This action might not be possible to undo. Are you sure you want to continue?
Advanced Series on Statistical Science & I Applied Probability ^^^A£J
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
. . . . 290 5 Regenerative simulation . . . . .. . . . 336 A3 Matrixexponentials . . . . .. . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . . . . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula .. . . . . . . . . . . . . . . . 350 Bibliography Index 363 383 . 326 Appendix 331 Al Renewal theory . . . .. . . . . . . . . . . . . . . .. . . .. . . . . . . . . . 306 4 The distribution of the aggregate claims . .. . . . . . . . . . . . . . . . . . . . .. . . . . 285 3 Importance sampling via Lundberg conjugation . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 6 Reservedependent premiums . . . . . . . . . 279 X Simulation methodology 281 1 Generalities .. 297 2 Further applications of martingales .. . . . . . . . . . . . .. . 340 A4 Some linear algebra .. . . . . . . 287 4 Importance sampling for the finite horizon case . . . . . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . . . . . . . . 251 2 The compound Poisson model . . .. . 261 4 Models with dependent input . . . . . . . . . . . . .. . . . . . 304 3 Large deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 3 The renewal model . . . . . . . 323 6 Reinsurance . . . . . . . .. . . . . . .. . . . . ... . . . . . . . . . . . . 344 AS Complements on phasetype distributions . . . . . . . . . . . . . . . . . . 292 6 Sensitivity analysis . . . . . . . .. . . . . . . . . .. . . . . . . . . . . . The twobarrier ruin problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 5 Principles for premium calculation .. . . . . . . . . .. . . . . . . . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . . . ... . . 331 A2 WienerHopf factorization . . . .
This page is intentionally left blank .
Thus. I have deliberately stayed away from discussing the practical relevance of the theory. and the series editor Ole BarndorffNielsen for their patience. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. and other projects absorbed my interest. and my belief was that this could be done rather quickly.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. However. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. if the formulations occasionally give a different impression. 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!). Apart from these remarks. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. As an excuse: many of these projects were related to the book. the book is basically mathematical in its flavour. this applies to longrange dependence which is intensely studied in the neighboring ix . I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. 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. University of Copenhagen. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. The course was never realized. But the pace was much slower than expected. Since I was to produce some handouts for the students anyway. 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. and has been an active area of research from the days of Lundberg all the way up to today. In particular. and the result is now that the book is much more related to my own research than the initial outline. the idea was close to expand these to a short book on the subject. it is not by intention. It has obviously not been possible to cover all subareas.
e.5. In the classical setting of CramerLundberg models.lth. incorporate 11. for which I apologize to the reader and the authors of the many papers who ought to have been on the list. A book like this can be organized in many ways. Good luck! I have tried to be fairly exhaustive in citing references close to the text. More recently.45. for the effects on tail probabilities. Asmussen. an area which is becoming increasingly important. Another interesting area which is not covered is dynamic control. see e.2.6 (to understand the PollaczeckKhinchine formula in 111. 111. In addition. The main motivation comes from statistical data for network traffic (e.g. Willinger et al.g. see in particular Michna [259].se Lund February 2000 Soren Asmussen . IX.13. see also Schmidli [325] and the references in Asmussen & Taksar [52].15.13 and IX. The present book is in between these two possibilities.13 and XI. [381]). Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. Resnick & Samorodnitsky [303] and references therein. 111.x PREFACE field of queueing theory. IV. For a brief orientation. VI. X.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. another by method. the first part of 11. One is by model. VII. I intend to keep a list of misprints and remarks posted on my web page. VII.maths .14. 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). IV. It is obvious that such a system involves a number of inconsistencies and omissions.2.lth.g. Finally.13. Here is a suggestion on how to get started with the book. some papers not cited in the text but judged to be of interest are included in the Bibliography. read Chapter I.1. the standard stochastic control setting of diffusion models has been considered. it has not been possible to incorporate more numerical examples than the few there are. Hojgaard & Taksar [206].3. For a second reading. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. The rest is up to your specific interests. IV.4a. VIII. http:// www. Concerning ruin probabilities.89.2 more properly). some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. I regret that due to time constraints.
4 from Asmussen. IV. Fig. Parts of X. 3 is reprinted from Asmussen & Nielsen [39] and parts of IX.8 .6 by my 1999 simulation class in Lund. . of which there are not many at this stage . many of which were pointed out by Hanspeter Schmidli .1 by Bjarne Hojgaard and the table in Example 111.2 by Rafal Kulik . were produced by Lone Juul Hansen . More substantial remarks. Parts of II. Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . 111 . A number of other figures were supplied by Christian Geisler Asmussen . as well as some additional references continue to be at the web page.6. 5. not least the more complicated ones. Schmidli & Schmidt [47] with the permission from Applied Probability Trust .5 from Asmussen [21] with permission from CRC Press.6 is reprinted from Asmussen & Schmidt [49] and parts of IX.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. Section VIII. 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers. Section VII . Fig.PREFACE xi The second printing differs from the first only by minor corrections. Aarhus.1 and X. supported by Center for Mathematical Physics and Stochastics (MaPhySto).
This page is intentionally left blank .
M = (1. respectively. For mathematical purposes. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero.2) (O<t<T Ro=ul. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}.1) We also refer to t/) ( u) and 0(u. We denote throughout the initial reserve by u = Ro. (1. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u .4) O<t<oo O<t<T 1 .Rt. MT = sup St. is a model for the time evolution of the reserves of an insurance company. (1.Chapter I Introduction 1 The risk process In this chapter .3) sup St.i(u. A risk reserve process { Rt}t>o. and give a very brief summary of some of the models.T) = P inf Rt < 0 I . They are the main topics of study of the present book. results and topics to be studied in the rest of the book. (1. we introduce some general notation and terminology. T) as ruin probabilities with infinite horizon and finite horizon . as defined in broad terms .
We denote the interarrival times of claims by T2. . respectively. per unit time. Putting things together.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. Thus.. Figure 1.1.i(u. T3. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. (1. However.6) Sofar we have not imposed any assumptions on the risk reserve process. and T1 is the time of the first claim.T) = F (MT > u) = P(r(u) < T). say. 1.1 . and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. the time of arrival of the nth claim is an = T1 + • • • + Tn. St = E Uk . (1.5) i.2 CHAPTER I.b(u) = P (r(u) < oo) = P(M > u). • Premiums flow in at rate p. That is.. (1. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon.E Uk. we see that Nt Nt Rt = u + pt . the number Nt of arrivals in [0. t] is finite. the ruin probabilities can then alternatively be written as .pt.
allowing a countable infinity of jumps on Fig.. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. If 77 < 0.b(u) = 1 for all u. We shall not deal with this case either. and hence O(u) < 1 for all sufficiently large u.1. since any modeling involves some approximative assumptions. and hence . say 10% .8) The interpretation of p is as the average amount of claim per unit time.s. t * oo.) V 0. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. though many results are straightforward to generalize from the compound Poisson model. 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. for example. If 77 > 0.1.. VII.20%. however. one could well replace Rt by Rtnr(u) or RtA. It would appear obvious.1.s.(. • Brownian motion or more general diffusions. we shall. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. not discuss whether this actually corresponds to practice. and the basic ruin probabilities are derived in XI. 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. then M = oo a. then M < oo a. We study this case in Ch. . of course. 1. one may well argue that Brownian motion in itself could be a reasonable model.e. (1. Thus. 1. however. that the insurance company should try to ensure 77 > 0. immaterial.1 the slope of {Rt} should depend also on the level). and in fact: Proposition 1.1 Assume that (1. However. We shall discuss Brownian motion somewhat in Chapter IV. For the purpose of studying ruin probabilities this distinction is. on Fig. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. but as an approximation to the risk process rather than as a model of intrinsic merit.8) holds. a basic references is Gerber [127].
. in connection with risk processes in a Markovian or periodic environment (Chapter VI). . 0 We shall only encounter a few instances of a Cox process..10) is a property which we will typically encounter. _ St __ k =1 Uk pt a4. corresponding to the Pdlya process.11) .v.Q (say) and U1.8).oo t 0 J (provided the limit exists).8) that F N.i(u.10) hold with p constant.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t. However.s.i.d.. However. with the most notable special case being V having a Gamma distribution. Nt)}. namely. M < oo a. rl > 0. This case is referred to as the mixed Poisson process.8) is given by ^t p = EU • lim it (3(s) ds t. St In concrete models. tb(u) = 1 for all u holds also when rl = 0.10) Again. are i.Tp). zP(u . namely that M = oo a. this needs to be verified in each separate case. U2.d.6EU (on the average. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). 0(u.. and independent of {Nt}. . If U1. are i. where {Nt} is a Poisson process with rate . and independent of {(0(t). k=1 (1.T) = i. Thus p may well be random for such processes.. The simplest example is 3(t) = V where V is a r . then similarly limSt/t < 0.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.i. and that . then this limit is > 0 which implies St a$ oo and hence M = oo a. INTRODUCTION Proof It follows from (1.T) for {Rt} is given by V)(u) = t/i (u). U2. and here (1. If u oo. . (1. if {(3(t)} is nonergodic. we obtain typically a somewhat stronger conclusion.s. (1. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model.b(u) < 1 for all u when rl > 0. Here it is easy to see that p = . t t p  p' t ^ oo. (1. not all models considered in the literature have this feature: Example 1. it is not too difficult to show that p as defined by (1. Proposition 1.3 Assume p 54 1 and define Rt = Rt1p. If 77 < 0.s.4 CHAPTER I.
. In the even more general area of nonlife insurance mathematics. De Vylder [110]. see e . but in probability and applied probability as a whole. Daykin et al. Hipp & Michel [198]. The Swedish school was pioneering not only in risk theory. Gerber [157]. [134]. Besides in standard journals in probability and applied probability. Taylor [364]. lighttailed distributions (sometimes the term . Note that when p = 1. the role of the result is to justify to take p = 1.g. Embrechts 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. another important early Swedish work is Tacklind [373]. Note that life insurance (e.g. Some of the main general ideas were laid down by Lundberg [250]. Schmidt & Teugels [307] and Seal [326].. in particular.. often referred to as collective risk theory or just risk theory. Daykin. the assumption > 0 is equivalent to p < 1. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Rolski. Straub [353]. Since { Rt } has premium rate 1. We roughly classify these into two groups . Segerdahl [334] and Philipson [289]. Schmidli. see also Chapter XI. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory.. [76]. Buhlmann [82]. the claim arrivals are Poisson or renewal at the same time). Gerber [159]) has a rather different flavour. the research literature is often published in journals like Astin Bulletin . while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). Grandell [171]. Cox processes are treated extensively in Grandell [171]. Some early surveys are given in Cramer [91]. Pentikainen & Pesonen [101]. U2. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. we shall be able to identify p with the traffic intensity of an associated queue. For mixed Poisson processes and Polya processes. Sundt [354]. [330]. the recent survey by Grandell [173] and references therein. Some main later textbooks are (in alphabetical order) Buhlmann [82]. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. Insurance: Mathematics and Economics. Heilmann [191]. [101].2. Notes and references The study of ruin probabilities. was largely initiated in Sweden in the first half of the century. and we do not get near to the topic anywhere in this book. in a number of models. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al.
f.1) The parameter 6 is referred to as the rate or the intensity.2) = 0. regularly varying (see below) or even regularly varying with infinite variance.B(x) satisfies B(x) = O(e8x) for some s > 0.2 and /LB is the mean of B.g. The crucial feature is the lack of memory: if X is exponential with rate 6.3) .f. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . 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'.O(u) can be found in closed form. Equivalently. B is heavytailed if b[s] = oo for all s > 0. Here lighttailed means that the tail B(x) = 1 . As in a number of other applied probability areas.u at the time of ruin given r(u) is again exponential u with rate 8.6 CHAPTER I.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. In contrast. 2a Lighttailed distributions Example 2. INTRODUCTION 'Cramertype conditions' is used). 6 has density r(p)xPleax b(x) P and m.8. Example 2 . if 1 °O AB Jbos x B(dx) > 0. (2. the m. a fact which turns out to contain considerable information. for the compound Poisson model with exponential claim sizes the ruin probability .1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. For example in the compound Poisson model.e. and heavytailed distributions. In particular. the exponential distribution is by far the simplest to deal with in risk theory as well. P B[s]= (8Is ) .x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). but different more restrictive definitions are often used: subexponential. i. then the conditional distribution of X . where B(bo.g. s<8. B[s] is finite for some s > 0. On the more heuristical side. and can also be interpreted as the (constant) failure rate b(x)/B(x).
2. .d. then X v Xl + • • • + X. and exponential with rate d. 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. 0 < ai < 1. we develop computationally tractable results mainly for the Erlang case (p = 1. u .1 Poisson events in [0. This special case is referred to as the Erlang distribution with p stages. one has r(bx.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. An important property of the hyperexponential distribution is that its s. where X1.... p). by Grandell & Segerdahl [175] and Thorin [369]... if p is integer and X has the gamma distribution p. . u Example 2 . > 1 for p < 1 and = 1 for p = 1 (the exponential case).y i=1 where >i ai = 1. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions.. p) = J tPletdt.ate (b2 ): L• i=o In the present text.1) (or the 1/pth root if p < 1).v.).c. or just the Erlang(p) distribution. among others.2. x] so that B(x) = r` e. are i. In particular. B(x) = r(p) Asymptotically. The exact form of the tail B(x) is given by the incomplete Gamma function r(x.) VarX1 (EX )2 p is < 1 for p > 1. i = 1. In particular. p) °° where r (x.2) can be considered as the pth power of the exponential density (2.i. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. . 0. the squared coefficient of variation (s. P b(x) = r` aibiea. X2..v.c. the Gamma density (2. is > 1. Ruin probabilities for the general case has been studied.
7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix .5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. This class of distributions is popular in older literature on both risk theory and queues.6. This class of distributions plays a major role in this book as the one within computationally tractable exact forms of the ruin probability z/)(u) can be obtained. 1)' is the column vector with 1 at all entries. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree.d.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. T) is called the representation.g. which is slightly smaller but more amenable to probabilistic reasoning. The parameters of a phasetype distribution is the set E of transient states. We give some theory for matrixu exponential distribution in VIII. Important special cases are the exponential. Example 2 . but the current trend in applied probability is to restrict attention to the class of phasetype distributions. the Erlang and the hyperexponential distributions. . INTRODUCTION Example 2 .f. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. T) or sometimes the triple (E.8) are realvalued. a.. See XI.8) j=1 j=1 j=1 where the parameters in (2. equivalently. (or. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. However. of which one is absorbing and the rest transient.(2. The couple (a. are b(x) = aeTxt.6. there exists a xo < oo such that B(x) = 0 for x > xo.8 CHAPTER I. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx..1 and defer further details to u Chapter VIII.xo)+ is covered by the reinsurer).e.7) are possibly complexvalued but the parameters in (2.f. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities. 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 . We give a more comprehensive treatment in VIII. B(x) = aeTxe where t = Te and e = (1 . The density and c. Example 2 . resp. q2 q3 (2.
12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed.2.p a 1 (2.u l b(x) = d dx or J ax lor 1 exp Asymptotically.10) The loinormal distribution has moments of all orders.1. the tail is B (x ) 2 x.N(0. (2. and then b(x) = 0. we obtain the Weibull distribution B(x) = eCx'.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2. a2). one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. All moments are finite. the exponential distribution representing the simplest example since here b(x) is constant. 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). b(x) = crx''le`xr. It follows that the density is 't (1ogX . CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2.9) which is heavytailed when 0 < r < I. or equivalently as the distribution of a°U+µ where U .pl = 1 W (logx . the mean u is eµ+a /2 and the second moment is e2µ+2o2. However. There are various variants of the definition around.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. a)/A)a+1' x > a. (2.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . x < a. Example 2 . Writing c = d/r.N(p. In particular.1). Here failure rates b(x) = b(x)/B(x) play an important role. p is defined as the distribution of ev where V . (2. u Example 2 .13) u .
e. 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 > 5.1)/p.2). The density is 8p(log x)pi b(x) . In general. the loggamma distribution is a Pareto distribution. x 4 oo (any L having a limit in (0. in particular. { s () 1s+3s29s3log(1+2s I p=3.12 (DISTRIBUTIONS WITH REGULARLY VARYING TAILS) The tail B(x) of a distribution B is said to be regularly varying with exponent a if B(x) . where Y is Pareto distributed with a = (p .16) 11 Example 2.10 CHAPTER I. satisfies L(xt)/L(x) 4 1. (2. Thus. another standard example is (log x)').L( x ). Choudhury & Whitt [1] as the class of distributions of r. i.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. The motivation for this class is the fact that the Laplace transform is explicit (which is not the case for the Pareto or other standard heavytailed distributions). oo) is slowly varying . The simplest examples correspond to p small and integervalued.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. u . For p = 1.17) where L (x) is slowly varying.'s of the form YX. B(x) = O(xP). INTRODUCTION Example 2. examples of distributions with regularly varying tails are the Pareto distribution (2. x + 00.12) (here L (x) * 1) and ( 2.(1 + 2x + 2x2)e2x) p = 2 (2. in particular.13). the density is { 3 (1 .v. A = 1 and X is standard exponential. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. (2.x6+lr(p) (2. u Example 2 .(1 + Zx + $ p = 3.15) x2 + 16x3 ) a3x/2) 3 (1 .
3 The arrival process For the purpose of modeling a risk process .1.3. the knowledge of the claim size distribution will typically be based upon statistical data.18) B(x) It can be proved (see IX. but the model also admits a natural interpretation : a large portfolio of insurance holders . 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. Also. However.. (2. this phenomenon represents one of the true controversies of the area. From a practical point of view. though the proof of this is nontrivial. When studying ruin probabilities.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. Namely. U2. and so is the Weibull distribution with 0 < r < 1. one may argue that this difficulty is not resticted to ruin probability theory alone. THE ARRIVAL PROCESS 11 Example 2. Thus.. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available.. the claim size distribution represents of course only one aspect (though a major one). Similar discussion applies to the distribution of the accumulated claims (XI. and based upon such information it seems questionable to extrapolate to tail behaviour. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. We return to a closer study in IX. for example the lognormal distribution is subexponential (but not regularly varying).. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. which each have a ( timehomogeneous) small rate of experiencing a . 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. The reason is in part mathematical since this model is the easiest to analyze. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed.1) that any distribution with a regularly varying tail is subexponential. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1.
d.12 CHAPTER I. The difficulty in such an approach lies in that it may be difficult or even impossible to imbed such a distribution into the continuous setup of {Nt } evolving over time . where {/3 (t)}too is an arbitrary stochastic process . when Jt = i. are i. The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. 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.. 5. and also that the ruin problem may be hard to analyze . The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). in Chapter VII). However . The compound Poisson model is studied in detail in Chapters III. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . Some of them have concentrated on the marginal distribution of NT (say T = one year ). T2.. its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. such that 8(t) = . not many detailed studies of the goodnessoffit of the Poisson model in insurance are available .(3. so that . too general and one neeed to specialize to more concrete assumptions . Mathematically. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications.6. I.i. but with a general not necessarily exponential distribution ). in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). Another one is Cox processes. A more appealing way to allow for inhomogeneity is by means of an intensity .e. e. In order to prove reasonably substantial and interesting results . Cox processes are.g. This model . which facilitate the analysis. with a common term {Nt} is a Markovmodulated Poisson process . has some mathematically appealing random walk features . gives rise to an arrival process which is very close to a Poisson process. we study this case in VI .3(t) fluctuating over time. epidemics in life insurance etc. INTRODUCTION claim . IV (and. to be studied in Chapter V. with the extension to premiums depending on the reserve. Nevertheless . see 11. This applies also to the case where the claim size distribution depends on the time of the year or . the negative binomial distribution. An obvious example is 3(t) depending on the time of the year (the season). the periodic and the Markov modulated models also have attractive features . it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. however. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . Historically. getting away from the simple Poisson process seems a crucial step in making the model more realistic. To the author 's knowledge ..8 (t) is a periodic function of t.. In others. in particular to allow for certain inhomogeneities.
1). A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. interacting particle systems. extreme value theory. stochastic geometry. ruin probabilities for risk processes with an input process which is renewal. this amounts to Vo having the stationary distribution of {Vt}). reliability. time series and Gaussian processes. Mathematically. 0(u) = P(V > u).6) . The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. that quite often the emphasis is on computing expected values like EV.0 (u. Thus. genetics models. with Poisson arrivals and constant release rule p(x) = 1.4. The study of the steady state is by far the most dominant topic of queueing and storage theory. It should be noted. queueing theory. others being branching processes. Similarly.1) holds as well provided the risk process has a premium rule depending on the reserve.T) = P(VT > u). Some of these have a certain resemblance in flavour and methodology.v. The ones which appear most related to risk theory are queueing theory and dam/storage processes. and the limit t 4 oo is the steadystate limit. the classical result is that the ruin probabilities for the compound Poisson model are related to the workload (virtual waiting time) process {Vt}too of an initially empty M/G/1 queue by means of . In fact. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. others are quite different. point processes and so on. stochastic differential equations. and a lot of information on steadystate r. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. In the setting of (4. Markovmodulated or periodic can be related to queues with similar characteristics. R = p(R) in between jumps. and here (4. it is desirable to have a set of formulas like (4. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory.1) where V is the limit in distribution of Vt as t + oo. and which seems well motivated from a practical point of view as well. this gives only f0 O°i (u)du which is of limited . 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. dam/storage processes. methods or modeling ideas developed in one area often has relevance for the other one as well. however. (4. More generally.'s like V is available.1) permitting to translate freely between risk theory and the queueing/storage setting.
T). A prototype of the duality results in this book is Theorem 11.1 .p(y) y^ Jo p(x) can be written in closed form.6. Here O(u) = peryu where 3 is the arrival intensity. The cases where this is possible are basically the following for the infinite horizon ruin probability 0(u): • The compound Poisson model with constant premium rate p = 1 and exponential claim size distribution B. .14 CHAPTER I. 4b Exact solutions Of course . see Corollary VII. Vi(u. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .3. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions.8.'s like the environmental process {Jt} in a Markovmodulated setting.1). as is typically the case. have to some extent a different flavour.g.3. The infinite horizon (steady state ) case is covered by letting T oo. Example VIII. the functions w x f d 1 exdx () . see Corollary III. Similarly. though overlapping.1.3.v. e . 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. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B.3.2).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 ). Here ?P(u) is explicit provided that . B(x) = ebx. 3. • The compound Poisson model with some rather special heavytailed claim size distributions. p = 0/8 and y = 8 . the two areas. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). see Boxma & Cohen [74] and Abate & Whitt [3].. • The compound Poisson model with a claim size distribution degenerate at one point.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. which gives a sample path version of (4. INTRODUCTION intrinsic interest . Thus . 3. which can be expanded into a sum of exponential terms by diagonalization (see. The fact that Theorem H. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII.
[s.f f 2µ(y)/a2(y) dy} dx  (4. a2 (x): Ip (u) = where S(u) = f °O exp {. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). Given this can be done. the only example of something like an explicit expression is the compound Poisson model with constant premium rate p = 1 and exponential claim size distribution . (u. but are somewhat out of the mainstream of the area . However. • An astable Levy process with drift . 1).S(u) 1S(oo) f °D exp {.Lef$er function. relevant references are Grubel [179]. the formulas ( IV. f {eXp U LX 2. T). esuTb( u. 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. For the finite horizon ruin probability 0(u. O(u. Here are some of the main approaches: Laplace transform inversion Often.7. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. A notable fact ( see again XI.2) is the natural scale.ff 2µ(y)/a2(y) dy} dx . T) themselves.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. 191). Also Brownian models or certain skip free random walks lead to explicit solutions (see XI . Embrechts. see VIII. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. it is easier to find the Laplace transforms = f e8 . Ab(u). .1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. Abate & Whitt [2].u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. We don't discuss Laplace transform inversion much.b(u)du .4. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. T) can then be calculated numerically by some method for transform inversion.
it states that i/i(u) .g.3) in the compound Poisson model which is an integral equation of Volterra type. T) as the solution to a differential. In the compound Poisson model with p = 1.1) and y > 0 is the solution of the Lundberg equation (4. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). U is explicit in terms of the model parameters.) B[s]. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. and in particular the naive idea of conditioning upon process behaviour in [0. and carry out the solution by some standard numerical method. (4. dt] most often leads to equations involving both differential and integral terms. which can equivalently be written as f3 [7] = 1 +13 .3) where C = (1 .16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). One example where this is feasible is the renewal equation for tl'(u) (Corollary III. .f. see VIII. u * oo. However. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!).7.3.p)/(13B'[ry] .4) 00['Y]1)'Y = 0.Ce"u. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. 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. and in quite a few cases (Chapter VIII). Differential. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. 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.or integral equation.and integral equations The idea is here to express 'O(u) or '(u.
2. other approaches are thus required. for the compound Poisson model ^(u) p pu In fact . In the case of heavytailed distributions. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. 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. often for all u > 0 and not just for large u. Diffusion approximations are easy to calculate. the claim size distribution should have an exponentially decreasing tail B(x).7 and IV. Approximations for O(u) as well as for 1(u. Large claims approximations In order for the CramerLundberg approximation to be valid. In particular. u > oo. some further possibilities are surveyed in 111 . However. in such cases the evaluation of C is more cumbersome. However.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. when the claim size distribution is of phasetype. In fact. . 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. It has generalizations to the models with renewal arrivals. but typically not very precise in their first naive implementation. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. incorporating correction terms may change the picture dramatically. corrected diffusion approximations (see IV. T). This list of approximations does by no means exhaust the topic. a Markovian environment or periodically varying parameters. For example. in some cases the results are even more complete than for light tails.4. (4. 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. T) for large u are available in most of the models we discuss.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. See Chapter IX.
8.k (U(`) . one expects a model with a deterministic claim size distribution B. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. . more importantly.18 CHAPTER I. of being somewhat easier to generalize beyond the compound Poisson setting. INTRODUCTION Compared to the CramerLundberg approximation (4. in the compound Poisson model. and to plot the empirical mean residual life 1 N . given NT. lower bounds etc. which is a standard statistical problem since the claim sizes Ui.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.U(k)) i =k+ i . We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. it has the advantage of not involving approximations and also. However. In practice.3). T]. How do we produce a confidence interval? And. obtained say by observing the risk process in [0.) at various places and in various settings. When comparing different risk models. fitting a parametric model to U1. 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. say degenerate at m. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. This procedure in itself is fairly straightforward. For example. it is a general principle that adding random variation to a model increases the risk. the difficulty comes in when drawing inference about the ruin probabilities. UNT are i. .. as a general rule. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail.. . .i.g. For example. they have however to be estimated from data. However.. This is proved for the compound Poisson model in 111. empirical evidence shows that the general principle holds in a broad variety of settings. The standard suggestion is to observe that the mean residual life E[U . 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. . to have smaller ruin probabilities than when B is nondegenerate with the same mean m. UNT may be viewed as an interpolation in or smoothing of the histogram). e.d. . though not too many precise mathematical results have been obtained.
However.2. Klnppelberg & Mikosch [134]. The infinite horizon case presents a difficulty. .3) or Section 3 of Chapter VI are referred to as Proposition VI. CONVENTIONS 19 as function of U(k). and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget.. (or a functional of the expectation of a set of r. Truncation to a finite horizon has been used. and look at them to see whether they exhibit the expected behaviour or some surprises come up...i. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. where U(1) < .. < U(N) are the order statistics based upon N i.v. See further Embrechts.. but is not very satisfying. to observe whether one or the other limiting behaviour is apparent in the tail. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. UN.v's) which can be generated by simulation.4. because it appears to require an infinitely long simulation. and in fact methods from that area can often be used in risk theory as well . We look at a variety of such methods in Chapter X.3) and Section VI. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. respectively. Still.5.3). For example. A main problem is that ruin is typically a rare event (i. in all other chapters than VI where we just write . this is a straightforward way to estimate finite horizon ruin probabilities.d. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics.(5. reference [14].e. Thus Proposition 4.2. claims U1. The chapter number is specified only when it is not the current one. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. . and of course the method is relevant in risk theory as well. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. in this book referred to as [APQ]). formula VI. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. formula (5.3 (or just VI.
The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. B(dy).g. w. Abbreviations c.d. E. IIGII the total mass (variation ) of a (signed ) measure G . E expectation. A different type of asymptotics: less precise.f. i. B(x) the tail 1 . References like Proposition A. cumulant generating function.4.20 CHAPTER L INTRODUCTION Proposition 4. infinitely often l. . (moment generating function) fm e82B(dx) of the distribution B. and for a defective probability distribution IIGII < 1. with respect to w.g. In particular. or a more precise one like eh .g.29) refer to the Appendix.c. see under b[s] below. i.s. say a heuristic approxi1 + h + h2/2. If. as for typical claim size distributions.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.f. n i oo. with probability Mathematical notation P probability. if B(x) . right hand side (of equation) r. EX2/(EX)2.f.3) or Section 3.f.f.v. random variable s. The Laplace transform is b[s]. squared coefficient of variation.i.s. cumulative distribution function P(X < x) c. n! 27r nn+1/2en.h.g.d. r. for a probability distribution IIGII = 1. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example.g.v. mation.t. moment generating function.h. B(x) = P(X < x) = fx. left hand side (of equation) m.2. then for 1s < 5).ceax.r. B is concentrated on [0.o.B(x) = P(X > x) of B. oo).f. formula (5. (A. B[s] the m. .e. independent identically distributed i.d. h + 0. log E[s] where b[s] is the m.p.
7r. the ith unit row vector is e'i. 21 D [0. (the dimension is usually clear from the context and left unspecified in the notation). often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. 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. i. Xt_ the left limit limstt X8f i.5. R(s) the real part of a complex number s. 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. and column vectors have lowercase Roman letters like t. the value just before t. a2) the normal distribution with mean p and variance oa2. only have finitely many jumps in each finite interval.e. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. A. xa. i. . oo) the space of Rvalued functions which are rightcontionuous and have left limits. 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). though slightly more complicated. Matrices and vectors are denoted by bold letters. . intensity interpretation. of numbers. N(it. In the Frenchinspired literature. the ith entry is 1 and all other 0. a. all stochastic processes considered in this book are assumed to have sample paths in this space. E[X.A] means E[XI(A)]. Thus. Notation like f3i and 3(t) in Chapter VI has a similar . the processes we consider are piecewise continuous.. row vectors have lowercase Greek letters like a. I(A) the indicator function of the event A.. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). F o r a given set x1. 0 marks the end of a proof.e. Unless otherwise stated.e. Usually. matrices have uppercase Roman or Greek letters like T.
cf.22 CHAPTER L INTRODUCTION B the claim size distribution. ry The adjustment coefficient. or quantities with a similar time average interpretation.1. FL. I. VI. interpretation. cf. J the rate parameter of B for the exponential case B(x) = eby. I.5. e. p the net amount /3pB of claims per unit time.g. cf.1. Notation like BE and B(t) in Chapter VI has a similar. . EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. 'q the safety loading .5. 111. though slightly more complicated.
in most cases via likelihood ratio arguments. 5) are. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here.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. Sections 4. however. The reader should therefore observe that it is possible to skip most of the chapter. Sections 4. however. The topic is. Due to the generality of the theory. somewhat more advanced than in the rest of the book. in part. strictly speaking. The duality results in Section 3 (and. used in Chapter VI on risk processes in a Markovian (or periodic) environment. More precisely. in particular at a first reading of the book. not crucial for the rest of the book. a parallel selfcontained treatment is given of the facts needed there. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. The general theory is. the level of the exposition is. 23 . 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. The likelihood ratio approach in Section 2 is basic for most of the models under study. All results are proved elsewhere . When encountered for the first time in connection with the compound Poisson model in Chapter III.
We get 1 = Ee7So = E e'Y S(.1 is basic for the study of the compound Poisson model in Chapter III. T(u) > T] . however. V) (u. Proposition 1. and in the limit (1. Thus N. T) = P(T(u) < T). the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem .1 Assume that (a) for some ry > 0. {e'YS° }t>0 is a martingale. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. and the ruin probabilities are ip(u) = P (T(u) < oo).2 Consider the compound Poisson model with Poisson arrival rate . T(u) < T] + E [eryST . Let e(u) = ST(u) . As usual.(u).QµB < 1. The more general Theorem 6. claim size distribution B and p = .24 CHAPTER II. Example 1 .. f1 . 1 Martingales We consider the claim surplus process {St} of a general risk process.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/.2) As T > oo.u denote the overshoot.0. (b) St a$ oo on {T(u) = oo}. 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.)AT = E [e7ST(°).2) takes the form 1 = E [e'ys(). StUit. the time to ruin r(u) is inf It > 0 : St > u}. using r(u) A T invokes no problems because r(u) A T is bounded by T).5 can be skipped. (1. T(u) < oo] + 0 = eryuE [e7Vu).
a = a . Under the conditions of Proposi Proof Just note that C(u) > 0. of the normal distribution.3/6 < 1. O(u ) < e7". Eeas° = e"(°) where n(a) = a2a2/2 ./3. Since {St} has stationary independent increments.f. the ruin probability is O(u) = pe.2. Thus.g.x. By standard formulas for the m. Now at the time r(u) of ruin {St} upcrosses level u by making a jump .= e"(') where K(a) = .i.Q. u Corollary 1.a it is immediately seen that y = S . Then {St } is Brownian motion with variance constant o2 and drift p < 0. and thus by the memoryless property of the exponential distribution .1. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.u + x is again just exponential with rate S. Example 1 . with common distribution B (and independent of {Nt}).1.Ft = a(S" : v < t). From this it is readily seen (see III. and p =.(„)_ = x is that of a claim size U given U > u . Thus.1) shows that Eels. and (b) follows from p < 1 and the law of large numbers (see Proposition III. 1. Since {St} has stationary independent increments. the martingale property now follows just as in Example 1.d.1) .Q and the U.Q(B[a] . MARTINGALES 25 where {Nt} is a Poisson process with rate .a. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx .1) . Proof Since c(a) = /3 (B[a] .1 is satisfied. and thus Ee7s° = 1.2 and drift p > 0.1.1 are satisfied. A simple calculation (see Proposition III. and thus Ee'rs° = 1. The available information on this jump is that the distribution given r(u) = t and S.ap. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .2(c)).5 For the compound Poisson model with B exponential. u Corollary 1. the conditional distribution of the overshoot e(u) = U .4 (LUNDBERG ' S INEQUALITY ) tion 1 . B(x) _ edx.. condition (a) of Proposition 1. are i. the conditions of Proposition 1.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists.3 Assume that {Rt} is Brownian motion with variance constant o.r" where y = S .
Example 2 .. Embrechts. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. F(S) = P(S) = 1. F). then S and S are disjoint .1 Let F.Ft}too and the Borel afield F. Thus the sets S = I tlim +oot t =. But if a $ ^ . Proof Just note that ^(u) = 0 by continuity of Brownian motion. A E Ft. The number Nt F) of jumps > e before time t is a (measurable) r.v. . P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. B. Grandell [171].3 below). on (DE. hence so is Nt = limfyo N2`i. Theorem 2. I. Two such processes may be represented by probability measures F. cf. More recent references are Dassios & Embrechts [98]. Grandell & Schmidli [131]. However. oo).6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. which we equip with the natural filtration {. then z/'(u) = e7" where 'y = 21A/a2. and is further exploited in his book [157].1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. P on (DE. the parameters of the two processes can be reconstructed from a single infinite path. P are then singular (concentrated on two disjoint measurable sets).6 N S = { lim Nt I t +00 t gJ are both in F.F). u Notes and references The first use of martingales in risk theory is due to Gerber [156]. as shown by the following example this setup is too restrictive: typically'.e. 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. (2. and by the law of large numbers for the Poisson process . and F. Delbaen & Haezendonck [103] and Schmidli [320]. 0 and claim size distributions B. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. A somewhat similar u argument gives singularity when B $ B. A]. [172].
A E Ft .F).Pt}adapted process {Lt}t>o (2. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. (i) If {Lt}t> o is a nonnegative martingale w. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE.r. Conversely. that the restriction of P to (DE.1) and nonnegativity by letting A = {Lt < 0}. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. ({Ft} . The truth of this for all A E Y.t. Lets < t. G ] = E [LT . _.t. Ft). Proposition 2. F the Borel o•field and P a given probability measure on (DE. P) such that LLt = 1.1) holds. . we have E [ LTIFT]1 = LT on {T < T}.r. ELt = 1 follows by taking A = DE in (2. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). G C {T < oo}. the restriction of P to (DE.Tt) is absolutely continuous w.1) holds. This proves (i). then {Lt} is a nonnegative martingale w. define P by Pt (A) = E[Lt.F such that (2. using the martingale property in the fourth step.Ft}.Y) such that P(A) = Pt(A). if for some probability measure P and some {.2 Let {Ft}t>o be the natural filtration on DE. P be as in Proposition 2.2. Proof Under the assumptions of (i). Lt < 0] can only be nonnegative if P(A) = 0. 1 J (2. Then P(A) = E[Lt. If r is a stopping time and G E PT.r. A]. A]. implies that E[LtI. (ii) Conversely. then { 1 P(G) = EG .3 Let {Lt}.Pt)) The following result gives the connection to martingales. A E F. A] = E[Lt.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). then there exists a unique probability measure Pon . A E F8. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE. 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 . Finally. F) such that ELt = 1. By the martingale property..2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. Hence E [_ . Then Ft (A) = E[Lt.2(i).t.F8] = L8 and the martingale property.e. . ({.
is nonnegative and has Ey Lt = 1 for all x.4) Proof Letting G = {r(u) < oo}. {St} = {u . (2. and this problem will now be studied. SOME GENERAL TOOLS AND RESULTS In the general case .O(u) = eryuE[e'YC(u).Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. St). First we ask when the Markov property is preserved.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). Consider a (timehomogeneous) Markov process {Xt} with state space E.t. For the definition. we assume for simplicity that {Xt} has Dpaths. To this end. u From Theorem 2.1.r.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1.2) by noting that 1 = ersr(„) = e1'ue7Ou). r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. one would typically have Xt = Rt. say. in continuous time (the discrete time case is parallel but slightly simpler). The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. Xt = St.Gn {r <T} .r. . Xt = (Jt.t. first in the Markov case and next (Sections 4.Ft} .1: Corollary 2.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . 1 Since everything is non negative.. In the context of ruin probabilities.h. t. Lr(u) 11 The advantage of (2. is Markov w. each F. The crucial step is to obtain information on the process evolving according to F. Rt) or Xt = (Jt.4) compared to (1. (2.28 CHAPTER II. and letting T * oo. we need the concept of a multiplicative functional. T(u) < oo]. we have F(G) = V )(u). the natural filtration {.4 Under condition (a) of Proposition 1.1).. A change of measure is performed by finding a process {Lt} which is a martingale w.2) follows by monotone convergence.s. applying (2. of (2. where {Rt} is the risk reserve process. Now just rewrite the r. 5) for processes with some randomwalklike structure.1) in Proposition 1. both sides are increasing in T.
t and let Px be the probability measure given by t. for all x. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.v.T9measurable Y8.Ft] = Ex.Y8f t < s. Proof Since both sides of (2.. Indeed.7) for any Ftmeasurable Zt and any . t] * [0. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.Ft }. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. s.5) Pxa. The precise meaning of this is the following: being . Similarly.(A) = Ex [Lt.5 Let {Xt} be Markov w. o 9t = V. and then L.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any .v. o 9tI.t. (2.2.v. ({Xt+u} 0<u<8) Theorem 2.Ftmeasurable.(Xtitl) with all t(i) < t + s. since Zt • (Y8 o Ot ) is . t. Zt.[Y.v. (2.5) are Tt+e measurable. where Ot is the shift operator. 0 .. the Markov property can be written E. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2. Y8.Ft]. nonnegative and Lt+8 = Lt•(Lso9t) (2. (2. A]. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].s.r. or. (2.7).'s of the form Zt • (Y8 o 0t) comprises all r. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x. The converse follows since the class of r.6) implies (2.v.Ft+8measurable. oo). the natural filtration {Ft} on DE. By definition of Px.8) which is the same as (2.6) for any .7). this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].F8measurable r.Pt+8measurable r.YB] for any Ftmeasurable r. Vt+e.'s of the form fl' f.
(using the Markov property in the second step) so that the martingale property is automatic. UN.. aN where or* = T UN_k+l. Thus R = Ro + f p(R8) ds . with a proof somewhat different from the present one. < aN < T.... In between jumps. . {Vt} ... then Remark 2. the premium rate is p(r) when the reserve is r (i.30 CHAPTER H. CN. The storage process {Vt }o<t<T is essentially defined by time reversion. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x.At where At = k: vk <t U. Ro = u (say). In between jumps. .. 3 Duality with other applied probability models In this section. R = p(R)).Ft] = LtE[L8 o 9t I.1) The initial condition is arbitrary.6 For {u .. the arrival epochs are Qi. t] = LtExt L8 = Lt. The corresponding claim sizes are Ul. Indeed.e.. t. . The result is a sample path relation. (3. see Dynkin [128] and Kunita [239]. The formulation has applications to virtually all the risk models studied in this book.5 can be found in Kuchler & Sorensen [240]. . and the time to ruin is 7(u) = inf {t > 0: Rt < 0}.. 0 < vl < . A more elementary version along the lines of Theorem 2. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. reflection at zero and initiar condition Vo = 0. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. More precisely . .. We work on a finite time interval [0. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. } E[Lt+B I. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process.
.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability. Then rt°) > rt°) for all t when u > v.1.e. {Vt} remains at 0 until the next arrival).T) = P(VT > u). :. 3. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i._.3. V)(u.____•_. Theorem 3.1) we have Vt = At  f P(Vs)ds where A= U= AT . V = p(V))..1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u. (3.x.11 4.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0... instead of (3.1 The events {T(u) < T} and {VT > u} coincide.__... In particular. (3.__.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u. . Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.___ ..AT_t. That is.. The sample path relation between these two processes is illustrated in Fig.._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3.
one may feel that the interaction between the different areas has been surprisingly limited even up to today. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. Nevertheless. Corollary 3. with distribution B. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . the distinction between right.1 and its proof is from Asmussen & Schock Petersen [50].2 Consider the compound Poisson risk model with a general premium rule p(r).1 with Ro = u = ul). The results can be viewed as special cases of Siegmund duality. .1 with Ro = u = u2). u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell.i. We get: Corollary 3. if and only if O(u) < 1 for all u. Proof Let T ^ oo in (3. and then '0 (u) = P(V > u).T l . and since ruin can only occur at the times of claims. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. we can repeat the argument and get VoN_1 > Ra2 and so on.2 from Harrison & Resnick [188].and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. and a general premium rule p(r) when the reserve is r. Then similarly VVN = r0. we have RQ„ < 0 so that indeed r(u) < T. Hence if n satisfies VVN_n+1 = 0 (such a n exists. Historically. and so on. If VaN > 0.U1 > roil . Resnick & Tweedie [79].32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. if nothing else n = N). Hence RQ„ > 0 for all n < N.3). this represents a model for storage.U1 = Rol. = r(VT) . Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. 3. 3. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo.3 and being i. Historically. say of water in a dam though other interpretations like the amount of goods stored are also possible. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51]. see Siegmund [344]. say V. Theorem 3. The arrival epochs correspond to rainfalls.d. Then the storage process {Vt} has a proper limit in distribution. we have r(u) > T.Ul < roil  Ul = RQ„ Va1V_1 < RQ2. Then Vo.
..1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}. .2) (for a rigorous proof. W1.. hits (oo..e. . ZN = . (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo... WN be the Lindley process generated by Z1 = YN. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN. Theorem 4. with common distribution F (say).. .. and is reset to 0 once the r.1.4.1.. n=0.. Z2 = YN_1 i ..YNn+1) n=0..2).1)). 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.. Let further N be fixed and let Wo. N min (Y1 + + Yn).2) satisfies the same recursion as in (4. = Xo + Y1 + • • • + Y. where the Yi are i . i.. as long as the random walk only takes nonnegative values. has a proper limit W in distribution as n + oo. I. Z2. Z2 .Yl min (YN .1..Y1 according to Wo = 0.. 0). .d..1 in terms of Lindley processes . W1.. For discrete time random walks .. of (4.. WN = YN .. For a given i. Proof By (4.. R valued sequence Z1. Then the events {r(u) < N} and {WN > u} coincide.. In particular. (c) The Lindley process {WN} generated by Zl = Y1. Z2 = Y2..1. N min (Y1 + • • • + YNn) n=0.. . the Lindley process Wo.. . . ..N From this the result immediately follows.1.h. generated by Z1.N (4.w... evolves as a random walk with increments Z1i Z2.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0.... Xo = 0. Most often.min n=0.. can be viewed as the reflected version of the random walk with increments Z1. there is an analogue of Theorem 3... W2. . Here F is a general probability distribution on R (the special case of F being concentrated on {1.. just verify that the r ..1....i... 0 Corollary 4.1) Thus {Wn}n=o.. {Wn}n=0. Z2.... is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4..d. 1} is often referred to as simple random walk or Bernoulli random walk).. ..s..
s.s .g.) sup n=0. Proof Since (YN. e.. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists.l. . there is a more general version of Corollary 4. .1 have the same distribution for n = 0..d. a Markovian change of measure as in Theorem 2. .. .. 176) but appears to be rather intractable.. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . (d) #. Clearly. Y1) has the same distribution as (Y1...5 does not necessarily lead to a random walk: if. 0 By the law of large numbers. then the restrictions of Fx. One then assumes Yn to be a stationary sequence.=o.3 The i... ZN or. Remark 4 .i. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p..l.1..1. (e).2 and Theorem 4..s. or M < oo a. either M = oo a. Thus the assertion of Theorem 4. the Lindley processes in Corollary 4. .1.. . the Y1.g.1 is equivalent to WN D MN = (Z1 + ... doubly u infinite (n'= 0. . Combining these facts gives easily the equivalence of (a)(d). YN). a sufficient condition for (e) is that EY is welldefined and > 0. w. + Z.. By Kolmogorov's 01 law. . 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. equivalently. ±1.2. In that case .the result is a sample path relation as is Theorem 3. For a random walk. Next consider change of measure via likelihood ratios..s. ±2.o. .ooa. Similarly...) and defines Zn = Yn. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. (Yi + • • • + Yn) > oo a.. . N.. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u )..N so that WN _P4 M = supra=0. assumption on the Z1. SOME GENERAL TOOLS AND RESULTS (d) m = inf. (e)Yi+•••+Yn 74 . YN in Theorem 4.. In general...1 is actually not necessary . . W v m and P(W > u) = P (m > u) = 0(u).1..34 CHAPTER II.
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.3) holds. h(Yn) (4.4.. (4. For n = 2. e. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. the changed increment distribution is F(x ) = E[h(Y).5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.. Breiman [78] p. 100 ). Then the change of measure in Theorem 2. Conversely. and so onforn =3. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) .3) holds for n = 1.4.g. Y) = h(Y ) a. of F). (a) = log F(a] is the c. u A particular important example is exponential change of measure (h(y) = e°y'(") where r. Then the change of measure in Theorem 2. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. In particular.. .g.4) ({Ln} is the familiar Wald martingale . for some function h with Eh(Y) = 1. Proof If (4.3) 1Pxa. cf. We get: Corollary 4.4).f. Since L1 has the form g (Xo. Y ) f (Y)] = E[g(O. Y) f (Y)] for all f and x.. Y1).s. Y < x].s.4 Let {Ln} be a multiplicative functional of a random walk with E_L. the random walk property implies Ex f (Y1) = Eo f (Y1 ). y ). In that case..Y2) = h(1'i)h(I'a). where h (y) = g(0.nrc(a )} (4.. implying g (x. and define Ln by (4. this means E(g(x.5 corresponds to a new random walk if and only if Ln = h(Y1) . = 1 for all n and x..
however. Xt =Xo+pt+oBt+Mt. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}.7) for all e > 0. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). {Xt} can be written as the independent sum of a pure drift {pt}. see Chapter V. (4. the tradition in the area is to use continuous time models. < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) .5) Note that the structure of such a process admits a complete description. corresponds to a process with stationary independent increments and u = p.6) More precisely. the pure jump process is given by its Levy measure v(dx). say the beginning of each month or year . the claim surplus process for the compound Poisson risk model .Xt)I. The simplest case is 3 = JhvMM < oo. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). 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. this description needs some amendments. In continuous time. {Xt} is a random walk. v2 = 0 and v = 3B).36 CHAPTER II. or imbedded into continuous time processes . However. (4.e. In discrete time. e f x:IxJ>e} v(dx) < oo (4. i. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . Xt (n)t(n1) being independent whenever t(O) < t(1 ) < .Ft] = Eof (X. Xt(2)_t(l).t(i . Roughly. . . say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples).. but we omit the details ).. a positive measure on R with the properties e J x2v(dx) < oo. given by a the increment distribution F(x) = P(Xn+l . An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . they arise as models for the reserve or claim surplus at a discrete sequence of instants. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular .. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo.1)..). with premium rate p. In risk theory.Xn < x). we are .
where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". has upwards jumps governed by B at the epochs of a Poisson process with rate . Now consider reflected versions of processes with stationary independent increments. defined as a system with a single server working at a unit rate.7 If {Xt} has stationary independent increments as in (4. jxJ v(dx) < oo.Q and distribution B of the service times of the arriving customers. T) = P(VT > u). i. Chapter III. 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). O(u.v.1)v(dx) (4.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. (ex . [The condition for V < oo a.3 and decreases linearly at rate 1 in between jumps.4. 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. Then the storage process {Vt} has constant release rate 1.8) O<t<T (assuming Wo = Xo = 0 for simplicity). 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) = 1.6). 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. and b(u) = P(V > u). then Ee'(xtxo) = Eoeaxt = etx(a). Proposition 4.1. having Poisson arrivals with rate . and the reflected version is then defined by means of the abstract reflection operator as in (4. Corollary 4. is easily seen to be f3pB < 1. Furthermore.6 In the compound Poisson risk model with constant premium rate p(r) . WT = XT .min Xt (4.s. 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. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. oo].2).10) .e. VT + V for some r. cf. V E [0.
By explicit calculation . Eea(µt + QBt) = et{aµ +a2OZ/2}. In particular.11) (eax . we use the characterization (4.5) and get E [f(Xt+B . Q2 = v2. Theorem 4 .g. e.Xt)L8 o 0tIFt] = E [f (Xt+s .1)eexv(dx).f.1 that E eaMt = exp fmoo In the general case . .Xt)I'Ftl = E [f(Xt+B . 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t. Chung [86]). This is of course no coincidence since the distribution of Xl .g.Xt)I Ftl = Eof (X8)g(s.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. Proof For the first statement . (4. if Lt = e9(xt . Then the Markov process given by Theorem 2. t.6) are µ = µ + Oo2 . Xt Xo) with E2Lt = 1 for all x. use the representation as limit of compound Poisson processes. v(dx) = e9xv (dx).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 . 5 has stationary independent increments as well. 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) .38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. Xt +B .10) is the LevyKhinchine representation of the c. u Note that (4.1 ) v(dx) .1. we show in the compound Poisson case ( IlvIl < oo) in Proposition III. then the changed parameters in the representation (4.Xt)g(s. of an infinitely divisible distribution (see.. X8) = Eof (X8)L8 = Eof (X8)• For the second. let e" (a ) = Eoeaxl.xo)tk ( e).
Recalling that U1.2.St)g(Jt+s)I. one reason is that in parts of the applied probability literature. a = 0. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . Thus (since µ = p = 1.l3 and claim u size distribution B.0 in the following. MARKOV ADDITIVE PROCESSES 39 Remark 4. the corresponding claim sizes . abbreviated as MAP in this section2.. is defined as a bivariate Markov process {Xt} = {(Jt. 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. B(dx) = B[9] B(dx). we write Pi. the structure of MAP's is completely understood when E is finite: 2and only there .10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .8.1) For shorthand . MAP stands for the Markovian arrival process discussed below. . Ei instead of P2. dB/dB = b/b when B.3 and claim size distribution B.. 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 . u 5 Markov additive processes A Markov additive processes.3B[B]. Example 4 .(3B(dx).4). As for processes with stationary independent increments . b with b(x) > 0 for all x such that b(x) > 0). Example 4 . and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate.g. . Then we can write v(dx) _ /3eOxB(dx) = / (dx). are the arrival times and U1.3 =.11 For an example of a likelihood ratio not covered by Theorem 4. U2. <t whenever the RadonNikodym derivative dB/dB exists (e.(3 = . where .. corresponding to p = 1.5.9 If X0 = 0.0. it is then easily seen that Lt = H dB(Ui) i:o. B have densities b.Ft] = Ejt. (5. v(dx) _ .3 and claim size distribution B # B.o[f (S8)g(J8)]... Ei. then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4. 0.
) If E is infinite a MAP may be much more complicated. Proposition 5.f.jEE• On an interval [t.40 CHAPTER H. . 1 J1 ='^])iJEE = (Fij[a])i . a MAP is the same as a semiMarkov or Markov renewal process. {Jt} is specified by its intensity matrix A = (Aij)i.[a) = (Ei[easl.g. with the Y„ being interpreted as interarrival times. Fn[a] = F[a]n where P[a] = P . the converse requires a proof. As an example. SOME GENERAL TOOLS AND RESULTS In discrete time. vi(dx) in (4.. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time. by generating Yn according to Hij when J„_1 = i. the distribution of which has some distribution Bij. v. {St} evolves like a process with stationary independent increments and the parameters pi.J1=j)= Fij (dx) Pij In simulation language. t+s) where Jt . In continuous time (assuming Dpaths).9 EE = (iii&ij[a])i j EE .jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.Sr_1.o(Ji = j.1 For a MAP in discrete time and with E finite.. consider the matrix Ft [a] with ijth element least Ei . Y2. let {Jt} be standard Brownian motion on the line. 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. If all Fij are concentrated on (0.. An alternative description is in terms of the transition matrix P = (piA. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. oo). As a generalization of the m. In addition.. (That a process with this description is a MAP is obvious. Y1 E dx) where Y„ = S„ . Jn = j.i.. which we omit and refer to Neveu [272] or cinlar [87].it = A.6) depending on i.
J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. Jt = k] { xk kEE j la] . which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . where K[a] = A+ (r.1) } (recall that qjj = 0). MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. vi(dx). Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . Sn ) yields Ei[easn+ '. u Proposition 5. vi(dx) (i E E). j E E) and So = 0. kEE Jn = k]Ek[e"Y" . Then the matrix Pt[a] with ijth element Ei [east. up to o( h) terms.4c). Bij (i.1 )v(dx). Jn+1 = A] = 41 Ei[ e 5„. pi. a= . we infer that in the discrete time case the .1)) . this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. By PerronFrobenius theory (see A.ijgij(Bij[a] . \ diag Ft[a] = Ft[a]K. 00 r(i) (a) = api + a2ot /2 + f (e° .5.(')(a)) diag + (). assume that the Markov chain/process {Jt} is ergodic.qkj + k?^j qkj Bkj [a] } = Ei [east. aSt h = (1 + Ajjh) Ei [east . Jt = k] { 1 . Then. Jt = j] is given by etK[a]. qij. u In the following. Jt = j] Ejesh'^ + E Ak j hEi [ease .2 Let E be finite and consider a continuous time Markov additive process with parameters A. In matrix formulation . 013 .
h(a)vva)etw(a). and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . Corollary 5. The corresponding left and right eigenvectors v("). Yrh(a ) = 1. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a).3 Ei [east. a. In particular. and we shall take V(a)h(a) = 1. Furthermore.7.tK(a)h(a) J jj it L o is a martingale. as will be seen from the following results. Proposition 5. we are free to impose two normalizations. The function ic(a) plays in many respects the same role as the cumulant g. 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. its derivatives are 'asymptotic cumulants'.Eikjt = ttc'(0) + ki . . Corollary 5.r. Jeast.e=e°tk.4 Eie"sth(a) = h=a)et?(").4c).5 EiSt = tK'(0) + ki . Jt = j] . of a random walk.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).f.42 CHAPTER II. Since v("). Proof For the first assertion. u Let k(a) denote the derivative of h() w. We also get an analogue of the Wald martingale for random walks: Proposition 5.4. and write k = k(°).c(a) (and h(")). cf. cf. Then h(°) = e.2) where 7r = v(°) is the stationary distribution.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . h(") are only given up to a constants. Proof By PerronFrobenius theory (see A. Corollary 5. (5. h(") may be chosen with strictly positive components. Eie"sth^a) = e'Pt[a]h( a) = e.t.
