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
. . . .. . . . 316 5 Principles for premium calculation . . .. . . . . . . . . . . . . . . . . .. . . . . . . 340 A4 Some linear algebra . . . . . . . . .. 297 2 Further applications of martingales . . . . . . . . . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. .. . . . . 350 Bibliography Index 363 383 . . . . . . . . . . . . . . . . 336 A3 Matrixexponentials . . . .. . . . .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . 290 5 Regenerative simulation . . . . . . . . . . 323 6 Reinsurance . . . . .. . . . . . . . . . . 344 AS Complements on phasetype distributions . . . . . . . . . . 271 6 Reservedependent premiums . . . . .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . .. . . . . .. . . . . . .. . . . . . . . . . . . . . .. . . . . . . 306 4 The distribution of the aggregate claims . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . . . . . . . . . . . . . . 304 3 Large deviations . . . . . . . . . . . . . . . . . . . The twobarrier ruin problem . . . . 259 3 The renewal model . . . . . . . . . . 285 3 Importance sampling via Lundberg conjugation . . . . . . .. . . . . . . . .. . . . . . 279 X Simulation methodology 281 1 Generalities . . . . . . . 264 5 Finitehorizon ruin probabilities . . . . .. . . . . . .. . . . . . . . . . . . . 326 Appendix 331 Al Renewal theory . . . . . . . 292 6 Sensitivity analysis . . . . . 261 4 Models with dependent input . . 251 2 The compound Poisson model . 287 4 Importance sampling for the finite horizon case . 331 A2 WienerHopf factorization ... . . . . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .
This page is intentionally left blank .
The course was never realized. 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!).Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. Thus. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. and my belief was that this could be done rather quickly. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. it is not by intention. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. In particular. if the formulations occasionally give a different impression. which has in particular removed one of the standard criticisms of the area. that it can only say something about very simple models and questions. However. and the result is now that the book is much more related to my own research than the initial outline. 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. Since I was to produce some handouts for the students anyway. and has been an active area of research from the days of Lundberg all the way up to today. this applies to longrange dependence which is intensely studied in the neighboring ix . Apart from these remarks. But the pace was much slower than expected. the idea was close to expand these to a short book on the subject. University of Copenhagen. and other projects absorbed my interest. I have deliberately stayed away from discussing the practical relevance of the theory. and the series editor Ole BarndorffNielsen for their patience. As an excuse: many of these projects were related to the book. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. the book is basically mathematical in its flavour. It has obviously not been possible to cover all subareas.
It is obvious that such a system involves a number of inconsistencies and omissions. In the classical setting of CramerLundberg models. The present book is in between these two possibilities.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . More recently.4a.g. the standard stochastic control setting of diffusion models has been considered. Hojgaard & Taksar [206].45.2. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. Good luck! I have tried to be fairly exhaustive in citing references close to the text. IV. http:// www. 111. One is by model. IV. e.15. Willinger et al. an area which is becoming increasingly important. For a second reading.2. Here is a suggestion on how to get started with the book.5. VII. I regret that due to time constraints.13.se Lund February 2000 Soren Asmussen . Asmussen.g. Another interesting area which is not covered is dynamic control. I intend to keep a list of misprints and remarks posted on my web page. the first part of 11. The rest is up to your specific interests. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. for the effects on tail probabilities. 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). some papers not cited in the text but judged to be of interest are included in the Bibliography.1. [381]).x PREFACE field of queueing theory. IX.lth.g. IV.6 (to understand the PollaczeckKhinchine formula in 111. for which I apologize to the reader and the authors of the many papers who ought to have been on the list.13.13 and XI.89.13 and IX. see in particular Michna [259]. The main motivation comes from statistical data for network traffic (e. another by method. VIII. In addition. VII. see e. Concerning ruin probabilities.lth. VI. 111.3. A book like this can be organized in many ways. Resnick & Samorodnitsky [303] and references therein. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. X.2 more properly).maths . Finally. incorporate 11. For a brief orientation. it has not been possible to incorporate more numerical examples than the few there are. read Chapter I. see also Schmidli [325] and the references in Asmussen & Taksar [52].14. Hojgaard & Taksar [35] and Paulsen & Gjessing [284].
Fig.6 by my 1999 simulation class in Lund. More substantial remarks.6. 111 .6 is reprinted from Asmussen & Schmidt [49] and parts of IX.8 . 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . of which there are not many at this stage .1 and X.2 by Rafal Kulik .5 from Asmussen [21] with permission from CRC Press. . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . Section VIII. IV. as well as some additional references continue to be at the web page.4 from Asmussen. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX.1 by Bjarne Hojgaard and the table in Example 111.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. not least the more complicated ones. supported by Center for Mathematical Physics and Stochastics (MaPhySto). many of which were pointed out by Hanspeter Schmidli . 5. Parts of II. A number of other figures were supplied by Christian Geisler Asmussen . Aarhus. Fig. were produced by Lone Juul Hansen .PREFACE xi The second printing differs from the first only by minor corrections. Section VII . Parts of X. 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers.
This page is intentionally left blank .
T) = P inf Rt < 0 I .2) (O<t<T Ro=ul. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. (1.Rt. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. results and topics to be studied in the rest of the book. (1. A risk reserve process { Rt}t>o.Chapter I Introduction 1 The risk process In this chapter . respectively. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t.3) sup St. For mathematical purposes.4) O<t<oo O<t<T 1 . we introduce some general notation and terminology. They are the main topics of study of the present book. MT = sup St. and give a very brief summary of some of the models.1) We also refer to t/) ( u) and 0(u. We denote throughout the initial reserve by u = Ro. as defined in broad terms . T) as ruin probabilities with infinite horizon and finite horizon . is a model for the time evolution of the reserves of an insurance company.i(u. M = (1. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . (1.
1. Figure 1. and T1 is the time of the first claim. Thus. . Putting things together. (1.6) Sofar we have not imposed any assumptions on the risk reserve process. the ruin probabilities can then alternatively be written as .. That is. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un.2 CHAPTER I.pt. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. • Premiums flow in at rate p.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig.E Uk. We denote the interarrival times of claims by T2. the time of arrival of the nth claim is an = T1 + • • • + Tn. St = E Uk .5) i.T) = F (MT > u) = P(r(u) < T). respectively. t] is finite. we see that Nt Nt Rt = u + pt .1. (1. (1.b(u) = P (r(u) < oo) = P(M > u).1 . However. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. per unit time.. say. T3. the number Nt of arrivals in [0.i(u.
However. and the basic ruin probabilities are derived in XI. and in fact: Proposition 1.. We shall not deal with this case either.8) holds.1.8) The interpretation of p is as the average amount of claim per unit time. that the insurance company should try to ensure 77 > 0.b(u) = 1 for all u.. Thus.s.1 the slope of {Rt} should depend also on the level). one could well replace Rt by Rtnr(u) or RtA.(. then M = oo a. since any modeling involves some approximative assumptions. a basic references is Gerber [127]. immaterial. It would appear obvious.s. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. For the purpose of studying ruin probabilities this distinction is. allowing a countable infinity of jumps on Fig. however.1.e. (1. We study this case in Ch. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. then M < oo a. not discuss whether this actually corresponds to practice. and hence .1. 1.20%. of course. . however. If 77 < 0. If 77 > 0.1 Assume that (1. 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. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. We shall discuss Brownian motion somewhat in Chapter IV. though many results are straightforward to generalize from the compound Poisson model. 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. on Fig. and hence O(u) < 1 for all sufficiently large u. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. • Brownian motion or more general diffusions. but as an approximation to the risk process rather than as a model of intrinsic merit. 1. say 10% .) V 0. t * oo. we shall. one may well argue that Brownian motion in itself could be a reasonable model. for example. VII.
This case is referred to as the mixed Poisson process. Here it is easy to see that p = . However. INTRODUCTION Proof It follows from (1. are i.i. tb(u) = 1 for all u holds also when rl = 0. .T) = i. _ St __ k =1 Uk pt a4. (1. and independent of {Nt}. namely.8) that F N. then this limit is > 0 which implies St a$ oo and hence M = oo a.i. Thus p may well be random for such processes. and that . corresponding to the Pdlya process. 0(u. it is not too difficult to show that p as defined by (1.. t t p  p' t ^ oo. .s.s.b(u) < 1 for all u when rl > 0. k=1 (1.6EU (on the average. Proposition 1. If U1. If u oo.v. we obtain typically a somewhat stronger conclusion. However.. (1.3 Assume p 54 1 and define Rt = Rt1p.d. The simplest example is 3(t) = V where V is a r . not all models considered in the literature have this feature: Example 1.T) for {Rt} is given by V)(u) = t/i (u).Q (say) and U1.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. are i.10) hold with p constant. in connection with risk processes in a Markovian or periodic environment (Chapter VI).8).. if {(3(t)} is nonergodic.d.i(u.10) is a property which we will typically encounter. . (1. 0 We shall only encounter a few instances of a Cox process. U2. with the most notable special case being V having a Gamma distribution.. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model.oo t 0 J (provided the limit exists). namely that M = oo a. M < oo a.11) .Tp).8) is given by ^t p = EU • lim it (3(s) ds t. Nt)}. zP(u . where {Nt} is a Poisson process with rate . and independent of {(0(t). U2. and here (1.10) Again. then similarly limSt/t < 0. rl > 0.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t. St In concrete models. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). If 77 < 0.s.. this needs to be verified in each separate case.4 CHAPTER I.
in a number of models. the role of the result is to justify to take p = 1. [330]. Cox processes are treated extensively in Grandell [171]. Insurance: Mathematics and Economics. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. De Vylder [110]. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. Since { Rt } has premium rate 1. Some of the main general ideas were laid down by Lundberg [250]. Grandell [171]. We roughly classify these into two groups . the recent survey by Grandell [173] and references therein. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. see also Chapter XI. Notes and references The study of ruin probabilities. For mixed Poisson processes and Polya processes. In the even more general area of nonlife insurance mathematics. Heilmann [191]. Daykin. Pentikainen & Pesonen [101]. Gerber [159]) has a rather different flavour. Buhlmann [82]. U2.. Taylor [364]. Gerber [157]. the research literature is often published in journals like Astin Bulletin . [76]. Schmidt & Teugels [307] and Seal [326]. [101].. Embrechts et al. The Swedish school was pioneering not only in risk theory. was largely initiated in Sweden in the first half of the century. [134]. Note that when p = 1. see e . Note that life insurance (e. Straub [353]. Daykin et al. the claim arrivals are Poisson or renewal at the same time). some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. 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. Hipp & Michel [198]. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). another important early Swedish work is Tacklind [373]. Some main later textbooks are (in alphabetical order) Buhlmann [82]. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91].g. Schmidli. but in probability and applied probability as a whole. Segerdahl [334] and Philipson [289]. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. Some early surveys are given in Cramer [91]. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. in particular. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. often referred to as collective risk theory or just risk theory. we shall be able to identify p with the traffic intensity of an associated queue. Rolski.. and we do not get near to the topic anywhere in this book. Sundt [354]. lighttailed distributions (sometimes the term .2..g. the assumption > 0 is equivalent to p < 1. Besides in standard journals in probability and applied probability.
f. Example 2 . Equivalently. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . a fact which turns out to contain considerable information.u at the time of ruin given r(u) is again exponential u with rate 8. the m. 2a Lighttailed distributions Example 2.g. For example in the compound Poisson model.e. B is heavytailed if b[s] = oo for all s > 0. Here lighttailed means that the tail B(x) = 1 . then the conditional distribution of X . INTRODUCTION 'Cramertype conditions' is used).x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). As in a number of other applied probability areas.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. The crucial feature is the lack of memory: if X is exponential with rate 6. but different more restrictive definitions are often used: subexponential.2) = 0.B(x) satisfies B(x) = O(e8x) for some s > 0.3) .1) The parameter 6 is referred to as the rate or the intensity. for the compound Poisson model with exponential claim sizes the ruin probability . where B(bo. 6 has density r(p)xPleax b(x) P and m. if 1 °O AB Jbos x B(dx) > 0. i.g. B[s] is finite for some s > 0.8. On the more heuristical side.f. s<8. the exponential distribution is by far the simplest to deal with in risk theory as well.O(u) can be found in closed form. and can also be interpreted as the (constant) failure rate b(x)/B(x).2 and /LB is the mean of B.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. In contrast. 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'. (2. and heavytailed distributions. P B[s]= (8Is ) . In particular. regularly varying (see below) or even regularly varying with infinite variance.6 CHAPTER I.
1) (or the 1/pth root if p < 1). . An important property of the hyperexponential distribution is that its s. is > 1.d. u Example 2 . CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. p).y i=1 where >i ai = 1. we develop computationally tractable results mainly for the Erlang case (p = 1. if p is integer and X has the gamma distribution p.1 Poisson events in [0.v. by Grandell & Segerdahl [175] and Thorin [369]. 0 < ai < 1. x] so that B(x) = r` e. . P b(x) = r` aibiea.) VarX1 (EX )2 p is < 1 for p > 1. .. In particular... An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p .). p) °° where r (x.c.c... i = 1.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. are i. 0. Ruin probabilities for the general case has been studied. p. B(x) = r(p) Asymptotically.ate (b2 ): L• i=o In the present text.2) can be considered as the pth power of the exponential density (2.i. X2.v. The exact form of the tail B(x) is given by the incomplete Gamma function r(x. or just the Erlang(p) distribution. u .. one has r(bx. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. In particular. > 1 for p < 1 and = 1 for p = 1 (the exponential case). the Gamma density (2. 2. the squared coefficient of variation (s. p) = J tPletdt.. and exponential with rate d. among others.2. then X v Xl + • • • + X. where X1. This special case is referred to as the Erlang distribution with p stages.
g. Example 2 . INTRODUCTION Example 2 .6. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. See XI. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities. .8) are realvalued. 1)' is the column vector with 1 at all entries. We give a more comprehensive treatment in VIII.d. T) is called the representation. 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.f. (or.e.1 and defer further details to u Chapter VIII. but the current trend in applied probability is to restrict attention to the class of phasetype distributions.(2. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. Example 2 . a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree.f.6. of which one is absorbing and the rest transient. q2 q3 (2.7) are possibly complexvalued but the parameters in (2.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i.. resp.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. This class of distributions is popular in older literature on both risk theory and queues. the Erlang and the hyperexponential distributions. Important special cases are the exponential. The couple (a.8 CHAPTER I. which is slightly smaller but more amenable to probabilistic reasoning. T) or sometimes the triple (E. there exists a xo < oo such that B(x) = 0 for x > xo.8) j=1 j=1 j=1 where the parameters in (2.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . B(x) = aeTxe where t = Te and e = (1 .xo)+ is covered by the reinsurer). The density and c. However. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx.. a. equivalently. then the claim size which is relevant from the point of view of the insurance company itself is U A xo rather than U u (the excess (U . The parameters of a phasetype distribution is the set E of transient states. We give some theory for matrixu exponential distribution in VIII. are b(x) = aeTxt.
one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0.N(0. the mean u is eµ+a /2 and the second moment is e2µ+2o2.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed.1).7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. However.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2. Example 2 .u l b(x) = d dx or J ax lor 1 exp Asymptotically.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. u Example 2 .9) which is heavytailed when 0 < r < I. Here failure rates b(x) = b(x)/B(x) play an important role. p is defined as the distribution of ev where V . or equivalently as the distribution of a°U+µ where U .10) The loinormal distribution has moments of all orders. a2). b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . 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). a)/A)a+1' x > a. x < a. It follows that the density is 't (1ogX . we obtain the Weibull distribution B(x) = eCx'. and then b(x) = 0. b(x) = crx''le`xr. the tail is B (x ) 2 x. the exponential distribution representing the simplest example since here b(x) is constant.13) u . All moments are finite. (2.2. Writing c = d/r. In particular.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2.p a 1 (2. There are various variants of the definition around. (2.pl = 1 W (logx .1.N(p. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. (2.
In general.(1 + 2x + 2x2)e2x) p = 2 (2. another standard example is (log x)').15) x2 + 16x3 ) a3x/2) 3 (1 .x6+lr(p) (2. For p = 1. examples of distributions with regularly varying tails are the Pareto distribution (2. satisfies L(xt)/L(x) 4 1. u . the density is { 3 (1 . 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).'s of the form YX.v.(1 + Zx + $ p = 3. where Y is Pareto distributed with a = (p .16) 11 Example 2. 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) .13). i. B(x) = O(xP). Choudhury & Whitt [1] as the class of distributions of r.12) (here L (x) * 1) and ( 2.e. x 4 oo (any L having a limit in (0. The density is 8p(log x)pi b(x) . the loggamma distribution is a Pareto distribution. (2. { s () 1s+3s29s3log(1+2s I p=3.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p.10 CHAPTER I. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. oo) is slowly varying .17) where L (x) is slowly varying.L( x ). x + 00. (2.2). in particular.1)/p. Thus. in particular. The simplest examples correspond to p small and integervalued. INTRODUCTION Example 2. u Example 2 . 6 is defined as the distribution of et' where V has the gamma density (2.14) The pth moment is finite if p < 5 and infinite if p > 5.
18) B(x) It can be proved (see IX. 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. Also. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. Thus. and so is the Weibull distribution with 0 < r < 1. though the proof of this is nontrivial.. We return to a closer study in IX.. The reason is in part mathematical since this model is the easiest to analyze. (2. for example the lognormal distribution is subexponential (but not regularly varying). one may argue that this difficulty is not resticted to ruin probability theory alone. Namely. From a practical point of view.3. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. the knowledge of the claim size distribution will typically be based upon statistical data.1) that any distribution with a regularly varying tail is subexponential. When studying ruin probabilities..13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available.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. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. THE ARRIVAL PROCESS 11 Example 2.1. we may know that such a process (with a covariance function estimated from data) is a reasonable description of the behaviour of the system under study in typical conditions. which each have a ( timehomogeneous) small rate of experiencing a .. Similar discussion applies to the distribution of the accumulated claims (XI. U2. the claim size distribution represents of course only one aspect (though a major one). the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. this phenomenon represents one of the true controversies of the area. but the model also admits a natural interpretation : a large portfolio of insurance holders . and based upon such information it seems questionable to extrapolate to tail behaviour. However. 3 The arrival process For the purpose of modeling a risk process .
i. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance .. A more appealing way to allow for inhomogeneity is by means of an intensity . e. the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 .. This applies also to the case where the claim size distribution depends on the time of the year or .3(t) fluctuating over time. so that . which facilitate the analysis. see 11. In order to prove reasonably substantial and interesting results . found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives .. T2. This model . However . where {/3 (t)}too is an arbitrary stochastic process . in Chapter VII). in particular to allow for certain inhomogeneities. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications.12 CHAPTER I. are i. Cox processes are. it may be used in a purely descriptive way when it is empirically observed that the claim arrivals are more bursty than allowed for by the simple Poisson process. too general and one neeed to specialize to more concrete assumptions . To the author 's knowledge .g. but with a general not necessarily exponential distribution ). getting away from the simple Poisson process seems a crucial step in making the model more realistic.d. The difficulty in such an approach lies in that it may be difficult or even impossible to imbed such a distribution into the continuous setup of {Nt } evolving over time . however. in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). gives rise to an arrival process which is very close to a Poisson process.6. we study this case in VI . the negative binomial distribution. when Jt = i. the periodic and the Markov modulated models also have attractive features . such that 8(t) = . The compound Poisson model is studied in detail in Chapters III.. Nevertheless . epidemics in life insurance etc. with the extension to premiums depending on the reserve. The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. has some mathematically appealing random walk features . The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ).8 (t) is a periodic function of t. and also that the ruin problem may be hard to analyze . not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. 5. IV (and. Another one is Cox processes. INTRODUCTION claim . Mathematically. In others. An obvious example is 3(t) depending on the time of the year (the season). I.e. Historically. with a common term {Nt} is a Markovmodulated Poisson process .(3. to be studied in Chapter V. Some of them have concentrated on the marginal distribution of NT (say T = one year ).
point processes and so on. More generally. R = p(R) in between jumps. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. 4 A summary of main results and methods 4a Duality with other applied probability models Risk theory may be viewed as one of many applied probability areas. 0(u) = P(V > u).T) = P(VT > u). and here (4. however.1) where V is the limit in distribution of Vt as t + oo.v. Thus. Markovmodulated or periodic can be related to queues with similar characteristics. interacting particle systems. that quite often the emphasis is on computing expected values like EV. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. ruin probabilities for risk processes with an input process which is renewal. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. time series and Gaussian processes. this gives only f0 O°i (u)du which is of limited . queueing theory. stochastic differential equations. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. In the setting of (4.0 (u. reliability.'s like V is available. and which seems well motivated from a practical point of view as well. (4. this amounts to Vo having the stationary distribution of {Vt}). others being branching processes.1) permitting to translate freely between risk theory and the queueing/storage setting. 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 . others are quite different. it is desirable to have a set of formulas like (4.1). A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. dam/storage processes. methods or modeling ideas developed in one area often has relevance for the other one as well. Similarly. In fact. Some of these have a certain resemblance in flavour and methodology. It should be noted. The ones which appear most related to risk theory are queueing theory and dam/storage processes. Mathematically.4. and a lot of information on steadystate r.1) holds as well provided the risk process has a premium rule depending on the reserve. with Poisson arrivals and constant release rule p(x) = 1.6) . genetics models. stochastic geometry. and the limit t 4 oo is the steadystate limit. The study of the steady state is by far the most dominant topic of queueing and storage theory. extreme value theory.
g.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 ). • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. much of the study of finite horizon problems (often referred to as transient behaviour) in queueing theory deals with busy period analysis which has no interpretation in risk theory at all.14 CHAPTER I.p(y) y^ Jo p(x) can be written in closed form. e . see Corollary VII. The cases where this is possible are basically the following for the infinite horizon ruin probability 0(u): • The compound Poisson model with constant premium rate p = 1 and exponential claim size distribution B. the two areas. see Boxma & Cohen [74] and Abate & Whitt [3].'s like the environmental process {Jt} in a Markovmodulated setting.v. 4b Exact solutions Of course .6. A prototype of the duality results in this book is Theorem 11. B(x) = ebx.. Example VIII. the functions w x f d 1 exdx () . which gives a sample path version of (4. Similarly. see Corollary III. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u).2). The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. Vi(u. . which can be expanded into a sum of exponential terms by diagonalization (see. 3. Here O(u) = peryu where 3 is the arrival intensity. • The compound Poisson model with some rather special heavytailed claim size distributions.1). 3.3. have to some extent a different flavour. Thus . though overlapping. p = 0/8 and y = 8 . The infinite horizon (steady state ) case is covered by letting T oo.3.3.3. INTRODUCTION intrinsic interest .1 .8. • The compound Poisson model with a claim size distribution degenerate at one point. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .1. The fact that Theorem H. as is typically the case. Here ?P(u) is explicit provided that .T).
a2 (x): Ip (u) = where S(u) = f °O exp {. O(u. A notable fact ( see again XI.Lef$er function. 1).u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. However. (u. Abate & Whitt [2]. . relevant references are Grubel [179]. Embrechts. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag.b(u)du .7. • An astable Levy process with drift . but are somewhat out of the mainstream of the area . Here are some of the main approaches: Laplace transform inversion Often. Given this can be done.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model.ff 2µ(y)/a2(y) dy} dx . esuTb( u. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. 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.f f 2µ(y)/a2(y) dy} dx  (4. it is easier to find the Laplace transforms = f e8 . the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. T) themselves. T). 191). Also Brownian models or certain skip free random walks lead to explicit solutions (see XI . T) can then be calculated numerically by some method for transform inversion.S(u) 1S(oo) f °D exp {. [s. We don't discuss Laplace transform inversion much.2) is the natural scale. Ab(u).1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). the formulas ( IV.4. f {eXp U LX 2. For the finite horizon ruin probability 0(u. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). 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 . see VIII.
1) and y > 0 is the solution of the Lundberg equation (4.3) in the compound Poisson model which is an integral equation of Volterra type. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). dt] most often leads to equations involving both differential and integral terms. In the compound Poisson model with p = 1. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). u * oo. (4.g.and integral equations The idea is here to express 'O(u) or '(u. T) as the solution to a differential. and in particular the naive idea of conditioning upon process behaviour in [0.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential).Ce"u.or integral equation. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m.7.f. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. Differential. However. U is explicit in terms of the model parameters. either as the iterative solution of a fixpoint problem or by finding the diagonal form in terms of the complex roots to certain transcendental equations. it states that i/i(u) . which can equivalently be written as f3 [7] = 1 +13 .) B[s].p)/(13B'[ry] .3) where C = (1 . see VIII.4) 00['Y]1)'Y = 0. and carry out the solution by some standard numerical method. 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. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. One example where this is feasible is the renewal equation for tl'(u) (Corollary III. and in quite a few cases (Chapter VIII). .3.
Large claims approximations In order for the CramerLundberg approximation to be valid. Diffusion approximations are easy to calculate. u > oo. In the case of heavytailed distributions. but typically not very precise in their first naive implementation. the claim size distribution should have an exponentially decreasing tail B(x). 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 exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. J B dx. T).6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. in such cases the evaluation of C is more cumbersome. See Chapter IX. In fact. For example. when the claim size distribution is of phasetype. This list of approximations does by no means exhaust the topic. 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.7 and IV. other approaches are thus required. for the compound Poisson model ^(u) p pu In fact . (4. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. a Markovian environment or periodically varying parameters. It has generalizations to the models with renewal arrivals.4. T) for large u are available in most of the models we discuss. some further possibilities are surveyed in 111 . in some cases the results are even more complete than for light tails.2. often for all u > 0 and not just for large u. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. However. corrected diffusion approximations (see IV. However. Approximations for O(u) as well as for 1(u.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. incorporating correction terms may change the picture dramatically. In particular.
.) at various places and in various settings.18 CHAPTER I. When comparing different risk models. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. they have however to be estimated from data. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. For example. However. of being somewhat easier to generalize beyond the compound Poisson setting. which is a standard statistical problem since the claim sizes Ui.i. in the compound Poisson model. as a general rule. T]. In practice. more importantly. it is a general principle that adding random variation to a model increases the risk. This procedure in itself is fairly straightforward.3).. empirical evidence shows that the general principle holds in a broad variety of settings. UNT may be viewed as an interpolation in or smoothing of the histogram). the difficulty comes in when drawing inference about the ruin probabilities. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. . obtained say by observing the risk process in [0. . .d.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.8.k (U(`) . UNT are i. For example. 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.g. one expects a model with a deterministic claim size distribution B. However. The standard suggestion is to observe that the mean residual life E[U . We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. fitting a parametric model to U1. .U(k)) i =k+ i . This is proved for the compound Poisson model in 111... lower bounds etc. INTRODUCTION Compared to the CramerLundberg approximation (4. How do we produce a confidence interval? And. and to plot the empirical mean residual life 1 N . it has the advantage of not involving approximations and also. given NT. e. though not too many precise mathematical results have been obtained. . say degenerate at m. 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.
< U(N) are the order statistics based upon N i. The infinite horizon case presents a difficulty. However. in all other chapters than VI where we just write . . Truncation to a finite horizon has been used.3).e.d.i.3) and Section VI. We look at a variety of such methods in Chapter X.5.3 (or just VI. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. because it appears to require an infinitely long simulation. Klnppelberg & Mikosch [134].. CONVENTIONS 19 as function of U(k).. in this book referred to as [APQ])..(5. A main problem is that ruin is typically a rare event (i. respectively. formula (5. but is not very satisfying. Thus Proposition 4. and in fact methods from that area can often be used in risk theory as well .. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics.2. and look at them to see whether they exhibit the expected behaviour or some surprises come up.v's) which can be generated by simulation. reference [14]. The chapter number is specified only when it is not the current one. UN. and of course the method is relevant in risk theory as well. formula VI. where U(1) < . . good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths.4.2. (or a functional of the expectation of a set of r. this is a straightforward way to estimate finite horizon ruin probabilities. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987.v. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. See further Embrechts. For example. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. claims U1. to observe whether one or the other limiting behaviour is apparent in the tail. Still..3) or Section 3 of Chapter VI are referred to as Proposition VI. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available.
E expectation. i.f. and for a defective probability distribution IIGII < 1. n! 27r nn+1/2en. EX2/(EX)2. or a more precise one like eh .v. References like Proposition A. say a heuristic approxi1 + h + h2/2.e.ceax. infinitely often l. right hand side (of equation) r.3) or Section 3. i. left hand side (of equation) m.d. if B(x) .g.t. random variable s. with probability Mathematical notation P probability.g. as for typical claim size distributions. A different type of asymptotics: less precise.29) refer to the Appendix. Abbreviations c. cumulant generating function.f. squared coefficient of variation.s. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example.d.2. moment generating function.g.f. for a probability distribution IIGII = 1.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c.g.B(x) = P(X > x) of B. cumulative distribution function P(X < x) c.i. . E. B(x) = P(X < x) = fx. (A. B[s] the m. w. In particular. oo).f.s.v. n i oo. B(x) the tail 1 .4.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.c.r. then for 1s < 5). The Laplace transform is b[s].h.o. formula (5. B is concentrated on [0. B(dy). mation. independent identically distributed i.20 CHAPTER L INTRODUCTION Proposition 4. r.h. IIGII the total mass (variation ) of a (signed ) measure G . If. log E[s] where b[s] is the m. h + 0.f. . see under b[s] below. with respect to w.d.p.f. (moment generating function) fm e82B(dx) of the distribution B.g.
Xt_ the left limit limstt X8f i. 7r. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. E[X. .. Matrices and vectors are denoted by bold letters. Notation like f3i and 3(t) in Chapter VI has a similar . i. A. Usually. 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. 0 marks the end of a proof. and column vectors have lowercase Roman letters like t.e. In the Frenchinspired literature. oo) the space of Rvalued functions which are rightcontionuous and have left limits. row vectors have lowercase Greek letters like a. i. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi).. N(it.e. Usually. . of numbers. Unless otherwise stated. the ith unit row vector is e'i. the ith entry is 1 and all other 0. a. xa.5. 21 D [0. a2) the normal distribution with mean p and variance oa2. intensity interpretation. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. Thus.e. though slightly more complicated. . F o r a given set x1. I(A) the indicator function of the event A. R(s) the real part of a complex number s. an example or a remark.oi denotes the column vector with the xi as components Special notation for risk processes /3 the arrival intensity (when the arrival process is Poisson). the value just before t.A] means E[XI(A)]. (the dimension is usually clear from the context and left unspecified in the notation). the processes we consider are piecewise continuous. matrices have uppercase Roman or Greek letters like T. only have finitely many jumps in each finite interval. 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. all stochastic processes considered in this book are assumed to have sample paths in this space.
p the net amount /3pB of claims per unit time. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. e. J the rate parameter of B for the exponential case B(x) = eby. . interpretation. FL. Notation like BE and B(t) in Chapter VI has a similar. ry The adjustment coefficient. I.5. cf.1.g. I.1.5. cf. though slightly more complicated. 'q the safety loading . or quantities with a similar time average interpretation. 111. cf.22 CHAPTER L INTRODUCTION B the claim size distribution. VI.
More precisely. 5) are. The topic is. somewhat more advanced than in the rest of the book. The duality results in Section 3 (and. Due to the generality of the theory. the level of the exposition is. The reader should therefore observe that it is possible to skip most of the chapter. 23 . the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. not crucial for the rest of the book. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model.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. however. When encountered for the first time in connection with the compound Poisson model in Chapter III. in particular at a first reading of the book. however. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. All results are proved elsewhere . however. The general theory is. The likelihood ratio approach in Section 2 is basic for most of the models under study. in most cases via likelihood ratio arguments. strictly speaking. in part. used in Chapter VI on risk processes in a Markovian (or periodic) environment. Sections 4. a parallel selfcontained treatment is given of the facts needed there. Sections 4.
1 Martingales We consider the claim surplus process {St} of a general risk process.5 can be skipped.24 CHAPTER II. As usual. and in the limit (1.1 Assume that (a) for some ry > 0. StUit. We get 1 = Ee7So = E e'Y S(.2 Consider the compound Poisson model with Poisson arrival rate . The more general Theorem 6. T(u) > T] . Let e(u) = ST(u) . SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. f1 .)AT = E [e7ST(°). claim size distribution B and p = . and the ruin probabilities are ip(u) = P (T(u) < oo). using r(u) A T invokes no problems because r(u) A T is bounded by T).2) takes the form 1 = E [e'ys(). the time to ruin r(u) is inf It > 0 : St > u}. however.2) As T > oo. (1. 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. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem .u denote the overshoot.1 is basic for the study of the compound Poisson model in Chapter III. Thus N. V) (u.0.(u). Proposition 1. T(u) < oo] + 0 = eryuE [e7Vu). (b) St a$ oo on {T(u) = oo}.. T) = P(T(u) < T). {e'YS° }t>0 is a martingale. T(u) < T] + E [eryST .QµB < 1.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. Example 1 . the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}).
5 For the compound Poisson model with B exponential. and thus by the memoryless property of the exponential distribution .2 and drift p > 0. B(x) _ edx.= e"(') where K(a) = .1. A simple calculation (see Proposition III./3. Then {St } is Brownian motion with variance constant o2 and drift p < 0.Q(B[a] .r" where y = S .3 Assume that {Rt} is Brownian motion with variance constant o. and (b) follows from p < 1 and the law of large numbers (see Proposition III. Under the conditions of Proposi Proof Just note that C(u) > 0.a.Q and the U. the martingale property now follows just as in Example 1. From this it is readily seen (see III.u + x is again just exponential with rate S.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists. The available information on this jump is that the distribution given r(u) = t and S.2.1) .1 is satisfied. condition (a) of Proposition 1. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where . By standard formulas for the m. the conditions of Proposition 1. MARTINGALES 25 where {Nt} is a Poisson process with rate .(„)_ = x is that of a claim size U given U > u .ap. the ruin probability is O(u) = pe. Eeas° = e"(°) where n(a) = a2a2/2 .1) . and thus Ee7s° = 1. Proof Since c(a) = /3 (B[a] .a = a . of the normal distribution.4 (LUNDBERG ' S INEQUALITY ) tion 1 . Thus. the conditional distribution of the overshoot e(u) = U . u Corollary 1.x.d.a it is immediately seen that y = S . u Corollary 1. and p =.Ft = a(S" : v < t).1) shows that Eels. and thus Ee'rs° = 1. Now at the time r(u) of ruin {St} upcrosses level u by making a jump . Since {St} has stationary independent increments.3/6 < 1. with common distribution B (and independent of {Nt}).1. Thus.g.2(c)). 1.f. are i. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx .i.Q. Since {St} has stationary independent increments. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.1. O(u ) < e7". Example 1 ..1 are satisfied.
Two such processes may be represented by probability measures F. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. Example 2 . But if a $ ^ . Grandell [171]. F). P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3.6 N S = { lim Nt I t +00 t gJ are both in F. Delbaen & Haezendonck [103] and Schmidli [320]. A E Ft. oo). A somewhat similar u argument gives singularity when B $ B. on (DE. More recent references are Dassios & Embrechts [98]. . Grandell & Schmidli [131]. Proof Just note that ^(u) = 0 by continuity of Brownian motion.1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes.F). and by the law of large numbers for the Poisson process . [172]. 0 and claim size distributions B. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. Theorem 2. However. B. I.. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. the parameters of the two processes can be reconstructed from a single infinite path. P are then singular (concentrated on two disjoint measurable sets). and F. then S and S are disjoint . A]. The number Nt F) of jumps > e before time t is a (measurable) r. cf. P on (DE. (2. Thus the sets S = I tlim +oot t =. F(S) = P(S) = 1.e. which we equip with the natural filtration {. then z/'(u) = e7" where 'y = 21A/a2. Embrechts.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. and is further exploited in his book [157]. hence so is Nt = limfyo N2`i.1 Let F. as shown by the following example this setup is too restrictive: typically'.3 below).Ft}too and the Borel afield F.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist.v. 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.
.F8] = L8 and the martingale property. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G).2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. if for some probability measure P and some {. G C {T < oo}. By the martingale property. . This proves (i). If r is a stopping time and G E PT.2(i). P) such that LLt = 1. Lets < t. Proof Under the assumptions of (i).Y) such that P(A) = Pt(A).3 Let {Lt}.r. we have E [ LTIFT]1 = LT on {T < T}.e. Ft).Pt)) The following result gives the connection to martingales. _. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. A] = E[Lt. A].r. 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 . Proposition 2. A E Ft . ({. The truth of this for all A E Y. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. that the restriction of P to (DE. A E F. F) such that ELt = 1. F the Borel o•field and P a given probability measure on (DE.F such that (2. Then Ft (A) = E[Lt.F).t. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE..Tt) is absolutely continuous w.r. then {Lt} is a nonnegative martingale w. ELt = 1 follows by taking A = DE in (2. the restriction of P to (DE. 1 J (2. Then P(A) = E[Lt.1) holds. then there exists a unique probability measure Pon .t. define P by Pt (A) = E[Lt.2 Let {Ft}t>o be the natural filtration on DE. P be as in Proposition 2. ({Ft} . Conversely. A E F8.1) and nonnegativity by letting A = {Lt < 0}.2.Ft}. Hence E [_ . (i) If {Lt}t> o is a nonnegative martingale w. Finally. (ii) Conversely. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8.t. then { 1 P(G) = EG .1) holds. A]. Lt < 0] can only be nonnegative if P(A) = 0. using the martingale property in the fourth step.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). implies that E[LtI.Pt}adapted process {Lt}t>o (2. G ] = E [LT .
both sides are increasing in T. in continuous time (the discrete time case is parallel but slightly simpler). First we ask when the Markov property is preserved. say. we have F(G) = V )(u). each F. we need the concept of a multiplicative functional. and letting T * oo.t. 1 Since everything is non negative.r. and this problem will now be studied. applying (2.4 Under condition (a) of Proposition 1. In the context of ruin probabilities. The crucial step is to obtain information on the process evolving according to F. St). 5) for processes with some randomwalklike structure. Consider a (timehomogeneous) Markov process {Xt} with state space E. Now just rewrite the r.s.1.1) in Proposition 1. Xt = St. SOME GENERAL TOOLS AND RESULTS In the general case .4) Proof Letting G = {r(u) < oo}. .h.2) by noting that 1 = ersr(„) = e1'ue7Ou). of (2. u From Theorem 2. the natural filtration {. (2. where {Rt} is the risk reserve process.t.1: Corollary 2. For the definition. T(u) < oo]. one would typically have Xt = Rt.r.2) follows by monotone convergence. is nonnegative and has Ey Lt = 1 for all x.28 CHAPTER II. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. is Markov w. we assume for simplicity that {Xt} has Dpaths.1). A change of measure is performed by finding a process {Lt} which is a martingale w. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1.4) compared to (1. first in the Markov case and next (Sections 4.Ft} . Rt) or Xt = (Jt.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 ... To this end. Lr(u) 11 The advantage of (2.O(u) = eryuE[e'YC(u). {St} = {u . t.Gn {r <T} . Xt = (Jt.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). (2.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian.
o 9t = V.(Xtitl) with all t(i) < t + s. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].YB] for any Ftmeasurable r.(A) = Ex [Lt.v. Zt. 0 . (2. Indeed. t.v.7). The converse follows since the class of r. By definition of Px.v.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any . Proof Since both sides of (2. t] * [0. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.s. o 9tI.6) for any . A]. the Markov property can be written E. where Ot is the shift operator.Pt+8measurable r. t and let Px be the probability measure given by t. or. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {. (2. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)]. since Zt • (Y8 o Ot ) is .7) for any Ftmeasurable Zt and any .7).[Y.5 Let {Xt} be Markov w.. for all x. The precise meaning of this is the following: being . nonnegative and Lt+8 = Lt•(Lso9t) (2. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2.8) which is the same as (2. ({Xt+u} 0<u<8) Theorem 2.5) are Tt+e measurable. oo). and then L.v.Ft] = Ex.T9measurable Y8. the natural filtration {Ft} on DE.Y8f t < s.Ft+8measurable.Ft }. Vt+e. (2. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.. s.6) implies (2.Ftmeasurable. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx.t.v.F8measurable r. (2.'s of the form fl' f. Y8. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x.r.Ft].5) Pxa.2. Similarly.'s of the form Zt • (Y8 o 0t) comprises all r.
In between jumps. More precisely . t] = LtExt L8 = Lt. } E[Lt+B I.At where At = k: vk <t U.. t. In between jumps. .e. The result is a sample path relation. (3. UN. < aN < T. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. Indeed. CN. the arrival epochs are Qi. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. We work on a finite time interval [0. ..30 CHAPTER H. The formulation has applications to virtually all the risk models studied in this book. see Dynkin [128] and Kunita [239].Ft] = LtE[L8 o 9t I.. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. aN where or* = T UN_k+l.. with a proof somewhat different from the present one. {Vt} .. reflection at zero and initiar condition Vo = 0.. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l.. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. R = p(R)). A more elementary version along the lines of Theorem 2.5 can be found in Kuchler & Sorensen [240]. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. The storage process {Vt }o<t<T is essentially defined by time reversion. 0 < vl < . Thus R = Ro + f p(R8) ds . the premium rate is p(r) when the reserve is r (i. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. The corresponding claim sizes are Ul. u Notes and references The results of the present section are essentially known in a very general Markov process formulation.6 For {u ... ..1) The initial condition is arbitrary.. then Remark 2. . (using the Markov property in the second step) so that the martingale property is automatic. 3 Duality with other applied probability models In this section. Ro = u (say). .
.____•_.. V = p(V)). {Vt} remains at 0 until the next arrival).__. :.x. instead of (3.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.T) = P(VT > u).1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.. Then rt°) > rt°) for all t when u > v._.1 The events {T(u) < T} and {VT > u} coincide. In particular..AT_t..11 4. V)(u.1.. (3. The sample path relation between these two processes is illustrated in Fig..__. .3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u.___ ..3.e._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3.. 3. (3. Theorem 3.1) we have Vt = At  f P(Vs)ds where A= U= AT . That is. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.
d. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. Then similarly VVN = r0. we can repeat the argument and get VoN_1 > Ra2 and so on. 3. Proof Let T ^ oo in (3. if nothing else n = N). we have r(u) > T. if and only if O(u) < 1 for all u.32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. We get: Corollary 3.1 with Ro = u = ul). Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions.1 with Ro = u = u2). we have RQ„ < 0 so that indeed r(u) < T.1 and its proof is from Asmussen & Schock Petersen [50]. say V.3). = r(VT) . Historically. Historically. with distribution B. and since ruin can only occur at the times of claims. If VaN > 0. Corollary 3.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . 3.3 and being i.2 from Harrison & Resnick [188]. say of water in a dam though other interpretations like the amount of goods stored are also possible. see Siegmund [344]. one may feel that the interaction between the different areas has been surprisingly limited even up to today.i. Then the storage process {Vt} has a proper limit in distribution. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. Theorem 3. the distinction between right. . Hence if n satisfies VVN_n+1 = 0 (such a n exists. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51]. and then '0 (u) = P(V > u). this represents a model for storage.Ul < roil  Ul = RQ„ Va1V_1 < RQ2.U1 > roil . u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell.2 Consider the compound Poisson risk model with a general premium rule p(r).T l . The results can be viewed as special cases of Siegmund duality. Hence RQ„ > 0 for all n < N. and a general premium rule p(r) when the reserve is r.U1 = Rol. The arrival epochs correspond to rainfalls. Resnick & Tweedie [79]. Then Vo. and so on. Nevertheless.
evolves as a random walk with increments Z1i Z2... has a proper limit W in distribution as n + oo.YNn+1) n=0....1.d.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0. Xo = 0. n=0.1... generated by Z1. . ...1. as long as the random walk only takes nonnegative values.e.s. Let further N be fixed and let Wo. Z2 = YN_1 i .h. and is reset to 0 once the r.1)).. just verify that the r . with common distribution F (say). 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.. WN be the Lindley process generated by Z1 = YN. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN. 0 Corollary 4.1. where the Yi are i .1 in terms of Lindley processes .. Most often...min n=0. WN = YN .4.....i. . . Here F is a general probability distribution on R (the special case of F being concentrated on {1..1. For discrete time random walks . the Lindley process Wo..1.. Then the events {r(u) < N} and {WN > u} coincide. .d... N min (Y1 + + Yn).. W2. Z2. 1} is often referred to as simple random walk or Bernoulli random walk).. ZN = .Y1 according to Wo = 0. Z2 = Y2.... . is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4... i.. . I. W1.Yl min (YN . For a given i.. N min (Y1 + • • • + YNn) n=0.w.N (4..... (c) The Lindley process {WN} generated by Zl = Y1.. W1.. of (4. 0).. {Wn}n=0. In particular. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo..2) (for a rigorous proof.. = Xo + Y1 + • • • + Y. R valued sequence Z1... Theorem 4.1) Thus {Wn}n=o... Z2 . hits (oo... . there is an analogue of Theorem 3.2). Proof By (4.N From this the result immediately follows. Z2. .1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}....2) satisfies the same recursion as in (4. . can be viewed as the reflected version of the random walk with increments Z1.
. a sufficient condition for (e) is that EY is welldefined and > 0. equivalently. ZN or. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p. 0 By the law of large numbers.l.. . ±1.. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists. YN in Theorem 4.. W v m and P(W > u) = P (m > u) = 0(u).s..) sup n=0.1.s . there is a more general version of Corollary 4.. Remark 4 .i. Similarly. 176) but appears to be rather intractable. . Combining these facts gives easily the equivalence of (a)(d).. + Z.l.s. .N so that WN _P4 M = supra=0. . YN). SOME GENERAL TOOLS AND RESULTS (d) m = inf. the Y1.1. . then the restrictions of Fx. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo.2 and Theorem 4. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). the Lindley processes in Corollary 4.g. Clearly.. ±2... (e).d.the result is a sample path relation as is Theorem 3.5 does not necessarily lead to a random walk: if. ..ooa.1 is equivalent to WN D MN = (Z1 + .1 have the same distribution for n = 0..=o. either M = oo a. (e)Yi+•••+Yn 74 .. .. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . e. . In that case .. w. 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. One then assumes Yn to be a stationary sequence.g.. For a random walk. . By Kolmogorov's 01 law. Next consider change of measure via likelihood ratios. doubly u infinite (n'= 0. assumption on the Z1...2...) and defines Zn = Yn..3 The i. (d) #. Thus the assertion of Theorem 4.o. In general..34 CHAPTER II. . or M < oo a. (Yi + • • • + Yn) > oo a. N. Y1) has the same distribution as (Y1...1.. Proof Since (YN.1 is actually not necessary ..1. . a Markovian change of measure as in Theorem 2.s.
. Y1). u A particular important example is exponential change of measure (h(y) = e°y'(") where r.s. the changed increment distribution is F(x ) = E[h(Y). In that case. Y < x]. Then the change of measure in Theorem 2. e. for some function h with Eh(Y) = 1. In particular.Y2) = h(1'i)h(I'a).3) holds. = 1 for all n and x. the random walk property implies Ex f (Y1) = Eo f (Y1 ). We get: Corollary 4.4). y ). Then the change of measure in Theorem 2.4 Let {Ln} be a multiplicative functional of a random walk with E_L. Conversely. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. Since L1 has the form g (Xo. this means E(g(x.. implying g (x.s. and so onforn =3.4) ({Ln} is the familiar Wald martingale .. 100 ).. of F).5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) .5 corresponds to a new random walk if and only if Ln = h(Y1) .g. For n = 2..5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.3) 1Pxa. Breiman [78] p. h(Yn) (4. Y) = h(Y ) a. where h (y) = g(0. cf..g. 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 define Ln by (4. (a) = log F(a] is the c.3) holds for n = 1.f. (4.. Y ) f (Y)] = E[g(O. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. Y) f (Y)] for all f and x.4.4. Proof If (4. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .nrc(a )} (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 . say the beginning of each month or year . Xt =Xo+pt+oBt+Mt.Xn < x). (4. we are .. the tradition in the area is to use continuous time models.1). the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. but we omit the details ).e. the pure jump process is given by its Levy measure v(dx). say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). or imbedded into continuous time processes . with premium rate p.5) Note that the structure of such a process admits a complete description. a positive measure on R with the properties e J x2v(dx) < oo.Xt)I. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). . An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . corresponds to a process with stationary independent increments and u = p. however. this description needs some amendments.7) for all e > 0..). In risk theory. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . see Chapter V. In discrete time. the claim surplus process for the compound Poisson risk model . < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) .6) More precisely. i.36 CHAPTER II. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. (4. However. In continuous time. v2 = 0 and v = 3B). Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . . they arise as models for the reserve or claim surplus at a discrete sequence of instants. {Xt} can be written as the independent sum of a pure drift {pt}.t(i . The simplest case is 3 = JhvMM < oo. e f x:IxJ>e} v(dx) < oo (4. given by a the increment distribution F(x) = P(Xn+l .. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). Xt(2)_t(l). Roughly.Ft] = Eof (X. {Xt} is a random walk. 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..
Then the storage process {Vt} has constant release rate 1.6). cf. and the reflected version is then defined by means of the abstract reflection operator as in (4. Corollary 4. O(u. has upwards jumps governed by B at the epochs of a Poisson process with rate . (ex . Now consider reflected versions of processes with stationary independent increments. VT + V for some r. 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.e. Proposition 4. i.10) . then Ee'(xtxo) = Eoeaxt = etx(a).] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. defined as a system with a single server working at a unit rate.min Xt (4.7 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.1. 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. 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).4. and b(u) = P(V > u).8) O<t<T (assuming Wo = Xo = 0 for simplicity).Q and distribution B of the service times of the arriving customers. jxJ v(dx) < oo.3 and decreases linearly at rate 1 in between jumps. is easily seen to be f3pB < 1.s. Furthermore. where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". T) = P(VT > u). [The condition for V < oo a.2). A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. Chapter III. WT = XT . having Poisson arrivals with rate . V E [0.1)v(dx) (4. RANDOM WALKS IN DISCRETE OR CONTINUOUS TIME 37 almost solely concerned with the compound Poisson case and shall therefore not treat the intricacies of unbounded Levy measures in detail.6 In the compound Poisson risk model with constant premium rate p(r) .v. oo].
Theorem 4 .5) and get E [f(Xt+B .11) (eax . In particular.g.Xt)g(s. 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t. This is of course no coincidence since the distribution of Xl . Xt +B .Xt)I Ftl = Eof (X8)g(s.1.xo)tk ( e).1 ) v(dx) . 5 has stationary independent increments as well. Then the Markov process given by Theorem 2. we use the characterization (4.Xt)I'Ftl = E [f(Xt+B .g. u Note that (4. then the changed parameters in the representation (4.. if Lt = e9(xt . Chung [86]).10) is the LevyKhinchine representation of the c. v(dx) = e9xv (dx). Proof For the first statement . (4. t. X8) = Eof (X8)L8 = Eof (X8)• For the second. e. let e" (a ) = Eoeaxl.Xt)L8 o 0tIFt] = E [f (Xt+s .Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. use the representation as limit of compound Poisson processes. we show in the compound Poisson case ( IlvIl < oo) in Proposition III. .6) are µ = µ + Oo2 .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 . Eea(µt + QBt) = et{aµ +a2OZ/2}.f. Q2 = v2. Xt Xo) with E2Lt = 1 for all x.1)eexv(dx). of an infinitely divisible distribution (see.1 that E eaMt = exp fmoo In the general case . 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) . By explicit calculation .38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution.
o[f (S8)g(J8)]. dB/dB = b/b when B. we write Pi.(3 = . . U2. it is then easily seen that Lt = H dB(Ui) i:o.3 and claim size distribution B # B.. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate .2. the corresponding claim sizes .l3 and claim u size distribution B. As for processes with stationary independent increments . B(dx) = B[9] B(dx).St)g(Jt+s)I.3 =..3B[B]. .0 in the following.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .Ft] = Ejt. where .9 If X0 = 0.3 and claim size distribution B. u 5 Markov additive processes A Markov additive processes. corresponding to p = 1. MAP stands for the Markovian arrival process discussed below.1) For shorthand . <t whenever the RadonNikodym derivative dB/dB exists (e. Recalling that U1. a = 0.. are the arrival times and U1. MARKOV ADDITIVE PROCESSES 39 Remark 4. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. 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. Then we can write v(dx) _ /3eOxB(dx) = / (dx). then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4. 0. Ei.. one reason is that in parts of the applied probability literature. 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. Thus (since µ = p = 1. B have densities b.4).(3B(dx). the structure of MAP's is completely understood when E is finite: 2and only there . Example 4 . (5. abbreviated as MAP in this section2. is defined as a bivariate Markov process {Xt} = {(Jt. Example 4 .8. Ei instead of P2. b with b(x) > 0 for all x such that b(x) > 0).11 For an example of a likelihood ratio not covered by Theorem 4..g. v(dx) _ .5.
Jn = j. vi(dx) in (4.[a) = (Ei[easl.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. As a generalization of the m. let {Jt} be standard Brownian motion on the line. v.g. t+s) where Jt . In continuous time (assuming Dpaths). oo). . As an example..6) depending on i.. which we omit and refer to Neveu [272] or cinlar [87]. the distribution of which has some distribution Bij. An alternative description is in terms of the transition matrix P = (piA.) If E is infinite a MAP may be much more complicated. In addition.Sr_1..1 For a MAP in discrete time and with E finite. by generating Yn according to Hij when J„_1 = i. 1 J1 ='^])iJEE = (Fij[a])i .i.40 CHAPTER H.. (That a process with this description is a MAP is obvious.9 EE = (iii&ij[a])i j EE . this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1. If all Fij are concentrated on (0. a MAP is the same as a semiMarkov or Markov renewal process.jEE• On an interval [t.. consider the matrix Ft [a] with ijth element least Ei . Y1 E dx) where Y„ = S„ . Proposition 5.f. 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. Y2. the converse requires a proof. with the Y„ being interpreted as interarrival times. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time.it = A.J1=j)= Fij (dx) Pij In simulation language. SOME GENERAL TOOLS AND RESULTS In discrete time. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. {Jt} is specified by its intensity matrix A = (Aij)i. {St} evolves like a process with stationary independent increments and the parameters pi. Fn[a] = F[a]n where P[a] = P .o(Ji = j.
up to o( h) terms.1)) . 013 . J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. vi(dx). \ diag Ft[a] = Ft[a]K. In matrix formulation . assume that the Markov chain/process {Jt} is ergodic. By PerronFrobenius theory (see A. where K[a] = A+ (r. Jt = j] Ejesh'^ + E Ak j hEi [ease .(')(a)) diag + ().4c). kEE Jn = k]Ek[e"Y" .1) } (recall that qjj = 0). vi(dx) (i E E). Sn ) yields Ei[easn+ '. this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. u In the following. MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. Then the matrix Pt[a] with ijth element Ei [east. Bij (i. Jt = k] { xk kEE j la] . Jt = j] is given by etK[a]. qij. a= .2 Let E be finite and consider a continuous time Markov additive process with parameters A. Jn+1 = A] = 41 Ei[ e 5„.ijgij(Bij[a] . aSt h = (1 + Ajjh) Ei [east . pi. 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 . u Proposition 5. we infer that in the discrete time case the .qkj + k?^j qkj Bkj [a] } = Ei [east. Then. 00 r(i) (a) = api + a2ot /2 + f (e° . j E E) and So = 0. Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi .5. Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east .1 )v(dx).
Since v("). The corresponding left and right eigenvectors v("). In particular. and we shall take V(a)h(a) = 1. h(") are only given up to a constants.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l .t.42 CHAPTER II.Eikjt = ttc'(0) + ki . just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a).2) where 7r = v(°) is the stationary distribution. We also get an analogue of the Wald martingale for random walks: Proposition 5. (5. Corollary 5. The function ic(a) plays in many respects the same role as the cumulant g. Jeast. cf. Furthermore.h(a)vva)etw(a). Eie"sth^a) = e'Pt[a]h( a) = e. . Proof By PerronFrobenius theory (see A. its derivatives are 'asymptotic cumulants'. of a random walk. u Let k(a) denote the derivative of h() w. 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.r.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).7.4.5 EiSt = tK'(0) + ki . Jt = j] .f.tK(a)h(a) J jj it L o is a martingale. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of .3 Ei [east. a. Corollary 5. Corollary 5. h(") may be chosen with strictly positive components. as will be seen from the following results.4 Eie"sth(a) = h=a)et?("). Yrh(a ) = 1.4c). we are free to impose two normalizations.c(a) (and h(")). and write k = k(°). Proposition 5.e=e°tk. Proof For the first assertion. Then h(°) = e. cf.
Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i. E=ST = tc'(0)E7. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. Corollary 5. tam E tSt a (0). summing and letting a = 0 yields E„ [St + 2Stkj. Remark 5 . t a oo. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.e.5.Eikjr .4) . 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . for a random walk: Corollary 5.5.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk .") }) .a) + ttc (a)2hia ) Multiplying by v=. the distribution of Jo). we differentiate (5.+ k. [E.3) to get Ej [St a " st h i(a ) + 2Ste"st k(a) + e"st k^a) J etI(a) (kia )' + ttc (a)ki") + t {ic"(a)h. subtraction yields Vary St = tic"(0) + O(1).2ttc (0)Evkjt + 0(i). In the same way.7 No matter the initial distribution v of Jo. (5. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x). Squaring in Corollary 5.g. 8 Also for E being infinite (possibly uncountable ).4. .. (5.5 yields + W (a)k. For the second .6 For any stopping time T with finite mean.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. one obtains a generalization of Wald's identity EST = Er • ES. More precisely. u The argument is slightly heuristic (e..
inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. however. let ha(i. In view of this discussion . St)} be a MAP and let 0 be such that h(Jt) OStt. some extra conditions are imposed.9 The condition that (5. Then {Lt } is a multiplicative functional. 1) (i.3b and Remark VI.5 defines a new MAP. h(Jo) Lo is a Px martingale for each x E E.6..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. First.6. 0) = n(a) h(i).(9) {Lt}t>o = . 0 Proposition 5.5) is a martingale . where {Jt} is deterministic period motion on E = [0. (5.10 Let {(Jt. St) } as follows. Jt = (s+t) mod 1 P8a. For t small . this is. this leads to h(i) + tcha( i.5.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. however..4 that { h(Jt) easttK(a) L o (5. in particular that f is bounded. G is defined as Gf (x) = lim Exf (Xt) .f (x) tyo t provided the limit exists.e. u forsEE). s) = ea8h(i).6) We shall not exploit this approach systematically.s. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i). xEE . Given a function h on E.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. From (5. Usually. 0) = h(i )( 1 + ttc(a)). gha(i. V.for the present purposes. we take the martingale property as our basic condition below (though this is automatic in the finite case). and the family {f LEE given by Theorem 2. An example beyond the finite case occurs for periodic risk processes in VI. Remark 5. see.
