) and (5.Po = 09µB6 . Then as 0 _+ 90.Pe. Furrer. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. and two further standard references in the area are Grandell [168].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. the simplicity of (5. However. B0 * Boo. In contrast. In view of the excellent fit of the CramerLundberg approximation. and which is much more precise. the claim size distribution B9 and the premium rate p9 depends on 0. on the premium rule involving interest. the B9. we have ^A. However. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory..6 of [APQ]. All material of this section can be found in these references. e.1 and Section VIII. [169]. as 0 * 00 and that the U2 are uniformly integrable w.120 CHAPTER IV. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. (5.
However . 77 = 1/p .Q (B[s] .4. and we are studying b(u. 9o T 0.6.T) = Peo(r(u) < T) for 90 < 0. which we have seen to play an important role for example for the CramerLundberg approximation . whereas there we let the given 3B./c(9 . Then r. this means the following: 1. Bo(dx) = B[eo]B(dx). 77 is close to zero.'(yo) = 0 and let 90 = 'Yo. .ao (0) _ /c(s + 9 . PB('r(u ) < oo) < 1 for 9 < 0.90) and the given risk process corresponds to Poo where 90 = 'yo. and we want to consider the limit 77 10 corresponding to Oo f 0.1 > 0. 0(0) = 0. Determine yo > 0 by r. In this setup. 3. 2. Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. Since Brownian motion is skip free. risk process with safety loading 77 > 0 correspond to 9 = 0 . P9(r (u) < oo) = 1 for 9 > 0. this is because in the regime of the diffusion approximation . The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III.1) . claim size distribution B .9(s) = Ico ( s + 9) . .6. For each 9.c(s) = . Let PO refer to the risk process with parameters e9oz Qo = QB[90]. In terms of the given risk process with Poisson intensity .s and p = /3µB < 1.90] B(dx). The objective of the corrected diffusion approximation is to take this and other deficits into consideration.90) . this idea ignores (among other things) the presence of the overshoot e(u). B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). 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.
(. tu2 ) i IG (t. bl IG(t81.C..() The idea of the proof is to improve upon this by an O (u1) term (in the following. The corrected diffusion approximation to be derived is (u.Varo S1 = f30Eo U2 = S1..2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. Vargo S. u) = euh(a .7(u)/u2} eh(A..1) IG(x.(y) = 0. C) = 2A + (2 . PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. for brevity. (01. 1) • Since L eatIG (dt. (U. the solution of r. and Si = QoEoU2 = Q B"'['Yo Eo U3 ].e.u. One has (6.3) this implies (take u = 1) Ego exp { .1) . means up to o(u1) terms): . 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. 0o to. 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. C. () where h (A.T) 1+u2 (6. (6.122 CHAPTER IV. The first step in the derivation is to note that µ = k (0) = r0 (00) . IGu+u2. Theorem 5. i.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. . write r = T(u). 9otc0" (0) = 0061 = ul. _ ^(u) = ST .2) .S. C . . (. u) = IG(x/u2..
4.5) according to (6.v. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.h. which is based upon exponential claims with mean µB = 1. A numerical illustration is given in Fig.6.1 below is exact.s.5) Once this is established . however . that whereas the proof of Proposition 6.1 + u2 I Indeed.2) is indeed o(u1).2 ). But the Laplace transform of such a r. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. p = 0. however. 9o T 0 in such as way that C = Sou is fixed. calculated using numerical integration and Proposition 1.'yu /2)(1 + b2/u)} + Aug 1I J . . .ry2 .z .7. is the c. (6. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6. we have p =. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} .1 + 629.1 As u + oo. In ( 1) and (2). 1. bl I IG I t +2 .exp { h(A. .52/u where Z has distribution IG (•.3). of a (defective) r. we get by formal Laplace transform inversion that C 2 u. The solid line represents the exact value . 6 . The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3). in (3) and (4).3 = 0. the r. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . .h. distributed as Z .s. To arrive at (6. just replace t by Tb1/u2. of (6. 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.f. Note. 1% in (2) and (4).2).d. and the dotted line the corrected diffusion approximation (6.3.v.
124 0. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small. (Inc 0s 0.00 0. A51 7(SAT 3 3 h(X.114 0.(061 0.T1 00.aa1 .01 0.u2 2u3 (e . OM 0.011 L1 60 T IM 11. and all of the numerical studies the author knows of indicate that its fit at p = 0.7. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].07 0.08 a.2 e.() Lemma 6. The proof of Proposition 6.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.19)2 11 20 20 i0 T 1n0 Figure 6. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p. For further numerical illustrations.^) . .T) 111 0.TI CHAPTER IV. the fit at p = 0..1 proceeds in several steps.W21 0.1 It is seen that the numerical fit is extraordinary for p = 0.7 or at values of Vi(u) like 1% is unsatisfying.199 0.T) 0..4 may not be outstanding but nevertheless.08 0. Similarly.0 0.1 W IU.111 W(U.EB 0 p ex p ( 7 S h ^)u . see Asmussen [12]. PROBABILITY OF RUIN IN FINITE TIME 0.OOIi O.05{ 0.
.C2 = 2). 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 .C)C/u .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .. () 62 Eeo exp u u2 J . () + C and note that 2 KO (0) = 102. (6.00)(u +C)  'r (. 1 = PB(T < oo) = Eo0 exp 125 {(B .s.6) u U3 Lemma 6 . (6.+ h (A.4) that the r. 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 4.4 Ea.2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration . the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .2u (B3 .7) 2 2 .6.h. in Lemma 6.T (co (8/u) .co ((/u)) } Let 8 = (2a + (2)1/2 = h().. CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.61a2T (B3 .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.1) h(A.(3)Eea LauT exp i 3J . + a1b2 + ..co (e) . exp ue } al 1J 3 exP I [2).
2 (^/2 + 3y9o + 390) + O(u3).6  d h(A. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6. letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2).2) for O(u) (indeed . yu/2). yields +90 62 0 + O(u 3) 2u2 +O(u 3).7) and using eh(a.126 CHAPTER IV.() I 1 + u2 ) y .2u [2A+ (2 3 .h (A. C) ( 1+ u2 The result follows by combining Lemma 6 . PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27. () . Thus a2 y = 290 + O (u2). There are two reasons for this : in this way.(2A + ()1/21 exp S h(A.h.() . () by h(\.S) d e 62 . Thus by Taylor expansion around ( = 90u.1 (y/2 + Oo)u .\ + () 1 2 / . 0 The last step is to replace h(A. yu/2) h(A.2. 5 exp { _h(A) (1 + / y u J)) exp 1. l Lemma 6 . [2+ (2 .e h(aS)h (^^ 262 exp {_h(.4.s.6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4)..x. 2 and (6. we get h(A. and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. 2 + 00 = .\+ (2 (3 e 2u [ (2. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. and inserting this and 9o 2 = S/u on the r.
'yu/2) 127 ( i+ M pz^ exP { h (A. i. HOW DOES RUIN OCCUR? exp { h (x. () (i+a ) 2A + (2 . ()} 3 h (A. 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. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. 0 1 Proof of Proposition 6.(i+ 62 exP{ h(A. () I 1 + u 2 ) } S 1 . with the translation to risk processes being carried out by the author [12]. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. The answer is similar: the process behaved as if it changed its whole distribution to FL. that is. the same as for the unconditional Lundberg process. ()} .T) has not been carried out and seems nontrivial.1: Just insert Lemma 6.4. In Siegmund's book [346]. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. the analogous analysis of finite horizon ruin probabilities O(u. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting.7. . We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin.5 in Lemma 6. Fuh [148] considers the closely related case of discrete time Markov additive processes. and to the Markovmodulated model of Chapter VI in Asmussen [16]. the 'typical' value (say in sense of the conditional mean) was umL.e. The adaptation to risk theory has not been carried out. the approach to the finite horizon case is in part different and uses local central limit theorems.
FL As example. so in the in the proof.r.EL[e7S. stating roughly that under F(u).3 to . r(u) < oo) .. we give a typical application of Theorem 7. ^(u) is exponential with rate 8 w.F.(u)_ and ^(u) are independent . Recall that . F(u)c] ti e' ru]PL (F(u)`) > 0.vi(u) Ce'Yu Corollary 7. Theorem 7 . + TMOO )..128 CHAPTER IV.(u)_ and similarly the denominator is exactly equal to Ce7u. In fact.(u) is not . F(u)c] P(r(u) < oo) ?P(U) < EL[e7u. then P(u) and FL coincide on .T. P(u) and rate = aL w.(u). In the exponential case. 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.3L and the claim size distribution from B to BL. .F.t. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7].(u)_ is that i.t.1 Let {F(u)}u>0 be any family of events with F(u) E F. Then also P(u)(F(u)) + 1.1. {ST(u)+t . the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . u * oo. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . the Poisson rate changes from .r. .TT(u) _measurable. Proof P(u) (F(u)c) = F(flu)c. (u) and satifying PL(F(u)) * 1.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. Note that basically the difference between FT(u) and . {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u).ST(u)}t> o is just an independent copy of {St}t>o).FT(u) = o' (T(u ). PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.2 If B is exponential. .
HOW DOES RUIN OCCUR? Corollary 7. Notes and references The results of the present section are part of a more general study carried out by the author [11]. the queueing results are of a somewhat different type because of the presence of reflection at 0. This is currently a very active area of research. take I(Tk < x) .(1 .eALx) M(u) k=1 u The proof of the second is similar.7. A somewhat similar study was carried out in the queueing setting by Anantharam [6]. however. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x).3 M(u) pcu) 1 .3. who also treated the heavytailed case. From a mathematical point of view. the subject treated in this section leads into the area of large deviations theory. . Proof For the first assertion.eaLx. see further XI.
This page is intentionally left blank .
the Tn are independent..Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. with the same distribution A (say) for T2.i.d.1 (T1 = a1). Proposition 1. with Nt = # {n: Un <t} the number of arrivals before t. AA t*oo lim St = lim ESt t tioo t = p . . A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. In the socalled zerodelayed case.1. are i.Then no matter the distribution Al of T1i B.7).Q.(1.. the one corresponding to the stationary case by 00)(u).1 Define p = !µ. We use much of the same notation as in Chapter I.1).. . the distribution Al of T1 is A as well. and M is the maximum of {St}. . and the one corresponding to T1 = s by 1/i8 (u). D'2. {St} is the claim surplus process given by I. Thus the premium rate is 1. T3.. r(u) the time to ruin. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). of the risk process form a renewal process: letting Tn = Qn .. . with common distribution B. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. see A. the claim sizes U1. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 .. U2..
but the arrival rate in the ON state is . Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently).1) ENt/t + 1/µA. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U.Nt] * a/PA. 2). The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. Proposition 1.0 > 0. s + t µA PA 0 Of course. CHAPTER V. Here are two special cases of the renewal model with a similar direct interpretation: Example 1.1) follows . Thus.1 of the safety loading appears reasonable here as well. However .1 gives the desired interpretation of the constant p as the expected claims per unit time. 3) follows similarly by Blackwell 's renewal theorem. Example 1 . say at a. such that no arrivals occur in the off state. From this ( 1. the definition 77 = 1/p . The renewal model is often referedd to as the Sparre Andersen process. and (1 . If the environment is Markovian with transition rate A from on to off and u from OFF to ON. after E. A. by the elementary renewal theorem (cf.132 Furthermore for any a > 0.2 (DETERMINISTIC ARRIVALS) If A is degenerate.1).t. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth.St] = a(p .3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. t 4oo Proof Obviously. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U.a is really the accumulated claims over a period u of length a. OFF. RENEWAL ARRIVALS lim E [St+a . the . Nt ESt = E E UI Nt t = ENt•pB . For (1 . Nt + EVar U. stating that E[Nt+a .
A is phasetype (Example 1.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.. (1.2. we have From this the result immediately follows. INTRODUCTION 133 interarrival times become i.} with {S(d)} a discrete time random walk with increments distributed as the independent difference U . 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^.4) fo Indeed. the first term represents the probability F(U1 ..r.y)B(dy).s.4) w.t. u For later use.1.i. arrival times. integrate (1. However. the fundamental connections to the theory of queues and random walks. and then the whole process repeats itself). (an arrival occurs necessarily in the ON state.1. Proof The essence of the argument is that ruin can only occur at claim times. if for nothing else then for the mathematical elegance of the subject. initial vector (1 0) and phase generator 11 However.oFF}..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. in general the mechanism generating a renewal arrival process appears much harder to understand. we feel it reasonable to present at least some basic features of the model.d. and the present author agrees to a large extent to this criticism. 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 . and for historical reasons.1.. Therefore. For the stationary case. the relevance of the model has been questioned repeatedly. 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 . .s < u).4) with phase space {oN. More precisely. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1.. U1 .T between a claim U and an interarrival time T. S o<t<oo n=0. Ao..
the remaining part of the prepayment (if any ) is made available to the company. then 0 * (u) = 1 for all u > 0. At the time of death . The initial reserve is obtained by prepayments from the policy holders. i.1) +ry. resp .3* pB. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion .Ut. St = t . A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution.134 CHAPTER V. 00). RENEWAL ARRIVALS 2 Exponential claims. Theorem 2 . U Figure 2.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . < 1. If . b=1 !=1 where {Nt } is a Poisson process with rate .1 If.1 r* (u) One situation where this model is argued to be relevant is life annuities. t.1. (2. 2.a*PB• > 1. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (. each of which receive a payment at constant rate during the lifetime .1) . with common distribution B* (say) concentrated on (0. That is . Using Lundberg conjugation . then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] .d. the claim surplus process are given by Nt Nt Rt = u+^U.3* (say ) and the U. are independent of {Nt} and i. the claims and the premium rate are negative so that the risk reserve process . A typical sample path of {Rt } is illustrated in Fig.
0 Now return to the renewal model.UB. and the Lundberg conjugate of {St} is { St } and vice versa. 2. Hence y exists and is unique. and thus 1 = P(T. (a) is*(a) (b) .1.2(a). > 1 . 0) and has typically the shape on Fig. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u).f. 2.2.(u ) < oo) = E {e7sr_ (u).(a) 7 Figure 2. B.0. the safety loading of { St} is > 0. Hence T_(u) < oo a.3*. of {St} is c(a) = is*(a7). If I3*pB* < 1.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. Then the function k* is defined on the whole of (oo. cf.. Let B(dx) = ^e7x B*(dx). B* [7] and let {St} be a compound Poisson risk process with parameters . then by Proposition 111. Then the c.Rt.g. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r.* (a) = log Ee'st I. Since ic'(0) < 0. St=Rtu=St. . T_ (u) = inf { t > 0 : St = u 'r* (u).2 Assume now . Fig. Proof Define 135 St =u . B*.s. Define T_ (u) = inf It > 0 : St = u} .2(b).
f.Tr+. we get Ee'M(d) = Ee°M* _ Y/(.a) = 1 .1 means that M* is exponentially distributed with rate ry.'s and noting that V)*(u) = P(M* > u) so that Theorem 2.Ti = U1..1 it is seen that ruin is equivalent to one of these values being > u. alternatively termination occurs at a jump time (having rate 8). 3* = 6.+Tn U1 Un.7r+ 7r EeTo b/(Sa) + +. with the probability that a particular jump time is not followed by any later maximum values being 1 . and from Fig ... which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII. the failure rate of this process is y.u+ and lr+. and 5PA > 1. and hence the failure rate . According to Theorem 2.Un } = max St = t>0 n=0. To + max {Ul+•••+UnTI..2 If B is exponential.1.Y a I. 2. Taking m..4. Now the value of {St*} just before the nth claim is To +T1* +.136 CHAPTER V. 1) means that 8(A[ry] .. with rate S (say).... However.. u Hence P(M(d) > u) _ 1r+e'r".• • • . the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 . respectively. then . A variant of the last part of the proof.Tn} n=0. and (2 ..e. Hence M* max {To + Ti + • • • + Tn .2).1) + ry = 0 which is easily seen to be the same as (2. To + M(d) in the notation of Proposition 1. RENEWAL ARRIVALS Theorem 2 ..1. T2 = U2.•.g.. Then B* = A.4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value.1.)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[.Y] (2.2) 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 .Ui ..
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] .2.. the imbedded discrete time random walk and Markov additive processes. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1.. This follow since. resp.7r+) and hence r+ = 1.6. 111.7r+). Hence the failure rate of M(') is 6(1 . 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^. we see that ry = 6(1. However.. hence exponential with rate b. which states that for a given a.3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1.5. B^d) where Aad> (dt) = ^[ a] A(dt). 3a The imbedded random walk The key steps have already been carried out in Corollary 11. Furthermore.B(dx). a ladder step is the overshoot of a claim size. 0 3 Change of measure via exponential families We shall discuss two points of view.3. The probability that the first ladder step is finite is 7r+. : S(d) > u} . the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U .2. Putting this equal to y. letting P(d) refer to the renewal risk model with these changed parameters .y/b. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. Bads (dx) = .7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r".4. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+..
ik.e.Ce"u where C = limu. Consider now the Lundberg case. O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) .1) is explicit given 7. we get: Proposition 3.2 In the zerodelayed case. (d) (7) _ 0. let 7 > 0 be the solution of r. E(d)e 1' (u).T is nonlattice.3 For the delayed case Tl = s. (a) '(u) < eryu. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. (b) V)(u) .. i . and claim (a) follows immediately from this and e (u) > 0. In fact..4. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0. 187) and thereby for ^(u) to be nonlattice w.1 For any a such that k(d)' (a) > 0. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form. just note that F7(d) is nonlattice when F is so . 7µA .. provided the distribution F of U .C(°)eryu where C(O) = C0[7] .C8e7u where Cs = Ce78B[7]. Proof Proposition 3.p)/($B'[7] .2 p. in the easiest nonexponential case where B is phasetype.138 CHAPTER V. VIII.1 implies Cu) = e«uE ( 7d)e«^(u) . We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . For claim (b).t. For the stationary case.r. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . 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 . Corollary 3.1).(u) . cf. 00)(u) .
y) = e°yh(s)..Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds).h'(s).dt ) eadt = h ( s) .0 ) = Eo[ha ( Jdt. 0 0 . To determine boundary 0.y) B(dy) 0 For the stationary case. B(x) = o(e7x) and dominated convergence. where G is the infinitesimal generator of {Xt} = {(Jt. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). According to Remark 11.a . IPA 0 Of course. E8h0 (Jdt..1) = C(O). Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . 0) = tc(a)h(s). The expressions are slightly more complicated and we omit the details.9. Sdt) = h(s . Let P8f E8 refer to the case Jo = s. 0) = ah (s) .(s. 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] .4).5. (s. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt.5. (u + s . delayed version of Lundberg's inequality can be obtained in a e7u./c.St)} can be defined by taking Jt as the residual time until the next arrival.dt(ah ( s) + h'(s)) so that Gha ( s. For s > 0. we look for a function h(s) and a k (both depending on a) such that Gh. we invoke the behavior at the 1 = h«(0.1) (normalizing by h(0) = 1). h(s) = e(a +x( a))8 (3.3. Here K.(°) ( Ce8B[7] Ao(ds) similar manner. 3b Markov additive processes We take the Markov additive point of view of II. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. St)} and h.
rc(a)] = 1.140 CHAPTER V. [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +.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. Further. RENEWAL ARRIVALS B[a]A[a .rc(a)] B[a] A[a .8. = J8 = T2.a . (3. . An easy extension of the argument shows that U1.2). Note that the changed distributions of A and B are in general not the same for Pa.e(«+k(a))t esy A(dt).rc(a)] = B = Ba[13]Aa[5].5. .4 The probability measure Pa. .2) As in 11. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J.c(a)] B[a] Proof Pa. T2 are independent with distributions Ba.s and P(d).S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east . however. 5 For the compound Poisson case where A is exponential with rate ...s governing {(Jt. Ease AU1+6T2 [ AU1+6T2 = Ea a LT.. A[a .c(a)] which shows that U1.tK( a)e. U. .13]A[b . the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same.. i. Ba(dx) = B(dx). resp. rc(a) = 0 (B[a] .2) means 1 = B[a]/3/(/3+a+rc (a)). Proposition 3..1)a in agreement u with Chapter III. Ba where Aa (dt) .( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. Remark 3 .a .. Aa as asserted . since JT. [a + /3] A[b . J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk. . (3. we can now for each a define a new probability measure Pa. An important exception is. resp. . T2. ] = E.e.s(Jo = s) = 1 follows trivially from Lo = 1.
say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims.s eaysr(")+r(u ) K(ay) h (s) .r.y yuAa y [ay + K(ay) . The virtual waiting time Vt at time t is the residual amount of work at time t. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue.g. which is the same as the asserted inequality for 0. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time).t. Then J(rr(u)) = TM(u)+1 and hence Ws(u.4. that is. Then "^ e(ay+w(aY))8 Ys(u. .(u. 2.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. Proof As in the proof of Theorem IV. The claim for the zerodelayed case follows by integration w. A(ds). the time from he arrives to the queue till he starts service. or FCFS = first come first served) queueing discipline and renewal interarrival times.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. not after time T.)+1 e J j e(ay+w(ay))8 e . yu). and define yy = ay . for the zerodelayed case zp8(u. XII. .yu ) e7vu A[ay . yu) = F'ay. see e. and U„ the service time of customer n. defined as the single server queue with first in first out (FIFO.ay+ray))TM(.. Using the Markov additive approach yields for example the following analogue of Theorem IV. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45].5.5 Proposition 3. and assume that T„ is the time between the arrivals of customers n . yu ) < e7yu..yx(ay).4. For the approach via Markov additive processes. Label the customers 1. . Let M(u) be the number of claims leading to ruin . 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. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. it is easily seen that ic(ay ) > 0. u The approach via the embedded random walk is standard. the amount of .4. that is. [APQ] Ch.1 and n.
and obviously z/'(u) = limN.4. (4. Then: (a) as n + oo. equivalently..• • • Tk ).2) (b) as t * oo. 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. If W1 = 0.. the proposition follows. 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. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.4. Let the T there be the random time UN.4): Proposition 4.1. and we have P(V > u) = ?/iiol(u).1. (u).+ Un.1) The following result shows that {Wn} is a Lindley process in the sense of II. whereas in [On . we have Wn = Van(left limit). 0 Applying Theorem 11. the waiting time a customer would have if he arrived at time t.4: Proposition 4.Tn)+. and we have P(W > u) = V.2. since customer n arrives at time on. Vt converges in distribution to a random variable V.n1 (U1 +• • •+Uk Tl .142 CHAPTER V. equivalently.1 and Corollary 11. Wn converges i n distribution to a random variable W. p < 1. then Wn v M. in which case {V} remains at zero until time on+1..1). It then jumps to VQ„ . Thus. an+1) = [on.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ . on + Tn) the residual work decreases linearly until possibly zero is hit.2 Let Mnd) = maxk=o. Thus Vos}1 _ = (Wn + Un .3.3 Assume rl > 0 or..1 Wn+1 = (Wn + U.4. . (4. but interchanging the set . The traffic intensity of the queue is p = EU/ET. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. and combining with (4. (4. but we shall present a slightly different proof via the duality result given in Theorem II.3) Proof Part (a) is contained in Theorem 11."^ Vi(N) (u). we get: Corollary 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). Then the corresponding queue is M/G/1. Ti) and similarly for the U.4. (4. Hence for x > 0. conditioning upon U* .Ao in (b). we get W = (W + U* . (4. THE DUALITY WITH QUEUEING THEORY 143 (T1. and we get: . resp . where U*. as WN. It follows that P(WN > u) =.e. which implies the convergence in distribution and (4. Corollary 4. Then the arrivals of {Rt} in [0. x > 0... convergence in distribution hold for arbitrary initial conditions . T1... u Now return to the Poisson case .1.T* = y yields K(x) = P ((W + U* .. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival.. hence (since the residual lifetime at 0 and the age at T have the same distribution . i.T*)+ < x) = P(W + U* . Then K(x) = J x00K(x . For part (b).2). but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.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.T*)+. TN) with (TN. we obtain: Corollary 4. T1 ..T* are independent and distributed as U1.4 The steadystate actual waiting time W has the same distribution as M(d). T] form a stationary renewal process with interarrival distribution A.5) Proof Letting n . namely W1 = 0 in (a) and Vo = 0. In fact .. cf. Letting n oo in Corollary 4.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). by an obvious reversibility argument this does not affect the distribution . K(x) = P(W < x).oo in Proposition 4.2..y)F(dy).(N)(u) has the limit tp(u) for all u. we let T be deterministic .5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x).le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before .. However. A.
The equation (4.5) looks like the convolution equation K = F * K but is not the same (one would need (4. VIII). RENEWAL ARRIVALS Corollary 4. 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. the actual and the virtual waiting time have the same distribution in the steady state. Cohen [88] or [APQ] Ch.g. Hence '(u) = Ali(°)(u). .5) to hold for all x E R and not just x > 0).6 For the M/G/1 queue with p < 1. the zerodelayed and the stationary renewal processes are identical. Note that (4. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . Asmussen [24] and references there. That is.g.5) is in fact a homogeneous WienerHopf equation.144 CHAPTER V. implying P(W > u) = P(V > u) for all u. Some early classical papers are Smith [350] and Lindley [246]. see e. Proof For the Poisson case. W v V.
Thus. and can be computed as the positive solution of WA = 0. {Jt} describes the environmental conditions for the risk process. As in Chapter I. 145 Oj( u.T) = Pi (T(u) < T).f pi. N. i=1 0 and r(u) = inf It > 0: St > u}. The intensity matrix governing {Jt} is denoted by A = (A.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. Ire = 1. ..(3i when Jt = i. • The premium rate when Jt = i is pi. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). t St = E Ui . here it exists whenever A is irreducible which is assumed throughout.)iJEE and its stationary limiting distribution by lr. M = supt>o St. {St} denotes the claim surplus process. dv.
this is no restriction when studying infinite horizon ruin probabilities.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). Example 1.2. one expects that 3i > on and presumably also that Bn # Bi.a('). meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. P = E 7riPi. Then the state space for the environment is the disjoint union of E(n) and E(i).4) with representation (E(i). /3 = Nn when j E E(n). assume that the sojourn time in the icy state has a more general distribution A(i). i and corresponding arrival intensities Qn.1 Consider car insurance. say. say. r^ = P (1.14. We let p Pi = /ji/AB. Unless otherwise stated. For example. and p is the overall average amount of claims per unit time.T(n)). An example of how such a mechanism could be relevant in risk theory follows. cf. and we have f3. respectively.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. in blockpartitioned form. According to Theorem A5.11 below. and assume that weather conditions play a major role for the occurence of accidents.5 below. = iii when j E E(i). leading to E having two states n. T(=)). u . f3i and claim size distributions Bn. the operational time argument given in Example 1. we shall assume that pi = 1. Bi. which is clearly unrealistic.146 CHAPTER VI. Proposition 1. t(i) = T(')e are the exit rates. 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. with rates Ani and Ain. we could distinguish between normal and icy road conditions. a(i). Thus. we can approximate A(i) with a phasetype distribution (cf. cf. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. Example 1 ..
. Indeed. 1 . i8f n1. . w.>. but assume now that the arrival intensity changes during the icy period. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). depending only on 77. This amounts to a family (A(")) ?CH Of sojourn time distributions.n.p. One way to model this would be to take A(') to be Coxian (cf.. and 1/ii(u) = t/ii(u).3i/pi.4) with states i1. MODEL AND EXAMPLES 147 Example 1 .3i. u Example 1 . The simplest model for the arrival intensity amounts to . T(9) +wggt(9)0.1. u Example 1. Approximating each A('?) by a phasetype distribution with representation (E('l). where W = (w. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. the parameters are ^ij = aid/pi. St = SB=(t).....J017.. say. u From now on.2. iq (visited in that order) and letfOil >.tEH is a transition matrix.a(n). let T 9(T) = f pi. dt.. Qi = . say it is larger initially. such that a sojourn time of type rt is followed by one of type c w.j = . 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. i ) : n E H. (9) where q = CHI.. A('^).. Example VIII.T(n)).Q. n8}. and similarly for the normal period. t(n) = T("i)e.1. i E E(n) }. Then for example wi... In the car insurance example.3 Consider again the alternating renewal model for car insurance in Example 1. it = Jel(t). one could for example have H = {i1.. the state space E for the environment is { ('q.3. and . 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. 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) .5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. is the probability that a long icy period is followed by a short normal one. resp.
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. . Next we note a semiMarkov structure of the arrival process: Proposition 1. t l=1 Note that the last statement of the proposition just means that in the limit.5. In particular.8 As t oo. N > 1(Ul < x) a4 B*(x). the empirical distribution of the claims is B*. . The key property for much of the analysis presented below is the following immediate observation: Proposition 1. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. we put )3* = E 7fi/3i.(Qi)diag)• More precisely. B* = 1 /^* Bi. )3*. Pi (Ti E dx. vi(dx) = .148 CHAPTER VI. one can associate in a natural way a standard Poisson one by averaging over the environment.e(A(Oi)d'sg)xe.A . 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). 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.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.7 The Pidistribution of T1 is phasetype with representation (ei. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. o = 0. Nt Nt a . qij = 0 in the notation of Chapter 11. JT1 = j) = Qj • e. dx. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.(3iBi(dx). More precisely.
Then {St)} + {St*} in D[0. zli( (u) . Nt a' t t iEE Also. Proof According to Proposition 1. Bi. Then it is standard that ti lt '4' iri as t > oo.aA . A. oo) as a 4 oo. and let {St °i} refer to the one with parameters Pi.9 Consider a Markovmodulated risk process {St} with param eters Ni. Proposition 1. B*. By Proposition A5. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. the limiting distribution of the first claim size U1 is B*.(/3i)aiag). we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti..1. In particular. given {Jt}0<t<0. 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 . this converges to the exponential distribution with rate 0* as a * oo. cf.. Hence Nt'> a ..* (u) for all u..x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). 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. {St} to the compound Poisson model with parameters 0 *.7.6.4. has distribution (7ri)3i //3*)iEE and is independent of Ti. However . e. ^j 7riNi. aA. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. Bi(x).. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. and furthermore in the limit JT. Example 11. Bi. y Ni) i Nti) t a. In particular. 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$.2. i. Conditioning .
0 Example 1. 132=2. e. B2=1E3+4E7. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. That is.. the company even suffers an average loss.. lines in the path of {St}.g. oo).2 +2 2 = 3. U.s = 1o in state 2.l3* and U2 having distribution B*. On Fig. Claims of type E3 arrive with intensity 2 . Computing the parameters of the averaged compound Poisson model.150 CHAPTER VI. T) + ?P* (u.).s = o in state 1 and with intensity 1 . the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. s 5 in state 2. 1.. the paths of the surplus process will exhibit the type of behaviour in Fig. 1. we first get that 3 (3* = 2.31µB 2 = 2 5 3 7 70 Thus in state 1 where p..FT. and (at least when a is small such that state changes of the environment are infrequent). U2) are independent of . 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 = . we may imagine that we have two types of claims such that the claim size distributions are E3 and E7.2. resp.1. T) for all u and T. which also yield O(a) (u. Continuing in this manner shows that the limiting distribution of (T. shows similarly that in the limit (T2. there are p = 2 background states of {Jt}.=1. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier). is as in {St }. with T2 being exponential with rate . 9 .. thick. those of type E7 with intensity z s = 5 in state 1 and with intensity z .2}. The fact that indeed 0(a) (U) 3 0* (u) follows. Thus.. marked by thin.. from Theorem 3. > 1.. B1=3E3+2E7. state 1 appears as more dangerous than state 2.2.10 Let E_{1.1 with periods with positive drift alternating with periods with negative drift.1 of [145]. MARKOVIAN ENVIRONMENT upon FT. From this the convergence in distribution follows by general facts on weak convergence in D[0. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary..
1. That is. 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). iEE . 01 /.. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. 0 The definition (1. For (b).s. (b) St/t * p . note first that EN Uk')/N a4' µgi.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 1.1. a fraction r.(3. the averaged compound Poisson model is the same as in III.11 (a ) ESt/t * p .8. Proof In the notation of Proposition 1. = P. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. MODEL AND EXAMPLES 151 Figure 1. t * oo.1 Thus. t + oo.1).1 a.
PB. + Xn SWn ](1 a . 136 or A.152 CHAPTER VI.a form a renewal process .. limit p . [302]. some early studies are in Janssen & Reinhard [211]. and a more comprehensive treatment in Asmussen [16]. and involves a version of the . n n Thus {SWn l is a discrete time random walk with mean zero. [APQ]. Proof The case 77 < 0 is trivial since then the a. let some state i be fixed and define w=wl=inf{t >0:Jt_#i. then M < oo a.1 of St / t is > 0. If 77 > 0.Jt=i}.. Proposition 1.1 and the Corollary are standard. . There seems still to be more to be done in this area.4. In risk theory... and hence 1/ii(u ) = 1 for all i and u. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. see [APQ] p.. EiX = 0. MARKOVIAN ENVIRONMENT Corollary 1. see the Notes to Section 7. w2=inf {t>w1:Jt_#i. s. 38) Eiw1 = 1/ir.0i(u) < 1 for all i and u. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.. [315]. and hence oscillates between 0o and oo so that also here M = oo. X 1 =Sty. having the Pidistribution of X.Eiw o'o Eiw • E ^ifjµs.1 jEE = (p .. Then by standard Markov process formulas (e..g.ld.s.s. 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. Eiw. The proof of Proposition 1.1)Eiw = 0. X3. [212]. . then M = 00 a. Now obviously the w. also + . and ..2(a) p. See Meier [258] and Ryden [314]. The mainstream of the present chapter follows [16].. Statistical aspects are not treated here. X2 =SW2 So.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes.\ i and EiX1 Ei f 13 J. Since the X„ are independent . Theorem II.Jt=i}. and hence wn /n a4. dt . and so on. and hence M = 00.12 If 77 < 0. Now let r) = 0. with X2.
.. by specializing results for general stationary risk processes (Theorem II . G+ is the matrix whose ijth element is E G +(i.4) we obtain the following result . That is.6.Jt=j)dt.3*B *(y)dy. k.dx). For measurevalued matrices. e. Proposition 2. oo) = J ao 0 G+(i. •). n=0 (2. j.6*.A) =ZI(St E. •)• kEE Also.j. see also Example II.(u) = Pi(M < u) = e' E G+ (u)(I . let G+(i.j E E.IIG +II)e. •) II = JG+(i. Let further R denote the preT+ occupation kernel. j.i.x). j. only with the product of real numbers replaced by convolution of measures. However . dx)/jBj(y . T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i. cf.1) 0 (b) G+ (y.j. Thus.5. B* in Section 1.2) R(dx)S((y . we define the convolution operation by the same rule as for multiplication of realvalued matrices. oo)). 6. T R(i. IIG+ II denotes the matrix with ijth element IIG+(i. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0).x.a/i. for i.1 irG+(dy)e =. j.EA. The form of G+ turns out to be explicit (or at least computable). Proposition 2.A) = Pt(ST+ E A. the definition of .Jr+ =j.2.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element.2 (a) The distribution of M is given by 00 1 . oo)) = f R(i. we get the same ladder height distribution as for the averaged compound Poisson model. 00 (2. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. but is substantially more involved than for the compound Poisson case . (y.g. •) * G +(k. j.
the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. The u proof of (2. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). {mx} is a non terminating Markov process on E. mx = j when for some (necessarily unique) t we have St = x.1) follows by summing over n and j.3 When q > 0. To make Proposition 2.1 for an illustration in the case of p = 2 environmental states of {Jt}.6. and that the environment is j at the nth when we start from i is e .1 The following observation is immediate: Proposition 2.2) is just the same as the proof of Lemma 11. To this end . JJ = j. 0  x Figure 2. see Figure 2.2 useful . G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. From this (2. marked by thin. . III to bring R and G+ on a more explicit form . resp. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value.IIG+II)e.3. we need to invoke the timereversed version {Jt } of {Jt} .154 CHAPTER VI. we need as in Chapters II.3) We let {St*} be defined as {St}. thick. lines in the path of {St}. hence uniquely specified by its intensity matrix Q (say). St < S* for u < t. That is.
( Q( n)) converges monotonically to Q. In general. {S. = x}. If there are no jumps in (t. the sequence {Q(n)} A* defined by Q(O) = . An excursion of {St*} above level x starts at time t if St = x. 2. and the excursion ends at time s = inf {v > t : S..2 . The definitions are illustrated on Fig.(/3i) diag. Otherwise each jump at a minimum level during the excursion starts a subexcursion.(/3i)diag + T S(dx) eQx. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. For example the excursion of depth 2 has one subexcursion which is of depth 1. } is a minimum value at v = t. we say that the excursion has depth 0.2 where there are three excursions of depth 1. s]. Figure 2.2. 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). we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. Furthermore.2. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2.*.0.and a jump (claim arrival) occurs at time t. corresponding to two subexcursions of depth 0. 0 mms1   ^O \ T. Proof The argument relies on an interpretation in terms of excursions. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. Q( n+l) _ ^.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .
