This action might not be possible to undo. Are you sure you want to continue?
Ruin Probabilities
Seren Asmussen
World Scientific
Ruin Probabilities
ADVANCED SERIES ON STATISTICAL SCIENCE & APPLIED PROBABILITY
Editor: Ole E. BarndorffNielsen
Published Vol. 1: Random Walks of Infinitely Many Particles by P. Revesz Vol. 2: Ruin Probabilities by S. Asmussen Vol. 3: Essentials of Stochastic Finance : Facts, Models, Theory by Albert N. Shiryaev Vol. 4: Principles of Statistical Inference from a NeoFisherian Perspective by L. Pace and A. Salvan Vol. 5: Local Stereology by Eva B. Vedel Jensen Vol. 6: Elementary Stochastic Calculus  With Finance in View by T. Mikosch Vol. 7: Stochastic Methods in Hydrology: Rain, Landforms and Floods eds. O. E. Barndorff Nielsen et al. Vol. 8: Statistical Experiments and Decisions : Asymptotic Theory by A. N. Shiryaev and V. G. Spokoiny
Ruin P robabilities
Soren Asmussen
Mathematical Statistics Centre for Mathematical Sciences Lund University
Sweden
World Scientific
Singapore • NewJersey • London • Hong Kong
Published by World Scientific Publishing Co. Pte. Ltd. P O Box 128, Fatter Road , Singapore 912805 USA office: Suite 1B, 1060 Main Street, River Edge, NJ 07661 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Library of Congress CataloginginPublication Data Asmussen, Soren
Ruin probabilities / Soren Asmussen. p. cm.  (Advanced series on statistical science and applied probability ; vol. 2) Includes bibliographical references and index. ISBN 9810222939 (alk. paper) 1. InsuranceMathematics. 2. Risk. I. Tide. II. Advanced series on statistical science & applied probability ; vol. 2. HG8781 .A83 2000 368'.01dc2l 00038176
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
First published 2000 Reprinted 2001
Copyright ® 2000 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
Printed by Fulsland Offset Printing (S) Pte Ltd, Singapore
Contents
Preface I ix
Introduction 1 1 The risk process . . . . . . . . . . . . . .. . . . .. .. . . . . 1 2 Claim size distributions .. . . . . . . . .. . . . . . . . . . . . 5 3 The arrival process . . . . . . . . . . . . . . . . . . . . . . . . 11 4 A summary of main results and methods . . . . .. . . . . . . 13 5 Conventions . .. . .. .. . . . . . . . . . . . . . . . . . . . . 19
II Some general tools and results 23 1 Martingales . .. . .. .. . . . . . .. . . . . . . . . . . . . . 24 2 Likelihood ratios and change of measure . . .. . . . . . .. . 26 3 Duality with other applied probability models . . .. . . . . . 30 4 Random walks in discrete or continuous time . . . . . . . . . . 33 5 Markov additive processes . . . . . . . .. . . . . . . . . . . . 39 6 The ladder height distribution . . . .. . .. .. . . . . . . . . 47
III The compound Poisson model 57 1 Introduction . . . . . . . . .. .. .. . .. .. . . . . . . 58 . . . . . . . . . . . . . . . 61 3 Special cases of the PollaczeckKhinchine formula . . . . . . . 62 4 Change of measure via exponential families . . . .... . .. . 67 5 Lundberg conjugation . .. . . . . . . . . . . . . . . . . . . . . 69 6 Further topics related to the adjustment coefficient .. . . . . 75 7 Various approximations for the ruin probability . . . . . . . . 79 8 Comparing the risks of different claim size distributions . . . . 83 9 Sensitivity estimates . . . . . . . . . . . . . . . . . . . . . . . 10 Estimation of the adjustment coefficient . . . . . . . . . . . . 86 93 2 The PollaczeckKhinchine formula
v
vi
CONTENTS
IV The probability of ruin within finite time 97 1 Exponential claims . . . . . . . . . . . . . . . . . . . . . . . . 98 2 The ruin probability with no initial reserve . . . . . . . . . . . 103 3 Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . 108 4 When does ruin occur? . . . . . . . . . . . . . . . . . . . . . . 110 5 Diffusion approximations . . . . . . . . . . . . .. . . .. . . . 117 6 Corrected diffusion approximations . . . . . . . . . . .. . . . 121 7 How does ruin occur ? . . .. . . . . . . . . . . . . . . . . . . . 127 V Renewal arrivals 131 1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . 131 2 Exponential claims. The compound Poisson model with negative claims . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 3 Change of measure via exponential families . . . . . . . . . . . 137 4 The duality with queueing theory .. .. .. . . . .. . . . . . 141 VI Risk theory in a Markovian environment 145 1 Model and examples . . . . . . . . . . . .. . .. . . . . . . . 145 2 The ladder height distribution . . . . . . . . . .. . . . . . . . 152 3 Change of measure via exponential families ........... 160 4 Comparisons with the compound Poisson model ........ 168 5 The Markovian arrival process . . . . . . .. .. . . ... . . . 173 6 Risk theory in a periodic environment .. . . . .. . . . . . . . 176 7 Dual queueing models .... ... ................ 185 VII Premiums depending on the current reserve 189 1 Introduction . . . . . . . . . . . . . . . . . . . .. . . . . . . . 189 2 The model with interest . . . . . .. . . . . . . . . . .. . . . 196 3 The local adjustment coefficient. Logarithmic asymptotics . . 201 VIII Matrixanalytic methods 215 1 Definition and basic properties of phasetype distributions .. 215 2 Renewal theory . . . . . . . . . . . . . . . . . . . . . . . . . . 223 3 The compound Poisson model . . . . . . . . . .. . . . . . . . 227 4 The renewal model . . . . . . . . . . . . . . . .. . . . . . . . 229 5 Markovmodulated input . . .. . . . . . . . . . . . . . . . . . 234 6 Matrixexponential distributions . . . . . . . . . . . .. . . . 240 7 Reservedependent premiums . . . . .. . . . .. . . . . . . . 244
. . . . . . . 336 A3 Matrixexponentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . . 340 A4 Some linear algebra . . . . . . . . . . 297 2 Further applications of martingales . . . . . . . 290 5 Regenerative simulation . . . .. . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261 4 Models with dependent input . . .. . . . . . . . . . . .. . . . . The twobarrier ruin problem . . . 292 6 Sensitivity analysis . . . 316 5 Principles for premium calculation . . . . . 344 AS Complements on phasetype distributions . . . . . . . . . . .. . . . .. 279 X Simulation methodology 281 1 Generalities . . . . . . . . .. . . . .. . . . . . .. . . . . . . . . 306 4 The distribution of the aggregate claims . . . . . . . . . 304 3 Large deviations . . . . . . . . . 259 3 The renewal model . . . . . . . . . . . . . . . . . . . . .. . . . . . . 331 A2 WienerHopf factorization . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . . . . .. . . . . . .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . .. .. . . . . 287 4 Importance sampling for the finite horizon case . . . . . . . 251 2 The compound Poisson model . . . ... . . . . . . . . . . . . . . .. . . . . 264 5 Finitehorizon ruin probabilities . . 350 Bibliography Index 363 383 . . . . . 271 6 Reservedependent premiums . . . . . . . . . . . . . . . . . . . . . . .. . . . . 285 3 Importance sampling via Lundberg conjugation . . . . . . . .. . . 326 Appendix 331 Al Renewal theory . . . . . . .. . . . . . . . .. . . . . . . . .. . . . .. . . . 323 6 Reinsurance . .
This page is intentionally left blank .
which has in particular removed one of the standard criticisms of the area. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. University of Copenhagen. it is not by intention. But the pace was much slower than expected. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. 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. this applies to longrange dependence which is intensely studied in the neighboring ix . However. and the result is now that the book is much more related to my own research than the initial outline. I have deliberately stayed away from discussing the practical relevance of the theory. and my belief was that this could be done rather quickly. and other projects absorbed my interest. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. but the handouts were written and the book was started (even a contract was signed with a deadline I do not dare to write here!). that it can only say something about very simple models and questions. The course was never realized. if the formulations occasionally give a different impression. As an excuse: many of these projects were related to the book. Apart from these remarks. the book is basically mathematical in its flavour. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. It has obviously not been possible to cover all subareas. and the series editor Ole BarndorffNielsen for their patience. the idea was close to expand these to a short book on the subject. Since I was to produce some handouts for the students anyway. In particular. Thus. 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.
Finally. 111.45. Good luck! I have tried to be fairly exhaustive in citing references close to the text. A book like this can be organized in many ways.89. read Chapter I.13 and XI.13. IV. For a brief orientation.g. see in particular Michna [259].4a. it has not been possible to incorporate more numerical examples than the few there are.5.14. 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 . The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. Asmussen.13.15.lth. Hojgaard & Taksar [206]. The present book is in between these two possibilities. For a second reading. Hojgaard & Taksar [35] and Paulsen & Gjessing [284].maths . another by method. [381]). X.x PREFACE field of queueing theory. I intend to keep a list of misprints and remarks posted on my web page. see also Schmidli [325] and the references in Asmussen & Taksar [52]. 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. The main motivation comes from statistical data for network traffic (e. More recently. Concerning ruin probabilities. In the classical setting of CramerLundberg models. VII. IX. Another interesting area which is not covered is dynamic control. 111. It is obvious that such a system involves a number of inconsistencies and omissions.2 more properly).lth. see e.2. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. I regret that due to time constraints.3.13 and IX. for the effects on tail probabilities. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. for which I apologize to the reader and the authors of the many papers who ought to have been on the list.6 (to understand the PollaczeckKhinchine formula in 111.2. VI. VII. the first part of 11. One is by model. some papers not cited in the text but judged to be of interest are included in the Bibliography.g. IV. In addition. http:// www. IV.g. e.se Lund February 2000 Soren Asmussen .1. Resnick & Samorodnitsky [303] and references therein. the standard stochastic control setting of diffusion models has been considered. VIII. an area which is becoming increasingly important. Willinger et al. Here is a suggestion on how to get started with the book.
many of which were pointed out by Hanspeter Schmidli .3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers.6. IV. as well as some additional references continue to be at the web page. 111 . 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . of which there are not many at this stage .6 by my 1999 simulation class in Lund.2 by Rafal Kulik .1 and X. Section VIII. Section VII . Parts of II. not least the more complicated ones. Fig. . Fig.4 from Asmussen. 5.1 by Bjarne Hojgaard and the table in Example 111.PREFACE xi The second printing differs from the first only by minor corrections. Parts of X.8 . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX.5 from Asmussen [21] with permission from CRC Press. Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . were produced by Lone Juul Hansen . supported by Center for Mathematical Physics and Stochastics (MaPhySto).6 is reprinted from Asmussen & Schmidt [49] and parts of IX. A number of other figures were supplied by Christian Geisler Asmussen . Aarhus. More substantial remarks.
This page is intentionally left blank .
Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. A risk reserve process { Rt}t>o.2) (O<t<T Ro=ul.1) We also refer to t/) ( u) and 0(u. results and topics to be studied in the rest of the book. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t. is a model for the time evolution of the reserves of an insurance company. We denote throughout the initial reserve by u = Ro. MT = sup St. They are the main topics of study of the present book. we introduce some general notation and terminology. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero.3) sup St. (1. as defined in broad terms . and give a very brief summary of some of the models. M = (1. respectively.4) O<t<oo O<t<T 1 .Chapter I Introduction 1 The risk process In this chapter .i(u.Rt. T) as ruin probabilities with infinite horizon and finite horizon .T) = P inf Rt < 0 I . (1. (1. For mathematical purposes.
(1. per unit time. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon.1 ..7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig.b(u) = P (r(u) < oo) = P(M > u). (1. and T1 is the time of the first claim. say.T) = F (MT > u) = P(r(u) < T). T3.i(u. (1.E Uk.6) Sofar we have not imposed any assumptions on the risk reserve process. the number Nt of arrivals in [0. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un.1.. Thus. Putting things together. we see that Nt Nt Rt = u + pt . St = E Uk . That is.2 CHAPTER I. . the ruin probabilities can then alternatively be written as .5) i. the time of arrival of the nth claim is an = T1 + • • • + Tn. However. • Premiums flow in at rate p. 1. We denote the interarrival times of claims by T2. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. t] is finite.pt. Figure 1. respectively.
1 the slope of {Rt} should depend also on the level).e. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. We shall discuss Brownian motion somewhat in Chapter IV. • Brownian motion or more general diffusions. a basic references is Gerber [127].1. We shall not deal with this case either. For the purpose of studying ruin probabilities this distinction is. If 77 > 0. and in fact: Proposition 1. say 10% . immaterial. 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. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. on Fig. however.b(u) = 1 for all u. since any modeling involves some approximative assumptions. for example. However. allowing a countable infinity of jumps on Fig. though many results are straightforward to generalize from the compound Poisson model.s. that the insurance company should try to ensure 77 > 0.. however. and hence O(u) < 1 for all sufficiently large u. then M = oo a. Thus.1. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. but as an approximation to the risk process rather than as a model of intrinsic merit. and the basic ruin probabilities are derived in XI. one may well argue that Brownian motion in itself could be a reasonable model. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i.1 Assume that (1.1.8) The interpretation of p is as the average amount of claim per unit time.. It would appear obvious.(. and hence . we shall. not discuss whether this actually corresponds to practice. one could well replace Rt by Rtnr(u) or RtA. If 77 < 0.s. 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. 1.) V 0. We study this case in Ch.20%. (1.8) holds. of course. VII. 1. . then M < oo a. t * oo.
3 Assume p 54 1 and define Rt = Rt1p. this needs to be verified in each separate case.v. . Nt)}.. 0(u. are i. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model.8) is given by ^t p = EU • lim it (3(s) ds t. (1.i(u. are i. k=1 (1. in connection with risk processes in a Markovian or periodic environment (Chapter VI). and independent of {Nt}. then this limit is > 0 which implies St a$ oo and hence M = oo a. t t p  p' t ^ oo. tb(u) = 1 for all u holds also when rl = 0.. .11) .10) is a property which we will typically encounter. M < oo a. and that . if {(3(t)} is nonergodic. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). This case is referred to as the mixed Poisson process. However. . and independent of {(0(t).T) = i. rl > 0.6EU (on the average. Thus p may well be random for such processes.. (1. with the most notable special case being V having a Gamma distribution. INTRODUCTION Proof It follows from (1.oo t 0 J (provided the limit exists). 0 We shall only encounter a few instances of a Cox process. not all models considered in the literature have this feature: Example 1.d.s.8) that F N.10) Again. then similarly limSt/t < 0. U2.T) for {Rt} is given by V)(u) = t/i (u). we obtain typically a somewhat stronger conclusion. Proposition 1. and here (1.. However.i.. The simplest example is 3(t) = V where V is a r . namely.Q (say) and U1.Tp). If U1. If u oo. (1.i. If 77 < 0. it is not too difficult to show that p as defined by (1. Here it is easy to see that p = .2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t. namely that M = oo a.4 CHAPTER I.10) hold with p constant. _ St __ k =1 Uk pt a4.s.b(u) < 1 for all u when rl > 0.8). where {Nt} is a Poisson process with rate .d.s. U2.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. St In concrete models. zP(u . corresponding to the Pdlya process.
we shall be able to identify p with the traffic intensity of an associated queue. Note that when p = 1. U2. Schmidt & Teugels [307] and Seal [326]. Daykin.g... while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. Note that life insurance (e. [330]. [134]. the recent survey by Grandell [173] and references therein. Insurance: Mathematics and Economics. Some early surveys are given in Cramer [91]. but in probability and applied probability as a whole. We roughly classify these into two groups . see also Chapter XI. in a number of models. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. Embrechts et al. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. in particular. Grandell [171]. Straub [353]. Pentikainen & Pesonen [101]. [101]. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. another important early Swedish work is Tacklind [373]. Some main later textbooks are (in alphabetical order) Buhlmann [82]. Besides in standard journals in probability and applied probability. Hipp & Michel [198]. Notes and references The study of ruin probabilities. the role of the result is to justify to take p = 1. De Vylder [110]. Buhlmann [82]. Schmidli. 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). The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. In the even more general area of nonlife insurance mathematics. the assumption > 0 is equivalent to p < 1. 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.g. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Taylor [364]. often referred to as collective risk theory or just risk theory. Gerber [157]. Heilmann [191]. Rolski. the research literature is often published in journals like Astin Bulletin . many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. [76]. see e .2. Daykin et al. lighttailed distributions (sometimes the term . the claim arrivals are Poisson or renewal at the same time). Segerdahl [334] and Philipson [289]. Cox processes are treated extensively in Grandell [171].. Since { Rt } has premium rate 1. Some of the main general ideas were laid down by Lundberg [250]. Gerber [159]) has a rather different flavour. For mixed Poisson processes and Polya processes.. Sundt [354]. and we do not get near to the topic anywhere in this book. The Swedish school was pioneering not only in risk theory.
B is heavytailed if b[s] = oo for all s > 0.2) = 0. and heavytailed distributions. For example in the compound Poisson model. for the compound Poisson model with exponential claim sizes the ruin probability .2 and /LB is the mean of B. but different more restrictive definitions are often used: subexponential.g.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant).u at the time of ruin given r(u) is again exponential u with rate 8. B[s] is finite for some s > 0.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. Example 2 . a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . The crucial feature is the lack of memory: if X is exponential with rate 6. if 1 °O AB Jbos x B(dx) > 0. In particular.f.B(x) satisfies B(x) = O(e8x) for some s > 0. In contrast. regularly varying (see below) or even regularly varying with infinite variance.6 CHAPTER I. 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. P B[s]= (8Is ) .2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. the m.1) The parameter 6 is referred to as the rate or the intensity. s<8. INTRODUCTION 'Cramertype conditions' is used). 2a Lighttailed distributions Example 2. Here lighttailed means that the tail B(x) = 1 .3) . then the conditional distribution of X . where B(bo. On the more heuristical side.e. Equivalently.f. (2. 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'. i.g.O(u) can be found in closed form.8. a fact which turns out to contain considerable information. and can also be interpreted as the (constant) failure rate b(x)/B(x). As in a number of other applied probability areas.
.y i=1 where >i ai = 1. P b(x) = r` aibiea.v.v.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. x] so that B(x) = r` e. by Grandell & Segerdahl [175] and Thorin [369]. 0 < ai < 1. p).2) can be considered as the pth power of the exponential density (2. where X1.c. u .. u Example 2 .) VarX1 (EX )2 p is < 1 for p > 1. B(x) = r(p) Asymptotically. we develop computationally tractable results mainly for the Erlang case (p = 1. The exact form of the tail B(x) is given by the incomplete Gamma function r(x. one has r(bx.).. is > 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 .d. . and exponential with rate d.1) (or the 1/pth root if p < 1). X2.ate (b2 ): L• i=o In the present text.. the squared coefficient of variation (s.. . 0. the Gamma density (2. i = 1. or just the Erlang(p) distribution.i. p. > 1 for p < 1 and = 1 for p = 1 (the exponential case). In particular.. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. 2. p) °° where r (x. if p is integer and X has the gamma distribution p. An important property of the hyperexponential distribution is that its s.2. . This special case is referred to as the Erlang distribution with p stages. then X v Xl + • • • + X. are i.. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. Ruin probabilities for the general case has been studied. among others.1 Poisson events in [0. In particular. p) = J tPletdt.c.
. This class of distributions is popular in older literature on both risk theory and queues. However.7) are possibly complexvalued but the parameters in (2. equivalently. We give a more comprehensive treatment in VIII. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities. the Erlang and the hyperexponential distributions. B(x) = aeTxe where t = Te and e = (1 . The density and c. 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.d. q2 q3 (2. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree..6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. See XI.1 and defer further details to u Chapter VIII.g. 1)' is the column vector with 1 at all entries.xo)+ is covered by the reinsurer).f. there exists a xo < oo such that B(x) = 0 for x > xo. Example 2 . Example 2 . 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 . T) is called the representation. of which one is absorbing and the rest transient. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. T) or sometimes the triple (E.e. a.(2.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . which is slightly smaller but more amenable to probabilistic reasoning. The parameters of a phasetype distribution is the set E of transient states. are b(x) = aeTxt. Important special cases are the exponential..5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. resp.f. INTRODUCTION Example 2 .8) are realvalued.6. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. The couple (a. We give some theory for matrixu exponential distribution in VIII. (or.6.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states.8 CHAPTER I.8) j=1 j=1 j=1 where the parameters in (2. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution.
Example 2 . a)/A)a+1' x > a.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. p is defined as the distribution of ev where V . In particular. the exponential distribution representing the simplest example since here b(x) is constant. b(x) = crx''le`xr.10) The loinormal distribution has moments of all orders.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory.u l b(x) = d dx or J ax lor 1 exp Asymptotically. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. (2. x < a.9) which is heavytailed when 0 < r < I.N(0. the mean u is eµ+a /2 and the second moment is e2µ+2o2. There are various variants of the definition around. (2.1.1). and then b(x) = 0. the tail is B (x ) 2 x.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2. (2. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . we obtain the Weibull distribution B(x) = eCx'.N(p. It follows that the density is 't (1ogX .13) u . u Example 2 . However.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. or equivalently as the distribution of a°U+µ where U .2. a2). Writing c = d/r. Here failure rates b(x) = b(x)/B(x) play an important role.p a 1 (2.pl = 1 W (logx . All moments are finite. 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). CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2.
B(x) = O(xP). 6 is defined as the distribution of et' where V has the gamma density (2.e. u . (2. u Example 2 .x6+lr(p) (2.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. in particular. x + 00. oo) is slowly varying .v. x 4 oo (any L having a limit in (0. The simplest examples correspond to p small and integervalued. The density is 8p(log x)pi b(x) .14) The pth moment is finite if p < 5 and infinite if p > 5.(1 + Zx + $ p = 3.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate.(1 + 2x + 2x2)e2x) p = 2 (2. In general. A = 1 and X is standard exponential.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) . where Y is Pareto distributed with a = (p .1)/p.2). 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).L( x ).16) 11 Example 2. the density is { 3 (1 . INTRODUCTION Example 2. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. satisfies L(xt)/L(x) 4 1. (2.15) x2 + 16x3 ) a3x/2) 3 (1 . For p = 1.10 CHAPTER I. another standard example is (log x)'). examples of distributions with regularly varying tails are the Pareto distribution (2. the loggamma distribution is a Pareto distribution. Choudhury & Whitt [1] as the class of distributions of r. { s () 1s+3s29s3log(1+2s I p=3. i. Thus.12) (here L (x) * 1) and ( 2.'s of the form YX.13). in particular.17) where L (x) is slowly varying.
By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1.. one may argue that this difficulty is not resticted to ruin probability theory alone. (2. However.3. but the model also admits a natural interpretation : a large portfolio of insurance holders . U2. Similar discussion applies to the distribution of the accumulated claims (XI. We return to a closer study in IX.1. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. 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. this phenomenon represents one of the true controversies of the area. 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. Namely. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available.18) B(x) It can be proved (see IX. Thus. the knowledge of the claim size distribution will typically be based upon statistical data.. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed..13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2.. and based upon such information it seems questionable to extrapolate to tail behaviour. and so is the Weibull distribution with 0 < r < 1. which each have a ( timehomogeneous) small rate of experiencing a . When studying ruin probabilities.1) that any distribution with a regularly varying tail is subexponential. for example the lognormal distribution is subexponential (but not regularly varying). 3 The arrival process For the purpose of modeling a risk process . From a practical point of view. THE ARRIVAL PROCESS 11 Example 2.4) or even to completely different applied probability areas like extreme value theory: if we are using a Gaussian process to predict extreme value behaviour. Also. The reason is in part mathematical since this model is the easiest to analyze. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. though the proof of this is nontrivial. the claim size distribution represents of course only one aspect (though a major one).
Historically. has some mathematically appealing random walk features . but with a general not necessarily exponential distribution ). To the author 's knowledge . however. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). epidemics in life insurance etc. Cox processes are. not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . INTRODUCTION claim . gives rise to an arrival process which is very close to a Poisson process. to be studied in Chapter V. In order to prove reasonably substantial and interesting results . found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives .i. T2.6. in particular to allow for certain inhomogeneities. too general and one neeed to specialize to more concrete assumptions . A more appealing way to allow for inhomogeneity is by means of an intensity . the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . getting away from the simple Poisson process seems a crucial step in making the model more realistic. and also that the ruin problem may be hard to analyze . The compound Poisson model is studied in detail in Chapters III. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . which facilitate the analysis. However . This applies also to the case where the claim size distribution depends on the time of the year or .. where {/3 (t)}too is an arbitrary stochastic process . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. I. Another one is Cox processes. with a common term {Nt} is a Markovmodulated Poisson process .8 (t) is a periodic function of t. Mathematically. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. with the extension to premiums depending on the reserve. when Jt = i.d.e. Some of them have concentrated on the marginal distribution of NT (say T = one year ). it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. so that .. Nevertheless . 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.12 CHAPTER I. its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. see 11.. 5. IV (and. in Chapter VII).3(t) fluctuating over time. This model . e. such that 8(t) = . 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 periodic and the Markov modulated models also have attractive features . In others..g. An obvious example is 3(t) depending on the time of the year (the season). the negative binomial distribution. are i.(3. 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 . we study this case in VI .
Thus. 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. that quite often the emphasis is on computing expected values like EV.1). it is desirable to have a set of formulas like (4.v. Markovmodulated or periodic can be related to queues with similar characteristics. others are quite different. (4. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. time series and Gaussian processes.6) . queueing theory. More generally. stochastic differential equations. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. The study of the steady state is by far the most dominant topic of queueing and storage theory. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. extreme value theory.1) holds as well provided the risk process has a premium rule depending on the reserve. and a lot of information on steadystate r.4. and which seems well motivated from a practical point of view as well.0 (u. methods or modeling ideas developed in one area often has relevance for the other one as well. and here (4. 0(u) = P(V > u). R = p(R) in between jumps. 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 . In fact. Some of these have a certain resemblance in flavour and methodology. genetics models. and the limit t 4 oo is the steadystate limit. Similarly. It should be noted. this gives only f0 O°i (u)du which is of limited .T) = P(VT > u). with Poisson arrivals and constant release rule p(x) = 1. Mathematically. In the setting of (4. others being branching processes. The ones which appear most related to risk theory are queueing theory and dam/storage processes.1) permitting to translate freely between risk theory and the queueing/storage setting. dam/storage processes. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. interacting particle systems.'s like V is available. ruin probabilities for risk processes with an input process which is renewal. stochastic geometry. reliability. this amounts to Vo having the stationary distribution of {Vt}). however.1) where V is the limit in distribution of Vt as t + oo. point processes and so on.
see Corollary VII.1 . • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B.. Example VIII. 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. though overlapping.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r.3. Similarly. 4b Exact solutions Of course . as is typically the case.'s like the environmental process {Jt} in a Markovmodulated setting.1. The fact that Theorem H. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .3. have to some extent a different flavour. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u).3. 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. the two areas. The infinite horizon (steady state ) case is covered by letting T oo.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 ). A prototype of the duality results in this book is Theorem 11.2). p = 0/8 and y = 8 . Thus . Here O(u) = peryu where 3 is the arrival intensity. e . see Corollary III.3. 3. • The compound Poisson model with a claim size distribution degenerate at one point. INTRODUCTION intrinsic interest . . 3. B(x) = ebx. which can be expanded into a sum of exponential terms by diagonalization (see.g.T). • The compound Poisson model with some rather special heavytailed claim size distributions. which gives a sample path version of (4. Vi(u. the functions w x f d 1 exdx () .8.14 CHAPTER I.p(y) y^ Jo p(x) can be written in closed form.v. see Boxma & Cohen [74] and Abate & Whitt [3].1). Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. Here ?P(u) is explicit provided that .
1).u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. relevant references are Grubel [179]. Here are some of the main approaches: Laplace transform inversion Often.b(u)du . [s.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. T). 191). A notable fact ( see again XI. but are somewhat out of the mainstream of the area . say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. 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. . However.4. We don't discuss Laplace transform inversion much. O(u. the formulas ( IV.S(u) 1S(oo) f °D exp {. it is easier to find the Laplace transforms = f e8 . Embrechts. see VIII. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. T) can then be calculated numerically by some method for transform inversion. f {eXp U LX 2. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI . Abate & Whitt [2]. • An astable Levy process with drift . a2 (x): Ip (u) = where S(u) = f °O exp {. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). (u.f f 2µ(y)/a2(y) dy} dx  (4. Ab(u). For the finite horizon ruin probability 0(u. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. 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 .2) is the natural scale.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. esuTb( u.Lef$er function. T) themselves.7.ff 2µ(y)/a2(y) dy} dx . Given this can be done.
and in quite a few cases (Chapter VIII). However. 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. see VIII. T) as the solution to a differential. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. and carry out the solution by some standard numerical method.4) 00['Y]1)'Y = 0. In the compound Poisson model with p = 1. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m.7.3. U is explicit in terms of the model parameters. and in particular the naive idea of conditioning upon process behaviour in [0. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. 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.and integral equations The idea is here to express 'O(u) or '(u. One example where this is feasible is the renewal equation for tl'(u) (Corollary III.g.f.p)/(13B'[ry] . which can equivalently be written as f3 [7] = 1 +13 .Ce"u. it states that i/i(u) . 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization.or integral equation.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). u * oo. Differential. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). .) B[s]. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole).3) where C = (1 .1) and y > 0 is the solution of the Lundberg equation (4. dt] most often leads to equations involving both differential and integral terms. (4.3) in the compound Poisson model which is an integral equation of Volterra type.
u > oo. (4. other approaches are thus required. For example.4. . It has generalizations to the models with renewal arrivals. Large claims approximations In order for the CramerLundberg approximation to be valid. often for all u > 0 and not just for large u. a Markovian environment or periodically varying parameters. T) for large u are available in most of the models we discuss. in some cases the results are even more complete than for light tails. but typically not very precise in their first naive implementation.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. Approximations for O(u) as well as for 1(u. See Chapter IX. In the case of heavytailed distributions. In particular. incorporating correction terms may change the picture dramatically. T). in such cases the evaluation of C is more cumbersome. However. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. In fact. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models.7 and IV. J B dx. However. when the claim size distribution is of phasetype.2.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. the claim size distribution should have an exponentially decreasing tail B(x). some further possibilities are surveyed in 111 . for the compound Poisson model ^(u) p pu In fact . 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. corrected diffusion approximations (see IV. Diffusion approximations are easy to calculate. This list of approximations does by no means exhaust the topic. 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.
. the difficulty comes in when drawing inference about the ruin probabilities.d. This is proved for the compound Poisson model in 111. one expects a model with a deterministic claim size distribution B. though not too many precise mathematical results have been obtained. However.. in the compound Poisson model. lower bounds etc. This procedure in itself is fairly straightforward. it has the advantage of not involving approximations and also. In practice. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. How do we produce a confidence interval? And.U(k)) i =k+ i .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.. of being somewhat easier to generalize beyond the compound Poisson setting. it is a general principle that adding random variation to a model increases the risk.g. . .. they have however to be estimated from data. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. 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.) at various places and in various settings. empirical evidence shows that the general principle holds in a broad variety of settings. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. . say degenerate at m.i.3). 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.8. However. which is a standard statistical problem since the claim sizes Ui. . e.18 CHAPTER I. When comparing different risk models. given NT. as a general rule. and to plot the empirical mean residual life 1 N . obtained say by observing the risk process in [0. more importantly. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. fitting a parametric model to U1. INTRODUCTION Compared to the CramerLundberg approximation (4. UNT may be viewed as an interpolation in or smoothing of the histogram). The standard suggestion is to observe that the mean residual life E[U . T]. . can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. UNT are i. For example.k (U(`) . For example.
and of course the method is relevant in risk theory as well.3) or Section 3 of Chapter VI are referred to as Proposition VI. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. The infinite horizon case presents a difficulty.2. We look at a variety of such methods in Chapter X. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget.d. See further Embrechts. CONVENTIONS 19 as function of U(k).5. Still. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. formula VI. < U(N) are the order statistics based upon N i. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. reference [14]. .. in all other chapters than VI where we just write . to observe whether one or the other limiting behaviour is apparent in the tail. in this book referred to as [APQ]). The chapter number is specified only when it is not the current one. (or a functional of the expectation of a set of r. However. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987.(5. where U(1) < .. Truncation to a finite horizon has been used.3 (or just VI.v.. claims U1. For example. Klnppelberg & Mikosch [134]. because it appears to require an infinitely long simulation. UN. this is a straightforward way to estimate finite horizon ruin probabilities. .2. Thus Proposition 4.4. A main problem is that ruin is typically a rare event (i.3) and Section VI.. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. respectively.e.3).. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. and in fact methods from that area can often be used in risk theory as well . and look at them to see whether they exhibit the expected behaviour or some surprises come up. but is not very satisfying.v's) which can be generated by simulation. formula (5.i.
h + 0.h. B(dy).d.g. oo). EX2/(EX)2. cumulative distribution function P(X < x) c.f.f.h.d. E. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c.g.4. left hand side (of equation) m. The Laplace transform is b[s]. random variable s.B(x) = P(X > x) of B.o. If.g.ceax. w.20 CHAPTER L INTRODUCTION Proposition 4.f. say a heuristic approxi1 + h + h2/2.f. E expectation.e. References like Proposition A. r.v. see under b[s] below. cumulant generating function.i. Abbreviations c.s. . independent identically distributed i.f. right hand side (of equation) r. . (moment generating function) fm e82B(dx) of the distribution B.v.r. formula (5. B(x) the tail 1 .p.g.c. n! 27r nn+1/2en. for a probability distribution IIGII = 1. B(x) = P(X < x) = fx. mation.f. with probability Mathematical notation P probability.s. IIGII the total mass (variation ) of a (signed ) measure G . In particular.d. i. infinitely often l.g.29) refer to the Appendix.3) or Section 3. squared coefficient of variation. if B(x) .t. or a more precise one like eh . B[s] the m. and for a defective probability distribution IIGII < 1. n i oo.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit. then for 1s < 5). log E[s] where b[s] is the m. with respect to w.2. (A. B is concentrated on [0. A different type of asymptotics: less precise. i. moment generating function. as for typical claim size distributions.
Thus. E[X. Unless otherwise stated. 7r. 21 D [0. 0 marks the end of a proof. 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. R(s) the real part of a complex number s.e. 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. 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. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). the processes we consider are piecewise continuous. . the value just before t. In the Frenchinspired literature. of numbers. intensity interpretation. a2) the normal distribution with mean p and variance oa2.. Xt_ the left limit limstt X8f i. A. a. I(A) the indicator function of the event A.e. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. Matrices and vectors are denoted by bold letters. N(it.A] means E[XI(A)]. i. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. . .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). F o r a given set x1. Usually. matrices have uppercase Roman or Greek letters like T. row vectors have lowercase Greek letters like a. the ith unit row vector is e'i. Usually. and column vectors have lowercase Roman letters like t. i.5. the ith entry is 1 and all other 0. (the dimension is usually clear from the context and left unspecified in the notation). xa. an example or a remark. all stochastic processes considered in this book are assumed to have sample paths in this space. though slightly more complicated.. only have finitely many jumps in each finite interval.
FL. cf. VI. though slightly more complicated. I. or quantities with a similar time average interpretation. e. 'q the safety loading . p the net amount /3pB of claims per unit time.22 CHAPTER L INTRODUCTION B the claim size distribution. ry The adjustment coefficient. 111. Notation like BE and B(t) in Chapter VI has a similar.1.5. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation.1.g. . interpretation. I. cf.5. cf. J the rate parameter of B for the exponential case B(x) = eby.
5) are. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. however. in most cases via likelihood ratio arguments. All results are proved elsewhere . Sections 4. somewhat more advanced than in the rest of the book. The general theory is. Due to the generality of the theory. Sections 4. used in Chapter VI on risk processes in a Markovian (or periodic) environment. The reader should therefore observe that it is possible to skip most of the chapter. More precisely. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. When encountered for the first time in connection with the compound Poisson model in Chapter III. 23 . a parallel selfcontained treatment is given of the facts needed there. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. in particular at a first reading of the book. the level of the exposition is. strictly speaking. not crucial for the rest of the book. The duality results in Section 3 (and. in part. however.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. The likelihood ratio approach in Section 2 is basic for most of the models under study. however. The topic is.
0. and the ruin probabilities are ip(u) = P (T(u) < oo). (b) St a$ oo on {T(u) = oo}. 1 Martingales We consider the claim surplus process {St} of a general risk process. StUit.(u).2) takes the form 1 = E [e'ys(). claim size distribution B and p = .QµB < 1. Proposition 1. T(u) < oo] + 0 = eryuE [e7Vu). The more general Theorem 6.5 can be skipped. T(u) > T] .1 Assume that (a) for some ry > 0. As usual. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). T(u) < T] + E [eryST . f1 . however.24 CHAPTER II. V) (u. Example 1 . Let e(u) = ST(u) . using r(u) A T invokes no problems because r(u) A T is bounded by T).)AT = E [e7ST(°).T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. 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..2 Consider the compound Poisson model with Poisson arrival rate . {e'YS° }t>0 is a martingale. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6.1 is basic for the study of the compound Poisson model in Chapter III. the time to ruin r(u) is inf It > 0 : St > u}. We get 1 = Ee7So = E e'Y S(. T) = P(T(u) < T). and in the limit (1. (1.u denote the overshoot. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem .2) As T > oo. Thus N.
r" where y = S . From this it is readily seen (see III. the conditions of Proposition 1. condition (a) of Proposition 1. u Corollary 1.4 (LUNDBERG ' S INEQUALITY ) tion 1 . Example 1 . O(u ) < e7".f.1) shows that Eels. Now at the time r(u) of ruin {St} upcrosses level u by making a jump .Q.Ft = a(S" : v < t).Q and the U.= e"(') where K(a) = . Thus. u Corollary 1. Then {St } is Brownian motion with variance constant o2 and drift p < 0..u + x is again just exponential with rate S. and (b) follows from p < 1 and the law of large numbers (see Proposition III.1 is satisfied. and thus by the memoryless property of the exponential distribution . B(x) _ edx.ap. MARTINGALES 25 where {Nt} is a Poisson process with rate . Since {St} has stationary independent increments.1. the ruin probability is O(u) = pe./3. 1. By standard formulas for the m. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx .1. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.d.a = a . Since {St} has stationary independent increments.3/6 < 1. and thus Ee7s° = 1. of the normal distribution.3 Assume that {Rt} is Brownian motion with variance constant o.2(c)). with common distribution B (and independent of {Nt}).g.1) .i. the conditional distribution of the overshoot e(u) = U .(„)_ = x is that of a claim size U given U > u .6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists. Under the conditions of Proposi Proof Just note that C(u) > 0.1 are satisfied.2 and drift p > 0. Thus. Proof Since c(a) = /3 (B[a] . and p =.Q(B[a] .1) . are i.1.a. and thus Ee'rs° = 1.x. the martingale property now follows just as in Example 1. The available information on this jump is that the distribution given r(u) = t and S. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .a it is immediately seen that y = S .2.5 For the compound Poisson model with B exponential. A simple calculation (see Proposition III. Eeas° = e"(°) where n(a) = a2a2/2 .
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. I. Two such processes may be represented by probability measures F. But if a $ ^ . The number Nt F) of jumps > e before time t is a (measurable) r. Embrechts. B. the parameters of the two processes can be reconstructed from a single infinite path.Ft}too and the Borel afield F. (2. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. Proof Just note that ^(u) = 0 by continuity of Brownian motion. then S and S are disjoint . oo). Grandell & Schmidli [131]. P on (DE. then z/'(u) = e7" where 'y = 21A/a2. and is further exploited in his book [157]. and F.F). More recent references are Dassios & Embrechts [98]..26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1.1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. Example 2 .6 N S = { lim Nt I t +00 t gJ are both in F. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. which we equip with the natural filtration {. . Delbaen & Haezendonck [103] and Schmidli [320]. 0 and claim size distributions B. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. cf. Grandell [171]. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random.3 below). [172].6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. on (DE. F). A]. A somewhat similar u argument gives singularity when B $ B. A E Ft.e. 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(S) = P(S) = 1. P are then singular (concentrated on two disjoint measurable sets).v. Thus the sets S = I tlim +oot t =.1 Let F. as shown by the following example this setup is too restrictive: typically'. and by the law of large numbers for the Poisson process . hence so is Nt = limfyo N2`i. Theorem 2. However.
if for some probability measure P and some {. we have E [ LTIFT]1 = LT on {T < T}. then { 1 P(G) = EG . Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE.r. A].1) holds. P be as in Proposition 2. This proves (i).A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A).Y) such that P(A) = Pt(A).. A]. define P by Pt (A) = E[Lt. Conversely. Then Ft (A) = E[Lt. (ii) Conversely.Tt) is absolutely continuous w.F).r.3 Let {Lt}. P) such that LLt = 1.2. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. Proposition 2.1) and nonnegativity by letting A = {Lt < 0}.F such that (2. Proof Under the assumptions of (i).r. G C {T < oo}. A E F8. using the martingale property in the fourth step. the restriction of P to (DE. The truth of this for all A E Y.Pt)) The following result gives the connection to martingales. A E Ft . LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i.2(i). under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. . F the Borel o•field and P a given probability measure on (DE. By the martingale property.t.1) holds. ({.Ft}. _. G ] = E [LT . then {Lt} is a nonnegative martingale w.F8] = L8 and the martingale property. Lt < 0] can only be nonnegative if P(A) = 0. ELt = 1 follows by taking A = DE in (2. . implies that E[LtI.e.Pt}adapted process {Lt}t>o (2. If r is a stopping time and G E PT. ({Ft} . Lets < t.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. F) such that ELt = 1. Finally. A E F. Then P(A) = E[Lt. (i) If {Lt}t> o is a nonnegative martingale w. 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 . A] = E[Lt.t. Ft). then there exists a unique probability measure Pon . that the restriction of P to (DE.2 Let {Ft}t>o be the natural filtration on DE. Hence E [_ . G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G).t. 1 J (2.
St).1) in Proposition 1. Consider a (timehomogeneous) Markov process {Xt} with state space E. .2) follows by monotone convergence.1). Xt = St.4) Proof Letting G = {r(u) < oo}. In the context of ruin probabilities.Gn {r <T} .O(u) = eryuE[e'YC(u). u From Theorem 2. Now just rewrite the r.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian.t.1. (2.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . of (2.. Rt) or Xt = (Jt. To this end. 5) for processes with some randomwalklike structure. Lr(u) 11 The advantage of (2. {St} = {u . SOME GENERAL TOOLS AND RESULTS In the general case .r. one would typically have Xt = Rt.r.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). T(u) < oo]. first in the Markov case and next (Sections 4. is Markov w. For the definition.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1. A change of measure is performed by finding a process {Lt} which is a martingale w.28 CHAPTER II. the natural filtration {.t. both sides are increasing in T. 1 Since everything is non negative. and this problem will now be studied. in continuous time (the discrete time case is parallel but slightly simpler).s.4 Under condition (a) of Proposition 1. is nonnegative and has Ey Lt = 1 for all x. (2. each F.h.4) compared to (1. First we ask when the Markov property is preserved. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. The crucial step is to obtain information on the process evolving according to F. and letting T * oo.1: Corollary 2. say. Xt = (Jt. t.2) by noting that 1 = ersr(„) = e1'ue7Ou). we have F(G) = V )(u). r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. applying (2.Ft} . we need the concept of a multiplicative functional. we assume for simplicity that {Xt} has Dpaths.. where {Rt} is the risk reserve process.
oo).v. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].7). (2.Ft }. for all x.v.(Xtitl) with all t(i) < t + s.v.Ft].YB] for any Ftmeasurable r. or.s. o 9t = V. o 9tI.8) which is the same as (2. (2. Similarly. Indeed. t] * [0. The converse follows since the class of r.7).6) implies (2. A]. Y8. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I. s. The precise meaning of this is the following: being . nonnegative and Lt+8 = Lt•(Lso9t) (2. Zt.2.r. Vt+e. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx.[Y.7) for any Ftmeasurable Zt and any .'s of the form Zt • (Y8 o 0t) comprises all r. and then L. where Ot is the shift operator. 0 .F8measurable r..Ft+8measurable. the natural filtration {Ft} on DE. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x. ({Xt+u} 0<u<8) Theorem 2.5) are Tt+e measurable. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2.T9measurable Y8. the Markov property can be written E.(A) = Ex [Lt. t. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.6) for any . since Zt • (Y8 o Ot ) is .Ftmeasurable.t.v.'s of the form fl' f..5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any . (2. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.5) Pxa. By definition of Px.Pt+8measurable r.v. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]]. t and let Px be the probability measure given by t.5 Let {Xt} be Markov w. (2.Y8f t < s. Proof Since both sides of (2.Ft] = Ex.
In between jumps. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. Ro = u (say). . (3. {Vt} . (using the Markov property in the second step) so that the martingale property is automatic... aN where or* = T UN_k+l. t. More precisely . R = p(R)). .5 can be found in Kuchler & Sorensen [240]. 3 Duality with other applied probability models In this section. reflection at zero and initiar condition Vo = 0. t] = LtExt L8 = Lt. The storage process {Vt }o<t<T is essentially defined by time reversion. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. ..6 For {u .. the premium rate is p(r) when the reserve is r (i. the arrival epochs are Qi.. } E[Lt+B I. We work on a finite time interval [0.1) The initial condition is arbitrary. Thus R = Ro + f p(R8) ds .e. then Remark 2. ... The result is a sample path relation.. CN. see Dynkin [128] and Kunita [239].. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. The corresponding claim sizes are Ul.30 CHAPTER H. UN. The formulation has applications to virtually all the risk models studied in this book. with a proof somewhat different from the present one. 0 < vl < . and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. A more elementary version along the lines of Theorem 2. In between jumps. Indeed.. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or.Ft] = LtE[L8 o 9t I.. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. < aN < T.At where At = k: vk <t U. .
DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.1..e.___ .T) = P(VT > u).2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. .x. The sample path relation between these two processes is illustrated in Fig. In particular. V = p(V)).. Then rt°) > rt°) for all t when u > v.1 The events {T(u) < T} and {VT > u} coincide.1) we have Vt = At  f P(Vs)ds where A= U= AT .T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability.. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous. Theorem 3.3. :..____•_.. 3..__.AT_t.11 4. instead of (3._. (3.__..1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u.. That is.. {Vt} remains at 0 until the next arrival). V)(u._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u. (3.
U1 = Rol. and since ruin can only occur at the times of claims. .T l . Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].3 and being i. with distribution B. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo.2 Consider the compound Poisson risk model with a general premium rule p(r).1 with Ro = u = ul). Resnick & Tweedie [79].2 from Harrison & Resnick [188]. and a general premium rule p(r) when the reserve is r. Hence if n satisfies VVN_n+1 = 0 (such a n exists. say V. and so on.1 and its proof is from Asmussen & Schock Petersen [50]. this represents a model for storage. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. The arrival epochs correspond to rainfalls. = r(VT) . see Siegmund [344].U1 > roil . Corollary 3. Then Vo. we have r(u) > T. we can repeat the argument and get VoN_1 > Ra2 and so on. Nevertheless. the distinction between right. Historically. Theorem 3. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. We get: Corollary 3. if nothing else n = N). u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . we have RQ„ < 0 so that indeed r(u) < T. Then similarly VVN = r0. Historically. Then the storage process {Vt} has a proper limit in distribution.3).Ul < roil  Ul = RQ„ Va1V_1 < RQ2. one may feel that the interaction between the different areas has been surprisingly limited even up to today. If VaN > 0.32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. Proof Let T ^ oo in (3. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. 3. say of water in a dam though other interpretations like the amount of goods stored are also possible. Hence RQ„ > 0 for all n < N.1 with Ro = u = u2).i. 3.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero).d. and then '0 (u) = P(V > u). The results can be viewed as special cases of Siegmund duality. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. if and only if O(u) < 1 for all u. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig.
1)).. is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4. 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. with common distribution F (say).1.d. .. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN.. n=0.i. Let further N be fixed and let Wo. Most often... has a proper limit W in distribution as n + oo. . ...1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}. can be viewed as the reflected version of the random walk with increments Z1. . 0 Corollary 4..N From this the result immediately follows. WN be the Lindley process generated by Z1 = YN... generated by Z1.. Z2..Yl min (YN . Z2 = Y2. Xo = 0.2)..d...w. Z2 = YN_1 i .1... ZN = .4. 0)....... Z2. .s. N min (Y1 + • • • + YNn) n=0. For discrete time random walks .... i.YNn+1) n=0. . hits (oo.Y1 according to Wo = 0.min n=0.e. 1} is often referred to as simple random walk or Bernoulli random walk). (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo.. W1. W2... where the Yi are i .h.. I. In particular.. WN = YN . evolves as a random walk with increments Z1i Z2.. R valued sequence Z1...1.. (c) The Lindley process {WN} generated by Zl = Y1. N min (Y1 + + Yn)...1... Proof By (4..1. = Xo + Y1 + • • • + Y.1 in terms of Lindley processes .. . just verify that the r .2) satisfies the same recursion as in (4.. the Lindley process Wo. . For a given i.1) Thus {Wn}n=o.. Z2 .. {Wn}n=0. as long as the random walk only takes nonnegative values.N (4..1. Here F is a general probability distribution on R (the special case of F being concentrated on {1.. W1. Theorem 4. and is reset to 0 once the r... Then the events {r(u) < N} and {WN > u} coincide. .2) (for a rigorous proof.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0. of (4. .. there is an analogue of Theorem 3..
One then assumes Yn to be a stationary sequence..N so that WN _P4 M = supra=0. . e.) and defines Zn = Yn.1 have the same distribution for n = 0. YN in Theorem 4. . In general. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ).. In that case .. .3 The i. there is a more general version of Corollary 4..the result is a sample path relation as is Theorem 3. . ZN or..2. Next consider change of measure via likelihood ratios. + Z. By Kolmogorov's 01 law.1. W v m and P(W > u) = P (m > u) = 0(u). doubly u infinite (n'= 0.l. (d) #.i.34 CHAPTER II. (e)Yi+•••+Yn 74 .. Y1) has the same distribution as (Y1. (Yi + • • • + Yn) > oo a...s. SOME GENERAL TOOLS AND RESULTS (d) m = inf... N.ooa. or M < oo a.l. w.s...1. equivalently. .s. 176) but appears to be rather intractable. Similarly. ±2.. assumption on the Z1... .s . .g.5 does not necessarily lead to a random walk: if. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo..g... ...d. a Markovian change of measure as in Theorem 2. For a random walk.1. either M = oo a. Clearly. Thus the assertion of Theorem 4. the Y1..1. 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.o.1 is actually not necessary .. (e). Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists. a sufficient condition for (e) is that EY is welldefined and > 0. Proof Since (YN. Remark 4 . ... ±1.. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . the Lindley processes in Corollary 4. then the restrictions of Fx.1 is equivalent to WN D MN = (Z1 + . .) sup n=0. 0 By the law of large numbers.. YN).2 and Theorem 4.=o. . Combining these facts gives easily the equivalence of (a)(d). the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.
Y2) = h(1'i)h(I'a). the changed increment distribution is F(x ) = E[h(Y). (4. Conversely. cf. this means E(g(x. Y) = h(Y ) a.3) holds.f.s.. Y) f (Y)] for all f and x..nrc(a )} (4.g.4). Proof If (4. Y < x].5 corresponds to a new random walk if and only if Ln = h(Y1) . The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .s. Since L1 has the form g (Xo. Y1). In that case.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.. .. the random walk property implies Ex f (Y1) = Eo f (Y1 ). of F). RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4.. We get: Corollary 4.. h(Yn) (4.3) 1Pxa.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. 100 ). In particular.4) ({Ln} is the familiar Wald martingale . = 1 for all n and x. 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. and so onforn =3. implying g (x. Y ) f (Y)] = E[g(O. For n = 2.. where h (y) = g(0. Breiman [78] p.4.g. (a) = log F(a] is the c. and define Ln by (4.4 Let {Ln} be a multiplicative functional of a random walk with E_L.4. for some function h with Eh(Y) = 1. u A particular important example is exponential change of measure (h(y) = e°y'(") where r. e. Then the change of measure in Theorem 2.3) holds for n = 1. y ). Then the change of measure in Theorem 2.
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.. they arise as models for the reserve or claim surplus at a discrete sequence of instants. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e .6) More precisely. i. the tradition in the area is to use continuous time models. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}.5) Note that the structure of such a process admits a complete description.36 CHAPTER II. this description needs some amendments. the claim surplus process for the compound Poisson risk model . Roughly. (4. However. In discrete time. v2 = 0 and v = 3B). the pure jump process is given by its Levy measure v(dx). < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . In continuous time. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . however. or imbedded into continuous time processes . (4. ..Ft] = Eof (X. corresponds to a process with stationary independent increments and u = p. The simplest case is 3 = JhvMM < oo.Xt)I. but we omit the details ). Xt(2)_t(l).1).7) for all e > 0. {Xt} is a random walk. .Xn < x). The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). In risk theory. given by a the increment distribution F(x) = P(Xn+l . with premium rate p. we are .).t(i . {Xt} can be written as the independent sum of a pure drift {pt}. see Chapter V. say the beginning of each month or year . Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . 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).e. Xt =Xo+pt+oBt+Mt. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). e f x:IxJ>e} v(dx) < oo (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 .
defined as a system with a single server working at a unit rate. [The condition for V < oo a. T) = P(VT > u).min Xt (4. oo].6 In the compound Poisson risk model with constant premium rate p(r) .6). 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. Corollary 4. is easily seen to be f3pB < 1. and the reflected version is then defined by means of the abstract reflection operator as in (4.e.4.s.3 and decreases linearly at rate 1 in between jumps. 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). then Ee'(xtxo) = Eoeaxt = etx(a).1. VT + V for some r. jxJ v(dx) < oo.1)v(dx) (4.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. having Poisson arrivals with rate . Chapter III. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. 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. Then the storage process {Vt} has constant release rate 1.10) . has upwards jumps governed by B at the epochs of a Poisson process with rate . 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. (ex . WT = XT . cf. i.2).7 If {Xt} has stationary independent increments as in (4. 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. Now consider reflected versions of processes with stationary independent increments. O(u. V E [0. and b(u) = P(V > u). where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f".8) O<t<T (assuming Wo = Xo = 0 for simplicity). Proposition 4.v.Q and distribution B of the service times of the arriving customers. Furthermore.