(5. summing and letting a = 0 yields E„ [St + 2Stkj.") }) . Squaring in Corollary 5. . one obtains a generalization of Wald's identity EST = Er • ES.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1..a) + ttc (a)2hia ) Multiplying by v=. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) .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.4. In the same way. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions..+ k. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. 8 Also for E being infinite (possibly uncountable ).St]2 = t2/c'(0 ) 2 + 2ttc (0)vk . tam E tSt a (0).2ttc (0)Evkjt + 0(i). MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5.Eikjr . Remark 5 . 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . we differentiate (5. Corollary 5. [E. t a oo. For the second .7 No matter the initial distribution v of Jo. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.4) .e.5 yields + W (a)k. the distribution of Jo). subtraction yields Vary St = tic"(0) + O(1).5.g. More precisely.6 For any stopping time T with finite mean. u The argument is slightly heuristic (e. E=ST = tc'(0)E7. (5.5. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x). for a random walk: Corollary 5. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1).
Usually.5 defines a new MAP. some extra conditions are imposed. and the family {f LEE given by Theorem 2.f (x) tyo t provided the limit exists. 0) = h(i )( 1 + ttc(a)). see. Given a function h on E. G is defined as Gf (x) = lim Exf (Xt) . In view of this discussion . xEE .. s) = ea8h(i).6) We shall not exploit this approach systematically. however. let ha(i. 0 Proposition 5.e. St)} be a MAP and let 0 be such that h(Jt) OStt. this is.6. St) } as follows.s.6. u forsEE).5. gha(i. From (5. Remark 5. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. V.3b and Remark VI. 1) (i. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i). 0) = n(a) h(i). (5. this leads to h(i) + tcha( i. in particular that f is bounded. we take the martingale property as our basic condition below (though this is automatic in the finite case). For t small .for the present purposes. however. First.5) is a martingale . h(Jo) Lo is a Px martingale for each x E E.10 Let {(Jt.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.. Jt = (s+t) mod 1 P8a.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E.9 The condition that (5.4 that { h(Jt) easttK(a) L o (5.(9) {Lt}t>o = . Then {Lt } is a multiplicative functional. An example beyond the finite case occurs for periodic risk processes in VI. where {Jt} is deterministic period motion on E = [0.
MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . 1 + q(b . u Theorem 5. Bi [0] Remark 5.7) In particular. (5. qij = r.10 is given by P = eK(e) Oh e) F[e]Oh('). 0<b<oo.7(dx) Bij [0] Bij(dx) in the continuous time case . this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0.. if vi(dx) is compound Poisson. Ai = µi + 0Q. one can directly verify that (5.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. That 0 < qij < 1 follows from the inequality qb <1. Bi(dx) = Bi(dx).St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. ^i = of qij Bij [0] 1 + qij ( Bij [0] . 0 < qij < 1 and Bij [0] > 0. Bi.11 below in the finite case. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). 0<q<1. .1) . then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].5. We omit the details. in the discrete time case.c(0)e = tc(0)e .ic(0)e = ic(0)Oh e) h(e) .1) eft ea' f ij (dx) = Hij (dx) Hij [0] .tc(0)e = 0 .. In the infinite case . Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. vi (dx) = f3 Bi(dx) with .Qi < oo and Bi a probability measure.1) holds for the P.11 Consider the irreducible case with E finite.(0)j. In particular. Then the MAP in Proposition 5. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .
tc(0)I. are probability measures .46 CHAPTER II. it follows that indeed the normalizing constant is H1 [0]. since Hij. (dx) of a process with stationary independent increments follows from Theorem 4. in continuous time ( 5. this implies k[a] = A 1 ) (K[a + 0] .tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] . H1. is absolutely continuous w. In matrix notation .K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) . Now we can write K[a] =A+A ) ( K[a + 0] .8). a = 0 in (5.8. F:j with a density proportional to eei . Jl = j] :(Yi E dx.. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. Here the stated formula for P follows immediately by letting t = 1. This shows that F. Yi E dx.13) for matrixexponentials .tc (') (0) corresponds to the stated parameters µ. Letting a = 0 yields the stated expression for A. v= .e) Consider first the discrete time case . Jt = j] = hie) .8) h(.8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. . v. 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' . Hence the same is true for H=j and H.r. Similarly.Bay [0]) That k(') (a + 0) .t.11.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() .. . Jl = j) = Ei[Lt.. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. (dx). this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. Further Fib (dx ) = P=(YI E dx.
< x) = 11 (S. [226] and Miller [260]. Conditions for analogues of Corollary 5.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . i.Bij[0]) = hjel)ijgijBij[0](Bij[a] . 0]. however. 7+ < oo). THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. Note that G+ is concentrated on (0. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S.7). For the Wald identity in Corollary 5. Though the literature on MAP's is extensive. 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]. 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). has no mass on (oo.. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. hardly a single comprehensive treatment. an extensive bibliography on aspects of the theory can be found in Asmussen [16].1) = Aij4ij(Bij[a] .)Ajjgij(Bij[a+0] . Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. and is typically defective.6.6.3 for an infinite E are given by Ney & Nummelin [266]. there is.e. which. is slightly less general than the present setting.. h. . see also Fuh & Lai [149] and Moustakides [264]. [262] in discrete time.1). oo). the literature on the continuous time case tends more to deal with special cases.+ < x. however. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. [261]. [225].
(3B(x ) = pbo(x) on (0.d. the sum of all the ladder steps (if rl > 0.00 ).5 below.T+ > t)dt = E f 0T+I(St E A) dt. G+ is given by the defective density g + (x) =.1. For the proof of Theorem 6.48 CHAPTER K.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+.ST+(1) and so on. To illustrate the ideas. In any case. 6. they have a semiMarkov structure (but in complete generality. Theorem 6 .e. 0].1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A.i. the second ladder height (step) is ST+(2) . a fact which turns out to be extremely useful. and the maximum M is the total height of the ladder. Thus. In simple cases like the compound Poisson model. at present we concentrate on the first ladder height.e. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr.1. The main result of this section is Theorem 6. 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+. R+ is concentrated on (oo. On Fig. it follows that for g > 0 measurable. the dependence structure seems too complicated to be useful). Here bo(x) _ B(x)/µB. we shall first consider the compound Poisson model in the notation of Example 1. the ladder heights are i. g(y)R+(dy) = E f g(St)dt. there are only finitely many).2.B(x) denotes the tail of B. The first ladder step is precisely ST+. i. o 00 (6. = ST+(1) Figure 6. see Fig. by approximation with step functions . where basically only stationarity is assumed. oo). has no mass on ( 0.. i. 6. In other cases like the Markovian environment model. Recall that B(x) = 1 .1.. Also. 0 f T+ (6.2) . which gives an explicit expression for G+ in a very general setting.
ST_t. since the distribution of the Poisson process is invariant under time reversion. has the same distribution as {St}o<t<T.O<t<T) = P(STEA.2. That is.ST<St.O<t<T). P(STEA.T+>T) = P(STEA. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. St S* t a Figure 6.0<t<T) = F(ST E A. THE LADDER HEIGHT DISTRIBUTION Lemma 6 . ST < ST_t.6.2 R+ is the restriction of Lebesgue measure to (00. 0]. . 0 < t < T. 49 Proof Let T be fixed and define St = ST .St<0. see Fig.ST<St.2(a): T+ > t Figure 6.2(b): r+ < t Thus. 0 < t < T) P(STEA. 6.
E A} precisely when r+ > t. and since the jump rate is /3. oo). The probability of this given { Su}u<t is B(A .. Fig. U + St_ E A.50 CHAPTER II. cf.r.. T+ > t] 0 _ /3 f E[B( A .St _)I(r+ > t). it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time .3 where the bold lines correspond to minimal values. But since St 4 oo a.2) in the last).St _). oo).St). this is just the Lebesgue measure of A.y) (here we used the fact that the probability of a jump at u t is zero in the second step. and (6. G+(A) = Q f 0 B(A .3 G+ is the restriction of /3R+*B to (0. we get G+ (A) = f 00 /3 dt E[B(A . SOME GENERAL TOOLS AND RESULTS Integrating w.T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A . That is.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS.3 Lemma 6 . for A C (0. 6. . s. Figure 6.t dT.
i. cf. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0.3 yields g+ (x) = . 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. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup.4) are (ak.e.T+ < oo). {St+8 . 0 Generalizing the setup..s > 0). assuming basically stationarity in time and space.1.e. as a point process on [0.4). THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. h]} /h (by stationarity. obviously. The marked point process . In the stationary case.s. Nt St =>Uk k=1 t where Nt = max{k = O. this does not depend on h).z)B(dz) _ f I(x < z)B(dz) _ f (x). 6 ..) where ak = Ti + • • • + Tk . oo)..M o 08 shifted by s is defined the obvious way. i. The traditional representation of the input sequence {(TT. . we define the arrival rate as E# { k : ak E [0 .Q f r+(x . 2.:T1 +•••+Tk <t}. this is equivalent to the risk process {St*} being stationary in the sense of (6. Uk) for those k for which ak . The points in the plane (marked by x on Fig. the first component representing time (the arrival time o. Uk) (k = 1.6. Fig. U k)} k=1 a is as a marked point process M *.S8 )t> o = {St }t>o for all s > 0. . 6 . 4 (the points in the plane are (ak . 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. Lemma 6.. oo) x (0..1 With r+(y) = I(y < 0) the density of R+.* ) and the second the mark (the claim size Uk ).
Sigman [348] for these and further aspects of Palm theory. o. This more or less gives a proof that indeed (6. Uk) k=1. = 0 . Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t.4 Given a stationary marked point process M*. Oh becomes the approximate probability F(ri < h) of an arrival in [0. Example 6 . where TI = 0.e. of (6. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . V(M* o eak ). As above . h] and the sum approximately ^o(M*)I(ul < h). e. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. .QiBi(dx). the r. The two fundamental formulas connecting M* and M are Eco(M) = aE E.2. k: vk E [0. i. 0. vi(dx) = . SOME GENERAL TOOLS AND RESULTS M* U.. See.g. We represent M by the sequence (Tk. h.52 CHAPTER II. letting h J. Section 5) which has pure jump structure corresponding to pi = a = 0. h] Eco(M*) = 1 E f co(M o Bt)dt.5) represents the conditional distribution of M* given vi = 0. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. most often one takes h = 1).4 Consider a finite Markov additive process (cf. and let T = T2 denote the first proper interarrival time .s. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6...5) does not depend on h..
*(0) with initial reserve u = 0 is p = /3EU0. the probability aij of Jt . 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 . Before giving the proof. .6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . 5 Consider a general stationary claim surplus process {St }t>o. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J.oo a. an arrival for M* occurs before time t + dt w. Note in particular that the Palm distribution of the mark size (i.p.= i. let U0 be a r.OF(x).O fo "o F(x)dx = . It follows that we can describe the Palm version M as follows .6 Under the assumptions of Theorem 6. and that p = 0EU0 < 1.e. Assume that St * . If Jt_ = i.p. aij for (i. Jt = j is iri(3i /. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution .O for i # j.s. qij when {Jt} jumps from i to j and have mark distribution Bij.6.OEU0. and by some additional arrivals which occur w. v. we note: Corollary 6. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. let the arrivals and their marks be generated by {Jt} starting from Jo = j. After that. A stationary marked point process M* is obtained by assigning Jo distribution Tr. the ruin probability .. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i.O for i = j and iriAijgij/.5. First choose (Jo_./.p. Jo) w. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = .
in (0.Mt). 105) shows that one can assume w.$St_ u.. oo)).0<u<t) = P(St EA.0<u<t) = P(StEA.. Proof of Theorem 6. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0. 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.. Let p(t) be the conditional probability that ST+ E A.(left limit) when 0 < it < t and is illustrated on Fig .s. oo) p(t) = P(St EA. oo) x (0 . 0).. are point processes on (oo . h. .Su_ <0. We then represent M by the mark (claim size ) Uo of the arrival at time 0. The sample path relation between { Su } and { Su } amounts to S„ = St ..e. Now conditionally upon At .Su< 0.0<u<tIAt) = P(St EA.Q_k and has size U_ k.Su<0.5). T+ = t given the event At that an arrival at t occurs . CHAPTER H. and the kth preceding claim arrives at time t . 2. the arrival times 0 < 0'1 < Q2 < . 6. . that M* and M have doubly infinite time (i.St<Su. has a very simple interpretation as the average amount of claims received per unit time ...0<u<tIAt) = P(St EA.o<u<t where a claim arrives at time t and has size Uo. oo ) and the arrival times 0 > 0_1 > a_2 > . in (oo..o.g.A.).5.54 By (6.5. the mark at time Qk is denoted by Uk. which makes an upwards jump at time .1] here the r . moves down linearly at a unit rate in between jumps and starts from S0 = U. (k = St}t>o 1.St*_ u. The last property is referred to as insensitivity in the applied probability literature. 0<u<t) = P(St EA.o. A standard argument for stationary processes ([78] p. { Su}0<u<t is distributed as a process {Su} . It follows that for A C (0. . Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt.St <.l.
NIt)dt . Uo]. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. Mt)dt = i3EL(A) o"o . 6. the left endpoint of the support is oo. Since So = U0. and we let L(dy) be the random measure L(A) = fo°° I(St E A. the support of L has right endpoint U0. Thus.6. A sample path inspection just as in the proof of Lemma 6 . THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6.5 where it = { St < Su. 6. In Fig. 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.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. G' (A) = 3 f P(St E A. and since by assumption St * oo a.5. cf.s.. Fig. t a oo.
2. 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. [147]. .56 CHAPTER II.5 is due to Schmidt & coworkers [48].6 is Bjork & Grandell [67]. [263] (a special case of the result appear in Proposition VI.1).
4 below . 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 . i. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . . For finite horizon ruin probabilities . and independent of {Nt}. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. St = uRt = EUi t. exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. i. say.d. U2.. 3). are i. Panjer's recursion ( Corollary XI. 4. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability.6) and simulation methods ( Chapter X). • the claim sizes U1. see Chapter IV.Chapter III The compound Poisson model We consider throughout this chapter a risk reserve process {Rt } t>o in the terminology and notation of Chapter I. and assume that • { Nt}t>o is a Poisson process with rate j3. • the premium rate is p = 1. Some possibilities are numerical Laplace transform inversion via Corollary 3.e. with common distribution B. Thus ..
of the claim surplus St .u . (b) Var St = t.f.t = E E [ U k k=1 k=1 Nt .t = t(p .Rt. cumulants .'s etc. [324].58 CHAPTER III.1 (a) ESt = t(13µ$ . 0 .1) = t(p .)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally.. (d) The kth cumulant of St is tf3p(k) for k > 2. We do not spell out in detail such generalizations. Furrer [150].1).g.t = fltpB . we get Ee8st = 00 e8t c` Ee8 (U1+. See e. m. e .1 is the expected claim surplus per unit time. (c) Ee8St = et" (8) where c(s) = f3(B[s] . For (c). we shall start by giving the basic formulas for moments.g. The same method yields also the variance as Nt Ne Nt Var St = Var E Uk = Var E ^ Uk Nt +EVar [ k=1 k=1 1 k=1 Uk Nt Var [Ntµs] + E[NtVar U] = 113µs + t13Var U = tf3pB2).+Uk)P(Nt = k) k=O e8t k=O B[s]k . where K(k) (0) is the kth derivative of is at 0. for (d) just note that the kth cumulant of St is tic(k) (0). Dufresne & Gerber [126]. and this immediately yields (a). Proof It was noted in Chapter I that p . THE COMPOUND POISSON MODEL Poisson risk process.t = E[Ntµs] . and that B(k)[0] = Pak). P = PAB = 1/(1 + rl) Proposition 1. A more formal proof goes as follows: Nt r Nt ESt = E > U k .6pBa). 1 Introduction For later reference. say stable Levy motion. Schmidli [319].. {Bt} a Brownian motion and {Rt} a pure jump process.1). Write pB^) = EUn' YB = Pali = EU.s.1) . and Schlegel [316].
2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. (a) No matter the value of 77. we need the following lemma: Lemma 1. v > 0. The connections to random walks are in fact fundamental. meaning that the increments are stationary and independent. then St 4 co. 1.Tk. (b) If 77 < 0. Here is one immediate application: Proposition 1. rather to view {St} directly as a random walk in continuous time. cf.3) is proved similarly. then St> SnhV>Snhh. so that {Sok } is a random walk with mean EUET = EU. obviously 0(u) = F(maxk Sok > u). however.Tk are i.Sok_l = Uk . where Tk is the time between the kth and the (k .V.1)th claim. . which is often used in the literature for obtaining information about {St} and the ruin probabilities. the Uk . The right hand inequality in (1. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1.1 = . St = oo. (d) If 17 = 0.1 is the same as if {St} was a random walk indexed by t = 0. then lien inft.1. then Snh . if t = nh + v with 0 < v < h. we have Sok . In particular.4. Indeed. Sn+0 .h < St < S(n+1)h + h. Proof We first note that for u.3 If nh < t < (n + 1)h. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. 2.i.S„ attains its minimal value when there are no arrivals in (u. II. we get a discrete time random walk imbedded in the claim surplus process {St}. and the value is then precisely v.. S„+V > S„ . u + v]. St = oo.. For example.3EU01 = 1µs where rt is the safety loading. Obviously... lim supt. We return to this approach in Chapter V. The point of view in the present chapter is. then St 00. For the proof.d. (c) If 77 > 0. In this way.
it is seen that upcrossing occurs at least twice. and < 1 for all u when 77 > 0. hence by induction i. u 307).. 169) stating that lim infra. h. There is also a central limit version of Proposition 1. if P(M > 0) = 1.s..p. THE COMPOUND POISSON MODEL Proof of Proposition 1. Thus using Lemma 1. h A similar argument for lim sup proves (a). or by a general result on discrete skeletons ([APQ] p. . and hence by the strong law of large numbers. p. (c) are immediate consequences of (a). The general case now follows either by another easy application of Lemma 1. This contradicts u St400.2. then {St} upcrosses level 0 a. Remark 1 . However.3. Considering the next downcrossing (which occurs w. Corollary 1. Notes and references All material of the present section is standard.1. 2h. lim supn_.. is a discretetime random walk for any h > 0.1(b)) that the assertion holds as t 4 oo through values of the form t = 0.3. it suffices to prove 4'(0) = F(M > 0) < 1. 0 Snh = 00. Snh/n a4' ESh = h(p . 1 since St 4 oo) and repeating the argument. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time.1.60 CHAPTER III. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. at least once. If rl > 0.4 The ruin probability 0(u) is 1 for all u when 77 < 0._.. Snh u = 00 (the lemma is not needed for (d)).o.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).1).2: Proposition 1.2. where the size of the portfolio at time t is M(t).. and (b).. Part (d) follows by a (slightly more intricate) general random walk result ([APQ]..5 The limiting distribution of St . is a discrete time random walk.. For any fixed h.t . {Snh}n=o. 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. 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 . {Snh}n=o.1.6 Often it is of interest to consider size fluctuations...
where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . i. which we henceforth refer to as the PollaczeckKhinchine formula. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. equivalently.6. the ladder heights are i. [APQ] Ch.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ).1) is not entirely satisfying because of the infinite sum of convolution powers. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . Thus . 11. The expression for g+ was proved in Theorem 11. Fig. Theorem 2..IIG +II)EG+ . 1 The distribution of M is (1. Summing over n. It is crucial to note that for the compound Poisson model. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ).just before ruin is again B0. the formula for the distribution of M follows . but we shall be able to extract substantial information from the formula. IV.6. Combined with i/i(u) = P ( M > u).e. nevertheless. We assume throughout rl > 0 or.1.34 or A. (2.1 provides a representation formula for 0(u). B(x)/aB. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime.1) representing the distribution of M as a geometric compound. p < 1. This follows simply by noting that the process repeats itself after reaching a relative maximum. As a vehicle for computing tIi(u).. 0 Alternatively. 1e.1. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). n=0 (2. d. cf. and we further get information about the joint conditional distribution of the surplus and the deficit. oo ).P) E PnBon(u) . Note that this . 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}.2. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 .
ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. . and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. Theorem 2 . see Schmidli [323] and references there. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. (c) the marginal distribution of ST(o)_ is Bo . (d) the marginal distribution of ST(o)_ is B0. in this setting there is no decomposition of M as a sum of i. (a) 11 (ST(o)_ > x. Theorem A1. ST(o )) given r (0) < oo is the same as the distribution of (VW.6. As shown in Theorem 11. In the risk theory literature.d. However. Beekman [61]. cf. W are independent. Asmussen & Schmidt [49]. 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 PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula.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. Again.6. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections.5.i. see for example [APQ]. cf. f +b (b) the joint distribution of (ST( o). ST(o) > y.2(a) is from Dufresne & Gerber [125]. Theorem 2. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. We assume rt > 0 throughout. V is uniform on (0. 2 The joint distribution of (ST(o )_. The proof of Theorem 11. where it requires slightly more calculation. there is a general marked point process version. (1 .V)W) where V.1 is traditionally carried out for the imbedded discrete time random walk. cf. 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. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). Feller [143] or Wolff [384].5. [62]. ST(o)) is given by the following four equivalent statements: B(z) dz.just after ruin.62 CHAPTER III.
p)pSe a ( l v)x = p( S . 1 . the formula for P(O) holds in a more general setting. also be seen probabilistically without summing infinite series . hence without memory.2 If B is exponential with rate 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.6. then. I. As shown in 11. a further relevant reference is Bjork & Grandell [67]. use Laplace transforms. 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. Thus r(x) = S(1 . Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 .1)1 00 ( 1 . however . B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations .p) E pn S n x n.p. 0 .p) = S .O)e(b0)x. Thus .1 e ax = n1 (n .e. the result follows . Let r ( x) be the failure rate of M at x > 0. and hence this overshoot has the same distribution as the claims themselves . Alternatively.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.3. Integrating from u to oo. But claims are exponential . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w.p. 3b Exponential claims Corollary 3. For a failure at x.0(u) = pe(aA)" Proof The distribution Bo of the ascending ladder height ( given that it is defined ) is the distribution of the overshoot of {St} at time r+ over level 0. The result can.
Corollary 3. then 24 1 V. (u) 35eu + 35e6u.p + f u Z(u . (3. E.y)G+(dy) For the last identity in (3. Then the first term on the r. (b) use stopped martingales . The case of (3.4) is similar (equivalently. 2 is one of the main classical early results in the area.4) zu P(M > u .p + G+ * Z(u) = 1 .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 .3).3.s.h.3)). T+ <00) (3. THE COMPOUND POISSON MODEL In VIII. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) . (Example VIII.1.3) Equivalently.2).2) Notes and references Corollary 3. 0 Proof Write o (u) as P(M>u) = P(S. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.3) below. u .g.3.4) can be derived by elementary algebra from (3.S. u + oo.+ = y yields P(M>u.+ <u. and conditioning upon S. (3. the survival probability Z(u) = 1 . (3.64 CHAPTER III. II.i(u) satisfies the defective renewal equation Z(u) = 1 . is ?7+ ( u). We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.1) For a heavytailed B. cf. (3.T+ <oo).y)f3 (y) dy.y)/3B (y) dy.1 p pBo(u).S.T+ <oo)+P(M> u.+ <U. just insert the explicit form of G+. A variety of proofs are available .+ >u. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx). and weights 1/2 for each.y)G+(dy ) = f U V(u .
p)s .p)2 3(1 .(3 . We omit the details (see. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. [APQ] pp. eau B(u) du = f PB 3PB SPB 0 o (3. (3.pBo[s] no (1 . 111112 or Feller [143].3. 191). Also (3. by analytical manipulations (L'Hospital's rule) from (3. Some relevant references are Abate & Whitt [2]. In fact.)3B[s]) (3. 0 Notes and references Corollary 3.s .4) can be derived by elementary but tedious manipulations. .s .1 Bo[s] = f oc. Corollary 3../3B[s] which is the same as (3.6) 00 = (I . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.7) and Corollary 3. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used.7) s +.g..5 The first two moments of M are 2 EM .PPB2) EM2 = PPB) + QZPBl 2(1 . Of course.p) E p"Bo[s]" = 1 . e. Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p. Griibel & Pitts [132]. see e.5 can be found in virtually any queueing book. [APQ] pp.3 is standard . which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3.5).8) Proof This can be shown.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 .4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du .Ps s(. g./3B[s] .f.7).5).(3B[s] 1 .3 .g.3 . for example.P)pB' (3.p = (1 . SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3. Bo of B0 as m e8u B(du) = B[s] . Embrechts. In view of (3.P)PB 2(1 . 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. 206207).p)s s /3 .
4) for Z( u) means f lhu Z(u) = 1. .u) [p(k .1)! k=1 u1 .u)]k1 ku+1) [/3( k .u) [N(k .9). Z^ =eR(k.u + 1 )]k = QZ(u) . THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3. eO('u) [)3(k .Q) k=0 k! E e0( = /32(u) .z/'(u) takes the form Z(u) L^J L.Q (k 1 k= n  [O(k .u)]k d 1 u) _ a) n ( du ( k! (1  .Q) 3e.)3(1 .9) shown for n .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.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 (3.3I( 0<y<1)dy Z(y)/3I(0<u.u)]k k! k0 The renewal equation (3.s. then p) 1: ep(k u/. differentiation yields Z(u) _ /3Z(u) ./32(u .66 CHAPTER III. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .9) follows for 0 < u < 1./3Z(u .h.3+ 18+ J0 Z(uy).u) a)Qea" + (1 . Assume (3.1 < u < n and let Z(u ) denote the r. of (3. For n < u < n + 1.6 If B is degenerate at p.1).1).
4 Change of measure via exponential families If X is a random variable with c.f.Qe(Bo[a] . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. 0 Notes and references Corollary 3. 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) .d. or equivalently BB[a] = B[^+ Repeating for t 54 1.rc(9) = .3B = . we set up . B9(dx) = B[9] B(dx).(9).r.g.4) . and thus (4.3) by t.a.4. F and c.g.3B[9].f. The answer is yes: inserting in (4.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11. of F9.a. co(a) = rc(a + 9) . it follows that Z(u) = 2(u) for n<u<n+1. (4. but will now be repeated for the sake of selfcontainedness. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.4.1 that c(a) = /3(B[a] .1) or equivalently.1) .(9) is welldefined.4) works as well.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. (4. We could first tentatively consider the claim surplus X = St for a single t.g. (4. we just have to multiply (4.2).2) (Here 9 is any such number such that r.f. say t = 1: recall from Proposition 1. and define rce by (4.2) shows that the solution is Ox [O]0]. in terms of the c.1) . Formalizing this for the purpose of studying the whole process {St}.f. The question then naturally arises whether ie is the c.
2 For any fixed T. and PBT) the restriction of PB to FT.r. then EBZ = E [Ze9ST _T"(9)I .d. (4. . The following result (Proposition 4. (4. for G E FT. (4. The identity (4. EeeBSt + tk(B) = 1.1) and multiply from 1 to n).6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} .3 Let T be any stopping time and let G E FT.68 CHAPTER III. it suffices to consider the case where Z is measurable w.g. Then the Xk are i. . v(Xi.t.2. and thus (4. Let FT = o(St : t < T) denote the o•algebra spanned by the St. Proposition 4.9) Proof We first note that for any fixed t.f..i. and dP(T) dP^T) That is. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . replications from Fe (replace x by xi in (4. Z is measurable w. But let Xk = SkT/n . (4. BB by (4.0e and claim size distribution Be.5) for the density. the corresponding expectation operator is E9.i.8) follows by discrete exponential family theory.d.(9)} (4.. Xn). with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .8) By standard measure theory. the PBT) are mutually equivalent on. G C {T < oo}. . Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity. in particular the expression (4. with common c. Ti(a)/n. t < T. and define 09. G].4).7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 .FT.7) Proof We must prove that if Z is FTmeasurable. (4. . G]. .r. THE COMPOUND POISSON MODEL Definition 4.3 and claim size distribution B.S(k_1)Tln. . = exp {BST .t.5) for the density of n i.FTn) = Q(SkTIn : k = 0. oo) governing a given compound Poisson risk process with arrival intensity.nr..1 Let P be the probability measure on D[0.1.10) . n) for a given n.Tic (0)} .
Now consider a general G.7 1 Some discussion further supporting this statement is given in the next section. .(Y) = 13(B['Y] . according to what has just been proved. Then G E FT. (4. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . The behaviour at zero is given by the first order Taylor expansion c(a) r.9) holds for G as well. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T. GT C_ Jr < T}.10).1) . 77 Thus. 5. c(a) is a convex function of a. Given FT. Ee [exp { BST +Trc(9)} I(G) FT)] = 1.r is deterministic. subject to the basic assumption ij > 0 of a positive safety loading. and hence (4. t = T .5. (0) + rc'(0)a = 0 + ES1 a = a (p . Thus. 5 Lundberg conjugation Being a c.f. the typical shape of rc is as in Fig.ST) + (T .g.1 It is seen that typically) a ry > 0 satisfying 0 = r. Letting T t oo and using monotone convergence then shows u that (4.9) holds with G replaced by GT. Thus by (4..FT]] = EB [exp { BST + Trc(9)} I(G)] .1) _ 1 + a.1(a). (a) rc (a) (b) KL(a) 'Y 'Y Figure 5.r)rc(9)}I .7) holds. Then GT = G n Jr < T} satisfies GT E FT.
2)) is 7 = 5/3. It is then readily seen that the nonzero solution of (5.3) cf.a = i(a + 7). THE COMPOUND POISSON MODEL exists . b[s] = 5/(b . Thus. e. Fig. Note that KL (a) = /L (BL [a] .4) yields /3L = b and that BL is again exponential with rate bL =.1 Consider the case of exponential claims. the claim surplus process has positive drift > 0. we write FL instead of F7.2 is B(7) = 1 + ^.s).70 CHAPTER III. 5.1(b).2) 7 Figure 5. An established terminology is to call y the adjustment coefficient but there are various alternatives around. Fig. Taking T = r(u). (5. G = {T(u) < oo} in Theorem 4. and (4.3.1(b). the Lundberg exponent. Equation (5. 5. we further note that ( 5.g.3.1) .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 . 5. Thus B[7] = 6/. u It is a crucial fact that when governed by FL. (5.2 s As support for memory. (5. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes.3.1) is precisely what is needed for one of the terms in the exponent .1) (or (5.4) ELS1 = #L(0) cf. an equivalent version illustrated in Fig. Example 5 . .
G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.3. Proof Just note that e(u) > 0 in (5.(u)} . we can rewrite this as 0(u) = e"ELe7^(u).3 takes a particular simple form.t.ascending ladder height distribution and µ+ its mean.(oo) (in the sense of weak convergence w.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .6 ). (5.7) 0 and all that is needed to check is that ( 5.7) is the same as (5.1p . take first 0 = ry.8) .6) Proof By renewal theory.5.P Y j o' xeryxOB (x) dx /3k [Y] . where C . 0 Theorem 5 .r. Since a7' is continuous and bounded. Letting e(u) = ST(u) .e7x)G+(dx). (5.4). we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 .G+L)(x)) dx ry^+L) J 00 f 0 (1 . (5.+ E A} in Theorem 4. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. ST+ E A] .1e.1 (5.Ce7u as u 4 oo. T(u) < oo] . To this end.2 (LUNDBERG'S INEQUALITY) For all u > 0.1. G = {S. see A .5). V)(u) < e7u. V) (u) = P(T(u) < oo) = EL [exp {ryS. T = T+. which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx. e(u) has a limit i.u be the overshoot and noting that PL(T(u) < oo) = 1 by (5. Then P(ST+ E A) = EL [exp { 7S?+} .5) Theorem 5 . PL ) with density 1 .
7). we get L where 00 (1 .11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.1) (5. or equivalently of how close the safety loading 77 is to zero. THE COMPOUND POISSON MODEL In principle.4 Consider first the exponential case b(x) = Seax.8) yields +L) J0 xel'B ( x) dx (5. but some tedious (though elementary) calculations remain to bring the expressions on a final form.4).")G + (dx ) = 1  J0 00 3B(x) dx = 1p. this solves the problem of evaluating (5. Noting that SIG(L)II = 1 because of (5.3 (this was found already in Example 5.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .72 CHAPTER III. u .1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one. that 7 = S . Then 0(u) = pe(a_Q)u where p = /3/S.12) Example 5 . 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. of course.1 above) and that C = p. From this it follows. Using (5.10) VW = JI c* e° (x) dx = a (B[a] .e.1 . (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)
.
Though of course it is based upon purely empirical grounds.1. Letting Bo be the stationary excess life distribution.80 CHAPTER III.3* /S.p . we shall represent this situation with a limit where /3 T fl but B is fixed./3)] 1 .Ps(/3max . That is. THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 . p* _ . numerical evidence (e.g. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. we have according to the PollaczeckKhinchine formula in the form (3./3)PBo PB . cf. Notes and references The approximation (7.P .(bA*)". Proposition 1. and hence the ruin probability approximation is b(u) e(bAln)u. but has an obvious interpretation also in risk theory: on the average. heavy traffic conditions mean that the safety loading q is positive but small.2) was suggested by De Vylder [109]. (/3max .(3)2 P PB 2µB 2µB Letting /3* = /3/P.p + p { 1 1p ti 1 .1. Mathematically. Proposition 7. 1924.PBo [s (/3max . or equivalently that /3 is only slightly smaller than /3max = 1/µ8.s(/3max .b(u) = p*e. the premiums exceed only slightly the expected claims.3 and Corollary 3.3 )1 } _ 1p 1 . [174]) shows that it may produce surprisingly good results.7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 ./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 . the approximating risk process has ruin probability z. Grandell [171] pp.p = (/3max 0)µB.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) .1 As /3 f Nmax.)3 )PBo µB  .
It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce.2. Numerical evidence shows that the fit of (7. light traffic is of some interest as a complement to heavy traffic . then P(u) 4 e6„ Proof Write z'(u) as P((/3max . We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . [APQ] Ch. we shall represent this situation with a limit where 3 10 but B is fixed./3)u * v. while the approximation may be far off for large u. VIII).l3)M > (/3max . but has an obvious interpretation also in risk theory: on the average . These results suggest the approximation Vi(u) e6(0_.Q T /3max. That is .ze a2unµB laB (7.B AB ) 6()3max _'3) = However .7. In the setting of risk theory. light traffic conditions mean that the safety loading rl is positive and large .1 1 . Mathematically.g. Of course. . in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . and hence 2µ2B 1 . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . the term light traffic comes from queueing theory. 7d The light traffic approximation As for heavy traffic . obviously Corollary 7.2 If . Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. However . This follows since rl = 1/p .p.3) is reasonable for g being say 1020% and u being small or moderate. the premiums are much larger than the expected claims ./3)u). or equivalently that 0 is small compared to µB . VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. the first results of heavy traffic type seem to be due to Hadwiger [184].ryu . as well as it is needed for the interpolation approximation to be studied in the next subsection.4) suggested by the Cramer Lundberg approximation and Proposition 6.0)u. 2 provides the better mathematical foundation. u * oo in such a way that (3max .p _ 2rl11B PB p.
5) follow by integration by parts.u)+. u Note that heuristically the light traffic approximation in Proposition 7. ( 3 J O B dx. Light traffic does not appear to have been studied in risk theory.(3 10. Again. Indeed. Omax max m. U > u] = /3iE(U .5) u Proof According to the PollaczeckKhinchine formula.T > u).u.T > u) = J o" B(x + u)/3eax dx . The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . 10 ( u The alternative expressions in (7. see Daley & Rolski [96].82 CHAPTER III. i. 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 . . the Poisson case is much easier than the renewal case.3 is the same which comes out by saying that basically ruin can only occur at the F(U . For a more comprehensive treatment. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. by monotone time T of the first claim .e. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. THE COMPOUND POISSON MODEL Proposition 7. En'=2 • • • = O(/32) so that only the first terms matters.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. 0(u) /3 J B(x)dx = /3E[U . 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. Asmussen [19] and references there. Sigman [347]. [97]. (7. z/' (u) convergence P(U . cf.3 As .= 1 aJ 1 a 0+ 1 = = p.
Thus . where further references can be found . 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. .Wmax f(x ) dx + pee6mQ. however.VHT) ( ax QmQ ) h (B) ( . the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ].3 and use similar notation for %(B) (u) = (u). Notes and references In the queueing setting . . ) M. f / Qmax B(x)dx 00 eQmaxxdx 4/ Qmax 00 QmaxQ amaze" and the approximation we suggest is J B(x) dx = cLT(v) (say). 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1).O0 M. we see that the following limits HT) (u'). (7. Let OLT) (u) denote the light traffic approximation given by Proposition 7.6) (1p) The particular features of this approximation is that it is exact for the exponential distribution and asymptotically correct both in light and heavy traffic. with rate 1/µB = /3max. z/i(E) (u) = pe(QmaxQ)u. (U).O(E)(u) 1 (1 . "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. The adaptation to risk theory is new./3)) . that is.3n. even if the safety loading is not very small. one may hope that some correction of the heavy traffic approximation has been obtained. no empirical study of the fit of (7. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say).8. Substituting v = u(.x .6) is . Instead. _(E) (u). COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . 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. [84]. to get nondegenerate limits . available.3). Another main queueing paper is Whitt [380]. Al . B(2).
Proposition 8. whereas (consider x2) B(2) has the larger variance. we shall need various ordering properties of distributions. Bill is said to be convexly smaller than B(2) (in symbols. A weaker concept is increasing convex ordering: B(1) is said to be 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. we can assume that 1) < St 2l for all t. Rather than measuring difference in size. then . u Of course.2 If B(') <j. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. Proposition 8. In the literature on risk theory. THE COMPOUND POISSON MODEL To this end. Proof According to the above characterization of stochastical ordering. then i.' 1)(u) < V)(2) (U) for all u. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. Taking probabilities. or the existence of random variables U(l). U(2) distribution B(2) and U(1) < U(2) a. the proof is complete. Recall that B(') is said to be stochastically smaller than B(2) (in symbols.84 CHAPTER III. Here convex ordering is useful: Proposition 8.1 If B(') <d B(2). In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. B(2) and PB(1) = µB(2).ill(u) < V)(2) (U) for all u.6. B(' <. U(2) such that U(l) has distribution B('). XI. Finally. this ordering measures difference in variability. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. we have the convex ordering. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. In terms of the time to ruin. B(2)) if f fdB(1) < f fdB(2) for any convex function f. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. B(') <i.1 is quite weak.s. . for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. 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 Bill = B(2). this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. cf.
. Hence by the PollaczeckKhinchine formula .5 If '0(1)(u) < p(2) (U) for all u and a. A general picture that emerges from these results and numerical studies like in Example 8.(1) (. Corollary 8. B(2).u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1.. A first attempt would of course be to identify 'variation' with variance. Bo1) <_d Bo2) which implies the same order relation for all convolution powers. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities. This u implies that D <. say to p.e. Example 8.4) certainly supports this view: noting that.1 and µB at 1 so that the safety loading 11 is 10%. B(2). COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal. we have by Jensen 's inequality that E f (U) > f ( EU).6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean).1.3 provides another instance of this.3 If B(1) <. with fixed mean.2 is the following: Proposition 8. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. Proof Consider the light traffic approximation in Proposition 7.8. The problem is to specify what 'variation' means. Proof If f is convex. then /'(')(u) < 0(2)(u) for all u.6 Fix /3 at 1/1. A partial converse to Proposition 8. larger variance is paramount to larger second moment. from which the result immediately follows. and here is one more result of the same flavor: Corollary 8. and consider the following claim size distributions: B1: the standard exponential distribution with density ay.. Then V. B. (D) (u) < O(B) (U ) for all u. then B(1) <. The heavy traffic approximation (7.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8.4 Let D refer to the distribution degenerate at 'LB .
1358.001 u0. We return to ordering of ruin probabilities in a special problem in VI. Note to make the figures comparable.9A2e'2r where A. 1/)(u.01%. = 0. and this is presumably a consequence of a heavier tail rather than larger variance.4142. Pellerey [287] and (for the convex ordering) Makowski [ 252].e'\1x + 0. 11 Notes and references Further relevant references are Goovaerts et al. Kluppelberg [234]. 0. B3 the comparison is as expected from the intutition concerning the variability of these distributions. In terms of variances o2.. A2 = 3. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. van Heerwarden [189]. Let ua denote the a fractile of the ruin function. in comparison to B2 the effect on the ua does not show before a = 0. B3: the Erlang distribution with density 4xe2x.4. For B1i B2. and consider a = 5%.01%. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable.000. i. However. [166]. B4: the Pareto distribution with density 3/(1 + 2x)5/2. 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. 0. 9 Sensitivity estimates In a broad setting. all distributions have mean 1.86 CHAPTER III.1%. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!).e. One then obtains the following table: U005 U0.) = a. 1%. B. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.lA. the behaviour of which is governed by a parameter 9. 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).0' U0. A standard example from queueing theory is .
Then a p ao = 00 Qa/. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . we may be interested in a'/ap for assesing the effects of a small change in the premium. a0 as ao 80 19P . In the present setting. obtained say in the natural way as the empirical arrival rate Nt/t in [0. For example.01/2u. with 0 the vector of service rates at different nodes and routing probabilities. and s(9) the expected sojourn time of a customer in the network.19P a/ . Then ib = Pe(613)u. Thus at p = 1.e. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively .3. a2/t).2 Consider a risk process { Rt} with a general premium rate p. while /3 = j3 is an estimate. Then the arrival rate /3(P) for { R(P) } is )31p. where Q2 = fl ( l2 1113 / _ Ou2v)2. say estimated from data. Assume for example that 8 is known. increasing in u. the distribution of %3 0 is approximatively normal N(0„ Q/t).. i.1.(u) for large u. SENSITIVITY ESTIMATES 87 a queueing network. Thus.9. a/3 0 . where the partial derivatives are evaluated at p = 1. Similar conclusions will be found below.Ap). and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. it follows that ' is approximatively normal N(0. 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. t]. u Proposition 9. In particular . if = a e(6A)u. Proof This is an easy time transformation argument in a similar way as in Proposition 1. Then if t is large . the premium rate p and the claim size distribution B. Example 9.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). Let R(P) = Rtli.
Proposition 9. u Now consider the ruin probability 0 = 0 (u) itself.4). Viei '0(. we cannot expect in general to find explicit expressions like in Example 9. ^) .6 below for some discussion of this assumption). (9.88 CHAPTER III.0 = t/'(u) and the CramerLundberg constant C. this intuition is indeed correct. /3 yields w e(e + Y. we can rewrite the Lundberg equation as w(9+ y. so that heuristically we obtain '00 50ryu = Coe"u . THE COMPOUND POISSON MODEL As a consequence. 5) (Q+'Y)[we(0+7.2) (see Remark 9. various parametric families of claim size distributions could be considered.5) are similar.()YC = 1 +y/ /3 \ Q2 From this (9. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. and the proofs of (9. (3+'y)PC (0+7. but must look for approximations for the sensitivities 0. Similar notation for partial derivatived are used below. (. Consider first the case of 8/8/3: . namely that of a twoparameter exponential family of the form Bo. Consider first the adjustment coefficient y as function of 3. However . for the ruin probabilities .3. 3) ( 9 . Of course.w(O.3 70 = 'Ye = = 7 /3(1 we(e +'y. 4) (9 .1 or Proposition 9. In the case of the claim size distribution B. (9.()(0 +'0) ' (9 . and write yp = 8y/8/3 and so on .g. ()} p(dx) .Owe (9. 9. () = log(1 + y//3).((dx ) = exp {Ox + (t(x) .3) follows by straightforward algebra. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation .^)] 1(/3+y)we(9+'y.()wC(e.(/3 + y)we(9 + 7. but we shall concentrate on a special structure covering a number of important cases. x > 0 (9. () Proof According to (9.r. Differentiating w.t.10) below.()^ 1 . it suffices to fix the premium at p = 1 and consider only the effects of changing .3 or/and B.6) As will be seen below. e.3.uypCe7u urypO.w(6.
p)/C'y.x)B(x) dx + J U W(u .g. O} (9. Z(u)/u a C//3PF where PF is the mean of F.10) (9. z2(u) _ 1 ^ e'ri`i7i( u .9.3 (see in particular (5. Hence by a variant of the key renewal theorem (Proposition A1. u 0 Then Z = z + F * Z and F is a proper probability distribution . w((9 + a.we(9 . the proof is complete.9) (9.w(9. SENSITIVITY ESTIMATES Proposition 9. () .([a] = exp {w(9 + a.w(9. F(dx) = e'yy/3B(x)dx.4 As u oo. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3. () exp {w(9 + a.x) F(dx ) f u J C F(dx) = C as u 4 oo.8) (Section 5). () . (9. we multiply by e7" and let Z(u) = elt" cp(u). () . z2(U) = e7" J u b(u .C).QB(x) dx. But from the proof of Theorem 5. we get p(u) = J "O B(x) dx + J U O(u . Z= zl + z2 where zl (u) = e7u J m B(x)dx. Combining these estimates .8) Letting cp = e0/e/3 and differentiating (9.x). ()} .3) for z/'(u).(e"U = = wS(O. Further write de = [we (9 +'y. and alsoo zl(u) + 0 because of B['y ] < oo. () . Be. 0(u) = /3 Ju"O B(x) dx + f 0 0(u .x)B(x)dx.2 of the Appendix ). u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9.St (U) Ee. By dominated convergence. 11 For the following. ()] exp {w (O + y.w(O.x). we note the formulas Ee.8).12)).3(x) dx. ()} .4t (U)e°`U = which are wellknown and easy to show (see e.11) Ee. PF = (1 . BarndorffNielsen [58]).
this implies Z = z + F * Z. ()]e7vB(dy) 'fCd 7 c .9)(9.w (9. z = zl + z2.e7x/3 f 00 [t(y) . C)] (1 + 7 ) Proposition 9. ()](e7v .90 CHAPTER III.6C do 89 1p 8( 1p Proof By straightforward differentiation. ^)} [wc (0 + 7.w( (0.w(e.5 Assume that (9.wc(9. ()} 1z(dy) = f [t(y) . THE COMPOUND POISSON MODEL [we(e+7. 8^ ue7u. By dominated convergence and (9.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = .wc(O. ())B(dy) dx.6e7u f "o f[t(y) . ()]B(dy) dx.12) f exp {O y + (t(y) . ()]B(dy) dx x 0 0C T ON O . 8 8() 8( (9. F(dx) = e7x.x)f3 f ^[t(y) . ^) . u Z2(U) = e7° f u ^/i(u .we (0.wc (O. Then as u > oo.1) B(dy) 'f '[t(y) . oo z2 (u) f C . 0 x Multiplying by e7" and letting Z(u) = e"uV(u). 01 (i+) do = +'Y. )}B(dy)• Letting cp it thus follows from (9.8) that cp(u) .w(0. 2 z 07P N ue7u (3C de . C) .x).11).w( (0.QB(x)dx.2) holds.
16) (9. by inserting in the above formulas.a/35a&y' ' (9. ())B(dy) < oo. and the proof of the first one u is similar.ry) 5a1 cry (5 .pa+1 . Here (9.. () ='I'(t. we (9.12) takes the form y) alp a .(log r(a) a log S)} • r(a) 1.2) holds with p(dx) = xldx.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 . 9 = S.rye) S 5rya. Z(u) /3C 91 o c'o e11(t (y) . () = log r(a) . Example 9.12) follows. () = C/9 = a/S.. that C = a./35' a/i'y + aryl 625ry.18) (05 + 57 _'3_y . .w((9. It follows after some elementary calculus that p = a)3/5 and. U 7µF from which the second assertion of (9. ue_Yu 'C2d( 8a 8( 1 p .) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.1 .Sry a/32 + a/37 + /37 .6 Consider the gamma density b (x ) = Sa xa.C log(9). SENSITIVITY ESTIMATES as u 3 oo. We get w( (0. < = a.13) (9.QS 1 . and also zj (u) 4 0 because of f Hence.a log S = log r(c) . a /(S .yu/3C2do u86 89 1p' az/) = 8z/.17) (9. w(e.15) (9.3ary tog('Finally.9.Y)a+1 ' (9. ( 9.. t(x) = logx.1edz = 1 exp {Sx + a log x .
1 = eXP {c(C .22. for a < a* = z (. Straightforward but tedious calculations . () = Cc . which we omit in part .2a) } Thus the condition B[a*] > 1 + a* /. Be.2 log (0.3Ee.7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9. w(e.w(9. C = .2 ."62 . t(x) _ .S[a] = exp {w (9 + a. THE COMPOUND POISSON MODEL Example 9.l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.CZ try)} 1 C C2 try . ()} = exp {c (C .2) with µ(dx) = 2x3zrdx.9) (() . 9 = .92 CHAPTER III. C) = B = Yc = de = do = . C) . further yield .log c = 2 In particular.21og 2.([Y] .1 16 +ry c C22ry 2( = + 70 We (e.3.
we have assumed k = 2 and ti (x) = x. Van Wouve et al. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3. Also.10. Thus. B are assumed to be completely unknown. To this end.7 and references there.g. the results presented here are new.oo. Comparatively less work seems to have been done in risk theory.12) takes the form a = a 93 ar.a. (9.. in u which case the extension just described applies. [379] consider a special problem related to reinsurance.2) is motivated as follows. 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). and we estimate y by means of the empirical solution ryT to the Lundberg equation.. However .1) . queueing networks) are typically much more complicated than the one considered here. sj=1 and let YT be defined by IKT('ryT) = 0. for which we refer to X. we can just fix k . Finally if k = 1. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. the models there (e. Note that if NT = 0. thus. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). However.3C2de 1p' z a = c . by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . the main tool is simulation.8 The specific form of (9. ae t 1lEY u S _ . ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. Thus. to our knowledge. then ryT < 0. in which case we can just let t(x) = 0. or Ct(x). let NT 16T = ^T . the exponent is either Ox. That it is no restriction to assume k < 2 follows since if k > 2.cue_7u)3C P Remark 9.2 of the parameters. BT [a]= NT ^` e"U. . kT (a) = /T (BT [a] .+UNT) > 1.. then BT and hence ryT is undefined. In general. and hence explicit or asymptotic estimates are in general not possible. if 1 PT = /3TNT(U1+.
: N 0.: N ()3.1)2 + E[27] . N ( n[7]. Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . B[27] . For the proof.2 As T * oo.3T ..a BT[7] I B[7] I + . a2 where a2 = /3r. More generally.)vl+ N CO.B[7]2 n Hence ( 10.B[7]) 0+ Iv/o(b[y]. If furthermore B[27] < oo. since NT /T . then (10. B[2'Y]  /3T ) . 1) r. THE COMPOUND POISSON MODEL Theorem 10 .T y .3) Proof Since Var(eryU) = we have B[7]. (10.'Y .b[Yp'V21 T { (E[7] . it is easy to see that we can write \ V1 1 l _ .Q and Anscombe 's theorem. 16T where V1.(27)/K'(7)2.94 CHAPTER III./^ B[27] .1) . B [7]2 (10.1 As T 4 oo.B[7]) + B [7] . V2 are independent N (0.1) .2) follows from NT/T a4' .3/ T).If . 7T a4' 7. we need a lemma.i3)(B[7] 1) + (3(BT[7]  .1) 'YT .B[7]2 V2 .v.7 + (.2) rT(7) N N (0. vfoVFB[2y]. . Lemma 10 .'s.B[7]2 }) ( T 0 .
and the truth of this for all e > 0 implies ryT at 'y. 6"Y (10.KT(7) kT(7) K'(7) .E) < 0 < kT(7 + E) for all sufficiently large T .2. 0 are estimated from data . If ryT E (7  we have KT(7 .(ry . OT a 95 u 4 /3.4) and Lemma 10.e. To this end . Let 0 < E < ry..Q. y + E) eventually.10. 7T E (y . I. Theorem 10.c'(7) N (0' T (2(7) / N (0.'(y). By the law of large numbers. Then r.1 By the law of large numbers.E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r. Proof of Theorem 10. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7). NT i =1 n'(a) for all a so that for all sufficiently large T K7 .e) < 0 < r. °7IT) . ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.4) + E). where ryT is some point between ryT and ry. lcT(a ) 4 /c(a).E) < 4T(7T) < 4T(7 + E). Combining ( 10.(ry + e) and hence KT(7 .1 can be used to obtain error bounds on the ruin probabilities when the parameters . NT BT [a] Hence r.e. first note that e7TU N (e7U u2e27Uo'2/T) 7 .3). BT[a] 3 B[a]. it follows that 7T7 KT(7T) .'T(a) = 1 E Uie°U' a$' EUe "u = B'[a].
0) < 5. [197]. various alternatives have been developed. ft.ueryuU ". For example .info< „< t S.. satisfies b(.. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. i. with a tail of the form P(Y > y) .e.3 or equivalently p > 1/2 or 11 < 100%..96 CHAPTER III. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. Vt = St .5%).f.g.Q. Griibel & Embrechts [292].) = a (e.. This approach in fact applies also for many models more general than the compound Poisson one. V. Herkenrath [192]... if B is exponential with rate 8 so that ry = 8 . > 0 for some t E [Wn_ 1.T VIT where r7ry. 6 < 2. 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.Wn) are i . Hipp [196]. the nth busy cycle is then [Wn1. Deheuvels & Steinebach [102].T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. For this reason . U2. .. i .g.1 is from Grandell [170]. it means 2 (8 .i.1 : Vt = 0. = 1. Asmussen [23]) can then be used to produce an estimate of ry. 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.C1e"a ( see e.96 if a = 2.. wn = inf{t > W.d. Csorgo & Teugels [95]. Letting Wo = 0. and the known fact that the Y„ = max Vt tE[W„1. t]}. Mammitzsch [253] and Pitts. Frees [146]. Embrechts & Mikosch [133].e. Wn). Notes and references Theorem 10.
defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] .1) . The notation is essentially as in Chapter III. B[•] and mean AB. 'y) where c(a) attains it minimum value.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. Unless otherwise stated. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. In particular. the Poisson intensity is 0 and the claim size distribution is B with m.f.1 (the role of ryy will be explained in Section 4b). 97 . 0. The safety loading is q = 1/p . T) = P( /r(u) <T) \ = PI inf Rt <OIRo=u1 /\0<t<T PI sup St>ul 0<t<T Only the compound Poisson case is treated. Further let 'Yo be the unique point in (0. the premium rate is 1. exists.s.g.1 where p = 13µB. generalizations to other models are either discussed in the Notes and References or in relevant chapters. See Fig.
the time of ruin is T(u) and ^(u) = ST(t&) . 1 Exponential claims Proposition 1.9). 1 FL. 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 (1.1 The claims surplus is {St}. PL = 6/0 = 1/p > 1). Var[T(u) I T(u) < 00] = VarL T( U) . .(.. 7.(U) < 00] = ELT(u)ke'YS. PROBABILITY OF RUIN IN FINITE TIME Figure 0.1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0.(4. using that the overshoot l. In particular.u is the overshoot. 2 that E [T(u)k.) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k.2) Proof Let as in Example 111. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .5 .98 CHAPTER IV.t. By the likelihood identity III.r. FL and independent of T(u). (u) is exponential with rate 0 w.1) (1. we have for k = 1. E[T(u) I T(u) < 00 ] = ELT (U).
h.(yo) = 2V ./3) . Let 0 > yo be determined by ^c(0 ) = a.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 . we have by Wald's identity that (note that ELSt = t(pL . where = eBu I 1 . u + ELe(u) _ PL .1)2VarLT(u) + 2 Ca 1I VarLT(u).h.1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.1) .(PL .2) is aLELT( u) . Since Sr (u) and (PL . 1).6."(ry) = 26//32.1//32 (6/)3 1)2 26(/3u + 1)/(6 .s./3 .12 Thus the l.B = a.1)) ELST(u) ELT(u) (PL . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u). This means that /3(6/(6 .1.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.I (1. T(u) < oo] fora > r. is V1rLSr( u) +VarL ((PL .1 (6)3)2 which is the same as the r. Wald's second moment identity yields 2 EL (Sr(u) .1)T(u) are independent with QL the same mean . .1)ELT(u). which leads to the quadratic 02 + (/3 . of (1.2).6 + a)0 .6a = 0 with solution 0 (the .1)T(u)) = VarLe(u) + (PL .1)T(u))2 = UL where = s.0) .s.s. EXPONENTIAL CLAIMS For (1 .h. 0 Proposition 1.V/ is as asserted. the 1.