156 CHAPTER VI. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. A) = f Pi(mx = j) dx eie4xej dx A u (2. StEA . It follows that qij = A.j +/3ipij. Similarly.5 R(i. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). 0)). Suppose mx = i. i. p1^) Define a further kernel U by f U(i. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. 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. (2. Then a jump to j (i. = j. By considering minimum values within the excursion..6) . mx+dx = j) occurs in two ways . A) = L' U(j. Similarly by induction . either due to a jump of {Jt } which occurs with intensity A= j. 7rE Proof We shall show that Fi(Jt=j. we first compute qij for i $ j. A).u< t).4). j. Writing out in matrix notation .5) A (note that we use A = {x : x E Al on the r. or through an arrival starting an excursion terminating with J.St EA.(01)diag = Q.4) To show Q = cp(Q).. of the definition to make U be concentrated on (co. e.j. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. Theorem 2 . It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* .s.Qi + )%pij) Now just note that t pij and insert (2.T+>t) _ ^iF 7ri (JJ =i.St <S*. Q = W(Q) follows. h.
We may then assume Ju=Jtu. St < St U.0<u<t) = P. K( n (d) the sequence converges monotonically to K.S„<0. St E A. (Jo = j. e. x < 0. G+((z. From Qe = 0.2. the CramerLundberg approximation (Section 3).. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx). we shall see that nevertheless we have enough information to derive. To this end..1 can be rederived using the more detailed form of G+ in Corollary 2.(Jt=j.=StSt.. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). Remark 2. oo)) = f o' eIXS((x + z. 0 < u < t) = 7rjPj(Jt =i..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. and get irPi(Jt =j. and to obtain a simple solution in the .z+>t) = P. Jt = i. {Jt }. S.(. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2. 0 +1) = cp (K( n)) defined by K(o) = A .6 (a) R(dx) = eKxdx.4]. and we let k be the corresponding right eigenvector normalized by Irk = 1.St EA. oo))dx.6 is hardly all that explicit in general.g.7 It is instructive to see how Proposition 2. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.StEA. St EA. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. consider stationary versions of {Jt}..Qi)diag.St <Su. (b) for z > 0. and this immediately yields (2.0<u<t.Jo=i.6). 0<u<t). dt. where A is the diagonal matrix with 7r on the diagonal: Corollary 2.t.r.
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
we let G+ * W(u) be the vector with ith component E(G+(i.5). r(u) yu o)(y)eavuEav.00 su e7( ( 3. 1 Similarly.y)G+(z.i [e*(u)K(av). exp {e() + r(u))} . as in the classical case (3.5) will produce the maximal ryy for which the argument works. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av). av 'i [h. dy)• o iEE jEE .i [eT(u)K(av ).3. However. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. (3.j) * coj)(u) _ f u ^Pj(u . if y > 1lk'(ry). we have ic(ay) < 0 and get 'i(u) . yu) f h(av)e v avuE«v. hj P . Chie ryu < Vi(u ) < C+hie 7u.7. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .9) For the proof. 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 I (a) exp {aye(u) + r(u)r. for a vector <p(u) = (cpi (u))iEE of functions . we shall need the matrices G+ and R of Section 2. 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. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i.j. (u.8 ) Then for all i E E and all u > 0..V)i(u.(ay)}.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). oo)) and.
3jhj // f 00 R(i. Then cpin)(u) sit (u) as n + oo. U = U".ery(&u+x)Bj (dy) Bj(u Bj (u . j. = Eo G+ G. 00 f C_ hj f e(Y)G+(i.dy). Lemma 3 . 00 Thus C+ > hj f"o e7(Yu)G +(i.x ) R(i. °O .u IMP:°) (u) I < oo. dy) : 1(u) < C+ > hj u e(1tL)G+(i. dy) = aj f Bj(dy .(0) ] (u) < sup Jt t. dx) f e7( vu)Bj (dy . Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. dx). and define W(n+1) (u) = G(u) + (G+ * tp(n))(u). Proof Write UN = EN G+ . _ To see that the ith component of U * G(u) equals ?Pi (u).x) jEE 00 u 0 //^^ C+E. dy) 00 C+ ijhj f R(i.x) x) jEE 0 E Qj f jEE R(i.13 For all i and u.12 Assume sup1. 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. j. jEE u 0 j.j. Hence lim cp(n) exists and equals U * G. n > oo.166 CHAPTER VI. dx) 100 C . j.& (u). j. dx ) Bj (u . MARKOVIAN ENVIRONMENT Lemma 3 .7. if r+ (n) is the nth ladder epoch.u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. we have G *(N +1) * ^.j. 0 G+(i.x ) = Gi(u).
13 Let first cp=°)(u) = C_ hie"u in Lemma 3.T) = Pi(M > u) .8) with y replaced by yo and hi by h=7o ).tpi(u.(u) < T ) to 0i (u) which is different from Theorem 3. j.ST).10 ) by Lemma 3 .M>u) = Ei [VGJT (u .M > u) = Pi(ST<u. this is obvious if n = 0. 13 and the second by the induction hypothesis . ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . +i .Pi(MT > u) = Pi(MT < u.10: Theorem 3 . T) = Pi (7. j. and assuming it shown for n. y]hj = C_ e7uhi.3. j. We claim by induction that then cpin) (u) > C_ hie7u for all n. T) < C+(')' o)hi7u)e7ou8T .11). taking cps°) (u) = 0.u)G+(i. 9 for the last equality in (3.11. let C+(yo) be as in (3.13. dy) (3. letting MT = maxo<t<T St. (3.13) Hence. we have Vii (u) . it follows that Vi(u) < C_(yo) h=70)e7ou. Indeed. from which the lower inequality follows by letting n * oo. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. u The proof of the upper inequality is similar .y)G+(i. 167 u Proof of Theorem 3. (3. Then 0< Vi (u )  0i(u.10) C_ 1 f hje7(y.12) Proof We first note that just as in the proof of Theorem 3. and using Lemma 3 . we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. MT < u. and let 8 = e'(70). and the proof of the lower one is similar.11) C_e7u 57 O+[i. dy) jEE u U +C_ hje7( yu)G jEE"" +(i.n) ( u . dy) jEE o (3. jEE estimating the first term in (3. 14 Let yo > 0 be the solution of 'c'(yo ) = 0.MT<u. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality. j.
o.3) to B = Bi does not depend on i. The results to be presented show that quite often this is so. <s. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. Further related discussion is given in Grigelionis [176]. [177]..3).0. this is not the case for (4. The Markov process {Jt} is stochastically monotone (4. For the notion of monotone Markov processes.. but that in general the picture is more diverse. in part from the folklore principle that any added stochastic variation increases the risk. we refer to . (4..2) (4. V)" if z/i'(u) <'c"" (u)...o. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44].168 CHAPTER VI. B2 <_s.. It was long conjectured that 0* Vi. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).1) Obviously.33 or Bi 0 Bj. (4. (4.o.2) alone just amounts to an ordering of the states. Occasionally we strengthen (4. The motivation that such a result should be true came in part from numerical studies.3) Bl <_s. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). we also assume that there exist i # j such that either /3i <. The conditions which play a role in the following are: .4) To avoid trivialities. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). and finally in part from queueing theory. this correponds to the usual stochastic ordering of the maxima M'. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and .. ".5) Note that whereas (4.3p. Bp. < . 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'.o.31:5)32 .. u > 0. we define the stochastic ordering by 0' < s.
6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. (4. Then V..6).x) dx u o i =1 i=1 (4..5 (cf..x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.9) follows by considering the increasing functions 3iBi (x) and Oi (u .. Comparing (4.9) (4..* For the proof. = aP or b1 = .2. 0 Here (4. (b) P.2)(4.x)B*(x) dx.4) say basically that if i < j .6. T(0) < oo) = Bi(x) dx/tcai . Lemma 4 . then P P P 7rjbj. Conditions (4. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = . Lemma 4 ..10) and (4.. then j is the more risky state . and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.r (JT(o) = i. Theorem 4 .1. Proposition 2. b1 < . it follows by a standard . 3 (a) P. 2 If al < .2. also Proposition 2.7) 7ri. = b.7) and Lemma 4. p).1 for the first term in (4.r (Sr(o) E dx Jr(o) = i.3 for the second) *(u) _ /3 *B* (u) +. < a. ^i 7ri = 1.. .3* f uB(x) z/^. E 7r i Wi(u . 7(0) < oo) = pirf+). where 7r2+) = QiµBilri/p.4. note that (4.4) is automatic in some simple examples like birthdeath processes or p = 2 . Proof of Theorem 4.1 Assume that conditions (4. Proschan & Walkup [140].8) ^j Tri/iBd(x) .3iBi(x)YPi(u . the second follows from an extension of Theorem I1.. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary.9 ) below). COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]....13* J0 u 0*(u .. we obtain (cf. we need two lemmas.r(u x)dx. . Section 4.. < bp and 7ri > 0 (i = 1.1) which with basically the same proof can be found in Asmussen & Schmidt [49].10) Q*B*(u)+.x) of i and using Lemma 4.4) hold. dx (4.2)(4. Conditioning upon the first ladder epoch.
1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . (u) is not in general true: Proposition 4.4 Assume that .0. they are at present not quite complete.4) is essential (the present author conjectures it is). i=1 i=1 7'r(0) _ EFioiwi(0) ...(0) = V. As is seen.4 is not vacuous.. (u) may fail for some u. u Here is a counterexample showing that the inequality tp* (u) < V).s.. µB2 = 104. this ruin probability is /3iPBi. and from this the claim follows. it will hold for all sufficiently large u.6).3i.170 CHAPTER VI. Proof Since 0. What is missing in relation to Theorem 4.h.4 is the understanding of whether the stochastic monotonicity condition (4.2.8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi . MARKOVIAN ENVIRONMENT argument from renewal theory that tk. = 102. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. it is sufficient to show that 0'.0*• i=1 But it is intuitively clear (see Theorem 3. Notes and references The results are from Asmussen.11) i=1` and that A has the form eAo for some fixed intensity matrix A0. u To see that Proposition 4.r (u ) fails for all sufficiently small e > 0. dominates the solution 0* to the renewal equation (4.3µi < 1 for all i.s..11) is of order 104 and the r.6).h. Using (4. of order 101. Frey.* (0)./3*. For u = 0. Bi as e J. 0. (4. 01 = 103.1 and Proposition 4. Rolski & Schmidt [32].(0) < b *'(0) for e small enough. let = ( 1/2 1/2 ) . except possibly for a very special situation . µB. Then i/i*(u) < . 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. Recall that . Then the l. of (4. Q2 = 1.
is nondegenerate unless bi does not depend on i E E. Hence if 5i 54 0 for some i E E.)a. Then {(Jt. it follows by Proposition A1. It is clear that the distribution of X.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.5.4(b) that the limit in (4.2 we have (Ei[e"X'. Further (see Corollary 11.. (4. Jt = i])' EE = vA+n(6.a = E irirci(a). and by Proposition II. (4. in particular .1) .ld) with generic cycle w = inf{t>0: Jt_54 k.12) iEE Theorem 4. Lemma 4.14) is nonzero so that A"(0) > 0. Asmussen [20]) as discussed in 11. cf.5. with strict inequality unless a = 0 or bi = 0 for all i E E. 0 . Now we can view {Xt} as a cumulative process (see A. Proof Define X= f &ids.a.13) (4. Xt)} is a Markov additive process (a socalled Markovian fluid model.13) implies A(a) > 0 for all 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] . This implies that A is strictly convex.14) A„(O) iioo varXt t t By convexity.4.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0.5 y < ry*.g. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E. with strict inequality unless rci (y*) does not depend on iEE.1) .5. e.g.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. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] .(a) > 0 for all a 0 0..
12) and rc*(y*) = 0. Then > risi = 0 because of (4. The corresponding adjustment coefficient is denoted by ry(e).15) once more and letting e = 0 we get . Further a(1) = rc(y*) by definition of A(. Frey. Rolski & Schmidt [32]. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h.. we have 7rh' = 0. If rci(y* ) is not a constant function of i E E.e7r)1 (Ici(Y*))diage.. h(0) = e. MARKOVIAN ENVIRONMENT Proof of Theorem 4. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. (4. Thus y(e) * y* as e 10. Since ic is convex with rc'(0) < 0 . Here we put a = 1/e.eir)h'(0). improving upon more incomplete results from Asmussen.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao .) and rc (•).5. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. y. the basic equation is (A + (rci(y))diag)h = 0. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. In the case of e. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'.5 is from Asmussen & O'Cinneide [40]. Hence rc (y*) > 0.172 CHAPTER VI.. Notes and references Theorem 4.15) Normalizing h by 7rh = 0. (4.16) Differentiating (4. a = 1 in Lemma 4. note that y(a) + mins=1.6.Qi and Bi are fixed . whereas the . h depend on the parameter (e or a). where A.. Let bi = rci(y*). we get rc (y*) > 0 which in a similar manner implies that u y < y*. Hence letting e = 0 in (4. h'(0) = (Ao .p yi and compute 8y 8a a=0 In both cases.
. 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..17) by 7r to the left to get (4. which has recently received much attention in the queueing literature.5. multiplying (4. i = 2. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity .8 If (4.18).i(7' *))diagh'(0).20) Letting a = 0 in (4. (4. We assume that 0 < y < 7i.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). . (4.20) 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).7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. (4.8 when ryi < 0 for some i is open. We get 0 = (aAo + ( lc&Y))diag)h.. p. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) .17) (4. Rolski & Schmidt [32].16) yields Proposition 4. Inserting (4. and may have some relevance in risk theory as well (though this still remains to be implemented). and we have proved: Proposition 4. 5 The Markovian arrival process We shall here briefly survey an extension of the model. Frey.19) holds.18) 0 = 27'(0)p+27r(rs. then 8a a=o All rci (0) Notes and references The results are from Asmussen. 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. The analogue of Proposition 4.
2). where qij is the probability that a transition i * j is accompanied by a claim with distribution.d. This is the only way in which arrivals can occur. Here are some main examples: Example 5 . Again . Note that the case that 0 < qij < 1. the definition of Bij is redundant for i i4 j. We then let (see the Appendix for the Kronecker E = E(1) x E(2). A(1'k) A(2 k1).2 for details). A(l) = T.2) A(1) = A(' 1) ® A(1.6i ) diag. let { Jt 1) }.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . B. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. II. Bij = B.i.174 CHAPTER VI. T). we use the convention that a1i = f3i where 3i is the Poisson rate in state i. j(2) } be two independent environmental processes and let E(k). the definition of Bi is redundant because of f3i = 0. A ( 2) = A (2`1 ) ® A.^) etc. Jt2)) (2. refer to notation) { Jt k) }. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. and the marked transitions are then the ones corresponding to arrivals. For i = j. In the above setting. A(1) = A . . A(l) = tv. Jt = (Jtl). is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. Thus . with common distribution B. the claim surplus is a Markov additive process (cf. Indeed. and thus 1i = 0. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process.4). Bii = Bi . the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i.(13i )diag. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. that Bii = Bi . u Example 5 . MARKOVIAN ENVIRONMENT f o r a transition i + j by A .
.5...1i2 . superpositions of renewal processes....iN C17 AilO. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. all Al i2. Thus. THE MARKOVIAN ARRIVAL PROCESS Bij. This means that the environmental states are of the form i1i2 • • • iN with il.1i2..kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. assume that there is a finite number N of policies. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. 11. iN.iil. possibly having a general phasetype sojourn time.iN = a2. u Example 5 . more recently.. after which it starts afresh.. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313].iN.. 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 .. MARRIED. with rate ai. Example 5 . In fact . 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.iil. claims occur only at state transitions for the environment so that AN2. the kth policy enters a recovering state. e.. In this way we can model. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. The versatility of the setup is even greater than for the Markovmodulated model. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate ..4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. Bilo. i2i .. Similarly.g. E 10.kl is redundant).. or... iN. • upon a claim.iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant.. say. E = { WORKING. iN = all BOi2..}.iN. DEAD etc.. and that the policy then expires. However . RETIRED. DIVORCED. INVALIDIZED. as the Markovian arrival process ( MAP).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).iN are zero and all Bi are redundant. WIDOWED..
• Claims arriving at time t of the year have distribution B(t). 1). we talk of s as the 'time of the year'. Without loss of generality. Lucantoni [248].1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . . we may assume that the functions /3(t). one needs to assume also (as a minimum) that they are measurable in t. 0 < t < 1.p)/p.176 CHAPTER VI. where i f00 xB(°) (dx) _ . [248]. for s E E = [0. continuity would hold in presumably all reasonable examples. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones.3*µs • p = f /3(v) dv 0 0 (6. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). from an application point of view. 1). Neuts [271] and Asmussen & Perry [42]. 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. Some main queueing references using the MAP are Ramaswami [298].2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. For the Markovmodulated model. )3 t 1 J (6. let the period be 1. Thus at time t the premium rate is p(s + t). Lucantoni et at. p * = 0 p(t) dt. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. Obviously. By periodic extension. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. • The premium rate at time t of the year is p(t). 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 . Sengupta [336]. B* = J f B(t) ((*) dt. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . p(t) and B(t) are defined also for t t [0. Let 1 1 /3* _ f /3(t) dt.
p* as an averaged version of the periodic model. 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. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. St = SeI(t). Section 4b). the conditional distribution .6.w(t)) dt). The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . In contrast. equivalently.t. one may think of the standard compound Poisson model with parameters 3*.(3.8. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. B*.3(t) = 3A(1 + sin 27rt). and thus the averaged standard compound Poisson models have the same risk for all A. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. In particular. of the periodic model as arising from the compound Poisson model by adding some extra variability. Thus . It is easily seen that .2 Define T 6(T) = p(t ) dt. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . (6. the discussion in 111. We u assume in the rest of this section that p(t) .3) Note that A enters just as a scaling factor of the time axis.1 As an example to be used for numerical illustration throughout this section. in agreement with the general principle of added variation increasing the risk (cf. Thus. In contrast. u Remark 6 . since the added variation is deterministic. let . respectively. Many of the results given below indicate that the averaged and the periodic model share a number of main features.w(t).1) and Example 1.1. it turns out that they have the same adjustment coefficient.10. 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 . the average compound Poisson model is the same as in III.3* = 3A. we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations.9). not random. or. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. The behaviour of the periodic model needs not to be seen as a violation of this principle.
1) .1) a = J8 . 3 E(8)eaSt = h(s. i. 0 (5)(u.a. As usual.a .(1 . let f 8+1 tc *(a) _ (B* [a] . a) is periodic on R.e.(8) [eaSt+dt I7t] = = (1 .(3(s + t)dt)e«St adt + /3(s + t)dt .east B(8+t) [a] east . see the Notes to Section 7).. of the claim surplus process. . and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00).3(v)(B(vl [a] .T) = P(8)(r(u) <T). Daykin et. of the averaged compound Poisson model (the last expression is independent of s by periodicity).5. and define h(s.g. we obtain E. we start by deriving formulas giving the m. 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.a) Proof Conditioning upon whether a claim occurs in [t.tc* (a)] dv then h (.g.s .1]) . Dassios & Embrechts [98] and Asmussen & Rolski [43]. Jt = (s + t) mod 1 P(8) .5 (see in particular Remark 11.. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive. The exposition of the present chapter is basically an extract from [44]. The claim surplus process {St} may be seen as a Markov additive process. (6.al.f. [101] .. r(u) _ inf It > 0 : St > u} is the time to ruin .1) dv .g. 1). To this end. Notes and references The model has been studied in risk theory by.adt +.Q(v) (B(„) [a] .3(s + t)dt[B(8+t)[a] . J Theorem 6 . t + dt] or not.4) At a first sight this point of view may appear quite artificial. e. with some variants in the proofs. a) etw*(a) h(s+t.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. given that the ith claim occurs at time t is B(8+t). 6b Lundberg conjugation Motivated by the discussion in Chapter II.^8 [.a) = exp { .f. with the underlying Markov process {Jt} being deterministic period motion on E = [0.a be the c.8).
s. a) Thus E(8)east = h(s + t. 0) P(8)a.(e) Let = h( h(Jo. 0) exist and are finite.1]) .t}t>o = h(s. a) . + v)(B([a] .4). Proof In the Markov additive sense of (6. a) h(s + t.3(s + t)[D(8 +t)[a] . a) = exp I f t3(v)(kv)[a) .9) eastt. St)} . RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et. h(s + t. a) = h(s. it then suffices to note that E(8)Le.* (a) h(s. 9) is a P ( 8)martingale with mean one.1)dv l og E(8) et where atetk•(a) h(t.2.3. u Remark 6. a) et. St)} and .t} is a multiplicative functional for the Markov process { (Jt. so that obviously {Lo. With g the infinitesimal generator of {Xt} = {(Jt..1]) .c* (e) {Le.0(s + t)dt[B(8+t)[a] .t = 1 by Theorem 6.5.1)dv  o h(t. According to Remark 11. E (8)east (a +.log h(s. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11.6. a) Corollary 6.4 For each 0 such that the integrals in the definition of h(t .5 The formula for h(s) = h(s. we can write Lo Jt.1].adt +. at + f log h(s + t. dt log E(8)east a + f3(s + t) [B(8+t) [a] .t. B) eoSt t.(8)east 179 = = = = = E(8)east (1 .9 as follows.6 . a).
3(s)h(s) + h'(s) +.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. J s [. That is.2. St)} with governing probability measures Fes).tc] dv} (normalizing by h(0) = 1).7 When a > yo. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + .f. (iv) finally.a .T.1) .'y). That is.g.3(s)dt • B(s)[a]h(s) = gha(s. Proposition 6. . cf. For each 0 satisfying the conditions of Corollary 6. MARKOVIAN ENVIRONMENT ha(s. Proof (i) Check that m. y solves n* (y) = 0. of St is as for the asserted periodic risk model. say. see [44] for 11 a formal proof. yo is determined by 0 = k* (70) = QB*. Bet)(dx) = ^ B(t ) (dx).0) = Kh(s). (ii) use Markovmodulated approximations (Section 6c).180 CHAPTER VI. as above E (s) ha(Jdt. However.4.3(s)B(s) [a]h(s). ry)) at which n* (a) attains its minimum. Lemma 6 . such that for any s and T < oo. it follows by Theorem II. we put for short h(s) = h(s. [70] .6 The P(s).(3(s)dt) +. A further important constant is the value yo (located in (0. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. Now define 'y as the positive solution of the Lundberg equation for the averaged model. When a = y. Proposition 6.y) = eayh(s).3.5 that we can define a new Markov process {(Jt.3(v)( Bi"i [a] . the requirement is cha(i. 0 < s < 1. 0) = h(s) + dt {ah(s) .60(t) = a(t)B(t)[0].1. P(s) (T(u) < oo) = 1 for all u > 0. B(s). ( iii) use approximations with piecewiese constant /3(s).6 ( s ) exp { 0( s )&s) [a] + tc . correspond to a new periodic risk model with parameters ex . Sdt) = h(s + dt) eadt (1 .
2).7) h(B(u).9(u))} u>0. The proof involves machinery from the ergodic theory of Markov chains on a general state space.6.9) 0(')(u) = h(s.9) and noting that weak convergence entails convergence of E f (^(u).2. has a unique stationary distribution. q) = eryx/h(q)). xEJ 0 (s)b(8)(x) > 0. the Markov process {(^(u). a) TI h(9(u). a) a > ry0 (6. the mean number of claims per unit time is p« 181 = Jo 1. Wu).1. e(cc)) Letting u > oo in (6. u which is > 1 by convexity. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11. 9(u)) for any bounded continuous function (e. (6. Here and in the following. T(u) < (6. B(oo)). 1). T) = h(s.9 Assume that there exist open intervals I C [0. considered with governing probability measures { E(8) }E[ . and no matter what is the initial season s.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. f (x.g. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u. have components with densities b(8)(x) satisfying inf sEI. Lemma 6 . Corollary 6.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. and we refer to [44].10) Then for each a.1) the distribution of (l: (oo). which is not used elsewhere in the book. we need the following auxiliary result . a) e«uE(8 ) e «^ .8) (6. we get: . ^(u) = ST(u) . s E I. RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. 0(u)) * (b(oo).4. say s0. J C R+ such that the B(8).
At this stage .1 In contrast to h.6 for the Markovmodulated model: Theorem 6 . Among other things.1.16. Vi(8) (u) . (6.ir) } Plots of h for different values of A are given in Fig.11) Note that ( 6. where e. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3.) C = E1 h(B(oo)) u + oo. For our basic Example 6 .1. 11 7/'O (u) < C+°)h(s) ery".10) of Lemma 3. A=1/4 A=1 A=4 0 Figure 6.182 CHAPTER VI. 10 shows that certainly ry is the correct Lundberg exponent. 1. illustrating that the effect of seasonality increases with A. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs . this provides an algorithm for computing C as a limit.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.10 Under the condition (6. where C(o) = 1 + info < t<i h(t) . which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c. 6. Theorem 6 .Ch(s)ery".W. MARKOVIAN ENVIRONMENT Theorem 6. Noting that ^(u) > 0 in ( 6.9). it does not seem within the range of our methods to compute C explicitly.
e7 ( yx)B(t)(dy) > Then for all $ E [0.(8) (u.42 so that 183 tp(8) (u) < 1.13) Elementary convexity arguments show that we always have ryy > Y and ay > ry. r.13 to our basic example. 1 (6.7x j dx _7x } _ 6w + 6(1 .16 In order to apply Theorem 6.11 as well as it supplements with a lower bound. we obtain Co) = 1.w)e4u dx 9w + 7(1 .3x + (1 . Consider first the timedependent version of Lundberg's inequality.7e . RISK THEORY IN A PERIODIC ENVIRONMENT Thus. the proofs are basically the same as in Section 3 and we refer to [44] for details. C_h(s)e7u < V. in our basic example with A = 1. we first note that the function fu° exu {w • 3e .w) . yu) 000 (u) . where ay is the unique solution of W(ay) =y• (6.w)e4u .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.(s)(u) < C+h(s)e7". #c( ay) < 0 when y > 1/tc'('y). Theorem 6. T) and replace the Lundberg exponent ry by ryy = ay . ay) • (6. we substitute T = yu in 0(u.47r sin 27rs + 167r cos 47rs .42 • exp {J_ cos 27rs .(ay) > 0 when y < 1/ic' (7). (6.0(8) (u+ yu) (6. We state the results below. Lundberg's inequality can be con siderably sharpened and extended.12) As for the Markovian environment model. Just as in IV.15) The next result improves upon the constant C+) in front of eryu in Theorem 6.g. 1 ) and all u > 0.yr.6.4. (ay). e.167r I Cu.17) (6.. Theorem 6 . .w ) • 7e u{w • 3e3x + ( 1 .12 Let 00)(y) 1 Then info < t<i h(t. whereas ay < y.14) < C+)(y)h(s) e7yu.
9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.. .T) < C+('Yo)h( s.1 sin 2irs + 16_ cos 47rs . n}.20 •exp { 2n cos 27rs .4^ sin 2irs + 16^ cos 41rs . Thus.19 } 0 <8<1 8 + cos 21rs Thus e.'Yo)e (6. Some of the present proofs are more elementary by avoiding the general point process machinery of [44].20). Then 7oudT . 1) for the environment). 1).1 sin 27rs + 1 cos 47rs . the nth Markovian environmental process {Jt} moves cyclically on {1. . MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0. 14 Let C+('yo) be as in (6.16. completing a cycle . such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0.\ = 0 . with s the initial season. and let 8 = er' (Y0). This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.cos 27rs . .184 CHAPTER VI.16) with 'y replaced by yo and h(t) by h(t.013. 0 <'p(8)(u ) . exp 2^ cos 21rs . where the environment at time t is (s + t) mod 1 E [0. yo). and in fact.g.18) Notes and references The material is from Asmussen & Rolski [44].\ 3 C+ = sup 6 exp { A (. Thus C_ = 2 inf ex cos 2irs . Finally.66.I eu. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. Of course.L sin 27rs + 1 I cos 47rs .0.. but thereby also slightly longer.181 s(u) < 1.013.19 I eu. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'. we have the following result: Theorem 6 . C+ = 1.66. 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.(8)(u. for A = 1 (where 3 e0. 1/i18 1 s (u) > 0.
(6. We let {Stn)} (6.7. This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. .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. 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. 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. it is desirable to have formulas permitting freely to translate from one setting into the other. but others are also possible.20) be the claim surplus process of t>o the nth approximating Markovmodulated model.jEE of the risk process. one simple choice is Oni = 0( i . To this end. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). Thus. since the settings are equivalent from a mathematical point of view. DUAL QUEUEING MODELS 185 within one unit of time on the average . M(n) = Supt>o Stn). Bi. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i.1 ((i 1)/n) ) and Bni = B . so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. Notes and references See Rolski [306]. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. z/'i (u.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. Let 0j.
• The queueing discipline is FIFO.1). Taking probabilities and using the stationarity yields 7riPi(T(u) < T. T) = 7ri 1 P. (VT > u I JT = 2). Jo = i.2).4) where 0* = >jEE 7rj/3j. Jo = j.T(V > u I J* = i). JT = j} and {VT > u. JT = Z).1) follows. ii (u) = it /3 P(W > u. let T . and for (7. I* = i)..n(VT > u. (7. J* = i) = P.3. JT = i) = P(V > u. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. JT = j) = LjPj (VT > u.. The actual waiting time process 1W1. JT = i} coincide. 2 .186 CHAPTER VI. The first conclusion of that result then states that the events {T(u) < T.=1 . JT = j) = 7rjPj(VT > u.1) over j. Proposition 7.P(V > u..3) 7ri where (V. JJ = i). Jt ).1 Assume V0 = 0. J*) is the steadystate limit of (Vt. Proposition 7. (7. J* = i). Proof Consider stationary versions of {Jt}o<t<T. JT = i) = 'P. (7.2) and use that limF (VT > u. Now let In denote the environment when customer n arrives and I* the steadystate limit. In particular. For (7.2) Oi(u) = 1.1) 7ri In particular.0i (u . 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. (7.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. {Jt }o<t<T• Then we may assume that Jt = JTt. Then Pi(T(u) < T. . . just sum (7. J* = i) for all j. and (7.3). I* )3i P(V > u.oo in u (7.
P(. 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.I.o. that /3(t). and one has PI'>(rr(u) < T) = P('_T)(VT > u).4). Lemoine [242]. (7. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. see Regterschot & van Doorn [123]. with (7. T].3). I*) with the timeaverage . we have 1: I(W. Taking the ratio yields (7. The relation (7. P(W >u. a paper relying heavily on classical complex plane methods. on average 0*T customers arrive in [0.I *=i).7) of that paper. B(t) have been periodically extended to negative t). and (7. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. and further references (to which we add Prabhu & Zhu [296]) can be found there.T)(T(u) <T) = P(8)(VT > u).g. . n=1 N However.7) (7. >u. see in particular Harrison & Lemoine [186].l. With {Vt} denoting the workload process of the periodic queue.5) follows from (7.4) can be found in Regterschot & de Smit [301]. [243]. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. Proposition 7.. In the setting of the periodic model of Section 6. N * oo. if T is large. a general formalism allowing this type of conclusion is 'conditional PASTA'. and of these. on average /32TP(V > u.6) (7. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. P(1')(r(u) < oo) = P(')(00) > u).8) For treatments of periodic M/G/1 queue.4) and (7.7.1 is from Asmussen [16].3) improving somewhat upon (2. I* = i. and Rolski [306]. J* = i) see W > u. A more probabilistic treatment was given by Asmussen [17].=i) a4.
This page is intentionally left blank .
t] are Nt At = Ui (1.T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite. and the evolution of the reserve may be described by the equation Rt = u . are i. Zt As earlier.d. 189 . Thus. finite horizon. resp . However . and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). with common distribution B and independent of {Nt}.1) (other terms are accumulated claims or total claims). the aggregate claims in [0. the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r.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 .i. and that the claim sizes U1. {Rt} moves according to the differential equation R = p(R). U2. z/i(u) = F IinffRt< 0IRo=u 1 (1.. Thus in between jumps.T) = F(T(u) < T). . i&(u.At + p(R8) ds.2) tk(u..6.
p2. Proposition 1. but when the reserve comes above v. Hence in terms of survival probabilities. it seems reasonable to assume monotonicity (p(r) is u . but assume now that the company borrows the deficit in the bank when the reserve goes negative. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . or o(u) < 1 for all u.'(u)) > 0 so that V'(v) < 1. oo) is given by i (u + p/S). the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. 1 . Now return to the general model. Example 1. say e.e. In this situation. pi > p2 and p(r) = One reason could be competition. Assume 0(u) < 1 for some u. rather than when the reserve itself becomes negative. If Ro = v < u. Thus at deficit x > 0 (meaning Rt = x). dividends are paid out at rate pi . we get p(r) = p + er. Example 1.4 Either i. when x > p/S.i(u) = 1 for all u. where one would try to attract new customers as soon as the business has become reasonably safe. there is positive probability.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1.p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. that {Rt} will reach level u before the first claim arrives. we can put Rt = Rt + p/S. That is. P(r) _ p + e(r . i. and the probability of absolute ruin with initial reserve u E [p/S.Vi(v) u > e(1 .2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. However. No tractable necessary and sufficient condition is known in complete generality of the model.2.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. say at interest rate b. Proof Obviously '(u) < ilb(v) when u > v.190 CHAPTER VII.
1.3µB for all sufficiently large r. if and only if V)(u) < 1 for all u. We next recall the following results. Hence ik(u) < 1 for all u by Proposition III. that u zPp(u . cf. 296297): Theorem 1.1. { Vt} remains at 0 until the next arrival).e. let uo be chosen such that p(r) < p = /3µB for r > uo.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1. In case (a). and hence by a geometric trials argument. This is basically covered by the following result (but note that the case p(r) .2(d)). let uo be chosen such that p(r) > p = 0ILB + e for r > uo.4. one can couple the risk process and the storage process on [0. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. Let Op(u) refer to the compound Poisson model with the same 0. appealing to Proposition 111. [APQ] pp. we have z/i(u) <p(u .1. instead of (1. (1. hence Rt < uo also for a whole sequence of is converging to oo. obviously infu<uo z/'(u) > 0.6) .2(d).T) = P(VT > u). In between jumps.5) and the process {Vt} has a proper limit in distribution .+ p(r) exists..2) for r sufficiently large so that p(oo) = limr. and P(Rt + oo) > 0. That is. Proof This follows by a simple comparison with the compound Poisson model. Starting from Ro = uo.1. {Vt} decreases at rate p(v) when Vt = v (i. B and (constant) premium rate p. then l/i(u) < 1 for all u.2) we have t Vt = At . say V. . In case (b).5 (a) If p(r) < /.uo) < 1.f p(Vs) ds.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. Proposition I1I. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. then ?(u) = 1 for all u. V = p(V)).3.2 once more.1 and increasing in Example 1.uo) and.b(u. In particular. (1. However. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. (1.6 For any T < oo. which was proved in 11. Then 0(u) = P(V > u).4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. Theorem 1. Then if u > no. INTRODUCTION 191 decreasing in Example 1.
6 applicable.s. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. say if p(r) goes to 0 at rate 1 /r or faster as r j 0.Sx}. Jo AX) (1.8) as g(x) = p 1 {yo13e_6x +.h. u Define ^x 1 w(x) Jo p(t) dt.8) is the rate of downcrossings (the event of an arrival in [t. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. say when {Vt} is in state y. It follows in particular that 0(u) = fg(Y)dy.8) Proof In stationarity. B(x) = e.6w(x) .Qw(x) . t + dt] if and only if Vt E [x.8 Assume that B is exponential with rate b. of (1. Now obviously.6x and that w(x) < oo for all x > 0.7 p(x)g(x) = tofB (x) + a f (x . It is intuitively obvious and not too hard to prove that G is a mixture of two components.y. Corollary 1. In view of the path structure of {V t }. say. and is succesful if the jump size is larger than x . (1. the flow of mass from [0. where g(x) = p( ^ exp {. oo). (1. x] to (x.h. Note that it may happen that w (x) = oo for all x > 0. 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. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. for the storage process {Vt}. of (1. one having an atom at 0 of size 'yo. Then the ruin probability is tp (u) = f' g(y)dy. Oeax f x e'Yg (y) dy } = p) eaxa(x) .8) as the rate of upcrossings. the l. yo ^ 1 + oo Q exp {. Considering the cases y = 0 and 0 < y < x separately.s.Sx} dx.7) Proposition 1. An attempt of an upcrossing occurs as result of an arrival. x + p(x)dt]). and the other being given by a density g(x) on (0.y)g(y) dy.192 CHAPTER VII.9) Proof We may rewrite (1. say. we arrive at the desired interpretation of the r. we thus need to look more into the stationary distribution G. oo) must be the same as the flow the other way.
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
with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.pa.3/E) By the memoryless property of the exponential distribution.3. and the c. Proof 2 We use Theorem 2.13 /E) r (. 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 . where V is Gamma(b.e.2) follows by elementary algebra. The process {St} corresponds to {Wt} so that c(a) or2a2/2 . r (j3/E) In particular. 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.f.V < x)]0 + f P(V > p/E ) + eby fv (p/E . of Z is IogH[a] = f ytc(y)dy = e fa (0. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C . RESERVEDEPENDENT PREMIUMS u from which (2.2) follows by elementary algebra.V.g. . From ic(a) = .2y +µ ) dy .x) dx e.a) .5 The analysis leading to Theorem 2. then {Rt} is the diffusion with drift function p+Ex and constant variance a2.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.200 CHAPTER VII.3a/ (5 . i. As an example./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . 13/E). assume that {Wt} is Brownian motion with drift µ and variance v2. /^ u Example 2 .pa.