Xt +B . use the representation as limit of compound Poisson processes.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments.. t. (4.6) are µ = µ + Oo2 .11) (eax . Q2 = v2. we show in the compound Poisson case ( IlvIl < oo) in Proposition III. Proof For the first statement . of an infinitely divisible distribution (see. let e" (a ) = Eoeaxl.1. if Lt = e9(xt .5) and get E [f(Xt+B .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 .10) is the LevyKhinchine representation of the c.Xt)g(s.1 ) v(dx) .38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution.g.f.Xt)L8 o 0tIFt] = E [f (Xt+s . then the changed parameters in the representation (4.1)eexv(dx).g. Chung [86]). e. 5 has stationary independent increments as well. . Xt Xo) with E2Lt = 1 for all x.1 that E eaMt = exp fmoo In the general case . By explicit calculation . Then the Markov process given by Theorem 2. 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) .Xt)I'Ftl = E [f(Xt+B . Eea(µt + QBt) = et{aµ +a2OZ/2}. 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t. In particular. This is of course no coincidence since the distribution of Xl . u Note that (4. X8) = Eof (X8)L8 = Eof (X8)• For the second. we use the characterization (4.xo)tk ( e). Theorem 4 .Xt)I Ftl = Eof (X8)g(s. v(dx) = e9xv (dx).
dB/dB = b/b when B. the corresponding claim sizes . 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 .0 in the following. Then we can write v(dx) _ /3eOxB(dx) = / (dx). Thus (since µ = p = 1. Example 4 .10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate . then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . As for processes with stationary independent increments . B have densities b. is defined as a bivariate Markov process {Xt} = {(Jt.3 =. the structure of MAP's is completely understood when E is finite: 2and only there . it is then easily seen that Lt = H dB(Ui) i:o.11 For an example of a likelihood ratio not covered by Theorem 4.5. b with b(x) > 0 for all x such that b(x) > 0). where . .4). a = 0. Example 4 .1) For shorthand .Ft] = Ejt.(3 = . B(dx) = B[9] B(dx). U2. v(dx) _ .l3 and claim u size distribution B. MAP stands for the Markovian arrival process discussed below.3 and claim size distribution B. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. .8. <t whenever the RadonNikodym derivative dB/dB exists (e..o[f (S8)g(J8)]. corresponding to p = 1.. abbreviated as MAP in this section2. MARKOV ADDITIVE PROCESSES 39 Remark 4..g.3 and claim size distribution B # B. (5.St)g(Jt+s)I. Ei instead of P2. 0. one reason is that in parts of the applied probability literature. Ei. u 5 Markov additive processes A Markov additive processes..2. Recalling that U1..(3B(dx).3B[B].0.9 If X0 = 0. are the arrival times and U1. we write Pi. 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.
{Jt} is specified by its intensity matrix A = (Aij)i.jEE• On an interval [t. Y1 E dx) where Y„ = S„ .it = A. with the Y„ being interpreted as interarrival times. An alternative description is in terms of the transition matrix P = (piA.) If E is infinite a MAP may be much more complicated.. by generating Yn according to Hij when J„_1 = i. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time. (That a process with this description is a MAP is obvious. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1. 1 J1 ='^])iJEE = (Fij[a])i .. As an example. As a generalization of the m.6) depending on i.. t+s) where Jt .g. vi(dx) in (4. Jn = j. Fn[a] = F[a]n where P[a] = P . which we omit and refer to Neveu [272] or cinlar [87]. In continuous time (assuming Dpaths).jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. the converse requires a proof.i.o(Ji = j. Then a Markov additive process can be defined by letting t St = lim 1 I(IJB1 < e)ds E1o 2d o be the local time at 0 up to time t. .[a) = (Ei[easl. v. If all Fij are concentrated on (0. let {Jt} be standard Brownian motion on the line.1 For a MAP in discrete time and with E finite. consider the matrix Ft [a] with ijth element least Ei . {St} evolves like a process with stationary independent increments and the parameters pi. Proposition 5. a MAP is the same as a semiMarkov or Markov renewal process. In addition.40 CHAPTER H. oo).Sr_1.f.. SOME GENERAL TOOLS AND RESULTS In discrete time.J1=j)= Fij (dx) Pij In simulation language.. Y2.9 EE = (iii&ij[a])i j EE . the distribution of which has some distribution Bij.
Sn ) yields Ei[easn+ '. Then. kEE Jn = k]Ek[e"Y" . Jt = j] is given by etK[a]. where K[a] = A+ (r. \ diag Ft[a] = Ft[a]K.2 Let E be finite and consider a continuous time Markov additive process with parameters A. pi.4c). MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. 00 r(i) (a) = api + a2ot /2 + f (e° . a= . we infer that in the discrete time case the .1 )v(dx). assume that the Markov chain/process {Jt} is ergodic. Jt = j] Ejesh'^ + E Ak j hEi [ease . aSt h = (1 + Ajjh) Ei [east . Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . 013 . qij.(')(a)) diag + (). vi(dx) (i E E). Jt = k] { xk kEE j la] . this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. Then the matrix Pt[a] with ijth element Ei [east. Jn+1 = A] = 41 Ei[ e 5„. u Proposition 5. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. Jt = k] { 1 . In matrix formulation .ijgij(Bij[a] . j E E) and So = 0. By PerronFrobenius theory (see A. vi(dx). up to o( h) terms. u In the following.qkj + k?^j qkj Bkj [a] } = Ei [east.5.1)) .1) } (recall that qjj = 0). J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. Bij (i. Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east .
c(a) (and h(")). Corollary 5.4. 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. of a random walk.tK(a)h(a) J jj it L o is a martingale. The corresponding left and right eigenvectors v("). Then h(°) = e.h(a)vva)etw(a).etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l .7. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). Jeast. and we shall take V(a)h(a) = 1. Furthermore. The function ic(a) plays in many respects the same role as the cumulant g.t.5 EiSt = tK'(0) + ki . Yrh(a ) = 1. a. Proof For the first assertion. we are free to impose two normalizations. cf. Jt = j] .Eikjt = ttc'(0) + ki . h(") are only given up to a constants. . We also get an analogue of the Wald martingale for random walks: Proposition 5. its derivatives are 'asymptotic cumulants'. Since v("). h(") may be chosen with strictly positive components. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . (5. Proof By PerronFrobenius theory (see A. as will be seen from the following results.4c). u Let k(a) denote the derivative of h() w.3 Ei [east. Corollary 5.f. cf.4 Eie"sth(a) = h=a)et?(").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). and write k = k(°).2) where 7r = v(°) is the stationary distribution. In particular. Eie"sth^a) = e'Pt[a]h( a) = e. Corollary 5.42 CHAPTER II. Proposition 5.e=e°tk.r.
4) . subtraction yields Vary St = tic"(0) + O(1). Remark 5 . [E. tam E tSt a (0). t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. (5.7 No matter the initial distribution v of Jo.") }) . u The argument is slightly heuristic (e. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) .g.4.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1.6 For any stopping time T with finite mean. we differentiate (5. Squaring in Corollary 5. (5.a) + ttc (a)2hia ) Multiplying by v=.5 yields + W (a)k. E=ST = tc'(0)E7.Eikjr .+ k.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk . Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.5. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x).2ttc (0)Evkjt + 0(i). for a random walk: Corollary 5. t a oo. summing and letting a = 0 yields E„ [St + 2Stkj. . the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions. Corollary 5. In the same way. MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. the distribution of Jo).. 8 Also for E being infinite (possibly uncountable ).. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). one obtains a generalization of Wald's identity EST = Er • ES. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . More precisely.e. For the second .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.5.
f (x) tyo t provided the limit exists.(9) {Lt}t>o = . 0) = n(a) h(i)..4 that { h(Jt) easttK(a) L o (5. however. G is defined as Gf (x) = lim Exf (Xt) .5) is a martingale .3b and Remark VI. see.6.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. An example beyond the finite case occurs for periodic risk processes in VI. h(Jo) Lo is a Px martingale for each x E E. St) } as follows.6. Jt = (s+t) mod 1 P8a. gha(i. First.for the present purposes. where {Jt} is deterministic period motion on E = [0.5 defines a new MAP. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i). however.10 Let {(Jt.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. xEE . (5. s) = ea8h(i).e. we take the martingale property as our basic condition below (though this is automatic in the finite case). in particular that f is bounded. Usually. Remark 5.5. In view of this discussion . let ha(i. this leads to h(i) + tcha( i. u forsEE). inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. this is. Then {Lt } is a multiplicative functional. For t small . From (5.s. Given a function h on E. V.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 5.6) We shall not exploit this approach systematically.. 1) (i. St)} be a MAP and let 0 be such that h(Jt) OStt. 0) = h(i )( 1 + ttc(a)). 0 Proposition 5.9 The condition that (5. some extra conditions are imposed. and the family {f LEE given by Theorem 2.
10 is given by P = eK(e) Oh e) F[e]Oh('). this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. We omit the details. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .1) . In the infinite case . Ai = µi + 0Q.11 below in the finite case. 0 < qij < 1 and Bij [0] > 0. Bi(dx) = Bi(dx). 0<q<1.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j.. if vi(dx) is compound Poisson.c(0)e = tc(0)e . .5. 0<b<oo. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). Then the MAP in Proposition 5. Bi [0] Remark 5. in the discrete time case.tc(0)e = 0 . Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. That 0 < qij < 1 follows from the inequality qb <1.(0)j. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0]. Bi.1) eft ea' f ij (dx) = Hij (dx) Hij [0] . MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . u Theorem 5.. qij = r.Qi < oo and Bi a probability measure. one can directly verify that (5.11 Consider the irreducible case with E finite. vi (dx) = f3 Bi(dx) with .ic(0)e = ic(0)Oh e) h(e) . 1 + q(b .St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. In particular.7(dx) Bij [0] Bij(dx) in the continuous time case .7) In particular. (5.1) holds for the P. ^i = of qij Bij [0] 1 + qij ( Bij [0] .
8. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . Letting a = 0 yields the stated expression for A.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . this implies k[a] = A 1 ) (K[a + 0] . Jl = j) = Ei[Lt. since Hij.8). Jl = j] :(Yi E dx. Jt = j] = hie) . .tc(0)I.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] . Here the stated formula for P follows immediately by letting t = 1. F:j with a density proportional to eei . it follows that indeed the normalizing constant is H1 [0]. in continuous time ( 5. (dx) of a process with stationary independent increments follows from Theorem 4.11.13) for matrixexponentials .46 CHAPTER II. v= .8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) ..e) Consider first the discrete time case .. a = 0 in (5. In matrix notation . this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. . (dx). Hence the same is true for H=j and H. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. This shows that F. Jt = A..Bay [0]) That k(') (a + 0) . Yi E dx. v.tc (') (0) corresponds to the stated parameters µ.8) h(. Similarly. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. Further Fib (dx ) = P=(YI E dx. Now we can write K[a] =A+A ) ( K[a + 0] .r. is absolutely continuous w.t. are probability measures . H1.
..7). 0]. i. 7+ < oo). hardly a single comprehensive treatment. oo). For the Wald identity in Corollary 5. however. [226] and Miller [260].3 for an infinite E are given by Ney & Nummelin [266].1) = Aij4ij(Bij[a] . Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. 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).+ < x. [261]. the literature on the continuous time case tends more to deal with special cases. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S. Note that G+ is concentrated on (0. which. Though the literature on MAP's is extensive. h. Conditions for analogues of Corollary 5. 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]. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. an extensive bibliography on aspects of the theory can be found in Asmussen [16].)Ajjgij(Bij[a+0] .Bij[0]) = hjel)ijgijBij[0](Bij[a] .6.1). there is. [262] in discrete time.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 ..e. and is typically defective. however. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. is slightly less general than the present setting. [225]. < x) = 11 (S. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. see also Fuh & Lai [149] and Moustakides [264]. has no mass on (oo.6.
Thus. 1 For the compound Poisson model with p = 01LB < 1. In other cases like the Markovian environment model.. The first ladder step is precisely ST+.ST+(1) and so on. The main result of this section is Theorem 6. On Fig. and the maximum M is the total height of the ladder. To illustrate the ideas. 0 f T+ (6.B(x) denotes the tail of B.e. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. at present we concentrate on the first ladder height. they have a semiMarkov structure (but in complete generality. o 00 (6. Also. it follows that for g > 0 measurable.2) . we shall first consider the compound Poisson model in the notation of Example 1. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. the sum of all the ladder steps (if rl > 0. there are only finitely many).(3B(x ) = pbo(x) on (0.2. the second ladder height (step) is ST+(2) ..5 below. 6. = ST+(1) Figure 6. oo). G+ is given by the defective density g + (x) =. For the proof of Theorem 6. In simple cases like the compound Poisson model.1.d. 0].1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. In any case. i. where basically only stationarity is assumed. the ladder heights are i.48 CHAPTER K.T+ > t)dt = E f 0T+I(St E A) dt. a fact which turns out to be extremely useful. Here bo(x) _ B(x)/µB. i. which gives an explicit expression for G+ in a very general setting. R+ is concentrated on (oo. 6.i. Recall that B(x) = 1 . Theorem 6 .1. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. has no mass on ( 0.00 ). by approximation with step functions . see Fig.1.e. the dependence structure seems too complicated to be useful). g(y)R+(dy) = E f g(St)dt.
St<0.2 R+ is the restriction of Lebesgue measure to (00.2. . St S* t a Figure 6.ST<St.ST<St. 0].T+>T) = P(STEA. THE LADDER HEIGHT DISTRIBUTION Lemma 6 . {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. 49 Proof Let T be fixed and define St = ST . since the distribution of the Poisson process is invariant under time reversion.O<t<T) = P(STEA. ST < ST_t. see Fig. P(STEA. 0 < t < T) P(STEA.0<t<T) = F(ST E A. That is.6. 6.2(a): T+ > t Figure 6. 0 < t < T. has the same distribution as {St}o<t<T.ST_t.2(b): r+ < t Thus.O<t<T).
t dT.r. E A} precisely when r+ > t. SOME GENERAL TOOLS AND RESULTS Integrating w. cf.3 where the bold lines correspond to minimal values. U + St_ E A. and (6. this is just the Lebesgue measure of A.. .3 G+ is the restriction of /3R+*B to (0. oo).y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. Figure 6.y) (here we used the fact that the probability of a jump at u t is zero in the second step. oo). and since the jump rate is /3. Fig. G+(A) = Q f 0 B(A .2) in the last). But since St 4 oo a. The probability of this given { Su}u<t is B(A . 6. s. T+ > t] 0 _ /3 f E[B( A . That is. it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time .50 CHAPTER II. for A C (0.St _)I(r+ > t).St).T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .3 Lemma 6 . we get G+ (A) = f 00 /3 dt E[B(A .St _)..
) where ak = Ti + • • • + Tk . assuming basically stationarity in time and space.z)B(dz) _ f I(x < z)B(dz) _ f (x).3 yields g+ (x) = . . The marked point process .1 With r+(y) = I(y < 0) the density of R+. Uk) for those k for which ak . 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.* ) and the second the mark (the claim size Uk ). We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. Lemma 6. .e..1.e.. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup. 6 . the first component representing time (the arrival time o. as a point process on [0. The traditional representation of the input sequence {(TT. The points in the plane (marked by x on Fig.. 0 Generalizing the setup.s > 0). Uk) (k = 1. i. U k)} k=1 a is as a marked point process M *. {St+8 . 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. cf.6. this is equivalent to the risk process {St*} being stationary in the sense of (6..s. oo).T+ < oo).4). i.. Fig.M o 08 shifted by s is defined the obvious way.4) are (ak. In the stationary case. oo) x (0.:T1 +•••+Tk <t}. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. obviously. this does not depend on h). we define the arrival rate as E# { k : ak E [0 . Nt St =>Uk k=1 t where Nt = max{k = O.S8 )t> o = {St }t>o for all s > 0. h]} /h (by stationarity. 4 (the points in the plane are (ak . 6 . 2.Q f r+(x .
Example 6 .. h] and the sum approximately ^o(M*)I(ul < h). vi(dx) = .4 Given a stationary marked point process M*. of (6.2. The two fundamental formulas connecting M* and M are Eco(M) = aE E. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. V(M* o eak )..5) does not depend on h. We represent M by the sequence (Tk. o.QiBi(dx). Sigman [348] for these and further aspects of Palm theory. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. h] Eco(M*) = 1 E f co(M o Bt)dt. the r.s. e. letting h J. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6.g. where TI = 0. SOME GENERAL TOOLS AND RESULTS M* U.e. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. and let T = T2 denote the first proper interarrival time . h.5) represents the conditional distribution of M* given vi = 0. This more or less gives a proof that indeed (6. As above . k: vk E [0.. = 0 .. i. most often one takes h = 1). Uk) k=1. See. 0.52 CHAPTER II. . Oh becomes the approximate probability F(ri < h) of an arrival in [0. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . Section 5) which has pure jump structure corresponding to pi = a = 0.4 Consider a finite Markov additive process (cf.
6 Under the assumptions of Theorem 6. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J. the ruin probability . an arrival for M* occurs before time t + dt w.6.O for i # j.O for i = j and iriAijgij/. let the arrivals and their marks be generated by {Jt} starting from Jo = j. qij when {Jt} jumps from i to j and have mark distribution Bij./. Jt = j is iri(3i /. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution .= i. aij for (i. A stationary marked point process M* is obtained by assigning Jo distribution Tr. v. Jo) w.O fo "o F(x)dx = . Then the ladder height distribution G+ is given by the (defective) density g+(x) = ..OEU0. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . and that p = 0EU0 < 1. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i. 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 .s.oo a.OF(x).6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . the probability aij of Jt . Before giving the proof. It follows that we can describe the Palm version M as follows . 5 Consider a general stationary claim surplus process {St }t>o.p.e. .p. If Jt_ = i. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise.p. First choose (Jo_. we note: Corollary 6.5. let U0 be a r.*(0) with initial reserve u = 0 is p = /3EU0. and by some additional arrivals which occur w. Assume that St * . Note in particular that the Palm distribution of the mark size (i. After that.
Q_k and has size U_ k. The last property is referred to as insensitivity in the applied probability literature.0<u<t) = P(StEA.1] here the r . . which makes an upwards jump at time .. A standard argument for stationary processes ([78] p. oo ) and the arrival times 0 > 0_1 > a_2 > . We then represent M by the mark (claim size ) Uo of the arrival at time 0.o.5.). 105) shows that one can assume w.e. { Su}0<u<t is distributed as a process {Su} . the arrival times 0 < 0'1 < Q2 < .5). h. moves down linearly at a unit rate in between jumps and starts from S0 = U.l.. oo) x (0 . . It follows that for A C (0.St*_ u.St <. the mark at time Qk is denoted by Uk.. (k = St}t>o 1. T+ = t given the event At that an arrival at t occurs . Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt. CHAPTER H. Now conditionally upon At .Su<0.(left limit) when 0 < it < t and is illustrated on Fig .St<Su. . 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. 2.o<u<t where a claim arrives at time t and has size Uo. has a very simple interpretation as the average amount of claims received per unit time . The sample path relation between { Su } and { Su } amounts to S„ = St .Su_ <0.g.0<u<tIAt) = P(St EA.Mt). SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0.. 0<u<t) = P(St EA.. Proof of Theorem 6.A. and the kth preceding claim arrives at time t . that M* and M have doubly infinite time (i. 6. oo)).0<u<t) = P(St EA.s.54 By (6. in (oo.o.. Let p(t) be the conditional probability that ST+ E A.Su< 0. 0). in (0. are point processes on (oo ..5. oo) p(t) = P(St EA.0<u<tIAt) = P(St EA..$St_ u.
NIt)dt . and we let L(dy) be the random measure L(A) = fo°° I(St E A. In Fig. t a oo. 0 < u < t } is the event that { Su } has a relative minimum at t . the support of L has right endpoint U0.. Mt)dt = i3EL(A) o"o . and since by assumption St * oo a. G' (A) = 3 f P(St E A.5.5 where it = { St < Su. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. the left endpoint of the support is oo. Thus. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. Uo]. 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.s. 6. Since So = U0. cf.6. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. A sample path inspection just as in the proof of Lemma 6 . Fig.
[263] (a special case of the result appear in Proposition VI. 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 6.6 is Bjork & Grandell [67].1).56 CHAPTER II. [147]. .2. A further relevant reference related to Corollary 6.5 is due to Schmidt & coworkers [48].
• the premium rate is p = 1. Thus . Some possibilities are numerical Laplace transform inversion via Corollary 3. and assume that • { Nt}t>o is a Poisson process with rate j3. Panjer's recursion ( Corollary XI. i. St = uRt = EUi t. • the claim sizes U1.4 below . 3). 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 . with common distribution B.e. say.6) and simulation methods ( Chapter X).d. i.. 4. see Chapter IV. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . are i. exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. For finite horizon ruin probabilities . . {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi.. U2.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. and independent of {Nt}.
we shall start by giving the basic formulas for moments. cumulants . 0 .6pBa). m.. and Schlegel [316].)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. for (d) just note that the kth cumulant of St is tic(k) (0). A more formal proof goes as follows: Nt r Nt ESt = E > U k . Write pB^) = EUn' YB = Pali = EU.1 is the expected claim surplus per unit time. We do not spell out in detail such generalizations. (b) Var St = t. 1 Introduction For later reference.g. and that B(k)[0] = Pak). Furrer [150].. and this immediately yields (a).1) = t(p . e .g.1) .u . P = PAB = 1/(1 + rl) Proposition 1.t = E E [ U k k=1 k=1 Nt . 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).t = t(p .1). THE COMPOUND POISSON MODEL Poisson risk process.'s etc.58 CHAPTER III. Schmidli [319].1 (a) ESt = t(13µ$ .s.f.+Uk)P(Nt = k) k=O e8t k=O B[s]k . we get Ee8st = 00 e8t c` Ee8 (U1+. (c) Ee8St = et" (8) where c(s) = f3(B[s] .t = fltpB . of the claim surplus St . [324]. For (c). say stable Levy motion.Rt. where K(k) (0) is the kth derivative of is at 0. See e.t = E[Ntµs] . Dufresne & Gerber [126]. {Bt} a Brownian motion and {Rt} a pure jump process. (d) The kth cumulant of St is tf3p(k) for k > 2. Proof It was noted in Chapter I that p .1).
Sn+0 . The connections to random walks are in fact fundamental.Tk. u + v]. We return to this approach in Chapter V. (d) If 17 = 0. In particular. rather to view {St} directly as a random walk in continuous time.h < St < S(n+1)h + h. (c) If 77 > 0. we get a discrete time random walk imbedded in the claim surplus process {St}. . lim supt.V.4.3) is proved similarly.. however. Indeed. we need the following lemma: Lemma 1. obviously 0(u) = F(maxk Sok > u). if t = nh + v with 0 < v < h.1 = . II. and the value is then precisely v. Here is one immediate application: Proposition 1. For example. v > 0. cf. the Uk . then St 4 co. 2.d. so that {Sok } is a random walk with mean EUET = EU. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. (b) If 77 < 0. where Tk is the time between the kth and the (k .3EU01 = 1µs where rt is the safety loading. 1.Tk are i.1 is the same as if {St} was a random walk indexed by t = 0. St = oo.. S„+V > S„ . (a) No matter the value of 77. which is often used in the literature for obtaining information about {St} and the ruin probabilities.i. For the proof. The point of view in the present chapter is.Sok_l = Uk . Obviously.. then Snh .1. Proof We first note that for u. then lien inft. meaning that the increments are stationary and independent. then St> SnhV>Snhh. The right hand inequality in (1. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. we have Sok . In this way.3 If nh < t < (n + 1)h. St = oo.1)th claim.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo.S„ attains its minimal value when there are no arrivals in (u. then St 00..
is a discrete time random walk. 2h. 1 since St 4 oo) and repeating the argument.4 The ruin probability 0(u) is 1 for all u when 77 < 0.. at least once.. There is also a central limit version of Proposition 1. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time..2.. h. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. or by a general result on discrete skeletons ([APQ] p. However. if P(M > 0) = 1.1. 0 Snh = 00. u 307). hence by induction i. The general case now follows either by another easy application of Lemma 1. and < 1 for all u when 77 > 0.6 Often it is of interest to consider size fluctuations. it suffices to prove 4'(0) = F(M > 0) < 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)...o. For any fixed h._. then {St} upcrosses level 0 a. and hence by the strong law of large numbers.s..5 The limiting distribution of St .. 169) stating that lim infra. Thus using Lemma 1.t . where the size of the portfolio at time t is M(t). Snh u = 00 (the lemma is not needed for (d)).1).3. THE COMPOUND POISSON MODEL Proof of Proposition 1. . and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1. Remark 1 . is a discretetime random walk for any h > 0. h A similar argument for lim sup proves (a). Part (d) follows by a (slightly more intricate) general random walk result ([APQ].2.p.1. Corollary 1. lim supn_..1(b)) that the assertion holds as t 4 oo through values of the form t = 0. Notes and references All material of the present section is standard.1. and (b). 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.60 CHAPTER III. Considering the next downcrossing (which occurs w.. This contradicts u St400.2: Proposition 1. it is seen that upcrossing occurs at least twice. Snh/n a4' ESh = h(p . {Snh}n=o. If rl > 0. {Snh}n=o. (c) are immediate consequences of (a). 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 .3.
P) E PnBon(u) . B(x)/aB. which we henceforth refer to as the PollaczeckKhinchine formula.1) is not entirely satisfying because of the infinite sum of convolution powers. the formula for the distribution of M follows . the ladder heights are i. Fig. IV. oo ). This follows simply by noting that the process repeats itself after reaching a relative maximum. It is crucial to note that for the compound Poisson model.1.1 provides a representation formula for 0(u). Combined with i/i(u) = P ( M > u). [APQ] Ch. n=0 (2. 1 The distribution of M is (1. but we shall be able to extract substantial information from the formula. 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}. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . 0 Alternatively..34 or A. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 .2.1) representing the distribution of M as a geometric compound. The expression for g+ was proved in Theorem 11. Theorem 2. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . cf. equivalently.1. (2. 11.. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ).e. and we further get information about the joint conditional distribution of the surplus and the deficit. nevertheless.just before ruin is again B0. 1e. Note that this .IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). Summing over n. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. p < 1. As a vehicle for computing tIi(u).6. d. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o).6.IIG +II)EG+ . where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. Thus . i. We assume throughout rl > 0 or.
5. cf. (d) the marginal distribution of ST(o)_ is B0. V is uniform on (0. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. Asmussen & Schmidt [49]. [62]. Beekman [61]. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„).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. Theorem 2. ST(o )) given r (0) < oo is the same as the distribution of (VW. The proof of Theorem 11. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. (1 . 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. W are independent. We assume rt > 0 throughout.62 CHAPTER III. Theorem 2 . 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).6.6. in this setting there is no decomposition of M as a sum of i. .d. ST(o) > y. there is a general marked point process version. Again.2(a) is from Dufresne & Gerber [125].just after ruin. 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. see Schmidli [323] and references there. cf. f +b (b) the joint distribution of (ST( o). where it requires slightly more calculation. In the risk theory literature.V)W) where V. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. (a) 11 (ST(o)_ > x. see for example [APQ]. ST(o)) is given by the following four equivalent statements: B(z) dz.1 is traditionally carried out for the imbedded discrete time random walk. 2 The joint distribution of (ST(o )_. cf. However. As shown in Theorem 11. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV.5. Feller [143] or Wolff [384]. Theorem A1. (c) the marginal distribution of ST(o)_ is Bo .i.
then.3.p. also be seen probabilistically without summing infinite series . SPECIAL CASES OF POLLACZECKKHINCHINE 3a The ruin probability when the initial reserve is zero 63 The case u = 0 is remarkable by giving a formula for V)(u) which depends only on the claim size distribution through its mean: Corollary 3.p) = S .2 If B is exponential with rate S. however . Alternatively.6. 1 . the result follows .e. The result can.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. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 . the formula for P(O) holds in a more general setting. B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations ..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. But claims are exponential . Thus r(x) = S(1 . use Laplace transforms. I.O)e(b0)x.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. and hence this overshoot has the same distribution as the claims themselves .p) E pn S n x n.p. For a failure at x. As shown in 11. a further relevant reference is Bjork & Grandell [67]. Let r ( x) be the failure rate of M at x > 0.1 e ax = n1 (n . 0 .1)1 00 ( 1 . Integrating from u to oo. hence without memory.p)pSe a ( l v)x = p( S . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w. 3b Exponential claims Corollary 3. Thus .
Then the first term on the r.+ = y yields P(M>u.1) For a heavytailed B. u . (b) use stopped martingales .T+ <oo)+P(M> u.y)G+(dy ) = f U V(u . the survival probability Z(u) = 1 .y)G+(dy) For the last identity in (3.2).64 CHAPTER III. (3.2) Notes and references Corollary 3.+ <U. then 24 1 V. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype.y)f3 (y) dy. (u) 35eu + 35e6u.1 p pBo(u).g.3) below.+ <u. is ?7+ ( u).p + f u Z(u . (3. 0 Proof Write o (u) as P(M>u) = P(S. and weights 1/2 for each.3.3) Equivalently.i(u) satisfies the defective renewal equation Z(u) = 1 . Corollary 3. (3.3). cf.S.T+ <oo).1. E. (3. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.y)/3B (y) dy.4) zu P(M > u .3. 2 is one of the main classical early results in the area. THE COMPOUND POISSON MODEL In VIII.3)). just insert the explicit form of G+. II. u + oo. and conditioning upon S. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx). T+ <00) (3.+ >u. A variety of proofs are available . if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.h.p + G+ * Z(u) = 1 .4) can be derived by elementary algebra from (3.4) is similar (equivalently.s. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .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 . (Example VIII. The case of (3.S.
111112 or Feller [143].p) E p"Bo[s]" = 1 . Some relevant references are Abate & Whitt [2].7).3. 206207).pBo[s] no (1 . In fact.f.7) and Corollary 3.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 . In view of (3..P)pB' (3. 0 Notes and references Corollary 3. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula.4) can be derived by elementary but tedious manipulations. 191).P)PB 2(1 . e.6) 00 = (I .g. by analytical manipulations (L'Hospital's rule) from (3.p)2 3(1 .p)s s /3 . ./3B[s] . Of course.g. eau B(u) du = f PB 3PB SPB 0 o (3.5 The first two moments of M are 2 EM .s ./3B[s] which is the same as (3. Corollary 3. see e.p)s .3 is standard . [APQ] pp.5).. The approach there is to condition upon the first claim occuring at time t and having size x .Ps s(. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used. Griibel & Pitts [132].3 .(3 . [APQ] pp.8) Proof This can be shown. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3.s . Embrechts.7) s +. numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.5) Proof We first find the m.1 Bo[s] = f oc.(3B[s] 1 . We omit the details (see.p = (1 .5).3 . Also (3.5 can be found in virtually any queueing book.PPB2) EM2 = PPB) + QZPBl 2(1 .4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . g. for example. Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3.)3B[s]) (3. (3. Bo of B0 as m e8u B(du) = B[s] .
Q (k 1 k= n  [O(k . differentiation yields Z(u) _ /3Z(u) .6 If B is degenerate at p.u)]k k! k0 The renewal equation (3.1).9) shown for n .Q) k=0 k! E e0( = /32(u) . eO('u) [)3(k .z/'(u) takes the form Z(u) L^J L./32(u .u) a)Qea" + (1 .)3(1 .u) [N(k ./3Z(u .4) for Z( u) means f lhu Z(u) = 1.u)]k k! (1 L3) 1: e_O(ku) NIN (k (k .y<1)dy 0<u<1 1 < u < oo uu ulhu 1a+/3 J0 uZ(y)dy U Z(y) dy 113+0 For 0 < u < 1. differentiation yields Z'(u) _ /3Z(u) which together with the boundary condition Z(0) = 1/3 yields Z(u) _ (1/3) eAu so that (3.66 CHAPTER III. Z^ =eR(k. of (3.1).h.3I( 0<y<1)dy Z(y)/3I(0<u.u)]k1 ku+1) [/3( k .3+ 18+ J0 Z(uy).9) follows for 0 < u < 1. THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.Q) 3e.1)! k=1 u1 .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary. .9).1 < u < n and let Z(u ) denote the r. then p) 1: ep(k u/. For n < u < n + 1. Assume (3.u)]k d 1 u) _ a) n ( du ( k! (1  .s. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .u) [p(k .u + 1 )]k = QZ(u) .
say t = 1: recall from Proposition 1.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. We could first tentatively consider the claim surplus X = St for a single t.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). (4.4) works as well. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. Formalizing this for the purpose of studying the whole process {St}. of F9.r. and thus (4.a.(9) is welldefined.1) .f. co(a) = rc(a + 9) .1) .f.(9).d. The question then naturally arises whether ie is the c.f. but will now be repeated for the sake of selfcontainedness. and define rce by (4.rc(9) = .4. we set up . 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) .2) shows that the solution is Ox [O]0]. or equivalently BB[a] = B[^+ Repeating for t 54 1. 0 Notes and references Corollary 3.g.1 that c(a) = /3(B[a] .Qe(Bo[a] .2). it follows that Z(u) = 2(u) for n<u<n+1.4.f. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a].g. we just have to multiply (4. (4. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references. B9(dx) = B[9] B(dx).3B[9].4) .3B = .1) or equivalently. The answer is yes: inserting in (4.g. F and c. in terms of the c.2) (Here 9 is any such number such that r.a.3) by t. (4. 4 Change of measure via exponential families If X is a random variable with c.
.i. with common c.FT. (4.d.r..t. and define 09. and thus (4.(9)} (4.d. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity. EeeBSt + tk(B) = 1. Then the Xk are i. then EBZ = E [Ze9ST _T"(9)I . . G].r. Let FT = o(St : t < T) denote the o•algebra spanned by the St.t. . it suffices to consider the case where Z is measurable w.Tic (0)} .8) By standard measure theory.8) follows by discrete exponential family theory.1 Let P be the probability measure on D[0.10) .9) Proof We first note that for any fixed t.. Ti(a)/n.3 Let T be any stopping time and let G E FT.g..6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} .1) and multiply from 1 to n). Proposition 4. v(Xi. . (4.7) Proof We must prove that if Z is FTmeasurable. and dP(T) dP^T) That is. But let Xk = SkT/n . for G E FT. the corresponding expectation operator is E9.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 .S(k_1)Tln.f. n) for a given n. t < T. Xn). Z is measurable w. The identity (4. . G C {T < oo}. = exp {BST . The following result (Proposition 4.0e and claim size distribution Be. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} .1. oo) governing a given compound Poisson risk process with arrival intensity.nr. in particular the expression (4. (4. (4. the PBT) are mutually equivalent on. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .4).FTn) = Q(SkTIn : k = 0. (4.3 and claim size distribution B. THE COMPOUND POISSON MODEL Definition 4.68 CHAPTER III.2. .5) for the density.2 For any fixed T. and PBT) the restriction of PB to FT.5) for the density of n i. BB by (4.i. G]. replications from Fe (replace x by xi in (4.
the typical shape of rc is as in Fig.ST) + (T . Then GT = G n Jr < T} satisfies GT E FT.1 It is seen that typically) a ry > 0 satisfying 0 = r.r)rc(9)}I . according to what has just been proved. The behaviour at zero is given by the first order Taylor expansion c(a) r.FT]] = EB [exp { BST + Trc(9)} I(G)] . t = T .f..g. Letting T t oo and using monotone convergence then shows u that (4.(Y) = 13(B['Y] . so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . 5 Lundberg conjugation Being a c. Now consider a general G.9) holds for G as well. Thus by (4. 77 Thus.10).9) holds with G replaced by GT.7) holds.7 1 Some discussion further supporting this statement is given in the next section. .1) . LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.5. GT C_ Jr < T}. and hence (4. subject to the basic assumption ij > 0 of a positive safety loading. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. Ee [exp { BST +Trc(9)} I(G) FT)] = 1.r is deterministic. c(a) is a convex function of a. Then G E FT. Given FT.1) _ 1 + a. 5. (0) + rc'(0)a = 0 + ES1 a = a (p . Thus. (4.1(a).
Thus.1(b).2 s As support for memory. (5. the claim surplus process has positive drift > 0. Taking T = r(u).s). Fig.a = i(a + 7). Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes.1) .2)) is 7 = 5/3.3.1) (or (5. 5. we write FL instead of F7.3.1) is known as the Lundberg equation and plays a fundamental role in risk theory . 5. Fig.1) is precisely what is needed for one of the terms in the exponent . G = {T(u) < oo} in Theorem 4. An established terminology is to call y the adjustment coefficient but there are various alternatives around. (5.3. an equivalent version illustrated in Fig.4) ELS1 = #L(0) cf. Example 5 .1(b).3) cf.QL instead of /37 and so on in the following . b[s] = 5/(b . Equation (5.2) 7 Figure 5. 5.70 CHAPTER III. .1 Consider the case of exponential claims. the Lundberg exponent. we further note that ( 5. and (4.4) yields /3L = b and that BL is again exponential with rate bL =. e. u It is a crucial fact that when governed by FL. It is then readily seen that the nonzero solution of (5. THE COMPOUND POISSON MODEL exists . Note that KL (a) = /L (BL [a] .2 is B(7) = 1 + ^. (5.g. Thus B[7] = 6/.
(oo) (in the sense of weak convergence w.1p . T = T+.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.1. 0 Theorem 5 .5. where C .t.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .(u)} .7) 0 and all that is needed to check is that ( 5.Ce7u as u 4 oo.r.5) Theorem 5 . V) (u) = P(T(u) < oo) = EL [exp {ryS. T(u) < oo] .7) is the same as (5.3.2 (LUNDBERG'S INEQUALITY) For all u > 0. (5. ST+ E A] . see A . (5.6) Proof By renewal theory. Since a7' is continuous and bounded.P Y j o' xeryxOB (x) dx /3k [Y] . PL ) with density 1 . 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. To this end. V)(u) < e7u. Proof Just note that e(u) > 0 in (5.1e. Letting e(u) = ST(u) . which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.6 ).1 (5.e7x)G+(dx).G+L)(x)) dx ry^+L) J 00 f 0 (1 .+ E A} in Theorem 4.ascending ladder height distribution and µ+ its mean. (5. take first 0 = ry. Then P(ST+ E A) = EL [exp { 7S?+} .4).3 takes a particular simple form.8) . G = {S. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4.5). e(u) has a limit i. we can rewrite this as 0(u) = e"ELe7^(u).
Noting that SIG(L)II = 1 because of (5. that 7 = S . .7).3 (this was found already in Example 5. From this it follows. u . we get L where 00 (1 . THE COMPOUND POISSON MODEL In principle.1) (5.4).1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] . (5. but some tedious (though elementary) calculations remain to bring the expressions on a final form. 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.1 .10) VW = JI c* e° (x) dx = a (B[a] .12) Example 5 .11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.72 CHAPTER III. or equivalently of how close the safety loading 77 is to zero.")G + (dx ) = 1  J0 00 3B(x) dx = 1p. Then 0(u) = pe(a_Q)u where p = /3/S. this solves the problem of evaluating (5.4 Consider first the exponential case b(x) = Seax.8) yields +L) J0 xel'B ( x) dx (5.e. Using (5.1 above) and that C = p. of course.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.
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)
.
Ps(/3max .p = (/3max 0)µB.1. and hence the ruin probability approximation is b(u) e(bAln)u. p* _ . we shall represent this situation with a limit where /3 T fl but B is fixed. heavy traffic conditions mean that the safety loading q is positive but small.p .3 )1 } _ 1p 1 .PBo [s (/3max .P .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . That is. the approximating risk process has ruin probability z. numerical evidence (e. we have according to the PollaczeckKhinchine formula in the form (3. but has an obvious interpretation also in risk theory: on the average.2) was suggested by De Vylder [109].3* /S. [174]) shows that it may produce surprisingly good results. THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 . (/3max ./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 .)3 )PBo µB  .1 As /3 f Nmax.g. Mathematically. Proposition 1.3 and Corollary 3. the premiums exceed only slightly the expected claims.1.80 CHAPTER III. Grandell [171] pp. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. Notes and references The approximation (7.p + p { 1 1p ti 1 . or equivalently that /3 is only slightly smaller than /3max = 1/µ8.b(u) = p*e.(bA*)". cf./3)] 1 ./3)PBo PB .s(/3max .8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . Letting Bo be the stationary excess life distribution. Though of course it is based upon purely empirical grounds.(3)2 P PB 2µB 2µB Letting /3* = /3/P. Proposition 7. 1924.
7. the premiums are much larger than the expected claims . 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.B AB ) 6()3max _'3) = However .l3)M > (/3max . 7d The light traffic approximation As for heavy traffic .p _ 2rl11B PB p. u * oo in such a way that (3max . but has an obvious interpretation also in risk theory: on the average . Of course.4) suggested by the Cramer Lundberg approximation and Proposition 6. . We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . the term light traffic comes from queueing theory.p. the first results of heavy traffic type seem to be due to Hadwiger [184].Q T /3max. It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. [APQ] Ch.2 If .ze a2unµB laB (7. This follows since rl = 1/p . That is . Numerical evidence shows that the fit of (7. Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]./3)u * v. in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . obviously Corollary 7.3) is reasonable for g being say 1020% and u being small or moderate. VIII).ryu . or equivalently that 0 is small compared to µB .2. However .1 1 . we shall represent this situation with a limit where 3 10 but B is fixed. 2 provides the better mathematical foundation. then P(u) 4 e6„ Proof Write z'(u) as P((/3max . while the approximation may be far off for large u. light traffic conditions mean that the safety loading rl is positive and large . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e ./3)u).g. and hence 2µ2B 1 .0)u. Mathematically. These results suggest the approximation Vi(u) e6(0_. light traffic is of some interest as a complement to heavy traffic . In the setting of risk theory.
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 .(3 10. U > u] = /3iE(U .3 is the same which comes out by saying that basically ruin can only occur at the F(U . Asmussen [19] and references there. Sigman [347]. by monotone time T of the first claim . THE COMPOUND POISSON MODEL Proposition 7.82 CHAPTER III. (7.e. En'=2 • • • = O(/32) so that only the first terms matters.u.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm.5) u Proof According to the PollaczeckKhinchine formula. cf.3 As . i. [97].= 1 aJ 1 a 0+ 1 = = p. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx.T > u). 10 ( u The alternative expressions in (7. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. the Poisson case is much easier than the renewal case.5) follow by integration by parts. Indeed. Omax max m. Again. u Note that heuristically the light traffic approximation in Proposition 7. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . see Daley & Rolski [96]. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. Light traffic does not appear to have been studied in risk theory.T > u) = J o" B(x + u)/3eax dx . ( 3 J O B dx. z/' (u) convergence P(U . 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. . 0(u) /3 J B(x)dx = /3E[U . For a more comprehensive treatment.u)+.
6) is . however. Thus . . no empirical study of the fit of (7.6) (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. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). The adaptation to risk theory is new. (7.O0 M./3)) . COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . Notes and references In the queueing setting . _(E) (u). where further references can be found .x . ) M. (U). 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. that is.VHT) ( ax QmQ ) h (B) ( . Let OLT) (u) denote the light traffic approximation given by Proposition 7. "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . Instead. ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. [84]. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1).Wmax f(x ) dx + pee6mQ. B(2).3).3 and use similar notation for %(B) (u) = (u). Al . 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. Another main queueing paper is Whitt [380]. . Substituting v = u(. even if the safety loading is not very small.O(E)(u) 1 (1 . to get nondegenerate limits .3n. the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. with rate 1/µB = /3max. z/i(E) (u) = pe(QmaxQ)u. we see that the following limits HT) (u'). one may hope that some correction of the heavy traffic approximation has been obtained. 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).8. available.
B(') <i. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. we have the convex ordering. B(2) and PB(1) = µB(2). equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. Rather than measuring difference in size. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}.ill(u) < V)(2) (U) for all u. the proof is complete. u Of course.6. U(2) distribution B(2) and U(1) < U(2) a. we shall need various ordering properties of distributions.' 1)(u) < V)(2) (U) for all u. U(2) such that U(l) has distribution B(').84 CHAPTER III. whereas (consider x2) B(2) has the larger variance. Here convex ordering is useful: Proposition 8. . B(' <. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. Proposition 8.s.1 is quite weak.2 If B(') <j. In terms of the time to ruin. then . THE COMPOUND POISSON MODEL To this end. B(2)) if f fdB(1) < f fdB(2) for any convex function f. this ordering measures difference in variability. Finally. XI. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. then Bill = B(2). most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. Proposition 8. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. Bill is said to be convexly smaller than B(2) (in symbols. or the existence of random variables U(l). for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. Taking probabilities. 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).1 If B(') <d B(2). Proof According to the above characterization of stochastical ordering. then i. cf. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. we can assume that 1) < St 2l for all t. In the literature on risk theory.
say to p.3 If B(1) <.. then B(1) <. larger variance is paramount to larger second moment.4) certainly supports this view: noting that. A partial converse to Proposition 8. Proof Consider the light traffic approximation in Proposition 7. Bo1) <_d Bo2) which implies the same order relation for all convolution powers. Example 8. A general picture that emerges from these results and numerical studies like in Example 8. A first attempt would of course be to identify 'variation' with variance. we have by Jensen 's inequality that E f (U) > f ( EU).5 If '0(1)(u) < p(2) (U) for all u and a. and here is one more result of the same flavor: Corollary 8. The heavy traffic approximation (7.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). . The problem is to specify what 'variation' means. (D) (u) < O(B) (U ) for all u.2 is the following: Proposition 8.4 Let D refer to the distribution degenerate at 'LB .u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1.1. Hence by the PollaczeckKhinchine formula . we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I.1 and µB at 1 so that the safety loading 11 is 10%. Proof If f is convex. B(2). B. from which the result immediately follows. and consider the following claim size distributions: B1: the standard exponential distribution with density ay. Corollary 8.3 provides another instance of this.6 Fix /3 at 1/1. with fixed mean. 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. then /'(')(u) < 0(2)(u) for all u. Then V.8.(1) (. B(2).p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities.e.. This u implies that D <. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal.
van Heerwarden [189]. B. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!).1%. Let ua denote the a fractile of the ruin function.01%. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. However. For B1i B2.001 u0. B3: the Erlang distribution with density 4xe2x.4142. Kluppelberg [234].4. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. In terms of variances o2. and consider a = 5%. 9 Sensitivity estimates In a broad setting. = 0. 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).000. in comparison to B2 the effect on the ua does not show before a = 0.e'\1x + 0.9A2e'2r where A.lA. B3 the comparison is as expected from the intutition concerning the variability of these distributions. i. 0. B4: the Pareto distribution with density 3/(1 + 2x)5/2. 1/)(u. 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. 11 Notes and references Further relevant references are Goovaerts et al. 1%. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.. Note to make the figures comparable. the behaviour of which is governed by a parameter 9. all distributions have mean 1.01%. We return to ordering of ruin probabilities in a special problem in VI. Pellerey [287] and (for the convex ordering) Makowski [ 252]. A standard example from queueing theory is .0' U0. 0.e.1358. A2 = 3.) = a. [166]. and this is presumably a consequence of a heavier tail rather than larger variance.86 CHAPTER III. One then obtains the following table: U005 U0.
and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. it follows that ' is approximatively normal N(0.Ap). the distribution of %3 0 is approximatively normal N(0„ Q/t).1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). Assume for example that 8 is known. the premium rate p and the claim size distribution B.01/2u. Proof This is an easy time transformation argument in a similar way as in Proposition 1. a2/t). a0 as ao 80 19P . the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively .2 Consider a risk process { Rt} with a general premium rate p. Similar conclusions will be found below. SENSITIVITY ESTIMATES 87 a queueing network.(u) for large u.e. u Proposition 9. increasing in u.3. Thus at p = 1. where Q2 = fl ( l2 1113 / _ Ou2v)2.1.9. 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. Example 9. For example. t]. obtained say in the natural way as the empirical arrival rate Nt/t in [0. a/3 0 . Then a p ao = 00 Qa/. while /3 = j3 is an estimate. where the partial derivatives are evaluated at p = 1. 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. i. and s(9) the expected sojourn time of a customer in the network. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. Then ib = Pe(613)u. Then if t is large . if = a e(6A)u. Thus. Let R(P) = Rtli. say estimated from data. In the present setting.19P a/ . In particular .. with 0 the vector of service rates at different nodes and routing probabilities.
and the proofs of (9. u Now consider the ruin probability 0 = 0 (u) itself.t.88 CHAPTER III. and write yp = 8y/8/3 and so on .Owe (9. Proposition 9.3.6) As will be seen below.((dx ) = exp {Ox + (t(x) .g.1 or Proposition 9.()^ 1 .4). In the case of the claim size distribution B.3) follows by straightforward algebra.3 70 = 'Ye = = 7 /3(1 we(e +'y. x > 0 (9. 4) (9 . (9. However .2) (see Remark 9. ^) .3.()wC(e. Viei '0(. (3+'y)PC (0+7. Similar notation for partial derivatived are used below. it suffices to fix the premium at p = 1 and consider only the effects of changing . () Proof According to (9. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable.w(6.()(0 +'0) ' (9 . various parametric families of claim size distributions could be considered.0 = t/'(u) and the CramerLundberg constant C. (9. e. but we shall concentrate on a special structure covering a number of important cases.()YC = 1 +y/ /3 \ Q2 From this (9.3 or/and B. /3 yields w e(e + Y. 3) ( 9 . so that heuristically we obtain '00 50ryu = Coe"u . Consider first the case of 8/8/3: . The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . (. but must look for approximations for the sensitivities 0. ()} p(dx) .uypCe7u urypO. we can rewrite the Lundberg equation as w(9+ y. this intuition is indeed correct.10) below. Consider first the adjustment coefficient y as function of 3.5) are similar.6 below for some discussion of this assumption). we cannot expect in general to find explicit expressions like in Example 9. Of course.(/3 + y)we(9 + 7. namely that of a twoparameter exponential family of the form Bo. for the ruin probabilities . Differentiating w.w(O. () = log(1 + y//3). 5) (Q+'Y)[we(0+7.^)] 1(/3+y)we(9+'y. THE COMPOUND POISSON MODEL As a consequence.r. 9.
Z= zl + z2 where zl (u) = e7u J m B(x)dx. By dominated convergence.x)B(x)dx. z2(U) = e7" J u b(u .QB(x) dx. u 0 Then Z = z + F * Z and F is a proper probability distribution .x). Further write de = [we (9 +'y. () .w(9.2 of the Appendix ).p)/C'y.8) (Section 5). PF = (1 .9.12)).8). Z(u)/u a C//3PF where PF is the mean of F.g. we multiply by e7" and let Z(u) = elt" cp(u).3) for z/'(u). we note the formulas Ee.3 (see in particular (5. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9. Be.10) (9. z2(u) _ 1 ^ e'ri`i7i( u .9) (9. Hence by a variant of the key renewal theorem (Proposition A1. and alsoo zl(u) + 0 because of B['y ] < oo.we(9 . SENSITIVITY ESTIMATES Proposition 9. the proof is complete. O} (9.C).3(x) dx.([a] = exp {w(9 + a. () . ()] exp {w (O + y.x)B(x) dx + J U W(u . it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.4 As u oo.(e"U = = wS(O.w(9. w((9 + a. () .w(O. But from the proof of Theorem 5. ()} . ()} .8) Letting cp = e0/e/3 and differentiating (9.St (U) Ee. we get p(u) = J "O B(x) dx + J U O(u .4t (U)e°`U = which are wellknown and easy to show (see e. 11 For the following. 0(u) = /3 Ju"O B(x) dx + f 0 0(u .x). (9.11) Ee. () . Combining these estimates . () exp {w(9 + a. BarndorffNielsen [58]).x) F(dx ) f u J C F(dx) = C as u 4 oo. F(dx) = e'yy/3B(x)dx.
C) .e7x/3 f 00 [t(y) .w(e.x)f3 f ^[t(y) .wc(O. z = zl + z2. 8 8() 8( (9. this implies Z = z + F * Z. )}B(dy)• Letting cp it thus follows from (9.we (0.90 CHAPTER III.w( (0. Then as u > oo.wc (O.6e7u f "o f[t(y) .w (9. ()} 1z(dy) = f [t(y) . u Z2(U) = e7° f u ^/i(u . ^) . ())B(dy) dx.QB(x)dx.wc(9. ()](e7v . 8^ ue7u.w(0.11).6C do 89 1p 8( 1p Proof By straightforward differentiation.12) f exp {O y + (t(y) .w( (0. By dominated convergence and (9.5 Assume that (9. 2 z 07P N ue7u (3C de .8) that cp(u) . ()]B(dy) dx.1) B(dy) 'f '[t(y) .lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . 01 (i+) do = +'Y. C)] (1 + 7 ) Proposition 9. THE COMPOUND POISSON MODEL [we(e+7.x). 0 x Multiplying by e7" and letting Z(u) = e"uV(u).2) holds.9)(9. oo z2 (u) f C . ()]e7vB(dy) 'fCd 7 c . ()]B(dy) dx x 0 0C T ON O . ^)} [wc (0 + 7. F(dx) = e7x.
9 = S. < = a.Sry a/32 + a/37 + /37 .14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 .3ary tog('Finally. we (9.rye) S 5rya.. It follows after some elementary calculus that p = a)3/5 and.yu/3C2do u86 89 1p' az/) = 8z/.Y)a+1 ' (9.. and the proof of the first one u is similar. Here (9. Z(u) /3C 91 o c'o e11(t (y) .pa+1 . ())B(dy) < oo. w(e./35' a/i'y + aryl 625ry.QS 1 . () = C/9 = a/S.12) takes the form y) alp a . .9. () = log r(a) .w((9.ry) 5a1 cry (5 .16) (9.6 Consider the gamma density b (x ) = Sa xa. Example 9.12) follows.(log r(a) a log S)} • r(a) 1. by inserting in the above formulas. that C = a.18) (05 + 57 _'3_y .1edz = 1 exp {Sx + a log x . t(x) = logx.17) (9. U 7µF from which the second assertion of (9. a /(S .15) (9.1 .2) holds with p(dx) = xldx.. () ='I'(t. and also zj (u) 4 0 because of f Hence. ( 9.C log(9).) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function. ue_Yu 'C2d( 8a 8( 1 p .a log S = log r(c) .13) (9. We get w( (0.a/35a&y' ' (9. SENSITIVITY ESTIMATES as u 3 oo.
THE COMPOUND POISSON MODEL Example 9.([Y] .3. t(x) _ . () = Cc . for a < a* = z (.2 .2a) } Thus the condition B[a*] > 1 + a* /.2) with µ(dx) = 2x3zrdx.92 CHAPTER III. 9 = ."62 .w(9.9) (() . w(e. Be. further yield .2 log (0.CZ try)} 1 C C2 try . Straightforward but tedious calculations .l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2. which we omit in part .S[a] = exp {w (9 + a.1 = eXP {c(C .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9.21og 2. C) = B = Yc = de = do = . ()} = exp {c (C . C = .3Ee.22.log c = 2 In particular. C) .1 16 +ry c C22ry 2( = + 70 We (e.
the models there (e. thus.+UNT) > 1.. in which case we can just let t(x) = 0. the results presented here are new. then ryT < 0. or Ct(x). 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3.1) .g.12) takes the form a = a 93 ar. in u which case the extension just described applies.7 and references there. to our knowledge. sj=1 and let YT be defined by IKT('ryT) = 0. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. we can just fix k . [379] consider a special problem related to reinsurance.10. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x).2) is motivated as follows. for which we refer to X. let NT 16T = ^T . Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. That it is no restriction to assume k < 2 follows since if k > 2. ae t 1lEY u S _ . then BT and hence ryT is undefined. However . However. and we estimate y by means of the empirical solution ryT to the Lundberg equation. (9. the exponent is either Ox. kT (a) = /T (BT [a] . Thus. . Comparatively less work seems to have been done in risk theory. Finally if k = 1. if 1 PT = /3TNT(U1+..8 The specific form of (9. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . B are assumed to be completely unknown. we have assumed k = 2 and ti (x) = x.3C2de 1p' z a = c .cue_7u)3C P Remark 9. the main tool is simulation. BT [a]= NT ^` e"U. 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).2 of the parameters. queueing networks) are typically much more complicated than the one considered here.oo. Note that if NT = 0. Van Wouve et al. and hence explicit or asymptotic estimates are in general not possible. Thus. To this end.a.. Also. In general.