St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5.tc(0)e = 0 . vi (dx) = f3 Bi(dx) with .7(dx) Bij [0] Bij(dx) in the continuous time case . In the infinite case .c(0)e = tc(0)e . Bi [0] Remark 5. . in the discrete time case.Qi < oo and Bi a probability measure. In particular.1) . That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .11 Consider the irreducible case with E finite.ic(0)e = ic(0)Oh e) h(e) . Then the MAP in Proposition 5. (5. 1 + q(b .(0)j. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. Bi(dx) = Bi(dx).7) In particular. That 0 < qij < 1 follows from the inequality qb <1. if vi(dx) is compound Poisson.11 below in the finite case. We omit the details. Bi. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. 0 < qij < 1 and Bij [0] > 0.. 0<b<oo. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx).10 is given by P = eK(e) Oh e) F[e]Oh(').1) eft ea' f ij (dx) = Hij (dx) Hij [0] . qij = r. Ai = µi + 0Q. one can directly verify that (5. u Theorem 5.5.. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ .12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. 0<q<1. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].1) holds for the P. ^i = of qij Bij [0] 1 + qij ( Bij [0] .
r.tc(0)I. F:j with a density proportional to eei . a = 0 in (5. (dx) of a process with stationary independent increments follows from Theorem 4.46 CHAPTER II. v. (dx). In matrix notation .13) for matrixexponentials .. Jt = j] = hie) . since Hij.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] ..K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) .11.8. Jl = j] :(Yi E dx.8) h(.Bay [0]) That k(') (a + 0) .t. Similarly. Yi E dx. Jt = A. . First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . are probability measures .8). . Jl = j) = Ei[Lt.tc (') (0) corresponds to the stated parameters µ. H1..tc(0)' )Ah() = Oh(o) K[a + 0]Oh() .8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. this implies k[a] = A 1 ) (K[a + 0] . Here the stated formula for P follows immediately by letting t = 1. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. in continuous time ( 5. Letting a = 0 yields the stated expression for A. it follows that indeed the normalizing constant is H1 [0].e) Consider first the discrete time case . This shows that F. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. Hence the same is true for H=j and H. Further Fib (dx ) = P=(YI E dx. is absolutely continuous w. this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. Now we can write K[a] =A+A ) ( K[a + 0] . v= .
IIG+II = G+(oo) = P(T+ < oo) = 0(0) < 1 when 77 > 0 (there is positive probability that {St} will never come above level 0).3 for an infinite E are given by Ney & Nummelin [266].(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. Though the literature on MAP's is extensive. [225]. the literature on the continuous time case tends more to deal with special cases.e.. [261].1) = Aij4ij(Bij[a] . 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. [226] and Miller [260]. 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]. has no mass on (oo. hardly a single comprehensive treatment. is slightly less general than the present setting. oo). Note that G+ is concentrated on (0. For the Wald identity in Corollary 5. < x) = 11 (S. . an extensive bibliography on aspects of the theory can be found in Asmussen [16]. see also Fuh & Lai [149] and Moustakides [264]..Bij[0]) = hjel)ijgijBij[0](Bij[a] . 0].+ < x. which.6.6.)Ajjgij(Bij[a+0] . h. however. Conditions for analogues of Corollary 5. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. [262] in discrete time. however. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5.1). i. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S. there is.7). and is typically defective. 7+ < oo).
48 CHAPTER K. see Fig.B(x) denotes the tail of B. Thus.1. the sum of all the ladder steps (if rl > 0. there are only finitely many). Recall that B(x) = 1 . 1 For the compound Poisson model with p = 01LB < 1. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. g(y)R+(dy) = E f g(St)dt.1. the second ladder height (step) is ST+(2) . it follows that for g > 0 measurable. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. Theorem 6 . Here bo(x) _ B(x)/µB.. i.2) . the dependence structure seems too complicated to be useful). o 00 (6. Also. 6.d.e.i.e. where basically only stationarity is assumed. they have a semiMarkov structure (but in complete generality. i. = ST+(1) Figure 6. In any case. For the proof of Theorem 6. has no mass on ( 0. which gives an explicit expression for G+ in a very general setting. 0 f T+ (6. 6. at present we concentrate on the first ladder height.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. a fact which turns out to be extremely useful.(3B(x ) = pbo(x) on (0. 0].. the ladder heights are i. On Fig. In other cases like the Markovian environment model. In simple cases like the compound Poisson model. oo). and the maximum M is the total height of the ladder. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. The main result of this section is Theorem 6. The first ladder step is precisely ST+.5 below. G+ is given by the defective density g + (x) =. To illustrate the ideas. by approximation with step functions . we shall first consider the compound Poisson model in the notation of Example 1.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+.T+ > t)dt = E f 0T+I(St E A) dt.00 ).2.ST+(1) and so on.1. R+ is concentrated on (oo.
6. 6. . 0]. 0 < t < T) P(STEA.T+>T) = P(STEA.St<0.2(a): T+ > t Figure 6. has the same distribution as {St}o<t<T. since the distribution of the Poisson process is invariant under time reversion.2(b): r+ < t Thus.O<t<T). 49 Proof Let T be fixed and define St = ST . St S* t a Figure 6. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. ST < ST_t. see Fig.ST<St.0<t<T) = F(ST E A. 0 < t < T.ST_t.ST<St. P(STEA. That is.2 R+ is the restriction of Lebesgue measure to (00. THE LADDER HEIGHT DISTRIBUTION Lemma 6 .2.O<t<T) = P(STEA.
oo).St _)I(r+ > t).T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A . and since the jump rate is /3. oo). and (6.3 Lemma 6 .t dT. U + St_ E A. this is just the Lebesgue measure of A. cf. E A} precisely when r+ > t. But since St 4 oo a. s. 6.3 G+ is the restriction of /3R+*B to (0.. for A C (0.3 where the bold lines correspond to minimal values.2) in the last). That is. Figure 6.St _). G+(A) = Q f 0 B(A .. The probability of this given { Su}u<t is B(A .St).50 CHAPTER II. SOME GENERAL TOOLS AND RESULTS Integrating w.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. . T+ > t] 0 _ /3 f E[B( A . it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time . Fig.r.y) (here we used the fact that the probability of a jump at u t is zero in the second step. we get G+ (A) = f 00 /3 dt E[B(A .
The traditional representation of the input sequence {(TT.z)B(dz) _ f I(x < z)B(dz) _ f (x). assuming basically stationarity in time and space. 0 Generalizing the setup.4) are (ak... 2.* ) and the second the mark (the claim size Uk ). 6 . Uk) (k = 1.Q f r+(x . 6 . Nt St =>Uk k=1 t where Nt = max{k = O. . We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. U k)} k=1 a is as a marked point process M *. as a point process on [0.4). oo) x (0. {St+8 . In the stationary case. The points in the plane (marked by x on Fig. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup. h]} /h (by stationarity. oo). we define the arrival rate as E# { k : ak E [0 . obviously.) where ak = Ti + • • • + Tk . 4 (the points in the plane are (ak .6.M o 08 shifted by s is defined the obvious way. Uk) for those k for which ak . i. Lemma 6..1. the first component representing time (the arrival time o.e.3 yields g+ (x) = . this does not depend on h).s.. Fig.T+ < oo).:T1 +•••+Tk <t}. this is equivalent to the risk process {St*} being stationary in the sense of (6.s > 0). i. The marked point process .e. 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. 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.. .S8 )t> o = {St }t>o for all s > 0. cf.1 With r+(y) = I(y < 0) the density of R+.
h] and the sum approximately ^o(M*)I(ul < h). where TI = 0. k: vk E [0. letting h J. V(M* o eak ).. As above . This more or less gives a proof that indeed (6. SOME GENERAL TOOLS AND RESULTS M* U. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6.2. vi(dx) = . of (6.s. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists.52 CHAPTER II. We represent M by the sequence (Tk. 0. h. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 .4 Consider a finite Markov additive process (cf.e. i. e. = 0 . The two fundamental formulas connecting M* and M are Eco(M) = aE E. Sigman [348] for these and further aspects of Palm theory. .g.5) does not depend on h. the r. Example 6 . Oh becomes the approximate probability F(ri < h) of an arrival in [0.4 Given a stationary marked point process M*.5) represents the conditional distribution of M* given vi = 0.. See. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t.. and let T = T2 denote the first proper interarrival time .QiBi(dx). o. most often one takes h = 1). where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. Uk) k=1.. Section 5) which has pure jump structure corresponding to pi = a = 0. h] Eco(M*) = 1 E f co(M o Bt)dt.
v. Jo) w. the probability aij of Jt . an arrival for M* occurs before time t + dt w. Note in particular that the Palm distribution of the mark size (i.O for i # j.= i. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution .p.OEU0./. A stationary marked point process M* is obtained by assigning Jo distribution Tr.. 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 .p.6. the ruin probability .s. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. Jt = j is iri(3i /. Before giving the proof. Then the ladder height distribution G+ is given by the (defective) density g+(x) = .e. .6 Under the assumptions of Theorem 6. 5 Consider a general stationary claim surplus process {St }t>o.OF(x). the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J. It follows that we can describe the Palm version M as follows . let U0 be a r.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . and by some additional arrivals which occur w.O for i = j and iriAijgij/.p. If Jt_ = i. First choose (Jo_. After that.*(0) with initial reserve u = 0 is p = /3EU0.5. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i.O fo "o F(x)dx = . qij when {Jt} jumps from i to j and have mark distribution Bij. we note: Corollary 6. and that p = 0EU0 < 1. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. Assume that St * . This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . let the arrivals and their marks be generated by {Jt} starting from Jo = j.oo a. aij for (i.
l. 0).(left limit) when 0 < it < t and is illustrated on Fig .. The sample path relation between { Su } and { Su } amounts to S„ = St . 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..0<u<t) = P(StEA. oo) p(t) = P(St EA.0<u<tIAt) = P(St EA.o. that M* and M have doubly infinite time (i. .s. Now conditionally upon At .. { Su}0<u<t is distributed as a process {Su} . h..0<u<t) = P(St EA. moves down linearly at a unit rate in between jumps and starts from S0 = U. Let p(t) be the conditional probability that ST+ E A. 105) shows that one can assume w. and the kth preceding claim arrives at time t . in (0. oo ) and the arrival times 0 > 0_1 > a_2 > .Q_k and has size U_ k. . 2.g.St <. oo) x (0 .54 By (6. (k = St}t>o 1. It follows that for A C (0. in (oo.. which makes an upwards jump at time . the arrival times 0 < 0'1 < Q2 < . Proof of Theorem 6. has a very simple interpretation as the average amount of claims received per unit time . 0<u<t) = P(St EA.o.5..Su< 0. T+ = t given the event At that an arrival at t occurs .A. CHAPTER H. are point processes on (oo .e. the mark at time Qk is denoted by Uk.. A standard argument for stationary processes ([78] p. 6.5.o<u<t where a claim arrives at time t and has size Uo.).Su<0.St*_ u.. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0.St<Su.5). We then represent M by the mark (claim size ) Uo of the arrival at time 0.$St_ u.Mt).0<u<tIAt) = P(St EA. The last property is referred to as insensitivity in the applied probability literature. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt.Su_ <0. .1] here the r . oo)).
5 where it = { St < Su.. In Fig. the support of L has right endpoint U0. Mt)dt = i3EL(A) o"o . 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. cf. NIt)dt . Fig. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}.s.6. and since by assumption St * oo a. the left endpoint of the support is oo. Uo]. 6. G' (A) = 3 f P(St E A. A sample path inspection just as in the proof of Lemma 6 . t a oo. and we let L(dy) be the random measure L(A) = fo°° I(St E A.5. 0 < u < t } is the event that { Su } has a relative minimum at t . 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. Thus. Since So = U0.
56 CHAPTER II.1).6 is Bjork & Grandell [67].5 is due to Schmidt & coworkers [48]. [263] (a special case of the result appear in Proposition VI. . A further relevant reference related to Corollary 6. 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 6.2. [147].
e. and assume that • { Nt}t>o is a Poisson process with rate j3. 4.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. Thus . i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. i. i.d. Panjer's recursion ( Corollary XI. with common distribution B.4 below . are i. 3). • the claim sizes U1. Some possibilities are numerical Laplace transform inversion via Corollary 3. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. see Chapter IV. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . and independent of {Nt}. • the premium rate is p = 1.. St = uRt = EUi t. say.. For finite horizon ruin probabilities .6) and simulation methods ( Chapter X). exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. U2. 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 . . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes .
's etc. (c) Ee8St = et" (8) where c(s) = f3(B[s] . P = PAB = 1/(1 + rl) Proposition 1.g. we shall start by giving the basic formulas for moments. Dufresne & Gerber [126].1).1) .6pBa). of the claim surplus St .t = fltpB . For (c).t = t(p . We do not spell out in detail such generalizations. Schmidli [319]. (d) The kth cumulant of St is tf3p(k) for k > 2. we get Ee8st = 00 e8t c` Ee8 (U1+. and Schlegel [316]...)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. say stable Levy motion.Rt. Proof It was noted in Chapter I that p . 0 . Furrer [150].1 is the expected claim surplus per unit time. [324]. where K(k) (0) is the kth derivative of is at 0. e .58 CHAPTER III. Write pB^) = EUn' YB = Pali = EU. A more formal proof goes as follows: Nt r Nt ESt = E > U k . for (d) just note that the kth cumulant of St is tic(k) (0). 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). cumulants . and this immediately yields (a).+Uk)P(Nt = k) k=O e8t k=O B[s]k .t = E E [ U k k=1 k=1 Nt . and that B(k)[0] = Pak).u .1 (a) ESt = t(13µ$ . {Bt} a Brownian motion and {Rt} a pure jump process. THE COMPOUND POISSON MODEL Poisson risk process.g.s.1) = t(p .1). See e. (b) Var St = t. m.f.t = E[Ntµs] . 1 Introduction For later reference.
meaning that the increments are stationary and independent. (a) No matter the value of 77. 1.i. then St> SnhV>Snhh. In particular. we need the following lemma: Lemma 1. For example. Proof We first note that for u. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. II. S„+V > S„ . lim supt. Sn+0 . The point of view in the present chapter is. Indeed.3) is proved similarly.3EU01 = 1µs where rt is the safety loading. obviously 0(u) = F(maxk Sok > u). The connections to random walks are in fact fundamental... rather to view {St} directly as a random walk in continuous time. u + v].1 is the same as if {St} was a random walk indexed by t = 0. v > 0. The right hand inequality in (1. we get a discrete time random walk imbedded in the claim surplus process {St}. and the value is then precisely v. the Uk .d. then lien inft. cf. St = oo. (d) If 17 = 0. In this way. 2.1.S„ attains its minimal value when there are no arrivals in (u.1 = . however. we have Sok . Obviously. which is often used in the literature for obtaining information about {St} and the ruin probabilities. so that {Sok } is a random walk with mean EUET = EU. (c) If 77 > 0. then St 00. then Snh .Sok_l = Uk . (b) If 77 < 0. where Tk is the time between the kth and the (k .1)th claim.Tk..3 If nh < t < (n + 1)h.h < St < S(n+1)h + h. For the proof. . Here is one immediate application: Proposition 1.V. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo.4.Tk are i. St = oo. then St 4 co. if t = nh + v with 0 < v < h.. We return to this approach in Chapter V.
. THE COMPOUND POISSON MODEL Proof of Proposition 1. {Snh}n=o..5 The limiting distribution of St .. u 307).o.. lim supn_.. Snh/n a4' ESh = h(p .1._.4 The ruin probability 0(u) is 1 for all u when 77 < 0.. is a discretetime random walk for any h > 0. If rl > 0. Considering the next downcrossing (which occurs w. then {St} upcrosses level 0 a. Snh u = 00 (the lemma is not needed for (d)).2.1.60 CHAPTER III. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1.s. The general case now follows either by another easy application of Lemma 1. hence by induction i.6 Often it is of interest to consider size fluctuations. Remark 1 .1. at least once. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time.t .. is a discrete time random walk.1). if P(M > 0) = 1.p.. 1 since St 4 oo) and repeating the argument.. 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 . However.2. Part (d) follows by a (slightly more intricate) general random walk result ([APQ].3. it suffices to prove 4'(0) = F(M > 0) < 1.2: Proposition 1. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. Thus using Lemma 1. and (b). For any fixed h. or by a general result on discrete skeletons ([APQ] p.. .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). There is also a central limit version of Proposition 1. p.1(b)) that the assertion holds as t 4 oo through values of the form t = 0. 0 Snh = 00. 2h. 169) stating that lim infra. 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. h A similar argument for lim sup proves (a). and < 1 for all u when 77 > 0. Corollary 1. {Snh}n=o. where the size of the portfolio at time t is M(t). This contradicts u St400. h. it is seen that upcrossing occurs at least twice.3. Notes and references All material of the present section is standard. and hence by the strong law of large numbers. (c) are immediate consequences of (a).
d. cf. i. [APQ] Ch. p < 1.1) representing the distribution of M as a geometric compound. 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}. 0 Alternatively. It is crucial to note that for the compound Poisson model. oo ). Summing over n. Fig.6. As a vehicle for computing tIi(u). The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . 1 The distribution of M is (1. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . Thus .just before ruin is again B0. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime.. and we shall here exploit the decomposition of the maximum M as sum of ladder heights.6.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). The expression for g+ was proved in Theorem 11. IV. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. equivalently. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). nevertheless. the ladder heights are i. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i.1) is not entirely satisfying because of the infinite sum of convolution powers.e.1..IIG +II)EG+ . Theorem 2. B(x)/aB.1 provides a representation formula for 0(u). This follows simply by noting that the process repeats itself after reaching a relative maximum. Note that this .34 or A. n=0 (2. 11. the formula for the distribution of M follows . Combined with i/i(u) = P ( M > u). and we further get information about the joint conditional distribution of the surplus and the deficit. but we shall be able to extract substantial information from the formula.2.P) E PnBon(u) . We assume throughout rl > 0 or. 1e. which we henceforth refer to as the PollaczeckKhinchine formula. (2.1.
cf. (a) 11 (ST(o)_ > x. see for example [APQ]. Theorem A1. 2 The joint distribution of (ST(o )_. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB.6. In the risk theory literature. Theorem 2. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. [62].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. (d) the marginal distribution of ST(o)_ is B0. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. Asmussen & Schmidt [49]. Theorem 2 . (c) the marginal distribution of ST(o)_ is Bo . cf.62 CHAPTER III.i. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. V is uniform on (0. The proof of Theorem 11.1 is traditionally carried out for the imbedded discrete time random walk. Again.d. cf.5.V)W) where V.6.2(a) is from Dufresne & Gerber [125]. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula.just after ruin. ST(o) > y. We assume rt > 0 throughout. . there is a general marked point process version. in this setting there is no decomposition of M as a sum of i. (1 . W are independent. However. Feller [143] or Wolff [384]. As shown in Theorem 11. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). ST(o )) given r (0) < oo is the same as the distribution of (VW. 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)_ = y is the overshoot distribution B(Y) given by Bov)(z) _ Bo (y + z )/Bo(y). 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. where it requires slightly more calculation. ST(o)) is given by the following four equivalent statements: B(z) dz.5. f +b (b) the joint distribution of (ST( o). Beekman [61]. see Schmidli [323] and references there.
B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . 0 .3. But claims are exponential . however . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w.6. I. For a failure at x. As shown in 11. Integrating from u to oo. Thus r(x) = S(1 .1)1 00 ( 1 . Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 . 1 . The result can. Thus .O)e(b0)x.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.p. 3b Exponential claims Corollary 3.2 If B is exponential with rate S.p) = S .e.. use Laplace transforms. a further relevant reference is Bjork & Grandell [67].p)pSe a ( l v)x = p( S .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.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 0.1 e ax = n1 (n . and hence this overshoot has the same distribution as the claims themselves . the formula for P(O) holds in a more general setting. Let r ( x) be the failure rate of M at x > 0.p. the result follows . 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. Alternatively. also be seen probabilistically without summing infinite series . hence without memory. then.p) E pn S n x n.
we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype.h. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .3).T+ <oo)+P(M> u.T+ <oo). and conditioning upon S.p + G+ * Z(u) = 1 .+ >u.2).y)G+(dy ) = f U V(u . Corollary 3.3 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 .s.3) below.4) is similar (equivalently.3)).1. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3. then 24 1 V. Then the first term on the r.S. (u) 35eu + 35e6u. and weights 1/2 for each.g. is ?7+ ( u). u + oo.4) can be derived by elementary algebra from (3. 2 is one of the main classical early results in the area. cf. The case of (3. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).y)f3 (y) dy.64 CHAPTER III.1 p pBo(u).3) Equivalently. u .4) zu P(M > u .3. T+ <00) (3.+ <U.p + f u Z(u .3. (b) use stopped martingales . THE COMPOUND POISSON MODEL In VIII.+ <u.+ = y yields P(M>u. just insert the explicit form of G+.y)/3B (y) dy. 0 Proof Write o (u) as P(M>u) = P(S.y)G+(dy) For the last identity in (3. (3.2) Notes and references Corollary 3. (3.1) For a heavytailed B.i(u) satisfies the defective renewal equation Z(u) = 1 . the survival probability Z(u) = 1 . II. A variety of proofs are available . E. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7. (3.S. (3. (Example VIII.
3 . 206207). [APQ] pp.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 . 191).3 is standard .(3 .(3B[s] 1 . either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. [APQ] pp.s ./3B[s] . by analytical manipulations (L'Hospital's rule) from (3. Bo of B0 as m e8u B(du) = B[s] .p) E p"Bo[s]" = 1 . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities. In view of (3./3B[s] which is the same as (3.5).4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p.6) 00 = (I .5 The first two moments of M are 2 EM . In fact. e.g. Embrechts.7).. We omit the details (see. The approach there is to condition upon the first claim occuring at time t and having size x .5) Proof We first find the m. . Of course.5 can be found in virtually any queueing book. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3.4) can be derived by elementary but tedious manipulations.7) and Corollary 3.3.p = (1 . 0 Notes and references Corollary 3.P)PB 2(1 . Also (3. 111112 or Feller [143].1 Bo[s] = f oc.)3B[s]) (3. Some relevant references are Abate & Whitt [2]. Griibel & Pitts [132]. (3.7) s +. see e.pBo[s] no (1 .p)2 3(1 .3 .PPB2) EM2 = PPB) + QZPBl 2(1 .8) Proof This can be shown.s . it is not surprising that such arguments are more cumbersome since the ladder height representation is not used. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3. Corollary 3.. g.p)s .p)s s /3 . for example.5).f. eau B(u) du = f PB 3PB SPB 0 o (3.g.Ps s(.P)pB' (3.
u)]k1 ku+1) [/3( k .6 If B is degenerate at p.4) for Z( u) means f lhu Z(u) = 1. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .1)./3Z(u .Q) k=0 k! E e0( = /32(u) .Q) 3e.9) shown for n .u + 1 )]k = QZ(u) .)3(1 . .u) [N(k . THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.z/'(u) takes the form Z(u) L^J L.u)]k d 1 u) _ a) n ( du ( k! (1  . of (3.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. differentiation yields Z(u) _ /3Z(u) .1).9) follows for 0 < u < 1.66 CHAPTER III.1 < u < n and let Z(u ) denote the r.Q (k 1 k= n  [O(k . eO('u) [)3(k . Assume (3. then p) 1: ep(k u/. For n < u < n + 1.1)! k=1 u1 . Z^ =eR(k.3+ 18+ J0 Z(uy).u) a)Qea" + (1 .9)./32(u .u)]k k! k0 The renewal equation (3.s.u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.u) [p(k .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k .3I( 0<y<1)dy Z(y)/3I(0<u.h.
4) . We could first tentatively consider the claim surplus X = St for a single t. we just have to multiply (4. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.d.1) .a. 4 Change of measure via exponential families If X is a random variable with c. or equivalently BB[a] = B[^+ Repeating for t 54 1. (4.f. in terms of the c. (4.g. and thus (4. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx).4) works as well.1 that c(a) = /3(B[a] . of F9.2) (Here 9 is any such number such that r.rc(9) = .g.1) or equivalently. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. B9(dx) = B[9] B(dx). (4.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11.(9).a. 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) .1) . we set up .6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139].2). F and c. it follows that Z(u) = 2(u) for n<u<n+1.4.4. The question then naturally arises whether ie is the c. and define rce by (4.(9) is welldefined. The answer is yes: inserting in (4. but will now be repeated for the sake of selfcontainedness. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. co(a) = rc(a + 9) .3B[9]. 0 Notes and references Corollary 3.f.f. say t = 1: recall from Proposition 1.r.g.2) shows that the solution is Ox [O]0].f. Formalizing this for the purpose of studying the whole process {St}.3) by t.3B = .Qe(Bo[a] .
(4. Z is measurable w. the corresponding expectation operator is E9.5) for the density of n i. oo) governing a given compound Poisson risk process with arrival intensity. Let FT = o(St : t < T) denote the o•algebra spanned by the St. and PBT) the restriction of PB to FT.3 and claim size distribution B. (4.r..FT. in particular the expression (4. BB by (4. replications from Fe (replace x by xi in (4. and define 09.1) and multiply from 1 to n). Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.f. = exp {BST . t < T.7) Proof We must prove that if Z is FTmeasurable. v(Xi.8) follows by discrete exponential family theory.t. THE COMPOUND POISSON MODEL Definition 4.i. But let Xk = SkT/n .3 Let T be any stopping time and let G E FT. . (4.(9)} (4. . . with common c.d. n) for a given n. .68 CHAPTER III.0e and claim size distribution Be.d. and dP(T) dP^T) That is. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . (4.nr.9) Proof We first note that for any fixed t. and thus (4. .t. G]. for G E FT.2.Tic (0)} . G C {T < oo}. G].1.4).1 Let P be the probability measure on D[0.FTn) = Q(SkTIn : k = 0. EeeBSt + tk(B) = 1. Then the Xk are i. . then EBZ = E [Ze9ST _T"(9)I .i.g.. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) . The identity (4. it suffices to consider the case where Z is measurable w. the PBT) are mutually equivalent on. Ti(a)/n. Proposition 4.10) .S(k_1)Tln. The following result (Proposition 4. Xn).r.8) By standard measure theory.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} ..2 For any fixed T.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 . (4.5) for the density.
ST) + (T . 5. Then G E FT. Given FT. GT C_ Jr < T}. 5 Lundberg conjugation Being a c. 77 Thus.FT]] = EB [exp { BST + Trc(9)} I(G)] . Thus by (4. Ee [exp { BST +Trc(9)} I(G) FT)] = 1.(Y) = 13(B['Y] . and hence (4.1 It is seen that typically) a ry > 0 satisfying 0 = r.9) holds for G as well. subject to the basic assumption ij > 0 of a positive safety loading. c(a) is a convex function of a. the typical shape of rc is as in Fig. Letting T t oo and using monotone convergence then shows u that (4.7) holds.r)rc(9)}I . (a) rc (a) (b) KL(a) 'Y 'Y Figure 5.5. Then GT = G n Jr < T} satisfies GT E FT. (4.7 1 Some discussion further supporting this statement is given in the next section.f. (0) + rc'(0)a = 0 + ES1 a = a (p .1) .1(a). Thus.1) _ 1 + a.g. The behaviour at zero is given by the first order Taylor expansion c(a) r. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . t = T . . according to what has just been proved.r is deterministic..9) holds with G replaced by GT.10). Now consider a general G. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.
THE COMPOUND POISSON MODEL exists .1(b). e.1 Consider the case of exponential claims. we further note that ( 5. Thus B[7] = 6/. Fig.2) 7 Figure 5. and (4.2)) is 7 = 5/3.3.4) yields /3L = b and that BL is again exponential with rate bL =. we write FL instead of F7.3) cf. (5. Example 5 . 5.s). the claim surplus process has positive drift > 0.2 is B(7) = 1 + ^.2 s As support for memory.1) . b[s] = 5/(b . 5. the Lundberg exponent.1) is precisely what is needed for one of the terms in the exponent . Note that KL (a) = /L (BL [a] . .a = i(a + 7).3. Thus. Fig.70 CHAPTER III. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes. G = {T(u) < oo} in Theorem 4.g. It is then readily seen that the nonzero solution of (5.1) (or (5. u It is a crucial fact that when governed by FL. Taking T = r(u).1(b). An established terminology is to call y the adjustment coefficient but there are various alternatives around. (5.1) is known as the Lundberg equation and plays a fundamental role in risk theory .QL instead of /37 and so on in the following .3. an equivalent version illustrated in Fig. (5. Equation (5. 5.4) ELS1 = #L(0) cf.
which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.1 (5. PL ) with density 1 . (5.1e.8) . take first 0 = ry.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) . LUNDBERG CONJUGATION 71 to vanish so that Theorem 4.1.4).r. see A . T = T+.Ce7u as u 4 oo.5).6) Proof By renewal theory. Then P(ST+ E A) = EL [exp { 7S?+} .7) is the same as (5.1p . Letting e(u) = ST(u) . Proof Just note that e(u) > 0 in (5.t.5.7) 0 and all that is needed to check is that ( 5. G = {S.+ E A} in Theorem 4. where C .6 ).G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL. V)(u) < e7u. ST+ E A] . To this end.(u)} . 0 Theorem 5 . Since a7' is continuous and bounded.ascending ladder height distribution and µ+ its mean. we can rewrite this as 0(u) = e"ELe7^(u).P Y j o' xeryxOB (x) dx /3k [Y] .e7x)G+(dx).3. V) (u) = P(T(u) < oo) = EL [exp {ryS. (5.(oo) (in the sense of weak convergence w.G+L)(x)) dx ry^+L) J 00 f 0 (1 .2 (LUNDBERG'S INEQUALITY) For all u > 0. T(u) < oo] .5) Theorem 5 .3 takes a particular simple form. (5. 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. e(u) has a limit i.
. Noting that SIG(L)II = 1 because of (5.1) (5. From this it follows. 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.e.8) yields +L) J0 xel'B ( x) dx (5.10) VW = JI c* e° (x) dx = a (B[a] .1 . u .3 (this was found already in Example 5.72 CHAPTER III.4 Consider first the exponential case b(x) = Seax.4). that 7 = S . or equivalently of how close the safety loading 77 is to zero. (5. this solves the problem of evaluating (5. THE COMPOUND POISSON MODEL In principle. but some tedious (though elementary) calculations remain to bring the expressions on a final form.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] . Then 0(u) = pe(a_Q)u where p = /3/S. we get L where 00 (1 . Using (5.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.")G + (dx ) = 1  J0 00 3B(x) dx = 1p.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.1 above) and that C = p.7). of course.12) Example 5 .
5. LUNDBERG CONJUGATION Remark 5.5 Noting that PL  1 = ,3LIBL  1 = #ci (0 ) = k (ry) _ ,QB' ['Y]  1 ,
73
we can rewrite the CramerLundberg constant C in the nice symmetrical form G, _'(0)1  1  p K'(7) PL1
(5.13)
In Chapter IV, we shall need the following result which follows by a variant of the calculations in the proof of Theorem 5.3: 1  aB[ry  a]  1 Lemma 5 . 6 For a # ry, ELea^ (°°) = 7 aK'(7) 7  a Proof Replacing 7 by a in (5.7) and using ( 5.8), we obtain 1 (I 1  ^ e('ra) x,3 (x)dx) (L ) ELea^*) = a \\\ f
using integration by parts as in (3.6) in the last step . Inserting (5.12), the result follows. u
Notes and references The results of this section are classical, with Lundberg's inequality being given first in Lundberg [251] and the CramerLundberg approximation in Cramer [91]. Therefore, extensions and generalizations are main topics in the area of ruin probabilities, and in particular numerous such results can be found later in this book; in particular, see Sections IV.4, V.3, VI.3, VI.6.
The mathematical approach we have taken is less standard in risk theory (some of the classical ones can be found in the next subsection). The techniques are basically standard ones from sequential analysis, see for example Wald [376] and Siegmund [346].
5a Alternative proofs
For the sake of completeness, we shall here give some classical proofs, first one of Lundberg's inequality which is slightly longer but maybe also slightly more elementary:
74 CHAPTER III. THE COMPOUND POISSON MODEL
Alternative proof of Lundberg 's inequality Let X the value of {St} just after the first claim , F(x) = P(X < x). Then , since X is the independent difference U  T between an interarrival time T and a claim U, ,3+ry F'[7} = Ee7 ( UT) = Ee7U • Ee7T = B['Y] a = 1' where the last equality follows from c(ry) = 1. Let 0( n) (u) denote the probability of ruin after at most n claims. Conditioning upon the value x of X and considering the cases x > u and x < u separately yields
,0(n +1) (u) = F(u) +
Ju
0 (n) (u  x) F(dx).
We claim that this implies /,(n) (u) < e 7u, which completes the proof since Vi(u) = limniw 1/J(n) (u). Indeed , this is obvious for n = 0 since 00)(u) = 0. Assuming it proved for n, we get
„/, (n+1)(u) <
F(u) + e7u
00
Ju
e7(u=) F(dx)
00
<
f
e7x F(dx)
+ fu
e  7(u z) F(dx)
u
o0
= e 7uE[ 'Y] = e 7u.
Of further proofs of Lundberg's inequality, we mention in particular the martingale approach, see II.1. Next consider the CramerLundberg approximation. Here the most standard proof is via the renewal equation in Corollary 3.3 (however, as will be seen, the calculations needed to identify the constant C are precisely the same as above): Alternative proof of the CramerLundberg's approximation Recall from Corollary
3.3 that
(u) = )3
J OO B(x) dx + J U Vi(u  x)/3 (x) dx.
u 0
Multiplying by e7u and letting Z(u) = e7" O(u), we can rewrite this as
u Z(u) =
z(u) = e7u/
J
B(x)dx, F(dx) = e7x,QB(x)dx,
u
z(u)
f +
J
e7(ux ),Y' 1 • l•(u  x) • e7'/B(x) dx,
0
= z(u) +
J0 u Z(u  x)F(dx),
6. MORE ON THE ADJUSTMENT COEFFICIENT 75
i.e. Z = z+F*Z. Note that by (5.11) and the Lundberg equation, ry is precisely the correct exponent which will ensure that F is a proper distribution (IIFII = 1). It is then a matter of routine to verify the conditions of the key renewal theorem (Proposition A1.1) to conclude that Z (u) has the limit C = f z(x)dx/µF, so that it only remains to check that C reduces to the expression given above. However, µF is immediately seen to be the same as a+ calculated in (5.10), whereas
L
00
z(u) du =
f
J
/3e7udu "o
J "o B(x) dx = J "o B(x)dx J y,0eludu
u 0 0
B(x)^ (e7x  1) dx = ^' (B[7]  1)  As] [0 µs] = l y P^
using the Lundberg equation and the calculations in (5.11). Easy calculus now gives (5.6). u
6 Further topics related to the adjustment coefficient
6a On the existence of y
In order that the adjustment coefficient y exists, it is of course necessary that B is lighttailed in the sense of I.2a, i.e. that b[a] < oo for some a > 0. This excludes heavytailed distributions like the lognormal or Pareto, but may in many other cases not appear all that restrictive, and the following possibilities then occur: 1. B[a] < oo for all a < oo. 2. There exists a* < oo such that b[a] < oo for all a < a* and b[a] = 00 for all a > a*. 3. There exists a* < oo such that fl[a] < oo for all a < a* and b[a] = 00 for all a > a*. In particular , monotone convergence yields b[a] T oo as a T oo in case 1, and B[a] T oo as a f a* in case 2 (in exponential family theory , this is often referred to as the steep case). Thus the existence of y is automatic in cases 1 , 2; standard examples are distributions with finite support or tail satisfying B(x) = o(eax)
76 CHAPTER III. THE COMPOUND POISSON MODEL
for all a in case 1, and phasetype or Gamma distributions in case 2. Case 3 may be felt to be rather atypical, but some nonpathological examples exist, for example the inverse Gaussian distribution (see Example 9.7 below for details). In case 3, y exists provided B[a*] > 1+a*/,3 and not otherwise, that is, dependent on whether 0 is larger or smaller than the threshold value a*/(B[a*]  1). Notes and references Ruin probabilities in case 3 with y nonexistent are studied, e.g., by Borovkov [73] p. 132 and Embrechts & Veraverbeeke [136]. To the present authors mind, this is a somewhat special situation and therefore not treated in this book.
6b Bounds and approximations for 'y
Proposition 6.1 ry <
2(1  aps) 2µs
OMB PB)
Proof From U > 0 it follows that B[a] = Eea' > 1 + µsa + pB2)a2/2. Hence 1 = a(B[7]  1) > Q (YPB +72µs)/2) = 3µs + OYµa2) 2 (6.1) 7 'Y from which the results immediately follows. u
The upper bound in Proposition 6.1 is also an approximation for small safety loadings (heavy traffic, cf. Section 7c): Proposition 6.2 Let B be fixed but assume that 0 = ,3(77) varies with the safety loading such that 0 = 1 Then as 77 .0, µB(1 +rl) 2) Y = Y(77) 277 PB Further, the CramerLundberg constant satisfies C = C(r1)  1. Proof Since O(u) + 1 as r7 , 0, it follows from Lundberg's inequality that y * 0. Hence by Taylor expansion, the inequality in (6.1) is also an approximation so that OAY]  1) N Q (711s + 72µB2) /2) = p + 3,,,(2) B 'y 7 2 2(1  p) _ 271µB
QPB PB)
6. MORE ON THE ADJUSTMENT COEFFICIENT 77
That C 4 1 easily follows from y 4 0 and C = ELe7V°O) (in the limit, b(oo) is distributed as the overshoot corresponding to q = 0 ). For an alternative analytic proof, note that C  1P = rlµB 73B' [7]  1 B' [ry)  1/0 711µB µB +7µB2 )  µB(1 +77 ) 'l = 1. 277q
77
7PBIPB
 77
13 Obviously, the approximation (6.2) is easier to calculate than y itself. However, it needs to be used with caution say in Lundberg's inequality or the CramerLundberg approximation, in particular when u is large.
6c A refinement of Lundberg 's inequality
The following result gives a sharpening of Lundberg 's inequality (because obviously C+ < 1) as well as a supplementary lower bound:
Theorem 6 .3 C_eryu < ,)(u) < C+ eryu where
= B(x) = C_ x>o f °° e7( Yx)B(dy )' C+
B(x) xuo f 0 e'r( vx)B(dy)
Proof Let H(dt, dx ) be the PLdistribution of the time r(u) of ruin and the reserve u  S7(„)_ just before ruin . Given r(u) = t, u  ST (u) = x, a claim occurs at time t and has distribution BL(dy)/BL(x), y > x. Hence ELe7£(u) 0
J
°o
H(dt, dx)
fX
eY(Y x) 00 f°° B(dy) x
BL dy
BL(x)
o
f
f H(dt, dx)
L ^ H(dt, dx) f e7B( x)B(dy) Jo oc, < C+
J0 0 o" H(dt, dx) = C. o" J
The upper bound then follows from ik(u) = e7uELeVu), and the proof of the u lower bound is similar.
78 CHAPTER III. THE COMPOUND POISSON MODEL
Example 6.4 If B(x) = eax, then an explicit calculation shows easily that B(x) _ e6X fz ° e7(Yx)B(dy) f x' e(6,6)(Yx)8esydy = 5 = P. Hence C_ = C+ = p so that the bounds in Theorem 6.3 collapse and yield the exact expression pey" for O(u). u The following concluding example illustrates a variety of the topics discussed above (though from a general point of view the calculations are deceivingly simple: typically, 7 and other quantities will have to be calculated numerically). Example 6.5 Assume as for (3.1) that /3 = 3 and b(x) = 2 .3e3x + 2 .7e7x, and recall that the ruin probability is 24 5su 5eu + 3e *(u) = 3 Since the dominant term is 24/35 • e", it follows immediately that 7 = 1 and C = 24/35 = 0.686 (also, bounding aS" by a" confirms Lundberg's inequality). For a direct verification, note that the Lundberg equation is
7 = /3(B['Y]1)
= 3\
2.337
+2.7771
which after some elementary algebra leads to the cubic equation 273  1472 + 127 = 0 with roots 0, 1, 6. Thus indeed 7 = 1 (6 is not in the domain of convergence of B[7] and therefore excluded). Further, 1P = B [7] 181B = 13 2.3+2.71 = 1 3 1 7 I 7'
_ 17
2 (3 a )2 + 2 (7  a)2 «=7=1 2 1p _ 7 _ 24
36 '
3.171 35* 36 For Theorem 6.3, note that the function QB[Y]1 f°°{L 3e_3x+
u
• 7e7x 1 dx
J
3 + 3e4u
f 0c, ex .
I 2 . 3e3x + 2 . 7e7x l dx
l J
9/2 + 7/2e4u
7. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 79
attains its minimum C_ = 2/3 = 0.667 for u = oo and its maximum C+ = 3/4 = 0.750 for u = 0, so that 0.667 < C < 0.750 in accordance with C = 0.686.
Notes and references Theorem 6.3 is from Taylor [360]. Closely related results are given in a queueing setting in Kingman [231], Ross [308] and Rossberg & Siegel [309]. Some further references on variants and extensions of Lundberg's inequality are Kaas & Govaaerts [217], Willmot [382], Dickson [114] and Kalashnikov [218], [220], all of which also go into aspects of the heavytailed case.
7 Various approximations for the ruin probabil
ity
7a The BeekmanBowers approximation
The idea is to write i (u) as F(M > u), fit a gamma distribution with parameters A, 6 to the distribution of M by matching the two first moments and use the approximation
0(u)
f
u
Sa
r(A)
xa  leax dx.
According to Corollary 3.5, this means that A, 8 are given by A/S = a1, 2A/52 = a2 (2) PIB3) ^ZP(B)2 __ PPB a2 al 2(1  P)PB 3(1  P)µ8 + 2(1  p)2' i.e. S = 2a1 /a2, A = 2a2 1/a2.
Notes and references The approximation was introduced by Beekman [60], with the present version suggested by Bowers in the discussion of [60].
7b De Vylder's approximation
Given a risk process with parameters ,(3, B, p = 1, the idea is to approximate the ruin probability with the one for a different process with exponential claims, say with rate parameter S, arrival intensity a and premium rate p. In order to make the processes look so much as possible alike, we make the first three cumulants match, which according to Proposition 1.1 means p=AUB1=P1,
2N
(2) 6^= =OP
,
/3,4)
.
the approximating risk process has ruin probability z. we have according to the PollaczeckKhinchine formula in the form (3./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 ./3)PBo PB .)3 )PBo µB  . 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. Notes and references The approximation (7.g.p = (/3max 0)µB. 1924. but has an obvious interpretation also in risk theory: on the average.PBo [s (/3max .8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . numerical evidence (e.3 )1 } _ 1p 1 .p + p { 1 1p ti 1 .Ps(/3max . cf./3)] 1 . That is.1.P .3* /S.7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . and hence the ruin probability approximation is b(u) e(bAln)u.1. heavy traffic conditions mean that the safety loading q is positive but small. [174]) shows that it may produce surprisingly good results.2) was suggested by De Vylder [109]. Proposition 1. Letting Bo be the stationary excess life distribution. or equivalently that /3 is only slightly smaller than /3max = 1/µ8. Though of course it is based upon purely empirical grounds.3 and Corollary 3. p* _ . the premiums exceed only slightly the expected claims. (/3max . THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 .1 As /3 f Nmax. Mathematically.(bA*)". Proposition 7. Grandell [171] pp.s(/3max .(3)2 P PB 2µB 2µB Letting /3* = /3/P. we shall represent this situation with a limit where /3 T fl but B is fixed.80 CHAPTER III.p .b(u) = p*e.
In the setting of risk theory. These results suggest the approximation Vi(u) e6(0_.2 If . as well as it is needed for the interpolation approximation to be studied in the next subsection. or equivalently that 0 is small compared to µB .7.Q T /3max.0)u.3) is reasonable for g being say 1020% and u being small or moderate. This follows since rl = 1/p . . light traffic conditions mean that the safety loading rl is positive and large . Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. the term light traffic comes from queueing theory. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. while the approximation may be far off for large u.p _ 2rl11B PB p.2.1 1 . VIII). light traffic is of some interest as a complement to heavy traffic .g. the premiums are much larger than the expected claims .p. u * oo in such a way that (3max .l3)M > (/3max . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . the first results of heavy traffic type seem to be due to Hadwiger [184]. However . Mathematically. but has an obvious interpretation also in risk theory: on the average . in risk theory heavy traffic is most often argued to be the typical case rather than light traffic .ze a2unµB laB (7./3)u * v./3)u). Numerical evidence shows that the fit of (7. we shall represent this situation with a limit where 3 10 but B is fixed. obviously Corollary 7.ryu . then P(u) 4 e6„ Proof Write z'(u) as P((/3max .B AB ) 6()3max _'3) = However . That is . and hence 2µ2B 1 . 7d The light traffic approximation As for heavy traffic . Of course. It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. 2 provides the better mathematical foundation.4) suggested by the Cramer Lundberg approximation and Proposition 6. [APQ] Ch. We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there .
3 is the same which comes out by saying that basically ruin can only occur at the F(U . i. 10 ( u The alternative expressions in (7. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. [97].5) follow by integration by parts. . z/' (u) convergence P(U . U > u] = /3iE(U .u)+. 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 .e. Indeed.5) u Proof According to the PollaczeckKhinchine formula.u. see Daley & Rolski [96].(3 10. Again.T > u). Omax max m.T > u) = J o" B(x + u)/3eax dx . 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined.3 As . (7. Asmussen [19] and references there. For a more comprehensive treatment. by monotone time T of the first claim . Light traffic does not appear to have been studied in risk theory. THE COMPOUND POISSON MODEL Proposition 7.82 CHAPTER III. 0(u) /3 J B(x)dx = /3E[U .Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. ( 3 J O B dx. Sigman [347]. En'=2 • • • = O(/32) so that only the first terms matters. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69].= 1 aJ 1 a 0+ 1 = = p. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . cf. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. u Note that heuristically the light traffic approximation in Proposition 7. the Poisson case is much easier than the renewal case.
_(E) (u). [84]. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). one may hope that some correction of the heavy traffic approximation has been obtained. The adaptation to risk theory is new. ) M. .Wmax f(x ) dx + pee6mQ. B(2). Thus . ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). Notes and references In the queueing setting .VHT) ( ax QmQ ) h (B) ( . where further references can be found .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. (U).O0 M. available. "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . however. Let OLT) (u) denote the light traffic approximation given by Proposition 7. that is. 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. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . f / Qmax B(x)dx 00 eQmaxxdx 4/ Qmax 00 QmaxQ amaze" and the approximation we suggest is J B(x) dx = cLT(v) (say).O(E)(u) 1 (1 . Another main queueing paper is Whitt [380].6) is . the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. Substituting v = u(./3)) . Instead.8. to get nondegenerate limits .3n.x . . with rate 1/µB = /3max. (7. z/i(E) (u) = pe(QmaxQ)u. even if the safety loading is not very small. 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. Al .3). ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. we see that the following limits HT) (u').3 and use similar notation for %(B) (u) = (u). no empirical study of the fit of (7.
Bill is said to be convexly smaller than B(2) (in symbols. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. U(2) distribution B(2) and U(1) < U(2) a. cf. then Bill = B(2). for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. XI. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. we have the convex ordering. U(2) such that U(l) has distribution B('). Proposition 8. whereas (consider x2) B(2) has the larger variance. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. B(') <i. the proof is complete. B(2)) if f fdB(1) < f fdB(2) for any convex function f. In terms of the time to ruin. B(2) and PB(1) = µB(2). . one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. we shall need various ordering properties of distributions.s. this ordering measures difference in variability. 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). then i.ill(u) < V)(2) (U) for all u. Rather than measuring difference in size. Proof According to the above characterization of stochastical ordering. we can assume that 1) < St 2l for all t. Here convex ordering is useful: Proposition 8.6. or the existence of random variables U(l). B(') <d B(2)) if B(1)(x) < B(2)(x) for all x.1 If B(') <d B(2).84 CHAPTER III. B(' <. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. then . Finally.1 is quite weak.2 If B(') <j. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. Proposition 8. u Of course. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. In the literature on risk theory. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. THE COMPOUND POISSON MODEL To this end.' 1)(u) < V)(2) (U) for all u. Taking probabilities.
say to p.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean).4 Let D refer to the distribution degenerate at 'LB . and consider the following claim size distributions: B1: the standard exponential distribution with density ay.2 is the following: Proposition 8. The heavy traffic approximation (7.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. Proof Consider the light traffic approximation in Proposition 7. Hence by the PollaczeckKhinchine formula . with fixed mean. Bo1) <_d Bo2) which implies the same order relation for all convolution powers. from which the result immediately follows. A partial converse to Proposition 8. . B. A first attempt would of course be to identify 'variation' with variance. 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.(1) (. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities.. The problem is to specify what 'variation' means..4) certainly supports this view: noting that. B(2).u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. Example 8. Proof If f is convex.3 provides another instance of this. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. A general picture that emerges from these results and numerical studies like in Example 8.8.1 and µB at 1 so that the safety loading 11 is 10%. Then V. B(2). then B(1) <. and here is one more result of the same flavor: Corollary 8.e.3 If B(1) <. we have by Jensen 's inequality that E f (U) > f ( EU). This u implies that D <.6 Fix /3 at 1/1.5 If '0(1)(u) < p(2) (U) for all u and a. larger variance is paramount to larger second moment. then /'(')(u) < 0(2)(u) for all u. (D) (u) < O(B) (U ) for all u. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal.1. Corollary 8.
9A2e'2r where A.lA.1358. In terms of variances o2. Let ua denote the a fractile of the ruin function. A standard example from queueing theory is . THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.001 u0. and consider a = 5%.000. [166]. and this is presumably a consequence of a heavier tail rather than larger variance.01%. Note to make the figures comparable. 11 Notes and references Further relevant references are Goovaerts et al. 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). 9 Sensitivity estimates In a broad setting.1%.4. = 0. 1/)(u. However.e. A2 = 3.01%. B3: the Erlang distribution with density 4xe2x. B. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. B3 the comparison is as expected from the intutition concerning the variability of these distributions. Pellerey [287] and (for the convex ordering) Makowski [ 252]. For B1i B2. 0. the behaviour of which is governed by a parameter 9. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). One then obtains the following table: U005 U0. i. 1%.0' U0. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable.e'\1x + 0. sensitivity analysis (or pertubation analysis) deals with the calculation of the derivative (the gradient in higher dimensions) of a performance measure s(O) of a stochastic or deterministic system.86 CHAPTER III.. We return to ordering of ruin probabilities in a special problem in VI. Kluppelberg [234]. 0. van Heerwarden [189]. B4: the Pareto distribution with density 3/(1 + 2x)5/2.) = a. in comparison to B2 the effect on the ua does not show before a = 0. all distributions have mean 1.4142.
Proof This is an easy time transformation argument in a similar way as in Proposition 1. with 0 the vector of service rates at different nodes and routing probabilities. a/3 0 . Then ib = Pe(613)u. t]. we may be interested in a'/ap for assesing the effects of a small change in the premium. Then if t is large . and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. if = a e(6A)u. increasing in u.. Similar conclusions will be found below. Thus at p = 1. Let R(P) = Rtli.01/2u. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . i.e. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. a0 as ao 80 19P . Then the arrival rate /3(P) for { R(P) } is )31p.2 Consider a risk process { Rt} with a general premium rate p. while /3 = j3 is an estimate.1. Thus. the premium rate p and the claim size distribution B.(u) for large u.3.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). the distribution of %3 0 is approximatively normal N(0„ Q/t).9.Ap). u Proposition 9. In the present setting. Example 9. Then a p ao = 00 Qa/. SENSITIVITY ESTIMATES 87 a queueing network. and s(9) the expected sojourn time of a customer in the network. a2/t). the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . where Q2 = fl ( l2 1113 / _ Ou2v)2.19P a/ . obtained say in the natural way as the empirical arrival rate Nt/t in [0. say estimated from data. Assume for example that 8 is known. For example. where the partial derivatives are evaluated at p = 1. In particular . 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. it follows that ' is approximatively normal N(0.
Consider first the case of 8/8/3: .88 CHAPTER III. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . Differentiating w. Similar notation for partial derivatived are used below. Viei '0(.g.3.3) follows by straightforward algebra.4). THE COMPOUND POISSON MODEL As a consequence. so that heuristically we obtain '00 50ryu = Coe"u .t.r. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. ^) . (9. 9. and write yp = 8y/8/3 and so on .6 below for some discussion of this assumption). namely that of a twoparameter exponential family of the form Bo. However .3 or/and B. Consider first the adjustment coefficient y as function of 3. but we shall concentrate on a special structure covering a number of important cases.3 70 = 'Ye = = 7 /3(1 we(e +'y. for the ruin probabilities . 5) (Q+'Y)[we(0+7. u Now consider the ruin probability 0 = 0 (u) itself. we can rewrite the Lundberg equation as w(9+ y.w(6.Owe (9.6) As will be seen below. Proposition 9. this intuition is indeed correct.2) (see Remark 9. and the proofs of (9. we cannot expect in general to find explicit expressions like in Example 9.3.5) are similar.()^ 1 . it suffices to fix the premium at p = 1 and consider only the effects of changing . but must look for approximations for the sensitivities 0.()wC(e. 3) ( 9 . (3+'y)PC (0+7.uypCe7u urypO.()(0 +'0) ' (9 . () Proof According to (9.0 = t/'(u) and the CramerLundberg constant C. e. In the case of the claim size distribution B.^)] 1(/3+y)we(9+'y.(/3 + y)we(9 + 7. (. x > 0 (9.()YC = 1 +y/ /3 \ Q2 From this (9. (9.w(O. () = log(1 + y//3). Of course.1 or Proposition 9.((dx ) = exp {Ox + (t(x) . 4) (9 . /3 yields w e(e + Y. ()} p(dx) .10) below. various parametric families of claim size distributions could be considered.
()] exp {w (O + y. By dominated convergence.w(9.x)B(x)dx. Z(u)/u a C//3PF where PF is the mean of F. we note the formulas Ee. () .9.QB(x) dx. we multiply by e7" and let Z(u) = elt" cp(u).3 (see in particular (5.x) F(dx ) f u J C F(dx) = C as u 4 oo. () exp {w(9 + a. () .8) (Section 5). 11 For the following. O} (9. Hence by a variant of the key renewal theorem (Proposition A1.10) (9.9) (9. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.we(9 . Be.8).([a] = exp {w(9 + a.St (U) Ee. () . ()} .x). 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. z2(U) = e7" J u b(u .x). u 0 Then Z = z + F * Z and F is a proper probability distribution . (9. PF = (1 .4 As u oo.(e"U = = wS(O. and alsoo zl(u) + 0 because of B['y ] < oo.12)).3(x) dx.C). F(dx) = e'yy/3B(x)dx. BarndorffNielsen [58]). () .g.w(9. Z= zl + z2 where zl (u) = e7u J m B(x)dx. Further write de = [we (9 +'y.w(O. 0(u) = /3 Ju"O B(x) dx + f 0 0(u . z2(u) _ 1 ^ e'ri`i7i( u .2 of the Appendix ). SENSITIVITY ESTIMATES Proposition 9. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9. ()} . But from the proof of Theorem 5.3) for z/'(u). Combining these estimates .8) Letting cp = e0/e/3 and differentiating (9.x)B(x) dx + J U W(u . the proof is complete.p)/C'y. w((9 + a.11) Ee.
0 x Multiplying by e7" and letting Z(u) = e"uV(u).9)(9. ()} 1z(dy) = f [t(y) .90 CHAPTER III. ()]B(dy) dx. )}B(dy)• Letting cp it thus follows from (9. 01 (i+) do = +'Y. C) . 8 8() 8( (9. ^) . C)] (1 + 7 ) Proposition 9.x)f3 f ^[t(y) .w( (0. ^)} [wc (0 + 7.w(0. ()]B(dy) dx x 0 0C T ON O .1) B(dy) 'f '[t(y) . 8^ ue7u.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . oo z2 (u) f C .6e7u f "o f[t(y) .8) that cp(u) . ()]e7vB(dy) 'fCd 7 c .12) f exp {O y + (t(y) . THE COMPOUND POISSON MODEL [we(e+7.11).6C do 89 1p 8( 1p Proof By straightforward differentiation.we (0.QB(x)dx.5 Assume that (9.wc(O.e7x/3 f 00 [t(y) . ()](e7v . By dominated convergence and (9.w( (0. z = zl + z2.wc(9.wc (O. Then as u > oo. F(dx) = e7x.x). 2 z 07P N ue7u (3C de . u Z2(U) = e7° f u ^/i(u . this implies Z = z + F * Z.w (9. ())B(dy) dx.w(e.2) holds.