Goldie & Griibel [167]. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) .8.p*. write Vi* (u) for the ruin probability etc. Corollary 2. for a martingale proof. The formula (2. Q2/2E). [357]. [282]. It must be noted. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. Some of these references also go into a stochastic interest rate.g..i. see e.1) . [129].d. of the form Ei° p"X" with the X„ i.3) was derived by Emanuel et at. that the analysis does not seem to carry over to general phasetype distributions. as in the proof of Proposition 2. it is also used as basis for a diffusion approximation by these authors.4 is classical. Paulsen & Gjessing [286] and Sundt & Teugels [356]. Paulsen [281]. or to nonlinear premium rules p(•). it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. Gerber [155].3 is from Harrison [185].3. Gerber [157] p.2 is a special case of a perpetuity. write y* for the solution of the Lundberg equation f3(B[ry *] . se e. A r. Emanuel et at. 3 The local adjustment coefficient.e. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. Z is normal (p/E. and recall Lundberg 's inequality .g.v. Further studies of the model with interest can be found in Boogaert & Crijns [71]. and since RT = 0 by the continuity of Brownian motion. Delbaen & Haezendonck [104]. [283]. 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). [129] and Harrison [185]. however. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case).Y*p* W*(u) < ery*u = 0. not even Erlang(3) or H3..
(3.ap(x). as solution of the equation n(x. and that p(x) * oo. Then lim sup u>oo u and e E''p(r) + 0. 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. (x. it holds that f3[s] T oo.4) we assume existence of y(x) for all x. B(x) > C(2)e(ao+f)x for all x.>o 7(x) > 0.i)eex. (3. and (for simplicity) that inf p(x) > (3µs .2) such that p(x) < c(.1. Let y* < So. 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). which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.202 CHAPTER VII..*(u) . i. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. i. x * oo.e.1 ).5) which implies inf.'y ( x)) = 0 where r. choose uo such that p( x) > p* when x > u0E. When u > uo. x>0 (3. 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) .. For the last asssertion . c(. 1) and for a given E > 0. as will hold under the steepness assumption of Theorem 3. e(1o+e)2 (x ) u > 00. If 60 s f 6o.1) . a first step is the following: Theorem 3 .e. 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 . Letting first E * 0 and next ry * T 5o yields the first statement of the theorem.w (u) J dt > c(3)eeu v 1 p(u+ t) .C*ef*". a) = f3(B[a] . choose c(.1.1 Assume that for some 0 < 5o < oo. Proof of Theorem 3. The steepness assumption and p(x) + oo ensure 'y(x) * So.3) When trying to extend these results to the model of this chapter where p(x) depends on x.log '(u)/u < ry*(1 . oo for all E > 0. log ?i(u) < < 00 JO . then log u (u) In the proof as well as in the remaining part of the section . we will use the local adjustment coefficient 'y(x).E). let p* be a in (3.
If p(x) = pis constant .2). THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. then Rte) = CRtie for all t so that V). or (b) Condition 3. Then .' (u) < eI("). (3.0 are the same. the asymptotics u * oo and c .3 Assume that either (a) p(r) is a non decreasing function of r.2 is also an approximation under appropriate conditions. Theorem 3 . (u) = O(u/e). However.1 only presents a first step. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 .6) The second main result to be derived states that the bound in Theorem 3.. The form of the result is superficially similar to the CramerLundberg approximation. ruin will occur if the claim is at least u + v. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. by cU2.e.3 to be reasonably precise and use e` (u) as approximation to 0 (u).13 is a technical condition on the claim size distribution B. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(.7) CIO Remarks: 1. UJU > x cannot have a much heavier tail than the claim U itself. 3.3. Condition 3. the result is not very informative if bo = oo. The rest of this section deals with tail estimates involving the local adjustment coefficient. For e > 0. which essentially says that an overshoot r. Then limelog l/ie (u) = I(u). .(u) > so. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V.13 below holds. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e.ea°/(ecf1)).v. one can then assume that e = 1 is small enough for Theorem 3. 2. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. Theorem 3.g. u Obviously.4)eE" Given such an arrival. Bucklew [81]). 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. {Rte)} defined as in (1. and in particular. (3. noting that in many cases the constant C is close to 1. I. 3) = (1 .
J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .6). One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) . we consider some simple examples.8 in terms of I(u) when the claims are exponential: Example 3 . As typical in large deviations theory.8.2. we show how to rewrite the explicit solution for ti(u) in Corollary 1.bx}]o + b /' oo exp low (x) .3. 3. Then y(x) = b . u .204 CHAPTER VII.(iw(x) . say.bx} dx fo 00 1 + [exp {/(3w(x) . and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.bu}.3.4 Consider again the exponential case B(x) = eax as in Corollary 1. the logaritmic form of (3.(3/p(x). 3a Examples Before giving the proofs of Theorems 3.bx} dx = 1 + J0 dodx(x) exp {. it is formally needed only for Theorem 3.bx} dx . 5.(x) dx.bx} dx 1+0. rather than eI(u)). we get = 1+ J" AX) exp {(3w (x) .1 + b f e. First.exp {/33w(u) . RESERVEDEPENDENT PREMIUMS 4.7) 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). However.
fory(x+u)dxdy ( 3. oo) with drift µ(x) and variance a2 (x) > 0 at x.5 Assume that {Rt} is a diffusion on [0. Be = e log U000 e. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.2. and (3. In particular.3. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. ry(x /b I u o e f0 °° e  e. For Theorem 3. (3.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3.e.10 or Karlin & Taylor [222] pp.3.BE. note first that the appropriate slow Markov walk assumption amounts to u..5) is infx>o 7(x) > 0 which implies that f °O .0.7) follows. 0.fo 7(x) dx /E dy > av 'yo /Edy = E (1 . u .2. BE * 0. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.I ( v )dy fo +u) dxdy . > lime log e = 0 and AE * 0. IE(u) = I(u)/e.2(X) = ev2(x) so that 7e(x) = 7(x)/e.fo 7(x)dx/Edy f . 3.f y(x)dxd y If 7(x) is increasing . we get r 00 e. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . (3.1.9 ) 11000 eI(v)dy f000 e. (X) = µ(x). and (3. 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.1/8 . (u) = I(u) + AE . Choosing yo. It is well known that (see Theorem XI.ev 0 O /E) J0 70 70 Yo This implies lim inf A.10) where AE = e log 000 e.. 70 > 0 such that 7(x) < 7o for y < yo.fa 7(x+u)dx/Edy o The analogue of (3.I ( u) fool. in the definition of AE converges to 0.9) yields e log . Similarly.3 in the particularly simple case of diffusions: Example 3.
RESERVEDEPENDENT PREMIUMS The analogue of Example 3. so our approach is to determine standard functions Gl (u). Nevertheless . I. Thus 7e(x) _7(x)/e and (3.. however . . + Gq(u) + o(G9(u))• Gi (u) It should be noted .Q/p*. lim sup Af < lim sup c log(1 . that the interchange of the slow Markov walk oo is not justified and in fact.6) exactly as in Example 3. _ .4. G..e.(u) oo. ..1 3. Ignoring 1/5 in the formula there.6/p* so that u 1 I (U) = bu ./3 1 AX dx. the results are suggestive in their form and much more explicit than anything else in the literature. G.6 Assume that B is exponential with rate S. . I(u ) = G1(u) + . this leads to (3. Then the solution of the Lundberg equation is y* = b . the slow Markov walk assumption means 5E = b/c.0/e. Further.5 for risk processes with exponential claims is as follows: Example 3 . 7(x) is typically not explicit. .5.5. we have 5 > 7o and get lim inf AE > lime log e .7o C 15 I I.5) and 7* = 5 .0) = 0. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. E+o e*O By (3.10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B.206 CHAPTER VII.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. > . .0.. 0 Now (3. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. 0.5. Of course.+1 (u) = o( 1). G. As in Example 3.
c4 = c2b 1/'/(1 .1/k). . .ry*°p*.1/a). This covers mixtures or convolutions of exponentials or. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .. phasetype distributions (Example 1. if the phase generator is irreducible ( Proposition VIII.c2 Su a dx ) Su a<1 Su .11) that b[s] * co as s f S and hence 7* T S as p* + oo. More precisely.4) or gamma distributions. y = 2 if B is uniform on (0.1) leads to .Y .8). more generally.:.12) with y > 1.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . (3. I(u) Pt. the typical case is a = 1 which holds .C2p* C2 = (3clr( a))11'.y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b. in the phasetype case . 77 = 1 if B is degenerate at 1. Here B[s] is defined for all s and B[s] .8 Assume next that B has bounded support. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 .11) with a > 0.s)C' f "o o as s T S. 1. x T 1.3. u Example 3 . say 1 is the upper limit and B(x) .7 Assume that B(x) . e. For example. 2.cs(1 .3cse7*I7(77) . ry* loge*+ g7loglogp*. It follows from (3. and hence (3. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0. Hence (3.g.1) leads to (S7T N Ocp a.x)n1.clxale5x 207 (3. u(logu + r7loglogu). fu I(u) Su .
Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.ru(TI)) .2 We first remark that the definition (3.4). (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3. g.g. assume that B(x) CO x2/2c7.c8 log . 1 0 3e. By convexity of the m .u .4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3.11) and (3.Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u). (3. of U1 + v .1) . ec78)2/2c7 dx C7 . h].css 2%rc7eC782/2.(t))dt.Ote7o( u)(u.B[7o (u)] .u is a nondecreasing function of u. 1 = E.sp(u).f.15) Proposition 3. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .4) is the formula h logEues ( Rhu) .13) We get b[s] .r„(Ti). I (u) c8u log u 0 where c8 = 2/c7. .14) for the m .1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .10 Assume that p(x) is a nondecreasing function of x.r^.. h 10. The assumption implies that ru(t) . (3. one could also have considered the increment ru (T1) . x f oo . 3b Proof of Theorem 3. RESERVEDEPENDENT PREMIUMS Example 3 .3 (B[s] .e7o ( u)(ul+u r.208 CHAPTER VII.9 As a case intermediate between (3. (3.•. 7 * .(T1)) > Ee7o(u)(ul+vr»(Ti)).f.log p*. of the increment in a small time interval [0.12). this is only possible if 7o(v) 2 7o(u)• . Hence for u<V.
u > tp(u). note that the assumption implies that ru(t) . We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3. Hence „/.3. (3.17) from which the theorem follows by letting n + oo. it is easily seen that fu x7o(y)dy < xyo (u).u[70(u)] fo eyo(x)dx . Then (u) < efo Yo(x)dx.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).ru(T1) < x).10(b): Theorem 3. fa 7o(y)dy < u7o(u) < xyo (u) for x > u. THE LOCAL ADJUSTMENT COEFFICIENT For (b). the case of 7 then follows immediately by Proposition 3.(n+l) (u) 1 . Hence 1 = EeYo(u)(U1+uru(T1)) < E. u We prove Theorem 3.e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] . this is only possible if yo(u) > 7(u).(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .17) shown for n and let Fu(x) = P(U1 + u . Separating after whether ruin occurs at the first claim or not.4) considered as function of 7 is convex and 0 for y = 0. Assume (3. Also.2 in terms of 7o. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.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.Fu(u ) + J ^(n)(u .7o (u)p(u)• Since (3.1) .11 Assume that p(x) is a nondecreasing function of x. we obtain „I.
2.11 is sharper than the one given by Theorem 3.2).n) pn niE (u /n) n n_1 n.15).n.10(a) for some of the inequalities.n = ku. 0.I. P k. {RtE)} (starting from u = un.. To this end. Y*u /E.n u k}1.n. Also. by €U=. For these reasons. Proof For ruin to occur. and here it is easily seen that yo(u) .n.x/n. Further.n (starting from u/n) without that 2u/n is upcrossed before ruin.E (u/ n) Y'E (un . 0. given downcrossing occurs. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3. the value of {R(E)} at the time of downcrossing is < unl.: y(u). let Op*. resp. for either of Theorems 3.e (u) = v'.12 lim sup4^o f log O. y* evaluated for p* = Pk.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. we have chosen to work with y(u) as the fundamental local adjustment coefficient.3 is required. and. pk n = uk_l.nbe C*.3). (un2.. 3. (3. 3c Proof of Theorem 3.3/e and U.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x . C*e where the first equality follows by an easy scaling argument and the approximation by (3. op*.3). Lemma 3..n inf n uk1..210 CHAPTER VII.n AX). (u) < I(u).. the probability that ruin occurs in the CramerLundberg model with p* = pn.n) > k =1 II v ^k n. in accordance with the notation i/iE (u). define uk.2. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3. yo(u) appears more difficult to evaluate than y(u). (u).n. we used also Proposition 3.n = sup p(x)..n so that n.E ( u/n) ^•e. However. The probability of this is at least n n. ryk.E (u/n) Now as e .10(b ) that the bound provided by Theorem 3.n <Z auk}l. 0 It follows from Proposition 3.E (u/n). Let Ck. in . W O .E (u) denote the ruin probability for the classical model with 0 replaced by .n) must first downcross unl.
18) (ii) the family of claim overshoot distributions is stochastically dominated by V. It follows that n log V'C (u) k =1 log Ypk.F (2u/n).nu /fn( 1 Ck  e. B(x) (3. .12 completes the proof. 11 Theorem 3.E (urn) < \ *I. Indeed . THE LOCAL ADJUSTMENT COEFFICIENT particular.n = sup ?'(x). In case (b). 211 Clearly. 3 now follows easily in case (a).. 0 with n and u fixed. in obvious notation one has tC (x) = y(x)/e. 40 Combining with the upper bound of Lemma 3.a( u)z. ne7k. uk_1. /' (u/n) 'T nk.n. since ry' is an increasing function of p'..E (u/n) OP +^p•. i. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). k=1 k=1 n u _ nE7 k.e.E (u/n) Op•.nu/en(1 + where o(1) refers to the limit e .! (u/n) n n m 7k. so that Theorem 3.n cE (2u/n) Ck ne7k.13 There exists a r.3.n .nu /En) o(1)). for all x .19) ..i.7k. v. *p•. we need the following condition: Condition 3.. also ryk.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).n + 0(1).nk=1 limsupelogv). (3. 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. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof.log Ck.n <X<Uk.
EV) = EiI 1 . u/n) < oo] . Write EO.18) for the last equality)... u /n) < oo] l = = < E [OE (u/n .E(E) (u.v.5) and the standard formula for b(0). T(E) (u.n V. u /en 0(i) _n so that E2 < e2ryl nu/En0(1).E (2u/n .of:>2 in n(x).( .nu/En0(1) .eV) • P (T(E) (u.^(E) (u. T() (u. u/n) < oo) . v ) = v . infx>2u /n P(x) . N with EN < 1 = infx>2u/nA(x) = 0(1). u/n)) .E (u/n . .EV) = e.EV) = El + E2. (R. RESERVEDEPENDENT PREMIUMS To complete the proof.^'' = E [ .n < ery1. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. (u/n . u/n) < oo] E [OE (u/n . u/n)) I T(E) (u. P (T(E) (u.1 n.V) = e71 nu/Eno(l) (using (3. Then the standard Lundberg inequality yields El < E?. V < u/En] + P(V > u/En) (u/En . Then Y'E (u) ^(E) (u. (3..nu /EnE [e71..212 CHAPTER VII. v ) = inf { t > 0 : R(c ) < v R) = u } . Ei + E2 < e71.E (0) cf. The probability of ruin in between two downcrossings is bounded by Epp .2y 1 ' . ) (u u /n)) . let v < u and define T(E) (u.R<) (u v). For E2.QEU 1 . u/n) < oo) EV). where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest. (u/n .
1. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . s).7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. l o JJJ o . Typically. T) is maximized over T by taking T as the time for (3.r.3. whereas the most probable path leading to ruin is the solution of r(x) _ k (x. Bucklew [81]).21) (the initial condition is r(0) = u in both cases).4) and the prime meaning differentiation w. Comparing these references with the present work shows that in the slow Markov walk setup.g. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u. s) as in (3. 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.T) = P „(info<t <T Rt < 0) via related large deviations techniques. Similarly. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u.21) to pass from u to 0. [89].t.J y(Rs)dR. 0 and b T 00 are interchangeable in the setting of [89]. one can in fact arrive at the optimal path by showing that the approximation for 0(u.)Ui } .7(x)) (3.13. it might be possible to show that the limits e .J r(Rs)p(R.20) (with ic(x.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof.) = exp . Whereas the result of [122] is given in terms of an action integral which does not look very explicit. the risk process itself is close to the solution of the differential equation r(x) _ r (x. 0 ) (= p(x) . the results are from Asmussen & Nielsen [39].=1 J An approximation similar to (3. u Notes and references With the exception of Theorem 3. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . the approximation (3. Djehiche [122] gives an approximation for tp(u.3EU) (3. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.)ds + Y(R2. .7) then comes out (at least heuristically) by analytical manipulations with the action integral.
214 CHAPTER VII. .3. however. the exponential distribution ). see XI. e.g. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . 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.. We should like. For different types of applications of large deviations to ruin probabilities .
and not in other cases. We often write p for the number of elements of E. More precisely. 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. if a problem can be solved explicitly when the relevant distributions are exponentials. on Eo = E U {A} where A is some extra state which is absorbing. refers to the case Jo = i. A distribution B on (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. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. Typically. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . P. that is. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient.1) is 'Here as usual . 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 . if v = (vi)iEE is a probability distribution. Note that since (1. 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.
that is. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A.e. an exponential distribution with rate parameter .0 = t11. and we have t = Te. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. i.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. The initial vector a is written as a row vector. (1. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. Then a = a1 = 1. T) (or sometimes just (a. B(t) = Fa(^ < t ). E = {i.216 CHAPTER VIII.1 The phase diagram of a phasetype distribution with 3 phases. C is the lifetime sup It > 0 : Jt E E} of {Jt}. Here are some important special cases: Example 1 . T is a subintensity matrix2. a. k}. i. 0 2this means that tii < 0. Equivalently. t1 = /3.2) The interpretation of the column vector t is as the exit rate vector. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. tij > 0 for i 54 j and EjEE tij < 0 .3. j.e. 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. We now say that B is of phasetype with representation (E. In particular.1 Suppose that p = 1 and write .
.3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51.1)!e Since this corresponds to a convolution of p exponential densities with the same rate S.. p}... 6. the EP distribution may be represented by the phase diagram (p = 3) Figure 1.. 0 •. 0 S 6 Example 1.1.. a = (1 0 0 . 00)) S s o . .x i=1 Thus E _ Si 0 T 0 S2 0 0 ... 0 0 0 0 S 6 . 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •..2 corresponding to E = {1...2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p. PHASETYPE DISTRIBUTIONS 217 Example 1. 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 .. . . 0 SP 0 and the phase diagram is (p = 2) . . . .. so that the density is P E ai6ie6..
the Erlang distribution is a special case of a Coxian distribution. dp. The basic analytical properties of phase type distributions are given by the following result . Then: (a) the c.f. Then for i . ds^ = ds' = ttlaj + tikpkj.1 tP1 1 Figure 1. (c) the m.218 CHAPTER VIII. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} .g. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3.4 For example. Theorem 1 . p:. 36) yields s d.e.d. T). a.3 . MATRIXANALYTIC METHODS Figure 1.t2 yt bP. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. the restriction of P8 to E.1)"n! aT"e. the backwards equation for {Jt} (e. 5 Let B be phasetype with representation (E. [APQ ] p.f is B (x) = 1 . i. j E E.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. (b) the density is b(x ) = B'(x) = aeTxt. .g.3 0 Example 1 .aeTxe. see A.
jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute).5) as hi(tii + s) = ti  t ij hj. the solution is P8 = eT8. B(n)[0] = _ Alternatively. .f. and since b[s] = ah. (Jx E E) = this proves (a).tii we go to A.p..s I .g.6) .f. this means in vector notation that (T + sI)h = t.. = aPxe. 1. Alternatively.n1t = (1)nn!aTn1Te (1)nn! aTne.5) Indeed . i. we arrive once more at the stated expression for B[s]. (1) n+1n!aT . After that.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . the rule (A. For (c). hj .n lt .g. PHASETYPE DISTRIBUTIONS 219 That is. define hi = Eie8S. and (b) then follows from 1: aipF.g. of the initial sojourn in state i. d8 P8 = TP8. d" dsn a (.T) 1t. Part (d) follows by differentiating the m.f. h = (T + sI)1t. ti/ .tii is the rate of the exponential holding time of state i and hence (tii)/(tii . j#i jEE tijhj + his = ti.s) is the m . Since 1 .f. for n = 1 we may put ki = Ei( and get as in (1.p. we i w. in which case the time to absorption is 0 with m . Rewriting ( 1.5) ki = 1 + tii L jj:Ai tii (1. Then h tit ti + ti3 h j .tii tii . tij / . or w. and since obviously P° = I.1 ) n +l n ! a (s I + T ) .T) 't = (.tii and have an additional time to absorption either go to state j which has m . i.B(x) = 1'a (( > x) = P.g.1.e.s j# tii i (1.
T= 2 111 so that 2 2 Then (cf. 0 Example 1. there are some examples where it is appealing to write T on diagonal form.h. another the case p = 2 where explicit diagonalization formulas are always available. are idempotent. 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.s. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x . Example A3.220 CHAPTER VIII. Consider for example 3 9 a= (2 2). MATRIXANALYTIC METHODS which is solved as above to get k = aTle. 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."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer. making the problem trivial. see the Appendix.
(1.11aDD. where the initial vector a is substochastic.4b for definitions and basic rules): Proposition 1. i. This is the traditional choice in the literature. a random variable U having a defective phasetype distribution with representation (a. then the matrix m. • The phasetype distribution B is zeromodified. i.29) and Proposition A4.1. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).e a mixture of a phasetype distribution with representation (a/llall.7 If B is phasetype with representation (v. . or one just lets U be undefined on this additional set. 0 Sometimes it is relevant also to consider phasetype distributions.T) with weight hall and an atom at zero with weight 1 .hall. 5 and serves at this stage to introduce Kronecker notation and calculus (see A.f. 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).T).7) Proof According to (A. There are two ways to interpret this: • The phasetype distribution B is defective.e 11BIJ = 1laDD < 1. and in fact one also most often there allows a to have a component ao at A.g. < 1. T) is then defined to be oo on a set of probability 1. hail = E=EE a.4. 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.
see his book [269] (a historical important intermediate step is Jensen [214]). .4c).8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. h can be chosen with strictly positive component.g. 77 > 0 and k = 0. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. the result follows (with C = (ah)(ve)). (1.f.8 Let B be phasetype with representation (a. 1. (or Laplace transform) are often used where one would now work instead with phasetype distributions. All material of the present section is standard. cf. No satisfying . Other expositions of the basic theory of phasetype distributions can be found in [APQ]. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . 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. cf. we give a criterion for asymptotical exponentiality of a phasetype distribution B. the text is essentially identical to Section 2 of Asmussen [26]. but in many practical cases. assume that T is irreducible . 0 Of course. Rolski. here k = p1).q be the eigenvalue of largest real part of T. 2.f. Notes and references The idea behind using phasetype distributions goes back to Erlang. See in particular the notes to Section 6. The Erlang distribution gives an example where k > 0 (in fact. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. In older literature. B(x) . Using B(x) = aeTxe . Lipsky [247]. the conditions of Proposition 1.8) Proof By PerronFrobenius theory (A. distributions with a rational m. not only in the tail but in the whole distribution. let v. Schmidli. Schmidt & Teugels [307] and Wolff [384].1 of the Appendix. Example A5.Ce7'. but todays interest in the topic was largely initiated by M... the Erlang case). Then the tail B(x) is asymptotically exponential. let . where C. Here is a sufficient condition: Proposition 1. x * oo. and we have eTx . but the relevant T is not irreducible.222 CHAPTER VIII.hve7x.F.8).g. Neuts. one has k = 0. T). i is real and positive. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. In Proposition A5. v.
or the density is available ) is. A related important unsolved problem deals with minimal representations: given a phasetype distribution . with common distribution B and define4 U(A) = E# {n = 0.1 of the Appendix. as the lifetimes of items (say electrical bulbs) which are replaced upon failure.. U1<t < U1+U2..1. For this reason. we refer to U as the renewal measure.. known.f.. .. (2.. 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 +. + U0 is 0 . Let U1.. .1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together.r. n=O We may think of the U. if U is absolutely continuous on (0. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. but is in part repeated below. and U(A) is then the expected number of replacements (renewals) in A.2. 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. JtJt1) Then { 0<t<U1 .. . The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. however.g. but nevertheless.d.: U1 + .T)....1 Consider a renewal process with interarrivals which are phasetype with representation (cr. Jt={Jt?ul}. the renewals form a Poisson process and we have u(x) = 0. Lebesgue measure. oo) w. If B is exponential with rate 0. Then the renewal density exists and is given by u(x) = ae(T+ta)xt. we denote the density by u(x) and refer to u as the renewal density.+UnEA).t. +UnEA} 00 = EEI(U1 +... U2. be i.i. is Markov and has two types of jumps .
2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. and let µB = aTle be the mean of B. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. T). and the distribution of Jx is ae ( T+t«)x. Corollary 2. + Uit_1 where s. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter .1) remains valid for that case. see Fig. fi(t) U2 U1 .. Equivalently. Then the lifetime is zeromodified phase type with representation (a. Proof Just note that { it } is a governing phase process for the lifetime. is the first k with Uk = 00. the density is veTxt = B(x)/µB. since Uk = oo with probability 1 . that is.1) follows by the law of total probability.T) where vt = ae (T+ta)t .e. This is defined as U1 + .224 CHAPTER VIII. However. IIafl < 1.. and the jumps of the first type are governed by T.T + ta). i.1 Corollary 2.T). as the time of the last renewal. B is defective . (b) £(t) has a limiting distribution as t * oo.3 Consider a renewal process with interarrivals which are phasetype with representation (a. this is welldefined. Hence ( 2.T) where v = aT1 /µB. u The argument goes through without change if the renewal process is terminating.1. The renewal density at x is now just the rate of jumps of the second type. 2.e. which is phase type with representation (v. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj . i.IIBII which is > 0 in the defective case. . and hence ( 2. u Returning to nonterminating renewal processes . define the excess life e(t) at time t as the time until the next renewal following t. which is ti in state i. . the lifetime of the renewal process.U1 U3 U2 U3 U4 Figure 2. Hence the intensity matrix is T + ta.
i. (2. u Example 2 . According to Example A3.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0.1.) ( t2 ) . Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. cf. The renewal density is then aeQtt = (al a2) ( 7i 7"2.4 Consider a nonterminating renewal process with two phases. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 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. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . Al. we first compute the stationary distribution of Q. (ii) First check the asserted identity for the density: since T. The time of the next renewal after t is the time of the next jump of the second type. hence e(t) is phasetype with representation (vt.T) where vt is the distribution of it which is obviously given by the expression in (a).2. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. the unique positive solution of ve = 1.e. T1 and eTx commute.e.q2.6.
A = 25. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .4 yields the renewal density as u(t) = 2 (1 . Then Q= 0 55 )+(1o)=( j ad ).`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . )t (51 .226 CHAPTER VIII.e2bt) 13 Example 2 . MATRIXANALYTIC METHODS e.(biaz + aza. and Example 2. . t1B 0 Example 2 . Hence 7r = (1/2 1/2).4 yields the renewal density as u(t) = 5152 e. The present treatment is somewhat more probabilistic.6 Let B be hyperexponential.t2) .5 Let B be Erlang(2). and Example 2.a27r1) (t1 .a27rl) (tl .t2) 1 + eat (a17r2 .52a1.tl) 7r2t2 + eat (a17r2 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].
T). Since the results is so basic. {St} the claim surplus process. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. For (b). The stars represent the ladder points ST+(k).3. which occurs at rate ti. with 0 denoting the Poisson intensity. and M is zeromodified phasetype with representation (a+. Thus the total rate is tip + tia+. 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.f3aT1.1 on the next page. B the claim size distribution. Corollary 3.1 Assume that the claim size distribution B is phasetype with representation (a. cf. 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. and if there is a subsequent ladder step starting in j whic occurs w. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). T + to+).i. T).3.2. marked by thin and thick lines on the figure. however. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. Then: (a) G+ is defective phasetype with representation (a+. Here we have taken the terminating Markov process underlying B with two states.e. . (b) V. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. and rewriting in matrix form yields the phase generator of {my} as T + ta+. Corollary 2. The essence is contained in Fig. Next. 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. use the phasetype representation of Bo.) = F(ST(o) E •. the Markov processes representing ladder steps can be pieced together to one {my}. i. 3. a+j. add a more selfcontained explanation of why of the phasetype structure is preserved. we shall.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. T(0) < oo) the ladder height distribution and M = supt>o St. Within ladder steps. Considering the first. T) where a+ is given by a+ = . itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. r(u) the time of ruin with initial reserve u. G+(. We asssume that B is phasetype with representation (a.p.
Q = 3 and b(x) = .QaT1. 0 Example 3. 3e3x + . This is in fact a simple consequence of the form of the excess distribution B0.1 This derivation is a complete proof except for the identification of a+ with . MATRIXANALYTIC METHODS t ..M {mx} ST+(2)  S . see Corollary 2.7)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 .2 Assume that .t t d kkt S. Figure 3.3. 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 )..228 CHAPTER VIII.
3. his derivation of +'(u) is different. THE RENEWAL MODEL This is the same matrix as is Example 1. The parameters of Example 3. and the argument for the renewal case starts in just the same way (cf. this was obtained in Section 3. see Section 6.and Markovmodulated models. The result carries over to B being matrixexponential.T). For further more or less explicit computations of ruin probabilities. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). For the compound Poisson model.6).4.j). 4 The renewal model We consider the renewal model in the notation of Chapter V. see Shin [340]. the duality result given in Corollary 11.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3. That is.4. see Stanford & Stroinski [351] . cf.1 can be found in Neuts [269] (in the setting of M/G/1 queues. 3. We assume p = PB/µA < 1 and that B is phasetype with representation (a. where a+ is the (defective) . T) for some vector a+ = (a+. In the next sections. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case).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+. For an attempt. Fig. T). (a) G+ is of phasetype with representation (a+.1 In the zerodelayed case. the discussion around Fig.2 are taken from Gerber [157]. with A denoting the interarrival distribution and B the service time distribution. 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 .6. We shall derive phasetype representations of the ruin probabilities V) (u). if we define {mz} just as for the Poisson case (cf.1): Proposition 4. we encounter similar expressions for the ruin probabilities in the renewal. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community.
where a(8) = aT1/PA.3 a+ satisfies a+ = V(a+). But by Corollary 2.3. the calculation of the first ladder height is simple in the stationary case: Proposition 4. cf. Also.6. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4.*'} is Markov with the same transition intensities as {mx}. (c) {mx } is a (terminating) Markov process on E. Then {m. We have now almost collected all pieces of the main result of this section: Theorem 4 .1) Proof We condition upon T1 = y and define {m. Proof Obviously. which for numerical purposes can be solved by iteration. Since the conditional distribution of my given T1 = y is ae4y. where B0 is the stationary excess life distribution corresponding to B. 4.Sy} in the same way as {mx} is defined from {St}.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).T). obviously mo = m. but with initial distribution a rather than a+.T)• Proposition 4. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. (4. The key difference from the Poisson case is that it is more difficult to evaluate a+. the Palm distribution of the claim size is just B.1).5. Then . In fact.230 distribution of mo.T). CHAPTER VIII. Fig.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy).1. Nevertheless.*} from {St+y . with intensity matrix Q given by Q = T + to+. G(') = pBo. Hence by Theorem 11. B0 is phasetype with representation (aT1/µa.
3).•. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case. It remains to prove convergence of the iteration scheme (4. (4.1 by noting that the distribution of mo is a+. Hence ^p(. .1) and a(8) _ aT.1/pA. thus .0.^(u) = a+e ( T+ta+)xe.3) Proof The first expression in (4..2. In particular .1 . by a+ = lim a +n) where a+°) .e.4. and that this is given by Proposition 4. THE RENEWAL MODEL 231 . i.2 ) follows from Proposition 4. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) ..1(b).3) (defined on the domain of subprobability vectors ... only with initial distribution a(*) for mo.0) is an increasing function of /3. i y ^ T1= y `•r Figure 4. . Furthermore . a+2) = ^p (a+l)) .^(8)(u) = a ( 8)e(T+ta +) xe. (4.2) where a+ satisfies (4.1). the maximum claim surplus for the stationary case has a similar representation as in Proposition 4. a+l ) = cp (a+°)) .M. a+ can be computed by iteration of (4. I {mx} .
which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . the normalization is equivalent to F(s) = 1.g.) = P(mTl = i. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. However.1.} can contain at most n ..2. a+ ) exists .1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). F[s] being interpreted in the sense of the analytical continuation of the m.5) Since s $ sp(T). and let &+". To this end. Thus by (4.4). n) &+n) T a+. Obviously. limn4oo a ) < a+.5 Let s be some complex number with k(s) > 0. Then each subexcursion of {St+Tl . (4. let F be the distribution of U1 . and hence we may assume that h has been normalized such that ahA[s] = 1. Then (4. For n = 0. so to complete the proof it suffices to show that &+ < a+) for all n. Similarly.T)'t • A[s] (4. 7+ ]. this implies that ahA[s] # 0.T1. Proof Suppose first Qh = sh. s ¢ sp(T).P[s] = A[s]B[s]. . To prove the converse inequality. we use an argument similar to the proof of Proposition VI. Assume the assertion shown for n . Then F[s] = a(sI .232 CHAPTER VIII. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn.ST. Theorem 4. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. 0 0 We next give an alternative algorithm.T)It. Thus .T)1t.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). (4. the corresponding right eigenvector may be taken as (sI . 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).f. Fn ).4.5) yields h = (sI . with B[s].4) whenever EeR(S)U < oo. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. In that case. both quantities are just 0 .
9) we have G+[s] = 1 which according to Theorem 1. pdhd.. (4.6. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1... yd satisfying R(ryi) > 0.type with representation (a+.6 Suppose u < 0.. Let d denote the number of phases. As in Corollary 4. . t(ry) > 0. THE RENEWAL MODEL 233 Suppose next F(s) = 1. 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 ). Q = CD1 where C is the matrix with columns hl.. Notes and references Results like those of the present section have a long history..1 has the d distinct eigenvalues .6) i=1 i=1 Proof Appealing to Theorem 4. in turn. the matrix Q in Theorem 2. 0).p1i .T)lt = sh. . In older literature . ..5(c) means that a+(sI T)1t = 1.. . . Pd with corresponding eigenvectors hl. Then G+ is phase.lt we get Qh = (T + to+)h = T(sI . hd. 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. Given T has been computed..T) = 1 ata+ = a+. Q has diagonal form d d Q = dpivi®hi = dpihivi.. we get at a(Q . This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . This gives d roots 'y.5. W v M(d) in the notation of Chapter V). 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.. Since R(s) > 0 and G _ is concentrated on (oo. we have IG_ [s] I < 1 .' that the equation F(s) = 1 has d distinct roots p1... . Pd in the domain ER(s) > 0 . . Corollary 4. Further... hd. and define hi = (piI . D that with columns p1 hl. T) with a+ = a(QT)/at.4.T)It.T)lt + t = s(sI . and hence by the WienerHopf factorization identity (A. and the solution is .. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . Hence with h = (sI T)..
234 then in transform terms CHAPTER VIII.4. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. The arrival rate in background state i is a.. For further explicit computations of ruin probabilities in the phasetype renewal case . 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI.F. and the distribution of an arrival claim is B. involving .type assumptions are basic. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. [270] and Latouche & Ramaswami [241].contained derivation). In risk theory. The matrix. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). The exposition here is based upon [18]. The number of elements of El=> is denoted by q. [119]. with representation say (a(' ).. and appears already in some early work by Wallace [377]. starting around in 1975. whereas the approach was introduced in queueing theory by Smith [350]. the fixpoint problems look like R=Ao+RAI+R2A2+ . the intensity matrix is A and the stationary row vector is ir . E(t)).) d (see. an alternative approach (the matrixgeometric method ) has been developed largely by M. T('). Here phase. see Neuts [269]. the background Markov process with p states is {Jt}. Numerical examples appear in Asmussen & Rolski [43].type assumption . That is . e. the ruin probability can be found in matrixexponential form just as for the renewal model. 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..exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. is phasetype. We assume that each B. Neuts and his students. In queueing theory. a pioneering paper in this direction is Tacklind [373]. where R is an unknown matrix. Asmussen & O'Cinneide [ 41] for a short self. It turns out that subject to the phase. see Dickson & Hipp [118]. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88].g. The solutions are based upon iterations schemes like in Theorem 4. The distribution of W comes out from the approach but in a rather complicated form . For surveys .
5. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. O. has states o. 5a Calculations via fluid models.1 In Fig.5.4. Section 5b then gives a representation along the lines of Theorem 4. The two environmental states are denoted o. 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).Vt)} obtained by time reversing the I component.1. the phase space E(°) for B. However. Diagonalization Consider a process {(It. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. and the one E(•) for B. 5. •. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. The key unknown is the matrix K.2. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. say with slope r(i) on intervals where It = i.1. states .6. 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. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. p = ql = Q2 = 2.