T+ Ti a t U T I 1 a i F.4.Y1 Y2 Figure 1.1 .. and M(u)+1 is the index of the ladder segment corresponding to T(u)..3) we have E [e«T(u ). (1. Using 5 = 6 . T(u) < oo] = e.v. But by the fundamental likelihood ratio identity ( Theorem 111. Ti.. T(u) < oo] = EB [exp {aT(u ) . T2 . Note that it follows from Proposition 1.v. Y(u) belonging to a convolution semigroup . PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0). Cf. . Y2.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. .100 CHAPTER IV..3 that we can write EeaT( u) = eeuEe 017(o).'s with rate 5.1 where Y1.9ST(u) +T(u)!c(0)} . 1.3. the result follows. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment .4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. are the lengths of of the ladder segments 2. Fig.0... are the ladder heights which form a terminating sequence of exponential r. More precisely... St Ti F.
If QT = N > 0. .v.T + • • • + UN.3 sin0 + 29) f3(0) = 1+/32/cos9. . 2..T. For j = 0.i (u.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. EN has an Erlang distribution with parameters (N.6(u) = Vfl/j l(Su.i. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . Let {Qt} be the queue length process of the queue (number in system.1.1 )!.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) .4. the conditional distribution of VT given QT = N is that of EN where the r.. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V. i. Proof We use the formula .I ex cos B cos j O dB fo " . then VT = U1. cf. T) to be evaluated by numerical integration: Proposition 1. 1). the following formula is convenient by allowing t. UN.T the service times of the customers awaiting service . let (cf. including the customer being currently served). T.e. . density xN lex/(N . . UN...6.3 Assume that claims are exponential with rate b = 1. Since U1 . U2.d.ST). where U1. Corollary 11.. Then V(u. .0. and exponential with rate S = 1..cos (u/. 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 . EXPONENTIAL CLAIMS 101 For numerical purposes .T.1(u.k + 1).T is the residual service time of the customer being currently served and U2 .T.T are conditionally i.1..T) 1 I fl(O)h(0) fdO where (1.
3(k +1)/ 2ei(k + l)6 (.i(k +1)e R [/3( klal/2e:0 (01 /2 e .112 l 1( k +1)/2 [ 31/ 2 cos(kO) . 912.1 00 ok+lR 00 j=k1 +1)/2e .cos((k + 1)0)] f3(0) 00 flk +1 > j=k1 3j/2 COS(jB) l)/2ei(k+1)e )3j/2eije = R)3(k+ (31 /2eie .1)] L _112 /(k+1)/2 [.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.8 ) yields F(QT > k + 1) .13(k +l)/2ei(k +1)9 R E .31 /2eie L 1)] 1 I/31/2eie . (1. f3(0) . in particular equations (1.1 R [.ie . 8789) 00 E aj j= 00 = 1.(31/2 cos (( k + 2)9) .)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. Then (see Prabhu [294] pp.cos((k + 2)9)] d9. and define tj = e(1+R)Taj/2Ij(2vT T).38).44). similar formulas are in [APQ] pp.102 CHAPTER IV.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) . let I _ j (x) = Ij (x).(31/2eie . k k2 + $k+1 E bj 00 t j . 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .
0(u. The first formula. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. equivalently. E Fk.. or. oo (u)31/2e^e)k = )3k z cos(k9) = R k.2. F(x. however.7) that _ [^ au ak+l (30 k L. going back to Cramer.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. 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). it follows as in (1. from the accumulated claim distribution N. k! k=O k0 i/z Co Uk ate" o'/z e . Ui < x I / (note that P(St < x ) = F(x + t.e = e' COS a cos(uf31/2 sin 0). Seal [327] gives a different numerical integration fomula for 1 . we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. 2 The ruin probability with no initial reserve In this section . the numerical examples in [12] are correct).T) which. is numerically unstable for large T. t )). Related formulas are in Takacs [359]. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. . u Notes and references Proposition 1. however. there are several misprints in the formula there. T) in terms of F(. The rest of the proof is easy algebra. T). and the next one (often called Seal's formula but originating from Prabhu [293]) shows how to reduce the case u 54 0 to this. expresses V)(0. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. t) = P .
2.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS.1 In formulas.(. . [v. ") } is at a minimum at time t.T)) 1 fT P(M(v.T))dv E^T I(M(v.t)= {Stv) < SM. T) = P(Tr(0) > T) = P(M(0. resp. f T lStv)} 0<t<T by a 'cyclic translation'.T))dv. v].b (0. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0.T)) does not {Stv)} depend on v.T].(6. co ). T].i.(0. See Fig. T T o where the second equality follows from II. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 . we define a new claim surplus process St StM NJ Figure 2. 1 1 .1. Stv^ _ Define M(v.0<w<t} St+v . Then 1 .3) with A = (0.T)dx.T) T F(x. 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. Proof For any v E [0.104 CHAPTER IV.
We claim that if M(0. cf. then M(v. If ST < 0. in which case there is a last time o where St downcrosses level u. or it occurs. v). T) = M(0. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .t)dt..T) = F(u+T.T)) dv f T I(M(0.ST on M (0. letting w = inf It > 0 : St_ = mino<w<T Sw}. v < t < T} n M(0.Sv. For example. 0<t <Tv}n{ST<ST Sv+St T+v. we can write M(v.2. this integral is 0 if STv) . T)) dv. T)). T) as {ST<St+ vS. T].2 10(u. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. T T o i =1 Let f (•. we can take v E (w E. Obviously. then i fT I(M(v.2. It follows that if M(0 . t) denote the density of F(•. Hence T TE f I( M(v. v<t<T}n{ST<STSv+St. ST > 0. T) occurs. v). THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v.T) occurs or not as long as ST < 0.xdx. t). there exist v such that M(v.T) occurs. v) = M(0. T Theorem 2 .Tt))f(u+t. T) occurs. T. . Indeed.T)f(I z /)(0. Fig 2. w) for some small E.v<t<T} = {ST<StSv.T) and Sv < 0 on M(0. v)) dv = ST T T o (note that the Lebesgue measure of the v for which {St} is at a minimum at v is exactly . v). where the last equality follows from ST < St on M(0. 0<t<v} = {ST < St . It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0..
106 CHAPTER IV. z > 0. The proof is combined with the proof of Theorem 111.2. The following representation of T(0) will be used in the next section.ST_ t_ and let A(z.T) = . 0 < t <T . {S t > z. E [t.(0.p. C*(z.2 Here o. For a fixed T > 0. ST_ _ z} .z.Tt))P(StE[u.3 Define r_ (z) = inf It > 0 : St = z}. which is independent of St and has the stationary excess distribution B0.T) = C(z. which occurs w.v. Hence P(ST<u) = 1 . t + dt] occurs if and only if St E [u.u+dt]). 2. Proposition 2.2.t). u + dt] and there is no upcrossing of level u after time t.2 .T)+ J0 T (1V. ST_ _ z}. Let Z be a r. {St > . u which is the same as the assertion of the theorem. ST_ _ z}. define St = ST . Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).T) = {St < 0. 0 < t <T. O(T .b(u. Proof of Theorem 111. 0 < t < T. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2.
Thus P(Sr(o)_>x. 2. Figure 2.3 But by sample path inspection (cf. (2. Fig. z + dz]. z + dz]) = P(A(z.T)) = P(Cx. {St }o<t<T have the same distribution . It follows by division by P(ST(o)_ E [z. we therefore have P(A(z. r(0) < oo) = 3R(z) dz JP(C(z.2. T + dT] I S7(o)_ E [z. z + dz]. u which is the assertion of Theorem 111.T). . THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T.1) yields P(ST(o)_ E [z.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z.T + dT]. ST(o)_ E [z. z + dz]. T(0) < oo) = OR(z) dz in (2.T))f3B(z) dz dT.2.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. z + dz].T)).3. and since {St}o<t<T.1) that P(T(0) E [T.ST(o) >y. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz.3).2. Proof of Proposition 2.T) = C*(z.1) z T . 7( 0) < oo) = P (C(z)) dT. A(z. Hence integrating (2.
I L Let ga(x) be the density of the measure E[ear(°).c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). cf.1) . who instead of the present direct proof gave two arguments. r(0) < oo.3. Tak'ecs [359]. (3.1 and the present proof is in the spirit of Ballot theorems.1 Eear( y) = eyr(a).3 was noted by Asmussen & Kl(ippelberg [36]. Let T_ (y) be defined as Proposition 2.2. Lemma 3.s. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T.T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = . 2. Proposition 2. z + dz]. Theorem 2.108 Hence CHAPTER IV. some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. see in addition to Prabhu [293] also Seal [326]. Note that T_ (y) < oo a. because of77>0.1.r(a).5 and one upon excursion theory for Markov processes (see IX.T+dT]). ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. Proof Optional stopping of the martingale I er (a) 9 t. T(0) < oo) = dTP(T_(Z) E [T.5a). 3 Laplace transforms Throughout in this section. z + dz]. Notes and references For Theorems 2. a martingale proof is in Delbaen & Haezendonck [103]. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z. Lemma 3 . In the setting of general Levy processes.1) where a > r.6.(3(B[r( a)] .(yo).2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . [329].
Corollary 3.x)ga (x) dx where za(u) = f.ga [b] 0 TO Using Lemma 3.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .ga [b] 1 . Z = y] = EeaT.5 f 00 o a/r(a) .g. b . (Laplace transform) of the ruin Corollary 3.3 ga[b] = c(b) Proof + b + a .3.ST(o)_ just before ruin .(v) = ev''(a).ic(b)/b x(b) + a eb"E[eaT(" ).r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a.r(a) oo Q f ex(br(a))dx f00 eyr(a)B(dy) x 0 Q f evraB(dy) e(a))dx 0 Q cc ev(br (a)) . (u .3. time T(u): u u Here is a classical result : the double m.4 E[eaT (o).r(a) = a [B[b] B[r(a)]] .°° ga(x)dx. Further by Theorem 111. r(u) < oo).1] evr(a)B(dy)[ b . It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.2. E[ear (o) I T(0) < oo . T(u) < oo] du = Proof Define Za(u) = E [eaT(" ). rr(0) < oo) = 1_ r(a) Proof Let b = 0.3.x)(a) B(dy)• Lemma 3 .r(a) b .f. Then by Proposition 2. Hence eb"du E[eaT(").2 P(Z E [y. the result follows after simple algebra. y + dy]. LAPLACE TRANSFORMS 109 Proof Let Z be the surplus . £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v. u .
mu ) ( 0 m < ML '(u) 1 m > rL.e..w ) v/. Theorem 4 . Then as u * oo. Proposition A1.6. the known results are even less explicit than for the exponential claims case. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .1) i. and take basically the form of approximations and inequalities. (u) t. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. For the second . i. uoo u using e.110 CHAPTER IV.mL > E T(u) < 00 ) 40.(u ) = o(u) a. That is. note that by Wald's identity u + EC(u) = ES. = (p . and hence a.. T(u)/u mL as u + oo.00 St = lim . Then given r(u) < 00. St/t 1 1/m. P = /3µB > 1. for any c > 0 P( Further.3). cf.1.1)Er(u) . we need the following auxiliary result: Proposition 4.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). By Proposition 111.s.2 Assume ri < 0.1 Assume 77 > 0.mu D 2 4 N(0. for any m T(u) u . PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model. The first main result of the present section is that the value umL. T(u) a.r(u) = Er(u) • ES.h(u. (4.2. This proves the first assertion of (4.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. Later results then deal with more precise and refined versions of this statement. For the proof.s.
the same conclusion holds with t replaced by r(u).2) follows immediately from u (4.^ N (o. Thus.t/m D (2) 111 . note first that ( Proposition 111. cf. T (u) < 00 J 0(u) e7'PL U \ I T u) . implying T(u) . though it is not easy to attribute priority to any particular author.s. apB ) . 1'r(U) . Theorem 7.6. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0.1 (by considering 0(u.mu (2) '• m3/2 µB 7 .1). For (4 .5) St .2 of [86]) and (4.mL >E By Proposition 4. which may be viewed both as a refinement of Theorem 4.h. According to Anscombe' s theorem (e.2.g.4. and as a timedependent version of the CramerLundberg approximation.1 The l. Tu) T( u) .3. 4).N(0. again Proposition A1. PL (•)+ 0.1 is standard. . proving (4.r(u)/m T(u) ti µB2) Z.3). of (4.1). T(u) < oo f / 00) e7uE L [e_7 (t1).1). If Z .7 6  11 Proof of Theorem 4. the result comes out not only by the present direct proof but also from any of the results in the following subsections. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl.1) is T (u)  U mL P( T (u) < I > E. Notes and references Theorem 4.1.mL U > E.mu m . T) for T which are close to the critical value umL). this can be rewritten as u + 1(u) . and (4.6µB2) Z v m (3µB2) Z.
Then the distribution of T(u) .4). one has 9 (r(u)_rnu) Ef (^(u)) * E.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u . with w2 as in (4. Let h(u) = E f (^(u)).w2) r. letting Z be a N(0. P because of ^(u') .5) For the proof. and thus in (4. g are continuous and bounded on [0. resp .e(u') oo w .um.r.3). and similarly as above we get E[f(^(u)) I Fr(u. using that ul/4 .t. PROBABILITY OF RUIN IN FINITE TIME Corollary 4. we need the following auxiliary result: Proposition 4.4 (SIAM'S LEMMA) If 71 < 0.L+YWLV'U) .)mu \ h(oo)Eg (r(ul) . oo ). e'°'/b (u. Using ( 4. Then h(u) 4 h(oo) = E f (6(oo)).6) whenever f.a C4'(y )• ( 4.ul/4. we get E[ T (u) .112 CHAPTER IV. S( u ) < ul/4] < ET(ul / 4) = O(ul/4). (oo. Proof Define u' = u .))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . E9(Z) (4.u1/4)I(S(u') > u1 /4) h(oo) + 0. Then for any y.mul h(oo)Eg(Z).VU T.f ( (oo)) .^(T(u')). then e(u) and r(u) are asymptotically independent in the sense that. we can replace T(u) by r(u').T ( u')] = E[ T ( ul /4 .l:(oo) (recall that rt < 0). oo). O .v.T(u') given F. Hence Ef (Vu )) 9 (T(u.6).3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).ST( u') = u1/4 .(u.
where we used Stain's lemma in the third step and (4. y u) < . The present proof is basically that of Siegmund [342].4. umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). Theorem 4. For refinements of Corollary 4.3 ery"z/i(u .3 in terms of Edgeworth expansions .7) to be valid is that T varies with u in such a way that y(T) has a limit in (.oo. 4b Gerber's time. see Asmussen [12] and Malinovskii [254]. define ay. e7v" y < ^'(7) (4 . y > k'(7) . just substitute T = umL + ywL in (4.7) To arrive at this .dependent version of Lundberg's inequality For y > 0. Notes and references Corollary 4 .7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . For practical purposes . ELe7E (") . Thus .5 '(u . also Hoglund [204]. PL(T(u ) < umL + ywL) 113 4 C4(y). . The precise condition for (4. Segerdahl 's result suggests the approximation b(u. WHEN DOES RUIN OCCUR? Proof of Corollary 4. 3 is due to Segerdahl [333]. see also von Bahr [55 ] and Gut [182].z/)(u .7) to be good.9) ( 4 . yy by 1 K. 0.1. in practice one would trust (4. oo ) as u * oo. Cf.4) in the last.T) Ce7"4 (T .8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML).yK(ay)• (4. y u) < e 7v" .5) and solve for y = y(T). u needs to be very large). 10) '5(u) .umL wI V"U u (4. CL Fig.(ay) = 17 7y = ay . however . that for the fit of (4. T(u) < umL + ywL f.
114 CHAPTER IV. a. dy) Notes and references Theorem 4 .Y' (u.8). Hoglund [203] treats the renewal case. Then ic(ay) > 0 (see Fig . yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). which shows that the correct rate of decay of tp(u.8). yu) is e 'Yyu/ . 5 is due to Gerber [156 ].1). For a different proof. .2. yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . see MartinLM [257] .yu ) = eayuEav [e . the bound a7y° turns out to be rather crude . From the proof it is seen that this amounts to that a should maximize ayic(a).t. we have rc(ay) < 0 and get (u) . yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). However. the point is that we want to select an a which produces the largest possible exponent in the inequalities.6. we arrive at the expression in (4. 0. yy is sometimes called the timedependent Lundberg exponent. Numerical comparisons are in Grandell [172 ].8 below . if y > 1/ic'(y). T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly.r. In view of Theorem 4.b (u.9): Proposition 4.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. and hence t.ay4(u)+ T(u)K(ay ).5.h(u. An easy combination with the proof of Theorem 111.3 yields easily the following sharpening of (4. u Differentiating w. who used a martingale argument. which may be understood from Theorem 4.7 i. and generalizations to more general models are given in Chapter VI. f Some urther discussion is given in XI.v"U. PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y).
For any a > yo.e. if we want EaT(u) . and in case of ruin probabilities the approach leads to the following result: Theorem 4 .c(&) = ic(ay) is < 0.8 If y < 1/ic'(ry). i.: T. (0) r1 (a) ' I.'(y ).e. yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma. yu) = e.(u. u 4 oo. then the relevant choice is precisely a = ay where y = T/u. T(u) suggests heuristically that l t/. [eT(u )K( ay).5. and ii(u) . Using Lemma 111.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. the choice of ay. T(u) < yu] . then ay > 0. As a motivation.i(u.ayC() . The traditional application of the saddlepoint method is to derive approximations.yu) c ay .6 with P replaced by Pay and FL by Pay. and b(u.ay a.^3 ]1/ Bay [lay .ay y 'Yay  ay .11) ' If y > 1/ r . Proposition 4.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. it is instructive to reinspect the choice of the change of measure in the proof. we have ryas = ay .4.2 yields EaT(u) u u r.13) . not inequalities. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations.. yu ) eaauEaye .. (4. (4. the formula 0(u.z. This idea is precisely what characterizes the saddlepoint method.. Ea . We thereby obtain that T is 'in the center' of the Padistribution of T(u).ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 . then the solution &y < ay of .ayuEay f eay^ ( u)+T(u)K(ay). (4.
The difficulties in making the proof precise is in part to show (4. Writing r(u) and W2 = I3ay{.B[ay] /ay &y y(ay .a)2 . V < 01 Ir 00 er(ay)"1'2"'x eyur.1)3 = (jB"[ay]l (Pay .ay) ay +. .4).1B[ay]1 ) y(ay . and the equation ic'(a) = 1/y is easily seen to have . The proof of (4.ay + ayl /BLay] .a) .1)3 = y3/3B"[ay].c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2. i B[7ay .c'(a) _ /3a/(8 . PROBABILITY OF RUIN IN FINITE TIME ry I i .13) rigorously.ay)K(ay) ay ayI&YI For the second term in (4.a.12) is 0 entirely similar.11) follows.3(5/(S .116 CHAPTER IV.I ay &y a ^c'(ay) a (1 +. it seems tempting to apply the normal approxiyu + ul/2wV.1. Then ic(a) = .13).1) under Pay mation (4.1) .7ruw2 Inserting these estimates in (4. a nr=. and in part that for the final calculation one needs a sharpened version of the CLT for t(u) (basically a local CLT with remainder term).ay ) r. Example 4.(j (1 .l'B)y /(Pay .9 Assume that B(x) = eay. (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 .13). (4. we get heuristically that Eay Ler (u)r(ay). where V is normal(0.(ay) _ y(ay . T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.
g. 2 = Var(Si ) the variance. (5. yu) when y < 1/ic'('y) = p/1 .1) .p. and next to note that such an approximation in particular implies that the first passage probabilities are close. 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 =. A related result appears in BarndorffNielsen & Schmidli [59]. . in discrete time: if p = ES..3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 .5.because the c.8 is from Arfwedson [9].i )( v s vc ('3 + s _2 / .ay)3 0 3/2 and (4. 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. 0 Notes and references Theorem 4. is undefined for a > 5)./4 ^y for 1/i (u.tcp) Lo {Wo ( t)}t>0 .11) gives the expression '31/4 ( ..1. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.f. then { __ . The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.. is the drift and o.= (s. c a 00. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .
Mathematically. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes). + {Wo(t ) . Indeed .e.1)) is inconvenient.p.p/c < St(p) < S((n+l)/ c + Pp/c. of which a particular case is the claim surplus process (see the proof of Theorem 5.1 below). where p is the critical premium rate APBTheorem 5 . we have o {i!t s: . 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. and this can be obtained under the assumption that the safety loading rt is small and positive.7c).tp). 0 .1 As p J. such that the claim size distribution B and the Poisson rate a are the same for all p (i.118 CHAPTER IV. However.3) whenever c = cp f oo as p 1 p.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.t} _ {W_1(t)} . We want an approximation of the claim surplus process itself. Lemma 111.z } {W_1(t )}t>o (5. for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.a = Snp) and the inequalities Sn )C . this is an easy consequence of (5.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}.1. St = EN` U= . (5. a2 =/3µB2) Proof The first step is to note that { WC (St P) ..1) with S. and consider the limit p j p. oo)). cf. 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.3.. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p. Letting c = a2/pp.2) t>o where p = pp = p . p. n/c < t < (n + 1)/c.
8 or [APQ] p. Corollary 5.5. 263) that the distribution IG(•.1 . has a continuous distribution.s. 1. (ua2 To2 op \ IPI > IG ( T .s.5). Because of the direct argument in Chapter III. (. and in fact some additional arguments are needed to justify (5. Corollary 5 .h.. since ti(u) has infinite horizon . 119 It is wellknown (Corollary XI.2 suggests the approximation u 0(u.4) Note that IG(. u).t. is IG(T.. u) =PIT( (u) < x) = 1 .f I \\\ J \ (5..7c. see Grandell [ 168]. . u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. ulpl /a2) = e2"1µl / or2. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. we obtain formally the approximation V. TS(u)=inf{t>0: WW(t)>u}. 196.(u) ti IG(oo.1. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. However. u) is defective when < 0. C. we omit the details . (5. and the r. the continuity argument above does not generalize immediately.s. ('.e. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. 199. 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.6) from Theorem 5. any probability measure concentrated on the continuous functions.6) This is the same as the heavy traffic approximation derived in III.u). w. (5.1. For practical purposes .2 As p j p.1 I 7= .( ^ I + e2( \ I .r.h.T) IG(Tp2/ a2).h. ulpI/a2). [169] or [APQ] pp.Ta2 /p2).5) Note that letting T * oo in ( 5. is 1/ip (ua2 /IpI.
120 CHAPTER IV. as 0 * 00 and that the U2 are uniformly integrable w. in particular for large u. However. In view of the excellent fit of the CramerLundberg approximation.. the simplicity of (5. the claim size distribution B9 and the premium rate p9 depends on 0. (5. The proof is a straightforward combination of the proof of Theorem 5. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. B0 * Boo. In contrast.t. Then as 0 _+ 90. However.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. See for example Billingsley [64].6 of [APQ]. as an example of such a generalization we mention the paper [129] by Emanuel et al. pt? 4 peo. Furrer. [169]. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75].r.00µB6 + 0.Pe. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . . pe . Theorem 5. on the premium rule involving interest. All material of this section can be found in these references.1 and Section VIII.6) therefore does not appear to of much practical relevance for the compound Poisson model.1. e. the B9. The first application in risk theory is Iglehart [207]. in Asmussen [12]. We conclude this section by giving a more general triangular array version of Theorem 5. that 00 4090. The picture which emerges is that the approximations are not terribly precise. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation.g. such that the Poisson rate Oe.5) and (5. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory.5) for the compound Poisson model which does not require much more computation. Assume further that 039µB6 < pe. we have ^A.5) combined with the fact that finite horizon ruin probabilities are so hard to deal with even for the compound Poisson model makes this approximation more appealing.Po = 09µB6 . Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. in the next subsection we shall derive a refinement of (5. For claims with infinite variance. and which is much more precise. and two further standard references in the area are Grandell [168].6) are presented.
/c(9 . 2. 3. and we want to consider the limit 77 10 corresponding to Oo f 0. this means the following: 1. 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]. it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. However . P9(r (u) < oo) = 1 for 9 > 0. .6.9(s) = Ico ( s + 9) . PB('r(u ) < oo) < 1 for 9 < 0.90] B(dx).T) = Peo(r(u) < T) for 90 < 0. 9o T 0. Since Brownian motion is skip free.6. 77 is close to zero. B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . this idea ignores (among other things) the presence of the overshoot e(u). 77 = 1/p .1 > 0. In this setup. 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. risk process with safety loading 77 > 0 correspond to 9 = 0 . Then r.Q (B[s] . The objective of the corrected diffusion approximation is to take this and other deficits into consideration. whereas there we let the given 3B. which we have seen to play an important role for example for the CramerLundberg approximation . Bo(dx) = B[eo]B(dx). this is because in the regime of the diffusion approximation . . In terms of the given risk process with Poisson intensity .'(yo) = 0 and let 90 = 'Yo.90) and the given risk process corresponds to Poo where 90 = 'yo.c(s) = . The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. and we are studying b(u.1) . claim size distribution B .s and p = /3µB < 1.ao (0) _ /c(s + 9 .90) .4. 0(0) = 0. For each 9. Determine yo > 0 by r.
.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process.(y) = 0. (6.S. IGu+u2.. The first step in the derivation is to note that µ = k (0) = r0 (00) . C . (.. 1) • Since L eatIG (dt. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]. C. i.3) this implies (take u = 1) Ego exp { . (01. bl IG(t81.122 CHAPTER IV. means up to o(u1) terms): . . 0o to.1) IG(x. _ ^(u) = ST .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. () where h (A. . tu2 ) i IG (t.e..u.2) . for brevity.7(u)/u2} eh(A.Varo S1 = f30Eo U2 = S1.C. (U. (.() The idea of the proof is to improve upon this by an O (u1) term (in the following. One has (6. 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. 9otc0" (0) = 0061 = ul. write r = T(u). 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. Vargo S. u) = euh(a .T) 1+u2 (6. C) = 2A + (2 . u) = IG(x/u2. the solution of r. PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x.1) . The corrected diffusion approximation to be derived is (u. Theorem 5.
3 = 0.2). of (6. p = 0. however. The solid line represents the exact value .z .h.1 + u2 I Indeed.3.h.exp { h(A.2 ). which is based upon exponential claims with mean µB = 1. and the dotted line the corrected diffusion approximation (6. A numerical illustration is given in Fig.7. we get by formal Laplace transform inversion that C 2 u.s. .d.1 below is exact. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r.v. To arrive at (6. we have p =. is the c. . .1 + 629.f.'yu /2)(1 + b2/u)} + Aug 1I J . 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. the r.5) according to (6.3).4.52/u where Z has distribution IG (•. of a (defective) r. 1. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} . But the Laplace transform of such a r. calculated using numerical integration and Proposition 1.ry2 . in (3) and (4). bl I IG I t +2 .2) is indeed o(u1). Note. however .5) Once this is established .1 As u + oo. In ( 1) and (2). The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3). the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6. (6.6.v. 6 .s. just replace t by Tb1/u2. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . distributed as Z . that whereas the proof of Proposition 6. 9o T 0 in such as way that C = Sou is fixed. 1% in (2) and (4). .
and all of the numerical studies the author knows of indicate that its fit at p = 0.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI..1 proceeds in several steps.7.124 0.EB 0 p ex p ( 7 S h ^)u .011 L1 60 T IM 11.aa1 .() Lemma 6. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.T1 00. (Inc 0s 0.W21 0.00 0.199 0.08 0. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34]. For further numerical illustrations.01 0. A51 7(SAT 3 3 h(X. see Asmussen [12].0 0..TI CHAPTER IV.^) .08 a. Similarly.111 W(U.1 W IU.4 may not be outstanding but nevertheless. OM 0. PROBABILITY OF RUIN IN FINITE TIME 0. the fit at p = 0. .114 0.05{ 0.u2 2u3 (e . Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small. The proof of Proposition 6.T) 111 0.OOIi O.1 It is seen that the numerical fit is extraordinary for p = 0.7 or at values of Vi(u) like 1% is unsatisfying.2 e.(061 0.T) 0.19)2 11 20 20 i0 T 1n0 Figure 6.07 0.
4 Ea. the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy . in Lemma 6. () + C and note that 2 KO (0) = 102.6) u U3 Lemma 6 . CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0. (6.C)C/u .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .6.co ((/u)) } Let 8 = (2a + (2)1/2 = h()..7) 2 2 . + a1b2 + .T (co (8/u) .(3)Eea LauT exp i 3J . 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.co (e) . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .4) that the r... 1 = PB(T < oo) = Eo0 exp 125 {(B . (6..1) h(A. 1 / Po(C(0) > y) dy EoC(0) x k EDUk + 1 k Eo[(0)k+1 EoC(0) _ (k + 1)EoU' EoC(^) _ (k + 1) Eo£(0) Lemma 6 .+ h (A.r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .2u (B3 .C2 = 2).2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S . () 62 Eeo exp u u2 J .h. exp ue } al 1J 3 exP I [2).(3) J t _ aa1T l + eh(A. the result follows.61a2T (B3 .s.00)(u +C)  'r (.
6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4).e h(aS)h (^^ 262 exp {_h(.2. 2 + 00 = .x.h. letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). yields +90 62 0 + O(u 3) 2u2 +O(u 3).. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6.h (A. 5 exp { _h(A) (1 + / y u J)) exp 1.() I 1 + u2 ) y .s. we get h(A. () by h(\. 2 and (6. and inserting this and 9o 2 = S/u on the r. l Lemma 6 . PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.() .4.\ + () 1 2 / . [2+ (2 .2 (^/2 + 3y9o + 390) + O(u3).\+ (2 (3 e 2u [ (2. () .6  d h(A.1 (y/2 + Oo)u .7) and using eh(a.2) for O(u) (indeed . yu/2). C) ( 1+ u2 The result follows by combining Lemma 6 . There are two reasons for this : in this way. Thus by Taylor expansion around ( = 90u. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6. 0 The last step is to replace h(A. and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6.2u [2A+ (2 3 .126 CHAPTER IV.S) d e 62 . Thus a2 y = 290 + O (u2).(2A + ()1/21 exp S h(A. yu/2) h(A.
4. ()} .7.(i+ 62 exP{ h(A.5 in Lemma 6. that is. the approach to the finite horizon case is in part different and uses local central limit theorems. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. The adaptation to risk theory has not been carried out.e.T) has not been carried out and seems nontrivial. 'yu/2) 127 ( i+ M pz^ exP { h (A. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. 0 1 Proof of Proposition 6. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. Fuh [148] considers the closely related case of discrete time Markov additive processes. the same as for the unconditional Lundberg process. i. () I 1 + u 2 ) } S 1 . () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. the 'typical' value (say in sense of the conditional mean) was umL. the analogous analysis of finite horizon ruin probabilities O(u. and to the Markovmodulated model of Chapter VI in Asmussen [16]. 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. In Siegmund's book [346]. The answer is similar: the process behaved as if it changed its whole distribution to FL. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. () (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. ()} 3 h (A. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. .1: Just insert Lemma 6.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. HOW DOES RUIN OCCUR? exp { h (x.
. In fact.r. then P(u) and FL coincide on .2 If B is exponential. r(u) < oo) .T. Recall that . (u) and satifying PL(F(u)) * 1.(u)_ and similarly the denominator is exactly equal to Ce7u. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u. ^(u) is exponential with rate 8 w. . u * oo.(u). Note that basically the difference between FT(u) and . and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + .(u)_ and ^(u) are independent .ST(u)}t> o is just an independent copy of {St}t>o).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.1 Let {F(u)}u>0 be any family of events with F(u) E F. .TT(u) _measurable. so in the in the proof.. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E .128 CHAPTER IV. FL As example. Proof P(u) (F(u)c) = F(flu)c.r. P(u) and rate = aL w.t. Theorem 7 . we give a typical application of Theorem 7.F.F. {ST(u)+t . stating roughly that under F(u).FT(u) = o' (T(u ).vi(u) Ce'Yu Corollary 7. the Poisson rate changes from .3 to . F(u)c] ti e' ru]PL (F(u)`) > 0. 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.3L and the claim size distribution from B to BL. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL. {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u).EL[e7S. In the exponential case. + TMOO ).t.(u)_ is that i. Then also P(u)(F(u)) + 1..(u) is not .
the queueing results are of a somewhat different type because of the presence of reflection at 0. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). Notes and references The results of the present section are part of a more general study carried out by the author [11]. Proof For the first assertion.(1 . who also treated the heavytailed case. see further XI. A somewhat similar study was carried out in the queueing setting by Anantharam [6].7.3. This is currently a very active area of research. From a mathematical point of view. however. take I(Tk < x) .eALx) M(u) k=1 u The proof of the second is similar.3 M(u) pcu) 1 .eaLx. the subject treated in this section leads into the area of large deviations theory. HOW DOES RUIN OCCUR? Corollary 7. .
This page is intentionally left blank .
A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA..1.1 (T1 = a1). Thus the premium rate is 1.i. D'2.7).Q. the one corresponding to the stationary case by 00)(u).1). r(u) the time to ruin. the claim sizes U1. . the Tn are independent. the distribution Al of T1 is A as well. . with the same distribution A (say) for T2. with Nt = # {n: Un <t} the number of arrivals before t. . We use much of the same notation as in Chapter I. and the one corresponding to T1 = s by 1/i8 (u). are i..Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . Proposition 1. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). see A.. U2. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory.. .. AA t*oo lim St = lim ESt t tioo t = p .Then no matter the distribution Al of T1i B.. with common distribution B.1 Define p = !µ. of the risk process form a renewal process: letting Tn = Qn .(1.. In the socalled zerodelayed case. {St} is the claim surplus process given by I. T3. and M is the maximum of {St}.d.
1) ENt/t + 1/µA. From this ( 1. the definition 77 = 1/p . If the environment is Markovian with transition rate A from on to off and u from OFF to ON.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. RENEWAL ARRIVALS lim E [St+a . 2). say at a. Thus. by the elementary renewal theorem (cf. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). CHAPTER V. t 4oo Proof Obviously.2 (DETERMINISTIC ARRIVALS) If A is degenerate. the .St] = a(p . OFF.132 Furthermore for any a > 0.Nt] * a/PA. For (1 .t. stating that E[Nt+a . This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). Here are two special cases of the renewal model with a similar direct interpretation: Example 1. and (1 . but the arrival rate in the ON state is . Proposition 1. 3) follows similarly by Blackwell 's renewal theorem.1) follows . Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth.1 of the safety loading appears reasonable here as well. Nt ESt = E E UI Nt t = ENt•pB .1 gives the desired interpretation of the constant p as the expected claims per unit time.a is really the accumulated claims over a period u of length a.0 > 0. Example 1 . after E. Nt + EVar U. The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. However . s + t µA PA 0 Of course. such that no arrivals occur in the off state. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. A.1). The renewal model is often referedd to as the Sparre Andersen process.
y)B(dy).4) fo Indeed. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. we feel it reasonable to present at least some basic features of the model. we note that the ruin probabilities for the delayed case T1 = s can be expressed as in terms of the ones for the zerodelayed case as u+8 z/i8(u) = B(u + s) + '( u + s ..} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .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.i. the first term represents the probability F(U1 . However. . Proof The essence of the argument is that ruin can only occur at claim times. S o<t<oo n=0.1. and then the whole process repeats itself).s < u)..s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo. in general the mechanism generating a renewal arrival process appears much harder to understand. the relevance of the model has been questioned repeatedly. A is phasetype (Example 1.. we have From this the result immediately follows. (an arrival occurs necessarily in the ON state.s. u For later use.d. INTRODUCTION 133 interarrival times become i.. initial vector (1 0) and phase generator 11 However. U1 .. integrate (1. (1.4) w.4) with phase space {oN.2.. 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 .T between a claim U and an interarrival time T. Therefore. the fundamental connections to the theory of queues and random walks. More precisely.r. and for historical reasons.oFF}. Ao.1. and the present author agrees to a large extent to this criticism.t. if for nothing else then for the mathematical elegance of the subject. arrival times. For the stationary case. 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.
The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . Using Lundberg conjugation .a*PB• > 1. b=1 !=1 where {Nt } is a Poisson process with rate . i.d. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . the claims and the premium rate are negative so that the risk reserve process . the claim surplus process are given by Nt Nt Rt = u+^U. 00). A typical sample path of {Rt } is illustrated in Fig. the remaining part of the prepayment (if any ) is made available to the company.1. 2. At the time of death . (2. The initial reserve is obtained by prepayments from the policy holders.3* (say ) and the U. with common distribution B* (say) concentrated on (0.1) +ry.Ut.3* pB.1) .1 If. St = t . each of which receive a payment at constant rate during the lifetime . resp .1 r* (u) One situation where this model is argued to be relevant is life annuities. then 0 * (u) = 1 for all u > 0. are independent of {Nt} and i.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . t. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. If . we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.134 CHAPTER V. < 1. Theorem 2 . U Figure 2. That is . RENEWAL ARRIVALS 2 Exponential claims.
Let B(dx) = ^e7x B*(dx). 0 Now return to the renewal model. . EXPONENTIAL OR NEGATIVE CLAIMS [Note that r.. B.2(a). B*.0.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. Define T_ (u) = inf It > 0 : St = u} .f. of {St} is c(a) = is*(a7).(a) 7 Figure 2. Then the c.Rt. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. and thus 1 = P(T. Since ic'(0) < 0. Fig. the safety loading of { St} is > 0.UB. 0) and has typically the shape on Fig. St=Rtu=St.2 Assume now . T_ (u) = inf { t > 0 : St = u 'r* (u). then by Proposition 111.s. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). > 1 .g.3*. Then the function k* is defined on the whole of (oo. B* [7] and let {St} be a compound Poisson risk process with parameters .1.(u ) < oo) = E {e7sr_ (u). Hence T_(u) < oo a. and the Lundberg conjugate of {St} is { St } and vice versa.2. (a) is*(a) (b) . 2. If I3*pB* < 1.* (a) = log Ee'st I.2(b). cf. 2. Proof Define 135 St =u . Hence y exists and is unique.
Tr+.•.Ti = U1. and from Fig . Taking m. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII...136 CHAPTER V.)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[. T2 = U2..Tn} n=0.7r+ 7r EeTo b/(Sa) + +.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 .Y] (2. Then B* = A.+Tn U1 Un..1.2 If B is exponential.f.1 it is seen that ruin is equivalent to one of these values being > u. To + M(d) in the notation of Proposition 1. with the probability that a particular jump time is not followed by any later maximum values being 1 . with rate S (say).1) + ry = 0 which is easily seen to be the same as (2... and 5PA > 1. To + max {Ul+•••+UnTI.e.g. u Hence P(M(d) > u) _ 1r+e'r". A variant of the last part of the proof. Hence M* max {To + Ti + • • • + Tn .Ui . the failure rate of this process is y.. then ..1..2). we get Ee'M(d) = Ee°M* _ Y/(..1. 3* = 6. According to Theorem 2..'s and noting that V)*(u) = P(M* > u) so that Theorem 2. alternatively termination occurs at a jump time (having rate 8). and (2 .• • • .Un } = max St = t>0 n=0... Now the value of {St*} just before the nth claim is To +T1* +.u+ and lr+. 2. 1) means that 8(A[ry] .4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 .a) = 1 .4.Y a I. respectively. and hence the failure rate . RENEWAL ARRIVALS Theorem 2 . However.1 means that M* is exponentially distributed with rate ry..
B(dx)..5. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II.6. resp. 3a The imbedded random walk The key steps have already been carried out in Corollary 11.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". a ladder step is the overshoot of a claim size.. Bads (dx) = . B^d) where Aad> (dt) = ^[ a] A(dt). Furthermore.2..7r+) and hence r+ = 1. The probability that the first ladder step is finite is 7r+.y/b.. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. 0 3 Change of measure via exponential families We shall discuss two points of view. 111. Hence the failure rate of M(') is 6(1 . However. : S(d) > u} . we see that ry = 6(1. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U .3. letting P(d) refer to the renewal risk model with these changed parameters .3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1.T to F(d)(x) = eK^d^(«) ^x e"vFidi(dy) 00 K(d) (a) = log F(d) [a] = log B[a] + log A[a] .2.4. hence exponential with rate b. which states that for a given a. This follow since. 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^. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+.7r+). the imbedded discrete time random walk and Markov additive processes. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. Putting this equal to y.
just note that F7(d) is nonlattice when F is so .(u) .1 For any a such that k(d)' (a) > 0. provided the distribution F of U . (b) V)(u) . ik. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. i .t. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . E(d)e 1' (u). It should be noted that the computation of the CramerLundberg constant C is much more complicated for the renewal case than for the compound Poisson case where C = (1 .Ce"u where C = limu. and claim (a) follows immediately from this and e (u) > 0. O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . Corollary 3.r. 187) and thereby for ^(u) to be nonlattice w. in the easiest nonexponential case where B is phasetype. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.C(°)eryu where C(O) = C0[7] . For the stationary case.p)/($B'[7] .1). 7µA . cf.e.4. (d) (7) _ 0. 00)(u) .C8e7u where Cs = Ce78B[7].T is nonlattice. Consider now the Lundberg case.. (a) '(u) < eryu. For claim (b).1 implies Cu) = e«uE ( 7d)e«^(u) .2 In the zerodelayed case. we get: Proposition 3... VIII. In fact. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.1) is explicit given 7.2 p. Proof Proposition 3.3 For the delayed case Tl = s. We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 .u the overshoot . let 7 > 0 be the solution of r.138 CHAPTER V.
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. y) = e°yh(s). Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ .dt(ah ( s) + h'(s)) so that Gha ( s. IPA 0 Of course. Let P8f E8 refer to the case Jo = s.(s. we invoke the behavior at the 1 = h«(0. St)} and h. According to Remark 11. B(x) = o(e7x) and dominated convergence. 3b Markov additive processes We take the Markov additive point of view of II.9./c.St)} can be defined by taking Jt as the residual time until the next arrival.5. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1..0 ) = Eo[ha ( Jdt. 0) = tc(a)h(s). To determine boundary 0. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). E8h0 (Jdt.(°) ( Ce8B[7] Ao(ds) similar manner.1) (normalizing by h(0) = 1). The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt.4). we look for a function h(s) and a k (both depending on a) such that Gh. For s > 0.a .Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds). delayed version of Lundberg's inequality can be obtained in a e7u. where G is the infinitesimal generator of {Xt} = {(Jt.dt ) eadt = h ( s) . (u + s . 0) = ah (s) .3.h'(s).1) = C(O). Here K. Sdt) = h(s . The expressions are slightly more complicated and we omit the details. (s..y) B(dy) 0 For the stationary case. 0 0 .5.
. since JT.c(a)] B[a] Proof Pa.. Proposition 3.140 CHAPTER V. [a + /3] A[b .s governing {(Jt.2) As in 11. 5 For the compound Poisson case where A is exponential with rate . An important exception is.8. resp. . . the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. (3. T2 are independent with distributions Ba. Further. Note that the changed distributions of A and B are in general not the same for Pa. .s(Jo = s) = 1 follows trivially from Lo = 1. RENEWAL ARRIVALS B[a]A[a .s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa. however.a .2) means 1 = B[a]/3/(/3+a+rc (a)).rc(a)] = B = Ba[13]Aa[5].a .13]A[b .S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east . T2.c(a)] which shows that U1.4 The probability measure Pa. resp. Ba where Aa (dt) . we can now for each a define a new probability measure Pa. = J8 = T2.e.. U. . Aa as asserted .2).1)a in agreement u with Chapter III. A[a . Remark 3 . (3. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. An easy extension of the argument shows that U1. Ba(dx) = B(dx). i. Ease AU1+6T2 [ AU1+6T2 = Ea a LT.rc(a)] = 1. rc(a) = 0 (B[a] . J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk.e(«+k(a))t esy A(dt).( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3.s and P(d)..5. ... ] = E.tK( a)e. .rc(a)] B[a] A[a . [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +.
A(ds). .yu ) e7vu A[ay .g.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. Proof As in the proof of Theorem IV. not after time T. Label the customers 1..5 Proposition 3. yu) = F'ay.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. and define yy = ay . Then J(rr(u)) = TM(u)+1 and hence Ws(u. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time).r.)+1 e J j e(ay+w(ay))8 e . defined as the single server queue with first in first out (FIFO. The claim for the zerodelayed case follows by integration w..s eaysr(")+r(u ) K(ay) h (s) . which is the same as the asserted inequality for 0. that is. 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. or FCFS = first come first served) queueing discipline and renewal interarrival times.t. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. [APQ] Ch. see e. . XII. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. and assume that T„ is the time between the arrivals of customers n . u The approach via the embedded random walk is standard. for the zerodelayed case zp8(u. For the approach via Markov additive processes.(u.y yuAa y [ay + K(ay) . say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims.ay+ray))TM(. yu). the amount of .1 and n. Then "^ e(ay+w(aY))8 Ys(u. Using the Markov additive approach yields for example the following analogue of Theorem IV. The virtual waiting time Vt at time t is the residual amount of work at time t. Let M(u) be the number of claims leading to ruin . the time from he arrives to the queue till he starts service. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. it is easily seen that ic(ay ) > 0. .4. and U„ the service time of customer n. that is.4. yu ) < e7yu. 2.4.yx(ay).5.
2) (b) as t * oo. It then jumps to VQ„ . Let the T there be the random time UN. (4. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival. then Wn v M. Vt converges in distribution to a random variable V.1. (u).Tn)+. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. (4. the proposition follows. equivalently.n1 (U1 +• • •+Uk Tl .4. and obviously z/'(u) = limN..1). but interchanging the set .3 Assume rl > 0 or. Thus. Then: (a) as n + oo. 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. since customer n arrives at time on. The traffic intensity of the queue is p = EU/ET. and combining with (4.• • • Tk ).4.3) Proof Part (a) is contained in Theorem 11.2.3. 0 Applying Theorem 11. (4. Thus Vos}1 _ = (Wn + Un .142 CHAPTER V.. whereas in [On . on + Tn) the residual work decreases linearly until possibly zero is hit.1. we get: Corollary 4.4.1 Wn+1 = (Wn + U. but we shall present a slightly different proof via the duality result given in Theorem II. Wn converges i n distribution to a random variable W..Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ .4): Proposition 4. RENEWAL ARRIVALS time the server will have to work until the system is empty provided no new customers arrive (for this reason often the term workload process is used) or. . p < 1.4: Proposition 4. equivalently. an+1) = [on."^ Vi(N) (u). in which case {V} remains at zero until time on+1. the waiting time a customer would have if he arrived at time t.1 and Corollary 11. If W1 = 0. we have Wn = Van(left limit). and we have P(V > u) = ?/iiol(u).1) The following result shows that {Wn} is a Lindley process in the sense of II. and we have P(W > u) = V.2 Let Mnd) = maxk=o..+ Un.
5) Proof Letting n . we get W = (W + U* . (4.T* are independent and distributed as U1. as WN. hence (since the residual lifetime at 0 and the age at T have the same distribution . conditioning upon U* .. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival.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. we obtain: Corollary 4.2. TN) with (TN. T1 . (4. Corollary 4.2). we let T be deterministic ..T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). However..le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before .4 The steadystate actual waiting time W has the same distribution as M(d).. For part (b). u Now return to the Poisson case .Ao in (b).. Letting n oo in Corollary 4.5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x). x > 0.1. convergence in distribution hold for arbitrary initial conditions . THE DUALITY WITH QUEUEING THEORY 143 (T1. Ti) and similarly for the U. It follows that P(WN > u) =. cf. T1. A. by an obvious reversibility argument this does not affect the distribution . K(x) = P(W < x).4. and we get: . which implies the convergence in distribution and (4.. resp .. In fact .. Then K(x) = J x00K(x .(N)(u) has the limit tp(u) for all u. but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit. Hence for x > 0. Then the arrivals of {Rt} in [0. Then the corresponding queue is M/G/1. i. namely W1 = 0 in (a) and Vo = 0.e.T*)+ < x) = P(W + U* . where U*..T* = y yields K(x) = P ((W + U* . T] form a stationary renewal process with interarrival distribution A.oo in Proposition 4.T*)+. { 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)..y)F(dy).
VIII). the actual and the virtual waiting time have the same distribution in the steady state. implying P(W > u) = P(V > u) for all u.5) looks like the convolution equation K = F * K but is not the same (one would need (4. Note that (4. RENEWAL ARRIVALS Corollary 4. Proof For the Poisson case.6 For the M/G/1 queue with p < 1.g. Cohen [88] or [APQ] Ch. the zerodelayed and the stationary renewal processes are identical. see e. W v V. Hence '(u) = Ali(°)(u).5) to hold for all x E R and not just x > 0). Asmussen [24] and references there. Some early classical papers are Smith [350] and Lindley [246]. 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. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . . That is. The equation (4.g.144 CHAPTER V.5) is in fact a homogeneous WienerHopf equation.
(3i when Jt = i.f pi.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. Ire = 1. i=1 0 and r(u) = inf It > 0: St > u}. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). {St} denotes the claim surplus process.. M = supt>o St. The intensity matrix governing {Jt} is denoted by A = (A. Thus. • Claims arriving when Jt = i have distribution Bi. N. and can be computed as the positive solution of WA = 0. . {Jt} describes the environmental conditions for the risk process.)iJEE and its stationary limiting distribution by lr. 145 Oj( u. t St = E Ui .T) = Pi (T(u) < T). • The premium rate when Jt = i is pi. dv. here it exists whenever A is irreducible which is assumed throughout.
For example. we could distinguish between normal and icy road conditions. We let p Pi = /ji/AB. P = E 7riPi. 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. Thus. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. this is no restriction when studying infinite horizon ruin probabilities.2. T(=)). with rates Ani and Ain. a(i). /3 = Nn when j E E(n). the operational time argument given in Example 1. say.T(n)). f3i and claim size distributions Bn. one expects that 3i > on and presumably also that Bn # Bi. Example 1.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. we can approximate A(i) with a phasetype distribution (cf. = iii when j E E(i).2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. leading to E having two states n. According to Theorem A5.11 below. Then the state space for the environment is the disjoint union of E(n) and E(i). Unless otherwise stated. t(i) = T(')e are the exit rates. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e.4) with representation (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('). Bi. we shall assume that pi = 1. r^ = P (1.14. and we have f3.. An example of how such a mechanism could be relevant in risk theory follows. Example 1 . meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones.146 CHAPTER VI.a(').5 below. assume that the sojourn time in the icy state has a more general distribution A(i). Proposition 1. say. and p is the overall average amount of claims per unit time.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. respectively. cf. and assume that weather conditions play a major role for the occurence of accidents. u .1 Consider car insurance. in blockpartitioned form. i and corresponding arrival intensities Qn. cf. which is clearly unrealistic.
we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t..1.J017.. Then for example wi.a(n). i ) : n E H. 1 .>.T(n)).p.. (9) where q = CHI. let T 9(T) = f pi. the parameters are ^ij = aid/pi.3 Consider again the alternating renewal model for car insurance in Example 1.n. dt.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. depending only on 77.. but assume now that the arrival intensity changes during the icy period. . t(n) = T("i)e. n8}. the state space E for the environment is { ('q. such that a sojourn time of type rt is followed by one of type c w. iq (visited in that order) and letfOil >. i8f n1. u Example 1 . where W = (w. is the probability that a long icy period is followed by a short normal one. say it is larger initially. and ..3i.3.. say. Qi = .1.tEH is a transition matrix.. and similarly for the normal period. One way to model this would be to take A(') to be Coxian (cf. In the car insurance example. u From now on.Q.. u Example 1.2. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. i E E(n) }. A('^). The simplest model for the arrival intensity amounts to .j = . Example VIII. St = SB=(t).3i/pi. w. resp. Approximating each A('?) by a phasetype distribution with representation (E('l). . 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. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L).. 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) . MODEL AND EXAMPLES 147 Example 1 . T(9) +wggt(9)0. Indeed. one could for example have H = {i1.4) with states i1. it = Jel(t). and 1/ii(u) = t/ii(u). This amounts to a family (A(")) ?CH Of sojourn time distributions....
Nt Nt a . vi(dx) = .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. one can associate in a natural way a standard Poisson one by averaging over the environment. Pi (Ti E dx.(3iBi(dx).6 The claim surplus process {St} of a risk process in a Markovian environment is a Markov additive process corresponding to the parameters µi = pi.7 The Pidistribution of T1 is phasetype with representation (ei. More precisely. 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.e(A(Oi)d'sg)xe. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.5. In particular. 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. the empirical distribution of the claims is B*. we put )3* = E 7fi/3i. Next we note a semiMarkov structure of the arrival process: Proposition 1. )3*. JT1 = j) = Qj • e. o = 0.8 As t oo. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. B* = 1 /^* Bi. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. N > 1(Ul < x) a4 B*(x).148 CHAPTER VI.(Qi)diag)• More precisely. The key property for much of the analysis presented below is the following immediate observation: Proposition 1. . 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).
. and furthermore in the limit JT. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. y Ni) i Nti) t a. Hence Nt'> a .x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). 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 . Proof According to Proposition 1. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. the limiting distribution of the first claim size U1 is B*. aA. Conditioning . and let {St °i} refer to the one with parameters Pi. i.aA . In particular.. By Proposition A5.1. Example 11. e.. oo) as a 4 oo. A. 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.* (u) for all u. {St} to the compound Poisson model with parameters 0 *.7.2. cf. has distribution (7ri)3i //3*)iEE and is independent of Ti.6. given {Jt}0<t<0. In particular. this converges to the exponential distribution with rate 0* as a * oo. zli( (u) . Bi.. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. Then {St)} + {St*} in D[0.. ^j 7riNi.9 Consider a Markovmodulated risk process {St} with param eters Ni. Bi(x). 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$. Then it is standard that ti lt '4' iri as t > oo. However . we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti. Nt a' t t iEE Also. B*.4. Proposition 1.(/3i)aiag). Bi.
since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier).g. Continuing in this manner shows that the limiting distribution of (T.l3* and U2 having distribution B*. Computing the parameters of the averaged compound Poisson model. with T2 being exponential with rate .. 132=2.2}. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary.1. 1. 9 . we may imagine that we have two types of claims such that the claim size distributions are E3 and E7..).. MARKOVIAN ENVIRONMENT upon FT.. thick. and (at least when a is small such that state changes of the environment are infrequent)..2. is as in {St }. oo). B2=1E3+4E7.s = o in state 1 and with intensity 1 . resp.=1. there are p = 2 background states of {Jt}. Thus. On Fig. That is. From this the convergence in distribution follows by general facts on weak convergence in D[0.s = 1o in state 2. the company even suffers an average loss.FT.. > 1.31µB 2 = 2 5 3 7 70 Thus in state 1 where p. marked by thin. the paths of the surplus process will exhibit the type of behaviour in Fig.. U.. Claims of type E3 arrive with intensity 2 .. U2) are independent of . 1.1 with periods with positive drift alternating with periods with negative drift.10 Let E_{1. e. which also yield O(a) (u. s 5 in state 2. state 1 appears as more dangerous than state 2. those of type E7 with intensity z s = 5 in state 1 and with intensity z . 0 Example 1. B1=3E3+2E7. lines in the path of {St}.1 of [145]. we first get that 3 (3* = 2. shows similarly that in the limit (T2.150 CHAPTER VI. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. from Theorem 3. A= (  a a ) \ a a 5 5 J 9 3 2 a1=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 = . T) + ?P* (u.2 +2 2 = 3.2. T) for all u and T. The fact that indeed 0(a) (U) 3 0* (u) follows.
a fraction r. note first that EN Uk')/N a4' µgi. t + oo. MODEL AND EXAMPLES 151 Figure 1.1 Thus. (b) St/t * p . t * oo. That is..(3. iEE .1 a. 0 The definition (1.1.1) 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. 01 /. Proof In the notation of Proposition 1.1). 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).8. the averaged compound Poisson model is the same as in III.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2.1.11 (a ) ESt/t * p . Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7. = P.s. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. For (b).