2) holds with p(dx) = xldx.18) (05 + 57 _'3_y . Z(u) /3C 91 o c'o e11(t (y) .Sry a/32 + a/37 + /37 .3ary tog('Finally.1edz = 1 exp {Sx + a log x .a/35a&y' ' (9. U 7µF from which the second assertion of (9. and the proof of the first one u is similar.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 .13) (9.rye) S 5rya. 9 = S. ue_Yu 'C2d( 8a 8( 1 p .) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function. by inserting in the above formulas.12) follows. Example 9. () = log r(a) .pa+1 . t(x) = logx.9.16) (9.6 Consider the gamma density b (x ) = Sa xa. .ry) 5a1 cry (5 .w((9.15) (9. we (9..12) takes the form y) alp a .yu/3C2do u86 89 1p' az/) = 8z/. Here (9./35' a/i'y + aryl 625ry. () = C/9 = a/S. that C = a.(log r(a) a log S)} • r(a) 1.QS 1 . < = a. SENSITIVITY ESTIMATES as u 3 oo. ())B(dy) < oo.. ( 9.1 . w(e. It follows after some elementary calculus that p = a)3/5 and.17) (9. a /(S . () ='I'(t. We get w( (0.a log S = log r(c) .C log(9). and also zj (u) 4 0 because of f Hence..Y)a+1 ' (9.
log c = 2 In particular.21og 2. C = . which we omit in part .S[a] = exp {w (9 + a. () = Cc .1 = eXP {c(C .2 .2a) } Thus the condition B[a*] > 1 + a* /.2) with µ(dx) = 2x3zrdx.92 CHAPTER III.l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.22.7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9.w(9."62 . w(e. C) . for a < a* = z (. ()} = exp {c (C .3. Straightforward but tedious calculations . Be.9) (() . THE COMPOUND POISSON MODEL Example 9. t(x) _ .CZ try)} 1 C C2 try . further yield .3Ee. 9 = .2 log (0. C) = B = Yc = de = do = .1 16 +ry c C22ry 2( = + 70 We (e.([Y] .
10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3. That it is no restriction to assume one of the ti(x) to be linear follows since the whole setup requires exponential moments to be finite (thus we can always extend the family if necessary by adding a term Ox).cue_7u)3C P Remark 9. the exponent is either Ox. to our knowledge.12) takes the form a = a 93 ar. sj=1 and let YT be defined by IKT('ryT) = 0.a. However. for which we refer to X. the results presented here are new. Note that if NT = 0. then ryT < 0. let NT 16T = ^T . in u which case the extension just described applies. we have assumed k = 2 and ti (x) = x. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). ae t 1lEY u S _ . However . we can just fix k . To this end.10. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . BT [a]= NT ^` e"U.2 of the parameters. kT (a) = /T (BT [a] . Thus. and we estimate y by means of the empirical solution ryT to the Lundberg equation. .. Finally if k = 1.2) is motivated as follows. (9.oo. Also.3C2de 1p' z a = c . the main tool is simulation. queueing networks) are typically much more complicated than the one considered here. Thus. That it is no restriction to assume k < 2 follows since if k > 2. the models there (e.g. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. then BT and hence ryT is undefined. In general.. [379] consider a special problem related to reinsurance.8 The specific form of (9.. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. in which case we can just let t(x) = 0.7 and references there. or Ct(x).+UNT) > 1. Comparatively less work seems to have been done in risk theory. Van Wouve et al. thus.1) . B are assumed to be completely unknown. and hence explicit or asymptotic estimates are in general not possible. if 1 PT = /3TNT(U1+.
1 As T 4 oo. since NT /T .2) follows from NT/T a4' .2 As T * oo. THE COMPOUND POISSON MODEL Theorem 10 . For the proof.'Y . (10.1) .B[7]2 }) ( T 0 .i3)(B[7] 1) + (3(BT[7]  .: N ()3. B[27] ./^ B[27] .Q and Anscombe 's theorem.2) rT(7) N N (0.3T .1)2 + E[27] . B[2'Y]  /3T ) . it is easy to see that we can write \ V1 1 l _ . N ( n[7]. then (10.B[7]2 V2 .b[Yp'V21 T { (E[7] .1) 'YT . 1) r.B[7]) + B [7] .3) Proof Since Var(eryU) = we have B[7].94 CHAPTER III.3/ T). we need a lemma.If . vfoVFB[2y].B[7]) 0+ Iv/o(b[y]. More generally.: N 0.(27)/K'(7)2.B[7]2 n Hence ( 10. 16T where V1.)vl+ N CO. Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] .1) .. V2 are independent N (0. Lemma 10 .7 + (. a2 where a2 = /3r.T y . . If furthermore B[27] < oo.v. B [7]2 (10.'s. 7T a4' 7.a BT[7] I B[7] I + .
E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r.E) < 0 < kT(7 + E) for all sufficiently large T .3).1 can be used to obtain error bounds on the ruin probabilities when the parameters . it follows that 7T7 KT(7T) .c'(7) N (0' T (2(7) / N (0. I. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10. °7IT) .10. OT a 95 u 4 /3.e.1 By the law of large numbers. 7T E (y . Combining ( 10.4) + E). NT i =1 n'(a) for all a so that for all sufficiently large T K7 . BT[a] 3 B[a]. Proof of Theorem 10. lcT(a ) 4 /c(a). first note that e7TU N (e7U u2e27Uo'2/T) 7 .Q.2.(ry . y + E) eventually. Theorem 10.4) and Lemma 10. NT BT [a] Hence r.E) < 4T(7T) < 4T(7 + E). where ryT is some point between ryT and ry. Then r.(ry + e) and hence KT(7 .'T(a) = 1 E Uie°U' a$' EUe "u = B'[a].'(y). If ryT E (7  we have KT(7 . 0 are estimated from data . 6"Y (10..KT(7) kT(7) K'(7) . and the truth of this for all e > 0 implies ryT at 'y.e. By the law of large numbers.e) < 0 < r. To this end . Let 0 < E < ry. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).
.i. U2.e. 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. t]}.T VIT where r7ry. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. Letting Wo = 0. ..d. > 0 for some t E [Wn_ 1. i.Wn) are i .Q. Hipp [196]. various alternatives have been developed.g. if B is exponential with rate 8 so that ry = 8 ...1 is from Grandell [170]. and the known fact that the Y„ = max Vt tE[W„1. [197]. i . 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.info< „< t S. V.. For example ..96 CHAPTER III.1 : Vt = 0. Notes and references Theorem 10.f. ft. Asmussen [23]) can then be used to produce an estimate of ry. satisfies b(.. Frees [146]. Vt = St . 6 < 2. This approach in fact applies also for many models more general than the compound Poisson one.ueryuU ".g.96 if a = 2.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. Mammitzsch [253] and Pitts. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94].C1e"a ( see e. Deheuvels & Steinebach [102]. the nth busy cycle is then [Wn1.. Wn).3 or equivalently p > 1/2 or 11 < 100%. with a tail of the form P(Y > y) . Griibel & Embrechts [292]. Embrechts & Mikosch [133]. For this reason . Csorgo & Teugels [95]. = 1.) = a (e.e. Herkenrath [192].5%).0) < 5. it means 2 (8 .
s.1) . The safety loading is q = 1/p .g. the premium rate is 1.1 (the role of ryy will be explained in Section 4b). defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . Unless otherwise stated. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y.1 where p = 13µB. exists. In particular. generalizations to other models are either discussed in the Notes and References or in relevant chapters. the Poisson intensity is 0 and the claim size distribution is B with m. Further let 'Yo be the unique point in (0. The notation is essentially as in Chapter III. B[•] and mean AB. See Fig. 97 .Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. 0.f. 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. 'y) where c(a) attains it minimum value.
. Var[T(u) I T(u) < 00] = VarL T( U) . FL and independent of T(u). PROBABILITY OF RUIN IN FINITE TIME Figure 0. E[T(u) I T(u) < 00 ] = ELT (U). the time of ruin is T(u) and ^(u) = ST(t&) .1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. .) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k. 1 FL.9).(U) < 00] = ELT(u)ke'YS. PL = 6/0 = 1/p > 1).t. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus . 7.(.1 The claims surplus is {St}.u is the overshoot.5 .r. we have for k = 1. using that the overshoot l. 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. In particular.(4. 2 that E [T(u)k. (u) is exponential with rate 0 w. By the likelihood identity III.98 CHAPTER IV. 1 Exponential claims Proposition 1.1) (1.2) Proof Let as in Example 111.
1) .h.1)T(u))2 = UL where = s.0) . Since Sr (u) and (PL . of (1.(PL . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).h. Let 0 > yo be determined by ^c(0 ) = a.1)T(u)) = VarLe(u) + (PL .(yo) = 2V .1//32 (6/)3 1)2 26(/3u + 1)/(6 .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.I (1.h./3) . T(u) < oo] fora > r. 0 Proposition 1."(ry) = 26//32. we have by Wald's identity that (note that ELSt = t(pL . Wald's second moment identity yields 2 EL (Sr(u) . EXPONENTIAL CLAIMS For (1 .B = a.6a = 0 with solution 0 (the . u + ELe(u) _ PL .2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0. This means that /3(6/(6 .1)T(u) are independent with QL the same mean .s. .6 + a)0 .1)ELT(u).2). the 1./3 .s.1.2) is aLELT( u) .1 (6)3)2 which is the same as the r.s. which leads to the quadratic 02 + (/3 . where = eBu I 1 . is V1rLSr( u) +VarL ((PL .12 Thus the l.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 .6.1)) ELST(u) ELT(u) (PL .1)2VarLT(u) + 2 Ca 1I VarLT(u). 1).V/ is as asserted.
are the lengths of of the ladder segments 2. T(u) < oo] = EB [exp {aT(u ) .. .1 where Y1. 1.. the result follows. Y2.. T2 . and M(u)+1 is the index of the ladder segment corresponding to T(u). Using 5 = 6 .100 CHAPTER IV. More precisely. Ti. Y(u) belonging to a convolution semigroup . T(u) < oo] = e.9ST(u) +T(u)!c(0)} .... M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment .4. are the ladder heights which form a terminating sequence of exponential r. . PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).3 that we can write EeaT( u) = eeuEe 017(o). St Ti F.0. Cf.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. (1.Y1 Y2 Figure 1.v. Fig.4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r.T+ Ti a t U T I 1 a i F..'s with rate 5. Note that it follows from Proposition 1.3) we have E [e«T(u ).v..3. But by the fundamental likelihood ratio identity ( Theorem 111.1 .
3 Assume that claims are exponential with rate b = 1. then VT = U1.ST). Hence 00 F(VT > u ) P(QT = N)P(EN > u) N=1 00 N1 k F(QT = N) eu N=1 k=1 °O u k! k Ee k=0 1t P(QT .T..T are conditionally i.I ex cos B cos j O dB fo " . i.1(u.T) 1 I fl(O)h(0) fdO where (1.. U2. ..i. EXPONENTIAL CLAIMS 101 For numerical purposes . density xN lex/(N .T the service times of the customers awaiting service . let (cf. Proof We use the formula ..6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) .e. Then V(u. If QT = N > 0. and exponential with rate S = 1.v. UN. T.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. . T) to be evaluated by numerical integration: Proposition 1. EN has an Erlang distribution with parameters (N.T + • • • + UN. UN.3 sin0 + 29) f3(0) = 1+/32/cos9. 1).4.T.1. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V.6(u) = Vfl/j l(Su. cf. . Since U1 . .T.6.cos (u/. .0. where U1.T is the residual service time of the customer being currently served and U2 .. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! .d.k + 1). 2. the following formula is convenient by allowing t.1 )!. Corollary 11. including the customer being currently served).. For j = 0.i (u.. the conditional distribution of VT given QT = N is that of EN where the r.1. Let {Qt} be the queue length process of the queue (number in system.
8 ) yields F(QT > k + 1) . 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = . in particular equations (1.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 .cos (( k + 1)0)] f3(9) Hence the integral expression in (1. PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.ie .i(k +1)e R [/3( klal/2e:0 (01 /2 e .cos((k + 2)9)] d9. 8789) 00 E aj j= 00 = 1.(31/2 cos (( k + 2)9) . f3(0) .31 /2eie L 1)] 1 I/31/2eie . (1.)3k +1 tj g'(QT >.k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.1 00 ok+lR 00 j=k1 +1)/2e .(31/2eie .112 l 1( k +1)/2 [ 31/ 2 cos(kO) .1 R [. similar formulas are in [APQ] pp. Then (see Prabhu [294] pp. and define tj = e(1+R)Taj/2Ij(2vT T). let I _ j (x) = Ij (x). 912.44).38).1)] L _112 /(k+1)/2 [. k k2 + $k+1 E bj 00 t j .)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .102 CHAPTER IV.3(k +1)/ 2ei(k + l)6 (.13(k +l)/2ei(k +1)9 R E .
. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. 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) in terms of F(. going back to Cramer. it follows as in (1. E Fk. from the accumulated claim distribution N. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St.0(u.2. or.e = e' COS a cos(uf31/2 sin 0). t )). expresses V)(0. Ui < x I / (note that P(St < x ) = F(x + t. however. Seal [327] gives a different numerical integration fomula for 1 . however. 2 The ruin probability with no initial reserve In this section . k! k=O k0 i/z Co Uk ate" o'/z e . equivalently.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. the numerical examples in [12] are correct).T) which. u Notes and references Proposition 1. T). The first formula. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. there are several misprints in the formula there. F(x. k=0 103 Cu) A further application of Euler's formulas yields cc k =0 k 'ese)k __ U #kJ2 cos((k + 2)9) = R eNO ^` (u^1 L k= = eup i/z L OI = =ateU161/2 e '0+2iO COS a cos(u(31/2 sin 9 + 20).7) that _ [^ au ak+l (30 k L. Related formulas are in Takacs [359]. is numerically unstable for large T. The rest of the proof is easy algebra. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. oo (u)31/2e^e)k = )3k z cos(k9) = R k. . t) = P .
t)= {Stv) < SM.T].T)dx.T)) 1 fT P(M(v.b (0. T) = P(Tr(0) > T) = P(M(0. 1 1 .T))dv E^T I(M(v. T]. Stv^ _ Define M(v.(6. f T lStv)} 0<t<T by a 'cyclic translation'. Proof For any v E [0.1 In formulas. 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.i. T T o where the second equality follows from II.1. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 .104 CHAPTER IV. ") } is at a minimum at time t. See Fig. [v.T)) does not {Stv)} depend on v.0<w<t} St+v .(.T))dv.3) with A = (0. Then 1 .(0.T) T F(x. resp. we define a new claim surplus process St StM NJ Figure 2. 2. .S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. co ). v].
It follows that if M(0 .v<t<T} = {ST<StSv. We claim that if M(0.T)) dv f T I(M(0.T) and Sv < 0 on M(0. Fig 2.Sv.ST on M (0. we can take v E (w E. we can write M(v. T.2. For example. v) = M(0. Hence T TE f I( M(v. v). v < t < T} n M(0. 0<t <Tv}n{ST<ST Sv+St T+v. where the last equality follows from ST < St on M(0. t) denote the density of F(•.T) occurs or not as long as ST < 0. T) occurs. cf. T Theorem 2 . w) for some small E. T T o i =1 Let f (•. T)) dv. If ST < 0.T) occurs. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. Indeed. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. in which case there is a last time o where St downcrosses level u. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T . v). ST > 0. letting w = inf It > 0 : St_ = mino<w<T Sw}. v).t)dt. this integral is 0 if STv) . there exist v such that M(v. Obviously. 0<t<v} = {ST < St . .T) = F(u+T.. T) = M(0.T)f(I z /)(0. then i fT I(M(v. T) as {ST<St+ vS. T]. T) occurs.2 10(u. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v.Tt))f(u+t.2. 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 .. v<t<T}n{ST<STSv+St. T)). or it occurs. then M(v. t).xdx.
0 < t < T. u + dt] and there is no upcrossing of level u after time t. ST_ _ z}.2 .3 Define r_ (z) = inf It > 0 : St = z}.t). 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. C*(z. t + dt] occurs if and only if St E [u.(0. Proof of Theorem 111. {St > .T)+ J0 T (1V.p. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).106 CHAPTER IV.Tt))P(StE[u.2.2 Here o. Let Z be a r.ST_ t_ and let A(z. u which is the same as the assertion of the theorem. 2. For a fixed T > 0. Hence P(ST<u) = 1 . which is independent of St and has the stationary excess distribution B0. which occurs w.T) = . define St = ST . 0 < t <T. The proof is combined with the proof of Theorem 111. O(T .T) = C(z. ST_ _ z}. {S t > z.b(u.T) = {St < 0.z. ST_ _ z} . Proposition 2. z > 0. 0 < t <T .u+dt]). E [t.2.v.
Fig.3.T) = C*(z.T). T + dT] I S7(o)_ E [z.T))f3B(z) dz dT. Thus P(Sr(o)_>x. z + dz].T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. A(z. Proof of Proposition 2.1) that P(T(0) E [T. ST(o)_ E [z. z + dz]. u which is the assertion of Theorem 111. 2. 7( 0) < oo) = P (C(z)) dT.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz.3 But by sample path inspection (cf. (2.1) z T . T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. we therefore have P(A(z.T)). {St }o<t<T have the same distribution . THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. and since {St}o<t<T. .T + dT]. z + dz]) = P(A(z. T(0) < oo) = OR(z) dz in (2.2.T)) = P(Cx.3).1) yields P(ST(o)_ E [z.2. It follows by division by P(ST(o)_ E [z. z + dz]. z + dz]. Hence integrating (2.2. Figure 2. r(0) < oo) = 3R(z) dz JP(C(z.ST(o) >y.
2. Theorem 2. see in addition to Prabhu [293] also Seal [326].r(a).2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . Notes and references For Theorems 2.6. Proof Optional stopping of the martingale I er (a) 9 t. because of77>0. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = . ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. T(0) < oo) = dTP(T_(Z) E [T.(3(B[r( a)] .(yo).T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. a martingale proof is in Delbaen & Haezendonck [103]. Note that T_ (y) < oo a. who instead of the present direct proof gave two arguments. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. Let T_ (y) be defined as Proposition 2.T+dT]). cf. 3 Laplace transforms Throughout in this section.1 Eear( y) = eyr(a).3 was noted by Asmussen & Kl(ippelberg [36]. r(0) < oo. Lemma 3.5 and one upon excursion theory for Markov processes (see IX.1) . T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.1 and the present proof is in the spirit of Ballot theorems.108 Hence CHAPTER IV.s.3.2. Tak'ecs [359]. [329]. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. I L Let ga(x) be the density of the measure E[ear(°). some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. z + dz].1.1) where a > r. Lemma 3 .c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). z + dz]. In the setting of general Levy processes. Proposition 2. (3.5a).
Corollary 3. u .3. time T(u): u u Here is a classical result : the double m.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)) .x)(a) B(dy)• Lemma 3 .r(a) b . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus . (Laplace transform) of the ruin Corollary 3.2. (u .(v) = ev''(a).r(a) = a [B[b] B[r(a)]] .r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a.3 ga[b] = c(b) Proof + b + a .3.x)ga (x) dx where za(u) = f. b .g.1] evr(a)B(dy)[ b .ic(b)/b x(b) + a eb"E[eaT(" ). rr(0) < oo) = 1_ r(a) Proof Let b = 0. Further by Theorem 111.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 . £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v. Hence eb"du E[eaT(").4 E[eaT (o).ST(o)_ just before ruin .ga [b] 1 . y + dy]. It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z. the result follows after simple algebra.°° ga(x)dx. Z = y] = EeaT.3. r(u) < oo).5 f 00 o a/r(a) . Then by Proposition 2. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ). E[ear (o) I T(0) < oo .ga [b] 0 TO Using Lemma 3.2 P(Z E [y.f.
w ) v/. (u) t. This proves the first assertion of (4.3). That is. we need the following auxiliary result: Proposition 4. Proposition A1.mL > E T(u) < 00 ) 40. i..6. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model. T(u) a.h(u. P = /3µB > 1. Then given r(u) < 00.00 St = lim . mu ) ( 0 m < ML '(u) 1 m > rL. = (p . u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). St/t 1 1/m.s. T(u)/u mL as u + oo. and hence a.2.1 Assume 77 > 0.110 CHAPTER IV. (4.2 Assume ri < 0.r(u) = Er(u) • ES. Theorem 4 . note that by Wald's identity u + EC(u) = ES. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . For the second . For the proof.1) i.(u ) = o(u) a. Later results then deal with more precise and refined versions of this statement.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo.s.e. cf. and take basically the form of approximations and inequalities.1. By Proposition 111. Then as u * oo.mu D 2 4 N(0.1)Er(u) . uoo u using e. for any m T(u) u . the known results are even less explicit than for the exponential claims case. The first main result of the present section is that the value umL. for any c > 0 P( Further.. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S.
t/m D (2) 111 .1. apB ) .g. For (4 . 1'r(U) .1 is standard.3). T(u) < oo f / 00) e7uE L [e_7 (t1).r(u)/m T(u) ti µB2) Z. of (4.1) is T (u)  U mL P( T (u) < I > E. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. If Z . which may be viewed both as a refinement of Theorem 4. and (4. According to Anscombe' s theorem (e. cf. T) for T which are close to the critical value umL). and as a timedependent version of the CramerLundberg approximation. Tu) T( u) .^ N (o. the same conclusion holds with t replaced by r(u).s. though it is not easy to attribute priority to any particular author.1). PL (•)+ 0.h. proving (4.1 The l.mL U > E. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. T (u) < 00 J 0(u) e7'PL U \ I T u) .2. Notes and references Theorem 4.2) follows immediately from u (4.2 of [86]) and (4. this can be rewritten as u + 1(u) . the result comes out not only by the present direct proof but also from any of the results in the following subsections.mu m . Theorem 7.7 6  11 Proof of Theorem 4.3.mL >E By Proposition 4.6.6µB2) Z v m (3µB2) Z. implying T(u) . note first that ( Proposition 111.mu (2) '• m3/2 µB 7 . Thus.5) St .1 (by considering 0(u. . again Proposition A1. 4).4.1).1).N(0.
(oo. we get E[ T (u) .l:(oo) (recall that rt < 0).)mu \ h(oo)Eg (r(ul) . P because of ^(u') .L+YWLV'U) . letting Z be a N(0.VU T. with w2 as in (4.T(u') given F. we can replace T(u) by r(u'). PROBABILITY OF RUIN IN FINITE TIME Corollary 4. Then the distribution of T(u) . E9(Z) (4. resp . Then for any y.3). then e(u) and r(u) are asymptotically independent in the sense that.(u.mul h(oo)Eg(Z).w2) r. Then h(u) 4 h(oo) = E f (6(oo)).6) whenever f. Hence Ef (Vu )) 9 (T(u.))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . S( u ) < ul/4] < ET(ul / 4) = O(ul/4).6).4).v. Proof Define u' = u . one has 9 (r(u)_rnu) Ef (^(u)) * E.5) For the proof. oo ). we need the following auxiliary result: Proposition 4.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u .112 CHAPTER IV.f ( (oo)) .ul/4. and thus in (4.t. Using ( 4. and similarly as above we get E[f(^(u)) I Fr(u.T ( u')] = E[ T ( ul /4 .4 (SIAM'S LEMMA) If 71 < 0.ST( u') = u1/4 . oo). using that ul/4 .e(u') oo w . O .a C4'(y )• ( 4. g are continuous and bounded on [0. Let h(u) = E f (^(u)).u1/4)I(S(u') > u1 /4) h(oo) + 0.r. e'°'/b (u.um.^(T(u')).3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).
where we used Stain's lemma in the third step and (4. 4b Gerber's time. that for the fit of (4.umL wI V"U u (4.z/)(u . u needs to be very large).yK(ay)• (4.4) in the last. e7v" y < ^'(7) (4 .7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . Segerdahl 's result suggests the approximation b(u. see also von Bahr [55 ] and Gut [182].7) to be valid is that T varies with u in such a way that y(T) has a limit in (.7) to be good. in practice one would trust (4. umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V ").T) Ce7"4 (T . 3 is due to Segerdahl [333]. PL(T(u ) < umL + ywL) 113 4 C4(y). also Hoglund [204]. T(u) < umL + ywL f.7) To arrive at this . yy by 1 K. For practical purposes .(ay) = 17 7y = ay . The present proof is basically that of Siegmund [342]. CL Fig. Theorem 4.9) ( 4 . just substitute T = umL + ywL in (4.4.3 ery"z/i(u .5) and solve for y = y(T). y u) < e 7v" . oo ) as u * oo. 0. The precise condition for (4.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML).5 '(u . 10) '5(u) . however . . see Asmussen [12] and Malinovskii [254]. Cf.dependent version of Lundberg's inequality For y > 0.oo. y > k'(7) .3 in terms of Edgeworth expansions . For refinements of Corollary 4.1. ELe7E (") . WHEN DOES RUIN OCCUR? Proof of Corollary 4. Thus . define ay. y u) < . Notes and references Corollary 4 .
b (u. Then ic(ay) > 0 (see Fig . 5 is due to Gerber [156 ]. the point is that we want to select an a which produces the largest possible exponent in the inequalities.h(u. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. From the proof it is seen that this amounts to that a should maximize ayic(a).7 i. and hence t. Hoglund [203] treats the renewal case.r. yu) is e 'Yyu/ .8).1).t. which may be understood from Theorem 4.3 yields easily the following sharpening of (4.8 below .Y' (u. a. For a different proof.2. and generalizations to more general models are given in Chapter VI. which shows that the correct rate of decay of tp(u. dy) Notes and references Theorem 4 . yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. f Some urther discussion is given in XI. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). In view of Theorem 4. who used a martingale argument.9): Proposition 4. we arrive at the expression in (4.v"U. . PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). 0. see MartinLM [257] .yu ) = eayuEav [e . Numerical comparisons are in Grandell [172 ].6. the bound a7y° turns out to be rather crude . yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . yy is sometimes called the timedependent Lundberg exponent. if y > 1/ic'(y). yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay).5. However. u Differentiating w. we have rc(ay) < 0 and get (u) . An easy combination with the proof of Theorem 111.ay4(u)+ T(u)K(ay ).114 CHAPTER IV.8).
WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations. and ii(u) . not inequalities.e..4. T(u) < yu] .5.13) .'(y ).6 with P replaced by Pay and FL by Pay.^3 ]1/ Bay [lay . yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.ay a. As a motivation.. T(u) suggests heuristically that l t/. yu ) eaauEaye . (4. then ay > 0.(u.. then the solution &y < ay of . This idea is precisely what characterizes the saddlepoint method. (4. the choice of ay. Using Lemma 111.c(&) = ic(ay) is < 0. and in case of ruin probabilities the approach leads to the following result: Theorem 4 . [eT(u )K( ay). Proposition 4.e. and b(u. it is instructive to reinspect the choice of the change of measure in the proof. For any a > yo. we have ryas = ay . then the relevant choice is precisely a = ay where y = T/u. The traditional application of the saddlepoint method is to derive approximations. (0) r1 (a) ' I. We thereby obtain that T is 'in the center' of the Padistribution of T(u).: T. if we want EaT(u) .ayC() . the formula 0(u.yu) c ay .ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 .z. (4.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. u 4 oo.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. i.i(u. yu) = e.11) ' If y > 1/ r .ay y 'Yay  ay .8 If y < 1/ic'(ry).ayuEay f eay^ ( u)+T(u)K(ay). Ea .2 yields EaT(u) u u r.
116 CHAPTER IV. V < 01 Ir 00 er(ay)"1'2"'x eyur. Then ic(a) = .13). Example 4.3(5/(S . (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 .1B[ay]1 ) y(ay . we get heuristically that Eay Ler (u)r(ay). .c'(a) _ /3a/(8 . it seems tempting to apply the normal approxiyu + ul/2wV.12) is 0 entirely similar.(j (1 . i B[7ay . and the equation ic'(a) = 1/y is easily seen to have . (4.7ruw2 Inserting these estimates in (4.a)2 .B[ay] /ay &y y(ay .1) under Pay mation (4.13).I ay &y a ^c'(ay) a (1 +. and in part that for the final calculation one needs a sharpened version of the CLT for t(u) (basically a local CLT with remainder term).13) rigorously.l'B)y /(Pay .1.a. where V is normal(0. The difficulties in making the proof precise is in part to show (4. a nr=.9 Assume that B(x) = eay.1)3 = y3/3B"[ay].ay + ayl /BLay] .(ay) _ y(ay . T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.4). PROBABILITY OF RUIN IN FINITE TIME ry I i .a) . The proof of (4.1) .11) follows. Writing r(u) and W2 = I3ay{.c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.ay)K(ay) ay ayI&YI For the second term in (4.ay) ay +.1)3 = (jB"[ay]l (Pay .ay ) r.
then { __ .g.5. 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.f.. is undefined for a > 5). y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.1) . The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.p. (5.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 .= (s..i )( v s vc ('3 + s _2 / . and next to note that such an approximation in particular implies that the first passage probabilities are close..ay)3 0 3/2 and (4. 2 = Var(Si ) the variance.11) gives the expression '31/4 ( . 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 =. .tcp) Lo {Wo ( t)}t>0 . 0 Notes and references Theorem 4./4 ^y for 1/i (u.1. in discrete time: if p = ES. yu) when y < 1/ic'('y) = p/1 . DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .because the c.8 is from Arfwedson [9]. c a 00. is the drift and o. A related result appears in BarndorffNielsen & Schmidli [59].
. of which a particular case is the claim surplus process (see the proof of Theorem 5. such that the claim size distribution B and the Poisson rate a are the same for all p (i. 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. St = EN` U= . However.1) with S. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes).1 As p J. where p is the critical premium rate APBTheorem 5 .2) t>o where p = pp = p .p. Mathematically. + {Wo(t ) . a2 =/3µB2) Proof The first step is to note that { WC (St P) . this is an easy consequence of (5. n/c < t < (n + 1)/c..1 below). We want an approximation of the claim surplus process itself. 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. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p. Indeed .3.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.1)) is inconvenient.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. and this can be obtained under the assumption that the safety loading rt is small and positive.tp). Lemma 111.1. (5. 0 .3) whenever c = cp f oo as p 1 p.z } {W_1(t )}t>o (5.t} _ {W_1(t)} . and consider the limit p j p. for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.p/c < St(p) < S((n+l)/ c + Pp/c.7c). Letting c = a2/pp. p. oo)).a = Snp) and the inequalities Sn )C . we have o {i!t s: . cf.e.118 CHAPTER IV.
h. 196.(u) ti IG(oo.e..u). DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. . and in fact some additional arguments are needed to justify (5.6) from Theorem 5. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a.s.4) Note that IG(.1. u). u) =PIT( (u) < x) = 1 . Corollary 5.( ^ I + e2( \ I .s.6) This is the same as the heavy traffic approximation derived in III. and the r. is IG(T. since ti(u) has infinite horizon . 119 It is wellknown (Corollary XI. 199.1. u) is defective when < 0.h. is 1/ip (ua2 /IpI. For practical purposes .7c.f I \\\ J \ (5. 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.2 As p j p. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r.h. (5. w.5). see Grandell [ 168]. (ua2 To2 op \ IPI > IG ( T . ('. the continuity argument above does not generalize immediately.1 I 7= . we obtain formally the approximation V. ulpI/a2). (5.1 . we omit the details .r. C. 263) that the distribution IG(•.Ta2 /p2).8 or [APQ] p..2 suggests the approximation u 0(u. 1. any probability measure concentrated on the continuous functions.T) IG(Tp2/ a2). Because of the direct argument in Chapter III. Corollary 5 . [169] or [APQ] pp.5) Note that letting T * oo in ( 5. has a continuous distribution. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. (. TS(u)=inf{t>0: WW(t)>u}.5.t.. However. ulpl /a2) = e2"1µl / or2.s.
as 0 * 00 and that the U2 are uniformly integrable w.t.6 of [APQ]. We conclude this section by giving a more general triangular array version of Theorem 5. we have ^A.. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75].5) and (5. In view of the excellent fit of the CramerLundberg approximation.6) therefore does not appear to of much practical relevance for the compound Poisson model. the claim size distribution B9 and the premium rate p9 depends on 0. The picture which emerges is that the approximations are not terribly precise.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. . a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory.00µB6 + 0. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation. Assume further that 039µB6 < pe. Furrer. See for example Billingsley [64]. pe .r.6) are presented. The proof is a straightforward combination of the proof of Theorem 5. in particular for large u. In contrast. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . All material of this section can be found in these references. as an example of such a generalization we mention the paper [129] by Emanuel et al. in the next subsection we shall derive a refinement of (5. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. and which is much more precise.Po = 09µB6 . e. [169]. on the premium rule involving interest. and two further standard references in the area are Grandell [168].120 CHAPTER IV. the B9.5) combined with the fact that finite horizon ruin probabilities are so hard to deal with even for the compound Poisson model makes this approximation more appealing. the simplicity of (5.Pe. Theorem 5. (5. The first application in risk theory is Iglehart [207].5) for the compound Poisson model which does not require much more computation.g. B0 * Boo. Then as 0 _+ 90. pt? 4 peo. However. such that the Poisson rate Oe. For claims with infinite variance. that 00 4090. in Asmussen [12]. However.1 and Section VIII.1. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion.
and we want to consider the limit 77 10 corresponding to Oo f 0. 2. B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . whereas there we let the given 3B. Then r. and we are studying b(u.s and p = /3µB < 1.4. For each 9. this idea ignores (among other things) the presence of the overshoot e(u).'(yo) = 0 and let 90 = 'Yo. . 0(0) = 0.c(s) = . risk process with safety loading 77 > 0 correspond to 9 = 0 . The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. claim size distribution B . 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. P9(r (u) < oo) = 1 for 9 > 0. Since Brownian motion is skip free.1) . this means the following: 1. 77 is close to zero. Determine yo > 0 by r.ao (0) _ /c(s + 9 . 3. it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). In this setup. Bo(dx) = B[eo]B(dx). let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o].90] B(dx).6. However . 77 = 1/p .T) = Peo(r(u) < T) for 90 < 0. 9o T 0. The objective of the corrected diffusion approximation is to take this and other deficits into consideration.Q (B[s] .90) .90) and the given risk process corresponds to Poo where 90 = 'yo.9(s) = Ico ( s + 9) . PB('r(u ) < oo) < 1 for 9 < 0.6. Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). Let PO refer to the risk process with parameters e9oz Qo = QB[90]. which we have seen to play an important role for example for the CramerLundberg approximation . this is because in the regime of the diffusion approximation .1 > 0. In terms of the given risk process with Poisson intensity ./c(9 . .
(. IGu+u2.Varo S1 = f30Eo U2 = S1. C.3) this implies (take u = 1) Ego exp { .. 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) = ST . Theorem 5. The corrected diffusion approximation to be derived is (u. write r = T(u).e. the solution of r.T) 1+u2 (6. . . PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. (U. tu2 ) i IG (t. () where h (A.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. C . The first step in the derivation is to note that µ = k (0) = r0 (00) . One has (6. Vargo S. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]. 9otc0" (0) = 0061 = ul. for brevity. bl IG(t81.C.3 applies and yields 1061 U61 Stdlu2/CZdi {W_1(t)}t>0 t>0 which easily leads to 1 StU2 {W( J(t)1t>0 { u S1 t>o Y'(u..() The idea of the proof is to improve upon this by an O (u1) term (in the following. u) = IG(x/u2.7(u)/u2} eh(A.122 CHAPTER IV. 0o to.1) .1) IG(x. u) = euh(a .2) . 1) • Since L eatIG (dt.u. (6.. (. C) = 2A + (2 . 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. i.(y) = 0. (01.S.
To arrive at (6. which is based upon exponential claims with mean µB = 1. 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. .s. 6 . p = 0.f. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.52/u where Z has distribution IG (•. however . But the Laplace transform of such a r.5) Once this is established .5) according to (6.v.h. 9o T 0 in such as way that C = Sou is fixed. In ( 1) and (2).4. A numerical illustration is given in Fig. (6. the r.1 + u2 I Indeed.2) is indeed o(u1). distributed as Z .d. The solid line represents the exact value . of (6.s. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.7.3). just replace t by Tb1/u2. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} . Note.1 below is exact. 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 . u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. 1% in (2) and (4). we have p =. . calculated using numerical integration and Proposition 1.1 + 629. . however. in (3) and (4).ry2 . that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] .exp { h(A.2).v.1 As u + oo.2 ).6. that whereas the proof of Proposition 6. and the dotted line the corrected diffusion approximation (6.3 = 0. of a (defective) r. we get by formal Laplace transform inversion that C 2 u. is the c.z .3. . bl I IG I t +2 .h. 1.
.7 or at values of Vi(u) like 1% is unsatisfying.00 0.05{ 0.1 It is seen that the numerical fit is extraordinary for p = 0. .0 0.01 0.aa1 .OOIi O.u2 2u3 (e .1 W IU. The proof of Proposition 6.2 e.(061 0.19)2 11 20 20 i0 T 1n0 Figure 6. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.199 0. OM 0.T1 00.7. PROBABILITY OF RUIN IN FINITE TIME 0.111 W(U.W21 0. and all of the numerical studies the author knows of indicate that its fit at p = 0..T) 0. Similarly. For further numerical illustrations. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.07 0.4 may not be outstanding but nevertheless.1 proceeds in several steps.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.114 0. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34]. (Inc 0s 0.08 a.011 L1 60 T IM 11.EB 0 p ex p ( 7 S h ^)u .TI CHAPTER IV.08 0. see Asmussen [12]. A51 7(SAT 3 3 h(X.124 0.() Lemma 6.T) 111 0. the fit at p = 0.^) .
(6.(3)Eea LauT exp i 3J . 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 . exp ue } al 1J 3 exP I [2).co (e) .C2 = 2). the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .. in Lemma 6.s.(3) J t _ aa1T l + eh(A. 1 = PB(T < oo) = Eo0 exp 125 {(B .4) that the r..2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 ..61a2T (B3 .1) h(A.2u (B3 .h..+ h (A. () + C and note that 2 KO (0) = 102. the result follows.4 Ea.r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .6. CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.C)C/u .co ((/u)) } Let 8 = (2a + (2)1/2 = h().6) u U3 Lemma 6 .T (co (8/u) . (6. 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 .00)(u +C)  'r (. + a1b2 + . () 62 Eeo exp u u2 J .
we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.h.6  d h(A.2.126 CHAPTER IV.\+ (2 (3 e 2u [ (2.() . PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27. and inserting this and 9o 2 = S/u on the r. l Lemma 6 . Thus by Taylor expansion around ( = 90u.e h(aS)h (^^ 262 exp {_h(. we get h(A.() I 1 + u2 ) y .4.6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4).2) for O(u) (indeed . 2 and (6. () . 2 + 00 = . () by h(\.2 (^/2 + 3y9o + 390) + O(u3).\ + () 1 2 / . letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2).(2A + ()1/21 exp S h(A. yields +90 62 0 + O(u 3) 2u2 +O(u 3). [2+ (2 . 0 The last step is to replace h(A.7) and using eh(a.s. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6.1 (y/2 + Oo)u .h (A. yu/2) h(A. 5 exp { _h(A) (1 + / y u J)) exp 1.x.2u [2A+ (2 3 . C) ( 1+ u2 The result follows by combining Lemma 6 . and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6.. Thus a2 y = 290 + O (u2). There are two reasons for this : in this way. yu/2).S) d e 62 .
4.1: Just insert Lemma 6. () (i+a ) 2A + (2 . this case is in part simpler than the general random walk case because the ladder height distribution G+ can be found explicitly (as pBo) which avoids the numerical integration involving characteristic functions which was used in [345] to determine the constants.5 in Lemma 6. The answer is similar: the process behaved as if it changed its whole distribution to FL. the 'typical' value (say in sense of the conditional mean) was umL. that is. the analogous analysis of finite horizon ruin probabilities O(u. The adaptation to risk theory has not been carried out. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. with the translation to risk processes being carried out by the author [12]. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs.(i+ 62 exP{ h(A. the approach to the finite horizon case is in part different and uses local central limit theorems. () I 1 + u 2 ) } S 1 . HOW DOES RUIN OCCUR? exp { h (x. 'yu/2) 127 ( i+ M pz^ exP { h (A.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. . the same as for the unconditional Lundberg process. ()} 3 h (A.e. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. i. 0 1 Proof of Proposition 6. ()} .T) has not been carried out and seems nontrivial. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. and to the Markovmodulated model of Chapter VI in Asmussen [16]. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. Fuh [148] considers the closely related case of discrete time Markov additive processes.7. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. In Siegmund's book [346].
.vi(u) Ce'Yu Corollary 7.(u). (u) and satifying PL(F(u)) * 1. We are concerned with describing the F(u) distribution of {St}o<t<T(u) (note that the behaviour after rr(u) is trivial: by the strong Markov property. In the exponential case. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL. r(u) < oo) .EL[e7S. . Then also P(u)(F(u)) + 1.TT(u) _measurable. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[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.2 If B is exponential..(u)_ is that i.(u)_ and ^(u) are independent .F.t. ^(u) is exponential with rate 8 w.1 Let {F(u)}u>0 be any family of events with F(u) E F. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u.(u) is not .ST(u)}t> o is just an independent copy of {St}t>o). Note that basically the difference between FT(u) and . so in the in the proof.F. stating roughly that under F(u).3 to .r. then P(u) and FL coincide on . Recall that .t.T.FT(u) = o' (T(u ). {ST(u)+t . + TMOO ). we give a typical application of Theorem 7. In fact. Proof P(u) (F(u)c) = F(flu)c. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . Theorem 7 .128 CHAPTER IV.r. F(u)c] ti e' ru]PL (F(u)`) > 0.(u)_ and similarly the denominator is exactly equal to Ce7u. the Poisson rate changes from . {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u).. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . . FL As example.3L and the claim size distribution from B to BL.1. u * oo. P(u) and rate = aL w.
3 M(u) pcu) 1 . HOW DOES RUIN OCCUR? Corollary 7. A somewhat similar study was carried out in the queueing setting by Anantharam [6]. however.7. Proof For the first assertion. From a mathematical point of view. Notes and references The results of the present section are part of a more general study carried out by the author [11].(1 .eALx) M(u) k=1 u The proof of the second is similar. the queueing results are of a somewhat different type because of the presence of reflection at 0. see further XI.3. . This is currently a very active area of research. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). who also treated the heavytailed case. take I(Tk < x) . the subject treated in this section leads into the area of large deviations theory.eaLx.
This page is intentionally left blank .
A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. see A.1). .i.. {St} is the claim surplus process given by I.1 Define p = !µ. In the socalled zerodelayed case. and M is the maximum of {St}. the distribution Al of T1 is A as well. . Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . We use much of the same notation as in Chapter I.(1. U2. . D'2. of the risk process form a renewal process: letting Tn = Qn . and the one corresponding to T1 = s by 1/i8 (u). AA t*oo lim St = lim ESt t tioo t = p .1. the Tn are independent. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u).d. r(u) the time to ruin. Proposition 1. with Nt = # {n: Un <t} the number of arrivals before t.. the one corresponding to the stationary case by 00)(u). .. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. with the same distribution A (say) for T2. with common distribution B.7)... are i.Q.Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1.. the claim sizes U1. T3.1 (T1 = a1).Then no matter the distribution Al of T1i B.. Thus the premium rate is 1.
we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. s + t µA PA 0 Of course. Nt + EVar U. such that no arrivals occur in the off state. OFF.Nt] * a/PA. The renewal model is often referedd to as the Sparre Andersen process. Here are two special cases of the renewal model with a similar direct interpretation: Example 1. The simplest case is of course the Poisson case where A and Al are both exponential with rate 0.2 (DETERMINISTIC ARRIVALS) If A is degenerate.132 Furthermore for any a > 0. Example 1 .1 of the safety loading appears reasonable here as well. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). If the environment is Markovian with transition rate A from on to off and u from OFF to ON.1) ENt/t + 1/µA. stating that E[Nt+a . Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). Thus.0 > 0. the .1) follows . RENEWAL ARRIVALS lim E [St+a . For (1 . From this ( 1. but the arrival rate in the ON state is . However . t 4oo Proof Obviously. Proposition 1. the definition 77 = 1/p . after E. CHAPTER V. Nt ESt = E E UI Nt t = ENt•pB . by the elementary renewal theorem (cf.1 gives the desired interpretation of the constant p as the expected claims per unit time. and (1 .a is really the accumulated claims over a period u of length a. 2).St] = a(p .3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. A.t. say at a.1). 3) follows similarly by Blackwell 's renewal theorem.
r.1.s < u).. the first term represents the probability F(U1 . integrate (1.t.. initial vector (1 0) and phase generator 11 However. the fundamental connections to the theory of queues and random walks. Ao.4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0.s. INTRODUCTION 133 interarrival times become i. arrival times. u For later use. the relevance of the model has been questioned repeatedly.y)B(dy). S o<t<oo n=0. (1.4) w. U1 . The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. Proof The essence of the argument is that ruin can only occur at claim times.2. (an arrival occurs necessarily in the ON state. and for historical reasons. . and the present author agrees to a large extent to this criticism.} with {S(d)} a discrete time random walk with increments distributed as the independent difference U . For the stationary case. and then the whole process repeats itself).. More precisely. 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^.1..i. 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 . 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 have From this the result immediately follows. Therefore.d. if for nothing else then for the mathematical elegance of the subject. However.4) with phase space {oN.oFF}.4) fo Indeed..1.T between a claim U and an interarrival time T. in general the mechanism generating a renewal arrival process appears much harder to understand. A is phasetype (Example 1.. we feel it reasonable to present at least some basic features of the model.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.
b=1 !=1 where {Nt } is a Poisson process with rate .0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . At the time of death . then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . are independent of {Nt} and i. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . That is . i. t. < 1. 2. Theorem 2 . the remaining part of the prepayment (if any ) is made available to the company.3* (say ) and the U. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. resp .Ut. RENEWAL ARRIVALS 2 Exponential claims. 00). then 0 * (u) = 1 for all u > 0. If .1) +ry. with common distribution B* (say) concentrated on (0. Using Lundberg conjugation . The initial reserve is obtained by prepayments from the policy holders.d.134 CHAPTER V.1. each of which receive a payment at constant rate during the lifetime .a*PB• > 1. (2. A typical sample path of {Rt } is illustrated in Fig.3* pB.1 r* (u) One situation where this model is argued to be relevant is life annuities.1 If. the claims and the premium rate are negative so that the risk reserve process . St = t .1) . U Figure 2. the claim surplus process are given by Nt Nt Rt = u+^U. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.
0) and has typically the shape on Fig. Define T_ (u) = inf It > 0 : St = u} . Let B(dx) = ^e7x B*(dx). Then the function k* is defined on the whole of (oo. (a) is*(a) (b) .g. > 1 . B*. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r.s.* (a) = log Ee'st I. B. then by Proposition 111. T_ (u) = inf { t > 0 : St = u 'r* (u).Rt.3*. If I3*pB* < 1.2 Assume now . cf..2.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. the safety loading of { St} is > 0. Hence y exists and is unique. of {St} is c(a) = is*(a7). T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). Fig. 2. St=Rtu=St.2(b). Hence T_(u) < oo a.(u ) < oo) = E {e7sr_ (u). and the Lundberg conjugate of {St} is { St } and vice versa. B* [7] and let {St} be a compound Poisson risk process with parameters . Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *.(a) 7 Figure 2.1. Since ic'(0) < 0. 2. Then the c.0.UB. . and thus 1 = P(T.f.2(a). 0 Now return to the renewal model. Proof Define 135 St =u .
. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.u+ and lr+. 3* = 6. 2. However. we get Ee'M(d) = Ee°M* _ Y/(.. and hence the failure rate .Y] (2.4. 1) means that 8(A[ry] . RENEWAL ARRIVALS Theorem 2 .2 If B is exponential.. with rate S (say).1 it is seen that ruin is equivalent to one of these values being > u..1. Now the value of {St*} just before the nth claim is To +T1* +.Tn} n=0. Then B* = A. alternatively termination occurs at a jump time (having rate 8).1. T2 = U2..136 CHAPTER V.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 . with the probability that a particular jump time is not followed by any later maximum values being 1 . u Hence P(M(d) > u) _ 1r+e'r".g. the failure rate of this process is y. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 .f. respectively.7r+ 7r EeTo b/(Sa) + +.a) = 1 . and from Fig ..•.. and 5PA > 1..Y a I..e.)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[... then ..1.1) + ry = 0 which is easily seen to be the same as (2. To + M(d) in the notation of Proposition 1.2).Tr+.'s and noting that V)*(u) = P(M* > u) so that Theorem 2..Ti = U1. According to Theorem 2. A variant of the last part of the proof. To + max {Ul+•••+UnTI.1 means that M* is exponentially distributed with rate ry. Hence M* max {To + Ti + • • • + Tn . Taking m.4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value.Un } = max St = t>0 n=0.. and (2 .+Tn U1 Un.Ui .• • • .
..4. Hence the failure rate of M(') is 6(1 . letting P(d) refer to the renewal risk model with these changed parameters . Furthermore. which states that for a given a. This follow since.2. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1.2.3. the imbedded discrete time random walk and Markov additive processes..3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U .7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". hence exponential with rate b.7r+) and hence r+ = 1. 111. Putting this equal to y. resp. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+. B^d) where Aad> (dt) = ^[ a] A(dt). 3a The imbedded random walk The key steps have already been carried out in Corollary 11. we see that ry = 6(1. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. 0 3 Change of measure via exponential families We shall discuss two points of view.5. a ladder step is the overshoot of a claim size..y/b.7r+). However. : S(d) > u} . we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. The probability that the first ladder step is finite is 7r+. Bads (dx) = .B(dx).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] . 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^.6.
T is nonlattice. Corollary 3. in the easiest nonexponential case where B is phasetype.p)/($B'[7] .C8e7u where Cs = Ce78B[7]. (d) (7) _ 0. 187) and thereby for ^(u) to be nonlattice w.1) is explicit given 7. and claim (a) follows immediately from this and e (u) > 0.r. just note that F7(d) is nonlattice when F is so . cf.138 CHAPTER V. Proof Proposition 3. ik. In fact.u the overshoot .2 In the zerodelayed case.3 For the delayed case Tl = s. For claim (b).(u) . O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.1 implies Cu) = e«uE ( 7d)e«^(u) . to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.. E(d)e 1' (u). Consider now the Lundberg case. provided the distribution F of U . VIII. We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . we get: Proposition 3.2 p. let 7 > 0 be the solution of r.1).C(°)eryu where C(O) = C0[7] . 7µA . RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) .1 For any a such that k(d)' (a) > 0.t. 00)(u) . This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. (b) V)(u) . (a) '(u) < eryu.4... i . 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 .Ce"u where C = limu.e. For the stationary case.
h'(s). IPA 0 Of course.0 ) = Eo[ha ( Jdt.1) (normalizing by h(0) = 1). 0) = ah (s) . The expressions are slightly more complicated and we omit the details. E8h0 (Jdt. 3b Markov additive processes We take the Markov additive point of view of II. Sdt) = h(s . B(x) = o(e7x) and dominated convergence.dt ) eadt = h ( s) .a . Let P8f E8 refer to the case Jo = s. 0 0 . delayed version of Lundberg's inequality can be obtained in a e7u.St)} can be defined by taking Jt as the residual time until the next arrival.dt(ah ( s) + h'(s)) so that Gha ( s.(°) ( Ce8B[7] Ao(ds) similar manner. St)} and h. (u + s .Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds).9. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. where G is the infinitesimal generator of {Xt} = {(Jt. y) = e°yh(s). For s > 0. we invoke the behavior at the 1 = h«(0..4). Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v).5. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. According to Remark 11.3.1) = C(O). (s.(s. To determine boundary 0. Here K. 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] . h(s) = e(a +x( a))8 (3./c. 0) = tc(a)h(s). we look for a function h(s) and a k (both depending on a) such that Gh.y) B(dy) 0 For the stationary case.5..
T2 are independent with distributions Ba. U. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same....c(a)] which shows that U1. RENEWAL ARRIVALS B[a]A[a . St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. . .rc(a)] B[a] A[a .2) means 1 = B[a]/3/(/3+a+rc (a)). J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk..1)a in agreement u with Chapter III. we can now for each a define a new probability measure Pa. however. resp. .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. T2. (3.tK( a)e. resp. = J8 = T2.rc(a)] = B = Ba[13]Aa[5]. Ba(dx) = B(dx).8..rc(a)] = 1.s and P(d). Remark 3 .13]A[b .2) As in 11. A[a . Proposition 3..2). Ba where Aa (dt) . ] = E. Aa as asserted . Note that the changed distributions of A and B are in general not the same for Pa.4 The probability measure Pa. An easy extension of the argument shows that U1.140 CHAPTER V. An important exception is.5.( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3.s(Jo = s) = 1 follows trivially from Lo = 1. Further.s governing {(Jt.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east . . 5 For the compound Poisson case where A is exponential with rate . . (3.e. Ease AU1+6T2 [ AU1+6T2 = Ea a LT. [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +. . rc(a) = 0 (B[a] . [a + /3] A[b .e(«+k(a))t esy A(dt). i.a .c(a)] B[a] Proof Pa.a . since JT.
and define yy = ay . Let M(u) be the number of claims leading to ruin .5. or FCFS = first come first served) queueing discipline and renewal interarrival times. For the approach via Markov additive processes.yu ) e7vu A[ay .rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. and U„ the service time of customer n. not after time T. Then "^ e(ay+w(aY))8 Ys(u. say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims. see e.5 Proposition 3. Then J(rr(u)) = TM(u)+1 and hence Ws(u.t. 2. Using the Markov additive approach yields for example the following analogue of Theorem IV.ay+ray))TM(. and assume that T„ is the time between the arrivals of customers n . [APQ] Ch.y yuAa y [ay + K(ay) . Proof As in the proof of Theorem IV.4.. A(ds).g. 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. XII. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). for the zerodelayed case zp8(u. The virtual waiting time Vt at time t is the residual amount of work at time t. the time from he arrives to the queue till he starts service. . defined as the single server queue with first in first out (FIFO.1 and n. .)+1 e J j e(ay+w(ay))8 e . see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. the amount of .. . The claim for the zerodelayed case follows by integration w.r.yx(ay). u The approach via the embedded random walk is standard.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. yu). it is easily seen that ic(ay ) > 0. yu) = F'ay. Label the customers 1. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. yu ) < e7yu.4. which is the same as the asserted inequality for 0. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v.(u. that is.4. that is.s eaysr(")+r(u ) K(ay) h (s) .
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.1 Wn+1 = (Wn + U.4. Wn converges i n distribution to a random variable W.142 CHAPTER V. Thus Vos}1 _ = (Wn + Un .4. the waiting time a customer would have if he arrived at time t.3) Proof Part (a) is contained in Theorem 11. an+1) = [on. and combining with (4. . we get: Corollary 4.• • • Tk ). equivalently. Then: (a) as n + oo. and we have P(W > u) = V. Vt converges in distribution to a random variable V. If W1 = 0.+ Un. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. (4. 0 Applying Theorem 11. p < 1. Thus.3 Assume rl > 0 or. and obviously z/'(u) = limN. Let the T there be the random time UN. (4.1).. 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.. in which case {V} remains at zero until time on+1.1 and Corollary 11. but interchanging the set .2) (b) as t * oo.. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.. and we have P(V > u) = ?/iiol(u). the proposition follows.1. since customer n arrives at time on. The traffic intensity of the queue is p = EU/ET.1.3. on + Tn) the residual work decreases linearly until possibly zero is hit. (u).4): Proposition 4.4. we have Wn = Van(left limit).2 Let Mnd) = maxk=o. whereas in [On .2.Tn)+.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ .1) The following result shows that {Wn} is a Lindley process in the sense of II. equivalently. then Wn v M. (4."^ Vi(N) (u). but we shall present a slightly different proof via the duality result given in Theorem II.4: Proposition 4.n1 (U1 +• • •+Uk Tl . It then jumps to VQ„ .
(4. we obtain: Corollary 4.. Then the corresponding queue is M/G/1. T1...1. convergence in distribution hold for arbitrary initial conditions . by an obvious reversibility argument this does not affect the distribution ..(N)(u) has the limit tp(u) for all u. which implies the convergence in distribution and (4.2. Then the arrivals of {Rt} in [0.y)F(dy).2). In fact .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. conditioning upon U* . K(x) = P(W < x).T* are independent and distributed as U1. Corollary 4. Then K(x) = J x00K(x .. hence (since the residual lifetime at 0 and the age at T have the same distribution . A. cf. For part (b). but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.T*)+. It follows that P(WN > u) =.Ao in (b)..T* = y yields K(x) = P ((W + U* . u Now return to the Poisson case .. T] form a stationary renewal process with interarrival distribution A. we get W = (W + U* . { 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). (4. THE DUALITY WITH QUEUEING THEORY 143 (T1. as WN. namely W1 = 0 in (a) and Vo = 0. i. T1 . we let T be deterministic . Ti) and similarly for the U. Hence for x > 0.. resp .le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before . x > 0.e. where U*. and we get: . TN) with (TN..oo in Proposition 4.4 The steadystate actual waiting time W has the same distribution as M(d). Letting n oo in Corollary 4.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival.5) Proof Letting n .4.5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x)..T*)+ < x) = P(W + U* . However.
Note that (4.g. the zerodelayed and the stationary renewal processes are identical.5) looks like the convolution equation K = F * K but is not the same (one would need (4.5) is in fact a homogeneous WienerHopf equation. W v V. implying P(W > u) = P(V > u) for all u. RENEWAL ARRIVALS Corollary 4. despite the fact that the extension from M/G/1 is of equally doubtful relevance as we argued in Section 1 to be the case in risk theory. Cohen [88] or [APQ] Ch. The equation (4. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . Some early classical papers are Smith [350] and Lindley [246]. VIII). Hence '(u) = Ali(°)(u). the actual and the virtual waiting time have the same distribution in the steady state. That is.144 CHAPTER V.5) to hold for all x E R and not just x > 0). see e.6 For the M/G/1 queue with p < 1. . Proof For the Poisson case. Asmussen [24] and references there.g.