236 CHAPTER VIII. Thus F = {o. In the general formulation . First. F is the disjoint union of E and the Eli). 2.1))diag ) a = sa and the eigenvector b = . i E E. r(i) _ 1. Let E denote the matrix .1 A complex number s satisfies 'A+ (f3i(Bi[s] . '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.31a(l) (/3i)diag . a) = 1. whereas Ee8s' = oo for all t and all s > so where so < oo.Vt)} is then obtained by changing the vertical jumps to segments with slope 1.1) if and only if s is an eigenvalue of E.1(b) {(It . Eli) + Proposition 5. consider the vector a satisfying (A + (13i(Bi[ s] .(Ni)diag r(i. of E into components indexed by E. If s is such a number. 4. Bi[s] = a(i)(T(i) + sI)it('). 4. A claim in state i can then be represented by an E()valued Markov process as on Fig. MATRIXANALYTIC METHODS 4. V.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . Recall that in the phasetype case. a E E(i) } . we have more martingales at our disposal.1))diag + sII = 0 (5. in the fluid model Eel'. F = E U { (i. 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. corresponding to the partitioning + Epp). The fluid model on Fig . 5. 4}. 5. resp. o.1(a). a) : i E E. •. This implies that in the fluid context.92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. a) of {It}. < oo for all s. The intensity matrix for { It} is (taking p = 3 for simplicity) I A . j = 1. Second. t.
Nla(1) 0 0 T 1.A . where c.(sI . MARKOVMODULATED INPUT 237 indexed by E.1).E22)1 E21a E21a . iEE (a> of 0* 1 AI.5.sI)1t)) iag I = 0 which is the same as (5. it follows that if Qla(1) 0 0 .sI ()3i)diag .E22 . Then E21c+E22d = E21a . and let d = (sI . c = a. d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .sI.sI) (sI . For the assertions on the eigenvectors.32a(2) (/3i)diag . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .sI + E12 (sI .E22)1 E21) a = 0. Then (up to a constant) c = a. assume that a is chosen as asserted which means (Ell . Noting that E11c + E12d = se by definition.E12E22 E21 I .sI 0 0 t(3) 0 0 = 0. resp .T('))1t(i) .sI 0 0 0 T(2) .A . with Eii replaced by Eii . E(1) + + E(P). 0 . t(1) 0 0 then also 0 t(2) 0 .E21a + sd = sd.sI 0 0 0 T(3) .sI+ ((3ia(i)(T(i) .E22)1 E21a. d = (sI E22)1E21a = E ai(sI .
sv)b(v) = 0. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . v.. v) yields C{V) = e8 .. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p..4 that {e"1b(v) is a martingale . a )d(a + e8 °vpi (u . j. j. Example 5 . B2 are both exponential with rates 51 i b2.a Solving for the pi(u. a). w(u)=inf{t >O:Vtu}. w(u. I' i( V P2 (w (u) < oo.pi(u. For u. j) pi( u ..upi(u.v) = j).( u. Example 5 . it follows by Proposition II. Proof Writing Or'Alb( v) = svb( v) as (AI . v > 0.v) = Optional stopping at time w (u.v) = (j. .3 Consider the Poisson model with exponential claims with rate 5. j.v}.O./' u = e' (esiuc ( 1) .. sq with $2s. the result u follows.. e89uc(e)) (d(1) .a)d^ ). Then we get V)i (u) as sum of two exponential terms where the rates s1. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s.. v..2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. a) = Pi (Vw(u. q. Here E has one state only.238 CHAPTER VIII. j. Thus 0(u) = esu/d = pe7 ° as u should be.Q. v = 1. .5.v) = v) I.4 Assume that E has two states and that B1. . a) = (j. . . Then . d("))1 e. u) Iw(u.. a)). pi(u. MATRIXANALYTIC METHODS Theorem 5.j. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b .. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u. To determine 0 (u). a) and noting that i1 (u) = >I j.. a)).j)c v . j..v) = = p i( u .. v. define w(u.. Iw(u. c j.v)=inf{t >0:Vtu orVt=.v. .
(y. (5.x) 00 f ° (') (j) eT (yy)edx .k. 0 Theorem 5 .b (u) = Pi(M > u) = 9(i)euue.h. (') a T( However .( 2. j. oo)) j)ye. 8^')IT(j)) where e 3^') =.3j eye. j.33(e = 0 a(j))(K ®T ( j))(ej (9 I). i.5 G+(i.5. is 0 /3 f R(i .s.6 For i E E.3) .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.y = to B k7 j # k In particular. dx)Bj(y . 9(').2) the l. Proof We must show that G+ (i. U) where t(j) + t(j)O(j j = k uja. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI. according to VI.2. In terms of K.xxej • a 00 oo el . j. the Pidistribution of M is phasetype with representation (E(1) + + E(P). •) is phasetype with representation (E(i). 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.
0 1)'. Numerical illustrations are given in Asmussen & Rolski [43]. if b* [0] = b1 +b20+b302 +. some square matrix T and some column vector t (the triple (a. we have the additional possibility of a phase change from a to ry within the ladder step.... T. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability.e. For a transition from (j.1 Let b(x) be an integrable function on [0. Associated with each ladder step is a phase process. i. bn1 bn).5). say. which occurs at rate t(i). we have sofar concentrated on a claim size distribution B of phasetype.. the current ladder step of type j must terminate. 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. and lifelength M. However.) which is rational. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . that the density b(x) can be written as aeTxt for some row vector a. the ratio between two polynomials (for the form of the density. +bn0i1 0n +a10n1 +.f. a m.2) . intensity matrix U. which occurs w. the initial value of (i. see Example 1. This yields the asserted form of uja. Starting from Jo = i. i.. For j = k. t = (0 0 . which occurs at rate t^^7. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. Then b*[0] is rational if and only b(x) is matrixexponential. Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). and a new ladder step of type k must start in phase y. equivalently. t) is the representation of the matrixexponential distribution/density): Proposition 6. a) is obviously chosen according to e(`). u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17].k y.y) to occur when j # k. (6. a) to (k. and it just remains to check that U has the asserted form..240 CHAPTER VIII.. Furthermore.2.g.e.p. .. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . Bk7 . An alternative characterization is that such a distribution is matrixexponential. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j).
. u giving b(x) = E 1 ciebiz/bY. t)... 0 0 . MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . we can always obtain a real one (a. bn of the denominator to be distinct and expand the r. The converse follows from the last statement of the theorem. s) is given by 27r i . Writing b(x) = c(e( 2ni1 ) y/2 . . T= 0 0 1 . S = f c/2 0 21ri . Thus. .2). since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.47x2 3 1 0 .T)1 is so.. but as follows from Proposition 6./(0 + bi). t= 0 .. where c = 1 + 1/47r 2..3) that we can take 0 1 0 0 a= (1 + 47r2 0 0).e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. of (6.47r2 3 .(6. S.1 0 0 )3 = (111).6.. shows that the distribution B with density b(x) = c(1 cos(21r x))ex. s = c/ 2 . .3) was suggested by Colm O'Cinneide.h. One of his elementary criteria.3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. b(x) > 0 for x > 0. see Asmussen & Bladt [29] (the representation (6. (6..1. T.1 0 . Namely. personal communication).an_3 an _ 4 .s. (6.1 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. 1 . matrixexponentiality implies a rational transform. cannot be phasetype. 0 1 an an1 an _2 . a2 a1 Proof If b(x) = aeTxt.2 A remarkable feature of Proposition 6. (6..4) 0 0 1 c This representation is complex. namely to asssume the roots 6l. Example 6 .. For a proof. u Remark 6. . then b*[0] = a(0I T)1t which is rational since each element of (01 .3) 0 0 0 0 0 .1) as E 1 c. 0 0 0 0 1 0 0 .2).
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)) .242 CHAPTER VIII. we use a representation (a.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . Consider the distribution with density = 15 ((2e2x . T the phase generator and t = Te. . that despite that the proof of (6.3.1)2 + 6). (6. For the first.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. leading to matrix calculus in high dimensions when b is small. Then (cf. Corollary 111.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. (6. T. We recall (see Section 3. MATRIXANALYTIC METHODS Example 6 .1 to get i (u) = f3esus. we take as starting point a representation of b* [0] as p( O)/q(9) where p. and that the minimal number of phases in a phasetype representation increases to 0o as 5 . q are polynomials without common roots. recall that t = Te) that if B is phasetype and (a.6) holds true also in the matrixexponential case.5) Thus. then: Proposition 6. For the second algorithm. T. and present two algorithms for calculating '(u) in that setting. As for the role of matrixexponential distributions in ruin probability calculations. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1. and can use this to invert by the method of Proposition 6. 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. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. t) a phasetype representation with a the initial vector. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. 0.6) The remarkable fact is.5 (6. t) of b(x). we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels).
Then in Laplace transform formulation .to+)1 = (BI . (6.T)1 + 1 ib* (91.6b* .T)1T 2 = and 1 = AB IT2 + 82T .1t = b* . we get b+ = 0aT1(9I T).t.1 + 82 (9I .to+)1T . xb(x) dx = aT2t.1 = ^(T1 + ( 91T)1).T)1 J0 00 b(x) dx = f aT1t. the assertion is equivalent to a+(BI .T)1 (91. we get (91. b+ = a+(9I .T)1t ( l .T)1t.'t.T . (91.1t du = .T)1T1t. . b+ = a +(BI . U =.1t = f3a (0I T)1T1t . but we shall give an algebraic proof. MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I . 519) (A + UBV ).6).T)1ta+(OI . Presumably.T .b* (6.T .to+)1T .T)1 so that b* b** b** a+(9I . with A = 91T.a+(9I . this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .A .5 ).T)1 + (6I .7) 9( cf.T). (6.6.1BVA1. From the general matrix identity ([331] p.1 + b+ = b++ 1 .1UB(B + BVA1UB).1 = A1 . since (91T)1T . Now.T)1t)1a +(9I .B=land V=a+.
7). For expositions on the general theory of matrixexponential distributions. of (6. 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. In Corollary VII. . 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. 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. The proof of Proposition 6. 3. VII. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. which is selfexplanatory given Fig.1. Much of the flavor of this classical approach and many examples are in Cohen [88].8.T)1)t = 8 (1 .1) is the same as the r.8 a(T1 + (01.1.h.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. see Asmussen & Bladt [29].3 is taken).b*).T)1T2t . 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0.T)1T.s. cf.1. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. see the Notes to VII. Lipsky [247] and Asmussen & O'Cinneide [41].244 CHAPTER VIII. T). MATRIXANALYTIC METHODS . some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). a key early paper is Cox [90] (from where the distribution in Example 6. /3aT1(0I .82b*. See Fig. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. but the argument of [286] does not apply in any reasonable generality). 0 Notes and references As noted in the references to section 4. the ruin probability(u) was found in explicit form for the case of B being exponential.la.1t = /3a (9I . 7.1.5 is similar to arguments used in [29] for formulas in renewal theory. a. From this it is straightforward to check that b**/(b+ .
0 < t < u. though still Markov.1z I. Let P(tl. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). t2) = exp where Q(t) = ds [P(t.1) where A(t) = v(u)P(0.7.1 The difference from the case p(r) = p is that {m2}. we obtain V)(u) = P(m„ E E) = v(u)P(0. Figure 7. Note that in general >iEE Vi (U) < 1. in contrast to Section 3. In fact. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. Proof The first statement is clear by definition.t)). P(tl. By general results on timeinhomogeneous Markov processes. t + s) .t2) be the matrix with ijth element P (mt2 =j I mtl = i). Given the v(t) have been computed. O<. Also. i. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. 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. is no longer timehomogeneous.u)e = A(u)e (7.tl < t2 < u. the definition of {m8} depends on the initial reserve u = Ro. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u .I] I 80 { tq f Q(v) dvl t1 1 . Ai(t) = P(mt = i).e.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . t) is the vector of state probabilities for mt.
t). those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . Given A'.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). or it stops between level u + p(u)dt and u.t) for the second. the probability of which is 1 . 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. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 .3dt. 0 Thus. Hence Q(t) _ T + tv(u . jEE . Proposition 7. A'(t) = A(t)Q(t) = A(t)(T + tv(u . the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt).(u) p ( u) = . MATRIXANALYTIC METHODS However. Given A. (7. given A. dt]. whereas in the second case the probability is p(u)dt • tjvi(u).t)).246 CHAPTER VIII. 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 this occurs. the probability that level u is downcrossed for the first time in phase i is ai. {mx} has jumps of two types. vi. 0 < t < u. two things can happen: either the current jump continues from u + p(u)dt to u. Thus.2 For i E E.t and being followed by a downcrossing. In the first case. from a computational point of view the remaining problem is to evaluate the v(t).Sj i)p(u)dt • tji = Sji + p(u)tji dt.Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. we get vi(u) = aidt + (1 .4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0.
V .. <oo..h.4) backwards for {va (t)}v>t>o.s.5).^ 0. say. (u) for any values of u and v such that u < v. after a certain level v.)P"(AnB. supRt>v l t<7 I where o.3 For any fixed u > 0. This yields v. When solving the differential equation in Proposition 7. Since the processes Rt and Rt coincide under level B. Let p" (t). (v) is given by the r. Then P(B.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. Rt .0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. (u) on both side and dividing by dt yields the asserted differential u equation. From Section 3. we face the difficulty that no boundary conditions is immediately available.. vi (U) = lim v= (u). we can first for a given v solve (7. To deal with this. . we have p(r) = p = vi (u) 0aTe. denotes the time of downcrossing level u . refer to the modified process. Thus. of (7.i7rT1/p. RESERVEDEPENDENT PREMIUMS 247 Subtracting v. starting from v"(v) = . Now since both P(A n Bv) 3 0 and P"(A n Bv) . P u which implies that v.7.) P"(AnBv) = P(AnB.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.. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. and similarly P"(Bv) . say. F" etc. then P(A n Bv) _ P"(A n BV').) + 0 as v + oo. consider a modification of the original process {Rt} by linearizing the process with some rate p.2.
e. Therefore u pl(vvueTa t 1.7) f o (the integral is the contribution from {R. where v = inf It > 0 : Rt < v}.zp1(u)/(1 . 1/n. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. cf. The trapezoidal rule used in [288] gives a precision of 0(n 3).z51(v)).1. where. The f iin in (7. 3u etc. the evaluation of Vi(u) requires q(u) = 1 .248 CHAPTER VIII. as well p1(u).q(v dx +( ) ) = ( ) ( q( )) vueTva (7. while the fourthorder RungeKutta method implemented in [30] gives 0(n5).10 that in addition to the O'(•). 2u. However. We recall from Propositon VII. The precision depends on the particular quadrature rule being employed. p(r) P. < 0}). such that Rt coincide with RI under level v and with Rt above level v. T). p2.. 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. (7. which is available since the z/i'(. say. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. let v(u) = a+2ieiT +ta+>)(uv).7) equals 01 (v . assuming u > v for the moment. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. > 0} and the last term the contribu tion from {R.v v(u)eTat 1 1 . r<v r > v. 2/n.V" M 0 . numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. 0 < u < v.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. v = u..x) dx f v(u)eT xt dx .1.1.RQ (defined for or < oo only) is defective phasetype with representation (v(u). To evaluate p1(u). the probability of ruin between a and the next upcrossing of v. i..9.. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII.1a. for u > v the distribution of v . (u)}.) are so. Corollary 3.
v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.e6u 35 . so we consider the nontrivial case example p2 = 4 and p1 = 1.1 from which we see that pl (u) = 1 + 1 249  1 . Example 7.01 (v .e6v Let Al = 3 + 2V'2.24ev . p2 < 3.2.x) dx 1 ^(v) ( 1 . (7.v(u)eTve). Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e . I.8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A..24e. 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .v(u)eTVe .(7 The arrival rate is (i = 3.to+))11 {e{T®(Ttoy+ ))}„ .jl (t ®e) Thus.^1(v) 1 . all quantities involved in the computation of b(u) have been found in matrix form.7.2.be the eigenvalues of T + to( 2 ).v(u ) eTV e J v(u)eTxtz/)l (v .v) + (2^ + 3v2 ea'(u " .e.21 = ? yields 0(u) = 1.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T . RESERVEDEPENDENT PREMIUMS 1 .u .8) The integral in (7.and A2 = 3 .2V"2. From Example 3.x) dx} V 1 1(v) f V v(u) eTxt. Since µB = 5/21. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.4 Let {Rt } be as in Example 3.
/2) ea 1(u .1 V2 = 4e5"+6 35e6v .1 Thus.v)esv + 7 4_ 2. MATRIXANALYTIC METHODS From (7. and one gets 12e5" .24es" . pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v . The analysis and the example are from Asmussen & Bladt .24es" .24e5v ..24e5v .2 35e6v .+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4. 192esv + 8 P1 . ) e sv + ( 2v/2.21(35e6v .1. Notes and references [30].1)' ?.250 CHAPTER VIII./2 ea1(u") .b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived. 21 3 In particular.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.
For the definition . a2)) with density 1 e(logyFh) 2/2az .g. we require that B is concentrated on (0.N(µ. x 2iror2 (c) the Weibull distribution with decreasing failure rate . see I.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[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.46 and at numerous later occasions require a light tail.4. For further examples. for all t > 0. III. (b) the lognormal distribution (the distribution of eu where U . 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. A rough distinction between light and heavy tails is that the m. the exponential change of measure techniques discussed in II. L(tx)/L(x) 4 1.2b. B(x) = ex0 with 0<0<1.B(x). Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. oo ) and say then that B is subexponential (B E S) if 251 . and instead we shall work within the class S of subexponential distributions . For example.f. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying.
P(max(Xi.v. Proof By the inclusionexclusion formula. Then: (a) P(max(Xi.B(x)2 .3 Consider the standard exponential distribution.1 Let B be any distribution on (0. oo). X2 but none of them exceeds x. that is. As contrast to Proposition 1. B(x) Here B*2 is the convolution square. if X1 + X2 is large . . the behaviour in the lighttailed case is illustrated in the following example: Example 1. 1/2 'typical' (with distribution B) and w. X2) > x} C {X1 + X2 > x}. That is.p.252 CHAPTER IX. HEAVY TAILS B*2\ 2.v. 1).'s X1. Thus . x 3 00. P(Xi <yI Xi+X2>x) 1B(y). then P(X1>xI X1+X2>x)* 2. That is. then (with high probau bility) so are both of X1.v. X2) > x}. Since B is concentrated on (0. one can check that x x where U is uniform on (0. the r. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. X2) > x) is P(X1 > x) + P(X2 > x) .2.1(b) is oo. X2) > u x)/B(x) = 2. To capture the intuition behind this definition. (b) liminf BB(() ) > 2. note first the following fact: Proposition 1. We later show: Proposition 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.1) then means P(X1 +X2 > x) 2P(Xi > x).2 If B E S. oo).2B(x). proving (a).p.'s. X2 with distribution B. In terms of r. B(x) ax. X2) > x) ^' 2B(x). X1 is w. (1. X2 > x) = 2B(x) . we have {max(Xi. the distribution of independent r. 1/2 it has the distribution of X1I X1 > x. given X1 + X2 > x. and thus the lim inf in (b) is at least lim inf P(max(Xi. Thus the liminf in Proposition 1. In contrast.F(X1 > x. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large.
[In terms of r. then B(B(x)y) * 1 uniformly in y E [0. and combining with Proposition u 1.B E S. we get limsupB*2(x)/B(x) < 2. We now turn to the mathematical theory of subexponential distributions. Hence lim sup a+oo B*2(x) 2B((1 .B*(n ) B(dz) (1. 1 < B(x ) B( x) Y) < B( 0).B*n(x . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . a contradiction. If lim sup B(x .1.6)x)/((1 . then the overshoot X .S)x. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.y)/B(x) > 1 since y > 0. B( 0 .B(y) = 2. then either one of the Xi exceeds (1 . Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 . or they both exceed Sx.yo].z B(x) .4 Any B with a regularly varying tail is subexponential.xIX > x converges in distribution tooo.v.S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . If X1 + X2 > x. yo] as X + 00.'s: if X . Finally lim inf B(x .B(y)) .2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. y] and (y. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. Let 0 < 5 < 1/2.] Proof Consider first a fixed y. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.y)/B(x) > 1. x].5 If B E S. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1.1(b) we get B*2(x)/B(x) * 2.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . Proposition 1.
This implies B(x) > c2e5x for all x.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . B*(n+1) (x I xy + Jxx y) W. Proof For 0 < 5 < e.B*2 (x) B(x) (x .nI < e for x > y. choose y such that IB*n(x)/B(x) . Proposition 1. then for any n B*n(x)/B(x) * n.5 that B(n) > e6B(n .B(x .2.z) B(dz) _y B(x) 111 Lx B .2). B(x) \Jo _ B(x . HEAVY TAILS Corollary 1. Given e > 0. 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 .. its intuitive content is the same as discussed in the case n = 2 above. 0 Proof of Proposition 1.6 If B E 8.5 and the induction hypothesis. b[c] = oo for all e > 0. Then by (1. Proof We use induction. then e"R(x) * oo. x oo.254 CHAPTER IX. The first integral is y B(x . and this immediately yields the desired conclusions.5 and dominated convergence.z) B(x) Here the second integral can be bounded by B*n(y) B(x) . O The following result is extremely important and is often taken as definition of the class S.z) B(dz). The case n = 2 is just the definition. so assume the proposition has been shown for n. we have by Proposition 1.z ) (x ) = 1 + (^ B(x .z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.7 If B E S.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1.1) for all large n so that B(n) > cle6n for all n.
0 completes the proof.y)Ai(dy) = (x)o(1) (1. Then by (1. Since P(X1+X2 > x. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .v. Then Al * A2 (x) . Then Al * A2(x) = P(X1 + X2 > x).'s such that Xi has distribution Ai. 0 Lemma 1.9 Let A1. X2 be independent r.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .3) Using the necessity part in the case Al = A2 = B yields f xv B(x . Proposition 1. A2 be distributions on (0. Xi <= v Ai (x .B(x .1. choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). For any fixed v. e > 0. Combining these estimates and letting a 4.(al + a2)B(x).5 easily yields P(X1 + X2 > x. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .y)Ai(dy) v) f o . oo) such that Ai (x) _ aiB(x) for some B E S and some constants al.X2 > xv) < A1(xv)A2(x v) .i).y))/B(x).ala2B(x)2 which can be neglected.y)B(dy) = B(x)ov (1)• v (1.z) B(x . a2 with a1 + a2 > 0.z) B(dz) < 1 + A + an(1 + d) . Proof Let X1. 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) .4) . an+1 fX B*n( *n(x .z) B(x) < 1 + A + an sup f x B(x .X1 > xv. an = supx>o B*n(x)/B(x). 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. 0 Proposition 1.z) B(dz) x .2).8 If B E S. Proof Define 5 > 0 by (1+5)2 = 1+e.
5) Here approximately the last term is B(x)o„(1) by ( 1. Then B E S provided fo "O exA(x) b(x) dx < oo. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1.13 Let B have density b and failure rate A(x) such that . of (1.11 Let B E S and let A be any distribution with a ligther tail.y)B(dy).3) follows if CHAPTER LX. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.aiB(v)) = B(x)o„(1). In the regularly varying case. HEAVY TAILS 'VV B(x . A(x) = o(B(x)).4).y)Ai(dy) = B(x)o„(1). We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290].9).2.s. then A E S. B1 * B2 (x) . the l. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown. a1 = a2 = a yields A*2(x) . However.(x) is decreasing for x > x0 with limit 0 at oo.Bl (x) + B2 (x) when B1. it should hold that B1 * B2 E S and B1 * B2 (x) . Then A * B E S and A * B(x) . u Corollary 1. u It is tempting to conjecture that S is closed under convolution. with a > 0 and L1. Proof Taking Al = A2 = A.2A(x). L2 slowly varying. a2 = 1.5) becomes x B(x . B2 E S. That is. u Corollary 1. f " By a change of variables.256 Now (1. A2 = B so that a1 = 0. Hence Corollary 1. V (1.Ai(x .v)Ai(v) . .h.B(x) Proof Take Al = A.v)B(v) + _'U Aq(x . L2 are slowly varying. if q(x) aB(x) for some B E S and some constant a > 0. it is easy to see that if L1. That is. whereas the two first yield B(x)(Ai(v) .2aB(x) . then so is L = L1 + L2.10 The class S is closed under tailequivalence. i = 1.12 Assume that Bi(x) = Li(x)lxa.
2). Since ) (x . Jo For y < x/2. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. In the regularly varying case. we first quote Karamata's theorem (Bingham. Then b(x) = Ox0lexp.U) /or) v 2x This yields easily that ex.A(y)a(y ) = ev'(y) b(y).A(y)\(y) dy + fox/ 2 eA(x ).`(x)b(x) is integrable. a(x) = ax01.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy )..y) dy.3 < 1.1 has limit 1 + 0. x . Thus A(x) is everywhere decreasing. Thus by dominated convergence .A(xy)A ( y).y) < yA(x . elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. an integrable function by assumption.1.1)xcl1 ..e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . Then B(x) = eA(x). 0 A(x) .A(x .(y) dy.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).13 works in this setting. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). we can use the same domination for the second integral but now the integrand has limit 0 . By (1.(x . Thus.y) * 0.14 Consider the DFR Weibull case B(x) = ex0 with 0 <.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . Further.15 In the lognormal distribution. Example 1. To illustrate how Proposition 1.12. the first integral has limit 1 . the u lognormal distribution is subexponential. The middle bound shows that it converges to b(y) for any fixed y since \ (x . and exa(x)b(x) = (3x01e(10)x9 is integrable.y) < A (y) for y < x/2. L(x) y° (a . Thus B*2(x )/ B(x) . f ' L(y) dy . proving B E S. Thus. Goldie & Teugels [66]): Proposition 1. the DFR Weibull distriu bution is subexponential. Example 1. B*2(x) .16 For L(x) slowly varying and a > 1. Define A(x) = fo . subexponentiality has alrady been proved in Corollary 1.
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 .258 From this we get CHAPTER IX. HEAVY TAILS Proposition 1.18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1. f O B(y) dy .17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.1)]) xa L(x) (x[1 + y/(a .13 may present a problem in some cases so that the direct proof in Proposition 1. .1) and P(X (. Then 7(x) . then B(x) . then 7(x) x/(a .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. (1 + y/(a . let X W = X . (c) Under the assumptions of either ( a) or (b). However.6) EX(x) . Thus exa(x)b(x) .a/x.1)X(x)/x > y) = P(X > x[1 + y/(a . yo] . Then: Proposition 1.xjX > x.1))a . we get (1.ea b(x) is integrable. the monotonicity condition in Proposition 1. We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.L(x)/x" and )t(x) . 'y(x) = EXix>. Proof ( a): Using Karamata's theorem.y(x)B(x).4 is necessary in full generality.)/Y(x) > y) (1 + y/(a .1)])a 1 1 .E(X .1)] I X > x) L(x[1 + y/(a . 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/A(x) and P(X ixil'Y (x) > y) * e'. More precisely.
d. The remaining statement (1. cf. .. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). with common distribution G E S and let K be an independent integervalued r. then Vi(u) P Bo(u).15..7) is referred to as 1/A(x) being selfneglecting. u a oo. 0 G(u) L G(u) ...14.+YK> u) = ^•P(K = n)G* n(u ) .A(x) I X > x) = exp {A(x) . THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) .t be the claim surplus at time t and M = sups>0 St. Theorem 2 . Y2.and lognormal distributions . . Examples 1. Lemma 2.2. Let St = Ei ` Ui .f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. be i.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .1/. i. 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. 1. We get p(yl+. nG(u). Notes and references A good general reference for subexponential distribution is Embrechts. Bo(x) = f0 B(y) dy / µB.v.n0 1•P(K= n)•n = EK.nn.(x).2 Let Y1.EK G(u). 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. r(u) = inf it > 0. Then P(Y1 + • • • + YK > u) .8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/.1 If Bo E S. with EzK < oo for some z > 1. Recall that B0 denotes the stationary excess distribution. It is trivially verified to hold for the Weibull. St > u}. Kliippelberg & Mikosch [134]. P The proof is based upon the following lemma (stated slightly more generally than needed at present).
For some numerical studies.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 ) . The problem is a very slow rate of convergence as u ^ oo. u The condition Bo E S is for all practical purposes equivalent to B E S. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . Proof of Theorem 2. Bo ¢ S. Note that in these examples . P(K = k) = (1. Bo E S. the result follows immediately from Lemma 2. (2.x^ ) B(x) _ f or ( lox . Bo E S is immediate in the regularly varying case.µB(01 .. .x400 PBB(x) PB Leta+oo.1 is notoriously not very accurate. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant.18. as well as examples where B ¢ S. in our three main examples (regular variation . Proof Since B(x + y)/B(x) * 1 uniformly in y E [0.2. Borovkov [73] and Pakes [280]. see Abate.p) and EzK < oo whenever pz < 1. mathematically one must note that there exist (quite intricate) examples where B E S.260 CHAPTER IX.?(xµ 8 (x). The approximation in Theorem 2. lognormal . r(1/Q) xlQexp B(x) = ex' From this .µ J ) . The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1.p)p'.. x 4 00. In general: Proposition 2.3 If B E S.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. a]. then Bo(x)/B(x) + 00. The PollaczeckKhinchine formula states that (in the setup of Lemma 2. Since EK = p/(1. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. See also Embrechts & Veraverbeeke [136]. u x+a Notes and references Theorem 2.13). However. Weibull) one has Bo(x ( B(x) .1) In particular .1.1 is essentially due to von Bahr [56]. Bo is more heavytailed than B .
1. (3. In [1]... We assume positive safety loading. THE RENEWAL MODEL 261 Choudhury & Whitt [1]. {n= 0. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. there is a representation of M similar to the PollaczeckKhinchine formula.g. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x .y)/B (x) > 1 uniformly on compact y internals. E. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA).Ti.. i. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. + Xn. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. Somewhat related work is in Omey & Willekens [278]. T1 the ith interarrival time and Xi = U.. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. To Bo(u) u + 00. [279].1 when u is small or moderately large.. Define further 0 = IIG+II = P(r9+ < oo). Based upon ideas of Hogan [200].3. i=1 . Then K M=EY. . also a second order term is introduced but unfortunately it does not present a great improvement.e. The main result is: Theorem 3 . G+ (A) = P(Sq+ E A. in [219] p. Kalashnikov [219] and Asmussen & Binswanger [27].] The proof is based upon the observation that also in the renewal setting.} Then ik(u) = F ( M > u) = P(i9 (u) < oo). p = iB /µA < 1.. Let U= be the ith claim .1 gives 1010.1) this end . T+ < oo) where r+ = T1 + • • • + T. Snd) = Xl +.+ E A. This shows that even the approximation is asymptotically correct in the tail. t9(u) = inf {n : Snd> > u} . M = sup s$ . 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V.y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}.9+ < oo) = P(S..
0] normalized by IPG_ I so that we should have to G+(x) .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_. Write G+( x) = G+ ( x. HEAVY TAILS where K is geometric with parameter 9. Lemma 3 .FI(u). cf.(. Proof Let R+(A) = E E'+ ' I(S. are independent of K and i. d+ < oo).. Let F denote the distribution of the Xi and F1 the integrated tail. Lemma 3 . u a 00.262 CHAPTER IX. Proof By dominated convergence and (b). x > 0. Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. and hence FI(x) . B(x) _ J O° B(B(x)y) A(dy) f 1 .oo.9)9'' and Y1.1 IPG_ I / F(x .d.B(x). 0 The lemma implies that (3. A(dy) = 1. x + oo.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. oo) = F(S.i. 0] to the integral is O(F(x)) = o(FI(x)). FI (x) _ fz ° F(y) dy.Ti).2). G_(A) = P(S.FI(x) /IPG_I.3) and we will prove it in this form (in the next Section.1) holds for a general random walk satisfying the analogues of (a).d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_. we will use the fact that the proof of (3. As for the compound Poisson model.3 G+ (x) . (3.1) is equivalent to P(M > u) " .Y2.PBBo(x). The heuristics is now that because of (b). P(K = k) = (1 .y) dy = 1 Pi (X) oo IPG_ I .g+ > x. x * oo.. U_ (dy) is close to Lebesgue measure on (. with distribution G+/9 (the distribution of S.2 F(x) . the contribution from the interval (. whereas for large y . A. (b) and does not rely on the structure Xi = Ui .y+ given r+ < oo).N. 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.. Then 0 0 F( x .
THE RENEWAL MODEL 263 We now make this precise. then by Blackwell 's renewal theorem U_ (n .UG_ I x. Given e.1.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup . u Proof of Theorem 3. By Lemma 3.1. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. In the lattice case. n] is just the probability of a renewal at n.9)IpG_ I Differentiating the WienerHopf factorization identity (A. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .2). we can assume that the span is 1 and then the same conclusion holds since then U(n .1. > (1 . (3. We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x .1. and in the last that FI is asymptotically proportional to Bo E S. choose N such that F(n .O[s])(1 .=1 BIp G_ I (1.0)0k k I(u) A.9) 1 .3.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10.y) U. n] + 1/I µG_ 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 0.3. n] < (1 + e)/1µc_ I for n > N.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.G+[s]) . the proof is complete.(dy) fN FI ( x) + lim sup ZY00 N F(x .2) yields 00 F F I (u) P(M > u) _ E(1 . If G_ is nonlattice. F(Y= > x) FI(x)/(OIp _ 1).1. and that U_(n . Hence using dominated convergence precisely as for the compound Poisson model.1 I .F[s] = (1 .y) U_ (dy) 00 FI (x) < lim sup F(x) U(N. Similarly.
Proof Let w(u) = inf {n : Sid) E (u .(1 .. Mn < u}. we have P(M E (u .4 on the joint distribution of (S. P(M > u. Therefore by Lemma 3. S+9(u) .IIG+II)µc_ = (1 . see Asmussen & Kliippelberg [36]. 10(0) But since P(M > u .So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . Note that substantially sharper statements than Lemma 3.yiui_1.a) N P(M > u). Sty(u) .0)ua_ . u)) > P(w(u) < oo)(i lp (0))• On the other hand.a.264 and letting s = 0 yields CHAPTER IX. u)). Then P(M E (u .1 is due to Embrechts & Veraverbeke [136].2. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .Se(u)_1 < a) = o(Fj(u)). 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.SS(u)}n=o.(u)+n . with roots in von Bahr [56] and Pakes [280].a.So(u)) are available. on the set {M > u..AB iP We conclude by a lemma needed in the next section: Lemma 3 .1)6+[0] . HEAVY TAILS µF = (1 .a. 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times..u)) = o(P (M > u)) = o(FI(u)). S+q(u) . .l.Sty(u)_I < a} we have w(u) < oo. allowing also for possible dependence between the arrival process and the claim sizes.a. must attain a maximum > 0 so that P(M > u.4 For any a < oo. Notes and references Theorem 3. and {Su.
M.1... examples and counterexamples.1) .1 = max k=0.1. Theorem 4. 2.. = Sx..X1 is the generic cycle.. {Sn}n=o. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i. {SX1+t .4.1 based upon a regenerative assumption. We let F* denote the Podistribution of Si. Define S. 4..F*(X) = P0(Si > x) . 4.1 Note that no specific sample path structure of {St} (like in Fig....4 below. 4... see [47]. G(x) (4. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen. Figure 4....1 except for the first one) is a random walk. E0.. For further approaches.X2 < . (corresponding to the filled circles on Fig. +1.i. t>0 S. 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 <.n n=0. 0o(u) etc. and the distribution of {Sxk+t . and apply it to the Markovmodulated model of Chapter VI. We give here one of them. < 0 and EoX < oo where X = X2 . Thus the assumption . The idea is now to observe that in the zerodelayed case...Sxk}o<t<xk+1xk is the same for all k = 1...1) is assumed.Sxi}0<t<x2Xl . The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.d.. M = sup St. . See Fig. M* = max S..1 where the filled circles symbolize a regeneration in the path. We return to this point in Example 4. such that {SXo+t  SXo}0<t< X 1Xo . Schmidli & Schmidt [47]. assume pp.
3) hold.2) to show F(M* > u) > 1. Fo(Si > X). HEAVY TAILS for some G such that both G E S and Go E S makes (3..* i o<t<xn+1x. Then '00 (u) = Fo(M > u) . Since clearly M(x) > Sl . See Fig.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. 4. u p 00...3) where Mnx) = sup o<t<xn +1 X. it suffices by (4. Sxn +t . The one we focus on is Fo (Mix) > x) .2 Theorem 4..Sxn = sup Sxn+t . the assumption means that Mix) and Sl are not too far away.S.2.1 Assume that (4.3) applicable so that F(M* > u) 141 F*(u).1) and (4. (4. (4. N N Xi=0 N Figure 4.266 CHAPTER IX. (4. Proof Since M > M*. jF11 F* (U).4) liminf u>oo F(M > u) .