If . 16T where V1.1) .1 As T 4 oo.7 + (. it is easy to see that we can write \ V1 1 l _ .b[Yp'V21 T { (E[7] . 7T a4' 7.1)2 + E[27] .1) .B[7]2 }) ( T 0 .i3)(B[7] 1) + (3(BT[7]  .3/ T)./^ B[27] .B[7]) 0+ Iv/o(b[y].1) 'YT . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . B[2'Y]  /3T ) . vfoVFB[2y]. a2 where a2 = /3r.v.3) Proof Since Var(eryU) = we have B[7]. B[27] . since NT /T . we need a lemma. For the proof. (10.94 CHAPTER III.a BT[7] I B[7] I + .'Y . B [7]2 (10.: N ()3.B[7]2 n Hence ( 10. Lemma 10 . THE COMPOUND POISSON MODEL Theorem 10 . 1) r. then (10. V2 are independent N (0.(27)/K'(7)2.B[7]) + B [7] .B[7]2 V2 .: N 0.)vl+ N CO.Q and Anscombe 's theorem.2 As T * oo.2) rT(7) N N (0. If furthermore B[27] < oo. More generally.'s.T y . .3T ..2) follows from NT/T a4' . N ( n[7].
4) + E). Theorem 10.1 can be used to obtain error bounds on the ruin probabilities when the parameters . I. first note that e7TU N (e7U u2e27Uo'2/T) 7 .1 By the law of large numbers. OT a 95 u 4 /3. and the truth of this for all e > 0 implies ryT at 'y.e) < 0 < r.(ry + e) and hence KT(7 . Let 0 < E < ry.2.4) and Lemma 10. °7IT) .'(y).E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r.. Proof of Theorem 10. where ryT is some point between ryT and ry.(ry . y + E) eventually.e. BT[a] 3 B[a].e.10. If ryT E (7  we have KT(7 . 0 are estimated from data . Then r.Q.KT(7) kT(7) K'(7) . ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10. it follows that 7T7 KT(7T) .c'(7) N (0' T (2(7) / N (0.3). To this end . NT BT [a] Hence r. 6"Y (10. By the law of large numbers.'T(a) = 1 E Uie°U' a$' EUe "u = B'[a]. lcT(a ) 4 /c(a). Combining ( 10. 7T E (y .E) < 0 < kT(7 + E) for all sufficiently large T . NT i =1 n'(a) for all a so that for all sufficiently large T K7 .E) < 4T(7T) < 4T(7 + E). Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).
T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. 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.. ft. Deheuvels & Steinebach [102].. U2. Letting Wo = 0. if B is exponential with rate 8 so that ry = 8 .e. Mammitzsch [253] and Pitts. i.ueryuU ".g. it means 2 (8 . Vt = St . . Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. and the known fact that the Y„ = max Vt tE[W„1..C1e"a ( see e..1 is from Grandell [170].i. For example . t]}. i .0) < 5. V.d. Hipp [196]. Asmussen [23]) can then be used to produce an estimate of ry.3 or equivalently p > 1/2 or 11 < 100%. Notes and references Theorem 10. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. 6 < 2. Griibel & Embrechts [292]. Csorgo & Teugels [95]. with a tail of the form P(Y > y) .info< „< t S.Wn) are i ..5%).T VIT where r7ry. [197].) = a (e. Embrechts & Mikosch [133]. = 1. various alternatives have been developed. wn = inf{t > W. 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.. satisfies b(.1 : Vt = 0.. Wn). This approach in fact applies also for many models more general than the compound Poisson one..g.Q. the nth busy cycle is then [Wn1. > 0 for some t E [Wn_ 1. Frees [146]. Herkenrath [192].e.f. For this reason .96 CHAPTER III.96 if a = 2.
exists. The safety loading is q = 1/p . 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. B[•] and mean AB.g. the Poisson intensity is 0 and the claim size distribution is B with m. 0. Further let 'Yo be the unique point in (0. the premium rate is 1. The notation is essentially as in Chapter III.f. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] .s. In particular. generalizations to other models are either discussed in the Notes and References or in relevant chapters. 97 .Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u.1 where p = 13µB. Unless otherwise stated.1 (the role of ryy will be explained in Section 4b). it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y.1) . See Fig. 'y) where c(a) attains it minimum value.
FL and independent of T(u). 1 Exponential claims Proposition 1. Var[T(u) I T(u) < 00] = VarL T( U) .) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k.2) Proof Let as in Example 111. . In particular. 7.(4. PL = 6/0 = 1/p > 1). we have for k = 1. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .(U) < 00] = ELT(u)ke'YS.1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0.5 .u is the overshoot..1) (1. By the likelihood identity III.t. using that the overshoot l.1 The claims surplus is {St}. 1 FL. 2 that E [T(u)k. 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&) .(.9). (u) is exponential with rate 0 w. E[T(u) I T(u) < 00 ] = ELT (U).r. PROBABILITY OF RUIN IN FINITE TIME Figure 0.
0) .I (1. is V1rLSr( u) +VarL ((PL . 1). which leads to the quadratic 02 + (/3 . Let 0 > yo be determined by ^c(0 ) = a./3) .V/ is as asserted.12 Thus the l.1//32 (6/)3 1)2 26(/3u + 1)/(6 .1 (6)3)2 which is the same as the r. u + ELe(u) _ PL . of (1.B = a.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0. where = eBu I 1 .h. This means that /3(6/(6 .1)) ELST(u) ELT(u) (PL .2).(yo) = 2V ."(ry) = 26//32.1)T(u))2 = UL where = s.1.1)ELT(u).s.2) is aLELT( u) . .6a = 0 with solution 0 (the . 0 Proposition 1. Wald's second moment identity yields 2 EL (Sr(u) . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u). T(u) < oo] fora > r. the 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 .h.6 + a)0 .6.s.h.1)T(u)) = VarLe(u) + (PL ./3 .(PL . EXPONENTIAL CLAIMS For (1 . Since Sr (u) and (PL .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.1) .1)T(u) are independent with QL the same mean .1)2VarLT(u) + 2 Ca 1I VarLT(u). we have by Wald's identity that (note that ELSt = t(pL .s.
. T(u) < oo] = EB [exp {aT(u ) .. the result follows..v. More precisely.. T(u) < oo] = e. and M(u)+1 is the index of the ladder segment corresponding to T(u).OuEee 04(u) = ee u be BB+B where we used that PB(T(u) < oo ) = 1 because 0 > ryo and hence E9S1 = K'(0) u > 0..T+ Ti a t U T I 1 a i F..Y1 Y2 Figure 1.4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0). T2 .'s with rate 5. Cf..3. Y(u) belonging to a convolution semigroup . (1. 1.3) we have E [e«T(u ). But by the fundamental likelihood ratio identity ( Theorem 111.1 . Fig.4. Note that it follows from Proposition 1. ..0. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment ..1 where Y1. Ti. are the lengths of of the ladder segments 2.9ST(u) +T(u)!c(0)} . St Ti F. are the ladder heights which form a terminating sequence of exponential r.100 CHAPTER IV.3 that we can write EeaT( u) = eeuEe 017(o). Using 5 = 6 . Y2.v.
.k + 1). EN has an Erlang distribution with parameters (N. .1.3 Assume that claims are exponential with rate b = 1. Then V(u.0.4. T. 1). .T.i (u. then VT = U1. UN.v.1. 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 . i. the following formula is convenient by allowing t.T the service times of the customers awaiting service . where U1. U2. .6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . including the customer being currently served).ST). let (cf. EXPONENTIAL CLAIMS 101 For numerical purposes .3 sin0 + 29) f3(0) = 1+/32/cos9. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . Since U1 .cos (u/... . cf.T + • • • + UN.1 )!.e.T are conditionally i. Proof We use the formula . 2..1(u. T) to be evaluated by numerical integration: Proposition 1.d. If QT = N > 0. the conditional distribution of VT given QT = N is that of EN where the r..T is the residual service time of the customer being currently served and U2 .i. Let {Qt} be the queue length process of the queue (number in system. density xN lex/(N . ..T) 1 I fl(O)h(0) fdO where (1..6. UN.I ex cos B cos j O dB fo " . Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V.T. Corollary 11. For j = 0.T.6(u) = Vfl/j l(Su.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. and exponential with rate S = 1.
i(k +1)e R [/3( klal/2e:0 (01 /2 e .13(k +l)/2ei(k +1)9 R E . PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.cos (( k + 1)0)] f3(9) Hence the integral expression in (1.)3k +1 tj g'(QT >. and define tj = e(1+R)Taj/2Ij(2vT T). in particular equations (1.1 00 ok+lR 00 j=k1 +1)/2e . let I _ j (x) = Ij (x).(31/2eie .k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.cos((k + 2)9)] d9.102 CHAPTER IV.1)] L _112 /(k+1)/2 [. 8789) 00 E aj j= 00 = 1.1 R [. Then (see Prabhu [294] pp.ie .44). k k2 + $k+1 E bj 00 t j . 912.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .38).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 .31 /2eie L 1)] 1 I/31/2eie . (1.3(k +1)/ 2ei(k + l)6 (.(31/2 cos (( k + 2)9) . 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = . similar formulas are in [APQ] pp. f3(0) .8 ) yields F(QT > k + 1) .112 l 1( k +1)/2 [ 31/ 2 cos(kO) .
k! k=O k0 i/z Co Uk ate" o'/z e . is numerically unstable for large T. T). We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St.e = e' COS a cos(uf31/2 sin 0).T) which. Related formulas are in Takacs [359]. The rest of the proof is easy algebra. or. equivalently.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. going back to Cramer. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1.0(u. Seal [327] gives a different numerical integration fomula for 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). from the accumulated claim distribution N. 2 The ruin probability with no initial reserve In this section . E Fk.2. there are several misprints in the formula there. the numerical examples in [12] are correct). however.. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. t) = P . 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. t )). F(x. expresses V)(0.7) that _ [^ au ak+l (30 k L. Ui < x I / (note that P(St < x ) = F(x + t. . it follows as in (1. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. T) in terms of F(. u Notes and references Proposition 1. however.
T) = P(Tr(0) > T) = P(M(0.b (0.3) with A = (0.1.(6.T) T F(x. . resp.(.t)= {Stv) < SM. T T o where the second equality follows from II. we define a new claim surplus process St StM NJ Figure 2. ") } is at a minimum at time t. f T lStv)} 0<t<T by a 'cyclic translation'. [v.1 In formulas.i.T)) does not {Stv)} depend on v.104 CHAPTER IV. See Fig. co ). 2.0<w<t} St+v . Proof For any v E [0. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 .T))dv E^T I(M(v.T)dx.(0.T]. Stv^ _ Define M(v.T))dv. Then 1 . and the third from the obvious fact (exchangeability properties of the Poisson process) that has the same distribution as St = { Si0)} so that P(M(v. T]. 1 1 . meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. v].T)) 1 fT P(M(v.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS.
we can take v E (w E.Sv.T) occurs or not as long as ST < 0.2 10(u.v<t<T} = {ST<StSv. then i fT I(M(v. .T)f(I z /)(0. v<t<T}n{ST<STSv+St. we can write M(v. where the last equality follows from ST < St on M(0. 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). It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. Obviously.T)) dv f T I(M(0. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. T) occurs. or it occurs. v).. v). T Theorem 2 .xdx. t) denote the density of F(•. T T o i =1 Let f (•. T]. Indeed. w) for some small E. cf. v) = M(0. If ST < 0. ST > 0. T. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0.T) and Sv < 0 on M(0. this integral is 0 if STv) ..2. For example.t)dt.Tt))f(u+t. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T . there exist v such that M(v. v < t < T} n M(0. 0<t <Tv}n{ST<ST Sv+St T+v. Fig 2. T)). It follows that if M(0 . in which case there is a last time o where St downcrosses level u. then M(v. v). T) as {ST<St+ vS. T) occurs.ST on M (0. We claim that if M(0. T) = M(0.T) = F(u+T.T) occurs.2. letting w = inf It > 0 : St_ = mino<w<T Sw}. Hence T TE f I( M(v. 0<t<v} = {ST < St . T)) dv.
0 < t <T. which occurs w.2 Here o. {St > .2 .106 CHAPTER IV. Proof of Theorem 111. ST_ _ z}.2. define St = ST . Hence P(ST<u) = 1 . t + dt] occurs if and only if St E [u.u+dt]). For a fixed T > 0.ST_ t_ and let A(z. The proof is combined with the proof of Theorem 111.T)+ J0 T (1V. {S t > z.T) = {St < 0.t).3 Define r_ (z) = inf It > 0 : St = z}. which is independent of St and has the stationary excess distribution B0. 2. O(T . Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).T) = .2.p. u + dt] and there is no upcrossing of level u after time t.Tt))P(StE[u. E [t.T) = C(z. z > 0. Let Z be a r.v. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. The following representation of T(0) will be used in the next section.(0. Proposition 2. 0 < t <T . u which is the same as the assertion of the theorem. C*(z. ST_ _ z}.b(u.z. 0 < t < T. ST_ _ z} .
T))f3B(z) dz dT. {St }o<t<T have the same distribution . THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T.T)).T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. It follows by division by P(ST(o)_ E [z.T) = C*(z. z + dz]. Figure 2.1) yields P(ST(o)_ E [z. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. u which is the assertion of Theorem 111.2. z + dz]) = P(A(z. .1) z T .3).ST(o) >y. r(0) < oo) = 3R(z) dz JP(C(z.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. z + dz]. (2.1) that P(T(0) E [T. 2. we therefore have P(A(z. ST(o)_ E [z. Fig. and since {St}o<t<T.3.T)) = P(Cx. Proof of Proposition 2. A(z.3 But by sample path inspection (cf. 7( 0) < oo) = P (C(z)) dT. Thus P(Sr(o)_>x. Hence integrating (2. T + dT] I S7(o)_ E [z.2.T + dT]. T(0) < oo) = OR(z) dz in (2. z + dz].2.T). z + dz].
T+dT]).2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . Tak'ecs [359]. cf.r(a). z + dz]. because of77>0. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .(3(B[r( a)] .1. who instead of the present direct proof gave two arguments. [329]. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. T(0) < oo) = dTP(T_(Z) E [T.2. Notes and references For Theorems 2. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.1) . r(0) < oo.(yo). see in addition to Prabhu [293] also Seal [326].1 Eear( y) = eyr(a).5 and one upon excursion theory for Markov processes (see IX.3 was noted by Asmussen & Kl(ippelberg [36]. I L Let ga(x) be the density of the measure E[ear(°).1) where a > r. Lemma 3.c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y).1 and the present proof is in the spirit of Ballot theorems. a martingale proof is in Delbaen & Haezendonck [103]. Theorem 2. some relevant references are Shtatland [338] and Gusak & Korolyuk [181].s.T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. Lemma 3 .6.108 Hence CHAPTER IV. 2. Proof Optional stopping of the martingale I er (a) 9 t. (3. Let T_ (y) be defined as Proposition 2. In the setting of general Levy processes. 3 Laplace transforms Throughout in this section.3. Proposition 2. Note that T_ (y) < oo a.5a). z + dz]. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx.
the result follows after simple algebra. time T(u): u u Here is a classical result : the double m.4 E[eaT (o). rr(0) < oo) = 1_ r(a) Proof Let b = 0.3. Then by Proposition 2. u . Hence eb"du E[eaT(").f.°° ga(x)dx. b . y + dy].1] evr(a)B(dy)[ b .2.ic(b)/b x(b) + a eb"E[eaT(" ). LAPLACE TRANSFORMS 109 Proof Let Z be the surplus . E[ear (o) I T(0) < oo . r(u) < oo).r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a.x)(a) B(dy)• Lemma 3 .3.3 ga[b] = c(b) Proof + b + a .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)) .ga [b] 1 .5 f 00 o a/r(a) .3. Further by Theorem 111. Corollary 3. Z = y] = EeaT. (Laplace transform) of the ruin Corollary 3.2 P(Z E [y.r(a) b .x)ga (x) dx where za(u) = f.(v) = ev''(a). (u .r(a) = a [B[b] B[r(a)]] .g. £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v.ST(o)_ just before ruin .T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 . T(u) < oo] du = Proof Define Za(u) = E [eaT(" ). It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.ga [b] 0 TO Using Lemma 3.
. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .2. Then given r(u) < 00. That is.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant).mu D 2 4 N(0.mL > E T(u) < 00 ) 40. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . For the second .h(u. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model.2 Assume ri < 0. note that by Wald's identity u + EC(u) = ES.110 CHAPTER IV. we need the following auxiliary result: Proposition 4. The first main result of the present section is that the value umL.(u ) = o(u) a. = (p .6. for any c > 0 P( Further.00 St = lim . St/t 1 1/m. P = /3µB > 1. for any m T(u) u .e.w ) v/. Theorem 4 .s. By Proposition 111.1. For the proof. T(u) a.1 Assume 77 > 0.. (4.3). uoo u using e. mu ) ( 0 m < ML '(u) 1 m > rL. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. i. and take basically the form of approximations and inequalities. This proves the first assertion of (4.1)Er(u) . Later results then deal with more precise and refined versions of this statement. the known results are even less explicit than for the exponential claims case.s. Then as u * oo.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. and hence a. (u) t. cf. Proposition A1. T(u)/u mL as u + oo.1) i.r(u) = Er(u) • ES.
1).7 6  11 Proof of Theorem 4. again Proposition A1. Notes and references Theorem 4.6µB2) Z v m (3µB2) Z. .6. note first that ( Proposition 111. of (4. cf.1. T) for T which are close to the critical value umL).1).1) is T (u)  U mL P( T (u) < I > E.2) follows immediately from u (4.mu (2) '• m3/2 µB 7 . PL (•)+ 0. apB ) .2 of [86]) and (4.s.2.1 The l. Tu) T( u) .1). the same conclusion holds with t replaced by r(u).t/m D (2) 111 . T(u) < oo f / 00) e7uE L [e_7 (t1).mu m . Thus. though it is not easy to attribute priority to any particular author. which may be viewed both as a refinement of Theorem 4. If Z .^ N (o. 4).mL >E By Proposition 4. proving (4. Theorem 7. and (4.h. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. According to Anscombe' s theorem (e.N(0. 1'r(U) .mL U > E. this can be rewritten as u + 1(u) .3.g. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. T (u) < 00 J 0(u) e7'PL U \ I T u) .1 (by considering 0(u.1 is standard. and as a timedependent version of the CramerLundberg approximation.5) St .3). For (4 . the result comes out not only by the present direct proof but also from any of the results in the following subsections.r(u)/m T(u) ti µB2) Z. implying T(u) .4.
3).w2) r. we need the following auxiliary result: Proposition 4.(u. Let h(u) = E f (^(u)). Then the distribution of T(u) .))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . then e(u) and r(u) are asymptotically independent in the sense that.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u .4). one has 9 (r(u)_rnu) Ef (^(u)) * E. (oo.112 CHAPTER IV. O .4 (SIAM'S LEMMA) If 71 < 0.um. Then for any y. using that ul/4 .6).3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).ul/4. E9(Z) (4.L+YWLV'U) . and thus in (4.t. resp .v. we get E[ T (u) .5) For the proof.)mu \ h(oo)Eg (r(ul) . Proof Define u' = u .T ( u')] = E[ T ( ul /4 . oo ).a C4'(y )• ( 4. e'°'/b (u. we can replace T(u) by r(u'). PROBABILITY OF RUIN IN FINITE TIME Corollary 4.r.VU T.6) whenever f.^(T(u')). Then h(u) 4 h(oo) = E f (6(oo)). letting Z be a N(0. oo). g are continuous and bounded on [0.f ( (oo)) . with w2 as in (4.T(u') given F.e(u') oo w . Using ( 4.l:(oo) (recall that rt < 0). and similarly as above we get E[f(^(u)) I Fr(u. Hence Ef (Vu )) 9 (T(u.ST( u') = u1/4 .u1/4)I(S(u') > u1 /4) h(oo) + 0. S( u ) < ul/4] < ET(ul / 4) = O(ul/4).mul h(oo)Eg(Z). P because of ^(u') .
oo ) as u * oo. just substitute T = umL + ywL in (4. CL Fig. . PL(T(u ) < umL + ywL) 113 4 C4(y). 10) '5(u) . Theorem 4. in practice one would trust (4. also Hoglund [204].(ay) = 17 7y = ay . 3 is due to Segerdahl [333]. u needs to be very large).oo.3 in terms of Edgeworth expansions . umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V ").yK(ay)• (4. 0. T(u) < umL + ywL f.7) to be good.5) and solve for y = y(T).5 '(u .1. however . Cf. Thus . For refinements of Corollary 4. y > k'(7) .7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . y u) < e 7v" . y u) < .3 ery"z/i(u . Notes and references Corollary 4 . ELe7E (") . see also von Bahr [55 ] and Gut [182].T) Ce7"4 (T .7) To arrive at this . 4b Gerber's time.z/)(u .9) ( 4 .umL wI V"U u (4.4) in the last. The precise condition for (4. The present proof is basically that of Siegmund [342]. For practical purposes . e7v" y < ^'(7) (4 .4. Segerdahl 's result suggests the approximation b(u. WHEN DOES RUIN OCCUR? Proof of Corollary 4.dependent version of Lundberg's inequality For y > 0. define ay. yy by 1 K. that for the fit of (4. where we used Stain's lemma in the third step and (4. see Asmussen [12] and Malinovskii [254].7) to be valid is that T varies with u in such a way that y(T) has a limit in (.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML).
which may be understood from Theorem 4.h(u.b (u. we arrive at the expression in (4.ay4(u)+ T(u)K(ay ).yu ) = eayuEav [e . the point is that we want to select an a which produces the largest possible exponent in the inequalities.2. In view of Theorem 4.7 i.3 yields easily the following sharpening of (4. we have rc(ay) < 0 and get (u) .8).1). yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. dy) Notes and references Theorem 4 . which shows that the correct rate of decay of tp(u.t.114 CHAPTER IV.Y' (u. f Some urther discussion is given in XI.9): Proposition 4. if y > 1/ic'(y). Then ic(ay) > 0 (see Fig . From the proof it is seen that this amounts to that a should maximize ayic(a). An easy combination with the proof of Theorem 111. u Differentiating w.8 below .r. the bound a7y° turns out to be rather crude . who used a martingale argument. 5 is due to Gerber [156 ]. see MartinLM [257] .6. yu) is e 'Yyu/ . PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y).v"U. a.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. 0. and hence t. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). Numerical comparisons are in Grandell [172 ]. yy is sometimes called the timedependent Lundberg exponent. However.5. . Hoglund [203] treats the renewal case. and generalizations to more general models are given in Chapter VI.8). yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. For a different proof.
we have ryas = ay .11) ' If y > 1/ r . yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.^3 ]1/ Bay [lay . then the solution &y < ay of .4. not inequalities. T(u) < yu] . [eT(u )K( ay). This idea is precisely what characterizes the saddlepoint method. (4. then ay > 0. the formula 0(u. For any a > yo. T(u) suggests heuristically that l t/..12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. Ea . u 4 oo.ay y 'Yay  ay . the choice of ay. (0) r1 (a) ' I. Using Lemma 111.(u. (4. As a motivation.8 If y < 1/ic'(ry). if we want EaT(u) .5. Proposition 4. then the relevant choice is precisely a = ay where y = T/u.ayC() .ay a.e.z. it is instructive to reinspect the choice of the change of measure in the proof.13) .e.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00.ayuEay f eay^ ( u)+T(u)K(ay).c(&) = ic(ay) is < 0. We thereby obtain that T is 'in the center' of the Padistribution of T(u). WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations.2 yields EaT(u) u u r. yu) = e..yu) c ay .6 with P replaced by Pay and FL by Pay. i.: T. and b(u. and ii(u) . yu ) eaauEaye .i(u. The traditional application of the saddlepoint method is to derive approximations. (4.. and in case of ruin probabilities the approach leads to the following result: Theorem 4 .'(y ).ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 .
B[ay] /ay &y y(ay .4).a) . .12) is 0 entirely similar.1)3 = y3/3B"[ay].13).a. The difficulties in making the proof precise is in part to show (4. Then ic(a) = . 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).l'B)y /(Pay .3(5/(S .1)3 = (jB"[ay]l (Pay .11) follows.I ay &y a ^c'(ay) a (1 +. PROBABILITY OF RUIN IN FINITE TIME ry I i . Writing r(u) and W2 = I3ay{. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV. (4.ay)K(ay) ay ayI&YI For the second term in (4. it seems tempting to apply the normal approxiyu + ul/2wV.7ruw2 Inserting these estimates in (4. V < 01 Ir 00 er(ay)"1'2"'x eyur. Example 4.13).c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.ay ) r.9 Assume that B(x) = eay.13) rigorously.1B[ay]1 ) y(ay .1) .a)2 . i B[7ay . where V is normal(0.116 CHAPTER IV.(ay) _ y(ay .c'(a) _ /3a/(8 . The proof of (4.ay + ayl /BLay] . a nr=. and the equation ic'(a) = 1/y is easily seen to have . we get heuristically that Eay Ler (u)r(ay).(j (1 .1.1) under Pay mation (4. (ay) J0 1 K(ay )u 1 00 c2(x) dx /2 w 1 ezcp(z /( k(ay)u1 /2w)) dz /O° _ 1 1 J e Z .ay) ay +.
1) ..p. in discrete time: if p = ES.11) gives the expression '31/4 ( . (5.tcp) Lo {Wo ( t)}t>0 . The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.= (s.. 0 Notes and references Theorem 4. and next to note that such an approximation in particular implies that the first passage probabilities are close. A related result appears in BarndorffNielsen & Schmidli [59]. then { __ .because the c. 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 =.g. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .f. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.ay)3 0 3/2 and (4./4 ^y for 1/i (u. 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.1. . 2 = Var(Si ) the variance. is the drift and o.. yu) when y < 1/ic'('y) = p/1 .8 is from Arfwedson [9].3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 .5. c a 00. is undefined for a > 5).i )( v s vc ('3 + s _2 / .
0 . 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. (5. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes). we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p... + {Wo(t ) .1.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.118 CHAPTER IV. p.t} _ {W_1(t)} .7c). We want an approximation of the claim surplus process itself.1 As p J. we have o {i!t s: . this is an easy consequence of (5. cf. and consider the limit p j p.1 below). Letting c = a2/pp. and this can be obtained under the assumption that the safety loading rt is small and positive.3) whenever c = cp f oo as p 1 p. oo)). n/c < t < (n + 1)/c.a = Snp) and the inequalities Sn )C . Indeed .tp). of which a particular case is the claim surplus process (see the proof of Theorem 5.2) t>o where p = pp = p .1) with S. a2 =/3µB2) Proof The first step is to note that { WC (St P) . Mathematically.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. where p is the critical premium rate APBTheorem 5 .z } {W_1(t )}t>o (5. such that the claim size distribution B and the Poisson rate a are the same for all p (i.p/c < St(p) < S((n+l)/ c + Pp/c.1)) is inconvenient.3.e. However. St = EN` U= . 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. Lemma 111.p. for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.
(ua2 To2 op \ IPI > IG ( T .2 As p j p. (.h. 119 It is wellknown (Corollary XI. any probability measure concentrated on the continuous functions.s. the continuity argument above does not generalize immediately.1.1 .5). 199. is 1/ip (ua2 /IpI. and in fact some additional arguments are needed to justify (5. (5. w. u) is defective when < 0.6) This is the same as the heavy traffic approximation derived in III. . ('.1. has a continuous distribution. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. Because of the direct argument in Chapter III. since ti(u) has infinite horizon .2 suggests the approximation u 0(u.5.e..8 or [APQ] p.r. Corollary 5 . 196. we obtain formally the approximation V. 263) that the distribution IG(•.. u).h.s. and the r. (5.1 I 7= . ulpI/a2).T) IG(Tp2/ a2). u) =PIT( (u) < x) = 1 . [169] or [APQ] pp. For practical purposes .(u) ti IG(oo.h. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. TS(u)=inf{t>0: WW(t)>u}.s.f I \\\ J \ (5. ulpl /a2) = e2"1µl / or2. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. we omit the details . C. 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) Note that letting T * oo in ( 5. 1.Ta2 /p2). see Grandell [ 168].( ^ I + e2( \ I . Corollary 5.6) from Theorem 5.. However.t.4) Note that IG(.u).7c. is IG(T. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a.
1. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. [169]. However. All material of this section can be found in these references. See for example Billingsley [64]. Assume further that 039µB6 < pe. in particular for large u. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. The first application in risk theory is Iglehart [207]. the B9. as 0 * 00 and that the U2 are uniformly integrable w. and two further standard references in the area are Grandell [168].00µB6 + 0.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9.6) therefore does not appear to of much practical relevance for the compound Poisson model. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. in the next subsection we shall derive a refinement of (5. the claim size distribution B9 and the premium rate p9 depends on 0.Pe. Furrer.120 CHAPTER IV. . the simplicity of (5.5) for the compound Poisson model which does not require much more computation. pt? 4 peo.6) are presented. as an example of such a generalization we mention the paper [129] by Emanuel et al. We conclude this section by giving a more general triangular array version of Theorem 5. In view of the excellent fit of the CramerLundberg approximation. In contrast. that 00 4090. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . we have ^A.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.5) and (5. The proof is a straightforward combination of the proof of Theorem 5..r. and which is much more precise. in Asmussen [12]. such that the Poisson rate Oe. For claims with infinite variance.t.Po = 09µB6 . Theorem 5. e. However. The picture which emerges is that the approximations are not terribly precise. pe .1 and Section VIII. (5. Then as 0 _+ 90. on the premium rule involving interest. B0 * Boo.6 of [APQ].g.
77 is close to zero. 0(0) = 0. Determine yo > 0 by r. .9(s) = Ico ( s + 9) . 3. For each 9.6. whereas there we let the given 3B.c(s) = . 9o T 0. .4.ao (0) _ /c(s + 9 . it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). The objective of the corrected diffusion approximation is to take this and other deficits into consideration. B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . 2. In terms of the given risk process with Poisson intensity .90) and the given risk process corresponds to Poo where 90 = 'yo. and we are studying b(u.6.'(yo) = 0 and let 90 = 'Yo.s and p = /3µB < 1. 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]. risk process with safety loading 77 > 0 correspond to 9 = 0 .1) .90] B(dx)./c(9 . and we want to consider the limit 77 10 corresponding to Oo f 0. P9(r (u) < oo) = 1 for 9 > 0. this idea ignores (among other things) the presence of the overshoot e(u). Then r. Bo(dx) = B[eo]B(dx). However . Since Brownian motion is skip free.1 > 0. this is because in the regime of the diffusion approximation . In this setup. which we have seen to play an important role for example for the CramerLundberg approximation .Q (B[s] . this means the following: 1.90) . Let PO refer to the risk process with parameters e9oz Qo = QB[90]. 77 = 1/p . claim size distribution B . PB('r(u ) < oo) < 1 for 9 < 0.T) = Peo(r(u) < T) for 90 < 0. The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. 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.
The first step in the derivation is to note that µ = k (0) = r0 (00) . IGu+u2. (6. the solution of r.1) IG(x. One has (6.. (. u) = euh(a . bl IG(t81.() The idea of the proof is to improve upon this by an O (u1) term (in the following.1) .e.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.u. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]. (U. (.7(u)/u2} eh(A. C . Vargo S. tu2 ) i IG (t. () where h (A.. . .. for brevity. 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 > 0.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. 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. means up to o(u1) terms): . u) = IG(x/u2.2) .C.3) this implies (take u = 1) Ego exp { . 1) • Since L eatIG (dt.Varo S1 = f30Eo U2 = S1. i. _ ^(u) = ST .T) 1+u2 (6.. C) = 2A + (2 . PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. 9otc0" (0) = 0061 = ul. Theorem 5.122 CHAPTER IV. C. The corrected diffusion approximation to be derived is (u.S. (01. write r = T(u). 0o to.(y) = 0.
s. 6 . however . the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.h. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.1 + 629.4. To arrive at (6. however. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . .5) according to (6. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} .2 ). .3 = 0.ry2 .7. The solid line represents the exact value .52/u where Z has distribution IG (•.v. 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. 1% in (2) and (4). p = 0.exp { h(A.s. in (3) and (4). A numerical illustration is given in Fig. distributed as Z . that whereas the proof of Proposition 6. Note. (6.v.1 As u + oo.5) Once this is established .f. bl I IG I t +2 . calculated using numerical integration and Proposition 1. But the Laplace transform of such a r.6. is the c.d.1 + u2 I Indeed.1 below is exact.3).z .3.2) is indeed o(u1). In ( 1) and (2).2). u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. 1. just replace t by Tb1/u2. and the dotted line the corrected diffusion approximation (6. which is based upon exponential claims with mean µB = 1. of a (defective) r. . the r. . 9o T 0 in such as way that C = Sou is fixed. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3).'yu /2)(1 + b2/u)} + Aug 1I J . we get by formal Laplace transform inversion that C 2 u.h. we have p =.
. For further numerical illustrations.7 or at values of Vi(u) like 1% is unsatisfying.T) 0.aa1 . The proof of Proposition 6. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.^) ..() Lemma 6.TI CHAPTER IV. the fit at p = 0.W21 0.114 0.01 0. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].08 a.1 W IU. OM 0.07 0.111 W(U. and all of the numerical studies the author knows of indicate that its fit at p = 0. PROBABILITY OF RUIN IN FINITE TIME 0. (Inc 0s 0.1 It is seen that the numerical fit is extraordinary for p = 0. Similarly.0 0.011 L1 60 T IM 11.7.124 0. see Asmussen [12]. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.T1 00.u2 2u3 (e . A51 7(SAT 3 3 h(X.2 e. .199 0.05{ 0.00 0.T) 111 0.EB 0 p ex p ( 7 S h ^)u .19)2 11 20 20 i0 T 1n0 Figure 6.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.OOIi O.1 proceeds in several steps.4 may not be outstanding but nevertheless.(061 0.08 0.
4) that the r.61a2T (B3 . exp ue } al 1J 3 exP I [2).6) u U3 Lemma 6 . the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy . (6.+ h (A..(3) J t _ aa1T l + eh(A. the result follows. () 62 Eeo exp u u2 J . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration . + a1b2 + . CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0. 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 .h. C) 1 1 + u2/ 111 + 2u CZ Z  (2A + ()1/2 J 1 Proof It follows by a suitable variant of Stam's lemma (Proposition 4.7) 2 2 .C)C/u .T (co (8/u) . in Lemma 6. 1 = PB(T < oo) = Eo0 exp 125 {(B .6. (6.r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .00)(u +C)  'r (.3 EoU2 + 103OoEoU3 + " 2 6 Using d2 . () + C and note that 2 KO (0) = 102..s.2u (B3 .(3)Eea LauT exp i 3J .2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .C2 = 2).co ((/u)) } Let 8 = (2a + (2)1/2 = h()..co (e) ..4 Ea.1) h(A.
h.2. PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.() .2u [2A+ (2 3 . [2+ (2 . 5 exp { _h(A) (1 + / y u J)) exp 1.e h(aS)h (^^ 262 exp {_h(. and inserting this and 9o 2 = S/u on the r. 2 and (6. yu/2). yields +90 62 0 + O(u 3) 2u2 +O(u 3).h (A. There are two reasons for this : in this way. () . and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6.2 (^/2 + 3y9o + 390) + O(u3). letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). 2 + 00 = . () by h(\.\+ (2 (3 e 2u [ (2. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6.6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4).\ + () 1 2 / . 0 The last step is to replace h(A. Thus by Taylor expansion around ( = 90u. Thus a2 y = 290 + O (u2).. C) ( 1+ u2 The result follows by combining Lemma 6 . we get h(A.x. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.2) for O(u) (indeed . yu/2) h(A.1 (y/2 + Oo)u . l Lemma 6 .s.6  d h(A.S) d e 62 .() I 1 + u2 ) y .(2A + ()1/21 exp S h(A.4.126 CHAPTER IV.7) and using eh(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. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. 0 1 Proof of Proposition 6. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. i. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. ()} 3 h (A. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. the 'typical' value (say in sense of the conditional mean) was umL. Fuh [148] considers the closely related case of discrete time Markov additive processes.e. 'yu/2) 127 ( i+ M pz^ exP { h (A. the approach to the finite horizon case is in part different and uses local central limit theorems. the same as for the unconditional Lundberg process.5 in Lemma 6. ()} . with the translation to risk processes being carried out by the author [12]. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. HOW DOES RUIN OCCUR? exp { h (x. In Siegmund's book [346]. and to the Markovmodulated model of Chapter VI in Asmussen [16].7. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. () I 1 + u 2 ) } S 1 .1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A.4.1: Just insert Lemma 6. the analogous analysis of finite horizon ruin probabilities O(u.(i+ 62 exP{ h(A. The adaptation to risk theory has not been carried out. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. The answer is similar: the process behaved as if it changed its whole distribution to FL.T) has not been carried out and seems nontrivial. () (i+a ) 2A + (2 . . that is.
FT(u) = o' (T(u ). Theorem 7 .EL[e7S. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. FL As example.2 If B is exponential. P(u) and rate = aL w.. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u.r. {ST(u)+t . Recall that . .(u) is not . the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E .t. (u) and satifying PL(F(u)) * 1. the Poisson rate changes from . . ^(u) is exponential with rate 8 w.(u)_ and ^(u) are independent .r.3 to . In fact. . 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. stating roughly that under F(u). u * oo. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . r(u) < oo) . Proof P(u) (F(u)c) = F(flu)c.(u)_ and similarly the denominator is exactly equal to Ce7u. + TMOO ).128 CHAPTER IV. {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u).TT(u) _measurable. we give a typical application of Theorem 7.1. so in the in the proof. Then also P(u)(F(u)) + 1.1 Let {F(u)}u>0 be any family of events with F(u) E F.3L and the claim size distribution from B to BL.t.(u). F(u)c] ti e' ru]PL (F(u)`) > 0.F.ST(u)}t> o is just an independent copy of {St}t>o).T. Note that basically the difference between FT(u) and ..(u)_ is that i. then P(u) and FL coincide on .vi(u) Ce'Yu Corollary 7.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.F. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL. In the exponential case.
(1 . the queueing results are of a somewhat different type because of the presence of reflection at 0. Notes and references The results of the present section are part of a more general study carried out by the author [11]. .3 M(u) pcu) 1 . From a mathematical point of view. This is currently a very active area of research. however. HOW DOES RUIN OCCUR? Corollary 7.3. the subject treated in this section leads into the area of large deviations theory. who also treated the heavytailed case. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x).eaLx. A somewhat similar study was carried out in the queueing setting by Anantharam [6].7. take I(Tk < x) .eALx) M(u) k=1 u The proof of the second is similar. see further XI. Proof For the first assertion.
This page is intentionally left blank .
The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). with the same distribution A (say) for T2. and the one corresponding to T1 = s by 1/i8 (u).Then no matter the distribution Al of T1i B. . .1 (T1 = a1)..d. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . Proposition 1.(1. D'2.. with common distribution B. the distribution Al of T1 is A as well. {St} is the claim surplus process given by I.. AA t*oo lim St = lim ESt t tioo t = p . U2. . are i. see A.. the one corresponding to the stationary case by 00)(u). A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. the claim sizes U1.Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1.. with Nt = # {n: Un <t} the number of arrivals before t.. T3. r(u) the time to ruin.Q. We use much of the same notation as in Chapter I.1. the Tn are independent. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory.1 Define p = !µ..1). . of the risk process form a renewal process: letting Tn = Qn .i. In the socalled zerodelayed case.7). and M is the maximum of {St}. Thus the premium rate is 1.
2 (DETERMINISTIC ARRIVALS) If A is degenerate. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U.1 of the safety loading appears reasonable here as well.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. CHAPTER V.132 Furthermore for any a > 0.t.1 gives the desired interpretation of the constant p as the expected claims per unit time.1) follows . we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U.St] = a(p . the definition 77 = 1/p . The simplest case is of course the Poisson case where A and Al are both exponential with rate 0.1). Proposition 1. t 4oo Proof Obviously. 3) follows similarly by Blackwell 's renewal theorem. RENEWAL ARRIVALS lim E [St+a . Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. say at a. A. and (1 . such that no arrivals occur in the off state. after E. s + t µA PA 0 Of course. Nt + EVar U.Nt] * a/PA. Example 1 . For (1 .0 > 0. the . This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). OFF.a is really the accumulated claims over a period u of length a. The renewal model is often referedd to as the Sparre Andersen process. but the arrival rate in the ON state is . From this ( 1. Here are two special cases of the renewal model with a similar direct interpretation: Example 1. However . Thus. by the elementary renewal theorem (cf.1) ENt/t + 1/µA. 2). stating that E[Nt+a . Nt ESt = E E UI Nt t = ENt•pB . If the environment is Markovian with transition rate A from on to off and u from OFF to ON.
INTRODUCTION 133 interarrival times become i.. 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 . we feel it reasonable to present at least some basic features of the model. we have From this the result immediately follows. Proof The essence of the argument is that ruin can only occur at claim times.4) with phase space {oN.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.. 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^.d.s. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. integrate (1.1.2. S o<t<oo n=0.. .y)B(dy).oFF}. if for nothing else then for the mathematical elegance of the subject.4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0.1.s < u).4) fo Indeed. the first term represents the probability F(U1 . in general the mechanism generating a renewal arrival process appears much harder to understand. Ao. A is phasetype (Example 1. and the present author agrees to a large extent to this criticism. More precisely. and then the whole process repeats itself). initial vector (1 0) and phase generator 11 However. Therefore. (1. u For later use.. However.i. the fundamental connections to the theory of queues and random walks..} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .r. U1 .T between a claim U and an interarrival time T. the relevance of the model has been questioned repeatedly.1. For the stationary case. and for historical reasons. arrival times. (an arrival occurs necessarily in the ON state.4) w..t. we note that the ruin probabilities for the delayed case T1 = s can be expressed as in terms of the ones for the zerodelayed case as u+8 z/i8(u) = B(u + s) + '( u + s .
1) . < 1. U Figure 2. That is . b=1 !=1 where {Nt } is a Poisson process with rate . (2. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.d. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. the remaining part of the prepayment (if any ) is made available to the company. RENEWAL ARRIVALS 2 Exponential claims. i. At the time of death . each of which receive a payment at constant rate during the lifetime . The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion .1) +ry. The initial reserve is obtained by prepayments from the policy holders.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) .3* (say ) and the U. are independent of {Nt} and i. 00). 2.134 CHAPTER V.1 r* (u) One situation where this model is argued to be relevant is life annuities. then 0 * (u) = 1 for all u > 0. t.a*PB• > 1.Ut. Using Lundberg conjugation . If . the claims and the premium rate are negative so that the risk reserve process .1. Theorem 2 . St = t .1 If. with common distribution B* (say) concentrated on (0. resp . the claim surplus process are given by Nt Nt Rt = u+^U. A typical sample path of {Rt } is illustrated in Fig. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] .3* pB.
2(a). 0 Now return to the renewal model. . Let B(dx) = ^e7x B*(dx). Define T_ (u) = inf It > 0 : St = u} .3*. then by Proposition 111.0. 2. Then the function k* is defined on the whole of (oo.g. > 1 . of {St} is c(a) = is*(a7). Fig. T_ (u) = inf { t > 0 : St = u 'r* (u). T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *.UB.2(b).(a) 7 Figure 2.s. Hence y exists and is unique.Rt. 2.1. cf. and the Lundberg conjugate of {St} is { St } and vice versa.2.(u ) < oo) = E {e7sr_ (u). (a) is*(a) (b) . Hence T_(u) < oo a..* (a) = log Ee'st I. B. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. If I3*pB* < 1. B*. St=Rtu=St. and thus 1 = P(T.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. Then the c. 0) and has typically the shape on Fig. Since ic'(0) < 0.2 Assume now . B* [7] and let {St} be a compound Poisson risk process with parameters . Proof Define 135 St =u .f. the safety loading of { St} is > 0.
f..)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[.1) + ry = 0 which is easily seen to be the same as (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 . Then B* = A.. u Hence P(M(d) > u) _ 1r+e'r".136 CHAPTER V. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 .. with rate S (say).1. alternatively termination occurs at a jump time (having rate 8). the failure rate of this process is y. and 5PA > 1..Un } = max St = t>0 n=0. RENEWAL ARRIVALS Theorem 2 . However.. respectively.. According to Theorem 2. 2.. then . Now the value of {St*} just before the nth claim is To +T1* +..'s and noting that V)*(u) = P(M* > u) so that Theorem 2. and from Fig . A variant of the last part of the proof..Y a I.e..+Tn U1 Un.7r+ 7r EeTo b/(Sa) + +.Tn} n=0.2 If B is exponential. 3* = 6. To + max {Ul+•••+UnTI..• • • . we get Ee'M(d) = Ee°M* _ Y/(. Taking m.Tr+.Y] (2.Ti = U1. with the probability that a particular jump time is not followed by any later maximum values being 1 . 1) means that 8(A[ry] .4..4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value.g.1. T2 = U2.1.u+ and lr+. and (2 . and hence the failure rate . which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.1 means that M* is exponentially distributed with rate ry.Ui . To + M(d) in the notation of Proposition 1..2)..a) = 1 .•.1 it is seen that ruin is equivalent to one of these values being > u. Hence M* max {To + Ti + • • • + Tn .
. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+. letting P(d) refer to the renewal risk model with these changed parameters .7r+) and hence r+ = 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^. the imbedded discrete time random walk and Markov additive processes.3. Putting this equal to y. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. 3a The imbedded random walk The key steps have already been carried out in Corollary 11.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". Hence the failure rate of M(') is 6(1 .3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1.7r+). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. resp. 111.. Furthermore.4. B^d) where Aad> (dt) = ^[ a] A(dt).6.5. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U . we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. a ladder step is the overshoot of a claim size. which states that for a given a. This follow since.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] . 0 3 Change of measure via exponential families We shall discuss two points of view. The probability that the first ladder step is finite is 7r+.2. hence exponential with rate b.. we see that ry = 6(1. : S(d) > u} .B(dx)..2. Bads (dx) = . However.y/b.
00)(u) . 7µA .p)/($B'[7] .2 In the zerodelayed case. For claim (b). (d) (7) _ 0..1 implies Cu) = e«uE ( 7d)e«^(u) .r. We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . Corollary 3.. Proof Proposition 3. VIII.1).2 p. 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 . cf. in the easiest nonexponential case where B is phasetype.C8e7u where Cs = Ce78B[7].Ce"u where C = limu. For the stationary case.T is nonlattice. and claim (a) follows immediately from this and e (u) > 0.. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) .3 For the delayed case Tl = s.1) is explicit given 7. Consider now the Lundberg case. In fact.e. let 7 > 0 be the solution of r.C(°)eryu where C(O) = C0[7] . 187) and thereby for ^(u) to be nonlattice w.4. we get: Proposition 3.(u) .t. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form. O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) .138 CHAPTER V. provided the distribution F of U .u the overshoot . i .1 For any a such that k(d)' (a) > 0. ik. just note that F7(d) is nonlattice when F is so . (a) '(u) < eryu. (b) V)(u) . E(d)e 1' (u). to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.
According to Remark 11. 3b Markov additive processes We take the Markov additive point of view of II. y) = e°yh(s). (s.dt(ah ( s) + h'(s)) so that Gha ( s. The expressions are slightly more complicated and we omit the details.1) (normalizing by h(0) = 1).dt ) eadt = h ( s) . we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). IPA 0 Of course. Let P8f E8 refer to the case Jo = s. we invoke the behavior at the 1 = h«(0.y) B(dy) 0 For the stationary case.(°) ( Ce8B[7] Ao(ds) similar manner.4). 0) = ah (s) . 0 0 .St)} can be defined by taking Jt as the residual time until the next arrival.(s. (u + s . where G is the infinitesimal generator of {Xt} = {(Jt. For s > 0./c.9. delayed version of Lundberg's inequality can be obtained in a e7u. B(x) = o(e7x) and dominated convergence. Here K. h(s) = e(a +x( a))8 (3.0 ) = Eo[ha ( Jdt. we look for a function h(s) and a k (both depending on a) such that Gh. 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] .5.a ..h'(s). St)} and h. Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt.Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds).. E8h0 (Jdt. To determine boundary 0. 0) = tc(a)h(s).5.3. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1.1) = C(O). Sdt) = h(s .
[e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +. Ba where Aa (dt) . Aa as asserted . (3.1)a in agreement u with Chapter III. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same.. . 5 For the compound Poisson case where A is exponential with rate .rc(a)] = 1..140 CHAPTER V. J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk. .. = J8 = T2.s governing {(Jt. . Remark 3 .c(a)] which shows that U1. [a + /3] A[b . rc(a) = 0 (B[a] . An easy extension of the argument shows that U1.rc(a)] = B = Ba[13]Aa[5]. resp. resp.13]A[b . Note that the changed distributions of A and B are in general not the same for Pa. we can now for each a define a new probability measure Pa. An important exception is.2) As in 11. ] = E. RENEWAL ARRIVALS B[a]A[a .2). Ease AU1+6T2 [ AU1+6T2 = Ea a LT.2) means 1 = B[a]/3/(/3+a+rc (a)). since JT. Further.8. .a . Proposition 3.e(«+k(a))t esy A(dt). St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. . T2 are independent with distributions Ba.tK( a)e.( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. i.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east .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. Ba(dx) = B(dx).s(Jo = s) = 1 follows trivially from Lo = 1..c(a)] B[a] Proof Pa..rc(a)] B[a] A[a . T2.e.. U. A[a .s and P(d). however. (3.4 The probability measure Pa. .a .5.
Then "^ e(ay+w(aY))8 Ys(u. the amount of . Then J(rr(u)) = TM(u)+1 and hence Ws(u.. A(ds). 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. [APQ] Ch. defined as the single server queue with first in first out (FIFO.5 Proposition 3. . yu ) < e7yu.yu ) e7vu A[ay .1 and n. . and assume that T„ is the time between the arrivals of customers n .5.t.y yuAa y [ay + K(ay) .)+1 e J j e(ay+w(ay))8 e . the time from he arrives to the queue till he starts service.g. The claim for the zerodelayed case follows by integration w. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. not after time T.4. it is easily seen that ic(ay ) > 0. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. Proof As in the proof of Theorem IV.r. that is. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue.. .(u. yu). and U„ the service time of customer n. yu) = F'ay. The virtual waiting time Vt at time t is the residual amount of work at time t. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). For the approach via Markov additive processes.yx(ay). Let M(u) be the number of claims leading to ruin . or FCFS = first come first served) queueing discipline and renewal interarrival times.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. and define yy = ay . XII.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular.s eaysr(")+r(u ) K(ay) h (s) .4. that is. which is the same as the asserted inequality for 0. say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims.ay+ray))TM(.4. Label the customers 1. u The approach via the embedded random walk is standard. for the zerodelayed case zp8(u. see e. 2. Using the Markov additive approach yields for example the following analogue of Theorem IV.
but we shall present a slightly different proof via the duality result given in Theorem II.1. and combining with (4.1) The following result shows that {Wn} is a Lindley process in the sense of II.1). and we have P(W > u) = V. .2. then Wn v M.1 Wn+1 = (Wn + U.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ ."^ Vi(N) (u). an+1) = [on. we get: Corollary 4.. p < 1.4: Proposition 4.+ Un. and obviously z/'(u) = limN. equivalently.1 and Corollary 11. the waiting time a customer would have if he arrived at time t. Let the T there be the random time UN. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival. Thus.3 Assume rl > 0 or.2) (b) as t * oo. equivalently.3. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. since customer n arrives at time on.3) Proof Part (a) is contained in Theorem 11. Thus Vos}1 _ = (Wn + Un .. the proposition follows.Tn)+. If W1 = 0. 0 Applying Theorem 11. (u). Vt converges in distribution to a random variable V. The traffic intensity of the queue is p = EU/ET.n1 (U1 +• • •+Uk Tl . 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 11.2 Let Mnd) = maxk=o. Then: (a) as n + oo..4): Proposition 4.4.1. (4.. Wn converges i n distribution to a random variable W.4.• • • Tk ).4. It then jumps to VQ„ .142 CHAPTER V. but interchanging the set . and we have P(V > u) = ?/iiol(u). (4. 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. (4. whereas in [On . on + Tn) the residual work decreases linearly until possibly zero is hit. in which case {V} remains at zero until time on+1. we have Wn = Van(left limit).
e. where U*.4. Then the corresponding queue is M/G/1. resp .oo in Proposition 4.T* are independent and distributed as U1.2). hence (since the residual lifetime at 0 and the age at T have the same distribution . Letting n oo in Corollary 4. Hence for x > 0. (4. namely W1 = 0 in (a) and Vo = 0. In fact . Corollary 4.4) 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. (4. we let T be deterministic . by an obvious reversibility argument this does not affect the distribution ..Ao in (b). However.T*)+. T1. K(x) = P(W < x). It follows that P(WN > u) =. For part (b).y)F(dy). x > 0. A.1. { 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). convergence in distribution hold for arbitrary initial conditions . THE DUALITY WITH QUEUEING THEORY 143 (T1.4 The steadystate actual waiting time W has the same distribution as M(d).. i.le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before . cf. Then the arrivals of {Rt} in [0.5) Proof Letting n . T] form a stationary renewal process with interarrival distribution A. and we get: .. Then K(x) = J x00K(x ..(N)(u) has the limit tp(u) for all u.. u Now return to the Poisson case .. Ti) and similarly for the U. as WN.T* = y yields K(x) = P ((W + U* .T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!).. which implies the convergence in distribution and (4.. we obtain: Corollary 4.T*)+ < x) = P(W + U* . we get W = (W + U* . T1 .. TN) with (TN. conditioning upon U* . and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival..2.5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x). but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.