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. Thus. i=1 0 and r(u) = inf It > 0: St > u}.)iJEE and its stationary limiting distribution by lr. dv. . • The premium rate when Jt = i is pi. t St = E Ui .T) = Pi (T(u) < T).f pi. • Claims arriving when Jt = i have distribution Bi. {Jt} describes the environmental conditions for the risk process. here it exists whenever A is irreducible which is assumed throughout. M = supt>o St.(3i when Jt = i. Ire = 1. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). {St} denotes the claim surplus process. The intensity matrix governing {Jt} is denoted by A = (A.. 145 Oj( u. and can be computed as the positive solution of WA = 0. N.
say. 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. u .1 Consider car insurance. t(i) = T(')e are the exit rates. Unless otherwise stated.a('). P = E 7riPi. one expects that 3i > on and presumably also that Bn # Bi. According to Theorem A5.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1.14. with rates Ani and Ain. leading to E having two states n. r^ = P (1.5 below. = iii when j E E(i). Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). f3i and claim size distributions Bn. Then the state space for the environment is the disjoint union of E(n) and E(i). We let p Pi = /ji/AB. T(=)). i and corresponding arrival intensities Qn. Bi. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. a(i). Example 1 . An example of how such a mechanism could be relevant in risk theory follows. this is no restriction when studying infinite horizon ruin probabilities. and assume that weather conditions play a major role for the occurence of accidents. assume that the sojourn time in the icy state has a more general distribution A(i).1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. we shall assume that pi = 1. we can approximate A(i) with a phasetype distribution (cf. cf.2.146 CHAPTER VI. Example 1. respectively.. which is clearly unrealistic. Proposition 1. For example. we could distinguish between normal and icy road conditions. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. the operational time argument given in Example 1.4) with representation (E(i). /3 = Nn when j E E(n). and we have f3.11 below. Thus. cf.T(n)).1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. and p is the overall average amount of claims per unit time. in blockpartitioned form. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. say.
it = Jel(t). 1 . . t(n) = T("i)e.>. resp. and similarly for the normal period. One way to model this would be to take A(') to be Coxian (cf.3 Consider again the alternating renewal model for car insurance in Example 1. Example VIII.1..4) with states i1.. depending only on 77.T(n)). 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. i ) : n E H. u From now on. i8f n1. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). u Example 1.3i/pi. (9) where q = CHI. Indeed.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. is the probability that a long icy period is followed by a short normal one. let T 9(T) = f pi. w.1. Then for example wi.n..2..Q. ..3.p. 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) .. but assume now that the arrival intensity changes during the icy period. A('^). and 1/ii(u) = t/ii(u). one could for example have H = {i1.. St = SB=(t).j = . where W = (w. i E E(n) }. T(9) +wggt(9)0.. dt.. 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. the parameters are ^ij = aid/pi.. n8}. MODEL AND EXAMPLES 147 Example 1 . This amounts to a family (A(")) ?CH Of sojourn time distributions. Approximating each A('?) by a phasetype distribution with representation (E('l). say.J017. iq (visited in that order) and letfOil >.. such that a sojourn time of type rt is followed by one of type c w.tEH is a transition matrix.3i. The simplest model for the arrival intensity amounts to . and . Qi = .a(n). the state space E for the environment is { ('q. In the car insurance example.. u Example 1 . say it is larger initially.
The key property for much of the analysis presented below is the following immediate observation: Proposition 1.(3iBi(dx). N > 1(Ul < x) a4 B*(x). o = 0. In particular. B* = 1 /^* Bi.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.e(A(Oi)d'sg)xe.148 CHAPTER VI. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation.(Qi)diag)• More precisely. vi(dx) = .A . Next we note a semiMarkov structure of the arrival process: Proposition 1. JT1 = j) = Qj • e. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. )3*. More precisely. one can associate in a natural way a standard Poisson one by averaging over the environment. Pi (Ti E 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. we put )3* = E 7fi/3i.5. Nt Nt a . t l=1 Note that the last statement of the proposition just means that in the limit. dx. qij = 0 in the notation of Chapter 11. 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). . the empirical distribution of the claims is B*. . 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. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.7 The Pidistribution of T1 is phasetype with representation (ei.8 As t oo.
the Fidistribution of T1 in {St(a ) } is phase type with representation (E. In particular.. has distribution (7ri)3i //3*)iEE and is independent of Ti.. 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$. Hence Nt'> a . i.aA . Example 11. ^j 7riNi.1. cf.6. oo) as a 4 oo. B*.. Nt a' t t iEE Also. However .7. {St} to the compound Poisson model with parameters 0 *. MODEL AND EXAMPLES 149 Proof Let ti = f1 I(JJ = i) ds be the time spent in state i up to time t and Nti) the number of claim arrivals in state i .(/3i)aiag). zli( (u) .. Bi(x). Proof According to Proposition 1. and let {St °i} refer to the one with parameters Pi. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size.* (u) for all u. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. By Proposition A5. Proposition 1. y Ni) i Nti) t a. Then it is standard that ti lt '4' iri as t > oo.2. Bi. A. and furthermore in the limit JT. e. Bi. In particular.. the limiting distribution of the first claim size U1 is B*. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti.9 Consider a Markovmodulated risk process {St} with param eters Ni. given {Jt}0<t<0.4. Conditioning .x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). Then {St)} + {St*} in D[0. this converges to the exponential distribution with rate 0* as a * oo. aA. 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.
1. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7. state 1 appears as more dangerous than state 2. 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 = . That is. resp. U.2 +2 2 = 3. s 5 in state 2.. T) for all u and T. there are p = 2 background states of {Jt}..2}.s = o in state 1 and with intensity 1 .150 CHAPTER VI. Claims of type E3 arrive with intensity 2 .). thick. oo). Thus. e..=1. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. 132=2.2. > 1.FT. 1.10 Let E_{1. we first get that 3 (3* = 2.2.l3* and U2 having distribution B*. those of type E7 with intensity z s = 5 in state 1 and with intensity z .. From this the convergence in distribution follows by general facts on weak convergence in D[0.. the paths of the surplus process will exhibit the type of behaviour in Fig.. 9 . with T2 being exponential with rate . 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary.. On Fig. T) + ?P* (u. U2) are independent of . A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. and (at least when a is small such that state changes of the environment are infrequent). Continuing in this manner shows that the limiting distribution of (T.. which also yield O(a) (u.1 with periods with positive drift alternating with periods with negative drift. lines in the path of {St}. the company even suffers an average loss.g. shows similarly that in the limit (T2.s = 1o in state 2. marked by thin.1 of [145]. B1=3E3+2E7. B2=1E3+4E7. is as in {St }. Computing the parameters of the averaged compound Poisson model. The fact that indeed 0(a) (U) 3 0* (u) follows..31µB 2 = 2 5 3 7 70 Thus in state 1 where p. 1. MARKOVIAN ENVIRONMENT upon FT. 0 Example 1. from Theorem 3. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier).
01 /.1. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs.(3. the averaged compound Poisson model is the same as in III.1 a.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2.s.1). MODEL AND EXAMPLES 151 Figure 1. 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). For (b).1 Thus.1. Proof In the notation of Proposition 1. 0 The definition (1. (b) St/t * p .8.1) 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. = P. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7. note first that EN Uk')/N a4' µgi.11 (a ) ESt/t * p . a fraction r.. t * oo. iEE . t + oo. That is.
and a more comprehensive treatment in Asmussen [16]. 38) Eiw1 = 1/ir.ld. and hence oscillates between 0o and oo so that also here M = oo. and hence wn /n a4. and hence 1/ii(u ) = 1 for all i and u.Jt=i}. and involves a version of the .1 jEE = (p . Now let r) = 0. also + . let some state i be fixed and define w=wl=inf{t >0:Jt_#i. s. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43].12 If 77 < 0. w2=inf {t>w1:Jt_#i. and so on. The mainstream of the present chapter follows [16]. See Meier [258] and Ryden [314].1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes.1)Eiw = 0.. . dt .s. There seems still to be more to be done in this area. then M = 00 a. The proof of Proposition 1. n n Thus {SWn l is a discrete time random walk with mean zero. X2 =SW2 So. Proof The case 77 < 0 is trivial since then the a. X3. Now obviously the w. Proposition 1. see [APQ] p.. Eiw.\ i and EiX1 Ei f 13 J. PB. some early studies are in Janssen & Reinhard [211].2(a) p.0i(u) < 1 for all i and u. X 1 =Sty. Theorem II. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. In risk theory.Jt=i}.. Then by standard Markov process formulas (e. and hence M = 00. [302]. [APQ].a form a renewal process . [212]. MARKOVIAN ENVIRONMENT Corollary 1...Eiw o'o Eiw • E ^ifjµs.. Statistical aspects are not treated here. and .. [315]. EiX = 0. having the Pidistribution of X. .4. then M < oo a.. with X2. The case 77 > 0 is similarly easy.1 and the Corollary are standard. If 77 > 0.1 of St / t is > 0.s. 136 or A.. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade. see the Notes to Section 7.g.. + Xn SWn ](1 a .152 CHAPTER VI. limit p . Since the X„ are independent .
The form of G+ turns out to be explicit (or at least computable). let G+(i. •) II = JG+(i. k.EA.(u) = Pi(M < u) = e' E G+ (u)(I . •) * G +(k.x.3*B *(y)dy. Proposition 2. n=0 (2. 6. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. •)• kEE Also. for i.IIG +II)e.A) =ZI(St E.2 (a) The distribution of M is given by 00 1 .4) we obtain the following result .g..1) 0 (b) G+ (y. For measurevalued matrices. . G+ is the matrix whose ijth element is E G +(i. e.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. Proposition 2. by specializing results for general stationary risk processes (Theorem II . j. That is. cf. However . j.x). (y.j E E. oo) = J ao 0 G+(i. •). IIG+ II denotes the matrix with ijth element IIG+(i.5.j.6*. Let further R denote the preT+ occupation kernel.j. T R(i.2. B* in Section 1. the definition of . see also Example II. Thus.a/i. oo)).Jr+ =j. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i.dx).1 irG+(dy)e =. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. 00 (2. only with the product of real numbers replaced by convolution of measures. j.6. we define the convolution operation by the same rule as for multiplication of realvalued matrices. oo)) = f R(i.Jt=j)dt. but is substantially more involved than for the compound Poisson case .A) = Pt(ST+ E A. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0).2) R(dx)S((y . j. we get the same ladder height distribution as for the averaged compound Poisson model.i. j. dx)/jBj(y .
2) is just the same as the proof of Lemma 11. lines in the path of {St}. To make Proposition 2. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej.6. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. St < S* for u < t.2 useful . we need to invoke the timereversed version {Jt } of {Jt} . III to bring R and G+ on a more explicit form .3 When q > 0.IIG+II)e.3. {mx} is a non terminating Markov process on E.3) We let {St*} be defined as {St}.1 for an illustration in the case of p = 2 environmental states of {Jt}. see Figure 2. 0  x Figure 2. From this (2. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). mx = j when for some (necessarily unique) t we have St = x. That is. marked by thin. we need as in Chapters II. thick. To this end .1) follows by summing over n and j.154 CHAPTER VI. .1 The following observation is immediate: Proposition 2. The u proof of (2. hence uniquely specified by its intensity matrix Q (say). and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value. resp. and that the environment is j at the nth when we start from i is e . JJ = j.
2. Proof The argument relies on an interpretation in terms of excursions. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. 0 mms1   ^O \ T. Furthermore. } is a minimum value at v = t. Q( n+l) _ ^.2. An excursion of {St*} above level x starts at time t if St = x.*. The definitions are illustrated on Fig.0.2 where there are three excursions of depth 1. we say that the excursion has depth 0. For example the excursion of depth 2 has one subexcursion which is of depth 1. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. If there are no jumps in (t. corresponding to two subexcursions of depth 0. the sequence {Q(n)} A* defined by Q(O) = .2 .(/3i) diag. and the excursion ends at time s = inf {v > t : S. ( Q( n)) converges monotonically to Q.. 2. {S. = x}. s].4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* . we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2.(/3i)diag + T S(dx) eQx. In general. Figure 2. Note that the integral in the definition of W(Q) is the matrix whose ith row is the ith row of _ 3 f e2Bi(dx).and a jump (claim arrival) occurs at time t. Otherwise each jump at a minimum level during the excursion starts a subexcursion.
i.(01)diag = Q. A) = L' U(j. mx+dx = j) occurs in two ways . j. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. of the definition to make U be concentrated on (co. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. Suppose mx = i.St EA.St <S*. A) = f Pi(mx = j) dx eie4xej dx A u (2. h.j. 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. or through an arrival starting an excursion terminating with J..u< t).4) To show Q = cp(Q).6) . Similarly. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* .s.5) A (note that we use A = {x : x E Al on the r. Then a jump to j (i.Qi + )%pij) Now just note that t pij and insert (2. (2.j +/3ipij. 7rE Proof We shall show that Fi(Jt=j. = j. Q = W(Q) follows. we first compute qij for i $ j. p1^) Define a further kernel U by f U(i. Theorem 2 . 0)). A).. StEA .156 CHAPTER VI. e.4). The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). either due to a jump of {Jt } which occurs with intensity A= j.T+>t) _ ^iF 7ri (JJ =i. It follows that qij = A.5 R(i. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . Similarly by induction . Writing out in matrix notation . By considering minimum values within the excursion.
0<u<t).Jo=i. Remark 2. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). (Jo = j. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2. {Jt }. (b) for z > 0.6 is hardly all that explicit in general.. St E A. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.2. 0 +1) = cp (K( n)) defined by K(o) = A .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. e. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx). G+((z. we shall see that nevertheless we have enough information to derive.6 (a) R(dx) = eKxdx. consider stationary versions of {Jt}. We may then assume Ju=Jtu. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. the CramerLundberg approximation (Section 3)..g. St < St U.St EA.4].. dt. To this end.(Jt=j. 0 < u < t) = 7rjPj(Jt =i.=StSt.St <Su. and to obtain a simple solution in the .7 It is instructive to see how Proposition 2.z+>t) = P. oo)) = f o' eIXS((x + z. x < 0. Jt = i. S.0<u<t) = P. St EA.(. where A is the diagonal matrix with 7r on the diagonal: Corollary 2.6). and this immediately yields (2. oo))dx..Qi)diag. K( n (d) the sequence converges monotonically to K.StEA. and we let k be the corresponding right eigenvector normalized by Irk = 1. and get irPi(Jt =j.S„<0.t. From Qe = 0.1 can be rederived using the more detailed form of G+ in Corollary 2.0<u<t.r..
158
CHAPTER VI. MARKOVIAN ENVIRONMENT
special case of phasetype claims (Chapter VIII). As preparation, we shall give at this place some simple consequences of Corollary 2.6. Lemma 2 .8 (I  IIG+II)e = (1  p)k. Proof Using Corollary 2.6(b) with z = 0, we get
IIG+II = feIxsux, oo dx.
In particular, multiplying by K and integrating by parts yields
0
(2.7)
I)S(dx) KIIG+II =  (eKx
T
= K  A + (,13i)diag 
Z
S(dx) = K A.
2.8)
0 OO
Let L = (kir  K)'. Then (k7r  K) k = k implies Lk = k. Now using (2.7), (2.8) and ireKx = ir, we get
kirIIG +IIe =
ao k f
7rS((x , oo))e = k (lri(3ips, ) rowe = pk,
0 KIIG+IIe = Ke,
(kirK)(I  IIG+II)e = kKepk+Ke = ( 1p)k.
Multiplying by L to the left, the proof is complete. u
Here is an alternative algorithm to the iteration scheme in Corollary 2.6 for computing K. Let IAI denote the determinant of the matrix A and d the number of states in E. Proposition 2.9 The following assertions are equivalent: (a) all d eigenvalues of K are distinct; (b) there exist d distinct solutions 8 1 , .. , sd E {s E C : its < 0} of (A + (131(Bi[s]  1))diag  sIl = 0. (2.9) I n that case , then Si, ... , sd are precisely the eigenvalues of K, and the corresponding left row eigenvectors al, ... , ad can be computed by
ai (A 
(fi(Bi[Si]

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

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

1
E(+(IIG+(e,j)IIG+[t,j;
In matrix formulation, this means that
C =
E L;i
e7f()
hj,r(_) L
 L
ryC M e
L
c+
(IIG+II  G +[ 7]) 0le
1
L
YC+M+e
'y(PL  1)
(ir KL) (I  G+[ y]) 0le,
164
CHAPTER VI. MARKOVIAN ENVIRONMENT
using Lemma 2.10 for the two last equalities. Inserting first Lemma 3.9 and next Lemma 2.8, this becomes 1 7r LA 1(YI  K)(I  IIG+II)e 'Y(PL  1) = 1 P 7r LA 1(yI  K) k = 1P 7rLO1k. Y(PL  1) (PL  1 ) Thus, to complete the proof it only remains to check that irL = vL A. The normalization vLhL = 1 ensures vLOe = 1. Finally, VLOAL = vLAA'K['Y]A = 0
since by definition vLK[y] = k(y)vL = 0.
u
3b Ramifications of Lundberg 's inequality
We consider first the timedependent version of Lundberg 's inequality, cf. IV.4. The idea is as there to substitute T = yu in 'Pi (u, T) and to replace the Lundberg exponent y by yy = ay  yk(ay ), where ay is the unique solution of rc(ay)= 1 Y Graphically, the situation is just as in Fig. 0.1 of Chapter IV. Thus, one has always yy > y, whereas ay > y, k( ay) > 0 when y < 1/k'(y), and ay < y, k(ay) < 0 when y > 1/k'(y). Theorem 3 .10 Let C+°) (y) _ 1
miniEE hiav)
Then 1 y< (y)
y>
Vi(u,yu)
Pi(u) 
C+°)(y)hiav)
e7vu,
(3.6)
V,i(u,yu)
< C+)(y)hiar )e 'Yvu,
(y) (3.7)
Proof Consider first the case y <
Then, since k (ay) > 0, (3 .1) yields
'12(u,yu)
hiav)]E'iav,i
h(ay ) J*(u)
exp {ayST(,L ) +r(u)k( ay)}; T(u) < yu
We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. (3. However.i [eT(u)K(av ).i I (a) exp {aye(u) + r(u)r. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).00 su e7( ( 3. 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 (3.9) For the proof. as in the classical case (3..8 ) Then for all i E E and all u > 0.5).7. exp {e() + r(u))} . if y > 1lk'(ry).(ay)}. 1 Similarly. 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.j. oo)) and. hj P . (u. for a vector <p(u) = (cpi (u))iEE of functions .j) * coj)(u) _ f u ^Pj(u . yu) f h(av)e v avuE«v.3. dy)• o iEE jEE .i [e*(u)K(av).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). yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav . r(u) yu o)(y)eavuEav.5) will produce the maximal ryy for which the argument works.y)G+(z. we have ic(ay) < 0 and get 'i(u) . Chie ryu < Vi(u ) < C+hie 7u. we let G+ * W(u) be the vector with ith component E(G+(i. av 'i [h.V)i(u. we shall need the matrices G+ and R of Section 2. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE.
U = U". j. Then cpin)(u) sit (u) as n + oo. dy) : 1(u) < C+ > hj u e(1tL)G+(i.ery(&u+x)Bj (dy) Bj(u Bj (u . if r+ (n) is the nth ladder epoch. dy) = aj f Bj(dy . 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. we have G *(N +1) * ^. MARKOVIAN ENVIRONMENT Lemma 3 . j. dx).12 Assume sup1. = Eo G+ G.u IMP:°) (u) I < oo.7. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However.x ) = Gi(u). 00 f C_ hj f e(Y)G+(i. _ To see that the ith component of U * G(u) equals ?Pi (u). dy) 00 C+ ijhj f R(i.& (u). n > oo.166 CHAPTER VI. j. °O . dx) f e7( vu)Bj (dy .13 For all i and u. dx) 100 C . Proof Write UN = EN G+ .3jhj // f 00 R(i. 0 G+(i. dx ) Bj (u . and define W(n+1) (u) = G(u) + (G+ * tp(n))(u).u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0.x) x) jEE 0 E Qj f jEE R(i. jEE u 0 j. j.x ) R(i. Lemma 3 .x) jEE 00 u 0 //^^ C+E.j.(0) ] (u) < sup Jt t.j. 00 Thus C+ > hj f"o e7(Yu)G +(i. Hence lim cp(n) exists and equals U * G.dy).
(3. 9 for the last equality in (3. taking cps°) (u) = 0. 167 u Proof of Theorem 3.u)G+(i. dy) jEE o (3.13. j. We claim by induction that then cpin) (u) > C_ hie7u for all n. Then 0< Vi (u )  0i(u.n) ( u .(u) < T ) to 0i (u) which is different from Theorem 3.Pi(MT > u) = Pi(MT < u. +i .10 ) by Lemma 3 . jEE estimating the first term in (3.13) Hence. 13 and the second by the induction hypothesis . letting MT = maxo<t<T St.10: Theorem 3 . (3. u The proof of the upper inequality is similar .13 Let first cp=°)(u) = C_ hie"u in Lemma 3. and assuming it shown for n. 14 Let yo > 0 be the solution of 'c'(yo ) = 0. T) < C+(')' o)hi7u)e7ou8T . Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. and the proof of the lower one is similar. MT < u.M > u) = Pi(ST<u. j. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp.MT<u. this is obvious if n = 0.ST). dy) jEE u U +C_ hje7( yu)G jEE"" +(i.y)G+(i. let C+(yo) be as in (3. j.3. y]hj = C_ e7uhi.T) = Pi(M > u) . T) = Pi (7. it follows that Vi(u) < C_(yo) h=70)e7ou. dy) (3.8) with y replaced by yo and hi by h=7o ).11) C_e7u 57 O+[i.12) Proof We first note that just as in the proof of Theorem 3.11. from which the lower inequality follows by letting n * oo. and using Lemma 3 .M>u) = Ei [VGJT (u .11). j. Indeed. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . we have Vii (u) .tpi(u. and let 8 = e'(70).10) C_ 1 f hje7(y. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.
but that in general the picture is more diverse. (4.31:5)32 . It was long conjectured that 0* Vi.. and finally in part from queueing theory. For the notion of monotone Markov processes. B2 <_s. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and . The Markov process {Jt} is stochastically monotone (4.o.. ". we refer to . this correponds to the usual stochastic ordering of the maxima M'. < .0. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. Occasionally we strengthen (4. in part from the folklore principle that any added stochastic variation increases the risk.33 or Bi 0 Bj.. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). [177]. we also assume that there exist i # j such that either /3i <.3p.. Further related discussion is given in Grigelionis [176]. The results to be presented show that quite often this is so. u > 0.4) To avoid trivialities.2) (4.o. The conditions which play a role in the following are: .2) alone just amounts to an ordering of the states.3) to B = Bi does not depend on i. this is not the case for (4..o.o.168 CHAPTER VI.1) Obviously. we define the stochastic ordering by 0' < s. <s.5) Note that whereas (4.. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).. (4. Bp..3).3) Bl <_s. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. V)" if z/i'(u) <'c"" (u). (4. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). The motivation that such a result should be true came in part from numerical studies.
COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]. Lemma 4 .4.2. 0 Here (4. 3 (a) P. (4. dx (4.r (Sr(o) E dx Jr(o) = i.r (JT(o) = i.2)(4.9 ) below). E 7r i Wi(u . also Proposition 2. the second follows from an extension of Theorem I1. then P P P 7rjbj..3iBi(x)YPi(u .3* f uB(x) z/^. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = . then j is the more risky state . we need two lemmas. Lemma 4 . Section 4.9) (4. Conditioning upon the first ladder epoch.2)(4..9) follows by considering the increasing functions 3iBi (x) and Oi (u . (b) P. < bp and 7ri > 0 (i = 1. Comparing (4..* For the proof.r(u x)dx.x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4. < a.10) and (4.4) say basically that if i < j . Proposition 2... it follows by a standard . and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. = aP or b1 = .1 for the first term in (4..13* J0 u 0*(u . Proschan & Walkup [140].2... we obtain (cf. 2 If al < . = b.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB.x)B*(x) dx. where 7r2+) = QiµBilri/p.5 (cf.x) of i and using Lemma 4. Conditions (4.6). note that (4. Theorem 4 . Proof of Theorem 4.7) 7ri...7) and Lemma 4. . p).x) dx u o i =1 i=1 (4..4) hold.1 Assume that conditions (4. b1 < .1.10) Q*B*(u)+. 7(0) < oo) = pirf+)..1) which with basically the same proof can be found in Asmussen & Schmidt [49]. T(0) < oo) = Bi(x) dx/tcai . .3 for the second) *(u) _ /3 *B* (u) +..6.8) ^j Tri/iBd(x) .. ^i 7ri = 1.4) is automatic in some simple examples like birthdeath processes or p = 2 . Then V.
(0) = V. Frey. except possibly for a very special situation ... Then i/i*(u) < .170 CHAPTER VI. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. µB. What is missing in relation to Theorem 4.h. Recall that . Q2 = 1.4) is essential (the present author conjectures it is).6). u To see that Proposition 4. For u = 0.3µi < 1 for all i.11) i=1` and that A has the form eAo for some fixed intensity matrix A0. Then the l. (u) may fail for some u.11) is of order 104 and the r.4 is the understanding of whether the stochastic monotonicity condition (4.(0) < b *'(0) for e small enough.3i. of (4..r (u ) fails for all sufficiently small e > 0.4 Assume that .0.0*• i=1 But it is intuitively clear (see Theorem 3. dominates the solution 0* to the renewal equation (4. and from this the claim follows. Bi as e J.h.1 and Proposition 4. u Here is a counterexample showing that the inequality tp* (u) < V)./3*.s. it is sufficient to show that 0'.. Proof Since 0. (u) is not in general true: Proposition 4. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4.6). they are at present not quite complete.8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi . 0.. it will hold for all sufficiently large u. Notes and references The results are from Asmussen. = 102. MARKOVIAN ENVIRONMENT argument from renewal theory that tk.* (0). (4. Using (4. i=1 i=1 7'r(0) _ EFioiwi(0) .s.4 is not vacuous. this ruin probability is /3iPBi. As is seen. let = ( 1/2 1/2 ) . of order 101.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . µB2 = 104. 01 = 103. Rolski & Schmidt [32].2.
and by Proposition II. It is clear that the distribution of X. Jt = i])' EE = vA+n(6. with strict inequality unless a = 0 or bi = 0 for all i E E.)a.g.4(b) that the limit in (4.5.6 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. 0 .2 we have (Ei[e"X'. (4.14) A„(O) iioo varXt t t By convexity.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.a = E irirci(a). Xt)} is a Markov additive process (a socalled Markovian fluid model.14) is nonzero so that A"(0) > 0.(a) > 0 for all a 0 0. it follows by Proposition A1.12) iEE Theorem 4. (4. Further (see Corollary 11. Proof Define X= f &ids. 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] . The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] . is nondegenerate unless bi does not depend on i E E. with strict inequality unless rci (y*) does not depend on iEE.13) implies A(a) > 0 for all a. (4..13) (4.a.4. Now we can view {Xt} as a cumulative process (see A.ld) with generic cycle w = inf{t>0: Jt_54 k. Hence if 5i 54 0 for some i E E.g.5. e. in particular .1) . Lemma 4. This implies that A is strictly convex.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E. Then {(Jt. Asmussen [20]) as discussed in 11.5.5 y < ry*.1) . cf..
Since ic is convex with rc'(0) < 0 . Thus y(e) * y* as e 10. Hence rc (y*) > 0.. Frey. improving upon more incomplete results from Asmussen. Here we put a = 1/e.) and rc (•). (4.p yi and compute 8y 8a a=0 In both cases.15) Normalizing h by 7rh = 0. the basic equation is (A + (rci(y))diag)h = 0. MARKOVIAN ENVIRONMENT Proof of Theorem 4. where A. Notes and references Theorem 4. note that y(a) + mins=1.5 is from Asmussen & O'Cinneide [40]. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. we have 7rh' = 0. Further a(1) = rc(y*) by definition of A(. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. h depend on the parameter (e or a). we get rc (y*) > 0 which in a similar manner implies that u y < y*.172 CHAPTER VI. Rolski & Schmidt [32]. In the case of e.15) once more and letting e = 0 we get . 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e.. Hence letting e = 0 in (4..16) Differentiating (4. whereas the .15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao .12) and rc*(y*) = 0. a = 1 in Lemma 4. The corresponding adjustment coefficient is denoted by ry(e). h(0) = e.e7r)1 (Ici(Y*))diage.. If rci(y* ) is not a constant function of i E E.6. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. (4. Let bi = rci(y*). Then > risi = 0 because of (4.Qi and Bi are fixed . h'(0) = (Ao . y.5.eir)h'(0).
which has recently received much attention in the queueing literature. .. 5 The Markovian arrival process We shall here briefly survey an extension of the model. p.16) yields Proposition 4. (4. (4.18) 0 = 27'(0)p+27r(rs. . the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity .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). 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). Rolski & Schmidt [32]. Inserting (4. and we have proved: Proposition 4.17) (4. multiplying (4.18).17) by 7r to the left to get (4.i(7' *))diagh'(0). and may have some relevance in risk theory as well (though this still remains to be implemented). We get 0 = (aAo + ( lc&Y))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.8 when ryi < 0 for some i is open. 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) .7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a.5.8 If (4.19) holds. (4.20) Letting a = 0 in (4. i = 2. Frey. We assume that 0 < y < 7i. The analogue of Proposition 4.
then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. B. A(1) = A . We then let (see the Appendix for the Kronecker E = E(1) x E(2). T). and thus 1i = 0. For i = j.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII.2) A(1) = A(' 1) ® A(1.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. the definition of Bi is redundant because of f3i = 0. A(l) = tv.4). This is the only way in which arrivals can occur.i. let { Jt 1) }. and the marked transitions are then the ones corresponding to arrivals. A(1'k) A(2 k1). the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. Jt = (Jtl). j(2) } be two independent environmental processes and let E(k).2).6i ) diag. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. the claim surplus is a Markov additive process (cf. Indeed. A ( 2) = A (2`1 ) ® A. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. Here are some main examples: Example 5 . with common distribution B. Bii = Bi . where qij is the probability that a transition i * j is accompanied by a claim with distribution. that Bii = Bi . Bij = B. the definition of Bij is redundant for i i4 j.(13i )diag. Thus . u Example 5 .^) etc. refer to notation) { Jt k) }. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . II.2 for details). In the above setting. A(l) = T. Note that the case that 0 < qij < 1. Jt2)) (2. Again .174 CHAPTER VI.d. . we use the convention that a1i = f3i where 3i is the Poisson rate in state i.
all Al i2.. claims occur only at state transitions for the environment so that AN2. and that the policy then expires. Bilo. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. Thus. iN. DIVORCED.}. u Example 5 . the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]...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). i2i . after which it starts afresh.5. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. RETIRED. iN. THE MARKOVIAN ARRIVAL PROCESS Bij. INVALIDIZED.iN C17 AilO. In fact .kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. E = { WORKING.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states.. say. However . 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... or... The versatility of the setup is even greater than for the Markovmodulated model..iil. E 10..kl is redundant). Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate ... with rate ai.iN. In this way we can model. 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 .1i2.iN = a2. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. possibly having a general phasetype sojourn time.... assume that there is a finite number N of policies.iN. the kth policy enters a recovering state. • upon a claim. iN = all BOi2. This means that the environmental states are of the form i1i2 • • • iN with il.iN are zero and all Bi are redundant..iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. Similarly.iil.. e. DEAD etc. 11. as the Markovian arrival process ( MAP). more recently. MARRIED. WIDOWED...g.1i2 ... superpositions of renewal processes. Example 5 .
For the Markovmodulated model. p(t) and B(t) are defined also for t t [0. we talk of s as the 'time of the year'. B* = J f B(t) ((*) dt. By periodic extension. Without loss of generality. Let 1 1 /3* _ f /3(t) dt.176 CHAPTER VI. . for s E E = [0. 1). • The premium rate at time t of the year is p(t). one needs to assume also (as a minimum) that they are measurable in t. MARKOVIAN ENVIRONMENT sense any arrival stream to a risk process can be approximated by a model of the type studied in this section : any marked point process is the weak limit of a sequence of such models . from an application point of view. [248]. p * = 0 p(t) dt. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. Obviously. Thus at time t the premium rate is p(s + t). Lucantoni [248]. Lucantoni et at. where i f00 xB(°) (dx) _ . )3 t 1 J (6. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. we may assume that the functions /3(t). • Claims arriving at time t of the year have distribution B(t). Sengupta [336]. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . Some main queueing references using the MAP are Ramaswami [298]. continuity would hold in presumably all reasonable examples. 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.3*µs • p = f /3(v) dv 0 0 (6. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t).1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* .p)/p. let the period be 1. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. 0 < t < 1. Neuts [271] and Asmussen & Perry [42]. 1).
9).3(t) = 3A(1 + sin 27rt). of the periodic model as arising from the compound Poisson model by adding some extra variability. In particular. in agreement with the general principle of added variation increasing the risk (cf. In contrast. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . 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 . Thus.6. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. equivalently. Thus . for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. the average compound Poisson model is the same as in III. The behaviour of the periodic model needs not to be seen as a violation of this principle. B*.w(t). the discussion in 111.1 As an example to be used for numerical illustration throughout this section. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . or. since the added variation is deterministic.2 Define T 6(T) = p(t ) dt.10. (6.1. one may think of the standard compound Poisson model with parameters 3*.w(t)) dt). let . St = SeI(t).8. respectively. u Remark 6 . and thus the averaged standard compound Poisson models have the same risk for all A.3) Note that A enters just as a scaling factor of the time axis. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. Section 4b).(3. 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. not random. it turns out that they have the same adjustment coefficient. We u assume in the rest of this section that p(t) . we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations. In contrast. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. Many of the results given below indicate that the averaged and the periodic model share a number of main features. p* as an averaged version of the periodic model. It is easily seen that . the conditional distribution .3* = 3A.t.1) and Example 1. Example 6 .
The exposition of the present chapter is basically an extract from [44]. r(u) _ inf It > 0 : St > u} is the time to ruin . let f 8+1 tc *(a) _ (B* [a] .f. The claim surplus process {St} may be seen as a Markov additive process.(1 .5 (see in particular Remark 11. e.tc* (a)] dv then h (. with the underlying Markov process {Jt} being deterministic period motion on E = [0. of the claim surplus process..a) = exp { .f.. Notes and references The model has been studied in risk theory by. 0 (5)(u. 1). To this end. we start by deriving formulas giving the m. Daykin et.s .1) a = J8 . [101] . we obtain E.e.1) .Q(v) (B(„) [a] .(8) [eaSt+dt I7t] = = (1 .8). of the averaged compound Poisson model (the last expression is independent of s by periodicity).g. i. t + dt] or not. see the Notes to Section 7).3(v)(B(vl [a] . and define h(s. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00). with some variants in the proofs.1]) .1) dv .4) At a first sight this point of view may appear quite artificial.a) Proof Conditioning upon whether a claim occurs in [t.a.g. J Theorem 6 .al. (6.^8 [.(3(s + t)dt)e«St adt + /3(s + t)dt . given that the ith claim occurs at time t is B(8+t).. 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. Jt = (s + t) mod 1 P(8) .178 CHAPTER VL MARKOVIAN ENVIRONMENT of U.T) = P(8)(r(u) <T).g.3(s + t)dt[B(8+t)[a] . 3 E(8)eaSt = h(s.5. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.a be the c.adt +. Dassios & Embrechts [98] and Asmussen & Rolski [43]. a) is periodic on R.east B(8+t) [a] east . As usual.a . . 6b Lundberg conjugation Motivated by the discussion in Chapter II. a) etw*(a) h(s+t.
t} is a multiplicative functional for the Markov process { (Jt.9 as follows.0(s + t)dt[B(8+t)[a] .1].log h(s. Proof In the Markov additive sense of (6. a) .5 The formula for h(s) = h(s. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et. so that obviously {Lo.t. With g the infinitesimal generator of {Xt} = {(Jt.c* (e) {Le. a) h(s + t. a).s.(8)east 179 = = = = = E(8)east (1 . 0) P(8)a.2. a) Corollary 6. B) eoSt t.4). it then suffices to note that E(8)Le. a) Thus E(8)east = h(s + t. St)} and . we can write Lo Jt.6. a) et. u Remark 6. a) = exp I f t3(v)(kv)[a) . + v)(B([a] .adt +.5. 0) exist and are finite. 9) is a P ( 8)martingale with mean one.1)dv l og E(8) et where atetk•(a) h(t.1)dv  o h(t.1]) . at + f log h(s + t.4 For each 0 such that the integrals in the definition of h(t .(e) Let = h( h(Jo. a) = h(s.1]) .3(s + t)[D(8 +t)[a] .t}t>o = h(s.3.t = 1 by Theorem 6. dt log E(8)east a + f3(s + t) [B(8+t) [a] .. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11.* (a) h(s.9) eastt. E (8)east (a +. According to Remark 11. St)} . h(s + t.6 .
3. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. Proposition 6. Lemma 6 . That is.3(s)dt • B(s)[a]h(s) = gha(s. of St is as for the asserted periodic risk model. (iv) finally.3(s)h(s) + h'(s) +.5 that we can define a new Markov process {(Jt.'y). correspond to a new periodic risk model with parameters ex .a . y solves n* (y) = 0. Proof (i) Check that m. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le.4.T. see [44] for 11 a formal proof.1) .3(s)B(s) [a]h(s).(3(s)dt) +. the requirement is cha(i.1. Sdt) = h(s + dt) eadt (1 . ry)) at which n* (a) attains its minimum. . 0) = h(s) + dt {ah(s) . Now define 'y as the positive solution of the Lundberg equation for the averaged model.tc] dv} (normalizing by h(0) = 1).180 CHAPTER VI.y) = eayh(s).0) = Kh(s). [70] . St)} with governing probability measures Fes). cf. That is. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + .6 The P(s). yo is determined by 0 = k* (70) = QB*. When a = y. Proposition 6.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. say.f. B(s). as above E (s) ha(Jdt. it follows by Theorem II.60(t) = a(t)B(t)[0]. (ii) use Markovmodulated approximations (Section 6c). For each 0 satisfying the conditions of Corollary 6. P(s) (T(u) < oo) = 1 for all u > 0. 0 < s < 1. However. J s [.3(v)( Bi"i [a] .6 ( s ) exp { 0( s )&s) [a] + tc . we put for short h(s) = h(s.2. ( iii) use approximations with piecewiese constant /3(s). Bet)(dx) = ^ B(t ) (dx). A further important constant is the value yo (located in (0.7 When a > yo. such that for any s and T < oo.g. MARKOVIAN ENVIRONMENT ha(s.
1) the distribution of (l: (oo). a) TI h(9(u). ^(u) = ST(u) . 9(u)) for any bounded continuous function (e. we get: . 0(u)) * (b(oo). 1).2. e(cc)) Letting u > oo in (6. J C R+ such that the B(8). xEJ 0 (s)b(8)(x) > 0.1. a) a > ry0 (6. T(u) < (6. we need the following auxiliary result . B(oo)). s E I. Corollary 6.9(u))} u>0.7) h(B(u). the Markov process {(^(u). has a unique stationary distribution. T) = h(s. Lemma 6 . (6.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. have components with densities b(8)(x) satisfying inf sEI. The proof involves machinery from the ergodic theory of Markov chains on a general state space.8) (6. a) e«uE(8 ) e «^ .4.10) Then for each a.2). and no matter what is the initial season s. u which is > 1 by convexity. considered with governing probability measures { E(8) }E[ . the mean number of claims per unit time is p« 181 = Jo 1.9 Assume that there exist open intervals I C [0. Wu).8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u.6.9) 0(')(u) = h(s.g. which is not used elsewhere in the book. and we refer to [44]. q) = eryx/h(q)). RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. say s0. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. f (x. Here and in the following. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11.9) and noting that weak convergence entails convergence of E f (^(u).
1. 6. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c.182 CHAPTER VI. Noting that ^(u) > 0 in ( 6.ir) } Plots of h for different values of A are given in Fig. MARKOVIAN ENVIRONMENT Theorem 6. it does not seem within the range of our methods to compute C explicitly. At this stage .) C = E1 h(B(oo)) u + oo.16. A=1/4 A=1 A=4 0 Figure 6. illustrating that the effect of seasonality increases with A. this provides an algorithm for computing C as a limit.1 In contrast to h. 1. (6.10 Under the condition (6.9).11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary. Theorem 6 . For our basic Example 6 .W. 10 shows that certainly ry is the correct Lundberg exponent.10) of Lemma 3. Vi(8) (u) .6 for the Markovmodulated model: Theorem 6 . elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs .Ch(s)ery". 11 7/'O (u) < C+°)h(s) ery". where e. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3.11) Note that ( 6.1. Among other things. where C(o) = 1 + info < t<i h(t) .
yu) 000 (u) . e7 ( yx)B(t)(dy) > Then for all $ E [0. Consider first the timedependent version of Lundberg's inequality. RISK THEORY IN A PERIODIC ENVIRONMENT Thus. Just as in IV. We state the results below.42 • exp {J_ cos 27rs .15) The next result improves upon the constant C+) in front of eryu in Theorem 6. we substitute T = yu in 0(u. we first note that the function fu° exu {w • 3e . in our basic example with A = 1. 1 (6.42 so that 183 tp(8) (u) < 1. Theorem 6.13 to our basic example.yr. (ay).(8) (u. T) and replace the Lundberg exponent ry by ryy = ay . Theorem 6 . e. .11 as well as it supplements with a lower bound.w)e4u . r.g.12 Let 00)(y) 1 Then info < t<i h(t. whereas ay < y.3x + (1 . #c( ay) < 0 when y > 1/tc'('y).14) < C+)(y)h(s) e7yu.0(8) (u+ yu) (6.17) (6. where ay is the unique solution of W(ay) =y• (6.7e .7x j dx _7x } _ 6w + 6(1 .13) Elementary convexity arguments show that we always have ryy > Y and ay > ry. the proofs are basically the same as in Section 3 and we refer to [44] for details.4.. Lundberg's inequality can be con siderably sharpened and extended.w ) • 7e u{w • 3e3x + ( 1 . 1 ) and all u > 0.(ay) > 0 when y < 1/ic' (7).47r sin 27rs + 167r cos 47rs . we obtain Co) = 1.167r I Cu. C_h(s)e7u < V.w)e4u dx 9w + 7(1 .12) As for the Markovian environment model.6.16 In order to apply Theorem 6.13 Let = 1 B(t) C o<tf i h(t) 2no f °O e'r(Yx)B( t) (dy)' (x) x 1 B(t) (x) C+ = sup sup o<t<i h ( t) xo J.(s)(u) < C+h(s)e7". (6. ay) • (6.w) .
. C+ = 1. for A = 1 (where 3 e0.013.\ = 0 .I eu.184 CHAPTER VI. . The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'. 1/i18 1 s (u) > 0.g. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. 1).18) Notes and references The material is from Asmussen & Rolski [44].66. yo). .20). completing a cycle ..4^ sin 2irs + 16^ cos 41rs .20 •exp { 2n cos 27rs . Of course.013.181 s(u) < 1.. where the environment at time t is (s + t) mod 1 E [0.cos 27rs . 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. the nth Markovian environmental process {Jt} moves cyclically on {1.0.L sin 27rs + 1 I cos 47rs .(8)(u. 1) for the environment). This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.T) < C+('Yo)h( s. 14 Let C+('yo) be as in (6.\ 3 C+ = sup 6 exp { A (. Thus C_ = 2 inf ex cos 2irs . Then 7oudT . with s the initial season. n}.66. and in fact. Some of the present proofs are more elementary by avoiding the general point process machinery of [44].9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.19 } 0 <8<1 8 + cos 21rs Thus e. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0.16. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.16) with 'y replaced by yo and h(t) by h(t. and let 8 = er' (Y0). exp 2^ cos 21rs . Finally. 0 <'p(8)(u ) . but thereby also slightly longer.'Yo)e (6. we have the following result: Theorem 6 . Thus.1 sin 2irs + 16_ cos 47rs .1 sin 27rs + 1 cos 47rs .19 I eu.
z/'i (u. DUAL QUEUEING MODELS 185 within one unit of time on the average . Thus.jEE of the risk process.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. 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.7.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. 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.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. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). one simple choice is Oni = 0( i . This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. (6. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model.1 ((i 1)/n) ) and Bni = B . M(n) = Supt>o Stn). We let {Stn)} (6. since the settings are equivalent from a mathematical point of view. Bi. To this end. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). Let 0j. . but others are also possible. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. it is desirable to have formulas permitting freely to translate from one setting into the other.
J* = i) for all j. (7. J*) is the steadystate limit of (Vt. I* )3i P(V > u.2) and use that limF (VT > u. Jo = i.n(VT > u. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. and (7.3). JT = Z).. Taking probabilities and using the stationarity yields 7riPi(T(u) < T. (VT > u I JT = 2).4) where 0* = >jEE 7rj/3j. T) = 7ri 1 P. 2 .1).2). Jo = j.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. Proposition 7. (7. The actual waiting time process 1W1. For (7.1 Assume V0 = 0. Jt ). JT = j) = 7rjPj(VT > u. .3. let T .oo in u (7. JT = j} and {VT > u. JT = i} coincide.3) 7ri where (V. JT = i) = P(V > u. J* = i).. (7. (7.P(V > u. just sum (7. Proposition 7. JT = i) = 'P. and for (7.0i (u . J* = i) = P.T(V > u I J* = i).=1 . JT = j) = LjPj (VT > u. Then Pi(T(u) < T.. JJ = i). In particular.1) 7ri In particular. ii (u) = it /3 P(W > u. I* = i). 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. Proof Consider stationary versions of {Jt}o<t<T. Now let In denote the environment when customer n arrives and I* the steadystate limit. . The first conclusion of that result then states that the events {T(u) < T.2) Oi(u) = 1.1) follows.186 CHAPTER VI. • The queueing discipline is FIFO. {Jt }o<t<T• Then we may assume that Jt = JTt.1) over j.
[243]. that /3(t).4) can be found in Regterschot & de Smit [301].l. if T is large. P(W >u. n=1 N However. on average /32TP(V > u. I*) with the timeaverage . and of these.I *=i).T)(T(u) <T) = P(8)(VT > u).. and Rolski [306]. P(1')(r(u) < oo) = P(')(00) > u). and (7. see Regterschot & van Doorn [123]. a paper relying heavily on classical complex plane methods. (7. and further references (to which we add Prabhu & Zhu [296]) can be found there. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. . Proposition 7. A more probabilistic treatment was given by Asmussen [17].4).=i) a4. Lemoine [242]. a general formalism allowing this type of conclusion is 'conditional PASTA'.4) and (7.7. 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.3). DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. >u. In the setting of the periodic model of Section 6. B(t) have been periodically extended to negative t).7) of that paper. T]. P(.5) follows from (7. see in particular Harrison & Lemoine [186].1 is from Asmussen [16]. N * oo. With {Vt} denoting the workload process of the periodic queue. I* = i.7) (7.g. on average 0*T customers arrive in [0. Taking the ratio yields (7. The relation (7.3) improving somewhat upon (2. we have 1: I(W.I. J* = i) see W > u. with (7.6) (7.o. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301].8) For treatments of periodic M/G/1 queue. and one has PI'>(rr(u) < T) = P('_T)(VT > u).
This page is intentionally left blank .
T) = F(T(u) < T).d. However . and that the claim sizes U1. 189 . Zt As earlier.i. and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). i&(u. . and the evolution of the reserve may be described by the equation Rt = u . Thus. are i. z/i(u) = F IinffRt< 0IRo=u 1 (1..6. t] are Nt At = Ui (1. 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. the aggregate claims in [0. with common distribution B and independent of {Nt}.. finite horizon. {Rt} moves according to the differential equation R = p(R). U2.2) tk(u.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 . Thus in between jumps. resp .At + p(R8) ds.1) (other terms are accumulated claims or total claims).
rather than when the reserve itself becomes negative. i. but when the reserve comes above v.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. that {Rt} will reach level u before the first claim arrives. or o(u) < 1 for all u.p2. pi > p2 and p(r) = One reason could be competition. No tractable necessary and sufficient condition is known in complete generality of the model. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . That is. Hence in terms of survival probabilities. it seems reasonable to assume monotonicity (p(r) is u . Now return to the general model. 1 .i(u) = 1 for all u. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. say at interest rate b. Proposition 1. Thus at deficit x > 0 (meaning Rt = x).p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. Example 1.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. dividends are paid out at rate pi . there is positive probability.Vi(v) u > e(1 . we get p(r) = p + er. If Ro = v < u. Assume 0(u) < 1 for some u. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p.e. However. Example 1. say e. where one would try to attract new customers as soon as the business has become reasonably safe.190 CHAPTER VII. and the probability of absolute ruin with initial reserve u E [p/S. P(r) _ p + e(r . but assume now that the company borrows the deficit in the bank when the reserve goes negative.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. when x > p/S. we can put Rt = Rt + p/S. Proof Obviously '(u) < ilb(v) when u > v. In this situation.'(u)) > 0 so that V'(v) < 1. oo) is given by i (u + p/S).4 Either i.2.
Let Op(u) refer to the compound Poisson model with the same 0.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.1. we have z/i(u) <p(u . In particular. one can couple the risk process and the storage process on [0. INTRODUCTION 191 decreasing in Example 1. {Vt} decreases at rate p(v) when Vt = v (i. (1.3. instead of (1.T) = P(VT > u). Hence ik(u) < 1 for all u by Proposition III. 296297): Theorem 1.1. Then if u > no.+ p(r) exists. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. obviously infu<uo z/'(u) > 0. In case (b).1 and increasing in Example 1. Proof This follows by a simple comparison with the compound Poisson model. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. { Vt} remains at 0 until the next arrival). Then 0(u) = P(V > u). (1. then ?(u) = 1 for all u.2 once more. let uo be chosen such that p(r) > p = 0ILB + e for r > uo.3µB for all sufficiently large r.5 (a) If p(r) < /. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. In case (a). hence Rt < uo also for a whole sequence of is converging to oo. V = p(V)).uo) and.b(u. [APQ] pp. then l/i(u) < 1 for all u.6) .1. that u zPp(u . However. if and only if V)(u) < 1 for all u. B and (constant) premium rate p.1. say V. That is. Proposition I1I. Theorem 1.. cf.2(d). We next recall the following results. In between jumps.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.2) for r sufficiently large so that p(oo) = limr. This is basically covered by the following result (but note that the case p(r) . which was proved in 11. (1.2(d)).uo) < 1. Starting from Ro = uo. and P(Rt + oo) > 0. and hence by a geometric trials argument. .5) and the process {Vt} has a proper limit in distribution .6 For any T < oo.f p(Vs) ds.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo.2) we have t Vt = At . Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0.4.e. let uo be chosen such that p(r) < p = /3µB for r > uo. appealing to Proposition 111.
Qw(x) .y)g(y) dy. (1.h. the l.6x and that w(x) < oo for all x > 0.h.7) Proposition 1. oo).8) as g(x) = p 1 {yo13e_6x +. the flow of mass from [0.8) as the rate of upcrossings.192 CHAPTER VII. B(x) = e. In view of the path structure of {V t }. one having an atom at 0 of size 'yo. Note that it may happen that w (x) = oo for all x > 0. say when {Vt} is in state y.8) is the rate of downcrossings (the event of an arrival in [t. Then the ruin probability is tp (u) = f' g(y)dy.Sx} dx. where g(x) = p( ^ exp {. oo) must be the same as the flow the other way. and is succesful if the jump size is larger than x .9) Proof We may rewrite (1. Considering the cases y = 0 and 0 < y < x separately.8 Assume that B is exponential with rate b. and the other being given by a density g(x) on (0. t + dt] if and only if Vt E [x. An attempt of an upcrossing occurs as result of an arrival. we arrive at the desired interpretation of the r. x + p(x)dt]). say. Oeax f x e'Yg (y) dy } = p) eaxa(x) . x] to (x. (1. of (1. Corollary 1. for the storage process {Vt}. of (1.6 applicable. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. we thus need to look more into the stationary distribution G.6w(x) . Jo AX) (1. It is intuitively obvious and not too hard to prove that G is a mixture of two components. It follows in particular that 0(u) = fg(Y)dy. yo ^ 1 + oo Q exp {. say if p(r) goes to 0 at rate 1 /r or faster as r j 0. Then w(x) is the time it takes for the reserve to reach level x provided it starts with Ro = 0 and no claims arrive.Sx}. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1.7 p(x)g(x) = tofB (x) + a f (x .y. u Define ^x 1 w(x) Jo p(t) dt. Now obviously. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t.s.8) Proof In stationarity.s. say.
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
3a/ (5 . of Z is IogH[a] = f ytc(y)dy = e fa (0. The process {St} corresponds to {Wt} so that c(a) or2a2/2 . then {Rt} is the diffusion with drift function p+Ex and constant variance a2. RESERVEDEPENDENT PREMIUMS u from which (2.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.a) .V. where V is Gamma(b. From ic(a) = .x) dx e.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2. i.2) follows by elementary algebra.3.5 The analysis leading to Theorem 2.V < x)]0 + f P(V > p/E ) + eby fv (p/E . As an example.pa. and the c. Proof 2 We use Theorem 2. . 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 .13 /E) r (.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments. 13/E).f.2y +µ ) dy ./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E .e. r (j3/E) In particular.200 CHAPTER VII.g.2) follows by elementary algebra.pa. /^ u Example 2 .3/E) By the memoryless property of the exponential distribution. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .
8. write y* for the solution of the Lundberg equation f3(B[ry *] . see e.3 is from Harrison [185]. that the analysis does not seem to carry over to general phasetype distributions.e.3. Paulsen [281].g. se e. Paulsen & Gjessing [286] and Sundt & Teugels [356]. Some of these references also go into a stochastic interest rate. Z is normal (p/E. Emanuel et at. Delbaen & Haezendonck [104].p*. The formula (2. as in the proof of Proposition 2. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case).1) . or to nonlinear premium rules p(•).. not even Erlang(3) or H3. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. Goldie & Griibel [167].4 is classical.v. for a martingale proof. [282]. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. it is also used as basis for a diffusion approximation by these authors. [357].. Further studies of the model with interest can be found in Boogaert & Crijns [71].d. and recall Lundberg 's inequality . 3 The local adjustment coefficient.3) was derived by Emanuel et at. [129]. [283]. The solution is in terms of Bessel functions for an Erlang(2) B and in terms of confluent hypergeometric functions for a H2 B (a mixture of two exponentials).g. Q2/2E). and since RT = 0 by the continuity of Brownian motion. Gerber [155]. however. [129] and Harrison [185].Y*p* W*(u) < ery*u = 0. A r. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . Corollary 2. write Vi* (u) for the ruin probability etc. It must be noted. of the form Ei° p"X" with the X„ i.2 is a special case of a perpetuity. Gerber [157] p.i. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2.
choose c(. Then lim sup u>oo u and e E''p(r) + 0.1 Assume that for some 0 < 5o < oo.C*ef*". The steepness assumption and p(x) + oo ensure 'y(x) * So. log ?i(u) < < 00 JO . as will hold under the steepness assumption of Theorem 3. 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). oo for all E > 0. a first step is the following: Theorem 3 .log '(u)/u < ry*(1 .ap(x).w (u) J dt > c(3)eeu v 1 p(u+ t) . B(x) > C(2)e(ao+f)x for all x.1 ).E).1. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.4) we assume existence of y(x) for all x..'y ( x)) = 0 where r.>o 7(x) > 0. and that p(x) * oo.1.3) When trying to extend these results to the model of this chapter where p(x) depends on x. a) = f3(B[a] .5) which implies inf. x>0 (3. e(1o+e)2 (x ) u > 00. x * oo. we will use the local adjustment coefficient 'y(x).1) . (3. then log u (u) In the proof as well as in the remaining part of the section . c(. If 60 s f 6o. it holds that f3[s] T oo. and (for simplicity) that inf p(x) > (3µs . choose uo such that p( x) > p* when x > u0E. 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 . Let y* < So. 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) .. (3. For the last asssertion .i)eex. i.e. When u > uo. as solution of the equation n(x. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. i. 1) and for a given E > 0. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. let p* be a in (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.e. Proof of Theorem 3. (x.*(u) .202 CHAPTER VII.2) such that p(x) < c(.
However. The form of the result is superficially similar to the CramerLundberg approximation. For e > 0. then Rte) = CRtie for all t so that V). (u) = O(u/e).' (u) < eI("). 3) = (1 . {Rte)} defined as in (1.2). Theorem 3 . (3. which essentially says that an overshoot r. and in particular.3. 2. (3. Then limelog l/ie (u) = I(u).v.7) CIO Remarks: 1. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. UJU > x cannot have a much heavier tail than the claim U itself.6) The second main result to be derived states that the bound in Theorem 3.e.. by cU2. . ruin will occur if the claim is at least u + v. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. one can then assume that e = 1 is small enough for Theorem 3.3 to be reasonably precise and use e` (u) as approximation to 0 (u). or (b) Condition 3. The rest of this section deals with tail estimates involving the local adjustment coefficient. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. the asymptotics u * oo and c . If p(x) = pis constant . I.3 Assume that either (a) p(r) is a non decreasing function of r. u Obviously. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U.g. Then .1 only presents a first step. 3. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 .13 below holds. 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. Theorem 3.4)eE" Given such an arrival. Condition 3.(u) > so. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx.2 is also an approximation under appropriate conditions.ea°/(ecf1)). noting that in many cases the constant C is close to 1. the result is not very informative if bo = oo.0 are the same. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(. Bucklew [81]).13 is a technical condition on the claim size distribution B.
bx} dx 1+0.bx}]o + b /' oo exp low (x) .(3/p(x).bx} dx fo 00 1 + [exp {/(3w(x) . the logaritmic form of (3.2.exp {/33w(u) . it is formally needed only for Theorem 3.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) . J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .bu}.6). First.bx} dx .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).3.4 Consider again the exponential case B(x) = eax as in Corollary 1.bx} dx = 1 + J0 dodx(x) exp {. Then y(x) = b . RESERVEDEPENDENT PREMIUMS 4. 5. 3. As typical in large deviations theory. However. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.3.1 + b f e. we show how to rewrite the explicit solution for ti(u) in Corollary 1.(x) dx. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3. rather than eI(u)). we consider some simple examples. say. we get = 1+ J" AX) exp {(3w (x) . u .(iw(x) .204 CHAPTER VII.8 in terms of I(u) when the claims are exponential: Example 3 . 3a Examples Before giving the proofs of Theorems 3.8.
fo 7(x) dx /E dy > av 'yo /Edy = E (1 . THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . (3.fory(x+u)dxdy ( 3.ev 0 O /E) J0 70 70 Yo This implies lim inf A.e. Similarly.. For Theorem 3..1/8 . we get r 00 e.9 ) 11000 eI(v)dy f000 e.2. 70 > 0 such that 7(x) < 7o for y < yo. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. It is well known that (see Theorem XI.10) where AE = e log 000 e.10 or Karlin & Taylor [222] pp. and (3. Be = e log U000 e. in the definition of AE converges to 0. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.I ( v )dy fo +u) dxdy . IE(u) = I(u)/e.fa 7(x+u)dx/Edy o The analogue of (3. u .3.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. 0. BE * 0. ry(x /b I u o e f0 °° e  e.5 Assume that {Rt} is a diffusion on [0. note first that the appropriate slow Markov walk assumption amounts to u. > lime log e = 0 and AE * 0.0. (3. and (3.7) follows.9) yields e log . 3. The appropriate definition of the local adjustment coefficient 7(x) is then as the one 2p(x)la2(x) for the locally approximating Brownian motion. oo) with drift µ(x) and variance a2 (x) > 0 at x.1.2(X) = ev2(x) so that 7e(x) = 7(x)/e.2.f y(x)dxd y If 7(x) is increasing .fo 7(x)dx/Edy f . In particular.5) is infx>o 7(x) > 0 which implies that f °O .BE. (u) = I(u) + AE . (X) = µ(x).3. Choosing yo.I ( u) fool.3 in the particularly simple case of diffusions: Example 3. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.