E (u ..: S.5) which follows since Po (M > u.4. 2. Given e > 0. Po(M* > u) .1 can be rewritten as 00 (U) (4. x > a. assume the path structure Nt St = EUit+Zt i=1 . Let a > 0 be fixed. u))/P(M* = 0) = o(Po(M* > u)).(1 .S. S.E) Po ( n max St u.a. Mn+l > a V (u .. /3(u) = inf{n=1.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). .e)Po (MMX> > x).( u)1 > a) 00 1: Po(Mn<u..4). M^xu)+l > a) . To this end.e)Po (M > U). Letting first u + oo and next e . Theorem 4.: Sn > u} .a.Po (M* > u..1 = limti00 St/t. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 .2.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u. u)} < P(M* E (u .4. )) > (1 .e)Po (M > u.Mn +1 >aV(u n=1 00 S. MW O(u)+1 < a) IN ( U n=1 A1... 0 yields (4.Sn 0<t<x„+j ( 1 . We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. choose a such that Po(Si > x ) > (1 .(u) . Under suitable conditions . Then by Lemma 3..6) 1 p pBo(u) u where B is the Palm distribution of claims and p .
cf.} and satisfying Zt/t N.6) holds with p = .v. Proof It is easily seen that the r. a4' 0.2 Assume that {St} is regenerative and satisfies (4. and also for Mix) since Nx FNX U.268 with {Zt} continuous. since the tail of Zx is lighter than B(x) by (iv).3PB.I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .Q = EoNx/EoX.7). Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) . HEAVY TAILS N` U. the proof of Lemma 4. (ii) EozNX < oo for some z > 1.4). The same is true for Sl. (iii) For some o field Y.p) Ju P Bo(u) 1p 0 .1 is in force. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . X and N.'s order EoNx • B(x). independent of {> CHAPTER IX. oX (see Proposition A1. we get 00 (u) 1 IPF.X both have tails of sup Zt. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.8) x Write .6 below. Corollary 4. Mix) < > UE + i=1 o<t<x Thus Theorem 4. Assume further that (i) both B and Bo are subexponential. and ENX Ui . i=1 (4.
9). we conclude just as in Example 4.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4.4 Assume that St = Zt .4. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. and for some constants ci < oo such that cl + • • • + c. In particular. 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).(NX). consider the periodic model of VI. More precisely.. Bo E S. X2 = 1.e. in particular lighttailed. .. and taking F = o. Taking again Xo = Xi = 0. i.6) u holds. < 1.t + EN'I Ui where {>N`1 Ui . Assume that B E S. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. (iii) is obvious.. X2 = 1. X3 = 2.. The number N.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. The regenerative assumption is satisfied if we take Xo = Xi = 0. (i) holds. Zt .3 As a first quick application.6) holds. .0 (thus (iv) is trivial). . then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. Theorem 4. Again . we assume that B E S. Example 4 . 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. 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 . MODELS WITH DEPENDENT INPUT 269 Example 4 . 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.3 that (4.. of claims arriving in [0. We consider the case where one or more of the claim size distributions Bi are heavytailed. The arrival rate is /3i and the claim size distribution Bi when Jt = i. Bo E S. Then (4. Thus we conclude that (4.6) holds. > 0. The key step of the proof is the following lemma. .
. 2 . i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo..ciG(x). . i=1 Proof Consider first the case X = 0. The same dominated convergence argument completes the proof.. Let {Fi}t=1 P be a family of distributions on [0.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 .. . . 1.X i=1 j=1 where conditionally upon F the Xi. are independent with distribution Fi for Xij. . i1 = E\ G(x) In the general case. X > 0 a r. HEAVY TAILS Lemma 4 . It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni.. . . cp with cl + > 0 it holds that Fi(x) .270 CHAPTER IX. oo) and define p Ni Yx = EEX'i . i=1 P(Yx > x ^) < P(Y0 > x I. NP ) and X are ."+Np ... and that for some + cp distribution G on [0. oo) such that G E S and some c1.^•) G(x) P ^ E ciNi = C.2. .}P. P P P(YX and > x I. NP ) be a random vector in {0. 6 Let (N1.5. u Proof of Theorem 4. An easy conditioning argument then yields the result when Jo is u random. For lighttailed distributions. Thus dominated convergence yields ( P(Yo>x P(Yo>x .G( x ) > ciNi . we can define the regenerations points as the times of returns to i. If Jo = i. as x a oo. and the rest of the argument is then just as the proof of Corollary 4. and F a aalgebra such that (N1. Assume EzN1+"'+Np < oo for some z > 1 and all i.F) < CG(x)zn'1+. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin .Fmeasurable.. Then P P(Yx > x) .v.c'(x) where c = ciENi .
cf.5 shows that basically only the tail dominant claim size distributions (those with c.. 5 was first proved by Asmussen.5. The present approach via Theorem 4. Theorem 5. We start by reviewing some general facts which are fundamental for the analysis.4.4. Then O(u) .7.. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. the discussion provides an alternative point of view to some results in Chapter IV. Essentially..p)Bo(u).1. 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. cf. see Schlegel [316]. As usual. i.T2. i. states that under mild additional conditions.6) to hold in a situation where the interclaim times (T1.4. For further studies of perturbations like in Corollary 4.2 and Example 4. we let PN"N = P(.. Theorem 2. I T(u) < oo). 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). IV. this should be compared with the normal limit for the lighttailed case. Schmidli & Schmidt [47]. An improvement was given in Asmussen & Hojgaard [33].. The main result of this section.e. cf.d. Within the class of risk processes in a Markovian environment. ). It follows from Theorem 4. In contrast.3. m is a (orfinite) .i. in particular Proposition 2. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. r(u) is the time of ruin and as in IV.1 is from Asmussen.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. as well as a condition for (4. and independent of (T1. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u.. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case.7). That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). VI.pl(1 . Combined with the approximation for O(u). and the final reduction by Jelenkovic & Lazar [213]. Theorem 4.T2.4. this is applied for example to risk processes with Poisson cluster arrivals. Notes and references Theorem 4. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. ) form a general stationary sequence and the U.
2) means ffh(a.s=j are the transition probabilities for {St}. For the present purposes it suffices . . {St} and {Rt} are in classical duality w. The simplest example is a discrete time discrete state space chain.). but the example of relevance for us is the following: Proposition 5. 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} . Let G denote the distribution of ENt U. Say {St} is reflected Brownian motion on [0. y = 0).z. Thus. a familiar case is time reversion (here m is the stationary distribution). and (5.s. to consider only the case Px(w(F`) = 0) 0. resp.z) dx G(dz) = ffh(y + z) k(y)dy G(dz).2) for all bounded measurable functions h.s.272 CHAPTER IX. the whole of R and not as usual impose the restrictions x > 0. St is distributed as y . k on E.y = Qx (.2) with t = 1 means m. in the terminology of general Markov process theory. u For F C E.t. j. t.t. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. say. w(Fc) < oo ) 'In general Markov process theory. follows by the substitution y = x .1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w .00). where we can take h.= y. Sw(F. . k as indicator functions. We let QS be the corresponding distribution and Qx. y to vary in.h.rij = mjsji where r13. and starting from So = y. Then (5. 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. to the r. x = 0+ and F = (0. m. {Rt}.)k(x . for states i. however . Proof Starting from Ro = x. Rt is distributed as x + t . HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. r.>N` Ui. Lebesgue measure. oo).h.r. (note that we allow x. The equality of the l.t + EI U.
..). Sw(F)1 = y) ...SS(F.. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. Sn+1 E Fc) nx.. 5. .. x = 0.5. 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). . oo) = r(0) x= St y (a) Figure 5.= y) Theorem 5 .3 The distribution of r(0) given r(0) < oo. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time.y = Qy Q. and we let Qy y refer to the time reversed excursion .1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. io = x. the one in (b) is the time reversed path. w(0. That is. 0]. in = y. The theorem is illustrated in Fig .itt) = P Px(w(Fc) < 00. /^s x (S1 = Z1.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. when p = . But in the risk theory example (corresponding to which the sample paths are drawn).y() = P ({SW(F`)t} 0<t<w(F °) E So = x. Sn = in = y. .2.y(2p21 . Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. [note that w(z) < oo a. S. Qx. We can then view Qy. Thus.13AB < 1] Proof of Theorem 5. z > 0.. In particular: Corollary 5. in with i0. Sw(Fo)_ should be interpreted as Sw(F^)_1). . QR and QRy are defined similarly. 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. i1.y as a measure on all strings of the form i0i1 .2 Qy. in E F.1 for the case F = (oo.s.
i„_iEF Similarly.. = in = y. ..rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 ... in = x.. Si11 S 1 ..... in with 20. Si l io E mjSjx.ii .... Silt' E SO k=1 i1. in E F.gilt' k=1 ii .TI( 2n2n _1 . R .. in)  Pt' (R1 = ii. 2p) when 20.y(inin _ E SYj jEF` 00 Sxik _1 ... MY Thus Qx(ioii .ik1EF Similarly but easier Sxin_1 . .J (i. HEAVY TAILS E E Px (Si = 21i . S. 2p). 2n) = Qx. in) = oo jEF^ Sxin1 . note first that Pt' (R l = il. .. 21 ...... Rn = in = x.in E F. Sn+1 E Fc) n=1 i1. i0 = y... in = x..... Si1y 00 jEF° E E 5xik_ 1 .. t' y and Qy x are measures on all strings of the form ipi l . Rn = in = x... .(F<)1 = Y) S S and Qx y( ipil ... .274 note that Fx(w(Fc) < 00.... . . Rn+1 E F`) F (w(Fc) < 00.... Si1y k=1 i1 .. Rn+1 E FC) TioilTili2. (Fc)1 = y) 00 CHAPTER IX.in1 .. R Qy x(2p21 ...... . ... S..ik_1EF .... i0) Q x..... in) = Qx.... 20 = y.ik_1EF Sxin_1 . To show Q y x (i0 i 1 . Silt' E Sxik_1 ..
5.(o) > y} = {T(0) < oo. 7(0) < oo. S.')density of Y is B(y)/[.p. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch .v. 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 .2 The distribution of (Y.5. that is. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . P(") = P(. 1 w . Now the P(u. y > u.B(a) +a PBBo(u) .')distribution of Yu is Bo"). that is. the case r (O) < oo. the distribution w.'s are defined w.UBBo(u)]. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . ST(o) > y. P(o) ). the P(u. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). Bo") is also the P(u.2. Y > y} . Z) is described in Theorem 111.r. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. That is.r. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.t.2. Y > u). see Fig.2. U T(O) = T (u) Y Figure 5.t.
o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. . Yn_1 'typical'. We let K(u) = inf In = 1.276 CHAPTER IX. i.3. HEAVY TAILS Let {w(z)}Z^.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b).. the duration T+ (n) . The idea is now to observe that if K(u) = n. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long..d. Z). and since its dominates the first n . > u with high probability. Since w(z)/z a$. Z = ZI but relative to the kth ladder segment. the random vectors (YI. In the proof. Fig. and distributed as (Y.3) holds..p). we get the same asymptotics as when n = 1. Then Corollary 5. Since the conditional distribution of Z is known (viz.i. cf.p) in Pi"'')distribution. . z ^ oo. then by the subexponential property Yn must be large. . Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo.T+(2). Zn). 2. 4 Assume that Bo E S and that (5. . Bo") ). r(u)/Z 4 1/(1 . Hence Z.e.p) in F(u) distribution. i... denote the ladder epochs and let Yk. this in principle determines the asymptotic behaviour of r(u). and YI..1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. more precisely. Then. P(Z < a I Y > u) 3 0. 1/(1 . Zk be defined similarly as Y = Y1... K(u) = n).. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 .. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. (Y. a slight rewriting may be more appealing. Z/'y(u) * W in Pi "' ')distribution .. it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . in particular of Z.. Then 7(u)/y(u) ^ W/(1 . However.p) then yields the final result T(u)/y(u) + W/(1 . let r+(1) = T(0)..e.18(c) Bo")(yY (u)) + P(W > y) ( 5.: r+ (n) < oo. It is straightforward that under the conditions of Proposition 1. .1. That is . Recall the definition of the auxiliary function y(x) in Section 1..3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). Z1).r+ (n . . . conditionally upon r+ (n) < oo.. 5. must be large and Z1.p). are i.
II ' II denotes the total variation norm between probability measures and ® product measure. A"(u) _ {K(u)=n} = {Y1+ P(.n).n) (y1. P (Yj.u) E •) . Proof We shall use the easily proved fact that if A'(u). +Yn1<u. ..5 Ilp(u.2. Further. .u) II 0. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1. then IIP( I A'(u)) Taking A'(u) = {Y... I A'(u)) = P(u.. P(. .. Lemma 5.Yl+ +Yf1>u}.Bo (ri1) ®B( .Yn1iYn ..5. > u}. .u) E • I A'(u)) = Bo (n1) ®Bou) . FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5. Y„1. suitably adapted). I A"(u ))II + 0.3 In the following. Yn . A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events)..
..p) < y).i.6 IIPIu'n ) CHAPTER IX.. Then according to Section 5a. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + . be independent random vectors such that the conditional distribution of Zk given Y. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit.. . +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . For Theorem 5. . .4).u has distribution Bout That is. HEAVY TAILS ((Z1'. Z' are arbitrary random vectors..278 Lemma 5 .. 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 •) .. and that Yk has marginal distribution B0 for k = 1.7 O (u. n_1 < u.. y > u.. copies of {w(z)}.t. The first step is to observe that K(u) has a proper limit distribution w. the marginal distribution of Zk is Bo for k < n. {wn(z)} be i..r. Y") u etc... n . Z11)..). Zn). .y(u)T) . the discussion just before the statement of Theorem 5.d. (Y.P(Z' E •)II > 0 (here Y. Y1 +...1. wk(Zk) has a proper limit distribution as u + oo for k < n.. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn).. 2. k = 1. .P(Y' E •)II * 0.2..1). . in our example Y = (Y1. then 11P(Z E •) . n. Proof Let (Y11. .. Zn) E •) .P) > y) Corollary 5. P(u) since by Theorem 2. Y'. the density of Yn is B(y)/[IBBO(u)]. Zn are independent.. Z.1 P PBo(u) • P(W/(1 . Let {wl(z)}.P) Bo(u) for n = 1.. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic. see Fitzsimmons [144])....Bo (n1) ®Bo' 0.. . Similarly (replace u by 0)... By Lemma 5. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +.. Proof of Theorem 5.4. in particular his Proposition (2.1 and Y„ . . ..' = y is BM.6. and clearly Zi. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). . .+y 1 p"F(Yn > u) P)Pn1 P/(1 .
3. nontrivial and we refer to Asmussen [22]. Asmussen & Teugels [53] studied approximations of i (u. More precisely. claim size distribution B. the probability that is exceeds u is then B(u . Proof of Theorem 6. The heuristic motivation is the usual in the heavytailed area. The rigorous proof is. however. Then P(MT > u) .y) . T) when T + oo with u fixed. that MQ becomes large as consequence of one big jump. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 .1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}.. .e.(u) = P(V > u) = f f (y) dy .6. and premium rate p(x) at level x of the reserve. 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 . p(Y) and the result follows. cf.(3 u u J B(y) dy .1 Assume that B is subexponential and that p(x) > 00. one expects the level y form which the big jump occurs to be 0(1). and define the cycle as a = inf{t>0: Vt=0. V.B(u).2 Define M. i. the results only cover the regularly varying case. that fo p(x)1 dx < oo. 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). Corollary II./3Ea B(u). Then 0 (u) Qf "O ^) dy. x > oo.1. u (6.2. 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. Theorem 6 . Assume for simplicity that {Vt} regenerates in state 0 . We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. = supo<t<0.
Then D(u) = f(u)p(u) and. 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.P(MT > u) $B(u) Ft µ(1 . Hence f (u)r(u) = D(u) = Do(u) . 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 . by regenerative process theory. D(u) = DQ(u)/µ.q(u) Now just use that p(x) * oo implies q (x) + 0. .280 CHAPTER IX. u Notes and references The results are from Asmussen [22]. It is also shown in that paper that typically. where also the (easier) case of p(x) having a finite limit is treated . there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential.
281 . vrN(z . topics of direct relevance for the study of ruin probabilities are treated in more depth. 4Z)... z) 2 = Zit NE ii ii According to standard central limit theory . and this is the form in which the result of the simulation experiment is commonly reported. ZN. where a2 = Var(Z ). Rubinstein [310] or Rubinstein & Melamed [311] for more detail . Hence 1. .i. Fox & Schrage [77].z) 4 N(0. The crude Monte Carlo ( CMC) method then amounts to simulating i.96s z f (1. Ripley [304].. 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. replicates Zl.Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . We shall be brief concerning general aspects and refer to standard textbooks like Bratley.2) is an asymptotic 95% confidence interval . estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2.d.
Z = I inf Rt < 0 (0<t<T = I('r(u) < T).282 CHAPTER X. and in most cases this modest increase of N is totally unproblematic. variance reduction is hardly worthwhile. generated at the same time as Z. 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'. and many sophisticated ideas have been developed. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . an added programming effort. so that Z' is a candidate for a Monte Carlo estimator of z. there are others which are widely used in other areas and potentially useful also for ruin probabilities. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. Further. 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. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . We mention in particular ( regression adjusted) control variates and common random numbers. Letting Z' = E[Z I Y]. and a longer CPU time to produce one replication. Say that Var(Z') = Var(Z)/2. one can argue that unless Var(Z') is considerable smaller than Var(Z). Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). Therefore. However. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. v. The situation is more intricate for the infinite horizon ruin probability 0(u). typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). we then have EZ = EZ = z. conditional Monte Carlo and importance sampling.b(u. 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. We survey two methods which are used below to study ruin probabilities. This is a classical area of the simulation literature. SIMULATION METHODOLOGY In the setting of ruin probabilities.
Nevertheless. it may often be impossible to describe P in such a way that it is straightforward to simulate from P).e. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio.v. L1).3) Thus. and the problem is to make an efficient choice. .E [Z Z]2 = z2 . a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. .z2 = 0. In order to achieve (1.3). using the CMC method one generates (Z1.[E(LZ)] = E Z2 Zz .zrs. Thus we cannot compute L = Z/z (further. it appears that we have produced an estimator with variance zero. This may also be difficult to assess . L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). Thus..1. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. Then z Var(LZ) = E(LZ)2 . (1.. the argument cheats because we are simulating since z is not avaliable analytically. (ZN. However. even if the optimal change of measure is not practical.96 sis v^ N 2 1 where srs = N j(LiZi . i. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. but tentatively. L such that z = EZ = E[LZ]. one would try to choose P to make large values of Z more likely. GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. 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. . To this end.
say 10%. i. say of the order 103 or less. if z is small..96oz /(zV) = 0. The CMC method leads to a variance of oZ = z(1 . To introduce these. 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. and let Z(u) be a Monte Carlo estimator of z(u). the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . In ruin probability theory. Thus. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency.1.5 or even much smaller . let z(u) = P(A(u)).z) 1001. SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . 10 .B = iP(AB) = P(BIA). a confidence interval of width 10 4 may look small. u + oo. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i.96 2Z ( 1 . This leads to the equation 1. We then . Z = I(A) and A is a rare event. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}).1. However. it does not help telling whether z is of the magnitude 104. N ..100 . I. An example where this works out nicely is given in Section 3.e.96 2 z2 z increases like z1 as z . However. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . we may try to make P look as much like P(•IA) as possible. assume that the A(u) are rare in the sense that z(u) * 0.e. as is the case of typical interest. z I. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. and further it is usually not practicable to simulate from P(•IA). Again. in terms of the halfwidth of the confidence interval. Z z V5 In other words .284 CHAPTER X. For each u. the optimal P is the conditional distribution given A. large sample sizes are required.z) which tends to zero as z ^ 0. but if the point estimate z is of the order 105.0.z) 1 > 00.e.
Generate K as geometric. 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. which gives a logarithmically efficient estimator . it is appealing to combine with some variance reduction method . F(K = k) = (1 . O (u) = z = EZ. with common density bo(x) = B(x)/µB and K is geometric with parameter p. This allows Var(Z(u)) to decrease slightly slower than z(u)2.X1 + + XK. 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. Otherwise. it is not efficient for large u .0. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. 3. However. If M > u. where M = X1 + • • • + XK.. the PollaczeckKhinchine formula III. Generate X1. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. 2. where X1.. Thus. Let M . this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded.p)pk.1) may be written as V) (u) = P(M > u). where Z = I(M > u) may be generated as follows: 1.. XK from the density bo(x). .d. see Asmussen & Rubinstein [45] and Heidelberger [190]. and in practice. . let Z +. i. . P(K = k) = (1 . are i.4).(2.e. We shall here present an algorithm developed by Asmussen & Binswanger [ 271.log Var(Z(u)) lim inf > 2 u+oo . The term logarithmic comes from the equivalent form .i.p)pk. The algorithm gives a solution to the infinite horizon problem . let Z +. According to the above discussion. logarithmic efficiency is almost as good as bounded relative error. but as a CMC method .4) for any e > 0.1. so that NE (u) may go to infinity.2. the mathematical definition puts certain restrictions on this growth rate. X2. Therefore .log z(u) of (1.. Notes and references For surveys on rare events simulation.
.... and let Z(2)(u) = _ P (SK B0((u > u I X(l)..1) V)(u) .XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0.. For the simulation..X(2). just note that EZ(1)(u ) 2 > E[Bo (x . 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. However. form the order statistics X(1) < X(2) < . SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail.X1 .XK_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.S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1)... XK... Then (cf. note first that To check the formula for the P(X(n) > x I X(1). we generate only X1.XK1] = EBo(uX1 . XK1..p/(l .X(2).2. . compute Y = u . conditional probability.. . assume in the following that Bo(x) . X1 + + XK_ 1 > x when X1 > x. y < 0). A first obvious idea is to use conditional Monte Carlo: write i.Xl ...X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) . Xl > x.X(K1)) ..p)Bo(x).. The idea of [27] is to avoid this problem by discarding the largest X. < X(K) throw away the largest one X(K). and considering only the remaining ones.. Z(1) (u) has a smaller variance than Zl (x)... . asymptotically it presents no improvement : the variance is of the same order of magnitude F(x). Z(1)(u) is defined as 0). K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X. and that Bo(y) = 1. Thus...b(u) = P (Xl +•••+XK>u) = EF[Xl + ....286 CHAPTER X..SK1)2. we thus generate K and X1i .+XK > uIXl... So. Theorem IX.L(x)/x`' with a > 0 and L(x) slowly varying. To see this. As a conditional Monte Carlo estimator .
1) . l)) . .modulated model P(r+ < oo) and G+ are not explicit ). 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7".. X(2). X(n1)) P(X(TZ) + S(.S (n1)) V X (. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . However .. l)) BO(X(n1)) Theorem 2 .. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. 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. . The algorithm is sofar the only efficient one which has been developed for the heavytailed case. For practical purposes. .. B. X (. BL instead of 0. and we refer to [27]. X(2). BL(dx) = e7sB(dx)/B[y]. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . ..1 is elementary but lengty. X(n1)) Bo((x . using 13L.Khinchine formula and importance sampling . Thus. the algorithm for generating Z = Z(u) is: 1. Then the algorithm given by { Z (2) (u) } is logarithmically efficient. ._1) > P(X(n) > _ X X(1).S(n_1) I X(1). 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. X(2). and simulate from FL. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1).y. Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . 111. that is.5). BL by I3L = /3B[y]. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. for the purpose of recording Z(u) = erysr(u). in the renewal or Markov.3. and define )3L. 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. Notes and references The proof of Theorem 2. X . Asmussen ..u is the representation 0(u) = e7sr(u) overshoot (cf. .
SIMULATION METHODOLOGY 3.288 2.. BL). with distribution F. The answer is no. + X. = X1 + . . 4. b different from . u It is tempting to ask whether choosing importance sampling parameters . be i.F. and the change of measure F r FL corresponds to B > BL.1) (simulated with parameters ^3. Let Sf0 CHAPTER X. X2.d. r(u) < oo) and FL (both measures restricted to. The algorithm generalizes easily to the renewal model . return to 3. Let X1. There are various intuitive reasons that this should be a good algorithm. and assume that µF < 0 and that F[y] = 1. A > AL as in Chapter V. Xi = U. Generate T as being exponential with parameter . B) is not logarithmically efficient when (/3.. More precisely. let Z F e_'s. let S.. P'[y] < oo for some ry > 0. If S > u. M(u) = inf {n : S„ > u}. 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.(u)) are asymptotically coincide on {r(u)} < oo. We formulate this in a slightly more general random walk setting '. Proof Just note that EZ(u)2 < e .2 The estimator (3. The proof is given below as a corollary to Theorem 3..i.S+U . namely ELery£("). Let FL (dx) = 'For the renewal model.2ryu _ z (u)2/C2. In fact: Theorem 3. so that changing the measure to FL is close to the optimal scheme for importance sampling .3.Q..l3 and U from B. one must restrict attention to the case 4µB > 1. the results of IV. . It resolves the infinite horizon problem since FL(.r(u) < oo) = 1. and avoid simulating the known part e7".. and we have: Theorem 3. Ti.T.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. Let S . b) # (/3L. to deal with the infinite horizon problem .QL. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.. Otherwise. BL could improve the variance of the estimator . In detail . the discussion at the end of Section 1b.7 tell that P(. cf.
.i. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). it thus follows that for 0 < e < e'/ELXi. 1. . F(XM(u)). + KM(u))} = exp {ELM(u)(E . are i. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) .. Here ELK. Proof The first statement is proved exactly as Theorem 3 .2 > 0.2ryELXi)} ... where e' = EL Iog dFL (Xi) > 0 by the information inequality. K2.P = FL.+KM(u)}.2ryELXi. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + .d..yu+elu u +oo etry' 1 > lim up C2e2..3 The estimator (3. . is not logarithmically efficient.2) dF Theorem 3. where Kl og (X) (j) 2 ) = log dFL (Xi) . write W(F IF) _ F(XI). When F # FL. EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.3. Since K1. = c'. (3. Since ELM(u)/u + 1/ELXi.yu = G.2) (simulated with distribution F of the X3 has bounded relative error when ... By the chain rule for RadonNikodym derivatives. 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+. For the second.2'X1 .. More generally.
optimality is discussed in a heavy traffic limit y 10 rather than when u + oo.4 indicate that we can expect a major difference according to whether y < 1/r. see e. First by the memoryless distribution of the exponential distribution . In [13]. all that needs to be shown is that if U' .4.T' D U" . with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45].1 If y > 1/ic'('y). The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L.T" has a left exponential tail with rate /3'. Then according to Theorem 3.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". generic interarrival times T' .'(y) or y > 1/r. T) with T < oo. The easy case is y > 1/k'(y) where O(u.290 which completes the proof. The results of IV. we write T = yu. As in IV. BL) has bounded relative error. . The optimality result Theorem 3. Further discussion is in Lehtonen & Nyrhinen [245]. SIMULATION METHODOLOGY u Proof of Theorem 3. Consider compound Poisson risk process with intensities /3'. In fact: Proposition 4. The queueing literature on related algorithms is extensive . 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 . U' . the references in Asmussen & Rubinstein [45] and Heidelberger [190]. claim size distributions B'.T". U' . we conclude by differentiation that Bo(x)=B' (x)forallx > 0.e.i. CHAPTER X.B'=B".T".T" > x) J /3"e0 yB (x + y) dy = . yu) is close to zk(u). from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. /3". T".'(y). Next.T' has a left exponential tail with rate /3' and U" .1 is from Lehtonen & Nyrhinen [244].2. then /3' B' = B". U".T' = U" . This immediately yields / = 3". /3'eQ'YR'( x + y) dy = . 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. so that one would expect the change of measure F 4 FL to produce close to optimal results. B" and generic claim sizes U'.3'eO'x f f P (U" .g.3.
The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu).log Var(Z(u)) _ . 7y (4.1).1) which is all that is needed here can be showed much easier . yu) is of order of magnitude a71. and in fact. Remark 4 .1). we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. yu)/z.O(u. lim inf uoo 27yu .4. .yk(ay) determines the order of magnitude of z'(u. T(u) < yu] . and that ryy > ry.2 The estimator (4. Bay) is logarithmically efficient.4. Further .8).yy> 2 .log 4')u) 4 u (Theorem IV. We next consider the case y < 1/r. We recall that ay is defined as the solution of a'(a) = 1/y.4.1) so that z(u) = zP(u. 3 Theorem IV. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.1. Let Qy2 = .(ay). yu) = eayu Eay Leay^(u)+r(u)K(ay).yu. (4. Bounding u ELZ(u)2 above by a7u.2) Since the definition of ay is equivalent to Eay r(u) . (4. the result follows as in the proof of Theorem 3.4.3) and we have: Theorem 4. yu) in the sense that .log Var(Z(u)) l im of .3) (simulated with parameters /gay.5) follows. T(u) < yu] e Hence by (4. T( u) < yu] e2ryyuEay le 2ay^(u). that ryy = ay .(u) * 1 (Theorem IV. Proof Since ryy > y.8 has a stronger conclusion than (4. one would expect that the change of measure F Pay is in some sense optimal.'(7).to g x ( u ) u u so that (1.
ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV.1) where the identity for Vi(u) requires that Vt has a limit in distribution V.yu)/(uyu1/2) . N(0.292 CHAPTER X. related discussion is given in a heavy traffic limit q J.Qyu1/2 < T(u) C yu e.1) (see Proposition IV.1): then by Proposition A1. there exists a dual process { V t} such that i. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . > u) = E f I(VV > u) dt 0 (5.4. In most of the simulation literature (say in queueing applications).o .yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.4.4).1) may be useful.2).(av)Eav l e. zi(u) = INV. 0 Notes and references The results of the present section are new.. Hence lira inf log ryyu + vyu 1/2 tc(ay) .u1/2 < r(u) < yu Le ] l = e. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W). 5 Regenerative simulation Our starting point is the duality representations in 11.T) = P O<t<T inf Rt < 0 = P(VT > u).2) . and (5. One main example is {Vt} being regenerative (see A. However. yu . 0 rather than when u 3 oo.uaoo U That lim sup < follows similarly but easier as when estimating En. '%(u) = P I info Rt < 0) = P(VV > u).1) is used to study Voo by simulating {Rt} (for example.o.7y x(u) > hm inf u+Oo U . the object of interest is {Vt} rather than {Rt}. Z (u)2 above.b(u. (5.a vt(u). yu . SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .3: for many risk processes {Rt }. In Asmussen [13]. we believe that there are examples also in risk theory where (5.3.
. Z2 = N (X21' + .d.. To derive confidence intervals . oh) for h : R2 ^ R and Ch = VhEVh.. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. . Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed.E). provides estimates for F ( V. 2. Therefore .5. Z2 . the regenerative estimator z%(u) is consistent. Z2 a4* z2. i (^(u) . EZ1'i = z1 = Ew. EZ2'i = z2 = E Thus. Then (Z1z1i Z2z2 ) 4 N2(0. record Zi'i = (Z1'). (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo.... REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. Thus.3) ... For details . and Z2'>) where Zi'i = w. Taking h(zl..+Z(N) z 1. + Z1N>) . letting J0 'o I (Vt > u) dt . i. consider first the case of independent cycles . which we survey below . Zl the LLN yields Z1 a$' Z(1) +. z2) z2/z1 yields Vh = (z2/z2 1/zl). 02) (5.. . let E denote the 2 x 2 covariance matrix of Z('). The method of regenerative simulation. > u) (and more general expectations Eg(V.t(u)) 4 N(0. is the cycle length. z2)) > N(O. Then Z(1). )). Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. Z1 = (Zl1i +. Vh = (8h/8z1 8h/ 8z2). Z(N) are i . Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). + Z2N)) . For the ith cycle. j = 1..h (zl.
2. 6 Sensitivity analysis We return to the problem of 111 . However . Before going into the complications of ruin probabilities . () depending on C.Z) ^Z(=) .g S12 (5. with distribution depending on a parameter (. in some situations it may be the only one resolving the infinite horizon problem . In 111.96s/v"N. 9.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . see e. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 . The regenerative method is not likely to be efficient for large u but rather a brute force one. v. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . Then z(() = f cp(x) f (x. () dx so that differentiation yields zS d( fco(x)f(x.v. () dx = E[SZ] f(X. Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . Let X have a density f (x. There is potential also for combining with some variance reduction method. Rubinstein [310] and Rubinstein & Melamed [311]. () dx f Ax) (dl d()f (x' () f ( z.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW .C)dx = f w(x) d( f ( x. the expectation z = EZ of a single r. Z of the form Z = ^p(X) where X is a r .294 where 01 2 CHAPTER X.9. Notes and references The literature on regenerative simulation is extensive. We here consider simulation algorithms which have the potential of applying to substantially more complex situations. consider an extremely simple example .0 . to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . asymptotic estimates were derived using the renewal equation for z /i(u).5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1.g.
For example . however . SZ is an unbiased Monte Carlo estimator of z(. just take cp as an indicator function . The likelihood ratio up to r(u) for two Poisson processes with rates /3. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. Let M(u) be the number of claims up to the time r(u) of ruin (thus. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. C)). Thus.() d( is the score function familiar from statistics . cp(h(U. So assume that a r. /3 is 0. () = . To see this. ()) h((U.r. IPA will estimate zS by 0 which is obviously not correct. = E [`d (h(U. for some Co = (o(U). For IPA there are.v. Then z(() = Ecp(h(U. () is increasing in C. Then . () = d log f (X. () can be generated as h(U. C) f(X. C). the Poisson rate /3 in the compound Poisson model.log U/(. () Thus. For the SF method. A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. The following example demonstrates how the SF method handles this situation.t. In the setting of ruin probabilities . Example 6 . Thus . zc = E [d co(h(U. 11 /3oeOoT.r.1 Consider the sensitivity tka(u) w. this is usually unproblematic and involves some application of dominated convergence . ()) is 0 w . if f (x. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . r(u) = Tl + • • • +TM(u)). with density f (x. () _ (eSx. ( where h( (u.t. I(r(u) . p. giving h( (U. () = (8/8()h (u. Infinitesimal perturbation analysis (IPA) uses sample path derivatives. one. one can take h (U. () where U is uniform(0.1). () = log U/(2.6. say W(x) = I(x > xo) and assume that h(U. ()) d( hc(U. ()). cp' (h(U. () is an unbiased Monte Carlo estimator of zS. /3o is M(u) Oe (3T: < oo) .
3 (u) (to generate Zp (u).296 CHAPTER X. 4) that V5.3L. In the setting of ruin probabilities.t. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. for different models and for the sensitivities w. different parameters. We then arrive at the estimator ZZ(u) = (M(u) . the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . BL).T(u)) e7uerVu) for ?P.3 (u) is of the order of magnitude ue7u. However. in part for different measures of risk than ruin probabilities.0(1) so that in fact the estimator Zf(u) has bounded relative error. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). whereas for the SF method we refer to Rubinstein & Shapiro [312].T(u)) I(T(u) < co) ] . we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . Example 6. There have been much work on resolving the difficulties associated with IPA pointed out above. We recall (Proposition 111. a relevant reference is VazquezAbad [374].r. SIMULATION METHODOLOGY Taking expectation.r. . change the measure to FL as when simulating tp(u).1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. j3 and letting flo = 0. Thus. the risk process should be simulated with parameters . differentiating w.9 .t. To resolve the infinite horizon problem .
(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. defined as Ro = u (with u E {0.. Y'a(U) = P(T (u) = r+(a)) = 1 ..i. '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 ).Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion..P(•r(u..1...g. as e.+• • •+X. R„ = u+X. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo.d. a) = r(u) A T+(a).. }). and {1. 297 . a) = r(u)). wherel T(u) = inf {t > 0 : Rt < 0} .(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. are i.. The twobarrier ruin problem The twobarrier ruin probability 0. .. T+(a) = inf It > 0 : Rt > al. in the Bernoulli random walk example below. X2. where X1. That is.. T(u. in most cases . either this makes no difference (P(R. with P(Xk = 1) = 9. Besides its intrinsic interest . Consider first a Bernoulli random walk.1}valued .
4) by ea(u+Xl+.1 For a Bernoulli random walk with 0 0 1/2. and insertion shows that ( 1..r(u... 7/la(a .298 CHAPTER XI.a) = 0) + zap ( R.1) = (19)4/'0(a3)+9ba(a1). where a is any number such that Ee°X = F[a] <oo. The Lundberg equation becomes 1=F[ry]=(19)+9z. Wald's exponential martingale is defined as in 11. i.1) is solution.(4.e..1. In a general random walk setting . The martingale is then {zuzXl+•••+X„ } = {zR° }.(1B)u oJ 0.+Xn) F[ a]n n=0.2) Oa(a . one elementary but difficult to generalize to other models. z and the solution is z = (1 .(u) I\ e = 1 oa ' ()i a = u.y] = 1.2). u + 1.0)/0. u Proof 2. By optional stopping.o)'t/1a(a ..o» = z°P (RT ( u..1) o If 0 = 1/ 2.. = (1 .o)T/la (1) + 8z/'u(3). then 'Oa(u) _ au a We give two proofs . We choose a = ry where ry is the Lundberg exponent. and in view of the discrete nature of a Bernoulli random walk we write z = e7. = z°Va(u) + za(1  . C1_0\a. MISCELLANEOUS TOPICS Proposition 1.a(u)).a) Y. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). tba(2) _ (1 . (1. zu = EzRO = EzRT(u.. the solution of F[.. Proof 1. and the other more advanced but applicable also in some other settings.