. [302]. Since the X„ are independent . then M < oo a. In risk theory. and hence oscillates between 0o and oo so that also here M = oo. some early studies are in Janssen & Reinhard [211]. s.1)Eiw = 0. with X2. and hence 1/ii(u ) = 1 for all i and u. X3.Jt=i}.12 If 77 < 0. Now let r) = 0. MARKOVIAN ENVIRONMENT Corollary 1. 136 or A. See Meier [258] and Ryden [314].1 jEE = (p . EiX = 0. see the Notes to Section 7. Proof The case 77 < 0 is trivial since then the a..1 of St / t is > 0. and . 38) Eiw1 = 1/ir. see [APQ] p. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. Then by standard Markov process formulas (e. [APQ]. w2=inf {t>w1:Jt_#i. . The mainstream of the present chapter follows [16]. [315]. dt .0i(u) < 1 for all i and u. also + . and so on.. n n Thus {SWn l is a discrete time random walk with mean zero.. If 77 > 0. Statistical aspects are not treated here. X 1 =Sty. and hence wn /n a4.g. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade. . [212].. The case 77 > 0 is similarly easy. PB..\ i and EiX1 Ei f 13 J.. Theorem II. + Xn SWn ](1 a .152 CHAPTER VI.s.4. X2 =SW2 So.. The proof of Proposition 1. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43].. then M = 00 a. There seems still to be more to be done in this area.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes.Jt=i}.Eiw o'o Eiw • E ^ifjµs.ld. and involves a version of the . limit p .1 and the Corollary are standard. having the Pidistribution of X..s. Proposition 1. Now obviously the w. let some state i be fixed and define w=wl=inf{t >0:Jt_#i. and hence M = 00. and a more comprehensive treatment in Asmussen [16]. Eiw.2(a) p.a form a renewal process .
1 irG+(dy)e =. j.. (y.1) 0 (b) G+ (y.4) we obtain the following result . For measurevalued matrices. Proposition 2.2.(u) = Pi(M < u) = e' E G+ (u)(I .A) = Pt(ST+ E A. oo) = J ao 0 G+(i. see also Example II. j. 6.x). but is substantially more involved than for the compound Poisson case .a/i.A) =ZI(St E. dx)/jBj(y .i. the definition of . for i. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element.6*. The form of G+ turns out to be explicit (or at least computable). let G+(i. k.5. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. cf. That is.3*B *(y)dy.dx). •) II = JG+(i.j. j. only with the product of real numbers replaced by convolution of measures.2 (a) The distribution of M is given by 00 1 .Jt=j)dt. Thus. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i. oo)) = f R(i. Proposition 2. j.2) R(dx)S((y . .x. we get the same ladder height distribution as for the averaged compound Poisson model. we define the convolution operation by the same rule as for multiplication of realvalued matrices.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument.EA. e. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. •) * G +(k.j. However . oo)). T R(i.Jr+ =j. j.j E E. by specializing results for general stationary risk processes (Theorem II . n=0 (2. IIG+ II denotes the matrix with ijth element IIG+(i.6. •). Let further R denote the preT+ occupation kernel. •)• kEE Also.IIG +II)e.g. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). 00 (2. G+ is the matrix whose ijth element is E G +(i. B* in Section 1.
2 useful . G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . JJ = j.1 for an illustration in the case of p = 2 environmental states of {Jt}. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. The u proof of (2. St < S* for u < t. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). To make Proposition 2.3) We let {St*} be defined as {St}. thick. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. we need to invoke the timereversed version {Jt } of {Jt} . lines in the path of {St}. we need as in Chapters II.154 CHAPTER VI. To this end . see Figure 2. That is. marked by thin. hence uniquely specified by its intensity matrix Q (say).3. From this (2. {mx} is a non terminating Markov process on E. .1) follows by summing over n and j. resp. III to bring R and G+ on a more explicit form .6.3 When q > 0.IIG+II)e. 0  x Figure 2. and that the environment is j at the nth when we start from i is e .1 The following observation is immediate: Proposition 2.2) is just the same as the proof of Lemma 11. mx = j when for some (necessarily unique) t we have St = x.
} is a minimum value at v = t.2 where there are three excursions of depth 1.(/3i)diag + T S(dx) eQx.*.0. An excursion of {St*} above level x starts at time t if St = x. and the excursion ends at time s = inf {v > t : S. = x}. In general. ( Q( n)) converges monotonically to Q.(/3i) diag.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .and a jump (claim arrival) occurs at time t. 0 mms1   ^O \ T. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. Otherwise each jump at a minimum level during the excursion starts a subexcursion. Furthermore. The definitions are illustrated on Fig. 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). Figure 2. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. If there are no jumps in (t. Q( n+l) _ ^.2 . s]. 2.2.2. the sequence {Q(n)} A* defined by Q(O) = . Proof The argument relies on an interpretation in terms of excursions. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2.. we say that the excursion has depth 0. For example the excursion of depth 2 has one subexcursion which is of depth 1. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. {S. corresponding to two subexcursions of depth 0.
Then a jump to j (i.156 CHAPTER VI. of the definition to make U be concentrated on (co. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). Writing out in matrix notation .4).6) . It follows that qij = A.. A). StEA . p1^) Define a further kernel U by f U(i.u< t). (2. j. i. 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.St EA.j. e.St <S*. Similarly. 7rE Proof We shall show that Fi(Jt=j.Qi + )%pij) Now just note that t pij and insert (2.s.T+>t) _ ^iF 7ri (JJ =i. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. Suppose mx = i. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. A) = L' U(j. h. or through an arrival starting an excursion terminating with J. Q = W(Q) follows. Similarly by induction .5) A (note that we use A = {x : x E Al on the r.4) To show Q = cp(Q).(01)diag = Q. either due to a jump of {Jt } which occurs with intensity A= j. Theorem 2 . mx+dx = j) occurs in two ways .. A) = f Pi(mx = j) dx eie4xej dx A u (2. By considering minimum values within the excursion.j +/3ipij.5 R(i. we first compute qij for i $ j. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . 0)). = j.
z+>t) = P. We may then assume Ju=Jtu. where A is the diagonal matrix with 7r on the diagonal: Corollary 2. Jt = i. e. (Jo = j.(Jt=j. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx). and this immediately yields (2.S„<0.StEA.0<u<t) = P.. consider stationary versions of {Jt}. the CramerLundberg approximation (Section 3).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.. To this end.. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A. and to obtain a simple solution in the . St < St U.6).St <Su.2. oo)) = f o' eIXS((x + z. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. G+((z.7 It is instructive to see how Proposition 2. oo))dx.Qi)diag.t. Remark 2. x < 0.6 is hardly all that explicit in general.. K( n (d) the sequence converges monotonically to K. and get irPi(Jt =j. St E A. (b) for z > 0. St EA. From Qe = 0. we shall see that nevertheless we have enough information to derive.Jo=i.St EA. S.r. dt. {Jt }. 0<u<t).6 (a) R(dx) = eKxdx. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1).4].g. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2.1 can be rederived using the more detailed form of G+ in Corollary 2. 0 +1) = cp (K( n)) defined by K(o) = A .(. 0 < u < t) = 7rjPj(Jt =i..0<u<t. and we let k be the corresponding right eigenvector normalized by Irk = 1.=StSt.
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
(ay)}. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i.5). if y > 1lk'(ry). hj P .5) will produce the maximal ryy for which the argument works.V)i(u. exp {e() + r(u))} . r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av). (3. (u. Chie ryu < Vi(u ) < C+hie 7u.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). for a vector <p(u) = (cpi (u))iEE of functions . 1 Similarly.i [eT(u)K(av ). r(u) yu o)(y)eavuEav. dy)• o iEE jEE . However. as in the classical case (3. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. we have ic(ay) < 0 and get 'i(u) .00 su e7( ( 3.8 ) Then for all i E E and all u > 0. yu < r(u) < 00] < hiav)C+o)( y)eavu+yuw(av) 0 Note that the proof appears to use less information than is inherent in the definition (3.i [e*(u)K(av)..3. we shall need the matrices G+ and R of Section 2.9) For the proof.i I (a) exp {aye(u) + r(u)r.j) * coj)(u) _ f u ^Pj(u . yu) f h(av)e v avuE«v.j. 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. oo)) and. we let G+ * W(u) be the vector with ith component E(G+(i.7. av 'i [h.y)G+(z.
00 f C_ hj f e(Y)G+(i. j. 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. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u).(0) ] (u) < sup Jt t.j. we have G *(N +1) * ^.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. dx) f e7( vu)Bj (dy . j.x ) = Gi(u). 00 Thus C+ > hj f"o e7(Yu)G +(i. if r+ (n) is the nth ladder epoch.u IMP:°) (u) I < oo. j.13 For all i and u. = Eo G+ G.166 CHAPTER VI. Lemma 3 . _ To see that the ith component of U * G(u) equals ?Pi (u). j.& (u). U = U". Hence lim cp(n) exists and equals U * G.x) x) jEE 0 E Qj f jEE R(i.j. dy) 00 C+ ijhj f R(i. 0 G+(i.12 Assume sup1.ery(&u+x)Bj (dy) Bj(u Bj (u . dy) : 1(u) < C+ > hj u e(1tL)G+(i. n > oo. dx) 100 C . Proof Write UN = EN G+ . °O . dx ) Bj (u . dx). jEE u 0 j. dy) = aj f Bj(dy .x ) R(i. Then cpin)(u) sit (u) as n + oo.x) jEE 00 u 0 //^^ C+E.3jhj // f 00 R(i.dy).u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. MARKOVIAN ENVIRONMENT Lemma 3 .
MT<u.u)G+(i. (3. T) = Pi (7.(u) < T ) to 0i (u) which is different from Theorem 3. from which the lower inequality follows by letting n * oo. 13 and the second by the induction hypothesis . and using Lemma 3 . jEE estimating the first term in (3. j.13 Let first cp=°)(u) = C_ hie"u in Lemma 3.Pi(MT > u) = Pi(MT < u.ST). y]hj = C_ e7uhi. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . dy) (3. T) < C+(')' o)hi7u)e7ou8T . and let 8 = e'(70). letting MT = maxo<t<T St. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. taking cps°) (u) = 0. (3.11. We claim by induction that then cpin) (u) > C_ hie7u for all n.10) C_ 1 f hje7(y. j. this is obvious if n = 0. dy) jEE o (3.10 ) by Lemma 3 .13. +i . it follows that Vi(u) < C_(yo) h=70)e7ou. and assuming it shown for n.M>u) = Ei [VGJT (u .12) Proof We first note that just as in the proof of Theorem 3.3. 167 u Proof of Theorem 3. j. u The proof of the upper inequality is similar . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.11).11) C_e7u 57 O+[i. and the proof of the lower one is similar.T) = Pi(M > u) . let C+(yo) be as in (3. we have Vii (u) . dy) jEE u U +C_ hje7( yu)G jEE"" +(i. Then 0< Vi (u )  0i(u. j. 9 for the last equality in (3. MT < u.10: Theorem 3 .8) with y replaced by yo and hi by h=7o ).n) ( u .tpi(u.M > u) = Pi(ST<u. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u.13) Hence.y)G+(i. 14 Let yo > 0 be the solution of 'c'(yo ) = 0. Indeed.
and finally in part from queueing theory.o..4) To avoid trivialities.3) to B = Bi does not depend on i. we refer to . 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). this correponds to the usual stochastic ordering of the maxima M'. ".31:5)32 . this is not the case for (4.3p. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. For the notion of monotone Markov processes. u > 0. (4.. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).o. <s... [177].168 CHAPTER VI.. in part from the folklore principle that any added stochastic variation increases the risk. The results to be presented show that quite often this is so. Further related discussion is given in Grigelionis [176]. we also assume that there exist i # j such that either /3i <.o.o..33 or Bi 0 Bj. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and .5) Note that whereas (4. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). Occasionally we strengthen (4. The conditions which play a role in the following are: . (4. but that in general the picture is more diverse.2) alone just amounts to an ordering of the states. V)" if z/i'(u) <'c"" (u). MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. The Markov process {Jt} is stochastically monotone (4..1) Obviously. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. B2 <_s. It was long conjectured that 0* Vi..3) Bl <_s.3). < . Bp.0. we define the stochastic ordering by 0' < s. (4. The motivation that such a result should be true came in part from numerical studies.2) (4.
Section 4.4) say basically that if i < j . The first is a standard result going back to Chebycheff and appearing in a more general form in Esary.6. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]..5 (cf.7) and Lemma 4.r(u x)dx. Comparing (4. 2 If al < .. where 7r2+) = QiµBilri/p.1.1 Assume that conditions (4. dx (4.8) ^j Tri/iBd(x) . we need two lemmas. then P P P 7rjbj. then j is the more risky state .x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.9) follows by considering the increasing functions 3iBi (x) and Oi (u . it follows by a standard .7) 7ri. Proschan & Walkup [140]. p).. < a.3* f uB(x) z/^. E 7r i Wi(u . the second follows from an extension of Theorem I1. Lemma 4 . and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.9 ) below)..3iBi(x)YPi(u .10) Q*B*(u)+.2)(4. Conditioning upon the first ladder epoch. = aP or b1 = .. Proof of Theorem 4. . T(0) < oo) = Bi(x) dx/tcai .4) hold.2)(4.* For the proof.. 7(0) < oo) = pirf+)..10) and (4.r (JT(o) = i. (b) P. ^i 7ri = 1.13* J0 u 0*(u .4...2. Then V. Theorem 4 .9) (4. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = .6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. Proposition 2.x)B*(x) dx.x) dx u o i =1 i=1 (4.4) is automatic in some simple examples like birthdeath processes or p = 2 .. 3 (a) P.r (Sr(o) E dx Jr(o) = i. Conditions (4. = b..1) which with basically the same proof can be found in Asmussen & Schmidt [49]. we obtain (cf.6). 0 Here (4. b1 < . (4. also Proposition 2. note that (4. < bp and 7ri > 0 (i = 1.1 for the first term in (4.x) of i and using Lemma 4.2. . Lemma 4 .3 for the second) *(u) _ /3 *B* (u) +....
Notes and references The results are from Asmussen.0*• i=1 But it is intuitively clear (see Theorem 3.4 Assume that . let = ( 1/2 1/2 ) .8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi ..6). i=1 i=1 7'r(0) _ EFioiwi(0) .1 and Proposition 4. Then i/i*(u) < .6).3µi < 1 for all i..11) i=1` and that A has the form eAo for some fixed intensity matrix A0.4 is not vacuous. µB2 = 104. What is missing in relation to Theorem 4. 01 = 103. Then the l.s. this ruin probability is /3iPBi.4) is essential (the present author conjectures it is).170 CHAPTER VI. For u = 0. except possibly for a very special situation .0. As is seen.r (u ) fails for all sufficiently small e > 0. of (4. (u) may fail for some u. MARKOVIAN ENVIRONMENT argument from renewal theory that tk. and from this the claim follows. µB.* (0).4 is the understanding of whether the stochastic monotonicity condition (4. dominates the solution 0* to the renewal equation (4./3*..s. Rolski & Schmidt [32]. Recall that . Q2 = 1. (u) is not in general true: Proposition 4.. Frey.2.(0) < b *'(0) for e small enough. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4.h.(0) = V. = 102.h. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. of order 101. Bi as e J. u To see that Proposition 4. (4.3i. it will hold for all sufficiently large u. Using (4. it is sufficient to show that 0'.. Proof Since 0. they are at present not quite complete. u Here is a counterexample showing that the inequality tp* (u) < V).11) is of order 104 and the r.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . 0.
5.12) iEE Theorem 4. Lemma 4.g. is nondegenerate unless bi does not depend on i E E. It is clear that the distribution of X.g.)a. (4. Hence if 5i 54 0 for some i E E. with strict inequality unless a = 0 or bi = 0 for all i E E. Now we can view {Xt} as a cumulative process (see A. and by Proposition II. (4.1) . 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] . e.ld) with generic cycle w = inf{t>0: Jt_54 k.2 we have (Ei[e"X'. it follows by Proposition A1.4(b) that the limit in (4.5 y < ry*.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0.13) implies A(a) > 0 for all 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] . Further (see Corollary 11. cf. This implies that A is strictly convex. with strict inequality unless rci (y*) does not depend on iEE.13) (4. Then {(Jt. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E..(a) > 0 for all a 0 0.4.5. (4. Xt)} is a Markov additive process (a socalled Markovian fluid model.a = E irirci(a).14) is nonzero so that A"(0) > 0.a.6 Let (di)iEE be a given set of constants satisfying EiEE iribi = 0 and define A(a) as the eigenvalue with maximal real part of the matrix A + a(bi)diag• Then )t(a) > 0.14) A„(O) iioo varXt t t By convexity. 0 . Asmussen [20]) as discussed in 11. in particular . Proof Define X= f &ids. Jt = i])' EE = vA+n(6.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.5.1) ..
15) Normalizing h by 7rh = 0. Frey. note that y(a) + mins=1. Then > risi = 0 because of (4. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. h depend on the parameter (e or a). a = 1 in Lemma 4. whereas the . h'(0) = (Ao .12) and rc*(y*) = 0. h(0) = e. Here we put a = 1/e.e7r)1 (Ici(Y*))diage. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. Further a(1) = rc(y*) by definition of A(. Let bi = rci(y*). this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*.5. where A. Thus y(e) * y* as e 10. (4. Hence letting e = 0 in (4.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao .15) once more and letting e = 0 we get .6. The corresponding adjustment coefficient is denoted by ry(e). the basic equation is (A + (rci(y))diag)h = 0. In the case of e. Rolski & Schmidt [32]. Notes and references Theorem 4...) and rc (•).16) Differentiating (4.Qi and Bi are fixed . we have 7rh' = 0. Since ic is convex with rc'(0) < 0 .. (4. If rci(y* ) is not a constant function of i E E. y. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. improving upon more incomplete results from Asmussen. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e..p yi and compute 8y 8a a=0 In both cases.5 is from Asmussen & O'Cinneide [40].172 CHAPTER VI.eir)h'(0). we get rc (y*) > 0 which in a similar manner implies that u y < y*. Hence rc (y*) > 0. MARKOVIAN ENVIRONMENT Proof of Theorem 4.
0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. i = 2. and may have some relevance in risk theory as well (though this still remains to be implemented).8 If (4. The analogue of Proposition 4. which has recently received much attention in the queueing literature. (4.17) by 7r to the left to get (4.20) and multiplying by el to the left we get 0 = All + 7'(0)rci (0) + 0 (here we used icl (ry(0)) = 0 to infer that the first component of K[7(0)]h'( 0) is 0).20) Letting a = 0 in (4. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) .16) yields Proposition 4.i(7' *))diagh'(0).. Inserting (4. p. multiplying (4. We assume that 0 < y < 7i. .18) 0 = 27'(0)p+27r(rs.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. then 8a a=o All rci (0) Notes and references The results are from Asmussen.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector).18). 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. (4. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . Rolski & Schmidt [32]. (4.8 when ryi < 0 for some i is open.5.19) holds.. . Frey.17) (4. 5 The Markovian arrival process We shall here briefly survey an extension of the model. and we have proved: Proposition 4.
Indeed. MARKOVIAN ENVIRONMENT f o r a transition i + j by A .1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. j(2) } be two independent environmental processes and let E(k). A(1'k) A(2 k1). then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. . but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. Here are some main examples: Example 5 . In the above setting. we use the convention that a1i = f3i where 3i is the Poisson rate in state i. Bii = Bi . and thus 1i = 0.2).i. u Example 5 . is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. This is the only way in which arrivals can occur. T). A(1) = A . A(l) = T. the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. let { Jt 1) }.174 CHAPTER VI. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. the definition of Bij is redundant for i i4 j. A ( 2) = A (2`1 ) ® A. Jt2)) (2.^) etc. the claim surplus is a Markov additive process (cf. Bij = B. that Bii = Bi . A(l) = tv.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. where qij is the probability that a transition i * j is accompanied by a claim with distribution. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. Thus .(13i )diag.d. II.6i ) diag.4). refer to notation) { Jt k) }. Note that the case that 0 < qij < 1. Again .2) A(1) = A(' 1) ® A(1. For i = j. with common distribution B. Jt = (Jtl). We then let (see the Appendix for the Kronecker E = E(1) x E(2). the definition of Bi is redundant because of f3i = 0.2 for details). and the marked transitions are then the ones corresponding to arrivals. B.
• upon a claim... RETIRED.. more recently. The versatility of the setup is even greater than for the Markovmodulated model. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. INVALIDIZED. DEAD etc.iN. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential.kl is redundant). This means that the environmental states are of the form i1i2 • • • iN with il. u Example 5 . 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. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. In this way we can model. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. iN = all BOi2.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. MARRIED.. Example 5 . Bilo.. E 10.iN C17 AilO. iN. DIVORCED. In fact ... as the Markovian arrival process ( MAP)... assume that there is a finite number N of policies.g. and that the policy then expires.. However .5.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. all Al i2. after which it starts afresh.. or. iN.iil. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . Thus.iil.iN are zero and all Bi are redundant.... WIDOWED.. 11. with rate ai.iN. claims occur only at state transitions for the environment so that AN2.iN = a2..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). say. 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 .. e.. i2i .. THE MARKOVIAN ARRIVAL PROCESS Bij. superpositions of renewal processes.}.. possibly having a general phasetype sojourn time.. Similarly. E = { WORKING.iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant.1i2. the kth policy enters a recovering state.1i2 .
one needs to assume also (as a minimum) that they are measurable in t. let the period be 1. Lucantoni et at. • Claims arriving at time t of the year have distribution B(t). Lucantoni [248]. 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 .3*µs • p = f /3(v) dv 0 0 (6.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. Obviously. • The premium rate at time t of the year is p(t). we talk of s as the 'time of the year'. 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. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. 1). 0 < t < 1. By periodic extension. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . where i f00 xB(°) (dx) _ .176 CHAPTER VI. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. Thus at time t the premium rate is p(s + t). Some main queueing references using the MAP are Ramaswami [298]. )3 t 1 J (6. . For the Markovmodulated model.p)/p. we may assume that the functions /3(t). Sengupta [336]. p * = 0 p(t) dt.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . 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). Without loss of generality. Neuts [271] and Asmussen & Perry [42]. [248]. B* = J f B(t) ((*) dt. continuity would hold in presumably all reasonable examples. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. from an application point of view. p(t) and B(t) are defined also for t t [0. Let 1 1 /3* _ f /3(t) dt. for s E E = [0.
it turns out that they have the same adjustment coefficient. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. the conditional distribution . and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. Thus . u Remark 6 .3) Note that A enters just as a scaling factor of the time axis.1) and Example 1.2 Define T 6(T) = p(t ) dt. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. respectively.t. In particular. the average compound Poisson model is the same as in III.1.9). Thus. Section 4b).3(t) = 3A(1 + sin 27rt). The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . In contrast.w(t)) dt).(3. In contrast. of the periodic model as arising from the compound Poisson model by adding some extra variability. not random.1 As an example to be used for numerical illustration throughout this section.w(t). one may think of the standard compound Poisson model with parameters 3*. p* as an averaged version of the periodic model. The behaviour of the periodic model needs not to be seen as a violation of this principle. equivalently. St = SeI(t). and thus the averaged standard compound Poisson models have the same risk for all A. Many of the results given below indicate that the averaged and the periodic model share a number of main features. the discussion in 111. we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations. 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 . 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. Example 6 . We u assume in the rest of this section that p(t) .6. B*. in agreement with the general principle of added variation increasing the risk (cf. or. let .3* = 3A.10.8. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . It is easily seen that . 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. (6. since the added variation is deterministic.
.g.1) dv . and define h(s.^8 [.s . i. let f 8+1 tc *(a) _ (B* [a] . To this end. we start by deriving formulas giving the m.g. e.a) = exp { .tc* (a)] dv then h (. but it turns out to have obvious benefits in terms of guidelining the analysis of the model as a parallel of the analysis for the Markovian environment risk process.5 (see in particular Remark 11..1) a = J8 . with some variants in the proofs.Q(v) (B(„) [a] .g. a) is periodic on R. we obtain E. The exposition of the present chapter is basically an extract from [44]. see the Notes to Section 7).al.f.east B(8+t) [a] east .(3(s + t)dt)e«St adt + /3(s + t)dt .8). Daykin et. Jt = (s + t) mod 1 P(8) . 0 (5)(u. The claim surplus process {St} may be seen as a Markov additive process.4) At a first sight this point of view may appear quite artificial. As usual.a.1) . of the averaged compound Poisson model (the last expression is independent of s by periodicity).a) Proof Conditioning upon whether a claim occurs in [t. a) etw*(a) h(s+t.adt +.(1 . with the underlying Markov process {Jt} being deterministic period motion on E = [0. of the claim surplus process.T) = P(8)(r(u) <T). 1). Dassios & Embrechts [98] and Asmussen & Rolski [43].(8) [eaSt+dt I7t] = = (1 . Notes and references The model has been studied in risk theory by.e. t + dt] or not.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U.3(v)(B(vl [a] .a .f.3(s + t)dt[B(8+t)[a] . 6b Lundberg conjugation Motivated by the discussion in Chapter II. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00).5. (6.1]) . r(u) _ inf It > 0 : St > u} is the time to ruin .. 3 E(8)eaSt = h(s. given that the ith claim occurs at time t is B(8+t). [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive. . [101] .a be the c. J Theorem 6 .
1]) .4). 0) P(8)a.6 .2. St)} .s. so that obviously {Lo. According to Remark 11.1]. a) . a) = exp I f t3(v)(kv)[a) . a) Thus E(8)east = h(s + t.(8)east 179 = = = = = E(8)east (1 . a) = h(s.0(s + t)dt[B(8+t)[a] . h(s + t.3. St)} and . at + f log h(s + t.1)dv  o h(t.1)dv l og E(8) et where atetk•(a) h(t. + v)(B([a] .t} is a multiplicative functional for the Markov process { (Jt.9) eastt. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.4 For each 0 such that the integrals in the definition of h(t .t = 1 by Theorem 6.. With g the infinitesimal generator of {Xt} = {(Jt. dt log E(8)east a + f3(s + t) [B(8+t) [a] .log h(s. it then suffices to note that E(8)Le.9 as follows. a) Corollary 6.5. we can write Lo Jt.3(s + t)[D(8 +t)[a] .6. a).(e) Let = h( h(Jo.* (a) h(s. u Remark 6. a) h(s + t.c* (e) {Le. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11.5 The formula for h(s) = h(s. 9) is a P ( 8)martingale with mean one.t. E (8)east (a +. B) eoSt t. a) et.t}t>o = h(s.adt +. 0) exist and are finite.1]) . Proof In the Markov additive sense of (6.
Proposition 6. we put for short h(s) = h(s.3. MARKOVIAN ENVIRONMENT ha(s. the requirement is cha(i.1. ( iii) use approximations with piecewiese constant /3(s).7 When a > yo.'y). P(s) (T(u) < oo) = 1 for all u > 0.3(s)dt • B(s)[a]h(s) = gha(s. 0 < s < 1. Proposition 6. Now define 'y as the positive solution of the Lundberg equation for the averaged model. 0) = h(s) + dt {ah(s) . Sdt) = h(s + dt) eadt (1 . cf. y solves n* (y) = 0. For each 0 satisfying the conditions of Corollary 6. B(s). ry)) at which n* (a) attains its minimum. (ii) use Markovmodulated approximations (Section 6c).1) . . Proof (i) Check that m. That is. A further important constant is the value yo (located in (0.0) = Kh(s). (iv) finally.tc] dv} (normalizing by h(0) = 1).3(v)( Bi"i [a] . the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. That is.4. Bet)(dx) = ^ B(t ) (dx). as above E (s) ha(Jdt.6 The P(s).3(s)B(s) [a]h(s). That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. such that for any s and T < oo. However.y) = eayh(s).5 that we can define a new Markov process {(Jt.f.g. see [44] for 11 a formal proof. St)} with governing probability measures Fes). yo is determined by 0 = k* (70) = QB*. J s [. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . it follows by Theorem II.2.180 CHAPTER VI.T.a .3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. [70] . of St is as for the asserted periodic risk model.(3(s)dt) +.6 ( s ) exp { 0( s )&s) [a] + tc . say. correspond to a new periodic risk model with parameters ex . When a = y.60(t) = a(t)B(t)[0].3(s)h(s) + h'(s) +. Lemma 6 .
9 Assume that there exist open intervals I C [0. q) = eryx/h(q)). Lemma 6 . Corollary 6.9) and noting that weak convergence entails convergence of E f (^(u).1) the distribution of (l: (oo).2.9(u))} u>0. ^(u) = ST(u) .9) 0(')(u) = h(s. u which is > 1 by convexity.6. 0(u)) * (b(oo). xEJ 0 (s)b(8)(x) > 0. f (x.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. which is not used elsewhere in the book.8) (6. e(cc)) Letting u > oo in (6. Wu). a) a > ry0 (6. B(oo)). Here and in the following. RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. a) TI h(9(u).u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. 1). and no matter what is the initial season s. say s0. considered with governing probability measures { E(8) }E[ . we need the following auxiliary result . T(u) < (6. we get: .7) h(B(u). has a unique stationary distribution. and we refer to [44].2). a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. T) = h(s. (6. The proof involves machinery from the ergodic theory of Markov chains on a general state space.10) Then for each a. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11. 9(u)) for any bounded continuous function (e. a) e«uE(8 ) e «^ . the Markov process {(^(u). the mean number of claims per unit time is p« 181 = Jo 1.4. J C R+ such that the B(8).g. s E I. have components with densities b(8)(x) satisfying inf sEI.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u.1.
ir) } Plots of h for different values of A are given in Fig. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c.9). elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs .16.1.11) Note that ( 6. illustrating that the effect of seasonality increases with A. Theorem 6 .182 CHAPTER VI. where e.) C = E1 h(B(oo)) u + oo.10) of Lemma 3. where C(o) = 1 + info < t<i h(t) . (6. A=1/4 A=1 A=4 0 Figure 6. 6.10 Under the condition (6. it does not seem within the range of our methods to compute C explicitly. Vi(8) (u) .11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. For our basic Example 6 .Ch(s)ery". MARKOVIAN ENVIRONMENT Theorem 6. Among other things. this provides an algorithm for computing C as a limit. At this stage .1.W.1 In contrast to h. 11 7/'O (u) < C+°)h(s) ery". Noting that ^(u) > 0 in ( 6. 1. 10 shows that certainly ry is the correct Lundberg exponent.6 for the Markovmodulated model: Theorem 6 .
C_h(s)e7u < V.(ay) > 0 when y < 1/ic' (7).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.11 as well as it supplements with a lower bound. Just as in IV. RISK THEORY IN A PERIODIC ENVIRONMENT Thus.6.w)e4u .3x + (1 . whereas ay < y.7x j dx _7x } _ 6w + 6(1 .w)e4u dx 9w + 7(1 .47r sin 27rs + 167r cos 47rs . We state the results below.13) Elementary convexity arguments show that we always have ryy > Y and ay > ry. we first note that the function fu° exu {w • 3e . T) and replace the Lundberg exponent ry by ryy = ay . (ay). Consider first the timedependent version of Lundberg's inequality.13 to our basic example.7e . r.0(8) (u+ yu) (6.(s)(u) < C+h(s)e7". (6. 1 (6.. e7 ( yx)B(t)(dy) > Then for all $ E [0.16 In order to apply Theorem 6. yu) 000 (u) .12 Let 00)(y) 1 Then info < t<i h(t.167r I Cu. 1 ) and all u > 0.w) . in our basic example with A = 1. we substitute T = yu in 0(u.yr.w ) • 7e u{w • 3e3x + ( 1 . the proofs are basically the same as in Section 3 and we refer to [44] for details.14) < C+)(y)h(s) e7yu.42 so that 183 tp(8) (u) < 1.g. where ay is the unique solution of W(ay) =y• (6. ay) • (6.4.17) (6. Theorem 6 . we obtain Co) = 1. e.15) The next result improves upon the constant C+) in front of eryu in Theorem 6. Theorem 6.(8) (u.12) As for the Markovian environment model. . #c( ay) < 0 when y > 1/tc'('y). Lundberg's inequality can be con siderably sharpened and extended.42 • exp {J_ cos 27rs .
013. Some of the present proofs are more elementary by avoiding the general point process machinery of [44]. yo).. 0 <'p(8)(u ) .181 s(u) < 1.. and in fact.19 } 0 <8<1 8 + cos 21rs Thus e. Then 7oudT .66. .cos 27rs . .66. we have the following result: Theorem 6 .L sin 27rs + 1 I cos 47rs .184 CHAPTER VI. 1/i18 1 s (u) > 0.20 •exp { 2n cos 27rs .19 I eu. exp 2^ cos 21rs .013. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment. . for A = 1 (where 3 e0. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'.\ = 0 .(8)(u. 1). 1) for the environment). Thus C_ = 2 inf ex cos 2irs . MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e. the nth Markovian environmental process {Jt} moves cyclically on {1.g.4^ sin 2irs + 16^ cos 41rs .18) Notes and references The material is from Asmussen & Rolski [44]. C+ = 1.\ 3 C+ = sup 6 exp { A (. Of course. but thereby also slightly longer.1 sin 2irs + 16_ cos 47rs .1 sin 27rs + 1 cos 47rs . where the environment at time t is (s + t) mod 1 E [0.16. with s the initial season. completing a cycle . 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. 14 Let C+('yo) be as in (6.20). Thus.'Yo)e (6. Finally.T) < C+('Yo)h( s.0. n}. and let 8 = er' (Y0).16) with 'y replaced by yo and h(t) by h(t. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0.I eu.
it is desirable to have formulas permitting freely to translate from one setting into the other. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. (6. 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.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. Let 0j.7. M(n) = Supt>o Stn). . This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade.jEE of the risk process.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. since the settings are equivalent from a mathematical point of view. one simple choice is Oni = 0( i . To this end. Thus.1 ((i 1)/n) ) and Bni = B . We let {Stn)} (6. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. Notes and references See Rolski [306]. Bi. but others are also possible. 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.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. DUAL QUEUEING MODELS 185 within one unit of time on the average . z/'i (u. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u).
3) 7ri where (V. . JT = i) = P(V > u. JJ = i). Taking probabilities and using the stationarity yields 7riPi(T(u) < T. Proposition 7. For (7. In particular. JT = Z).4) where 0* = >jEE 7rj/3j. • The queueing discipline is FIFO. I* = i). Now let In denote the environment when customer n arrives and I* the steadystate limit. just sum (7.3). J*) is the steadystate limit of (Vt. JT = j} and {VT > u. ii (u) = it /3 P(W > u.186 CHAPTER VI. and for (7. 2 .2). J* = i) = P.n(VT > u. Proposition 7. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi.. Proof Consider stationary versions of {Jt}o<t<T. (7.1) 7ri In particular..3.. let T . (VT > u I JT = 2). and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V.1) follows. JT = j) = LjPj (VT > u.1) over j. 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. Jo = j. J* = i).2) Oi(u) = 1. .1).T(V > u I J* = i). JT = i} coincide. and (7.0i (u . T) = 7ri 1 P. (7. JT = j) = 7rjPj(VT > u. The actual waiting time process 1W1. Jo = i. J* = i) for all j.1 Assume V0 = 0. JT = i) = 'P. Jt ). The first conclusion of that result then states that the events {T(u) < T.2) and use that limF (VT > u.oo in u (7.=1 . Then Pi(T(u) < T. (7. I* )3i P(V > u. {Jt }o<t<T• Then we may assume that Jt = JTt.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u.P(V > u. (7.
the dual queueing model is a periodic M/G/1 queue with arrival rate 0(t) and service time distribution B(') at time t of the year (assuming w.I *=i). In the setting of the periodic model of Section 6. and of these. Lemoine [242]. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. n=1 N However. P(. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. if T is large.6) (7.=i) a4.4).I..7) of that paper.5) follows from (7. I*) with the timeaverage .T)(T(u) <T) = P(8)(VT > u). The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. and (7.1 is from Asmussen [16].g.3). and one has PI'>(rr(u) < T) = P('_T)(VT > u).4) can be found in Regterschot & de Smit [301]. [243]. on average 0*T customers arrive in [0.7. Proposition 7. see in particular Harrison & Lemoine [186]. a general formalism allowing this type of conclusion is 'conditional PASTA'. that /3(t). with (7. Taking the ratio yields (7. P(1')(r(u) < oo) = P(')(00) > u).7) (7.l. T].3) improving somewhat upon (2.4) and (7. N * oo. J* = i) see W > u. . >u. and further references (to which we add Prabhu & Zhu [296]) can be found there.o. a paper relying heavily on classical complex plane methods. and Rolski [306]. With {Vt} denoting the workload process of the periodic queue.8) For treatments of periodic M/G/1 queue. B(t) have been periodically extended to negative t). P(W >u. I* = i. we have 1: I(W. on average /32TP(V > u. see Regterschot & van Doorn [123]. A more probabilistic treatment was given by Asmussen [17]. The relation (7. (7.
This page is intentionally left blank .
. the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r. with common distribution B and independent of {Nt}.2) tk(u. Zt As earlier. {Rt} moves according to the differential equation R = p(R). i&(u. Thus in between jumps. and the evolution of the reserve may be described by the equation Rt = u .T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite. t] are Nt At = Ui (1.1) (other terms are accumulated claims or total claims). and that the claim sizes U1. z/i(u) = F IinffRt< 0IRo=u 1 (1. resp .i.d. Thus...T) = F(T(u) < T). and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). are i. finite horizon. However . U2.6. 189 .At + p(R8) ds.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 . the aggregate claims in [0.
Thus at deficit x > 0 (meaning Rt = x). where one would try to attract new customers as soon as the business has become reasonably safe.4 Either i. Proof Obviously '(u) < ilb(v) when u > v.'(u)) > 0 so that V'(v) < 1. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 .p2.p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. Proposition 1. pi > p2 and p(r) = One reason could be competition. that {Rt} will reach level u before the first claim arrives. and the probability of absolute ruin with initial reserve u E [p/S. Now return to the general model.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. or o(u) < 1 for all u. Example 1.e. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. Assume 0(u) < 1 for some u. 1 . there is positive probability.i(u) = 1 for all u. rather than when the reserve itself becomes negative. 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. In this situation. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r.Vi(v) u > e(1 . However. i. Example 1. say at interest rate b. we can put Rt = Rt + p/S.2. we get p(r) = p + er. it seems reasonable to assume monotonicity (p(r) is u .190 CHAPTER VII. Hence in terms of survival probabilities. No tractable necessary and sufficient condition is known in complete generality of the model. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. That is. but when the reserve comes above v. P(r) _ p + e(r . dividends are paid out at rate pi . oo) is given by i (u + p/S). say e. when x > p/S. If Ro = v < u.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v.
which was proved in 11. V = p(V)). and hence by a geometric trials argument.2(d). In between jumps. That is. INTRODUCTION 191 decreasing in Example 1.T) = P(VT > u).5 (a) If p(r) < /. Hence ik(u) < 1 for all u by Proposition III. Starting from Ro = uo. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. Theorem 1. Then if u > no. Then 0(u) = P(V > u). (1.+ p(r) exists.1.1. hence Rt < uo also for a whole sequence of is converging to oo. Proof This follows by a simple comparison with the compound Poisson model. appealing to Proposition 111. we have z/i(u) <p(u .b(u.4.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.uo) and. that u zPp(u . cf. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0.f p(Vs) ds. (1. We next recall the following results. In case (a).1.5) and the process {Vt} has a proper limit in distribution . In particular. instead of (1. if and only if V)(u) < 1 for all u. say V.1 and increasing in Example 1. . [APQ] pp.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. one can couple the risk process and the storage process on [0. This is basically covered by the following result (but note that the case p(r) .6 For any T < oo. 296297): Theorem 1. {Vt} decreases at rate p(v) when Vt = v (i. B and (constant) premium rate p.uo) < 1. let uo be chosen such that p(r) > p = 0ILB + e for r > uo.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.2 once more.2(d)). and P(Rt + oo) > 0. the probability that Rt < uo for some t is at least tp(0) = 1 (cf.6) .2) we have t Vt = At .2) for r sufficiently large so that p(oo) = limr. obviously infu<uo z/'(u) > 0. Let Op(u) refer to the compound Poisson model with the same 0. then l/i(u) < 1 for all u. let uo be chosen such that p(r) < p = /3µB for r > uo.. then ?(u) = 1 for all u.3µB for all sufficiently large r. In case (b). (1. Proposition I1I.1. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. { Vt} remains at 0 until the next arrival).e.3. However.
t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. oo) must be the same as the flow the other way.s. say. Note that it may happen that w (x) = oo for all x > 0. (1. B(x) = e.h. An attempt of an upcrossing occurs as result of an arrival. x + p(x)dt]). Considering the cases y = 0 and 0 < y < x separately. and the other being given by a density g(x) on (0. In view of the path structure of {V t }.192 CHAPTER VII. Now obviously.Qw(x) . for the storage process {Vt}. we arrive at the desired interpretation of the r. say when {Vt} is in state y.6 applicable.6w(x) . RESERVEDEPENDENT PREMIUMS In order to make Theorem 1.h. It follows in particular that 0(u) = fg(Y)dy.9) Proof We may rewrite (1.y. where g(x) = p( ^ exp {. Then the ruin probability is tp (u) = f' g(y)dy.8) as the rate of upcrossings.8 Assume that B is exponential with rate b. of (1. the l. say. oo). the flow of mass from [0.8) is the rate of downcrossings (the event of an arrival in [t. of (1. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings.8) as g(x) = p 1 {yo13e_6x +. u Define ^x 1 w(x) Jo p(t) dt.7) Proposition 1. (1. say if p(r) goes to 0 at rate 1 /r or faster as r j 0.Sx}. It is intuitively obvious and not too hard to prove that G is a mixture of two components. t + dt] if and only if Vt E [x. Jo AX) (1.s.7 p(x)g(x) = tofB (x) + a f (x .6x and that w(x) < oo for all x > 0.Sx} dx. and is succesful if the jump size is larger than x .y)g(y) dy. we thus need to look more into the stationary distribution G. yo ^ 1 + oo Q exp {. Corollary 1. Oeax f x e'Yg (y) dy } = p) eaxa(x) . one having an atom at 0 of size 'yo. Then w(x) is the time it takes for the reserve to reach level x provided it starts with Ro = 0 and no claims arrive.8) Proof In stationarity. x] to (x.
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
then {Rt} is the diffusion with drift function p+Ex and constant variance a2.x) dx e.3a/ (5 . assume that {Wt} is Brownian motion with drift µ and variance v2. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E. where V is Gamma(b. i.pa. RESERVEDEPENDENT PREMIUMS u from which (2. /^ u Example 2 .13 /E) r (.3/E) By the memoryless property of the exponential distribution. .2) follows by elementary algebra.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments.V < x)]0 + f P(V > p/E ) + eby fv (p/E . it follows that logH[a] = f 1 c(y)dy = 1 f '(pa/(a +y))dy f 0 0 Ey R/E 1 [pa + )3log 8 . As an example.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . The process {St} corresponds to {Wt} so that c(a) or2a2/2 .5 The analysis leading to Theorem 2. of Z is IogH[a] = f ytc(y)dy = e fa (0. 13/E).pa.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2. Proof 2 We use Theorem 2.V.f.e. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .2y +µ ) dy . and the c. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.2) follows by elementary algebra. From ic(a) = .g.a) .3. r (j3/E) In particular.200 CHAPTER VII.
Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. Further studies of the model with interest can be found in Boogaert & Crijns [71]. A r. [283]. se e. Goldie & Griibel [167]. It must be noted.3.1) .i. [129]. see e.e. 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). Gerber [155]. 3 The local adjustment coefficient.3 is from Harrison [185]. Q2/2E).v. as in the proof of Proposition 2.Y*p* W*(u) < ery*u = 0. write y* for the solution of the Lundberg equation f3(B[ry *] . [282].4 is classical. not even Erlang(3) or H3. or to nonlinear premium rules p(•). however. Gerber [157] p. Paulsen & Gjessing [286] and Sundt & Teugels [356]. write Vi* (u) for the ruin probability etc. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) .2 is a special case of a perpetuity. Corollary 2. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. Some of these references also go into a stochastic interest rate.3) was derived by Emanuel et at. and since RT = 0 by the continuity of Brownian motion. The formula (2.d. of the form Ei° p"X" with the X„ i. that the analysis does not seem to carry over to general phasetype distributions.p*. Delbaen & Haezendonck [104]. Emanuel et at. it is also used as basis for a diffusion approximation by these authors. and recall Lundberg 's inequality .g.g.. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case). [129] and Harrison [185]. for a martingale proof. Paulsen [281].8. [357].. Z is normal (p/E.
x>0 (3.C*ef*". we will use the local adjustment coefficient 'y(x).2) such that p(x) < c(.1) . then log u (u) In the proof as well as in the remaining part of the section .1. a first step is the following: Theorem 3 .e. For the last asssertion .3) When trying to extend these results to the model of this chapter where p(x) depends on x. c(.i)eex.1 ).w (u) J dt > c(3)eeu v 1 p(u+ t) . When u > uo. (3. If 60 s f 6o.1.e..'y ( x)) = 0 where r. oo for all E > 0. 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 . (3. Let y* < So.202 CHAPTER VII. e(1o+e)2 (x ) u > 00. and (for simplicity) that inf p(x) > (3µs . RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. log ?i(u) < < 00 JO . 1) and for a given E > 0.>o 7(x) > 0. (x. i. The steepness assumption and p(x) + oo ensure 'y(x) * So. B(x) > C(2)e(ao+f)x for all x. Then lim sup u>oo u and e E''p(r) + 0. choose c(. as solution of the equation n(x. and that p(x) * oo. as will hold under the steepness assumption of Theorem 3. Proof of Theorem 3.. choose uo such that p( x) > p* when x > u0E.E).5) which implies inf. let p* be a in (3. a) = f3(B[a] . Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. it holds that f3[s] T oo.log '(u)/u < ry*(1 .ap(x). which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„. x * oo. 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). i.*(u) .1 Assume that for some 0 < 5o < oo.4) we assume existence of y(x) for all x. The intuitive idea behind introducing local adjustment coefficients is that the classical risk model with premium rate p* = p(x) serves as a 'local approximation ' at level x for the general model when the reserve is close to x. 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) .
The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx.g. (u) = O(u/e). . The form of the result is superficially similar to the CramerLundberg approximation.' (u) < eI("). and in particular. one can then assume that e = 1 is small enough for Theorem 3. u Obviously.3. Condition 3.. 3) = (1 . Then .e. noting that in many cases the constant C is close to 1. ruin will occur if the claim is at least u + v.0 are the same.v. 3. For e > 0.6) The second main result to be derived states that the bound in Theorem 3. Then limelog l/ie (u) = I(u). (3.1 only presents a first step. the result is not very informative if bo = oo. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c.3 Assume that either (a) p(r) is a non decreasing function of r.4)eE" Given such an arrival.2). I. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. Theorem 3 . the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. 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. {Rte)} defined as in (1. UJU > x cannot have a much heavier tail than the claim U itself. Theorem 3.13 is a technical condition on the claim size distribution B. If p(x) = pis constant . 2.ea°/(ecf1)).7) CIO Remarks: 1. which essentially says that an overshoot r.13 below holds. However. or (b) Condition 3. (3. Bucklew [81]). Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(.(u) > so.3 to be reasonably precise and use e` (u) as approximation to 0 (u). by cU2.2 is also an approximation under appropriate conditions. the asymptotics u * oo and c . then Rte) = CRtie for all t so that V). 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.
First.3.(iw(x) .bx} dx fo 00 1 + [exp {/(3w(x) . we show how to rewrite the explicit solution for ti(u) in Corollary 1. it is formally needed only for Theorem 3. u . rather than eI(u)).8 in terms of I(u) when the claims are exponential: Example 3 .bx} dx 1+0.bu}. we consider some simple examples.bx} dx .2.bx}]o + b /' oo exp low (x) .bx} dx = 1 + J0 dodx(x) exp {. 3a Examples Before giving the proofs of Theorems 3. 5.6).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). say. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) . However.(x) dx.1 + b f e.4 Consider again the exponential case B(x) = eax as in Corollary 1. Then y(x) = b .204 CHAPTER VII. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts. the logaritmic form of (3.(3/p(x).3.exp {/33w(u) . RESERVEDEPENDENT PREMIUMS 4. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3. 3.8. As typical in large deviations theory. we get = 1+ J" AX) exp {(3w (x) .
(3. In particular. u . 3.2. (X) = µ(x). BE * 0.ev 0 O /E) J0 70 70 Yo This implies lim inf A. (u) = I(u) + AE .fo 7(x)dx/Edy f .9 ) 11000 eI(v)dy f000 e.2(X) = ev2(x) so that 7e(x) = 7(x)/e.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.10 or Karlin & Taylor [222] pp. we get r 00 e.BE.3 in the particularly simple case of diffusions: Example 3. IE(u) = I(u)/e. For Theorem 3. ry(x /b I u o e f0 °° e  e.7) follows.. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.2. and (3. 70 > 0 such that 7(x) < 7o for y < yo.10) where AE = e log 000 e.fa 7(x+u)dx/Edy o The analogue of (3. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0. Be = e log U000 e. 0.I ( u) fool.fory(x+u)dxdy ( 3.1/8 .. note first that the appropriate slow Markov walk assumption amounts to u. oo) with drift µ(x) and variance a2 (x) > 0 at x. > lime log e = 0 and AE * 0.f y(x)dxd y If 7(x) is increasing .I ( v )dy fo +u) dxdy .e. and (3. It is well known that (see Theorem XI.0. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . Choosing yo.3. Similarly.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3.5) is infx>o 7(x) > 0 which implies that f °O . in the definition of AE converges to 0.5 Assume that {Rt} is a diffusion on [0.1. (3.9) yields e log .fo 7(x) dx /E dy > av 'yo /Edy = E (1 .
Ignoring 1/5 in the formula there. . this leads to (3. 7(x) is typically not explicit.5.7) follows just as in Example We next investigate what the upper bound / approximation aI (°) looks like in the case p(x) = a + bx (interest) subject to various forms of the tail B(x) of B.5. . Of course.(u) oo. that the interchange of the slow Markov walk oo is not justified and in fact. As in Example 3. lim sup Af < lim sup c log(1 .5. I(u ) = G1(u) + . 0 Now (3./3 1 AX dx. so our approach is to determine standard functions Gl (u). G.0/e. Then the solution of the Lundberg equation is y* = b .0.6 Assume that B is exponential with rate S. 0. _ . . > . RESERVEDEPENDENT PREMIUMS The analogue of Example 3. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B.Q/p*..6/p* so that u 1 I (U) = bu .206 CHAPTER VII. .e. G.7o C 15 I I. + Gq(u) + o(G9(u))• Gi (u) It should be noted . we have 5 > 7o and get lim inf AE > lime log e .+1 (u) = o( 1). the slow Markov walk assumption means 5E = b/c.1 3.5) and 7* = 5 . Nevertheless .0) = 0. I. E+o e*O By (3... .5 for risk processes with exponential claims is as follows: Example 3 .6) exactly as in Example 3. Thus 7e(x) _7(x)/e and (3..4. Further. the results are suggestive in their form and much more explicit than anything else in the literature. G. however . (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large.
fu I(u) Su . in the phasetype case . u(logu + r7loglogu). say 1 is the upper limit and B(x) . Here B[s] is defined for all s and B[s] .Y .c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . if the phase generator is irreducible ( Proposition VIII.:. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S . (3. I(u) Pt.ry*°p*.x)n1. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . For example. more generally.3cse7*I7(77) . c4 = c2b 1/'/(1 .y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc. u Example 3 .8 Assume next that B has bounded support. This covers mixtures or convolutions of exponentials or. . 1.C2p* C2 = (3clr( a))11'. e.4) or gamma distributions.7 Assume that B(x) . 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.1) leads to (S7T N Ocp a. and hence (3. Hence (3. More precisely. ry* loge*+ g7loglogp*.g. phasetype distributions (Example 1.8). It follows from (3.11) with a > 0.3. the typical case is a = 1 which holds .clxale5x 207 (3.s)C' f "o o as s T S.cs(1 . 2.c2 Su a dx ) Su a<1 Su .12) with y > 1. x T 1.11) that b[s] * co as s f S and hence 7* T S as p* + oo.. .1) leads to . 77 = 1 if B is degenerate at 1.1/a). y = 2 if B is uniform on (0.1/k).
(b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3. this is only possible if 7o(v) 2 7o(u)• . h 10.e7o ( u)(ul+u r.13) We get b[s] . 3b Proof of Theorem 3. h].B[7o (u)] .f.sp(u).4) is the formula h logEues ( Rhu) .r„(Ti). 1 0 3e. 7 * . Then: (a) y(x) and 7o(x) are also nondecreasing functions of x. 1 = E..208 CHAPTER VII. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .log p*. I (u) c8u log u 0 where c8 = 2/c7. The assumption implies that ru(t) .1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .(T1)) > Ee7o(u)(ul+vr»(Ti)). x f oo . (3.10 Assume that p(x) is a nondecreasing function of x.(t))dt.12).Ote7o( u)(u.11) and (3.3 (B[s] .css 2%rc7eC782/2. . Hence for u<V.f.9 As a case intermediate between (3.14) for the m .Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u). of U1 + v . of the increment in a small time interval [0.c8 log . (3.4).1) . ec78)2/2c7 dx C7 . assume that B(x) CO x2/2c7.g. one could also have considered the increment ru (T1) . g. By convexity of the m . (3.ru(TI)) .r^.u is a nondecreasing function of u.•.u .2 We first remark that the definition (3.4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3.15) Proposition 3. RESERVEDEPENDENT PREMIUMS Example 3 .
this is only possible if yo(u) > 7(u).u[70(u)] fo eyo(x)dx .Fu(u ) + J ^(n)(u . Also.ru(T1) < x). Then (u) < efo Yo(x)dx.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] .11 Assume that p(x) is a nondecreasing function of x.17) from which the theorem follows by letting n + oo.3.2 in terms of 7o. THE LOCAL ADJUSTMENT COEFFICIENT For (b). it is easily seen that fu x7o(y)dy < xyo (u).7o (u)p(u)• Since (3. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3. Separating after whether ruin occurs at the first claim or not. fa 7o(y)dy < u7o(u) < xyo (u) for x > u. note that the assumption implies that ru(t) .(n+l) (u) 1 .u > tp(u). the case of 7 then follows immediately by Proposition 3.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j. we obtain „I.1) .4) considered as function of 7 is convex and 0 for y = 0. Assume (3. u We prove Theorem 3. (3.10(b): Theorem 3. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.17) shown for n and let Fu(x) = P(U1 + u . Hence „/. Hence 1 = EeYo(u)(U1+uru(T1)) < E.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.
x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3.I. resp. The probability of this is at least n n.n. 3.210 CHAPTER VII. in .nbe C*.: y(u). in accordance with the notation i/iE (u). we have chosen to work with y(u) as the fundamental local adjustment coefficient. and here it is easily seen that yo(u) . For these reasons.3/e and U.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x . C*e where the first equality follows by an easy scaling argument and the approximation by (3.E (u) denote the ruin probability for the classical model with 0 replaced by .x/n.n so that n.n AX). However. 0 It follows from Proposition 3. pk n = uk_l.E (u/n).n. 0.n) must first downcross unl.E (u/ n) Y'E (un . P k. Proof For ruin to occur.2. Lemma 3. op*.2). RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3.n) pn niE (u /n) n n_1 n. 0.3).. given downcrossing occurs.. {RtE)} (starting from u = un.10(b ) that the bound provided by Theorem 3..3).10(a) for some of the inequalities. To this end. let Op*. we used also Proposition 3.E ( u/n) ^•e. Further. (u).n. yo(u) appears more difficult to evaluate than y(u).n. the value of {R(E)} at the time of downcrossing is < unl. by €U=..e (u) = v'.n u k}1.n <Z auk}l. W O .n = ku.12 lim sup4^o f log O. Also. y* evaluated for p* = Pk.11 is sharper than the one given by Theorem 3. the probability that ruin occurs in the CramerLundberg model with p* = pn.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. (u) < I(u)..3 is required.n) > k =1 II v ^k n.2. Let Ck. for either of Theorems 3. ryk.E (u/n) Now as e . (un2.n inf n uk1.n = sup p(x). (3. 3c Proof of Theorem 3.. define uk.15). and. Y*u /E.n (starting from u/n) without that 2u/n is upcrossed before ruin.
E (urn) < \ *I..e. k=1 k=1 n u _ nE7 k. so that Theorem 3. v.3. in obvious notation one has tC (x) = y(x)/e. B(x) (3.n. 3 now follows easily in case (a).i..nu /fn( 1 Ck  e. we need the following condition: Condition 3.n <X<Uk. 211 Clearly.n = sup ?'(x).n cE (2u/n) Ck ne7k... also ryk. In case (b). .E (u/n) Op•. ne7k. It follows that n log V'C (u) k =1 log Ypk. since ry' is an increasing function of p'.log Ck. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y).18) (ii) the family of claim overshoot distributions is stochastically dominated by V. Indeed . 40 Combining with the upper bound of Lemma 3. V < oo such that (i) for any u < oo there exist Cu < oo and a (u) > supy <„ 7(x) such that P(V > x) < Cue.7k. (3.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).n + 0(1).n .F (2u/n). THE LOCAL ADJUSTMENT COEFFICIENT particular.12 completes the proof.13 There exists a r. for all x .nu /En) o(1)).19) .a( u)z.nu/en(1 + where o(1) refers to the limit e .! (u/n) n n m 7k. *p•. uk_1. /' (u/n) 'T nk. 11 Theorem 3. 0 with n and u fixed. i.E (u/n) OP +^p•.nk=1 limsupelogv).