144 CHAPTER V.5) is in fact a homogeneous WienerHopf equation. The equation (4. implying P(W > u) = P(V > u) for all u. That is.6 For the M/G/1 queue with p < 1. Note that (4. see e. W v V.5) looks like the convolution equation K = F * K but is not the same (one would need (4. Asmussen [24] and references there. RENEWAL ARRIVALS Corollary 4.g. . VIII). Hence '(u) = Ali(°)(u). 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e .5) to hold for all x E R and not just x > 0). the zerodelayed and the stationary renewal processes are identical. despite the fact that the extension from M/G/1 is of equally doubtful relevance as we argued in Section 1 to be the case in risk theory. the actual and the virtual waiting time have the same distribution in the steady state. Some early classical papers are Smith [350] and Lindley [246]. Cohen [88] or [APQ] Ch. Proof For the Poisson case.g.
here it exists whenever A is irreducible which is assumed throughout. • Claims arriving when Jt = i have distribution Bi. and can be computed as the positive solution of WA = 0. .f pi. M = supt>o St. t St = E Ui . Ire = 1. {Jt} describes the environmental conditions for the risk process. 145 Oj( u.)iJEE and its stationary limiting distribution by lr.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 . As in Chapter I. The intensity matrix governing {Jt} is denoted by A = (A. dv. {St} denotes the claim surplus process.T) = Pi (T(u) < T). • The premium rate when Jt = i is pi.. Thus. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). N.(3i when Jt = i. i=1 0 and r(u) = inf It > 0: St > u}.
According to Theorem A5. with rates Ani and Ain. = iii when j E E(i). Thus. Example 1. this is no restriction when studying infinite horizon ruin probabilities. /3 = Nn when j E E(n). assume that the sojourn time in the icy state has a more general distribution A(i). Bi. and p is the overall average amount of claims per unit time. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. 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. we can approximate A(i) with a phasetype distribution (cf. T(=)). one expects that 3i > on and presumably also that Bn # Bi. r^ = P (1. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. respectively. in blockpartitioned form. Then the state space for the environment is the disjoint union of E(n) and E(i). An example of how such a mechanism could be relevant in risk theory follows.2. We let p Pi = /ji/AB. f3i and claim size distributions Bn.5 below. say. cf. For example.14. Proposition 1.. Example 1 .1 Consider car insurance.11 below.a('). i and corresponding arrival intensities Qn.4) with representation (E(i). which is clearly unrealistic. Unless otherwise stated. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). we shall assume that pi = 1.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. and assume that weather conditions play a major role for the occurence of accidents. a(i).2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. we could distinguish between normal and icy road conditions. t(i) = T(')e are the exit rates. the operational time argument given in Example 1. cf.146 CHAPTER VI. leading to E having two states n. say. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential.T(n)). u . P = E 7riPi. and we have f3.
St = SB=(t).J017. and . we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. the state space E for the environment is { ('q.. iq (visited in that order) and letfOil >. Approximating each A('?) by a phasetype distribution with representation (E('l).. Indeed. Example VIII. MODEL AND EXAMPLES 147 Example 1 . and similarly for the normal period.3 Consider again the alternating renewal model for car insurance in Example 1.. A('^).j = .4) with states i1..Q. u Example 1 . n8}. it = Jel(t).. In the car insurance example.n. dt..tEH is a transition matrix. such that a sojourn time of type rt is followed by one of type c w. but assume now that the arrival intensity changes during the icy period.T(n)). u From now on. and 1/ii(u) = t/ii(u).2. is the probability that a long icy period is followed by a short normal one. . u Example 1. 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. resp.. Then for example wi. i E E(n) }.1.p.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. say it is larger initially. say. where W = (w..3i/pi. i8f n1.. let T 9(T) = f pi.3i. T(9) +wggt(9)0. depending only on 77. t(n) = T("i)e..3.>.. One way to model this would be to take A(') to be Coxian (cf.a(n). one could for example have H = {i1.1. (9) where q = CHI. The simplest model for the arrival intensity amounts to . the parameters are ^ij = aid/pi. 1 .. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. This amounts to a family (A(")) ?CH Of sojourn time distributions. 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) . Qi = . . such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). i ) : n E H. w.
one can associate in a natural way a standard Poisson one by averaging over the environment. Pi (Ti E dx. N > 1(Ul < x) a4 B*(x). B* = 1 /^* Bi. Next we note a semiMarkov structure of the arrival process: Proposition 1. More precisely. 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).148 CHAPTER VI.e(A(Oi)d'sg)xe. )3*. qij = 0 in the notation of Chapter 11. 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. we put )3* = E 7fi/3i. . dx.(3iBi(dx). 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. . Nt Nt a . vi(dx) = . the empirical distribution of the claims is B*.(Qi)diag)• More precisely. JT1 = j) = Qj • e.8 As t oo.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.5. In particular. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.A . the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. t l=1 Note that the last statement of the proposition just means that in the limit. o = 0. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. The key property for much of the analysis presented below is the following immediate observation: Proposition 1.7 The Pidistribution of T1 is phasetype with representation (ei.
given {Jt}0<t<0. ^j 7riNi.7. Bi(x). B*. cf.* (u) for all u. 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$.. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. 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. A. and let {St °i} refer to the one with parameters Pi.. Hence Nt'> a . Then it is standard that ti lt '4' iri as t > oo. has distribution (7ri)3i //3*)iEE and is independent of Ti. In particular. aA. oo) as a 4 oo. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <..(/3i)aiag).. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. Example 11. e. Conditioning . this converges to the exponential distribution with rate 0* as a * oo. and furthermore in the limit JT. {St} to the compound Poisson model with parameters 0 *.6. By Proposition A5.2. Nt a' t t iEE Also. Bi.1. In particular.9 Consider a Markovmodulated risk process {St} with param eters Ni. y Ni) i Nti) t a. 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 . Bi. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti. Proposition 1.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). Proof According to Proposition 1. zli( (u) .. However . the limiting distribution of the first claim size U1 is B*.4. Then {St)} + {St*} in D[0.aA . i.
.1 of [145]. 0 Example 1. That is.s = o in state 1 and with intensity 1 . A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier).2 +2 2 = 3. T) + ?P* (u. On Fig. state 1 appears as more dangerous than state 2. U2) are independent of .. the company even suffers an average loss.l3* and U2 having distribution B*.2. U. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary.. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7.).FT. we first get that 3 (3* = 2. resp. MARKOVIAN ENVIRONMENT upon FT. > 1.. and (at least when a is small such that state changes of the environment are infrequent).31µB 2 = 2 5 3 7 70 Thus in state 1 where p. marked by thin. B1=3E3+2E7. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7..1 with periods with positive drift alternating with periods with negative drift. s 5 in state 2.2}. shows similarly that in the limit (T2. 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 = . B2=1E3+4E7. is as in {St }. T) for all u and T. Claims of type E3 arrive with intensity 2 . oo). e.1. the paths of the surplus process will exhibit the type of behaviour in Fig. From this the convergence in distribution follows by general facts on weak convergence in D[0..g. from Theorem 3. The fact that indeed 0(a) (U) 3 0* (u) follows.10 Let E_{1. which also yield O(a) (u. lines in the path of {St}.. Thus.=1. Computing the parameters of the averaged compound Poisson model. with T2 being exponential with rate . Continuing in this manner shows that the limiting distribution of (T. 1. 1.2. 132=2..s = 1o in state 2. 9 . there are p = 2 background states of {Jt}. those of type E7 with intensity z s = 5 in state 1 and with intensity z .150 CHAPTER VI.. thick.
1).1. = P. note first that EN Uk')/N a4' µgi. iEE .8.(3. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs.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. 0 The definition (1. Proof In the notation of Proposition 1.11 (a ) ESt/t * p .1 Thus. t * oo. For (b). t + oo. a fraction r. 01 /. the averaged compound Poisson model is the same as in III. That is.1.1 a. 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. 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). MODEL AND EXAMPLES 151 Figure 1.s. (b) St/t * p ..
EiX = 0. + Xn SWn ](1 a . let some state i be fixed and define w=wl=inf{t >0:Jt_#i.s. Since the X„ are independent . X2 =SW2 So. Theorem II. then M < oo a.s.12 If 77 < 0.0i(u) < 1 for all i and u.Jt=i}. Now let r) = 0. The proof of Proposition 1... also + .a form a renewal process . The case 77 > 0 is similarly easy.. Proposition 1. [212].4.. and a more comprehensive treatment in Asmussen [16]. then M = 00 a. Then by standard Markov process formulas (e. Now obviously the w. MARKOVIAN ENVIRONMENT Corollary 1.ld. and hence wn /n a4.. and . X 1 =Sty.1 jEE = (p . [315]. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. PB. see the Notes to Section 7. and hence 1/ii(u ) = 1 for all i and u. . limit p . [302]. dt .1)Eiw = 0. and involves a version of the . n n Thus {SWn l is a discrete time random walk with mean zero. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model.. If 77 > 0.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes. and hence M = 00. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.. see [APQ] p. ..1 of St / t is > 0.2(a) p.g.. having the Pidistribution of X. 38) Eiw1 = 1/ir. In risk theory. [APQ]. w2=inf {t>w1:Jt_#i. s. with X2.152 CHAPTER VI..Eiw o'o Eiw • E ^ifjµs.1 and the Corollary are standard. The mainstream of the present chapter follows [16]. Proof The case 77 < 0 is trivial since then the a. X3. 136 or A. some early studies are in Janssen & Reinhard [211]. and hence oscillates between 0o and oo so that also here M = oo.Jt=i}. Statistical aspects are not treated here. Eiw. There seems still to be more to be done in this area. and so on.\ i and EiX1 Ei f 13 J. See Meier [258] and Ryden [314].
dx). n=0 (2. T R(i. j. •) * G +(k. oo)). but is substantially more involved than for the compound Poisson case .6. Proposition 2. j.1) 0 (b) G+ (y. j.2 (a) The distribution of M is given by 00 1 . e.. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. That is.i.j. (y. However . For measurevalued matrices. let G+(i. the definition of . T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i. we define the convolution operation by the same rule as for multiplication of realvalued matrices.A) = Pt(ST+ E A.2) R(dx)S((y . j.1 irG+(dy)e =.j. B* in Section 1. oo)) = f R(i. IIG+ II denotes the matrix with ijth element IIG+(i. Let further R denote the preT+ occupation kernel. •)• kEE Also.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument.3*B *(y)dy.x.x). 6.Jt=j)dt. k. dx)/jBj(y .(u) = Pi(M < u) = e' E G+ (u)(I .Jr+ =j.IIG +II)e.2. cf. for i. •) II = JG+(i. .A) =ZI(St E. by specializing results for general stationary risk processes (Theorem II . we get the same ladder height distribution as for the averaged compound Poisson model.EA. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0).5. see also Example II. oo) = J ao 0 G+(i. j. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. The form of G+ turns out to be explicit (or at least computable). G+ is the matrix whose ijth element is E G +(i. only with the product of real numbers replaced by convolution of measures.g.j E E.a/i.4) we obtain the following result . Thus. •). Proposition 2. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2.6*. 00 (2.
St < S* for u < t. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. JJ = j. The u proof of (2. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. From this (2.6.1 The following observation is immediate: Proposition 2. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). marked by thin. resp.154 CHAPTER VI.2) is just the same as the proof of Lemma 11. thick. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value. we need to invoke the timereversed version {Jt } of {Jt} . 0  x Figure 2. III to bring R and G+ on a more explicit form . see Figure 2. we need as in Chapters II. .2 useful .3 When q > 0.3.IIG+II)e. To make Proposition 2. mx = j when for some (necessarily unique) t we have St = x. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . lines in the path of {St}. and that the environment is j at the nth when we start from i is e . To this end .3) We let {St*} be defined as {St}. {mx} is a non terminating Markov process on E.1) follows by summing over n and j.1 for an illustration in the case of p = 2 environmental states of {Jt}. That is. hence uniquely specified by its intensity matrix Q (say).
Q( n+l) _ ^.2 where there are three excursions of depth 1. we say that the excursion has depth 0.(/3i) diag. 2.0. } is a minimum value at v = t. An excursion of {St*} above level x starts at time t if St = x. 0 mms1   ^O \ T. the sequence {Q(n)} A* defined by Q(O) = . If there are no jumps in (t. s]. {S. and the excursion ends at time s = inf {v > t : S.*. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2. Furthermore. Proof The argument relies on an interpretation in terms of excursions. Otherwise each jump at a minimum level during the excursion starts a subexcursion. The definitions are illustrated on Fig. ( Q( n)) converges monotonically to Q. = x}.2.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .2 . For example the excursion of depth 2 has one subexcursion which is of depth 1..and a jump (claim arrival) occurs at time t. and the excursion is said to have depth 1 if each of these subexcursions have depth 0.2. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. Figure 2.(/3i)diag + T S(dx) eQx. In general. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. 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). corresponding to two subexcursions of depth 0.
A) = f Pi(mx = j) dx eie4xej dx A u (2. Similarly.T+>t) _ ^iF 7ri (JJ =i. h.5 R(i. we first compute qij for i $ j.j +/3ipij. It follows that qij = A.j. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . mx+dx = j) occurs in two ways . The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). 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.6) . or through an arrival starting an excursion terminating with J. Writing out in matrix notation . either due to a jump of {Jt } which occurs with intensity A= j. Similarly by induction . 0)).5) A (note that we use A = {x : x E Al on the r. StEA . j. p1^) Define a further kernel U by f U(i. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs ..4). of the definition to make U be concentrated on (co. e. = j.u< t).Qi + )%pij) Now just note that t pij and insert (2. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. 7rE Proof We shall show that Fi(Jt=j. Then a jump to j (i. Theorem 2 . A). By considering minimum values within the excursion. Suppose mx = i. (2.St <S*.(01)diag = Q..4) To show Q = cp(Q).St EA.156 CHAPTER VI. A) = L' U(j.s. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. Q = W(Q) follows. i. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii.
Jo=i. (b) for z > 0. oo)) = f o' eIXS((x + z... we shall see that nevertheless we have enough information to derive. and this immediately yields (2.6 is hardly all that explicit in general.StEA.7 It is instructive to see how Proposition 2.. To this end. e. oo))dx.2. G+((z. x < 0.r.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. dt. and get irPi(Jt =j. We may then assume Ju=Jtu.6). and we let k be the corresponding right eigenvector normalized by Irk = 1.0<u<t) = P.. From Qe = 0. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2.St <Su. St E A. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w.S„<0.Qi)diag.(Jt=j. St < St U.z+>t) = P.0<u<t.(. Jt = i. where A is the diagonal matrix with 7r on the diagonal: Corollary 2. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.g. K( n (d) the sequence converges monotonically to K. (Jo = j. consider stationary versions of {Jt}.t.St EA. {Jt }. Remark 2.6 (a) R(dx) = eKxdx. St EA.4].1 can be rederived using the more detailed form of G+ in Corollary 2. 0<u<t). the CramerLundberg approximation (Section 3). 0 < u < t) = 7rjPj(Jt =i. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). S. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx).. 0 +1) = cp (K( n)) defined by K(o) = A .=StSt. and to obtain a simple solution in the .
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
j.V)i(u.i I (a) exp {aye(u) + r(u)r. However. 1 Similarly. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i.(ay)}. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .9) For the proof. yu) f h(av)e v avuE«v. we shall need the matrices G+ and R of Section 2. Chie ryu < Vi(u ) < C+hie 7u. we have ic(ay) < 0 and get 'i(u) .7. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE.3. (3. av 'i [h. if y > 1lk'(ry). as in the classical case (3. exp {e() + r(u))} . r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).00 su e7( ( 3.i [eT(u)K(av ). for a vector <p(u) = (cpi (u))iEE of functions .y)G+(z.j) * coj)(u) _ f u ^Pj(u .5).5) will produce the maximal ryy for which the argument works. 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 3.8 ) Then for all i E E and all u > 0. oo)) and. dy)• o iEE jEE . (u. hj P .11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy). we let G+ * W(u) be the vector with ith component E(G+(i. 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. r(u) yu o)(y)eavuEav.i [e*(u)K(av)..
dy) 00 C+ ijhj f R(i. j.x ) R(i.x) x) jEE 0 E Qj f jEE R(i.x ) = Gi(u). just note that the recursion <p(n+1) = G + G+ * (p(n) holds for the particular case where cpin)(u) is the probability of ruin after at most n ladder steps and that then obviously u cp2n) (u) + t. = Eo G+ G. dx) f e7( vu)Bj (dy . U = U". dx). 00 Thus C+ > hj f"o e7(Yu)G +(i. jEE u 0 j. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u).j. dy) : 1(u) < C+ > hj u e(1tL)G+(i.(0) ] (u) < sup Jt t.166 CHAPTER VI. MARKOVIAN ENVIRONMENT Lemma 3 .3jhj // f 00 R(i.7. Proof Write UN = EN G+ .dy). 0 G+(i. n > oo.x) jEE 00 u 0 //^^ C+E.13 For all i and u. j. dx) 100 C . if r+ (n) is the nth ladder epoch. Lemma 3 . j. dy) = aj f Bj(dy .u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. dx ) Bj (u . we have G *(N +1) * ^. 00 f C_ hj f e(Y)G+(i. j. Hence lim cp(n) exists and equals U * G.& (u). _ To see that the ith component of U * G(u) equals ?Pi (u).ery(&u+x)Bj (dy) Bj(u Bj (u . °O . Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However.u IMP:°) (u) I < oo.j.12 Assume sup1. Then cpin)(u) sit (u) as n + oo.
Pi(MT > u) = Pi(MT < u. letting MT = maxo<t<T St. it follows that Vi(u) < C_(yo) h=70)e7ou. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.3. Then 0< Vi (u )  0i(u.y)G+(i. T) < C+(')' o)hi7u)e7ou8T .u)G+(i. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. We claim by induction that then cpin) (u) > C_ hie7u for all n.10: Theorem 3 . (3. j.MT<u. MT < u. we have Vii (u) . and let 8 = e'(70). u The proof of the upper inequality is similar .10 ) by Lemma 3 . and assuming it shown for n.13 Let first cp=°)(u) = C_ hie"u in Lemma 3. 13 and the second by the induction hypothesis . 167 u Proof of Theorem 3.(u) < T ) to 0i (u) which is different from Theorem 3.13) Hence. jEE estimating the first term in (3. and using Lemma 3 . and the proof of the lower one is similar.8) with y replaced by yo and hi by h=7o ). j. j. Indeed.10) C_ 1 f hje7(y. from which the lower inequality follows by letting n * oo. taking cps°) (u) = 0.M > u) = Pi(ST<u.ST).13. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. (3. y]hj = C_ e7uhi.11). this is obvious if n = 0. 14 Let yo > 0 be the solution of 'c'(yo ) = 0.M>u) = Ei [VGJT (u .11.11) C_e7u 57 O+[i. let C+(yo) be as in (3.tpi(u.T) = Pi(M > u) .12) Proof We first note that just as in the proof of Theorem 3. j. 9 for the last equality in (3. dy) jEE o (3. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . dy) jEE u U +C_ hje7( yu)G jEE"" +(i. +i . dy) (3.n) ( u . T) = Pi (7.
The motivation that such a result should be true came in part from numerical studies. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. [177]..33 or Bi 0 Bj.3) to B = Bi does not depend on i.3) Bl <_s.5) Note that whereas (4.. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. The results to be presented show that quite often this is so. we also assume that there exist i # j such that either /3i <. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). Occasionally we strengthen (4. Bp.o. we define the stochastic ordering by 0' < s. V)" if z/i'(u) <'c"" (u)..1) Obviously.. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u). (4. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'.0.4) To avoid trivialities. Further related discussion is given in Grigelionis [176]. ". we refer to . B2 <_s. and finally in part from queueing theory.3p. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and .o. this correponds to the usual stochastic ordering of the maxima M'.2) alone just amounts to an ordering of the states. It was long conjectured that 0* Vi.168 CHAPTER VI. For the notion of monotone Markov processes.2) (4. (4. The Markov process {Jt} is stochastically monotone (4. in part from the folklore principle that any added stochastic variation increases the risk. this is not the case for (4. < . (4.o.. <s. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). but that in general the picture is more diverse. u > 0. The conditions which play a role in the following are: .3).o.31:5)32 ....
. . (b) P.9 ) below). Theorem 4 ..6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. Lemma 4 ..9) (4. Conditioning upon the first ladder epoch. note that (4... Then V.2)(4..5 (cf.4) say basically that if i < j . = b. < a.2.r(u x)dx. then j is the more risky state . dx (4. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]. where 7r2+) = QiµBilri/p. Proof of Theorem 4. E 7r i Wi(u . Comparing (4..6.6).. (4.2)(4.3 for the second) *(u) _ /3 *B* (u) +.1. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.3iBi(x)YPi(u . then P P P 7rjbj. b1 < . the second follows from an extension of Theorem I1. 7(0) < oo) = pirf+).1) which with basically the same proof can be found in Asmussen & Schmidt [49]. Section 4.4. Proposition 2. Lemma 4 . ^i 7ri = 1. we obtain (cf..x)B*(x) dx.10) Q*B*(u)+.* For the proof.r (Sr(o) E dx Jr(o) = i. we need two lemmas. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = ..10) and (4.x) of i and using Lemma 4. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. also Proposition 2.. 2 If al < . it follows by a standard ..x) dx u o i =1 i=1 (4.3* f uB(x) z/^... < bp and 7ri > 0 (i = 1. 3 (a) P.4) hold.7) and Lemma 4. = aP or b1 = .1 Assume that conditions (4.1 for the first term in (4.9) follows by considering the increasing functions 3iBi (x) and Oi (u . p). 0 Here (4. T(0) < oo) = Bi(x) dx/tcai ..8) ^j Tri/iBd(x) . Conditions (4.r (JT(o) = i. Proschan & Walkup [140].x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.2.13* J0 u 0*(u .7) 7ri.4) is automatic in some simple examples like birthdeath processes or p = 2 .
s.h.6). let = ( 1/2 1/2 ) .* (0). except possibly for a very special situation . (u) is not in general true: Proposition 4.3µi < 1 for all i. Frey. (4. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. What is missing in relation to Theorem 4. of order 101. u To see that Proposition 4. Using (4. (u) may fail for some u. µB2 = 104. this ruin probability is /3iPBi. of (4.3i. i=1 i=1 7'r(0) _ EFioiwi(0) . µB. Then the l. As is seen.4 is the understanding of whether the stochastic monotonicity condition (4.h.4) is essential (the present author conjectures it is). 01 = 103.(0) < b *'(0) for e small enough. Then i/i*(u) < . Proof Since 0. it will hold for all sufficiently large u.../3*. u Here is a counterexample showing that the inequality tp* (u) < V)..0.8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi . Bi as e J.0*• i=1 But it is intuitively clear (see Theorem 3. Recall that .. 0.1 and Proposition 4.2.170 CHAPTER VI. it is sufficient to show that 0'. they are at present not quite complete. Q2 = 1. Rolski & Schmidt [32].1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . and from this the claim follows.r (u ) fails for all sufficiently small e > 0.4 is not vacuous. dominates the solution 0* to the renewal equation (4. For u = 0. MARKOVIAN ENVIRONMENT argument from renewal theory that tk. Notes and references The results are from Asmussen. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *.4 Assume that . = 102.11) i=1` and that A has the form eAo for some fixed intensity matrix A0.6).s..(0) = V.11) is of order 104 and the r.
1) . Then {(Jt. (4. Lemma 4.2 we have (Ei[e"X'. It is clear that the distribution of X. Asmussen [20]) as discussed in 11.14) is nonzero so that A"(0) > 0. 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] . (4.a = E irirci(a).5 y < ry*. it follows by Proposition A1.13) (4.)a.5. Xt)} is a Markov additive process (a socalled Markovian fluid model. Now we can view {Xt} as a cumulative process (see A.4(b) that the limit in (4..g. and by Proposition II.. 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.5.5.(a) > 0 for all a 0 0.1) .14) A„(O) iioo varXt t t By convexity.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) > 0.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. with strict inequality unless a = 0 or bi = 0 for all i E E. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state. with strict inequality unless rci (y*) does not depend on iEE. is nondegenerate unless bi does not depend on i E E. 0 .ld) with generic cycle w = inf{t>0: Jt_54 k. Further (see Corollary 11. Jt = i])' EE = vA+n(6. This implies that A is strictly convex. (4. e. cf.g. Proof Define X= f &ids.13) implies A(a) > 0 for all a. in particular . Hence if 5i 54 0 for some i E E.4.12) iEE Theorem 4.
h'(0) = (Ao . whereas the .. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. y. Frey. the basic equation is (A + (rci(y))diag)h = 0..15) Normalizing h by 7rh = 0.12) and rc*(y*) = 0. h depend on the parameter (e or a). 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. Notes and references Theorem 4. In the case of e. note that y(a) + mins=1.5. Then > risi = 0 because of (4.. If rci(y* ) is not a constant function of i E E. (4. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. where A.e7r)1 (Ici(Y*))diage. Here we put a = 1/e. Let bi = rci(y*).172 CHAPTER VI.) and rc (•).. MARKOVIAN ENVIRONMENT Proof of Theorem 4.Qi and Bi are fixed . improving upon more incomplete results from Asmussen. a = 1 in Lemma 4.5 is from Asmussen & O'Cinneide [40]. (4. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. h(0) = e.p yi and compute 8y 8a a=0 In both cases. we get rc (y*) > 0 which in a similar manner implies that u y < y*. Thus y(e) * y* as e 10. Hence rc (y*) > 0. 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(.16) Differentiating (4. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e.6.eir)h'(0). we have 7rh' = 0. Rolski & Schmidt [32].15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . Hence letting e = 0 in (4. Since ic is convex with rc'(0) < 0 .
5. i = 2..8 If (4. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . We assume that 0 < y < 7i. and we have proved: Proposition 4.19) holds.18).17) by 7r to the left to get (4. which has recently received much attention in the queueing literature. Rolski & Schmidt [32]. (4. The analogue of Proposition 4. Inserting (4.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). . 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. 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. We get 0 = (aAo + ( lc&Y))diag)h. and may have some relevance in risk theory as well (though this still remains to be implemented).20) Letting a = 0 in (4. (4.18) 0 = 27'(0)p+27r(rs. (4. Frey.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).8 when ryi < 0 for some i is open.17) (4. p..16) yields Proposition 4. then 8a a=o All rci (0) Notes and references The results are from Asmussen. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) .i(7' *))diagh'(0). 5 The Markovian arrival process We shall here briefly survey an extension of the model. . multiplying (4.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a.
II.2) A(1) = A(' 1) ® A(1. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. B. j(2) } be two independent environmental processes and let E(k). Note that the case that 0 < qij < 1. This is the only way in which arrivals can occur. the definition of Bij is redundant for i i4 j. In the above setting. Again . The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. T).4). .6i ) diag.2). refer to notation) { Jt k) }.2 for details). and thus 1i = 0. Indeed. Jt2)) (2. Jt = (Jtl). A(1) = A . Thus . we use the convention that a1i = f3i where 3i is the Poisson rate in state i.174 CHAPTER VI.i. Bij = B.^) etc.(13i )diag. where qij is the probability that a transition i * j is accompanied by a claim with distribution. Here are some main examples: Example 5 . with common distribution B. let { Jt 1) }. For i = j.d. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . u Example 5 . the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. the claim surplus is a Markov additive process (cf. the definition of Bi is redundant because of f3i = 0. A(l) = tv. Bii = Bi . A(1'k) A(2 k1). we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. and the marked transitions are then the ones corresponding to arrivals. that Bii = Bi . A(l) = T.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . We then let (see the Appendix for the Kronecker E = E(1) x E(2). A ( 2) = A (2`1 ) ® A. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered.
after which it starts afresh.1i2 . In this way we can model... as the Markovian arrival process ( MAP).iN = a2. superpositions of renewal processes. DIVORCED.. 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. Thus. The versatility of the setup is even greater than for the Markovmodulated model..kl is redundant). u Example 5 . This means that the environmental states are of the form i1i2 • • • iN with il. the kth policy enters a recovering state. i2i . Similarly. THE MARKOVIAN ARRIVAL PROCESS Bij. DEAD etc. 11. more recently. e..iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. Example 5 .iN are zero and all Bi are redundant.3 (AN INDIVIDUAL MODEL) In contrast to the collective assumptions (which underly most of the topics treated sofar in this book and lead to Poisson arrivals). assume that there is a finite number N of policies. and that the policy then expires.iil. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. iN = all BOi2. with rate ai. E = { WORKING.1i2. INVALIDIZED... iN.. claims occur only at state transitions for the environment so that AN2. WIDOWED. RETIRED. MARRIED.iil. • upon a claim. 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. possibly having a general phasetype sojourn time.}..iN.g.. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. all Al i2. E 10...kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. say. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential... However ...iN. or..5.iN C17 AilO.. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. Bilo.. 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 .. iN.
We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process.p)/p.176 CHAPTER VI. Lucantoni et at. Neuts [271] and Asmussen & Perry [42]. • Claims arriving at time t of the year have distribution B(t). Some main queueing references using the MAP are Ramaswami [298]. Sengupta [336]. we may assume that the functions /3(t).1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . 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 . Lucantoni [248]. where i f00 xB(°) (dx) _ . Obviously. 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. one needs to assume also (as a minimum) that they are measurable in t. 1). • The premium rate at time t of the year is p(t).3*µs • p = f /3(v) dv 0 0 (6. p(t) and B(t) are defined also for t t [0. for s E E = [0. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . B* = J f B(t) ((*) dt. By periodic extension. . continuity would hold in presumably all reasonable examples. Let 1 1 /3* _ f /3(t) dt. let the period be 1. 6 Risk theory in a periodic environment 6a The model We assume as in the previous part of the chapter that the arrival mechanism has a certain timeinhomogeneity. we talk of s as the 'time of the year'. )3 t 1 J (6. Thus at time t the premium rate is p(s + t). one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. 1). p * = 0 p(t) dt. Without loss of generality.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. [248]. 0 < t < 1. For the Markovmodulated model. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones.
and thus the averaged standard compound Poisson models have the same risk for all A. B*. respectively. Example 6 . in agreement with the general principle of added variation increasing the risk (cf.t. Thus . we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations. Many of the results given below indicate that the averaged and the periodic model share a number of main features. p* = A whereas B* is a mixture of exponential distributions with intensities 3 and 7 and weights 1/2 for each (1/2 = ff w(t)dt = f o (1. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 .3) Note that A enters just as a scaling factor of the time axis. Section 4b).1) and Example 1.1. the conditional distribution . In particular. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. equivalently.3(t) = 3A(1 + sin 27rt). the average compound Poisson model is the same as in III. one may think of the standard compound Poisson model with parameters 3*.1 As an example to be used for numerical illustration throughout this section. of the periodic model as arising from the compound Poisson model by adding some extra variability. the discussion in 111. let .w(t)) dt). (6.w(t).3* = 3A. 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 . St = SeI(t).9). The behaviour of the periodic model needs not to be seen as a violation of this principle. it turns out that they have the same adjustment coefficient. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . We u assume in the rest of this section that p(t) . for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. not random.10. or.2 Define T 6(T) = p(t ) dt. since the added variation is deterministic. In contrast. It is easily seen that .8. p* as an averaged version of the periodic model. Thus. u Remark 6 .(3. In contrast.6.
Daykin et.al.1) . see the Notes to Section 7). a) etw*(a) h(s+t.f. 3 E(8)eaSt = h(s.(3(s + t)dt)e«St adt + /3(s + t)dt . let f 8+1 tc *(a) _ (B* [a] . 1). and define h(s.8). 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.^8 [. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00). with some variants in the proofs.tc* (a)] dv then h (. Notes and references The model has been studied in risk theory by.. . we obtain E.a be the c.east B(8+t) [a] east .1]) . The claim surplus process {St} may be seen as a Markov additive process.3(v)(B(vl [a] .a) = exp { .(1 .5. As usual.(8) [eaSt+dt I7t] = = (1 .. (6.a.5 (see in particular Remark 11. i. we start by deriving formulas giving the m.Q(v) (B(„) [a] .a) Proof Conditioning upon whether a claim occurs in [t. a) is periodic on R.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.g.s .f. J Theorem 6 .1) dv .3(s + t)dt[B(8+t)[a] .. of the averaged compound Poisson model (the last expression is independent of s by periodicity).g.e.g. To this end. [101] .1) a = J8 . r(u) _ inf It > 0 : St > u} is the time to ruin .a . 0 (5)(u.T) = P(8)(r(u) <T). e. with the underlying Markov process {Jt} being deterministic period motion on E = [0. Dassios & Embrechts [98] and Asmussen & Rolski [43]. 6b Lundberg conjugation Motivated by the discussion in Chapter II. Jt = (s + t) mod 1 P(8) . of the claim surplus process. t + dt] or not. The exposition of the present chapter is basically an extract from [44].adt +.4) At a first sight this point of view may appear quite artificial. given that the ith claim occurs at time t is B(8+t).
dt log E(8)east a + f3(s + t) [B(8+t) [a] .6.t} is a multiplicative functional for the Markov process { (Jt.6 . E (8)east (a +. a) Corollary 6. it then suffices to note that E(8)Le.t. a) = h(s.3. a).4).9 as follows.log h(s.c* (e) {Le. so that obviously {Lo. + v)(B([a] .5.1]. we can write Lo Jt. h(s + t. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.4 For each 0 such that the integrals in the definition of h(t .3(s + t)[D(8 +t)[a] . With g the infinitesimal generator of {Xt} = {(Jt.0(s + t)dt[B(8+t)[a] . a) Thus E(8)east = h(s + t. at + f log h(s + t. St)} .(e) Let = h( h(Jo.adt +.2. 0) exist and are finite..t}t>o = h(s. a) h(s + t.5 The formula for h(s) = h(s. 0) P(8)a. a) . 9) is a P ( 8)martingale with mean one.(8)east 179 = = = = = E(8)east (1 .1]) .1]) . Proof In the Markov additive sense of (6.* (a) 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.t = 1 by Theorem 6. According to Remark 11.s.1)dv l og E(8) et where atetk•(a) h(t. a) et.1)dv  o h(t. St)} and . a) = exp I f t3(v)(kv)[a) . B) eoSt t. u Remark 6.9) eastt.
3(s)dt • B(s)[a]h(s) = gha(s. the requirement is cha(i.T. cf. Proposition 6. 0) = h(s) + dt {ah(s) . ry)) at which n* (a) attains its minimum. we put for short h(s) = h(s. St)} with governing probability measures Fes). the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le.0) = Kh(s).180 CHAPTER VI.60(t) = a(t)B(t)[0]. such that for any s and T < oo. That is. Now define 'y as the positive solution of the Lundberg equation for the averaged model.6 The P(s). Bet)(dx) = ^ B(t ) (dx). B(s). Proof (i) Check that m. For each 0 satisfying the conditions of Corollary 6. ( iii) use approximations with piecewiese constant /3(s).g. . However.a . Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . J s [.2.3(s)h(s) + h'(s) +.5 that we can define a new Markov process {(Jt. correspond to a new periodic risk model with parameters ex .f.1. A further important constant is the value yo (located in (0.(3(s)dt) +.3(v)( Bi"i [a] .1) . y solves n* (y) = 0. (iv) finally. 0 < s < 1. (ii) use Markovmodulated approximations (Section 6c).3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. Proposition 6.4. That is. yo is determined by 0 = k* (70) = QB*. see [44] for 11 a formal proof.3(s)B(s) [a]h(s).3. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. Lemma 6 . say.'y). it follows by Theorem II.tc] dv} (normalizing by h(0) = 1). MARKOVIAN ENVIRONMENT ha(s. of St is as for the asserted periodic risk model. P(s) (T(u) < oo) = 1 for all u > 0.y) = eayh(s). [70] .6 ( s ) exp { 0( s )&s) [a] + tc . as above E (s) ha(Jdt. Sdt) = h(s + dt) eadt (1 . When a = y.7 When a > yo.
The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11. and we refer to [44]. have components with densities b(8)(x) satisfying inf sEI. 0(u)) * (b(oo).9) 0(')(u) = h(s. a) TI h(9(u). q) = eryx/h(q)). has a unique stationary distribution.1) the distribution of (l: (oo). which is not used elsewhere in the book. s E I.10) Then for each a. 9(u)) for any bounded continuous function (e. Lemma 6 . B(oo)). we need the following auxiliary result . say s0.7) h(B(u).6. the mean number of claims per unit time is p« 181 = Jo 1.4. ^(u) = ST(u) .8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u. Wu). u which is > 1 by convexity. Corollary 6. J C R+ such that the B(8). considered with governing probability measures { E(8) }E[ . a) a > ry0 (6. xEJ 0 (s)b(8)(x) > 0.g. Here and in the following.9) and noting that weak convergence entails convergence of E f (^(u).2).9(u))} u>0.9 Assume that there exist open intervals I C [0. we get: .6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. e(cc)) Letting u > oo in (6. T(u) < (6. The proof involves machinery from the ergodic theory of Markov chains on a general state space. T) = h(s. RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. the Markov process {(^(u). f (x.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. 1).8) (6. and no matter what is the initial season s. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3.2.1. a) e«uE(8 ) e «^ . (6.
ir) } Plots of h for different values of A are given in Fig. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3.10 Under the condition (6. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c. Among other things. illustrating that the effect of seasonality increases with A. it does not seem within the range of our methods to compute C explicitly.W. 11 7/'O (u) < C+°)h(s) ery". (6. 10 shows that certainly ry is the correct Lundberg exponent.1 In contrast to h.182 CHAPTER VI. A=1/4 A=1 A=4 0 Figure 6. 1.16. Vi(8) (u) .10) of Lemma 3.1. this provides an algorithm for computing C as a limit.Ch(s)ery". elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs . 6. where C(o) = 1 + info < t<i h(t) . Noting that ^(u) > 0 in ( 6. Theorem 6 . MARKOVIAN ENVIRONMENT Theorem 6. where e.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.6 for the Markovmodulated model: Theorem 6 .) C = E1 h(B(oo)) u + oo.11) Note that ( 6. At this stage .1.9). For our basic Example 6 .
13) Elementary convexity arguments show that we always have ryy > Y and ay > ry.(s)(u) < C+h(s)e7". (6. Consider first the timedependent version of Lundberg's inequality. 1 (6.16 In order to apply Theorem 6.yr.0(8) (u+ yu) (6. r. we obtain Co) = 1.7e .w) . the proofs are basically the same as in Section 3 and we refer to [44] for details.17) (6.11 as well as it supplements with a lower bound.42 so that 183 tp(8) (u) < 1.7x j dx _7x } _ 6w + 6(1 . Just as in IV.w)e4u . Theorem 6.42 • exp {J_ cos 27rs .6. where ay is the unique solution of W(ay) =y• (6. #c( ay) < 0 when y > 1/tc'('y). We state the results below. yu) 000 (u) . in our basic example with A = 1.47r sin 27rs + 167r cos 47rs . T) and replace the Lundberg exponent ry by ryy = ay .w ) • 7e u{w • 3e3x + ( 1 . we substitute T = yu in 0(u. (ay).167r I Cu. C_h(s)e7u < V.4. we first note that the function fu° exu {w • 3e .15) The next result improves upon the constant C+) in front of eryu in Theorem 6.3x + (1 . whereas ay < y.12 Let 00)(y) 1 Then info < t<i h(t.(8) (u. ay) • (6. e. .w)e4u dx 9w + 7(1 .12) As for the Markovian environment model.13 to our basic example. 1 ) and all u > 0.(ay) > 0 when y < 1/ic' (7). e7 ( yx)B(t)(dy) > Then for all $ E [0.13 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.14) < C+)(y)h(s) e7yu.. Theorem 6 . RISK THEORY IN A PERIODIC ENVIRONMENT Thus.g. Lundberg's inequality can be con siderably sharpened and extended.
1). Thus C_ = 2 inf ex cos 2irs .18) Notes and references The material is from Asmussen & Rolski [44].4^ sin 2irs + 16^ cos 41rs . such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0.013.20 •exp { 2n cos 27rs .66.(8)(u.19 } 0 <8<1 8 + cos 21rs Thus e.g. 1) for the environment).66. Some of the present proofs are more elementary by avoiding the general point process machinery of [44].1 sin 27rs + 1 cos 47rs . 14 Let C+('yo) be as in (6. C+ = 1.cos 27rs .013.19 I eu. and in fact. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'. Of course.\ 3 C+ = sup 6 exp { A (. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.. but thereby also slightly longer.20). for A = 1 (where 3 e0. 0 <'p(8)(u ) . yo).184 CHAPTER VI. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. where the environment at time t is (s + t) mod 1 E [0.. Finally. we have the following result: Theorem 6 . Then 7oudT . completing a cycle .9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.T) < C+('Yo)h( s. .'Yo)e (6.L sin 27rs + 1 I cos 47rs . MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.I eu.1 sin 2irs + 16_ cos 47rs .0. 1/i18 1 s (u) > 0. .181 s(u) < 1. Thus.\ = 0 . with s the initial season. exp 2^ cos 21rs .16) with 'y replaced by yo and h(t) by h(t. n}.16. 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. and let 8 = er' (Y0). . the nth Markovian environmental process {Jt} moves cyclically on {1.
z/'i (u.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.7. To this end. DUAL QUEUEING MODELS 185 within one unit of time on the average . AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = 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. M(n) = Supt>o Stn). (6. since the settings are equivalent from a mathematical point of view. This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. Notes and references See Rolski [306]. 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. Let 0j.jEE of the risk process.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. Bi. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. one simple choice is Oni = 0( i .20) be the claim surplus process of t>o the nth approximating Markovmodulated model. Thus. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). We let {Stn)} (6.1 ((i 1)/n) ) and Bni = B . 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). . but others are also possible. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6.
JT = j) = LjPj (VT > u.. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi.2). JT = j} and {VT > u.1) 7ri In particular.4) where 0* = >jEE 7rj/3j. I* )3i P(V > u. J* = i) for all j. Jt ). J* = i) = P. In particular. JT = i) = P(V > u.2) Oi(u) = 1.3). and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. Jo = j.3) 7ri where (V. For (7. • The queueing discipline is FIFO. J* = i). and for (7.186 CHAPTER VI. and (7. . (7. JJ = i). JT = i) = 'P. Proof Consider stationary versions of {Jt}o<t<T.1).0i (u . let T .. (7. JT = Z).. ii (u) = it /3 P(W > u.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u.P(V > u. JT = j) = 7rjPj(VT > u. .1) over j.T(V > u I J* = i). just sum (7. The actual waiting time process 1W1. T) = 7ri 1 P.=1 . J*) is the steadystate limit of (Vt. Proposition 7. I* = i).n(VT > u.1 Assume V0 = 0. Taking probabilities and using the stationarity yields 7riPi(T(u) < T. Jo = i. Proposition 7. 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. 2 . Now let In denote the environment when customer n arrives and I* the steadystate limit.2) and use that limF (VT > u. JT = i} coincide. Then Pi(T(u) < T.3. (VT > u I JT = 2). The first conclusion of that result then states that the events {T(u) < T. (7. {Jt }o<t<T• Then we may assume that Jt = JTt.1) follows.oo in u (7. (7.
and one has PI'>(rr(u) < T) = P('_T)(VT > u).4). I* = i. P(W >u.l. if T is large. and (7. In the setting of the periodic model of Section 6.g. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W.o.3) improving somewhat upon (2. see Regterschot & van Doorn [123]. T].5) follows from (7. .=i) a4. on average /32TP(V > u. P(1')(r(u) < oo) = P(')(00) > u).3). see in particular Harrison & Lemoine [186].8) For treatments of periodic M/G/1 queue. I*) with the timeaverage .4) and (7. and Rolski [306]. Taking the ratio yields (7. J* = i) see W > u. n=1 N However. a paper relying heavily on classical complex plane methods. we have 1: I(W.1 is from Asmussen [16].7) (7. Proposition 7. a general formalism allowing this type of conclusion is 'conditional PASTA'. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84].T)(T(u) <T) = P(8)(VT > u).6) (7. (7. and further references (to which we add Prabhu & Zhu [296]) can be found there. P(.7) of that paper. N * oo. With {Vt} denoting the workload process of the periodic queue. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. and of these.I *=i). B(t) have been periodically extended to negative t). with (7. >u. [243].7. 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. on average 0*T customers arrive in [0.. that /3(t). The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. A more probabilistic treatment was given by Asmussen [17]. The relation (7.4) can be found in Regterschot & de Smit [301]. Lemoine [242].
This page is intentionally left blank .
with common distribution B and independent of {Nt}. the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r..T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite.1) (other terms are accumulated claims or total claims). and the evolution of the reserve may be described by the equation Rt = u . 189 . Zt As earlier.At + p(R8) ds. i&(u. Thus in between jumps. the aggregate claims in [0.T) = F(T(u) < T).2) tk(u. z/i(u) = F IinffRt< 0IRo=u 1 (1. However . are i.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 .6. resp . and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). . finite horizon. t] are Nt At = Ui (1. Thus.i.d. {Rt} moves according to the differential equation R = p(R). U2. and that the claim sizes U1..
However. but when the reserve comes above v.i(u) = 1 for all u. If Ro = v < u. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. In this situation.e. No tractable necessary and sufficient condition is known in complete generality of the model. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. or o(u) < 1 for all u.p2. oo) is given by i (u + p/S). that {Rt} will reach level u before the first claim arrives. P(r) _ p + e(r . dividends are paid out at rate pi . pi > p2 and p(r) = One reason could be competition. rather than when the reserve itself becomes negative. 1 .1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v.'(u)) > 0 so that V'(v) < 1. Example 1. say e.2. it seems reasonable to assume monotonicity (p(r) is u .190 CHAPTER VII. Example 1. Now return to the general model. Assume 0(u) < 1 for some u. Thus at deficit x > 0 (meaning Rt = x). the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p.p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above.4 Either i. we get p(r) = p + er. That is. where one would try to attract new customers as soon as the business has become reasonably safe. we can put Rt = Rt + p/S. there is positive probability. Proof Obviously '(u) < ilb(v) when u > v. but assume now that the company borrows the deficit in the bank when the reserve goes negative.Vi(v) u > e(1 . say at interest rate b.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . Proposition 1. Hence in terms of survival probabilities. when x > p/S.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. i. and the probability of absolute ruin with initial reserve u E [p/S.
1. In particular. and P(Rt + oo) > 0. one can couple the risk process and the storage process on [0. if and only if V)(u) < 1 for all u.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. We next recall the following results. B and (constant) premium rate p. appealing to Proposition 111. cf. Proposition I1I.2(d).3. Hence ik(u) < 1 for all u by Proposition III. However.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo.6 For any T < oo. This is basically covered by the following result (but note that the case p(r) .b(u. we have z/i(u) <p(u .2(d)).1.5 (a) If p(r) < /. {Vt} decreases at rate p(v) when Vt = v (i. then ?(u) = 1 for all u.+ p(r) exists. Then if u > no. [APQ] pp.uo) and. say V.f p(Vs) ds. and hence by a geometric trials argument. In between jumps. Then 0(u) = P(V > u).uo) < 1. instead of (1.2 once more. In case (a). let uo be chosen such that p(r) > p = 0ILB + e for r > uo.e. which was proved in 11. 296297): Theorem 1. That is. (1. let uo be chosen such that p(r) < p = /3µB for r > uo. In case (b).1. V = p(V)). then l/i(u) < 1 for all u.5) and the process {Vt} has a proper limit in distribution .1. { Vt} remains at 0 until the next arrival).3µB for all sufficiently large r. (1.2) we have t Vt = At .6) . Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. . Starting from Ro = uo.4.2) for r sufficiently large so that p(oo) = limr.T) = P(VT > u). INTRODUCTION 191 decreasing in Example 1.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1. Theorem 1. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0.. that u zPp(u . Let Op(u) refer to the compound Poisson model with the same 0.1 and increasing in Example 1. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. obviously infu<uo z/'(u) > 0. (1. Proof This follows by a simple comparison with the compound Poisson model. hence Rt < uo also for a whole sequence of is converging to oo.
7 p(x)g(x) = tofB (x) + a f (x . RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. Now obviously. say.Sx} dx. (1.Qw(x) . for the storage process {Vt}. oo). In view of the path structure of {V t }. we thus need to look more into the stationary distribution G. of (1. and the other being given by a density g(x) on (0. oo) must be the same as the flow the other way. the flow of mass from [0. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. one having an atom at 0 of size 'yo. yo ^ 1 + oo Q exp {. Corollary 1. we arrive at the desired interpretation of the r. It follows in particular that 0(u) = fg(Y)dy. Oeax f x e'Yg (y) dy } = p) eaxa(x) . t + dt] if and only if Vt E [x.9) Proof We may rewrite (1.8) Proof In stationarity.192 CHAPTER VII. say when {Vt} is in state y. An attempt of an upcrossing occurs as result of an arrival.s. Then the ruin probability is tp (u) = f' g(y)dy.6w(x) . (1. say.8) as the rate of upcrossings. say if p(r) goes to 0 at rate 1 /r or faster as r j 0.h. B(x) = e.6 applicable. 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.8 Assume that B is exponential with rate b.h.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. x] to (x. the l. of (1.y)g(y) dy. and is succesful if the jump size is larger than x . x + p(x)dt]). Jo AX) (1.6x and that w(x) < oo for all x > 0.s.Sx}.8) as g(x) = p 1 {yo13e_6x +. Note that it may happen that w (x) = oo for all x > 0.7) Proposition 1. It is intuitively obvious and not too hard to prove that G is a mixture of two components. Considering the cases y = 0 and 0 < y < x separately.y. where g(x) = p( ^ exp {. u Define ^x 1 w(x) Jo p(t) dt.
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
RESERVEDEPENDENT PREMIUMS u from which (2.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2.pa.g. r (j3/E) In particular.a) .3.x) dx e.V.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E. The process {St} corresponds to {Wt} so that c(a) or2a2/2 .2y +µ ) dy .3a/ (5 .2) follows by elementary algebra. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E. and the c.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments. /^ u Example 2 .2) follows by elementary algebra.3/E) By the memoryless property of the exponential distribution. then {Rt} is the diffusion with drift function p+Ex and constant variance a2. where V is Gamma(b.pa.f. assume that {Wt} is Brownian motion with drift µ and variance v2. 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 .e. From ic(a) = . i.5 The analysis leading to Theorem 2. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0. As an example. Proof 2 We use Theorem 2.V < x)]0 + f P(V > p/E ) + eby fv (p/E ./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E .200 CHAPTER VII. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .13 /E) r (. 13/E). of Z is IogH[a] = f ytc(y)dy = e fa (0. .
Further studies of the model with interest can be found in Boogaert & Crijns [71].g. [129] and Harrison [185]. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1.4 is classical. and since RT = 0 by the continuity of Brownian motion. [129]..d. Delbaen & Haezendonck [104]. Q2/2E).v.3) was derived by Emanuel et at. A r. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I.3 is from Harrison [185].1) . Corollary 2. 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). write Vi* (u) for the ruin probability etc. It must be noted.p*. Emanuel et at. Gerber [157] p. Paulsen [281]. 3 The local adjustment coefficient. Paulsen & Gjessing [286] and Sundt & Teugels [356]. for a martingale proof. that the analysis does not seem to carry over to general phasetype distributions.2 is a special case of a perpetuity. or to nonlinear premium rules p(•). [357].Y*p* W*(u) < ery*u = 0. see e.3. Z is normal (p/E.e.i. however. The formula (2. Gerber [155]. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. and recall Lundberg 's inequality . [283]. it is also used as basis for a diffusion approximation by these authors. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case).8. of the form Ei° p"X" with the X„ i. Goldie & Griibel [167]. se e. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . [282]..g. write y* for the solution of the Lundberg equation f3(B[ry *] . Some of these references also go into a stochastic interest rate. as in the proof of Proposition 2. not even Erlang(3) or H3.
as will hold under the steepness assumption of Theorem 3. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V.1.1 ). (3.i)eex.e.e. let p* be a in (3. we will use the local adjustment coefficient 'y(x). which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.1. oo for all E > 0. i.1) . 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).2) such that p(x) < c(. e(1o+e)2 (x ) u > 00. 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 . and that p(x) * oo. as solution of the equation n(x. For the last asssertion .E). log ?i(u) < < 00 JO .w (u) J dt > c(3)eeu v 1 p(u+ t) . Proof of Theorem 3. 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. 1) and for a given E > 0. and (for simplicity) that inf p(x) > (3µs . a first step is the following: Theorem 3 .1 Assume that for some 0 < 5o < oo. choose uo such that p( x) > p* when x > u0E. Let y* < So. a) = f3(B[a] . If 60 s f 6o.3) When trying to extend these results to the model of this chapter where p(x) depends on x.5) which implies inf. x * oo. then log u (u) In the proof as well as in the remaining part of the section . B(x) > C(2)e(ao+f)x for all x.ap(x). (x. When u > uo.'y ( x)) = 0 where r. c(. Then lim sup u>oo u and e E''p(r) + 0.*(u) . Letting first E * 0 and next ry * T 5o yields the first statement of the theorem.4) we assume existence of y(x) for all x.C*ef*". The steepness assumption and p(x) + oo ensure 'y(x) * So..log '(u)/u < ry*(1 . (3. i.. 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) . choose c(. x>0 (3.202 CHAPTER VII. it holds that f3[s] T oo.>o 7(x) > 0.
Then limelog l/ie (u) = I(u). Bucklew [81]).13 is a technical condition on the claim size distribution B. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. Theorem 3. The rest of this section deals with tail estimates involving the local adjustment coefficient. and in particular. which essentially says that an overshoot r. one can then assume that e = 1 is small enough for Theorem 3. (3. {Rte)} defined as in (1. Condition 3. For e > 0. (u) = O(u/e). the result is not very informative if bo = oo. then Rte) = CRtie for all t so that V). or (b) Condition 3.7) CIO Remarks: 1. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V.3 Assume that either (a) p(r) is a non decreasing function of r.3. UJU > x cannot have a much heavier tail than the claim U itself. ruin will occur if the claim is at least u + v.3 to be reasonably precise and use e` (u) as approximation to 0 (u).(u) > so. Theorem 3 . THE LOCAL ADJUSTMENT COEFFICIENT 203 where c.0 are the same.ea°/(ecf1)). 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. u Obviously. The form of the result is superficially similar to the CramerLundberg approximation. If p(x) = pis constant . I.v.4)eE" Given such an arrival. the asymptotics u * oo and c .. 3) = (1 .' (u) < eI("). noting that in many cases the constant C is close to 1. Then .e.2 is also an approximation under appropriate conditions. However. .g.13 below holds. 3. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. 2.6) The second main result to be derived states that the bound in Theorem 3.1 only presents a first step. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. (3. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(. by cU2.2).