. Ignoring 1/5 in the formula there.6) exactly as in Example 3. E+o e*O By (3. 0 Now (3. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.6 Assume that B is exponential with rate S.Q/p*.5) and 7* = 5 .206 CHAPTER VII./3 1 AX dx. Then the solution of the Lundberg equation is y* = b . Of course. the results are suggestive in their form and much more explicit than anything else in the literature. lim sup Af < lim sup c log(1 . Nevertheless . > .6/p* so that u 1 I (U) = bu . G. RESERVEDEPENDENT PREMIUMS The analogue of Example 3.e.1 3. the slow Markov walk assumption means 5E = b/c.0/e. .5. however . ..+1 (u) = o( 1).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. we have 5 > 7o and get lim inf AE > lime log e . .(u) oo.. that the interchange of the slow Markov walk oo is not justified and in fact.5 for risk processes with exponential claims is as follows: Example 3 ..5. _ . 0.10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. Thus 7e(x) _7(x)/e and (3. As in Example 3. . Further. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. .0. + Gq(u) + o(G9(u))• Gi (u) It should be noted . G. so our approach is to determine standard functions Gl (u).5. I(u ) = G1(u) + . G. 7(x) is typically not explicit. this leads to (3.7o C 15 I I.4.0) = 0. I.
y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.cs(1 .7 Assume that B(x) .3.clxale5x 207 (3.3cse7*I7(77) .:. 2. in the phasetype case . x T 1. This covers mixtures or convolutions of exponentials or.12) with y > 1.11) with a > 0. c4 = c2b 1/'/(1 . Here B[s] is defined for all s and B[s] .1/k). if the phase generator is irreducible ( Proposition VIII. .8 Assume next that B has bounded support. y = 2 if B is uniform on (0. ry* loge*+ g7loglogp*.Y . 1. I(u) Pt. It follows from (3.4) or gamma distributions.1) leads to .ry*°p*. phasetype distributions (Example 1. More precisely. fu I(u) Su .11) that b[s] * co as s f S and hence 7* T S as p* + oo. Hence (3. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S . e. 77 = 1 if B is degenerate at 1. and hence (3. more generally.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . say 1 is the upper limit and B(x) .C2p* C2 = (3clr( a))11'. u(logu + r7loglogu).c2 Su a dx ) Su a<1 Su . 1) and 17 = k + 1 if B is the convolution of k uniforms on (0. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 .. .x)n1.g.8).1/a).1) leads to (S7T N Ocp a.s)C' f "o o as s T S.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b. u Example 3 . For example. the typical case is a = 1 which holds . (3.
Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.ru(TI)) .12).1) . (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3. of the increment in a small time interval [0.sp(u).g. 1 = E.4) is the formula h logEues ( Rhu) . x f oo . 3b Proof of Theorem 3.B[7o (u)] . . h]. this is only possible if 7o(v) 2 7o(u)• . 7 * .3 (B[s] .14) for the m .(t))dt.11) and (3..f. ec78)2/2c7 dx C7 .log p*.15) Proposition 3. (3.Ote7o( u)(u. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u . (3. By convexity of the m .r^. g. of U1 + v .4). (3.u is a nondecreasing function of u.•.10 Assume that p(x) is a nondecreasing function of x. Hence for u<V. assume that B(x) CO x2/2c7.css 2%rc7eC782/2.208 CHAPTER VII.f. one could also have considered the increment ru (T1) .4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3.Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u).u . RESERVEDEPENDENT PREMIUMS Example 3 .13) We get b[s] . The assumption implies that ru(t) .2 We first remark that the definition (3.(T1)) > Ee7o(u)(ul+vr»(Ti)). 1 0 3e. h 10. I (u) c8u log u 0 where c8 = 2/c7.r„(Ti).1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .9 As a case intermediate between (3.c8 log .e7o ( u)(ul+u r.
e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] .17) shown for n and let Fu(x) = P(U1 + u .u > tp(u). Also.11 Assume that p(x) is a nondecreasing function of x. the case of 7 then follows immediately by Proposition 3. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0. Assume (3.7o (u)p(u)• Since (3.2 in terms of 7o.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).4) considered as function of 7 is convex and 0 for y = 0.3. Hence 1 = EeYo(u)(U1+uru(T1)) < E.u[70(u)] fo eyo(x)dx .(n+l) (u) 1 . Hence „/. fa 7o(y)dy < u7o(u) < xyo (u) for x > u. note that the assumption implies that ru(t) .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.17) from which the theorem follows by letting n + oo. 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 (3.1) . we obtain „I.ru(T1) < x).Fu(u ) + J ^(n)(u . Then (u) < efo Yo(x)dx.(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J . it is easily seen that fu x7o(y)dy < xyo (u). this is only possible if yo(u) > 7(u).es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j. u We prove Theorem 3.10(b): Theorem 3. (3. THE LOCAL ADJUSTMENT COEFFICIENT For (b).
n so that n. define uk. given downcrossing occurs. we used also Proposition 3. ryk. Proof For ruin to occur.3).e (u) = v'. Also.n AX).x/n.n.n <Z auk}l. (3. y* evaluated for p* = Pk. 3. 0.n = ku. 0 It follows from Proposition 3.n.12 lim sup4^o f log O.n.: y(u).3 The idea of the proof is to bound { R( f) } above and below in a small interval [x .2)..E ( u/n) ^•e. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3.2. For these reasons.nbe C*. op*.n (starting from u/n) without that 2u/n is upcrossed before ruin. in accordance with the notation i/iE (u). in .210 CHAPTER VII.E (u) denote the ruin probability for the classical model with 0 replaced by ..10(b ) that the bound provided by Theorem 3.. and. the value of {R(E)} at the time of downcrossing is < unl.I..n) pn niE (u /n) n n_1 n. To this end. by €U=.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3.15). 3c Proof of Theorem 3. and here it is easily seen that yo(u) .n) > k =1 II v ^k n. yo(u) appears more difficult to evaluate than y(u).3). pk n = uk_l..3/e and U. (u). W O . {RtE)} (starting from u = un.n u k}1. 0. resp. Y*u /E.n.E (u/n). However.E (u/ n) Y'E (un . P k. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3.11 is sharper than the one given by Theorem 3.2.n) must first downcross unl. let Op*.n inf n uk1.E (u/n) Now as e . (u) < I(u). Let Ck. C*e where the first equality follows by an easy scaling argument and the approximation by (3. Lemma 3..n = sup p(x). Further. The probability of this is at least n n.10(a) for some of the inequalities. for either of Theorems 3. the probability that ruin occurs in the CramerLundberg model with p* = pn.3 is required. we have chosen to work with y(u) as the fundamental local adjustment coefficient. (un2.
11 Theorem 3.n cE (2u/n) Ck ne7k.12 completes the proof. Indeed . 3 now follows easily in case (a).7k. *p•. 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. ne7k.i. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).log Ck.nu /fn( 1 Ck  e.n <X<Uk.. uk_1.n.. for all x .n = sup ?'(x).nu /En) o(1)). k=1 k=1 n u _ nE7 k.E (u/n) OP +^p•.a( u)z.E (u/n) Op•.nk=1 limsupelogv).nu/en(1 + where o(1) refers to the limit e . in obvious notation one has tC (x) = y(x)/e.3.n + 0(1). /' (u/n) 'T nk. 211 Clearly. .F (2u/n).18) (ii) the family of claim overshoot distributions is stochastically dominated by V. so that Theorem 3..E (urn) < \ *I. (3.19) . B(x) (3..e.! (u/n) n n m 7k. i.n . v. we need the following condition: Condition 3. It follows that n log V'C (u) k =1 log Ypk. 0 with n and u fixed. also ryk. since ry' is an increasing function of p'. 40 Combining with the upper bound of Lemma 3. THE LOCAL ADJUSTMENT COEFFICIENT particular. In case (b).13 There exists a r.
Then the standard Lundberg inequality yields El < E?.E (0) cf.QEU 1 . u/n) < oo) EV).v.of:>2 in n(x).nu/En0(1) . u /en 0(i) _n so that E2 < e2ryl nu/En0(1). (u/n . . u/n) < oo) .18) for the last equality).n V. Ei + E2 < e71.n < ery1.2y 1 ' .nu /EnE [e71.eV) • P (T(E) (u. where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest.E (u/n .EV) = e.EV) = El + E2.E (2u/n . N with EN < 1 = infx>2u/nA(x) = 0(1). RESERVEDEPENDENT PREMIUMS To complete the proof. V < u/En] + P(V > u/En) (u/En .212 CHAPTER VII. u/n) < oo] E [OE (u/n . Write EO.. infx>2u /n P(x) .R<) (u v). u/n)) I T(E) (u... The probability of ruin in between two downcrossings is bounded by Epp .EV) = EiI 1 .^'' = E [ . v ) = v .( . T(E) (u. v ) = inf { t > 0 : R(c ) < v R) = u } . T() (u.E(E) (u. u/n)) . let v < u and define T(E) (u. u/n) < oo] . (3.5) and the standard formula for b(0). For E2. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r.^(E) (u. u /n) < oo] l = = < E [OE (u/n . Then Y'E (u) ^(E) (u. (R. P (T(E) (u..1 n.V) = e71 nu/Eno(l) (using (3. ) (u u /n)) . (u/n .
=1 J An approximation similar to (3.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof. 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 3.1. Djehiche [122] gives an approximation for tp(u. T) is maximized over T by taking T as the time for (3. s). [89].t.4) and the prime meaning differentiation w. whereas the most probable path leading to ruin is the solution of r(x) _ k (x. Bucklew [81]). Whereas the result of [122] is given in terms of an action integral which does not look very explicit.7) then comes out (at least heuristically) by analytical manipulations with the action integral. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . the approximation (3.T) = P „(info<t <T Rt < 0) via related large deviations techniques.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. Typically.13.21) to pass from u to 0.) = exp .3. Comparing these references with the present work shows that in the slow Markov walk setup.J r(Rs)p(R.7(x)) (3. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u.)ds + Y(R2. the risk process itself is close to the solution of the differential equation r(x) _ r (x. one can in fact arrive at the optimal path by showing that the approximation for 0(u. Similarly.)Ui } . 0 ) (= p(x) .g. . l o JJJ o .3EU) (3. the results are from Asmussen & Nielsen [39]. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali. u Notes and references With the exception of Theorem 3. it might be possible to show that the limits e . 0 and b T 00 are interchangeable in the setting of [89].20) (with ic(x.21) (the initial condition is r(0) = u in both cases). where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . s) as in (3.r. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.J y(Rs)dR.
214 CHAPTER VII. For different types of applications of large deviations to ruin probabilities . 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. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . see XI. We should like. . the exponential distribution ).3.. e.g. however.
P. and not in other cases. refers to the case Jo = i. that is. 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<. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. Typically. if a problem can be solved explicitly when the relevant distributions are exponentials. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. Note that since (1. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 .1) is 'Here as usual . on Eo = E U {A} where A is some extra state which is absorbing. 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. More precisely.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. oo) is said to be of phasetype if B is the distribution of the lifetime of a terminating Markov process {Jt}t>o with finitely many states and time homogeneous transition rates. A distribution B on (0. We often write p for the number of elements of E. 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.
C is the lifetime sup It > 0 : Jt E E} of {Jt}. a.e. The initial vector a is written as a row vector. j. In particular. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. B(t) = Fa(^ < t ). We now say that B is of phasetype with representation (E. Then a = a1 = 1. T is a subintensity matrix2. 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. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. 0 2this means that tii < 0. k}. tij > 0 for i 54 j and EjEE tij < 0 . Here are some important special cases: Example 1 .e.3. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. i.0 = t11. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. that is. i.2) The interpretation of the column vector t is as the exit rate vector. Equivalently.1 Suppose that p = 1 and write . T) (or sometimes just (a. an exponential distribution with rate parameter . E = {i.1 The phase diagram of a phasetype distribution with 3 phases. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. (1. and we have t = Te. t1 = /3.216 CHAPTER VIII.
. . 6.. 0 •.. .1.x i=1 Thus E _ Si 0 T 0 S2 0 0 . .3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51.2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p.. so that the density is P E ai6ie6.....2 corresponding to E = {1.1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. the EP distribution may be represented by the phase diagram (p = 3) Figure 1... PHASETYPE DISTRIBUTIONS 217 Example 1. 0 S 6 Example 1.. . 0 0 0 0 S 6 . . ... 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 . 00)) S s o . 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •.. 0 SP 0 and the phase diagram is (p = 2) . p}. a = (1 0 0 .
the Erlang distribution is a special case of a Coxian distribution. the restriction of P8 to E. 5 Let B be phasetype with representation (E.g. see A. Then: (a) the c. 36) yields s d.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. [APQ ] p. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . . the backwards equation for {Jt} (e.1 tP1 1 Figure 1.3 0 Example 1 .f.aeTxe. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2.e. j E E. The basic analytical properties of phase type distributions are given by the following result . T). Theorem 1 .g. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . (b) the density is b(x ) = B'(x) = aeTxt. a.4 For example.d. dp.t2 yt bP. MATRIXANALYTIC METHODS Figure 1. i. p:. Then for i .3 .1)"n! aT"e. (c) the m.218 CHAPTER VIII.f is B (x) = 1 . B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. ds^ = ds' = ttlaj + tikpkj.
we i w. hj . Then h tit ti + ti3 h j .1 ) n +l n ! a (s I + T ) . B(n)[0] = _ Alternatively. .T) 't = (. Part (d) follows by differentiating the m.5) as hi(tii + s) = ti  t ij hj. the solution is P8 = eT8. d" dsn a (.tii tii .p.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). for n = 1 we may put ki = Ei( and get as in (1.g.6) .B(x) = 1'a (( > x) = P.s I .tii we go to A.p.1.g. j#i jEE tijhj + his = ti. (Jx E E) = this proves (a)..f. h = (T + sI)1t. d8 P8 = TP8.5) Indeed .s j# tii i (1.tii and have an additional time to absorption either go to state j which has m . this means in vector notation that (T + sI)h = t. (1) n+1n!aT .f. or w. 1. of the initial sojourn in state i. and since obviously P° = I.f.tii is the rate of the exponential holding time of state i and hence (tii)/(tii .e. = aPxe.f. Since 1 . and (b) then follows from 1: aipF. PHASETYPE DISTRIBUTIONS 219 That is.g. and since b[s] = ah. Alternatively.s) is the m . i. tij / . Rewriting ( 1. After that. we arrive once more at the stated expression for B[s].g. the rule (A. define hi = Eie8S.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI .5) ki = 1 + tii L jj:Ai tii (1. ti/ . i.n1t = (1)nn!aTn1Te (1)nn! aTne.T) 1t.n lt .. in which case the time to absorption is 0 with m . For (c).
there are some examples where it is appealing to write T on diagonal form. Example A3."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. T= 2 111 so that 2 2 Then (cf. are idempotent. another the case p = 2 where explicit diagonalization formulas are always available. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x .s. One obvious instance is the hyperexponential distribution. 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. Consider for example 3 9 a= (2 2).h. 0 Example 1.7) 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.220 CHAPTER VIII. see the Appendix. making the problem trivial.
T) with weight hall and an atom at zero with weight 1 .T). and in fact one also most often there allows a to have a component ao at A.g. a random variable U having a defective phasetype distribution with representation (a. (1. .7 If B is phasetype with representation (v.4b for definitions and basic rules): Proposition 1.29) and Proposition A4. • The phasetype distribution B is zeromodified.1.4. There are two ways to interpret this: • The phasetype distribution B is defective.f. i. This is the traditional choice in the literature.hall. 5 and serves at this stage to introduce Kronecker notation and calculus (see A. i. 0 Sometimes it is relevant also to consider phasetype distributions.7) Proof According to (A. then the matrix m. or one just lets U be undefined on this additional set. 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.e 11BIJ = 1laDD < 1. < 1. T) is then defined to be oo on a set of probability 1.e a mixture of a phasetype distribution with representation (a/llall.11aDD. hail = E=EE a. where the initial vector a is substochastic. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I). 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).
v. See in particular the notes to Section 6.q be the eigenvalue of largest real part of T. and we have eTx . the result follows (with C = (ah)(ve)). No satisfying . the Erlang case). let . Neuts. .. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form.4c). Schmidt & Teugels [307] and Wolff [384].g. Lipsky [247]. Schmidli. Example A5. assume that T is irreducible . one has k = 0. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. the conditions of Proposition 1. 2.F. Using B(x) = aeTxe . but the relevant T is not irreducible. Then the tail B(x) is asymptotically exponential. see his book [269] (a historical important intermediate step is Jensen [214]). distributions with a rational m. Here is a sufficient condition: Proposition 1.f.8 Let B be phasetype with representation (a. x * oo. h can be chosen with strictly positive component.8).f. In Proposition A5. In older literature. but in many practical cases. All material of the present section is standard.. 0 Of course. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. we give a criterion for asymptotical exponentiality of a phasetype distribution B. but todays interest in the topic was largely initiated by M. the text is essentially identical to Section 2 of Asmussen [26]. i is real and positive. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . where C.8) Proof By PerronFrobenius theory (A. Rolski.hve7x. cf. (or Laplace transform) are often used where one would now work instead with phasetype distributions.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail.1 of the Appendix.222 CHAPTER VIII. cf. 77 > 0 and k = 0. 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. Notes and references The idea behind using phasetype distributions goes back to Erlang. let v. not only in the tail but in the whole distribution. (1.Ce7'. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. The Erlang distribution gives an example where k > 0 (in fact. here k = p1).g. B(x) . T). 1.
+UnEA} 00 = EEI(U1 +. the jumps of the j(k) and the it } k) to the next J( k+l) A jump jumps corresponding to a transition from one Jt 4Here the empty sum U1 +.1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together. A related important unsolved problem deals with minimal representations: given a phasetype distribution .1 Consider a renewal process with interarrivals which are phasetype with representation (cr. as the lifetimes of items (say electrical bulbs) which are replaced upon failure. . For this reason.. is Markov and has two types of jumps .. Jt={Jt?ul}. U1<t < U1+U2. however.: U1 + .2. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. known.r.1. Then the renewal density exists and is given by u(x) = ae(T+ta)xt. U2.. with common distribution B and define4 U(A) = E# {n = 0. the renewals form a Poisson process and we have u(x) = 0. n=O We may think of the U. Lebesgue measure.. and U(A) is then the expected number of replacements (renewals) in A.. .. JtJt1) Then { 0<t<U1 .g. be i.1 of the Appendix.d.+UnEA). we refer to U as the renewal measure. + U0 is 0 . oo) w.f. but is in part repeated below..t. or the density is available ) is.T)... RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m.i. If B is exponential with rate 0.. .... Let U1. we denote the density by u(x) and refer to u as the renewal density. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. if U is absolutely continuous on (0. 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. (2. but nevertheless.
and let µB = aTle be the mean of B. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . define the excess life e(t) at time t as the time until the next renewal following t. and the jumps of the first type are governed by T.e.e. u The argument goes through without change if the renewal process is terminating.U1 U3 U2 U3 U4 Figure 2. since Uk = oo with probability 1 . and the distribution of Jx is ae ( T+t«)x.T) where vt = ae (T+ta)t . see Fig. as the time of the last renewal. Proof Just note that { it } is a governing phase process for the lifetime.1) remains valid for that case. this is welldefined. Then the lifetime is zeromodified phase type with representation (a. This is defined as U1 + . that is. IIafl < 1. Equivalently. . is the first k with Uk = 00. i. T).. which is phase type with representation (v.3 Consider a renewal process with interarrivals which are phasetype with representation (a. the lifetime of the renewal process. Hence the intensity matrix is T + ta.T). u Returning to nonterminating renewal processes . MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .1.224 CHAPTER VIII.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt.T) where v = aT1 /µB.1 Corollary 2. + Uit_1 where s.T + ta). However. the density is veTxt = B(x)/µB.. 2. Corollary 2. Hence ( 2. The renewal density at x is now just the rate of jumps of the second type.1) follows by the law of total probability. i. . which is ti in state i.IIBII which is > 0 in the defective case. B is defective . (b) £(t) has a limiting distribution as t * oo. fi(t) U2 U1 . and hence ( 2.
T) where vt is the distribution of it which is obviously given by the expression in (a). cf. The renewal density is then aeQtt = (al a2) ( 7i 7"2.e. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql .) ( t2 ) . u Example 2 .2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. According to Example A3. 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). Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. The time of the next renewal after t is the time of the next jump of the second type.2. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. (ii) First check the asserted identity for the density: since T.e.q2. hence e(t) is phasetype with representation (vt. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. i. Al. we first compute the stationary distribution of Q. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2.2) v(T + ta) = 0.4 Consider a nonterminating renewal process with two phases. (2.1. T1 and eTx commute.6. the unique positive solution of ve = 1.
5 Let B be Erlang(2).(biaz + aza.52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].226 CHAPTER VIII.4 yields the renewal density as u(t) = 5152 e.a27r1) (t1 . .t2) 1 + eat (a17r2 . Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 . MATRIXANALYTIC METHODS e. t1B 0 Example 2 .a27rl) (tl .4 yields the renewal density as u(t) = 2 (1 . The present treatment is somewhat more probabilistic. A = 25.`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . Then Q= 0 55 )+(1o)=( j ad ). and Example 2.e2bt) 13 Example 2 . and Example 2.t2) .52a1. )t (51 .tl) 7r2t2 + eat (a17r2 . Hence 7r = (1/2 1/2).6 Let B be hyperexponential.
the Markov processes representing ladder steps can be pieced together to one {my}.) = F(ST(o) E •. We asssume that B is phasetype with representation (a. The essence is contained in Fig. Thus the total rate is tip + tia+. T(0) < oo) the ladder height distribution and M = supt>o St. 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. Considering the first. marked by thin and thick lines on the figure. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). Since the results is so basic. and rewriting in matrix form yields the phase generator of {my} as T + ta+. 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. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2.f3aT1. and if there is a subsequent ladder step starting in j whic occurs w. r(u) the time of ruin with initial reserve u.3. Then each claim (jump) corresponds to one (finite) sample path of the Markov process.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. i. however.p. 3.e. which occurs at rate ti. Corollary 3. Next. use the phasetype representation of Bo. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_.3. Within ladder steps. and M is zeromodified phasetype with representation (a+. add a more selfcontained explanation of why of the phasetype structure is preserved. Corollary 2. we shall.1 Assume that the claim size distribution B is phasetype with representation (a. T). Then: (a) G+ is defective phasetype with representation (a+. The stars represent the ladder points ST+(k). T) where a+ is given by a+ = .2. B the claim size distribution. a+j. G+(. T + to+). Here we have taken the terminating Markov process underlying B with two states. cf. T). with 0 denoting the Poisson intensity. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. .1 on the next page. (b) V. 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. {St} the claim surplus process.i. For (b).
3.t t d kkt S. 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ). MATRIXANALYTIC METHODS t . 0 Example 3.M {mx} ST+(2)  S .2 Assume that ..1 This derivation is a complete proof except for the identification of a+ with ..1 . T = (3 . see Corollary 2.Q = 3 and b(x) = .. 3e3x + .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 .228 CHAPTER VIII. Figure 3.QaT1. This is in fact a simple consequence of the form of the excess distribution B0.
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 .1 In the zerodelayed case. cf. We shall derive phasetype representations of the ruin probabilities V) (u). see Stanford & Stroinski [351] . with A denoting the interarrival distribution and B the service time distribution.6).4. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). THE RENEWAL MODEL This is the same matrix as is Example 1. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration).6. T). Fig. For the compound Poisson model.2 are taken from Gerber [157].1 can be found in Neuts [269] (in the setting of M/G/1 queues. this was obtained in Section 3. 4 The renewal model We consider the renewal model in the notation of Chapter V.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+. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. In the next sections. That is.and Markovmodulated models.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. For further more or less explicit computations of ruin probabilities. his derivation of +'(u) is different.T). 3.4. The result carries over to B being matrixexponential.1): Proposition 4.j). and the argument for the renewal case starts in just the same way (cf. see Shin [340]. For an attempt. the discussion around Fig. we encounter similar expressions for the ruin probabilities in the renewal. where a+ is the (defective) . 3. The parameters of Example 3. T) for some vector a+ = (a+. the duality result given in Corollary 11. (a) G+ is of phasetype with representation (a+. if we define {mz} just as for the Poisson case (cf. see Section 6. We assume p = PB/µA < 1 and that B is phasetype with representation (a.
3.2 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).1. Since the conditional distribution of my given T1 = y is ae4y. with intensity matrix Q given by Q = T + to+. obviously mo = m. Also.5.1) Proof We condition upon T1 = y and define {m. G(') = pBo. The key difference from the Poisson case is that it is more difficult to evaluate a+. which for numerical purposes can be solved by iteration. Hence by Theorem 11.T).230 distribution of mo. Proof Obviously. In fact. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). 4. Then . cf. where B0 is the stationary excess life distribution corresponding to B.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a.1). it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. But by Corollary 2. Fig.*'} is Markov with the same transition intensities as {mx}. but with initial distribution a rather than a+. CHAPTER VIII. (4. Nevertheless. the calculation of the first ladder height is simple in the stationary case: Proposition 4. where a(8) = aT1/PA. B0 is phasetype with representation (aT1/µa.T)• Proposition 4.3 a+ satisfies a+ = V(a+).T). the Palm distribution of the claim size is just B. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}.6. (c) {mx } is a (terminating) Markov process on E. Then {m.*} from {St+y . We have now almost collected all pieces of the main result of this section: Theorem 4 .Sy} in the same way as {mx} is defined from {St}. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+).
Hence ^p(. a+2) = ^p (a+l)) . The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3.1(b). and that this is given by Proposition 4.. i.1/pA.2 ) follows from Proposition 4.. .. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . (4.1 .2. thus .. a+ can be computed by iteration of (4. a+l ) = cp (a+°)) . i y ^ T1= y `•r Figure 4.0.3) Proof The first expression in (4.•.3) (defined on the domain of subprobability vectors . I {mx} . only with initial distribution a(*) for mo.4. Furthermore .2) where a+ satisfies (4. (4. It remains to prove convergence of the iteration scheme (4.^(8)(u) = a ( 8)e(T+ta +) xe.0) is an increasing function of /3. THE RENEWAL MODEL 231 .3).1). the maximum claim surplus for the stationary case has a similar representation as in Proposition 4. by a+ = lim a +n) where a+°) .1) and a(8) _ aT. 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+. .M.^(u) = a+e ( T+ta+)xe. In particular .e.
232 CHAPTER VIII. Thus by (4. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1..4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. 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).ST. and hence we may assume that h has been normalized such that ahA[s] = 1. Then F[s] = a(sI .4.1. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. the corresponding right eigenvector may be taken as (sI .g. Assume the assertion shown for n . let F be the distribution of U1 . Fn ). (4. Theorem 4.5 Let s be some complex number with k(s) > 0. 7+ ]. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). we use an argument similar to the proof of Proposition VI. s ¢ sp(T).f. To prove the converse inequality. F[s] being interpreted in the sense of the analytical continuation of the m. so to complete the proof it suffices to show that &+ < a+) for all n. (4. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =.T)It.5) yields h = (sI .} can contain at most n . Similarly. Proof Suppose first Qh = sh. 0 0 We next give an alternative algorithm. Obviously. . 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . Thus .) = P(mTl = i.T)1t. this implies that ahA[s] # 0. To this end.1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ).4) whenever EeR(S)U < oo. with B[s]. the normalization is equivalent to F(s) = 1.5) Since s $ sp(T). a+ ) exists . For n = 0. In that case. both quantities are just 0 .4).P[s] = A[s]B[s]. However.T)'t • A[s] (4. limn4oo a ) < a+. and let &+".2. Then each subexcursion of {St+Tl .T1. Then (4. n) &+n) T a+.
we get at a(Q . (4. we have IG_ [s] I < 1 .6. Hence with h = (sI T).. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1... . Q = CD1 where C is the matrix with columns hl.. In older literature .. 0). . . letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 ..6) i=1 i=1 Proof Appealing to Theorem 4. 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). . hd. pdhd...5(c) means that a+(sI T)1t = 1.lt we get Qh = (T + to+)h = T(sI .. yd satisfying R(ryi) > 0.6 Suppose u < 0.. and the solution is .. This gives d roots 'y. Pd in the domain ER(s) > 0 .' that the equation F(s) = 1 has d distinct roots p1. Pd with corresponding eigenvectors hl. Q has diagonal form d d Q = dpivi®hi = dpihivi. in turn. t(ry) > 0.1 has the d distinct eigenvalues .T)It. Given T has been computed.. hd. Further.p1i .type with representation (a+. T) with a+ = a(QT)/at.4. Then G+ is phase. the matrix Q in Theorem 2..T) = 1 ata+ = a+. Corollary 4.T)lt + t = s(sI . This immediately implies that Q has the form CD1 and the last assertion on the diagonal form .. THE RENEWAL MODEL 233 Suppose next F(s) = 1. Since R(s) > 0 and G _ is concentrated on (oo. and hence by the WienerHopf factorization identity (A. and define hi = (piI . Notes and references Results like those of the present section have a long history. .9) we have G+[s] = 1 which according to Theorem 1... The roots are counted and located by Rouche' s theorem (a classical result from complex analysis giving a criterion for two complex functions to have the same number of zeros within the unit circle ). As in Corollary 4.T)lt = sh.5. Let d denote the number of phases. .. . 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. D that with columns p1 hl.
F. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. an alternative approach (the matrixgeometric method ) has been developed largely by M. starting around in 1975. [270] and Latouche & Ramaswami [241].4. The number of elements of El=> is denoted by q. where R is an unknown matrix. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). That is . and the distribution of an arrival claim is B. the background Markov process with p states is {Jt}.exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. It turns out that subject to the phase. see Neuts [269]. see Dickson & Hipp [118]. The solutions are based upon iterations schemes like in Theorem 4. is phasetype.type assumptions are basic. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. involving .type assumption . e.. [119]. whereas the approach was introduced in queueing theory by Smith [350]. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. In risk theory. the ruin probability can be found in matrixexponential form just as for the renewal model.contained derivation). Here phase. the intensity matrix is A and the stationary row vector is ir . The matrix. The exposition here is based upon [18]. We assume that each B. E(t)).. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. The distribution of W comes out from the approach but in a rather complicated form .234 then in transform terms CHAPTER VIII. The arrival rate in background state i is a. a pioneering paper in this direction is Tacklind [373]. and appears already in some early work by Wallace [377]. 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. Neuts and his students.. For further explicit computations of ruin probabilities in the phasetype renewal case .g. In queueing theory. the fixpoint problems look like R=Ao+RAI+R2A2+ .) d (see. T('). For surveys . Asmussen & O'Cinneide [ 41] for a short self. with representation say (a(' ). Numerical examples appear in Asmussen & Rolski [43].
1. Diagonalization Consider a process {(It. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. the phase space E(°) for B.2. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It.6. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. 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). 5.5. However. The key unknown is the matrix K. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. and the one E(•) for B. O. The version of the process obtained by imposing reflection on the V component is denoted a Markovian fluid and is of considerable interest in telecommunications engineering as model for an ATM (Asynchronuous Transfer Mode) switch. 5. has states o. say with slope r(i) on intervals where It = i. states . The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig.Vt)} obtained by time reversing the I component. 5a Calculations via fluid models.1 In Fig. Section 5b then gives a representation along the lines of Theorem 4.1. •. p = ql = Q2 = 2. for which the relevant fixpoint problem and iteration scheme has already been studied in VI.4. The two environmental states are denoted o.
j = 1. in the fluid model Eel'.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. i E E. 4. t. whereas Ee8s' = oo for all t and all s > so where so < oo. a E E(i) } . Bi[s] = a(i)(T(i) + sI)it('). A claim in state i can then be represented by an E()valued Markov process as on Fig. a) : i E E. r(i) _ 1. F is the disjoint union of E and the Eli). First. 5. < oo for all s. a) of {It}. o. corresponding to the partitioning + Epp). The fluid model on Fig . a) = 1. If s is such a number. 4. consider the vector a satisfying (A + (13i(Bi[ s] . Second. The intensity matrix for { It} is (taking p = 3 for simplicity) I A .(Ni)diag r(i. •.31a(l) (/3i)diag . 4}. 5.1))diag ) a = sa and the eigenvector b = .1) if and only if s is an eigenvalue of E. of E into components indexed by E.1 A complex number s satisfies 'A+ (f3i(Bi[s] .A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 .236 CHAPTER VIII. V. Eli) + Proposition 5. we have more martingales at our disposal. Let E denote the matrix . Thus F = {o. 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.92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. MATRIXANALYTIC METHODS 4. In the general formulation .Vt)} is then obtained by changing the vertical jumps to segments with slope 1. This implies that in the fluid context. 2.1(a). resp. Recall that in the phasetype case. F = E U { (i.1))diag + sII = 0 (5.
E22)1 E21a E21a .T('))1t(i) . and let d = (sI . d = (sI E22)1E21a = E ai(sI . it follows that if Qla(1) 0 0 .(sI . Then (up to a constant) c = a. c = a.sI) (sI . t(1) 0 0 then also 0 t(2) 0 .sI. with Eii replaced by Eii . resp .1).E22)1 E21) a = 0.A . E(1) + + E(P).32a(2) (/3i)diag . Then E21c+E22d = E21a . assume that a is chosen as asserted which means (Ell . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .E22)1 E21a.sI 0 0 t(3) 0 0 = 0.A . 0 .sI+ ((3ia(i)(T(i) . where c.5. Noting that E11c + E12d = se by definition. iEE (a> of 0* 1 AI.sI)1t)) iag I = 0 which is the same as (5.sI + E12 (sI .Nla(1) 0 0 T 1. MARKOVMODULATED INPUT 237 indexed by E. it follows that Ell E12 ( E 21 E22) (d) = s 1 d I . For the assertions on the eigenvectors.sI 0 0 0 T(2) .E22 .E12E22 E21 I .sI ()3i)diag .E21a + sd = sd.sI 0 0 0 T(3) .
e89uc(e)) (d(1) . j. a) = (j. MATRIXANALYTIC METHODS Theorem 5.. j. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. Example 5 .. j. a) and noting that i1 (u) = >I j. v > 0.j.sv)b(v) = 0. sq with $2s. v = 1. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . ..Q. define w(u.5. a)). d("))1 e. Proof Writing Or'Alb( v) = svb( v) as (AI . q. v.O. Then . v.4 Assume that E has two states and that B1. Thus 0(u) = esu/d = pe7 ° as u should be. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . Here E has one state only. the result u follows.j)c v .v) = (j. I' i( V P2 (w (u) < oo. a )d(a + e8 °vpi (u .. ./' u = e' (esiuc ( 1) .. c j. a).v) = v) I. a) = Pi (Vw(u.238 CHAPTER VIII. it follows by Proposition II. Then we get V)i (u) as sum of two exponential terms where the rates s1.a Solving for the pi(u.upi(u. . Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u.( u.v..v}. Example 5 . . v) yields C{V) = e8 .. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s. v..2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si.v) = Optional stopping at time w (u.4 that {e"1b(v) is a martingale . w(u.v) = j). j) pi( u .a)d^ ).3 Consider the Poisson model with exponential claims with rate 5.. . a)). To determine 0 (u).. . u) Iw(u.v)=inf{t >0:Vtu orVt=. pi(u.. j. Iw(u. j. B2 are both exponential with rates 51 i b2.. For u. w(u)=inf{t >O:Vtu}.pi(u..v) = = p i( u ..
3j eye. 9('). Proof We must show that G+ (i.2) the l. the Pidistribution of M is phasetype with representation (E(1) + + E(P). In terms of K. 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. U) where t(j) + t(j)O(j j = k uja. oo)) j)ye. (') a T( However . i. 8^')IT(j)) where e 3^') =. 0 Theorem 5 . j.h. according to VI.s. •) is phasetype with representation (E(i).x) 00 f ° (') (j) eT (yy)edx .k. j. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.6 For i E E. is 0 /3 f R(i .5.2.5 G+(i. j. dx)Bj(y .b (u) = Pi(M > u) = 9(i)euue.y = to B k7 j # k In particular.3) .33(e = 0 a(j))(K ®T ( j))(ej (9 I).( 2. (y.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. (5.xxej • a 00 oo el .
0 1)'. bn1 bn).240 CHAPTER VIII. (6.. Associated with each ladder step is a phase process. t = (0 0 . if b* [0] = b1 +b20+b302 +.. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability.g. Bk7 . oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. Starting from Jo = i. say. and a new ladder step of type k must start in phase y. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . intensity matrix U.. the current ladder step of type j must terminate. For j = k. and lifelength M.. which occurs at rate t^^7. Furthermore. the ratio between two polynomials (for the form of the density.. 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. This yields the asserted form of uja. we have sofar concentrated on a claim size distribution B of phasetype. An alternative characterization is that such a distribution is matrixexponential.y) to occur when j # k. a m. ..) which is rational. which occurs w. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17].k y. i. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . Piecing together these phase processes yields a terminating Markov process with state space EiEE E(').. equivalently. a) is obviously chosen according to e(`). we have the additional possibility of a phase change from a to ry within the ladder step..e. Numerical illustrations are given in Asmussen & Rolski [43]. that the density b(x) can be written as aeTxt for some row vector a. Then b*[0] is rational if and only b(x) is matrixexponential. However. which occurs at rate t(i).2) . +bn0i1 0n +a10n1 +.f.1 Let b(x) be an integrable function on [0. the initial value of (i. see Example 1. a) to (k. t) is the representation of the matrixexponential distribution/density): Proposition 6.e. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). T. and it just remains to check that U has the asserted form. some square matrix T and some column vector t (the triple (a.5).2.p. i. For a transition from (j.
but as follows from Proposition 6.1 is that it gives an explicit Laplace tranform inversion which may appear more appealing than the first attempt to invert b* [0] one would do.an_3 an _ 4 .T)1 is so. (6. S = f c/2 0 21ri ./(0 + bi). t). Writing b(x) = c(e( 2ni1 ) y/2 . personal communication).h.3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. see Asmussen & Bladt [29] (the representation (6. T.. shows that the distribution B with density b(x) = c(1 cos(21r x))ex. S.47x2 3 1 0 .1) as E 1 c. Example 6 .. since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.47r2 3 .e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. where c = 1 + 1/47r 2. .4) 0 0 1 c This representation is complex.. 0 0 . One of his elementary criteria. we can always obtain a real one (a. . a2 a1 Proof If b(x) = aeTxt. (6.2).2). . For a proof. cannot be phasetype.s. u Remark 6. bn of the denominator to be distinct and expand the r. 0 0 0 0 1 0 0 .1.. namely to asssume the roots 6l... t= 0 . s) is given by 27r i ...(6.3) was suggested by Colm O'Cinneide. then b*[0] = a(0I T)1t which is rational since each element of (01 . 1 ..3) 0 0 0 0 0 .1 0 . of (6.6. b(x) > 0 for x > 0.1 0 0 )3 = (111). MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . matrixexponentiality implies a rational transform. Thus. 0 1 an an1 an _2 . (6.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). . Namely. s = c/ 2 .2 A remarkable feature of Proposition 6. u giving b(x) = E 1 ciebiz/bY. . The converse follows from the last statement of the theorem. T= 0 0 1 .
242 CHAPTER VIII. and can use this to invert by the method of Proposition 6. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1.5) Thus. We recall (see Section 3. (6. that despite that the proof of (6. and present two algorithms for calculating '(u) in that setting.6) holds true also in the matrixexponential case. we take as starting point a representation of b* [0] as p( O)/q(9) where p. we use a representation (a.3. leading to matrix calculus in high dimensions when b is small. . For the first. t) a phasetype representation with a the initial vector. Then (cf. As for the role of matrixexponential distributions in ruin probability calculations. then: Proposition 6.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. and that the minimal number of phases in a phasetype representation increases to 0o as 5 . t) of b(x).6) The remarkable fact is. 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. MATRIXANALYTIC METHODS Example 6 . 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0.1)2 + 6).5 (6.4) the Laplace transform of the ruin probability is /g(e)PO 0*[e] _ /' eeu^G(u)dU = 0 9(/3a0p(9)ap (9)/q(9)) . (6. T. Consider the distribution with density = 15 ((2e2x . T the phase generator and t = Te. 0. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels).1 to get i (u) = f3esus. q are polynomials without common roots. recall that t = Te) that if B is phasetype and (a. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. Corollary 111.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. T. For the second algorithm.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons .
1 + b+ = b++ 1 . b+ = a +(BI .b* (6.T). we get (91. we get b+ = 0aT1(9I T).5 ).1BVA1.t.1UB(B + BVA1UB). 519) (A + UBV ).7) 9( cf. MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .T)1 (91.to+)1 = (BI . (6. the assertion is equivalent to a+(BI .T .T)1T 2 = and 1 = AB IT2 + 82T . Then in Laplace transform formulation . From the general matrix identity ([331] p. (6.1 + 82 (9I . .a+(9I .T)1T1t.T)1t ( l .T .to+)1T .to+)1T .T)1 so that b* b** b** a+(9I .'t. (91.6.B=land V=a+.1t = f3a (0I T)1T1t .T)1t)1a +(9I .1t = b* .T .T)1ta+(OI . U =.6b* .6).T)1t. Now. but we shall give an algebraic proof.1t du = .T)1 J0 00 b(x) dx = f aT1t. xb(x) dx = aT2t. with A = 91T.T)1 + 1 ib* (91.A . b+ = a+(9I . Presumably.1 = A1 .1 = ^(T1 + ( 91T)1). this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .T)1 + (6I . since (91T)1T .
1.1.5 is similar to arguments used in [29] for formulas in renewal theory.3 is taken).7). premium rate p(r) at level r of the reserve {Rt} and claim size distribution B which we assume to be of phasetype with representation (E. See Fig.b*). 3. VII.la. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. see the Notes to VII.244 CHAPTER VIII. a key early paper is Cox [90] (from where the distribution in Example 6. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). /3aT1(0I . 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4.s. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. a.1.T)1)t = 8 (1 . 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.1.1t = /3a (9I .T)1T. In Corollary VII. which is selfexplanatory given Fig. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. . 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0.T)1T2t . Lipsky [247] and Asmussen & O'Cinneide [41]. of (6. The proof of Proposition 6.h. cf. see Asmussen & Bladt [29].82b*. Much of the flavor of this classical approach and many examples are in Cohen [88].8. T).1) is the same as the r.8 a(T1 + (01. 0 Notes and references As noted in the references to section 4. MATRIXANALYTIC METHODS . but the argument of [286] does not apply in any reasonable generality). 7. For expositions on the general theory of matrixexponential distributions.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. From this it is straightforward to check that b**/(b+ . the ruin probability(u) was found in explicit form for the case of B being exponential.
1) where A(t) = v(u)P(0. Ai(t) = P(mt = i).I] I 80 { tq f Q(v) dvl t1 1 . O<. we obtain V)(u) = P(m„ E E) = v(u)P(0.tl < t2 < u.1 The difference from the case p(r) = p is that {m2}. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}.t2) be the matrix with ijth element P (mt2 =j I mtl = i). Note that in general >iEE Vi (U) < 1. in contrast to Section 3.1z I. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). t2) = exp where Q(t) = ds [P(t.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . t + s) . By general results on timeinhomogeneous Markov processes. i. though still Markov. Let P(tl. P(tl. Proof The first statement is clear by definition.7.u)e = A(u)e (7. Given the v(t) have been computed.e. 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.t)). In fact. 0 < t < u. is no longer timehomogeneous. Figure 7. the definition of {m8} depends on the initial reserve u = Ro. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . t) is the vector of state probabilities for mt. Also.
Given A'. vi. 0 < t < u.3dt. jEE . (7. the probability that level u is downcrossed for the first time in phase i is ai. MATRIXANALYTIC METHODS However. Thus. {mx} has jumps of two types.Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. from a computational point of view the remaining problem is to evaluate the v(t).(u) p ( u) = . or it stops between level u + p(u)dt and u.2 For i E E. those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . dt].Sj i)p(u)dt • tji = Sji + p(u)tji dt. the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 . 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. 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.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0.t) for the second.t). 0 Thus. we get vi(u) = aidt + (1 . given A. Given A. the probability of which is 1 . the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt).t and being followed by a downcrossing.t)). two things can happen: either the current jump continues from u + p(u)dt to u.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). Given this occurs.246 CHAPTER VIII. In the first case. Proposition 7. whereas in the second case the probability is p(u)dt • tjvi(u). A'(t) = A(t)Q(t) = A(t)(T + tv(u . Hence Q(t) _ T + tv(u .
F" etc. refer to the modified process. This yields v. denotes the time of downcrossing level u .2. vi (U) = lim v= (u).) P"(AnBv) = P(AnB. (v) is given by the r.)P"(AnB.s. Then P(B. of (7.5).^ 0.h. we have p(r) = p = vi (u) 0aTe. and similarly P"(Bv) . When solving the differential equation in Proposition 7. From Section 3.. P u which implies that v. To deal with this. . starting from v"(v) = . (u) for any values of u and v such that u < v.4) backwards for {va (t)}v>t>o. say. Let p" (t). say. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. we face the difficulty that no boundary conditions is immediately available.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. consider a modification of the original process {Rt} by linearizing the process with some rate p.3 For any fixed u > 0. Now since both P(A n Bv) 3 0 and P"(A n Bv) . we can first for a given v solve (7.00 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. RESERVEDEPENDENT PREMIUMS 247 Subtracting v. Rt . V . (u) on both side and dividing by dt yields the asserted differential u equation.. Thus. then P(A n Bv) _ P"(A n BV').7.i7rT1/p. supRt>v l t<7 I where o..) + 0 as v + oo. Since the processes Rt and Rt coincide under level B.. <oo.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. after a certain level v.
) are so. 2/n. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v.10 that in addition to the O'(•). v = u.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. p2.zp1(u)/(1 .. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII.1a. (u)}. Therefore u pl(vvueTa t 1. cf.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. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII.RQ (defined for or < oo only) is defective phasetype with representation (v(u). p(r) P. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.1. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. let v(u) = a+2ieiT +ta+>)(uv). where. say.V" M 0 . 1/n. To evaluate p1(u).e. The precision depends on the particular quadrature rule being employed.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. (7. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. which is available since the z/i'(. Corollary 3. > 0} and the last term the contribu tion from {R.248 CHAPTER VIII. However.x) dx f v(u)eT xt dx .v v(u)eTat 1 1 . assuming u > v for the moment. The f iin in (7.. such that Rt coincide with RI under level v and with Rt above level v. We recall from Propositon VII. 3u etc.7) f o (the integral is the contribution from {R.7) equals 01 (v .z51(v)). < 0}). Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. The trapezoidal rule used in [288] gives a precision of 0(n 3). as well p1(u). numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. 2u.. r<v r > v. the evaluation of Vi(u) requires q(u) = 1 . while the fourthorder RungeKutta method implemented in [30] gives 0(n5).. T).9. where v = inf It > 0 : Rt < v}.1.. for u > v the distribution of v . 0 < u < v. the probability of ruin between a and the next upcrossing of v. i.
v(u ) eTV e J v(u)eTxtz/)l (v . (7.7.and A2 = 3 .x) dx 1 ^(v) ( 1 .4 Let {Rt } be as in Example 3. all quantities involved in the computation of b(u) have been found in matrix form.8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .2.21 = ? yields 0(u) = 1.. 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 . B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.2.u . I.v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.v(u)eTVe . Example 7.e6u 35 .to+))11 {e{T®(Ttoy+ ))}„ . From Example 3. Since µB = 5/21. RESERVEDEPENDENT PREMIUMS 1 .v(u)eTve).be the eigenvalues of T + to( 2 ).24e.jl (t ®e) Thus.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .x) dx} V 1 1(v) f V v(u) eTxt.24ev .2V"2.e.(7 The arrival rate is (i = 3.8) The integral in (7.v) + (2^ + 3v2 ea'(u " . p2 < 3.^1(v) 1 .e6v Let Al = 3 + 2V'2. so we consider the nontrivial case example p2 = 4 and p1 = 1.1 from which we see that pl (u) = 1 + 1 249  1 .01 (v .
24es" . ) e sv + ( 2v/2./2 ea1(u") . 21 3 In particular.v)esv + 7 4_ 2.1 Thus../2) ea 1(u .21(35e6v . pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v . Notes and references [30].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 .1.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.250 CHAPTER VIII. 192esv + 8 P1 .2 35e6v .+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.24e5v . The analysis and the example are from Asmussen & Bladt .24es" . MATRIXANALYTIC METHODS From (7. and one gets 12e5" .1)' ?.24e5v .
g.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. For example. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. we require that B is concentrated on (0. B(x) = ex0 with 0<0<1.4. and instead we shall work within the class S of subexponential distributions . 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.N(µ. for all t > 0. A rough distinction between light and heavy tails is that the m. see I. 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. x 2iror2 (c) the Weibull distribution with decreasing failure rate . oo ) and say then that B is subexponential (B E S) if 251 .2b. a2)) with density 1 e(logyFh) 2/2az . x 4 oo.f. (b) the lognormal distribution (the distribution of eu where U .B(x).46 and at numerous later occasions require a light tail. L(tx)/L(x) 4 1. For further examples. the exponential change of measure techniques discussed in II. III. For the definition .
That is. x 3 00.2 If B E S. X2) > x) ^' 2B(x). To capture the intuition behind this definition.B(x)2 . oo). then P(X1>xI X1+X2>x)* 2.v.'s. proving (a). In contrast. X2 but none of them exceeds x. HEAVY TAILS B*2\ 2. Since B is concentrated on (0.2. one can check that x x where U is uniform on (0. Proof By the inclusionexclusion formula. X2) > x}. note first the following fact: Proposition 1.F(X1 > x. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. . 1). given X1 + X2 > x. That is. The proof shows that the condition for B E S is that the probability of the set {X1 + X2 > x} is asymptotically the same as the probability of its subset {max(Xi. X1 is w. 1/2 it has the distribution of X1I X1 > x.1(b) is oo. if X1 + X2 is large . oo). the distribution of independent r.2B(x).v. 1/2 'typical' (with distribution B) and w. X2 with distribution B. B(x) Here B*2 is the convolution square. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex.'s X1. B(x) ax. X2) > x} C {X1 + X2 > x}. (b) liminf BB(() ) > 2. (1.p.3 Consider the standard exponential distribution.252 CHAPTER IX. P(max(Xi. X2) > u x)/B(x) = 2.1 Let B be any distribution on (0. and thus the lim inf in (b) is at least lim inf P(max(Xi.p. Then: (a) P(max(Xi. then (with high probau bility) so are both of X1.1) then means P(X1 +X2 > x) 2P(Xi > x). the behaviour in the lighttailed case is illustrated in the following example: Example 1. As contrast to Proposition 1. that is. We later show: Proposition 1. Thus the liminf in Proposition 1. Thus . the r. X2 > x) = 2B(x) . P(Xi <yI Xi+X2>x) 1B(y). we have {max(Xi. X2) > x) is P(X1 > x) + P(X2 > x) .v. In terms of r.
We now turn to the mathematical theory of subexponential distributions. and combining with Proposition u 1.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10. Hence lim sup a+oo B*2(x) 2B((1 .B E S.y)/B(x) > 1. If lim sup B(x . [In terms of r. If X1 + X2 > x. Finally lim inf B(x . we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . yo] as X + 00. x].4 Any B with a regularly varying tail is subexponential. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.'s: if X . y] and (y.v. or they both exceed Sx.] Proof Consider first a fixed y.yo].2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.B*(n ) B(dz) (1.6)x)/((1 .xIX > x converges in distribution tooo. Let 0 < 5 < 1/2. we get limsupB*2(x)/B(x) < 2.S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . then the overshoot X .1(b) we get B*2(x)/B(x) * 2. a contradiction.1.z B(x) .S)x.y)/B(x) > 1 since y > 0. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .B*n(x . 1 < B(x ) B( x) Y) < B( 0). then either one of the Xi exceeds (1 . Proposition 1. then B(B(x)y) * 1 uniformly in y E [0.5 If B E S.B(y)) . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) .B(y) = 2. B( 0 .
x oo. This implies B(x) > c2e5x for all x. so assume the proposition has been shown for n.5 that B(n) > e6B(n . Proposition 1. Given e > 0. Proof For 0 < 5 < e. choose y such that IB*n(x)/B(x) . HEAVY TAILS Corollary 1.1) for all large n so that B(n) > cle6n for all n.254 CHAPTER IX. The first integral is y B(x .z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. B*(n+1) (x I xy + Jxx y) W.B*2 (x) B(x) (x . Then by (1. The case n = 2 is just the definition.5 and the induction hypothesis. Proof We use induction. then for any n B*n(x)/B(x) * n. then e"R(x) * oo. its intuitive content is the same as discussed in the case n = 2 above.z) B(dz). we have by Proposition 1.2). b[c] = oo for all e > 0. O The following result is extremely important and is often taken as definition of the class S.. 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.B(x . B(x) \Jo _ B(x .z) B(dz) _y B(x) 111 Lx B .2.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) .5 and dominated convergence. and this immediately yields the desired conclusions.6 If B E 8.7 If B E S.z) B(x) Here the second integral can be bounded by B*n(y) B(x) .nI < e for x > y.z ) (x ) = 1 + (^ B(x .
Then Al * A2 (x) .1.i).z) B(dz) x .(al + a2)B(x). Combining these estimates and letting a 4. X2 be independent r.y)Ai(dy) v) f o . Then Al * A2(x) = P(X1 + X2 > x).3) Using the necessity part in the case Al = A2 = B yields f xv B(x .y)B(dy) = B(x)ov (1)• v (1. Since P(X1+X2 > x.0 completes the proof.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x . For any fixed v.2). 0 Proposition 1.ala2B(x)2 which can be neglected. choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). Then by (1. Xi <= v Ai (x .X1 > xv. an = supx>o B*n(x)/B(x). an+1 fX B*n( *n(x .'s such that Xi has distribution Ai.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .v. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.4) .5 easily yields P(X1 + X2 > x. x>T o B(x) The truth of this for all n together with al = 1 implies an < K(1 + 5)2n where K = (1 + A)/e. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x . A2 be distributions on (0.z) B(dz) < 1 + A + an(1 + d) .z) B(x) < 1 + A + an sup f x B(x . Proof Define 5 > 0 by (1+5)2 = 1+e.X2 > xv) < A1(xv)A2(x v) .y)Ai(dy) = (x)o(1) (1.B(x .z) B(x . 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) . e > 0. Proposition 1. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al.8 If B E S. a2 with a1 + a2 > 0.y))/B(x). 0 Lemma 1. Proof Let X1.9 Let A1.
Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. u Corollary 1.aiB(v)) = B(x)o„(1). However.4). whereas the two first yield B(x)(Ai(v) . then A E S. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.v)B(v) + _'U Aq(x . a2 = 1.10 The class S is closed under tailequivalence. a1 = a2 = a yields A*2(x) . of (1.s. A2 = B so that a1 = 0.11 Let B E S and let A be any distribution with a ligther tail. it is easy to see that if L1. A(x) = o(B(x)). B2 E S. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown.2A(x).2. V (1. if q(x) aB(x) for some B E S and some constant a > 0. L2 are slowly varying.13 Let B have density b and failure rate A(x) such that .3) follows if CHAPTER LX. f " By a change of variables. Then B E S provided fo "O exA(x) b(x) dx < oo. That is.Bl (x) + B2 (x) when B1. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. u It is tempting to conjecture that S is closed under convolution.y)Ai(dy) = B(x)o„(1). That is. Proof Taking Al = A2 = A.12 Assume that Bi(x) = Li(x)lxa.v)Ai(v) . HEAVY TAILS 'VV B(x . then so is L = L1 + L2.256 Now (1. u Corollary 1.2aB(x) . with a > 0 and L1.Ai(x . i = 1.9). In the regularly varying case. Then A * B E S and A * B(x) . L2 slowly varying.5) Here approximately the last term is B(x)o„(1) by ( 1. it should hold that B1 * B2 E S and B1 * B2 (x) . Hence Corollary 1.h.B(x) Proof Take Al = A.5) becomes x B(x .(x) is decreasing for x > x0 with limit 0 at oo. the l. . B1 * B2 (x) .y)B(dy).