1 n.eV) • P (T(E) (u.QEU 1 . u/n) < oo] . The probability of ruin in between two downcrossings is bounded by Epp . For E2.. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. infx>2u /n P(x) ..V) = e71 nu/Eno(l) (using (3. Write EO. (3.R<) (u v). let v < u and define T(E) (u. N with EN < 1 = infx>2u/nA(x) = 0(1). Ei + E2 < e71. v ) = v . V < u/En] + P(V > u/En) (u/En . . u/n) < oo) EV)..E (u/n . ) (u u /n)) . (R. T(E) (u.n < ery1. (u/n .EV) = e.. u/n) < oo] E [OE (u/n .E(E) (u. u /n) < oo] l = = < E [OE (u/n .( . u/n) < oo) .EV) = El + E2. Then the standard Lundberg inequality yields El < E?. v ) = inf { t > 0 : R(c ) < v R) = u } .18) for the last equality).^'' = E [ .212 CHAPTER VII. (u/n .EV) = EiI 1 . Then Y'E (u) ^(E) (u. P (T(E) (u.2y 1 ' .n V.E (2u/n . RESERVEDEPENDENT PREMIUMS To complete the proof. T() (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. u /en 0(i) _n so that E2 < e2ryl nu/En0(1).v.of:>2 in n(x). u/n)) I T(E) (u. u/n)) .^(E) (u.5) and the standard formula for b(0).nu/En0(1) .E (0) cf.nu /EnE [e71.
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.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof. one can in fact arrive at the optimal path by showing that the approximation for 0(u. Comparing these references with the present work shows that in the slow Markov walk setup. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u. Djehiche [122] gives an approximation for tp(u. 0 ) (= p(x) . .13.21) (the initial condition is r(0) = u in both cases). 0 and b T 00 are interchangeable in the setting of [89]. Typically.1.=1 J An approximation similar to (3. s) as in (3.J y(Rs)dR. the risk process itself is close to the solution of the differential equation r(x) _ r (x.g. whereas the most probable path leading to ruin is the solution of r(x) _ k (x.7) then comes out (at least heuristically) by analytical manipulations with the action integral.J r(Rs)p(R.3EU) (3. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e .21) to pass from u to 0. 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 3.20) (with ic(x. Bucklew [81]). it might be possible to show that the limits e .7(x)) (3. Similarly.) = exp .r.)Ui } . Whereas the result of [122] is given in terms of an action integral which does not look very explicit. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.4) and the prime meaning differentiation w. u Notes and references With the exception of Theorem 3.t. s).3. T) is maximized over T by taking T as the time for (3. the results are from Asmussen & Nielsen [39]. l o JJJ o .7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al.T) = P „(info<t <T Rt < 0) via related large deviations techniques.)ds + Y(R2. [89].
3. however. 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. . For different types of applications of large deviations to ruin probabilities . RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . We should like. the exponential distribution ). see XI. e.214 CHAPTER VII.g..
Typically. Note that since (1. 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. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . 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. P. on Eo = E U {A} where A is some extra state which is absorbing. refers to the case Jo = i. We often write p for the number of elements of E. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. if v = (vi)iEE is a probability distribution.1) is 'Here as usual . A distribution B on (0. More precisely. if a problem can be solved explicitly when the relevant distributions are exponentials. 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<. that is. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. and not in other cases.
A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. i. that is. i. Then a = a1 = 1. 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.2) The interpretation of the column vector t is as the exit rate vector. We now say that B is of phasetype with representation (E. and we have t = Te. B(t) = Fa(^ < t ). k}. t1 = /3. In particular. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process.e.1 The phase diagram of a phasetype distribution with 3 phases.e. T is a subintensity matrix2. j. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. T) (or sometimes just (a.0 = t11. E = {i. The initial vector a is written as a row vector. tij > 0 for i 54 j and EjEE tij < 0 . the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. Equivalently.3. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. 0 2this means that tii < 0. C is the lifetime sup It > 0 : Jt E E} of {Jt}.1 Suppose that p = 1 and write .T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}.216 CHAPTER VIII. a. Here are some important special cases: Example 1 . (1. an exponential distribution with rate parameter .
1)!e Since this corresponds to a convolution of p exponential densities with the same rate S.. 0 SP 0 and the phase diagram is (p = 2) ... 0 0 0 0 S 6 ..3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51. so that the density is P E ai6ie6. .1.. 00)) S s o . 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •. .. p}. 6.x i=1 Thus E _ Si 0 T 0 S2 0 0 . 0 •.2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p. the EP distribution may be represented by the phase diagram (p = 3) Figure 1.. a = (1 0 0 ... .. 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 .. .. .. PHASETYPE DISTRIBUTIONS 217 Example 1. 0 S 6 Example 1.2 corresponding to E = {1. ..
(c) the m. i. the backwards equation for {Jt} (e.218 CHAPTER VIII. the restriction of P8 to E. see A. dp. MATRIXANALYTIC METHODS Figure 1.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature.1 tP1 1 Figure 1. j E E. ds^ = ds' = ttlaj + tikpkj. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3.3 0 Example 1 . 36) yields s d.1)"n! aT"e.4 For example.aeTxe. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . Then for i . T). p:. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . a.e.g. The basic analytical properties of phase type distributions are given by the following result . Theorem 1 . the Erlang distribution is a special case of a Coxian distribution. 5 Let B be phasetype with representation (E.d.t2 yt bP.3 . Then: (a) the c. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. [APQ ] p.f. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. (b) the density is b(x ) = B'(x) = aeTxt. .f is B (x) = 1 .g.
T) 1t.tii is the rate of the exponential holding time of state i and hence (tii)/(tii . and (b) then follows from 1: aipF. Part (d) follows by differentiating the m. and since b[s] = ah. = aPxe. After that. we i w. .s I . Rewriting ( 1. d8 P8 = TP8. For (c).jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). this means in vector notation that (T + sI)h = t.6) .tii we go to A.f. i.g.n1t = (1)nn!aTn1Te (1)nn! aTne. and since obviously P° = I..1.f.5) Indeed . 1. the rule (A.s) is the m .5) as hi(tii + s) = ti  t ij hj.B(x) = 1'a (( > x) = P. Alternatively. hj .p.T) 't = (. (Jx E E) = this proves (a). Then h tit ti + ti3 h j .12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI .tii tii .5) ki = 1 + tii L jj:Ai tii (1. we arrive once more at the stated expression for B[s]. of the initial sojourn in state i. define hi = Eie8S. h = (T + sI)1t. (1) n+1n!aT . i.1 ) n +l n ! a (s I + T ) . tij / . for n = 1 we may put ki = Ei( and get as in (1.g.f. j#i jEE tijhj + his = ti.p. Since 1 .f.s j# tii i (1.e. d" dsn a (.. in which case the time to absorption is 0 with m .g.tii and have an additional time to absorption either go to state j which has m . PHASETYPE DISTRIBUTIONS 219 That is.g. or w. the solution is P8 = eT8. B(n)[0] = _ Alternatively. ti/ .n lt .
"n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly.s. One obvious instance is the hyperexponential distribution.7) the diagonal form of T is 9 9 1 9 T 10 7 10 70 1 10 6 10 7 0 70 9 1 10 where the two matrices on the r. 0 Example 1. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. T= 2 111 so that 2 2 Then (cf. Consider for example 3 9 a= (2 2). Example A3. 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 .6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer.220 CHAPTER VIII. 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.h. there are some examples where it is appealing to write T on diagonal form. making the problem trivial. see the Appendix. are idempotent.
11aDD. < 1. PHASETYPE DISTRIBUTIONS 1 10 7 10 221 9 6 70 7 9 10 2 +e 6x (1 11 2 2 35ex + 18e6x 35 The following result becomes basic in Sections 4.4b for definitions and basic rules): Proposition 1.f. 0 Sometimes it is relevant also to consider phasetype distributions. and in fact one also most often there allows a to have a component ao at A. (1. hail = E=EE a. T) is then defined to be oo on a set of probability 1. then the matrix m. or one just lets U be undefined on this additional set. 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). This is the traditional choice in the literature. i. .7 If B is phasetype with representation (v.7) Proof According to (A.4.g.T).29) and Proposition A4. a random variable U having a defective phasetype distribution with representation (a.e a mixture of a phasetype distribution with representation (a/llall. i.T) with weight hall and an atom at zero with weight 1 .hall. • The phasetype distribution B is zeromodified. There are two ways to interpret this: • The phasetype distribution B is defective. 5 and serves at this stage to introduce Kronecker notation and calculus (see A. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).1. where the initial vector a is substochastic.e 11BIJ = 1laDD < 1.
it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx.222 CHAPTER VIII. Here is a sufficient condition: Proposition 1. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . Neuts. cf. but in many practical cases.g. i is real and positive. 0 Of course.g. the result follows (with C = (ah)(ve)). Schmidli. cf. . 77 > 0 and k = 0.4c). we give a criterion for asymptotical exponentiality of a phasetype distribution B. In Proposition A5.q be the eigenvalue of largest real part of T.8) Proof By PerronFrobenius theory (A. the text is essentially identical to Section 2 of Asmussen [26].8 Let B be phasetype with representation (a. where C. 2. one has k = 0. Rolski. h can be chosen with strictly positive component. 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..F. In older literature. Lipsky [247]. here k = p1). MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. v. the Erlang case). but the relevant T is not irreducible. not only in the tail but in the whole distribution. Notes and references The idea behind using phasetype distributions goes back to Erlang. Example A5. All material of the present section is standard. B(x) .f. and we have eTx .8). Schmidt & Teugels [307] and Wolff [384]. let . No satisfying . Using B(x) = aeTxe ..Ce7'. 1. x * oo.1 of the Appendix. see his book [269] (a historical important intermediate step is Jensen [214]). The Erlang distribution gives an example where k > 0 (in fact.f. let v.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. but todays interest in the topic was largely initiated by M. distributions with a rational m. (or Laplace transform) are often used where one would now work instead with phasetype distributions.hve7x. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. assume that T is irreducible . See in particular the notes to Section 6. (1. Then the tail B(x) is asymptotically exponential. T). the conditions of Proposition 1.
Then the renewal density exists and is given by u(x) = ae(T+ta)xt. For this reason... ... we refer to U as the renewal measure. but nevertheless.f. U1<t < U1+U2. known. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. Jt={Jt?ul}. and U(A) is then the expected number of replacements (renewals) in A.+UnEA). If B is exponential with rate 0. n=O We may think of the U.. Let U1.1 of the Appendix. (2... is Markov and has two types of jumps . A related important unsolved problem deals with minimal representations: given a phasetype distribution . we denote the density by u(x) and refer to u as the renewal density.T). the problem has an algorithmically tractable solution if B is phasetype: Theorem 2.. if U is absolutely continuous on (0.r..1 Consider a renewal process with interarrivals which are phasetype with representation (cr. however. or the density is available ) is. .i.t. but is in part repeated below. . 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 +.. JtJt1) Then { 0<t<U1 ..1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together. with common distribution B and define4 U(A) = E# {n = 0. the renewals form a Poisson process and we have u(x) = 0. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. be i.. Lebesgue measure. + U0 is 0 ..d.1. +UnEA} 00 = EEI(U1 +.g.: U1 + . oo) w. U2. as the lifetimes of items (say electrical bulbs) which are replaced upon failure. 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.
T + ta).IIBII which is > 0 in the defective case.1 Corollary 2. B is defective . (b) £(t) has a limiting distribution as t * oo. the density is veTxt = B(x)/µB. define the excess life e(t) at time t as the time until the next renewal following t. fi(t) U2 U1 . i.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. 2. . and the distribution of Jx is ae ( T+t«)x. . and the jumps of the first type are governed by T. The renewal density at x is now just the rate of jumps of the second type. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj . Hence the intensity matrix is T + ta. is the first k with Uk = 00.3 Consider a renewal process with interarrivals which are phasetype with representation (a.T) where v = aT1 /µB. IIafl < 1. T). Proof Just note that { it } is a governing phase process for the lifetime. since Uk = oo with probability 1 .e. Then the lifetime is zeromodified phase type with representation (a. this is welldefined. that is.T).U1 U3 U2 U3 U4 Figure 2. Equivalently. However.. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt.1. see Fig. Corollary 2. the lifetime of the renewal process. u Returning to nonterminating renewal processes . which is ti in state i.. as the time of the last renewal.224 CHAPTER VIII. This is defined as U1 + .T) where vt = ae (T+ta)t . Hence ( 2. and let µB = aTle be the mean of B.1) follows by the law of total probability. + Uit_1 where s.e. and hence ( 2. u The argument goes through without change if the renewal process is terminating. i. which is phase type with representation (v.1) remains valid for that case.
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. the unique positive solution of ve = 1. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. T1 and eTx commute. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2. cf.T) where vt is the distribution of it which is obviously given by the expression in (a). i. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB.1. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . hence e(t) is phasetype with representation (vt.e. The renewal density is then aeQtt = (al a2) ( 7i 7"2.) ( t2 ) .e.q2. we first compute the stationary distribution of Q.2. Al. The formulas involve the matrixexponential of the intensity matrix Q = T + to = ( tll + tlal t12 + t2al tlz + tlaz _ q1 ql t22 + t2a2 q2 q2 (say). (2.4 Consider a nonterminating renewal process with two phases. u Example 2 . (ii) First check the asserted identity for the density: since T. According to Example A3.6.2) v(T + ta) = 0.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt.
`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 .52a1. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .a27r1) (t1 . MATRIXANALYTIC METHODS e.t2) .5 Let B be Erlang(2).4 yields the renewal density as u(t) = 5152 e.t2) 1 + eat (a17r2 . Then Q= 0 55 )+(1o)=( j ad ). Hence 7r = (1/2 1/2). .4 yields the renewal density as u(t) = 2 (1 .tl) 7r2t2 + eat (a17r2 .e2bt) 13 Example 2 .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.a27rl) (tl . )t (51 . and Example 2. and Example 2. t1B 0 Example 2 . The present treatment is somewhat more probabilistic.6 Let B be hyperexponential.(biaz + aza. A = 25.
) = F(ST(o) E •. 3. 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. add a more selfcontained explanation of why of the phasetype structure is preserved. Corollary 2. a+j.p. use the phasetype representation of Bo.1 on the next page. 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. T). Since the results is so basic. T + to+).1 Assume that the claim size distribution B is phasetype with representation (a. marked by thin and thick lines on the figure. however. The stars represent the ladder points ST+(k). and rewriting in matrix form yields the phase generator of {my} as T + ta+. which occurs at rate ti. . we shall. with 0 denoting the Poisson intensity. B the claim size distribution. r(u) the time of ruin with initial reserve u. (b) V. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. and M is zeromodified phasetype with representation (a+.2.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. Corollary 3. {St} the claim surplus process. Within ladder steps. Considering the first. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). THE COMPOUND POISSON MODEL 227 3 The compound Poisson model 3a Phasetype claims Consider the compound Poisson (CramerLundberg) model in the notation of Section 1.3.i. cf. The essence is contained in Fig. Here we have taken the terminating Markov process underlying B with two states. T) where a+ is given by a+ = . the Markov processes representing ladder steps can be pieced together to one {my}.e. Next. We asssume that B is phasetype with representation (a. Then: (a) G+ is defective phasetype with representation (a+. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. G+(. i.f3aT1. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. For (b). Thus the total rate is tip + tia+. and if there is a subsequent ladder step starting in j whic occurs w.3. T(0) < oo) the ladder height distribution and M = supt>o St. T).
 Figure 3. 0 Example 3..3.Q = 3 and b(x) = ..QaT1. T = (3 .t t d kkt S.1 This derivation is a complete proof except for the identification of a+ with . 3e3x + .2 Assume that .M {mx} ST+(2)  S .228 CHAPTER VIII. This is in fact a simple consequence of the form of the excess distribution B0.7)diag so that a+ = QaT 1 = 3 ( 3 2 2) 0 3 9 2 14 7 2 11 2 T+ta+ = 3 0 07/+( 7I \ 2 14 . MATRIXANALYTIC METHODS t . see Corollary 2. 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).1 ..
his derivation of +'(u) is different.T). see Shin [340]. see Stanford & Stroinski [351] .^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3.6). with A denoting the interarrival distribution and B the service time distribution. The result carries over to B being matrixexponential. That is. the discussion around Fig. we encounter similar expressions for the ruin probabilities in the renewal. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). cf.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+. the duality result given in Corollary 11. We assume p = PB/µA < 1 and that B is phasetype with representation (a.1 In the zerodelayed case. For further more or less explicit computations of ruin probabilities. 4 The renewal model We consider the renewal model in the notation of Chapter V. (a) G+ is of phasetype with representation (a+.4. Fig. The parameters of Example 3. T) for some vector a+ = (a+. For an attempt. this was obtained in Section 3.1 can be found in Neuts [269] (in the setting of M/G/1 queues.2 are taken from Gerber [157].1): Proposition 4. and the argument for the renewal case starts in just the same way (cf.4. T). For the compound Poisson model. where a+ is the (defective) . In the next sections. so that as there 229 9 9 e(T+ta+)u 1 9 e_u 10 70 10 70 7 10 Thus 1 7 9 10 ) + e6'4 ( 10 10 .6. 3. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration).and Markovmodulated models. THE RENEWAL MODEL This is the same matrix as is Example 1.j). We shall derive phasetype representations of the ruin probabilities V) (u). see Section 6. if we define {mz} just as for the Poisson case (cf. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. 3.
with intensity matrix Q given by Q = T + to+. Hence by Theorem 11. In fact.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). cf. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. (4. (c) {mx } is a (terminating) Markov process on E. but with initial distribution a rather than a+.*} from {St+y . where a(8) = aT1/PA. Proof Obviously. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}.1) Proof We condition upon T1 = y and define {m. the calculation of the first ladder height is simple in the stationary case: Proposition 4. Then .T). The key difference from the Poisson case is that it is more difficult to evaluate a+. the Palm distribution of the claim size is just B. Since the conditional distribution of my given T1 = y is ae4y. Nevertheless. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). G(') = pBo.T).T)• Proposition 4. Then {m. Fig. Also. 4.1). We have now almost collected all pieces of the main result of this section: Theorem 4 . the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). CHAPTER VIII. obviously mo = m.6. B0 is phasetype with representation (aT1/µa.5.Sy} in the same way as {mx} is defined from {St}.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a.1. which for numerical purposes can be solved by iteration.3. where B0 is the stationary excess life distribution corresponding to B. But by Corollary 2.*'} is Markov with the same transition intensities as {mx}.3 a+ satisfies a+ = V(a+).230 distribution of mo.
.^(8)(u) = a ( 8)e(T+ta +) xe.0. (4..0) is an increasing function of /3. i. (4. It remains to prove convergence of the iteration scheme (4.e. Furthermore .M.1(b). the maximum claim surplus for the stationary case has a similar representation as in Proposition 4.1) and a(8) _ aT.1 . a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . a+ can be computed by iteration of (4.. a+l ) = cp (a+°)) . I {mx} .4.1/pA.2.1 by noting that the distribution of mo is a+. thus . THE RENEWAL MODEL 231 . .•.^(u) = a+e ( T+ta+)xe. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case.2 ) follows from Proposition 4. only with initial distribution a(*) for mo.2) where a+ satisfies (4.3) Proof The first expression in (4...3) (defined on the domain of subprobability vectors . Hence ^p(. by a+ = lim a +n) where a+°) .1). i y ^ T1= y `•r Figure 4. In particular . and that this is given by Proposition 4.3). a+2) = ^p (a+l)) .
4. F[s] being interpreted in the sense of the analytical continuation of the m. a+ ) exists .P[s] = A[s]B[s]. To this end.. this implies that ahA[s] # 0.5 Let s be some complex number with k(s) > 0.ST. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes).T)It.4). (4. s ¢ sp(T).4) whenever EeR(S)U < oo. let F be the distribution of U1 . It follows that n1) so that on Fn the feedback to {mz} after each ladder step cannot exceed &+ a+ n) < a f ^ e(T+ t&+ 1))YA(dy) o < a is e(T+t«+1')YA(dy) _ w (a+1 )) = a+n).f.} can contain at most n . both quantities are just 0 . Thus by (4. 0 0 We next give an alternative algorithm. Then F[s] = a(sI . Obviously.5) Since s $ sp(T).1.4) 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.T1. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1.232 CHAPTER VIII. Then (4.2.T)1t. 7+ ]. Proof Suppose first Qh = sh. n) &+n) T a+. Assume the assertion shown for n . 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n .1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). Theorem 4. Then each subexcursion of {St+Tl . so to complete the proof it suffices to show that &+ < a+) for all n. we use an argument similar to the proof of Proposition VI.g. In that case.5) yields h = (sI . Similarly. (4.T)'t • A[s] (4. However.) = P(mTl = i. To prove the converse inequality. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. . and hence we may assume that h has been normalized such that ahA[s] = 1. the normalization is equivalent to F(s) = 1. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. and let &+". For n = 0. with B[s]. limn4oo a ) < a+. Fn ). Thus . the corresponding right eigenvector may be taken as (sI .
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. Q = CD1 where C is the matrix with columns hl.1 has the d distinct eigenvalues .. THE RENEWAL MODEL 233 Suppose next F(s) = 1. D that with columns p1 hl. .... This gives d roots 'y. Let d denote the number of phases. we get at a(Q . and define hi = (piI . 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 ).T)lt = sh. (4. . hd. In older literature . 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. and hence by the WienerHopf factorization identity (A. Pd with corresponding eigenvectors hl. . Pd in the domain ER(s) > 0 .4. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . Then G+ is phase..... .T)It.p1i .. yd satisfying R(ryi) > 0. W v M(d) in the notation of Chapter V). the matrix Q in Theorem 2. As in Corollary 4. hd. . Given T has been computed. and the solution is . Since R(s) > 0 and G _ is concentrated on (oo. pdhd. Further.type with representation (a+. t(ry) > 0.9) we have G+[s] = 1 which according to Theorem 1. Corollary 4.5(c) means that a+(sI T)1t = 1..T)lt + t = s(sI . we have IG_ [s] I < 1 . Q has diagonal form d d Q = dpivi®hi = dpihivi.6. T) with a+ = a(QT)/at. Notes and references Results like those of the present section have a long history..5.. 0). in turn. .T) = 1 ata+ = a+. .' that the equation F(s) = 1 has d distinct roots p1.lt we get Qh = (T + to+)h = T(sI ...6) i=1 i=1 Proof Appealing to Theorem 4. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 ..6 Suppose u < 0... Hence with h = (sI T). the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1.
Here phase. an alternative approach (the matrixgeometric method ) has been developed largely by M. We assume that each B. In risk theory. the ruin probability can be found in matrixexponential form just as for the renewal model. The distribution of W comes out from the approach but in a rather complicated form . That is . starting around in 1975. In queueing theory.. a pioneering paper in this direction is Tacklind [373]. The exposition here is based upon [18]. [270] and Latouche & Ramaswami [241]. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. The number of elements of El=> is denoted by q. the fixpoint problems look like R=Ao+RAI+R2A2+ . the background Markov process with p states is {Jt}. It turns out that subject to the phase. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme.type assumptions are basic. Asmussen & O'Cinneide [ 41] for a short self. see Neuts [269].) d (see. Neuts and his students.type assumption .contained derivation). The matrix. Numerical examples appear in Asmussen & Rolski [43].. the intensity matrix is A and the stationary row vector is ir . involving .F. The solutions are based upon iterations schemes like in Theorem 4. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t.g. 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. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. The arrival rate in background state i is a. For surveys . with representation say (a(' ). see Dickson & Hipp [118]. where R is an unknown matrix.. E(t)). T('). and the distribution of an arrival claim is B. For further explicit computations of ruin probabilities in the phasetype renewal case .4.exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. whereas the approach was introduced in queueing theory by Smith [350]. and appears already in some early work by Wallace [377]. [119]. e.234 then in transform terms CHAPTER VIII. is phasetype.
•. The key unknown is the matrix K.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. has states o. However.5. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. the analysis involves new features like an equivalence with first passage problems for Markovian fluids and the use of martingales (these ideas also apply to phasetype renewal models though we have not given the details). This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. The two environmental states are denoted o. say with slope r(i) on intervals where It = i. states . the phase space E(°) for B.Vt)} obtained by time reversing the I component.1.2. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. 5a Calculations via fluid models. 5. and the one E(•) for B. 5. p = ql = Q2 = 2.1 In Fig. Section 5b then gives a representation along the lines of Theorem 4. 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. O. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4.6.4.
(Ni)diag r(i. Thus F = {o. In the general formulation . •.1) if and only if s is an eigenvalue of E. F = E U { (i. Second. F is the disjoint union of E and the Eli). we have more martingales at our disposal.1))diag + sII = 0 (5.Vt)} is then obtained by changing the vertical jumps to segments with slope 1.92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. The intensity matrix for { It} is (taking p = 3 for simplicity) I A . 5. consider the vector a satisfying (A + (13i(Bi[ s] .1 A complex number s satisfies 'A+ (f3i(Bi[s] .31a(l) (/3i)diag . First. resp.1(b) {(It . 5. r(i) _ 1. i E E.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . MATRIXANALYTIC METHODS 4. A claim in state i can then be represented by an E()valued Markov process as on Fig. corresponding to the partitioning + Epp). < oo for all s. Eli) + Proposition 5. 2.1(a). of E into components indexed by E. j = 1. 4}. a E E(i) } . The fluid model on Fig . '31a(1) 0 0 f32a(2) 0 0 AI = t(1) 0 0 0 t(2) 0 0 0 t(3) 0 T1 0 0 0 0 T(2) 0 '33a(3) 0 0 T(3) The reasons for using the fluid representation are twofold. a) : i E E. If s is such a number. whereas Ee8s' = oo for all t and all s > so where so < oo. o. 4. 4.1))diag ) a = sa and the eigenvector b = . t. in the fluid model Eel'. 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. Let E denote the matrix . a) of {It}. Bi[s] = a(i)(T(i) + sI)it('). Recall that in the phasetype case.236 CHAPTER VIII. V. a) = 1. This implies that in the fluid context.
d = (sI E22)1E21a = E ai(sI . t(1) 0 0 then also 0 t(2) 0 . c = a. For the assertions on the eigenvectors. assume that a is chosen as asserted which means (Ell .32a(2) (/3i)diag . resp .E22)1 E21a. Then (up to a constant) c = a.Nla(1) 0 0 T 1.5.sI ()3i)diag . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .(sI .E22)1 E21) a = 0. and let d = (sI .E12E22 E21 I . 0 . Noting that E11c + E12d = se by definition.E21a + sd = sd.A .1).T('))1t(i) . MARKOVMODULATED INPUT 237 indexed by E.A . iEE (a> of 0* 1 AI. it follows that Ell E12 ( E 21 E22) (d) = s 1 d I . it follows that if Qla(1) 0 0 . Then E21c+E22d = E21a .sI 0 0 0 T(2) .sI.sI 0 0 t(3) 0 0 = 0.sI+ ((3ia(i)(T(i) . where c. with Eii replaced by Eii .sI)1t)) iag I = 0 which is the same as (5.sI) (sI .E22)1 E21a E21a .E22 .sI + E12 (sI .sI 0 0 0 T(3) . E(1) + + E(P).
v. I' i( V P2 (w (u) < oo. For u.pi(u. Thus 0(u) = esu/d = pe7 ° as u should be.a)d^ ). We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p.. a).. Here E has one state only.4 that {e"1b(v) is a martingale . .Q. a )d(a + e8 °vpi (u . Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u. Proof Writing Or'Alb( v) = svb( v) as (AI . j.v) = j). v > 0.v}.v) = (j. w(u. Then .O. . j) pi( u ./' u = e' (esiuc ( 1) . d("))1 e.. c j. Iw(u. B2 are both exponential with rates 51 i b2.v.upi(u. u) Iw(u.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. q. v) yields C{V) = e8 .v) = = p i( u .j.3 Consider the Poisson model with exponential claims with rate 5.v)=inf{t >0:Vtu orVt=. Example 5 .4 Assume that E has two states and that B1. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s. To determine 0 (u). it follows by Proposition II. e89uc(e)) (d(1) . Example 5 . define w(u. Then we get V)i (u) as sum of two exponential terms where the rates s1.. v = 1. j. . a) = Pi (Vw(u. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b .v) = v) I..v) = Optional stopping at time w (u. a) and noting that i1 (u) = >I j. a) = (j. a)).a Solving for the pi(u.. the result u follows..... . .5. j. pi(u. . MATRIXANALYTIC METHODS Theorem 5...( u.sv)b(v) = 0. a)). w(u)=inf{t >O:Vtu}. j.. v. sq with $2s.. v.j)c v . j.238 CHAPTER VIII.
5 G+(i. i.6 For i E E. dx)Bj(y . (') a T( However .b (u) = Pi(M > u) = 9(i)euue. •) is phasetype with representation (E(i).h.2) the l.3j eye.k.s. (5. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.2.33(e = 0 a(j))(K ®T ( j))(ej (9 I). 9('). In terms of K.3) .x) 00 f ° (') (j) eT (yy)edx . according to VI. U) where t(j) + t(j)O(j j = k uja.y = to B k7 j # k In particular.Qj eie 0 f e (j) T(') x T(j)y ej a e dx e e 00 00 eKx ® e T(')' dx (ej (& I)e T(')ye eKa®T(')x dx (ej (9 I)eT(') Ye e(i)eT(')ye. the Pidistribution of M is phasetype with representation (E(1) + + E(P). j.xxej • a 00 oo el . 8^')IT(j)) where e 3^') =. oo)) j)ye. 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. is 0 /3 f R(i . (y.( 2. Proof We must show that G+ (i. j.5. 0 Theorem 5 . j.
Numerical illustrations are given in Asmussen & Rolski [43].y) to occur when j # k. a) is obviously chosen according to e(`). MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps .k y. some square matrix T and some column vector t (the triple (a.2.p.e. However.. the current ladder step of type j must terminate. the ratio between two polynomials (for the form of the density. This yields the asserted form of uja.. and a new ladder step of type k must start in phase y. T. 0 1)'.e. which occurs at rate t^^7. Associated with each ladder step is a phase process.... which occurs w. if b* [0] = b1 +b20+b302 +. t) is the representation of the matrixexponential distribution/density): Proposition 6. bn1 bn). with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). and lifelength M. i..240 CHAPTER VIII. For j = k. see Example 1.) which is rational. . +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. Furthermore. that the density b(x) can be written as aeTxt for some row vector a. Bk7 . equivalently. Then b*[0] is rational if and only b(x) is matrixexponential. 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. the initial value of (i. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. and it just remains to check that U has the asserted form. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. which occurs at rate t(i). a m.g. (6. intensity matrix U..f.2) . t = (0 0 . For a transition from (j.1 Let b(x) be an integrable function on [0. Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). +bn0i1 0n +a10n1 +. Starting from Jo = i. An alternative characterization is that such a distribution is matrixexponential. a) to (k..5). say. i. we have sofar concentrated on a claim size distribution B of phasetype. we have the additional possibility of a phase change from a to ry within the ladder step.
s = c/ 2 .. S = f c/2 0 21ri . T= 0 0 1 ./(0 + bi).2).1 0 .3) was suggested by Colm O'Cinneide. shows that the distribution B with density b(x) = c(1 cos(21r x))ex.1. MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . Thus. bn of the denominator to be distinct and expand the r. u giving b(x) = E 1 ciebiz/bY. . where c = 1 + 1/47r 2.an_3 an _ 4 ..1 0 0 )3 = (111).. . S..2).e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3.. a2 a1 Proof If b(x) = aeTxt. cannot be phasetype. namely to asssume the roots 6l.. s) is given by 27r i . we can always obtain a real one (a. (6.4) 0 0 1 c This representation is complex.1) as E 1 c. 0 1 an an1 an _2 .3) 0 0 0 0 0 .h.. .3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.47r2 3 .. but as follows from Proposition 6. of (6.T)1 is so. matrixexponentiality implies a rational transform.(6. The converse follows from the last statement of the theorem.s. see Asmussen & Bladt [29] (the representation (6. b(x) > 0 for x > 0. . personal communication). (6. t= 0 .6. then b*[0] = a(0I T)1t which is rational since each element of (01 . (6. 1 . For a proof.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. t). Writing b(x) = c(e( 2ni1 ) y/2 . Example 6 . 0 0 0 0 1 0 0 .47x2 3 1 0 . 0 0 .3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276].. Namely. u Remark 6.2 A remarkable feature of Proposition 6. T. One of his elementary criteria. .
we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. Corollary 111.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. T.1 to get i (u) = f3esus. recall that t = Te) that if B is phasetype and (a. and can use this to invert by the method of Proposition 6. As for the role of matrixexponential distributions in ruin probability calculations. that despite that the proof of (6. (6.1)2 + 6). MATRIXANALYTIC METHODS Example 6 .242 CHAPTER VIII. We recall (see Section 3. t) of b(x). we take as starting point a representation of b* [0] as p( O)/q(9) where p. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). 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. . T the phase generator and t = Te.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . Consider the distribution with density = 15 ((2e2x . Then (cf. t) a phasetype representation with a the initial vector. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1.6) holds true also in the matrixexponential case.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. For the first. and that the minimal number of phases in a phasetype representation increases to 0o as 5 .3. T. then: Proposition 6. we use a representation (a. q are polynomials without common roots. For the second algorithm.5) Thus.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) The remarkable fact is. (6. leading to matrix calculus in high dimensions when b is small. and present two algorithms for calculating '(u) in that setting.5 (6. 0.
Then in Laplace transform formulation .1BVA1. we get (91.T .A . xb(x) dx = aT2t. this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .T)1 (91.T)1ta+(OI .B=land V=a+. From the general matrix identity ([331] p.6.T)1 J0 00 b(x) dx = f aT1t.T)1 + 1 ib* (91. Presumably.'t. since (91T)1T .5 ).1UB(B + BVA1UB). with A = 91T. the assertion is equivalent to a+(BI .7) 9( cf.a+(9I . Now. U =. 519) (A + UBV ).T .1 + b+ = b++ 1 .1 = A1 .1 + 82 (9I . b+ = a+(9I .1t = f3a (0I T)1T1t . (6. we get b+ = 0aT1(9I T).T)1t ( l .6b* .6).T)1t)1a +(9I . b+ = a +(BI . MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .T)1t.to+)1 = (BI .1t du = .to+)1T .T).T)1 so that b* b** b** a+(9I .T)1T 2 = and 1 = AB IT2 + 82T .t.1t = b* .b* (6. (6.1 = ^(T1 + ( 91T)1).to+)1T .T)1 + (6I . but we shall give an algebraic proof.T .T)1T1t. (91. .
h.5 is similar to arguments used in [29] for formulas in renewal theory.8. . but the argument of [286] does not apply in any reasonable generality).7). From this it is straightforward to check that b**/(b+ .T)1)t = 8 (1 .T)1T. 0 Notes and references As noted in the references to section 4. See Fig. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. VII. a. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization.1.1t = /3a (9I . 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. /3aT1(0I . For expositions on the general theory of matrixexponential distributions. which is selfexplanatory given Fig. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. 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. of (6.b*).3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. 7.1. MATRIXANALYTIC METHODS .1) is the same as the r. The proof of Proposition 6.82b*.3 is taken).la. see Asmussen & Bladt [29]. a key early paper is Cox [90] (from where the distribution in Example 6.s. cf.8 a(T1 + (01.1. Lipsky [247] and Asmussen & O'Cinneide [41]. T).244 CHAPTER VIII. In Corollary VII. the ruin probability(u) was found in explicit form for the case of B being exponential. 3.T)1T2t . Much of the flavor of this classical approach and many examples are in Cohen [88]. see the Notes to VII.
>iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7.I] I 80 { tq f Q(v) dvl t1 1 .1) where A(t) = v(u)P(0. 0 < t < u.e. Proof The first statement is clear by definition. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. t + s) . though still Markov.t)).1z I.1 The difference from the case p(r) = p is that {m2}. i. we obtain V)(u) = P(m„ E E) = v(u)P(0. Ai(t) = P(mt = i).t2) be the matrix with ijth element P (mt2 =j I mtl = i). in contrast to Section 3. Given the v(t) have been computed. 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. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u .tl < t2 < u. O<.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . Also. t2) = exp where Q(t) = ds [P(t. is no longer timehomogeneous. Note that in general >iEE Vi (U) < 1. By general results on timeinhomogeneous Markov processes.u)e = A(u)e (7. Figure 7.7. In fact. P(tl. the definition of {m8} depends on the initial reserve u = Ro. Let P(tl. t) is the vector of state probabilities for mt.
(7. 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. Given A'. Given A. Thus. vi.2 For i E E. Given this occurs.t). the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). MATRIXANALYTIC METHODS However. we get vi(u) = aidt + (1 . 0 Thus. whereas in the second case the probability is p(u)dt • tjvi(u). A'(t) = A(t)Q(t) = A(t)(T + tv(u . {mx} has jumps of two types. 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 probability that level u is downcrossed for the first time in phase i is ai. 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. jEE .t) for the second. 0 < t < u.246 CHAPTER VIII.t and being followed by a downcrossing. two things can happen: either the current jump continues from u + p(u)dt to u.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0.Sj i)p(u)dt • tji = Sji + p(u)tji dt.(u) p ( u) = . Hence Q(t) _ T + tv(u . or it stops between level u + p(u)dt and u.3dt.t)). Proposition 7. given A. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . In the first case. from a computational point of view the remaining problem is to evaluate the v(t). the probability of which is 1 .(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 .Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. dt].
This yields v. V . we have p(r) = p = vi (u) 0aTe... refer to the modified process.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7.) + 0 as v + oo. . denotes the time of downcrossing level u . Since the processes Rt and Rt coincide under level B.s. RESERVEDEPENDENT PREMIUMS 247 Subtracting v.^ 0. starting from v"(v) = . supRt>v l t<7 I where o.3 For any fixed u > 0. we can first for a given v solve (7.. then P(A n Bv) _ P"(A n BV'). we face the difficulty that no boundary conditions is immediately available. say. vi (U) = lim v= (u). say. after a certain level v. Thus. (v) is given by the r. F" etc. When solving the differential equation in Proposition 7. Then P(B. of (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. consider a modification of the original process {Rt} by linearizing the process with some rate p.7. From Section 3.4) backwards for {va (t)}v>t>o. To deal with this. <oo.h.i7rT1/p.5)..) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. (u) on both side and dividing by dt yields the asserted differential u equation. (u) for any values of u and v such that u < v. P u which implies that v.2. Now since both P(A n Bv) 3 0 and P"(A n Bv) . Let p" (t). and similarly P"(Bv) .)P"(AnB. Rt .) P"(AnBv) = P(AnB.
e. v = u. cf. Thus we obtain a convergent sequence of solutions that converges to {vi(t)}u>t>o• Notes and references The exposition is based upon Asmussen & Bladt [30] which also contains numerical illustrations.1a. the probability of ruin between a and the next upcrossing of v. We recall from Propositon VII. while the fourthorder RungeKutta method implemented in [30] gives 0(n5)..10 that in addition to the O'(•).q(v dx +( ) ) = ( ) ( q( )) vueTva (7. as well p1(u). let v(u) = a+2ieiT +ta+>)(uv). MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.1.248 CHAPTER VIII. p(r) P. 2u. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. (u)}. (7.. for u > v the distribution of v . To evaluate p1(u).V" M 0 . such that Rt coincide with RI under level v and with Rt above level v. p2.zp1(u)/(1 .7) f o (the integral is the contribution from {R. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. < 0}).1.1. i. which is available since the z/i'(. the evaluation of Vi(u) requires q(u) = 1 . The algorithm based upon numerical solution of a Volterra integral equation (Remark VII. where. T). r<v r > v. 3u etc. However.9.. Therefore u pl(vvueTa t 1. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. 1/n.. Corollary 3.z51(v)). The precision depends on the particular quadrature rule being employed. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. say.7) equals 01 (v . > 0} and the last term the contribu tion from {R.RQ (defined for or < oo only) is defective phasetype with representation (v(u). The trapezoidal rule used in [288] gives a precision of 0(n 3). 0 < u < v. assuming u > v for the moment.) are so. The f iin in (7. where v = inf It > 0 : Rt < v}..x) dx f v(u)eT xt dx .v v(u)eTat 1 1 . 2/n.
Since µB = 5/21.v) + (2^ + 3v2 ea'(u " .u .8) The integral in (7. RESERVEDEPENDENT PREMIUMS 1 .21 = ? yields 0(u) = 1.2. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .and A2 = 3 .v(u)eTVe .v(u)eTve).01 (v .2V"2.x) dx 1 ^(v) ( 1 .e6u 35 . 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 . (7. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t. so we consider the nontrivial case example p2 = 4 and p1 = 1.v(u ) eTV e J v(u)eTxtz/)l (v .e6v Let Al = 3 + 2V'2.x) dx} V 1 1(v) f V v(u) eTxt.24ev . Example 7.1 from which we see that pl (u) = 1 + 1 249  1 . I.(7 The arrival rate is (i = 3..7.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T . all quantities involved in the computation of b(u) have been found in matrix form.jl (t ®e) Thus.2.8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A.be the eigenvalues of T + to( 2 ).24e.4 Let {Rt } be as in Example 3.e.to+))11 {e{T®(Ttoy+ ))}„ . p2 < 3.^1(v) 1 .v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u. From Example 3.
+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.1.250 CHAPTER VIII.1 V2 = 4e5"+6 35e6v . The analysis and the example are from Asmussen & Bladt . MATRIXANALYTIC METHODS From (7.2 35e6v .24es" . 192esv + 8 P1 .1 Thus.21(35e6v . pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .24e5v . and one gets 12e5" .1)' ?. ) e sv + ( 2v/2. Notes and references [30].v)esv + 7 4_ 2.24e5v ./2 ea1(u") ..b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived./2) ea 1(u . 21 3 In particular.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.24es" .
and instead we shall work within the class S of subexponential distributions .4. III. (b) the lognormal distribution (the distribution of eu where U . The definition b[s] = oo for all s > 0 of heavy tails is too general to allow for a general nontrivial results on ruin probabilities. see I. the exponential change of measure techniques discussed in II.N(µ.46 and at numerous later occasions require a light tail. For the definition . 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. a2)) with density 1 e(logyFh) 2/2az . For further examples. we require that B is concentrated on (0. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. for all t > 0. For example.2b.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. x 2iror2 (c) the Weibull distribution with decreasing failure rate .g.f. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. x 4 oo. oo ) and say then that B is subexponential (B E S) if 251 .B(x). B(x) = ex0 with 0<0<1. L(tx)/L(x) 4 1. A rough distinction between light and heavy tails is that the m.
1).1 Let B be any distribution on (0. the r. 1/2 it has the distribution of X1I X1 > x. X2) > x) ^' 2B(x). We later show: Proposition 1. if X1 + X2 is large .F(X1 > x.1(b) is oo. HEAVY TAILS B*2\ 2. Thus .p. As contrast to Proposition 1. P(max(Xi. X2 with distribution B.'s X1.252 CHAPTER IX. Thus the liminf in Proposition 1. X2) > u x)/B(x) = 2.1) then means P(X1 +X2 > x) 2P(Xi > x). then (with high probau bility) so are both of X1.v. (1. we have {max(Xi. 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. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex.2.B(x)2 . (b) liminf BB(() ) > 2. Then: (a) P(max(Xi. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. B(x) ax. . In terms of r. oo). X2) > x}. that is. X1 is w. proving (a). That is. note first the following fact: Proposition 1. Proof By the inclusionexclusion formula.2B(x). P(Xi <yI Xi+X2>x) 1B(y). the distribution of independent r.p. oo). B(x) Here B*2 is the convolution square. given X1 + X2 > x. X2 but none of them exceeds x. x 3 00.3 Consider the standard exponential distribution. To capture the intuition behind this definition. In contrast. the behaviour in the lighttailed case is illustrated in the following example: Example 1. then P(X1>xI X1+X2>x)* 2.'s. That is. and thus the lim inf in (b) is at least lim inf P(max(Xi.v.v. 1/2 'typical' (with distribution B) and w. X2) > x} C {X1 + X2 > x}. X2 > x) = 2B(x) . X2) > x) is P(X1 > x) + P(X2 > x) . one can check that x x where U is uniform on (0.2 If B E S. Since B is concentrated on (0.
Hence lim sup a+oo B*2(x) 2B((1 . B( 0 .xIX > x converges in distribution tooo. then the overshoot X . Let 0 < 5 < 1/2. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.yo]. yo] as X + 00.] Proof Consider first a fixed y. We now turn to the mathematical theory of subexponential distributions. y] and (y. then B(B(x)y) * 1 uniformly in y E [0.6)x)/((1 .'s: if X . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . Finally lim inf B(x .B(y) = 2. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0. 1 < B(x ) B( x) Y) < B( 0). then either one of the Xi exceeds (1 .1.B E S. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. and combining with Proposition u 1. [In terms of r.1(b) we get B*2(x)/B(x) * 2. Proposition 1.4 Any B with a regularly varying tail is subexponential.v. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .5 If B E S. If lim sup B(x .z B(x) . or they both exceed Sx.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0.B(y)) .S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo .y)/B(x) > 1. we get limsupB*2(x)/B(x) < 2. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 .S)x. a contradiction. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. x].B*n(x .y)/B(x) > 1 since y > 0. If X1 + X2 > x.B*(n ) B(dz) (1.
5 and dominated convergence. 0 Proof of Proposition 1. The case n = 2 is just the definition. The first integral is y B(x .z ) (x ) = 1 + (^ B(x . B*(n+1) (x I xy + Jxx y) W.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. so assume the proposition has been shown for n.254 CHAPTER IX. B(x) \Jo _ B(x .z) B(dz). Proof For 0 < 5 < e.2). we have by Proposition 1.z) B(x) Here the second integral can be bounded by B*n(y) B(x) .. x oo.z) B(dz) _y B(x) 111 Lx B . then e"R(x) * oo. Proposition 1.6 If B E 8.B(x . Then by (1. Given e > 0.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . HEAVY TAILS Corollary 1. and this immediately yields the desired conclusions.5 and the induction hypothesis.B*2 (x) B(x) (x . then for any n B*n(x)/B(x) * n. This implies B(x) > c2e5x for all x.5 that B(n) > e6B(n .2. O The following result is extremely important and is often taken as definition of the class S. its intuitive content is the same as discussed in the case n = 2 above. 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 . choose y such that IB*n(x)/B(x) . b[c] = oo for all e > 0.7 If B E S.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.nI < e for x > y.1) for all large n so that B(n) > cle6n for all n. Proof We use induction.
Proposition 1. choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). Then Al * A2 (x) .z) B(x .0 completes the proof. Then by (1. an = supx>o B*n(x)/B(x). Proof Let X1. A2 be distributions on (0.'s such that Xi has distribution Ai.i). e > 0.y)Ai(dy) v) f o .ala2B(x)2 which can be neglected.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .B(x . Combining these estimates and letting a 4. Then Al * A2(x) = P(X1 + X2 > x).4) .y)B(dy) = B(x)ov (1)• v (1. X2 be independent r. 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) .3) Using the necessity part in the case Al = A2 = B yields f xv B(x . Xi <= v Ai (x .z) B(x) < 1 + A + an sup f x B(x . an+1 fX B*n( *n(x .9 Let A1. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x . oo) such that Ai (x) _ aiB(x) for some B E S and some constants al.X2 > xv) < A1(xv)A2(x v) . Since P(X1+X2 > x. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x. Proof Define 5 > 0 by (1+5)2 = 1+e.y))/B(x).X1 > xv. For any fixed v.5 easily yields P(X1 + X2 > x.(al + a2)B(x). 0 Proposition 1.y)Ai(dy) = (x)o(1) (1.2). 0 Lemma 1.z) B(dz) x .z) B(dz) < 1 + A + an(1 + d) .1.v. 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.8 If B E S.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x . a2 with a1 + a2 > 0.
9). Then A * B E S and A * B(x) .256 Now (1. then so is L = L1 + L2.5) Here approximately the last term is B(x)o„(1) by ( 1. u It is tempting to conjecture that S is closed under convolution. Proof Taking Al = A2 = A. with a > 0 and L1.11 Let B E S and let A be any distribution with a ligther tail. a2 = 1. .h. Hence Corollary 1. V (1. u Corollary 1.y)B(dy). u Corollary 1.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1.2A(x).s.5) becomes x B(x . B1 * B2 (x) .12 Assume that Bi(x) = Li(x)lxa.3) follows if CHAPTER LX. L2 are slowly varying.v)Ai(v) . We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. HEAVY TAILS 'VV B(x . B2 E S.v)B(v) + _'U Aq(x .y)Ai(dy) = B(x)o„(1). B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown. That is. That is. i = 1. then A E S. In the regularly varying case. it should hold that B1 * B2 E S and B1 * B2 (x) . a1 = a2 = a yields A*2(x) .aiB(v)) = B(x)o„(1). it is easy to see that if L1. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«.13 Let B have density b and failure rate A(x) such that .(x) is decreasing for x > x0 with limit 0 at oo.2.4). Then B E S provided fo "O exA(x) b(x) dx < oo.B(x) Proof Take Al = A.Ai(x . the l. However. A2 = B so that a1 = 0.Bl (x) + B2 (x) when B1. f " By a change of variables. L2 slowly varying. of (1. if q(x) aB(x) for some B E S and some constant a > 0. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.10 The class S is closed under tailequivalence.2aB(x) . whereas the two first yield B(x)(Ai(v) . A(x) = o(B(x)).
15 In the lognormal distribution.1)xcl1 .A(y)\(y) dy + fox/ 2 eA(x ). Example 1.1. Thus.16 For L(x) slowly varying and a > 1. Thus A(x) is everywhere decreasing. Goldie & Teugels [66]): Proposition 1. the DFR Weibull distriu bution is subexponential. elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing.y) dy. Further. L(x) y° (a .3 < 1. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. the u lognormal distribution is subexponential.14 Consider the DFR Weibull case B(x) = ex0 with 0 <.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x .1 has limit 1 + 0. a(x) = ax01. the first integral has limit 1 . an integrable function by assumption.(x . B*2(x) . f ' L(y) dy .y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).A(x ..13 works in this setting.. Jo For y < x/2. Example 1. we first quote Karamata's theorem (Bingham. Then b(x) = Ox0lexp. we can use the same domination for the second integral but now the integrand has limit 0 . Since ) (x . Thus.e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . Thus B*2(x )/ B(x) .(y) dy.U) /or) v 2x This yields easily that ex. To illustrate how Proposition 1. proving B E S.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ).y) < A (y) for y < x/2. The middle bound shows that it converges to b(y) for any fixed y since \ (x .y) * 0. subexponentiality has alrady been proved in Corollary 1. In the regularly varying case. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). 0 A(x) .2). x . Define A(x) = fo . Then B(x) = eA(x).y) < yA(x . By (1.12.`(x)b(x) is integrable.A(xy)A ( y).A(y)a(y ) = ev'(y) b(y). and exa(x)b(x) = (3x01e(10)x9 is integrable. Thus by dominated convergence .
1) and P(X (.17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.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 . the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull. Then 7(x) . (1 + y/(a . Then: Proposition 1. (c) Under the assumptions of either ( a) or (b). the monotonicity condition in Proposition 1.4 is necessary in full generality.ea b(x) is integrable.1)] I X > x) L(x[1 + y/(a .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0.a/x.258 From this we get CHAPTER IX. then 7(x) x/(a .13 may present a problem in some cases so that the direct proof in Proposition 1. let X W = X . . Thus exa(x)b(x) .1)])a 1 1 . More precisely.6) EX(x) .y(x)B(x).1)X(x)/x > y) = P(X > x[1 + y/(a . 'y(x) = EXix>. HEAVY TAILS Proposition 1.1)]) xa L(x) (x[1 + y/(a .)/Y(x) > y) (1 + y/(a .L(x)/x" and )t(x) .1/A(x) and P(X ixil'Y (x) > y) * e'.E(X .xjX > x. However. we get (1. We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions. then B(x) . yo] .1))a . f O B(y) dy . Proof ( a): Using Karamata's theorem.18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.
Y2. Lemma 2.A(x) I X > x) = exp {A(x) .n0 1•P(K= n)•n = EK. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B.1/. nG(u). Kliippelberg & Mikosch [134].14. cf.nn.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . r(u) = inf it > 0. It is trivially verified to hold for the Weibull. Recall that B0 denotes the stationary excess distribution.v..+YK> u) = ^•P(K = n)G* n(u ) . Notes and references A good general reference for subexponential distribution is Embrechts. We get p(yl+. i. d.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. u a oo.2 Let Y1. 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.and lognormal distributions . P The proof is based upon the following lemma (stated slightly more generally than needed at present).8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. with EzK < oo for some z > 1. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). Then P(Y1 + • • • + YK > u) . 1. The remaining statement (1.. .2.15. then Vi(u) P Bo(u).EK G(u). Examples 1. 0 G(u) L G(u) .(x).7) is referred to as 1/A(x) being selfneglecting. be i. with common distribution G E S and let K be an independent integervalued r. St > u}..1 If Bo E S. .t be the claim surplus at time t and M = sups>0 St. Let St = Ei ` Ui . Theorem 2 . Bo(x) = f0 B(y) dy / µB.
2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. Bo E S. as well as examples where B ¢ S..1) In particular .?(xµ 8 (x). and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. Since EK = p/(1. a].p) and EzK < oo whenever pz < 1. x 4 00. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . Proof of Theorem 2.1 is notoriously not very accurate. Bo is more heavytailed than B .18. Note that in these examples .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 ) .1 is essentially due to von Bahr [56].µB(01 . r(1/Q) xlQexp B(x) = ex' From this .x^ ) B(x) _ f or ( lox . u The condition Bo E S is for all practical purposes equivalent to B E S. then Bo(x)/B(x) + 00. P(K = k) = (1. (2. mathematically one must note that there exist (quite intricate) examples where B E S. Bo E S is immediate in the regularly varying case. see Abate. the result follows immediately from Lemma 2. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant. The approximation in Theorem 2. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1.. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . in our three main examples (regular variation . u x+a Notes and references Theorem 2.p)p'. For some numerical studies. However.13). The PollaczeckKhinchine formula states that (in the setup of Lemma 2.x400 PBB(x) PB Leta+oo.3 If B E S. .1. See also Embrechts & Veraverbeeke [136]. Weibull) one has Bo(x ( B(x) . Proof Since B(x + y)/B(x) * 1 uniformly in y E [0.2. In general: Proposition 2. Bo ¢ S.260 CHAPTER IX. lognormal .µ J ) . The problem is a very slow rate of convergence as u ^ oo. Borovkov [73] and Pakes [280].