3.6).bx} dx fo 00 1 + [exp {/(3w(x) . say. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.3. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3.204 CHAPTER VII.(x) dx.exp {/33w(u) .8 in terms of I(u) when the claims are exponential: Example 3 .bx} dx = 1 + J0 dodx(x) exp {. 3a Examples Before giving the proofs of Theorems 3.4 Consider again the exponential case B(x) = eax as in Corollary 1. First.7) is only captures 'the main term in the exponent' but is not precise to describe the asymptotic form of O(u) in terms of ratio limit theorems (the precise asymptotics could be logI(u)e1(U) or I(u)"e_I(u).bx} dx 1+0. 3. u . Then y(x) = b . 5. However.bu}. we consider some simple examples.bx}]o + b /' oo exp low (x) .8.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) . it is formally needed only for Theorem 3.1 + b f e. the logaritmic form of (3.bx} dx .(iw(x) . As typical in large deviations theory. RESERVEDEPENDENT PREMIUMS 4. we get = 1+ J" AX) exp {(3w (x) .(3/p(x). rather than eI(u)).2. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) . we show how to rewrite the explicit solution for ti(u) in Corollary 1.
3.2.7) follows.3.I ( u) fool. note first that the appropriate slow Markov walk assumption amounts to u. (3. Be = e log U000 e. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.fo 7(x)dx/Edy f .9) yields e log . (3. For Theorem 3.5 Assume that {Rt} is a diffusion on [0. Choosing yo.3 in the particularly simple case of diffusions: Example 3. BE * 0. in the definition of AE converges to 0. 70 > 0 such that 7(x) < 7o for y < yo.fo 7(x) dx /E dy > av 'yo /Edy = E (1 .5) is infx>o 7(x) > 0 which implies that f °O .9 ) 11000 eI(v)dy f000 e. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3.3.f y(x)dxd y If 7(x) is increasing . Similarly.2.fory(x+u)dxdy ( 3.10 or Karlin & Taylor [222] pp.1/8 .I ( v )dy fo +u) dxdy . (X) = µ(x). and (3..BE. oo) with drift µ(x) and variance a2 (x) > 0 at x. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy .e. 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. IE(u) = I(u)/e. ry(x /b I u o e f0 °° e  e.fa 7(x+u)dx/Edy o The analogue of (3. (u) = I(u) + AE . 0.10) where AE = e log 000 e. > lime log e = 0 and AE * 0. It is well known that (see Theorem XI. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.0. In particular..8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3.ev 0 O /E) J0 70 70 Yo This implies lim inf A.1. we get r 00 e. and (3.2(X) = ev2(x) so that 7e(x) = 7(x)/e. u .
lim sup Af < lim sup c log(1 . I(u ) = G1(u) + .6 Assume that B is exponential with rate S. RESERVEDEPENDENT PREMIUMS The analogue of Example 3. G...0..Q/p*. G. the slow Markov walk assumption means 5E = b/c. I. G. 0 Now (3. .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. the results are suggestive in their form and much more explicit than anything else in the literature. . Thus 7e(x) _7(x)/e and (3.4.+1 (u) = o( 1). the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. that the interchange of the slow Markov walk oo is not justified and in fact.6/p* so that u 1 I (U) = bu .5 for risk processes with exponential claims is as follows: Example 3 . ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.5. however .5.. Further. we have 5 > 7o and get lim inf AE > lime log e .5) and 7* = 5 . Then the solution of the Lundberg equation is y* = b .6) exactly as in Example 3. so our approach is to determine standard functions Gl (u).0/e.1 3./3 1 AX dx. 0. + Gq(u) + o(G9(u))• Gi (u) It should be noted . this leads to (3.206 CHAPTER VII. . As in Example 3. .10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. Nevertheless . . E+o e*O By (3.e.0) = 0. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. _ . Of course.7o C 15 I I.5.(u) oo. > . 7(x) is typically not explicit. Ignoring 1/5 in the formula there.
fu I(u) Su . For example. I(u) Pt.4) or gamma distributions.7 Assume that B(x) .11) that b[s] * co as s f S and hence 7* T S as p* + oo.c2 Su a dx ) Su a<1 Su . THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . Here B[s] is defined for all s and B[s] . Hence (3. This covers mixtures or convolutions of exponentials or. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .C2p* C2 = (3clr( a))11'. and hence (3.:. e. more generally. 2. More precisely.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b. (3.8).1/a). in the phasetype case . It follows from (3.1/k).3cse7*I7(77) . x T 1. phasetype distributions (Example 1.Y ..y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.12) with y > 1. if the phase generator is irreducible ( Proposition VIII.3. . y = 2 if B is uniform on (0.clxale5x 207 (3.s)C' f "o o as s T S.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . 1. c4 = c2b 1/'/(1 . u Example 3 . 77 = 1 if B is degenerate at 1.cs(1 .x)n1. ry* loge*+ g7loglogp*. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.ry*°p*. u(logu + r7loglogu).8 Assume next that B has bounded support.11) with a > 0. say 1 is the upper limit and B(x) .g.1) leads to (S7T N Ocp a. the typical case is a = 1 which holds .1) leads to . .
RESERVEDEPENDENT PREMIUMS Example 3 . one could also have considered the increment ru (T1) .r^.g.. By convexity of the m . This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .f. .sp(u).(t))dt.11) and (3.12). 1 = E.c8 log .•.13) We get b[s] .e7o ( u)(ul+u r. this is only possible if 7o(v) 2 7o(u)• .u . h 10.B[7o (u)] . ec78)2/2c7 dx C7 . h]. 3b Proof of Theorem 3.1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .css 2%rc7eC782/2. g. I (u) c8u log u 0 where c8 = 2/c7. (3.(T1)) > Ee7o(u)(ul+vr»(Ti)).4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3. x f oo .10 Assume that p(x) is a nondecreasing function of x. (3. The assumption implies that ru(t) .f. (3. 1 0 3e. (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3.Ote7o( u)(u.r„(Ti).3 (B[s] .log p*.9 As a case intermediate between (3.4).u is a nondecreasing function of u. of U1 + v . of the increment in a small time interval [0.15) Proposition 3.ru(TI)) . assume that B(x) CO x2/2c7.Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u). Hence for u<V. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.208 CHAPTER VII.4) is the formula h logEues ( Rhu) .1) .14) for the m . 7 * .2 We first remark that the definition (3.
4) considered as function of 7 is convex and 0 for y = 0. Then (u) < efo Yo(x)dx.7o (u)p(u)• Since (3. Also. Separating after whether ruin occurs at the first claim or not. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3. the case of 7 then follows immediately by Proposition 3. Assume (3.17) shown for n and let Fu(x) = P(U1 + u .17) from which the theorem follows by letting n + oo.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. THE LOCAL ADJUSTMENT COEFFICIENT For (b).11 Assume that p(x) is a nondecreasing function of x.u[70(u)] fo eyo(x)dx . it is easily seen that fu x7o(y)dy < xyo (u).3.(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn). Hence 1 = EeYo(u)(U1+uru(T1)) < E. fa 7o(y)dy < u7o(u) < xyo (u) for x > u. Hence „/.10(b): Theorem 3. u We prove Theorem 3.1) .e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] . this is only possible if yo(u) > 7(u). (3.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j. we obtain „I.(n+l) (u) 1 .u > tp(u).Fu(u ) + J ^(n)(u .ru(T1) < x). note that the assumption implies that ru(t) .2 in terms of 7o. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.
The probability of this is at least n n. define uk. the value of {R(E)} at the time of downcrossing is < unl.n <Z auk}l.3). {RtE)} (starting from u = un. Let Ck. Y*u /E.E ( u/n) ^•e. yo(u) appears more difficult to evaluate than y(u).3 is required. W O . Lemma 3.n AX).x/n.12 lim sup4^o f log O. 0. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3. 3c Proof of Theorem 3. and. let Op*.. y* evaluated for p* = Pk.n.2. Further...n) pn niE (u /n) n n_1 n.11 is sharper than the one given by Theorem 3. P k.10(b ) that the bound provided by Theorem 3. the probability that ruin occurs in the CramerLundberg model with p* = pn.: y(u). C*e where the first equality follows by an easy scaling argument and the approximation by (3. for either of Theorems 3. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3. 0 It follows from Proposition 3.nbe C*. in accordance with the notation i/iE (u). 0. To this end.n) must first downcross unl..3/e and U. Proof For ruin to occur.E (u/ n) Y'E (un .E (u/n) Now as e .15). given downcrossing occurs.n u k}1.n = sup p(x). (3.n.. resp. Also. we have chosen to work with y(u) as the fundamental local adjustment coefficient.n.n.10(a) for some of the inequalities. 3. However.3).2).n inf n uk1. For these reasons. (u) < I(u)..2. pk n = uk_l. we used also Proposition 3.E (u) denote the ruin probability for the classical model with 0 replaced by .210 CHAPTER VII.e (u) = v'.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3.n = ku. (un2.n (starting from u/n) without that 2u/n is upcrossed before ruin.E (u/n).n) > k =1 II v ^k n. (u). in . ryk. by €U=.n so that n. and here it is easily seen that yo(u) .3 The idea of the proof is to bound { R( f) } above and below in a small interval [x . op*.I.
e. also ryk.i.n <X<Uk. ne7k.n cE (2u/n) Ck ne7k.nu/en(1 + where o(1) refers to the limit e . since ry' is an increasing function of p'. 211 Clearly. 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. THE LOCAL ADJUSTMENT COEFFICIENT particular.. B(x) (3..n + 0(1). i. 11 Theorem 3. *p•.nk=1 limsupelogv). 0 with n and u fixed. 3 now follows easily in case (a). we need the following condition: Condition 3.3.! (u/n) n n m 7k.7k.nu /fn( 1 Ck  e.nu /En) o(1)).n . (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof. Indeed . . for all x . /' (u/n) 'T nk. In case (b). (3.E (urn) < \ *I..12 completes the proof.a( u)z.13 There exists a r. v. 40 Combining with the upper bound of Lemma 3. It follows that n log V'C (u) k =1 log Ypk. so that Theorem 3.n.F (2u/n). k=1 k=1 n u _ nE7 k. in obvious notation one has tC (x) = y(x)/e.19) .2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u)..18) (ii) the family of claim overshoot distributions is stochastically dominated by V.E (u/n) Op•.n = sup ?'(x).log Ck. uk_1. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y).E (u/n) OP +^p•.
^'' = E [ .v.eV) • P (T(E) (u.E(E) (u.V) = e71 nu/Eno(l) (using (3.EV) = EiI 1 . v ) = inf { t > 0 : R(c ) < v R) = u } . u/n) < oo] E [OE (u/n . we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. (u/n .E (0) cf..n < ery1.. u/n)) . RESERVEDEPENDENT PREMIUMS To complete the proof.212 CHAPTER VII. Write EO. u/n)) I T(E) (u.E (u/n .nu/En0(1) . u/n) < oo] . ) (u u /n)) . (R. infx>2u /n P(x) . . (3.18) for the last equality). Then the standard Lundberg inequality yields El < E?.R<) (u v).of:>2 in n(x).5) and the standard formula for b(0).. u /en 0(i) _n so that E2 < e2ryl nu/En0(1). Ei + E2 < e71.QEU 1 .n V.( .2y 1 ' . N with EN < 1 = infx>2u/nA(x) = 0(1)..^(E) (u. Then Y'E (u) ^(E) (u. (u/n . T(E) (u. u /n) < oo] l = = < E [OE (u/n .nu /EnE [e71. 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. V < u/En] + P(V > u/En) (u/En . let v < u and define T(E) (u.E (2u/n .1 n.EV) = El + E2. T() (u. P (T(E) (u.EV) = e. The probability of ruin in between two downcrossings is bounded by Epp . v ) = v . u/n) < oo) . For E2.
J y(Rs)dR.7) then comes out (at least heuristically) by analytical manipulations with the action integral. [89]. the risk process itself is close to the solution of the differential equation r(x) _ r (x.T) = P „(info<t <T Rt < 0) via related large deviations techniques.13. s) as in (3. Typically. u Notes and references With the exception of Theorem 3. .t.=1 J An approximation similar to (3. 0 ) (= p(x) . THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.3. Whereas the result of [122] is given in terms of an action integral which does not look very explicit.20) (with ic(x. s).21) to pass from u to 0. l o JJJ o .21) (the initial condition is r(0) = u in both cases).7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al.) = exp . the results are from Asmussen & Nielsen [39].3EU) (3. Djehiche [122] gives an approximation for tp(u. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u. Similarly.r. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . 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.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof.1.J r(Rs)p(R. it might be possible to show that the limits e .4) and the prime meaning differentiation w. one can in fact arrive at the optimal path by showing that the approximation for 0(u. the approximation (3. Bucklew [81]). T) is maximized over T by taking T as the time for (3.g.)Ui } .7(x)) (3. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . whereas the most probable path leading to ruin is the solution of r(x) _ k (x. Comparing these references with the present work shows that in the slow Markov walk setup.)ds + Y(R2. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. 0 and b T 00 are interchangeable in the setting of [89].
For different types of applications of large deviations to ruin probabilities . RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding .g. see XI. We should like. the exponential distribution ). 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.214 CHAPTER VII.. however. . e.3.
on Eo = E U {A} where A is some extra state which is absorbing. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . 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<. if a problem can be solved explicitly when the relevant distributions are exponentials. refers to the case Jo = i. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. Note that since (1. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . if v = (vi)iEE is a probability distribution. that is.1) is 'Here as usual . 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. 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.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. We often write p for the number of elements of E. Typically. P. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. and not in other cases. A distribution B on (0.
Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. i. Then a = a1 = 1.3.e. T is a subintensity matrix2. tij > 0 for i 54 j and EjEE tij < 0 . i.216 CHAPTER VIII. We now say that B is of phasetype with representation (E.1 Suppose that p = 1 and write . Equivalently.1 The phase diagram of a phasetype distribution with 3 phases. 0 2this means that tii < 0. The initial vector a is written as a row vector.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. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. In particular. t1 = /3.0 = t11. Here are some important special cases: Example 1 .T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. k}. (1. that is. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. T) (or sometimes just (a. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. B(t) = Fa(^ < t ). the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1.2) The interpretation of the column vector t is as the exit rate vector. and we have t = Te. C is the lifetime sup It > 0 : Jt E E} of {Jt}. j. an exponential distribution with rate parameter . E = {i. a.
1)!e Since this corresponds to a convolution of p exponential densities with the same rate S.. .... 0 SP 0 and the phase diagram is (p = 2) . 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 .. so that the density is P E ai6ie6. . 0 •..2 corresponding to E = {1.x i=1 Thus E _ Si 0 T 0 S2 0 0 .. 00)) S s o . 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •.. 6.2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p. 0 0 0 0 S 6 .1. . the EP distribution may be represented by the phase diagram (p = 3) Figure 1.. p}.3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51. .. 0 S 6 Example 1. . .. a = (1 0 0 . PHASETYPE DISTRIBUTIONS 217 Example 1....
3 . see A.g.f is B (x) = 1 . . the backwards equation for {Jt} (e. 5 Let B be phasetype with representation (E. i. ds^ = ds' = ttlaj + tikpkj.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. (b) the density is b(x ) = B'(x) = aeTxt. The basic analytical properties of phase type distributions are given by the following result . E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . the restriction of P8 to E.f. [APQ ] p. Theorem 1 . B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. T).aeTxe.3 0 Example 1 . Then for i . 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. MATRIXANALYTIC METHODS Figure 1. dp.1 tP1 1 Figure 1.4 For example.e.d. a. p:.218 CHAPTER VIII. Then: (a) the c. 36) yields s d. (c) the m.t2 yt bP. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2.1)"n! aT"e. the Erlang distribution is a special case of a Coxian distribution. j E E. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3.
h = (T + sI)1t. or w.g.f.5) ki = 1 + tii L jj:Ai tii (1.e.tii we go to A. of the initial sojourn in state i.g.tii is the rate of the exponential holding time of state i and hence (tii)/(tii .T) 't = (. i.. the solution is P8 = eT8.n lt .g.. Then h tit ti + ti3 h j .p.B(x) = 1'a (( > x) = P. .T) 1t. (1) n+1n!aT . tij / . 1.s I . = aPxe.s j# tii i (1. Rewriting ( 1. in which case the time to absorption is 0 with m . j#i jEE tijhj + his = ti.tii and have an additional time to absorption either go to state j which has m . d8 P8 = TP8. d" dsn a (. this means in vector notation that (T + sI)h = t.5) as hi(tii + s) = ti  t ij hj.f. we arrive once more at the stated expression for B[s].tii tii .1.s) is the m .12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . hj . and since obviously P° = I.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). (Jx E E) = this proves (a).1 ) n +l n ! a (s I + T ) . ti/ . Alternatively. i.f. For (c). for n = 1 we may put ki = Ei( and get as in (1. B(n)[0] = _ Alternatively.n1t = (1)nn!aTn1Te (1)nn! aTne. and since b[s] = ah.5) Indeed . define hi = Eie8S. PHASETYPE DISTRIBUTIONS 219 That is. we i w. and (b) then follows from 1: aipF. Part (d) follows by differentiating the m.p. the rule (A.g. After that.f.6) . Since 1 .
0 Example 1. making the problem trivial. another the case p = 2 where explicit diagonalization formulas are always available.h. see the Appendix.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly.s. This implies that we can compute the nth moment as (1)"n! aT "e 1"n! 1 1 22 9 9 10 70 7 1 10 10 1 9 +6. 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. Consider for example 3 9 a= (2 2). T= 2 111 so that 2 2 Then (cf. there are some examples where it is appealing to write T on diagonal form.220 CHAPTER VIII. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. are idempotent. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x . Example A3.
and in fact one also most often there allows a to have a component ao at A. T) is then defined to be oo on a set of probability 1. i.11aDD. then the matrix m. 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). 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. a random variable U having a defective phasetype distribution with representation (a. .7 If B is phasetype with representation (v. i.T). • The phasetype distribution B is zeromodified. This is the traditional choice in the literature.4b for definitions and basic rules): Proposition 1. (1.hall.f.e a mixture of a phasetype distribution with representation (a/llall. There are two ways to interpret this: • The phasetype distribution B is defective.1. < 1.29) and Proposition A4.7) Proof According to (A. hail = E=EE a. where the initial vector a is substochastic. 5 and serves at this stage to introduce Kronecker notation and calculus (see A. 0 Sometimes it is relevant also to consider phasetype distributions.e 11BIJ = 1laDD < 1.4.T) with weight hall and an atom at zero with weight 1 . B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).g. or one just lets U be undefined on this additional set.
8 Let B be phasetype with representation (a. v. but in many practical cases. Rolski. x * oo.hve7x. In Proposition A5.Ce7'. but todays interest in the topic was largely initiated by M. 77 > 0 and k = 0.1 of the Appendix. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. see his book [269] (a historical important intermediate step is Jensen [214]). cf. In older literature. the conditions of Proposition 1.F. cf. 1.f.q be the eigenvalue of largest real part of T. the result follows (with C = (ah)(ve)). it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. where C. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve .8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. See in particular the notes to Section 6. No satisfying . and we have eTx . we give a criterion for asymptotical exponentiality of a phasetype distribution B. 2. not only in the tail but in the whole distribution. i is real and positive. Lipsky [247].4c). 0 Of course. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. Notes and references The idea behind using phasetype distributions goes back to Erlang. assume that T is irreducible . let . Schmidt & Teugels [307] and Wolff [384]. Neuts.222 CHAPTER VIII.g.. All material of the present section is standard. . Using B(x) = aeTxe . but the relevant T is not irreducible. B(x) . Here is a sufficient condition: Proposition 1. Then the tail B(x) is asymptotically exponential. one has k = 0.. The Erlang distribution gives an example where k > 0 (in fact.8). distributions with a rational m. let v.g. 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. the text is essentially identical to Section 2 of Asmussen [26]. T). (1. Example A5. h can be chosen with strictly positive component. here k = p1). MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. Schmidli.8) Proof By PerronFrobenius theory (A. the Erlang case). (or Laplace transform) are often used where one would now work instead with phasetype distributions.f.
.: U1 + . Jt={Jt?ul}. JtJt1) Then { 0<t<U1 . if U is absolutely continuous on (0. or the density is available ) is..1 of the Appendix. A related important unsolved problem deals with minimal representations: given a phasetype distribution . 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 +. known.. be i. n=O We may think of the U. U1<t < U1+U2.d.. but is in part repeated below. . U2. 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.1 Consider a renewal process with interarrivals which are phasetype with representation (cr.g.+UnEA).f.. If B is exponential with rate 0. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. (2... +UnEA} 00 = EEI(U1 +.1.i. + U0 is 0 .1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together... with common distribution B and define4 U(A) = E# {n = 0.. oo) w. is Markov and has two types of jumps . we refer to U as the renewal measure. 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... the problem has an algorithmically tractable solution if B is phasetype: Theorem 2.r. Lebesgue measure. Let U1. as the lifetimes of items (say electrical bulbs) which are replaced upon failure. . however. but nevertheless.2.T). For this reason. 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 renewals form a Poisson process and we have u(x) = 0. and U(A) is then the expected number of replacements (renewals) in A.t.
and let µB = aTle be the mean of B.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. and the distribution of Jx is ae ( T+t«)x. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . Corollary 2.3 Consider a renewal process with interarrivals which are phasetype with representation (a. which is phase type with representation (v.IIBII which is > 0 in the defective case.U1 U3 U2 U3 U4 Figure 2.1 Corollary 2. which is ti in state i. This is defined as U1 + ..1. T). fi(t) U2 U1 .224 CHAPTER VIII.. B is defective . the lifetime of the renewal process. Equivalently. as the time of the last renewal.T). 2. i. define the excess life e(t) at time t as the time until the next renewal following t. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj . the density is veTxt = B(x)/µB. u Returning to nonterminating renewal processes .1) follows by the law of total probability. The renewal density at x is now just the rate of jumps of the second type.T + ta). i. IIafl < 1. that is. u The argument goes through without change if the renewal process is terminating.e. this is welldefined. Then the lifetime is zeromodified phase type with representation (a. since Uk = oo with probability 1 . and hence ( 2. see Fig. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. (b) £(t) has a limiting distribution as t * oo. However. is the first k with Uk = 00. Hence the intensity matrix is T + ta.T) where v = aT1 /µB. + Uit_1 where s. Proof Just note that { it } is a governing phase process for the lifetime. Hence ( 2. . and the jumps of the first type are governed by T.T) where vt = ae (T+ta)t .e. .1) remains valid for that case.
2) v(T + ta) = 0.T) where vt is the distribution of it which is obviously given by the expression in (a). T1 and eTx commute. (ii) First check the asserted identity for the density: since T. u Example 2 . RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2. the unique positive solution of ve = 1. i. The renewal density is then aeQtt = (al a2) ( 7i 7"2.) ( t2 ) . (2. The time of the next renewal after t is the time of the next jump of the second type. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. cf. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. Al. According to Example A3.2. hence e(t) is phasetype with representation (vt.q2.e. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB.6. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . 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).e.4 Consider a nonterminating renewal process with two phases. we first compute the stationary distribution of Q.1.
52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].t2) . Then Q= 0 55 )+(1o)=( j ad ).a27r1) (t1 . t1B 0 Example 2 . The present treatment is somewhat more probabilistic.`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 .6 Let B be hyperexponential.a27rl) (tl . A = 25. and Example 2.4 yields the renewal density as u(t) = 2 (1 .52a1. MATRIXANALYTIC METHODS e.e2bt) 13 Example 2 .5 Let B be Erlang(2).tl) 7r2t2 + eat (a17r2 .(biaz + aza.4 yields the renewal density as u(t) = 5152 e. Hence 7r = (1/2 1/2). and Example 2. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .t2) 1 + eat (a17r2 . .226 CHAPTER VIII. )t (51 .
1 on the next page.p. Then: (a) G+ is defective phasetype with representation (a+.f3aT1. with 0 denoting the Poisson intensity. and if there is a subsequent ladder step starting in j whic occurs w. Corollary 2. use the phasetype representation of Bo. The essence is contained in Fig. Here we have taken the terminating Markov process underlying B with two states.3. r(u) the time of ruin with initial reserve u. We asssume that B is phasetype with representation (a. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. Within ladder steps. Next. Corollary 3. we shall.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. The stars represent the ladder points ST+(k). . For (b). Then each claim (jump) corresponds to one (finite) sample path of the Markov process.) = F(ST(o) E •. and M is zeromodified phasetype with representation (a+.1 Assume that the claim size distribution B is phasetype with representation (a. the Markov processes representing ladder steps can be pieced together to one {my}. however. T + to+). 3. {St} the claim surplus process. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. add a more selfcontained explanation of why of the phasetype structure is preserved.i. and rewriting in matrix form yields the phase generator of {my} as T + ta+. 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. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. i. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). Considering the first. Since the results is so basic. T) where a+ is given by a+ = . which occurs at rate ti.2. T). G+(. (b) V.3. T(0) < oo) the ladder height distribution and M = supt>o St. cf. a+j. Thus the total rate is tip + tia+.e. 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. T). 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. marked by thin and thick lines on the figure. B the claim size distribution.
. MATRIXANALYTIC METHODS t . see Corollary 2. 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).t t d kkt S.228 CHAPTER VIII.Q = 3 and b(x) = . 0 Example 3.7)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 . T = (3 .1 .1 This derivation is a complete proof except for the identification of a+ with ..M {mx} ST+(2)  S . This is in fact a simple consequence of the form of the excess distribution B0. 3e3x + .3. Figure 3.2 Assume that .QaT1..
1): Proposition 4.T). 3. That is.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+. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. we encounter similar expressions for the ruin probabilities in the renewal.1 In the zerodelayed case. We shall derive phasetype representations of the ruin probabilities V) (u). the discussion around Fig. this was obtained in Section 3. For the compound Poisson model. see Section 6.2 are taken from Gerber [157]. see Shin [340]. The result carries over to B being matrixexponential. cf. We assume p = PB/µA < 1 and that B is phasetype with representation (a. where a+ is the (defective) . In the next sections. T). THE RENEWAL MODEL This is the same matrix as is Example 1. The parameters of Example 3. if we define {mz} just as for the Poisson case (cf.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3. with A denoting the interarrival distribution and B the service time distribution. 3. 4 The renewal model We consider the renewal model in the notation of Chapter V. see Stanford & Stroinski [351] .j). but there the vector a+ is not explicit but needs to be calculated (typically by an iteration).6.6).and Markovmodulated models. the duality result given in Corollary 11. 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 .4.4. For further more or less explicit computations of ruin probabilities. Fig. his derivation of +'(u) is different.1 can be found in Neuts [269] (in the setting of M/G/1 queues. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. (a) G+ is of phasetype with representation (a+. For an attempt. and the argument for the renewal case starts in just the same way (cf. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). T) for some vector a+ = (a+.
1.Sy} in the same way as {mx} is defined from {St}.*} from {St+y . (4. Since the conditional distribution of my given T1 = y is ae4y. B0 is phasetype with representation (aT1/µa. with intensity matrix Q given by Q = T + to+. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). Hence by Theorem 11. where B0 is the stationary excess life distribution corresponding to B. cf.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. The key difference from the Poisson case is that it is more difficult to evaluate a+.T)• Proposition 4. Proof Obviously. But by Corollary 2.*'} is Markov with the same transition intensities as {mx}.1) Proof We condition upon T1 = y and define {m. the Palm distribution of the claim size is just B. Nevertheless.T). it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. 4. obviously mo = m.1).3 a+ satisfies a+ = V(a+).3. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. which for numerical purposes can be solved by iteration. Also. CHAPTER VIII. the calculation of the first ladder height is simple in the stationary case: Proposition 4. G(') = pBo. (c) {mx } is a (terminating) Markov process on E. where a(8) = aT1/PA.5. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy).6.2 The distribution G(s) of the first ladder height of the claim surplus process {Ste) } for the stationary case is phase type with representation (a(8). but with initial distribution a rather than a+. We have now almost collected all pieces of the main result of this section: Theorem 4 . Fig. In fact.T). Then {m.230 distribution of mo. Then .
3) (defined on the domain of subprobability vectors .1/pA.e. and that this is given by Proposition 4. i y ^ T1= y `•r Figure 4.0) is an increasing function of /3. thus .4. only with initial distribution a(*) for mo. the maximum claim surplus for the stationary case has a similar representation as in Proposition 4.. a+l ) = cp (a+°)) . i.. by a+ = lim a +n) where a+°) . a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . It remains to prove convergence of the iteration scheme (4.1) and a(8) _ aT.1).•.1(b). (4. In particular .2 ) follows from Proposition 4. . Furthermore . a+2) = ^p (a+l)) . Hence ^p(.. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case.1 by noting that the distribution of mo is a+. I {mx} . (4.3) Proof The first expression in (4.1 .0. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3.^(u) = a+e ( T+ta+)xe. .^(8)(u) = a ( 8)e(T+ta +) xe. THE RENEWAL MODEL 231 .2) where a+ satisfies (4.3). a+ can be computed by iteration of (4.M..2.
Assume the assertion shown for n .4) whenever EeR(S)U < oo.ST. the corresponding right eigenvector may be taken as (sI .4).T)'t • A[s] (4.T)1t.P[s] = A[s]B[s]. we use an argument similar to the proof of Proposition VI. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. 7+ ]. To prove the converse inequality. Similarly. 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). the normalization is equivalent to F(s) = 1.4. To this end.T1. limn4oo a ) < a+.5) Since s $ sp(T). MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. with B[s].f.1. Obviously.2.232 CHAPTER VIII.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). s ¢ sp(T). Fn ). In that case. and hence we may assume that h has been normalized such that ahA[s] = 1. (4. Proof Suppose first Qh = sh.) = P(mTl = i. a+ ) exists . so to complete the proof it suffices to show that &+ < a+) for all n. For n = 0.} can contain at most n . Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. Theorem 4.1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ).5 Let s be some complex number with k(s) > 0. Then (4. 0 0 We next give an alternative algorithm. F[s] being interpreted in the sense of the analytical continuation of the m. and let &+". Then each subexcursion of {St+Tl . However.5) yields h = (sI . both quantities are just 0 . Thus .. Thus by (4. (4.T)It. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . . let F be the distribution of U1 .g. Then F[s] = a(sI . n) &+n) T a+. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). this implies that ahA[s] # 0.
1 has the d distinct eigenvalues .T)lt = sh.9) we have G+[s] = 1 which according to Theorem 1. ..p1i . t(ry) > 0. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . Pd in the domain ER(s) > 0 .. 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.. In older literature . and define hi = (piI . As in Corollary 4.type with representation (a+. pdhd.T)It.T) = 1 ata+ = a+. hd. . 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 ).. and hence by the WienerHopf factorization identity (A. . the matrix Q in Theorem 2..6) i=1 i=1 Proof Appealing to Theorem 4.6 Suppose u < 0. hd. This gives d roots 'y. Hence with h = (sI T). 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. W v M(d) in the notation of Chapter V). . T) with a+ = a(QT)/at.4.T)lt + t = s(sI .5(c) means that a+(sI T)1t = 1. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . Q has diagonal form d d Q = dpivi®hi = dpihivi... Corollary 4. Since R(s) > 0 and G _ is concentrated on (oo. (4. Further. Pd with corresponding eigenvectors hl. Let d denote the number of phases. THE RENEWAL MODEL 233 Suppose next F(s) = 1. . Notes and references Results like those of the present section have a long history.6. in turn.lt we get Qh = (T + to+)h = T(sI .' that the equation F(s) = 1 has d distinct roots p1.. . and the solution is . yd satisfying R(ryi) > 0. Q = CD1 where C is the matrix with columns hl....... Given T has been computed.. Then G+ is phase..5. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. we get at a(Q . we have IG_ [s] I < 1 . 0).. D that with columns p1 hl. .
Neuts and his students. starting around in 1975. That is . Here phase. The exposition here is based upon [18]. the fixpoint problems look like R=Ao+RAI+R2A2+ . a pioneering paper in this direction is Tacklind [373].type assumptions are basic. the intensity matrix is A and the stationary row vector is ir . see Dickson & Hipp [118]. [270] and Latouche & Ramaswami [241]. The distribution of W comes out from the approach but in a rather complicated form . but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). For further explicit computations of ruin probabilities in the phasetype renewal case . The arrival rate in background state i is a. the ruin probability can be found in matrixexponential form just as for the renewal model. [119]. where R is an unknown matrix.type assumption . In risk theory. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. We assume that each B. E(t)). T('). 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI.. For surveys .. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. The solutions are based upon iterations schemes like in Theorem 4. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88].4. The number of elements of El=> is denoted by q. with representation say (a(' ). In queueing theory. Numerical examples appear in Asmussen & Rolski [43]. an alternative approach (the matrixgeometric method ) has been developed largely by M. The matrix. Asmussen & O'Cinneide [ 41] for a short self. involving . whereas the approach was introduced in queueing theory by Smith [350].exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. and the distribution of an arrival claim is B.. see Neuts [269]. 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. It turns out that subject to the phase.) d (see. e.g. and appears already in some early work by Wallace [377].234 then in transform terms CHAPTER VIII. the background Markov process with p states is {Jt}. is phasetype.F.contained derivation).
has states o. 5.Vt)} obtained by time reversing the I component. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. 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. Section 5b then gives a representation along the lines of Theorem 4.1. say with slope r(i) on intervals where It = i. p = ql = Q2 = 2. the phase space E(°) for B.1.1 In Fig. 5. 5a Calculations via fluid models. The key unknown is the matrix K.2. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. and the one E(•) for B. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. 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). This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims.6. However.5. •.4. O. states . Diagonalization Consider a process {(It. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. The two environmental states are denoted o.
F = E U { (i.1(a). a) of {It}. a) : i E E. whereas Ee8s' = oo for all t and all s > so where so < oo. we have more martingales at our disposal. consider the vector a satisfying (A + (13i(Bi[ s] . 4}. resp. 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. Eli) + Proposition 5.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . The intensity matrix for { It} is (taking p = 3 for simplicity) I A . Thus F = {o. 4. •. This implies that in the fluid context. 4. a E E(i) } . Let E denote the matrix . of E into components indexed by E. a) = 1.Vt)} is then obtained by changing the vertical jumps to segments with slope 1. r(i) _ 1. If s is such a number. o. 5. Recall that in the phasetype case. A claim in state i can then be represented by an E()valued Markov process as on Fig.1) if and only if s is an eigenvalue of E. 5.1))diag ) a = sa and the eigenvector b = . The fluid model on Fig .31a(l) (/3i)diag .1 A complex number s satisfies 'A+ (f3i(Bi[s] .92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. t. Bi[s] = a(i)(T(i) + sI)it('). First. MATRIXANALYTIC METHODS 4.(Ni)diag r(i. i E E. V.1))diag + sII = 0 (5. corresponding to the partitioning + Epp). j = 1. 2. F is the disjoint union of E and the Eli). < oo for all s.236 CHAPTER VIII. In the general formulation . Second.1(b) {(It . '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. in the fluid model Eel'.
sI 0 0 0 T(3) . resp .sI) (sI . and let d = (sI .E22)1 E21a.sI)1t)) iag I = 0 which is the same as (5. For the assertions on the eigenvectors.A . 0 . t(1) 0 0 then also 0 t(2) 0 .E22)1 E21a E21a . it follows that if Qla(1) 0 0 . with Eii replaced by Eii .(sI .sI.sI+ ((3ia(i)(T(i) . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .32a(2) (/3i)diag .5. Then (up to a constant) c = a.E12E22 E21 I . Then E21c+E22d = E21a .sI ()3i)diag . iEE (a> of 0* 1 AI.1). MARKOVMODULATED INPUT 237 indexed by E. E(1) + + E(P).E21a + sd = sd. d = (sI E22)1E21a = E ai(sI .T('))1t(i) .A . assume that a is chosen as asserted which means (Ell .sI + E12 (sI . Noting that E11c + E12d = se by definition. where c.sI 0 0 0 T(2) . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .E22 .E22)1 E21) a = 0.Nla(1) 0 0 T 1. c = a.sI 0 0 t(3) 0 0 = 0.
B2 are both exponential with rates 51 i b2..sv)b(v) = 0.. To determine 0 (u). a). MATRIXANALYTIC METHODS Theorem 5..238 CHAPTER VIII. a )d(a + e8 °vpi (u . j) pi( u ... . pi(u. v = 1.. . Proof Writing Or'Alb( v) = svb( v) as (AI . Then . v. u) Iw(u.a Solving for the pi(u. e89uc(e)) (d(1) .. Iw(u.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si.O.pi(u. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . a) = (j.. For u. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b .v) = v) I. . a) and noting that i1 (u) = >I j.3 Consider the Poisson model with exponential claims with rate 5.j)c v . Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u. Example 5 . c j. I' i( V P2 (w (u) < oo.4 that {e"1b(v) is a martingale . w(u)=inf{t >O:Vtu}.. j. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. .v)=inf{t >0:Vtu orVt=. j.v) = (j. j.4 Assume that E has two states and that B1. v > 0.v) = Optional stopping at time w (u.. v) yields C{V) = e8 .j. v. Thus 0(u) = esu/d = pe7 ° as u should be. . Example 5 . w(u. . d("))1 e.v}. q. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s.( u.upi(u. sq with $2s./' u = e' (esiuc ( 1) . j. Here E has one state only.Q. a))... v.v.v) = j).5. j. define w(u.v) = = p i( u . a) = Pi (Vw(u... it follows by Proposition II.a)d^ ). a)). Then we get V)i (u) as sum of two exponential terms where the rates s1. the result u follows.
2) the l.xxej • a 00 oo el . 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.s.33(e = 0 a(j))(K ®T ( j))(ej (9 I). (5. 0 Theorem 5 . is 0 /3 f R(i . Proof We must show that G+ (i. j.5.x) 00 f ° (') (j) eT (yy)edx . dx)Bj(y .( 2. according to VI.2.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. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.h. 8^')IT(j)) where e 3^') =. 9('). j. (y. •) is phasetype with representation (E(i). (') a T( However .6 For i E E. In terms of K. U) where t(j) + t(j)O(j j = k uja. the Pidistribution of M is phasetype with representation (E(1) + + E(P).b (u) = Pi(M > u) = 9(i)euue. i.5 G+(i. oo)) j)ye.3) .3j eye. j.k.
g. the initial value of (i. equivalently. we have the additional possibility of a phase change from a to ry within the ladder step. in many cases where such expressions are available there are classical results from the prephasetypeera which give alternative solutions under the slightly more general assumption that B has a Laplace transform (or. An alternative characterization is that such a distribution is matrixexponential.. bn1 bn).. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform.) which is rational. For a transition from (j. .5). say. see Example 1. a) is obviously chosen according to e(`).. which occurs w. This yields the asserted form of uja.240 CHAPTER VIII. However. i. 0 1)'. Bk7 . a) to (k.f. Numerical illustrations are given in Asmussen & Rolski [43]. (6.e. if b* [0] = b1 +b20+b302 +. the ratio between two polynomials (for the form of the density. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability.2.2) . and a new ladder step of type k must start in phase y.e. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . Associated with each ladder step is a phase process.. Starting from Jo = i. i.1 Let b(x) be an integrable function on [0. the current ladder step of type j must terminate.. which occurs at rate t(i).. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j).. intensity matrix U. Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). t) is the representation of the matrixexponential distribution/density): Proposition 6.k y. we have sofar concentrated on a claim size distribution B of phasetype. For j = k. and lifelength M. some square matrix T and some column vector t (the triple (a. t = (0 0 . T.y) to occur when j # k.. Furthermore. which occurs at rate t^^7. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . a m. and it just remains to check that U has the asserted form.p. Then b*[0] is rational if and only b(x) is matrixexponential. that the density b(x) can be written as aeTxt for some row vector a. +bn0i1 0n +a10n1 +.
2). S. bn of the denominator to be distinct and expand the r.s. 0 0 0 0 1 0 0 . cannot be phasetype.47x2 3 1 0 .. then b*[0] = a(0I T)1t which is rational since each element of (01 . . (6. but as follows from Proposition 6. shows that the distribution B with density b(x) = c(1 cos(21r x))ex. personal communication). Namely. we can always obtain a real one (a. u Remark 6. (6. MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 .(6. namely to asssume the roots 6l.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. . . s = c/ 2 .3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. One of his elementary criteria.6.1..1) as E 1 c.3) 0 0 0 0 0 .. The converse follows from the last statement of the theorem. t= 0 .47r2 3 . Example 6 . u giving b(x) = E 1 ciebiz/bY./(0 + bi)..3) was suggested by Colm O'Cinneide. T= 0 0 1 .2 A remarkable feature of Proposition 6. t).1 0 .h. b(x) > 0 for x > 0. .2). of (6. (6. For a proof. since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.. see Asmussen & Bladt [29] (the representation (6. T. s) is given by 27r i . Thus. 0 1 an an1 an _2 . where c = 1 + 1/47r 2. ..T)1 is so. Writing b(x) = c(e( 2ni1 ) y/2 .1 0 0 )3 = (111).an_3 an _ 4 . S = f c/2 0 21ri . 1 ..4) 0 0 1 c This representation is complex. matrixexponentiality implies a rational transform. a2 a1 Proof If b(x) = aeTxt.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.. 0 0 ..
we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). that despite that the proof of (6.5) Thus. For the first.242 CHAPTER VIII. and that the minimal number of phases in a phasetype representation increases to 0o as 5 . 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. Corollary 111.1 to get i (u) = f3esus.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)) . and can use this to invert by the method of Proposition 6.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . (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 6.5 (6. q are polynomials without common roots. T. we take as starting point a representation of b* [0] as p( O)/q(9) where p. t) a phasetype representation with a the initial vector. T. t) of b(x). Consider the distribution with density = 15 ((2e2x . leading to matrix calculus in high dimensions when b is small. 0. recall that t = Te) that if B is phasetype and (a. (6. As for the role of matrixexponential distributions in ruin probability calculations. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1. Then (cf.6) The remarkable fact is. we use a representation (a. and present two algorithms for calculating '(u) in that setting. then: Proposition 6. We recall (see Section 3.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way.3.1)2 + 6). T the phase generator and t = Te. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. For the second algorithm. MATRIXANALYTIC METHODS Example 6 .1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J.6) holds true also in the matrixexponential case. .
1BVA1.T)1t)1a +(9I .T)1 (91.6). 519) (A + UBV ). Now.T)1T 2 = and 1 = AB IT2 + 82T .T .T)1T1t.6b* .T)1 J0 00 b(x) dx = f aT1t. this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .1t = b* .1 = A1 .T).7) 9( cf.to+)1T .to+)1 = (BI .t.a+(9I .T)1 so that b* b** b** a+(9I .T)1ta+(OI .1 + b+ = b++ 1 .T)1t.B=land V=a+. .T .b* (6. Presumably.1UB(B + BVA1UB).to+)1T . (91.T)1 + (6I . we get b+ = 0aT1(9I T). the assertion is equivalent to a+(BI . but we shall give an algebraic proof.'t. b+ = a+(9I . From the general matrix identity ([331] p. since (91T)1T . xb(x) dx = aT2t. with A = 91T.1 = ^(T1 + ( 91T)1). b+ = a +(BI . U =.T .5 ).6.T)1t ( l . (6.A .1t = f3a (0I T)1T1t .1t du = .T)1 + 1 ib* (91. Then in Laplace transform formulation . MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .1 + 82 (9I . (6. we get (91.
T). a. 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.1t = /3a (9I . For expositions on the general theory of matrixexponential distributions.3 is taken). a key early paper is Cox [90] (from where the distribution in Example 6. 3. cf. The proof of Proposition 6. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0.1. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. of (6.8 a(T1 + (01.7). which is selfexplanatory given Fig. 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4.1. In Corollary VII. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory).1) is the same as the r. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization.h.1. the ruin probability(u) was found in explicit form for the case of B being exponential. VII. See Fig.5 is similar to arguments used in [29] for formulas in renewal theory.T)1T. 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. but the argument of [286] does not apply in any reasonable generality). Much of the flavor of this classical approach and many examples are in Cohen [88].s. 7.T)1)t = 8 (1 . Lipsky [247] and Asmussen & O'Cinneide [41]. /3aT1(0I .82b*. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. 0 Notes and references As noted in the references to section 4.la.8. From this it is straightforward to check that b**/(b+ .T)1T2t .1.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7.244 CHAPTER VIII. see Asmussen & Bladt [29]. . see the Notes to VII. MATRIXANALYTIC METHODS .b*).
i.u)e = A(u)e (7. t) is the vector of state probabilities for mt.7. Given the v(t) have been computed. t2) = exp where Q(t) = ds [P(t.t2) be the matrix with ijth element P (mt2 =j I mtl = i). Ai(t) = P(mt = i). Proof The first statement is clear by definition. is no longer timehomogeneous. Note that in general >iEE Vi (U) < 1.t)). in contrast to Section 3.e. t + s) . though still Markov. By general results on timeinhomogeneous Markov processes.1z I. Also. Figure 7. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . P(tl. we obtain V)(u) = P(m„ E E) = v(u)P(0.tl < t2 < u.1) where A(t) = v(u)P(0. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. Define further vi(u) as the probability that the risk process starting from RD = u downcrosses level u for the first time in phase i.1 The difference from the case p(r) = p is that {m2}. O<.I] I 80 { tq f Q(v) dvl t1 1 . >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). In fact. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. the definition of {m8} depends on the initial reserve u = Ro. 0 < t < u. Let P(tl.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u .
Thus. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u .Sj i)p(u)dt • tji = Sji + p(u)tji dt. Given A'. the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). from a computational point of view the remaining problem is to evaluate the v(t). 0 Thus.t)). the probability of which is 1 . the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 . those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . {mx} has jumps of two types. In the first case. Given A. jEE .t). given A. two things can happen: either the current jump continues from u + p(u)dt to u.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. 0 < t < u.t) for the second. A'(t) = A(t)Q(t) = A(t)(T + tv(u . Given this occurs. whereas in the second case the probability is p(u)dt • tjvi(u). 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.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). Proposition 7.(u) p ( u) = .2 For i E E. we get vi(u) = aidt + (1 . dt]. Hence Q(t) _ T + tv(u .t and being followed by a downcrossing. the probability that level u is downcrossed for the first time in phase i is ai. vi.3dt. (7. MATRIXANALYTIC METHODS However. or it stops between level u + p(u)dt and u. 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.Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}.246 CHAPTER VIII.
2. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7.4) backwards for {va (t)}v>t>o. F" etc. Let p" (t). we face the difficulty that no boundary conditions is immediately available. V . .) P"(AnBv) = P(AnB. (u) on both side and dividing by dt yields the asserted differential u equation. supRt>v l t<7 I where o.7. we can first for a given v solve (7. refer to the modified process. Now since both P(A n Bv) 3 0 and P"(A n Bv) .) + 0 as v + oo.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. of (7. (u) for any values of u and v such that u < v. (v) is given by the r.3 For any fixed u > 0.. and similarly P"(Bv) . after a certain level v.h. Thus.i7rT1/p.^ 0.. Since the processes Rt and Rt coincide under level B. denotes the time of downcrossing level u . This yields v. RESERVEDEPENDENT PREMIUMS 247 Subtracting v. then P(A n Bv) _ P"(A n BV').. consider a modification of the original process {Rt} by linearizing the process with some rate p. Rt . <oo. say.5). When solving the differential equation in Proposition 7. P u which implies that v. vi (U) = lim v= (u).) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo..0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. we have p(r) = p = vi (u) 0aTe.s. From Section 3. starting from v"(v) = . Then P(B. say. To deal with this.)P"(AnB.
p2. 1/n.1a. cf. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v. 0 < u < v.7) equals 01 (v . We recall from Propositon VII.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. Therefore u pl(vvueTa t 1. assuming u > v for the moment. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. 3u etc. for u > v the distribution of v . while the fourthorder RungeKutta method implemented in [30] gives 0(n5).e. p(r) P. let v(u) = a+2ieiT +ta+>)(uv). > 0} and the last term the contribu tion from {R. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII.1. r<v r > v. 2u. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. v = u.v v(u)eTat 1 1 . say. (7.7) f o (the integral is the contribution from {R. To evaluate p1(u). which is available since the z/i'(. However..10 that in addition to the O'(•). Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII.. the evaluation of Vi(u) requires q(u) = 1 .RQ (defined for or < oo only) is defective phasetype with representation (v(u). the probability of ruin between a and the next upcrossing of v.) are so. T).z51(v)).q(v dx +( ) ) = ( ) ( q( )) vueTva (7. such that Rt coincide with RI under level v and with Rt above level v.248 CHAPTER VIII.zp1(u)/(1 .. 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. 2/n. i. (u)}.V" M 0 . as well p1(u). Corollary 3. The f iin in (7. where v = inf It > 0 : Rt < v}. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. < 0}). The precision depends on the particular quadrature rule being employed. The trapezoidal rule used in [288] gives a precision of 0(n 3).9.1...1.x) dx f v(u)eT xt dx . where. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u.
1 from which we see that pl (u) = 1 + 1 249  1 .v) + (2^ + 3v2 ea'(u " .v(u ) eTV e J v(u)eTxtz/)l (v .^1(v) 1 .4 Let {Rt } be as in Example 3.21 = ? yields 0(u) = 1.24e.2.7. I.e6v Let Al = 3 + 2V'2.u . 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .x) dx} V 1 1(v) f V v(u) eTxt. RESERVEDEPENDENT PREMIUMS 1 . all quantities involved in the computation of b(u) have been found in matrix form.e. (7. Example 7.v(u)eTVe .v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u. so we consider the nontrivial case example p2 = 4 and p1 = 1.2V"2. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.8) The integral in (7.be the eigenvalues of T + to( 2 ). Since µB = 5/21.to+))11 {e{T®(Ttoy+ ))}„ .2. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .24ev .8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .x) dx 1 ^(v) ( 1 .(7 The arrival rate is (i = 3.and A2 = 3 . p2 < 3.e6u 35 .01 (v .v(u)eTve).. From Example 3.jl (t ®e) Thus.
21 3 In particular.b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived.1 V2 = 4e5"+6 35e6v .21(35e6v .24e5v .7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v../2 ea1(u") . Notes and references [30]./2) ea 1(u .250 CHAPTER VIII.2 35e6v .24es" .1)' ?. ) e sv + ( 2v/2. The analysis and the example are from Asmussen & Bladt .24e5v .24es" .+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .1 Thus. and one gets 12e5" . MATRIXANALYTIC METHODS From (7.1. 192esv + 8 P1 .v)esv + 7 4_ 2.
and instead we shall work within the class S of subexponential distributions . For further examples. we require that B is concentrated on (0. a2)) with density 1 e(logyFh) 2/2az .N(µ.2b. For the definition . x 2iror2 (c) the Weibull distribution with decreasing failure rate . for all t > 0.4. B(x) = ex0 with 0<0<1.B(x). For example. 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.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. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. x 4 oo. see I. 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. III. (b) the lognormal distribution (the distribution of eu where U . L(tx)/L(x) 4 1.g.46 and at numerous later occasions require a light tail.f. oo ) and say then that B is subexponential (B E S) if 251 . A rough distinction between light and heavy tails is that the m. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. the exponential change of measure techniques discussed in II.
then (with high probau bility) so are both of X1.v. 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. B(x) Here B*2 is the convolution square.2. X2 with distribution B. (b) liminf BB(() ) > 2. Thus the liminf in Proposition 1.1 Let B be any distribution on (0. X2) > u x)/B(x) = 2. one can check that x x where U is uniform on (0. proving (a).v.v. note first the following fact: Proposition 1.'s. P(max(Xi. In terms of r.p. X2) > x}. the behaviour in the lighttailed case is illustrated in the following example: Example 1. the distribution of independent r.2 If B E S.2B(x). To capture the intuition behind this definition. We later show: Proposition 1. That is. Thus .B(x)2 . Then: (a) P(max(Xi. 1). 1/2 'typical' (with distribution B) and w.252 CHAPTER IX. given X1 + X2 > x. 1/2 it has the distribution of X1I X1 > x. X1 is w. . X2) > x} C {X1 + X2 > x}. As contrast to Proposition 1. X2) > x) is P(X1 > x) + P(X2 > x) . P(Xi <yI Xi+X2>x) 1B(y). oo). in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. X2) > x) ^' 2B(x).1(b) is oo. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. we have {max(Xi. Proof By the inclusionexclusion formula. X2 > x) = 2B(x) . oo). X2 but none of them exceeds x.3 Consider the standard exponential distribution. That is. In contrast. (1. x 3 00. B(x) ax. if X1 + X2 is large .F(X1 > x. then P(X1>xI X1+X2>x)* 2.p. the r. Since B is concentrated on (0.1) then means P(X1 +X2 > x) 2P(Xi > x).'s X1. that is. and thus the lim inf in (b) is at least lim inf P(max(Xi. HEAVY TAILS B*2\ 2.
y] and (y.'s: if X . Proposition 1. Hence lim sup a+oo B*2(x) 2B((1 .y)/B(x) > 1 since y > 0. yo] as X + 00.y)/B(x) > 1. If X1 + X2 > x. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 . [In terms of r.v.] Proof Consider first a fixed y.B*(n ) B(dz) (1. then B(B(x)y) * 1 uniformly in y E [0.5 If B E S.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10. B( 0 . or they both exceed Sx. and combining with Proposition u 1. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0. We now turn to the mathematical theory of subexponential distributions. then either one of the Xi exceeds (1 .B(y) = 2.z B(x) . 1 < B(x ) B( x) Y) < B( 0).1(b) we get B*2(x)/B(x) * 2.S)x. we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) .B E S. If lim sup B(x .6)x)/((1 . a contradiction. x]. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1.yo]. we get limsupB*2(x)/B(x) < 2.4 Any B with a regularly varying tail is subexponential. Finally lim inf B(x . The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.B*n(x . This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1.S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . Let 0 < 5 < 1/2. then the overshoot X .2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0.xIX > x converges in distribution tooo.B(y)) .1.
z) B(dz).. choose y such that IB*n(x)/B(x) .z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1. Proof For 0 < 5 < e.1) for all large n so that B(n) > cle6n for all n. HEAVY TAILS Corollary 1.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . we have by Proposition 1. and this immediately yields the desired conclusions. O The following result is extremely important and is often taken as definition of the class S.nI < e for x > y.254 CHAPTER IX. The first integral is y B(x . b[c] = oo for all e > 0.B(x . x oo.z) B(x) Here the second integral can be bounded by B*n(y) B(x) . Proof We use induction.7 If B E S.z ) (x ) = 1 + (^ B(x . then e"R(x) * oo. its intuitive content is the same as discussed in the case n = 2 above.6 If B E 8.2. then for any n B*n(x)/B(x) * n. 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 . 0 Proof of Proposition 1. This implies B(x) > c2e5x for all x. so assume the proposition has been shown for n.5 and dominated convergence. Proposition 1.5 that B(n) > e6B(n . B(x) \Jo _ B(x . Given e > 0.B*2 (x) B(x) (x . The case n = 2 is just the definition. Then by (1.5 and the induction hypothesis.z) B(dz) _y B(x) 111 Lx B .y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. B*(n+1) (x I xy + Jxx y) W.2).
an = supx>o B*n(x)/B(x). X2 be independent r.z) B(x) < 1 + A + an sup f x B(x .3) Using the necessity part in the case Al = A2 = B yields f xv B(x .y)Ai(dy) v) f o . it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .i). Then by (1.8 If B E S. Since 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. A2 be distributions on (0.z) B(x .B(x . Then Al * A2(x) = P(X1 + X2 > x).9 Let A1.1.X1 > xv.z) B(dz) x .y)Ai(dy) = (x)o(1) (1. Then Al * A2 (x) . Proof Define 5 > 0 by (1+5)2 = 1+e. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al. 0 Proposition 1. 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. 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) .y))/B(x).4) .5 easily yields P(X1 + X2 > x. Proposition 1.v.X2 > xv) < A1(xv)A2(x v) . an+1 fX B*n( *n(x . choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). Combining these estimates and letting a 4. a2 with a1 + a2 > 0.y)B(dy) = B(x)ov (1)• v (1.'s such that Xi has distribution Ai. e > 0.(al + a2)B(x).z) B(dz) < 1 + A + an(1 + d) .0 completes the proof.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x . For any fixed v. 0 Lemma 1.2).ala2B(x)2 which can be neglected.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 . Xi <= v Ai (x .