A(y)\(y) dy + fox/ 2 eA(x ). Further. Example 1. f ' L(y) dy .A(y)a(y ) = ev'(y) b(y).U) /or) v 2x This yields easily that ex.1. we can use the same domination for the second integral but now the integrand has limit 0 . replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Goldie & Teugels [66]): Proposition 1. x . subexponentiality has alrady been proved in Corollary 1..A(x .y) dy.1 has limit 1 + 0. Thus by dominated convergence . To illustrate how Proposition 1. 0 A(x) . Thus A(x) is everywhere decreasing. L(x) y° (a . By (1. an integrable function by assumption. the u lognormal distribution is subexponential. a(x) = ax01. we first quote Karamata's theorem (Bingham. Define A(x) = fo . Thus.12. and exa(x)b(x) = (3x01e(10)x9 is integrable.y) < A (y) for y < x/2. Thus.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . In the regularly varying case.2). The middle bound shows that it converges to b(y) for any fixed y since \ (x . Example 1.(x . B*2(x) .y) * 0. Then b(x) = Ox0lexp. elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. the DFR Weibull distriu bution is subexponential.A(xy)A ( y).`(x)b(x) is integrable. Then B(x) = eA(x).(y) dy.13 works in this setting. Thus B*2(x )/ B(x) .14 Consider the DFR Weibull case B(x) = ex0 with 0 <.y) < yA(x .15 In the lognormal distribution.1)xcl1 .. proving B E S. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise.16 For L(x) slowly varying and a > 1.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x .3 < 1.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ). the first integral has limit 1 . Since ) (x . Jo For y < x/2.
f O B(y) dy .13 may present a problem in some cases so that the direct proof in Proposition 1.)/Y(x) > y) (1 + y/(a . then B(x) .xjX > x. let X W = X . (c) Under the assumptions of either ( a) or (b).4 is necessary in full generality.1/A(x) and P(X ixil'Y (x) > y) * e'. yo] .E(X .6) EX(x) .1))a .a/x.L(x)/x" and )t(x) . . the monotonicity condition in Proposition 1. then 7(x) x/(a .258 From this we get CHAPTER IX. Then: Proposition 1.ea b(x) is integrable.18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 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 conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.y(x)B(x). 'y(x) = EXix>.1)]) xa L(x) (x[1 + y/(a .17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.1)X(x)/x > y) = P(X > x[1 + y/(a .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. More precisely. we get (1.1)] I X > x) L(x[1 + y/(a . Then 7(x) . (1 + y/(a . Proof ( a): Using Karamata's theorem.1) and P(X (.1)])a 1 1 .x)+ _ 1 °° P(X > x) P(X>x )J L PX >y)dy 1 x L(y)/ydy L(x)/((a1)x'1) x )l ° J °° ( ()l a x a1 Further P ((a . However. HEAVY TAILS Proposition 1. Thus exa(x)b(x) .
Notes and references A good general reference for subexponential distribution is Embrechts.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P . 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. Lemma 2.nn.t be the claim surplus at time t and M = sups>0 St. Bo(x) = f0 B(y) dy / µB..A(x) I X > x) = exp {A(x) . We get p(yl+. with common distribution G E S and let K be an independent integervalued r.n0 1•P(K= n)•n = EK. be i.2 Let Y1. d. Examples 1. u a oo.. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo).7) is referred to as 1/A(x) being selfneglecting. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) .8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/.14..2. The remaining statement (1.(x).v. then Vi(u) P Bo(u). 1. 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. nG(u).and lognormal distributions . . . Kliippelberg & Mikosch [134].1/.1 If Bo E S. cf. St > u}. i.+YK> u) = ^•P(K = n)G* n(u ) . Y2..f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1.15. It is trivially verified to hold for the Weibull. P The proof is based upon the following lemma (stated slightly more generally than needed at present). Theorem 2 . Recall that B0 denotes the stationary excess distribution. Let St = Ei ` Ui . r(u) = inf it > 0.EK G(u). with EzK < oo for some z > 1. 0 G(u) L G(u) . Then P(Y1 + • • • + YK > u) .
260 CHAPTER IX.x400 PBB(x) PB Leta+oo. For some numerical studies. Since EK = p/(1.1. x 4 00. Proof of Theorem 2. The approximation in Theorem 2. lognormal . P(K = k) = (1.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. Weibull) one has Bo(x ( B(x) . u The condition Bo E S is for all practical purposes equivalent to B E S. Note that in these examples . r(1/Q) xlQexp B(x) = ex' From this .13). In general: Proposition 2. then Bo(x)/B(x) + 00. in our three main examples (regular variation ..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 ) . (2.2. the result follows immediately from Lemma 2. u x+a Notes and references Theorem 2. Bo is more heavytailed than B . Bo E S is immediate in the regularly varying case. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant.18. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) .1 is notoriously not very accurate. The problem is a very slow rate of convergence as u ^ oo. Borovkov [73] and Pakes [280].x^ ) B(x) _ f or ( lox .p)p'.1) In particular . a].µB(01 . The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. Bo ¢ S. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. mathematically one must note that there exist (quite intricate) examples where B E S. The PollaczeckKhinchine formula states that (in the setup of Lemma 2. as well as examples where B ¢ S. . see Abate. See also Embrechts & Veraverbeeke [136].µ J ) .3 If B E S.?(xµ 8 (x).p) and EzK < oo whenever pz < 1.1 is essentially due to von Bahr [56]. Bo E S.. However. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1.
1. p = iB /µA < 1. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. [279].. in [219] p.y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. there is a representation of M similar to the PollaczeckKhinchine formula..1 when u is small or moderately large. (3.. . also a second order term is introduced but unfortunately it does not present a great improvement. t9(u) = inf {n : Snd> > u} . Let U= be the ith claim . T1 the ith interarrival time and Xi = U. We assume positive safety loading. To Bo(u) u + 00. + Xn. E. i=1 . The main result is: Theorem 3 . 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. i.9+ < oo) = P(S. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x . Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA). Then K M=EY. G+ (A) = P(Sq+ E A..e. Kalashnikov [219] and Asmussen & Binswanger [27].3.. Based upon ideas of Hogan [200]. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. Snd) = Xl +.] The proof is based upon the observation that also in the renewal setting. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2.Ti.g.. M = sup s$ . Somewhat related work is in Omey & Willekens [278]. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2.y)/B (x) > 1 uniformly on compact y internals.1) this end .1 gives 1010.} Then ik(u) = F ( M > u) = P(i9 (u) < oo).. This shows that even the approximation is asymptotically correct in the tail. T+ < oo) where r+ = T1 + • • • + T.+ E A. Define further 0 = IIG+II = P(r9+ < oo). In [1]. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. {n= 0. THE RENEWAL MODEL 261 Choudhury & Whitt [1].
FI(x) /IPG_I.PBBo(x). Lemma 3 .3 G+ (x) . Let F denote the distribution of the Xi and F1 the integrated tail..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_.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.. B(x) _ J O° B(B(x)y) A(dy) f 1 . A(dy) = 1.2 F(x) .262 CHAPTER IX.1 IPG_ I / F(x . Write G+( x) = G+ ( x. u a 00.Ti).d.FI(u).9)9'' and Y1.y+ given r+ < oo). 0 The lemma implies that (3. U_ (dy) is close to Lebesgue measure on (.1) holds for a general random walk satisfying the analogues of (a). As for the compound Poisson model. Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. HEAVY TAILS where K is geometric with parameter 9.g+ > x. and hence FI(x) . The heuristics is now that because of (b). oo) = F(S. x > 0. (3. with distribution G+/9 (the distribution of S.y) dy = 1 Pi (X) oo IPG_ I .i. d+ < oo).B(x). 0] normalized by IPG_ I so that we should have to G+(x) . FI (x) _ fz ° F(y) dy.3) and we will prove it in this form (in the next Section. (b) and does not rely on the structure Xi = Ui . cf. P(K = k) = (1 .1) is equivalent to P(M > u) " .2).. x * oo.N. x + oo.Y2. A. 0] to the integral is O(F(x)) = o(FI(x)). Proof By dominated convergence and (b). whereas for large y . the contribution from the interval (.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. are independent of K and i. Lemma 3 . Then 0 0 F( x .oo. G_(A) = P(S. Proof Let R+(A) = E E'+ ' I(S. 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.(.
2) yields 00 F F I (u) P(M > u) _ E(1 . In the lattice case.UG_ I x.3.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N. (3.1.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup . the proof is complete. THE RENEWAL MODEL 263 We now make this precise. choose N such that F(n . n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG.y) U. and that U_(n . > (1 . then by Blackwell 's renewal theorem U_ (n . and in the last that FI is asymptotically proportional to Bo E S. we can assume that the span is 1 and then the same conclusion holds since then U(n .0)0k k I(u) A. u Proof of Theorem 3. Hence using dominated convergence precisely as for the compound Poisson model.1.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.1 I . F(Y= > x) FI(x)/(OIp _ 1). We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x . By Lemma 3.1.9) 1 .(dy) fN FI ( x) + lim sup ZY00 N F(x . n] is just the probability of a renewal at n.3. Given e.O[s])(1 .2).1.F[s] = (1 .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] < (1 + e)/1µc_ I for n > N.1. If G_ is nonlattice.=1 BIp G_ I (1.G+[s]) . Similarly. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .9)IpG_ I Differentiating the WienerHopf factorization identity (A. n] + 1/I µG_ I.
Se(u)_1 < a) = o(Fj(u)).2. Sty(u) . and {Su.264 and letting s = 0 yields CHAPTER IX. ...IIG+II)µc_ = (1 .Sty(u)_I < a} we have w(u) < oo. S+q(u) . Notes and references Theorem 3. with roots in von Bahr [56] and Pakes [280]. see Asmussen & Kliippelberg [36]. must attain a maximum > 0 so that P(M > u.(u)+n . u).yiui_1. Therefore by Lemma 3. u)) > P(w(u) < oo)(i lp (0))• On the other hand. on the set {M > u.a.AB iP We conclude by a lemma needed in the next section: Lemma 3 . 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.u)) = o(P (M > u)) = o(FI(u)).a. Mn < u}. Note that substantially sharper statements than Lemma 3. we have P(M E (u .So(u)) are available.a.4 on the joint distribution of (S. HEAVY TAILS µF = (1 .1)6+[0] . S+9(u) . 10(0) But since P(M > u .(1 .a) N P(M > u).So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u .1 is due to Embrechts & Veraverbeke [136].a. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .0)ua_ . Proof Let w(u) = inf {n : Sid) E (u . allowing also for possible dependence between the arrival process and the claim sizes. 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times. P(M > u.SS(u)}n=o.. u)).4 For any a < oo. Then P(M E (u .l.
0o(u) etc. M = sup St..1.X1 is the generic cycle. assume pp.1) . Theorem 4. (corresponding to the filled circles on Fig. Thus the assumption .. and the distribution of {Sxk+t . See Fig..1 based upon a regenerative assumption.. Schmidli & Schmidt [47]. examples and counterexamples. Figure 4. = Sx.. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.. M* = max S.1) is assumed.. G(x) (4. < 0 and EoX < oo where X = X2 ..1 Note that no specific sample path structure of {St} (like in Fig. {SX1+t . such that {SXo+t  SXo}0<t< X 1Xo . 4.n n=0. see [47]. We return to this point in Example 4. Define S.4 below.. {Sn}n=o.1 = max k=0. We let F* denote the Podistribution of Si. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i.Sxi}0<t<x2Xl . For further approaches. +1.X2 < ..... and apply it to the Markovmodulated model of Chapter VI... We give here one of them. The idea is now to observe that in the zerodelayed case.d. t>0 S. M..1 except for the first one) is a random walk.Sxk}o<t<xk+1xk is the same for all k = 1.F*(X) = P0(Si > x) . ..... 2..1 where the filled circles symbolize a regeneration in the path.1. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.4. 4. 4.i. E0. 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 <.
Since clearly M(x) > Sl .. u p 00. Then '00 (u) = Fo(M > u) .2 Theorem 4. the assumption means that Mix) and Sl are not too far away. 4.* i o<t<xn+1x. Fo(Si > X).. HEAVY TAILS for some G such that both G E S and Go E S makes (3.1) and (4.1 Assume that (4. it suffices by (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.3) where Mnx) = sup o<t<xn +1 X. N N Xi=0 N Figure 4. See Fig. (4. (4.4) liminf u>oo F(M > u) . jF11 F* (U)..Sxn = sup Sxn+t . Proof Since M > M*.S.. (4. Sxn +t . The one we focus on is Fo (Mix) > x) .3) applicable so that F(M* > u) 141 F*(u).266 CHAPTER IX.3) hold.2.2) to show F(M* > u) > 1.
To this end. Theorem 4. /3(u) = inf{n=1.4).(1 ..a.( u)1 > a) 00 1: Po(Mn<u.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}).Mn +1 >aV(u n=1 00 S.a. assume the path structure Nt St = EUit+Zt i=1 ..e)Po (M > U).Po (M* > u.2.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u. 2. 0 yields (4. u)} < P(M* E (u . MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 . Then by Lemma 3.1 = limti00 St/t.E) Po ( n max St u. M^xu)+l > a) . Po(M* > u) . Given e > 0.. E (u ..5) which follows since Po (M > u.4. Letting first u + oo and next e ..(u) .1 can be rewritten as 00 (U) (4. .: Sn > u} . x > a. MW O(u)+1 < a) IN ( U n=1 A1.. Under suitable conditions .S.6) 1 p pBo(u) u where B is the Palm distribution of claims and p .: S.e)Po (MMX> > x).Sn 0<t<x„+j ( 1 . u))/P(M* = 0) = o(Po(M* > u)). choose a such that Po(Si > x ) > (1 .e)Po (M > u. We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4.. S. Let a > 0 be fixed. )) > (1 . Mn+l > a V (u .4.
X and N. Corollary 4. Mix) < > UE + i=1 o<t<x Thus Theorem 4.6 below.X both have tails of sup Zt.} and satisfying Zt/t N. Assume further that (i) both B and Bo are subexponential. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.268 with {Zt} continuous. the proof of Lemma 4. and also for Mix) since Nx FNX U. a4' 0. we get 00 (u) 1 IPF. since the tail of Zx is lighter than B(x) by (iv).6) holds with p = .8) x Write .v. HEAVY TAILS N` U. cf. (ii) EozNX < oo for some z > 1.1 is in force. (iii) For some o field Y. i=1 (4.7).Q = EoNx/EoX. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) . and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . oX (see Proposition A1. and ENX Ui . Proof It is easily seen that the r. independent of {> CHAPTER IX.3PB.2 Assume that {St} is regenerative and satisfies (4. The same is true for Sl.4).p) Ju P Bo(u) 1p 0 .I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .'s order EoNx • B(x).
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 . 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. . . Bo E S. More precisely. In particular. Assume that B E S.6) holds. we conclude just as in Example 4. of claims arriving in [0.(NX). i. The arrival rate is /3i and the claim size distribution Bi when Jt = i.. X2 = 1. and for some constants ci < oo such that cl + • • • + c. Taking again Xo = Xi = 0. > 0.0 (thus (iv) is trivial). Zt . We consider the case where one or more of the claim size distributions Bi are heavytailed.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).. consider the periodic model of VI. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. Thus we conclude that (4. X2 = 1.4. Bo E S.4 Assume that St = Zt . 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. The key step of the proof is the following lemma. MODELS WITH DEPENDENT INPUT 269 Example 4 . Theorem 4.6) u holds.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. X3 = 2. .e.t + EN'I Ui where {>N`1 Ui .5 Consider the Markovmodulated risk model with claim size distributions satisfying (4. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. X3 = 2. Again . (i) holds.3 that (4. . The regenerative assumption is satisfied if we take Xo = Xi = 0.. < 1. in particular lighttailed. Example 4 . Then (4. and taking F = o. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. (iii) is obvious..6) holds. The number N.. 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. we assume that B E S.9).
and F a aalgebra such that (N1. For lighttailed distributions. 2 . NP ) and X are .^•) G(x) P ^ E ciNi = C.2. and the rest of the argument is then just as the proof of Corollary 4.c'(x) where c = ciENi ...270 CHAPTER IX. .. cp with cl + > 0 it holds that Fi(x) . i1 = E\ G(x) In the general case.}P.. and that for some + cp distribution G on [0.. ..F) < CG(x)zn'1+. .5.. 6 Let (N1. i=1 Proof Consider first the case X = 0. Then P P(Yx > x) . 1. The same dominated convergence argument completes the proof. are independent with distribution Fi for Xij. Let {Fi}t=1 P be a family of distributions on [0. oo) such that G E S and some c1. oo) and define p Ni Yx = EEX'i .F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 .ciG(x). . P P P(YX and > x I. NP ) be a random vector in {0. .X i=1 j=1 where conditionally upon F the Xi. we can define the regenerations points as the times of returns to i. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin . . Assume EzN1+"'+Np < oo for some z > 1 and all i.Fmeasurable. u Proof of Theorem 4. X > 0 a r. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.. An easy conditioning argument then yields the result when Jo is u random. Thus dominated convergence yields ( P(Yo>x P(Yo>x . It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. ."+Np . i=1 P(Yx > x ^) < P(Y0 > x I.v.G( x ) > ciNi . If Jo = i. HEAVY TAILS Lemma 4 . as x a oo.
this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5.5.6) to hold in a situation where the interclaim times (T1. see Schlegel [316]. IV. Theorem 5.7. Within the class of risk processes in a Markovian environment. cf. this is applied for example to risk processes with Poisson cluster arrivals. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. As usual.p)Bo(u).4. The present approach via Theorem 4. Schmidli & Schmidt [47].2 and Example 4. 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..5 shows that basically only the tail dominant claim size distributions (those with c. Theorem 2. we let PN"N = P(.pl(1 . cf. Essentially.1. cf. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form.. Combined with the approximation for O(u). It follows from Theorem 4.4. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u. in particular Proposition 2.4. VI. as well as a condition for (4.e. ). Theorem 4. The main result of this section.T2. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). i. states that under mild additional conditions.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.1 is from Asmussen. We start by reviewing some general facts which are fundamental for the analysis.4. and independent of (T1. I T(u) < oo). > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. m is a (orfinite) . For further studies of perturbations like in Corollary 4.i.. this should be compared with the normal limit for the lighttailed case. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. Notes and references Theorem 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). ) form a general stationary sequence and the U. An improvement was given in Asmussen & Hojgaard [33]. r(u) is the time of ruin and as in IV.T2. 5 was first proved by Asmussen. the discussion provides an alternative point of view to some results in Chapter IV.3.7). i.d.. In contrast. Then O(u) . and the final reduction by Jelenkovic & Lazar [213]..
y = 0). for states i. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0.h. w(Fc) < oo ) 'In general Markov process theory. .t + EI U. Then (5. follows by the substitution y = x . and starting from So = y.2) for all bounded measurable functions h.= y.>N` Ui.2) means ffh(a. Proof Starting from Ro = x.s=j are the transition probabilities for {St}.).rij = mjsji where r13. y to vary in. resp.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w .t. St is distributed as y . Sw(F. Thus. oo). to the r.)k(x . say.00).h. r. the whole of R and not as usual impose the restrictions x > 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. and (5.z) dx G(dz) = ffh(y + z) k(y)dy G(dz).z. but the example of relevance for us is the following: Proposition 5. m. . however .s. a familiar case is time reversion (here m is the stationary distribution). u For F C E. k as indicator functions. We let QS be the corresponding distribution and Qx. j. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1.y = Qx (.r. (note that we allow x. in the terminology of general Markov process theory. The equality of the l.s.2) with t = 1 means m. Lebesgue measure. where we can take h. Say {St} is reflected Brownian motion on [0. The simplest example is a discrete time discrete state space chain. t. 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} .t. k on E. {Rt}. x = 0+ and F = (0. For the present purposes it suffices .272 CHAPTER IX. Let G denote the distribution of ENt U. to consider only the case Px(w(F`) = 0) 0. {St} and {Rt} are in classical duality w. Rt is distributed as x + t .
.. The theorem is illustrated in Fig .1 for the case F = (oo. Thus. 5. S. . the one in (b) is the time reversed path. i1.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. io = x.. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit..y as a measure on all strings of the form i0i1 . [note that w(z) < oo a.3 The distribution of r(0) given r(0) < oo.. QR and QRy are defined similarly.y = Qy Q.. z > 0.itt) = P Px(w(Fc) < 00.y() = P ({SW(F`)t} 0<t<w(F °) E So = x.s.). . w(0. We can then view Qy.y(2p21 .1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. 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. when p = .. 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). in with i0. That is.2. in E F. .SS(F.5. and we let Qy y refer to the time reversed excursion . FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time.= y) Theorem 5 .13AB < 1] Proof of Theorem 5. Sn = in = y. But in the risk theory example (corresponding to which the sample paths are drawn). Qx.2 Qy. 0]. x = 0. /^s x (S1 = Z1. . Sn+1 E Fc) nx. in = y. 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.. Sw(F)1 = y) . Sw(Fo)_ should be interpreted as Sw(F^)_1). In particular: Corollary 5.
.ik_1EF .274 note that Fx(w(Fc) < 00.. . Si l io E mjSjx. S.. ....gilt' k=1 ii . Si1y 00 jEF° E E 5xik_ 1 .. note first that Pt' (R l = il... . in E F.. t' y and Qy x are measures on all strings of the form ipi l . 2n) = Qx... Rn+1 E F`) F (w(Fc) < 00... Sn+1 E Fc) n=1 i1.. in) = Qx. Si1y k=1 i1 . in = x. in)  Pt' (R1 = ii... .in E F.y(inin _ E SYj jEF` 00 Sxik _1 .i„_iEF Similarly..... Si11 S 1 .... Silt' E Sxik_1 ....rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 ... Rn = in = x.TI( 2n2n _1 .ik1EF Similarly but easier Sxin_1 . in) = oo jEF^ Sxin1 . 20 = y. MY Thus Qx(ioii . R Qy x(2p21 . i0) Q x.... i0 = y.... Silt' E SO k=1 i1... . S. ..... = in = y. (Fc)1 = y) 00 CHAPTER IX.. ... Rn+1 E FC) TioilTili2... in = x.. HEAVY TAILS E E Px (Si = 21i ...ik_1EF Sxin_1 . Rn = in = x...ii .J (i.in1 .. in with 20. . To show Q y x (i0 i 1 .. R ..(F<)1 = Y) S S and Qx y( ipil ... ... 2p).. 21 . 2p) when 20.
y > u. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. that is. Y > y} . Z) is described in Theorem 111. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.t. 5. the P(u.B(a) +a PBBo(u) . the case r (O) < oo.r. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment .2. P(") = P(.'s are defined w. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). FINITEHORIZON RUIN PROBABILITIES 275 5b The time to ruin Our approach to the study of the asymptotic distribution of the ruin time is to decompose the path of { St} in ladder segments .2.5. S. see Fig. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. Y > u).2. U T(O) = T (u) Y Figure 5.r. That is. ST(o) > y. P(o) ).v.')distribution of Yu is Bo"). that is. 1 w .p.')density of Y is B(y)/[. Bo") is also the P(u. 7(0) < oo.(o) > y} = {T(0) < oo.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch .UBBo(u)]. Now the P(u. the distribution w.2 The distribution of (Y.t.
5.. more precisely. Bo") ). r(u)/Z 4 1/(1 . Zk be defined similarly as Y = Y1. Since the conditional distribution of Z is known (viz..p) then yields the final result T(u)/y(u) + W/(1 .T+(2). Recall the definition of the auxiliary function y(x) in Section 1.i. a slight rewriting may be more appealing.3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z).. HEAVY TAILS Let {w(z)}Z^. z ^ oo.. and YI. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long.p) in F(u) distribution. in particular of Z. P(Z < a I Y > u) 3 0.3) holds. 1/(1 . Then Corollary 5. the duration T+ (n) . Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. K(u) = n). Z = ZI but relative to the kth ladder segment. Z). 4 Assume that Bo E S and that (5. Z/'y(u) * W in Pi "' ')distribution . In the proof.p) in Pi"'')distribution. Yn_1 'typical'. cf.276 CHAPTER IX.. Then 7(u)/y(u) ^ W/(1 . However. and since its dominates the first n .p). Fig..r+ (n . this in principle determines the asymptotic behaviour of r(u).: r+ (n) < oo.. 2. i.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. i. Then.. . The idea is now to observe that if K(u) = n. then by the subexponential property Yn must be large.3. . . We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 . Zn).o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. must be large and Z1. the random vectors (YI. > u with high probability.p). are i.18(c) Bo")(yY (u)) + P(W > y) ( 5.e.e. It is straightforward that under the conditions of Proposition 1.. . Since w(z)/z a$. let r+(1) = T(0). and distributed as (Y. Hence Z.1. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b).. (Y. we get the same asymptotics as when n = 1. it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . We let K(u) = inf In = 1. Z1). That is .. conditionally upon r+ (n) < oo. ....d. .. . denote the ladder epochs and let Yk.
. .Yl+ +Yf1>u}.n) (y1. then IIP( I A'(u)) Taking A'(u) = {Y. suitably adapted).. Lemma 5. Proof We shall use the easily proved fact that if A'(u). Further. > u}. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events). FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5. .u) II 0. +Yn1<u.n). .. Y„1. II ' II denotes the total variation norm between probability measures and ® product measure.3 In the following. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1.Bo (ri1) ®B( . I A'(u)) = P(u.5. A"(u) _ {K(u)=n} = {Y1+ P(.. Yn . . I A"(u ))II + 0.5 Ilp(u.u) E • I A'(u)) = Bo (n1) ®Bou) ..Yn1iYn ..u) E •) .2. P (Yj.. P(.
.+y 1 p"F(Yn > u) P)Pn1 P/(1 . ..P) > y) Corollary 5. Z11). Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) .. and that Yk has marginal distribution B0 for k = 1. then 11P(Z E •) . (Y. HEAVY TAILS ((Z1'.. wk(Zk) has a proper limit distribution as u + oo for k < n.P) Bo(u) for n = 1. n_1 < u.y(u)T) .. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic.). Proof Let (Y11. ... Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .r. ..1 and Y„ .... see Fitzsimmons [144]). copies of {w(z)}.. the discussion just before the statement of Theorem 5. ... . By Lemma 5. Z' are arbitrary random vectors.d.i.1 P PBo(u) • P(W/(1 .6 IIPIu'n ) CHAPTER IX. Y'.. {wn(z)} be i.. . Zn are independent.. be independent random vectors such that the conditional distribution of Zk given Y. . Let {wl(z)}.4).... y > u.2. +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . whereas wn(Zn) has the same limit behaviour as when n = 1 (cf.4. Then according to Section 5a. For Theorem 5. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. Zn) E •) . in particular his Proposition (2. Z. P(u) since by Theorem 2. ..7 O (u. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit. The first step is to observe that K(u) has a proper limit distribution w.Bo (n1) ®Bo' 0.u has distribution Bout That is. .. Zn). .. ..' = y is BM. Y") u etc.p) < y). 2. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). k = 1. the density of Yn is B(y)/[IBBO(u)].P(Y' E •)II * 0.. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn)..278 Lemma 5 .1). .6.1. and clearly Zi.. n..t. in our example Y = (Y1. n ..P(Z' E •)II > 0 (here Y. Y1 +. Proof of Theorem 5. Similarly (replace u by 0). the marginal distribution of Zk is Bo for k < n.
one expects the level y form which the big jump occurs to be 0(1). the probability that is exceeds u is then B(u . 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). however.(u) = P(V > u) = f f (y) dy . x > oo. 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 .2 Define M. 3. and premium rate p(x) at level x of the reserve. cf. p(Y) and the result follows.e. Theorem 6 . .y) . More precisely. the results only cover the regularly varying case. nontrivial and we refer to Asmussen [22]. Asmussen & Teugels [53] studied approximations of i (u.1 Assume that B is subexponential and that p(x) > 00. Corollary II. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. Then 0 (u) Qf "O ^) dy. Assume for simplicity that {Vt} regenerates in state 0 . Proof of Theorem 6.B(u). that MQ becomes large as consequence of one big jump.(3 u u J B(y) dy . i./3Ea B(u). and define the cycle as a = inf{t>0: Vt=0.1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. = supo<t<0. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 .6.. Then P(MT > u) . u (6. V.2. that fo p(x)1 dx < oo.1. The rigorous proof is. T) when T + oo with u fixed. Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. The heuristic motivation is the usual in the heavytailed area. claim size distribution B. 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3.
Hence f (u)r(u) = D(u) = Do(u) . where also the (easier) case of p(x) having a finite limit is treated . It is also shown in that paper that typically. 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. there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. Further the conditional distribution of the number of downcrossings of u during a cycle given Mo > u is geometric with parameter q(u) = P(Mo > u I Vo = u). D(u) = DQ(u)/µ. u Notes and references The results are from Asmussen [22]. Then D(u) = f(u)p(u) and.q ( u)) 1 . by regenerative process theory.q(u) Now just use that p(x) * oo implies q (x) + 0.P(MT > u) $B(u) Ft µ(1 .280 CHAPTER IX. .
. 281 . vrN(z . ZN. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . Hence 1. . We shall be brief concerning general aspects and refer to standard textbooks like Bratley. The crude Monte Carlo ( CMC) method then amounts to simulating i.d.z) 4 N(0. Ripley [304].i.Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . 4Z).2) is an asymptotic 95% confidence interval . where a2 = Var(Z ).. z) 2 = Zit NE ii ii According to standard central limit theory . and this is the form in which the result of the simulation experiment is commonly reported. topics of direct relevance for the study of ruin probabilities are treated in more depth. replicates Zl. 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.96s z f (1.. Fox & Schrage [77].
b(u. However. conditional Monte Carlo and importance sampling. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . Z = I inf Rt < 0 (0<t<T = I('r(u) < T). The situation is more intricate for the infinite horizon ruin probability 0(u). and many sophisticated ideas have been developed.282 CHAPTER X. one can argue that unless Var(Z') is considerable smaller than Var(Z). Therefore. we then have EZ = EZ = z. Say that Var(Z') = Var(Z)/2. variance reduction is hardly worthwhile. We mention in particular ( regression adjusted) control variates and common random numbers. This is a classical area of the simulation literature. generated at the same time as Z. We survey two methods which are used below to study ruin probabilities. 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. 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. SIMULATION METHODOLOGY In the setting of ruin probabilities. v. Further. so that Z' is a candidate for a Monte Carlo estimator of z. an added programming effort. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). 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. and a longer CPU time to produce one replication. 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'. Letting Z' = E[Z I Y]. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. and in most cases this modest increase of N is totally unproblematic. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r .
L1).e.z2 = 0.v. . This may also be difficult to assess . L such that z = EZ = E[LZ]. even if the optimal change of measure is not practical.. Then z Var(LZ) = E(LZ)2 .E [Z Z]2 = z2 . However. (1.[E(LZ)] = E Z2 Zz . the argument cheats because we are simulating since z is not avaliable analytically. Thus. (ZN. To this end. Nevertheless. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. Thus we cannot compute L = Z/z (further. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. but tentatively. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible.. . i.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). . Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. In order to achieve (1.3).96 sis v^ N 2 1 where srs = N j(LiZi .3) Thus. and the problem is to make an efficient choice. one would try to choose P to make large values of Z more likely. it may often be impossible to describe P in such a way that it is straightforward to simulate from P). it appears that we have produced an estimator with variance zero. using the CMC method one generates (Z1. 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.1.
However.5 or even much smaller . Again.e. Z = I(A) and A is a rare event. z I..z) which tends to zero as z ^ 0. In ruin probability theory.96 2Z ( 1 .96oz /(zV) = 0. The CMC method leads to a variance of oZ = z(1 . a confidence interval of width 10 4 may look small. This leads to the equation 1. the optimal P is the conditional distribution given A. N . Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency.100 . say 10%.1. 10 . Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . i. An example where this works out nicely is given in Section 3. in terms of the halfwidth of the confidence interval. say of the order 103 or less. let z(u) = P(A(u)).e. large sample sizes are required. and let Z(u) be a Monte Carlo estimator of z(u).1. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large.z) 1 > 00. SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}).z) 1001.284 CHAPTER X. To introduce these. Z z V5 In other words . I.B = iP(AB) = P(BIA).0.. as is the case of typical interest. it does not help telling whether z is of the magnitude 104. u + oo.96 2 z2 z increases like z1 as z . We then . We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. assume that the A(u) are rare in the sense that z(u) * 0. we may try to make P look as much like P(•IA) as possible. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . For each u. 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. Thus. However. if z is small.e. but if the point estimate z is of the order 105. and further it is usually not practicable to simulate from P(•IA).
i.log Var(Z(u)) lim inf > 2 u+oo . and in practice. Generate K as geometric.4). logarithmic efficiency is almost as good as bounded relative error. Notes and references For surveys on rare events simulation. where M = X1 + • • • + XK. . Logarithmic efficiency is defined by the slightly weaker requirement that one can get as close to the power 2 as desired: Var(Z(u)) should go to 0 as least as fast as z(u)2E. However. the PollaczeckKhinchine formula III. P(K = k) = (1 . The term logarithmic comes from the equivalent form . Thus. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. 2. let Z +. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. X2. it is not efficient for large u . According to the above discussion. so that NE (u) may go to infinity.log z(u) of (1. but as a CMC method . Therefore .X1 + + XK. with common density bo(x) = B(x)/µB and K is geometric with parameter p. This allows Var(Z(u)) to decrease slightly slower than z(u)2.0. 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 X1.1) may be written as V) (u) = P(M > u). The algorithm gives a solution to the infinite horizon problem . 3. Otherwise. F(K = k) = (1 .. .i..p)pk.1.2.d. are i. see Asmussen & Rubinstein [45] and Heidelberger [190]. Let M .4) for any e > 0. it is appealing to combine with some variance reduction method . where X1. let Z +.e..(2. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. the mathematical definition puts certain restrictions on this growth rate. O (u) = z = EZ. XK from the density bo(x). where Z = I(M > u) may be generated as follows: 1.. If M > u. which gives a logarithmically efficient estimator . 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. .p)pk.
S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1). Then (cf..XK_1). Thus.. conditional probability. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x).Xl .+XK > uIXl.. So.. just note that EZ(1)(u ) 2 > E[Bo (x ..X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) .1) V)(u) .. A first obvious idea is to use conditional Monte Carlo: write i.. ..b(u) = P (Xl +•••+XK>u) = EF[Xl + .X(K1)) .. XK1. Z(1) (u) has a smaller variance than Zl (x).X(2).. . and considering only the remaining ones. form the order statistics X(1) < X(2) < . Xl > x.... To see this.. The idea of [27] is to avoid this problem by discarding the largest X.. Theorem IX.. we thus generate K and X1i . < X(K) throw away the largest one X(K).. K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X..SK1)2. XK.2.. For the simulation.p)Bo(x).. X1 + + XK_ 1 > x when X1 > x.p/(l . note first that To check the formula for the P(X(n) > x I X(1).. 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. and let Z(2)(u) = _ P (SK B0((u > u I X(l).XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0.X1 . and that Bo(y) = 1. we generate only X1.X(2). 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. . SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail... . y < 0).XK1] = EBo(uX1 ... assume in the following that Bo(x) . Z(1)(u) is defined as 0). compute Y = u .L(x)/x`' with a > 0 and L(x) slowly varying. However..286 CHAPTER X... As a conditional Monte Carlo estimator .
_1) > P(X(n) > _ X X(1). . BL by I3L = /3B[y]. X(n1)) P(X(TZ) + S(. l)) . use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") .S (n1)) V X (. that is... . using 13L. X(2). 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". the algorithm for generating Z = Z(u) is: 1. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. X(n1)) Bo((x .S(n_1) I X(1).Khinchine formula and importance sampling . BL(dx) = e7sB(dx)/B[y]. Notes and references The proof of Theorem 2. 111. Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . X(2). the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals.. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). X(2). . and define )3L. X (. Then the algorithm given by { Z (2) (u) } is logarithmically efficient. .1) . The algorithm is sofar the only efficient one which has been developed for the heavytailed case. in the renewal or Markov. Asmussen .u is the representation 0(u) = e7sr(u) overshoot (cf. .5)..y. it must be noted that a main restriction of both algorithms is that they are so intimately tied up with the compound Poisson model because the explicit form of the PollaczeckKhinchine formula is crucial (say. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. and we refer to [27]. B.1 is elementary but lengty. Thus. 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. . l)) BO(X(n1)) Theorem 2 .modulated model P(r+ < oo) and G+ are not explicit ).3. and simulate from FL. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . However . BL instead of 0.. For practical purposes.. X . for the purpose of recording Z(u) = erysr(u).
Q. . + X. If S > u. one must restrict attention to the case 4µB > 1. cf. Otherwise.l3 and U from B. to deal with the infinite horizon problem .. BL could improve the variance of the estimator . let S.i. In fact: Theorem 3.3. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.T. There are various intuitive reasons that this should be a good algorithm.d. Let FL (dx) = 'For the renewal model. let Z F e_'s. It resolves the infinite horizon problem since FL(. X2. BL).1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. and avoid simulating the known part e7". A > AL as in Chapter V. the results of IV. with distribution F. r(u) < oo) and FL (both measures restricted to. P'[y] < oo for some ry > 0. SIMULATION METHODOLOGY 3. the discussion at the end of Section 1b.7 tell that P(. b different from . Proof Just note that EZ(u)2 < e . . be i. so that changing the measure to FL is close to the optimal scheme for importance sampling . The algorithm generalizes easily to the renewal model . and the change of measure F r FL corresponds to B > BL. Generate T as being exponential with parameter . Let Sf0 CHAPTER X..S+U . Ti. b) # (/3L..r(u) < oo) = 1.288 2. M(u) = inf {n : S„ > u}.2 The estimator (3.. and we have: Theorem 3.(u)) are asymptotically coincide on {r(u)} < oo.1) (simulated with parameters ^3.F. The answer is no. return to 3. More precisely.. namely ELery£("). Let X1..2ryu _ z (u)2/C2. 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. We formulate this in a slightly more general random walk setting '. Xi = U. = X1 + . u It is tempting to ask whether choosing importance sampling parameters .. The proof is given below as a corollary to Theorem 3. 4. and assume that µF < 0 and that F[y] = 1.QL. Let S . In detail . B) is not logarithmically efficient when (/3.
= c'. Since ELM(u)/u + 1/ELXi.2 > 0.+KM(u)}. The importance sampling estimator is then Z( u) = e'rSM( ). are i. When F # FL. where Kl og (X) (j) 2 ) = log dFL (Xi) ... 1. Since K1. EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.. Here ELK.2) (simulated with distribution F of the X3 has bounded relative error when . K2..i. Proof The first statement is proved exactly as Theorem 3 . . By the chain rule for RadonNikodym derivatives.yu+elu u +oo etry' 1 > lim up C2e2.... it thus follows that for 0 < e < e'/ELXi. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . is not logarithmically efficient.3 The estimator (3.2) dF Theorem 3.P = FL. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . More generally. + KM(u))} = exp {ELM(u)(E . EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2. For the second.3.d.. (3. . where e' = EL Iog dFL (Xi) > 0 by the information inequality.2ryELXi)} .2ryELXi..yu = G. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). write W(F IF) _ F(XI). F(XM(u)).2'X1 .
T' has a left exponential tail with rate /3' and U" . . The results of IV.'(y) or y > 1/r.2. we conclude by differentiation that Bo(x)=B' (x)forallx > 0.e. In fact: Proposition 4. Next.290 which completes the proof. 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 . yu) is close to zk(u). generic interarrival times T' .T" > x) J /3"e0 yB (x + y) dy = .g.3'eO'x f f P (U" . see e. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. BL) has bounded relative error. we write T = yu. so that one would expect the change of measure F 4 FL to produce close to optimal results.T".T' D U" .T". As in IV.3. The optimality result Theorem 3.T' = U" .3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". /3'eQ'YR'( x + y) dy = . Further discussion is in Lehtonen & Nyrhinen [245]. Consider compound Poisson risk process with intensities /3'. the references in Asmussen & Rubinstein [45] and Heidelberger [190].'(y). from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. U". In [13]. T) with T < oo.4. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L.1 is from Lehtonen & Nyrhinen [244].T" has a left exponential tail with rate /3'. /3".B'=B". First by the memoryless distribution of the exponential distribution . T". with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. U' . all that needs to be shown is that if U' . U' . The queueing literature on related algorithms is extensive .i. Then according to Theorem 3. then /3' B' = B". SIMULATION METHODOLOGY u Proof of Theorem 3.4 indicate that we can expect a major difference according to whether y < 1/r. The easy case is y > 1/k'(y) where O(u. CHAPTER X. This immediately yields / = 3". The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. claim size distributions B'.1 If y > 1/ic'('y). B" and generic claim sizes U'.
3) and we have: Theorem 4. yu) is of order of magnitude a71.yu. We next consider the case y < 1/r.1) so that z(u) = zP(u. one would expect that the change of measure F Pay is in some sense optimal. yu) in the sense that .yk(ay) determines the order of magnitude of z'(u.1). lim inf uoo 27yu . and in fact. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u. 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 . T(u) < yu] .'(7).log Var(Z(u)) _ . 3 Theorem IV. yu) = eayu Eay Leay^(u)+r(u)K(ay).8 has a stronger conclusion than (4. Remark 4 .to g x ( u ) u u so that (1. Let Qy2 = .(u) * 1 (Theorem IV.4. We recall that ay is defined as the solution of a'(a) = 1/y. Proof Since ryy > y.O(u. Bay) is logarithmically efficient.8).1.4.yy> 2 .log 4')u) 4 u (Theorem IV.2) Since the definition of ay is equivalent to Eay r(u) . (4. Further .1). we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r.(ay). (4. 7y (4.4. T(u) < yu] e Hence by (4.4.2 The estimator (4. .5) follows. T( u) < yu] e2ryyuEay le 2ay^(u). Bounding u ELZ(u)2 above by a7u. and that ryy > ry.log Var(Z(u)) l im of . yu)/z. that ryy = ay . the result follows as in the proof of Theorem 3.3) (simulated with parameters /gay.
the object of interest is {Vt} rather than {Rt}.4).yu)/(uyu1/2) .7y x(u) > hm inf u+Oo U .u1/2 < r(u) < yu Le ] l = e. we believe that there are examples also in risk theory where (5.(av)Eav l e. zi(u) = INV. '%(u) = P I info Rt < 0) = P(VV > u). 5 Regenerative simulation Our starting point is the duality representations in 11.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu .Qyu1/2 < T(u) C yu e. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).3: for many risk processes {Rt }. Hence lira inf log ryyu + vyu 1/2 tc(ay) . 0 Notes and references The results of the present section are new.b(u.yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.a vt(u).4.o.3. yu .. > u) = E f I(VV > u) dt 0 (5. related discussion is given in a heavy traffic limit q J.T) = P O<t<T inf Rt < 0 = P(VT > u). N(0.1) may be useful. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .1): then by Proposition A1.uaoo U That lim sup < follows similarly but easier as when estimating En.1) is used to study Voo by simulating {Rt} (for example. Z (u)2 above.ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV.1) (see Proposition IV. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W). there exists a dual process { V t} such that i. In most of the simulation literature (say in queueing applications).2). One main example is {Vt} being regenerative (see A.o . In Asmussen [13].4.2) . However. yu .292 CHAPTER X. and (5. 0 rather than when u 3 oo. (5.1) where the identity for Vi(u) requires that Vt has a limit in distribution V.
Then (Z1z1i Z2z2 ) 4 N2(0. Z1 = (Zl1i +. letting J0 'o I (Vt > u) dt . record Zi'i = (Z1'). )). j = 1... EZ1'i = z1 = Ew.. Then Z(1).d.3) . oh) for h : R2 ^ R and Ch = VhEVh. . Thus. To derive confidence intervals . Z2 = N (X21' + . the regenerative estimator z%(u) is consistent.. 02) (5.t(u)) 4 N(0. Zl the LLN yields Z1 a$' Z(1) +. and Z2'>) where Zi'i = w. REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. + Z2N)) .. z2)) > N(O. Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). let E denote the 2 x 2 covariance matrix of Z('). EZ2'i = z2 = E Thus. + Z1N>) . i. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. .h (zl. is the cycle length.. 2. Taking h(zl. z2) z2/z1 yields Vh = (z2/z2 1/zl). which we survey below . Vh = (8h/8z1 8h/ 8z2). Therefore . provides estimates for F ( V. Z2 a4* z2.. The method of regenerative simulation.. consider first the case of independent cycles . i (^(u) ..E).+Z(N) z 1. For the ith cycle. For details . > u) (and more general expectations Eg(V. Z2 ..5. Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed.. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. Z(N) are i .
2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . 6 Sensitivity analysis We return to the problem of 111 . However . to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . see e. In 111.Z) ^Z(=) . () dx = E[SZ] f(X. asymptotic estimates were derived using the renewal equation for z /i(u).g. v. Notes and references The literature on regenerative simulation is extensive. with distribution depending on a parameter (. Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method .2. () dx so that differentiation yields zS d( fco(x)f(x. Rubinstein [310] and Rubinstein & Melamed [311].g S12 (5. () dx f Ax) (dl d()f (x' () f ( z. () depending on C.9. consider an extremely simple example . We here consider simulation algorithms which have the potential of applying to substantially more complex situations. There is potential also for combining with some variance reduction method.C)dx = f w(x) d( f ( x. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims .294 where 01 2 CHAPTER X. Before going into the complications of ruin probabilities .0 . 9. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 . Z of the form Z = ^p(X) where X is a r . Let X have a density f (x. The regenerative method is not likely to be efficient for large u but rather a brute force one. Then z(() = f cp(x) f (x.v. the expectation z = EZ of a single r.96s/v"N.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 .5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. in some situations it may be the only one resolving the infinite horizon problem .
For the SF method.r. r(u) = Tl + • • • +TM(u)). ( where h( (u. cp(h(U. () where U is uniform(0. For IPA there are. Thus .t. however . I(r(u) . one can take h (U. So assume that a r. one. ()) d( hc(U. In the setting of ruin probabilities . () = . zc = E [d co(h(U. () = (8/8()h (u. SZ is an unbiased Monte Carlo estimator of z(. The following example demonstrates how the SF method handles this situation.t. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. () Thus.log U/(. just take cp as an indicator function . () is increasing in C. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. for some Co = (o(U). ()). = E [`d (h(U. C) f(X. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . ()) h((U. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified.() d( is the score function familiar from statistics . () = d log f (X. with density f (x. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. For example . () can be generated as h(U. Thus. /3 is 0. Example 6 . Infinitesimal perturbation analysis (IPA) uses sample path derivatives. giving h( (U. To see this. Then . cp' (h(U. say W(x) = I(x > xo) and assume that h(U. /3o is M(u) Oe (3T: < oo) . ()) is 0 w . p.r. 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 . if f (x. C)).1). 11 /3oeOoT.1 Consider the sensitivity tka(u) w. The likelihood ratio up to r(u) for two Poisson processes with rates /3. the Poisson rate /3 in the compound Poisson model. C). IPA will estimate zS by 0 which is obviously not correct. () is an unbiased Monte Carlo estimator of zS.v. Then z(() = Ecp(h(U. () = log U/(2. Let M(u) be the number of claims up to the time r(u) of ruin (thus. () _ (eSx.6.
We recall (Proposition 111. Thus. In the setting of ruin probabilities. whereas for the SF method we refer to Rubinstein & Shapiro [312]. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). To resolve the infinite horizon problem . since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. .t. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) .3 (u) is of the order of magnitude ue7u. differentiating w.0(1) so that in fact the estimator Zf(u) has bounded relative error. change the measure to FL as when simulating tp(u).t. 4) that V5.r. different parameters.3 (u) (to generate Zp (u). We then arrive at the estimator ZZ(u) = (M(u) . the risk process should be simulated with parameters . Example 6. a relevant reference is VazquezAbad [374]. However. SIMULATION METHODOLOGY Taking expectation. for different models and for the sensitivities w. in part for different measures of risk than ruin probabilities.T(u)) I(T(u) < co) ] .3L. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu .296 CHAPTER X. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation .r.1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. j3 and letting flo = 0.T(u)) e7uerVu) for ?P.9 . There have been much work on resolving the difficulties associated with IPA pointed out above. BL).
})... either this makes no difference (P(R. That is. T(u.(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. 297 . where X1. T+(a) = inf It > 0 : Rt > al. in the Bernoulli random walk example below. a) = r(u) A T+(a).P(•r(u.. defined as Ro = u (with u E {0. as e. Consider first a Bernoulli random walk.. with P(Xk = 1) = 9..i.d... are i. Besides its intrinsic interest .Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion.g. The twobarrier ruin problem The twobarrier ruin probability 0. wherel T(u) = inf {t > 0 : Rt < 0} . in most cases .1}valued . Y'a(U) = P(T (u) = r+(a)) = 1 . .(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. 'Note that in the definition of r(u ) differs from the rest of the book where we use r(u) = inf {t > 0 : Rt < 0} ( two sharp inequalities ).1.. and {1.+• • •+X. R„ = u+X.. a) = r(u)). X2.. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo.
and the other more advanced but applicable also in some other settings.2) Oa(a . tba(2) _ (1 ... one elementary but difficult to generalize to other models.2). By optional stopping.(u) I\ e = 1 oa ' ()i a = u. The Lundberg equation becomes 1=F[ry]=(19)+9z..4) by ea(u+Xl+. where a is any number such that Ee°X = F[a] <oo. In a general random walk setting .. u Proof 2. zu = EzRO = EzRT(u.1) = (19)4/'0(a3)+9ba(a1).r(u..+Xn) F[ a]n n=0..1) o If 0 = 1/ 2.1 For a Bernoulli random walk with 0 0 1/2. 7/la(a .a) Y. and in view of the discrete nature of a Bernoulli random walk we write z = e7. The martingale is then {zuzXl+•••+X„ } = {zR° }.o» = z°P (RT ( u. Wald's exponential martingale is defined as in 11.. Proof 1. u + 1.298 CHAPTER XI. = z°Va(u) + za(1  .o)'t/1a(a . C1_0\a. i.1. (1.(4. z and the solution is z = (1 . the solution of F[..a) = 0) + zap ( R.e. then 'Oa(u) _ au a We give two proofs .y] = 1. and insertion shows that ( 1..(1B)u oJ 0. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2).a(u)). MISCELLANEOUS TOPICS Proposition 1.o)T/la (1) + 8z/'u(3).0)/0.1) is solution. We choose a = ry where ry is the Lundberg exponent. = (1 .
3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . Corollary 1.1 If p = 0. RANDOM WALK.. and solving for 9/la(u) yields Z/)a(u) = (e 76 . 1h (u) = a el u \1 If 9 < 1/ 2. u Proposition 1.u)/u. then Vi(u) = 1.2 For a Bernoulli random walk with 9 > 1/2. pa( u) _ u Corollary 1. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. (1.5) . then Proof Since 'Oa (U)  au a Eea(R°. However. If 9 = 1/2.1).e7u)/(e7° . Proof Let a+ oo in (1. {R.• a2µa e2µu . If p<0. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . (1. If p = 0.1 yields 't/la(u) = (a .0a(u)). Applying optional stopping to the exponential martingale {e7R.2) is trivial (z = 1).zu)/(za . } yields e7u = Ee7R° = e°Wa(u) + e7a(1 . BROWNIAN MOTION. Then for p 0 0. .1).} is then itself a martingale and we get in a similar manner u = ER° = ER ra( u) = 0 • Y'a (u) + all  au Y'a( u)).1) for p # 0.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. i1(u) = e211 .ba(u) = e2µa . thenz1 (u)=1.1.4 For a Brownian motion with drift u > 0.
0 (u) (where u p =.a) = a on {r (u.7) .7/la(u)).7).300 Proof Let a * oo in (1.. (u) _ O(u) . passing to even more general cases the method quickly becomes unfeasible (see. It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1. however. say. VIII. However.a) = a ) + e ' ° ( 1 .a) = r+ (a)} and similarly for the boundary 0.4). 1 . (1. CHAPTER XI. this immediately yields (1.a ) < 0) + e 7aF ( R (u.5a).3.+^a(u))^(a) If 7k(a) < 1.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. Ic 5ry 'pa(u) Using y = 6 .e7a (u) = 6 /0 . the paths are upwards skipfree but not downwards. 5).vi(a) Proof By the upwards skipfree property. 7/'(u) = 1). implying R(u. For most standard risk processes .5 Consider the compound Poisson model with exponential claims (with rate.616). letting a * oo yields the standard expression pe7u for . Here the undershoot under 0 is exponential with rate 5. a) I R(u a ) < 0] P (R(u . 0.e7a Again . 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.a) = a) = 5 y = P (R (u.0(a) 0 < u < a. 7O(u) = 7/la(u) + (1 . valid if p < 1 (otherwise . . Here is one more case where this is feasible: Example 1. we obtain 'Oa a7u . and thus one encounters the problem of controlling the undershoot under level 0. and hence e7u = Ee7Ro E [e7R(.a) < 0) + e7°P (R(u.
11 ) is the same as (1. T(u) E dT. (1.1a for computing ruin probabilities for a twostep premium function.ST>U).µ%T (1.9) = 2P(ST > u).. P(MT > u) = P(ST > u) + P(ST < u. (1.T) P(MT > u) where MT = maxo<t<T St. the density dPµ / dP0 of St is eµsttµ2/2.µ T I + e2µ"4) ( . BROWNIAN MOTION. (i).2 .. 10) follows then by straightforward differentiation.11) VIT ) Proof For p = 0. TWO BARRIERS 301 Note thas this argument has already been used in VII.d. and (1 .f. Corollary 1..7 For Brownian motion with drift 0. Then the density and c.µ so that {St} is Brownian motion with drift µ .1. we have ili(u. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu . For µ # 0.. Hence P(MT>u. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T . MT > U) = P(ST > u) + P(ST > u) (1.8 Let {Rt} be Brownian motion with drift .4) I = . = 1 .ST<u) = P(MT>u. MT > u) = P (ST > u) + P (ST > u. and hence Pµ('r(u) E dT) = Eo [e µsr(.8) Proof In terms of the claim surplus process { St} = {u . 0(u.Rt}. T ) = P(T(u) < T ) = 241. RANDOM WALK. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. ( 1. 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 .8 ).)_ _( u)µ2 /2. Here {St } is Brownian motion with drift 0 (starting from 0). + µ2T) } .10) Pµ (T(u) < T) !.r(u).. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities.
as defined in (1. u Small modifications also apply to Bernoulli random walks: Proposition 1.h.T (1.12) P(ST = v) = 0 otherwise. is (1. oo) with drift µ(x) and variance a2 (x) at x.T+2. T are integervalued and nonnegative. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 .s.302 CHAPTER XI.T2. (1. Thus. S(x) = f x s(y)dy.. Here {2T( (v}TT)/2) v=T. If this assumption fails.11) then follows by checking that the derivative of the r..13) with 0 as lower limit of integration. 226).10). Breiman [78] or Karlin & Taylor [222] p.T) = P(ST = u) + 2P (ST > u). close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). that 0(u)... S(oo) = f c s(y)dy. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. the behaviour at the boundary 0 is more complicated and it may happen.12) is the same as ( 1.T)dx.9) goes through unchanged. and (1.g. The same argument as used for Corollary 1. Proof The argument leading to ( 1. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.g.10) and that the value at 0 is 0.9 For Bernoulli random walk with 9 = 1/2.9). 0 0 (1. Theorem 1.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). Let s(y) = ef0 ry(.8 also applies to the case 9 54 1/2. Vi(u. see e. MISCELLANEOUS TOPICS which is the same as (1. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. e. whenever u.10 Consider a diffusion process {Rt} on [0. and in a similar spirit as in VII. oo). We finally consider a general diffusion {Rt} on [0.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion. as defined above as the probability of actually hitting 0. but we omit the details. is finite for all x > 0.
14) fails.16).b(u) = S(a) .(u) < 1 for all u > 0 and ^ S^ Conversely. The obvious boundary conditions '0a.b(b) = 1. Using s'/ s = 2p/a2. the function S(x) is . b = 0.11 Let 0 < b < u < a and let t&0. If (1.b = 0 implies that VQ b/s is constant. S(oo) < oo separately u completes the proof.16) yields 4b (u) = 1 .b('u) = Eu . if (1. Lemma 1. see in particular pp.b = a+/3S. Assume further that S (x) as defined in (1.10. For generalizations of Proposition 1. A good introduction to diffusions is in Karlin & Taylor [222]. 1'. O.e LVa. and we get Wo.13) is finite for all x > 0. 0 in (1. BROWNIAN MOTION. see Asmussen & Perry [42]. Wa.S(u)/S(a).b (Rdt) = Oa. TWO BARRIERS 303 for x > 0. then. [117].ba.b(u) + L.ba. 191195 for material related to Theorem 1. Then YIa. If b < u < a. so that Y)n. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .14) S(oo) < 00. i. we can ignore the possibility of ruin or hitting the upper barrier a before dt. (1. 15) i.b('u) = Eu &0.b(u) be the probability that {Rt} hits b before a starting from u.e.b(a) = 0 then yield the result.1.b(u)dt. A classical reference for further aspects of Bernoulli random walks is Feller [142].b(Rdt). 0 Proof of Theorem 1.10. E„ q(Rdt) = q(u)+Lq(u)dt.S(u) (1.17) Hence L. In view of (1. Letting a T oo and considering the cases S(oo) = oo.16) S(a) .S(b) Proof Recall that under mild conditions on q.0(u) = 1 for all u > 0. RANDOM WALK. then 0 < 2l. Notes and references All material of the present section is standard. Further references on twobarrier ruin problems include Dickson & Gray [116]. .6 to Markovmodulated models . Letting b J. (1 .
4) I.)AT . with the drift and the variance depending on an underlying Markov process .aR.5) A martingale proof of (2.o•K(a) = Ee .2. equivalently. (2.304 CHAPTER XI. defined by the density 1/va(u)s(u) showing up in (1.(T(u)AT) r. 111 . is currently an extremely active area of research.13)).6.(.4.5. Another basic quantity is the speed measure M .4. yu) '+/1(u) .1 ) was given already in II. (2. (2. (2.(7) .17). variance 0.3.3) < e 7yu.6) . The emphasis is often on stationary distributions . 7y = ay . but by duality. Markovmodulated Brownian models . 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. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. 1 y < k (y).5): _ z/'(u) < e 7u. much of the literature dels with the pure drift case.(a) (2.2) C_e7u < t(u) < C+e _7u. correponding to piecewise linear paths or . See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.1. one works instead with a lower limit 5 > 0 of integration in (1.ytc (ay). Lo is a martingale (cf.1) (2. Remark 11. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. Lo I.9 ) and optional stopping applied to the stopping time r(u) A T.. y > . MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0.aRo .t&(u. IV. yu) where W (ay) = y. information on ruin probabilities can be obtained . and here are alternative martingale proofs of the rest . yielding eau = Ee. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.