+ Xn.. We assume positive safety loading. Based upon ideas of Hogan [200].] The proof is based upon the observation that also in the renewal setting. Then K M=EY. Define further 0 = IIG+II = P(r9+ < oo)... [279]. t9(u) = inf {n : Snd> > u} .3.1. Kalashnikov [219] and Asmussen & Binswanger [27]. T+ < oo) where r+ = T1 + • • • + T. Snd) = Xl +. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. also a second order term is introduced but unfortunately it does not present a great improvement.1 when u is small or moderately large.1 gives 1010. there is a representation of M similar to the PollaczeckKhinchine formula. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. Somewhat related work is in Omey & Willekens [278].} Then ik(u) = F ( M > u) = P(i9 (u) < oo). 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. p = iB /µA < 1.y)/B (x) > 1 uniformly on compact y internals.1) this end .+ E A.Ti. G+ (A) = P(Sq+ E A. This shows that even the approximation is asymptotically correct in the tail. THE RENEWAL MODEL 261 Choudhury & Whitt [1]. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }.. E. The main result is: Theorem 3 . i. To Bo(u) u + 00. M = sup s$ .. Let U= be the ith claim .e. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA). {n= 0. T1 the ith interarrival time and Xi = U. in [219] p. i=1 . Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2.g. In [1].. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x . (3.9+ < oo) = P(S.y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. .. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S.
with distribution G+/9 (the distribution of S. (3.1) is equivalent to P(M > u) " . 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. The heuristics is now that because of (b).N. G_(A) = P(S.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_.3 G+ (x) .FI(u).2). (b) and does not rely on the structure Xi = Ui .Ti). A(dy) = 1.PBBo(x).. 0] normalized by IPG_ I so that we should have to G+(x) . whereas for large y . As for the compound Poisson model.262 CHAPTER IX. Let F denote the distribution of the Xi and F1 the integrated tail. A. cf. u a 00.g+ > x.3) and we will prove it in this form (in the next Section. 0] to the integral is O(F(x)) = o(FI(x)).d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.9)9'' and Y1. d+ < oo).. FI (x) _ fz ° F(y) dy. HEAVY TAILS where K is geometric with parameter 9. x * oo. Then 0 0 F( x . we will use the fact that the proof of (3.2 F(x) . Proof Let R+(A) = E E'+ ' I(S. Write G+( x) = G+ ( x.d.oo. the contribution from the interval (. oo) = F(S. Proof By dominated convergence and (b).y) dy = 1 Pi (X) oo IPG_ I . 0 The lemma implies that (3.. are independent of K and i.B(x).1 IPG_ I / F(x .(. x > 0.i.y) R+(dy ) _ j (x_y)U_(dY) G+ (x) = J 00 00 (the first identity is obvious and the second follows since an easy time reversion argument shows that R+ = U_. B(x) _ J O° B(B(x)y) A(dy) f 1 . U_ (dy) is close to Lebesgue measure on (.y+ given r+ < oo). and hence FI(x) .FI(x) /IPG_I. x + oo.Y2. Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. P(K = k) = (1 . Lemma 3 . Lemma 3 .1) holds for a general random walk satisfying the analogues of (a).
then by Blackwell 's renewal theorem U_ (n .I n=N (1 E)2 r00 F(x + y) dy + e) lim sup . By Lemma 3. Similarly. F(Y= > x) FI(x)/(OIp _ 1). the proof is complete. We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x .9) 1 .1.9)IpG_ I Differentiating the WienerHopf factorization identity (A. n] < (1 + e)/1µc_ I for n > N.1.3. u Proof of Theorem 3. and in the last that FI is asymptotically proportional to Bo E S.1 I .1.2). we can assume that the span is 1 and then the same conclusion holds since then U(n .oo Fj(x) N J (1 +6)2 I {IC_ I lim sup X400 FI(x + N) _ (1 + e)z (x) I Pi µ G_ I Here in the third step we used that (b) implies B(x)/Bo(x) + 0 and hence F(x)/FI(x) 4 0.3.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N. n] + 1/I µG_ I. THE RENEWAL MODEL 263 We now make this precise. In the lattice case. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n . Hence using dominated convergence precisely as for the compound Poisson model.UG_ I x.G+[s]) .2) yields 00 F F I (u) P(M > u) _ E(1 .(dy) fN FI ( x) + lim sup ZY00 N F(x .O[s])(1 .1.=1 BIp G_ I (1. (3. n] is just the probability of a renewal at n.F[s] = (1 . > (1 .0)0k k I(u) A.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10. Given e.y) U. If G_ is nonlattice.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.1. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. choose N such that F(n . and that U_(n .
must attain a maximum > 0 so that P(M > u. u)) > P(w(u) < oo)(i lp (0))• On the other hand.yiui_1. 10(0) But since P(M > u .1 is due to Embrechts & Veraverbeke [136].(u)+n .u)) = o(P (M > u)) = o(FI(u)). S+q(u) . Proof Let w(u) = inf {n : Sid) E (u . u).2.Sty(u)_I < a} we have w(u) < oo. 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. P(M > u. and {Su. Note that substantially sharper statements than Lemma 3.a. S+9(u) .0)ua_ .So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . Then P(M E (u ... 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times..SS(u)}n=o. Notes and references Theorem 3.a. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .4 on the joint distribution of (S.a.l. Mn < u}. we have P(M E (u .AB iP We conclude by a lemma needed in the next section: Lemma 3 . see Asmussen & Kliippelberg [36].1)6+[0] .(1 .IIG+II)µc_ = (1 . allowing also for possible dependence between the arrival process and the claim sizes.264 and letting s = 0 yields CHAPTER IX.a. with roots in von Bahr [56] and Pakes [280]. HEAVY TAILS µF = (1 .Se(u)_1 < a) = o(Fj(u)). u)). Therefore by Lemma 3. on the set {M > u.a) N P(M > u).4 For any a < oo. . Sty(u) .So(u)) are available.
.d... Theorem 4.. 4... (viewed as random elements of the space of Dfunctions with finite lifelengths) are i. {Sn}n=o.1 Note that no specific sample path structure of {St} (like in Fig.. Thus the assumption .1 where the filled circles symbolize a regeneration in the path. M = sup St. see [47]. The idea is now to observe that in the zerodelayed case...n n=0.Sxi}0<t<x2Xl . We give here one of them..F*(X) = P0(Si > x) . Figure 4.. M.1) is assumed. {SX1+t ..1 based upon a regenerative assumption. and apply it to the Markovmodulated model of Chapter VI.. assume pp. For further approaches. E0.. We return to this point in Example 4.1 = max k=0. Schmidli & Schmidt [47].X2 < .1. and the distribution of {Sxk+t . 4.X1 is the generic cycle. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.. 2. (corresponding to the filled circles on Fig. See Fig. . such that {SXo+t  SXo}0<t< X 1Xo .1 except for the first one) is a random walk. We let F* denote the Podistribution of Si.i.1) . G(x) (4... M* = max S.Sxk}o<t<xk+1xk is the same for all k = 1. Define S.4 below. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.4. +1. 0o(u) etc. t>0 S. examples and counterexamples. 4...1. 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 <... < 0 and EoX < oo where X = X2 . = Sx.
4) liminf u>oo F(M > u) ..* i o<t<xn+1x. jF11 F* (U).3) applicable so that F(M* > u) 141 F*(u).266 CHAPTER IX. Sxn +t .2) to show F(M* > u) > 1. See Fig. Since clearly M(x) > Sl . Fo(Si > X). 4.. (4.2. u p 00.1 Assume that (4. The one we focus on is Fo (Mix) > x) .S.2 Theorem 4.Sxn = sup Sxn+t . (4. it suffices by (4.. Proof Since M > M*. Then '00 (u) = Fo(M > u) .3) hold.3) where Mnx) = sup o<t<xn +1 X. N N Xi=0 N Figure 4.. (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. HEAVY TAILS for some G such that both G E S and Go E S makes (3.1) and (4. the assumption means that Mix) and Sl are not too far away.
.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). Let a > 0 be fixed. choose a such that Po(Si > x ) > (1 ..e)Po (MMX> > x). Letting first u + oo and next e .a. E (u .Mn +1 >aV(u n=1 00 S. )) > (1 . S.1 can be rewritten as 00 (U) (4.2.Sn 0<t<x„+j ( 1 .(1 ..e)Po (M > u. Then by Lemma 3. u)} < P(M* E (u . u))/P(M* = 0) = o(Po(M* > u)). Under suitable conditions . Mn+l > a V (u .E) Po ( n max St u. .: Sn > u} . 2.5) which follows since Po (M > u.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u..1 = limti00 St/t. 0 yields (4.6) 1 p pBo(u) u where B is the Palm distribution of claims and p .4. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 .e)Po (M > U). We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4.a. Po(M* > u) . Theorem 4. x > a.4. MW O(u)+1 < a) IN ( U n=1 A1.4).: S.(u) .Po (M* > u. M^xu)+l > a) .S.( u)1 > a) 00 1: Po(Mn<u.... assume the path structure Nt St = EUit+Zt i=1 . /3(u) = inf{n=1. To this end. Given e > 0.
Assume further that (i) both B and Bo are subexponential.3PB. and also for Mix) since Nx FNX U. since the tail of Zx is lighter than B(x) by (iv).6 below.8) x Write .I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .2 Assume that {St} is regenerative and satisfies (4. we get 00 (u) 1 IPF.'s order EoNx • B(x).} and satisfying Zt/t N. independent of {> CHAPTER IX.7). (ii) EozNX < oo for some z > 1. the proof of Lemma 4.1 is in force. X and N. HEAVY TAILS N` U. The same is true for Sl. Mix) < > UE + i=1 o<t<x Thus Theorem 4.v. a4' 0.p) Ju P Bo(u) 1p 0 . Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .Q = EoNx/EoX. cf. i=1 (4.4). Corollary 4.X both have tails of sup Zt. Proof It is easily seen that the r. 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.6) holds with p = . and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . and ENX Ui . (iii) For some o field Y. oX (see Proposition A1.
e. Example 4 .6) u holds. Bo E S. (i) holds.3 that (4. More precisely. Zt . Bo E S. Thus we conclude that (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. Theorem 4. we will assume that lim B2(x) = ci x+oo G(x) for some distribution G such that both G and the integrated tail fx°O G(y) dy are subexponential .5 Consider the Markovmodulated risk model with claim size distributions satisfying (4.4 Assume that St = Zt .4. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. . X2 = 1.t + EN'I Ui where {>N`1 Ui . MODELS WITH DEPENDENT INPUT 269 Example 4 .. < 1.. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i.9). Taking again Xo = Xi = 0. and for some constants ci < oo such that cl + • • • + c. 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.. consider the periodic model of VI. We consider the case where one or more of the claim size distributions Bi are heavytailed. In particular.3 As a first quick application. The arrival rate is /3i and the claim size distribution Bi when Jt = i. The key step of the proof is the following lemma. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. X3 = 2. > 0.6 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). X2 = 1.(NX).6) holds.6) holds.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. i..0 (thus (iv) is trivial). . we assume that B E S. Again . The regenerative assumption is satisfied if we take Xo = Xi = 0. of claims arriving in [0. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. X3 = 2. and taking F = o. Assume that B E S. we conclude just as in Example 4. . Then (4. The number N.. (iii) is obvious. . in particular lighttailed.
.2.c'(x) where c = ciENi . i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.v.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 .. oo) and define p Ni Yx = EEX'i . . .270 CHAPTER IX.. are independent with distribution Fi for Xij. as x a oo."+Np . . It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni.. and F a aalgebra such that (N1. 1. P P P(YX and > x I.^•) G(x) P ^ E ciNi = C. i=1 P(Yx > x ^) < P(Y0 > x I. and that for some + cp distribution G on [0. X > 0 a r.G( x ) > ciNi .}P. u Proof of Theorem 4.Fmeasurable..ciG(x). An easy conditioning argument then yields the result when Jo is u random. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin . The same dominated convergence argument completes the proof.F) < CG(x)zn'1+. we can define the regenerations points as the times of returns to i. i=1 Proof Consider first the case X = 0. HEAVY TAILS Lemma 4 . Assume EzN1+"'+Np < oo for some z > 1 and all i. 2 . Then P P(Yx > x) . and the rest of the argument is then just as the proof of Corollary 4.. Let {Fi}t=1 P be a family of distributions on [0. . NP ) be a random vector in {0. If Jo = i. . .. For lighttailed distributions.X i=1 j=1 where conditionally upon F the Xi.5. 6 Let (N1. oo) such that G E S and some c1. i1 = E\ G(x) In the general case. Thus dominated convergence yields ( P(Yo>x P(Yo>x . cp with cl + > 0 it holds that Fi(x) .. NP ) and X are ..
6) to hold in a situation where the interclaim times (T1.d.7. We start by reviewing some general facts which are fundamental for the analysis. we let PN"N = P(.3. Within the class of risk processes in a Markovian environment... An improvement was given in Asmussen & Hojgaard [33]. cf. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. I T(u) < oo).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..T2. and the final reduction by Jelenkovic & Lazar [213]. cf. Theorem 4. i. IV.5 shows that basically only the tail dominant claim size distributions (those with c. Theorem 2. The main result of this section. Then O(u) . Notes and references Theorem 4. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5. the discussion provides an alternative point of view to some results in Chapter IV. VI..4.7). as well as a condition for (4..2 and Example 4. m is a (orfinite) . and independent of (T1.T2.4. states that under mild additional conditions. Theorem 5. The present approach via Theorem 4.e.1. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. 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. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u. Schmidli & Schmidt [47].p)Bo(u).4. ). i. r(u) is the time of ruin and as in IV.. this is applied for example to risk processes with Poisson cluster arrivals.pl(1 . see Schlegel [316]. in particular Proposition 2. For further studies of perturbations like in Corollary 4. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. It follows from Theorem 4. this should be compared with the normal limit for the lighttailed case.5. ) form a general stationary sequence and the U.1 is from Asmussen. there exist constants Y(u) such that the F(u)distribution of r(u)/y(u) has a limit which is either Pareto (when B is regularly varying) or exponential (for B's such as the lognormal or DFR Weibull). As usual. cf. Combined with the approximation for O(u). Essentially.4. In contrast. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). 5 was first proved by Asmussen. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case.i.
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} .2) means ffh(a. j.)k(x . Let G denote the distribution of ENt U. Proof Starting from Ro = x. say.t + EI U.z) dx G(dz) = ffh(y + z) k(y)dy G(dz).s.r.h. x = 0+ and F = (0.rij = mjsji where r13. Lebesgue measure. Then (5.).t. Rt is distributed as x + t . Sw(F. w(Fc) < oo ) 'In general Markov process theory. Say {St} is reflected Brownian motion on [0. however .>N` Ui. . (note that we allow x.00). For the present purposes it suffices . t.272 CHAPTER IX.= y. St is distributed as y . y to vary in. k as indicator functions. follows by the substitution y = x . the whole of R and not as usual impose the restrictions x > 0. for states i. where we can take h. oo). The equality of the l. {Rt}. a familiar case is time reversion (here m is the stationary distribution). Thus. {St} and {Rt} are in classical duality w. to consider only the case Px(w(F`) = 0) 0. and starting from So = y.h. but the example of relevance for us is the following: Proposition 5.2) with t = 1 means m.s. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0.y = Qx (. m. u For F C E.t. resp.z. r.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . to the r. . and (5.2) for all bounded measurable functions h. Then there is a Markov process {Rt} on E such that fE m(dx)h(x)Exk(Rt) = Lm(dy)k(y)Eyh(St) (5. y = 0). in the terminology of general Markov process theory. The simplest example is a discrete time discrete state space chain. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. We let QS be the corresponding distribution and Qx. k on E.s=j are the transition probabilities for {St}.
Sw(F)1 = y) . in with i0. and we let Qy y refer to the time reversed excursion . . io = x. oo) = r(0) x= St y (a) Figure 5. 0]. 5. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. Qx. S. That is. /^s x (S1 = Z1. w(0. the one in (b) is the time reversed path.s. But in the risk theory example (corresponding to which the sample paths are drawn).2 Qy. in E F.y = Qy Q. We can then view Qy. The theorem is illustrated in Fig . Sn+1 E Fc) nx.. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F.. Sw(Fo)_ should be interpreted as Sw(F^)_1).3 The distribution of r(0) given r(0) < oo. z > 0. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time..2.. Thus.. . . QR and QRy are defined similarly. x = 0.).itt) = P Px(w(Fc) < 00. i1.1 for the case F = (oo. Sn = in = y..(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}.5. .y as a measure on all strings of the form i0i1 .y(2p21 . when p = . [note that w(z) < oo a.y() = P ({SW(F`)t} 0<t<w(F °) E So = x..= y) Theorem 5 .SS(F. in = y. In particular: Corollary 5.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. 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. 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)..13AB < 1] Proof of Theorem 5.
note first that Pt' (R l = il...ik1EF Similarly but easier Sxin_1 . Sn+1 E Fc) n=1 i1. MY Thus Qx(ioii ....TI( 2n2n _1 . . in E F. Si l io E mjSjx.. R . . . 2p) when 20.in1 ..(F<)1 = Y) S S and Qx y( ipil . t' y and Qy x are measures on all strings of the form ipi l .. To show Q y x (i0 i 1 .. ....in E F. (Fc)1 = y) 00 CHAPTER IX. .. i0 = y. ... 21 ... Si11 S 1 ..y(inin _ E SYj jEF` 00 Sxik _1 . = in = y. Rn+1 E F`) F (w(Fc) < 00.ik_1EF Sxin_1 .. i0) Q x. 2p). in = x.... in) = oo jEF^ Sxin1 .ii ....ik_1EF .. Si1y k=1 i1 .... R Qy x(2p21 .i„_iEF Similarly.gilt' k=1 ii ... S.... in) = Qx. Silt' E Sxik_1 . .... Silt' E SO k=1 i1...274 note that Fx(w(Fc) < 00. Rn = in = x.. . in with 20. Si1y 00 jEF° E E 5xik_ 1 . Rn+1 E FC) TioilTili2..rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 .....J (i... Rn = in = x. in = x... 20 = y.. . HEAVY TAILS E E Px (Si = 21i . S...... in)  Pt' (R1 = ii.. 2n) = Qx....
UBBo(u)].r. y > u. ST(o) > y.B(a) +a PBBo(u) .r. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . S. 1 w . Y > u). see Fig. the case r (O) < oo.t.2. that is. that is. P(") = P(. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). the P(u. 7(0) < oo.2 The distribution of (Y. U T(O) = T (u) Y Figure 5.2.2.')density of Y is B(y)/[. Now the P(u. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo.(o) > y} = {T(0) < oo. Bo") is also the P(u.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy .t. P(o) ). That is.')distribution of Yu is Bo"). The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. Y > y} .'s are defined w.5. 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 . the distribution w.p. Z) is described in Theorem 111. 5. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch .v.
Hence Z. HEAVY TAILS Let {w(z)}Z^. Bo") ). . denote the ladder epochs and let Yk. 2.i. cf. In the proof.. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long.p) then yields the final result T(u)/y(u) + W/(1 . and since its dominates the first n . K(u) = n). 5. Z1). Z). > u with high probability... conditionally upon r+ (n) < oo. we get the same asymptotics as when n = 1. . 4 Assume that Bo E S and that (5.. a slight rewriting may be more appealing... this in principle determines the asymptotic behaviour of r(u). . let r+(1) = T(0).p). However. Fig.3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z).d. ... . Z/'y(u) * W in Pi "' ')distribution . Since the conditional distribution of Z is known (viz.T+(2). Then Corollary 5. Then 7(u)/y(u) ^ W/(1 ..r+ (n . Then. Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo.p) in F(u) distribution.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}.1. z ^ oo. are i.: r+ (n) < oo. P(Z < a I Y > u) 3 0.e. r(u)/Z 4 1/(1 . in particular of Z. Since w(z)/z a$.p). We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 . Yn_1 'typical'. must be large and Z1. The idea is now to observe that if K(u) = n. (Y. Zk be defined similarly as Y = Y1.p) in Pi"'')distribution. Recall the definition of the auxiliary function y(x) in Section 1. the random vectors (YI.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. We let K(u) = inf In = 1.. It is straightforward that under the conditions of Proposition 1.18(c) Bo")(yY (u)) + P(W > y) ( 5. more precisely. i.. then by the subexponential property Yn must be large. the duration T+ (n) .. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. 1/(1 . it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 ....3.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). and YI. and distributed as (Y. .276 CHAPTER IX. Zn). . i.3) holds.e. Z = ZI but relative to the kth ladder segment. That is .
Y„1. Further.Yn1iYn . I A"(u ))II + 0. suitably adapted). Proof We shall use the easily proved fact that if A'(u)..u) II 0.n).Yl+ +Yf1>u}. P(. > u}.. Yn .. II ' II denotes the total variation norm between probability measures and ® product measure.. Lemma 5. P (Yj. . A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events).2. +Yn1<u.3 In the following.u) E • I A'(u)) = Bo (n1) ®Bou) . the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1.n) (y1. . then IIP( I A'(u)) Taking A'(u) = {Y. A"(u) _ {K(u)=n} = {Y1+ P(.5 Ilp(u. . I A'(u)) = P(u.5.u) E •) .Bo (ri1) ®B( . FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5.. ...
Z.t..1 and Y„ . .P(Z' E •)II > 0 (here Y.1. By Lemma 5. . Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) .... Z11).r. Y1 +.. ..). in particular his Proposition (2. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .1). (Y..' = y is BM. Zn).6 IIPIu'n ) CHAPTER IX. the discussion just before the statement of Theorem 5.7 O (u.. y > u..P) > y) Corollary 5.. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit. n_1 < u.p) < y). be independent random vectors such that the conditional distribution of Zk given Y.. ... .. {wn(z)} be i. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf. Let {wl(z)}. and that Yk has marginal distribution B0 for k = 1..Bo (n1) ®Bo' 0. Y'.1 P PBo(u) • P(W/(1 .. 2. .. the marginal distribution of Zk is Bo for k < n.2. . Then according to Section 5a. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. .4. Proof Let (Y11. n. in our example Y = (Y1.. copies of {w(z)}.. Zn are independent. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic. then 11P(Z E •) . The first step is to observe that K(u) has a proper limit distribution w. Y") u etc.y(u)T) . Zn) E •) .. HEAVY TAILS ((Z1'.278 Lemma 5 . .6. P(u) since by Theorem 2... .. see Fitzsimmons [144]). For Theorem 5. Similarly (replace u by 0).. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou)...P(Y' E •)II * 0. Z' are arbitrary random vectors. and clearly Zi. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). wk(Zk) has a proper limit distribution as u + oo for k < n. k = 1..d. . . Proof of Theorem 5.. n . +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . the density of Yn is B(y)/[IBBO(u)].+y 1 p"F(Yn > u) P)Pn1 P/(1 ..u has distribution Bout That is.4).i.P) Bo(u) for n = 1.
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). Assume for simplicity that {Vt} regenerates in state 0 . 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. Corollary II. 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 . The heuristic motivation is the usual in the heavytailed area. that fo p(x)1 dx < oo.6. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V.2 Define M.1. nontrivial and we refer to Asmussen [22]. Then 0 (u) Qf "O ^) dy.(3 u u J B(y) dy . claim size distribution B. however./3Ea B(u). cf. u (6. 3.1 Assume that B is subexponential and that p(x) > 00. V. . that MQ becomes large as consequence of one big jump. one expects the level y form which the big jump occurs to be 0(1). Proof of Theorem 6. Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. x > oo. = supo<t<0. and premium rate p(x) at level x of the reserve..2. Theorem 6 . The rigorous proof is. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 . the results only cover the regularly varying case.1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}.e. i.y) . More precisely. and define the cycle as a = inf{t>0: Vt=0.B(u). Asmussen & Teugels [53] studied approximations of i (u. the probability that is exceeds u is then B(u . p(Y) and the result follows. Then P(MT > u) .(u) = P(V > u) = f f (y) dy . T) when T + oo with u fixed.
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). It is also shown in that paper that typically. u Notes and references The results are from Asmussen [22]. by regenerative process theory.q(u) Now just use that p(x) * oo implies q (x) + 0. Then D(u) = f(u)p(u) and. HEAVY TAILS Define D(u) as the steadystate rate of downcrossings of {Vt} of level u and Da (u) as the expected number of downcrossings of level u during a cycle. there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential.P(MT > u) $B(u) Ft µ(1 . D(u) = DQ(u)/µ. Hence f (u)r(u) = D(u) = Do(u) . where also the (easier) case of p(x) having a finite limit is treated .280 CHAPTER IX.q ( u)) 1 . .
where a2 = Var(Z ). The crude Monte Carlo ( CMC) method then amounts to simulating i. vrN(z .2) is an asymptotic 95% confidence interval . topics of direct relevance for the study of ruin probabilities are treated in more depth. ZN. la The crude Monte Carlo method Let Z be some random variable and assume that we want to evaluate z = EZ in a situation where z is not available analytically but Z can be simulated. 281 . .. replicates Zl. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . Fox & Schrage [77].. Hence 1.. We shall be brief concerning general aspects and refer to standard textbooks like Bratley. Ripley [304]. 4Z). estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2.Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . and this is the form in which the result of the simulation experiment is commonly reported.d.z) 4 N(0. z) 2 = Zit NE ii ii According to standard central limit theory .i.96s z f (1.
variance reduction is hardly worthwhile. However. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . there are others which are widely used in other areas and potentially useful also for ruin probabilities. one can argue that unless Var(Z') is considerable smaller than Var(Z). Then replacing the number of replications N by 2N will give the same precision for the CMC method as when simulating N' = N replications of Z'. we then have EZ = EZ = z. Further. and a longer CPU time to produce one replication. The difficulty in the naive choice Z = I(T(u) < oo) is that Z can not be simulated in finite time: no finite segment of {St} can tell whether ruin will ultimately occur or not. The situation is more intricate for the infinite horizon ruin probability 0(u). Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. conditional Monte Carlo and importance sampling.282 CHAPTER X. SIMULATION METHODOLOGY In the setting of ruin probabilities. generated at the same time as Z. Therefore. an added programming effort. v. and many sophisticated ideas have been developed. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). so that Z' is a candidate for a Monte Carlo estimator of z.b(u. Letting Z' = E[Z I Y]. 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. Say that Var(Z') = Var(Z)/2. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). We survey two methods which are used below to study ruin probabilities. This is a classical area of the simulation literature. 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. We mention in particular ( regression adjusted) control variates and common random numbers. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i.
This may also be difficult to assess . Then z Var(LZ) = E(LZ)2 . L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). using the CMC method one generates (Z1. L such that z = EZ = E[LZ]. one would try to choose P to make large values of Z more likely.3) Thus.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . (1. i.z2 = 0. and the problem is to make an efficient choice.. Thus we cannot compute L = Z/z (further. but tentatively..E [Z Z]2 = z2 . . GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. Nevertheless.1. (ZN. . L1). it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. it appears that we have produced an estimator with variance zero.v. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P.3). In order to achieve (1.e. . even if the optimal change of measure is not practical. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. 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. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. To this end.96 sis v^ N 2 1 where srs = N j(LiZi . Thus.zrs. it may often be impossible to describe P in such a way that it is straightforward to simulate from P). the argument cheats because we are simulating since z is not avaliable analytically. However.[E(LZ)] = E Z2 Zz .
An example where this works out nicely is given in Section 3. say 10%. z I.. We then . in terms of the halfwidth of the confidence interval. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem.1.B = iP(AB) = P(BIA).z) 1001. u + oo. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . let z(u) = P(A(u)). the optimal P is the conditional distribution given A.96 2 z2 z increases like z1 as z . assume that the A(u) are rare in the sense that z(u) * 0. However. we may try to make P look as much like P(•IA) as possible. and let Z(u) be a Monte Carlo estimator of z(u).e. To introduce these. Again.96 2Z ( 1 . In ruin probability theory. For each u. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. if z is small.100 . assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. However. Z z V5 In other words . i. it does not help telling whether z is of the magnitude 104. large sample sizes are required. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. as is the case of typical interest. N . Thus.e. Z = I(A) and A is a rare event.. The CMC method leads to a variance of oZ = z(1 .z) which tends to zero as z ^ 0.5 or even much smaller . Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision .284 CHAPTER X.1. and further it is usually not practicable to simulate from P(•IA).96oz /(zV) = 0. This leads to the equation 1.e. I.z) 1 > 00. but if the point estimate z is of the order 105. 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. 10 . say of the order 103 or less. a confidence interval of width 10 4 may look small.0.
. O (u) = z = EZ. P(K = k) = (1 .log z(u) of (1.p)pk.1.d. Generate K as geometric. so that NE (u) may go to infinity. However.4) for any e > 0. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. The algorithm gives a solution to the infinite horizon problem . 2. with common density bo(x) = B(x)/µB and K is geometric with parameter p. logarithmic efficiency is almost as good as bounded relative error. where M = X1 + • • • + XK. let Z +. X2. Let M . and in practice. let Z +. 3.p)pk. it is appealing to combine with some variance reduction method .. F(K = k) = (1 . are i.. where X1.2.1) may be written as V) (u) = P(M > u). Notes and references For surveys on rare events simulation. Thus.X1 + + XK.. Therefore .4).0. . see Asmussen & Rubinstein [45] and Heidelberger [190]. This allows Var(Z(u)) to decrease slightly slower than z(u)2. i. the PollaczeckKhinchine formula III. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. it is not efficient for large u . If M > u. 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. where Z = I(M > u) may be generated as follows: 1. We shall here present an algorithm developed by Asmussen & Binswanger [ 271.log Var(Z(u)) lim inf > 2 u+oo . this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. XK from the density bo(x).(2. the mathematical definition puts certain restrictions on this growth rate. . Otherwise. which gives a logarithmically efficient estimator .e. The term logarithmic comes from the equivalent form . According to the above discussion. 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.i. . Generate X1. but as a CMC method .
. .XK_1). .X1 .. XK... Z(1) (u) has a smaller variance than Zl (x).X(K1)) ..+XK > uIXl. compute Y = u . 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.. and that Bo(y) = 1. A first obvious idea is to use conditional Monte Carlo: write i. we generate only X1. K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X. just note that EZ(1)(u ) 2 > E[Bo (x .. and considering only the remaining ones.X(2).. So.X(2).. Xl > x.. conditional probability. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail.Xl . However. note first that To check the formula for the P(X(n) > x I X(1).286 CHAPTER X. and let Z(2)(u) = _ P (SK B0((u > u I X(l). To see this.S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1)..p)Bo(x).. Theorem IX.. XK1. < X(K) throw away the largest one X(K).XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0..p/(l .L(x)/x`' with a > 0 and L(x) slowly varying. 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.b(u) = P (Xl +•••+XK>u) = EF[Xl + . Thus.. assume in the following that Bo(x) . form the order statistics X(1) < X(2) < . Then (cf..X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) .. ... Z(1)(u) is defined as 0)..SK1)2. The idea of [27] is to avoid this problem by discarding the largest X...1) V)(u) . As a conditional Monte Carlo estimator . .XK1] = EBo(uX1 .. X1 + + XK_ 1 > x when X1 > x. y < 0).....2. For the simulation.. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x). we thus generate K and X1i .
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.. X . BL(dx) = e7sB(dx)/B[y]. 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. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions .3. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . The algorithm is sofar the only efficient one which has been developed for the heavytailed case. X(2). l)) .. BL instead of 0.5). for the purpose of recording Z(u) = erysr(u). 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". . the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals.modulated model P(r+ < oo) and G+ are not explicit ). Thus.S(n_1) I X(1). and we refer to [27]. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1).u is the representation 0(u) = e7sr(u) overshoot (cf.Khinchine formula and importance sampling .1) .. X (. . X(n1)) P(X(TZ) + S(. Asmussen . the algorithm for generating Z = Z(u) is: 1. X(n1)) Bo((x . However .1 is elementary but lengty. B. Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] .. .y.. X(2). that is. For practical purposes. 111. . and define )3L.. in the renewal or Markov. and simulate from FL. Notes and references The proof of Theorem 2.S (n1)) V X (. using 13L. . X(2). 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. Then the algorithm given by { Z (2) (u) } is logarithmically efficient. BL by I3L = /3B[y]. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. . l)) BO(X(n1)) Theorem 2 ._1) > P(X(n) > _ X X(1).
S+U .l3 and U from B. the results of IV. let S. Proof Just note that EZ(u)2 < e . Otherwise.7 tell that P(.Q. let Z F e_'s.r(u) < oo) = 1. u It is tempting to ask whether choosing importance sampling parameters .1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. The proof is given below as a corollary to Theorem 3.. be i. Let FL (dx) = 'For the renewal model. = X1 + . X2. There are various intuitive reasons that this should be a good algorithm. It resolves the infinite horizon problem since FL(. M(u) = inf {n : S„ > u}.d. so that changing the measure to FL is close to the optimal scheme for importance sampling . cf.. In fact: Theorem 3. More precisely. and avoid simulating the known part e7". The algorithm generalizes easily to the renewal model .... + X. with distribution F. to deal with the infinite horizon problem . A > AL as in Chapter V.T.F. P'[y] < oo for some ry > 0.QL. Ti. and the change of measure F r FL corresponds to B > BL.3. b different from .1) (simulated with parameters ^3. Generate T as being exponential with parameter . BL). Let X1.288 2. Xi = U. the discussion at the end of Section 1b. B) is not logarithmically efficient when (/3. SIMULATION METHODOLOGY 3. We formulate this in a slightly more general random walk setting '. If S > u. In detail . .2 The estimator (3.(u)) are asymptotically coincide on {r(u)} < oo. return to 3. and we have: Theorem 3. one must restrict attention to the case 4µB > 1. and assume that µF < 0 and that F[y] = 1.2ryu _ z (u)2/C2. 4. The answer is no. r(u) < oo) and FL (both measures restricted to.. namely ELery£("). Let S .. .i. b) # (/3L. BL could improve the variance of the estimator . The estimator is then M(u) /3eQT' dB Z(u) (Ui) j=1 )3 e $Ti dB where M(u) is the number of claims leading to ruin. Let Sf0 CHAPTER X. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.
. When F # FL.2) dF Theorem 3.3.P = FL. Here ELK.. Since K1. Proof The first statement is proved exactly as Theorem 3 . (3. More generally.. write W(F IF) _ F(XI).. are i. Since ELM(u)/u + 1/ELXi..yu = G. where Kl og (X) (j) 2 ) = log dFL (Xi) . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx).i. where e' = EL Iog dFL (Xi) > 0 by the information inequality. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2. By the chain rule for RadonNikodym derivatives.+KM(u)}..2 > 0. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . .3 The estimator (3. K2. F(XM(u)). 1.d.2'X1 . .. + KM(u))} = exp {ELM(u)(E . For the second.. The importance sampling estimator is then Z( u) = e'rSM( ). is not logarithmically efficient.2ryELXi)} .2ryELXi.yu+elu u +oo etry' 1 > lim up C2e2.. = c'.2) (simulated with distribution F of the X3 has bounded relative error when . EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+. it thus follows that for 0 < e < e'/ELXi.
/3'eQ'YR'( x + y) dy = . with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45].1 is from Lehtonen & Nyrhinen [244].g. U' . B" and generic claim sizes U'. so that one would expect the change of measure F 4 FL to produce close to optimal results. In fact: Proposition 4. we conclude by differentiation that Bo(x)=B' (x)forallx > 0.e. 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 .'(y).T' has a left exponential tail with rate /3' and U" . The queueing literature on related algorithms is extensive . see e. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. T". claim size distributions B'. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L.i. BL) has bounded relative error.290 which completes the proof. we write T = yu. Further discussion is in Lehtonen & Nyrhinen [245].2.'(y) or y > 1/r. Next. U' . then /3' B' = B".T' D U" . all that needs to be shown is that if U' . U". SIMULATION METHODOLOGY u Proof of Theorem 3. yu) is close to zk(u).3. This immediately yields / = 3". . generic interarrival times T' . As in IV.T' = U" . In [13]. CHAPTER X. First by the memoryless distribution of the exponential distribution .T" has a left exponential tail with rate /3'. The easy case is y > 1/k'(y) where O(u. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u.T".T".B'=B".1 If y > 1/ic'('y).3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3".3'eO'x f f P (U" . The results of IV. /3". Then according to Theorem 3. Consider compound Poisson risk process with intensities /3'.4. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. The optimality result Theorem 3. T) with T < oo.4 indicate that we can expect a major difference according to whether y < 1/r.T" > x) J /3"e0 yB (x + y) dy = . the references in Asmussen & Rubinstein [45] and Heidelberger [190].
The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu). T(u) < yu] . Remark 4 . 7y (4. yu)/z.3) and we have: Theorem 4. (4.4.8 has a stronger conclusion than (4.log Var(Z(u)) l im of .5) follows. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.4.(ay).3) (simulated with parameters /gay. and in fact. (4. yu) in the sense that . T(u) < yu] e Hence by (4. Further . Let Qy2 = .4.(u) * 1 (Theorem IV.1). the result follows as in the proof of Theorem 3. We next consider the case y < 1/r. Bay) is logarithmically efficient. one would expect that the change of measure F Pay is in some sense optimal.1.8).log Var(Z(u)) _ . T( u) < yu] e2ryyuEay le 2ay^(u).yu. yu) is of order of magnitude a71.4.to g x ( u ) u u so that (1.1) which is all that is needed here can be showed much easier .yk(ay) determines the order of magnitude of z'(u. and that ryy > ry.1). 3 Theorem IV. Bounding u ELZ(u)2 above by a7u. Proof Since ryy > y.log 4')u) 4 u (Theorem IV.'(7). yu) = eayu Eay Leay^(u)+r(u)K(ay). We recall that ay is defined as the solution of a'(a) = 1/y.2 The estimator (4.O(u. that ryy = ay . .1) so that z(u) = zP(u.yy> 2 . we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. lim inf uoo 27yu .2) Since the definition of ay is equivalent to Eay r(u) .
SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .2).4).b(u. N(0. the object of interest is {Vt} rather than {Rt}.T) = P O<t<T inf Rt < 0 = P(VT > u). Z (u)2 above.Qyu1/2 < T(u) C yu e. In most of the simulation literature (say in queueing applications).ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV. 0 Notes and references The results of the present section are new.3. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).o .292 CHAPTER X. One main example is {Vt} being regenerative (see A.a vt(u).a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . In Asmussen [13].4.7y x(u) > hm inf u+Oo U . we believe that there are examples also in risk theory where (5. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W). > u) = E f I(VV > u) dt 0 (5.1) is used to study Voo by simulating {Rt} (for example.2) . yu . yu .yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r. 0 rather than when u 3 oo.1) may be useful. However.1): then by Proposition A1.uaoo U That lim sup < follows similarly but easier as when estimating En.(av)Eav l e.o. '%(u) = P I info Rt < 0) = P(VV > u). and (5.4. zi(u) = INV.1) (see Proposition IV.3: for many risk processes {Rt }. 5 Regenerative simulation Our starting point is the duality representations in 11.1) where the identity for Vi(u) requires that Vt has a limit in distribution V.u1/2 < r(u) < yu Le ] l = e. there exists a dual process { V t} such that i. related discussion is given in a heavy traffic limit q J.. Hence lira inf log ryyu + vyu 1/2 tc(ay) .yu)/(uyu1/2) . (5.
Then (Z1z1i Z2z2 ) 4 N2(0. + Z2N)) .. let E denote the 2 x 2 covariance matrix of Z('). + Z1N>) . For the ith cycle.. z2) z2/z1 yields Vh = (z2/z2 1/zl). j = 1. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo.. )). Zl the LLN yields Z1 a$' Z(1) +. The method of regenerative simulation. Then Z(1).. Therefore . 2. For details . . provides estimates for F ( V. Z2 a4* z2. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. 02) (5. EZ2'i = z2 = E Thus. z2)) > N(O. Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ').. which we survey below .5. i (^(u) .h (zl. Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. and Z2'>) where Zi'i = w. EZ1'i = z1 = Ew. To derive confidence intervals . Taking h(zl. Z2 .. i.. Thus.. the regenerative estimator z%(u) is consistent.t(u)) 4 N(0. Z(N) are i . Vh = (8h/8z1 8h/ 8z2).d.. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. consider first the case of independent cycles .. Z2 = N (X21' + .3) . oh) for h : R2 ^ R and Ch = VhEVh.E).+Z(N) z 1.. letting J0 'o I (Vt > u) dt . is the cycle length. . > u) (and more general expectations Eg(V. record Zi'i = (Z1'). Z1 = (Zl1i +.
Z of the form Z = ^p(X) where X is a r .g S12 (5.9. Notes and references The literature on regenerative simulation is extensive.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method .C)dx = f w(x) d( f ( x.294 where 01 2 CHAPTER X. 9.96s/v"N. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . Rubinstein [310] and Rubinstein & Melamed [311]. () depending on C.g. The regenerative method is not likely to be efficient for large u but rather a brute force one. v. in some situations it may be the only one resolving the infinite horizon problem . to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . In 111. () dx f Ax) (dl d()f (x' () f ( z. Then z(() = f cp(x) f (x. the expectation z = EZ of a single r.Z) ^Z(=) . However . SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. with distribution depending on a parameter (. see e. () dx so that differentiation yields zS d( fco(x)f(x.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW .v. We here consider simulation algorithms which have the potential of applying to substantially more complex situations.0 .2. Let X have a density f (x. () dx = E[SZ] f(X. consider an extremely simple example . 6 Sensitivity analysis We return to the problem of 111 . asymptotic estimates were derived using the renewal equation for z /i(u). There is potential also for combining with some variance reduction method. Before going into the complications of ruin probabilities .
v. C) f(X.() d( is the score function familiar from statistics . () can be generated as h(U.log U/(. To see this. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. For example . Example 6 . () = log U/(2. ()). /3 is 0. C). For IPA there are. So assume that a r.1 Consider the sensitivity tka(u) w. giving h( (U. 11 /3oeOoT. Thus. Infinitesimal perturbation analysis (IPA) uses sample path derivatives. say W(x) = I(x > xo) and assume that h(U. just take cp as an indicator function . with density f (x. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. cp' (h(U. SZ is an unbiased Monte Carlo estimator of z(. A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. C)). ( where h( (u. /3o is M(u) Oe (3T: < oo) . the Poisson rate /3 in the compound Poisson model. The likelihood ratio up to r(u) for two Poisson processes with rates /3. ()) is 0 w . zc = E [d co(h(U.t. () is increasing in C. Then z(() = Ecp(h(U.t. if f (x. For the SF method. Let M(u) be the number of claims up to the time r(u) of ruin (thus.6. ()) h((U. = E [`d (h(U. for some Co = (o(U). The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. however . () _ (eSx. r(u) = Tl + • • • +TM(u)). this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . ()) d( hc(U. p. one can take h (U. The following example demonstrates how the SF method handles this situation. this is usually unproblematic and involves some application of dominated convergence .1). cp(h(U. () Thus. () = (8/8()h (u. In the setting of ruin probabilities . one. () where U is uniform(0. IPA will estimate zS by 0 which is obviously not correct. Thus . I(r(u) . () is an unbiased Monte Carlo estimator of zS. () = d log f (X.r. () = . Then .r.
since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u.1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. Thus. the risk process should be simulated with parameters . There have been much work on resolving the difficulties associated with IPA pointed out above. for different models and for the sensitivities w. We then arrive at the estimator ZZ(u) = (M(u) . 4) that V5. However. In the setting of ruin probabilities. .9 . a relevant reference is VazquezAbad [374]. To resolve the infinite horizon problem . 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial).3 (u) (to generate Zp (u). whereas for the SF method we refer to Rubinstein & Shapiro [312].3L.T(u)) I(T(u) < co) ] . Example 6. SIMULATION METHODOLOGY Taking expectation. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . differentiating w. change the measure to FL as when simulating tp(u).3 (u) is of the order of magnitude ue7u. different parameters.296 CHAPTER X.r. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . BL).t.0(1) so that in fact the estimator Zf(u) has bounded relative error. in part for different measures of risk than ruin probabilities.r. We recall (Proposition 111. j3 and letting flo = 0. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu .t.T(u)) e7uerVu) for ?P.
defined as Ro = u (with u E {0... The twobarrier ruin problem The twobarrier ruin probability 0. X2. in the Bernoulli random walk example below. where X1. and {1.1}valued .. are i. wherel T(u) = inf {t > 0 : Rt < 0} .+• • •+X..(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. as e.. Besides its intrinsic interest . 'Note that in the definition of r(u ) differs from the rest of the book where we use r(u) = inf {t > 0 : Rt < 0} ( two sharp inequalities ). Consider first a Bernoulli random walk. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo. R„ = u+X.g.. 297 .P(•r(u. a) = r(u)).1. a) = r(u) A T+(a). T+(a) = inf It > 0 : Rt > al.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion. with P(Xk = 1) = 9. T(u. That is. in most cases . }). Y'a(U) = P(T (u) = r+(a)) = 1 ..(u) = 0 ) = 0) or it is trivial to translate from one setup to the other.d. either this makes no difference (P(R. .i....
MISCELLANEOUS TOPICS Proposition 1. 7/la(a . The Lundberg equation becomes 1=F[ry]=(19)+9z.(u) I\ e = 1 oa ' ()i a = u. and the other more advanced but applicable also in some other settings. (1. u + 1.1) o If 0 = 1/ 2.(1B)u oJ 0. and in view of the discrete nature of a Bernoulli random walk we write z = e7. Wald's exponential martingale is defined as in 11.4) by ea(u+Xl+... = (1 .a) Y.e.1..2) Oa(a ..0)/0.o)'t/1a(a . = z°Va(u) + za(1  . then 'Oa(u) _ au a We give two proofs . where a is any number such that Ee°X = F[a] <oo.a) = 0) + zap ( R.o)T/la (1) + 8z/'u(3).1) is solution. i. z and the solution is z = (1 .1) = (19)4/'0(a3)+9ba(a1). In a general random walk setting ..o» = z°P (RT ( u. By optional stopping.+Xn) F[ a]n n=0.r(u. We choose a = ry where ry is the Lundberg exponent. tba(2) _ (1 . one elementary but difficult to generalize to other models.. u Proof 2. The martingale is then {zuzXl+•••+X„ } = {zR° }. zu = EzRO = EzRT(u.. the solution of F[. C1_0\a..(4.1 For a Bernoulli random walk with 0 0 1/2. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). and insertion shows that ( 1.a(u)).2). Proof 1.y] = 1.298 CHAPTER XI..
1). BROWNIAN MOTION.} 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)).. then Proof Since 'Oa (U)  au a Eea(R°. thenz1 (u)=1.zu)/(za . However. (1. and solving for 9/la(u) yields Z/)a(u) = (e 76 . {R.2) is trivial (z = 1). then Vi(u) = 1. Corollary 1. u Proposition 1. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za .u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. pa( u) _ u Corollary 1.0a(u)). Then for p 0 0. 1h (u) = a el u \1 If 9 < 1/ 2.1 yields 't/la(u) = (a . If 9 = 1/2.1). {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1.1 If p = 0.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . Proof Let a+ oo in (1.• a2µa e2µu .1) for p # 0.5) . i1(u) = e211 . If p = 0.e7u)/(e7° . RANDOM WALK. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .u)/u. If p<0.4 For a Brownian motion with drift u > 0. .1. (1.ba(u) = e2µa .2 For a Bernoulli random walk with 9 > 1/2. Applying optional stopping to the exponential martingale {e7R.
5a).e7a (u) = 6 /0 .7) .4). It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1.0(a) 0 < u < a. we obtain 'Oa a7u .300 Proof Let a * oo in (1. passing to even more general cases the method quickly becomes unfeasible (see. letting a * oo yields the standard expression pe7u for . For most standard risk processes . Here the undershoot under 0 is exponential with rate 5. However.a) < 0) + e7°P (R(u. Ic 5ry 'pa(u) Using y = 6 . Here is one more case where this is feasible: Example 1. CHAPTER XI.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. and hence e7u = Ee7Ro E [e7R(.5 Consider the compound Poisson model with exponential claims (with rate. 7/'(u) = 1). implying R(u. (u) _ O(u) .0 (u) (where u p =.+^a(u))^(a) If 7k(a) < 1. 5). 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. however. 0.7/la(u)). 1 .a) = a on {r (u. say.3. VIII. valid if p < 1 (otherwise .7).616).a) = r+ (a)} and similarly for the boundary 0. this immediately yields (1. (1.vi(a) Proof By the upwards skipfree property.a ) < 0) + e 7aF ( R (u. and thus one encounters the problem of controlling the undershoot under level 0. ..a) = a) = 5 y = P (R (u. a) I R(u a ) < 0] P (R(u .a) = a ) + e ' ° ( 1 . the paths are upwards skipfree but not downwards.e7a Again . 7O(u) = 7/la(u) + (1 .
MT > u) = P (ST > u) + P (ST > u.2 .1.11 ) is the same as (1.10) Pµ (T(u) < T) !.. T ) = P(T(u) < T ) = 241.1a for computing ruin probabilities for a twostep premium function.f. Here {St } is Brownian motion with drift 0 (starting from 0).. Then the density and c. For µ # 0. (i). TWO BARRIERS 301 Note thas this argument has already been used in VII. P(MT > u) = P(ST > u) + P(ST < u. T(u) E dT.d. 10) follows then by straightforward differentiation. 0(u. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T .Rt}. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities. MT > U) = P(ST > u) + P(ST > u) (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 .ST<u) = P(MT>u.)_ _( u)µ2 /2. ( 1. BROWNIAN MOTION. we have ili(u. Hence P(MT>u. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1.. the density dPµ / dP0 of St is eµsttµ2/2.µ so that {St} is Brownian motion with drift µ . (1.µ T I + e2µ"4) ( .T) P(MT > u) where MT = maxo<t<T St. and (1 .8 ).4) I = .. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu . (1.8) Proof In terms of the claim surplus process { St} = {u . Corollary 1.8 Let {Rt} be Brownian motion with drift . = 1 ..7 For Brownian motion with drift 0. RANDOM WALK.r(u). and hence Pµ('r(u) E dT) = Eo [e µsr(.9) = 2P(ST > u). + µ2T) } .11) VIT ) Proof For p = 0.ST>U).µ%T (1.
T2. 226).10 Consider a diffusion process {Rt} on [0. and in a similar spirit as in VII. u Small modifications also apply to Bernoulli random walks: Proposition 1. Thus. 0 0 (1.. If this assumption fails.9).g.h. and (1. T are integervalued and nonnegative. (1. Proof The argument leading to ( 1. Vi(u. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. S(oo) = f c s(y)dy.T)dx. Here {2T( (v}TT)/2) v=T. e. Breiman [78] or Karlin & Taylor [222] p. The same argument as used for Corollary 1.12) is the same as ( 1.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion.8 also applies to the case 9 54 1/2. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. MISCELLANEOUS TOPICS which is the same as (1. is finite for all x > 0.12) P(ST = v) = 0 otherwise.s. see e.9) goes through unchanged.T (1.302 CHAPTER XI. Theorem 1.13) with 0 as lower limit of integration. is (1. the behaviour at the boundary 0 is more complicated and it may happen. oo). S(x) = f x s(y)dy.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). oo) with drift µ(x) and variance a2 (x) at x. Let s(y) = ef0 ry(.11) then follows by checking that the derivative of the r. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 .. that 0(u). as defined above as the probability of actually hitting 0. as defined in (1.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). We finally consider a general diffusion {Rt} on [0. whenever u. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.. but we omit the details..g.10).9 For Bernoulli random walk with 9 = 1/2.T+2.10) and that the value at 0 is 0.
b('u) = Eu &0.(u) < 1 for all u > 0 and ^ S^ Conversely.b = 0 implies that VQ b/s is constant. (1 . 15) i.S(u) (1. Further references on twobarrier ruin problems include Dickson & Gray [116]. see in particular pp.S(u)/S(a). Assume further that S (x) as defined in (1.b('u) = Eu .b = a+/3S. For generalizations of Proposition 1.b (Rdt) = Oa. 0 in (1. and we get Wo.10. b = 0. Letting b J.13) is finite for all x > 0. if (1. E„ q(Rdt) = q(u)+Lq(u)dt.0(u) = 1 for all u > 0. Letting a T oo and considering the cases S(oo) = oo.16).b(u)dt. then. A good introduction to diffusions is in Karlin & Taylor [222].ba. .b(a) = 0 then yield the result.b(u) be the probability that {Rt} hits b before a starting from u. so that Y)n.b(u) + L. Notes and references All material of the present section is standard.17) Hence L. (1. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion.b(b) = 1.16) S(a) . RANDOM WALK. A classical reference for further aspects of Bernoulli random walks is Feller [142]. In view of (1. then 0 < 2l.11 Let 0 < b < u < a and let t&0. 191195 for material related to Theorem 1. Wa.S(b) Proof Recall that under mild conditions on q. 0 Proof of Theorem 1. 1'.1. If (1. S(oo) < oo separately u completes the proof. see Asmussen & Perry [42].14) fails. i.6 to Markovmodulated models .e LVa. the function S(x) is .e. The obvious boundary conditions '0a. O.b(Rdt).ba. [117]. If b < u < a.14) S(oo) < 00. BROWNIAN MOTION. Then YIa.10.16) yields 4b (u) = 1 . Lemma 1. TWO BARRIERS 303 for x > 0. we can ignore the possibility of ruin or hitting the upper barrier a before dt.b(u) = S(a) . elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] . Using s'/ s = 2p/a2.
MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0..)AT .17). information on ruin probabilities can be obtained . one works instead with a lower limit 5 > 0 of integration in (1.2. but by duality. 7y = ay .2) C_e7u < t(u) < C+e _7u.o•K(a) = Ee .3) < e 7yu. 111 . (2.t&(u.304 CHAPTER XI. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.5): _ z/'(u) < e 7u. Remark 11.1.ytc (ay).9 ) and optional stopping applied to the stopping time r(u) A T.aRo . y > . and here are alternative martingale proofs of the rest . 1 y < k (y).5) A martingale proof of (2.4.4. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.3. defined by the density 1/va(u)s(u) showing up in (1. correponding to piecewise linear paths or . much of the literature dels with the pure drift case. yielding eau = Ee. The emphasis is often on stationary distributions . Another basic quantity is the speed measure M . They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu.6. IV. yu) '+/1(u) .4) I. yu) where W (ay) = y. with the drift and the variance depending on an underlying Markov process . equivalently. (2. Lo is a martingale (cf.1 ) was given already in II. variance 0.13)). Markovmodulated Brownian models . which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications.(a) (2. is currently an extremely active area of research.(.(T(u)AT) r.aR. (2. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u. (2.1) (2.6) . Lo I.(7) .5.
(u.. Equivalently.)r(u)r.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e.6) below by 1 E Le7Rr(.B(r))/B(r). dr) denote the conditional distribution of (T(u). RT(u)_) given r(u) < oo. A claim leading to ruin at time t has c.yu) Y Similarly for (2. For (2. Rt has distribution B(r + dy)/B(r). Let H(dt.6) with = 'y that eyu .6).4): We take a = ay in (2. eyuk (ay) = e7yu e > eyu"(ay ) ij(u. we have tc(ay) > 0 and we can bound (2.1.d. u Proof of (2.7R.f.3). Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt. y > r. (2. and the proof of the lower inequality is similar. . Proof of ( 2.yuk (ay)(u&(u.( u ) I T(U) < 00] . 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). so that i/1(uL yu) < eayu .3).2): As noted in Proposition II.2.yu))• b(u. we have ic(ay ) < 0 and use the lower bound E [e7Rr („). dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.T)  V. when Rt_ = r.4). dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0). it follows easily from (2. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II. (B(y) .E [e.1 . dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows.1.
2). (3.1 We will go into some more detail concerning (3.gn if nioo lim 109 fn = 1 log gn (later in this section. cle . v2 later. e. and that a considerable body of theory has been developed. Example 3. large deviations results have usually a weaker form.1) does not capture the \ in (3. its generality.1) is an example of sharp asymptotics : . Accordingly.1).. + X. in being capable of treating many models beyond simple random walks which are not easily treated by other models .(B) = log EeOX 1 is defined for sufficiently many 0.^ e nn 1 > x n 0o 2xn (3. gn with fn + 0 . .?n typically only give the dominant term in an asymptotic expression .the correct sharp asymptotics might as well have +. The classical result in the area is Cramer's theorem. 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. logarithmic asymptotics .2) can be rewritten as F (Sn/n > x) 1g a'fin.means (as at other places in the book) that the ratio is one in the limit (here n * oo).1) where we return to the values of 0../n E I) for intervals I C R. Thus . However . ri. we will write fn 1. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . such that the cumulant generating function r. For example.306 CHAPTER XI. the parameter will be u rather than n). og For sequences fn.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. gn 4 0. (3. The limit result (3. large deviations results been. and gave sharp asymptotics for probabilities of the form P (S.nn or C2e. not quite so much in insurance risk.. if x > EX1. 1) but only the dominant exponential term . nroo n n /// Note in particular that (3. however . Cramer considered a random walk Sn = X1 + .3na with a < 1.g. . The last decades have seen a boom in the area and a considerable body of applications in queueing theory. then P C S. Thus. which in the setting of (3. The advantage of the large deviations approach is..