We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. In the regularly varying case. B2 E S.aiB(v)) = B(x)o„(1). u Corollary 1. Hence Corollary 1. HEAVY TAILS 'VV B(x . whereas the two first yield B(x)(Ai(v) .9).y)Ai(dy) = B(x)o„(1). Proof Taking Al = A2 = A. then so is L = L1 + L2. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1. u Corollary 1. B1 * B2 (x) .12 Assume that Bi(x) = Li(x)lxa. then A E S.2aB(x) .v)B(v) + _'U Aq(x . i = 1. That is. Then A * B E S and A * B(x) .3) follows if CHAPTER LX.Ai(x . V (1.y)B(dy).h.11 Let B E S and let A be any distribution with a ligther tail.10 The class S is closed under tailequivalence. L2 slowly varying. a2 = 1.Bl (x) + B2 (x) when B1.2. f " By a change of variables.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1. if q(x) aB(x) for some B E S and some constant a > 0.5) Here approximately the last term is B(x)o„(1) by ( 1. it is easy to see that if L1. it should hold that B1 * B2 E S and B1 * B2 (x) . L2 are slowly varying.(x) is decreasing for x > x0 with limit 0 at oo. Then B E S provided fo "O exA(x) b(x) dx < oo. .5) becomes x B(x . A2 = B so that a1 = 0. the l. of (1.4). That is. with a > 0 and L1. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. A(x) = o(B(x)). u It is tempting to conjecture that S is closed under convolution.2A(x).256 Now (1. However.s.B(x) Proof Take Al = A.13 Let B have density b and failure rate A(x) such that .v)Ai(v) . a1 = a2 = a yields A*2(x) . B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown.
A(x .12. Then B(x) = eA(x). To illustrate how Proposition 1. Thus A(x) is everywhere decreasing.16 For L(x) slowly varying and a > 1.(x . we can use the same domination for the second integral but now the integrand has limit 0 ..3 < 1. an integrable function by assumption.y) * 0.1)xcl1 .`(x)b(x) is integrable. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Thus. Since ) (x .A(y)a(y ) = ev'(y) b(y).(y) dy. Thus by dominated convergence . the first integral has limit 1 .. Jo For y < x/2. elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. we first quote Karamata's theorem (Bingham. the DFR Weibull distriu bution is subexponential.y) dy. Then b(x) = Ox0lexp.14 Consider the DFR Weibull case B(x) = ex0 with 0 <. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. x .1. the u lognormal distribution is subexponential. Example 1. Goldie & Teugels [66]): Proposition 1. subexponentiality has alrady been proved in Corollary 1. Thus.1 has limit 1 + 0.y) < yA(x . and exa(x)b(x) = (3x01e(10)x9 is integrable.13 works in this setting. proving B E S. Further. a(x) = ax01. By (1. Define A(x) = fo . B*2(x) .y) < A (y) for y < x/2.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ).A(y)\(y) dy + fox/ 2 eA(x ). L(x) y° (a . In the regularly varying case. f ' L(y) dy .A(xy)A ( y). Thus B*2(x )/ B(x) .1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . 0 A(x) . The middle bound shows that it converges to b(y) for any fixed y since \ (x .e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx .U) /or) v 2x This yields easily that ex.15 In the lognormal distribution. Example 1.2).y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).
then B(x) . Then: Proposition 1. the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull. we get (1.xjX > x. Then 7(x) . 'y(x) = EXix>.1)] I X > x) L(x[1 + y/(a .ea b(x) is integrable. . yo] .258 From this we get CHAPTER IX.a/x.4 is necessary in full generality. Thus exa(x)b(x) .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. f O B(y) dy . then 7(x) x/(a .y(x)B(x). We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.L(x)/x" and )t(x) .E(X .18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.)/Y(x) > y) (1 + y/(a . More precisely. let X W = X .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 . Proof ( a): Using Karamata's theorem.1)]) xa L(x) (x[1 + y/(a .1) and P(X (.17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.13 may present a problem in some cases so that the direct proof in Proposition 1. HEAVY TAILS Proposition 1. (1 + y/(a . the monotonicity condition in Proposition 1.6) EX(x) . However.1)X(x)/x > y) = P(X > x[1 + y/(a .1)])a 1 1 .1/A(x) and P(X ixil'Y (x) > y) * e'.1))a . (c) Under the assumptions of either ( a) or (b).
Let St = Ei ` Ui .1 If Bo E S.t be the claim surplus at time t and M = sups>0 St. We get p(yl+. 0 G(u) L G(u) . . P The proof is based upon the following lemma (stated slightly more generally than needed at present).2. Then P(Y1 + • • • + YK > u) . r(u) = inf it > 0.. i. nG(u). be i. u a oo. with common distribution G E S and let K be an independent integervalued r. 1.15.(x). Theorem 2 .8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. .+YK> u) = ^•P(K = n)G* n(u ) .. Recall that B0 denotes the stationary excess distribution. cf. then Vi(u) P Bo(u).A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .and lognormal distributions ..n0 1•P(K= n)•n = EK. Kliippelberg & Mikosch [134]. Bo(x) = f0 B(y) dy / µB. The remaining statement (1.1/. Lemma 2.. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. Y2. Examples 1.nn.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. Notes and references A good general reference for subexponential distribution is Embrechts.14.v. with EzK < oo for some z > 1. d.EK G(u). We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo).A(x) I X > x) = exp {A(x) .2 Let Y1. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . It is trivially verified to hold for the Weibull.7) is referred to as 1/A(x) being selfneglecting. St > u}. and that for each Proof Recall from Section 1 that G*n (u) z > 1 there is a D < oo such that G*n(u) < G(u)Dzn for all u.
HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant. Proof of Theorem 2.. in our three main examples (regular variation . Bo E S is immediate in the regularly varying case. mathematically one must note that there exist (quite intricate) examples where B E S.µ J ) . For some numerical studies. See also Embrechts & Veraverbeeke [136]. The approximation in Theorem 2. a]..13).2. lognormal . (2.p)p'. Borovkov [73] and Pakes [280]. However.1 is notoriously not very accurate. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1.µB(01 . Weibull) one has Bo(x ( B(x) . Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. . then Bo(x)/B(x) + 00. see Abate. the result follows immediately from Lemma 2. P(K = k) = (1. r(1/Q) xlQexp B(x) = ex' From this . Bo ¢ S.260 CHAPTER IX. The problem is a very slow rate of convergence as u ^ oo.18. The PollaczeckKhinchine formula states that (in the setup of Lemma 2.?(xµ 8 (x).2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. u x+a Notes and references Theorem 2.p) and EzK < oo whenever pz < 1. Bo E S.1 is essentially due to von Bahr [56]. Bo is more heavytailed than B . Note that in these examples .x400 PBB(x) PB Leta+oo. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) .3 If B E S. as well as examples where B ¢ S. u The condition Bo E S is for all practical purposes equivalent to B E S. x 4 00. Since EK = p/(1.1) In particular .x^ ) B(x) _ f or ( lox .1. In general: Proposition 2. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) .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 ) .
. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V..g. in [219] p. i=1 .Ti.. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x .y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. E. + Xn. Based upon ideas of Hogan [200]. G+ (A) = P(Sq+ E A. We assume positive safety loading. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. p = iB /µA < 1. To Bo(u) u + 00. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S..} Then ik(u) = F ( M > u) = P(i9 (u) < oo). {n= 0. .9+ < oo) = P(S. i.1) this end .. T+ < oo) where r+ = T1 + • • • + T. T1 the ith interarrival time and Xi = U. In [1].1 when u is small or moderately large.] The proof is based upon the observation that also in the renewal setting. Define further 0 = IIG+II = P(r9+ < oo). there is a representation of M similar to the PollaczeckKhinchine formula. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA).1. This shows that even the approximation is asymptotically correct in the tail.+ E A. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. Snd) = Xl +. Somewhat related work is in Omey & Willekens [278]. M = sup s$ . also a second order term is introduced but unfortunately it does not present a great improvement. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. Let U= be the ith claim . Kalashnikov [219] and Asmussen & Binswanger [27].1 gives 1010.e. (3. Then K M=EY.. [279].3.y)/B (x) > 1 uniformly on compact y internals. The main result is: Theorem 3 . t9(u) = inf {n : Snd> > u} . THE RENEWAL MODEL 261 Choudhury & Whitt [1]..
Proof By dominated convergence and (b).Ti).2).d..y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. Proof Let R+(A) = E E'+ ' I(S. P(K = k) = (1 . (b) and does not rely on the structure Xi = Ui ..FI(x) /IPG_I. FI (x) _ fz ° F(y) dy.1) holds for a general random walk satisfying the analogues of (a). (3. The heuristics is now that because of (b). u a 00.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_. A(dy) = 1. are independent of K and i.FI(u). 0] normalized by IPG_ I so that we should have to G+(x) . HEAVY TAILS where K is geometric with parameter 9. d+ < oo). Then 0 0 F( x .1 IPG_ I / F(x .1) is equivalent to P(M > u) " . Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. 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. x > 0.PBBo(x).262 CHAPTER IX. the contribution from the interval (.2 F(x) . G_(A) = P(S.. Let F denote the distribution of the Xi and F1 the integrated tail.3) and we will prove it in this form (in the next Section.y) dy = 1 Pi (X) oo IPG_ I . B(x) _ J O° B(B(x)y) A(dy) f 1 .9)9'' and Y1. whereas for large y .N.3 G+ (x) . cf. Lemma 3 . 0 The lemma implies that (3. Write G+( x) = G+ ( x.y+ given r+ < oo). 0] to the integral is O(F(x)) = o(FI(x)). oo) = F(S.Y2. Lemma 3 . As for the compound Poisson model.i. and hence FI(x) .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. x + oo.oo.(. A. x * oo.g+ > x. U_ (dy) is close to Lebesgue measure on (.B(x). with distribution G+/9 (the distribution of S.
I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .9) 1 .1.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.3.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. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. n] + 1/I µG_ I.(dy) fN FI ( x) + lim sup ZY00 N F(x .2) yields 00 F F I (u) P(M > u) _ E(1 . and that U_(n .G+[s]) .y) U.1. Similarly.1.1 I . then by Blackwell 's renewal theorem U_ (n .e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10. F(Y= > x) FI(x)/(OIp _ 1). n] < (1 + e)/1µc_ I for n > N. We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x .3. (3. n] is just the probability of a renewal at n. choose N such that F(n . THE RENEWAL MODEL 263 We now make this precise.1. u Proof of Theorem 3.F[s] = (1 .=1 BIp G_ I (1.2). In the lattice case.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N. If G_ is nonlattice. By Lemma 3. and in the last that FI is asymptotically proportional to Bo E S.1. > (1 . the proof is complete.9)IpG_ I Differentiating the WienerHopf factorization identity (A. Hence using dominated convergence precisely as for the compound Poisson model. we can assume that the span is 1 and then the same conclusion holds since then U(n . 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .0)0k k I(u) A.UG_ I x. Given e.O[s])(1 .
P(M > u.Se(u)_1 < a) = o(Fj(u)).(u)+n . we have P(M E (u .. and {Su.So(u)) are available.a. 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. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA . HEAVY TAILS µF = (1 .Sty(u)_I < a} we have w(u) < oo. Therefore by Lemma 3.1 is due to Embrechts & Veraverbeke [136].0)ua_ .264 and letting s = 0 yields CHAPTER IX.2. on the set {M > u.a. u).l.4 For any a < oo. with roots in von Bahr [56] and Pakes [280].4 on the joint distribution of (S. Note that substantially sharper statements than Lemma 3.SS(u)}n=o.a.So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . u)) > P(w(u) < oo)(i lp (0))• On the other hand. Proof Let w(u) = inf {n : Sid) E (u .. Sty(u) . Notes and references Theorem 3. Mn < u}. Then P(M E (u .a. . S+9(u) .u)) = o(P (M > u)) = o(FI(u)).1)6+[0] .(1 .. S+q(u) .AB iP We conclude by a lemma needed in the next section: Lemma 3 . allowing also for possible dependence between the arrival process and the claim sizes. u)). see Asmussen & Kliippelberg [36].yiui_1. 10(0) But since P(M > u . must attain a maximum > 0 so that P(M > u.IIG+II)µc_ = (1 .a) N P(M > u). 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times.
For further approaches. t>0 S. G(x) (4.. We let F* denote the Podistribution of Si.1) . . examples and counterexamples. = Sx.. We give here one of them. 0o(u) etc.. We return to this point in Example 4.. and the distribution of {Sxk+t . 2.Sxi}0<t<x2Xl . Theorem 4... E0. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.1. and apply it to the Markovmodulated model of Chapter VI.4 below.i.... assume pp.F*(X) = P0(Si > x) . Figure 4.... Schmidli & Schmidt [47].4. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.Sxk}o<t<xk+1xk is the same for all k = 1.1 except for the first one) is a random walk.. {SX1+t . +1.X1 is the generic cycle. 4. (corresponding to the filled circles on Fig.. Thus the assumption . The idea is now to observe that in the zerodelayed case.. {Sn}n=o. Define S.. 4. See Fig. M. 4... 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 <. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i.1) is assumed. see [47].1 Note that no specific sample path structure of {St} (like in Fig.d.1 based upon a regenerative assumption.n n=0. < 0 and EoX < oo where X = X2 . such that {SXo+t  SXo}0<t< X 1Xo . M* = max S..1.1 where the filled circles symbolize a regeneration in the path.X2 < .. M = sup St..1 = max k=0.
Since clearly M(x) > Sl . jF11 F* (U).. the assumption means that Mix) and Sl are not too far away. Fo(Si > X). u p 00.3) where Mnx) = sup o<t<xn +1 X. 4.. The one we focus on is Fo (Mix) > x) .1 Assume that (4.2) 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. N N Xi=0 N Figure 4. Sxn +t .2) to show F(M* > u) > 1. it suffices by (4..2. See Fig. (4. Proof Since M > M*.4) liminf u>oo F(M > u) .* i o<t<xn+1x.2 Theorem 4. Then '00 (u) = Fo(M > u) . (4.266 CHAPTER IX. (4.Sxn = sup Sxn+t .S. HEAVY TAILS for some G such that both G E S and Go E S makes (3.3) hold..3) applicable so that F(M* > u) 141 F*(u).1) and (4.
u))/P(M* = 0) = o(Po(M* > u)). /3(u) = inf{n=1.(u) .6) 1 p pBo(u) u where B is the Palm distribution of claims and p . MW O(u)+1 < a) IN ( U n=1 A1.1 = limti00 St/t.. .1 can be rewritten as 00 (U) (4.4.4. Let a > 0 be fixed. Po(M* > u) .5) which follows since Po (M > u. We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. Theorem 4.2.4).. To this end..S.( u)1 > a) 00 1: Po(Mn<u.: Sn > u} . MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 . x > a.(1 . E (u .e)Po (M > u.a. Under suitable conditions . 0 yields (4. 2.. )) > (1 ..Mn +1 >aV(u n=1 00 S. assume the path structure Nt St = EUit+Zt i=1 . Mn+l > a V (u ..a. Letting first u + oo and next e . choose a such that Po(Si > x ) > (1 .e)Po (M > U). Then by Lemma 3. u)} < P(M* E (u .E) Po ( n max St u. Given e > 0.Sn 0<t<x„+j ( 1 .Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). S. M^xu)+l > a) .e)Po (MMX> > x).: S..Po (M* > u.
X both have tails of sup Zt. a4' 0. we get 00 (u) 1 IPF.2 Assume that {St} is regenerative and satisfies (4.} and satisfying Zt/t N. Corollary 4.6) holds with p = . (iii) For some o field Y. X and N.4).1 is in force.Q = EoNx/EoX. oX (see Proposition A1.I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 . Proof It is easily seen that the r.7). Mix) < > UE + i=1 o<t<x Thus Theorem 4. i=1 (4. since the tail of Zx is lighter than B(x) by (iv).8) x Write . and ENX Ui . independent of {> CHAPTER IX. and also for Mix) since Nx FNX U.p) Ju P Bo(u) 1p 0 .268 with {Zt} continuous. cf.6 below. HEAVY TAILS N` U. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) . are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4. (ii) EozNX < oo for some z > 1.v. The same is true for Sl. Assume further that (i) both B and Bo are subexponential. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ .3PB.'s order EoNx • B(x). the proof of Lemma 4.
The number N. in particular lighttailed...(NX).t + EN'I Ui where {>N`1 Ui . i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB.0 (thus (iv) is trivial). .5 Consider the Markovmodulated risk model with claim size distributions satisfying (4. MODELS WITH DEPENDENT INPUT 269 Example 4 . X3 = 2.4 Assume that St = Zt . Again . > 0. X2 = 1. we assume that B E S. Theorem 4.3 that (4. The key step of the proof is the following lemma. consider the periodic model of VI. of claims arriving in [0.6) holds.. (i) holds. 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.4. X2 = 1. Bo E S. and for some constants ci < oo such that cl + • • • + c. 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 . Zt . Assume that B E S... 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. The arrival rate is /3i and the claim size distribution Bi when Jt = i. Taking again Xo = Xi = 0. More precisely. and taking F = o. we conclude just as in Example 4. We consider the case where one or more of the claim size distributions Bi are heavytailed.9). X3 = 2. < 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.e. In particular.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).t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2.6) holds. Example 4 . 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds.6) u holds. i. . Then (4. . The regenerative assumption is satisfied if we take Xo = Xi = 0. (iii) is obvious. Thus we conclude that (4. Bo E S. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1.3 As a first quick application.
X i=1 j=1 where conditionally upon F the Xi. and the rest of the argument is then just as the proof of Corollary 4.ciG(x).F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 .}P. X > 0 a r. are independent with distribution Fi for Xij. NP ) and X are . It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni..2.Fmeasurable.. we can define the regenerations points as the times of returns to i. If Jo = i.. 2 . HEAVY TAILS Lemma 4 . . 6 Let (N1.c'(x) where c = ciENi . 1. . i1 = E\ G(x) In the general case. ."+Np . Let {Fi}t=1 P be a family of distributions on [0. NP ) be a random vector in {0. as x a oo. cp with cl + > 0 it holds that Fi(x) .F) < CG(x)zn'1+. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin .^•) G(x) P ^ E ciNi = C. For lighttailed distributions. P P P(YX and > x I.. i=1 Proof Consider first the case X = 0. . The same dominated convergence argument completes the proof.5. .270 CHAPTER IX.. i=1 P(Yx > x ^) < P(Y0 > x I.. u Proof of Theorem 4. Assume EzN1+"'+Np < oo for some z > 1 and all i. oo) and define p Ni Yx = EEX'i . and F a aalgebra such that (N1.G( x ) > ciNi . Thus dominated convergence yields ( P(Yo>x P(Yo>x .. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.v. Then P P(Yx > x) . and that for some + cp distribution G on [0.. . oo) such that G E S and some c1. . An easy conditioning argument then yields the result when Jo is u random.
we let PN"N = P(. Essentially. Theorem 4.i. 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.. Within the class of risk processes in a Markovian environment. Then O(u) .6) to hold in a situation where the interclaim times (T1.4. 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). as well as a condition for (4..e. the discussion provides an alternative point of view to some results in Chapter IV. cf.T2.pl(1 . this is applied for example to risk processes with Poisson cluster arrivals. ) form a general stationary sequence and the U. An improvement was given in Asmussen & Hojgaard [33].2 and Example 4.7)..d. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u.5. It follows from Theorem 4. Notes and references Theorem 4.. i. r(u) is the time of ruin and as in IV.3. In contrast. Schmidli & Schmidt [47]. As usual.T2.5 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. I T(u) < oo).. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). Combined with the approximation for O(u).p)Bo(u). m is a (orfinite) . The present approach via Theorem 4.4. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. i. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. Theorem 2. For further studies of perturbations like in Corollary 4. cf.1. and the final reduction by Jelenkovic & Lazar [213]. VI. cf.. this should be compared with the normal limit for the lighttailed case.4. 5 was first proved by Asmussen. see Schlegel [316]. The main result of this section.7. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5. ). Theorem 5.5 shows that basically only the tail dominant claim size distributions (those with c.4.1 is from Asmussen. IV. in particular Proposition 2. and independent of (T1. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. states that under mild additional conditions. We start by reviewing some general facts which are fundamental for the analysis.
h. in the terminology of general Markov process theory. j. for states i.s. follows by the substitution y = x . a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1.2) for all bounded measurable functions h. to consider only the case Px(w(F`) = 0) 0.). 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. m. y to vary in. however .1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . {St} and {Rt} are in classical duality w. 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} . Rt is distributed as x + t . the whole of R and not as usual impose the restrictions x > 0. y = 0). Lebesgue measure. w(Fc) < oo ) 'In general Markov process theory. The equality of the l.>N` Ui. k on E. u For F C E.t + EI U.2) with t = 1 means m. Thus.= y. Then (5. Sw(F. oo).z. . and starting from So = y. Let G denote the distribution of ENt U. and (5.rij = mjsji where r13. Proof Starting from Ro = x.y = Qx (. k as indicator functions.s.272 CHAPTER IX. say. a familiar case is time reversion (here m is the stationary distribution). The simplest example is a discrete time discrete state space chain. r. resp. Say {St} is reflected Brownian motion on [0.)k(x .z) dx G(dz) = ffh(y + z) k(y)dy G(dz). x = 0+ and F = (0. to the r. but the example of relevance for us is the following: Proposition 5.s=j are the transition probabilities for {St}.00). {Rt}. t. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. St is distributed as y .2) means ffh(a. where we can take h.h.t.t.r. (note that we allow x. For the present purposes it suffices . We let QS be the corresponding distribution and Qx. .
in E F. 0]. 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. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. z > 0. /^s x (S1 = Z1.3 The distribution of r(0) given r(0) < oo.s. In particular: Corollary 5. That is. 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. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. We can then view Qy. io = x. Sn+1 E Fc) nx.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. Thus. QR and QRy are defined similarly. and we let Qy y refer to the time reversed excursion . in = y. . i1. Sn = in = y.y = Qy Q. w(0. in with i0.y as a measure on all strings of the form i0i1 . x = 0.. Qx. . Sw(F)1 = y) .y() = P ({SW(F`)t} 0<t<w(F °) E So = x. [note that w(z) < oo a. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time.1 for the case F = (oo.. 5. Sw(Fo)_ should be interpreted as Sw(F^)_1). the one in (b) is the time reversed path. when p = ..13AB < 1] Proof of Theorem 5.5. But in the risk theory example (corresponding to which the sample paths are drawn).).2 Qy.. S.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. ..itt) = P Px(w(Fc) < 00.2..= y) Theorem 5 . The theorem is illustrated in Fig .SS(F.y(2p21 .
.. Silt' E SO k=1 i1.. in = x... (Fc)1 = y) 00 CHAPTER IX.....gilt' k=1 ii ... Sn+1 E Fc) n=1 i1. R Qy x(2p21 ... i0) Q x. .. 21 . ... Rn+1 E F`) F (w(Fc) < 00. in = x... Rn+1 E FC) TioilTili2.i„_iEF Similarly.. = in = y.in1 ... To show Q y x (i0 i 1 . Si l io E mjSjx. in)  Pt' (R1 = ii. MY Thus Qx(ioii .ii .ik_1EF ..... .TI( 2n2n _1 . . 2p). i0 = y. Si11 S 1 .... S. HEAVY TAILS E E Px (Si = 21i .. in) = Qx. .. Silt' E Sxik_1 .ik1EF Similarly but easier Sxin_1 .(F<)1 = Y) S S and Qx y( ipil ... .ik_1EF Sxin_1 . Si1y 00 jEF° E E 5xik_ 1 .. 20 = y. 2n) = Qx.rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 .in E F. 2p) when 20... note first that Pt' (R l = il. in E F.... Rn = in = x..y(inin _ E SYj jEF` 00 Sxik _1 ... in) = oo jEF^ Sxin1 . Si1y k=1 i1 . . S.. .........274 note that Fx(w(Fc) < 00..... Rn = in = x.... R . in with 20..J (i... t' y and Qy x are measures on all strings of the form ipi l .
')density of Y is B(y)/[.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . that is.')distribution of Yu is Bo"). that is.2 The distribution of (Y. see Fig. 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 .p.2. y > u. 7(0) < oo. Y > y} . That is.v.UBBo(u)].t. P(o) ).r.2. P(") = P(. 5. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y).r.(o) > y} = {T(0) < oo. Z) is described in Theorem 111. the distribution w. the case r (O) < oo. Y > u).2. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . ST(o) > y. U T(O) = T (u) Y Figure 5. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.'s are defined w.5. Now the P(u. the P(u.t.B(a) +a PBBo(u) . 1 w . S. Bo") is also the P(u. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y.
i. 2.r+ (n . HEAVY TAILS Let {w(z)}Z^. Recall the definition of the auxiliary function y(x) in Section 1. The idea is now to observe that if K(u) = n. cf.... Since w(z)/z a$. Zn).3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b).3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). Z/'y(u) * W in Pi "' ')distribution . Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. and YI. . .3. > u with high probability. z ^ oo. We let K(u) = inf In = 1. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 .. denote the ladder epochs and let Yk. .p). Now Bo E S implies that the Bo ")(a) + 0 for any fixed a.p) in Pi"'')distribution. and distributed as (Y.1. (Y. Yn_1 'typical'. Z). Bo") ). 1/(1 . .i. Fig...e. a slight rewriting may be more appealing.. It is straightforward that under the conditions of Proposition 1.3) holds. are i.p) then yields the final result T(u)/y(u) + W/(1 .d.T+(2). Then Corollary 5. .: r+ (n) < oo. it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . 5. . 4 Assume that Bo E S and that (5.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}.. Zk be defined similarly as Y = Y1. we get the same asymptotics as when n = 1. Then.. . and since its dominates the first n . must be large and Z1. let r+(1) = T(0). r(u)/Z 4 1/(1 .18(c) Bo")(yY (u)) + P(W > y) ( 5. Hence Z. P(Z < a I Y > u) 3 0. K(u) = n). then by the subexponential property Yn must be large. conditionally upon r+ (n) < oo. this in principle determines the asymptotic behaviour of r(u). Z = ZI but relative to the kth ladder segment. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long. the duration T+ (n) .. i.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. the random vectors (YI.. However. Z1). more precisely...p). That is . Then 7(u)/y(u) ^ W/(1 . Since the conditional distribution of Z is known (viz. In the proof. in particular of Z.276 CHAPTER IX..p) in F(u) distribution.e..
P(..u) E • I A'(u)) = Bo (n1) ®Bou) .5 Ilp(u..Bo (ri1) ®B( .Yn1iYn .. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events). A"(u) _ {K(u)=n} = {Y1+ P(. . suitably adapted). Y„1. II ' II denotes the total variation norm between probability measures and ® product measure. .3 In the following. Proof We shall use the easily proved fact that if A'(u). I A"(u ))II + 0. P (Yj. .Yl+ +Yf1>u}.n) (y1.n). . Further. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1.2... FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5. I A'(u)) = P(u. then IIP( I A'(u)) Taking A'(u) = {Y.5. +Yn1<u.u) II 0.u) E •) . Lemma 5. Yn .. > u}..
{wn(z)} be i. . . + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +..+y 1 p"F(Yn > u) P)Pn1 P/(1 . Zn) E •) .6.. ..P) Bo(u) for n = 1.y(u)T) ..p) < y). the density of Yn is B(y)/[IBBO(u)]. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic. the discussion just before the statement of Theorem 5. Let {wl(z)}. .u has distribution Bout That is. y > u. Proof of Theorem 5. (Y.. see Fitzsimmons [144]). the marginal distribution of Zk is Bo for k < n. .. Similarly (replace u by 0)... the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). Proof Let (Y11.. k = 1....P(Y' E •)II * 0. copies of {w(z)}. Z' are arbitrary random vectors.... For Theorem 5.2. Y1 +. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit.t. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf.. Y'.' = y is BM. .4). By Lemma 5.1..r. . and that Yk has marginal distribution B0 for k = 1.7 O (u. . Zn are independent.. in particular his Proposition (2.P(Z' E •)II > 0 (here Y. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + . in our example Y = (Y1. . and clearly Zi. n . +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 .278 Lemma 5 .d.. HEAVY TAILS ((Z1'. ... Then according to Section 5a.4. be independent random vectors such that the conditional distribution of Zk given Y.P) > y) Corollary 5. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). P(u) since by Theorem 2.6 IIPIu'n ) CHAPTER IX.i.1).. .. 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 •) . 2. Z. Y") u etc. Zn).).. Z11).1 and Y„ . The first step is to observe that K(u) has a proper limit distribution w..... n. n_1 < u.. then 11P(Z E •) . .Bo (n1) ®Bo' 0. wk(Zk) has a proper limit distribution as u + oo for k < n.1 P PBo(u) • P(W/(1 .
2. Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 . 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. Assume for simplicity that {Vt} regenerates in state 0 .B(u). one expects the level y form which the big jump occurs to be 0(1). Asmussen & Teugels [53] studied approximations of i (u. More precisely. = supo<t<0. and premium rate p(x) at level x of the reserve. and define the cycle as a = inf{t>0: Vt=0. i. 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). cf.. however. Theorem 6 .1.2 Define M. T) when T + oo with u fixed.1 Assume that B is subexponential and that p(x) > 00./3Ea B(u). Proof of Theorem 6. x > oo. that MQ becomes large as consequence of one big jump.y) . Corollary II. V. Then 0 (u) Qf "O ^) dy. claim size distribution B. nontrivial and we refer to Asmussen [22].6. . the results only cover the regularly varying case. that fo p(x)1 dx < oo.e. 3. the probability that is exceeds u is then B(u . Then P(MT > u) . p(Y) and the result follows. u (6. The heuristic motivation is the usual in the heavytailed area. 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 .(u) = P(V > u) = f f (y) dy . We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. The rigorous proof is.(3 u u J B(y) dy .1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}.
P(MT > u) $B(u) Ft µ(1 . Hence f (u)r(u) = D(u) = Do(u) . Then D(u) = f(u)p(u) and. by regenerative process theory.q ( u)) 1 .280 CHAPTER IX. It is also shown in that paper that typically.q(u) Now just use that p(x) * oo implies q (x) + 0. 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. where also the (easier) case of p(x) having a finite limit is treated . . 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). there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. u Notes and references The results are from Asmussen [22]. D(u) = DQ(u)/µ.
z) 2 = Zit NE ii ii According to standard central limit theory . We shall be brief concerning general aspects and refer to standard textbooks like Bratley.... Hence 1.2) is an asymptotic 95% confidence interval . Fox & Schrage [77].d.z) 4 N(0.96s z f (1. 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. The crude Monte Carlo ( CMC) method then amounts to simulating i. .Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . and this is the form in which the result of the simulation experiment is commonly reported. 281 . topics of direct relevance for the study of ruin probabilities are treated in more depth. 4Z). vrN(z . where a2 = Var(Z ). ZN. Ripley [304]. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . replicates Zl. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2.i.
Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). one can argue that unless Var(Z') is considerable smaller than Var(Z). The situation is more intricate for the infinite horizon ruin probability 0(u). Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. generated at the same time as Z. we then have EZ = EZ = z.b(u. Therefore. Then replacing the number of replications N by 2N will give the same precision for the CMC method as when simulating N' = N replications of Z'. SIMULATION METHODOLOGY In the setting of ruin probabilities. Say that Var(Z') = Var(Z)/2. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z.282 CHAPTER X. Letting Z' = E[Z I Y]. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). Further. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . However. and in most cases this modest increase of N is totally unproblematic. We mention in particular ( regression adjusted) control variates and common random numbers. This is a classical area of the simulation literature. We survey two methods which are used below to study ruin probabilities. and many sophisticated ideas have been developed. 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. there are others which are widely used in other areas and potentially useful also for ruin probabilities. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . so that Z' is a candidate for a Monte Carlo estimator of z. conditional Monte Carlo and importance sampling. and a longer CPU time to produce one replication. 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. v. an added programming effort. variance reduction is hardly worthwhile.
E [Z Z]2 = z2 . it appears that we have produced an estimator with variance zero.e. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . To this end. i. L such that z = EZ = E[LZ].3). . (1. but tentatively. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. even if the optimal change of measure is not practical. Nevertheless.3) Thus. Then z Var(LZ) = E(LZ)2 . GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. .1.96 sis v^ N 2 1 where srs = N j(LiZi . one would try to choose P to make large values of Z more likely. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). In order to achieve (1. (ZN. .v. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. However.. Thus we cannot compute L = Z/z (further. 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.zrs. and the problem is to make an efficient choice. L1). This may also be difficult to assess . the argument cheats because we are simulating since z is not avaliable analytically. it may often be impossible to describe P in such a way that it is straightforward to simulate from P). LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P.z2 = 0. Thus.. using the CMC method one generates (Z1.[E(LZ)] = E Z2 Zz .
This leads to the equation 1. large sample sizes are required. assume that the A(u) are rare in the sense that z(u) * 0. say of the order 103 or less. In ruin probability theory. N . u + oo. An example where this works out nicely is given in Section 3. and further it is usually not practicable to simulate from P(•IA). I. The CMC method leads to a variance of oZ = z(1 . a confidence interval of width 10 4 may look small. However. let z(u) = P(A(u)).5 or even much smaller .100 .B = iP(AB) = P(BIA).96 2 z2 z increases like z1 as z ..96oz /(zV) = 0. We then . Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . For each u.96 2Z ( 1 . i.0. Again. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem.z) which tends to zero as z ^ 0. and let Z(u) be a Monte Carlo estimator of z(u). Z z V5 In other words . A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. the optimal P is the conditional distribution given A. SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . as is the case of typical interest. Thus.284 CHAPTER X. if z is small.e. Z = I(A) and A is a rare event. in terms of the halfwidth of the confidence interval.e. say 10%. but if the point estimate z is of the order 105.z) 1001.e.1.1. 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. However.z) 1 > 00. we may try to make P look as much like P(•IA) as possible. 10 . it does not help telling whether z is of the magnitude 104. z I. To introduce these. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i..
p)pk. so that NE (u) may go to infinity. . Generate K as geometric. let Z +. Otherwise. Let M . 2.0.4) for any e > 0. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. This allows Var(Z(u)) to decrease slightly slower than z(u)2. and in practice. However. Thus. where M = X1 + • • • + XK.e.. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. O (u) = z = EZ. X2. Therefore . i. SIMULATION VIA THE POLLACZECKKHINCHINE FORMULA 285 say that {Z(u)} has bounded relative error if Var(Z(u))/z(u)2 remains bounded as u 3 oo. Generate X1. the PollaczeckKhinchine formula III. it is not efficient for large u . . 3.1) may be written as V) (u) = P(M > u). which gives a logarithmically efficient estimator . The algorithm gives a solution to the infinite horizon problem .i. According to the above discussion. with common density bo(x) = B(x)/µB and K is geometric with parameter p. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. but as a CMC method . Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. . where Z = I(M > u) may be generated as follows: 1.log z(u) of (1. XK from the density bo(x). F(K = k) = (1 . where X1.2. If M > u. see Asmussen & Rubinstein [45] and Heidelberger [190]. are i.log Var(Z(u)) lim inf > 2 u+oo ..4).. 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.d..X1 + + XK. Notes and references For surveys on rare events simulation. let Z +.p)pk.(2. The term logarithmic comes from the equivalent form .1. logarithmic efficiency is almost as good as bounded relative error. P(K = k) = (1 . it is appealing to combine with some variance reduction method . the mathematical definition puts certain restrictions on this growth rate.
XK_1).XK1] = EBo(uX1 .. < X(K) throw away the largest one X(K). 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. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail. we thus generate K and X1i . y < 0). we generate only X1... compute Y = u ..+XK > uIXl.X(2). asymptotically it presents no improvement : the variance is of the same order of magnitude F(x).p/(l . So.SK1)2..X(K1)) ... X1 + + XK_ 1 > x when X1 > x. just note that EZ(1)(u ) 2 > E[Bo (x . and considering only the remaining ones.286 CHAPTER X.. The idea of [27] is to avoid this problem by discarding the largest X.Xl . . K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X.. As a conditional Monte Carlo estimator . . However... and let Z(2)(u) = _ P (SK B0((u > u I X(l). Z(1)(u) is defined as 0).. . Xl > x.. . A first obvious idea is to use conditional Monte Carlo: write i..b(u) = P (Xl +•••+XK>u) = EF[Xl + . XK. Then (cf...X1 ..p)Bo(x). and that Bo(y) = 1.. and the problem is to produce an estimator Z(u) with a variance going to zero not slower (in the logarithmic sense ) than Bo(u)2. Thus.L(x)/x`' with a > 0 and L(x) slowly varying.. Theorem IX.....XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. assume in the following that Bo(x) .X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) .S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1). conditional probability.X(2). note first that To check the formula for the P(X(n) > x I X(1).1) V)(u) . XK1... Z(1) (u) has a smaller variance than Zl (x). form the order statistics X(1) < X(2) < .. For the simulation...2. To see this.
111.u is the representation 0(u) = e7sr(u) overshoot (cf. Asmussen . BL(dx) = e7sB(dx)/B[y]..modulated model P(r+ < oo) and G+ are not explicit ). Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . For practical purposes. X(2). .1) .S (n1)) V X (.Khinchine formula and importance sampling . X(2).. using 13L. BL instead of 0. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). in the renewal or Markov. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . and we refer to [27]. X(n1)) Bo((x ..S(n_1) I X(1). Then the algorithm given by { Z (2) (u) } is logarithmically efficient.5). and define )3L. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. X . the algorithm for generating Z = Z(u) is: 1.1 is elementary but lengty. . X(n1)) P(X(TZ) + S(. l)) . 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". Notes and references The proof of Theorem 2. 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. X(2). The algorithm is sofar the only efficient one which has been developed for the heavytailed case..._1) > P(X(n) > _ X X(1). Thus. X (.y. However . l)) BO(X(n1)) Theorem 2 . 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. . for the purpose of recording Z(u) = erysr(u). and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. . that is. B.. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. BL by I3L = /3B[y]. . and simulate from FL. .3.
Otherwise.Q. Let Sf0 CHAPTER X.l3 and U from B. Xi = U. + X. Let X1. let S. P'[y] < oo for some ry > 0. . r(u) < oo) and FL (both measures restricted to. let Z F e_'s.T. More precisely. and assume that µF < 0 and that F[y] = 1. Let FL (dx) = 'For the renewal model. one must restrict attention to the case 4µB > 1. The answer is no. Proof Just note that EZ(u)2 < e . the discussion at the end of Section 1b. BL could improve the variance of the estimator . namely ELery£(").2ryu _ z (u)2/C2. A > AL as in Chapter V. Generate T as being exponential with parameter . SIMULATION METHODOLOGY 3. 4. Let S .i. with distribution F. BL).(u)) are asymptotically coincide on {r(u)} < oo.. and avoid simulating the known part e7".. b different from . In fact: Theorem 3.7 tell that P(. = X1 + . so that changing the measure to FL is close to the optimal scheme for importance sampling . b) # (/3L. to deal with the infinite horizon problem . If S > u.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. M(u) = inf {n : S„ > u}. cf.QL. The algorithm generalizes easily to the renewal model . 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.1) (simulated with parameters ^3. There are various intuitive reasons that this should be a good algorithm. u It is tempting to ask whether choosing importance sampling parameters . In detail .d.2 The estimator (3.. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown. X2.. We formulate this in a slightly more general random walk setting '. . the results of IV.. B) is not logarithmically efficient when (/3.3. return to 3. be i. It resolves the infinite horizon problem since FL(.. and we have: Theorem 3.S+U . The proof is given below as a corollary to Theorem 3.F. and the change of measure F r FL corresponds to B > BL..288 2.r(u) < oo) = 1. Ti.
.3.yu = G.3 The estimator (3. When F # FL.. The importance sampling estimator is then Z( u) = e'rSM( ).. write W(F IF) _ F(XI). Here ELK..2 > 0.+KM(u)}.2ryELXi. where e' = EL Iog dFL (Xi) > 0 by the information inequality. it thus follows that for 0 < e < e'/ELXi. F(XM(u)).2ryELXi)} . EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+. Proof The first statement is proved exactly as Theorem 3 . EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2...i.d. By the chain rule for RadonNikodym derivatives. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). = c'. are i.P = FL. For the second. 1. K2.yu+elu u +oo etry' 1 > lim up C2e2.2'X1 . . (3. where Kl og (X) (j) 2 ) = log dFL (Xi) . More generally. is not logarithmically efficient. Since ELM(u)/u + 1/ELXi.2) dF Theorem 3.. + KM(u))} = exp {ELM(u)(E .. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) .2) (simulated with distribution F of the X3 has bounded relative error when .. Since K1. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + ..
This immediately yields / = 3". In [13].g. we conclude by differentiation that Bo(x)=B' (x)forallx > 0.2.3. U' .T' has a left exponential tail with rate /3' and U" .e. generic interarrival times T' . Further discussion is in Lehtonen & Nyrhinen [245]. so that one would expect the change of measure F 4 FL to produce close to optimal results. all that needs to be shown is that if U' . /3'eQ'YR'( x + y) dy = .3'eO'x f f P (U" . As in IV. the references in Asmussen & Rubinstein [45] and Heidelberger [190].290 which completes the proof.1 is from Lehtonen & Nyrhinen [244]. The optimality result Theorem 3.1 If y > 1/ic'('y). . BL) has bounded relative error.T". then /3' B' = B". The queueing literature on related algorithms is extensive .4.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". we write T = yu. 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.T' D U" .'(y) or y > 1/r. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L.'(y). T) with T < oo. 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 .T". see e.T" > x) J /3"e0 yB (x + y) dy = .T" has a left exponential tail with rate /3'. Then according to Theorem 3. First by the memoryless distribution of the exponential distribution . U' . In fact: Proposition 4.i.B'=B". with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45].T' = U" . The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. T". Next. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. Consider compound Poisson risk process with intensities /3'. CHAPTER X. yu) is close to zk(u). claim size distributions B'. B" and generic claim sizes U'. The results of IV. U". /3". The easy case is y > 1/k'(y) where O(u. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. SIMULATION METHODOLOGY u Proof of Theorem 3.
(ay). yu) in the sense that . Bay) is logarithmically efficient. (4. (4. T( u) < yu] e2ryyuEay le 2ay^(u). and in fact. T(u) < yu] e Hence by (4. 7y (4. Remark 4 .8). IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.2) Since the definition of ay is equivalent to Eay r(u) .1). We next consider the case y < 1/r.4.log Var(Z(u)) _ .4.(u) * 1 (Theorem IV. we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r.'(7). yu) is of order of magnitude a71.yk(ay) determines the order of magnitude of z'(u.O(u.3) and we have: Theorem 4.yu. one would expect that the change of measure F Pay is in some sense optimal. 3 Theorem IV. and that ryy > ry.3) (simulated with parameters /gay. Bounding u ELZ(u)2 above by a7u. . T(u) < yu] . yu)/z. Further . that ryy = ay .log 4')u) 4 u (Theorem IV. Proof Since ryy > y.yy> 2 .4. lim inf uoo 27yu .8 has a stronger conclusion than (4.4.1).2 The estimator (4. Let Qy2 = . We recall that ay is defined as the solution of a'(a) = 1/y. yu) = eayu Eay Leay^(u)+r(u)K(ay).log Var(Z(u)) l im of . The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu).1) which is all that is needed here can be showed much easier .5) follows. the result follows as in the proof of Theorem 3.1) so that z(u) = zP(u.to g x ( u ) u u so that (1.1.
ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV.b(u. (5.7y x(u) > hm inf u+Oo U .T) = P O<t<T inf Rt < 0 = P(VT > u). we believe that there are examples also in risk theory where (5.yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r. related discussion is given in a heavy traffic limit q J. yu .1) where the identity for Vi(u) requires that Vt has a limit in distribution V. yu .1) may be useful. In Asmussen [13].. 0 rather than when u 3 oo. 0 Notes and references The results of the present section are new. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).o .yu)/(uyu1/2) . In most of the simulation literature (say in queueing applications). and (5. One main example is {Vt} being regenerative (see A.u1/2 < r(u) < yu Le ] l = e.(av)Eav l e.4).1) (see Proposition IV. 5 Regenerative simulation Our starting point is the duality representations in 11. However.2) .2). N(0.1) is used to study Voo by simulating {Rt} (for example.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu .Qyu1/2 < T(u) C yu e.1): then by Proposition A1. Hence lira inf log ryyu + vyu 1/2 tc(ay) . '%(u) = P I info Rt < 0) = P(VV > u). zi(u) = INV. > u) = E f I(VV > u) dt 0 (5. there exists a dual process { V t} such that i. Z (u)2 above.3.4.292 CHAPTER X.3: for many risk processes {Rt }. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W).uaoo U That lim sup < follows similarly but easier as when estimating En.a vt(u). the object of interest is {Vt} rather than {Rt}.o.4. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .
. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. the regenerative estimator z%(u) is consistent.. letting J0 'o I (Vt > u) dt . j = 1. record Zi'i = (Z1').h (zl.. oh) for h : R2 ^ R and Ch = VhEVh.5.. To derive confidence intervals . which we survey below . Then (Z1z1i Z2z2 ) 4 N2(0. z2) z2/z1 yields Vh = (z2/z2 1/zl). Then Z(1). provides estimates for F ( V. i. > u) (and more general expectations Eg(V. Taking h(zl. and Z2'>) where Zi'i = w. Zl the LLN yields Z1 a$' Z(1) +. .. For the ith cycle. 2.. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi.3) . Z2 = N (X21' + .. z2)) > N(O. Thus. i (^(u) . )). Vh = (8h/8z1 8h/ 8z2).. EZ1'i = z1 = Ew. . For details .+Z(N) z 1.. The method of regenerative simulation. + Z2N)) .d. + Z1N>) . REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. Z(N) are i . Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ').t(u)) 4 N(0. 02) (5. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. Therefore .. Z1 = (Zl1i +.. EZ2'i = z2 = E Thus.E). Z2 a4* z2. is the cycle length. consider first the case of independent cycles . let E denote the 2 x 2 covariance matrix of Z('). Z2 . Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed.
Z of the form Z = ^p(X) where X is a r . consider an extremely simple example .5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. Rubinstein [310] and Rubinstein & Melamed [311]. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . the expectation z = EZ of a single r. with distribution depending on a parameter (. However .0 .g S12 (5. Before going into the complications of ruin probabilities . () dx so that differentiation yields zS d( fco(x)f(x. () dx = E[SZ] f(X. Then z(() = f cp(x) f (x.2. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .9.v.96s/v"N. v. In 111. We here consider simulation algorithms which have the potential of applying to substantially more complex situations. Notes and references The literature on regenerative simulation is extensive. 6 Sensitivity analysis We return to the problem of 111 . There is potential also for combining with some variance reduction method.Z) ^Z(=) . 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. to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process .294 where 01 2 CHAPTER X.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . in some situations it may be the only one resolving the infinite horizon problem .2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . asymptotic estimates were derived using the renewal equation for z /i(u). () depending on C. 9. The regenerative method is not likely to be efficient for large u but rather a brute force one.C)dx = f w(x) d( f ( x. () dx f Ax) (dl d()f (x' () f ( z.g. see e.
11 /3oeOoT. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. () = (8/8()h (u. ()). () Thus. To see this. ()) h((U. C) f(X. The following example demonstrates how the SF method handles this situation. cp(h(U. I(r(u) . = E [`d (h(U. For example . C)). SZ is an unbiased Monte Carlo estimator of z(.log U/(. () can be generated as h(U. for some Co = (o(U). /3o is M(u) Oe (3T: < oo) . () = . Then . cp' (h(U. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. For IPA there are.t. the Poisson rate /3 in the compound Poisson model. () is increasing in C. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation.v.() d( is the score function familiar from statistics . IPA will estimate zS by 0 which is obviously not correct. Thus.r. Then z(() = Ecp(h(U. say W(x) = I(x > xo) and assume that h(U. () is an unbiased Monte Carlo estimator of zS. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators .1 Consider the sensitivity tka(u) w. In the setting of ruin probabilities . one can take h (U.6. () where U is uniform(0. C). r(u) = Tl + • • • +TM(u)). ( where h( (u. Example 6 . A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. this is usually unproblematic and involves some application of dominated convergence . one.r. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. ()) d( hc(U. p. if f (x. with density f (x. For the SF method. Infinitesimal perturbation analysis (IPA) uses sample path derivatives. () _ (eSx. Let M(u) be the number of claims up to the time r(u) of ruin (thus. ()) is 0 w . Thus .t. /3 is 0. zc = E [d co(h(U. () = log U/(2. The likelihood ratio up to r(u) for two Poisson processes with rates /3. So assume that a r.1). () = d log f (X. just take cp as an indicator function . however . giving h( (U.
Thus. There have been much work on resolving the difficulties associated with IPA pointed out above.3 (u) is of the order of magnitude ue7u. BL).3 (u) (to generate Zp (u). different parameters.0(1) so that in fact the estimator Zf(u) has bounded relative error. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. We recall (Proposition 111. a relevant reference is VazquezAbad [374].T(u)) I(T(u) < co) ] . j3 and letting flo = 0. We then arrive at the estimator ZZ(u) = (M(u) . we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) .9 . 4) that V5. whereas for the SF method we refer to Rubinstein & Shapiro [312]. differentiating w.1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. change the measure to FL as when simulating tp(u).3L.t. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu .296 CHAPTER X. in part for different measures of risk than ruin probabilities.r.t.r. However. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . Example 6. for different models and for the sensitivities w. In the setting of ruin probabilities. . 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial).T(u)) e7uerVu) for ?P. the risk process should be simulated with parameters . To resolve the infinite horizon problem . SIMULATION METHODOLOGY Taking expectation.
}). a) = r(u) A T+(a). either this makes no difference (P(R.1}valued . a) = r(u)).1..d.. That is. The twobarrier ruin problem The twobarrier ruin probability 0.. defined as Ro = u (with u E {0. in most cases . Besides its intrinsic interest . Y'a(U) = P(T (u) = r+(a)) = 1 .i.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion. X2... R„ = u+X.. as e. and {1..(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo. 297 . wherel T(u) = inf {t > 0 : Rt < 0} . T+(a) = inf It > 0 : Rt > al.P(•r(u. . where X1. in the Bernoulli random walk example below.+• • •+X. are 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 ). Consider first a Bernoulli random walk. T(u.g. with P(Xk = 1) = 9..(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u..
one elementary but difficult to generalize to other models. and in view of the discrete nature of a Bernoulli random walk we write z = e7. We choose a = ry where ry is the Lundberg exponent.r(u..y] = 1. u + 1.. zu = EzRO = EzRT(u.(4.2).2) Oa(a .1) is solution.1) = (19)4/'0(a3)+9ba(a1).1 For a Bernoulli random walk with 0 0 1/2. Wald's exponential martingale is defined as in 11. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2).1.o)'t/1a(a . i.o)T/la (1) + 8z/'u(3).. MISCELLANEOUS TOPICS Proposition 1.0)/0. In a general random walk setting . By optional stopping. z and the solution is z = (1 .1) o If 0 = 1/ 2. = (1 .a) Y.. The martingale is then {zuzXl+•••+X„ } = {zR° }. then 'Oa(u) _ au a We give two proofs .e.a(u)). C1_0\a.4) by ea(u+Xl+. where a is any number such that Ee°X = F[a] <oo. tba(2) _ (1 .(u) I\ e = 1 oa ' ()i a = u. and the other more advanced but applicable also in some other settings. u Proof 2... (1.+Xn) F[ a]n n=0. the solution of F[.298 CHAPTER XI.a) = 0) + zap ( R.. The Lundberg equation becomes 1=F[ry]=(19)+9z.o» = z°P (RT ( u. and insertion shows that ( 1..(1B)u oJ 0. Proof 1. = z°Va(u) + za(1  . 7/la(a ..
then Vi(u) = 1.u)/u. pa( u) _ u Corollary 1.0a(u)).3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance .1) for p # 0.} 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)). u Proposition 1.1).• a2µa e2µu .1 yields 't/la(u) = (a .5) . Corollary 1. i1(u) = e211 . 1h (u) = a el u \1 If 9 < 1/ 2. RANDOM WALK. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .zu)/(za . {R.1).1 If p = 0. Applying optional stopping to the exponential martingale {e7R. Proof Let a+ oo in (1. and solving for 9/la(u) yields Z/)a(u) = (e 76 .e7u)/(e7° . .ba(u) = e2µa .1. Then for p 0 0.4 For a Brownian motion with drift u > 0. If p = 0.2 For a Bernoulli random walk with 9 > 1/2.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. thenz1 (u)=1. However. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . If 9 = 1/2. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. If p<0. then Proof Since 'Oa (U)  au a Eea(R°. (1. BROWNIAN MOTION. (1.2) is trivial (z = 1)..
6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. For most standard risk processes .. Here is one more case where this is feasible: Example 1. this immediately yields (1. valid if p < 1 (otherwise . 5).a) = a) = 5 y = P (R (u.7). It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1.5 Consider the compound Poisson model with exponential claims (with rate.a) < 0) + e7°P (R(u. 1 . (u) _ O(u) .a) = r+ (a)} and similarly for the boundary 0.616). 0.a) = a on {r (u. 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 Again . Here the undershoot under 0 is exponential with rate 5.+^a(u))^(a) If 7k(a) < 1. we obtain 'Oa a7u . 7O(u) = 7/la(u) + (1 .4). Ic 5ry 'pa(u) Using y = 6 . However.e7a (u) = 6 /0 .7) . and hence e7u = Ee7Ro E [e7R(.a ) < 0) + e 7aF ( R (u. a) I R(u a ) < 0] P (R(u .0(a) 0 < u < a.7/la(u)). passing to even more general cases the method quickly becomes unfeasible (see.300 Proof Let a * oo in (1. say. letting a * oo yields the standard expression pe7u for . CHAPTER XI. 7/'(u) = 1). however. VIII. implying R(u.0 (u) (where u p =. and thus one encounters the problem of controlling the undershoot under level 0. (1.vi(a) Proof By the upwards skipfree property.a) = a ) + e ' ° ( 1 .5a).3. the paths are upwards skipfree but not downwards.