} yields e7u = Ee7R° = e°Wa(u) + e7a(1 . u Proposition 1.} 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)).4 For a Brownian motion with drift u > 0. If p = 0. Corollary 1.ba(u) = e2µa .0a(u)).zu)/(za . thenz1 (u)=1. If p<0.1 If p = 0.1. then Vi(u) = 1.1 yields 't/la(u) = (a . i1(u) = e211 . {R.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . (1. 1h (u) = a el u \1 If 9 < 1/ 2. then Proof Since 'Oa (U)  au a Eea(R°.2 For a Bernoulli random walk with 9 > 1/2.5) . and solving for 9/la(u) yields Z/)a(u) = (e 76 . BROWNIAN MOTION..u)/u. RANDOM WALK. Proof Let a+ oo in (1.e7u)/(e7° . If 9 = 1/2.1) for p # 0.1). {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. (1. Then for p 0 0. . However.1). TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za .u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. pa( u) _ u Corollary 1.• a2µa e2µu .2) is trivial (z = 1). Applying optional stopping to the exponential martingale {e7R.
this immediately yields (1. letting a * oo yields the standard expression pe7u for . 1 . and thus one encounters the problem of controlling the undershoot under level 0.a) = a on {r (u.0 (u) (where u p =.3.0(a) 0 < u < a. Here is one more case where this is feasible: Example 1.+^a(u))^(a) If 7k(a) < 1. however. Here the undershoot under 0 is exponential with rate 5. 5). CHAPTER XI.a) < 0) + e7°P (R(u. (u) _ O(u) . 7/'(u) = 1).a ) < 0) + e 7aF ( R (u. implying R(u. .616). It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1. Ic 5ry 'pa(u) Using y = 6 .e7a Again . (1.7) . passing to even more general cases the method quickly becomes unfeasible (see. For most standard risk processes .7/la(u)).5 Consider the compound Poisson model with exponential claims (with rate..5a). 0. valid if p < 1 (otherwise . 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.e7a (u) = 6 /0 . However. a) I R(u a ) < 0] P (R(u .a) = a ) + e ' ° ( 1 .300 Proof Let a * oo in (1.vi(a) Proof By the upwards skipfree property. we obtain 'Oa a7u . VIII.a) = a) = 5 y = P (R (u.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. the paths are upwards skipfree but not downwards. 7O(u) = 7/la(u) + (1 .4). say.7). and hence e7u = Ee7Ro E [e7R(.a) = r+ (a)} and similarly for the boundary 0.
P(MT > u) = P(ST > u) + P(ST < u.2 .8) Proof In terms of the claim surplus process { St} = {u . Hence P(MT>u.ST<u) = P(MT>u.d.r(u).. = 1 . of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .ST>U).11 ) is the same as (1.8 Let {Rt} be Brownian motion with drift . For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. Corollary 1. RANDOM WALK.10) Pµ (T(u) < T) !. + µ2T) } . and hence Pµ('r(u) E dT) = Eo [e µsr(. ( 1.. MT > U) = P(ST > u) + P(ST > u) (1.. BROWNIAN MOTION. Here {St } is Brownian motion with drift 0 (starting from 0). (1. T ) = P(T(u) < T ) = 241..8 ). we have ili(u.1. 0(u. T(u) E dT.µ so that {St} is Brownian motion with drift µ . = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T . (1.9) = 2P(ST > u).)_ _( u)µ2 /2..4) I = .Rt}.f.1a for computing ruin probabilities for a twostep premium function. (i).µ T I + e2µ"4) ( .T) P(MT > u) where MT = maxo<t<T St. 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 . 10) follows then by straightforward differentiation. Then the density and c. For µ # 0.7 For Brownian motion with drift 0. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities. the density dPµ / dP0 of St is eµsttµ2/2. MT > u) = P (ST > u) + P (ST > u. and (1 .11) VIT ) Proof For p = 0.µ%T (1. TWO BARRIERS 301 Note thas this argument has already been used in VII.
Here {2T( (v}TT)/2) v=T.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). see e.302 CHAPTER XI. Vi(u.T) = P(ST = u) + 2P (ST > u). such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 . (1. Breiman [78] or Karlin & Taylor [222] p. u Small modifications also apply to Bernoulli random walks: Proposition 1. If this assumption fails. oo).T+2.9). e. 0 0 (1. is (1. S(oo) = f c s(y)dy. and (1. that 0(u). We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.10 Consider a diffusion process {Rt} on [0.8 also applies to the case 9 54 1/2.11) then follows by checking that the derivative of the r. We finally consider a general diffusion {Rt} on [0.13) with 0 as lower limit of integration.. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. as defined in (1.10). and in a similar spirit as in VII. T are integervalued and nonnegative.T (1. Let s(y) = ef0 ry(. Theorem 1. MISCELLANEOUS TOPICS which is the same as (1. Thus. The same argument as used for Corollary 1.T)dx. whenever u.T2.12) is the same as ( 1.g. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x).g.9) goes through unchanged. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. Proof The argument leading to ( 1. as defined above as the probability of actually hitting 0. S(x) = f x s(y)dy.10) and that the value at 0 is 0..3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion. 226)..h.s. but we omit the details.9 For Bernoulli random walk with 9 = 1/2. the behaviour at the boundary 0 is more complicated and it may happen. is finite for all x > 0. oo) with drift µ(x) and variance a2 (x) at x..12) P(ST = v) = 0 otherwise.
S(u) (1. Letting a T oo and considering the cases S(oo) = oo. if (1. 0 in (1. Using s'/ s = 2p/a2. Lemma 1. i. (1 . Then YIa. we can ignore the possibility of ruin or hitting the upper barrier a before dt. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] . the function S(x) is .16) yields 4b (u) = 1 .b(Rdt). 191195 for material related to Theorem 1.b = 0 implies that VQ b/s is constant. then.10.16). 1'.10. then 0 < 2l. A classical reference for further aspects of Bernoulli random walks is Feller [142]. S(oo) < oo separately u completes the proof.e LVa. E„ q(Rdt) = q(u)+Lq(u)dt. TWO BARRIERS 303 for x > 0.14) S(oo) < 00.17) Hence L.6 to Markovmodulated models .b (Rdt) = Oa. If b < u < a.b(u) + L. [117].b('u) = Eu &0.b('u) = Eu . RANDOM WALK.b(u) be the probability that {Rt} hits b before a starting from u. Further references on twobarrier ruin problems include Dickson & Gray [116]. and we get Wo.b(u) = S(a) . where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. A good introduction to diffusions is in Karlin & Taylor [222].16) S(a) . Notes and references All material of the present section is standard.b = a+/3S.11 Let 0 < b < u < a and let t&0. 0 Proof of Theorem 1. see Asmussen & Perry [42]. BROWNIAN MOTION. O.S(b) Proof Recall that under mild conditions on q. The obvious boundary conditions '0a. In view of (1.13) is finite for all x > 0.14) fails.ba.b(u)dt.S(u)/S(a).e. Assume further that S (x) as defined in (1.1. Wa.ba. (1.(u) < 1 for all u > 0 and ^ S^ Conversely. If (1.b(b) = 1. For generalizations of Proposition 1. see in particular pp. 15) i.b(a) = 0 then yield the result. Letting b J. . so that Y)n. b = 0.0(u) = 1 for all u > 0.
. 111 .4. y > . with the drift and the variance depending on an underlying Markov process .1) (2.4.13)).9 ) and optional stopping applied to the stopping time r(u) A T.(a) (2. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.2) C_e7u < t(u) < C+e _7u. Lo is a martingale (cf.4) I. information on ruin probabilities can be obtained . IV. yu) '+/1(u) .5) A martingale proof of (2. and here are alternative martingale proofs of the rest . Lo I. defined by the density 1/va(u)s(u) showing up in (1.17).aRo . but by duality.2. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.)AT . yu) where W (ay) = y. Remark 11. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0.5.o•K(a) = Ee .1 ) was given already in II.ytc (ay).aR. equivalently. Markovmodulated Brownian models .(7) . (2.304 CHAPTER XI.5): _ z/'(u) < e 7u.6) . They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. (2.(T(u)AT) r.3. yielding eau = Ee.(. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature. The emphasis is often on stationary distributions . correponding to piecewise linear paths or . Another basic quantity is the speed measure M . (2. 1 y < k (y). (2. variance 0.3) < e 7yu.t&(u. much of the literature dels with the pure drift case.1. is currently an extremely active area of research. 7y = ay . one works instead with a lower limit 5 > 0 of integration in (1.6.
d. Equivalently.4): We take a = ay in (2. Proof of ( 2.yu))• b(u.2.E [e. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.2): As noted in Proposition II. (2. For (2.yuk (ay)(u&(u. dr) denote the conditional distribution of (T(u). it follows easily from (2. u Proof of (2. and the proof of the lower inequality is similar.T)  V. RT(u)_) given r(u) < oo.3). when Rt_ = r.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. eyuk (ay) = e7yu e > eyu"(ay ) ij(u.1. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.yu) Y Similarly for (2..(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0). Let H(dt. 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).f.(u. y > r.1 .6) below by 1 E Le7Rr(. so that i/1(uL yu) < eayu .7R.4). .6).3). we have ic(ay ) < 0 and use the lower bound E [e7Rr („). dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.)r(u)r.B(r))/B(r). we have tc(ay) > 0 and we can bound (2.1. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.( u ) I T(U) < 00] .6) with = 'y that eyu . Rt has distribution B(r + dy)/B(r). (B(y) . A claim leading to ruin at time t has c.
(B) = log EeOX 1 is defined for sufficiently many 0. og For sequences fn. The classical result in the area is Cramer's theorem./n E I) for intervals I C R. However . Cramer considered a random walk Sn = X1 + . which in the setting of (3. large deviations results have usually a weaker form.306 CHAPTER XI. then P C S.. Thus. (3. + X. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative .. cle .1) does not capture the \ in (3. not quite so much in insurance risk. For example. large deviations results been. Accordingly. nroo n n /// Note in particular that (3.1).^ e nn 1 > x n 0o 2xn (3.1 We will go into some more detail concerning (3. 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. The limit result (3. . . 1) but only the dominant exponential term . and gave sharp asymptotics for probabilities of the form P (S.means (as at other places in the book) that the ratio is one in the limit (here n * oo).. we will write fn 1. e. gn with fn + 0 . if x > EX1. the parameter will be u rather than n).1) is an example of sharp asymptotics : . ri.g.2).?n typically only give the dominant term in an asymptotic expression .1) where we return to the values of 0.. logarithmic asymptotics .1) amounts to the weaker statement lim 1 log P I Sn > x I = 17.the correct sharp asymptotics might as well have +. The last decades have seen a boom in the area and a considerable body of applications in queueing theory. Thus .3na with a < 1. however .2) can be rewritten as F (Sn/n > x) 1g a'fin. in being capable of treating many models beyond simple random walks which are not easily treated by other models . gn 4 0. The advantage of the large deviations approach is. Example 3.gn if nioo lim 109 fn = 1 log gn (later in this section.nn or C2e. v2 later. its generality. (3. such that the cumulant generating function r. and that a considerable body of theory has been developed.
the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox . we get P(Sn/n > x) E [e9nx +nK(9)9" '. S rtn > x 1. In fact. Define .4) n Next. i.sseo f which in conjunction with (3.960/) * 0.2). exponential change of measure is a key tool in large deviations methods. More precisely. nx < Sn < nx + 1. since Sn is asymptotically normal w. V > 0 e. 2 where o2 = o2(x) = rc"(0).3) is put equal to x.3. Since P nn > x) = E {e_8 ' ( 9). Most often.r.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn .425.t. if we replace Sn by nx + o / V where V is N(0. and hence for large n P(Sn/n > x) > E [e.1). the LegendreFenchel transform or just the Legendre transform or the large deviations rate function).4 enn +1. which is a saddlepoint equation .q = rc* (x).4) immediately yields (3.96o /] > 0.(0)) e 307 (other names are the entropy.rc(0) where 0 = 0(x) is the solution of x = rc'(0). (3.r. of P(X1 E dx) = E[e9X1K. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3.(e)i XI E dx]. LARGE DEVIATIONS Define rc* as the convex conjugate of rc. P with mean nx and variance no.e.9S„+n' ( 9). we have P(nx < Sn < nx + 1.the mean rc'(0) of the distribution of X1 exponentially tilted with 0. rc*(x) = sup(Ox .
Ee9X n < oo for e < 0 < e. and the WentzellFreidlin theory of slow Markov walks. . commonly denoted as is the saddlepoint approximation. The substitution by V needs.. is differentiable at ry with 0 < K'(y) < 00. Xn given by Fn(dxl.e < 8 < y + e. that is..s.3 For each i > 0. be a sequence of r.. which is of similar spirit as the dicussion in VII. We further write µ = tc'(ry).. (iii) #c (8) = limn.e < 8 < y + e. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. to be made rigorous. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 .. integrates to 1 by the definition of Icn)..'(u) )Ng a"u.. .. (iv) tc(ry) = 0 and r. 1].308 CHAPTER XI. 1) and no such that Sn . . Xn) and sn = x1 + • • • + xn (note that the r. there exists z E (0.2 (GLYNN & WHITT [163]) Let X1. Sanov's theorem which give rare events asymptotics for empirical distributions. Further main results in large deviations theory are the GartnerEllis theorem. (ii) lim supn. Then i/. MISCELLANEOUS TOPICS which is the same as (3. For the proof.3... We shall need: Lemma 3 . and write Sn = X1 + • • • + Xn.'s.p > 7 < zn. however. Assume that there exists 'y. 260 for details..o log Ee9Sn /n.1). see Jensen u [215] or [APQ] p...dxn) where Fn is the distribution of (X1i ..v. . X2. In the application of large deviations to ruin probabilities. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y . Mogulskii's theorem which gives path asymptotics. n Icn(0) exists and is finite for ry ... r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo)./^ >7 < zn n for n n0. we introduce a change of measure for X1. Pn Sn1 .dxn) = 05nKn(7)Fn(dx1. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. ..h.
µ?7 .s. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1. 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.3. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0... h. The rest of the argument is as before. 0. can be chosen strictly negative by taking 9 small enough.h. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3. The corresponding claim for Pn(Sn/n < µ .. This proves the existence of z < 1 and no such that Pn (Sn/n > µ. ( U) P(S.YS.m(7). the r .n m µ 1 + rl . This establishes the first claim of the lemma . We first show that lim inf„_. in particular the r.91) + o(O ) as 0 J.n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.Bµ .077 n^oo n and by Taylor expansion and (iv ).. > 1 +17] m(7).> . S.n > u ) = [ Em [em Em 1e. S.71 < e and jq9j < e. h.y) .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 +.s. for Sn. is of order .+r. Then V. mµ Sm > u] km e7Sm+n.ne(p+ 17). log zl'(u)/u > 'y. it is easy to see that the r.2. Since I EeqOX „ ] 1/q is bounded for large n by (ii).ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) .]1/q = e. Clearly.2. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7)..s.77) follows by symmetry (note that the argument did not use µ > 0). 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.r (7) n = e.W. u Proof of Theorem 3.+r7) < zn for n > no.
I2 = F(T(u) = n).6) for some z < 1 and all n > n(E).I < µl1 1+77 I M 1_ 1+277 S. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (.(Y). For lim supu.+wn(7).YS +^c CHAPTER XI.. Pn \ > la+ 8 I < zn (3. P(T(u) = n) < P(Sn > u) = En [e7S.7) so that n(b) I1 < e'Yu E en..log z) /2 and Sn Fn\ n >lb+S) <Zn. n=1 . Obviously.. logO(u)/u > ry..n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E..310 ]Em I e.. Sn > u] < eYu+Kn(7)pn(Sn > u) (3. 0 yields liminfu __.. MISCELLANEOUS TOPICS (7).n(ry)/u 4 0andm/u* (1 + r7)/µ. I > IL exp `S. 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). 3.3. we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n). and since Ic.(•) goes to 1 by Lemma 3. this is possible by (iii).0 log i'(u )/u < 'y. (iv) and Lemma 3.
S. Sn1 C U. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) . eryu en logz/2p n nt n. 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 . C 26u `p / +1 I e6u(1+6)/µ (3. u . we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.10) 00 I4 < E F(Sn_1 < u.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.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).' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^.3. 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 .
9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . (7 + a) < 2arc'(7).('+'Y). 4 there is an aj > 0 and a cj < oo such that Ij < c3e.3ui where . IV. e'. 2. ryue«iu .7' a"ju. we replace the bound P(Sn > u ) < 1 used in (3. the last steps of (3. 13 = P(T (u) E (u(1 b)l^ (7). this is straightforward since the last inequality in (3. cf.2. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. 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.xl/2 to give the desired conclusion.u(1+b)/rc'(7)).b)/i(7).11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 ..(u) = I1+I2+I3+I4'^ ery( u). For 12.4. u(1 + b)/i(7)) Proof Since V. Corollary 3. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. say n n1. I2. u . it suffices to show that for j = 1.4 Under the assumptions of Theorem 3. For I. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7).312 CHAPTER XI. the typical time is u/rc'(7) just as for the compound Poisson model.8) by P(S.4/3rc'(y) > 0.Q is so small that w = 1 .z 1/z For I1. Then for n large. For 14.. Letting c11 = maxn<n. we need to redefine n(b) as L. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .
. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * . Xk+l) k=1 00 naoo n provided the sum converges absolutely.12) k=0.f.3(s) at time s.'(y) > 0.. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3. Thus the total reward in the interval [0. To verify these in concrete examples may well present considerable difficulties. An event occuring at time s is rewarded by a r.3. r.2 shows that the discrete time structure is used in an essential way. If {St}t> 0 is the claims surplus process.v. 2 is in force with y = 2p/wz. Theorem 3..g. but nevertheless. criteria are given in Duffield & O'Connell [124].1..1. and we conclude that Theorem 3 ..2 then immediately yields the estimate log F( sup Skh > u) a7u (3.. we shall give two continuous time examples and tacitly assume that this can be done. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process. V(s) with m. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0. Assuming that the further regularity conditions can be verified. Obviously many of the most interesting examples have a continuous time scale.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. t] is Rt = E V (Un) n: o„ <t .e. 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.(O) = 9µ+02 for all 9 E R. i.13) One would expect this to hold in considerable generality. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate . 09(9). 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... and in fact. 11 Inspection of the proof of Theorem 3. the key condition similar to (iii).14) is needed in both examples .LARGE DEVIATIONS 313 Example 3 .
we conclude that Cu) log e7 u (cf. 0 Example 3 .14) (to see this .s). the above discussion of discrete skeletons). O'n +S] is a r . derive . = U„ ( t .9t = /3 J t (Ee8U° i8l . nondecreasing and with finite limits Un as s T oo ( thus. the best estimator of /3µB based upon Ft.14). . 0 and since EeOUn(8) + Ee°U^ as s * oo. Thus.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. assuming a continuous premium inflow at unit rate. MISCELLANEOUS TOPICS are the event times.1) ds rt (3. n: o.1) . e. Example 3..15) . the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. More precisely.Lundberg model has the larger ruin probability. <t which is a shotnoise process. one would take p(t) = (1 + rt)At/ t.g. e > 0 such that ic('y) = 0 and that r. Then logEeOR° = J0 /3(s)(^8(9) .'`1 U.It. Thus. Most obviously.1) ds . It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot . i... (3. Thus by (3. this is not realistic . then the payments from the company in [on. Of course . Since the remaining conditions of Theorem 3. (9) < oo for 9 < 'y + C. a differential equation in t). We further assume that the processes {U1(s)}8>0 are i. Un(s). If the nth claim arrives at time Qn = s. An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics .0 and assume there are y. we have S. it contributes to St by the amount Un(t . where Ft = a(A8 : 0 < s < t). the Cramer.Q„) .v.1) ds .6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 .2 are trivial to verify.9t. At = . if the nth claim arrives at time a. leading to St = At(1+77) Joo t S8 ds.. Of course. is At . We let ic (9) = 3(EeWU° . we have rct (9)/t 4 ic (9). Un represents the total payment for the nth claim). but that a claim is not settled immediately.d.t. Kt (0) t (Ee9U"it8i J0 . 7 Given the safety loading 77.314 where the an CHAPTER XI.
(1 + r7) log t (3.16) i=1 o i=1 Let ict (a) = log Eeast . and since the remaining conditions are trivial to verify.18) Thus (iii) of Theorem 3.b(u) IN a'Yu (cf. uniform (0. the solution of /3(Eelu . we have Nt t N.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 .2 hold. standard exponential . LARGE DEVIATIONS With the Qi the arrival times.d./3. Thus.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3. 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. one has y > y' (3.1) or . It then follows from (3. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 . we conclude that t.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3. i. (3.(1 + 17)0µB = 0. (3.19) with equality if and only if U is degenerate.log Oi are i.3.i.1) .e. Indeed. 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 . typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .(1 +i) f > i= 1 s ds = E Ui 1 .d. equivalently.i. the Vi = . Ui Nt / t 01i 315 St = Ui .20) (3. To see this .
(1 + ri)y*x is convex with k(oo) = 00. For Example 3. x > x0. Further.d. the study is motivated from the formulas in IV.7. we are interested in estimating P(A > x) for large x.2. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. Lehtonen & Nyrhinen [244]. MartinL6f [256]. Further. see Nyrhinen [275] and Asmussen [25]. This implies n(y*) < 0. Therefore e7'U _ k(U) E [1+(1+77)y*U] . at time t. k(0) = 0. Further applications of large deviations idea in risk theory occur in Djehiche [122].2 expressing the finite horizon ruin probabilities in terms of the distribution of A. and since tc(s). In particular. a* (s) are convex with tc'(0) < 0 . MISCELLANEOUS TOPICS Next. with common distribution B and independent of Nt. though we do not always spell this out. k'(0) < 0. using that Ek(U) = 0 because of (3. much of the analysis carries over to more general cases. and k(x) < 0.1 . = P(N = n) = e(3an However.316 CHAPTER XI. this in turn yields y > y*. the function k(x) = e7*x . 0 < x < x0. 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. [257] and Nyrhinen [275]. . say one year. see also Nyrhinen [275] for Theorem 3. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. the proof of (3.20) is due to Tatyana Turova. The main example is Nt being Poisson with rate fit. For notational simplicity.i. are i.xo. [245]. rc*' (0 ) < 0. we then take t = 1 so that p.19). y = y* can only occur if U .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.. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. In addition to Glynn & Whitt [163]. assuming that the U. Dembo & Zeitouni [105] and Shwartz & Weiss [339].
Proposition 4. we define the saddlepoint 9 = 9(x) by EBA = x. only with 0 replaced by a9 and B by B9. e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. 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 . Then as x * oo.2) implies that the limiting Pedistribution of (A . i. The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9).9(Ax).3e(bo[a] .x)//3B"[9] is standard normal.1). This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution. Then Ee"A = e'(") where x(a) _ 0(B[a] . Vare(A) = s. A E dx] . THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. (4.1) where )30 = . Hence P(A > x) = E e [e9A+ ic(9). In particular. B"' [s] lim (B".1 Assume that lim8T8. For a given x. no(a) = logE9e'A = rc(a + 9) . 818' where s' = sup{s : B[s] < oo}.1). K'(0) _ ic'(9) = x.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx).ic(9) = .3B"[9].1. The analysis largely follows Example 3. B"[s] = oo. A > x)] = eex+K( e)E9 [e ."(0) = .[s])3/2 = 0.4.e.
then P(A > x) . Y satisfies 9(u) ti eu2/2(1 + ibu3) (4.318 CHAPTER XI. Notes and references Proposition 4. 4b The NP approximation In many cases .4) . Furthermore 00 b(x)Sdx < oo for some ( E (1. b is gammalike. 1 .3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion.v. The present proof is somewhat heuristical in the CLT steps.e. 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 . Remark 4 .2 If B is subexponential and EzN < oo for some z > 1. some regularity of the density b(x) of B is required.1 goes all the way back to Esscher [141]. Jensen [215] and references therein.1).x') where x' = sup {x : b(x) > 0}. For example. the distribution of A is approximately normal . [138]. leading to P(A > x) :. A covers the exponential distribution and phasetype distributions. i. under the Poisson assumption (4.3) and related results u for the case of main interest .EN B(x). b is logconcave. large x. it holds that EA = .(D X . B covers distributions with finite support or with a density not too far from ax° with a > 1. b(x) = q(x)eh(z).l3pB. Var(A) _ ^3p.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. bounded with b(x) .Q{AB (4. In particular. Thus .1 yields: Proposition 4. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. or. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. For a rigorous proof. see Embrechts et al. it is quite questionable to use (4. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. and (4.ycix °ie6x B.2i and that (A .2) is often referred to as the Esscher approximation. For example. more generally. For details. 2). In fact. either of the following is sufficient: A. just the same dominated convergence argument as in the proof of Theorem 2.
are small.f.3& (y).l = EY. the CLT for Y = Y6 is usually derived via expanding the ch. resp. .1). then P(Y < y) 4(y) .equantile in the distribution of Y.c2i. in particular. K3 = E(Y .2 2 . Var(Y) = 1 as above . Let Y = (A . If this holds ... so that 1(u) 3 exp { .EY)3.EA)/ Var(A) and let yl_E.2 ^ \1 .6(1 .5) may be negative and is not necessarily an increasing function of y for jyj large. (4.. Heuristically. u5.: EA + zl_E Var(A) . A particular case is a.2X2 . of (4. however. If the distribution of Y is close to N(0.2K3 + 4i 64 + .3!). K4 . s. zl_e be the 1 .99. one expects the u3 term to dominate the terms of order u4. Rather than with the tail probabilities F(A > x). which is often denoted VaR (the Value at Risk).4. are the cumulants . THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter..5) is obtained by noting that by Fourier inversion. Thus if EY = 0. ylE should be close to zl_E (cf. . where Kl .6) .i 3 K3 } Pt^ exp . the density of Y is 1 °° _ eiuy f(u) du 2x _. and from this (4. the NP (normal power) approximation deals with the quantile al_E.y2)^P(y)• 319 Note as a further warning that the r. defined as the the solution of P(A < yle) = 1 ..e. the standard normal distribution.s.i 6 r 1 3 so that we should take b = ic3/6 in (4.5 (y3 . In concrete examples .h. one needs to show that 163.. as u2 u3 u4 9(u) = Ee'uY = exp {iuci ... Remark 4. . (4.5) follows by integration. and so as a first approximation we obtain a1_E = EA + yle Var(A) . K2 = Var (Y)..5). f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .
. this yields the NP approximation 6(Z1 _E .y2)cp( y) term. Using Y = (A . b such that EN 1 U%.1)! n ^eQ .1).EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.S(1 .5(1 . as required . the kth cumulant of A is /3PBk' and so s.E )Azl E) 4(z1E) + ( ylE . Another main reference is Daykin et at.EA ) / Var(A)..k = /3µB^1 / (.E .zlE)W(zlE) 1 .1)EY3.E)A1 l E) 1 E 4)(yl E) ^' .(y) terms dominate the S(1 .E(/3PB^1 )1^2 + s(z1E .zl E)^o(zl E) .3ni /3 . this holds with a = 0.. This leads to t( yl E) . K5 .zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 . that [101] distinguishes between the NP and Edgeworth approximations. [101]. In particular ..1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. MISCELLANEOUS TOPICS A correction term may be computed from (4.S(1 .6 (1 . b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n . k3 is small for large /3 but dominates 1c4.E = z1E + S(zi_E .320 CHAPTER XI.1) E (A . 4c Panjer 's recursion Consider A = constants a. n = 1.E + (yl.7) as 1 (3) a1E = Qµa +z1 . and assume that there exist n ) Pn_i ...zl E)V(zl_E) . 21 .zi. however. For example. We can rewrite (4. Note.yi.5) by noting that the 4. ..6pBki) d/2. let pn Pn = (a+ = P(N = n).
4 is that the algorithm is much faster than the naive method. .. .13) but only O(j2) for Proposition 4. fj = P(A = j). u Proof of Proposition 4. . . if go = 0. j = 1.11) Remark 4.4. 2. .4. the value of (4. j = 0..5 The crux of Proposition 4.12). 2. E[a +bU=I >Ui =j l i=1 J (4. Hence by (4. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4. By symmetry.. Since the sum over i is na + b. Then fo = >20 9onpn and fi = 1 E In particular.14) is independent of i = 1. .. fj = E (a+ b k =1 )9kfi_k .9)... . (4.10) f o = po.. . (4. n = k=n1 9k(n1 )9j k • (4.4 Assume that B is concentrated on {0. the complexity (number of arithmetic operations required) is O(j3) for (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 .14) is therefore a + b/n. and calculating the gj*n recursively by 9*1 = 9j. j = 1. 2. . j1 g. n. (4.12) where g*n is the nth convolution power of g.} and write gj = 2 ...k . The expression for fo is obvious.. (4.. then j (a + b!) 1ag k_1 3 gkfj. 1.13) Namely.1.. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4.4.
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).
h(Ui)P JJJ l:='l {ry ] or.d. we get N 1 = Eexp ry E[Ui ii . X2 = U . however. we get EX = EN • EU.i. Proof As in the proof of Proposition 6.b)+ with b determined by E(U . and so on. the excess ofloss rule hl (u) = (u .3). Then for any local retention function u .6 Assume the Ui are i.4.5) holds trivially.4).4). N E X .6). Remark 6.d.5 Because of the independence assumptions . y = Ei [Ui . (6. as often local as global.5) reduce quite a lot.4) and u g(x) = e7x in Ohlin's lemma.6) u where C[ry] = Ee'r(u4(u)). Eexp 7 [E [Ui .h * (X) . then (6.h(Ui)] . this implies 7* > 7.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. Local reinsurance with h(u) = (u .P.h(Ui)] . (6.h(U) (as in the proof of Proposition 6.h(Ui)] . u But since ry > 0.P > EexP{7[X .4). it suffices to show that Eexp {ry ii 'UiAb. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') . appealing to (6.h(U)]. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.h( UU) = EN • E[U .b)+ = Eh(U) (and the same P) satisfies 71 > ry..P } < 1 = Eexp E[Ui.P I = EC [7]N. i.h(u) and any P satisfying (6. Assuming for simplicity that the Ui are i. .6. ry* > 0 because of (6. expectations like those in (6. This follows by taking Xl = U A b. ' ii (6. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x . The arrangement used in practice is.P]}.
. See further Hesselager [194] and Dickson & Waters [120]. Bowers et at. [76]. Heilman [191] and Sundt [354].many texts on insurance mathematics.330 CHAPTER XI. see also Sundt [354]. The original reference for Ohlin's lemma is Ohlin [277].g. e. The present proof is from van Dawen [99]. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in.
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. t 00 (A.1) (here U(t) = U([0.. Y. the renewal process is called zerodelayed.r.U(t) is the expected number of renewals in (t. t]) so that U(t + a) . . Then Blackwell 's renewal theorem holds. The number max k : Tk_j < t of renewals in [0. Technically..T„_1). Lebesgue measure dt normalized by the mean to of F. i. the distribution of Yo is called the delay distribution. . are independent and Y1. not concentrated on {h..} for any h > 0.. Y2. The mathematical representation is either the ordered set 0 < To < T1 < ..t. denoted by F in the following and referred to as the interarrival distribution. some condition is needed: that F is nonlattice.. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process.. That is. Y2. If Yo = 0. of epochs or the set Y1. t] is denoted by Nt. 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. t +a]). .e. The point process is called a renewal process if Yo.. note in particular that U({0}) = 1.. then Stone 's decomposition holds : U = U... when t is large. of interarrival times and the time Yo = To of the first arrival (that is. stating that U(t+a)U (t) ^ a. Y1. all have the same distribution. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . . 2h. The renewal theorem asserts that U(dt) is close to dt/µ. = T„ . If F satisfies the stronger condition of being spreadout (F*' is nonsingular w .
Note in particular that F is spreadout if F has a density f.i". U Z(u . Under weak regularity conditions (see [APQJ Ch. then Z(u) i f0 z(x)dx .1 if F is nonlattice and z (u) is directly Riemann integrable (d. Then Z(u) 4 z(oo).EN(t) .2) has the unique solution Z = U * z. 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 . that z(u) has a limit z(oo) (say) as u 4 oo.R.5) 2This condition can be weakened considerably . oo).2) Z(u) = J0 u z(x)U(dx). the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. and F(dx) a known probability measure . see [APQ] Ch. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. Equivalently.x)F(dx). resp. (A. In 111. z(u) a known function. i. and that F has a bounded density2.9.4) If F is spread. z(x) = 0. IV). (A. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. µF (A. Both result are valid for delayed renewal processes. the statements being EN(t + a) .a.2).2 Assume that Z solves the renewal equation (A. then it suffices for (A. (A..4) that z is Lebesgue integrable with limZ.e. but suffices for the present purposes .i. in convolution notation Z = z + F * Z. IV). stating that U(t)/t > 1/p.332 APPENDIX u(t) has limit 1/µ as t 4 oo.3) Further.out. (A. u u PF 4 00.
this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. Note. a basic reason that renewal theory is relevant is the renewal equation II.. 0 PF µF 11 In risk theory. .(3. Z(u) U = 1 u 1 u f z(u ..5a. To this end.} be a renewal process.x)u(x) dx = z(u( 1 . . The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. Assuming that y can be chosen such that f °° Ox F(dx) = 1.r.e. refer to the zerodelayed case.. Tk (or. . T1. the postTk process {XT. • .2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x)... 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. .. z(x) = e7xz(x). that the existence of y may fail for heavytailed F. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). or many queueing processes. Hence by dominated convergence.. equivalently. The kth cycle is defined as {XTk+t}o<t<Yk . A stochastic process {Xt}t>0 with a general state space E is called regenerative w. cycles. {Tn} if for any k. A regenerative process converges in distribution under very mild conditions: . Yk ). Y1 .d. asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. This program has been carried out in III.3) satisfied by the ruin probability for the compound Poisson model. .i. We let FO.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). Here the relevant F does not have mass one (F is defective). The distribution F of Y1. however. we let µ denote its mean. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. i..k+t }t>o is independent of To. 1c Regenerative processes Let {T. multiply (A. Y2. the present more general definition is needed to deal with say Harris recurrent Markov chains. and its distribution does not depend on k. F(dx) = e7xF(dx). of Yo. that F is a probability measure. is called the cycle length distribution and as before. However. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. T1.t. Tk and {Xt }o<t<Tk • For example. The simplest case is when {Xt} has i.. Eo etc.
0<t<Yi then Zt /t a$• EU1/µ.i.6) id Cumulative processes Let {Tn} be a renewal process with i. fi (t) = inf {Tk . e(t )) . cycles (we allow a different distribution of the first cycle). {Tn}. This is the case considered in [APQ] V. If p = oo. and q(t) = sup It .v.i.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1.334 APPENDIX Proposition A1. Then {Zt}t^. oo). (A. in total variation.0 be cumulative w. under the condition of Blackwell's renewal theorem. {i7(t)} are Markov with state spaces (0.. assume that p < 00 and define Un = ZT}1 .. oo). An example is Zt = fo f (X8) ds where {Xt} is regenerative w.e.. then Xt .r. Then it (ii.. [0.r. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p.ZTOI < 00.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. where the distribution of X. We denote the limiting r. {Tn}. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt... are i.t. (b) If in addition Var(Ul ) < oo.0 is called cumulative w. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. but in fact. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t.. for n = 1. µ 0 If F is spreadout. resp .t.3. 2. Otherwise .ZT }0<t<Y„+. then e (t) . and we have: holds more generally that (rl(t). Then Xt Di X..+ X.'s by e. just the same proof as there carries over to show: Proposition A1. then (Zt .. {Tn} if the processes {ZT +t . r.Tk : t < Tk} as the age.e.t : t < Tk}..ZT Then: (a) If E sup I ZTo+t ..d. Then {e(t)}.. i.r.oo (i.t.4 Let {Zt}t^. C).d.
assume first the renewal process is zerodelayed.i..^(t))} as a regenerative process.APPENDIX 335 Theorem A1.4. W are independent. . Hence t t lt ) = f U(dy)z(t .t.d. and the conditional distribution of ri given l. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 .y) = f U(t . Proof The number Nt of renewal before t satisfies Nt/t a4' p. V is uniform on (0. the joint distribution of (rl. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). Hence for t large enough.U(x) < U( 1)). and the equivalence of (a) with (b)(d) is an easy exercise. if in addition EYo < oo.t. ^) is given by the following four equivalent statements: (a) P (77 > x. are not i. Y1 > t] 4 0. but governed by a Markov chain {Jn} (we . and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y). Since the maximum Mn of n i. For the second. Y1i Y2. ^ > y) = 1 f +Y (z)dz. EC(t)/t + 0. r. the first statement follows. In IV. (c) the marginal distribution of q is FO.d. Yo > 0] + f Eo^ (t . = z is Foz) The proof of (a) is straightforward by viewing {(r.y)P(Yo E dy) . we used: Proposition A1.U(x) (c < oo because it is easily seen that U(x + 1) . (1 V)W) where V.v..5 Under the condition of Blackwell's renewal theorem. the sum is o(t) so that Eo£(t)/t + 0 .i. In the general case. (b) the joint distribution of (ri. l:) is the same as the distribution of (VW.6 Consider a renewal process with µ < oo. (d) the marginal distribution of ^ is FO. U(x + 1) . Yl > t]. Then fi(t)/t a4' 0 and.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. use t E^(t)/t = E[Yo .. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . 1) and W has distribution Fw given by dFw/dF(x) = x/pF.(t). Since z ( k) < E[Yi .
jEE is a family of distributions on (0. the conditional distribution of {XT„+t}t>o given Yo. For example. . T_=inf{n>0: Sn<0}.g.t.T_ < oo). is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. Jn_1. J1 i . Jo..r.. .7 Consider a nonlattice semiregenerative process.. . Yn.. .. = io for some arbitrary but fixed reference state io E E. the semiregenerative process is called nonlattice if {T.i . We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions.}.. . . the Markov renewal process if for any n. and define r+=inf{n>0: Sn>0}. A Markov renewal process {Tn} contains an imbedded renewal process.. Notes and references Renewal theory and regenerative processes are treated. Sn = X1 + • • • + Xn the associated random walk. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo... be i. X2.and regenerative processes.) and (Fij )i. r+ < oo).. Jn +1=j} where J = a(JO. .. .. The semiregenerative process is then regenerative w. Jn = i is the same as the P.d. in [APQ]. These facts allow many definitions and results to be reduced to ordinary renewal. Further: Proposition A1. with common distribution F. oo). G+(x) = P(S.t... Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE.. .} is nonlattice (it is easily seen that this definition does not depend on i). Let X1. 0] or (0 . namely {Twk } where {Wk } is the sequence of instants w where Jo.+ < x. < yIJ) = Fij( y) on {Jn= i. e. Y1. IT. .r. Then Xt 4 Xo. oo). distribution ofjXt}t>o itself where Pi refers to the case Jo = i. where the distribution of X.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. A stochastic process {Xt}t>o is called semiregenerative w. G_(x) = P(ST_ < x. Alsmeyer [5] and Thorisson [372].