4) immediately yields (3.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn .2). i. which is a saddlepoint equation . P with mean nx and variance no.96o /] > 0. V > 0 e.4) n Next.(0)) e 307 (other names are the entropy. since Sn is asymptotically normal w. Since P nn > x) = E {e_8 ' ( 9).1). exponential change of measure is a key tool in large deviations methods.e. nx < Sn < nx + 1. we get P(Sn/n > x) E [e9nx +nK(9)9" '.r. and hence for large n P(Sn/n > x) > E [e. More precisely.the mean rc'(0) of the distribution of X1 exponentially tilted with 0. rc*(x) = sup(Ox . the LegendreFenchel transform or just the Legendre transform or the large deviations rate function).3) is put equal to x. Most often.rc(0) where 0 = 0(x) is the solution of x = rc'(0).(e)i XI E dx].9S„+n' ( 9).425. 2 where o2 = o2(x) = rc"(0).4 enn +1. of P(X1 E dx) = E[e9X1K.t. In fact. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3. we have P(nx < Sn < nx + 1.sseo f which in conjunction with (3. Define .3. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox . if we replace Sn by nx + o / V where V is N(0.r. S rtn > x 1. (3. LARGE DEVIATIONS Define rc* as the convex conjugate of rc.q = rc* (x).960/) * 0.
dxn) where Fn is the distribution of (X1i .e < 8 < y + e... n Icn(0) exists and is finite for ry . .. (ii) lim supn. Sanov's theorem which give rare events asymptotics for empirical distributions. Xn) and sn = x1 + • • • + xn (note that the r. Xn given by Fn(dxl. .o log Ee9Sn /n. 260 for details.1).h. Further main results in large deviations theory are the GartnerEllis theorem. commonly denoted as is the saddlepoint approximation. however. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .dxn) = 05nKn(7)Fn(dx1. Ee9X n < oo for e < 0 < e..v. and write Sn = X1 + • • • + Xn.. (iv) tc(ry) = 0 and r. integrates to 1 by the definition of Icn). We shall need: Lemma 3 . see Jensen u [215] or [APQ] p. In the application of large deviations to ruin probabilities. 1) and no such that Sn . We further write µ = tc'(ry). be a sequence of r.. that is.'(u) )Ng a"u. Then i/..3 For each i > 0. The substitution by V needs. Assume that there exists 'y.s.. For the proof.e < 8 < y + e.. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn.3. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. Pn Sn1 . there exists z E (0.'s. which is of similar spirit as the dicussion in VII.2 (GLYNN & WHITT [163]) Let X1. Mogulskii's theorem which gives path asymptotics. to be made rigorous.. .308 CHAPTER XI.. and the WentzellFreidlin theory of slow Markov walks. .. X2... 1]. . MISCELLANEOUS TOPICS which is the same as (3. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 .. we introduce a change of measure for X1.../^ >7 < zn n for n n0. (iii) #c (8) = limn.p > 7 < zn. is differentiable at ry with 0 < K'(y) < 00. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo).
is of order .Bµ . 0. S..2. h. in particular the r. ( U) P(S. the r .ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . h. Then V. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.n m µ 1 + rl .µ?7 .y) . can be chosen strictly negative by taking 9 small enough.s.ne(p+ 17). for Sn. 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.YS.3. mµ Sm > u] km e7Sm+n.r (7) n = e.077 n^oo n and by Taylor expansion and (iv ).77) follows by symmetry (note that the argument did not use µ > 0). log zl'(u)/u > 'y. This proves the existence of z < 1 and no such that Pn (Sn/n > µ.m(7).71 < e and jq9j < e. S. it is easy to see that the r. u Proof of Theorem 3.. We first show that lim inf„_. This establishes the first claim of the lemma .h.s. Clearly.+r7) < zn for n > no.. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1.s. The rest of the argument is as before..n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.91) + o(O ) as 0 J.]1/q = e. 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. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7). Since I EeqOX „ ] 1/q is bounded for large n by (ii).2. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3.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 +.n > u ) = [ Em [em Em 1e.W.> .. > 1 +17] m(7).+r. The corresponding claim for Pn(Sn/n < µ .
I > IL exp `S.. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).. I2 = F(T(u) = n).. n=1 .YS +^c CHAPTER XI. and since Ic. (iv) and Lemma 3.(Y)..(•) goes to 1 by Lemma 3. Sn > u] < eYu+Kn(7)pn(Sn > u) (3. MISCELLANEOUS TOPICS (7). For lim supu. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. logO(u)/u > ry. Pn \ > la+ 8 I < zn (3. we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J..log z) /2 and Sn Fn\ n >lb+S) <Zn.3.n(ry)/u 4 0andm/u* (1 + r7)/µ.+wn(7). this is possible by (iii).n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. P(T(u) = n) < P(Sn > u) = En [e7S.310 ]Em I e.7) so that n(b) I1 < e'Yu E en... we write P(T(u) = n) = Il + I2 + I3 + I4 'i/I(u) _ E00 n=1 where n(b) Lu(10/µJ Ii = 1: F(T(u) = n). 3. 0 yields liminfu __.0 log i'(u )/u < 'y. Obviously.I < µl1 1+77 I M 1_ 1+277 S.6) for some z < 1 and all n > n(E).
10) 00 I4 < E F(Sn_1 < u. eryu en logz/2p n nt n. C 26u `p / +1 I e6u(1+6)/µ (3. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. 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 . S.3. µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 .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).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. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0. u .' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. Sn1 C U. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .
(u) = I1+I2+I3+I4'^ ery( u). IV.u(1+b)/rc'(7)).3ui where .312 CHAPTER XI. Corollary 3. Then for n large. For 14. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 . For 12. we need to redefine n(b) as L. 4 there is an aj > 0 and a cj < oo such that Ij < c3e.Q is so small that w = 1 . 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. we replace the bound P(Sn > u ) < 1 used in (3. ryue«iu . say n n1. cf. 13 = P(T (u) E (u(1 b)l^ (7).. 2.('+'Y).z 1/z For I1.2.b)/i(7).4/3rc'(y) > 0.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7). For I.xl/2 to give the desired conclusion. I2.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . it suffices to show that for j = 1. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. the typical time is u/rc'(7) just as for the compound Poisson model.4 Under the assumptions of Theorem 3. Letting c11 = maxn<n. this is straightforward since the last inequality in (3. e'.8) by P(S.4..7' a"ju. u . the last steps of (3. u(1 + b)/i(7)) Proof Since V. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. (7 + a) < 2arc'(7).
. and in fact.3(s) at time s. 09(9).2 shows that the discrete time structure is used in an essential way. Assuming that the further regularity conditions can be verified.LARGE DEVIATIONS 313 Example 3 .e. we shall give two continuous time examples and tacitly assume that this can be done. V(s) with m. and we conclude that Theorem 3 .1. i. Thus the total reward in the interval [0. t] is Rt = E V (Un) n: o„ <t .. whether P ( sup St > u ltg a ^" 0<t<oo // (3. 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. r. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * . An event occuring at time s is rewarded by a r..'(y) > 0. the key condition similar to (iii). Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .v. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0.12) k=0. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0. but nevertheless.g.1.f....14) is needed in both examples . 11 Inspection of the proof of Theorem 3.(O) = 9µ+02 for all 9 E R. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process. Obviously many of the most interesting examples have a continuous time scale. If {St}t> 0 is the claims surplus process..2 then immediately yields the estimate log F( sup Skh > u) a7u (3. Xk+l) k=1 00 naoo n provided the sum converges absolutely. 2 is in force with y = 2p/wz. criteria are given in Duffield & O'Connell [124].13) One would expect this to hold in considerable generality.3.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. Theorem 3.. To verify these in concrete examples may well present considerable difficulties.
It.v. n: o. assuming a continuous premium inflow at unit rate.. one would take p(t) = (1 + rt)At/ t. At = . An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . the Cramer. <t which is a shotnoise process. MISCELLANEOUS TOPICS are the event times.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . it contributes to St by the amount Un(t .. Then logEeOR° = J0 /3(s)(^8(9) .9t = /3 J t (Ee8U° i8l . then the payments from the company in [on.14). where Ft = a(A8 : 0 < s < t). 7 Given the safety loading 77.9t. we have rct (9)/t 4 ic (9). a differential equation in t). O'n +S] is a r .1) ds rt (3..Q„) . We let ic (9) = 3(EeWU° . Kt (0) t (Ee9U"it8i J0 . the best estimator of /3µB based upon Ft. Since the remaining conditions of Theorem 3. . Thus.. 0 and since EeOUn(8) + Ee°U^ as s * oo. derive .Lundberg model has the larger ruin probability. e. If the nth claim arrives at time Qn = s.s). i.0 and assume there are y. Most obviously. we have S. Thus by (3. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot . this is not realistic . More precisely.'`1 U.t. we conclude that Cu) log e7 u (cf. Of course . We further assume that the processes {U1(s)}8>0 are i. leading to St = At(1+77) Joo t S8 ds. Thus. (3. Un represents the total payment for the nth claim).g.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un.1) ds . Example 3. Un(s).14) (to see this . is At .314 where the an CHAPTER XI. e > 0 such that ic('y) = 0 and that r. if the nth claim arrives at time a. Of course.d.1) . (9) < oo for 9 < 'y + C. the above discussion of discrete skeletons).15) . = U„ ( t .1) ds . the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. but that a claim is not settled immediately. nondecreasing and with finite limits Un as s T oo ( thus.2 are trivial to verify. 0 Example 3 .
LARGE DEVIATIONS With the Qi the arrival times. the Vi = .2 hold. again the above discussion of discrete skeletons) where y solves ic('y) = 0 It is interesting to compare the adjustment coefficient y with the one y* of the CramerLundberg model.16) i=1 o i=1 Let ict (a) = log Eeast .d. we conclude that t.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 .d.18) Thus (iii) of Theorem 3. (3. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 . one has y > y' (3.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.log Oi are i. Indeed.(1 + 17)0µB = 0.3. we have Nt t N. Ui Nt / t 01i 315 St = Ui . 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 . and since the remaining conditions are trivial to verify.19) with equality if and only if U is degenerate.i. equivalently. uniform (0. standard exponential . (3.1) or . typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model . Thus. the solution of /3(Eelu .17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.(1 + r7) log t (3./3.(1 +i) f > i= 1 s ds = E Ui 1 .i.e.b(u) IN a'Yu (cf.1) . To see this . i.20) (3. It then follows from (3.
and since tc(s). the proof of (3. [257] and Nyrhinen [275].1 . . the function k(x) = e7*x . 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. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. Therefore e7'U _ k(U) E [1+(1+77)y*U] . assuming that the U.19). and k(x) < 0. MISCELLANEOUS TOPICS Next. This implies n(y*) < 0.xo.2 expressing the finite horizon ruin probabilities in terms of the distribution of A. we are interested in estimating P(A > x) for large x.20) is due to Tatyana Turova.316 CHAPTER XI. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. using that Ek(U) = 0 because of (3. we then take t = 1 so that p.i. In addition to Glynn & Whitt [163]. MartinL6f [256]. rc*' (0 ) < 0. though we do not always spell this out.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. For Example 3. much of the analysis carries over to more general cases. y = y* can only occur if U .. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. Further.(1 + ri)y*x is convex with k(oo) = 00. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. Further. this in turn yields y > y*. the study is motivated from the formulas in IV. Further applications of large deviations idea in risk theory occur in Djehiche [122]. For notational simplicity.7. at time t.2. k(0) = 0. see also Nyrhinen [275] for Theorem 3. k'(0) < 0. a* (s) are convex with tc'(0) < 0 . see Nyrhinen [275] and Asmussen [25]. [245]. say one year. Lehtonen & Nyrhinen [244].d. = P(N = n) = e(3an However. x > x0. 0 < x < x0. are i. In particular. with common distribution B and independent of Nt. The main example is Nt being Poisson with rate fit.
ic(9) = . Proposition 4.1).1) where )30 = .3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution.x)//3B"[9] is standard normal. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. B"[s] = oo. no(a) = logE9e'A = rc(a + 9) . we define the saddlepoint 9 = 9(x) by EBA = x.1 Assume that lim8T8. The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). B"' [s] lim (B". Vare(A) = s. Then as x * oo.1). A E dx] . Hence P(A > x) = E e [e9A+ ic(9). only with 0 replaced by a9 and B by B9. 818' where s' = sup{s : B[s] < oo}. A > x)] = eex+K( e)E9 [e . (4."(0) = . The analysis largely follows Example 3. For a given x.2) implies that the limiting Pedistribution of (A .1.3B"[9].[s])3/2 = 0. 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 .9(Ax).4. In particular. i.3e(bo[a] . K'(0) _ ic'(9) = x. Then Ee"A = e'(") where x(a) _ 0(B[a] .e.
Furthermore 00 b(x)Sdx < oo for some ( E (1.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. [138]. 1 . Thus .EN B(x).v.318 CHAPTER XI. i. bounded with b(x) . or. b is logconcave. Notes and references Proposition 4. In fact.2i and that (A . B covers distributions with finite support or with a density not too far from ax° with a > 1. and (4.1 goes all the way back to Esscher [141].x') where x' = sup {x : b(x) > 0}. Jensen [215] and references therein. leading to P(A > x) :. Remark 4 . For example. just the same dominated convergence argument as in the proof of Theorem 2. A covers the exponential distribution and phasetype distributions. then P(A > x) . b(x) = q(x)eh(z). Var(A) _ ^3p. under the Poisson assumption (4. large x.Q{AB (4. For details.(D X . 2). For a rigorous proof. the distribution of A is approximately normal .2) is often referred to as the Esscher approximation. b is gammalike. In particular.1). it is quite questionable to use (4. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. some regularity of the density b(x) of B is required. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. The present proof is somewhat heuristical in the CLT steps.l3pB.2 If B is subexponential and EzN < oo for some z > 1. more generally.e. either of the following is sufficient: A. see Embrechts et al.ycix °ie6x B. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward.4) . it holds that EA = .1 yields: Proposition 4. 4b The NP approximation In many cases . For example. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. 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 .3) and related results u for the case of main interest .3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion.
s. are small. In concrete examples .. are the cumulants .y2)^P(y)• 319 Note as a further warning that the r.. If the distribution of Y is close to N(0. .5 (y3 .e.i 3 K3 } Pt^ exp . in particular. the density of Y is 1 °° _ eiuy f(u) du 2x _. K2 = Var (Y).i 6 r 1 3 so that we should take b = ic3/6 in (4. and from this (4. .. the standard normal distribution. (4.EY)3. zl_e be the 1 .3!).99. and so as a first approximation we obtain a1_E = EA + yle Var(A) .5) follows by integration.equantile in the distribution of Y.f..2 ^ \1 . K4 . Var(Y) = 1 as above .EA)/ Var(A) and let yl_E.4. so that 1(u) 3 exp { . K3 = E(Y .. If this holds . Heuristically.2K3 + 4i 64 + . one needs to show that 163.5).h. which is often denoted VaR (the Value at Risk). THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. one expects the u3 term to dominate the terms of order u4. the NP (normal power) approximation deals with the quantile al_E. the CLT for Y = Y6 is usually derived via expanding the ch. defined as the the solution of P(A < yle) = 1 . Let Y = (A .5) may be negative and is not necessarily an increasing function of y for jyj large.6) . resp.. Thus if EY = 0..l = EY. u5.: EA + zl_E Var(A) .2X2 . (4. Rather than with the tail probabilities F(A > x). A particular case is a. of (4. however. then P(Y < y) 4(y) .2 2 .s.c2i. where Kl .5) is obtained by noting that by Fourier inversion.. ylE should be close to zl_E (cf.6(1 . ..1).3& (y). as u2 u3 u4 9(u) = Ee'uY = exp {iuci . Remark 4. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .
.E = z1E + S(zi_E .3ni /3 .zlE)W(zlE) 1 .E ..1)! n ^eQ .E )Azl E) 4(z1E) + ( ylE .(y) terms dominate the S(1 .320 CHAPTER XI. as required .E(/3PB^1 )1^2 + s(z1E . b such that EN 1 U%. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .zi. that [101] distinguishes between the NP and Edgeworth approximations.k = /3µB^1 / (. n = 1. however.zl E)^o(zl E) .6pBki) d/2.y2)cp( y) term..6 (1 .yi.5(1 .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.S(1 . Note. Using Y = (A .7) as 1 (3) a1E = Qµa +z1 . MISCELLANEOUS TOPICS A correction term may be computed from (4.. K5 . Another main reference is Daykin et at. let pn Pn = (a+ = P(N = n). this yields the NP approximation 6(Z1 _E .zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 . This leads to t( yl E) ..1) E (A .1).. 21 . [101].zl E)V(zl_E) . For example. In particular .E)A1 l E) 1 E 4)(yl E) ^' .E + (yl.. the kth cumulant of A is /3PBk' and so s.1)EY3.. We can rewrite (4. k3 is small for large /3 but dominates 1c4.EA ) / Var(A).5) by noting that the 4. this holds with a = 0. 4c Panjer 's recursion Consider A = constants a.1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. and assume that there exist n ) Pn_i .S(1 .
. ..1. 2..13) but only O(j2) for Proposition 4. j1 g. 1. n = k=n1 9k(n1 )9j k • (4. (4. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4. j = 1.12) 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 . fj = E (a+ b k =1 )9kfi_k .. E[a +bU=I >Ui =j l i=1 J (4..9)..4 is that the algorithm is much faster than the naive method. the value of (4.14) is independent of i = 1. Hence by (4.4.. 2. if go = 0. . . Then fo = >20 9onpn and fi = 1 E In particular.11) Remark 4.4. By symmetry. n. then j (a + b!) 1ag k_1 3 gkfj. . which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4.4 Assume that B is concentrated on {0. The expression for fo is obvious. . . u Proof of Proposition 4.4..k . 2. fj = P(A = j).12). Since the sum over i is na + b.. j = 0. (4. and calculating the gj*n recursively by 9*1 = 9j.} and write gj = 2 ..10) f o = po. j = 1. . the complexity (number of arithmetic operations required) is O(j3) for (4. (4.13) Namely.12) where g*n is the nth convolution power of g...14) is therefore a + b/n.. .5 The crux of Proposition 4. (4.
322
00 J
CHAPTER XI. MISCELLANEOUS TOPICS
EE (a + bk I gkg3 _ k lieni n=ik=0 (a+bk l gkE g j'`kpn = E (a+b!)9kfi_k n=0 k=0 k=0 ^I 1 E(a+b. agofj+ k Jgkfjk, k=i /
and and (4.9) follows . (4.11) is a trivial special case.
u
If the distribution B of the Ui is nonlattice , it is natural to use a discrete approximation . To this end, let U(;+, U(h) be U; rounded upwards, resp. downwards , to the nearest multiple of h and let A}h) = EN U. An obvious modification of Proposition 4.4 applies to evaluate the distribution F(h) of A(h) letting f( ) = P(A() = jh) and
g(h) gkh+
= P (U(h2 = kh) = B((k + 1)h)  B(kh ), k = 0, 1, 2, ... , = P (U4;+ = kh) = B(kh)  B (( k  1)h) = gk  l,, k = 1, 2, ... .
Then the error on the tail probabilities (which can be taken arbitrarily small by choosing h small enough ) can be evaluated by
00 00
< P(A > x ) f (h) j=Lx/hl j=Lx/hl
Further examples ( and in fact the only ones , cf. Sundt & Jewell [355]) where (4.9) holds are the binomial distribution and the negative binomial (in particular, geometric ) distribution . The geometric case is of particular importance because of the following result which immediately follows from by combining Proposition 4.4 and the PollaczeckKhinchine representation: Corollary 4.6 Consider a compound Poisson risk process with Poisson rate 0 and claim size distribution B. Then for any h > 0, the ruin probability zb(u) satisfies 00 00
f^,h) Cu) < E ff,+, j=Lu/hJ j=Lu/hJ (4.15)
f! h)
5. PRINCIPLES FOR PREMIUM CALCULATION
where f^ +, f^ h) are given by the recursions
(h) 3 (h) (h)
323
fj,+ = P 9k fjk,+ ' I = 17 2, .. .
k=1 3 (h)
(h)
=
P
(h)
f9,  (h) gk,fAk, e 1  ago, k=1
j = 1+2,
starting from fo + = 1  p, f(h) = (1  p)/(1  pgoh) and using 07
g(kh) 1 (k+1)h
=
Bo((k + 1 ) h)  Bo(kh ) =  f
AB
kh
B(x) dx, k = 0, 1, 2, ... , k = 1,2 .....
gkh+
Bo(kh )  Bo((k  1 ) h) = 9kh)1 ,
Notes and references The literature on recursive algorithms related to Panjer's recursion is extensive, see e.g. Dickson [115] and references therein.
5 Principles for premium calculation
The standard setting for discussing premium calculation in the actuarial literature does not involve stochastic processes, but only a single risk X > 0. By this we mean that X is a r.v. representing the random payment to be made (possibly 0). A premium rule is then a [0, oo)valued function H of the distribution of X, often written H(X), such that H(X) is the premium to be paid, i.e. the amount for which the company is willing to insure the given risk. The standard premium rules discussed in the literature (not necessarily the same which are used in practice!) are the following: The net premium principle H(X) = EX (also called the equivalence principle). As follows from the fluctuation theory of r.v.'s with mean, this principle will lead to ruin if many independent risks are insured. This motivates the next principle, The expected value principle H(X) = (1 + 77)EX where 77 is a specified safety loading. For 77 = 0, we are back to the net premium principle. A criticism of the expected value principle is that it does not take into account the variability of X which leads to The variance principle H(X) = EX+77Var(X). A modification (motivated from EX and Var(X) not having the same dimension) is
324
CHAPTER XI. MISCELLANEOUS TOPICS
Var(X).
The standard deviation principle H(X) = EX +rl
The principle of zero utility. Here v(x) is a given utility function, assumed to be concave and increasing with (w.lo.g) v(O) = 0; v(x) represents the utility of a capital of size x . The zero utility principle then means v(0) = Ev (H(X)  X); (5.1)
a generalization v(u) = Ev (u + H(X)  X ) takes into account the initial reserve u of the company. By Jensen 's inequality, v(H(X)  EX) > Ev(H(X)  X) = 0 so that H(X) > EX. For v(x) = x, we have equality and are back to the net premium principle. There is also an approximate argument leading to the variance principle as follows. Assuming that the Taylor approximation
v(H(X)  X) ^ 0 +v'(0)(H (X)  X) + v 0 (H(X)  X)2 ,/2
is reasonable , taking expectations leads to the quadratic v"H(X )2 + H(X) (2v'  2v"EX) + v"EX2  2v'EX = 0 (with v', v" evaluated at 0) with solution
H(X)=EXv^±V( ^ )2Var(X).
Write
( vI ) 2 \
Var(X) v^  2v^Var(X)/ I  (
, Var(X) )2
If v"/v' is small, we can ignore the last term. Taking +f then yields H(X) ,:: EX 
2v'(0) VarX;
since v"(0) < 0 by concavity, this is approximately the variance principle. The most important special case of the principle of zero utility is The exponential principle which corresponds to v(x) = (1  e6x)/a for some a > 0. Here (5.1) is equivalent to 0 = 1  e0H(X)EeaX, and we get
H(X) = 1 log Ee 0X .
a
5. PRINCIPLES FOR PREMIUM CALCULATION
325
Since m.g.f.'s are logconcave, it follows that H,, (X) = H(X) is increasing as function of a. Further, limQyo Ha (X) = EX (the net premium princiHa (X) = b (the premium ple) and, provided b = ess supX < oo, lim,, H(X) = b is called the maximal loss principle but is clearly not principle very realistic). In view of this, a is called the risk aversion The percentile principle Here one chooses a (small ) number a, say 0.05 or 0.01, and determines H(X) by P(X < H(X)) = 1  a (assuming a continuous distribution for simplicity). Some standard criteria for evaluating the merits of premium rules are 1. 77 > 0, i .e. H(X) > EX. 2. H(X) < b when b (the ess sup above ) is finite 3. H(X + c) = H(X) + c for any constant c
4. H(X + Y) = H(X) + H(Y) when X, Y are independent
5. H(X) = H(H(XIY)). For example , if X = EN U= is a random sum with the U; independent of N, this yields
H
C^
U; I = H(H(U)N)
(where, of course, H(U) is a constant). Note that H(cX) = cH(X) is not on the list! Considering the examples above, the net premium principle and the exponential principle can be seen to the only ones satisfying all five properties. The expected value principle fails to satisy, e.g., 3), whereas (at least) 4) is violated for the variance principle, the standard deviation principle, and the zero utility principle (unless it is the exponential or net premium principle). For more detail, see e.g. Gerber [157] or Sundt [354]. Proposition 5.1 Consider the compound Poisson case and assume that the premium p is calculated using the exponential principle with time horizon h > 0. That is,
N,,
Ev I P  E U;
i =1
= 0 where
v(x) = 1(1  e°x
a
Then ry = a, i.e. the adjustment coefficient 'y coincides with the risk aversion a.
326
Proof The assumption means
CHAPTER XI. MISCELLANEOUS TOPICS
0 a (1  eareo (B[a11)
l
i.e. /3(B[a]  1)  ap = 0 which is the same as saying that a solves the Lundberg u equation. Notes and references The theory exposed is standard and can be found in many texts on insurance mathematics, e.g. Gerber [157], Heilman [191] and Sundt [354]. For an extensive treatment, see Goovaerts et al. [165].
6 Reinsurance
Reinsurance means that the company (the cedent) insures a part of the risk at another insurance company (the reinsurer). Again, we start by formulation the basic concepts within the framework of a single risk X _> 0. A reinsurance arrangement is then defined in terms of a function h(x) with the property h(x) < x. Here h(x) is the amount of the claim x to be paid by the reinsurer and x  h(x) by the the amount to be paid by the cedent. The function x  h(x) is referred to as the retention function. The most common examples are the following two: Proportional reinsurance h(x) = Ox for some 0 E (0, 1). Also called quota share reinsurance. Stoploss reinsurance h(x) = (x  b)+ for some b E (0, oo), referred to as the retention limit. Note that the retention function is x A b. Concerning terminology, note that in the actuarial literature the stoploss transform of F(x) = P(X < x) (or, equivalently, of X), is defined as the function
b * E(X  b)+ =
f
(s  b)F(dx) _ f
6 00
(x) dx.
An arrangement closely related to stoploss reinsurance is excessofloss reinsurance, see below.
Stoploss reinsurance and excessofloss reinsurance have a number of nice optimality properties. The first we prove is in terms of maximal utility: Proposition 6.1 Let X be a given risk, v a given concave nondecreasing utility function and h a given retention function. Let further b be determined by E(X b)+ = Eh(X). Then for any x,
Ev(x  {X  h(X)}) < Ev(x  X A b).
6. REINSURANCE
327
Remark 6 .2 Proposition 6.1 can be interpreted as follows. Assume that the cedent charges a premium P > EX for the risk X and is willing to pay P1 < P for reinsurance. If the reinsurer applies the expected value principle with safety loading q, this implies that the cedent is looking for retention functions with Eh(X) = P2 = P1/(1 + 77). The expected utility after settling the risk is thus
Ev(u + P  P1  {X  h(X)})
where u is the initial reserve . Letting x = u + P  P1, Proposition 6.1 shows that the stoploss rule h (X) = (X  b)+ with b chosen such that E(X  b)+ u = P2 maximizes the expected utility. For the proof of Proposition 6.1, we shall need the following lemma: Lemma 6 .3 (OHLIN'S LEMMA) Let X1, X2 be two risks with the same mean, such that Fj(x) < F2 (x), x < b, Fi(x) ? F2(x), x > b for some b where Fi(x) = P(Xi < x). Then Eg(X1) < g(X2) for any convex function g. Proof Let Yi=XiAb, Zi=Xivb.
Then
P(Yl < x) _ Fi(x) <_ F2 (x) = P(Y2 < x) x < b 1=P(Y2<x) x>b so that Y1 is larger than Y2 in the sense of stochastical ordering . Similarly, P(Zl < x) _ 0 = P(Z2 < x) x < b Fi(x) > F2(x) = P(Z2 < x) x > b
so that Z2 is larger than Zl in stochastical ordering. Since by convexity, v(x) = g(x)  g(b)  g'(b)(x  b) is nonincreasing on [0, b] and nondecreasing on [b, oo), it follows that Ev(Y1) < Ev(Y2), Ev(Zi) < Ev(Z2). Using v(Yi) + v(Zi) = v(Xi), it follows that
0 < Ev(X2)  Ev(Xi) = Eg(X2)  Eg(X1),
using EX1 = EX2 in the last step. u
Proof of Proposition 6.1. It is easily seen that the asssumptions of Ohlin' s lemma hold when X1 = X A b, X2 = X  h(X); in particular, the requirement EX1
328
CHAPTER XI. MISCELLANEOUS TOPICS
= EX2 is then equivalent to E(X  b)+ = Eh(X). Now just note that v is convex. u
We now turn to the case where the risk can be written as N
X = Ui
i=1
with the Ui independent; N may be random but should then be independent of the Ui. Typically, N could be the number of claims in a given period, say a year, and the Ui the corresponding claim sizes. A reinsurance arrangement of the form h(X) as above is called global; if instead h is applied to the individual claims so that the reinsurer pays the amount EN h(Ui), the arrangement is called local (more generally, one could consider EN hi(Ui) but we shall not discuss this). The following discussion will focus on maximizing the adjustment coefficient. For a global rule with retention function h* (x) and a given premium P* charged for X  h* (X), the cedents adjustment coefficient y* is determined by
1 = Eexp {ry*[X  h*(X)  P*]},
for a local rule corresponding to h(u) and premium P for X look instead for the ry solving
J _f
(6.2) N 1 h (Ui), we
[ X_P_^
1 = Eexp
[ Ei  h(Ui)] P [U
= Eexp{ry
h(Ui)]
l (6.3) This definition of the adjustment coefficients is motivated by considering ruin at a sequence of equally spaced time points, say consecutive years, such that N is the generic number of claims in a year and P, P* the total premiums charged in a year, and referring to the results of V.3a. The following result shows that if we compare only arrangements with P = P*, a global rule if preferable to a local one. Proposition 6.4 To any local rule with retention function h(u) and any
N
J}
P > E X  N h(Ui)
4 =1
(6.4)
there is a global rule with retention function h* (x) such that
N
Eh*(X) = Eh(U1)
i=1
and 'y* > ry where ry* is evaluated with P* = P in (6.3).
4).. This follows by taking Xl = U A b.P } < 1 = Eexp E[Ui.4). expectations like those in (6. however.d. i. (6. Local reinsurance with h(u) = (u .h(Ui)] . then (6.h(Ui)P JJJ l:='l {ry ] or. y = Ei [Ui . u But since ry > 0. ry* > 0 because of (6. Eexp 7 [E [Ui . REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x . we get N 1 = Eexp ry E[Ui ii .3). X2 = U .h( UU) = EN • E[U .P.4.d.6 Assume the Ui are i.5) reduce quite a lot. appealing to (6. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.b)+ = Eh(U) (and the same P) satisfies 71 > ry. N E X .4).b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6.6).h(Ui)] . Proof As in the proof of Proposition 6.4) and u g(x) = e7x in Ohlin's lemma. it suffices to show that Eexp {ry ii 'UiAb.5 Because of the independence assumptions . Then for any local retention function u . this implies 7* > 7.h * (X) . . we get EX = EN • EU.6. Assuming for simplicity that the Ui are i.P]}.h(U)].i. and so on.P I = EC [7]N.h(u) and any P satisfying (6. the excess ofloss rule hl (u) = (u .6) u where C[ry] = Ee'r(u4(u)). ' ii (6. The arrangement used in practice is. Remark 6.5) holds trivially.P > EexP{7[X .b)+ with b determined by E(U . as often local as global.h(Ui)] . (6.h(U) (as in the proof of Proposition 6. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .
see also Sundt [354]. The original reference for Ohlin's lemma is Ohlin [277]. The present proof is from van Dawen [99]. e.330 CHAPTER XI. [76].many texts on insurance mathematics. . See further Hesselager [194] and Dickson & Waters [120]. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in.g. Bowers et at. Heilman [191] and Sundt [354].
Technically. Y2. not concentrated on {h.. of interarrival times and the time Yo = To of the first arrival (that is. t] is denoted by Nt.t. . If F satisfies the stronger condition of being spreadout (F*' is nonsingular w ..e.. .. The number max k : Tk_j < t of renewals in [0.} for any h > 0. note in particular that U({0}) = 1. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. of epochs or the set Y1. Y...T„_1)..U(t) is the expected number of renewals in (t. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . the distribution of Yo is called the delay distribution. the renewal process is called zerodelayed. stating that U(t+a)U (t) ^ a. some condition is needed: that F is nonlattice. 2h.. If Yo = 0. i. are independent and Y1. Y1. t 00 (A. = T„ . Then Blackwell 's renewal theorem holds... .r. The renewal theorem asserts that U(dt) is close to dt/µ. then Stone 's decomposition holds : U = U. denoted by F in the following and referred to as the interarrival distribution. when t is large. That is. Lebesgue measure for some n > 1).Appendix Al Renewal theory la Renewal processes and the renewal theorem By a simple point process on the line we understand a random collection of time epochs without accumulation points and without multiple points. t +a]). The point process is called a renewal process if Yo. t]) so that U(t + a) . U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. The mathematical representation is either the ordered set 0 < To < T1 < .1) (here U(t) = U([0. all have the same distribution. Y2. . Lebesgue measure dt normalized by the mean to of F..
4) that z is Lebesgue integrable with limZ.i". stating that U(t)/t > 1/p. Both result are valid for delayed renewal processes.4) If F is spread. z(u) a known function.out. oo).2) has the unique solution Z = U * z. (A. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. see [APQ] Ch. IV). u u PF 4 00. 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 . and that F has a bounded density2. i. the statements being EN(t + a) .332 APPENDIX u(t) has limit 1/µ as t 4 oo. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. Under weak regularity conditions (see [APQJ Ch.R. In 111. (A. then it suffices for (A. IV). (A. Note in particular that F is spreadout if F has a density f. Equivalently.x)F(dx). z(x) = 0.i. and F(dx) a known probability measure .2 Assume that Z solves the renewal equation (A. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem.e. Then Z(u) 4 z(oo).3) Further. resp.5) 2This condition can be weakened considerably . that z(u) has a limit z(oo) (say) as u 4 oo.2).1 if F is nonlattice and z (u) is directly Riemann integrable (d.2) Z(u) = J0 u z(x)U(dx). then Z(u) i f0 z(x)dx . (A.a. U Z(u .. in convolution notation Z = z + F * Z.EN(t) . but suffices for the present purposes .9. µF (A.
} be a renewal process. that the existence of y may fail for heavytailed F. Tk and {Xt }o<t<Tk • For example. multiply (A. However. i. T1. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). is called the cycle length distribution and as before.... .r. Note. however. • . A stochastic process {Xt}t>0 with a general state space E is called regenerative w.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). . asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. T1.d. Assuming that y can be chosen such that f °° Ox F(dx) = 1. Y1 . The kth cycle is defined as {XTk+t}o<t<Yk . The simplest case is when {Xt} has i. Yk ). or many queueing processes..t. cycles. .APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. of Yo..5a. and its distribution does not depend on k. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles).. However. 1c Regenerative processes Let {T. Z(u) U = 1 u 1 u f z(u . z(x) = e7xz(x). . Y2. Tk (or. To this end. We let FO. The distribution F of Y1. {Tn} if for any k.. Here the relevant F does not have mass one (F is defective).x)u(x) dx = z(u( 1 . the postTk process {XT. that F is a probability measure. .. refer to the zerodelayed case. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. . 0 PF µF 11 In risk theory. F(dx) = e7xF(dx). Eo etc.k+t }t>o is independent of To.e.3) satisfied by the ruin probability for the compound Poisson model. A regenerative process converges in distribution under very mild conditions: . This program has been carried out in III. equivalently. the present more general definition is needed to deal with say Harris recurrent Markov chains. Hence by dominated convergence. we let µ denote its mean.i.. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. a basic reason that renewal theory is relevant is the renewal equation II. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i.(3.
e.4 Let {Zt}t^. If p = oo.+ X.r. under the condition of Blackwell's renewal theorem.ZT }0<t<Y„+. 2. then (Zt . C).ZTOI < 00. Then {Zt}t^. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt.d... Otherwise . {Tn}.. assume that p < 00 and define Un = ZT}1 .i.3. fi (t) = inf {Tk .e.t. {Tn}.t.Tk : t < Tk} as the age. resp . 0<t<Yi then Zt /t a$• EU1/µ. {Tn} if the processes {ZT +t ..r. Then it (ii. [0.6) id Cumulative processes Let {Tn} be a renewal process with i.. and q(t) = sup It . just the same proof as there carries over to show: Proposition A1.d. r. oo). C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p. e(t )) . We denote the limiting r. for n = 1. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t. This is the case considered in [APQ] V.ZT Then: (a) If E sup I ZTo+t .v..... An example is Zt = fo f (X8) ds where {Xt} is regenerative w.0 be cumulative w. but in fact.oo (i.. i.t : t < Tk}.0 is called cumulative w.'s by e. {i7(t)} are Markov with state spaces (0. then e (t) ... are i. Then {e(t)}. oo). and we have: holds more generally that (rl(t).t. (b) If in addition Var(Ul ) < oo. cycles (we allow a different distribution of the first cycle).tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1. where the distribution of X.i. then Xt . in total variation.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo.r. (A. Then Xt Di X.334 APPENDIX Proposition A1. µ 0 If F is spreadout.
U(x) < U( 1)).4.APPENDIX 335 Theorem A1. Then fi(t)/t a4' 0 and.U(x) (c < oo because it is easily seen that U(x + 1) .d.i. Yl > t]. . Y1 > t] 4 0.5 Under the condition of Blackwell's renewal theorem.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). ^ > y) = 1 f +Y (z)dz. Hence t t lt ) = f U(dy)z(t .i. Since z ( k) < E[Yi .^(t))} as a regenerative process. Hence for t large enough. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.6 Consider a renewal process with µ < oo. U(x + 1) . For the second. we used: Proposition A1. Proof The number Nt of renewal before t satisfies Nt/t a4' p. the first statement follows. Y1i Y2.v.t. In IV.t. and the conditional distribution of ri given l. and the equivalence of (a) with (b)(d) is an easy exercise. Since the maximum Mn of n i. but governed by a Markov chain {Jn} (we .y) = f U(t . EC(t)/t + 0. = z is Foz) The proof of (a) is straightforward by viewing {(r... if in addition EYo < oo.(t).y)P(Yo E dy) .dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. l:) is the same as the distribution of (VW. are not i. the sum is o(t) so that Eo£(t)/t + 0 . ^) is given by the following four equivalent statements: (a) P (77 > x.. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . (b) the joint distribution of (ri. r. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y). the joint distribution of (rl.d. (d) the marginal distribution of ^ is FO. (1 V)W) where V. 1) and W has distribution Fw given by dFw/dF(x) = x/pF. use t E^(t)/t = E[Yo . assume first the renewal process is zerodelayed. Yo > 0] + f Eo^ (t . V is uniform on (0. W are independent. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 . In the general case. (c) the marginal distribution of q is FO.
. oo).r. A stochastic process {Xt}t>o is called semiregenerative w. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE.... 0] or (0 . . A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. . < yIJ) = Fij( y) on {Jn= i. G+(x) = P(S. ..jEE is a family of distributions on (0. Yn. Jn +1=j} where J = a(JO.t. Y1. distribution ofjXt}t>o itself where Pi refers to the case Jo = i. X2. the conditional distribution of {XT„+t}t>o given Yo. r+ < oo). Then Xt 4 Xo. . Jo.) and (Fij )i. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. Alsmeyer [5] and Thorisson [372]. namely {Twk } where {Wk } is the sequence of instants w where Jo. Further: Proposition A1..}. with common distribution F. Jn = i is the same as the P. and define r+=inf{n>0: Sn>0}. ..and regenerative processes. = io for some arbitrary but fixed reference state io E E. A Markov renewal process {Tn} contains an imbedded renewal process.T_ < oo).d. be i.g.... G_(x) = P(ST_ < x. For example. T_=inf{n>0: Sn<0}.. in [APQ].} is nonlattice (it is easily seen that this definition does not depend on i). the semiregenerative process is called nonlattice if {T. e. oo). . The semiregenerative process is then regenerative w.i . Jn_1. Sn = X1 + • • • + Xn the associated random walk. Let X1. . Notes and references Renewal theory and regenerative processes are treated...r. IT. where the distribution of X. . These facts allow many definitions and results to be reduced to ordinary renewal.+ < x. ..t.. . the Markov renewal process if for any n. J1 i .7 Consider a nonlattice semiregenerative process.
=n w=m i Figure A. A C (0.. A C (0.r. 0]). Sr_ _1 is at its minimum . 0<j<m. A C (oo. S. F(A) + (G+ * G_)(A).g. F(A .7). (A..7) follows since G+(A) = 0 when A C (oo. G+. oo).S.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F..1 . 0). On {T_ > 2}. define w as the time where the preT_ path S1. we consider the last such time (to make w unique) so that {w=m.7) (A. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). n=0 The basic identities are the following: Theorem A2.1 (a) F = G+ + G_ . n=0 n=0 00 00 and the T+.x)R+(dx). In (A. oo). n 0 R_(A) = E I(Sn E A). (c) G+(A) = f °. More rigorously. the renewal measures U+=>G+. m<j<n}. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A).=EGn. oo) (A.x)R_ (dx).G+ * G_: (b) G_ (A) = f °° F(A . (e) R_ = U+. Proof Considering the restrictions of measures to (oc.>0. (d) R+ = U_. >0. G_.T_=n} = {S. 0]. . u .S. 0] and (0. .8) (e. A C (oo. U.
and the proof of (A.m. .3 8 APPENDIX Reversing the time points 0.3. 0 < k < n._ = n . A..u) f0m m=1 n=m+1 00 J0 OO P(S. It follows that for n > 2 F (7. SmEdu) = P(T+=m. clearly (Sj Sm>0...Sn_1Edx.XnEAx) 00 f 0 f 0 00 00 1: F(A .. ST+Edu). 0<j<m. m it follows (see Fig.0<k<ri .x)R+(dx).. Aso. S.= n. ST_ E A . E du) = P(T_=nm. (A.8) is similar. and reversing the order of summation yields P(T_ > 2.1). . m=1 f S mming over n = 2.1. Sn1 E dx) n=1  F(A . r+ = n) n=1 n=1 0  C0 E fF(Sk< 0._ E A . SnEAIS.7) follows. .x)P(Sk < 0._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu)...+ E du)P(S.1) that P(Sj Sn. (b) follows from 00 G+ (A) _ E F(Sn E A. ST_ E A) P(T+ = m. m < j <n. Sr_ E Adu) (s ee again Fig .>0.du) (G+ * G)(A)• C llecting terms. A.+ E du) E P(S. S..F(r_n_mSrEA_u).
APPENDIX 339 and the proof of (c) is similar. is based upon representing G+ as in (b).T+> n) = P(Sk < O.F[s] = (1 . Then for A C (oo. For example. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure).0<k<n. . and similarly H_ (s) = 1 .O<k<n. and G+. the derivation of the form of G+ for the compound Poisson model (Theorem 11. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1).SnEA) = P(Sn<Sk.O<k<n. oo). E.1.4). cf. it serves as model and motivation for a number of results and arguments in continuous time. which is basic for the PollaczeckKhinchine formula.. In discrete time. WienerHopf theory is only used at a few places in this book.SnEA) = P(SnSn_ k.6.0+[s])(1 . Another main extension of the theory deals with Markov dependence. Summing over n yields R+ (A) = U_ (A). Again.SnEA) = P(Sn<Sk. G_ are trivial.Sn_k. and sometimes in a larger strip. Sk = X1 + • • • + Xk = Sn . then T+ = inf It > 0 : St = 0} is 0 a.1.G_[s]) (A. the survey [15] by the author and the extensive list of references there. see e. there are direct analogues of Theorem A2. P(SnEA . In continuous time. The classical analytical form of the WienerHopf problem is to write 1 . consider a fixed n and let Xk = Xn_k+l. if {St} is Brownian motion. and the proof of (e) is similar. u Notes and references In its above discrete time version.g. such developments motivate the approach in Chapter VI on the Markovian environment model. we can rewrite (a) as 1 .1(a) is from Kennedy [228]. 6+ [s].P as a product H+H_ of functions with such properties.1).g. see for example Bingham [65]. However. The present proof of Theorem A2. 11.9) whenever F[s]. there is no direct analogue of Theorem A2. 0].. being concentrated at 0.2 In terms of m.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. Since G+ is concentrated on (0.s. u Remark A2. a number of related identities can be derived. G_ [s] are defined at the same time.SnEA) is the probability that n is a weak descending ladder point with Sn E A. this holds always on the line its = 0. In this generality of.g. Nevertheless.f.'s. For (d).0<k<n. the analogue of a random walk is a process with stationary independent increments (a Levy process.
11) A f eAtdt = eA. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step). To circumvent this. however . Here are. three of the c rrently most widely used ones: xample A3. Eo Kn/n! converges rapidly and can be evaluated without p oblems.12) eA'AO = Ale AA (A.1 (SCALING AND SQUARING) The difficulty in directly applying t e series expansion eQ = Eo Q"/n! arises when the elements of Q are large. It is seen from Theorem VIII. ere A is the eigenvalue of largest absolute value. if m is s fficiently large. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum).10) d dteAt = AeAt = eAtA (A.5 that when handling phase type distributi ons. whereas there is no similar single established a proach in the case of matrix exponentials. 1. _I 0 (A. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. 0 .340 APPENDIX 3 Matrixexponentials T e exponential eA of a p x p matrix A is defined by the usual series expansion 00 An eA n=0 n! he series is always convergent because A' = O(nk Ialn) for some integer k < p. Thus.13) henever A is a diagonal matrix with all diagonal elements nonzero. 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). Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . and eQ can then be computed as the mth power (by squaring if = 2).
i.15) Then it is easily checked that P is a transition matrix .14) E n n=0 which is easily seen to be valid as a consequence of eqt = en(Pr)t = entenpt The idea which lies behind is uniformization of a Markov process {Xt}. Ap. vp be the corresponding left . . p different eigenvalues Aj i ..4 (DIAGONALIZATION) Assume that Q has diagonal form. . i. To this end.e.. some jumps are dummy in the sense that no state transition occurs ). 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 = q22.2 (UNIFORMIZATION) Formally. Let vi. However .. In practice. Here is a further method which appears quite appealing at a first sight: Example A3 .e.3 i (A. Zo = a (Z = QZ.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. 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.. condition upon the number n of Poisson events in [Olt])  Example A3.APPENDIX 341 Example A3.]t)n (A.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (. Zo = h). the procedure consists in choosing some suitable i > 0. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. 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 . One then can reduce to p linear differential equations by noting that k = ZQ. 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.
Qhi = vihi. Then vihj = 0. Then P P Q = > Aihivi = E Aihi (9 vi.18) contains terms which almost cancel and the loss of digits may be disasterous. i # j. not all ai are real... In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials.. i=1 i=1 Thus.g H1.16) (A. and hence A2 is so because of A2 = tr(Q). v5Q = Aivi. (A..5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. hi have been computed. we can take H as the matrix with columns hl. i= 1 i=1 P P (A. (A. say Al. under the conditions of the PerronFrobenius theorem). There are. The phenomenon occurs not least when the dimension p is large. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. Nevertheless.342 APPENDIX (row) eigenvectors and hl.. and vihi ¢ 0. Complex calculus : Typically. and writing eQt as eQt = He°tH1 = H (e\it)di. hp.17) eQt = E e\`thivi = E ea:thi ® vi. we have an explicit formula for eQt once the A j.. vi. however. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. say A = (Ai)diag. and we may adapt some normalization convention ensuring vihi = 1. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close. of largest real part is often real (say.. some cases remain where diagonalization may still be appealing.. D = ) 2 2 .18) Namely. Example A3. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. hp the corresponding right (column) eigenvectors. the eigenvalue.
b are any constants ensuring//Irk = 1.e.k1).Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector. 1) . l ab (g12g21 + (A1  411) 2) = 1. u Example A3. replacing ai by A2. However.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. The other eigenvalue is A = A2 = q1 . v2 and h2 can be computed in just the same way.20) ir = q2 ql qi +q 2 9l +q2 (A. Then 7r = (ir1 7r2 ) = a (q21 Al . eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A. h2 = Thus.q. i. Of course. where (A. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 .6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. k  C k2 ) =b ( A1 q 1 Q11 / where a .19) Example A3 .7 Let 3 9 2 14 7 11 2 2 .
.5 .satisfying AAA = A.23) .11/2 . Generalized inverses play an important role in statistics. 2 2 1=ab(142+(1+2)2 ) = tab. but only that dimensions match . 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_.344 Then D= 2+ 11)' 7 T4 2 =52. APPENDIX x1 3/2 .6. (AA+)' = AA+.11/2 + 5 1. A2 = 3/2 . A+AA+ = A+. and a generalized inverse may not unique. e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. (A.22) Note that in this generality it is not assumed that A is necessarily square.. (A. for example AA+A = A. (A+A)' = A+A. They are most often constructed by imposing some additional properties .
lt o eAx dx = te7r + D(eAt .23) is called the MoorePenrose inverse of A.e ® 7r)1.eir)1 = I . = 0 where m < p is the rank of A..g. .P). Here is a typical result on the role of such matrices in applied probability: Proposition A4. and define D = (A .g.. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .I) (A.ew. one is also faced with singular matrices . E. Assume that a unique stationary distribution w exists .. Am > 0.. .= (I .P + e7r)1 (here ( I . are ordered such that Al > 0. These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . _ A. 0 01 In applied probability.eir )1.24) = te7r ..25) .1Q = Q(Q .eir ).D + O(ebt). one then works with Q = (Q . ( Q .. Am+1 = .1 Let A be an irreducible intensity matrix with stationary row vector it. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A .P + e7r ).APPENDIX 345 A matrix A+ satisfying (A. Then for some b > 0. Rather than with generalized inverses . (I . most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. if A is a possibly singular covariance matrix (nonnegative definite). (A.1 goes under the name fundamental matrix of the Markov chain). and exists and is unique (see for example Rao [300]).e.
Note that h ® it has rank 1. h as 1 x m and k x 1 matrices.eir)eAt = eAt = A'(t). B'(t) = e7r + DAeAt = eir + (I .D + D2 + O(ebt). in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. the rows are proportional to it. see below. . For example.24).26) 2 = 2 e7r + tD . the formulas involving O(e6t) follow by PerronFrobenius theory.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). I.I) .27) Proof Let A(t). (A. 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. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4.J {xe^r + D(e .s. the r.h. . o Finally. of (A. resp. respectively. B(t) denote the l.2e7r .346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . Interpreting 7r.I)} dx.91a(2) . Then A(O) _ B(O) = 0.I)}.. (A.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .h.e.s. Equivalently.2 Let it be a row vector with m components and h a column vector with k components. h ® it reduces to hit in standard matrix notation.I) (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. and the columns to h. it follows that h ® it is the k x m matrix with ijth element hi7rj . and in fact any rank 1 matrix can be written on this form.DZ(ent .
and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) . and the number of such factors is precisely given by the relevant binomial coefficient.29) If A and B are both square (k1 = ml and k2 = m2). if Al = vi. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I . Using (A.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^. A2 = v2 are row vectors and C1 = h1. (A. each of which is A ® I or I ® B.5v/. then v1B1h1 and v2B2h2 are real numbers. (A B)' = eA®B e! L 1=0 0 . then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).APPENDIX 347 Then A®B = 2 f 20.3vV/72f 20.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2).A9.29). (A. if A ® I occurs k times.4 eA® B = eA ®eB.28) In particular. 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.3v'6.50 6 7 6 4f 4.3f 4v/. such a factor is Ak (&B 1k according to (A.31) Indeed.31).3V8.4vf.(A. C2 = h2 are column vectors.
{ 1't(1) }. first term on the r . where transition matrix of the bivariate Markov chain {X n1).32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). Thus . Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . Yt(2 ) }. X ) }.33) . k any column vectors. 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 +.348 APPENDIX Remark A4. resp .I)(h ® k). Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .32).4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines .s. represents ces Q( 1). and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. independent Markov chains. h.3 < 0 Lemma A4 . {Yt(1). A special case of Proposition A4. Let further it. and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . P(2). From what has been said about matrices of {Yt( 1).5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. Q(2). p = P(1) ® {X }. P(t) Yt(2) }. Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. in the definition (A. the same time. we have P8 = Pal) ® p(2). the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). Let P8f P(Sl). P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . { On the other hand. (A.6 Suppose that A and of B.
. = j and atk_li. in such that io = i. Similarly.7 Let A be a p x pmatrix with nonnegative elements. which can be found in a great number of books.. 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). . h such that vh = 1. That is. then IN < Ao for all A E sp(A). > 0 for k = 1. j = 1. [APQ] X. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. and if we normalize v. and the corresponding left and right eigenvectors v. .12). then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. h = e and v = 7r (the stationary row vector). i.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. il. Here is the PerronFrobenius theorem. (A. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A.g. E (0. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. .. .29).34) Note that for a transition matrix.3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. Then the eigenvalue Ao with largest real part is simple and real.. Now note that the eigenvalues of A ® B are of the form a +. and appeal to (A. see e. (b) if in addition A is aperiodic.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements.. f o r each i. n. . and the corresponding left and right eigenvectors v. h can be chosen with strictly positive elements. so that by asssumption A ® B is u invertible. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). p there should exist io.. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements.APPENDIX 349 Proof According to (A.. . . h can be chosen with 3By this. ..
Then for any (3.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. not only in the tail but in the whole distribution. let {Yti°i } be a Markov process with initial distribution a and intensity . but is an easy consequence of the PerronFrobenius theorem.(3.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). if we normalize v. h such that vh = 1. 10) and use the formula me at e Bt = e 00 Antn = e . The next result gives a condition for asymptotical exponentiality. Ao).e. 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). then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. h = e and v = 7r (the stationary row vector). we have A0 = 0.8 is most often not stated explicitly in textbooks. Furthermore.(ti)ding. the analogy of this procedure with unformization. For example.35) for some p E (oo. the condition is that t is small compared to Q.8. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. note that we can write the phase generator T as Q . Example A3. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . To this end. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j).1. the phasetype distribution B(a) with representation (..350 APPENDIX strictly positive elements. Note that for an intensity matrix.n t AL n=0 n! (cf. 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.2). Corollary A4. Proposition A5. I.
Hence O ((a) aa.YQ(av) = j) Pi ( ci(a'V) > x. and that Yt(a) = Yat for all t. it states that the state.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. a' = a . Let further V be exponential with intensity V and independent of everything else.1. from which the phase process is terminated .x (1 . a'/a + 1. J(()) _ = i) + at•x t tt' . dx/ti] or not. We shall . 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. and this easily yields a(x)/x a' 1/t*. Since JJ(. By the law of large numbers for Markov processes .aE where 0 < e < 1).jEE.(a) > x . Then {Ix} is a Markov process with to = Yo.9. {t Y( a) } v>0 . and write Yt = Yt(1).Yj(av) = j f . v/ t. Conditioning upon whether { Yt} changes state in [0. t < (a). a .g. has a limit distribution: Proposition A5. In addition to the asymptotic exponentiality. fo tY dv/t a$' t*. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .bij) Hence the intensity matrix of { Ix} is (qij/ti)i. J^O)_ = j) Pi (v(aaV) > x.a' + oo (e. 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}. We can assume that Jta) = Yt(°). Then a(a'V)/a (aV) a' 1. in fact . = YQ(x).APPENDIX 351 ((1) etc. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. prove a somewhat more general result which was used in the proof of Proposition VI. Proof Assume first ti > 0 for all i and let I.2 Pi (c(a) > x.
Et II I a(a^V) > x) at' . Gnedenko & Kovalenko [164] and Glasserman & Kou [162]). However. See also Korolyuk. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case...Pe. Keilson [223]. > 0}. k>1. A distribution B on {1.} is said to be discrete phasetype with representation (E. Indeed.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .2 do not appear to be in the literature. a). u Notes and references Propositions A5.. (c) the nth moment k 1 k"bkis 1)"n!aP"p.g. so we shall be brief. . .1 and A5.. = 0 for one or more i.. is discrete phasetype.j) and initial distribution a. ' pk 0 k>1 11 Theorem A5.x k > K.. . say bk = 0. 2. P. Example A5. K}. a = b = (bk)k=1. these results are in the spirit of rare events theory for regenerative processes (e. the simplest discrete phasetype distribution: here E has only one element.. (b) the generating function b[z] _ E' . 1 k=1 1 0 otherwise.. Example A5. Penev & Turbin [238].3 As the exponential distribution is the simplest continuous phasetype distribution..p)k1 p. so is the geometric distribution. with point probabilities bk = (1 . let E and Pkj j=k1.zP)'p.4 Any discrete distribution B with finite support. Then P is substochastic and the vector of exit probabilities is p = e . k = 1.5 Let B be discrete phasetype with representation (P.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t.. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. Then: (a) The point probabilities are bk = aPklp.. an easy modification of the argument yields finally the result for the case where t. 2. zkbk is za(I . a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p.