7 For Brownian motion with drift 0. BROWNIAN MOTION. MT > u) = P (ST > u) + P (ST > u. Then the density and c.µ T I + e2µ"4) ( . (i). and hence Pµ('r(u) E dT) = Eo [e µsr(. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. the density dPµ / dP0 of St is eµsttµ2/2.8 Let {Rt} be Brownian motion with drift . RANDOM WALK. MT > U) = P(ST > u) + P(ST > u) (1..r(u). we have ili(u.11 ) is the same as (1.T) P(MT > u) where MT = maxo<t<T St.1a for computing ruin probabilities for a twostep premium function. ( 1.d. Hence P(MT>u.. (1.11) VIT ) Proof For p = 0.. + µ2T) } .4) I = . T ) = P(T(u) < T ) = 241.8 ).µ%T (1.8) Proof In terms of the claim surplus process { St} = {u . 10) follows then by straightforward differentiation.µ so that {St} is Brownian motion with drift µ . For µ # 0..Rt}.ST>U). 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 . P(MT > u) = P(ST > u) + P(ST < u.10) Pµ (T(u) < T) !. Here {St } is Brownian motion with drift 0 (starting from 0). = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T . TWO BARRIERS 301 Note thas this argument has already been used in VII. Corollary 1. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities.2 . = 1 .f. and (1 . T(u) E dT. 0(u. (1.1.ST<u) = P(MT>u..9) = 2P(ST > u).)_ _( u)µ2 /2. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .
T+2. Proof The argument leading to ( 1. 226). T are integervalued and nonnegative. Here {2T( (v}TT)/2) v=T. see e.T) = P(ST = u) + 2P (ST > u)..g. oo).3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion. Breiman [78] or Karlin & Taylor [222] p. MISCELLANEOUS TOPICS which is the same as (1.13) with 0 as lower limit of integration. e. as defined in (1. is (1.8 also applies to the case 9 54 1/2. is finite for all x > 0. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.T)dx. S(oo) = f c s(y)dy. the behaviour at the boundary 0 is more complicated and it may happen. Theorem 1.10). We finally consider a general diffusion {Rt} on [0.9) goes through unchanged. The same argument as used for Corollary 1. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). Vi(u..12) is the same as ( 1. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 .. and (1.12) P(ST = v) = 0 otherwise.10 Consider a diffusion process {Rt} on [0. The expression for F ( ST = v) is just a standard formula for the u binomial distribution.s. u Small modifications also apply to Bernoulli random walks: Proposition 1.h.T (1.9 For Bernoulli random walk with 9 = 1/2. and in a similar spirit as in VII. whenever u.9).T2. oo) with drift µ(x) and variance a2 (x) at x. as defined above as the probability of actually hitting 0.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). If this assumption fails. Let s(y) = ef0 ry(.g. that 0(u). Thus. 0 0 (1.11) then follows by checking that the derivative of the r. S(x) = f x s(y)dy. (1.10) and that the value at 0 is 0.. but we omit the details. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions.302 CHAPTER XI.
b(u) = S(a) . 191195 for material related to Theorem 1.ba. Then YIa.e LVa. 0 Proof of Theorem 1.b(u) be the probability that {Rt} hits b before a starting from u.b(Rdt).(u) < 1 for all u > 0 and ^ S^ Conversely. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. [117].b(u) + L. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .b(b) = 1.b = 0 implies that VQ b/s is constant. Lemma 1. (1 .14) S(oo) < 00.ba.16). If (1.6 to Markovmodulated models .11 Let 0 < b < u < a and let t&0. If b < u < a.10. and we get Wo. E„ q(Rdt) = q(u)+Lq(u)dt.b(u)dt. so that Y)n. S(oo) < oo separately u completes the proof. if (1. In view of (1. . Letting a T oo and considering the cases S(oo) = oo. Assume further that S (x) as defined in (1.b(a) = 0 then yield the result. (1. Further references on twobarrier ruin problems include Dickson & Gray [116].S(u) (1. Using s'/ s = 2p/a2. the function S(x) is . Letting b J. then. 0 in (1.16) S(a) . BROWNIAN MOTION. Wa.13) is finite for all x > 0.14) fails.b = a+/3S. i. RANDOM WALK. 1'.10.17) Hence L.1.b (Rdt) = Oa. A good introduction to diffusions is in Karlin & Taylor [222]. A classical reference for further aspects of Bernoulli random walks is Feller [142]. 15) i. see in particular pp.b('u) = Eu . we can ignore the possibility of ruin or hitting the upper barrier a before dt.e. Notes and references All material of the present section is standard.S(u)/S(a). For generalizations of Proposition 1.0(u) = 1 for all u > 0.b('u) = Eu &0.S(b) Proof Recall that under mild conditions on q. The obvious boundary conditions '0a. TWO BARRIERS 303 for x > 0. O. b = 0. then 0 < 2l. see Asmussen & Perry [42].16) yields 4b (u) = 1 .
(2.o•K(a) = Ee .5. yu) '+/1(u) .ytc (ay). Markovmodulated Brownian models . is currently an extremely active area of research.3) < e 7yu.13)).aR. much of the literature dels with the pure drift case.(. variance 0. Lo is a martingale (cf. 1 y < k (y). See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.6.1 ) was given already in II.(T(u)AT) r. 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.)AT .17). Another basic quantity is the speed measure M . Remark 11. correponding to piecewise linear paths or . The emphasis is often on stationary distributions .1.1) (2. yielding eau = Ee. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. (2. Lo I.(7) .6) . which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. 111 .4. one works instead with a lower limit 5 > 0 of integration in (1.. and here are alternative martingale proofs of the rest . defined by the density 1/va(u)s(u) showing up in (1.3.5): _ z/'(u) < e 7u.(a) (2. (2. equivalently. with the drift and the variance depending on an underlying Markov process . where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.t&(u.304 CHAPTER XI.aRo .2) C_e7u < t(u) < C+e _7u. yu) where W (ay) = y.5) A martingale proof of (2. IV.9 ) and optional stopping applied to the stopping time r(u) A T. (2.4. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. y > . 7y = ay .4) I. information on ruin probabilities can be obtained . but by duality.
eyuk (ay) = e7yu e > eyu"(ay ) ij(u. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).f. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.yuk (ay)(u&(u. and the proof of the lower inequality is similar.d.1 . y > r.2): As noted in Proposition II.7R. For (2. Let H(dt. A claim leading to ruin at time t has c. u Proof of (2. dr) denote the conditional distribution of (T(u). when Rt_ = r.)r(u)r. it follows easily from (2. (B(y) . . dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows.(u. Rt has distribution B(r + dy)/B(r).4).T)  V.( u ) I T(U) < 00] . Equivalently. (2.yu))• b(u.6) with = 'y that eyu .E [e. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.3). RT(u)_) given r(u) < oo. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.yu) Y Similarly for (2.6).. Proof of ( 2.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0).6) below by 1 E Le7Rr(. 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).1.B(r))/B(r). we have tc(ay) > 0 and we can bound (2.3).1. so that i/1(uL yu) < eayu .2. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e.4): We take a = ay in (2.
ri.. gn 4 0.3na with a < 1. nroo n n /// Note in particular that (3.1).1) does not capture the \ in (3. in being capable of treating many models beyond simple random walks which are not easily treated by other models .1 We will go into some more detail concerning (3. and gave sharp asymptotics for probabilities of the form P (S. Cramer considered a random walk Sn = X1 + . not quite so much in insurance risk. cle . logarithmic asymptotics .the correct sharp asymptotics might as well have +./n E I) for intervals I C R.. we will write fn 1. however .(B) = log EeOX 1 is defined for sufficiently many 0. . The last decades have seen a boom in the area and a considerable body of applications in queueing theory. .?n typically only give the dominant term in an asymptotic expression . og For sequences fn. and that a considerable body of theory has been developed.2) can be rewritten as F (Sn/n > x) 1g a'fin. The limit result (3. + X. However . if x > EX1.^ e nn 1 > x n 0o 2xn (3. v2 later.nn or C2e.g. Example 3..1) where we return to the values of 0.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17.. The classical result in the area is Cramer's theorem. which in the setting of (3. (3. large deviations results have usually a weaker form.means (as at other places in the book) that the ratio is one in the limit (here n * oo).2). then P C S.gn if nioo lim 109 fn = 1 log gn (later in this section. 1) but only the dominant exponential term . (3. For example. Thus . Accordingly. large deviations results been. gn with fn + 0 . The advantage of the large deviations approach is. Thus. the parameter will be u rather than n).306 CHAPTER XI. its generality. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative .1) is an example of sharp asymptotics : . 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. e. such that the cumulant generating function r.
4) n Next. Since P nn > x) = E {e_8 ' ( 9). the LegendreFenchel transform or just the Legendre transform or the large deviations rate function).3) is put equal to x. In fact. P with mean nx and variance no. of P(X1 E dx) = E[e9X1K.3.sseo f which in conjunction with (3. i. V > 0 e. nx < Sn < nx + 1.4) immediately yields (3.2). exponential change of measure is a key tool in large deviations methods.r. S rtn > x 1. which is a saddlepoint equation . rc*(x) = sup(Ox .4 enn +1. Define .e.425. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .960/) * 0. we get P(Sn/n > x) E [e9nx +nK(9)9" '. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3. since Sn is asymptotically normal w. Most often.96o /] > 0. (3.r.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.q = rc* (x). if we replace Sn by nx + o / V where V is N(0.rc(0) where 0 = 0(x) is the solution of x = rc'(0). and hence for large n P(Sn/n > x) > E [e.1). LARGE DEVIATIONS Define rc* as the convex conjugate of rc.(e)i XI E dx]. we have P(nx < Sn < nx + 1.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn .t.9S„+n' ( 9).(0)) e 307 (other names are the entropy. 2 where o2 = o2(x) = rc"(0). More precisely.
. We further write µ = tc'(ry).. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. there exists z E (0... 260 for details. MISCELLANEOUS TOPICS which is the same as (3.s.3 For each i > 0...p > 7 < zn. Then i/.v./^ >7 < zn n for n n0. Pn Sn1 .. and write Sn = X1 + • • • + Xn.. (ii) lim supn. 1) and no such that Sn .. . In the application of large deviations to ruin probabilities. to be made rigorous. Xn) and sn = x1 + • • • + xn (note that the r. .e < 8 < y + e. which is of similar spirit as the dicussion in VII. integrates to 1 by the definition of Icn).dxn) = 05nKn(7)Fn(dx1.. see Jensen u [215] or [APQ] p..h. . (iv) tc(ry) = 0 and r. X2..dxn) where Fn is the distribution of (X1i . Further main results in large deviations theory are the GartnerEllis theorem. that is.3. n Icn(0) exists and is finite for ry . Mogulskii's theorem which gives path asymptotics.. (iii) #c (8) = limn.e < 8 < y + e. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .. and the WentzellFreidlin theory of slow Markov walks. . commonly denoted as is the saddlepoint approximation. is differentiable at ry with 0 < K'(y) < 00. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. The substitution by V needs. Xn given by Fn(dxl. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo).'(u) )Ng a"u.. we introduce a change of measure for X1..'s. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 .o log Ee9Sn /n. .308 CHAPTER XI.. be a sequence of r. We shall need: Lemma 3 .2 (GLYNN & WHITT [163]) Let X1. Assume that there exists 'y.1). Sanov's theorem which give rare events asymptotics for empirical distributions. 1]. however. For the proof. Ee9X n < oo for e < 0 < e.
s. u Proof of Theorem 3..ne(p+ 17). P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).W.71 < e and jq9j < e. S. the r . is of order .77) follows by symmetry (note that the argument did not use µ > 0). > 1 +17] m(7).n m µ 1 + rl .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 +. ( U) P(S.2. it is easy to see that the r.. S. 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. h. Then V.n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . This establishes the first claim of the lemma .y) . h.077 n^oo n and by Taylor expansion and (iv ).> . The corresponding claim for Pn(Sn/n < µ .+r.µ?7 . We first show that lim inf„_. Clearly.Bµ .+r7) < zn for n > no. for Sn.. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0. log zl'(u)/u > 'y. we get lim sup 1 log Pn (Sn1 /n > µ + r7) < 0(1i + r7) + i(p(0 +'Y))/p n+oo n and by Taylor expansion.s.2.r (7) n = e.m(7). Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1.. 0. This proves the existence of z < 1 and no such that Pn (Sn/n > µ. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3.n > u ) = [ Em [em Em 1e. The rest of the argument is as before. mµ Sm > u] km e7Sm+n. in particular the r. can be chosen strictly negative by taking 9 small enough.]1/q = e.s.91) + o(O ) as 0 J.. Since I EeqOX „ ] 1/q is bounded for large n by (ii).3.YS.h.
. Pn \ > la+ 8 I < zn (3. and since Ic. Sn > u] < eYu+Kn(7)pn(Sn > u) (3. (iv) and Lemma 3.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E.3. For lim supu. P(T(u) = n) < P(Sn > u) = En [e7S.0 log i'(u )/u < 'y. 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).. I > IL exp `S.YS +^c CHAPTER XI. 0 yields liminfu __.310 ]Em I e. MISCELLANEOUS TOPICS (7).7) so that n(b) I1 < e'Yu E en.. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).n(ry)/u 4 0andm/u* (1 + r7)/µ.(•) goes to 1 by Lemma 3.. Obviously. 3. I2 = F(T(u) = n)..+wn(7).6) for some z < 1 and all n > n(E).log z) /2 and Sn Fn\ n >lb+S) <Zn.I < µl1 1+77 I M 1_ 1+277 S. logO(u)/u > ry.(Y). we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J.. this is possible by (iii).. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. n=1 .
S.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. Sn1 C U. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) . 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 . eryu en logz/2p n nt n. C 26u `p / +1 I e6u(1+6)/µ (3.11) [u(1+6)/µJ+1 1  Thus an upper bound for z/'(u) is n(6) e'Yu n=1 eKn (7) + 2 + (28U + 1) e6u(1+6)/µ Fi 1 zl /2 and using (i). u . LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.10) 00 I4 < E F(Sn_1 < u. 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 .3.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3.
> u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r.b)/i(7). this is straightforward since the last inequality in (3.('+'Y). the typical time is u/rc'(7) just as for the compound Poisson model.7' a"ju..11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . e'. the last steps of (3. we replace the bound P(Sn > u ) < 1 used in (3.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . For 14.z 1/z For I1. For 12. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7). u(1 + b)/i(7)) Proof Since V. Then for n large. say n n1. I2.2. ryue«iu . cf. (7 + a) < 2arc'(7). For I.3ui where . we need to redefine n(b) as L.. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. u . 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. it suffices to show that for j = 1. 2. Letting c11 = maxn<n.(u) = I1+I2+I3+I4'^ ery( u). Corollary 3. IV. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .xl/2 to give the desired conclusion.4 Under the assumptions of Theorem 3.4/3rc'(y) > 0.4. 4 there is an aj > 0 and a cj < oo such that Ij < c3e. 13 = P(T (u) E (u(1 b)l^ (7).Q is so small that w = 1 .312 CHAPTER XI.8) by P(S.u(1+b)/rc'(7)).
Assuming that the further regularity conditions can be verified..g. r. Thus the total reward in the interval [0.13) One would expect this to hold in considerable generality. criteria are given in Duffield & O'Connell [124]..12) k=0. To verify these in concrete examples may well present considerable difficulties. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.LARGE DEVIATIONS 313 Example 3 .1. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * . 11 Inspection of the proof of Theorem 3.3.. Obviously many of the most interesting examples have a continuous time scale.2 shows that the discrete time structure is used in an essential way..1. whether P ( sup St > u ltg a ^" 0<t<oo // (3. the key condition similar to (iii). (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0. An event occuring at time s is rewarded by a r.. Xk+l) k=1 00 naoo n provided the sum converges absolutely..e. Theorem 3.'(y) > 0. V(s) with m. 2 is in force with y = 2p/wz. 09(9).2 then immediately yields the estimate log F( sup Skh > u) a7u (3. and in fact. and we conclude that Theorem 3 ..3(s) at time s.14) is needed in both examples .v. 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.(O) = 9µ+02 for all 9 E R. i. If {St}t> 0 is the claims surplus process. we shall give two continuous time examples and tacitly assume that this can be done.. but nevertheless.f. t] is Rt = E V (Un) n: o„ <t . The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .
Un represents the total payment for the nth claim).9t = /3 J t (Ee8U° i8l . 0 Example 3 . MISCELLANEOUS TOPICS are the event times. n: o.It. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot .314 where the an CHAPTER XI. Most obviously. leading to St = At(1+77) Joo t S8 ds. Thus by (3. (3.14) (to see this .0 and assume there are y. if the nth claim arrives at time a. 7 Given the safety loading 77. An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . this is not realistic .noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . e..14). Then logEeOR° = J0 /3(s)(^8(9) . (9) < oo for 9 < 'y + C. the above discussion of discrete skeletons). one would take p(t) = (1 + rt)At/ t.v. is At . then the payments from the company in [on.15) .2 are trivial to verify. it contributes to St by the amount Un(t . At = . i.t. nondecreasing and with finite limits Un as s T oo ( thus. 0 and since EeOUn(8) + Ee°U^ as s * oo. the Cramer.9t.1) . We further assume that the processes {U1(s)}8>0 are i. where Ft = a(A8 : 0 < s < t). Of course . Kt (0) t (Ee9U"it8i J0 . Since the remaining conditions of Theorem 3.1) ds rt (3. = U„ ( t ..g. we conclude that Cu) log e7 u (cf. Of course. Thus. but that a claim is not settled immediately.. We let ic (9) = 3(EeWU° . the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. derive .d. we have S. assuming a continuous premium inflow at unit rate.. Example 3. More precisely. Thus.1) ds . we have rct (9)/t 4 ic (9).s).Lundberg model has the larger ruin probability. e > 0 such that ic('y) = 0 and that r.1) ds . a differential equation in t). If the nth claim arrives at time Qn = s. <t which is a shotnoise process. Un(s). . the best estimator of /3µB based upon Ft.Q„) .'`1 U. O'n +S] is a r .
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 .20) (3. Ui Nt / t 01i 315 St = Ui . rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 . we conclude that t. and since the remaining conditions are trivial to verify.b(u) IN a'Yu (cf.(1 + 17)0µB = 0. 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. To see this ./3.d. one has y > y' (3.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.1) .i. standard exponential .log Oi are i. (3.e.16) i=1 o i=1 Let ict (a) = log Eeast .3. 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 . It then follows from (3. uniform (0. (3.18) Thus (iii) of Theorem 3. Indeed.i.19) with equality if and only if U is degenerate.(1 +i) f > i= 1 s ds = E Ui 1 .(1 + r7) log t (3. we have Nt t N. equivalently.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.2 hold.1) or . typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model . Thus. the solution of /3(Eelu .d. LARGE DEVIATIONS With the Qi the arrival times. i. the Vi = .
rc*' (0 ) < 0. and since tc(s). Further. Dembo & Zeitouni [105] and Shwartz & Weiss [339].xo. 0 < x < x0. using that Ek(U) = 0 because of (3.20) is due to Tatyana Turova. . = P(N = n) = e(3an However. [245]. the study is motivated from the formulas in IV. see Nyrhinen [275] and Asmussen [25]. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. at time t. assuming that the U. Lehtonen & Nyrhinen [244]. say one year. Further applications of large deviations idea in risk theory occur in Djehiche [122]. are i.. see also Nyrhinen [275] for Theorem 3.1 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. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. 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. much of the analysis carries over to more general cases. In particular. [257] and Nyrhinen [275]. with common distribution B and independent of Nt. the proof of (3.d. the function k(x) = e7*x .7. and k(x) < 0. we then take t = 1 so that p. though we do not always spell this out. This implies n(y*) < 0. Therefore e7'U _ k(U) E [1+(1+77)y*U] . k(0) = 0. this in turn yields y > y*.2. For notational simplicity. x > x0. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0.1 . MISCELLANEOUS TOPICS Next. k'(0) < 0. a* (s) are convex with tc'(0) < 0 . In addition to Glynn & Whitt [163]. MartinL6f [256]. Further. y = y* can only occur if U . we are interested in estimating P(A > x) for large x.(1 + ri)y*x is convex with k(oo) = 00.i. The main example is Nt being Poisson with rate fit. For Example 3.2 expressing the finite horizon ruin probabilities in terms of the distribution of A.19).316 CHAPTER XI.
x)//3B"[9] is standard normal. Hence P(A > x) = E e [e9A+ ic(9). Vare(A) = s.2) implies that the limiting Pedistribution of (A . This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution.1 Assume that lim8T8.1). A E dx] . The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). 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 .e.4. only with 0 replaced by a9 and B by B9. i. Then Ee"A = e'(") where x(a) _ 0(B[a] . Proposition 4. B"[s] = oo.3e(bo[a] ."(0) = . K'(0) _ ic'(9) = x.3B"[9]. For a given x.9(Ax). THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. Then as x * oo.1). A > x)] = eex+K( e)E9 [e . 818' where s' = sup{s : B[s] < oo}. The analysis largely follows Example 3.1.ic(9) = . we define the saddlepoint 9 = 9(x) by EBA = x.1) where )30 = . no(a) = logE9e'A = rc(a + 9) . e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. In particular.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx).[s])3/2 = 0. B"' [s] lim (B". (4.
Q{AB (4. For example. B covers distributions with finite support or with a density not too far from ax° with a > 1. Thus . Remark 4 .2i and that (A . 2). Furthermore 00 b(x)Sdx < oo for some ( E (1.e. In fact.v. Notes and references Proposition 4. b is gammalike. For details. leading to P(A > x) :. For example.1).x') where x' = sup {x : b(x) > 0}.4) . large x.(D X .EN B(x). it is quite questionable to use (4. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. In particular. then P(A > x) . b is logconcave. Jensen [215] and references therein. bounded with b(x) . 4b The NP approximation In many cases . the distribution of A is approximately normal . see Embrechts et al.1 yields: Proposition 4. or. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo.3) and related results u for the case of main interest .2) is often referred to as the Esscher approximation. Var(A) _ ^3p. and (4. just the same dominated convergence argument as in the proof of Theorem 2. 1 . b(x) = q(x)eh(z).318 CHAPTER XI.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion.2 If B is subexponential and EzN < oo for some z > 1. it holds that EA = . For a rigorous proof. The present proof is somewhat heuristical in the CLT steps.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. A covers the exponential distribution and phasetype distributions. under the Poisson assumption (4. some regularity of the density b(x) of B is required. either of the following is sufficient: A. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r.ycix °ie6x B. [138]. i.1 goes all the way back to Esscher [141]. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4.l3pB. more generally. 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 .
y2)^P(y)• 319 Note as a further warning that the r. the CLT for Y = Y6 is usually derived via expanding the ch. . Heuristically.2 ^ \1 .i 6 r 1 3 so that we should take b = ic3/6 in (4.2X2 .i 3 K3 } Pt^ exp . so that 1(u) 3 exp { .3& (y). then P(Y < y) 4(y) .. of (4. as u2 u3 u4 9(u) = Ee'uY = exp {iuci .h. u5.6) .. In concrete examples . resp..5) follows by integration.f. the standard normal distribution. Rather than with the tail probabilities F(A > x). THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. the NP (normal power) approximation deals with the quantile al_E. one expects the u3 term to dominate the terms of order u4.3!).1).l = EY.c2i.EY)3. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .. Let Y = (A . ..2K3 + 4i 64 + . Var(Y) = 1 as above . s. defined as the the solution of P(A < yle) = 1 .5) may be negative and is not necessarily an increasing function of y for jyj large. If this holds . Thus if EY = 0. are the cumulants . are small.. zl_e be the 1 . If the distribution of Y is close to N(0.. where Kl .EA)/ Var(A) and let yl_E.5 (y3 .6(1 . Remark 4.2 2 . one needs to show that 163.5). and from this (4. A particular case is a. however.: EA + zl_E Var(A) .5) is obtained by noting that by Fourier inversion.4.99. (4. K4 .. which is often denoted VaR (the Value at Risk). K2 = Var (Y). and so as a first approximation we obtain a1_E = EA + yle Var(A) . . the density of Y is 1 °° _ eiuy f(u) du 2x _. (4.. ylE should be close to zl_E (cf.s.equantile in the distribution of Y. K3 = E(Y .e. in particular.
1).zl E)^o(zl E) .3ni /3 . this holds with a = 0.6 (1 .1)! n ^eQ .EA ) / Var(A).E )Azl E) 4(z1E) + ( ylE . let pn Pn = (a+ = P(N = n).zi. n = 1.E(/3PB^1 )1^2 + s(z1E . that [101] distinguishes between the NP and Edgeworth approximations.. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .1) E (A .6pBki) d/2.(y) terms dominate the S(1 . We can rewrite (4.S(1 .zlE)W(zlE) 1 .zl E)V(zl_E) .. 21 .1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. however. . Using Y = (A .5(1 .zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 .E .y2)cp( y) term. b such that EN 1 U%.yi. For example. MISCELLANEOUS TOPICS A correction term may be computed from (4..E = z1E + S(zi_E . [101].. k3 is small for large /3 but dominates 1c4. the kth cumulant of A is /3PBk' and so s.5) by noting that the 4. and assume that there exist n ) Pn_i .1)EY3. this yields the NP approximation 6(Z1 _E . Another main reference is Daykin et at.E + (yl. This leads to t( yl E) .E)A1 l E) 1 E 4)(yl E) ^' .. Note.7) as 1 (3) a1E = Qµa +z1 . In particular . K5 .320 CHAPTER XI.EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.k = /3µB^1 / (. as required .S(1 ... 4c Panjer 's recursion Consider A = constants a.
and calculating the gj*n recursively by 9*1 = 9j.11) Remark 4.1. Since the sum over i is na + b. . Hence by (4. . E[a +bU=I >Ui =j l i=1 J (4... (4. fj = E (a+ b k =1 )9kfi_k .12) where g*n is the nth convolution power of g. . 2..13) but only O(j2) for Proposition 4. u Proof of Proposition 4.13) Namely. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4. 2.. (4. then j (a + b!) 1ag k_1 3 gkfj. Then fo = >20 9onpn and fi = 1 E In particular. . j = 0.. . j1 g.4. if go = 0. the complexity (number of arithmetic operations required) is O(j3) for (4.4. (4.. .. By symmetry. fj = P(A = j). j = 1.4 is that the algorithm is much faster than the naive method. ...14) is independent of i = 1. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4.. n = k=n1 9k(n1 )9j k • (4.. .10) f o = po.5 The crux of Proposition 4..4. n. j = 1. the value of (4.14) is therefore a + b/n.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 . The expression for fo is obvious. 2.} and write gj = 2 .9).12).k . (4. 1.4 Assume that B is concentrated on {0. .
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).
4.h * (X) .4).4) and u g(x) = e7x in Ohlin's lemma. Assuming for simplicity that the Ui are i. Then for any local retention function u . and so on. it suffices to show that Eexp {ry ii 'UiAb.6 Assume the Ui are i. ry* > 0 because of (6.d.h(U) (as in the proof of Proposition 6.d. Remark 6.P.5) reduce quite a lot. The arrangement used in practice is. then (6.h(Ui)] . (6.h(U)]. This follows by taking Xl = U A b. Proof As in the proof of Proposition 6. expectations like those in (6. N E X .b)+ with b determined by E(U .h(Ui)P JJJ l:='l {ry ] or.6) u where C[ry] = Ee'r(u4(u)).P > EexP{7[X .6).5 Because of the independence assumptions . y = Ei [Ui . Eexp 7 [E [Ui . X2 = U .4).4). . Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy. however. (6.b)+ = Eh(U) (and the same P) satisfies 71 > ry. we get EX = EN • EU.i. Local reinsurance with h(u) = (u . ' ii (6.h(Ui)] . u But since ry > 0. i.P } < 1 = Eexp E[Ui. appealing to (6.h(Ui)] .b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. this implies 7* > 7. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .5) holds trivially.6..P I = EC [7]N.3). we get N 1 = Eexp ry E[Ui ii . the excess ofloss rule hl (u) = (u .h(u) and any P satisfying (6. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .P]}.h( UU) = EN • E[U . as often local as global.
[76]. Bowers et at.330 CHAPTER XI.many texts on insurance mathematics. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. See further Hesselager [194] and Dickson & Waters [120].g. The original reference for Ohlin's lemma is Ohlin [277]. e. The present proof is from van Dawen [99]. see also Sundt [354]. . Heilman [191] and Sundt [354].
= T„ . not concentrated on {h. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. The renewal theorem asserts that U(dt) is close to dt/µ. all have the same distribution. of interarrival times and the time Yo = To of the first arrival (that is.e.T„_1). stating that U(t+a)U (t) ^ a. of epochs or the set Y1. Lebesgue measure for some n > 1). t +a])..1) (here U(t) = U([0. are independent and Y1. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 .r. Y2. Y1.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. Lebesgue measure dt normalized by the mean to of F.... Then Blackwell 's renewal theorem holds. The point process is called a renewal process if Yo. t] is denoted by Nt. That is.t. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. The mathematical representation is either the ordered set 0 < To < T1 < . If Yo = 0.U(t) is the expected number of renewals in (t. Technically. i. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w .. some condition is needed: that F is nonlattice. . Y2. the distribution of Yo is called the delay distribution. denoted by F in the following and referred to as the interarrival distribution. Y. note in particular that U({0}) = 1. 2h.. . . the renewal process is called zerodelayed.....} for any h > 0. t]) so that U(t + a) . . when t is large.. t 00 (A. then Stone 's decomposition holds : U = U. The number max k : Tk_j < t of renewals in [0.
2 Assume that Z solves the renewal equation (A. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. that z(u) has a limit z(oo) (say) as u 4 oo. IV).i.e. 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 . µF (A. (A. u u PF 4 00.4) that z is Lebesgue integrable with limZ.2). wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. then Z(u) i f0 z(x)dx .3) Further. and F(dx) a known probability measure . in convolution notation Z = z + F * Z. and that F has a bounded density2. resp. i.332 APPENDIX u(t) has limit 1/µ as t 4 oo. z(u) a known function.x)F(dx). but suffices for the present purposes .2) Z(u) = J0 u z(x)U(dx).9. Under weak regularity conditions (see [APQJ Ch. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. see [APQ] Ch. stating that U(t)/t > 1/p.5) 2This condition can be weakened considerably .EN(t) . z(x) = 0..out.i". In 111. oo). (A. Note in particular that F is spreadout if F has a density f. IV). Then Z(u) 4 z(oo).4) If F is spread.R. Equivalently. (A. then it suffices for (A. the statements being EN(t + a) .a.2) has the unique solution Z = U * z. Both result are valid for delayed renewal processes.1 if F is nonlattice and z (u) is directly Riemann integrable (d. (A. U Z(u .
however. that the existence of y may fail for heavytailed F. Y2.x)u(x) dx = z(u( 1 . Z(u) U = 1 u 1 u f z(u . the postTk process {XT. F(dx) = e7xF(dx). cycles. asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. • . i. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. To this end. ... this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. A regenerative process converges in distribution under very mild conditions: . refer to the zerodelayed case. The simplest case is when {Xt} has i. Y1 . the present more general definition is needed to deal with say Harris recurrent Markov chains.(3. Eo etc. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles).t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). . However. . Tk (or. or many queueing processes. Yk ). A stochastic process {Xt}t>0 with a general state space E is called regenerative w. 1c Regenerative processes Let {T.... However. that F is a probability measure.3) satisfied by the ruin probability for the compound Poisson model. Tk and {Xt }o<t<Tk • For example. . 0 PF µF 11 In risk theory. .t. {Tn} if for any k. Hence by dominated convergence. of Yo. Here the relevant F does not have mass one (F is defective).5a. This program has been carried out in III. z(x) = e7xz(x). and its distribution does not depend on k. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z.e.r.. The kth cycle is defined as {XTk+t}o<t<Yk .i.. is called the cycle length distribution and as before. We let FO. equivalently.d..} be a renewal process. we let µ denote its mean. T1. multiply (A..2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). Note. T1.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. . Assuming that y can be chosen such that f °° Ox F(dx) = 1.k+t }t>o is independent of To. a basic reason that renewal theory is relevant is the renewal equation II. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. The distribution F of Y1.
cycles (we allow a different distribution of the first cycle).6) id Cumulative processes Let {Tn} be a renewal process with i.. (A..+ X. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. i.ZTOI < 00. This is the case considered in [APQ] V. {Tn}. We denote the limiting r. {i7(t)} are Markov with state spaces (0.. in total variation. assume that p < 00 and define Un = ZT}1 .e. [0. Then {e(t)}.. An example is Zt = fo f (X8) ds where {Xt} is regenerative w.3. Then {Zt}t^. C).r.d. {Tn}. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p.t : t < Tk}. for n = 1.'s by e. and we have: holds more generally that (rl(t). e(t )) . are i... µ 0 If F is spreadout. under the condition of Blackwell's renewal theorem.r.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1.. 0<t<Yi then Zt /t a$• EU1/µ.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. then (Zt . resp ..e.ZT }0<t<Y„+.0 be cumulative w.oo (i. where the distribution of X.v.334 APPENDIX Proposition A1. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo.. then Xt .i.4 Let {Zt}t^. oo). Then it (ii. just the same proof as there carries over to show: Proposition A1.0 is called cumulative w.. but in fact. fi (t) = inf {Tk . (b) If in addition Var(Ul ) < oo.i. Otherwise . then e (t) . r. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t. 2.d..t..ZT Then: (a) If E sup I ZTo+t .r.t. and q(t) = sup It . oo). {Tn} if the processes {ZT +t . Then Xt Di X.Tk : t < Tk} as the age.t. If p = oo.
(c) the marginal distribution of q is FO. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y). In the general case. if in addition EYo < oo.(t). 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . Yo > 0] + f Eo^ (t . and the conditional distribution of ri given l. Y1 > t] 4 0. the first statement follows.U(x) < U( 1)). (b) the joint distribution of (ri. For the second. ^) is given by the following four equivalent statements: (a) P (77 > x. assume first the renewal process is zerodelayed. we used: Proposition A1. Hence for t large enough. (d) the marginal distribution of ^ is FO.v.APPENDIX 335 Theorem A1. the joint distribution of (rl. l:) is the same as the distribution of (VW.d. Then fi(t)/t a4' 0 and.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). and the equivalence of (a) with (b)(d) is an easy exercise. V is uniform on (0. .d.6 Consider a renewal process with µ < oo.U(x) (c < oo because it is easily seen that U(x + 1) .y)P(Yo E dy) . W are independent. Proof The number Nt of renewal before t satisfies Nt/t a4' p. ^ > y) = 1 f +Y (z)dz. In IV. 1) and W has distribution Fw given by dFw/dF(x) = x/pF.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. Hence t t lt ) = f U(dy)z(t .5 Under the condition of Blackwell's renewal theorem.y) = f U(t . Y1i Y2. = z is Foz) The proof of (a) is straightforward by viewing {(r.. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.^(t))} as a regenerative process. Since the maximum Mn of n i.i. use t E^(t)/t = E[Yo . r.i.. (1 V)W) where V. Yl > t]. the sum is o(t) so that Eo£(t)/t + 0 .t. Since z ( k) < E[Yi . but governed by a Markov chain {Jn} (we . U(x + 1) . EC(t)/t + 0. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 ..4.t. are not i.
e.. A stochastic process {Xt}t>o is called semiregenerative w.. Notes and references Renewal theory and regenerative processes are treated. . oo)..} is nonlattice (it is easily seen that this definition does not depend on i)..g. A Markov renewal process {Tn} contains an imbedded renewal process.. in [APQ]. Jn_1. X2.7 Consider a nonlattice semiregenerative process. . G_(x) = P(ST_ < x. with common distribution F. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. . Yn. .+ < x. the semiregenerative process is called nonlattice if {T.r. Sn = X1 + • • • + Xn the associated random walk.T_ < oo).t.i . A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y.. Then Xt 4 Xo. distribution ofjXt}t>o itself where Pi refers to the case Jo = i.. Y1. Jo. G+(x) = P(S. .}. namely {Twk } where {Wk } is the sequence of instants w where Jo.) and (Fij )i. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. . where the distribution of X. the conditional distribution of {XT„+t}t>o given Yo.r. IT. = io for some arbitrary but fixed reference state io E E.. These facts allow many definitions and results to be reduced to ordinary renewal. T_=inf{n>0: Sn<0}. J1 i ..t.. r+ < oo). . . Jn = i is the same as the P. be i. oo).. Further: Proposition A1. Let X1. Jn +1=j} where J = a(JO.d.and regenerative processes. 0] or (0 . The semiregenerative process is then regenerative w.. the Markov renewal process if for any n. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE. and define r+=inf{n>0: Sn>0}.. < yIJ) = Fij( y) on {Jn= i. Alsmeyer [5] and Thorisson [372].. .jEE is a family of distributions on (0. For example. ..
r.=n w=m i Figure A. 0] and (0. m<j<n}.8) (e.x)R_ (dx). we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A). F(A) + (G+ * G_)(A).7) follows since G+(A) = 0 when A C (oo. In (A. Proof Considering the restrictions of measures to (oc. n=0 The basic identities are the following: Theorem A2. A C (0. . A C (0. 0].7).T_=n} = {S.S.7) (A. G+. Sr_ _1 is at its minimum . n=0 n=0 00 00 and the T+. >0. we consider the last such time (to make w unique) so that {w=m. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. 0]).>0. More rigorously. define w as the time where the preT_ path S1.g. G_. 0)..x)R+(dx). oo) (A..APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. U. oo).. F(A .S. n 0 R_(A) = E I(Sn E A).1 . (A. On {T_ > 2}. . 0<j<m. the renewal measures U+=>G+. (d) R+ = U_.=EGn.G+ * G_: (b) G_ (A) = f °° F(A .1 (a) F = G+ + G_ . oo). A C (oo. A C (oo. u . (e) R_ = U+. S. (c) G+(A) = f °.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A).
It follows that for n > 2 F (7.m. m=1 f S mming over n = 2. and the proof of (A. Sn1 E dx) n=1  F(A . 0<j<m.1... ST_ E A .. . A.1) that P(Sj Sn..3 8 APPENDIX Reversing the time points 0.3.= n.7) follows.+ E du)P(S.du) (G+ * G)(A)• C llecting terms. S. . A. and reversing the order of summation yields P(T_ > 2. SmEdu) = P(T+=m. E du) = P(T_=nm.XnEAx) 00 f 0 f 0 00 00 1: F(A . ST+Edu)._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu)..0<k<ri . m it follows (see Fig. SnEAIS..8) is similar.x)R+(dx). 0 < k < n.x)P(Sk < 0._ = n .. S.Sn_1Edx.1).F(r_n_mSrEA_u). Aso. r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. . clearly (Sj Sm>0. ST_ E A) P(T+ = m.. m < j <n.u) f0m m=1 n=m+1 00 J0 OO P(S. Sr_ E Adu) (s ee again Fig . (A.>0.+ E du) E P(S._ E A . (b) follows from 00 G+ (A) _ E F(Sn E A.
1. Sk = X1 + • • • + Xk = Sn .2 In terms of m. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). there are direct analogues of Theorem A2. Again.0<k<n. However. and sometimes in a larger strip. WienerHopf theory is only used at a few places in this book. u Notes and references In its above discrete time version.0+[s])(1 .F[s] = (1 . Since G+ is concentrated on (0. In continuous time.g. there is no direct analogue of Theorem A2. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). and the proof of (e) is similar. it serves as model and motivation for a number of results and arguments in continuous time.g.f.9) whenever F[s]. For example.O<k<n..SnEA) = P(Sn<Sk. For (d). Then for A C (oo.1).1(a) is from Kennedy [228]. see for example Bingham [65].P as a product H+H_ of functions with such properties.APPENDIX 339 and the proof of (c) is similar.g. Nevertheless. G_ [s] are defined at the same time.G_[s]) (A. such developments motivate the approach in Chapter VI on the Markovian environment model. we can rewrite (a) as 1 . Another main extension of the theory deals with Markov dependence.1. this holds always on the line its = 0. Summing over n yields R+ (A) = U_ (A).0<k<n. In discrete time.Sn_k. u Remark A2. being concentrated at 0. E.s. G_ are trivial.O<k<n. then T+ = inf It > 0 : St = 0} is 0 a. cf. In this generality of. the analogue of a random walk is a process with stationary independent increments (a Levy process. is based upon representing G+ as in (b).SnEA) is the probability that n is a weak descending ladder point with Sn E A. .G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. The present proof of Theorem A2. see e.SnEA) = P(Sn<Sk. The classical analytical form of the WienerHopf problem is to write 1 .SnEA) = P(SnSn_ k.6. which is basic for the PollaczeckKhinchine formula. and similarly H_ (s) = 1 . and G+.'s. the derivation of the form of G+ for the compound Poisson model (Theorem 11. a number of related identities can be derived.T+> n) = P(Sk < O.. 11. 0]. P(SnEA . oo). if {St} is Brownian motion.4). consider a fixed n and let Xk = Xn_k+l. the survey [15] by the author and the extensive list of references there. 6+ [s].
_I 0 (A. Eo Kn/n! converges rapidly and can be evaluated without p oblems. Thus. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). Here are. 1.340 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.12) eA'AO = Ale AA (A. 0 . if m is s fficiently large.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.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. It is seen from Theorem VIII.5 that when handling phase type distributi ons. however . 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 . and eQ can then be computed as the mth power (by squaring if = 2). To circumvent this. three of the c rrently most widely used ones: xample A3.11) A f eAtdt = eA. whereas there is no similar single established a proach in the case of matrix exponentials. ere A is the eigenvalue of largest absolute value. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). 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).10) d dteAt = AeAt = eAtA (A.
3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt. One then can reduce to p linear differential equations by noting that k = ZQ. .7t) n=0 n! u °O n Pn (to see this. The probabilistic reason that (A. the procedure consists in choosing some suitable i > 0. However . In practice. 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 . The approach is in particular convenient if one wants eQt for many different u values of t. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1.e. Zo = a (Z = QZ.4 (DIAGONALIZATION) Assume that Q has diagonal form. the intensity matrix Q is the same as the one Q for {Xt} since a jump from i to j 11 i occurs at rate qij = 77pij = q22. Here is a further method which appears quite appealing at a first sight: Example A3 . 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. p different eigenvalues Aj i .. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77...APPENDIX 341 Example A3. . i..]t)n (A. vp be the corresponding left . To this end. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. Let vi.3 i (A.2 (UNIFORMIZATION) Formally. some jumps are dummy in the sense that no state transition occurs ). condition upon the number n of Poisson events in [Olt])  Example A3. i. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(.e. Ap.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (.15) Then it is easily checked that P is a transition matrix ..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}. Zo = h).
18) Namely.18) contains terms which almost cancel and the loss of digits may be disasterous. we can take H as the matrix with columns hl. some cases remain where diagonalization may still be appealing.. not all ai are real. and hence A2 is so because of A2 = tr(Q). of largest real part is often real (say.. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421... and writing eQt as eQt = He°tH1 = H (e\it)di.342 APPENDIX (row) eigenvectors and hl. hp the corresponding right (column) eigenvectors. the eigenvalue. hp. Complex calculus : Typically.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. and vihi ¢ 0. we have an explicit formula for eQt once the A j.. under the conditions of the PerronFrobenius theorem). There are. i=1 i=1 Thus. The phenomenon occurs not least when the dimension p is large. D = ) 2 2 . and we may adapt some normalization convention ensuring vihi = 1. i # j.. say Al. say A = (Ai)diag. (A. i= 1 i=1 P P (A. hi have been computed. Nevertheless. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close. Then P P Q = > Aihivi = E Aihi (9 vi. vi. Then vihj = 0. however.16) (A. v5Q = Aivi. Qhi = vihi. Example A3.17) eQt = E e\`thivi = E ea:thi ® vi.g H1. (A.. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix.. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials.
eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A.e.7 Let 3 9 2 14 7 11 2 2 . k  C k2 ) =b ( A1 q 1 Q11 / where a .6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. However.20) ir = q2 ql qi +q 2 9l +q2 (A. replacing ai by A2. i. where (A. b are any constants ensuring//Irk = 1. Of course. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 .Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF.q.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. The other eigenvalue is A = A2 = q1 . Then 7r = (ir1 7r2 ) = a (q21 Al . v2 and h2 can be computed in just the same way. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e.k1). 1) .APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector. u Example A3. l ab (g12g21 + (A1  411) 2) = 1.19) Example A3 . h2 = Thus.
2 2 1=ab(142+(1+2)2 ) = tab..satisfying AAA = A. They are most often constructed by imposing some additional properties .23) . (A+A)' = A+A. Generalized inverses play an important role in statistics. 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_.11/2 + 5 1. for example AA+A = A. but only that dimensions match . e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. and a generalized inverse may not unique. A2 = 3/2 . A+AA+ = A+.5 . APPENDIX x1 3/2 .11/2 .344 Then D= 2+ 11)' 7 T4 2 =52. (A. (AA+)' = AA+.6..22) Note that in this generality it is not assumed that A is necessarily square. (A.
most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix..= (I .e. Here is a typical result on the role of such matrices in applied probability: Proposition A4. (I . = 0 where m < p is the rank of A. lt o eAx dx = te7r + D(eAt .P + e7r ). then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . Rather than with generalized inverses .25) . one is also faced with singular matrices ..1 Let A be an irreducible intensity matrix with stationary row vector it. and define D = (A .e ® 7r)1. one then works with Q = (Q . Am+1 = . . 0 01 In applied probability. Then for some b > 0.24) = te7r . and exists and is unique (see for example Rao [300]).APPENDIX 345 A matrix A+ satisfying (A.g.eir )1. _ A.23) is called the MoorePenrose inverse of A.P).1 goes under the name fundamental matrix of the Markov chain).1Q = Q(Q .. if A is a possibly singular covariance matrix (nonnegative definite)... ( Q .P + e7r)1 (here ( I . E. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .I) (A. .eir )..D + O(ebt). are ordered such that Al > 0.g. Assume that a unique stationary distribution w exists . These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . Am > 0. (A.ew.eir)1 = I .
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.91a(2) . Then A(O) _ B(O) = 0. of (A.s.s.26) 2 = 2 e7r + tD .I)} dx. I.I) (A.e.J {xe^r + D(e . Note that h ® it has rank 1.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). the rows are proportional to it. Equivalently. B(t) denote the l. resp.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx . .D + D2 + O(ebt).24). ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4.2 Let it be a row vector with m components and h a column vector with k components. For example.DZ(ent . in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. h as 1 x m and k x 1 matrices.. and the columns to h. and in fact any rank 1 matrix can be written on this form. . see below. o Finally.I) .346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . the formulas involving O(e6t) follow by PerronFrobenius theory.h. h ® it reduces to hit in standard matrix notation. B'(t) = e7r + DAeAt = eir + (I .h. (A. Interpreting 7r. it follows that h ® it is the k x m matrix with ijth element hi7rj . respectively. 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. (A.27) Proof Let A(t).eir)eAt = eAt = A'(t).I)}. the r.2e7r .
(A. (A. each of which is A ® I or I ® B.50 6 7 6 4f 4. and the number of such factors is precisely given by the relevant binomial coefficient. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I .28) In particular.4 eA® B = eA ®eB.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2).k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.29).3v'6. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) .5v/.4vf. 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. if A ® I occurs k times.29) If A and B are both square (k1 = ml and k2 = m2).3f 4v/.3vV/72f 20. (A B)' = eA®B e! L 1=0 0 . C2 = h2 are column vectors. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2). then v1B1h1 and v2B2h2 are real numbers.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.A9.3V8.APPENDIX 347 Then A®B = 2 f 20. A2 = v2 are row vectors and C1 = h1. if Al = vi. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.31) Indeed. Using (A.(A.31). such a factor is Ak (&B 1k according to (A.
where transition matrix of the bivariate Markov chain {X n1). A special case of Proposition A4. independent Markov chains. v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. P(2). the same time.I)(h ® k). (A. we have P8 = Pal) ® p(2). resp .32). and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. h.4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . p = P(1) ® {X }. {Yt(1). in the definition (A. Let P8f P(Sl). P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity . { 1't(1) }. P(t) Yt(2) }.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. k any column vectors.s.6 Suppose that A and of B. Yt(2 ) }. X ) }. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1).3 < 0 Lemma A4 .33) . 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 +. { On the other hand.5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. Let further it. first term on the r . the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . From what has been said about matrices of {Yt( 1). and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . Thus . Q(2).348 APPENDIX Remark A4. represents ces Q( 1). Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba .
il. p there should exist io. . ao).7 Let A be a p x pmatrix with nonnegative elements. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). f o r each i..g... (A. . 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements.34) Note that for a transition matrix. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. . then IN < Ao for all A E sp(A). then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. . i. see e. .29). h can be chosen with strictly positive elements. h such that vh = 1. and the corresponding left and right eigenvectors v. Here is the PerronFrobenius theorem. in such that io = i. j = 1.. and if we normalize v. h = e and v = 7r (the stationary row vector). E (0. and appeal to (A.. h can be chosen with 3By this.. (b) if in addition A is aperiodic. .3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. > 0 for k = 1. which can be found in a great number of books. That is.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. = j and atk_li.. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. Now note that the eigenvalues of A ® B are of the form a +. so that by asssumption A ® B is u invertible. n. . 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.12). [APQ] X. we have AO = 1. and the corresponding left and right eigenvectors v. Similarly.APPENDIX 349 Proof According to (A.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements.
Corollary A4. The next result gives a condition for asymptotical exponentiality. For example.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ .1. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo.(ti)ding. if we normalize v.. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential.e. Proposition A5. the condition is that t is small compared to Q. we have A0 = 0. relate the eigenvalues of B to those of B via (A. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength.35) for some p E (oo. Then for any (3.8 is most often not stated explicitly in textbooks. 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). Note that for an intensity matrix. To this end.(3. the phasetype distribution B(a) with representation (. I. h = e and v = 7r (the stationary row vector). but is an easy consequence of the PerronFrobenius theorem. Ao). note that we can write the phase generator T as Q . 10) and use the formula me at e Bt = e 00 Antn = e . the analogy of this procedure with unformization. h such that vh = 1.8.n t AL n=0 n! (cf. Example A3. Furthermore. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. not only in the tail but in the whole distribution. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII.350 APPENDIX strictly positive elements.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). let {Yti°i } be a Markov process with initial distribution a and intensity .2).
We can assume that Jta) = Yt(°). and that Yt(a) = Yat for all t.x (1 . has a limit distribution: Proposition A5. J^O)_ = j) Pi (v(aaV) > x. and this easily yields a(x)/x a' 1/t*. it states that the state. {t Y( a) } v>0 .Yj(av) = j f . Then {Ix} is a Markov process with to = Yo. By the law of large numbers for Markov processes . Let further V be exponential with intensity V and independent of everything else. 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. We shall . prove a somewhat more general result which was used in the proof of Proposition VI. dx/ti] or not.a' + oo (e. Hence O ((a) aa. v/ t. in fact . Proof Assume first ti > 0 for all i and let I.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. In addition to the asymptotic exponentiality.1. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .bij) Hence the intensity matrix of { Ix} is (qij/ti)i.APPENDIX 351 ((1) etc. = YQ(x). 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}. Since JJ(. a' = a .aE where 0 < e < 1).g. a'/a + 1.YQ(av) = j) Pi ( ci(a'V) > x. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt.9. from which the phase process is terminated .jEE. Then a(a'V)/a (aV) a' 1.(a) > x .2 Pi (c(a) > x. a . Conditioning upon whether { Yt} changes state in [0. t < (a). and write Yt = Yt(1). fo tY dv/t a$' t*. J(()) _ = i) + at•x t tt' .
' pk 0 k>1 11 Theorem A5. (c) the nth moment k 1 k"bkis 1)"n!aP"p. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. Indeed.3 As the exponential distribution is the simplest continuous phasetype distribution. with point probabilities bk = (1 .+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. Example A5. so we shall be brief. k>1.g. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]). 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. the simplest discrete phasetype distribution: here E has only one element.zP)'p. However. k = 1..2 do not appear to be in the literature. let E and Pkj j=k1.Pe.. these results are in the spirit of rare events theory for regenerative processes (e.} is said to be discrete phasetype with representation (E.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) . Then P is substochastic and the vector of exit probabilities is p = e . P. . Example A5.p)k1 p.. A distribution B on {1. zkbk is za(I . 2. > 0}. 2. a). Penev & Turbin [238]. so is the geometric distribution. a = b = (bk)k=1.1 and A5.. See also Korolyuk... Keilson [223].5 Let B be discrete phasetype with representation (P.. u Notes and references Propositions A5. say bk = 0.. .x k > K. 1 k=1 1 0 otherwise. K}. Then: (a) The point probabilities are bk = aPklp.j) and initial distribution a. (b) the generating function b[z] _ E' .. . = 0 for one or more i. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. is discrete phasetype.. Et II I a(a^V) > x) at' . an easy modification of the argument yields finally the result for the case where t.4 Any discrete distribution B with finite support..
B2 be phasetype with representations (E(1).a(2).6 (CONVOLUTIONS) Let B1. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). Then the convolution B = B1 * B2 is phasetype with representation (E. _ i E E(1) T(1) t(1)a(2) i E E(2) ..6. A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A. and a=1)..T(1)).36) in blockpartitioned notation (where we could also write a as (a (1) 0)). { Jt 2) } with lifetimes U1 . T= ( 0 T(2) ) (A. resp.T(2)).a(1).2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. a. Then {Jt} has lifetime U1 + U2 .. resp. (E(2). r .1 This corresponds to a convolution of r geometric distributions with the same parameter p. The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r. Jt t > U1 + U2. and hence the negative binomial distribution is discrete phaseu type.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U. as is seen by minor modifications of Example A5.{ 0. 11 Example A5. initial distribution a and phase generator T.6 is the Erlang distribution Er which is the convolution of r exponential distributions. . U2.APPENDIX 353 5c Closure properties Example A5. A. a' .r + 1..