On {T_ > 2}. we consider the last such time (to make w unique) so that {w=m. G_. n=0 The basic identities are the following: Theorem A2.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. F(A) + (G+ * G_)(A). .7) (A. 0].=EGn. (A.7). oo) (A. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. u .1 (a) F = G+ + G_ ..T_=n} = {S. 0] and (0.S.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). A C (oo.7) follows since G+(A) = 0 when A C (oo. >0. 0<j<m. (d) R+ = U_. (c) G+(A) = f °. F(A . n 0 R_(A) = E I(Sn E A). A C (0. oo). .=n w=m i Figure A. the renewal measures U+=>G+. m<j<n}. (e) R_ = U+.x)R+(dx).r.g. Proof Considering the restrictions of measures to (oc. n=0 n=0 00 00 and the T+.x)R_ (dx). we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A).. S.1 . In (A.S.8) (e.>0. oo).G+ * G_: (b) G_ (A) = f °° F(A . A C (0. define w as the time where the preT_ path S1. G+.. 0]). Sr_ _1 is at its minimum . More rigorously. U. A C (oo. 0).
r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. S. 0<j<m.+ E du)P(S.XnEAx) 00 f 0 f 0 00 00 1: F(A .... . ST+Edu). S.>0.x)R+(dx). It follows that for n > 2 F (7. and the proof of (A.m.0<k<ri . Sr_ E Adu) (s ee again Fig .1).7) follows. A. SmEdu) = P(T+=m. . A.u) f0m m=1 n=m+1 00 J0 OO P(S.1) that P(Sj Sn. 0 < k < n.3. E du) = P(T_=nm. and reversing the order of summation yields P(T_ > 2.+ E du) E P(S. Sn1 E dx) n=1  F(A . (A. ST_ E A) P(T+ = m. m=1 f S mming over n = 2.du) (G+ * G)(A)• C llecting terms. (b) follows from 00 G+ (A) _ E F(Sn E A.._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu).Sn_1Edx. m it follows (see Fig. ST_ E A ..1.= n.8) is similar. .F(r_n_mSrEA_u)..x)P(Sk < 0.3 8 APPENDIX Reversing the time points 0. clearly (Sj Sm>0. Aso._ E A ... SnEAIS. m < j <n._ = n .
For example. The present proof of Theorem A2.g. G_ are trivial.s. a number of related identities can be derived. being concentrated at 0. which is basic for the PollaczeckKhinchine formula.SnEA) is the probability that n is a weak descending ladder point with Sn E A.SnEA) = P(Sn<Sk. is based upon representing G+ as in (b). such developments motivate the approach in Chapter VI on the Markovian environment model.T+> n) = P(Sk < O. see e. see for example Bingham [65]. the survey [15] by the author and the extensive list of references there. However. Nevertheless. and the proof of (e) is similar. G_ [s] are defined at the same time. u Remark A2.. 0]. Then for A C (oo. we can rewrite (a) as 1 . E.SnEA) = P(SnSn_ k.SnEA) = P(Sn<Sk. 6+ [s].1. In this generality of. P(SnEA .F[s] = (1 . and G+.O<k<n. Again. Another main extension of the theory deals with Markov dependence.APPENDIX 339 and the proof of (c) is similar. Summing over n yields R+ (A) = U_ (A).0<k<n.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}.0<k<n.4).9) whenever F[s]. and sometimes in a larger strip.2 In terms of m.G_[s]) (A. there are direct analogues of Theorem A2.O<k<n. oo). it serves as model and motivation for a number of results and arguments in continuous time.1). u Notes and references In its above discrete time version.'s. cf.0+[s])(1 . H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1).. For (d). The classical analytical form of the WienerHopf problem is to write 1 . 11.6. and similarly H_ (s) = 1 . In discrete time. In continuous time.1(a) is from Kennedy [228]. consider a fixed n and let Xk = Xn_k+l. then T+ = inf It > 0 : St = 0} is 0 a. the derivation of the form of G+ for the compound Poisson model (Theorem 11. the analogue of a random walk is a process with stationary independent increments (a Levy process. there is no direct analogue of Theorem A2.g. Since G+ is concentrated on (0.1.P as a product H+H_ of functions with such properties. this holds always on the line its = 0. . and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). Sk = X1 + • • • + Xk = Sn .Sn_k. if {St} is Brownian motion.g. WienerHopf theory is only used at a few places in this book.f.
write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step). Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . _I 0 (A. Eo Kn/n! converges rapidly and can be evaluated without p oblems.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. whereas there is no similar single established a proach in the case of matrix exponentials.13) henever A is a diagonal matrix with all diagonal elements nonzero. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. if m is s fficiently large. and eQ can then be computed as the mth power (by squaring if = 2). Here are. 1. 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).11) A f eAtdt = eA. 0 . one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). Thus. To circumvent this. ere A is the eigenvalue of largest absolute value.10) d dteAt = AeAt = eAtA (A.12) eA'AO = Ale AA (A.5 that when handling phase type distributi ons.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. It is seen from Theorem VIII. however . three of the c rrently most widely used ones: xample A3. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum).
To this end.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt..]t)n (A. 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.. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. Ap. i.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (. i..15) Then it is easily checked that P is a transition matrix .3 i (A. condition upon the number n of Poisson events in [Olt])  Example A3.. some jumps are dummy in the sense that no state transition occurs ). Here is a further method which appears quite appealing at a first sight: Example A3 . One then can reduce to p linear differential equations by noting that k = ZQ. p different eigenvalues Aj i .. vp be the corresponding left . assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. Zo = h).APPENDIX 341 Example A3. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. The probabilistic reason that (A. Zo = a (Z = QZ. the procedure consists in choosing some suitable i > 0. Let vi. 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 .e. However .7t) n=0 n! u °O n Pn (to see this. The approach is in particular convenient if one wants eQt for many different u values of t.e.4 (DIAGONALIZATION) Assume that Q has diagonal form. . In practice. .2 (UNIFORMIZATION) Formally. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector.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}. 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.
and writing eQt as eQt = He°tH1 = H (e\it)di. i # j. not all ai are real. some cases remain where diagonalization may still be appealing. vi.18) Namely. say A = (Ai)diag.. Then P P Q = > Aihivi = E Aihi (9 vi. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. hp the corresponding right (column) eigenvectors.17) eQt = E e\`thivi = E ea:thi ® vi. and hence A2 is so because of A2 = tr(Q)... the eigenvalue.. Example A3. i= 1 i=1 P P (A. we have an explicit formula for eQt once the A j. The phenomenon occurs not least when the dimension p is large. under the conditions of the PerronFrobenius theorem). two serious drawbacks of this approach: u Numerical instability : If the A5 are too close.. Qhi = vihi. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts... v5Q = Aivi. Complex calculus : Typically. we can take H as the matrix with columns hl. say Al.342 APPENDIX (row) eigenvectors and hl.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. (A. however.18) contains terms which almost cancel and the loss of digits may be disasterous.g H1. D = ) 2 2 . hp. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix.16) (A. of largest real part is often real (say. Nevertheless. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. (A. and vihi ¢ 0. There are. i=1 i=1 Thus.. Then vihj = 0. hi have been computed. and we may adapt some normalization convention ensuring vihi = 1.
6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. v2 and h2 can be computed in just the same way. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A. k  C k2 ) =b ( A1 q 1 Q11 / where a . l ab (g12g21 + (A1  411) 2) = 1.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. 1) .e.q. However. The other eigenvalue is A = A2 = q1 .20) ir = q2 ql qi +q 2 9l +q2 (A. replacing ai by A2. where (A. Of course. Then 7r = (ir1 7r2 ) = a (q21 Al . h2 = Thus.19) Example A3 .Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. i.k1). u Example A3. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . b are any constants ensuring//Irk = 1.7 Let 3 9 2 14 7 11 2 2 .
11/2 .11/2 + 5 1.22) Note that in this generality it is not assumed that A is necessarily square.satisfying AAA = A. (A+A)' = A+A. 2 2 1=ab(142+(1+2)2 ) = tab. Generalized inverses play an important role in statistics.. for example AA+A = A.6. A2 = 3/2 . (A.344 Then D= 2+ 11)' 7 T4 2 =52. (AA+)' = AA+..23) .5 . A+AA+ = A+. (A. and a generalized inverse may not unique. APPENDIX x1 3/2 . They are most often constructed by imposing some additional properties . e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix 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_. but only that dimensions match .
P).25) .I) (A.1 Let A be an irreducible intensity matrix with stationary row vector it. .. and define D = (A .. and exists and is unique (see for example Rao [300]).1Q = Q(Q . Here is a typical result on the role of such matrices in applied probability: Proposition A4..eir ). E.= (I . one is also faced with singular matrices . . and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' . Rather than with generalized inverses . Am > 0. _ A.eir )1. (A. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A .e.ew. one then works with Q = (Q .eir)1 = I . 0 01 In applied probability. (I . most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. = 0 where m < p is the rank of A.D + O(ebt)..P + e7r)1 (here ( I . Then for some b > 0.. lt o eAx dx = te7r + D(eAt .g. Assume that a unique stationary distribution w exists .24) = te7r .g.1 goes under the name fundamental matrix of the Markov chain).P + e7r ). are ordered such that Al > 0.. if A is a possibly singular covariance matrix (nonnegative definite).APPENDIX 345 A matrix A+ satisfying (A. These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . Am+1 = . ( Q .e ® 7r)1.23) is called the MoorePenrose inverse of A.
Equivalently. h ® it reduces to hit in standard matrix notation. Then A(O) _ B(O) = 0.s. the r. I. . .h. 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.eir)eAt = eAt = A'(t). o Finally.J {xe^r + D(e .e.26) 2 = 2 e7r + tD . B'(t) = e7r + DAeAt = eir + (I . of (A.. the rows are proportional to it.2 Let it be a row vector with m components and h a column vector with k components. For example. B(t) denote the l. the formulas involving O(e6t) follow by PerronFrobenius theory.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4. and in fact any rank 1 matrix can be written on this form. resp.h. h as 1 x m and k x 1 matrices.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .DZ(ent .s. (A. (A.I) .27) Proof Let A(t).D + D2 + O(ebt).2e7r .346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. 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. it follows that h ® it is the k x m matrix with ijth element hi7rj . and the columns to h. see below.24).I)}.I)} dx.91a(2) . Note that h ® it has rank 1.I) (A. Interpreting 7r. respectively.
31) Indeed. 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.3f 4v/.29) If A and B are both square (k1 = ml and k2 = m2). C2 = h2 are column vectors.4 eA® B = eA ®eB.3vV/72f 20. such a factor is Ak (&B 1k according to (A.A9. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) .(A.3V8.APPENDIX 347 Then A®B = 2 f 20.31). (A B)' = eA®B e! L 1=0 0 . then v1B1h1 and v2B2h2 are real numbers. each of which is A ® I or I ® B. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I .28) In particular.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^. Using (A.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.29).3v'6.4vf. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).5v/. A2 = v2 are row vectors and C1 = h1. and the number of such factors is precisely given by the relevant binomial coefficient.50 6 7 6 4f 4. if Al = vi. if A ® I occurs k times. (A. (A.
I)(h ® k). P(t) Yt(2) }. in the definition (A. and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A.6 Suppose that A and of B. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } .348 APPENDIX Remark A4. Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. the same time. {Yt(1). { On the other hand. P(2). the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . p = P(1) ® {X }. { 1't(1) }. where transition matrix of the bivariate Markov chain {X n1). independent Markov chains. we have P8 = Pal) ® p(2).4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines .3 < 0 Lemma A4 . Q(2). A special case of Proposition A4. Let P8f P(Sl). 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 +. From what has been said about matrices of {Yt( 1). Let further it. (A. 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 . h.33) . resp . Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }.s.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). X ) }. Yt(2 ) }.32). represents ces Q( 1). n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). k any column vectors.5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. Thus . Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .
and the corresponding left and right eigenvectors v.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4.34) Note that for a transition matrix. Similarly. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. h = e and v = 7r (the stationary row vector). f o r each i. which can be found in a great number of books.. j = 1. and appeal to (A. .APPENDIX 349 Proof According to (A. = j and atk_li. so that by asssumption A ® B is u invertible.29). ao). [APQ] X. > 0 for k = 1. Here is the PerronFrobenius theorem.g. E (0.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). . we have AO = 1. .. . . p there should exist io. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. Then the eigenvalue Ao with largest real part is simple and real. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. (A. . . and the corresponding left and right eigenvectors v. n.7 Let A be a p x pmatrix with nonnegative elements. h can be chosen with 3By this.. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. That is.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. h such that vh = 1..3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. in such that io = i... see e. and if we normalize v.. h can be chosen with strictly positive elements. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. (b) if in addition A is aperiodic.12). il. Now note that the eigenvalues of A ® B are of the form a +. then IN < Ao for all A E sp(A). . i. .
Note that for an intensity matrix.1. we have A0 = 0. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. Example A3. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). 10) and use the formula me at e Bt = e 00 Antn = e . not only in the tail but in the whole distribution.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). h = e and v = 7r (the stationary row vector). the analogy of this procedure with unformization. let {Yti°i } be a Markov process with initial distribution a and intensity . h such that vh = 1.(ti)ding.35) for some p E (oo. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo.. 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). A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. relate the eigenvalues of B to those of B via (A. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. Then for any (3. but is an easy consequence of the PerronFrobenius theorem. note that we can write the phase generator T as Q .8. if we normalize v. To this end. Proposition A5. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ .2). I. The next result gives a condition for asymptotical exponentiality. Ao).(3. For example.350 APPENDIX strictly positive elements.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. Furthermore. the condition is that t is small compared to Q.8 is most often not stated explicitly in textbooks. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. Corollary A4.e.n t AL n=0 n! (cf. the phasetype distribution B(a) with representation (.
= YQ(x). from which the phase process is terminated .9. and that Yt(a) = Yat for all t.YQ(av) = j) Pi ( ci(a'V) > x. In addition to the asymptotic exponentiality. Let further V be exponential with intensity V and independent of everything else.aE where 0 < e < 1). and write Yt = Yt(1). Proof Assume first ti > 0 for all i and let I. By the law of large numbers for Markov processes .jEE.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. fo tY dv/t a$' t*.APPENDIX 351 ((1) etc. 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}. Hence O ((a) aa. J^O)_ = j) Pi (v(aaV) > x. prove a somewhat more general result which was used in the proof of Proposition VI. Conditioning upon whether { Yt} changes state in [0. dx/ti] or not. Then a(a'V)/a (aV) a' 1. has a limit distribution: Proposition A5.bij) Hence the intensity matrix of { Ix} is (qij/ti)i. a' = a . in fact . Then {Ix} is a Markov process with to = Yo.1. a'/a + 1. it states that the state. We shall .2 Pi (c(a) > x. a . t < (a). J(()) _ = i) + at•x t tt' .Yj(av) = j f .a' + oo (e. 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(°). Since JJ(. and this easily yields a(x)/x a' 1/t*.(a) > x . {t Y( a) } v>0 . 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. v/ t.x (1 .g.
g. A distribution B on {1. these results are in the spirit of rare events theory for regenerative processes (e.5 Let B be discrete phasetype with representation (P. k>1. . P. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. Example A5. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. (b) the generating function b[z] _ E' . 2..352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .} is said to be discrete phasetype with representation (E. ' pk 0 k>1 11 Theorem A5.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. . Penev & Turbin [238].j) and initial distribution a. Then P is substochastic and the vector of exit probabilities is p = e . (c) the nth moment k 1 k"bkis 1)"n!aP"p. However.x k > K.. = 0 for one or more i. a = b = (bk)k=1.. so we shall be brief. k = 1.2 do not appear to be in the literature. let E and Pkj j=k1. Example A5. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).Pe. See also Korolyuk... with point probabilities bk = (1 . .p)k1 p.4 Any discrete distribution B with finite support. is discrete phasetype.3 As the exponential distribution is the simplest continuous phasetype distribution. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. K}. Then: (a) The point probabilities are bk = aPklp. a). u Notes and references Propositions A5. an easy modification of the argument yields finally the result for the case where t... Et II I a(a^V) > x) at' . 2.zP)'p. Keilson [223].1 and A5. zkbk is za(I . so is the geometric distribution. Indeed. > 0}... the simplest discrete phasetype distribution: here E has only one element. 1 k=1 1 0 otherwise... say bk = 0.
{ 0.T(1)). r .a(2). The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r.a(1). B2 be phasetype with representations (E(1). T= ( 0 T(2) ) (A. A. as is seen by minor modifications of Example A5. Then the convolution B = B1 * B2 is phasetype with representation (E. and hence the negative binomial distribution is discrete phaseu type. initial distribution a and phase generator T. resp. . 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.6.. _ i E E(1) T(1) t(1)a(2) i E E(2) .. a' .6 (CONVOLUTIONS) Let B1. 11 Example A5. U2.1 This corresponds to a convolution of r geometric distributions with the same parameter p.. { Jt 2) } with lifetimes U1 . and a=1). (E(2)..2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }.T(2)). Jt t > U1 + U2. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.36) in blockpartitioned notation (where we could also write a as (a (1) 0)).7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5.APPENDIX 353 5c Closure properties Example A5. Then {Jt} has lifetime U1 + U2 .r + 1. a. resp.
T(2)).T(1)).0)ai2). we need to restart the phase process for B w. (E(2). a.0)a(2) E(2) Figure A. Then the mixture B = 9B1 + (1 . then C is the distribution of Ul + • • • + UN.T. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E.p.4 .E) where a(°) = fAa(a)v(da). with common distribution and N is independent of the Uk and geometrically distributed with parameter p. B2 be phasetype with representations (E(1). Then it is trivial to see that B(") is u phasetype with representation (a(").O)B2 (0 < 0 < 1) is phasetype with representation (E. resp. Example A5. Example A5. A reduced phase diagram is 0a(1) E(1) A .a(2). a mixture of more than two phasetype distributions is seen to be phasetype.d. In risk theory.354 APPENDIX Example A5.37) (1) (1 . U2. Equivalently..0)a(2))). P(N = n) = (1 . a.p)pn1B*n. To obtain a phase process for C..3 In exactly the same way.p)pn1. Thus. i E E(2) 0 T(2) =IT (in blockpartitioned notation. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims. and consider B(") = fA B(a) v(da) where v is a probability measure on A.a(1). i E E(1) T 0 I (A. if U1..i. this means that a = (Oa(1) (1 .').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. are i. and o'i Oa. T) and C = EO°_1(1 . Let B(") be the corresponding phasetype distribution. a reduced phase diagram is f a E t Figure A. p at each termination.8 (FINITE MIXTURES) Let B1.
let {Jtl)}. +UN 2. then U1 + • • + UN is zeromodified phasetype with representation (a. i. U2. a(2).aF[T]. If U1 has a different initial vector. T).v.1. j E F}. a. X independent of U. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). 13 (MINIMA AND MAXIMA ) Let U1.a(1). T + ta.X)+ is zeromodified phasetype with representation (E. For U1 A U2.. Equivalently. Example A5..x)+.. . T + pta).d.. Indeed.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. then C is the distribution of U1 + • • • + UN. T(1) ® T(2)).T + pta).°_1 f„ B*?l. To see this.7.9) that (U . { Jt2) } be independent with lifetimes U1.2. resp.. . 12 (PHASETYPE COMPOUNDS ) Let fl. v. T(2) ). be the point probabilities of a discrete phasetype distribution with representation (E. a. but the same T. Minor modifications of the argument show that 1..TWWW).. cf. of F.f. Example A5 .T) where F[T] = J0 "o eTx F(dx) u is the matrix m.2. say with distribution F. U2. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype.°. f2. resp. are i. U2 be random variables with distributions B1. it follows by mixing (Example A5.aeTx. Note that this was exactly the structure of the lifetime of a terminating renewal u process. if U1. v.. B2 of phasetype with representations (E('). say v. { 4 } as exit of {Jt}. Thus the representation is (E(1) x E(2). if B is defective and N + 1 is the first n with U„ = oo. It is zeromodified phasetype with representation (E. cf. Example A5 . (E(2). let the phase space be E x F = {i j : i E E. P). with common distribution B and N is independent of the Uk with P(N = n) = f. Corollary VIII. then Jy has distribution aeTx.g. Proposition VIII. a(1) ® a(2 ). E). T) and C = F. a.T) if U is phasetype with representation (E. To obtain a phase representation for C . let B be a continuous phasetype distribution with representation (F. If we replace x by a r. then U1 +• is phasetype with representation (E. if {Jt} is a phase process for U.APPENDIX 355 and C is phasetype with representation (E.
(n) = D.(bk) + B(bk) for all k.. we can assume that ID. Thus the state space is E(1 ) x E(2) U E(1) U E( 2)....B(bk) I < 1/n for n > k.n = I:pi(n)Er v ( __ ) n) ) a= 1 . Now we can find first a sequence {Dm} of distributions with finite support such that D. Let the support of Dn be {xl(n). elementary) Let {bk} be any dense sequence of continuity points for B(x).(bk)'. any distribution B on (0. i= 1 C. relies more on matrix algebra than the probabilistic interpretation exploited here). 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 . there is a sequence {B. oo). That is. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5.(bk) + B(bk) for all k as n * oo.(Sn) with Sn = n/b. with weight pi(n) for xi(n). Then we must find phasetype distributions Bn with B.. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . By the diagonal argument (subsequent thinnings). and let Bn be the Erlang distribution E. say degenerate at b. the initial vector is (a(1) (& a (2) 0 0).. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. cf.} of phasetype distributions such that Bn 3 B as n + oo.xq(n)(n)}. The general case now follows easily from this.2) } to go on (on E(2)) when { i 1) } exits. Then from above.. see Neuts [269] (where the proof. Here are the details at two somewhat different levels of abstraction: (diagonal argument . and the closedness of the class of phasetype distributions under the formation of finite mixtures. Example A5. q(n) q(n) pi(n)a .8. Proof Assume first that B is a onepoint distribution. however. we need to allow { Jt. Hence it is immediate that Bn 4 B. and vice versa..14 To a given distribution B on (0.356 APPENDIX For U1 V U2. r # oo.
oo). oo) approximation Assume that we can compute a functional W(B) when B is phasetype. f2. oo) * [0. k < n. But To is the class G of all distributions on [0.d. i. we can then approximate Bo by a phasetype B. Then ICr( n )..n. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large.(x)Bf.i. one would use the B given by some statistical fitting procedure (see below). x 4 oo. u 2 (abstract topological ) The essence of the argument above is that the closure (w. k < n. and that cp is known to be continuous.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. say on the claim size distribution B in risk theory..B(bk )I < . 2. then it is immediate that WI(B) = p2(B) for all distributions B on [0. for some a < oo.15 To a given distribution B on (0 .14 is fundamental and can motivate phasetype assumptions. the class CO of all discrete distributions.. oo) and any fl. For a general Bo..t. i = 1. If Cpl (B) and ^02(B) are weakly continuous.. Hence G C PET and L = PIT. Let E be the class of functions f : [0. Corollary A5... PIT contains all finite mixtures of onepoint distributions. In particular.n (bk) . if information on Bo is given in terms of observations (i. however. .( dx) * f r f{(x)B(dx). E E. and we can take Bn = Cr(n).n( b k ) . replications).. 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. oo) such that f (x) = O(e«x). It should be noted. compute W(B) and use this quantity as an approximation to cp(B0).r. u Theorem A5. Since PET is closed under the continuous operation of formation of finite mixtures.D(bk)I < n.e.
358 Proof By Fatou' s lemma. for each i.. i=1. and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ). n B=az.. Now returning to the proof of (A. By (A. liminf B.39). 2....f (z) = f = 1 1 1 1n/ o . .38 ).n(dx) < 1+. i = 1..oo J fi(x)B. i = 1.. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge..38) We first show that for each f E E. . TO (A.. .(dx) > J fi(x)B(dx). oo). n. i = 1. and the case of a general f then follows from the definition of the class E and a uniform integrability argument.. n.f (x)B(dx).n(dx) + f 0 fi(x)Dn(dx). f00 fi(x)Cr...39) Indeed.2 .f ' f (x)B(dx).  APPENDIX B implies that 00 o o 00 n. if f (x ) = e°x.16 To a given distribution B on (0 .. . we may assume that in the proof of Theorem A5.14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx). \\ 0 Corollary A5. and hence we may choose r(n) such that L 9l) f (x)Cr(n). then cc f (x)Bn ( dx) = (?!c ) e'= . f° xtBn(dx ) * f °° x`B( dx). Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. .
(N. For practical purposes. e ) and ei J. from a more conceptual . the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. the remaining results may be slightly stronger than those given in the literature./3) is defined as the unique solution > 0 of B[y] = l+y/j3.l3µb < 1. there is a sequence {B.. oo) with B[y +e] < oo for some e > y = 7(B. However. . lim inf > is proved similarly. . . . O We state without proof the following result: Corollary A5.17 To a given /3 > 0 and a given distribution B on (0. Notes and references Theorem A5. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions./3). This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. I. . one can obtain 7(Bn.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . The present section is a survey of some of the available approaches and software for inplementing this. but are certainly not unexpected.3). there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. lim sup ryn < 7. 0 as i * oo.14 is classical.. If ei > 0. the adjustment coefficient 'y = 7(B. The adjustment coefficient is a fundamental quantity. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i ..18 In the setting of Corollary A5.> y for some sequence {ei} with ei E (0. Proof Let fi(x) = el'r+E. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . . 5e Phasetype fitting As has been mentioned a number of times already. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. .16.e. /3) = ry for all n. (N or a given distribution Bo. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn.
the L1 distance between the c . and as fitted distribution we may take B. defined by the absence of loops in the phase diagram . g. we do not not want to perform matrix calculus in hundreds or thousands dimensions). The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . . Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . [70]) restrict attention to acyclic phase type distributions . The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter.g.. for some suitable large n.. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . we have constructed a sequence { B. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). risk theory. reliability or queueing theory. 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.} of phasetype distribution such that Bo. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood.d.f.g. and used a nonlinear programming approach . [317] ) has considered an extension of this setup. [202]..f. Of course. and in practice this sets a limitation to the usefulness (the curse of dimensionality . It seems therefore a key issue to develop methods allowing for a more general phase diagram. at a a number of selected points . d. In a series of papers (e. a program package written in C for the SUN workstation or the PC is available as shareware. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . The characteristics of all of these methods is that even the number of parameters may be low (e.'s).g. A method developed by Bobbio and coworkers (see e. Schmickler (the MEDA package. B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly. [216] ). the number of phases required for a good fit will typically be much larger. 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.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . (N is the empirical distribution Be.g . giving mass 1 /N to each S=. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. e . . three for a mixture of two Erlangs ).. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. cf.
x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations.g. . one is lead to an iterative scheme. Nii = = . e.. E.. (n+1) _ Ea (n). (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. ..g.. since this is parameterdependent. Thus.T(n) (Nik IC1.T (n)(TiI(1. . N Ti = I(J= i) dt.. (N) tJk Ea ( n). eieT(n)((k. it seems open whether the restriction to the acyclic case is a severe loss of generality. . The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations.. In practice. it is easy to see that N (k Ea(n). the methods of [70] and [38] appear to produce almost identical results.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j).. = j) f k=1 k =1 tE[0. EN where ai = N 1 I ((k) = i) tii=i iEE. then the estimators would be of simple occurenceexposure type. In fact..(N) = E Ea(n).. jEEA. .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) (Ti ^^ 1.APPENDIX 361 tion Sk was available.
This page is intentionally left blank .
B. Aktuar Tidskr. Aktuar Tidskr. ed. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Borokov. G. 143170.A. Abate & W.G. Abate & W. [5] G. Asmussen (1987) Applied Probability and Queues. Dover. [9] G. Appl. ibid. 3157 . pp.). Chichester New York. with applications to risk reserve processes and the GI /G/1 queue. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. [8] G. [2] J. 57. Abramowitz & I. Asmussen (1985) Conjugate processes and the simulation of ruin problems. [6] V. 253267. Transactions XVth International Congress of Actuaries. Optimizations Software. Scand. New York. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. 38. [3] J. Proc. J. 213229. [14] S. Queueing Systems 5 . 311338. 345368. The Mathematical Scientist 14. [11) S. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions .). Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . Skand. Teubner. 1984 . New York. Preprint. Alsmeyer (1991) Erneuerungstheorie. 14. 191223. 1985. [7] E. [4] M. The case of equal risk sums. [12] S. Queueing Systems 10. AT&T. Choudhury & W.Bibliography [1] J. In: Limit Theorems and Related Problems (A. Stuttgart. Appl. Probab. Stoch. [15] S. Asmussen (1984) Approximations for the probability of ruin within finite time . New York. Act. Arfwedson ( 1954) Research in collective risk theory. 37. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . 219229. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Skand. [10] K. Adv.L. 587. 53100. [13] S. Arfwedson ( 1955) Research in collective risk theory. 101116. Abate. John Wiley & Sons. Queueing Systems 16. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. II. 363 . 20.
[32] S. [33] S.M. Astin Bulletin 27. [31] S. Scand. Asmussen & B. Appl. 69100. stationary distributions and first passage probabilities. 303322. Probabilistic Analysis of Rare Events (V. [23] S. Act. 37. 1936. Asmussen. [28] S. [22] S. 20. Probab. Rolski & V. Appl. New York. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. [18] S.S. 354374. Extremes 1 . 2. [30] S.364 BIBLIOGRAPHY [16] S. 1999 . Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. Stoch. Stochastic Models 11. J. Advances in Queueing: Models. Probab. Act. Statist. Alfa & S. [27] S. J. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Proc.). 297318. Chakravarty. 193226. Kalashnikov & A.). [25] S. eds. K. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. SIAM Review 40. 2149. Asmussen & B.). Appl. Asmussen. 1989 . Binswanger & B. Hojgaard (2000) Rare events simulation for heavytailed distributions. Asmussen. 54. Asmussen (1992b) Light traffic equivalence in single server queues. 555574. Floe Henriksen & C. 2943. 25. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. Frey. Asmussen & M. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Asmussen (2000) Matrixanalytic models and their analysis. eds. Proc.K. Random Processes 2. Stoch. 189201. 313341. 106119. MatrixAnalytic Methods in Stochastic Models (A. Bernoulli 6. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. Asmussen (1999) On the ruin problem for some adapted premium rules. Asmussen & M. J. 96107. Asmussen & K. Asmussen (1998c) A probabilistic look at the WienerHopf equation. 4966. 137168. [24] S. Asmussen (1992a) Phasetype representations in random walk and queueing problems. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Scand. CRC Press. Ann. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. 313326. Marcel Dekker. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Scand. Scand. [21] S. Asmussen (1995) Stationary distributions via first passage times. Florida. 27. [20] S. Ann. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. [29] S. 315. . 8. [26] S. [19] S. Act. 1996. Probab. [17] S. Riga Aviation University. Appl. L. Andronov. Th. 772789. 79102. Boca Raton. T. A. Ann. Methods V Problems (J. J. [34] S. Dshalalow ed. Asmussen (1989b) Risk theory in a Markovian environment.
419441. Asmussen & C.BIBLIOGRAPHY 365 [35] S. Probab. Adv. Prob 32 . Insurance: Mathematics and Economics 10. Asmussen & S. Wiley. first passage problems and extreme value theory for queues. Koole (1993). Sc. Finance and Stochastics 4. Read eds. Rubinstein (1999) Sensitivity analysis of insurance risk models. [43] S. Johnson. 115. 103125. 259274. Appl. J. J. Res. Asmussen & C. 365372. Taksar (1997) Controlled diffusion models for optimal dividend payout. O. [44] S. Proc. Stoch. [45] S. Eng.M. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Asmussen & G. Boca Raton. Asmussen & R. 58. Asmussen. Asmussen & D. Oper. CRC Press. 913916. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. Schmidt (1995) Ladder height distributions with marks. Supplementary Volume (Kotz. Appl. Methods 8 Problems (J. Asmussen & C. Asmussen. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Asmussen & T. Florida. [42] S. B. 20. Dshalalow ed. Appl. Statist. 410433. 421458. 23. Insurance: Mathematics and Economics 20. Example of excessoffloss reinsurance for an insurance corporation. [49] S.Y. [52] S. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. 422447. Asmussen & T. [40] S. [38] S. . Advances in Queueing: Models. with applications to insurance risk Stoch. Management Science 45. Opns. 30. Stochastic Models 8. 736755. [48] S. 31.). 11251141. [50] S. Inf. [41] S. Appl. [46] S. Asmussen & V. 19. 64. [51] S. Th. [39] S. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. Probab. [37] S. Asmussen & R. 429466. 10.A. Hojgaard & M. Asmussen & M. Olsson (1996). 299324. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Asmussen . Res. [36] S. H. Adv. Statistica Neerlandica 47. Kliippelberg (1996) Large deviations results for subexponential tails.). Fitting phasetype distributions via the EM algorithm.A. (to appear). 120. Asmussen & H. Probab. J.Y. Asmussen & K. Math. Nerman & M. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Proc. Asmussen & V. Taksar (2000) Optimal risk control and dividend distribution policies . Probab. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. Appl. [47] S. Sigman (1996) Monotone stochastic recursions and their duals. Appl. Scand. Schmidli & V. Perry (1992) On cycle maxima. Marked point processes as limits of Markovian arrival streams. 105119.
T. J. J. 77111. Bjork & J. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. Scand.A. Probab. Boogaert & A. Insurance: Mathematics and Economics 6. [58] O. Cambridge University Press. Insurance: Mathematics and Economics 4. New York. Soc. Bloomfield & D. [71] P.A. New York. J. Scand. 1974.B . [67] T. . BarndorffNielsen & H. [66] N. [59] O. Bingham. 275279. Scand. [61] J. 169186. Bjork & J. Scand. Grandell (1985) An insensitivity property of the ruin probability. Asmussen & J. 190204. Wiley. 705766. [73] A. [57] C. [55] B.J.L. Baker (1977) The Numerical Solution of Integral Equations. Stochastic Models 10. Probab. New York. Teugels (1987) Regular Variation. 33. [54] K. Appl. Wiley. [63] A. Act. Cox (1972) A low traffic approximation for queues. 4148. 1988 . de Waegenaere (1990) Simulation of ruin probabilities. 661667. [70] A. Actuaries 21. Probab. J.R. Berlin. Cambridge. Borovkov (1976) Asymptotic Methods in Queueing Theory. 11811190. [68] T. 1975. Act. [72] P. [69] P. Beekman (1985) A series for infinite time ruin probabilities. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. Goldie & J. Halsted Press. Trans. Act. 9599. 1995. [56] B. Beekman (1969) A ruin function approximation. Billingsley (1968) Convergence of Probability Measures. Springer . 129134. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH.H. Act. Bobbio & M. Act. Clarendon Press. Ney ( 1972) Branching Processes . Adv.L. J.M. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. Athreya & P. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Appl. J. 221232. [60] J. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. Insurance: Mathematics and Economics 9. Appl. Bingham (1975) Fluctuation theory in continuous time. Berman & R. Scand. [64] P.H. J. 7. SIAM. Boogaert & V. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. [62] J. SpringerVerlag. C. 9. 148156. Beekman (1974) Two Stochastic Processes. 610. [65] N. 1985 . 832840.366 BIBLIOGRAPHY [53] S. Oxford. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions.H.