T= ( 0 T(2) ) (A. { Jt 2) } with lifetimes U1 .. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). B2 be phasetype with representations (E(1)...36) in blockpartitioned notation (where we could also write a as (a (1) 0)).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. Jt t > U1 + U2. as is seen by minor modifications of Example A5. Then {Jt} has lifetime U1 + U2 . a' . . (E(2). U2. 11 Example A5. resp.r + 1.6 is the Erlang distribution Er which is the convolution of r exponential distributions.. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U.a(1).T(1)). A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.T(2)).APPENDIX 353 5c Closure properties Example A5. and hence the negative binomial distribution is discrete phaseu type. a. A. and a=1).2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. Then the convolution B = B1 * B2 is phasetype with representation (E.a(2). r .6.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5.6 (CONVOLUTIONS) Let B1. _ i E E(1) T(1) t(1)a(2) i E E(2) . initial distribution a and phase generator T. resp.
U2. i E E(2) 0 T(2) =IT (in blockpartitioned notation. Then the mixture B = 9B1 + (1 . this means that a = (Oa(1) (1 .0)ai2). Example A5. Then it is trivial to see that B(") is u phasetype with representation (a(").a(1). A reduced phase diagram is 0a(1) E(1) A . p at each termination.T(1)).p.0)a(2))).p)pn1. are i.354 APPENDIX Example A5.O)B2 (0 < 0 < 1) is phasetype with representation (E.. and consider B(") = fA B(a) v(da) where v is a probability measure on A.37) (1) (1 .T(2)). To obtain a phase process for C. a.0)a(2) E(2) Figure A.T. Thus. with common distribution and N is independent of the Uk and geometrically distributed with parameter p.i.p)pn1B*n. a mixture of more than two phasetype distributions is seen to be phasetype. In risk theory. resp. T) and C = EO°_1(1 . Example A5. (E(2).a(2). P(N = n) = (1 . Let B(") be the corresponding phasetype distribution.E) where a(°) = fAa(a)v(da). a. Equivalently.8 (FINITE MIXTURES) Let B1.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E.'). one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims. B2 be phasetype with representations (E(1). we need to restart the phase process for B w.d. if U1. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).4 .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.. i E E(1) T 0 I (A. then C is the distribution of Ul + • • • + UN. and o'i Oa. a reduced phase diagram is f a E t Figure A..3 In exactly the same way.
Thus the representation is (E(1) x E(2). let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). i. For U1 A U2.d. If we replace x by a r.. say v. Example A5 . Corollary VIII. 12 (PHASETYPE COMPOUNDS ) Let fl.. resp.. but the same T.. B2 of phasetype with representations (E(').2. X independent of U. P). of F. a. say with distribution F. T). let the phase space be E x F = {i j : i E E. T + ta. T + pta). Example A5 .aF[T]. .g.T) if U is phasetype with representation (E.TWWW). U2 be random variables with distributions B1. be the point probabilities of a discrete phasetype distribution with representation (E. U2. { 4 } as exit of {Jt}. then U1 +• is phasetype with representation (E.. f2. Proposition VIII. { Jt2) } be independent with lifetimes U1. resp. Minor modifications of the argument show that 1. if B is defective and N + 1 is the first n with U„ = oo.°_1 f„ B*?l. let {Jtl)}.APPENDIX 355 and C is phasetype with representation (E.1. then C is the distribution of U1 + • • • + UN. 13 (MINIMA AND MAXIMA ) Let U1.x)+.2. Equivalently. To see this. let B be a continuous phasetype distribution with representation (F. if U1.°. T(2) ). are i.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . Indeed.. E). we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. If U1 has a different initial vector. if {Jt} is a phase process for U. v. cf.T + pta). with common distribution B and N is independent of the Uk with P(N = n) = f.a(1).X)+ is zeromodified phasetype with representation (E.. a(1) ® a(2 ).. T) and C = F. To obtain a phase representation for C .T) where F[T] = J0 "o eTx F(dx) u is the matrix m.9) that (U . then U1 + • • + UN is zeromodified phasetype with representation (a.v. a. cf. (E(2). . then Jy has distribution aeTx. j E F}. Example A5. T(1) ® T(2)). +UN 2. U2.7. it follows by mixing (Example A5.aeTx. v. a(2). a. Note that this was exactly the structure of the lifetime of a terminating renewal u process. It is zeromodified phasetype with representation (E.f.
..(bk) + B(bk) for all k as n * oo.. Here are the details at two somewhat different levels of abstraction: (diagonal argument . however... relies more on matrix algebra than the probabilistic interpretation exploited here). and the phase generator is T(1) ®T(2) T(1) ®t(2) t(1) ® T(2) 0 T(1) 0 0 0 T(2) Notes and references The results of the present section are standard . Then we must find phasetype distributions Bn with B. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). cf.. elementary) Let {bk} be any dense sequence of continuity points for B(x).2) } to go on (on E(2)) when { i 1) } exits. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness .} of phasetype distributions such that Bn 3 B as n + oo. we need to allow { Jt. and the closedness of the class of phasetype distributions under the formation of finite mixtures. Proof Assume first that B is a onepoint distribution. Let the support of Dn be {xl(n). say degenerate at b.xq(n)(n)}. we can assume that ID. there is a sequence {B.(bk)'.B(bk) I < 1/n for n > k. Example A5.(n) = D. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5. Then from above. with weight pi(n) for xi(n). oo). r # oo. Now we can find first a sequence {Dm} of distributions with finite support such that D. and vice versa. By the diagonal argument (subsequent thinnings)..(Sn) with Sn = n/b. q(n) q(n) pi(n)a . the initial vector is (a(1) (& a (2) 0 0).8.n = I:pi(n)Er v ( __ ) n) ) a= 1 ..14 To a given distribution B on (0. Hence it is immediate that Bn 4 B. see Neuts [269] (where the proof. any distribution B on (0.(bk) + B(bk) for all k. The general case now follows easily from this. and let Bn be the Erlang distribution E. i= 1 C.356 APPENDIX For U1 V U2. That is.
Corollary A5. if information on Bo is given in terms of observations (i.. E E. k < n.n. Then ICr( n ). i.B(bk )I < . and we can take Bn = Cr(n). i = 1. It should be noted.. oo) and any fl. however..( dx) * f r f{(x)B(dx). . But To is the class G of all distributions on [0. Hence G C PET and L = PIT.i. PIT contains all finite mixtures of onepoint distributions. the class CO of all discrete distributions. then it is immediate that WI(B) = p2(B) for all distributions B on [0.14 is fundamental and can motivate phasetype assumptions. Let E be the class of functions f : [0.n (bk) .. oo) such that f (x) = O(e«x).D(bk)I < n. one would use the B given by some statistical fitting procedure (see below). In particular. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. 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.d. 2.n( b k ) . and that cp is known to be continuous. we can then approximate Bo by a phasetype B.. oo). u 2 (abstract topological ) The essence of the argument above is that the closure (w. say on the claim size distribution B in risk theory..(x)Bf. for some a < oo.t. For a general Bo.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). oo) approximation Assume that we can compute a functional W(B) when B is phasetype.. replications)..15 To a given distribution B on (0 . x 4 oo. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. f2. If Cpl (B) and ^02(B) are weakly continuous.e. k < n. Since PET is closed under the continuous operation of formation of finite mixtures.r. u Theorem A5. compute W(B) and use this quantity as an approximation to cp(B0). oo) * [0. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large.
.. and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ).f (z) = f = 1 1 1 1n/ o . . and hence we may choose r(n) such that L 9l) f (x)Cr(n).. By (A.. TO (A.. Now returning to the proof of (A..39) Indeed. \\ 0 Corollary A5.38) We first show that for each f E E. for each i. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. i = 1.  APPENDIX B implies that 00 o o 00 n. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. . liminf B. . n.n(dx) < 1+. f° xtBn(dx ) * f °° x`B( dx). i = 1. i = 1.n(dx) + f 0 fi(x)Dn(dx).. oo)..358 Proof By Fatou' s lemma. i=1.14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx)..39).2 .. and the case of a general f then follows from the definition of the class E and a uniform integrability argument.38 ). n.f (x)B(dx).... n B=az.(dx) > J fi(x)B(dx). then cc f (x)Bn ( dx) = (?!c ) e'= . 2.oo J fi(x)B.16 To a given distribution B on (0 .. . if f (x ) = e°x. we may assume that in the proof of Theorem A5.f ' f (x)B(dx). f00 fi(x)Cr.
The adjustment coefficient is a fundamental quantity. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. ..17 To a given /3 > 0 and a given distribution B on (0. (N.> y for some sequence {ei} with ei E (0. the remaining results may be slightly stronger than those given in the literature.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . . the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. from a more conceptual . the adjustment coefficient 'y = 7(B./3).l3µb < 1. . the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. I. but are certainly not unexpected. Proof Let fi(x) = el'r+E. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions.14 is classical./3) is defined as the unique solution > 0 of B[y] = l+y/j3. (N or a given distribution Bo.16.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn. .e. lim inf > is proved similarly. e ) and ei J. . 0 as i * oo. Notes and references Theorem A5. one can obtain 7(Bn.18 In the setting of Corollary A5.. oo) with B[y +e] < oo for some e > y = 7(B. /3) = ry for all n. However. The present section is a survey of some of the available approaches and software for inplementing this.. there is a sequence {B. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . For practical purposes. . 5e Phasetype fitting As has been mentioned a number of times already. If ei > 0.3). O We state without proof the following result: Corollary A5. lim sup ryn < 7. .
and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. [317] ) has considered an extension of this setup.. the L1 distance between the c . we have constructed a sequence { B. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . A method developed by Bobbio and coworkers (see e. d.d. The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter.} of phasetype distribution such that Bo. at a a number of selected points . the number of phases required for a good fit will typically be much larger. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly..g.'s). cf. we do not not want to perform matrix calculus in hundreds or thousands dimensions).. and in practice this sets a limitation to the usefulness (the curse of dimensionality . giving mass 1 /N to each S=. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . e . [202]. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . 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. . reliability or queueing theory. (N is the empirical distribution Be. and used a nonlinear programming approach ..f. The characteristics of all of these methods is that even the number of parameters may be low (e. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. defined by the absence of loops in the phase diagram .g. three for a mixture of two Erlangs ).g . Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments .360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . [216] ). and as fitted distribution we may take B.f. Schmickler (the MEDA package. In a series of papers (e. Of course. . one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). risk theory. for some suitable large n. g. [70]) restrict attention to acyclic phase type distributions .g. The constraints were the exact fit of the two first moments and the objective function to be minimized involved the deviation of the empirical and fitted c. a program package written in C for the SUN workstation or the PC is available as shareware. It seems therefore a key issue to develop methods allowing for a more general phase diagram.
then the estimators would be of simple occurenceexposure type. EN where ai = N 1 I ((k) = i) tii=i iEE.. (N) tJk Ea ( n). In practice. Thus. E. . since this is parameterdependent. Nii = = .T(n) (Ti ^^ 1. e. one is lead to an iterative scheme.. eieT(n)((k. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations. it seems open whether the restriction to the acyclic case is a severe loss of generality.APPENDIX 361 tion Sk was available. jEEA.. = j) f k=1 k =1 tE[0.. the methods of [70] and [38] appear to produce almost identical results.T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei .(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j)....T (n)(TiI(1. ..x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations.. . (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations.g. (n+1) _ Ea (n).g. In fact. it is easy to see that N (k Ea(n). N Ti = I(J= i) dt. .(N) = E Ea(n). .T(n) (Nik IC1.
This page is intentionally left blank .
[8] G. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Abramowitz & I. 219229. Abate. John Wiley & Sons. 101116. Arfwedson ( 1954) Research in collective risk theory. Queueing Systems 5 . Queueing Systems 16. ed. [7] E. J. G.G. 587. Aktuar Tidskr. 345368. [15] S.). Transactions XVth International Congress of Actuaries. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . New York. 37. AT&T. [9] G. Stoch. 38. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions .L. In: Limit Theorems and Related Problems (A. 311338. Choudhury & W. Aktuar Tidskr. Queueing Systems 10. Skand. ibid. 3157 . Asmussen (1987) Applied Probability and Queues. [3] J. [12] S. The Mathematical Scientist 14. Appl. Adv. Proc. Probab. Borokov. Asmussen (1985) Conjugate processes and the simulation of ruin problems. B. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. [10] K. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. Act. 57. with applications to risk reserve processes and the GI /G/1 queue. Teubner. New York. 213229.). 143170. Dover. Arfwedson ( 1955) Research in collective risk theory. [11) S. [14] S. 363 . II. Abate & W. [13] S. 1985. Asmussen (1984) Approximations for the probability of ruin within finite time . 191223. 253267. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. Preprint. pp. [4] M. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . 20. 14.A. Skand.Bibliography [1] J. New York. Appl. The case of equal risk sums. [6] V. [2] J. [5] G. Optimizations Software. 53100. Stuttgart. Alsmeyer (1991) Erneuerungstheorie. Abate & W. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . 1984 . Scand. Chichester New York.
Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. 297318. Th. 106119. [27] S. Stoch. Appl. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Asmussen & K. Ann. J. [26] S. Asmussen (1992b) Light traffic equivalence in single server queues. Proc. Stoch. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. New York. Hojgaard (2000) Rare events simulation for heavytailed distributions. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. 20. [34] S. Probab. Dshalalow ed. [33] S. 189201. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. 555574. Advances in Queueing: Models.).M. 772789. Bernoulli 6. Random Processes 2. Asmussen & B. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Stochastic Models 11. 2149. SIAM Review 40. A. [30] S. J. Andronov. MatrixAnalytic Methods in Stochastic Models (A. 1989 . Asmussen (1992a) Phasetype representations in random walk and queueing problems. eds. Act. Probab. [19] S. Asmussen (1999) On the ruin problem for some adapted premium rules. Marcel Dekker. Chakravarty. Methods V Problems (J.). 96107. 137168.364 BIBLIOGRAPHY [16] S. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. J. Scand.). Asmussen & M. 354374. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. [20] S. Asmussen (1995) Stationary distributions via first passage times. Asmussen & B. Probab. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. [31] S. Asmussen.S. 4966. [32] S. Appl. Ann. Probabilistic Analysis of Rare Events (V. 193226. Frey. 1996. Rolski & V. Floe Henriksen & C. Act. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. stationary distributions and first passage probabilities. Asmussen. Ann. 25. Alfa & S. [22] S. Scand. L. Act. 2943. [24] S. Kalashnikov & A. Scand. Asmussen (1989b) Risk theory in a Markovian environment. Binswanger & B. [17] S. . [21] S. 1936. Asmussen. K. Boca Raton. [28] S. CRC Press. Riga Aviation University.K. 79102. 315. 1999 . [18] S. 27. Asmussen & M. [25] S. Scand. 8. Proc. Statist. Appl. Asmussen (1998c) A probabilistic look at the WienerHopf equation. Astin Bulletin 27. [29] S. 313341. Asmussen (2000) Matrixanalytic models and their analysis. 54. 2. J. 313326. 303322. T. eds. 37. Extremes 1 . Florida. 69100. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. [23] S. Appl.
J. 20. Asmussen & H. Fitting phasetype distributions via the EM algorithm. Statistica Neerlandica 47. Sc. Asmussen & T. 365372. Insurance: Mathematics and Economics 20. Asmussen & C. Schmidt (1995) Ladder height distributions with marks. Asmussen & R. Finance and Stochastics 4. 11251141. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. first passage problems and extreme value theory for queues. with applications to insurance risk Stoch. Sigman (1996) Monotone stochastic recursions and their duals. Taksar (1997) Controlled diffusion models for optimal dividend payout. Asmussen & K. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. 259274.M. Koole (1993). Oper. Read eds. Nerman & M. (to appear). Probab. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Adv. Eng. [52] S.BIBLIOGRAPHY 365 [35] S. CRC Press. [40] S. [49] S. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Proc. Dshalalow ed. [50] S. Math. Prob 32 . 299324. Asmussen. Supplementary Volume (Kotz. Asmussen & R. 419441. 421458. 23. Kliippelberg (1996) Large deviations results for subexponential tails. Rubinstein (1999) Sensitivity analysis of insurance risk models. Stochastic Models 8. 58. Asmussen & M. Appl. Adv. [51] S. Methods 8 Problems (J. Appl. Appl. Schmidli & V. [47] S. Res. O. 429466.A. Asmussen & C. 10. Asmussen & D. J. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. Appl. 19. H. Boca Raton. Th. Proc. Management Science 45. Asmussen . 30. Perry (1992) On cycle maxima. 105119. Hojgaard & M. Wiley. Appl. Probab.A. Scand. Advances in Queueing: Models. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. 913916. Asmussen & V. J. Florida.). 120. Insurance: Mathematics and Economics 10. [41] S. Asmussen & G. [44] S. Inf. 64. 31. [36] S. Marked point processes as limits of Markovian arrival streams. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue.Y. [39] S. Taksar (2000) Optimal risk control and dividend distribution policies . 410433. 103125. . [43] S. [46] S. [38] S. [37] S. Asmussen. 115. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. [48] S.). Example of excessoffloss reinsurance for an insurance corporation.Y. Asmussen & T. [42] S. Asmussen & C. Asmussen & V. Stoch. Asmussen & S. Res. Probab. Appl. Johnson. Opns. B. Probab. 422447. [45] S. 736755. Olsson (1996). Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Statist.
Wiley. J. Boogaert & V. 129134.H. 190204. [56] B. Boogaert & A. Athreya & P. Berman & R. Act. J. J. [67] T.R. New York.L. 4148. [63] A.366 BIBLIOGRAPHY [53] S. J. Wiley. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. 1995. Appl. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. Scand. Appl.B . [65] N. Bingham.A. Insurance: Mathematics and Economics 9. Probab. Act. 275279. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Scand. de Waegenaere (1990) Simulation of ruin probabilities. 11811190. Beekman (1974) Two Stochastic Processes. 77111. Teugels (1987) Regular Variation. Asmussen & J. Insurance: Mathematics and Economics 4. [62] J. [55] B. [68] T. Borovkov (1976) Asymptotic Methods in Queueing Theory. [69] P. J. Stochastic Models 10. [72] P.H. 148156. Cambridge University Press. [57] C. [58] O. SIAM. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. J. Bloomfield & D. 1974. 705766. 221232. SpringerVerlag. Act. Goldie & J. Probab. Bjork & J. Appl. Ney ( 1972) Branching Processes . Halsted Press. New York. [71] P. Soc. Beekman (1985) A series for infinite time ruin probabilities. Actuaries 21. BarndorffNielsen & H. Baker (1977) The Numerical Solution of Integral Equations. Trans. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. New York. Beekman (1969) A ruin function approximation. Scand. 1988 .J. [73] A. Probab. 9. Clarendon Press. 661667. 1975. Billingsley (1968) Convergence of Probability Measures. [59] O. [64] P. 9599. Scand. [60] J.A. [54] K. 7.L.M. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. 832840. 1985 . .H. Bjork & J. Bobbio & M. [70] A. Act. Act. 610. 169186.T. Insurance: Mathematics and Economics 6. Adv. Bingham (1975) Fluctuation theory in continuous time. Oxford. Scand. Berlin. Grandell (1985) An insensitivity property of the ruin probability. Springer . 33. Cox (1972) A low traffic approximation for queues. Cambridge. [66] N. J. C. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. [61] J.
New York.R. Resnick & R. Cohen (1982) The Single Server Queue (2nd ed. Aut. Veraverbeke (1990). Qinlar (1972) Markov additive processes. [90] D. SpringerVerlag. Berlin. IEEE Trans.J. 749763. Stockholm. Res. 62.L. Fox & L. AddisonWesley. . Proc.W. Tweedie (1982) Storage processes with general release rule and additive inputs . [77] P. Itasca. IEEE J. [76] N.L.M. Hickman. Reiser eds. Bowers . Herzog (1977) The phase concept : approximations of measured data and performance analysis. Nesbitt (1986) Actuarial Mathematics .W. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. J. 907920.L. [87] E. 3539. Cramer (1955) Collective risk theory. Springer.A. Bratley. Appl.Y. Area Commun. [83] D.BIBLIOGRAPHY 367 [74] O. verve. Brockwell . Chandy & M. Chung (1974) A Course in Probability Theory (2nd ed. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Croux & N. Jr. Burman & D. M. [93] K. [81] J. Geb. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions .J. Broeck . NorthHolland. 313319. 127130. Illinois. Fort & G. [80] F. Biihlmann ( 1970) Mathematical Methods in Risk Theory. Boxma & J. 105119. Wahrscheinlichkeitsth. Schrage (1987) A Guide to Simulation. Insurance : Mathematics and Economics 9. D. 93121. [86] K.W.L. [79] P. [92] H.C. Cambr. Computer Performance (K. J. Jones & C. II. Soc. [89] M. New York San Francisco London. Philos. 392433. J. 24.). Tech. 51. Oper. Insurance : Mathematics and Economics 5. [82] H.U. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Control AC28. Breiman ( 1968) Probability. Simulation and Estimation.R. Stockholm.) NorthHolland. H. Cottrell. B. Goovaerts & F. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Boxma & J. Queueing Systems 33. Burman & D . [85] W. 14331453. Cramer (1930) On the Mathematical Theory of Risk. Academic Press. Adv.I. Wiley. Z. The Jubilee volume of Forsakringsbolaget Skandia. [91] H. Reading. Syst. Skandia Jubilee Volume. Nonparametric estimators for the probability of ruin. Amsterdam. New York. Bucklew ( 1990) Large Deviation Techniques in Decision. Probab.C. Cox (1955) Use of complex probabilities in the theory of stochastic processes. 2338. The Society of Actuaries . [78] L.R. [84] D. 34. [75] O. Gerber. Bell. De Vylder (1986) Ordering of risks and ruin probabilities. Amsterdam. 177204. [88] J.Y .A. S. 14.. Bux & U. 16.). Sel.
1990 . 435436. [110] F. 624628. Haezendonck (1985) Inversed martingales in risk theory.. Dempster. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). Laird & D. [98] A. De Vylder (1977) A new proof of a known result in risk theory. Res. Oper. [107] L. Comp. 3. Devroye (1986) Nonuniform Random Variate Generation. T. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Daykin. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . Scand. 85116. [109] F. Insurance: Mathematics and Economics 3. [102] P. N. [112] F. 201206. Math. Statist. Scand. 5771. Chapman & Hall. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. 135159. [103] F. Pentikainen & E. [108] F. Deheuvels & J. 181217. Oper. Springer. Probab.B. 3750. Appl. Embrechts (1989).M. Dembo & O. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . 88101. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. Csorgo & J. Boston. [111] F. 17. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. 227229. Csorgo & J. . van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. [99] R. [95] M. 7083 (1969). Actuarial J. De Vylder & M. Suppl. Insurance: Mathematics and Economics 6. Goovaerts (1984) Bounds for classical ruin probabilities. Daley & T. 22. [96] D. 121131. Editions de 1'Universite de Bruxelles. Soc. Tidskr. Pesonen (1994) Practical Risk Theory for Actuaries. Stochastic Models 5. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Rolski (1984) A light traffic approximation for a single server queue. 27. [105] A. 114119. [101] C. Act. Jones and Bartlett. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. De Vylder (1996) Advanced Risk Theory. [97] D. Delbaen & J. English version published in Skand. J. Res. Insurance: Mathematics and Economics 4. [106] A. [100] A.P.. De Vylder & M.D. Stockholm. Insurance: Mathematics and Economics 10. 16. Appl. Zeitouni (1992) Large Deviations Techniques and Applications. Math. 583602. J. Aktuar. Haezendonck (1987) Classical risk theory in an economic environment. [104] F. Dassios & P. Daley & T. Martingales and insurance risk. Math. 9 . Rolski (1991) Light traffic approximations in queues.L. Insurance: Mathematics and Economics 7. J.368 BIBLIOGRAPHY [94] M. Roy. Delbaen & J. J. New York.
C. Oper.M. Gerber & E. J. Emanuel . with applications. [120] D. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. 191207. Philos. Dickson & H. J. and the amount of claim causing ruin. 229232.R. Harrison & A. Springer . 1984 . [127] F. J.C. [125] F. 177192.R.M. Dufresne & H. 4962.R.U. Berlin Gottingen Heidelberg.C. Insurance: Mathematics and Economics 7. Scand. Dickson (1992) On the distribution of the surplus prior to ruin .C. 118. 1975 . [121] D. Dickson & C. Dickson (1994) An upper bound for the probability of ultimate ruin. Camb. 174186. 193199. 269274. Scand. [124] N. Letters 7. Insurance : Mathematics and Economics 10. 4259. J. J.C. Dickson & C. [123] E. 147167.U. [128] E. Scand. [119] D. [115] D. Proc. Scand. 131138. Duffield & N. Res.M. H. [116] D.M. Act. 3745.J. Astin Bulletin 21.R. Embrechts . British Actuarial J.C. 1993 . Scand. Act. Act. Dickson & J. Waters (1999) Ruin probabilities wih compounding. 1984 . O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue.M. Djehiche (1993) A large deviation estimate for ruin probabilities. J. [117] D. [114] D. J. van Doorn & G. Dufresne. 1994 . 363374.U. [118] D.C. 5159. Insurance : Mathematics and Economics 7. Waters (1996 ) Reinsurance and ruin . Embrechts ( 1988 ). Gerber (1988) The surpluses immediately before and at ruin. Math.G. Scand. Hipp (1999) Ruin problems for phasetype(2) risk processes. Scand. Regterschot (1988) Conditional PASTA. 1741. Insurance : Mathematics and Economics 11.BIBLIOGRAPHY 369 [113] D.M. 1993 . Grandell & H.C. Shiu (1991) Risk theory with the Gamma process.M. Dickson (1995) A review of Panjer's recursion formula and its applications.B. Ruin estimates for large claims . J. Act. Act.J. J. 6180. Insurance : Mathematics and Economics 22.M. 1.S. Act. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. Hipp (1998) Ruin probabilities for Erlang (2) risk processes.W. Insurance: Mathematics and Economics 25. Insurance: Mathematics and Economics 19.C. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. . [126] F. 2000 . 251262. Dynkin (1965) Markov Processes I. Dufresne & H. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. Act. [129] D .C.M. Dickson & H.K. [122] B. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . 107124. 105115. [130] P.M. Dickson & J. Soc. [131] P.
Probab. J. Wiley. Ann. Scand. Appl. 35. Mikosch (1991). Nyt Tidsskrift for Matematik B20.D. Wiley. J. 1998 . Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . Frees (1986) Nonparametric estimation of the probability of ruin. [148] C. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. 269274. 159178. Esscher (1932) On the probability function in the collective theory of risk.). Statistica Neerlandica 47. Appl. 5974. Trans. 8190. F. Fuh & T. Esary. [138] P. [137] P. Lai (1998) Wald's equations.D. Schmidli (1994) Ruin estimation for a general insurance risk model. [150] H. with applications. Akt. Fitzsimmons (1987). 3339. Embrechts & N.). 26. 695712. On the excursions of Markov processes in classical duality. Danish Academy Tech. R. 29. Th. 175195. [141] F. Erlang 2. Heidelberg. [143] W. Frenz & V. Math. Furrer (1998) Risk processes perturbed by astable Levy motion.A. A bootstrap procedure for estimating the adjustment coefficient. Appl. Fields 75. [149] C. Probab. C. Embrechts & T. Walkup (1967) Association of random variables. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. [142] W.L. 5572. 14661474. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. [133] P. [140] J. Proschan & D. Insurance: Mathematics and Economics 7. Schmidt (1992) An insensitivity property of ladder height distributions. Embrechts. Embrechts & J. Probab. [144] P. Probab. Insurance : Mathematics and Economics 1. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Skand. Maejima & J. Appl.W. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Act. 17. . Embrechts & H. Teugels (1985). Astin Bulletin 16. New York. 38. Konig. New York. Villasenor (1988) Ruin estimates for large claims. Kliippelberg & T.K. Schmidt (1982) Queues and Point Processes.J. first passage times and moments of ladder variables in Markov random walks. [139] A. Embrechts. Embrechts. J. Rel. Statist. J. [134] P. [135] P.L. Sci. [145] P. Arndt & V. Springer. D. Approximations for compound Poisson and Polya processes. 131137.W. 623637. 181190.K. Tidsskr.D. 566580. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. [136] P. Insurance: Mathematics and Economics 10. Franken. Appl. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. 404422. Griibel & S. Wiley. Adv. 616624. Adv.370 BIBLIOGRAPHY [132] P. Adv. Probab.M. 5975. Jensen. U. Probab. [147] M. [146] E. M. 29.
131156. [165] M. NorthHolland. ETH Zurich. Gerber (1986). Gani. University of Pennsylvania. contained in the authors PhD thesis.J. Kaas. [154] H. Birkhauser . Vers. Insurance : Mathematics and Economics 7. Insurance: Mathematics and Economics 20. J. Suppl.J. Probab. Glasserman & S. R. 205216. Appl. Gerber (1979) An Introduction to Mathematical Risk Theory. 1523. Dordrecht. [161] P. Griibel (1996) Perpetuities with thin tails. 1978 . J. Furrer. eds. Goldie & R. 3752. Probab. de Vylder & J. Galambos & J. Appl. [159] H. 205210. [166] M. J. Skand. Goovaerts. Math.U.U.M. NorthHolland. Grandell (1977) A class of approximations of ruin probabilities. Unpublished. 73. Amsterdam. [163] P. Ver.). Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. 1981 . Scand. Aktuarietidskrift 1970 . Basel. Ver. 5. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Gerber (1988) Mathematical fun with ruin theory.. 105115. Insurance: Mathematics and Economics 15.U. Math.BIBLIOGRAPHY 371 [151] H. . Scand. J. Life Insurance Mathematics. 2336. A.N. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'.G. [169] J. 1977. Schweiz. Furrer & H. 463480. Boston. Vers. Bauwelinckx (1990) Effective Actuarial Methods.E. [152] H. Huebner Foundation Monographs. Haezendonck (1984) Insurance Premiums. Mitt. Michna & A.. Kou (1995) Limits of first passage times to rare sets in regenerative processes.U. Z. Ann. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Weron (1997) Stable Levy motion approximation in collective risk theory. Act. Glynn & W. 28. van Heerwarden & T. Springer. Mitt. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. [158] H. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. London. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. In Studies in Applied Probability (J.U. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. Act. Goovaerts. 31A. 7778. S. [156] H.W.71. Adv. Gnedenko & I. Scand. 97114.S. Probab. Kluwer.V. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier.U. Gerber (1973) Martingales in risk theory. [167] C. Appl. [157] H. F. 424445. 6370. [160] H. [155] H. [164] B.U. [153] H. [162] P. [168] J. Schweiz. Amsterdam. Act.).
347358. ed. Statistical Algorithm 265. Grandell ( 1992) Finite time ruin probabilities and martingales . Theory and Applications.243255. Probab. SpringerVerlag. [176] B. Informatica 2. 54. Res. Stoch. Proc. 8. [177] B. Proc. Math. Appl. [188] J. Opns . 7590. 197214. 14. Opns .and Sozialforschung 6. 6779. 143158. [174] J. 566576. Lemoine (1977) Limit theorems for periodic queues. [171] J. Chapman & Hall. Harrison & A. mathematische Wirtschaft . Segerdahl (1971) A comparison of some approximations of ruin probabilities.131135.). [181] D. J. Math. [173] J. 3041. Resnick (1977) The recurrence classification of risk and storage processes . [189] A. M. Gusak & V. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment.I. New York. Skand. M. Res. [182] A. 355365. [172] J. Probability Theory and Mathematical Statistics (I. [175] J. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . [187] J. Insurance: Mathematics and Economics 26. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . Proc . Appl. Aktuar Tidskr. Grandell (1997) Mixed Poisson Processes. Astin Bulletin 29. Winter School on Stochastic Analysis and Appl . [179] R. E.I. Grandell (1999) Simple approximations of ruin functions. Preprint. Liet. Harrison & S. Grandell (1990) Aspects of Risk Theory. AkademieVerlag. [183] M. Appl. 3 . Appl. 33. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. [185] J.A. Th. Probab. Ibragimov. Archiv v. [184] H . 332. Berlin.. M. Gyllenberg & D. 167176. Applied Statistics 40. Grandell (1979) Empirical bounds for ruin probabilities.S. Tinbergen Institute Research Series 20. Harrison (1977) Ruin problems with compounding assets . Griibel & R. Grigelionis (1996) Lundbergtype stochastic processes. Matem. 5766. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . 400409. [180] R.372 BIBLIOGRAPHY [170] J. 5 . KTH. Amsterdam. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. Grnbel (1991) G/G/1 via FFT. Harrison & S. M. Stoch. Rink. SpringerVerlag.V. . 1. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications.O. Grandell & C. [186] J. [178] B.J. Gut (1988 ) Stopped Random Walks . 14.
285292. 8095. Nachrichten 30. [207] D. [201] L. J. Act. eds. Probab. Ann. Hipp (1989a) Efficient estimators for ruin probabilities. . Act.. Iglehart (1969) Diffusion approximations in collective risk theory. [193] U. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Hogg & S. [203] T. Geb. ACM TOMACS 6.a program for fitting phasetype distributions.L. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve.A. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. [196] C. Versicherungswirtschaft. J. on Asymptotic Statistics (P. 25. Nerman (1992) EMPHT . Ann. Insurance: Mathematics and Economics 5. Klugman (1984) Loss Distributions. Asmussen & 0. Heilmann (1987) Grundbegriffe der Risikotheorie. Scand. 285293. Proc. Hipp & R. Horvath & E. J. [206] B. [204] T. verve. Probab. 18. 166180.V. [202] 0. 5770. [197] C. Studies in Statistical Quality Control and Reliability 1992 :4.M. 8996. Karlsruhe. 3. 378389. J. [198] C. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. 4385. Iversen & L. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. [195] B. Appl. Hoglund (1991). [192] U. Wahrscheinlichkeitsth. 11631174. Statist. 305313. 259268. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden.BIBLIOGRAPHY 373 [190] P. 123151. Ann.V.). Probab. Karlsruhe. 1990 . Z. Hojgaard & M. [208] V. S. Electronic Letters 35. Hill (1975) A simple general approach to inference about the tail of a distribution. Appl. [191] W. Wiley. Math. Chalmers University of Technology and the University of Goteborg. [205] T. The ruin problem for finite Markov chains. 23. 19. [194] 0. Mandl & M. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. New York. Verlag Versicherungswirtschaft e. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. [199] R. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Mathematical Statistics. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. No. Insurance: Mathematics and Economics 5. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. Huskova. 1998 .B.L. Haggstrom. 6. 12981310. Prob.R. Astin Bulletin 19. Scand. 29. Foruth Prague Symp. [200] M. 377381.
[213] P. 61. 116129. Jensen (1995) Saddle Point Approximations. Appl. Scand. Astin Bull. [221] E. Janssen & J. Copenhagen. 11. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. [217] R. [212] J. Proc. Act. [226] J. [214] A. Soc. 259281. 6. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories.L. ibid. Proc. Jagerman (1985) Certain Volterra integral equations arising in queueing. 325247. Cambridge Philos. Clarendon Press. 35. [222] S. NEC Research Institute. Manuscript. Kluwer. Wishart (1964).R. Kemperman (1961) The Passage Problem for a Markov Chain. 187193. J. Goovaerts (1986) General bound on ruin probabilities. 15. Amer. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. [227] J. 8793. Jelenkovic & A. Appl. Insurance: Mathematics and Economics 5. Probab. 239256. 60. Soc. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. Ann. [219] V. [215] J. 547567. Keilson & D.H.M. Oxford. University of Chicago Press. N. [223] J.B. Cambridge Philos. [224] J. 118. Stochastic Models 1.J. ibid.374 BIBLIOGRAPHY [209] D. Cambridge Philos. Soc. Jensen (1953) A Distribution Model Applicable to Economics.J. Keilson & D.P. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk.A.C.G. 165167. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application.M. Johnson & M. Karlin & H. Probab. [225] J. [228] J. J. [216] M. Kao (1988) Computing the phasetype renewal and related functions. . Proc. J. Math. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain.M. 711743. Kaas & M. Statist. Keilson & D. 4151. 173190. 3. [220] V. New York. Taaffe (1989/90) Matching moments to phase distributions. Princeton. 561563. 1996. Stochastic Models 5. Astin Bull. J. Act. Chicago. Academic Press. [211] J. Addenda to for processes defined on a finite Markov chain. N. 6.G. Technometrics 30. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. 31. Munksgaard.M. 283306. 123133. 37. 63. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens.M.G. 866870. [210] D. Kalashnikov (1996) Twosided bounds of ruin probabilities. Taylor (1981) A Second Course in Stochastic Processes. [218] V.
Kunita (1976) Absolute continuity of Markov processes. Kingman (1964) A martingale inequality in the theory of queues. [245] T. Mikosch (1995) Modelling delay in claim settlement.F. Act. Stochastic Models 7.J. 25. Klnppelberg & T. Probab.J.C. Philos. Proc. [247] L. J. Proc. Kliippelberg & U. Appl. SpringerVerlag. 889900. SpringerVerlag.F. I. 48. 259264. [246] D. Roy. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . 24. Scand. 359361. Seminaire de Probabilties X . Cybernetica 4. [242] A. [233] C. Math. Latouche & V. [248] D. Soc. Probab. Philos. 1995 . 277289. 390397. 60 .BIBLIOGRAPHY 375 [229] J. Prob. Statist.. [239] H. Cambr. Insurance : Mathematics and Economics 12. Lecture Notes in Mathematics 511.F. Sorensen (1997) Exponential Families of Stochastic Processes. [237] C. Lemoine (1981) On queues with periodic Poisson input. 132141.C. Lindley (1952) The theory of queues with a single server. [243] A. Camb. 4958. Scand. Korolyuk. 146. [240] U. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. 279285.F. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. J. Quart. 154168. [236] C. J. 383392. SIAM.P. 1998 . 26. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. [238] V. Probab. . Scand.C. Act. [232] C. 6075. Kliippelberg (1988) Subexponential distributions and integrated tails. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. Lipsky (1992) Queueing Theory . Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Bernoulli 1 . Kingman (1962) On queues in heavy traffic. Actuarial J. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. J. Insurance: Mathematics and Economics 8. J. Soc. Appl. Lehtonen & H. Appl. J. Kuchler & M. B24. Appl. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. Macmillan.S. Quart. [234] C. 133135 (in Russian). J.a Linear Algebraic Approach. 858874. [241] G. Klnppelberg & T. New York. 125147. [231] J. Adv. Penev & A. [230] J. Lehtonen & H. Kingman (1961) A convexity property of positive matrices . 283284. 18. [244] T. (235] C. Soc. Oxford 12. 4477.
Appl. 834841. Probab.D. 147149. Europ. Cambridge Philos. J. 58. Opns. [252] A.D. [264] G. Anal. Schmidt ( 1993) On ladder height distributions of general risk processes . 22. . Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. Statist. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Appl. Almqvist & Wiksell. J. Res. J. a useful concept in risk theory. Nummelin (1987) Markov additive processes I. [261] H. Probability and Mathematical Statistics (A. J. 32. 763776. Soc. J. K. Hoist eds. 378406. Soc. 1986 . 4859. Lucantoni .S. 15. Scand. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. Makowski ( 1994) On an elementary characterization of the increasing convex order . [266] P. 3.V.K. Englunds Boktryckeri AB. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. Scand.D. Math. J. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. Act. Act. Proc. Ney & E. 429448. [253] V. 29 . Michna ( 1998) Selfsimilar processes in collective risk theory. Cambridge Philos.S. MartinL3f (1986) Entropy. J. Math. 2. Probab. [257] A. 676705. Appl. 268285. Appl. Proc. Uppsala. Probab. 1994. MeierHellstern & M . 1261 (0?)1270. F. with an application . [262] H . Stockholm. 286298. [256] A. .M. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . [263] M . [265] S. Miyazawa & V. MartinL6f (1983) Entropy estimates for ruin probabilities . Eigenvalue properties and limit theorems . Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Act. Appl. Meier ( 1984). 11 . Scand. 124147. Th.F.K. 223235. Nagaev (1957) Some limit theorems for stationary Markov chains. Probab. Ann. [250] F. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. Insurance : Mathematics and Economics 5. Appl. Adv. 36. [255] V. II Aterforsdkring av Kollektivrisker . Ann. Stoch. [251] F. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. 161174. [254] V . 31. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . [258] K . Ann. [259] Z . [260] H.V. 1996 . Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes .). 2939. 561592. Gut & J. Probab . 58 . Probab. 370377.376 BIBLIOGRAPHY [249] D.
Baltimore .F. Neuts (1977) A versatile Markovian point process. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. [271] M.A. Norberg ( 1990) Risk theory and its statistics environment . Omey & E. 249266. 733746. Neuts (1992) Models based on the Markovian arrival process. Neuts ( 1978) Renewal processes of phase type. Appl. 135148. [279] E. [275] H. Proc. Appl. . [269] M . [270] M . 30. 123144. Proc. New York. Paulsen & H. Appl. Johns Hopkins University Press. [280] A. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. E75B . Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. 764779. Adv. 316. 445454. Marcel Dekker. Paulsen (1998) Ruin theory with compounding assets . [273] R. O'Cinneide (1990) Characterization of phasetype distributions. Probab. Nyrhinen (1999) Large deviations for the time of ruin.G. 46. 311342.F. Probab. Math. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. Appl.F. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications.K. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. 215223. 71. Stoch. Commun. 36. [284] J. Appl. Paulsen & H. J. 21. Abh. 157. Stoch. Insurance: Mathematics and Economics 22. [281] J. Omey & E. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Naval Research Logistics Quarterly 25.F. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Stoch. London. [274] H . J. 555564.K. [278] E.BIBLIOGRAPHY 377 [267] M . Appl. 273299. Insurance: Mathematics and Economics 20. Appl. 12551265. 75. [285] J. 16. Sem.F. Astin Bulletin 5 . Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . [277] J. [272] J . Probab. Paulsen (1993) Risk theory in a stochastic economic environment. J. 339353. Proc. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Statistics 21. 12. IEICE Trans. Appl. 10081026. Proc. 327361. Models 6. [282] J . Hamburg. Probab. [268] M . Stochastic Models 3 . Pakes (1975) On the tail of waiting time distributions. Stoch.a survey. Stoch. [283] J. [276] C.
147159. 537555. New York. 337347. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. [289] C. Berlin. Res. Philipson (1968) A review of the collective theory of risk. [302] J. 4568. Inst. Math. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services.M. 46. Heidelberg. Ann. 30. Queues. Adv. [292] S. Queueing Systems 5. [299] C. 691707. [287] F.L. Math. Adv. Tidskr. Stat. Rao (1965 ) Linear Statistical Inference and Its Applications. 757764. Resnick & G. Act. R. J. Appl. Prabhu (1961). [297] R. [301] G. Wiley. 1989 . [295] N. Math. Probab . [298] V. Prabhu (1965) Queues and Inventories. Soc. 12. 2343. 4.378 BIBLIOGRAPHY [286] J. 222261. 215246. Pyke (1959) The supremum and infimum of the Poisson process. and Dams. 32.A. Griibel & P. [294] N. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Ann. 568576. [293] N. J.J. Probab. . Ann. Insurance: Mathematics and Economics 3. 29A. Astin Bulletin XIV.G.K. J. Pellerey (1995) On the preservation of some orderings of risks under convolution. Statist. 465483.U. Prabhu & Zhu (1989) Markovmodulated queueing systems. Wiley. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Gjessing (1997c) Ruin theory with stochastic return on investments. Res. Austr.U. 19. [288] S. 2330. Statist. Appl. [290] E.R. [305] C. Skand. [291] S. Insurance: Mathematics and Economics 16. Insurance Risk. 117133.J. Paulsen & H. Math.. 28. Appl. Adv. Appl. Aktuar. Pitman (1980) Subexponential distribution functions. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. Embrechts (1996) Confidence bounds for the adjustment coefficient. Ripley (1987) Stochastic Simulation . Pitts.U. Probab. [306] T. Wiley. On the ruin problem of collective risk theory. Regterschot & J. 29.K. [303] S.G. 235243. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. 139143. Oper. Prabhu (1980) Stochastic Storage Processes. Opns. 61. Appl. 390413. [304] B. [296] N.H. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Probab. 965985. Scand.M. Math. 11. [300] C. Ann.M. New York.U. Probab. Springer. Rolski (1987) Approximation of periodic queues. 820827.
L. Wiley. Melamed (1998) Modern Simulation and Modelling. Data Anal. 431447 [316] S. . 417421. [323] H. Statist.J. Act. 11. 21. [324] H . Ann. Science Culture Technology Publishing . Rolski. 687699. Statist. Probab. Appl. Insurance: Mathematics and Economics 16. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . Scand. [322] H. Insurance: Mathematics and Economics 22. Schlegel ( 1998) Ruin probabilities in perturbed risk models. Operationsforsch. 93104. Math. Schmidli ( 1999b) Perturbed risk processes: a review . Wiley. [310] R. 145165. 5. Wiley. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. 4068. Schmidli. J.Y. H. [313] M. J. [311] R. Probab. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . 155188. [308] S. 131156. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Wiley. Rubinstein (1981) Simulation and the Monte Carlo Method. [321] 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. 365388.M. Adv. Appl.BIBLIOGRAPHY 379 [307] T. Stochastic Models 6. Teugels (1999) Stochastic Processes for Insurance and Finance. 1997. Scand. Probab . 4857. Rubinstein & B. V. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Rubinstein & A. Appi. 2001 . Rudemo (1973) Point processes generated by transitions of a Markov chain. Schmidt & J. Astin Bulletin 29. [318] H. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . Act. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. 5. [317] L. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. [326] H. 135149. Seal ( 1969) The Stochastic Theory of a Risk Business. Stoch. Schmidli ( 1996) Martingales and insurance risk. Singapore.Y. Comp. 5. 121133. 7. J. [320] H. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. 262286. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . [315] T. 795829. [314] T. 227244. Wiley. Th. Stochastic Models 10 . Proc. [319] H. Stochastic Models 10. [309] H. [312] R. Ross (1974) Bounds on the delay distribution in GI/G/1 queues.Y. [325] H .
Probab . Chapman & Hall. Th. [341] E. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. Seal ( 1972). SpringerVerlag. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. [345] D. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. Shin (1989) Ruin probability by operational calculus. Math. t). Act. Aktuar Tidsskr. [348] K .the Harald Cramer volume (ed. [342] D. Seal (1974) The numerical calculation of U(w. . Chapman and Hall. [328] H. Queueing Systems 11. t). Stochastic Models 6. 21. 75. Statist. 673684. 914924. Scand. 2236. Aktuar Tidsskr. J. Math. Probab. [338] E. 77100. 429442. Segerdahl (1959) A survey of results in the collective theory of risk. Mitteil. Sigman (1992) Light traffic for workload in queues. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. [332] C : O. Act. 1955. 4. Scand.A. 1974. Versich. [333] C: O. [330] H. [340] E. 72.L. 61. Siegmund (1975) The time until ruin in collective risk theory.S. 701719. Shantikumar (1993) Stochastic Orders and Their Applications. [331] G.S.S. 11. Ann. Shwartz & A. Shaked & J. Adv. Math. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . 191197. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. 72. Mitt. 11.L. 4. 1987. Seal (1978) Survival Probabilities. 4383. Stockholm. Verein Schweiz. [344] D. [334] C: O. Wiley. Probab. Versich. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. 243249. 483487. Shin (1987) Convolution of uniform distributions and ruin probability. Adv. Academic Press. Risk teory and the single server queue .W. . 383413. Almqvist & Wiksell. 121139. [329] H. Sengupta (1990) The semiMarkov queue: theory and applications. the probability of nonruin in an interval (0. [336] B. In Probability and statistics .F. Siegmund ( 1985) Sequential Analysis . J. Grenander). Appl. [347] K. Insurance: Mathematics and Economics 8. New York.L. Appl. [346) D . [343] D. L. Verein Schweiz. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. [339] A. [335] B . 279299.380 BIBLIOGRAPHY [327] H. pp. [337] M. Mitt.W. Seber (1984) Multivariate Observations. 159180. Probab. Verein Schweiz Versich. Wiley. 171178. 157166. Weiss (1995) Large Deviations for Performance Analysis. Skand. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Appl.G. Ann.
J. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. 725741. Straub (1988) NonLife Insurance Mathematics. [368] O. Probab. [357] B. Scand. 5776. [362] G.L. John Wiley & Sons. 8199. [354] B. [361] G. 8594. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . Springer. 722. 1980 .). Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. 114137. Insurance: Mathematics and Economics 1. Insurance: Mathematics and Economics 19. 197208. [365] J.L. Wiley. Proc. Insurance: Mathematics and Economics 16.BIBLIOGRAPHY 381 [349] E. J. Soc 49. Adv.and largedeviations probabilities in actuarial risk theory. Sundt & W. Scand. [369] O. NorthHolland.C. Appl. Astin Bulletin 24. 1976. [367] O. [363] G. Astin Bulletin 12. Act. Thorin (1982) Probabilities of ruin.L. Sundt & J. Act. Scand. 65102. Taylor (1979) Probability of ruin with variable premium rate. 163175. J.V. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. [350] W. Act.C. Hoesman ( 1989) Moderate. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . 1982 .J. Unpublished manuscript. Thorin (1977). 65102.S. [366] O. Scand. 149162.C. 449461. 1974. Proceeedings of the IEEE 77.A. [352] D. 21.. Karlsruhe.L. Scand. J. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . [359] L. Cambridge Philos. [356] B . Verlag Versicherungswirtschaft e. 2739.). Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . . Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. [358] R. New York. Sundt & J.. 235254. Stanford & K.C. Slud & C. Actuarial J. Daley ed. [364] G. Ruin probabilities prepared for numerical calculations. Act. Act. J. Smith ( 1953) Distribution of queueing times. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Teugels ( 1995 ) Ruin estimates under interest force . [355] B . Scand.C. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. [351] D . Taylor (1986) Claims reserving in NonLife Insurance. New York. J. 1978. Teugels (1982) Estimation of ruin probabilities . [360] G. Astin Bulletin 16 . 118. [353] E.
Tacklind (1942) Sur le risque dans les jeux inequitables. Scand. Statistical Science 10. 31. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. [379] M.). Wolff (1990) Stochastic Modeling and the Theory of Queues.382 BIBLIOGRAPHY [370] 0. Van Wouve. [372] H. J. M. [384] R. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. 231246. 142. J. Reidel. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. Thorin & N. Wiley. Stationarity and Regeneration. Res. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. 37. Haezendonck eds. [378] H. D. Aktuar. Thorisson (2000) Coupling. [376] A. Insurance : Mathematics and Economics 13. F. PrenticeHall. [382] G. 1983 . Willinger. De Vylder & M. 4963.W. [383] G. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. 148164. 936952. Willmot & X. Tidskr. . [380] W. Skand. SpringerVerlag. Probab. [375] N. M. Unpublished Ph. Thorin & N. Dordrecht Boston Lancaster. [377] V.E. [374] F. University of Michigan. Waters (1983) Probability of ruin for a risk process with claims cost inflation. In: Premium Calculation in Insurance (F. 743756. Lin (1994) Lundberg bounds on the tails of compound distributions. 1942. Astin Bulletin IX. Act. [373] S. J.E. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. Appl. De Vylder. Astin Bulletin VII. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. Insurance: Mathematics and Economics 15. Sherman & D. Oper. 137153. Wald (1947) Sequential Analysis. thesis. 6785. R. [381] W. Goovaerts. [371] 0.R. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Taqqu. 5762. Submitted.
110113. 361 diffusion 3.228229.259261. 170173. 7079. 5. 283. 301 Kronecker product. 91. 9396.160167. 135. 245248. 332333 Volterra 192194.135. 227229.293294.137141. 8283 hyperexponential distribution 7. 278 gamma distribution 67. 12 CramerLundberg approximation 1617.150. 9293.272.285292. 360 excursion 155156. 117128.Index adjustment coefficient 17.242. 141144. 2425. 201 Brownian motion 3 . 9496. 180182. 207 heavytailed distribution 6. 7879. 318319 Erlang distribution 7. 196201 inverse Gaussian distribution 76.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16. 302303 diffusion approximation 17.299.200201. 3334. 80 81.249.318320 change of measure 2630.182. 1415.185187.314316.287292.and sum 221. 37. 117127 corrected 121127 duality 1314. 162164.251280 heavy traffic 76.203. 111117. 39.100. 308.346349 383 . 7179. 3436. 226. 14.9899. 248 WienerHopf 144 interest rate 190.203. 79. 7475.292293 Edgeworth expansion 113. 97. 205. 189. 218 Cox process 4.4447. 17. 239. 217. 89. 341.301 central limit theorem 60 . 316323 Bessel function 102.121129.328330. 271274. 15.281.178184.86.217. 323 Coxian distribution 147. 17.359 aggregate claims 103106. 3839. 119. 4851. 86. 40. 1819. 3032. 2526.6779. 138139.226. 5796.249250 integral equation 16 Lindley 143 renewal 64.269.307312 compound Poisson model 4. 122. 201214. 97129. 1112.
288290. 15.336339 Laplace transform 15.123. 112113.152160. see also sensitivity analysis phasetype distribution 8.285287 queue 14 .218221. 149. 306316 Levy process 3. 144.350361 Poisson process Markovmodulated 12 periodic 12.178182.139141. 35.348 terminating 215216. 269 PerronFrobenius theory 4142. 251. 108109. 25. 5758. 134135.336339 . 176185.238. 7179. 138.160161.340350 multiplicative functional 2830.297299.315 inequality 1718. 260 Lundberg conjugation 6979 . 141144. 86 periodicity 12. 52 53. 7576. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 44.161. 42. 80.240244. 38. 213214. 65. 41.339 large deviations 129.384 ladder heights 4756. 99. 245 M/G/1 13.259261. 133. 39. 234 matrixexponential distribution 240244 matrixexponentials 14. 203204.261264. 16.349 350 perturbation 172173. 134. 179 NP approximation 318320 Palm distribution 5253. 14. 145187. 6162.161164.161. 267269 Panjer's recursion 320323 Pareto distribution 910. 71.302. 227228. 96. 175 light traffic 8183 Lindley integral equation 143 process 3334. 44. 230. 132133. 3639. 171. 108 life insurance 5.287. 162.174. 16. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 4446. 154. 271274. 6970. 295.227230. nonlinear 155.304 process 2830.298299. 113114. 261264. 3947. 37. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 32. 2730.287291 INDEX matrix equation . 133. 25.275278. 38.146148.128129. 142 likelihood ratio : see change of measure lognormal distribution 9. 106108.201.234. 178 modulation 12. 203 Markov additive process 12.134135.234240. 157. 257.180. 304305 random walk 3336.108. 100. 137139.269271. 9899. 59. 35. 3947.148.215250.
292294. 131144. 9693.244250.279280 Rouche roots 158. 251. 244. see also matrixexponential distribution regenerative process 264 268. 89. 147. 4950. 3032. 189214. 281296 stable process 15. 1819. 5455. 146. 260 reinsurance 8. 60. 223226.262263. 11. 160. 186187 renewal process 131. 233234. 172173. 174. 233. 251. 279280 subexponential distribution 11. 253. 120 statistics x. 213. 191192. 222. 240. 7475. 229234. 107.273274. 123. 87. 307308.336339 workload 13. 326330 Weibull distribution 9. 280. 37. 332333 model 12. 333334 regular variation 10. 261264 reservedependent premiums 14. 152. 335336 sensitivity analysis 8693. 331336 equation 64.154157.314. 338 utility 324. 31. 141144.359361 stochastic control x stochastic ordering 18. 162. 186187 virtual: see workload rational Laplace transform 8. 168172 storage process 13. 294296 shotnoise process 314 simulation 19. 238 saddlepoint method 115117. 327 . 251280 time change 4. 8386. 12. 257.186. 260 WienerHopf theory 144.INDEX 385 waiting time 141. 177 timereversion 14. 256258. 317318 semiMarkov 147.
"This book is a must for anybody working in applied probability. I 1! Ruin Probabilities . phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory.. It is a comprehensive treatment of the known results on ruin probabilities. Special features of the book are the emphasis on change of measure techniques. the ^W A l \ i l ' ''' CramerLundberg approximation. y finite horizon ruin probabilities.com 2779 he 9 "789810ll22293211 .Advanced Series on Statistical Science & Applied Probability . for heavytailed claim size distributions). extensions of the classical compound Poisson model to allow f o r reservedependent premiums. Some i (l I JL I J r of the topics are Lundberg's inequality." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. 2 A I 11 JjVb l' i  i Yj . Markovmodulation or periodicity..g. worldscientific. P'i yfliother approximations (e.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory.Vol. exact solutions.