T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.1 .1.4). (2.yuk (ay)(u&(u. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt. .d.B(r))/B(r).( u ) I T(U) < 00] . eyuk (ay) = e7yu e > eyu"(ay ) ij(u. and the proof of the lower inequality is similar.T)  V. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).7R. For (2. dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows.3). Rt has distribution B(r + dy)/B(r). it follows easily from (2. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.yu) Y Similarly for (2.2): As noted in Proposition II. Equivalently. Let H(dt.3).1.2.f. RT(u)_) given r(u) < oo. we have tc(ay) > 0 and we can bound (2.6) below by 1 E Le7Rr(.)r(u)r. 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).4): We take a = ay in (2.E [e. A claim leading to ruin at time t has c.yu))• b(u. dr) denote the conditional distribution of (T(u).(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0). Proof of ( 2.(u. y > r. so that i/1(uL yu) < eayu . when Rt_ = r.6).6) with = 'y that eyu . (B(y) . yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II. u Proof of (2..
such that the cumulant generating function r. The limit result (3. For example. in being capable of treating many models beyond simple random walks which are not easily treated by other models . Cramer considered a random walk Sn = X1 + . . (3. However .306 CHAPTER XI.2) can be rewritten as F (Sn/n > x) 1g a'fin. then P C S. v2 later. . however .1). cle . Thus.nn or C2e. its generality. e. Thus .1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. Example 3. gn with fn + 0 .means (as at other places in the book) that the ratio is one in the limit (here n * oo). if x > EX1.1 We will go into some more detail concerning (3.?n typically only give the dominant term in an asymptotic expression .3na with a < 1. large deviations results have usually a weaker form.the correct sharp asymptotics might as well have +. logarithmic asymptotics .g. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . we will write fn 1. which in the setting of (3.1) does not capture the \ in (3.. the parameter will be u rather than n).2). and that a considerable body of theory has been developed.gn if nioo lim 109 fn = 1 log gn (later in this section. and gave sharp asymptotics for probabilities of the form P (S. The classical result in the area is Cramer's theorem. og For sequences fn. 1) but only the dominant exponential term . not quite so much in insurance risk. large deviations results been.. ri. nroo n n /// Note in particular that (3.1) where we return to the values of 0.(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. 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.1) is an example of sharp asymptotics : . gn 4 0./n E I) for intervals I C R. Accordingly.. The advantage of the large deviations approach is..^ e nn 1 > x n 0o 2xn (3. (3. + X.
which is a saddlepoint equation .3) is put equal to x. S rtn > x 1.r.3.4) n Next.425. 2 where o2 = o2(x) = rc"(0).2). since Sn is asymptotically normal w.9S„+n' ( 9).t. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . we get P(Sn/n > x) E [e9nx +nK(9)9" '.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.960/) * 0. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3. (3. Define . of P(X1 E dx) = E[e9X1K. and hence for large n P(Sn/n > x) > E [e. if we replace Sn by nx + o / V where V is N(0. nx < Sn < nx + 1.4) immediately yields (3. LARGE DEVIATIONS Define rc* as the convex conjugate of rc. In fact.r.sseo f which in conjunction with (3. exponential change of measure is a key tool in large deviations methods. rc*(x) = sup(Ox . i. P with mean nx and variance no.rc(0) where 0 = 0(x) is the solution of x = rc'(0).4 enn +1.1).96o /] > 0.q = rc* (x).(0)) e 307 (other names are the entropy. More precisely. V > 0 e. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function).(e)i XI E dx]. we have P(nx < Sn < nx + 1.e. Most often. Since P nn > x) = E {e_8 ' ( 9).
which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn.. which is of similar spirit as the dicussion in VII. .v. For the proof. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y . Further main results in large deviations theory are the GartnerEllis theorem. 1) and no such that Sn .. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0.1). Assume that there exists 'y. there exists z E (0.3. and the WentzellFreidlin theory of slow Markov walks.h. see Jensen u [215] or [APQ] p.. Xn given by Fn(dxl. MISCELLANEOUS TOPICS which is the same as (3. Pn Sn1 .. 1]. Xn) and sn = x1 + • • • + xn (note that the r. n Icn(0) exists and is finite for ry . be a sequence of r.'s..dxn) where Fn is the distribution of (X1i .3 For each i > 0. we introduce a change of measure for X1. Then i/.2 (GLYNN & WHITT [163]) Let X1.308 CHAPTER XI.. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo)../^ >7 < zn n for n n0. integrates to 1 by the definition of Icn).. (iii) #c (8) = limn. .o log Ee9Sn /n... In the application of large deviations to ruin probabilities. The substitution by V needs.. is differentiable at ry with 0 < K'(y) < 00. that is. Ee9X n < oo for e < 0 < e. 260 for details. however. Mogulskii's theorem which gives path asymptotics. We shall need: Lemma 3 .s. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 . .. and write Sn = X1 + • • • + Xn. to be made rigorous..e < 8 < y + e..e < 8 < y + e.. We further write µ = tc'(ry). . commonly denoted as is the saddlepoint approximation.dxn) = 05nKn(7)Fn(dx1. X2... .'(u) )Ng a"u.p > 7 < zn. (ii) lim supn. (iv) tc(ry) = 0 and r. Sanov's theorem which give rare events asymptotics for empirical distributions.
for Sn.+r. h. u Proof of Theorem 3...3. The rest of the argument is as before.2. S.YS.> . 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.Bµ . LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . This establishes the first claim of the lemma . in particular the r. can be chosen strictly negative by taking 9 small enough. Since I EeqOX „ ] 1/q is bounded for large n by (ii). Then V.77) follows by symmetry (note that the argument did not use µ > 0)..]1/q = e. 0. is of order . mµ Sm > u] km e7Sm+n. the r .n m µ 1 + rl .s. We first show that lim inf„_. > 1 +17] m(7).2. This proves the existence of z < 1 and no such that Pn (Sn/n > µ.n > u ) = [ Em [em Em 1e.h..W.91) + o(O ) as 0 J. h.71 < e and jq9j < e. The corresponding claim for Pn(Sn/n < µ . log zl'(u)/u > 'y. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).r (7) n = e.077 n^oo n and by Taylor expansion and (iv ). Clearly.m(7).y) .s.µ?7 .n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX..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 +.ne(p+ 17).s. it is easy to see that the r. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1.+r7) < zn for n > no. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0. ( U) P(S. S. 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.
0 log i'(u )/u < 'y. I > IL exp `S. I2 = F(T(u) = n). (iv) and Lemma 3. MISCELLANEOUS TOPICS (7).n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (.. P(T(u) = n) < P(Sn > u) = En [e7S.. Pn \ > la+ 8 I < zn (3.. 3.6) for some z < 1 and all n > n(E).(Y).. this is possible by (iii). n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n). Sn > u] < eYu+Kn(7)pn(Sn > u) (3. Obviously.310 ]Em I e.(•) goes to 1 by Lemma 3..I < µl1 1+77 I M 1_ 1+277 S. we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J. and since Ic.YS +^c CHAPTER XI.7) so that n(b) I1 < e'Yu E en. logO(u)/u > ry... n=1 .+wn(7).3.n(ry)/u 4 0andm/u* (1 + r7)/µ. 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). For lim supu.log z) /2 and Sn Fn\ n >lb+S) <Zn. 0 yields liminfu __.
> u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .zl/z en6 [u(1 +6)/µJ 1u (1 +6) /µJ ekn(7) < e' 13 < C" E Yu l u(16)/lij+1 Lu(16)/µJ+l1 < e7U Finally.11) [u(1+6)/µJ+1 1  Thus an upper bound for z/'(u) is n(6) e'Yu n=1 eKn (7) + 2 + (28U + 1) e6u(1+6)/µ Fi 1 zl /2 and using (i). 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 . we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. u .3.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. C 26u `p / +1 I e6u(1+6)/µ (3.10) 00 I4 < E F(Sn_1 < u. S. Sn1 C U. Sn > U] [ e(u(1+6)/µJ+l < eYu (u(1+6)/µJ+1 7u r 0 0 e L^ en('Y ) fPn (I Sn 1 . eryu en logz/2p n nt n.
the last steps of (3. we replace the bound P(Sn > u ) < 1 used in (3. ryue«iu . 2.312 CHAPTER XI. 13 = P(T (u) E (u(1 b)l^ (7). cf. say n n1.3ui where . For 14. I2.b)/i(7).z 1/z For I1.8) by P(S.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . Corollary 3.(u) = I1+I2+I3+I4'^ ery( u). this is straightforward since the last inequality in (3. we get Lou] E exp {( 7 + a)u + Kn(a +7)} n=1 Il Lou] exp {(y + a)u} { 111 + exp {4narc'(7)} n=1 exp {('y + a)u} c1 exp {4/3uarc'(7)} = clewhere a1 = aw. Letting c11 = maxn<n. e'.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 .('+'Y). u(1 + b)/i(7)) Proof Since V..4/3rc'(y) > 0. it suffices to show that for j = 1.4. we need to redefine n(b) as L. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. u .Q is so small that w = 1 . Then for n large. For I. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 . > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. the typical time is u/rc'(7) just as for the compound Poisson model.xl/2 to give the desired conclusion.4 Under the assumptions of Theorem 3. 4 there is an aj > 0 and a cj < oo such that Ij < c3e.7' a"ju. For 12.. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7). (7 + a) < 2arc'(7).2.u(1+b)/rc'(7)). IV.
An event occuring at time s is rewarded by a r. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate . To verify these in concrete examples may well present considerable difficulties.f. Obviously many of the most interesting examples have a continuous time scale. If {St}t> 0 is the claims surplus process.. V(s) with m. criteria are given in Duffield & O'Connell [124]..14) is needed in both examples . for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0. 11 Inspection of the proof of Theorem 3.. and we conclude that Theorem 3 .. t] is Rt = E V (Un) n: o„ <t . 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.2 shows that the discrete time structure is used in an essential way..13) One would expect this to hold in considerable generality. the key condition similar to (iii). Assuming that the further regularity conditions can be verified.3(s) at time s. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * . and in fact.2 then immediately yields the estimate log F( sup Skh > u) a7u (3.'(y) > 0.12) k=0.(O) = 9µ+02 for all 9 E R. we shall give two continuous time examples and tacitly assume that this can be done. Theorem 3.1.g.. 09(9). Xk+l) k=1 00 naoo n provided the sum converges absolutely. Thus the total reward in the interval [0.v. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3.3.e.1. r. 2 is in force with y = 2p/wz. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process.LARGE DEVIATIONS 313 Example 3 ..5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. but nevertheless. i. whether P ( sup St > u ltg a ^" 0<t<oo // (3..
noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. i. . if the nth claim arrives at time a..Q„) . Most obviously. Then logEeOR° = J0 /3(s)(^8(9) . the Cramer. it contributes to St by the amount Un(t . More precisely.9t. Un represents the total payment for the nth claim).14).6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . n: o. Un(s).s). We let ic (9) = 3(EeWU° . (3. e > 0 such that ic('y) = 0 and that r. we conclude that Cu) log e7 u (cf. 0 and since EeOUn(8) + Ee°U^ as s * oo. 7 Given the safety loading 77.v. this is not realistic .1) ds .1) . is At . An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . we have rct (9)/t 4 ic (9).14) (to see this . = U„ ( t . one would take p(t) = (1 + rt)At/ t.g.1) ds rt (3.'`1 U. 0 Example 3 . We further assume that the processes {U1(s)}8>0 are i. derive . e. Thus. Thus by (3. the above discussion of discrete skeletons). Kt (0) t (Ee9U"it8i J0 . At = . the best estimator of /3µB based upon Ft.Lundberg model has the larger ruin probability. O'n +S] is a r . then the payments from the company in [on. but that a claim is not settled immediately. Of course .15) . Of course.2 are trivial to verify. MISCELLANEOUS TOPICS are the event times. we have S. Thus.. leading to St = At(1+77) Joo t S8 ds.t.It. Example 3. Since the remaining conditions of Theorem 3.d.. nondecreasing and with finite limits Un as s T oo ( thus. <t which is a shotnoise process.. assuming a continuous premium inflow at unit rate. If the nth claim arrives at time Qn = s.0 and assume there are y. where Ft = a(A8 : 0 < s < t). It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot . (9) < oo for 9 < 'y + C. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. a differential equation in t).314 where the an CHAPTER XI.1) ds .9t = /3 J t (Ee8U° i8l .
and since the remaining conditions are trivial to verify. It then follows from (3.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.d. the Vi = .1) .3.2 hold.e. LARGE DEVIATIONS With the Qi the arrival times.(1 +i) f > i= 1 s ds = E Ui 1 .1) or . i. 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 .i.(1 + 17)0µB = 0. the solution of /3(Eelu .18) Thus (iii) of Theorem 3. we have Nt t N.i. standard exponential . uniform (0.b(u) IN a'Yu (cf. (3. Ui Nt / t 01i 315 St = Ui .16) i=1 o i=1 Let ict (a) = log Eeast . rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 .20) (3./3.19) with equality if and only if U is degenerate. we conclude that t.(1 + r7) log t (3. typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .d. (3.log Oi are i. Thus.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )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 . one has y > y' (3. equivalently.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 . Indeed.
MartinL6f [256]. and k(x) < 0.7. Further. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. The main example is Nt being Poisson with rate fit. much of the analysis carries over to more general cases. . with common distribution B and independent of Nt. the function k(x) = e7*x . this in turn yields y > y*. a* (s) are convex with tc'(0) < 0 . In particular. the study is motivated from the formulas in IV. 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. though we do not always spell this out. rc*' (0 ) < 0.d.1 . are i. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U.2. [257] and Nyrhinen [275]. y = y* can only occur if U .20) is due to Tatyana Turova. the proof of (3.2 expressing the finite horizon ruin probabilities in terms of the distribution of A. In addition to Glynn & Whitt [163].(1 + ri)y*x is convex with k(oo) = 00. For notational simplicity. and since tc(s). MISCELLANEOUS TOPICS Next. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. = P(N = n) = e(3an However. Lehtonen & Nyrhinen [244]. x > x0. Further applications of large deviations idea in risk theory occur in Djehiche [122].19). say one year.xo. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. k(0) = 0. see Nyrhinen [275] and Asmussen [25]. Therefore e7'U _ k(U) E [1+(1+77)y*U] . This implies n(y*) < 0. k'(0) < 0. Further. 0 < x < x0. [245]. we then take t = 1 so that p. For Example 3. assuming that the U. at time t.316 CHAPTER XI.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. see also Nyrhinen [275] for Theorem 3.i. we are interested in estimating P(A > x) for large x. using that Ek(U) = 0 because of (3..
4.1. we define the saddlepoint 9 = 9(x) by EBA = x. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4.9(Ax). Proposition 4. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution. only with 0 replaced by a9 and B by B9. B"' [s] lim (B". For a given x.1) where )30 = . 818' where s' = sup{s : B[s] < oo}.2) implies that the limiting Pedistribution of (A . A > x)] = eex+K( e)E9 [e . e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x.3B"[9]. In particular. A E dx] .ic(9) = .3e(bo[a] .x)//3B"[9] is standard normal. Vare(A) = s. B"[s] = oo.1).e.1 Assume that lim8T8."(0) = . K'(0) _ ic'(9) = x. A > x) eex+K(e ) ee AB°[ely 1 ev2/2 dy 0 2^ 00 9x+p(e) e ezez2/(2BZpB „[9)) dz 9 27r/3B" [9] fo eex+w ( e) oo z x)] ] 0 27r /3B" [9] o e 9 2 /3B" [9] J eex+w(B) dz . Hence P(A > x) = E e [e9A+ ic(9). The analysis largely follows Example 3.1). Then Ee"A = e'(") where x(a) _ 0(B[a] . The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). (4.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). Then as x * oo. no(a) = logE9e'A = rc(a + 9) . i.[s])3/2 = 0.
For example.1 goes all the way back to Esscher [141]. The present proof is somewhat heuristical in the CLT steps. [138].3) and related results u for the case of main interest . Thus . i. then P(A > x) . or. Var(A) _ ^3p. either of the following is sufficient: A. 4b The NP approximation In many cases .(D X . b is gammalike.EN B(x). For details.2i and that (A .l3pB.Q{AB (4.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. the distribution of A is approximately normal .e. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. In fact. see Embrechts et al.v. large x. just the same dominated convergence argument as in the proof of Theorem 2.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. it holds that EA = .x') where x' = sup {x : b(x) > 0}.318 CHAPTER XI. b is logconcave. For a rigorous proof. more generally.4) . For example. 2). b(x) = q(x)eh(z). Remark 4 . In particular.ycix °ie6x B. bounded with b(x) .1). Jensen [215] and references therein. it is quite questionable to use (4. Furthermore 00 b(x)Sdx < oo for some ( E (1. 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 . leading to P(A > x) :.2) is often referred to as the Esscher approximation. B covers distributions with finite support or with a density not too far from ax° with a > 1.1 yields: Proposition 4.2 If B is subexponential and EzN < oo for some z > 1. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. A covers the exponential distribution and phasetype distributions. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. some regularity of the density b(x) of B is required. and (4. 1 . Notes and references Proposition 4. under the Poisson assumption (4.
2 ^ \1 . (4. Let Y = (A .5) may be negative and is not necessarily an increasing function of y for jyj large..2K3 + 4i 64 + . . of (4.6(1 .: EA + zl_E Var(A) . in particular.h. the standard normal distribution. In concrete examples . (4.equantile in the distribution of Y. then P(Y < y) 4(y) .l = EY..c2i. where Kl .2X2 . so that 1(u) 3 exp { . . Rather than with the tail probabilities F(A > x)...3& (y). f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) . Heuristically. zl_e be the 1 . u5. the CLT for Y = Y6 is usually derived via expanding the ch. and from this (4.. are small. K2 = Var (Y). resp. Var(Y) = 1 as above .y2)^P(y)• 319 Note as a further warning that the r.f.4. and so as a first approximation we obtain a1_E = EA + yle Var(A) . Thus if EY = 0. are the cumulants .5) follows by integration. Remark 4.i 6 r 1 3 so that we should take b = ic3/6 in (4. K3 = E(Y .5) is obtained by noting that by Fourier inversion. ylE should be close to zl_E (cf. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. one expects the u3 term to dominate the terms of order u4. s. however.EY)3.s.5). the density of Y is 1 °° _ eiuy f(u) du 2x _. one needs to show that 163.i 3 K3 } Pt^ exp .3!).. If the distribution of Y is close to N(0.1). K4 .2 2 .6) . as u2 u3 u4 9(u) = Ee'uY = exp {iuci ..EA)/ Var(A) and let yl_E.. If this holds .99.e. . A particular case is a.5 (y3 . defined as the the solution of P(A < yle) = 1 .. the NP (normal power) approximation deals with the quantile al_E. which is often denoted VaR (the Value at Risk).
[101]. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n ..7) as 1 (3) a1E = Qµa +z1 .E )Azl E) 4(z1E) + ( ylE .6 (1 .3ni /3 .1) E (A .E .E = z1E + S(zi_E .S(1 .zl E)V(zl_E) . let pn Pn = (a+ = P(N = n). This leads to t( yl E) . this holds with a = 0..zlE)W(zlE) 1 .320 CHAPTER XI. the kth cumulant of A is /3PBk' and so s.E)A1 l E) 1 E 4)(yl E) ^' .zi.. 4c Panjer 's recursion Consider A = constants a.1)! n ^eQ . k3 is small for large /3 but dominates 1c4. n = 1.1)EY3. as required .. For example.(y) terms dominate the S(1 . Note.y2)cp( y) term.. and assume that there exist n ) Pn_i .zl E)^o(zl E) .k = /3µB^1 / (.E(/3PB^1 )1^2 + s(z1E .. .6pBki) d/2. however.5) by noting that the 4. MISCELLANEOUS TOPICS A correction term may be computed from (4. In particular .5(1 . 21 .1).E + (yl. b such that EN 1 U%.yi.S(1 . We can rewrite (4. K5 . Another main reference is Daykin et at.1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. that [101] distinguishes between the NP and Edgeworth approximations. Using Y = (A . this yields the NP approximation 6(Z1 _E .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.EA ) / Var(A)..zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 .
which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4.1. 2.k . By symmetry.5 The crux of Proposition 4..14) is therefore a + b/n. u Proof of Proposition 4. .4 Assume that B is concentrated on {0..12) where g*n is the nth convolution power of g. j = 0.13) Namely. then j (a + b!) 1ag k_1 3 gkfj. Then fo = >20 9onpn and fi = 1 E In particular. (4.12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 .4.14) is independent of i = 1.11) Remark 4. n = k=n1 9k(n1 )9j k • (4. .. and calculating the gj*n recursively by 9*1 = 9j. fj = P(A = j). if go = 0. (4. . THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4.4 is that the algorithm is much faster than the naive method. . Since the sum over i is na + b. The expression for fo is obvious.} and write gj = 2 . . E[a +bU=I >Ui =j l i=1 J (4...4. . . . 2.12)..9). fj = E (a+ b k =1 )9kfi_k . 2.. ..13) but only O(j2) for Proposition 4. (4. the value of (4. 1. j1 g.4. Hence by (4...10) f o = po... j = 1. (4. n. j = 1. the complexity (number of arithmetic operations required) is O(j3) for (4.
322
00 J
CHAPTER XI. MISCELLANEOUS TOPICS
EE (a + bk I gkg3 _ k lieni n=ik=0 (a+bk l gkE g j'`kpn = E (a+b!)9kfi_k n=0 k=0 k=0 ^I 1 E(a+b. agofj+ k Jgkfjk, k=i /
and and (4.9) follows . (4.11) is a trivial special case.
u
If the distribution B of the Ui is nonlattice , it is natural to use a discrete approximation . To this end, let U(;+, U(h) be U; rounded upwards, resp. downwards , to the nearest multiple of h and let A}h) = EN U. An obvious modification of Proposition 4.4 applies to evaluate the distribution F(h) of A(h) letting f( ) = P(A() = jh) and
g(h) gkh+
= P (U(h2 = kh) = B((k + 1)h)  B(kh ), k = 0, 1, 2, ... , = P (U4;+ = kh) = B(kh)  B (( k  1)h) = gk  l,, k = 1, 2, ... .
Then the error on the tail probabilities (which can be taken arbitrarily small by choosing h small enough ) can be evaluated by
00 00
< P(A > x ) f (h) j=Lx/hl j=Lx/hl
Further examples ( and in fact the only ones , cf. Sundt & Jewell [355]) where (4.9) holds are the binomial distribution and the negative binomial (in particular, geometric ) distribution . The geometric case is of particular importance because of the following result which immediately follows from by combining Proposition 4.4 and the PollaczeckKhinchine representation: Corollary 4.6 Consider a compound Poisson risk process with Poisson rate 0 and claim size distribution B. Then for any h > 0, the ruin probability zb(u) satisfies 00 00
f^,h) Cu) < E ff,+, j=Lu/hJ j=Lu/hJ (4.15)
f! h)
5. PRINCIPLES FOR PREMIUM CALCULATION
where f^ +, f^ h) are given by the recursions
(h) 3 (h) (h)
323
fj,+ = P 9k fjk,+ ' I = 17 2, .. .
k=1 3 (h)
(h)
=
P
(h)
f9,  (h) gk,fAk, e 1  ago, k=1
j = 1+2,
starting from fo + = 1  p, f(h) = (1  p)/(1  pgoh) and using 07
g(kh) 1 (k+1)h
=
Bo((k + 1 ) h)  Bo(kh ) =  f
AB
kh
B(x) dx, k = 0, 1, 2, ... , k = 1,2 .....
gkh+
Bo(kh )  Bo((k  1 ) h) = 9kh)1 ,
Notes and references The literature on recursive algorithms related to Panjer's recursion is extensive, see e.g. Dickson [115] and references therein.
5 Principles for premium calculation
The standard setting for discussing premium calculation in the actuarial literature does not involve stochastic processes, but only a single risk X > 0. By this we mean that X is a r.v. representing the random payment to be made (possibly 0). A premium rule is then a [0, oo)valued function H of the distribution of X, often written H(X), such that H(X) is the premium to be paid, i.e. the amount for which the company is willing to insure the given risk. The standard premium rules discussed in the literature (not necessarily the same which are used in practice!) are the following: The net premium principle H(X) = EX (also called the equivalence principle). As follows from the fluctuation theory of r.v.'s with mean, this principle will lead to ruin if many independent risks are insured. This motivates the next principle, The expected value principle H(X) = (1 + 77)EX where 77 is a specified safety loading. For 77 = 0, we are back to the net premium principle. A criticism of the expected value principle is that it does not take into account the variability of X which leads to The variance principle H(X) = EX+77Var(X). A modification (motivated from EX and Var(X) not having the same dimension) is
324
CHAPTER XI. MISCELLANEOUS TOPICS
Var(X).
The standard deviation principle H(X) = EX +rl
The principle of zero utility. Here v(x) is a given utility function, assumed to be concave and increasing with (w.lo.g) v(O) = 0; v(x) represents the utility of a capital of size x . The zero utility principle then means v(0) = Ev (H(X)  X); (5.1)
a generalization v(u) = Ev (u + H(X)  X ) takes into account the initial reserve u of the company. By Jensen 's inequality, v(H(X)  EX) > Ev(H(X)  X) = 0 so that H(X) > EX. For v(x) = x, we have equality and are back to the net premium principle. There is also an approximate argument leading to the variance principle as follows. Assuming that the Taylor approximation
v(H(X)  X) ^ 0 +v'(0)(H (X)  X) + v 0 (H(X)  X)2 ,/2
is reasonable , taking expectations leads to the quadratic v"H(X )2 + H(X) (2v'  2v"EX) + v"EX2  2v'EX = 0 (with v', v" evaluated at 0) with solution
H(X)=EXv^±V( ^ )2Var(X).
Write
( vI ) 2 \
Var(X) v^  2v^Var(X)/ I  (
, Var(X) )2
If v"/v' is small, we can ignore the last term. Taking +f then yields H(X) ,:: EX 
2v'(0) VarX;
since v"(0) < 0 by concavity, this is approximately the variance principle. The most important special case of the principle of zero utility is The exponential principle which corresponds to v(x) = (1  e6x)/a for some a > 0. Here (5.1) is equivalent to 0 = 1  e0H(X)EeaX, and we get
H(X) = 1 log Ee 0X .
a
5. PRINCIPLES FOR PREMIUM CALCULATION
325
Since m.g.f.'s are logconcave, it follows that H,, (X) = H(X) is increasing as function of a. Further, limQyo Ha (X) = EX (the net premium princiHa (X) = b (the premium ple) and, provided b = ess supX < oo, lim,, H(X) = b is called the maximal loss principle but is clearly not principle very realistic). In view of this, a is called the risk aversion The percentile principle Here one chooses a (small ) number a, say 0.05 or 0.01, and determines H(X) by P(X < H(X)) = 1  a (assuming a continuous distribution for simplicity). Some standard criteria for evaluating the merits of premium rules are 1. 77 > 0, i .e. H(X) > EX. 2. H(X) < b when b (the ess sup above ) is finite 3. H(X + c) = H(X) + c for any constant c
4. H(X + Y) = H(X) + H(Y) when X, Y are independent
5. H(X) = H(H(XIY)). For example , if X = EN U= is a random sum with the U; independent of N, this yields
H
C^
U; I = H(H(U)N)
(where, of course, H(U) is a constant). Note that H(cX) = cH(X) is not on the list! Considering the examples above, the net premium principle and the exponential principle can be seen to the only ones satisfying all five properties. The expected value principle fails to satisy, e.g., 3), whereas (at least) 4) is violated for the variance principle, the standard deviation principle, and the zero utility principle (unless it is the exponential or net premium principle). For more detail, see e.g. Gerber [157] or Sundt [354]. Proposition 5.1 Consider the compound Poisson case and assume that the premium p is calculated using the exponential principle with time horizon h > 0. That is,
N,,
Ev I P  E U;
i =1
= 0 where
v(x) = 1(1  e°x
a
Then ry = a, i.e. the adjustment coefficient 'y coincides with the risk aversion a.
326
Proof The assumption means
CHAPTER XI. MISCELLANEOUS TOPICS
0 a (1  eareo (B[a11)
l
i.e. /3(B[a]  1)  ap = 0 which is the same as saying that a solves the Lundberg u equation. Notes and references The theory exposed is standard and can be found in many texts on insurance mathematics, e.g. Gerber [157], Heilman [191] and Sundt [354]. For an extensive treatment, see Goovaerts et al. [165].
6 Reinsurance
Reinsurance means that the company (the cedent) insures a part of the risk at another insurance company (the reinsurer). Again, we start by formulation the basic concepts within the framework of a single risk X _> 0. A reinsurance arrangement is then defined in terms of a function h(x) with the property h(x) < x. Here h(x) is the amount of the claim x to be paid by the reinsurer and x  h(x) by the the amount to be paid by the cedent. The function x  h(x) is referred to as the retention function. The most common examples are the following two: Proportional reinsurance h(x) = Ox for some 0 E (0, 1). Also called quota share reinsurance. Stoploss reinsurance h(x) = (x  b)+ for some b E (0, oo), referred to as the retention limit. Note that the retention function is x A b. Concerning terminology, note that in the actuarial literature the stoploss transform of F(x) = P(X < x) (or, equivalently, of X), is defined as the function
b * E(X  b)+ =
f
(s  b)F(dx) _ f
6 00
(x) dx.
An arrangement closely related to stoploss reinsurance is excessofloss reinsurance, see below.
Stoploss reinsurance and excessofloss reinsurance have a number of nice optimality properties. The first we prove is in terms of maximal utility: Proposition 6.1 Let X be a given risk, v a given concave nondecreasing utility function and h a given retention function. Let further b be determined by E(X b)+ = Eh(X). Then for any x,
Ev(x  {X  h(X)}) < Ev(x  X A b).
6. REINSURANCE
327
Remark 6 .2 Proposition 6.1 can be interpreted as follows. Assume that the cedent charges a premium P > EX for the risk X and is willing to pay P1 < P for reinsurance. If the reinsurer applies the expected value principle with safety loading q, this implies that the cedent is looking for retention functions with Eh(X) = P2 = P1/(1 + 77). The expected utility after settling the risk is thus
Ev(u + P  P1  {X  h(X)})
where u is the initial reserve . Letting x = u + P  P1, Proposition 6.1 shows that the stoploss rule h (X) = (X  b)+ with b chosen such that E(X  b)+ u = P2 maximizes the expected utility. For the proof of Proposition 6.1, we shall need the following lemma: Lemma 6 .3 (OHLIN'S LEMMA) Let X1, X2 be two risks with the same mean, such that Fj(x) < F2 (x), x < b, Fi(x) ? F2(x), x > b for some b where Fi(x) = P(Xi < x). Then Eg(X1) < g(X2) for any convex function g. Proof Let Yi=XiAb, Zi=Xivb.
Then
P(Yl < x) _ Fi(x) <_ F2 (x) = P(Y2 < x) x < b 1=P(Y2<x) x>b so that Y1 is larger than Y2 in the sense of stochastical ordering . Similarly, P(Zl < x) _ 0 = P(Z2 < x) x < b Fi(x) > F2(x) = P(Z2 < x) x > b
so that Z2 is larger than Zl in stochastical ordering. Since by convexity, v(x) = g(x)  g(b)  g'(b)(x  b) is nonincreasing on [0, b] and nondecreasing on [b, oo), it follows that Ev(Y1) < Ev(Y2), Ev(Zi) < Ev(Z2). Using v(Yi) + v(Zi) = v(Xi), it follows that
0 < Ev(X2)  Ev(Xi) = Eg(X2)  Eg(X1),
using EX1 = EX2 in the last step. u
Proof of Proposition 6.1. It is easily seen that the asssumptions of Ohlin' s lemma hold when X1 = X A b, X2 = X  h(X); in particular, the requirement EX1
328
CHAPTER XI. MISCELLANEOUS TOPICS
= EX2 is then equivalent to E(X  b)+ = Eh(X). Now just note that v is convex. u
We now turn to the case where the risk can be written as N
X = Ui
i=1
with the Ui independent; N may be random but should then be independent of the Ui. Typically, N could be the number of claims in a given period, say a year, and the Ui the corresponding claim sizes. A reinsurance arrangement of the form h(X) as above is called global; if instead h is applied to the individual claims so that the reinsurer pays the amount EN h(Ui), the arrangement is called local (more generally, one could consider EN hi(Ui) but we shall not discuss this). The following discussion will focus on maximizing the adjustment coefficient. For a global rule with retention function h* (x) and a given premium P* charged for X  h* (X), the cedents adjustment coefficient y* is determined by
1 = Eexp {ry*[X  h*(X)  P*]},
for a local rule corresponding to h(u) and premium P for X look instead for the ry solving
J _f
(6.2) N 1 h (Ui), we
[ X_P_^
1 = Eexp
[ Ei  h(Ui)] P [U
= Eexp{ry
h(Ui)]
l (6.3) This definition of the adjustment coefficients is motivated by considering ruin at a sequence of equally spaced time points, say consecutive years, such that N is the generic number of claims in a year and P, P* the total premiums charged in a year, and referring to the results of V.3a. The following result shows that if we compare only arrangements with P = P*, a global rule if preferable to a local one. Proposition 6.4 To any local rule with retention function h(u) and any
N
J}
P > E X  N h(Ui)
4 =1
(6.4)
there is a global rule with retention function h* (x) such that
N
Eh*(X) = Eh(U1)
i=1
and 'y* > ry where ry* is evaluated with P* = P in (6.3).
d.h * (X) .6. u But since ry > 0.h(U) (as in the proof of Proposition 6.. Proof As in the proof of Proposition 6.P > EexP{7[X . expectations like those in (6. Then for any local retention function u . then (6. The arrangement used in practice is.h(Ui)] . ry* > 0 because of (6.4) and u g(x) = e7x in Ohlin's lemma. Local reinsurance with h(u) = (u .6). appealing to (6.b)+ with b determined by E(U . this implies 7* > 7.P.4).5) holds trivially.4).P I = EC [7]N.4). REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x . however. (6.h( UU) = EN • E[U .h(Ui)] .b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. Remark 6. ' ii (6. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy. the excess ofloss rule hl (u) = (u . This follows by taking Xl = U A b. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') . we get EX = EN • EU. and so on. as often local as global. it suffices to show that Eexp {ry ii 'UiAb. .6) u where C[ry] = Ee'r(u4(u)). we get N 1 = Eexp ry E[Ui ii .6 Assume the Ui are i. X2 = U . i. Assuming for simplicity that the Ui are i.h(Ui)P JJJ l:='l {ry ] or. (6.d.3).h(Ui)] .i. N E X .P } < 1 = Eexp E[Ui. y = Ei [Ui .5 Because of the independence assumptions .5) reduce quite a lot.h(u) and any P satisfying (6. Eexp 7 [E [Ui .4.b)+ = Eh(U) (and the same P) satisfies 71 > ry.P]}.h(U)].
MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. The original reference for Ohlin's lemma is Ohlin [277].many texts on insurance mathematics. See further Hesselager [194] and Dickson & Waters [120].330 CHAPTER XI. Heilman [191] and Sundt [354]. [76]. see also Sundt [354]. Bowers et at. .g. The present proof is from van Dawen [99]. e.
e. some condition is needed: that F is nonlattice.. t] is denoted by Nt. when t is large. then Stone 's decomposition holds : U = U. t]) so that U(t + a) ... all have the same distribution. i.. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. Technically. Lebesgue measure for some n > 1). t 00 (A. t +a]). If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . The renewal theorem asserts that U(dt) is close to dt/µ.. = T„ . of interarrival times and the time Yo = To of the first arrival (that is. . note in particular that U({0}) = 1. Y. Y2.. of epochs or the set Y1. Lebesgue measure dt normalized by the mean to of F. If Yo = 0. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. Y2.. .. the renewal process is called zerodelayed.. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . That is. denoted by F in the following and referred to as the interarrival distribution. not concentrated on {h. stating that U(t+a)U (t) ^ a. the distribution of Yo is called the delay distribution.r.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. The point process is called a renewal process if Yo. The number max k : Tk_j < t of renewals in [0.. are independent and Y1. Then Blackwell 's renewal theorem holds.} for any h > 0. 2h.T„_1). Y1.t.. . The mathematical representation is either the ordered set 0 < To < T1 < .1) (here U(t) = U([0. .U(t) is the expected number of renewals in (t.
e.EN(t) . but suffices for the present purposes . wee shall need the following less standard parallel to the key renewal theorem: Proposition A1.332 APPENDIX u(t) has limit 1/µ as t 4 oo. (A.3) Further.4) If F is spread. In 111. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. (A. IV). that z(u) has a limit z(oo) (say) as u 4 oo. Both result are valid for delayed renewal processes. Then Z(u) 4 z(oo). µF (A.i". oo). resp. (A.4) that z is Lebesgue integrable with limZ. see [APQ] Ch.1 if F is nonlattice and z (u) is directly Riemann integrable (d. Equivalently.a. z(x) = 0. IV). i.x)F(dx). in convolution notation Z = z + F * Z. ENt 4 1 lb Renewal equations and the key renewal theorem The renewal equation is the convolution equation Z(u) = z(u) + f where Z(u) is an unknown function of u E [0 . U Z(u . the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1.2) has the unique solution Z = U * z.out. z(u) a known function.2) Z(u) = J0 u z(x)U(dx). and F(dx) a known probability measure . (A. then Z(u) i f0 z(x)dx ..i. the statements being EN(t + a) . then it suffices for (A. u u PF 4 00.2).9. Under weak regularity conditions (see [APQJ Ch. stating that U(t)/t > 1/p.R.5) 2This condition can be weakened considerably . and that F has a bounded density2. Note in particular that F is spreadout if F has a density f.2 Assume that Z solves the renewal equation (A.
i. F(dx) = e7xF(dx).. Here the relevant F does not have mass one (F is defective).} be a renewal process. however. multiply (A. Z(u) U = 1 u 1 u f z(u .. the present more general definition is needed to deal with say Harris recurrent Markov chains. . of Yo. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. T1. 0 PF µF 11 In risk theory.e. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo.r. . is called the cycle length distribution and as before. that the existence of y may fail for heavytailed F.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). Assuming that y can be chosen such that f °° Ox F(dx) = 1.x)u(x) dx = z(u( 1 . or many queueing processes.. • . 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. a basic reason that renewal theory is relevant is the renewal equation II. The kth cycle is defined as {XTk+t}o<t<Yk . we let µ denote its mean. that F is a probability measure. . z(x) = e7xz(x). refer to the zerodelayed case.. A regenerative process converges in distribution under very mild conditions: .3) satisfied by the ruin probability for the compound Poisson model. Eo etc. Yk ). The simplest case is when {Xt} has i. {Tn} if for any k. .(3. equivalently. Y1 . This program has been carried out in III.t. Hence by dominated convergence.k+t }t>o is independent of To. Note. T1. .. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles).. the postTk process {XT. and its distribution does not depend on k. Tk and {Xt }o<t<Tk • For example. . Y2. We let FO. A stochastic process {Xt}t>0 with a general state space E is called regenerative w. However.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO).... To this end. Tk (or.d. However.5a. The distribution F of Y1. 1c Regenerative processes Let {T. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. cycles. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z.
. C). assume that p < 00 and define Un = ZT}1 .. If p = oo.t. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo.d. (b) If in addition Var(Ul ) < oo.334 APPENDIX Proposition A1. (A.. r.t.oo (i.. µ 0 If F is spreadout. This is the case considered in [APQ] V. are i.3.ZT }0<t<Y„+.4 Let {Zt}t^. cycles (we allow a different distribution of the first cycle).. An example is Zt = fo f (X8) ds where {Xt} is regenerative w.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. where the distribution of X.. Then {Zt}t^.6) id Cumulative processes Let {Tn} be a renewal process with i. We denote the limiting r.. {Tn}. {Tn} if the processes {ZT +t . C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p. and q(t) = sup It . oo). but in fact. i.t.'s by e.0 be cumulative w. Then {e(t)}. Otherwise .. Then Xt Di X.+ X. for n = 1. under the condition of Blackwell's renewal theorem.d. and we have: holds more generally that (rl(t).. just the same proof as there carries over to show: Proposition A1..0 is called cumulative w.Tk : t < Tk} as the age. Then it (ii.i. oo). 2. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t. [0. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. {Tn}.r.r. 0<t<Yi then Zt /t a$• EU1/µ. {i7(t)} are Markov with state spaces (0.e. then (Zt . fi (t) = inf {Tk .e.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1..v.i. in total variation. e(t )) .t : t < Tk}.ZTOI < 00.ZT Then: (a) If E sup I ZTo+t . then Xt ..r. resp . then e (t) .
. In IV. In the general case. use t E^(t)/t = E[Yo . ^) is given by the following four equivalent statements: (a) P (77 > x.. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y). U(x + 1) .5 Under the condition of Blackwell's renewal theorem.4. the first statement follows. Yo > 0] + f Eo^ (t .U(x) < U( 1)).^(t))} as a regenerative process. r. Hence for t large enough.i.y) = f U(t . (b) the joint distribution of (ri. and the equivalence of (a) with (b)(d) is an easy exercise. (d) the marginal distribution of ^ is FO. the sum is o(t) so that Eo£(t)/t + 0 . 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo .U(x) (c < oo because it is easily seen that U(x + 1) .APPENDIX 335 Theorem A1.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. EC(t)/t + 0.. are not i.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). For the second.v. (c) the marginal distribution of q is FO. Since the maximum Mn of n i. Y1i Y2. and the conditional distribution of ri given l. Y1 > t] 4 0. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 . we can bound e(t) by M(t) = max {Yk : k < 2t/p}.i. if in addition EYo < oo.d. .t. = z is Foz) The proof of (a) is straightforward by viewing {(r. W are independent. Hence t t lt ) = f U(dy)z(t .(t). the joint distribution of (rl. Proof The number Nt of renewal before t satisfies Nt/t a4' p. (1 V)W) where V. Then fi(t)/t a4' 0 and. 1) and W has distribution Fw given by dFw/dF(x) = x/pF.6 Consider a renewal process with µ < oo.y)P(Yo E dy) . but governed by a Markov chain {Jn} (we . Since z ( k) < E[Yi . l:) is the same as the distribution of (VW.d. assume first the renewal process is zerodelayed. Yl > t]. ^ > y) = 1 f +Y (z)dz.t. V is uniform on (0. we used: Proposition A1.
.and regenerative processes.. Jn +1=j} where J = a(JO. For example. . T_=inf{n>0: Sn<0}.t. G+(x) = P(S... distribution ofjXt}t>o itself where Pi refers to the case Jo = i. Further: Proposition A1.+ < x. = io for some arbitrary but fixed reference state io E E.jEE is a family of distributions on (0. r+ < oo).g. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE. in [APQ]. Notes and references Renewal theory and regenerative processes are treated. where the distribution of X. oo). X2. with common distribution F.d. Let X1. The semiregenerative process is then regenerative w. A Markov renewal process {Tn} contains an imbedded renewal process. < yIJ) = Fij( y) on {Jn= i.. G_(x) = P(ST_ < x. . . the semiregenerative process is called nonlattice if {T. IT. A stochastic process {Xt}t>o is called semiregenerative w. Sn = X1 + • • • + Xn the associated random walk.. .) and (Fij )i. Then Xt 4 Xo. .. Jn_1. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. oo).} is nonlattice (it is easily seen that this definition does not depend on i)...r. . . Alsmeyer [5] and Thorisson [372].t.T_ < oo). namely {Twk } where {Wk } is the sequence of instants w where Jo. 0] or (0 . the Markov renewal process if for any n. These facts allow many definitions and results to be reduced to ordinary renewal. e.r.i .. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj.7 Consider a nonlattice semiregenerative process.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. Yn. the conditional distribution of {XT„+t}t>o given Yo. . J1 i . be i..... Jn = i is the same as the P..}. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. Y1.. . Jo. and define r+=inf{n>0: Sn>0}.
0<j<m. F(A . . u .G+ * G_: (b) G_ (A) = f °° F(A .g. Sr_ _1 is at its minimum .=n w=m i Figure A.7) follows since G+(A) = 0 when A C (oo. the renewal measures U+=>G+. 0]).. A C (0. G_. 0] and (0. define w as the time where the preT_ path S1. In (A.S.r. (e) R_ = U+.7).x)R_ (dx). oo). n=0 The basic identities are the following: Theorem A2.T_=n} = {S. A C (oo. .7) (A. 0). G+.x)R+(dx). n 0 R_(A) = E I(Sn E A). (c) G+(A) = f °. m<j<n}.S. (A. F(A) + (G+ * G_)(A). n=0 n=0 00 00 and the T+.. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A). oo). S.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). F(A) is the contribution from the event {T_ = 1} = {X1 < 0}.1 . More rigorously. On {T_ > 2}. Proof Considering the restrictions of measures to (oc. (d) R+ = U_. >0.1 (a) F = G+ + G_ . oo) (A.>0. U.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. A C (0. A C (oo. we consider the last such time (to make w unique) so that {w=m. 0].=EGn..8) (e.
+ E du)P(S. ST_ E A) P(T+ = m. and reversing the order of summation yields P(T_ > 2.XnEAx) 00 f 0 f 0 00 00 1: F(A .>0. SnEAIS.m.0<k<ri . .. Sr_ E Adu) (s ee again Fig .du) (G+ * G)(A)• C llecting terms. Aso.3 8 APPENDIX Reversing the time points 0. .x)R+(dx). m < j <n. (A.Sn_1Edx.x)P(Sk < 0.._ E A .8) is similar.7) follows. m it follows (see Fig..= n..1).1) that P(Sj Sn. clearly (Sj Sm>0.+ E du) E P(S. 0 < k < n... m=1 f S mming over n = 2. SmEdu) = P(T+=m.. S. (b) follows from 00 G+ (A) _ E F(Sn E A.u) f0m m=1 n=m+1 00 J0 OO P(S. S. .1. 0<j<m. Sn1 E dx) n=1  F(A . A. and the proof of (A.F(r_n_mSrEA_u)._ = n . ST+Edu).. E du) = P(T_=nm._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu). A. ST_ E A . r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. It follows that for n > 2 F (7.3.
the analogue of a random walk is a process with stationary independent increments (a Levy process. being concentrated at 0. which is basic for the PollaczeckKhinchine formula. such developments motivate the approach in Chapter VI on the Markovian environment model.. it serves as model and motivation for a number of results and arguments in continuous time. 11. Since G+ is concentrated on (0.O<k<n. there is no direct analogue of Theorem A2.T+> n) = P(Sk < O..0+[s])(1 . and G+. WienerHopf theory is only used at a few places in this book. oo). u Remark A2.APPENDIX 339 and the proof of (c) is similar.4). Sk = X1 + • • • + Xk = Sn . In discrete time. . G_ are trivial.0<k<n. there are direct analogues of Theorem A2.1(a) is from Kennedy [228]. the survey [15] by the author and the extensive list of references there. In this generality of. However.f. G_ [s] are defined at the same time. Again. P(SnEA . and sometimes in a larger strip. For (d).G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}.1). The present proof of Theorem A2. and similarly H_ (s) = 1 .SnEA) is the probability that n is a weak descending ladder point with Sn E A. In continuous time. is based upon representing G+ as in (b).0<k<n.g.SnEA) = P(Sn<Sk. cf.SnEA) = P(SnSn_ k. consider a fixed n and let Xk = Xn_k+l.g. u Notes and references In its above discrete time version. a number of related identities can be derived. and the proof of (e) is similar. Nevertheless.O<k<n. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure).F[s] = (1 . then T+ = inf It > 0 : St = 0} is 0 a.6. Another main extension of the theory deals with Markov dependence. Then for A C (oo. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). this holds always on the line its = 0. we can rewrite (a) as 1 .P as a product H+H_ of functions with such properties. Summing over n yields R+ (A) = U_ (A).2 In terms of m. 0]. E. The classical analytical form of the WienerHopf problem is to write 1 .Sn_k.1. see e. if {St} is Brownian motion. the derivation of the form of G+ for the compound Poisson model (Theorem 11.s.9) whenever F[s].1.g.G_[s]) (A.SnEA) = P(Sn<Sk. 6+ [s]. For example.'s. see for example Bingham [65].
and eQ can then be computed as the mth power (by squaring if = 2).11) A f eAtdt = eA. ere A is the eigenvalue of largest absolute value. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step).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. Thus. _I 0 (A. 0 .10) d dteAt = AeAt = eAtA (A.13) henever A is a diagonal matrix with all diagonal elements nonzero. Eo Kn/n! converges rapidly and can be evaluated without p oblems. It is seen from Theorem VIII. 1. if m is s fficiently large. whereas there is no similar single established a proach in the case of matrix exponentials. Here are. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). three of the c rrently most widely used ones: xample A3. To circumvent this. 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).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. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting .5 that when handling phase type distributi ons. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). however .12) eA'AO = Ale AA (A.
]t)n (A. To this end. Zo = a (Z = QZ. 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. Zo = h). construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77.2 (UNIFORMIZATION) Formally.. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector.. The approach is in particular convenient if one wants eQt for many different u values of t.7t) n=0 n! u °O n Pn (to see this.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (.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}. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. The probabilistic reason that (A. 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.. vp be the corresponding left .. i. One then can reduce to p linear differential equations by noting that k = ZQ. 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 procedure consists in choosing some suitable i > 0. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(.4 (DIAGONALIZATION) Assume that Q has diagonal form.e.e. In practice. Here is a further method which appears quite appealing at a first sight: Example A3 . p different eigenvalues Aj i .APPENDIX 341 Example A3. some jumps are dummy in the sense that no state transition occurs ). .15) Then it is easily checked that P is a transition matrix . condition upon the number n of Poisson events in [Olt])  Example A3. .3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt. Let vi. Ap.. However .3 i (A. i.
the eigenvalue.... of largest real part is often real (say. hp..17) eQt = E e\`thivi = E ea:thi ® vi.. some cases remain where diagonalization may still be appealing.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. we can take H as the matrix with columns hl. hi have been computed. v5Q = Aivi. Complex calculus : Typically.16) (A.18) contains terms which almost cancel and the loss of digits may be disasterous.. under the conditions of the PerronFrobenius theorem). Then P P Q = > Aihivi = E Aihi (9 vi. say A = (Ai)diag. Example A3. i= 1 i=1 P P (A. we have an explicit formula for eQt once the A j. (A. and vihi ¢ 0. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. vi. D = ) 2 2 . The phenomenon occurs not least when the dimension p is large. i # j. and hence A2 is so because of A2 = tr(Q). Nevertheless.. (A.g H1.342 APPENDIX (row) eigenvectors and hl. hp the corresponding right (column) eigenvectors.18) Namely. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. Then vihj = 0. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. not all ai are real. There are. 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. i=1 i=1 Thus. however. and we may adapt some normalization convention ensuring vihi = 1. Qhi = vihi. say Al..
1) . where (A.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity.20) ir = q2 ql qi +q 2 9l +q2 (A.k1). However. Then 7r = (ir1 7r2 ) = a (q21 Al .q. u Example A3.7 Let 3 9 2 14 7 11 2 2 .APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. h2 = Thus. i. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . l ab (g12g21 + (A1  411) 2) = 1.19) Example A3 . Of course. k  C k2 ) =b ( A1 q 1 Q11 / where a . Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. b are any constants ensuring//Irk = 1. replacing ai by A2.e.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A. v2 and h2 can be computed in just the same way. The other eigenvalue is A = A2 = q1 .
satisfying AAA = A. APPENDIX x1 3/2 . (A. and a generalized inverse may not unique.23) .11/2 . They are most often constructed by imposing some additional properties . (AA+)' = AA+. for example AA+A = A. (A+A)' = A+A. A2 = 3/2 .22) Note that in this generality it is not assumed that A is necessarily square. Generalized inverses play an important role in statistics. 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... 2 2 1=ab(142+(1+2)2 ) = tab.5 .11/2 + 5 1.6. (A. A+AA+ = A+.344 Then D= 2+ 11)' 7 T4 2 =52. 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_.
E. . Assume that a unique stationary distribution w exists . are ordered such that Al > 0. Then for some b > 0. _ A.g.g.e.D + O(ebt). Here is a typical result on the role of such matrices in applied probability: Proposition A4..P + e7r ).eir ). and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' . and exists and is unique (see for example Rao [300]).I) (A.25) . and define D = (A .eir)1 = I .1Q = Q(Q .1 Let A be an irreducible intensity matrix with stationary row vector it.eir )1.= (I . Rather than with generalized inverses . one is also faced with singular matrices .23) is called the MoorePenrose inverse of A. most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. ( Q .ew.P + e7r)1 (here ( I . These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . = 0 where m < p is the rank of A.1 goes under the name fundamental matrix of the Markov chain). (I . (A... Am > 0.P). one then works with Q = (Q .. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . lt o eAx dx = te7r + D(eAt .e ® 7r)1. ... 0 01 In applied probability.APPENDIX 345 A matrix A+ satisfying (A. if A is a possibly singular covariance matrix (nonnegative definite). Am+1 = .24) = te7r .
h ® it reduces to hit in standard matrix notation. of (A. 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. the rows are proportional to it. resp. 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.s. the r.I) . (A. and in fact any rank 1 matrix can be written on this form.eir)eAt = eAt = A'(t). respectively. .h. Interpreting 7r.I)}. in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. (A.26) 2 = 2 e7r + tD . For example. it follows that h ® it is the k x m matrix with ijth element hi7rj .2 Let it be a row vector with m components and h a column vector with k components.346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4..91a(2) . the formulas involving O(e6t) follow by PerronFrobenius theory.DZ(ent . B'(t) = e7r + DAeAt = eir + (I .3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). Then A(O) _ B(O) = 0. .D + D2 + O(ebt).24).h.I)} dx. h as 1 x m and k x 1 matrices. o Finally. B(t) denote the l.e.27) Proof Let A(t). I. Equivalently. see below.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .I) (A.J {xe^r + D(e .s.2e7r . and the columns to h. Note that h ® it has rank 1.
and the number of such factors is precisely given by the relevant binomial coefficient. 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.50 6 7 6 4f 4.APPENDIX 347 Then A®B = 2 f 20.4 eA® B = eA ®eB.A9. C2 = h2 are column vectors.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.3f 4v/. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) . (A B)' = eA®B e! L 1=0 0 . (A.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2).5v/. if Al = vi. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.29) If A and B are both square (k1 = ml and k2 = m2).30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4. such a factor is Ak (&B 1k according to (A. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I . each of which is A ® I or I ® B.31) Indeed.(A. (A.4vf.29). if A ® I occurs k times.28) In particular. Using (A. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).3V8. A2 = v2 are row vectors and C1 = h1. then v1B1h1 and v2B2h2 are real numbers.31).3vV/72f 20.3v'6.
v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. (A. and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. first term on the r . k any column vectors. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. the same time. P(t) Yt(2) }.s. {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 . Let P8f P(Sl). Q(2). Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . p = P(1) ® {X }. A special case of Proposition A4. P(2). the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . Yt(2 ) }.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). { 1't(1) }. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .33) . Let further it. where transition matrix of the bivariate Markov chain {X n1). { On the other hand. represents ces Q( 1). X ) }. h. From what has been said about matrices of {Yt( 1).348 APPENDIX Remark A4.4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines .3 < 0 Lemma A4 . n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). we have P8 = Pal) ® p(2). in the definition (A.32). resp .6 Suppose that A and of B. independent Markov chains.I)(h ® k). 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 +.5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. Thus .
g.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. Here is the PerronFrobenius theorem.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). see e. ..APPENDIX 349 Proof According to (A. il. . . and appeal to (A. i. and the corresponding left and right eigenvectors v. n. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. That is. = j and atk_li. then IN < Ao for all A E sp(A). Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. h such that vh = 1.34) 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. [APQ] X. . h can be chosen with 3By this. (b) if in addition A is aperiodic.7 Let A be a p x pmatrix with nonnegative elements.3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B.. p there should exist io.. so that by asssumption A ® B is u invertible.29). and the corresponding left and right eigenvectors v. Then the eigenvalue Ao with largest real part is simple and real. h = e and v = 7r (the stationary row vector). . we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix... f o r each i. > 0 for k = 1.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. (A. h can be chosen with strictly positive elements. Similarly. which can be found in a great number of books. .. j = 1. we have AO = 1. . We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. ao). and if we normalize v. . in such that io = i. . then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. E (0.12).. Now note that the eigenvalues of A ® B are of the form a +.
Proposition A5.e. Ao). For example. To this end.35) for some p E (oo.350 APPENDIX strictly positive elements. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. but is an easy consequence of the PerronFrobenius theorem. let {Yti°i } be a Markov process with initial distribution a and intensity . The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. relate the eigenvalues of B to those of B via (A. I. if we normalize v.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. The next result gives a condition for asymptotical exponentiality.(ti)ding.(3. Corollary A4. the phasetype distribution B(a) with representation (.2). h such that vh = 1. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. note that we can write the phase generator T as Q . the analogy of this procedure with unformization. 10) and use the formula me at e Bt = e 00 Antn = e . Then for any (3. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ .8. the condition is that t is small compared to Q. we have A0 = 0. not only in the tail but in the whole distribution. Note that for an intensity matrix.n t AL n=0 n! (cf. Furthermore. Example A3.1. 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).. h = e and v = 7r (the stationary row vector).(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0).8 is most often not stated explicitly in textbooks.
)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. a .aE where 0 < e < 1). In addition to the asymptotic exponentiality. Hence O ((a) aa. {t Y( a) } v>0 .(a) > x .Yj(av) = j f .YQ(av) = j) Pi ( ci(a'V) > x. from which the phase process is terminated . t < (a). Let further V be exponential with intensity V and independent of everything else. in fact .1. We can assume that Jta) = Yt(°).g. and this easily yields a(x)/x a' 1/t*. a'/a + 1. 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}. Then a(a'V)/a (aV) a' 1.x (1 . = YQ(x). a' = a .a' + oo (e. 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. dx/ti] or not. J^O)_ = j) Pi (v(aaV) > x. We shall .bij) Hence the intensity matrix of { Ix} is (qij/ti)i. Conditioning upon whether { Yt} changes state in [0.2 Pi (c(a) > x. Since JJ(.9. v/ t. Proof Assume first ti > 0 for all i and let I. and that Yt(a) = Yat for all t. has a limit distribution: Proposition A5. and write Yt = Yt(1). By the law of large numbers for Markov processes . We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .APPENDIX 351 ((1) etc. Then {Ix} is a Markov process with to = Yo. J(()) _ = i) + at•x t tt' . it states that the state.jEE. prove a somewhat more general result which was used in the proof of Proposition VI. fo tY dv/t a$' t*.