T) and C = EO°_1(1 . To obtain a phase process for C.0)a(2))). Then the mixture B = 9B1 + (1 .E) where a(°) = fAa(a)v(da). and consider B(") = fA B(a) v(da) where v is a probability measure on A.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E. a mixture of more than two phasetype distributions is seen to be phasetype.'). i E E(2) 0 T(2) =IT (in blockpartitioned notation.a(1).37) (1) (1 . i E E(1) T 0 I (A. (E(2).a(2).0)a(2) E(2) Figure A. p at each termination. B2 be phasetype with representations (E(1).p)pn1. are i. Equivalently. Example A5.T(1)).4 . this means that a = (Oa(1) (1 ...8 (FINITE MIXTURES) Let B1. Let B(") be the corresponding phasetype distribution. Thus. Example A5.d.354 APPENDIX Example A5. a. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).p.p)pn1B*n. P(N = n) = (1 . A reduced phase diagram is 0a(1) E(1) A .O)B2 (0 < 0 < 1) is phasetype with representation (E.T.i. In risk theory. a reduced phase diagram is f a E t Figure A. we need to restart the phase process for B w. with common distribution and N is independent of the Uk and geometrically distributed with parameter p.. resp. and o'i Oa.T(2)). Then it is trivial to see that B(") is u phasetype with representation (a("). U2. a. then C is the distribution of Ul + • • • + UN. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims.0)ai2).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. if U1.3 In exactly the same way.
let B be a continuous phasetype distribution with representation (F. then C is the distribution of U1 + • • • + UN. It is zeromodified phasetype with representation (E. (E(2). If we replace x by a r. of F.f. i. a.°_1 f„ B*?l. To obtain a phase representation for C . U2 be random variables with distributions B1.aF[T].T) if U is phasetype with representation (E. E)... T) and C = F..2. it follows by mixing (Example A5.2. P). cf.. a.. T).. 12 (PHASETYPE COMPOUNDS ) Let fl. resp.x)+. then U1 + • • + UN is zeromodified phasetype with representation (a.g. f2. For U1 A U2. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. then U1 +• is phasetype with representation (E. resp. U2.d. a(1) ® a(2 ). let {Jtl)}. are i.TWWW).. { 4 } as exit of {Jt}. +UN 2. Example A5 .T + pta). cf. a(2). say with distribution F.T) where F[T] = J0 "o eTx F(dx) u is the matrix m. v. if {Jt} is a phase process for U. j E F}. with common distribution B and N is independent of the Uk with P(N = n) = f. 13 (MINIMA AND MAXIMA ) Let U1.X)+ is zeromodified phasetype with representation (E. T(2) ).1. if B is defective and N + 1 is the first n with U„ = oo. Proposition VIII.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . T + ta.aeTx. B2 of phasetype with representations (E('). To see this. Example A5. let the phase space be E x F = {i j : i E E. T + pta).9) that (U . if U1. but the same T. then Jy has distribution aeTx. If U1 has a different initial vector. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). X independent of U. v. Example A5 . T(1) ® T(2)). . U2. { Jt2) } be independent with lifetimes U1. say v.APPENDIX 355 and C is phasetype with representation (E. . a. Minor modifications of the argument show that 1. Indeed.°. Equivalently. be the point probabilities of a discrete phasetype distribution with representation (E.a(1).7. Corollary VIII. Note that this was exactly the structure of the lifetime of a terminating renewal u process..v. Thus the representation is (E(1) x E(2).
. say degenerate at b.(Sn) with Sn = n/b.B(bk) I < 1/n for n > k. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). Then we must find phasetype distributions Bn with B.14 To a given distribution B on (0. cf.. Then from above.n = I:pi(n)Er v ( __ ) n) ) a= 1 . Hence it is immediate that Bn 4 B. Example A5. there is a sequence {B.356 APPENDIX For U1 V U2.... The general case now follows easily from this. elementary) Let {bk} be any dense sequence of continuity points for B(x). That is. see Neuts [269] (where the proof. Proof Assume first that B is a onepoint distribution.2) } to go on (on E(2)) when { i 1) } exits. 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 . the initial vector is (a(1) (& a (2) 0 0). we need to allow { Jt. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . and let Bn be the Erlang distribution E.(bk) + B(bk) for all k.(n) = D. any distribution B on (0. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5. relies more on matrix algebra than the probabilistic interpretation exploited here). i= 1 C. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. and vice versa.. however. oo).. with weight pi(n) for xi(n). By the diagonal argument (subsequent thinnings).(bk) + B(bk) for all k as n * oo. and the closedness of the class of phasetype distributions under the formation of finite mixtures.(bk)'. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. Let the support of Dn be {xl(n). Now we can find first a sequence {Dm} of distributions with finite support such that D.xq(n)(n)}. q(n) q(n) pi(n)a . r # oo.} of phasetype distributions such that Bn 3 B as n + oo. Here are the details at two somewhat different levels of abstraction: (diagonal argument .8. we can assume that ID..
..14 is fundamental and can motivate phasetype assumptions. u Theorem A5.. Hence G C PET and L = PIT.. oo) approximation Assume that we can compute a functional W(B) when B is phasetype. then it is immediate that WI(B) = p2(B) for all distributions B on [0. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. But To is the class G of all distributions on [0. however. and we can take Bn = Cr(n).t.n.15 To a given distribution B on (0 .. In particular.. compute W(B) and use this quantity as an approximation to cp(B0).i.n( b k ) .. in at least two ways: insensitivity Suppose we are able to verify a specific result when B is of phasetype say that two functionals Cpl (B) and W2 (B) coincide. If Cpl (B) and ^02(B) are weakly continuous. PIT contains all finite mixtures of onepoint distributions. u 2 (abstract topological ) The essence of the argument above is that the closure (w. 2. say on the claim size distribution B in risk theory. x 4 oo.( dx) * f r f{(x)B(dx). the class CO of all discrete distributions.e.(x)Bf. Let E be the class of functions f : [0. oo).D(bk)I < n. k < n. It should be noted. Then ICr( n ).B(bk )I < .r. i = 1. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. we can then approximate Bo by a phasetype B. i. replications).APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). E E. For a general Bo. oo) * [0..d. oo) and any fl. for some a < oo. oo) such that f (x) = O(e«x). if information on Bo is given in terms of observations (i. one would use the B given by some statistical fitting procedure (see below). f2. Since PET is closed under the continuous operation of formation of finite mixtures. k < n. and that cp is known to be continuous.n (bk) . there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. Corollary A5..
. and the case of a general f then follows from the definition of the class E and a uniform integrability argument..14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx).n(dx) + f 0 fi(x)Dn(dx). . i=1. . n B=az. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. then cc f (x)Bn ( dx) = (?!c ) e'= .. i = 1. Now returning to the proof of (A. .f (z) = f = 1 1 1 1n/ o ..39) Indeed. oo)..358 Proof By Fatou' s lemma.38 ). By (A. liminf B. i = 1..n(dx) < 1+.. . \\ 0 Corollary A5. n. we may assume that in the proof of Theorem A5..16 To a given distribution B on (0 . f00 fi(x)Cr..  APPENDIX B implies that 00 o o 00 n. f° xtBn(dx ) * f °° x`B( dx). 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 ). 2..2 .(dx) > J fi(x)B(dx). .oo J fi(x)B..39).. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. TO (A.f (x)B(dx). and hence we may choose r(n) such that L 9l) f (x)Cr(n).f ' f (x)B(dx).38) We first show that for each f E E. n. if f (x ) = e°x.. for each i.
and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. . We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . from a more conceptual .18 In the setting of Corollary A5. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. (N. lim inf > is proved similarly. O We state without proof the following result: Corollary A5. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal./3) is defined as the unique solution > 0 of B[y] = l+y/j3. the adjustment coefficient 'y = 7(B. there is a sequence {B. . . . e ) and ei J. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. one can obtain 7(Bn. but are certainly not unexpected.> y for some sequence {ei} with ei E (0.. If ei > 0.3).l3µb < 1. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions.14 is classical. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . . (N or a given distribution Bo. /3) = ry for all n. . . I. Proof Let fi(x) = el'r+E.e.16. lim sup ryn < 7. The adjustment coefficient is a fundamental quantity.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn. the remaining results may be slightly stronger than those given in the literature. 5e Phasetype fitting As has been mentioned a number of times already. 0 as i * oo.. However. The present section is a survey of some of the available approaches and software for inplementing this. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying .. For practical purposes. Notes and references Theorem A5. oo) with B[y +e] < oo for some e > y = 7(B./3).17 To a given /3 > 0 and a given distribution B on (0.
's). Schmickler (the MEDA package. In a series of papers (e.g. for some suitable large n. Of course. risk theory. cf.f. we have constructed a sequence { B.} of phasetype distribution such that Bo.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 .g. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. and used a nonlinear programming approach . Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . 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. giving mass 1 /N to each S=. (N is the empirical distribution Be. g. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . The characteristics of all of these methods is that even the number of parameters may be low (e. d..d. 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. It seems therefore a key issue to develop methods allowing for a more general phase diagram. [202].. . one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). at a a number of selected points . [216] ).. three for a mixture of two Erlangs ). a program package written in C for the SUN workstation or the PC is available as shareware. we do not not want to perform matrix calculus in hundreds or thousands dimensions).. defined by the absence of loops in the phase diagram . The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . [70]) restrict attention to acyclic phase type distributions . e . and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. [317] ) has considered an extension of this setup. reliability or queueing theory.f. A method developed by Bobbio and coworkers (see e. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory.g. and as fitted distribution we may take B. . A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly. the L1 distance between the c . The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter.g . and in practice this sets a limitation to the usefulness (the curse of dimensionality .
. N Ti = I(J= i) dt. one is lead to an iterative scheme.g. .x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations.g. . since this is parameterdependent. = j) f k=1 k =1 tE[0.. it is easy to see that N (k Ea(n). EN where ai = N 1 I ((k) = i) tii=i iEE. then the estimators would be of simple occurenceexposure type. .. eieT(n)((k..T(n) (Nik IC1. .APPENDIX 361 tion Sk was available. Thus. (n+1) _ Ea (n). In fact. E..(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j). (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. the methods of [70] and [38] appear to produce almost identical results. . Nii = = .T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei . it seems open whether the restriction to the acyclic case is a severe loss of generality.T(n) (Ti ^^ 1. e. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations...(N) = E Ea(n). (N) tJk Ea ( n).T (n)(TiI(1... jEEA. In practice.
This page is intentionally left blank .
213229. Appl. ed. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . II. AT&T. 311338. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. J. Abate & W. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . [2] J. G. Dover. [7] E. Stuttgart. Teubner. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . 57. pp. [10] K. Chichester New York. Asmussen (1985) Conjugate processes and the simulation of ruin problems. [5] G. 53100. Adv. ibid. 253267. Arfwedson ( 1955) Research in collective risk theory. 14. Skand. Scand. New York. 143170. New York. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. with applications to risk reserve processes and the GI /G/1 queue. The case of equal risk sums. Abate & W. 345368.A.G. New York. [9] G. Appl. Skand.Bibliography [1] J. 219229. Queueing Systems 5 . [13] S. Choudhury & W. Probab. Borokov. Act. [12] S. [8] G. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . 587. Abramowitz & I. 3157 . [15] S. Arfwedson ( 1954) Research in collective risk theory. Queueing Systems 10. [11) S. B. [6] V. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. 191223. Aktuar Tidskr. 363 . Asmussen (1987) Applied Probability and Queues. The Mathematical Scientist 14. [4] M.). Proc. Aktuar Tidskr. [3] J. Abate. John Wiley & Sons. Optimizations Software. 37.L. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. 1985.). 101116. 20. Alsmeyer (1991) Erneuerungstheorie. Transactions XVth International Congress of Actuaries. Stoch. Preprint. 1984 . Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Queueing Systems 16. [14] S. 38. In: Limit Theorems and Related Problems (A. Asmussen (1984) Approximations for the probability of ruin within finite time .
Stoch. Act. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. CRC Press. Appl. 354374. Random Processes 2. Asmussen (1998c) A probabilistic look at the WienerHopf equation. 4966. Probabilistic Analysis of Rare Events (V.364 BIBLIOGRAPHY [16] S. Scand. [21] S. Extremes 1 . 313326. Dshalalow ed. 96107. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. New York. Rolski & V. Florida. 1999 . Boca Raton. Asmussen & B. Asmussen. Scand. [32] S. 315. Scand. [27] S. Asmussen (1995) Stationary distributions via first passage times. Scand. Methods V Problems (J. Asmussen & M. 8. Proc. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. [22] S. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. Binswanger & B. Asmussen. 2943. [26] S. [17] S. J. 313341. [31] S. Asmussen (1989b) Risk theory in a Markovian environment. 106119. eds. L. [20] S. [34] S. [19] S. Ann. 20. Proc. Bernoulli 6. Asmussen & K. Hojgaard (2000) Rare events simulation for heavytailed distributions. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Marcel Dekker. Asmussen. [28] S. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. stationary distributions and first passage probabilities. Riga Aviation University. 37. Asmussen & M. Astin Bulletin 27. 69100. eds. 772789. [30] S. J.). Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Andronov. 1996. [24] S. 297318. 1989 . K. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. 1936. Floe Henriksen & C. 555574. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve.). Frey. Statist. Asmussen (2000) Matrixanalytic models and their analysis. 193226. 303322. [25] S. [29] S. Asmussen & B. 25. J. A.S. 54. Appl. Advances in Queueing: Models. Appl. 27. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Alfa & S. Act. Stochastic Models 11. 2. Asmussen (1992b) Light traffic equivalence in single server queues. Ann. Chakravarty. [33] S. 2149. Appl. Ann. 79102. Asmussen (1998b) Extreme value theory for queues via cycle maxima. . Probab. 189201. Probab. Act. Asmussen (1999) On the ruin problem for some adapted premium rules. SIAM Review 40. Th. [23] S.K. MatrixAnalytic Methods in Stochastic Models (A.M. J.). Asmussen (1992a) Phasetype representations in random walk and queueing problems. Stoch. [18] S. Probab. 137168. T. Kalashnikov & A.
419441. Management Science 45. Adv. Probab. Asmussen & H. 58. Nerman & M. Schmidli & V. Hojgaard & M. first passage problems and extreme value theory for queues. with applications to insurance risk Stoch. Proc. Asmussen & T. Wiley. Koole (1993). Asmussen & C. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. [37] S. 422447. [47] S. [36] S. 299324. 103125.A. 64. Proc. J. [44] S. Appl. Statist. [43] S. [48] S. [42] S. 120. Fitting phasetype distributions via the EM algorithm. B. Asmussen & C. [51] S. 410433. Finance and Stochastics 4. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. H. Appl. (to appear). Stochastic Models 8.Y. Kliippelberg (1996) Large deviations results for subexponential tails. Florida. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. 115. Probab. Appl. Asmussen & M. Asmussen & G. Example of excessoffloss reinsurance for an insurance corporation. Asmussen & S.). Prob 32 . Math. 30. Taksar (1997) Controlled diffusion models for optimal dividend payout. 23. 19. Olsson (1996).A. 11251141. [41] S. Rubinstein (1999) Sensitivity analysis of insurance risk models. [49] S. Scand. Res. Th.M. J. Nielsen (1995) Ruin probabilities via local adjustment coefficients . [50] S. Dshalalow ed. [45] S. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. Asmussen & R. Statistica Neerlandica 47. Asmussen & V. [52] S. Eng. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. 20. Insurance: Mathematics and Economics 10. [40] S. Stoch. Schmidt (1995) Ladder height distributions with marks. Perry (1992) On cycle maxima. 31. J. Read eds. Asmussen & K. Adv. 259274. CRC Press. Sc. Supplementary Volume (Kotz. Boca Raton. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Taksar (2000) Optimal risk control and dividend distribution policies . Asmussen. Asmussen & R. Johnson.BIBLIOGRAPHY 365 [35] S. 913916. O.Y. 10. Inf. . [39] S. Probab. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. [38] S. Marked point processes as limits of Markovian arrival streams. Res. 429466. Opns. Appl. Asmussen & C. Asmussen & V. 736755. Asmussen & T. Insurance: Mathematics and Economics 20. Asmussen. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Asmussen & D. Asmussen . 421458.). 105119. Advances in Queueing: Models. Oper. Appl. Appl. Sigman (1996) Monotone stochastic recursions and their duals. Methods 8 Problems (J. Probab. [46] S. 365372.
H. Trans. Clarendon Press. Act.T. Asmussen & J. [69] P. [67] T. J.A. [60] J. Halsted Press. Cox (1972) A low traffic approximation for queues. J. Act. J. SIAM. [66] N. 1985 . von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Oxford. [57] C. Bloomfield & D. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Scand. Goldie & J. 190204.A. Grandell (1985) An insensitivity property of the ruin probability. Bingham. 33. [55] B. [65] N. Cambridge University Press. [54] K. de Waegenaere (1990) Simulation of ruin probabilities. [61] J.366 BIBLIOGRAPHY [53] S. Soc. 148156. New York. Bjork & J. 77111. Adv.L. J. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. New York. 1975. Beekman (1969) A ruin function approximation. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. Stochastic Models 10. Insurance: Mathematics and Economics 6. Appl. Bobbio & M. Scand. 9. Insurance: Mathematics and Economics 4. [70] A. BarndorffNielsen & H. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. Probab. Springer . 705766. [73] A. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. New York. [72] P. 9599. Bingham (1975) Fluctuation theory in continuous time. [68] T. [59] O. Borovkov (1976) Asymptotic Methods in Queueing Theory. [58] O. 275279. Act. Baker (1977) The Numerical Solution of Integral Equations. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. 832840. Scand. Wiley. 129134. 661667.M. Beekman (1985) A series for infinite time ruin probabilities. 11811190.J. SpringerVerlag.R. J. Cambridge. [56] B. Act. 4148. [63] A. Appl. .H. J. Appl. Act. 1974. Beekman (1974) Two Stochastic Processes. Actuaries 21. Wiley.B . Probab. 610. Scand. 221232. Boogaert & A. Teugels (1987) Regular Variation. 1995. 7. Billingsley (1968) Convergence of Probability Measures. 169186. Athreya & P. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. J. Berlin. [64] P. Insurance: Mathematics and Economics 9. Berman & R. Ney ( 1972) Branching Processes . Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Bjork & J. [71] P. Scand. 1988 . Probab. [62] J.L. Boogaert & V. C.H.
Cramer (1930) On the Mathematical Theory of Risk. [90] D. 34. Skandia Jubilee Volume.R.R. [93] K. 392433.W. [92] H. Proc. The Society of Actuaries . Bell. Bratley. Burman & D.U.I. [75] O. [89] M. New York. Queueing Systems 33.L. J. Probab.M.A. Bucklew ( 1990) Large Deviation Techniques in Decision.J. IEEE J. D. Resnick & R. Amsterdam. Jr. Simulation and Estimation.C.L. Goovaerts & F.W.BIBLIOGRAPHY 367 [74] O. . Burman & D . Springer. Broeck .L. Veraverbeke (1990). Berlin. Oper. 14. 177204. Reading. 3539. [85] W. Sel. Bux & U. 749763. Insurance : Mathematics and Economics 9. Adv. 127130. [91] H. Amsterdam. Qinlar (1972) Markov additive processes. Aut.Y. 24. Geb. Control AC28. Nonparametric estimators for the probability of ruin. verve. IEEE Trans. Chung (1974) A Course in Probability Theory (2nd ed. M. Computer Performance (K.J. J. New York. [86] K. 93121. S. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Stockholm.. Cottrell. 907920.) NorthHolland. 2338.).C. 16. Insurance : Mathematics and Economics 5. De Vylder (1986) Ordering of risks and ruin probabilities. Hickman. [81] J. Academic Press. Appl. Illinois. [88] J. Gerber. Soc. SpringerVerlag. Wiley. Nesbitt (1986) Actuarial Mathematics .L. [84] D. 62. 313319. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. H. II. [80] F.A. Jones & C. Res. Boxma & J. 51. Chandy & M.Y . Cambr. AddisonWesley. Cox (1955) Use of complex probabilities in the theory of stochastic processes. Croux & N. [79] P. Tweedie (1982) Storage processes with general release rule and additive inputs . J. Breiman ( 1968) Probability. Tech.). Boxma & J. [87] E. [78] L. [82] H. B. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Wahrscheinlichkeitsth. Schrage (1987) A Guide to Simulation. Cramer (1955) Collective risk theory. Reiser eds. Herzog (1977) The phase concept : approximations of measured data and performance analysis. 14331453. 105119. Z. [76] N. Biihlmann ( 1970) Mathematical Methods in Risk Theory. Stockholm. Brockwell . Fox & L. Cohen (1982) The Single Server Queue (2nd ed.W. New York San Francisco London. The Jubilee volume of Forsakringsbolaget Skandia. NorthHolland. Syst. Bowers . [83] D. [77] P. Philos. Fort & G. Itasca.R. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Area Commun.
J. Devroye (1986) Nonuniform Random Variate Generation. Dempster. Boston. Haezendonck (1985) Inversed martingales in risk theory. [95] M. Actuarial J. Appl. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. [106] A. Embrechts (1989). 227229. 9 . [108] F. Daley & T. 85116.368 BIBLIOGRAPHY [94] M. N. Act. [111] F.D. Csorgo & J. 435436.. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Pesonen (1994) Practical Risk Theory for Actuaries. 7083 (1969). 5771. Statist. Zeitouni (1992) Large Deviations Techniques and Applications. Insurance: Mathematics and Economics 6. Scand. Math. 3. Daley & T. Math. Rolski (1991) Light traffic approximations in queues. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. 583602. Delbaen & J. Editions de 1'Universite de Bruxelles. De Vylder & M. Stochastic Models 5. Goovaerts (1984) Bounds for classical ruin probabilities. Jones and Bartlett. Laird & D. Deheuvels & J. [110] F. Insurance: Mathematics and Economics 4. [96] D. 201206. 27. Insurance: Mathematics and Economics 3. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. 16.P. 181217. T. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Soc.M. Martingales and insurance risk. De Vylder & M. [105] A. Aktuar. 135159. [98] A. Springer.. [99] R. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). Daykin. English version published in Skand. 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. Dembo & O. Oper. Tidskr. [104] F. [109] F. 114119. J. Appl. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. Chapman & Hall. [112] F. Dassios & P. De Vylder (1996) Advanced Risk Theory. New York. Suppl. Haezendonck (1987) Classical risk theory in an economic environment. [103] F. Stockholm. [107] L. . [97] D. Delbaen & J.L.B. Rolski (1984) A light traffic approximation for a single server queue. Math. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . J. 1990 . J. Pentikainen & E. Oper. Res. Roy. Comp. [100] A. 22. 88101. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . [101] C. Insurance: Mathematics and Economics 7. Scand. Probab. 3750. [102] P. 17. Res. 121131. Csorgo & J. 624628. Insurance: Mathematics and Economics 10.
Scand. Ruin estimates for large claims .G. Hipp (1999) Ruin problems for phasetype(2) risk processes. J. Gerber & E. Letters 7. J. Waters (1999) Ruin probabilities wih compounding. Proc. Act. 229232. Soc. Dickson & H.M.M. Dickson (1992) On the distribution of the surplus prior to ruin .BIBLIOGRAPHY 369 [113] D. Grandell & H.R. Shiu (1991) Risk theory with the Gamma process.U. 1984 . H.M.M.C. Dickson & H. [129] D . 269274. with applications. Insurance: Mathematics and Economics 7. 107124. [119] D. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . 105115.C. Res. 1993 .C. [127] F.R. Scand.C. Dufresne.J. Act.S. J.C. 251262.M. [116] D. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. 3745. Act. Hipp (1998) Ruin probabilities for Erlang (2) risk processes. 1741. Math. J.C. Dufresne & H. Embrechts . Insurance : Mathematics and Economics 7. Emanuel .W. Scand. J. [117] D.R. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. Insurance : Mathematics and Economics 22. 118. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. 1. 1975 . Act. Oper. Scand. [130] P. Astin Bulletin 21. Berlin Gottingen Heidelberg.R. Scand. 2000 . 147167. Philos. 4962. Dickson (1994) An upper bound for the probability of ultimate ruin. van Doorn & G.J. [122] B. Djehiche (1993) A large deviation estimate for ruin probabilities.M.B. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. [131] P. 174186.U. [114] D. [128] E.C. 1984 . Dynkin (1965) Markov Processes I.C. 193199. Gerber (1988) The surpluses immediately before and at ruin.K. [125] F. Act. Dickson & J. Harrison & A. 177192. Insurance : Mathematics and Economics 10. Waters (1996 ) Reinsurance and ruin . Springer . Regterschot (1988) Conditional PASTA. British Actuarial J.C. [124] N. Dickson & J.U. J. [118] D.M. J. [126] F. J. Dickson & C.M. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. and the amount of claim causing ruin. Insurance : Mathematics and Economics 11. 1994 . 4259. . [120] D. 5159. Dickson (1995) A review of Panjer's recursion formula and its applications. 6180. Dufresne & H. Act. [115] D. 191207. Insurance: Mathematics and Economics 19. J.C. Insurance: Mathematics and Economics 25. Scand. Duffield & N. [121] D. 1993 . Embrechts ( 1988 ). Act. 131138. [123] E. Scand. Dickson & C. Camb.M. 363374.M.
Appl. Teugels (1985). 1998 . Embrechts & J. Lai (1998) Wald's equations. Scand. Adv. 26. 5572. Frees (1986) Nonparametric estimation of the probability of ruin. Probab.370 BIBLIOGRAPHY [132] P. Griibel & S. Statist. [147] M. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. Fuh & T. 623637. [144] P. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . . Mikosch (1991). D. 3339.K. Fields 75. 14661474. Embrechts. Insurance: Mathematics and Economics 7. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . J. R. [139] A. [146] E. Act. Arndt & V. [135] P. Appl. 181190. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Proschan & D. Furrer (1998) Risk processes perturbed by astable Levy motion. [141] F. Appl.A. Skand. 38. Probab. Walkup (1967) Association of random variables. Esscher (1932) On the probability function in the collective theory of risk. 29. New York. Wiley. [133] P. Villasenor (1988) Ruin estimates for large claims. Schmidli (1994) Ruin estimation for a general insurance risk model.M. Springer. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. Akt. with applications. [140] J. Math.K. Frenz & V. Approximations for compound Poisson and Polya processes.L. Nyt Tidsskrift for Matematik B20. [138] P.). Trans. [148] C. 5974. J. Appl. Probab. Wiley. M. 5975. Astin Bulletin 16. Probab.D. 269274.D. first passage times and moments of ladder variables in Markov random walks. [143] W. Insurance : Mathematics and Economics 1. [142] W. Embrechts.L. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. 695712. 29. Sci. 175195. 17. Schmidt (1992) An insensitivity property of ladder height distributions. Embrechts & N.W. Rel. A bootstrap procedure for estimating the adjustment coefficient. [150] H. Danish Academy Tech. Heidelberg. Kliippelberg & T. Maejima & J. Fitzsimmons (1987). Probab. Esary. Appl. F. J. Wiley. [134] P. C. Insurance: Mathematics and Economics 10. Adv. 616624. Jensen. U. On the excursions of Markov processes in classical duality. Adv. [136] P. Schmidt (1982) Queues and Point Processes. Tidsskr.J. Statistica Neerlandica 47. Embrechts & T. 566580. Ann. Franken. New York. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. 35. Embrechts. Konig.W. [137] P. 131137. 404422. J. [145] P. 159178. Embrechts & H. 8190.D. [149] C. Probab. Th.). Erlang 2.
London. ETH Zurich. [154] H. Ver. Probab.N. Amsterdam. 424445. [163] P. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Haezendonck (1984) Insurance Premiums. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Michna & A. Probab. [168] J. F. Appl.). 3752. 28. NorthHolland. [157] H. [164] B. S. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk.W. 6370. Griibel (1996) Perpetuities with thin tails. Adv. 5. Math. [162] P. Appl. Scand.U. contained in the authors PhD thesis.G. Aktuarietidskrift 1970 . 1523. 1978 . Suppl. J. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue.U. 205210. Furrer & H. R.U. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Life Insurance Mathematics. Goldie & R. Galambos & J. Mitt. Gerber (1973) Martingales in risk theory. Scand. 463480. Amsterdam.U. [158] H. 131156. Springer.U. Vers. 1981 . Insurance : Mathematics and Economics 7. Appl. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. Kluwer. Glasserman & S. Scand. [161] P. Furrer. J. Insurance: Mathematics and Economics 20. NorthHolland. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'.S. In Studies in Applied Probability (J. Huebner Foundation Monographs. [159] H. Glynn & W. Birkhauser . Mitt. Dordrecht.. [155] H. 73. Act.71.U. [153] H. Gerber (1979) An Introduction to Mathematical Risk Theory. Ann. Schweiz. Math. Goovaerts. Insurance: Mathematics and Economics 15. J.). Goovaerts.BIBLIOGRAPHY 371 [151] H. van Heerwarden & T. Act. Unpublished. Kaas. 2336. [160] H. . de Vylder & J. Kou (1995) Limits of first passage times to rare sets in regenerative processes. Bauwelinckx (1990) Effective Actuarial Methods. Z. [165] M. Gerber (1988) Mathematical fun with ruin theory.. A.J. 1977. Skand.M. [166] M.V. [152] H. Act. [156] H. Schweiz.U. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Ver. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. J. eds. [167] C. Gani. University of Pennsylvania. 31A. Vers. Grandell (1977) A class of approximations of ruin probabilities. Boston. 105115.J. Gerber (1986). Weron (1997) Stable Levy motion approximation in collective risk theory. Basel. Gnedenko & I. 97114. Probab.E. [169] J. 7778. 205216.
Gusak & V. [174] J. Segerdahl (1971) A comparison of some approximations of ruin probabilities. Grandell (1990) Aspects of Risk Theory. J. Opns . 566576. [183] M. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Insurance: Mathematics and Economics 26. 14. Grandell & C. 5 . [188] J. Skand.O. Math. Amsterdam. M. 347358. Lemoine (1977) Limit theorems for periodic queues.I. SpringerVerlag. M. Rink. Matem. Gyllenberg & D. Res. Probab. 197214. mathematische Wirtschaft .). . van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. Stoch. 355365. 143158. [176] B. Appl. Math. Harrison (1977) Ruin problems with compounding assets .and Sozialforschung 6.I. Grigelionis (1996) Lundbergtype stochastic processes. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . Informatica 2. Preprint. Theory and Applications. Applied Statistics 40.243255. SpringerVerlag.V. 167176. Grnbel (1991) G/G/1 via FFT. Archiv v. [180] R. [182] A. Probab. Proc . M.131135. [185] J. Winter School on Stochastic Analysis and Appl . Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . Harrison & S. 5766. 54. 1. Proc. AkademieVerlag. KTH. 14. [187] J. [172] J. 3 . Res.. Statistical Algorithm 265. [171] J. [184] H . Opns . Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. Astin Bulletin 29. Grandell (1997) Mixed Poisson Processes.372 BIBLIOGRAPHY [170] J.S. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . Resnick (1977) The recurrence classification of risk and storage processes . 8. 400409. Appl. Grandell (1999) Simple approximations of ruin functions. [173] J. 3041. Th. 332. [179] R. Appl. Harrison & S. [186] J. M.A. Appl. Griibel & R. Proc. New York. [181] D. Harrison & A. Stoch. 33. E. Chapman & Hall. Probability Theory and Mathematical Statistics (I. Aktuar Tidskr. Gut (1988 ) Stopped Random Walks . Liet. Grandell ( 1992) Finite time ruin probabilities and martingales . 7590. [177] B. Ibragimov. [189] A. [175] J. Grandell (1979) Empirical bounds for ruin probabilities. 6779.J. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. [178] B. Tinbergen Institute Research Series 20. ed. Berlin. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum .
285293.M. Haggstrom. [192] U. Hogg & S. Versicherungswirtschaft. [199] R. 11631174. Wahrscheinlichkeitsth. Horvath & E. 259268. 4385. Hipp & R. The ruin problem for finite Markov chains. Nachrichten 30. Electronic Letters 35. J. [195] B. Klugman (1984) Loss Distributions. 123151. [197] C. 6.A. 1990 .B. Insurance: Mathematics and Economics 5. Ann. S. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. Statist. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Prob. Insurance: Mathematics and Economics 5.V. Mathematical Statistics. Act.L. Verlag Versicherungswirtschaft e. 12981310. 377381. Astin Bulletin 19. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. Chalmers University of Technology and the University of Goteborg.R. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. Z. New York. No. Nerman (1992) EMPHT . 285292. Scand.). Karlsruhe. Asmussen & 0. verve. 378389. 305313. Proc. [207] D. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. Act. Hojgaard & M. Karlsruhe. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. Hoglund (1991). on Asymptotic Statistics (P. [194] 0. 5770. 8996. Hill (1975) A simple general approach to inference about the tail of a distribution. Probab. Ann. [200] M. [191] W. Scand.V. 18.. [206] B. [203] T. 23. eds. [201] L. Heilmann (1987) Grundbegriffe der Risikotheorie. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. [208] V. 3. Huskova.BIBLIOGRAPHY 373 [190] P. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. 19. ACM TOMACS 6. 166180. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. [204] T. J. [205] T. Studies in Statistical Quality Control and Reliability 1992 :4. 29. J. . Foruth Prague Symp. [198] C. Ann. Iglehart (1969) Diffusion approximations in collective risk theory. Mandl & M. [196] C. Geb. Appl.L. Probab. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. 8095. Probab. Math. [202] 0.a program for fitting phasetype distributions. 1998 . J. [193] U. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. 25. Wiley. Iversen & L. Hipp (1989a) Efficient estimators for ruin probabilities. Appl.
165167. Appl.G. [216] M.R. 60. Statist. Soc. [210] D. Jagerman (1985) Certain Volterra integral equations arising in queueing. [226] J. 547567. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk.M. J. ibid. Astin Bull. 561563. Soc. Munksgaard. 37. [220] V. 8793. ibid. Stochastic Models 1. N. Jelenkovic & A. 259281. [218] V. Johnson & M. Cambridge Philos. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. 3. 866870. [217] R. [223] J. Cambridge Philos. Appl. 35. Act. 11. University of Chicago Press. Princeton. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application.C. Proc. Keilson & D. Soc. 711743. Kao (1988) Computing the phasetype renewal and related functions. J. Stochastic Models 5. Janssen & J. Cambridge Philos. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Taylor (1981) A Second Course in Stochastic Processes. Kluwer.A. Proc. New York.M. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain.J. Goovaerts (1986) General bound on ruin probabilities. Act. Jensen (1995) Saddle Point Approximations. 118. Probab. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. 239256. [227] J. Insurance: Mathematics and Economics 5. 187193. N. Keilson & D. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. Kaas & M.G. Kalashnikov (1996) Twosided bounds of ruin probabilities. Keilson & D. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues.B. 61. Taaffe (1989/90) Matching moments to phase distributions. Manuscript.M. [213] P. 63. 15. 4151. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. 1996.M. Kemperman (1961) The Passage Problem for a Markov Chain. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. Proc. Jensen (1953) A Distribution Model Applicable to Economics. [215] J. Karlin & H.374 BIBLIOGRAPHY [209] D.J. Math. Scand.P. . Oxford. [221] E. [225] J. Clarendon Press. Probab. [212] J. [214] A. Academic Press. 31. [228] J. J. 116129. J. 283306. Addenda to for processes defined on a finite Markov chain. 6. Ann. [211] J.L. NEC Research Institute. Copenhagen.G. 123133. Astin Bull.H. 173190.M. [219] V. 6. Amer. Technometrics 30. [224] J. Chicago. Wishart (1964). [222] S. 325247.
125147. . Insurance: Mathematics and Economics 8.F. 26. Actuarial J. Appl. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. [240] U. [237] C.F. Macmillan. B24. (235] C. [231] J.F. Philos. [234] C. [236] C. Proc. [245] T. Lehtonen & H. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. 133135 (in Russian). J. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. Philos. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. [244] T. [246] D. Quart. J. Kingman (1962) On queues in heavy traffic.F. SIAM. Kliippelberg (1988) Subexponential distributions and integrated tails. Statist. Oxford 12. Quart. Soc. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input.. 6075. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. 25. Klnppelberg & T. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Kunita (1976) Absolute continuity of Markov processes. 60 . Lipsky (1992) Queueing Theory . Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Prob. 132141. Act. Mikosch (1995) Modelling delay in claim settlement. Soc. Appl. [239] H. Scand. 390397. Klnppelberg & T. [241] G.BIBLIOGRAPHY 375 [229] J. [243] A.J. [232] C. Sorensen (1997) Exponential Families of Stochastic Processes. Lehtonen & H. Penev & A. [233] C. Kingman (1961) A convexity property of positive matrices . J. SpringerVerlag. Scand. Kuchler & M. 1998 . [230] J. 279285. 48. Camb.S. Adv. 259264.P. Latouche & V. 283284.a Linear Algebraic Approach. Soc. Probab. 889900.C.C. Kliippelberg & U. Appl. Korolyuk. 154168. Scand. J. Act. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities.C. 24. 18. Cybernetica 4. Seminaire de Probabilties X . I. Appl. New York. Kingman (1964) A martingale inequality in the theory of queues. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . 277289. [247] L. 4477. Roy. [242] A. SpringerVerlag. 146. Insurance : Mathematics and Economics 12. J. [248] D. Lindley (1952) The theory of queues with a single server. Lecture Notes in Mathematics 511. 858874. 359361. J. [238] V. Probab. Bernoulli 1 . 4958. 383392. Stochastic Models 7. Cambr.J. Lemoine (1981) On queues with periodic Poisson input. 1995 . Math. J. Proc. Probab.
Appl. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. Probab. Soc. [253] V.K. Nummelin (1987) Markov additive processes I. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. [259] Z .D. Scand. . J. Eigenvalue properties and limit theorems . Scand. Nagaev (1957) Some limit theorems for stationary Markov chains. Scand. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. 561592. Probab . J. Cambridge Philos. [264] G. Act. Lucantoni . Proc.S. Probab. 429448. J. 1986 . 161174. 58. Appl. II Aterforsdkring av Kollektivrisker . Probab. [258] K . Stoch. 763776. [250] F. Ney & E. J.D. Act.376 BIBLIOGRAPHY [249] D. 3. Ann. Uppsala. Appl. [262] H . 1261 (0?)1270. 2939. Insurance : Mathematics and Economics 5. Proc. Appl. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. [255] V. Ann. Math. Opns. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. Appl. with an application . K. 223235. a useful concept in risk theory. [266] P. [265] S. 22. 4859. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. [256] A.). Europ. 32. Statist. 1994. Miyazawa & V.M. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Anal. MartinL3f (1986) Entropy. F. Hoist eds. [257] A. Ann. Gut & J. 378406. Probab. J. 147149. 1996 . Adv. J. Almqvist & Wiksell. Meier ( 1984). Michna ( 1998) Selfsimilar processes in collective risk theory. [260] H. . 676705.V. 286298. 11 . Makowski ( 1994) On an elementary characterization of the increasing convex order . MeierHellstern & M .V. Th. Appl. [261] H. 834841. 58 . [263] M . [251] F. 36. Cambridge Philos. Englunds Boktryckeri AB. 15. Stockholm. 370377. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . [252] A. 31. 268285.F. Probability and Mathematical Statistics (A. Act. 29 . Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . Soc. MartinL6f (1983) Entropy estimates for ruin probabilities .S.K. Res. J. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. Math. 124147.D. 2. Probab. Schmidt ( 1993) On ladder height distributions of general risk processes . [254] V .
Neuts (1977) A versatile Markovian point process.F.F. Neuts (1992) Models based on the Markovian arrival process. Probab. . Abh. Models 6. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Johns Hopkins University Press. 46. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Appl. IEICE Trans. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. 339353. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. Stoch. 10081026. [272] J . 316. Appl. Appl. Paulsen (1998) Ruin theory with compounding assets . Paulsen (1993) Risk theory in a stochastic economic environment. 21. 327361. Astin Bulletin 5 . Naval Research Logistics Quarterly 25. Stochastic Models 3 . Appl. 135148. 12.F. 123144.a survey. Appl.K. 71. 157. London.F. 764779. 445454. 273299. Norberg ( 1990) Risk theory and its statistics environment . E75B . [276] C. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Appl. 30. 12551265. Proc. Probab. Probab. [275] H. [285] J. Insurance: Mathematics and Economics 22.G. Paulsen & H.F.BIBLIOGRAPHY 377 [267] M . Omey & E. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. [280] A. Baltimore .K. [274] H . Probab. Adv. Stoch. 16. O'Cinneide (1990) Characterization of phasetype distributions. Appl. J. 36. 215223. Insurance: Mathematics and Economics 20. 311342. Commun. [270] M . Proc. 733746. Stoch. Nyrhinen (1999) Large deviations for the time of ruin. New York. [278] E. J. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Paulsen & H. Proc.A. Marcel Dekker. [277] J. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. 249266. Neuts ( 1978) Renewal processes of phase type. Appl. J. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Sem. Omey & E. [269] M . 75. [279] E. [282] J . [268] M . Hamburg. Proc. [271] M. Math. [283] J. [281] J. Stoch. Statistics 21. Pakes (1975) On the tail of waiting time distributions. [284] J. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. 555564. Stoch. [273] R.
Statist. Stat. Embrechts (1996) Confidence bounds for the adjustment coefficient. 222261.A. Appl. 215246. Oper. Philipson (1968) A review of the collective theory of risk. Rao (1965 ) Linear Statistical Inference and Its Applications. [295] N. 2330. Prabhu & Zhu (1989) Markovmodulated queueing systems. Paulsen & H. [292] S. Ann. [301] G. Resnick & G.. Probab .G.G. J. [293] N. 337347. Insurance Risk.378 BIBLIOGRAPHY [286] J. Aktuar. Opns. Skand. Appl. [290] E. Queueing Systems 5. Astin Bulletin XIV.U. Gjessing (1997c) Ruin theory with stochastic return on investments. 29. 568576.L. Ripley (1987) Stochastic Simulation . Queues. Springer. Wiley.K. and Dams. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. Adv. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Prabhu (1980) Stochastic Storage Processes. Math. 965985. Act.J. Pitman (1980) Subexponential distribution functions. 4568. 235243.R. 757764. [289] C. Adv. 117133. Berlin. 2343. 537555. Griibel & P. Math. Adv. [302] J. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. [298] V.J.M. 30. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Prabhu (1961). [294] N.M. Scand. [299] C. 28. Probab. Math. [305] C. Probab. Probab. 19. Austr. 61. Pyke (1959) The supremum and infimum of the Poisson process. 147159. [297] R. Pitts. Res. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. J. Rolski (1987) Approximation of periodic queues. Regterschot & J. Prabhu (1965) Queues and Inventories. 4. Soc. [291] S. 12.H. Appl. [287] F. Appl. 465483. Statist. Appl. Wiley.U. Ann. 139143. Insurance: Mathematics and Economics 16. Math. . 29A. 1989 . Insurance: Mathematics and Economics 3. [288] S. New York. 691707. New York. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. Res. Tidskr. [296] N. Ann. 46. 32. 390413. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. J. Probab. Ann. On the ruin problem of collective risk theory.K. 11.U. [300] C.U.M. [303] S. Pellerey (1995) On the preservation of some orderings of risks under convolution. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Heidelberg. [304] B. R. 820827. Math. [306] T. Wiley. Inst.
155188. Statist. [314] T. [310] R. Stochastic Models 6. Rubinstein & A. Wiley. [319] H. 1997. 227244. 135149. 365388. 121133. Rudemo (1973) Point processes generated by transitions of a Markov chain. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. [311] R. 145165. [321] H. . Schmidli ( 1999b) Perturbed risk processes: a review . 5. Th. Act. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . 11. 4857. Rubinstein & B. Rubinstein (1981) Simulation and the Monte Carlo Method. Scand. [313] M. J. Act. Rolski. Probab. V. [320] H. Melamed (1998) Modern Simulation and Modelling. Ann. Proc.BIBLIOGRAPHY 379 [307] T. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. [326] H. 21. Appi. 262286. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . 5. [312] R. Schmidli. 5. 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.Y. [318] H. 687699. Stochastic Models 10 . J. Schmidli ( 1996) Martingales and insurance risk. 431447 [316] S. Wiley. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. H. Data Anal. Comp. Probab .J. [315] T. 795829. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. Stoch. Teugels (1999) Stochastic Processes for Insurance and Finance. [324] H . [309] H. 93104. Appl. [322] H. 2001 . Ryden (1994) Parameter estimation for Markov modulated Poisson processes. 7. 4068. Scand. Probab. Adv.M. Schmidt & J. Math. Insurance: Mathematics and Economics 22. [325] H . Stochastic Models 10. Insurance: Mathematics and Economics 16. Statist. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . 131156.Y. Seal ( 1969) The Stochastic Theory of a Risk Business. Appl. Astin Bulletin 29.L.Y. [323] H. [317] L. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. Singapore. Science Culture Technology Publishing . [308] S. Wiley. Schlegel ( 1998) Ruin probabilities in perturbed risk models. 417421. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Wiley. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Wiley. J. Operationsforsch.
Sengupta (1990) The semiMarkov queue: theory and applications. [332] C : O. Verein Schweiz Versich. Shin (1987) Convolution of uniform distributions and ruin probability. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . [346) D . Ann. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. [328] H. Appl. Shaked & J. . 21. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. [338] E. 75. Appl. 483487. Chapman and Hall. Stochastic Models 6. Act. [348] K . Chapman & Hall. Skand. 77100. Shwartz & A. Almqvist & Wiksell.A. Verein Schweiz. 121139. 11. [337] M. [329] H.L. Appl. 72. 157166. Risk teory and the single server queue .F. Scand. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. [340] E. 4383. [336] B. 61. [343] D. Ann. the probability of nonruin in an interval (0. Math. 171178. Versich. Adv. 429442. L. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. Stockholm. Insurance: Mathematics and Economics 8. New York.S. 914924. 701719. 159180. Mitteil. [341] E. 673684. Seber (1984) Multivariate Observations. 1955. Act. Th. In Probability and statistics .the Harald Cramer volume (ed. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Wiley.L. t). 1987. Segerdahl (1959) A survey of results in the collective theory of risk. 11. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. 191197. [347] K. Versich. Siegmund (1975) The time until ruin in collective risk theory. Probab. Shin (1989) Ruin probability by operational calculus. Sigman (1992) Light traffic for workload in queues. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Scand. . Mitt. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. [344] D. Academic Press. Probab.G. Grenander). t).W. [342] D. [335] B . Statist. pp. SpringerVerlag. Shantikumar (1993) Stochastic Orders and Their Applications. [334] C: O. 4. 279299. Verein Schweiz. Aktuar Tidsskr. 383413. [333] C: O. Seal ( 1972). J. 4.380 BIBLIOGRAPHY [327] H. Queueing Systems 11.S. 72. Seal (1974) The numerical calculation of U(w. Aktuar Tidsskr. Wiley. Mitt. Probab . [331] G. J. [345] D. Weiss (1995) Large Deviations for Performance Analysis. 243249.L. Siegmund ( 1985) Sequential Analysis . 2236.W. Seal (1978) Survival Probabilities. Math. Adv. Math. Probab.S. 1974. [330] H. [339] A.
Act. Astin Bulletin 16 . Thorin (1977).V.L. [365] J. [364] G. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Teugels (1982) Estimation of ruin probabilities . Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. Soc 49. 1978. Scand. [368] O. [369] O. Ruin probabilities prepared for numerical calculations. Scand. .and largedeviations probabilities in actuarial risk theory. Insurance: Mathematics and Economics 16. Teugels ( 1995 ) Ruin estimates under interest force . Unpublished manuscript. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . 1974.A. Slud & C. J. Karlsruhe. 2739. [353] E. New York. Scand. [352] D. 8199. Act. Proceeedings of the IEEE 77.C. 118. [354] B. 21. 8594. Act. Scand. 65102. Act. Appl. 1976. Taylor (1979) Probability of ruin with variable premium rate. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. Straub (1988) NonLife Insurance Mathematics. Smith ( 1953) Distribution of queueing times.S. Cambridge Philos. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities .. 114137. 197208. Verlag Versicherungswirtschaft e. Thorin (1982) Probabilities of ruin. Stanford & K. J.). Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Insurance: Mathematics and Economics 19. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. Adv. Insurance: Mathematics and Economics 1.. Astin Bulletin 24. J. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . [359] L. 722. NorthHolland.C. 163175. [350] W. J. John Wiley & Sons. 235254. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . [360] G. New York. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. Wiley.L.C. Probab. [351] D . 1982 . [356] B . Sundt & J. Astin Bulletin 12.C. Daley ed. [358] R. Sundt & J.L.).J. [357] B. 1980 . [363] G.J. 5776. J.L.C. Scand. 65102. Springer. Actuarial J. Scand. [367] O. 449461. Sundt & W. Act. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities.BIBLIOGRAPHY 381 [349] E. Proc. 149162. [355] B . [366] O. Taylor (1986) Claims reserving in NonLife Insurance. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. Hoesman ( 1989) Moderate. [362] G. [361] G. 725741.
In: Premium Calculation in Insurance (F. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. thesis. J. Appl. Lin (1994) Lundberg bounds on the tails of compound distributions. 5762. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. J.). [376] A. University of Michigan. Stationarity and Regeneration. 37. Unpublished Ph. 1983 . Scand. Dordrecht Boston Lancaster. Waters (1983) Probability of ruin for a risk process with claims cost inflation. Res. J. [377] V. 743756. 31. M. Goovaerts. . Skand. Astin Bulletin VII. D. Statistical Science 10. R. [374] F. Sherman & D. Reidel. [375] N. [371] 0. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Haezendonck eds. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. Astin Bulletin IX.E. Wiley. Oper. F. [384] R. PrenticeHall. [378] H. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. 142. 137153. Willmot & X. Van Wouve. [372] H. 936952.W. Thorisson (2000) Coupling. De Vylder & M. [373] S. M. 231246. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. Tidskr. [383] G. 4963. 1942. Submitted.382 BIBLIOGRAPHY [370] 0. [382] G. SpringerVerlag. Tacklind (1942) Sur le risque dans les jeux inequitables. Act.R. 6785. De Vylder. Wolff (1990) Stochastic Modeling and the Theory of Queues. Insurance: Mathematics and Economics 15. Thorin & N. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Aktuar. Thorin & N. Insurance : Mathematics and Economics 13. [381] W. 148164.E. Taqqu. [380] W. Wald (1947) Sequential Analysis. [379] M. Probab. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Willinger.
40. 117128.359 aggregate claims 103106.249. 226.160167.249250 integral equation 16 Lindley 143 renewal 64. 135. 110113. 245248. 308. 5796.259261. 316323 Bessel function 102. 7079. 7879.307312 compound Poisson model 4. 201 Brownian motion 3 . 97129. 138139.121129.6779.346349 383 .200201.135. 4851. 117127 corrected 121127 duality 1314. 97.9899.251280 heavy traffic 76. 239. 7475. 7179.137141.185187. 3032.301 central limit theorem 60 . 89. 3334. 205. 39.217.203.178184. 332333 Volterra 192194.203. 80 81. 207 heavytailed distribution 6. 227229. 2526. 180182. 301 Kronecker product.100. 119.226. 318319 Erlang distribution 7. 122.281.269. 14. 271274. 17. 218 Cox process 4. 111117. 360 excursion 155156.293294. 86. 15.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16. 278 gamma distribution 67.and sum 221.150. 341.314316.272. 189. 323 Coxian distribution 147. 283. 302303 diffusion approximation 17. 248 WienerHopf 144 interest rate 190. 2425. 12 CramerLundberg approximation 1617. 141144.4447.318320 change of measure 2630. 170173. 1819.328330. 17.86. 79. 37. 162164.287292. 361 diffusion 3.285292.299. 1415. 5. 3436. 9396.292293 Edgeworth expansion 113. 8283 hyperexponential distribution 7. 1112.182.Index adjustment coefficient 17. 217.228229. 91. 3839.242. 9293. 9496. 196201 inverse Gaussian distribution 76. 201214.
261264. 132133.304 process 2830.108. 113114. 4446. 42. 144.178182. 203204. 261264.350361 Poisson process Markovmodulated 12 periodic 12. 14.287. 234 matrixexponential distribution 240244 matrixexponentials 14. 7576.152160. 176185.146148.128129. 35.298299. 15. 65. 145187.297299.215250.160161. 16. 16. 52 53. 100. 257.288290. 175 light traffic 8183 Lindley integral equation 143 process 3334. 99. 86 periodicity 12. 203 Markov additive process 12.269271. 37. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 230.285287 queue 14 . 138. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 3947. 133. 295.275278.161. 134135.139141. 71. 41. 32.384 ladder heights 4756. 154. 6162. 142 likelihood ratio : see change of measure lognormal distribution 9.234. 134. 171. 141144. 59.336339 . 149. 178 modulation 12. 5758. 6970. 137139. nonlinear 155.339 large deviations 129. 108109. 108 life insurance 5. 213214. 260 Lundberg conjugation 6979 . 306316 Levy process 3.348 terminating 215216. 44.287291 INDEX matrix equation . 39. 267269 Panjer's recursion 320323 Pareto distribution 910. 96.161164. 269 PerronFrobenius theory 4142. 25. 7179. 162. 133. 35.227230. see also sensitivity analysis phasetype distribution 8. 271274. 38.349 350 perturbation 172173. 80. 106108. 3947.340350 multiplicative functional 2830. 304305 random walk 3336.134135. 38.174.234240.302.259261. 2730.240244.180.238. 157. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 25. 44. 112113. 3639.161.148. 227228.218221.336339 Laplace transform 15. 245 M/G/1 13. 179 NP approximation 318320 Palm distribution 5253.315 inequality 1718.201. 251.123. 9899.
31. 172173. 186187 renewal process 131. 253. 4950. 317318 semiMarkov 147. 174. 131144.INDEX 385 waiting time 141. 238 saddlepoint method 115117. 162. 332333 model 12. 87. 294296 shotnoise process 314 simulation 19. 213. 233. 335336 sensitivity analysis 8693. 251280 time change 4.279280 Rouche roots 158. 223226. 1819. 326330 Weibull distribution 9. 37.186. 292294. 11.154157. 260 WienerHopf theory 144. 60. 107. 240.273274. 168172 storage process 13. 5455. 260 reinsurance 8. 257. 280. 338 utility 324. 89. 251. 8386. 191192. 244. 123. 233234.314. 12. 186187 virtual: see workload rational Laplace transform 8. 327 . 160.359361 stochastic control x stochastic ordering 18.262263. see also matrixexponential distribution regenerative process 264 268. 251. 279280 subexponential distribution 11. 256258. 7475.244250. 307308. 152. 333334 regular variation 10. 281296 stable process 15. 147. 189214. 222.336339 workload 13. 141144. 229234. 331336 equation 64. 120 statistics x. 3032. 9693. 261264 reservedependent premiums 14. 177 timereversion 14. 146.
worldscientific. Some i (l I JL I J r of the topics are Lundberg's inequality.. P'i yfliother approximations (e. the ^W A l \ i l ' ''' CramerLundberg approximation. "This book is a must for anybody working in applied probability. It is a comprehensive treatment of the known results on ruin probabilities. for heavytailed claim size distributions).com 2779 he 9 "789810ll22293211 . exact solutions. Markovmodulation or periodicity. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory.g.Advanced Series on Statistical Science & Applied Probability . Special features of the book are the emphasis on change of measure techniques. y finite horizon ruin probabilities." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www.Vol.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. extensions of the classical compound Poisson model to allow f o r reservedependent premiums. 2 A I 11 JjVb l' i  i Yj .. I 1! Ruin Probabilities .
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.