Academic Press. Cox (1955) Use of complex probabilities in the theory of stochastic processes. H. Queueing Systems 33. Goovaerts & F. [84] D. AddisonWesley. Probab. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Breiman ( 1968) Probability. II. [86] K.BIBLIOGRAPHY 367 [74] O. 51. [91] H. [82] H.U. Insurance : Mathematics and Economics 5. 24. Nonparametric estimators for the probability of ruin.L. SpringerVerlag. [80] F. Bux & U. J. Chandy & M. Proc. Adv. [92] H. Gerber. Bell.). 177204.Y. 2338. Appl. [81] J. Chung (1974) A Course in Probability Theory (2nd ed. 127130. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals .L. [77] P. Fort & G. Herzog (1977) The phase concept : approximations of measured data and performance analysis. Nesbitt (1986) Actuarial Mathematics . Berlin. Croux & N. Tweedie (1982) Storage processes with general release rule and additive inputs .R. J. Springer. Schrage (1987) A Guide to Simulation.C.L. Cambr. 907920.C.L. M.. Soc. 14. Tech. Veraverbeke (1990). Reading. Insurance : Mathematics and Economics 9.M. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. Aut. [90] D. Illinois. Philos. 16. [89] M.). [75] O. 749763. 3539. Cramer (1955) Collective risk theory. The Society of Actuaries . Res. Z. Skandia Jubilee Volume. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Boxma & J. Simulation and Estimation. IEEE Trans. Fox & L. D. Itasca. 93121. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Computer Performance (K. Stockholm. Hickman. New York.R. [83] D.A. Sel. Jr. NorthHolland. Bucklew ( 1990) Large Deviation Techniques in Decision. Qinlar (1972) Markov additive processes. De Vylder (1986) Ordering of risks and ruin probabilities. [76] N. Cottrell. Stockholm. Amsterdam. Bowers . Geb. Amsterdam. 105119. Wahrscheinlichkeitsth. B. S. [78] L. Burman & D. [79] P. The Jubilee volume of Forsakringsbolaget Skandia.J. 14331453. 62. verve.W. [88] J. Oper. IEEE J. Bratley.W. Jones & C. Resnick & R.W. . 313319.I. [93] K. Broeck . Burman & D .J.R. New York. Brockwell . [85] W. Cramer (1930) On the Mathematical Theory of Risk. Syst. Wiley. Area Commun. 34. J. Cohen (1982) The Single Server Queue (2nd ed. Boxma & J. [87] E. Biihlmann ( 1970) Mathematical Methods in Risk Theory. 392433. New York San Francisco London.A.Y . Control AC28. Reiser eds.) NorthHolland.
Csorgo & J. Goovaerts (1984) Bounds for classical ruin probabilities. 16. [106] A. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Pentikainen & E. [99] R. [103] F.M. Insurance: Mathematics and Economics 3. Res. 583602.. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. 181217. Csorgo & J. Roy. Dembo & O. Laird & D. Insurance: Mathematics and Economics 7. Dempster. 3750. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . 27. [104] F. Daley & T. [111] F. Haezendonck (1987) Classical risk theory in an economic environment. 17. 85116. Probab. Stockholm. [107] L. J. [101] C. 9 . Soc. Appl. Act. Math. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). [102] P. 114119. [108] F. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. 624628. Insurance: Mathematics and Economics 4. Dassios & P. Math. Zeitouni (1992) Large Deviations Techniques and Applications. Springer. Insurance: Mathematics and Economics 6. Rolski (1984) A light traffic approximation for a single server queue. J. Martingales and insurance risk. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities.L. De Vylder & M. [96] D.368 BIBLIOGRAPHY [94] M.. Delbaen & J. Rolski (1991) Light traffic approximations in queues. J. Actuarial J. Suppl. Delbaen & J. English version published in Skand. Devroye (1986) Nonuniform Random Variate Generation.P. Haezendonck (1985) Inversed martingales in risk theory. Pesonen (1994) Practical Risk Theory for Actuaries. Tidskr. Scand. Math. [112] F. De Vylder (1977) A new proof of a known result in risk theory. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Oper. Aktuar. [98] A. De Vylder & M. Oper. Editions de 1'Universite de Bruxelles. [95] M. N. Insurance: Mathematics and Economics 10. Boston. Scand. New York. Stochastic Models 5. 3. 201206. [109] F. 1990 . J. [105] A. . Res. [110] F. 88101. 135159. T. [100] A. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Appl. 5771. [97] D. Statist. De Vylder (1996) Advanced Risk Theory. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. 7083 (1969). Jones and Bartlett. Deheuvels & J.D. Daley & T. Daykin. Embrechts (1989). 22. 121131. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics .B. Comp. 435436. Chapman & Hall. 227229.
Djehiche (1993) A large deviation estimate for ruin probabilities. 4962. 251262. 1984 . Harrison & A. Dufresne. Gerber (1988) The surpluses immediately before and at ruin. J. Emanuel . J.C.C. Act. Act. Astin Bulletin 21. Embrechts ( 1988 ). J. Dufresne & H.C. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. Soc.J. Dickson & H. 177192. Dickson (1994) An upper bound for the probability of ultimate ruin. Insurance: Mathematics and Economics 19.S. 118. [125] F. 1. 193199. .M. 1984 . Scand. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier.R. 363374. [114] D. H. Act. [117] D. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. Ruin estimates for large claims . Letters 7. J. 105115. 2000 . [120] D. Camb.B.BIBLIOGRAPHY 369 [113] D.M. 269274.M. Embrechts .M. Regterschot (1988) Conditional PASTA. 1741.W. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. Act. [124] N. Dickson & C.C. [123] E. Shiu (1991) Risk theory with the Gamma process. Duffield & N. Dickson & H.M. [122] B. Dickson & J. Insurance : Mathematics and Economics 7. 174186. [131] P. [130] P. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. Dickson & J. Berlin Gottingen Heidelberg. and the amount of claim causing ruin. [126] F. Waters (1999) Ruin probabilities wih compounding. Insurance : Mathematics and Economics 10. 191207. Res. 3745. [116] D. Math. Scand. Insurance : Mathematics and Economics 11.C. Act. Dickson (1992) On the distribution of the surplus prior to ruin .M. Scand. 107124. Dufresne & H. Act. 1993 . [118] D. J. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . [129] D . Act. Dickson & C. Dickson (1995) A review of Panjer's recursion formula and its applications.C. Scand. Scand. Gerber & E. 1994 .M.U. 1975 . [119] D.U.C. [121] D. Scand.R. Proc. Dynkin (1965) Markov Processes I. 4259. J. J. Hipp (1998) Ruin probabilities for Erlang (2) risk processes. 229232. British Actuarial J. J. van Doorn & G. Insurance: Mathematics and Economics 7. Insurance: Mathematics and Economics 25.K. [128] E.G. 5159.M. Springer .R.U. 6180.M. [115] D.C. 131138. Waters (1996 ) Reinsurance and ruin . J.J. Scand. 147167. Grandell & H.C. with applications. Oper.R. Insurance : Mathematics and Economics 22.C. Hipp (1999) Ruin problems for phasetype(2) risk processes. [127] F. Philos. 1993 .M.
Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. 616624. 14661474. Proschan & D. Schmidli (1994) Ruin estimation for a general insurance risk model. A bootstrap procedure for estimating the adjustment coefficient. Probab. 5572. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks.K. U.L. Adv. [141] F. [148] C. Danish Academy Tech. Erlang 2.W. Lai (1998) Wald's equations. Probab. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Insurance: Mathematics and Economics 7. Erlang (1909) Sandsynlighedsregning og telefonsamtaler .D. 159178. Heidelberg. [150] H. On the excursions of Markov processes in classical duality. Probab.D. [137] P. Adv. [146] E. 695712. [136] P. Esary. Kliippelberg & T. [143] W.L. [135] P. Insurance: Mathematics and Economics 10. Embrechts & H. Scand.). Astin Bulletin 16. Math. Fields 75. 26. Wiley. Approximations for compound Poisson and Polya processes. 269274. Appl. R. [144] P. Probab. Ann. Th. J. Konig. Statist. with applications.D. Wiley. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Sci. Griibel & S. [149] C. 38. Probab.W. 131137.J. Akt. 29. D. Appl. Fuh & T. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. Walkup (1967) Association of random variables. [133] P. 29. first passage times and moments of ladder variables in Markov random walks. 404422. [145] P. [138] P. C. Probab. F. Furrer (1998) Risk processes perturbed by astable Levy motion.K. Tidsskr. 623637. [140] J. 17. New York.). 8190. Statistica Neerlandica 47. J. Schmidt (1982) Queues and Point Processes. M. Franken. Appl. 175195. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . 5974. Maejima & J. Schmidt (1992) An insensitivity property of ladder height distributions. Arndt & V.370 BIBLIOGRAPHY [132] P. Frenz & V. 181190. 1998 . J. Esscher (1932) On the probability function in the collective theory of risk. Nyt Tidsskrift for Matematik B20. 5975. Appl. Insurance : Mathematics and Economics 1. Mikosch (1991). 35. [134] P. Rel. J.M. Skand. 3339. Embrechts & N. Springer. Fitzsimmons (1987).A. Appl. Embrechts. [139] A. Act. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. Teugels (1985). Jensen. Frees (1986) Nonparametric estimation of the probability of ruin. Embrechts & J. Villasenor (1988) Ruin estimates for large claims. New York. [147] M. Embrechts. Wiley. Embrechts. Trans. Adv. . Embrechts & T. 566580. [142] W.
3752. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. [166] M. 1978 . Haezendonck (1984) Insurance Premiums. Kou (1995) Limits of first passage times to rare sets in regenerative processes. Kaas. Adv. Griibel (1996) Perpetuities with thin tails. [158] H.U. Scand. Goovaerts. Gerber (1979) An Introduction to Mathematical Risk Theory. Act. Amsterdam. Grandell (1977) A class of approximations of ruin probabilities. Appl. Mitt. A. Birkhauser .U.U. Life Insurance Mathematics. Schweiz. Aktuarietidskrift 1970 .BIBLIOGRAPHY 371 [151] H. Galambos & J. [164] B. Appl. Gerber (1988) Mathematical fun with ruin theory. 2336. [155] H. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. University of Pennsylvania. 105115. Glasserman (1991) Gradient Estimation via Perturbation Analysis.G. Gerber (1973) Martingales in risk theory. Gani.M.U. Probab. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. 1977.N. Ver. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. Z. Huebner Foundation Monographs. 7778. [161] P.W. 424445. Vers. [159] H.).U. Amsterdam. Weron (1997) Stable Levy motion approximation in collective risk theory.V. [163] P. 6370. J. Ver. J. de Vylder & J.71. Act. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Glasserman & S. 31A.. [156] H. Insurance : Mathematics and Economics 7. .U.U. Skand. Probab. 131156. J. Glynn & W. Scand.). 5. Schweiz. [168] J. Furrer & H. contained in the authors PhD thesis. NorthHolland. Mitt. 28. Ann. S. Michna & A. Probab. [160] H. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. Basel. [153] H. Goovaerts. Kluwer. [165] M. [162] P. F. Gerber (1986). Math. Insurance: Mathematics and Economics 15.J. Dordrecht. Math. 73. Act. Vers.S. Gnedenko & I. Unpublished. [169] J. Insurance: Mathematics and Economics 20. Furrer. NorthHolland. 1981 . Appl. [157] H. In Studies in Applied Probability (J. [167] C. Boston. [154] H. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Springer. London. J. Scand. R. 205210. 463480. 97114. Suppl. [152] H.J. Bauwelinckx (1990) Effective Actuarial Methods. Goldie & R.E.. van Heerwarden & T. 1523. ETH Zurich. eds. 205216.
Aktuar Tidskr. Proc. Segerdahl (1971) A comparison of some approximations of ruin probabilities. Resnick (1977) The recurrence classification of risk and storage processes . Stoch. Proc. Skand. M. 54. [179] R. M.372 BIBLIOGRAPHY [170] J. [184] H . Opns . [181] D. Tinbergen Institute Research Series 20. Res. 1. Math. J. [171] J..and Sozialforschung 6. 566576. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. Harrison & S. 143158. 8. 167176. Lemoine (1977) Limit theorems for periodic queues. Grandell & C. M. New York. 14.S. Gut (1988 ) Stopped Random Walks . 3041. Proc . Insurance: Mathematics and Economics 26. Math. E. Winter School on Stochastic Analysis and Appl . Matem. Astin Bulletin 29. Informatica 2. Archiv v. [174] J. [176] B. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . [178] B. [180] R. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . Grandell (1979) Empirical bounds for ruin probabilities. Probab. 355365. Appl. . 14. [188] J. Harrison (1977) Ruin problems with compounding assets . Grnbel (1991) G/G/1 via FFT. Grandell (1997) Mixed Poisson Processes. Berlin. 197214.243255. Applied Statistics 40.O. 6779. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. 400409. 3 . Gyllenberg & D. Res. 33. Theory and Applications. Gusak & V.J. [173] J. Rink.I. Grandell (1990) Aspects of Risk Theory. Appl. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. 5766. mathematische Wirtschaft . Th. 347358. [182] A. AkademieVerlag. SpringerVerlag. Ibragimov. Grandell ( 1992) Finite time ruin probabilities and martingales . 5 . 332. [177] B. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. [186] J. [189] A. [187] J. Probability Theory and Mathematical Statistics (I. Grandell (1999) Simple approximations of ruin functions.A. Statistical Algorithm 265. Griibel & R. Amsterdam. ed. Probab. M. Appl. Grigelionis (1996) Lundbergtype stochastic processes. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . [172] J. [175] J. Opns .).V.I. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . KTH. 7590. Chapman & Hall.131135. Preprint. [183] M. Stoch. [185] J. Harrison & A. Harrison & S. SpringerVerlag. Appl. Liet.
Wahrscheinlichkeitsth. Appl. Act. Hoglund (1991). [206] B. ACM TOMACS 6. Hogg & S. 8996.A. [195] B. [204] T. [202] 0. Ann. . 166180. Insurance: Mathematics and Economics 5. [203] T. Act. 25. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. 5770. Wiley. 12981310. Mathematical Statistics. 259268. 378389. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Horvath & E. S. verve. [207] D. Astin Bulletin 19. Hojgaard & M. 18. Karlsruhe.a program for fitting phasetype distributions. 8095. [200] M. Hipp (1989a) Efficient estimators for ruin probabilities. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. 1998 . Haggstrom. eds. Probab. 305313. 23. J. Hill (1975) A simple general approach to inference about the tail of a distribution. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. No. 3. on Asymptotic Statistics (P. Prob. [192] U. Heilmann (1987) Grundbegriffe der Risikotheorie. Verlag Versicherungswirtschaft e. Scand.). The ruin problem for finite Markov chains.B. Probab. 285292. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Iglehart (1969) Diffusion approximations in collective risk theory. [194] 0.V. Appl. Proc.L. Scand. [199] R. [205] T. Mandl & M.R. [201] L. Karlsruhe. Foruth Prague Symp. Insurance: Mathematics and Economics 5. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems.. Iversen & L.V. Huskova. [196] C. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Electronic Letters 35. [193] U. Hipp & R. [197] C. [191] W. Nerman (1992) EMPHT . Probab. 4385. Math. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. Geb. 11631174. 19.M. Chalmers University of Technology and the University of Goteborg. Klugman (1984) Loss Distributions. Z.L. J. 123151. Nachrichten 30. [208] V. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. 377381. Ann. Ann. Taksar (2000) Optimal proportional reinsurance policies for diffusion models.BIBLIOGRAPHY 373 [190] P. 285293. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. 29. J. 6. Studies in Statistical Quality Control and Reliability 1992 :4. New York. 1990 . [198] C. J. Asmussen & 0. Versicherungswirtschaft. Statist.
Astin Bull. 547567. Keilson & D. Jagerman (1985) Certain Volterra integral equations arising in queueing. 116129. Kalashnikov (1996) Twosided bounds of ruin probabilities. University of Chicago Press. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. [227] J. Jelenkovic & A. 239256. 37. 6. 6. [217] R. ibid. Munksgaard. 3. Cambridge Philos. Kemperman (1961) The Passage Problem for a Markov Chain. Manuscript. Keilson & D. [225] J. NEC Research Institute. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. [214] A. Princeton. 325247. Scand. [220] V. [221] E. Statist.R. Probab.P. 35. 61. Jensen (1953) A Distribution Model Applicable to Economics. Proc.B.M. Johnson & M. Proc. [211] J. Astin Bull. 4151.A. Soc. Kluwer. Act.L.M. 118. [222] S. Taylor (1981) A Second Course in Stochastic Processes. . Probab.G.G. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. Stochastic Models 1. [223] J. [224] J. J. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Wishart (1964). 866870. Math.J. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. [218] V. N. 60. 711743. 173190.M. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. Ann. [228] J. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. N. Technometrics 30. Copenhagen. 63. [219] V. 11. 561563. 187193.M. Proc. Appl. [210] D.H. Cambridge Philos. ibid. Goovaerts (1986) General bound on ruin probabilities. 1996. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. Keilson & D.M. Taaffe (1989/90) Matching moments to phase distributions. Janssen & J. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. Kao (1988) Computing the phasetype renewal and related functions. Academic Press. J. J.J. New York. Jensen (1995) Saddle Point Approximations. 15. Insurance: Mathematics and Economics 5. [215] J. 259281.374 BIBLIOGRAPHY [209] D. 8793. J. Amer. 123133. Act. [213] P. Oxford. Cambridge Philos. 165167. Karlin & H. Soc. Soc. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories.G. Kaas & M. Clarendon Press. 31. Stochastic Models 5. [212] J.C. [226] J. Chicago. Appl. 283306. Addenda to for processes defined on a finite Markov chain. [216] M.
J. 259264. [239] H. Insurance: Mathematics and Economics 8. Math. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. [244] T. Cybernetica 4. Lemoine (1981) On queues with periodic Poisson input. [247] L. Latouche & V. Scand. Kliippelberg & U. J. Scand. Macmillan. Appl. [240] U. J. J. Kingman (1961) A convexity property of positive matrices . Proc. . [241] G. 48. Probab. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. 4477. Lehtonen & H. Camb. Lipsky (1992) Queueing Theory . J. Actuarial J. [246] D. 26. 1995 . 4958. Oxford 12. 18.F. Mikosch (1995) Modelling delay in claim settlement. Bernoulli 1 .C. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Soc. J.F. Appl. 383392. 279285. [242] A. Act. Klnppelberg & T.S. Quart. 25.J. Act. Kingman (1964) A martingale inequality in the theory of queues. 125147. 154168. Scand. Lindley (1952) The theory of queues with a single server. [243] A. Cambr. Kunita (1976) Absolute continuity of Markov processes. 146. [236] C. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. Proc. Philos. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling.F. Insurance : Mathematics and Economics 12. Soc. 359361. 390397. Appl. Statist. [248] D.F. 24. [234] C. 858874. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. Penev & A.J.P. I. [231] J. 283284. SpringerVerlag. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. Quart. Philos. New York. Kuchler & M. Probab. Roy. Prob.a Linear Algebraic Approach. Seminaire de Probabilties X . [238] V. Lehtonen & H. 133135 (in Russian). SIAM. Sorensen (1997) Exponential Families of Stochastic Processes. (235] C. [230] J. Probab. Adv. Soc.BIBLIOGRAPHY 375 [229] J. [233] C. 60 . Korolyuk. [245] T. Klnppelberg & T. Kingman (1962) On queues in heavy traffic. 6075. J. Lecture Notes in Mathematics 511. 889900. 132141. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. SpringerVerlag. 1998 .C. [237] C.. B24. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. [232] C. Kliippelberg (1988) Subexponential distributions and integrated tails. Appl.C. Stochastic Models 7. 277289.
[259] Z . Scand.S. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . 429448. Almqvist & Wiksell. a useful concept in risk theory. Miyazawa & V. [260] H. Nummelin (1987) Markov additive processes I. Nagaev (1957) Some limit theorems for stationary Markov chains. Th. J.S. 31. [264] G. 11 . [255] V. 1994. Probab. Appl. Math. [261] H.D. Gut & J. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. 378406. [252] A. [256] A. 15. Probab. MartinL6f (1983) Entropy estimates for ruin probabilities . Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. Act. 286298. J. Appl. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Cambridge Philos. Statist. MartinL3f (1986) Entropy. J. 223235. Englunds Boktryckeri AB.M. Act. Ann. Res. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. Act.376 BIBLIOGRAPHY [249] D. Appl. 4859. Ney & E. Probab. Scand.V. 124147. K. 1996 . A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Stockholm. Scand. [263] M . Uppsala. [262] H .D.). 58 . 32.F. Eigenvalue properties and limit theorems .K. Probab . Ann. Hoist eds. 676705. 763776. Probab. Probability and Mathematical Statistics (A. 58.D. Makowski ( 1994) On an elementary characterization of the increasing convex order . Schmidt ( 1993) On ladder height distributions of general risk processes . 147149.V. 29 .K. Meier ( 1984). Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. 1261 (0?)1270. 3. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. Adv. 36. Proc. Insurance : Mathematics and Economics 5. Stoch. II Aterforsdkring av Kollektivrisker . Ann. [266] P. J. Europ. Anal. Soc. Appl. 1986 . Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. [250] F. Soc. [265] S. Appl. 161174. J. [258] K . Opns. 22. Appl. 370377. . 2939. [257] A. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . [254] V . Probab. with an application . J. MeierHellstern & M . Proc. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Lucantoni . Michna ( 1998) Selfsimilar processes in collective risk theory. J. 834841. [253] V. 561592. Cambridge Philos. F. 268285. Math. . 2. [251] F.
Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. 327361. 273299. 36. 10081026. [276] C. Proc. 16. 75. . Paulsen & H. Appl. 733746. Insurance: Mathematics and Economics 22. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. 555564.F. 764779. [285] J.F. Abh. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. Sem. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Probab. Neuts (1992) Models based on the Markovian arrival process. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Paulsen (1998) Ruin theory with compounding assets . IEICE Trans. [278] E. London. [282] J . [274] H . Probab. Math. Neuts (1977) A versatile Markovian point process. [283] J. Statistics 21. [279] E. [280] A.K. [270] M . [271] M. 71. Adv. [268] M . Hamburg. 316. Appl. [281] J. Probab. [272] J . 445454. Appl. 339353. 12551265. 21. [284] J.G. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. [277] J.K. Marcel Dekker. Stoch. 46. Baltimore . Appl. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. Paulsen & H.A. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Norberg ( 1990) Risk theory and its statistics environment . Proc. 215223.F. J. Naval Research Logistics Quarterly 25. Commun. 30.a survey. Astin Bulletin 5 . Models 6. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. New York. Omey & E. 157. Insurance: Mathematics and Economics 20. Paulsen (1993) Risk theory in a stochastic economic environment. [273] R. Johns Hopkins University Press. Omey & E. Proc.F. 123144. [269] M . 12. Stoch. Stochastic Models 3 . J. Stoch. Nyrhinen (1999) Large deviations for the time of ruin. Appl. Neuts ( 1978) Renewal processes of phase type. 311342. J.BIBLIOGRAPHY 377 [267] M . Appl. Pakes (1975) On the tail of waiting time distributions. Proc. O'Cinneide (1990) Characterization of phasetype distributions.F. E75B . Probab. [275] H. Appl. 249266. Appl. Stoch. Stoch. 135148.
Opns. Astin Bulletin XIV. Springer. [304] B. [295] N.G. Regterschot & J. Math. Prabhu (1961). [294] N.K. [291] S. 117133. [296] N. Stat. Adv. Wiley. Statist. Ann. Soc.R. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve.U. . Inst. [300] C. 4568. Scand. [289] C.L. Probab. Queues. Rolski (1987) Approximation of periodic queues. Math. Aktuar. 757764. Resnick & G. Act..M. Math. [287] F. 2343. Berlin. Austr. Probab. 139143. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. On the ruin problem of collective risk theory. Ripley (1987) Stochastic Simulation . 4. 215246. Queueing Systems 5. 28. 11. [290] E. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. New York.U. 465483. Gjessing (1997c) Ruin theory with stochastic return on investments. [288] S. Ann. Tidskr. [298] V. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. 820827. Oper. Wiley. 390413. [297] R. 965985. [301] G. Appl. Res. Adv. Probab . Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. Probab. [302] J.J. Prabhu & Zhu (1989) Markovmodulated queueing systems. Appl. [299] C. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. Prabhu (1980) Stochastic Storage Processes. Appl. [293] N. Appl. 19. J. Wiley. 12. Rao (1965 ) Linear Statistical Inference and Its Applications. Ann. Insurance: Mathematics and Economics 3. J. Prabhu (1965) Queues and Inventories. Paulsen & H. [305] C. 30. Math.M.J. New York. 32. Pitts. Philipson (1968) A review of the collective theory of risk.A. Math. [292] S. Skand. 61. 1989 . 568576. 2330. Res. Statist. Pitman (1980) Subexponential distribution functions. J. Pyke (1959) The supremum and infimum of the Poisson process. 537555. and Dams. 147159. [306] T.G. [303] S. Heidelberg. Griibel & P. 337347. Adv. 235243. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. 29.U. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Appl.U. 46. Probab. 29A. Insurance Risk.378 BIBLIOGRAPHY [286] J. R. Ann. Embrechts (1996) Confidence bounds for the adjustment coefficient.K.M. 222261. Pellerey (1995) On the preservation of some orderings of risks under convolution.H. 691707. Insurance: Mathematics and Economics 16.
H. J. 431447 [316] S. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . Stochastic Models 6. 21.BIBLIOGRAPHY 379 [307] T. 795829. [324] H .M. Probab. Act. 227244. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . Scand. 4857. Melamed (1998) Modern Simulation and Modelling. Rubinstein & A. [313] M. Statist. 365388. Rubinstein (1981) Simulation and the Monte Carlo Method. 4068. Rolski. Adv. J.J. Singapore. Astin Bulletin 29. Appl. [322] H. 417421. Wiley. [308] S. Science Culture Technology Publishing . Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. 5. [309] H. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. Teugels (1999) Stochastic Processes for Insurance and Finance. . Appl. Wiley. Probab . Schmidli ( 1999b) Perturbed risk processes: a review . 11. Scand. [312] R. 135149. Stochastic Models 10 .Y. 262286. [326] H. [318] H. 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.L. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. 121133. Statist. Probab. [325] H . Schmidli. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Math. Wiley. J. 7. Rudemo (1973) Point processes generated by transitions of a Markov chain. [315] T. 155188. [319] H. 131156.Y. Schlegel ( 1998) Ruin probabilities in perturbed risk models. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Data Anal. Rubinstein & B. Insurance: Mathematics and Economics 22. Comp. Stoch. Ann. 1997. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. Wiley. [314] T. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . Act. Schmidli ( 1996) Martingales and insurance risk. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Wiley. 5. Seal ( 1969) The Stochastic Theory of a Risk Business. Operationsforsch.Y. 687699. [317] L. V. 2001 . Insurance: Mathematics and Economics 16. Appi. Stochastic Models 10. 93104. [320] H. Th. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. 145165. Proc. 5. [310] R. Schmidt & J. [311] R. [323] H. [321] H.
Queueing Systems 11. 11. Chapman and Hall. . Ann. 429442. Shaked & J. Grenander). 191197. Shin (1989) Ruin probability by operational calculus. [335] B . 77100. Skand. 157166. 483487. [337] M. New York. Scand.F. Probab. Mitteil. 243249. [346) D . t). [333] C: O. Academic Press. Sigman (1992) Light traffic for workload in queues. Versich. Appl. pp. 11. 914924. 4. 4. Segerdahl (1959) A survey of results in the collective theory of risk. Shantikumar (1993) Stochastic Orders and Their Applications. [344] D. 75. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. Statist.W. J. Math.the Harald Cramer volume (ed. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. In Probability and statistics . Probab. 61.S. [348] K .L. Verein Schweiz. Math. SpringerVerlag. t). 1974. Stockholm. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen .A. 383413. Seal (1974) The numerical calculation of U(w. Wiley.S. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. J. the probability of nonruin in an interval (0. Seal (1978) Survival Probabilities. [336] B. Chapman & Hall. Adv. 279299. [332] C : O.380 BIBLIOGRAPHY [327] H. Stochastic Models 6. Weiss (1995) Large Deviations for Performance Analysis. Probab . 21. Adv. L. [329] H. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. [339] A. Sengupta (1990) The semiMarkov queue: theory and applications. 1987. 1955. [334] C: O. 171178. [338] E. 4383. [331] G. Versich. Shin (1987) Convolution of uniform distributions and ruin probability. Appl. 673684. Wiley. Seal ( 1972). Insurance: Mathematics and Economics 8. [343] D.S. [328] H. [330] H. 2236. 159180. Math. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes .L. 121139. Aktuar Tidsskr. [342] D.L. Appl.G. Verein Schweiz Versich. [340] E. [345] D.W. Aktuar Tidsskr. Siegmund ( 1985) Sequential Analysis . Seber (1984) Multivariate Observations. 72. Act. 701719. Shwartz & A. Mitt. Verein Schweiz. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. Ann. [341] E. Risk teory and the single server queue . Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Scand. Almqvist & Wiksell. [347] K. Th. Act. Siegmund (1975) The time until ruin in collective risk theory. 72. Probab. . Mitt.
Act. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. New York. [363] G. Soc 49. [350] W.C. Wiley. 235254. Proceeedings of the IEEE 77. 163175.). 149162. 1978. [366] O. 1982 . Scand. Sundt & W. 118. [368] O. [356] B . J. John Wiley & Sons. [369] O. Act. Teugels (1982) Estimation of ruin probabilities . [355] B .). 1976. Act. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. Astin Bulletin 24.V. . Act. [357] B.L. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. Cambridge Philos. Ruin probabilities prepared for numerical calculations. [367] O. Scand.. [358] R. Scand.and largedeviations probabilities in actuarial risk theory.S. New York. Springer.C. Straub (1988) NonLife Insurance Mathematics. Scand. 449461. 5776. 8594. Act. 65102. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . 65102. Daley ed. J. J. 1974.J. Taylor (1986) Claims reserving in NonLife Insurance. NorthHolland. [353] E. Thorin (1977). 21. [354] B. 1980 . Teugels (1997) The adjustment coefficient in ruin estimates under interest force . [351] D . [364] G. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings .A.C. Appl. 197208. Slud & C.J. Karlsruhe. [359] L. Insurance: Mathematics and Economics 16. 722. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Insurance: Mathematics and Economics 1. 114137.BIBLIOGRAPHY 381 [349] E. Hoesman ( 1989) Moderate. Sundt & J. [360] G. [365] J. [362] G. Taylor (1979) Probability of ruin with variable premium rate. J. 2739. Probab. Thorin (1982) Probabilities of ruin.L. Teugels ( 1995 ) Ruin estimates under interest force . Adv.C. Astin Bulletin 12. Unpublished manuscript. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. [361] G. Stanford & K. Insurance: Mathematics and Economics 19.C. Actuarial J. Scand.L. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process.. J. Smith ( 1953) Distribution of queueing times. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . 8199. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Astin Bulletin 16 . Scand. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. Proc. [352] D. Verlag Versicherungswirtschaft e.L. Sundt & J. 725741.
E. Scand. Taqqu. [372] H. Goovaerts. [383] G. 148164.W. J. Willinger. Dordrecht Boston Lancaster. Aktuar. [378] H. Astin Bulletin VII. Thorisson (2000) Coupling. thesis. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. 31. R. . 231246. Insurance : Mathematics and Economics 13. Statistical Science 10. [381] W. [371] 0. D. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Tacklind (1942) Sur le risque dans les jeux inequitables. In: Premium Calculation in Insurance (F.E. 936952. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. [375] N. University of Michigan. 137153. [377] V. Submitted. 6785. 4963. [376] A. Wolff (1990) Stochastic Modeling and the Theory of Queues. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Astin Bulletin IX. J. Willmot & X.). [379] M. SpringerVerlag. De Vylder. Appl. Sherman & D. Lin (1994) Lundberg bounds on the tails of compound distributions. Insurance: Mathematics and Economics 15. Wald (1947) Sequential Analysis. Tidskr. Unpublished Ph. [373] S. 37. 1942. PrenticeHall. Wiley. Probab. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. [374] F. Van Wouve.R. 743756. M. F. 142. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. Skand. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. Act. Res. 1983 . 5762. [382] G. Oper. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Haezendonck eds. J. Reidel. Thorin & N. De Vylder & M. [380] W. Stationarity and Regeneration. [384] R. M. Thorin & N.382 BIBLIOGRAPHY [370] 0. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. Waters (1983) Probability of ruin for a risk process with claims cost inflation.
150. 7879. 39.299. 1819. 91. 283. 80 81. 180182. 111117. 17.249. 17.203.9899. 3436.242.281. 40. 9496. 9293. 278 gamma distribution 67. 170173. 79. 2526. 3839.135. 361 diffusion 3.and sum 221.285292. 318319 Erlang distribution 7. 37. 3032. 14.185187. 97. 196201 inverse Gaussian distribution 76. 7179. 218 Cox process 4. 138139. 207 heavytailed distribution 6. 226. 117127 corrected 121127 duality 1314. 15.121129. 271274. 89. 122.137141. 2425. 308.160167. 332333 Volterra 192194.Index adjustment coefficient 17.359 aggregate claims 103106. 245248. 8283 hyperexponential distribution 7.287292. 248 WienerHopf 144 interest rate 190.272.217. 7475.86.328330.226. 189. 97129. 217.269.200201.301 central limit theorem 60 . 162164. 205.4447.178184. 227229. 323 Coxian distribution 147. 1415.292293 Edgeworth expansion 113.251280 heavy traffic 76.307312 compound Poisson model 4.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.228229. 201214. 301 Kronecker product. 1112. 86. 117128.249250 integral equation 16 Lindley 143 renewal 64. 341.346349 383 . 302303 diffusion approximation 17.100. 119. 5796.6779. 201 Brownian motion 3 . 239.259261.182. 135. 9396. 141144.203. 316323 Bessel function 102. 12 CramerLundberg approximation 1617. 7079. 3334.318320 change of measure 2630. 4851.293294. 110113.314316. 5. 360 excursion 155156.
230.336339 .161164. nonlinear 155. 171. 141144. 16.160161. 38. 38. 4446. 132133. 35.161.108. 295.259261. 37. 162. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 3947. 149. 2730. 42.339 large deviations 129. 9899. 15.302. 261264.178182.218221. 86 periodicity 12.227230. 96. 203204. 25.201. 16. 99. 142 likelihood ratio : see change of measure lognormal distribution 9. 227228.180.152160. 134135. 306316 Levy process 3. 3639.384 ladder heights 4756. 39. 3947.234. 145187. 25. 59. 6970.128129.348 terminating 215216. 234 matrixexponential distribution 240244 matrixexponentials 14. 154.148. 41. 52 53. 108 life insurance 5. 157. 179 NP approximation 318320 Palm distribution 5253. 108109. 112113.261264. 176185. 133.139141.269271. 203 Markov additive process 12. 133.297299. 100.336339 Laplace transform 15.350361 Poisson process Markovmodulated 12 periodic 12.174. 5758. 185187 GI/G/1 141144 M/D/1 6667 equation 16.340350 multiplicative functional 2830.123.298299. 213214. 267269 Panjer's recursion 320323 Pareto distribution 910. 65. see also sensitivity analysis phasetype distribution 8. 7179.287291 INDEX matrix equation . 35. 175 light traffic 8183 Lindley integral equation 143 process 3334. 113114.234240. 106108.349 350 perturbation 172173. 260 Lundberg conjugation 6979 .275278. 6162.161. 137139.146148.238. 178 modulation 12. 7576.304 process 2830.315 inequality 1718. 80.240244. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167.134135. 257.288290. 44.215250. 32. 44.285287 queue 14 . 71. 144. 251. 304305 random walk 3336. 134. 269 PerronFrobenius theory 4142. 14. 138. 245 M/G/1 13.287. 271274.
257. 107. 260 WienerHopf theory 144. 141144. 12. 37. 251. 307308. 222. 240. 294296 shotnoise process 314 simulation 19. 335336 sensitivity analysis 8693. 189214. 260 reinsurance 8. 160.279280 Rouche roots 158. 233.186. 147. 89. 251. 233234. 3032. 331336 equation 64. 229234. see also matrixexponential distribution regenerative process 264 268. 186187 virtual: see workload rational Laplace transform 8. 261264 reservedependent premiums 14. 317318 semiMarkov 147. 256258. 1819. 213. 168172 storage process 13. 280. 327 . 60. 5455. 186187 renewal process 131. 191192. 31. 131144. 162.359361 stochastic control x stochastic ordering 18. 7475.262263. 177 timereversion 14. 251280 time change 4. 238 saddlepoint method 115117. 338 utility 324.273274. 174. 9693. 253. 279280 subexponential distribution 11. 172173.154157. 152. 4950. 120 statistics x. 292294. 326330 Weibull distribution 9. 333334 regular variation 10. 332333 model 12. 8386.336339 workload 13. 123.INDEX 385 waiting time 141. 244.244250. 223226. 87. 281296 stable process 15.314. 11. 146.
exact solutions.com 2779 he 9 "789810ll22293211 . It is a comprehensive treatment of the known results on ruin probabilities. I 1! Ruin Probabilities .. y finite horizon ruin probabilities. extensions of the classical compound Poisson model to allow f o r reservedependent premiums. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. Special features of the book are the emphasis on change of measure techniques.Vol. 2 A I 11 JjVb l' i  i Yj . P'i yfliother approximations (e.Advanced Series on Statistical Science & Applied Probability . the ^W A l \ i l ' ''' CramerLundberg approximation. worldscientific. for heavytailed claim size distributions). Some i (l I JL I J r of the topics are Lundberg's inequality.g." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. Markovmodulation or periodicity.. "This book is a must for anybody working in applied probability.
This action might not be possible to undo. Are you sure you want to continue?