. so is the geometric distribution. .5 Let B be discrete phasetype with representation (P.3 As the exponential distribution is the simplest continuous phasetype distribution.j) and initial distribution a.g.. However.4 Any discrete distribution B with finite support. is discrete phasetype. . Example A5.. Example A5. A distribution B on {1.Pe... = 0 for one or more i. let E and Pkj j=k1. K}. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]). See also Korolyuk. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case.zP)'p. (b) the generating function b[z] _ E' . 1 k=1 1 0 otherwise. k>1. say bk = 0.x k > K.. P. a)... Then: (a) The point probabilities are bk = aPklp. Et II I a(a^V) > x) at' .1 and A5. an easy modification of the argument yields finally the result for the case where t. zkbk is za(I . 2. with point probabilities bk = (1 .p)k1 p. these results are in the spirit of rare events theory for regenerative processes (e. Penev & Turbin [238]. so we shall be brief. Indeed.} is said to be discrete phasetype with representation (E. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. Then P is substochastic and the vector of exit probabilities is p = e .352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .. ' pk 0 k>1 11 Theorem A5. . the simplest discrete phasetype distribution: here E has only one element.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t.2 do not appear to be in the literature.. a = b = (bk)k=1.. (c) the nth moment k 1 k"bkis 1)"n!aP"p. Keilson [223]. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. > 0}. 2. u Notes and references Propositions A5. k = 1.
. U2.36) in blockpartitioned notation (where we could also write a as (a (1) 0)). as is seen by minor modifications of Example A5.6. Jt t > U1 + U2. initial distribution a and phase generator T. resp.a(2). and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U. a' . and hence the negative binomial distribution is discrete phaseu type.. resp.r + 1. (E(2).6 is the Erlang distribution Er which is the convolution of r exponential distributions. _ i E E(1) T(1) t(1)a(2) i E E(2) . Then the convolution B = B1 * B2 is phasetype with representation (E...7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. a. and a=1). Then {Jt} has lifetime U1 + U2 . 11 Example A5.a(1).6 (CONVOLUTIONS) Let B1.T(1)). . A. T= ( 0 T(2) ) (A.2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. B2 be phasetype with representations (E(1). T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). { Jt 2) } with lifetimes U1 . A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A. r .1 This corresponds to a convolution of r geometric distributions with the same parameter p.{ 0. The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r.T(2)).APPENDIX 353 5c Closure properties Example A5.
are i.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E. if U1. this means that a = (Oa(1) (1 . with common distribution and N is independent of the Uk and geometrically distributed with parameter p. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). Example A5.9 (INFINITE MIXTURES WITH T FIXED) Assume that a = a(°) depends on a parameter a E A whereas E and T are the same for all a. A reduced phase diagram is 0a(1) E(1) A .0)a(2) E(2) Figure A. To obtain a phase process for C.354 APPENDIX Example A5.4 .. a reduced phase diagram is f a E t Figure A. we need to restart the phase process for B w.a(2).0)a(2))). P(N = n) = (1 . resp. then C is the distribution of Ul + • • • + UN.p)pn1.i.d. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims. Thus.a(1). i E E(2) 0 T(2) =IT (in blockpartitioned notation. a mixture of more than two phasetype distributions is seen to be phasetype. (E(2).E) where a(°) = fAa(a)v(da). Equivalently.T(1)). and o'i Oa.3 In exactly the same way..'). Let B(") be the corresponding phasetype distribution.. Example A5. and consider B(") = fA B(a) v(da) where v is a probability measure on A. a.8 (FINITE MIXTURES) Let B1. B2 be phasetype with representations (E(1).0)ai2). a. Then the mixture B = 9B1 + (1 .37) (1) (1 . In risk theory. Then it is trivial to see that B(") is u phasetype with representation (a(").p)pn1B*n.p.T(2)). U2.O)B2 (0 < 0 < 1) is phasetype with representation (E. T) and C = EO°_1(1 . i E E(1) T 0 I (A. p at each termination.T.
.. be the point probabilities of a discrete phasetype distribution with representation (E. If U1 has a different initial vector. +UN 2.T + pta). If we replace x by a r. 12 (PHASETYPE COMPOUNDS ) Let fl. Equivalently. T(1) ® T(2)). are i. Note that this was exactly the structure of the lifetime of a terminating renewal u process. a. Example A5 .APPENDIX 355 and C is phasetype with representation (E. f2. To obtain a phase representation for C . Example A5 . U2.. Proposition VIII.2. j E F}. it follows by mixing (Example A5. T(2) ).v.d. cf..7. let B be a continuous phasetype distribution with representation (F.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . T). To see this. U2 be random variables with distributions B1. if B is defective and N + 1 is the first n with U„ = oo.TWWW). Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. a.a(1). a(1) ® a(2 ). { Jt2) } be independent with lifetimes U1. v. but the same T. For U1 A U2. say with distribution F. then U1 + • • + UN is zeromodified phasetype with representation (a. Minor modifications of the argument show that 1. X independent of U.9) that (U .X)+ is zeromodified phasetype with representation (E. of F. U2.°.x)+.°_1 f„ B*?l. Thus the representation is (E(1) x E(2). say v.aF[T]. E). resp. .. then C is the distribution of U1 + • • • + UN. then U1 +• is phasetype with representation (E.T) where F[T] = J0 "o eTx F(dx) u is the matrix m. Corollary VIII. T + ta. let {Jtl)}.. then Jy has distribution aeTx. 13 (MINIMA AND MAXIMA ) Let U1. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. It is zeromodified phasetype with representation (E. Example A5. P).T) if U is phasetype with representation (E. { 4 } as exit of {Jt}. a.. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). Indeed. (E(2). let the phase space be E x F = {i j : i E E.1. a(2). cf. with common distribution B and N is independent of the Uk with P(N = n) = f. i. . if U1.f. T + pta). B2 of phasetype with representations (E('). if {Jt} is a phase process for U. T) and C = F.g. v.2.aeTx. resp..
the initial vector is (a(1) (& a (2) 0 0).(bk)'.2) } to go on (on E(2)) when { i 1) } exits. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. Here are the details at two somewhat different levels of abstraction: (diagonal argument .n = I:pi(n)Er v ( __ ) n) ) a= 1 . i= 1 C.(Sn) with Sn = n/b.... see Neuts [269] (where the proof. Example A5. That is. Then we must find phasetype distributions Bn with B. By the diagonal argument (subsequent thinnings). the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support.(bk) + B(bk) for all k as n * oo. Now we can find first a sequence {Dm} of distributions with finite support such that D. say degenerate at b. Proof Assume first that B is a onepoint distribution. The general case now follows easily from this.} of phasetype distributions such that Bn 3 B as n + oo. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5. and the closedness of the class of phasetype distributions under the formation of finite mixtures.8. r # oo. relies more on matrix algebra than the probabilistic interpretation exploited here). and vice versa. there is a sequence {B. Hence it is immediate that Bn 4 B.(n) = D.. we can assume that ID. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . we need to allow { Jt.. and let Bn be the Erlang distribution E. however. cf.14 To a given distribution B on (0.. Then from above. any distribution B on (0. oo). elementary) Let {bk} be any dense sequence of continuity points for B(x).356 APPENDIX For U1 V U2. q(n) q(n) pi(n)a .B(bk) I < 1/n for n > k.xq(n)(n)}. with weight pi(n) for xi(n).(bk) + B(bk) for all k. Thus the state space is E(1 ) x E(2) U E(1) U E( 2).. Let the support of Dn be {xl(n). 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 ..
oo) and any fl. Corollary A5. u 2 (abstract topological ) The essence of the argument above is that the closure (w. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions.(x)Bf.t..D(bk)I < n.15 To a given distribution B on (0 . for some a < oo.i. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. PIT contains all finite mixtures of onepoint distributions..n (bk) . however. Hence G C PET and L = PIT. one would use the B given by some statistical fitting procedure (see below). 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. .B(bk )I < ..( dx) * f r f{(x)B(dx).. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. i = 1. 2. Then ICr( n ).d. then it is immediate that WI(B) = p2(B) for all distributions B on [0. It should be noted.e. replications). oo) approximation Assume that we can compute a functional W(B) when B is phasetype.n( b k ) . But To is the class G of all distributions on [0. f2. oo). k < n. i. compute W(B) and use this quantity as an approximation to cp(B0). we can then approximate Bo by a phasetype B. say on the claim size distribution B in risk theory. and that cp is known to be continuous.. the class CO of all discrete distributions. and we can take Bn = Cr(n). x 4 oo. if information on Bo is given in terms of observations (i. oo) * [0. In particular. u Theorem A5. oo) such that f (x) = O(e«x). If Cpl (B) and ^02(B) are weakly continuous.... For a general Bo. k < n.14 is fundamental and can motivate phasetype assumptions. Let E be the class of functions f : [0.n.r. E E.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). Since PET is closed under the continuous operation of formation of finite mixtures.
if f (x ) = e°x. i = 1.. i = 1...16 To a given distribution B on (0 . \\ 0 Corollary A5.. . Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A.(dx) > J fi(x)B(dx)..f (z) = f = 1 1 1 1n/ o .. f00 fi(x)Cr. Now returning to the proof of (A. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. we may assume that in the proof of Theorem A5.. liminf B..n(dx) + f 0 fi(x)Dn(dx).f ' f (x)B(dx). i=1. then cc f (x)Bn ( dx) = (?!c ) e'= . . i = 1. 2... .39) Indeed.14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx). f° xtBn(dx ) * f °° x`B( dx).f (x)B(dx). n B=az. oo). TO (A.2 . and the case of a general f then follows from the definition of the class E and a uniform integrability argument.38) We first show that for each f E E. . and hence we may choose r(n) such that L 9l) f (x)Cr(n). and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ).oo J fi(x)B..  APPENDIX B implies that 00 o o 00 n.n(dx) < 1+.38 ). n. for each i. By (A. n.39).. .358 Proof By Fatou' s lemma..
We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i ./3) is defined as the unique solution > 0 of B[y] = l+y/j3. . The adjustment coefficient is a fundamental quantity.> y for some sequence {ei} with ei E (0. oo) with B[y +e] < oo for some e > y = 7(B.18 In the setting of Corollary A5. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions. lim sup ryn < 7. .} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn. However. I. 0 as i * oo. O We state without proof the following result: Corollary A5. If ei > 0.3). 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. . . from a more conceptual .16.e. Proof Let fi(x) = el'r+E. there is a sequence {B. . . e ) and ei J. the adjustment coefficient 'y = 7(B. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1.. one can obtain 7(Bn. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. (N. For practical purposes. (N or a given distribution Bo. The present section is a survey of some of the available approaches and software for inplementing this.14 is classical. Notes and references Theorem A5.l3µb < 1.. but are certainly not unexpected. /3) = ry for all n. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. lim inf > is proved similarly.. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n .APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying .17 To a given /3 > 0 and a given distribution B on (0./3). .
and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). (N is the empirical distribution Be. we do not not want to perform matrix calculus in hundreds or thousands dimensions). [202]. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory.d. e . risk theory. and in practice this sets a limitation to the usefulness (the curse of dimensionality ..f. [70]) restrict attention to acyclic phase type distributions . d.'s). Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . the number of phases required for a good fit will typically be much larger.g. 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. three for a mixture of two Erlangs ). In a series of papers (e. Of course. a program package written in C for the SUN workstation or the PC is available as shareware. The characteristics of all of these methods is that even the number of parameters may be low (e. . g. giving mass 1 /N to each S=.f. The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter. [216] ).360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 .} of phasetype distribution such that Bo.g. and as fitted distribution we may take B. 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..g . Schmickler (the MEDA package. cf. A method developed by Bobbio and coworkers (see e. at a a number of selected points . The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . defined by the absence of loops in the phase diagram . reliability or queueing theory..g. It seems therefore a key issue to develop methods allowing for a more general phase diagram. we have constructed a sequence { B. and used a nonlinear programming approach . the L1 distance between the c . . for some suitable large n. [317] ) has considered an extension of this setup.. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm .
. then the estimators would be of simple occurenceexposure type. Thus.T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei . e. .T (n)(TiI(1. EN where ai = N 1 I ((k) = i) tii=i iEE. In fact. .g. one is lead to an iterative scheme. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations.. .. eieT(n)((k. (N) tJk Ea ( n). (n+1) _ Ea (n).. E.x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations....T(n) (Nik IC1.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j)... . Nii = = . .g.T(n) (Ti ^^ 1. the methods of [70] and [38] appear to produce almost identical results. it seems open whether the restriction to the acyclic case is a severe loss of generality. In practice.(N) = E Ea(n).APPENDIX 361 tion Sk was available. since this is parameterdependent. it is easy to see that N (k Ea(n). (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. jEEA. = j) f k=1 k =1 tE[0. N Ti = I(J= i) dt.
This page is intentionally left blank .
The Mathematical Scientist 14. Dover. Arfwedson ( 1955) Research in collective risk theory. Scand. [6] V. Asmussen (1987) Applied Probability and Queues.G. [5] G.). 101116. [3] J. J. Probab. B. Queueing Systems 10. The case of equal risk sums. Aktuar Tidskr. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. Teubner. Abate & W. 57. Asmussen (1984) Approximations for the probability of ruin within finite time . Borokov. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. AT&T. Adv. Chichester New York. [13] S. [12] S. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . Choudhury & W. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Queueing Systems 16. Stoch. John Wiley & Sons. [11) S. 363 . Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Abate & W. [4] M. Alsmeyer (1991) Erneuerungstheorie. 37. [15] S. pp. 14. 53100. 191223. ibid. New York. Arfwedson ( 1954) Research in collective risk theory. 311338. [2] J.Bibliography [1] J.). Transactions XVth International Congress of Actuaries. Aktuar Tidskr. II. Abramowitz & I. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . 38. with applications to risk reserve processes and the GI /G/1 queue. G. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Skand. [7] E. 143170.A. Stuttgart. 213229. Proc. New York. 1985. New York. Abate. Optimizations Software.L. 1984 . Appl. Preprint. ed. [10] K. Act. 587. Appl. Skand. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. [9] G. In: Limit Theorems and Related Problems (A. Queueing Systems 5 . [14] S. 20. 219229. 345368. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Asmussen (1985) Conjugate processes and the simulation of ruin problems. 3157 . [8] G. 253267.
772789.K. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. [26] S. 106119. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Extremes 1 . [28] S. Asmussen. Asmussen. Asmussen (1998c) A probabilistic look at the WienerHopf equation. eds. Scand. Asmussen & B. Proc. [18] S. Rolski & V. Act. 313341. Binswanger & B. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Marcel Dekker. [27] S. 2943. 37. 79102. 54. Andronov. [31] S. stationary distributions and first passage probabilities. 354374. Scand. J. Bernoulli 6. Statist. Act. Scand. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Probabilistic Analysis of Rare Events (V. J.). [22] S. Act. Asmussen (1999) On the ruin problem for some adapted premium rules. Asmussen & M. Asmussen & M. Stoch. Appl. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. [32] S. Hojgaard (2000) Rare events simulation for heavytailed distributions. 193226. . Th. Methods V Problems (J. 2149. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. [19] S.M. Probab. 96107. Advances in Queueing: Models. 297318. 1996. 189201. 69100. T. Asmussen (2000) Matrixanalytic models and their analysis. Ann. Random Processes 2.364 BIBLIOGRAPHY [16] S. 1989 . Chakravarty.). SIAM Review 40. Asmussen (1992b) Light traffic equivalence in single server queues. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. [25] S. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. L. Alfa & S. [23] S. Kalashnikov & A. A. 20. Appl. [21] S. K. Proc. Floe Henriksen & C. Asmussen (1989b) Risk theory in a Markovian environment. Riga Aviation University. 1999 . 25. 4966. Probab. Asmussen (1992a) Phasetype representations in random walk and queueing problems. 1936. 137168. Appl. Stoch. 2. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Boca Raton. Asmussen. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. 8. Ann. Stochastic Models 11. Appl. Frey. Astin Bulletin 27. Asmussen & K. 313326. [29] S. Asmussen & B. Ann. 555574. [33] S. 303322. Scand.S. J. Florida. eds.). MatrixAnalytic Methods in Stochastic Models (A. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. Dshalalow ed. [17] S. CRC Press. New York. 27. [30] S. Asmussen (1995) Stationary distributions via first passage times. Probab. [34] S. 315. J. [20] S. [24] S.
Scand. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. [44] S. Appl. Proc. Insurance: Mathematics and Economics 20. CRC Press. Res. Asmussen & D. Asmussen . [37] S. Oper. 410433. [49] S. J. Appl. Asmussen & C. 23. Probab. . 429466. [47] S. Res. Koole (1993). Asmussen & M. Florida. Wiley. B. [36] S. Asmussen & H. Adv. Asmussen.). Eng. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Probab. Boca Raton. [45] S. with applications to insurance risk Stoch. 365372. first passage problems and extreme value theory for queues. Johnson.A. Asmussen & T. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. [38] S. Asmussen & R. Opns. 20. [43] S. Sc. [46] S. [50] S. Insurance: Mathematics and Economics 10. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Advances in Queueing: Models. Asmussen & C. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Th. 421458.A. Asmussen & V. Example of excessoffloss reinsurance for an insurance corporation. Asmussen & G.Y. 11251141. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. Read eds. Methods 8 Problems (J. [40] S. [48] S. 64. Schmidt (1995) Ladder height distributions with marks. Dshalalow ed. J. Statist. 10. 30. Supplementary Volume (Kotz. Math. Taksar (1997) Controlled diffusion models for optimal dividend payout. (to appear). 913916. J. [39] S. 115. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Olsson (1996). Marked point processes as limits of Markovian arrival streams. Appl.Y. Probab. Perry (1992) On cycle maxima. Stoch. 58. Asmussen & K. Asmussen & C.M. 299324. Rubinstein (1999) Sensitivity analysis of insurance risk models.BIBLIOGRAPHY 365 [35] S. 19. 259274. Hojgaard & M. Adv. Appl. 736755. Kliippelberg (1996) Large deviations results for subexponential tails. Appl. Proc.). [42] S. [52] S. Asmussen & T. Asmussen & V. Schmidli & V. Inf. 31. Asmussen & R. Taksar (2000) Optimal risk control and dividend distribution policies . O. Nerman & M. H. Asmussen. Probab. [51] S. Appl. 422447. Prob 32 . Fitting phasetype distributions via the EM algorithm. [41] S. Stochastic Models 8. 103125. 120. Finance and Stochastics 4. Management Science 45. Sigman (1996) Monotone stochastic recursions and their duals. 105119. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Asmussen & S. Statistica Neerlandica 47. 419441. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue.
M. Probab. [70] A. Oxford. [69] P. Act. Athreya & P. Boogaert & A. Wiley. Appl. 1985 . [68] T. 7. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. Grandell (1985) An insensitivity property of the ruin probability. J. 77111. Cambridge University Press. Bjork & J. Act. Scand. [61] J. Appl. 1975. Ney ( 1972) Branching Processes . Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. J. J. 148156. [62] J. Bjork & J. Act. Wiley. 11811190. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Appl. Asmussen & J. Beekman (1985) A series for infinite time ruin probabilities. 1988 . Clarendon Press. Bingham. [66] N.B . [72] P. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. Actuaries 21. [56] B. Stochastic Models 10. C.R. SpringerVerlag. Probab. [58] O. [73] A. New York. Springer .H. 4148. [63] A. 129134. Insurance: Mathematics and Economics 4.T.366 BIBLIOGRAPHY [53] S. Adv. [54] K.J. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. Borovkov (1976) Asymptotic Methods in Queueing Theory. New York. 1974. J. Bobbio & M. [57] C. 9599. Scand. 221232. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. 610. Scand. Halsted Press. Trans. Scand.A. Beekman (1969) A ruin function approximation. J. Bingham (1975) Fluctuation theory in continuous time. Berman & R. Soc. [55] B. Insurance: Mathematics and Economics 6. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Act. Cox (1972) A low traffic approximation for queues. [67] T. [65] N. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . [60] J. 661667.A. 275279. Baker (1977) The Numerical Solution of Integral Equations. 705766. Scand. 832840. Berlin. J. Act.L. Billingsley (1968) Convergence of Probability Measures. Insurance: Mathematics and Economics 9. New York. 33.H. 1995. 169186. Bloomfield & D. Boogaert & V. SIAM. [71] P. J. Cambridge. 9. Teugels (1987) Regular Variation. Beekman (1974) Two Stochastic Processes. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. . de Waegenaere (1990) Simulation of ruin probabilities.H. [59] O. Probab.L. Goldie & J. [64] P. BarndorffNielsen & H. 190204.
Bowers . 105119. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Wiley. Proc. Stockholm. D. [82] H. Oper. Nonparametric estimators for the probability of ruin. Aut. 51. Burman & D .C. Jr. M. [85] W. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Cox (1955) Use of complex probabilities in the theory of stochastic processes. Jones & C. Qinlar (1972) Markov additive processes. [83] D.). New York. [89] M. 34. Itasca. verve. Springer. Stockholm. Control AC28. [80] F. Soc. [86] K. 14. 16. New York. J. [90] D.C. [91] H. H. Cottrell. Skandia Jubilee Volume. Chandy & M. Fort & G. Geb. Area Commun.A. Academic Press. 24. Simulation and Estimation. Z. Brockwell .BIBLIOGRAPHY 367 [74] O. [88] J. J. Fox & L. Resnick & R.W. [76] N. Hickman. Syst. Boxma & J. Appl. Chung (1974) A Course in Probability Theory (2nd ed. [75] O. 62.L.R. Cambr. NorthHolland. Reading. Gerber. 14331453. 127130. 2338. New York San Francisco London. Cohen (1982) The Single Server Queue (2nd ed. [78] L.U. [84] D.W.I. Cohen (1998) The M/G/1 queue with heavytailed service time distribution.L.A. Res. Adv. Amsterdam. [81] J. Illinois.W.J.Y. Cramer (1955) Collective risk theory. Breiman ( 1968) Probability. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. [77] P. 313319.R. Berlin. Biihlmann ( 1970) Mathematical Methods in Risk Theory. J. [87] E. The Jubilee volume of Forsakringsbolaget Skandia. Probab. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Philos. Schrage (1987) A Guide to Simulation. Bux & U. Tweedie (1982) Storage processes with general release rule and additive inputs . The Society of Actuaries . Goovaerts & F. [92] H. Sel. Croux & N. IEEE Trans. Reiser eds.L.R. B. . Insurance : Mathematics and Economics 9. Cramer (1930) On the Mathematical Theory of Risk. Bucklew ( 1990) Large Deviation Techniques in Decision. Queueing Systems 33. AddisonWesley. De Vylder (1986) Ordering of risks and ruin probabilities. II. Computer Performance (K. Insurance : Mathematics and Economics 5.) NorthHolland. 3539.L. Bratley. Boxma & J. [79] P. S. Herzog (1977) The phase concept : approximations of measured data and performance analysis. Wahrscheinlichkeitsth. Veraverbeke (1990).Y . Bell. [93] K. Tech. SpringerVerlag. 177204. IEEE J. 749763.M.). Burman & D.J. Broeck . 392433. Amsterdam. Nesbitt (1986) Actuarial Mathematics . 907920. 93121..
16. Statist. Martingales and insurance risk. English version published in Skand. J. 227229. 435436. [98] A. [99] R. Dembo & O. 114119. Jones and Bartlett. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . 7083 (1969). [103] F.M. T. [109] F. Embrechts (1989). Insurance: Mathematics and Economics 3.D. Delbaen & J. Insurance: Mathematics and Economics 4. Haezendonck (1985) Inversed martingales in risk theory.B. Appl. 3. Delbaen & J. [112] F. Insurance: Mathematics and Economics 7. Oper. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. 583602. New York. Suppl. 121131. Rolski (1991) Light traffic approximations in queues. Appl.P. Comp. [111] F. Scand. Math. Stochastic Models 5.. [100] A. Dempster. Editions de 1'Universite de Bruxelles. Stockholm. 1990 . Res. Boston. Oper. [104] F. Soc. Probab. J. [107] L. Csorgo & J. [96] D. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Daykin. . 88101. [95] M. Tidskr. Actuarial J. 181217. Chapman & Hall. Haezendonck (1987) Classical risk theory in an economic environment. J. Daley & T. Deheuvels & J. 3750. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. 85116.368 BIBLIOGRAPHY [94] M. Act. De Vylder (1977) A new proof of a known result in risk theory. [97] D. [102] P. [106] A. Goovaerts (1984) Bounds for classical ruin probabilities. 5771. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). J. 624628. De Vylder (1996) Advanced Risk Theory. [110] F.. Insurance: Mathematics and Economics 10. [101] C. De Vylder & M.L. Dassios & P. Insurance: Mathematics and Economics 6. 22. Daley & T. 17. Laird & D. Roy. [105] A. Scand. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. De Vylder & M. [108] F. Csorgo & J. 9 . Pesonen (1994) Practical Risk Theory for Actuaries. 27. N. Springer. Math. Pentikainen & E. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . Devroye (1986) Nonuniform Random Variate Generation. Math. Rolski (1984) A light traffic approximation for a single server queue. 135159. Aktuar. Zeitouni (1992) Large Deviations Techniques and Applications. Res. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. 201206. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm.
Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . Act. 191207. [123] E. Soc. Waters (1996 ) Reinsurance and ruin .C.C. 1975 . Dickson (1994) An upper bound for the probability of ultimate ruin.M.M. Act. [122] B. J. Duffield & N. Oper. Scand.G. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. Grandell & H. Hipp (1999) Ruin problems for phasetype(2) risk processes. Insurance: Mathematics and Economics 19.R.R. Insurance : Mathematics and Economics 22.J. [116] D. Dickson & J.U. [120] D. Dufresne. [118] D. Act. 2000 . Scand. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. J. 4962. Scand. Letters 7. Act. [121] D. with applications.U. J. Springer . and the amount of claim causing ruin. [119] D.J.S. [131] P.M.M. Insurance : Mathematics and Economics 7. [126] F. 118. 107124. Proc. J. Waters (1999) Ruin probabilities wih compounding. Dickson & C.BIBLIOGRAPHY 369 [113] D. [124] N. Scand. [117] D. Res. Dufresne & H.C. British Actuarial J. Act.M. Insurance : Mathematics and Economics 11. Dickson & H.C. Dickson (1995) A review of Panjer's recursion formula and its applications.C. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. Insurance : Mathematics and Economics 10.R. 1741. 6180. 1993 . H. Ruin estimates for large claims . 193199. 5159. Dickson & H. 147167. Gerber & E. [115] D. Dufresne & H. Emanuel .M.C.C. [129] D . Act. Embrechts . [125] F. 1. [130] P. 3745. Philos. Insurance: Mathematics and Economics 25. Scand. Scand.W. van Doorn & G. Djehiche (1993) A large deviation estimate for ruin probabilities. Gerber (1988) The surpluses immediately before and at ruin. Astin Bulletin 21. 269274. 363374. Act.M.U. Regterschot (1988) Conditional PASTA.K. 177192. 251262.C. 105115. . 1994 . 131138. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. [127] F. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. Dynkin (1965) Markov Processes I. Insurance: Mathematics and Economics 7. Harrison & A.R. J.C. 1993 . Dickson & C. 174186. Embrechts ( 1988 ). [128] E. J. Camb. Berlin Gottingen Heidelberg. J.M.M. Hipp (1998) Ruin probabilities for Erlang (2) risk processes. 4259. 229232. Math. Dickson & J. Scand. [114] D. Dickson (1992) On the distribution of the surplus prior to ruin .M. Shiu (1991) Risk theory with the Gamma process.C. J.B. J. 1984 . 1984 .
Adv. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Trans. Appl.L. Scand. New York. New York. Embrechts & H. Appl. Teugels (1985).D. Embrechts. Schmidli (1994) Ruin estimation for a general insurance risk model. [140] J. Fields 75. Astin Bulletin 16.M. Embrechts & J. Adv. Fuh & T. first passage times and moments of ladder variables in Markov random walks. .W. [135] P. Appl. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . Tidsskr. [139] A. On the excursions of Markov processes in classical duality. Embrechts. 5572. 5974. 14661474. 38. Springer. Esary.). Statistica Neerlandica 47. [150] H. 181190.K. J. Insurance: Mathematics and Economics 7. Konig. [149] C. 623637. Approximations for compound Poisson and Polya processes. Erlang 2. C. Mikosch (1991). Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. Wiley. A bootstrap procedure for estimating the adjustment coefficient.K. Fitzsimmons (1987). Kliippelberg & T. [148] C. Danish Academy Tech. Skand. 695712. Probab. Maejima & J. Jensen. Probab. [147] M. 17. Wiley. Esscher (1932) On the probability function in the collective theory of risk. J. Adv. Insurance: Mathematics and Economics 10. Furrer (1998) Risk processes perturbed by astable Levy motion. Arndt & V. 566580. 269274. 35. Probab. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. M. F. Ann.370 BIBLIOGRAPHY [132] P.L. 175195. Probab. [141] F. [144] P. Schmidt (1992) An insensitivity property of ladder height distributions. 616624. Franken. Math. [134] P. 1998 . Frees (1986) Nonparametric estimation of the probability of ruin. Sci. Appl. D. Appl. [136] P. Act. Probab.). [133] P. Akt. 29. 159178. R. 29. Proschan & D. Probab. Insurance : Mathematics and Economics 1. [138] P. [145] P. 131137. Heidelberg. Embrechts & T. Statist. U.D. Nyt Tidsskrift for Matematik B20. with applications. [143] W. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks.J. Walkup (1967) Association of random variables. Rel. Villasenor (1988) Ruin estimates for large claims. Griibel & S. 8190. [142] W. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. Embrechts. Frenz & V. 5975.A. Embrechts & N. J. Wiley.W. Lai (1998) Wald's equations. Th. [137] P.D. J. Schmidt (1982) Queues and Point Processes. 404422. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. [146] E. 26. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . 3339.
E. Birkhauser . Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Goldie & R. London. de Vylder & J. Mitt. 3752. 105115. Insurance : Mathematics and Economics 7. [156] H. Furrer & H. Gerber (1988) Mathematical fun with ruin theory. [166] M.71.U. Appl. Ver. [168] J.M. Haezendonck (1984) Insurance Premiums. [157] H. Huebner Foundation Monographs. Furrer.U. [155] H. Vers.. Vers. Ver. Mitt. Kou (1995) Limits of first passage times to rare sets in regenerative processes. Weron (1997) Stable Levy motion approximation in collective risk theory.U. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. [161] P. Griibel (1996) Perpetuities with thin tails. Goovaerts. [162] P. 1977. 131156. NorthHolland. Probab. 97114. Grandell (1977) A class of approximations of ruin probabilities. . Amsterdam. 28.G. 205216. F. Act. Scand. Act. Gnedenko & I. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. 6370. J. J. Schweiz. A. R.U. Scand. Michna & A. S. Gerber (1979) An Introduction to Mathematical Risk Theory. Math. Z. Appl. 73.). Insurance: Mathematics and Economics 20.U. Glynn & W. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. [154] H. Act. 1978 . ETH Zurich. Adv. 5. University of Pennsylvania. contained in the authors PhD thesis. Basel. Suppl. NorthHolland. Glasserman (1991) Gradient Estimation via Perturbation Analysis.U. 1523. Math.). Gerber (1986). Amsterdam. Springer. Kaas. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. 31A. [163] P. [159] H.. In Studies in Applied Probability (J. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Skand. J. 424445. Kluwer. Probab. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. eds.U. Galambos & J. Glasserman & S. Ann. 2336. 7778. 1981 . Life Insurance Mathematics. van Heerwarden & T. [164] B.J. Schweiz.W. Scand. [158] H. [167] C.J.S. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. Gani. Aktuarietidskrift 1970 . Unpublished. Goovaerts. J. [152] H. Probab. Dordrecht.N. 205210. [165] M. Insurance: Mathematics and Economics 15. Boston.V. Appl. Bauwelinckx (1990) Effective Actuarial Methods. 463480. [160] H.BIBLIOGRAPHY 371 [151] H. [153] H. [169] J. Gerber (1973) Martingales in risk theory.
ed. Applied Statistics 40. 1. Segerdahl (1971) A comparison of some approximations of ruin probabilities. Astin Bulletin 29. SpringerVerlag. Math. . Harrison & S. [180] R. Aktuar Tidskr. [185] J. [179] R. Stoch. [178] B. mathematische Wirtschaft . Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. E. Stoch. [186] J. M. Skand. Griibel & R.O. Grandell & C. 566576.131135. Grandell (1990) Aspects of Risk Theory. Harrison & S. Liet. 54. [174] J. Gyllenberg & D. Lemoine (1977) Limit theorems for periodic queues. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. Res. Opns . Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. 14. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. Probab. 5 . 7590. [182] A. 332.J. [189] A. [176] B. [177] B. Tinbergen Institute Research Series 20. [187] J. Statistical Algorithm 265. [184] H . M. Harrison & A. Opns . Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . Res.. [183] M. Gut (1988 ) Stopped Random Walks . [171] J. Matem. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . SpringerVerlag. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . 167176. 3 . Appl. 197214. 143158. Proc . Insurance: Mathematics and Economics 26. 3041. 6779. Grandell (1997) Mixed Poisson Processes. Rink. Appl. 33. M. Theory and Applications.). KTH. 14. Grandell ( 1992) Finite time ruin probabilities and martingales . 5766. Appl.I. Harrison (1977) Ruin problems with compounding assets . Grandell (1999) Simple approximations of ruin functions. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . Appl. Ibragimov. Informatica 2. Archiv v. Chapman & Hall. 355365. Th. Resnick (1977) The recurrence classification of risk and storage processes . Probab. Grigelionis (1996) Lundbergtype stochastic processes.V. Grandell (1979) Empirical bounds for ruin probabilities. Math. Proc. 8. [181] D. Preprint.S. [175] J. M. Grnbel (1991) G/G/1 via FFT. [188] J. [172] J.372 BIBLIOGRAPHY [170] J. [173] J. Winter School on Stochastic Analysis and Appl .A. Berlin. 347358. AkademieVerlag.243255. J. Amsterdam. Proc. Gusak & V. New York. Probability Theory and Mathematical Statistics (I.I.and Sozialforschung 6. 400409.
[191] W. Probab. 377381. Proc. Statist. . Verlag Versicherungswirtschaft e. [198] C. 8996. [201] L. [197] C.V. Huskova. J. Nerman (1992) EMPHT . J. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. Probab. [200] M.B. Iglehart (1969) Diffusion approximations in collective risk theory. The ruin problem for finite Markov chains. [196] C. [205] T. 12981310. Studies in Statistical Quality Control and Reliability 1992 :4. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. S.V. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Horvath & E. 3. Scand. 285292. [202] 0. Versicherungswirtschaft. Iversen & L. [195] B. No. 166180. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. 1990 . Wahrscheinlichkeitsth. Hojgaard & M. Geb. Haggstrom. eds. Appl. 378389. Prob. verve. [204] T. 123151. 11631174. 18. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities.). Klugman (1984) Loss Distributions. Chalmers University of Technology and the University of Goteborg. Appl. Ann.L. Hoglund (1991). Probab. [199] R. Foruth Prague Symp.A.BIBLIOGRAPHY 373 [190] P. [206] B. 8095.R.L. [207] D. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Z. Ann. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. 6. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. 305313. Act. 4385. Hipp (1989a) Efficient estimators for ruin probabilities. Karlsruhe. J. Karlsruhe. [193] U. [208] V. Scand. Heilmann (1987) Grundbegriffe der Risikotheorie. Hipp & R. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve.a program for fitting phasetype distributions. Insurance: Mathematics and Economics 5. Insurance: Mathematics and Economics 5. [192] U. 29. 285293. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. [203] T.. 259268. Nachrichten 30. ACM TOMACS 6. Mathematical Statistics. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. on Asymptotic Statistics (P. Math.M. Hogg & S. Wiley. 25. 23. Hill (1975) A simple general approach to inference about the tail of a distribution. Mandl & M. 1998 . 5770. Astin Bulletin 19. Act. J. Ann. New York. [194] 0. Electronic Letters 35. 19. Asmussen & 0.
J. Insurance: Mathematics and Economics 5.M. Jelenkovic & A. 561563. Act. Clarendon Press. Kaas & M. Wishart (1964). J. 283306. Karlin & H.H. [222] S. 3. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. [221] E. 31. NEC Research Institute. Statist. Jensen (1953) A Distribution Model Applicable to Economics. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes.P. J. ibid. 60. [213] P. Appl. [225] J. Scand. 6.M.374 BIBLIOGRAPHY [209] D. Proc. N. Appl. Cambridge Philos. Janssen & J. [224] J. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Jagerman (1985) Certain Volterra integral equations arising in queueing. 63. Academic Press. Keilson & D. Taylor (1981) A Second Course in Stochastic Processes. 711743. Probab. [220] V. Chicago. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications.B. N. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application.G.R.G.L. New York. Keilson & D. ibid. 15. Kao (1988) Computing the phasetype renewal and related functions. Goovaerts (1986) General bound on ruin probabilities. Munksgaard. 866870. Ann. Princeton. 4151. 37. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. Technometrics 30. [219] V. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. [215] J. [223] J.M. Taaffe (1989/90) Matching moments to phase distributions. [217] R. 173190. 239256.J. Stochastic Models 1. Proc. 325247. Jensen (1995) Saddle Point Approximations. Addenda to for processes defined on a finite Markov chain. Keilson & D. [226] J. 165167.A. J. [218] V. Soc. 547567. 35. 118. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. Math.J. Stochastic Models 5. Copenhagen. [216] M. Proc. Astin Bull.G. 6. Soc. [227] J. 8793. Astin Bull. 187193. [212] J. 61.C. Soc. Kalashnikov (1996) Twosided bounds of ruin probabilities. Amer. [214] A. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. Kemperman (1961) The Passage Problem for a Markov Chain. Cambridge Philos. 259281. University of Chicago Press. [228] J.M. Cambridge Philos. 1996. [210] D. Act. [211] J. Manuscript. Probab. Oxford.M. 116129. 123133. Johnson & M. 11. Kluwer. .
Appl. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. 48. Lehtonen & H. 390397. Klnppelberg & T. Act. Kliippelberg (1988) Subexponential distributions and integrated tails. [246] D. Macmillan. J. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. J. [245] T. [247] L. Scand. Math. 26. Actuarial J. Kliippelberg & U. . Scand.S. 4477. 18. [239] H. Seminaire de Probabilties X . Mikosch (1995) Modelling delay in claim settlement. Penev & A. [243] A. J. 259264. [242] A. Insurance : Mathematics and Economics 12. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. Kingman (1964) A martingale inequality in the theory of queues. B24. Cybernetica 4. 125147. Act.BIBLIOGRAPHY 375 [229] J. [234] C. Soc. Oxford 12. 146. 283284. SIAM. Soc. Statist.P. Appl. SpringerVerlag. [244] T. 277289. [241] G. [240] U. Lemoine (1981) On queues with periodic Poisson input. [237] C.C. Lindley (1952) The theory of queues with a single server.F. Probab. (235] C. 24. J.F. Scand. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . Philos. Lipsky (1992) Queueing Theory . 4958. Soc. Cambr. Insurance: Mathematics and Economics 8.. Prob. Probab.J. Roy. Appl. Kunita (1976) Absolute continuity of Markov processes. [236] C. [230] J.F. 359361. Bernoulli 1 . Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. Kingman (1961) A convexity property of positive matrices . 383392. Korolyuk. Quart. New York.a Linear Algebraic Approach. J. Klnppelberg & T. 889900. Proc. Sorensen (1997) Exponential Families of Stochastic Processes. Adv. SpringerVerlag. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. Lehtonen & H. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. [231] J. 132141. Kuchler & M. J.F. Lecture Notes in Mathematics 511. 6075. 133135 (in Russian). 25. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. 60 . [233] C. 1998 . I. Quart. Probab. Stochastic Models 7. 1995 . Appl.J. 154168.C.C. Kingman (1962) On queues in heavy traffic. Proc. Latouche & V. [232] C. Philos. Camb. 858874. [248] D. J. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. 279285. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. [238] V.
Probab. 15. Gut & J. Appl. 223235. [263] M . Scand. [258] K . 124147. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . J. Act. Ann. MartinL3f (1986) Entropy. Appl. Anal. J. [260] H.S. 2939. Act. Proc. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . 58 . Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Probability and Mathematical Statistics (A. Appl. [265] S. Miyazawa & V. Almqvist & Wiksell. Probab. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. 268285. 429448. Math. Probab. [253] V. Makowski ( 1994) On an elementary characterization of the increasing convex order . Proc. K. J.M. Stoch. [259] Z . Schmidt ( 1993) On ladder height distributions of general risk processes . Adv. Cambridge Philos. Insurance : Mathematics and Economics 5. J. Englunds Boktryckeri AB. 1986 . [266] P. 161174. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. Act. 32. Appl. Lucantoni . [257] A. Michna ( 1998) Selfsimilar processes in collective risk theory.K. Europ. MartinL6f (1983) Entropy estimates for ruin probabilities . J. 147149. 370377.V. Ney & E. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. Uppsala. Nummelin (1987) Markov additive processes I. 22. 1261 (0?)1270. 1996 . 676705. 11 . Scand.K. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . with an application . Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. MeierHellstern & M . Soc. 36. Soc. Res.). II Aterforsdkring av Kollektivrisker . Probab. Scand. Eigenvalue properties and limit theorems . [256] A. 378406. Appl. J. J. Th. Math. 286298. Stockholm. F. Probab. 763776. [251] F. 834841. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. [252] A.D.F.D.376 BIBLIOGRAPHY [249] D. 58.V. Ann. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Statist. [261] H. . [254] V . . 31. Probab . Cambridge Philos. 3. [255] V. 2. [262] H . Ann.S. [264] G. 4859. Nagaev (1957) Some limit theorems for stationary Markov chains. a useful concept in risk theory. 1994. Opns. Hoist eds. 29 . Meier ( 1984). 561592.D. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Appl. [250] F.
Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . [284] J. Proc. 135148. Stoch. 123144. New York. [272] J . Probab. Probab. 21. Sem. Naval Research Logistics Quarterly 25. [273] R. 75. 311342. 339353. J. [268] M . Commun. IEICE Trans. Abh. [269] M . 16. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. [275] H.F. Paulsen & H. . Marcel Dekker. Astin Bulletin 5 .F. Appl. 12. [282] J . Adv. [270] M . Proc. [281] J. Appl. [277] J. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. Johns Hopkins University Press. 249266. 316. Stoch.a survey. Math. Proc. Models 6. Probab. [271] M. Paulsen & H.F. Insurance: Mathematics and Economics 22.F. Norberg ( 1990) Risk theory and its statistics environment . 71. Appl.K. Appl. 764779.F. 273299.K. Nyrhinen (1999) Large deviations for the time of ruin. Baltimore . [283] J. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. 157. J. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. Statistics 21. Neuts (1992) Models based on the Markovian arrival process. E75B . [274] H .BIBLIOGRAPHY 377 [267] M . Omey & E. Stoch. J. Appl. Pakes (1975) On the tail of waiting time distributions. Paulsen (1993) Risk theory in a stochastic economic environment. [276] C. Probab. 12551265. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. 10081026. 46. Hamburg.A. Stochastic Models 3 . Insurance: Mathematics and Economics 20. Neuts (1977) A versatile Markovian point process. [278] E. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Paulsen (1998) Ruin theory with compounding assets . 215223. [279] E. 36. 445454. London. 30.G. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. [285] J. 327361. Appl. [280] A. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. O'Cinneide (1990) Characterization of phasetype distributions. 555564. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Neuts ( 1978) Renewal processes of phase type. Omey & E. 733746. Appl. Proc. Stoch. Appl. Stoch.
2330. 965985.M.G. [306] T. [297] R. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Res. Rolski (1987) Approximation of periodic queues. Wiley. Probab. Appl.G. 757764. Ann. [300] C. 61.U. [299] C. Pellerey (1995) On the preservation of some orderings of risks under convolution. Prabhu (1980) Stochastic Storage Processes. 117133. Paulsen & H. Appl.U. Scand. Ann.K. Soc.M. 222261. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. Appl. Ann. Ann.J. Math. [303] S.378 BIBLIOGRAPHY [286] J. [292] S. Adv. 19. J. Probab. 2343.J. [302] J. Prabhu (1965) Queues and Inventories. Heidelberg. 691707. 32. Insurance Risk. 568576. New York. 4.U. 215246. . 390413. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services.K. 11. Pyke (1959) The supremum and infimum of the Poisson process. Insurance: Mathematics and Economics 3. Philipson (1968) A review of the collective theory of risk. 139143. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Probab. J. Oper. [294] N. Opns. Queues. [287] F. Griibel & P. 337347.A. Insurance: Mathematics and Economics 16. 28. Probab. [288] S. Math. Wiley. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Inst. Pitts. [305] C. Wiley. Appl. Rao (1965 ) Linear Statistical Inference and Its Applications. Regterschot & J. [295] N. Math. On the ruin problem of collective risk theory. [304] B. Skand. Prabhu (1961). Probab . Statist.M. Adv. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. 537555.H. Embrechts (1996) Confidence bounds for the adjustment coefficient.U. Math. [298] V. Berlin. 12. Springer. 46. 29. [296] N. 465483. J. Prabhu & Zhu (1989) Markovmodulated queueing systems.R. 4568. Adv. Queueing Systems 5.. Aktuar. Stat. R. 235243. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Appl. Pitman (1980) Subexponential distribution functions. 820827.L. [289] C. 30. 29A. [290] E. [293] N. 147159. Statist. Tidskr. Math. 1989 . Austr. Ripley (1987) Stochastic Simulation . Gjessing (1997c) Ruin theory with stochastic return on investments. [301] G. Act. [291] S. Astin Bulletin XIV. New York. Res. Resnick & G. and Dams.
Proc. Schmidli ( 1999b) Perturbed risk processes: a review . [319] H. [318] H. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. 5. Ann. Act. Appl. Adv. 262286. Wiley. 155188. Seal ( 1969) The Stochastic Theory of a Risk Business. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. [308] S. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. [310] R. Stochastic Models 10.Y. 21. [326] H. Insurance: Mathematics and Economics 16. [321] H. . Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna).J. Rubinstein (1981) Simulation and the Monte Carlo Method. [315] T. Scand. Teugels (1999) Stochastic Processes for Insurance and Finance. 5. J. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. 2001 . Melamed (1998) Modern Simulation and Modelling. Wiley. Rudemo (1973) Point processes generated by transitions of a Markov chain. Rubinstein & A. 7. Scand. 795829. J. [314] T. [309] H. Math. 121133. 4857. H. Probab. Stochastic Models 10 . 227244. 131156. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. Stochastic Models 6. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . 135149. Rolski. Schmidli. Statist. Singapore. Data Anal. 431447 [316] S. [322] H. 417421. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . Operationsforsch. [313] M. Schlegel ( 1998) Ruin probabilities in perturbed risk models. Th. [317] L.M. [311] R. [324] H .Y. Schmidli ( 1996) Martingales and insurance risk. Probab. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. Appi. 11. 5. Appl.Y. Science Culture Technology Publishing . 145165. Wiley. Schmidt & J. Comp. Rubinstein & B. 687699. Stoch. 93104. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Wiley. J. V. [320] H. Statist. Insurance: Mathematics and Economics 22. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . 4068. Astin Bulletin 29. Act. 1997. [312] R. [323] H.BIBLIOGRAPHY 379 [307] T. 365388. Probab .L. Wiley. [325] H . Rossberg & GSiegel (1974) Die Bedeutung von Kingmans Integralgleichungen bei der Approximation der stationiiren Wartezeitverteilung im Modell GI/C/1 mit and ohne Verzogerung beim Beginn einer Beschiiftigungsperiode.
11. Chapman & Hall. Scand. Ann. In Probability and statistics .L. [335] B . Verein Schweiz Versich. 383413. Verein Schweiz. Aktuar Tidsskr. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. Grenander). [336] B. Mitt. J. Adv. Adv. Seal (1974) The numerical calculation of U(w. 483487. Seber (1984) Multivariate Observations. [342] D. Shaked & J. 279299.380 BIBLIOGRAPHY [327] H.W. [345] D. Shwartz & A. Sengupta (1990) The semiMarkov queue: theory and applications. 2236. Verein Schweiz. [347] K. Seal (1978) Survival Probabilities. 159180. Appl. . [328] H. Act. Mitt. 121139. Siegmund (1979) Corrected diffusion approximations in certain random walk problems.S.F. 4383. 171178. 243249.S. Math. Math. Appl. 77100. Versich. Insurance: Mathematics and Economics 8. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. 191197. 1987.W. 1955. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. [333] C: O. [348] K . Statist.the Harald Cramer volume (ed. Queueing Systems 11. Appl.A.L. 4. 673684. Act. t). pp.G. Chapman and Hall. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Stockholm. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. 914924. Risk teory and the single server queue . 21. [334] C: O. Segerdahl (1959) A survey of results in the collective theory of risk. Th. [337] M. t). L. Siegmund (1975) The time until ruin in collective risk theory. [341] E. Wiley. 701719. Math. Probab. 72. 11.L. 75. 4. [338] E. [330] H. New York. 157166. Mitteil. Seal ( 1972). Probab . Versich. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. Sigman (1992) Light traffic for workload in queues. 1974. [329] H. 61. Wiley. Ann. . Academic Press. the probability of nonruin in an interval (0. Shantikumar (1993) Stochastic Orders and Their Applications. Stochastic Models 6. [346) D . Probab. 72. [344] D. J. [343] D. 429442. Almqvist & Wiksell. Aktuar Tidsskr. Shin (1989) Ruin probability by operational calculus. Scand. Shin (1987) Convolution of uniform distributions and ruin probability. [339] A. Skand. SpringerVerlag. [332] C : O. Probab. Siegmund ( 1985) Sequential Analysis . Weiss (1995) Large Deviations for Performance Analysis. [340] E. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen .S. [331] G. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests.
. Slud & C. Straub (1988) NonLife Insurance Mathematics. Hoesman ( 1989) Moderate. Thorin (1977). John Wiley & Sons. 1978. 65102. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. Sundt & J.L. New York. Scand. Daley ed. 449461. Astin Bulletin 16 . Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. [353] E. Springer. [361] G. 5776. 1980 . [356] B . 1982 . 235254. [365] J. Karlsruhe. Insurance: Mathematics and Economics 1. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Insurance: Mathematics and Economics 16. . [357] B. Smith ( 1953) Distribution of queueing times. 2739. [369] O. Astin Bulletin 12. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. 163175. Act.C.C. Teugels ( 1995 ) Ruin estimates under interest force . [366] O.L. [351] D .).S. [350] W. Scand. 149162. Astin Bulletin 24. [358] R. Insurance: Mathematics and Economics 19. [363] G. Teugels (1982) Estimation of ruin probabilities . Taylor (1986) Claims reserving in NonLife Insurance. [360] G. [368] O. [359] L. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . [354] B. Sundt & J. Actuarial J. Act. 1974. J. [355] B . J. Sundt & W. 114137. 21. 65102. NorthHolland. 8199.V.C. 1976.. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions .J.A.BIBLIOGRAPHY 381 [349] E. J. Scand. Scand. 197208.). Cambridge Philos. [352] D. Stanford & K.and largedeviations probabilities in actuarial risk theory. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. Proceeedings of the IEEE 77.L. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . New York. Thorin (1982) Probabilities of ruin. Act. Unpublished manuscript. Proc. Act. 722. Wiley. Adv. [364] G. Act. 725741. [362] G.C. J. Taylor (1979) Probability of ruin with variable premium rate. J. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . Scand. [367] O. Soc 49.J. Ruin probabilities prepared for numerical calculations. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. 118.C. Scand. Appl.L. Verlag Versicherungswirtschaft e. 8594. Probab. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue.
Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Statistical Science 10. [382] G. Wald (1947) Sequential Analysis. [376] A. Res. Stationarity and Regeneration. R. 4963. [378] H. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. M. Probab. [379] M. M. Astin Bulletin VII. Willmot & X.). Dordrecht Boston Lancaster.E. 37. Act. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Lin (1994) Lundberg bounds on the tails of compound distributions.382 BIBLIOGRAPHY [370] 0. Astin Bulletin IX. 936952.E. J. PrenticeHall. Insurance: Mathematics and Economics 15. Scand. Haezendonck eds. 5762. Taqqu. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. [377] V. Reidel. [384] R. D. Wolff (1990) Stochastic Modeling and the Theory of Queues. Oper. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. SpringerVerlag. 31. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. Tidskr.R. Van Wouve. thesis. De Vylder. Thorin & N. 142. [372] H. 231246. Sherman & D. Thorin & N. Unpublished Ph. [380] W. 1942. [373] S. 148164. F. Thorisson (2000) Coupling. 743756. [371] 0. In: Premium Calculation in Insurance (F. Submitted. 1983 . . Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Wiley. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. University of Michigan. Willinger. [383] G. Goovaerts. [374] F. 137153. Waters (1983) Probability of ruin for a risk process with claims cost inflation. Appl. Tacklind (1942) Sur le risque dans les jeux inequitables. [381] W. Insurance : Mathematics and Economics 13. 6785. Aktuar. [375] N. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements.W. De Vylder & M. J. J. Skand.
203. 308. 226.346349 383 . 4851.185187. 201 Brownian motion 3 .121129. 341.287292. 117128.307312 compound Poisson model 4. 5.359 aggregate claims 103106. 3334. 7879. 9496.203. 39. 117127 corrected 121127 duality 1314. 7179. 138139. 245248. 170173. 196201 inverse Gaussian distribution 76. 2526.249250 integral equation 16 Lindley 143 renewal 64. 1415. 15. 3839.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.4447. 14. 248 WienerHopf 144 interest rate 190. 3436. 37. 283. 110113.269. 180182.150.100. 201214. 302303 diffusion approximation 17. 360 excursion 155156. 2425.301 central limit theorem 60 .285292. 122. 111117. 86.251280 heavy traffic 76.137141.299.318320 change of measure 2630. 5796.281. 8283 hyperexponential distribution 7. 80 81. 217.160167. 3032.Index adjustment coefficient 17.182.242. 227229. 1112. 318319 Erlang distribution 7. 278 gamma distribution 67.292293 Edgeworth expansion 113.228229. 323 Coxian distribution 147. 9293. 7079. 239. 17.293294. 9396. 97. 79.6779. 7475.200201. 316323 Bessel function 102. 162164. 218 Cox process 4.and sum 221.86. 17.226. 207 heavytailed distribution 6. 141144.272.178184. 205. 97129. 332333 Volterra 192194. 361 diffusion 3. 91. 40. 1819. 189.249.314316. 89.9899. 135.328330. 271274. 119. 12 CramerLundberg approximation 1617.217.259261. 301 Kronecker product.135.
39. 15.287. 234 matrixexponential distribution 240244 matrixexponentials 14.275278.287291 INDEX matrix equation .160161. 134. 3947. 245 M/G/1 13.227230. 16. 108109. 3639. see also sensitivity analysis phasetype distribution 8. 25.349 350 perturbation 172173. 213214. 16. 162.340350 multiplicative functional 2830. 175 light traffic 8183 Lindley integral equation 143 process 3334. 7179.339 large deviations 129.288290.161.123.238. 261264.180. 112113. 203 Markov additive process 12. 138. 132133. 6162.261264. 108 life insurance 5. 178 modulation 12. 38. 5758. 134135. 133. 203204.201. 227228. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 96.161164. 145187.297299. 257.304 process 2830. 157. 304305 random walk 3336.148.146148. 9899.139141.234. 32. 38. 4446.350361 Poisson process Markovmodulated 12 periodic 12.134135. nonlinear 155. 2730. 52 53. 71. 144. 269 PerronFrobenius theory 4142.174. 35. 6970. 44.298299. 80. 14. 141144.285287 queue 14 . 185187 GI/G/1 141144 M/D/1 6667 equation 16. 25. 106108. 260 Lundberg conjugation 6979 . 133. 37.259261. 41. 251.269271.348 terminating 215216. 154. 171. 7576. 179 NP approximation 318320 Palm distribution 5253. 176185.108.336339 . 44. 142 likelihood ratio : see change of measure lognormal distribution 9. 3947. 295.128129.152160. 271274. 149. 230. 267269 Panjer's recursion 320323 Pareto distribution 910. 99.178182.384 ladder heights 4756. 100.234240. 113114.240244.218221. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 306316 Levy process 3.302. 137139. 86 periodicity 12.215250.336339 Laplace transform 15.161. 35. 59. 65.315 inequality 1718. 42.
292294. 191192. 327 . 120 statistics x. 141144. 335336 sensitivity analysis 8693. 123. 152. 260 reinsurance 8. 12. 307308. 3032. 280. 168172 storage process 13. 256258. see also matrixexponential distribution regenerative process 264 268. 186187 renewal process 131. 251. 31. 7475. 107. 229234. 213. 60.154157. 317318 semiMarkov 147. 281296 stable process 15. 260 WienerHopf theory 144. 223226.336339 workload 13. 222. 11. 177 timereversion 14.INDEX 385 waiting time 141. 37. 1819.186. 87. 294296 shotnoise process 314 simulation 19. 233234. 5455.279280 Rouche roots 158. 172173. 4950. 331336 equation 64. 244. 131144. 147. 189214.273274. 253. 326330 Weibull distribution 9. 146.244250. 174. 238 saddlepoint method 115117. 9693.314. 160. 251280 time change 4. 162. 251. 333334 regular variation 10. 257.262263.359361 stochastic control x stochastic ordering 18. 261264 reservedependent premiums 14. 8386. 233. 186187 virtual: see workload rational Laplace transform 8. 332333 model 12. 89. 279280 subexponential distribution 11. 240. 338 utility 324.
Special features of the book are the emphasis on change of measure techniques. I 1! Ruin Probabilities . P'i yfliother approximations (e. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. worldscientific. It is a comprehensive treatment of the known results on ruin probabilities.. exact solutions. the ^W A l \ i l ' ''' CramerLundberg approximation. Markovmodulation or periodicity.Advanced Series on Statistical Science & Applied Probability .g.Vol. y finite horizon ruin probabilities. 2 A I 11 JjVb l' i  i Yj .com 2779 he 9 "789810ll22293211 . "This book is a must for anybody working in applied probability. extensions of the classical compound Poisson model to allow f o r reservedependent premiums. for heavytailed claim size distributions).." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. Some i (l I JL I J r of the topics are Lundberg's inequality.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory.